diff --git "a/stackoverflow_DL-related_questions_data/data15.json" "b/stackoverflow_DL-related_questions_data/data15.json"
new file mode 100644--- /dev/null
+++ "b/stackoverflow_DL-related_questions_data/data15.json"
@@ -0,0 +1,3000 @@
+{"QuestionId":52586526,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><caffe>","CreationDate":"2018-10-01T07:40:08.327","AcceptedAnswerId":52596030.0,"OwnerUserId":10345968.0,"Title":"Caffe model fails to learn","Body":"<p>I have the following convolutional model implemented in Keras, where after training for 100,000 epoch, it shows excellent performance with greate accuracy.<\/p>\n\n<pre><code>img_rows, img_cols = 24, 15\ninput_shape = (img_rows, img_cols, 1)\nnb_filters = 32\npool_size = (2, 2)\nkernel_size = (3, 3)\n\nmodel = Sequential()\nmodel.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1],\n                        border_mode='valid',\n                        input_shape=input_shape))\nmodel.add(Activation('relu'))\nmodel.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1]))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=pool_size))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(128))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(nb_classes))\nmodel.add(Activation('softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adadelta',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>However after trying to implement the same model in Caffe, it fails to train with an almost fixed loss value >=2.1 &amp;&amp; &lt;=2.6. \nHere is my Caffe prototext implementation:<\/p>\n\n<pre><code>name: \"FneishNet\"\nlayer {\n  name: \"inlayer1\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TRAIN\n  }\n  data_param {\n    source: \"examples\/fneishnet_numbers\/fneishnet_numbers_train_lmdb\"\n    batch_size: 128\n    backend: LMDB\n  }\n}\nlayer {\n  name: \"inlayer1\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TEST\n  }\n  data_param {\n    source: \"examples\/fneishnet_numbers\/fneishnet_numbers_val_lmdb\"\n    batch_size: 64\n    backend: LMDB\n  }\n}\nlayer {\n  name: \"conv1\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"conv1\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  convolution_param {\n    num_output: 32\n    kernel_size: 3\n    stride: 1\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"relu1\"\n  type: \"ReLU\"\n  bottom: \"conv1\"\n  top: \"conv1\"\n}\nlayer {\n  name: \"conv2\"\n  type: \"Convolution\"\n  bottom: \"conv1\"\n  top: \"conv2\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  convolution_param {\n    num_output: 32\n    kernel_size: 3\n    stride: 1\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"relu2\"\n  type: \"ReLU\"\n  bottom: \"conv2\"\n  top: \"conv2\"\n}\nlayer {\n  name: \"pool1\"\n  type: \"Pooling\"\n  bottom: \"conv2\"\n  top: \"pool1\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 2\n    stride: 1\n  }\n}\nlayer {\n  name: \"drop1\"\n  type: \"Dropout\"\n  bottom: \"pool1\"\n  top: \"pool1\"\n  dropout_param {\n    dropout_ratio: 0.25\n  }\n}\nlayer {\n  name: \"flatten1\"\n  type: \"Flatten\"\n  bottom: \"pool1\"\n  top: \"flatten1\"\n}\nlayer {\n  name: \"fc1\"\n  type: \"InnerProduct\"\n  bottom: \"flatten1\"\n  top: \"fc1\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 128\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"relu3\"\n  type: \"ReLU\"\n  bottom: \"fc1\"\n  top: \"fc1\"\n}\nlayer {\n  name: \"drop2\"\n  type: \"Dropout\"\n  bottom: \"fc1\"\n  top: \"fc1\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayer {\n  name: \"fc2\"\n  type: \"InnerProduct\"\n  bottom: \"fc1\"\n  top: \"fc2\"\n  param {\n    lr_mult: 1\n  }\n  param {\n    lr_mult: 2\n  }\n  inner_product_param {\n    num_output: 11\n    weight_filler {\n      type: \"xavier\"\n    }\n    bias_filler {\n      type: \"constant\"\n    }\n  }\n}\nlayer {\n  name: \"accuracy\"\n  type: \"Accuracy\"\n  bottom: \"fc2\"\n  bottom: \"label\"\n  top: \"accuracy\"\n  include {\n    phase: TEST\n  }\n}\nlayer {\n  name: \"loss\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"fc2\"\n  bottom: \"label\"\n  top: \"loss\"\n}\n<\/code><\/pre>\n\n<p>And here is my model solver (hyper-parameters):<\/p>\n\n<pre><code>net: \"models\/fneishnet_numbers\/train_val.prototxt\"\ntest_iter: 1000\ntest_interval: 4000\ntest_initialization: false\ndisplay: 40\naverage_loss: 40\nbase_lr: 0.01\ngamma: 0.1\nlr_policy: \"poly\"\npower: 0.5\nmax_iter: 3000000\nmomentum: 0.9\nweight_decay: 0.0005\nsnapshot: 100000\nsnapshot_prefix: \"models\/fneishnet_numbers\/fneishnet_numbers_quick\"\nsolver_mode: CPU\n<\/code><\/pre>\n\n<p>I believe that if i have no problem translating the model into Caffe, then it should performs the same way it do in Keras, so i think i had missed something.\nAny help would be appreciated, thanks.<\/p>\n","answers":[{"AnswerId":"52596030","CreationDate":"2018-10-01T17:27:00.517","ParentId":null,"OwnerUserId":"3802483","Title":null,"Body":"<p>poly: the effective learning rate follows a polynomial decay, to be\n\/\/      zero by the max_iter. return base_lr (1 - iter\/max_iter) ^ (power)<\/p>\n\n<p>So basically, are you sure you want to keep power set to 0.5 in\nreturns base_lr (1 - iter\/max_iter) ^ (power)? I think that might be the problem as you are decaying to minus something, try 2?<\/p>\n"}]}
+{"QuestionId":52586787,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2018-10-01T07:58:27.707","AcceptedAnswerId":null,"OwnerUserId":6933148.0,"Title":"How to solve the tensorflow's Conv2d layer error: \"self.kernel_size[i], IndexError: tuple index out of range\"?","Body":"<p>I'm running my programs with <code>tensorflow 1.8.0<\/code> and <code>keras 2.2.2<\/code> \nI have the following model architecture:<\/p>\n\n<pre><code>in1 = Input(name='in1', shape=(None, 1))\nin2 = Input(name='in2', shape=(None, 1))\nembedding = Embedding(1000, 50)\ne_in1 = embedding(in1)\ne_in2 = embedding(in2)\ncross = Dot(axes=[3, 3], normalize=False)([e_in1, e_in2])\ncross = Conv2D(1, 3, activation='relu', name=\"conv\", padding='same')(cross)\n<\/code><\/pre>\n\n<p>But arriving at the <code>Conv2D<\/code> layer, I've got the following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <blockquote>\n    <p>File\n    \"\/usr\/local\/lib\/python3.5\/dist-packages\/IPython\/core\/interactiveshell.py\",\n    line 2910, in run_code\n        exec(code_obj, self.user_global_ns, self.user_ns)   File \"\", line 1, in \n        cross = Conv2D(1, 3, activation='relu', name=\"conv\", padding='same')(cross)   File\n    \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/base_layer.py\",\n    line 474, in <strong>call<\/strong>\n        output_shape = self.compute_output_shape(input_shape)   File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/layers\/convolutional.py\",\n    line 195, in compute_output_shape\n        self.kernel_size[i], IndexError: tuple index out of range<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>I saw <a href=\"https:\/\/github.com\/titu1994\/Neural-Style-Transfer\/issues\/35\" rel=\"nofollow noreferrer\">here<\/a> solutions related to the keras version, but I don't think it's the same here, because my keras version worked well with another program that uses the same layer <code>Conv2D<\/code>.\n Anyone can give me some tips, please?\nThanks in advance<\/p>\n","answers":[]}
+{"QuestionId":52586853,"AnswerCount":1,"Tags":"<keras><tensorboard>","CreationDate":"2018-10-01T08:03:07.277","AcceptedAnswerId":null,"OwnerUserId":785041.0,"Title":"Batchnormalization nodes wrongfully linked with each other","Body":"<p>I'm training a Keras network using BatchNormalization layers and saw a strange thing looking at the TensorBoard graph. My network consists of a stack of 1D convolutions followed by BatchNormalization layers. Most of the graph seems fine, but the very first BatchNormalization layer is - according to TensorBoard - sending information to all other BatchNormalization layers. Is this normal?<\/p>\n\n<p>Here's the output of the network according to Keras <code>model.summary()<\/code><\/p>\n\n<pre><code>| Layer (type)                    | Output Shape      | Param # | Connected to        |\n|---------------------------------|-------------------|---------|---------------------|\n| pt_cloud_0 (InputLayer)         | (None, None, 39)  | 0       |                     |\n| pt_cloud_1 (InputLayer)         | (None, None, 39)  | 0       |                     |\n| conv1d_0_0 (Conv1D)             | (None, None, 64)  | 2560    | pt_cloud_0[0][0]    |\n| conv1d_1_0 (Conv1D)             | (None, None, 64)  | 2560    | pt_cloud_1[0][0]    |\n| batchnorm_0_0 (BatchNormalizati | (None, None, 64)  | 256     | conv1d_0_0[0][0]    |\n| batchnorm_1_0 (BatchNormalizati | (None, None, 64)  | 256     | conv1d_1_0[0][0]    |\n| conv1d_0_1 (Conv1D)             | (None, None, 64)  | 4160    | batchnorm_0_0[0][0] |\n| conv1d_1_1 (Conv1D)             | (None, None, 64)  | 4160    | batchnorm_1_0[0][0] |\n| batchnorm_0_1 (BatchNormalizati | (None, None, 64)  | 256     | conv1d_0_1[0][0]    |\n| batchnorm_1_1 (BatchNormalizati | (None, None, 64)  | 256     | conv1d_1_1[0][0]    |\n| conv1d_0_2 (Conv1D)             | (None, None, 316) | 20540   | batchnorm_0_1[0][0] |\n| conv1d_1_2 (Conv1D)             | (None, None, 316) | 20540   | batchnorm_1_1[0][0] |\n| batchnorm_0_2 (BatchNormalizati | (None, None, 316) | 1264    | conv1d_0_2[0][0]    |\n| batchnorm_1_2 (BatchNormalizati | (None, None, 316) | 1264    | conv1d_1_2[0][0]    |\n| conv1d_0_3 (Conv1D)             | (None, None, 316) | 100172  | batchnorm_0_2[0][0] |\n| conv1d_1_3 (Conv1D)             | (None, None, 316) | 100172  | batchnorm_1_2[0][0] |\n| aux_in (InputLayer)             | (None, 46)        | 0       | 0                   |\n| batchnorm_0_3 (BatchNormalizati | (None, None, 316) | 1264    | conv1d_0_3[0][0]    |\n| batchnorm_1_3 (BatchNormalizati | (None, None, 316) | 1264    | conv1d_1_3[0][0]    |\n| aux_dense_0 (Dense)             | (None, 384)       | 18048   | aux_in[0][0]        |\n| global_max_0 (GlobalMaxPooling1 | (None, 316)       | 0       | batchnorm_0_3[0][0] |\n| global_max_1 (GlobalMaxPooling1 | (None, 316)       | 0       | batchnorm_1_3[0][0] |\n| aux_dense_1 (Dense)             | (None, 384)       | 147840  | aux_dense_0[0][0]   |\n| concatenate_1 (Concatenate)     | (None, 1016)      | 0       | global_max_0[0][0]  |\n|                                 |                   |         | global_max_1[0][0]  |\n|                                 |                   |         | aux_dense_1[0][0]   |\n| dense_0 (Dense)                 | (None, 384)       | 390528  | concatenate_1[0][0] |\n| dropout_0 (Dropout)             | (None, 384)       | 0       | dense_0[0][0]       |\n| dense_1 (Dense)                 | (None, 384)       | 147840  | dropout_0[0][0]     |\n| prediction (Dense)              | (None, 101)       | 38885   | dense_1[0][0]       |\n<\/code><\/pre>\n\n<p>And here's (part of) the graph shown in TensorBoard <img src=\"https:\/\/imgur.com\/a\/G74uIWE\" alt=\"graph\">\n(If the image is not visible, please go to this link: <a href=\"https:\/\/imgur.com\/a\/G74uIWE\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/G74uIWE<\/a>)\nZoomed version: <img src=\"https:\/\/imgur.com\/a\/vtF3VWb\" alt=\"zoomed_graph\"> or this link: <a href=\"https:\/\/imgur.com\/a\/vtF3VWb\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/vtF3VWb<\/a> <\/p>\n\n<p>The red-outlined layer is the very first batch normalization layer I made in the network (batchnorm_0_0). I don't know much about the inner workings of a batchnormalization layer but I find it odd that it is linked to all other BN-layers, while those other BN-layers do not (they just are connected to the input\/output I assigned them).\nI'm wondering if this is a bug in my code, in keras, or maybe in TensorBoard?<\/p>\n\n<p><strong>Update:<\/strong> model's code below; it's written in a way I can easily experiment with the number of convolution layers\/filters, etc... but should be rather explanatory.<\/p>\n\n<pre><code>def _build(self, conv_filter_counts, dense_counts, dense_dropout_rates=None):\n    \"\"\"\n    Builds the model. The model will have the following architecture:\n      (1) [Per pointcloud] N 1D convolution layers (with possibly different depths) followed by BatchNormalization\n                           layers.\n      (2) [Per pointcloud] A global max pooling layer (calculating a 'global feature' of the point cloud).\n      (3) [Once] M dense layers (with possibly different amounts of neurons), optionally followed by DropOut layers.\n      (4) [Once] A final dense layer with `self.class_count` neurons and softmax activation.\n\n    Arguments:\n      conv_filter_counts: A list (length N) containing the succesive 1D convolution filter depths in (1)\n      dense_counts: A list (length M) containing the amount of succesive neurons in (3)\n      dense_dropout_rates: Optional. If specified, must be a list of length M containing the dropout rates\n                           for each corresponding dense layer specified by `dense_counts`. Individual entries\n                           can be set to None to disable dropout.\n                           If not specified, dropout is applied nowhere.\n    \"\"\"\n    inputs = [Input(shape=(None, self.pt_dim), name='pt_cloud_{}'.format(i)) for i in range(self.input_count)]\n    if self.aux_input_count &gt; 0:\n        aux_input = Input(shape=(self.aux_input_count,), name='aux_in')\n\n    if self.spatial_subnet:\n        # Predict and apply spatial transform for each pointcloud.\n        spatial_transforms = [transform_subnet(i, [64, 128, 256], [256, 64]) for i in inputs]\n        inputs_tr = [apply_transform_layer(i, tr, self.pt_dim) for i, tr in zip(inputs, spatial_transforms)]\n    else:\n        inputs_tr = inputs\n\n    global_feats = []\n    for i, input_pts in enumerate(inputs_tr):\n       x = input_pts\n\n       # Convolution stack\n       for j, c in enumerate(conv_filter_counts):\n           x = Convolution1D(c, 1, activation='relu', name='conv1d_{}_{}'.format(i, j))(x)\n           x = BatchNormalization(name='batchnorm_{}_{}'.format(i, j))(x)\n\n       global_feats += [GlobalMaxPooling1D(name='global_max_{}'.format(i))(x)]\n\n    # Concatenate features and possibly auxiliary input\n    if self.aux_input_count &gt; 0:\n        x = aux_input\n\n        # Create a dense subnetwork just for the auxiliary inpuy\n        for i, (c, d) in enumerate(zip(dense_counts, dense_dropout_rates)):\n            x = Dense(c, activation='relu', name='aux_dense_{}'.format(i))(x)\n\n        x = Concatenate()(global_feats + [x])\n    elif len(global_feats) &gt; 1:\n        x = Concatenate()(global_feats)\n    else:\n        x = global_feats[0]\n\n    # Dense stack with optional dropout\n    if dense_dropout_rates is None:\n        dense_dropout_rates = [None] * len(dense_counts)\n\n    for i, (c, d) in enumerate(zip(dense_counts, dense_dropout_rates)):\n        x = Dense(c, activation='relu', name='dense_{}'.format(i))(x)\n        if d is not None:\n            x = Dropout(rate=d, name='dropout_{}'.format(i))(x)\n\n    # Final prediction\n    prediction = Dense(self.class_count, activation='softmax', name='prediction')(x)\n\n    # Link all up in a model\n    if self.aux_input_count &gt; 0:\n        inputs.append(aux_input)\n\n    if len(inputs) == 1:\n        inputs = inputs[0]\n\n    return Model(inputs=inputs, outputs=prediction)\n<\/code><\/pre>\n\n<p>Kind regards,<\/p>\n\n<p>steven<\/p>\n","answers":[{"AnswerId":"52984880","CreationDate":"2018-10-25T08:37:17.550","ParentId":null,"OwnerUserId":"785041","Title":null,"Body":"<p>A cautious answer to my own question, @Mike, I think (hope?) this is indeed a bug on the tensorboard side as I can't explain it otherwise. <\/p>\n\n<p>I plotted the architecture using <code>keras.utils.plot_model<\/code> and this also doesn't show any links between the BatchNormalization layers.<\/p>\n"}]}
+{"QuestionId":52587151,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T08:24:03.867","AcceptedAnswerId":null,"OwnerUserId":10439898.0,"Title":"tensorflow - how to input a directory by dataset api","Body":"<p>I am new to tensorflow, and here is my situation: I have lots of folders and each contains several images. I need my training input to be folders(each time 2 folders), and each time 4 images inside a folder be selected for training.\nI have tried Dataset api, and tried to use the <code>map<\/code> or <code>flat_map<\/code> function, but I failed to read images inside a folder.\nHere is part of my codes:<\/p>\n\n<pre><code>def parse_function(filename):\n    print(filename)\n    batch_data = []\n    batch_label = []\n    dir_path = os.path.join(data_path, str(filename))\n    imgs_list = os.listdir(dir_path)\n    random.shuffle(imgs_list)\n    imgs_list = imgs_list * 4 #each time select 4 images\n    for i in range(img_num):\n        img_path = os.path.join(dir_path, imgs_list[i])\n        image_string = tf.read_file(img_path)  \n        image_decoded = tf.image.decode_image(image_string)  \n        image_resized = tf.image.resize_images(image_decoded, [224, 224])\n        batch_data.append(image_resized)\n        batch_label.append(label)\n    return batch_data, batch_label\ndataset = tf.data.Dataset.from_tensor_slices((filenames, labels)) \ndataset = dataset.map(_parse_function) \n<\/code><\/pre>\n\n<p>where filename is a list of folder name like '123456', labels is list of label like 0 or 1.<\/p>\n","answers":[]}
+{"QuestionId":52587227,"AnswerCount":1,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-10-01T08:29:47.663","AcceptedAnswerId":null,"OwnerUserId":9900971.0,"Title":"how to install pytorch version 0.1.12 in anaconda 3.6 windows 10?","Body":"<p>Tried to install with this command but still didn't work-> conda install -c peterjc123 pytorch=0.1.12<\/p>\n\n<p>Also tried installing using this command:\nconda install pytorch=0.1.12 -c pytorch\nHow can it be installed with python anaconda 3.6?<\/p>\n","answers":[{"AnswerId":"52588937","CreationDate":"2018-10-01T10:14:33.370","ParentId":null,"OwnerUserId":"6390175","Title":null,"Body":"<p>First of all, make sure that <code>Python 3.5 or later<\/code> is installed as well.<\/p>\n\n<ul>\n<li><em>Go<\/em> to <a href=\"https:\/\/drive.google.com\/drive\/folders\/0B-X0-FlSGfCYdTNldW02UGl4MXM\" rel=\"nofollow noreferrer\">this link<\/a> and <em>download<\/em> \"pytorch-0.1.12-py36_0.1.12cu80.tar.bz2\".<\/li>\n<li>Now <em>navigate<\/em> to the <strong><em>downloaded file<\/em><\/strong> and <em>install<\/em> it through <code>Anaconda console<\/code> <code>conda install [path to file]\/pytorch-0.1.12-py36_0.1.12cu80.tar.bz2<\/code><\/li>\n<\/ul>\n\n<p><strong><em>Note:<\/em><\/strong>\n<em>If it requires some additional modules, use conda or pip to install those addition modules via:<\/em><\/p>\n\n<pre><code>conda install (module)\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>pip install (module)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52589043,"AnswerCount":0,"Tags":"<tensorflow><parallel-processing><hdfs><tfrecord>","CreationDate":"2018-10-01T10:21:07.167","AcceptedAnswerId":null,"OwnerUserId":4844279.0,"Title":"How to read TFRecord files in TensorFlow from a remote HDFS file system in parallel and efficiently","Body":"<p>I am reading TFRecords data files from a remote HDFS file system through tf.data.TFRecordDataset API in TensorFlow. However, when I increase the num_parallel_reads, say from 1 to 32, not only I don't see improvement in reading speed but I see it takes longer to read the data. Are there any tips to set num_parallel_reads properly to enhance the performance? I know tf.data.TFRecordDataset  uses parallel interleaves which is supposed to be good.<\/p>\n","answers":[]}
+{"QuestionId":52590317,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensorboard><object-detection-api>","CreationDate":"2018-10-01T11:41:09.727","AcceptedAnswerId":null,"OwnerUserId":9709132.0,"Title":"How do we know the model converged while training a neural network for object classification\/detection?","Body":"<p>I am using Tensorflow's Object Detection API to detect cars. It should detect the cars as one class \"car\".<\/p>\n\n<p>I followed sentdex's following series:<\/p>\n\n<p><a href=\"https:\/\/pythonprogramming.net\/introduction-use-tensorflow-object-detection-api-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/pythonprogramming.net\/introduction-use-tensorflow-object-detection-api-tutorial\/<\/a><\/p>\n\n<p>System information:<\/p>\n\n<p>OS - Ubuntu 18.04 LTS<\/p>\n\n<p>GPU - Nvidia 940M (VRAM : 2GB)<\/p>\n\n<p>Tensorflow : 1.10<\/p>\n\n<p>Python - 3.6<\/p>\n\n<p>CPU - Intel i5<\/p>\n\n<p>Problem:\nThe training process runs pretty fine. In order to know when the model converges and when I should stop training, I observe the <strong>loss during the training per step<\/strong> in the terminal where the training is running and <strong>also observe the Total Loss graph in Tensorboard<\/strong> via running the following command in another terminal,<\/p>\n\n<pre><code>$tensorboard --logdit=\"training\"\n<\/code><\/pre>\n\n<p>But even after training till 60k steps, the loss fluctuates between 2.1 to 1.2. If I stop the training and export the inference graph from the last checkpoint(saved in the training\/ folder), it detects cars in some cases and in some it gives false positives.<\/p>\n\n<p>I also tried running <strong>eval.py<\/strong> like below,<\/p>\n\n<pre><code>python3 eval.py     --logtostderr     --pipeline_config_path=training\/ssd_mobilenet_v1_pets.config     --checkpoint_dir=training\/     --eval_dir=eval\/\n<\/code><\/pre>\n\n<p>but it gives out an error that indicates that the GPU memory fails to run this script <strong>along with<\/strong> train.py.<\/p>\n\n<p>So, I stop the training to make sure the GPU is free and then run <strong>eval.py<\/strong> but it creates only one eval point in eval\/ folder. Why?<\/p>\n\n<p>Also, how do I understand from the Precision graphs in Tensorboard that the training needs to be stopped?<\/p>\n\n<p>I could also post screenshots if anyone wants.\nShould I keep training till the loss stays on an average around 1?<\/p>\n\n<p>Thanks.<\/p>\n\n<p>PS: Added Total Loss graph below till 66k steps.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vhVBF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vhVBF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>PS2: After 2 days training(and still on) this is the total loss graph below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/enXvM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/enXvM.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52590942","CreationDate":"2018-10-01T12:16:48.303","ParentId":null,"OwnerUserId":"2832960","Title":null,"Body":"<p>Usually, one uses a separate set of data to measure the error and generalisation abilities of the model. So, one would have the following sets of data to train and evaluate a model:<\/p>\n\n<ul>\n<li><strong>Training set:<\/strong> The data used to train the model. <\/li>\n<li><strong>Validation set:<\/strong> A separate set of data which will be used to measure the error during training. The data of this set is not used to perform any weight updates.<\/li>\n<li><strong>Test set:<\/strong> This set is used to measure the model's performance after the training.<\/li>\n<\/ul>\n\n<p>In your case, you would have to define a separate set of data, the <strong>validation set<\/strong> and run an evaluation repeadingly after a fixed number of batches\/steps and log the error or accuracy. What usually happens is, that the error on that data will decrease in the beginning and increase at a certain point during training. So it's important to keep track of that error and to generate a checkpoint whenever this error is decreases. The checkpoint with the lowest error on your validation data is one that you want to use. This technique is called <strong>Early Stopping<\/strong>.<\/p>\n\n<p>The reason why the error increases after a certain point during training is called <strong>Overfitting<\/strong>. It tells you that the model losses it's ability to generalize to unseen data.<\/p>\n\n<p><strong>Edit:<\/strong>\nHere's an example of a training loop with early stopping procedure:<\/p>\n\n<pre><code> for step in range(1, _MAX_ITER):\n     if step % _TEST_ITER == 0:\n         sample_count = 0\n         while True:\n                try:\n                    test_data = sess.run(test_batch)\n                    test_loss, summary = self._model.loss(sess, test_data[0], self._assign_target(test_data), self._merged_summary)\n                    sess.run(self._increment_loss_opt, feed_dict={self._current_loss_pl: test_loss})\n                    sample_count += 1\n                except tf.errors.OutOfRangeError:\n                    score = sess.run(self._avg_batch_loss, feed_dict={self._batch_count_pl: sample_count})\n                    best_score =sess.run(self._best_loss)\n                    if score &lt; best_score:\n                        '''\n                        Save your model here...\n                        '''\n<\/code><\/pre>\n"}]}
+{"QuestionId":52590441,"AnswerCount":1,"Tags":"<tensorflow><keras><convolution>","CreationDate":"2018-10-01T11:48:20.700","AcceptedAnswerId":null,"OwnerUserId":6263397.0,"Title":"tensorflow keras conv1d : ValueError: len(dilation_rate)=1 but should be -1","Body":"<p>I try to build a custom Keras regularize with tensorflow as backend.\nExecuting the following piece of code gives me an exception :<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras    \ninputs = keras.Input(shape=(10,))\nx  = keras.backend.conv1d(inputs, tf.constant([-1,1]), padding = 'same', dilation_rate=None)\n\n\n x  = keras.backend.conv1d(inputs, tf.constant([-1,1]), padding = 'same', dilation_rate=None)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/arthur\/miniconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/backend.py\", line 3775, in conv1d\n    data_format=tf_data_format)\n  File \"\/home\/arthur\/miniconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 779, in convolution\n    data_format=data_format)\n  File \"\/home\/arthur\/miniconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 842, in __init__\n    num_spatial_dims, strides, dilation_rate)\n  File \"\/home\/arthur\/miniconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 625, in _get_strides_and_dilation_rate\n    (len(dilation_rate), num_spatial_dims))\nValueError: len(dilation_rate)=1 but should be -1\n<\/code><\/pre>\n\n<p>i can't understand what i am doing wrong.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"52962655","CreationDate":"2018-10-24T06:54:49.760","ParentId":null,"OwnerUserId":"5367279","Title":null,"Body":"<p>I think the problem is in <code>tf.constant([-1,1])<\/code>. This is a place for kernel which should have dimension like <code>input_length,in_channel,out_channel<\/code>.<\/p>\n"}]}
+{"QuestionId":52590711,"AnswerCount":1,"Tags":"<android><opencv><tensorflow><arcore><tensorflow-lite>","CreationDate":"2018-10-01T12:04:09.480","AcceptedAnswerId":52658182.0,"OwnerUserId":1402551.0,"Title":"Image classification through Tensorflow gives the exact same prediction","Body":"<p>Bear with me while I try to elucidate.<\/p>\n\n<p>I have an Android Application which uses OpenCV to convert a YUV420 image into a bitmap and transfers it to an Interpreter. The problem is, every time I run it, I get the exact same class prediction with the exact same confidence values irrelevant of what I point at.<\/p>\n\n<pre><code>...\nRecognitions : [macbook pro: 0.95353276, cello gripper: 0.023749515]. \nRecognitions : [macbook pro: 0.95353276, cello gripper: 0.023749515]. \nRecognitions : [macbook pro: 0.95353276, cello gripper: 0.023749515]. \nRecognitions : [macbook pro: 0.95353276, cello gripper: 0.023749515].\n...\n<\/code><\/pre>\n\n<p>Now before you mention my model is not trained enough, I've tested the exact same <code>.tflite<\/code> file in TFLite example provided in the Tensorflow Codelab-2. It works as it should and recognizes all 4 of my classes with 90%+ accuracy. In addition, I used a <code>label_image.py<\/code> script to test the <code>.pb<\/code> file using which my <code>.tflite<\/code> was derived from and it works as it should. I've trained the model on nearly 5000+ images of each class. Since it works on other apps, I'm guessing there's no problem with the model but my implementation. Though I just can't pinpoint it.<\/p>\n\n<p>Following code is used to Create Mat(s) from the image bytes :<\/p>\n\n<pre><code>\/\/Retrieve the camera Image from ARCore\nval cameraImage = frame.acquireCameraImage()\nval cameraPlaneY = cameraImage.planes[0].buffer\nval cameraPlaneUV = cameraImage.planes[1].buffer\n\n\/\/ Create a new Mat with OpenCV. One for each plane - Y and UV\nval y_mat = Mat(cameraImage.height, cameraImage.width, CvType.CV_8UC1, cameraPlaneY)\nval uv_mat = Mat(cameraImage.height \/ 2, cameraImage.width \/ 2, CvType.CV_8UC2, cameraPlaneUV)\nvar mat224 = Mat()\nvar cvFrameRGBA = Mat()\n\n\/\/ Retrieve an RGBA frame from the produced YUV\nImgproc.cvtColorTwoPlane(y_mat, uv_mat, cvFrameRGBA, Imgproc.COLOR_YUV2BGRA_NV21)\n\/\/ I've tried the following in the above line\n\/\/ Imgproc.COLOR_YUV2RGBA_NV12\n\/\/ Imgproc.COLOR_YUV2RGBA_NV21\n\/\/ Imgproc.COLOR_YUV2BGRA_NV12\n\/\/ Imgproc.COLOR_YUV2BGRA_NV21\n<\/code><\/pre>\n\n<p>Following code is used to add Image data into a ByteBuffer :<\/p>\n\n<pre><code>\/\/ imageFrame is a Mat object created from OpenCV by processing a YUV420 image received from ARCore\noverride fun setImageFrame(imageFrame: Mat) {\n    ...\n    \/\/ Convert mat224 into a float array that can be sent to Tensorflow\n    val rgbBytes: ByteBuffer = ByteBuffer.allocate(1 * 4 * 224 * 224 * 3)\n    rgbBytes.order(ByteOrder.nativeOrder())\n\n    val frameBitmap = Bitmap.createBitmap(imageFrame.cols(), imageFrame.rows(), Bitmap.Config.ARGB_8888, true)\n    \/\/ convert Mat to Bitmap\n    Utils.matToBitmap(imageFrame, frameBitmap, true)\n    frameBitmap.getPixels(intValues, 0, frameBitmap.width, 0, 0, frameBitmap.width, frameBitmap.height)\n\n    \/\/ Iterate over all pixels and retrieve information of RGB channels\n    intValues.forEach { packedPixel -&gt;\n        rgbBytes.putFloat((((packedPixel shr 16) and 0xFF) - 128) \/ 128.0f)\n        rgbBytes.putFloat((((packedPixel shr 8) and 0xFF) - 128) \/ 128.0f)\n        rgbBytes.putFloat(((packedPixel and 0xFF) - 128) \/ 128.0f)\n    }\n}\n\n.......\nprivate var labelProb: Array&lt;FloatArray&gt;? = null\n.......\n\/\/ and classify \nlabelProb?.let { interpreter?.run(rgbBytes, it) }\n.......\n<\/code><\/pre>\n\n<p>I checked the bitmap that gets converted from Mat. It shows up quite as best as it possibly can. <\/p>\n\n<p>Any ideas anyone?<\/p>\n\n<h1>Update One<\/h1>\n\n<p>I changed the implementation of <code>setImageFrame<\/code> method slightly to match an implementation <a href=\"https:\/\/github.com\/Rohithkvsp\/MNIST-on-Android-with-TensorflowLite-and-OpenCV\/blob\/master\/app\/src\/main\/java\/com\/example\/rohithkvsp\/handwittendigit\/Classifier.java\" rel=\"nofollow noreferrer\">here<\/a>. Since it works for him, I hoped it would work for me as well. It still doesn't.<\/p>\n\n<pre><code>override fun setImageFrame(imageFrame: Mat) {\n    \/\/ Reset the rgb bytes buffer\n    rgbBytes.rewind()\n\n    \/\/ Iterate over all pixels and retrieve information of RGB channels only\n    for(rows in 0 until imageFrame.rows())\n        for(cols in 0 until imageFrame.cols()) {\n            val imageData = imageFrame.get(rows, cols)\n            \/\/ Type of Mat is 24\n            \/\/ Channels is 4\n            \/\/ Depth is 0\n            rgbBytes.putFloat(imageData[0].toFloat())\n            rgbBytes.putFloat(imageData[1].toFloat())\n            rgbBytes.putFloat(imageData[2].toFloat())\n        }\n}\n<\/code><\/pre>\n\n<h1>Update Two<\/h1>\n\n<p>Suspicious of my float model, I changed it to a pre-built MobileNet Quant model just to eliminate a possibility. The problem persists in this as well.<\/p>\n\n<pre><code>...\nRecognitions : [candle: 18.0, otterhound: 15.0, syringe: 13.0, English foxhound: 11.0]\nRecognitions : [candle: 18.0, otterhound: 15.0, syringe: 13.0, English foxhound: 11.0]\nRecognitions : [candle: 18.0, otterhound: 15.0, syringe: 13.0, English foxhound: 11.0]\nRecognitions : [candle: 18.0, otterhound: 15.0, syringe: 13.0, English foxhound: 11.0]\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"52658182","CreationDate":"2018-10-05T04:00:42.507","ParentId":null,"OwnerUserId":"1402551","Title":null,"Body":"<p>Okay. so After 4 days, I was able to finally solve this. The Issue was how The <code>ByteBuffer<\/code> is initiated. I was doing :<\/p>\n\n<pre><code>private var rgbBytes: ByteBuffer = ByteBuffer.allocate(1 * 4 * 224 * 224 * 3)\n<\/code><\/pre>\n\n<p>instead of what I ought to be doing :<\/p>\n\n<pre><code>private val rgbBytes: ByteBuffer = ByteBuffer.allocateDirect(1 * 4 * 224 * 224 * 3)\n<\/code><\/pre>\n\n<p>I tried to understand what is the difference between <code>ByteBuffer.allocate()<\/code> and <code>ByteBuffer.allocateDirect()<\/code> <a href=\"https:\/\/docs.oracle.com\/javase\/7\/docs\/api\/java\/nio\/ByteBuffer.html\" rel=\"nofollow noreferrer\">here<\/a> but to no avail. <\/p>\n\n<p>I'd be glad if someone can answer two further questions :<\/p>\n\n<ol>\n<li>Why does Tensorflow need a Direct Byte Buffer rather than a Non Direct buffer?<\/li>\n<li>What is the difference between Direct and Non Direct ByteBuffer in a simplified description?<\/li>\n<\/ol>\n"}]}
+{"QuestionId":52590716,"AnswerCount":0,"Tags":"<python><pandas><tensorflow>","CreationDate":"2018-10-01T12:04:26.440","AcceptedAnswerId":null,"OwnerUserId":8877753.0,"Title":"Panda dataframe -> Float64 or npy?","Body":"<p>I'm currently running through the Tensorflow housing price regression <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_regression#the_boston_housing_prices_dataset\" rel=\"nofollow noreferrer\">tutorial<\/a>, with the same method as in the tutorial, but with real life data.<\/p>\n\n<p>I have all the data sliced up thanks to iloc and pandas dataframe. This leaves me in a position where I have 4 different dataframes.<\/p>\n\n<ol>\n<li><p>Train data - (852,2)<\/p><\/li>\n<li><p>Train labels - (852,1)<\/p><\/li>\n<li><p>Test data - (364,2)<\/p><\/li>\n<li><p>Test labels - (364,1)<\/p><\/li>\n<\/ol>\n\n<p>Now these are all listed as DataFrames. See figure 1 below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Kr3O7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Kr3O7.png\" alt=\"Figure 1\"><\/a><\/p>\n\n<p>This does not work!<\/p>\n\n<p>So I tried running the official tutorial through and I realised that the output of the tutorials' train and test data\/labels, are floats. As seen here in figure 2.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2dAGx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2dAGx.png\" alt=\"Figure 2\"><\/a><\/p>\n\n<p>So I realized its probably related to the top of the columns being str's, and the rest ints\/floats. But even upon calling the dtype out, I get an error and can't actually determine its origin.<\/p>\n\n<p>How would suggest I proceed? I'm not sure if I should abandon the Dataframe format here and stuff it in np arrays?<\/p>\n","answers":[]}
+{"QuestionId":52590787,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T12:09:05.050","AcceptedAnswerId":52593694.0,"OwnerUserId":7447265.0,"Title":"Inspect the .tfrecords file","Body":"<p>I created the images.tfrecoreds file using following code<\/p>\n\n<pre><code>from PIL import Image\nimport numpy as np\nimport tensorflow as tf\nimport glob\n\nimages = glob.glob('E:\\Projects\/FYPT\/vehicle\/bus\/*.jpg')\n\ndef _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef _int64_feature(value):\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\ntfrecords_filename = 'E:\\Projects\/FYPT\/vehicle\/images.tfrecords'\n\nwriter = tf.python_io.TFRecordWriter(tfrecords_filename)\n\noriginal_images = []\n\nfor img_path in images:\n    img = np.array(Image.open(img_path))\n\n\nheight = img.shape[0]\nwidth = img.shape[1]\n\n# Put in the original images into array\n# Just for future check for correctness\noriginal_images.append((img))\n\nimg_raw = img.tostring()\n\nexample = tf.train.Example(features=tf.train.Features(feature={\n    'height': _int64_feature(height),\n    'width': _int64_feature(width),\n    'image_raw': _bytes_feature(img_raw)\n    }))\n\nwriter.write(example.SerializeToString())\n\nwriter.close()\n<\/code><\/pre>\n\n<p>Then I tried to Inspect the output of the tf.TFRecordReader() by printing the output of the \"serialized_example\" <\/p>\n\n<pre><code>import tensorflow as tf\nimport skimage.io as io\n\n\nreader = tf.TFRecordReader()\ntfrecords_filename = 'E:\\Projects\/FYPT\/vehicle\/images.tfrecords'\nfilename_queue = tf.train.string_input_producer([tfrecords_filename],num_epochs=10)\n_,serialized_example = reader.read(filename_queue)\n\nsess= tf.Session()\nprint(sess.run(serialized_example))\n<\/code><\/pre>\n\n<p>but it gives me following warning and not give any output of the \"serialized_example\" <a href=\"https:\/\/i.stack.imgur.com\/LIuF5.png\" rel=\"nofollow noreferrer\">this<\/a> is the screen shot of the command line<\/p>\n\n<p>What is the mistake have I done and how should I print the output of the \"serialized_example\"<\/p>\n","answers":[{"AnswerId":"52593694","CreationDate":"2018-10-01T14:49:03.250","ParentId":null,"OwnerUserId":"4606149","Title":null,"Body":"<p>You get that warning, because you are using <code>tf.train.string_input_producer()<\/code> which returns a queue, but input pipelines based on <code>QueueRunner API<\/code> are deprecated and not supported in future versions.<\/p>\n\n<p><strong>Queue-based solution - not recommended!<\/strong><\/p>\n\n<p><code>serialized_example<\/code> is just a string object (the same that was written with the <code>tf.python_io.TFRecordWriter<\/code> to <code>images.tfrecords<\/code> file for each example). <\/p>\n\n<p>You need to parse each single example to get its features. In your case:<\/p>\n\n<pre><code>features = tf.parse_single_example(serialized_example,\n                                   features={\"image_raw\": tf.FixedLenFeature([], tf.string),\n                                             \"height\": tf.FixedLenFeature([], tf.int64) }\n\nimg_raw = tf.image.decode_jpeg(features[\"image_raw\"])\nimg_height = features[\"height\"]\n\n# initialize global and local variables\ninit_op = tf.group(tf.local_variables_initializer(),\n                   tf.global_variables_initializer())\n\nwith tf.Session() as sess:\n  sess.run(init_op)\n\n  # start a number of threads\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n  try:\n    while not coord.should_stop():\n     img_raw_value, img_height_value = sess.run([img_raw, img_height])\n     print(img_raw_value.shape)\n     print(img_height_value)\n  except tf.errors.OutOfRangeError:\n     print(\"End of data\")\n  finally:\n     coord.request_stop()\n\n  # wait for all threads to terminate\n  coord.join(threads)\n  sess.close()\n<\/code><\/pre>\n\n<p><strong>Dataset API - highly recommended!<\/strong><\/p>\n\n<p>A detailed description how to build a input pipeline can be found <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/reading_data#_QueueRunner_\" rel=\"nofollow noreferrer\">here: TensorFlow API<\/a>.<\/p>\n\n<p>In your case you should define a <code>_parse_function<\/code> like this:<\/p>\n\n<pre><code>def _parse_function(example_proto):\n  features={\"imgage_raw\": tf.FixedLenFeature([], tf.string),\n            \"height\": tf.FixedLenFeature([], tf.int64),\n            \"width\": tf.FixedLenFeature([], tf.int64)}\n\n  parsed_features = tf.parse_single_example(example_proto, features)\n\n  img_raw = tf.image.decode_jpeg(parsed_features[\"img_raw\"])\n  height = parsed_features[\"height\"]\n  width = parsed_features[\"width\"]\n\n  return img_raw, height, width\n<\/code><\/pre>\n\n<p>Than create a dataset that reads all of the examples from TFRecord file, and extract the features:<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset([tfrecords_filename])\ndataset = dataset.map(_parse_function)\n# here you could batch and shuffle\n\niterator = dataset.make_one_shot_iterator()\n\nnext_element = iterator.get_next()\n\nwith tf.Session() as sess: \n  while True:\n    try:\n      val = sess.run(next_element)\n      print(\"img_raw:\", val[0].shape)\n      print(\"height:\", val[1])\n      print(\"width:\", val[2])\n    except tf.errors.OutOfRangeError:\n      print(\"End of dataset\")\n      break \n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"}]}
+{"QuestionId":52590880,"AnswerCount":2,"Tags":"<tensorflow><memory><resources><gpu><mask>","CreationDate":"2018-10-01T12:13:50.590","AcceptedAnswerId":52602692.0,"OwnerUserId":10438842.0,"Title":"Mask RCNN Resource exhausted (OOM) on my own dataset","Body":"<p>Help needed for Mask RCNN Resource Exhausted -<\/p>\n\n<p>H\/W - i7-8700, 32G RAM, single ASUS ROG STRIX 1080ti (11GB)<\/p>\n\n<p>Virtual env setup - tensorflow-gpu==1.5.0, python==3.6.6, Cuda==9.0.176, cudnn==7.2.1 <\/p>\n\n<p>image resolution - maximum width=900 pixels, maximum height=675pixels, minimum width=194 pixels, minimum height=150 pixels, 11 images for training<\/p>\n\n<p>S\/W - IMAGES_PER_GPU = 1 (in class xxConfig(Config), xxx.py), BACKBONE = \"resnet50\", POST_NMS_ROIS_TRAINING = 1000, POST_NMS_ROIS_INFERENCE = 500, IMAGE_RESIZE_MODE = \"square\", IMAGE_MIN_DIM = 400, IMAGE_MAX_DIM = 512, TRAIN_ROIS_PER_IMAGE = 100<\/p>\n\n<p>What strange to me was, nvidia-smi showed &lt; 300MB used for python, the terminal showed the following, however, <\/p>\n\n<p>ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[3,3,256,256] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n     [[Node: fpn_p5\/random_uniform\/RandomUniform = RandomUniformT=DT_INT32, dtype=DT_FLOAT, seed=87654321, seed2=5038409, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WPdQK.png\" rel=\"nofollow noreferrer\">nvidia-smi<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PfJU2.png\" rel=\"nofollow noreferrer\">error-log when running the code<\/a><\/p>\n","answers":[{"AnswerId":"52602692","CreationDate":"2018-10-02T05:53:26.727","ParentId":null,"OwnerUserId":"10438842","Title":null,"Body":"<p>After replacing cudnn 7.2.1 with 7.0.5, I am now able to train Mask-RCNN using 1080ti gpu without a resource exhausted (OOM) issue.<\/p>\n"},{"AnswerId":"52600388","CreationDate":"2018-10-01T23:59:08.283","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>Tensorflow by default allocates all GPU memory. So if you only see 700 MB allocated in <code>nvidia-smi<\/code> then you have most likely set some option in Tensorflow to limit GPU memory such as:<\/p>\n\n<pre><code>config.gpu_options.allow_growth = True\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>config.gpu_options.per_process_gpu_memory_fraction = 0.4\n<\/code><\/pre>\n\n<p>Remove this options and try again. See also: <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/using_gpu<\/a><\/p>\n"}]}
+{"QuestionId":52591014,"AnswerCount":1,"Tags":"<javascript><tensorflow><broadcast><tensorflow.js>","CreationDate":"2018-10-01T12:21:08.117","AcceptedAnswerId":null,"OwnerUserId":2603925.0,"Title":"How to broadcast matrix\/vector dot ptoduct in tensorflow.js","Body":"<p>How can I broadcast the multiplication of a matrix A with multiple vectors in Tensorflow with Javascript ?\nLet's define A:<\/p>\n\n<pre><code>A = tf.tensor2d([1, 0, 0, 2], [2, 2])\nA.print()\nTensor\n    [[1, 0],\n     [0, 2]]\n<\/code><\/pre>\n\n<p>Let's define the b vectors as column vector (matrix):<\/p>\n\n<pre><code>b = Array.apply({}, new Array(10)).map(_ =&gt; tf.tensor2d([[Math.floor(Math.random() * 100)], [Math.floor(Math.random() * 100)]]));\nb[0].print()\nTensor\n    [[90 ],\n     [122]]\n<\/code><\/pre>\n\n<p>Trying <code>.matMul<\/code>:<\/p>\n\n<pre><code>A.matMul(b)\ntfjs@latest:2 Uncaught Error: Error in matMul: inputs must have the same rank of at least 2, got ranks 2 and 1.\n<\/code><\/pre>\n\n<p>or <code>.dot<\/code>:<\/p>\n\n<pre><code>A.dot(b)\ntfjs@latest:2 Uncaught Error: Error in dot: inner dimensions of inputs must match, but got 2 and 10\n<\/code><\/pre>\n\n<p>Also tried with b as regular 1d tensor:<\/p>\n\n<pre><code>b = Array.apply({}, new Array(10)).map(_ =&gt; tf.tensor1d([Math.floor(Math.random() * 100), Math.floor(Math.random() * 100)]));\n<\/code><\/pre>\n\n<p>Results are the same. Moreover, looking at the doc it seems that those operations are not broadcastable. I don't see why they can't be as <code>numpy<\/code> is able to broadcast them.<\/p>\n\n<p>I saw some answer to the same question for python suggesting the use of <code>einsum<\/code>. But it seems to be available only for python, not javascript.<\/p>\n","answers":[{"AnswerId":"52596800","CreationDate":"2018-10-01T18:25:05.810","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>This is not a broadcasting issue. <code>b<\/code> is a normal JavaScript array. In this case it is an array of tensor. \nTo use b in <code>dot<\/code> or <code>matMul<\/code> operation, b needs to be a tensor. That is not the case here.\nFor a <code>matMul<\/code> operation, given two tensors <code>x (shape [p,q])<\/code>, <code>y (shape [r, s])<\/code>, q should be equal to r. This is the very definition of a matrix multiplication, ie there is no brodcasting for matMul either in tensorflow.js or in tensorflow.py. The dot operation is very similar. The only difference is that one operand can be a matrix and the second a vector.<\/p>\n\n<p>Consider the following: for the sake of simplicity, what is in brackets is a tensor of that shape.<\/p>\n\n<pre><code>[4, 5].matMul([5, 6]) = [4, 5].dot([5, 4]) =&gt; [4, 6]\n<\/code><\/pre>\n\n<pre><code>[4, 5].matMul([4, 5]) =&gt; error \n<\/code><\/pre>\n\n<p>because of 4 and 5 (ie why to do the square of a matrix is sometimes defines as the product between the matrix and its transpose. For given a matrix A,  A*A could not always be computed unless A is a square matrix)<\/p>\n\n<pre><code>[5].dot([5, 4]) =&gt; [4]\n[5, 4].dot([4]) =&gt; [5, 1]\n<\/code><\/pre>\n\n<pre><code>[5].dot([4, 5]) =&gt; error\n[5, 4].dot([5]) =&gt; error\n<\/code><\/pre>\n\n<p>To compute the product of b and a, you will need to iterate over b and do the matMul operation.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>a = tf.tensor2d([1, 0, 0, 2], [2, 2])\r\n\r\n\/\/ there is much simpler way to create an array of 10 tensors without apply function\r\nconst b = Array.from({length:10}, _ =&gt; tf.tensor2d([[Math.floor(Math.random() * 100)], [Math.floor(Math.random() * 100)]]))\r\n\/\/ Matmul operation\r\nconst prod = b.map(t =&gt; a.matMul(t))\r\nprod.forEach(t =&gt; t.print())<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.12.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":52591291,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-estimator>","CreationDate":"2018-10-01T12:35:53.257","AcceptedAnswerId":null,"OwnerUserId":3079777.0,"Title":"Tensorflow Estimator: loss not decreasing when using tf.feature_column.embedding_column for a list of categorical variables","Body":"<p>I'm very new to Tensorflow Estimator. I wonder if it's possible to pass an array of categorical variables as feature to the estimator and it automatically converts it to an array of embeddings. For example, the following is a record in a CSV file. It contains 2 lists of categorical variables(enclosed in brackets), \"country\" and \"watch\", 2 categorical variables, \"day_of_week\" and \"day_period\" and one target, \"movie_id\" in this case.<\/p>\n\n<pre><code>day_of_week,day_period,country,movie_id,watched\nSUNDAY,EVENING,[USA,UK],B2JO1owWbeLn,[WGdZ5qZmLw0,abcdef]\nMONDAY,EVENING,[China],xxx,[abc,def,ijk]\n<\/code><\/pre>\n\n<p>According to the doc <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column<\/a>, \"day_of_week\" and \"day_period\" can be represented as \"categorical_column_with_vocabulary_list\". This is straightforward. However, \"country\", and \"watched\" are a list of categorical variables. I'd like to merge each categorical variable in a list to an embedding. From the same doc, a \"tf.feature_column.embedding_column\" would do the trick.<\/p>\n\n<p>The following function builds columns representing the above input.<\/p>\n\n<pre><code>def build_model_columns():\n    day_of_week = tf.feature_column.categorical_column_with_vocabulary_list('day_of_week', day_of_weeks)\n    day_period = tf.feature_column.categorical_column_with_vocabulary_list('day_period', day_periods)\n    country = tf.feature_column.categorical_column_with_vocabulary_list('country', countries)\n    watched = tf.feature_column.categorical_column_with_vocabulary_list('watched', movie_emb_ids)\n\n    columns = [\n        tf.feature_column.indicator_column(day_of_week),\n        tf.feature_column.indicator_column(day_period),\n        tf.feature_column.embedding_column(country, 8),\n        tf.feature_column.embedding_column(watched, 32)\n    ]\n    return columns\n<\/code><\/pre>\n\n<p>The following is a function generating training dataset<\/p>\n\n<pre><code>def tensor_to_array(tensor):\n    length = tf.size(tf.string_split([tensor], \"\"))\n    sub = tf.substr(tensor, 1, length-2) # remove the leading '[' and trailing ']'\n    splits = tf.string_split([sub], delimiter=',')\n    return splits\n\ndef train_input_fn():\n    train_files = \"train.csv\"\n    target_files = \"target.csv\"\n    target_table, target_ids = read_table_lookup(target_files, \"movie\")\n\n    def preprocess(day_of_week, day_period, country, movie_id, watched):\n\n        features = {\n            'day_of_week': day_of_week,\n            'day_period': day_period,\n            'country': tensor_to_array(country),\n            'watched': tensor_to_array(watched)\n\n        }\n        # target_table is a lookup table converting \"movie_id\" to integer \"id\"\n        return features, target_table.lookup(movie_id) \n\n    dataset = (tf.contrib.data.CsvDataset(train_files, record_defaults, header=True)\n           .map(preprocess, num_parallel_calls=5)\n           .batch(batch_size=batch_size, drop_remainder=False)\n           .repeat()\n          )\n\n    # iterator = dataset.make_initializable_iterator()\n    # tf.add_to_collection(tf.GraphKeys.TABLE_INITIALIZERS, iterator.initializer)\n\n    return dataset \n<\/code><\/pre>\n\n<p>The following is a code snippet to create and train an estimator.<\/p>\n\n<pre><code>hidden_units = [512, 512]\nrecord_defaults = [[\"\"]] * 5\ncolumns = build_model_columns()\nestimator = tf.estimator.DNNClassifier(model_dir=\"dir\",\n                                   feature_columns=columns,\n                                   hidden_units=hidden_units,\n                                   n_classes=len(target_ids)) # length of all targets\n\nestimator.train(input_fn=train_input_fn)\n<\/code><\/pre>\n\n<p>I got no error and it seems like everything should work as expected but the training loss is so huge and fluctuating around 3,xxx and never decreasing. See below<\/p>\n\n<pre><code>INFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Create CheckpointSaverHook.\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Saving checkpoints for 0 into \/home\/cocoza4\/movie_models\/deep\/model.ckpt.\nINFO:tensorflow:loss = 6538.0645, step = 0\nINFO:tensorflow:global_step\/sec: 17.353\nINFO:tensorflow:loss = 3596.562, step = 100 (5.764 sec)\nINFO:tensorflow:global_step\/sec: 17.434\nINFO:tensorflow:loss = 3504.936, step = 200 (5.736 sec)\nINFO:tensorflow:global_step\/sec: 17.4234\nINFO:tensorflow:loss = 3500.0488, step = 300 (5.739 sec)\nINFO:tensorflow:global_step\/sec: 17.5321\nINFO:tensorflow:loss = 3480.702, step = 400 (5.705 sec)\nINFO:tensorflow:global_step\/sec: 17.4534\nINFO:tensorflow:loss = 3517.599, step = 500 (5.729 sec)\nINFO:tensorflow:global_step\/sec: 17.3421\nINFO:tensorflow:loss = 3446.142, step = 600 (5.769 sec)\nINFO:tensorflow:global_step\/sec: 17.313\nINFO:tensorflow:loss = 3281.3088, step = 700 (5.776 sec)\nINFO:tensorflow:global_step\/sec: 17.4421\nINFO:tensorflow:loss = 3326.7336, step = 800 (5.731 sec)\nINFO:tensorflow:global_step\/sec: 17.3619\nINFO:tensorflow:loss = 3464.902, step = 900 (5.762 sec)\nINFO:tensorflow:global_step\/sec: 17.2013\nINFO:tensorflow:loss = 3364.2153, step = 1000 (5.813 sec)\nINFO:tensorflow:global_step\/sec: 17.4429\nINFO:tensorflow:loss = 3410.449, step = 1100 (5.734 sec)\nINFO:tensorflow:global_step\/sec: 17.0483\nINFO:tensorflow:loss = 3351.018, step = 1200 (5.866 sec)\nINFO:tensorflow:global_step\/sec: 17.4214\nINFO:tensorflow:loss = 3386.995, step = 1300 (5.740 sec)\nINFO:tensorflow:global_step\/sec: 17.7965\nINFO:tensorflow:loss = 3263.6074, step = 1400 (5.617 sec)\nINFO:tensorflow:global_step\/sec: 17.6944\nINFO:tensorflow:loss = 3321.574, step = 1500 (5.652 sec)\nINFO:tensorflow:global_step\/sec: 17.3603\nINFO:tensorflow:loss = 3234.7761, step = 1600 (5.760 sec)\n<\/code><\/pre>\n\n<p>I wonder if I've done something wrong when preparing the training data?<\/p>\n\n<p>thanks<\/p>\n\n<p>Peeranat F.<\/p>\n","answers":[{"AnswerId":"53356321","CreationDate":"2018-11-17T22:53:58.270","ParentId":null,"OwnerUserId":"10240511","Title":null,"Body":"<p>The first thing that caught my attention is the number of hidden units. I would try to adjust your hidden units. Usually the layers should decrease in size, so I would try <code>[512,256,128]<\/code><\/p>\n"}]}
+{"QuestionId":52591657,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-01T12:55:26.483","AcceptedAnswerId":52592308.0,"OwnerUserId":8334261.0,"Title":"How to replace feed_dict when using an input pipeline?","Body":"<p>Suppose you have an network that has worked with <code>feed_dict<\/code> so far to inject data into a graph. Every few epochs, I evaluated the training and test loss by feeding a batch from either dataset to my graph.<\/p>\n\n<p>Now, for performance reasons, I decided to use an input pipeline. Take a look at this dummy example:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndataset_size = 200\nbatch_size= 5\ndimension = 4\n\n# create some training dataset\ndataset = tf.data.Dataset.\\\n    from_tensor_slices(np.random.normal(2.0,size=(dataset_size,dimension)).\n    astype(np.float32))\n\ndataset = dataset.batch(batch_size) # take batches\n\niterator = dataset.make_initializable_iterator()\nx = tf.cast(iterator.get_next(),tf.float32)\nw = tf.Variable(np.random.normal(size=(1,dimension)).astype(np.float32))\n\nloss_func = lambda x,w: tf.reduce_mean(tf.square(x-w)) # notice that the loss function is a mean!\nloss = loss_func(x,w) # this is the loss that will be minimized\ntrain_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n    # train one epoch\n    sess.run(iterator.initializer)\n    for i in range(dataset_size\/\/batch_size):\n        # the training step will update the weights based on ONE batch of examples each step\n        loss1,_ = sess.run([loss,train_op])\n        print('train step {:d}.  batch loss {:f}.'.format(i,loss1))\n\n        # I want to print the loss from another dataset (test set) here\n<\/code><\/pre>\n\n<p><strong>Printing the loss of the training data is no problem, but how do I do this for another dataset?<\/strong> When using <code>feed_dict<\/code>, I simply got a batch from said set and fed it a value for x.<\/p>\n","answers":[{"AnswerId":"52592308","CreationDate":"2018-10-01T13:31:12.277","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>There are several things you can do for that. One simple option could be something like having two datasets and iterators and use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\"><code>tf.cond<\/code><\/a> to switch between them. However, the more powerful way of doing it is to use an iterator that supports this directly. See the guide on <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#creating_an_iterator\" rel=\"nofollow noreferrer\">how to create iterators<\/a> for a description of the various iterator types. For example, using a reinitializable iterator you could have something like this:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndataset_size = 200\ndataset_test_size = 20\nbatch_size= 5\ndimension = 4\n\n# create some training dataset\ndataset = tf.data.Dataset.\\\n    from_tensor_slices(np.random.normal(2.0,size=(dataset_size,dimension)).\n    astype(np.float32))\n\ndataset = dataset.batch(batch_size) # take batches\n\n# create some test dataset\ndataset_test = tf.data.Dataset.\\\n    from_tensor_slices(np.random.normal(2.0,size=(dataset_test_size,dimension)).\n    astype(np.float32))\n\ndataset_test = dataset_test.batch(batch_size) # take batches\n\niterator = tf.data.Iterator.from_structure(dataset.output_types,\n                                           dataset.output_shapes)\n\ndataset_init_op = iterator.make_initializer(dataset)\ndataset_test_init_op = iterator.make_initializer(dataset_test)\n\nx = tf.cast(iterator.get_next(),tf.float32)\nw = tf.Variable(np.random.normal(size=(1,dimension)).astype(np.float32))\n\nloss_func = lambda x,w: tf.reduce_mean(tf.square(x-w)) # notice that the loss function is a mean!\nloss = loss_func(x,w) # this is the loss that will be minimized\ntrain_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n    # train one epoch\n    sess.run(dataset_init_op)\n    for i in range(dataset_size\/\/batch_size):\n        # the training step will update the weights based on ONE batch of examples each step\n        loss1,_ = sess.run([loss,train_op])\n        print('train step {:d}.  batch loss {:f}.'.format(i,loss1))\n\n    # print test loss\n    sess.run(dataset_test_init_op)\n    for i in range(dataset_test_size\/\/batch_size):\n        loss1 = sess.run(loss)\n        print('test step {:d}.  batch loss {:f}.'.format(i,loss1))\n<\/code><\/pre>\n\n<p>You can do something similar with a feedable iterator, depending on what you find more convenient, and I suppose even with an initializable iterator, for example making a boolean dataset that then you map to some data with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\"><code>tf.cond<\/code><\/a>, although that would not be a very natural way to do it.<\/p>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>Here is how you can do it with an initializable iterator, actually in a cleaner way than what I was initially thinking, so maybe you actually like this more:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndataset_size = 200\ndataset_test_size = 20\nbatch_size= 5\ndimension = 4\n\n# create data\ndata = tf.constant(np.random.normal(2.0,size=(dataset_size,dimension)), tf.float32)\ndata_test = tf.constant(np.random.normal(2.0,size=(dataset_test_size,dimension)), tf.float32)\n# choose data\ntesting = tf.placeholder_with_default(False, ())\ncurrent_data = tf.cond(testing, lambda: data_test, lambda: data)\n# create dataset\ndataset = tf.data.Dataset.from_tensor_slices(current_data)\ndataset = dataset.batch(batch_size)\n# create iterator\niterator = dataset.make_initializable_iterator()\n\nx = tf.cast(iterator.get_next(),tf.float32)\nw = tf.Variable(np.random.normal(size=(1,dimension)).astype(np.float32))\n\nloss_func = lambda x,w: tf.reduce_mean(tf.square(x-w)) # notice that the loss function is a mean!\nloss = loss_func(x,w) # this is the loss that will be minimized\ntrain_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n    # train one epoch\n    sess.run(iterator.initializer)\n    for i in range(dataset_size\/\/batch_size):\n        # the training step will update the weights based on ONE batch of examples each step\n        loss1,_ = sess.run([loss,train_op])\n        print('train step {:d}.  batch loss {:f}.'.format(i,loss1))\n\n    # print test loss\n    sess.run(iterator.initializer, feed_dict={testing: True})\n    for i in range(dataset_test_size\/\/batch_size):\n        loss1 = sess.run(loss)\n        print('test step {:d}.  batch loss {:f}.'.format(i,loss1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52591862,"AnswerCount":0,"Tags":"<python><python-3.x><keras><save>","CreationDate":"2018-10-01T13:06:43.490","AcceptedAnswerId":null,"OwnerUserId":10171000.0,"Title":"How can I save an object containing keras models?","Body":"<p>Here is my code skeleton:<\/p>\n\n<pre><code>def build_model(x, y):\n    model = tf.keras.models.Sequential()\n    model.add(tf.keras.layers.Dense(1, activation='relu'))\n    model.compile(loss='mean_squared_error', optimizer='adam')\n    model.fit(x, y)\n    return model\n\nclass MultiModel(object):\n    def __init__(self, x1, y1, x2, y2):\n        self.model1 = build_model(x1, y1)\n        self.model2 = build_model(x2, y2)\n\nif __name__ == '__main__':\n    # code that builds x1, x2, y1, y2\n    mm = MultiModel(x1, y1, x2, y2)  # How to save mm ?\n<\/code><\/pre>\n\n<p>The issue is I don't know how to save the <em>mm<\/em> object that contains several Keras models.<\/p>\n\n<p>The Keras built-in save method enables only to save Keras model, so it is unusable in that case.\nThe pickle module can not save <em>_thread.RLock objects<\/em>, so it is also unusable.<\/p>\n\n<p>It is maybe possible to save each model independently with the Keras save method, then to regroup them and to save them as a whole. But I do not know how to proceed. <\/p>\n\n<p>Any ideas ?<\/p>\n","answers":[]}
+{"QuestionId":52593481,"AnswerCount":1,"Tags":"<python><tensorflow><homography>","CreationDate":"2018-10-01T14:37:58.437","AcceptedAnswerId":52779070.0,"OwnerUserId":6211109.0,"Title":"Implementation of homography in Tensorflow","Body":"<p>I am trying to implement a network in TensorFlow that learns to predict homography (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Homography_(computer_vision)\" rel=\"nofollow noreferrer\">Wiki<\/a>). My network will output a 4-dimensional vector, which will be used to translate 2 images horizontally and vertically. Then, I calculate the error with a ground truth using a central crop (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/central_crop\" rel=\"nofollow noreferrer\"><code>tf.image.central_crop<\/code><\/a>)of these warped images. I tried implementing it, and for the translation part, I used <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/translate\" rel=\"nofollow noreferrer\"><code>tf.contrib.image.translate<\/code><\/a>. But, the gradients are not flowing to the variables of the network. How can I fix this problem? This is the error that I am getting:<\/p>\n\n<pre><code>ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables [\"&lt;tf.Variable 'get_tr\/w1:0' shape=(3, 3, 6, 64) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/b1:0' shape=(20, 298, 298, 64) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/w2:0' shape=(3, 3, 64, 64) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/b2:0' shape=(20, 296, 296, 64) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/w3:0' shape=(3, 3, 64, 128) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/b3:0' shape=(20, 147, 147, 128) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/w4:0' shape=(3, 3, 128, 128) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/b4:0' shape=(20, 73, 73, 128) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/w5:0' shape=(5, 5, 128, 128) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc1\/kernel:0' shape=(609408, 512) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc1\/bias:0' shape=(512,) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc2\/kernel:0' shape=(512, 1024) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc2\/bias:0' shape=(1024,) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc_o\/kernel:0' shape=(1024, 4) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'get_tr\/fc_o\/bias:0' shape=(4,) dtype=float32_ref&gt;\"] and loss Tensor(\"mean_squared_error\/value:0\", shape=(), dtype=float32).\n<\/code><\/pre>\n\n<p>Here is the code for getting this transformation vector.<\/p>\n\n<pre><code>def get_transform_vectors(self):\n    # Start of transformation prediction network\n\n    image = tf.concat((self.img_train_1, self.img_train_2), 3)\n\n    with tf.variable_scope('get_tr'):\n        w1 = tf.Variable(tf.truncated_normal(shape=[3, 3, 6, 64], stddev=0.1),\n                    name='w1')\n        conv1 = tf.nn.conv2d(image,\n                                    filter=w1,\n                                    strides=(1,1,1,1),\n                                    padding=\"VALID\",\n                                    name='conv1')\n\n        b1 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv1), stddev=0.1),\n                    name='b1')\n        conv1_ = tf.nn.relu(conv1+b1, name='conv1_')\n\n        w2 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1),\n                    name='w2')\n        conv2 = tf.nn.conv2d(conv1_,\n                                    filter=w2,\n                                    strides=(1,1,1,1),\n                                    padding=\"VALID\",\n                                    name='conv2')\n        b2 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv2), stddev=0.1),\n                    name='b2')\n        conv2_ = tf.nn.relu(conv2+b2, name='conv2_')\n\n        w3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 128], stddev=0.1),\n                    name='w3')\n        conv3 = tf.nn.conv2d(conv2_,\n                                    filter=w3,\n                                    strides=(1,2,2,1),\n                                    padding=\"VALID\",\n                                    name='conv3')\n        b3 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv3), stddev=0.1),\n                    name='b3')\n        conv3_ = tf.nn.relu(conv3+b3, name='conv3_')\n\n        w4 = tf.Variable(tf.truncated_normal(shape=[3, 3, 128, 128], stddev=0.1),\n                    name='w4')\n        conv4 = tf.nn.conv2d(conv3_,\n                                    filter=w4,\n                                    strides=(1,2,2,1),\n                                    padding=\"VALID\",\n                                    name='conv4')\n\n        b4 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv4), stddev=0.1),\n                    name='b4')\n        conv4_ = tf.nn.relu(conv4+b4, name='conv4_')\n\n        w5 = tf.Variable(tf.truncated_normal(shape=[5, 5, 128, 128], stddev=0.1),\n                    name='w5')\n        conv5 = tf.nn.conv2d(conv4_,\n                                    filter=w5,\n                                    strides=(1,1,1,1),\n                                    padding=\"VALID\",\n                                    name='conv5')\n\n        conv5_ = tf.contrib.layers.flatten(conv5)\n\n        fc1 = tf.layers.dense(conv5_, 512, activation=tf.nn.relu, name='fc1')\n        fc2 = tf.layers.dense(fc1, 1024, activation=tf.nn.relu, name='fc2')\n        fc_o = tf.layers.dense(fc2, 4, name='fc_o')\n\n    return fc_o\n<\/code><\/pre>\n\n<p>This is the code for translating the image and calculating the loss.<\/p>\n\n<pre><code>self.img_o = tf.contrib.image.translate(self.img_train_1,\n                                            tf.cast(tf.reshape(self.tr_vector[:, 0:2], [self.batch_size,2]),\n                                                    dtype=tf.float32)) +\\\n                 tf.contrib.image.translate(self.img_train_2,\n                                            tf.cast(tf.reshape(self.tr_vector[:, 2:4], [self.batch_size,2]),\n                                                    dtype=tf.float32))\n\nself.loss = tf.losses.mean_squared_error(self.img_o, self.img_label)\nself.optim = tf.train.AdamOptimizer().minimize(self.loss)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52779070","CreationDate":"2018-10-12T11:58:45.160","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Since you are only interested in optimizing translations, I came up with this function to perform a translation and have gradients for the vectors, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/custom_gradient\" rel=\"nofollow noreferrer\"><code>@tf.custom_gradient<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.custom_gradient\ndef my_img_translate(imgs, translates):\n    # Interpolation model has to be fixed due to limitations of tf.custom_gradient\n    interpolation = 'NEAREST'\n    imgs_translated = tf.contrib.image.translate(imgs, translates, interpolation=interpolation)\n    def grad(img_translated_grads):\n        translates_x = translates[:, 0]\n        translates_y = translates[:, 1]\n        translates_zero = tf.zeros_like(translates_x)\n        # X gradients\n        imgs_x_grad = (imgs[:, :, :-2] - imgs[:, :, 2:]) \/ 2\n        imgs_x_grad = tf.concat([(imgs[:, :, :1] - imgs[:, :, 1:2]),\n                                 imgs_x_grad,\n                                 (imgs[:, :, -2:-1] - imgs[:, :, -1:])], axis=2)\n        imgs_x_grad_translated = tf.contrib.image.translate(\n            imgs_x_grad, tf.stack([translates_x, translates_zero], axis=1),\n            interpolation=interpolation)\n        translates_x_grad = tf.reduce_sum(img_translated_grads * imgs_x_grad_translated, axis=(1, 2, 3))\n        # Y gradients\n        imgs_y_grad = (imgs[:, :-2] - imgs[:, 2:]) \/ 2\n        imgs_y_grad = tf.concat([(imgs[:, :1] - imgs[:, 1:2]),\n                                 imgs_y_grad,\n                                 (imgs[:, -2:-1] - imgs[:, -1:])], axis=1)\n        imgs_y_grad_translated = tf.contrib.image.translate(\n            imgs_y_grad, tf.stack([translates_zero, translates_y], axis=1),\n            interpolation=interpolation)\n        translates_y_grad = tf.reduce_sum(img_translated_grads * imgs_y_grad_translated, axis=(1, 2, 3))\n        # Complete gradient\n        translates_grad = tf.stack([translates_x_grad, translates_y_grad], axis=1)\n        return None, translates_grad\n    return imgs_translated, grad\n<\/code><\/pre>\n\n<p>Note that in this case I am not returning any gradient for the images, since these will not be optimized (but if you wanted that, you could in principle use the builtin gradient for the translation operation).<\/p>\n\n<p>I tested this for a simple use case of translating an image so its center had the highest value:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n@tf.custom_gradient\ndef my_img_translate(imgs, translates):\n    # Interpolation model has to be fixed due to limitations of tf.custom_gradient\n    interpolation = 'NEAREST'\n    imgs_translated = tf.contrib.image.translate(imgs, translates, interpolation=interpolation)\n    def grad(img_translated_grads):\n        translates_x = translates[:, 0]\n        translates_y = translates[:, 1]\n        translates_zero = tf.zeros_like(translates_x)\n        # X gradients\n        imgs_x_grad = (imgs[:, :, :-2] - imgs[:, :, 2:]) \/ 2\n        imgs_x_grad = tf.concat([(imgs[:, :, :1] - imgs[:, :, 1:2]),\n                                 imgs_x_grad,\n                                 (imgs[:, :, -2:-1] - imgs[:, :, -1:])], axis=2)\n        imgs_x_grad_translated = tf.contrib.image.translate(\n            imgs_x_grad, tf.stack([translates_x, translates_zero], axis=1),\n            interpolation=interpolation)\n        translates_x_grad = tf.reduce_sum(img_translated_grads * imgs_x_grad_translated, axis=(1, 2, 3))\n        # Y gradients\n        imgs_y_grad = (imgs[:, :-2] - imgs[:, 2:]) \/ 2\n        imgs_y_grad = tf.concat([(imgs[:, :1] - imgs[:, 1:2]),\n                                 imgs_y_grad,\n                                 (imgs[:, -2:-1] - imgs[:, -1:])], axis=1)\n        imgs_y_grad_translated = tf.contrib.image.translate(\n            imgs_y_grad, tf.stack([translates_zero, translates_y], axis=1),\n            interpolation=interpolation)\n        translates_y_grad = tf.reduce_sum(img_translated_grads * imgs_y_grad_translated, axis=(1, 2, 3))\n        # Complete gradient\n        translates_grad = tf.stack([translates_x_grad, translates_y_grad], axis=1)\n        return None, translates_grad\n    return imgs_translated, grad\n\n# Test operations\nimgs = tf.placeholder(tf.float32, [None, None, None, None])\ntranslates = tf.Variable([0, 0], dtype=tf.float32)\ntranslates_tiled = tf.tile(translates[tf.newaxis], (tf.shape(imgs)[0], 1))\nimgs_translated = my_img_translate(imgs, translates_tiled)\nimgs_midpoint = imgs_translated[:, tf.shape(imgs_translated)[1] \/\/ 2, tf.shape(imgs_translated)[2] \/\/ 2]\nloss = -tf.reduce_sum(tf.square(imgs_midpoint))\ntrain_op = tf.train.GradientDescentOptimizer(10).minimize(loss)\ninit_op = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    # Make test image\n    xs, ys = np.meshgrid(np.linspace(-2, 1, 100), np.linspace(-1, 2, 75))\n    test_img = (1 \/ (1 + np.square(xs))) * (1 \/ (1 + np.square(ys)))\n    test_img \/= np.max(test_img)\n    test_img_batch = test_img[np.newaxis, :, :, np.newaxis]\n    # Train\n    sess.run(init_op)\n    for _ in range(100):\n        sess.run(train_op, feed_dict={imgs: test_img_batch})\n    # Show result\n    test_img_result = sess.run(imgs_translated, feed_dict={imgs: test_img_batch})[0, :, :, 0]\n    plt.figure()\n    plt.subplot(121)\n    plt.imshow(test_img)\n    plt.subplot(122)\n    plt.imshow(test_img_result)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TDovA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TDovA.png\" alt=\"Test result\"><\/a><\/p>\n"}]}
+{"QuestionId":52593828,"AnswerCount":1,"Tags":"<caffe2>","CreationDate":"2018-10-01T14:57:22.570","AcceptedAnswerId":null,"OwnerUserId":2225895.0,"Title":"Caffe2: make ops fails when installing densepose","Body":"<p>I am trying to install <a href=\"https:\/\/github.com\/facebookresearch\/DensePose\" rel=\"nofollow noreferrer\">DensePose<\/a> but run into problems when executing <code>make ops<\/code>. <\/p>\n\n<p>First I get error <code>Caffe2Config.cmake<\/code> not found, but after following the instructions here <a href=\"https:\/\/caffe2.ai\/docs\/getting-started.html?platform=ubuntu&amp;configuration=compile\" rel=\"nofollow noreferrer\">Caffe2 build from source<\/a> the <code>Caffe2Config.cmake<\/code> ends up under the <code>build<\/code> directory.<\/p>\n\n<p>Adding the path to the <code>build<\/code> folder gives then the following error: <\/p>\n\n<pre><code>  CMake Error at \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/Caffe2Config.cmake:14 (include):\n  include could not find load file:\n\n    \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/public\/utils.cmake\nCall Stack (most recent call first):\n  CMakeLists.txt:8 (find_package)\n\n\nCMake Error at \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/Caffe2Config.cmake:17 (include):\n  include could not find load file:\n\n    \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/public\/threads.cmake\nCall Stack (most recent call first):\n  CMakeLists.txt:8 (find_package)\n\n\nCMake Error at \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/Caffe2Config.cmake:88 (include):\n  include could not find load file:\n\n    \/home\/erikbylow\/Code\/LocalLibs\/pytorch\/build\/public\/cuda.cmake\nCall Stack (most recent call first):\n  CMakeLists.txt:8 (find_package)\n<\/code><\/pre>\n\n<p>The paths are clearly wrong since the requested <code>cmake-<\/code>files are under <code>pytorch\/cmake\/public<\/code> instead of <code>pytorch\/build\/public<\/code>. <\/p>\n\n<p>How can one resolve that? I have just followed the instructions and the tests one makes along the way works. It seems to be the installation of <code>Caffe2<\/code> that causes the problem?<\/p>\n","answers":[{"AnswerId":"53314201","CreationDate":"2018-11-15T07:16:06.073","ParentId":null,"OwnerUserId":"9441075","Title":null,"Body":"<p>In my case, setting the environment variable <code>Caffe2_DIR<\/code> to the directory of the missing files solved the problem.<\/p>\n\n<p>As you mentioned, these files are in the <code>pytorch\/cmake\/public<\/code> directory instead of <code>pytorch\/build\/public<\/code>. To define the path of Caffe2 manually, open the <code>CMakeLists.txt<\/code> in the densepose folder, add one line in the beginning:<\/p>\n\n<pre><code>set(Caffe2_DIR \"\/home\/erikbylow\/Code\/LocalLibs\/pytorch\/cmake\/\")\n<\/code><\/pre>\n\n<p>Run <code>make ops<\/code> again and the issue should be solved. If not, also try<\/p>\n\n<pre><code>set(Caffe2_DIR \"\/home\/erikbylow\/Code\/LocalLibs\/pytorch\/torch\/share\/cmake\/Caffe2\/\")\n<\/code><\/pre>\n\n<p>Also, after DensePose is successfully installed, remember to append its path to PYTHONPATH:<\/p>\n\n<pre><code>export PYTHONPATH=$PYTHONPATH:\/path\/to\/densepose\/\n<\/code><\/pre>\n"}]}
+{"QuestionId":52594382,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-01T15:31:46.653","AcceptedAnswerId":null,"OwnerUserId":9504664.0,"Title":"Installing Tensorflow for Python 3.6 (Windows 10)","Body":"<p>It could be a replicate question, but none of the existing questiones have been useful for me. I've tried all the solutions given here (<a href=\"https:\/\/stackoverflow.com\/questions\/38896424\/tensorflow-not-found-using-pip\">stackoverflow<\/a>)\nI get this error:<\/p>\n\n<blockquote>\n  <p>Could not find a version that satisfies the requirement tensorflow (from versions: )\n  No matching distribution found for tensorflow<\/p>\n<\/blockquote>\n\n<p>I have this version of the pip on my Pycharm machine with Python 3.6 (Windows 10):<\/p>\n\n<blockquote>\n  <p>Package    Version ------- \n  pip        18.0 --\n  setuptools 40.4.3 -- \n  wheel      0.32.0<\/p>\n<\/blockquote>\n\n<p>I would like to install <em>tensorflow<\/em> and I have tried to take a look at this <a href=\"https:\/\/storage.googleapis.com\/tensorflow\" rel=\"nofollow noreferrer\">link<\/a> as most of the people suggested.\nDoes anybody want to guide me in my installation?<\/p>\n","answers":[{"AnswerId":"52594638","CreationDate":"2018-10-01T15:46:08.337","ParentId":null,"OwnerUserId":"10394971","Title":null,"Body":"<h3>Installation Commands<\/h3>\n\n<p>The following commands are based on the following installation guide <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<h3>using cmd<\/h3>\n\n<pre><code>C:&gt; pip3 install --upgrade tensorflow \/\/ cpu\nC:&gt; pip3 install --upgrade tensorflow-gpu \/\/ gpu\n<\/code><\/pre>\n\n<h3>using Anaconda<\/h3>\n\n<pre><code>C:&gt; conda create -n tensorflow python=3.5 \nC:&gt; activate tensorflow\n(tensorflow)C:&gt; pip install --ignore-installed --upgrade tensorflow\n(tensorflow)C:&gt; pip install --ignore-installed --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p>You can get step by step instructions <a href=\"https:\/\/www.tensorflow.org\/install\/source_windows\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"57761310","CreationDate":"2019-09-02T18:22:58.213","ParentId":null,"OwnerUserId":"12011204","Title":null,"Body":"<p>below command worked for me.\npip3 install --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.0-py3-none-any.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.0-py3-none-any.whl<\/a><\/p>\n"}]}
+{"QuestionId":52594468,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-01T15:36:49.750","AcceptedAnswerId":52595261.0,"OwnerUserId":8909870.0,"Title":"Tensorflow graph with multiple inputs wihtout tf.placeholder for validation","Body":"<p>I use the <code>tf.data<\/code> API for my models. For now I define the output of the <code>tf.data<\/code> iterator as input for my network. After I got rid of the <code>feed_dict<\/code> method, my performance increased significantly. <\/p>\n\n<p>Now I want to implement a validation set that runs at least after each training once. Is there a way to implement a validation run for <code>tf.data<\/code> or do I have to set a placeholder and manually switch the tf.datasets and use <code>feed_dicts<\/code> again? What is the recommended way for validation tests?<\/p>\n","answers":[{"AnswerId":"52595261","CreationDate":"2018-10-01T16:29:28.737","ParentId":null,"OwnerUserId":"748102","Title":null,"Body":"<h1>The hack-ish way - node replacement<\/h1>\n\n<p>The most trivial way, although definitely not the most beautiful, would just be to use the node created by the <code>tf.data<\/code> API as an input to the feed_dict - this is because in Tensorflow you can replace the value of any node in the computation graph by feeding it's value directly to the feed_dict.<\/p>\n\n<p>So this would be something like<\/p>\n\n<pre><code>batch_input = tf_train_data_foo()\nvalidation_input = tf_validation_data_foo()\n\nmodel = build_model(batch_input)\noptimization_step = some_optimization_foo(model)\n\n# Regular train\nsession.run(optimization_step)\n\n# Validation run\nvalidation_data = session.run(validation_input)\nsession.run(model, {batch_input: validation_data})\n<\/code><\/pre>\n\n<h1>The better way - variable reuse<\/h1>\n\n<p>If all the construction uses <code>tf.get_variable<\/code> instead of creating new variables, and the scopes are all set to enable obtaining an existing variable, you can just call your <code>build_model<\/code> function twice - once with the train data (from <code>tf.data<\/code>) and once with your validation data. You can see more details on variable re-use on <a href=\"https:\/\/stackoverflow.com\/a\/36184935\/748102\">this answer<\/a><\/p>\n"}]}
+{"QuestionId":52594650,"AnswerCount":1,"Tags":"<python><pandas><numpy><tensorflow><keras>","CreationDate":"2018-10-01T15:46:58.020","AcceptedAnswerId":52599385.0,"OwnerUserId":10430178.0,"Title":"Generating Sequence\/Batches with specific lengths for RNN for grouped data","Body":"<p>The problem arrises when I wish to pass data from different groups into RNN - most examples assume elegant timeseries, but when adding groups, we can't simply window over the dataframe, we need to jump when the group changes, so that the data comes from within the group. <\/p>\n\n<p>These groups are just different people, so I wanted to keep their sequences to themselves. E.g. a user browsing a website and us collecting pageview data. Or it could be different stocks and their associated price movements.<\/p>\n\n<pre><code>import pandas as pd\ndata = {\n    'group_id': [1,1,1,1,2,2],\n    'timestep': [1,2,3,4,1,2],\n    'x': [6,5,4,3,2,1],\n    'y': [0,1,1,1,0,1]\n}\ndf = pd.DataFrame(data=data)\n\n\n   group_id  timestep  x  y\n0         1         1  6  0\n1         1         2  5  1\n2         1         3  4  1\n3         1         4  3  1\n4         2         1  2  0\n5         2         2  1  1\n<\/code><\/pre>\n\n<p>Let's assume we would like to use the batch size of 2 samples and each of the samples will have 3 timesteps. <code>RNNSequence.__len__ = 3<\/code> (below) batches, but this is not possible, because we can at most get 2 samples from the 1st group (that makes 1 batch). The 2nd group has only 2 time steps, so the iteration is not possible. <\/p>\n\n<pre><code>from keras.utils import Sequence\n\nclass RNNSequence(Sequence):\n\n    def __init__(self, x_set, y_set, batch_size, seq_length):\n        self.x, self.y = x_set, y_set\n        self.batch_size = batch_size\n        self.seq_length = seq_length\n\n    def __len__(self):\n        return int(np.ceil(len(self.x) \/ float(self.batch_size)))\n\n    def __getitem__(self, idx):\n        # get_batch to be coded\n        return get_batch(idx, self.x, self.y, self.batch_size, self.seq_length)\n<\/code><\/pre>\n\n<p>What is the most efficient way of getting these batches using a sequence? <\/p>\n\n<p>My solution was to actually not use the Sequence and instead use a custom generator that spits out data without knowing how many batches will there be in advance. and use <code>fit_generator(custom_generator, max_queue_size=batch_size)<\/code> instead. Is this the most efficient way? The problem here is that there is no shuffling and it could be a problem?<\/p>\n\n<p>Desired output for batchsize=2, seq_length=3 is:<\/p>\n\n<pre><code>X = [ \n        [ [6], [5], [4] ], \n        [ [5], [4], [3] ] \n    ]\n\nY = [ 1, 1 ]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52599385","CreationDate":"2018-10-01T21:53:58.027","ParentId":null,"OwnerUserId":"1050233","Title":null,"Body":"<p>It seems you need to not only know the number of batches but also be able to output any batch given just the batch number. You could create an index of all samples in <code>RNNSequence.__init__<\/code> or earlier, then assemble batches from this. In <code>__getitem__<\/code> you can then output the batches accordingly.<\/p>\n\n<p>This quick and dirty pseudocode should illustrate the concept of the sample index. If needed you might decide to use functions in pandas or numpy instead etc.<\/p>\n\n<pre><code># Pseuducode for generating indexes for where samples start.\nseq_len = 3\nsample_start_ids = []\nfor group_id, group in enumerate(groups):\n    for timestep_id, timestep in enumerate(group_timesteps):\n        # Only add as sample if it is the first\n        # timestep in the group or if a full sample fits.\n        if timestep == 1 or timestep &lt;= len(group_timesteps) - seq_len+1:\n            sample_start_ids.append((group_id, timestep_id))\n\nnum_samples = len(sample_start_ids)\n\n# Group the samples into batches of appropriate size.\npass\n\nnum_batches = len(your_batches)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52594736,"AnswerCount":1,"Tags":"<python-3.x><numpy><machine-learning><keras>","CreationDate":"2018-10-01T15:53:19.700","AcceptedAnswerId":52595372.0,"OwnerUserId":7779411.0,"Title":"Keras expects 2d array and a shape of (1,) simultaneously","Body":"<p>I load a model of keras by<\/p>\n\n<pre><code>model = keras.models.load_model('myModel.h5')\nmodel.fit(x,y, batch_size=1, epochs=2)\n<\/code><\/pre>\n\n<p>Here at the last line it shows an error such that it wants y to be a 2D input. My y file's sample is\n[0,0,-1]\nwhich can be changed to [[0,0,-1]] or [[0],[0],[-1]] in 2Ds\nIn the first case,\nError: got array with shape(3,) where expected shape(1,) \nand in second case\nError: There are 3 samples where expected 1<\/p>\n\n<p>Before this training part I run model.predict() during which it gives an output of format [[0.00251, -0.00526, 0.005476]]<\/p>\n\n<p>What am I missing?<\/p>\n\n<p>Model Definition<\/p>\n\n<pre><code>inputs = Input(shape=(32,32,1))\ndenseModel = DenseNet121(include_top=False, weights=None, input_shape=(32,32,1), pooling='avg')(inputs)\ndenseModel = Dense(32, activation='relu')(denseModel)\ndenseModel = BatchNormalization()(denseModel)\ndenseModelPrediction = Dense(3, activation='tanh', name='predictions')(denseModel)\n\nmyModel = Model(inputs=inputs, outputs=denseModelPrediction)\nmyModel.compile(loss='sparse_categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\nmyModel.save('myModel.h5')\n<\/code><\/pre>\n\n<p>Merely initializing it here<\/p>\n","answers":[{"AnswerId":"52595372","CreationDate":"2018-10-01T16:39:11.950","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>Your loss function (sparse_categorical_crossentropy) is expecting your y values to be one hot encoded. As you have three output nodes, the loss is expecting a y value with a (3,) shape, hence your first error.<\/p>\n\n<p>If you are trying to regress three independent probabilities, I'd suggest switching your loss to something that compares the numbers (eg <code>loss='mse'<\/code> mean squared error). This should fix your input shape issue. You might also want to consider setting your outputs activation to sigmoid, so that it is forced to be between zero and one.<\/p>\n"}]}
+{"QuestionId":52594934,"AnswerCount":1,"Tags":"<tensorflow><tensor>","CreationDate":"2018-10-01T16:05:53.360","AcceptedAnswerId":52596260.0,"OwnerUserId":5443617.0,"Title":"How to replicate a column in tensorflow?","Body":"<p>I have a tensor with shape (k,1) and I want to replicate the first column n times. That is, the resulting tensor should have the shape (k,n) with matrix values copied from the first column. How to achieve this using tf.tile?<\/p>\n","answers":[{"AnswerId":"52596260","CreationDate":"2018-10-01T17:45:29.143","ParentId":null,"OwnerUserId":"8315956","Title":null,"Body":"<p>You can use <code>tf.tile()<\/code> operation. For example:<\/p>\n\n<pre><code>a = tf.constant([[1], [2], [3]], dtype = tf.float32)\nc = tf.tile(a, [1,3])\nwith tf.Session() as sess:\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>returns:<\/p>\n\n<pre><code>[[1. 1. 1.]\n [2. 2. 2.]\n [3. 3. 3.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52594956,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><rotation><elementtree>","CreationDate":"2018-10-01T16:07:37.167","AcceptedAnswerId":null,"OwnerUserId":284199.0,"Title":"How can I find new coordinates of boundary box of rotated image to modify its xml file for Tensorflow data augmentation?","Body":"<p>I'm trying to make more dataset to train my model in Tensorflow for data augmantaion. I added the labels of boundary boxes to original image. I want to rotate image 45 degree and modify the xml file for the new exact boundary box(rectangle) to label new created image. It's resizing and fetching to window to not to loose anything on image. <\/p>\n\n<p>Let me show you how I try:<\/p>\n\n<pre><code>def rotateImage(mat, angle):\n    height, width = mat.shape[:2]\n    image_center = (width \/ 2, height \/ 2)\n\n    rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1)\n\n    radians = math.radians(angle)\n    sin = math.sin(radians)\n    cos = math.cos(radians)\n    bound_w = int((height * abs(sin)) + (width * abs(cos)))\n    bound_h = int((height * abs(cos)) + (width * abs(sin)))\n\n    rotation_mat[0, 2] += ((bound_w \/ 2) - image_center[0])\n    rotation_mat[1, 2] += ((bound_h \/ 2) - image_center[1])\n\n    rotated_mat = cv2.warpAffine(mat, rotation_mat, (bound_w, bound_h))\n    return rotated_mat\n\n\nimage = cv2.imread(\"test.jpg\")\n\nangle = 45\n\nrotated_45_image = image.copy()\n\nrotated_45_image = rotateImage(rotated_45_image, angle=45)\n\ntree_for_45_rotated = ET.parse(file_name + \".xml\")\nroot = tree_for_xml.getroot()\n\nfor object in root.iter(\"object\"):\n    xmin = object.find(\"bndbox\").find(\"xmin\")\n    ymin = object.find(\"bndbox\").find(\"ymin\")\n    xmax = object.find(\"bndbox\").find(\"xmax\")\n    ymax = object.find(\"bndbox\").find(\"ymax\")\n    print(xmin.text, ymin.text, xmax.text, ymax.text)\n    print(\"new\")\n    new_xmin = math.cos(angle) * int(xmin.text) - math.sin(angle) * int(ymin.text)\n    new_xmax = math.cos(angle) * int(xmax.text) - math.sin(angle) * int(ymin.text)\n    new_ymin = math.sin(angle) * int(xmin.text) + math.cos(angle) * int(ymin.text)\n    new_ymax = math.sin(angle) * int(xmin.text) + math.cos(angle) * int(ymax.text)\n    print(new_xmin, new_ymin, new_xmax, new_ymax)\n<\/code><\/pre>\n\n<p>After rotation the image is like this:\n<a href=\"https:\/\/i.stack.imgur.com\/v7ZZZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/v7ZZZ.jpg\" alt=\"Rotated 45\"><\/a><\/p>\n\n<p>By the way, I'm using Python and OpenCV. I can't calculate the exact new coordinates to label the image.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52598357","CreationDate":"2018-10-01T20:22:36.103","ParentId":null,"OwnerUserId":"9671732","Title":null,"Body":"<p>I cannot add a comment in the post above, so sorry for the post.  All you need is print after rotation values from <em>corners<\/em><\/p>\n\n<pre><code>img = cv2.imread(\"test.jpg\")\nrotated, corners = rotateImage(img, 30)\nprint(corners)\n<\/code><\/pre>\n\n<p>and if you want a specific value just use <\/p>\n\n<pre><code>print(corners[0])\nprint(corners[1])\nprint(corners[2])\nprint(corners[3])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52594994,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><loss-function>","CreationDate":"2018-10-01T16:10:04.937","AcceptedAnswerId":null,"OwnerUserId":9855596.0,"Title":"How to get the horizontal and vertical gradient of the difference between y_true and y_pred?","Body":"<p>I want to define a custom loss function using Keras, which contains the gradient of the difference between y_true and y_pred.\nI found <code>numpy.gradient<\/code> can help me get the gradient of an array.\nSo part of my code for loss function looks like this:<\/p>\n\n<pre><code>def loss(y_true, y_pred):\n    d   = y_true - y_pred\n    gradient_x = np.gradient(d, axis=0)\n    gradient_y = np.gradient(d, axis=1)\n<\/code><\/pre>\n\n<p>but it turns out <code>d<\/code> is a Tensorflow tensor class and <code>numpy.gradient<\/code> can't process it.\nI'm kind of new to Keras and Tensorflow. <\/p>\n\n<p>Is there any other function can help me do this? Or I have to compute the gradient by myself?<\/p>\n","answers":[{"AnswerId":"52596025","CreationDate":"2018-10-01T17:26:52.960","ParentId":null,"OwnerUserId":"6937913","Title":null,"Body":"<p>Tensorflow tensors are not arrays at all when they are executed, they are only references to a computational graph that is being built. You might want to review the <a href=\"https:\/\/www.tensorflow.org\/guide\/graphs\" rel=\"nofollow noreferrer\">tutorial on how Tensorflow builds graphs<\/a>.<\/p>\n\n<p>You have two problems with your loss function: first, collapsing on either axis will not yield a scalar, so it won't be possible to take a derivative and second, <code>np.gradient<\/code> doesn't appear to exist in Tensorflow.<\/p>\n\n<p>For the first problem, you can solve it by reducing along the remaining axis of <code>gradient_y<\/code> or <code>gradient_x<\/code>. I don't know which function you might want to use because I don't know your application.<\/p>\n\n<p>The second problem can be fixed in two ways:<\/p>\n\n<ol>\n<li>You could wrap <code>np.gradient<\/code> using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\"><code>py_func<\/code><\/a>, but you plan to use this as a loss function, so you will want to take the gradient of that function, and defining the gradient of a <code>py_func<\/code> call is <a href=\"https:\/\/stackoverflow.com\/a\/41535880\/6937913\">complicated<\/a>.<\/li>\n<li>Write your own version of <code>np.gradient<\/code> using pure Tensorflow.<\/li>\n<\/ol>\n\n<p>For example, here's a 1D <code>np.gradient<\/code> in tensorflow (<em>untested<\/em>):<\/p>\n\n<pre><code>def gradient(x):\n    d = x[1:]-x[:-1]\n    fd = tf.concat([x,x[-1]], 0).expand_dims(1)\n    bd = tf.concat([x[0],x], 0).expand_dims(1)\n    d = tf.concat([fd,bd], 1)\n    return tf.reduce_mean(d,1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52595477,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><tensor>","CreationDate":"2018-10-01T16:48:25.800","AcceptedAnswerId":52595548.0,"OwnerUserId":626537.0,"Title":"Why is this Keras Conv2D layer not compatible with the input?","Body":"<p>I am having trouble understanding what input shapes my first convolutional neural network expects. <\/p>\n\n<p>My training set is 500 grayscale images of 50x50 pixels. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nvMSN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nvMSN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The network starts with a <code>Conv2D<\/code> layer. Documentation for the argument <code>input_shape<\/code> says:<\/p>\n\n<pre><code>Input shape:\n    4D tensor with shape:\n    `(samples, channels, rows, cols)` if data_format='channels_first'\n    or 4D tensor with shape:\n    `(samples, rows, cols, channels)` if data_format='channels_last'.\n<\/code><\/pre>\n\n<p>So I expected that I need to supply my images (which are so far stored in a column of a <code>pandas.DataFrame<\/code>) as a <code>numpy.array<\/code> of the shape <code>(500, 1, 50, 50)<\/code>, since I only have one \"color\" channel in the images. I reshaped it as follows:<\/p>\n\n<pre><code>X = np.array([img for img in imgs[\"img_res\"]])\nX = X.reshape(-1, 1, img_size, img_size)\n<\/code><\/pre>\n\n<p><code>X.shape<\/code> is now: (500, 1, 50, 50). I supplied that as an arguent to <code>Conv2D<\/code>.<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\n    tf.keras.layers.Conv2D(filters=64, \n                           kernel_size=(3,3), \n                           input_shape=X.shape[1:], \n                           activation=\"relu\"),\n])\n<\/code><\/pre>\n\n<p>This produces the following error. Can you point out what is wrong here?<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)\n   1566   try:\n-&gt; 1567     c_op = c_api.TF_FinishOperation(op_desc)\n   1568   except errors.InvalidArgumentError as e:\n\nInvalidArgumentError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d\/Conv2D' (op: 'Conv2D') with input shapes: [?,1,50,50], [3,3,50,64].\n\nDuring handling of the above exception, another exception occurred:\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-24-0b665136e60b&gt; in &lt;module&gt;()\n      3                            kernel_size=(3,3),\n      4                            input_shape=X.shape[1:],\n----&gt; 5                            activation=\"relu\"),\n      6     #tf.keras.layers.MaxPool2D(pool_size=(2,2)),\n      7     #tf.keras.layers.Conv2D(filters=64,\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/sequential.py in __init__(self, layers, name)\n     99     if layers:\n    100       for layer in layers:\n--&gt; 101         self.add(layer)\n    102 \n    103   @property\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/sequential.py in add(self, layer)\n    162           # and create the node connecting the current layer\n    163           # to the input layer we just created.\n--&gt; 164           layer(x)\n    165           set_inputs = True\n    166         else:\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    312     \"\"\"\n    313     # Actually call the layer (optionally building it).\n--&gt; 314     output = super(Layer, self).__call__(inputs, *args, **kwargs)\n    315 \n    316     if args and getattr(self, '_uses_inputs_arg', True):\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/base.py in __call__(self, inputs, *args, **kwargs)\n    715 \n    716         if not in_deferred_mode:\n--&gt; 717           outputs = self.call(inputs, *args, **kwargs)\n    718           if outputs is None:\n    719             raise ValueError('A layer\\'s `call` method should return a Tensor '\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/convolutional.py in call(self, inputs)\n    166 \n    167   def call(self, inputs):\n--&gt; 168     outputs = self._convolution_op(inputs, self.kernel)\n    169 \n    170     if self.use_bias:\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py in __call__(self, inp, filter)\n    866 \n    867   def __call__(self, inp, filter):  # pylint: disable=redefined-builtin\n--&gt; 868     return self.conv_op(inp, filter)\n    869 \n    870 \n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py in __call__(self, inp, filter)\n    518 \n    519   def __call__(self, inp, filter):  # pylint: disable=redefined-builtin\n--&gt; 520     return self.call(inp, filter)\n    521 \n    522 \n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py in __call__(self, inp, filter)\n    202         padding=self.padding,\n    203         data_format=self.data_format,\n--&gt; 204         name=self.name)\n    205 \n    206 \n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_nn_ops.py in conv2d(input, filter, strides, padding, use_cudnn_on_gpu, data_format, dilations, name)\n    954         \"Conv2D\", input=input, filter=filter, strides=strides,\n    955         padding=padding, use_cudnn_on_gpu=use_cudnn_on_gpu,\n--&gt; 956         data_format=data_format, dilations=dilations, name=name)\n    957     _result = _op.outputs[:]\n    958     _inputs_flat = _op.inputs\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    785         op = g.create_op(op_type_name, inputs, output_types, name=scope,\n    786                          input_types=input_types, attrs=attr_protos,\n--&gt; 787                          op_def=op_def)\n    788       return output_structure, op_def.is_stateful, op\n    789 \n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py in create_op(self, op_type, inputs, dtypes, input_types, name, attrs, op_def, compute_shapes, compute_device)\n   3390           input_types=input_types,\n   3391           original_op=self._default_original_op,\n-&gt; 3392           op_def=op_def)\n   3393 \n   3394       # Note: shapes are lazily computed with the C API enabled.\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py in __init__(self, node_def, g, inputs, output_types, control_inputs, input_types, original_op, op_def)\n   1732           op_def, inputs, node_def.attr)\n   1733       self._c_op = _create_c_op(self._graph, node_def, grouped_inputs,\n-&gt; 1734                                 control_input_ops)\n   1735     else:\n   1736       self._c_op = None\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)\n   1568   except errors.InvalidArgumentError as e:\n   1569     # Convert to ValueError for backwards compatibility.\n-&gt; 1570     raise ValueError(str(e))\n   1571 \n   1572   return c_op\n\nValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d\/Conv2D' (op: 'Conv2D') with input shapes: [?,1,50,50], [3,3,50,64].\n<\/code><\/pre>\n","answers":[{"AnswerId":"52595548","CreationDate":"2018-10-01T16:54:07.280","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>Specify that you are not using the default data format by passing <code>data_format='channels_first'<\/code> to Conv2D.<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\n     tf.keras.layers.Conv2D(filters=64, \n                       kernel_size=(3,3), \n                       input_shape=X.shape[1:], \n                       activation=\"relu\", \n                       data_format='channels_first'),\n     ])\n<\/code><\/pre>\n"},{"AnswerId":"52596165","CreationDate":"2018-10-01T17:37:16.760","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You are using TensorFlow, which by default uses the \"channels last\" input format, meaning that the channels dimension should go at the end:<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\ntf.keras.layers.Conv2D(filters=64, \n                       kernel_size=(3,3), \n                       input_shape=(50, 50, 1), \n                       activation=\"relu\"),\n])\n<\/code><\/pre>\n\n<p>The error happens because the 1 in your input shape was being interpreted as one of the spatial dimensions, producing a negative dimension after performing convolution.<\/p>\n"}]}
+{"QuestionId":52595819,"AnswerCount":1,"Tags":"<testing><input><conv-neural-network><caffe><pycaffe>","CreationDate":"2018-10-01T17:13:04.790","AcceptedAnswerId":52722353.0,"OwnerUserId":3802483.0,"Title":"Caffe always returns one label","Body":"<p>I have trained a model with caffe tools under bin and now I am trying to do testing using python script, I read in an image and preprocess it myself (as I did for my training dataset) and I load the pretrained weights to the net, but I am almost always (99.99% of the time) receiving the same result -0- for every test image. I did consider that my model might be overfitting but after training a few models, I have come to realize the labels I get from predictions are most likely the cause. I have also increased dropout and took random crops to overcome overfitting and I have about 60K for training. The dataset is also roughly balanced. I get between 77 to 87 accuracy during evaluation step of training (depending on how I process data, what architecture I use etc)<\/p>\n\n<p>Excuse my super hacky code, I have been distant to caffe testing for some time so I suspect the problem is how I pass the input data to the network, but I can't put my finger on it:<\/p>\n\n<pre><code>import h5py, os\nimport sys\nsys.path.append(\"\/home\/X\/Desktop\/caffe-caffe-0.16\/python\")\nfrom caffe.io import oversample\nfrom caffe.io import resize_image\nimport caffe\nfrom random import randint\nimport numpy as np\nimport cv2\nimport matplotlib.pyplot as plt\nfrom collections import Counter as Cnt\n\nmeanImg = cv2.imread('\/home\/caffe\/data\/Ch\/Final_meanImg.png')\n\nmodel_def = '\/home\/X\/Desktop\/caffe-caffe-0.16\/models\/bvlc_googlenet\/deploy.prototxt'\nmodel_weights = '\/media\/X\/DATA\/SDet\/Google__iter_140000.caffemodel'\n\n# load the model\n#caffe.set_mode_gpu()\n#caffe.set_device(0)\n\nnet = caffe.Net(model_def,      # defines the structure of the model\n                model_weights,  # contains the trained weights\n                caffe.TEST)     # use test mode (e.g., don't perform dropout)\n\n\nwith open( '\/home\/caffe\/examples\/sdet\/SDet\/test_random.txt', 'r' ) as T, open('\/media\/X\/DATA\/SDet\/results\/testResults.txt','w') as testResultsFile:\n\n\n    readImgCounter = 0\n    runningCorrect = 0\n    runningAcc = 0.0\n\n    #testResultsFile.write('filename'+' '+'prediction'+' '+'GT')\n    lines = T.readlines()\n\n    for i,l in enumerate(lines):\n\n\n        sp = l.split(' ')\n\n        video = sp[0].split('_')[0]\n\n        impath =  '\/home\/caffe\/data\/Ch\/images\/'+video+'\/'+sp[0] +'.jpg' \n\n        img = cv2.imread(impath)\n\n        resized_img = resize_image(img, (255,255))\n\n        oversampledImages = oversample([resized_img], (224,224)) #5 crops x 2 mirror flips = return 10 images\n\n        transposed_img = np.zeros( (10, 3, 224, 224), dtype='f4' ) \n        tp =  np.zeros( (1, 3, 224, 224), dtype='f4' ) \n\n        predictedLabels = []\n\n        for j in range(0,oversampledImages.shape[0]-1):\n            transposed_img[j] = oversampledImages[j].transpose((2,0,1))\n\n\n            tp[0] = transposed_img[j]\n            net.blobs['data'].data[0]  = tp\n\n            pred =  net.forward(data=tp)\n\n            predictedLabels.append(pred['prob'].argmax())\n\n\n        print(predictedLabels)\n\n        prediction,num_most_common = Cnt(predictedLabels).most_common(1)[0]\n\n\n        print(prediction)\n\n        readImgCounter = readImgCounter + 1\n\n        if (prediction == int(sp[1])):\n            runningCorrect = runningCorrect + 1\n\n        runningAcc = runningCorrect \/ readImgCounter\n        print('runningAcc:')\n        print(runningAcc)\n        print('-----------')\n        print('runningCorrect:')\n        print(runningCorrect)\n        print('-----------')\n        print('totalImgRead:')\n        print(readImgCounter)\n        print('-----------')\n\n        testResultsFile.write(sp[0]+' '+str(prediction)+' '+sp[1])\n        testResultsFile.write('\\n')\n<\/code><\/pre>\n","answers":[{"AnswerId":"52722353","CreationDate":"2018-10-09T13:32:25.683","ParentId":null,"OwnerUserId":"3802483","Title":null,"Body":"<p>I have fixed this problem eventually. I am not 100% sure what worked but it was most likely changing the bias to 0 while learning. <\/p>\n"}]}
+{"QuestionId":52596139,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-10-01T17:35:24.130","AcceptedAnswerId":52596617.0,"OwnerUserId":1050648.0,"Title":"mapping labels to RGB in tensorflow","Body":"<p>How can we map integer labels to different pseudo-colors in tensorflow. I tried:<\/p>\n\n<pre><code>with sess.as_default():\na = tf.random_uniform(shape=[3,3,3],minval=0,maxval=5,dtype=tf.int32)\nkeys = [0,1,2,3,4,5]\nvalues = [0,10,20,30,40,50]\ntable = tf.contrib.lookup.HashTable(tf.contrib.lookup.KeyValueTensorInitializer(keys, values), -1)\nout = table.lookup(a)\ntable.init.run()\nprint(out.eval())\n<\/code><\/pre>\n\n<p>But this option doesn't solve my use-case. For each unique label in the input tensor [height x width x channels=1], I want to map it to [height x width x channels =3 ]<\/p>\n\n<pre><code>labels                LUT\n-------               ----\n   0                 [0,0,0]\n   1                 [128,64,128]\n   2                 [64,128,256]\n   3                 [255,64,128]\n<\/code><\/pre>\n\n<p>Appreciate some help.<\/p>\n","answers":[{"AnswerId":"52596617","CreationDate":"2018-10-01T18:12:17.110","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>You can use <code>tf.nn.embedding_lookup<\/code> to map the <code>labels<\/code> to <code>LUT<\/code> table:<\/p>\n\n<pre><code>LUT = tf.constant([[0,0,0],[128,64,128],[64,128,256],[255,64,128]], tf.int32)\nlabels = [0, 1, 2, 3]\nout = tf.nn.embedding_lookup(LUT, labels)\n\nwith tf.Session() as sess:\n   print(sess.run(out))\n\n#[[  0   0   0]\n# [128  64 128]\n# [ 64 128 256]\n# [255  64 128]]\n<\/code><\/pre>\n"},{"AnswerId":"57605592","CreationDate":"2019-08-22T08:56:59.103","ParentId":null,"OwnerUserId":"1822293","Title":null,"Body":"<p>A complete example using Google Colab (Links on the bottom)<\/p>\n\n<pre><code>import tensorflow as tf\nfrom google.colab import drive\nfrom matplotlib import pyplot as plt\nimport numpy as np\nimport cv2\n\ndrive.mount('\/content\/drive')\n\nsess = tf.Session()\n\n# an input image\n\nimg = cv2.imread('\/content\/drive\/My Drive\/Colab Notebooks\/test.png')\nimg = img[:, :, 0]\nprint(img.shape)\n\nlabel = tf.placeholder(tf.int32,shape=[None,None])\n\nLUTval = np.zeros((256,3)) # lookup table must span all possible labels, classes are 0 to 19, 255 is unlabelled in this example\nLUTval[0:19, :] = [[128, 64, 128],[244, 35, 232],[70, 70, 70],[102, 102, 156],[190, 153, 153],[153, 153, 153],[250, 170, 30],[220, 220, 0],[107, 142, 35],[152, 251, 152],[70, 130, 180],[220, 20, 60],[255, 0, 0],[0, 0, 142],[0, 0, 70],[0, 60, 100],[0, 80, 100],[0, 0, 230],[119, 11, 32]]\nLUT = tf.constant(LUTval, tf.int32)\ncol = tf.nn.embedding_lookup(LUT, label)\n\nres_col = sess.run(col, feed_dict = {label: img})\n\nprint(res_col.shape)\n\n#img2 = img[:,:,::-1]\nplt.imshow(cv2.cvtColor(np.array(res_col).astype(np.uint8), cv2.COLOR_BGR2RGB))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1bCHOUwLJMvFjxJMqcy9M6de8d5_jXlCW\" rel=\"nofollow noreferrer\">Snippet on Google Colab<\/a>\n<a href=\"https:\/\/drive.google.com\/open?id=1mdz0K_YAEvR5xVdSQUQlTnfnq9HqvtjC\" rel=\"nofollow noreferrer\">Test Image<\/a><\/p>\n"}]}
+{"QuestionId":52596891,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T18:31:16.820","AcceptedAnswerId":null,"OwnerUserId":10441876.0,"Title":"How to score model saved using Tensorflow estimator?","Body":"<p>All,<\/p>\n\n<p>I built a customized model for binary image classification. I managed to successfully save model using tf estimator to .pb format. My jpg image files have images in various sizes, so I have a image transformation step to transform the images to 224x224. Here is how I define serving_input_fn():<\/p>\n\n<pre><code>def parse_function_test(filename):\n    image_string=tf.read_file(filename)\n    image=tf.image.decode_jpeg(image_string, channels=3)\n    shape = tf.shape(image)\n    init_width = shape[0]\n    init_height = shape[1]\n    max_size = 224\n    resized_image = resize_image_keep_aspect(image,init_width,init_height,max_size)\n    image_padded = tf.image.resize_image_with_crop_or_pad(resized_image,max_size,max_size)\n    final_image_padded=tf.image.convert_image_dtype(image_padded,dtype=tf.float32)\n\n    return final_image_padded\n\n\ndef serving_input_fn():\n    serialized_tf_example=tf.placeholder(dtype=tf.string,shape=1,name='input_tensor')\n    receiver_tensors={'inputs':serialized_tf_example}\n    feature_spec ={'image\/encoded':tf.FixedLenFeature([],dtype=tf.string)}\n    features=tf.parse_example(serialized_tf_example,feature_spec)\n    jpegs=features['image\/encoded']\n    images=tf.map_fn(parse_function_test,jpegs,dtype=tf.float32)\n\n    return tf.estimator.export.ServingInputReceiver(images,receiver_tensors)\n<\/code><\/pre>\n\n<p>This is the scoring script. I tried to score on one image file. <\/p>\n\n<pre><code>exported_path='.\/1538070515'\ntestimg = '.\/test.jpg'\n\nfilename = tf.constant([testimg])\n\nimage= tf.map_fn(parse_function_test,filename,dtype=tf.float32)\ndef _bytes_feature(value):\n  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef main():\n    with tf.Session() as sess:\n         value=sess.run(image)\n         tf.saved_model.loader.load(sess, \n             [tf.saved_model.tag_constants.SERVING], exported_path)\n         model_input=tf.train.Example(features=tf.train.Features(feature={'image\/encoded':_bytes_feature(value.tostring())}))\n         predictor= tf.contrib.predictor.from_saved_model(exported_path)\n         input_tensor=tf.get_default_graph().get_tensor_by_name(\"input_tensor:0\")\n         print(input_tensor)\n         model_input = model_input.SerializeToString()\n         output_dict=predictor({'inputs':[model_input]})\n         print(\"probability is\",output_dict['probabilities'])\n\nif __name__==\"__main__\":\n    main()\n<\/code><\/pre>\n\n<p>Got error on output_dict=predictor({'inputs':[model_input]}) line says \"<strong>NotFoundError: NewRandomAccessFile failed to Create\/Open:<\/strong>\" <\/p>\n\n<p>Where did I do wrong? I am not sure if it was correct to transformation image in the scoring script first then make it a bytes_features... Or maybe I did something wrong in serving_input_fn(). <\/p>\n","answers":[{"AnswerId":"52610098","CreationDate":"2018-10-02T14:06:37.400","ParentId":null,"OwnerUserId":"10441876","Title":null,"Body":"<p>I figured it out... <\/p>\n\n<p>All the image transformation inside the map function were built to the output graph already. In the scoring script, just need to encode image name string to bytes, then use this as input. <\/p>\n\n<pre><code>model_input=tf.train.Example(features=tf.train.Features(feature={'image\/encoded':_bytes_feature(testimg.encode())}))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52597523,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-01T19:18:42.820","AcceptedAnswerId":null,"OwnerUserId":9731282.0,"Title":"How to load_weights to a Keras model from a Tensorflow checkpoint","Body":"<p>I have some python code to train a network using Tensorflow's TFRecords and Dataset APIs. I have built the network using tf.Keras.layers, this being arguably the easiest and fastest way. The handy function model_to_estimator()<\/p>\n\n<pre><code>modelTF = tf.keras.estimator.model_to_estimator(\n    keras_model=model,\n    custom_objects=None,\n    config=run_config,\n    model_dir=checkPointDirectory\n)\n<\/code><\/pre>\n\n<p>converts a Keras model to an estimator, which allows us to take advantage of the Dataset API nicely, and automatically save checkpoints to checkPointDirectory during training, and upon training completion. The estimator API presents some invaluable features, such as automatically distributing the workload over multiple GPUs, with, e.g.<\/p>\n\n<pre><code>distribution = tf.contrib.distribute.MirroredStrategy()\nrun_config = tf.estimator.RunConfig(train_distribute=distribution)\n<\/code><\/pre>\n\n<p>Now for big models and lots of data, it is often useful to execute predictions after training using some form of saved model. It seems that as of Tensorflow 1.10 (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19295\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19295<\/a>), a tf.keras.model object supports load_weights() from a Tensorflow checkpoint. This is mentioned briefly in the Tensorflow docs, but not the Keras docs, and I can't find anyone showing an example of this. After defining the model layers again in some new .py, I have tried <\/p>\n\n<pre><code>checkPointPath = os.path.join('.', 'tfCheckPoints', 'keras_model.ckpt.index')\nmodel.load_weights(filepath=checkPointPath, by_name=False)\n<\/code><\/pre>\n\n<p>but this gives a NotImplementedError:<\/p>\n\n<pre><code>Restoring a name-based tf.train.Saver checkpoint using the object-based restore API. This mode uses global names to match variables, and so is somewhat fragile. It also adds new restore ops to the graph each time it is called when graph building. Prefer re-encoding training checkpoints in the object-based format: run save() on the object-based saver (the same one this message is coming from) and use that checkpoint in the future.\n\n2018-10-01 14:24:49.912087:\nTraceback (most recent call last):\n  File \"C:\/Users\/User\/PycharmProjects\/python\/mercury.classifier reductions\/V3.2\/wikiTestv3.2\/modelEvaluation3.2.py\", line 141, in &lt;module&gt;\n    model.load_weights(filepath=checkPointPath, by_name=False)\n  File \"C:\\Users\\User\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 1526, in load_weights\n    checkpointable_utils.streaming_restore(status=status, session=session)\n  File \"C:\\Users\\User\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\checkpointable\\util.py\", line 880, in streaming_restore\n    \"Streaming restore not supported from name-based checkpoints. File a \"\nNotImplementedError: Streaming restore not supported from name-based checkpoints. File a feature request if this limitation bothers you.\n<\/code><\/pre>\n\n<p>I would like to do as suggested by the Warning and use the 'object-based saver' instead, but I haven't found a way to do this via a RunConfig passed to estimator.train(). <\/p>\n\n<p>So is there a better way to get the saved weights back into an estimator for use in prediction? The github thread seems to suggest that this is already implemented (though based on the error, probably in a different way than I am attempting above). Has anyone successfully used load_weights() on a TF checkpoint? I haven't been able to find any tutorials\/examples on how this can be done, so any help is appreciated. <\/p>\n","answers":[{"AnswerId":"52918442","CreationDate":"2018-10-21T18:17:59.120","ParentId":null,"OwnerUserId":"10537190","Title":null,"Body":"<p>I'm not sure but maybe you can change <code>keras_model.ckpt.index<\/code> to <code>keras_model.ckpt<\/code> for test.<\/p>\n"}]}
+{"QuestionId":52597637,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-10-01T19:27:28.590","AcceptedAnswerId":52618633.0,"OwnerUserId":10027028.0,"Title":"Installing Tensorflow for Python 3.6.0","Body":"<p>can somebody explain to me what to do step by step for installation of Tensorflow for Python 3.7.0! (I'm using ipython)!\nMy system is windows 10 64 bit system.\nI do not understand what to do even with all the explanation I find so far.\nSo can somebody pointed me to the exact link I have to click on to install tensor flow? Tnx<\/p>\n\n<pre><code>ImportError                               Traceback (most recent call last)\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in &lt;module&gt;()\n     57 \n---&gt; 58   from tensorflow.python.pywrap_tensorflow_internal import *\n     59   from tensorflow.python.pywrap_tensorflow_internal import __version__\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in &lt;module&gt;()\n     27             return _mod\n---&gt; 28     _pywrap_tensorflow_internal = swig_import_helper()\n     29     del swig_import_helper\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in swig_import_helper()\n     23             try:\n---&gt; 24                 _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n     25             finally:\n\n~\\Anaconda3\\lib\\imp.py in load_module(name, file, filename, details)\n    242         else:\n--&gt; 243             return load_dynamic(name, filename, file)\n    244     elif type_ == PKG_DIRECTORY:\n\n~\\Anaconda3\\lib\\imp.py in load_dynamic(name, path, file)\n    342             name=name, loader=loader, origin=path)\n--&gt; 343         return _load(spec)\n    344 \n\nImportError: DLL load failed: Een initialisatieroutine van de dynamic link library (DLL-bestand) is mislukt.\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-2-fae406f752e9&gt; in &lt;module&gt;()\n      6 from PIL import Image\n      7 from scipy import ndimage\n----&gt; 8 import tensorflow as tf\n      9 from tensorflow.python.framework import ops\n     10 from cnn_utils import *\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\__init__.py in &lt;module&gt;()\n     20 \n     21 # pylint: disable=g-bad-import-order\n---&gt; 22 from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n     23 \n     24 try:\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\__init__.py in &lt;module&gt;()\n     47 import numpy as np\n     48 \n---&gt; 49 from tensorflow.python import pywrap_tensorflow\n     50 \n     51 # Protocol buffers\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in &lt;module&gt;()\n     72 for some common reasons and solutions.  Include the entire stack trace\n     73 above this error message when asking for help.\"\"\" % traceback.format_exc()\n---&gt; 74   raise ImportError(msg)\n     75 \n     76 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long\n\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\gstevens\\Anaconda3\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\gstevens\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: Een initialisatieroutine van de dynamic link library (DLL-bestand) is mislukt.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n","answers":[{"AnswerId":"52618633","CreationDate":"2018-10-03T01:48:34.330","ParentId":null,"OwnerUserId":"5939172","Title":null,"Body":"<p>Summarizing installation steps from the <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">official tensorflow installation guide<\/a> specifically for your case.<\/p>\n\n<p>Since you already have Python 3.6 and Conda installed on your machine (which is good!) we can move on to confirming the other stuffs.<\/p>\n\n<h2>1. Microsoft Visual C++ 2015 Redistributable Update 3<\/h2>\n\n<p>Install the <em>Microsoft Visual C++ 2015 Redistributable Update 3<\/em>. This comes with <em>Visual Studio 2015<\/em> but can be installed separately:<\/p>\n\n<ol>\n<li>Go to the <a href=\"https:\/\/visualstudio.microsoft.com\/vs\/older-downloads\/\" rel=\"nofollow noreferrer\">Visual Studio downloads<\/a>,<\/li>\n<li>Select <em>Redistributables and Build Tools<\/em>, (it is well hidden, use search to look for it)<\/li>\n<li>Download and install the <em>Microsoft Visual C++ 2015 Redistributable Update 3<\/em>.<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QEqnP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QEqnP.png\" alt=\"Microsoft Visual C++ 2015 Redistributable Update 3\"><\/a><\/p>\n\n<h2>2. Install Tensorflow in your Conda virtual environment<\/h2>\n\n<p>Open your Command Prompt. Since you already have the environment created with the name <code>tensorflow<\/code> you just need to:<\/p>\n\n<ol>\n<li>In case you wish to create a new environment, run <code>conda create -n venv pip python=3.6<\/code>. Replace <code>venv<\/code> with a name of your choice for the environment.<\/li>\n<li><code>source activate tensorflow<\/code>. <code>tensorflow<\/code> in this line is the name of the Conda environment you named when you ran <code>conda create -n tensorflow python=3.6<\/code>. Unless you've followed step 1 and made a new environment, replace <code>tensorflow<\/code> with the name you used in step 1.<\/li>\n<li>Now you have activated your virtual environment. Within the virtual environment, install the Tensorflow pip package using its <a href=\"https:\/\/www.tensorflow.org\/install\/pip#package-location\" rel=\"nofollow noreferrer\">complete URL<\/a>, so for your case, just run:<\/li>\n<\/ol>\n\n<p><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.11.0-cp36-cp36m-win_amd64.whl<\/code><\/p>\n\n<p>Please update the URL in the command above according to the URL list if you're installing this again in the future. The command above will become outdated eventually.<\/p>\n\n<h2>3. Verifying the install<\/h2>\n\n<p>Restart your Command Prompt, and again activate your conda environment, then run <code>python -c \"import tensorflow as tf; print(tf.__version__)\"<\/code>, and it should print the version of Tensorflow you have installed, in this case 1.11.0. If it does, then you have successfully installed Tensorflow in your Conda virtual environment. Congratulations!<\/p>\n\n<p>If it doesn't work, comment here so I can modify my answer to further assist you.<\/p>\n"},{"AnswerId":"52600356","CreationDate":"2018-10-01T23:54:27.293","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>Tensorflow does not support Python 3.7 yet (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022<\/a>).<\/p>\n\n<p>Use Python 3.6.<\/p>\n"}]}
+{"QuestionId":52598117,"AnswerCount":1,"Tags":"<python><machine-learning><keras><nlp><keras-layer>","CreationDate":"2018-10-01T20:04:24.927","AcceptedAnswerId":52598591.0,"OwnerUserId":7115193.0,"Title":"How to add addition in concatenate Layers keras","Body":"<pre><code>Y = Dense(2)(Y)\nZ = LSTM(128, return_sequences=False)(X)\nZ = Dense(2)(Z)\nM = concatenate([Y, Z,Y+Z])\nM=Dense(4)(M)\nM = Dense(2)(M)\n# Add a softmax activation\nM = Activation('softmax')(M)\n\n# Create Model instance which converts sentence_indices into X.\nmodel = Model(inputs=sentence_indices, outputs=M)\n\nreturn model\n<\/code><\/pre>\n\n<p>The given block is my code. Here, I have given the partial code. What I want is I want to merge layers using <code>M = concatenate([Y, Z])<\/code>, it is working fine. Then I thought of adding more variables to the <code>Dense<\/code> layer so I add <code>M = concatenate([Y, Z,Y+Z])<\/code>; however, it's not working. It gave me this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/sathiyakugan\/PycharmProjects\/internal-apps\/apps\/support-tools\/EscalationApp\/IMDBmodified.py\", line 213, in &lt;module&gt;\n    model = buildModel((maxLen,), word_to_vec_map, word_to_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/internal-apps\/apps\/support-tools\/EscalationApp\/IMDBmodified.py\", line 206, in buildModel\n    model = Model(inputs=sentence_indices, outputs=M)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 91, in __init__\n    self._init_graph_network(*args, **kwargs)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 235, in _init_graph_network\n    self.inputs, self.outputs)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1412, in _map_graph_network\n    tensor_index=tensor_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1399, in build_map\n    node_index, tensor_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1399, in build_map\n    node_index, tensor_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1399, in build_map\n    node_index, tensor_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1399, in build_map\n    node_index, tensor_index)\n  File \"\/home\/sathiyakugan\/PycharmProjects\/Python\/venv\/lib\/python3.5\/site-packages\/keras\/engine\/network.py\", line 1371, in build_map\n    node = layer._inbound_nodes[node_index]\nAttributeError: 'NoneType' object has no attribute '_inbound_nodes' \n<\/code><\/pre>\n\n<p>Could you please help me to resolve this problem?<\/p>\n","answers":[{"AnswerId":"52598591","CreationDate":"2018-10-01T20:39:52.157","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can use <a href=\"https:\/\/keras.io\/layers\/merge\/#add_1\" rel=\"nofollow noreferrer\"><code>add<\/code><\/a> functional interface (or its equivalent <a href=\"https:\/\/keras.io\/layers\/merge\/#add\" rel=\"nofollow noreferrer\"><code>Add<\/code><\/a> layer):<\/p>\n\n<pre><code>A = add([Y, Z])\nM = concatenate([Y, Z, A])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52598416,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T20:26:40.057","AcceptedAnswerId":52600329.0,"OwnerUserId":9376476.0,"Title":"Tensorflow to optimize a list of tensors","Body":"<p>I have created two lists of tensors that are multiple sizes, for example:<\/p>\n\n<pre><code>n=5\nx=[placeholder(tf.float64,[i*2,None]) for i in range(n)]\ny=[placeholder(tf.float64,[i*2,None]) for i in range(n)]\n<\/code><\/pre>\n\n<p>I want to use x as the input to an auto-encoder, and y as the output:\n1) Is there a better way to group my tensors instead of using a list?\n2) How can I ask tesnsorflow to learn simultaneously from all the tensors within x? should I use tf.group? or is there a better way?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52600329","CreationDate":"2018-10-01T23:51:11.037","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>Pad them into the same size with <code>tf.pad<\/code> and then stack them into a single tensor with <code>tf.stack<\/code>?<\/p>\n\n<p>I don't think there is enough information in your question to give you a much better answer. So I would request your give more details.<\/p>\n"}]}
+{"QuestionId":52598605,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T20:40:46.653","AcceptedAnswerId":52655253.0,"OwnerUserId":9192218.0,"Title":"Tensor had NaN values - TensorFlow faster-rccn training error","Body":"<p>This is the cmd output I get. It varies with how many steps it does before the error but is always less than 20.<\/p>\n\n<pre><code>C:\\Users\\edupt\\Documents\\GitHub\\Project&gt;python object_detection\/train.py \\ --logtostderr \\ --train_dir=train \\ --pipeline_config_path=faster_rcnn_resnet101.config\nINFO:tensorflow:Scale of 0 disables regularizer.\nINFO:tensorflow:Scale of 0 disables regularizer.\nWARNING:tensorflow:From C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\trainer.py:176: create_global_step (from tensorflow.contrib.framework.python.ops.variables) is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease switch to tf.train.create_global_step\nWARNING:tensorflow:From C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\core\\preprocessor.py:1922: calling squeeze (from tensorflow.python.ops.array_ops) with squeeze_dims is deprecated and will be removed in a future version.\nInstructions for updating:\nUse the `axis` argument instead\nINFO:tensorflow:Scale of 0 disables regularizer.\nINFO:tensorflow:Scale of 0 disables regularizer.\nINFO:tensorflow:Scale of 0 disables regularizer.\nINFO:tensorflow:Scale of 0 disables regularizer.\nWARNING:tensorflow:From C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\core\\box_predictor.py:371: calling reduce_mean (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.\nInstructions for updating:\nkeep_dims is deprecated, use keepdims instead\nWARNING:tensorflow:From C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\core\\losses.py:269: softmax_cross_entropy_with_logits (from tensorflow.python.ops.nn_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\n\nFuture major versions of TensorFlow will allow gradients to flow\ninto the labels input on backprop by default.\n\nSee @{tf.nn.softmax_cross_entropy_with_logits_v2}.\n\nWARNING:tensorflow:From C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\builders\\optimizer_builder.py:105: get_or_create_global_step (from tensorflow.contrib.framework.python.ops.variables) is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease switch to tf.train.get_or_create_global_step\nINFO:tensorflow:Summary name Learning Rate is illegal; using Learning_Rate instead.\nINFO:tensorflow:Summary name \/clone_loss is illegal; using clone_loss instead.\nC:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\gradients_impl.py:100: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n  \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\nWARNING:tensorflow:From C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py:737: Supervisor.__init__ (from tensorflow.python.training.supervisor) is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease switch to tf.train.MonitoredTrainingSession\n2018-10-01 21:26:27.032708: I T:\\src\\github\\tensorflow\\tensorflow\\core\\platform\\cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n2018-10-01 21:26:27.350000: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1392] Found device 0 with properties:\nname: GeForce GTX 1060 6GB major: 6 minor: 1 memoryClockRate(GHz): 1.7085\npciBusID: 0000:01:00.0\ntotalMemory: 6.00GiB freeMemory: 4.96GiB\n2018-10-01 21:26:27.356938: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1471] Adding visible gpu devices: 0\n2018-10-01 21:26:29.297942: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:952] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-01 21:26:29.300781: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:958]      0\n2018-10-01 21:26:29.302611: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:971] 0:   N\n2018-10-01 21:26:29.305150: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1084] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4726 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:01:00.0, compute capability: 6.1)\nINFO:tensorflow:Restoring parameters from train\\model.ckpt-0\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Starting Session.\nINFO:tensorflow:Saving checkpoint to path train\\model.ckpt\nINFO:tensorflow:Starting Queues.\nINFO:tensorflow:global_step\/sec: 0\nINFO:tensorflow:Recording summary at step 0.\nINFO:tensorflow:global step 1: loss = 4.7042 (19.631 sec\/step)\nINFO:tensorflow:global step 2: loss = 4.7257 (0.878 sec\/step)\nINFO:tensorflow:global step 3: loss = 4.4725 (0.851 sec\/step)\nINFO:tensorflow:global step 4: loss = 4.2467 (0.832 sec\/step)\nINFO:tensorflow:global step 5: loss = 4.0482 (0.922 sec\/step)\nINFO:tensorflow:global step 6: loss = 3.8669 (0.647 sec\/step)\nINFO:tensorflow:global step 7: loss = 3.7094 (0.731 sec\/step)\nINFO:tensorflow:global step 8: loss = 3.2892 (0.629 sec\/step)\nINFO:tensorflow:global step 9: loss = 3.6964 (0.608 sec\/step)\nINFO:tensorflow:Error reported to Coordinator: &lt;class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'&gt;, LossTensor is inf or nan. : Tensor had NaN values\n         [[Node: CheckNumerics = CheckNumerics[T=DT_FLOAT, message=\"LossTensor is inf or nan.\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](clone_loss\/_3493)]]\n\nCaused by op 'CheckNumerics', defined at:\n  File \"object_detection\/train.py\", line 198, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"object_detection\/train.py\", line 194, in main\n    worker_job_name, is_chief, FLAGS.train_dir)\n  File \"C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\trainer.py\", line 227, in train\n    total_loss = tf.check_numerics(total_loss, 'LossTensor is inf or nan.')\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 968, in check_numerics\n    \"CheckNumerics\", tensor=tensor, message=message, name=name)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3414, in create_op\n    op_def=op_def)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1740, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): LossTensor is inf or nan. : Tensor had NaN values\n         [[Node: CheckNumerics = CheckNumerics[T=DT_FLOAT, message=\"LossTensor is inf or nan.\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](clone_loss\/_3493)]]\n\nTraceback (most recent call last):\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1322, in _do_call\n    return fn(*args)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1307, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1409, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: LossTensor is inf or nan. : Tensor had NaN values\n         [[Node: CheckNumerics = CheckNumerics[T=DT_FLOAT, message=\"LossTensor is inf or nan.\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](clone_loss\/_3493)]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"object_detection\/train.py\", line 198, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"object_detection\/train.py\", line 194, in main\n    worker_job_name, is_chief, FLAGS.train_dir)\n  File \"C:\\Users\\edupt\\Documents\\GitHub\\Project\\object_detection\\trainer.py\", line 296, in train\n    saver=saver)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py\", line 770, in train\n    sess, train_op, global_step, train_step_kwargs)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py\", line 487, in train_step\n    run_metadata=run_metadata)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 900, in run\n    run_metadata_ptr)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1135, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1316, in _do_run\n    run_metadata)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1335, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: LossTensor is inf or nan. : Tensor had NaN values\n         [[Node: CheckNumerics = CheckNumerics[T=DT_FLOAT, message=\"LossTensor is inf or nan.\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](clone_loss\/_3493)]]\n\nCaused by op 'CheckNumerics', defined at:\n  File \"object_detection\/train.py\", line 198, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"object_detection\/train.py\", line 194, in main\n    worker_job_name, is_chief, FLAGS.train_dir)\n  File \"C:\\Users\\edupt\\Documents\\GitHub\\LEGO-ID-Project\\object_detection\\trainer.py\", line 227, in train\n    total_loss = tf.check_numerics(total_loss, 'LossTensor is inf or nan.')\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 968, in check_numerics\n    \"CheckNumerics\", tensor=tensor, message=message, name=name)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3414, in create_op\n    op_def=op_def)\n  File \"C:\\Users\\edupt\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1740, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): LossTensor is inf or nan. : Tensor had NaN values\n         [[Node: CheckNumerics = CheckNumerics[T=DT_FLOAT, message=\"LossTensor is inf or nan.\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](clone_loss\/_3493)]]\n<\/code><\/pre>\n\n<p>I have tried reducing the learning rate and increasing the batch size but this doesn't help. I don't think it is my annotations as I have error checking with those. I have tried a lot of the other stuff that people with similar errors get with no success.<\/p>\n","answers":[{"AnswerId":"52655253","CreationDate":"2018-10-04T21:16:33.247","ParentId":null,"OwnerUserId":"9192218","Title":null,"Body":"<p>Turns out the issue was my annotations after all. This became clear when I noticed it crashing on the same step but changing this position when I would recreate the random ordered TF Record files.<\/p>\n\n<p>The error with was with some of my annotaiton files having the Maximums and Minimums the wrong way around.<\/p>\n"}]}
+{"QuestionId":52598903,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow><loss-function><softmax>","CreationDate":"2018-10-01T21:06:18.103","AcceptedAnswerId":52602472.0,"OwnerUserId":9540764.0,"Title":"Map the output values of a network in 0 or 1 TensorFlow","Body":"<p>I have a network with the output layer of size <code>[3, 13000, 3, 1] (B,H,W,C)<\/code> and I transformed it using <code>tf.reduce_mean<\/code> to obtain an output size <code>[3, 13000, 1]<\/code>.<\/p>\n\n<p>Graphically is this <a href=\"https:\/\/i.stack.imgur.com\/tKbIz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tKbIz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is right?<\/p>\n\n<p>My labels are in the size of <code>[3, 13000, 1]<\/code> as my new output and are all values 0 or 1.<\/p>\n\n<p>Now I have to compute the loss with the labels. To compute this loss I use this formula <code>tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=predict, labels=y))<\/code>, but first I have to transform all the values in the output in 0 or 1. I'm using the <code>tf.nn.softmax<\/code> function but I get all 1.<\/p>\n\n<p>How can I implement a function that maps all the values under a threshold to 0 and above 1? And this threshold should be for example <code>(max value - min value) \/ 2<\/code>. This should also work with the gradient in the backprop step.<\/p>\n","answers":[{"AnswerId":"52602472","CreationDate":"2018-10-02T05:25:42.253","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>Since your prediction is a single class value, when you apply <code>softmax<\/code> on it, its going to be always 1 irrespective of the value: <code>(exp(predict)\/sum(exp(predict)) = exp(predict)\/exp(predict) = 1)<\/code>. Either convert the input to <code>one-hot<\/code> and make the model predict two classes: [0, 1] or use <code>sigmoid cross entropy<\/code> instead.<\/p>\n"}]}
+{"QuestionId":52598959,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-01T21:13:15.177","AcceptedAnswerId":null,"OwnerUserId":8926217.0,"Title":"Constant Validation Accuracy with a high loss in machine learning","Body":"<p>I'm currently trying to do create an image classification model using Inception V3 with 2 classes. I have 1428 images which are balanced about 70\/30. When I run my model I get a pretty high loss of as well as a constant validation accuracy. What might be causing this constant value?<\/p>\n\n<pre><code>data = np.array(data, dtype=\"float\")\/255.0\nlabels = np.array(labels,dtype =\"uint8\")\n\n(trainX, testX, trainY, testY) = train_test_split(\n                            data,labels, \n                            test_size=0.2, \n                            random_state=42) \n\nimg_width, img_height = 320, 320 #InceptionV3 size\n\ntrain_samples =  1145 \nvalidation_samples = 287\nepochs = 20\n\nbatch_size = 32\n\nbase_model = keras.applications.InceptionV3(\n        weights ='imagenet',\n        include_top=False, \n        input_shape = (img_width,img_height,3))\n\nmodel_top = keras.models.Sequential()\nmodel_top.add(keras.layers.GlobalAveragePooling2D(input_shape=base_model.output_shape[1:], data_format=None)),\nmodel_top.add(keras.layers.Dense(350,activation='relu'))\nmodel_top.add(keras.layers.Dropout(0.2))\nmodel_top.add(keras.layers.Dense(1,activation = 'sigmoid'))\nmodel = keras.models.Model(inputs = base_model.input, outputs = model_top(base_model.output))\n\n\nfor layer in model.layers[:30]:\n  layer.trainable = False\n\nmodel.compile(optimizer = keras.optimizers.Adam(\n                    lr=0.00001,\n                    beta_1=0.9,\n                    beta_2=0.999,\n                    epsilon=1e-08),\n                    loss='binary_crossentropy',\n                    metrics=['accuracy'])\n\n#Image Processing and Augmentation \ntrain_datagen = keras.preprocessing.image.ImageDataGenerator(\n          zoom_range = 0.05,\n          #width_shift_range = 0.05, \n          height_shift_range = 0.05,\n          horizontal_flip = True,\n          vertical_flip = True,\n          fill_mode ='nearest') \n\nval_datagen = keras.preprocessing.image.ImageDataGenerator()\n\n\ntrain_generator = train_datagen.flow(\n        trainX, \n        trainY,\n        batch_size=batch_size,\n        shuffle=True)\n\nvalidation_generator = val_datagen.flow(\n                testX,\n                testY,\n                batch_size=batch_size)\n\nhistory = model.fit_generator(\n    train_generator, \n    steps_per_epoch = train_samples\/\/batch_size,\n    epochs = epochs, \n    validation_data = validation_generator, \n    validation_steps = validation_samples\/\/batch_size,\n    callbacks = [ModelCheckpoint])\n<\/code><\/pre>\n\n<p>This is my log when I run my model:<\/p>\n\n<pre><code>Epoch 1\/20\n35\/35 [==============================]35\/35[==============================] - 52s 1s\/step - loss: 0.6347 - acc: 0.6830 - val_loss: 0.6237 - val_acc: 0.6875\n\nEpoch 2\/20\n35\/35 [==============================]35\/35 [==============================] - 14s 411ms\/step - loss: 0.6364 - acc: 0.6756 - val_loss: 0.6265 - val_acc: 0.6875\n\nEpoch 3\/20\n35\/35 [==============================]35\/35 [==============================] - 14s 411ms\/step - loss: 0.6420 - acc: 0.6743 - val_loss: 0.6254 - val_acc: 0.6875\n\nEpoch 4\/20\n35\/35 [==============================]35\/35 [==============================] - 14s 414ms\/step - loss: 0.6365 - acc: 0.6851 - val_loss: 0.6289 - val_acc: 0.6875\n\nEpoch 5\/20\n35\/35 [==============================]35\/35 [==============================] - 14s 411ms\/step - loss: 0.6359 - acc: 0.6727 - val_loss: 0.6244 - val_acc: 0.6875\n\nEpoch 6\/20\n35\/35 [==============================]35\/35 [==============================] - 15s 415ms\/step - loss: 0.6342 - acc: 0.6862 - val_loss: 0.6243 - val_acc: 0.6875\n<\/code><\/pre>\n","answers":[{"AnswerId":"52759335","CreationDate":"2018-10-11T11:48:45.713","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I think you have too low learning rate and too few epochs. try with <code>lr = 0.001<\/code> and <code>epochs = 100<\/code>. <\/p>\n"},{"AnswerId":"52812040","CreationDate":"2018-10-15T07:49:56.310","ParentId":null,"OwnerUserId":"602721","Title":null,"Body":"<p>Your accuracy is 68.25%. Given that your classes are split roughly 70\/30 it is likely that your model is just predicting the same thing every time, ignoring the input. That would give the accuracy you are seeing. Your model has not yet learned from your data. <\/p>\n\n<p>As Novak said, your learning rate seems very low, so maybe try increasing that first to see if that helps. <\/p>\n"}]}
+{"QuestionId":52599110,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2018-10-01T21:27:00.673","AcceptedAnswerId":null,"OwnerUserId":10443246.0,"Title":"Tensorflow TocoConverter gives toco_from_protos error","Body":"<p>I basically used freeze_graph to freeze the model and then tried to use    TocoConverter to convert the model to tflite, however it gives me the error:<\/p>\n\n<pre><code>RuntimeError: TOCO failed see console for info.\nb'\/bin\/sh: 1: toco_from_protos: not found\\n'\nNone \n<\/code><\/pre>\n\n<p>Below is my code:<\/p>\n\n<pre><code>graph_def_file = '.\/frozen_model2.pb'\ninput_arrays = ['IteratorGetNext']\noutput_arrays = ['model\/fc_result\/prediction\/BiasAdd']\n\nconverter = tf.contrib.lite.TocoConverter.from_frozen_graph(\ngraph_def_file, input_arrays, output_arrays)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>Any help would be much appreciated!!!<\/p>\n","answers":[{"AnswerId":"56283420","CreationDate":"2019-05-23T21:55:11.810","ParentId":null,"OwnerUserId":"5727285","Title":null,"Body":"<p>You have two options <\/p>\n\n<ol>\n<li>Download and build tensorflow from source<\/li>\n<li>Use google collab if you don't want to install tensorflow from\nsource <a href=\"https:\/\/youtu.be\/IJt_veyTlv8\" rel=\"nofollow noreferrer\">Here i made video on how to fix this Error<\/a><\/li>\n<\/ol>\n"},{"AnswerId":"53075836","CreationDate":"2018-10-31T03:26:38.440","ParentId":null,"OwnerUserId":"3700490","Title":null,"Body":"<p>The error <code>b'\/bin\/sh: 1: toco_from_protos: not found\\n'<\/code> indicates that the shell command <code>toco_from_protos<\/code> is not found. I'm guessing <code>toco --help<\/code> doesn't work for you either.<\/p>\n\n<p>I built Tensorflow 1.9.0 from source so I assume I must have screwed up a step where shell commands like <code>toco<\/code> and <code>toco_from_protos<\/code> get set up.<\/p>\n\n<p>Anyway, my traceback was this:<\/p>\n\n<pre><code>  File \"\/home\/casey\/anaconda3\/envs\/mnist\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", line 330, in convert\n    dump_graphviz_video=self.dump_graphviz_video)\n  File \"\/home\/casey\/anaconda3\/envs\/mnist\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 263, in toco_convert\n    input_data.SerializeToString())\n  File \"\/home\/casey\/anaconda3\/envs\/mnist\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 107, in toco_convert_protos\n    (stdout, stderr))\nRuntimeError: TOCO failed see console for info.\nb'\/bin\/sh: 1: toco_from_protos: not found\\n'\nNone\n<\/code><\/pre>\n\n<p>If you follow the traceback to line 107 in <code>[python path]\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py<\/code>, you'll see there is some logic in the definition for <code>toco_convert_protos()<\/code> to run <code>toco_from_protos<\/code> as a shell command: <\/p>\n\n<pre><code># TODO(aselle): When toco does not use fatal errors for failure, we can\n# switch this on.\nif not _toco_from_proto_bin:\n    return _toco_python.TocoConvert(\n        model_flags_str, toco_flags_str, input_data_str)\n\nwith _tempfile.NamedTemporaryFile() as fp_toco, \\\n           _tempfile.NamedTemporaryFile() as fp_model, \\\n           _tempfile.NamedTemporaryFile() as fp_input, \\\n           _tempfile.NamedTemporaryFile() as fp_output:\n    fp_model.write(model_flags_str)\n    fp_toco.write(toco_flags_str)\n    fp_input.write(input_data_str)\n    fp_model.flush()\n    fp_toco.flush()\n    fp_input.flush()\n\n    cmd = [\n        _toco_from_proto_bin, fp_model.name, fp_toco.name, fp_input.name,\n        fp_output.name\n    ]\n    cmdline = \" \".join(cmd)\n    proc = _subprocess.Popen(\n        cmdline,\n        shell=True,\n        stdout=_subprocess.PIPE,\n        stderr=_subprocess.STDOUT,\n        close_fds=True)\n    stdout, stderr = proc.communicate()\n    exitcode = proc.returncode\n    if exitcode == 0:\n      stuff = fp_output.read()\n      return stuff\n    else:\n      raise RuntimeError(\"TOCO failed see console for info.\\n%s\\n%s\\n\" %\n                         (stdout, stderr))\n<\/code><\/pre>\n\n<p>If your <code>toco<\/code> and <code>toco_from_protos<\/code> shell commands aren't working then obviously this step will fail.<\/p>\n\n<p>I made this change as a workaround in <code>[python path]\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py<\/code>, to force the first case to return:<\/p>\n\n<pre><code>  if _toco_from_proto_bin:\n  # if not _toco_from_proto_bin: # comment out this line to force return\n    return _toco_python.TocoConvert(\n        model_flags_str, toco_flags_str, input_data_str)\n<\/code><\/pre>\n\n<p>Janky, but it got the job done for me. The proper solution would probably be to reinstall a version of Tensorflow that has working <code>toco<\/code> or redo the bazel build if you are building from source.<\/p>\n"},{"AnswerId":"55125313","CreationDate":"2019-03-12T15:33:13.887","ParentId":null,"OwnerUserId":"11192158","Title":null,"Body":"<p>Make sure that the toco scripts are in your search path. Check PATH variable.<\/p>\n\n<p>It seems that they are installed in <code>\/home\/user\/.local\/bin<\/code> so add this to <code>PATH<\/code> variable.<\/p>\n"},{"AnswerId":"58988445","CreationDate":"2019-11-22T06:10:41.803","ParentId":null,"OwnerUserId":"3829943","Title":null,"Body":"<p>I just experienced this problem when quantizing a model into integer-precision.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>RuntimeError: TOCO failed see console for info.\nb'\/bin\/sh: 1: toco_from_protos: not found\\n'\n<\/code><\/pre>\n\n<p>The key is to find where does <code>toco_from_protos<\/code> actually reside.<\/p>\n\n<p>I use <code>locate<\/code> command to search across ALL files in my Linux system. (newly installed Linux may need to call <code>updatedb<\/code> on the first time, to create an initial database).<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>bash   &gt; locate toco_from\n\nresult &gt; \/usr\/local\/intelpython3\/bin\/toco_from_protos\n<\/code><\/pre>\n\n<p>Bingo, let see if there are more things in this folder<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>ls \/usr\/local\/intelpython3\/bin\/toco*\n\n\/usr\/local\/intelpython3\/bin\/toco\n\/usr\/local\/intelpython3\/bin\/toco_from_protos\n<\/code><\/pre>\n\n<p><strong>My custom python installation requires manual configuration in order to be properly exposed to PATH<\/strong><\/p>\n\n<p>So I added the following lines in my <code>~\/.profile<\/code><\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>PYTHONBIN=\/usr\/local\/intelpython3\/bin\nexport PATH=$PATH:$PYTHONBIN\n<\/code><\/pre>\n\n<p>and I made it in effect immediately by:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>source ~\/.profile\n<\/code><\/pre>\n\n<p>Then everything goes well again.<\/p>\n"}]}
+{"QuestionId":52599146,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-10-01T21:29:48.417","AcceptedAnswerId":null,"OwnerUserId":1996842.0,"Title":"Loading a Pytorch model into C++ Using Dev Pytorch 1.0","Body":"<p>Pytorch 1.0 has a feature to convert a model into a torch script program (serialized in a way) to enable its execution in C++ without Python dependencies. <\/p>\n\n<p>The details are in this tutorial. \n<a href=\"https:\/\/pytorch.org\/tutorials\/advanced\/cpp_export.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/advanced\/cpp_export.html<\/a><\/p>\n\n<p>This is how it is done:<\/p>\n\n<pre><code>import torch\nimport torchvision\n\n# An instance of your model.\nmodel = A UNET MODEL FROM FASTAI which has hooks as required by UNET\n\n# An example input you would normally provide to your model's forward() method.\nexample = torch.rand(1, 3, 224, 224)\n\n# Use torch.jit.trace to generate a torch.jit.ScriptModule via tracing.\ntraced_script_module = torch.jit.trace(model, example)\n<\/code><\/pre>\n\n<p>In my use case, I am using a UNET Model for semantic segmentation. However, I trace the model using this method, I get the following error.<\/p>\n\n<pre><code>Forward or backward hooks can't be compiled \n<\/code><\/pre>\n\n<p>UNET Model uses hooks to save intermediate features which is used at later layers in the network. Is there a way around it? or This is still a limitation of this new method that it cannot work with Models using such hooks. <\/p>\n","answers":[{"AnswerId":"53643392","CreationDate":"2018-12-06T01:42:03.753","ParentId":null,"OwnerUserId":"1490517","Title":null,"Body":"<p>Maybe you could rewrite the model in C++, since c++ API has almost the same interface as python version.<\/p>\n"}]}
+{"QuestionId":52600127,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><deep-learning><generator>","CreationDate":"2018-10-01T23:20:41.973","AcceptedAnswerId":null,"OwnerUserId":10443041.0,"Title":"Keras fit_generator and fit results are different","Body":"<p>I am training a Convolutional Neural Network using face images dataset. The dataset has 10,000 images of dimensions 700 x 700. My model has 12 layers. I am using a generator function to read images into Keras fit_generator function as below.<\/p>\n\n<p>train_file_names ==> Python list containing filenames of training instances<br\/>\ntrain_class_labels ==> Numpy array of one-hot encoded class lables ([0, 1, 0], [0, 0, 1] etc.)<br\/>\ntrain_data ==> Numpy array of training instances<br\/>\ntrain_steps_epoch ==> 16 (Batch size is 400 and I have 6400 instances for training. Hence it takes 16 iterations for a single pass through the whole dataset)<br\/>\nbatch_size ==> 400<br\/>\ncalls_made ==> When generator reaches end of training instances, it resets indexes to load data from first index in next epoch.<br\/><\/p>\n\n<p>I am passing this generator as an argument to keras 'fit_generator' function to generate new batch of data for each epoch.<br\/><\/p>\n\n<p>val_data, val_class_labels ==> Validation data numpy arrays<br\/>\nepochs ==> No. of epochs<br\/><\/p>\n\n<p><b> Using Keras fit_generator <\/b>:<\/p>\n\n<pre><code>model.fit_generator(generator=train_generator, steps_per_epoch=train_steps_per_epoch, epochs=epochs, use_multiprocessing=False, validation_data=[val_data, val_class_labels], verbose=True, callbacks=[history, model_checkpoint], shuffle=True, initial_epoch=0) \n<\/code><\/pre>\n\n<p><b>Code<\/b><\/p>\n\n<pre><code>def train_data_generator(self):     \n    index_start = index_end = 0 \n    temp = 0\n    calls_made = 0\n\n    while temp &lt; train_steps_per_epoch:\n        index_end = index_start + batch_size\n        for temp1 in range(index_start, index_end):\n            index = 0\n            # Read image\n            img = cv2.imread(str(TRAIN_DIR \/ train_file_names[temp1]), cv2.IMREAD_GRAYSCALE).T\n            train_data[index]  = cv2.resize(img, (self.ROWS, self.COLS), interpolation=cv2.INTER_CUBIC)\n            index += 1       \n        yield train_data, self.train_class_labels[index_start:index_end]\n        calls_made += 1\n        if calls_made == train_steps_per_epoch:\n            index_start = 0\n            temp = 0\n            calls_made = 0\n        else:\n            index_start = index_end\n            temp += 1  \n        gc.collect()\n<\/code><\/pre>\n\n<p><b> Output of fit_generator <\/b><\/p>\n\n<p>Epoch 86\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.5739 - acc: 0.2991 - val_loss: 12.0076 - val_acc: 0.2110<br\/>\nEpoch 87\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.6010 - acc: 0.2549 - val_loss: 11.6689 - val_acc: 0.2016<br\/>\nEpoch 88\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.5750 - acc: 0.2391 - val_loss: 10.2663 - val_acc: 0.2004<br\/>\nEpoch 89\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.5526 - acc: 0.2641 - val_loss: 11.8809 - val_acc: 0.2249<br\/>\nEpoch 90\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.5867 - acc: 0.2602 - val_loss: 12.0392 - val_acc: 0.2010<br\/>\nEpoch 91\/300<br\/>\n16\/16 [==============================] - 16s 1s\/step - loss: 1.5524 - acc: 0.2609 - val_loss: 12.0254 - val_acc: 0.2027<br\/><\/p>\n\n<p>My problem is, while using 'fit_generator' with above generator function as above, my model loss is not at all improving and validation accuracy is very poor. But when I use keras 'fit' function as below, the model loss decreases and validation accuracy is far better. <\/p>\n\n<p><b> Using Keras fit function without using a generator <\/b><\/p>\n\n<pre><code>model.fit(self.train_data, self.train_class_labels, batch_size=self.batch_size, epochs=self.epochs, validation_data=[self.val_data, self.val_class_labels], verbose=True, callbacks=[history, model_checkpoint])    \n<\/code><\/pre>\n\n<p><b> Output when trained using fit function <\/b><\/p>\n\n<p>Epoch 25\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0207 - acc: 0.9939 - val_loss: 4.1009 - val_acc: 0.4916<br\/>\nEpoch 26\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0197 - acc: 0.9948 - val_loss: 2.4758 - val_acc: 0.5568<br\/>\nEpoch 27\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0689 - acc: 0.9800 - val_loss: 1.2843 - val_acc: 0.7361<br\/>\nEpoch 28\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0207 - acc: 0.9947 - val_loss: 5.6979 - val_acc: 0.4560<br\/>\nEpoch 29\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0353 - acc: 0.9908 - val_loss: 1.0801 - val_acc: 0.7817<br\/>\nEpoch 30\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0362 - acc: 0.9896 - val_loss: 3.7851 - val_acc: 0.5173<br\/>\nEpoch 31\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0481 - acc: 0.9896 - val_loss: 1.1152 - val_acc: 0.7795<br\/>\nEpoch 32\/300<br\/>\n6400\/6400 [==============================] - 20s 3ms\/step - loss: 0.0106 - acc: 0.9969 - val_loss: 1.4803 - val_acc: 0.7372<br\/><\/p>\n","answers":[{"AnswerId":"54423329","CreationDate":"2019-01-29T14:33:46.077","ParentId":null,"OwnerUserId":"949251","Title":null,"Body":"<p>You have to make sure that your data generator shuffles the data between epochs. I would suggest you create a list of possible indices outside of your loop, randomize it with random.shuffle and then iterate over that inside your loop.<\/p>\n\n<p>Source: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2389\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/2389<\/a> and own experience.<\/p>\n"}]}
+{"QuestionId":52600215,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><object-detection>","CreationDate":"2018-10-01T23:36:20.910","AcceptedAnswerId":null,"OwnerUserId":3042444.0,"Title":"How do I know the total number of steps while training using Tensorflow Object Detection API?","Body":"<p>I've been running a training job for the last 3 hours on GPU powered cloud machine with the following command:<\/p>\n\n<pre><code>python legacy\/train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_mobilenet_v1_pets.config\n<\/code><\/pre>\n\n<p>and after running that, the log says this for example:<\/p>\n\n<pre><code> INFO:tensorflow:global step 14455: loss = 0.5896 (0.775 sec\/step)\nI1001 19:27:43.575182 140054916601600 tf_logging.py:116] global step 14455: loss = 0.5896 (0.775 sec\/step)\n<\/code><\/pre>\n\n<p>How do I know how many steps are there to be done or how many steps are there in total?<\/p>\n","answers":[{"AnswerId":"52603930","CreationDate":"2018-10-02T07:41:32.773","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>In the <code>ssd_mobilenet_v1_pets.config<\/code> it says in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/samples\/configs\/ssd_mobilenet_v1_pets.config\" rel=\"nofollow noreferrer\">line 163<\/a>:<\/p>\n\n<pre><code>num_steps: 200000\n<\/code><\/pre>\n\n<p>This is the number of total steps, the training script will perform if you did not make any changes.<\/p>\n"},{"AnswerId":"52600311","CreationDate":"2018-10-01T23:48:54.037","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>If you're using a pre-defined model topology, you look up the training period (in epochs or steps) in the documentation that comes with the model.  If you've made your own model, you determine the training period by watching the test results.  When the accuracy reaches an acceptable level and then starts to drop, you're likely over-training.  Back up to the high point of accuracy.  Repeat this experiment a few times to determine the \"sweet spot\" for your model.<\/p>\n"}]}
+{"QuestionId":52600230,"AnswerCount":2,"Tags":"<python><python-3.x><multithreading><tensorflow><python-multithreading>","CreationDate":"2018-10-01T23:38:01.637","AcceptedAnswerId":52616453.0,"OwnerUserId":2934305.0,"Title":"Reusing Tensorflow session in multiple threads causes crash","Body":"<p><strong>Background:<\/strong> <\/p>\n\n<p>I have some complex reinforcement learning algorithm that I want to run in multiple threads. <\/p>\n\n<p><strong>Problem<\/strong><\/p>\n\n<p>When trying to call <code>sess.run<\/code> in a thread I get the following error message:<\/p>\n\n<p><code>RuntimeError: The Session graph is empty.  Add operations to the graph before calling run().<\/code><\/p>\n\n<p><strong>Code reproducing the error:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\n\nimport threading\n\ndef thread_function(sess, i):\n    inn = [1.3, 4.5]\n    A = tf.placeholder(dtype=float, shape=(None), name=\"input\")\n    P = tf.Print(A, [A])\n    Q = tf.add(A, P)\n    sess.run(Q, feed_dict={A: inn})\n\ndef main(sess):\n\n    thread_list = []\n    for i in range(0, 4):\n        t = threading.Thread(target=thread_function, args=(sess, i))\n        thread_list.append(t)\n        t.start()\n\n    for t in thread_list:\n        t.join()\n\nif __name__ == '__main__':\n\n    sess = tf.Session()\n    main(sess)\n<\/code><\/pre>\n\n<p>If I run the same code outside a thread it works properly.<\/p>\n\n<p>Can someone give some insight on how to use Tensorflow sessions properly with python threads?<\/p>\n","answers":[{"AnswerId":"52616453","CreationDate":"2018-10-02T21:11:16.663","ParentId":null,"OwnerUserId":"8676953","Title":null,"Body":"<p>Not only can the Session be the current thread default, but also the graph.\nWhile you pass in the session and call <code>run<\/code> on it, the default graph will be a different one.<\/p>\n\n<p>You can ammend your <em>thread_function<\/em> like this to make it work:<\/p>\n\n<pre><code>def thread_function(sess, i):\n    with sess.graph.as_default():\n        inn = [1.3, 4.5]\n        A = tf.placeholder(dtype=float, shape=(None), name=\"input\")\n        P = tf.Print(A, [A])\n        Q = tf.add(A, P)\n        sess.run(Q, feed_dict={A: inn})\n<\/code><\/pre>\n\n<p>However, I wouldn't hope for any significant speedup. Python threading isn't what it means in some other languages, only certain operations, like io, would run in parallel. For CPU heavy operations it's not very useful. Multiprocessing can run code truely in parallel, but you wouldn't share the same session.<\/p>\n"},{"AnswerId":"57397201","CreationDate":"2019-08-07T14:54:20.110","ParentId":null,"OwnerUserId":"4960855","Title":null,"Body":"<p>Extending de1's answer with another resource on github:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/28287#issuecomment-495168033\" rel=\"nofollow noreferrer\">tensorflow\/tensorflow#28287 (comment)<\/a><\/p>\n\n<p>The following resolved tf's multithreading compatibility for me:<\/p>\n\n<pre><code># on thread 1\nsession = tf.Session(graph=tf.Graph())\nwith session.graph.as_default():\n    k.backend.set_session(session)\n    model = k.models.load_model(filepath)\n\n# on thread 2\nwith session.graph.as_default():\n    k.backend.set_session(session)\n    model.predict(x)\n<\/code><\/pre>\n\n<p>This keeps both the <code>Session<\/code> and the <code>Graph<\/code> for other threads.<br>\nThe model is loaded in their \"context\" (instead of the default ones) and kept for other threads to use.<br>\n(By default the model is loaded to the default <code>Session<\/code> and the default <code>Graph<\/code>)<br>\nAnother plus is that they're kept in the same object - easier to handle.<\/p>\n"}]}
+{"QuestionId":52600729,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2018-10-02T00:57:54.177","AcceptedAnswerId":null,"OwnerUserId":8517573.0,"Title":"How to use random number in user defined tensorflow op?","Body":"<p>How to use random number in user defined tensorflow op?<\/p>\n\n<p>I am writing a op in cpp which need random number in the Compute function.<\/p>\n\n<p>but It seems I should not use cpp random library directly, since that cannot control by <code>tf.set_random_seed<\/code>.<\/p>\n\n<p>My current code is something like the following, what should I do in function <code>some_interesting_random_function<\/code> ?<\/p>\n\n<pre><code>#include \"tensorflow\/core\/framework\/op.h\"\n#include \"tensorflow\/core\/framework\/op_kernel.h\"\n#include \"tensorflow\/core\/framework\/common_shape_fns.h\"\n#include &lt;iostream&gt;\n#include &lt;typeinfo&gt;\n#include &lt;random&gt;\n\nusing namespace tensorflow;\n\nREGISTER_OP(\"MyRandom\")\n.Output(\"random: int32\")\n.SetShapeFn([](::tensorflow::shape_inference::InferenceContext* c) {\n        c-&gt;set_output(0, c-&gt;Scalar());\n        return Status::OK();\n});\n\nint some_interesting_random_function(){\n    return 10;\n}\n\nclass MyRandomOp : public OpKernel {\n    public:\n        explicit MyRandomOp(OpKernelConstruction* context) : OpKernel(context) {}\n        void Compute(OpKernelContext* context) override {\n            Tensor* res;\n            TensorShape shape;\n            int dims[] = {};\n            TensorShapeUtils::MakeShape(dims, 0, &amp;shape);\n            OP_REQUIRES_OK(context, context-&gt;allocate_output(0, shape,\n                        &amp;res));\n            auto out1 = res-&gt;flat&lt;int32&gt;();\n            out1(0) = some_interesting_random_function();\n        }\n};\n\nREGISTER_KERNEL_BUILDER(Name(\"MyRandom\").Device(DEVICE_CPU), MyRandomOp);\n<\/code><\/pre>\n","answers":[{"AnswerId":"52607937","CreationDate":"2018-10-02T12:02:30.330","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The core of all random number generation in TensorFlow is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/lib\/random\/philox_random.h\" rel=\"nofollow noreferrer\"><code>PhiloxRandom<\/code><\/a>, generally accessed through its wrapper <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/util\/guarded_philox_random.h\" rel=\"nofollow noreferrer\"><code>GuardedPhiloxRandom<\/code><\/a>. As explained in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/set_random_seed\" rel=\"nofollow noreferrer\"><code>tf.set_random_seed<\/code><\/a>, there are graph-level and op-level seeds, both of which may or may not be set. If you want to have this in your op too, you need to do a couple of things. First, your op should be declared with two optional attributes, <code>seed<\/code> and <code>seed2<\/code>; see the existing ops in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/ops\/random_ops.cc\" rel=\"nofollow noreferrer\"><code>random_ops.cc<\/code><\/a>. Then, in Python, you have some user API wrapping your op that makes these two values using <code>tensorflow.python.framework.random_seed<\/code>, which you have to import as <code>tensorflow.python.framework import random_seed<\/code>, and do <code>seed1, seed2 = random_seed.get_seed(seed)<\/code>; this will correctly create the two seed values using the graph's seed and an optional <code>seed<\/code> parameter to the function (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/python\/ops\/random_ops.py\" rel=\"nofollow noreferrer\"><code>random_ops.py<\/code><\/a>). These <code>seed1<\/code> and <code>seed2<\/code> values are then passed as <code>seed<\/code> and <code>seed2<\/code> attributes to your op, obviously. If you do all that, then <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/util\/guarded_philox_random.h\" rel=\"nofollow noreferrer\"><code>GuardedPhiloxRandom<\/code><\/a> will take care of properly initializing the random number generator using the right seeds.<\/p>\n\n<p>Now, to the kernel implementation. In addition to the things I mentioned above, you will need to combine two things: the struct template <code>FillPhiloxRandom<\/code>, declared in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/random_op.h\" rel=\"nofollow noreferrer\"><code>core\/kernels\/random_op.h<\/code><\/a>, which will help you fill a tensor with random data; and a <code>Distribution<\/code>, which is just an object that can be called with a random number generator to produce a value (see existing implementations in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/lib\/random\/random_distributions.h\" rel=\"nofollow noreferrer\"><code>core\/lib\/random\/random_distributions.h<\/code><\/a>). Now it is mostly a matter of looking at how it is done in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/kernels\/random_op.cc\" rel=\"nofollow noreferrer\"><code>core\/kernels\/random_op.cc<\/code><\/a>, and copy the bits you need. Most kernels in there are based on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/kernels\/random_op.cc#L182-L214\" rel=\"nofollow noreferrer\"><code>PhiloxRandomOp<\/code><\/a> (which is not publicly declared, but you can copy or adapt). This essentially holds a random number generator, allocates space in the output tensor (it assumes the first input is the desired shape) and calls <code>FillPhiloxRandom<\/code> to do the work. If this is the kind of op you are trying to create (generate some data according to some distribution), then you are all set! Your code could look something like this:<\/p>\n\n<pre><code>\/\/ Required for thread pool device\n#define EIGEN_USE_THREADS\n\n#include \"tensorflow\/core\/framework\/op_kernel.h\"\n#include \"tensorflow\/core\/framework\/register_types.h\"\n#include \"tensorflow\/core\/framework\/tensor.h\"\n#include \"tensorflow\/core\/framework\/tensor_shape.h\"\n#include \"tensorflow\/core\/kernels\/random_op.h\"\n#include \"tensorflow\/core\/util\/guarded_philox_random.h\"\n\n\/\/ Helper function to convert an 32-bit integer to a float between [0..1).\n\/\/ Copied from core\/lib\/random\/random_distributions.h\nPHILOX_DEVICE_INLINE float Uint32ToFloat(uint32 x) {\n  \/\/ IEEE754 floats are formatted as follows (MSB first):\n  \/\/    sign(1) exponent(8) mantissa(23)\n  \/\/ Conceptually construct the following:\n  \/\/    sign == 0\n  \/\/    exponent == 127  -- an excess 127 representation of a zero exponent\n  \/\/    mantissa == 23 random bits\n  const uint32 man = x &amp; 0x7fffffu;  \/\/ 23 bit mantissa\n  const uint32 exp = static_cast&lt;uint32&gt;(127);\n  const uint32 val = (exp &lt;&lt; 23) | man;\n\n  \/\/ Assumes that endian-ness is same for float and uint32.\n  float result;\n  memcpy(&amp;result, &amp;val, sizeof(val));\n  return result - 1.0f;\n}\n\n\/\/ Template class for your custom distribution\ntemplate &lt;class Generator, typename RealType&gt;\nclass MyDistribution;\n\n\/\/ Implementation for tf.float32\ntemplate &lt;class Generator&gt;\nclass MyDistribution&lt;Generator, float&gt; {\n public:\n  \/\/ The number of elements that will be returned (see below).\n  static const int kResultElementCount = Generator::kResultElementCount;\n  \/\/ Cost of generation of a single element (in cycles) (see below).\n  static const int kElementCost = 3;\n  \/\/ Indicate that this distribution may take variable number of samples\n  \/\/ during the runtime (see below).\n  static const bool kVariableSamplesPerOutput = false;\n  typedef Array&lt;float, kResultElementCount&gt; ResultType;\n  typedef float ResultElementType;\n\n  PHILOX_DEVICE_INLINE\n  ResultType operator()(Generator* gen) {\n    typename Generator::ResultType sample = (*gen)();\n    ResultType result;\n    for (int i = 0; i &lt; kResultElementCount; ++i) {\n      float r = Uint32ToFloat(sample[i]);\n      \/\/ Example distribution logic: produce 1 or 0 with 50% probability\n      result[i] = 1.0f * (r &lt; 0.5f);\n    }\n    return result;\n  }\n};\n\n\/\/ Could add implementations for other data types...\n\n\/\/ Base kernel\n\/\/ Copied from core\/kernels\/random_op.cc\nstatic Status AllocateOutputWithShape(OpKernelContext* ctx, const Tensor&amp; shape,\n                                      int index, Tensor** output) {\n  TensorShape tensor_shape;\n  TF_RETURN_IF_ERROR(ctx-&gt;op_kernel().MakeShape(shape, &amp;tensor_shape));\n  return ctx-&gt;allocate_output(index, tensor_shape, output);\n}\n\ntemplate &lt;typename Device, class Distribution&gt;\nclass PhiloxRandomOp : public OpKernel {\n public:\n  typedef typename Distribution::ResultElementType T;\n  explicit PhiloxRandomOp(OpKernelConstruction* ctx) : OpKernel(ctx) {\n    OP_REQUIRES_OK(ctx, generator_.Init(ctx));\n  }\n  void Compute(OpKernelContext* ctx) override {\n    const Tensor&amp; shape = ctx-&gt;input(0);\n    Tensor* output;\n    OP_REQUIRES_OK(ctx, AllocateOutputWithShape(ctx, shape, 0, &amp;output));\n    auto output_flat = output-&gt;flat&lt;T&gt;();\n    tensorflow::functor::FillPhiloxRandom&lt;Device, Distribution&gt;()(\n        ctx, ctx-&gt;eigen_device&lt;Device&gt;(),\n        \/\/ Multiplier 256 is the same as in FillPhiloxRandomTask; do not change\n        \/\/ it just here.\n        generator_.ReserveRandomOutputs(output_flat.size(), 256),\n        output_flat.data(), output_flat.size(), Distribution());\n  }\n private:\n  GuardedPhiloxRandom generator_;\n};\n\n\/\/ Register kernel\ntypedef Eigen::ThreadPoolDevice CPUDevice;\ntemplate struct functor::FillPhiloxRandom&lt;\n    CPUDevice, MyDistribution&lt;tensorflow::random::PhiloxRandom, float&gt;&gt;;\nREGISTER_KERNEL_BUILDER(\n    Name(\"MyDistribution\")\n        .Device(DEVICE_CPU)\n        .HostMemory(\"shape\")\n        .TypeConstraint&lt;float&gt;(\"dtype\"),\n    PhiloxRandomOp&lt;CPUDevice, MyDistribution&lt;tensorflow::random::PhiloxRandom, float&gt;&gt;);\n\n\/\/ Register kernels for more types, can use macros as in core\/kernels\/random_op.cc...\n<\/code><\/pre>\n\n<p>There are a few extra bits and pieces here. First you need to understand that <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/lib\/random\/philox_random.h\" rel=\"nofollow noreferrer\"><code>PhiloxRandom<\/code><\/a> generally produces four unsigned 32-bit integers on each step, and you have to make your random values from these. <code>Uint32ToFloat<\/code> is a helper to get a <code>float<\/code> between zero and one from one of this numbers. There are a few constants in there too. <code>kResultElementCount<\/code> is the number of values your distribution produces on each step. If you produce one value per random number form the generator, you can set it too <code>Generator::kResultElementCount<\/code>, like here (which is 4). However, for example if you want to produce <code>double<\/code> values (that is, <code>tf.float64<\/code>), you may want to use two 32-bit integers per value, so maybe you would produce <code>Generator::kResultElementCount \/ 2<\/code> in that case. <code>kElementCost<\/code> is supposed to indicate how many cycles it takes your distribution to produce an element. I do not know how this is measured by the TensorFlow team, but it is just a hint to distribute the generation work among tasks (used by <code>FillPhiloxRandom<\/code>), so you can just guess something, or copy it from a similarly expensive distribution. <code>kVariableSamplesPerOutput<\/code> determines whether each call to your distribution may produce a different number of outputs; again, when this is <code>false<\/code> (which should be the common case), <code>FillPhiloxRandom<\/code> will make the value generation more efficient. <code>PHILOX_DEVICE_INLINE<\/code> (defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/core\/lib\/random\/philox_random.h\" rel=\"nofollow noreferrer\"><code>core\/lib\/random\/philox_random.h<\/code><\/a>) is a compiler hint to inline the function. You can add then additional implementations and kernel registrations for other data types and, if you are supporting it, for <code>DEVICE_GPU<\/code> <code>GPUDevice<\/code> (with <code>typedef Eigen::GpuDevice GPUDevice<\/code>) or even <code>DEVICE_SYCL<\/code> (with <code>typedef Eigen::SyclDevice SYCLDevice<\/code>), if you want. And about that, <code>EIGEN_USE_THREADS<\/code> is just to enable the thread pool execution device in Eigen, to make CPU implementation multi-threaded.<\/p>\n\n<p>If your use case is different, though (for example, you want to generate some random numbers and do some other computation in addition to that), <code>FillPhiloxRandom<\/code> may not be useful to you (or it may be, but then you also need to do something else). Having a look at <code>core\/kernels\/random_op.cc<\/code> and the headers of the different classes should help you figure out how to use them for your problem.<\/p>\n"}]}
+{"QuestionId":52600822,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><deep-learning>","CreationDate":"2018-10-02T01:14:44.330","AcceptedAnswerId":null,"OwnerUserId":5890355.0,"Title":"How to improve the under-fitting issue of my RNN Autoencoder on random squences?","Body":"<p>I am trying to use RNN (seq-to-seq) as Autoencoder to reconstruct random sequences (classification). However, the model is suffering from under-fitting issue in which it can't improve the training loss (settles around 2.6) and predicts mostly 0s (length variation is high so I padded with a lot 0s while doing min-batch training).<\/p>\n\n<p>The random sequences are initiated with the following code:<\/p>\n\n<pre><code>vocab = 200\nraw = np.array(range(vocab))+3 # 200 Unique Events=3-203 + [pad=0],[EOS=1],[SOS=2]\nEventOptions = [str(i) for i in raw]\n\nclass data_simulator(object):\n\n   def __init__(self,seq_list,seq_min,seq_max):\n    self.seq_min = seq_min\n    self.seq_max = seq_max\n    self.seq_list = seq_list\n    print(\"Data Generator Registered...\")\n\n   def Gen(self,volume):\n    return pd.DataFrame.from_dict({'Instance':range(volume),\n                               'Sequence':[' | '.join(np.random.choice(self.seq_list, np.random.randint(self.seq_min, self.seq_max))) for i in xrange(volume)]\n                               })\n\nGenerator = data_simulator(EventOptions,20,200) # lengths varies from 20 to 200\nRawData_Normal = Generator.Gen(5000) # 5000 sequences\n<\/code><\/pre>\n\n<p>I started with an initial learning to identify the learning rate and optimizer for this data. So, I initiate an basic architecture of RNN autoencoder as:<\/p>\n\n<ul>\n<li>Single layer GRU cell with tanh activation<\/li>\n<li>Basic Encoder-Decoder seq-to-seq<\/li>\n<li>100 neurons in hidden layer<\/li>\n<li>20 embedding features<\/li>\n<li>50 random sequences + batch_size = 10 + n_epochs = 100<\/li>\n<\/ul>\n\n<p>to test the learning rate range with various optimizers:<\/p>\n\n<pre><code>learning_vector = [0.000001,0.000003,0.000006,0.000008,\n                   0.00001,0.00003,0.00006,0.00008,\n                   0.0001,0.0003,0.0006,0.0008,\n                   0.001,0.003,0.006,0.008,\n                   0.01,0.03,0.06,0.08,\n                   0.1, 0.3, 0.6, 0.8]\n<\/code><\/pre>\n\n<p>The result is below with increasing learning rate [0.000001-0.8] and all optimizers settled around 2.6:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iJBmt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iJBmt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Then, I picked \"Adam\" with the best learning rate on previous plot - 0.0005 to retrain the model with more capacities like below:<\/p>\n\n<ul>\n<li>3 layers GRU cell with He initialization and tanh activation<\/li>\n<li>Encoder-Decoder seq-to-seq with attention<\/li>\n<li>250 neurons in each hidden layer<\/li>\n<li>20 embedding features<\/li>\n<li>5000 random sequences + batch_size = 200 + n_epochs = 3000<\/li>\n<\/ul>\n\n<p>Then, I have got very similar \"underfitting\" result like below which settles around 2.6 for both train and validation loss (make sense since both are random sequences):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yRmwY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yRmwY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, I was expecting the increased model capacity should enables the model to \"memorize\" the random noises so that we can further improve training performance till overfitting the random sequences. Yet, this wasn't happen and the training loss was settled around 2.6 no matter how many more epochs I train the model. Is it because the random sequences variations are too informative to be learnt given my current model capacity and train time? Or the high variations in the sequence length resulted a lot of 0 pads which can skewed the prediction? How can I resolve the underfitting issues in this case?<\/p>\n","answers":[]}
+{"QuestionId":52600836,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><tfrecord>","CreationDate":"2018-10-02T01:17:01.813","AcceptedAnswerId":null,"OwnerUserId":10443639.0,"Title":"something went wrong during video to numpy conversion","Body":"<p>I use video2record to deal with some short video, I want to turn them into tfrecord then I can input them into tensorflow. \nThen I use videorecord on github: <a href=\"https:\/\/github.com\/ferreirafabio\/video2tfrecord\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ferreirafabio\/video2tfrecord<\/a>\nBut when I run it, the error said numpy wrong\nhere are my code(python):<\/p>\n\n<pre><code>from video2tfrecord import convert_videos_to_tfrecord\n\nconvert_videos_to_tfrecord(source_path=\"video\", destination_path=\"video\", n_videos_in_record= 1 , n_frames_per_video = 5 , file_suffix=\"*.mp4\") \n<\/code><\/pre>\n\n<p>here are the error:<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>AssertionError                            Traceback (most recent call last)  in ()\n          1 from video2tfrecord import convert_videos_to_tfrecord\n          2 \n    ----> 3 convert_videos_to_tfrecord(source_path=\"video\", destination_path=\"video\", n_videos_in_record= 1 , n_frames_per_video =\n    5 , file_suffix=\"*.mp4\")<\/p>\n  <\/blockquote>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/video2tfrecord\/video2tfrecord.py\n  in convert_videos_to_tfrecord(source_path, destination_path,\n  n_videos_in_record, n_frames_per_video, file_suffix,\n  dense_optical_flow, width, height, color_depth, video_filenames)\n      174       total_batch_number = int(math.ceil(len(filenames) \/ n_videos_in_record))\n      175     print('Batch ' + str(i + 1) + '\/' + str(total_batch_number) + \" completed\")\n  --> 176     assert data.size != 0, 'something went wrong during video to numpy conversion'\n      177     save_numpy_to_tfrecords(data, destination_path, 'batch_',\n      178                             n_videos_in_record, i + 1, total_batch_number,<\/p>\n  \n  <p>AssertionError: something went wrong during video to numpy conversion<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52601631,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-10-02T03:25:58.317","AcceptedAnswerId":null,"OwnerUserId":10443985.0,"Title":"Back-propagation with complex parameters in Tensorflow","Body":"<p>Currently I'm trying to train a complex data generated from telecom engineering models. The weights and biases are also complex. I have used the relu activation for the hidden layers as follows at the l-th layer: <\/p>\n\n<pre><code>A_l = tf.complex(tf.nn.relu(tf.real(Z_l)), tf.nn.relu(tf.imag(Z_l))) \n<\/code><\/pre>\n\n<p>But how to do it for the cost and optimizer, please? I am really confused because I'm a beginner in machine learning. I have gone through some papers about non-analytic functions, but none of them helped to use the Tensorflow API. For example: how do I rewrite the functions below?  <\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = Z_out, labels = y)) \noptimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)\n<\/code><\/pre>\n\n<p>I have seen a recommendation to split the cost as real and imaginary parts as: <\/p>\n\n<pre><code>cost_R = .., cost_I = ... \n<\/code><\/pre>\n\n<p>but I didn't try it because I think the optimizer will be split and the optimization not be working. I used the relu activation for the hidden layers as follows at the l-th layer: <\/p>\n\n<pre><code>A_l = tf.complex(tf.nn.relu(tf.real(Z_l)), tf.nn.relu(tf.imag(Z_l))) \n<\/code><\/pre>\n\n<p>But how to the cost and optimizer? \nAny help is much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52602034,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2018-10-02T04:27:37.663","AcceptedAnswerId":null,"OwnerUserId":1155409.0,"Title":"Can I get other metric values using model history file already ran though model.fit()?","Body":"<p>I thought that in the model history file, there were all information such as y_pred, y_true values as model trains, so that I thoguth even if I add other metrics which I customaized, I could reuse the history file not running model.fit() again. Is this possible?<\/p>\n","answers":[]}
+{"QuestionId":52602071,"AnswerCount":1,"Tags":"<dataframe><nlp><pytorch><torchtext>","CreationDate":"2018-10-02T04:33:11.700","AcceptedAnswerId":null,"OwnerUserId":4096212.0,"Title":"Dataframe as datasource in torchtext","Body":"<p>I have a dataframe, which has two columns (review and sentiment). I am using pytorch and torchtext library for preprocessing data.\nIs it possible to use dataframe as source to read data from, in torchtext?\nI am looking for something similar to, but not  <\/p>\n\n<pre><code>data.TabularDataset.splits(path='.\/data')\n<\/code><\/pre>\n\n<p>I have performed some operation (clean, change to required format) on data and final data is in a dataframe.<\/p>\n\n<p>If not torchtext, what other package would you suggest that would help in preprocessing text data present in a datarame. I could not find anything online. Any help would be great.<\/p>\n","answers":[{"AnswerId":"52655787","CreationDate":"2018-10-04T22:01:52.900","ParentId":null,"OwnerUserId":"5173924","Title":null,"Body":"<p>Adapting the <code>Dataset<\/code> and <code>Example<\/code> classes from <code>torchtext.data<\/code><\/p>\n\n<pre><code>from torchtext.data import Field, Dataset, Example\nimport pandas as pd\n\n class DataFrameDataset(Dataset):\n     \"\"\"Class for using pandas DataFrames as a datasource\"\"\"\n     def __init__(self, examples, fields, filter_pred=None):\n         \"\"\"\n         Create a dataset from a pandas dataframe of examples and Fields\n         Arguments:\n             examples pd.DataFrame: DataFrame of examples\n             fields {str: Field}: The Fields to use in this tuple. The\n                 string is a field name, and the Field is the associated field.\n             filter_pred (callable or None): use only exanples for which\n                 filter_pred(example) is true, or use all examples if None.\n                 Default is None\n         \"\"\"\n         self.examples = examples.apply(SeriesExample.fromSeries, args=(fields,), axis=1).tolist()\n         if filter_pred is not None:\n             self.examples = filter(filter_pred, self.examples)\n         self.fields = dict(fields)\n         # Unpack field tuples\n         for n, f in list(self.fields.items()):\n             if isinstance(n, tuple):\n                 self.fields.update(zip(n, f))\n                 del self.fields[n]\n\n class SeriesExample(Example):\n     \"\"\"Class to convert a pandas Series to an Example\"\"\"\n\n     @classmethod\n     def fromSeries(cls, data, fields):\n         return cls.fromdict(data.to_dict(), fields)\n\n     @classmethod\n     def fromdict(cls, data, fields):\n         ex = cls()\n\n         for key, field in fields.items():\n             if key not in data:\n                 raise ValueError(\"Specified key {} was not found in \"\n                 \"the input data\".format(key))\n             if field is not None:\n                 setattr(ex, key, field.preprocess(data[key]))\n             else:\n                 setattr(ex, key, data[key])\n    return ex\n<\/code><\/pre>\n\n<p>Then, if you have your two datasets handy <code>train_df<\/code>, <code>valid_df<\/code>, just load them in the Dataset object with:<\/p>\n\n<pre><code>train_ds = DataFrameDataset(train_df, fields)\nvalid_ds = DataFrameDataset(valid_df, fields)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52603085,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><pytorch><reinforcement-learning>","CreationDate":"2018-10-02T06:32:20.793","AcceptedAnswerId":null,"OwnerUserId":8759752.0,"Title":"PPO \/ TRPO Implementation","Body":"<p>So, I recently watched this video on PPO and want to upgrade my actor-critic algorithm written in <code>PyTorch<\/code> with PPO, but I'am not sure how the new parameters \/ thetas are calculated.<\/p>\n\n<p>Given algorithm in the paper:<br>\n<a href=\"https:\/\/i.stack.imgur.com\/bVkQH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bVkQH.png\" alt=\"enter image description here\"><\/a><br>\nsays to run <code>pi_theta_old<\/code>, compute advantage estimates and optimize the objective, but how can we calculate <code>pi_theta<\/code> for the objective since we have not updated the <code>policy_old<\/code> yet?<\/p>\n\n<p>EDIT : To all those looking for an clear answer, I have posted the same question on Stack Exchange Artificial Intelligence (<a href=\"https:\/\/ai.stackexchange.com\/questions\/8212\/ppo-trpo-implementation\">https:\/\/ai.stackexchange.com\/questions\/8212\/ppo-trpo-implementation<\/a>).<\/p>\n","answers":[]}
+{"QuestionId":52603302,"AnswerCount":1,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2018-10-02T06:53:56.883","AcceptedAnswerId":null,"OwnerUserId":9766649.0,"Title":"Tensorflow Lite GPU acceleration - does it work for all OS versions or only for 8.1?","Body":"<p>Does Tensorflow Lite GPU acceleration work for all Android APIs or only for 8.1 (and mb only for specific devices like Nexus with specific chip)?<\/p>\n\n<p>Does framework require any settings to activate GPU acceleration (in Java\/Kotlin) or does it automatically choose between CPU\/GPU and just CPU<\/p>\n\n<p>There some official tensorflow examples and I see that minimal version for lite example is 15 API<\/p>\n","answers":[{"AnswerId":"53082218","CreationDate":"2018-10-31T11:23:37.790","ParentId":null,"OwnerUserId":"8049751","Title":null,"Body":"<p>As I found at the moment gpu is used only on android 8.1, when the use_nnapi flag is set in the tflite interpreter<\/p>\n"}]}
+{"QuestionId":52603944,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-02T07:42:10.007","AcceptedAnswerId":null,"OwnerUserId":10444749.0,"Title":"loading tensorflow on Windows with new RTX-2080 NVIDIA card","Body":"<p>Trying to load tensorflow with Windows 10 on a new RTX-2080 NVIDIA card (using Python). Are there existing binaries? Do I have to built from source? What versions of CUDA and Cudnn do I need?<\/p>\n\n<p>Thanks,<\/p>\n\n<p>Rich<\/p>\n","answers":[]}
+{"QuestionId":52604040,"AnswerCount":1,"Tags":"<linux><python-3.x><tensorflow><caffe>","CreationDate":"2018-10-02T07:49:23.527","AcceptedAnswerId":null,"OwnerUserId":10444822.0,"Title":"Error while Installing tensorflow-gpu on MX150","Body":"<p>I am trying to install tensorflow gpu on my laptop running MX150. I updated my NVIDIA driver, installed CUDA , installed CUDnn. But still not able to import Tensorflow. getting following error<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\nTraceback (most recent call last):\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 72, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/rohan\/anaconda3\/envs\/tensorflow_env\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n","answers":[{"AnswerId":"52605017","CreationDate":"2018-10-02T08:51:31.063","ParentId":null,"OwnerUserId":"5106245","Title":null,"Body":"<p>It looks like there is a mismatch between your python version and your CUDA version.\nplease check your versions by:<\/p>\n\n<pre><code>python --version\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>nvcc --version\n<\/code><\/pre>\n\n<p>What is yuor OS? What version of Tensorflow your are trying to run?\nI suggest to uninstall tensorflow and reinstall it after you have the currect version of Python and CUDA. You can follow <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">Tesnoeflow with gpu documentation<\/a>.<\/p>\n\n<p>Also check this issue which is similar to yours\n<a href=\"https:\/\/stackoverflow.com\/questions\/44865253\/libcublas-so-8-0-error-with-tensorflow\">libcublas.so.8.0 error with tensorflow<\/a><\/p>\n"}]}
+{"QuestionId":52604876,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-02T08:42:56.007","AcceptedAnswerId":null,"OwnerUserId":10164729.0,"Title":"is the speed is only problem for training tensorflow model with cpu","Body":"<p>I am training to train tensorflow with cpu,it is slow, i have no problem with this but my question is <\/p>\n\n<blockquote>\n  <p>\"will my mdel be trained as like gpu, is the only difference between\n  cpu and gpu in deep learning is speed?<\/p>\n<\/blockquote>\n\n<p>my train code <\/p>\n\n<pre><code>python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_mobilenet_v1_pets.config\n<\/code><\/pre>\n\n<p>I am using windows 10 <\/p>\n","answers":[]}
+{"QuestionId":52604879,"AnswerCount":1,"Tags":"<python><tensorflow><gradient>","CreationDate":"2018-10-02T08:43:12.740","AcceptedAnswerId":52605552.0,"OwnerUserId":6127030.0,"Title":"How excute custom gradient with tf.multiply?","Body":"<p>I define custom gradient mapper with tensorflow package. <\/p>\n\n<p>When i use tf.multiply with custom gradient, it did not work . <\/p>\n\n<p>whole code is here<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.RegisterGradient(\"MyopGrad\")\ndef frop_grad(op, grad):\n    x = op.inputs[0] \n    return 1000.0 * x \n\ninput = tf.Variable([4.0], dtype=tf.float32)\nx = tf.constant(5.0)\ng = tf.get_default_graph()\n\nwith g.gradient_override_map({\"Multiply\": \"MyopGrad\"}): \n  output1 = tf.multiply(input, x , name = 'multiply')\ngrad1 = tf.gradients(output1, input)\n\n# output without gradient clipping in the backwards pass for comparison:\noutput1_ori = tf.multiply(input , x)\ngrad1_ori = tf.gradients(output1_ori, input)\n\nwith tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  print(\"with custom:\", sess.run(grad1)[0])\n  print(\"without custom:\", sess.run(grad1_ori)[0])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52605552","CreationDate":"2018-10-02T09:23:33.453","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The TensorFlow op name for <code>tf.multiply<\/code> is just <code>Mul<\/code>, not <code>Multiply<\/code>. Also, <code>tf.multiply<\/code> has two inputs, so its gradients should have two outputs. So your code could look something like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.RegisterGradient(\"MyopGrad\")\ndef frop_grad(op, grad):\n    x = op.inputs[0]\n    y = op.inputs[1]\n    return 1000.0 * x, 1000.0 * y\n\ninput = tf.Variable([4.0], dtype=tf.float32)\nx = tf.constant(5.0)\ng = tf.get_default_graph()\n\nwith g.gradient_override_map({\"Mul\": \"MyopGrad\"}): \n  output1 = tf.multiply(input, x , name = 'multiply')\ngrad1 = tf.gradients(output1, input)\n\n# output without gradient clipping in the backwards pass for comparison:\noutput1_ori = tf.multiply(input , x)\ngrad1_ori = tf.gradients(output1_ori, input)\n\nwith tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  print(\"with custom:\", sess.run(grad1)[0])\n  print(\"without custom:\", sess.run(grad1_ori)[0])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>with custom: [4000.]\nwithout custom: [5.]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52605067,"AnswerCount":0,"Tags":"<python><tensorflow><seq2seq><encoder-decoder>","CreationDate":"2018-10-02T08:54:24.807","AcceptedAnswerId":null,"OwnerUserId":10223227.0,"Title":"how to make decoder in rnn, tensorflow, feeded previous output","Body":"<p>I'm wondering how to make decoder in tensorflow rnn, feed it's i th output to (i+1)th input<\/p>\n\n<p>my inputs have 20 sequence and 3680 dementions\nand my outputs have 39 sequence and 3680 dementions\nall data is 0~1 number<\/p>\n\n<p>here is my model <\/p>\n\n<pre><code>with tf.variable_scope('encoder'):    \n    enc_input = tf.placeholder(tf.float32,[None, input_sequence_length, input_dim])\n\n    enc_cell = tf.contrib.rnn.BasicLSTMCell(num_units = input_sequence_length)\n\n    _ , encoder_states = tf.nn.dynamic_rnn(enc_cell, enc_input ,  dtype=tf.float32)\n\nwith tf.variable_scope('decoder'):\n    dec_input = tf.placeholder(tf.float32,[None, output_sequence_length, output_dim])\n    dec_output = tf.placeholder(tf.float32,[None, output_sequence_length, output_dim])\n\n    dec_cell = tf.contrib.rnn.BasicLSTMCell(num_units = output_sequence_length)\n\n    outputs , _ = tf.nn.dynamic_rnn(dec_cell, dec_input,dtype = tf.float32,\n                                    initial_state = encoder_states)\n<\/code><\/pre>\n\n<p>how can i make decoder model that feed previous outputs to next input?<\/p>\n\n<p>P.S<\/p>\n\n<p>I make my self-answer code like this <\/p>\n\n<pre><code>with tf.variable_scope('decoder'):\n    dec_input = tf.placeholder(tf.float32,[None, 1, output_dim])\n    dec_output = tf.placeholder(tf.float32,[None, output_sequence_length, output_dim])\n\n    outputs = []\n    state = encoder_states\n\n    dec_cell = tf.contrib.rnn.BasicLSTMCell(num_units = dec_hidden_size)\n\n    for i in range(output_sequence_length):\n        if i==0:\n            output , state = tf.nn.dynamic_rnn(dec_cell, dec_input, initial_state = state, dtype = tf.float32)\n            outputs.append(output)\n        else:\n            output , state = tf.nn.dynamic_rnn(dec_cell, \n                                                 output, \n                                                 initial_state = state, \n                                                 dtype = tf.float32)\n            outputs.append(output)\n\noutputs = tf.reshape(outputs,[-1,output_dim])\noutputs = tf.reshape(outputs,[-1,output_sequence_length,output_dim])\n<\/code><\/pre>\n\n<p>I think this code's output is different from upper code's output \nbut I'm not sure it worked properly.<\/p>\n\n<p>so still I wonder how to make decoder that has loop function((i)output->(i+1)input) with tensorflow method, \nbecause it takes more memory allocation than the upper code. (in my thought it has same cell count)<\/p>\n","answers":[]}
+{"QuestionId":52605783,"AnswerCount":1,"Tags":"<python><tensorflow><bessel-functions>","CreationDate":"2018-10-02T09:39:12.940","AcceptedAnswerId":null,"OwnerUserId":8632298.0,"Title":"How do I implement BesselJ functions in tensorflow","Body":"<p>tensorflow inner functions supports the Bessel function BesselI <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/bessel-i0e\" rel=\"nofollow noreferrer\">tensorflow besseli0<\/a>.<\/p>\n\n<p>However, it does not have BesselJ functions (especially BesselJ[2,x]).<\/p>\n\n<p>If I were to approximate this function, I would have to use <code>If<\/code> in tensorflow, which is not efficient. <\/p>\n\n<p>Does tensorflow have any replacement or any other methods that can be recommended?<\/p>\n","answers":[{"AnswerId":"52618260","CreationDate":"2018-10-03T00:53:48.157","ParentId":null,"OwnerUserId":"8632298","Title":null,"Body":"<p>After many searches, I found a just-satisfactory solution:<\/p>\n\n<pre><code>def BesselJ0(x):\n    PI=3.1415926\n    temp = tf.cast(tf.less(x,tf.ones_like(x)*4.7),dtype=tf.float32)\n    temp1 = tf.abs(temp-1)\n    left_part = tf.multiply(1-x**2\/4+x**4\/64-x**6\/2304+x**8\/147456-x**10\/14745600,temp)\n    right_part = tf.multiply(tf.sqrt(2\/PI\/x)*tf.cos(x-PI\/4),temp1)\n    return left_part+right_part\ndef BesselJ1(x):\n    PI=3.1415926\n    temp = tf.cast(tf.less(x,tf.ones_like(x)*3.8),dtype=tf.float32)\n    temp1 = tf.abs(temp-1)\n    left_part = tf.multiply(x\/2-x**3\/16+x**5\/384-x**7\/18432+x**9\/1474560-x**11\/176947200+x**13\/29727129600,temp)\n    right_part = tf.multiply(tf.sqrt(2\/PI\/x)*tf.cos(x-PI\/2-PI\/4),temp1)\n    return left_part+right_part\n<\/code><\/pre>\n\n<p>This is not very fast, though.<\/p>\n"}]}
+{"QuestionId":52605876,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-02T09:45:22.647","AcceptedAnswerId":null,"OwnerUserId":7143865.0,"Title":"Create placeholders for tensors of dynamic size at every iteration","Body":"<p>I am creating the placeholder for a tensor in the following way:<\/p>\n\n<pre><code>self.sample_set = tf.placeholder(tf.int32, [None,None,self.max_input_right]) --- (1)\n<\/code><\/pre>\n\n<p>Now I want to slice this tensor of shape [None,self.max_input_right]. Basically the above tensor consist of list of values and now I want to extract those values from the list. But with every iteration of feed_dict, the size of this list changes.\nSo after step (1), I was looking if I can perform something like this:<\/p>\n\n<pre><code>for i in range(self.k):\n        self.ans_set = tf.slice(self.sample_set, [0,i])\n        self.ans_sets.append(self.ans_set) \n<\/code><\/pre>\n\n<p>where the value of k would change with every feed_dict. But I am unable to understand how should I pass the changing value of k while building the graph.<\/p>\n\n<p>This question might look similar to previously posted question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/37621841\/tensorflow-dynamically-create-placeholders\">tensorflow dynamically create placeholders<\/a><\/p>\n\n<p>But this answer didn't help me to find the answer. <\/p>\n","answers":[{"AnswerId":"52607045","CreationDate":"2018-10-02T11:02:33.257","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>For the use case in your example, you can simply use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/unstack\" rel=\"nofollow noreferrer\"><code>tf.unstack<\/code><\/a>:<\/p>\n\n<pre><code>self.sample_set = tf.placeholder(tf.int32, [None,None,self.max_input_right])\nself.ans_sets = tf.unstack(self.sample_set)\n<\/code><\/pre>\n\n<p>This will give you the list of 3D tensors. By default <code>unstack<\/code> works on axis 0 but you can also change this.<\/p>\n"}]}
+{"QuestionId":52606075,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-10-02T09:59:05.390","AcceptedAnswerId":52609081.0,"OwnerUserId":6590555.0,"Title":"Getting nan value in Output in Tensorflow","Body":"<p>Can you please help me in below code ?\nGetting nan value in Output in Tensorflow, when trying to get the values\nof w and b in tensorflow session<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport numpy.random as rand\n\n\ntrainX = np.array([[2.5,5.6,7.8,8.9]],dtype=np.float32)\n\ntrainY = np.array([[6.7,6.7,8.9,5.4]],dtype=np.float32)\nnum_samples = trainX.shape[0]\nnum_samples2 = trainY.shape[0]\nX = tf.placeholder(dtype=tf.float32)\nY = tf.placeholder(dtype=tf.float32)\ntype(X)\n w = tf.Variable(rand.randn(1),dtype=tf.float32)\nb = tf.Variable(rand.randn(1),dtype=tf.float32)\ntype(w)\ntf.shape(w)\ntf.shape(b)\nw = tf.convert_to_tensor(w,dtype=tf.float32)\nb = tf.convert_to_tensor(b,dtype=tf.float32)\nnum_iter = 10000 \nlearning_rate = 0.01\ndef model(X,w,b):\n    return tf.multiply(X,w) + b\ntype(X)\ntype(w)\npred = model(X,w,b)\ncost = tf.square(Y-pred)\noptimizer = tf.train.GradientDescentOptimizer(learning_rate)\ntrain = optimizer.minimize(cost)\ntext_file = open(\"Output.txt\", \"w\")\nmodel = tf.global_variables_initializer()\nwith tf.Session() as session:\n    session.run(model)\n    for i in range(num_iter):\n        session.run(train, feed_dict={X: trainX , Y: trainY})\n    print(session.run(w))\n    print(session.run(b))\n<\/code><\/pre>\n\n<p>Thanks and Regards,\nSubho<\/p>\n","answers":[{"AnswerId":"52609081","CreationDate":"2018-10-02T13:12:28.243","ParentId":null,"OwnerUserId":"7328893","Title":null,"Body":"<p>It is \"diverged\".<\/p>\n\n<p>Change learning rate lower.<\/p>\n\n<pre><code>#learning_rate = 0.01\nlearning_rate = 0.001\n<\/code><\/pre>\n\n<p>I confirmed below result.<\/p>\n\n<pre><code>[0.00044938]\n[6.922184]\n<\/code><\/pre>\n"},{"AnswerId":"52609933","CreationDate":"2018-10-02T13:58:27.023","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>One way is to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/clip_by_value\" rel=\"nofollow noreferrer\">\ntf.clip_by_value<\/a><\/p>\n\n<p>If the cost is calculated like this<\/p>\n\n<pre><code>cost = -tf.reduce_sum(Y*tf.log(tf.clip_by_value(pred,1e-10,1.0)))\n<\/code><\/pre>\n\n<p>you get a value for the weight.<\/p>\n\n<p>The following code is essentially yours. I have tried to get the weights from the list of trainable variables. You can get bias too.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport numpy.random as rand\n\n\ntrainX = np.array([[2.5,5.6,7.8,8.9]],dtype=np.float32)\n\ntrainY = np.array([[6.7,6.7,8.9,5.4]],dtype=np.float32)\nnum_samples = trainX.shape[0]\nnum_samples2 = trainY.shape[0]\nX = tf.placeholder(dtype=tf.float64)\nY = tf.placeholder(dtype=tf.float64)\nw = tf.get_variable(initializer=rand.randn(1),dtype=tf.float64,name=\"weights\")\nb = tf.Variable(rand.randn(1),dtype=tf.float64,name=\"bias\")\n\nnum_iter = 10000\n\nlearning_rate = 0.01\n\ndef model(X,w,b):\n    return tf.multiply(X,w) + b\n\npred = model(X,w,b)\n\ncost = -tf.reduce_sum(Y*tf.log(tf.clip_by_value(pred,1e-10,1.0)))\n\noptimizer = tf.train.GradientDescentOptimizer(learning_rate)\ntrain = optimizer.minimize(cost)\nmodel = tf.global_variables_initializer()\nwith tf.Session() as session:\n    session.run(model)\n    for i in range(num_iter):\n        session.run(train, feed_dict={X: trainX , Y: trainY})\n    var = [v for v in tf.trainable_variables() if v.name == \"weights:0\"]\n\n    #Both print the value\n    print(session.run(var))\n    print(session.run(w))\n\n    #print(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52606473,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning>","CreationDate":"2018-10-02T10:23:08.840","AcceptedAnswerId":null,"OwnerUserId":6643984.0,"Title":"Load pb model with python keras","Body":"<p>Is there a way to load .pb model made with keras specifically to know input and output names?<\/p>\n","answers":[]}
+{"QuestionId":52606751,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-02T10:42:46.863","AcceptedAnswerId":null,"OwnerUserId":10445598.0,"Title":"InvalidArgumentError: input_1:0 is both fed and fetched","Body":"<p>I found this GitHub example about <a href=\"https:\/\/github.com\/philipperemy\/keras-visualize-activations\/blob\/master\/read_activations.py\" rel=\"nofollow noreferrer\">print activation maps<\/a>. Code is quite simple. All I did was copy pasted the function. <\/p>\n\n<pre><code>def get_activations(model, model_inputs, print_shape_only=False, layer_name=None):\n    print('----- activations -----')\n    activations = []\n    inp = model.input\n\n    model_multi_inputs_cond = True\n    if not isinstance(inp, list):\n        # only one input! let's wrap it in a list.\n        inp = [inp]\n        model_multi_inputs_cond = False\n\n    outputs = [layer.output for layer in model.layers if\n               layer.name == layer_name or layer_name is None]  # all layer outputs\n\n    funcs = [K.function(inp + [K.learning_phase()], [out]) for out in outputs]  # evaluation functions\n\n    if model_multi_inputs_cond:\n        list_inputs = []\n        list_inputs.extend(model_inputs)\n        list_inputs.append(0.)\n    else:\n        list_inputs = [model_inputs, 0.]\n\n\n    print list_inputs\n    layer_outputs = [func(list_inputs)[0] for func in funcs]\n    for layer_activations in layer_outputs:\n        activations.append(layer_activations)\n        if print_shape_only:\n            print(layer_activations.shape)\n        else:\n            print(layer_activations)\n    return activations\n<\/code><\/pre>\n\n<p>And then I passed my model and inputs. However, it generates this error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"test_cnn_128.py\", line 80, in &lt;module&gt;\n    get_activations(model, test_x)\n  File \"test_cnn_128.py\", line 45, in get_activations\n    layer_outputs = [func(list_inputs)[0] for func in funcs]\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 2666, in __call__\n    return self._call(inputs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 2635, in _call\n    session)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 2587, in _make_callable\n    callable_fn = session._make_callable_from_options(callable_opts)\n  File \"\/home\/fatima\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1414, in _make_callable_from_options\n    return BaseSession._Callable(self, callable_options)\n  File \"\/home\/fatima\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1368, in __init__\n    session._session, options_ptr, status)\n  File \"\/home\/fatima\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/errors_impl.py\", line 519, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: input_1:0 is both fed and fetched.\nException tensorflow.python.framework.errors_impl.InvalidArgumentError: InvalidArgumentError() in &lt;bound method _Callable.__del__ of &lt;tensorflow.python.client.session._Callable object at 0x7f44de0cd210&gt;&gt; ignored\n<\/code><\/pre>\n\n<p>I am unsure of how to fix this.<\/p>\n","answers":[{"AnswerId":"55418198","CreationDate":"2019-03-29T13:12:15.810","ParentId":null,"OwnerUserId":"5838189","Title":null,"Body":"<p>As I posted on the thread <a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\/55417982#55417982\">Keras, How to get the output of each layer?<\/a>, the way to solve this is to replace the line <code>outputs = [layer.output for layer in model.layers if\n               layer.name == layer_name or layer_name is None]<\/code>\nwith <\/p>\n\n<p><code>outputs = [layer.output for layer in model.layers if\n               layer.name == layer_name or layer_name is None][1:]<\/code>  <\/p>\n"},{"AnswerId":"58112384","CreationDate":"2019-09-26T08:15:12.543","ParentId":null,"OwnerUserId":"12123574","Title":null,"Body":"<p>Exclude the input layer from your output\/fetch layers (which is why the error says \"fed &amp; fetch\" - can't get output(fetch) for input(feed) from the input layer (feed) itself)<\/p>\n\n<pre><code>    outputs = [layer.output for layer in model.layers if\n               layer.name == layer_name or layer_name is None][1:]\n<\/code><\/pre>\n\n<p>OR by dropping layer name stuff altogether;<\/p>\n\n<pre><code>    outputs = [layer.output for layer in model.layers][1:]\n<\/code><\/pre>\n\n<p>should fix the issue (notice --> <code>[1:]<\/code> at the end which is the simple fix for your issue)<\/p>\n"}]}
+{"QuestionId":52606828,"AnswerCount":1,"Tags":"<python><loops><session><tensorflow><keras>","CreationDate":"2018-10-02T10:47:18.860","AcceptedAnswerId":52607067.0,"OwnerUserId":10159479.0,"Title":"Keras set_session in for loop problem, not run and stop","Body":"<p>I want to reproduce result for <code>keras<\/code>, the backend is <code>tensorflow<\/code>.\nSo I use <code>ParameterGrid<\/code> for check that.\nWhen I run <code>K.set_session(Session)<\/code> second time in a loop, the program \nstop, and nothing to print. No error! How to fix it? Thanks.\nHere is my code.<\/p>\n\n<pre><code>'''Trains a simple convnet on the MNIST dataset.\nGets to 99.25% test accuracy after 12 epochs\n(there is still a lot of margin for parameter tuning).\n16 seconds per epoch on a GRID K520 GPU.\n'''\n\nfrom __future__ import print_function\nimport keras\nfrom keras.datasets import mnist\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras import backend as K\nimport numpy\nimport tensorflow\nfrom sklearn.model_selection import ParameterGrid\n\n##  TuneParameter\nTuneParameter = {}\nTuneParameter[\"Batch\"] = [50, 50, 50, 50, 50, 50]\nTuneParameter[\"Epoch\"] = [2]\nTuneParameter = ParameterGrid(TuneParameter)\n\n##  For each pair of parameter\nfor p in TuneParameter:\n\n    ##  Initial session\n    numpy.random.seed(2018)\n    tensorflow.set_random_seed(2018)\n    Session = tensorflow.Session(graph=tensorflow.get_default_graph())\n    K.set_session(Session)\n\n    # input image dimensions\n    img_rows, img_cols = 28, 28\n\n    # the data, split between train and test sets\n    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n\n    if K.image_data_format() == 'channels_first':\n        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n        input_shape = (1, img_rows, img_cols)\n    else:\n        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)\n        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)\n        input_shape = (img_rows, img_cols, 1)\n\n        x_train = x_train.astype('float32')\n        x_test = x_test.astype('float32')\n        x_train \/= 255\n        x_test \/= 255\n        print('x_train shape:', x_train.shape)\n        print(x_train.shape[0], 'train samples')\n        print(x_test.shape[0], 'test samples')\n\n    # convert class vectors to binary class matrices\n    y_train = keras.utils.to_categorical(y_train, 10)\n    y_test = keras.utils.to_categorical(y_test, 10)\n\n    model = Sequential()\n    model.add(Conv2D(32, kernel_size=(3, 3),\n                     activation='relu',\n                     input_shape=input_shape))\n    model.add(Conv2D(64, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.25))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(10, activation='softmax'))\n\n    model.compile(loss=keras.losses.categorical_crossentropy,\n                  optimizer=keras.optimizers.Adadelta(),\n                  metrics=['accuracy'])\n\n    model.fit(x_train, y_train,\n              batch_size=p[\"Batch\"],\n              epochs=p[\"Epoch\"],\n              verbose=1,\n              validation_data=(x_test, y_test))\n    score = model.evaluate(x_test, y_test, verbose=0)\n\n    K.clear_session()\n    print('Test loss:', score[0])\n    print('Test accuracy:', score[1])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52607067","CreationDate":"2018-10-02T11:04:06.843","ParentId":null,"OwnerUserId":"8484384","Title":null,"Body":"<p>Try<\/p>\n\n<pre><code>session = K.get_session()\n<\/code><\/pre>\n\n<p>Removing the:<\/p>\n\n<pre><code>Session = tensorflow.Session(graph=tensorflow.get_default_graph())\nK.set_session(Session)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52607063,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><warnings>","CreationDate":"2018-10-02T11:03:57.640","AcceptedAnswerId":52607304.0,"OwnerUserId":1259561.0,"Title":"Tensorflow warning: The graph couldn't be sorted in topological order?","Body":"<p>When I run <code>tensorflow<\/code> training (with custom defined graph, closed source), it outputs the warning:<\/p>\n\n<blockquote>\n  <p>2018-10-03 14:29:24.352895: E tensorflow\/core\/grappler\/optimizers\/dependency_optimizer.cc:666] Iteration = 0, topological sort failed with message: The graph couldn't be sorted in topological order.<\/p>\n<\/blockquote>\n\n<p>What does it mean? What could likely cause this problem and how to avoid it?<\/p>\n\n<p><strong>Update:<\/strong> For the record, in my case, <code>tensorflow<\/code> still works correctly despite this warning. So I think it just means there are more loops in the computation graph than what <code>tensorflow<\/code> expects, not necessarily infinite loop. I fixed this by avoid using manual loop in code, but instead using tensor manipulation (<code>stack, concat, slice, reshape...<\/code>)<\/p>\n","answers":[{"AnswerId":"52607304","CreationDate":"2018-10-02T11:18:57.787","ParentId":null,"OwnerUserId":"4378613","Title":null,"Body":"<p>A <a href=\"https:\/\/en.wikipedia.org\/wiki\/Topological_sorting\" rel=\"nofollow noreferrer\">topological ordering<\/a> of a directed graph is an ordering of its vertices in such a way that whenever there is an edge from vertex <code>u<\/code> to vertex <code>v<\/code>, vertex <code>u<\/code> comes before vertex <code>v<\/code> in the ordering.<\/p>\n\n<p>This kind of ordering is possible for every directed <strong>acyclic<\/strong> graph but not for arbitrary graphs. There is most likely some kind of cycle in your graph that prevents the sorting algorithm from succeeding. So, the way to go is searching for the cycle and removing it in some way.<\/p>\n\n<p>As an example, consider a very small graph with two vertices, <code>u<\/code> and <code>v<\/code>, and two edges, <code>u -&gt; v<\/code> and <code>v -&gt; u<\/code>. There is no way to sort u and v in accordance with the requirements given above. <\/p>\n\n<p>Without further information on your code it is hard to say exactly what is going on.<\/p>\n"}]}
+{"QuestionId":52609243,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-02T13:21:36.310","AcceptedAnswerId":null,"OwnerUserId":9242453.0,"Title":"TensorFlow slice Tensor","Body":"<p>I want to do something like this:<\/p>\n\n<pre><code>def f(ind, region, matrix):\n\n    ...\n<\/code><\/pre>\n\n<p>Parameter ind is placeholder or Tensor and contains coordinate x, y of central element, region is tuple of integers that represents shape of region. And you need to calculate indices and get region from matrix.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>ind = [2, 2]\nregion = [3, 3]\nmatrix = [[1, 1, 1, 1, 1], \n          [2, 2, 2, 2, 2], \n          [3, 3, 3, 3, 3], \n          [4, 4, 4, 4, 4], \n          [5, 5, 5, 5, 5]]\n\nindices = tf.placeholder(tf.in32, shape=...)\nregion = [3, 3]\nmat = tf.placeholder(tf.float32, shape=...)\n\nres = f(indices, region, mat)\nwith tf.Session() as sess:\n    r = sess.run(res, feed_dict={ indices: ind, mat: matrix })\n<\/code><\/pre>\n\n<p>Result is:<\/p>\n\n<pre><code>   r = [[2, 2, 2], \n        [3, 3, 3], \n        [4, 4, 4]]\n<\/code><\/pre>\n\n<p>because central element on (2, 2) is 3, and region of size 3 by 3 around is given by result r.<\/p>\n","answers":[]}
+{"QuestionId":52609374,"AnswerCount":1,"Tags":"<tensorflow><tensorflow.js>","CreationDate":"2018-10-02T13:28:11.867","AcceptedAnswerId":52609915.0,"OwnerUserId":7992070.0,"Title":"How to classify data for Tensorflow.js as 2d, 3d etc?","Body":"<p>Before I begin, I am completely new to machine learning and just starting to get my feet wet. I am sufficiently comfortable at JavaScript, and hence I thought I would give tensorflow.js a run. So please forgive my naivety.<\/p>\n\n<p>I have data which contains information for an individual sample as:<\/p>\n\n<pre><code>Probes  Signal\n1       150\n2       320\n3       15\n4       40\n.       .\n.       .\n10      200\n<\/code><\/pre>\n\n<p>Say I have 10 such known samples with different Signal levels for the probes. The probes and their order remain the same i.e. 1,2,3...10. This would be 3rd dimension?<\/p>\n\n<p>My eventual aim is to create a model using these known samples. Then for the Unknown sample I would also have Probe (same order) and Signal data as given above, and the model should predict what is the sample (or closest related sample).<\/p>\n\n<p>I am a little confused whether to represent this as a <code>tf.tensor2d<\/code> or <code>tf.tensor3d<\/code> <\/p>\n\n<p>If it is tensor2d can I represent it in the manner so that each individual array will contain the Signal data:<\/p>\n\n<p><code>tf.tensor2d([[150, 320, 15, 40, ....., 200], [&lt;data for sample2&gt;], [&lt;data for sample 3].....);<\/code><\/p>\n\n<p>If it is a tensor3d can I represent it in the manner so that it contains both the Probe number and the Signal data:<\/p>\n\n<p><code>tf.tensor3d([[[1,150], [2,320]], [[3,15], [4, 40],....,[10,200], [&lt;data for sample2&gt;], [&lt;data for sample 3]);<\/code><\/p>\n\n<p>On one hand I feel a 2D tensor would be sufficient as the probe order remains the same, however please correct me if I am wrong in that assumption. \nAny help is appreciated, thanks.<\/p>\n","answers":[{"AnswerId":"52609915","CreationDate":"2018-10-02T13:57:27.823","ParentId":null,"OwnerUserId":"9507270","Title":null,"Body":"<p>Since the probe ids(1-10) of your input data are consistant over each sample, you don't need to pass them to your model.<\/p>\n\n<p>So your data is only one-dimensional, more specific it has the shape: <code>[10]<\/code>, which is also the shape of your first layer.<\/p>\n\n<p>But since <code>model.fit()<\/code> and <code>model.predict()<\/code> take multiple samples as inputs they need  either one dimension more(<code>tensor2D<\/code>) or an array of tensors of the input shape of your first layer(<code>tensor1D[]<\/code>).<\/p>\n"}]}
+{"QuestionId":52609595,"AnswerCount":1,"Tags":"<python><tensorflow><normalization><word-embedding><fasttext>","CreationDate":"2018-10-02T13:40:40.430","AcceptedAnswerId":null,"OwnerUserId":7332158.0,"Title":"Normalize Fasttext word embedding vector generated by model","Body":"<p>I'm currently using FastText to get the word embeddings for some input text data in order to catch similarity among them and give those embedding as input to a neural network for a NER task.\nI've first tried to use the fastText .vec file and I was able to get good similarity result from them, then I've tried to use the .bin model file so that I can get embeddings also from out-of-vocabulary terms but I've noticed that the vectors generated from the .vec and the .bin file are different.\nThe .vec file contains embeddings clipped to a range of [-1,1] while this is not true for the embedding contained in the .bin file.\nUsing the .bin file I get very bad results so I'm trying to undestand if this is due to the fact the these vectors (.bin) are not normalized.<\/p>\n\n<p>Is there a good way to normalize these vectors in a range [-1, 1] using Python?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"52609856","CreationDate":"2018-10-02T13:54:50.223","ParentId":null,"OwnerUserId":"2271091","Title":null,"Body":"<p>Just re-scale the data into range [-1, 1]. Wish it may help.<\/p>\n\n<pre><code>def norm(data):\n    min_v, max_v = min(data), max(data)\n    mid_v = (max_v + min_v) &gt;&gt; 1\n    len_v = float((max_v - min_v) &gt;&gt; 1)\n    return [(v-mid_v)\/len_v for v in data]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52609798,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow><keras><linode>","CreationDate":"2018-10-02T13:51:53.197","AcceptedAnswerId":null,"OwnerUserId":3177617.0,"Title":"Keras creating three classes instead of two","Body":"<p>I am trying to train a model to identify images containing fire VS images that contain forests. I am training the model on a remote server using Linode. I am using Python 2.7 and Ubuntu 16.04.5.<\/p>\n\n<p>When i run the following code locally or in Jupyter notebooks it will create 2 classes, but when i want to run it on the server it creates 3 classes.<\/p>\n\n<p>The code that classifies the model:<\/p>\n\n<pre><code>def onehot(x): return np.array(OneHotEncoder().fit_transform(x.reshape(-1,1)).todense())\n\n\nmodel = keras.applications.InceptionV3(weights='imagenet', include_top=False)\n\n\nbatch_size=16\n\ntrain_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n\n\ntrain_generator = train_datagen.flow_from_directory(\n        'datareal\/train',  # this is the target directory\n        batch_size=batch_size,\n        target_size=(224, 224),\n        class_mode='binary')\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\nvalidation_generator = test_datagen.flow_from_directory(\n        'datareal\/valid',\n        batch_size=batch_size,\n        target_size=(224, 224),\n        class_mode='binary')\n\n\nx = model.output\nx = GlobalAveragePooling2D()(x)\n# let's add a fully-connected layer\nx = Dense(1024, activation='relu')(x)\n# and a logistic layer -- let's say we have 2 classes\npredictions = Dense(2, activation='softmax')(x)\nnewmodel = Model(inputs=model.input, outputs=predictions)\n\nnewmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['acc'])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Found 173 images belonging to 3 classes.\nFound 40 images belonging to 3 classes.\n<\/code><\/pre>\n\n<p>The directory containing all the images is structured as follows:<\/p>\n\n<pre><code>datareal\n        valid\n                forest\n                fire\n        train\n                forest\n                fire\n<\/code><\/pre>\n\n<p>How do i get the model to label only 2 classes instead of 3? <\/p>\n","answers":[{"AnswerId":"52610442","CreationDate":"2018-10-02T14:25:41.063","ParentId":null,"OwnerUserId":"3177617","Title":null,"Body":"<p>I figured it out. the problem was that i still had some hidden folders in my train and validation folder. <\/p>\n"}]}
+{"QuestionId":52610577,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-02T14:32:24.177","AcceptedAnswerId":52610821.0,"OwnerUserId":354722.0,"Title":"Feed both tf.data.Dataset and NumPy array to model","Body":"<p>I have a class representing a model that is set up as follows:<\/p>\n\n<pre><code>class Model:\n  def __init__(self):\n    self.setup_graph()\n\n  def setup_graph():\n    # sets up the model\n    ....\n\n  def train(self, dataset):\n    # dataset is a tf.data.Dataset iterator, from which I can get \n    # tf.Tensor objects directly, which become part of the graph\n    ....\n\n  def predict(self, sample):\n    # sample is a single NumPy array representing a sample,\n    # which could be fed to a tf.placeholder using feed_dict\n    ....\n<\/code><\/pre>\n\n<p>During training I want to make use of the efficiency of TensorFlow's <code>tf.data.Dataset<\/code>, but I still want to be able to get the output of the model on a single sample. It seems to me that this requires recreating the graph for prediction. Is this true, or can I create a TF graph where I can either run with a sample from a <code>tf.data.Dataset<\/code>, or with a given sample I feed to a <code>tf.placeholder<\/code>?<\/p>\n","answers":[{"AnswerId":"52610821","CreationDate":"2018-10-02T14:46:22.670","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can create your model with a dataset, iterator, etc as usual. Then, if you want to pass some custom data with <code>feed_dict<\/code>, you can just by passing values to the tensors produced by <code>get_next()<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndataset = (tf.data.Dataset\n    .from_tensor_slices(np.ones((100, 3), dtype=np.float32))\n    .batch(5))\niterator = dataset.make_one_shot_iterator()\nbatch = iterator.get_next()\n\noutput = 2 * batch\n\nwith tf.Session() as sess:\n    print('From iterator:')\n    print(sess.run(output))\n    print('From feed_dict:')\n    print(sess.run(output, feed_dict={batch: [[1, 2, 3]]}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>From iterator:\n[[2. 2. 2.]\n [2. 2. 2.]\n [2. 2. 2.]\n [2. 2. 2.]\n [2. 2. 2.]]\nFrom feed_dict:\n[[2. 4. 6.]]\n<\/code><\/pre>\n\n<p>In principle you could achieve the same effect with initializable, reinitializable or feedable iterators, but if you really just want to test single samples of data I think this is the quickest and less intrusive way.<\/p>\n"}]}
+{"QuestionId":52610740,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2018-10-02T14:41:36.730","AcceptedAnswerId":52649885.0,"OwnerUserId":9361210.0,"Title":"Can't import tensorflow in Python 3.6.6","Body":"<p>I installed tensorflow with <code>pip3 install tensorflow<\/code> and then I opened the terminal typed <code>python<\/code> and <code>import tensorflow<\/code> <br>\nBut when I hit <code>Enter<\/code> it gives me an Error: <code>Invalid machine command (memory dump written)<\/code> I already reinstalled tensorflow multiple times but it won't work. Could someone please help me? Thank you :D<\/p>\n","answers":[{"AnswerId":"52610809","CreationDate":"2018-10-02T14:45:43.693","ParentId":null,"OwnerUserId":"3216980","Title":null,"Body":"<p>Not sure what's causing your problem exactly but if you search for same problem and possible solutions online, you reach here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17411\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17411<\/a>.<\/p>\n\n<p>Basically a solution would be to install an earlier tensorflow version.<\/p>\n\n<p>Try:<\/p>\n\n<pre><code>pip uninstall tensorflow\npip install tensorflow==1.5\n<\/code><\/pre>\n"},{"AnswerId":"52649885","CreationDate":"2018-10-04T15:00:20.257","ParentId":null,"OwnerUserId":"9361210","Title":null,"Body":"<ol>\n<li>install anaconda or miniconda<\/li>\n<li>create a virtualenv with <code>conda create -n tensorflow python=3.6<\/code><br>\n2.5 activate the environement with <code>source activate tensorflow<\/code><\/li>\n<li>in the virtualenv <code>(tensorflow)name@name$<\/code> type <code>conda install tensorflow<\/code><\/li>\n<li>after installation type (always in the virtualenv) <code>python3<\/code> and then <code>import tensorflow<\/code><\/li>\n<li>Always launch tensorflow in the virtualenv. To deactivate the virtualenv type <code>exit<\/code>\nto reactivate it type <code>source activate tensorflow<\/code><\/li>\n<\/ol>\n"}]}
+{"QuestionId":52610922,"AnswerCount":0,"Tags":"<python><tensorflow><nvidia><ubuntu-18.04>","CreationDate":"2018-10-02T14:51:17.687","AcceptedAnswerId":null,"OwnerUserId":6167402.0,"Title":"Cant get tensorflow-gpu to work on Ubuntu MATE 18.04","Body":"<p>I tried to install tensorflow-gpu as per instructions given on Tensorflow website. My system configuration is as follows:-<br>\n    - CPU-Intel(R) Xeon(R) CPU E5-2650<br>\n    - GPU-Nvidia GTX 1080 ti<br>\n    - OS- Ubuntu Mate 18.04    <\/p>\n\n<ol>\n<li><p>For installing Nvidia Graphics card driver, i did following-<\/p>\n\n<ul>\n<li>sudo ubuntu-drivers autoinstall<\/li>\n<\/ul><\/li>\n<li><p>Running <code>sudo lshw -c display | grep configuration<\/code>  shows:-<\/p>\n\n<ul>\n<li>configuration: driver=nouveau latency=0<\/li>\n<\/ul><\/li>\n<li>Running  <code>gpu-manager shows<\/code>  following:-\n\n<ul>\n<li><a href=\"http:\/\/pasted.co\/93f21f9e\" rel=\"nofollow noreferrer\">http:\/\/pasted.co\/93f21f9e<\/a><\/li>\n<\/ul><\/li>\n<li><p>Output of nvcc -V is:-<\/p>\n\n<ul>\n<li>nvcc: NVIDIA (R) Cuda compiler driver<\/li>\n<li>Copyright (c) 2005-2017 NVIDIA Corporation<\/li>\n<li>Built on Fri_Sep__1_21:08:03_CDT_2017<\/li>\n<li>Cuda compilation tools, release 9.0, V9.0.176<\/li>\n<\/ul><\/li>\n<li><p>Finally, i installed CUDA 9.0 , cuDNN 7.3, CUPTI, as recommended. Tensorflow runs but uses CPU.  <\/p>\n\n<ul>\n<li>Error shown in runtime is:-  \n\n<ul>\n<li>2018-10-03 06:25:30.562101: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:300] failed call to cuInit: CUDA_ERROR_UNKNOWN: unknown error  <\/li>\n<li>2018-10-03 06:25:30.562186: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:150] kernel driver does not appear to be running on this host (rk39096-HP-Z840-Workstation): \/proc\/driver\/nvidia\/version does not exist  <\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<li><p>To disable noveau driver, I followed following steps:-  <\/p>\n\n<ul>\n<li><a href=\"https:\/\/linuxconfig.org\/how-to-disable-nouveau-nvidia-driver-on-ubuntu-18-04-bionic-beaver-linux\" rel=\"nofollow noreferrer\">https:\/\/linuxconfig.org\/how-to-disable-nouveau-nvidia-driver-on-ubuntu-18-04-bionic-beaver-linux<\/a>  <\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>This is my first question on SO. So please guide me if i didnt ask question as per specific format.<\/p>\n","answers":[]}
+{"QuestionId":52611080,"AnswerCount":1,"Tags":"<docker><tensorflow><tensorflow-serving>","CreationDate":"2018-10-02T15:00:31.257","AcceptedAnswerId":52632716.0,"OwnerUserId":6208328.0,"Title":"Tensor flow serving docker invalid field","Body":"<p>When running the sample example in tensorflow model server [<a href=\"https:\/\/www.tensorflow.org\/serving\/docker](Serving\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/serving\/docker](Serving<\/a> example part)<\/p>\n\n<p><code>docker run -p 8501:8501   --mount type=bind,  source=serving\/tensorflow_serving\/servables\/tensorflow\/testdata\/saved_model_half_plus_two_cpu,  target=\/models\/half_plus_two   -e MODEL_NAME=half_plus_two -t tensorflow\/serving &amp;<\/code><\/p>\n\n<p>the following error comes <\/p>\n\n<p><code>invalid argument \"type=bind,\" for \"--mount\" flag: invalid field '' must be a key=value pair\nSee 'docker run --help'.\n[1]+  Exit 125\n<\/code><\/p>\n\n<p>docker --version gives \nDocker version 18.06.1-ce<\/p>\n\n<p>The model server docker image is the latest version as well. <\/p>\n","answers":[{"AnswerId":"52632716","CreationDate":"2018-10-03T17:35:44.037","ParentId":null,"OwnerUserId":"10163556","Title":null,"Body":"<p>The command shouldn't have spaces after the commas.<\/p>\n\n<p><code>\ndocker run -p 8501:8501 --mount type=bind,source=serving\/tensorflow_serving\/servables\/tensorflow\/testdata\/saved_model_half_plus_two_cpu,target=\/models\/half_plus_two -e MODEL_NAME=half_plus_two -t tensorflow\/serving &amp;\n<\/code><\/p>\n"}]}
+{"QuestionId":52611575,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-02T15:27:37.337","AcceptedAnswerId":null,"OwnerUserId":9854132.0,"Title":"tf.while_loop gives wrong result when it runs in parallel","Body":"<p>I want to update a two dimensional <code>tf.variable<\/code> inside a <code>tf.while_loop<\/code> in tensorflow by row. For this reason, I use the <code>tf.assign<\/code> method. The problem is that with my implementation and <code>parallel_iterations&gt;1<\/code> the result is wrong. With <code>parallel_iterations=1<\/code> the result is correct. The code is like this:<\/p>\n\n<pre><code>a = tf.Variable(tf.zeros([100, 100]), dtype=tf.int64)\n\ni = tf.constant(0)\ndef condition(i, var):\n    return tf.less(i, 100)\n\ndef body(i, var):\n    updated_row = method() # This method returns a [1, 100] tensor which is the updated row for the variable\n    temp = tf.assign(a[i], updated_row)\n    return [tf.add(i, 1), temp]\n\nz = tf.while_loop(condition, body, [i, a], back_prop=False, parallel_iterations=10)\n<\/code><\/pre>\n\n<p>The iterations are completely independent and I do not know what is the problem.<\/p>\n\n<p>Strangely If I change the code like this:<\/p>\n\n<pre><code>a = tf.Variable(tf.zeros([100, 100]), dtype=tf.int64)\n\ni = tf.constant(0)\ndef condition(i, var):\n    return tf.less(i, 100)\n\ndef body(i, var):\n    zeros = lambda: tf.zeros([100, 100], dtype=tf.int64)\n    temp = tf.Variable(initial_value=zeros, dtype=tf.int64)\n    updated_row = method() # This method returns a [1, 100] tensor which is the updated row for the variable\n    temp = tf.assign(temp[i], updated_row)\n    return [tf.add(i, 1), temp]\n\nz = tf.while_loop(condition, body, [i, a], back_prop=False, parallel_iterations=10)\n<\/code><\/pre>\n\n<p>the code gives the correct outcome for <code>parallel_iterations&gt;1<\/code>. Can someone explain me what is going on here and give me an efficient solution to update the variable because the original variable I want to update is huge and the solution I found is very inefficient.<\/p>\n","answers":[{"AnswerId":"52611988","CreationDate":"2018-10-02T15:50:04.053","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You do not need to use variables for this, you can just generate the row-updated tensor on the loop body:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef method(i):\n    # Placeholder logic\n    return tf.cast(tf.range(i, i + 100), tf.float32)\n\ndef condition(i, var):\n    return tf.less(i, 100)\n\ndef body(i, var):\n    # Produce new row\n    updated_row = method(i)\n    # Index vector that is 1 only on the row to update\n    idx = tf.equal(tf.range(tf.shape(a)[0]), i)\n    idx = tf.cast(idx[:, tf.newaxis], var.dtype)\n    # Compose the new tensor with the old one and the new row\n    var_updated = (1 - idx) * var + idx * updated_row\n    return [tf.add(i, 1), var_updated]\n\n# Start with zeros\na = tf.zeros([100, 100], tf.float32)\ni = tf.constant(0)\ni_end, a_updated = tf.while_loop(condition, body, [i, a], parallel_iterations=10)\n\nwith tf.Session() as sess:\n    print(sess.run(a_updated))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[  0.   1.   2. ...  97.  98.  99.]\n [  1.   2.   3. ...  98.  99. 100.]\n [  2.   3.   4. ...  99. 100. 101.]\n ...\n [ 97.  98.  99. ... 194. 195. 196.]\n [ 98.  99. 100. ... 195. 196. 197.]\n [ 99. 100. 101. ... 196. 197. 198.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52612052,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-02T15:54:21.590","AcceptedAnswerId":null,"OwnerUserId":10022495.0,"Title":"Using particular size for the image in Keras resNet50","Body":"<p>I am trying to use RestNet50 (keras). I want to execute it with several images \n of varying sizes since the default size of ResNet50 is 224x244. I have tried to change the argument <code>include_top<\/code> to <code>False<\/code> but it still does not work.<\/p>\n","answers":[{"AnswerId":"52612930","CreationDate":"2018-10-02T16:51:36.710","ParentId":null,"OwnerUserId":"2723405","Title":null,"Body":"<p>You can use the preprocessing function included in applications.resnet50 as described here: <a href=\"https:\/\/stackoverflow.com\/a\/50136397\/2723405\">preprocessing images generated using keras function ImageDataGenerator() to train resnet50 model<\/a><\/p>\n\n<p>This will generate images form your inputs that are compatible with  ResNet50<\/p>\n\n<pre><code>from keras.applications.resnet50 import preprocess_input\n\ntrain_datagen=ImageDataGenerator(preprocessing_function=preprocess_input)\ntrain_generator = train_datagen.flow_from_directory(directory, batch_size, shuffle=True, target_size, class_mode)  \n<\/code><\/pre>\n"}]}
+{"QuestionId":52612435,"AnswerCount":2,"Tags":"<python><string><dictionary><tensorflow>","CreationDate":"2018-10-02T16:18:04.277","AcceptedAnswerId":52613014.0,"OwnerUserId":10447167.0,"Title":"How to convert string without quotes to dictionary in python","Body":"<p>I have to convert string without quotes into dictionary.<\/p>\n\n<pre><code>device: 0, name: GeForce GTX 1080 8GB, pci bus id: 0000:01:00.0\n<\/code><\/pre>\n\n<p>the 'device', 'name' and 'pci bus id' have to be keys,<\/p>\n\n<p>and '0', 'GeForce GTX 1080 8GB', '0000:01:00.0' have to be values.<\/p>\n\n<p>I get this from tensorflow.python.client.list_local_devices()<\/p>\n","answers":[{"AnswerId":"52613014","CreationDate":"2018-10-02T16:58:43.867","ParentId":null,"OwnerUserId":"10255652","Title":null,"Body":"<p>Using, two <code>.split()<\/code>'s and <em>dictionary comprehension<\/em>, first <code>.split(', ')<\/code> divides up the entire string, the second <code>split(': ')<\/code> divides up the items of list to be cast as <code>keys and values<\/code>    <\/p>\n\n<pre><code>s = \"device: 0, name: GeForce GTX 1080 8GB, pci bus id: 0000:01:00.0\"\nd = {i.split(': ')[0]: i.split(': ')[1] for i in s.split(', ')}\n<\/code><\/pre>\n\n<blockquote>\n<pre><code>{'device': '0', 'name': 'GeForce GTX 1080 8GB', 'pci bus id': '0000:01:00.0'}\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"52612645","CreationDate":"2018-10-02T16:32:30.080","ParentId":null,"OwnerUserId":"8726235","Title":null,"Body":"<p>First, you have to split the string by ', ' (comma and space) to separate each key : value in your string. After, for each string with 'key: value' you must split it for ': ' (colon and space) to get key and value separately to build your dictionary.<\/p>\n\n<pre><code>    dict = {}\n    s = \"device: 0, name: GeForce GTX 1080 8GB, pci bus id: 0000:01:00.0\"\n\n    '''split to separate the main string in 'key: value' substrings'''\n    key_value = s.split(\", \")\n\n'''each substring is separated in key and value to be appended into dictionary'''\n\n    for v in key_value:\n       aux = v.split(\": \")\n       dict[aux[0]] = aux[1]\n\n    print(dict)\n<\/code><\/pre>\n\n<p>The outuput:<\/p>\n\n<pre><code>{'name': 'GeForce GTX 1080 8GB', 'device': '0', 'pci bus id': '0000:01:00.0'}\n<\/code><\/pre>\n"}]}
+{"QuestionId":52612455,"AnswerCount":0,"Tags":"<python><tensorflow><keras><nlp><lstm>","CreationDate":"2018-10-02T16:19:18.653","AcceptedAnswerId":null,"OwnerUserId":6057371.0,"Title":"tensorflow.python.framework.errors_impl.InvalidArgumentError: in KERAS LSTM model","Body":"<p>I have the following lstm model:<\/p>\n\n<pre><code>class LSTM_model():\n\ndef __init__(self):\n    w2v_model = gensim.models.Word2Vec(sentences, size=150, window=10, min_count=2, workers=10)\n    pretrained_weights = w2v_model.wv.syn0\n    vocab_size, emdedding_size = pretrained_weights.shape\n    self.w2v_model = w2v_model\n    self.keras_lstm_model = Sequential()\n    self.keras_lstm_model.add(Embedding(input_dim = vocab_size, output_dim = emdedding_size, weights = [pretrained_weights]))\n    self.keras_lstm_model.add(LSTM(units = emdedding_size))\n    self.keras_lstm_model.add(Dense(units = vocab_size))\n    self.keras_lstm_model.add(Activation('sigmoid'))\n    self.keras_lstm_model.compile(optimizer = 'adam', loss = 'sparse_categorical_crossentropy', metrics = ['mae','acc'])\n<\/code><\/pre>\n\n<p>My aim is to predict probablity of a given word in a context.\nI have list of sentences and I want to train this model: <\/p>\n\n<pre><code>   def train_lstm_model(self, sentences):\n\n        sentences_as_indexes = []\n\n\n        filtered_sentences = list(filter(lambda x : all([w in self.w2v_model.wv.vocab for w in x]) , sentences)) #filter to take only sentences with no OOV words\n        for sentence in filtered_sentences:\n            if all([w in self.w2v_model.wv.vocab for w in sentence]): #todo use filter\n                indexes_row = []\n                for word in sentence:\n\n                    idx = self.w2v_model.wv.vocab.get(word).index\n                    indexes_row.append(idx)\n\n                sentences_as_indexes.append(indexes_row)\n        X = pd.DataFrame([sentence[:-1] for sentence in sentences_as_indexes])\n        y = pd.DataFrame([sentence[-1] for sentence in sentences_as_indexes])\n        print(datetime.datetime.now(), \": Fitting LSTM model , size of X is \", X.shape)\n\n        self.keras_lstm_model.fit(X, y) #HERE the error\n<\/code><\/pre>\n\n<p>Where sentences is list of around 1M sentences (my training set).\nI Get the following error in the fit():<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: indices[7,38] = -2147483648 is not in [0, 694415)\n       [[Node: embedding_1\/embedding_lookup = GatherV2[Taxis=DT_INT32, Tindices=DT_INT32, Tparams=DT_FLOAT, _class=[\"loc:@training\/Adam\/Assign_2\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](embedding_1\/embeddings\/read, embedding_1\/Cast, training\/Adam\/gradients\/embedding_1\/embedding_lookup_grad\/concat\/axis)]]<\/p>\n<\/blockquote>\n\n<p>I though this might be because I have data that is OOV but as you can see - I filtered the sentences to take only those with all the word in the vocab.<\/p>\n\n<p>What may cause this error?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52612933,"AnswerCount":1,"Tags":"<tensorflow><linear-regression>","CreationDate":"2018-10-02T16:51:52.497","AcceptedAnswerId":52614875.0,"OwnerUserId":2026722.0,"Title":"What are good MSE and RMSE for my normalized dataset to between 0 and 1","Body":"<p>I using a training source CSV file raed into a master dataframe that I split into 80% training data and 20% test data. Before I split the data I normalized all columns of the dataframe to have all the independent and dependent data to be between 0 and 1, including the targets (dependent variables). In my results after training my predicted values all read between 0 and 1. I then de-normalize a single prediction to see what value I get and compare to the expected value. My question is I'm measuring the model by MSE (mean squared error) and RMSE (root mean squared error). My MSE and RMSE on my training data are 0.03 and 0.16, respectively. Are these acceptable values with a normalized data source? If not, with my normalized data source, what would be acceptable values? Or should I be normalizing my data because I don't have huge differences of range differences between my independent variables? If I don't normalize my data should I then use a Normalized RMSE to interpret the metric? If I normalize the RMSE when not normalizing the training and test data, what would be an acceptable value for the Normalized RMSE? Thanks in advance for any replies.<\/p>\n","answers":[{"AnswerId":"52614875","CreationDate":"2018-10-02T19:07:22.233","ParentId":null,"OwnerUserId":"5728493","Title":null,"Body":"<p>\u201cGood\u201d should be measure relative to a naive forecast (such as a random walk). That benchmark will vary according to the degree of volatility in the data. .5 might be terrible for one forecast and excellent for another. <\/p>\n"}]}
+{"QuestionId":52612984,"AnswerCount":2,"Tags":"<tensorflow><neural-network><conv-neural-network><cudnn>","CreationDate":"2018-10-02T16:56:29.230","AcceptedAnswerId":null,"OwnerUserId":4209599.0,"Title":"How to extract Tensorflow trained weights from graph.pbtxt to raw data","Body":"<p>I have trained a custom neural network with the function:<\/p>\n\n<p><code>tf.estimator.train_and_evaluate<\/code><\/p>\n\n<p>After correct training, it contains the following files:<\/p>\n\n<ul>\n<li>checkpoint  <\/li>\n<li>events.out.tfevents.1538489166.ti   <\/li>\n<li>model.ckpt-0.data-00000-of-00002 <\/li>\n<li>model.ckpt-0.index   <\/li>\n<li>model.ckpt-10.data-00000-of-00002      <\/li>\n<li>model.ckpt-10.index eval      <\/li>\n<li>graph.pbtxt                           <\/li>\n<li>model.ckpt-0.data-00001-of-00002   <\/li>\n<li>model.ckpt-0.meta       <\/li>\n<li>model.ckpt-10.data-00001-of-00002   <\/li>\n<li>model.ckpt-10.meta<\/li>\n<\/ul>\n\n<p>Now I need to export the <em>weights<\/em> and <em>biases<\/em> of every layer, into a raw data structure, e.g. an array, <code>numpy<\/code>.<\/p>\n\n<p>I have read multiple pages on TensorFlow, and on other topics, but neither can find this question. The first thing I would assume to put the fils together into <code>graph.pd<\/code> with the freeze.py as suggested here: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/45864363\/tensorflow-how-to-convert-meta-data-and-index-model-files-into-one-graph-pb?rq=1\">Tensorflow: How to convert .meta, .data and .index model files into one graph.pb file<\/a><\/p>\n\n<p>But then still the main question is unsolved.<\/p>\n","answers":[{"AnswerId":"52615013","CreationDate":"2018-10-02T19:17:55.027","ParentId":null,"OwnerUserId":"9437711","Title":null,"Body":"<p>If saving weights and biases in a numpy array is your strict requirement, you can follow this example: <\/p>\n\n<pre><code># In a TF shell, define all requirements and call the model function \ny = model(x, is_training=False, reuse=tf.AUTO_REUSE) # For example\n<\/code><\/pre>\n\n<p>Once you call this function, you can see all the variables in the graph by running<\/p>\n\n<pre><code>tf.global_variables()\n<\/code><\/pre>\n\n<p>You need to restore all these variables from the latest checkpoint (say <strong>ckpt_dir<\/strong>) and then execute each of these variables to get the latest values.<\/p>\n\n<pre><code>checkpoint = tf.train.latest_checkpoint('.\/model_dir\/')\nfine_tune = tf.contrib.slim.assign_from_checkpoint_fn(checkpoint,\n                                                      tf.global_variables(),\n                                                      ignore_missing_vars=True)\n\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\ngv = sess.run(tf.global_variables())\n<\/code><\/pre>\n\n<p>Now <strong>gv<\/strong> will be a list of all the values of your variables (weights and biases); You can access any individual component via indexing - <strong>gv[5]<\/strong> etc. Or you can convert the entire thing into an array and save using numpy.<\/p>\n\n<pre><code>np.save('my_weights', np.array(gv))\n<\/code><\/pre>\n\n<p>This will save all your weights and biases in your current working directory as a numpy array - 'my_weights.npy'. <\/p>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"52613333","CreationDate":"2018-10-02T17:21:53.077","ParentId":null,"OwnerUserId":"7078425","Title":null,"Body":"<p>If you wish to evaluate tensors alone, you can check out <a href=\"https:\/\/stackoverflow.com\/questions\/33633370\/how-to-print-the-value-of-a-tensor-object-in-tensorflow\">this question<\/a>. But if you wish to e.g. deploy your network, you can take a look at TensorFlow <a href=\"https:\/\/www.tensorflow.org\/serving\/\" rel=\"nofollow noreferrer\">serving<\/a>, which is probably the most performant one right now. Or if you want to export this network to other frameworks and use them there, you can actually use <a href=\"https:\/\/onnx.ai\" rel=\"nofollow noreferrer\">ONNX<\/a> for this purpose.<\/p>\n"}]}
+{"QuestionId":52613383,"AnswerCount":1,"Tags":"<python><gpu><cpu><pytorch>","CreationDate":"2018-10-02T17:24:58.043","AcceptedAnswerId":null,"OwnerUserId":8592877.0,"Title":"A better way to make pytorch code agnostic to running on a CPU or GPU?","Body":"<p>The Migration <a href=\"https:\/\/pytorch.org\/blog\/pytorch-0_4_0-migration-guide\/\" rel=\"noreferrer\">guide<\/a> recommends the following to make code CPU\/GPU agnostic:<\/p>\n\n<pre><code>&gt; # at beginning of the script\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n...\n# then whenever you get a new Tensor or Module\n# this won't copy if they are already on the desired device\ninput = data.to(device)\nmodel = MyModule(...).to(device)\n<\/code><\/pre>\n\n<p>I did this and ran my code on a CPU-only device, but my model crashed when fed an input array, as saying it was expecting a CPU tensor not a GPU one.  Somehow my model was automatically converting the a CPU-input array to a GPU array.  Finally I traced it down to this command in my code:<\/p>\n\n<pre><code>model = torch.nn.DataParallel(model).to(device)\n<\/code><\/pre>\n\n<p>Even though I convert the model to 'cpu', the nn.DataParallel overrides this.  The best solution I came up with was a conditional:<\/p>\n\n<pre><code>if device.type=='cpu':\n    model = model.to(device)\nelse:\n    model = torch.nn.DataParallel(model).to(device)\n<\/code><\/pre>\n\n<p>This does not seem elegant.  Is there a better way?<\/p>\n","answers":[{"AnswerId":"52664444","CreationDate":"2018-10-05T11:18:00.393","ParentId":null,"OwnerUserId":"4892874","Title":null,"Body":"<p>How about<\/p>\n\n<pre><code>if torch.cuda.device_count() &gt; 1:\n    model = torch.nn.DataParallel(model)\nmodel = model.to(device)\n<\/code><\/pre>\n\n<p>?<\/p>\n\n<p>You don't need <code>DataParallel<\/code> if you have only one GPU.<\/p>\n"}]}
+{"QuestionId":52613883,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><caffe><object-detection>","CreationDate":"2018-10-02T18:00:07.723","AcceptedAnswerId":null,"OwnerUserId":4339694.0,"Title":"training images? Considerations for selection","Body":"<p>I'm relatively new and am still learning the basics. I've used NVIDIA DIGITS in the past, and am now looking at Tensorflow. While I've been able to fumble my way around creating some models for a few projects I'm working on, I really want to start diving deeper into what I'm doing, how I'm doing it, and ultimately a better understanding of why.<\/p>\n\n<p>One area that I would like to start with is the Images that I'm using for training and testing. Can anyone point me to a blog, an article, a paper, or give me some insight in what I need to consider when selecting images to train a new model on. Up until recently, I've been using datasets that have already been selected and that are available for download. Lets say I'm going to start working on a project that involves object detection of ships from a variety of distances and angles.<\/p>\n\n<p>So my thoughts would be<\/p>\n\n<p>1) I need a large quantity of images.\n2) The images need to contain ships of the different types I would like to detect. (lets just say one class, ships, don't care what type of ships)\n3) I also need to have images that have a great variety of distance perspective for the different types of ships.<\/p>\n\n<p>Ultimately, my thoughts are that the images need to reflect the distance, perspective, and types of ships I would ideally want to identify from the video. Seems simple enough.<\/p>\n\n<p>However, there are a number of questions<\/p>\n\n<p>Does the images need to be the same\/similar resolution as the camera I'll be using, for best results?\nDoes the images all need to be the same resolution?\nCan I use a single image and just digitally zoom out on the image to give the illusion of different distances?<\/p>\n\n<p>I'm sure there are a number of other questions that I'm not asking, or should be asking. Are there any guide lines available for creating a solid collection of images to use when creating the collection of images for training and validation?<\/p>\n","answers":[{"AnswerId":"52614147","CreationDate":"2018-10-02T18:18:14.150","ParentId":null,"OwnerUserId":"3802483","Title":null,"Body":"<p>I recommend thinking through end to end, like would you need to classify ship models as a next step? I recommend going through well known public datasets and actually work with the structure, how to store data, labels, how to handle preprocessing etc.<\/p>\n\n<p>More importantly, what are you trying to achieve? Talking to experts in the topic does help greatly while preparing your own dataset. <\/p>\n\n<p>Use open source images if you can, e.g. flickr, google, imagenet.<\/p>\n\n<p>No, you don't need them to be the same resolution. <\/p>\n\n<p>It is not ideal to zoom in\/out images to use in different categories. Preprocessing images and data augmentation already does this to create more distant representations of the same class. This is why I would recommend hands on approach with an existing dataset first.<\/p>\n\n<p>Yes, what you need is many, different representations of classes, and a roughly balanced dataset of classes. If you define your data structure well in the beginning, it will save you a ton of time as you won't have to make changes often.<\/p>\n"}]}
+{"QuestionId":52613977,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-02T18:06:08.557","AcceptedAnswerId":null,"OwnerUserId":4713057.0,"Title":"Data augmentation: recuperate the transformed and original images alongside","Body":"<p>I am working with MNIST images and I want to perform some data augmentation techniques using Keras <code>ImageDataGenerator<\/code>.<\/p>\n\n<p>I m wondering if I can get the original images alongside with the transformed ones.\nHere is the code so far. Actually, I don't know how to recuperate the original images corresponding to the transformed ones.<\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nimport matplotlib.pyplot as plt\n\ndef load_mnist():\n    # the data, shuffled and split between train and test sets\n    from tensorflow.keras.datasets import mnist\n    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n    x = np.concatenate((x_train, x_test))\n    y = np.concatenate((y_train, y_test))\n    x = x.reshape([-1, 28, 28, 1]) \/ 255.0\n    print('MNIST samples', x.shape)\n    return x, y\n\ndef show_images(X_original, X_transformed, nb_images=50, img_h=28, img_w=28):\n        plt.figure(figsize=(40, 4))\n        for i in range(nb_images):\n                # display original\n                ax = plt.subplot(2, nb_images, i + 1)\n                plt.imshow(X_original[i].reshape(28, 28))\n                plt.gray()\n                ax.get_xaxis().set_visible(False)\n                ax.get_yaxis().set_visible(False)\n\n                # display reconstruction\n                ax = plt.subplot(2, nb_images, i + 1 + nb_images)\n                plt.imshow(X_transformed[i].reshape(28, 28))\n                plt.gray()\n                ax.get_xaxis().set_visible(False)\n                ax.get_yaxis().set_visible(False)\n        plt.show()\n\ndatagen = ImageDataGenerator(width_shift_range=0.1, height_shift_range=0.1, rotation_range=10, zoom_range=[0.8, 1.2])\nX, Y = load_mnist()\ngen0 = datagen.flow(X, Y, shuffle=True, batch_size=256)\nX1, Y1 = gen0.next()\nshow_images(X_original=?, X_transformed=X1, nb_images=50, img_h=28, img_w=28)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52614606,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-10-02T18:49:48.933","AcceptedAnswerId":null,"OwnerUserId":4233984.0,"Title":"Constructing SparseTensors from dataset elements and lookup table for estimator","Body":"<p>I am trying to create a single tf.SparseTensor to feed into a model. I want to construct a sparse tensor per input then concat them row-wise and then batch them. The inputs are strings which are turned into indices via a lookup table from the contrib.lookup module. The model and input pipeline are constructed via the estimator paradigm and the dataset API. My example protos are in TFRecord files like so:<\/p>\n\n<pre><code>{'user_id': ['user123'],\n 'item_id':['item123'],\n 'label': [.12]}\n<\/code><\/pre>\n\n<p>my <code>input_fn<\/code> looks like this:<\/p>\n\n<pre><code>def dataset_input_fn(tfrecord_pattern, batch_size,\n                     n_epochs, user_lookup, item_lookup):\n\n    def _make_sparse_features(features):\n        keys_to_features = {\n            'user_id': tf.FixedLenFeature([1], tf.string),\n            'item_id': tf.FixedLenFeature([1], tf.string),\n            'label': tf.FixedLenFeature([1], tf.float32)\n         }\n        features = tf.parse_single_example(features, keys_to_features)\n        user_tensor = sparse_tensor_from_lookup(user_lookup, features['user_id'])\n\n        item_tensor = sparse_tensor_from_lookup(item_lookup, features['item_id'])\n\n        sp_inputs = [user_tensor, item_tensor]\n        sparse_features = tf.sparse_concat(axis=0, values=sp_inputs)\n\n        return sparse_features, features['label']\n\n    dataset = tf.data.Dataset.from_tensor_slices(glob.glob(tfrecord_pattern))\n    dataset = dataset.shuffle(buffer_size=1028)\n    dataset = dataset.flat_map(tf.data.TFRecordDataset)\n    dataset = dataset.map(_make_sparse_features)\n\n    dataset = dataset.repeat(num_epochs)\n    dataset = dataset.batch(batch_size)\n\n    return dataset\n<\/code><\/pre>\n\n<p>where <code>sparse_tensor_from_lookup<\/code> is defined like so (I have tried both <code>tf.SparseTensor<\/code> and <code>tf.SparseTensorValue<\/code>):<\/p>\n\n<pre><code>def sparse_tensor_from_lookup(lookup, key_tensor):\n    indices = lookup.lookup(key_tensor)\n    sparse_tensor = tf.SparseTensorValue(indices=indices,\n                                         values=tf.ones_like(indices),\n                                         dense_shape=lookup.size())\n    return sparse_tensor\n<\/code><\/pre>\n\n<p>The lookup tables are instantiated in my <code>main<\/code> training\/eval function:<\/p>\n\n<pre><code>def main_training_and_eval_loop(model_function,\n                            training_epochs=10000, batch_size=256,\n                            max_training_steps=8000000, learning_rate=1e-2, l2_weight=1e-3,\n                            reader_n_threads=2, parser_n_threads=2,\n                            save_checkpoint_every_n_steps=2000, factor_order=2,\n                            max_steps_without_decrease=20000, shuffle_buffer_size=100000,\n                            **context):\n\n    model_directory = '\/tmp\/'\n\n    users_id_file_path = os.path.join(model_directory, 'vocab\/users.txt')\n    item_id_file_path = os.path.join(model_directory, 'vocab\/items.txt')\n\n    tfrecord_dir = 'airflow_tmp\/tfrecords_backfill_2018-09-29T00:43:32.624007\/'\n\n    item_lookup = index_table_from_file(item_id_file_path,\n                                              num_oov_buckets=1)\n    user_lookup = index_table_from_file(users_id_file_path,\n                                     num_oov_buckets=1)\n\n    training_input_fn = functools.partial(dataset_input_fn,\n                                          os.path.join(tfrecord_dir, '*.tfrecord'), \n                                          batch_size=batch_size,\n                                          n_epochs=training_epochs,\n                                          user_lookup=user_lookup,\n                                          item_lookup=item_lookup)\n\n\n... etc\n<\/code><\/pre>\n\n<p>The error I receive seems to be with regard to the values being passed to <code>tf.SparseTensor<\/code> when trying to construct said tensor:<\/p>\n\n<pre><code>&lt;ipython-input-5-ec313e9d7732&gt; in _make_sparse_features(features)\n 17         }\n 18         features = tf.parse_single_example(features, keys_to_features)\n---&gt; 19         course_tensor = sparse_tensor_from_lookup(course_lookup, features['target_course_id'])\n     20 \n     21         course_tags = tf.sparse_tensor_to_dense(features['course_tags'], default_value='')\n\n~\/Code\/learning-recommendations\/airflow\/dags\/utils\/tensorflow\/sparse.py in sparse_tensor_from_lookup(lookup, key_tensor)\n      6     sparse_tensor = tf.SparseTensor(indices=indices,\n      7                                     values=tf.ones_like(indices),\n----&gt; 8                                     dense_shape=lookup.size())\n      9     return sparse_tensor\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/sparse_tensor.py in __init__(self, indices, values, dense_shape)\n    132     self._dense_shape = dense_shape\n    133 \n--&gt; 134     indices_shape = indices.get_shape().with_rank(2)\n    135     values_shape = values.get_shape().with_rank(1)\n    136     dense_shape_shape = dense_shape.get_shape().with_rank(1)\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py in with_rank(self, rank)\n    745       return self.merge_with(unknown_shape(ndims=rank))\n    746     except ValueError:\n--&gt; 747       raise ValueError(\"Shape %s must have rank %d\" % (self, rank))\n    748 \n    749   def with_rank_at_least(self, rank):\n\nValueError: Shape (1,) must have rank 2\n<\/code><\/pre>\n\n<p>but this differs from the error I get with <code>tf.SparseTensorValue<\/code>:<\/p>\n\n<pre><code>&lt;ipython-input-5-ec313e9d7732&gt; in _make_sparse_features(features)\n     30 \n     31         sp_inputs = [course_tensor, course_tags_tensor, user_tensor, user_interests_tensor]\n---&gt; 32         sparse_features = tf.concat(axis=0, values=sp_inputs)\n     33 \n     34         return sparse_features, features['label']\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py in concat(values, axis, name)\n   1112               tensor_shape.scalar())\n   1113       return identity(values[0], name=scope)\n-&gt; 1114   return gen_array_ops.concat_v2(values=values, axis=axis, name=name)\n   1115 \n   1116 \n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py in concat_v2(values, axis, name)\n   1028     _attr_N = len(values)\n   1029     _, _, _op = _op_def_lib._apply_op_helper(\n-&gt; 1030         \"ConcatV2\", values=values, axis=axis, name=name)\n   1031     _result = _op.outputs[:]\n   1032     _inputs_flat = _op.inputs\n\n\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    481                                 (prefix, dtype.name))\n    482               else:\n--&gt; 483                 raise TypeError(\"%s that don't all match.\" % prefix)\n    484             else:\n    485               raise TypeError(\"%s that are invalid.\" % prefix)\n\nTypeError: Tensors in list passed to 'values' of 'ConcatV2' Op have types [&lt;NOT CONVERTIBLE TO TENSOR&gt;, &lt;NOT CONVERTIBLE TO TENSOR&gt;] that don't all match.\n<\/code><\/pre>\n\n<p>Am I off the res here? I've struggled to find materials on how to work this kind of workflow with the dataset api and an even harder time finding material on the differences between <code>tf.SparseTensor<\/code> and <code>tf.SparseTensorValue<\/code> or much info on sparse tensors at all. <\/p>\n","answers":[]}
+{"QuestionId":52614665,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><serialization><protocol-buffers><tfrecord>","CreationDate":"2018-10-02T18:53:52.477","AcceptedAnswerId":52616053.0,"OwnerUserId":3753275.0,"Title":"Tensorflow parses the record incorrectly","Body":"<p>I trying to create tfrecords for my sematic segmentation dataset (rgb_image_in -> binary_raycast_out). <\/p>\n\n<p>Below is my code to write the list of images to a train.tfrecord.<\/p>\n\n<pre><code>    def _process_image_files(image_names, raycast_names):\n\n        writer = tf.python_io.TFRecordWriter('train')\n\n        #My implementation of decoding jpeg\/png image\n        coder = ImageCoder()\n\n        for i in range(len(image_names)):\n            print('{}\\n{}\\n\\n'.format(image_names[i], raycast_names[i]))\n\n            image_buffer, im_height, im_width, im_channels = _process_image(image_names[i], coder)\n\n            raycast_buffer, rc_height, rc_width, rc_channels = _process_image(raycast_names[i], coder)\n\n            example = _convert_to_example(image_names[i], raycast_names[i], image_buffer, raycast_buffer, \\\n                                          im_height, im_width, im_channels)\n\n            writer.write(example.SerializeToString())\n        writer.close()\n        sys.stdout.flush() \n\ndef _process_image(filename, coder):\n    with tf.gfile.FastGFile(filename, 'rb') as f:\n        image_data = f.read()\n\n    # Convert any PNG to JPEG's for consistency.\n    if _is_png(filename):\n        print('Converting PNG to JPEG for %s' % filename)\n        image_data = coder.png_to_jpeg(image_data)\n\n    # Decode the RGB JPEG.\n    image = coder.decode_jpeg(image_data)\n\n    # Check that image converted to RGB\n    assert len(image.shape) == 3\n    height = image.shape[0]\n    width = image.shape[1]\n    channels = image.shape[2]\n    assert channels == 3\n\n    return image_data, height, width, channels\n\n\ndef _convert_to_example(image_name, raycast_name, image_buffer, raycast_buffer, sample_height, sample_width, sample_channels):\n\n    example = tf.train.Example(features=tf.train.Features(feature={\n        'height': _int64_feature(sample_height),\n        'width': _int64_feature(sample_width),\n        'channels': _int64_feature(sample_channels),\n        'image\/filename': _bytes_feature(tf.compat.as_bytes(image_name)),\n        'image\/encoded': _bytes_feature(tf.compat.as_bytes(image_buffer)),\n        'raycast\/filename': _bytes_feature(tf.compat.as_bytes(raycast_name)),\n        'raycast\/encoded': _bytes_feature(tf.compat.as_bytes(raycast_buffer))}))\n\n    return example\n<\/code><\/pre>\n\n<p>The above code works fine in creating the tfrecord file. I put some print statements inside the <code>_convert_to_example<\/code> method to make sure the corresponding filenames (image_file &amp; raycast_file) are getting written in one example.<\/p>\n\n<p>However, when I read the examples from tfrecord and print the image names, it looks like the image_file &amp; raycast_file names do not correspond. The pair of images read by the tfRecordReader() is wrong.<\/p>\n\n<p>Below is my code to read the record:<\/p>\n\n<pre><code>def parse_example_proto(example_serialized):\n\n    feature_map = {\n                    'image\/encoded': tf.FixedLenFeature([], dtype=tf.string, default_value=''),\n                    'raycast\/encoded': tf.FixedLenFeature([], dtype=tf.string, default_value=''),\n                    'height': tf.FixedLenFeature([1], dtype=tf.int64, default_value=-1),\n                    'width': tf.FixedLenFeature([1], dtype=tf.int64, default_value=-1),\n                    'channels': tf.FixedLenFeature([1], dtype=tf.int64, default_value=-1),\n                    'image\/filename': tf.FixedLenFeature([], dtype=tf.string, default_value=''),\n                    'raycast\/filename': tf.FixedLenFeature([], dtype=tf.string, default_value='')\n                    }\n\n    features = tf.parse_single_example(example_serialized, feature_map)\n\n    return features['image\/encoded'], features['raycast\/encoded'], \\\n           features['height'], features['width'], features['channels'],\\\n           features['image\/filename'], features['raycast\/filename']\n\n\n\ndef retirieve_samples():\n\n    with tf.name_scope('batch_processing'):\n        data_files = ['train']\n\n        filename_queue = tf.train.string_input_producer(data_files, shuffle=False)\n\n        reader = tf.TFRecordReader()\n\n        _, example_serialized = reader.read(filename_queue)\n\n        image_buffer, raycast_buffer, height, width, channels, image_name, raycast_name = parse_example_proto(example_serialized)            \n\n        orig_image = tf.image.resize_images(tf.image.decode_jpeg(image_buffer, channels=3), \n                                            [480, 856])\n        orig_raycast = tf.image.resize_images(tf.image.decode_jpeg(raycast_buffer, channels=3), \n                                              [480, 856])\n\n        return image_name, raycast_name\n<\/code><\/pre>\n\n<p>Below is my code to print a pair of filenames<\/p>\n\n<pre><code>image_name, raycast_name = retirieve_samples()\nwith tf.Session() as sess:    \n    for i in range(1):\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n        print(sess.run(image_name))\n        print(sess.run(raycast_name))\n        coord.request_stop()\n        coord.join(threads)\n<\/code><\/pre>\n\n<p>I have spent few days on this. I am not able to identify why I am not able to retrieve the correct pair. An example being retrieved should have the same data as the example being created right ? Why am I seeing different name pairs when I read and write ?<\/p>\n\n<p>Any help would be appriciated<\/p>\n","answers":[{"AnswerId":"52616053","CreationDate":"2018-10-02T20:37:58.100","ParentId":null,"OwnerUserId":"8676953","Title":null,"Body":"<p>A smaller example would be better.<\/p>\n\n<p>Each <code>session.run<\/code> will evaluate the tensor and run the graph. That means if you evaluate <code>image_name<\/code> and <code>raycast_name<\/code> separately, then you will get them from different runs and they won't be a pair.<\/p>\n\n<p>You could get the pair by evaluating both at the same time, e.g.:<\/p>\n\n<pre><code>current_image_name, current_raycast_name = session.run([\n    image_name, raycast_name\n])\n<\/code><\/pre>\n\n<p>I would also recommend to use the newer <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">Dataset API<\/a> over the queues.<\/p>\n"}]}
+{"QuestionId":52614690,"AnswerCount":0,"Tags":"<python-3.x><keras><lstm>","CreationDate":"2018-10-02T18:55:29.580","AcceptedAnswerId":null,"OwnerUserId":10447830.0,"Title":"ValueError: Input arrays should have the same number of samples as target arrays. LSTM Keras","Body":"<p>I know multiple similar problems have been posted on the forum already, but I couldn't derive the solution from them... Anyway here is my problem: <\/p>\n\n<p>I have an Pandas DataFrame of shape (10240, 3). The first two columns contain numeric values, while the third and last column holds labels for each row.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code># Selecting and reshaping the X_train set\r\nX_train=df.loc[0:8191,df.columns != 'Label']\r\nX_train=X_train.values.reshape([1, 8192, 2])\r\n\r\n# Selecting and converting to categorical y_train\r\ntemp = df['Label']\r\ny_train = temp[0:8192]\r\ny_train = np_utils.to_categorical(y_train, num_classes = 5)\r\n\r\n# Selecting and reshaping X_test\r\nX_test = df.loc[8192:10240,df.columns != 'Label']\r\nX_test = X_test.values.reshape([1, 2048, 2])\r\n\r\n# Selecting and coverting to categorical y_test\r\ny_test = temp[8192:10240]\r\ny_test = np_utils.to_categorical(y_test, num_classes = 5)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Next, I try to insert the data with 5 different classes in my LSTM model, which looks the following:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>model = Sequential()\r\n\r\nmodel.add(LSTM(100, input_shape = (8192, 2), return_sequences = True))\r\nmodel.add(TimeDistributed(Dense(50)))\r\nmodel.add(GlobalAvgPool1D())\r\nmodel.add(Dense(5, activation = 'softmax'))<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>model.compile(loss = 'categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])\r\nprint(model.summary())\r\nhistory = model.fit(X_train, y_train, validation_data = (X_test, y_test), batch_size = 64, epochs = 40)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Upon compiling I then get the following error: ValueError: Input arrays should have the same number of samples as target arrays. Found 1 input samples and 8192 target samples.<\/p>\n\n<p>I checked whether the shapes of the different sets match up, and to my best of knowledge they do. Furthermore I think the input follows the [sample, timesteps, feature] \"rule of thumb\". And also the batch size seems to be ok. Anyhone has any idea how to get the input to work?<\/p>\n\n<p>I'm running Python 3.x and Keras version 2.2.2<\/p>\n","answers":[]}
+{"QuestionId":52614786,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><conv-neural-network><pytorch>","CreationDate":"2018-10-02T19:01:10.193","AcceptedAnswerId":null,"OwnerUserId":10447809.0,"Title":"How to implement a location specific convolutional filters in Tensorflow or Pytorch?","Body":"<p>I want to implement a convolutional layer, with a different convolutional filter for each output location. Specifically, think of the case when the output is of 16*16*128 (W * H * C). Instead of having a 3*3*128 filter we have 16*16 filters; each with size 3*3*128. This would lead to huge amount of parameters, but it can the case be that each of the 3*3*128 filter may be the same except scaled by a different constant, and the constants can be learned through a side network. In this way the number of parameters won't be too much.<\/p>\n\n<p>The similar idea is briefly in <a href=\"https:\/\/arxiv.org\/abs\/1605.09673\" rel=\"nofollow noreferrer\">Dynamic Filter Networks<\/a>, but I cannot find an implementation of location specific filters. My question is, if we want a location specific convolutional filter how do I implement it in Tensorflow or Pytorch? Do I need to write my own operation or there is some smart way to use the functions provided? If I have to write an OP is there any trick that can easily achieve this idea? Any help is appreciated!<\/p>\n","answers":[{"AnswerId":"52623576","CreationDate":"2018-10-03T09:12:27.577","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>A convolution, by definition, is <strong>not<\/strong> location specific - this is what makes it a convolution. If you wish to generalize convolution, bear in mind that eventually a convolution is a special case of a simple linear operation. Therefore, you can implement your \"location specific\" convolution as a fully-connected layer (<code>nn.Linear<\/code>) with a very specific sparse weights.  <\/p>\n"}]}
+{"QuestionId":52614922,"AnswerCount":1,"Tags":"<tensorflow><scikit-learn><neural-network><keras><roc>","CreationDate":"2018-10-02T19:11:04.443","AcceptedAnswerId":52615050.0,"OwnerUserId":10265949.0,"Title":"How to plot training loss and accuracy curves for a MLP model in Keras?","Body":"<p>I am modeling a neural network using Keras and I am trying to evaluate it with a graph of <code>acc<\/code> and <code>val_acc<\/code>. I have 3 errors in the following lines of code:<\/p>\n\n<ol>\n<li>In <code>print(history.keys())<\/code> The error is <code>function' object has not attribute 'keys'<\/code><\/li>\n<li>In <code>y_pred = classifier.predict(X_test)<\/code> The error is <code>name 'classifier' is not defined<\/code><\/li>\n<li>In <code>plt.plot(history.history['acc'])<\/code> The error is  <code>'History' object is not subscriptable<\/code><\/li>\n<\/ol>\n\n<p>I'm also trying to graph the ROC curve, how could I do it?<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom sklearn import cross_validation\nfrom matplotlib import pyplot\nfrom keras.utils import plot_model\n\ndataset = pd.read_csv('Data_BP.csv')\nX = dataset.iloc[:, 0:11].values\ny = dataset.iloc[:, -1].values\n\nfrom sklearn.model_selection import train_test_split\nX_train, X_test, y_train, y_test = cross_validation.train_test_split(X, y, test_size = 0.2, random_state = 0)\n\nfrom sklearn.preprocessing import StandardScaler\nsc = StandardScaler()\nX_train = sc.fit_transform(X_train)\nX_test = sc.transform(X_test)\n\ndef Model():\n    classifier = Sequential()\n    classifier.add(Dense(units = 12, kernel_initializer = 'uniform', activation = 'relu', input_dim = 11))\n    classifier.add(Dense(units = 8, kernel_initializer = 'uniform', activation = 'relu'))\n    classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n    classifier.compile(optimizer = 'adam', loss = 'mean_squared_error', metrics = ['mse', 'acc'])\n    return classifier\n\nclassifier = Model()\nhistory = classifier.fit(X_train, y_train, validation_split=0.25, batch_size = 10, epochs = 5)\n\nprint('\\n', history.history.keys())\n\ny_pred = classifier.predict(X_test)\ny_pred = (y_pred &gt; 0.5)\n\nfrom sklearn.metrics import recall_score, classification_report, auc, roc_curve\ncm = confusion_matrix(y_test, y_pred)\nprint(cm)\n\n\nplt.plot(history.history['acc'])\nplt.plot(history.history['val_acc'])\nplt.title('Model accuracy')\nplt.ylabel('Accuracy')\nplt.xlabel('Epoch')\nplt.legend(['Train', 'Test'], loc='upper left')\nplt.show()\n<\/code><\/pre>\n\n<p>What functions should be added?<\/p>\n","answers":[{"AnswerId":"52615050","CreationDate":"2018-10-02T19:21:18.880","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Change <code>history<\/code> to <code>classifier<\/code> in the following lines (actually <code>History<\/code> object is the return value of <code>fit<\/code> method called on <code>Model<\/code> object) like this:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>classifier = Model()\nhistory = classifier.fit(...)\n<\/code><\/pre>\n\n<p>Don't confuse the return value of <code>fit<\/code> method with your model. The <code>History<\/code> object, as its name suggests, only contains the history of training. However, your model is <code>classifier<\/code> and <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">it is the one that has methods<\/a> like <code>fit()<\/code>, <code>predict()<\/code>, <code>evaluate()<\/code>, <code>compile()<\/code>, etc.<\/p>\n\n<p>Plus, the <code>History<\/code> object has an attribute called <code>history<\/code> which is a dictionary containing the values of loss and metrics during the training. Therefore you need to use <code>print(history.history.keys())<\/code> instead.<\/p>\n\n<p>Now, if you would like to for example plot loss curve during training (i.e. loss at the end of each epoch) you can do it like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loss_values = history.history['loss']\nepochs = range(1, len(loss_values)+1)\n\nplt.plot(epochs, loss_values, label='Training Loss')\nplt.xlabel('Epochs')\nplt.ylabel('Loss')\nplt.legend()\n\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52614936,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><tensorboard><sentiment-analysis>","CreationDate":"2018-10-02T19:11:48.670","AcceptedAnswerId":null,"OwnerUserId":7179000.0,"Title":"How to build embedding layer with tensorflow for the following model?","Body":"<p>This is a keras model for sentiment analysis i need to convert it to tensorflow i couldn\u2019t build embedding layer with tensorflow and using confusion matrix to evaluate this model? And I asked if tf-learn is the-same as tensorflow<\/p>\n\n<pre><code>import os     \nimport numpy as np  \nimport pandas as pd  \nimport tensorflow as tf  \nfrom tensorflow import set_random_seed\nset_random_seed(2)\nfrom nltk.tokenize import word_tokenize\nfrom sklearn.utils import shuffle\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import confusion_matrix\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras.preprocessing.text import Tokenizer\nfrom sklearn.preprocessing import  LabelEncoder\nfrom keras.utils import to_categorical\nfrom keras.models import Sequential\nfrom keras.layers.embeddings import Embedding\nfrom keras.layers import Flatten\nfrom keras.layers import Conv1D, MaxPooling1D\nfrom keras.layers import Dense,Activation\nfrom keras.layers import Dropout\nfrom keras.callbacks import TensorBoard, ModelCheckpoint\nimport re\nimport string\nimport collections\nimport time\nseed = 10\n<\/code><\/pre>\n\n<p><strong>Read CSV Files<\/strong><\/p>\n\n<pre><code>df=pd.read_csv('tweets-pos-neg.csv', usecols = ['text','airline_sentiment'])\ndf = df.reindex(['text','airline_sentiment'], axis=1) #reorder columns\ndf=df.apply(lambda x: x.astype(str).str.lower()) \n<\/code><\/pre>\n\n<p><strong>Normalize Text<\/strong><\/p>\n\n<pre><code>def normalize(text):\n    text= re.sub(r\"http\\S+\", r'', text) \n    text= re.sub(r\"@\\S+\", r'', text)\n    punctuation = re.compile(r'[!\"#$%&amp;()*+,-.\/:;&lt;=&gt;?@[\\]^_`{|}~|0-9]')\n    text = re.sub(punctuation, ' ', text)\n    text= re.sub(r'(.)\\1\\1+', r'\\1', text) \n    return text\n<\/code><\/pre>\n\n<p><strong>Cleaned Text<\/strong><\/p>\n\n<pre><code>def prepareDataSets(df):\n     sentences=[]\n     for index, r in df.iterrows():\n         text= normalize(r['text'])\n         sentences.append([text,r['airline_sentiment']])\n          df_sentences=pd.DataFrame(sentences,columns= \n          ['text','airline_sentiment'])\n     return df_sentences\nedit_df=prepareDataSets(df)\nedit_df=shuffle(edit_df)\nX=edit_df.iloc[:,0]\nY=edit_df.iloc[:,1]\n<\/code><\/pre>\n\n<p><strong>Split reviews to tokens<\/strong>  <\/p>\n\n<pre><code> max_features = 50000\n tokenizer = Tokenizer(num_words=max_features, split=' ')\n tokenizer.fit_on_texts(X.values)\n #convert review tokens to integers\n X_seq = tokenizer.texts_to_sequences(X)\n<\/code><\/pre>\n\n<p><strong>Padding Sequence to make all vectors with the same size according to MAX-length of reviews<\/strong> <\/p>\n\n<pre><code>seq_len=35\nX_pad = pad_sequences(X_seq,maxlen=seq_len)   \n<\/code><\/pre>\n\n<p><strong>Convert target value from string to integer<\/strong><\/p>\n\n<pre><code>le=LabelEncoder()\nY_le=le.fit_transform(Y)\nY_le_oh=to_categorical(Y_le)\n<\/code><\/pre>\n\n<p><strong>Train-Test-Split<\/strong><\/p>\n\n<pre><code>X_train, X_test, Y_train, Y_test = train_test_split(X_pad,Y_le_oh, test_size \n= 0.33, random_state = 42)\nX_train, X_Val, Y_train, Y_Val = train_test_split(X_train,Y_train, test_size \n= 0.1, random_state = 42)\nprint(X_train.shape,Y_train.shape)\nprint(X_test.shape,Y_test.shape)\nprint(X_Val.shape,Y_Val.shape) \n<\/code><\/pre>\n\n<p><strong>Create the model<\/strong><\/p>\n\n<pre><code>embedding_vecor_length = 32    #no of vector columns\nmodel_cnn = Sequential()\nmodel_cnn.add(Embedding(max_features, embedding_vecor_length, \ninput_length=seq_len))\nmodel_cnn.add(Conv1D(filters=100, kernel_size=2, padding='valid', \nactivation='relu', strides=1))\nmodel_cnn.add(MaxPooling1D(2))\nmodel_cnn.add(Flatten())\nmodel_cnn.add(Dense(256, activation='relu'))\nmodel_cnn.add(Dense(2, activation='softmax'))\nopt=tf.keras.optimizers.Adam(lr=0.001, decay=1e-6)\nmodel_cnn.compile(loss='binary_crossentropy', optimizer=opt, metrics= \n['accuracy'])\nprint(model_cnn.summary())  \n<\/code><\/pre>\n\n<p><strong>Evaluate model<\/strong><\/p>\n\n<pre><code>history=model_cnn.fit(X_train, Y_train, epochs=3, batch_size=32, callbacks=[tensorboard], validation_data=(X_Val, Y_Val))\nscores = model_cnn.evaluate(X_test, Y_test, verbose=0)\nprint(\"Accuracy: %.2f%%\" % (scores[-1]*100))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52615017","CreationDate":"2018-10-02T19:18:19.663","ParentId":null,"OwnerUserId":"187343","Title":null,"Body":"<p>If you just need to use Tensorflow APIs to train \/ evaluate, you can build an Estimator using <code>model_to_estimator<\/code> function.<\/p>\n\n<p>Here's the <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators#creating_estimators_from_keras_models\" rel=\"nofollow noreferrer\">documentation<\/a> with an example.<\/p>\n"}]}
+{"QuestionId":52615296,"AnswerCount":2,"Tags":"<python-3.x><pip><pycharm><pytorch>","CreationDate":"2018-10-02T19:41:00.917","AcceptedAnswerId":null,"OwnerUserId":3311406.0,"Title":"PyCharm won't install PyTorch - Building wheel failed","Body":"<p>I have tried for some time to install PyTorch and TorchVision to aid in some neural networks \/ ML projects I'm doing.<\/p>\n\n<p>My issue is that I can' seem to get the installation to work, neither in PyCharm pip, Anaconda prompt or the cmd\/with the pip3 command.<\/p>\n\n<p>The output I get when trying to run the following command through PyCharm:<\/p>\n\n<pre><code>pip install --user pytorch==0.1.2\n<\/code><\/pre>\n\n<p>is this:<\/p>\n\n<pre><code>Collecting pytorch\n  Using cached https:\/\/files.pythonhosted.org\/packages\/a9\/41\/4487bc23e3ac4d674943176f5aa309427b011e00607eb98899e9d951f67b\/pytorch-0.1.2.tar.gz\nBuilding wheels for collected packages: pytorch\n  Running setup.py bdist_wheel for pytorch: started\n  Running setup.py bdist_wheel for pytorch: finished with status 'error'\n  Complete output from command C:\\Users\\Even\\Anaconda3\\envs\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Even\\\\AppData\\\\Local\\\\Temp\\\\pycharm-packaging\\\\pytorch\\\\setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" bdist_wheel -d C:\\Users\\Even\\AppData\\Local\\Temp\\pip-wheel-2sfgqmm7 --python-tag cp36:\n  Traceback (most recent call last):\n    File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    File \"C:\\Users\\Even\\AppData\\Local\\Temp\\pycharm-packaging\\pytorch\\setup.py\", line 17, in &lt;module&gt;\n      raise Exception(message)\n  Exception: You should install pytorch from http:\/\/pytorch.org\n\n  ----------------------------------------\n  Running setup.py clean for pytorch\nFailed to build pytorch\nInstalling collected packages: pytorch\n  Running setup.py install for pytorch: started\n    Running setup.py install for pytorch: finished with status 'error'\n    Complete output from command C:\\Users\\Even\\Anaconda3\\envs\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Even\\\\AppData\\\\Local\\\\Temp\\\\pycharm-packaging\\\\pytorch\\\\setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record C:\\Users\\Even\\AppData\\Local\\Temp\\pip-record-1jx5e1ke\\install-record.txt --single-version-externally-managed --compile --user --prefix=:\n    Traceback (most recent call last):\n      File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n      File \"C:\\Users\\Even\\AppData\\Local\\Temp\\pycharm-packaging\\pytorch\\setup.py\", line 13, in &lt;module&gt;\n        raise Exception(message)\n    Exception: You should install pytorch from http:\/\/pytorch.org\n\n    ----------------------------------------\n\n  Failed building wheel for pytorch\nCommand \"C:\\Users\\Even\\Anaconda3\\envs\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Even\\\\AppData\\\\Local\\\\Temp\\\\pycharm-packaging\\\\pytorch\\\\setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record C:\\Users\\Even\\AppData\\Local\\Temp\\pip-record-1jx5e1ke\\install-record.txt --single-version-externally-managed --compile --user --prefix=\" failed with error code 1 in C:\\Users\\Even\\AppData\\Local\\Temp\\pycharm-packaging\\pytorch\\\n<\/code><\/pre>\n\n<p>I've also tried this command (CMD):<\/p>\n\n<pre><code>pip3 install http:\/\/download.pytorch.org\/whl\/cpu\/torch-0.4.1-cp36-cp36m- \nwin_amd64.whl\n<\/code><\/pre>\n\n<p>Which returned in this output:<\/p>\n\n<pre><code>torch-0.4.1-cp36-cp36m-win_amd64.whl is not a supported wheel on this \nplatform.\n<\/code><\/pre>\n\n<p>I might be missing some crucial information, but I've been googling, searching forums, and asked people I know about this. No one has an answer. I hope some of you may have encountered the same issue.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"52620282","CreationDate":"2018-10-03T05:34:10.747","ParentId":null,"OwnerUserId":"9427337","Title":null,"Body":"<p>Firstly, verify if the installation is complete in the local terminal. Look out for windows specific installation from <a href=\"https:\/\/pytorch.org\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/<\/a><\/p>\n\n<p>Once this is done, make sure that you select the correct python interpreter in pycharm. Refer the link: <a href=\"https:\/\/www.jetbrains.com\/help\/pycharm\/installing-uninstalling-and-upgrading-packages.html\" rel=\"nofollow noreferrer\">https:\/\/www.jetbrains.com\/help\/pycharm\/installing-uninstalling-and-upgrading-packages.html<\/a> to  manage packages for a particular Python interpreter.<\/p>\n"},{"AnswerId":"54258600","CreationDate":"2019-01-18T17:11:42.167","ParentId":null,"OwnerUserId":"9532057","Title":null,"Body":"<ol>\n<li><p>I solved this on <a href=\"https:\/\/pytorch.org\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/<\/a> as told by error. And suffered from <\/p>\n\n<p>torch-0.4.1-cp36-cp36m-win_amd64.whl is not a supported wheel on this \nplatform<\/p><\/li>\n<\/ol>\n\n<p>. as well. But I figured out it is my CUDA version that cause the error. I thought it was 8\/9, but it's 10 actually. You can check pip and python version with <code>pip --version<\/code>.<\/p>\n\n<ol start=\"2\">\n<li>There is another point that we should notice. Pip and Anaconda install packages in different ways. Previously, Pytorch was installed by Anaconda, but I cannot import in Anaconda Prompt (<code>No module named torch<\/code>). I have to use pip to install to fix the error.<\/li>\n<\/ol>\n\n<p>Hope this would help.<\/p>\n"}]}
+{"QuestionId":52615715,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><tensorboard>","CreationDate":"2018-10-02T20:11:54.793","AcceptedAnswerId":52655272.0,"OwnerUserId":10425539.0,"Title":"Understanding Keras LSTM Tensorboard Graph","Body":"<p>I am confused about the graph I am getting in Tensborboad for my Keras LSTM network. I have defined my Keras LSTM network like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(neurons, return_sequences=True,input_shape=(look_back,2)))\n#model.add(Bidirectional(LSTM(neurons, return_sequences=True),input_shape=(look_back,2)))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(LSTM(neurons,return_sequences=True,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))   \nmodel.add(LSTM(20,return_sequences=False,recurrent_regularizer=l2(weight_decay),\n          kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),dropout=dropout,recurrent_dropout=dropout))\nmodel.add(Dense(outputs,kernel_regularizer=l2(weight_decay),bias_regularizer=l2(weight_decay),activation='linear'))\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I thought that would give me a sequential model, where each LSTM takes the output of the previous LSTM. I sort of get that. But I also get one of the LSTM layers as the input to every single subsequent layer:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rCYWL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rCYWL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In the graph, it looks like <strong>lstm_2<\/strong> feeds into every layer. I wouldn't have expected that. So my question is, is this expected? And if so, why? <\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"52655272","CreationDate":"2018-10-04T21:17:43.813","ParentId":null,"OwnerUserId":"10425539","Title":null,"Body":"<p>I figured out why it is showing like that. It turns out that Keras creates a learning_phase placeholder and it places it in the second hidden layer. The learning_phase object branches out to every single layer, but the LSTM itself does not. I refer to <a href=\"https:\/\/stackoverflow.com\/questions\/46259467\/tensorboard-and-dropout-layers\">this answer<\/a> for more details. <\/p>\n\n<p>Here's what the insides of my LSTM_1 layer looked like in my Tensorboard graph:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wwmUx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wwmUx.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":52616110,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-02T20:43:02.960","AcceptedAnswerId":null,"OwnerUserId":3602093.0,"Title":"Training on a single class car but predicting some other class","Body":"<p>I am using Deformable fpn to train on a single class for e.g car and i want to predict car in my testing script, but its predicting for e.g airplane. What should i do.<\/p>\n","answers":[]}
+{"QuestionId":52616287,"AnswerCount":1,"Tags":"<python><machine-learning><keras><gcloud><google-cloud-ml>","CreationDate":"2018-10-02T20:56:24.033","AcceptedAnswerId":52620076.0,"OwnerUserId":4187329.0,"Title":"How to set an environment variable when using gcloud ML engine?","Body":"<p>All,<\/p>\n\n<p>(Environments: Windows 7, Python 3.6, Keras &amp; tensorflow libs, gcloud ml engine)<\/p>\n\n<p>I am running certain Keras ML model examples using gcloud ml engine as introduced <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/getting-started-training-prediction\" rel=\"nofollow noreferrer\">here<\/a>. Everything was fine but I just got various results among multiple runs although I was using the same training and validation data. My goal is to make reproductive training results from multiple runs.<\/p>\n\n<p>I googled it for a while and found some solutions in this <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-reproducible-results-using-keras-during-development\" rel=\"nofollow noreferrer\">Keras Q&amp;A<\/a> regarding making reproductive results. Basically they first suggested this: <\/p>\n\n<blockquote>\n  <p>First, you need to set the PYTHONHASHSEED environment variable to 0 before the program starts (not within the program itself). <\/p>\n<\/blockquote>\n\n<p>I know I could set the variables locally on my own machine, or I could set it when I deploy gcloud function as introduced <a href=\"https:\/\/cloud.google.com\/functions\/docs\/env-var\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>But, I just do not know how to set environment variables when I was using <strong>gcloud ML engine<\/strong> (on server side but NOT on local). So I cannot set \"PYTHONHASHSEED=0\" on the gcloud server when my model programs running there.<\/p>\n\n<p>BTW, in general I know the randomness is a useful nature in the ML field but I am not very familiar about the topic of making reproductive results yet, so any thoughts regarding this topic are also welcome. Thanks!<\/p>\n\n<p>Daqi<\/p>\n\n<p>PS:\nI have tried to setting the environment variables in running time below:<\/p>\n\n<pre><code>import os\nos.environ[\"PYTHONHASHSEED\"] = \"0\" \nprint(hash(\"keras\"))\n<\/code><\/pre>\n\n<p>But it cannot make the effect that \"setting the variables before the program starts\". So by having this code, I still cannot get the same hash results from the multiple runs. On the other hand, on local, if I set \"PYTHONHASHSEED=0\" before running the code, I may get the same hash results.<\/p>\n","answers":[{"AnswerId":"52620076","CreationDate":"2018-10-03T05:13:39.113","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>I don't believe the Cloud ML Engine API provides a mechanism to set environment variables. However, you might be able to workaround this by writing a wrapper script (NB: UNTESTED CODE):<\/p>\n\n<pre><code>import os\nimport subprocess\n\nenv = os.environ.copy()\nenv[\"PYTHONHASHSEED\"] = \"0\"\n\nsubprocess.check_call(['python', 'main.py'], env=env)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52616650,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><tensor>","CreationDate":"2018-10-02T21:27:25.717","AcceptedAnswerId":52617244.0,"OwnerUserId":6145828.0,"Title":"'Tensor' object has no attribute '_keras_history'","Body":"<p>I wanted to extract and apply independently a <code>Conv2D<\/code> layer on the columns of my input tensors, but after adding the code:<\/p>\n\n<pre><code>accelerometer_input = Input(shape=(1400, 3))\n\nfor i in range(3):\n    out = Lambda(lambda x: x[:,:, i:i+1])(accelerometer_input) # Extracting the ith channel\n    out = K.expand_dims(out, axis=1)\n    out = Conv2D(64, (30, 1), data_format=\"channels_first\")(out)  \n    branch_outputs.append(out)\nout_put = K.concatenate(branch_outputs)\n<\/code><\/pre>\n\n<p>it gives me the error in the title. I think it is due to the <code>Lambda<\/code> layer or the <em>extraction<\/em> which is not differentiable. <\/p>\n\n<p>But how can I do without it?<\/p>\n","answers":[{"AnswerId":"52617244","CreationDate":"2018-10-02T22:22:17.930","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>That's because you are directly applying a backend function (i.e. <code>K.expand_dims()<\/code>) on a Keras Tensor (i.e. <code>out<\/code>) and therefore the result would be a Tensor (and not a Keras Tensor). Actually, a Keras Tensor is an augmented version of Tensor and have additional attributes (e.g. <code>_keras_history<\/code>) which helps Keras to build the model. Now, to resolve the issue, you just need to put the backend function in a <code>Lambda<\/code> layer to have a Keras Tensor as output:<\/p>\n\n<pre><code>out = Lambda(lambda x: K.expand_dims(x, axis=1))(out)\n<\/code><\/pre>\n\n<p>The same thing applies to using <code>K.concatenate()<\/code>. However in this case, there is a specific layer for it in Keras:<\/p>\n\n<pre><code>from keras.layers import concatenate, Concatenate\n\n# use functional interface\nout_put = concatenate(branch_outputs)\n\n# or use layer class\nout_put = Concatenate()(branch_outputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52616670,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-02T21:28:33.433","AcceptedAnswerId":null,"OwnerUserId":9660268.0,"Title":"LSTM print gradients keras","Body":"<p>I am trying to print the gradients and I have this so far. <\/p>\n\n<pre><code>my_model=load_model('word2vec_lstm.h5')\n\nweights = my_model.weights # weight tensors\n\ngradients = my_model.optimizer.get_gradients(my_model.total_loss, weights) # gradient tensors\n\ninput_tensors = [my_model.inputs[0], # input data\n                 my_model.sample_weights[0], # sample weights\n                 my_model.targets[0], # labels\n                 K.learning_phase(), # train or test mode\n]\n\nget_gradients = K.function(inputs=input_tensors, outputs=gradients)\n\ninputs = [[trainDataVecs], # X input data\n          [1], # sample weights\n          [Y_train], # y labels\n          0.005 # learning phase in TEST mode\n]\n\nprint [a for a in zip(weights, get_gradients(inputs))]\n<\/code><\/pre>\n\n<p>However, this gives me the following error <code>InvalidArgumentError: transpose expects a vector of size 4. But input(1) is a vector of size 3<\/code><\/p>\n","answers":[]}
+{"QuestionId":52616941,"AnswerCount":1,"Tags":"<machine-learning><neural-network><computer-vision><pytorch><convolution>","CreationDate":"2018-10-02T21:52:46.663","AcceptedAnswerId":52621058.0,"OwnerUserId":1415817.0,"Title":"Combining Neural Networks Pytorch","Body":"<p>I have <strong>2 images<\/strong> as input, <strong>x1<\/strong> and <strong>x2<\/strong>  and try to use convolution as a <strong>similarity measure<\/strong>. The idea is that the learned weights substitute more traditional measure of similarity (cross correlation, NN, ...). Defining my forward function as follows:<\/p>\n\n<pre><code>def forward(self,x1,x2):\n    out_conv1a = self.conv1(x1)\n    out_conv2a = self.conv2(out_conv1a)\n    out_conv3a = self.conv3(out_conv2a)\n\n    out_conv1b = self.conv1(x2)\n    out_conv2b = self.conv2(out_conv1b)\n    out_conv3b = self.conv3(out_conv2b)\n<\/code><\/pre>\n\n<p>Now for the similarity measure:<\/p>\n\n<pre><code>out_cat = torch.cat([out_conv3a, out_conv3b],dim=1)\nfuther_conv = nn.Conv2d(out_cat)\n<\/code><\/pre>\n\n<p><strong>My question is as follows:<\/strong><\/p>\n\n<p>1) Would Depthwise\/Separable Convolutions as in the google <a href=\"https:\/\/arxiv.org\/pdf\/1706.03059.pdf\" rel=\"nofollow noreferrer\">paper<\/a> yield any advantage over 2d convolution of the concatenated input. For that matter can convolution be a similarity measure, cross correlation and convolution are very similar.<\/p>\n\n<p>2) It is my understanding that the groups=2 option in conv2d would provide 2 separate inputs to train weights with, in this case each of the previous networks weights. How are these combined afterwards?<\/p>\n\n<p>For a basic concept see <a href=\"https:\/\/www.quora.com\/What-are-Siamese-neural-networks-what-applications-are-they-good-for-and-why\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"52621058","CreationDate":"2018-10-03T06:37:46.963","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Using a <code>nn.Conv2d<\/code> layer you assume weights are trainable parameters. However, if you want to filter one feature map with another, you can dive deeper and use <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.functional.conv2d\" rel=\"nofollow noreferrer\"><code>torch.nn.functional.conv2d<\/code><\/a> to explicitly define both input and filter yourself:<\/p>\n\n<pre><code>out = torch.nn.functional.conv2d(out_conv3a, out_conv3b)  \n<\/code><\/pre>\n"}]}
+{"QuestionId":52617765,"AnswerCount":2,"Tags":"<callback><keras>","CreationDate":"2018-10-02T23:32:30.497","AcceptedAnswerId":null,"OwnerUserId":2105525.0,"Title":"Keras get model outputs after each batch","Body":"<p>I'm using a generator to make sequential training data for a hierarchical recurrent model, which needs the outputs of the previous batch to generate the inputs for the next batch. This is a similar situation to the Keras argument <code>stateful=True<\/code> which saves the hidden states for the next batch, except it's more complicated so I can't just use that as-is.<\/p>\n\n<p>So far I tried putting a hack in the loss function:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n    global output_ref\n    output_ref[0] = y_pred[0].eval(session=K.get_session())\n    output_ref[1] = y_pred[1].eval(session=K.get_session())\n<\/code><\/pre>\n\n<p>but that didn't compile and I hope there's a better way. Will Keras callbacks be of any help?<\/p>\n","answers":[{"AnswerId":"58457754","CreationDate":"2019-10-18T20:28:02.240","ParentId":null,"OwnerUserId":"1331851","Title":null,"Body":"<p>I use this in the Tensorflow version of Keras, but it should work in Keras without Tensorflow <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nclass ModelOutput:\n    ''' Class wrapper for a metric that stores the output passed to it '''\n    def __init__(self, name):\n        self.name = name\n        self.y_true = None\n        self.y_pred = None\n\n    def save_output(self, y_true, y_pred):\n        self.y_true = y_true\n        self.y_pred = y_pred\n        return tf.constant(True)\n\nclass ModelOutputCallback(tf.keras.callbacks.Callback):\n  def __init__(self, model_outputs):\n    tf.keras.callbacks.Callback.__init__(self)\n    self.model_outputs = model_outputs\n\n  def on_train_batch_end(self, batch, logs=None):\n    #use self.model_outputs to get the outputs here\n\nmodel_outputs = [\n                ModelOutput('rbox_score_map'),\n                ModelOutput('rbox_shapes'),\n                ModelOutput('rbox_angles')\n            ]\n\n# Note the extra [] around m.save_output, this example is for a model with \n# 3 outputs, metrics must be a list of lists if you type it out\nmodel.compile( ..., metrics=[[m.save_output] for m in self.model_outputs])\n\nmodel.fit(..., callbacks=[ModelOutputCallback(model_outputs)])\n\n<\/code><\/pre>\n"},{"AnswerId":"53983523","CreationDate":"2018-12-31T04:16:05.113","ParentId":null,"OwnerUserId":"10850699","Title":null,"Body":"<p>Learned from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3469\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><code>model.compile(optimizer='adam')\n# hack after compile\noutput_layers = [ 'gru' ]\ns_name = 's'\nmodel.metrics_names += [s_name]\nmodel.metrics_tensors += [layer.output for layer in model.layers if layer.name in output_layers]\n\nclass my_callback(Callback):\n    def on_batch_end(self, batch, logs=None):\n        s_pred = logs[s_name]\n        print('s_pred:', s_pred)\n        return\n\nmodel.fit(..., callbacks=[my_callback()])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52618644,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-03T01:50:14.790","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Missing clear_session, set_session, and get_session?","Body":"<p>I'm using Keras 2.2.0 and am trying to do something like the following:<\/p>\n\n<pre><code>import keras.backend as K\nK.clear_session()\nsess = tf.Session()\nK.set_session(sess)\n...\nwith K.get_session() as sess:\n<\/code><\/pre>\n\n<p>However, I get errors saying <code>AttributeError: 'module' object has no attribute 'clear_session'<\/code>.  So it seems this functionality is no longer in <code>keras.backend<\/code>?<\/p>\n\n<p>For instance, if I do <code>dir(keras.backend)<\/code>, I get:<\/p>\n\n<pre><code>['Function', 'NAME_SCOPE_STACK', 'Print', 'RandomStreams', 'T', 'T_softsign', '_BACKEND', '__builtins__', '__doc__', '__file__', '__name__', '__package__', '__path__', '_backend', '_config', '_config_path', '_epsilon', '_floatx', '_image_data_format', '_keras_base_dir', '_keras_dir', 'abs', 'absolute_import', 'all', 'any', 'arange', 'argmax', 'argmin', 'backend', 'batch_dot', 'batch_flatten', 'batch_get_value', 'batch_normalization', 'batch_set_value', 'bias_add', 'binary_crossentropy', 'cast', 'cast_to_floatx', 'categorical_crossentropy', 'clip', 'common', 'concatenate', 'constant', 'contextmanager', 'conv1d', 'conv2d', 'conv2d_transpose', 'conv3d', 'conv3d_transpose', 'cos', 'count_params', 'ctc_batch_cost', 'ctc_cost', 'ctc_create_skip_idxs', 'ctc_interleave_blanks', 'ctc_path_probs', 'ctc_update_log_p', 'cumprod', 'cumsum', 'defaultdict', 'depthwise_conv2d', 'division', 'dot', 'dropout', 'dtype', 'elu', 'epsilon', 'equal', 'eval', 'exp', 'expand_dims', 'eye', 'f', 'flatten', 'floatx', 'foldl', 'foldr', 'function', 'gather', 'get_uid', 'get_value', 'get_variable_shape', 'gradients', 'greater', 'greater_equal', 'hard_sigmoid', 'has_arg', 'identity', 'ifelse', 'image_data_format', 'image_dim_ordering', 'importlib', 'in_test_phase', 'in_top_k', 'in_train_phase', 'int_shape', 'is_keras_tensor', 'is_placeholder', 'is_sparse', 'is_tensor', 'json', 'l2_normalize', 'learning_phase', 'less', 'less_equal', 'local_conv1d', 'local_conv2d', 'log', 'logsumexp', 'map_fn', 'max', 'maximum', 'mean', 'min', 'minimum', 'moving_average_update', 'name_scope', 'ndim', 'normalize_batch_in_training', 'not_equal', 'np', 'one_hot', 'ones', 'ones_like', 'os', 'pattern_broadcast', 'permute_dimensions', 'placeholder', 'pool', 'pool2d', 'pool3d', 'pow', 'print_function', 'print_tensor', 'prod', 'py_all', 'py_any', 'py_slice', 'py_sum', 'random_binomial', 'random_normal', 'random_normal_variable', 'random_uniform', 'random_uniform_variable', 'relu', 'repeat', 'repeat_elements', 'reset_uids', 'reshape', 'resize_images', 'resize_volumes', 'reverse', 'rnn', 'round', 'separable_conv1d', 'separable_conv2d', 'set_epsilon', 'set_floatx', 'set_image_data_format', 'set_image_dim_ordering', 'set_learning_phase', 'set_value', 'shape', 'sigmoid', 'sign', 'sin', 'slice', 'softmax', 'softplus', 'softsign', 'sparse_categorical_crossentropy', 'spatial_2d_padding', 'spatial_3d_padding', 'sqrt', 'square', 'squeeze', 'stack', 'std', 'stop_gradient', 'sum', 'switch', 'sys', 'tanh', 'temporal_padding', 'th_sparse_module', 'theano', 'theano_backend', 'tile', 'to_dense', 'transpose', 'truncated_normal', 'update', 'update_add', 'update_sub', 'var', 'variable', 'zeros', 'zeros_like']\n<\/code><\/pre>\n\n<p>and don't see any of those 3 in there.  <\/p>\n\n<p>How should I be writing this code in modern Keras?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"56351698","CreationDate":"2019-05-29T00:44:56.183","ParentId":null,"OwnerUserId":"8495866","Title":null,"Body":"<p>EDIT: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/11015\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/11015<\/a>\nSeems like it is not available any may have to downgrade <\/p>\n\n<p>It might be that your backend is set to using Theano (I believe clear_session is only available through the Tensorflow backend with Keras). You can change these settings in your keras.json to TF and clear_session should be available to you.<\/p>\n"}]}
+{"QuestionId":52618700,"AnswerCount":1,"Tags":"<python><keras><mnist>","CreationDate":"2018-10-03T01:59:06.833","AcceptedAnswerId":52618746.0,"OwnerUserId":4616611.0,"Title":"How to sort MNIST digits into each class label?","Body":"<p>I'm importing mnist dataset from Keras using <code>(x_train, y_train), (x_test, y_test) = mnist.load_data()<\/code> and what I want to do is sort each sample by it's corresponding digit. I'm imagining some trivial way to do this but I can't seem to find any label attribute of the data. Any simple way to do this? <\/p>\n","answers":[{"AnswerId":"52618746","CreationDate":"2018-10-03T02:06:57.310","ParentId":null,"OwnerUserId":"540259","Title":null,"Body":"<p><code>y_train<\/code> and <code>y_test<\/code> are the vectors containing the label associated with each image in x_train and x_test respectively. That will tell you the digit shown in each image. So just get the indices that will sort these vectors using <code>np.argsort<\/code> and then use these indices to re-order the corresponding matrix. <\/p>\n\n<pre><code>import numpy as np\n\nidx = np.argsort(y_train)\nx_train_sorted = x_train[idx]\ny_train_sorted = y_train[idx]\n<\/code><\/pre>\n\n<p>So if you want all the images for a particular digit, you can simply grab them by indexing the corresponding matrix<\/p>\n\n<pre><code>x_train_zeros = x_train[y_train == 0]\nx_train_ones = x_train[y_train == 1]\n# and so on...\n<\/code><\/pre>\n\n<p>Notice that in this case you don't need to pre-sort the data.<\/p>\n"}]}
+{"QuestionId":52619166,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><vgg-net>","CreationDate":"2018-10-03T03:08:17.903","AcceptedAnswerId":52619418.0,"OwnerUserId":2701933.0,"Title":"Issue with transfer learning with Tensorflow and Keras","Body":"<p>I've been trying to recreate the work done in <a href=\"https:\/\/medium.com\/tensorflow\/neural-style-transfer-creating-art-with-deep-learning-using-tf-keras-and-eager-execution-7d541ac31398\" rel=\"nofollow noreferrer\">this blog post<\/a>. The writeup is very comprehensive and <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/research\/nst_blogpost\/4_Neural_Style_Transfer_with_Eager_Execution.ipynb#scrollTo=ES9dC6ZyJBD2\" rel=\"nofollow noreferrer\">code<\/a> is shared via a collab.<\/p>\n\n<p>What I'm trying to do is extract layers from the pretrained VGG19 network and create a new network with these layers as the outputs. However, when I assemble the new network, it highly resembles the VGG19 network and seems to contain layers that I didn't extract. An example is below.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.keras import models\n\n## Create network based on VGG19 arch with pretrained weights\nvgg = tf.keras.applications.vgg19.VGG19(include_top=False, weights='imagenet')\nvgg.trainable = False\n<\/code><\/pre>\n\n<p>When we look at the summary of the VGG19, we see an architecture that we expect.<\/p>\n\n<pre><code>vgg.summary()\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_2 (InputLayer)         (None, None, None, 3)     0         \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, None, None, 64)    1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, None, None, 64)    36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, None, None, 64)    0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, None, None, 128)   73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, None, None, 128)   147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, None, None, 128)   0         \n_________________________________________________________________\nblock3_conv1 (Conv2D)        (None, None, None, 256)   295168    \n_________________________________________________________________\nblock3_conv2 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_conv3 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_conv4 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_pool (MaxPooling2D)   (None, None, None, 256)   0         \n_________________________________________________________________\nblock4_conv1 (Conv2D)        (None, None, None, 512)   1180160   \n_________________________________________________________________\nblock4_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_conv3 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_conv4 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_pool (MaxPooling2D)   (None, None, None, 512)   0         \n_________________________________________________________________\nblock5_conv1 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock5_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock5_conv3 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock5_conv4 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock5_pool (MaxPooling2D)   (None, None, None, 512)   0         \n=================================================================\nTotal params: 20,024,384\nTrainable params: 0\nNon-trainable params: 20,024,384\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Then, we extract the layers and create a new model<\/p>\n\n<pre><code>## Layers to extract\ncontent_layers = ['block5_conv2'] \nstyle_layers = ['block1_conv1','block2_conv1','block3_conv1','block4_conv1','block5_conv1']\n## Get output layers corresponding to style and content layers \nstyle_outputs = [vgg.get_layer(name).output for name in style_layers]\ncontent_outputs = [vgg.get_layer(name).output for name in content_layers]\nmodel_outputs = style_outputs + content_outputs\n\nnew_model = models.Model(vgg.input, model_outputs)\n<\/code><\/pre>\n\n<p>When <code>new_model<\/code> is created, I believe we should have a much smaller model. However, a summary on the model shows that the new model is very close to the original model (it contains 19 of the 22 layers from VGG19) and it contains layers that we didn't extract.<\/p>\n\n<pre><code>new_model.summary()\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_2 (InputLayer)         (None, None, None, 3)     0         \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, None, None, 64)    1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, None, None, 64)    36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, None, None, 64)    0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, None, None, 128)   73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, None, None, 128)   147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, None, None, 128)   0         \n_________________________________________________________________\nblock3_conv1 (Conv2D)        (None, None, None, 256)   295168    \n_________________________________________________________________\nblock3_conv2 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_conv3 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_conv4 (Conv2D)        (None, None, None, 256)   590080    \n_________________________________________________________________\nblock3_pool (MaxPooling2D)   (None, None, None, 256)   0         \n_________________________________________________________________\nblock4_conv1 (Conv2D)        (None, None, None, 512)   1180160   \n_________________________________________________________________\nblock4_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_conv3 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_conv4 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock4_pool (MaxPooling2D)   (None, None, None, 512)   0         \n_________________________________________________________________\nblock5_conv1 (Conv2D)        (None, None, None, 512)   2359808   \n_________________________________________________________________\nblock5_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n=================================================================\nTotal params: 15,304,768\nTrainable params: 15,304,768\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>So my questions are...<\/p>\n\n<ol>\n<li>Why are layers that I didn't extract showing up in <code>new_model<\/code>. Are these being inferred by the model's instantiation process <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">per the docs<\/a>? This just seems too close to the VGG19 architecture to be inferred.<\/li>\n<li>From how I understand Keras' <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">Model (functional API)<\/a>, passing multiple output layers should create a model with multiple outputs, however, it seems that the new model is sequential and only has a single output layer. Is this the case?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"52619418","CreationDate":"2018-10-03T03:47:37.537","ParentId":null,"OwnerUserId":"826970","Title":null,"Body":"<blockquote>\n  <ol>\n  <li>Why are layers that I didn't extract showing up in <code>new_model<\/code>.<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>That's because when you create a model with <code>models.Model(vgg.input, model_outputs)<\/code> the \"intermediate\" layers between <code>vgg.input<\/code> and the output layers are included as well. This is the intended way as VGG is constructed this way. <\/p>\n\n<p>For example if you were to create a model this way: <code>models.Model(vgg.input, vgg.get_layer('block2_pool')<\/code> every intermediate layer between the <code>input_1<\/code> and <code>block2_pool<\/code> would be included since the input <strong>has to<\/strong> flow through them before reaching <code>block2_pool<\/code>. Below is a partial graph of VGG that could help with that.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/B4ICp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/B4ICp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now, -if I've not misunderstood- if you want to create a model that doesn't include those intermediate layers (which would probably work poorly), you have to create one yourself. Functional API is very useful on this. There are examples on the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">documentation<\/a> but the gist of what you want to do is as below:<\/p>\n\n<pre><code>from keras.layers import Conv2D, Input\n\nx_input = Input(shape=(28, 28, 1,))\nblock1_conv1 = Conv2D(64, (3, 3), padding='same')(x_input)\nblock2_conv2 = Conv2D(128, (3, 3), padding='same')(x_input)\n...\n\nnew_model = models.Model(x_input, [block1_conv1, block2_conv2, ...])\n<\/code><\/pre>\n\n<blockquote>\n  <ol start=\"2\">\n  <li>... however, it seems that the new model is sequential and only has a single output layer. Is this the case?<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>No, your model has multiple outputs as you intended to. <code>model.summary()<\/code> should have display which layers are connected to what (which would help understanding the structure), but I believe there is a small bug with some versions that prevents that. In any case you can see that your model have multiple outputs by checking <code>new_model.output<\/code>, that should give you:<\/p>\n\n<pre><code>[&lt;tf.Tensor 'block1_conv1\/Relu:0' shape=(?, ?, ?, 64) dtype=float32&gt;,\n &lt;tf.Tensor 'block2_conv1\/Relu:0' shape=(?, ?, ?, 128) dtype=float32&gt;,\n &lt;tf.Tensor 'block3_conv1\/Relu:0' shape=(?, ?, ?, 256) dtype=float32&gt;,\n &lt;tf.Tensor 'block4_conv1\/Relu:0' shape=(?, ?, ?, 512) dtype=float32&gt;,\n &lt;tf.Tensor 'block5_conv1\/Relu:0' shape=(?, ?, ?, 512) dtype=float32&gt;,\n &lt;tf.Tensor 'block5_conv2\/Relu:0' shape=(?, ?, ?, 512) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>Printing it sequentially in <code>new_model.summary()<\/code> is just a design choice, as it would get hairy with complex models.<\/p>\n"}]}
+{"QuestionId":52619535,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><rnn>","CreationDate":"2018-10-03T04:03:37.730","AcceptedAnswerId":null,"OwnerUserId":4293811.0,"Title":"keras RNN with characters input fails to predict","Body":"<p>I am working on an RNN that must predict a number from a string.\nBasically, I have a list of strings, and a float is associated to each string. The strings, which are not words and may seem random, are in fact ordered and have logic. Thus the associated float is completely predictable.<\/p>\n\n<p>My strategy for this problem is to convert each character of the string in a number, with <code>ord(character)<\/code> and put all the strings in a NumPy array.<\/p>\n\n<p>As a first test, I created a very easy set of strings, only containing \"a\" and \"b\" characters. The logic is very simple. The more there are \"a\" in the string, the higher the associated number. Unfortunately, the RNN completely fails.<\/p>\n\n<p><strong>Training data preparation:<\/strong><\/p>\n\n<pre><code>a = \"abbbbb\"\nb = \"aabbbb\"\nc = \"aaabbb\"\nd = \"aaaabb\"\ne = \"aaaaab\"\nf = \"aaaaaa\"\n\nX_train = []\nX_train.append([ord(letter) for letter in a]) # Converts \"abbbbb\" to [97, 98, 98, 98, 98, 98]\nX_train.append([ord(letter) for letter in b])\nX_train.append([ord(letter) for letter in c])\nX_train.append([ord(letter) for letter in d])\nX_train.append([ord(letter) for letter in e])\nX_train.append([ord(letter) for letter in f])\n\nX_train = np.array(X_train)\nX_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[1], 1))\n\ny_train = [0, 5, 15, 30, 50, 80]\n<\/code><\/pre>\n\n<p><strong>RNN implementation and training<\/strong><\/p>\n\n<p>This is one of many RNN I tried. I tried changing the number of layers, the number of units per layer, and the Dropout. Without success.<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, LSTM, Dropout\nimport math\nfrom sklearn.metrics import mean_squared_error\n\nregressor = Sequential()\n\nregressor.add(LSTM(units=256, return_sequences=True, input_shape=(X_train.shape[1], 1)))\nregressor.add(Dropout(0.2))\nregressor.add(LSTM(units=256, return_sequences=True))\nregressor.add(Dropout(0.2))\nregressor.add(LSTM(units=256, return_sequences=True))\nregressor.add(Dropout(0.2))\nregressor.add(LSTM(units=256, return_sequences=True))\nregressor.add(Dropout(0.2))\nregressor.add(LSTM(units=256, return_sequences=True))\nregressor.add(Dropout(0.2))\nregressor.add(LSTM(units=256, return_sequences=False))\nregressor.add(Dropout(0.2))\nregressor.add(Dense(units=1))\n\nregressor.compile(optimizer = \"adam\", loss = \"mean_squared_error\")\n\nregressor.fit(X_train, y_train, epochs=100, batch_size=32)\n<\/code><\/pre>\n\n<p><strong>Test data preparation and prediction<\/strong><\/p>\n\n<pre><code>X_test = list(ord(letter) for letter in f)\nX_test = np.array(X_test)\nX_test = np.reshape(X_test, (1, X_test.shape[0], 1))\n\nprediction = regressor.predict(X_test)\n<\/code><\/pre>\n\n<p><strong>THE PROBLEM<\/strong><\/p>\n\n<p>I tested the RNN with different strings, even with the 6 strings from the training set. The RNN prediction is always the same, 30.0094, no matter what is the input. The two last decimals can vary, but basically, the output will always be 30.00xx.<\/p>\n\n<p>The number 30 is not any number since it is the mean of the y_train list, [0, 5, 15, 30, 50, 80]<\/p>\n\n<p>I tried with different y_train lists, I always obtain the mean...<\/p>\n\n<p>Any idea on what is going on ?<\/p>\n","answers":[{"AnswerId":"52619806","CreationDate":"2018-10-03T04:41:26.300","ParentId":null,"OwnerUserId":"7499583","Title":null,"Body":"<p>Can you try with a way simple architecture?<\/p>\n\n<pre><code>regressor = Sequential()\n\nregressor.add(LSTM(units=1, return_sequences=True, input_shape=(X_train.shape[1], 1)))\nregressor.compile(optimizer = \"adam\", loss = \"mean_squared_error\")  \nregressor.fit(X_train, y_train, epochs=5, batch_size=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52619697,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><artificial-intelligence>","CreationDate":"2018-10-03T04:27:30.127","AcceptedAnswerId":null,"OwnerUserId":883418.0,"Title":"Any advantage to checkpoint model very frequently","Body":"<p>I'm currently learning Tensorflow, and found checkpointing interesting. <\/p>\n\n<p>I'm wondering if you think there is any advantage if we can checkpoint the model literally every step? Would that help debugging and finding the best model?<\/p>\n","answers":[{"AnswerId":"52638161","CreationDate":"2018-10-04T02:34:34.853","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>You want to save the best checkpoint in general. As a model evolves quite slowly during training every X training steps you can validate your model and then save it if it is better than the previous one.<\/p>\n"}]}
+{"QuestionId":52620141,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-03T05:19:49.250","AcceptedAnswerId":null,"OwnerUserId":9768291.0,"Title":"Keras pre-trained model switching to varying input size","Body":"<p>Very much similar to <a href=\"https:\/\/stackoverflow.com\/a\/41092113\/9768291\">this question<\/a> except I am wondering how I could take my <strong>pre-trained model<\/strong> which had an input size of <code>(128, 128, 3)<\/code> images, keep its weights, and use it to <code>predict<\/code> on images of varying input size.<\/p>\n\n<p>I get this, as it is, when I try to input an image of arbitrary size:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"arg_test.py\", line 127, in &lt;module&gt;\n    predict(args)\n  File \"arg_test.py\", line 71, in predict\n    predictions.append(model.predict(input_img)[0])  # returns a list of lists, one for each image in the batch\n  File \"C:\\Users\\payne\\Anaconda3\\envs\\ml-gpu\\lib\\site-packages\\keras\\engine\\training.py\", line 1147, in predict\n    x, _, _ = self._standardize_user_data(x)\n  File \"C:\\Users\\payne\\Anaconda3\\envs\\ml-gpu\\lib\\site-packages\\keras\\engine\\training.py\", line 749, in _standardize_user_data\n    exception_prefix='input')\n  File \"C:\\Users\\payne\\Anaconda3\\envs\\ml-gpu\\lib\\site-packages\\keras\\engine\\training_utils.py\", line 137, in standardize_input_data\n    str(data_shape))\nValueError: Error when checking input: expected input_1 to have shape (128, 128, 3) but got array with shape (2736, 3648, 3)\n<\/code><\/pre>\n\n<p>Here is <a href=\"https:\/\/github.com\/payne911\/SR-ResCNN-Keras-\/blob\/master\/model.py\" rel=\"nofollow noreferrer\">my model<\/a>:<\/p>\n\n<pre><code>def setUpModel(x_train, y_train):\n    filters = 256\n    kernel_size = 3\n    strides = 1\n\n    # Head module\n    input = Input(shape=(img_height\/\/scale_fact, img_width\/\/scale_fact, img_depth))\n    conv0 = Conv2D(filters, kernel_size, strides=strides, padding='same')(input)\n\n    # Body module\n    res = Conv2D(filters, kernel_size, strides=strides, padding='same')(conv0)\n    act = ReLU()(res)\n    res = Conv2D(filters, kernel_size, strides=strides, padding='same')(act)\n    res_rec = Add()([conv0, res])\n\n    for i in range(res_blocks):\n        res1 = Conv2D(filters, kernel_size, strides=strides, padding='same')(res_rec)\n        act  = ReLU()(res1)\n        res2 = Conv2D(filters, kernel_size, strides=strides, padding='same')(act)\n        res_rec = Add()([res_rec, res2])\n\n    conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(res_rec)\n    add  = Add()([conv0, conv])\n\n    # Tail module\n    conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(add)\n    act  = ReLU()(conv)\n    up   = UpSampling2D(size=scale_fact if scale_fact != 4 else 2)(act)  # TODO: try \"Conv2DTranspose\"\n    # mul = Multiply([np.zeros((img_width,img_height,img_depth)).fill(0.1), up])(up)\n\n    # When it's a 4X factor, we want the upscale split in two procedures\n    if(scale_fact == 4):\n        conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(up)\n        act  = ReLU()(conv)\n        up   = UpSampling2D(size=2)(act)  # TODO: try \"Conv2DTranspose\"\n\n    output = Conv2D(filters=3,\n                    kernel_size=1,\n                    strides=1,\n                    padding='same')(up)\n\n    model = Model(inputs=input, outputs=output)\n<\/code><\/pre>\n\n<p>This was only the architecture of the model that was used during training, but tha training is behind: I have my <code>model.h5<\/code> file obtained through <code>model.save()<\/code>.<\/p>\n\n<p>Here is how I get predictions:<\/p>\n\n<pre><code>import argparse\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport skimage.io\n\nfrom keras.models import load_model\nfrom keras.optimizers import Adam\nfrom keras.optimizers import Adadelta\n\nfrom constants import save_dir\nfrom constants import model_name\nfrom constants import crops_p_img\nfrom constants import tests_path\nfrom constants import img_height\nfrom constants import img_width\nfrom constants import scale_fact\nfrom utils import float_im\nfrom utils import crop_center\n\n\nparser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n\nparser.add_argument('-a', '--amount', type=int, default=crops_p_img,\n                    help='how many (cropped to 128x128) samples to predict from within the image')\nparser.add_argument('image', type=str,\n                    help='image name (example: \"bird.png\") that must be inside the \".\/input\/\" folder')\nparser.add_argument('-m', '--model', type=str, default=model_name,\n                    help='model name (in the \".\/save\/\" folder), followed by \".h5\"')\nparser.add_argument('-r', '--random', action=\"store_true\",  # if var is in args, set to TRUE, else, set to FALSE\n                    help='flag that will select a random 128x128 area in the input image instead of the center')\nparser.add_argument('-f', '--full', action=\"store_true\",  # if var is in args, set to TRUE, else, set to FALSE\n                    help='(WIP) flag that will get the whole image to be processed by the network')\n\nargs = parser.parse_args()\n\n\ndef predict(args):\n    model = load_model(save_dir + '\/' + args.model)\n\n    # Setting up the proper optimizer       TODO: needed?\n    if args.model == \"my_full_model.h5\":\n        optimizer = Adadelta(lr=1.0,\n                             rho=0.95,\n                             epsilon=None,\n                             decay=0.0)\n    else:\n        optimizer = Adam(lr=0.001,\n                         beta_1=0.9,\n                         beta_2=0.999,\n                         epsilon=None,\n                         decay=0.0,\n                         amsgrad=False)\n\n    model.compile(optimizer=optimizer,\n                  loss='mean_squared_error')\n\n    image = skimage.io.imread(tests_path + args.image)\n\n    if image.shape[0] == 128:\n        args.amount = 1\n\n    predictions = []\n    images = []\n\n    # TODO: integrate FULL IMAGE\n    # if args.full:\n    #     images.append(image)\n    #     # Hack because GPU can only handle one image at a time\n    #     input_img = (np.expand_dims(images[0], 0))  # Add the image to a batch where it's the only member\n    #     predictions.append(model.predict(input_img)[0])  # returns a list of lists, one for each image in the batch\n    # else:\n    if True:\n        for i in range(args.amount):\n            # Cropping to fit input size\n            if (args.random or args.amount &gt; 1) and image.shape[0] &gt; 128:\n                images.append(random_crop(image))\n            else:\n                images.append(crop_center(image, img_width\/\/scale_fact, img_height\/\/scale_fact))\n\n            input_img = (np.expand_dims(images[i], 0))\n            predictions.append(model.predict(input_img)[0])\n\n    for i in range(len(predictions)):\n        show_pred_output(images[i], predictions[i])\n\n\n# adapted from: https:\/\/stackoverflow.com\/a\/52463034\/9768291\ndef random_crop(img):\n    crop_h, crop_w = img_width\/\/scale_fact, img_height\/\/scale_fact\n    print(\"Shape of input image to crop:\", img.shape[0], img.shape[1])\n\n    if (img.shape[0] &gt;= crop_h) and (img.shape[1] &gt;= crop_w):\n        # Cropping a random part of the image\n        rand_h = np.random.randint(0, img.shape[0]-crop_h)\n        rand_w = np.random.randint(0, img.shape[1]-crop_w)\n        print(\"Random position for the crop:\", rand_h, rand_w)\n        tmp_img = img[rand_h:rand_h+crop_h, rand_w:rand_w+crop_w]\n\n        new_img = float_im(tmp_img)  # From [0,255] to [0.,1.]\n    else:\n        return img\n\n    return new_img\n\n\ndef show_pred_output(input, pred):\n    plt.figure(figsize=(20, 20))\n    plt.suptitle(\"Results\")\n\n    plt.subplot(1, 2, 1)\n    plt.title(\"Input: 128x128\")\n    plt.imshow(input, cmap=plt.cm.binary).axes.get_xaxis().set_visible(False)\n\n    plt.subplot(1, 2, 2)\n    plt.title(\"Output: 512x512\")\n    plt.imshow(pred, cmap=plt.cm.binary).axes.get_xaxis().set_visible(False)\n\n    plt.show()\n\n\nif __name__ == '__main__':\n    print(\"   -  \", args)\n    predict(args)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52620780","CreationDate":"2018-10-03T06:16:51.497","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You should replace this line:<\/p>\n\n<pre><code>input = Input(shape=(None, None, img_depth))\n<\/code><\/pre>\n\n<p><code>None<\/code> in an shape means variable size. Since the model is just convolutions it should work with images of any size.<\/p>\n"}]}
+{"QuestionId":52620964,"AnswerCount":0,"Tags":"<python><keras><hdf5><large-data>","CreationDate":"2018-10-03T06:31:54.950","AcceptedAnswerId":null,"OwnerUserId":5019517.0,"Title":"Input\/output error using fit_generator in Keras","Body":"<p>I have a large training matrix storing images and corresponding classes in h5 format. I want to read the images and their labels for training a model in Keras. Since the dataset is too large, I created my own data generator function <em>imageLoader()<\/em> and used it in <em>model.fit_generator<\/em>. The <em>imageLoader()<\/em> reads the training data in batches of size <em>batch_size<\/em> and use it for training. <\/p>\n\n<p>The training matrix (h5 file) contains data as follows: <\/p>\n\n<p>Train_features: (58160 x 25 x 25 x 50) => no. of images x row x col x channels<\/p>\n\n<p>Train_class: (58160 x 1) => no. of images x label<\/p>\n\n<p>However, I am still getting memory error when using fit_generator in Keras. Following is my code:<\/p>\n\n<pre><code>import os\nos.environ['PYTHONHASHSEED'] = '0'\nimport numpy as np\nnp.random.seed(7)  \nfrom tensorflow import set_random_seed\nset_random_seed(2)\nimport keras\nfrom keras.layers import Input, Dense, Dropout, BatchNormalization\nfrom keras.models import Model, Sequential\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D, Flatten, Activation\nfrom keras.callbacks import ReduceLROnPlateau\nimport h5py\n\n\nh5f = h5py.File('path to h5 file', 'r')\n\n#Data Generator\ndef imageLoader(h5f, batch_size):\n    L= len(h5f['train_class'])\n    while True:\n        batch_start=0\n        batch_end= batch_size\n        while batch_start&lt;L:\n            limit = min(batch_end, L)\n            X= h5f['train_features'][batch_start:limit]\n            Y= h5f['train_class'][batch_start:limit]\n            print(batch_start, batch_end)\n\n            yield (X,keras.utils.to_categorical(Y,2)) #I have two classes\n\n            batch_start +=batch_size\n            batch_end +=batch_size\n\n# Building Model\nmodel= Sequential()\nmodel.add(Conv2D(80,(4,4), activation='relu', input_shape= (25,25,50)))\nmodel.add(MaxPooling2D(pool_size= (2,2)))\nmodel.add(Conv2D(150,(4,4), activation = 'relu'))\nmodel.add(MaxPooling2D(pool_size= (2,2)))\nmodel.add(Conv2D(200,(4,4), activation = 'relu')) \nmodel.add(Flatten())\nmodel.add(Dense(2, activation = 'softmax'))\nmodel.summary()\n\n\nmodel.compile(loss = 'binary_crossentropy', optimizer = 'adam', metrics = \n['accuracy'])\n\nmodel.fit_generator(imageLoader(h5f,100), \nsteps_per_epoch=len(h5f['train_class'])\/100, epochs=3)\n<\/code><\/pre>\n\n<p>Here is the error log:<\/p>\n\n<pre><code>Epoch 1\/3\n0 100\n100 200\n200 300\n300 400\n400 500\n500 600\n600 700\n700 800\n800 900\n900 1000\n1000 1100\n1\/54 [..............................] - ETA: 22s - loss: 0.9183 - acc: \n0.11001100 1200\n1200 1300\n1300 1400\n1400 1500\n5\/54 [=&gt;............................] - ETA: 4s - loss: 0.9579 - acc: 0.0380 \n1500 1600\n7\/54 [==&gt;...........................] - ETA: 42s - loss: 0.8970 - acc: \n0.1429\nException in thread Thread-28:\n\n Traceback (most recent call last):\n File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/threading.py\", line 916, in _bootstrap_inner\n self.run()\n File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/threading.py\", line 864, in run\n self._target(*self._args, **self._kwargs)\n\nFile \"\/home\/saror\/.pyenv\/versions\/anaconda3- 4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/keras\/utils\/data_utils.py\", \nline 579, in data_generator_task\ngenerator_output = next(self._generator)\n  File \"&lt;ipython-input-119-7a1d852e803c&gt;\", line 10, in imageLoader\n    X= h5f['test_features'][batch_start:limit]\n  File \"h5py\/_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\n  File \"h5py\/_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\n  File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/h5py\/_hl\/dataset.py\", line 496, in __getitem__\nself.id.read(mspace, fspace, arr, mtype, dxpl=self._dxpl)\n  File \"h5py\/_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\n  File \"h5py\/_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\n  File \"h5py\/h5d.pyx\", line 181, in h5py.h5d.DatasetID.read\n  File \"h5py\/_proxy.pyx\", line 130, in h5py._proxy.dset_rw\n  File \"h5py\/_proxy.pyx\", line 84, in h5py._proxy.H5PY_H5Dread\nOSError: Can't read data (file read failed: time = Wed Oct  3 23:43:01 2018\n, filename = '\/mnt\/hdd\/shreya\/change_detection_files\/san_test_data.h5', file \ndescriptor = 86, errno = 5, error message = 'Input\/output error', buf = \n0x7f53b2990c80, total read size = 6415424, bytes this sub-read = 6415424, \nbytes actually read = 18446744073709551615, offset = 418586624)\n\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-120-e74f6fee0af9&gt;\", line 15, in &lt;module&gt;\n    model.fit_generator(imageLoader(h5f,100), \nsteps_per_epoch=len(h5f['test_class'])\/100, epochs=3)\n\n  File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py\", line 87, in wrapper\nreturn func(*args, **kwargs)\n\n  File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/keras\/models.py\", line 1156, in fit_generator\ninitial_epoch=initial_epoch)\n\n  File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py\", line 87, in wrapper\nreturn func(*args, **kwargs)\n\n  File \"\/home\/saror\/.pyenv\/versions\/anaconda3-4.4.0\/envs\/py3dl2\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 2046, in fit_generator\ngenerator_output = next(output_generator)\n\nStopIteration   \n<\/code><\/pre>\n\n<p>Could anyone please tell what is wrong in the above code? How can I read large data efficiently for model training?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":52621257,"AnswerCount":1,"Tags":"<python><keras><lstm>","CreationDate":"2018-10-03T06:51:31.197","AcceptedAnswerId":52624702.0,"OwnerUserId":3647970.0,"Title":"Understanding apparently different Keras LSTM API calls","Body":"<p>From the <a href=\"https:\/\/keras.io\/layers\/recurrent\/#lstm\" rel=\"nofollow noreferrer\">official documentation<\/a> i don't see any arguements like <code>batch_size<\/code> and <code>input_shape<\/code> for LSTM. However i have seen declarations like these <code>model.add(LSTM(batch_size, input_shape=(time_steps, features)))<\/code> in <a href=\"https:\/\/medium.freecodecamp.org\/applied-introduction-to-lstms-for-text-generation-380158b29fb3\" rel=\"nofollow noreferrer\">this<\/a> medium article as well as this <a href=\"https:\/\/stackoverflow.com\/questions\/38714959\/understanding-keras-lstms\">SO post<\/a> which uses <code>model.add(LSTM(4, batch_input_shape=(batch_size, look_back, 1), stateful=True))<\/code>. <\/p>\n\n<p>Q1) Can someone please elucidate how is that possible?<br>\nQ2) Another thing i am not able to understand is the concept of <code>batch_size<\/code> especially for text analysis. If i want my model to learn from sentences then should i use a batch_size of 1 referring to 1 sentence sample? In general the arguments passed to LSTMs don't seem definitive. Is there a proper guide to using LSTMs correctly?<\/p>\n","answers":[{"AnswerId":"52624702","CreationDate":"2018-10-03T10:11:36.227","ParentId":null,"OwnerUserId":"2276831","Title":null,"Body":"<p>In your question, you mentioned the use of<\/p>\n\n<pre><code>model.add(LSTM(batch_size, input_shape=(time_steps, features)))\n<\/code><\/pre>\n\n<p>I think the author mislead you as <code>batch_size<\/code> is the incorrect term to be used here. As you can see in the Keras API spec, the first parameter defines the number of hidden states\/ units in this 'layer'. I looked at the medium post you linked and I believe what likely happened is that the <code>batch_size<\/code> was equal to the number of units, and the author was lazy\/ uniformed and decided to use the same constant for both. There has been no change to Keras since May of this year (when the post was written) that can explain their mistake.<\/p>\n\n<p>As for the SO post, <code>batch_input_shape<\/code> is only applicable for stateful LSTM layers. From the documentation:<\/p>\n\n<blockquote>\n  <p>You can set RNN layers to be 'stateful', which means that the states\n  computed for the samples in one batch will be reused as initial states\n  for the samples in the next batch. This assumes a one-to-one mapping\n  between samples in different successive batches.<\/p>\n  \n  <p>To enable statefulness: - specify <code>stateful=True<\/code> in the layer\n  constructor. - specify a fixed batch size for your model, by passing\n  if sequential model: <code>batch_input_shape=(...)<\/code> to the first layer in\n  your model. else for functional model with 1 or more Input layers:\n  <code>batch_shape=(...)<\/code> to all the first layers in your model. This is the\n  expected shape of your inputs including the batch size. It should be a\n  tuple of integers, e.g. <code>(32, 10, 100)<\/code>. - specify  <code>shuffle=False<\/code> when\n  calling <code>fit()<\/code>.<\/p>\n<\/blockquote>\n\n<p>As for your question regarding the nature of <code>batch_size<\/code> in text analysis, it simply refers to the number of samples to propagate through the network. Therefore, if you wanted to pass only a single sentence at a time, you could set it to 1. The problem with this is that your gradient estimation will be less accurate. If you can (memory constraints come into play here) you should use a larger <code>batch_size<\/code>.<\/p>\n"}]}
+{"QuestionId":52621570,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-10-03T07:10:11.663","AcceptedAnswerId":52621807.0,"OwnerUserId":2530674.0,"Title":"pytorch out of GPU memory","Body":"<p>I am trying to implement Yolo-v2 in pytorch. However, I seem to be running out of memory just passing data through the network. The model is large and is shown below. However, I feel like I'm doing something stupid here with my network (like not freeing memory somewhere). The network works as expected on cpu.<\/p>\n\n<p>The test code (where memory runs out) is:<\/p>\n\n<pre><code>x = torch.rand(32,3,416, 416).cuda()\nmodel = Yolov2().cuda()\ny = model(x.float())\n<\/code><\/pre>\n\n<p><strong>Question<\/strong><\/p>\n\n<ol>\n<li>Anything that is obviously wrong with my model?<\/li>\n<li>How do I make this more efficient with memory?<\/li>\n<li>Other comments?<\/li>\n<\/ol>\n\n<p>The model:<\/p>\n\n<pre><code>import torch\nfrom torch import nn\nimport torch.nn.functional as F\n\nclass Yolov2(nn.Module):\n\n    def __init__(self):\n        super(Yolov2, self).__init__()\n\n        self.conv1 = nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm1 = nn.BatchNorm2d(32)\n\n        self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm2 = nn.BatchNorm2d(64)\n\n        self.conv3 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm3 = nn.BatchNorm2d(128)\n        self.conv4 = nn.Conv2d(in_channels=128, out_channels=64, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm4 = nn.BatchNorm2d(64)\n        self.conv5 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm5 = nn.BatchNorm2d(128)\n\n        self.conv6 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm6 = nn.BatchNorm2d(256)\n        self.conv7 = nn.Conv2d(in_channels=256, out_channels=128, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm7 = nn.BatchNorm2d(128)\n        self.conv8 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm8 = nn.BatchNorm2d(256)\n\n        self.conv9 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm9 = nn.BatchNorm2d(512)\n        self.conv10 = nn.Conv2d(in_channels=512, out_channels=256, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm10 = nn.BatchNorm2d(256)\n        self.conv11 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm11 = nn.BatchNorm2d(512)\n        self.conv12 = nn.Conv2d(in_channels=512, out_channels=256, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm12 = nn.BatchNorm2d(256)\n        self.conv13 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm13 = nn.BatchNorm2d(512)\n\n        self.conv14 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm14 = nn.BatchNorm2d(1024)\n        self.conv15 = nn.Conv2d(in_channels=1024, out_channels=512, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm15 = nn.BatchNorm2d(512)\n        self.conv16 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm16 = nn.BatchNorm2d(1024)\n        self.conv17 = nn.Conv2d(in_channels=1024, out_channels=512, kernel_size=1, stride=1, padding=0, bias=False)\n        self.batchnorm17 = nn.BatchNorm2d(512)\n        self.conv18 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm18 = nn.BatchNorm2d(1024)\n\n        self.conv19 = nn.Conv2d(in_channels=1024, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm19 = nn.BatchNorm2d(1024)\n        self.conv20 = nn.Conv2d(in_channels=1024, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm20 = nn.BatchNorm2d(1024)\n\n        self.conv21 = nn.Conv2d(in_channels=3072, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)\n        self.batchnorm21 = nn.BatchNorm2d(1024)\n\n        self.conv22 = nn.Conv2d(in_channels=1024, out_channels=125, kernel_size=1, stride=1, padding=0)\n\n    def reorg_layer(self, x):\n        stride = 2\n        batch_size, channels, height, width = x.size()\n        new_ht = int(height\/stride)\n        new_wd = int(width\/stride)\n        new_channels = channels * stride * stride\n\n#         from IPython.core.debugger import Tracer; Tracer()()\n        passthrough = x.permute(0, 2, 3, 1)\n        passthrough = passthrough.contiguous().view(-1, new_ht, stride, new_wd, stride, channels)\n        passthrough = passthrough.permute(0, 1, 3, 2, 4, 5)\n        passthrough = passthrough.contiguous().view(-1, new_ht, new_wd, new_channels)\n        passthrough = passthrough.permute(0, 3, 1, 2)\n        return passthrough\n\n    def forward(self, x):\n        out = F.max_pool2d(F.leaky_relu(self.batchnorm1(self.conv1(x)), negative_slope=0.1), 2, stride=2)\n        out = F.max_pool2d(F.leaky_relu(self.batchnorm2(self.conv2(out)), negative_slope=0.1), 2, stride=2)\n\n        out = F.leaky_relu(self.batchnorm3(self.conv3(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm4(self.conv4(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm5(self.conv5(out)), negative_slope=0.1)\n        out = F.max_pool2d(out, 2, stride=2)\n\n        out = F.leaky_relu(self.batchnorm6(self.conv6(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm7(self.conv7(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm8(self.conv8(out)), negative_slope=0.1)\n        out = F.max_pool2d(out, 2, stride=2)\n\n        out = F.leaky_relu(self.batchnorm9(self.conv9(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm10(self.conv10(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm11(self.conv11(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm12(self.conv12(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm13(self.conv13(out)), negative_slope=0.1)\n#         from IPython.core.debugger import Tracer; Tracer()()\n        passthrough = self.reorg_layer(out)\n        out = F.max_pool2d(out, 2, stride=2)\n\n        out = F.leaky_relu(self.batchnorm14(self.conv14(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm15(self.conv15(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm16(self.conv16(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm17(self.conv17(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm18(self.conv18(out)), negative_slope=0.1)\n\n        out = F.leaky_relu(self.batchnorm19(self.conv19(out)), negative_slope=0.1)\n        out = F.leaky_relu(self.batchnorm20(self.conv20(out)), negative_slope=0.1)\n\n        out = torch.cat([passthrough, out], 1)\n        out = F.leaky_relu(self.batchnorm21(self.conv21(out)), negative_slope=0.1)\n        out = self.conv22(out)\n\n        return out\n<\/code><\/pre>\n\n<h2>Additional Info:<\/h2>\n\n<ul>\n<li>Torch version is <code>'0.4.1.post2'<\/code><\/li>\n<li>Running on aws p2.xlarge (limit 12gb GPU memory).<\/li>\n<li>The number of parameters for this model is 67137565. Which would occupy &lt;500MB.<\/li>\n<li><a href=\"https:\/\/discuss.pytorch.org\/t\/cuda-out-of-memory\/449\" rel=\"nofollow noreferrer\">this thread from pytorch<\/a> might be related. <\/li>\n<\/ul>\n","answers":[{"AnswerId":"56978258","CreationDate":"2019-07-10T20:35:21.837","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Here is the cycle you may try out:<\/p>\n\n<pre><code>ma = torch.cuda.memory_allocated()\nprint(ma)\n\nmc = torch.cuda.memory_cached()\nprint(mc)\n\ntorch.cuda.empty_cache()\n\nma = torch.cuda.memory_allocated()\nprint(ma)\n\nmc = torch.cuda.memory_cached()\nprint(mc)\n\n# 653475518\n# 952107008\n# 383533568\n# 385875968\n<\/code><\/pre>\n\n<p>See how memory can be freed. The other technique is called on demand loading of batches to GPU. So not the whole dataset, but a single batch or several batches at once.<\/p>\n\n<p>Loading a single batch into GPU you basically measure in \u03bc seconds, while tensor operations like matrix multiply are m seconds.<\/p>\n"},{"AnswerId":"52621807","CreationDate":"2018-10-03T07:23:04.033","ParentId":null,"OwnerUserId":"8868801","Title":null,"Body":"<p>I would try to use smaller batch sizes. Start from 1 and then check what is your maximum.\nI can also try to reduce your input Tensor dimensions.\nYour network is not so small for your GPU <\/p>\n"}]}
+{"QuestionId":52621627,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><deployment><google-cloud-platform><google-bigquery>","CreationDate":"2018-10-03T07:13:16.120","AcceptedAnswerId":null,"OwnerUserId":8531952.0,"Title":"Tensorflow Model Deployment in GCP without Tensorflow Serving","Body":"<blockquote>\n  <p>Machine Learning Model: Tensorflow Based (version 1.9) &amp; Python version 3.6<\/p>\n  \n  <p>Data Input: From Bigquery<\/p>\n  \n  <p>Data Output: To Bigquery<\/p>\n  \n  <p>Production prediction frequency: Monthly<\/p>\n<\/blockquote>\n\n<p>I have a developed a Tensorflow based machine learning model. I have trained it locally and want to deploy it in Google Cloud Platform for predictions.<\/p>\n\n<p>The model reads input data from Google Bigquery and the output predictions has to be written in Google Bigquery. There are some data preparation scripts which has to be run before the model prediction is run. Currently I cannot use BigQuery ML in Production as it is in Beta stage. Additionally as it is a batch prediction I don't think Tensorflow Serving will be a good choice.<\/p>\n\n<p>Strategies which I have tried for deployment:<\/p>\n\n<ol>\n<li><p><strong>Use Google ML Engine for prediction:<\/strong> This approach creates output part files on GCS. These have to be combined and written to Google Bigquery. So in this approach I have to spin up a VM just to execute the data preparation script and ML Engine output to Google Bigquery script. This adds up to 24x7 cost of VM just for running two scripts in a month.<\/p><\/li>\n<li><p><strong>Use Dataflow for data preparation script execution along with Google ML Engine<\/strong>: Dataflow uses python 2.7 while the model is developed in Tensorflow version 1.9 and python version 3.6. So this approach cannot be used.<\/p><\/li>\n<li><p><strong>Google App Engine<\/strong>: Using this approach a complete web application has to be developed in order to serve predictions. As the predictions are in batch this approach is not suitable. Additionally flask\/django has to be integrated with the code in order to use it.<\/p><\/li>\n<li><p><strong>Google Compute Engine<\/strong>: Using this approach the VM would be running 24x7 just for running monthly predictions and running two scripts. The would cause a lot of cost overhead.<\/p><\/li>\n<\/ol>\n\n<p>I would like to know what is best deployment approach for Tensorflow models which has some pre and post processing scripts.<\/p>\n","answers":[{"AnswerId":"56247666","CreationDate":"2019-05-21T23:55:08.463","ParentId":null,"OwnerUserId":"9687748","Title":null,"Body":"<p>Regarding the option 3, Dataflow can read from BigQuery and store the prepared data in BigQuery at the end of the job.<\/p>\n\n<p>Then you can have Tensorflow use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/cloud\/BigQueryReader\" rel=\"nofollow noreferrer\"><code>BigQueryReader<\/code><\/a> to data from BigQuery.<\/p>\n\n<p>Another that you can use is <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/distributed-tensorflow-mnist-cloud-datalab\" rel=\"nofollow noreferrer\">Datalab<\/a>, this is a notebook in which you can prepare your data and then use it for your prediction. <\/p>\n"},{"AnswerId":"56861715","CreationDate":"2019-07-03T01:14:13.160","ParentId":null,"OwnerUserId":"220997","Title":null,"Body":"<p>I've also not found this process flow easy or intuitive. There are two new updates which might help in your project:<\/p>\n\n<ul>\n<li>BigQuery ML now allows you to import TensorFlow models <a href=\"https:\/\/cloud.google.com\/bigquery-ml\/docs\/reference\/standard-sql\/bigqueryml-syntax-create-tensorflow\" rel=\"nofollow noreferrer\">link<\/a> - there are some limitations but this may eliminate some of the back and forth data movement between BQ and cloud storage or other environments.<\/li>\n<li>Cloud DataFlow supports Python 3 in alpha (check the Apache Beam roadmap - <a href=\"https:\/\/beam.apache.org\/roadmap\/#python-3-support\" rel=\"nofollow noreferrer\">link<\/a> )<\/li>\n<\/ul>\n"}]}
+{"QuestionId":52621974,"AnswerCount":0,"Tags":"<tensorflow><cuda>","CreationDate":"2018-10-03T07:33:09.383","AcceptedAnswerId":null,"OwnerUserId":5957168.0,"Title":"tensorflow failed to create session: CUDA driver version is insufficient for CUDA runtime version","Body":"<p>UPDATE: I installed CUDA-9.0 and upon running deviceQuery I see that the CUDA Driver Version and CUDA Runtime Version both are 9.0 now. <\/p>\n\n<p>CUDA Device Query (Runtime API) version (CUDART static linking)<\/p>\n\n<p>Detected 1 CUDA Capable device(s)<\/p>\n\n<p>Device 0: \"GeForce GTX 1070\"\n  CUDA Driver Version \/ Runtime Version          9.0 \/ 9.0\n  CUDA Capability Major\/Minor version number:    6.1\n  Total amount of global memory:                 8112 MBytes (8505655296 bytes)\n  (16) Multiprocessors, (128) CUDA Cores\/MP:     2048 CUDA Cores\n  GPU Max Clock rate:                            1645 MHz (1.64 GHz)\n  Memory Clock rate:                             4004 Mhz\n  Memory Bus Width:                              256-bit\n  L2 Cache Size:                                 2097152 bytes\n  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)\n  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers\n  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers\n  Total amount of constant memory:               65536 bytes\n  Total amount of shared memory per block:       49152 bytes\n  Total number of registers available per block: 65536\n  Warp size:                                     32\n  Maximum number of threads per multiprocessor:  2048\n  Maximum number of threads per block:           1024\n  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)\n  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)\n  Maximum memory pitch:                          2147483647 bytes\n  Texture alignment:                             512 bytes\n  Concurrent copy and kernel execution:          Yes with 2 copy engine(s)\n  Run time limit on kernels:                     Yes\n  Integrated GPU sharing Host Memory:            No\n  Support host page-locked memory mapping:       Yes\n  Alignment requirement for Surfaces:            Yes\n  Device has ECC support:                        Disabled\n  Device supports Unified Addressing (UVA):      Yes\n  Supports Cooperative Kernel Launch:            Yes\n  Supports MultiDevice Co-op Kernel Launch:      Yes\n  Device PCI Domain ID \/ Bus ID \/ location ID:   0 \/ 1 \/ 0\n  Compute Mode:\n     &lt; Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) ><\/p>\n\n<p>deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 9.0, CUDA Runtime Version = 9.0, NumDevs = 1\nResult = PASS<\/p>\n\n<p>But I still get the same error when I create a Session in tensorflow. (I uninstalled tensorflow and installed again) My tensorflow version is '1.10.0'.<\/p>\n\n<p>I have installed tensorflow-gpu using conda <\/p>\n\n<pre><code>2018-10-03 11:07:50.532573: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n<\/code><\/pre>\n\n<p>2018-10-03 11:07:50.657654: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:897] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-03 11:07:50.658351: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 0 with properties: \nname: GeForce GTX 1070 major: 6 minor: 1 memoryClockRate(GHz): 1.645\npciBusID: 0000:01:00.0\ntotalMemory: 7.92GiB freeMemory: 7.46GiB\n2018-10-03 11:07:50.658367: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1484] Adding visible gpu devices: 0\n2018-10-03 11:07:50.658495: E tensorflow\/core\/common_runtime\/direct_session.cc:158] Internal: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version\nTraceback (most recent call last):\n  File \"\", line 1, in \n  File \"\/home\/uday\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1494, in <strong>init<\/strong>\n    super(Session, self).<strong>init<\/strong>(target, graph, config=config)\n  File \"\/home\/uday\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 626, in <strong>init<\/strong>\n    self._session = tf_session.TF_NewSession(self._graph._c_graph, opts)\ntensorflow.python.framework.errors_impl.InternalError: Failed to create session.<\/p>\n\n<p>nvidia driver detail<\/p>\n\n<pre><code>&gt;&gt;nvidia-smi\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 384.130                Driver Version: 384.130                   |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  GeForce GTX 1070    Off  | 00000000:01:00.0  On |                  N\/A |\n| N\/A   43C    P8     9W \/  N\/A |    459MiB \/  8111MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0      1048      G   \/usr\/lib\/xorg\/Xorg                           322MiB |\n|    0      1984      G   compiz                                        65MiB |\n|    0      2509      G   ...-token=CFC23952F1B00BFEF0F52F62675D8896    69MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52621995,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-03T07:34:34.227","AcceptedAnswerId":52712125.0,"OwnerUserId":4625843.0,"Title":"Keras custom metrics with more than two inputs","Body":"<p>I have a VAE model that I've broken down into the encoder and decoder parts, and implemented a custom loss.  A simplified example is as below<\/p>\n\n<pre><code>input = Input(shape=(self.image_height, self.image_width, self.image_channel))\nencoded = build_encoder(input)\ndecoded = build_decoder(encoded)\nmodel = Model(input, decoded)\n<\/code><\/pre>\n\n<p>The loss (simplified) is<\/p>\n\n<pre><code>loss = K.mean(decoded[0] + decoded[1] + encoded[0]**2)\nmodel.add_loss(loss)\nmodel.compile(optimizer=self.optimizer)\n<\/code><\/pre>\n\n<p>My main problem is that I want to use Keras' modelcheckpoint function, which would then require me to set custom metrics.  However, everything I have seen online is similar to <a href=\"https:\/\/keras.io\/metrics\/#custom_metrics\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/metrics\/#custom_metrics<\/a>.  This only takes in y_true and y_pred, and modify the validation loss from there.  How would I implement it in my example model, where the loss is calculated from multiple inputs, not only the final output of \"decoded\"?<\/p>\n","answers":[{"AnswerId":"52712125","CreationDate":"2018-10-09T01:31:10.677","ParentId":null,"OwnerUserId":"4625843","Title":null,"Body":"<p>Well apparently you can still use the variables (keras layers) without passing those into the custom loss function.<\/p>\n\n<p>So for my example, the loss can be calculated as<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n    return K.mean(decoded[0] + decoded[1] + encoded[0]**2)\n\nmodel.compile(optimizer=self.optimizer, loss=custom_loss)\n<\/code><\/pre>\n\n<p>y_true and y_pred is never used, but the actual required inputs can still be called (as long as they are in the same scope as the custom loss function of course).<\/p>\n"}]}
+{"QuestionId":52622343,"AnswerCount":1,"Tags":"<python><tensorflow><keras><gradient>","CreationDate":"2018-10-03T07:56:05.087","AcceptedAnswerId":52626700.0,"OwnerUserId":6541612.0,"Title":"Tensorflow: What gradients needed to be defined for custom operation?","Body":"<p>Although there are many references showing how to register a gradient, but I'm still not very clear what exactly kind of gradient need to be defined.<\/p>\n\n<p>Some similar topics: \n<a href=\"https:\/\/stackoverflow.com\/questions\/43256517\/how-to-register-a-custom-gradient-for-a-operation-composed-of-tf-operations\">How to register a custom gradient for a operation composed of tf operations<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/36456436\/how-can-i-define-only-the-gradient-for-a-tensorflow-subgraph\">How Can I Define Only the Gradient for a Tensorflow Subgraph?<\/a><\/p>\n\n<hr>\n\n<p>Okay, here comes my question:<\/p>\n\n<p>I have a forward function <code>y = f(A,B)<\/code>, where the size of each of them are:<\/p>\n\n<pre><code>y: (batch_size, m, n)\nA: (batch_size, a, a)\nB: (batch_size, b, b)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7vlyA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7vlyA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Suppose I can write down the mathematical partial derivatives of every element of y with respect every element of A and B.  <code>dy\/dA, dy\/dB<\/code>. My question is what should I return in the gradient function?<\/p>\n\n<pre><code>@ops.RegisterGradient(\"f\")\ndef f_grad(op, grad):\n    ...\n    return ???, ???\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op#implement_the_gradient_in_python\" rel=\"nofollow noreferrer\">Here<\/a> says that <em>The result of the gradient function must be a list of Tensor objects representing the gradients with respect to each input.<\/em><\/p>\n\n<p>It is very easy to understand the gradient to be defined when <code>y<\/code> is scalar and <code>A<\/code>, <code>B<\/code> are matrix. But when <code>y<\/code> is matrix and <code>A<\/code>, <code>B<\/code> are also matrix, what should that gradient be?<\/p>\n","answers":[{"AnswerId":"52626700","CreationDate":"2018-10-03T12:07:23.450","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gradients\" rel=\"nofollow noreferrer\"><code>tf.gradients<\/code><\/a> computes the gradient of the sum of each output tensor with respect to each value in the input tensors. A gradient operation receives the op for which you are computing the gradient, <code>op<\/code>, and the gradient accumulated at this point, <code>grad<\/code>. In your example, <code>grad<\/code> would be a tensor with the same shape as <code>y<\/code>, and each value would be the gradient of the corresponding value in <code>y<\/code> - that is, if <code>grad[0, 0] == 2<\/code>, it means that increasing <code>y[0, 0]<\/code> by 1 will increase the sum of the output tensor by 2 (I know, you probably are already clear on this). Now you have to compute the same thing for <code>A<\/code> and <code>B<\/code>. Let's say you figure out that increasing <code>A[2, 3]<\/code> by 1 will increase <code>y[0, 0]<\/code> by 3 and have no effect over any other value in <code>y<\/code>. That means that would increase the sum of the output value by 3 \u00d7 2 = 6, so the gradient for <code>A[2, 3]<\/code> would be 6.<\/p>\n\n<p>As an example, let's take the gradient of the matrix multiplication (op <code>MatMul<\/code>), which you can find in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.11.0\/tensorflow\/python\/ops\/math_grad.py#L1087-L1107\" rel=\"nofollow noreferrer\"><code>tensorflow\/python\/ops\/math_grad.py<\/code><\/a>:<\/p>\n\n<pre><code>@ops.RegisterGradient(\"MatMul\")\ndef _MatMulGrad(op, grad):\n  \"\"\"Gradient for MatMul.\"\"\"\n\n  t_a = op.get_attr(\"transpose_a\")\n  t_b = op.get_attr(\"transpose_b\")\n  a = math_ops.conj(op.inputs[0])\n  b = math_ops.conj(op.inputs[1])\n  if not t_a and not t_b:\n    grad_a = gen_math_ops.mat_mul(grad, b, transpose_b=True)\n    grad_b = gen_math_ops.mat_mul(a, grad, transpose_a=True)\n  elif not t_a and t_b:\n    grad_a = gen_math_ops.mat_mul(grad, b)\n    grad_b = gen_math_ops.mat_mul(grad, a, transpose_a=True)\n  elif t_a and not t_b:\n    grad_a = gen_math_ops.mat_mul(b, grad, transpose_b=True)\n    grad_b = gen_math_ops.mat_mul(a, grad)\n  elif t_a and t_b:\n    grad_a = gen_math_ops.mat_mul(b, grad, transpose_a=True, transpose_b=True)\n    grad_b = gen_math_ops.mat_mul(grad, a, transpose_a=True, transpose_b=True)\n  return grad_a, grad_b\n<\/code><\/pre>\n\n<p>We will focus on the case where <code>transpose_a<\/code> and <code>transpose_b<\/code> are both <code>False<\/code>, and so we are in the first branch , <code>if not t_a and not t_b:<\/code> (also ignore the <code>conj<\/code>, which is meant for complex values). 'a' and 'b' are the operands here and, as said before, <code>grad<\/code> has the gradient of the sum of the output with respect to each value in the multiplication result. So how would things change if I increase <code>a[0, 0]<\/code> by one? Basically, each element in the first row of the product matrix would be increased by the values in the first row of <code>b<\/code>. So the gradient for <code>a[0, 0]<\/code> is the dot product of the first row of <code>b<\/code> and the first row of <code>grad<\/code> - that is, how much I would increase each output value multiplied by the accumulated gradient of each of these. If you think about it, the line <code>grad_a = gen_math_ops.mat_mul(grad, b, transpose_b=True)<\/code> is doing exactly that. <code>grad_a[0, 0]<\/code> will be the dot product of the first row of <code>grad<\/code> and the first row of <code>b<\/code> (because we are transposing <code>b<\/code> here), and, in general, <code>grad_a[i, j]<\/code> will be the dot product of the <code>i<\/code>-th row of <code>grad<\/code> and the <code>j<\/code>-th row of <code>b<\/code>. You can follow a similar reasoning for <code>grad_b<\/code> too.<\/p>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>As an example, see how <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gradients\" rel=\"nofollow noreferrer\"><code>tf.gradients<\/code><\/a> and the registered gradient relate to each other:<\/p>\n\n<pre><code>import tensorflow as tf\n# Import gradient registry to lookup gradient functions\nfrom tensorflow.python.framework.ops import _gradient_registry\n\n# Gradient function for matrix multiplication\nmatmul_grad = _gradient_registry.lookup('MatMul')\n# A matrix multiplication\na = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)\nb = tf.constant([[6, 7, 8], [9, 10, 11]], dtype=tf.float32)\nc = tf.matmul(a, b)\n# Gradient of sum(c) wrt each element of a\ngrad_c_a_1, = tf.gradients(c, a)\n# The same is obtained by backpropagating an all-ones matrix\ngrad_c_a_2, _ = matmul_grad(c.op, tf.ones_like(c))\n# Multiply each element of c by itself, but stopping the gradients\n# This should scale the gradients by the values of c\ncc = c * tf.stop_gradient(c)\n# Regular gradients computation\ngrad_cc_a_1, = tf.gradients(cc, a)\n# Gradients function called with c as backpropagated gradients\ngrad_cc_a_2, _ = matmul_grad(c.op, c)\nwith tf.Session() as sess:\n    print('a:')\n    print(sess.run(a))\n    print('b:')\n    print(sess.run(b))\n    print('c = a * b:')\n    print(sess.run(c))\n    print('tf.gradients(c, a)[0]:')\n    print(sess.run(grad_c_a_1))\n    print('matmul_grad(c.op, tf.ones_like(c))[0]:')\n    print(sess.run(grad_c_a_2))\n    print('tf.gradients(c * tf.stop_gradient(c), a)[0]:')\n    print(sess.run(grad_cc_a_1))\n    print('matmul_grad(c.op, c)[0]:')\n    print(sess.run(grad_cc_a_2))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>a:\n[[1. 2.]\n [3. 4.]]\nb:\n[[ 6.  7.  8.]\n [ 9. 10. 11.]]\nc = a * b:\n[[24. 27. 30.]\n [54. 61. 68.]]\ntf.gradients(c, a)[0]:\n[[21. 30.]\n [21. 30.]]\nmatmul_grad(c.op, tf.ones_like(c))[0]:\n[[21. 30.]\n [21. 30.]]\ntf.gradients(c * tf.stop_gradient(c), a)[0]:\n[[ 573.  816.]\n [1295. 1844.]]\nmatmul_grad(c.op, c)[0]:\n[[ 573.  816.]\n [1295. 1844.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52622518,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><pytorch><spatial-pooling>","CreationDate":"2018-10-03T08:06:50.727","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"how to convert pytorch adaptive_avg_pool2d method to keras or tensorflow","Body":"<p>I don't know how to convert the PyTorch method <code>adaptive_avg_pool2d<\/code> to Keras or TensorFlow. Anyone can help?\nPyTorch mehod is <\/p>\n\n<pre><code>adaptive_avg_pool2d(14,[14])\n<\/code><\/pre>\n\n<p>I tried to use the average pooling, the reshape the tensor in Keras, but got the error:<\/p>\n\n<blockquote>\n  <p>ValueError: total size of new array must be unchanged<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52625680","CreationDate":"2018-10-03T11:06:04.340","ParentId":null,"OwnerUserId":"604734","Title":null,"Body":"<p>I'm not sure if I understood your question, but in PyTorch, you pass the spatial dimensions to <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.AdaptiveAvgPool2d\" rel=\"nofollow noreferrer\"><code>AdaptiveAvgPool2d<\/code><\/a>. For instance, if you want to have an output sized 5x7, you can use <code>nn.AdaptiveAvgPool2d((5,7))<\/code>.<\/p>\n\n<p>If you want a global average pooling layer, you can use <code>nn.AdaptiveAvgPool2d(1)<\/code>. In Keras you can just use <a href=\"https:\/\/keras.io\/layers\/pooling\/#globalaveragepooling2d\" rel=\"nofollow noreferrer\"><code>GlobalAveragePooling2D<\/code><\/a>.<\/p>\n\n<p>For other output sizes in Keras, you need to use <a href=\"https:\/\/keras.io\/layers\/pooling\/#averagepooling2d\" rel=\"nofollow noreferrer\"><code>AveragePooling2D<\/code><\/a>, but you can't specify the output shape directly. You need to calculate\/define the <code>pool_size<\/code>, <code>stride<\/code>, and <code>padding<\/code> parameters depending on how you want the output shape. If you need help with the calculations, check <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/#pool\" rel=\"nofollow noreferrer\">this page of CS231n course<\/a>.<\/p>\n"}]}
+{"QuestionId":52623005,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><deep-learning><computer-vision>","CreationDate":"2018-10-03T08:40:25.907","AcceptedAnswerId":null,"OwnerUserId":9579075.0,"Title":"pixel wise softmax with crossentropy for multiclass segmentation","Body":"<p>I am trying to implement a multiclass semantic segmentation model with 2\nclasses ( human, car). here is my modified implementation of unet architecture. I number of output channels to 3 (3 classes - human, car, background). How do i get pixel wise classification?\nhere are 2 examples from my ground truth masks.<\/p>\n\n<p>i am using 1 channel for each object class ie.<\/p>\n\n<ul>\n<li>channel 1 for class=car<\/li>\n<li>channel 2 for class=background<\/li>\n<li>channel 3 for class=human<\/li>\n<\/ul>\n\n<p><img src=\"https:\/\/raw.githubusercontent.com\/srihari-humbarwadi\/person-segmentation-keras\/master\/outputs\/car.png\" alt=\"class 1\">\n<img src=\"https:\/\/raw.githubusercontent.com\/srihari-humbarwadi\/person-segmentation-keras\/master\/outputs\/human.png\" alt=\"class 2\"><\/p>\n\n<pre><code>def conv_block(tensor, nfilters, size=3, padding='same', initializer=\"he_normal\"):\n    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(tensor)\n    x = BatchNormalization()(x)\n    x = Activation(\"relu\")(x)\n    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(x)\n    x = BatchNormalization()(x)\n    x = Activation(\"relu\")(x)\n    return x\n\n\ndef deconv_block(tensor, residual, nfilters, size=3, padding='same', strides=(2, 2)):\n    y = Conv2DTranspose(nfilters, kernel_size=(size, size), strides=strides, padding=padding)(tensor)\n    y = concatenate([y, residual], axis=3)\n    y = conv_block(y, nfilters)\n    return y\n\ndef Unet(img_height, img_width, nclasses=3, filters=64):\n    input_layer = Input(shape=(img_height, img_width, 3), name='image_input')\n    conv1 = conv_block(input_layer, nfilters=filters)\n    conv1_out = MaxPooling2D(pool_size=(2, 2))(conv1)\n    conv2 = conv_block(conv1_out, nfilters=filters*2)\n    conv2_out = MaxPooling2D(pool_size=(2, 2))(conv2)\n    conv3 = conv_block(conv2_out, nfilters=filters*4)\n    conv3_out = MaxPooling2D(pool_size=(2, 2))(conv3)\n    conv4 = conv_block(conv3_out, nfilters=filters*8)\n    conv4_out = MaxPooling2D(pool_size=(2, 2))(conv4)\n    conv4_out = Dropout(0.5)(conv4_out)\n    conv5 = conv_block(conv4_out, nfilters=filters*16)\n    conv5 = Dropout(0.5)(conv5)\n\n    deconv6 = deconv_block(conv5, residual=conv4, nfilters=filters*8)\n    deconv6 = Dropout(0.5)(deconv6)\n    deconv7 = deconv_block(deconv6, residual=conv3, nfilters=filters*4)\n    deconv7 = Dropout(0.5)(deconv7) \n    deconv8 = deconv_block(deconv7, residual=conv2, nfilters=filters*2)\n    deconv9 = deconv_block(deconv8, residual=conv1, nfilters=filters)\n\n    output_layer = Conv2D(filters=nclasses, kernel_size=(1, 1))(deconv9)\n    output_layer = BatchNormalization()(output_layer)\n    output_layer = Reshape((img_height*img_width, nclasses), input_shape=(img_height, img_width, nclasses))(output_layer)\n    output_layer = Activation('softmax')(output_layer)\n\n    model = Model(inputs=input_layer, outputs=output_layer, name='Unet')\n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"52624843","CreationDate":"2018-10-03T10:19:27.090","ParentId":null,"OwnerUserId":"6398153","Title":null,"Body":"<p>You are almost done, now backpropagate the network error with:<\/p>\n\n<pre><code>loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=output_layer, labels=labels))   \ntf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss) \n<\/code><\/pre>\n\n<p>You don't have to convert your ground truth into the one-hot format, sparse_softmax will dot it for you.<\/p>\n"}]}
+{"QuestionId":52623614,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T09:14:18.600","AcceptedAnswerId":52637848.0,"OwnerUserId":9367924.0,"Title":"Stuck on tensorflow advanced indexing","Body":"<p>Given an input tensor of shape (?,5,5) I need to find the max sum for each example by summing elements as specified by an index tensor of shape (120,5,2). The index tensor lists 120 ways to sum an example's 5x5 matrix.\nE.g.:<\/p>\n\n<pre><code>Input tensor (?,5,5):\n[\n  [\n    [0,1,0,0,0],\n    [0,0,0,0,1],\n    [1,0,0,0,0],\n    [0,0,0,1,0],\n    [0,0,1,0,0]\n  ],\n  [\n    ...\n  ],\n  ...\n]\n\nIndex tensor(120,5,2):\n[\n  [\n    [0,1], \n    [1,4], \n    [2,2], \n    [3,0], \n    [4,3]  \n  ],\n  [\n    ...\n  ],\n...\n]\n<\/code><\/pre>\n\n<p>Here, the result for the <em>first<\/em> summation would be 1+1+0+0+0 = 2.\nI need to find the max sum for all 120 ways given by the index array for each example.<\/p>\n\n<p>In numpy I would use advanced indexing with integer index arrays but unfortunately tf doesn't support this. I found tf.gather_nd but it seems that I this function assumes I know the indices for each example in the batch which I don't.<\/p>\n","answers":[{"AnswerId":"52637848","CreationDate":"2018-10-04T01:47:40.593","ParentId":null,"OwnerUserId":"9367924","Title":null,"Body":"<p>Solved it.\nThe trick is to transpose the axes. That way the unknown dimensions can be pushed to the end and gather_nd can select all slices that come before the unknown dimension.<\/p>\n\n<p>Here's the full code if anybody cares...<\/p>\n\n<pre><code>def permute(a, l, r):\n    if l==r:\n        yield list(zip([0,1,2,3,4],a))\n    else:\n        for i in range(l,r+1):\n            a[l], a[i] = a[i], a[l]\n            yield from permute(a, l+1, r)\n            a[l], a[i] = a[i], a[l]\n\ndef multi_class_acc_positions(pred, target, input):\n    pred_5x5 = tf.reshape(pred, [-1, 5, 5])\n    target_5x5 = tf.reshape(target, [-1, 5, 5])\n    pred_5x5_T = tf.transpose(pred_5x5, (1,2,0))\n    all_perms = tf.constant(list(permute([0,1,2,3,4],0,4)))\n    selected_elemens_per_example = tf.gather_nd(pred_5x5_T, all_perms)\n    sums_per_example = tf.reduce_sum(selected_elemens_per_example, axis=1)\n    best_perm_per_example_index = tf.argmax(sums_per_example, axis=0)\n    best_perms = tf.gather_nd(all_perms, best_perm_per_example_index[:,tf.newaxis])[:,:,1]\n    pred_5x5_one_hot = tf.reshape(tf.one_hot(best_perms, depth=5), (-1, 5, 5))\n    correct_prediction = tf.equal(tf.argmax(pred_5x5_one_hot, axis=2), tf.argmax(target_5x5, axis=2))\n    all_correct = tf.reduce_min(tf.cast(correct_prediction, tf.float32), 1)\n    acc = tf.reduce_mean(all_correct)\n    return acc\n<\/code><\/pre>\n"}]}
+{"QuestionId":52623706,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><tflearn>","CreationDate":"2018-10-03T09:19:05.860","AcceptedAnswerId":null,"OwnerUserId":10169965.0,"Title":"Im getting an error when I try to run this file. tf learn problem?","Body":"<p>I found something that i wanted to run from GitHub but when i do i get this error.<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Traceback (most recent call last):\n  File \"C:\\Users\\Samuel\\Desktop\\GTA\\pygta5-master\\pygta5-master\\Tutorial Codes\\Part 14-15\\train_model.py\", line 4, in &lt;module&gt;\n    from alexnet import alexnet\n  File \"C:\\Users\\Samuel\\Desktop\\GTA\\pygta5-master\\pygta5-master\\Tutorial Codes\\Part 14-15\\alexnet.py\", line 11, in &lt;module&gt;\n    import tflearn\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tflearn\\__init__.py\", line 4, in &lt;module&gt;\n    from . import config\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tflearn\\config.py\", line 3, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 52, in &lt;module&gt;\n    from tensorflow.core.framework.graph_pb2 import *\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"C:\\Users\\Samuel\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in &lt;module&gt;\n    from google.protobuf.pyext import _message\nImportError: DLL load failed: The specified procedure could not be found.\n<\/code><\/pre>\n\n<p>This is the code <\/p>\n\n<pre><code># train_model.py\n\nimport numpy as np\nfrom alexnet import alexnet\nWIDTH = 160\nHEIGHT = 120\nLR = 1e-3\nEPOCHS = 10\nMODEL_NAME = 'pygta5-car-fast-{}-{}-{}-epochs-300K-data.model'.format(LR, 'alexnetv2',EPOCHS)\n\nmodel = alexnet(WIDTH, HEIGHT, LR)\n\nhm_data = 22\nfor i in range(EPOCHS):\n    for i in range(1,hm_data+1):\n        train_data = np.load('training_data-{}-balanced.npy'.format(i))\n\n        train = train_data[:-100]\n        test = train_data[-100:]\n\n        X = np.array([i[0] for i in train]).reshape(-1,WIDTH,HEIGHT,1)\n        Y = [i[1] for i in train]\n\n        test_x = np.array([i[0] for i in test]).reshape(-1,WIDTH,HEIGHT,1)\n        test_y = [i[1] for i in test]\n\n        model.fit({'input': X}, {'targets': Y}, n_epoch=1, validation_set=({'input': test_x}, {'targets': test_y}), \n            snapshot_step=500, show_metric=True, run_id=MODEL_NAME)\n\n        model.save(MODEL_NAME)\n\n\n\n# tensorboard --logdir=foo:C:\/path\/to\/log\n<\/code><\/pre>\n\n<p>it looks like its trying to get access to the alexnet.py which is<\/p>\n\n<pre><code># alexnet.py\n\n\"\"\" AlexNet.\nReferences:\n    - Alex Krizhevsky, Ilya Sutskever &amp; Geoffrey E. Hinton. ImageNet\n    Classification with Deep Convolutional Neural Networks. NIPS, 2012.\nLinks:\n    - [AlexNet Paper](http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf)\n\"\"\"\n\nimport tflearn\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\nfrom tflearn.layers.normalization import local_response_normalization\n\ndef alexnet(width, height, lr):\n    network = input_data(shape=[None, width, height, 1], name='input')\n    network = conv_2d(network, 96, 11, strides=4, activation='relu')\n    network = max_pool_2d(network, 3, strides=2)\n    network = local_response_normalization(network)\n    network = conv_2d(network, 256, 5, activation='relu')\n    network = max_pool_2d(network, 3, strides=2)\n    network = local_response_normalization(network)\n    network = conv_2d(network, 384, 3, activation='relu')\n    network = conv_2d(network, 384, 3, activation='relu')\n    network = conv_2d(network, 256, 3, activation='relu')\n    network = max_pool_2d(network, 3, strides=2)\n    network = local_response_normalization(network)\n    network = fully_connected(network, 4096, activation='tanh')\n    network = dropout(network, 0.5)\n    network = fully_connected(network, 4096, activation='tanh')\n    network = dropout(network, 0.5)\n    network = fully_connected(network, 3, activation='softmax')\n    network = regression(network, optimizer='momentum',\n                         loss='categorical_crossentropy',\n                         learning_rate=lr, name='targets')\n\n    model = tflearn.DNN(network, checkpoint_path='model_alexnet',\n                        max_checkpoints=1, tensorboard_verbose=0, tensorboard_dir='log')\n\n    return model\n<\/code><\/pre>\n\n<p>After looking around a lot I tried a lot of things and then i tired to import tflearn on a python cmd and it wouldn't import when i ran the command import tflearn. So is it an Tflearn problem?<\/p>\n\n<p>Im on python 3.6 and here is everything else installed<\/p>\n\n<pre><code>absl-py             0.5.0\nastor               0.7.1\ngast                0.2.0\ngrpcio              1.15.0\nh5py                2.8.0\nKeras-Applications  1.0.6\nKeras-Preprocessing 1.0.5\nMarkdown            3.0.1\nnumpy               1.15.2\nopencv-python       3.4.3.18\npandas              0.23.4\nPillow              5.3.0\npip                 18.0\nprotobuf            3.6.1\npypiwin32           223\npython-dateutil     2.7.3\npytz                2018.5\npywin32             224\nsetuptools          28.8.0\nsix                 1.11.0\ntensorboard         1.11.0\ntensorflow          1.11.0\ntensorflow-gpu      1.11.0\ntermcolor           1.1.0\ntflearn             0.3.2\nvirtualenv          16.0.0\nWerkzeug            0.14.1\nwheel               0.32.0\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52623733,"AnswerCount":1,"Tags":"<python><tensorflow><computer-vision>","CreationDate":"2018-10-03T09:20:14.117","AcceptedAnswerId":null,"OwnerUserId":3042444.0,"Title":"Tensorflow Incompatible shapes error while training","Body":"<p>I have been training an object detector, was able to do this successfully with 1 class, now I have added 2 classes and a few more images to the dataset, I have regenerated xml to csv and tf records by deleting old ones, pbtxt file is correct and also deleted old checkpoints inside the training directory but when I run the training job like this:<\/p>\n\n<pre><code>python legacy\/train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_mobilenet_v1_pets.config\n<\/code><\/pre>\n\n<p>following this <a href=\"https:\/\/pythonprogramming.net\/training-custom-objects-tensorflow-object-detection-api-tutorial\/?completed=\/creating-tfrecord-files-tensorflow-object-detection-api-tutorial\/\" rel=\"nofollow noreferrer\">tutorial<\/a> . I get the following error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Incompatible shapes: [2,1917] vs. [4,1]\n [[Node: Loss\/Match\/cond\/mul_4 = Mul[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](Loss\/Match\/cond\/one_hot, Loss\/Match\/cond\/Cast_2)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52916700","CreationDate":"2018-10-21T15:05:42.260","ParentId":null,"OwnerUserId":"5484902","Title":null,"Body":"<p>If you are training Mobilenet SSD following the Obstacle Detection API tutorial you could try commenting out the \"ssd_random_crop\" augmentation in your config file. This worked for me. Was initially pointed out by rky0930 <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5423\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52624333,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-03T09:52:30.077","AcceptedAnswerId":null,"OwnerUserId":581442.0,"Title":"TensorFlow Lite Inference on a PC very slow","Body":"<p>I'm experimenting with TensorFlow Lite on the PC:<\/p>\n\n<pre><code>from tensorflow.contrib.lite.python import interpreter as interpreter_wrapper\n\n\nmodel_path = os.path.join(ROOT_DIR, 'model', 'yolov3.tflite')\n\ninterpreter = interpreter_wrapper.Interpreter(model_path=model_path)\ninterpreter.allocate_tensors()\n\ninput_details = interpreter.get_input_details()\noutput_details = interpreter.get_output_details()\n\nif input_details[0]['dtype'] == np.float32:\n    floating_model = True\n\norig = cv2.imread('data\/dog-cycle-car.png')\n\nheight = input_details[0]['shape'][1]\nwidth = input_details[0]['shape'][2]\n\nimage, image_data = preprocess_image(orig, (height, width))\n\nstart = time.time()\ninterpreter.set_tensor(input_details[0]['index'], image_data)\ninterpreter.invoke()\nend = time.time()\n\noutput_data = interpreter.get_tensor(output_details[0]['index'])\n\n# Takes around 30 seconds on a PC\nprint(\"Inference time: {:.2f}s\".format((end - start)))\n<\/code><\/pre>\n\n<p>However inference takes around 30 seconds, which seems slightly abnormal. Am  I missing something?<\/p>\n","answers":[{"AnswerId":"54584433","CreationDate":"2019-02-08T00:26:59.993","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>For issues on performance, please file a tracking issue on Github since there may be a bug.<\/p>\n\n<p>The other question is how much of a performance difference is there? One factor is that the optimizations in Tensorflow Lite don't target x86 (where as Tensorflow does more so) and we target architectures more associated with mobile\/embedded (e.g. arm).<\/p>\n"}]}
+{"QuestionId":52624457,"AnswerCount":1,"Tags":"<python><keras><hyperparameters>","CreationDate":"2018-10-03T09:58:21.103","AcceptedAnswerId":52633721.0,"OwnerUserId":3261292.0,"Title":"Keras hyperparameter tuning with hyperas using manual metric","Body":"<p>I'm using the <a href=\"https:\/\/github.com\/maxpumperla\/hyperas\" rel=\"nofollow noreferrer\">hyperas document example<\/a> to tune the network parameters but based on f1 score instead of accuracy.<\/p>\n\n<p>I'm using the following implementation for f1 score:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef f1(y_true, y_pred):\n    def recall(y_true, y_pred):\n        \"\"\"Recall metric.\n        Only computes a batch-wise average of recall.\n        Computes the recall, a metric for multi-label classification of\n        how many relevant items are selected.\n        \"\"\"\n        true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))\n        possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))\n        recall = true_positives \/ (possible_positives + K.epsilon())\n        return recall\n\n    def precision(y_true, y_pred):\n        \"\"\"Precision metric.\n        Only computes a batch-wise average of precision.\n        Computes the precision, a metric for multi-label classification of\n        how many selected items are relevant.\n        \"\"\"\n        true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))\n        predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))\n        precision = true_positives \/ (predicted_positives + K.epsilon())\n        return precision\n    precision = precision(y_true, y_pred)\n    recall = recall(y_true, y_pred)\n    return 2*((precision*recall)\/(precision+recall+K.epsilon()))\n<\/code><\/pre>\n\n<p>with updating the metric parameter for compile function in following code line:<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy', metrics=['accuracy'],\n                  optimizer={{choice(['rmsprop', 'adam', 'sgd'])}})\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy', metrics=[f1],\n                  optimizer={{choice(['rmsprop', 'adam', 'sgd'])}})\n<\/code><\/pre>\n\n<p>the above metric works perfectly without using hyperas, while when I try to use it with the tuning process, I get the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/path\/test.py\", line 96, in &lt;module&gt;\n    trials=Trials())\n  File \"C:\\Python35\\lib\\site-packages\\hyperas\\optim.py\", line 67, in minimize\n    verbose=verbose)\n  File \"C:\\Python35\\lib\\site-packages\\hyperas\\optim.py\", line 133, in base_minimizer\n    return_argmin=True),\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\fmin.py\", line 367, in fmin\n    return_argmin=return_argmin,\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\base.py\", line 635, in fmin\n    return_argmin=return_argmin)\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\fmin.py\", line 385, in fmin\n    rval.exhaust()\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\fmin.py\", line 244, in exhaust\n    self.run(self.max_evals - n_done, block_until_done=self.asynchronous)\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\fmin.py\", line 218, in run\n    self.serial_evaluate()\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\fmin.py\", line 137, in serial_evaluate\n    result = self.domain.evaluate(spec, ctrl)\n  File \"C:\\Python35\\lib\\site-packages\\hyperopt\\base.py\", line 840, in evaluate\n    rval = self.fn(pyll_rval)\n  File \"D:\\path\\temp_model.py\", line 86, in keras_fmin_fnct\nNameError: name 'f1' is not defined\n<\/code><\/pre>\n","answers":[{"AnswerId":"52633721","CreationDate":"2018-10-03T18:40:07.667","ParentId":null,"OwnerUserId":"1536437","Title":null,"Body":"<p>If you are following the code example you linked to, you are not making hyperas aware of the custom f1 function.  The package author provides an <a href=\"https:\/\/github.com\/maxpumperla\/hyperas\/blob\/master\/examples\/hyperas_in_intermediate_fns.py\" rel=\"nofollow noreferrer\">example to do that<\/a>, as well.<\/p>\n\n<p>In short, you need to add an additional <code>functions<\/code> argument to your <code>optim.minimize()<\/code> call.  Something like<\/p>\n\n<pre><code>    best_run, best_model = optim.minimize(model=model,\n     data=data,\n     functions=[f1],\n     algo=tpe.suggest,\n     max_evals=5,\n     trials=Trials())\n<\/code><\/pre>\n\n<p>I actually just implemented it today, so I'm confident you can get it working, too :)<\/p>\n"}]}
+{"QuestionId":52624703,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T10:11:44.403","AcceptedAnswerId":52627105.0,"OwnerUserId":5937211.0,"Title":"Difference between installation libraries of Tensorflow GPU vs CPU","Body":"<p>Recently, I wanted to move my python libraries to a pendrive to keep all the libraries constant while switching between my Workstation and Laptop. (Also so that if I update one, its updated on other also)<\/p>\n\n<p>For this, I have installed a tensorflow-gpu version on pendrive (My laptop doesnt have a GPU). Everything works fine without a problem on both PC (It detects and uses my GPU without a problem) and Laptop (It automatically uses my CPU).<\/p>\n\n<p>That's where my question lies. What is the difference between a <\/p>\n\n<pre><code>tensorflow-gpu \n<\/code><\/pre>\n\n<p>AND just<\/p>\n\n<pre><code>tensorflow\n<\/code><\/pre>\n\n<p>? (Because when no GPU is found, tensorflow-gpu automatically uses the CPU version.)<\/p>\n\n<p>Does the difference lie only in the GPU support? Then why at all have a non GPU version of tensorflow?<\/p>\n\n<p>Also, is it alright to proceed like this? Or should I create virtual environments to keep separate installations for CPU and GPU?<\/p>\n\n<hr>\n\n<p><strong>Edit:<\/strong> \nThe closest answer i can find is \n<a href=\"https:\/\/stackoverflow.com\/questions\/45290326\/how-to-develope-for-tensor-flow-with-gpu-without-a-gpu\">How to develope for tensor flow with gpu without a gpu<\/a><\/p>\n\n<p>But it only specifies that its completely okay to use tensorflow-gpu on a CPU platform, but it still does not answer my first question. Also the answer might be outdated as tensorflow keeps releasing new updates.<\/p>\n\n<hr>\n\n<p><strong>Edit 2:<\/strong>\nI had installed the tensorflow-gpu version on my workstation with GTX 1070 (Thus a successful install).<\/p>\n\n<p>Also i understand the difference is <code>pip install tensorflow-gpu<\/code> will require CUDA enabled device to install, but my question is more towards the usage of the libraries because I am not getting any problems when using the <code>tensorflow-gpu<\/code> version on my laptop (with no GPU) and all my scripts run without any error.<\/p>\n\n<p>(Also removed pip install from above to avoid confusion)<\/p>\n\n<hr>\n\n<p><strong>Edit 3:<\/strong><\/p>\n\n<p>Also isn't running <code>tensorflow-gpu<\/code> on a system with no GPU same as setting <code>CUDA_VISIBLE_DEVICES=-1<\/code><\/p>\n","answers":[{"AnswerId":"52624984","CreationDate":"2018-10-03T10:26:28.153","ParentId":null,"OwnerUserId":"2653663","Title":null,"Body":"<p><code>tensorflow-gpu<\/code> requires cuda\/cudnn. <code>tensorflow<\/code> does not. <code>pip<\/code> doesn't install cuda for you (<code>conda<\/code> does), so <code>pip install tensorflow-gpu<\/code> won't work out of the box on most systems without a nvidia gpu.<\/p>\n"},{"AnswerId":"52627105","CreationDate":"2018-10-03T12:31:18.373","ParentId":null,"OwnerUserId":"5276428","Title":null,"Body":"<p>One thing to Note: CUDA can be installed even if you don't have a GPU in your system.<\/p>\n\n<p>For packages <code>tensorflow<\/code> and <code>tensorflow-gpu<\/code> I hope this clears the confusion. <strong>yes\/no<\/strong> means \"<em>Will the package work out of the box when executing<\/em> <code>import tensorflow as tf<\/code>\"? Here are the differences:<\/p>\n\n<pre><code>| Support for TensorFlow libraries | tensorflow | tenosorflow-gpu |\n|----------------------------------|------------|-----------------|\n| cpu-only                         |    yes     |       no        |\n| gpu with cuda+cudnn installed    |    yes     |       yes       |\n| gpu without cuda+cudnn installed |    yes     |       no        |\n<\/code><\/pre>\n\n<p>Edit: Confirmed the <code>no<\/code> answers on a <code>cpu-only<\/code> system and the <code>gpu without cuda+cudnn installed<\/code> (by removing CUDA+CuDNN env variables).<\/p>\n"}]}
+{"QuestionId":52624870,"AnswerCount":0,"Tags":"<python><machine-learning><keras><embedding><doc2vec>","CreationDate":"2018-10-03T10:21:05.617","AcceptedAnswerId":null,"OwnerUserId":10418811.0,"Title":"Doc2Vec Distributed Memory Model of Paragraph Vectors (PV-DM) doesn't classified","Body":"<p>I'm using Keras to code DM according to Mikolov article.<\/p>\n\n<p>The documents we are facing are very particular : we have a lot of different documents (around 20 000) but very few different words (only 100).<\/p>\n\n<p>When I use the document embedding in order to classify my documents, it doesn't seem to work ...<\/p>\n\n<p>I don't know if this is due to the very specific data or to my way to code DM.<\/p>\n\n<p>Any clue would be greatly appreciated !<\/p>\n\n<p>This is my code :<\/p>\n\n<pre><code>## Libraries\n\nimport numpy as np\nimport pandas as pd\nimport keras.backend as K\nfrom keras.models import Model\nfrom keras.layers import Dense, Embedding, Lambda,concatenate, Input, Flatten\nfrom keras.utils.data_utils import get_file\nfrom keras.utils import np_utils\nfrom keras.preprocessing import sequence\nfrom keras.preprocessing.text import Tokenizer\nfrom keras import optimizers\nfrom keras.constraints import non_neg\nimport random as rd\n\n\n## Parameters\n\nseuil = 0.0001 # Mikolov threshold\nloss = 'categorical_crossentropy'\ndim=20 # dimension of words and documents embeddings\nlr = 0.005 # learning rate\nepochs = 15\nshuffle=True\nsteps_per_epoch = 10000\nwindow_size = 7\nseed=123\nrd.seed(seed)\n\nV1 = 100 # number of words\nV2 = 20000 # number of documents\n<\/code><\/pre>\n\n<p>This what my corpus looks like :<\/p>\n\n<pre><code>new_corpus =\n[[2],\n [33, 33, 11],\n [8, 16],\n [27, 10],\n [13],\n [],\n [21, 21, 32, 10, 1],\n [],\n [1, 27],\n ... ]\n<\/code><\/pre>\n\n<p>The neural network used :<\/p>\n\n<pre><code>input_words=Input((window_size*2,))\ncbow_words = Embedding(input_dim=V1, output_dim=dim, input_length=window_size*2 ,embeddings_constraint=non_neg())(input_words)\n\ninput_texts=Input((V2,))\ncbow_texts = Embedding(input_dim=V2, output_dim=dim, input_length=V2 ,embeddings_constraint=non_neg())(input_texts)\n\nconcat = concatenate([cbow_words,cbow_texts],axis=1)\nlambd = Lambda(lambda x: K.mean(x, axis=1), output_shape=(dim,))(concat)\noutput = Dense(V1+1, activation='softmax')(lambd) # V1+1 because words are indexed from 1\n\ncbow = Model(inputs=[input_words,input_texts], outputs=output)\n\ncbow.compile(loss=loss, optimizer=optimizers.Adadelta(lr=lr, rho=0.95, epsilon=None, decay=0.0))\ncbow.summary()\n<\/code><\/pre>\n\n<p>It prints :<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_47 (InputLayer)           (None, 14)           0                                            \n__________________________________________________________________________________________________\ninput_48 (InputLayer)           (None, 20000)        0                                            \n__________________________________________________________________________________________________\nembedding_47 (Embedding)        (None, 14, 20)       2160        input_47[0][0]                   \n__________________________________________________________________________________________________\nembedding_48 (Embedding)        (None, 20000, 20)    400000      input_48[0][0]                   \n__________________________________________________________________________________________________\nconcatenate_24 (Concatenate)    (None, 20000, 20)    0           embedding_47[0][0]               \n                                                                 embedding_48[0][0]               \n__________________________________________________________________________________________________\nlambda_24 (Lambda)              (None, 20)           0           concatenate_24[0][0]             \n__________________________________________________________________________________________________\ndense_24 (Dense)                (None, 101)          2121        lambda_24[0][0]                  \n==================================================================================================\n<\/code><\/pre>\n\n<p>This is the way the data is transformed into a genetator :<\/p>\n\n<pre><code>generatedData=generate_data(new_corpus, window_size, V1)\n\nfor x, y in generatedData:\n    print(x,y)\n<\/code><\/pre>\n\n<p>Fisrt element represents the context.<\/p>\n\n<p>Second element represents the id of the document with a one hot vector.<\/p>\n\n<p>Third element represents the predicted word as a one hot vector.<\/p>\n\n<p>These are two first lines<\/p>\n\n<pre><code>[array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 4]]), array([[0., 0., 0., ..., 0., 0., 0.]], dtype=float32)] [[... 0. 0. 0. 0. 1. ... 0. 0. 0. 0.]]\n[array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 23]]), array([[0., 0., 0., ..., 0., 0., 0.]], dtype=float32)] [[... 0. 0. 1. .... 0. 0. 0. 0.]]\n...\n<\/code><\/pre>\n\n<p>Then I fit the model :<\/p>\n\n<pre><code>cbow.fit_generator(generatedData,steps_per_epoch=steps_per_epoch,epochs=epochs,shuffle=shuffle)\n<\/code><\/pre>\n\n<p>The model converges.<\/p>\n\n<pre><code>E1=pd.DataFrame(cbow.layers[2].get_weights()[0]) # words embedding\nE2=pd.DataFrame(cbow.layers[3].get_weights()[0]) # docs embedding\n<\/code><\/pre>\n\n<p>Documents are ultimately classified as random ...<\/p>\n","answers":[]}
+{"QuestionId":52625358,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network><tf-idf>","CreationDate":"2018-10-03T10:46:24.363","AcceptedAnswerId":null,"OwnerUserId":8047559.0,"Title":"Getting an error while giving input to Conv1D layer in a Keras model","Body":"<p>I am using tf-idf vector data as an input for my Keras model. tf-idf vectors has the following shape:<\/p>\n\n<pre><code>&lt;class 'scipy.sparse.csr.csr_matrix'&gt; (25000, 310617)\n<\/code><\/pre>\n\n<p>Code:<\/p>\n\n<pre><code>inputs = Input((X_train.shape[1],))\nconvnet1=Conv1D(128,3,padding='same',activation='relu')(inputs)\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer conv1d_25: expected ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>When I am converting the input to <code>Input(None,X_train.shape[1],)<\/code> then I am getting an error while fitting because the input dimension has been changed to 3.<\/p>\n","answers":[{"AnswerId":"52635348","CreationDate":"2018-10-03T20:40:24.427","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As mentioned in the error (i.e. <code>expected ndim=3, found ndim=2<\/code>), <code>Conv1D<\/code> takes a 3D array as input. So if you would like to feed this array to it you first need to reshape it:<\/p>\n\n<pre><code>X_train = X_train.reshape((X_train.shape[0], X_train.shape[1], 1))\n<\/code><\/pre>\n\n<p>And also set the input layer shape accordingly:<\/p>\n\n<pre><code>inputs = Input(X_train.shape[1:])\n<\/code><\/pre>\n\n<hr>\n\n<p>However, <code>Conv1D<\/code> is usually used for processing sequences (like sequence of words in a sentence) or temporal data (like timeseries of weather temperature). And that's exactly why it takes inputs with shape of <code>(num_samples, num_timesteps or sequence_len, num_features)<\/code>. Applying it on a tf-idf representation, which does not have any sequential order, may not be efficient that much. Instead, I suggest you to use a <code>Dense<\/code> layer. Or alternatively, instead of using tf-idf you can also feed the raw data (i.e. texts or sentences) directly into an <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\"><code>Embedding<\/code><\/a> layer and use <code>Conv1D<\/code> or <code>LSTM<\/code> layer(s) after it.<\/p>\n"}]}
+{"QuestionId":52625844,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><object-detection>","CreationDate":"2018-10-03T11:15:38.030","AcceptedAnswerId":null,"OwnerUserId":9493219.0,"Title":"TypeError: int() argument must be a string in tensorflow","Body":"<p>I had trained a model and generated its inference graph. when i tried to test it on a single image this error appeared to me<\/p>\n\n<pre><code>&gt; Traceback (most recent call last):   File\n&gt; \"C:\\tensorflow1\\models\\research\\object_detection\\Object_detection_image.py\",\n&gt; line 98, in &lt;module&gt;\n&gt;     feed_dict={image_tensor: image_expanded})   File \"C:\\Users\\Hamada Fathy\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\client\\session.py\",\n&gt; line 877, in run\n&gt;     run_metadata_ptr)   File \"C:\\Users\\Hamada Fathy\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\client\\session.py\",\n&gt; line 1069, in _run\n&gt;     np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)   File \"C:\\Users\\Hamada\n&gt; Fathy\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\numpy\\core\\numeric.py\",\n&gt; line 492, in asarray\n&gt;     return array(a, dtype, copy=False, order=order) TypeError: int() argument must be a string, a bytes-like object or a number, not\n&gt; 'NoneType'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52627196,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-03T12:36:37.337","AcceptedAnswerId":null,"OwnerUserId":7584763.0,"Title":"How to use multiple keras models in a single thread?","Body":"<p>I want to use pretrained and saved keras models for prediction. I trained and saved the models to json files in different threads with:<\/p>\n\n<pre><code>model_json = model.to_json()\nwith open(\"model.json\", \"w\") as json_file:\njson_file.write(model_json)\n<\/code><\/pre>\n\n<p>When I tried to use these models, I need to read them one by one and get prediction results. The first one works well. However when it comes to the second model, I read it from json file and run <code>model.predict(dataToPredict)<\/code>,\nit returns an error indicating that the layers are not the element of the graph. Here is the error:<\/p>\n\n<p>TypeError: Cannot interpret feed_dict key as Tensor: Tensor Tensor(\"conv_layer1_input:0\", shape=(?, 1182, 1), dtype=float32) is not an element of this graph.<\/p>\n\n<p>I used <\/p>\n\n<pre><code>json_file = open(path, 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nwith tf.Session():\n    model = model_from_json(loaded_model_json)\n    predictionResult = model.predict(dataToPredict)\n<\/code><\/pre>\n\n<p>for each model and it didn't raise the error this time. However, it seems <strong>weights of the models are randomly generated, not read from the file<\/strong> or I initialized them by mistake with this <code>with tf.Session():<\/code>. That's why they produce different results every time I run them with the same data. I also tried using:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n<\/code><\/pre>\n\n<p>without <code>with tf.Session():<\/code> but it raised the same error again.\nI am confused about \"graph\" and \"session\" terms. I thought keras manages both graph and session. If so, what is the problem? Here is my prediction code:<\/p>\n\n<pre><code>predictions = []\nfor modelFilePath in modelFilePaths:\nwith tf.Graph().as_default():\n    json_file = open(modelFilePath , 'r')\n    loaded_model_json = json_file.read()\n    json_file.close()\n    model = model_from_json(loaded_model_json)\n    predictions.append(model.predict(dataToPredict))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52627519,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T12:55:05.343","AcceptedAnswerId":null,"OwnerUserId":2956077.0,"Title":"Batch Prediction in Tensorflow session run","Body":"<p>I am pretty new to Tensorflow and I would like to know when I load a frozen graph for prediction, is it possible to do a batch prediction like in keras.<\/p>\n\n<p>My code looks like the following in Tensorflow :<\/p>\n\n<pre><code>path='some_path.pb'\nwith dgraph.as_default():\n     od_graph_def=tf.GraphDef()\n     with tf.gfile.GFile(path,'rb') as fid:\n          serialized_graph=fid.read()\n          od_graph_def.ParseFromString(serialized_graph)\n          tf.import_graph_def(od_graph_def,name='')\nsess = tf.Session(graph=dgraph)   \n\nfeat_5=detection_graph.get_tensor_by_name('Relu6:0')\ninput=np.random.randn(100,300,300,3)\n\nsess.run([feat_5],feed_dict={image_tensor: input}) **want to add batches of data here instead of input directly\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52627564,"AnswerCount":1,"Tags":"<deep-learning><classification><pytorch>","CreationDate":"2018-10-03T12:57:05.873","AcceptedAnswerId":null,"OwnerUserId":5223033.0,"Title":"What should be the loss function for classification problem in pytorch if sigmoid is used in the output layer","Body":"<p>I am trying to implement a model for binary classification problem. Up to now, I was using softmax function (at the output layer) together with torch.NLLLoss function to calculate the loss. However, now I want to use the sigmoid function (instead of softmax) at the output layer. If I do that, should I also change the loss function (to BCELoss or <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.functional.binary_cross_entropy\" rel=\"nofollow noreferrer\">binary_cross_entropy<\/a>) or may I still use torch.NLLLoss function ?<\/p>\n","answers":[{"AnswerId":"52633402","CreationDate":"2018-10-03T18:19:21.447","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>If you use <code>sigmoid<\/code> function, then you can only do binary classification. It's not possible to do a multi-class classification. The reason for this is because <code>sigmoid<\/code> function always returns a value in the range between 0 and 1. So, for instance one can threshold the value at 0.5 and separate (or classify) it into two classes based on the obtained values.<\/p>\n\n<p>Regarding the objective function <code>NLLLoss<\/code> - Negative Log Likelihood Loss. It just learns the data distribution. So, it's not a problem as long as that what you're trying to achieve during training.<\/p>\n"}]}
+{"QuestionId":52627739,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-03T13:07:04.693","AcceptedAnswerId":52629902.0,"OwnerUserId":4729764.0,"Title":"How to Merge Numerical and Embedding Sequential Models to treat categories in RNN","Body":"<p>I would like to build a one layer LSTM model with embeddings for my categorical features. I currently have numerical features and a few categorical features, such as Location, which can't be one-hot encoded e.g. using <code>pd.get_dummies()<\/code> due to computational complexity, which is what I originally intended to do. <\/p>\n\n<p>Let's visualise an example:<\/p>\n\n<h3>Sample Data<\/h3>\n\n<pre><code>data = {\n    'user_id': [1,1,1,1,2,2,3],\n    'time_on_page': [10,20,30,20,15,10,40],\n    'location': ['London','New York', 'London', 'New York', 'Hong Kong', 'Tokyo', 'Madrid'],\n    'page_id': [5,4,2,1,6,8,2]\n}\nd = pd.DataFrame(data=data)\nprint(d)\n   user_id  time_on_page   location  page_id\n0        1            10     London        5\n1        1            20   New York        4\n2        1            30     London        2\n3        1            20   New York        1\n4        2            15  Hong Kong        6\n5        2            10      Tokyo        8\n6        3            40     Madrid        2\n<\/code><\/pre>\n\n<p>Let's look at the person visiting a website. I'm tracking numerical data such as time on page and others. Categorical data includes: Location (over 1000 uniques), Page_id (> 1000 uniques), Author_id (100+ uniques). The simplest solution would be to one-hot encoding everything and put this into LSTM with variable sequence lengths, each timestep corresponding to a different page view.<\/p>\n\n<p>The above DataFrame will generate 7 training samples, with variable sequence lengths. For example, for <code>user_id=2<\/code> I will have 2 training samples:<\/p>\n\n<pre><code>[ ROW_INDEX_4 ] and [ ROW_INDEX_4, ROW_INDEX_5 ]\n<\/code><\/pre>\n\n<p>Let <code>X<\/code> be the training data, and let's look at the first training sample <code>X[0]<\/code>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/uAATZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uAATZ.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From the picture above, my categorical features are <code>X[0][:, n:]<\/code>.<\/p>\n\n<p>Before creating sequences, I factorized the categorical variables into <code>[0,1... number_of_cats-1]<\/code>, using <code>pd.factorize()<\/code> so the data in <code>X[0][:, n:]<\/code> is numbers corresponding to their index. <\/p>\n\n<p>Do I need to create an <code>Embedding<\/code> for each of the Categorical Features separately? E.g. an embedding for each of <code>x_*n, x_*n+1, ..., x_*m<\/code>?<\/p>\n\n<p>If so, how do I put this into Keras code?<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Embedding(?, ?, input_length=variable)) # How do I feed the data into this embedding? Only the categorical inputs.\n\nmodel.add(LSTM())\nmodel.add(Dense())\nmodel.add.Activation('sigmoid')\nmodel.compile()\n\nmodel.fit_generator() # fits the `X[i]` one by one of variable length sequences.\n<\/code><\/pre>\n\n<p><strong>My solution idea:<\/strong><\/p>\n\n<p>Something that looks like:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FVNWi.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FVNWi.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I can train a Word2Vec model on every single categorical feature (m-n) to vectorise any given value. E.g. London will be vectorised in 3 dimensions. Let's suppose I use 3 dimensional embeddings. Then I will put everything back into the X matrix, which will now have n + 3(n-m), and use the LSTM model to train it? <\/p>\n\n<p>I just think there should be an easier\/smarter way.<\/p>\n","answers":[{"AnswerId":"52629902","CreationDate":"2018-10-03T14:54:30.507","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One solution, as you mentioned, is to one-hot encode the categorical data (or even use them as they are, in index-based format) and feed them along the numerical data to an LSTM layer. Of course, you can also have two LSTM layers here, one for processing the numerical data and another for processing categorical data (in one-hot encoded format or index-based format) and then merge their outputs.<\/p>\n\n<p>Another solution is to have one separate embedding layer for each of those categorical data. Each embedding layer may have its own embedding dimension (and as suggested above, you may have more than one LSTM layer for processing numerical and categorical features separately):<\/p>\n\n<pre><code>num_cats = 3 # number of categorical features\nn_steps = 100 # number of timesteps in each sample\nn_numerical_feats = 10 # number of numerical features in each sample\ncat_size = [1000, 500, 100] # number of categories in each categorical feature\ncat_embd_dim = [50, 10, 100] # embedding dimension for each categorical feature\n\nnumerical_input = Input(shape=(n_steps, n_numerical_feats), name='numeric_input')\ncat_inputs = []\nfor i in range(num_cats):\n    cat_inputs.append(Input(shape=(n_steps,1), name='cat' + str(i+1) + '_input'))\n\ncat_embedded = []\nfor i in range(num_cats):\n    embed = TimeDistributed(Embedding(cat_size[i], cat_embd_dim[i]))(cat_inputs[i])\n    cat_embedded.append(embed)\n\ncat_merged = concatenate(cat_embedded)\ncat_merged = Reshape((n_steps, -1))(cat_merged)\nmerged = concatenate([numerical_input, cat_merged])\nlstm_out = LSTM(64)(merged)\n\nmodel = Model([numerical_input] + cat_inputs, lstm_out)\nmodel.summary()\n<\/code><\/pre>\n\n<p>Here is the model summary:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ncat1_input (InputLayer)         (None, 100, 1)       0                                            \n__________________________________________________________________________________________________\ncat2_input (InputLayer)         (None, 100, 1)       0                                            \n__________________________________________________________________________________________________\ncat3_input (InputLayer)         (None, 100, 1)       0                                            \n__________________________________________________________________________________________________\ntime_distributed_1 (TimeDistrib (None, 100, 1, 50)   50000       cat1_input[0][0]                 \n__________________________________________________________________________________________________\ntime_distributed_2 (TimeDistrib (None, 100, 1, 10)   5000        cat2_input[0][0]                 \n__________________________________________________________________________________________________\ntime_distributed_3 (TimeDistrib (None, 100, 1, 100)  10000       cat3_input[0][0]                 \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 100, 1, 160)  0           time_distributed_1[0][0]         \n                                                                 time_distributed_2[0][0]         \n                                                                 time_distributed_3[0][0]         \n__________________________________________________________________________________________________\nnumeric_input (InputLayer)      (None, 100, 10)      0                                            \n__________________________________________________________________________________________________\nreshape_1 (Reshape)             (None, 100, 160)     0           concatenate_1[0][0]              \n__________________________________________________________________________________________________\nconcatenate_2 (Concatenate)     (None, 100, 170)     0           numeric_input[0][0]              \n                                                                 reshape_1[0][0]                  \n__________________________________________________________________________________________________\nlstm_1 (LSTM)                   (None, 64)           60160       concatenate_2[0][0]              \n==================================================================================================\nTotal params: 125,160\nTrainable params: 125,160\nNon-trainable params: 0\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>Yet there is another solution which you can try: just have one embedding layer for all the categorical features. It involves some preprocessing though: you need to re-index all the categories to make them distinct from each other. For example, the categories in first categorical feature would be numbered from 1 to <code>size_first_cat<\/code> and then the categories in the second categorical feature would be numbered from <code>size_first_cat + 1<\/code> to <code>size_first_cat + size_second_cat<\/code> and so on. However, in this solution all the categorical features would have the same embedding dimension since we are using only one embedding layer. <\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong> Now that I think about it, you can also reshape the categorical features in data preprocessing stage or even in the model to get rid of <code>TimeDistributed<\/code> layers and the <code>Reshape<\/code> layer (and this may increase the training speed as well):<\/p>\n\n<pre><code>numerical_input = Input(shape=(n_steps, n_numerical_feats), name='numeric_input')\ncat_inputs = []\nfor i in range(num_cats):\n    cat_inputs.append(Input(shape=(n_steps,), name='cat' + str(i+1) + '_input'))\n\ncat_embedded = []\nfor i in range(num_cats):\n    embed = Embedding(cat_size[i], cat_embd_dim[i])(cat_inputs[i])\n    cat_embedded.append(embed)\n\ncat_merged = concatenate(cat_embedded)\nmerged = concatenate([numerical_input, cat_merged])\nlstm_out = LSTM(64)(merged)\n\nmodel = Model([numerical_input] + cat_inputs, lstm_out)\n<\/code><\/pre>\n\n<p>As for fitting the model, you need to feed each input layer separately with its own corresponding numpy array, for example:<\/p>\n\n<pre><code>X_tr_numerical = X_train[:,:,:n_numerical_feats]\n\n# extract categorical features: you can use a for loop to this as well.\n# note that we reshape categorical features to make them consistent with the updated solution\nX_tr_cat1 = X_train[:,:,cat1_idx].reshape(-1, n_steps) \nX_tr_cat2 = X_train[:,:,cat2_idx].reshape(-1, n_steps)\nX_tr_cat3 = X_train[:,:,cat3_idx].reshape(-1, n_steps)\n\n# don't forget to compile the model ...\n\n# fit the model\nmodel.fit([X_tr_numerical, X_tr_cat1, X_tr_cat2, X_tr_cat3], y_train, ...)\n\n# or you can use input layer names instead\nmodel.fit({'numeric_input': X_tr_numerical,\n           'cat1_input': X_tr_cat1,\n           'cat2_input': X_tr_cat2,\n           'cat3_input': X_tr_cat3}, y_train, ...)\n<\/code><\/pre>\n\n<p>If you would like to use <code>fit_generator()<\/code> there is no difference:<\/p>\n\n<pre><code># if you are using a generator\ndef my_generator(...):\n\n    # prep the data ...\n\n    yield [batch_tr_numerical, batch_tr_cat1, batch_tr_cat2, batch_tr_cat3], batch_tr_y\n\n    # or use the names\n    yield {'numeric_input': batch_tr_numerical,\n           'cat1_input': batch_tr_cat1,\n           'cat2_input': batch_tr_cat2,\n           'cat3_input': batch_tr_cat3}, batch_tr_y\n\nmodel.fit_generator(my_generator(...), ...)\n\n# or if you are subclassing Sequence class\nclass MySequnece(Sequence):\n    def __init__(self, x_set, y_set, batch_size):\n        # initialize the data\n\n    def __getitem__(self, idx):\n        # fetch data for the given batch index (i.e. idx)\n\n        # same as the generator above but use `return` instead of `yield`\n\nmodel.fit_generator(MySequence(...), ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52628101,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T13:25:55.210","AcceptedAnswerId":null,"OwnerUserId":8380638.0,"Title":"Discard dataset iterators when saving model in Tensorflow","Body":"<p>I am trying to save my model using the following instruction:<\/p>\n\n<pre><code>saver = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Vgg16'))  # Init saver\n<\/code><\/pre>\n\n<p>My model is a VGG16 CNN network which layers are organised under the Vgg16 scope:\n<a href=\"https:\/\/i.stack.imgur.com\/9QMHJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9QMHJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Problem is that after recovering the model I still can see the DataPipe module (between others). What I am doing wrong?\nThese is the code I use to recover the trained model:<\/p>\n\n<pre><code>checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)\nsaver = tf.train.import_meta_graph(\"{}.meta\".format(checkpoint_file))\nwith tf.Session() as sess:\n    saver.restore(sess, checkpoint_file)\n<\/code><\/pre>\n\n<p>Any idea?<\/p>\n","answers":[]}
+{"QuestionId":52628270,"AnswerCount":3,"Tags":"<python><pytorch><torchvision>","CreationDate":"2018-10-03T13:35:08.173","AcceptedAnswerId":52784628.0,"OwnerUserId":8085890.0,"Title":"Is there any way I can download the pre-trained models available in PyTorch to a specific path?","Body":"<p>I am referring to the models that can be found here: <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/models.html#torchvision-models\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/torchvision\/models.html#torchvision-models<\/a> <\/p>\n","answers":[{"AnswerId":"52784628","CreationDate":"2018-10-12T17:44:21.150","ParentId":null,"OwnerUserId":"8085890","Title":null,"Body":"<p>As, <a href=\"https:\/\/stackoverflow.com\/users\/3607203\/dennlinger\">@dennlinger<\/a> mentioned in his <a href=\"https:\/\/stackoverflow.com\/a\/52631758\/8085890\">answer<\/a> : <a href=\"https:\/\/pytorch.org\/docs\/stable\/model_zoo.html#module-torch.utils.model_zoo\" rel=\"nofollow noreferrer\"><code>torch.utils.model_zoo<\/code><\/a>, is being internally called when you load a pre-trained model.<\/p>\n\n<p>More specifically, the method: <code>torch.utils.model_zoo.load_url()<\/code> is being called every time a pre-trained model is loaded. The documentation for the same, mentions:<\/p>\n\n<blockquote>\n  <p>The default value of <code>model_dir<\/code> is <code>$TORCH_HOME\/models<\/code> where\n  <code>$TORCH_HOME<\/code> defaults to <code>~\/.torch<\/code>. <\/p>\n  \n  <p>The default directory can be overridden with the <code>$TORCH_HOME<\/code>\n  environment variable.<\/p>\n<\/blockquote>\n\n<p>This can be done as follows:<\/p>\n\n<pre><code>import torch \nimport torchvision\nimport os\n\n# Suppose you are trying to load pre-trained resnet model in directory- models\\resnet\n\nos.environ['TORCH_HOME'] = 'models\\\\resnet' #setting the environment variable\nresnet = torchvision.models.resnet18(pretrained=True)\n<\/code><\/pre>\n\n<p>I came across the above solution by raising an issue in the PyTorch's GitHub repository:\n<a href=\"https:\/\/github.com\/pytorch\/vision\/issues\/616\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/vision\/issues\/616<\/a><\/p>\n\n<p>This led to an improvement in the documentation i.e. the solution mentioned above.<\/p>\n"},{"AnswerId":"52633236","CreationDate":"2018-10-03T18:07:33.990","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Yes, you can simply copy the urls and use <code>wget<\/code> to download it to the desired path. Here's an illustration:<\/p>\n\n<p>For <strong>AlexNet<\/strong>:<\/p>\n\n<pre><code>$ wget -c https:\/\/download.pytorch.org\/models\/alexnet-owt-4df8aa71.pth\n<\/code><\/pre>\n\n<p>For <strong>Google Inception (v3)<\/strong>:<\/p>\n\n<pre><code>$ wget -c https:\/\/download.pytorch.org\/models\/inception_v3_google-1a9a5a14.pth\n<\/code><\/pre>\n\n<p>For <strong>SqueezeNet<\/strong>:<\/p>\n\n<pre><code>$ wget -c https:\/\/download.pytorch.org\/models\/squeezenet1_1-f364aa15.pth\n<\/code><\/pre>\n\n<hr>\n\n<p>If you want to do it in Python, then use something like:<\/p>\n\n<pre><code>In [11]: from six.moves import urllib\n\n# resnet 101 host url\nIn [12]: url = \"https:\/\/download.pytorch.org\/models\/resnet101-5d3b4d8f.pth\"\n\n# download and rename the file to `resnet_101.pth`\nIn [13]: urllib.request.urlretrieve(url, \"resnet_101.pth\")\nOut[13]: ('resnet_101.pth', &lt;http.client.HTTPMessage at 0x7f7fd7f53438&gt;)\n<\/code><\/pre>\n\n<p>P.S: You can find all the download URLs in the individual python files in <a href=\"https:\/\/github.com\/pytorch\/vision\/tree\/master\/torchvision\/models\" rel=\"nofollow noreferrer\">torchvision.models<\/a><\/p>\n"},{"AnswerId":"52631758","CreationDate":"2018-10-03T16:31:44.287","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>TL;DR: No, it is not possible directly, but you can easily adapt it.<\/p>\n\n<p>I think what you want to do is to look at <a href=\"https:\/\/pytorch.org\/docs\/stable\/model_zoo.html#module-torch.utils.model_zoo\" rel=\"nofollow noreferrer\"><code>torch.utils.model_zoo<\/code><\/a>, which is internally called when you load a pre-trained model:<\/p>\n\n<p>If we look at the code for the pre-trained models, for example AlexNet <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/models\/alexnet.py\" rel=\"nofollow noreferrer\">here<\/a>, we can see that it simply calls the previously mentioned <code>model_zoo<\/code> function, but without the saved location. You can either modify the PyTorch source to specify this (that would actually be a great addition IMO, so maybe open a pull request for that), or else simply adopt the code in the second link to your own liking (and save it to a custom location under a different name), and then manually insert the relevant location there.<\/p>\n\n<p>If you want to regularly update PyTorch, I would heavily recommend the second method, since it doesn't involve directly altering PyTorch's code base, and potentially throw errors during updates.<\/p>\n"}]}
+{"QuestionId":52628487,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning><keras-layer>","CreationDate":"2018-10-03T13:45:23.197","AcceptedAnswerId":null,"OwnerUserId":6643984.0,"Title":"Changing keras model input name","Body":"<p>When i'm creating model ensemble consisting of 6 models in keras i do the following:<\/p>\n\n<pre><code>out = keras.layers.Average(name=\"output\")(outputs) #6 models\nmodel = Model(input, out, name='ensemble')\n<\/code><\/pre>\n\n<p>i would like to set the input to be <\/p>\n\n<pre><code>input = Input(SHAPE, name='my_input')\n<\/code><\/pre>\n\n<p>but then i get the error message because from what i understand when making an avarage model it's input was automatically set to 'input_6', so i get an error: <strong>Graph disconnected: cannot obtain value for tensor Tensor(\"input_6:0\", shape=(?, 256, 256, 1), dtype=float32)<\/strong>. Is there a way to change model input name?<\/p>\n","answers":[]}
+{"QuestionId":52629410,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-03T14:31:06.503","AcceptedAnswerId":null,"OwnerUserId":4450159.0,"Title":"tf.dynamic_partition get non-empty tensors","Body":"<p>I'm using <code>tf.dynamic_partition<\/code> to get a dynamic split of a tensor, which can not infer the size. Using the answer (Choice 2) of this  <a href=\"https:\/\/stackoverflow.com\/questions\/45404056\/tf-unstack-with-dynamic-shape\">tf.unstack with dynamic shape<\/a> , I create this experimental code, where the idea is that after processing with <code>tf.dynamic_partition<\/code> create some sort of filtering when the tensor has <code>tf.size<\/code> == 0, so a new list would be created that obtains only the non-empty tensors <code>tf.size<\/code>> 0<\/p>\n\n<p>I need to have the list of non-empty tensors at the moment of creating the operations, because afterwards I need to create new operations on this list.<\/p>\n\n<pre><code>import tensorflow as tf\n\n\ndata = tf.placeholder(tf.float32, (None, 2))\npartitions = tf.placeholder(tf.int32, (None))\nmax_partitions = 10\n\ndp = tf.dynamic_partition(\n    data,\n    partitions,\n    max_partitions,\n    name=\"DynamicPartition\"\n)\nnew_dp = []\nfor tensor in dp:\n    with tf.control_dependencies([tf.assert_rank_at_least(tensor, 1), tf.assert_greater(tf.size(tensor), 0)]):\n          new_dp.append(tensor)\n\narray = [[1,1],\n         [1,1],\n         [1,1],\n         [1,1],\n         [1,1],\n         [2,2],\n         [2,2],\n         [2,2]]\norder = [1,1,1,1,1,0,0,0]\n\n\nwith tf.Session() as sess:\n    result = sess.run([dp], \n                      feed_dict={\n                        data:array,\n                        partitions:order\n                      })\n    print(result)\n<\/code><\/pre>\n\n<p>The problem is that this experiment is not working, because the output is the list created by <code>tf.dynamic_partition<\/code> in which you can see all the tensors, both the non-empty and the empty<\/p>\n\n<pre><code>[[array([[2., 2.],\n   [2., 2.],\n   [2., 2.]], dtype=float32), array([[1., 1.],\n   [1., 1.],\n   [1., 1.],\n   [1., 1.],\n   [1., 1.]], dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32), array([], shape=(0, 2), dtype=float32)]]\n<\/code><\/pre>\n\n<p>Does anyone have a solution for this problem?<\/p>\n","answers":[]}
+{"QuestionId":52630376,"AnswerCount":1,"Tags":"<tensorflow><mnist><tensorflow-datasets>","CreationDate":"2018-10-03T15:17:22.410","AcceptedAnswerId":null,"OwnerUserId":3837778.0,"Title":"TensorFlow IOError: No route to host","Body":"<p>I am trying the following codes in Juypter notebook:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.examples.tutorials.mnist.input_data as input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n<\/code><\/pre>\n\n<p>but got the following error:<\/p>\n\n<pre><code>Extracting MNIST_data\/train-images-idx3-ubyte.gz\n---------------------------------------------------------------------------\nIOError                                   Traceback (most recent call last)\n&lt;ipython-input-9-aeeb0aa14838&gt; in &lt;module&gt;()\n      3 import tensorflow.examples.tutorials.mnist.input_data as input_data\n----&gt; 4 mnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n      5 print(\"MNIST downloaded.\")\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/util\/deprecation.pyc in new_func(*args, **kwargs)\n    248               'in a future version' if date is None else ('after %s' % date),\n    249               instructions)\n--&gt; 250       return func(*args, **kwargs)\n    251     return tf_decorator.make_decorator(\n    252         func, new_func, 'deprecated',\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.pyc in read_data_sets(train_dir, fake_data, one_hot, dtype, reshape, validation_size, seed, source_url)\n    263 \n    264   local_file = base.maybe_download(TRAIN_LABELS, train_dir,\n--&gt; 265                                    source_url + TRAIN_LABELS)\n    266   with gfile.Open(local_file, 'rb') as f:\n    267     train_labels = extract_labels(f, one_hot=one_hot)\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/util\/deprecation.pyc in new_func(*args, **kwargs)\n    248               'in a future version' if date is None else ('after %s' % date),\n    249               instructions)\n--&gt; 250       return func(*args, **kwargs)\n    251     return tf_decorator.make_decorator(\n    252         func, new_func, 'deprecated',\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/base.pyc in maybe_download(filename, work_directory, source_url)\n    250   filepath = os.path.join(work_directory, filename)\n    251   if not gfile.Exists(filepath):\n--&gt; 252     temp_file_name, _ = urlretrieve_with_retry(source_url)\n    253     gfile.Copy(temp_file_name, filepath)\n    254     with gfile.GFile(filepath) as f:\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/util\/deprecation.pyc in new_func(*args, **kwargs)\n    248               'in a future version' if date is None else ('after %s' % date),\n    249               instructions)\n--&gt; 250       return func(*args, **kwargs)\n    251     return tf_decorator.make_decorator(\n    252         func, new_func, 'deprecated',\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/base.pyc in wrapped_fn(*args, **kwargs)\n    203       for delay in delays():\n    204         try:\n--&gt; 205           return fn(*args, **kwargs)\n    206         except Exception as e:  # pylint: disable=broad-except\n    207           if is_retriable is None:\n\n\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/base.pyc in urlretrieve_with_retry(url, filename)\n    231 @_internal_retry(initial_delay=1.0, max_delay=16.0, is_retriable=_is_retriable)\n    232 def urlretrieve_with_retry(url, filename=None):\n--&gt; 233   return urllib.request.urlretrieve(url, filename)\n    234 \n    235 \n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/urllib.pyc in urlretrieve(url, filename, reporthook, data, context)\n     96     else:\n     97         opener = _urlopener\n---&gt; 98     return opener.retrieve(url, filename, reporthook, data)\n     99 def urlcleanup():\n    100     if _urlopener:\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/urllib.pyc in retrieve(self, url, filename, reporthook, data)\n    243             except IOError:\n    244                 pass\n--&gt; 245         fp = self.open(url, data)\n    246         try:\n    247             headers = fp.info()\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/urllib.pyc in open(self, fullurl, data)\n    211         try:\n    212             if data is None:\n--&gt; 213                 return getattr(self, name)(url)\n    214             else:\n    215                 return getattr(self, name)(url, data)\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/urllib.pyc in open_https(self, url, data)\n    441             if realhost: h.putheader('Host', realhost)\n    442             for args in self.addheaders: h.putheader(*args)\n--&gt; 443             h.endheaders(data)\n    444             errcode, errmsg, headers = h.getreply()\n    445             fp = h.getfile()\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/httplib.pyc in endheaders(self, message_body)\n   1036         else:\n   1037             raise CannotSendHeader()\n-&gt; 1038         self._send_output(message_body)\n   1039 \n   1040     def request(self, method, url, body=None, headers={}):\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/httplib.pyc in _send_output(self, message_body)\n    880             msg += message_body\n    881             message_body = None\n--&gt; 882         self.send(msg)\n    883         if message_body is not None:\n    884             #message_body was not a string (i.e. it is a file) and\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/httplib.pyc in send(self, data)\n    842         if self.sock is None:\n    843             if self.auto_open:\n--&gt; 844                 self.connect()\n    845             else:\n    846                 raise NotConnected()\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/httplib.pyc in connect(self)\n   1253             \"Connect to a host on a given (SSL) port.\"\n   1254 \n-&gt; 1255             HTTPConnection.connect(self)\n   1256 \n   1257             if self._tunnel_host:\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/httplib.pyc in connect(self)\n    819         \"\"\"Connect to the host and port specified in __init__.\"\"\"\n    820         self.sock = self._create_connection((self.host,self.port),\n--&gt; 821                                            self.timeout, self.source_address)\n    822 \n    823         if self._tunnel_host:\n\n\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/socket.pyc in create_connection(address, timeout, source_address)\n    573 \n    574     if err is not None:\n--&gt; 575         raise err\n    576     else:\n    577         raise error(\"getaddrinfo returns an empty list\")\n\nIOError: [Errno socket error] [Errno 65] No route to host\n<\/code><\/pre>\n\n<p>Any idea why? Thanks.<\/p>\n","answers":[{"AnswerId":"52631856","CreationDate":"2018-10-03T16:38:57.763","ParentId":null,"OwnerUserId":"187343","Title":null,"Body":"<p>Tensorflow downloads the mnist dataset from mnist website. Make sure your internet connection is working.<\/p>\n\n<p>If you're in a corporate network, you might have to configure proxy settings.<\/p>\n"}]}
+{"QuestionId":52631353,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T16:06:10.750","AcceptedAnswerId":null,"OwnerUserId":7181986.0,"Title":"Tensorflow get validation loss issue","Body":"<p>I have confused ideas on how to calculate the validation loss in my model. I've a problem of classifying the pixels of the image.<\/p>\n\n<p>So, after training, in validation statement, i can get my prediction image with this code:<\/p>\n\n<pre><code>output_image = sess.run(network,feed_dict={net_input:input_image})\n<\/code><\/pre>\n\n<p>But, if i run:<\/p>\n\n<pre><code> cost , output_image  = sess.run([loss,network],feed_dict={net_input: input_image, net_output: output_image})\n<\/code><\/pre>\n\n<p>I get this exception: <code>Cannot feed value of shape (1, 480, 480, 3) for Tensor 'Placeholder_1:0', which has shape '(?, ?, ?, 2)'<\/code><\/p>\n\n<p>I don't understand why, the shape of input and output images are the same (1, 480, 480, 3). \nWhat is the way to get the cost of validation image?<\/p>\n\n<p><strong>EDIT<\/strong>:\nThis is my code:<\/p>\n\n<pre><code> net_input = tf.placeholder(tf.float32,shape=[None,None,None,3])\n    net_output = tf.placeholder(tf.float32,shape=[None,None,None,num_classes]) \n\n    network, _ = model_builder.build_model(args.model, net_input=net_input, num_classes=num_classes, crop_width=args.crop_width, crop_height=args.crop_height)\n\nsess.run(tf.global_variables_initializer())\n\n\n        loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=network, labels=net_output))\n\n        opt = tf.train.RMSPropOptimizer(learning_rate=0.0001, decay=0.995).minimize(loss, var_list=[var for var in tf.trainable_variables()])\n\n        saver=tf.train.Saver(max_to_keep=1000)\n        sess.run(tf.global_variables_initializer())\n\n\n        # Set random seed to make sure models are validated on the same validation images.\n        # So you can compare the results of different models more intuitively.\n        val_indices=random.sample(range(0,len(val_input_names)),num_vals)\n\n\n        # Do the training here\n        for epoch in range(args.epoch_start_i, args.num_epochs):\n\n            current_losses = []\n\n            # Equivalent to shuffling\n            id_list = np.random.permutation( len(train_input_names) )\n\n            num_iters = int(np.floor(len(id_list) \/ args.batch_size))\n\n            #for i in range(num_iters):\n            description_train = '[i]Train Epoch {:&gt;2}\/{}'.format(epoch + 1, args.num_epochs)\n            for i in tqdm(range(num_iters), desc=description_train, unit='batch'):\n\n                input_image_batch = []\n                output_image_batch = []\n\n                # Collect a batch of images\n                for j in range(args.batch_size):\n                    index = i*args.batch_size + j\n                    id = id_list[index]\n                    input_image = utils.load_image(train_input_names[id])\n                    output_image = utils.load_image(train_output_names[id])\n\n                    with tf.device('\/cpu:0'):\n                        input_image, output_image = data_augmentation(input_image, output_image)\n\n\n                        # Prep the data. Make sure the labels are in one-hot format\n                        input_image = np.float32(input_image) \/ 255.0\n                        output_image = np.float32(helpers.one_hot_it(label=output_image, label_values=label_values))\n\n                        input_image_batch.append(np.expand_dims(input_image, axis=0))\n                        output_image_batch.append(np.expand_dims(output_image, axis=0))\n\n                if args.batch_size == 1:\n                    input_image_batch = input_image_batch[0]\n                    output_image_batch = output_image_batch[0]\n                else:\n                    input_image_batch = np.squeeze(np.stack(input_image_batch, axis=1))\n                    output_image_batch = np.squeeze(np.stack(output_image_batch, axis=1))\n\n                # Do the training\n                _,current=sess.run([opt,loss],feed_dict={net_input:input_image_batch,net_output:output_image_batch})\n\n\n                # Do the validation on a small set of validation images\n                description_val = '[i]Validazione {:&gt;2}\/{}'.format(epoch + 1, args.num_epochs)\n                loss_val = [];\n                for ind in tqdm(val_indices, total=len(val_indices), desc=description_val, unit='img'):\n\n                    input_image = np.expand_dims(np.float32(utils.load_image( val_input_names[ind] )[:args.crop_height, :args.crop_width]),axis=0)\/255.0\n\n                    output_image = np.expand_dims(np.float32(utils.load_image(val_output_names[ind])[:args.crop_height, :args.crop_width]), axis=0) \/ 255.0\n\n                   #Do the validation\n                    output_image =sess.run(network,feed_dict{net_input:input_image})\n<\/code><\/pre>\n","answers":[{"AnswerId":"55147800","CreationDate":"2019-03-13T17:19:28.027","ParentId":null,"OwnerUserId":"3766568","Title":null,"Body":"<p>Looks like the number of classes (<code>num_classes<\/code>) is two in your case. So <code>output_image<\/code> you are feeding to <code>sess.run()<\/code> as <code>net_output<\/code> should have only two channels. But in your case, you have three channels and that's why you are getting this error.<\/p>\n\n<p>Use <code>helpers.one_hot_it()<\/code> for getting a binary mask of your output image. You will have to expand dimension using <code>np.expand_dim()<\/code> to make it a batch of one image since the <code>network<\/code> accepts one batch at a time, not one image at a time.<\/p>\n\n<p>You can make use of the following code snippet to get validation loss:<\/p>\n\n<pre><code># Do the validation on a small set of validation images\ndescription_val = '[i]Validazione {:&gt;2}\/{}'.format(epoch + 1, args.num_epochs)\nloss_val = [];\nfor ind in tqdm(val_indices, total=len(val_indices), desc=description_val, unit='img'):\n    input_image = np.expand_dims(np.float32(utils.load_image( val_input_names[ind] )[:args.crop_height, :args.crop_width]),axis=0)\/255.0\n    output_image = utils.load_image(val_output_names[ind])[:args.crop_height, :args.crop_width]\n    output_one_hot = helpers.one_hot_it(output_image, label_values)\n\n    #Do the validation\n    output_image, loss_value = sess.run([network, loss], feed_dict={net_input: input_image, net_output: np.expand_dims(output_one_hot, axis=0)})\n<\/code><\/pre>\n"},{"AnswerId":"52652737","CreationDate":"2018-10-04T17:52:24.203","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>The problem lies in  <code>net_output = tf.placeholder(tf.float32,shape=[None,None,None,num_classes])<\/code>. If this is the known classification of the image then it should be of the like (if those are already logits):<\/p>\n\n<p><code>net_output = tf.placeholder(tf.float32,shape=[None, num_classes])<\/code><\/p>\n\n<p>To train your network you would do:<\/p>\n\n<pre><code>_, eval_loss = sess.run([opt, loss], feed_dict={net_input: input_image_batch, net_output:output_image_batch}})`\n<\/code><\/pre>\n\n<p>because you want to run the optimizer you have <code>opt<\/code> and because you want to monitor the loss you do <code>loss<\/code>.<\/p>\n\n<p>To get the classification of an image you would do (after training):<\/p>\n\n<p><code>classified_images = sess.run([network], feed_dict={net_input: input_image_batch})<\/code><\/p>\n"}]}
+{"QuestionId":52632253,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2018-10-03T17:05:51.373","AcceptedAnswerId":null,"OwnerUserId":9266457.0,"Title":"Error in py_call_impl(callable, dots$args, dots$keywords)","Body":"<p>I am working with keras library in R, and we obtain the following error:<\/p>\n\n<pre><code>Error in py_call_impl(callable, dots$args, dots$keywords): \nTypeError: Failed to convert object of type &lt;class 'keras.engine.sequential.Sequential'&gt; to Tensor. Contents: &lt;keras.engine.sequential.Sequential object at 0x000000005491C2E8&gt;. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>The source code in which I get this failure is as follows:<\/p>\n\n<pre><code>conv_base &lt;- application_vgg16(include_top = FALSE, input_shape = c(150, 150, 3))\n<\/code><\/pre>\n\n<p>Anybody could help me to fix it??\n   Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":52632263,"AnswerCount":1,"Tags":"<pytorch><reinforcement-learning>","CreationDate":"2018-10-03T17:06:32.773","AcceptedAnswerId":null,"OwnerUserId":9679270.0,"Title":"Why isn't my RL model behaving the same after being loaded in pytorch?","Body":"<p>I'm training some simple neural networks for Reinforcement Learning in Pytorch. At the end of the training, I save the model like so:<\/p>\n\n<pre><code>torch.save(self.policy_NN.state_dict(), self.model_fname)\n<\/code><\/pre>\n\n<p>It's doing pretty good at this point. Then later, in another script, I load it again, like so:<\/p>\n\n<pre><code>self.policy_NN.load_state_dict(torch.load(model_fname))\n<\/code><\/pre>\n\n<p>And then just play out the episode, as if the training never stopped (except I'm not doing DQN learning anymore, it's just taking the greedy action at each point). So I'd expect it to behave basically as it did when I saved it. <\/p>\n\n<p>However, whenever I load it, it behaves completely differently, to the point that it seems like it didn't learn at all before I saved it. For example, if I look at the last 1000 time steps of the training session, it will get many rewards, but after loading it, it gets basically none.<\/p>\n\n<p>I've verified (by doing <code>print(self.policy_NN.state_dict())<\/code>) that the weights and biases are in fact the same when I save the model and when I load it again.<\/p>\n\n<p>What could be going on? Is there something else to the network that might not be getting saved somehow?<\/p>\n","answers":[{"AnswerId":"52644008","CreationDate":"2018-10-04T10:05:27.040","ParentId":null,"OwnerUserId":"3419728","Title":null,"Body":"<p>Dropout and some other layers behave differently in <code>eval<\/code> and <code>train<\/code> mode. You can switch between these by <code>model.train()<\/code> and <code>model.eval()<\/code>.<\/p>\n\n<p>I remember reading that RL typically suffers from brittle learning where changing the input even slightly causes wildly different performance. The example was when training an algorithm on an Atari game and simply by shifting the screen right by one pixel the network lost all performance gains.<\/p>\n\n<p>You might want to check that in both modes your environment behaves similarly.<\/p>\n"}]}
+{"QuestionId":52632301,"AnswerCount":0,"Tags":"<python><tensorflow><memory-management><cuda><anaconda>","CreationDate":"2018-10-03T17:08:59.480","AcceptedAnswerId":null,"OwnerUserId":10367554.0,"Title":"Can I program to avoid GPU memory errors? like cuMemAlloc results in CUDA_ERROR_LAUNCH_FAILED","Body":"<p>I am doing some fun testing of my GPU Nvidia video card.   Set up a nice program that does operations on the intel CPU and then does them by vectorizing them on the GPU.   It is cool, because I get about an 80x speed improvement when I use the GPU.   However, with some simple programs, I have been able to just blow out the memory constraint and one of my programs, trashes the kernal. <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom timeit import default_timer as timer\nfrom numba import vectorize\nimport gc\n@vectorize(['float32(float32, float32)'], target='cuda')\ndef loopGPU(vStart, vEnd):\n    vPosition=vStart\n    vOriginal=vStart\n    for vStart in range(vEnd):\n        vPosition=vPosition+1\n    print('Counted from: ', vOriginal, ' to: ', vEnd, ' by: 1 ' )\n    return(vPosition)\ndef loop2aTrillion(vStart, vEnd):\n    start = timer()\n    vNewStart=loopGPU(vStart, vEnd)\n    duration = timer() - start\n    print('GPU Loop Time',duration)\n<\/code><\/pre>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>loop2aTrillion(1,100000000)\n      Counted from:  1.000000  to:  100000000.000000  by: 1\n      GPU Loop Time 3.7991681657728336<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>So it works great to about 100 million..   But If I go to a Billion or even 500 million..  I get this amazing error from CUDA<\/p>\n\n<p>See my try to see where the breaking point is..<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>loop2aTrillion(1,100)\n      Counted from:  1.000000  to:  100.000000  by: 1\n      GPU Loop Time 0.10281342493556167\n      loop2aTrillion(1,1000000)\n      Counted from:  1.000000  to:  1000000.000000  by: 1\n      GPU Loop Time 0.048263507262902294\n      loop2aTrillion(1,100000000)\n      Counted from:  1.000000  to:  100000000.000000  by: 1\n      GPU Loop Time 3.7804056272377693\n      loop2aTrillion(1,100000000)\n      Counted from:  1.000000  to:  100000000.000000  by: 1\n      GPU Loop Time 3.797257584627573\n      loop2aTrillion(1,1000000000)\n      Traceback (most recent call last):\n        File \"\", line 1, in \n        File \"\", line 3, in loop2aTrillion\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\dispatcher.py\", line 88, in <strong>call<\/strong>\n          return CUDAUFuncMechanism.call(self.functions, args, kws)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\npyufunc\\deviceufunc.py\", line 311, in call\n          return devout.copy_to_host().reshape(outshape)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\cudadrv\\devices.py\", line 212, in _require_cuda_context\n          return fn(*args, **kws)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\cudadrv\\devicearray.py\", line 252, in copy_to_host\n          _driver.device_to_host(hostary, self, self.alloc_size, stream=stream)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\cudadrv\\driver.py\", line 1776, in device_to_host\n          fn(host_pointer(dst), device_pointer(src), size, *varargs)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\cudadrv\\driver.py\", line 288, in safe_cuda_api_call\n          self._check_error(fname, retcode)\n        File \"D:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\numba\\cuda\\cudadrv\\driver.py\", line 323, in _check_error\n          raise CudaAPIError(retcode, msg)\n      numba.cuda.cudadrv.driver.CudaAPIError: [719] Call to cuMemcpyDtoH results in CUDA_ERROR_LAUNCH_FAILED<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>What is fun about this error, is the kernal is trash and I have to start over..   \"it must be destroyed\"  love it!!!!<\/p>\n\n<p>CUDA_ERROR_LAUNCH_FAILED An exception occurred on the device while executing a kernel. Common\ncauses include dereferencing an invalid device pointer and accessing out of bounds shared memory. The\ncontext cannot be used, <strong>so it must be destroyed<\/strong> (and a new one should be created). All existing device\nmemory allocations from this context are invalid and must be reconstructed if the program is to continue\nusing CUDA.<\/p>\n\n<p>OK, so now you know my background.  Here are my questions.\n1.  How do I now how much is too much ( off line )?   Is there some type of python math or library that is good at knowing where I am memory wise while I am interactive? <\/p>\n\n<ol start=\"2\">\n<li><p>Can I check for this situation \"on line\"?    So either before I call the vectorized fucntion, can I estimate live, if I am going to blow up memory and trash my kernal?    Or can I do it in the Vectorized procedure and have more control or insight to the memory allocation?<\/p><\/li>\n<li><p>Will a Try \/ Exception block save me?   Or is the memory allocation too deep in the CUDA that once it blows up, it is gone, and the pythong wrapper cannot save the kernal..  \"It MUST BE DESTROYED\"...  <\/p><\/li>\n<\/ol>\n\n<p>Thanks, as you can see, just trying to push how many calculations I can run through on these GPUs..<\/p>\n","answers":[]}
+{"QuestionId":52632388,"AnswerCount":2,"Tags":"<tensorflow><amazon-ec2><keras><amazon-sagemaker>","CreationDate":"2018-10-03T17:14:39.113","AcceptedAnswerId":52643287.0,"OwnerUserId":5586214.0,"Title":"Is it possible to deploy a already trained Keras to Sagemaker?","Body":"<p>I was playing with the AWS instances and trying to deploy some locally trained Keras models, but I find no documentation on that. Has anyone already been able to do it? <\/p>\n\n<p>I tried to use a similar approach to <a href=\"https:\/\/aws.amazon.com\/pt\/blogs\/machine-learning\/bring-your-own-pre-trained-mxnet-or-tensorflow-models-into-amazon-sagemaker\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/pt\/blogs\/machine-learning\/bring-your-own-pre-trained-mxnet-or-tensorflow-models-into-amazon-sagemaker\/<\/a>, but I had no success. I also found some examples for training keras models in the cloud, but I was not able to get the entry_point + artifacts right. <\/p>\n\n<p>Thanks for your time!<\/p>\n","answers":[{"AnswerId":"52643287","CreationDate":"2018-10-04T09:22:37.793","ParentId":null,"OwnerUserId":"1538258","Title":null,"Body":"<p>Yes, it is possible, and yes, the official documentation is not much of help.\nHowever, I wrote an <a href=\"https:\/\/gnomezgrave.com\/2018\/07\/05\/using-a-custom-model-for-ml-inference-with-amazon-sagemaker\" rel=\"nofollow noreferrer\">article on that<\/a>, and I hope it will help you.<\/p>\n\n<p>Let me know if you need more details. Cheers!<\/p>\n"},{"AnswerId":"54880108","CreationDate":"2019-02-26T07:10:40.183","ParentId":null,"OwnerUserId":"1450312","Title":null,"Body":"<p>AWS recently released a <a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/deploy-trained-keras-or-tensorflow-models-using-amazon-sagemaker\/\" rel=\"nofollow noreferrer\">tutorial<\/a> on this exact same topic which I find very easy and quicker than the docker image route. Hope this helps.<\/p>\n"}]}
+{"QuestionId":52632567,"AnswerCount":2,"Tags":"<debugging><keras>","CreationDate":"2018-10-03T17:26:19.873","AcceptedAnswerId":null,"OwnerUserId":10336700.0,"Title":"Keras fit and fit_generator returns quite different results","Body":"<p>Keras fit and fit_generator returns quite different results, fit_generator drops precision almost 20%. I do use shuffle in the data generator. I have attached my data_generator below. Thank you !<\/p>\n\n<pre><code>def data_generator(input_x, input_y, batch_size = BATCH_SIZE):\n    loopcount = len(input_x) \/\/ batch_size\n    while True:\n        i = random.randint(0, loopcount - 1)\n        x_batch = input_x[i*batch_size:(i+1)*batch_size]\n        y_batch = input_y[i*batch_size:(i+1)*batch_size]\n        yield x_batch, y_batch\n<\/code><\/pre>\n\n<p>and my model.fit_generator is showing below:<\/p>\n\n<pre><code>    model.fit_generator(generator = data_generator(x_train, y_train, batch_size = BATCH_SIZE),steps_per_epoch = len(x_train) \/\/ BATCH_SIZE, epochs = 20, validation_data = data_generator(x_val, y_val, batch_size = BATCH_SIZE), validation_steps = len(x_val) \/\/ BATCH_SIZE)\n<\/code><\/pre>\n","answers":[{"AnswerId":"54422652","CreationDate":"2019-01-29T13:53:55.103","ParentId":null,"OwnerUserId":"949251","Title":null,"Body":"<p>It looks like your generator picks batches randomly, which will likely end up with duplicate or unused batches within the same epoch.<\/p>\n\n<p>To avoid this, you could something like this (untested):<\/p>\n\n<pre><code>    def data_generator(input_x, input_y, batch_size):\n        loopcount = len(input_x) \/\/ batch_size\n        batches = range(loopcount)\n        random.shuffle(batches)\n        i = 0\n        while True:\n            b = batches[i]\n            x_batch = input_x[b*batch_size:(b+1)*batch_size]\n            y_batch = input_y[b*batch_size:(b+1)*batch_size]\n            i += 1\n            yield x_batch, y_batch\n<\/code><\/pre>\n\n<p>However, you will lose some patches if your data amount is not a multiple of batch_size. You could code around that though and return a final batch that is smaller than batch_size.<\/p>\n\n<p>Or, you could use an Iterator:<\/p>\n\n<pre><code>class MyIterator(Iterator):\n    def __init__(self, x, y, batch_size, shuffle=True, seed=None):\n        self.x = x\n        self.y = y\n        super(MyIterator, self).__init__(x.shape[0], batch_size, shuffle, seed)\n\n    def _get_batches_of_transformed_samples(self, index_array):\n        return self.x[index_array], self.y[index_array]\n<\/code><\/pre>\n\n<p>Then start the training like this:<\/p>\n\n<pre><code>    train_iterator = MyIterator(x_train, y_train, batch_size)\n    val_iterator = MyIterator(x_val, y_val, batch_size)\n    model.fit_generator(generator=iterator, \n                        steps_per_epoch=len(train_iterator),  \n                        validation_data=val_iterator, \n                        validation_steps=len(val_iterator),\n                        epochs=20)\n<\/code><\/pre>\n\n<p>The iterator will handle the remaining patches for you, effectively creating a last batch that may be smaller than batch_size.<\/p>\n\n<p>Edit: Following the discussion over at <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2389\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/2389<\/a>, it is very important to shuffle the data when you write your own custom generators.<\/p>\n"},{"AnswerId":"52637506","CreationDate":"2018-10-04T00:52:20.073","ParentId":null,"OwnerUserId":"10443041","Title":null,"Body":"<p>Note: This is not the solution but a similar problem. I think both of these problems are related and resolving one can be the solution to other.  <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/52600127\/keras-fit-generator-and-fit-results-are-different\">fit_generator is not working properly<\/a><\/p>\n"}]}
+{"QuestionId":52632612,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-03T17:29:19.707","AcceptedAnswerId":52633773.0,"OwnerUserId":8737789.0,"Title":"ImportError: cannot import name 'relu6' and AttributeError: module 'keras.applications.mobilenet' has no attribute 'relu6'","Body":"<p>I am trying to convert a h5 keras model to a <code>.mlmodel<\/code> file type, using this code:<\/p>\n\n<pre><code>from keras.models import load_model\nimport keras\nfrom keras.applications.mobilenet import MobileNet\nfrom keras.layers import DepthwiseConv2D\n# convert the model to coreml format\nprint(\"[INFO] converting model\")\n\nfrom keras.utils.generic_utils import CustomObjectScope\n\nwith CustomObjectScope({'relu6': keras.applications.mobilenet.relu6,'DepthwiseConv2D':     keras.applications.mobilenet.DepthwiseConv2D}):\n   model = load_model('\/Users\/nikhil.c\/aslModel.h5', custom_objects={\n               'relu6': MobileNet})\ncoreml_model = coremltools.converters.keras.convert(\"\/Users\/nikhil.c        \/aslModel.h5\",\n        input_names=\"image\",\n        image_input_names=\"image\",\n        image_scale=1\/255.0,\n        class_labels= [\"hello\", \"hi\", \"you\"],\n        is_bgr=True)\n\n# save the model to disk\ncoremltools.utils.save_spec(coreml_model, 'aslModel.mlmodel')\n<\/code><\/pre>\n\n<p>I have received this error initially before using <code>CustomObjectScope<\/code>:<\/p>\n\n<pre><code>ImportError: cannot import name 'relu6'\n<\/code><\/pre>\n\n<p>I fixed it through <code>CustomObjectScope<\/code>, but now I am getting the error:<\/p>\n\n<pre><code>AttributeError: module 'keras.applications.mobilenet' has no attribute 'relu6'. \n<\/code><\/pre>\n\n<p>I do not usually post in stack overflow, so please tell me if you need more information.<\/p>\n","answers":[{"AnswerId":"52633773","CreationDate":"2018-10-03T18:43:20.127","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The code you have is for an older Keras version, the latest Keras I checked (2.2.2) has the ReLU and DepthWiseConv2D already integrated inside <code>keras.layers<\/code>, so you only need to impor this to use MobileNet:<\/p>\n\n<pre><code>import keras\nfrom keras.applications import MobileNet\n<\/code><\/pre>\n\n<p>There is also <code>MobileNetV2<\/code> in the same package for a newer version of MobileNet.<\/p>\n"}]}
+{"QuestionId":52632647,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><neural-network><conv-neural-network>","CreationDate":"2018-10-03T17:31:35.663","AcceptedAnswerId":52638112.0,"OwnerUserId":1561358.0,"Title":"How to add convolution layer to custom estimator","Body":"<p>I made a basic custom estimator for classifying short texts.<\/p>\n\n<p>It's pretty much lifted from tutorial on www.tensorflow.org, with few minor changes, mostly for easier tweaking.<\/p>\n\n<p>The feature column is categorical_column_with_hashbucket, wrapped in embedding_column.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/categorical_column_with_hash_bucket\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/categorical_column_with_hash_bucket<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/embedding_column\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/embedding_column<\/a><\/p>\n\n<pre><code>def text_classifier(\nfeatures, # This is batch_features from input_fn\nlabels,   # This is batch_labels from input_fn\nmode,     # An instance of tf.estimator.ModeKeys\nparams):  # Additional configuration\n\n# Use `input_layer` to apply the feature columns.\n\nnet = tf.feature_column.input_layer(\n    features, \n    feature_columns=params['feature_columns'])\n\nif params['dropout']:\n            net = tf.layers.dropout(net,rate=0.0001,training=True)\n# Build the hidden layers, sized according to the 'hidden_units' param.\ntry:\n    activation=params['activation']\nexcept KeyError:\n    activation=tf.nn.relu\nif 'hidden_dense_units' in params:\n    for units in params['hidden_dense_units']:\n        net = tf.layers.dense(net,\n                    units=units,\n                    activation=activation,\n                    kernel_initializer=tf.glorot_uniform_initializer()\n                    )\n        #Dropout layer\n        if params['dropout']:\n            net = tf.layers.dropout(net,rate=0.5,training=True)\n        #Normalization\n        if params['normalization']:\n            net = tf.layers.batch_normalization(\n                net,\n                momentum=0.999,\n                training=True\n                )\n\n# Compute logits (1 per class).\nlogits = tf.layers.dense(net, params['n_classes'], activation=None)\n\n# Compute predictions.\npredicted_classes = tf.argmax(logits, 1)\nif mode == tf.estimator.ModeKeys.PREDICT:\n    predictions = {\n        'class_ids': predicted_classes[:, tf.newaxis],\n        'probabilities': tf.nn.softmax(logits),\n        'logits': logits,\n    }\n    return tf.estimator.EstimatorSpec(mode, predictions=predictions)\n\n# Compute loss. \nloss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)\n\n# Compute evaluation metrics.\naccuracy = tf.metrics.accuracy(labels=labels,\n                                predictions=predicted_classes,\n                                name='acc_op')\n\nmetrics = {'accuracy': accuracy}\n\ntf.summary.scalar('accuracy', accuracy[1])\n\nif mode == tf.estimator.ModeKeys.EVAL:\n    return tf.estimator.EstimatorSpec(\n        mode, loss=loss, eval_metric_ops=metrics)\n\n#Training OP\nassert mode == tf.estimator.ModeKeys.TRAIN\n\n#Initiating optimizer\ntry:\n    optimizer = params['optimizer']\nexcept KeyError:\n    optimizer = tf.train.AdamOptimizer(learning_rate = 0.001)\n\ntrain_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n\nreturn tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)\n<\/code><\/pre>\n\n<p>After reaching, what I assumed to be, best possible performance of roughly 0.6 accuracy on testing dataset. I decided to add one dimensional convolution layer from Tensorflow library. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv1d\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv1d<\/a><\/p>\n\n<p>Unfortunately, the documentation doesn't specify, that it requires 3 dimensional tensor as an input. If I simply use the input_layer as it's input, for example like this:<\/p>\n\n<pre><code>net = tf.layers.conv1d(net,filters=1,kernel_size=1)\n<\/code><\/pre>\n\n<p>I get following exception:<\/p>\n\n<pre><code>ValueError: Input 0 of layer conv1d_1 is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: [None, 1024]\n<\/code><\/pre>\n\n<p>The 1024 is embedding dimension btw, but it doesn't matter how large it is, the process crashes anyway.<\/p>\n\n<p>I could expand the tensor, adding dimension like so:<\/p>\n\n<pre><code>more_dim = tf.expand_dims(input=net,axis=-1)\nnet = tf.layers.conv1d(more_dim,filters=1,kernel_size=1)\n<\/code><\/pre>\n\n<p>But that only gives me another exception, this time when computing softmax cross entropy loss:<\/p>\n\n<pre><code>ValueError(\"Can not squeeze dim[2], expected a dimension of 1, got 18 for 'sparse_softmax_cross_entropy_loss\/remove_squeezable_dimensions\/Squeeze' (op: 'Squeeze') with input shapes: [?,1024,18].\",)\n<\/code><\/pre>\n\n<p>The last dimension, 18, is number of labels. I could resize the tensor again. But I feel like it might be a loss of time. <\/p>\n\n<p>And since my, admitedly naive, approach doesn't work. I have to ask, how to do it properly. <\/p>\n","answers":[{"AnswerId":"52638112","CreationDate":"2018-10-04T02:26:48.453","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>I think Tensorflow requires an additional channel dimension. Try<\/p>\n\n<pre><code>net = tf.expand_dims(net, -1) # Adding a channel as last dimension.\nnet = tf.layers.conv1d(net,filters=1,kernel_size=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52633421,"AnswerCount":0,"Tags":"<performance><tensorflow>","CreationDate":"2018-10-03T18:20:58.900","AcceptedAnswerId":null,"OwnerUserId":10452211.0,"Title":"Tensorflow dataset API slow to get data on large batch size","Body":"<p>I find that getting one batch from tensorflow dataset API can be super slow when the batch size is large even when all data is in memory. Following is one example. Does anyone have any insight ?<\/p>\n\n<pre><code>FEATURE_NUM = 500\ntf_X = tf.placeholder(dtype=tf.float32, shape=[None, FEATURE_NUM], name=\"X\")\ntf_Y = tf.placeholder(dtype=tf.float32, shape=[None, 1], name=\"Y\")\n\nbatch_size = 1000000\ndataset = tf.data.Dataset.from_tensor_slices((tf_X, tf_Y)).batch(batch_size)\niterator = dataset.make_initializable_iterator()\nnext_element = iterator.get_next()\n\nse = tf.Session()\nse.run(tf.global_variables_initializer())\nse.run(iterator.initializer, feed_dict={tf_X : numpy_array_X, tf_Y : numpy_array_Y})\n\nwhile True:\n    data = se.run(next_element) # This takes more than 5 seconds per call\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52634226,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><ensemble-learning>","CreationDate":"2018-10-03T19:17:35.277","AcceptedAnswerId":52634607.0,"OwnerUserId":583464.0,"Title":"ensemble model with different inputs (Expected to see 2 array(s))","Body":"<p>I have trained 2 models.<\/p>\n\n<p>First model is UNet:<\/p>\n\n<pre><code>print(model_unet.summary())\n\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_4 (InputLayer)            (None, 128, 128, 1)  0                                            \n__________________________________________________________________________________________________\nconv2d_26 (Conv2D)              (None, 128, 128, 32) 320         input_4[0][0]                    \n__________________________________________________________________________________________________\nconv2d_27 (Conv2D)              (None, 128, 128, 32) 9248        conv2d_26[0][0]  \n.....\n.....\nconv2d_44 (Conv2D)              (None, 128, 128, 1)  33          zero_padding2d_4[0][0]           \n==================================================================================================\nTotal params: 7,846,081\nTrainable params: 7,846,081\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Second is ResNet:<\/p>\n\n<pre><code>print(model_resnet.summary())\n\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_3 (InputLayer)            (None, 128, 128, 3)  0                                            \n__________________________________________________________________________________________________\nconv1_pad (ZeroPadding2D)       (None, 134, 134, 3)  0           input_3[0][0]                    \n....\n....\nconv2d_25 (Conv2D)              (None, 128, 128, 3)  99          zero_padding2d_3[0][0]           \n==================================================================================================\nTotal params: 24,186,915\nTrainable params: 24,133,795\nNon-trainable params: 53,120\n<\/code><\/pre>\n\n<p>The UNet has 1 channel (gray) and the ResNet has 3 channels.<\/p>\n\n<p>Then, I am trying to create an ensemble model:<\/p>\n\n<pre><code>def ensemble(models, models_input):\n\n    outputs = [model(models_input[idx]) for idx, model in enumerate(models)]\n    x = Average()(outputs)\n\n    model_inputs = [model for model in models_input]\n    model = Model(model_inputs, x)\n\n    return model\n\nmodels = [model_unet, model_resnet]\nmodels_input = [Input((128,128,1)), Input((128,128, 3))]\n\nensemble_model = ensemble(models, models_input)\n<\/code><\/pre>\n\n<p>When I try to predict on validation data:<\/p>\n\n<pre><code>pred_val = ensemble_model.predict(X_val)\n<\/code><\/pre>\n\n<p>I am receiving the error:<\/p>\n\n<pre><code>Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[[[0.46755977],\n         [0.52268691],\n         [0.52766109],\n         ....\n\nX_val.shape is : (800, 128, 128, 1)\n<\/code><\/pre>\n\n<p>I think the problem is the channels but I don't know how to overcome this.<\/p>\n","answers":[{"AnswerId":"52634607","CreationDate":"2018-10-03T19:45:03.217","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If your training data are grayscale images and considering that your ResNet model takes as input RGB images, then you should ask yourself how do you want to go from grayscale to RGB? One answer is to repeat the grayscale image 3 times to get the RBG image. Then you can easily define a model with <strong>one<\/strong> input layer that takes your grayscale images and feed them accordingly to the models you have defined:<\/p>\n\n<pre><code>from keras import backend as K\n\ninput_image = Input(shape=(128,128,1))\n\nunet_out = model_unet(input_image)\nrgb_image = Lambda(lambda x: K.repeat_elements(x, 3, -1))(input_image)\nresnet_out = model_resnet(rgb_image)\n\noutput = Average()([unet_out, resnet_out])\n\nensemble_model = Model(input_image, output)\n<\/code><\/pre>\n\n<p>Then you can easily call <code>predict<\/code> with <strong>one<\/strong> input array:<\/p>\n\n<pre><code>pred_val = ensemble_model.predict(X_val)\n<\/code><\/pre>\n\n<p>One alternative to this solution is to go with the solution you have used in your question. However, you first need to convert your images from grayscale to RGB and then pass both of the arrays to <code>predict<\/code> method:<\/p>\n\n<pre><code>X_val_rgb = np.repeat(X_val, 3, -1)\n\npred_val = ensemble_model.predict([X_val, X_val_rgb])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52634553,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-03T19:41:02.570","AcceptedAnswerId":null,"OwnerUserId":7819329.0,"Title":"CNN architecture predicts always the same class","Body":"<p>I am using this CNN architecture :<\/p>\n\n<pre><code>def createModel():\n\n\nmodel = Sequential()\nmodel.add(Conv2D(96, (11, 11), strides=(4,4), padding='same', activation='relu', input_shape=(224,224,3))) #, input_shape=input_shape\nmodel.add(MaxPooling2D(pool_size=(3, 3), strides=(2,2), padding='same'))\n\n\nmodel.add(Conv2D(256, (5, 5), strides=(1,1), padding='same', activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(3, 3), strides=(2,2), padding='same'))\n\n\nmodel.add(Conv2D(384, (3, 3), strides=(1,1), padding='same', activation='relu'))\nmodel.add(Conv2D(384, (3, 3), strides=(1,1), padding='same', activation='relu'))\nmodel.add(Conv2D(256, (3, 3), strides=(1,1), padding='same', activation='relu'))\n\n\n\nmodel.add(Flatten())\nmodel.add(Dense(4096, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(4096, activation='relu'))\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(num_classes, activation='softmax'))\n\nreturn model\n<\/code><\/pre>\n\n<p>At the total, it makes 226,065,795 trainable parameters... I train my model on a quite small dataset. Indeed, I have three classes, each one contains 1000 images, that are split between training and validation set. I used OneHotEncoding. <\/p>\n\n<p>However, when I train the model, the overall accuracy tends to a very low result : 32%. I am using <code>categorical_crossentropy<\/code> loss function. When looking deeper, it seems that the model always predicts <code>1,0,0<\/code> vector, i.e. the first class and never the others, even for the trained data that thus have been shown to the model.<\/p>\n\n<p>What's the problem there ? Why do I have poor results ? My dataset is balanced tough. Is it because my number of parameters is too high compared to the amount of data? I don't understand...<\/p>\n\n<p>EDIT : removing Dropout still keeps the bad thing<\/p>\n","answers":[]}
+{"QuestionId":52634720,"AnswerCount":1,"Tags":"<linux><ubuntu><tensorflow><tensorboard>","CreationDate":"2018-10-03T19:53:05.923","AcceptedAnswerId":null,"OwnerUserId":10453306.0,"Title":"Suggestions for how to host a public TensorBoard to present an ML prototype?","Body":"<p>My current TensorBoard is only available locally, I understand I can ssh into my ML server and access the TensorBoard remotely. I am looking for a way to host the TensorBoard files on my webserver that doesn't have a dependency on the ML server, so that I may present the results of my ML prototype to those that have access to a web url. Ideas? <\/p>\n","answers":[{"AnswerId":"52638089","CreationDate":"2018-10-04T02:22:27.560","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>You can just install Tensorflow (which also installs Tensorboard) on your web server and then just copy the Tensorboard files over from your ML server to the web server either manually or using <code>rsync<\/code>. This will not expose your ML server to the outside at all.<\/p>\n"}]}
+{"QuestionId":52634828,"AnswerCount":0,"Tags":"<path><ubuntu-14.04><caffe><ros><catkin>","CreationDate":"2018-10-03T20:01:55.270","AcceptedAnswerId":null,"OwnerUserId":9942542.0,"Title":"Cannot compile GPD in Ubuntu 14.04","Body":"<p>I cannot compile gpd in Ubuntu 14.04, I compiled it before but no luck this time. When I tried to compile, I keep getting this error:<\/p>\n\n<p><strong>In file included from \/home\/tawfiq\/catkin_ws\/src\/gpd\/src\/gpd\/caffe_classifier.cpp:1:0:\n\/home\/tawfiq\/catkin_ws\/src\/gpd\/src\/gpd\/..\/..\/include\/gpd\/caffe_classifier.h:41:27: fatal error: caffe\/caffe.hpp: No such file or directory\n #include \"caffe\/caffe.hpp\"<\/strong><\/p>\n\n<p>The following lines are already in caffe_classifier.h file:<\/p>\n\n<pre><code>\/\/ Caffe\n#include \"caffe\/caffe.hpp\"\n#include \"caffe\/layers\/memory_data_layer.hpp\"\n#include \"caffe\/util\/io.hpp\"\n<\/code><\/pre>\n\n<p>In caffe classifier.cpp file which is inside the gpd folder in my catkin_ws src directory, the following line is included:<\/p>\n\n<pre><code>#include \"..\/..\/include\/gpd\/caffe_classifier.h\"\n<\/code><\/pre>\n\n<p>The caffe folder is in home, inside caffe folder, there is include folder, then caffe folder, then caffe.hpp in caffe.<\/p>\n\n<p>It is working fine in others' computers. I installed cuda 7.5. Then caffe and gpd from this link:<\/p>\n\n<p><a href=\"https:\/\/github.com\/atenpas\/gpd\" rel=\"nofollow noreferrer\">https:\/\/github.com\/atenpas\/gpd<\/a><\/p>\n\n<p>and followed what they said. Cuda and caffe successfully compiled. It is so frustrating, I tried updating the bashrc or putting the absolute path in caffe folder, nothing worked. <\/p>\n","answers":[]}
+{"QuestionId":52635537,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-10-03T20:53:16.997","AcceptedAnswerId":null,"OwnerUserId":10453487.0,"Title":"Tensorflow Data pipeline","Body":"<p>I am new to both Tensorflow and StackOverflow. I have a question related to <a href=\"https:\/\/stackoverflow.com\/questions\/50781373\/using-feed-dict-is-more-than-5x-faster-than-using-dataset-api\">Using feed_dict is more than 5x faster than using dataset API?<\/a>.<\/p>\n\n<p>I see that the recommended solution using the tf.data-based pipeline involves using repeat() before batch(). If one is not going to shuffle between epochs - as seems to be the case here - wouldn't it much faster to cache() after map() and follow that with repeat().<\/p>\n","answers":[]}
+{"QuestionId":52635866,"AnswerCount":3,"Tags":"<python><tensorflow><keras><jupyter-notebook><conda>","CreationDate":"2018-10-03T21:21:54.920","AcceptedAnswerId":null,"OwnerUserId":613326.0,"Title":"How to add a conda environment in jupyter","Body":"<p>I'm try to enable a keras environmnt into jupyter.<br>\nUsing below commands I added the conda tf environment for Keras :<\/p>\n\n<pre><code>C:&gt;conda create --name tf python=3.5 \nC:&gt;activate tf (tf)\nC:\\Keras\\Test&gt;pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.4.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>Next I downloaded Keras-2.1.3-py2.py3-none-any.whl (from : <a href=\"https:\/\/repo.continuum.io\/archive\/\" rel=\"nofollow noreferrer\">https:\/\/repo.continuum.io\/archive\/<\/a>)<br>\nAnd successfully installed it. <\/p>\n\n<pre><code>(tf) C:\\Keras&gt;pip install --ignore-installed --upgrade C:\\Keras\\Keras-2.1.3-py2.py3-none-any.whl\n<\/code><\/pre>\n\n<p>But then when i type <\/p>\n\n<pre><code>(tf) c:\\Keras\\Jupyter Notebook\n<\/code><\/pre>\n\n<p>.<\/p>\n\n<p>I would have thought to be able to 'switch' to the tf keras environment.\nWhen starting up the web browser\nBut all I see in below jupyter page is the python 3 environment which doesn't know of keras. So shouldn't I see a <strong>\"tf\"<\/strong> environment here as well?\nfrom both environments i can start the jupyter notebook, but not sure if that should make a difference, both startups dont show tf, am i missing something here?. (in the picture i also show both conda consoles)\n<a href=\"https:\/\/i.stack.imgur.com\/d3h9o.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d3h9o.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When i try to launch a notebook that makes use of keras.<br>\nThen this is the error i see like if it wasnt installed ?. (i'm sure it did though).\n<a href=\"https:\/\/i.stack.imgur.com\/NUvy0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NUvy0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>.<br>\nHowever a small test in python ran directly from the console\nproves keras is installed and working \n<a href=\"https:\/\/i.stack.imgur.com\/HsgdT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HsgdT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>(but why not in jupyter??)<\/p>\n","answers":[{"AnswerId":"54056539","CreationDate":"2019-01-05T21:42:00.057","ParentId":null,"OwnerUserId":"1432481","Title":null,"Body":"<p>Create kernel.json file in  <code>~\/local\/share\/jupyter\/kernels\/&lt;YOUR_KERNEL_NAME\/<\/code>. Take an example for the content from:<code>\/usr\/local\/share\/jupyter\/kernels\/python3\/kernel.json<\/code> <\/p>\n\n<p>see: <a href=\"https:\/\/jupyter-client.readthedocs.io\/en\/stable\/kernels.html\" rel=\"nofollow noreferrer\">https:\/\/jupyter-client.readthedocs.io\/en\/stable\/kernels.html<\/a>\n     <a href=\"https:\/\/ipython.org\/ipython-doc\/3\/development\/kernels.html\" rel=\"nofollow noreferrer\">https:\/\/ipython.org\/ipython-doc\/3\/development\/kernels.html<\/a> <\/p>\n"},{"AnswerId":"52636522","CreationDate":"2018-10-03T22:25:35.207","ParentId":null,"OwnerUserId":"613326","Title":null,"Body":"<p>Well when i was in the (tf) environment made earlier and typed \"conda list\"\nI got a fairly short list of installed packages (just tensorflow) about 26 packages.\nThen I noted that jupyter package wasnt in this environment.\nNot sure if it should i added it with <\/p>\n\n<pre><code>conda install jupyter\n<\/code><\/pre>\n\n<p>After a while (that command added about 8 Gig of python code...) it got installed.\nWith the message:<br>\nEnabling notebook extension jupyter-js-widgets\/extension...\n      - Validating: ok<\/p>\n\n<p>next i tried <\/p>\n\n<pre><code>(tf) C:\\Keras&gt;python -m ipykernel install --user --name tf --display-name \"Python3 tensorflow\"\n<\/code><\/pre>\n\n<p>with a message \"Installed kernelspec tf in C:\\Users\\Peter\\AppData\\Roaming\\jupyter\\kernels\\tf\"<\/p>\n\n<blockquote>\n  <p>I'm not sure though if this is the proper solution, because would it\n  always be required to add 8Giga just to launch it in a webpage.\n  (seams overkill, but i just dont know if that's normal for conda-jupyter. (while jupyther was allready in the 'plain' python 3\n  enviroment. (or it was available to conda)...not sure i'm thinking\n  perhaps it be better if conda was part of jupyter but maybe its just\n  the otherway around.<\/p>\n<\/blockquote>\n"},{"AnswerId":"52636245","CreationDate":"2018-10-03T21:55:51.583","ParentId":null,"OwnerUserId":"10365609","Title":null,"Body":"<p>Fundamentally the problem is usually rooted in the fact that the Jupyter kernels are disconnected from Jupyter's shell; in other words, the installer points to a different Python version than is being used in the notebook.<\/p>\n\n<p>You might first try installing with conda. <\/p>\n\n<pre><code>conda install -c conda-forge keras \n<\/code><\/pre>\n\n<p>Otherwise you can try installing from within the notebook itself:<\/p>\n\n<pre><code># Install a pip package in the current Jupyter kernel\nimport sys\n!{sys.executable} -m pip install keras \n<\/code><\/pre>\n\n<p>As tempting as it might be... DO NOT DO:<\/p>\n\n<pre><code># DON'T DO THIS\n!pip install keras \n<\/code><\/pre>\n"}]}
+{"QuestionId":52636127,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-10-03T21:45:58.093","AcceptedAnswerId":null,"OwnerUserId":7656080.0,"Title":"How to use Keras generator with tf.data API","Body":"<p>I am trying to use the generator found in Keras preprocessing library. I wanted to experiment with this since Keras provides great functions for image augmentation. However, I am not sure if this is actually possible. <\/p>\n\n<p>Here is how I made a tf dataset from the Keras generator:<\/p>\n\n<pre><code>def make_generator():\n    train_datagen = ImageDataGenerator(rescale=1. \/ 255)\n    train_generator = \n    train_datagen.flow_from_directory(train_dataset_folder,target_size=(224, 224), class_mode='categorical', batch_size=32)\n    return train_generator\n\ntrain_dataset = tf.data.Dataset.from_generator(make_generator,(tf.float32, tf.float32)).shuffle(64).repeat().batch(32)\n<\/code><\/pre>\n\n<p>Note that if you try to directly give <code>train_generator<\/code> as the argument to <code>tf.data.Dataset.from_generator<\/code> there will be an error. However, the above method doesn't produce an error.<\/p>\n\n<p>When I run it within a session to check the output from the dataset I get the following error.<\/p>\n\n<pre><code>iterator = train_dataset.make_one_shot_iterator()\nnext_element = iterator.get_next()\nsess = tf.Session()\nfor i in range(100):\n    sess.run(next_element)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Found 1000 images belonging to 2 classes.\n  --------------------------------------------------------------------------- InvalidArgumentError                      Traceback (most recent call\n  last)\n  \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\n  in _do_call(self, fn, *args)    1291     try:\n  -> 1292       return fn(*args)    1293     except errors.OpError as e:<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\n  in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata) \n  1276       return self._call_tf_sessionrun(\n  -> 1277           options, feed_dict, fetch_list, target_list, run_metadata)    1278 <\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\n  in _call_tf_sessionrun(self, options, feed_dict, fetch_list,\n  target_list, run_metadata)    1366         self._session, options,\n  feed_dict, fetch_list, target_list,\n  -> 1367         run_metadata)    1368 <\/p>\n  \n  <p>InvalidArgumentError: Cannot batch tensors with different shapes in\n  component 0. First element had shape [32,224,224,3] and element 29 had\n  shape [8,224,224,3].   [[{{node IteratorGetNext_2}} =\n  IteratorGetNextoutput_shapes=[, ],\n  output_types=[DT_FLOAT, DT_FLOAT],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n<\/blockquote>\n\n<p>Please let me know if anyone has any experience with this or know any alternate way.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I was able to solve the problem after using the suggestion by J.E.K.<\/p>\n\n<pre><code>train_dataset = tf.data.Dataset.from_generator(make_generator,(tf.float32, tf.float32))\n<\/code><\/pre>\n\n<p>However when I give <code>train_dataset<\/code> to a Keras <code>.fit<\/code> method I get the following error.<\/p>\n\n<pre><code>model_regular.fit(train_dataset,steps_per_epoch=1000,epochs=2)\n<\/code><\/pre>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- ValueError                                Traceback (most recent call\n  last)  in ()\n  ----> 1 model_regular.fit(train_dataset,steps_per_epoch=1000,epochs=2)<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\n  in fit(self, x, y, batch_size, epochs, verbose, callbacks,\n  validation_split, validation_data, shuffle, class_weight,\n  sample_weight, initial_epoch, steps_per_epoch, validation_steps,\n  **kwargs)    1507         steps_name='steps_per_epoch',    1508         steps=steps_per_epoch,\n  -> 1509         validation_split=validation_split)    1510     1511     # Prepare validation data.<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\n  in _standardize_user_data(self, x, y, sample_weight, class_weight,\n  batch_size, check_steps, steps_name, steps, validation_split)\n      948           x = self._dataset_iterator_cache[x]\n      949         else:\n  --> 950           iterator = x.make_initializable_iterator()\n      951           self._dataset_iterator_cache[x] = iterator\n      952           x = iterator<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\n  in make_initializable_iterator(self, shared_name)\n      119     with ops.colocate_with(iterator_resource):\n      120       initializer = gen_dataset_ops.make_iterator(self._as_variant_tensor(),\n  --> 121                                                   iterator_resource)\n      122     return iterator_ops.Iterator(iterator_resource, initializer,\n      123                                  self.output_types, self.output_shapes,<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gen_dataset_ops.py\n  in make_iterator(dataset, iterator, name)    2542   if _ctx is None or\n  not _ctx._eager_context.is_eager:    2543     _, _, _op =\n  _op_def_lib._apply_op_helper(\n  -> 2544         \"MakeIterator\", dataset=dataset, iterator=iterator, name=name)    2545     return _op    2546     _result = None<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\n  in _apply_op_helper(self, op_type_name, name, **keywords)\n      348       # Need to flatten all the arguments into a list.\n      349       # pylint: disable=protected-access\n  --> 350       g = ops._get_graph_from_inputs(_Flatten(keywords.values()))\n      351       # pylint: enable=protected-access\n      352     except AssertionError as e:<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\n  in _get_graph_from_inputs(op_input_list, graph)    5659         graph\n  = graph_element.graph    5660       elif original_graph_element is not None:\n  -> 5661         _assert_same_graph(original_graph_element, graph_element)    5662       elif graph_element.graph is not graph:<br>\n  5663         raise ValueError(\"%s is not from the passed-in graph.\" %\n  graph_element)<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\n  in _assert_same_graph(original_item, item)    5595   if\n  original_item.graph is not item.graph:    5596     raise\n  ValueError(\"%s must be from the same graph as %s.\" % (item,\n  -> 5597                                                                 original_item))    5598     5599 <\/p>\n  \n  <p>ValueError: Tensor(\"IteratorV2:0\", shape=(), dtype=resource) must be\n  from the same graph as Tensor(\"FlatMapDataset:0\", shape=(),\n  dtype=variant).<\/p>\n<\/blockquote>\n\n<p>Is this a bug or is Keras fit method not meant to be used this way?<\/p>\n","answers":[{"AnswerId":"52649515","CreationDate":"2018-10-04T14:41:41.550","ParentId":null,"OwnerUserId":"4606149","Title":null,"Body":"<p>I have tried to reproduce your results with a simple example and I found out that you get different output shapes when one uses batching within the generator function and <code>tf.data<\/code>.<\/p>\n\n<p>The Keras function <code>train_datagen.flow_from_directory(batch_size=32)<\/code> already returns the data with shape <code>[batch_size, width, height, depth]<\/code>. If one uses <code>tf.data.Dataset().batch(32)<\/code> the output data is batched again into shape <code>[batch_size, batch_size, width, height, depth]<\/code>.<\/p>\n\n<p>This could have caused your issue for some reason.<\/p>\n"},{"AnswerId":"55517035","CreationDate":"2019-04-04T13:20:47.900","ParentId":null,"OwnerUserId":"3387851","Title":null,"Body":"<p>Shouldn't <\/p>\n\n<pre><code>model_regular.fit(train_dataset,steps_per_epoch=1000,epochs=2)\n<\/code><\/pre>\n\n<p>be <\/p>\n\n<pre><code>model_regular.fit(train_dataset.make_one_shot_iterator(),steps_per_epoch=1000,epochs=2)\n<\/code><\/pre>\n\n<p>as per <a href=\"https:\/\/stackoverflow.com\/a\/48134765\/3387851\">this answer<\/a>.<\/p>\n"}]}
+{"QuestionId":52636328,"AnswerCount":1,"Tags":"<python><lambda><neural-network><keras><nonetype>","CreationDate":"2018-10-03T22:04:23.973","AcceptedAnswerId":52636813.0,"OwnerUserId":9843744.0,"Title":"NoneType' object has no attribute '_inbound_nodes'","Body":"<p>Hi I am trying to build a Mixture-of-experts neural network. I found a code here: <a href=\"http:\/\/blog.sina.com.cn\/s\/blog_dc3c53e90102x9xu.html\" rel=\"nofollow noreferrer\">http:\/\/blog.sina.com.cn\/s\/blog_dc3c53e90102x9xu.html<\/a>. My goal is that the gate and expert come from different data, but with same dimensions.<\/p>\n\n<pre><code>def sliced(x,expert_num):\n    return x[:,:,:expert_num]\n\ndef reduce(x, axis):\n    return K.sum(x, axis=axis, keepdims=True)\n\ndef gatExpertLayer(inputGate, inputExpert, expert_num, nb_class):\n    #expert_num=30\n    #nb_class=10\n    input_vector1 = Input(shape=(inputGate.shape[1:]))\n    input_vector2 = Input(shape=(inputExpert.shape[1:]))\n\n    #The gate\n    gate = Dense(expert_num*nb_class, activation='softmax')(input_vector1)\n    gate = Reshape((1,nb_class, expert_num))(gate)\n    gate = Lambda(sliced, output_shape=(nb_class, expert_num), arguments={'expert_num':expert_num})(gate)\n\n    #The expert\n    expert = Dense(nb_class*expert_num, activation='sigmoid')(input_vector2)\n    expert = Reshape((nb_class, expert_num))(expert)\n\n    #The output\n    output = tf.multiply(gate, expert)\n    #output = keras.layers.merge([gate, expert], mode='mul')\n    output = Lambda(reduce, output_shape=(nb_class,), arguments={'axis': 2})(output)\n\n    model = Model(input=[input_vector1, input_vector2], output=output)\n\n    model.compile(loss='mean_squared_error', metrics=['mse'], optimizer='adam')\n\n    return model\n<\/code><\/pre>\n\n<p>However, I got \"'NoneType' object has no attribute '_inbound_nodes'\". I checked other similar questions here: <a href=\"https:\/\/stackoverflow.com\/questions\/52448652\/attributeerror-nonetype-object-has-no-attribute-inbound-nodes-while-trying\">AttributeError: &#39;NoneType&#39; object has no attribute &#39;_inbound_nodes&#39; while trying to add multiple keras Dense layers<\/a> but the problem is fixed with the Lambda function of keras to convert into a layer. <\/p>\n","answers":[{"AnswerId":"52636813","CreationDate":"2018-10-03T23:04:35.280","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Well, you need to put <code>tf.multiply()<\/code> inside a <code>Lambda<\/code> layer to get a Keras Tensor as output (and not a Tensor):<\/p>\n\n<pre><code>output = Lambda(lambda x: tf.multiply(x[0], x[1]))([gate, expert])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52636437,"AnswerCount":1,"Tags":"<r><tensorflow><keras>","CreationDate":"2018-10-03T22:16:25.077","AcceptedAnswerId":null,"OwnerUserId":1883795.0,"Title":"How can I keep track of which weights correspond to which layers and inputs in Keras\/Tensorflow?","Body":"<p>What is a good method of keeping track of what weights go where?  When I do <code>get_weights(model)<\/code>, I get a list of arrays, and I have to guess which ones correspond to which part of my model by their dimensionality.  How can I figure out which layers they're emanating from and connecting to, so that I can manipulate them programmatically?<\/p>\n\n<p>I'm working in R, but answers in python can probably be translated.<\/p>\n","answers":[{"AnswerId":"52640208","CreationDate":"2018-10-04T06:17:48.293","ParentId":null,"OwnerUserId":"5403074","Title":null,"Body":"<p>I would suggest using <a href=\"https:\/\/www.tensorflow.org\/guide\/debugger\" rel=\"nofollow noreferrer\">tensorflow debugger<\/a> in combination with naming your layers with names that make sense. After that you can execute a certain amount of iterations and check the values of the weights for a layer by using its name. Also you can check gradients for these layers.  <\/p>\n"}]}
+{"QuestionId":52636869,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2018-10-03T23:13:09.603","AcceptedAnswerId":52638071.0,"OwnerUserId":8402528.0,"Title":"How to display the contents of a tf.Variable","Body":"<p>I am new to Tensorflow and I appreciate knowing how to visualize the contents of a tf.Variable, tried% f,% s but it is not shown, where is my mistake.\nI put the code that I am using, I appreciate your response.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\ninit = tf.global_variables_initializer()\nsess.run(init)\nstartIter = 2000\nglobalStep = tf.Variable(startIter, trainable=False)\nx = tf.Variable(5.0, name=\"counter\") \nfor i in range(startIter): \n    totalLoss = x**2-20.0*x+1.0 \n    opt = tf.train.AdamOptimizer(learning_rate=0.0001) \n    grads = opt.compute_gradients(totalLoss)\n    grads = [(None, var) if grad is None else (tf.clip_by_value(grad, -1.0, 1.0), var) for grad, var in grads]\n    applyGradientOp = opt.apply_gradients(grads, global_step=globalStep)\n    #print(\"opt.get_name(): \",opt.get_name(),\"opt._lr: \",opt._lr,\"opt._lr_t: %f \"% (sess.run(opt._lr_t)))  #jll1\n    print(\"opt.get_slot_names: \",opt.get_slot_names())\n    print('  ', opt.get_slot(var,'m'))  # here\n    print('  ', opt.get_slot(var,'v'))  # here\n    assign_op = tf.assign(x, x + 1)\n<\/code><\/pre>\n\n<p>This result is displayed<\/p>\n\n<pre><code>('opt.get_slot_names: ', ['m', 'v'])\n('  ', &lt;tf.Variable 'counter\/Adam_614:0' shape=() dtype=float32_ref&gt;)\n('  ', &lt;tf.Variable 'counter\/Adam_615:0' shape=() dtype=float32_ref&gt;)\n<\/code><\/pre>\n\n<p>but I would like to visualize a value, of course if it is possible.<\/p>\n\n<p>I understand that they are AdamOptimizer slots, and I seek to show the learning rate in each step. I have reviewed other answers but they do not work.\nUse:<\/p>\n\n<pre><code>print (\"opt.get_name ():\", opt.get_name (), \"opt._lr:\", opt._lr, \"opt._lr_t:\", opt._lr_t) # jll1 \n<\/code><\/pre>\n\n<p>before the print and the same result.<\/p>\n","answers":[{"AnswerId":"52648456","CreationDate":"2018-10-04T13:50:17.477","ParentId":null,"OwnerUserId":"8402528","Title":null,"Body":"<p>Thank you very much for the help, I reordered the code.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()   #jll2\nstartIter = 2000\nglobalStep = tf.Variable(startIter, trainable=False)\nx = tf.Variable(5.0, name=\"counter\")\nfor i in range(startIter): \n    totalLoss = x**2-20.0*x+1.0  \n    opt = tf.train.AdamOptimizer(learning_rate=0.0001)  \n#    print(\"opt.get_name(): \",opt.get_name(),\"opt._lr: \",opt._lr,\"opt._lr_t: \",opt._lr_t)  #jll1\n    grads = opt.compute_gradients(totalLoss)\n    grads = [(None, var) if grad is None else (tf.clip_by_value(grad, -1.0, 1.0), var) for grad, var in grads]\n    applyGradientOp = opt.apply_gradients(grads, global_step=globalStep)\n#    print('Learning rate opt._lr_t: %f' % (sess.run(opt._lr_t)))    #jll3             \n#    print(\"opt.get_name(): \",opt.get_name(),\"opt._lr: \",opt._lr,\"opt._lr_t: %f \"% (sess.run(opt._lr_t)))  #jll1\n    print(\"opt.get_slot_names: \",opt.get_slot_names())\n    #### **by matwilso**\n    with sess.as_default():\n        sess.run(tf.global_variables_initializer())\n        print('m:',sess.run(opt.get_slot(var,'m')),'v:',sess.run(opt.get_slot(var,'v')))\n    #### **by Andreas Pasternak**\n    with sess.as_default():\n        sess.run(tf.global_variables_initializer())\n        print_node1 = tf.Print(opt.get_slot(var,'m'), [opt.get_slot(var,'m')], 'm')\n        print_node2 = tf.Print(opt.get_slot(var,'v'), [opt.get_slot(var,'v')], 'v')\n        print(\"['m', 'v']:\",sess.run([print_node1,print_node2]))\n    assign_op = tf.assign(x, x + 1)\n# thanks\n<\/code><\/pre>\n\n<p>And now I get:<\/p>\n\n<pre><code>('opt.get_slot_names: ', ['m', 'v'])\n('m:', 0.0, 'v:', 0.0)\n(\"['m', 'v']:\", [0.0, 0.0])\n<\/code><\/pre>\n"},{"AnswerId":"52637463","CreationDate":"2018-10-04T00:45:26.000","ParentId":null,"OwnerUserId":"7211137","Title":null,"Body":"<p>To view the values in a tf.Variable, you need to run it with a TensorFlow session.  This should work:<\/p>\n\n<pre><code>print('  ', sess.run(opt.get_slot(var,'m')))  # here\nprint('  ', sess.run(opt.get_slot(var,'v')))  # here\n<\/code><\/pre>\n"},{"AnswerId":"52638071","CreationDate":"2018-10-04T02:18:32.733","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>You can use <code>tf.Print()<\/code> but this requires you to add it to your computational graph:<\/p>\n\n<pre><code>print_node = tf.Print(opt.get_slot(var,'m'), [opt.get_slot(var,'m')], 'm')\nsess.run([print_node], ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52636943,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-03T23:23:01.457","AcceptedAnswerId":52643573.0,"OwnerUserId":1047543.0,"Title":"What is a \"stateful object\" in tensorflow?","Body":"<p>In several parts of the documentation (e.g. <code>Dataset Iterators<\/code> <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#consuming_values_from_an_iterator\" rel=\"nofollow noreferrer\">here<\/a>) there are references to <code>Stateful Objects<\/code>. What exactly are they and what role do they play in the graph?<\/p>\n\n<p>To clarify, in the Dataset documentation there's an example with the <code>one_shot_iterator<\/code> that works because it's stateless:  <\/p>\n\n<pre><code>dataset = tf.data.Dataset.range(100)\niterator = dataset.make_one_shot_iterator()\n<\/code><\/pre>\n\n<p>what makes the iterator stateless?<\/p>\n","answers":[{"AnswerId":"52643573","CreationDate":"2018-10-04T09:39:58.237","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>As others have mentioned, stateful objects are those holding a state. <strong>Now, a state, in TensorFlow terms, is some value or data that is saved between different calls to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"noreferrer\"><code>tf.Session.run<\/code><\/a><\/strong>. The most common and basic kind of stateful objects are variables. You can call <code>run<\/code> once to update a model's parameters, which are variables, and they will maintain their assigned value for the next call to <code>run<\/code>. This is different to most operations; for example, if you have an addition operation, which takes two tensors and outputs a third one, the output value that it computes in one call to <code>run<\/code> is not saved. Indeed, even if your graph consists only on operations with constant values, tensor operations will be evaluated every time you call <code>run<\/code>, even though the result will always be the same. <strong>When you <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"noreferrer\">assign<\/a> a value to a variable, however, it will \"stick\"<\/strong> (and, by the way, take the corresponding memory and, if you choose so, be serialized on checkpoints).<\/p>\n\n<p>Dataset iterators are also stateful. When you get a piece of data in one run then it is consumed, and then in the next run you get a different piece of data; the iterator \"remembers\" where it was between runs. That is why, similarly to how you initialize variables, you can initialize iterators (when they support it), to reset them back to a known state.<\/p>\n\n<p>Technically speaking, another kind of stateful objects is random operations. One usually regards random operations as, well, random, but in reality they hold a random number generator that does have a state which is held between runs, and if you provide a seed then they will be in a well-defined state when you start the session. However there is not, as far as I know, any way to reset random operations to their initial state within the same session.<\/p>\n\n<p>Note that frequently (when one is not referring to TensorFlow in particular) the term \"stateful\" is used in a slightly different sense, or at a different level of abstraction. For example, <a href=\"https:\/\/en.wikipedia.org\/wiki\/Recurrent_neural_network\" rel=\"noreferrer\">recurrent neural networks (RNNs)<\/a> are generally said to be stateful, because, conceptually, they have an internal state that changes with every input they receive. When you make a RNN in TensorFlow, however, that internal state does not necessarily have to be in a stateful object! Like any other kind of neural network, RNNs in TensorFlow will in principle have some parameters or weights, typically stored in trainable variables - so, in TensorFlow terms, all trainable models, RNN or not, have stateful objects for the trained parameters. However, the internal state of the RNN is represented in TensorFlow with an input state value and and output state value that you get on each run (see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"noreferrer\"><code>tf.nn.dynamic_rnn<\/code><\/a>), and you can just start with a \"zero\" state on each run and forget about the final output state. Of course, you can also, if you want, take the input state to be the value of a variable, and the write the output state back to that variable, and then your RNN internal state will be \"stateful\" for TensorFlow; that is, you would be able to process some data in one run and the \"pick up where you left it\" in the next run (which may or may not make sense depending on the case). I understand this can be a bit confusing but I hope it makes sense.<\/p>\n"}]}
+{"QuestionId":52637206,"AnswerCount":0,"Tags":"<tensorflow><raspberry-pi3><tensorflow-lite>","CreationDate":"2018-10-04T00:05:46.197","AcceptedAnswerId":null,"OwnerUserId":581442.0,"Title":"How to use libtensorflow-lite.a on Raspi 3?","Body":"<p>I want to install TensorFlow Lite on the Raspi. <\/p>\n\n<p>I assume I can first <code>pip install<\/code> the prebuilt TensorFlow for the Raspi. I'm reading the instructions to cross compile TensorFlow Lite <a href=\"https:\/\/www.tensorflow.org\/lite\/rpi\" rel=\"noreferrer\">here<\/a>, but I have no idea what to do after generating <code>libtensorflow-lite.a<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":52637211,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><deep-learning><computer-vision><pytorch>","CreationDate":"2018-10-04T00:07:16.140","AcceptedAnswerId":52639188.0,"OwnerUserId":8482467.0,"Title":"Find all ReLU layer in a torchvision model","Body":"<p>After I fetch a pre-trained model from <code>torchvision.models<\/code>, I want all the <code>ReLU<\/code> instance to <code>register_backward_hook(f)<\/code>,which is like this:<\/p>\n\n<pre><code>for pos, module in self.model.features._modules.items():\n    for sub_module in module:\n        if isinstance(module, ReLU):\n            module.register_backward_hook(f)\n<\/code><\/pre>\n\n<p>The problems for me is how to find all <code>ReLU<\/code> in a model. For <code>densenet161<\/code>, the <code>ReLU<\/code> exists not only in <code>model.features._modules<\/code> but also in self-defined dense layer, eg. <code>model.features._modules['denseblock1'][0]<\/code>. For  <code>resnet151<\/code>, the <code>ReLU<\/code> exists in <code>model._modules<\/code> and its self-define layer, eg <code>model._modules['layer1']<\/code>.<\/p>\n\n<p>Is there any way to find all <code>ReLU<\/code> inside a model?<\/p>\n","answers":[{"AnswerId":"52639188","CreationDate":"2018-10-04T04:46:36.730","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>A more elegant way to iterate over all components of a model is using <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.modules\" rel=\"nofollow noreferrer\"><code>modules()<\/code><\/a>\nmethod:<\/p>\n\n<pre><code>from torch import nn\n\nfor module in self.model.modules():\n  if isinstance(module, nn.ReLU):\n    module.register_backward_hook(f)\n<\/code><\/pre>\n\n<p>If you do not want to get all sub-modules, only the immediate ones, you may consider using <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.children\" rel=\"nofollow noreferrer\"><code>children()<\/code><\/a> method instead of <code>modules()<\/code>. You can also get the name of the sub module using <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.named_modules\" rel=\"nofollow noreferrer\"><code>named_modules()<\/code><\/a> method.<\/p>\n"}]}
+{"QuestionId":52637493,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T00:50:54.873","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Save Model In Suitable Format for Google Cloud ML Engine (missing functions)","Body":"<p>I'm trying to deploy a recently trained Keras model on Google Cloud ML Engine.  I googled around to see what format the saved model needs to be for ML Engine and found this:<\/p>\n\n<pre><code>import keras.backend as K\nimport tensorflow as tf\nfrom keras.models import load_model, Sequential\nfrom tensorflow.python.saved_model import builder as saved_model_builder\nfrom tensorflow.python.saved_model import tag_constants, signature_constants\nfrom tensorflow.python.saved_model.signature_def_utils_impl import predict_signature_def\n\n# reset session\nK.clear_session()\nsess = tf.Session()\nK.set_session(sess)\n\n# disable loading of learning nodes\nK.set_learning_phase(0)\n\n# load model\nmodel = load_model('model.h5')\nconfig = model.get_config()\nweights = model.get_weights()\nnew_Model = Sequential.from_config(config)\nnew_Model.set_weights(weights)\n\n# export saved model\nexport_path = 'YOUR_EXPORT_PATH' + '\/export'\nbuilder = saved_model_builder.SavedModelBuilder(export_path)\n\nsignature = predict_signature_def(inputs={'NAME_YOUR_INPUT': new_Model.input},\n                                  outputs={'NAME_YOUR_OUTPUT': new_Model.output})\n\nwith K.get_session() as sess:\n    builder.add_meta_graph_and_variables(sess=sess,\n                                         tags=[tag_constants.SERVING],\n                                         signature_def_map={\n                                             signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature})\n    builder.save()\n<\/code><\/pre>\n\n<p>However, in Keras 2.1.3, <code>keras.backend<\/code> no longer seems to have <code>clear_session(), set_session()<\/code>, or <code>get_session()<\/code>.  What's the modern way of handling this issue?  Do those functions live elsewhere now?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52637970,"AnswerCount":2,"Tags":"<tensorflow><tensor>","CreationDate":"2018-10-04T02:04:08.760","AcceptedAnswerId":52639622.0,"OwnerUserId":10454184.0,"Title":"How to perform tensor slice update, like in pytorch?","Body":"<p>In Pytorch, you can easily update tensor as example below:<\/p>\n\n<pre><code> for i in range(x_len):\n     tensor_abc[:, i, i] = 0\n<\/code><\/pre>\n\n<p>How can we update a tensor like that in tensorflow coding?\nI tried tf.assign, can't do slice update..\ntried tf.scatter_update, not working as well...<\/p>\n","answers":[{"AnswerId":"52639622","CreationDate":"2018-10-04T05:31:19.577","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>This answer only pertains to variables.<\/p>\n\n<pre><code>import tensorflow as tf\n\nsess = tf.InteractiveSession()\nv = tf.zeros((5,5,5))\nvar = tf.Variable(initial_value=v)\n\n\ninit = tf.variables_initializer([var])\nsess.run(init)\n\n\nvar = var[ 1 : 2 ,\n           1 : 2 ,\n           1 : 2 ].assign(tf.ones((1,1,1)))\n\nprint(sess.run(var))\n<\/code><\/pre>\n\n<p>This produces<\/p>\n\n<pre><code>[[[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[0. 0. 0. 0. 0.]\n  [0. 1. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]]\n<\/code><\/pre>\n\n<p>And this<\/p>\n\n<pre><code>var = var[ 1 : 2 ,\n           0 : 1 ,\n           0 : 1 ].assign(tf.ones((1,1,1)))\n<\/code><\/pre>\n\n<p>produces<\/p>\n\n<pre><code>  [[[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[1. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n  ....\n  ....]]\n<\/code><\/pre>\n\n<p>Another example is<\/p>\n\n<pre><code>var = var[ 1 : 2 ,\n             : 2 ,\n             : 2 ].assign(tf.ones((1,2,2)))\n\n[[[0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n [[1. 1. 0. 0. 0.]\n  [1. 1. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]\n\n      ....\n      ....]]\n<\/code><\/pre>\n\n<p>You should explore <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/manip\/scatter_nd\" rel=\"nofollow noreferrer\">tf.scatter_nd<\/a> for tensors.<\/p>\n"},{"AnswerId":"52651056","CreationDate":"2018-10-04T16:00:51.457","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p><code>tf.Variable<\/code>s are the only tensors that can be updated (<a href=\"https:\/\/www.tensorflow.org\/guide\/variables\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/variables<\/a>). With variables, you would use code like <code>gather<\/code> and <code>scatter_update<\/code> for slicing.<\/p>\n\n<p>Note that other tensors are not amenable to assignment. If this what you're trying to do, I would wonder why its necessary. However, it's still possible to create new tensors with the values you want (instead of assigning in place), with code that is a bit convoluted. For example, the following doesn't work:<\/p>\n\n<pre><code>index = ... tensor = tf.constant([0,1,2,3,4]) \ntensor[i] = 0  \n## Doesn't work (TypeError: `Tensor` object does not support item assignment)\n<\/code><\/pre>\n\n<p>But this can do the equivalent:<\/p>\n\n<pre><code>tensor = tf.constant([0,1,2,3,4]) \ntensor = tf.concat([tensor[:i], tf.zeros_like(tensor[i:i+1]), tensor[i+1:]], 0)  \n## This works, creates a new tensor\n<\/code><\/pre>\n\n<p>OR:<\/p>\n\n<pre><code>tensor = tf.constant([0,1,2,3,4]) \ntensor = tf.concat([tensor[:i], tf.fill([1], 0), tensor[i+1:]], 0)  \n## This works, creates a new tensor\n<\/code><\/pre>\n"}]}
+{"QuestionId":52638191,"AnswerCount":0,"Tags":"<python><linux><tensorflow><ubuntu>","CreationDate":"2018-10-04T02:38:26.940","AcceptedAnswerId":null,"OwnerUserId":9969622.0,"Title":"Tensorflow session failed to start! Cuda virtual environment","Body":"<p>So im trying to make tensorflow run on my computer. I successfully installed ubuntu, cuda, cudnn and anaconda and set a environment for tensorflow GPU. I have the lastest nvidia drivers available for my gtx 970m. \nThis is the code i ran to test my program:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;pre&gt;&lt;font color=\"#8AE234\"&gt;&lt;b&gt;alejandro@alejandro-GL502VT&lt;\/b&gt;&lt;\/font&gt;:&lt;font color=\"#729FCF\"&gt;&lt;b&gt;~&lt;\/b&gt;&lt;\/font&gt;$ conda activate tensorflow_gpuenv\r\n(tensorflow_gpuenv) &lt;font color=\"#8AE234\"&gt;&lt;b&gt;alejandro@alejandro-GL502VT&lt;\/b&gt;&lt;\/font&gt;:&lt;font color=\"#729FCF\"&gt;&lt;b&gt;~&lt;\/b&gt;&lt;\/font&gt;$ python\r\nPython 3.6.6 |Anaconda, Inc.| (default, Jun 28 2018, 17:14:51) \r\n[GCC 7.2.0] on linux\r\nType &amp;quot;help&amp;quot;, &amp;quot;copyright&amp;quot;, &amp;quot;credits&amp;quot; or &amp;quot;license&amp;quot; for more information.\r\n&amp;gt;&amp;gt;&amp;gt; import tensorflow as tf\r\n&amp;gt;&amp;gt;&amp;gt; hello = tf.constant(&amp;apos;Hello Stack Overflow!!&amp;apos;)\r\n&amp;gt;&amp;gt;&amp;gt; sess = tf.Session()\r\n2018-10-03 23:23:05.722821: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\r\n2018-10-03 23:23:05.804766: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:897] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2018-10-03 23:23:05.805564: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 0 with properties: \r\nname: GeForce GTX 970M major: 5 minor: 2 memoryClockRate(GHz): 1.038\r\npciBusID: 0000:01:00.0\r\ntotalMemory: 2.95GiB freeMemory: 2.57GiB\r\n2018-10-03 23:23:05.805579: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1484] Adding visible gpu devices: 0\r\n2018-10-03 23:23:05.805780: E tensorflow\/core\/common_runtime\/direct_session.cc:158] Internal: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version\r\nTraceback (most recent call last):\r\n  File &amp;quot;&amp;lt;stdin&amp;gt;&amp;quot;, line 1, in &amp;lt;module&amp;gt;\r\n  File &amp;quot;\/home\/alejandro\/anaconda3\/envs\/tensorflow_gpuenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py&amp;quot;, line 1494, in __init__\r\n    super(Session, self).__init__(target, graph, config=config)\r\n  File &amp;quot;\/home\/alejandro\/anaconda3\/envs\/tensorflow_gpuenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py&amp;quot;, line 626, in __init__\r\n    self._session = tf_session.TF_NewSession(self._graph._c_graph, opts)\r\ntensorflow.python.framework.errors_impl.InternalError: Failed to create session.\r\n&lt;\/pre&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>and here is my driver info<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;pre&gt;+-----------------------------------------------------------------------------+\r\n| NVIDIA-SMI 390.48                 Driver Version: 390.48                    |\r\n|-------------------------------+----------------------+----------------------+\r\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\r\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\r\n|===============================+======================+======================|\r\n|   0  GeForce GTX 970M    Off  | 00000000:01:00.0 Off |                  N\/A |\r\n| N\/A   51C    P8     7W \/  N\/A |    334MiB \/  3024MiB |      0%      Default |\r\n+-------------------------------+----------------------+----------------------+\r\n                                                                               \r\n+-----------------------------------------------------------------------------+\r\n| Processes:                                                       GPU Memory |\r\n|  GPU       PID   Type   Process name                             Usage      |\r\n|=============================================================================|\r\n|    0       980      G   \/usr\/lib\/xorg\/Xorg                           227MiB |\r\n|    0      1248      G   \/usr\/bin\/gnome-shell                         102MiB |\r\n|    0      2671      G   \/usr\/lib\/firefox\/firefox                       2MiB |\r\n+-----------------------------------------------------------------------------+&lt;\/pre&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[]}
+{"QuestionId":52638722,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2018-10-04T03:49:45.550","AcceptedAnswerId":null,"OwnerUserId":258274.0,"Title":"How does pytorch nn.module keep computation graph and when&how release it?","Body":"<p>I got a piece of code from:\n<a href=\"https:\/\/github.com\/liorshk\/facenet_pytorch\/blob\/master\/train_triplet.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/liorshk\/facenet_pytorch\/blob\/master\/train_triplet.py<\/a><\/p>\n\n<pre><code>def train(train_loader, model, optimizer, epoch):\n# switch to train mode\nmodel.train()\n\npbar = tqdm(enumerate(train_loader))\nlabels, distances = [], []\n\n\nfor batch_idx, (data_a, data_p, data_n,label_p,label_n) in pbar:\n\n    data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()\n    data_a, data_p, data_n = Variable(data_a), Variable(data_p), \\\n                             Variable(data_n)\n\n    # compute output\n    out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)\n\n    # Choose the hard negatives\n    d_p = l2_dist.forward(out_a, out_p)\n    d_n = l2_dist.forward(out_a, out_n)\n    all = (d_n - d_p &lt; args.margin).cpu().data.numpy().flatten()\n    hard_triplets = np.where(all == 1)\n    if len(hard_triplets[0]) == 0:\n        continue\n    out_selected_a = Variable(torch.from_numpy(out_a.cpu().data.numpy()[hard_triplets]).cuda())\n    out_selected_p = Variable(torch.from_numpy(out_p.cpu().data.numpy()[hard_triplets]).cuda())\n    out_selected_n = Variable(torch.from_numpy(out_n.cpu().data.numpy()[hard_triplets]).cuda())\n\n    selected_data_a = Variable(torch.from_numpy(data_a.cpu().data.numpy()[hard_triplets]).cuda())\n    selected_data_p = Variable(torch.from_numpy(data_p.cpu().data.numpy()[hard_triplets]).cuda())\n    selected_data_n = Variable(torch.from_numpy(data_n.cpu().data.numpy()[hard_triplets]).cuda())\n\n    selected_label_p = torch.from_numpy(label_p.cpu().numpy()[hard_triplets])\n    selected_label_n= torch.from_numpy(label_n.cpu().numpy()[hard_triplets])\n    triplet_loss = TripletMarginLoss(args.margin).forward(out_selected_a, out_selected_p, out_selected_n)\n\n    cls_a = model.forward_classifier(selected_data_a)\n    cls_p = model.forward_classifier(selected_data_p)\n    cls_n = model.forward_classifier(selected_data_n)\n\n    criterion = nn.CrossEntropyLoss()\n    predicted_labels = torch.cat([cls_a,cls_p,cls_n])\n    true_labels = torch.cat([Variable(selected_label_p.cuda()),Variable(selected_label_p.cuda()),Variable(selected_label_n.cuda())])\n\n    cross_entropy_loss = criterion(predicted_labels.cuda(),true_labels.cuda())\n\n    loss = cross_entropy_loss + triplet_loss\n    # compute gradient and update weights\n    optimizer.zero_grad()\n    loss.backward()\n    optimizer.step()\n\n    # update the optimizer learning rate\n    adjust_learning_rate(optimizer)\n\n    # log loss value\n    logger.log_value('triplet_loss', triplet_loss.data[0]).step()\n    logger.log_value('cross_entropy_loss', cross_entropy_loss.data[0]).step()\n    logger.log_value('total_loss', loss.data[0]).step()\n    if batch_idx % args.log_interval == 0:\n        pbar.set_description(\n            'Train Epoch: {} [{}\/{} ({:.0f}%)]\\tLoss: {:.6f} \\t # of Selected Triplets: {}'.format(\n                epoch, batch_idx * len(data_a), len(train_loader.dataset),\n                100. * batch_idx \/ len(train_loader),\n                loss.data[0],len(hard_triplets[0])))\n\n\n    dists = l2_dist.forward(out_selected_a,out_selected_n) #torch.sqrt(torch.sum((out_a - out_n) ** 2, 1))  # euclidean distance\n    distances.append(dists.data.cpu().numpy())\n    labels.append(np.zeros(dists.size(0)))\n\n\n    dists = l2_dist.forward(out_selected_a,out_selected_p)#torch.sqrt(torch.sum((out_a - out_p) ** 2, 1))  # euclidean distance\n    distances.append(dists.data.cpu().numpy())\n    labels.append(np.ones(dists.size(0)))\n\nlabels = np.array([sublabel for label in labels for sublabel in label])\ndistances = np.array([subdist[0] for dist in distances for subdist in dist])\n\ntpr, fpr, accuracy, val, val_std, far = evaluate(distances,labels)\nprint('\\33[91mTrain set: Accuracy: {:.8f}\\n\\33[0m'.format(np.mean(accuracy)))\nlogger.log_value('Train Accuracy', np.mean(accuracy))\n\nplot_roc(fpr,tpr,figure_name=\"roc_train_epoch_{}.png\".format(epoch))\n\n# do checkpointing\ntorch.save({'epoch': epoch + 1, 'state_dict': model.state_dict()},\n           '{}\/checkpoint_{}.pth'.format(LOG_DIR, epoch))\n<\/code><\/pre>\n\n<p>as you can see,there are a bunch of model.forward(..) invoking:  <\/p>\n\n<pre><code>out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)\n....\ncls_a = model.forward_classifier(selected_data_a)\ncls_p = model.forward_classifier(selected_data_p)\ncls_n = model.forward_classifier(selected_data_n)\n(model.forward_classifier() will invoke model.forward() also)\n<\/code><\/pre>\n\n<p>due to loss.backward() invoked,before second iteration beginning,a little bit gpu memory was release,but more than 30G gpu memory not(I have 4x1080ti),then training crashed with run out of GPU memory,so question are:<\/p>\n\n<p>1,In a nn.module,how(or where) is computation graph saved,how to release gpu memory?<\/p>\n\n<p>2,Must to call backward()?if yes,what should be in .backward() method?<\/p>\n\n<p>3,In my case,which part of GPU memory is released,which part of them is kept?Is there any way to see it?<\/p>\n","answers":[]}
+{"QuestionId":52639165,"AnswerCount":1,"Tags":"<python><tensorflow><mask>","CreationDate":"2018-10-04T04:44:09.363","AcceptedAnswerId":null,"OwnerUserId":6514350.0,"Title":"TensorFlow\uff1aHow to generate differentiable mask","Body":"<p>I want to generate a mask from a tensor. For example:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\ninput = tf.Variable(np.array([[0, 0.1], [0.9, 1.0]]))\n<\/code><\/pre>\n\n<p>And from <code>input<\/code>, I want to get a mask tensor with value:<\/p>\n\n<pre><code>[[0, 1.0], [1.0, 1.0]]\n<\/code><\/pre>\n\n<p>A way to do this is:<\/p>\n\n<pre><code>output = tf.cast(tf.cast(input, dtype=tf.bool), dtype=tf.float32)\n<\/code><\/pre>\n\n<p>However, <code>tf.cast<\/code> is not differentiable and the gradient can not back-propagate while I hope that the operation is differentiable. \nI also search other methods such as <code>tf.floor<\/code>, <code>tf.where<\/code> but they are all not differentiable. <\/p>\n\n<p>A possible method is to use these undifferentiable methods and define the gradient for them. However, I wonder what their gradients should be in order to make the back-propagation feasible (maybe just set the gradient to be 1 ?) and why?<\/p>\n","answers":[{"AnswerId":"52641127","CreationDate":"2018-10-04T07:19:55.957","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>As you suggested, you can artificially stick a gradient to a masking operation that is not differentiable. Here is an example, where the gradient is always one:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef my_mask(x):\n  return tf.to_float(tf.greater(x, 0))\n\ndef diff_mask(mask_op):\n  @tf.custom_gradient\n  def _diff_mask(x):\n    def grad(dy):\n      return dy * tf.ones_like(x)\n    return mask_op(x), grad\n  return _diff_mask\n\nx = tf.random_normal((5,))\nw = tf.random_normal((5,))\nm = diff_mask(my_mask)(w)\nloss = x * m\ng = tf.gradients([loss], [x, w])\n\nsess = tf.InteractiveSession()\nsess.run([x, loss] + g)\n<\/code><\/pre>\n\n<p>You could of course choose another gradient, for example the gradient of the sigmoid. It really depends on what you want to do; any choice will be wrong since thresholding is not differentiable, but hopefully one choice will be useful for you.<\/p>\n\n<p>A pitfall of this kind of technique is that now the gradient that you use is not the gradient of the loss you are minimizing. That is why people generally tends to use soft thresholds, rather than hard threshold. For example if you intend to use the gradient of the sigmoid because you feel you need a threshold in the feedback, then why not using the sigmoid itself as a mask?<\/p>\n"}]}
+{"QuestionId":52639289,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-04T04:57:12.180","AcceptedAnswerId":52807572.0,"OwnerUserId":602721.0,"Title":"Accessing Estimator evaluation results via SessionRunHooks","Body":"<p>I'm trying to modify a program that uses the Estimator class in TensorFlow (v1.10) and I would like to access the evaluation metric results every time evaluation occurs so that I can copy the checkpoint files only when a new maximum has been achieved.<\/p>\n\n<p>One idea I had was to create a class inheriting from <code>SessionRunHook<\/code>, doing the work I want in the <code>after_run<\/code> method. According to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.10\/api_docs\/python\/tf\/train\/SessionRunHook\" rel=\"nofollow noreferrer\">documentation<\/a> I can specify what is passed to <code>after_run<\/code> using <code>before_run<\/code>. However I cannot find a way to access the evaluation metrics results I want from the information passed in to <code>before_run<\/code>. <\/p>\n\n<p>I looked into the <code>Estimator<\/code> code and it appears that it is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4dcfddc5d12018a5a0fdca652b9221ed95e9eb23\/tensorflow\/python\/estimator\/estimator.py#L1669\" rel=\"nofollow noreferrer\">writing the results to a summary file<\/a> so another idea I had was to read this back in the <code>after_run<\/code> method, but the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.10\/api_docs\/python\/tf\/summary\" rel=\"nofollow noreferrer\">summary api<\/a> doesn't seem to provide any read operations.<\/p>\n\n<p>Are there any other ways I can achieve what I want to do? Not using the <code>Estimator<\/code> class is not an option as that would involve drastic changes to the code I'm working with.<\/p>\n","answers":[{"AnswerId":"52807572","CreationDate":"2018-10-14T22:21:37.957","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>Checkpoints are not the same as exporting. Checkpoints are about fault-recovery and involve saving the complete training state (weights, global step number, etc.). <\/p>\n\n<p>In your case I would recommend exporting. The exported model will written to a directory called \u201cexporter\u201d and the serving input function specifies what the end-user will be expected to provide to the prediction service. <\/p>\n\n<p>You can use the class \"Best Exporter\" to just export the models that are perfoming best:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/BestExporter\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/BestExporter<\/a><\/p>\n<\/blockquote>\n\n<p>This class exports the serving graph and checkpoints of the best models.<\/p>\n\n<p>Also, it performs a model export everytime when the new model is better than any exsiting model.<\/p>\n"}]}
+{"QuestionId":52640036,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T06:05:58.107","AcceptedAnswerId":null,"OwnerUserId":4015544.0,"Title":"keras with tensorflow backend: layer implementation when batch_size is used, but it is None during graph construction","Body":"<p>I am implementing a keras layer <code>AnchorTargets<\/code> for retinanet in object detection, it inputs <code>anchors<\/code> and <code>annotations<\/code> (i.e. ground-truth bounding-boxes). The ideas for the codes are motivated by <a href=\"https:\/\/github.com\/fizyr\/keras-retinanet\" rel=\"nofollow noreferrer\">keras-retinanet<\/a> in github. <\/p>\n\n<p>However, I got error in <code>labels_shape = (input_shape[0]*input_shape[1], num_classes)<\/code> since input_shape[0] denotes batch size, which is unavailable during network construction. Is there any suggestions for this issue? Thanks.<\/p>\n\n<pre><code>def batch_anchor_targets_bbox(\nanchors,\nannotations,\nnum_classes,\nmask_shape=None,\nnegative_overlap=0.4,\npositive_overlap=0.5,\n**kwargs):\n\"\"\" Generate anchor targets for bbox detection.\n\nArgs\n    anchors: np.array of annotations of shape (B, N, 4) for (x1, y1, x2, y2).\n    annotations: np.array of shape (B, M, 5) for (x1, y1, x2, y2, label).\n    num_classes: Number of classes to predict.\n    mask_shape: If the image is padded with zeros, mask_shape can be used to mark the relevant part of the image.\n    negative_overlap: IoU overlap for negative anchors (all anchors with overlap &lt; negative_overlap are negative).\n    positive_overlap: IoU overlap or positive anchors (all anchors with overlap &gt; positive_overlap are positive).\n\nReturns\n    labels: np.array of shape (B, N, num_classes) where a row consists of 0 for negative and 1 for positive for a certain class.\n    annotations: np.array of shape (B, N, 5) for (x1, y1, x2, y2, label) containing the annotations corresponding to each anchor or 0 if there is no corresponding anchor.\n    anchor_states: np.array of shape (B, N) containing the state of an anchor (-1 for ignore, 0 for bg, 1 for fg).\n\"\"\"\n# anchor states: 1 is positive, 0 is negative, -1 is dont care    \nanchors     = keras.backend.cast(anchors, 'float32')\nannotations = keras.backend.cast(annotations, 'float32')\ninput_shape = keras.backend.int_shape(anchors)\nannos_shape = keras.backend.int_shape(annotations)    \n\noverlaps = batch_compute_overlap(anchors, annotations)   \nargmax_overlaps_inds = keras.backend.argmax(overlaps, axis=2)\nmax_overlaps         = keras.backend.max(overlaps, axis=2)\n\n# assign \"dont care\" labels\nflag_pos = keras.backend.greater_equal(max_overlaps, positive_overlap)\nanchor_states = keras.backend.cast(flag_pos, 'float32')\nflag = backend.logical_and(keras.backend.greater_equal(max_overlaps, negative_overlap), \\\n                               keras.backend.less(max_overlaps, positive_overlap))\nanchor_states = anchor_states - keras.backend.cast(flag, 'float32')\n\n# reshape the values and indices of the max output\nargmax_overlaps_inds = keras.backend.reshape(argmax_overlaps_inds,(-1,))\nflag_pos = keras.backend.reshape(flag_pos,(-1,))\n\n# compute box regression targets\nannotations_o = keras.backend.reshape(annotations, (-1, annos_shape[2]))\nannotations_o = backend.gather(annotations_o,argmax_overlaps_inds)\nannotations = keras.backend.reshape(annotations_o, (-1, input_shape[1],annos_shape[2]))\n\n# compute target class labels\nind_x = backend.where(flag_pos)\nind_y = keras.backend.cast(backend.gather(annotations_o[:,4], ind_x), 'int64')\nindices = keras.backend.concatenate([ind_x,ind_y], axis=1)\nlabels_shape = (input_shape[0]*input_shape[1], num_classes)\nvalues = keras.backend.ones(keras.backend.shape(ind_x)[0])\ndelta = backend.SparseTensor(indices, values, labels_shape)\nlabels = backend.sparse_tensor_to_dense(delta)\nlabels = keras.backend.reshape(labels, (-1, input_shape[1], num_classes))\nreturn indices, annotations, anchor_states\n\n\nclass AnchorTargets(keras.layers.Layer):\n\ndef __init__(self, num_classes, negative_overlap=0.4, positive_overlap=0.5, *args, **kwargs):\n    self.num_classes = num_classes\n    self.negative_overlap = negative_overlap\n    self.positive_overlap = positive_overlap\n\n    super(AnchorTargets, self).__init__(*args, **kwargs)\n\ndef call(self, inputs, **kwargs): \n    # anchors (batch_size x N x 4): pre-defined anchors \n    # annotations (batch_size x #gt x 4): gt_bboxes.\n\n    anchors, annotations = inputs \n\n    # calculate regression and labels\n    labels, annotations, anchor_states = batch_anchor_targets_bbox(anchors, annotations,\n                            self.num_classes, self.negative_overlap, self.positive_overlap)\n    anchor_states = keras.backend.expand_dims(anchor_states,axis=2)\n    labels = keras.backend.concatenate([labels, anchor_states], axis=2)\n    regressions = keras.backend.concatenate([annotations, anchor_states], axis=2)\n    return [regressions, labels]\n\ndef compute_output_shape(self, input_shape):\n    return [(input_shape[0][0],input_shape[0][1],input_shape[0][2]+1),\\\n            (input_shape[0][0],input_shape[0][1],self.num_classes+2)]\n\ndef get_config(self):\n    config = super(AnchorTargets, self).get_config()\n    config.update({\n        'num_classes':      self.num_classes,\n        'negative_overlap': self.negative_overlap,\n        'positive_overlap': self.positive_overlap\n    })\n\n    return config\n<\/code><\/pre>\n","answers":[{"AnswerId":"52654806","CreationDate":"2018-10-04T20:38:40.793","ParentId":null,"OwnerUserId":"4015544","Title":null,"Body":"<p>Never mind. I have figured it out. The basic idea is to avoid the use of input_shape[0] (i.e. batch size) in <code>call<\/code> function of <code>keras.layers.Layer<\/code>. My implementation is as follows if anyone is interested in it.<\/p>\n\n<pre><code>def batch_anchor_targets_bbox(\nanchors,\nannotations,\nnum_classes,\nmask_shape=None,\nnegative_overlap=0.4,\npositive_overlap=0.5,\n**kwargs):\n\"\"\" Generate anchor targets for bbox detection.\n\nArgs\n    anchors: np.array of annotations of shape (B, N, 4) for (x1, y1, x2, y2).\n    annotations: np.array of shape (B, M, 5) for (x1, y1, x2, y2, label).\n    num_classes: Number of classes to predict.\n    mask_shape: If the image is padded with zeros, mask_shape can be used to mark the relevant part of the image.\n    negative_overlap: IoU overlap for negative anchors (all anchors with overlap &lt; negative_overlap are negative).\n    positive_overlap: IoU overlap or positive anchors (all anchors with overlap &gt; positive_overlap are positive).\n\nReturns\n    labels: np.array of shape (B, N, num_classes) where a row consists of 0 for negative and 1 for positive for a certain class.\n    annotations: np.array of shape (B, N, 5) for (x1, y1, x2, y2, label) containing the annotations corresponding to each anchor or 0 if there is no corresponding anchor.\n    anchor_states: np.array of shape (B, N) containing the state of an anchor (-1 for ignore, 0 for bg, 1 for fg).\n\"\"\"\n# anchor states: 1 is positive, 0 is negative, -1 is dont care    \nanchors     = keras.backend.cast(anchors, 'float32')\nannotations = keras.backend.cast(annotations, 'float32')\ninput_shape = keras.backend.int_shape(anchors)\nannos_shape = keras.backend.int_shape(annotations)    \n\noverlaps = batch_compute_overlap(anchors, annotations)   \nargmax_overlaps_inds = keras.backend.argmax(overlaps, axis=2)\nmax_overlaps         = keras.backend.max(overlaps, axis=2)\n\n# assign \"dont care\" labels\nflag_pos = keras.backend.greater_equal(max_overlaps, positive_overlap)\nanchor_states = keras.backend.cast(flag_pos, 'float32')\nflag = backend.logical_and(keras.backend.greater_equal(max_overlaps, negative_overlap), \\\n                               keras.backend.less(max_overlaps, positive_overlap))\nanchor_states = anchor_states - keras.backend.cast(flag, 'float32')\n\n# reshape the values and indices of the max output\nargmax_overlaps_inds = keras.backend.reshape(argmax_overlaps_inds,(-1,))\nflag_pos = keras.backend.reshape(flag_pos,(-1,))\n\n# compute box regression targets\nannotations_o = keras.backend.reshape(annotations, (-1, annos_shape[2]))\nannotations_o = backend.gather(annotations_o,argmax_overlaps_inds)\nannotations = keras.backend.reshape(annotations_o, (-1, input_shape[1],annos_shape[2]))\n\n# compute target class labels\nanchor_states_o = -1*keras.backend.cast(keras.backend.less(max_overlaps, positive_overlap), 'float32')\nanchor_states_o = keras.backend.reshape(anchor_states_o, (-1,))\nind_y = backend.where(keras.backend.greater_equal(anchor_states_o,-0.5), annotations_o[:,4], anchor_states_o)\nlabels = keras.backend.one_hot(keras.backend.cast(ind_y,'int32'), num_classes)\nlabels = keras.backend.reshape(labels, (-1, input_shape[1], num_classes))\nreturn labels, annotations, anchor_states\n<\/code><\/pre>\n"}]}
+{"QuestionId":52641501,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T07:42:07.710","AcceptedAnswerId":null,"OwnerUserId":4015544.0,"Title":"tensorflow assign values of 2D tensor by indices when 1st dimension of the tensor is unknown [in keras.layers.Layer]","Body":"<p>I am implementing a keras layer, and its <code>call<\/code> function needs to achieve 2D Tensor value assign by indices. The function is demonstrated by python numpy as follows:<\/p>\n\n<pre><code>import numpy as np\nA = np.arange(12).reshape(6,2)\nind_x = [2,4,5]\nind_y = [0,1,0]\nA[(ind_x,ind_y)] = -1 # assign value \nA = array([[ 0,  1],\n           [ 2,  3],\n           [-1,  5],\n           [ 6,  7],\n           [ 8, -1],\n           [-1, 11]])\n<\/code><\/pre>\n\n<p>If <code>A.shape = (6,2)<\/code> is available, then <code>tf.SparseTensor<\/code> is perfect for this purpose. <\/p>\n\n<p>However in my case, A is the input of keras.layers.Layer, thus its first dimension (i.e. batch size) is undetermined during graph construction. In the end, <code>A.shape = (None,2)<\/code>. Is there any other tensorflow function can achieve this? Thanks.<\/p>\n\n<p>P.S. A question was asked with full keras implementation <a href=\"https:\/\/stackoverflow.com\/questions\/52640036\/keras-with-tensorflow-backend-layer-implementation-when-batch-size-is-used-but\">here<\/a>. However, I think the question might be clearer and more fundamental here.<\/p>\n","answers":[]}
+{"QuestionId":52641518,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-04T07:43:03.177","AcceptedAnswerId":null,"OwnerUserId":10333833.0,"Title":"How to print filter after training the network? Just want to see the values of weights of the filter after training the network","Body":"<p>I have a CNN network with one convolution, maximum pooling and one fully connected layer. After training the network I want to see the final value of weights that filter has adopted. What I have to do for this purpose? Should I have to print W matrix of fully connected layer? Should I have to replace some weights?I am training the network using a batch size of 30. <\/p>\n\n<pre><code>`\ndef weight_variable(shape):\n    initial = tf.truncated_normal(shape, mean=0, stddev=0.1)\nreturn tf.Variable(initial)\n#network\nx = tf.placeholder(tf.float32, [None, \nFLAGS.image_height*FLAGS.image_width])\ny_ = tf.placeholder(tf.float32, [None, 2])\n\n\n\ninput=tf.reshape(x, \n   [-1,FLAGS.image_height,FLAGS.image_width,FLAGS.input_channel])\nfilter = weight_variable([FLAGS.filter_size, FLAGS.filter_size, \n   FLAGS.input_channel, FLAGS.filter_channel])\nconv_out = tf.nn.sigmoid(conv2d(input, filter))\n\npool_out = max_pool(conv_out)\n\n\npool_list = pool_out.get_shape().as_list()\ninput_dim = pool_list[1]* pool_list[2]* pool_list[3]\npool_2D = tf.reshape(pool_out, [-1, input_dim])\nW_fc = weight_variable([input_dim, 2])\n\nlogits = tf.matmul(pool_2D, W_fc)   #(batch_size,2)\ny_conv=tf.nn.softmax(logits)\ns=tf.argmax(y_conv,axis=1)\ntest = tf.reduce_sum(s)\n\ncross_entropy = \ntf.nn.softmax_cross_entropy_with_logits(logits=logits,labels=y_)\nloss = tf.reduce_mean(cross_entropy)\noptimizer = tf.train.GradientDescentOptimizer(FLAGS.rLearn).minimize(loss)\n\n\ncorrect_prediction = tf.equal(tf.argmax(y_conv,axis=1), tf.argmax(y_, \naxis=1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n#training loop\nfor i in range(200):`\n<\/code><\/pre>\n","answers":[{"AnswerId":"52651491","CreationDate":"2018-10-04T16:26:53.037","ParentId":null,"OwnerUserId":"9179860","Title":null,"Body":"<p>I don't understand the way you initialize the weight matrix.<\/p>\n\n<p>Assuming that you initialize weigh matrix in this way:<\/p>\n\n<pre><code>def new_weights(shape):\n    return tf.Variable(tf.truncated_normal(shape, stddev=0.05)) \n<\/code><\/pre>\n\n<p>After training you can retrieve the values of the weight-variables in this way:<\/p>\n\n<pre><code># Retrieve the values of the weight-variables from TensorFlow.\n# A feed-dict is not necessary because nothing is calculated.\nw = session.run(weights)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52641737,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-04T07:55:36.463","AcceptedAnswerId":52642619.0,"OwnerUserId":5623899.0,"Title":"TensorFlow 1.10+ custom estimator early stopping with train_and_evaluate","Body":"<p>Suppose you are training a custom <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"noreferrer\"><code>tf.estimator.Estimator<\/code><\/a> with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/train_and_evaluate\" rel=\"noreferrer\"><code>tf.estimator.train_and_evaluate<\/code><\/a> using a validation dataset in a setup similar to that of <a href=\"https:\/\/stackoverflow.com\/a\/47228832\/5623899\">@simlmx's<\/a>:<\/p>\n\n<pre><code>classifier = tf.estimator.Estimator(\n    model_fn=model_fn,\n    model_dir=model_dir,\n    params=params)\n\ntrain_spec = tf.estimator.TrainSpec(\n    input_fn = training_data_input_fn,\n)\n\neval_spec = tf.estimator.EvalSpec(\n    input_fn = validation_data_input_fn,\n)\n\ntf.estimator.train_and_evaluate(\n    classifier,\n    train_spec,\n    eval_spec\n)\n<\/code><\/pre>\n\n<p>Often, one uses a validation dataset to cut off training to prevent over-fitting when the loss continues to improve for the training dataset but not for the validation dataset.<\/p>\n\n<p>Currently the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EvalSpec\" rel=\"noreferrer\"><code>tf.estimator.EvalSpec<\/code><\/a> allows one to specify after how many <code>steps<\/code> (defaults to 100) to evaluate the model. <\/p>\n\n<p>How can one (if possible not using <code>tf.contrib<\/code> functions) designate to terminate training after <code>n<\/code> number of evaluation calls (<code>n * steps<\/code>) where the evaluation loss does not improve and then save the \"best\" model \/ checkpoint  (determined by validation dataset) to a unique file name (e.g. <code>best_validation.checkpoint<\/code>)<\/p>\n","answers":[{"AnswerId":"52642619","CreationDate":"2018-10-04T08:46:46.863","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<p>I understand your confusion now. The documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/estimator\/stop_if_no_decrease_hook\" rel=\"noreferrer\"><code>stop_if_no_decrease_hook<\/code><\/a> states (emphasis mine):<\/p>\n\n<blockquote>\n  <p>max_steps_without_decrease: int, maximum number of <strong>training steps<\/strong> with\n  no decrease in the given metric. <\/p>\n  \n  <p>eval_dir: If set, directory\n  containing summary files with eval metrics. By default,\n  estimator.eval_dir() will be used.<\/p>\n<\/blockquote>\n\n<p>Looking through <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/contrib\/estimator\/python\/estimator\/early_stopping.py\" rel=\"noreferrer\">the code of the hook (version 1.11)<\/a>, though, you find:<\/p>\n\n<pre><code>def stop_if_no_metric_improvement_fn():\n    \"\"\"Returns `True` if metric does not improve within max steps.\"\"\"\n\n    eval_results = read_eval_metrics(eval_dir) #&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;\n\n    best_val = None\n    best_val_step = None\n    for step, metrics in eval_results.items(): #&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;&lt;\n      if step &lt; min_steps:\n        continue\n      val = metrics[metric_name]\n      if best_val is None or is_lhs_better(val, best_val):\n        best_val = val\n        best_val_step = step\n      if step - best_val_step &gt;= max_steps_without_improvement: #&lt;&lt;&lt;&lt;&lt;\n        tf_logging.info(\n            'No %s in metric \"%s\" for %s steps, which is greater than or equal '\n            'to max steps (%s) configured for early stopping.',\n            increase_or_decrease, metric_name, step - best_val_step,\n            max_steps_without_improvement)\n        return True\n    return False\n<\/code><\/pre>\n\n<p>What the code does is load the evaluation results (produced with your <code>EvalSpec<\/code> parameters) and extract the eval results and the <code>global_step<\/code> (or whichever other custom step you use to count) associated with the specific evaluation record.<\/p>\n\n<p>This is the source of the <code>training steps<\/code> part of the docs: the early stopping is not triggered according to the number of non-improving evaluations, but to the number of non-improving evals in a certain step range (which IMHO is a bit counter-intuitive).<\/p>\n\n<p>So, to recap: <strong>Yes<\/strong>, the early-stopping hook uses the evaluation results to decide when it's time to cut the training, <strong>but<\/strong> you need to pass in the number of training steps you want to monitor and keep in mind how many evaluations will happen in that number of steps.<\/p>\n\n<h3>Examples with numbers to hopefully clarify more<\/h3>\n\n<p>Let's assume you're training indefinitely long having an evaluation every 1k steps. The specifics of how the evaluation runs is not relevant, as long as it runs every 1k steps producing a metric we want to monitor.<\/p>\n\n<p>If you set the hook as <code>hook = tf.contrib.estimator.stop_if_no_decrease_hook(my_estimator, 'my_metric_to_monitor', 10000)<\/code> the hook will consider the evaluations happening in a range of 10k steps. <\/p>\n\n<p>Since you're running 1 eval every 1k steps, this boils down to early-stopping if there's a sequence of 10 consecutive evals without any improvement.\nIf then you decide to rerun with evals every 2k steps, the hook will only consider a sequence of 5 consecutive evals without improvement.<\/p>\n\n<h3>Keeping the best model<\/h3>\n\n<p>First of all, an important note: <strong>this has nothing to do with early stopping<\/strong>, the issue of keeping a copy of the best model through the training and the one of stopping the training once performance start degrading are completely unrelated.<\/p>\n\n<p>Keeping the best model can be done very easily defining a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/BestExporter\" rel=\"noreferrer\"><code>tf.estimator.BestExporter<\/code><\/a> in your <code>EvalSpec<\/code> (snippet taken from the link):<\/p>\n\n<pre><code>  serving_input_receiver_fn = ... # define your serving_input_receiver_fn\n  exporter = tf.estimator.BestExporter(\n      name=\"best_exporter\",\n      serving_input_receiver_fn=serving_input_receiver_fn,\n      exports_to_keep=5) # this will keep the 5 best checkpoints\n\n  eval_spec = [tf.estimator.EvalSpec(\n    input_fn=eval_input_fn,\n    steps=100,\n    exporters=exporter,\n    start_delay_secs=0,\n    throttle_secs=5)]\n<\/code><\/pre>\n\n<p>If you don't know how to define the <code>serving_input_fn<\/code> <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#using_savedmodel_with_estimators\" rel=\"noreferrer\">have a look here<\/a><\/p>\n\n<p>This allows you to keep the overall best 5 models you obtained, stored as <code>SavedModel<\/code>s (which is the preferred way to store models at the moment).<\/p>\n"}]}
+{"QuestionId":52642756,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-04T08:54:24.803","AcceptedAnswerId":52649588.0,"OwnerUserId":583464.0,"Title":"memory error in predict_on_batch() on large data set","Body":"<p>I have a test set of 18000 examples.<\/p>\n\n<pre><code>\u03a7_test.shape: (18000, 128, 128, 1)\n<\/code><\/pre>\n\n<p>I have trained my model and want to use predict on X_test.<\/p>\n\n<p>If I try to just use:<\/p>\n\n<pre><code>pred = model.predict_on_batch(X_test)\n<\/code><\/pre>\n\n<p>it gives a memory error.<\/p>\n\n<p>I tried something like:<\/p>\n\n<pre><code>X_test_split = X_test.flatten()\nX_test_split = np.array_split(X_test_split, 562) # batch size is 32\npred = np.empty(len(X_test_split), dtype=np.float32)\n\nfor idx, _ in enumerate(X_test_split):\n    pred[idx] = model.predict_on_batch(X_test_split[idx].reshape(32, 128, 128, 1))\n<\/code><\/pre>\n\n<p>but it either gives me memory error again or it gives me error about the reshape (depending on variations I am trying in the above code)<\/p>\n\n<p>I have the same problem using <code>predict_generator<\/code> also.<\/p>\n","answers":[{"AnswerId":"52649588","CreationDate":"2018-10-04T14:45:46.883","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As requested by OP, I am posting my comment as an answer and try to elaborate more:<\/p>\n\n<p>It seems your model size is big and therefore you need to either use a smaller batch size (&lt; 32, since you mentioned it does not work with 32) or modify the model and decrease the number of parameters (e.g. removing some layers, decreasing the number of filter or units, etc.).<\/p>\n"}]}
+{"QuestionId":52643025,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-04T09:08:31.730","AcceptedAnswerId":52645553.0,"OwnerUserId":7167328.0,"Title":"How to use ModelCheckpoint() in Keras with weighted validation loss","Body":"<p>I train a DNN in Keras which has high imbalanced classes. So I used class_weight in fit_generator to correct this. Now I want to save the model with the lowest weighted validation loss using the ModelCheckpoint() function. I am trying but I can't figure out the way to achieve this. Would any one have a simple example?   <\/p>\n","answers":[{"AnswerId":"52645553","CreationDate":"2018-10-04T11:27:52.953","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<pre><code>ModelCheckpoint(\"checkpoint.hdf5\", monitor='val_loss', mode = 'min', verbose=1, save_best_only = True)\nmodel.fit_genetor(....)\n<\/code><\/pre>\n\n<p>I think you are asking for this piece of code.<\/p>\n"}]}
+{"QuestionId":52643439,"AnswerCount":0,"Tags":"<c++><tensorflow><cmake>","CreationDate":"2018-10-04T09:31:30.567","AcceptedAnswerId":null,"OwnerUserId":3370378.0,"Title":"Using libtensorflow_cc.so in an existing C++ cmake project","Body":"<p>I have an existing C++ cmake project and I want to use tensorflow in it as a third party library. I searched extensively but all the solutions involve installing tensorflow at the end or copying the files to \/usr\/local directory.<\/p>\n\n<p>My problem is I don't want to make any changes to the system directories and everything should happen within my project directory.<\/p>\n\n<p>So, I compiled libtensorflow_cc.so and libtensorflow_framework.so from tensorflow source code and copied into my project directory. I added below code in my CMake to link these libraries.<\/p>\n\n<pre><code> find_library(tensorflowlibfwk libtensorflow_framework.so \n PATHS ${PROJECT_SOURCE_DIR}\/thirdparty\/lib)\n find_library(tensorflowlibcc libtensorflow_cc.so\n PATHS ${PROJECT_SOURCE_DIR}\/thirdparty\/lib)\n target_link_libraries(tensorflowtest ${tensorflowlibfwk} ${tensorflowlibcc})\n<\/code><\/pre>\n\n<p>I have no idea about which headers should I include to use the tensorflow in my project ??\nAny kind of help please<\/p>\n","answers":[]}
+{"QuestionId":52643550,"AnswerCount":1,"Tags":"<android><tensorflow><android-bitmap>","CreationDate":"2018-10-04T09:38:33.147","AcceptedAnswerId":52644698.0,"OwnerUserId":4560470.0,"Title":"Android normalize bitmap","Body":"<p>I am integrating a Tensorflow model in an Android app and I need to preprocess my image before feeding it to my model.<\/p>\n\n<p>For that I need my image to :<\/p>\n\n<ol>\n<li><p>crop my image to the right dimensions (no problem with this step) : <\/p>\n\n<p><code>Bitmap bitmap_cropped = Bitmap.createScaledBitmap(bitmap, 224, 224,false);<\/code><\/p><\/li>\n<li><p>get RGB chanels from 0-255 values to 0-1 range. So basically divide all pixels by 255. I tried the code below but I am not sure this does what I need :<\/p>\n\n<pre><code>bitmap_cropped.getPixels(imageBitmapPixels, 0, bitmap_cropped.getWidth(), 0, 0, bitmap_cropped.getWidth(), bitmap_cropped.getHeight());\nfor (int i = 0 ; i &lt; imageBitmapPixels.length; i++) {\n    imageBitmapPixels[i] = imageBitmapPixels[i]\/255;\n}\n<\/code><\/pre>\n\n<p>But when I display the values of each pixels I get something like that <\/p>\n\n<pre><code>-12898260\n-12963540\n-12963540\n-11254211\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>I guess the 3 chanels are mixed so I can't check if got what I wanted.. Is this the right way to do it ?<\/p>\n","answers":[{"AnswerId":"52644698","CreationDate":"2018-10-04T10:39:38.137","ParentId":null,"OwnerUserId":"2349435","Title":null,"Body":"<p>This can give you some idea of what you need to do<\/p>\n\n<pre><code>val rgbs = Array(bitmap.width * bitmap.height){\n    val x = it % bitmap.height\n    val y = it \/ bitmap.height\n\n    val pixel = bitmap.getPixel(x, y)\n    val red = (pixel and 0x00FF0000 shr 16) \/ 255f\n    val green = (pixel and 0x0000FF00 shr 8) \/ 255f\n    val blue = (pixel and 0x000000FF) \/ 255f\n\n    RGB(red, green, blue)\n}\n\ndata class RGB(val r: Float, val g: Float, val b: Float)\n<\/code><\/pre>\n\n<p>Excuse my kotlin<\/p>\n\n<p>disclamer 1: this code is not performant<\/p>\n\n<p>disclamer 2: there is probably a lib which doing pretty much the same but faster<\/p>\n\n<p>disclamer 3: the code above relies on your bitmap being ARGB_8888 format <\/p>\n"}]}
+{"QuestionId":52643866,"AnswerCount":0,"Tags":"<python><deep-learning><caffe2>","CreationDate":"2018-10-04T09:56:27.490","AcceptedAnswerId":null,"OwnerUserId":4108160.0,"Title":"AttributeError: module 'caffe2.python._import_c_extension' has no attribute 'get_cudnn_version'","Body":"<p>I'm struggling with this error for a few days now and I have no clue how to solve this. This is the error I get when I try to import the core module from the Caffe2 package.<\/p>\n\n<pre><code>from caffe2.python import core\n\nAttributeError: module 'caffe2.python._import_c_extension' has no attribute 'get_cudnn_version'\n<\/code><\/pre>\n\n<p>I searched a lot on Google but couldn't find any answers that could solve this problem. Does someone know how to solve this?<\/p>\n","answers":[]}
+{"QuestionId":52644160,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T10:13:08.943","AcceptedAnswerId":null,"OwnerUserId":462794.0,"Title":"Simple tensorflow example :","Body":"<p>So i am trying to make a micro example of how to use tensorflow :<\/p>\n\n<p>Here are my data : <\/p>\n\n<pre><code>train_images = numpy.array([[1,1,1],[2,2,2],[3,3,3],[1,2,2],[2,1,3],[3,21]]) \n<\/code><\/pre>\n\n<p>The main goal is to find that the first number of each array is equal to the label.<\/p>\n\n<p>Here is my code :<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport numpy\n\n\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(3, 3)),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(),\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nclass_names = ['chiffre 1',' chiffre 2',' chiffre 3']\n\ntrain_images = numpy.array([[1,1,1],[2,2,2],[3,3,3],[1,2,2],[2,1,3],[3,21]])\n\ntrain_labels = numpy.array(['1','2','3','1','2','3'])\n\nmodel.fit(train_images, train_labels, epochs=5)\n\ntest_images =numpy.array([[1,1,1],[2,2,2],[3,3,3],[1,2,2]])\n\ntest_labels = numpy.array(['1','2','3','1'])\n\ntest_loss, test_acc = model.evaluate(test_images, test_labels)\n\ntest_image =numpy.array([[1,1,1]])\npredictions = model.predict(test_image)\n\npred = numpy.argmax(predictions[0])\nprint(pred)\nprint(class_names[pred])\nprint(predictions)\n\n#print('Test accuracy:', test_acc)\n<\/code><\/pre>\n\n<p>I know that this line is false :<\/p>\n\n<pre><code>keras.layers.Flatten(input_shape=(3, 3)),\n<\/code><\/pre>\n\n<p>but how to make it correct ?\nRegards<\/p>\n","answers":[{"AnswerId":"52645327","CreationDate":"2018-10-04T11:14:25.557","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Start with deleting your Flatten layer, your data points are 1d vectors, flattening 1d vectors is meaningless. <\/p>\n\n<pre><code>model = keras.Sequential([\nkeras.layers.Dense(128, activation=tf.nn.relu , input_shape=(3,)),\nkeras.layers.Dense(3, activation=tf.nn.softmax)])\n<\/code><\/pre>\n\n<p>first dense layer takes three inputs, this is the size of your each data point.<\/p>\n\n<p>Since there are three classes, we need to change the output shape of the second layer too.<\/p>\n\n<pre><code>train_labels = numpy.array([0,1,2,0,1,2])\ntest_labels = numpy.array([0,1,2,0])\n<\/code><\/pre>\n\n<p>sparse_categorical_crossentropy loss function expects train and test labels to be in range of [0, 3). so we map 1 to 0 , 2 to 1 and 3 to 2. <\/p>\n\n<p>model, runs and learns with these changes, I suggest Increasing epoch number since you have very limited data.<\/p>\n"}]}
+{"QuestionId":52645302,"AnswerCount":0,"Tags":"<r><neural-network><keras><loss-function>","CreationDate":"2018-10-04T11:13:04.163","AcceptedAnswerId":null,"OwnerUserId":902265.0,"Title":"Custom Loss and Slicing (Keras R)","Body":"<p>I am trying to write a simple custom loss in Keras [R] and experiencing a bit of trouble. Consider a simple network that output a vector of dimension 2, for example<\/p>\n\n<pre><code>  model &lt;- keras_model_sequential() \n  model %&gt;% layer_dense(units = 2, input_shape = c(1))\n<\/code><\/pre>\n\n<p>Now, let us suppose (I know, that is silly) that one would like a loss that reads:<\/p>\n\n<pre><code>                      LOSS = (y_truth - y_pred[1])**2\n<\/code><\/pre>\n\n<p>where y_truth is 1-dimensional, and y_pred, the output of the network, is 2-dimensional (so, one completely ignore the second dimension). The following does not seem to work:<\/p>\n\n<pre><code>my_loss &lt;- function(y_true, y_pred){\n  K &lt;- backend()\n  loss &lt;- K$mean( K$pow(y_true - y_pred[,1],2))) \n  return(loss)\n}\n<\/code><\/pre>\n\n<p>I have tried a few variants, without success. Any idea how to do that?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>PS: what I'm really trying to do is write a loss function that would give me a Gaussian log-likelihood. The output of the network would be the mean and log_variance...<\/p>\n\n<p>[EDIT]\nit seems that the following variant works:<\/p>\n\n<pre><code>my_loss &lt;- function(y_true, y_pred){\n  K &lt;- backend()\n  loss &lt;- K$mean( K$pow(y_true - y_pred[,1:2],2))) \n  return(loss)\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52645962,"AnswerCount":0,"Tags":"<c++><tensorflow><build><cmake><eigen3>","CreationDate":"2018-10-04T11:50:09.963","AcceptedAnswerId":null,"OwnerUserId":3370378.0,"Title":"Using original eigen3 library for tensorflow","Body":"<p>I have an existing cmake C++ project and I want to use tensorflow as a shared library in it. Since my project required eigen3, I already have that library in my system (\/usr\/include\/eigen3). Now the problem is that tensorflow also uses eigen3 but the one present in third_party folder (tensorflow\/third_party\/eigen3\/) in its own directory. So, is it possible to use the original eigen3 library present on my system and not the one in tensorflow repository.<\/p>\n\n<p>I checked the build scripts but I could not find anything where I can specify to use the eigen3 present in the system<\/p>\n","answers":[]}
+{"QuestionId":52646836,"AnswerCount":0,"Tags":"<python><python-3.x><numpy><tensorflow>","CreationDate":"2018-10-04T12:33:40.153","AcceptedAnswerId":null,"OwnerUserId":2261278.0,"Title":"Tensor Flow Import Error : ModuleNotFoundError: No module named 'google'","Body":"<p>Kindly help with this error: <img src=\"https:\/\/i.stack.imgur.com\/8QsjX.png\" alt=\"enter image description here\"><\/p>\n\n<p>While importing tensorflow , i am facing above error as<\/p>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'google'<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52647019,"AnswerCount":1,"Tags":"<tensorflow><keras><pooling>","CreationDate":"2018-10-04T12:42:25.530","AcceptedAnswerId":null,"OwnerUserId":4203012.0,"Title":"Simple Average Pooling with Tensorflow","Body":"<p>I'm trying to do some very simple average pooling on a Keras \/ Tensorflow Tensor (not a layer in a network). I have a list of 18 embeddings (embedding = 2D vector) and want to average pool them with a pool-size of 3 with no overlap. So the first 3 embeddings should be averaged to an embedding, then the next 3 and so on. In the end, I need a list of 18\/3 = 6 embeddings.<\/p>\n\n<p>Here's my code so far:<\/p>\n\n<pre><code>    sequence_len = 3 # The number of embeddings that should be averaged\n    feature_dim = 2 # the dimensions of embeddings\n    input = [\n        [1, 3],  # Anchor 1\n        [2, 2],\n        [3, 1],\n        [4, 1],  # Anchor 2\n        [6, 1],\n        [8, 1],\n        [12, 3],  # Positive 1\n        [0, 6],\n        [0, 0],\n        [1, 1],  # Positive 2\n        [1, 1],\n        [1, 1],\n        [5, 0],  # Negative 1\n        [5, 12],\n        [5, 0],\n        [1, 1],  # Negative 2\n        [1, 1],\n        [1, 1],\n    ]\n    expected_out = [\n        [2, 2],  # Anchor 1\n        [6, 1],  # Anchor 2\n        [4, 3],  # Positive 1\n        [1, 1],  # Positive 2\n        [5, 4],  # Negative 1\n        [1, 1],  # Negative 2\n    ]\n    input_tensor = K.variable(input)\n    output_tensor = a_function_i_cant_figure_out(input_tensor) # What's the API to call?\n    assert numpy.array_equal(K.eval(output_tensor), expected_out)\n<\/code><\/pre>\n\n<p>I've spent over 2 hours on it now and I can't figure it out. I don't understand how I would have to reshape my data to get it to work with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/pool\" rel=\"nofollow noreferrer\">pooling function<\/a>.<\/p>\n\n<p>The closes I've come to a working solution is wrapping my input in another 2 lists (making it shape <code>(1,1,18,2)<\/code>) and calling this<\/p>\n\n<pre><code>output_tensor = K.pool2d(foo, (3, 2), (3, 2), pool_mode=\"avg\", data_format=\"channels_first\")\n<\/code><\/pre>\n\n<p>But this also averages the embeddings themselves. So instead of <code>[[2,2], [6,1], ...]<\/code> I'd receive <code>[2, 3.5, ...]<\/code><\/p>\n\n<p>This seems like such a trivial operation - how can I do this?<\/p>\n","answers":[{"AnswerId":"52647985","CreationDate":"2018-10-04T13:29:17.330","ParentId":null,"OwnerUserId":"4203012","Title":null,"Body":"<p>Ok, figured it out:<\/p>\n\n<pre><code>output_tensor = K.tf.nn.pool(input_tensor, [1, sequence_len], strides=(1, sequence_len), pooling_type=\"AVG\", padding=\"VALID\")\n<\/code><\/pre>\n\n<p>Think of my data as a 2D matrix.<\/p>\n\n<p><code>[1, sequence_len]<\/code> is the pool\/window size. So in X direction, we have a width of 1 and in Y direction a height of <code>sequence_len<\/code>. This means: The first value of <code>sequence_len<\/code> rows will be combined to one value. Then the stride of <code>(1, sequence_len)<\/code> means: We move the pool by 1 in X direction and once we're done with the row we move the pool down by <code>sequence_len<\/code>. <\/p>\n\n<p>We go with a width of 1 and a X-step of 1 column by column through the data which means our output has the same width. In Y direction we pool <code>sequence_len<\/code> rows to 1 row and then take the next <code>sequence_len<\/code> rows with no overlap. So the height of our output is <code>original_height \/ sequence_len<\/code><\/p>\n"}]}
+{"QuestionId":52647115,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-04T12:47:48.900","AcceptedAnswerId":52650092.0,"OwnerUserId":1115237.0,"Title":"python - Implementing an LSTM network with Keras and TensorFlow","Body":"<p>With limited knowledge, I've built an LSTM network. I would like to validate my assumptions and better understand the Keras API.<\/p>\n\n<p>Network Code:<\/p>\n\n<pre><code>#...\nmodel.add(LSTM(8, batch_input_shape=(None, 100, 4), return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(4, return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(1, return_sequences=False, activation='softmax'))\n#...\n<\/code><\/pre>\n\n<p>I have tried to build a network with 4 features input, 2 hidden layers: the first one with 8 neurons, second one with 4 neurons and 1 neuron on the output layer.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GPj1d.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GPj1d.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The activation I wanted was LeakyReLU.<\/p>\n\n<p>Q: <\/p>\n\n<ol>\n<li>Is the implementation correct?<br>\ni.e.: does the code reflects what I planned?<\/li>\n<li>When using LeakyReLU should I add linear activation on the previous layer?<br>\ni.e.: Do I need to add <code>activation='linear'<\/code> to the LSTM layers?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"52650092","CreationDate":"2018-10-04T15:10:53.530","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As for the first question: \"correct\" in what sense? i.e. It depends on the problem you are modeling and therefore more details need to be provided.<\/p>\n\n<p><code>softmax<\/code> is not used as the activation function when the last layer has only one output unit. That's because softmax normalizes the output to make the sum of its elements be one, i.e. to resemble a probability distribution. Therefore, if you use it on a layer with only one output unit it would always have an output of 1. Instead, either <code>linear<\/code> (in case of regression, i.e. predicting real values) or <code>sigmoid<\/code> (in case of binary classification) is used. Additionally, commonly a <code>Dense<\/code> layer is used as the last layer which acts as the final regressor or classifier. For example:<\/p>\n\n<pre><code>model.add(LSTM(8, batch_input_shape=(None, 100, 4), return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(4, return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(1, return_sequences=False))\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>As for the layers and number of units (according to the figure): it is a bit ambiguous, but I think there are three LSTM layers, the first one has 4 units, the second one has 8 units and the last one has 4 units. As for the final layer it seems to be a Dense layer. So the model would look like this (assuming <code>LeakyReLU<\/code> is applied on the output of LSTM layers):<\/p>\n\n<pre><code>model.add(LSTM(4, batch_input_shape=(None, 100, 4), return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(8, return_sequences=True))\nmodel.add(LeakyReLU())\nmodel.add(LSTM(4, return_sequences=False))\nmodel.add(Dense(1, activation='sigmoid')) # or activation='linear' if it is a regression problem\n<\/code><\/pre>\n\n<p>As for using the <code>LeakyReLU<\/code> layer: I guess you are right that <code>linear<\/code> activation should be used as the activation of its previous layer (as also suggested <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/117#issuecomment-100962004\" rel=\"nofollow noreferrer\">here<\/a>, though a<code>Dense<\/code> layer has been used there). That's because by default the activation of LSTM layer is hyperbolic tangent (i.e. <code>tanh<\/code>) and therefore it squashes the outputs to the range [-1,1] which I think may not be efficient when you apply <code>LeakyReLU<\/code> on it; however, I am not sure about this since I am not completely familiar with leaky relu's practical and recommended usage.<\/p>\n"}]}
+{"QuestionId":52647743,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-10-04T13:17:16.013","AcceptedAnswerId":null,"OwnerUserId":6623493.0,"Title":"unable to call keras module in python","Body":"<p>I am working on keras. I am new to kears. Here is my code for creating movie vectors and user vectors. After applying flatten its gives empty tensor.<\/p>\n\n<pre><code>movie_input = keras.layers.Input(shape=[1])\nmovie_vec = keras.layers.Flatten()(keras.layers.Embedding(n_movies + 1, 32) \n(movie_input))\nmovie_vec = keras.layers.Dropout(0.5)(movie_vec)\nprint(movie_vec)\n\n# Same thing for the users\nuser_input = keras.layers.Input(shape=[1])\nuser_vec = keras.layers.Flatten()(keras.layers.Embedding(n_users + 1, 32) \n(user_input))\nuser_vec = keras.layers.Dropout(0.5)(user_vec)\nprint(user_vec)\n\ninput_vecs = keras.layers.merge([movie_vec, user_vec], mode='concat')\nnn = keras.layers.Dropout(0.5)(keras.layers.Dense(128, activation='relu')    (input_vecs))\nnn = keras.layers.normalization.BatchNormalization()(nn)\nnn = keras.layers.Dropout(0.5)(keras.layers.Dense(128, activation='relu')(nn))\nnn = keras.layers.normalization.BatchNormalization()(nn)\nnn = keras.layers.Dense(128, activation='relu')(nn)\n<\/code><\/pre>\n\n<p>It gives me this error<\/p>\n\n<pre><code>&gt;TypeError                                 Traceback (most recent call last)\n&gt;cc&lt;ipython-input-27-10f282af0460&gt; in &lt;module&gt;()\n 18 # Next, we join them all together and put them\n 19 # through a pretty standard deep learning architecture\n 20 input_vecs = keras.layers.merge([movie_vec, user_vec], mode='concat')\n 21 nn = keras.layers.Dropout(0.5)(keras.layers.Dense(128,activation='relu')(input_vecs))\n 22 nn = keras.layers.normalization.BatchNormalization()(nn)\n &gt;TypeError: 'module' object is not callable\n<\/code><\/pre>\n","answers":[{"AnswerId":"52647880","CreationDate":"2018-10-04T13:23:11.900","ParentId":null,"OwnerUserId":"9114815","Title":null,"Body":"<p><code>keras.layers.merge<\/code> is not a callable. It is a module. Check out its <a href=\"https:\/\/keras.io\/layers\/merge\/\" rel=\"nofollow noreferrer\">documentation<\/a><\/p>\n"},{"AnswerId":"52648587","CreationDate":"2018-10-04T13:56:33.183","ParentId":null,"OwnerUserId":"7130689","Title":null,"Body":"<p>Use keras 2.1.5<\/p>\n\n<p>keras.layers.merge method available in it<\/p>\n"}]}
+{"QuestionId":52647884,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-04T13:23:19.253","AcceptedAnswerId":null,"OwnerUserId":10456792.0,"Title":"How to convert pytorch lstmcell to keras lstm or lstmcell","Body":"<p>This is example for pytorch lstmcell:<\/p>\n\n<pre><code>rnn = nn.LSTMCell(10, 20)\ninput = torch.randn(6, 3, 10)\nhx = torch.randn(3, 20)\ncx = torch.randn(3, 20)\noutput = []\nhx, cx = rnn(input[0], (hx, cx))\noutput.append(hx)\n<\/code><\/pre>\n\n<p>Not sure how to convert it to keras lstm\/lstmcell<\/p>\n","answers":[{"AnswerId":"52674255","CreationDate":"2018-10-05T23:01:59.643","ParentId":null,"OwnerUserId":"10456792","Title":null,"Body":"<p>Origial Pytorch code: <br\/>\nself.att_lstm = nn.LSTMCell(1536, 512)<br\/> \nh_att, c_att = self.att_lstm(att_lstm_input, (state[0][0], state[1][0]))<br\/>\nstate[0][0], state[1][0] is the tensor(10,512) \nWhat I am tried in keras:\n<code>inputs = Input(shape=(10, 1536))<\/code><br\/>\n<code>lstm, h_att, c_att = LSTM(units=512, input_shape=(10,1536), name='core.att_lstm', return_state=True)(inputs)<\/code> <br\/>\nso I am not sure if it is correct.<\/p>\n"}]}
+{"QuestionId":52648516,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-04T13:52:59.720","AcceptedAnswerId":53109929.0,"OwnerUserId":6299772.0,"Title":"How to get only the first n layers of a network loaded from a .pb file","Body":"<p><strong>What I want:<\/strong> A protobuf file which contains all the layers of a pretrained AlexNet up until the pool5 layer.<\/p>\n\n<p><strong>What I have:<\/strong> I downloaded the file of the weights of AlexNet <a href=\"https:\/\/zenodo.org\/record\/1208034\" rel=\"nofollow noreferrer\">here<\/a> and \nconverted it to a protobuf file of the model and a frozen protobuf file with <a href=\"https:\/\/github.com\/huanzhang12\/tensorflow-alexnet-model\" rel=\"nofollow noreferrer\">this<\/a> code.\nI loaded the resulting protobuf file with this code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.platform import gfile\nfrom tensorflow.contrib import graph_editor as editor\n\nGRAPH_PB_PATH = 'alexnet.pb'\nwith tf.Session() as sess:\n   print(\"load graph\")\n   with gfile.FastGFile(GRAPH_PB_PATH,'rb') as f:\n       graph_def = tf.GraphDef()\n   graph_def.ParseFromString(f.read())\n   sess.graph.as_default()\n   tf.import_graph_def(graph_def, name='')\n   writer = tf.summary.FileWriter('logs', sess.graph)\n   writer.close()\n   graph_nodes=[n for n in graph_def.node]\n   names = []\n   for t in graph_nodes:\n      names.append(t.name)\n   print(names)\n<\/code><\/pre>\n\n<p>Now I want to throw away all layers subsequent to the <code>pool5<\/code> layer, such that the input of the network is an image and the output is whatever <code>pool5<\/code> return (i.e. some vector). I would like to save the resulting, and now way smaller network into a protobuf file again.\nSo how do I delete the unnecessary layers?\nThanks in advance!<\/p>\n","answers":[{"AnswerId":"53109929","CreationDate":"2018-11-01T21:56:56.960","ParentId":null,"OwnerUserId":"9255615","Title":null,"Body":"<pre><code>graph_def = tf.GraphDef()\nwith open('alexnet.pb', 'rb') as f:\n    graph_def.ParseFromString(f.read())\n\nwith tf.Graph().as_default() as graph:\n    importer.import_graph_def(graph_def, name='')\n\nnew_model = tf.GraphDef()\n\nwith tf.Session(graph=graph) as sess:    \n    for n in sess.graph_def.node:            \n        nn = new_model.node.add()\n        nn.CopyFrom(n)\n        if n.op.name == 'pool5':\n            break;\n\ntf.train.write_graph(new_model, '.', 'cut_model.pb', as_text=False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52649144,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2018-10-04T14:24:28.080","AcceptedAnswerId":null,"OwnerUserId":4454990.0,"Title":"python.exe - Entry Point Not Found while running a code using tensorflow in Jupyter notebook","Body":"<p>I am a beginner in python and tensorflow programming. I am following this code <a href=\"https:\/\/gist.github.com\/simoninithomas\/7611db5d8a6f3edde269e18b97fa4d0c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/simoninithomas\/7611db5d8a6f3edde269e18b97fa4d0c<\/a>\n<a href=\"https:\/\/gist.github.com\/simoninithomas\/7611db5d8a6f3edde269e18b97fa4d0c\" rel=\"nofollow noreferrer\"><\/a>\n<a href=\"https:\/\/gist.github.com\/simoninithomas\/7611db5d8a6f3edde269e18b97fa4d0c\" rel=\"nofollow noreferrer\"><\/a> to use DQN but while running the part <code>In[8]<\/code> I face with a window saying that <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>The procedure entry point\n\n?AllocateAligned@Arenalmpl@internal@protobuf@google@QEAAPEAX_K@Z could not be located in the dynamic link library\n\nC:\\anaconda3\\envs\\python35\\lib\\site-packages\\tensorflow\\contrib\\data\\_dataset_ops.so.\n<\/code><\/pre>\n\n<p>while there is a file named <code>_dataset_ops.so<\/code> in that location.\nI should add that I have windows 10 and I use anaconda and jupyter notebook and cpu version of Tensorflow and no CUDA.\nI will appreciate if anyone can help about the issue.<\/p>\n","answers":[]}
+{"QuestionId":52649188,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-04T14:26:32.977","AcceptedAnswerId":null,"OwnerUserId":9683148.0,"Title":"How to derive weight and bias in a neural network?","Body":"<p>This is the Network:<\/p>\n\n<pre><code>    class Net(torch.nn.Module):\n        def __init__(self, n_feature, n_hidden, n_output):  \n            super(Net, self).__init__()  \n            self.hidden = torch.nn.Linear(n_feature, n_hidden)  \n            self.predict = torch.nn.Linear(n_hidden, n_output)   # output layer  \n        def forward(self, x): \n            x = F.relu(self.hidden(x))      # activation function for hidden layer \n            x = self.predict(x)             # linear output\n            return x\n\n    net = Net(n_feature=1, n_hidden=10, n_output=1)   \n    pytorch_total_params = sum(p.numel() for p in net.parameters())\n    print(pytorch_total_params)\n\n    w = list(net.parameters())\n    print(w)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>This is the runnung result:\n31\n[Parameter containing:\ntensor([[ 0.9534],\n        [-0.0309],\n        [-0.9570],\n        [-0.4179],\n        [-0.3757],\n        [-0.4227],\n        [-0.8866],\n        [ 0.2107],\n        [ 0.0222],\n        [ 0.2531]], requires_grad=True), Parameter containing:\ntensor([-0.0358, -0.2533,  0.2979,  0.9777,  0.9606,  0.9460,  0.9059,  0.7582,\n        -0.5286,  0.3367], requires_grad=True), Parameter containing:\ntensor([[-0.2863, -0.3157,  0.2086, -0.0011, -0.0415, -0.2574, -0.0683, -0.0788,\n         -0.0339, -0.0195]], requires_grad=True), Parameter containing:\ntensor([0.2031], requires_grad=True)]\n<\/code><\/pre>\n\n<p>I don't know why is the number of parameters be 31?\nAnd also don't understand the numbers printed above.(whether it is weight or bias<\/p>\n\n<p>Because I thought in Relu function, there will only be (2 parameters*10) which is weight and bias multiply 10 hidden layers.<\/p>\n","answers":[{"AnswerId":"52664868","CreationDate":"2018-10-05T11:40:52.600","ParentId":null,"OwnerUserId":"7752110","Title":null,"Body":"<p>If you print the named parameters you can see to which layer a parameter belongs.\nPrinting the named parameters:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for p in net.named_parameters():\n    print(p)\n<\/code><\/pre>\n\n<p>Creates the following output:<\/p>\n\n<pre><code>('hidden.weight', Parameter containing:\ntensor([[ 0.8324],\n        [ 0.2166],\n        [-0.9786],\n        [ 0.3977],\n        [ 0.9008],\n        [-0.3102],\n        [ 0.5052],\n        [ 0.6589],\n        [ 0.0828],\n        [ 0.6505]], requires_grad=True))\n('hidden.bias', Parameter containing:\ntensor([ 0.6715,  0.5503, -0.6043,  0.1102, -0.2700,  0.7203, -0.6524, -0.6332,\n        -0.2513, -0.1316], requires_grad=True))\n('predict.weight', Parameter containing:\ntensor([[ 0.1486,  0.1528, -0.0835, -0.3050,  0.1184, -0.0422, -0.2786, -0.2549,\n         -0.1532, -0.0255]], requires_grad=True))\n('predict.bias', Parameter containing:\ntensor([0.2878], requires_grad=True))\n<\/code><\/pre>\n\n<p>As you can see the layers are connected by 10 weights each, as you expected, but there is one bias per neuron on the right side of a 'connection'.\nSo you have 10 bias-parameters between your input and your hidden layer and just one for the calculation of your final prediction.<\/p>\n\n<p>You are calculating the input to each neuron in the l-th layer like this weighted sum:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iE1OD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iE1OD.png\" alt=\"weighted sum\"><\/a><\/p>\n\n<p>So you need a weight for every connection between the neurons of the two layers, but only one bias per neuron in the l-th layer.<\/p>\n\n<p>In your case:<\/p>\n\n<ul>\n<li>input to hidden: 10 weights and 10 bias, because your hidden layer has 10 neurons<\/li>\n<li>hidden to output\/predict: 10 weights and 1 bias, because you output a single value<\/li>\n<\/ul>\n\n<p>sums up to 31 parameters.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PO7yy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PO7yy.png\" alt=\"neural network\"><\/a><\/p>\n"}]}
+{"QuestionId":52649760,"AnswerCount":0,"Tags":"<tensorflow><deconvolution>","CreationDate":"2018-10-04T14:54:11.890","AcceptedAnswerId":null,"OwnerUserId":7104060.0,"Title":"How Tensorflow Handles 'SAME' padding in transpose convolutional?","Body":"<p>I would like to know how 'SAME' padding is calculated when using a transpose convolutional layer in Tensorflow. I find the way it is calculated in this <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/nn#convolution\" rel=\"nofollow noreferrer\">link<\/a> in a convolutional layer with this formula in this link, but I wonder how it is done in the deconvolutional layer. Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52649850,"AnswerCount":0,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2018-10-04T14:58:49.730","AcceptedAnswerId":null,"OwnerUserId":3285421.0,"Title":"How to pass weights and mask to weighted_masked_objective in keras?","Body":"<p>I am wondering how can I compute a pixel-wise weighted loss for image segmentation. For each image I have a ground-truth mask and a weighted mask to give more importance to certain pixels in an image. Note that, my weighted mask is unique for each image so I can not pass this mask in a user defined loss function as an input. I can see a function <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/58fd1f0589d33aeb33c4129cfedfb7737495efc0\/keras\/engine\/training_utils.py#L352\" rel=\"nofollow noreferrer\">\"weighted_masked_objective\"<\/a> in keras\/engine\/training_utils.py which can handle a mask but I din not find any way to pass weighted mask to that function.<\/p>\n\n<p>Is there any way to pass weighted mask to a loss function as part of standard input parameters e,g. y_true, y_pred? <\/p>\n","answers":[]}
+{"QuestionId":52650842,"AnswerCount":2,"Tags":"<tensorflow><keras><object-detection>","CreationDate":"2018-10-04T15:48:51.653","AcceptedAnswerId":null,"OwnerUserId":2243184.0,"Title":"How to convert hdf5 to tensorflow checkpoint","Body":"<p>I am working on Mask-rcnn (<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN<\/a>) and want to convert the Keras checkpoints in h5 format to Tensorflow checkpoints. How to do that?<\/p>\n","answers":[{"AnswerId":"58237028","CreationDate":"2019-10-04T13:04:11.603","ParentId":null,"OwnerUserId":"7867536","Title":null,"Body":"<p>This code worked for me in <code>tensorflow 2.x<\/code><\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.models import load_model\n\nsaver = tf.train.Checkpoint()\nmodel = load_model('model.hdf5', compile=False)\nsess = tf.compat.v1.keras.backend.get_session()\n\nsave_path = saver.save(\"model.ckpt\")\n<\/code><\/pre>\n"},{"AnswerId":"56535529","CreationDate":"2019-06-11T02:05:50.717","ParentId":null,"OwnerUserId":"2699907","Title":null,"Body":"<p>You can try to load the keras model, then use the tensorflow session to save it as usual:<\/p>\n\n<pre><code>model = keras.models.load_model(\"mask_rcnn_coco.h5\")\nsession = keras.backend.get_session()\n\nsaver = tf.train.Saver()\nsaver.save(session, \"checkpoints\/mask_rcnn.ckpt\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":52651749,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2018-10-04T16:42:19.433","AcceptedAnswerId":null,"OwnerUserId":9137396.0,"Title":"How to print summary of only last n layers of a model in keras","Body":"<p><code>model.summary()<\/code> prints details of the entire model. Is there a way to just print the last n layer(s) summary only?<\/p>\n\n<p>If not, can I create a new model from the last n layers of an existing pre-trained model and print its summary instead.<\/p>\n\n<p>I tried the following but it gives an error probably because of shared inputs:<\/p>\n\n<pre><code>temp_model = Model(inputs=base_model.layers[-4].input, outputs = base_model.layers[-1].output)\nprint(temp_model.summary())\n<\/code><\/pre>\n\n<p>Any help will be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52652214,"AnswerCount":2,"Tags":"<google-cloud-platform><pytorch><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-04T17:15:03.947","AcceptedAnswerId":58545925.0,"OwnerUserId":4106826.0,"Title":"Using TPUs with PyTorch","Body":"<p>I am trying to use Google Cloud's TPU from Colab. I was able to do it following the tutorial by using Tensorflow.<\/p>\n\n<p>Does anybody know if it is possible to make use of the TPUs using PyTorch?\nIf so how can I do it? Do you have any example?<\/p>\n","answers":[{"AnswerId":"52673106","CreationDate":"2018-10-05T20:48:02.787","ParentId":null,"OwnerUserId":"5091538","Title":null,"Body":"<p>Currently, it's not possible to use Cloud TPU with PyTorch since it's designed specifically for Tensorflow. <\/p>\n\n<p>But, according to <a href=\"https:\/\/cloud.google.com\/blog\/products\/ai-machine-learning\/introducing-pytorch-across-google-cloud\" rel=\"noreferrer\">this product news posted three days ago in the Google Cloud blog<\/a>, \"engineers on Google\u2019s TPU team are actively collaborating with core PyTorch developers to connect PyTorch to Cloud TPUs\".<\/p>\n"},{"AnswerId":"58545925","CreationDate":"2019-10-24T16:56:02.057","ParentId":null,"OwnerUserId":"10199571","Title":null,"Body":"<p>Check out our repository <a href=\"https:\/\/github.com\/pytorch\/xla\" rel=\"nofollow noreferrer\">pytorch\/xla<\/a> where you can start training PyTorch models on TPUs. <\/p>\n\n<p>Also, you can even use free TPUs on Colab with PyTorch with these <a href=\"https:\/\/github.com\/pytorch\/xla\/tree\/master\/contrib\/colab\" rel=\"nofollow noreferrer\">Colab notebooks<\/a>.<\/p>\n"}]}
+{"QuestionId":52652479,"AnswerCount":0,"Tags":"<c++><tensorflow><gpu>","CreationDate":"2018-10-04T17:33:53.800","AcceptedAnswerId":null,"OwnerUserId":1682075.0,"Title":"How to reuse tensorflow after cudaDeviceReset() in C++?","Body":"<p>I am working on a large cuda app in C++ that runs various models and needs to completely release all GPU memory or the other operations will fail.<\/p>\n\n<p>I am able to release all the memory after closing all tf sessions and running cudaDeviceReset(). But afterwards I cannot run any new tensorflow code and session creation will return nullptrs. I tried cudaDeviceSynchronize() before and after thinking that would help but no luck.<\/p>\n\n<p>I figured the call to InitMain would re-initialize tensorflow but it seems not to if I call again after the reset. Is there a specific entry-point I can call to basically \"start over\" and have tensorflow once again able to use the GPU and produce Sessions?<\/p>\n\n<p>Using C++, not python. Below is an example showing what tensorflow shows after I run the cudeDeviceReset() and then attempt to open a new Tensorflow Session in C++.<\/p>\n\n<pre><code>2018-10-04 17:01:19.225505: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1435] Adding visible gpu devices: 0\n2018-10-04 17:01:19.326074: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-04 17:01:19.326091: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:929]      0 \n2018-10-04 17:01:19.326095: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:942] 0:   N \n2018-10-04 17:01:19.326215: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1053] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 9446 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)\n2018-10-04 17:01:19.326554: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:785] failed to memset memory: CUDA_ERROR_INVALID_VALUE\n2018-10-04 17:01:19.326593: E tensorflow\/core\/common_runtime\/direct_session.cc:154] Failed precondition: Failed to memcopy into scratch buffer for device 0\n<\/code><\/pre>\n\n<p>May be related, but the first run has extra lines at the beginning as if a one-time initialization perhaps has ran.... The second run lacks those lines.. Below is what the first looks like.<\/p>\n\n<pre><code>2018-10-04 17:01:17.253809: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-04 17:01:17.254173: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1356] Found device 0 with properties: \nname: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6325\npciBusID: 0000:01:00.0\ntotalMemory: 10.92GiB freeMemory: 9.76GiB\n2018-10-04 17:01:17.254185: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1435] Adding visible gpu devices: 0\n2018-10-04 17:01:17.413712: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-04 17:01:17.413733: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:929]      0 \n2018-10-04 17:01:17.413737: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:942] 0:   N \n2018-10-04 17:01:17.413888: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1053] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 9446 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)\n<\/code><\/pre>\n\n<p>I have at least figured out how to release the memory properly and nvidia-smi shows my application drop the memory as expected after the reset. But it's still not useful if I can't do anything afterwards... Appreciate any help!<\/p>\n\n<p>EDIT: Had a question about which code I'm using to establish the session. I am not using any custom session options...<\/p>\n\n<pre><code>tensorflow::NewSession(tensorflow::SessionOptions());\n<\/code><\/pre>\n\n<p>A basic reproduction is to simply load a graph into a session, close the session and graph, free the pointers, call cudaDeviceReset() and again try to open another session which will throw nullptrs and the above error messages.<\/p>\n","answers":[]}
+{"QuestionId":52653060,"AnswerCount":1,"Tags":"<machine-learning><keras><conv-neural-network><flow><tensor>","CreationDate":"2018-10-04T18:14:02.913","AcceptedAnswerId":null,"OwnerUserId":6703017.0,"Title":"InvalidArgumentError: input must be 4-dimensional[8,6171,4]","Body":"<p>I'm running a 2d convolution network. My input has 3 dimension however I am getting this 4-dimension error: \n<a href=\"https:\/\/i.stack.imgur.com\/PiyuF.png\" rel=\"nofollow noreferrer\">dimension error<\/a><\/p>\n\n<p>As you can see my input has the correct dimensions:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hIgUR.png\" rel=\"nofollow noreferrer\">correct input dimension<\/a><\/p>\n\n<p>here is my code:<\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(8,6171,4)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu'))\n\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(64, activation='relu'))\nmodel.add(layers.Dense(10, activation='softmax'))\n<\/code><\/pre>\n\n<p>Why then is it asking me for 4-D when I only built a 3-D input layer?<\/p>\n\n<p>Please help.\nThanks.<\/p>\n","answers":[{"AnswerId":"55884186","CreationDate":"2019-04-27T19:34:10.103","ParentId":null,"OwnerUserId":"3633891","Title":null,"Body":"<p>You need to use <\/p>\n\n<pre><code>new_image = tf.expand_dims(image,0)\n<\/code><\/pre>\n\n<p>because the model expects a dataset instead of single images.<\/p>\n"}]}
+{"QuestionId":52653320,"AnswerCount":1,"Tags":"<python><machine-learning><keras><probability>","CreationDate":"2018-10-04T18:48:04.223","AcceptedAnswerId":52653638.0,"OwnerUserId":2130515.0,"Title":"predict_proba() method of Keras model does not exist","Body":"<p>I am trying to generate class scores by calling <code>predict_proba()<\/code> of Keras model, but it seems that this function does not exist! Is it deprecated because I see some examples in Google? I am using Keras 2.2.2.<\/p>\n","answers":[{"AnswerId":"52653638","CreationDate":"2018-10-04T19:12:47.840","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The <code>predict_proba()<\/code> and <code>predict_classes()<\/code> methods <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2524#issuecomment-218590463\" rel=\"nofollow noreferrer\">are not well-defined for models created using functional API<\/a> (i.e. <code>keras.models.Model()<\/code>). That's because the models created using functional API may have multiple output layers each with different configurations. Therefore predicting probabilities in this case is not meaningful, even if your model outputs probabilities. The method you referred to, as well as <code>predict_classes()<\/code>, is only defined for Sequential models (i.e. <code>keras.models.Sequential()<\/code>).<\/p>\n"}]}
+{"QuestionId":52653497,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T19:02:52.453","AcceptedAnswerId":52828599.0,"OwnerUserId":2985049.0,"Title":"Use tf.Keras with Tensorflow optimizer","Body":"<p>I'm trying to use tf.keras with the new AdamW optimizer in tensorflow and am running into issues. A toy version of the code is as follows:<\/p>\n\n<pre><code>from tensorflow.contrib.opt import AdamWOptimizer\nfrom tensorflow.python.keras.optimizers import TFOptimizer\n\nmodel = Sequential()\nmodel.add(Dense(2, activation=\"tanh\", input_shape=(3,)))\n\ntfopt = AdamWOptimizer(weight_decay=0.1, learning_rate=.004)\noptimizer = TFOptimizer(tfopt)\n\nmodel.compile(optimizer=optimizer, loss='mean_squared_error')\nmodel.fit(np.random.random((5, 3)),\n          np.random.random((5, 2)),\n          epochs=5, batch_size=5)\n<\/code><\/pre>\n\n<p>Error is as follows:<\/p>\n\n<pre><code>..\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py:1605: in fit\n    validation_steps=validation_steps)\n..\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py:153: in fit_loop\n    outs = f(ins)\n..\/python3.6\/site-packages\/tensorflow\/python\/keras\/backend.py:2978: in __call__\n    run_metadata=self.run_metadata)\n..\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py:1399: in __call__\n    run_metadata_ptr)\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \n\nself = &lt;tensorflow.python.framework.errors_impl.raise_exception_on_not_ok_status object at 0x11ecde550&gt;\ntype_arg = None, value_arg = None, traceback_arg = None\n\n    def __exit__(self, type_arg, value_arg, traceback_arg):\n      try:\n        if c_api.TF_GetCode(self.status.status) != 0:\n          raise _make_specific_exception(\n              None, None,\n              compat.as_text(c_api.TF_Message(self.status.status)),\n&gt;             c_api.TF_GetCode(self.status.status))\nE             tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value training\/TFOptimizer\/beta2_power\nE                [[{{node training\/TFOptimizer\/beta2_power\/read}} = Identity[T=DT_FLOAT, _class=[\"loc:@training\/TFOptimizer\/AdamW\/Assign\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](training\/TFOptimizer\/beta2_power)]]\n\n..\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py:526: FailedPreconditionError\n<\/code><\/pre>\n","answers":[{"AnswerId":"52828599","CreationDate":"2018-10-16T05:40:31.950","ParentId":null,"OwnerUserId":"2985049","Title":null,"Body":"<p>Turns out <code>TFOptimizer<\/code> won't work in python 3.6 but it does work in 2.7.<\/p>\n\n<p>However, you don't actually need to use TFOptimizer. Plugging the tensorflow optimizer <code>AdamWOptimizer<\/code> directly into the optimizer argument of fit runs perfectly.<\/p>\n"},{"AnswerId":"58391473","CreationDate":"2019-10-15T09:28:56.627","ParentId":null,"OwnerUserId":"3293408","Title":null,"Body":"<p>For me, this runs, and converges very fast to high accuracy for most of my classification goals, but there is one problem: the margins are very small, with all scores being close to 0.5. In addition, one classifier does not converge. With Adam, the results are better.<\/p>\n"}]}
+{"QuestionId":52653656,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2018-10-04T19:14:18.670","AcceptedAnswerId":null,"OwnerUserId":1738210.0,"Title":"How Can I Plot a TensorBoard Diagram?","Body":"<p>I have some diagrams using TensorBoard, and I want to plot them and then save them as images. How can I do this? Is it possible? <\/p>\n\n<p>I tried:<\/p>\n\n<pre><code>data = np.genfromtxt(\"C:\\Users\\maedeh\\Downloads\\val_model_31_mean_absolute_error.csv\",\n                 delimiter=',', skip_header=10,\n                 skip_footer=10, names=['x', 'y', 'z'])\n<\/code><\/pre>\n\n<p>but it produced this error: <\/p>\n\n<pre><code>SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in position 2-3: truncated \\UXXXXXXXX escape\n<\/code><\/pre>\n\n<p>Can you please help?<\/p>\n","answers":[]}
+{"QuestionId":52654032,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-04T19:40:46.507","AcceptedAnswerId":null,"OwnerUserId":9997976.0,"Title":"Trying to wrap a py function with tf.py_func but have an error","Body":"<p>I am trying to use tf.py_func to wrap a python function. But it gives me error. My code is:<\/p>\n\n<pre><code>def label_energy(energy):\n    return scipy.ndimage.measurements.label(energy)\nlabels_gt, _ = tf.py_func(label_energy, [gt_single_energy], tf.float32)\n<\/code><\/pre>\n\n<p>And I have error:<\/p>\n\n<pre><code>\"Tensor objects are not iterable when eager execution is not \"\nTypeError: Tensor objects are not iterable when eager execution is not \nenabled. To iterate over this tensor use tf.map_fn.\n<\/code><\/pre>\n\n<p>I can wrap other functions, for example <\/p>\n\n<pre><code>def gaussianblur(energy, tu, x):\n    return cv2.GaussianBlur(energy, tu, x)\n<\/code><\/pre>\n\n<p>But not this <code>label_energy()<\/code><\/p>\n","answers":[]}
+{"QuestionId":52654438,"AnswerCount":0,"Tags":"<tensorflow><pyspark>","CreationDate":"2018-10-04T20:10:59.660","AcceptedAnswerId":null,"OwnerUserId":9020032.0,"Title":"PySpark, Tensorflow and Tensorframes - ClassNotFoundException","Body":"<p>I tried to run a minimalistic code from the <a href=\"https:\/\/github.com\/databricks\/tensorframes\" rel=\"nofollow noreferrer\">sample repository<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorframes as tfs\nfrom pyspark.sql import Row\nfrom pyspark.sql import SparkSession\n\nspark = SparkSession\\\n    .builder\\\n    .appName(\"example-spark\")\\\n    .master(\"local[*]\")\\\n    .getOrCreate()\n\ndata = [Row(x=float(x)) for x in range(10)]\ndf = spark.createDataFrame(data)\nwith tf.Graph().as_default() as g:\n    x = tfs.block(df, \"x\")\n    z = tf.add(x, 3, name='z')\n    df2 = tfs.map_blocks(z, df)\n\ndf2.collect()\n<\/code><\/pre>\n\n<p>Unfortunately, I came across the error (I think there may be something wrong with the configuration). I tried different options (mainly setting paths and attaching files), but I still can not deal with it.<\/p>\n\n<pre><code>   Traceback (most recent call last):\n  File \"spark-ml.py\", line 80, in &lt;module&gt;\n    x = tfs.block(df, \"x\")\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/tensorframes\/core.py\", line 413, in block\n    return _auto_placeholder(df, col_name, tf_name, block = True)\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/tensorframes\/core.py\", line 431, in _auto_placeholder\n    info = _java_api().extra_schema_info(df._jdf)\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/tensorframes\/core.py\", line 34, in _java_api\n    return _jvm.Thread.currentThread().getContextClassLoader().loadClass(javaClassName) \\\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/py4j\/java_gateway.py\", line 1257, in __call__\n    answer, self.gateway_client, self.target_id, self.name)\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/pyspark\/sql\/utils.py\", line 63, in deco\n    return f(*a, **kw)\n  File \"\/home\/btw\/.local\/lib\/python3.5\/site-packages\/py4j\/protocol.py\", line 328, in get_return_value\n    format(target_id, \".\", name), value)\npy4j.protocol.Py4JJavaError: An error occurred while calling o55.loadClass.\n: java.lang.ClassNotFoundException: org.tensorframes.impl.DebugRowOps\n        at java.net.URLClassLoader.findClass(URLClassLoader.java:381)\n        at java.lang.ClassLoader.loadClass(ClassLoader.java:424)\n        at java.lang.ClassLoader.loadClass(ClassLoader.java:357)\n        at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)\n        at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)\n        at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)\n        at java.lang.reflect.Method.invoke(Method.java:498)\n        at py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:244)\n        at py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:357)\n        at py4j.Gateway.invoke(Gateway.java:282)\n        at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132)\n        at py4j.commands.CallCommand.execute(CallCommand.java:79)\n        at py4j.GatewayConnection.run(GatewayConnection.java:238)\n        at java.lang.Thread.run(Thread.java:748)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52654520,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-10-04T20:17:05.420","AcceptedAnswerId":52680874.0,"OwnerUserId":7368263.0,"Title":"Adding a layer to a trained Tensorflow estimator","Body":"<p>I have a Tensorflow model that was trained using the Estimator API. Now I want to load this model, add a new layer to it, and train only the new layer (i.e. freezing all other parameters). <\/p>\n\n<p>What would be the best way to do that?<\/p>\n\n<p>I managed to load the model from a checkpoint and add the new layer, but it's not clear to me how to feed examples as input and perform training. Specifically, I couldn't locate a placeholder for the input.<\/p>\n","answers":[{"AnswerId":"52680874","CreationDate":"2018-10-06T15:59:29.657","ParentId":null,"OwnerUserId":"7368263","Title":null,"Body":"<p>I have found a simple way to do that, and posting it here in case someone will find it useful:<\/p>\n\n<ol>\n<li>Load the Estimator checkpoint.<\/li>\n<li>Create a placeholder and a copy of the model graph, under a new name scope.<\/li>\n<li>Extract all trainable variables under the two scopes.<\/li>\n<li>Create <code>assign<\/code> ops for every variable.<\/li>\n<\/ol>\n\n<hr>\n\n<p>Code:<\/p>\n\n<pre><code>    # Load the trained model from checkpoint.\n    new_saver = tf.train.import_meta_graph('{}.meta'.format(config.ckpt_fullpath))\n    new_saver.restore(sess, config.ckpt_fullpath)\n\n    # Create new graph with a placeholder for input.\n    new_model_scope = 'new_scope'\n    trained_model_scope = 'old_scope' # this should be taken from the original model function of the estimator.\n    with tf.name_scope(new_model_scope):\n        model = Model(config)\n        input_tensor = tf.placeholder(tf.float32, \n                                      [None, config.img_size[0], config.img_size[1], 3])\n        model.build_model(input_tensor)\n\n    # Initialize the new graph variables with trained parameters.\n    trained_params = [t for t in tf.trainable_variables()\n                      if t.name.startswith(trained_model_scope)]\n    trained_params = sorted(trained_params, key=lambda v: v.name)\n    new_params = [t for t in tf.trainable_variables() \n                  if t.name.startswith(new_model_scope)]\n    new_params = sorted(new_params, key=lambda v: v.name)\n\n    update_ops = []\n    for trained_v, new_v in zip(trained_params, new_params):\n        op = new_v.assign(trained_v)\n        update_ops.append(op)\n    sess.run(update_ops)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52654587,"AnswerCount":2,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2018-10-04T20:21:26.043","AcceptedAnswerId":null,"OwnerUserId":10027028.0,"Title":"Type error: must be real number not Tensor","Body":"<p>my model to code a convolutional neural network is as follow:\nBelow you will find the code when I use the input examples as a whole.\nThereunder code for minibatches that is working. I do not see the difference!<\/p>\n\n<pre><code>    # model\n\n    def model(train_x, train_y, test_x, test_y, learning_rate = 0.009,\n              num_iterations = 100, print_cost = True):\n        \"\"\"\n        Implements a three-layer ConvNet in Tensorflow:\n        CONV2D -&gt; RELU -&gt; MAXPOOL -&gt; CONV2D -&gt; RELU -&gt; MAXPOOL -&gt; FLATTEN -&gt; FULLYCONNECTED\n\n        Arguments:\n        X_train -- training set, of shape (None, 64, 64, 3)\n        Y_train -- test set, of shape (None, n_y = 6)\n        X_test -- training set, of shape (None, 64, 64, 3)\n        Y_test -- test set, of shape (None, n_y = 6)\n        learning_rate -- learning rate of the optimization\n        num_epochs -- number of epochs of the optimization loop\n        minibatch_size -- size of a minibatch\n        print_cost -- True to print the cost every 100 epochs\n\n        Returns:\n        train_accuracy -- real number, accuracy on the train set (X_train)\n        test_accuracy -- real number, testing accuracy on the test set (X_test)\n        parameters -- parameters learnt by the model. They can then be used to predict.\n        \"\"\"\n\n        ops.reset_default_graph()                         # to be able to rerun the model without overwriting tf variables\n        tf.set_random_seed(1)                             # to keep results consistent (tensorflow seed)\n        seed = 3                                          # to keep results consistent (numpy seed)\n        (m, n_H0, n_W0, n_C0) = train_x.shape             \n        n_y = train_y.shape[1]                            \n        costs = []                                        # To keep track of the cost\n\n        # Create Placeholders of the correct shape\n        ### START CODE HERE ### (1 line)\n        X, Y = create_placeholders(n_H0, n_W0, n_C0, n_y)\n        ### END CODE HERE ###\n\n        # Initialize parameters\n        ### START CODE HERE ### (1 line)\n        parameters = initialize_parameters()\n        ### END CODE HERE ###\n        for i in range(0, num_iterations):\n            # Forward propagation: Build the forward propagation in the tensorflow graph\n            ### START CODE HERE ### (1 line)\n            Z3 = forward_propagation(X, parameters)\n            ### END CODE HERE ###\n\n            # Cost function: Add cost function to tensorflow graph\n            ### START CODE HERE ### (1 line)\n            cost = compute_cost(Z3, Y)\n            ### END CODE HERE ###\n\n            # Backpropagation: Define the tensorflow optimizer. Use an AdamOptimizer that minimizes the cost.\n            ### START CODE HERE ### (1 line)\n            optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)\n            ### END CODE HERE ###\n\n            # Initialize all the variables globally\n            init = tf.global_variables_initializer()\n\n            # Start the session to compute the tensorflow graph\n            with tf.Session() as sess:\n\n                # Run the initialization\n                sess.run(init)\n\n                # Do the training loop\n\n                _ , temp_cost = sess.run([optimizer, cost], feed_dict = {X: train_x, Y: train_y})\n\n                ### END CODE HERE ###\n\n            # Print the cost every 5 itterrations\n            print(\"cost =\" +str(cost))\n            if print_cost == True and i % 5 == 0:\n                print (\"Cost after num_iterations %i: %f\" % (i, cost))\n            if print_cost == True and i % 1 == 0:\n                costs.append(cost)\n\n\n            # plot the cost\n            plt.plot(np.squeeze(costs))\n            plt.ylabel('cost')\n            plt.xlabel('iterations (per tens)')\n            plt.title(\"Learning rate =\" + str(learning_rate))\n            plt.show()\n\n            # Calculate the correct predictions\n            predict_op = tf.argmax(Z3, 1)\n            correct_prediction = tf.equal(predict_op, tf.argmax(Y, 1))\n\n            # Calculate accuracy on the test set\n            accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n            print(accuracy)\n            train_accuracy = accuracy.eval({X: train_x, Y: train_y})\n            test_accuracy = accuracy.eval({X: test_x, Y: test_y})\n            print(\"Train Accuracy:\", train_accuracy)\n            print(\"Test Accuracy:\", test_accuracy)\n\n            return train_accuracy, test_accuracy, parameters\n\nHowever when I call the function model via\n\n    _, _, parameters = model(train_x, train_y, test_x, test_y)\n\nI run into following error:\n\n    cost =Tensor(\"Mean:0\", shape=(), dtype=float32)\n\n    ---------------------------------------------------------------------------\n    TypeError                                 Traceback (most recent call last)\n    &lt;ipython-input-78-8bde5fad21ba&gt; in &lt;module&gt;()\n    ----&gt; 1 _, _, parameters = model(train_x, train_y, test_x, test_y)\n\n    &lt;ipython-input-77-68bdfb51274c&gt; in model(train_x, train_y, test_x, test_y, learning_rate, num_iterations, print_cost)\n         73         print(\"cost =\" +str(cost))\n         74         if print_cost == True and i % 5 == 0:\n    ---&gt; 75             print (\"Cost after num_iterations %i: %f\" % (i, cost))\n         76         if print_cost == True and i % 1 == 0:\n         77             costs.append(cost)\n\n    TypeError: must be real number, not Tensor\n\nCan somebody explain me how to resolve this error? \n<\/code><\/pre>\n\n<p><strong>My code with minibatches that is working:<\/strong>\n<strong>Added this code so you can see this is working. I don't see the difference with code above and why code above is not working. Maybe you can help me out. Tnx<\/strong><\/p>\n\n<pre><code># GRADED FUNCTION: model\n\ndef model(train_x, train_y, test_x, test_y, learning_rate = 0.009,\n          num_epochs = 50, minibatch_size = 64, print_cost = True):\n    \"\"\"\n    Implements a three-layer ConvNet in Tensorflow:\n    CONV2D -&gt; RELU -&gt; MAXPOOL -&gt; CONV2D -&gt; RELU -&gt; MAXPOOL -&gt; FLATTEN -&gt; FULLYCONNECTED\n\n    Arguments:\n    X_train -- training set, of shape (None, 64, 64, 3)\n    Y_train -- test set, of shape (None, n_y = 6)\n    X_test -- training set, of shape (None, 64, 64, 3)\n    Y_test -- test set, of shape (None, n_y = 6)\n    learning_rate -- learning rate of the optimization\n    num_epochs -- number of epochs of the optimization loop\n    minibatch_size -- size of a minibatch\n    print_cost -- True to print the cost every 100 epochs\n\n    Returns:\n    train_accuracy -- real number, accuracy on the train set (X_train)\n    test_accuracy -- real number, testing accuracy on the test set (X_test)\n    parameters -- parameters learnt by the model. They can then be used to predict.\n    \"\"\"\n\n    ops.reset_default_graph()                         # to be able to rerun the model without overwriting tf variables\n    tf.set_random_seed(1)                             # to keep results consistent (tensorflow seed)\n    seed = 3                                          # to keep results consistent (numpy seed)\n    (m, n_H0, n_W0, n_C0) = train_x.shape             \n    n_y = train_y.shape[1]                            \n    costs = []                                        # To keep track of the cost\n\n    # Create Placeholders of the correct shape\n    ### START CODE HERE ### (1 line)\n    X, Y = create_placeholders(n_H0, n_W0, n_C0, n_y)\n    ### END CODE HERE ###\n\n    # Initialize parameters\n    ### START CODE HERE ### (1 line)\n    parameters = initialize_parameters()\n    ### END CODE HERE ###\n\n    # Forward propagation: Build the forward propagation in the tensorflow graph\n    ### START CODE HERE ### (1 line)\n    Z3 = forward_propagation(X, parameters)\n    ### END CODE HERE ###\n\n    # Cost function: Add cost function to tensorflow graph\n    ### START CODE HERE ### (1 line)\n    cost = compute_cost(Z3, Y)\n    ### END CODE HERE ###\n\n    # Backpropagation: Define the tensorflow optimizer. Use an AdamOptimizer that minimizes the cost.\n    ### START CODE HERE ### (1 line)\n    optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)\n    ### END CODE HERE ###\n\n    # Initialize all the variables globally\n    init = tf.global_variables_initializer()\n\n    # Start the session to compute the tensorflow graph\n    with tf.Session() as sess:\n\n        # Run the initialization\n        sess.run(init)\n\n        # Do the training loop\n        for epoch in range(num_epochs):\n            #print(\"m = \" + str(m))\n            minibatch_cost = 0.\n            num_minibatches = int(m \/ minibatch_size) # number of minibatches of size minibatch_size in the train set\n            #print(\" minibatch_size =\" + str(minibatch_size))\n            seed = seed + 1\n            minibatches = random_mini_batches(train_x, train_y, minibatch_size, seed)\n            #print(\"minibatch =\" + str(minibatch))\n            #print(\"minibatches = \" + str(minibatches))\n\n            for minibatch in minibatches:\n\n                # Select a minibatch\n                (minibatch_X, minibatch_Y) = minibatch\n                # IMPORTANT: The line that runs the graph on a minibatch.\n                # Run the session to execute the optimizer and the cost, the feedict should contain a minibatch for (X,Y).\n                ### START CODE HERE ### (1 line)\n                _ , temp_cost = sess.run([optimizer, cost], feed_dict = {X: minibatch_X, Y: minibatch_Y})\n                ### END CODE HERE ###\n\n                minibatch_cost += temp_cost \/ num_minibatches\n\n            # Print the cost every 5 epoch\n            if print_cost == True and epoch % 5 == 0:\n                print (\"Cost after epoch %i: %f\" % (epoch, minibatch_cost))\n            if print_cost == True and epoch % 1 == 0:\n                costs.append(minibatch_cost)\n\n\n        # plot the cost\n        plt.plot(np.squeeze(costs))\n        plt.ylabel('cost')\n        plt.xlabel('iterations (per tens)')\n        plt.title(\"Learning rate =\" + str(learning_rate))\n        plt.show()\n\n        # Calculate the correct predictions\n        predict_op = tf.argmax(Z3, 1)\n        correct_prediction = tf.equal(predict_op, tf.argmax(Y, 1))\n\n        # Calculate accuracy on the test set\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n        print(accuracy)\n        train_accuracy = accuracy.eval({X: train_x, Y: train_y})\n        test_accuracy = accuracy.eval({X: test_x, Y: test_y})\n        print(\"Train Accuracy:\", train_accuracy)\n        print(\"Test Accuracy:\", test_accuracy)\n\n        return train_accuracy, test_accuracy, parameters\n<\/code><\/pre>\n\n<p><strong>And below the output:<\/strong><\/p>\n\n<pre><code>_, _, parameters = model(X_train, Y_train, X_test, Y_test)\n\nCost after epoch 0: 1.917929\nCost after epoch 5: 1.506757\nCost after epoch 10: 0.955359\nCost after epoch 15: 0.845802\nCost after epoch 20: 0.701174\nCost after epoch 25: 0.571977\nCost after epoch 30: 0.518435\nCost after epoch 35: 0.495806\nCost after epoch 40: 0.429827\nCost after epoch 45: 0.407291\n<\/code><\/pre>\n","answers":[{"AnswerId":"52733847","CreationDate":"2018-10-10T06:22:54.207","ParentId":null,"OwnerUserId":"10027028","Title":null,"Body":"<p>problem is solved.\nput the itteration <code>for i in range(0, num_iterations):<\/code><\/p>\n\n<p>before the line <code>_ , temp_cost = sess.run([optimizer, cost], feed_dict = {X: train_x, Y: train_y})<\/code><\/p>\n\n<p>In stead of before my forward propagation step!<\/p>\n"},{"AnswerId":"52654836","CreationDate":"2018-10-04T20:41:24.720","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>Already line 73 is broken. To print a tensor you need to use <code>tf.Print<\/code>,\nbut this requires you to add it to your computational graph:<\/p>\n\n<pre><code>print_node = tf.Print(opt.get_slot(var,'m'), [opt.get_slot(var,'m')], 'm')\nsess.run([print_node], ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52655087,"AnswerCount":1,"Tags":"<python><tensorflow><installation>","CreationDate":"2018-10-04T21:01:45.113","AcceptedAnswerId":null,"OwnerUserId":8634746.0,"Title":"how to install tensorflow on python3.7","Body":"<p>I tried instruction like this but failed:<\/p>\n\n<pre><code>pip3 install -t \/projects\/****\/users\/****\/python\/3.7.0 --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.10.0-cp37-cp37m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>the error message is just cannot find the url. If I change version like cp37 to cp36 which will also failed since the version of python is 3.7. <\/p>\n\n<p>If I just use the instruction:<\/p>\n\n<pre><code>pip3 install -t \/projects\/****\/users\/****\/python\/3.7.0  tensorflow\n<\/code><\/pre>\n\n<p>it will have such an error message:<\/p>\n\n<pre><code>Cache entry deserialization failed, entry ignored\nCould not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow\n<\/code><\/pre>\n\n<p>So how exactly to install tensorflow on python3.7?<\/p>\n\n<p>I tried the instructions in the other post but none of them works.<\/p>\n","answers":[{"AnswerId":"52656096","CreationDate":"2018-10-04T22:38:33.080","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>Tensorflow does not support Python 3.7 yet (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022<\/a>).<\/p>\n\n<p>Use Python 3.6.<\/p>\n"}]}
+{"QuestionId":52655899,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-04T22:12:19.163","AcceptedAnswerId":null,"OwnerUserId":10458992.0,"Title":"How to implement Attention using Keras","Body":"<p>I have a keras model that has sequence of inputs and sequence of outputs where each input has an associated output(Label). <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Masking(mask_value=5, input_shape= (Seq_in.shape[1],1)))\nmodel.add(Bidirectional(LSTM(256,  return_sequences=True)))\nmodel.add(Dropout(0.2))\nmodel.add(Bidirectional(LSTM(256, return_sequences=True))) \nmodel.add(Dropout(0.2))\nmodel.add(Dense(n_Labels, activation='softmax'))  # n_Labels is the number of labels which is 15\nsgd = optimizers.SGD(lr=.1,momentum=0.9,decay=1e-3,nesterov=True)\nmodel.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\nmodel.fit(X_train,Y_train,epochs=2, validation_data=(X_val, Y_val),verbose=1)\n<\/code><\/pre>\n\n<p>Now I want to implement attention mechanism following this work Zhou et al: \n\"Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification\".<\/p>\n\n<blockquote>\n  <p>For each output, we compute the tanh for each output states in the\n  sequence(Eq# 9 in the paper), then we compute the softmax for each\n  output states in regard to the current output (Eq#10), then we\n  multiply each output states with the corrsponding softmax (attention)\n  (Eq# 11), then we take the sum of the weighted states, then we take\n  the tanh of the final output which represent. Finally we concatenate\n  the attention vector with the output states.<\/p>\n<\/blockquote>\n\n<p>How can I do that? is it possiple to that with the keras API or I have to come up with my own custom layer? any help ?<\/p>\n\n<p>Thank you in advance.... <\/p>\n","answers":[{"AnswerId":"52659185","CreationDate":"2018-10-05T06:00:55.487","ParentId":null,"OwnerUserId":"5367279","Title":null,"Body":"<p>There is no Keras API yet. But a lot of hard working programmers made some good implementations using Keras. You can try looking at the code in <a href=\"https:\/\/github.com\/andreyzharkov\/keras-monotonic-attention\" rel=\"nofollow noreferrer\">keras-monotonic-attention<\/a>.<\/p>\n"}]}
+{"QuestionId":52655925,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-data-validation>","CreationDate":"2018-10-04T22:15:33.190","AcceptedAnswerId":null,"OwnerUserId":10459023.0,"Title":"pip install tensorflow-data-validation fails on OS X El Capitan (10.11.6)","Body":"<p>I am not able to install tensorflow-data-validation using pip. when I use:<\/p>\n\n<p><code>pip install tensorflow-data-validation<\/code><\/p>\n\n<p>I get this error:<\/p>\n\n<blockquote>\n  <p>Could not find a version that satisfies the requirement\n  tensorflow-data-validation<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52727380","CreationDate":"2018-10-09T18:37:07.167","ParentId":null,"OwnerUserId":"6447747","Title":null,"Body":"<p>Are you using Python 3? Because TensorFlow Data Validation currently requires Python 2.7. <\/p>\n"}]}
+{"QuestionId":52655955,"AnswerCount":2,"Tags":"<docker><tensorflow><gpu><nvidia><pytorch>","CreationDate":"2018-10-04T22:19:19.063","AcceptedAnswerId":null,"OwnerUserId":10458980.0,"Title":"Does Docker image supporting Nvidia gpu on a cpu machine","Body":"<p>I want to run a deep neural network on a Docker container. I can use TensorFlow or PyTorch or CNTK images available publicly on Docker Hub that support a GPU. And I can get an NVIDA GPU in the cloud. <\/p>\n\n<p>What happens when I run those GPU-based containers on my PC that does not have a GPU (or an Intel GPU)? Will the Docker container based on NVIDIA GPU gracefully run on my a CPU (albeit slower)?<\/p>\n\n<p>It would be great if I could develop on locally without a GPU and scale up with GPU when I am ready to run at scale.<\/p>\n\n<p>Will this work? Thoughts?<\/p>\n","answers":[{"AnswerId":"52824979","CreationDate":"2018-10-15T21:25:52.840","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>Tensorflow has an option for allowing soft placement, which means that if the device doesn't exist, it automatically chooses an existing and supported device to run the operations. Quoting from the documentation about <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">Using GPUs<\/a> in tensorflow, <\/p>\n\n<blockquote>\n  <p>If you would like TensorFlow to automatically choose an existing and\n  supported device to run the operations in case the specified one\n  doesn't exist, you can set allow_soft_placement to True in the\n  configuration option when creating the session.<\/p>\n\n<pre><code># Creates a graph.\nwith tf.device('\/device:GPU:2'):\n  a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n  b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n  c = tf.matmul(a, b)\n# Creates a session with allow_soft_placement and log_device_placement set\n# to True.\nsess = tf.Session(config=tf.ConfigProto(\n      allow_soft_placement=True, log_device_placement=True))\n# Runs the op.\nprint(sess.run(c))\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"52656042","CreationDate":"2018-10-04T22:31:20.907","ParentId":null,"OwnerUserId":"2934305","Title":null,"Body":"<p>If Tensorflow does not find a GPU it will just use the CPU only so this should work without problems.<\/p>\n"}]}
+{"QuestionId":52656088,"AnswerCount":0,"Tags":"<numpy><keras>","CreationDate":"2018-10-04T22:37:39.790","AcceptedAnswerId":null,"OwnerUserId":7280300.0,"Title":"Keras: How to get class label from model.predict when using sigmoid for binary classsification","Body":"<p>I have implemented a basic neural network in <code>keras<\/code> with <code>sigmoid<\/code> activation at the output layer. However, when I use <code>model.predict(x_new)<\/code> on my unseen data, it returns me a <code>numpy<\/code> array with values ranging <code>0-1<\/code>. I want a binary class output. How to do that?<\/p>\n","answers":[]}
+{"QuestionId":52656293,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-04T23:07:45.257","AcceptedAnswerId":null,"OwnerUserId":1031417.0,"Title":"How does Keras set the dimensions in this network which has CNN and dense layers?","Body":"<p>I need some help to understand what's going on here. <\/p>\n\n<p>My goal is to have a network that receives <code>sizeXsize<\/code> images and returns <code>sizeXsize<\/code> binary matrices. The output of the network should be a binary <code>sizeXsize<\/code> matrix that indicates if a pixel has a feature or not. <\/p>\n\n<p>For example, think of a corner detection network where the output layer tells if a pixel is exactly a tip of the corner. Namely, we want to detect only the pixel of this corner:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/N92Gp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/N92Gp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The first layers in the networks are defined as follows:<\/p>\n\n<pre><code>from keras import models, layers\nimport numpy as np\n\nsize=5\n\ninput_image = layers.Input(shape=(size, size, 1))\n\nb = layers.Conv2D(5, (3,3), activation='relu', padding='same')(input_image)\nb = layers.MaxPooling2D((2,2), strides=1,  padding='same')(b)\nb = layers.Conv2D(5, (3,3), activation='relu', padding='same')(b)\nb_out = layers.MaxPooling2D((2,2),strides=1 ,padding='same')(b)\n<\/code><\/pre>\n\n<p>Until now I maintained the dimensions of the original input layer (<code>sizeXsize<\/code>).<\/p>\n\n<p>Now I would like to have a dense layer as an output layer with <code>sizeXsize<\/code> pixels. <\/p>\n\n<p>If I use <code>output = layers.Dense(size, activation='sigmoid')(b_out)<\/code> the layer built is <code>sizeXsizeXsize<\/code>, and if I do <code>output = layers.Dense(1, activation='sigmoid')(b_out)<\/code> the size is <code>sizeXsize<\/code>, how comes?!<\/p>\n\n<p>This is the building and the compilation part of the code:<\/p>\n\n<pre><code>model = models.Model(input_image, output)\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.summary()\n<\/code><\/pre>\n\n<p>What do I miss here? Isn't <code>output = layers.Dense(1, activation='sigmoid')(b_out)<\/code> just a single neuron? <\/p>\n\n<p>The thing is that if I train:<\/p>\n\n<pre><code>n_images=100\ndata = np.random.randint(0,2,(n_images,size,size,1))\nlabels = np.random.randint(0,2,(n_images,size,size,1))\nlabels = data\n\nmodel.fit(data, labels, verbose=1, batch_size=4, epochs=20)\n<\/code><\/pre>\n\n<p>and if I test it:<\/p>\n\n<pre><code>data1 = np.random.randint(0,2,(n_images,size,size,1))\nscore, acc = model.evaluate(data1,data1, verbose=1)\n\n\nprint('Test score:', score)\nprint('Test accuracy:', acc)\n\na=np.random.randint(0,2,(1,size,size,1))\nprediction = model.predict(a)\n\nprint(a==np.round(prediction))\n<\/code><\/pre>\n\n<p>I get a good accuracy, and it seems the sizes are correct for the output layer:<\/p>\n\n<pre><code>100\/100 [==============================] - 0s 349us\/step\nTest score: 0.187119951248\nTest accuracy: 0.926799981594\n[[[[ True]\n   [ True]\n   [ True]\n   [ True]\n   [ True]]\n\n  [[ True]\n   [ True]\n   [ True]\n   [ True]\n   [ True]]\n\n  [[ True]\n   [ True]\n   [ True]\n   [ True]\n   [ True]]\n\n  [[ True]\n   [ True]\n   [ True]\n   [ True]\n   [ True]]\n\n  [[ True]\n   [ True]\n   [ True]\n   [ True]\n   [ True]]]]\n<\/code><\/pre>\n\n<hr>\n\n<p>If I read <a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\">Dense<\/a> documentation:<\/p>\n\n<blockquote>\n  <p>units: Positive integer, dimensionality of the output space.<\/p>\n<\/blockquote>\n\n<p>So how comes if I put <code>layers.Dense(1, activation='sigmoid')(b_out)<\/code> I get an output layer of <code>sizeXsize<\/code>?<\/p>\n","answers":[{"AnswerId":"52662175","CreationDate":"2018-10-05T09:12:17.207","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Your confusion stems from the fact that <code>Dense<\/code> layer is currently implemented such that <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">it is applied on the last axis of input data<\/a>. That's why when you feed the output of MaxPooling layer (i.e. <code>b_out<\/code>), which has a shape of <code>(size, size, 5)<\/code>, to a Dense layer with one unit you get an output of shape <code>(size, size, 1)<\/code>. In this case, the single neuron in the Dense layer is connected to each of 5 elements in the output array, though with the same weights (that's why if you take a look at the <code>summary()<\/code> output, you would see that the Dense layer has 6 parameters, 5 weights plus one bias parameter). <\/p>\n\n<p>You can either use a Dense layer (with one unit) or a Conv2D layer (with one filter) as the last layer. If you ask which one works better the answer is It depends on the specific problem you are working on and the data you have. However,  you can take some ideas from image segmentation networks where first the image is processed with a combination of Conv2D and MaxPooling2D layers (and its dimension is reduced as we go forward in the model) and then some upsample layers and Conv2D layers are used to get back the image with the same size as input image. <a href=\"https:\/\/stackoverflow.com\/a\/52563566\/2099607\">Here<\/a> is a sketch (though, you don't need to use <code>TimeDistributed<\/code> and <code>LSTM<\/code> layers for your case). <\/p>\n"},{"AnswerId":"52658555","CreationDate":"2018-10-05T04:54:52.693","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>The trick is not to use the conventional <code>Dense<\/code> layer, but use a convolutional layer with kernel size <code>(1,1)<\/code>, i.e. you need something like below:<\/p>\n\n<pre><code>b = layers.Conv2D(5, (3,3), activation='relu', padding='same')(input_image)\nb = layers.MaxPooling2D((2,2), strides=1,  padding='same')(b)\nb = layers.Conv2D(5, (3,3), activation='relu', padding='same')(b)\nb = layers.MaxPooling2D((2,2),strides=1 ,padding='same')(b)\n# not use Dense, but Conv2D\nbinary_out = layers.Conv2D(1, (1,1), activation='sigmoid', padding='same')(b)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52656445,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-04T23:34:01.680","AcceptedAnswerId":null,"OwnerUserId":569313.0,"Title":"Keras Conv1D on Multiple Time Series : One at a time","Body":"<p>I have a numpy array that is shaped as (number of observations, number of time steps, number of features). This equates to say 59 features (i.e. time series) each of length 24 (number of time steps above). This is the standard setup for input into an RNN.<\/p>\n\n<p>What I would like to do is feed this into a model where the first step is a 1D convolution on each series separately and then concatenate the 59 outputs into the next layer (dense layer or an LSTM and then dense layer).<\/p>\n\n<p>If I loop through the number of features and apply a convolution to each series, how should the input X be structured? <\/p>\n\n<p>For example, say my data is as follows:<\/p>\n\n<pre><code>var=np.array([[[1., 3., 5., 7.],\n        [2., 4., 6., 8.]],\n\n       [[9., 11., 13., 15.],\n        [10., 12., 14., 16.]],\n\n       [[17., 19., 21., 23.],\n        [18., 20., 22., 24.]]])\n<\/code><\/pre>\n\n<p>where the first observation, time series 1 is [1,2] and the second time series is [3,4]....and the first observations fourth time series is [7,8].<\/p>\n\n<pre><code>nb_filters=(8,4)\nfilter_length=(5,5)\ndense=(16,8)\n\n\n    conv_layers=[]\n    input=Input(shape=(None,no_timesteps,no_features))\n\n    for i in range(no_features):\n        series=Lambda(lambda x: x[:, i])(input)\n        print(series.shape)\n        conv_layer=Conv1D(nb_filters[0], kernel_size=filter_length[0], padding='valid',activation='relu')(series)\n        conv_layer = MaxPooling1D()(conv_layer)\n        conv_layer=Conv1D(nb_filters[1], kernel_size=filter_length[1], padding='valid',activation='relu')(conv_layer)\n        conv_layer = MaxPooling1D()(conv_layer)\n        conv_layers.append(conv_layer)\n    merge_layer=Concatenate(axis=1)(conv_layers)\n    merge_layer= Flatten()(merge_layer)\n    output=Dense(dense[0], activation='sigmoid')(merge_layer)\n\n    nn=Model(inputs=inputs,outputs=output)\n    nn.compile(optimizer=RMSprop(), loss='binary_crossentropy')\n    nn.summary()\n    nn.fit(X,y, epochs=5)\n<\/code><\/pre>\n\n<p>First, this doesn't work as expected since the print statement gives a shape of (?, 24, 59) which doesnt appear that I am actually slicing in the lambda. I expected each slice to be a specific time series of shape (1,24) where 24 is the length of each series. The idea is to have the 1D convolution ran over this single time series (repeat for all 59 series) and concatenate them together.<\/p>\n","answers":[{"AnswerId":"52662384","CreationDate":"2018-10-05T09:23:04.890","ParentId":null,"OwnerUserId":"583464","Title":null,"Body":"<p>You must use:<\/p>\n\n<p><code>input=Input(shape=(None, no_timesteps,no_features))<\/code><\/p>\n\n<p>or <\/p>\n\n<p><code>input=Input(shape=(number_of_samples, no_timesteps,no_features))<\/code><\/p>\n\n<p>The input layer is a tensor which must have the same shape as your training data.<\/p>\n"}]}
+{"QuestionId":52656692,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><tensor><google-colaboratory>","CreationDate":"2018-10-05T00:15:14.080","AcceptedAnswerId":null,"OwnerUserId":5157633.0,"Title":"Debugging in Google Colab","Body":"<p>I am running the following code snippet in google colab in a single cell:<\/p>\n\n<pre><code>%debug\n# Create tensors of shape (10, 3) and (10, 2).\nx = torch.randn(10, 3)\ny = torch.randn(10, 2)\n\n# Build a fully connected layer.\nlinear = nn.Linear(3, 2)\nprint ('w: ', linear.weight)\nprint ('b: ', linear.bias)\n<\/code><\/pre>\n\n<p>I wish to debug a piece of code (step through it line by line) to understand what is going on. I wish to step inside the function nn.Linear. <\/p>\n\n<p>However, when I step through, it does not enter the function at  all. Is there a way to step through nn.Linear line by line? Also, how exactly do I set a breakpoint in nn.Linear? Besides, I wish to step though the snippet line by line as well. However, as the picture shows, the step command automatically steps through and executes the print statement as well.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PSBvV.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PSBvV.png\" alt=\"Step_though_collab\"><\/a><\/p>\n","answers":[{"AnswerId":"55191078","CreationDate":"2019-03-15T21:46:20.667","ParentId":null,"OwnerUserId":"4113721","Title":null,"Body":"<p>Since Python 3.7 you can use a built-in <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0553\/\" rel=\"noreferrer\">breakpoint function<\/a>. If this is not available, you can use <code>import pdb; pdb.set_trace()<\/code> instead.<\/p>\n\n<p>If you want to execute the next line you can try <code>n<\/code> (next) instead of <code>s<\/code> (step).<\/p>\n"}]}
+{"QuestionId":52656737,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-05T00:23:26.157","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Export Model File is Incorrect","Body":"<p>I have a Keras model called <code>model<\/code> using a Tensorflow backend and I'm trying to save it in an ML Engine-friendly format.  After googling around, here's my attempt:<\/p>\n\n<pre><code>    from tensorflow import saved_model\n    import tensorflow.keras.backend as K\n\n    K.clear_session()\n    sess = tf.Session()\n    K.set_session(sess)\n\n    # build a copy but without the learning nodes\n    K.set_learning_phase(0)  # disables creation of dropout\n\n    model_builder = saved_model.builder.SavedModelBuilder(\"exported_model\")\n    inputs = {\n        'input': saved_model.utils.build_tensor_info(model.input)\n    }\n    outputs = {\n        'output': saved_model.utils.build_tensor_info(model.output)\n    }\n    signature_def = saved_model.signature_def_utils.build_signature_def(\n        inputs=inputs,\n        outputs=outputs,\n        method_name=saved_model.signature_constants.PREDICT_METHOD_NAME\n    )\n    model_builder.add_meta_graph_and_variables(\n        K.get_session(),\n        tags=[saved_model.tag_constants.SERVING],\n        signature_def_map={saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_def\n                           })\n    model_builder.save()\n<\/code><\/pre>\n\n<p>The output is <code>saved_model.pb<\/code> but it's less than 1 MB when the <code>model.h5<\/code> file is over 100MB.  So I suspect it isn't saving the weights of my model?<\/p>\n\n<p>Any thoughts on what I'm doing incorrectly?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52657421,"AnswerCount":1,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2018-10-05T02:10:04.100","AcceptedAnswerId":52659708.0,"OwnerUserId":7155684.0,"Title":"Keras custom loss error : Unknown loss function","Body":"<p>I have tried to customize a loss function in Keras.<\/p>\n\n<p>I have tried two approaches:<\/p>\n\n<pre><code>import keras.backend as K\nfrom keras.losses import mean_absolute_error\n\ndef mae_in_minute(y_true, y_pred):\n    temp = K.mean(K.abs(y_pred - y_true), axis=-1)\/60\n    return temp\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>import keras.backend as K\nfrom keras.losses import mean_absolute_error\n\ndef mae_in_minute(y_true, y_pred):\n    return mean_absolute_error(y_true, y_pred)\/60\n<\/code><\/pre>\n\n<p>My model structure is:<\/p>\n\n<pre><code>input_layer = Input(shape=training.shape[1:len(training.shape)])\nadded = Conv2D(128, (3, training.shape[2]),activation=\"relu\")(input_layer)\nadded = Flatten()(added)\nadded = Dense(600, activation='relu')(added)\nadded = Dense(400, activation='relu')(added)\nadded = Dense(256, activation='relu')(added)\nadded = Dense(256, activation='relu')(added)\nadded = Dense(256, activation='relu')(added)\nadded = Dense(200, activation='relu')(added)\nadded = Dense(100, activation='relu')(added)\nadded = Dense(50, activation='relu')(added)\noutput_temp = Dense(2,activation='softmax', name=\"temp_output\")(added)\noutput_time = Dense(1,activation='relu', name=\"time_output\")(added)\nmodel = Model(input=input_layer, output=[output_temp,output_time])\nlosses = {\n    \"temp_output\": \"categorical_crossentropy\",\n    \"time_output\": \"mae_in_minute\",\n}\nlossWeights = {\"temp_output\": 1.0, \"time_output\": 1.0}\nmodel.compile(optimizer='adam',loss=losses, loss_weights=lossWeights)\nmodel.summary()\n<\/code><\/pre>\n\n<p>But I get this error message with both custom loss approaches: <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>Unknown loss function:mae_in_minute<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>How do I fix this problem?<\/p>\n\n<p>I have found one solution <a href=\"https:\/\/stackoverflow.com\/questions\/45576576\/keras-unknown-loss-function-error-after-defining-custom-loss-function\">here<\/a>.<\/p>\n\n<p>But is this the only way to use a custom loss? To save my model in advance and load it?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"52659708","CreationDate":"2018-10-05T06:41:22.363","ParentId":null,"OwnerUserId":"7155684","Title":null,"Body":"<p>Just remove the quoation of custom loss, and it should run prefectly.<\/p>\n\n<h1>My_loss<\/h1>\n\n<pre><code>import keras.backend as K\nfrom keras.losses import mean_absolute_error\n\ndef mae_in_minute(y_true, y_pred):\n    return mean_absolute_error(y_true, y_pred)\/60\n<\/code><\/pre>\n\n<h2>Before<\/h2>\n\n<pre><code>losses = {\n    \"temp_output\": \"categorical_crossentropy\",\n    \"time_output\": \"mae_in_minute\",\n}\nlossWeights = {\"temp_output\": 1.0, \"time_output\": 1.0}\nmodel.compile(optimizer='adam',loss=losses, loss_weights=lossWeights)\nmodel.summary()\n<\/code><\/pre>\n\n<h2>After<\/h2>\n\n<pre><code>losses = {\n    \"temp_output\": \"categorical_crossentropy\",\n    \"time_output\": mae_in_minute,\n}\nlossWeights = {\"temp_output\": 1.0, \"time_output\": 1.0}\nmodel.compile(optimizer='adam',loss=losses, loss_weights=lossWeights)\nmodel.summary()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52657599,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2018-10-05T02:33:19.330","AcceptedAnswerId":52657849.0,"OwnerUserId":8133867.0,"Title":"Is there any way to verify that I have successfully frozen certain variables\/parameters in my Tensorflow model?","Body":"<p>I attempted to freeze some of the variables by only passing the variables that I would like to fine-tune on via the<\/p>\n\n<pre><code> tf.train.AdamOptimizer.compute_gradients(var_list=[])\n<\/code><\/pre>\n\n<p>approach.<\/p>\n\n<p>But when I proceeded with fine-tuning my model, the loss is going up and the accuracy is going down so I am having doubt that I have successfully frozen the initial few layers.<\/p>\n\n<p>Any way to check if I'm freezing variables properly? Thanks!<\/p>\n","answers":[{"AnswerId":"52657849","CreationDate":"2018-10-05T03:09:04.933","ParentId":null,"OwnerUserId":"9415043","Title":null,"Body":"<p>Try <code>tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)<\/code>.<\/p>\n"}]}
+{"QuestionId":52657786,"AnswerCount":0,"Tags":"<tensorflow><restore><production>","CreationDate":"2018-10-05T02:59:32.527","AcceptedAnswerId":null,"OwnerUserId":2801404.0,"Title":"Tensorflow: how to restore only specific hidden layers from checkpoint and use them to build a different computational graph for inference?","Body":"<p>Let's say I trained a model with a very complex computational graph tailored for training. After a lot of training, the best model was saved to a checkpoint file. Now, I want to use the learned parameters of this best model for inference. However, the computational graph used for training is not exactly the same as the one I intend to use for inference. Concretely, there is a module in the graph with several layers in charge of outputting embedding vectors for items (recommender system context). However, for the sake of computational performance, during inference time I would like to have all the item embedding vectors precomputed in advance, so that the only computation required per request would just involve a couple of hidden layers.<\/p>\n\n<p>Therefore, what I would like to know how to do is:<\/p>\n\n<ol>\n<li><p>How to just restore the part of the network that outputs item embedding vectors, in order to precompute these vectors for all items (this would happen in some pre-processing script off-line)<\/p><\/li>\n<li><p>Once all item embedding vectors are precomputed, during on-line inference time how to just restore the hidden layers in the later parts of the network and make them receive the precomputed item embedding vectors instead.<\/p><\/li>\n<\/ol>\n\n<p>How can the points above be accomplished? I think point 1. is easier to get done. But my biggest concern is with point 2. In the computational graph used for training, in order to evaluate any layer I would have to provide values for the input placeholders. However, during on-line inference these placeholders would be obsolete because a lot of stuff would be precomputed and I don't know how to tell hidden layers in the later parts of the network that they should no longer depend on these obsolete placeholders but depend on the precomputed stuff instead.<\/p>\n","answers":[]}
+{"QuestionId":52657811,"AnswerCount":0,"Tags":"<tensorflow><localhost><tensorboard>","CreationDate":"2018-10-05T03:02:41.057","AcceptedAnswerId":null,"OwnerUserId":6002424.0,"Title":"tensorboard localhost connection","Body":"<p>I can't seem to find exact question about what I am about to ask here. I just started following a Tensorflow tutorial on <a href=\"https:\/\/www.youtube.com\/watch?v=9kC836XhICU&amp;index=2&amp;list=PLIDllPt3EQZoS8gCP3cw273Cq9puuPLTg\" rel=\"nofollow noreferrer\">YouTube<\/a> and got stuck at the very beginning. I wrote in my spyder IDE the below code:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant(2)\nb = tf.constant(3)\n\nx = tf.add(a,b)\n#writer = tf.summary.FileWriter('.\/graphs', tf.get_default_graph())\nwith tf.Session() as sess:\n    writer = tf.summary.FileWriter('.\/graphs', sess.graph)\n    print(sess.run(x))\n\nwriter.close()\n<\/code><\/pre>\n\n<p>And via anaconda terminal I activated my env (which I newly created, installed all packages required, spyder as well) I typed <code>python tftuts.py<\/code> and got <code>2018-10-05 11:50:49.431174: I T:\\src\\github\\tensorflow\\tensorflow\\core\\platform\\cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n5<\/code><\/p>\n\n<p>Then I typed <code>tensorboard --logdir=\".\/graphs\" --port 6006<\/code> as suggested in tutorial I am watching.\nNow, when I go to <code>http:\/\/localhost:6006\/<\/code> the page shows\n<a href=\"https:\/\/i.stack.imgur.com\/k3RCZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/k3RCZ.jpg\" alt=\"this\"><\/a><\/p>\n\n<p>I am on Win10, using python 3.6.6 in Anaconda env, tensorflow 1.10.0. <\/p>\n\n<p>How can solve this issue?<\/p>\n","answers":[]}
+{"QuestionId":52658204,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2018-10-05T04:03:44.103","AcceptedAnswerId":52662650.0,"OwnerUserId":4942590.0,"Title":"How the parameters of softmax are obtained in TensorFlow if the softmax is not included in training?","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/wvjvR.png\" rel=\"nofollow noreferrer\">Here is a tensorflow graph<\/a>, as we can see that one of the inputs of cross entropy is the output of the logit layer, not the output of softmax in the graph.<\/p>\n\n<p>I searched about it and found \"WARNING: This op expects unscaled logits, since it performs a softmax on logits internally for efficiency. Do not call this op with the output of softmax, as it will produce incorrect results.\" on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">this webpage<\/a>.<\/p>\n\n<p>My question is how the parameters of the softmax are obtained if this softmax is not trained? <\/p>\n","answers":[{"AnswerId":"52662650","CreationDate":"2018-10-05T09:38:39.043","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/YLeRi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YLeRi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Softmax is a parameter free activation function like RELU, Tanh or Sigmoid: it doesn't need to be trained. It only computes the exponential of every logit and then normalize the output vector by the sum of the exponentials.<\/p>\n"}]}
+{"QuestionId":52658755,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-05T05:16:24.610","AcceptedAnswerId":null,"OwnerUserId":10333833.0,"Title":"How to get filter weights after training?","Body":"<p>I want to see what values filter has adopted after training? If after the training loop I mention filter.eval() will it get me the values of filter weights that the filter has adopted after complete training? I don't think so I can get the filter weights in this way because filter variable calling a function weight_variable which is picking some values from normal distribution. I think calling filter.eval() command after training loop is simply like printing filter before training. So how can I get the values of filter weights that filter has adopted after training?<\/p>\n\n<pre><code>`def weight_variable(shape):\n    initial = tf.truncated_normal(shape, mean=0, stddev=0.1)\n    return tf.Variable(initial)\n#network\nx = tf.placeholder(tf.float32, [None, \nFLAGS.image_height*FLAGS.image_width])\ny_ = tf.placeholder(tf.float32, [None, 2])\n\n\n\ninput=tf.reshape(x, \n    [-1,FLAGS.image_height,FLAGS.image_width,FLAGS.input_channel])\nfilter = weight_variable([FLAGS.filter_size, FLAGS.filter_size, \n    FLAGS.input_channel, FLAGS.filter_channel])\nconv_out = tf.nn.sigmoid(conv2d(input, filter))\n\npool_out = max_pool(conv_out)\n\n\npool_list = pool_out.get_shape().as_list()\ninput_dim = pool_list[1]* pool_list[2]* pool_list[3]\npool_2D = tf.reshape(pool_out, [-1, input_dim])\nW_fc = weight_variable([input_dim, 2])\n\nlogits = tf.matmul(pool_2D, W_fc)   #(batch_size,2)\ny_conv=tf.nn.softmax(logits)`\n<\/code><\/pre>\n\n<h1>after checking correct predictions apply training loop<\/h1>\n\n<pre><code>`\nfor i in range(max.training_step):\n    #Check training and test accuracy\nprint(filter.eval())`\n<\/code><\/pre>\n","answers":[{"AnswerId":"52661604","CreationDate":"2018-10-05T08:38:46.117","ParentId":null,"OwnerUserId":"241013","Title":null,"Body":"<p>First, make sure you use the same graph and session. Graph contains the setup of your model and session contains the values of all the weights in the model. If you don't use the same graph\/session you will not get the same weights. I think in your example you use some default graph and session. It should work but it's better if you are more explicit about them.<\/p>\n\n<p>It looks like <code>filter<\/code> should be a tensor. If you do <code>print(filter)<\/code> you should see something like <code>&lt;tensor object shape=(..) &gt;<\/code>. In this case <code>print(filter.eval())<\/code> will return you the correct weights.<\/p>\n\n<p>If not, you can use the graph to get the tensor. <code>graph.get_tensor_by_name('filter_weights:0')<\/code> will give you a tensor that you can look at. You can get the name from tensorboard (from the graph) or you can run <code>tf.trainable_variables()<\/code> to get a list of all the variables defined in the graph (and pick the one you want).<\/p>\n\n<p>An alternative to <code>filter.eval()<\/code> is to do <code>session.run(filter)<\/code>. They are equivalent.<\/p>\n"}]}
+{"QuestionId":52658848,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-05T05:27:12.107","AcceptedAnswerId":null,"OwnerUserId":1090562.0,"Title":"How to remove\/rename the functions from tensorboard which were generated by TFRecordDataset parse_single_example","Body":"<p>I have a tf model which reads the data from <code>tf.data.TFRecordDataset<\/code>. So I have something like this:<\/p>\n\n<pre><code>keys_to_map = {\n    \"d\": tf.FixedLenFeature([], tf.string),\n    \"s\": tf.FixedLenFeature([], tf.int64),\n}\n\ndef parser(record):\n    parsed = tf.parse_single_example(record, keys_to_map)\n    return parsed[\"d\"], parsed[\"s\"] \n\ntf.data.TFRecordDataset(files_train).shuffle(buffer_size=10000).map(parser).batch(batch_size, drop_remainder=True).repeat(-1).make_initializable_iterator()\n<\/code><\/pre>\n\n<p>when I visualize the graph, one part of it looks like this (<em>tf_data_structured_function_wrapper<\/em>):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/V2c0g.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/V2c0g.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there a way to rename\/remove these functions into something more descriptive. Now it is clear what they mean and how they appeared in the code (in hindsight after I localized the problem), but when I had a huge graph it was very confusing.<\/p>\n","answers":[]}
+{"QuestionId":52659074,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-05T05:50:56.127","AcceptedAnswerId":null,"OwnerUserId":10027028.0,"Title":"Installing Tensorflow for Latest Python 3 Release - Python 3.6.0 (follow up)","Body":"<p>On request of @Jialer Chew new question posted as follow up from \nInstalling Tensorflow for Latest Python 3 Release - Python 3.6.0!\nPrevious question rases out of border.\nMy problem installing Tensorflow is now as follow:<\/p>\n\n<p><code>\n    --------------------------------------------------------------------------- ImportError Traceback (most recent call last) ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in swig_import_helper() 13 try: ---&gt; 14 return importlib.import_module(mname) 15 except ImportError: ~\\Anaconda3\\lib\\importlib\\__init__.py in import_module(name, package) 125 level += 1 --&gt; 126 return _bootstrap._gcd_import(name[level:], package, level) 127 ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in _gcd_import(name, package, level) ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in _find_and_load(name, import_) ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in _find_and_load_unlocked(name, import_) ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in _load_unlocked(spec) ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in module_from_spec(spec) ~\\Anaconda3\\lib\\importlib\\_bootstrap_external.py in create_module(self, spec) ~\\Anaconda3\\lib\\importlib\\_bootstrap.py in _call_with_frames_removed(f, *args, **kwds) ImportError: DLL load failed: Een initialisatieroutine van de dynamic link library (DLL-bestand) is mislukt. During handling of the above exception, another exception occurred: ModuleNotFoundError Traceback (most recent call last) ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in &lt;module&gt;() 57 ---&gt; 58 from tensorflow.python.pywrap_tensorflow_internal import * 59 from tensorflow.python.pywrap_tensorflow_internal import __version__ ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in &lt;module&gt;() 16 return importlib.import_module('_pywrap_tensorflow_internal') ---&gt; 17 _pywrap_tensorflow_internal = swig_import_helper() 18 del swig_import_helper ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in swig_import_helper() 15 except ImportError: ---&gt; 16 return importlib.import_module('_pywrap_tensorflow_internal') 17 _pywrap_tensorflow_internal = swig_import_helper() ~\\Anaconda3\\lib\\importlib\\__init__.py in import_module(name, package) 125 level += 1 --&gt; 126 return _bootstrap._gcd_import(name[level:], package, level) 127 ModuleNotFoundError: No module named '_pywrap_tensorflow_internal' During handling of the above exception, another exception occurred: ImportError Traceback (most recent call last) &lt;ipython-input-1-fae406f752e9&gt; in &lt;module&gt;() 6 from PIL import Image 7 from scipy import ndimage ----&gt; 8 import tensorflow as tf 9 from tensorflow.python.framework import ops 10 from cnn_utils import * ~\\Anaconda3\\lib\\site-packages\\tensorflow\\__init__.py in &lt;module&gt;() 20 21 # pylint: disable=g-bad-import-order ---&gt; 22 from tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import 23 24 try: ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\__init__.py in &lt;module&gt;() 47 import numpy as np 48 ---&gt; 49 from tensorflow.python import pywrap_tensorflow 50 51 # Protocol buffers ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in &lt;module&gt;() 72 for some common reasons and solutions. Include the entire stack trace 73 above this error message when asking for help.\"\"\" % traceback.format_exc() ---&gt; 74 raise ImportError(msg) 75 76 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long ImportError: Traceback (most recent call last): File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 14, in swig_import_helper return importlib.import_module(mname) File \"C:\\Users\\gstevens\\Anaconda3\\lib\\importlib\\__init__.py\", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File \"&lt;frozen importlib._bootstrap&gt;\", line 994, in _gcd_import File \"&lt;frozen importlib._bootstrap&gt;\", line 971, in _find_and_load File \"&lt;frozen importlib._bootstrap&gt;\", line 955, in _find_and_load_unlocked File \"&lt;frozen importlib._bootstrap&gt;\", line 658, in _load_unlocked File \"&lt;frozen importlib._bootstrap&gt;\", line 571, in module_from_spec File \"&lt;frozen importlib._bootstrap_external&gt;\", line 922, in create_module File \"&lt;frozen importlib._bootstrap&gt;\", line 219, in _call_with_frames_removed ImportError: DLL load failed: Een initialisatieroutine van de dynamic link library (DLL-bestand) is mislukt. During handling of the above exception, another exception occurred: Traceback (most recent call last): File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt; from tensorflow.python.pywrap_tensorflow_internal import * File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 17, in &lt;module&gt; _pywrap_tensorflow_internal = swig_import_helper() File \"C:\\Users\\gstevens\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 16, in swig_import_helper return importlib.import_module('_pywrap_tensorflow_internal') File \"C:\\Users\\gstevens\\Anaconda3\\lib\\importlib\\__init__.py\", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) ModuleNotFoundError: No module named '_pywrap_tensorflow_internal' Failed to load the native TensorFlow runtime. See https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n<\/code><\/p>\n","answers":[]}
+{"QuestionId":52659556,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-05T06:31:07.560","AcceptedAnswerId":null,"OwnerUserId":6180818.0,"Title":"Tensorflow Type Error when generate tf record","Body":"<p>This is part of my tensorflow generate tf record<\/p>\n\n<pre><code>def class_text_to_int(row_label):\n    if row_label == \"110kmh\": \n        return 1\n    else:\n        return 0\n<\/code><\/pre>\n\n<p>However I am getting this eror<\/p>\n\n<pre><code> File \"\/usr\/local\/lib\/python3.5\/dist-packages\/object_detection-0.1-py3.5.egg\/object_detection\/utils\/dataset_util.py\", line 34, in bytes_list_feature\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))\n<\/code><\/pre>\n\n<p>Is anyone familiar with this error? I been searching through the net whole day but still not getting any solution.<\/p>\n","answers":[{"AnswerId":"55051683","CreationDate":"2019-03-07T19:43:40.270","ParentId":null,"OwnerUserId":"9770490","Title":null,"Body":"<p>If you need other (working) <code>generate_tfrecord.py<\/code> file:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\"\"\"\nUsage:\n# Create train data:\npython generate_tfrecord.py --label_map=&lt;PATH_TO_LABEL_MAP_FILE&gt; --csv_input=&lt;PATH_TO_ANNOTATIONS_FOLDER&gt;\/\ntrain_labels.csv --output_path=&lt;PATH_TO_ANNOTATIONS_FOLDER&gt;\/train.record\n\n# Create test data:\npython generate_tfrecord.py --label_map=&lt;PATH_TO_LABEL_MAP_FILE&gt; --csv_input=&lt;PATH_TO_ANNOTATIONS_FOLDER&gt;\/\ntest_labels.csv --output_path=&lt;PATH_TO_ANNOTATIONS_FOLDER&gt;\/test.record\n\"\"\"\nfrom __future__ import division\nfrom __future__ import print_function\nfrom __future__ import absolute_import\nimport os\nimport io\nimport pandas as pd\nimport tensorflow as tf\nimport sys\n\nsys.path.append(\"..\/..\/models\/research\")\nfrom PIL import Image\nfrom object_detection.utils import dataset_util\nfrom collections import namedtuple, OrderedDict\n\nflags = tf.app.flags\nflags.DEFINE_string('csv_input', '', 'Path to the CSV input')\nflags.DEFINE_string('output_path', '', 'Path to output TFRecord')\nflags.DEFINE_string('img_path', '', 'Path to images')\nflags.DEFINE_string('label_map', '', 'Path to label map (.pbtxt) file')\n\n# if your image has more labels input them as\n# flags.DEFINE_string('label0', '', 'Name of class[0] label')\n# flags.DEFINE_string('label1', '', 'Name of class[1] label')\n# and so on.\n\nFLAGS = flags.FLAGS\n\n\ndef class_text_to_int(row_label):\n    for label_id, label_name in get_label_info():\n        if row_label == label_name:\n            return label_id\n    return 0\n\n\ndef get_label_info():\n    \"\"\"\n    Generate label info from label map (.pbtxt) file\n    :return: id, name\n    \"\"\"\n    label_info = []\n    with open(FLAGS.label_map) as fp:\n        for _, line in enumerate(fp):\n            if \"id\" in line:\n                label_id = int(line.split(\":\")[1])\n                label_info.append(label_id)\n            elif \"name\" in line:\n                label_name = line.split(\":\")[1].strip().replace(\"'\", \"\")\n                label_info.append(label_name)\n\n    for i in range(0, len(label_info), 2):\n        yield label_info[i:i + 2]\n\n\ndef split(df, group):\n    data = namedtuple('data', ['filename', 'object'])\n    gb = df.groupby(group)\n    return [data(filename, gb.get_group(x)) for filename, x in zip(gb.groups.keys(), gb.groups)]\n\n\ndef create_tf_example(group, path):\n    with tf.gfile.GFile(os.path.join(path, '{}'.format(group.filename)), 'rb') as fid:\n        encoded_jpg = fid.read()\n\n    encoded_jpg_io = io.BytesIO(encoded_jpg)\n    image = Image.open(encoded_jpg_io)\n    width, height = image.size\n    filename = group.filename.encode('utf8')\n    image_format = b'jpg'\n\n    # check if the image format is matching with your images.\n    xmins = []\n    xmaxs = []\n    ymins = []\n    ymaxs = []\n    classes_text = []\n    classes = []\n\n    for index, row in group.object.iterrows():\n        xmins.append(row['xmin'] \/ width)\n        xmaxs.append(row['xmax'] \/ width)\n        ymins.append(row['ymin'] \/ height)\n        ymaxs.append(row['ymax'] \/ height)\n        classes_text.append(row['class'].encode('utf8'))\n        classes.append(class_text_to_int(row['class']))\n\n    tf_example = tf.train.Example(features=tf.train.Features(feature={\n        'image\/height': dataset_util.int64_feature(height),\n        'image\/width': dataset_util.int64_feature(width),\n        'image\/filename': dataset_util.bytes_feature(filename),\n        'image\/source_id': dataset_util.bytes_feature(filename),\n        'image\/encoded': dataset_util.bytes_feature(encoded_jpg),\n        'image\/format': dataset_util.bytes_feature(image_format),\n        'image\/object\/bbox\/xmin': dataset_util.float_list_feature(xmins),\n        'image\/object\/bbox\/xmax': dataset_util.float_list_feature(xmaxs),\n        'image\/object\/bbox\/ymin': dataset_util.float_list_feature(ymins),\n        'image\/object\/bbox\/ymax': dataset_util.float_list_feature(ymaxs),\n        'image\/object\/class\/text': dataset_util.bytes_list_feature(classes_text),\n        'image\/object\/class\/label': dataset_util.int64_list_feature(classes),\n    }))\n    return tf_example\n\n\ndef main(_):\n    writer = tf.python_io.TFRecordWriter(FLAGS.output_path)\n    path = os.path.join(os.getcwd(), FLAGS.img_path)\n    examples = pd.read_csv(FLAGS.csv_input)\n    grouped = split(examples, 'filename')\n\n    for group in grouped:\n        tf_example = create_tf_example(group, path)\n        writer.write(tf_example.SerializeToString())\n\n    writer.close()\n    output_path = os.path.join(os.getcwd(), FLAGS.output_path)\n    print('Successfully created the TFRecords: {}'.format(output_path))\n\n\nif __name__ == '__main__':\n    tf.app.run()\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":52660985,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2018-10-05T08:01:33.573","AcceptedAnswerId":52671057.0,"OwnerUserId":10020793.0,"Title":"PyTorch - How to get learning rate during training?","Body":"<p>While training, I'd like to know the value of learning_rate.\nWhat should I do?<\/p>\n\n<p>It's my code, like this:<\/p>\n\n<pre><code>my_optimizer = torch.optim.SGD(my_model.parameters(), \n                               lr=0.001, \n                               momentum=0.99, \n                               weight_decay=2e-3)\n<\/code><\/pre>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"52671057","CreationDate":"2018-10-05T18:00:23.650","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>For only one parameter group like in the example you've given, you can use this function and call it during training to get the current learning rate:<\/p>\n\n<pre><code>def get_lr(optimizer):\n    for param_group in optimizer.param_groups:\n        return param_group['lr']\n<\/code><\/pre>\n"}]}
+{"QuestionId":52661518,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-05T08:32:50.187","AcceptedAnswerId":52678197.0,"OwnerUserId":1703586.0,"Title":"Keras Custom Loss","Body":"<p>I am kinda new to keras. I managed to build a network which has two outputs:<\/p>\n\n<pre><code>q_dot_P : &lt;tf.Tensor 'concatenate_1\/concat:0' shape=(?, 7) dtype=float32&gt;\nq_dot_N : &lt;tf.Tensor 'concatenate_2\/concat:0' shape=(?, 10) dtype=float32&gt;\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HhJB1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HhJB1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I wish to compute the above expression, q_dot_P is \\delta^{q}_P and q_dot_N is \\delta^{q}_P. <\/p>\n\n<p>Here is my attempt: <\/p>\n\n<pre><code>nN = 10 \nnP = 7    \n__a = keras.layers.RepeatVector(nN)( q_dot_P ) #OK, same as 1 . q_dot_P\n__b = keras.layers.RepeatVector(nP)( q_dot_N ) #OK, same as 1 . q_dot_N\nminu = keras.layers.Subtract()( [keras.layers.Permute( (2,1) )( __b ), __a ] )   \nminu = keras.layers.Lambda( lambda x: x + 0.1)( minu )      \nminu = keras.layers.Maximum()( [ minu, K.zeros(nN, nP) ] ) #this fails\n<\/code><\/pre>\n\n<p>The <code>keras.layers.Maximum()<\/code> fails.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"noveou_train_netvlad.py\", line 226, in &lt;module&gt;\n    minu = keras.layers.Maximum()( [ minu, K.zeros(nN, nP) ] )\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/base_layer.py\", line 457, in __call__\n    output = self.call(inputs, **kwargs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/layers\/merge.py\", line 115, in call\n    return self._merge_function(reshaped_inputs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/layers\/merge.py\", line 301, in _merge_function\n    output = K.maximum(output, inputs[i])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 1672, in maximum\n    return tf.maximum(x, y)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/ops\/gen_math_ops.py\", line 4707, in maximum\n    \"Maximum\", x=x, y=y, name=name)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 546, in _apply_op_helper\n    inferred_from[input_arg.type_attr]))\nTypeError: Input 'y' of 'Maximum' Op has type string that does not match type float32 of argument 'x'.\n<\/code><\/pre>\n\n<p>What is the simplest way to achieve this objective? <\/p>\n\n<hr>\n\n<p>After following the suggestion from @rvinas<\/p>\n\n<p>I have a time distributed model in keras. See <a href=\"https:\/\/stackoverflow.com\/questions\/52686173\/keras-timedistributed-layer-without-lstm\/52686507#52686507\">Keras TimeDistributed layer without LSTM<\/a> <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/B9WUe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/B9WUe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n    nP = 2\n    nN = 2\n\n    # y_pred.shape = shape=(?, 5, 512)\n    q = y_pred[:,0:1,:]    # shape=(?, 1, 512)\n    P = y_pred[:,1:1+nP,:] # shape=(?, 2, 512)\n    N = y_pred[:,1+nP:,:]  # shape=(?, 2, 512)\n    q_dot_P = keras.layers.dot( [q,P], axes=-1 )  # shape=(?, 1, 2)\n    q_dot_N = keras.layers.dot( [q,N], axes=-1 )  # shape=(?, 1, 2)\n\n    epsilon = 0.1  # Your epsilon here\n\n    zeros = K.zeros((nP, nN), dtype='float32')\n    ones_m = K.ones(nP, dtype='float32')\n    ones_n = K.ones(nN, dtype='float32')\n    code.interact( local=locals() , banner='custom_loss')\n    aux = ones_m[None, :, None] * q_dot_N[:, None, :] \\\n          - q_dot_P[:, :, None] * ones_n[None, None, :] \\\n          + epsilon * ones_m[:, None] * ones_n[None, :]\n\n    return K.maximum(zeros, aux)\n<\/code><\/pre>\n\n<p>Here is the main: <\/p>\n\n<pre><code># In __main__\n#---------------------------------------------------------------------------\n# Setting Up core computation\n#---------------------------------------------------------------------------\ninput_img = Input( shape=(image_nrows, image_ncols, image_nchnl ) )\ncnn = make_vgg( input_img )    \nout = NetVLADLayer(num_clusters = 16)( cnn )\nmodel = Model( inputs=input_img, outputs=out )\n\n#--------------------------------------------------------------------------\n# TimeDistributed\n#--------------------------------------------------------------------------\nt_input = Input( shape=(1+nP+nN, image_nrows, image_ncols, image_nchnl ) )\nt_out = TimeDistributed( model )( t_input )\nt_model = Model( inputs=t_input, outputs=t_out )\n\nt_model.compile( loss=custom_loss, optimizer='sgd' )\n<\/code><\/pre>\n","answers":[{"AnswerId":"52678197","CreationDate":"2018-10-06T10:38:58.793","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You could define your loss function as follows:<\/p>\n\n<pre><code>import keras.backend as K\n\nnN = 10\nnP = 7\n\n\ndef custom_loss(y_true, y_pred):\n    q_dot_P = ...  # Extract q_dot_P from y_pred\n    q_dot_N = ...  # Extract q_dot_N from y_pred\n    epsilon = ...  #\u00a0Your epsilon here\n\n    zeros = K.zeros((nP, nN), dtype='float32')\n    ones_m = K.ones(nP, dtype='float32')\n    ones_n = K.ones(nN, dtype='float32')\n    aux = ones_m[None, :, None] * q_dot_N[:, None, :] \\\n          - q_dot_P[:, :, None] * ones_n[None, None, :] \\\n          + epsilon * ones_m[:, None] * ones_n[None, :]\n\n    return K.maximum(zeros, aux)\n<\/code><\/pre>\n\n<p>and pass this function to <a href=\"https:\/\/keras.io\/models\/model\/#compile\" rel=\"nofollow noreferrer\">model.compile()<\/a>.<\/p>\n\n<p><strong>NOTE<\/strong>: Not tested.<\/p>\n"}]}
+{"QuestionId":52662153,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-05T09:11:04.913","AcceptedAnswerId":52700181.0,"OwnerUserId":7819329.0,"Title":"How to make a generic function that adapts to the model?","Body":"<p>I am developing a CNN network with Keras<\/p>\n\n<p>I have a function that I defined where I call <code>preprocess_input<\/code> function of Keras models (VGG16, VGG19,...) <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>So I am doing something like <\/p>\n\n<pre><code>preprocess_input (inputs)\n<\/code><\/pre>\n\n<p>However, I want to test several models, resulting in different <code>preprocess_input<\/code> in my function. How can I do to say in my homemade function that I want to use the <code>preprocess_input<\/code> corresponding to the model I am testing ?<\/p>\n\n<p>I thought about giving the name of my model as a parameter of my function but then, I don't know how to transform this string into a way that allows to call<\/p>\n\n<pre><code>modelThatIwant.preprocess_input(...)\n<\/code><\/pre>\n\n<p>Thanks for help<\/p>\n","answers":[{"AnswerId":"52700181","CreationDate":"2018-10-08T10:22:15.480","ParentId":null,"OwnerUserId":"7819329","Title":null,"Body":"<p>Actually, the answer is quite easy : you just need to add an argument <code>fct<\/code> in the homemade function you created and then, pass the adequate <code>preprocessing_inputs<\/code> corresponding to the model you want to train... Thanks to @Matias Valdenegro<\/p>\n"}]}
+{"QuestionId":52662197,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-05T09:13:25.553","AcceptedAnswerId":null,"OwnerUserId":1019455.0,"Title":"PyTorch derivative has no degree","Body":"<p>I am studying the tutorial from <code>PyTorch<\/code> official <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/autograd_tutorial.html\" rel=\"nofollow noreferrer\">docs<\/a>. I am trying to understand the content. Start from <code>You can do many crazy things with autograd!<\/code><\/p>\n\n<pre><code>x = torch.randn(3, requires_grad=True)\ny = x * 2\n\ni = 0\nwhile y.data.norm() &lt; 100:\n    y = y * 2\n    i+= 1\n\nprint(x)\nprint(y)\nprint(i)\n<\/code><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<pre><code>tensor([-0.6933,  0.1126,  0.3913], requires_grad=True)\ntensor([-88.7455,  14.4082,  50.0871], grad_fn=&lt;MulBackward&gt;)\n6\n<\/code><\/pre>\n\n<p>Find the derivative w.r.t to <code>x<\/code> at point <code>[0.1, 1.0, 0.0001]<\/code><\/p>\n\n<pre><code>gradients = torch.tensor([0.1, 1.0, 0.0001], dtype=torch.float)\ny.backward(gradients)\nprint(x.grad)\n<\/code><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<pre><code>tensor([ 12.8000, 128.0000,   0.0128])\n<\/code><\/pre>\n\n<p>In my understanding <code>i<\/code> equals to <code>6<\/code>. Then <code>y = (2x)^7<\/code> and derivative is different from <code>PyTorch<\/code>. It has <code>7<\/code> as a factor when substitute the value\nto my derivative.<\/p>\n\n<p>The answer from the <code>PyTorch<\/code> is simply substitute <code>x<\/code> with given point to the <code>dy\/dx = 2^7 * x<\/code><\/p>\n\n<p><strong>Question:<\/strong><\/p>\n\n<p>How to derive the derivative?<\/p>\n\n<p><strong>References:<\/strong><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/51500010\/how-to-use-pytorch-to-calculate-partial-derivatives\">How to use PyTorch to calculate partial derivatives?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/51054627\/pytorch-autograd-automatic-differentiation-feature\/51055391#51055391\">PyTorch Autograd automatic differentiation feature<\/a><\/p>\n","answers":[{"AnswerId":"52668874","CreationDate":"2018-10-05T15:24:49.593","ParentId":null,"OwnerUserId":"8715323","Title":null,"Body":"<p>If you look closely at the expressions, it'd turn out that <code>y = x * (2^7)<\/code>, the derivative of which is <code>2^7 * x<\/code> .<\/p>\n"}]}
+{"QuestionId":52662450,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-05T09:26:32.067","AcceptedAnswerId":null,"OwnerUserId":8208658.0,"Title":"Delete a single row(Free memory on GPU) from a 2D tensorflow variable","Body":"<p>I have a NxM size tensorflow variable stored in GPU. At some point, I would like to delete some rows i.e. free the memory occupied by them in GPU. \nI don't want to release all memory resources by doing a <code>sess.close()<\/code> or deleting the session. \nIs this possible in tensorflow to just pick random row from the variable and completely erase it? I have read <a href=\"https:\/\/stackoverflow.com\/questions\/42631417\/how-to-release-tensors-from-memory-when-not-in-need-anymore\">this<\/a> answer about releasing memory but its for a tensorflow op.<\/p>\n","answers":[]}
+{"QuestionId":52662479,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-05T09:28:31.703","AcceptedAnswerId":null,"OwnerUserId":10454025.0,"Title":"Py charm freeze graph not converted 8 variable to constant ops","Body":"<pre><code>import tensorflow as tf\n\nfrom tensorflow.python.tools import freeze_graph,optimize_for_inference_lib\n\n# Saving the graph as a pb file taking data from pbtxt and ckpt files and providing a few operations\nfreeze_graph.freeze_graph('advanced_mnist.pbtxt',\n                          '',\n                          True,\n                          'advanced_mnist.ckpt',\n                          'h_readout1',\n                          'save\/restore_all',\n                          'save\/Const:0',\n                          'frozen_advance_mnist.pb'\n                          ,True,\n                          '')\n# Read the data form the frozen graph pb file\n\ninput_graph_def = tf.GraphDef()\nwith tf.gfile.Open('frozen_advance_mnist.pb','rb') as f:\n    data = f.read()\n    input_graph_def.ParseFromString(data)\n# Optimize the graph with input and output nodes\noutput_graph_def = optimize_for_inference_lib.optimize_for_inference(input_graph_def,['x_input','keep_prob'],['h_readout1'],tf.float32.as_datatype_enum)\n# Save the optimized graph to the optimized pb file\nf = tf.gfile.FastGFile('optimized_advanced_mnist.pb','w')\nf.write(output_graph_def.SerializeToString())\n<\/code><\/pre>\n\n<p>I try run the input and output but the file not convert the variables into any constant ops instead just process finished with exit code 0.<\/p>\n","answers":[]}
+{"QuestionId":52662727,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-05T09:42:53.067","AcceptedAnswerId":52665472.0,"OwnerUserId":9316652.0,"Title":"Tensor shape modification using slicing and None","Body":"<p>I am bit puzzled by how to read and understand a simple line of code:<\/p>\n\n<ul>\n<li>I have a tensor <code>input<\/code> of shape <code>(19,4,64,64,3)<\/code>.<\/li>\n<li>The line of code <code>input[:, None]<\/code> returns a tensor of shape <code>(19, 1,\n4, 64, 64, 3)<\/code>.<\/li>\n<\/ul>\n\n<p>How should I understand the behavior of that line? It seems that <code>None<\/code> is adding a dimension, with a size of 1. But why is this added on that specific position (between 19 and 4)?<\/p>\n","answers":[{"AnswerId":"52665472","CreationDate":"2018-10-05T12:18:24.137","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>Indeed, <code>None<\/code> adds a new dimension. You can also use <code>tf.newaxis<\/code> for this which is a bit more explicit IMHO. <\/p>\n\n<p>The new dimension is added in axis 1 because that's where it appears in the index. E.g. <code>input[:, :, None]<\/code> should result in shape (19, 4, 1, 64, 64, 3) and so on.<br>\nIt might get clearer if we write all the dimensions in the slicing: <code>input[:, None, :, :, :, :]<\/code>. In slicing, <code>:<\/code> simply means taking all elements of the dimension. So by using one <code>:<\/code>, we take all elements of dimension 0 and then \"move on\" to dimension 1. Since <code>None<\/code> appears here, we know that the new size-1 axis should be in dimension 1. Accordingly, the remaining dimensions get \"pushed back\".<\/p>\n"}]}
+{"QuestionId":52662999,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-05T09:57:47.180","AcceptedAnswerId":52678189.0,"OwnerUserId":4713057.0,"Title":"tf.cond : adding an element to a list","Body":"<p>I need a conditional control flow in a very simple graph which enables to add an element to a list. In my case, the list should be declared using the <code>tf.Variable([])<\/code>.<\/p>\n\n<p>I am getting a weird error : <code>UnboundLocalError: local variable 'list1' referenced before assignment<\/code>.<\/p>\n\n<p>Here is a toy example:<\/p>\n\n<pre><code>pred = tf.placeholder(tf.bool, shape=[])\nlist1 = tf.Variable([])\n\ndef f1():\n  e1 = tf.constant(1.0)\n  list1 = tf.concat([list1, [e1]], 0)\n\ndef f2():\n  e2 = tf.constant(2.0)\n  list1 = tf.concat([list1, [e2]], 0)\n\ny = tf.cond(pred, f1, f2)\nwith tf.Session() as session:\n  session.run(tf.global_variables_initializer())\n  print(y.eval(feed_dict={pred: False}))  # ==&gt; [1]\n  print(y.eval(feed_dict={pred: True}))   # ==&gt; [2]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52678189","CreationDate":"2018-10-06T10:38:30.743","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_variable\" rel=\"nofollow noreferrer\">tf.get_variable<\/a> Gets an existing variable with these parameters or creates a new one. <\/p>\n\n<p>So this pattern avoids the problem that you face.<\/p>\n\n<p>In addition to that I had to use this line after switching off shape validation as that caused validation errors.<\/p>\n\n<pre><code>       list1 = tf.assign( list1, tf.concat([list1, [e1]], 0), validate_shape=False)\n<\/code><\/pre>\n\n<p>The working code is this. <\/p>\n\n<pre><code>   def f1():\n        with tf.variable_scope(\"reuse\", reuse=tf.AUTO_REUSE):\n            list1 = tf.get_variable(initializer=[], dtype=tf.float32, name='list1')\n            e1 = tf.constant(1.)\n            print(tf.shape(list1))\n            list1 = tf.assign( list1, tf.concat([list1, [e1]], 0), validate_shape=False)\n            return list1\n\n\n    def f2():\n        with tf.variable_scope(\"reuse\", reuse=tf.AUTO_REUSE):\n            list1 = tf.get_variable(  dtype=tf.float32, name='list1')\n            e2 = tf.constant(2.)\n            list1 = tf.assign( list1, tf.concat([list1, [e2]], 0), validate_shape=False)\n            return list1\n\n\n    y = tf.cond(pred, f1,  f2)\n\n    with tf.Session() as session:\n\n      with tf.variable_scope(\"reuse\", reuse=tf.AUTO_REUSE):\n          session.run(tf.global_variables_initializer())\n          print(session.run([y], feed_dict={pred: False}))  # ==&gt; [1]\n          print(session.run([y], feed_dict={pred: True}))  # ==&gt; [2]\n\n          # Gets the updated variable\n          list1 = tf.get_variable(dtype=tf.float32, name='list1')\n          print(session.run(list1))\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>[array([2.], dtype=float32)]\n[array([2., 1.], dtype=float32)]\n[2. 1.]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52663313,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-05T10:13:49.300","AcceptedAnswerId":55414203.0,"OwnerUserId":3549680.0,"Title":"Use tensor to set a specific seed?","Body":"<p>I would like to build graph with some random values in it. Then I would like to evaluate the graph with different seeds. Here is an example what I'm trying to achieve:<\/p>\n\n<pre><code>import tensorflow as tf\nseed = tf.placeholder(dtype=tf.int32, shape=(), name=\"seed\")\nrandoms = tf.random_normal(shape=[8], seed=seed)\n<\/code><\/pre>\n\n<p>Then I was hoping to do something like this where first two calls would return same result:<\/p>\n\n<pre><code>sess = tf.InteractiveSession()\nsess.run(randoms, {seed: 1})\nsess.run(randoms, {seed: 1})\nsess.run(randoms, {seed: 2})\n<\/code><\/pre>\n\n<p>Is there any way to work around this problem?<\/p>\n","answers":[{"AnswerId":"55414203","CreationDate":"2019-03-29T09:25:58.550","ParentId":null,"OwnerUserId":"3549680","Title":null,"Body":"<p>This can be done with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random\/stateless_normal\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random\/stateless_normal<\/a><\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nseed = tf.placeholder(dtype=tf.int32, shape=[2], name=\"seed\")\nrandoms = tf.random.stateless_normal(shape=[3], seed=seed)\n\nsess = tf.InteractiveSession()\nprint(sess.run(randoms, {seed: [1, 0]}))\nprint(sess.run(randoms, {seed: [1, 0]}))\nprint(sess.run(randoms, {seed: [2, 0]}))\n<\/code><\/pre>\n\n<pre><code>[ 0.1266503  -0.49301657  0.6311907 ]\n[ 0.1266503  -0.49301657  0.6311907 ]\n[-0.6394294  -0.18700573 -0.82845527]\n<\/code><\/pre>\n\n<hr>\n\n<p>However, warning-note from documentation:<\/p>\n\n<blockquote>\n  <p>The output is consistent across multiple runs on the same hardware (and between CPU and GPU), but may change between versions of TensorFlow or on non-CPU\/GPU hardware.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":52663358,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-05T10:16:20.327","AcceptedAnswerId":null,"OwnerUserId":9246640.0,"Title":"Problems about torch.nn.DataParallel","Body":"<p>I am new in deep learning area. Now I am reproducing a paper\u2019s codes. since they use several GPUs, there is a command <code>torch.nn.DataParallel(model, device_ids= args.gpus).cuda()<\/code> in codes. But I only have one GPU, what\nshould I change this code to match up my GPU?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"52664209","CreationDate":"2018-10-05T11:04:17.187","ParentId":null,"OwnerUserId":"7752110","Title":null,"Body":"<p><code>DataParallel<\/code> should work on a single GPU as well, but you should check if <code>args.gpus<\/code> only contains the id of the device that is to be used (should be 0) or <code>None<\/code>.\nChoosing <code>None<\/code> will make the module use all available devices.<\/p>\n\n<p>Also you could remove <code>DataParallel<\/code> as you do not need it and move the model to GPU only by calling <code>model.cuda()<\/code> or, as I prefer, <code>model.to(device)<\/code> where <code>device<\/code> is the device's name.<\/p>\n\n<p><strong>Example:<\/strong><\/p>\n\n<p>This example shows how to use a model on a single GPU, setting the device using <code>.to()<\/code> instead of <code>.cuda()<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch import nn\nimport torch\n\n# Set device to cuda if cuda is available\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\n# Create model\nmodel = nn.Sequential(\n  nn.Conv2d(1,20,5),\n  nn.ReLU(),\n  nn.Conv2d(20,64,5),\n  nn.ReLU()\n)\n\n# moving model to GPU\nmodel.to(device)\n<\/code><\/pre>\n\n<p>If you want to use <code>DataParallel<\/code> you could do it like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Optional DataParallel, not needed for single GPU usage\nmodel1 = torch.nn.DataParallel(model, device_ids=[0]).to(device)\n# Or, using default 'device_ids=None'\nmodel1 = torch.nn.DataParallel(model).to(device)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52663701,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-05T10:36:21.513","AcceptedAnswerId":null,"OwnerUserId":10414305.0,"Title":"ValueError: Only call `sigmoid_cross_entropy_with_logits` with named arguments (labels=..., logits=..., ...)","Body":"<p>I have just build my first model using Phyton, Keras and Tensorflow. <\/p>\n\n<p>My model looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2D(32, 3, 3, input_shape=(img_width, img_height,3)))    \nmodel.add(Activation('relu'))   \nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Convolution2D(32, 3, 3)) \nmodel.add(Activation('relu'))    \nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Convolution2D(64, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))   \n\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(1))    \nmodel.add(Activation('sigmoid'))\n<\/code><\/pre>\n\n<p>Now i'm trying to compile it, <\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy',\n\n              optimizer='rmsprop',\n\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>but i got this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Only call sigmoid_cross_entropy_with_logits with named\n  arguments (labels=..., logits=..., ...)<\/p>\n<\/blockquote>\n\n<p>I can't seem to find what the problem is. Please let me know what I'm doing wrong..<\/p>\n","answers":[]}
+{"QuestionId":52663976,"AnswerCount":1,"Tags":"<python><macos><hadoop><tensorflow><hdfs>","CreationDate":"2018-10-05T10:52:02.550","AcceptedAnswerId":null,"OwnerUserId":3534616.0,"Title":"Tensorflow Read from HDFS mac : java.lang.NoSuchFieldError: LOG","Body":"<p>I am trying to read from external hadoop from tensorflow on my mac. I have built tf with hadoop support from source, and also build hadoop with native library support on my mac. I am getting the following error ,<\/p>\n\n<pre><code>hdfsBuilderConnect(forceNewInstance=0, nn=192.168.60.53:9000, port=0, kerbTicketCachePath=(NULL), userName=(NULL)) error:\njava.lang.NoSuchFieldError: LOG\n    at org.apache.hadoop.ipc.ClientCache.getClient(ClientCache.java:62)\n    at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.&lt;init&gt;(ProtobufRpcEngine.java:145)\n    at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.&lt;init&gt;(ProtobufRpcEngine.java:133)\n    at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.&lt;init&gt;(ProtobufRpcEngine.java:119)\n    at org.apache.hadoop.ipc.ProtobufRpcEngine.getProxy(ProtobufRpcEngine.java:102)\n    at org.apache.hadoop.ipc.RPC.getProtocolProxy(RPC.java:579)\n    at org.apache.hadoop.hdfs.NameNodeProxies.createNNProxyWithClientProtocol(NameNodeProxies.java:418)\n    at org.apache.hadoop.hdfs.NameNodeProxies.createNonHAProxy(NameNodeProxies.java:314)\n    at org.apache.hadoop.hdfs.NameNodeProxies.createProxy(NameNodeProxies.java:176)\n    at org.apache.hadoop.hdfs.DFSClient.&lt;init&gt;(DFSClient.java:678)\n    at org.apache.hadoop.hdfs.DFSClient.&lt;init&gt;(DFSClient.java:619)\n    at org.apache.hadoop.hdfs.DistributedFileSystem.initialize(DistributedFileSystem.java:149)\n    at org.apache.hadoop.fs.FileSystem.createFileSystem(FileSystem.java:2669)\n    at org.apache.hadoop.fs.FileSystem.access$200(FileSystem.java:94)\n    at org.apache.hadoop.fs.FileSystem$Cache.getInternal(FileSystem.java:2703)\n    at org.apache.hadoop.fs.FileSystem$Cache.get(FileSystem.java:2685)\n    at org.apache.hadoop.fs.FileSystem.get(FileSystem.java:373)\n    at org.apache.hadoop.fs.FileSystem$1.run(FileSystem.java:162)\n    at org.apache.hadoop.fs.FileSystem$1.run(FileSystem.java:159)\n    at java.security.AccessController.doPrivileged(Native Method)\n    at javax.security.auth.Subject.doAs(Subject.java:422)\n    at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1698)\n    at org.apache.hadoop.fs.FileSystem.get(FileSystem.java:159)\n2018-10-05 16:01:21.867554: W tensorflow\/core\/kernels\/queue_base.cc:277] _0_input_producer: Skipping cancelled enqueue attempt with queue not closed\nTraceback (most recent call last):\n<\/code><\/pre>\n\n<p>This is my code:<\/p>\n\n<pre><code>import tensorflow as tf\n\n\ndef create_file_reader_ops(filename_queue):\n    reader = tf.TextLineReader(skip_header_lines=1)\n    _, csv_row = reader.read(filename_queue)\n    record_defaults = [[\"\"], [\"\"], [0], [0]]\n    country, code, gold, silver = tf.decode_csv(csv_row, record_defaults=record_defaults)\n    features = tf.stack([gold, silver])\n    return features, country\n\n\nfilename_queue = tf.train.string_input_producer([\n    \"hdfs:\/\/192.168.60.53:9000\/iris_data_multiclass.csv\",\n])\nexample, country = create_file_reader_ops(filename_queue)\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n    while True:\n        try:\n            example_data, country_name = sess.run([example, country])\n            print(example_data, country_name)\n        except tf.errors.OutOfRangeError:\n            break\n<\/code><\/pre>\n\n<p>I have build hadoop from source on mac. <\/p>\n\n<pre><code>$ hadoop version\nHadoop 2.7.3\nSubversion https:\/\/github.com\/apache\/hadoop.git -r baa91f7c6bc9cb92be5982de4719c1c8af91ccff\nCompiled by himaprasoon on 2018-10-04T11:09Z\nCompiled with protoc 2.5.0\nFrom source with checksum 2e4ce5f957ea4db193bce3734ff29ff4\nThis command was run using \/Users\/himaprasoon\/git\/hadoop\/hadoop-dist\/target\/hadoop-2.7.3\/share\/hadoop\/common\/hadoop-common-2.7.3.jar\n<\/code><\/pre>\n\n<p><strong>hadoop checknative output<\/strong><\/p>\n\n<pre><code>$ hadoop checknative\n18\/10\/05 16:15:05 INFO bzip2.Bzip2Factory: Successfully loaded &amp; initialized native-bzip2 library libbz2.dylib\n18\/10\/05 16:15:05 INFO zlib.ZlibFactory: Successfully loaded &amp; initialized native-zlib library\nNative library checking:\nhadoop:  true \/Users\/himaprasoon\/git\/hadoop\/hadoop-dist\/target\/hadoop-2.7.3\/lib\/native\/libhadoop.dylib\nzlib:    true \/usr\/lib\/libz.1.dylib\nsnappy:  true \/usr\/local\/lib\/libsnappy.1.dylib\nlz4:     true revision:99\nbzip2:   true \/usr\/lib\/libbz2.1.0.dylib\nopenssl: true \/usr\/local\/lib\/libcrypto.dylib\n<\/code><\/pre>\n\n<p>tf version : <strong>1.10.1<\/strong>\nAny ideas what I might be doing wrong?<\/p>\n\n<p>here are my <strong>environment variables<\/strong>.<\/p>\n\n<pre><code>HADOOP_HOME=\/Users\/himaprasoon\/git\/hadoop\/hadoop-dist\/target\/hadoop-2.7.3\/\nHADOOP_MAPRED_HOME=$HADOOP_HOME\nHADOOP_COMMON_HOME=$HADOOP_HOME\nHADOOP_HDFS_HOME=$HADOOP_HOME\nYARN_HOME=$HADOOP_HOME\nHADOOP_COMMON_LIB_NATIVE_DIR=$HADOOP_HOME\/lib\/native\nPATH=$PATH:$HADOOP_HOME\/sbin:$HADOOP_HOME\/bin\nHADOOP_INSTALL=$HADOOP_HOME\nOPENSSL_ROOT_DIR=\"\/usr\/local\/opt\/openssl\"\nLDFLAGS=\"-L${OPENSSL_ROOT_DIR}\/lib\"\nCPPFLAGS=\"-I${OPENSSL_ROOT_DIR}\/include\"\nPKG_CONFIG_PATH=\"${OPENSSL_ROOT_DIR}\/lib\/pkgconfig\"\nOPENSSL_INCLUDE_DIR=\"${OPENSSL_ROOT_DIR}\/include\"\nPATH=\"\/usr\/local\/opt\/protobuf@2.5\/bin:$PATH\nHADOOP_OPTS=\"-Djava.library.path=${HADOOP_HOME}\/lib\/native\"\nLD_LIBRARY_PATH=$LD_LIBRARY_PATH:${HADOOP_HOME}\/lib\/native\nJAVA_LIBRARY_PATH=$JAVA_LIBRARY_PATH:${HADOOP_HOME}\/lib\/native\n<\/code><\/pre>\n\n<p>this is how I am running my program<\/p>\n\n<pre><code>CLASSPATH=$($HADOOP_HDFS_HOME\/bin\/hdfs classpath --glob) python3.6 myfile.py\n<\/code><\/pre>\n\n<p>references used to build tf and hadoop<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/37800605\/hadoop-native-libraries-not-found-on-os-x\">Hadoop native libraries not found on OS\/X<\/a><\/li>\n<li><a href=\"https:\/\/medium.com\/@s.matthew.english\/build-hadoop-from-source-on-macos-a3fb2b958b6c\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@s.matthew.english\/build-hadoop-from-source-on-macos-a3fb2b958b6c<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/42614426\/can-tensorflow-read-from-hdfs-on-mac\">Can Tensorflow read from HDFS on Mac?<\/a><\/li>\n<li><a href=\"https:\/\/gist.github.com\/zedar\/f631ace0759c1d512573\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/zedar\/f631ace0759c1d512573<\/a><\/li>\n<\/ul>\n","answers":[{"AnswerId":"52807801","CreationDate":"2018-10-14T22:59:12.967","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>Have you read this post?<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/stackoverflow.com\/questions\/38678371\/tensorflow-enqueue-operation-was-cancelled\">Tensorflow Enqueue operation was cancelled<\/a><\/p>\n<\/blockquote>\n\n<p>It seems there is a workaround for the same error message there:<\/p>\n\n<blockquote>\n  <p>The problem happens at the very last stage when python tries to kill threads. \n  To do this properly you should create a train.Coordinator and pass it to your \n  queue_runner (no need to pass sess, as default session will be used> <\/p>\n<\/blockquote>\n\n<pre><code>with tf.Session() as sess:\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n    \/\/ do your things\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>The last two lines should be added to your while loop to make sure all threads are properly killed.<\/p>\n"}]}
+{"QuestionId":52663982,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><conv-neural-network><tensorflow-datasets>","CreationDate":"2018-10-05T10:52:08.763","AcceptedAnswerId":null,"OwnerUserId":3176696.0,"Title":"Convolution NN architecture, shaping error","Body":"<p>I've been trying to get a CNN working for the <a href=\"https:\/\/github.com\/brendenlake\/omniglot\" rel=\"nofollow noreferrer\">Omligot dataset<\/a> (105 x 105 x 1 images) via two tutorials I found: <a href=\"https:\/\/pythonprogramming.net\/cnn-tensorflow-convolutional-nerual-network-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">CNN tutorial 1<\/a> and <a href=\"https:\/\/www.datacamp.com\/community\/tutorials\/cnn-tensorflow-python\" rel=\"nofollow noreferrer\">CNN tutorial 2<\/a>, working on the usual MNIST dataset (28 x 28 x 1 images). <\/p>\n\n<p>I'm still struggling with a shaping conflict in the implementation (after a week with on- &amp; off time for debugging) No one was able to provide any help thus far, and in the mean time I could debug in a way, where I think I can provide better description of the shaping error.<\/p>\n\n<p>The majority of my code is as follows (just skipping few irrelevant stuff here and there). So my placeholders are defined as follows right:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, 105, 105, 1) )  # placeholder for train data\ny = tf.placeholder(tf.float32, shape=(None, 20) )           # placeholder for labels\nlr = tf.placeholder(tf.float32,shape=(), name=\"learnRate\")  # for varying learning rates during training\n<\/code><\/pre>\n\n<p>label dimension of 20 given that there's 20 different characters per alphabet. So one-hot-encoded vector of length 20.<\/p>\n\n<p>From here on out my model is defined as follow (where I commented the dimension results from each output):<\/p>\n\n<pre><code># weight and bias dimension definitions\nself.ConFltSize = 3\nself.ConOutSize = 7\nself.weights = {\n    'wc1': tf.Variable(tf.random_normal([self.ConFltSize,self.ConFltSize,1, 32], stddev=0.01, name='W0')), \n    'wc2': tf.Variable(tf.random_normal([self.ConFltSize,self.ConFltSize,32, 64], stddev=0.01, name='W1')), \n    'wc3': tf.Variable(tf.random_normal([self.ConFltSize,self.ConFltSize,64, 128], stddev=0.01, name='W2')), \n    'wd1': tf.Variable(tf.random_normal([self.ConOutSize * self.ConOutSize * 128, 128], stddev=0.01, name='W3')), \n    'out': tf.Variable(tf.random_normal([128, self.InLabels.shape[1]], stddev=0.01, name='W4')), \n    }\n\nself.biases = {\n    'bc1': tf.Variable(tf.random_normal([32], stddev=0.01, name='B0')), \n    'bc2': tf.Variable(tf.random_normal([64], stddev=0.01, name='B1')), \n    'bc3': tf.Variable(tf.random_normal([128], stddev=0.01, name='B2')),\n    'bd1': tf.Variable(tf.random_normal([128], stddev=0.01, name='B3')),\n    'out': tf.Variable(tf.random_normal([self.InLabels.shape[1]], stddev=0.01, name='B4')),\n    }\n\n# Model definition + shaping results\n# x = provide the input data\n# weights = dictionary variables for weights\n# biases = dictionary variables for biases\ndef Architecture(self, x, weights, biases): \n\n    conv1 = self.conv(x, weights['wc1'], biases['bc1'])      # convolution layer 1\n    conv1 = self.maxPool(conv1)                              # max pool layer 1\n    # out shape -&gt; [None, 53, 53, 32]\n\n    conv2 = self.conv(conv1, weights['wc2'], biases['bc2'])  # convolution layer 2\n    conv2 = self.maxPool(conv2)                              # max pool layer 2\n    # out shape -&gt; [None, 27, 27, 64]\n\n    conv3 = self.conv(conv2, weights['wc3'], biases['bc3'])  # convolution layer 3\n    conv3 = self.maxPool(conv3)                              # max pool layer 3\n    # out shape -&gt; [None, 14, 14, 128]\n\n    flayer = tf.reshape(conv3, [-1, weights['wd1'].shape[0]])   # flatten the output from convo layer\n    # for 7 x 7 x 128 this is -&gt; [None, 6272] \n    flayer = tf.add(tf.matmul(flayer, weights['wd1']), biases['bd1'])           # fully connected layer 1\n    flayer = tf.nn.relu(flayer)\n    # out shape -&gt; [None, 128]\n\n    out = tf.add( tf.matmul(flayer, weights['out']), biases['out'] )   # do last set of output weight * vals + bias  \n    # out shape -&gt; [None, 20]       \n\n    return out      # net input to output layer\n<\/code><\/pre>\n\n<p>But so now, in my main program I feed input data to my model in batches, basically with:<\/p>\n\n<pre><code>    out = self.Architecture(x, self.weights, self.biases)    # Implement network architecture, and get output tensor (net input to output layer)\n\n    # normalize, softmax and entropy the net input, in comparison with provided labels\n    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=out, labels=y) )   # cost function\n    optimizer = tf.train.GradientDescentOptimizer(learning_rate=lr).minimize(cost)                  # gradient descent optimizer\n\n    pred = tf.equal(tf.argmax(out, 1), tf.argmax(y , 1)) # output true \/ false if predicted value matches label\n    accuracy = tf.reduce_mean(tf.cast(pred, tf.float32))             # percentage value of correct predictions\n\n    for i in range(iters):          \n            [BX, _, BY, _] = batch.split(trainX, trainY, Bsize) # random split in batch size\n            # data shapes: BX -&gt; [160, 105, 105, 1], BY -&gt; [160, 20]\n\n            # Code bombs out after feeding with input data\n            opt = sess.run(optimizer, feed_dict={lr:learnr, x:BX, y:BY } )\n<\/code><\/pre>\n\n<p>The exception what I then get after with the sess.run command is:<\/p>\n\n<blockquote>\n  <p>'logits and labels must be broadcastable: logits_size=[640,20] labels_size=[160,20]\\n\\t [[Node: softmax_cross_entropy_with_logits = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Add_1, softmax_cross_entropy_with_logits\/Reshape_1)]]'<\/p>\n<\/blockquote>\n\n<p>From this I interpret that the softmax is getting [640, 20] as input while it's expecting [160, 20]... I do not understand how, where the data could be shaped to [640, 20]???<\/p>\n\n<p>Please show me if I'm missing something, or misinterpreting the error? <\/p>\n","answers":[]}
+{"QuestionId":52664110,"AnswerCount":4,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-05T10:58:44.177","AcceptedAnswerId":null,"OwnerUserId":6859553.0,"Title":"AttributeError: 'Sequential' object has no attribute 'output_names'","Body":"<p>I have got some problem for the below code \nof the following line \n<strong>new_model = load_model('124446.model', custom_objects=None, compile=True)<\/strong>\nHere is the code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.models import load_model\n\nmnist = tf.keras.datasets.mnist\n\n(x_train,y_train), (x_test,y_test) = mnist.load_data()\n\nx_train = tf.keras.utils.normalize(x_train,axis=1)\nx_test = tf.keras.utils.normalize(x_test,axis=1)\n\nmodel = tf.keras.models.Sequential()\n\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(128,activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(128,activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(10,activation=tf.nn.softmax))\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(x_train,y_train,epochs=3)\n\n\ntf.keras.models.save_model(model,'124446.model')\n\n\nval_loss, val_acc = model.evaluate(x_test,y_test)\nprint(val_loss, val_acc)\n\n\nnew_model = load_model('124446.model', custom_objects=None, compile=True)\n\n\nprediction = new_model.predict([x_test])\nprint(prediction)\n<\/code><\/pre>\n\n<p>Errors are:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"C:\/Users\/TanveerIslam\/PycharmProjects\/DeepLearningPractice\/1.py\",\n  line 32, in \n      new_model = load_model('124446.model', custom_objects=None, compile=True)   File\n  \"C:\\Users\\TanveerIslam\\PycharmProjects\\DeepLearningPractice\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\saving.py\",\n  line 262, in load_model\n      sample_weight_mode=sample_weight_mode)   File \"C:\\Users\\TanveerIslam\\PycharmProjects\\DeepLearningPractice\\venv\\lib\\site-packages\\tensorflow\\python\\training\\checkpointable\\base.py\",\n  line 426, in _method_wrapper\n      method(self, *args, **kwargs)   File \"C:\\Users\\TanveerIslam\\PycharmProjects\\DeepLearningPractice\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\",\n  line 525, in compile\n      metrics, self.output_names)<\/p>\n  \n  <p><strong>AttributeError: 'Sequential' object has no attribute 'output_names'<\/strong><\/p>\n<\/blockquote>\n\n<p>So can any one give me ant solution.<\/p>\n\n<p>Note: I use pycharm as IDE.<\/p>\n","answers":[{"AnswerId":"52711716","CreationDate":"2018-10-09T00:21:23.640","ParentId":null,"OwnerUserId":"1438845","Title":null,"Body":"<p>If this is run on Windows then the issue is that currently toco is not supported on Windows - <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20975\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20975<\/a><\/p>\n"},{"AnswerId":"56291662","CreationDate":"2019-05-24T11:21:30.903","ParentId":null,"OwnerUserId":"8156353","Title":null,"Body":"<pre><code>import tensorflow as tf    \ntf.keras.models.save_model(\n    model,\n    \"epic_num_reader.model\",\n    overwrite=True,\n    include_optimizer=True\n) \n\nnew_model = tf.keras.models.load_model('epic_num_reader.model', custom_objects=None, compile=False)\n\npredictions = new_model.predict(x_test)\nprint(predictions)\n\nimport numpy as np\n\nprint(np.argmax(predictions[0]))\nplt.imshow(x_test[0],cmap=plt.cm.binary)\nplt.show()\n<\/code><\/pre>\n"},{"AnswerId":"53980337","CreationDate":"2018-12-30T18:33:15.963","ParentId":null,"OwnerUserId":"6805747","Title":null,"Body":"<p>As @Shinva said to set the \"compile\" attribute of the load_model function to \"False\".\nThen after loading the model, compile it separately.<\/p>\n\n<pre><code>from tensorflow.keras.models import save_model, load_model\nsave_model(model,'124446.model')\n<\/code><\/pre>\n\n<p>Then for loading the model again do:<\/p>\n\n<pre><code>saved_model = load_model('124446.model', compile=False)\nsaved_model.compile(optimizer='adam',\n          loss='sparse_categorical_crossentropy',\n          metrics=['accuracy'])\nsaved_model.predict([x_test])\n<\/code><\/pre>\n\n<p><strong>Update<\/strong>: For some unknown reasons, I started to get the same errors as the question states. After trying to find different solutions it seems using the \"keras\" library directly instead of \"tensorflow.keras\" works properly.  <\/p>\n\n<p>My setup is on \"Windows 10\" with python:'3.6.7', tensorflow:'1.11.0' and keras:'2.2.4'<\/p>\n\n<p>As per my knowledge, there are three different ways in which you can save and restore your model; provided you have used keras directly to make your model.<\/p>\n\n<p><strong>Option1:<\/strong><\/p>\n\n<pre><code>import json\nfrom keras.models import model_from_json, load_model\n\n# Save Weights + Architecture\nmodel.save_weights('model_weights.h5')\nwith open('model_architecture.json', 'w') as f:\n    f.write(model.to_json())\n\n# Load Weights + Architecture\nwith open('model_architecture.json', 'r') as f:\n    new_model = model_from_json(f.read())\nnew_model.load_weights('model_weights.h5')\n<\/code><\/pre>\n\n<p><strong>Option2:<\/strong><\/p>\n\n<pre><code>from keras.models import save_model, load_model\n\n# Creates a HDF5 file 'my_model.h5' \nsave_model(model, 'my_model.h5') # model, [path + \"\/\"] name of model\n\n# Deletes the existing model\ndel model  \n\n# Returns a compiled model identical to the previous one\nnew_model = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p><strong>Option 3<\/strong><\/p>\n\n<pre><code># using model's methods\nmodel.save(\"my_model.h5\")\n\n# deletes the existing model\ndel model\n\n# load the saved model back\nnew_model = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>Option 1 requires the new_model to be <strong>compiled<\/strong> before using.  <\/p>\n\n<p>Option 2 and 3 are almost similar in syntax. <\/p>\n\n<p>Codes used from:<br>\n 1. <a href=\"https:\/\/jovianlin.io\/saving-loading-keras-models\/\" rel=\"nofollow noreferrer\">Saving &amp; Loading Keras Models<\/a><br>\n 2. <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model<\/a><\/p>\n"},{"AnswerId":"53474769","CreationDate":"2018-11-26T04:25:51.990","ParentId":null,"OwnerUserId":"1055144","Title":null,"Body":"<p>I was able to load the model by setting <strong>compile=False<\/strong> in <strong>load_model()<\/strong><\/p>\n"}]}
+{"QuestionId":52664390,"AnswerCount":0,"Tags":"<python><tensorflow><cv2>","CreationDate":"2018-10-05T11:15:09.223","AcceptedAnswerId":null,"OwnerUserId":6343916.0,"Title":"Tensor Flow CCTV IP Camera","Body":"<p>I have compiled an object deteection system using TensorFlow on a Windows 10 Machine, using Python 3.6. The model works with my Web Cam as well as a video stored on my PC.<\/p>\n\n<p>I have attemped to get a feed from an IP CCTV camera as below:<\/p>\n\n<pre><code>cap = cv2.VideoCapture(\"http:\/\/username:password@camera ip\")\n<\/code><\/pre>\n\n<p>This results in the error:<\/p>\n\n<pre><code>TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'\n<\/code><\/pre>\n\n<p>Could someone please guide me in the right direction.<\/p>\n","answers":[]}
+{"QuestionId":52664626,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2018-10-05T11:27:09.300","AcceptedAnswerId":null,"OwnerUserId":10396769.0,"Title":"Autoencoder Conv1D Wrong shape","Body":"<p>I am trying to train a model based on a number of txt. file with 2623 lines of numbers. The model is as follows:<\/p>\n\n<pre><code>input_img = Input(shape=(2623,1), name='input')\n\nx = Conv1D(32, 3, activation='relu', padding='same', use_bias=False)(input_img)\nx = BatchNormalization(axis=-1)(x)\nx = MaxPooling1D(2, padding='same')(x)\nx = Conv1D(16, 3, activation='relu', padding='same', use_bias=False, input_shape=(2623,1))(x)\nx = BatchNormalization(axis=-1)(x)\nx = MaxPooling1D(2, padding='same')(x)\nx = Conv1D(16,3, activation='relu', padding='same', use_bias=False, input_shape=(2623,1))(x)\nx = BatchNormalization(axis=-1)(x)\nencoded = MaxPooling1D(2, padding='same')(x)\n\n\nx = Conv1D(16,3, activation='relu', padding='same', use_bias=False, input_shape=(2623,1))(encoded)\nx = BatchNormalization(axis=-1)(x)\nx = UpSampling1D(2)(x)\nx = Conv1D(16,3, activation='relu', padding='same', use_bias=False, input_shape=(2623,1))(x)\nx = BatchNormalization(axis=-1)(x)\nx = UpSampling1D(2)(x)\nx = Conv1D(32, 3, activation='relu', padding='same', use_bias=False, input_shape=(2623,1))(x)  #input_shape=(30, 1))\nx = BatchNormalization(axis=-1)(x)  \nx = UpSampling1D(2)(x)\ndecoded = Conv1D(3, 3, activation='sigmoid', padding='same', use_bias=False,input_shape=(2623,1))(x)\n\nautoencoder = Model(input_img, decoded)\nautoencoder.compile(optimizer='rmsprop', loss='binary_crossentropy')\n\n\nx = Input(shape=(16, 300), name=\"input\")\nh = x\nh = Conv1D(filters=300, kernel_size=16,\n             activation=\"relu\", padding='same', name='Conv1',input_shape=(2623,1))(h)\nh = MaxPooling1D(pool_size=16, name='Maxpool1')(h)\n<\/code><\/pre>\n\n<p>Whenever i try to fit the model with the following code i get an error saying: \n<strong>Error when checking target: expected conv1d_283 to have shape (2624, 1) but got array with shape (2623, 1)<\/strong><\/p>\n\n<p>I tried different padding and setting the input space smaller but it does not seem to work. I do not get the error because whatever i do in the end i have an error for different dimensions. Any ideas?<\/p>\n\n<pre><code>history = autoencoder.fit(train, train,\n                epochs=100,\n                batch_size=256,\n                shuffle=True,\n                validation_data=(test, test))                   \n<\/code><\/pre>\n","answers":[{"AnswerId":"52666519","CreationDate":"2018-10-05T13:17:27.117","ParentId":null,"OwnerUserId":"10396769","Title":null,"Body":"<p>Thank you for the help. In the end adding a croplayer in the end did the trick, i.e:<\/p>\n\n<pre><code>x = Cropping1D(cropping=(0, 1))(x) # Crop nothing from input but crop 1 elemnt from the end \n<\/code><\/pre>\n"}]}
+{"QuestionId":52664763,"AnswerCount":1,"Tags":"<java><scala><tensorflow><save>","CreationDate":"2018-10-05T11:34:36.903","AcceptedAnswerId":null,"OwnerUserId":3549680.0,"Title":"Tensorflow: saving SavedModelBundle model from Java","Body":"<p>Is it possible to save a SavedModelBundle from Java API?<\/p>\n\n<p>My use case is that I would like to <\/p>\n\n<ol>\n<li>Load existing <code>SavedModelBundle<\/code><\/li>\n<li>Assign new values for some Variables (by calling assign-op already defined in the graph)<\/li>\n<li>Save the model again<\/li>\n<\/ol>\n\n<p>Steps 1. and 2. I can do but I can't figure out how to do step 3.<\/p>\n\n<p>Any way to achieve this<\/p>\n\n<p>(I would like to do this in Java instead of Python because my whole application is in Java (Scala) so side-stepping to Python will add lot of complications.)<\/p>\n","answers":[{"AnswerId":"52666286","CreationDate":"2018-10-05T13:04:18.017","ParentId":null,"OwnerUserId":"10347276","Title":null,"Body":"<p>Unfortunately, there is no API to save the Tensorflow model in JAVA environments.<\/p>\n\n<p>Basically, in Java environments, you can only load the existed freezed Tensorflow model that you can't add new variables or modify the value in variables.<\/p>\n\n<p>Step2, what you said is that you assign new values to java variable not Tensorflow model's variable.<\/p>\n\n<p>In many case, build the model using the Python and freeze the model (the output file is \"***.pb\" in general), then using the model in Java.<\/p>\n"}]}
+{"QuestionId":52665441,"AnswerCount":0,"Tags":"<c++><tensorflow><cuda><nvcc>","CreationDate":"2018-10-05T12:16:15.107","AcceptedAnswerId":null,"OwnerUserId":133374.0,"Title":"nvcc error: string_view.h: constexpr function return is non-constant","Body":"<p>I'm trying to compile some TensorFlow native op with CUDA code.\nI'm using Python 3.6 and TensorFlow 1.11 here, and CUDA 8.0, on Ubuntu 16.<\/p>\n\n<p>More specifically:<\/p>\n\n<pre><code>gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.10) \nGNU C++11 (Ubuntu 5.4.0-6ubuntu1~16.04.10) version 5.4.0 20160609 (x86_64-linux-gnu)\n        compiled by GNU C version 5.4.0 20160609, GMP version 6.1.0, MPFR version 3.1.4, MPC version 1.0.3\n<\/code><\/pre>\n\n<p><strong>Edit:<\/strong> I repeated the experiment with CUDA 9.0 and CUDA 9.1, and also various compiler backends, via <code>-compiler-bindir clang++<\/code> or <code>-compiler-bindir gcc-4.8<\/code>, and it's always the same error.<\/p>\n\n<p>test.cu:<\/p>\n\n<pre><code>\/\/ For Eigen::GpuDevice.\n#define EIGEN_USE_GPU 1\n\n\/\/ For Eigen::ThreadPoolDevice.\n#define EIGEN_USE_THREADS 1\n\n#include \"tensorflow\/core\/framework\/op.h\"\n#include \"tensorflow\/core\/framework\/shape_inference.h\"\n#include \"tensorflow\/core\/framework\/op_kernel.h\"\n#include \"tensorflow\/core\/common_runtime\/device.h\"\n\n#include &lt;cuda.h&gt;\n#include &lt;cuda_runtime.h&gt;\n#include &lt;cublas_v2.h&gt;\n#include &lt;math_constants.h&gt;\n\n#include \"tensorflow\/core\/platform\/stream_executor.h\"\n<\/code><\/pre>\n\n<p>Compiling:<\/p>\n\n<pre><code>$ \/usr\/local\/cuda-8.0\/bin\/nvcc -shared -O2 -std=c++11 -I \/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include -I \/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/external\/nsync\/public -I \/usr\/local\/cuda-8.0\/include -L \/usr\/local\/cuda-8.0\/lib64 -x cu -DGOOGLE_CUDA=1 -Xcompiler -fPIC -D_GLIBCXX_USE_CXX11_ABI=0 -g test.cu -o test.so -lblas -lf77blas -L\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/numpy\/.libs -lopenblasp-r0-8dca6697.3.0.dev -L\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow -ltensorflow_framework\n<\/code><\/pre>\n\n<p>(Add <code>-v -Xcompiler -v<\/code> to get some more compiler version info, although that does not seem related here.)<\/p>\n\n<p>Output:<\/p>\n\n<pre><code>nvcc warning : The 'compute_20', 'sm_20', and 'sm_21' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/google\/protobuf\/arena_impl.h(55): warning: integer conversion resulted in a change of sign\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/google\/protobuf\/arena_impl.h(309): warning: integer conversion resulted in a change of sign\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/google\/protobuf\/arena_impl.h(310): warning: integer conversion resulted in a change of sign\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/absl\/strings\/string_view.h(501): error: constexpr function return is non-constant\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1255): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"abs\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1255): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"abs\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1260): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"abs\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1260): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"abs\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1285): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1286): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1287): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1293): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1294): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/..\/..\/Eigen\/src\/Core\/MathFunctions.h(1295): warning: calling a constexpr __host__ function(\"imag\") from a __host__ __device__ function(\"exp\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/lib\/bfloat16\/bfloat16.h(72): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"bfloat16\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/lib\/bfloat16\/bfloat16.h(75): warning: calling a constexpr __host__ function(\"real\") from a __host__ __device__ function(\"bfloat16\") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/util\/tensor_format.h(124): warning: missing return statement at end of non-void function \"tensorflow::GetTensorSpatialDims\"\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/util\/tensor_format.h(148): warning: missing return statement at end of non-void function \"tensorflow::GetTensorDimsFromSpatialDims\"\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/framework\/function.h(618): warning: integer conversion resulted in a change of sign\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/tensorflow\/core\/framework\/function.h(619): warning: integer conversion resulted in a change of sign\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::VALUE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::VALUE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::VALUE]\" \n(2096): here\n            instantiation of \"Eigen::internal::igamma_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(34): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::VALUE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::VALUE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::VALUE]\" \n(2096): here\n            instantiation of \"Eigen::internal::igamma_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(34): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\" \n(2102): here\n            instantiation of \"Eigen::internal::igamma_der_a_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma_der_a(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(38): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::DERIVATIVE]\" \n(2102): here\n            instantiation of \"Eigen::internal::igamma_der_a_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma_der_a(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(38): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(2108): here\n            instantiation of \"Eigen::internal::gamma_sample_der_alpha_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::gamma_sample_der_alpha(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(42): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=float, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(2108): here\n            instantiation of \"Eigen::internal::gamma_sample_der_alpha_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::gamma_sample_der_alpha(const Scalar &amp;, const Scalar &amp;) [with Scalar=float]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsHalf.h(42): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::VALUE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::VALUE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::VALUE]\" \n(2096): here\n            instantiation of \"Eigen::internal::igamma_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(120): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::VALUE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::VALUE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::VALUE]\" \n(2096): here\n            instantiation of \"Eigen::internal::igamma_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(120): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\" \n(2102): here\n            instantiation of \"Eigen::internal::igamma_der_a_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma_der_a(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(135): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::DERIVATIVE]\" \n(2102): here\n            instantiation of \"Eigen::internal::igamma_der_a_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::igamma_der_a(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(135): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(651): warning: missing return statement at end of non-void function \"Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igammac_cf_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(855): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(2108): here\n            instantiation of \"Eigen::internal::gamma_sample_der_alpha_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::gamma_sample_der_alpha(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(154): here\n\n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/SpecialFunctionsImpl.h(712): warning: missing return statement at end of non-void function \"Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\"\n          detected during:\n            instantiation of \"Scalar Eigen::internal::igamma_series_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(863): here\n            instantiation of \"Scalar Eigen::internal::igamma_generic_impl&lt;Scalar, mode&gt;::run(Scalar, Scalar) [with Scalar=double, mode=Eigen::internal::SAMPLE_DERIVATIVE]\" \n(2108): here\n            instantiation of \"Eigen::internal::gamma_sample_der_alpha_retval&lt;Eigen::internal::global_math_functions_filtering_base&lt;Scalar, void&gt;::type&gt;::type Eigen::numext::gamma_sample_der_alpha(const Scalar &amp;, const Scalar &amp;) [with Scalar=double]\" \n\/u\/zeyer\/py-envs\/py36-tf111\/lib\/python3.6\/site-packages\/tensorflow\/include\/unsupported\/Eigen\/CXX11\/..\/src\/SpecialFunctions\/arch\/CUDA\/CudaSpecialFunctions.h(154): here\n\n1 error detected in the compilation of \"\/var\/tmp\/tmpxft_000021f3_00000000-9_test.cpp1.ii\".\n<\/code><\/pre>\n\n<p>The main error seems to be this:<\/p>\n\n<pre><code>absl\/strings\/string_view.h(501): error: constexpr function return is non-constant\n<\/code><\/pre>\n\n<p>Looking at that code in <code>string_view.h<\/code>:<\/p>\n\n<pre><code>static constexpr size_type CheckLengthInternal(size_type len) {\n  return ABSL_ASSERT(len &lt;= kMaxSize), len;\n}\n<\/code><\/pre>\n\n<p>Where <code>ABSL_ASSERT<\/code> is defined (in <code>absl\/base\/macros.h<\/code>) as:<\/p>\n\n<pre><code>\/\/ ABSL_ASSERT()\n\/\/\n\/\/ In C++11, `assert` can't be used portably within constexpr functions.\n\/\/ ABSL_ASSERT functions as a runtime assert but works in C++11 constexpr\n\/\/ functions.  Example:\n\/\/\n\/\/ constexpr double Divide(double a, double b) {\n\/\/   return ABSL_ASSERT(b != 0), a \/ b;\n\/\/ }\n\/\/\n\/\/ This macro is inspired by\n\/\/ https:\/\/akrzemi1.wordpress.com\/2017\/05\/18\/asserts-in-constexpr-functions\/\n#if defined(NDEBUG)\n#define ABSL_ASSERT(expr) (false ? (void)(expr) : (void)0)\n#else\n#define ABSL_ASSERT(expr)              \\\n  (ABSL_PREDICT_TRUE((expr)) ? (void)0 \\\n                             : [] { assert(false &amp;&amp; #expr); }())  \/\/ NOLINT\n#endif\n<\/code><\/pre>\n\n<p>Any suggestion what the problem might be? And how to resolve it?<\/p>\n\n<p>Compiling the same file directly with <code>g++<\/code> works. So this is specifically a problem with the nvcc frontend (<code>cudafe<\/code>).<\/p>\n\n<p><strong>Edit:<\/strong> I think this might actually be a bug in TensorFlow (or the absl library), so I reported it <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22766\" rel=\"nofollow noreferrer\">here<\/a>. Or it might be a bug in <code>cudafe<\/code>. (But actually, I think the error looks valid, as this is really not a constant expression, and the <code>ABSL_ASSERT<\/code> workaround\/hack described <a href=\"https:\/\/akrzemi1.wordpress.com\/2017\/05\/18\/asserts-in-constexpr-functions\/\" rel=\"nofollow noreferrer\">here<\/a> might not work here.)<\/p>\n","answers":[]}
+{"QuestionId":52665476,"AnswerCount":1,"Tags":"<machine-learning><conv-neural-network><pytorch>","CreationDate":"2018-10-05T12:18:37.553","AcceptedAnswerId":52793492.0,"OwnerUserId":1415817.0,"Title":"CNN weights at input image location","Body":"<p>Given a CNN, say AlexNet: \n<a href=\"https:\/\/i.stack.imgur.com\/xxsaB.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xxsaB.jpg\" alt=\"AlexNet\"><\/a><\/p>\n\n<p>How could one relate kernel locations at the 3rd conv block, i.e 13x13 filter size to the input image. I want to compare the filters at different locations. \nI was thinking of just bilinearly upsampling the location, from 13x13 to 224x224,<\/p>\n\n<pre><code>nn.Upsample(size = (224,224), mode='bilinear')\n<\/code><\/pre>\n\n<p>however it's hard to justify local correspondence. <\/p>\n","answers":[{"AnswerId":"52793492","CreationDate":"2018-10-13T13:35:00.867","ParentId":null,"OwnerUserId":"1415817","Title":null,"Body":"<p>Using adaptive pooling, with adaptive pooling, one can reduce any feature map size.<\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/pooling.html\" rel=\"nofollow noreferrer\">Adaptive max pooling<\/a><\/p>\n"}]}
+{"QuestionId":52666624,"AnswerCount":2,"Tags":"<python><tensorflow><cudnn>","CreationDate":"2018-10-05T13:23:00.623","AcceptedAnswerId":52667018.0,"OwnerUserId":9516820.0,"Title":"Tensorflow-gpu not installed properly in windows machine","Body":"<p>I am trying to install tensorflow-gpu in my new windows system. But am encountering this error<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\nTraceback (most recent call last):\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\user\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n\n<p>This question has been asked before and I have gone through those but I am not able to figure why this is not working.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/44503603\/tensorflow-on-windows-importerror-dll-load-failed-the-specified-module-could\">Tensorflow on windows - ImportError: DLL load failed: The specified module could not be found<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10033\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10033<\/a><\/p>\n\n<p>I have added all the paths to the environment variables based on this post\n<a href=\"https:\/\/medium.com\/@viveksingh.heritage\/how-to-install-tensorflow-gpu-version-with-jupyter-windows-10-in-8-easy-steps-8797547028a4\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@viveksingh.heritage\/how-to-install-tensorflow-gpu-version-with-jupyter-windows-10-in-8-easy-steps-8797547028a4<\/a><\/p>\n\n<p>The versions I am using are<\/p>\n\n<pre><code>Python: 3.6.7 \nCUDA: 10.0.130\ncuDNN: 7.3.1\n<\/code><\/pre>\n","answers":[{"AnswerId":"53539051","CreationDate":"2018-11-29T12:26:18.187","ParentId":null,"OwnerUserId":"4066577","Title":null,"Body":"<p>I suggest you try using Anaconda<\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc\" rel=\"nofollow noreferrer\">See this tutorial<\/a><\/p>\n\n<p>Summarizing:\n1. Install Miniconda or Anaconda and then go to \"Anaconda Prompt\": simple use Windows and start to write \"anaconda\" after installing: it should appear.\n2. Now, simply run:<\/p>\n\n<p>conda create --name tf_gpu tensorflow-gpu<\/p>\n\n<p>Done! To be sure, you can test it: <\/p>\n\n<p>In \"Anaconda Prompt\":<\/p>\n\n<pre><code>Activate environment using \u2018activate tf_env\u2019.\nGo to python console using \u2018python\u2019\n<\/code><\/pre>\n\n<p>In Python, just write:<\/p>\n\n<p>import tensorflow as tf\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))<\/p>\n\n<p>Info about your card should appear.\nNow, just create a new project in your IDE, choosing \"Existing Interpreter\", selecting the Virtual Envirment that we just created (\"tf_gpu\"), and your new files inside the project will run on the gpu.<\/p>\n\n<p>Hope I helped! Thanks to Harveen Singh (I tried many tutorials, before seeing one that worked and was easy)<\/p>\n"},{"AnswerId":"52667018","CreationDate":"2018-10-05T13:43:40.850","ParentId":null,"OwnerUserId":"7303319","Title":null,"Body":"<p>TF supports CUDA 9.0. <\/p>\n\n<p>Follow these instructions<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a><\/p>\n\n<p><a href=\"https:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-microsoft-windows\/\" rel=\"nofollow noreferrer\">https:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-microsoft-windows\/<\/a><\/p>\n"}]}
+{"QuestionId":52668857,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection><tensorflow-serving><object-detection-api>","CreationDate":"2018-10-05T15:23:51.227","AcceptedAnswerId":52670172.0,"OwnerUserId":7495008.0,"Title":"Tensorflow Object Detection API: How to ignore regions during training?","Body":"<p>I'm using the object detection API from the models\/research python repo on Ubuntu 16.04, and I wanted to fine-tune a pre-trained model (at the moment I'm interested in SSD with MobileNet or Inception backbones) on the UA-DETRAC dataset.<\/p>\n\n<p>The problem is that there are specific regions, with their bounding boxes, which are marked as \"ignored regions\" and I wouldn't want the model to train on what he thinks are some false positives which are true, just not annotated (included in those regions).\nI thought of cropping the images to exclude those regions, but I would lose some information.<\/p>\n\n<p>Is there the built-in possibility to mark them as \"don't care\" boxes or should I modify the code?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52670172","CreationDate":"2018-10-05T16:52:13.970","ParentId":null,"OwnerUserId":"5276428","Title":null,"Body":"<p>If those <em>regions to ignore<\/em> remain static, as in, the contents of the region doesn't change throughout the dataset, then the model can be learnt to ignore those regions. <\/p>\n\n<p>If you really want the model to ignore them during training, then mask them with a constant value.<\/p>\n"}]}
+{"QuestionId":52669719,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2018-10-05T16:20:15.313","AcceptedAnswerId":null,"OwnerUserId":4516769.0,"Title":"Keras single Image prediction not working with correct dims","Body":"<p>I want to predict single images with the functional API (keras version 2.2.2, tensorflow backend v1.7) . I load my model: <\/p>\n\n<pre><code># Loading Model\nbase_model = VGG16(include_top = False,  weights=None,\n              input_shape=(224,224,3))\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\nx = Dropout(0.7)(x)\nx = Dense(1020, activation='relu')(x)\npredictions = Dense(1, activation='sigmoid')(x)\nmodel = Model(inputs=base_model.input, outputs=predictions)\n\nmodel.load_weights(\"model.h5\")\n<\/code><\/pre>\n\n<p>Then I load an Image and transform it into the input format and try to predict: <\/p>\n\n<pre><code>from keras.preprocessing.image import img_to_array, load_img\nimg = load_img(\"data\/my_image.png\")  # this is a PIL image\narray = img_to_array(img)  # this is a Numpy array with shape \narrayresized = cv2.resize(array, (244,244))*1.\/255 \ninputarray = np.expand_dims(arrayresized, axis=0)\n\n# Predicting\nprediction = model.predict(inputarray, batch_size = 1)\n<\/code><\/pre>\n\n<p>Then I get this error back: <\/p>\n\n<pre><code>ValueError: Error when checking input: expected input_4 to have shape \n(224, 224, 3) but got array with shape (244, 244, 3)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52670060,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensor>","CreationDate":"2018-10-05T16:44:34.307","AcceptedAnswerId":52670174.0,"OwnerUserId":8344009.0,"Title":"KERAS: Get a SLICE of RNN timesteps with return_sequence = True","Body":"<p>My questions is rather simple but seemed to be unsolved. <\/p>\n\n<p><strong>Input<\/strong>: (bs, timesteps, input_dim) --> Tensor(\"stack:0\", shape=(?, <strong>4<\/strong>, 400), dtype=float32)<\/p>\n\n<p><strong>Layer<\/strong>:\noutput = LSTM(100, input_shape = (timesteps, input_feature), return_sequence = True) (input)<\/p>\n\n<p><strong>Expect<\/strong>: (bs, timesteps, output_dim) --> Tensor(\"gru_20\/transpose_1:0\", shape=(?, <strong>4<\/strong>, 100), dtype=float32)<\/p>\n\n<p><strong>Output<\/strong>: Tensor(\"gru_20\/transpose_1:0\", shape=(?, <strong>?<\/strong>, 100), dtype=float32)<\/p>\n\n<p>Why does Keras not infer the number of timesteps, even it receives an input_shape? When I use the model summary the result it shows has the correct output shape:<\/p>\n\n<hr>\n\n<p>lstm_2 (LSTM)                (None, 4, 100)             3232     <\/p>\n\n<p>But not during construction. So, when I want to unstacked the Tensor to a list of Tensors for every timesteps * (bs, 10) by using unstack(output, axis=1)] I receive ofc this error: ValueError: Cannot infer num from shape (?, ?, 100)<\/p>\n\n<p>Where is my mistake? <\/p>\n\n<p>BTW. Adding TimeDistributed(Dense(100))(questions) results in the correct output dim: Tensor(\"time_distributed_17\/Reshape_1:0\", shape=(?, <strong>4<\/strong>, 100), dtype=float32) but not an option because of shared weights. If not, what is the workaround? <\/p>\n","answers":[{"AnswerId":"52670174","CreationDate":"2018-10-05T16:52:27.227","ParentId":null,"OwnerUserId":"8344009","Title":null,"Body":"<p>Work-around might be a multiplication by 1 (no change). <\/p>\n\n<p>workaround = TimeDistributed(Lambda(lambda x: x * 1.0))(output)<\/p>\n\n<p>Inference works here: Tensor(\"time_distributed_17\/Reshape_1:0\", shape=(?, 4, 100), dtype=float32)<\/p>\n\n<p>Is a TimeDistributed Layer always necessary when using return_sequences=True?<\/p>\n"}]}
+{"QuestionId":52670408,"AnswerCount":2,"Tags":"<python><pandas><tensorflow><google-cloud-ml>","CreationDate":"2018-10-05T17:10:09.940","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Reading a pandas pickle file in Tensorflow in CloudML","Body":"<p>I'm getting an error trying to read a pandas pickle e.g. <code>df.to_pickle()<\/code> method, which is stored in Google Cloud storage. I'm trying to do the following:<\/p>\n\n<pre><code>path_to_gcs_file = 'gs:\/\/xxxxx'\nf = file_io.FileIO(path_to_gcs_file, mode='r').read()\ntrain_df = pd.read_pickle(f)\nf.close()\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte\n<\/code><\/pre>\n\n<p>Alternatively I tried:<\/p>\n\n<pre><code>f = BytesIO(file_io.read_file_to_string(path_to_gcs_file, binary_mode=True))\ntrain_df = pd.read_pickle(f)\n<\/code><\/pre>\n\n<p>Which works locally but not on CloudML!<\/p>\n\n<pre><code>f = file_io.read_file_to_string(path_to_gcs_file, binary_mode=True)\ntrain_df = pd.read_pickle(f)\n<\/code><\/pre>\n\n<p>Gives me an error:\nAttributeError: 'bytes' object has no attribute 'seek'<\/p>\n","answers":[{"AnswerId":"52671497","CreationDate":"2018-10-05T18:35:31.363","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p><code>pandas.read_pickle<\/code> accepts a path as the first argument; you are passing a <code>File<\/code> object (<code>file.FileIO<\/code>) and a <code>bytes<\/code> object (<code>read_to_string<\/code>).<\/p>\n\n<p>So far I have not found a way to read a pickle object directly from GCS using pandas, so you will have to copy it to the machine. You could use <code>file_io.copy<\/code> for that:<\/p>\n\n<pre><code>file_io.copy('gs:\/\/xxxx', '\/tmp\/x.pkl')\ntrain_df = pd.read_pickle('\/tmp\/x.pkl')\n<\/code><\/pre>\n"},{"AnswerId":"52670512","CreationDate":"2018-10-05T17:16:37.727","ParentId":null,"OwnerUserId":"3806464","Title":null,"Body":"<p>You should be able to get away with using a context manager, but I think you're pulling the end of the certificate using this way, so you should instead download the file through the api<\/p>\n\n<pre><code>pip install --upgrade google-cloud-storage\n<\/code><\/pre>\n\n<p>Then<\/p>\n\n<pre><code># Initialise a client\nstorage_client = storage.Client(\"[Your project name here]\")\n# Create a bucket object for our bucket\nbucket = storage_client.get_bucket(bucket_name)\n# Create a blob object from the filepath\nblob = bucket.blob(\"folder_one\/foldertwo\/filename.extension\")\n# Download the file to a destination\nblob.download_to_filename(path_to_gcs_file)\nwith open(path_to_gcs_file, \"rb\" as f:\n    train_df = = pickle.load(f)\n<\/code><\/pre>\n\n<p>Much was taken from this answer:\n<a href=\"https:\/\/stackoverflow.com\/questions\/42555142\/downloading-a-file-from-google-cloud-storage-inside-a-folder\">Downloading a file from google cloud storage inside a folder<\/a><\/p>\n"}]}
+{"QuestionId":52670667,"AnswerCount":0,"Tags":"<python><python-3.x><numpy><tensorflow>","CreationDate":"2018-10-05T17:29:00.467","AcceptedAnswerId":null,"OwnerUserId":7897991.0,"Title":"How do I save a tensorflow ML model without a session?","Body":"<p>When I look this problem up I find solutions using a session variable. I am not using a session variable so these do not work. I have also tried <code>tf.keras.models.save_model<\/code> and <code>tf.keras.models.load_model<\/code> but when I run the load_model I get an error relating to not being able to access the file. <\/p>\n\n<p>Is there a best solution for this?<\/p>\n\n<p>My error:\n<code>KeyError: 0<\/code><\/p>\n\n<p>My training code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(512, activation=tf.nn.relu),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=1)\n\ntf.keras.models.save_model(model,\"model.h5\",overwrite=True,\n  include_optimizer=True)\n<\/code><\/pre>\n\n<p>My testing code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom keras.models import load_model\nimport numpy as np\n\nmodel = load_model(\"model.h5\")\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmnist = tf.keras.datasets.mnist\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nplt.figure(figsize=(20,20))\nfor i in range(25):\n  plt.subplot(5,5,i+1)\n  plt.xticks([])\n  plt.yticks([])\n  plt.grid(False)\n  plt.imshow(x_test[i])\n  plt.xlabel(np.argmax(model.predict(x_test)[i]))  \n\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52671138,"AnswerCount":4,"Tags":"<rest><tensorflow><machine-learning><tensorflow-serving>","CreationDate":"2018-10-05T18:08:04.730","AcceptedAnswerId":52748087.0,"OwnerUserId":1766088.0,"Title":"Tensorflow Serving: Rest API returns \"Malformed request\" error","Body":"<p>Tensorflow Serving server (run with docker) responds to my GET (and POST) requests with this:<\/p>\n\n<p><code>{ \"error\": \"Malformed request: POST \/v1\/models\/saved_model\/\" }<\/code><\/p>\n\n<p>Precisely the same problem was already reported but never solved (supposedly, this is a StackOverflow kind of question, not a GitHub issue):<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/1085\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/issues\/1085<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/1095\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/issues\/1095<\/a><\/p>\n\n<p>Any ideas? Thank you very much.<\/p>\n","answers":[{"AnswerId":"53230125","CreationDate":"2018-11-09T16:54:04.037","ParentId":null,"OwnerUserId":"2226322","Title":null,"Body":"<p>I verified that this does not work pre-v12 and does indeed work post-v12.<\/p>\n\n<pre><code>&gt; docker run -it -p 127.0.0.1:9000:8500 -p 127.0.0.1:9009:8501 -v \/models\/55:\/models\/55 -e MODEL_NAME=55 --rm tensorflow\/serving\n&gt; curl http:\/\/localhost:9009\/v1\/models\/55\n   { \"error\": \"Malformed request: GET \/v1\/models\/55\" }\n<\/code><\/pre>\n\n<p>Now try with v12:<\/p>\n\n<pre><code>&gt; docker run -it -p 127.0.0.1:9000:8500 -p 127.0.0.1:9009:8501 -v \/models\/55:\/models\/55 -e MODEL_NAME=55 --rm tensorflow\/serving:1.12.0\n&gt; curl http:\/\/localhost:9009\/v1\/models\/55\n{\n \"model_version_status\": [\n  {\n   \"version\": \"1541703514\",\n   \"state\": \"AVAILABLE\",\n   \"status\": {\n    \"error_code\": \"OK\",\n    \"error_message\": \"\"\n   }\n  }\n ]\n}\n<\/code><\/pre>\n"},{"AnswerId":"52718715","CreationDate":"2018-10-09T10:22:42.600","ParentId":null,"OwnerUserId":"9594598","Title":null,"Body":"<p>Depends on your model, but this is what my body looks like:<\/p>\n\n<p>{\"inputs\": {\"text\": [\"Hello\"]}}<\/p>\n\n<p>I used Postman to help me out so that it knew it was a JSON.<\/p>\n\n<p>This is for predict API, so the url ends in  \":predict\"\nAgain, that depends on what API you're trying to use.<\/p>\n"},{"AnswerId":"52748087","CreationDate":"2018-10-10T20:12:55.883","ParentId":null,"OwnerUserId":"1766088","Title":null,"Body":"<p>There were two issues with my approach:<\/p>\n\n<p>1) The status check request wasn't supported in my Tensorflow_model_server (see <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/1085\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/issues\/1085<\/a> for details)<\/p>\n\n<p>2) More importantly, when using Windows you must <strong>escape quotation marks in JSON<\/strong>. So instead of:<\/p>\n\n<pre><code>curl -XPOST http:\/\/localhost:8501\/v1\/models\/saved_model:predict -d \"{\"instances\":[{\"features\":[1,1,1,1,1,1,1,1,1,1]}]}\"\n<\/code><\/pre>\n\n<p>I should have used this:<\/p>\n\n<pre><code>curl -XPOST http:\/\/localhost:8501\/v1\/models\/saved_model:predict -d \"{\\\"instances\\\":[{\\\"features\\\":[1,1,1,1,1,1,1,1,1,1]}]}\"\n<\/code><\/pre>\n"},{"AnswerId":"52750799","CreationDate":"2018-10-11T01:17:23.390","ParentId":null,"OwnerUserId":"6007480","Title":null,"Body":"<p>Model status API is only supported in master branch. There is no TF serving release that supports it yet (the API is slated for upcoming 1.12 release). You can use the nightly docker image (tensorflow\/serving:nightly) to test on master branch builds.<\/p>\n\n<p>This solution gived by netf in <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/1128\" rel=\"nofollow noreferrer\">issue:1128 in tensorflow\/serving<\/a>.\nI already try this solution, it's done and i can get the model status.<a href=\"https:\/\/i.stack.imgur.com\/qgUTG.png\" rel=\"nofollow noreferrer\">Getting Model status img<\/a>(this is the img for model status demo).<\/p>\n\n<p>Hope I can help you.<\/p>\n\n<p>If you not clear the master branch builds, you can contact me.<\/p>\n\n<p>I can give your instruction.<\/p>\n\n<p>Email:mizeshuang@gmail.com<\/p>\n"}]}
+{"QuestionId":52671193,"AnswerCount":1,"Tags":"<r><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-05T18:13:44.123","AcceptedAnswerId":52671921.0,"OwnerUserId":2840286.0,"Title":"R Keras flatten layer - got an array of shape 1","Body":"<p>I am trying to build a model in R Keras with TensorFlow that just has a flatten layer. Here is a snippet of the code:<\/p>\n\n<pre><code>model &lt;- keras_model_sequential() %&gt;%\n  layer_flatten(input_shape = c(lookback, dim(train.data)[-1]))\n\nmodel %&gt;% compile(\n  optimizer = optimizer_rmsprop(),\n  loss = \"mae\"\n)\n\nhistory &lt;- model %&gt;% fit_generator(\n  train_gen,\n  steps_per_epoch = 500,\n  epochs = 20\n)\n<\/code><\/pre>\n\n<p><code>lookback<\/code> is 1200 and <code>dim(train.data)<\/code> is <code>(13155, 3)<\/code>. The input to the flatten layer is <code>(1200, 3)<\/code> and I expect it should output a 1D vector of 3600.<\/p>\n\n<p><code>train_gen<\/code> returns a list of 2. The first is a 3D matrix with dimensions <code>(129, 1200, 3)<\/code> and the second one is a 1D vector with dimension <code>(129,)<\/code>.<\/p>\n\n<p>However, I get the error:<\/p>\n\n<pre><code>Error in py_call_impl(callable, dots$args, dots$keywords) : \n  ValueError: Error when checking target: expected dense_15 to have shape (3600,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>I don't know why this is happening. If I add a <code>layer_dense(units = 1)<\/code> it works but I don't understand why.<\/p>\n","answers":[{"AnswerId":"52671921","CreationDate":"2018-10-05T19:07:17.533","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The target shapes are incompatible: you provide as target a vector of size of 1 (i.e. <code>(129,)<\/code> means 129 sample labels with size one), however the model's output shape is <code>(None, 3600)<\/code> so it expects vectors of size 3600. And that's why when you add a Dense layer with one unit the problem is resolved: the Dense layer's output shape is <code>(None, 1)<\/code> and you provide <code>(129,)<\/code> and they match with each other, hence no issue to complain about.<\/p>\n"}]}
+{"QuestionId":52671474,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><lstm>","CreationDate":"2018-10-05T18:33:57.490","AcceptedAnswerId":52672082.0,"OwnerUserId":10462235.0,"Title":"Keras: Input 0 is incompatible with layer lstm_26: expected ndim=3, found ndim=2","Body":"<p>I have a text classification problem and I'm using a LSTM layer to get the context vectors, using Keras.<\/p>\n\n<p>I used the <code>tokenizer<\/code> and padded the sequences so my <code>X_train<\/code> has the shape <code>(21226, 1500)<\/code> where 1500 is the max sequence length.<\/p>\n\n<p>How do I directly pass this sequence to an LSTM with 64 units? I tried doing it directly but there appears to be an issue with the dimensions.<\/p>\n\n<blockquote>\n  <p>Input 0 is incompatible with layer lstm_26: expected ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>Here is the code, in Keras:<\/p>\n\n<pre><code>input = Input(shape=(1500,))\nlstm1 = LSTM(units=64, dropout=0.2, recurrent_dropout=0.2)\nout = lstm1(input)\n<\/code><\/pre>\n\n<p>Edit:<\/p>\n\n<p>Here is the updated code:<\/p>\n\n<pre><code>def getm():\n    inp = Input(shape=(1500,1))\n    lstm1 = LSTM(units=64, dropout=0.2, recurrent_dropout=0.2)\n    out = lstm1(inp)\n    model = Model(inputs=inp, outputs=out)\n    model.compile(optimizer='rmsprop', loss=custom_loss, metrics=[custom_loss])\n    return model\n<\/code><\/pre>\n\n<p>Now, X_train's shape is (21226, 1500), after reshaping with np.expand_dims, it is (21226, 1500, 1).<\/p>\n\n<p>I read the Keras docs and in X_train's shape, the first attribute should be the batch size, which should be None as Keras' predict and fit functions take care of that.<\/p>\n\n<p>EDIT2:<\/p>\n\n<p>Here's the complete code:<\/p>\n\n<pre><code>def getm():\n    inp = Input(shape=(1500,1))\n    lstm1 = LSTM(units=64, dropout=0.2, recurrent_dropout=0.2)\n    out = lstm1(inp)\n    model = Model(inputs=inp, outputs=out)\n    model.compile(optimizer='rmsprop', loss=custom_loss, metrics=[custom_loss])\n    return model\n\nmodel = getm()\n\nX_train.shape = (21226, 500)\n<\/code><\/pre>\n\n<p>I reshape this using<\/p>\n\n<pre><code>my_data = X_train.reshape(X_train.shape + (1,))\n<\/code><\/pre>\n\n<p>Now,<\/p>\n\n<pre><code>my_data.shape = (21226, 1500, 1)\n<\/code><\/pre>\n\n<p>I'm trying to predict the output for one data point:<\/p>\n\n<pre><code> model.predict(my_data[0])\n<\/code><\/pre>\n\n<p>And this error is thrown:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected input_25 to have 3 dimensions, but got array with shape (1500, 1) \n<\/code><\/pre>\n\n<p>EDIT3:<\/p>\n\n<p>This is the model summary<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_25 (InputLayer)        (None, 1500, 1)           0         \n_________________________________________________________________\nlstm_16 (LSTM)               (None, 64)                16896     \n=================================================================\nTotal params: 16,896\nTrainable params: 16,896\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n","answers":[{"AnswerId":"52672082","CreationDate":"2018-10-05T19:22:11.813","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>LSTM<\/code> layer expects a 3D input array of shape <code>(batch_size, num_timesteps or sequence_length, num_features)<\/code> (that's why you are getting <code>expected ndim=3, found ndim=2<\/code> error). So if you would like to feed your tokenized and word-indexed sentences directly to a LSTM layer you need to reshape it such that it also has a feature axis (which in this case would contain the index of the words in the dictionary):<\/p>\n\n<pre><code># using np.reshape\nmy_data = my_data.reshape(my_data.shape + (1,))\n\n# using np.expand_dims\nmy_data = np.expand_dims(my_data, axis=-1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52671481,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-05T18:34:50.090","AcceptedAnswerId":52878788.0,"OwnerUserId":10458815.0,"Title":"Why are variables defined with 'self' automatically given a ListWrapper() while inheriting from tf.keras.Model?","Body":"<p>I am not familiar with <code>ListWrapper()<\/code>, but it is being applied to all list variables created with <code>self<\/code> when my class inherits from <code>tf.keras.Model<\/code>. \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/models\/Model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/models\/Model<\/a><\/p>\n\n<p>This is bad because it is causing an <code>IndexError<\/code> when I use it in certain functions, or even by just passing it through my Tensorflow model. (I am using eager execution)<\/p>\n\n<p>A small reproduction of the problem can be seen with this code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nclass my_class(tf.keras.Model):\n\n    def __init__(self):\n        super(my_class, self).__init__()\n\n        self.x = [0]\n        print(self.x)\n\nmodel = my_class()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>ListWrapper([0])\n<\/code><\/pre>\n\n<p>Switching the inheritance to be from <code>object<\/code> solves the issue, which is how I know its <code>tf.keras.Model<\/code> that is causing this.<\/p>\n\n<p>I tried looking it up but can't find anything on this. Any tips? Thanks!<\/p>\n","answers":[{"AnswerId":"52878788","CreationDate":"2018-10-18T16:41:08.363","ParentId":null,"OwnerUserId":"10458815","Title":null,"Body":"<p>Turns out this was a bug in Tensorflow between tf.keras.Model and eager execution. \nThis was not \"how the tensorflow has 'patched' setting attributes\", as suggested by a comment.<\/p>\n\n<p>This is a link to the closed issue on Tensorflow:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22853\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22853<\/a><\/p>\n\n<p>If you have this problem, it should be fixed in the next Tensorflow update.\nThis bug was in version 1.11.0<\/p>\n"}]}
+{"QuestionId":52671948,"AnswerCount":0,"Tags":"<android><tensorflow><image-recognition><arcore><android-augmented-reality>","CreationDate":"2018-10-05T19:09:39.053","AcceptedAnswerId":null,"OwnerUserId":1858047.0,"Title":"How Can it connect TensorFlow lite and ArCore SceneForm Sdk on Android Project?","Body":"<p>I want to train a custom model to detect some  object using Image Classifier models or Object Detection Api from TensorFlow  when the object will be detected, it can show image 3D related with the result, for this, I want to use ArCore SceneForm SDK, these technologies about machine learning and Reality Augmented in a Android Project.<\/p>\n\n<p>I have problems to use ArCore SceneForm SDK to get frames per seconds on Android camera to scan from some TensorFlow Lite class using custom detection model and then shows a 3D image. Can you please help about this?<\/p>\n\n<p>TensorFlow lite\n<a href=\"https:\/\/www.tensorflow.org\/lite\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/lite\/<\/a><\/p>\n\n<p>ArCore SceneForm SDK\n<a href=\"https:\/\/developers.google.com\/ar\/develop\/java\/sceneform\/\" rel=\"noreferrer\">https:\/\/developers.google.com\/ar\/develop\/java\/sceneform\/<\/a><\/p>\n","answers":[]}
+{"QuestionId":52673610,"AnswerCount":1,"Tags":"<python><debugging><machine-learning><pytorch>","CreationDate":"2018-10-05T21:41:40.750","AcceptedAnswerId":52675560.0,"OwnerUserId":5157633.0,"Title":"Printing all the contents of a tensor","Body":"<p>I came across <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/neural_networks_tutorial.html#sphx-glr-beginner-blitz-neural-networks-tutorial-py\" rel=\"nofollow noreferrer\">this<\/a> PyTorch tutorial (in neural_networks_tutorial.py) where they construct a simple neural network and run an inference. I would like to print the contents of the entire input tensor for debugging purposes. What I get when I try to print the tensor is something like this and not the entire tensor:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zsLP1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zsLP1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I saw a similar <a href=\"https:\/\/stackoverflow.com\/questions\/1987694\/how-to-print-the-full-numpy-array\">link<\/a> for numpy but was not sure about what would work for PyTorch. I can convert it to numpy and may be view it, but would like to avoid the extra overhead. Is there a way for me to print the entire tensor?<\/p>\n","answers":[{"AnswerId":"52675560","CreationDate":"2018-10-06T03:53:52.417","ParentId":null,"OwnerUserId":"5326903","Title":null,"Body":"<p>Though I don't suggest to do that, if you want, then<\/p>\n\n<pre><code>In [18]: torch.set_printoptions(edgeitems=1)\n\nIn [19]: a\nOut[19]:\ntensor([[-0.7698,  ..., -0.1949],\n        ...,\n        [-0.7321,  ...,  0.8537]])\n\nIn [20]: torch.set_printoptions(edgeitems=3)\n\nIn [21]: a\nOut[21]:\ntensor([[-0.7698,  1.3383,  0.5649,  ...,  1.3567,  0.6896, -0.1949],\n        [-0.5761, -0.9789, -0.2058,  ..., -0.5843,  2.6311, -0.0008],\n        [ 1.3152,  1.8851, -0.9761,  ...,  0.8639, -0.6237,  0.5646],\n        ...,\n        [ 0.2851,  0.5504, -0.9471,  ...,  0.0688, -0.7777,  0.1661],\n        [ 2.9616, -0.8685, -1.5467,  ..., -1.4646,  1.1098, -1.0873],\n        [-0.7321,  0.7610,  0.3182,  ...,  2.5859, -0.9709,  0.8537]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52673692,"AnswerCount":1,"Tags":"<python><tensorflow><keras><pickle><deep-copy>","CreationDate":"2018-10-05T21:52:09.787","AcceptedAnswerId":52677957.0,"OwnerUserId":8125119.0,"Title":"deepcopy class containing keras models","Body":"<p>In my python script, I've created a class which, among others, contains <code>keras<\/code> models like so:<\/p>\n\n<pre><code>from keras.layers import Input, Activation, Dense\nfrom keras.models import Model\n\n\nclass Klass:\n\n    def __init__(self, input_dims, output_dims, hidden_dims, optimizer, a, b):\n\n        self.input_dims = input_dims\n        self.output_dims = output_dims\n        self.hidden_dims = hidden_dims\n        self.optimizer = optimizer\n        self.a = a\n        self.b = b\n\n        self.__build_nn()\n\n    def __build_nn(self):\n\n        inputs = Input(shape=(self.input_dims,))\n        net = inputs\n        for h_dim in self.hidden_dims:\n            net = Dense(h_dim, kernel_initializer='he_uniform')(net)\n            net = Activation(\"relu\")(net)\n\n        outputs = Dense(self.output_dims)(net)\n        outputs = Activation(\"linear\")(outputs)\n        self.nn1 = Model(inputs=inputs, outputs=outputs)\n        self.nn2 = Model(inputs=inputs, outputs=outputs)\n        self.nn1.compile(optimizer=self.optimizer, loss='mean_squared_error')\n        self.nn2.compile(optimizer=self.optimizer, loss='mean_squared_error')\n<\/code><\/pre>\n\n<p>After creating a <code>Klass<\/code> instance, I would like to make a deep copy of it:<\/p>\n\n<pre><code>import copy\nobj = Klass(10, 10, (20, 20), Adam(), 1, 2)\nobj_dc = copy.deepcopy(obj)\n<\/code><\/pre>\n\n<p>However, this throws a <code>TypeError: can't pickle _thread.RLock objects<\/code>. I am pretty sure that the error is related to the <code>keras<\/code> models in the class object since I was able to get a deep copy of a similar class without the <code>keras<\/code> models.<\/p>\n\n<p>Unfortunately, I was unable to find a solution to this on the internet, since most questions concerning deep copying a <code>keras<\/code> model were trying to clone a <code>keras<\/code> model like <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/1765\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>So, how can I get a deep copy of a class containing the <code>keras<\/code> models?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>These three questions (<a href=\"https:\/\/stackoverflow.com\/questions\/48303166\/keras-typeerror-cant-pickle-thread-lock-objects-with-kerasclassifier\">1<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/44855603\/typeerror-cant-pickle-thread-lock-objects-in-seq2seq\">2<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/45081204\/keras-2-typeerror-cant-pickle-thread-lock-objects?noredirect=1&amp;lq=1\">3<\/a>) mention a similar error under different circumstances. Yet, the solutions offered there do not apply in my case.<\/p>\n\n<p><strong>EDIT 2<\/strong><\/p>\n\n<p>As suggested in the comments, I've added a <code>copy<\/code> method in the class. Would that be a viable solution?<\/p>\n\n<pre><code>class Klass:\n\n    def __init__(self, input_dims, output_dims, hidden_dims, optimizer, a, b):\n\n        self.input_dims = input_dims\n        self.output_dims = output_dims\n        self.hidden_dims = hidden_dims\n        self.optimizer = optimizer\n        self.a = a\n        self.b = b\n\n        self.__build_nn()\n\n    # [...]\n\n    def copy(self):\n\n        new = Klass(self.input_dims, self.output_dims, self.hidden_dims,\n                    self.optimizer, self.a, self.b)\n        new.nn1.set_weights(self.nn1.get_weights())\n        new.nn2.set_weights(self.nn2.get_weights())\n\n        return new\n<\/code><\/pre>\n","answers":[{"AnswerId":"52677957","CreationDate":"2018-10-06T10:11:01.493","ParentId":null,"OwnerUserId":"8125119","Title":null,"Body":"<p>Solved in comments: Added a <code>copy<\/code> method for <code>Klass<\/code> which copies the weights from the old <code>Klass<\/code> instance to the newly created one.<\/p>\n"}]}
+{"QuestionId":52673987,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-05T22:27:42.340","AcceptedAnswerId":null,"OwnerUserId":2243184.0,"Title":"Keras model.get_config() raises error in deepcopy (MaskRCNN)","Body":"<p>I am working with the MaskRCNN repo \n<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/mrcnn\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/mrcnn\/model.py<\/a><\/p>\n\n<p>Here in line 2064 in model.py before returning the built model back I want to extract the config using <code>model.get_config()<\/code><\/p>\n\n<p>I get the following error. Since the entire network is built before <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \".\/samples\/coco\/coco.py\", line 460, in &lt;module&gt;\n    model_dir=args.logs)\n  File \"\/mask_rcnn_root\/Mask_RCNN\/mrcnn\/model.py\", line 2063, in build\n    print(self.keras_model.get_config())\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/network.py\", line 926, in get_config\n    return copy.deepcopy(config)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in _deepcopy_tuple\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in &lt;listcomp&gt;\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in _deepcopy_tuple\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in &lt;listcomp&gt;\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in _deepcopy_tuple\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in &lt;listcomp&gt;\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 166, in deepcopy\n    y = copier(memo)\nTypeError: cannot deepcopy this pattern object\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52674455,"AnswerCount":0,"Tags":"<python-3.x><numpy><tensorflow><keras><sentiment-analysis>","CreationDate":"2018-10-05T23:40:06.357","AcceptedAnswerId":null,"OwnerUserId":6184683.0,"Title":"ValueError: Error when checking target: expected dense_2 to have shape (1,) but got array with shape (500,) [Sentiment Analysis]","Body":"<p>Research done before asking this question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/47873374\/error-when-checking-target-expected-dense-2-to-have-shape-none-256-but-got-a\">Error when checking target: expected dense_2 to have shape (None, 256) but got array with shape (16210, 4096)<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/49623436\/valueerror-error-when-checking-target-expected-dense-2-to-have-3-dimensions-b\">ValueError: Error when checking target: expected dense_2 to have 3 dimensions, but got array with shape (10000, 1)<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/51755160\/error-when-checking-target-expected-dense-3-to-have-shape-2-but-got-array-wi\">Error when checking target: expected dense_3 to have shape (2,) but got array with shape (1,)<\/a><\/p>\n\n<p>I have searched for a solution to this problem for days now. Please help me figure this out.<\/p>\n\n<pre><code>vocab_size = 5000\ndim = 32 \ninput_length_var = 500\nmodel = Sequential() \nmodel.add(Embedding(vocab_size, dim, input_length=input_length_var)) \nmodel.add(LSTM(100)) \nmodel.add(Dense(1, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy',optimizer='adam', metrics=['accuracy']) \nprint(model.summary()) \n<\/code><\/pre>\n\n<p>The above code is my model. I will now print the summary of above model:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #\n=================================================================\nembedding_1 (Embedding)      (None, 500, 1)            500\n_________________________________________________________________\nlstm_1 (LSTM)                (None, 100)               40800\n_________________________________________________________________\ndense_1 (Dense)              (None, 1)                 101\n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 2\n=================================================================\nTotal params: 41,403\nTrainable params: 41,403\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>And finally I will show you the result of np.shape():<\/p>\n\n<pre><code>(1117228, 500)\n(1117228, 500)\n<\/code><\/pre>\n\n<p>I have tried everything from Reshape() to adding input_shape to the dense layers but the result is always the same. What am I doing wrong and how to I fix this? My task is sentiment analysis.<\/p>\n\n<p>EDIT: I was told that the dimensions of output needed to be (1117228,1) and I needed sentiment scores in train_test_split for the labels. The first half of my csv is negative sentiment and the other half is positive sentiment. How would I use this?<\/p>\n","answers":[]}
+{"QuestionId":52674760,"AnswerCount":0,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-10-06T00:40:29.023","AcceptedAnswerId":null,"OwnerUserId":9997976.0,"Title":"Assign value to tensor according to another tensor","Body":"<p>I tried to use tf.assign to assign value to my tensor according to another tensor which has boolean value. The code is <\/p>\n\n<pre><code>self.labels_gt = self.labels_gt[self.gt_single_mask == False].assign(-1)\n<\/code><\/pre>\n\n<p>What I try to do is simple, assign <code>-1<\/code> to <code>self.labels_gt<\/code> if the the corresponding index in <code>self.gt_single_mask<\/code> is <code>False<\/code>. But it throws an error<\/p>\n\n<pre><code>TypeError: Value passed to parameter 'begin' has DataType bool not in list of allowed values: int32, int64\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52674779,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-06T00:47:59.187","AcceptedAnswerId":null,"OwnerUserId":2426955.0,"Title":"Convolution with complex Kernel Tensorflow","Body":"<p>I would like to perform a convolution with a complex input and complex kernel.<\/p>\n\n<p>I do not need back propagation, so I don't care whether gradients can be calculated or not.<\/p>\n\n<p>Is this possible in Tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":52674892,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-06T01:17:23.177","AcceptedAnswerId":52677111.0,"OwnerUserId":1090562.0,"Title":"How to group learnable parameters for tensorboard?","Body":"<p>I usually structure my learnable parameters in the following way in tensorflow:<\/p>\n\n<pre><code>learnable_weights = {\n  'w1': tf.get_variable(...),\n  ...\n  'wn': tf.get_variable(...),\n} \n\nlearnable_biases = {\n  'bc1': tf.get_variable(...),\n  ...\n  'bd3': tf.get_variable(...)\n}\n<\/code><\/pre>\n\n<p>The problem I started to recently encounter is the congested tensorboard graph, where I have a lot of weights in the auxilary nodes (this is a part of a big graph, and the number of these nodes is way bigger): <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5yR1a.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5yR1a.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I tried to group them with <code>tf.name_scope<\/code>. Something like this:<\/p>\n\n<pre><code>with tf.name_scope('learnable_params'):\n  learnable_weights = {...}\n  learnable_biases = {...}\n<\/code><\/pre>\n\n<p>But this has no effect on the graph in tensorboard.<\/p>\n\n<p>Any reason why or better any suggestions how to group the learnable parameters so that they will not clutter the tensorboard?<\/p>\n","answers":[{"AnswerId":"52677111","CreationDate":"2018-10-06T08:19:24.890","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>You could try using <code>variable_scope<\/code> instead of <code>name_scope<\/code>. AFAIK variables created via <code>get_variable<\/code> ignore <code>name_scope<\/code>, and I wouldn't be surprised if this applies to the graph organization in Tensorboard as well. I only use <code>variable_scope<\/code> to wrap anything that creates variables and I've never had these issues with \"unorganized\" variables.<\/p>\n"}]}
+{"QuestionId":52674895,"AnswerCount":2,"Tags":"<python><tensorflow><image-processing>","CreationDate":"2018-10-06T01:17:57.067","AcceptedAnswerId":null,"OwnerUserId":2137798.0,"Title":"My Tensorflow neural network is still pretty stupid, after much remedial education","Body":"<p>I am fairly new at image classification and tensor flow and have been engaged in a little toy project for self-education. It is an image classification project that sorts photographs of cars into 16 different classes. The photographs are taken from the Kaggle Carvana competition, which had different goals and evaluated according to a different criteria. <\/p>\n\n<p>What I am trying to do is find out if I can tell that an image of a car is taken with the car pointed directly at the camera, or at any one of 16 radial points around the vehicle. This should be a fairly straightforward image classification problem.<\/p>\n\n<p>What I am doing in essence is to see the images into a neural network that does two relu feed-forwards, runs an AdamOptimizer (or others, no better) then scores them.\nThe problem I have been having is that the network never seems to learn-it always shows a loss of right around 2.80 (sometimes a little higher, sometimes a little lower) and never ever seems to wise up. The performance is never much better than right around random guessing. I have been tweaking, muttering, and staring for a long time. I fear I am losing my sanity.<\/p>\n\n<p>The crucial code for the loss, training and evaluation operations is posted below. When I call this, the vital parameters are:<\/p>\n\n<pre><code>--target carvana --numclasses 16   --batch_size 64  --show True --learning_rate 0.01 --epochs 10 --scale 8.0 --hidden1_units 128 --hidden2_units 32\n<\/code><\/pre>\n\n<p>The complete code is at <a href=\"https:\/\/github.com\/WascallyWabbit\/estimator\" rel=\"nofollow noreferrer\">https:\/\/github.com\/WascallyWabbit\/estimator<\/a> . <\/p>\n\n<p>Thanks, \nWabbit<\/p>\n\n<pre><code>(image_placeholder, label_placeholder) = target.get_graph_placeholders(img_shape=IMG_SHAPE,\n                                                                       batch_size=FLAGS.batch_size)\n\nlogits_op = target.inference(images_placeholder=image_placeholder,\n                             hidden1_units=FLAGS.hidden1_units,\n                             hidden2_units=FLAGS.hidden2_units)\n\nloss_op = target.loss(logits=logits_op,\n                      labels=label_placeholder)\n\ntrain_op = target.training(learning_rate=FLAGS.learning_rate,\n                           loss_op=loss_op)\n\nevaluation_op = target.evaluation(logits=logits_op,\n                                  labels=label_placeholder)\n\n# blah, blah...\ndef get_graph_placeholders(self, img_shape=None, batch_size=10, num_classes=16):\n    if img_shape == None:\n        img_shape = self.img_shape\n\n    pixel_num = ut.pixnum_from_img_shape(img_shape)\n\n    images_placeholder = tf.placeholder(tf.float32, shape=(batch_size, pixel_num), name='Images')\n    labels_placeholder = tf.placeholder(tf.int32, shape=(batch_size), name='Labels')\n\n    return (images_placeholder, labels_placeholder)\n\ndef inference(self,\n              images_placeholder,\n              hidden1_units,\n              hidden2_units,\n              num_classes = 16,\n              img_shape=None):\n\n    if img_shape == None:\n        img_shape = self.img_shape\n    pixel_num = ut.pixnum_from_img_shape(img_shape)\n\n\n    with tf.name_scope('inference'):\n        #       with tf.name_scope('inference'):\n        #       display_tensor = tf.reshape(tensor=images_placeholder, shape=[100,28,28,1])\n        #      tf.summary.image(tensor=display_tensor, max_outputs=3,name=\"Carvana_images\")\n        display_tensor = tf.reshape(tensor=images_placeholder, shape=[64,img_shape[1],img_shape[0],1])\n        tf.summary.image(tensor=display_tensor, max_outputs=320,name=\"Carvana_images\")\n\n        with tf.name_scope('hidden1'):\n            weights1 = tf.Variable(\n                tf.truncated_normal([pixel_num, hidden1_units],\n                                          stddev=1.0\/math.sqrt(float(pixel_num))),\n                                          name='weights1')\n            biases1 = tf.Variable(tf.zeros([hidden1_units]),\n                                 name='biases1')\n\n            hidden1 = tf.nn.relu(tf.matmul(images_placeholder, weights1) + biases1)\n            tf.summary.histogram(name='weights1', values=weights1)\n            tf.summary.histogram(name='biases1', values=biases1)\n\n\n        with tf.name_scope('hidden2'):\n            weights2 = tf.Variable(\n                tf.truncated_normal([hidden1_units, hidden2_units],\n                                          stddev=1.0 \/ math.sqrt(float(hidden1_units))),\n                                          name='weights2')\n            biases2 = tf.Variable(tf.zeros([hidden2_units]),\n                                 name='biases2')\n\n            hidden2 = tf.nn.relu(tf.matmul(hidden1, weights2) + biases2)\n            tf.summary.histogram(name='weights2', values=weights2)\n            tf.summary.histogram(name='biases2', values=biases2)\n\n        with tf.name_scope('softmax_linear'):\n            weights3 = tf.Variable(\n                tf.truncated_normal([hidden2_units, num_classes],\n                                          stddev=1.0 \/ math.sqrt(float(hidden2_units))),\n                name='weights3')\n            biases3 = tf.Variable(tf.zeros([num_classes]),\n                                 name='biases3')\n\n            logits = tf.nn.softmax(tf.matmul(hidden2, weights3) + biases3)\n\n            tf.summary.histogram(name='weights3', values=weights3)\n            tf.summary.histogram(name='biases3', values=biases3)\n            tf.summary.histogram(name='logits', values=logits)\n\n        return logits\n\ndef loss(self,logits,labels):\n    with tf.name_scope('loser'):\n        labels=tf.to_int64(labels)\n        cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(\n            labels=labels, logits=logits, name='xentropy')\n\n        rm = tf.reduce_mean(cross_entropy, name='xentropy_mean')\n        tf.summary.scalar('xentropy_reduced_mean', rm)\n        return rm\n\ndef evaluation(self, logits, labels):\n    with tf.name_scope('evaluation'):\n        correct = tf.nn.in_top_k(logits,labels,1,name='correct_evaluation')\n    #    tf.summary.scalar('Evaluation', correct)\n        rs = tf.reduce_sum(tf.cast(correct,tf.int32), name='Reduce_sum')\n        tf.summary.scalar('Reduced sum', rs)\n        return correct\n\ndef training(self, loss_op, learning_rate):\n    with tf.name_scope('training'):\n        tf.summary.scalar('Training loss_op', loss_op)\n        optimizer = tf.train.AdamOptimizer(learning_rate, name='Adam_Optimizer')\n        global_step = tf.Variable(0, name='global_step', trainable=False)\n        tf.summary.scalar('Training global_step', global_step)\n        train_op = optimizer.minimize(loss_op, global_step=global_step)\n        return train_op\n\ndef init_weights(self,pixel_num,hidden1_units,hidden2_units,num_classes):\n    w1 = tf.Variable(\n        tf.truncated_normal([pixel_num, hidden1_units],\n                                  stddev=1.0 \/ math.sqrt(float(pixel_num)),\n                                  name='weights1')\n    )\n\n    w2 = tf.Variable(\n        tf.truncated_normal([hidden1_units, hidden2_units],\n                                  stddev=1.0 \/ math.sqrt(float(pixel_num)),\n                                  name='weights2')\n    )\n\n    w3 = tf.Variable(\n        tf.truncated_normal([hidden2_units, num_classes],\n                                  stddev=1.0 \/ math.sqrt(float(hidden2_units)),\n                                  name='weights3')\n    )\n\n    return (w1,w2,w3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52676667","CreationDate":"2018-10-06T07:19:18.740","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>It could be that your Network is too small. You are just using two hidden layers. The number of hidden units seem quite low. Try increasing your layer number and size. <\/p>\n\n<p>Another idea is to use CNN layers for your classification problem, since they tend to be better for image classification then normal feed forward neural networks. \nTake a look at vgg 16 for example<\/p>\n"},{"AnswerId":"52677143","CreationDate":"2018-10-06T08:24:42.380","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>While I agree with the other answer mentioning that such a small neural network will have a hard time with such a task, <em>especially<\/em> a fully-connected one, I think the major issue here is that you apply the <code>softmax<\/code> activation yourself to the output layer. This is not correct; when using the <code>cross_entropy_with_logits<\/code> functions, you need to pass the <em>pre-softmax<\/em> values. The softmax is done internally inside the cost function in a numerically stable way. Thus, you should remove <code>tf.nn.softmax<\/code> from the <code>logits<\/code> in the <code>inference<\/code> function.<\/p>\n"}]}
+{"QuestionId":52675607,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2018-10-06T04:06:59.593","AcceptedAnswerId":52675720.0,"OwnerUserId":5349476.0,"Title":"Keras utilises less CPU when number of workers grows and numpy generates a large array","Body":"<p>My code uses a relatively extensive augmentation strategy, but I've noticed CPU utilisation isn't proportionate when <code>N<\/code> in <code>fit_generator(...workers=N)<\/code> increases. I have a 4-core CPU.<\/p>\n\n<ul>\n<li>When <code>N=1<\/code>, <code>htop<\/code> shows around 105% usage<\/li>\n<li>When <code>N=2<\/code>, <code>htop<\/code> shows around 202% usage<\/li>\n<li>When <code>N=3<\/code>, <code>htop<\/code> shows around 287% usage<\/li>\n<li>When <code>N=4<\/code>, <code>htop<\/code> shows around 342% usage<\/li>\n<\/ul>\n\n<p>GPU usage is less than 40% throughout.<\/p>\n\n<p>If I trim down the augmentation strategy to omit noise addition, I can achieve around 360% and higher GPU usage when <code>N=4<\/code>. Noise is added by<\/p>\n\n<p><code>x += numpy.random.normal(0, noise_sigma, x.shape) \/ 255.0<\/code><\/p>\n\n<p>where <code>x<\/code> is a 640x480 BGR input image. It is a slow call, averaging around 24.3ms per call, but shouldn't the CPU still do the work when <code>N=4<\/code>? How come numpy seems to be blocking other threads when it generates a large array of random numbers?<\/p>\n","answers":[{"AnswerId":"52675720","CreationDate":"2018-10-06T04:32:34.713","ParentId":null,"OwnerUserId":"4754037","Title":null,"Body":"<p><code>normal<\/code> calls <code>cont2_array*<\/code><\/p>\n\n<p><a href=\"https:\/\/github.com\/numpy\/numpy\/blob\/master\/numpy\/random\/mtrand\/mtrand.pyx#L1651\" rel=\"nofollow noreferrer\">https:\/\/github.com\/numpy\/numpy\/blob\/master\/numpy\/random\/mtrand\/mtrand.pyx#L1651<\/a><\/p>\n\n<p>and there is a <code>lock<\/code> <\/p>\n\n<p>Is this the reason?<\/p>\n\n<p>Can you try to use individual <code>RandomState<\/code> to generate random numbers?<\/p>\n\n<pre><code>r = numpy.random.RandomState() \n\n.....\n\nfor ... :\n    x += r.normal(0, noise_sigma, x.shape) \/ 255.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":52677319,"AnswerCount":0,"Tags":"<theano><pymc3>","CreationDate":"2018-10-06T08:51:07.647","AcceptedAnswerId":null,"OwnerUserId":6717054.0,"Title":"Why isn\u2019t NUTS sampling with tt.dot or pm.math.dot?","Body":"<p>I am trying to implement parts of Facebook's prophet with some help from this example. <\/p>\n\n<p><a href=\"https:\/\/github.com\/luke14free\/pm-prophet\/blob\/master\/pmprophet\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/luke14free\/pm-prophet\/blob\/master\/pmprophet\/model.py<\/a><\/p>\n\n<p>This goes well :), but I am having some problems with the dot product I don't understand. Note that I am implementing the linear trends.<\/p>\n\n<pre><code>ds = pd.to_datetime(df['dagindex'], format='%d-%m-%y')\n\n\nm = pm.Model()\nchangepoint_prior_scale = 0.05\nn_changepoints = 25\nchangepoints = pd.date_range(\n    start=pd.to_datetime(ds.min()),\n    end=pd.to_datetime(ds.max()),\n    periods=n_changepoints + 2\n)[1: -1]\n\nwith m:\n\n    # priors\n    sigma = pm.HalfCauchy('sigma', 10, testval=1)\n    #trend\n    growth = pm.Normal('growth', 0, 10)\n\n    prior_changepoints = pm.Laplace('changepoints', 0, changepoint_prior_scale, shape=len(changepoints))\n\n    y = np.zeros(len(df))\n\n    # indexes x_i for the changepoints. \n    s = [np.abs((ds - i).values).argmin() for i in changepoints]\n\n    g = growth\n    x = np.arange(len(ds))\n    # delta\n    d = prior_changepoints\n\n    regression = x * g\n\n    base_piecewise_regression = []\n\n    for i in s:\n        local_x = x.copy()[:-i]\n        local_x = np.concatenate([np.zeros(i), local_x])\n        base_piecewise_regression.append(local_x)\n\n    piecewise_regression = np.array(base_piecewise_regression)\n\n#  this dot product doesn't work?\n    piecewise_regression = pm.math.dot(theano.shared(piecewise_regression).T, d)\n\n#  If I comment out this line and use that one as dot product. It works fine\n#     piecewise_regression = (piecewise_regression.T * d[None, :]).sum(axis=-1)\n    regression += piecewise_regression\n\n    y += regression\n\n    obs = pm.Normal('y',\n                   mu=(y - df.gebruikers.mean()) \/ df.gebruikers.std(),\n                   sd=sigma,\n                   observed=(df.gebruikers - df.gebruikers.mean()) \/ df.gebruikers.std())\n\n    start = pm.find_MAP(maxeval=10000)\n    trace = pm.sample(500, step=pm.NUTS(), start=start)\n<\/code><\/pre>\n\n<p>If I run the snippet above with <\/p>\n\n<p><code>piecewise_regression = (piecewise_regression.T * d[None, :]).sum(axis=-1)<\/code><\/p>\n\n<p>the model works as expected. However I cannot get it to work with a dot product. The NUTS sampler doesn't sample at all.<\/p>\n\n<p><code>piecewise_regression = pm.math.dot(theano.shared(piecewise_regression).T, d)<\/code><\/p>\n\n<h1>EDIT<\/h1>\n\n<p>Ive got a minimal working example<\/p>\n\n<p>The problem still occurs with theano.shared. I\u2019ve got a minimal working example:<\/p>\n\n<pre><code>np.random.seed(5)\n\nn_changepoints = 10\nt = np.arange(1000)\ns = np.sort(np.random.choice(t, size=n_changepoints, replace=False))\na = (t[:, None] &gt; s) * 1\n\nreal_delta = np.random.normal(size=n_changepoints)\ny = np.dot(a, real_delta) * t\n\nwith pm.Model():\n    sigma = pm.HalfCauchy('sigma', 10, testval=1)\n    delta = pm.Laplace('delta', 0, 0.05, shape=n_changepoints)\n    g = tt.dot(a, delta) * t\n    obs = pm.Normal('obs',\n                   mu=(g - y.mean()) \/ y.std(),\n                   sd=sigma,\n                   observed=(y - y.mean()) \/ y.std())\n    trace = pm.sample(500)\n<\/code><\/pre>\n\n<p>It seems to have something to do with the size of matrix a. NUTS doesnt\u2019t sample if I start with<\/p>\n\n<p>t = np.arange(1000)<\/p>\n\n<p>however the example above does sample when I reduce the size of t to:<\/p>\n\n<p>t = np.arange(100)<\/p>\n","answers":[]}
+{"QuestionId":52677852,"AnswerCount":2,"Tags":"<tensorflow><machine-learning>","CreationDate":"2018-10-06T09:58:10.267","AcceptedAnswerId":null,"OwnerUserId":9719840.0,"Title":"Evaluation and Visualization of DeeplabV3 completed or not?","Body":"<p>Disclaimer: First time i am trying Machine Learning!\nWe have a requirement of Automatic segmentation of a objects in an image from Background. Through internet we found that \"Deep lab\" will solve our purpose. we downloaded the deeplab from their offical site and followed all the instructions that they have mentioned. we trained the pascal_voc_2012 dataset with below command<\/p>\n\n<blockquote>\n  <p>python deeplab\/train.py \\\n      --logtostderr \\\n      --training_number_of_steps=30000 \\\n      --train_split=\"train\" \\\n      --model_variant=\"xception_65\" \\\n      --atrous_rates=6 \\\n      --atrous_rates=12 \\\n      --atrous_rates=18 \\\n      --output_stride=16 \\\n      --decoder_output_stride=4 \\\n      --train_crop_size=513 \\\n      --train_crop_size=513 \\\n      --train_batch_size=1 \\\n      --dataset=\"pascal_voc_seg\" \\\n      --tf_initial_checkpoint=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/checkpoint\n  \\\n      --train_logdir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/train$\n  \\\n      --dataset_dir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/tfrecord<\/p>\n<\/blockquote>\n\n<p>Training is done after 50 hours. Then i started the Evaluation using below command<\/p>\n\n<blockquote>\n  <p>python deeplab\/eval.py \\\n      --logtostderr \\\n      --eval_split=\"val\" \\\n      --model_variant=\"xception_65\" \\\n      --atrous_rates=6 \\\n      --atrous_rates=12 \\\n      --atrous_rates=18 \\\n      --output_stride=16 \\\n      --decoder_output_stride=4 \\\n      --eval_crop_size=513 \\\n      --eval_crop_size=513 \\\n      --dataset=\"pascal_voc_seg\" \\\n      --checkpoint_dir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/train\/\n  \\\n      --eval_logdir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/eval\/\n  \\\n      --dataset_dir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/tfrecord<\/p>\n<\/blockquote>\n\n<p>After executing the above command, it found one checkpoint correctly, but after that it stays with this message <\/p>\n\n<blockquote>\n  <p>\"Waiting for checkpoint at\n  home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/train\/\"<\/p>\n<\/blockquote>\n\n<p>So i terminated the execution of Eval after 2 hours and started the visualization with below command<\/p>\n\n<blockquote>\n  <p>python deeplab\/vis.py \\\n      --logtostderr \\\n      --vis_split=\"val\" \\\n      --model_variant=\"xception_65\" \\\n      --atrous_rates=6 \\\n      --atrous_rates=12 \\\n      --atrous_rates=18 \\\n      --output_stride=16 \\\n      --decoder_output_stride=4 \\\n      --vis_crop_size=513 \\\n      --vis_crop_size=513 \\\n      --dataset=\"pascal_voc_seg\" \\\n      --checkpoint_dir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/train\/\n  \\\n      --vis_logdir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/vis\/\n  \\\n      --dataset_dir=\/home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/tfrecord\/<\/p>\n<\/blockquote>\n\n<p>visualization also executed for one checkpoint and then again got the same message like Eval.<\/p>\n\n<blockquote>\n  <p>\"Waiting for checkpoint at\n  home\/ktpl13\/Desktop\/models-master\/research\/deeplab\/datasets\/pascal_voc_seg\/exp\/train_on_train_set\/train\/\"<\/p>\n<\/blockquote>\n\n<p>Again i terminated the execution of vis. There is a folder generated under vis with name \"segmentation_results\" which contains the \"prediction.png\" for each input image. which is \"completly black image\".<\/p>\n\n<p>Now My questions are.<\/p>\n\n<ol>\n<li>Did My Evaluation and visualization are done? or am i doing something wrong? <\/li>\n<li>Why the predicted images all are Black?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53124223","CreationDate":"2018-11-02T18:43:58.050","ParentId":null,"OwnerUserId":"2328141","Title":null,"Body":"<p>For future reference, I ran into the same problem. After I found out what happened I laughed so hard.<\/p>\n\n<p>Both eval and vis ran as expected.<\/p>\n\n<p>For eval, right above your output of \"waiting for checkpoints,\" there should be a line that says \"miou[your model accuracy here]\" It is a tiny line and easy to miss.<\/p>\n\n<p>For vis, you will find your segmented result in the vis logdir you provided in your vis command.<\/p>\n\n<p>More in depth, both eval and vis have successfully analyze the network your trained, and as a feature they are waiting for more checkpoints in case you decided to train more networks to compare. <\/p>\n"},{"AnswerId":"58307468","CreationDate":"2019-10-09T15:25:54.377","ParentId":null,"OwnerUserId":"7928366","Title":null,"Body":"<p>About the eval waiting for another checkpoint, it's because the default expects to run along with the train process. To run the eval script only once, after training, add this flag to the eval.sh script:<\/p>\n\n<p><code>--max_number_of_evaluations = 1<\/code><\/p>\n\n<p>And you can view the value using TensorBoard.<\/p>\n\n<p>The vis.sh script appears to be running correctly as it's saving the images to the directory. The issue with all black images is a different problem (e.g: dataset configuration, label weights, colormap removal, etc).<\/p>\n"}]}
+{"QuestionId":52677861,"AnswerCount":0,"Tags":"<python><pytorch><multi-gpu>","CreationDate":"2018-10-06T09:59:25.057","AcceptedAnswerId":null,"OwnerUserId":2444520.0,"Title":"Arguments are located on different GPUs when using nn.DataParallel(model)","Body":"<p>Torch 0.4.1<\/p>\n\n<p>Python 2.7.12<\/p>\n\n<p>I was adapting <a href=\"https:\/\/github.com\/priba\/nmp_qc\" rel=\"nofollow noreferrer\">NMP QC code<\/a> <a href=\"https:\/\/github.com\/priba\/nmp_qc\/pull\/5\/commits\/b59efc524827d62ec46f8c85c19849c30fadb652\" rel=\"nofollow noreferrer\">(with some compatibility issues ironed out)<\/a> to use multiple GPUs since my GPU couldn't handle the workload (crashed after running out of VRAM)<\/p>\n\n<p>I'm new to pytorch, but I found <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/data_parallel_tutorial.html\" rel=\"nofollow noreferrer\">a tutorial on using nn.DataParallel(model)<\/a> to implement multi-gpu use<\/p>\n\n<p>I <a href=\"https:\/\/pastebin.com\/w5fvQ3VS\" rel=\"nofollow noreferrer\">modified main.py to use nn.DataParallel(model).<\/a> Areas I changed have \"#NEW\" stuck to them.<\/p>\n\n<p>The code runs fine even in multi-gpu mode if running on a single GPU, but gets a \"arguments are located on different GPUs\" error when running on 2 or more GPU<\/p>\n\n<pre><code>libibverbs: Warning: couldn't open config directory '\/etc\/libibverbs.d'.\nlibibverbs: Warning: no userspace device-specific driver found for \/sys\/class\/infiniband_verbs\/uverbs3\nlibibverbs: Warning: no userspace device-specific driver found for \/sys\/class\/infiniband_verbs\/uverbs2\nlibibverbs: Warning: no userspace device-specific driver found for \/sys\/class\/infiniband_verbs\/uverbs1\nlibibverbs: Warning: no userspace device-specific driver found for \/sys\/class\/infiniband_verbs\/uverbs0\nUnexpected end of \/proc\/mounts line `overlay \/ overlay rw,relatime,lowerdir=\/var\/lib\/docker\/overlay2\/l\/QKSBQ5PAFDDC3OMBEELQQETALQ:\/var\/lib\/docker\/overlay2\/l\/WWYI3IDQPNXGON7AHODBPSTVXL:\/var\/lib\/docker\/overlay2\/l\/Q54I2HYS4TKH4LDJKBTVTGWWO6:\/var\/lib\/docker\/overlay2\/l\/IUV2LFPNMPOS3MREOTT52TKL54:\/var\/lib\/docker\/overlay2\/l\/DB5GBUCI3DCBPX6TJG3O337YVB:\/var\/lib\/docker\/overlay2\/l\/DNYKXCZJH5FMFNJLNGYJJ2ITPI:\/var\/lib\/docker\/overlay2\/l\/7DZCTDVNSTPJISGW65UG7U3F75:\/var\/lib\/docker\/overlay2\/l\/VOEQO652VS63NLDLZZ4TCIJLO6:\/var\/lib\/docker\/overlay2\/l\/4SI6ZCRUIORG5'\nTraceback (most recent call last):\n  File \"main.py\", line 332, in &lt;module&gt;\n    main()\n  File \"main.py\", line 190, in main\n    train(train_loader, model, criterion, optimizer, epoch, evaluation, logger)\n  File \"main.py\", line 251, in train\n    output = model(g, h, e)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/torch\/nn\/parallel\/data_parallel.py\", line 123, in forward\n    outputs = self.parallel_apply(replicas, inputs, kwargs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/torch\/nn\/parallel\/data_parallel.py\", line 133, in parallel_apply\n    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 77, in parallel_apply\n    raise output\nRuntimeError: arguments are located on different GPUs at \/pytorch\/aten\/src\/THC\/generic\/THCTensorMathBlas.cu:236\n<\/code><\/pre>\n\n<p>Since I was sending the inputs one at a time instead of at once like in the tutorial, I checked by using .get_device(), which confirmed all 4 arguments being sent (g, h, e, target) were on the same device (device 0)<\/p>\n","answers":[]}
+{"QuestionId":52678048,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><regression>","CreationDate":"2018-10-06T10:22:28.897","AcceptedAnswerId":null,"OwnerUserId":9545198.0,"Title":"Reproducing results using keras with tensorflow background","Body":"<p>I am building a simple Neural Network using keras with tensorflow background. I am new to ML. <\/p>\n\n<p>My aim is to get reproducible results. I have read various blogs to do this and have implemented something like below:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport pandas as pd\n\nfrom numpy.random import seed\nseed(1)\nfrom tensorflow import set_random_seed\nset_random_seed(2)\n\n# Parameters\nlearning_rate = .001\ntraining_epochs = 2000\ndisplay_step = 5\n\ntrain = pd.read_csv('T:\/2.0 Pricing\/45. Machine learning\/R Training\/Training-2017\/Prince\/Kaggle House Price Competition\/House Prices - Advanced Regression Techniques\/Data\/train.csv')\ntest = pd.read_csv('T:\/2.0 Pricing\/45. Machine learning\/R Training\/Training-2017\/Prince\/Kaggle House Price Competition\/House Prices - Advanced Regression Techniques\/Data\/test.csv')\n\ntrain=train.loc[train['GrLivArea'] &lt;4000]\ntrain=train.loc[train['TotalBsmtSF'] &lt;5000]\ntrain=train.loc[train['1stFlrSF'] &lt;3000]\ntrain=train.loc[train['2ndFlrSF'] &lt;2000]\ntrain=train.loc[train['LotArea'] &lt;80000]\ntrain=train.loc[train['BsmtFinSF1'] &lt;5000]\n\ntrain_numeric=train.select_dtypes(exclude=['object'])\ntrain_catagorical=train.select_dtypes(include=['object'])\ntrain_catagorical=train_catagorical.apply(lambda col: pd.factorize(col, sort=True)[0])\ntrain_variables=train_catagorical.join(train_numeric, how='outer')\n#train=pd.get_dummies(train)\n#train_variables=train.select_dtypes(include=variables)\n#train_variables=train[variables]\ntrain_variables=train_variables.fillna(train_variables.mean())\ntrain_variables = train_variables.drop(labels=['SalePrice','Id'], axis=1)\ntrain_variables=pd.DataFrame(train_variables,dtype='float32')\n#train_variables=pd.DataFrame(train_variables[variablestouse],dtype='float32')\ntrain_response=pd.DataFrame(train[['SalePrice']],dtype='float32')\n\ntest_numeric=test.select_dtypes(exclude=['object'])\ntest_catagorical=test.select_dtypes(include=['object'])\ntest_catagorical=test_catagorical.apply(lambda col: pd.factorize(col, sort=True)[0])\ntest_variables=test_catagorical.join(test_numeric, how='outer')\n#test=pd.get_dummies(test)\n#variables=list(test)\n#test_variables=test.select_dtypes(exclude=[variables])\ntest_variables=test_variables.fillna(test_variables.mean())\ntest_variables = test_variables.drop(labels=['Id'], axis=1)\n#test_variables=pd.DataFrame(test_variables[variablestouse],dtype='float32')\ntest_variables=pd.DataFrame(test_variables,dtype='float32')\n\nn_samples = train.shape[0]\n\n# Model\nmodel = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Dense(100, input_dim=79, kernel_initializer='normal'))\nmodel.add(tf.keras.layers.Dense(50, kernel_initializer='normal', activation='relu'))\nmodel.add(tf.keras.layers.Dense(25, kernel_initializer='normal', activation='relu'))\nmodel.add(tf.keras.layers.Dense(1, kernel_initializer='normal'))\n\n# Compile model\nmodel.compile(loss='mean_squared_error', optimizer=tf.keras.optimizers.Adam())\n\nfeature_cols = train_variables\nlabels = train_response\n\n#np.random.seed(42)\nmodel.fit(train_variables, train_response, epochs=50, batch_size=100)\n\n# Predictions\nfeature_cols_test = test_variables\n\ny = model.predict(np.array(feature_cols_test))\n<\/code><\/pre>\n\n<p>However, I am still unable to reproduce results. Any suggestions on reproducing the results?<\/p>\n","answers":[]}
+{"QuestionId":52678215,"AnswerCount":5,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-06T10:41:16.133","AcceptedAnswerId":null,"OwnerUserId":4547005.0,"Title":"Find input that maximises output of a neural network using Keras and TensorFlow","Body":"<p>I have used Keras and TensorFlow to classify the <a href=\"https:\/\/github.com\/zalandoresearch\/fashion-mnist\" rel=\"nofollow noreferrer\">Fashion MNIST<\/a> following <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">this tutorial<\/a> .<\/p>\n\n<p>It uses the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">AdamOptimizer<\/a> to find the value for model parameters that minimize the loss function of the network. The input for the network is a 2-D tensor with shape [28, 28], and output is a 1-D tensor with shape [10] which is the result of a softmax function.<\/p>\n\n<p>Once the network has been trained, I want to use the optimizer for another task: find an input that maximizes one of the elements of the output tensor. How can this be done? Is it possible to do so using Keras or one have to use a lower level API?<\/p>\n\n<p>Since the input is not unique for a given output, it would be even better if we could impose some constraints on the values the input can take.<\/p>\n\n<p>The trained model has the following format<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Flatten(input_shape=(28, 28)),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52678529","CreationDate":"2018-10-06T11:21:57.343","ParentId":null,"OwnerUserId":"6145828","Title":null,"Body":"<p>Interesting. Maybe a solution would be to feed all your data to the network and for each sample save the <code>output_layer<\/code> after <code>softmax<\/code>.<\/p>\n\n<p>This way, for <em>3 classes<\/em>, where you want to find the best input for <em>class 1<\/em>, you are looking for outputs where the first component is high. For example: <code>[1 0 0]<\/code><\/p>\n\n<p>Indeed the output means the probability, or the confidence of the network, for the sample being one of the classes.<\/p>\n"},{"AnswerId":"52678833","CreationDate":"2018-10-06T11:59:22.770","ParentId":null,"OwnerUserId":"5055862","Title":null,"Body":"<p>I feel you would want to backprop with respect to the input freezing all the weights to your model. What you could do is:<\/p>\n\n<ol>\n<li>Add a dense layer after the input layer with the same dimensions as input and set it as trainable<\/li>\n<li>Freeze all the other layers of your model. (except the one you added) <\/li>\n<li>As an input, feed an identity matrix and train your model based on whatever output you desire.<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html\" rel=\"nofollow noreferrer\">This<\/a> article and <a href=\"https:\/\/stackoverflow.com\/questions\/43353851\/tune-input-features-using-backprop-in-keras\">this<\/a> post might be able to help you if you want to backprop based on the input instead. It's a bit like what you are aiming for but you can get the intuition. <\/p>\n"},{"AnswerId":"52683222","CreationDate":"2018-10-06T20:48:25.817","ParentId":null,"OwnerUserId":"5707705","Title":null,"Body":"<p>Funny coincident I was just working on the same \"problem\". I'm interested in the direction of adversarial training etc. What I did was to insert a <code>LocallyConnected2D<\/code> Layer after the input and then train with data which is all one and has as targets the class of interest.<\/p>\n\n<p>As model I use<\/p>\n\n<pre><code>batch_size = 64\nnum_classes = 10\nepochs = 20\ninput_shape = (28, 28, 1)\n\n\ninp = tf.keras.layers.Input(shape=input_shape)\nconv1 = tf.keras.layers.Conv2D(32, kernel_size=(3, 3),activation='relu',kernel_initializer='he_normal')(inp)\npool1 = tf.keras.layers.MaxPool2D((2, 2))(conv1)\ndrop1 = tf.keras.layers.Dropout(0.20)(pool1)\nflat  = tf.keras.layers.Flatten()(drop1)\nfc1   = tf.keras.layers.Dense(128, activation='relu')(flat)\nnorm1 = tf.keras.layers.BatchNormalization()(fc1)\ndropfc1 = tf.keras.layers.Dropout(0.25)(norm1)\nout   = tf.keras.layers.Dense(num_classes, activation='softmax')(dropfc1)\n\nmodel = tf.keras.models.Model(inputs = inp , outputs = out)\n\nmodel.compile(loss=tf.keras.losses.categorical_crossentropy,\n              optimizer=tf.keras.optimizers.RMSprop(),\n              metrics=['accuracy'])\nmodel.summary()\n<\/code><\/pre>\n\n<p>after training I insert the new layer<\/p>\n\n<pre><code>def insert_intermediate_layer_in_keras(model,position, before_layer_id):\n    layers = [l for l in model.layers]\n\n    if(before_layer_id==0) :\n        x = new_layer\n    else:\n        x = layers[0].output\n    for i in range(1, len(layers)):\n        if i == before_layer_id:\n            x = new_layer(x)\n            x = layers[i](x)\n\n        else:\n            x = layers[i](x)\n\n    new_model = tf.keras.models.Model(inputs=layers[0].input, outputs=x)\n    return new_model\n\ndef fix_model(model):\n    for l in model.layers:\n        l.trainable=False\n\n\nfix_model(model)    \nnew_layer = tf.keras.layers.LocallyConnected2D(1, kernel_size=(1, 1),\n                                               activation='linear',\n                                               kernel_initializer='he_normal',\n                                                use_bias=False)\nnew_model = insert_intermediate_layer_in_keras(model,new_layer,1)\nnew_model.compile(loss=tf.keras.losses.categorical_crossentropy,\n              optimizer=tf.keras.optimizers.RMSprop(),\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>and finally rerun training with my fake data. <\/p>\n\n<pre><code>X_fake = np.ones((60000,28,28,1))\nprint(Y_test.shape)\ny_fake = np.ones((60000))\nY_fake = tf.keras.utils.to_categorical(y_fake, num_classes)\nnew_model.fit(X_fake, Y_fake, epochs=100)\nweights = new_layer.get_weights()[0]\n\nimshow(weights.reshape(28,28))\nplt.show()\n<\/code><\/pre>\n\n<p>Results are not yet satisfying but I'm confident of the approach and guess I need to play around with the optimiser.<\/p>\n"},{"AnswerId":"52689582","CreationDate":"2018-10-07T14:42:33.110","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>It would be very similar to the way that filters of a Convolutional Network is visualized: we would do <strong>gradient ascent<\/strong> optimization in <strong>input space<\/strong> to <strong>maximize the response<\/strong> of a particular filter.<\/p>\n\n<p>Here is how to do it: after training is finished, first we need to specify the output and define a loss function that we want to maximize:<\/p>\n\n<pre><code>from keras import backend as K\n\noutput_class = 0 # the index of the output class we want to maximize\noutput = model.layers[-1].output\nloss = K.mean(output[:,output_class]) # get the average activation of our desired class over the batch\n<\/code><\/pre>\n\n<p>Next, we need to take the gradient of the loss we have defined above with respect to the input layer:<\/p>\n\n<pre><code>grads = K.gradients(loss, model.input)[0] # the output of `gradients` is a list, just take the first (and only) element\n\ngrads = K.l2_normalize(grads) # normalize the gradients to help having an smooth optimization process\n<\/code><\/pre>\n\n<p>Next, we need to define a backend function that takes the initial input image and gives the values of loss and gradients as outputs, so that we can use it in the next step to implement the optimization process:<\/p>\n\n<pre><code>func = K.function([model.input], [loss, grads])\n<\/code><\/pre>\n\n<p>Finally, we implement the gradient ascent optimization process:<\/p>\n\n<pre><code>import numpy as np\n\ninput_img = np.random.random((1, 28, 28)) # define an initial random image\n\nlr = 1.  # learning rate used for gradient updates\nmax_iter = 50  # number of gradient updates iterations\nfor i in range(max_iter):\n    loss_val, grads_val = func([input_img])\n    input_img += grads_val * lr  # update the image based on gradients\n<\/code><\/pre>\n\n<p>Note that, after this process is finished, to display the image you may need to make sure that all the values in the image are in the range [0, 255] (or [0,1]).<\/p>\n"},{"AnswerId":"52680776","CreationDate":"2018-10-06T15:49:07.920","ParentId":null,"OwnerUserId":"4547005","Title":null,"Body":"<p>After the hints Saket Kumar Singh gave in his answer, I wrote the following that seems to solve the question.<\/p>\n\n<p>I create two custom layers. Maybe Keras offers already some classes that are equivalent to them.<\/p>\n\n<p>The first on is a trainable input:<\/p>\n\n<pre><code>class MyInputLayer(keras.layers.Layer):\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(MyInputLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.kernel = self.add_weight(name='kernel',\n                                      shape=self.output_dim,\n                                      initializer='uniform',\n                                      trainable=True)\n        super(MyInputLayer, self).build(input_shape)\n\n    def call(self, x):\n        return self.kernel\n\n    def compute_output_shape(self, input_shape):\n        return self.output_dim\n<\/code><\/pre>\n\n<p>The second one gets the probability of the label of interest:<\/p>\n\n<pre><code>class MySelectionLayer(keras.layers.Layer):\n    def __init__(self, position, **kwargs):\n        self.position = position\n        self.output_dim = 1\n        super(MySelectionLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        super(MySelectionLayer, self).build(input_shape)\n\n    def call(self, x):\n        mask = np.array([False]*x.shape[-1])\n        mask[self.position] = True\n        return tf.boolean_mask(x, mask,axis=1)\n\n    def compute_output_shape(self, input_shape):\n        return self.output_dim\n<\/code><\/pre>\n\n<p>I used them in this way:<\/p>\n\n<pre><code># Build the model\nlayer_flatten =  keras.layers.Flatten(input_shape=(28, 28))\nlayerDense1 = keras.layers.Dense(128, activation=tf.nn.relu)\nlayerDense2 = keras.layers.Dense(10, activation=tf.nn.softmax)\nmodel = keras.Sequential([\n    layer_flatten,\n    layerDense1,\n    layerDense2\n])\n\n# Compile the model\nmodel.compile(optimizer=tf.train.AdamOptimizer(),\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\n# Train the model\n# ...\n\n# Freeze the model\nlayerDense1.trainable = False\nlayerDense2.trainable = False\n\n# Build another model\nclass_index = 7\n\nlayerInput =  MyInputLayer((1,784))\nlayerSelection = MySelectionLayer(class_index)\n\nmodel_extended = keras.Sequential([\n    layerInput,\n    layerDense1,\n    layerDense2,\n    layerSelection\n])\n\n# Compile it\nmodel_extended.compile(optimizer=tf.train.AdamOptimizer(),\n              loss='mean_absolute_error')\n\n# Train it\ndummyInput = np.ones((1,1))\ntarget = np.ones((1,1))\nmodel_extended.fit(dummyInput, target,epochs=300)\n\n# Retrieve the weights of layerInput\nlayerInput.get_weights()[0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52679454,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2018-10-06T13:17:01.683","AcceptedAnswerId":null,"OwnerUserId":10465762.0,"Title":"ModuleNotFoundError: No module named 'numpy.testing._private","Body":"<p>I'm trying to run tensorflow:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Chnoq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Chnoq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>it turned out <\/p>\n\n<pre><code>from numpy.testing._private.pytesttester import PytestTester\n<\/code><\/pre>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'numpy.testing._private'<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QJ6Vc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QJ6Vc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>numpy\/<strong>init<\/strong>.py look like this:\n<a href=\"https:\/\/i.stack.imgur.com\/gCZ8N.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gCZ8N.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52781476","CreationDate":"2018-10-12T14:16:45.813","ParentId":null,"OwnerUserId":"7458186","Title":null,"Body":"<p>I had the same problem when I installed tensorflow. Solved the issue by upgrading pip :<\/p>\n\n<pre><code>python.exe -m pip install --upgrade pip\n<\/code><\/pre>\n\n<p>and installing numpy again :<\/p>\n\n<pre><code>pip uninstall numpy\npip install numpy\n<\/code><\/pre>\n"}]}
+{"QuestionId":52679712,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-06T13:48:00.690","AcceptedAnswerId":null,"OwnerUserId":569313.0,"Title":"Correct Shape for Keras - Non Image 1D CNN","Body":"<p>I am trying to cement my understanding of using a 1D for text and time series classification using Keras \/ TF. Can anyone confirm or correct my understanding?<\/p>\n\n<p><strong>Text Embedding<\/strong><\/p>\n\n<p>In the example using a 1D cnn with an embedding <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/imdb_cnn.py\" rel=\"nofollow noreferrer\">keras example<\/a>, the input going into the 1D CNN is (samples, vocab size, embedding size). Here I think of this as for a given sample\/observation, the input is a matrix where the rows are the terms\/words and the columns are the series that describe the words (i.e. the embedding representation). Concretely, lets say there are 10 words and the embeddings are of length 5. So each sample is 10 x 5.<\/p>\n\n<p>Then, each kernel is  x 5. Here, filter length is a misnomer it seems because it describes the height of the kernel. The length of the kernel is always the same size as the embeddings (columns of the input matrix for each sample).<\/p>\n\n<p>The kernel moves down the input matrix and thus looks at groups of contiguous words.<\/p>\n\n<p><strong>Is this Correct?<\/strong><\/p>\n\n<p><strong>Time Series<\/strong><\/p>\n\n<p>In this setup, the input for each sample\/observation is also a matrix, but where the columns are the series and the rows are the values of the series (cell 1,1 is the first temporal value of the first series. The cell 3,8 is the 3rd temporal value of the 8th series etc.). As a concrete example, say there are 36 values of 10 series. The input matrix for each sample\/observation is 36 x 10.<\/p>\n\n<p>Then, given a kernel that is  x 10 (where 10 is fixed to the number of series) the kernel moves down the matrix, where unlike the text embedding, where word representations are combined, here it combines the temporal representation of the series. By this I mean if the kernel is 2 x 10 and the series are measured per day, it combines the first 2 days of all 10 series at once and then continues...<\/p>\n\n<p><strong>Is this correct?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":52680435,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-06T15:11:40.667","AcceptedAnswerId":null,"OwnerUserId":104668.0,"Title":"Tensorflow 1.11 hangs on \"sess = tf.Session()\"","Body":"<p>I updated to TensorFlow 1.11 from 1.10 and was unable to run a health-check. The below code hangs on <code>sess = tf.Session()<\/code>.<\/p>\n\n<p>Config: Windows 10 x64, latest GPU drivers, Cuda 9.0, cudnn 7.3, Python 3.6.6. This same setup used to work for TF 1.10, and all of a sudden everything stopped.<\/p>\n\n<p>Did anybody else encounter a similar situation?<\/p>\n\n<pre><code>import os\nimport numpy as np\nimport tensorflow as tf\n\ndef tftest(a, b):\n    # Create 100 phony x, y data points in NumPy, y = x * 0.1 + 0.3\n    x_data = np.random.rand(100).astype(np.float32)\n    y_data = x_data * a + b\n\n    # Try to find values for W and b that compute y_data = W * x_data + b\n    # (We know that W should be 0.1 and b 0.3, but TensorFlow will\n    # figure that out for us.)\n    W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))\n    b = tf.Variable(tf.zeros([1]))\n    y = W * x_data + b\n\n    # Minimize the mean squared errors.\n    loss = tf.reduce_mean(tf.square(y - y_data))\n    optimizer = tf.train.GradientDescentOptimizer(0.5)\n    train = optimizer.minimize(loss)\n\n    # Before starting, initialize the variables.  We will 'run' this first.\n    init = tf.global_variables_initializer()\n\n    # Launch the graph.\n    sess = tf.Session()   #&lt;&lt; the script hangs here!\n    sess.run(init)\n\n    # Fit the line.\n    for step in range(201):\n        sess.run(train)\n        if step % 20 == 0:\n            print(step, sess.run(W), sess.run(b))\n\nif __name__ == \"__main__\":\n    tftest(20, 10)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52908208","CreationDate":"2018-10-20T17:21:04.107","ParentId":null,"OwnerUserId":"104668","Title":null,"Body":"<p>I think it's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18652\" rel=\"nofollow noreferrer\">this bug<\/a>. The issue is still open, seems to be related to an issue in Eigen when built with MSVC.<\/p>\n\n<p>The good news is that TensorFlow does not hang indefinitely: one has to wait for 3-4-5 minutes and then the codes resumes. This only happens once, on first startup of TensorFlow.<\/p>\n"}]}
+{"QuestionId":52680886,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><neural-network><regression><loss-function>","CreationDate":"2018-10-06T16:01:08.573","AcceptedAnswerId":null,"OwnerUserId":10466241.0,"Title":"Mean absolute relative error as loss function in tensorflow (regression problem)","Body":"<p>I am solving a regression problem using neural network. \nMy target output ranges from 10-2 to 10-7. After running a lot of cases I was wondering if the mean squared error loss function can be replaced by a loss function which considers mean absolute relative error. I searched the tensorflow website and found the metric - <code>tf.metrics.mean_relative_error<\/code> but couldn't find anything under tf.losses. Is there any way to implement a loss function which uses relative error within tensorflow?<\/p>\n","answers":[{"AnswerId":"52680989","CreationDate":"2018-10-06T16:15:31.440","ParentId":null,"OwnerUserId":"5763590","Title":null,"Body":"<blockquote>\n  <p>Almost in all tensorflow tutorials they use custom functions. <\/p>\n<\/blockquote>\n\n<p>For example in the very <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">beginning tutorial<\/a> they write a custom function, which\nsums the squares of the deltas between the current model and the provided data<\/p>\n\n<pre><code>squared_deltas = tf.square(linear_model - y)\nloss = tf.reduce_sum(squared_deltas)\n<\/code><\/pre>\n\n<p>As you see it is not that hard at all: you just need to encode your function in a tensor-format and use their basic functions.<\/p>\n\n<p><em>Note:<\/em> Just as a little bonus, you might want to think about normalizing your inputs and expected outputs in a more suitable range e.g. [-0.5, 0.5] which works better for a handful of activation functions which otherwise need to work in a saturated area.<\/p>\n"}]}
+{"QuestionId":52680925,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-06T16:07:27.150","AcceptedAnswerId":null,"OwnerUserId":445114.0,"Title":"Tensorflow: How to create a Tensor with specific values?","Body":"<p>I have a Tesnor with shape NxM.<\/p>\n\n<p>I'd like to create another Tensor with the same shape, filled with ones up until a certain column (might be different for each row) and the rest of it filled with another value (let's say 10 for the example).<\/p>\n\n<p>How I do that?\nThanks.<\/p>\n","answers":[{"AnswerId":"52682835","CreationDate":"2018-10-06T19:55:39.233","ParentId":null,"OwnerUserId":"8315956","Title":null,"Body":"<p>Something like this can help you:<\/p>\n\n<pre><code>input = tf.Variable([[1, 2, 3, 4, 5], [1, 2, 3, 4, 5], [1, 2, 3, 4, 5]], dtype = tf.float32)\nindices = tf.constant([1, 4, 2])\nX = tf.ones_like(input)\nY = tf.constant(10, dtype = tf.float32, shape = input.shape)\nresult = tf.where(tf.sequence_mask(indices, tf.shape(input)[1]), X, Y)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(input))\n    print(sess.run(indices))\n    print(sess.run(result))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52681601,"AnswerCount":1,"Tags":"<python><machine-learning><keras><time-series><forecasting>","CreationDate":"2018-10-06T17:28:23.073","AcceptedAnswerId":52681743.0,"OwnerUserId":9261771.0,"Title":"Keras 2D input to 2D output","Body":"<p>First, I have read <a href=\"https:\/\/stackoverflow.com\/questions\/42154792\/keras-custom-layer-2d-input-2d-output\">this<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/36293208\/keras-2d-input-2d-output\">this<\/a> questions with similar names to mine and still do not have an answer.<\/p>\n\n<p>I want to build a feedforward network for sequence prediction. (I realize that RNNs are more suitable for this task, but I have my reasons). The sequences are of length 128 and each element is a vector with 2 entries, so each batch should be of shape <code>(batch_size, 128, 2)<\/code> and the target is the next step in the sequence, so the target tensor should be of shape <code>(batch_size, 1, 2)<\/code>.<\/p>\n\n<p>The network architecture is something like this:<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Dense(50, batch_input_shape=(None, 128, 2), kernel_initializer=\"he_normal\" ,activation=\"relu\"))\n    model.add(Dense(20, kernel_initializer=\"he_normal\", activation=\"relu\"))\n    model.add(Dense(5, kernel_initializer=\"he_normal\", activation=\"relu\"))\n    model.add(Dense(2))\n<\/code><\/pre>\n\n<p>But trying to train I get the following error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_4 to have shape (128, 2) but got array with shape (1, 2)\n<\/code><\/pre>\n\n<p>I've tried variations like:<\/p>\n\n<pre><code>model.add(Dense(50, input_shape=(128, 2), kernel_initializer=\"he_normal\" ,activation=\"relu\"))\n<\/code><\/pre>\n\n<p>but get the same error.<\/p>\n","answers":[{"AnswerId":"52681743","CreationDate":"2018-10-06T17:44:53.400","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you take a look at the <code>model.summary()<\/code> output you will see that what the issue is:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ndense_13 (Dense)             (None, 128, 50)           150       \n_________________________________________________________________\ndense_14 (Dense)             (None, 128, 20)           1020      \n_________________________________________________________________\ndense_15 (Dense)             (None, 128, 5)            105       \n_________________________________________________________________\ndense_16 (Dense)             (None, 128, 2)            12        \n=================================================================\nTotal params: 1,287\nTrainable params: 1,287\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>As you can see, the output of the model is <code>(None, 128,2)<\/code> and not <code>(None, 1, 2)<\/code> (or <code>(None, 2)<\/code>) as you expected. So, you may or may not know that <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">Dense layer is applied on the last axis of its input array<\/a> and as a result, as you see above, the time axis and dimension is preserved until the end.<\/p>\n\n<p>How to resolve this? You mentioned you don't want to use a RNN layer, therefore you have two options: you need to either use <code>Flatten<\/code> layer somewhere in the model or you can also use some Conv1D + Pooling1D layers or even a GlobalPooling layer. For example (these are just for demonstration, you may do it differently):<\/p>\n\n<p><strong>using <code>Flatten<\/code> layer<\/strong><\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(Dense(50, batch_input_shape=(None, 128, 2), kernel_initializer=\"he_normal\" ,activation=\"relu\"))\nmodel.add(Dense(20, kernel_initializer=\"he_normal\", activation=\"relu\"))\nmodel.add(Dense(5, kernel_initializer=\"he_normal\", activation=\"relu\"))\nmodel.add(Flatten())\nmodel.add(Dense(2))\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>Model summary:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense_17 (Dense)             (None, 128, 50)           150       \n_________________________________________________________________\ndense_18 (Dense)             (None, 128, 20)           1020      \n_________________________________________________________________\ndense_19 (Dense)             (None, 128, 5)            105       \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 640)               0         \n_________________________________________________________________\ndense_20 (Dense)             (None, 2)                 1282      \n=================================================================\nTotal params: 2,557\nTrainable params: 2,557\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p><strong>using <code>GlobalAveragePooling1D<\/code> layer<\/strong><\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(Dense(50, batch_input_shape=(None, 128, 2), kernel_initializer=\"he_normal\" ,activation=\"relu\"))\nmodel.add(Dense(20, kernel_initializer=\"he_normal\", activation=\"relu\"))\nmodel.add(GlobalAveragePooling1D())\nmodel.add(Dense(5, kernel_initializer=\"he_normal\", activation=\"relu\"))\nmodel.add(Dense(2))\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>\u200bModel summary:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense_21 (Dense)             (None, 128, 50)           150       \n_________________________________________________________________\ndense_22 (Dense)             (None, 128, 20)           1020      \n_________________________________________________________________\nglobal_average_pooling1d_2 ( (None, 20)                0         \n_________________________________________________________________\ndense_23 (Dense)             (None, 5)                 105       \n_________________________________________________________________\ndense_24 (Dense)             (None, 2)                 12        \n=================================================================\nTotal params: 1,287\nTrainable params: 1,287\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Note that in both cases above you need to reshape the labels (i.e. targets) array to <code>(n_samples, 2)<\/code> (or you may want to use a <code>Reshape<\/code> layer at the end).<\/p>\n"}]}
+{"QuestionId":52682457,"AnswerCount":0,"Tags":"<python-3.x><keras>","CreationDate":"2018-10-06T19:07:19.530","AcceptedAnswerId":null,"OwnerUserId":9045339.0,"Title":"Data types appropriate for Keras .flow_from_directory?","Body":"<p>I was hoping to find a list of \"data types\" (not sure of appropriate wording but example .jpg, .tif, etc) that can be fed into the .flow_from_directory in keras. Right now I am at work and cant test this but my \"images\" are saved as tensors size (#, width, length, height, other(pet or ct w\/ contrast)) in \".mat\" file. . pet =  , ct w\/ contrast = 0<\/p>\n\n<p>example(1, 48, 48, 32, 1) which is a volume from a PET scan.  <\/p>\n\n<p>As an aside to further complicate the matter to load the data manually I have to do something like below to get the array so that might also complicate the matter later if \".mat\" files are compatible with .flow_from_directory.<\/p>\n\n<pre><code>test = scipy.io.loadmat(directory)['1.mat']\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52682511,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><neural-network>","CreationDate":"2018-10-06T19:13:46.020","AcceptedAnswerId":null,"OwnerUserId":10466771.0,"Title":"expected dense_43_input to have shape (3,) but got array with shape (1,)","Body":"<p>i'm trying to get into NN, and i was making the simplest model to show it in the university. But now i'm really bad in it and only learning.<\/p>\n\n<p>I got an error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected dense_43_input to have shape (3,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>Model:<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Dense(500, input_shape=(3,), activation='relu'))\nmodel.add(Dense(500, activation='relu'))\nmodel.add(Dense(1))\nmodel.add(Activation('linear'))\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(X, y, batch_size=32, epochs=3, validation_split=0.3)\n<\/code><\/pre>\n\n<p>Predict:<\/p>\n\n<pre><code>t = np.array([-0.969308, 224.257508, 95.873082])\npredict = model.predict(t)\n<\/code><\/pre>\n\n<p>The shape of the prediction is (3,).<\/p>\n\n<p>What do i miss?<\/p>\n","answers":[{"AnswerId":"52683481","CreationDate":"2018-10-06T21:26:49.113","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Change your array to <code>np.array([[-0.969308, 224.257508, 95.873082]])<\/code>. If you do not add another dimension to your data then tf thinks you are asking to it to predict 3 inputs of 1 instead of 1 input of 3<\/p>\n\n<p>Edit: to comment on the other answer, your output dimensions are fine as the model is definitely complaining about the inputs. The predict function expects an array of inputs<\/p>\n"},{"AnswerId":"52682547","CreationDate":"2018-10-06T19:18:54.867","ParentId":null,"OwnerUserId":"9784039","Title":null,"Body":"<p>Why don\u2019t you change the model.add(Dense(1)) to model.add(Dense(3))?<\/p>\n"}]}
+{"QuestionId":52682521,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><reinforcement-learning><low-level-api>","CreationDate":"2018-10-06T19:15:06.210","AcceptedAnswerId":52682718.0,"OwnerUserId":7779411.0,"Title":"tf.gradients application on a function","Body":"<pre><code>x = tf.Placeholder(shape=[1,31,5,1])\ndef func(x):\n    operations...\n    return output\n\nconvolutionFunction = func(x)\nsess = tf.Session()\ngradientConv1 = gradientConv1 + sess.run(tf.gradients(tf.square(reward-convolutionFunction), weightsConv1))\n<\/code><\/pre>\n\n<p>gradientConv1 (numpy array of shape [2,2,1,32])\nweightsConv1 (tensor variable of shape [2,2,1,32])<\/p>\n\n<p>I'm getting an error such that \"Placeholder should have a dtype of float and shape of [1,31,5,1]\". It seems that it is showing me that I have not given a feed_dict to the function in sess.run? Please point me out to the error.\nAlso is my way of differentiating with respect to each value correct.<\/p>\n\n<p>reward is a scalar<\/p>\n","answers":[{"AnswerId":"52682718","CreationDate":"2018-10-06T19:38:21.443","ParentId":null,"OwnerUserId":"7779411","Title":null,"Body":"<pre><code>gradientConv1 = gradientConv1 + sess.run(tf.gradients(tf.square(reward-convolutionFunction), weightsConv1), feed_dict={x: &lt;valueOfPlaceholder&gt; })\n<\/code><\/pre>\n\n<p>where valueOfPlaceholder is the point at which we wish to evaluate the function<\/p>\n\n<p>Thanks to Andrey Akhmetov for pointing this out!<\/p>\n"}]}
+{"QuestionId":52682671,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-10-06T19:33:44.660","AcceptedAnswerId":null,"OwnerUserId":2243184.0,"Title":"Unable to save the model made with Keras functional API","Body":"<p>Specifically trying to save the entire MaskRCNN model implemented here \n<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN<\/a><\/p>\n\n<p>In line 2343 of <a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/mrcnn\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/mrcnn\/model.py<\/a> basically changed <code>save_weights_only<\/code> to <code>False<\/code> from <code>True<\/code> so that entire model is saved.<\/p>\n\n<pre><code>keras.callbacks.ModelCheckpoint(self.checkpoint_path, verbose=0, save_weights_only=False),\n<\/code><\/pre>\n\n<p>Stack trace of error is below <\/p>\n\n<pre><code>  File \".\/samples\/coco\/coco.py\", line 509, in &lt;module&gt;\n    augmentation=augmentation)\n  File \"\/mask_rcnn_root\/Mask_RCNN\/mrcnn\/model.py\", line 2374, in train\n    use_multiprocessing=True,\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/training.py\", line 1415, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/training_generator.py\", line 247, in fit_generator\n    callbacks.on_epoch_end(epoch, epoch_logs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/callbacks.py\", line 77, in on_epoch_end\n    callback.on_epoch_end(epoch, logs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/callbacks.py\", line 455, in on_epoch_end\n    self.model.save(filepath, overwrite=True)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/network.py\", line 1085, in save\n    save_model(self, filepath, overwrite, include_optimizer)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/saving.py\", line 116, in save_model\n    'config': model.get_config()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/network.py\", line 926, in get_config\n    return copy.deepcopy(config)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 218, in _deepcopy_list\n    y.append(deepcopy(a, memo))\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in _deepcopy_tuple\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in &lt;listcomp&gt;\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in _deepcopy_tuple\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 223, in &lt;listcomp&gt;\n    y = [deepcopy(a, memo) for a in x]\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 182, in deepcopy\n    y = _reconstruct(x, rv, 1, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 297, in _reconstruct\n    state = deepcopy(state, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 155, in deepcopy\n    y = copier(x, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 243, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n  File \"\/usr\/lib\/python3.5\/copy.py\", line 174, in deepcopy\n    rv = reductor(4)\nTypeError: can't pickle SwigPyObject objects keras \n<\/code><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"52804580","CreationDate":"2018-10-14T16:06:13.513","ParentId":null,"OwnerUserId":"2243184","Title":null,"Body":"<p>Basically the reason is improper use of Lambda layers in Keras messes up the saving of the model. Weights can still be saved though using <code>model.save_weights(\"my_model.h5\")<\/code> but if you try to save the entire model or extract the graph structure then the you get the crash. So all of the below failed for my case<\/p>\n\n<pre><code>model.save('my_model.h5') \njson_string = model.to_json() \nyaml_string = model.to_yaml() \n<\/code><\/pre>\n\n<p>More details here\n<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/11020#issuecomment-427638145\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/11020#issuecomment-427638145<\/a><\/p>\n"}]}
+{"QuestionId":52682979,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-06T20:14:32.777","AcceptedAnswerId":null,"OwnerUserId":10412109.0,"Title":"Is it possible to change the loss function dynamically during training?","Body":"<p>I am working on a machine learning project and I am wondering whether it is possible to change the loss function while the network is training. I'm not sure how to do it exactly in code.<\/p>\n\n<p>For example, start training with <code>cross entropy loss<\/code> and then halfway through training, switch to <code>0-1 loss<\/code>.<\/p>\n","answers":[{"AnswerId":"52695712","CreationDate":"2018-10-08T05:09:12.990","ParentId":null,"OwnerUserId":"10419477","Title":null,"Body":"<p>You have to implement your own algorithm. This is mostly possible with Tensorflow.<\/p>\n"}]}
+{"QuestionId":52683540,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-06T21:36:16.787","AcceptedAnswerId":null,"OwnerUserId":9893480.0,"Title":"In TensorFlow, for an image, how do I set value outside a circle be zero?","Body":"<p>What I want to do is like this.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GPkaf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GPkaf.png\" alt=\"enter image description here\"><\/a><a href=\"https:\/\/i.stack.imgur.com\/jluMR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jluMR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Below is code using numpy. How do I do it using tensorflow functions?<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\nimport tomopy\n\noutput_size = 64\n[X, Y] = np.mgrid[0:output_size, 0:output_size]\nxpr = X - int(output_size) \/\/ 2\nypr = Y - int(output_size) \/\/ 2\n\nreconstructed = tomopy.misc.phantom.barbara(size=64, dtype='float32')\nreconstructed = np.squeeze(reconstructed)\nplt.imshow(reconstructed, cmap='gray')\nplt.show()\n\nradius = output_size \/\/ 2\nreconstruction_circle = (xpr ** 2 + ypr ** 2) &lt;= radius ** 2\nreconstructed[~reconstruction_circle] = 0.\nplt.imshow(reconstructed, cmap='gray')\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53537150","CreationDate":"2018-11-29T10:42:59.287","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>you can do it by multiplication the image tensor with mask tensor. Mask tensor should be a tensor with 1. inside circle and 0. outside. You can pass this mask using placeholder or if the size of all images is fixed, you can define this tensor as tf.constant.<\/p>\n"}]}
+{"QuestionId":52683723,"AnswerCount":1,"Tags":"<javascript><python><tensorflow><keras><tensorflowjs-converter>","CreationDate":"2018-10-06T22:04:27.473","AcceptedAnswerId":53105743.0,"OwnerUserId":2730032.0,"Title":"Different results for tensorflowjs and keras on same model and tensor","Body":"<p>I trained a CNN model on some images following the example of <a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html<\/a>. My model code is identical, I just trained it on another image dataset: also for classification between two classes.<\/p>\n\n<p>The results are what you'd expect on the training set: images are classified correctly either 0 or 1.<\/p>\n\n<p>I saved the model in a tensorflowjs-friendly format following the \"Alternative: Use the Python API to export directly to TF.js Layers format\" section of  <a href=\"https:\/\/js.tensorflow.org\/tutorials\/import-keras.html\" rel=\"nofollow noreferrer\">https:\/\/js.tensorflow.org\/tutorials\/import-keras.html<\/a><\/p>\n\n<p>However when I try to access the results in an html page with javascript I get 1 for pretty much every image (or close to it): even if the image gives 0 in Keras.<\/p>\n\n<p>I have even saved an image as a tensor in JSON and I get 0 in Keras and 1 in TensorflowJS. Is this a bug or have I made a mistake somewhere ?<\/p>\n\n<p>Here is my code in TensorflowJS to access the json:<\/p>\n\n<pre><code>&lt;html&gt;\n  &lt;head&gt;\n    &lt;!-- Load TensorFlow.js --&gt;\n    &lt;meta http-equiv=\"content-type\" content=\"text\/html; charset=utf-8\" \/&gt;\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt;&lt;\/script&gt;\n\n    &lt;script&gt;\n      \/\/ https:\/\/stackoverflow.com\/a\/18324384\/2730032\n      function callAjax(url, callback){\n        var xmlhttp;\n        \/\/ compatible with IE7+, Firefox, Chrome, Opera, Safari\n        xmlhttp = new XMLHttpRequest();\n        xmlhttp.onreadystatechange = function(){\n            if (xmlhttp.readyState == 4 &amp;&amp; xmlhttp.status == 200){\n                callback(xmlhttp.responseText);\n            }\n        }\n        xmlhttp.open(\"GET\", url, true);\n        xmlhttp.send();\n      }\n\n      tf.loadModel('\/model.json').then(model =&gt; {\n        callAjax('\/tensor.json', res =&gt; {\n          arr = JSON.parse(res);\n          const example = tf.tensor(arr).reshape([1, 150, 150, 3]);\n          const prediction = model.predict(example);\n          prediction.data().then(res =&gt; {\n            console.log('PREDICTION JS', res[0]);\n          })\n        });\n      })\n    &lt;\/script&gt;\n  &lt;\/head&gt;\n  &lt;body&gt;\n  &lt;\/body&gt;\n&lt;\/html&gt;\n<\/code><\/pre>\n\n<p>And here is my python code for the same:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import json\nimport numpy as np\nimport tensorflowjs as tfjs\n\nmodel = tfjs.converters.load_keras_model('model.json')\n\nwith open('tensor.json', 'r') as f:\n    r = json.load(f)\narr = np.array([np.array([np.array(v) for v in l]) for l in r])\nprint('PREDICTION PYTHON', model.predict(arr[np.newaxis,...])[0][0])\n<\/code><\/pre>\n\n<p>I get PREDICTION JS 1 and PREDICTION PYTHON 0.0, for exactly the same data and the same model: does anybody see any issue in my code ?<\/p>\n\n<p><strong>EDIT1:<\/strong> I'm on Xubuntu 18.04.1 LTS and use the following software versions:<\/p>\n\n<pre><code>Python 3.6.6\nKeras 2.2.4\ntensorflow 1.11.0\ntensorflowjs 0.6.2\nnumpy 1.15.2\n<\/code><\/pre>\n\n<p><strong>EDIT2:<\/strong> I opened the following issue <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/issues\/776\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tfjs\/issues\/776<\/a> and it has since been fixed.<\/p>\n","answers":[{"AnswerId":"53105743","CreationDate":"2018-11-01T16:48:30.730","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>Upgrading to the latest version of tfjs (currently 0.13.3) solves the issue.\nMore context to the question can be viewed <a href=\"https:\/\/mail.google.com\/mail\/u\/0\/#inbox\/KtbxLxgNMwvBbhCpWgwfCfRHRtzlqtQMXV\" rel=\"nofollow noreferrer\">here<\/a> and  <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/issues\/776\" rel=\"nofollow noreferrer\">there<\/a><\/p>\n\n<pre><code>&lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.3\"&gt;&lt;\/script&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":52684127,"AnswerCount":1,"Tags":"<amazon-web-services><tensorflow><keras><gpu><tmux>","CreationDate":"2018-10-06T23:14:22.123","AcceptedAnswerId":null,"OwnerUserId":258483.0,"Title":"No module name `keras` under tmux on AWS instance","Body":"<p>I am trying to use Amazon AWS instance to train my network. To run it under keras, I need to run <\/p>\n\n<pre><code>source activate tensorflow_p36\n<\/code><\/pre>\n\n<p>first and it works. Unfortunately, if I do the same from under <code>tmux<\/code>, it says it can't find <code>keras<\/code> module.<\/p>\n\n<p>Why and how to overcome?<\/p>\n","answers":[{"AnswerId":"53754074","CreationDate":"2018-12-13T02:07:19.390","ParentId":null,"OwnerUserId":"5318904","Title":null,"Body":"<p>You can refer to the solution suggested in <a href=\"https:\/\/stackoverflow.com\/questions\/32543783\/tmux-session-wont-import-python-module?rq=1\">TMUX Session Won&#39;t Import Python Module<\/a>. If you start the tmux session first and then import tensorflow it should work. At least my issue was resolved when I used this sequence other wise I got an error saying tensorflow module was not found.<\/p>\n"}]}
+{"QuestionId":52684581,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-07T00:53:52.317","AcceptedAnswerId":null,"OwnerUserId":6476721.0,"Title":"Keras: How to concatenate input layer with output layer?","Body":"<p>I am trying to replicate the network from:\n<a href=\"https:\/\/arxiv.org\/pdf\/1604.07176.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1604.07176.pdf<\/a><\/p>\n\n<p>I see their implementation in\n<a href=\"https:\/\/github.com\/wentaozhu\/protein-cascade-cnn-lstm\/blob\/master\/cb6133.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wentaozhu\/protein-cascade-cnn-lstm\/blob\/master\/cb6133.py<\/a><\/p>\n\n<p>I am trying to train on the Q8 task only, with no solvent task. I am trying to concatenate the [512, 50] embedding layer to the [512, 22] output layer, but I keep getting various errors. This is how I am trying to concatenate at the moment:<\/p>\n\n<pre><code>main_input = Input(shape=(maxlen_seq,), dtype='int32', name='main_input')\n# Defining an embedding layer mapping from the words (n_words) to a vector of len 50\n# input_orig = K.reshape(input, (maxlen_seq, n_words))\nx = Embedding(input_dim=n_words, output_dim=50, input_length=maxlen_seq)(main_input)\n\naux_input = Input(shape=(maxlen_seq, n_words), name='aux_input')\nx = concatenate([x, aux_input], axis=-1)\n# ... rest of model ...\n<\/code><\/pre>\n\n<p>The model compiles fine with model.compile():<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\nmain_input (InputLayer)         (None, 512)          0                                            \n__________________________________________________________________________________________________\nembedding_27 (Embedding)        (None, 512, 50)      1100        main_input[0][0]                 \n__________________________________________________________________________________________________\naux_input (InputLayer)          (None, 512, 22)      0                                            \n__________________________________________________________________________________________________\nconcatenate_54 (Concatenate)    (None, 512, 72)      0           embedding_27[0][0]               \n                                                                 aux_input[0][0]                  \n________________________________________________________________________________________\n...\n<\/code><\/pre>\n\n<p>But I get a <code>ValueError: Error when checking input: expected aux_input to have 3 dimensions, but got array with shape (4464, 512)<\/code><\/p>\n\n<p>The model is defined as:<\/p>\n\n<pre><code>model = Model([main_input, aux_input], [y1, y2])\n<\/code><\/pre>\n\n<p>And fit as:<\/p>\n\n<pre><code>model.fit({'main_input':X_train,\n           'aux_input': X_train},\n          {'main_output': y_train,\n           'aux_output': y_train},\n          batch_size=128, epochs=20, callbacks=[early, best_model],\n          validation_data=({'main_input':X_val,\n                            'aux_input': X_val},\n                           {'main_output': y_val,\n                            'aux_output': y_val}),\n                            verbose=1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52684788,"AnswerCount":1,"Tags":"<python><python-3.x><pytorch><caffe2><onnx>","CreationDate":"2018-10-07T01:39:34.093","AcceptedAnswerId":null,"OwnerUserId":7452343.0,"Title":"ValueError: Don't know how to translate op Unsqueeze when running converted PyTorch Model","Body":"<p>I'm running into problems trying to use a PyTorch model exported as an ONNX model with Caffe2. Here is my export code <\/p>\n\n<pre><code>the_model = torchvision.models.densenet121(pretrained=True)\ngarbage, model_inputs = preprocessing(\"test.jpg\")\ntorch_out = torch.onnx._export(the_model,             \n                           model_inputs,                       \n                           \"model_weights\/chexnet-py.onnx\",\n                           export_params=True)\n<\/code><\/pre>\n\n<p>Now here is my testing code <\/p>\n\n<pre><code>model = onnx.load(\"model_weights\/chexnet-py.onnx\")\ngarbage, model_inputs = preprocessing(\"text.jpg\")\nprepared_backend = onnx_caffe2.backend.prepare(model)\nW = {model.graph.input[0].name: model_inputs.numpy()}\nc2_out = prepared_backend.run(W)[0]\n<\/code><\/pre>\n\n<p>This is returning the following error <\/p>\n\n<pre><code>ValueError: Don't know how to translate op Unsqueeze when running    converted PyTorch Model\n<\/code><\/pre>\n\n<p>Additional information \npytorch version 1.0.0a0+6f664d3\nCaffe2 is latest version (attempted building from source, pip, and conda). All gave same result. <\/p>\n","answers":[{"AnswerId":"55594529","CreationDate":"2019-04-09T13:52:38.983","ParentId":null,"OwnerUserId":"7878195","Title":null,"Body":"<p>Try looking into this, if you have to edit package called onnx-caffe2 to add the mapping b\/w Unsqueeze to ExpandDims\n<a href=\"https:\/\/github.com\/onnx\/onnx\/issues\/1481\" rel=\"nofollow noreferrer\">https:\/\/github.com\/onnx\/onnx\/issues\/1481<\/a><\/p>\n\n<p>Look for the answer:<\/p>\n\n<p>I found that the Caffe2 equivalence for Unsqueeze in ONNX is ExpandDims, and there is a special mapping in onnx_caffe2\/backend.py around line 121 for those operators that are different only in their names and attribute names, but somehow Unsqueeze isn't presented there (have no idea why). So I manually added the mapping rules for it in the _renamed_operators and _per_op_renamed_attrs dicts and the code would look like:<\/p>\n\n<pre><code>_renamed_operators = {\n    'Caffe2ConvTranspose':   'ConvTranspose',\n    'GlobalMaxPool':         'MaxPool',\n    'GlobalAveragePool':     'AveragePool',\n    'Pad':                   'PadImage',\n    'Neg':                   'Negative',\n    'BatchNormalization':    'SpatialBN',\n    'InstanceNormalization': 'InstanceNorm',\n    'MatMul':                'BatchMatMul',\n    'Upsample':              'ResizeNearest',\n    'Equal':                 'EQ',\n    'Unsqueeze':             'ExpandDims',  # add this line\n}\n\n_global_renamed_attrs = {'kernel_shape': 'kernels'}\n_per_op_renamed_attrs = {\n    'Squeeze':              {'axes': 'dims'},\n    'Transpose':            {'perm': 'axes'},\n    'Upsample':             {'mode': ''},\n    'Unsqueeze':            {'axes': 'dims'},  # add this line\n}\n<\/code><\/pre>\n\n<p>And everything works as expected.<\/p>\n\n<p>I am not the OP, thanks to OP though.<\/p>\n"}]}
+{"QuestionId":52684831,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-07T01:49:48.503","AcceptedAnswerId":null,"OwnerUserId":1152977.0,"Title":"fit_generator in Keras -- One minibatch at a time","Body":"<p>I have a model that operates on a large dataset - not all that large by Big Data standards but significantly more than my home server can hold in memory. As such, I'm using <code>fit_generator<\/code> to load it a chunk at a time, so that it never has to hold more than one minibatch in memory at once.<\/p>\n\n<p>... At least, that's the <em>theory<\/em>. But when Keras hung at <code>Epoch 1\/10<\/code> without even starting the \"training\" animation and I (eventually) got an Out of Memory exception -- the minibatches are large but I can still hold a couple of them in memory at once without trouble -- I got suspicious and threw in a bunch of testing <code>print<\/code> statements into my generator. Lo and behold, Keras was invoking the generator three or four times before even starting (appearing to start?).<\/p>\n\n<p>So... what's going on here? Is this normal, or did I implement my generator wrong somehow? How can I get it not to try and load more than one batch at once?<\/p>\n\n<p>Code follows, in case it helps:<\/p>\n\n<pre><code>def data_gen(directory):\n    def epsilon_div(x, y):\n        return (x + K.epsilon()) \/ (y + K.epsilon())\n\n    while(True):\n        filelist = os.listdir(directory + \"\/data\")\n        order = np.random.permutation(len(filelist))\n\n        for i in order:\n            dataf = directory + \"\/data\/\" + filelist[i]\n            labelf = directory + \"\/labels\/\" + filelist[i]\n\n            with open(dataf, 'rb') as f:\n                databook = sb.Songbook.FromString(f.read())\n\n            with open(labelf, 'rb') as f:\n                labelbook = sb.Songbook.FromString(f.read())\n\n            print('Booked')\n\n            l, _, r, _ = sb_np_extract(databook)\n            ll, _, lr, _ = sb_np_extract(labelbook)\n\n            databook = None\n            labelbook = None\n\n            print('Extracted')\n\n            l = l.transpose([0, 2, 1])\n            r = r.transpose([0, 2, 1])\n            ll = ll.transpose([0, 2, 1])\n            lr = lr.transpose([0, 2, 1])\n\n            print('Chosen')\n\n            mask_l = epsilon_div(ll, l)\n            mask_r = epsilon_div(lr, r)\n\n            print('Done')\n\n            yield [[l, r], [mask_l, mask_r]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52690882","CreationDate":"2018-10-07T17:05:39.113","ParentId":null,"OwnerUserId":"9179860","Title":null,"Body":"<p>I suggest you save the filenames in an array then shuffle it. I call the file list with the full path and shuffled <code>listname_data<\/code> and <code>listname_labels<\/code>. <\/p>\n\n<p>steps = number of files<\/p>\n\n<p>def generator(steps):<\/p>\n\n<pre><code>i = 1\n\nwhile True:\n\n     dataf = filelist_data[i]\n     labelf = filelist_labels[i]\n\n     ...\n\n     if i == steps:\n\n        i = 1\n        c = list(zip(listname_data,listname_data))\n\n        shuffle(c)\n\n        listname_data, listname_data = zip(*c)\n\n     else:\n\n        i +=1\n\n     yield [[l, r], [mask_l, mask_r]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52685768,"AnswerCount":2,"Tags":"<python><tensorflow><dataset><tensorflow-datasets>","CreationDate":"2018-10-07T05:39:51.933","AcceptedAnswerId":null,"OwnerUserId":10278331.0,"Title":"Tensorflow: Each row in the training dataset contains 99% of the previous rows data - can I optimize it before running the training?","Body":"<p>I am searching for a way to make my training and testing data smaller in file size.<\/p>\n\n<p><strong>The model I want to end up with<\/strong><\/p>\n\n<p>I want to train a model that predicts whether or not a crypto coin price is making and x% (0.4 or so) jump within the next 10 minutes (i.e. I want the model to answer with a Yes or No). <\/p>\n\n<p>Every minute I will the model with the last 3 hours of price and volume data (that means 180 datapoints each containing 5 values open, close, high, low prices and volume).<\/p>\n\n<p><strong>My current training and testing sets are BIG<\/strong><\/p>\n\n<p>My training and testing sets are therefore row in a csv file where each row contains 5 x 180 = 900 numbers plus one label (Yes or No) and with about 100k rows I guess this is a very huge dataset.<\/p>\n\n<p><strong>But each row in csv contains mostly redundant data<\/strong><\/p>\n\n<p>But each \"neighbor\" row in the csv file only contain 1 new data point since every next row is only 1 minute \"older\" and therefore only has skipped the data point of oldest minute and instead introduced a new point of the next minute.<\/p>\n\n<p><em>Is it possible to setup up the traing code so the csv file only needs to have the latest minute data point in each row?<\/em><\/p>\n","answers":[{"AnswerId":"52807275","CreationDate":"2018-10-14T21:31:51.143","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>If you use Data API then you can cache the input. Also maybe TF's support for Kafka might be a help here as you could model it as a stream of data. <\/p>\n\n<p>Another approach would be to reuse some data between session calls. Then you would have to use resource variable (in the current <code>Variable()<\/code> spec it means using flag <code>use_resource<\/code> in the constructor). This way your csv could contain only minute level data and you would just add it to the variable - create a kind of circle-buffer from it.<\/p>\n"},{"AnswerId":"52999192","CreationDate":"2018-10-25T22:53:51.093","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>Yes, you can set this up so that the CSV file needs to only have the latest minute of data in each row.<\/p>\n\n<p>If your CSV file contains one data point per row, you can create sliding window batches as follows:<\/p>\n\n<p>In TF 1.12 onwards:<\/p>\n\n<pre><code>window_size = 180\ndataset = tf.data.experimental.CsvDataset(filenames, defaults)\ndataset = dataset.window(window_size=window_size, shift=1)\ndataset = dataset.flat_map(lambda x: x.batch(window_size))\n<\/code><\/pre>\n\n<p>For example, if you had a dataset that contained the elements  <code>{1, 2, 3, 4, 5, 6, 7, 8, 9, 10}<\/code><\/p>\n\n<pre><code># Contains {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}\ndataset = ...\n\ndataset = dataset.window(window_size=3, shift=1)\ndataset = dataset.flat_map(lambda x: x.batch(3))\nit = dataset.make_one_shot_iterator()\nnxt = it.get_next()\nwith tf.Session() as sess:\n  sess.run(nxt)  # [1, 2, 3]\n  sess.run(nxt)  # [2, 3, 4]\n  sess.run(nxt)  # [3, 4, 5]\n  ...etc...\n<\/code><\/pre>\n\n<p>In TF 1.11 and earlier:<\/p>\n\n<pre><code>window_size = 180\ndataset = tf.contrib.data.CsvDataset(filenames, defaults)\ndataset = dataset.apply(tf.contrib.data.sliding_window_batch(window_size=window_size))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52685774,"AnswerCount":1,"Tags":"<python><tensorflow><keras><mnist>","CreationDate":"2018-10-07T05:40:53.323","AcceptedAnswerId":52715115.0,"OwnerUserId":10467819.0,"Title":"loading fashion_mnist data takes too much time","Body":"<p>I am using python 3.6.6, tensorflow 1.11.0, and keras 2.1.6 on Windows 10\nFollowing the below blog as a guideline to import and load the data.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a><\/p>\n\n<p>The shell IDLE prints too much output and continues indefinitely.  <\/p>\n\n<p>This line in my code seems to be the issue:<\/p>\n\n<pre><code>(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()\n<\/code><\/pre>\n\n<p>Below are a few lines of output from my shell IDLE execution:<\/p>\n\n<p>1.11.0<\/p>\n\n<pre><code>Downloading data from https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/train-images-idx3-ubyte.gz\n\n    8192\/26421880 [..............................] - ETA: 0s\n   16384\/26421880 [..............................] - ETA: 2:05\n  172032\/26421880 [..............................] - ETA: 21s \n  450560\/26421880 [..............................] - ETA: 11s\n  933888\/26421880 [&gt;.............................] - ETA: 7s \n 1507328\/26421880 [&gt;.............................] - ETA: 5s\n 2056192\/26421880 [=&gt;............................] - ETA: 4s\n 2670592\/26421880 [==&gt;...........................] - ETA: 4s\n 3358720\/26421880 [==&gt;...........................] - ETA: 3s\n 3833856\/26421880 [===&gt;..........................] - ETA: 3s\n 4259840\/26421880 [===&gt;..........................] - ETA: 3s\n 4489216\/26421880 [====&gt;.........................] - ETA: 3s\n 4931584\/26421880 [====&gt;.........................] - ETA: 3s\n 4997120\/26421880 [====&gt;.........................] - ETA: 3s\n 5005312\/26421880 [====&gt;.........................] - ETA: 3s\n 5013504\/26421880 [====&gt;.........................] - ETA: 4s\n 5021696\/26421880 [====&gt;.........................] - ETA: 4s\n 5029888\/26421880 [====&gt;.........................] - ETA: 4s\n 5038080\/26421880 [====&gt;.........................]\n<\/code><\/pre>\n\n<p>Why the shell does not stop?\nI appreciate any help.<\/p>\n","answers":[{"AnswerId":"52715115","CreationDate":"2018-10-09T07:04:04.600","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I believe IDLE's shell is not fast enough and might be slowing down the file download, try running the same python code from a terminal, as it will download the file and this needs to be done only once.<\/p>\n"}]}
+{"QuestionId":52686135,"AnswerCount":0,"Tags":"<pytorch><caffe2>","CreationDate":"2018-10-07T06:49:02.610","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Cannot open lmdb db file in Caffe2","Body":"<p>I installed Caffe2 from source for Ubuntu.\nI don't use Anaconda and my python is python2.<\/p>\n\n<p><code>which python<\/code> gives me<\/p>\n\n<pre><code>\/usr\/bin\/python\n<\/code><\/pre>\n\n<p>I can import lmdb in python.\nI can run tutorials from this <a href=\"https:\/\/caffe2.ai\/docs\/tutorials\" rel=\"nofollow noreferrer\">page<\/a> <\/p>\n\n<p>My problem in testing one of the examples is \"Can't open lmdb file\".<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-19-c9a97f268658&gt; in &lt;module&gt;()\n      1 # The parameter initialization network only needs to be run once.\n      2 # Now all the parameter blobs are initialized in the workspace.\n----&gt; 3 workspace.RunNetOnce(train_model.param_init_net)\n      4 \n      5 # Creating an actual network as a C++ object in memory.\n\n\/home\/pytorch\/build\/caffe2\/python\/workspace.py in RunNetOnce(net)\n    199         C.Workspace.current._last_failed_op_net_position,\n    200         GetNetName(net),\n--&gt; 201         StringifyProto(net),\n    202     )\n    203 \n\n\/home\/pytorch\/build\/caffe2\/python\/workspace.py in CallWithExceptionIntercept(func, op_id_fetcher, net_name, *args, **kwargs)\n    178 def CallWithExceptionIntercept(func, op_id_fetcher, net_name, *args, **kwargs):\n    179     try:\n--&gt; 180         return func(*args, **kwargs)\n    181     except Exception:\n    182         op_id = op_id_fetcher()\n\nRuntimeError: [enforce fail at db.h:190] db_. Cannot open db: \/home\/caffe2_notebooks\/tutorial_data\/mnist\/mnist-train-nchw-lmdb of type lmdbError from operator: \noutput: \"dbreader_\/home\/caffe2_notebooks\/tutorial_data\/mnist\/mnist-train-nchw-lmdb\" name: \"\" type: \"CreateDB\" arg { name: \"db_type\" s: \"lmdb\" } arg { name: \"db\" s: \"\/home\/caffe2_notebooks\/tutorial_data\/mnist\/mnist-train-nchw-lmdb\" }\n<\/code><\/pre>\n\n<p>I have all files and have read access.<\/p>\n\n<p>I compiled pytorch with <code>USE_LMDB=ON<\/code>.<\/p>\n\n<p>I see the same issues but can't find how they solved the issues.\nMy laptop system is small with <code>Ubuntu 16.4, with Nvidia Cuda GeForce GT 750M<\/code>, so quite old model.<\/p>\n\n<p>Then I found in one page. They solved the issue with changing LMDB_MAP_SIZE in lbdb.cc. So I changed to 4G, but still have error.<\/p>\n\n<pre><code>constexpr size_t LMDB_MAP_SIZE = 4294967296;\/\/4gb 1099511627776;  \/\/ 1 TB\n<\/code><\/pre>\n\n<p>The problem is at this line<\/p>\n\n<pre><code>workspace.RunNetOnce(train_model.param_init_net)\n<\/code><\/pre>\n\n<p>How to solve the issue?<\/p>\n","answers":[]}
+{"QuestionId":52686173,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-07T06:55:18.573","AcceptedAnswerId":52686507.0,"OwnerUserId":1703586.0,"Title":"Keras TimeDistributed layer without LSTM","Body":"<p>I am kind of new to keras. Here is what I am trying to achieve. I have a keras-model which takes as input an image to produce 512 vector. I create this as: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Ca2OB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ca2OB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>input_img = keras.layers.Input( shape=(240, 320, 3 ) )\ncnn = make_vgg( input_img )\nout = NetVLADLayer(num_clusters = 16)( cnn )\nmodel = keras.models.Model( inputs=input_img, outputs=out )\n<\/code><\/pre>\n\n<p>Now, for the training, each of my samples are actually 13 images. Say I have 2500 samples then my data's dimensions are 2500x13x240x320x3. \nI want the model to be applied independently to the 13 images. I came across the TimeDistributed layer in keras and wondering how can I use it to achieve my objective. <\/p>\n\n<pre><code>t_input = Input( shape=(13,240,320,3) )\n# How to use TimeDistributed with model? \nt_out = TimeDistributed( out )\nt_model = Model( inputs=t_input, outputs=t_out )\n<\/code><\/pre>\n\n<p>I am expecting t_out of size: None,13,512. The above code, however, throws a ValueError. Can anyone help my understanding? <\/p>\n","answers":[{"AnswerId":"52686507","CreationDate":"2018-10-07T07:53:39.077","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The error occurs in this line:<\/p>\n\n<pre><code>t_out = TimeDistributed(out)\n<\/code><\/pre>\n\n<p>It happens because <code>out<\/code> is a tensor, but <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"nofollow noreferrer\">TimeDistributed<\/a> expects a layer as argument. This layer will be applied to every temporal slice (dimension of index one) of the input. You could instead do the following:<\/p>\n\n<pre><code>t_input = Input(shape=(13, 240, 320, 3))\nt_out = TimeDistributed(model)(t_input)\nt_model = Model(inputs=t_input, outputs=t_out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52686275,"AnswerCount":2,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2018-10-07T07:11:27.623","AcceptedAnswerId":52686555.0,"OwnerUserId":2350784.0,"Title":"Tensorflow - Cannot feed value of shape (256, 784) for Tensor which has shape '(?, 28, 28, 1)'","Body":"<p>Here's my simple covnet.\nMy inputs are 28 * 28 * 1 (Grayscale).\nI've declared input placeholder to be (None, 28, 28, 1) to accomodate batches, and to work with conv filter shapes.\nWhen I evaluate the model, i think it is expecting images of size 28*28*1 but its getting a row vector. But even when I reshaped the image to 28*28*1 before feeding into the feed_dict, it still gives me the error. Can you pls help me correct it?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.examples.tutorials.mnist import input_data\n\n\ndef batches(batch_size, features, labels):\n    \"\"\"\n    Create batches of features and labels\n    :param batch_size: The batch size\n    :param features: List of features\n    :param labels: List of labels\n    :return: Batches of (Features, Labels)\n    \"\"\"\n    assert len(features) == len(labels)\n    outout_batches = []\n\n    sample_size = len(features)\n    for start_i in range(0, sample_size, batch_size):\n        end_i = start_i + batch_size\n        batch = [features[start_i:end_i], labels[start_i:end_i]]\n        outout_batches.append(batch)\n\n    return outout_batches\n\n\nmnist = input_data.read_data_sets('\/datasets\/ud730\/mnist', one_hot=True)\ntrain_features = mnist.train.images\ntest_features = mnist.test.images\n\ntrain_labels = mnist.train.labels.astype(np.float32)\ntest_labels = mnist.test.labels.astype(np.float32)\n\nbatch_size = 256\nn_input = 28 * 28\nn_labels = 10\nlearning_rate = 0.001\nn_hidden_layer = 256\nn_hidden_layer2 = 64\nmu = 0.0\nsigma = 0.1\n\n########### CONVOLUTION LAYER NETWORK START ########################################################################\n### INPUT --&gt; LAYER 1 : CONV1(5*5 FILTER, DEPTH=6, S=1, P=YES\/'SAME') --&gt; RELU --&gt; MAXPOOL --&gt;\n###           LAYER 2 :  --&gt; CONV2(5*5 FILTER, DEPTH=16, S=1, P=YES\/'SAME') --&gt; RELU --&gt; MAXPOOL --&gt;\n###           LAYER 3, 4, OUTPUT: --&gt; DENSE --&gt; DENSE --&gt; OUTPUT --&gt; ACTIVATION\n\n# INPUT         : 28 * 28\n# LAYER 1, d=6  : (28-5+1\/1 = 24) = 24 * 24 * 6\n# MAXPOOL1      : 12 * 12 * 6\n# LAYER 2, d=16 : (12-5+1\/1 = 8) = 8 * 8 * 16\n# MAXPOOL2      : 4 * 4 * 16\n# DENSE 1       : 256 --&gt; 128\n# DENSE 2       : 128 --&gt; 64\n# DENSE 3       : 64  --&gt; 10\n\n#LAYER1\n#w1 is the convolution filter weight, not the complete weight \n\nfeatures = tf.placeholder(tf.float32, [None, 28, 28, 1])\nlabels = tf.placeholder(tf.float32, [None])\n\n\nw1 = tf.Variable(tf.truncated_normal(shape=[5, 5, 1, 6], mean=mu, stddev=sigma))\nb1 = tf.Variable(tf.zeros(6))\nconv1 = tf.nn.conv2d(features, w1, strides=[1,1,1,1], padding='VALID')\nconv1 = tf.nn.relu(conv1)\nconv1 = tf.nn.max_pool(conv1, strides=[1,2,2,1], ksize=[1,2,2,1], padding='SAME')\n\nw2 = tf.Variable(tf.truncated_normal(shape=[5,5,6,16], mean=mu, stddev=sigma))\nb2 = tf.Variable(tf.zeros(16))\nconv2 = tf.nn.conv2d(conv1, w2, strides=[1,1,1,1], padding='VALID')\nconv2 = tf.nn.relu(conv2)\nconv2 = tf.nn.max_pool(conv2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\n\nflattened_vector = tf.contrib.layers.flatten(conv2)\n\nw3 = tf.Variable(tf.truncated_normal(shape=[256, 128], mean=mu, stddev=sigma))\nb3 = tf.Variable(tf.zeros(128))\nlayer3 = tf.add(tf.matmul(flattened_vector, w3), b3)\nlayer3 = tf.nn.relu(layer3)\n\nw4 = tf.Variable(tf.truncated_normal(shape=[128, 64], mean=mu, stddev=sigma))\nb4 = tf.Variable(tf.zeros(64))\nlayer4 = tf.add(tf.matmul(layer3, w4), b4)\nlayer4 = tf.nn.relu(layer4)\n\nw5 = tf.Variable(tf.truncated_normal(shape=[64, 10], mean=mu, stddev=sigma))\nb5 = tf.Variable(tf.zeros(10))\nlayer5 = tf.add(tf.matmul(layer4, w5), b5)\nlogits = tf.nn.relu(layer5)\n\n########### CONVOLUTION LAYER NETWORK END ########################################################################\n\n\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))\noptimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)\n\nprediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels,1))\naccuracy = tf.reduce_mean(tf.cast(prediction, tf.float32))\n\n\n############ EVALUATING THE MODEL ############\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    for batch_images, batch_labels in batches(batch_size, train_features, train_labels):\n        sess.run(optimizer, feed_dict={features:batch_images, labels:batch_labels})\n\n    total_accuracy = sess.run(accuracy, feed_dict={features:test_features, labels:test_labels})\n    print(total_accuracy)\n<\/code><\/pre>\n\n<p>Error <\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-14-3f341fec1373&gt; in &lt;module&gt;\n    106     sess.run(init)\n    107     for batch_images, batch_labels in batches(batch_size, train_features, train_labels):\n--&gt; 108         sess.run(optimizer, feed_dict={features:batch_images, labels:batch_labels})\n    109 \n    110     total_accuracy = sess.run(accuracy, feed_dict={features:test_features, labels:test_labels})\n\nF:\\ProgramFiles\\miniconda\\envs\\IntroToTensorFlow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in run(self, fetches, feed_dict, options, run_metadata)\n    875     try:\n    876       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 877                          run_metadata_ptr)\n    878       if run_metadata:\n    879         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nF:\\ProgramFiles\\miniconda\\envs\\IntroToTensorFlow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1074                              'which has shape %r' %\n   1075                              (np_val.shape, subfeed_t.name,\n-&gt; 1076                               str(subfeed_t.get_shape())))\n   1077           if not self.graph.is_feedable(subfeed_t):\n   1078             raise ValueError('Tensor %s may not be fed.' % subfeed_t)\n\nValueError: Cannot feed value of shape (256, 784) for Tensor 'Placeholder_26:0', which has shape '(?, 28, 28, 1)'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52686555","CreationDate":"2018-10-07T08:02:25.293","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>There were two issues in your code.<\/p>\n\n<p>You should reshape your inputs. For example by doing:<\/p>\n\n<pre><code>mnist = input_data.read_data_sets('\/tmp\/datasets\/ud730\/mnist', one_hot=True)\ntrain_features = [d.reshape(28, 28, 1) for d in mnist.train.images]\ntest_features = [d.reshape(28, 28, 1) for d in mnist.test.images]\n<\/code><\/pre>\n\n<p>Also the dimension of your labels should be:<\/p>\n\n<pre><code>...\nfeatures = tf.placeholder(tf.float32, [None, 28, 28, 1])\nlabels = tf.placeholder(tf.float32, [None, 10])\n...\n<\/code><\/pre>\n"},{"AnswerId":"52686364","CreationDate":"2018-10-07T07:28:08.217","ParentId":null,"OwnerUserId":"1259561","Title":null,"Body":"<p>I don't see where you reshape the input data. As the error message said, it expects the input to be the same shape as placeholder. So you either need to <\/p>\n\n<pre><code>np.reshape(batch_images, (-1, 28, 28, 1))\n<\/code><\/pre>\n\n<p>before putting it in <code>sess.run()<\/code> or <\/p>\n\n<pre><code>features = tf.reshape(features, (-1, 28, 28, 1))\n<\/code><\/pre>\n\n<p>in model definition.<\/p>\n"}]}
+{"QuestionId":52686639,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-07T08:11:46.513","AcceptedAnswerId":55166711.0,"OwnerUserId":9345284.0,"Title":"tensorflow lstm convert to tensor tpu model?","Body":"<p>what issue can be?<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation, Input, LSTM, Dense\ndef create_model():\n# create a small LSTM network\n    model = Sequential()\n    model.add(LSTM(20, input_shape=(X.shape[1], X.shape[2]), return_sequences=True))\n    model.add(LSTM(20, return_sequences=True))\n    model.add(LSTM(10, return_sequences=True))\n    model.add(Dropout(0.2))\n    model.add(LSTM(4, return_sequences=False))\n    model.add(Dense(4, kernel_initializer='uniform', activation='relu'))\n    model.add(Dense(1, kernel_initializer='uniform', activation='relu'))\n\nimport tensorflow as tf\n\ntf.logging.set_verbosity(tf.logging.DEBUG)\n\ntf.keras.backend.clear_session()\nmodel=create_model()\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    model,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))\n<\/code><\/pre>\n\n<blockquote>\n  <p>WARNING:tensorflow:tpu_model (from tensorflow.contrib.tpu.python.tpu.keras_support) is experimental and may change or be removed at any time, and without warning.<\/p>\n  \n  <p>ValueError: Fetch argument  cannot be interpreted as a Tensor. (Tensor Tensor(\"lstm_13\/kernel:0\", shape=(79, 320), dtype=float32_ref) is not an element of this graph.)<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"55166711","CreationDate":"2019-03-14T15:44:55.110","ParentId":null,"OwnerUserId":"11127105","Title":null,"Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/WDX9y.png\" rel=\"nofollow noreferrer\">First, change the backend to include a TPU using the notebook settings available in the Edit -> Notebook settings menu if it's not selected.<\/a> <\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.contrib.keras.api.keras.models import Sequential\nfrom tensorflow.python.keras.layers import Dense, Dropout, Activation, Input, LSTM, Dense\nimport os\ndef create_model(X):\n# create a small LSTM network\n    model = Sequential()\n    model.add(tf.keras.layers.LSTM(20, input_shape=(X.shape[1],X.shape[2]),return_sequences=True))\n    model.add(tf.keras.layers.LSTM(20, return_sequences=True))\n    model.add(tf.keras.layers.LSTM(10, return_sequences=True))\n    model.add(tf.keras.layers.Dropout(0.2))\n    model.add(tf.keras.layers.LSTM(4, return_sequences=False))\n    model.add(tf.keras.layers.Dense(4, kernel_initializer='uniform', activation='relu'))\n    model.add(tf.keras.layers.Dense(1, kernel_initializer='uniform', activation='relu'))\n    model.compile(loss=tf.keras.losses.categorical_crossentropy,\n                  optimizer=tf.train.RMSPropOptimizer(learning_rate=0.01),\n                  metrics=['accuracy'])\n# This address identifies the TPU we'll use when configuring TensorFlow.\ntpu='grpc:\/\/' + os.environ['COLAB_TPU_ADDR']\ntf.logging.set_verbosity(tf.logging.DEBUG)\ntf.keras.backend.clear_session()\nmodel=create_model(X)\n#Convert Keras model to TPU model\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    model,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(tpu)))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52686692,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-07T08:19:39.613","AcceptedAnswerId":null,"OwnerUserId":7036833.0,"Title":"LSTM using Keras, dimension error of input fed to the first layer","Body":"<p>as said above I have a problem with the input data dimension in my LSTM network, which gave me the next error: <\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>My code:<\/p>\n\n<pre><code>input_data = Input((20,))\nz = Embedding(5000, input_length=20, output_dim=80, \nweights[embedding_matrix])(input_data)\n\n#z:(?,25,100)-&gt;(batch_size?,25-sentence length,100-embedding_size)\n\nz = BatchNormalization(axis=-1)(z)\nz = LSTM(256,return_sequences=True)(inputs)\nz = LSTM(128)(z)\nz = Dense(64, activation='relu')(z)\nout = Dense(1, activation='sigmoid')(z)\n\nmodel = Model(input_data,out)\n<\/code><\/pre>\n\n<p>I'm getting the error in the first LSTM layer.<\/p>\n\n<p>I know that the problem is the <code>\"?\"<\/code> in <code>axis=0<\/code>, tried to solve this.\nThose are the things I tried:<\/p>\n\n<ol>\n<li><p>Inspired by this solution:\n<a href=\"http:\/\/%22https:\/\/stackoverflow.com\/questions\/49721810\/lstm-keras-value-input-dimension-error%22\" rel=\"nofollow noreferrer\">\"https:\/\/stackoverflow.com\/questions\/49721810\/lstm-keras-value-input-dimension-error\"<\/a><\/p>\n\n<p>x = x.reshape(1, x.shape[0], x.shape<a href=\"http:\/\/%22https:\/\/stackoverflow.com\/questions\/49721810\/lstm-keras-value-input-dimension-error%22\" rel=\"nofollow noreferrer\">1<\/a>)<\/p>\n\n<p>But I got this error:<\/p>\n\n<blockquote>\n  <p>IndexError: tuple index out of range<\/p>\n<\/blockquote><\/li>\n<li><p>Also Tried:<\/p>\n\n<blockquote>\n  <p>x = np.expand_dims(x,axis=0)<\/p>\n<\/blockquote>\n\n<p>But didn't because the types didn't match.<\/p><\/li>\n<\/ol>\n\n<p>I tried a few more solutions but with no luck, I'll be happy if someone can help me figure it out.<\/p>\n\n<p>Thanks!!<\/p>\n","answers":[]}
+{"QuestionId":52687780,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-10-07T10:51:09.823","AcceptedAnswerId":null,"OwnerUserId":8089695.0,"Title":"How to refer to the output of fully connected layer in tensorflow while restoring model?","Body":"<p>If i am connecting a fully-connected layer like below and the output of this is the logits that i want to refer to while restoring the models but, there is no <code>name<\/code> argument available in  <code>tf.contrib.layers.fully_connected<\/code>. Below is what i want to do but it throws error<\/p>\n\n<pre><code>logits = tf.contrib.layers.fully_connected(outl[str(i)], self.classes, activation_fn = None, \n                                                           weights_initializer = self.initializer,\n                                                          name = 'mlogits')\n\nfully_connected() got an unexpected keyword argument 'name'\n<\/code><\/pre>\n\n<p>What is the tensor name of the output of this fully connected layer?<\/p>\n","answers":[{"AnswerId":"52688085","CreationDate":"2018-10-07T11:35:10.897","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>You can just access to the complete name of the output tensor using the <code>.name<\/code> property:<\/p>\n\n<pre><code>print(logits.name)\n<\/code><\/pre>\n\n<p>This is the <strong>complete<\/strong> name of your output tensor<\/p>\n"}]}
+{"QuestionId":52688628,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-07T12:46:21.450","AcceptedAnswerId":null,"OwnerUserId":10413069.0,"Title":"Accessing a dict in Python results in TypeError: unhashable type: 'slice'","Body":"<pre><code>step = 0\nbatch_size = 128\niters = 2618730\nwhile step * batch_size &lt; iters:\n    batch_x = train_data[(step * batch_size):((step+1) * batch_size)]\n    batch_y = labels[(step * batch_size):((step+1) * batch_size)]\n    sess.run(optimizer, feed_dict={x: batch_x, y: batch_y keep_prob: dropout})\n<\/code><\/pre>\n\n<p>In the above snippet I'm trying to access train_data which is of 133 x 229498 (in dict form) in batches of 128, but it is throwing the following error<\/p>\n\n<pre><code>batch_x = train_data[(step * size):((step+1) * size)]\nTypeError: unhashable type: 'slice'\n<\/code><\/pre>\n\n<p>I'm using Python 3.6<\/p>\n","answers":[]}
+{"QuestionId":52688997,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-07T13:32:50.073","AcceptedAnswerId":null,"OwnerUserId":10165064.0,"Title":"Keras prediction time measurement: First value outlier","Body":"<p>I'm currently trying to measure the time, which a Keras (TF Back) model needs to make predictions. As this time might vary, i'm doing a sample of predicitons and calculate the mean and sample standard deviation of them. The code looks like this:<\/p>\n\n<pre><code> timekeras = []\n for i in range(int(ANALYSIS_SAMPLE_SIZE)):\n     timedummy = time.time()\n     predskeras = model.predict ( np.array(test_data_list) )\n     timekeras.append(time.time() - timedummy)\n<\/code><\/pre>\n\n<p>Now I noticed, that the sample standard deviation is always larger than the mean, which is a sign for some vast outliers. After looking at the list with the time measurements, I know now that the first prediction always takes about 10 times as long as those that follow. I guess the algorithm is somehow \"remembering\" that he did the exact same calculations right before and uses some of the already-calculated results?<\/p>\n\n<p>Anyway, my question is now: In a real application scenario, the model won't make the same prediction multiple times right one after another, so my measured values are seemingly useless except for the first one. Does anybody have an idea how to do the same prediction a number of times, but always as if it wasn't done before, in order to do meaningful time measurements? And generally, just because I'm curious: Does anybody know why Keras is doing this, taking 10 times less time for doing a prediction again?<\/p>\n","answers":[]}
+{"QuestionId":52689625,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-07T14:47:19.263","AcceptedAnswerId":null,"OwnerUserId":2424554.0,"Title":"Meet ValueError: Tensor must be from the same graph as Tensor when I predict twice","Body":"<p>Recently, I'm building a seq2seq model with Esmiator API, but it comes out a problem that has been bothering me a few days. <\/p>\n\n<p>It will come out a <code>ValueError<\/code>  when I predict twice.<\/p>\n\n<pre><code>   ...\n   predictions = list(mymodel.predict(\n       lambda: ev_func_with_sentence(last_result)))\n   predictions = list(mymodel.predict(\n       lambda: ev_func_with_sentence(last_result)))\n   ...\n   ...\nValueError: Tensor(\"embedding\/Cast:0\", shape=(?, 3000), dtype=int32) must be from the same graph as Tensor(\"embedding\/embeddings\/Read\/ReadVariableOp:0\", shape=(103727, 200), dtype=float32).\n<\/code><\/pre>\n\n<p><code>mymodel<\/code> was defined like this:<\/p>\n\n<pre><code>mymodel = tf.estimator.Estimator(\n    model_fn=model_fn,\n    model_dir=model_dir,\n    params={\n        'features': _get_features(),\n        'learning_rate': learning_rate\n    },\n    config=tf.estimator.RunConfig(\n        save_checkpoints_steps=1,\n        save_summary_steps=1,\n    )\n)\n<\/code><\/pre>\n\n<p>Here is my <code>input_func<\/code>:<\/p>\n\n<pre><code>def ev_data_with_sentence(last_results=None):\n    model = Word2Vec.load(configs.WORD2VEC_FILE)\n    datas = ev_set()\n    if last_results is None:\n        # A [batch * 30] array\n        last_results = numpy.zeros([len(datas), configs.MAX_TITLE_WORD])\n\n    for data, result in zip(datas, last_results):\n        content = data.get('content')\n        xs = tokenization(content, is_sentence=True)\n        X = index_word_with_sentence(xs, model)\n        # X.shape == (?, 50, 60)\n        yield X, result\n\n\ndef ev_func_with_sentence(last_results):\n    def _warp(data, labels):\n        return {'x1': data, 'x2': labels}\n\n    data = tf.data.Dataset().from_generator(\n        lambda: ev_data_with_sentence(last_results),\n        output_types=((tf.float32, tf.int64)))\n    data = data.batch(configs.batch_size)\n    data = data.map(_warp)\n    data = data.make_one_shot_iterator()\n    inputs = data.get_next()\n    return inputs\n<\/code><\/pre>\n\n<p>It makes me crazy. Help! Please!<\/p>\n","answers":[]}
+{"QuestionId":52689796,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-07T15:04:43.130","AcceptedAnswerId":null,"OwnerUserId":10460663.0,"Title":"ValueError: Cannot feed value of shape (256, 0) for Tensor 'Placeholder_2:0', which has shape '(?, 500)'","Body":"<pre><code>#Import TsnsorFlow\nimport tensorflow as tf\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\n#Imort data\ndata = pd.read_csv(\"C:\\\\testdata1004.csv\")\n\n#Drop data variable\ndata = data.drop(['DATE'],1)\n\n#Dimensions of dataset\nn = data.shape[0]\np = data.shape[1]\n\n#Make data a numpy array\ndata = data.values\n\n\n#Training and test data\ntrain_start = 0\ntrain_end = int(np.floor(0.8*n))\ntest_start = train_end\ntest_end = n\ndata_train = data[np.arange(train_start, train_end), :]\ndata_test = data[np.arange(test_start, test_end), :]\n\n#Scale data\nfrom sklearn.preprocessing import MinMaxScaler\nscaler = MinMaxScaler()\nscaler.fit(data_train)\ndata_train = scaler.transform(data_train)\ndata_test = scaler.transform(data_test)\n\n#Build X and Y\nX_train = data_train[:, 1:]\ny_train = data_train[:, 0]\nX_test = data_test[:, 1:]\ny_test = data_test[:, 0]\n\n\n#Define a and b as placeholders\na = tf.placeholder(dtype=tf.int8)\nb = tf.placeholder(dtype=tf.int8)\n\n#Define a and b as placeholders\nc = tf.add(a,b)\n\n#Initialize the graph\ngraph = tf.Session()\n\n#Run the graph\ngraph.run(c, feed_dict={a: 5, b: 4})\n\n#Model architecture parameters\nn_stocks = 500\nn_neurons_1 = 1024\nn_neurons_2 = 512\nn_neurons_3 = 256\nn_neurons_4 = 128\nn_target = 1\n\n#Placeholder\nX = tf.placeholder(dtype=tf.float32, shape=[None, n_stocks])\nY = tf.placeholder(dtype=tf.float32, shape=[None])\n\n#Initializers\nsigma = 1\nweight_initializer = tf.variance_scaling_initializer(mode=\"fan_avg\",distribution=\"uniform\", scale=sigma)\nbias_initializer = tf.zeros_initializer()\n\n#Layer 1: Variables for hidden weights and biases\nW_hidden_1 = tf.Variable(weight_initializer([n_stocks,n_neurons_1]))\nbias_hidden_1 = tf.Variable(bias_initializer([n_neurons_1]))\n\n#Layer 2: VAriables for hidden weights and biases\nW_hidden_2 = tf.Variable(weight_initializer([n_neurons_1,n_neurons_2]))\nbias_hidden_2 = tf.Variable(bias_initializer([n_neurons_2]))\n\n#Layer 3: VAriables for hidden weights and biases\nW_hidden_3 = tf.Variable(weight_initializer([n_neurons_2,n_neurons_3]))\nbias_hidden_3 = tf.Variable(bias_initializer([n_neurons_3]))\n\n#Layer 4: VAriables for hidden weights and biases\nW_hidden_4 = tf.Variable(weight_initializer([n_neurons_3,n_neurons_4]))\nbias_hidden_4 = tf.Variable(bias_initializer([n_neurons_4]))\n\n#Output layer : Variables for output weights and biases\nW_out = tf.Variable(weight_initializer([n_neurons_4,n_target]))\nbias_out = tf.Variable(bias_initializer([n_target]))\n\n#Hidden layer\nhidden_1 = tf.nn.relu(tf.add(tf.matmul(X, W_hidden_1),bias_hidden_1))\nhidden_2 = tf.nn.relu(tf.add(tf.matmul(hidden_1,W_hidden_2),bias_hidden_2))\nhidden_3 = tf.nn.relu(tf.add(tf.matmul(hidden_2,W_hidden_3),bias_hidden_3))\nhidden_4 = tf.nn.relu(tf.add(tf.matmul(hidden_3,W_hidden_4),bias_hidden_4))\n\n#Output layer (must be transposed)\nout = tf.transpose(tf.add(tf.matmul(hidden_4, W_out),bias_out))\n\n#Cost function\nmse = tf.reduce_mean(tf.squared_difference(out, Y))\n\n#Optimizer\nopt = tf.train.AdamOptimizer().minimize(mse)\n\n#Make Session\nnet = tf.Session()\n\n#Run initializer\nnet.run(tf.global_variables_initializer())\n\n#Setup interactive plot\nplt.ion()\nfig = plt.figure()\nax1 = fig.add_subplot(111)\nline1, = ax1.plot(y_test)\nline2, = ax1.plot(y_test*0.5)\n\nplt.show()\n\n#Number of epochs and batch size\nepochs = 10\nbatch_size = 256\n\nfor e in range(epochs):\n\n    #suffle training data\n    shuffle_indices = np.random.permutation(np.arange(len(y_train)))\n    X_train = X_train[shuffle_indices]\n    y_train = y_train[shuffle_indices]\n\n    #Minibatch training\n    for i in range(0, len(y_train) \/\/ batch_size) : \n        start = i * batch_size\n        batch_x = X_train[start:start + batch_size]\n        batch_y = y_train[start:start + batch_size]\n        #Run optimizer with batch\n        net.run(opt, feed_dict={X:batch_x, Y: batch_y})\n\n        #Show progress\n        if np.mod(i,5) == 0:\n            #Prediction\n            pred = net.run(out, feed_dict={X: X_test})\n            line2.set_ydata(pred)\n            plt.title('Epoch ' + str(e) + ', Batch ' + str(i))\n            file_name = 'img\/epoch_' + str(e) + '_batch_' + str(i) + '.jpg'\n            plt.savefig(file_name)\n            plt.pause(0.01)\n\n#Print final MSE after Training\nmse_final = net.run(mse, feed_dict= {X: X_test, Y: y_test})\nprint(mse_final)\n\u200b\n<\/code><\/pre>\n\n<p>I implemented tensorflow code, but it doesn't work and i got error \nthis code is just an example, that i have. I dont know what is the reason i got the error and the way I can fix. could you tell me how to fix it? Thank you\nI implemented tensorflow code, but it doesn't work and i got error \nthis code is just an example, that i have. I dont know what is the reason i got the error and the way I can fix. could you tell me how to fix it? Thank you<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/d2fXI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d2fXI.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52690309","CreationDate":"2018-10-07T15:59:18.530","ParentId":null,"OwnerUserId":"10469355","Title":null,"Body":"<p>The None is for batch size, but then you need another dimension for length of the data you input into y. Try changing:<\/p>\n\n<pre><code>Y = tf.placeholder(dtype=tf.float32, shape=[None,1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52689970,"AnswerCount":1,"Tags":"<python><pip><segmentation-fault><pytorch><coredump>","CreationDate":"2018-10-07T15:20:57.480","AcceptedAnswerId":null,"OwnerUserId":10469065.0,"Title":"Segmentation fault (Core dumped) on importing torch with python3.5 in virtualenv","Body":"<p>I have installed <code>torch=0.4.1<\/code> without <code>cuda<\/code> in a <code>virtualenv<\/code>. I am using <code>python3.5<\/code> on ubuntu 16.04. Whenever I <code>import torch<\/code> in an interactive python shell, it quits the python program by showing <code>Segmentation fault (core dumped)<\/code>. \n<a href=\"https:\/\/i.stack.imgur.com\/V2yP4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/V2yP4.png\" alt=\"Error on importing torch\"><\/a><\/p>\n\n<p>While, surprisingly I had earlier initialized a <code>jupyter notebook<\/code> and tried importing torch there and it was running fine. Can someone please help? I could not find a solution on official `PyTorch github' discussions.\n<a href=\"https:\/\/i.stack.imgur.com\/EtFaR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/EtFaR.png\" alt=\"Jupyter notebook torch import working fine\"><\/a><\/p>\n","answers":[{"AnswerId":"55891294","CreationDate":"2019-04-28T14:33:42.670","ParentId":null,"OwnerUserId":"11423285","Title":null,"Body":"<p>I had the same issue, the solution was given on <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/926#issuecomment-284258385\" rel=\"nofollow noreferrer\">GitHub<\/a>:<\/p>\n\n<pre><code>import cv2        # first import cv2\nimport torch\n<\/code><\/pre>\n"}]}
+{"QuestionId":52690283,"AnswerCount":0,"Tags":"<python><ssl><tensorflow><ubuntu-18.04>","CreationDate":"2018-10-07T15:56:07.337","AcceptedAnswerId":null,"OwnerUserId":6599735.0,"Title":"SSL Certificate Error when running tensorflow gpu on ubuntu","Body":"<p>Tensorflow: 1.11.0\nUbuntu: 18.04\nPython: 3.6.6<\/p>\n\n<p>I used pyenv to create an environment with python 3.6.6 and installed the tensorflow gpu into it.  <\/p>\n\n<p>Verified tensorflow GPU installed with:<\/p>\n\n<pre><code>python -c \"import tensorflow as tf; print(tf.__version__)\"\n<\/code><\/pre>\n\n<p>and it returned: 1.11.0<\/p>\n\n<p>then I tried running the sample code that was on: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/<\/a><\/p>\n\n<p>I encountered the following errors:<\/p>\n\n<pre><code>ython 3.6.6 (default, Oct  6 2018, 00:35:42) \n[GCC 6.4.0 20180424] on linux\nType \"copyright\", \"credits\" or \"license()\" for more information.\n&gt;&gt;&gt; \n=============== RESTART: \/home\/chubuntu\/tf366\/tf-tutorial-1.py ===============\nDownloading data from https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/mnist.npz\nTraceback (most recent call last):\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 1318, in do_open\n    encode_chunked=req.has_header('Transfer-encoding'))\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 1239, in request\n    self._send_request(method, url, body, headers, encode_chunked)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 1285, in _send_request\n    self.endheaders(body, encode_chunked=encode_chunked)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 1234, in endheaders\n    self._send_output(message_body, encode_chunked=encode_chunked)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 1026, in _send_output\n    self.send(msg)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 964, in send\n    self.connect()\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/http\/client.py\", line 1400, in connect\n    server_hostname=server_hostname)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/ssl.py\", line 407, in wrap_socket\n    _context=self, _session=session)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/ssl.py\", line 814, in __init__\n    self.do_handshake()\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/ssl.py\", line 1068, in do_handshake\n    self._sslobj.do_handshake()\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/ssl.py\", line 689, in do_handshake\n    self._sslobj.do_handshake()\nssl.SSLError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:841)\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 243, in get_file\n    urlretrieve(origin, fpath, dl_progress)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 248, in urlretrieve\n    with contextlib.closing(urlopen(url, data)) as fp:\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 223, in urlopen\n    return opener.open(url, data, timeout)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 526, in open\n    response = self._open(req, data)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 544, in _open\n    '_open', req)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 504, in _call_chain\n    result = func(*args)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 1361, in https_open\n    context=self._context, check_hostname=self._check_hostname)\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/urllib\/request.py\", line 1320, in do_open\n    raise URLError(err)\nurllib.error.URLError: &lt;urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:841)&gt;\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/chubuntu\/tf366\/tf-tutorial-1.py\", line 4, in &lt;module&gt;\n    (x_train, y_train),(x_test, y_test) = mnist.load_data()\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/datasets\/mnist.py\", line 49, in load_data\n    file_hash='8a61469f7ea1b51cbae51d4f78837e45')\n  File \"\/home\/chubuntu\/.pyenv\/versions\/3.6.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 245, in get_file\n    raise Exception(error_msg.format(origin, e.errno, e.reason))\nException: URL fetch failure on https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/mnist.npz: None -- [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:841)\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>help!<\/p>\n","answers":[]}
+{"QuestionId":52690293,"AnswerCount":2,"Tags":"<python><tensorflow><jupyter-notebook><python-3.6><debian-stretch>","CreationDate":"2018-10-07T15:57:44.330","AcceptedAnswerId":null,"OwnerUserId":6275044.0,"Title":"Tensorflow: AttributeError: 'NoneType' object has no attribute 'original_name_scope'","Body":"<p>I am trying to run some python tensorflow code on a debian 9.5 stretch system on google cloud. I am Using tensorflow GPU version of this (the latest version) with the approriate CODA and cuDNN software installed. Here is my code:<\/p>\n\n<pre><code>     import tensorflow as tf\n\nmnist = tf.keras.datasets.mnist\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\n\n\nx_train = tf.keras.utils.normalize(x_train, axis=1) # scales all values between 0 and 1 on pixel image\nx_test = tf.keras.utils.normalize(x_test, axis=1)\n\nmodel = tf.keras.models.Sequential()\nmodel.add (tf.keras.layers.Flatten()) # flattens the 28x28 pixels into long flat vector\nmodel.add(tf.layers.Dense(128, activation=tf.nn.relu))# builds hidden layer 128 neurons (we can tweak) and activation func - use this as default, sigmoid function\nmodel.add(tf.layers.Dense(128, activation=tf.nn.relu)) # second layer\nmodel.add(tf.layers.Dense(10, activation=tf.nn.softmax)) # output layer - number of output neurons =classes\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics = ['accuracy'])  # adam is least squares test\n\nmodel.fit(x_train, y_train, epochs=3)\n#model.fit(x_train, y_train, epochs=3)\n<\/code><\/pre>\n\n<p>However, I get the following error:<\/p>\n\n<hr>\n\n<p>AttributeError                            Traceback (most recent call last)\n&lt;<\/p>\n\n<pre><code>ipython-input-1-3604b3cbf07d&gt; in &lt;module&gt;\n     16 model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics = ['accuracy'])  # adam is least squares test\n     17 \n---&gt; 18 model.fit(x_train, y_train, epochs=3)\n     19 #model.fit(x_train, y_train, epochs=3)\n     20 \n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1507         steps_name='steps_per_epoch',\n   1508         steps=steps_per_epoch,\n-&gt; 1509         validation_split=validation_split)\n   1510 \n   1511     # Prepare validation data.\n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split)\n    991       x, y = next_element\n    992     x, y, sample_weights = self._standardize_weights(x, y, sample_weight,\n--&gt; 993                                                      class_weight, batch_size)\n    994     return x, y, sample_weights\n    995 \n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_weights(self, x, y, sample_weight, class_weight, batch_size)\n   1027       if not self.inputs:\n   1028         is_build_called = True\n-&gt; 1029         self._set_inputs(x)\n   1030 \n   1031     if y is not None:\n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/checkpointable\/base.py in _method_wrapper(self, *args, **kwargs)\n    424     self._setattr_tracking = False  # pylint: disable=protected-access\n    425     try:\n--&gt; 426       method(self, *args, **kwargs)\n    427     finally:\n    428       self._setattr_tracking = previous_value  # pylint: disable=protected-access\n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in _set_inputs(self, inputs, training)\n   1220       else:\n   1221         input_shape = (None,) + inputs.shape[1:]\n-&gt; 1222         self.build(input_shape=input_shape)\n   1223     if context.executing_eagerly():\n   1224       self._eager_set_inputs(inputs)\n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/sequential.py in build(self, input_shape)\n    219       for layer in self.layers:\n    220         if not layer.built:\n--&gt; 221           with ops.name_scope(layer._name_scope()):\n    222             layer.build(shape)\n    223           layer.built = True\n\n~\/Python-3.6.4\/py_36_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/base.py in _name_scope(self)\n    139   def _name_scope(self):\n    140     \"\"\"Determines op naming for the Layer.\"\"\"\n--&gt; 141     return self._current_scope.original_name_scope\n    142 \n    143   def _set_scope(self, scope=None):\n\nAttributeError: 'NoneType' object has no attribute 'original_name_scope'\n<\/code><\/pre>\n\n<p>This code works fine on my laptop using the CPU version of tensorflow, however I am getting this error on the virtual machine. Any ideas?<\/p>\n","answers":[{"AnswerId":"52729393","CreationDate":"2018-10-09T21:11:48.163","ParentId":null,"OwnerUserId":"6762463","Title":null,"Body":"<p>I had the same issue, and resolved it by setting the <code>input_size<\/code> parameter in the first layer of the sequential model.<\/p>\n\n<p>this line:\n<code>model.add (tf.keras.layers.Flatten())<\/code><\/p>\n\n<p>should be:\n<code>model.add (tf.keras.layers.Flatten(input_size=(28, 28))<\/code><\/p>\n"},{"AnswerId":"53964885","CreationDate":"2018-12-28T22:20:22.310","ParentId":null,"OwnerUserId":"536051","Title":null,"Body":"<p>i had the similar issue and the solution is to use<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(128, activation=tf.nn.relu))\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>model.add(tf.layers.Dense(128, activation=tf.nn.relu))\n<\/code><\/pre>\n\n<p>answer by @Josh will be required for saving the model.<\/p>\n"}]}
+{"QuestionId":52690303,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-10-07T15:58:59.823","AcceptedAnswerId":null,"OwnerUserId":5511906.0,"Title":"How to solve InvalidArgumentError for dynamic_partition in tensorflow?","Body":"<p>I was trying to use <code>tf.dynamic_partition()<\/code> to replace <code>tf.gather()<\/code> in order to aviod converting sparse representation into dense matrix implicitly.<\/p>\n\n<p>Here is my code,<\/p>\n\n<pre><code># edge_source_states = tf.gather(params=node_states_per_layer[-1], indices=edge_sources)\npartitions0 = tf.reduce_sum(tf.one_hot(edge_sources, tf.shape(node_states_per_layer[-1])[0], dtype='int32'),\n                                                        0)\nedge_source_states = tf.dynamic_partition(node_states_per_layer[-1], partitions0, 2)\nedge_source_states = edge_source_states[1]\n<\/code><\/pre>\n\n<p>The annotation is the orignal use of <code>tf.gather()<\/code>. There was no mistake when using <code>tf.gather()<\/code>, the only problem is that it converts sparse representation into dense matrix and thus consumes a lot of memory.\n<br>\nHowever, when I use <code>tf.dynamic_partition()<\/code> method instead, I got an error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): partitions[21] = 2 is not in [0, 2)\n<\/code><\/pre>\n\n<p>And according to the traceback, this error is caused by the sentence:<\/p>\n\n<pre><code>edge_source_states = tf.dynamic_partition(node_states_per_layer[-1], partitions0, 2)\n<\/code><\/pre>\n\n<p>As a novice, I really can't figure it out. <\/p>\n\n<p>My problems are:<\/p>\n\n<p>1) I thought my new codes using <code>tf.dynamic_partition()<\/code> is functionally equivalent to the original code using <code>tf.gather()<\/code>. So why there is an error?<\/p>\n\n<p>2) Is <code>tf.dynamic_partition()<\/code> a solution for avoiding the implicit conversion from sparse to dense as <code>tf.gather()<\/code> does? Are there any other solutions? I really need to control the memory consumption strictly.<\/p>\n","answers":[]}
+{"QuestionId":52690837,"AnswerCount":1,"Tags":"<keras><conv-neural-network><convolution>","CreationDate":"2018-10-07T17:00:40.800","AcceptedAnswerId":52690909.0,"OwnerUserId":6916919.0,"Title":"Output shape of a convolutional layer with multiple filters","Body":"<p>I built a ConvNet in keras and these are 2 of the layers<\/p>\n\n<pre><code>model.add(Conv2D(8 , 3 , input_shape = (28,28,1)))\nmodel.add(Activation(act))\n\nmodel.add(Conv2D(16 , 3))\nmodel.add(Activation(act))\n<\/code><\/pre>\n\n<p>The output of the first layer of size <code>26x26x8<\/code> which I completely understand since there are 8 filters of size 3x3 and each of them is applied to produce a separate feature map hence <code>26x26x8<\/code><\/p>\n\n<p>The output of the second layer is of size <code>24x24x16<\/code> which I do not understand. Shouldn't the output be of size <code>24x24x128<\/code> since each of the filters of the second layer will act on each feature map of the output of the first layer ?<\/p>\n\n<p>Basically, I do not understand how the output of a layer is fed to the input of the other layer<\/p>\n","answers":[{"AnswerId":"52690909","CreationDate":"2018-10-07T17:09:27.630","ParentId":null,"OwnerUserId":"3049753","Title":null,"Body":"<p>No, it's a <a href=\"https:\/\/www.coursera.org\/lecture\/convolutional-neural-networks\/convolutions-over-volume-ctQZz\" rel=\"nofollow noreferrer\">convolutions over volume<\/a>. Each filter is applied for all channels.<\/p>\n"}]}
+{"QuestionId":52690881,"AnswerCount":1,"Tags":"<python><deep-learning><mathematical-optimization><pytorch><loss-function>","CreationDate":"2018-10-07T17:05:38.710","AcceptedAnswerId":52693253.0,"OwnerUserId":7580944.0,"Title":"How to extend a Loss Function Pytorch","Body":"<p>I would like to create my own custom Loss function as a weighted combination of 3 Loss Function, something similar to:<\/p>\n\n<pre><code>criterion = torch.nn.CrossEntropyLoss(out1, lbl1) + \\\n            torch.nn.CrossEntropyLoss(out2, lbl2) + \\\n            torch.nn.CrossEntropyLoss(out3, lbl3)\n<\/code><\/pre>\n\n<p>I am doing it to address a multi-class multi-label classification problem.\nDoes it make sense? How to implement correctly such Loss Function in Pytorch?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52693253","CreationDate":"2018-10-07T21:55:03.777","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Your way of approaching the problem seems correct but there's a typo in your code. Here's a fix for that:<\/p>\n\n<pre><code>loss1 = torch.nn.CrossEntropyLoss()(out1, lbl1)\nloss2 = torch.nn.CrossEntropyLoss()(out2, lbl2) \nloss3 = torch.nn.CrossEntropyLoss()(out3, lbl3)\n\nfinal_loss = loss1 + loss2 + loss3\n<\/code><\/pre>\n\n<p>Then you can call <code>.backward<\/code> on <code>final_loss<\/code> which should then compute the gradients and backpropagate them.<\/p>\n\n<p>Also, it's possible to <em>weight<\/em> each of the component losses where the <em>weights<\/em> are itself <em>learned<\/em> during the training process.<\/p>\n\n<p>You can refer the discussions of <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-combine-multiple-criterions-to-a-loss-function\/348\/3\" rel=\"nofollow noreferrer\">combine-multiple-criterions-to-a-loss-function<\/a> for more information.<\/p>\n"}]}
+{"QuestionId":52690947,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-07T17:13:46.427","AcceptedAnswerId":null,"OwnerUserId":2880477.0,"Title":"Computing capability of GTX 870M","Body":"<p>I was trying to run tensorflow-gpu on an ASUS laptop with a GTX 870M card on Ubuntu 16.40, but got an error message <\/p>\n\n<p>018-10-07 16:54:50.537324: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1482] Ignoring visible gpu device (device: 0, name: GeForce GTX 870M, pci bus id: 0000:01:00.0, compute capability: 3.0) with Cuda compute capability 3.0. The minimum required Cuda capability is 3.5.<\/p>\n\n<p>However, GTX 870M's computing capability is listed as 5.0 (Maxwell).  My questions are (1) what is GTX 870M's computing capability, and (2) Can I run tensorflow-gpu (latest or nightly) with GTX 870M?  Thanks, CC.<\/p>\n","answers":[{"AnswerId":"56916097","CreationDate":"2019-07-06T17:03:07.913","ParentId":null,"OwnerUserId":"7748395","Title":null,"Body":"<p>I have a Razer Blade 14 2014 with the same GPU on Windows and run on the same problems. Unfortunately the CUDA compute capability <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">is listed as 3.0<\/a> for this GPU.<\/p>\n\n<p>In the context on Linux you have an advantage, for computing capability of 3.0 you can build the code from source using the official instructions, they recommend docker when doing this: <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/source<\/a><\/p>\n\n<p>A post I found a post where <a href=\"http:\/\/www.programmersought.com\/article\/1316940889\/\" rel=\"nofollow noreferrer\">the problem is solved<\/a> can help as well.<\/p>\n\n<p>Please let me know if you could make it work, I will try to install it using latest stable <a href=\"https:\/\/ubuntu-mate.org\/download\/\" rel=\"nofollow noreferrer\">Ubuntu Mate 18.04.2 LTS<\/a>.<\/p>\n"}]}
+{"QuestionId":52690971,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-10-07T17:16:45.000","AcceptedAnswerId":null,"OwnerUserId":10453487.0,"Title":"Passing input tensor size to via tensorflow pipeline","Body":"<p>I am getting the following error<\/p>\n\n<p>ValueError: The last dimension of the inputs to <code>Dense<\/code> should be defined.     Found <code>None<\/code>.<\/p>\n\n<p>when passing a tensor using the tf.data pipeline to tf.layers.dense. The relevant parts of my code are:<\/p>\n\n<pre><code>def _parse_function(example_proto):\n    features = {'X': tf.VarLenFeature(tf.float32),\n                'Y': tf.VarLenFeature(tf.float32)}\n    parsed_features = tf.parse_example(example_proto, features)\n    X = tf.sparse_tensor_to_dense(parsed_features['X'])\n    Y = tf.sparse_tensor_to_dense(parsed_features['Y'])\n    return X, Y\n\ndataset = tf.data.TFRecordDataset(fin)\ndataset = dataset.batch(100)\ndataset = dataset.map(_parse_function)\ndataset = dataset.cache()\ndataset = dataset.repeat()\niterator = dataset.make_one_shot_iterator()\nX, Y = iterator.get_next()\n\nhidden_0 = tf.layers.dense(X, N_HIDDEN_0, activation=tf.nn.crelu, use_bias=False)\n<\/code><\/pre>\n\n<p>Is this because passing the tensor shape to tf.layers.dense is not possible with the VarLenFeature or the sparse_tensor_to_dense operation? Is there a way to fix it without using sess.run(X,Y) and feeding the output using feed_dict? <\/p>\n\n<p>Am wondering if this is similar to the issue that @mrry addressed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/13348\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/13348<\/a>.<\/p>\n\n<p>Any insights would be appreciated!<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52708441","CreationDate":"2018-10-08T18:52:45.670","ParentId":null,"OwnerUserId":"10453487","Title":null,"Body":"<p>I figured out how to fix the problem by changing my def _parse_function to the following:<\/p>\n\n<pre><code>def _parse_function(example_proto):\n    features = {'X': tf.FixedLenFeature([18],tf.float32), 'Y': tf.FixedLenFeature([3],tf.float32)}\n    parsef = tf.parse_example(example_proto, features)\n    return parsef['X'], parsef['Y']\n<\/code><\/pre>\n\n<p>So I guess the problem was because with tf.VarLenFeature in my original code, the number of input and output nodes could not be communicated to tf.layers.dense. I am still unclear as to why I have to specify the number of columns of X and Y ([18] and [3] in the code above), instead of specifying them as [ ] - but I can live with that.<\/p>\n"}]}
+{"QuestionId":52691280,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-07T17:54:53.673","AcceptedAnswerId":52705946.0,"OwnerUserId":1185117.0,"Title":"TF \/ Keras error: InputLayer not a Checkpointable","Body":"<p>I am trying out the newly added TPU support on Google Colab with the simple cats vs dogs dataset. <\/p>\n\n<p>After creating a simple CNN, I tried to export the model to TPU. But it failed with error<\/p>\n\n<pre><code>TypeError: Checkpointable._track_checkpointable() passed type &lt;class 'keras.engine.topology.InputLayer'&gt;, not a Checkpointable.\n<\/code><\/pre>\n\n<p>Here's the code that I wrote on Colab.<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(layers.Conv2D(32, (3,3), activation='relu', input_shape=(150, 150, 3)))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Conv2D(64, (3,3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Dense(512, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))\nmodel.summary()\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(train_dir, target_size=(150,150), batch_size=20, class_mode='binary')\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(model, strategy=tf.contrib.tpu.TPUDistributionStrategy(tf.contrib.cluster_resolver.TPUClusterResolver(tpu=\"grpc:\/\/\" + os.environ['COLAB_TPU_ADDR'])))\n<\/code><\/pre>\n\n<p>My guess is I am doing something wrong in <code>train_generator<\/code>. But I am not sure what it is. Any help would be highly appreciated.<\/p>\n","answers":[{"AnswerId":"52705946","CreationDate":"2018-10-08T15:51:37.313","ParentId":null,"OwnerUserId":"6170271","Title":null,"Body":"<p>If you're using or import <code>layers<\/code> from <code>Keras<\/code> instead of <code>TensorFlow<\/code> like this:<\/p>\n\n<pre><code>from keras import layers,models\nfrom keras.preprocessing.image import ImageDataGenerator\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>You will get error like you mention above :<\/p>\n\n<pre><code>TypeError: Checkpointable._track_checkpointable() passed type &lt;class 'keras.engine.topology.InputLayer'&gt;, not a Checkpointable.\n<\/code><\/pre>\n\n<p>So, you can import <code>layers<\/code> directly from <code>TensorFlow<\/code> like my code below:<\/p>\n\n<pre><code>from tensorflow.keras import layers,models\nfrom keras.preprocessing.image import ImageDataGenerator\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>or you can see my full code here:\n<a href=\"https:\/\/gist.github.com\/ilmimris\/8218e397dd35ab693404e95db32dc574\" rel=\"noreferrer\">https:\/\/gist.github.com\/ilmimris\/8218e397dd35ab693404e95db32dc574<\/a><\/p>\n"}]}
+{"QuestionId":52691534,"AnswerCount":0,"Tags":"<r><tensorflow><keras><rstudio><ubuntu-18.04>","CreationDate":"2018-10-07T18:22:43.020","AcceptedAnswerId":null,"OwnerUserId":6599735.0,"Title":"ImportError: libcublas when loading dataset on Rstudio running Keras with Tensorflow GPU","Body":"<p>Ubuntu 18.04\nRstudio 3.4.4\nTensorflow 1.11.0<\/p>\n\n<p>When I run <code>install_keras(tensorflow=\"gpu\")<\/code>\nall was good with the last few lines being:<\/p>\n\n<pre><code>tensorflow-gpu 1.11.0 has requirement setuptools&lt;=39.1.0, but you'll have setuptools 40.4.3 which is incompatible.\nInstalling collected packages: keras, tensorflow-hub\nSuccessfully installed keras-2.2.4 tensorflow-hub-0.1.1\nYou are using pip version 10.0.1, however version 18.1 is available.\nYou should consider upgrading via the 'pip install --upgrade pip' command.\n\nInstallation complete.\n\n\nRestarting R session...\n<\/code><\/pre>\n\n<p>I don't think the issues are related to what's coming next.\nI loaded the IMDB dataset: <code>&gt; imdb &lt;- dataset_imdb(num_words = 10000)<\/code>\nAnd these errors popped up:<\/p>\n\n<pre><code>Error: ImportError: Traceback (most recent call last):\n  File \"\/home\/chubuntu\/anaconda3\/envs\/r-tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/chubuntu\/anaconda3\/envs\/r-tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/chubuntu\/anaconda3\/envs\/r-tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/chubuntu\/anaconda3\/envs\/r-tensorflow\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/chubuntu\/anaconda3\/envs\/r-tensorflow\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcublas\n<\/code><\/pre>\n\n<p>I can load those dataset in keras with tensorflow without GPU.  So what could be wrong here?<\/p>\n","answers":[]}
+{"QuestionId":52691538,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-07T18:22:57.877","AcceptedAnswerId":52691920.0,"OwnerUserId":9749306.0,"Title":"training a multi-output keras model","Body":"<p>I have 10,000 images, each of which are labeled with 20 tags. For each image, the tag is either true or false. I'm trying to train a multi-output model to perform all these 20 binary classifications with one network. <\/p>\n\n<p>The network is a Residual Network. After the flatten layer, the network branches out into 20 branches. Each branch has 2 fully connected layers, each of which are followed by a drop out layer. And finally a dense layer with one node and sigmoid activation in the end.<\/p>\n\n<p>The labels for each image and the image name are stored in a text file, for both train and validation set. Like this:\n1.jpg 1 -1 1 -1 -1 1 -1.........<\/p>\n\n<p>I wrote my own generator, but I can't get them to work. I keep getting this error:<\/p>\n\n<pre><code>Error when checking model target: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 20 array(s), but instead got the following list of 1 arrays.\n<\/code><\/pre>\n\n<p>Function explanations: <code>get_input<\/code> function reads an image and resizes it.\n<code>get_output<\/code> prepares the labels for each image. The labels are stored in a list and returned in the end. <code>preprocess_input<\/code> performs preprocessing and converting images into arrays. <code>train_generator<\/code> and <code>validation_generator<\/code> generate batches with size 32 to be fed to the model.<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><code>def get_input(img_name):\n    path = os.path.join(\"images\", img_name)\n    img = image.load_img(path, target_size=(224, 224))\n\n    return img\n\n\ndef get_output(img_name, file_path):\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n\n    img_id = img_name.split(\".\")[0]\n    img_id = img_id.lstrip(\"0\")\n    img_id = int(img_id)\n\n    labels = data.loc[img_id - 1].values\n    labels = labels[1:]\n\n    labels = list(labels)\n    label_arrays = []\n    for i in range(20):\n        val = np.zeros((1))\n        val[0] = labels[i]\n        label_arrays.append(val)\n\n    return label_arrays\n\n\ndef preprocess_input(img_name):\n    img = get_input(img_name)\n    x = image.img_to_array(img)\n    x = np.expand_dims(x, axis=0)    \n    return x\n\ndef train_generator(batch_size):\n    file_path = \"train.txt\"\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n\n    while True:\n        for i in range(math.floor(8000\/batch_size)):\n            x_batch = np.zeros(shape=(32, 224, 224, 3))\n            y_batch = np.zeros(shape=(32, 20))\n            for j in range(batch_size):\n                img_name = data.loc[i * batch_size + j].values\n                img_name = img_name[0]\n                x = preprocess_input(img_name)\n                y = get_output(img_name, file_path)\n                x_batch[j, :, :, :] = x\n                y_batch[j] = y\n            yield(x_batch, y_batch)\n\ndef val_generator(batch_size):\n    file_path = \"val.txt\"\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n\n    while True:\n        for i in range(math.floor(2000\/batch_size)):\n            x_batch = np.zeros(shape=(32, 224, 224, 3))\n            y_batch = np.zeros(shape=(32, 20))\n            for j in range(batch_size):\n                img_name = data.loc[i * batch_size + j].values\n                img_name = img_name[0]\n                x = preprocess_input(img_name)\n                y = get_output(img_name, file_path)\n                x_batch[j, :, :, :] = x\n                y_batch[j] = y\n            yield(x_batch, y_batch)\n<\/code><\/pre>\n\n<p>Edit:\nOne quick question. What's the difference between this loop and the one in your answer:<\/p>\n\n<pre><code>ys = []\nfor i in range(batch_size):\n    ys.append(y_batch[i, :])\n\nyield(x_batch, ys)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52691752","CreationDate":"2018-10-07T18:49:16.267","ParentId":null,"OwnerUserId":"9179860","Title":null,"Body":"<p>You can test the generator output dimensions initializing the generator and call the function next() to check the dimensions. For example with the train_generator:<\/p>\n\n<pre><code>train_gen = train_generator(batch_size)\nx_batch, y_batch = next(train_gen)\n<\/code><\/pre>\n\n<p>Then check x_batch and y_batch dimensions and datatype<\/p>\n\n<p>I would make the generator in this way:<\/p>\n\n<pre><code>def train_generator(batch_size):\n    file_path = \"train.txt\"\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n    # Initialize empty list\n    x_batch = []\n    y_batch = []\n\n    while True:\n        for i in range(math.floor(8000\/batch_size)):\n            for j in range(batch_size):\n                img_name = data.loc[i * batch_size + j].values\n                img_name = img_name[0]\n                x = preprocess_input(img_name)\n                y = get_output(img_name, file_path)\n                x_batch.append(x)\n                y_batch.append(y)\n\n            yield(np.array(x_batch), np.array(y_batch))\n<\/code><\/pre>\n"},{"AnswerId":"52691920","CreationDate":"2018-10-07T19:07:08.570","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If your model has 20 outputs then you must provide a list of 20 arrays as target. One way of doing this is to modify the generator (for both training and validation):<\/p>\n\n<pre><code>ys = []\nfor i in range(20):\n    ys.append(y_batch[:,i])\n\nyield(x_batch, ys)\n<\/code><\/pre>\n\n<p>As a side note, you mentioned that you have 20 tags per sample then why have you specified 40 in the input shape?<\/p>\n\n<pre><code>y_batch = np.zeros(shape=(32, 40))\n<\/code><\/pre>\n\n<p>Further, I don't know about the specific problem you are working on but alternatively you could only have one output of size 20 instead of 20 outputs with size one.<\/p>\n"}]}
+{"QuestionId":52691566,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-07T18:25:31.773","AcceptedAnswerId":null,"OwnerUserId":4769035.0,"Title":"Keras: \"An operation has `None` for gradient\" in a simple network","Body":"<p>I am getting started with Keras and trying to implement a simple network to test it on CIFAR10 (32x32 images, 10 classes), but the model fails to <code>fit<\/code> the dataset.<\/p>\n\n<p>Model code:<\/p>\n\n<pre><code>import keras, keras.layers as L\n\ninput_shape = [32, 32, 3]\nn_inputs = 32 * 32 * 3\n\nmodel = keras.models.Sequential()\nmodel.add(L.InputLayer(input_shape))\nmodel.add(L.Flatten())\nmodel.add(L.Dense(n_inputs, activation='relu'))\nmodel.add(L.Dense(10, activation='softmax'))\n\nmodel.compile(optimizer='adam', \n              loss=keras.metrics.sparse_categorical_accuracy,\n              metrics=['accuracy'])\nmodel.fit(x=X_train, y=y_train,\n          epochs=10)\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-16-be14b685c358&gt; in &lt;module&gt;\n      4 model.fit(x=X_train, y=y_train,\n      5 #           validation_data=[X_val, y_val],\n----&gt; 6           epochs=10)\n\n...\\keras\\engine\\training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n-&gt; 1010         self._make_train_function()\n\n...\\keras\\engine\\training.py in _make_train_function(self)\n--&gt; 509                         loss=self.total_loss)\n\n...\\keras\\legacy\\interfaces.py in wrapper(*args, **kwargs)\n---&gt; 91             return func(*args, **kwargs)\n\n...\\keras\\optimizers.py in get_updates(self, loss, params)\n    473     @interfaces.legacy_get_updates_support\n    474     def get_updates(self, loss, params):\n--&gt; 475         grads = self.get_gradients(loss, params)\n    476         self.updates = [K.update_add(self.iterations, 1)]\n    477 \n\n...\\keras\\optimizers.py in get_gradients(self, loss, params)\n---&gt; 91             raise ValueError('An operation has `None` for gradient. '\n\nValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.\n<\/code><\/pre>\n\n<p>I have seen similar snippets working correctly in tutorials and cannot find the mistake. What is wrong with this model?<\/p>\n","answers":[]}
+{"QuestionId":52691569,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-07T18:26:38.543","AcceptedAnswerId":null,"OwnerUserId":8365437.0,"Title":"keras model.fit() fed with initializable iterator of tf.Dataset object","Body":"<p>I am using the tf.keras API to build my CNN model with the use of the tf.Dataset API to create a input pipeline for my model. The mnist dataset from the <code>tf.keras.datasets<\/code> is used for testing and prepared in the memory by executing the code:<\/p>\n\n<pre><code>(train_images,train_labels),(test_images,test_labels) = tf.keras.datasets.mnist.load_data()\n<\/code><\/pre>\n\n<p>and also some preprocessing to be compatible with my keras model:<\/p>\n\n<pre><code>Train_images = np.expand_dims(train_images,3).astype('float')\/255.0\nTest_images = np.expand_dims(test_images,3).astype('float')\/255.0\n\nTrain_labels = tf.keras.utils.to_categorical(train_labels)\nTest_labels = tf.keras.utils.to_categorical(test_labels)\n<\/code><\/pre>\n\n<p>Those data are stored in the memory as arrays and there are two option for creating a Dataset object. The first one is simply using <code>tf.data.Dataset.from_tensor_slices<\/code>:<\/p>\n\n<pre><code>image = tf.data.Dataset.from_tensor_slices((Train_images,Train_labels))\n<\/code><\/pre>\n\n<p>And input this resulting object to the model.fit():<\/p>\n\n<pre><code>model.fit(x=image,steps_per_epoch=1000)\n<\/code><\/pre>\n\n<p>OR input this dataset's iterator by:<\/p>\n\n<pre><code>iterator = image.make_one_shot_iterator()\n\nmodel.fit(x=iterator,steps_per_epoch=1000)\n<\/code><\/pre>\n\n<p>Both of these two options just work fine since the dataset named image here is created using the data in the memory. However, according the the <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">Importing Data<\/a> here that we may like to avoid doing this because it copies data several times and takes up memory. So another option is creating such a dataset object based on <code>tf.placeholder<\/code> as well as the initialiable iterator:<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32,shape = [60000,28,28,1])\nY = tf.placeholder(tf.float32,shape = [60000,10])\nimage2 = tf.data.Dataset.from_tensor_slices((X,Y))\niterator2 = image.make_initializable_iterator()\n\nwith tf.Session() as sess:\n  sess.run(iterator2.initializer,feed_dict={X:Train_images,Y:Train_labels}\n  sess.run(iterator2.get_next())\n<\/code><\/pre>\n\n<p>This kind of iterator works fine when using <code>tf.Session()<\/code> when fed with the data in the memory and avoids the multiple copies of the data. But I cannot find way to let it work with <code>keras.model.fit()<\/code> since you cannot really call the <code>iterator.initializer<\/code> or feed any data there. Is there a way to use this kind of iterator?<\/p>\n","answers":[{"AnswerId":"55140535","CreationDate":"2019-03-13T11:12:15.733","ParentId":null,"OwnerUserId":"2209299","Title":null,"Body":"<p>I don't think keras officially supports a case of passing initializable iterators, as you noted, there is no place to provide the placeholders and values mappings.<\/p>\n\n<p>However a workaround is possible using <a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">keras callbacks<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport pandas as pd\n\n# Make sure only tensorflow.keras is imported, don't mix with keras\nfrom tensorflow.keras import layers\nimport tensorflow.keras.backend as K\n\n# example data\nx_values = np.random.randn(200, 100).astype(np.float32)\ny_labels = np.random.randint(low=0, high=9, size=200)\n\ngraph = tf.Graph()\nwith graph.as_default():\n    # make datasets from placeholders as in https:\/\/www.tensorflow.org\/guide\/datasets#reading_input_data\n    # X:\n    features_placeholder = tf.placeholder(tf.float32, x_values.shape, name='features')\n    dataset_x = tf.data.Dataset.from_tensor_slices({'x': features_placeholder})\n    # Y:\n    labels_placeholder = tf.placeholder(tf.float32, [None], name='labels')\n    dataset_y = tf.data.Dataset.from_tensor_slices({'y': labels_placeholder})\n\n    # compose datasets to make X-Y pairs for training\n    dataset0 = tf.data.Dataset.zip((dataset_x, dataset_y))\n    dataset0 = dataset0.batch(16).repeat()\n\n    # build model with keras\n    inputs = tf.keras.Input(name='x', shape=(x_values.shape[1],))\n    mlp1 = layers.Dense(16, name='mlp-1', activation='relu')\n    mlp1_out = mlp1(inputs)\n    output = layers.Dense(1, name='y', activation='linear')\n    output_out = output(mlp1_out)\n    model = tf.keras.Model(inputs=inputs, outputs=output_out)\n    # The compile step specifies the training configuration.\n    model.compile(optimizer=tf.train.RMSPropOptimizer(0.001), loss='mse', metrics=['mse'])\n\n    iterator = dataset0.make_initializable_iterator()\n    feed_dict = { labels_placeholder: y_labels, features_placeholder: x_values }\n\n    class InitIteratorCallback(tf.keras.callbacks.Callback):\n        \"\"\"\n        Ensures that placeholders in dataset are initialized before each epoch begins\n        \"\"\"\n\n        def on_epoch_begin(self, epoch, logs=None):\n            sess = K.get_session()\n            sess.run(iterator.initializer, feed_dict=feed_dict)\n\n\n    model.fit(iterator, callbacks=[InitIteratorCallback()],\n              epochs=10, steps_per_epoch=300)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52691611,"AnswerCount":1,"Tags":"<python><anaconda><spyder><pytorch>","CreationDate":"2018-10-07T18:30:33.550","AcceptedAnswerId":null,"OwnerUserId":8787039.0,"Title":"spyder doesn't launch after installing pytorch","Body":"<p>I installed pytorch but after that Spyder can no longer be launched. Here are the terminal info:<\/p>\n\n<blockquote>\n  <p>conda install pytorch torchvision -c pytorch Solving environment: done<\/p>\n  \n  <p>==> WARNING: A newer version of conda exists. &lt;==   current version: 4.5.10   latest version: 4.5.11<\/p>\n  \n  <p>Please update conda by running<\/p>\n\n<pre><code>$ conda update -n base -c defaults conda\n<\/code><\/pre>\n  \n  <h2>Package Plan<\/h2>\n  \n  <p>environment location: \/anaconda3\/envs\/base_py36<\/p>\n  \n  <p>added \/ updated specs: \n      - pytorch\n      - torchvision<\/p>\n  \n  <p>The following packages will be downloaded:<\/p>\n\n<pre><code>package                    |            build\n---------------------------|-----------------\ntorchvision-0.2.1          |           py36_1         123 KB  pytorch\nscipy-1.1.0                |   py36hf1f7d93_0        15.4 MB\nscikit-learn-0.20.0        |   py36h4f467ca_1         5.4 MB\nnumpy-base-1.15.2          |   py36h8a80b8c_1         4.1 MB\nnumpy-1.11.3               |   py36heee0a97_5         3.4 MB\nninja-1.8.2                |   py36h04f5b5a_1          93 KB\npytorch-0.4.1              |py36_cuda0.0_cudnn0.0_1        10.0 MB  pytorch\n------------------------------------------------------------\n                                       Total:        38.5 MB\n<\/code><\/pre>\n  \n  <p>The following NEW packages will be INSTALLED:<\/p>\n\n<pre><code>ninja:        1.8.2-py36h04f5b5a_1                 \npytorch:      0.4.1-py36_cuda0.0_cudnn0.0_1 pytorch\ntorchvision:  0.2.1-py36_1                  pytorch\n<\/code><\/pre>\n  \n  <p>The following packages will be REMOVED:<\/p>\n\n<pre><code>accelerate:   2.3.1-np111py36_0                    \n<\/code><\/pre>\n  \n  <p>The following packages will be UPDATED:<\/p>\n\n<pre><code>mkl:          11.3.3-0                              --&gt; 2019.0-118           \nnumexpr:      2.6.7-py36hde7755b_0                  --&gt; 2.6.8-py36h1dc9127_0 \nnumpy:        1.11.3-py36_nomklh8ecaf62_5           --&gt; 1.11.3-py36heee0a97_5\nnumpy-base:   1.15.0-py36he97cb71_0                 --&gt; 1.15.2-py36h8a80b8c_1\nscikit-learn: 0.19.1-py36_nomklhde7755b_0           --&gt; 0.20.0-py36h4f467ca_1\nscipy:        1.1.0-py36_nomklh7cd7d8e_0            --&gt; 1.1.0-py36hf1f7d93_0 \n<\/code><\/pre>\n  \n  <p>The following packages will be DOWNGRADED:<\/p>\n\n<pre><code>blas:         1.0-openblas                          --&gt; 1.0-mkl              \n<\/code><\/pre>\n  \n  <p>Proceed ([y]\/n)? y<\/p>\n  \n  <p>Downloading and Extracting Packages torchvision-0.2.1    | 123 KB    |<\/p>\n  \n  ############################### | 100%  scipy-1.1.0          | 15.4 MB   | ##################################### | 100%  scikit-learn-0.20.0  | 5.4 MB    |\n  \n  ############################### | 100%  numpy-base-1.15.2    | 4.1 MB    | ##################################### | 100%  numpy-1.11.3         | 3.4 MB    | ##################################### | 100%\n  \n  <p>ninja-1.8.2          | 93 KB     |<\/p>\n  \n  ############################### | 100%  pytorch-0.4.1        | 10.0 MB   | ##################################### | 100%  Preparing transaction: done\n<\/blockquote>\n","answers":[{"AnswerId":"56301053","CreationDate":"2019-05-25T01:45:00.373","ParentId":null,"OwnerUserId":"11553146","Title":null,"Body":"<p>I don't know if you have solved your issue, but in case you have and someone else comes across this question or you haven't and are still waiting:<\/p>\n\n<p>I came across your post as a result of having the same thing happen to me... <\/p>\n\n<p>It would seem that all I had to do was \"conda update all\" for it to start working again!<\/p>\n"}]}
+{"QuestionId":52691837,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-07T18:58:30.573","AcceptedAnswerId":null,"OwnerUserId":6484358.0,"Title":"How to train a Deeplab model from scratch in TensorFlow?","Body":"<p>I am using <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/deeplab\" rel=\"nofollow noreferrer\">deepLab<\/a> to generate semantic segmentation masked images for a video in cityscapes datasets. So, I started with the pre-trained model <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/model_zoo.md\" rel=\"nofollow noreferrer\">xception65_cityscapes_trainfine<\/a> provided on the modelzoo and trained it further on the dataset.<\/p>\n\n<p>I am curious to know How I can start training it from scratch? and not end up just using the pre-trained model? could anyone suggest a direction on How I can achieve it?<\/p>\n\n<p>Any contribution from the community will be helpful and appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52692622,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-07T20:28:52.977","AcceptedAnswerId":52692667.0,"OwnerUserId":5069471.0,"Title":"Error - No module named '_pywrap_tensorflow'","Body":"<p>I have seen multiple questions for the same issue. I went through all the answers and tried all of them. I updated pip, tensorflow, python etc to the latest versions or as suggested in the answers and still I am facing this issue.<\/p>\n\n<p>Pip version 18.0,\nPython 3.7<\/p>\n","answers":[{"AnswerId":"52692667","CreationDate":"2018-10-07T20:35:09.530","ParentId":null,"OwnerUserId":"3852441","Title":null,"Body":"<p>Please downgrade Python to 3.6.x and try again. I had faced similar issue while using Python 3.7.x. Once I downgraded it, it worked. Make sure you adjust your path variable accordingly. \"Pip\" also may have to be modified and the corresponding path variable.<\/p>\n"}]}
+{"QuestionId":52694261,"AnswerCount":1,"Tags":"<python><tensorflow><import><module><spyder>","CreationDate":"2018-10-08T01:12:05.780","AcceptedAnswerId":null,"OwnerUserId":10470641.0,"Title":"Python module loading with Spyder","Body":"<p>I am new to using Python for machine learning and I am trying to learn ZhuSuan using Spyder. <\/p>\n\n<p>I have downloaded and installed Zhusuan as descibed here:  <a href=\"https:\/\/zhusuan.readthedocs.io\/en\/latest\/\" rel=\"nofollow noreferrer\">https:\/\/zhusuan.readthedocs.io\/en\/latest\/<\/a>.<\/p>\n\n<p>I have also tried installing the additional dependencies required for the examples by following the instructions here: <a href=\"https:\/\/github.com\/wmyw96\/ZhuSuan\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wmyw96\/ZhuSuan<\/a>.<\/p>\n\n<p>I then try and run an example script (eg <a href=\"https:\/\/github.com\/thu-ml\/zhusuan\/blob\/master\/examples\/bayesian_neural_nets\/bayesian_nn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/thu-ml\/zhusuan\/blob\/master\/examples\/bayesian_neural_nets\/bayesian_nn.py<\/a>), importing the modules as with the code:<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import print_function\nfrom __future__ import division\nimport os\n\nimport tensorflow as tf\nfrom six.moves import range, zip\nimport numpy as np\nimport zhusuan as zs\n\nfrom examples import conf\nfrom examples.utils import dataset\n<\/code><\/pre>\n\n<p>However, on execution, I receive the following error:<\/p>\n\n<pre><code>  File \"C:\/*******\/bayesian_nn.py\", line **, in &lt;module&gt;\n    from examples import conf\n\nModuleNotFoundError: No module named 'examples'\n<\/code><\/pre>\n\n<p>I would be very grateful if anyone could anyone help identify where I have gone wrong with loading or importing the example modules.<\/p>\n\n<p>Many thanks in advance.<\/p>\n\n<p>Ross<\/p>\n","answers":[{"AnswerId":"52695240","CreationDate":"2018-10-08T04:03:32.443","ParentId":null,"OwnerUserId":"10470641","Title":null,"Body":"<p>I have solved the problem.  The additional dependencies had not installed correctly for some reason.  Running a full reinstall fixed the problem, and the program now runs successfully with the original code.  Thanks to Carlos for his comment.<\/p>\n"}]}
+{"QuestionId":52694753,"AnswerCount":1,"Tags":"<keras><backpropagation>","CreationDate":"2018-10-08T02:46:12.573","AcceptedAnswerId":null,"OwnerUserId":3963013.0,"Title":"Does customized loss function impact backpropagation in Keras?","Body":"<p>I am doing one project by using Keras with tensorflow back-end. For business reason, I need to give high weight to some kind of errors, so I implement one customized loss function. For example:<\/p>\n\n<pre><code>error = np.abs(y_true - predict)\nif error &gt; low_limit:\n    error = error * 10\n<\/code><\/pre>\n\n<p>I found that this customized loss function really changed the error value displayed during the training. What I am wondering is whether this new loss function really can change the behavior of backpropagation during the training? Because I did not see too much difference from the weights of my model. <\/p>\n","answers":[{"AnswerId":"52695220","CreationDate":"2018-10-08T04:00:45.043","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>Short answer: yes, this loss function does change the behavior of BP, but I quite surprise that this loss function works... <\/p>\n\n<p>Anyway, I feel the following loss function makes more sense<\/p>\n\n<pre><code>def my_mae( y_true, y_pred, low_bound=1e-3, coef=10. ) :\n    raw_mae = keras.losses.mae( y_true, y_pred )\n    mask = K.cast( raw_mae &lt;= low_bound, dtype='float32' )\n    #new_mae = mask * raw_mae * coef + (1-mask) * raw_mae\n    new_mae = (1 + mask * (coef-1)) *raw_mae\n    return new_mae\n<\/code><\/pre>\n\n<p>which gives extra loss when a sample loss is too low.<\/p>\n"}]}
+{"QuestionId":52695286,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-08T04:10:15.773","AcceptedAnswerId":52992960.0,"OwnerUserId":1090562.0,"Title":"Visualize the edge thickness in tensorboard","Body":"<p>I want the graph visualizer to label edges with tensor dimensions, and edge thickness to reflect total tensor size. Basically exactly the same as written in <a href=\"https:\/\/www.tensorflow.org\/guide\/graph_viz#tensor_shape_information\" rel=\"nofollow noreferrer\">this doc<\/a>:<\/p>\n\n<blockquote>\n  <p>When the serialized GraphDef includes tensor shapes, the graph\n  visualizer labels edges with tensor dimensions, and edge thickness\n  reflects total tensor size. To include tensor shapes in the GraphDef\n  pass the actual graph object (as in sess.graph) to the FileWriter when\n  serializing the graph. The images below show the CIFAR-10 model with\n  tensor shape information:<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KdEeo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KdEeo.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I pass the graph object to my summary.FileWriter:<\/p>\n\n<p><code>writer = tf.summary.FileWriter(_dir_tensorboard, graph=sess.graph, flush_secs=300)<\/code><\/p>\n\n<p>But I do not get any information about the <em>thickness<\/em> (all the lines are of the same size). I have just information about the shape of a tensor and information about the number of tensors.<\/p>\n\n<p>How can I achive the same visual effect as the tutorial claims?<\/p>\n","answers":[{"AnswerId":"52992960","CreationDate":"2018-10-25T15:25:34.393","ParentId":null,"OwnerUserId":"10558597","Title":null,"Body":"<p>There was a regression impacting 1.11 and 1.12. It should be fixed with <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/pull\/1544\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/pull\/1544<\/a>. Sorry about it :( Please file a GitHub issue next time this happens.<\/p>\n"}]}
+{"QuestionId":52695421,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm><multiclass-classification>","CreationDate":"2018-10-08T04:32:57.317","AcceptedAnswerId":null,"OwnerUserId":4556250.0,"Title":"Using Keras LSTM to predict 10+ classes: inconsistent label outputs by model.predict","Body":"<p>I am using keras bi-lstm to predict 14 classes.\nHere is my label index:<\/p>\n\n<pre><code>label_index=`{'1':0,'2':1,'D':2,'7':3,'8':4,'9':5,'10':6,'11':7,'12':8,'13':9,'14':10,'15':11,'16':12,'17':13}`\n<\/code><\/pre>\n\n<p>Here is part of my model setting:<\/p>\n\n<pre><code>model.add(Bidirectional(LSTM(64)))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(14, activation='softmax'))\n\nmodel.compile('adam', 'sparse_categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>The \"14\" in the Dense function is the number of classes. When I looked at the label distribution of <code>pred = model.predict(x_test, verbose=0)<\/code>, I found this:<\/p>\n\n<pre><code>Counter({'1': 54,\n         '10': 88,\n         '11': 5,\n         '12': 2,\n         '13': 6,\n         '14': 5,\n         '3': 65,\n         '4': 5,\n         '5': 3,\n         '6': 28,\n         '7': 10,\n         '8': 3,\n         '9': 3})\n<\/code><\/pre>\n\n<p>Where did this \"14\" come from???? It's not in my index. And there was no \"0\"!  I guess in the multi-class situation, kera.predict cannot output zero. Should I deduct 1 for every element in my \"pred\"?<\/p>\n","answers":[]}
+{"QuestionId":52695534,"AnswerCount":3,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-08T04:48:37.733","AcceptedAnswerId":null,"OwnerUserId":3963013.0,"Title":"Randomness of LSTM model","Body":"<p>I have one LSTM model like below:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv1D(3, 32, input_shape=(60, 12)))\nmodel.add(LSTM(units=256, return_sequences=False, dropout=0.25))\nmodel.add(Dense(256, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\nmodel.compile(optimizer='adam', loss='mse')\nmodel.summary() \n<\/code><\/pre>\n\n<p>Each time I use the same dataset to train it, I will get a different model. Most of the time, the performance of the trained model is acceptable, but sometime is really bad. I think that there are some randomness during the training or initialization. So how can I fix everything to get same model for each training? <\/p>\n","answers":[{"AnswerId":"52695714","CreationDate":"2018-10-08T05:09:53.423","ParentId":null,"OwnerUserId":"2846062","Title":null,"Body":"<p>You can save it<\/p>\n\n<pre><code>from keras.models import load_model\nmodel.save(\"lstm_model.h5\")\n<\/code><\/pre>\n\n<p>And load it later on<\/p>\n\n<pre><code>model = model.load(\"lstm_model.h5\")\n<\/code><\/pre>\n"},{"AnswerId":"52696193","CreationDate":"2018-10-08T05:59:38.617","ParentId":null,"OwnerUserId":"2906838","Title":null,"Body":"<p>I've experienced this problem with Keras as well, it has to do with the random seed, you can fix your random seed like this before importing the Keras, so that you could get the consistent result. <\/p>\n\n<pre><code>import numpy as np\nnp.random.seed(1000)\n\nimport os\nimport random\n\nos.environ['PYTHONHASHSEED'] = '0'\nrandom.seed(12345)\n\n# Also set the tf randomness to some fixed values like this if you need:\ntf.set_random_seed(1234)\n<\/code><\/pre>\n\n<p>This worked for me. <\/p>\n"},{"AnswerId":"52706269","CreationDate":"2018-10-08T16:09:58.153","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>Weights are initialized randomly in neural networks, so it is possibly to get different results by design. If you think about how backpropagation works and how the cost function is minimized, you will notice that you don\u00b4t have any guarantee that your network will find the \"global minima\". Fixing the seed is one idea to get reproducible results, but on the other hand you limit your network to a fixed starting position, where it probably will never reach the global minima.<\/p>\n\n<p>A lot of complex models, especially LSTMs are unstable. You could look at convolutional approaches. I noticed, they are performing almost equally and are much more stable.\n<a href=\"https:\/\/arxiv.org\/pdf\/1803.01271.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1803.01271.pdf<\/a><\/p>\n"}]}
+{"QuestionId":52695647,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><deep-learning><cloud><gcloud>","CreationDate":"2018-10-08T05:01:04.040","AcceptedAnswerId":null,"OwnerUserId":4512724.0,"Title":"Google cloud GPU machines reboot abruptly","Body":"<p>When training a model on the GPU machine, it get interrupted due to some system patch process. Since Google cloud GPU machines do not have an option of live migration, it is painful task to restart the training every time this happens. Google has clearly mentioned that there is no way around this but to restart the machines in this <a href=\"https:\/\/cloud.google.com\/compute\/docs\/gpus\/add-gpus#host-maintenance\" rel=\"nofollow noreferrer\">Doc<\/a>.<br>\nIs there a clever way to detect if the machine is rebooted and resume the training automatically.<br>\nSometimes it also happens that due to some kernel update, the CUDA drivers stop working and GPU is not visible and CUDA drivers need a re-installation. So writing a startup script to resume the training is also not a bulletproof solution.<\/p>\n","answers":[{"AnswerId":"53733347","CreationDate":"2018-12-11T22:31:44.160","ParentId":null,"OwnerUserId":"4674921","Title":null,"Body":"<p>Yes there is. If you use tensorflow, you can use it's checkpointing feature to save your progress and pick up where you left off.<\/p>\n\n<p>One great example of this is provided here: <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/ml-on-gcp\/blob\/master\/gce\/survival-training\/README-tf-estimator.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/GoogleCloudPlatform\/ml-on-gcp\/blob\/master\/gce\/survival-training\/README-tf-estimator.md<\/a><\/p>\n"}]}
+{"QuestionId":52695805,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network><tensorflow.js>","CreationDate":"2018-10-08T05:20:05.620","AcceptedAnswerId":null,"OwnerUserId":2558016.0,"Title":"Save a Tensorflow Model and Load it in Tensorflow.js","Body":"<p>In my regular python code, I have a CNN implemented. I save it using model.save and produce some four files (checkpoint, meta, index, and some other file). However, I cannot load these four files directly to tensorflow.js. Here is the sample CNN:<\/p>\n\n<pre><code>import tflearn\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\n\nconvnet = input_data(shape=[None, IMG_SIZE, IMG_SIZE, 1], name='input')\n\nconvnet = conv_2d(convnet, FIRST_NUM_CHANNEL, FILTER_SIZE, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = conv_2d(convnet, FIRST_NUM_CHANNEL*2, FILTER_SIZE, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = conv_2d(convnet, FIRST_NUM_CHANNEL*4, FILTER_SIZE, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = fully_connected(convnet, FIRST_NUM_CHANNEL*8, activation='relu')\nconvnet = dropout(convnet, 0.7)\n\nconvnet = fully_connected(convnet, NUM_OUTPUT, activation='softmax')\nconvnet = regression(convnet, optimizer='adam', learning_rate=LR, loss='categorical_crossentropy', name='targets')\n\nmodel = tflearn.DNN(convnet, tensorboard_dir='log')\n\ntrain = train_data[:7000]\ntest = train_data[-1000:]\n\n\nX = np.array([i[0] for i in train]).reshape(-1,IMG_SIZE,IMG_SIZE,1)\nY = [i[1] for i in train]\n\ntest_x = np.array([i[0] for i in test]).reshape(-1,IMG_SIZE,IMG_SIZE,1)\ntest_y = [i[1] for i in test]\n\n\nmodel.fit({'input': X}, {'targets': Y}, n_epoch=NUM_EPOCHS, validation_set=({'input': test_x}, {'targets': test_y}), \n    snapshot_step=500, show_metric=True, run_id=MODEL_NAME)\n\nmodel.save(MODEL_NAME)\nprint('MODEL SAVED:', MODEL_NAME)\n<\/code><\/pre>\n\n<p>The last two lines of this code snippet is for saving the model. I can just load the model in a flask app but I want to port it to tensorflow.js. Can anyone give me a tutorial on how to do this?<\/p>\n","answers":[{"AnswerId":"53098051","CreationDate":"2018-11-01T09:05:00.227","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p><code>tensorflowjs_converted<\/code> outputs among other files a weight file <code>weights_manifest.json<\/code> and a model topology file <code>tensorflowjs_model.pb<\/code>. To load this model in tensorflow.js, follow the steps outllined below.<\/p>\n\n<ul>\n<li>serve the folder containing the files using a server<\/li>\n<\/ul>\n\n<pre class=\"lang-none prettyprint-override\"><code>\/\/ cd to the directory containing the files\n\n\/\/ then launch the python server\npython3 -m http-server\n\n\/\/ or install and launch npm module http-server\nnpm install -g http-server\nhttp-server --cors -c1 .\n<\/code><\/pre>\n\n<ul>\n<li>create a js script to load the model<\/li>\n<\/ul>\n\n\n\n<pre><code>(async () =&gt; {\n   ...\n   const model = await tf.loadFrozenModel('http:\/\/localhost:8080\/tensorflowjs_model.pb', 'http:\/\/localhost:8080\/weights_manifest.json')\n })()\n<\/code><\/pre>\n\n<hr>\n\n<p>There is a difference between <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.3\/#loadModel\" rel=\"nofollow noreferrer\">loadModel<\/a> and <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.3\/#loadFrozenModel\" rel=\"nofollow noreferrer\">loadFrozenModel<\/a>. <\/p>\n\n<ul>\n<li><p><code>loadModel<\/code> is for loading model that were saved locally. The model could either be retrieved from the indexDB of the browser or from the localStorage. \nMaybe it can be used to retrieve a model saved by another tensorflow API different than Js, but the user will be required to select the file by using <code>tf.io.browserFiles<\/code> ( I haven't tried it)<\/p><\/li>\n<li><p><code>loadFrozenModel<\/code> is for loading model that are served by a server<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":52695829,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><object-detection><tflearn>","CreationDate":"2018-10-08T05:23:20.830","AcceptedAnswerId":null,"OwnerUserId":10278996.0,"Title":"Object Detection using TFLearn or Tensorflow object detection API","Body":"<p>I want to build a object detection model and cannot figure out whether to go for tensorflow object detection API or use TFLearn to build a model.Can someone tell me the pros and cons of both.<\/p>\n","answers":[{"AnswerId":"54467079","CreationDate":"2019-01-31T18:28:44.460","ParentId":null,"OwnerUserId":"9441041","Title":null,"Body":"<p>Both are fine. Object detection API is a standard builtin library you can quickly clone and use. It's good to learn it since it's popular across the industry because of the following reasons.<\/p>\n\n<ol>\n<li>It's less error prone.<\/li>\n<li>It saves your time<\/li>\n<li>You can easily use tensorboard with it, to analyze the training stats.<\/li>\n<li>It has inbuilt architectures like faster RCNN and SSD. You just have to understand using them. Understanding various hyperparameters in the config file and knowing to tune them is also a skill.<\/li>\n<\/ol>\n\n<p>On the other hand if you want to make your own custom made models, you can go for learning tensorflow. Actually these both are not alternatives and should be learnt separately. In fact I would also suggest reading the research papers for the object detection algos available in the API. <\/p>\n\n<p>Besides object detection tensorflow will also help you in long run, whether you are working on segmentation algos or autoencoders or anything else in DNNs.<\/p>\n"}]}
+{"QuestionId":52695913,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning><gradient-descent><loss-function>","CreationDate":"2018-10-08T05:33:10.430","AcceptedAnswerId":null,"OwnerUserId":2129212.0,"Title":"Custom loss function in Keras to penalize false negatives","Body":"<p>I am working on a medical dataset where I am trying to have as less false negatives as possible. A prediction of \"disease when actually no disease\" is okay for me but a prediction \"no disease when actually a disease\" is not. That is, I am okay with <code>FP<\/code> but not <code>FN<\/code>.<\/p>\n\n<p>After doing some research, I found out ways like <code>Keeping higher learning rate for one class<\/code>, <code>using class weights<\/code>,<code>ensemble learning with specificity\/sensitivity<\/code> etc.<\/p>\n\n<p>I achieved the near desired result using class weights like <code>class_weight = {0 : 0.3,1: 0.7}<\/code> and then calling the <code>model.fit(class_weights=class_weight)<\/code>. This gave me very low FN but a pretty high FP. I am trying to reduce FP as much as possible keeping FN very low.<\/p>\n\n<p>I am struggling to write a <strong>custom loss function<\/strong> using <code>Keras<\/code> which will help me to penalize the false negatives. Thanks for the help.<\/p>\n","answers":[{"AnswerId":"55787295","CreationDate":"2019-04-21T22:25:40.487","ParentId":null,"OwnerUserId":"8272600","Title":null,"Body":"<p>I'll briefly introduce the concepts we're trying to tackle. <\/p>\n\n<h2>Recall<\/h2>\n\n<p><em>From <strong>all<\/strong> that were <strong>positive<\/strong>, how many did our model predict as positive?<\/em><\/p>\n\n<p>All that were positive = <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=TP\" alt=\"positive\"><\/p>\n\n<p>What our model said were positive = <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=P%20%3D%20TP%2BFN\" alt=\"said positive\"><\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=Recall%20%3D%20%5Cfrac%7BTP%7D%7BP%7D%20%3D%20%5Cfrac%7BTP%7D%7BTP%2BFN%7D\" alt=\"recall\"><\/p>\n\n<p>Since recall is inversely proportional to FN, improving it decreases FN.<\/p>\n\n<h2>Specificity<\/h2>\n\n<p><em>From <strong>all<\/strong> that were <strong>negative<\/strong>, how many did our model predict as negative?<\/em><\/p>\n\n<p>All that were negative = <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=TN\" alt=\"negative\"><\/p>\n\n<p>What our model said were negative = <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=N%20%3D%20TN%2BFP\" alt=\"said negative\"><\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=Specificity%20%3D%20%5Cfrac%7BTN%7D%7BN%7D%20%3D%20%5Cfrac%7BTN%7D%7BTN%2BFP%7D\" alt=\"specificity\"><\/p>\n\n<p>Since recall is inversely proportional to FP, improving it decreases FP.<\/p>\n\n<p>In your next searches, or whatever classification-related activity you perform, knowing these is going to give you an extra edge in communication and understanding.<\/p>\n\n<hr>\n\n<h2>A Solution<\/h2>\n\n<p>So. These two concepts, as you mas have figured out already, are opposites. This means that <strong>increasing one is likely to decrease the other<\/strong>.<\/p>\n\n<p>Since you want <strong>priority<\/strong> on recall, but don't want to loose too much in specificity, you can combine both of those and attribute weights. Following what's clearly explained in <a href=\"https:\/\/datascience.stackexchange.com\/questions\/33587\/keras-custom-loss-function-as-true-negatives-by-true-negatives-plus-false-posit\">this answer<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport keras.backend as K\n\ndef binary_recall_specificity(y_true, y_pred, recall_weight, spec_weight):\n\n    TN = np.logical_and(K.eval(y_true) == 0, K.eval(y_pred) == 0)\n    TP = np.logical_and(K.eval(y_true) == 1, K.eval(y_pred) == 1)\n\n    FP = np.logical_and(K.eval(y_true) == 0, K.eval(y_pred) == 1)\n    FN = np.logical_and(K.eval(y_true) == 1, K.eval(y_pred) == 0)\n\n    # Converted as Keras Tensors\n    TN = K.sum(K.variable(TN))\n    FP = K.sum(K.variable(FP))\n\n    specificity = TN \/ (TN + FP + K.epsilon())\n    recall = TP \/ (TP + FN + K.epsilon())\n\n    return 1.0 - (recall_weight*recall + spec_weight*specificity)\n\n<\/code><\/pre>\n\n<p>Notice <code>recall_weight<\/code> and <code>spec_weight<\/code>? They're weights we're attributing to each of the metrics. For distribution convention, they should always add to <code>1.0<\/code>\u00b9, e.g. <code>recall_weight=0.9<\/code>, <code>specificity_weight=0.1<\/code>. The intention here is for you to see what proportion best suits your needs.<\/p>\n\n<p>But Keras' loss functions must only receive <code>(y_true, y_pred)<\/code> as arguments, so let's define a wrapper:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Our custom loss' wrapper\ndef custom_loss(recall_weight, spec_weight):\n\n    def recall_spec_loss(y_true, y_pred):\n        return binary_recall_specificity(y_true, y_pred, recall_weight, spec_weight)\n\n    # Returns the (y_true, y_pred) loss function\n    return recall_spec_loss\n<\/code><\/pre>\n\n<p>And onto using it, we'd have<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Build model, add layers, etc\nmodel = my_model\n# Getting our loss function for specific weights\nloss = custom_loss(recall_weight=0.9, spec_weight=0.1)\n# Compiling the model with such loss\nmodel.compile(loss=loss)\n<\/code><\/pre>\n\n<p>\u00b9 The weights, added, must total <code>1.0<\/code>, because in case both <code>recall=1.0<\/code> and <code>specificity=1.0<\/code> (the perfect score), the formula<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=loss%20%3D%201.0%20-%20%28w1%2Arecall%20%2B%20w2%2Aspecificity%29\" alt=\"loss1\"><\/p>\n\n<p>Shall give us, for example,<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=loss%20%3D%201.0%20-%20%280.9%2A1.0%20%2B%200.1%2A1.0%29%20%3D%200.0\" alt=\"loss2\"><\/p>\n\n<p>Clearly, if we got the perfect score, we'd want our loss to equal 0.<\/p>\n"}]}
+{"QuestionId":52696039,"AnswerCount":2,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2018-10-08T05:45:16.193","AcceptedAnswerId":null,"OwnerUserId":1381666.0,"Title":"Get a single value prediction from a 2-class softmax output","Body":"<p>I have a binary classification problem but with a <code>Dense(2, activation='softmax')<\/code> layer at the end. So I get predictions like <code>[0.2, 0.8]<\/code> or <code>[1.0, 0.0]<\/code> for instance. Am I calculating a resulting single-value-prediction correctly?<\/p>\n\n<p><code>single_value_prediction = predictions[predicted_class] if predictions[1] &gt; predictions[0] else 1 - predictions[predicted_class]\n<\/code><\/p>\n\n<p>So <code>[0.2, 0.8]<\/code> would result in <code>0.8<\/code> and <code>[1.0, 0.0]<\/code> in <code>0.0<\/code>?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52710701","CreationDate":"2018-10-08T22:03:39.343","ParentId":null,"OwnerUserId":"9415043","Title":null,"Body":"<p>If you want \"[0.2, 0.8] would result in 0.8 and [1.0, 0.0] in 0.0\", you could try <code>prediction = model.predict(data)[:, 1]<\/code>. <\/p>\n\n<p>If you want to turn the probabilities into classes at the same time, try <code>prediction = model.predict(data)[:, 1] &gt; 0.5<\/code>. Or use np.argmax, as @today suggested.<\/p>\n"},{"AnswerId":"52700456","CreationDate":"2018-10-08T10:37:56.520","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>After getting the predictions using:<\/p>\n\n<pre><code>preds = model.predict(test_data)\n<\/code><\/pre>\n\n<p>the <code>preds<\/code> array would have a shape of <code>(n_samples, 2)<\/code>. Now to find the predicted class you can easily use <code>np.argmax()<\/code>:<\/p>\n\n<pre><code>preds_classes = np.argmax(preds, axis=-1)\n<\/code><\/pre>\n\n<p>Alternatively, if your model is a Sequential model, i.e. <code>keras.models.Sequential<\/code>, (and not a model defined using functional API, i.e. <code>keras.model.Model<\/code>) you can use <code>predict_classes()<\/code> method directly:<\/p>\n\n<pre><code>preds_classes = model.predict_classes(test_data)\n<\/code><\/pre>\n\n<p>If you would like to find the probability of predicted class (i.e. class with the highest probability) then you can use <code>np.max()<\/code>:<\/p>\n\n<pre><code>preds_class_prob = np.max(preds, axis=-1) \n<\/code><\/pre>\n"}]}
+{"QuestionId":52696396,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T06:17:09.980","AcceptedAnswerId":null,"OwnerUserId":2290560.0,"Title":"Is it average loss for tf.estimator.DNNRegressor in tensorflow the same with MSE?","Body":"<p>In <code>model.evaluate of tensorflow<\/code>, there is a variable called \"<strong>average_loss<\/strong>\". Is it the same as MSE between labels and predictions? However in <code>tf.losses<\/code>, there is another function <code>mean_squared_error<\/code> as well. Which one is the correct MSE?<\/p>\n","answers":[{"AnswerId":"56315682","CreationDate":"2019-05-26T17:07:57.897","ParentId":null,"OwnerUserId":"5118339","Title":null,"Body":"<p>The loss is computed depending on the model (<a href=\"https:\/\/developers.google.com\/machine-learning\/glossary\/#loss\" rel=\"nofollow noreferrer\">see here<\/a>).<\/p>\n\n<p>Very likely your model uses mean squared error for the loss function, thus the <strong>mean_squared_error<\/strong> is also the <strong>loss<\/strong> variable. <\/p>\n\n<p>Instead, the <strong>average_loss<\/strong> would be the loss averaged over the last N iterations, because sometimes the training is performed with small batch size and the loss might be noisy.<\/p>\n"}]}
+{"QuestionId":52696408,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-10-08T06:17:57.983","AcceptedAnswerId":null,"OwnerUserId":9265440.0,"Title":"Tensorflow Object Detection API - Detecting the humans not wearing Helmet","Body":"<p>*** Please note, my previous problem of detecting withouthelmet as NA is resolved.<\/p>\n\n<p>Now I have a new issue, I used 1000 images for detecting humans not wearing helmets and also 1000 images for humans wearing helmet and 1000 images for only humans. I used SSD_mobilenet_v1_pets.config file for training.<\/p>\n\n<p>Here is my pbtxt file  <\/p>\n\n<pre><code>item {\n    id: 1\n    name: 'withouthelmet'\n}\nitem {\n    id: 2\n    name: 'withhelmet'\n}\nitem {\n    id: 3\n    name: 'person'\n}\n<\/code><\/pre>\n\n<p>sample training Image\n<a href=\"https:\/\/i.stack.imgur.com\/sBt0l.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sBt0l.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After the training my model detect every car as person..<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/n2Ahk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n2Ahk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Is that because of using ssd_mobilenet model(id: 1 for person but I used id: 1 as withouthelmet and id:3 for car but I used id:3 for person)<\/strong><\/p>\n\n<p>Pls help me to solve this problem<\/p>\n","answers":[{"AnswerId":"52698252","CreationDate":"2018-10-08T08:29:42.590","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Have you used the pre-trained SSD-MobileNetV1 model trained on the pets dataset?\nI think you better use a model trained on COCO dataset since it has persons, in contrast to pets.\nOf course that if you train your model it will learn to detect the person as well, but since you don't have a lot of examples of persons without a helmet, it would probably be better to start with a model which knows what a person is.<\/p>\n\n<p>Regarding your questions, if you only want to detect people without helmet, you can simply drop everything else in the pbtxt file, only put <\/p>\n\n<pre><code>item {\n  id: 1\n  name: 'withouthelmet'\n  display_name: 'withouthelmet'\n}\n<\/code><\/pre>\n\n<p>change the number of categories in the config file to 1, and fine-tune the model.<\/p>\n"},{"AnswerId":"52785359","CreationDate":"2018-10-12T18:39:42.500","ParentId":null,"OwnerUserId":"9356310","Title":null,"Body":"<ol>\n<li>Have you set num_classes to 1 in your config? <\/li>\n<li>Please note that min_negatives_per_image means min # of negative anchors (instead of images) so you data mix has nothing to do with this parameter.<\/li>\n<\/ol>\n\n<p>I had to modify earlier answer - if you add a background image(image with no gt boxes) to the dataset, it should help reduce false positives. Sorry I got confused with some other stuff.<\/p>\n"}]}
+{"QuestionId":52697586,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-08T07:44:09.467","AcceptedAnswerId":null,"OwnerUserId":8776654.0,"Title":"I have a trouble when I use Tensorflow to save model. File too large","Body":"<pre><code>2018-10-08 07:03:40.303219 Saving checkpoint of model...\nTraceback (most recent call last):\n  File \"train_reconstruction_conv5.py\", line 188, in &lt;module&gt;\n    save_path = saver.save(sess, checkpoint_name)\n  File \"\/home\/fanghb\/python35\/lib\/python3.5\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1645, in save\n    meta_graph_filename, strip_default_attrs=strip_default_attrs)\n  File \"\/home\/fanghb\/python35\/lib\/python3.5\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1683, in export_meta_graph\n    graph_def=ops.get_default_graph().as_graph_def(add_shapes=True),\n  File \"\/home\/fanghb\/python35\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3005, in as_graph_def\n    result, _ = self._as_graph_def(from_version, add_shapes)\n  File \"\/home\/fanghb\/python35\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2963, in _as_graph_def\n    c_api.TF_GraphToGraphDef(self._c_graph, buf)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot serialize protocol buffer of type tensorflow.GraphDef as the serialized size (2950754208bytes) would be larger than the limit (2147483647 bytes)\n<\/code><\/pre>\n\n<p>When I train my model, I raise above exceptions.Moreover, I found my meta file is very large in my checkpoints.<\/p>\n","answers":[{"AnswerId":"52702471","CreationDate":"2018-10-08T12:36:53.453","ParentId":null,"OwnerUserId":"8776654","Title":null,"Body":"<p>tf.data.Dataset.from_tensor_slices((features, labels)) embed the features and labels arrays in your TensorFlow graph as tf.constant() operations. \nsee more <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#consuming_numpy_arrays\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/datasets#consuming_numpy_arrays<\/a><\/p>\n"}]}
+{"QuestionId":52697809,"AnswerCount":1,"Tags":"<tensorflow><dll><site-packages>","CreationDate":"2018-10-08T07:58:54.183","AcceptedAnswerId":null,"OwnerUserId":6002424.0,"Title":"import tensorflow as tf gives errors even after installing without errors","Body":"<p>I've been trying to solve this issue for long time. After many attempts in my existing python=3.5 env (in Anaconda), I deleted it and created a new env again just to properly install tensorflow with gpu support. To do that I did <code>conda create -n FALL python=3.5<\/code> and then <code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.11.0-cp35-cp35m-win_amd64.whl<\/code> as suggested in tensorflow site, because I am on Windows 10 and have Anaconda. There was no error during installation. And I have been using Cuda 8.0 and cuDNN for my other deep learning practices without any hinders. So I could type <code>python<\/code> in my activated env. To check if tensorflow is installed accurately I did <code>import tensorflow as tf<\/code> and got<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\username\\Anaconda3\\envs\\FALL\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n\n<p>I would like someone could help me as possible.<\/p>\n","answers":[{"AnswerId":"53821718","CreationDate":"2018-12-17T19:18:34.430","ParentId":null,"OwnerUserId":"10802834","Title":null,"Body":"<p>Not sure if you are still looking for this question to be answered, but seeing as I just battled this very issue for the past few days, I figured I'd leave my two cents here in case it helps anyone moving forward.<\/p>\n\n<p>As an initial note, I was not installing tensorflow through anaconda, but rather just via a pip install in a python virtualenv. I expect the solution to this issue would be similar using either anaconda or just pip.<\/p>\n\n<p>It really just turned out in the end that I had one or more programs installed in incompatible versions for tensorflow.<\/p>\n\n<p>As of my writing of this, tensorflow-gpu requires CUDA 9.0, CUPTI (should be included with the CUDA install), and cuDNN version greater than 7.2 (but note that this version should be compatible with your specific install of CUDA). My problem was that I was unwittingly using CUDA 10.0, which I think is theoretically compatible with tensorflow-gpu (based on cursory searches, it seems to work for some people), but didn't work for me at all.<\/p>\n\n<p>Long story short, here is exactly the setup\/process that worked for me:<\/p>\n\n<ol>\n<li><p>CUDA compatible GPU (<a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-gpus<\/a> to check yours specifically<\/p><\/li>\n<li><p>python 3.5.4 <\/p><\/li>\n<li><p>tensorflow-gpu version 1.12.0 <\/p><\/li>\n<li><p>CUDA 9.0<\/p><\/li>\n<li>CUPTI (again, this is included with your CUDA 9.0 install)<\/li>\n<li>cuDNN 9.0 version 7.4.1.5. To install this correctly, make sure the .dll in \/bin gets copied into the \/bin folder of Cuda 9, the .h file from \/include gets copied into the Cuda 9 \/include folder, and the .lib file from \/lib\/x64 folder gets copied into the \/lib\/x64 folder for Cuda 9 (this all seems self evident, but oddly some installation guides I've seen only reference copying the .dll over).<\/li>\n<li>visual c++ 2017<\/li>\n<li>Microsoft Visual C++ 2015 Redistributable Update 3 (this can be found under 'Redistributables and Build Tools' here: <a href=\"https:\/\/visualstudio.microsoft.com\/vs\/older-downloads\/\" rel=\"nofollow noreferrer\">https:\/\/visualstudio.microsoft.com\/vs\/older-downloads\/<\/a>). Do note that to install this I had to go into programs and settings and manually uninstall the visual c++ 2017 redistributable that was installed when I installed visual c++ to begin with.<\/li>\n<li>Include both CUDA 9 and CUPTI in your path (CUDA (and cuDNN) for me was in C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v9.0\\bin, and CUPTI was in C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v9.0\\extras\\CUPTI\\libx64)<\/li>\n<\/ol>\n\n<p>After all this importing tensorflow worked fine for me.<\/p>\n\n<p>I honestly hope this helps someone out there, as this process was incredibly arduous for me personally and I'd love to be able to lessen that for someone else.<\/p>\n"}]}
+{"QuestionId":52697871,"AnswerCount":1,"Tags":"<pytorch><torch>","CreationDate":"2018-10-08T08:03:17.007","AcceptedAnswerId":52770738.0,"OwnerUserId":7475297.0,"Title":"what does THCudaTensor_data ( and THC in general ) do?","Body":"<p>The program I am inspecting uses pytorch to load weights and cuda code to do the computations with the weights. My understanding of THC library is how tensors are implemented in the backend of pytorch ( and torch, maybe? ).  <\/p>\n\n<ol>\n<li>how is THC implemented ( I would really appreiciate some details if possible )?<\/li>\n<li>what does THCudaTensor_data( THC_state, THCudaTensor* ) do? ( from the way it is used in the code, it seems like it is used to convert pytorch's tensor to an array in cuda. if this is the case, then would the function preserve all elements and the length of the array?)<\/li>\n<\/ol>\n","answers":[{"AnswerId":"52770738","CreationDate":"2018-10-12T00:53:15.103","ParentId":null,"OwnerUserId":"7475297","Title":null,"Body":"<p>I am still not exactly sure of the inner-workings of THCudaTensor_data, but the behaviour that was tripping me up was: for n-dimensional tensor, THCudaTensor_data returns a flattened 1D array of the tensor.<\/p>\n\n<p>Hope this helps<\/p>\n"}]}
+{"QuestionId":52697996,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-08T08:12:21.867","AcceptedAnswerId":52700437.0,"OwnerUserId":3830495.0,"Title":"Why is there a `Const` and a `Mean` operation in metrics node in keras?","Body":"<p>I try to implement a custom loss function using <code>tensorflow<\/code>'s <code>keras<\/code> interface.<\/p>\n\n<p>I added the same function to the <code>loss<\/code> and to the <code>metrics<\/code>:<\/p>\n\n<pre><code>def one_class_loss(y_true, y_pred):\n    return tf.norm(y_pred, name='distance')\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(0.001),\n              loss=one_class_loss,\n              metrics=[one_class_loss])\n<\/code><\/pre>\n\n<p>In <code>TensorBoard<\/code>, I see the following as part of the network:\n<a href=\"https:\/\/i.stack.imgur.com\/EdafE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/EdafE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Why is there a <code>Mean<\/code> and a <code>Const<\/code> before the training? What do they represent?<\/p>\n","answers":[{"AnswerId":"52700437","CreationDate":"2018-10-08T10:36:37.837","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>The <code>Const<\/code> part is used to avoid numerical instabilities. The derivative of the sqrt is (1 \/ sqrt) * inner_derivative and therefore if the sqrt is 0 (e.g. minimal), the derivative will result to an <code>inf<\/code> value and blow up your model. However, if you add a const to the sum the derivative is 0, due to the inner_derivative being 0. In all other cases the <code>Const<\/code> has minimal impact on your performance.<\/p>\n\n<p>The <code>Mean<\/code> is just used to make the optimizer work properly. It makes no sense to optimize a multi valued output, so the optimizer reduces your loss to a single value automatically by applying a mean operation, across the batch size, e.g. Mean Squared Error. (<a href=\"https:\/\/datascience.stackexchange.com\/questions\/23257\/gradient-descent-with-vector-valued-loss\">https:\/\/datascience.stackexchange.com\/questions\/23257\/gradient-descent-with-vector-valued-loss<\/a>) <\/p>\n"}]}
+{"QuestionId":52698202,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><floating-accuracy><tfrecord>","CreationDate":"2018-10-08T08:26:26.863","AcceptedAnswerId":null,"OwnerUserId":10471887.0,"Title":"Error preprocessing the input data when using Tensorflow Dataset API","Body":"<p>I have images of [64,512,5] stored in *.npy files which I convert into *.tfrecords files.<\/p>\n\n<p>I have verified that the reading of said records corresponds correctly with what is present in the *.npy files. However, when I perform some operation on the parser, like adding 1 to each pixel of the image, the result is not the expected one. The result should be 65*512*5 = 163840 but it is 163839.99980013957 (not always the same)<br>\nI have tried to perform different operations like tf.subtract, but the results are the same.<\/p>\n\n<p>Could someone tell me what is wrong?<\/p>\n\n<pre><code>import re\nimport ast\nimport sys, select\nimport random as rn\nfrom glob import glob\nfrom tqdm import tqdm\nfrom datetime import datetime\nfrom configparser import SafeConfigParser\n\nimport numpy as np\nimport numpy.ma as ma\nimport scipy.misc\n\nimport os.path\nfrom os import mkdir, stat\nfrom os.path import exists, dirname, abspath\nfrom os.path import join as dir_join\nimport tensorflow as tf\n\n''' File hierarchy\n'''\n_code_dir             = dirname(abspath(__file__))\n_python_dir           = dirname(_code_dir)\n_model_dir            = dirname(_python_dir)\n_project_dir          = dirname(_model_dir)\n_ml_dir               = dirname(_project_dir)\n_srv_dir              = dirname(_ml_dir)\n_root_datasets_dir    = dir_join(_srv_dir,'machine_learning','data_sets\/ssd_prepared')\n_config_dir           = dir_join(_python_dir, 'config')\n\n'''Data sets directories\n'''\nTHIS_DATA_SET_DIR     = 'Sph_50m' #WARNING: Global variable also used in helper.py\n_data_dir             = dir_join(_root_datasets_dir, THIS_DATA_SET_DIR)\n_data_set_dir         = dir_join(_data_dir,'ImageSet')\n_data_npy_dir         = dir_join(_data_dir,'data')\n_data_tfRecord_dir    = dir_join(_data_dir,'tfRecord')\n\n''' Configuration parser\n'''\ncfg_parser = SafeConfigParser()\ncfg_parser.read(dir_join(_config_dir,'cfg_model.ini'))\n\n''' Private variables\n'''\n_batch_size        = cfg_parser.getint(section='train', option='batch_size')\n_max_epoch         = cfg_parser.getint(section='train', option='max_epoch')\n_standarize        = cfg_parser.getboolean(section='train', option='standarize_input')\n\n_input_shape       = ast.literal_eval(cfg_parser.get(section='data_shape', option='input_shape'))\n_label_channel     = cfg_parser.getint(section='data_shape', option='label_channel')\n_track_channel     = cfg_parser.getint(section='data_shape', option='track_channel')\n_mask_channel      = cfg_parser.getint(section='data_shape', option='mask_channel')\n\n_data_train        = cfg_parser.get(section='data_set', option='data_train')\n_data_val          = cfg_parser.get(section='data_set', option='data_val')\n_data_test         = cfg_parser.get(section='data_set', option='data_test')\n\ndef _int64_feature(value):\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=value.reshape(-1)))\n\ndef _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef _floats_feature(value):\n    return tf.train.Feature(float_list=tf.train.FloatList(value=value.reshape(-1)))\n\ndef numpy_to_TFRecord():\n    if not exists(_data_tfRecord_dir): mkdir(_data_tfRecord_dir)\n\n    for dataset in [_data_train, _data_val, _data_test]:\n        tfRecord_folder = dir_join(_data_tfRecord_dir, dataset)\n        if not exists(tfRecord_folder): mkdir(tfRecord_folder)\n\n        #Retrieve list of files\n        projections_dir=[]\n        file_ = open(dir_join(_data_set_dir, dataset+'.txt'), 'r')\n        for x in file_.readlines():\n            file_nat = x.strip()+'.npy'\n            filename = dir_join(_data_npy_dir, file_nat)\n            assert exists(filename), \"{} doesn't exist\".format(filename)\n            projections_dir.append(filename)\n        file_.close()\n\n        totaltfRecordSize = 0\n        numFile = 0\n\n        for projection_dir in tqdm(projections_dir, ncols= 100, desc = 'TFRecord {}'.format(dataset)):\n            scanName = projection_dir.split('\/')[-1].split('.')[0]\n            if totaltfRecordSize &gt; 100*(10**6) or totaltfRecordSize == 0:\n                # address to save the TFRecords file\n                train_filename = dir_join(tfRecord_folder, \\\n                                        str(numFile) + '_' + dataset +'.tfrecords')\n                # open the TFRecords file                \n                writer = tf.python_io.TFRecordWriter(train_filename)\n\n                numFile += 1\n                totaltfRecordSize = 0\n\n            # Load the image\n            projection = np.load(projection_dir)\n            image = projection[:,:,:_label_channel]\n            label = projection[:,:,_label_channel].astype(int)\n            mask  = projection[:,:,_mask_channel].astype(int)\n            track = projection[:,:,_track_channel].astype(int)\n\n            # Create a feature\n            feature = {'image': _floats_feature(image),\n                       'label': _int64_feature(label),\n                       'mask' : _int64_feature(mask),\n                       'track': _int64_feature(track),\n                       'scanName': _bytes_feature(tf.compat.as_bytes(scanName))}\n\n            # Create an example protocol buffer\n            example = tf.train.Example(features=tf.train.Features(feature=feature))\n\n            # Serialize to string and write on the file\n            writer.write(example.SerializeToString())\n\n            fileSize = stat(train_filename).st_size\n            totaltfRecordSize += fileSize\n\n        writer.close()\n        sys.stdout.flush()\n\ndef readTFRecord():\n    # Transforms a scalar string `example_proto` into a pair of a scalar string and\n    # a scalar integer, representing an image and its label, respectively.\n    image_dim = _input_shape[0] * _input_shape[1] * _label_channel\n    label_dim = _input_shape[0] * _input_shape[1]\n    mean = np.load(dir_join(_data_dir,'mean.npy'))\n    std  = np.load(dir_join(_data_dir,'std.npy'))\n    mean_tf = tf.convert_to_tensor(mean, dtype=tf.float32, name='mean')\n    std_tf = tf.convert_to_tensor(std,  dtype=tf.float32, name='std')\n\n    with tf.variable_scope('TFRecord'):\n        def _parse_function(example_proto):\n            with tf.variable_scope('parser'):\n                features = {'image': tf.FixedLenFeature([image_dim], tf.float32),\n                            'label': tf.FixedLenFeature([label_dim], tf.int64),\n                            'mask' : tf.FixedLenFeature([label_dim], tf.int64),\n                            'track': tf.FixedLenFeature([label_dim], tf.int64),\n                            'scanName': tf.FixedLenFeature([], tf.string)}\n\n                parsed_features = tf.parse_single_example(example_proto, features)\n\n                # Reshape image data into the original shape\n                image = tf.reshape(parsed_features['image'], [_input_shape[0], _input_shape[1], _label_channel], name='image')\n                label = tf.reshape(parsed_features['label'], _input_shape, name='lable_reshape')\n                mask  = tf.reshape(parsed_features['mask'],  _input_shape, name='mask_reshape')\n                track = tf.reshape(parsed_features['track'], _input_shape, name='track_reshape')\n                scanName = parsed_features['scanName']\n\n                image = image + tf.constant(1., dtype=tf.float32)\n\n                return image, label, mask, track, scanName\n\n        training_filenames = glob(dir_join(_data_tfRecord_dir, _data_train, '*.tfrecords'))\n        validation_filenames = glob(dir_join(_data_tfRecord_dir, _data_val, '*.tfrecords'))\n\n        filenames = tf.placeholder(tf.string, shape=[None], name='filenames')\n        dataset = tf.data.TFRecordDataset(filenames)\n\n        dataset = dataset.map(_parse_function, num_parallel_calls=20)  # Parse the record into tensors.\n        dataset = dataset.shuffle(buffer_size=10000)\n        dataset = dataset.batch(_batch_size, drop_remainder=True)\n        dataset = dataset.prefetch(buffer_size=10)\n        iterator = dataset.make_initializable_iterator()\n\n        next = iterator.get_next()\n        sess = tf.Session()\n\n        while True:\n            sess.run(iterator.initializer, feed_dict={filenames: training_filenames})\n            try:\n                img, _, _, _, scanX = sess.run(next)\n                for i, scan in enumerate(scanX):\n                    print(scan.decode(\"utf-8\"))\n                    projection = np.load(dir_join(_data_npy_dir, scan.decode(\"utf-8\") + '.npy'))\n                    imagenp = projection[:,:,:_label_channel]\n\n                    if np.abs(np.sum(img[i,...] - imagenp)) &gt; 0.:\n                        print(np.sum(img[i,...] - imagenp))\n\n            except tf.errors.OutOfRangeError:\n                break\n\n    return training_filenames, validation_filenames, filenames, iterator\n\nif __name__ == '__main__':\n    numpy_to_TFRecord()\n    readTFRecord()\n<\/code><\/pre>\n\n<p>The test I'm doing in the previous code is to convert the *.npy files to *.tfrecords. Then, I compare the *.trecords with the *.npy. The value should be 0 if both images were identical.<\/p>\n\n<pre><code>img, _, _, _, scanX = sess.run(next)\n    for i, scan in enumerate(scanX):\n         print(scan.decode(\"utf-8\"))\n         projection = np.load(dir_join(_data_npy_dir, scan.decode(\"utf-8\") + '.npy'))\n         imagenp = projection[:,:,:_label_channel]\n\n         print(np.sum(img[i,...] - imagenp))\n<\/code><\/pre>\n\n<p>If the data is not preprocessed, these images are the same, however, if we perform some kind of transformation, the results do not match what was expected. In this case we are adding 1 to each pixel of the image, so the total difference should be 64 * 512 * 5.<\/p>\n\n<pre><code>image = image + tf.constant(1., dtype=tf.float32)\n<\/code><\/pre>\n\n<p>I would like to solve this error, since so far I have not been able to obtain the results obtained by my neural network using feed_dict instead of Tensorflow Dataset API, and this is the only point where I can observe a difference in the input data.<\/p>\n","answers":[]}
+{"QuestionId":52698425,"AnswerCount":0,"Tags":"<android><opencv><tensorflow>","CreationDate":"2018-10-08T08:40:41.310","AcceptedAnswerId":null,"OwnerUserId":10081622.0,"Title":"Opencv and Tensorflow not working in one Android project","Body":"<p>I am currently working on an android app.<\/p>\n\n<p>I am using Android NDK to run opencv in C++<\/p>\n\n<p>Everything was working until when I imported tensorflow to detect rotation of images (an image classifier). The tensorflow code is running fine (I have tested it in another android project) but when I am combing both the codes I am getting error. Please help my project is lagging due to this.<\/p>\n\n<p>I have checked tensorflow github repo similar issues have stated but no solution found there.<\/p>\n\n<pre><code>    --------- beginning of crash\n<\/code><\/pre>\n\n<p>10-09 22:21:15.572 7387-7387\/? E\/AndroidRuntime: FATAL EXCEPTION: main\n    Process: com.baurasia.pranav.digimother, PID: 7387\n    java.lang.UnsatisfiedLinkError: dalvik.system.PathClassLoader[DexPathList[[zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/base.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_dependencies_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_resources_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_0_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_1_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_2_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_3_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_4_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_5_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_6_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_7_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_8_apk.apk\", zip file \"\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_9_apk.apk\"],nativeLibraryDirectories=[\/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/lib\/arm64, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/base.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_dependencies_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_resources_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_0_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_1_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_2_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_3_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_4_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_5_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_6_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_7_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_8_apk.apk!\/lib\/arm64-v8a, \/data\/app\/com.baurasia.pranav.digimother-0NgVxe-pLiV2LPUwElaVZQ==\/split_lib_slice_9_apk.apk!\/lib\/arm64-v8a, \/system\/lib64, \/vendor\/lib64]]] couldn't find \"libopencv_java3.so\"\n        at java.lang.Runtime.loadLibrary0(Runtime.java:1011)\n        at java.lang.System.loadLibrary(System.java:1657)\n        at com.scanlibrary.ScanActivity.(ScanActivity.java:126)\n        at java.lang.Class.newInstance(Native Method)\n        at android.app.Instrumentation.newActivity(Instrumentation.java:1174)\n        at android.app.ActivityThread.performLaunchActivity(ActivityThread.java:2891)\n        at android.app.ActivityThread.handleLaunchActivity(ActivityThread.java:3078)\n        at android.app.ActivityThread.-wrap11(Unknown Source:0)\n        at android.app.ActivityThread$H.handleMessage(ActivityThread.java:1794)\n        at android.os.Handler.dispatchMessage(Handler.java:106)\n        at android.os.Looper.loop(Looper.java:187)\n        at android.app.ActivityThread.main(ActivityThread.java:7025)\n        at java.lang.reflect.Method.invoke(Native Method)\n        at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:514)\n        at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:888)<\/p>\n","answers":[]}
+{"QuestionId":52698444,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-08T08:41:44.293","AcceptedAnswerId":52736086.0,"OwnerUserId":205147.0,"Title":"Minimize angle in pytorch","Body":"<p>I'm trying to figure out how to minimize a scalar in pytorch that represents the angle of an axis\/angle rotation. My <code>target<\/code> is a sample set of 3D vectors, my <code>input<\/code> is the <code>target<\/code> rotated by a particular axis\/angle rotation (plus some gaussian noise). The axis is known and fixed. I want to find the angle using pytorch. <\/p>\n\n<p>What I have so far is this <code>nn.Module<\/code>:<\/p>\n\n<pre><code>class AngleModel(torch.nn.Module):\n    def __init__(self):\n        super(AngleModel, self).__init__()\n        self.angle = nn.Parameter(Variable(torch.Tensor([0.0]), requires_grad=True))\n        self.qw = torch.cos(self.angle \/ 2.)\n        self.qx = 0.0\n        self.qy = 0.0\n        self.qz = torch.sin(self.angle \/ 2.)\n\n    def forward(self, input):\n        matrix_np = transform_from_pq([0, 0, 0, self.qw, self.qx, self.qy, self.qz])\n        matrix = torch.from_numpy(matrix_np)\n        input_ext = torch.cat((input, torch.ones(input.size(0)).reshape(-1, 1)), 1)\n        output = torch.matmul(input_ext.float(), matrix.float())\n        return output[:, :3]\n<\/code><\/pre>\n\n<p>However using this in an optimizer fails because <code>model.parameters()<\/code> returns an empty list (i.e. \"empty\" generator). <\/p>\n\n<p>I'm completely new to pytorch. What am I doing wrong? <\/p>\n\n<h1>Relevant code<\/h1>\n\n<p>Initialization: <\/p>\n\n<pre><code>def _init_model(self):\n    self.model = AngleModel()\n    self.crit = torch.nn.MSELoss()\n    l_rate = 0.01\n    print(list(self.model.parameters()))\n    self.optim = torch.optim.SGD(self.model.parameters(), lr = l_rate)\n    self.epochs = 2000\n<\/code><\/pre>\n\n<p>Learning function: <\/p>\n\n<pre><code>for epoch in range(self.epochs):\n    inputs = torch.from_numpy(normalize(input))\n    labels = torch.from_numpy(normalize(target))\n    _x = Variable(inputs, requires_grad = True) # without it complains about no grad_fn\n    _y = Variable(labels)\n    self.optim.zero_grad()\n\n    outputs = self.model.forward(_x)\n    loss = self.crit(outputs, _y)\n    loss.backward()\n    self.optim.step()\n    print(\"Epoch %6d loss %05.3f; %s\" % (epoch, loss.data[0], self.model.angle))\n<\/code><\/pre>\n\n<h1>Update 1:<\/h1>\n\n<ul>\n<li>Code of the Module updated<\/li>\n<li>Surrounding code added<\/li>\n<li><strong>This at least does not throw any errors anymore but also does not optimize <code>angle<\/code><\/strong><\/li>\n<\/ul>\n","answers":[{"AnswerId":"52736086","CreationDate":"2018-10-10T08:42:40.557","ParentId":null,"OwnerUserId":"205147","Title":null,"Body":"<p>Here are the necessary changes to the AngleModel class to make it work:<\/p>\n\n<pre><code>class AngleModel(nn.Module):\n    def __init__(self):\n        super(AngleModel, self).__init__()\n        self.angle = nn.Parameter(torch.tensor(0.0))\n\n    def forward(self, input):\n        qw = torch.cos(self.angle \/ 2.)\n        qx = 0.0\n        qy = 0.0\n        qz = torch.sin(self.angle \/ 2.)\n\n        matrix = torch.zeros(3, 3)\n\n        matrix[0, 0] = 1. - 2. * qy ** 2 - 2. * qz ** 2\n        matrix[1, 1] = 1. - 2. * qx ** 2 - 2. * qz ** 2\n        matrix[2, 2] = 1. - 2. * qx ** 2 - 2. * qy ** 2\n\n        matrix[0, 1] = 2. * qx * qy - 2. * qz * qw\n        matrix[1, 0] = 2. * qx * qy + 2. * qz * qw\n\n        matrix[0, 2] = 2. * qx * qz + 2 * qy * qw\n        matrix[2, 0] = 2. * qx * qz - 2 * qy * qw\n\n        matrix[1, 2] = 2. * qy * qz - 2. * qx * qw\n        matrix[2, 1] = 2. * qy * qz + 2. * qx * qw\n\n        output = torch.matmul(input, matrix)\n        return output\n<\/code><\/pre>\n\n<p>Among others the main reason that it hadn't been working before was that the graph wasn't consistent due to the intermediate use of numpy. Furthermore the quaternion components qw..qz had been calculated only once in the <code>__init__<\/code> method. Their calculation had to move to the forward step too. <\/p>\n"}]}
+{"QuestionId":52698468,"AnswerCount":1,"Tags":"<python><numpy><keras><lstm>","CreationDate":"2018-10-08T08:43:02.500","AcceptedAnswerId":52736058.0,"OwnerUserId":8113093.0,"Title":"Error when feeding numpy sequence to bidirectional LSTM in Keras","Body":"<p>I am trying to feed the features extracted from 2 fine-tuned VGG16 (each on a different stream), then for each sequence of 9 data pairs, concatenate their numpy arrays and feed the sequence of 9 outputs (concatenated) to a bi-directional LSTM in Keras. <\/p>\n\n<p>The problem is that I am running into an error when trying to build the LSTM part. The following shows the generator I wrote to read both RGB and Optical flow streams, extract features and concatenate each pair :<\/p>\n\n<pre><code>def generate_generator_multiple(generator,dir1, dir2, batch_rgb, batch_flow, img_height,img_width):\n\n    print(\"Processing inside generate multiple\")\n\n    genX1 = generator.flow_from_directory(dir1,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_rgb,\n                                          shuffle=False \n                                          )\n\n    genX2 = generator.flow_from_directory(dir2,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_flow,\n                                          shuffle=False    \n                                          )\n\n    while True:\n\n            imgs, labels = next(genX1)\n            X1i = RGB_model.predict(imgs, verbose=0)\n\n            imgs2, labels2 = next(genX2)            \n            X2i = FLOW_model.predict(imgs2,verbose=0)\n\n            Xi = []\n            for i in range(9):\n                Xi.append(np.concatenate([X1i[i+1],X2i[i]]))\n\n            Xi = np.asarray(Xi)\n\n            if np.array_equal(labels[1:],labels2)==False:\n                print(\"ERROR !! problem of labels matching: RGB and FLOW have different labels\")\n\n            yield Xi, labels2[2]\n<\/code><\/pre>\n\n<p>I am expecting the generator to yield a sequence of 9 arrays, so the shape of Xi when I force the loop to run twice is: (9, 14, 7, 512) <\/p>\n\n<p>When I use while True (like in the code above) and try to call the method to check what it returs, after 3 iterations I get the error:<\/p>\n\n<pre><code>ValueError: too many values to unpack (expected 2)\n<\/code><\/pre>\n\n<p>Now, assuming that there is no problem with the generator, I try to feed the data returned by the generator to the bidirectional LSTM like the following:<\/p>\n\n<pre><code>n_frames = 9\nseq = 100\n\nBi_LSTM = Sequential()\nBi_LSTM.add(Bidirectional(LSTM(seq, return_sequences=True, dropout=0.25, recurrent_dropout=0.1),input_shape=(n_frames,14,7,512)))\nBi_LSTM.add(GlobalMaxPool1D())\nBi_LSTM.add(TimeDistributed(Dense(100, activation=\"relu\")))\nBi_LSTM.add(layers.Dropout(0.25))\nBi_LSTM.add(Dense(4, activation=\"relu\"))\n\nmodel.compile(Adam(lr=.00001), loss='categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>But I keep getting the following error: (the error log is a bit long)<\/p>\n\n<pre><code>InvalidArgumentError: Shape must be rank 4 but is rank 2 for 'bidirectional_2\/Tile_1' (op: 'Tile') with input shapes: [?,7,512,1], [2].\n<\/code><\/pre>\n\n<p>It seems to be caused by this line:<\/p>\n\n<pre><code>Bi_LSTM.add(Bidirectional(LSTM(seq, return_sequences=True, dropout=0.25, recurrent_dropout=0.1),input_shape=(n_frames,14,7,512)))\n<\/code><\/pre>\n\n<p>I am not sure anymore if the problem is the way I try to build the LSTM, the way I return the data from the generator, or the way I define the input of LSTM.<\/p>\n\n<p>Thanks a lot for any help you can provide.<\/p>\n","answers":[{"AnswerId":"52736058","CreationDate":"2018-10-10T08:41:15.543","ParentId":null,"OwnerUserId":"8113093","Title":null,"Body":"<p>It seems like this error specifically is cause by the following line:<\/p>\n\n<pre><code>input_shape=(n_frames,14,7,512)\n<\/code><\/pre>\n\n<p>I was confused about the input for LSTM. Instead to explicitly giving the shape of the input, we just need to specify the dimensions of the input. In my case, this is 3 since the input is a 3D np array. I still have other problems with my code, but for this specific error, the solution is changing that part with:<\/p>\n\n<pre><code>input_shape=(n_frames,3) \n<\/code><\/pre>\n\n<p><strong>Edit:<\/strong>\nWhen predicting, We need to get the mean of the prediction since LSTM expects a 1D input.<\/p>\n\n<p>Another issue in my code was the shape of Xi. It needs to be reshaped before yielding it so that it matches the input expected by LSTM.<\/p>\n"}]}
+{"QuestionId":52698563,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2018-10-08T08:48:47.867","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"Got error: Output tensors to a Model must be the output of a Keras `Layer`","Body":"<pre><code>for i in range(seq.size(1) - 1):&lt;br&gt;\n    it = seq[:, i].clone()&lt;br&gt;  \n    # break if all the sequences end &lt;br&gt;\n    if i &gt;= 1 and seq[:, i].sum() == 0:&lt;br&gt;\n        break &lt;br&gt;\n    output, state = self.get_logprobs_state(it, fc_feats, att_feats, &lt;br&gt;      \n                    pp_att_feats, att_masks, state) &lt;br&gt;\n    outputs[:, i].assign(output) &lt;br&gt;\noutputs = tf.convert_to_tensor(outputs) &lt;br&gt;\nmodel = Model([inputs_fc, inputs_att, inputs_seq], outputs) &lt;br&gt;\n<\/code><\/pre>\n\n<p>I tried to use Lambda, also not work.\nHere is the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Output tensors to a Model must be the output of a Keras\n  <code>Layer<\/code> (thus holding past layer metadata). Found:\n  Tensor(\"Variable_1\/read:0\", shape=(10, 15, 9488), dtype=float32)<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52698670,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-08T08:55:54.617","AcceptedAnswerId":52703339.0,"OwnerUserId":10386005.0,"Title":"How add new channels in keras?","Body":"<pre><code>from keras.layers import Conv2D, Input\n\n# input tensor for a 3-channel 256x256 image\nx = Input(shape=(256, 256, 3))\n# 3x3 conv with 2 output channels (same as input channels)\ny = Conv2D(2, (3, 3), padding='same')(x)\nz = keras.layers.add([x, y])\n<\/code><\/pre>\n\n<p>I need to create a neural network that not only classifies images, but also counts objects inside it.So, I have to create two channels, is the code above right? How can I train and predict in this case?<\/p>\n","answers":[{"AnswerId":"52703339","CreationDate":"2018-10-08T13:23:54.943","ParentId":null,"OwnerUserId":"8858832","Title":null,"Body":"<p>After a bit of after thought it seems to me that what you are trying to achieve is not the main purpose of keras , and while it is entirely possible to do this in keras , creating the custom loss function and maintaining its flow goes out of scope of keras , I recommend that you switch to tensorflow entirely and if needed use the tf.layers for simplicity , I have also created the custom loss function , though i recommend that you try changing it to suite your needs better.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx=np.random.randn(5,5,3)\nx=x.reshape(1,*x.shape)\ny_val=np.random.randn(1,3)\n\ninputv=tf.placeholder(tf.float32,shape=(None,5,5,3))\ny=tf.placeholder(tf.float32,shape=(None,3))\n\nw=tf.Variable(tf.truncated_normal([1,3,3,1],stddev=0.1),name=\"W\")\nb=tf.Variable(tf.constant(0.1,shape=[3]),name=\"B\")\nconv=tf.nn.conv2d(inputv,w,strides=[1,1,1,1],padding=\"SAME\")\nact=tf.nn.relu(conv+b)\n\nflat=tf.contrib.layers.flatten(act)\nw1=tf.Variable(tf.truncated_normal([75,3],stddev=0.1),name=\"W\")\nb1=tf.Variable(tf.constant(0.1,shape=[3]),name=\"B\")\nlogits=tf.nn.relu(tf.matmul(flat,w1)+b1)\n\nloss=tf.reduce_mean(tf.concat([tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits[0][1:],labels=y[0][1:]),tf.losses.mean_squared_error(y[0][0],logits[0][0])],0))\ntrain_step=tf.train.AdamOptimizer(1e-4).minimize(loss)\nsess=tf.Session()\nsess.run(tf.global_variables_initializer())\nval=sess.run((logits,loss,train_step),feed_dict={inputv:x,y:y_val})\nsess.close()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52699163,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T09:25:30.803","AcceptedAnswerId":null,"OwnerUserId":8568787.0,"Title":"Convert conv2d into tf.nn.separable_conv2d running into memory error","Body":"<p>I am currently trying to make my network a little bit faster by using separable convolutions instead of regular ones. One remark dont get confused by the additional code snippets. They are put there to emphesise the problem. My guess is I build the depthwise and pointwise conv the wrong way but I have no clue how I can use it otherwise. THanks in advance!<\/p>\n\n<p>I used the code below with my regular convolutions as an input I give a (1,800,800,3) input and use a 3x3 kernel and 32 filters. Both methods have the same syntax<\/p>\n\n<pre><code>def conv(input, kernel_size, filters, stride, is_training, batchnorm=True, activation=tf.nn.relu, name=\"conv\"):\n<\/code><\/pre>\n\n<p>afterwards I define my variables w and b and batch norm.<\/p>\n\n<pre><code>with tf.variable_scope(name):\n    weights = tf.get_variable('weights',\n                              shape=[kernel_size, kernel_size, input.get_shape()[3], filters],\n                              initializer=tf.contrib.layers.xavier_initializer()\n                              )\n    biases = tf.get_variable('biases',\n                             shape=[filters],\n                             initializer=tf.constant_initializer(0.0)\n                             )\n    conv = tf.nn.conv2d(input, weights, strides=[1, stride, stride, 1], padding='SAME')\n    output = tf.nn.bias_add(conv, biases)\n\n    if batchnorm:\n        output = tf.layers.batch_normalization(output, training=is_training, center=True, scale=False)\n\n    if activation is not None:\n        output = activation(output)\n    return output\n<\/code><\/pre>\n\n<p>Until here everything works as expected and when I run my code i can train my network.<\/p>\n\n<p>Now I tried to derive a separate convolution using the tf.nn.separable_conv2d<\/p>\n\n<p>Therefore I used<\/p>\n\n<pre><code>with tf.variable_scope(name):\n        weights = tf.get_variable('weights',\n                                  shape=[kernel_size, kernel_size, input.get_shape()[3], filters],\n                                  initializer=tf.contrib.layers.xavier_initializer())\n\n        pointwise_filter = tf.get_variable(\"pointwise_filter\",shape = [1,1,input.get_shape()[3]*filters, filters],\n                                    initializer = tf.contrib.layers.xavier_initializer())\n\n        biases = tf.get_variable('biases',\n                                 shape=[filters],\n                                 initializer=tf.constant_initializer(0.0)\n                                 )\n        conv = tf.nn.separable_conv2d(input, weights, pointwise_filter, strides=[1, stride, stride, 1], padding='SAME')\n        output = tf.nn.bias_add(conv, biases)\n\n        if batchnorm:\n            output = tf.layers.batch_normalization(output, training=is_training, center=True, scale=False)\n\n        if activation is not None:\n            output = activation(output)\n        return output\n<\/code><\/pre>\n\n<p>The network is building fine and then I run into memory issues with my Gtx1080ti.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xxpYk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xxpYk.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":52699598,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><time-series>","CreationDate":"2018-10-08T09:48:34.833","AcceptedAnswerId":52701095.0,"OwnerUserId":5585432.0,"Title":"Keras TimeDistributed with multiple Inputs in different shapes","Body":"<p>I've got a pretrained model with multiple inputs which have different shapes. So I can call the model on new inputs that have a matching shape like this:<\/p>\n\n<pre><code>new_output = model([input_1, input2])\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>input_1.shape = (400, 200)\ninput_2.shape = (400, 200, 10)\n<\/code><\/pre>\n\n<p>I want to reuse the model to train it on a data-series, so my new inputs would have the following shapes:<\/p>\n\n<pre><code>input_1.shape = (100, 400, 200)\ninput_2.shape = (100, 400, 200, 10)\n<\/code><\/pre>\n\n<p>To reuse the model I want to use the <code>TimeDistributed<\/code> layer from Keras like this:<\/p>\n\n<pre><code>output = TimeDistributed(model)([input_1, input_2])\n<\/code><\/pre>\n\n<p>This doesn't work because <code>TimeDistributed<\/code> doesn't accept lists as input. All workarounds that I found so far are using merged inputs to solve this problem, but this solution doesn't work in my case because the input shapes are different. <\/p>\n\n<p>I also tried to use some <code>Lambda<\/code> layers like this:<\/p>\n\n<pre><code>lambda_0 = Lambda(lambda x: x)\nlambda_1 = Lambda(lambda x: [TimeDistributed(lambda_0)(x[0]), TimeDistributed(lambda_0)(x[1])])([input_1, input_2])\noutput = model(lambda_1)\n<\/code><\/pre>\n\n<p>But this isn't working because it just uses the <code>TimeDistributed<\/code> on the <code>lambda_0<\/code> but not on the model. Maybe there is a solution with <code>Lambda<\/code> wrappers for this problem or maybe I have to define a custom Keras layer but at the moment I'm stuck with this problem so I would appreciate any help.<\/p>\n","answers":[{"AnswerId":"52701095","CreationDate":"2018-10-08T11:16:37.373","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can easily reshape <code>(?, 400, 200)<\/code> to <code>(?, 400, 200, 1)<\/code> and then concatenate the two inputs to have a Tensor of shape <code>(?, 400, 200, 11)<\/code>, then feed this tensor to a <code>Lambda<\/code> layer which is wrapped in a <code>TimeDistributed<\/code> layer and therefore applies your model on each timestep:<\/p>\n\n<pre><code>in1 = Input((400, 200))\nin2 = Input((400, 200, 10))\n\nr_in1 = Reshape((400, 200, 1))(in1) # you can also use `K.expand_dims()` in a Lambda layer\nconcat = concatenate([r_in1, in2])\n\nout_model = TimeDistributed(Lambda(lambda x: model([x[:,:,0], x[:,:,1:]])))(concat)\n\nnew_model = Model([in1, in2], [out_model])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52699654,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><object-detection-api>","CreationDate":"2018-10-08T09:51:24.467","AcceptedAnswerId":null,"OwnerUserId":10377244.0,"Title":"Object detection api training issue","Body":"<p>System information<\/p>\n\n<pre><code>OS WIN 10 Processor : i5-4440 3.10GHz RAM 8\n\nTensorFlow installed from (source or binary): pip install TensorFlow\nversion (use command below): b'v1.10.0-rc1-19-g656e7a2b34' 1.10.0\nPython version: 3.6.6\n<\/code><\/pre>\n\n<p>Describe the problem<\/p>\n\n<p>I am trying to train object detection api using <code>faster_rcnn_inception_v2_coco_2018_01_28<\/code>.<\/p>\n\n<p>I am facing with issue with object detection training. After ~400 steps the training stops because of a path problem while the script create the path:<\/p>\n\n<pre><code>path - C:\\Users\\z\\Anaconda3\\envs\\tensorflow\\Lib\\site-packages\\tensorflow\\models\\research\\object_detection\\training_dataturk\\models\\model\\export\\Servo\\temp-b'1538897129'\n<\/code><\/pre>\n\n<p>----the error screen shot that happened twice before stopping\n<a href=\"https:\/\/i.stack.imgur.com\/jum9p.png\" rel=\"nofollow noreferrer\">screen shot<\/a><\/p>\n","answers":[{"AnswerId":"56989914","CreationDate":"2019-07-11T13:03:16.957","ParentId":null,"OwnerUserId":"4663377","Title":null,"Body":"<p>After ~400 steps you are saving your checkpoint to a particular location make sure the path you mentioned is correct: <\/p>\n\n<p>You can try<\/p>\n\n<p><code>C:\\\\Users\\\\z\\\\Anaconda3\\\\envs\\\\tensorflow\\\\Lib\\\\site-packages\\\\tensorflow\\\\models\\\\research\\\\object_detection\\\\training_dataturk\\\\models\\\\model\\\\export\\\\Servo<\/code><\/p>\n"}]}
+{"QuestionId":52700417,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><python-imaging-library>","CreationDate":"2018-10-08T10:35:26.493","AcceptedAnswerId":52702710.0,"OwnerUserId":7826511.0,"Title":"Convert red line in image to numpy list","Body":"<p>I want to measure the height difference to ground of the detected object in this pipe. The lower red line shall be the marker for minimum height. I thought I might first convert the lower red line into a numpy list but how would I do that? The red circles are drawn with <code>cv2.circle()<\/code> function.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HK3WM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HK3WM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>Thanks to ZdaR I am closer to solving my problem. This is his solution rewritten to work with python3:<\/p>\n\n<pre><code>import cv2\nimport numpy as np\n\n\ndef get_center(arr):\n    return sum(arr)\/len(arr)\n\n\ndef get_cluster_centers(arr, tolerance):\n    clusters = [[arr[0]]]\n\n    for ele in arr[1:]:\n        if abs(clusters[-1][0] - ele) &lt; tolerance:\n            clusters[-1].append(ele)\n        else:\n            clusters.append([ele])\n\n    clusters_centers = map(get_center, clusters)\n    return clusters_centers\n\n\nimg = cv2.imread(\"\/home\/artur\/Desktop\/0.png\")\n\n# Segment the red color\nmask = cv2.inRange(img, np.array([0, 0, 255]), np.array([0, 0, 255]))\n\nfor i in mask:\n    print(i)\n\n# Let's iterate the middle column and get the distance between the two red lines.\nhalf_width = int(mask.shape[1]\/2)\nmiddle_column = mask[:, half_width]\n\nidx = np.where(middle_column == 255)\n\n# Since the width of line is not 1 px so we need to cluster the pixels to get a single center value.\ncenters = list(get_cluster_centers(idx[0], 5))\n\nif len(centers) == 2:\n    print(\"Distance between lines:\", centers[1] - centers[0], \"px\")\n<\/code><\/pre>\n\n<p>It measures the pixel distance between upper and lower red line with the help of the middle column of the image. How would I iterate over all columns to determine the minimum distance between those two line or better, between detected object and lower red line? Do I get it right that this solution only takes the middle column into account?<\/p>\n","answers":[{"AnswerId":"52702710","CreationDate":"2018-10-08T12:48:40.217","ParentId":null,"OwnerUserId":"3051961","Title":null,"Body":"<p>You can first segment the red color from the input image to get a binary mask then assuming that your red line is centered to the input image, we take the center column of that image and iterate over the column to find the red dot locations then simply find the distance in pixels as:<\/p>\n\n<pre><code>import cv2\nimport numpy as np\n\n\ndef get_center(arr):\n    return sum(arr)\/len(arr)\n\n\ndef get_cluster_centers(arr, tolerance):\n    clusters = [[arr[0]]]\n\n    for ele in arr[1:]:\n        if abs(clusters[-1][0] - ele) &lt; tolerance:\n            clusters[-1].append(ele)\n        else:\n            clusters.append([ele])\n\n    clusters_centers = map(get_center, clusters)\n    return clusters_centers\n\n\nimg = cv2.imread(\"\/home\/anmol\/Downloads\/HK3WM.png\")\n\n# Segment the red color\nmask = cv2.inRange(img, np.array([0, 0, 255]), np.array([0, 0, 255]))\n\n# Let's iterate the middle column and get the distance between the two red lines.\nhalf_width = mask.shape[1]\/2\nmiddle_column = mask[:, half_width]\n\nidx = np.where(middle_column == 255)\n\n# Since the width of line is not 1 px so we need to cluster the pixels to get a single center value.\ncenters = get_cluster_centers(idx[0], 5)\n\nif len(centers) == 2:\n    print \"Distance between lines:\", centers[1] - centers[0], \"px\"\n<\/code><\/pre>\n\n<p>PS: I am kind on in hurry if this does not explain something feel free to ask in comments.<\/p>\n"}]}
+{"QuestionId":52700483,"AnswerCount":0,"Tags":"<amazon-web-services><tensorflow><gpu>","CreationDate":"2018-10-08T10:39:24.893","AcceptedAnswerId":null,"OwnerUserId":9344209.0,"Title":"It is possible to run Tensorflow with elastic GPU in AWS?","Body":"<p>It is possible to run Tensorflow with elastic GPU (not CPU) in AWS?\nIf yes how?<\/p>\n","answers":[]}
+{"QuestionId":52700621,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T10:47:51.490","AcceptedAnswerId":53055899.0,"OwnerUserId":1692060.0,"Title":"Tensorflow server: I don't want to initialize global variables for every session","Body":"<p>EDIT2: Github link below contains possible solutions to the problem of calling TF model from process. They include eager execution and dedicated server process, serving TF model predictions via http requests. I wonder if with custom server and requests I win any time compared to initializing global variables each time and calling <code>tf.train.Server<\/code>, but it seems to be more elegant way.<\/p>\n\n<p>I will investigate memory leak, and if it is gone, close this question. <\/p>\n\n<p>EDIT: Added simple reproducible example of the problem:<\/p>\n\n<p><a href=\"https:\/\/github.com\/hcl14\/Tensorflow-server-launched-from-child-process\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hcl14\/Tensorflow-server-launched-from-child-process<\/a><\/p>\n\n<hr>\n\n<p><strong>Background<\/strong>: I am running Tensorflow server, and connecting to it from 'forked' processes. Creating (and destroying) processes dynamically is essential for me - I moved highly loaded part of code there because of a <a href=\"http:\/\/www.paulsprogrammingnotes.com\/2014\/10\/large-dictionaries-not-released-from.html\" rel=\"nofollow noreferrer\">weird memory leak<\/a>, not visible to Python profilers (threads do not solve the issue). Therefore, I want processes to be initialized fast and immediately start working. Memory is freed only when process is destroyed.<\/p>\n\n<p>Doing experiments, I found a solution when loaded model and graph are saved into global variable, then taken by child process (which uses 'fork' mode by default) and then server is called.<\/p>\n\n<p><strong>Problem<\/strong>: The strange thing for me is that, after loading keras models, I cannot lock graph which I do not expect to modify, and I need to run <code>tf.global_variables_initializer()<\/code> each time I open new session in child process. However, dummy run in the main flow without any session creation works Ok. I know that in this case tensorflow uses default session, but all the variables on a graph should be initialized after model run, so I expected new session to work Ok with previously defined graph.<\/p>\n\n<p>Thus, I think that modifying model makes Python to pickle a lot to the child process ('fork' mode), which creates computational and memory overhead.<\/p>\n\n<p>Please, excuse me for a lot of code. The model I use is legacy and black box for me, so it is possible that my problem is related to it. <strong>Tensorflow version is 1.2<\/strong> (I cannot upgrade it, model is not compatible), <strong>Python 3.6.5<\/strong>.<\/p>\n\n<p>Also, maybe my solution is inefficient and there is better one, I would be grateful for your advice.<\/p>\n\n<p>My setup is the following:<\/p>\n\n<p>1.Tensorflow server started in main process:<\/p>\n\n<p>Initialize the server:<\/p>\n\n<pre><code>def start_tf_server():\n    import tensorflow as tf\n    cluster = tf.train.ClusterSpec({\"local\": [tf_server_address]})\n    server = tf.train.Server(cluster, job_name=\"local\", task_index=0)    \n    server.join() # block process from exiting\n<\/code><\/pre>\n\n<p>In main process:<\/p>\n\n<pre><code>p = multiprocessing.Process(target=start_tf_server)\np.daemon=True\np.start() # this process never ends, unless tf server crashes\n\n# WARNING! Graph initialization must be made only after Tf server start!\n# Otherwise everything will hang\n# I suppose this is because of another session will be \n# created before the server one\n\n# init model graph before branching processes\n# share graph in the current process scope\ninterests = init_interests_for_process()\nglobal_vars.multiprocess_globals[\"interests\"] = interests\n<\/code><\/pre>\n\n<p>2.<code>init_interests_for_process()<\/code> is a model initializer, which loads my legacy model and shares it in the global variable. I do one dummy model pass to have everything initialized on the graph, and then want to lock the graph. But it is not working:<\/p>\n\n<pre><code>def init_interests_for_process():\n    # Prevent errors on my GPU and disable tensorflow \n    # complaining about CPU instructions\n    import os\n    os.environ[\"CUDA_VISIBLE_DEVICES\"]= \"\"\n    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n\n    import tensorflow as tf\n\n    from tensorflow.contrib.keras import models\n\n    # create tensorflow graph\n    graph = tf.get_default_graph()\n\n    with graph.as_default():\n\n        TOKENIZER = joblib.load(TOKENIZER_FILE)\n\n        NN1_MODEL = models.load_model(NN1_MODEL_FILE)\n\n        with open(NN1_CATEGORY_NAMES_FILE, 'r') as f:\n            NN1_CATEGORY_NAMES = f.read().splitlines()\n\n        NN2_MODEL = models.load_model(NN2_MODEL_FILE)\n\n        with open(NN2_CATEGORY_NAMES_FILE, 'r') as f:\n            NN2_CATEGORY_NAMES = f.read().splitlines()\n        # global variable with all the data to be shared\n        interests = {}\n\n        interests[\"TOKENIZER\"] = TOKENIZER\n        interests[\"NN1_MODEL\"] = NN1_MODEL\n        interests[\"NN1_CATEGORY_NAMES\"] = NN1_CATEGORY_NAMES\n        interests[\"NN2_MODEL\"] = NN2_MODEL\n        interests[\"NN2_CATEGORY_NAMES\"] = NN2_CATEGORY_NAMES\n        interests['all_category_names'] = NN1_CATEGORY_NAMES + \\\n                                          NN2_CATEGORY_NAMES\n        # Reconstruct a Python object from a file persisted with joblib.dump.\n        interests[\"INTEREST_SETTINGS\"] = joblib.load(INTEREST_SETTINGS_FILE)\n\n        # dummy run to create graph\n        x = tf.contrib.keras.preprocessing.sequence.pad_sequences(\n                         TOKENIZER.texts_to_sequences(\"Dummy srting\"),\n                         maxlen=interests[\"INTEREST_SETTINGS\"][\"INPUT_LENGTH\"]\n                         )\n        y1 = NN1_MODEL.predict(x)\n        y2 = NN2_MODEL.predict(x)\n\n        # PROBLEM: I want, but cannot lock graph, as child process \n        # wants to run its own tf.global_variables_initializer()\n        # graph.finalize()\n\n        interests[\"GRAPH\"] = graph\n\n        return interests\n<\/code><\/pre>\n\n<p>3.Now I spawn the process (actually, the process is spawned from another process - hierarchy is complicated):<\/p>\n\n<pre><code>def foo(q):\n     result = call_function_which_uses_interests_model(some_data) \n     q.put(result)\n     return # I've read it is essential for destroying local variables\nq = Queue()\np = Process(target=foo,args=(q,))\np.start()\np.join()\nresult = q.get() # retrieve data\n<\/code><\/pre>\n\n<p>4.And inside this process I am calling the model:<\/p>\n\n<pre><code># retrieve model from global variable\ninterests = global_vars.multiprocess_globals[\"interests\"]\n\ntokenizer = interests[\"TOKENIZER\"]\nnn1_model = interests[\"NN1_MODEL\"]\nnn1_category_names = interests[\"NN1_CATEGORY_NAMES\"]\nnn2_model = interests[\"NN2_MODEL\"]\nnn2_category_names = interests[\"NN2_CATEGORY_NAMES\"]\ninput_length = interests[\"INTEREST_SETTINGS\"][\"INPUT_LENGTH\"]\n\n# retrieve graph\ngraph = interests[\"GRAPH\"]\n\n# open session for server\nlogger.debug('Trying tf server at ' + 'grpc:\/\/'+tf_server_address)\nsess = tf.Session('grpc:\/\/'+tf_server_address, graph=graph)\n\n# PROBLEM: and I need to run variables initializer:\nsess.run(tf.global_variables_initializer())\n\n\ntf.contrib.keras.backend.set_session(sess)\n\n# finally, make a call to server:\nwith sess.as_default():        \n    x = tf.contrib.keras.preprocessing.sequence.pad_sequences(\n                            tokenizer.texts_to_sequences(input_str),\n                            maxlen=input_length)\n    y1 = nn1_model.predict(x)\n    y2 = nn2_model.predict(x)\n<\/code><\/pre>\n\n<p>Everything works Ok if I don't lock the graph and run variable initializer each time new process is spawned. (Except, there is a memory leak of about 30-90 MB for each call, not visible to python memory profilers). When I want to lock the graph, I get errors about uninitialized variables:<\/p>\n\n<pre><code>FailedPreconditionError (see above for traceback): \nAttempting to use uninitialized value gru_1\/bias\n       [[Node: gru_1\/bias\/read = Identity[T=DT_FLOAT, _class=[\"loc:@gru_1\/bias\"],\n       _device=\"\/job:local\/replica:0\/task:0\/cpu:0\"](gru_1\/bias)]]\n<\/code><\/pre>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"52956563","CreationDate":"2018-10-23T19:30:52.593","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>I am not sure if this could help you, but one needs to know that in tensorflow, variables are only initialized for a given <code>Session<\/code>. A variable needs to be initialized in each <code>Session<\/code> it is used -- even in the simpliest possible scenario:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.Variable(0.)\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    # x is initialized -- no issue here\n    x.eval()\n\nwith tf.Session() as sess:\n    x.eval()\n    # Error -- x was never initialized in this session, even though\n    # it has been initialized before in another session\n<\/code><\/pre>\n"},{"AnswerId":"53055899","CreationDate":"2018-10-30T00:53:26.607","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Have you considered TensorFlow Serving? <a href=\"https:\/\/www.tensorflow.org\/serving\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/serving\/<\/a><\/p>\n\n<p>Generally you'd want to cache Sessions, which I believe is the strategy TF Serving uses. That will be by far the best experience for deploying a TF model into a datacenter.<\/p>\n\n<p>You could also go the other direction and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/enable_eager_execution\" rel=\"nofollow noreferrer\"><code>tf.enable_eager_execution()<\/code><\/a>, which eliminates the need for Sessions. Variables still get initialized, although it happens as soon as the Python variable objects are created.<\/p>\n\n<p>But if you really want to create and destroy Sessions, you could replace variables in the graph with constants (<a href=\"https:\/\/www.tensorflow.org\/extend\/tool_developers\/#freezing\" rel=\"nofollow noreferrer\">\"freeze\" it<\/a>). I'd also consider disabling graph optimizations in this case, as the first <code>session.run<\/code> call with a new set of feeds and fetches will by default spend some time optimizing the graph (configured through a <code>RewriterConfig<\/code> inside a <code>GraphOptions<\/code> proto).<\/p>\n\n<p>(Expanded from a comment on the question)<\/p>\n"}]}
+{"QuestionId":52701315,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-08T11:31:03.197","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Unable to Save Keras checkpoints on CloudML (1.8) error: ImportError: `save_model` requires h5py","Body":"<p>After each epoch, I have the following callbacks:<\/p>\n\n<ol>\n<li>Create a TensorBoard.<\/li>\n<li>Save the model checkpoint.<\/li>\n<\/ol>\n\n<p>However, after the 1st epoch of training I get the following traceback. I am assuming this is to do with the checkpoint callback.<\/p>\n\n<p>Is this normal behaiviour? <\/p>\n\n<p>My <strong>callbacks.py<\/strong> where all callbacks are created in <code>create_callbacks()<\/code><\/p>\n\n<pre><code>def create_callbacks(job_dir, logs_path):\n\n    checkpoint_path = 'checkpoint.{epoch:04d}-{val_loss:.9f}.hdf5'\n\n    if not job_dir.startswith(\"gs:\/\/\"):  # then local\n        checkpoint_path = os.path.join(job_dir + 'checkpoints', checkpoint_path)\n\n    checkpoint = tf.keras.callbacks.ModelCheckpoint(checkpoint_path, monitor='val_loss', verbose=0, save_best_only=True,\n                                 save_weights_only=False,\n                                 mode='auto', period=1)\n\n    tb = tf.keras.callbacks.TensorBoard(log_dir=logs_path, batch_size=None, histogram_freq=0, write_graph=False)\n\n    # Continuous eval callback\n    export = ContinuousExport(eval_frequency=1, job_dir=job_dir)\n\n    return [checkpoint, tb, export]\n\n\nclass ContinuousExport(tf.keras.callbacks.Callback):\n    \"\"\"Continuous eval callback to evaluate the checkpoint once every so many epochs.\"\"\"\n\n    def __init__(self, eval_frequency, job_dir,):\n        self.eval_frequency = eval_frequency\n        self.job_dir = job_dir\n\n    def on_epoch_end(self, epoch, logs={}):\n        print('Epoch number is {}'.format(epoch))\n        print('Frequency is {}'.format(self.eval_frequency))\n        if epoch &gt; 0 and epoch % self.eval_frequency == 0:\n            # Unhappy hack to work around h5py not being able to write to GCS.\n            # Force snapshots and saves to local filesystem, then copy them over to GCS.\n            model_path_glob = 'checkpoint.*'\n            if not self.job_dir.startswith(\"gs:\/\/\"):\n                model_path_glob = os.path.join(self.job_dir + 'checkpoints', model_path_glob)\n            checkpoints = sorted(glob.glob(model_path_glob), key=os.path.getmtime)\n            print('Path is {}'.format(model_path_glob))\n            print('Length of cp is {}'.format(len(checkpoints)))\n            if len(checkpoints) &gt; 0:\n                print(checkpoints[-1])\n                if self.job_dir.startswith(\"gs:\/\/\"):\n                    print('Copying the model to {}'.format(self.job_dir + '\/checkpoints\/'))\n                    copy_file_to_gcs(self.job_dir + '\/checkpoints\/', checkpoints[-1])\n                else:\n                    print('Using local storage, not saving to GCS')\n        else:\n            print('\\nEvaluation epoch[{}] (no checkpoints found)'.format(epoch))\n\n\ndef copy_file_to_gcs(job_dir, file_path):\n    with file_io.FileIO(file_path, mode='rb') as input_f:\n        with file_io.FileIO(os.path.join(job_dir, file_path), mode='w+') as output_f:\n            output_f.write(input_f.read())\n<\/code><\/pre>\n\n<blockquote>\n  <p>INFO    2018-10-08 12:17:30 +0100       master-replica-0<br>\n  Module completed; cleaning up. INFO    2018-10-08 12:17:30 +0100<br>\n  master-replica-0                Clean up finished. ERROR   2018-10-08\n  12:18:23 +0100       service         The replica master 0 exited with\n  a non-zero status of 1.  ERROR   2018-10-08 12:18:23 +0100<br>\n  service         Traceback (most recent call last): ERROR   2018-10-08\n  12:18:23 +0100       service           File\n  \"\/usr\/lib\/python3.5\/runpy.py\", line 184, in _run_module_as_main ERROR \n  2018-10-08 12:18:23 +0100       service             \"<strong>main<\/strong>\",\n  mod_spec) ERROR   2018-10-08 12:18:23 +0100       service<br>\n  File \"\/usr\/lib\/python3.5\/runpy.py\", line 85, in _run_code ERROR<br>\n  2018-10-08 12:18:23 +0100       service             exec(code,\n  run_globals) ERROR   2018-10-08 12:18:23 +0100       service<br>\n  File \"\/root\/.local\/lib\/python3.5\/site-packages\/trainer\/model.py\", line\n  167, in  ERROR   2018-10-08 12:18:23 +0100       service<br>\n  train_model(train_file=train_file, test_file=test_file,\n  job_dir=job_dir, **arguments) ERROR   2018-10-08 12:18:23 +0100<br>\n  service           File\n  \"\/root\/.local\/lib\/python3.5\/site-packages\/trainer\/model.py\", line 59,\n  in train_model ERROR   2018-10-08 12:18:23 +0100       service<br>\n  model = fit_model(model, train_g, test_g, callbacks) ERROR<br>\n  2018-10-08 12:18:23 +0100       service           File\n  \"\/root\/.local\/lib\/python3.5\/site-packages\/trainer\/model.py\", line 124,\n  in fit_model ERROR   2018-10-08 12:18:23 +0100       service<br>\n  model.fit_generator(**params) ERROR   2018-10-08 12:18:23 +0100<br>\n  service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training.py\",\n  line 1598, in fit_generator ERROR   2018-10-08 12:18:23 +0100<br>\n  service             initial_epoch=initial_epoch) ERROR   2018-10-08\n  12:18:23 +0100       service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training_generator.py\",\n  line 231, in fit_generator ERROR   2018-10-08 12:18:23 +0100<br>\n  service             callbacks.on_epoch_end(epoch, epoch_logs) ERROR<br>\n  2018-10-08 12:18:23 +0100       service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/callbacks.py\",\n  line 95, in on_epoch_end ERROR   2018-10-08 12:18:23 +0100<br>\n  service             callback.on_epoch_end(epoch, logs) ERROR<br>\n  2018-10-08 12:18:23 +0100       service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/callbacks.py\",\n  line 468, in on_epoch_end ERROR   2018-10-08 12:18:23 +0100<br>\n  service             self.model.save(filepath, overwrite=True) ERROR<br>\n  2018-10-08 12:18:23 +0100       service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/network.py\",\n  line 1126, in save ERROR   2018-10-08 12:18:23 +0100       service<br>\n  save_model(self, filepath, overwrite, include_optimizer) ERROR<br>\n  2018-10-08 12:18:23 +0100       service           File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/saving.py\",\n  line 75, in save_model ERROR   2018-10-08 12:18:23 +0100       service\n  raise ImportError('<code>save_model<\/code> requires h5py.') ERROR   2018-10-08\n  12:18:23 +0100       service         ImportError: <code>save_model<\/code>\n  requires h5py.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52701406","CreationDate":"2018-10-08T11:36:43.853","ParentId":null,"OwnerUserId":"9179860","Title":null,"Body":"<p>Yes, you need install the package h5py.<\/p>\n\n<p>h5py file is a container to store the trained model. If you don't have installed h5py package, you cannot save models.<\/p>\n\n<p>Pre-built h5py wheels can be installed via pip from PyPI<\/p>\n\n<pre><code>$ pip install h5py\n<\/code><\/pre>\n"}]}
+{"QuestionId":52701713,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-08T11:55:02.283","AcceptedAnswerId":52717490.0,"OwnerUserId":7867665.0,"Title":"No gradients to be propagated for Lambda layer","Body":"<p>I want to add a <code>Lamda<\/code> layer <code>sum = Lambda(lambda x: K.sum(x, axis=1)(inputs)<\/code> but I don't want the gradients to be propagated. I want it to be treated like an <code>Input<\/code> layer. How can I do this in Keras? Something like a <code>no_grad<\/code> feature.<\/p>\n","answers":[{"AnswerId":"52717490","CreationDate":"2018-10-09T09:20:22.460","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>This should work<\/p>\n\n<pre><code>sum = Lambda(lambda x: K.sum(x, axis=1)(inputs)\nsum_stop_grad = Lambda(lambda x: K.stop_gradient(x))(sum)\ndense_1 = Dense(16)(sum_stop_grad)\n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":52702183,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T12:21:26.580","AcceptedAnswerId":null,"OwnerUserId":10472944.0,"Title":"Difference of tf.Variable() and tf.get_variable()","Body":"<p>I have following lines of code in Tensorflow:<\/p>\n\n<pre><code>y = tf.Variable(name='y', initial_value=[1.0])\n\ny = tf.get_variable(name='y', initializer=[1.0]\n\nwith tf.variable_scope(tf.get_variable_scope(), reuse=True):\n    y = tf.get_variable(name='y', initializer=[1.0])\n<\/code><\/pre>\n\n<p>First line creates <code>&lt;tf.Variable 'y:0'&gt;<\/code>. Second line creates <code>&lt;tf.Variable 'y_1:0'&gt;<\/code>. The last lines do not create variable and just reuse <code>&lt;tf.Variable 'y_1:0'&gt;<\/code>. <\/p>\n\n<p>Also, the following lines of code lead to ValueError message:<\/p>\n\n<pre><code>with tf.variable_scope(tf.get_variable_scope(), reuse=True):\n    y = tf.get_variable(name='y_1', initializer=[1.0])\n<\/code><\/pre>\n\n<p>How Tensorflow determines that it should reuse <code>&lt;tf.Variable 'y_1:0'&gt;<\/code> and not <code>&lt;tf.Variable 'y:0'&gt;<\/code>? How does Tensorflow store information about the way the Variable was created? Cause names of created Variables appear to have no information about this.<\/p>\n\n<p>Edit: The main question is how Tensorflow understands in the example above that one should reuse the latter Variable <code>&lt;tf.Variable 'y_1:0'&gt;<\/code> and not the former? I did not specify here variable scope - so maybe there is an implicit variable scope (that is not displayed in name of Variable) for Variables created with <code>tf.get_variable()<\/code>?<\/p>\n","answers":[]}
+{"QuestionId":52702220,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2018-10-08T12:23:10.353","AcceptedAnswerId":52707826.0,"OwnerUserId":10440891.0,"Title":"Access the weight matrix in Tensorflow in order to make apply changes to it in Numpy","Body":"<p>I need a way to have access to the weight matrix in TensorFlow or Keras within each iteration, so that I can convert it into a format that I can use in Numpy to carry out certain operations on it, then send it back to TensorFlow.<\/p>\n\n<p>For example, I want to change my filter such that some of the neurons are specified by others neuron of the filter. They have to be obtained as solutions of linear systems with other neurons as its coefficients, not by the learning process. As I could not find a way to do this in TensorFlow or Keras, I have to use Numpy.<\/p>\n\n<p>I have found many questions with the same or similar titles, but none of them helped. I would appreciate any hints.<\/p>\n\n<p><strong>EDIT<\/strong>\nlet me explain the problem more clearly \nconsider the following code<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.placeholder(tf.float32, (1, 5, 5, 1))\ny = tf.placeholder(tf.float32, (1))\n\n# create variable\nweights = {\n    \"my_filter\": tf.Variable(tf.truncated_normal([3, 3, 1, 1]), name=\"my_filter\"),\n    \"f_c\": tf.Variable(tf.truncated_normal([25,1]), name=\"f_c\") }\n\n\nconv = tf.nn.conv2d(x, weights[\"my_filter\"], [1,1,1,1], padding='SAME')\nflatten= tf.reshape(conv,[1,25])\nlogits= tf.matmul(flatten,weights[\"f_c\"])\n\n\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits,labels= y))\noptmize = tf.train.AdamOptimizer()\ngrads_and_vars = optmize.compute_gradients(cost)\n#In this part before applying gradient I have to apply some complicated mathematical operation\ntrain_op=optmize.apply_gradients(grads_and_vars)\n\n\n\ntrain_epochs=10\ninput_x = np.arange(25).reshape([1,5,5,1])\ninput_y = np.arange(1)\n\nwith tf.Session() as sess:\n\n    sess.run(tf.global_variables_initializer())\n\n    for epoch in range(train_epochs):\n\n        sess.run(train_op, feed_dict={x: input_x, y: input_y})\n<\/code><\/pre>\n\n<p>I have a 5*5 filter named my_filter and I want all its elements to be trained except for one of them, for exaple the (1,1) element, and I need the latter element to be determined by the rest of the elemnets. This has to be done in each iteration. This is exactly where my problem is. I know how to access the weight matrix after the training is finished, but I do not know how to do this within each iteration.<\/p>\n\n<p>In my code, I have first computed the gradients, then made the changes,  and then applied the gradients. But the problem is that the gradients are tuples of types, for exapmle tensor, which are not easy to work with in Numpy. I need some method to convert these data to more familiar Numpy types.<\/p>\n","answers":[{"AnswerId":"52707826","CreationDate":"2018-10-08T18:04:02.767","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Keras layers, and tf.keras.layers layers, support get_weights \/ set_weights methods, which return numpy arrays for the weights. So you can call get_weights, modify the result in numpy, and call set_weights back to put the new numpy values into tensorflow.<\/p>\n\n<p>Something like this:<\/p>\n\n<pre><code>model = tf.keras.Sequential(...)\nfor batch in data:\n  model.fit(batch)\n  if ...:\n    weights_as_numpy = model.get_weights()\n    # modify the weights\n    model.set_weights(weights_as_numpy)\n<\/code><\/pre>\n"},{"AnswerId":"52705993","CreationDate":"2018-10-08T15:54:03.530","ParentId":null,"OwnerUserId":"7691129","Title":null,"Body":"<p>For that, you will need to be able to access to the weights. Instead of defining layer using tf.layers which automatically allocate variable, you can first get a variable yourself, and then call tf.nn instead. <\/p>\n\n<pre><code># input\nx = tf.placeholder(tf.float32, (1, 5, 5, 1))\ndummy_input = np.arange(25).reshape([1,5,5,1])\n\n# create variable \nw = tf.get_variable('weight',[3,3,1,1])\n# assign the variable to layer e.g. conv\ny = tf.nn.conv2d(x, w, [1,1,1,1], padding='SAME')\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    # read the weight\n    random_weight  = sess.run(w, feed_dict={x:dummy_input})\n    print('random weight', random_weight)\n\n    # create some new values for weight\n    new_weight = np.arange(9).reshape([3,3,1,1])\n    # load it into the variable\n    w.load(new_weight,sess)\n\n    # read back and print to verify\n    new_weight  = sess.run(w, feed_dict={x:dummy_input})\n    print('new weight', new_weight)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52702431,"AnswerCount":1,"Tags":"<python><python-2.7><keras><nlp><conv-neural-network>","CreationDate":"2018-10-08T12:34:08.573","AcceptedAnswerId":null,"OwnerUserId":10472174.0,"Title":"Text classification on CNN model","Body":"<p>The problem is i need to load the file which i have in h5 format as below<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('my_model.h5')\nmodel.compile(loss='categorical_crossentropy',optimizer='rmsprop',metrics=['acc'])\nclasses = model.predict_classes(\"How is the weather today\")\nprint classes\n<\/code><\/pre>\n\n<p>And also i need that percentage value of the prediction to be printed<\/p>\n\n<p><a href=\"https:\/\/github.com\/jatana-research\/Text-Classification\/blob\/master\/CNN.ipynb\" rel=\"nofollow noreferrer\">Here is the link that i refered to while generating this model and saving the file<\/a><\/p>\n","answers":[{"AnswerId":"52702931","CreationDate":"2018-10-08T13:00:51.607","ParentId":null,"OwnerUserId":"10472174","Title":null,"Body":"<pre><code>import numpy as np\nimport pandas as pd\nimport pickle\nfrom collections import defaultdict\nimport re\nfrom bs4 import BeautifulSoup\nimport sys\nimport os\nos.environ['KERAS_BACKEND']='theano'\nfrom keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras.utils.np_utils import to_categorical\nfrom keras.layers import Embedding\nfrom keras.layers import Dense, Input, Flatten\nfrom keras.layers import Conv1D, MaxPooling1D, Embedding, Dropout\nfrom keras.models import Model\nfrom keras.callbacks import ModelCheckpoint\nimport matplotlib.pyplot as plt\nplt.switch_backend('agg')\n%matplotlib inline\nfrom pandas import compat\ncompat.PY3 = True\n\n\ndef clean_str(string):\n    string = re.sub(r\"\\\\\", \"\", string)\n    string = re.sub(r\"\\'\", \"\", string)\n    string = re.sub(r\"\\\"\", \"\", string)\n    return string.strip().lower()\n\n\nMAX_SEQUENCE_LENGTH = 1000\nMAX_NB_WORDS = 20000\nEMBEDDING_DIM = 100\nVALIDATION_SPLIT = 0.2\n\n\nweather_data = np.load('.\/data\/weatherdata.npy')\nmusic_data = np.load('.\/data\/musicdata.npy')\nsearch_data = np.load('.\/data\/searchdata.npy')\n\n\nprint(weather_data.shape)\nprint(music_data.shape)\nprint(search_data.shape)\ndata = np.concatenate((weather_data, music_data), axis=0)\n# data = np.concatenate((data, search_data), axis=0)\n\n\ndf = pd.DataFrame()\ndf = pd.DataFrame(data)\ndf = df.dropna()\ndf = df.reset_index(drop=True)\ndf[0].astype(str).values\ndf[1].astype(str).values\ndf.columns = ['message', 'class']\nprint('Shape of dataset ',df.shape)\nprint(df.columns)\nprint('No. of unique classes',len(set(df['class'])))\n\n\n\nmacronum=sorted(set(df['class']))\nmacro_to_id = dict((note, number) for number, note in enumerate(macronum))\n\ndef fun(i):\n    return macro_to_id[i]\n\ndf['class']=df['class'].apply(fun)\nprint(df.head(5))\n\n\ntexts = []\nlabels = []\n\n\nfor idx in range(df.message.shape[0]):\n    text = BeautifulSoup(df.message[idx])\n    texts.append(clean_str(str(text.get_text().encode())))\n\nfor idx in df['class']:\n    labels.append(idx)\n\n\n\ntokenizer = Tokenizer(num_words=MAX_NB_WORDS)\ntokenizer.fit_on_texts(texts)\nsequences = tokenizer.texts_to_sequences(texts)\n\nword_index = tokenizer.word_index\nprint('Number of Unique Tokens',len(word_index))\n\n\n\ndata = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)\n\nlabels = to_categorical(np.asarray(labels))\nprint('Shape of Data Tensor:', data.shape)\nprint('Shape of Label Tensor:', labels.shape)\n\nindices = np.arange(data.shape[0])\nnp.random.shuffle(indices)\ndata = data[indices]\nlabels = labels[indices]\nnb_validation_samples = int(VALIDATION_SPLIT * data.shape[0])\n\nx_train = data[:-nb_validation_samples]\ny_train = labels[:-nb_validation_samples]\nx_val = data[-nb_validation_samples:]\ny_val = labels[-nb_validation_samples:]\n\n\n\nembeddings_index = {}\n# f = open('glove.6B.100d.txt',encoding='utf8')\nimport io\nf = io.open('glove.6B.100d.txt', encoding='utf-8')\nfor line in f:\n    values = line.split()\n    word = values[0]\n    coefs = np.asarray(values[1:], dtype='float32')\n    embeddings_index[word] = coefs\nf.close()\n\nprint('Total %s word vectors in Glove 6B 100d.' % len(embeddings_index))\n\n\n\nembedding_matrix = np.random.random((len(word_index) + 1, EMBEDDING_DIM))\nfor word, i in word_index.items():\n    embedding_vector = embeddings_index.get(word)\n    if embedding_vector is not None:\n        # words not found in embedding index will be all-zeros.\n        embedding_matrix[i] = embedding_vector\n\nembedding_layer = Embedding(len(word_index) + 1,\n                            EMBEDDING_DIM,weights=[embedding_matrix],\n                            input_length=MAX_SEQUENCE_LENGTH,trainable=True)\n\n\n\nsequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')\nembedded_sequences = embedding_layer(sequence_input)\nl_cov1= Conv1D(128, 5, activation='relu')(embedded_sequences)\nl_pool1 = MaxPooling1D(5)(l_cov1)\nl_cov2 = Conv1D(128, 5, activation='relu')(l_pool1)\nl_pool2 = MaxPooling1D(5)(l_cov2)\nl_cov3 = Conv1D(128, 5, activation='relu')(l_pool2)\nl_pool3 = MaxPooling1D(35)(l_cov3)  # global max pooling\nl_flat = Flatten()(l_pool3)\nl_dense = Dense(128, activation='relu')(l_flat)\npreds = Dense(len(macronum), activation='softmax')(l_dense)\n\nmodel = Model(sequence_input, preds)\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='rmsprop',\n              metrics=['acc'])\n\nprint(\"Simplified convolutional neural network\")\nmodel.summary()\ncp=ModelCheckpoint('model_cnn.hdf5',monitor='val_acc',verbose=1,save_best_only=True)\n\n\n\nhistory=model.fit(x_train, y_train, validation_data=(x_val, y_val),epochs=15, batch_size=2,callbacks=[cp])\n\n\nfrom keras.models import load_model\n\nmodel.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'\n# del model  # deletes the existing model\n\n# returns a compiled model\n# identical to the previous one\n# model = load_model('my_model.h5')\n\nfig1 = plt.figure()\nplt.plot(history.history['loss'],'r',linewidth=3.0)\nplt.plot(history.history['val_loss'],'b',linewidth=3.0)\nplt.legend(['Training loss', 'Validation Loss'],fontsize=18)\nplt.xlabel('Epochs ',fontsize=16)\nplt.ylabel('Loss',fontsize=16)\nplt.title('Loss Curves :CNN',fontsize=16)\nfig1.savefig('loss_cnn.png')\nplt.show()\n\n\nfig2=plt.figure()\nplt.plot(history.history['acc'],'r',linewidth=3.0)\nplt.plot(history.history['val_acc'],'b',linewidth=3.0)\nplt.legend(['Training Accuracy', 'Validation Accuracy'],fontsize=18)\nplt.xlabel('Epochs ',fontsize=16)\nplt.ylabel('Accuracy',fontsize=16)\nplt.title('Accuracy Curves : CNN',fontsize=16)\nfig2.savefig('accuracy_cnn.png')\nplt.show()\n\n\n\nfrom keras.utils.vis_utils import plot_model\nplot_model(model, to_file='cnn_model.png', show_shapes=True, show_layer_names=True)\n\n\nfrom PIL import Image\ndisplay(Image.open('cnn_model.png'))\n\n\n **The problem is here....\n Now in a new file, I am trying to load the file and predict the classes as below**\n\n\nfrom keras.models import load_model\nimport numpy as np\n\nmodel = load_model('my_model.h5')\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='rmsprop',\n              metrics=['acc'])\n\n\nclasses = model.predict_classes(\"How is the weather today\")\n\nprint classes\n<\/code><\/pre>\n\n<p>This is the code for above..<\/p>\n"}]}
+{"QuestionId":52702653,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><nlp><lstm>","CreationDate":"2018-10-08T12:45:51.293","AcceptedAnswerId":null,"OwnerUserId":2886527.0,"Title":"Finding the sum or mean of a 3d matrix with variable length in tensorflow","Body":"<p>I have to do a averaging of 3d Tensor, where first dimension represents batch_size , second dimension reporesents max_length of sentence ( time axis ) in the batch and last dimension represents the embedding dimension. Those who are familiar with lstm, it is obtained by <code>tf.nn.emebedding_lookup<\/code><\/p>\n\n<p>For example:<\/p>\n\n<pre><code>Assume I have 3 sentences\n[ [i, love, you,], [i, don't, love, you,], [i, always, love, you, so, much ]]\n<\/code><\/pre>\n\n<p>Here <code>batch_size = 3<\/code>, <code>max_length = 6<\/code> (3rd sentence ) and assume <code>embedding dimension = 100<\/code>. Normally, we will pad the first 2 sentences to match the max_length. Now, I need to average the word embeddings of each word. But, if I am using tf.reduce_sum, it will consider those padded vectors into consideration for the first two sentences, which is wrong. Is there an efficient way to do this in tensorflow.  <\/p>\n","answers":[{"AnswerId":"52711208","CreationDate":"2018-10-08T23:10:07.517","ParentId":null,"OwnerUserId":"8337524","Title":null,"Body":"<p>A possible solution consists in passing the lengths of the original sentences (the ones without padding) to the model. In this way we can compute correct average embedding for each sentence.<\/p>\n\n<p>In the preprocessing phase (when you generate your sentences), keep track of the length of each sentence. Suppose you generate the sentences with a <code>generate_batch<\/code> function, then:<\/p>\n\n<pre><code>batch = generate_batch(...)\nbatch_sentences = batch[\"sentences\"] # [[i, love, you], [i, don't, love, you], ...]\nbatch_sentence_lengths = batch[\"sentence_lengths\"] # [3, 4, ...]\n<\/code><\/pre>\n\n<p>Now you can feed the sentences and their lengths to the model:<\/p>\n\n<pre><code>with tf.Session(...) as sess:\n    ...\n    (loss, ) = sess.run(\n        [loss],\n        feed_dict = {\n            sentences: batch_sentences,\n            sentence_lengths: batch_sentence_lengths,\n            ...\n        })\n    ...\n<\/code><\/pre>\n\n<p>You can use the length of each sentence in your model now:<\/p>\n\n<pre><code>...\n# sentence_lengths is a sequence of integers: convert it to a sequence of floats\n# sentence_lengths_float.shape = sentence_lengths.shape = (batch_size, )\nsentence_lengths_float = tf.cast(sentence_lengths, tf.float32)\n\n# Compute the sum of the embeddings for each sentence.\n# If sentence_embeddings.shape = (batch_size, max_sentence_length, embedding_size), then sentence_axis = 1\n# embeddings_sum_for_each_sentence.shape = (batch_size, embeddings_size)\nembeddings_sum_for_each_sentence = tf.reduce_sum(sentence_embeddings, axis=sentence_axis)\n\n# tf.div(a, b) divides each element of the last dimension of a by each element of b as long as the a.shape[-1] = n and b.shape = (1, n). See broadcasting in tf.\n# If a is matrix, then tf.div divides each element of a row by the corresponding element in b. But we want a column-wise division, so we need to transpose a first.\n# embeddings_avg_for_each_sentence_t.shape = (embedding_size, batch_size)\nembeddings_avg_for_each_sentence_t = tf.div(tf.transpose(embeddings_sum_for_each_sentence), sentence_lengths_float)\n\n# Finally we need to tranpose the result again.\n# embeddings_avg_for_each_sentence.shape = (batch_size, embedding_size)\nembeddings_avg_for_each_sentence = tf.tranpose(embeddings_avg_for_each_sentence_t)\n...\n<\/code><\/pre>\n"},{"AnswerId":"52771983","CreationDate":"2018-10-12T03:50:39.550","ParentId":null,"OwnerUserId":"2886527","Title":null,"Body":"<p>The way to do this is as follows. It is a bit complicated but works fine. <\/p>\n\n<p>Some functions are obtained from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/rnn.py<\/a> . I recommend to do the imports same as above code. The main code is as follows<\/p>\n\n<pre><code>def _dynamic_average_loop(inputs,\n                  initial_state,\n                  parallel_iterations,\n                  swap_memory,\n                  sequence_length=None,\n                  dtype=None):\n\n        state = initial_state\n        assert isinstance(parallel_iterations, int), \"parallel_iterations must be int\"\n        flat_input = nest.flatten(inputs)\n        embedding_dimension = tf.shape(inputs)[2]\n        flat_output_size = [embedding_dimension]\n\n        # Construct an initial output\n        input_shape = array_ops.shape(flat_input[0])\n        time_steps = input_shape[0]\n        batch_size = _best_effort_input_batch_size(flat_input)\n\n        inputs_got_shape = tuple(input_.get_shape().with_rank_at_least(3)\n                               for input_ in flat_input)\n\n        const_time_steps, const_batch_size = inputs_got_shape[0].as_list()[:2]\n\n        for shape in inputs_got_shape:\n            if not shape[2:].is_fully_defined():\n              raise ValueError(\n                  \"Input size (depth of inputs) must be accessible via shape inference,\"\n                  \" but saw value None.\")\n            got_time_steps = shape[0].value\n            got_batch_size = shape[1].value\n            if const_time_steps != got_time_steps:\n              raise ValueError(\n                  \"Time steps is not the same for all the elements in the input in a \"\n                  \"batch.\")\n            if const_batch_size != got_batch_size:\n              raise ValueError(\n        \"Batch_size is not the same for all the elements in the input.\")\n        # Prepare dynamic conditional copying of state &amp; output\n        def _create_zero_arrays(size):\n            size = _concat(batch_size, size)\n            return array_ops.zeros(\n                array_ops.stack(size), _infer_state_dtype(dtype, state))\n\n        flat_zero_output = tuple(_create_zero_arrays(output)\n                               for output in flat_output_size)\n        zero_output = nest.pack_sequence_as(structure=embedding_dimension,\n                                          flat_sequence=flat_zero_output)\n\n        if sequence_length is not None:\n            min_sequence_length = math_ops.reduce_min(sequence_length)\n            max_sequence_length = math_ops.reduce_max(sequence_length)\n        else:\n            max_sequence_length = time_steps\n\n        time = array_ops.constant(0, dtype=dtypes.int32, name=\"time\")\n\n        with ops.name_scope(\"dynamic_rnn\") as scope:\n            base_name = scope\n\n        def _create_ta(name, element_shape, dtype):\n            return tensor_array_ops.TensorArray(dtype=dtype,\n                                                size=time_steps,\n                                                element_shape=element_shape,\n                                                tensor_array_name=base_name + name)\n\n        in_graph_mode = not context.executing_eagerly()\n        if in_graph_mode:\n            output_ta = tuple(\n                _create_ta(\n                    \"output_%d\" % i,\n                    element_shape=(tensor_shape.TensorShape([const_batch_size])\n                                   .concatenate(\n                                       _maybe_tensor_shape_from_tensor(out_size))),\n                    dtype=_infer_state_dtype(dtype, state))\n                for i, out_size in enumerate(flat_output_size))\n            input_ta = tuple(\n                _create_ta(\n                    \"input_%d\" % i,\n                    element_shape=flat_input_i.shape[1:],\n                    dtype=flat_input_i.dtype)\n                for i, flat_input_i in enumerate(flat_input))\n            input_ta = tuple(ta.unstack(input_)\n                             for ta, input_ in zip(input_ta, flat_input))\n        else:\n            output_ta = tuple([0 for _ in range(time_steps.numpy())]\n                              for i in range(len(flat_output_size)))\n            input_ta = flat_input\n\n\n        def tf_average(A, B):\n\n            return A+B \n\n\n        def _time_step(time, output_ta_t, state):\n\n            input_t = tuple(ta.read(time) for ta in input_ta)\n            # Restore some shape information\n            for input_, shape in zip(input_t, inputs_got_shape):\n                input_.set_shape(shape[1:])\n\n            input_t = nest.pack_sequence_as(structure=inputs, flat_sequence=input_t)\n\n            flat_state = nest.flatten(state)\n            flat_zero_output = nest.flatten(zero_output)\n\n            # Vector describing which batch entries are finished.\n            copy_cond = time &gt;= sequence_length\n\n\n\n            def _copy_one_through(output, new_output):\n                # Otherwise propagate the old or the new value.\n                with ops.colocate_with(new_output):\n                    return array_ops.where(copy_cond, output, new_output)\n\n            the_average = tf_average(input_t, state)\n            the_average_updated = _copy_one_through(zero_output, the_average)\n            the_average_last_state = _copy_one_through(state, the_average)\n\n            for output, flat_output in zip([the_average_updated], flat_zero_output):\n                output.set_shape(flat_output.get_shape())\n\n            final_output = nest.pack_sequence_as(structure=zero_output, flat_sequence=[the_average_updated])\n            output_ta_t = tuple(ta.write(time, out) for ta, out in zip(output_ta_t, [final_output]))\n            return (time + 1, output_ta_t, the_average_last_state)\n\n\n        if in_graph_mode:\n            # Make sure that we run at least 1 step, if necessary, to ensure\n            # the TensorArrays pick up the dynamic shape.\n            loop_bound = math_ops.minimum(\n                time_steps, math_ops.maximum(1, max_sequence_length))\n        else:\n        # Using max_sequence_length isn't currently supported in the Eager branch.\n            loop_bound = time_steps\n\n        _, output_final_ta, final_state = control_flow_ops.while_loop(\n              cond=lambda time, *_: time &lt; loop_bound,\n              body=_time_step,\n              loop_vars=(time, output_ta, state),\n              parallel_iterations=parallel_iterations,\n              maximum_iterations=time_steps,\n             swap_memory=swap_memory)\n\n        final_outputs = tuple(ta.stack() for ta in output_final_ta)\n            # Restore some shape information\n        for output, output_size in zip(final_outputs, flat_output_size):\n            shape = _concat(\n              [const_time_steps, const_batch_size], output_size, static=True)\n            output.set_shape(shape)\n\n        final_outputs = nest.pack_sequence_as(structure=embedding_dimension,\n                                              flat_sequence=final_outputs)\n\n        return final_outputs , final_state\n\n\n\ndef dynamic_average(inputs, sequence_length=None, initial_state=None,\n            dtype=None, parallel_iterations=None, swap_memory=False,\n            time_major=False, scope=None):\n\n        with vs.variable_scope(scope or \"rnn\") as varscope:\n        # Create a new scope in which the caching device is either\n        # determined by the parent scope, or is set to place the cached\n        # Variable using the same placement as for the rest of the RNN.\n            if _should_cache():\n              if varscope.caching_device is None:\n                varscope.set_caching_device(lambda op: op.device)\n\n            # By default, time_major==False and inputs are batch-major: shaped\n            #   [batch, time, depth]\n            # For internal calculations, we transpose to [time, batch, depth]\n        flat_input = nest.flatten(inputs)\n        embedding_dimension = tf.shape(inputs)[2]\n\n        if not time_major:\n          # (B,T,D) =&gt; (T,B,D)\n          flat_input = [ops.convert_to_tensor(input_) for input_ in flat_input]\n          flat_input = tuple(_transpose_batch_time(input_) for input_ in flat_input)\n\n        parallel_iterations = parallel_iterations or 32\n        if sequence_length is not None:\n          sequence_length = math_ops.to_int32(sequence_length)\n          if sequence_length.get_shape().ndims not in (None, 1):\n            raise ValueError(\n                \"sequence_length must be a vector of length batch_size, \"\n                \"but saw shape: %s\" % sequence_length.get_shape())\n          sequence_length = array_ops.identity(  # Just to find it in the graph.\n              sequence_length, name=\"sequence_length\")\n\n        batch_size = _best_effort_input_batch_size(flat_input)\n\n        state = tf.zeros(shape=(batch_size, embedding_dimension))\n\n        def _assert_has_shape(x, shape):\n          x_shape = array_ops.shape(x)\n          packed_shape = array_ops.stack(shape)\n          return control_flow_ops.Assert(\n              math_ops.reduce_all(math_ops.equal(x_shape, packed_shape)),\n              [\"Expected shape for Tensor %s is \" % x.name,\n               packed_shape, \" but saw shape: \", x_shape])\n\n        if not context.executing_eagerly() and sequence_length is not None:\n          # Perform some shape validation\n          with ops.control_dependencies(\n              [_assert_has_shape(sequence_length, [batch_size])]):\n            sequence_length = array_ops.identity(\n                sequence_length, name=\"CheckSeqLen\")\n\n        inputs = nest.pack_sequence_as(structure=inputs, flat_sequence=flat_input)\n\n        (outputs, final_state) = _dynamic_average_loop(\n            inputs,\n            state,\n            parallel_iterations=parallel_iterations,\n            swap_memory=swap_memory,\n            sequence_length=sequence_length,\n    dtype=dtype)\n\n        if not time_major:\n            outputs = nest.map_structure(_transpose_batch_time, outputs)\n        return outputs, final_state\n<\/code><\/pre>\n\n<p>This is the main code. So to find the sum of a 3D matrix, of variable length as in RNN, we can test it as follows<\/p>\n\n<pre><code>    tf.reset_default_graph()\n\n    the_inputs = np.random.uniform(-1,1,(30,50,111)).astype(np.float32)\n    the_length = np.random.randint(50, size=30)\n    the_input_tensor = tf.convert_to_tensor(the_inputs)\n    the_length_tensor = tf.convert_to_tensor(the_length)\n\n    outputs, final_state = dynamic_average(inputs=the_input_tensor, \nsequence_length=the_length_tensor)\n\n    sess = tf.InteractiveSession()\n    sess.run(tf.global_variables_initializer())\n    outputs_result , final_state_result = sess.run((outputs, final_state))\n\n    print(\"Testing\")\n    for index in range(len(the_inputs)):\n\n        print(the_inputs[index,:,:][:the_length[index]].sum(axis=0) == final_state_result[index])\n    print('------------------------------------------------------------------')\n<\/code><\/pre>\n"}]}
+{"QuestionId":52703047,"AnswerCount":2,"Tags":"<tensorflow><filesystems><local><google-cloud-tpu><tpu>","CreationDate":"2018-10-08T13:07:28.223","AcceptedAnswerId":null,"OwnerUserId":1273813.0,"Title":"TPU local Filesystem doesn't exist?","Body":"<p>I wrote a NN model that analyze an image and extract 8 floating numbers at the end. The model is working fine (but slowly) on my computer so I try it on the TPU cloud and there BAM! I have an error:<\/p>\n\n<blockquote>\n  <p>I1008 12:58:47.077905 140221679261440 tf_logging.py:115] Error recorded from training_loop: File system scheme '[local]' not implemented (file: '\/home\/gcloud_iba\/Data\/CGTR\/model\/GA_subset\/model.ckpt-0_temp_e840841d93124a67b54074b1c0fd7ae4')\n       [[{{node save\/SaveV2}} = SaveV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64], _device=\"\/job:worker\/replica:0\/task:0\/device:CPU:0\"](save\/ShardedFilename, save\/SaveV2\/tensor_names, save\/SaveV2\/shape_and_slices, batch_normalization\/beta\/Read\/ReadVariableOp, batch_normalization\/beta\/Momentum\/Read_1\/ReadVariableOp, batch_normalization\/gamma\/Read\/ReadVariableOp, batch_normalization\/gamma\/Momentum\/Read_1\/ReadVariableOp, batch_normalization\/moving_mean\/Read\/ReadVariableOp, batch_normalization\/moving_variance\/Read\/ReadVariableOp, batch_normalization_1\/beta\/Read\/ReadVariableOp, batch_normalization_1\/beta\/Momentum\/Read_1\/ReadVariableOp, batch_normalization_1\/gamma\/Read\/ReadVariableOp, batch_normalization_1\/gamma\/Momentum\/Read_1\/ReadVariableOp, batch_normalization_1\/moving_mean\/Read\/ReadVariableOp, batch_normalization_1\/moving_variance\/Read\/ReadVariableOp, conv2d\/kernel\/Read\/ReadVariableOp, conv2d\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_1\/kernel\/Read\/ReadVariableOp, conv2d_1\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_2\/kernel\/Read\/ReadVariableOp, conv2d_2\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_3\/kernel\/Read\/ReadVariableOp, conv2d_3\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_4\/kernel\/Read\/ReadVariableOp, conv2d_4\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_5\/kernel\/Read\/ReadVariableOp, conv2d_5\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_6\/kernel\/Read\/ReadVariableOp, conv2d_6\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_7\/kernel\/Read\/ReadVariableOp, conv2d_7\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_8\/kernel\/Read\/ReadVariableOp, conv2d_8\/kernel\/Momentum\/Read_1\/ReadVariableOp, conv2d_9\/kernel\/Read\/ReadVariableOp, conv2d_9\/kernel\/Momentum\/Read_1\/ReadVariableOp, dense\/bias\/Read\/ReadVariableOp, dense\/bias\/Momentum\/Read_1\/ReadVariableOp, dense\/kernel\/Read\/ReadVariableOp, dense\/kernel\/Momentum\/Read_1\/ReadVariableOp, dense_1\/bias\/Read\/ReadVariableOp, dense_1\/bias\/Momentum\/Read_1\/ReadVariableOp, dense_1\/kernel\/Read\/ReadVariableOp, dense_1\/kernel\/Momentum\/Read_1\/ReadVariableOp, dense_2\/bias\/Read\/ReadVariableOp, dense_2\/bias\/Momentum\/Read_1\/ReadVariableOp, dense_2\/kernel\/Read\/ReadVariableOp, dense_2\/kernel\/Momentum\/Read_1\/ReadVariableOp, dense_3\/bias\/Read\/ReadVariableOp, dense_3\/bias\/Momentum\/Read_1\/ReadVariableOp, dense_3\/kernel\/Read\/ReadVariableOp, dense_3\/kernel\/Momentum\/Read_1\/ReadVariableOp, global_step\/Read\/ReadVariableOp)]]<\/p>\n<\/blockquote>\n\n<p>I checked that the TPU has access to the hard drive and it works (I have another piece of code that reads the same dataset with another model).\nI do not cache my data (yet) but I do some prefetching. Aside this, I don't see what isn't working?<\/p>\n\n<p>Thank you for any help you could provide!<\/p>\n\n<p>Pi-r<\/p>\n","answers":[{"AnswerId":"52880454","CreationDate":"2018-10-18T18:35:04.177","ParentId":null,"OwnerUserId":"3870291","Title":null,"Body":"<p>In absence of Google Cloud Storage, write your model using Keras API (<a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/<\/a>). <\/p>\n"},{"AnswerId":"52860914","CreationDate":"2018-10-17T17:54:42.460","ParentId":null,"OwnerUserId":"4716286","Title":null,"Body":"<p>The local filesystem is not available on Cloud TPU's. Model directories (checkpoints etc) and input data should be stored in Google Cloud Storage (and prefixed with \"gs:\/\/\").<\/p>\n\n<p>More details here<\/p>\n\n<p><a href=\"https:\/\/cloud.google.com\/tpu\/docs\/storage-buckets\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/tpu\/docs\/storage-buckets<\/a><\/p>\n"}]}
+{"QuestionId":52703096,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2018-10-08T13:10:09.650","AcceptedAnswerId":null,"OwnerUserId":3737468.0,"Title":"Calculate\/Visualize Tensorflow Keras Dense model layer relative connection weights w.r.t output classes","Body":"<p>Here is my tensorflow keras model,(you can ignore dropout layer if it makes things tough)<\/p>\n\n<pre><code>import tensorflow as tf\noptimizers = tf.keras.optimizers\nSequential = tf.keras.models.Sequential\nDense = tf.keras.layers.Dense\nDropout = tf.keras.layers.Dropout\nto_categorical = tf.keras.utils.to_categorical\n\nmodel = Sequential()\nmodel.add(Dense(256, input_shape=(20,), activation=\"relu\"))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(256, activation=\"relu\"))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(256, activation=\"relu\"))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(3, activation=\"softmax\"))\n\nadam = optimizers.Adam(lr=1e-3) # I don't mind rmsprop either\nmodel.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])  \nmodel.summary()\n<\/code><\/pre>\n\n<p>and I have saved the model structure and weights as<\/p>\n\n<pre><code>model.save(\"sim_score.h5\", overwrite=True)\nmodel.save_weights('sim_score_weights.h5', overwrite=True)\n<\/code><\/pre>\n\n<p>on doing <code>model.predict(X_test)<\/code> I get <code>[0.23, 0.63, 0.14]<\/code> which is the prediction probability of the 3 output classes.<\/p>\n\n<blockquote>\n  <p>How would I visualize how much weight\/importance each of my initial 20\n  features have in this model w.r.t the 3 output softmax?<\/p>\n<\/blockquote>\n\n<p>For example, my 2nd column could have negligible impact on the end result, while the 5th column could have 3x more impact on the output prediction than the 20th column. It does not matter what the absolute effect of the 5th column is, just figuring out the relative importance is good enough, for example <code>5th column = 0.3, 20th column = 0.1<\/code> and so on for a <code>20 x 3 matrix<\/code>.<\/p>\n\n<p>See this <a href=\"https:\/\/cdn-images-1.medium.com\/max\/1760\/0*Atm4E9LZrhecmuGB.gif\" rel=\"nofollow noreferrer\">animation<\/a> for intuition or <a href=\"https:\/\/playground.tensorflow.org\/#activation=tanh&amp;batchSize=10&amp;dataset=circle&amp;regDataset=reg-plane&amp;learningRate=0.03&amp;regularizationRate=0&amp;noise=0&amp;networkShape=4,2&amp;seed=0.45952&amp;showTestData=false&amp;discretize=false&amp;percTrainData=50&amp;x=true&amp;y=true&amp;xTimesY=false&amp;xSquared=false&amp;ySquared=false&amp;cosX=false&amp;sinX=false&amp;cosY=false&amp;sinY=false&amp;collectStats=false&amp;problem=classification&amp;initZero=false&amp;hideText=false\" rel=\"nofollow noreferrer\">Tensorflow playground<\/a>. The visualization doesn't have to show how the weights change during training, but can just show a snapshot image of how it looks at the end of training. <\/p>\n\n<blockquote>\n  <p>In fact, the solution does not even have to be a visualization, it can\n  even be an array of 20 elements x 3 outputs having the relative importance of each\n  feature w.r.t the 3 output softmax and importance relative to the other features.<\/p>\n<\/blockquote>\n\n<p>Getting the importance of intermediate layers is just an added bonus.<\/p>\n\n<p><strong>The reason I want to visualize the 20 features is for transparency purposes(currently the model feels like a black box).<\/strong> I am comfortable in matplotlib, pyplot, seaborn. \nI am aware of Tensorboard as well, but couldn't find any examples for simple Dense Relu network with Softmax. <\/p>\n\n<p>I feel a time-consuming way to get the 20 x 3 weights would be to do a domain search of <code>20 features<\/code> from <code>0 - 1<\/code> with delta of <code>0.5<\/code> by sending various inputs and trying to deduce the importance of features based off of that(which would have <code>3 to the power of 20<\/code> ~= <code>3.4 billion<\/code> possible sample space and gets exponentially worse the more features I add) and then applying conditional probability to reverse engineer the relative weights, but I am not sure if there is a simpler\/optimized way via TensorBoard or some custom logic.<\/p>\n\n<p>Can someone help visualize the model\/calculate the 20 x 3 = 60 relative weights of the 20 features w.r.t the 3 outputs with code snippets or provide a reference on how to achieve this? <\/p>\n","answers":[{"AnswerId":"53813389","CreationDate":"2018-12-17T10:35:55.993","ParentId":null,"OwnerUserId":"9114634","Title":null,"Body":"<p>I had come across a similar problem, but i was concerned more about the visualization of the model parameters(weights and biases) rather than the model features [<strong>since i wanted to explore and view the black box as well<\/strong>].<\/p>\n\n<p>For example, the following is a snippet of a shallow neural network with 2 hidden layers.  <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(128, input_dim=13, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(64, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(64, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(8, kernel_initializer='uniform', activation='softmax'))\n\n# Compile model\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# Using TensorBoard to visualise the Model\nks=TensorBoard(log_dir=\"\/your_full_path\/logs\/{}\".format(time()), histogram_freq=1, write_graph=True, write_grads=True, batch_size=10)\n\n# Fit the model   \nmodel.fit(X, Y, epochs = 64, shuffle = True, batch_size=10, verbose = 2, validation_split=0.2, callbacks=[ks])\n<\/code><\/pre>\n\n<p>For one to be able to visualize the parameters, there are few important things to be kept in mind:<\/p>\n\n<ol>\n<li><p>Always ensure to have a validation_split in the model.fit() function[else histograms cannot be visualised].<\/p><\/li>\n<li><p>Make sure the value of <strong>histogram_freq > 0<\/strong> always!![histograms won't be computed otherwise].<\/p><\/li>\n<li><p>The callbacks to TensorBoard have to be specified in the model.fit() as a list.<\/p><\/li>\n<\/ol>\n\n<p>Once, this is done; goto cmd and type the folllowing command:<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>tensorboard --logdir=logs\/<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>This gives you a local address with which you can access TensorBoard on your web browser. \nAll histograms, distributions, loss and accuracy functions will be available as plots and can be selected from the menu bar at the top.<\/p>\n\n<p>Hope this answer gives a hint about the procedure to visualize model parameters(I myself had a bit of a struggle as the above points weren't available together).<\/p>\n\n<p>Do let me know if it helped.<\/p>\n\n<p>Below is the keras documentation link for your reference:<\/p>\n\n<p><a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#tensorboard<\/a><\/p>\n"}]}
+{"QuestionId":52703204,"AnswerCount":1,"Tags":"<keras><deep-learning><conv-neural-network><loss-function>","CreationDate":"2018-10-08T13:15:49.083","AcceptedAnswerId":null,"OwnerUserId":9027070.0,"Title":"Why my loss function or accuracy does not improve?","Body":"<p>I have 3D CNN U-net architecture to solve segmentation problem. I am using Adam optimisation together with binary cross entropy and the metric is \"accuracy\". I try to understand why it does not improve.  <\/p>\n\n<pre><code>Train on 2774 samples, validate on 694 samples\nEpoch 1\/20\n2774\/2774 [==============================] - 166s 60ms\/step - loss: 0.5189 - acc: 0.7928 - val_loss: 0.5456 - val_acc: 0.7674\n\nEpoch 00001: val_loss improved from inf to 0.54555, saving model to model-tgs-salt-1.h5\nEpoch 2\/20\n2774\/2774 [==============================] - 170s 61ms\/step - loss: 0.5170 - acc: 0.7928 - val_loss: 0.5485 - val_acc: 0.7674\n\nEpoch 00002: val_loss did not improve from 0.54555\nEpoch 3\/20\n2774\/2774 [==============================] - 169s 61ms\/step - loss: 0.5119 - acc: 0.7928 - val_loss: 0.5455 - val_acc: 0.7674\n\nEpoch 00003: val_loss improved from 0.54555 to 0.54549, saving model to model-tgs-salt-1.h5\nEpoch 4\/20\n2774\/2774 [==============================] - 170s 61ms\/step - loss: 0.5117 - acc: 0.7928 - val_loss: 0.5715 - val_acc: 0.7674\n\nEpoch 00004: val_loss did not improve from 0.54549\nEpoch 5\/20\n2774\/2774 [==============================] - 169s 61ms\/step - loss: 0.5126 - acc: 0.7928 - val_loss: 0.5566 - val_acc: 0.7674\n\nEpoch 00005: val_loss did not improve from 0.54549\nEpoch 6\/20\n2774\/2774 [==============================] - 169s 61ms\/step - loss: 0.5138 - acc: 0.7928 - val_loss: 0.5503 - val_acc: 0.7674\n\nEpoch 00006: val_loss did not improve from 0.54549\nEpoch 7\/20\n2774\/2774 [==============================] - 170s 61ms\/step - loss: 0.5103 - acc: 0.7928 - val_loss: 0.5444 - val_acc: 0.7674\n\nEpoch 00007: val_loss improved from 0.54549 to 0.54436, saving model to model-tgs-salt-1.h5\nEpoch 8\/20\n2774\/2774 [==============================] - 169s 61ms\/step - loss: 0.5137 - acc: 0.7928 - val_loss: 0.5454 - val_acc: 0.7674\n<\/code><\/pre>\n","answers":[{"AnswerId":"52709496","CreationDate":"2018-10-08T20:14:19.113","ParentId":null,"OwnerUserId":"8079690","Title":null,"Body":"<p>If you use batch size in your network. let's try to increase that. I think it could be useful in speed of train.<\/p>\n"}]}
+{"QuestionId":52703283,"AnswerCount":0,"Tags":"<python><keras><nlp><word-embedding><glove>","CreationDate":"2018-10-08T13:21:04.853","AcceptedAnswerId":null,"OwnerUserId":7115193.0,"Title":"Use Glove vectors without Embedding layers in LSTM","Body":"<p>I want to use the Glove vectors in Language Modeling. But the problem is if I use Embedding layer in the model, I can't predict the output vector and match the word. What I mean here is, I want to give the glove vector representation for my sentences as the Input. and get them out from the Lstm layer and get the vectors and match it with Glove vectors I want to use the glove vectors without embedding layer.  can someone propose a method to do it? I am using the keras and python3<\/p>\n\n<p>What I want is to use the embedding layer as one model1 and return the output vector and give it to another LSTM model2 as the input. where it gives the index of the word vector.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MlLz9.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MlLz9.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":52703703,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T13:44:57.287","AcceptedAnswerId":52704220.0,"OwnerUserId":8522251.0,"Title":"How to pass output of one graph as input to another graph in tensorflow","Body":"<p>I want to do transfer learning with INCEPTION_V4 model as feature extractor \nand downloaded the code and checkpoint file from \n<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim#pre-trained-models\" rel=\"nofollow noreferrer\">GitHub repository of Tensorflow<\/a><\/p>\n\n<p>Then I added my own layer for classification in 5 classes. But during model restore using tf.train.Saver, it showing error that it can not find variable values for my layer that I added.<\/p>\n\n<p>To solve this, I created two separate graphs, one for loading pre-trained model and one for my classification layers. But I can't pass output of one graph as input to the second graph. \nCan you suggest any other way to do transfer learning or solve the problem<\/p>\n","answers":[{"AnswerId":"52704220","CreationDate":"2018-10-08T14:13:51.267","ParentId":null,"OwnerUserId":"748102","Title":null,"Body":"<p>The easy solution is to just construct the inception model (without your layers), then create the saver and use it for restore, and only then create your layers.<\/p>\n\n<p>There are more complex solutions (you can pass the <code>var_list<\/code> parameter of <code>tf.Saver<\/code> with the list of all variables you want to restore, and initialize that to be the list of all the inception variables), but this one should be straightforward, and I do it with my transfer learning models.<\/p>\n"}]}
+{"QuestionId":52704988,"AnswerCount":1,"Tags":"<keras><nlp><lstm><word-embedding><sentence-similarity>","CreationDate":"2018-10-08T14:57:02.547","AcceptedAnswerId":52706464.0,"OwnerUserId":10280273.0,"Title":"How to use my own sentence embeddings in Keras?","Body":"<p>I am new to Keras and I created my own tf_idf sentence embeddings with shape (no_sentences, embedding_dim). I am trying to add this matrix as input to an LSTM layer. My network looks something like this:<\/p>\n\n<pre><code>q1_tfidf = Input(name='q1_tfidf', shape=(max_sent, 300))\nq2_tfidf = Input(name='q2_tfidf', shape=(max_sent, 300))\n\nq1_tfidf = LSTM(100)(q1_tfidf)\nq2_tfidf = LSTM(100)(q2_tfidf)\ndistance2 = Lambda(preprocessing.exponent_neg_manhattan_distance, output_shape=preprocessing.get_shape)(\n        [q1_tfidf, q2_tfidf])\n<\/code><\/pre>\n\n<p>I'm struggling with how the matrix should be shaped. I am getting this error: <\/p>\n\n<pre><code>ValueError: Error when checking input: expected q1_tfidf to have 3 dimensions, but got array with shape (384348, 300)\n<\/code><\/pre>\n\n<p>I already checked this post: <a href=\"https:\/\/stackoverflow.com\/questions\/45001437\/sentence-embedding-in-keras\">Sentence Embedding Keras<\/a> but still can't figure it out. It seems like I'm missing something obvious.<\/p>\n\n<p>Any idea how to do this?<\/p>\n","answers":[{"AnswerId":"52706464","CreationDate":"2018-10-08T16:24:52.060","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>Ok as far as I understood, you want to predict the difference between two sentences.\nWhat about reusing the LSTM layer (the language model should be the same) and just learn a single sentence embedding and use it twice:<\/p>\n\n<pre><code>q1_tfidf = Input(name='q1_tfidf', shape=(max_sent, 300))\nq2_tfidf = Input(name='q2_tfidf', shape=(max_sent, 300))\n\nlstm = LSTM(100)\n\nlstm_out_q1= lstm (q1_tfidf)\nlstm_out_q2= lstm (q2_tfidf)\npredict = concatenate([lstm_out_q1, lstm_out_q2])\nmodel = Model(inputs=[q1_tfidf ,q1_tfidf ], outputs=predict)\n\npredict = concatenate([q1_tfidf , q2_tfidf])\n<\/code><\/pre>\n\n<p>You could also introduce your custom distance in an additional lambda layer, but therefore you need to use a different reshaping in concatenation.<\/p>\n"}]}
+{"QuestionId":52705471,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><machine-learning><binary-image>","CreationDate":"2018-10-08T15:23:30.133","AcceptedAnswerId":52708634.0,"OwnerUserId":3753275.0,"Title":"Tensorflow finding pixels with matching value","Body":"<p>I am trying to create data batches for training 2 class semantic segmentation network. The target segmented image has 2 layers, first layer with 1 for all pixels of class-1 and 0 otherwise. The second layer has pixels inverted.<\/p>\n\n<p>In the dataset I have the output images are 3 channel rgb images with <code>[255,255,255]<\/code> and <code>[0,0,0]<\/code>. The input and output images are stored in tf-record files.<\/p>\n\n<p>When I was experimenting in numpy I created a 2 channel binary image with code below:<\/p>\n\n<pre><code>c1_pix = np.all(op_img == np.array([255,255,255]), axis=2)\nc1_pix = c1_pix.reshape(*(h,w), 1)\nop_arr = np.concatenate((c1_pix, np.invert(c1_pix)), axis=2)\n<\/code><\/pre>\n\n<p>This gave me a 2 layer 1s and 0s images I wanted.<\/p>\n\n<p>I am trying to repeat it in tensorflow and I am new to it. I tried <code>c1_pix = tf.where(tf.equal(op_img, [[[255,255,255]]]))<\/code>. It seems to work but it returns 3 channel int64 tensor of 1s and 0s and I am not able to inver it.<\/p>\n\n<p>Could someone please help me figure this out ?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"52708634","CreationDate":"2018-10-08T19:08:34.107","ParentId":null,"OwnerUserId":"3753275","Title":null,"Body":"<p>After going through tf documentation for some time, I came up with the following solution<\/p>\n\n<pre><code>c1_pix, _, _ = tf.split(tf.equal(op_img,[[[255,255,255]]]), 3, axis=-1)\nc2_pix = tf.logical_not(c1_pix)\nnew_op_img = tf.concat([c1_pix, c2_pix], -1)\n<\/code><\/pre>\n\n<p>This worked for me. <\/p>\n"}]}
+{"QuestionId":52705863,"AnswerCount":0,"Tags":"<r><tensorflow><machine-learning><google-cloud-ml>","CreationDate":"2018-10-08T15:45:57.350","AcceptedAnswerId":null,"OwnerUserId":7690338.0,"Title":"Using the predict function in the cloudml R package","Body":"<p>I'm using the <a href=\"https:\/\/tensorflow.rstudio.com\/tools\/cloudml\/articles\/getting_started.html\" rel=\"nofollow noreferrer\">cloudml<\/a> package for R, in combination with the <a href=\"https:\/\/tensorflow.rstudio.com\/tfestimators\/\" rel=\"nofollow noreferrer\">tfestimators<\/a> package, to build and deploy a DNN classifier using Tensorflow so I can use it to predict a binary response (the \"Response\" column below).<\/p>\n\n<p>I have gotten as far as successfully deploying my model to Google Cloud ML Engine. It hasn't been easy.<\/p>\n\n<p>I am stuck on the last piece of the puzzle - getting predictions from my cloud model. Specifically, I need help structuring my data for \"instances\" argument to the \"cloudml_predict\" function. <a href=\"https:\/\/tensorflow.rstudio.com\/tools\/cloudml\/articles\/deployment.html\" rel=\"nofollow noreferrer\">This<\/a> is the only package documentation with an example I have found online for useRs but I don't find it applicable to my situation.<\/p>\n\n<p>Example data set:<\/p>\n\n<pre><code>library(dplyr)\n\nmy_data &lt;- data_frame(\n    Response = sample(c(\"Y\", \"N\"), 1000, replace = TRUE),\n    Col1 = sample(c(\"A\", \"B\"), 1000, replace = TRUE),\n    Col2 = sample(c(\"Happy\", \"Sad\"), 1000, replace = TRUE),\n    Col3 = sample(1:100, 1000, replace = TRUE),\n    col4 = sample(500:1000, 1000, replace = TRUE),\n    Col5 = runif(1000, 0, 1)\n)\n<\/code><\/pre>\n\n<p>This is my \"train.R\" script:<\/p>\n\n<pre><code>library(tfestimators)\nlibrary(caret)        # for even partitioning of dataset\n\n# Create feature columns ----\n    FLAGS &lt;- flags(\n        flag_numeric(\"num_epochs\", 100)\n    )\n\nnumeric_cols &lt;- colnames(my_data)[unlist(lapply(my_data, is.numeric))]\n\ncols &lt;- feature_columns(\n    column_numeric(numeric_cols),\n    column_indicator(\n        column_categorical_with_vocabulary_list(\n            \"Col1\",\n            vocabulary_list = unique(my_data$Col1),\n            dtype = tf$string\n        )\n    ),\n    column_indicator(\n        column_categorical_with_vocabulary_list(\n            \"Col2\",\n            vocabulary_list = unique(my_data$Col2),\n            dtype = tf$string\n        )\n    ),\n    column_indicator(\n        column_crossed(\n            keys = c(\"Col1\", \"Col2\"),\n            hash_bucket_size = 20L\n        )\n    )\n)\n\n\n\n# Split the data into training &amp; test sets ----\nset.seed(3456)\n\ninTrain &lt;- createDataPartition(my_data$Response,\n                               p = 0.8,\n                               list = FALSE,\n                               times = 1)\n\ntraining &lt;- my_data[inTrain, ]\ntesting  &lt;- my_data[-inTrain, ]\n\n#' Now we need to create an input function with the listing of input and output\n#' variables.\n\npred_fn &lt;- function(data, num_epochs = 10) {\n    input_fn(data,\n             features = setdiff(names(data), \"Response\"),\n             response = \"Response\",\n             num_epochs = num_epochs)\n}\n\n#' Build a deep learning classifier\n#' We will create 3 hidden layers, with 80, 40 and 30 nodes, respectively\n\nclassifier &lt;- dnn_classifier(\n    feature_columns = cols,\n    hidden_units = c(80, 40, 30),\n    n_classes = 2L,\n    label_vocabulary = c(\"N\", \"Y\"),\n    model_dir = \".\/tmp\",\n    activation_fn = \"relu\"\n)\n\nclassifier %&gt;% tfestimators::train(input_fn = pred_fn(training, num_epochs = FLAGS$num_epochs))\nclassifier %&gt;% tfestimators::evaluate(input_fn = pred_fn(testing))\nclassifier %&gt;% predict(input_fn = pred_fn(testing))\n\n#' Help on serving_input_receiver_fn:\n#' https:\/\/rdrr.io\/github\/rstudio\/tfdeploy\/src\/tests\/testthat\/models\/tfestimators-example.R\ntfestimators::export_savedmodel(classifier,\n                                export_dir_base = \"savedmodel1\",\n                                serving_input_receiver_fn = tf$estimator$export$build_parsing_serving_input_receiver_fn(\n                                    classifier_parse_example_spec(\n                                        feature_columns = cols,\n                                        weight_column = NULL,\n                                        label_key = \"label\"\n                                    )),\n                                as_text = FALSE)\n<\/code><\/pre>\n\n<p>Deployment script:<\/p>\n\n<pre><code>library(cloudml)\n\n# Train ----\njob &lt;- cloudml_train(\"train.R\")\n\n# Find directory of exported trained model\nlatest_run()$run_dir\n\n# Deploy the model\n# Use result from latest_run()$run_dir to populate file path here\ncloudml_deploy(\"runs\/cloudml_2018_10_08_113833889\/savedmodel1\", name = \"test_model\")\n\n\n# Prediction (where I need help) ----\n\ninstances &lt;- list()\ninstances[[\"instances\"]] &lt;- list()\ninstances$instances[[\"inputs\"]] &lt;- apply(testing[1:1, -1], 1, as.list)\n\ncloudml_predict(name = \"test_model\",\n                version = \"test_model_1\",\n                instances = instances,\n                verbose = TRUE)\n<\/code><\/pre>\n\n<p>The closest I've gotten is the following error message:<\/p>\n\n<pre><code>Error in cloudml_predict(name = \"test_model\", version = \"test_model_1\",  : \n  Prediction failed: Error processing input: Expected string, got... \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52706923,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T16:57:13.707","AcceptedAnswerId":null,"OwnerUserId":7886651.0,"Title":"Error Running Tensorflow Model while Using tf.data api and customized optimizer","Body":"<p>In the following code, I wanted to train a model in tensorflow. the model is a ResNet model, a deep one, hence the batch should be small for data\/all activations to fit in memory. For this reason, I have implemented a custom optimizer that accumulates the gradients over the different fed mini-batchs, and finally apply the gradient descent once. In addition, I have used <code>tf.data<\/code> api to fetch data from <code>tfrecords<\/code> which I created. Please note that my input data are video frames; and the detected variable used indicated whether a face is detected in a certain frame or not. Hence, detected is used only for <code>MSE<\/code> (Just for clarification).<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport csv\nimport os\n\nnum_epoch = 100\nlatent_dim = 100\ncell_size = 100\n\n# for each input frame, I have 3 outputs.\nnum_classes = 3\n\ncommon = \"C:\/Users\/user\/Documents\/SEWA_db\/tfrecords_db\/\"\n\nfilenames_train = []\nfilenames_dev = []\n\nfor i in range(1, 35):\n    filenames_train.append(common + \"Train_DE_{num:02d}.tfrecords\".format(num=i))\n\nfor i in range(1, 15):\n    filenames_dev.append(common + \"Devel_DE_{num:02d}.tfrecords\".format(num=i))\n\nphase_train = tf.placeholder_with_default(True, shape=(), name='phase')\n\ntrain_batch_size = 5\ntest_batch_size = 5\n\ntf.set_random_seed(123)\nmseed = 123\n\n# this method is used within the model()...\ndef create_variables(name, shape, initializer=tf.contrib.layers.xavier_initializer(), weight_decay=0.0001):\n    '''\n    :param name: A string. The name of the new variable\n    :param shape: A list of dimensions\n    :param initializer: User Xavier as default.\n    :param is_fc_layer: Want to create fc layer variable? May use different weight_decay for fc\n    layers.\n    :return: The created variable\n    '''\n\n    ## TODO: to allow different weight decay to fully connected layer and conv layer\n    regularizer = tf.contrib.layers.l2_regularizer(scale=weight_decay)\n\n    new_variables = tf.get_variable(name, shape=shape, initializer=initializer,\n                                    regularizer=regularizer)\n    return new_variables\n\ndef model(inputs, n):\n    ....\n    # predictions shape: (batch_size, 3)\n    return predictions\n\n# loss function:\nsummaries_while_testing = []\nsummaries_while_training = []\n\ndef loss(predictions, labels, detected, name_scope, train_test):\n    # MSE\n    with tf.name_scope(name_scope):\n\n        MSE = tf.square(tf.subtract(predictions, labels))\n        MSE = tf.boolean_mask(MSE, detected)\n        MSE = tf.reduce_mean(MSE)\n\n        if train_test == 'Train':\n            reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\n            MSE += tf.reduce_sum(reg_losses)\n            loss_s = tf.summary.scalar('MSE', MSE)\n            summaries_while_training.append(loss_s)\n        else:\n            loss_s = tf.summary.scalar('MSE', MSE)\n            summaries_while_testing.append(loss_s)\n\n    return MSE\n\n# optimizer:\ndef optimize(mse):\n     with tf.name_scope('Optimizer'):\n        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n        with tf.control_dependencies(update_ops):\n            optimizer = tf.train.AdamOptimizer(learning_rate=0.001)\n\n            trainable_variables = tf.trainable_variables()\n\n            accum_vars = [tf.Variable(tf.zeros_like(single_tr_variable.value()), trainable=False)\n                          for single_tr_variable in trainable_variables]\n\n            # This is used as a rest mode between different training iterations...\n            zero_ops = [tv.assign(tf.zeros_like(tv)) for tv in accum_vars]\n\n            grads_vars = optimizer.compute_gradients(mse, trainable_variables)\n\n            accum_ops = [accum_vars[i].assign_add(gv[0]) for i, gv in enumerate(grads_vars) if gv[0] is not None]\n            train_step = optimizer.apply_gradients([(accum_vars[i], gv[1]) for i, gv in enumerate(grads_vars)])\n\n            return train_step, accum_ops, zero_ops\n\n# retrieve data section\ndef _parse_function(example_proto):\n\n    # The annotation contains the following features: timestamp; arousal; valence; liking\n    features = {\n        'height': tf.FixedLenFeature([], tf.int64),\n        'width': tf.FixedLenFeature([], tf.int64),\n        'image_raw': tf.FixedLenFeature([], tf.string),\n        'frame_number': tf.FixedLenFeature([1], tf.int64),\n        'detected': tf.FixedLenFeature([1], tf.int64),\n        'arousal': tf.FixedLenFeature([1], tf.float32),\n        'valence': tf.FixedLenFeature([1], tf.float32),\n        'liking': tf.FixedLenFeature([1], tf.float32)\n    }\n\n    parsed_features = tf.parse_single_example(example_proto, features)\n\n    # This is how we create one example, that is, extract one example from the database.\n    image = tf.decode_raw(parsed_features['image_raw'], tf.uint8)\n    # The height and the weights are used to\n    height = tf.cast(parsed_features['height'], tf.int32)\n    width = tf.cast(parsed_features['width'], tf.int32)\n\n    # The image is reshaped since when stored as a binary format, it is flattened. Therefore, we need the\n    # height and the weight to restore the original image back.\n    # Tensor(\"Reshape:0\", shape=(112, 112, 3), dtype=uint8)\n    image = tf.reshape(image, [112, 112, 3])\n\n    detected = parsed_features['detected']\n    arousal = parsed_features['arousal']\n    valence = parsed_features['valence']\n    liking = parsed_features['liking']\n\n    return detected, arousal, valence, liking, image\n\n###############################      TRAINING      ###################################\n\ndatasets_train_iterators = []\n\nfor file_name in filenames_train:\n    dataset_train = tf.data.TFRecordDataset(file_name).map(_parse_function).batch(train_batch_size)\n    datasets_train_iterators.append(dataset_train)\n\ndataset_train_all = tf.data.Dataset.zip(tuple(datasets_train_iterators))\niterator_train_all = dataset_train_all.make_initializable_iterator()\n\ndef retrieve_inputs_train():\n\n    next_batch = iterator_train_all.get_next()\n\n    detected = []\n    arousal = []\n    valence = []\n    liking = []\n    images = []\n\n    for n in next_batch:\n        detected.append(n[0])\n        arousal.append(n[1])\n        valence.append(n[2])\n        liking.append(n[3])\n        images.append(n[4])\n\n    detected = tf.concat(detected, axis=0)\n    arousal = tf.concat(arousal, axis=0)\n    valence = tf.concat(valence, axis=0)\n    liking = tf.concat(liking, axis=0)\n    images = tf.concat(images, axis=0)\n\n    return detected, arousal, valence, liking, images\n\n###############################      TESTING      ###################################\ndatasets_dev_iterators = []\n\nfor file_name in filenames_dev:\n    dataset_dev = tf.data.TFRecordDataset(file_name).map(_parse_function).batch(test_batch_size)\n    datasets_dev_iterators.append(dataset_dev)\n\ndataset_dev_all = tf.data.Dataset.zip(tuple(datasets_dev_iterators))\niterator_dev_all = dataset_dev_all.make_initializable_iterator()\n\ndef retrieve_inputs_dev():\n\n    next_batch = iterator_dev_all.get_next()\n\n    detected = []\n    arousal = []\n    valence = []\n    liking = []\n    images = []\n\n    for n in next_batch:\n        detected.append(n[0])\n        arousal.append(n[1])\n        valence.append(n[2])\n        liking.append(n[3])\n        images.append(n[4])\n\n    detected = tf.concat(detected, axis=0)\n    arousal = tf.concat(arousal, axis=0)\n    valence = tf.concat(valence, axis=0)\n    liking = tf.concat(liking, axis=0)\n    images = tf.concat(images, axis=0)\n\n    return detected, arousal, valence, liking, images\n\n# preparing model before training\ndetected, arousal, valence, liking, images = tf.cond(phase_train,\n                                                     lambda: retrieve_inputs_train(),\n                                                     lambda: retrieve_inputs_dev())\n\nimages_casted = tf.cast(images, tf.float32)\nwith tf.name_scope('image_normal'):\n    images_casted_normalized = tf.map_fn(lambda img: tf.image.per_image_standardization(img), images_casted)\n\n# shape of predictions: (680, 3) -&gt; 3 since we are outputing arousal, valence and liking\n# the n parameter is for Resnet configuration... Not important for now\npredictions = model(images_casted_normalized, n=[3, 4, 6, 3])\n\npredicted_arousal = tf.slice(predictions, begin=[0, 0], size=[-1, 1], name='predicted_arousal')\npredicted_valence = tf.slice(predictions, begin=[0, 1], size=[-1, 1], name='predicted_valence')\npredicted_liking = tf.slice(predictions, begin=[0, 2], size=[-1, 1], name='predicted_liking')\n\nMSE_a = tf.cond(phase_train, \n                lambda: loss(predicted_arousal, arousal, detected, 'MSE_arousal_Train', 'Train'),\n                lambda: loss(predicted_arousal, arousal, detected, 'MSE_arousal_Devel', 'Devel'))\n\nMSE_v = tf.cond(phase_train, \n                lambda: loss(predicted_valence, valence, detected, 'MSE_valence_Train', 'Train'),\n                lambda: loss(predicted_valence, valence, detected, 'MSE_valence_Devel', 'Devel'))\n\nMSE_l = tf.cond(phase_train,\n                lambda: loss(predicted_liking, liking, detected, 'MSE_liking_Train', 'Train'),\n                lambda: loss(predicted_liking, liking, detected, 'MSE_liking_Devel', 'Devel'))\n\nMSE = MSE_a + MSE_v + MSE_l\n\ntrain_step, accum_ops, zero_ops = optimize(MSE)\n\ninit_op = tf.global_variables_initializer()\n\nmodel_path = \"C:\/Users\/user\/Documents\/f24\/model\"\nevents_path = \"C:\/Users\/user\/Documents\/f24\/event_files\/34_layers\"\nwith tf.Session() as sess:\n\n    sess.run(init_op)\n\n    train_writer = tf.summary.FileWriter(events_path, sess.graph)\n\n    merged_train = tf.summary.merge(summaries_while_training)\n    merged_val = tf.summary.merge(summaries_while_testing)\n\n    sess.run(iterator_train_all.initializer)\n    sess.run(iterator_dev_all.initializer)\n<\/code><\/pre>\n\n<p>Finally, I am getting the following error:<\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value conv3_1\/conv2_in_block\/conv\n     [[Node: conv3_1\/conv2_in_block\/conv\/read = Identity[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](conv3_1\/conv2_in_block\/conv)]]\n\nDuring handling of the above exception, another exception occurred:\n\nFailedPreconditionError                   Traceback (most recent call last)\n&lt;ipython-input-11-dbe6d12c67ce&gt; in &lt;module&gt;()\n      7 \n      8     for v in accum_vars:\n----&gt; 9         sess.run(v.initializer)\n     10 \n     11     sess.run(init_op)\n<\/code><\/pre>\n\n<p>...<\/p>\n\n<pre><code>File \"&lt;ipython-input-10-8d7d7b4aa814&gt;\", line 10, in &lt;module&gt;\n    predictions = model(images_casted_normalized, n=[3, 4, 6, 3])\n  File \"&lt;ipython-input-5-fae307f9536f&gt;\", line 25, in model\n    conv3 = residual_block(layers[-1], 256, is_training=phase_train)\n  File \"&lt;ipython-input-4-d8a2d1403f18&gt;\", line 97, in residual_block\n    conv2 = bn_relu_conv_layer(conv1, [3, 3, output_channel, output_channel], 1, is_training=is_training)\n  File \"&lt;ipython-input-4-d8a2d1403f18&gt;\", line 61, in bn_relu_conv_layer\n    filter = create_variables(name='conv', shape=filter_shape)\n  File \"&lt;ipython-input-4-d8a2d1403f18&gt;\", line 15, in create_variables\n    regularizer=regularizer)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\variable_scope.py\", line 1317, in get_variable\n    constraint=constraint)\n<\/code><\/pre>\n\n<p>Now when I remove these 2 lines in the <code>optimize()<\/code>, my code works fine, but I know that this is wrong.<\/p>\n\n<pre><code>update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n    with tf.control_dependencies(update_ops):\n<\/code><\/pre>\n\n<p>Or, if I use the following code for the optimizer, my code runs fine.<\/p>\n\n<pre><code>def optimize(mse):\n     with tf.name_scope('Optimizer'):\n        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n        with tf.control_dependencies(update_ops):\n            optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\n            train_step = optimizer.minimize(mse)\n            return train_step\n<\/code><\/pre>\n\n<p>To me, this is weird and strange. I would love to know the reason for the error I am getting.<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[]}
+{"QuestionId":52706996,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm><rnn>","CreationDate":"2018-10-08T17:02:32.603","AcceptedAnswerId":null,"OwnerUserId":4616785.0,"Title":"Keras LSTM Input Dimension understanding each other","Body":"<p>but I have been trying to play around with it for awhile. I've seen a lot of guides on how Keras is used to build LSTM models and how people feed in the inputs and get expected outputs. But what I have never seen yet is, for example stock data, how we can make the LSTM model understand patterns between different dimensions, say close price is much higher than normal because volume is low.<\/p>\n\n<p>Point of this is that I want to do a test with stock prediction, but make it so that each dimensions are not reliant on previous time steps, but also reliant on other dimensions it haves as well.<\/p>\n\n<p>Sorry if I am not asking the question correctly, please ask more questions if I am not explaining it clearly.<\/p>\n","answers":[{"AnswerId":"52786576","CreationDate":"2018-10-12T20:26:04.793","ParentId":null,"OwnerUserId":"10491777","Title":null,"Body":"<p>First: Regressors will replicate if you input a feature that gives some direct intuition about the predicted input might be to secure the error is minimized, rather than trying to actually predict it. Try to focus on binary classification or multiclass classification, whether the closing price go up\/down or how much. <\/p>\n\n<p>Second: Always engineer the raw features to give more explicit patterns to the ML algorithm. Think on inputs as Volume(t) - Volume(t-1), close(t)^2 - close(t-1)^2, technical indicators(RSI, CCI, OBV etc.) Create your own features. You can use the pyti library for technical indicators. <\/p>\n"}]}
+{"QuestionId":52707600,"AnswerCount":2,"Tags":"<tensorflow><keras><lstm><lm><stock>","CreationDate":"2018-10-08T17:47:58.947","AcceptedAnswerId":52709422.0,"OwnerUserId":10474207.0,"Title":"Why the LSTM can't reduce the loss","Body":"<p>i'm using lstm to train a model to predict the stock price, i used the many to one mode like this:<\/p>\n\n<pre><code>                                    O \n                                    | \n                              O O O O \n                              | | | | \n                              O O O O \n<\/code><\/pre>\n\n<p>and i used keras framework to build the network, but it seems the NN can't be build up easily...<\/p>\n\n<p>here is my lstm NN source code of python:<\/p>\n\n<pre><code>def lstm_rls(num_in,num_out=1, batch_size=128, step=1,dim=1):\n    model = Sequential()\n\n    model.add(LSTM(\n        1024,\n        input_shape=(step, num_in),\n        return_sequences=True))\n    model.add(Dropout(0.2))\n\n    model.add(LSTM(\n        512,\n        return_sequences=False))\n    model.add(Dropout(0.2))\n\n    model.add(Dense(1))\n    model.add(Activation('linear'))\n    model.compile(loss='mse', optimizer='rmsprop')\n    return model\n<\/code><\/pre>\n\n<p>i trained about 512 epochs, but the loss is always around the 7<\/p>\n\n<blockquote>\n  <p>Epoch 1\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 4s 978us\/step - loss: 48.6274<\/p>\n  \n  <p>Epoch 2\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 220us\/step - loss: 11.1913<\/p>\n  \n  <p>Epoch 3\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 221us\/step - loss: 6.8059<\/p>\n  \n  <p>Epoch 4\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 220us\/step - loss: 6.7905<\/p>\n  \n  <p>Epoch 5\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 221us\/step - loss: 6.8151<\/p>\n  \n  <p>Epoch 6\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 219us\/step - loss: 6.7907<\/p>\n  \n  <p>Epoch 7\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 220us\/step - loss: 6.8060<\/p>\n  \n  <p>Epoch 8\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 221us\/step - loss: 6.7824<\/p>\n<\/blockquote>\n\n<p>...\n...\n...<\/p>\n\n<blockquote>\n  <p>Epoch 509\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 222us\/step - loss: 6.7807<\/p>\n  \n  <p>Epoch 510\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 223us\/step - loss: 6.8199<\/p>\n  \n  <p>Epoch 511\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 222us\/step - loss: 6.7726<\/p>\n  \n  <p>Epoch 512\/512<\/p>\n  \n  <p>3968\/3968 [==============================] - 1s 222us\/step - loss: 6.7715<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52709422","CreationDate":"2018-10-08T20:09:07.933","ParentId":null,"OwnerUserId":"8079690","Title":null,"Body":"<p>I have had your problem for image captioning(I had four Gpu tesla k20). It take long time to train the LSTM(about several day). I think your problem is natural. One way that I used was increasing batch size. In other side, increasing of batch size is relative with performance and memory usage.    <\/p>\n"},{"AnswerId":"52747008","CreationDate":"2018-10-10T18:55:01.557","ParentId":null,"OwnerUserId":"8079690","Title":null,"Body":"<p>Today I watched one lecture about debugging tips. It said that one way that you can analysis your code is:\nVerify that object function decreases on small training set, not all at first time. <\/p>\n"}]}
+{"QuestionId":52708764,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-08T19:19:26.853","AcceptedAnswerId":null,"OwnerUserId":10171009.0,"Title":"Applying Boundary Conditions to Keras Output","Body":"<p>I have a periodic value (an angle) with period = <code>pi<\/code> (i.e. <code>pi+a<\/code> = <code>a<\/code>). I am training a neural network in Keras to output this value, given a set of inputs. The network performs very well, except for at the boundaries <code>0<\/code> and <code>pi<\/code> (if the truth value is <code>0<\/code> or <code>pi<\/code>, its prediction is random). This can be seen in <a href=\"https:\/\/i.stack.imgur.com\/z2SUT.png\" rel=\"nofollow noreferrer\">this histogram<\/a>. Does anyone know why that is? Can I somehow include the boundaries in the training or tell it that if it predicts <code>pi+a<\/code>, when the truth value is <code>a<\/code>, then it predicted correctly and doesn't need to update its weights and biases? Thanks.<\/p>\n","answers":[{"AnswerId":"52710032","CreationDate":"2018-10-08T20:59:43.610","ParentId":null,"OwnerUserId":"2084944","Title":null,"Body":"<p>Neuronal networks do not perform well on a task where you have a non-continuous behavior in your data. Especially on The angular data like yours. To circumvent your problem, you can first do a trigonometric transformation of your target value. This is actually nothing but applying a kernel to your data. Let's say your periodic output data is <code>y=0...\u03c0<\/code>. Then your new target is a bi-valued vector <code>y_hat=[sin(2y), cos(2y)]<\/code>. Let your network train against <code>y_hat<\/code>, while you perform the back transformation <code>y=.5 atan2(y_hat)<\/code> during inference to get your original value.<\/p>\n"}]}
+{"QuestionId":52708821,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-08T19:23:48.577","AcceptedAnswerId":null,"OwnerUserId":9633432.0,"Title":"Binary Classification using softmax LSTM","Body":"<p>I am trying to design bi-directional LSTM using word2vec features for binary classification problem.<\/p>\n\n<pre><code>my_model=Sequential()\n\nmy_model.add(Embedding(words,10,input_length=trainDataVecs.shape[1],weights=[embedding_matrix],trainable=True))\nmy_model.add(Bidirectional(LSTM(20,activation='tanh',init='glorot_uniform',recurrent_dropout = 0.2, dropout = 0.2)))\n<\/code><\/pre>\n\n<p>Since there are two classes,<\/p>\n\n<pre><code>my_model.add(Dense(2, activation='softmax'))\n\ndef auc(y_true, y_pred):\n    auc = tf.metrics.auc(y_true, y_pred)[1]\n    K.get_session().run(tf.local_variables_initializer())\n    return auc\n<\/code><\/pre>\n\n<p>Using AUC as the metrics<\/p>\n\n<pre><code>print \"Compiling...\"\noptimizer=RMSprop(lr=0.0001, rho=0.9, epsilon=1e-08)\nmy_model.compile(optimizer=optimizer,\n               loss='categorical_crossentropy',\n                  metrics=[auc])\n\nmy_model.fit(trainDataVecs, Y_train, shuffle = True, batch_size = 10, epochs=20)\n<\/code><\/pre>\n\n<p>But I am getting : <code>ValueError: Error when checking target: expected dense_2 to have shape (2,) but got array with shape (1,)<\/code> as <code>Y_train<\/code> is 1-D.\nI can use <code>sigmoid<\/code> instead of <code>softmax<\/code> but that will give me the probability when I will predict. My labels are <code>0<\/code> or <code>1<\/code>. Hence, I want to see the F-score between the predicted and the real values.<\/p>\n","answers":[]}
+{"QuestionId":52709073,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T19:43:29.990","AcceptedAnswerId":52718151.0,"OwnerUserId":10474989.0,"Title":"How to vectorize indexing operation in tensorflow","Body":"<p>I have a tensor A of shape (2, 4, 2), and a tensor B of shape (4, 4), all the values are int. Entries in A are from 0 to 3.<\/p>\n\n<p>I want to create a tensor C of shape(2, 4, 2). <\/p>\n\n<p>The for loop code is like:<\/p>\n\n<pre><code>\n\nfor i in range(2):\n    for j in range(2):\n        for k in range(4):\n            C[i][k][j] = B[k][A[i][k][j]]\n\n<\/code><\/pre>\n\n<p>How can I create such tensor C in tensorflow?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"52718151","CreationDate":"2018-10-09T09:54:27.880","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is how you can do it with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input values\nA = tf.placeholder(tf.int32, [None, None, None])\nB = tf.placeholder(tf.int32, [None, None])\n# Make indices for first dimension of B\nidx = tf.range(tf.shape(B)[0], dtype=A.dtype)[tf.newaxis, :,  tf.newaxis]\n# Tile first dimension indices to match the size of A\nidx = tf.tile(idx, (tf.shape(A)[0], 1, tf.shape(A)[2]))\n# Stack first dimension indices with A to complete index tensor\nidx = tf.stack([idx, A], axis=-1)\n# Make result gathering from B\nC = tf.gather_nd(B, idx)\n<\/code><\/pre>\n\n<p>Here is an example, testing that the result matches your code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Non-TensorFlow implementation for result comparison\nA_value = np.random.randint(0, 4, size=(2, 4, 2))\nB_value = np.random.randint(100, size=(4, 4))\nC_value = np.empty(A_value.shape, dtype=B_value.dtype)\nfor i in range(A_value.shape[0]):\n    for j in range(A_value.shape[2]):\n        for k in range(A_value.shape[1]):\n            C_value[i][k][j] = B_value[k][A_value[i][k][j]]\n\n# TensorFlow implementation\nA = tf.placeholder(tf.int32, [None, None, None])\nB = tf.placeholder(tf.int32, [None, None])\nidx = tf.range(tf.shape(B)[0], dtype=A.dtype)[tf.newaxis, :,  tf.newaxis]\nidx = tf.tile(idx, (tf.shape(A)[0], 1, tf.shape(A)[2]))\nidx = tf.stack([idx, A], axis=-1)\nC = tf.gather_nd(B, idx)\n\n# Check result\nwith tf.Session() as sess:\n    C_value_tf = sess.run(C, feed_dict={A: A_value, B: B_value})\n    print(np.all(np.equal(C_value_tf, C_value)))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>True\n<\/code><\/pre>\n"}]}
+{"QuestionId":52709701,"AnswerCount":0,"Tags":"<machine-learning><keras><deep-learning><classification><loss-function>","CreationDate":"2018-10-08T20:32:03.763","AcceptedAnswerId":null,"OwnerUserId":2129212.0,"Title":"Penalizing false negatives in a binary classification problem","Body":"<p>I am trying to penalize the False Negatives more than False Positives in a binary classification problem.<\/p>\n\n<p>The custom loss function looks like: inspired from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2115#issuecomment-204060456\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<pre><code>def w_categorical_crossentropy(y_true, y_pred, weights):\n    nb_cl = len(weights)\n    final_mask = K.zeros_like(y_pred[:, 0])\n    y_pred_max = K.max(y_pred, axis=1)\n    y_pred_max = K.reshape(y_pred_max, (K.shape(y_pred)[0], 1))\n    y_pred_max_mat = K.cast(K.equal(y_pred, y_pred_max), K.floatx())\n    for c_p, c_t in product(range(nb_cl), range(nb_cl)):\n        final_mask += (weights[c_t, c_p] * y_pred_max_mat[:, c_p] * y_true[:, c_t])\n    return K.categorical_crossentropy(y_pred, y_true) * final_mask\n<\/code><\/pre>\n\n<p>Following are the weights being passed to the weights param in the above function:<\/p>\n\n<pre><code>w_array = np.ones((2,2))\nw_array[1,0] = 2.5 # penalizing FN\nw_array[0,1] = 2.5 # penalizing FP\n<\/code><\/pre>\n\n<p>In my understanding, first line is penalizing the FN and the second one is penalizing the FP. When I try to run this code, the number of FN\/FP is pretty much the same as in the case with equal weights as shown below<\/p>\n\n<pre><code>w_array = np.ones((2,2))\n#w_array[1,0] = 2.5 # penalizing FN\n#w_array[0,1] = 2.5 # penalizing FP\n<\/code><\/pre>\n\n<p>FOLLOW UP: If i simply comment out the second assignment in w_array, does it mean I am only penalizing the FN and not the FP. What is the importance of <code>final_mask<\/code> here?<\/p>\n\n<p>Function call and usage:<\/p>\n\n<pre><code>loss = lambda y_true, y_pred: w_categorical_crossentropy(y_true, y_pred, weights=w_array)\n\nclassifier.compile(optimizer = sgd, loss = loss, metrics = ['accuracy'])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52710348,"AnswerCount":3,"Tags":"<python><tensorflow><deep-learning><pytorch><google-colaboratory>","CreationDate":"2018-10-08T21:27:50.553","AcceptedAnswerId":null,"OwnerUserId":5157633.0,"Title":"How to take and restore snapshots of model training on another VM in Google Colab?","Body":"<p>There is a 12 hour <a href=\"https:\/\/www.kaggle.com\/c\/jigsaw-toxic-comment-classification-challenge\/discussion\/48320\" rel=\"nofollow noreferrer\">time limit<\/a> for training DL models on GPU, according to google colab. Other people have had similar <a href=\"https:\/\/stackoverflow.com\/questions\/49469697\/how-can-i-keep-the-gpu-running-for-more-than-12-hours-in-google-colaboratory?rq=1\">questions<\/a> in the past, but there has been no clear answer on how to save and load models halfway through training when the 12 hour limits get exceeeded, including saving the number of epochs that has been completed\/saving other parameters. Is there an automated script for me to save the relevant parameters and resume operations on another VM? I am a complete noob; clear cut answers will be much appreciated. <\/p>\n","answers":[{"AnswerId":"52710834","CreationDate":"2018-10-08T22:20:04.097","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>Mount Drive, and save and load persistent data from there.<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/gdrive')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/colab.research.google.com\/notebooks\/io.ipynb#scrollTo=RWSJpsyKqHjH\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/notebooks\/io.ipynb#scrollTo=RWSJpsyKqHjH<\/a><\/p>\n"},{"AnswerId":"52719834","CreationDate":"2018-10-09T11:21:18.860","ParentId":null,"OwnerUserId":"8337524","Title":null,"Body":"<p>As far as I know, there is no way to automatically reconnect to another VM whenever you reach the 12 hours limit. So in any case, you have to manually reconnect when the time is up.<\/p>\n\n<p>As Bob Smith points out, you can mount Google Drive in Colab VM so that you can save and load data from there. In particular, you can periodically save model checkpoints so that you can load the most recent one whenever you connect to a new Colab VM.<\/p>\n\n<ol>\n<li><p>Mount Drive in your Colab VM:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/gdrive')\n<\/code><\/pre><\/li>\n<li><p>Create a <code>saver<\/code> in your graph:<\/p>\n\n<pre><code>saver = tf.train.Saver()\n<\/code><\/pre><\/li>\n<li><p>Periodically (e.g. every epoch) save a checkpoint in Drive:<\/p>\n\n<pre><code>saver.save(session, CHECKPOINT_PATH)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>When you connect to a new Colab VM (because of the timeout), mount Drive again in your VM and restore the most recent checkpoint before the training phase:<\/p>\n\n<pre><code>saver.restore(session, CHECKPOINT_PATH)\n...\n# Start training with the restored model.\n<\/code><\/pre>\n\n<p>Take a look at the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">documentation<\/a> to read more about <code>tf.train.Saver<\/code>.<\/p>\n"},{"AnswerId":"52733404","CreationDate":"2018-10-10T05:50:54.573","ParentId":null,"OwnerUserId":"667883","Title":null,"Body":"<p>From colab you can access github, which makes it possible to save your model checkpoints to github periodically. When a session is ended, you can start another session and load the checkpoint back from your github repo.<\/p>\n"}]}
+{"QuestionId":52710671,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><recurrent-neural-network>","CreationDate":"2018-10-08T22:00:53.777","AcceptedAnswerId":52712165.0,"OwnerUserId":7184403.0,"Title":"Why there will be a baseline for LSTM and how to improve the performance?","Body":"<p>I have a LSTM network with the following configuration:<\/p>\n\n<pre><code>model3 = tf.keras.Sequential()\nmodel3.add(tf.keras.layers.LSTM(15, input_shape=(1, 10), return_sequences=True)) \nmodel3.add(tf.keras.layers.LSTM(5))\nmodel3.add(tf.keras.layers.Dense(1, activation=None))\nmodel3.compile(loss='mean_squared_error', optimizer='adam')\nmodel3.fit(trainX, trainY, epochs=100, batch_size=30)\n<\/code><\/pre>\n\n<p>The result can predict the tendency but there will be a stable baseline, which is abnormal. The figure is like the following: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iAEZN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iAEZN.png\" alt=\"prediction results\"><\/a><\/p>\n\n<p>So I wonder why there is a stable line even if there is no activation function in my last layer. And also how to improve performance. <\/p>\n","answers":[{"AnswerId":"52712165","CreationDate":"2018-10-09T01:37:19.393","ParentId":null,"OwnerUserId":"10475762","Title":null,"Body":"<p>You are using two quite small LSTM layers to predict for what I assume is limited data. The relationship learned between the input and output will therefore be quite basic (as opposed to a deep model). <\/p>\n\n<p>In regards to your question, you must analyze the outputs according to your cost function. Using a cost function such as MSE will penalize the model for having large discrepancies between the output and true values, and it appears that what the model has learned here is that the value (around 900) is the value which minimizes MSE using the small amount of parameters available. So is this \"abnormal\" as you described? I would say it was abnormal if the line did not appear to almost be a line of best fit throughout the lower values of the data.<\/p>\n\n<p>What I would suggest to you is using a simpler network if your training data is not sufficient (see possibly CNNs, or windowed ANNs), or seeing if it is possible to obtain more training data. <\/p>\n"}]}
+{"QuestionId":52710878,"AnswerCount":1,"Tags":"<python><deep-learning><conda><pytorch>","CreationDate":"2018-10-08T22:25:38.073","AcceptedAnswerId":null,"OwnerUserId":3899975.0,"Title":"how to install pytorch and fastai with conda","Body":"<p>I created conda environment and install pytorch and fastai (Mac OS Mojave) as below:<\/p>\n\n<pre><code>conda create -n fai_course python=3.7\nsource activate fai_course\nconda install -c pytorch pytorch-nightly-cpu\nconda install -c fastai torchvision-nightly-cpu \njupyter notebook\n<\/code><\/pre>\n\n<p>When I import a package from jupyter notebook, I get OSError as shown below:<\/p>\n\n<pre><code>from fastai.imports import *\n<\/code><\/pre>\n\n<p>-<\/p>\n\n<pre><code>--------------------------------------------------------------------------\nFileNotFoundError                         Traceback (most recent call last)\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in &lt;module&gt;\n   1352         try:\n-&gt; 1353             fontManager = json_load(_fmcache)\n   1354             if (not hasattr(fontManager, '_version') or\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in json_load(filename)\n    887     \"\"\"\n--&gt; 888     with open(filename, 'r') as fh:\n    889         return json.load(fh, object_hook=_json_decode)\n\nFileNotFoundError: [Errno 2] No such file or directory: '\/Users\/user\/.matplotlib\/fontlist-v300.json'\n\nDuring handling of the above exception, another exception occurred:\n\nOSError                                   Traceback (most recent call last)\n&lt;ipython-input-5-9f9378ae0f2a&gt; in &lt;module&gt;\n----&gt; 1 from fastai.imports import *\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/fastai\/__init__.py in &lt;module&gt;\n----&gt; 1 from .basic_train import *\n      2 from .callback import *\n      3 from .callbacks import *\n      4 from .core import *\n      5 from .data import *\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/fastai\/basic_train.py in &lt;module&gt;\n      1 \"Provides basic training and validation with `Learner`\"\n----&gt; 2 from .torch_core import *\n      3 from .data import *\n      4 from .callback import *\n      5 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/fastai\/torch_core.py in &lt;module&gt;\n      1 \"Utility functions to help deal with tensors\"\n----&gt; 2 from .imports.torch import *\n      3 from .core import *\n      4 \n      5 AffineMatrix = Tensor\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/fastai\/imports\/__init__.py in &lt;module&gt;\n----&gt; 1 from .core import *\n      2 from .torch import *\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/fastai\/imports\/core.py in &lt;module&gt;\n      1 import csv, gc, gzip, os, pickle, shutil, sys, warnings\n----&gt; 2 import math, matplotlib.pyplot as plt, numpy as np, pandas as pd, random\n      3 import scipy.stats, scipy.special\n      4 import abc, collections, hashlib, itertools, json, operator\n      5 import mimetypes, inspect, typing, functools\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/pyplot.py in &lt;module&gt;\n     30 from cycler import cycler\n     31 import matplotlib\n---&gt; 32 import matplotlib.colorbar\n     33 import matplotlib.image\n     34 from matplotlib import rcsetup, style\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/colorbar.py in &lt;module&gt;\n     30 import matplotlib.collections as collections\n     31 import matplotlib.colors as colors\n---&gt; 32 import matplotlib.contour as contour\n     33 import matplotlib.cm as cm\n     34 import matplotlib.gridspec as gridspec\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/contour.py in &lt;module&gt;\n     16 import matplotlib.colors as mcolors\n     17 import matplotlib.collections as mcoll\n---&gt; 18 import matplotlib.font_manager as font_manager\n     19 import matplotlib.text as text\n     20 import matplotlib.cbook as cbook\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in &lt;module&gt;\n   1361             raise\n   1362         except Exception:\n-&gt; 1363             _rebuild()\n   1364     else:\n   1365         _rebuild()\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in _rebuild()\n   1342         global fontManager\n   1343 \n-&gt; 1344         fontManager = FontManager()\n   1345 \n   1346         if _fmcache:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in __init__(self, size, weight)\n    976         self.defaultFont = {}\n    977 \n--&gt; 978         ttffiles = findSystemFonts(paths) + findSystemFonts()\n    979         self.defaultFont['ttf'] = next(\n    980             (fname for fname in ttffiles\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in findSystemFonts(fontpaths, fontext)\n    268             # check for OS X &amp; load its fonts if present\n    269             if sys.platform == 'darwin':\n--&gt; 270                 fontfiles.update(OSXInstalledFonts(fontext=fontext))\n    271 \n    272     elif isinstance(fontpaths, str):\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in OSXInstalledFonts(directories, fontext)\n    216         directories = OSXFontDirectories\n    217     return [path\n--&gt; 218             for directory in directories\n    219             for ext in get_fontext_synonyms(fontext)\n    220             for path in list_fonts(directory, ext)]\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in &lt;listcomp&gt;(.0)\n    218             for directory in directories\n    219             for ext in get_fontext_synonyms(fontext)\n--&gt; 220             for path in list_fonts(directory, ext)]\n    221 \n    222 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in list_fonts(directory, extensions)\n    155     extensions = [\".\" + ext for ext in extensions]\n    156     return [str(path)\n--&gt; 157             for path in filter(Path.is_file, Path(directory).glob(\"**\/*.*\"))\n    158             if path.suffix in extensions]\n    159 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/matplotlib\/font_manager.py in &lt;listcomp&gt;(.0)\n    154     \"\"\"\n    155     extensions = [\".\" + ext for ext in extensions]\n--&gt; 156     return [str(path)\n    157             for path in filter(Path.is_file, Path(directory).glob(\"**\/*.*\"))\n    158             if path.suffix in extensions]\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in glob(self, pattern)\n   1080             raise NotImplementedError(\"Non-relative patterns are unsupported\")\n   1081         selector = _make_selector(tuple(pattern_parts))\n-&gt; 1082         for p in selector.select_from(self):\n   1083             yield p\n   1084 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _select_from(self, parent_path, is_dir, exists, scandir)\n    541             try:\n    542                 successor_select = self.successor._select_from\n--&gt; 543                 for starting_point in self._iterate_directories(parent_path, is_dir, scandir):\n    544                     for p in successor_select(starting_point, is_dir, exists, scandir):\n    545                         if p not in yielded:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n--&gt; 533                     for p in self._iterate_directories(path, is_dir, scandir):\n    534                         yield p\n    535         except PermissionError:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/pathlib.py in _iterate_directories(self, parent_path, is_dir, scandir)\n    529             entries = list(scandir(parent_path))\n    530             for entry in entries:\n--&gt; 531                 if entry.is_dir() and not entry.is_symlink():\n    532                     path = parent_path._make_child_relpath(entry.name)\n    533                     for p in self._iterate_directories(path, is_dir, scandir):\n\nOSError: [Errno 62] Too many levels of symbolic links: '.Trash\/NETGEARGenie.app\/Contents\/Frameworks\/QtPrintSupport.framework\/Versions\/5\/5'\n<\/code><\/pre>\n\n<p>Can you please let me know what am I missing?\nThanks<\/p>\n","answers":[{"AnswerId":"52722963","CreationDate":"2018-10-09T14:02:32.533","ParentId":null,"OwnerUserId":"3438366","Title":null,"Body":"<p>Admittedly, the error is very confusing. Looking at the library paths, however:<\/p>\n\n<p>\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/<\/p>\n\n<p>I would guess that you aren't using Jupyter installed through conda, but rather another local install.<\/p>\n\n<p>Try checking which Jupyter you're using:<\/p>\n\n<pre><code>    which jupyter\n<\/code><\/pre>\n\n<p>This should be something like:<\/p>\n\n<p>~\/anaconda3\/bin\/jupyter<\/p>\n\n<p>If it's not, you're using a different jupyter installation which doesn't appear to have those libraries installed.<\/p>\n"}]}
+{"QuestionId":52710931,"AnswerCount":1,"Tags":"<opencv><caffe><ros>","CreationDate":"2018-10-08T22:34:06.090","AcceptedAnswerId":null,"OwnerUserId":10475556.0,"Title":"Having issues with opencv\/dnn with caffe model","Body":"<p>I am trying to make use of openpose example in opencv using caffe model and opencv\/dnn.hpp<\/p>\n\n<p>tutorial I have been following - <a href=\"https:\/\/www.learnopencv.com\/deep-learning-based-human-pose-estimation-using-opencv-cpp-python\/\" rel=\"nofollow noreferrer\">https:\/\/www.learnopencv.com\/deep-learning-based-human-pose-estimation-using-opencv-cpp-python\/<\/a><\/p>\n\n<p>we require 2 files for the network as said in the tutorial : \n1 - prototxt - <a href=\"https:\/\/github.com\/spmallick\/learnopencv\/blob\/master\/OpenPose\/pose\/coco\/pose_deploy_linevec.prototxt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/spmallick\/learnopencv\/blob\/master\/OpenPose\/pose\/coco\/pose_deploy_linevec.prototxt<\/a><\/p>\n\n<p>2 - caffe model - posefs1.perception.cs.cmu.edu\/OpenPose\/models\/pose\/coco\/pose_iter_440000.caffemodel<\/p>\n\n<p>ros node that I made following the tutorial : <\/p>\n\n<pre><code>#include &lt;opencv2\/imgproc\/imgproc.hpp&gt;\n#include &lt;opencv2\/highgui\/highgui.hpp&gt;\n#include &lt;opencv2\/dnn\/dnn.hpp&gt;\n#include &lt;sensor_msgs\/image_encodings.h&gt;\n#include &lt;ros\/ros.h&gt;\n#include &lt;image_transport\/image_transport.h&gt;\n#include &lt;cv_bridge\/cv_bridge.h&gt;\n#include &lt;iostream&gt;\n\n\n using namespace std;\n using namespace cv;\n using namespace cv::dnn;\n\n  static const std::string OPENCV_WINDOW = \"Image window\";\n\n\n  #define COCO\n\n  #ifdef COCO\n  const int POSE_PAIRS[17][2] = \n  {   \n            {1,2}, {1,5}, {2,3},\n             {3,4}, {5,6}, {6,7},\n             {1,8}, {8,9}, {9,10},\n             {1,11}, {11,12}, {12,13},\n             {1,0},{0,14},\n             {14,16}, {0,15}, {15,17}\n     };\n\nstatic const std::string protoFile = \"pose\/coco\/pose_deploy_linevec.prototxt\";\nstatic const std::string weightsFile = \"pose\/coco\/pose_iter_440000.caffemodel\"; \n\n\n        int nPoints = 18;\n        #endif\n\n         class ImageConverter\n      {\n                ros::NodeHandle nh_;\n                image_transport::ImageTransport it_;\n                image_transport::Subscriber image_sub_;\n\n        public:\n                        ImageConverter()\n                                             : it_(nh_)\n     {\n\n            image_sub_ = it_.subscribe(\"\/zed\/rgb\/image_raw_color\", 1, &amp;ImageConverter::imageCb, this);\n      }\n\n\n     ~ImageConverter()\n     {\n           cv::destroyWindow(OPENCV_WINDOW);\n       }\n\n\n\nvoid imageCb(const sensor_msgs::ImageConstPtr&amp; msg)\n{\n     cv_bridge::CvImagePtr cv_ptr;\n     try\n     {\n     cv_ptr = cv_bridge::toCvCopy(msg, sensor_msgs::image_encodings::BGR8);\n     }  \n     catch (cv_bridge::Exception&amp; e)\n     {\n     ROS_ERROR(\"cv_bridge exception: %s\", e.what());\n     return;\n     }\n\n\nif (cv_ptr-&gt;image.rows &gt; 60 &amp;&amp; cv_ptr-&gt;image.cols &gt; 60)\n\ndetect_people(cv_ptr-&gt;image); \n\ncv::waitKey(3);\n\n\n\n }\n\n\n\n\n\nvoid detect_people(cv::Mat msg)\n{\nint inWidth = msg.cols;\nint inHeight = msg.rows;\nfloat thresh = 0.1;   \n\ncv::Mat frame;\nmsg.copyTo(frame);\ncv::Mat frameCopy = frame.clone();\nint frameWidth = frame.cols;\nint frameHeight = frame.rows;\n\ncv::dnn::Net net = cv::dnn::readNetFromCaffe(\"pose_deploy_linevec.prototxt\" ,\"pose_iter_440000.caffemodel\");\n\ncv::Mat inpBlob = blobFromImage(frame, 1.0\/255, cv::Size(inWidth, inHeight), cv::Scalar(0, 0, 0), false, false);  \n\nnet.setInput(inpBlob);\n\ncv::Mat output = net.forward();\n\nint H = output.size[2];\nint W = output.size[3];\n\nstd::vector&lt;cv::Point&gt; points(nPoints);\n\nfor (int n=0; n &lt; nPoints; n++)\n{\n    \/\/ Probability map of corresponding body's part.\n    cv::Mat probMap(H, W, CV_32F, output.ptr(0,n));\n\n    cv::Point2f p(-1,-1);\n    cv::Point maxLoc;\n    double prob;\n    cv::minMaxLoc(probMap, 0, &amp;prob, 0, &amp;maxLoc);\n    if (prob &gt; thresh)\n    {\n        p = maxLoc;\n        p.x *= (float)frameWidth \/ W ;\n        p.y *= (float)frameHeight \/ H ;\n\n        cv::circle(frameCopy, cv::Point((int)p.x, (int)p.y), 8, Scalar(0,255,255), -1);\n        cv::putText(frameCopy, cv::format(\"%d\", n), cv::Point((int)p.x, (int)p.y), cv::FONT_HERSHEY_COMPLEX, 1,      cv::Scalar(0, 0, 255), 2);\n\n    }\n    points[n] = p;\n}\n\n\n\nint nPairs = sizeof(POSE_PAIRS)\/sizeof(POSE_PAIRS[0]);\n\nfor (int n = 0; n &lt; nPairs; n++)\n{\n    \/\/ lookup 2 connected body\/hand parts\n    Point2f partA = points[POSE_PAIRS[n][0]];\n    Point2f partB = points[POSE_PAIRS[n][1]];\n\n    if (partA.x&lt;=0 || partA.y&lt;=0 || partB.x&lt;=0 || partB.y&lt;=0)\n        continue;\n\n    cv::line(frame, partA, partB, cv::Scalar(0,255,255), 8);\n    cv::circle(frame, partA, 8, cv::Scalar(0,0,255), -1);\n    cv::circle(frame, partB, 8, cv::Scalar(0,0,255), -1);\n}\n\n   cv::imshow(\"Output-Skeleton\", frame);\n\n\n   }\n\n   };\n\n     int main(int argc, char** argv)\n     {\n     ros::init(argc, argv, \"image_converter\");\n     ros::NodeHandle nh_;\n     ros::Publisher pub; \n     ImageConverter ic;\n     ros::spin();\n     return 0;\n    }\n<\/code><\/pre>\n\n<p>The code is compiled without any errors, but while I run the code it gives the following error msg : <\/p>\n\n<p>I get the following error when I run the node\nerror -  OpenCV Error: Unspecified error (FAILED: fs.is_open(). Can't open \"pose_deploy_linevec.prototxt\") in ReadProtoFromTextFile, file \/tmp\/binarydeb\/ros-kinetic-opencv3-3.3.1\/modules\/dnn\/src\/caffe\/caffe_io.cpp, line 1119\nterminate called after throwing an instance of 'cv::Exception'\n  what():  \/tmp\/binarydeb\/ros-kinetic-opencv3-3.3.1\/modules\/dnn\/src\/caffe\/caffe_io.cpp:1119: error: (-2) FAILED: fs.is_open(). Can't open \"pose_deploy_linevec.prototxt\" in function ReadProtoFromTextFile<\/p>\n\n<p>Aborted (core dumped)<\/p>\n\n<p>please help me solve this issue. <\/p>\n","answers":[{"AnswerId":"56860478","CreationDate":"2019-07-02T21:54:13.810","ParentId":null,"OwnerUserId":"8267304","Title":null,"Body":"<p>This issue is probably with the windows users only.\nSolve the issue by:<\/p>\n\n<ul>\n<li>Using\/Adding absolute path when calling prototxt file.<\/li>\n<li><p>Add the extension too. For Example: <\/p>\n\n<blockquote>\n  <p>\"pose\/coco\/pose_deploy_linevec.prototxt.txt\"<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p><em>Spent 3 hours debugging this myself. Hope it helped someone else.<\/em><\/p>\n"}]}
+{"QuestionId":52710975,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-08T22:38:36.483","AcceptedAnswerId":null,"OwnerUserId":4138044.0,"Title":"How do I accept a list of tensors in a placeholder in tensorflow?","Body":"<p>I want to do the following:<\/p>\n\n<pre><code>out = sess.run(operation, feed_dict={x: x_value, , y: y_value})\n<\/code><\/pre>\n\n<p>Somewhere in the model definition, I have used this:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, None, 3))\ny = tf.placeholder(tf.float32, shape=(None, None, 3))\noperation = some_function(x, y)\n<\/code><\/pre>\n\n<p>where <code>x<\/code> and <code>y<\/code> are images. How do I rewrite my code so that I can evaluate multiple <code>y<\/code>s with a single <code>x<\/code>. I want to do something like this:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, None, 3))\ny = []\nx = []\nfor _y in y:\n    operation = some_function(x, _y)\n    z.append(operation)\n<\/code><\/pre>\n\n<p>Right now I just do this:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, None, 3))\ny = tf.placeholder(tf.float32, shape=(None, None, None, 3)) #add extra dimension\nz = tf.unstack(tf.map_fn(some_function, y))\n<\/code><\/pre>\n\n<p>But the problem is when I do that, <code>tf.unpack<\/code> throws an error says its not aware of the size of y and fails. <\/p>\n\n<p>This is solved by doing this, where I initialize <code>y<\/code>'s 0 index with the total number of images. <\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, None, 3))\ny = tf.placeholder(tf.float32, shape=(2, None, None, 3)) #add extra dimension\nz = tf.unstack(tf.map_fn(some_function, y))\n<\/code><\/pre>\n\n<p>and I eval the values using:<\/p>\n\n<pre><code>y_values = np.stack([y_value1, y_value2])\nout = sess.run(operation, feed_dict={x: x_value, , y: y_values })\n<\/code><\/pre>\n\n<p>Ideally I want this to be variable and me not having to initialize the value of the placeholder. Lets say I want it to do a weighted sum of all the different <code>y<\/code> values. Each time just 1<code>y<\/code> or 5<code>y<\/code>s, etc. <\/p>\n","answers":[]}
+{"QuestionId":52711895,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-09T00:53:37.900","AcceptedAnswerId":52716460.0,"OwnerUserId":3259896.0,"Title":"How to run define Tensorflow graph were all variables are in float16 instead instead of float32","Body":"<p>By default, the variables Tensorflow is in float32. To save memory, I'm trying to run in float16. In my graph, every place where I could define the datatype as float16, I did. However, I get an error when I run the code <\/p>\n\n<p>Here's my code below. <\/p>\n\n<pre><code>import math\nimport numpy as np\nimport tensorflow as tf\n\nvocabulary_size = 10\nbatch_size = 64 \nembedding_size = 100 \nnum_inputs =4\nnum_sampled = 128 \n\ngraph = tf.Graph()\n\nwith graph.as_default(): #took out \" , tf.device('\/cpu:0')\"\n\n\n    train_dataset = tf.placeholder(tf.int32, shape=[batch_size, num_inputs ])\n    train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])\n\n    embeddings = tf.get_variable( 'embeddings', dtype=tf.float16,\n        initializer= tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0, dtype=tf.float16) )\n\n    softmax_weights = tf.get_variable( 'softmax_weights', dtype=tf.float16,\n        initializer= tf.truncated_normal([vocabulary_size, embedding_size],\n                             stddev=1.0 \/ math.sqrt(embedding_size), dtype=tf.float16 ) )\n\n    softmax_biases = tf.get_variable('softmax_biases', dtype=tf.float16,\n        initializer= tf.zeros([vocabulary_size], dtype=tf.float16),  trainable=False )\n\n    embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n    embed_reshaped = tf.reshape( embed, [batch_size*num_inputs, embedding_size] )\n\n    segments= np.arange(batch_size).repeat(num_inputs)\n\n    averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n    sam_sof_los = tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds,\n                                   labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)\n\n    loss = tf.reduce_mean( sam_sof_los )\n\n    optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) \n\n    saver = tf.train.Saver()\n<\/code><\/pre>\n\n<p>And this is this is the error message<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    509                 as_ref=input_arg.is_ref,\n--&gt; 510                 preferred_dtype=default_dtype)\n    511           except TypeError as err:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx)\n   1143     if ret is None:\n-&gt; 1144       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n   1145 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py in _TensorTensorConversionFunction(t, dtype, name, as_ref)\n    980         \"Tensor conversion requested dtype %s for Tensor with dtype %s: %r\" %\n--&gt; 981         (dtype.name, t.dtype.name, str(t)))\n    982   return t\n\nValueError: Tensor conversion requested dtype float16 for Tensor with dtype float32: 'Tensor(\"sampled_softmax_loss\/Log:0\", shape=(64, 1), dtype=float32)'\n\nDuring handling of the above exception, another exception occurred:\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-2-12d508b9e5d7&gt; in &lt;module&gt;()\n     46 \n     47     sam_sof_los = tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds,\n---&gt; 48                                    labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)\n     49 \n     50     loss = tf.reduce_mean( sam_sof_los )\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py in sampled_softmax_loss(weights, biases, labels, inputs, num_sampled, num_classes, num_true, sampled_values, remove_accidental_hits, partition_strategy, name, seed)\n   1347       partition_strategy=partition_strategy,\n   1348       name=name,\n-&gt; 1349       seed=seed)\n   1350   labels = array_ops.stop_gradient(labels, name=\"labels_stop_gradient\")\n   1351   sampled_losses = nn_ops.softmax_cross_entropy_with_logits_v2(\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py in _compute_sampled_logits(weights, biases, labels, inputs, num_sampled, num_classes, num_true, sampled_values, subtract_log_q, remove_accidental_hits, partition_strategy, name, seed)\n   1126     if subtract_log_q:\n   1127       # Subtract log of Q(l), prior probability that l appears in sampled.\n-&gt; 1128       true_logits -= math_ops.log(true_expected_count)\n   1129       sampled_logits -= math_ops.log(sampled_expected_count)\n   1130 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/math_ops.py in binary_op_wrapper(x, y)\n    860     with ops.name_scope(None, op_name, [x, y]) as name:\n    861       if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):\n--&gt; 862         return func(x, y, name=name)\n    863       elif not isinstance(y, sparse_tensor.SparseTensor):\n    864         try:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gen_math_ops.py in sub(x, y, name)\n   8316   if _ctx is None or not _ctx._eager_context.is_eager:\n   8317     _, _, _op = _op_def_lib._apply_op_helper(\n-&gt; 8318         \"Sub\", x=x, y=y, name=name)\n   8319     _result = _op.outputs[:]\n   8320     _inputs_flat = _op.inputs\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    544                   \"%s type %s of argument '%s'.\" %\n    545                   (prefix, dtypes.as_dtype(attrs[input_arg.type_attr]).name,\n--&gt; 546                    inferred_from[input_arg.type_attr]))\n    547 \n    548           types = [values.dtype]\n\nTypeError: Input 'y' of 'Sub' Op has type float32 that does not match type float16 of argument 'x'.\n<\/code><\/pre>\n\n<p>The error comes from line <code>tf.nn.sampled_softmax_loss<\/code>.<\/p>\n\n<p>At first I thought perhaps tf.segment_mean may cast the output as a float32, so I tried casting averaged_embeds to float16 but I still get the same error. <\/p>\n\n<p>From the documentation, there doesn't seem to be a way to define any data types in sampled_softmax_loss<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sampled_softmax_loss\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sampled_softmax_loss<\/a><\/p>\n","answers":[{"AnswerId":"52716460","CreationDate":"2018-10-09T08:24:05.363","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>As far as I can tell, you can only do it using a hack.<\/p>\n\n<p>The issue comes from the call to:<\/p>\n\n<pre><code>  if sampled_values is None:\n      sampled_values = candidate_sampling_ops.log_uniform_candidate_sampler(\n          true_classes=labels,\n          num_true=num_true,\n          num_sampled=num_sampled,\n          unique=True,\n          range_max=num_classes,\n          seed=seed)\n<\/code><\/pre>\n\n<p>which outputs an object of this type:<\/p>\n\n<pre><code>LogUniformCandidateSampler(\n    sampled_candidates=&lt;tf.Tensor 'LogUniformCandidateSampler:0' shape=(128,) dtype=int64&gt;,\n    true_expected_count=&lt;tf.Tensor 'LogUniformCandidateSampler:1' shape=(64, 1) dtype=float32&gt;,\n    sampled_expected_count=&lt;tf.Tensor 'LogUniformCandidateSampler:2' shape=(128,) dtype=float32&gt;\n)\n<\/code><\/pre>\n\n<p>The hack would be to generate yourself the <code>LogUniformCandidateSampler<\/code>, to cast its result as <code>tf.float16<\/code> and pass it to <code>tf.nn.sampled_softmax_loss<\/code>.<\/p>\n\n<pre><code># Redefine it as the tensorflow one is not exposed.\nLogUniformCandidateSampler = namedtuple(\"namedtuple\", [\"sampled_candidates\", \"true_expected_count\", \"sampled_expected_count\"]) \nsampled_values = tf.nn.log_uniform_candidate_sampler(\n      true_classes=tf.cast(train_labels, tf.int64), num_sampled=num_sampled,\n      num_true=1,\n      unique=True,\n      range_max=vocabulary_size,\n      seed=None)\n\nsampled_value_16 = LogUniformCandidateSampler(\n    sampled_values.sampled_candidates,\n    tf.cast(sampled_values.true_expected_count, tf.float16),\n    tf.cast(sampled_values.sampled_expected_count, tf.float16))\n\nsam_sof_los = tf.nn.sampled_softmax_loss(\n    weights=softmax_weights,\n    biases=softmax_biases,\n    inputs=averaged_embeds,\n    labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size, \n    sampled_values=sampled_value_16)\n<\/code><\/pre>\n\n<p>But this is really a hack and it might have unexpected consequences (an expected one would be that the <code>tf.cast<\/code> operation is not differentiable).<\/p>\n"}]}
+{"QuestionId":52712301,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-10-09T01:58:43.740","AcceptedAnswerId":52881394.0,"OwnerUserId":2302912.0,"Title":"Keras merge VS concatenate, can't update my code","Body":"<p>I have a Keras functional model for a CNN. I'm trying to implement a triplet-loss function. I found some posts about who to do that using \"merge\", which is now deprecated, but I'm not able to use \"concatenate\" as I was using merge.<\/p>\n\n<p>The original code looks like this:<\/p>\n\n<pre><code>def triplet_loss(x):\n    anchor, positive, negative = x\n    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), 1)\n    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), 1)\n\n    basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), 0.05)\n    loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)\n    return loss\n\n\n\ndef build_model(img_x, img_y):\n    input_shape = Input(shape=(img_x, img_y, 3))\n    c0 = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), activation='relu') (input_shape)\n    m0 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2)) (c0)\n    f = Flatten()(m0)\n    d1 = Dense(4024, activation='relu')(f)\n    d2 = Dense(512, activation='sigmoid')(d1)\n\n    anchor = Input(shape=(128, 254, 3))\n    positive = Input(shape=(128, 254, 3))\n    negative = Input(shape=(128, 254, 3))\n\n    reid_model = Model(inputs=[input_shape], outputs=[d2])\n\n    anchor_embed = reid_model(anchor)\n    positive_embed = reid_model(positive)\n    negative_embed = reid_model(negative)\n\n    loss = merge([anchor_embed, positive_embed, negative_embed],\n             mode=triplet_loss, output_shape=(1,))\n\n    model = Model(inputs=[anchor, positive, negative], outputs=loss)\n    model.compile(optimizer='Adam', loss='mean_absolute_error')\n    return model\n<\/code><\/pre>\n\n<p>I was using <code>loss = merge([anchor_embed, positive_embed, negative_embed], mode=triplet_loss, output_shape=(1,))<\/code> as a way to transform the output of the function <code>triplet_loss<\/code> into a keras layer output (as suggested in <a href=\"https:\/\/codepad.co\/snippet\/F1uVDD5N\" rel=\"nofollow noreferrer\">https:\/\/codepad.co\/snippet\/F1uVDD5N<\/a>). The function <code>concatenate<\/code> doesn't have a parameter \"mode\". HIs there any way to adapt my code to get the result of the loss function as a Keras layer output?<\/p>\n","answers":[{"AnswerId":"52881394","CreationDate":"2018-10-18T19:43:43.820","ParentId":null,"OwnerUserId":"2302912","Title":null,"Body":"<p>I finally found a way to compute the value of the <code>triplet_loss<\/code> function keeping the original architecture of my code by adding a lambda layer. <\/p>\n\n<pre><code>def triplet_loss(x):\n    anchor, positive, negative = x\n    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), 1)\n    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), 1)\n\n    basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), 0.05)\n    loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)\n    return loss\n\ndef build_model(img_x, img_y):\n    input_shape = Input(shape=(img_x, img_y, 3))\n    c0 = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), activation='relu') \n(input_shape)\n    m0 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2)) (c0)\n    f = Flatten()(m0)\n    d1 = Dense(4024, activation='relu')(f)\n    d2 = Dense(512, activation='sigmoid')(d1)\n\n    anchor = Input(shape=(128, 254, 3))\n    positive = Input(shape=(128, 254, 3))\n    negative = Input(shape=(128, 254, 3))\n\n    reid_model = Model(inputs=[input_shape], outputs=[d2])\n\n    anchor_embed = reid_model(anchor)\n    positive_embed = reid_model(positive)\n    negative_embed = reid_model(negative)\n\n    merged_output = concatenate([anchor_embed, positive_embed, \nnegative_embed])\n    loss = Lambda(triplet_loss, (1,))(merged_output)\n\n    model = Model(inputs=[anchor, positive, negative], outputs=loss)\n    model.compile(optimizer='Adam', loss='mse',\n                  metrics=[\"mae\"])\n    return model\n<\/code><\/pre>\n"}]}
+{"QuestionId":52712655,"AnswerCount":1,"Tags":"<python><keras><embedding>","CreationDate":"2018-10-09T02:48:46.433","AcceptedAnswerId":null,"OwnerUserId":10456792.0,"Title":"How to set embedding layer name in keras","Body":"<pre><code>inputs_bedding = Input(shape=(it.shape))   \nembedding = Embedding(9488, 512, trainable=False)(inputs_bedding)\n<\/code><\/pre>\n\n<p>There is no <code>name<\/code> parameter in keras Embedding layer. How to set the name to the layer?<\/p>\n","answers":[{"AnswerId":"52714863","CreationDate":"2018-10-09T06:47:16.743","ParentId":null,"OwnerUserId":"5143695","Title":null,"Body":"<p>You can set the name of the embedding layer just like any other layer.<\/p>\n\n<pre><code>from keras.layers import Embedding, Input\nfrom keras import Model\n\n\ninputs_bedding = Input(shape=(32,)) \nembedding = Embedding(9488, 512, trainable=False, name=\"test\")(inputs_bedding)\n\nmodel = Model(inputs=inputs_bedding, outputs=embedding)\n<\/code><\/pre>\n\n<p>model.summary() gives you:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_3 (InputLayer)         (None, 32)                0         \n_________________________________________________________________\ntest (Embedding)             (None, 32, 512)           4857856   \n=================================================================\nTotal params: 4,857,856\nTrainable params: 0\nNon-trainable params: 4,857,856\n_________________________________________________________________\n<\/code><\/pre>\n"}]}
+{"QuestionId":52712782,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><loss-function>","CreationDate":"2018-10-09T03:08:01.863","AcceptedAnswerId":null,"OwnerUserId":7892871.0,"Title":"How to create a loss function parameter that is dependent on epoch number in Keras?","Body":"<p>I have a custom loss function with a hyperparameter <code>alpha<\/code> that I want to change every 20 epochs over training. The loss function is something like: <\/p>\n\n<pre><code>def custom_loss(x, x_pred): \n    loss1 = binary_crossentropy(x, x_pred)\n    loss2 = (x, x_pred)\n    return (alpha)* loss1 + (1-alpha)*loss2\n<\/code><\/pre>\n\n<p>From my research, creating a custom callback is the way to go. I have looked at the solution for a similar question <a href=\"https:\/\/stackoverflow.com\/questions\/42995711\/keras-epoch-dependant-loss-function\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/40055225\/how-to-create-a-loss-function-which-changes-over-epoch-in-keras\">here<\/a> but the solutions do not implement the callback solution which is what I want to accomplish. <\/p>\n\n<p>I've attempted to create a custom callback by altering the <code>LearningRateScheduler<\/code> callback from the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L633\" rel=\"nofollow noreferrer\">keras repo<\/a> <\/p>\n\n<pre><code>class changeAlpha(Callback):\n    def __init__(self, alpha):\n        super(changeAlpha, self).__init__()\n        self.alpha = alpha \n\n    def on_epoch_begin(self, epoch, logs={}):\n        if epoch%20 == 0:   \n             K.set_value(self.alpha, K.get_value(self.alpha) * epoch**0.95)\n             print(\"Setting alpha to =\", str(alpha))\n<\/code><\/pre>\n\n<p>However, I am not certain the alpha value actually corresponds to the alpha value in my loss function. In any case, when I put the <code>changeAlpha<\/code> callback in the <code>model.fit<\/code> method, I receive an <code>attribute error<\/code>.<\/p>\n\n<p>Can someone help me edit the callback such that it alters my <code>alpha<\/code> parameter after a certain number of epochs? <\/p>\n","answers":[{"AnswerId":"52713046","CreationDate":"2018-10-09T03:41:14.837","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>I understood your idea. I think that the problem is that the alpha in the loss function is not referred to the member of changeAlpha class. You can try like this:<\/p>\n\n<pre><code>instance = changeAlpha()\ndef custom_loss(x, x_pred): \n    loss1 = binary_crossentropy(x, x_pred)\n    loss2 = (x, x_pred)\n    return (instance.alpha*)* loss1 + (1-instance.alpha)*loss2\n<\/code><\/pre>\n\n<p>Or, you can make the alpha as the class variable not the install variable, and then change the loss function as below:<\/p>\n\n<pre><code>def custom_loss(x, x_pred): \n    loss1 = binary_crossentropy(x, x_pred)\n    loss2 = (x, x_pred)\n    return (changeAlpha.alpha*)* loss1 + (1-changeAlpha.alpha)*loss2\n<\/code><\/pre>\n\n<p>Hope it can help you.<\/p>\n"}]}
+{"QuestionId":52713304,"AnswerCount":0,"Tags":"<python><tensorflow><masking>","CreationDate":"2018-10-09T04:16:28.067","AcceptedAnswerId":null,"OwnerUserId":2815551.0,"Title":"Adding a Mask in TensorFlow for training\/inference","Body":"<p>I am using <code>tf.contrib.slim<\/code> to train and evaluate state-of-the-art CNNs. I was wondering if there is a method that I can use <code>tf.select<\/code> or <code>tf.where<\/code> to create a mask to ignore certain weights' computation during training (or even inference) certain. Say, I would like to ignore the weights having value 0 and keep up training my model. Is it possible? <\/p>\n\n<p>Any suggestions\/codes would be appreciated. <\/p>\n","answers":[]}
+{"QuestionId":52713614,"AnswerCount":0,"Tags":"<python><tensorflow><keras><time-series><lstm>","CreationDate":"2018-10-09T04:56:48.250","AcceptedAnswerId":null,"OwnerUserId":7221923.0,"Title":"rescale and inverse_transform not returning predicted and actual values in original scale for LSTM multivariate time-series forecast","Body":"<p>I'm currently building an LSTM multivariate time-series model to predict one output at current time (t) using 22 features from the previous timestamp (t-1) as inputs. I've been following the instructions from <a href=\"https:\/\/machinelearningmastery.com\/multivariate-time-series-forecasting-lstms-keras\/\" rel=\"nofollow noreferrer\">this example<\/a>, and everything seems to be working correctly, but the one issue I'm having involves the final inverse_transform() functions.<\/p>\n\n<p>Specifically, I'm expecting that once I invert the scaling for both the forecasted values and the actual values, the output will be in the same units as the original data before I normalized and transformed my dataset. The variable I'm trying to predict has a mean of around 29.186, but when I run inverse_transform() and plot the predictions against the actuals, the output of my results ranges from 3.30 - 3.55.<\/p>\n\n<p>It could be that I've made a simple error slicing the array, or it could be something completely different, but I can't seem to identify the root cause of why my predicted and actual values aren't in the same units as the data before I transformed everything.<\/p>\n\n<p>For additional context, here's the printed output of train_X.shape, train_y.shape, test_X.shape, test_y.shape: <\/p>\n\n<pre><code>(1804950, 1, 22) (1804950,) (849389, 1, 22) (849389,)\n<\/code><\/pre>\n\n<p>Here's the relevant parts of my code:<\/p>\n\n<pre><code># ensure all data is float\nvalues = ips_data.values.astype('float32')\n\n# normalize features\nscaler = MinMaxScaler(feature_range=(0, 1))\nscaled = scaler.fit_transform(values)\n\n# frame as supervised learning\nips_reframed = series_to_supervised(scaled, 1, 1)\n\n# drop a bunch of columns I don't want to predict\nips_reframed.drop(ips_reframed.columns[[22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,42,43]], axis=1, inplace=True)\n# print(ips_reframed.head())\n\n# split into train and test sets\nvalues = ips_reframed.values\n# n_train_hours = (365 * 24 * 60) * 3\nn_train_hours = int(len(values) * 0.68)\ntrain = values[:n_train_hours, :]\ntest = values[n_train_hours:, :]\n\n# split into input and outputs\ntrain_X, train_y = train[:, :-1], train[:, -1]\ntest_X, test_y = test[:, :-1], test[:, -1]\n\n# reshape input to be 3D [samples, timesteps, features]\ntrain_X = train_X.reshape((train_X.shape[0], 1, train_X.shape[1]))\ntest_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1]))\n# print(train_X.shape, train_y.shape, test_X.shape, test_y.shape)\n\n# design network\nmodel = Sequential()\nmodel.add(LSTM(50, input_shape=(train_X.shape[1], train_X.shape[2]), return_sequences=True))\nmodel.add(Dropout(0.5))\nmodel.add(LSTM(50))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\nmodel.add(Activation(\"linear\"))\nmodel.compile(loss='mae', optimizer='adam')\n\n# fit network\nhistory = model.fit(train_X, train_y, epochs=100, batch_size=1000, \nvalidation_data=(test_X, test_y), verbose=2, shuffle=False)\n\n# plot history\nplt.plot(history.history['loss'], label='train')\nplt.plot(history.history['val_loss'], label='test')\nplt.legend()\nplt.show()\n\n# make a prediction\nyhat = model.predict(test_X)\n\n# invert scaling for forecast\ntest_X = test_X.reshape((test_X.shape[0], test_X.shape[2]))\ninv_yhat = concatenate((yhat, test_X[:, 1:]), axis=1)\ninv_yhat = scaler.inverse_transform(inv_yhat)\ninv_yhat = inv_yhat[:,0]\n\n# invert scaling for actual\ntest_y = test_y.reshape((len(test_y), 1))\ninv_y = concatenate((test_y, test_X[:, 1:]), axis=1)\ninv_y = scaler.inverse_transform(inv_y)\ninv_y = inv_y[:,0]\n\n# calculate RMSE\nrmse = sqrt(mean_squared_error(inv_y, inv_yhat))\n# print('Test RMSE: %.3f' % rmse)\n\n# plotting the predictions\nplt.plot(inv_yhat[-100:], label='predictions')\nplt.plot(inv_y[-100:], label='actual')\nplt.title(\"Prediction vs. Actual\")\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<p>Please let me know if you need any additional information or context, and thank you so much for your help!<\/p>\n","answers":[]}
+{"QuestionId":52714369,"AnswerCount":2,"Tags":"<tensorflow><gpu><nvidia><nsight><nvprof>","CreationDate":"2018-10-09T06:09:54.253","AcceptedAnswerId":55940784.0,"OwnerUserId":4238247.0,"Title":"nv-nsight-cu-cli caused Tensorflow to fail","Body":"<p>I've downloaded the newest Nsight Compute profiling tool and I want to use it to benchmark Tensorflow applications. The code I'm using is <a href=\"https:\/\/github.com\/valohai\/tensorflow-example\/blob\/master\/train.py\" rel=\"nofollow noreferrer\">here<\/a>. It runs perfectly fine when I execute it and when I benchmark it with <code>nvprof .\/mnist.py<\/code> it had no problem at all. However, when I try to run it with command <code>sudo .\/nv-nsight-cu-cli [path to the file]<\/code> I get the following error:<\/p>\n\n<p><code>ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory<\/code><\/p>\n\n<p>I suspect that <code>nv-nsight-cu-cli<\/code> somehow didn't recognized the environment variable at all. Is there any fix around?<\/p>\n","answers":[{"AnswerId":"55940784","CreationDate":"2019-05-01T18:20:39.160","ParentId":null,"OwnerUserId":"4238247","Title":null,"Body":"<p>It appears that GP100 is not supported by the tool at this moment.\nThe answer is found <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/1045430\/nsight-compute-\/partial-profiling\/\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>Nsight Compute only supports Pascal (other than GP100) and later GPUs.<\/p>\n<\/blockquote>\n"},{"AnswerId":"55933787","CreationDate":"2019-05-01T08:53:49.400","ParentId":null,"OwnerUserId":"2626244","Title":null,"Body":"<p>You need to search for differences in both environments:<\/p>\n\n<ul>\n<li>env variables<\/li>\n<li><code>LD_LIBRARY_PATH<\/code><\/li>\n<li><code>\/etc\/ld.so.conf<\/code><\/li>\n<li><code>\/etc\/ld.so.conf.d\/*<\/code><\/li>\n<li>cuBLAS\n\n<ul>\n<li>Is installation complete\/not broken?<\/li>\n<li>Is it installed at the same location on both machines?<\/li>\n<li>Versions<\/li>\n<\/ul><\/li>\n<li>...<\/li>\n<\/ul>\n\n<p>You can start with <code>locate libcublas.so<\/code> on both machines to see if there's a difference. Alternatively, you can <code>strace -f -e open<\/code> the program to check where it tries to <code>libcublas.so<\/code> from.<\/p>\n\n<p>Your error has (for now) nothing to do with GPUs: <code>libcublas.so.9.0<\/code> can just not be found. Find it, find why Tensorflow can not find it and your problem will be solved.<\/p>\n"}]}
+{"QuestionId":52714986,"AnswerCount":1,"Tags":"<machine-learning><keras><recurrent-neural-network><keras-layer>","CreationDate":"2018-10-09T06:55:42.203","AcceptedAnswerId":52719046.0,"OwnerUserId":9266500.0,"Title":"ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2 [keras]","Body":"<p>I got the error: <code>ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2<\/code> with the following code:<\/p>\n\n<pre><code>def make_model():\n  model = Sequential()      \n\n  model.add(Conv2D(20,(5,5), input_shape = (24,48,30), activation = \"relu\", strides = 1, padding = \"valid\"))\n  model.add(MaxPooling2D(pool_size=(2,2)))        \n  model.add(Conv2D(50, (5,5), use_bias = 50))    \n  model.add(MaxPooling2D(pool_size=(2,2)))    \n  model.add(Flatten())\n  model.add(Dense(20, activation = \"relu\"))\n  model.add(LSTM(50, activation=\"relu\", return_sequences=True))\n\n  return model\n<\/code><\/pre>\n\n<p>My input is 30 matrices of size 24*48 each. <\/p>\n","answers":[{"AnswerId":"52719046","CreationDate":"2018-10-09T10:40:04.870","ParentId":null,"OwnerUserId":"583464","Title":null,"Body":"<p>The problem lies in the fact that after the last Dense layer (before the lstm layer), the output shape is (?, 20) and the lstm layer expects 3D tensor, not 2D.So, you can expand the dimensions in order to add one more before feeding to lstm layer.<\/p>\n\n<p>You can expand dimensions using tf.expand_dims (assuming you use tensorflow as backend) <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims\" rel=\"nofollow noreferrer\">tf expand<\/a><\/p>\n\n<pre><code>input_layer = Input((30,24,48))\n\nmodel = Conv2D(20,(5,5), input_shape = (30,24,48), activation = \"relu\", strides = 1, padding = \"valid\")(input_layer)\nmodel = MaxPooling2D(pool_size=(2,2))(model)        \nmodel = Conv2D(50, (5,5), use_bias = 50)(model)    \nmodel = MaxPooling2D(pool_size=(2,2))(model)  \nmodel = Flatten()(model)\nmodel = Dense(20, activation = \"relu\")(model)\nmodel = tf.expand_dims(model, axis=-1)\nmodel = LSTM(50, activation=\"relu\", return_sequences=True)(model)\n<\/code><\/pre>\n\n<p>(I didn't use Sequential mode, I am using <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">functional api<\/a> since it is more flexible)<\/p>\n\n<p>If you want to use sequential model:<\/p>\n\n<pre><code>    model = Sequential()      \n\n    model.add(Conv2D(20,(5,5), input_shape = (30, 24, 48), activation = \"relu\", strides = 1, padding = \"valid\"))\n    model.add(MaxPooling2D(pool_size=(2,2)))        \n    model.add(Conv2D(50, (5,5), use_bias = 50))    \n    model.add(MaxPooling2D(pool_size=(2,2)))    \n    model.add(Flatten())\n    model.add(Dense(20, activation = \"relu\"))\n    model.add(Lambda(lambda x: tf.expand_dims(model.output, axis=-1)))\n    model.add(LSTM(50, activation=\"relu\", return_sequences=True))\n<\/code><\/pre>\n\n<p>you must use expand dims inside <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\">Lambda<\/a><\/p>\n"}]}
+{"QuestionId":52715477,"AnswerCount":0,"Tags":"<apache-spark><tensorflow><keras><pyspark><spark-streaming>","CreationDate":"2018-10-09T07:25:48.143","AcceptedAnswerId":null,"OwnerUserId":3782444.0,"Title":"Import statements taking time on spark executors (Pyspark Executors)","Body":"<p>I am developing a python prediction script using Spark (PySpark) streaming and Keras. The prediction is happening on the executor where I am calling model.predict().<\/p>\n\n<p>Modules that I have imported are<\/p>\n\n<pre><code>from keras.layers.core import Dense, Activation, Dropout \nfrom keras.layers.recurrent import LSTM \nfrom keras.models import Sequential\n<\/code><\/pre>\n\n<p>I have checked and these imports are taking 2.5 seconds on Spark driver(2 cor + 2gb) to load. What is surprising for me is that each time executor gets the job, it automatically do these imports again. The reason I am sure that these imports are happening each time a job is submitted to the executor is because I see below statements in executor logs per job which are only comes when I do imports of above said modules.<\/p>\n\n<pre><code>\/opt\/conda\/lib\/python3.6\/site-packages\/h5py\/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype fromfloattonp.floatingis deprecated. In future, it will be treated asnp.float64 == np.dtype(float).type. from ._conv import register_converters as _register_converters Using TensorFlow backend.\n<\/code><\/pre>\n\n<p>My target is to make the prediction in 1 sec but imports itself are taking 2.5 seconds (each time imports happens on the spark executor). Is this is the intended behavior? Is their anything I can do to minimize this time say in milliseconds ?<\/p>\n\n<p><strong>Update 1<\/strong><\/p>\n\n<p>For the past few days I was analyzing it and found out that there are 2 main issues.<\/p>\n\n<ol>\n<li>I found a way to wrap keras model and pickle it on driver and unpickle on executor. So that improves time by 1 sec.<\/li>\n<li>But on driver when mapPartition() occur for each batch the whole keras and tensorflow initialization happens each time on executor (for every job) which is taking 2.5 secs. Is there a way to initialize these imports once on each executor and not per job. May be some file in pyspark where can I put these imports(assuming that these files executed once when executor comes up)  <\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":52715499,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving><tensorflow-hub>","CreationDate":"2018-10-09T07:27:11.287","AcceptedAnswerId":52718488.0,"OwnerUserId":9594598.0,"Title":"How to serve a tensorflow-module, specifically Universal Sentence Encoder?","Body":"<p>I have spent several hours trying to set up Tensorflow serving of the Tensorflow-hub module, \"Universal Sentence Encoder.\" There is a similar question here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50788080\/how-to-make-the-tensorflow-hub-embeddings-servable-using-tensorflow-serving\">How to make the tensorflow hub embeddings servable using tensorflow serving?<\/a> <\/p>\n\n<p>I have been doing this on a Windows machine.<\/p>\n\n<p>This is the code I used to build the model:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_hub as hub\n\nMODEL_NAME = 'test'\nVERSION = 1\nSERVE_PATH = '.\/models\/{}\/{}'.format(MODEL_NAME, VERSION)\n\nwith tf.Graph().as_default():\n  module = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence- \n  encoder\/1\")\n  text = tf.placeholder(tf.string, [None])\n  embedding = module(text)\n\n  init_op = tf.group([tf.global_variables_initializer(), \n  tf.tables_initializer()])\n  with tf.Session() as session:\n  session.run(init_op)\n    tf.saved_model.simple_save(\n    session,\n    SERVE_PATH,\n    inputs = {\"text\": text},\n    outputs = {\"embedding\": embedding},\n    legacy_init_op = tf.tables_initializer()        \n   )\n<\/code><\/pre>\n\n<p>I have gotten to the point where running the following line:<\/p>\n\n<pre><code>saved_model_cli show --dir ${PWD}\/models\/test\/1 --tag_set serve --signature_def serving_default\n<\/code><\/pre>\n\n<p>gives me the following result:<\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\ninputs['text'] tensor_info:\n  dtype: DT_STRING\n  shape: (-1)\n  name: Placeholder:0\nThe given SavedModel SignatureDef contains the following output(s):\n outputs['embedding'] tensor_info:\n  dtype: DT_FLOAT\n  shape: (-1, 512)\n  name: module_apply_default\/Encoder_en\/hidden_layers\/l2_normalize:0\n<\/code><\/pre>\n\n<p>I have then tried running:<\/p>\n\n<pre><code>saved_model_cli run --dir ${PWD}\/models\/test\/1 --tag_set serve --signature_def serving_default --input_exprs 'text=[\"what this is\"]'\n<\/code><\/pre>\n\n<p>which gives the error:<\/p>\n\n<pre><code>  File \"&lt;string&gt;\", line 1\n[what this is]\n         ^\nSyntaxError: invalid syntax\n<\/code><\/pre>\n\n<p>I've tried changing the format of the part 'text=[\"what this is\"]', but nothing worked for me.<\/p>\n\n<p>Regardless of if this part works, the main goal is to set up the module for serving and create a callable API. <\/p>\n\n<p>I've tried with docker, the following line:<\/p>\n\n<pre><code>docker run -p 8501:8501 --name tf-serve -v ${PWD}\/models\/:\/models -t tensorflow\/serving --model_base_path=\/models\/test\n<\/code><\/pre>\n\n<p>Things appear to be set up properly: <\/p>\n\n<pre><code>Building single TensorFlow model file config:  model_name: model model_base_path: \/models\/test\n2018-10-09 07:05:08.692140: I tensorflow_serving\/model_servers\/server_core.cc:462] Adding\/updating models.\n2018-10-09 07:05:08.692301: I tensorflow_serving\/model_servers\/server_core.cc:517]  (Re-)adding model: model\n2018-10-09 07:05:08.798733: I tensorflow_serving\/core\/basic_manager.cc:739] Successfully reserved resources to load servable {name: model version: 1}\n2018-10-09 07:05:08.798841: I tensorflow_serving\/core\/loader_harness.cc:66] Approving load for servable version {name: model version: 1}\n2018-10-09 07:05:08.798870: I tensorflow_serving\/core\/loader_harness.cc:74] Loading servable version {name: model version: 1}\n2018-10-09 07:05:08.798904: I external\/org_tensorflow\/tensorflow\/contrib\/session_bundle\/bundle_shim.cc:360] Attempting to load native SavedModelBundle in bundle-shim from: \/models\/test\/1\n2018-10-09 07:05:08.798947: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/reader.cc:31] Reading SavedModel from: \/models\/test\/1\n2018-10-09 07:05:09.055822: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/reader.cc:54] Reading meta graph with tags { serve }\n2018-10-09 07:05:09.338142: I external\/org_tensorflow\/tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2018-10-09 07:05:09.576751: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/loader.cc:162] Restoring SavedModel bundle.\n2018-10-09 07:05:28.975611: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/loader.cc:138] Running MainOp with key saved_model_main_op on SavedModel bundle.\n2018-10-09 07:06:30.941577: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/loader.cc:259] SavedModel load for tags { serve }; Status: success. Took 82120946 microseconds.\n2018-10-09 07:06:30.990252: I tensorflow_serving\/servables\/tensorflow\/saved_model_warmup.cc:83] No warmup data file found at \/models\/test\/1\/assets.extra\/tf_serving_warmup_requests\n2018-10-09 07:06:31.046262: I tensorflow_serving\/core\/loader_harness.cc:86] Successfully loaded servable version {name: model version: 1}\n2018-10-09 07:06:31.184541: I tensorflow_serving\/model_servers\/server.cc:285] Running gRPC ModelServer at 0.0.0.0:8500 ...\n[warn] getaddrinfo: address family for nodename not supported\n2018-10-09 07:06:31.221644: I tensorflow_serving\/model_servers\/server.cc:301] Exporting HTTP\/REST API at:localhost:8501 ...\n[evhttp_server.cc : 235] RAW: Entering the event loop ...\n<\/code><\/pre>\n\n<p>I've tried <\/p>\n\n<pre><code>curl http:\/\/localhost:8501\/v1\/models\/test\n<\/code><\/pre>\n\n<p>which gives<\/p>\n\n<pre><code>{ \"error\": \"Malformed request: GET \/v1\/models\/test:predict\" }\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>curl -d '{\"text\": \"Hello\"}' -X POST http:\/\/localhost:8501\/v1\/models\/test:predict\n<\/code><\/pre>\n\n<p>which gives <\/p>\n\n<pre><code>{ \"error\": \"JSON Parse error: Invalid value. at offset: 0\" }\n<\/code><\/pre>\n\n<p>A similar question is here <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/52671138\/tensorflow-serving-rest-api-returns-malformed-request-error\">Tensorflow Serving: Rest API returns &quot;Malformed request&quot; error<\/a><\/p>\n\n<p>Just looking for any way to get this module serving. Thanks.<\/p>\n","answers":[{"AnswerId":"52718488","CreationDate":"2018-10-09T10:11:14.747","ParentId":null,"OwnerUserId":"9594598","Title":null,"Body":"<p>I was finally able to figure things out. I'll post what I did here in case someone else is trying to do the same thing.<\/p>\n\n<p>My issue with the saved_model_cli run command was with the quotes (using Windows command prompt). Change <code>'text=[\"what this is\"]'<\/code> to <code>\"text=['what this is']\"<\/code><\/p>\n\n<p>The issue with the POST request was two-fold. One, I noticed that the model's name is model, so should have been <a href=\"http:\/\/localhost:8501\/v1\/models\/model:predict\" rel=\"nofollow noreferrer\">http:\/\/localhost:8501\/v1\/models\/model:predict<\/a><\/p>\n\n<p>Secondly, the input format was not correct. I used Postman, and the body of the request looks like this:\n<code>{\"inputs\": {\"text\": [\"Hello\"]}}<\/code><\/p>\n"}]}
+{"QuestionId":52715626,"AnswerCount":1,"Tags":"<python><tensorflow><scipy>","CreationDate":"2018-10-09T07:35:30.610","AcceptedAnswerId":52719522.0,"OwnerUserId":8654291.0,"Title":"How to use a scipy function on each element of a tensor using Keras?","Body":"<p>I am new to tensorflow, i wish to apply a scipy gamma function to an existing tensor. When I try this <\/p>\n\n<pre><code>from scipy.special import gamma\ngamma_t = K.map_fn(lambda x:gamma(1.0 + 1.0 \/ x) ,b)\n<\/code><\/pre>\n\n<p>Where b is an existing tensor I get <\/p>\n\n<pre><code>TypeError: ufunc 'gamma' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''\n<\/code><\/pre>\n\n<p>How Can I solve This?<\/p>\n","answers":[{"AnswerId":"52719522","CreationDate":"2018-10-09T11:03:57.097","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You cannot use SciPy functions (or NumPy based functions, in general) directly on a TensorFlow tensor. You could do it with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\"><code>tf.py_func<\/code><\/a>, but in general the best option is to use TensorFlow operations only. In this case, neither the Keras backend abstraction nor TensorFlow have the gamma function, but TensorFlow has <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/lgamma\" rel=\"nofollow noreferrer\"><code>tf.lgamma<\/code><\/a>, which is its logarithm (well, the logarithm of its absolute value, to be precise). You can then get what you want with:<\/p>\n\n<pre><code>gamma_t = K.map_fn(lambda x: K.exp(tf.lgamma(1.0 + 1.0 \/ x)), b)\n<\/code><\/pre>\n\n<p>PS: Note that generally it is recommended to use only <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">backend functions<\/a> when manipulating Keras tensors, but since this is a rather specific feature and it is not exposed there (also, although Theano does have a gamma function implementation, CNTK currently does not, so it would not be possible to implement it for all backends).<\/p>\n"}]}
+{"QuestionId":52715744,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><loss-function>","CreationDate":"2018-10-09T07:43:12.860","AcceptedAnswerId":52716454.0,"OwnerUserId":7155684.0,"Title":"How to define a keras custom loss function in simple mathematical operation","Body":"<p>I define a custom function <code>my_sigmoid<\/code> as following:<\/p>\n\n<pre><code>import math\ndef my_sigmoid(x):\n    a =  1\/  ( 1+math.exp( -(x-300)\/30 ) )\n    return a\n<\/code><\/pre>\n\n<p>And then define a custom loss function called <code>my_cross_entropy<\/code>:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef my_cross_entropy(y_true, y_pred):\n    diff = abs(y_true-y_pred)\n    y_pred_transform = my_sigmoid(diff)\n    return K.categorical_crossentropy(0, y_pred_transform)\n<\/code><\/pre>\n\n<p>My keras backend is using tensorflow.\nAnd the error shows <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>TypeError: must be real number, not Tensor<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>I'm not familiar with tensorflow and don't know how to use custom loss.<\/p>\n\n<p>The following are my model structure and error message:<\/p>\n\n<pre><code>import keras.backend as K\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D, Dropout, Flatten, Dense\n\nmodel=Sequential()\nmodel.add(Conv2D(512,(5,X_train.shape[2]),input_shape=X_train.shape[1:4],activation=\"relu\"))\nmodel.add(Flatten())\nmodel.add(Dropout(0.1))\nmodel.add(Dense(100,activation=\"relu\"))\nmodel.add(Dense(100,activation=\"relu\"))\nmodel.add(Dense(50,activation=\"relu\"))\nmodel.add(Dense(10,activation=\"relu\"))\nmodel.add(Dense(1,activation=\"relu\"))\nmodel.compile(optimizer='adam', loss=my_cross_entropy)\nmodel.fit(X_train,Y_train,batch_size = 10,epochs=200,validation_data=(X_test,Y_test))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/otLc8.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/otLc8.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And the shape of <code>X_train<\/code> and <code>Y_train<\/code> is :\n<code>(120, 30, 80, 1)<\/code>\nand\n<code>(120,)<\/code><\/p>\n","answers":[{"AnswerId":"52716454","CreationDate":"2018-10-09T08:23:43.703","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Change <\/p>\n\n<pre><code>diff = abs(y_true-y_pred)\n<\/code><\/pre>\n\n<p>into<\/p>\n\n<pre><code>diff = K.abs(y_true-y_pred)\n<\/code><\/pre>\n\n<p>same for <\/p>\n\n<pre><code>math.exp()\n<\/code><\/pre>\n\n<p>change that into<\/p>\n\n<pre><code>K.exp()\n<\/code><\/pre>\n\n<p>abs and Math.exp are functions that cannot handle Tensors. If you still have problems refer to : <a href=\"https:\/\/stackoverflow.com\/questions\/45279085\/custom-loss-function-keras-tensorflow\">Custom Loss function Keras Tensorflow<\/a><\/p>\n"}]}
+{"QuestionId":52715884,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-10-09T07:51:17.080","AcceptedAnswerId":null,"OwnerUserId":5413768.0,"Title":"How performance for prediction with tensorflow session.run() and prediction with tf-serving differs?","Body":"<p>System configuration: 32core, 128GB RAM<\/p>\n\n<p>I have successfully deployed model with tensorflow-serving and model inference is done with gRPC request. It might be calling session.run() internally while prediction. With this approach, it takes 3sec for batch of 100 document.<\/p>\n\n<p>In case of direct deployment of tensorflow model (not with tf-serving), we pass batch of documents to session.run(model.prediction, feed_dict={}) and do inference.<\/p>\n\n<p>Question is, in which way tensorflow-serving differs from direct deployment in case of performance and optimization? What type of optimization is done in tf-serving while inference?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":52716136,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-09T08:04:54.497","AcceptedAnswerId":null,"OwnerUserId":2147347.0,"Title":"Tensorflow estimator: Different epoch number gives the same result","Body":"<p>I want to create a simple model using TF Estimator. Below is my code:<\/p>\n\n<pre><code>num_epoch = 4\nnum_train = 186\nnum_test = 81\n\n# 1 Define input function\ndef input_function(x, y, is_train):\n\n    dict_x = {\n        \"featurename\" : x,\n    }\n\n    dataset = tf.data.Dataset.from_tensor_slices((\n        dict_x, y\n    ))\n\n    if is_train:\n        dataset = dataset.shuffle(num_train, seed=113).repeat(num_epoch).batch(num_train)\n    else:   \n        dataset = dataset.batch(num_test)\n    return dataset\n\ndef my_serving_input_fn2():\n    input_data = {\n        \"featurename\" : tf.placeholder(tf.float32, [None, 15], name='inputtensors')\n    }\n    return tf.estimator.export.ServingInputReceiver(input_data, input_data)\n\n\ndef main(argv):\n    tf.set_random_seed(113)\n    np.random.seed(113)\n    run_config = tf.estimator.RunConfig(save_summary_steps=None, save_checkpoints_secs=None, tf_random_seed=113)\n\n    # 2 Define feature columns\n    feature_columns = [\n        tf.feature_column.numeric_column(key=\"featurename\",shape=15),\n    ]\n\n    # 3 Define an estimator\n    classifier = tf.estimator.DNNClassifier(\n        feature_columns=feature_columns,\n        hidden_units=[15, 15],\n        n_classes=3,\n        optimizer=tf.train.AdamOptimizer(),\n        activation_fn=tf.nn.relu,\n        model_dir = 'models\/',\n        config=run_config\n    )\n\n    # Train the model\n    classifier.train(\n        input_fn=lambda:input_function(X_train, y_train, True)\n    )\n\n    # Evaluate the model\n    eval_result = classifier.evaluate(\n        input_fn=lambda:input_function(X_test, y_test, False)\n    )\n\n    print('\\nTest set accuracy: {accuracy:0.3f}\\n'.format(**eval_result))\n\nif __name__ == \"__main__\":\n    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' \n    tf.app.run(main)\n<\/code><\/pre>\n\n<p>I believe the model will be trained <code>num_epoch<\/code> times as <code>num_epoch<\/code> affect dataset's <code>repeat()<\/code>. I didn't define <code>steps<\/code> in estimator so it will be trained forever until <code>input_fn<\/code> generates the <code>tf.errors.OutOfRange<\/code>.<\/p>\n\n<p>But if I change <code>num_epoch<\/code> value it always gives the same result. For example, when I run this program with <code>num_epoch=4<\/code> and <code>num_epoch=400<\/code> the test set accuracy will be the same.<\/p>\n\n<p>Am I misunderstanding the training process here? or it's just a coincidence? I hope someone can help explain why this happened<\/p>\n","answers":[{"AnswerId":"52716323","CreationDate":"2018-10-09T08:15:37.700","ParentId":null,"OwnerUserId":"5487345","Title":null,"Body":"<p>It may be because you don't have enough data. What are you doing is training it over the same data and hits a peak of \"intelligence\". Increase your dataset basically and you will see changes in your <code>num-epoch<\/code>.  <\/p>\n"}]}
+{"QuestionId":52716164,"AnswerCount":0,"Tags":"<python><pytorch><torchtext>","CreationDate":"2018-10-09T08:06:12.917","AcceptedAnswerId":null,"OwnerUserId":2174310.0,"Title":"Extending the vocabulary on deployment","Body":"<p>When doing training, I initialize my embedding matrix, using the pretrained embeddings picked for words in training set vocabulary.<\/p>\n\n<pre><code>import torchtext as tt\ncontexts = tt.data.Field(lower=True, sequential=True, tokenize=tokenizer, use_vocab=True)\ncontexts.build_vocab(data, vectors=\"fasttext.en.300d\",\n                         vectors_cache=config[\"vectors_cache\"])\n<\/code><\/pre>\n\n<p>In my model I pass contexts.vocab as parameter and initialize embeddings:<\/p>\n\n<pre><code>embedding_dim = vocab.vectors.shape[1]\nself.embeddings = nn.Embedding(len(vocab), embedding_dim)\nself.embeddings.weight.data.copy_(vocab.vectors)\nself.embeddings.weight.requires_grad=False\n<\/code><\/pre>\n\n<p>I train my model and during training I save its 'best' state via <code>torch.save(model, f)<\/code>.<\/p>\n\n<p>Then I want to test\/create demo for model in separate file for evaluation. I load the model via <code>torch.load<\/code>. How do I extend the embedding matrix to contain test vocabulary? I tried to replace embedding matrix<\/p>\n\n<pre><code># data is TabularDataset with test data\ncontexts.build_vocab(data, vectors=\"fasttext.en.300d\",\n                     vectors_cache=config[\"vectors_cache\"])\n\nmodel.embeddings = torch.nn.Embedding(len(contexts.vocab), contexts.vocab.vectors.shape[1])\nmodel.embeddings.weight.data.copy_(contexts.vocab.vectors)\nmodel.embeddings.weight.requires_grad = False\n<\/code><\/pre>\n\n<p>But the results are terrible (almost 0 accuracy). Model was doing good during training. What is the 'correct way' of doing this?<\/p>\n","answers":[]}
+{"QuestionId":52716998,"AnswerCount":1,"Tags":"<r><keras><r-caret>","CreationDate":"2018-10-09T08:53:42.317","AcceptedAnswerId":null,"OwnerUserId":10223141.0,"Title":"mlpKerasDropout in caret","Body":"<p>I have been trying to implement the mlpKerasDropout method when training using the caret R package.<\/p>\n\n<p>My code seems to cycle through 10 of 10 epochs continuously, and doesnt seem to converge. I have studied <a href=\"https:\/\/github.com\/topepo\/caret\/blob\/master\/models\/files\/mlpKerasDropout.R\" rel=\"nofollow noreferrer\">mlpKerasDropout.R<\/a> but am struggling to understand how this function works.<\/p>\n\n<p>Has anyone out there got a minimal example that they can share of how to use this function?<\/p>\n\n<p>Many thanks,<\/p>\n","answers":[{"AnswerId":"55752114","CreationDate":"2019-04-18T18:48:10.070","ParentId":null,"OwnerUserId":"1954810","Title":null,"Body":"<p>For each resampling (driven by trainControl()), it is running a fit. This is what you're seeing with the 10 of 10 epochs cycling continuously. Each cycle was a resample\/fold being fit. You can change the number of epochs used while you're hyperparameter tuning by setting the epochs argument to train, which will be passed to the training method \"mlpKerasDropout\" via dot args (...)<\/p>\n\n<p>See the code for mlpKerasDropout here: <a href=\"https:\/\/github.com\/topepo\/caret\/blob\/master\/models\/files\/mlpKerasDropout.R\" rel=\"nofollow noreferrer\">https:\/\/github.com\/topepo\/caret\/blob\/master\/models\/files\/mlpKerasDropout.R<\/a><\/p>\n\n<p>By default, the search argument for hyperparameters is set to 'grid' but you may want to set it to 'random' so that it tries different activation functions other than relu, or provide your own tuning grid.<\/p>\n\n<p>Here's a code sample showing the usage of tuneLength with search='random', and utilizing early stopping as well as epochs arguments passed to keras.<\/p>\n\n<pre><code>tune_model &lt;- train(x, y, \n           method = \"mlpKerasDropout\",\n           preProc = c('center', 'scale', 'spatialSign'),\n           trControl = trainControl(search = 'random', classProbs = T, \n              summaryFunction = mnLogLoss, allowParallel = TRUE),\n           metric = 'logLoss',\n           tuneLength = 20, \n# keras arguments following\nvalidation_split = 0.25,\n           callbacks = list(\n                keras::callback_early_stopping(monitor = \"val_loss\", mode = \"auto\", \n                  patience = 20, restore_best_weights = TRUE)\n              ),\n            epochs = 500)\n<\/code><\/pre>\n\n<p>Keep in mind you want re-fit your model on the training data after completing the hyperparameter tuning CV.<\/p>\n"}]}
+{"QuestionId":52717164,"AnswerCount":1,"Tags":"<tensorflow><keras><jupyter-notebook><google-colaboratory>","CreationDate":"2018-10-09T09:02:24.957","AcceptedAnswerId":null,"OwnerUserId":10392741.0,"Title":"Save files\/pictures in Google Colaboratory","Body":"<p>at the moment, I work with 400+ images and upload them with<\/p>\n\n<pre><code>from google.colab import files\nuploaded = files.upload()\n<\/code><\/pre>\n\n<p>This one's working fine but I have to reupload all the images every time I leave my colaboratory. Pretty annoying because the upload takes like 5-10 minutes.<\/p>\n\n<p>Any possibilities to prevent this? It seems like Colaboratory is saving the files only temporarily.<\/p>\n\n<p>I have to use Google Colaboratory because I need their GPU.<\/p>\n\n<p>Thanks in advance :) <\/p>\n","answers":[{"AnswerId":"52719351","CreationDate":"2018-10-09T10:55:15.133","ParentId":null,"OwnerUserId":"8337524","Title":null,"Body":"<p>As far as I know, there is no way to permanently store data on a Google Colab VM, but there are faster ways to upload data on Colab than files.upload().<\/p>\n\n<p>For example you can upload your images on Google Drive once and then 1) mount Google Drive directly in your VM or 2) use PyDrive to download your images on your VM. Both of these options should be way faster than uploading your images from your local drive.<\/p>\n\n<p><strong>Mounting Drive in your VM<\/strong><\/p>\n\n<ol>\n<li><p>Mount Google Drive:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/gdrive')\n<\/code><\/pre><\/li>\n<li><p>Print the contents of <code>foo.txt<\/code> located in the root directory of Drive:<\/p>\n\n<pre><code>with open('\/gdrive\/foo.txt') as f:\n    for line in f:\n        print(line)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p><strong>Using PyDrive<\/strong><\/p>\n\n<p>Take a look at the first answer to this <a href=\"https:\/\/stackoverflow.com\/questions\/48376580\/google-colab-how-to-read-data-from-my-google-drive\">question<\/a>.<\/p>\n"}]}
+{"QuestionId":52717181,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-09T09:03:23.087","AcceptedAnswerId":52721502.0,"OwnerUserId":4560470.0,"Title":"Keras : Create MobileNet_V2 model \"AttributeError\"","Body":"<p>I have successfully built several model based on mobileNet using keras. I noticed that MobileNet_V2 as been added in Keras 2.2.0, but I could not manage to make it work : <\/p>\n\n<pre><code>from keras.applications.mobilenet_v2 import mobilenet_v2\n\nbase_model = mobilenet_v2.MobileNetV2(weights='imagenet', include_top=False)\n<\/code><\/pre>\n\n<p>I get the following error : <code>AttributeError: 'NoneType' object has no attribute 'image_data_format'<\/code> on this line from mobilenet_v2.py <code>data_format=backend.image_data_format()<\/code><\/p>\n\n<p>It seems to me that <code>backend<\/code>has a definition problem... I am using Tensorflow backend, maybe it does not work with this one ?<\/p>\n","answers":[{"AnswerId":"52721502","CreationDate":"2018-10-09T12:49:51.117","ParentId":null,"OwnerUserId":"4560470","Title":null,"Body":"<p>The problem is from the import. The proper way to do this is to do the following :<\/p>\n\n<pre><code>from keras.applications import MobileNetV2\nm = MobileNetV2(weights='imagenet', include_top=False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52717630,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><tensorflow-datasets>","CreationDate":"2018-10-09T09:27:54.287","AcceptedAnswerId":55536270.0,"OwnerUserId":8188358.0,"Title":"Training crash due to saving model: \"tensorflow.GraphDef was modified concurrently during serialization\"","Body":"<p>I am currently trying to train a model, and my input pipeline is constructed as of this answer <a href=\"https:\/\/stackoverflow.com\/a\/47967475\/8188358\">here<\/a>. I want to save my model after each epochs. But after training for some epochs the training crash. I have read that it is because it adds the input as a constant tensor to the graph. There are suggested solutions <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19657\" rel=\"nofollow noreferrer\">here<\/a> to use a <code>tf.placeholder<\/code> to solve the problem. Unfortunately it doesn't solve the problem for me. The input pipeline looks as follows <\/p>\n\n<pre><code>....\nfilenames = [P_1]\ndataset = tf.data.TFRecordDataset(filenames)\ndef _parse_function(example_proto):\n       keys_to_features = { 'data':tf.VarLenFeature(tf.float32)},\n       parsed_features = tf.parse_single_example(example_proto,  keys_to_features)\n       return tf.sparse_tensor_to_dense(parsed_features['data'\n# Parse the record into tensors.\ndataset = dataset.map(_parse_function)\n# Shuffle the dataset\ndataset = dataset.shuffle(buffer_size=1000)\n# Repeat the input indefinitly \ndataset = dataset.repeat()      \n# Generate batches     \ndataset = dataset.batch(Batch_size) \n# Create a one-shot iterator\niterator = dataset.make_one_shot_iterator()\ndata = iterator.get_next()   \n....\nfor i in range(epochs):\n    for ii in range(iteration):\n        image = sess.run(data)\n        ....\n     saver.save(sess, 'filename')\n<\/code><\/pre>\n\n<p>The error message looks as follows <\/p>\n\n<pre><code>[libprotobuf FATAL external\/protobuf_archive\/src\/google\/protobuf\/message_lite.cc:68] CHECK failed: (byte_size_before_serialization) == (byte_size_after_serialization): tensorflow.GraphDef was modified concurrently during serialization.\nterminate called after throwing an instance of 'google::protobuf::FatalException'  \nwhat():  CHECK failed: (byte_size_before_serialization) == (byte_size_after_serialization): tensorflow.GraphDef was modified concurrently during serialization.\nAborted\n<\/code><\/pre>\n","answers":[{"AnswerId":"55536270","CreationDate":"2019-04-05T13:15:39.860","ParentId":null,"OwnerUserId":"11317239","Title":null,"Body":"<p>The problem looks like it is in the <code>_parse_function<\/code>. Make sure the parser is doing in the same way when you create the TFrecord file. For example if they have the same data type or so<\/p>\n"}]}
+{"QuestionId":52718451,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-09T10:09:18.527","AcceptedAnswerId":52718566.0,"OwnerUserId":7353867.0,"Title":"expected conv2d_1_input to have 4 dimensions, but got array with shape (15936, 64)","Body":"<p>Hello I'm new to Deep Learning and Keras and I was doing a project in order to learn Deep Learning and Keras. Here I've made a model.<\/p>\n\n<pre><code>Model = Sequential()\nModel.add(Conv2D(32, (3, 3) , input_shape = (100,64,64,), padding = 'same', \nactivation='relu'))\nModel.add(Conv2D(32, (3, 3), activation='relu', padding='same'))\nModel.add(MaxPooling2D(pool_size=(2, 2))) \nModel.add(Flatten()) #Conversion to Neurons\nModel.add(Dense(512, activation='relu')) \nModel.add(Dense(1, activation='softmax'))\n<\/code><\/pre>\n\n<p>For training and  fitting.<\/p>\n\n<pre><code>X = signalBuffer.transpose()\nY = np.ones([19920, 1], dtype = int)\nx_train, x_test, y_train, y_test = train_test_split(X, Y, \ntest_size=0.20,shuffle=True)\nModel.fit(x_train, y_train,batch_size=100,epochs=epochs,validation_data= \n(x_test, y_test),shuffle=True)\n<\/code><\/pre>\n\n<p>Here the X has 19920 rows and 64 columns and Y has 19920 rows and 1 column.\nThe training and testing splitting is executing without errors. The error is coming in the last line when I try to fit in the model.\nThe error is<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (15936, 64)<\/p>\n<\/blockquote>\n\n<p>Pardon me if this is a silly question or answer is very easy, but I'm trying to understand the model and I've tried a few solutions but its still giving errors. Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"52718566","CreationDate":"2018-10-09T10:14:52.190","ParentId":null,"OwnerUserId":"5143695","Title":null,"Body":"<p>From the keras <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>2D convolution layer (e.g. spatial convolution over images).<\/p>\n  \n  <p>This layer creates a convolution kernel that is convolved with the layer input to produce a tensor of outputs. If  use_bias is True, a bias vector is created and added to the outputs. Finally, if activation is not None, it is applied to the outputs as well.<\/p>\n  \n  <p>When using this layer as the first layer in a model, provide the keyword argument input_shape (tuple of integers, does not include the sample axis), e.g. input_shape=(128, 128, 3) for 128x128 RGB pictures in  data_format=\"channels_last\".<\/p>\n<\/blockquote>\n\n<p>What you have is 19920 samples and 64 features, assuming that this is correct, you should probably go for a 1D convolutional layer instead. The 1D convolutional layer takes a 2D input, it is the kernel itself that is 1D<\/p>\n\n<p>you will probably need to do some changes to ensure that your data is in the correct format as well, as the Conv1D takes the following: <\/p>\n\n<blockquote>\n  <p>Input shape\n  3D tensor with shape: (batch, steps, channels)<\/p>\n<\/blockquote>\n"},{"AnswerId":"52722610","CreationDate":"2018-10-09T13:44:37.963","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Convolutions are used to exploit the fact that spatial locality matters, e.g. near pixel are important to each other to find edges or classify something. This is seldomn the case with tabular data. If you still want to use NN to solve this you might want to use MLPs, e.g. dense layer for your task. Then you do not have to do anything apart from removing the convolutional part.<\/p>\n\n<p>If your data is spatially connected then you might want to use Conv1D layers. As a previos post explains the input shape is a 3D tensor with (Batch, steps, channels). The amount of parameters create is dependant on the number of input channels and number of output channels and irrelevant of the step parameter. E.g. reshaping your data to be (19920, 1, 64) will produce a dense layer. The other extreme is (19920, 64, 1) where the amount of parameters is solely dependand on kernel size and outgoing channel maps.<\/p>\n"}]}
+{"QuestionId":52719877,"AnswerCount":0,"Tags":"<image><image-processing><keras>","CreationDate":"2018-10-09T11:23:35.277","AcceptedAnswerId":null,"OwnerUserId":8444226.0,"Title":"How to customize .flow method of ImageDataGenerator (in Image processing) in keras","Body":"<p>.flow function <strong>reads in all the images at once<\/strong> and then generates augmented data in batches. I have a memory concern and is it possible to manipulate the .flow function's source code in keras and then use it, <strong>so that it could read in images in batches and then augment them and generate a batch of images for training and validation purposes.<\/strong> I want to use it with train_on_batch function for training.<\/p>\n","answers":[]}
+{"QuestionId":52719906,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><google-cloud-platform>","CreationDate":"2018-10-09T11:25:15.957","AcceptedAnswerId":null,"OwnerUserId":5068335.0,"Title":"Memory allocation error on Google cloud platform VM Instance for Keras Model","Body":"<p>I am trying to deploy my trained model on GCP VM Instance. It is around 94mb in size. The code worked perfectly on my local machine. I was able to start the flask server,send an image through cURL call and get the result also.\nBut on GCP, whenever i try to load my model, I get this error <code>allocation of 52953088 exceeds 10% of system memory<\/code> and the process gets killed.<\/p>\n\n<p>The same model loads safely and runs in my local machine. \nShould I increase the memory of the VM instance or is there any other method?<\/p>\n","answers":[{"AnswerId":"52739264","CreationDate":"2018-10-10T11:31:27.720","ParentId":null,"OwnerUserId":"5068335","Title":null,"Body":"<p>I was able to solve the error by creating a new instance with more memory. Increasing the memory from micro(0.6gb apprx) to small(1.7gb) resolved the error.<\/p>\n"}]}
+{"QuestionId":52720886,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-09T12:17:31.080","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Does tf.contrib.nn.sampled_sparse_softmax_loss allow for float16 training?","Body":"<p>Currently, tf.nn.sampled_softmax_loss does not allow for using float16. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/52711895\/how-to-run-define-tensorflow-graph-were-all-variables-are-in-float16-instead-ins\">How to run define Tensorflow graph were all variables are in float16 instead instead of float32<\/a><\/p>\n\n<p>I'm looking at <\/p>\n\n<p>tf.contrib.nn.sampled_sparse_softmax_loss <\/p>\n\n<p>from <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/nn\/sampled_sparse_softmax_loss\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/nn\/sampled_sparse_softmax_loss<\/a><\/p>\n\n<p>And it seems that this may allow float16 values. This code for this function is here<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/contrib\/nn\/python\/ops\/sampling_ops.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/contrib\/nn\/python\/ops\/sampling_ops.py<\/a><\/p>\n\n<p>Which seems to use tf.nn.sparse_softmax_cross_entropy_with_logits<\/p>\n\n<p>Which seems to have support for float16<\/p>\n\n<blockquote>\n  <p>logits: Unscaled log probabilities of shape [d_0, d_1, ..., d_{r-1}, num_classes] and dtype float16, float32, or float64.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sparse_softmax_cross_entropy_with_logits<\/a><\/p>\n\n<p>However, when I try to use it, I get an error<\/p>\n\n<p>My code<\/p>\n\n<pre><code>import math\nimport numpy as np\nimport tensorflow as tf\n\nvocabulary_size = 10\nbatch_size = 64 \nembedding_size = 100 \nnum_inputs =4\nnum_sampled = 128 \n\ngraph = tf.Graph()\n\nwith graph.as_default(): #took out \" , tf.device('\/cpu:0')\"\n\n\n    train_dataset = tf.placeholder(tf.int32, shape=[batch_size, num_inputs ])\n    train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])\n\n    embeddings = tf.get_variable( 'embeddings', dtype=tf.float16,\n        initializer= tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0, dtype=tf.float16) )\n\n    softmax_weights = tf.get_variable( 'softmax_weights', dtype=tf.float16,\n        initializer= tf.truncated_normal([vocabulary_size, embedding_size],\n                             stddev=1.0 \/ math.sqrt(embedding_size), dtype=tf.float16 ) )\n\n    softmax_biases = tf.get_variable('softmax_biases', dtype=tf.float16,\n        initializer= tf.zeros([vocabulary_size], dtype=tf.float16),  trainable=False )\n\n    embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n    embed_reshaped = tf.reshape( embed, [batch_size*num_inputs, embedding_size] )\n\n    segments= np.arange(batch_size).repeat(num_inputs)\n\n    averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n    sam_sof_los = tf.contrib.nn.sampled_sparse_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds,\n                                   labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)\n\n    loss = tf.reduce_mean( sam_sof_los )\n\n    optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) \n\n    saver = tf.train.Saver()\n<\/code><\/pre>\n\n<p>My error message<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    509                 as_ref=input_arg.is_ref,\n--&gt; 510                 preferred_dtype=default_dtype)\n    511           except TypeError as err:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx)\n   1143     if ret is None:\n-&gt; 1144       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n   1145 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py in _TensorTensorConversionFunction(t, dtype, name, as_ref)\n    980         \"Tensor conversion requested dtype %s for Tensor with dtype %s: %r\" %\n--&gt; 981         (dtype.name, t.dtype.name, str(t)))\n    982   return t\n\nValueError: Tensor conversion requested dtype float16 for Tensor with dtype float32: 'Tensor(\"sampled_sparse_softmax_loss\/Log:0\", shape=(64, 1), dtype=float32)'\n\nDuring handling of the above exception, another exception occurred:\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-11-b68afd94e9bd&gt; in &lt;module&gt;()\n     41 \n     42     sam_sof_los = tf.contrib.nn.sampled_sparse_softmax_loss(weights=softmax_weights , biases=softmax_biases , inputs=averaged_embeds,\n---&gt; 43                                    labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)\n     44 \n     45 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/contrib\/nn\/python\/ops\/sampling_ops.py in sampled_sparse_softmax_loss(weights, biases, labels, inputs, num_sampled, num_classes, sampled_values, remove_accidental_hits, partition_strategy, name)\n    331       remove_accidental_hits=remove_accidental_hits,\n    332       partition_strategy=partition_strategy,\n--&gt; 333       name=name)\n    334 \n    335   # There is only one true label. _compute_sampled_logits puts the true logit\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py in _compute_sampled_logits(weights, biases, labels, inputs, num_sampled, num_classes, num_true, sampled_values, subtract_log_q, remove_accidental_hits, partition_strategy, name, seed)\n   1126     if subtract_log_q:\n   1127       # Subtract log of Q(l), prior probability that l appears in sampled.\n-&gt; 1128       true_logits -= math_ops.log(true_expected_count)\n   1129       sampled_logits -= math_ops.log(sampled_expected_count)\n   1130 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/math_ops.py in binary_op_wrapper(x, y)\n    860     with ops.name_scope(None, op_name, [x, y]) as name:\n    861       if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):\n--&gt; 862         return func(x, y, name=name)\n    863       elif not isinstance(y, sparse_tensor.SparseTensor):\n    864         try:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gen_math_ops.py in sub(x, y, name)\n   8316   if _ctx is None or not _ctx._eager_context.is_eager:\n   8317     _, _, _op = _op_def_lib._apply_op_helper(\n-&gt; 8318         \"Sub\", x=x, y=y, name=name)\n   8319     _result = _op.outputs[:]\n   8320     _inputs_flat = _op.inputs\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    544                   \"%s type %s of argument '%s'.\" %\n    545                   (prefix, dtypes.as_dtype(attrs[input_arg.type_attr]).name,\n--&gt; 546                    inferred_from[input_arg.type_attr]))\n    547 \n    548           types = [values.dtype]\n\nTypeError: Input 'y' of 'Sub' Op has type float32 that does not match type float16 of argument 'x'.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52721018,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2018-10-09T12:24:42.637","AcceptedAnswerId":null,"OwnerUserId":1978965.0,"Title":"Keras fit_generator raise You must compile your model before using it Error","Body":"<p>I try to build a CNN + LSTM model by Keras to train a model for video classification task. Firstly, A simple model was built and trained with mock data, 'fit()' api, also it works!<\/p>\n\n<p>But actually, what is used to train this model is a  video dataset, it's so big that can't be loaded into memory. so I need a generator, in this place, I also write a mock generator which generate data with shape same as mock data method. Also, fit_generator API replaces fit.<\/p>\n\n<p>When I run train_gen function, I get following error:<\/p>\n\n<pre><code>File \"lstm.py\", line 48, in &lt;module&gt;\n    train_gen()\n  File \"lstm.py\", line 45, in train_gen\n    model.fit_generator(data_generator.mock_generator(batch_size=32, num_classes=16), steps_per_epoch=1000, epochs=20)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training.py\", line 1415, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training_generator.py\", line 39, in fit_generator\n    model._make_train_function()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training.py\", line 485, in _make_train_function\n    raise RuntimeError('You must compile your model before using it.')\nRuntimeError: You must compile your model before using it.\n<\/code><\/pre>\n\n<p>I tried to solve this problem by searching in stack overflow and google, and do not find exact solution. There are some similar questions can be solved by specifying LSTM <code>input_shape<\/code> and <code>output size<\/code>, or add <code>model.compile()<\/code>.<\/p>\n\n<p>The following snippet code is runnable, and completely same as that mentioned above.<\/p>\n\n<pre><code>import keras\nfrom keras.models import Sequential\nfrom keras.layers import Input, Embedding, LSTM, Dense, Reshape\nfrom keras.layers import Dense, Dropout, Flatten, Activation\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.layers.wrappers import Bidirectional\n\nfrom keras.optimizers import Adam\nimport numpy as np\nimport os, random\n\nos.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"   # see issue #152\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"\"\n\nuse_dropout = True\nmetrics = ['accuracy']\n\n#mock data generator\ndef mock_generator(batch_size, input_length, input_dims, num_classes=16):\n    while True:\n        yield np.random.random((batch_size, input_length, input_dims)), keras.utils.to_categorical(np.random.randint(num_classes, size=(batch_size, 1)), num_classes=num_classes)\n\n#mock data with shape as data generator\ndef mock_data(batch_size, input_length, input_dims, num_classes=16):\n    if True:\n        return np.random.random((batch_size, input_length, input_dims)), keras.utils.to_categorical(np.random.randint(num_classes, size=(batch_size, 1)), num_classes=num_classes)\n\n#construct model, lstm units is fixed\ndef bi_lstm(input_shape, num_classes=16):\n    model = Sequential()\n    model.add(Bidirectional(LSTM(100, return_sequences=True, activation='relu', input_shape=input_shape), merge_mode='concat'))\n    model.add(Bidirectional(LSTM(100, activation='relu', input_shape=(input_shape[0],100), return_sequences=False), merge_mode='concat'))\n    if use_dropout:\n        model.add(Dropout(0.5))\n    model.add(Dense(num_classes))\n    model.add(BatchNormalization())\n    model.add(Activation('softmax'))\n    optimizer = Adam(lr=1e-5, decay=1e-6)\n    model.compile(loss='categorical_crossentropy', optimizer=optimizer,\n                           metrics=metrics)\n    return model\n\n#fit api run successfully\ndef train():\n    input_length=10\n    input_dims=128\n    num_classes=10\n    model = bi_lstm((input_length, input_dims), num_classes)\n    x_train, y_train = mock_data(32, input_length, input_dims, num_classes)\n    model.fit(x_train, y_train, epochs=20, batch_size=32)\n\n#fit_generator api raise error!\ndef train_gen():\n    input_length=10\n    input_dims=128\n    num_classes=10\n    model = bi_lstm((input_length, input_dims), num_classes)\n    generator = mock_generator(32, input_length, input_dims, num_classes)\n    model.fit_generator(generator, steps_per_epoch=1000, epochs=20)\n\n#test mock generator function\ndef test_mock_gen():\n    result  = mock_generator(32,10,128,16)\n    for i in range(2):\n        x, y = result.next()\n        print x.shape\n        print y.shape\n\nif __name__ == '__main__':\n    train()\n    train_gen()\n    #test_mock_gen()\n<\/code><\/pre>\n","answers":[{"AnswerId":"54168021","CreationDate":"2019-01-13T10:36:34.280","ParentId":null,"OwnerUserId":"3103059","Title":null,"Body":"<p>The error you are getting couldn't be more explicit: <strong>model<\/strong> has been declared as a <strong>local object on the train() function<\/strong> and <strong>is not know by the function train_gen()<\/strong>. Define your model as a global variable (ex. in main) and it will work.<\/p>\n"}]}
+{"QuestionId":52721359,"AnswerCount":0,"Tags":"<caffe><importerror>","CreationDate":"2018-10-09T12:42:11.327","AcceptedAnswerId":null,"OwnerUserId":10385423.0,"Title":"Caffe Import error: libopencv_highgui.so.3.3: cannot open shared object file: No such file or directory","Body":"<p>I created a virtual environment for python2.7 and I installed caffe using the command <strong>conda install -c conda-forge caffe<\/strong>. But when I try to import caffe, its throwing the following error.<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nFile \"\/data\/home\/eex608\/.conda\/envs\/MyPython2.7Caffe\/lib\/python2.7\/site-packages\/caffe\/__init__.py\", line 1, in &lt;module&gt;\nfrom .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer\nFile \"\/data\/home\/eex608\/.conda\/envs\/MyPython2.7Caffe\/lib\/python2.7\/site-packages\/caffe\/pycaffe.py\", line 13, in &lt;module&gt;\nfrom ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\\nImportError: libopencv_highgui.so.3.3: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>Can anyone tell me how to resolve this error?<\/p>\n\n<p>The package info after I installed caffe (if this info is useful to debug)<\/p>\n\n<pre><code>2018-10-08 23:07:03  (rev 1)\n anaconda  {5.3.0 -&gt; custom}\n blas  {1.0 -&gt; 1.1 (conda-forge)}\n ca-certificates  {2018.03.07 -&gt; 2018.8.24 (conda-forge)}\n certifi  {2018.8.24 -&gt; 2018.8.24 (conda-forge)}\n h5py  {2.8.0 -&gt; 2.8.0 (conda-forge)}\n harfbuzz  {1.8.8 -&gt; 1.7.6}\n hdf5  {1.10.2 -&gt; 1.10.1 (conda-forge)}\n matplotlib  {2.2.3 -&gt; 2.2.3 (conda-forge)}\n mkl-service  {1.1.2 -&gt; 1.1.2}\n mkl_fft  {1.0.4 -&gt; 1.0.6 (conda-forge)}\n mkl_random  {1.0.1 -&gt; 1.0.1 (conda-forge)}\n numexpr  {2.6.8 -&gt; 2.6.6 (conda-forge)}\n numpy  {1.15.1 -&gt; 1.15.2 (conda-forge)}\n numpy-base  {1.15.1 -&gt; 1.15.2}\n openssl  {1.0.2p -&gt; 1.0.2p (conda-forge)}\n pyqt  {5.9.2 -&gt; 5.6.0}\n pytables  {3.4.4 -&gt; 3.4.4 (conda-forge)}\n qt  {5.9.6 -&gt; 5.6.3}\n qtconsole  {4.4.1 -&gt; 4.4.1 (conda-forge)}\n scikit-learn  {0.19.2 -&gt; 0.20.0 (conda-forge)}\n scipy  {1.1.0 -&gt; 1.1.0 (conda-forge)}\n+boost-1.66.0 (conda-forge)\n+boost-cpp-1.66.0 (conda-forge)\n+caffe-1.0 (conda-forge)\n+ffmpeg-3.2.4 (conda-forge)\n+gflags-2.2.1 (conda-forge)\n+giflib-5.1.4 (conda-forge)\n+glog-0.3.5 (conda-forge)\n+jasper-1.900.1 (conda-forge)\n+leveldb-1.18 (conda-forge)\n+libgcc-7.2.0 (conda-forge)\n+libgfortran-3.0.0 (conda-forge)\n+libiconv-1.15 (conda-forge)\n+libopenblas-0.3.3\n+libprotobuf-3.5.2 (conda-forge)\n+libwebp-0.5.2 (conda-forge)\n+lmdb-0.9.22 (conda-forge)\n+openblas-0.2.20 (conda-forge)\n+opencv-3.4.1 (conda-forge)\n+protobuf-3.5.2 (conda-forge)\n+python-gflags-3.1.2 (conda-forge)\n+python-leveldb-0.193 (conda-forge)\n+x264-1!152.20180717 (conda-forge)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52721616,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-09T12:55:57.933","AcceptedAnswerId":52723992.0,"OwnerUserId":4729764.0,"Title":"Dimensionality of the `Y` input for LSTM in Keras","Body":"<p>What dimension should the LSTM output be? Currently I'm using <code>return_sequences=False<\/code> and <code>model.summary()<\/code> suggests that the <code>Dense(1, activation='sigmoid')<\/code> layer should have output:<\/p>\n\n<p><code>(None, 1)<\/code><\/p>\n\n<p>Where <code>None<\/code> is the batch size. <\/p>\n\n<p>Does it mean that my <code>y<\/code> also has to be 2 dimensional? <\/p>\n\n<p>Currently I'm using a 1d array of the shape <code>y.shape = (None,)<\/code> which seems to work fine, but I wonder if I should add a new axis to match the dimensions as:\n<code>y = y[:, np.newaxis]<\/code><\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Below is my model summary.<\/p>\n\n<p>Essentially I'm modelling a binary classifier (0 or 1) and I am feeding batches of variable size in both <code>batch_size<\/code> and <code>time steps<\/code>. Hence the dimension of <code>X.shape = (None, None, no_of_features)<\/code>. <\/p>\n\n<p>As we can see below, the output of the <code>Dense()<\/code> layer is <code>(None, 1)<\/code> where as my <code>Y<\/code> is actually <code>(Batch_size, )<\/code> , a 1d numpy array. Should I add a new axis so that <code>Y<\/code> becomes 2d? E.g. <code>(batch_size, 1)<\/code>? <\/p>\n\n<p>I tried both, and both of them seem to work. So what's the difference?<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nconcatenate_8 (Concatenate)     (None, None, 285)    0           numeric_input[0][0]              \n                                                                 concatenate_7[0][0]              \n__________________________________________________________________________________________________\ndropout_3 (Dropout)             (None, None, 285)    0           concatenate_8[0][0]              \n__________________________________________________________________________________________________\nlstm_4 (LSTM)                   (None, 64)           89600       dropout_3[0][0]                  \n__________________________________________________________________________________________________\nmain_output (Dense)             (None, 1)            65          lstm_4[0][0]                     \n==================================================================================================\n<\/code><\/pre>\n","answers":[{"AnswerId":"52723992","CreationDate":"2018-10-09T14:55:06.217","ParentId":null,"OwnerUserId":"1097517","Title":null,"Body":"<p>Both are equivalent and either way works. The reason being is that the  two shapes -- <code>(None,1)<\/code> and <code>(None,)<\/code> -- are compatible and are broadcasted to the same shape automatically. Which is equivalent to you manually adding a new dimension of size 1. <a href=\"https:\/\/learningtensorflow.com\/broadcasting\/\" rel=\"nofollow noreferrer\">Here<\/a> is some additional reading on broadcasting in tensorflow. <\/p>\n"}]}
+{"QuestionId":52723253,"AnswerCount":0,"Tags":"<python><keras><embedding>","CreationDate":"2018-10-09T14:18:20.267","AcceptedAnswerId":null,"OwnerUserId":10456792.0,"Title":"Why keras embeding layer has no keys","Body":"<pre><code>self.embed = Sequential([Embedding(9488, output_dim=512,input_length=14),\n   Activation('relu'),\n  Dropout(0.5)], name='embed.0')\n\ninputs_bedding = Input(shape=(10,)) \nxt = self.embed(inputs_bedding) \n\n=============================\nkeras_model.save('temp.h5')\nwith h5py.File('temp.h5', 'a') as f:\nmodel_weights = f['model_weights']\nparams = util.dig_to_params(model_weights[layer])\n\ndef dig_to_params(keras_h5_layer): \n    while not _contains_weights(keras_h5_layer): \n        keras_h5_layer = keras_h5_layer[list(keras_h5_layer.keys())[0]]\n\n    return keras_h5_layer\n<\/code><\/pre>\n\n<blockquote>\n  <p>Error: <code>keras_h5_layer = keras_h5_layer[list(keras_h5_layer.keys())[0]]<\/code><\/p>\n  \n  <p>AttributeError: 'Dataset' object has no attribute 'keys'<br><\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52723416,"AnswerCount":0,"Tags":"<python><tensorflow><keras><har>","CreationDate":"2018-10-09T14:27:03.563","AcceptedAnswerId":null,"OwnerUserId":10479062.0,"Title":"How to import data window sensor data into python to be fed to a CNN for HAR","Body":"<p>I am currently working on a project to classify different upper body exercises using 2 gyroscopes and 1 accelerometer (9 channels of data). My approach is to use a sliding window approach for 6 exercises, and feed the windows to the CNN. The data is collected at 25Hz, and I created windows of 64 samples (2.56 seconds, with 50% overlap -> 32 samples) 9 channels. The dimensions of the windows are thereby 9x64 (Rows correspond to the different sensors and the columns to the readings), the reason for this is to be able to perform the convolution operations in the time domain (horizontal axis). With that being said I have created a small database with 60 windows per exercise, for a total of 360 windows. The data set contains 2 folders (Train, Test), where each of these two folders has 6 folders within (Each corresponding to the different exercises). In each exercise folder I have 48 windows for the Training samples and 12 windows for the testing samples. Each window is saved as a CSV file, with the dimensions of 9x64. My question is, what is the best solution to import the data using keras into the CNN. I have done a similar project using JPEG images, using the ImageDataGenerator prepossessing class from keras. For this project I had two folders (Test, Train) and each folder contain a folder for Cat and Dog pictures. My question is if there's a similar approach to import the data using CSV files :) <\/p>\n","answers":[]}
+{"QuestionId":52723671,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-09T14:39:52.843","AcceptedAnswerId":52725027.0,"OwnerUserId":10479075.0,"Title":"TensorFlow the variable cant be initialized","Body":"<pre><code>X=tf.placeholder(tf.float32,[None,32,32,3])\ny=tf.placeholder(tf.int64,[None])\nis_training=tf.placeholder(tf.bool)\n\ndef simple_model(X,y):\n\n    Wconv1=tf.get_variable(\"Wconv1\",shape=[7,7,3,32],use_resource=True)\n    bconv1=tf.get_variable('bconv1',shape=[32])\n    W1=tf.get_variable('W1',shape=[5408,10])\n    b1=tf.get_variable('b1',shape=[10])\n\n    a1=tf.nn.conv2d(X,Wconv1,[1,2,2,1],'VALID')+bconv1\n    h1=tf.nn.relu(a1)\n\n    h1_flat=tf.reshape(h1,[-1,5408])\n    y_out=tf.matmul(h1_flat,W1)+b1\n    return y_out\n\ninit=tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n\n    sess.run(simple_model(X,y),feed_dict={X:X_train,y:y_train})\n<\/code><\/pre>\n\n<p>error is<\/p>\n\n<blockquote>\n  <p>PreconditionError Attempting to use uninitialized variable <code>Wconv1<\/code><\/p>\n<\/blockquote>\n\n<p>I dont know what wrong with code?<\/p>\n","answers":[{"AnswerId":"52725027","CreationDate":"2018-10-09T15:57:37.783","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/global_variables_initializer\" rel=\"nofollow noreferrer\"><code>tf.global_variables_initializer<\/code><\/a> makes an initialization op for all the global variables created <strong>up to that point<\/strong>. This means that if you create other variables later, they will not be initialized by the operation. This is because variable initializers just hold a list of the variables they have to initialize, and this does not change as you add more variables (in fact, <code>tf.global_variables_initializer()<\/code> is just a shortcut for <code>tf.variables_initializer(tf.global_variables())<\/code> or <code>tf.variables_initializer(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES))<\/code>). In your case, the variables are being created in the second call to <code>sess.run<\/code>, after you create <code>init<\/code> earlier. You need to create the initialization operation <em>after<\/em> you have created your model with the variables:<\/p>\n\n<pre><code>X=tf.placeholder(tf.float32,[None,32,32,3])\ny=tf.placeholder(tf.int64,[None])\nis_training=tf.placeholder(tf.bool)\n\ndef simple_model(X,y):\n\n    Wconv1=tf.get_variable(\"Wconv1\",shape=[7,7,3,32],use_resource=True)\n    bconv1=tf.get_variable('bconv1',shape=[32])\n    W1=tf.get_variable('W1',shape=[5408,10])\n    b1=tf.get_variable('b1',shape=[10])\n\n    a1=tf.nn.conv2d(X,Wconv1,[1,2,2,1],'VALID')+bconv1\n    h1=tf.nn.relu(a1)\n\n    h1_flat=tf.reshape(h1,[-1,5408])\n    y_out=tf.matmul(h1_flat,W1)+b1\n    return y_out\n\nmy_model = simple_model(X,y)\ninit=tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n\n    sess.run(my_model, feed_dict={X:X_train,y:y_train})\n<\/code><\/pre>\n"}]}
+{"QuestionId":52723753,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-09T14:43:38.633","AcceptedAnswerId":52724475.0,"OwnerUserId":9618242.0,"Title":"Keras model not saving correctly","Body":"<p>I am training a neural network in Keras but when new data comes and I try to retrain it, the loss in the epochs it's as high as the first time I trained my model.  <\/p>\n\n<pre><code> checkpoint = ModelCheckpoint('my_model.h5', monitor='loss', verbose=1, save_best_only=True, mode='min')\n callbacks_list = [checkpoint]\n model.fit(X_train,y_train, batch_size = batch_size, epochs = epochs, callback = callbacks_list)\n new_model = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>As suggested here <a href=\"https:\/\/stackoverflow.com\/questions\/45393429\/keras-how-to-save-model-and-continue-training?noredirect=1&amp;lq=1\">Keras: How to save model and continue training?<\/a> I tried to predict the same data both in <code>model<\/code> and <code>new_model<\/code> and measure the differences with: <\/p>\n\n<pre><code>assert_allclose(model.predict(x_train),\n            new_model.predict(x_train),\n            1e-5)\n<\/code><\/pre>\n\n<p>In fact, I got the Assertion Error and I did even with <code>tol = 1e-2<\/code> so that makes me think my model is not loading as it should. Anyone has a clue of why is this happening?<\/p>\n","answers":[{"AnswerId":"52724475","CreationDate":"2018-10-09T15:24:12.310","ParentId":null,"OwnerUserId":"9179860","Title":null,"Body":"<p><code>ModelCheckpoint<\/code> saves the model weighs that has had less loss in training.<\/p>\n\n<p>Your <code>model<\/code> has saved the last epoch weighs.<\/p>\n\n<p>If the last epoch of your model is not the one that has had less loss, the weights of the saved model (<code>new_model<\/code>) do not match those of the original model and the prediction is not the same.<\/p>\n"}]}
+{"QuestionId":52724006,"AnswerCount":0,"Tags":"<python><scikit-learn><keras>","CreationDate":"2018-10-09T14:55:49.300","AcceptedAnswerId":null,"OwnerUserId":10478172.0,"Title":"Replicating sklearn.neural_network.MPLRegressor with Keras","Body":"<p>I'm currently working with sklearn.neural_network.MPLRegressor but looking for more fexibility in the definition of the neural network, I tried Keras.\nI create the following dataset <\/p>\n\n<pre><code>import numpy as np\nx = np.linspace(-0.5,5,2000)\ny = np.tanh(x)\ny=y.reshape(-1, )\nx=x.reshape(-1, 1)\n<\/code><\/pre>\n\n<p>and implement the trivial neural network with sklearn<\/p>\n\n<pre><code>from sklearn.neural_network import MLPRegressor\n    model = MLPRegressor(hidden_layer_sizes=(1,),activation='tanh',solver='lbfgs',verbose=True,validation_fraction = 0.1) \nmodel.fit(x,y)\n<\/code><\/pre>\n\n<p>which is solved with a high degree of precision <code>MSE 7.252229995056021e-10<\/code><\/p>\n\n<p>I replicate the same nn with Keras<\/p>\n\n<pre><code>from keras.layers import Input, Dense\nfrom keras.models import Model       \nimport keras\nfrom keras.wrappers.scikit_learn import KerasRegressor    \nkeras.initializers.RandomUniform()    \nvisible = Input(shape=(1,))\nhidden1 = Dense(1, activation='tanh')(visible)    \noutput = Dense(1,activation=\"linear\")(hidden1)    \nmodel = Model(inputs=visible, outputs=output)\nmodel.compile(optimizer='adam',metrics=['mean_squared_error'],loss =['mean_squared_error'] )\nmodel.fit(x, y,validation_split=0.1)\n<\/code><\/pre>\n\n<p>but I obtain <code>MSE 0.02130996414874087<\/code> (the result doesn't change a lot using different optimizers).<\/p>\n\n<p>Since the nn can be solved analytically, I would expect a higher degree of precision.\nDoes anyone know the reason behind such difference, if any (apart from the different optimizer)?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52724022,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-09T14:56:57.730","AcceptedAnswerId":52893880.0,"OwnerUserId":5902284.0,"Title":"Model fits on a single GPU but script crashes when trying to fit on multiple GPUs","Body":"<p>I have a model that can train fine on a single GPU, but when I try to fit it using multi_gpu_model, I get this CUDA error before the script exits:<\/p>\n\n<pre><code>F tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:521] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 0 feature_map_count: 16 spatial: 128 128 128  value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX}\n<\/code><\/pre>\n\n<p>I tried to pass both the compiled and not-compiled version of the model instance to the multi_gpu_model function, but it did not change anything. I call it like this:<\/p>\n\n<pre><code>multi_model = multi_gpu_model(model, gpus=4)\n<\/code><\/pre>\n\n<p>The compilation is done this way and does not raise any errors:<\/p>\n\n<pre><code>multi_model.compile(\n    optimizer=keras.optimizers.Adam(5e-4),\n    loss=dice_coefficient_loss,\n    metrics=[dice_coefficient]\n            + get_label_wise_dice_coefficient_functions(n_labels))\n\ndef dice_coefficient(y_true, y_pred, smooth=1.):\n    y_true_f = K.flatten(y_true)\n    y_pred_f = K.flatten(y_pred)\n    intersection = K.sum(y_true_f * y_pred_f)\n    return ((2. * intersection + smooth)\n            \/ (K.sum(y_true_f) + K.sum(y_pred_f) + smooth))\n\n\ndef dice_coefficient_loss(y_true, y_pred):\n    return -dice_coefficient(y_true, y_pred)\n\n\ndef label_wise_dice_coefficient(y_true, y_pred, label_index):\n    return dice_coefficient(y_true[:, label_index], y_pred[:, label_index])\n\n\ndef get_label_dice_coefficient_function(label_index):\n    f = functools.partial(label_wise_dice_coefficient, label_index=label_index)\n    f.__setattr__('__name__', 'label_{0}_dice_coef'.format(label_index))\n    return f\n\n\ndef get_label_wise_dice_coefficient_functions(n_labels):\n    return [get_label_dice_coefficient_function(i) for i in range(n_labels)]\n<\/code><\/pre>\n\n<p>(most of these functions and the model architecture have been stolen <a href=\"https:\/\/github.com\/ellisdg\/3DUnetCNN\/blob\/master\/unet3d\/metrics.py\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<p>I am using python 3.6.6, tensorflow-gpu 1.10.0, cudatoolkit 9.2, cudnn 7.2.1 from conda main repo and keras-contrib 2.0.8 installed with pip\/git on top of a 64-bits CentOS 7.4.1708<\/p>\n\n<p>Looking at the previous log lines, it looks like the multiple GPUs are properly detected:<\/p>\n\n<pre><code>2018-10-09 16:30:19.977993: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 0 with properties:\nname: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\npciBusID: 0000:20:00.0\ntotalMemory: 10.92GiB freeMemory: 10.74GiB\n2018-10-09 16:30:20.318137: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 1 with properties:\nname: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\npciBusID: 0000:21:00.0\ntotalMemory: 10.92GiB freeMemory: 10.74GiB\n2018-10-09 16:30:20.595428: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 2 with properties:\nname: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\npciBusID: 0000:22:00.0\ntotalMemory: 10.92GiB freeMemory: 10.74GiB\n2018-10-09 16:30:20.953619: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 3 with properties:\nname: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\npciBusID: 0000:23:00.0\ntotalMemory: 10.92GiB freeMemory: 10.74GiB\n2018-10-09 16:30:20.967429: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1484] Adding visible gpu devices: 0, 1, 2, 3\n2018-10-09 16:30:22.415906: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:965] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-09 16:30:22.415957: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:971]      0 1 2 3\n2018-10-09 16:30:22.415965: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] 0:   N Y Y Y\n2018-10-09 16:30:22.415971: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] 1:   Y N Y Y\n2018-10-09 16:30:22.415982: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] 2:   Y Y N Y\n2018-10-09 16:30:22.415988: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] 3:   Y Y Y N\n2018-10-09 16:30:22.416681: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1097] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 10393 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:20:00.0, compute capability: 6.1)\n2018-10-09 16:30:22.536003: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1097] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:1 with 10393 MB memory) -&gt; physical GPU (device: 1, name: GeForce GTX 1080 Ti, pci bus id: 0000:21:00.0, compute capability: 6.1)\n2018-10-09 16:30:22.637811: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1097] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:2 with 10393 MB memory) -&gt; physical GPU (device: 2, name: GeForce GTX 1080 Ti, pci bus id: 0000:22:00.0, compute capability: 6.1)\n2018-10-09 16:30:22.747698: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1097] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:3 with 10393 MB memory) -&gt; physical GPU (device: 3, name: GeForce GTX 1080 Ti, pci bus id: 0000:23:00.0, compute capability: 6.1)\n2018-10-09 16:30:25,557.557:__main__:INFO:Compiling model\n2018-10-09 16:30:25,634.634:__main__:INFO:Fitting model\n2018-10-09 16:31:31.773355: F tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:521] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 0 feature_map_count: 16 spatial: 128 128 128  value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX}\n\/bin\/bash: line 1: 160691 Aborted\n<\/code><\/pre>\n\n<p>Any help on what I did wrong would be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"52893880","CreationDate":"2018-10-19T13:58:31.317","ParentId":null,"OwnerUserId":"5902284","Title":null,"Body":"<p>It turned out that the <code>.fit()<\/code> method of a <code>multi_model_gpu<\/code> did not like it when the number of samples in the dataset was not a multiple of the <code>batch_size<\/code>, i.e., the number of GPUs in my case. Ditching a sample from my dataset solved my issue. I reported this bug <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/11434\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52724147,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-09T15:04:14.550","AcceptedAnswerId":null,"OwnerUserId":2815551.0,"Title":"Creating a subgraph of a graph in TensorFlow given a condition","Body":"<p>In <code>tf.slim<\/code> or the standard <code>TensorFlow<\/code>; can we save a subgraph of a graph given a condition? Say, we would like to generate (or save as a new graph) from all non-zero elements within each layer? Say <code>conv1<\/code> has this weight array:<\/p>\n\n<pre><code>conv2 = [ 1 0 2 0 0 3 4]\n<\/code><\/pre>\n\n<p>These weights are interconnected with the preceding and the succeeding layer. Can I downsize the layer to:<\/p>\n\n<pre><code>conv2_new = [ 1 2 3 4]\n<\/code><\/pre>\n\n<p>As if I dropped the connections and the weights of certain elements in each layer and the executing flow adapts itself given the new dimension of the layer? Any suggestions\/code would be appreciated. <\/p>\n","answers":[]}
+{"QuestionId":52725099,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><nvidia>","CreationDate":"2018-10-09T16:01:47.200","AcceptedAnswerId":null,"OwnerUserId":2130515.0,"Title":"Tensorflow cannot start a session on GPU","Body":"<p>I am running Tensorflow on Conda with GPU. Each time I get this error<\/p>\n\n<pre><code>E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:353] Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED\n2018-10-09 11:58:43.850743: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:361] Possibly insufficient driver version: 390.48.0\n<\/code><\/pre>\n\n<p>I am on ubuntu 18.04 Lambda<\/p>\n\n<p>It seems a driver problem. How can I resolve that?<\/p>\n","answers":[{"AnswerId":"52725315","CreationDate":"2018-10-09T16:16:07.697","ParentId":null,"OwnerUserId":"51685","Title":null,"Body":"<blockquote>\n  <p>Possibly insufficient driver version: 390.48.0<\/p>\n<\/blockquote>\n\n<p>sounds like you need to upgrade your Nvidia drivers. The way to do that depends on your distribution and how you've installed the drivers in the first place.<\/p>\n\n<p>You can also try to downgrade your Tensorflow installation.<\/p>\n"}]}
+{"QuestionId":52725712,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><object-detection><object-detection-api><lucid>","CreationDate":"2018-10-09T16:42:59.190","AcceptedAnswerId":null,"OwnerUserId":10479464.0,"Title":"How to use Lucid Interpretability tools on a Tensorflow Detection model?","Body":"<p>I want to use <a href=\"https:\/\/github.com\/tensorflow\/lucid\" rel=\"nofollow noreferrer\">Lucid<\/a> to analyze the feature extraction of a detection model I trained using the tensorflow Object Detection API on my own dataset. The model used is one from the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">Tensorflow Object Detection Zoo<\/a>, namely <code>faster_rcnn_resnet101<\/code>.<\/p>\n\n<p>I followed the <a href=\"https:\/\/colab.research.google.com\/drive\/1PPzeZi5sBN2YRlBmKsdvZPbfYtZI-pHl#scrollTo=EeOS7Fzu4a-p\" rel=\"nofollow noreferrer\">Lucid tutorial<\/a> to import my own model and saved a frozen graph of the model with the node <code>\/all_class_predictions_with_background<\/code> as <code>output_node<\/code>.<\/p>\n\n<p>I'm having trouble finding the input node of the graph to make Lucid run on it.<\/p>\n\n<p>Furthermore I don't really think I have the right approach. Maybe I should first extract all the classification part of the detection model and freeze a new graph with only this part before going to Lucid.<\/p>\n\n<p>Or maybe I should just import a <code>resnet_101<\/code> classification model and copy\/paste the correct weights from the detection model on it?<\/p>\n\n<p>But I don't really know how to do those kind of things.<\/p>\n\n<p>Can someone help me? I really want to try running Lucid on my detection network.<\/p>\n","answers":[{"AnswerId":"58186743","CreationDate":"2019-10-01T14:15:27.120","ParentId":null,"OwnerUserId":"5576984","Title":null,"Body":"<p>Yes, you should export an inference (frozen) graph to work with in Lucid.<\/p>\n\n<p>I use the following script to export a graph from the training checkpoint files.\nUseful information about the nodes in the exported file is logged to the console.<\/p>\n\n<pre><code>training_model=\"ssd_mnet_v2_ppn_512x288.config\"\n\nmodel_signature=\"eb_13_v09_ppmn2_13_256_adam_512x288_tf_1.14_200k\"\n\n# the specific checkpoint to export from\ncheckpoint_path=\"\/TRAIN\/models\/model\/train\/model.ckpt-200000\"\n\n# directory to export into\noutput_path=\"\/XYZ\/graphs\/${model_signature}\"\n\n# ensure these graph nodes are exported, and everything in between\nadditional_output_tensor_names=\"Preprocessor\/sub,concat_1\"\n\n# \npython export_inference_graph.py \\\n   --input_type=image_tensor \\\n   --pipeline_config_path \/TRAIN\/models\/model\/$training_model \\\n   --trained_checkpoint_prefix=$checkpoint_path \\\n   --output_directory=$output_path \\\n   --additional_output_tensor_names=$additional_output_tensor_names\n<\/code><\/pre>\n\n<p>I found it convenient to make my own Lucid Model class, after reviewing the examples in the Lucid model zoo.\nYou have to examine your graph carefully as you need to specify the input node, and provide a list of layers that Lucid can work with.<\/p>\n\n<pre><code>from lucid.modelzoo.vision_base import Model, _layers_from_list_of_dicts\n\n\n# the input node \"Preprocessor\/sub\" is appropriate for image injection\nclass SSD_Mnet2_PPN( Model ):\n\n    def __init__(self, image_shape=None, graph_path=None, labels_path=None ):        \n        self.model_path = graph_path\n        self.labels_path = labels_path\n\n        self.image_shape = image_shape\n        self.image_value_range = (-1, 1) \n        self.input_name = \"Preprocessor\/sub\"\n\n        super().__init__()\n\n# a hand-crafted list of layers - by inspection of the graph\nSSD_Mnet2_PPN.layers = _layers_from_list_of_dicts(SSD_Mnet2_PPN, [\n  { 'id':  0, 'tags': ['conv'], 'name': 'FeatureExtractor\/MobilenetV2\/expanded_conv_2\/add', 'depth': 24, 'shape': [ 1, 72, 128, 24 ], 'transform_id': 2 },  \n  { 'id':  2, 'tags': ['conv'], 'name': 'FeatureExtractor\/MobilenetV2\/expanded_conv_5\/add', 'depth': 32, 'shape': [ 1, 36, 64, 32 ], 'transform_id': 2 },  \n  { 'id':  5, 'tags': ['conv'], 'name': 'FeatureExtractor\/MobilenetV2\/expanded_conv_9\/add', 'depth': 64, 'shape': [ 1, 18, 32, 64 ], 'transform_id': 2 },  \n  { 'id':  7, 'tags': ['conv'], 'name': 'FeatureExtractor\/MobilenetV2\/expanded_conv_12\/add', 'depth': 96, 'shape': [ 1, 18, 32, 96 ], 'transform_id': 2 },  \n  { 'id':  9, 'tags': ['conv'], 'name': 'FeatureExtractor\/MobilenetV2\/expanded_conv_15\/add', 'depth': 160, 'shape': [ 1, 9, 16, 160 ], 'transform_id': 2 },  \n  { 'id': 11, 'tags': ['concat'], 'name': 'concat_1', 'depth': 13, 'shape': [ 1, 1212, 13 ], 'transform_id': 4 },\n])\n\n\ndef model_for_version( version=None, path=None ):\n\n    if \"320x180\" in version:\n        return SSD_Mnet2_PPN( graph_path=path, image_shape=[ 320, 180, 3 ] )\n\n    if \"480x270\" in version:\n        return SSD_Mnet2_PPN( graph_path=path, image_shape=[ 480, 270, 3 ] )\n\n    if \"512x288\" in version:\n        return SSD_Mnet2_PPN( graph_path=path, image_shape=[ 512, 288, 3 ] )        \n\n    if \"720x405\" in version:\n        return SSD_Mnet2_PPN( graph_path=path, image_shape=[ 720, 405, 3 ] )\n\n    raise ValueError( \"No model for graph_version: {}\".format( version ) )\n<\/code><\/pre>\n\n<p>Then you can write code as follows:<\/p>\n\n<pre><code>from lucid.optvis import render\n\nmodel = model_for_version( \n    version = \"eb_13_v09_ppmn2_13_256_adam_512x288_tf_1.14\", \n    path = \"\/XYZ\/graphs\/eb_13_v09_ppmn2_13_256_adam_512x288_tf_1.14_200k\/frozen_inference_graph.pb\" \n)\n\nmodel.load_graphdef()\n\n_ = render.render_vis( model, \"FeatureExtractor\/MobilenetV2\/expanded_conv_15\/add:17\", thresholds=( 32, 256, 1024 ) )    \n<\/code><\/pre>\n\n<p>Inevitably, one has to experiment quite a bit.<\/p>\n"}]}
+{"QuestionId":52725772,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-09T16:46:36.433","AcceptedAnswerId":null,"OwnerUserId":3577054.0,"Title":"Keras: why is `load_model` getting slower each time?","Body":"<p>It seems Keras is slower with each <code>load_model<\/code> call when running multiple of those calls. See related question:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47455397\/extremely-slow-model-load-with-keras\">Extremely slow model load with keras<\/a><\/p>\n\n<p>It can be reproduced with a simple code snippet:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.models import load_model\nfrom keras.layers import Dense\n\n\nmodel = Sequential()\n\nmodel.add(Dense(units=64, activation='relu', input_dim=100))\nmodel.add(Dense(units=10, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='sgd',\n              metrics=['accuracy'])\n\nmodel.save('asdf.h5')\ndel model\n\nfor i in range(100):\n    t0 = time.time()\n    load_model('asdf.h5')\n    print('%.2f' % (time.time() - t0))\n<\/code><\/pre>\n\n<p>Why?<\/p>\n\n<p>I am not interested in knowing how to avoid it, just why it happens.<\/p>\n","answers":[]}
+{"QuestionId":52726225,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-09T17:16:37.130","AcceptedAnswerId":52727000.0,"OwnerUserId":7280300.0,"Title":"Keras class weights","Body":"<p><code>keras<\/code> documentation is a little vague about <code>class_weight<\/code> parameter in <code>model.fit()<\/code>. I have a dataset with binary labels but it's instances per class are imbalanced. <code>Class 0<\/code> has <code>1000<\/code> instances while <code>class 1<\/code> has <code>15,000<\/code> instances. How would I set <code>class_weight<\/code> dictionary in this case? Would it be like <code>class_weight = {0: 0.07, 1: 0.93}<\/code> or there's some other method to calculate these values?<\/p>\n","answers":[{"AnswerId":"52727000","CreationDate":"2018-10-09T18:09:04.883","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>If you set the class weight to {0: 0.07, 1: 0.93} you are telling your model to make the correct classification of class: 1 15 times more important. You can think about it as you would just use the same training instance 15 times to train the model. <\/p>\n"}]}
+{"QuestionId":52726632,"AnswerCount":1,"Tags":"<android><tensorflow><machine-learning><tensorflow-lite><toco>","CreationDate":"2018-10-09T17:43:55.027","AcceptedAnswerId":null,"OwnerUserId":2328964.0,"Title":"TFLite Conversion changing model weights","Body":"<p>I have a custom built tensorflow graph implementing MobileNetV2-SSDLite which I implemented myself. It is working fine on the PC.<\/p>\n\n<p>However, when I convert the model to TFLite (all float, no quantization), the model weights are changed drastically. <\/p>\n\n<p>To give an example, a filter which was initially - \n 0.13172674179077148,\n 2.3185202252437188e-32,\n -0.003990101162344217<\/p>\n\n<p>becomes-\n4.165565013885498,\n-2.3981268405914307,\n-1.1919032335281372<\/p>\n\n<p>The large weight values are completely throwing off my on-device inferences. Need help! :(<\/p>\n","answers":[{"AnswerId":"52748373","CreationDate":"2018-10-10T20:33:29.507","ParentId":null,"OwnerUserId":"1204287","Title":null,"Body":"<p>What command are you using to convert to tflite? For instance are you using toco, and if so what parameters are you using? While I haven't been looking at the filters, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22106#issuecomment-428409506\" rel=\"nofollow noreferrer\">here are my default instructions<\/a> for finetuning a MobileNetV2-SSD and SSDLite graphs and the model has been performing well.<\/p>\n"}]}
+{"QuestionId":52727243,"AnswerCount":0,"Tags":"<python><tensorflow><numpy-broadcasting>","CreationDate":"2018-10-09T18:29:05.450","AcceptedAnswerId":null,"OwnerUserId":1691725.0,"Title":"Covariance Matrix from 2D vectors - Tensorflow, Numpy","Body":"<p>I'm trying to  generate a kernel function for GP using only Matrix operations (no loops).\nVectors where no problem taking advantage of broadcasting<\/p>\n\n<pre><code>def kernel(A,B):\n    return 1\/np.exp(np.linalg.norm(A-B.T))**2\n<\/code><\/pre>\n\n<p>A and B are both [n,1] vectors, but with [n,m] shaped matrices It just doesn't work. (Also tried reshaping to [1,n,m])\nI'm interested on computing an X Matrix where every ij-th element is defined by Ai-Bj.<\/p>\n\n<p>Now I'm working on Numpy but my final objective is implement this on Tensorflow.\nThanks in Advance.<\/p>\n","answers":[]}
+{"QuestionId":52727814,"AnswerCount":2,"Tags":"<python><keras><classification>","CreationDate":"2018-10-09T19:08:39.700","AcceptedAnswerId":52729196.0,"OwnerUserId":345660.0,"Title":"probit regression in keras","Body":"<p>I've been playing around with activation functions for a binary classification model in keras.  For example, I've implemented a logit model with <code>Dense(1, activation='sigmoid')<\/code>, and I've implemented a cloglog model as follows:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\nimport keras.backend as K\nnp.random.seed(42)\n\n# Some Data\nnrows = 1000\nncols = 20\nX = np.random.rand(nrows, ncols) - 0.5\nCF = np.random.rand(ncols, 1)\ny = np.sign(X.dot(CF))\ny[np.where(y == -1)] = 0\n\n# cloglog model\ndef cloglog(x):\n    return -(K.exp(-K.exp(x))-1)\n\ninput = Input(shape=(ncols,))\noutput = Dense(1, activation=cloglog)(input)\nmodel = Model(input, output)\nmodel.compile(loss='binary_crossentropy', optimizer='adam')\nmodel.fit(X, y, epochs=10)\n<\/code><\/pre>\n\n<p>I'd really like to try out a probit model, but I'm struggling to implement the link function with just the operations available in <code>keras.backend<\/code>.<\/p>\n\n<p>Is this going to be possible? Or do I really need the distribution functions (pdf\/cdf) available in <code>scipy<\/code>?<\/p>\n","answers":[{"AnswerId":"52729196","CreationDate":"2018-10-09T20:54:30.107","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>If you use TensorFlow as your Keras backend, you could leverage <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/distributions\" rel=\"nofollow noreferrer\">tf.distributions<\/a> to compute the CDF of the normal distribution:<\/p>\n\n<pre><code>from tensorflow.distributions import Normal\n\ndef probit(x):\n    normal = Normal(loc=0.,\n                    scale=1.)\n    return normal.cdf(x)\n<\/code><\/pre>\n"},{"AnswerId":"57506865","CreationDate":"2019-08-15T08:26:54.097","ParentId":null,"OwnerUserId":"10723149","Title":null,"Body":"<p>I'm dealing with another problem related to approximating a normal CDF \/ inverse-normal CDF. The latter is also called quantile funxtion, probit function or percent point function, which can be found in scipy. So this may be just what you needed a few months ago. :) <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from scipy.stats import norm\nnorm.ppf(x,0,1) \n<\/code><\/pre>\n"}]}
+{"QuestionId":52727882,"AnswerCount":0,"Tags":"<c++><multithreading><tensorflow>","CreationDate":"2018-10-09T19:13:35.140","AcceptedAnswerId":null,"OwnerUserId":2679594.0,"Title":"TensorFlow custom operation: Thread safety of output tensor","Body":"<p>I have a custom operation for TensorFlow which benefits from doing a certain computation in parallel.\nHowever, one user reported a <a href=\"https:\/\/github.com\/githubharald\/CTCWordBeamSearch\/issues\/8\" rel=\"nofollow noreferrer\">crash<\/a> when using this parallel computation (on my computer everything works fine).<\/p>\n\n<ul>\n<li>Is there anything I have to take care in the Compute method when working with threads?<\/li>\n<li>Is reading\/writing tensors already thread-safe or do I have to take care?<\/li>\n<li>Especially, do I have to synchronize my threads when writing to the output tensor? See simplified code snippet below.<\/li>\n<\/ul>\n\n<pre>\n    void Compute(OpKernelContext* context) override \n    {\n        \/\/ output tensor\n        Tensor* outputTensor = nullptr;\n        OP_REQUIRES_OK(context, context->allocate_output(0, TensorShape({maxI, maxJ}), &outputTensor));\n        auto outputMapped = outputTensor->tensor();\n\n        \/\/ do some computations for index (i, j) and write result to index (i, j) of output tensor\n        \/\/ do this for each (i, j) combination\n        outputMapped(i, j) = result; \n\n        \/\/ computing the result for each index (i, j) in parallel would speed up the computation\n        \/\/ Question: is it allowed to write to outputMapped(i, j) from different threads without synchronization?\n    }\n<\/pre>\n\n<p>This is the relevant code:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/githubharald\/CTCWordBeamSearch\/blob\/master\/cpp\/src\/TFWordBeamSearch.cpp#L204\" rel=\"nofollow noreferrer\">Compute(...) method: output tensor<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/githubharald\/CTCWordBeamSearch\/blob\/master\/cpp\/src\/TFWordBeamSearch.cpp#L215\" rel=\"nofollow noreferrer\">Start parallel work<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/githubharald\/CTCWordBeamSearch\/blob\/master\/cpp\/src\/TFWordBeamSearch.cpp#L125\" rel=\"nofollow noreferrer\">Write to output tensor from each thread in parallel<\/a><\/li>\n<\/ul>\n\n<p>I looked through the <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">TF docs<\/a>, but it only mentions that TF itself may call Compute(...) in parallel. But I could not find any information about thread-safety of the tensors. Does anyone have more information about this? <\/p>\n","answers":[]}
+{"QuestionId":52727944,"AnswerCount":1,"Tags":"<python><tensorflow><model><neural-network>","CreationDate":"2018-10-09T19:19:00.173","AcceptedAnswerId":52807262.0,"OwnerUserId":5759234.0,"Title":"restore Tensorflow model without extracting from directory","Body":"<p>I'm currently saving and restoring neural network models using Tensorflow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"noreferrer\"><code>Saver<\/code><\/a> class, as shown below:<\/p>\n\n<p><code>saver.save(sess, checkpoint_prefix, global_step=step)<\/code><\/p>\n\n<p><code>saver.restore(sess, checkpoint_file)<\/code><\/p>\n\n<p>This saves <code>.ckpt<\/code> files of the model to a specified path. Because I am running multiple experiments, I have limited space to save these models. <\/p>\n\n<p>I would like to know if there is a way to save these models without saving content in specified directories.<\/p>\n\n<p>Ex. can I just pass some object at the last checkpoint to some evaluate() function and restore the model from that object? <\/p>\n\n<p>So far as I see, the <code>save_path<\/code> parameter in <code>tf.train.Saver.restore()<\/code> is not optional.<\/p>\n\n<p>Any insight would be much appreciated.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52807262","CreationDate":"2018-10-14T21:29:09.837","ParentId":null,"OwnerUserId":"1097517","Title":null,"Body":"<p>You can use the loaded graph and weights to evaluate in the same way that you train. You just need to change the input to be from your evaluation set. Here is some pseudo code of a train loop with an evaluation loop every <code>1000<\/code> iterations (assumes you have created a <code>tf.Session<\/code> called <code>sess<\/code>):<\/p>\n\n<pre><code>x = tf.placeholder(...)\nloss, train_step = model(x)\nfor i in range(num_step):\n    input_x = get_train_data(i)\n    sess.run(train_step, feed_dict={x: input_x})\n    if i % 1000 == 0 and i != 0:\n        eval_loss = 0\n        for j in range(num_eval):\n            input_x = get_eval_data(j)\n            eval_loss += sess.run(loss, feed_dict={x: input_x})\n        print(eval_loss\/num_eval)\n<\/code><\/pre>\n\n<p>If you're using <code>tf.data<\/code> for your input then you can just create a <code>tf.cond<\/code> to select which input to use:<\/p>\n\n<pre><code>is_training = tf.placeholder(tf.bool)\nnext_element = tf.cond(is_training,\n                        lambda: get_next_train(),\n                        lambda: get_next_eval())\n<\/code><\/pre>\n\n<p><code>get_next_train<\/code> and <code>get_next_eval<\/code> would have to create all ops that are used for reading the dataset, or else there will be side affects of running the above code.<\/p>\n\n<p>This way you don't have to save anything to disc if you don't want to.<\/p>\n"}]}
+{"QuestionId":52728131,"AnswerCount":0,"Tags":"<python><tensorflow><gpu><tesla>","CreationDate":"2018-10-09T19:32:41.767","AcceptedAnswerId":null,"OwnerUserId":9632288.0,"Title":"Is it training too slow or at normal speed? GPU + python + tensorflow-gpu","Body":"<p>I am training a \"faster_rcnn_inception_resnet_v2_atrous_coco\" for custom object detection using tensorflow's api.<\/p>\n\n<p>I set up a machine at azure with following configuration:<\/p>\n\n<p>Intel cxeon CPU E5-2690 v3 @ 2.60GHz\nRAM 56GB\nwindows10 64bit\nGPU tesla k80 total memory 11.18GB<\/p>\n\n<p>When i run train.py I get the following speed per second:<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:global step 458: loss = 0.5601 (3.000 sec\/step)\n  I1009 19:30:13.254615  5916 tf_logging.py:115] global step 458: loss = 0.5601 (3.000 sec\/step)\n  INFO:tensorflow:global step 459: loss = 0.5724 (3.077 sec\/step)\n  I1009 19:30:16.331734  5916 tf_logging.py:115] global step 459: loss = 0.5724 (3.077 sec\/step)\n  INFO:tensorflow:global step 460: loss = 0.8615 (3.018 sec\/step)\n  I1009 19:30:19.350132  5916 tf_logging.py:115] global step 460: loss = 0.8615 (3.018 sec\/step)\n  INFO:tensorflow:global step 461: loss = 0.6021 (3.062 sec\/step)\n  I1009 19:30:22.428256  5916 tf_logging.py:115] global step 461: loss = 0.6021 (3.062 sec\/step)<\/p>\n<\/blockquote>\n\n<p>Is it fast enough or it should be faster as it is using a GPU? batchsize of the config file is 1. When I change it to 2 or higher it runs out of memory.<\/p>\n\n<p>it takes 3 seconds per step in a dataset of 93 images. Ok... but after training, when I load the frozen graph and try to predict it over all images, it takes 1 second per image.. with the GPU... seems too slow.. what I am doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":52728417,"AnswerCount":3,"Tags":"<python><macos><machine-learning><caffe>","CreationDate":"2018-10-09T19:55:11.293","AcceptedAnswerId":null,"OwnerUserId":6191118.0,"Title":"OSX Caffe Compilation Fails With Expected Expression Error","Body":"<p>I have been trying to install Caffe on my mac running OSX 10.13.6<\/p>\n\n<p>I have followed several guides, including the installation guide on the caffe site. I have erased everything and restarted several times. I get the same problem no matter what. When I go to compile everything I am getting this<\/p>\n\n<p><code>Scanning dependencies of target caffeproto\n[  1%] Building CXX object src\/caffe\/CMakeFiles\/caffeproto.dir\/__\/__\/include\/caffe\/proto\/caffe.pb.cc.o\nIn file included from \/Users\/Name\/Documents\/Programming\/PythonEnv\/caffe\/build\/include\/caffe\/proto\/caffe.pb.cc:4:\nIn file included from \/Users\/Name\/Documents\/Programming\/PythonEnv\/caffe\/build\/include\/caffe\/proto\/caffe.pb.h:9:\n\/usr\/local\/include\/google\/protobuf\/stubs\/common.h:209:17: error: expected\n      expression\n  OnShutdownRun([](const void* p) { delete static_cast&lt;const T*&gt;(p); }, p);\n                ^\nIn file included from \/Users\/Name\/Documents\/Programming\/PythonEnv\/caffe\/build\/include\/caffe\/proto\/caffe.pb.cc:4:\nIn file included from \/Users\/Name\/Documents\/Programming\/PythonEnv\/caffe\/build\/include\/caffe\/proto\/caffe.pb.h:25:\nIn file included from \/usr\/local\/include\/google\/protobuf\/generated_message_table_driven.h:34:\nIn file included from \/usr\/local\/include\/google\/protobuf\/map.h:49:\nIn file included from \/usr\/local\/include\/google\/protobuf\/map_type_handler.h:35:\nIn file included from \/usr\/local\/include\/google\/protobuf\/wire_format_lite_inl.h:43:\n\/usr\/local\/include\/google\/protobuf\/message_lite.h:117:3: error: unknown type\n      name 'constexpr'\n  constexpr const T&amp; get() const { return reinterpret_cast&lt;const T&amp;&gt;(union_); }\n<\/code><\/p>\n\n<p>I have also tried using CMake and run into the same problem. I'm not sure where to go from here. I am not incredibly knowledgable about building processes so I apologize if this is vague and will be happy to provide whatever other information might help fix this. <\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"53549977","CreationDate":"2018-11-30T01:30:30.573","ParentId":null,"OwnerUserId":"10725480","Title":null,"Body":"<p>You just need to replace the newest version of protobuf with protobuf v3.5.1.<\/p>\n\n<pre><code>wget https:\/\/github.com\/protocolbuffers\/protobuf\/archive\/v3.5.1.zip\n<\/code><\/pre>\n\n<p>You need to download to source code, and compile it yourself.<\/p>\n"},{"AnswerId":"54394139","CreationDate":"2019-01-28T00:14:45.503","ParentId":null,"OwnerUserId":"1895479","Title":null,"Body":"<p>If compiling with <code>make<\/code>, change <code>Makefile<\/code> as follows:<\/p>\n\n<p><code>- CXXFLAGS += -pthread -fPIC $(COMMON_FLAGS) $(WARNINGS)<\/code><\/p>\n\n<p><code>+ CXXFLAGS += -pthread -fPIC $(COMMON_FLAGS) $(WARNINGS) -std=c++11<\/code><\/p>\n"},{"AnswerId":"53172967","CreationDate":"2018-11-06T13:31:17.603","ParentId":null,"OwnerUserId":"173515","Title":null,"Body":"<p>This is the same problem as reported here: <a href=\"https:\/\/trac.macports.org\/ticket\/57093#comment:1\" rel=\"nofollow noreferrer\">https:\/\/trac.macports.org\/ticket\/57093#comment:1<\/a><\/p>\n\n<p>The compiler needs to be using C++11. Try making this change in <code>CMakeLists.txt<\/code>:<\/p>\n\n<p><code>\n if(UNIX OR APPLE)\n-  set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -fPIC -Wall\")\n+  set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -fPIC -Wall -std=c++11\")\n endif()\n<\/code><\/p>\n"}]}
+{"QuestionId":52728509,"AnswerCount":0,"Tags":"<tensorflow><inference>","CreationDate":"2018-10-09T20:02:35.313","AcceptedAnswerId":null,"OwnerUserId":10480638.0,"Title":"Is it possible to replace Batchnorm with FusedBatchNorm on a frozen inference graph?","Body":"<p>I am trying to replace a BatchNorm layer with a FusedBatchNorm layer in a pretrained model (a .pb file) in tensorflow. Is this actually possible without changing the dataflow of the model and if so, is there a working example of someone doing so?<\/p>\n","answers":[]}
+{"QuestionId":52728620,"AnswerCount":2,"Tags":"<tensorflow><machine-learning>","CreationDate":"2018-10-09T20:11:01.587","AcceptedAnswerId":null,"OwnerUserId":10479285.0,"Title":"Multiple trained models vs Multple features and one model","Body":"<p>I'm trying to build a regression based M\/L model using tensorflow.<\/p>\n\n<p>I am trying to estimate an <strong>object's ETA<\/strong> based on the following:<\/p>\n\n<ul>\n<li>distance from target<\/li>\n<li>distance from target (X component)<\/li>\n<li>distance from target (Y component)<\/li>\n<li>speed<\/li>\n<\/ul>\n\n<p>The object travels on specific journeys. This could be represented as from A->B or from A->C or from D->F (POINT 1 -> POINT 2). There are 500 specific journeys (between a set of points).\nThese journeys aren't completely straight lines, and every journey is different (ie. the shape of the route taken).<\/p>\n\n<p>I have two ways of getting around this problem:<\/p>\n\n<ol>\n<li>I can have 500 different models with 4 features and one label(the training ETA data). <\/li>\n<li>I can have 1 model with 5 features and one label.<\/li>\n<\/ol>\n\n<p>My dilemma is that if I use option 1, that's added complexity, but will be more accurate as every model will be specific to each journey.<\/p>\n\n<p>If I use option 2, the model will be pretty simple, but I don't know if it would work properly. The new feature that I would add are <strong>originCode<\/strong>+ <strong>destinationCode<\/strong>. Unfortunately these are not quantifiable in order to make any numerical sense or pattern - they're just text that define the journey (journey A->B, and the feature would be 'AB').<\/p>\n\n<p>Is there some way that I can use one model, and categorize the features so that one feature is just a 'grouping' feature (in order separate the training data with respect to the journey.<\/p>\n","answers":[{"AnswerId":"52731478","CreationDate":"2018-10-10T01:45:51.733","ParentId":null,"OwnerUserId":"10475762","Title":null,"Body":"<p>In ML, I believe that option 2 is <strong>generally<\/strong> the better option. We prefer general models rather than tailoring many models to specific tasks, as that gets dangerously close to hardcoding, which is what we're trying to get away from by using ML!<\/p>\n\n<p>I think that, depending on the training data you have available, and the model size, a one-hot vector could be used to describe the starting\/end points for the model. Eg, say we have 5 points (ABCDE), and we are going from position B to position C, this could be represented by the vector:<\/p>\n\n<blockquote>\n  <p>0100000100<\/p>\n<\/blockquote>\n\n<p>as in, the first five values correspond to the origin spot whereas the second five are the destination. It is also possible to combine these if you want to reduce your input feature space to:<\/p>\n\n<blockquote>\n  <p>01100<\/p>\n<\/blockquote>\n\n<p>There are other things to consider, as Scott has said in the comments:<\/p>\n\n<ul>\n<li>How much data do you have? Maybe the feature space will be too big this way, I can't be sure. If you have enough data, then the model will intuitively learn the general distances (not actually, but intrinsically in the data) between datapoints. \nIf you have enough data, you might even be able to accurately predict between two points you don't have data for! <\/li>\n<li>If it does come down to not having enough data, then finding representative features of the journey will come into use, ie. length of journey, shape of the journey, elevation travelled etc. Also a metric for distance travelled <strong>from<\/strong> the origin could be useful.<\/li>\n<\/ul>\n\n<p>Best of luck!<\/p>\n"},{"AnswerId":"52740876","CreationDate":"2018-10-10T13:00:49.047","ParentId":null,"OwnerUserId":"2886575","Title":null,"Body":"<p>I would be inclined to lean toward individual models.  This is because, for a given position along a given route and a constant speed, the ETA is a deterministic function of time.  If one moves monotonically closer to the target along the route, it is also a deterministic function of distance to target.  Thus, there is no information to transfer from one route to the next, i.e. \"lumping\" their parameters offers no a priori benefit.  This is assuming, of course, that you have several \"trips\" worth of data along each route (i.e. (distance, speed) collected once per minute, or some such).  If you have only, say, one datum per route then lumping the parameters is a must.  However, in such a low-data scenario, I believe that including a dummy variable for \"which route\" would ultimately be fruitless, since that would introduce a number of parameters that rivals the size of your dataset.<\/p>\n\n<p>As a side note, NEITHER of the models you describe could handle new routes.  I would be inclined to build an individual model per route, data quantity permitting, and a single model neglecting the route identity entirely just for handling new routes, until sufficient data is available to build a model for that route.<\/p>\n"}]}
+{"QuestionId":52728958,"AnswerCount":0,"Tags":"<python><tensorflow><scikit-learn><keras><lstm>","CreationDate":"2018-10-09T20:37:27.433","AcceptedAnswerId":null,"OwnerUserId":10420766.0,"Title":"How to use Scikit Learn Wrapper around Keras Bi-directional LSTM Model","Body":"<p>I have created a Keras LSTM model that does sequence classification.\nI have 27 sequences in the Training set and 18 sequences in the Test set. Each sequence has 4000 time-steps that I have achieved by padding with zeroes. Each sequence is a combination of 2499 parallel series. This means I have 2499 Features. <\/p>\n\n<ul>\n<li>Dimensions of X_Train is (27 x 4000 x 2499)\n:27-Sequences, 4000 Timesteps in each sequence and 2499 features.<\/li>\n<li>Dimension of Y_Train is (27 x 4000 x 1)<\/li>\n<li>Dimension of X_Test is (18 x 4000 x 2499)<\/li>\n<li>Dimension of Y_Test is (18 x 4000 x 1)<\/li>\n<\/ul>\n\n<p>I am using Bi-directional LSTM Model with return_sequences set to True<\/p>\n\n<p>My ultimate goal is to get Feature Importances using ELI5 Package's Permutation Importance.<\/p>\n\n<p>Since ELI5 package does not support Keras framework, I want to use a Scikit learn Wrapper around keras to get it to behave like scikit learn.<\/p>\n\n<p>Then I can finally use ELI5 package on my model to get Important features.<\/p>\n\n<p>I have used these parameters:<\/p>\n\n<ul>\n<li>layer1_units = 40 <\/li>\n<li>layer1_act = 'tanh' <\/li>\n<li>go_backwards = False <\/li>\n<li>return_sequences = True<\/li>\n<li>merge_mode = 'concat'<\/li>\n<li>lr = 0.01 <\/li>\n<li>epochs = 2 <\/li>\n<li>batch_size = 200<\/li>\n<\/ul>\n\n<p>I am going to use this lstm model in the build_fn attribute of KerasClassifier(). <\/p>\n\n<pre><code>model = KerasClassifier(build_fn= lstm_Trial.model(),  epochs=3, batch_size=40, verbose=1)\n<\/code><\/pre>\n\n<p>Then I am trying to use .fit() method.<\/p>\n\n<pre><code>model.fit(x = X_Train, y = Y_Train_Ori)\n<\/code><\/pre>\n\n<p>This throws an error.<\/p>\n\n<hr>\n\n<blockquote>\n  <p>ValueError                                Traceback (most recent call last)\n   in ()\n  ----> 1 model.fit(x = X_Train, y = Y_Train_Ori)\n   ~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/wrappers\/scikit_learn.py in fit(self, x, y, sample_weight, **kwargs)\n      203             y = np.searchsorted(self.classes_, y)\n      204         else:\n  --> 205             raise ValueError('Invalid shape for y: ' + str(y.shape))\n      206         self.n_classes_ = len(self.classes_)\n      207         if sample_weight is not None:<\/p>\n  \n  <p>ValueError: Invalid shape for y: (27, 4000, 1)<\/p>\n<\/blockquote>\n\n<p>How do I use KerasClassifier properly so that Ultimately I am able to use ELI5 package?<\/p>\n","answers":[]}
+{"QuestionId":52728981,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><jupyter-notebook>","CreationDate":"2018-10-09T20:39:52.727","AcceptedAnswerId":52731283.0,"OwnerUserId":7636396.0,"Title":"Jupyter Notebook Tensorflow MNIST import Error: ImportError: cannot import name batching","Body":"<p>I tried to execute this line in my application in Jupyter notebook.<\/p>\n\n<pre><code>from tensorflow.contrib.learn.python.learn.datasets.mnist import read_data_sets\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>.conda\/envs\/py27\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/__init__.py in &lt;module&gt;()\n     23 \n     24 # Add projects here, they will show up under tf.contrib.\n---&gt; 25 from tensorflow.contrib import batching\n     26 from tensorflow.contrib import bayesflow\n     27 from tensorflow.contrib import cloud\n\nImportError: cannot import name batching\n<\/code><\/pre>\n\n<p>I do not get this error if I ran it like this:<\/p>\n\n<pre><code> python -c \"from tensorflow.examples.tutorials.mnist import input_data\"\n<\/code><\/pre>\n\n<p>Can someone tell me what's causing the error in Jupyter Notebook<\/p>\n","answers":[{"AnswerId":"52731283","CreationDate":"2018-10-10T01:13:26.447","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>I have run into similar issue with importing libraries from jupyter notebook, and the problem has always been that I forgot to install <code>jupyter<\/code> in my new conda environment, but jupyter is installed in the root environment. <\/p>\n\n<p>Based on the error given, make sure you have jupyter installed within the <code>py27<\/code> conda environment. If it not installed, it will default to open the one in the root environment. <\/p>\n\n<p>So in the terminal: <\/p>\n\n<pre><code>source activate py27  # activate py27 if on windows\nconda install jupyter\n<\/code><\/pre>\n"}]}
+{"QuestionId":52729117,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-09T20:49:31.573","AcceptedAnswerId":52738354.0,"OwnerUserId":6169238.0,"Title":"Tensorflow With Context Manager vs. With Session","Body":"<p>I'm trying to understand how a <code>with<\/code> block behaves in the above scenarios. I  assume to always have one graph and one session only. I understand that I have (at least?) 2 ways to work using a session in a <code>with<\/code> block:  <\/p>\n\n<p><strong>Example 1 :<\/strong> using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>as_default()<\/code><\/a> that creates a context manager:<br>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\">Same documentation<\/a> says:  <\/p>\n\n<blockquote>\n  <p>Use with the <code>with<\/code> keyword to specify that calls to\n  <code>tf.Operation.run<\/code> or <code>tf.Tensor.eval<\/code> should be executed in this\n  session.<\/p>\n<\/blockquote>\n\n<pre><code># Create session\nsess = tf.Session()\n# Enter with block on a new context manager\nwith sess.as_default():\n  # Train: Following line should result calling tf.Operation.run\n  sess.run([optimizer, loss], feed_dict={x: x_train, y: y_train)\n  # Eval: Following line should result calling tf.Tensor.eval\n  sess.run([loss], feed_dict={x: x_eval, y: y_eval)\n<\/code><\/pre>\n\n<p><strong>Example 2 :<\/strong> with block on session as stated in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\">same documentation in a  lower section<\/a>:  <\/p>\n\n<blockquote>\n  <p>Alternatively, you can use <code>with tf.Session():<\/code> to create a session that\n  is automatically closed on exiting the context, including when an\n  uncaught exception is raised.<\/p>\n<\/blockquote>\n\n<pre><code># Enter with block on session instead of Context Manager\nwith tf.Session() as sess:\n  # Train: Following line seems calling tf.Operation.run as per my test\n  sess.run([optimizer, loss], feed_dict={x: x_train, y: y_train)\n  # Eval: Following is unclear\n  sess.run([loss], feed_dict={x: x_eval, y: y_eval)\n<\/code><\/pre>\n\n<p>I would like to understand what is the correct usage as I see GitHub examples of both cases but of course without the results. In my tests both Example 1 and 2 works for training. For evaluation it seems there is a difference that I can't understand. It exceeds my knowledge to browse the Tensorflow source. Can someone please explain?<\/p>\n","answers":[{"AnswerId":"52738354","CreationDate":"2018-10-10T10:40:24.120","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>They do slightly different things, so each may or may not be correct depending on the usage. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>tf.Session.as_default()<\/code><\/a> will just ensure that the session is set as the default ones, so calls to <code>eval<\/code> and <code>run<\/code> will use that session by default:<\/p>\n\n<pre><code>import tensorflow as tf\n\nsess = tf.Session()\nwith sess.as_default():\n    print(sess is tf.get_default_session())  # True\n<\/code><\/pre>\n\n<p>However, as stated in the documentation, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>tf.Session.as_default()<\/code><\/a> will <em>not<\/em> automatically close the session after the <code>with<\/code> block. If you want that, you can use the session itself as the context manager:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    # Do something with the session\nsess.run([])  # RuntimeError: Attempted to use a closed Session.\n<\/code><\/pre>\n\n<p>However, although (from my point of view) not clearly documented, using a session as context manager <em>also<\/em> makes it the default session.<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    print(sess is tf.get_default_session())  # True\n<\/code><\/pre>\n\n<p>What is the point of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>tf.Session.as_default()<\/code><\/a>, then? Well, simply when you want to temporarily make a session the default one but you do not one to close it after that (in which case you should manually close it later, or use it as an outer context manager). This is probably most relevant when you have more than one open session. Consider the following case:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess1, tf.Session() as sess2:\n    print(sess2 is tf.get_default_session())  # True\n<\/code><\/pre>\n\n<p>Here <code>sess2<\/code> is the default because its context was added later (it can be considered to be \"inner\" to the context created by <code>sess1<\/code>). But now maybe you want to make <code>sess1<\/code> the default for a while. However you can not use <code>sess1<\/code> itself as context manager again:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess1, tf.Session() as sess2:\n    # Do something with sess2\n    with sess1:\n        # RuntimeError: Session context managers are not re-entrant.\n        #               Use `Session.as_default()` if you want to enter\n        #               a session multiple times.\n<\/code><\/pre>\n\n<p>So you can switch between one and other default sessions with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>tf.Session.as_default()<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess1, tf.Session() as sess2:\n    with sess1.as_default():\n        # Do something with sess1\n    with sess2.as_default():\n        # Do something with sess2\n        # This is not really needed because sess2 was the default\n        # in the outer context but you can add it to be explicit\n# Both sessions are closed at the end of the outer context\n<\/code><\/pre>\n\n<p>Of course, you can be extra explicit even with one session, if you want:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess, sess.as_default():\n    # ...\n<\/code><\/pre>\n\n<p>Personally, I have never used <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#as_default\" rel=\"nofollow noreferrer\"><code>tf.Session.as_default()<\/code><\/a> in my actual code, but then again I have rarely needed to use multiple sessions, and I prefer to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\"><code>tf.Session.run()<\/code><\/a> instead of relying on the default session anyway, but that is mostly a matter of personal taste I suppose.<\/p>\n"}]}
+{"QuestionId":52729163,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2018-10-09T20:52:20.440","AcceptedAnswerId":null,"OwnerUserId":6442169.0,"Title":"NCCL install \/ Tensorflow complie error with RTX 2080ti","Body":"<p>I am installing Cuda 10, CuDnn 7.3 and NCCL 2.3 on Ubuntu 18.04 with a 2080ti. I am having trouble with the NCCL part. When i try to compile Tensorflow it says error and cannot find NCCL-SLA.txt file. But when i search for that file i cant find it anywhere either? So when i look online people say you can move it from another directory but its not there for me so i do not know what to do.<\/p>\n\n<p>Here is the error code i am getting:<\/p>\n\n<pre><code>ERROR: missing input file '@local_config_nccl\/\/:nccl\/NCCL-SLA.txt'\nERROR: \/home\/josh\/tensorflow\/tensorflow\/tools\/pip_package\/BUILD:166:1: \/\/tensorflow\/tools\/pip_package:build_pip_package: missing input file '@local_config_nccl\/\/:nccl\/NCCL-SLA.txt'\nTarget \/\/tensorflow\/tools\/pip_package:build_pip_package failed to build\nUse --verbose_failures to see the command lines of failed build steps.\nERROR: \/home\/josh\/tensorflow\/tensorflow\/tools\/pip_package\/BUILD:166:1 1 input file(s) do not exist\nINFO: Elapsed time: 17.209s, Critical Path: 11.30s\nINFO: 353 processes, local.\nFAILED: Build did NOT complete successfully\n<\/code><\/pre>\n","answers":[{"AnswerId":"52729902","CreationDate":"2018-10-09T21:56:15.940","ParentId":null,"OwnerUserId":"4828285","Title":null,"Body":"<p>If you can\u2019t find the file NCCL-SLA.txt, try copy the file LICENSE.txt to a new file named NCCL-SLA.txt in the same directory<\/p>\n"}]}
+{"QuestionId":52729443,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><classification><tensorflow.js>","CreationDate":"2018-10-09T21:16:04.510","AcceptedAnswerId":52749515.0,"OwnerUserId":3209558.0,"Title":"new shape and old shape must have the same number of elements","Body":"<p>For learning purpose, I am using Tensorflow.js, and I experience an error while trying to use the <code>fit<\/code> method with a batched dataset (10 by 10) to learn the process of batch training.<\/p>\n\n<p>I have got a few images 600x600x3 that I want to classify (2 outputs, either 1 or 0)<\/p>\n\n<p>Here is my training loop:<\/p>\n\n<pre><code>  const batches = await loadDataset()\n\n  for (let i = 0; i &lt; batches.length; i++) {\n    const batch = batches[i]\n    const xs = batch.xs.reshape([batch.size, 600, 600, 3])\n    const ys = tf.oneHot(batch.ys, 2)\n\n    console.log({\n      xs: xs.shape,\n      ys: ys.shape,\n    })\n    \/\/ { xs: [ 10, 600, 600, 3 ], ys: [ 10, 2 ] }\n\n    const history = await model.fit(\n      xs, ys,\n      {\n        batchSize: batch.size,\n        epochs: 1\n      }) \/\/ &lt;----- The code throws here\n\n    const loss = history.history.loss[0]\n    const accuracy = history.history.acc[0]\n\n    console.log({ loss, accuracy })\n  }\n<\/code><\/pre>\n\n<p>Here is how I define the dataset<\/p>\n\n<pre><code>const chunks = chunk(examples, BATCH_SIZE)\n\nconst batches = chunks.map(\n  batch =&gt; {\n    const ys = tf.tensor1d(batch.map(e =&gt; e.y), 'int32')\n    const xs = batch\n      .map(e =&gt; imageToInput(e.x, 3))\n      .reduce((p, c) =&gt; p ? p.concat(c) : c)\n    return { size: batch.length, xs , ys }\n  }\n)\n<\/code><\/pre>\n\n<p>Here is the model:<\/p>\n\n<pre><code>const model = tf.sequential()\nmodel.add(tf.layers.conv2d({\n  inputShape: [600, 600, 3],\n  kernelSize: 60,\n  filters: 50,\n  strides: 20,\n  activation: 'relu',\n  kernelInitializer: 'VarianceScaling'\n}))\nmodel.add(tf.layers.maxPooling2d({\n  poolSize: [20, 20],\n  strides: [20, 20]\n}))\nmodel.add(tf.layers.conv2d({\n  kernelSize: 5,\n  filters: 100,\n  strides: 20,\n  activation: 'relu',\n  kernelInitializer: 'VarianceScaling'\n}))\n\nmodel.add(tf.layers.maxPooling2d({\n  poolSize: [20, 20],\n  strides: [20, 20]\n}))\nmodel.add(tf.layers.flatten())\nmodel.add(tf.layers.dense({\n  units: 2,\n  kernelInitializer: 'VarianceScaling',\n  activation: 'softmax'\n}))\n<\/code><\/pre>\n\n<p>I get an error during the first iteration in the for-loop, from the <code>.fit<\/code> which is the following:<\/p>\n\n<pre><code>Error: new shape and old shape must have the same number of elements.\n    at Object.assert (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/util.js:36:15)\n    at reshape_ (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/ops\/array_ops.js:271:10)\n    at Object.reshape (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/ops\/operation.js:23:29)\n    at Tensor.reshape (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/tensor.js:273:26)\n    at Object.derB [as $b] (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/ops\/binary_ops.js:32:24)\n    at _loop_1 (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/tape.js:90:47)\n    at Object.backpropagateGradients (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/tape.js:108:9)\n    at \/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/engine.js:334:20\n    at \/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/engine.js:91:22\n    at Engine.scopedRun (\/Users\/person\/nn\/node_modules\/@tensorflow\/tfjs-core\/dist\/engine.js:101:23)\n<\/code><\/pre>\n\n<p>I don't know what to understand from that and found no documentation or help on that specific error, any idea?<\/p>\n","answers":[{"AnswerId":"52749515","CreationDate":"2018-10-10T22:16:53.093","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The issue of the model lies in the way the <code>convolution<\/code> is applied along with the <code>maxPooling<\/code><\/p>\n\n<p>The first layer is doing a convolution of kernelSize 60 with a strides of [20, 20] and 50 filters.\nThe output of this layer will have the approximate shape <code>[600 \/ 20, 600 \/ 20, 50] = [30, 30, 50]<\/code><\/p>\n\n<p>The max pooling is applied with a stride of <code>[20, 20]<\/code>. The output of this layer will also have the approximate shape <code>[30 \/ 20, 30 \/ 20, 50] =[1, 1, 50 ]<\/code><\/p>\n\n<p>From this step, the model can no longer perform a convolution with a kernelSize 5.  For the kernel shape <code>[5, 5]<\/code> is bigger than the input shape <code>[1, 1]<\/code> resulting in the error that is thrown. The only convolution the model can perform is that of a kernel whose size is 1. Obviously, that convolution will output the input without any transformation.<\/p>\n\n<p>The same rule applies to the last <code>maxPooling<\/code> whose <code>poolingSize<\/code> cannot be different from 1, otherwise an error will be thrown.<\/p>\n\n<p>Here is a snippet:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential()\r\nmodel.add(tf.layers.conv2d({\r\n  inputShape: [600, 600, 3],\r\n  kernelSize: 60,\r\n  filters: 50,\r\n  strides: 20,\r\n  activation: 'relu',\r\n  kernelInitializer: 'VarianceScaling'\r\n}))\r\nmodel.add(tf.layers.maxPooling2d({\r\n  poolSize: [20, 20],\r\n  strides: [20, 20]\r\n}))\r\nmodel.add(tf.layers.conv2d({\r\n  kernelSize: 1,\r\n  filters: 100,\r\n  strides: 20,\r\n  activation: 'relu',\r\n  kernelInitializer: 'VarianceScaling'\r\n}))\r\n\r\nmodel.add(tf.layers.maxPooling2d({\r\n  poolSize: 1,\r\n  strides: [20, 20]\r\n}))\r\nmodel.add(tf.layers.flatten())\r\nmodel.add(tf.layers.dense({\r\n  units: 2,\r\n  kernelInitializer: 'VarianceScaling',\r\n  activation: 'softmax'\r\n}))\r\n\r\nmodel.compile({optimizer: 'sgd', loss: 'meanSquaredError'});\r\nmodel.fit(tf.ones([10, 600, 600, 3]), tf.ones([10, 2]), {batchSize: 4});\r\n\r\nmodel.predict(tf.ones([1, 600, 600, 3])).print()<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":52729769,"AnswerCount":1,"Tags":"<tensorflow><apache-beam><tensorflow-transform>","CreationDate":"2018-10-09T21:44:48.837","AcceptedAnswerId":null,"OwnerUserId":2226322.0,"Title":"TFX Transform Rank Mismatch While Loading\/Applying TFX Beam Transform Graph","Body":"<p>I've already successfully fit a TFTransformOutput to some data (in this case, the Census dataset from UCI common amongst the TF and TFX examples.) I try to apply the transformer with the method transform_raw_features(raw_features) but keep getting the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Node 'transform\/transform\/inputs\/workclass_copy' has an\n  _output_shapes attribute inconsistent with the GraphDef for output #0: Shapes must be equal rank, but are 0 and 1<\/p>\n<\/blockquote>\n\n<p>Digging into the source code, it seems the error originates in saved_transform_io in the method _partially_apply_saved_transform_impl while doing:<\/p>\n\n<pre><code>saver = tf_saver.import_meta_graph(meta_graph_def, import_scope=import_scope,\ninput_map=input_map)\n<\/code><\/pre>\n\n<p>I examined the meta_graph_def produced by TFX TFTransform and Beam and notice that the graph indeed has a series of copied variables with input\/output rank differences. However, that is nothing I have control over. <\/p>\n\n<p>The column in the error message is \"workclass\" which is a simple categorical column. What might I be doing incorrectly? What is the best way to debug this? At this point, I've already dug deep into the TF source code but the error seems to originate with how the TFTransform graph was written, not sure what levers I have to change\/fix that.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZI5yB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZI5yB.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>This is using TF Transform v0.9 and the corresponding TF v1.9<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"\/home\/sahmed\/workspace\/ml_playground\/TFX-TFT\/trainers.py\", line 449,\n  in parse_csv\n      transformed_stuff=xformer.transform_raw_features(raw_features)   File\n  \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow_transform\/output_wrapper.py\",\n  line 122, in transform_raw_features\n      self.transform_savedmodel_dir, raw_features))   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow_transform\/saved\/saved_transform_io.py\",\n  line 360, in partially_apply_saved_transform_internal\n      saved_model_dir, logical_input_map, tensor_replacement_map)   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow_transform\/saved\/saved_transform_io.py\",\n  line 218, in _partially_apply_saved_transform_impl\n      input_map=input_map)   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/saver.py\",\n  line 1960, in import_meta_graph\n      **kwargs)   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/meta_graph.py\",\n  line 744, in import_scoped_meta_graph\n      producer_op_list=producer_op_list)   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\",\n  line 432, in new_func\n      return func(*args, **kwargs)   File \"\/home\/sahmed\/miniconda3\/envs\/kml2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/importer.py\",\n  line 422, in import_graph_def\n      raise ValueError(str(e)) ValueError: Node 'transform\/transform\/inputs\/workclass_copy' has an _output_shapes\n  attribute inconsistent with the GraphDef for output #0: Shapes must be\n  equal rank, but are 0 and 1<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53052532","CreationDate":"2018-10-29T19:31:24.540","ParentId":null,"OwnerUserId":"1486949","Title":null,"Body":"<p>The issue is likely that the shape of the workclass tensor is incompatible with what transform_raw_features expects.<\/p>\n\n<p>TFTransformOutput.transform_raw_features() expects these features to have the same characteristics as described in the metadata given to tft.AnalyzeDataset() similarly to how it's done in this example:\n<a href=\"https:\/\/github.com\/tensorflow\/transform\/blob\/master\/examples\/simple_example.py#L63\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/transform\/blob\/master\/examples\/simple_example.py#L63<\/a><\/p>\n\n<p>Could you take a look at the metadata used in your pipeline and see that it is compatible with the data fed into TFTransformOutput.transform_raw_features()?<\/p>\n"}]}
+{"QuestionId":52730013,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-09T22:08:02.797","AcceptedAnswerId":null,"OwnerUserId":1030294.0,"Title":"Import Error in installation: Error installing Horovod and Tensorflow","Body":"<p>I am trying to install Tensorflow and Horovod<\/p>\n\n<pre><code>pip install tensorflow\n\nHOROVOD_WITH_TENSORFLOW=1 pip install horovod\n<\/code><\/pre>\n\n<p>Then I ran a sample code<\/p>\n\n<pre><code>import tensorflow as tf\nimport horovod.tensorflow as hvd\n<\/code><\/pre>\n\n<p>When I run this code, I get the error <\/p>\n\n<pre><code>ImportError: Extension horovod.tensorflow has not been built.  If this is not expected, reinstall Horovod with HOROVOD_WITH_TENSORFLOW=1 to debug the build error.\n<\/code><\/pre>\n","answers":[{"AnswerId":"53480982","CreationDate":"2018-11-26T12:18:21.917","ParentId":null,"OwnerUserId":"9426998","Title":null,"Body":"<p>If you need to install tensorflow and horovod , you can use the following steps:<\/p>\n\n<p>1)Create a conda environment to avoid the mismatch of package versions.<\/p>\n\n<pre><code>conda create -n test_hvd -c intel python=3.6\n<\/code><\/pre>\n\n<p>2)Activate the environment<\/p>\n\n<pre><code>source activate test_hvd\n<\/code><\/pre>\n\n<p>(You can use any name instead of <strong>test_hvd<\/strong>, which is an environment name.)<\/p>\n\n<p>3)Install tensorflow in the activated environment:<\/p>\n\n<pre><code>pip install  https:\/\/storage.googleapis.com\/intel-optimized-tensorflow\/tensorflow-1.10.0-cp36-cp36m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>4)Finally install horovod<\/p>\n\n<pre><code>pip install --no-cache-dir horovod\n<\/code><\/pre>\n\n<p>Note: 1. Kindly confirm if you are using the latest versions of GCC (gcc (GCC) 6.4.0 works fine).<\/p>\n\n<ol start=\"2\">\n<li>These steps are tested on Linux OS<\/li>\n<\/ol>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":52730187,"AnswerCount":0,"Tags":"<keras><lstm><rnn>","CreationDate":"2018-10-09T22:27:16.447","AcceptedAnswerId":null,"OwnerUserId":1780570.0,"Title":"Does Stateful=False in Keras means only last batch matters","Body":"<p>According to Keras's source code, stateful means the \"state\" is assigned to the <code>initial_state<\/code> of next batch, otherwise <code>initial_state<\/code> would be reinitialized.<\/p>\n\n<p>What does <code>initial_state<\/code> mean? After reading the source code, my guess is it would correspond to all U, V and W in the below image (<a href=\"http:\/\/www.wildml.com\/2015\/09\/recurrent-neural-networks-tutorial-part-1-introduction-to-rnns\/\" rel=\"nofollow noreferrer\">source<\/a>), but I have nothing to back that up. <a href=\"https:\/\/i.stack.imgur.com\/Ar5zx.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ar5zx.jpg\" alt=\"below image\"><\/a><\/p>\n\n<p>Moreover, if <code>stateful=False<\/code>, does that mean only last batch matters?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":52730405,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><pytorch><random-seed>","CreationDate":"2018-10-09T23:01:29.477","AcceptedAnswerId":52736022.0,"OwnerUserId":6789651.0,"Title":"Does PyTorch seed affect dropout layers?","Body":"<p>I came across the idea of seeding my neural network for reproducible results, and was wondering if pytorch seeding affects dropout layers and what is the proper way to seed my training\/testing?<\/p>\n\n<p>I'm reading the documentation <a href=\"https:\/\/pytorch.org\/docs\/master\/cuda.html?highlight=manual_seed_all#torch.cuda.manual_seed\" rel=\"nofollow noreferrer\">here<\/a>, and wondering if just placing these lines will be enough?<\/p>\n\n<pre><code>torch.manual_seed(1)\ntorch.cuda.manual_seed(1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52736022","CreationDate":"2018-10-10T08:39:36.547","ParentId":null,"OwnerUserId":"604734","Title":null,"Body":"<p>You can easily answer your question with some lines of code:<\/p>\n\n<pre><code>import torch\nfrom torch import nn\n\ndropout = nn.Dropout(0.5)\ntorch.manual_seed(9999)\na = dropout(torch.ones(1000))\ntorch.manual_seed(9999)\nb = dropout(torch.ones(1000))\nprint(sum(abs(a - b)))\n# &gt; tensor(0.)\n<\/code><\/pre>\n\n<p>Yes, using <code>manual_seed<\/code> is enough.<\/p>\n"}]}
+{"QuestionId":52730645,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-09T23:35:36.550","AcceptedAnswerId":52733531.0,"OwnerUserId":1316501.0,"Title":"Keras: Callbacks Requiring Validation Split?","Body":"<p>I'm working on a multi-class classification problem with Keras 2.1.3 and a Tensorflow backend. I have two numpy arrays, <code>x<\/code> and <code>y<\/code> and I'm using <code>tf.data.Dataset<\/code> like this:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({\"sequence\": x}, y))\ndataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(self.batch_size))\ndataset = dataset.repeat()\nxt, yt = dataset.make_one_shot_iterator().get_next()\n<\/code><\/pre>\n\n<p>Then I make my Keras model (omitted for brevity), compile, and fit:<\/p>\n\n<pre><code>model.compile(\n    loss='categorical_crossentropy',\n    optimizer='adam',\n    metrics=['accuracy'],\n)\nmodel.fit(xt, yt, steps_per_epoch=100, epochs=10)\n<\/code><\/pre>\n\n<p>This works perfectly well.  But when I add in callbacks, I run into issues.  Specifically, if I do this:<\/p>\n\n<pre><code>callbacks = [\n    tf.keras.callbacks.ModelCheckpoint(\"model_{epoch:04d}_{val_acc:.4f}.h5\",\n                    monitor='val_acc',\n                    verbose=1,\n                    save_best_only=True,\n                    mode='max'),\n    tf.keras.callbacks.TensorBoard(os.path.join('.', 'logs')),\n    tf.keras.callbacks.EarlyStopping(monitor='val_acc', patience=5, min_delta=0, mode='max')\n]\nmodel.fit(xt, yt, steps_per_epoch=10, epochs=100, callbacks=callbacks)\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<pre><code>KeyError: 'val_acc'\n<\/code><\/pre>\n\n<p>Also, if I include <code>validation_split=0.1<\/code> in my <code>model.fit(...)<\/code> call, I'm told:<\/p>\n\n<p><code>ValueError: If your data is in the form of symbolic tensors, you cannot use validation_split<\/code>.`<\/p>\n\n<p>What is the normal way to use callbacks and validation splits with <code>tf.data.Dataset<\/code> (tensors)?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52731695","CreationDate":"2018-10-10T02:19:47.880","ParentId":null,"OwnerUserId":"6170271","Title":null,"Body":"<p>I think your problem is type of your datasets, where <code>xt<\/code> and <code>yt<\/code> is not a list or array, so it can not to slice or split.<\/p>\n\n<p>Let assume <code>x<\/code> is array with shape <code>(1000,2)<\/code> fill with random number using numpy and <code>y<\/code> is class with 10 classes.<\/p>\n\n<pre><code># Define x and y\nimport numpy as np\nnp.random.seed(2018)\nn_cat = 10\nx = np.random.rand(1000,2)  # Generate random 2 variables and 1000 rows\ny = [np.random.randint(1,n_cat) for i in range(len(x))]     # Make 10 class\n<\/code><\/pre>\n\n<p>Then you assemble the model and compile<\/p>\n\n<pre><code>model.compile(\n    loss='categorical_crossentropy',\n    optimizer='adam',\n    metrics=['accuracy'],\n)\n<\/code><\/pre>\n\n<p>Create callbacks for checkpoint model<\/p>\n\n<pre><code>callbacks = [\n    tf.keras.callbacks.ModelCheckpoint(\"model_{epoch:04d}_{val_acc:.4f}.h5\",\n                    monitor='val_acc',\n                    verbose=1,\n                    save_best_only=True,\n                    mode='max'),\n    tf.keras.callbacks.TensorBoard(os.path.join('.', 'logs')),\n    tf.keras.callbacks.EarlyStopping(monitor='val_acc', patience=5, min_delta=0, mode='max')\n]\n<\/code><\/pre>\n\n<p>fit the model<\/p>\n\n<pre><code>model.fit(x, y, batch_size=32, epochs=100, callbacks=callbacks, validation_split=0.1)\n<\/code><\/pre>\n\n<p>it's works for me.<\/p>\n"},{"AnswerId":"52733531","CreationDate":"2018-10-10T06:00:41.617","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>Using the tensorflow keras API, you can provide a <code>Dataset<\/code> for training and another for validation.<\/p>\n\n<p>First some imports<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.layers import Dense\nimport numpy as np\n<\/code><\/pre>\n\n<p>define the function which will split the numpy arrays into training\/val<\/p>\n\n<pre><code>def split(x, y, val_size=50):\n    idx = np.random.choice(x.shape[0], size=val_size, replace=False)\n    not_idx = list(set(range(x.shape[0])).difference(set(idx)))\n\n    x_val = x[idx]\n    y_val = y[idx]\n    x_train = x[not_idx]\n    y_train = y[not_idx]\n    return x_train, y_train, x_val, y_val\n<\/code><\/pre>\n\n<p>define numpy arrays and the train\/val tensorflow <code>Datasets<\/code><\/p>\n\n<pre><code>x = np.random.randn(150, 9)\ny = np.random.randint(0, 10, 150)\n\nx_train, y_train, x_val, y_val = split(x, y)\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices((x_train, tf.one_hot(y_train, depth=10)))\ntrain_dataset = train_dataset.batch(32).repeat()\n\nval_dataset = tf.data.Dataset.from_tensor_slices((x_val, tf.one_hot(y_val, depth=10)))\nval_dataset = val_dataset.batch(32).repeat()\n<\/code><\/pre>\n\n<p>Make the model (notice we are using the tensorflow keras API)<\/p>\n\n<pre><code>model = keras.models.Sequential([Dense(64, input_shape=(9,), activation='relu'),\n                                 Dense(64, activation='relu'),\n                                 Dense(10, activation='softmax')\n                                ])\nmodel.compile(optimizer='Adam', loss='categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.summary()\n\n\nmodel.fit(train_dataset,\n          epochs=10, \n          steps_per_epoch=int(100\/32)+1,\n          validation_data=val_dataset,\n          validation_steps=2)\n<\/code><\/pre>\n\n<p>and the model trains, kind of (output):<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense (Dense)                (None, 64)                640       \n_________________________________________________________________\ndense_1 (Dense)              (None, 64)                4160      \n_________________________________________________________________\ndense_2 (Dense)              (None, 10)                650       \n=================================================================\nTotal params: 5,450\nTrainable params: 5,450\nNon-trainable params: 0\n_________________________________________________________________\nEpoch 1\/10\n4\/4 [==============================] - 0s 69ms\/step - loss: 2.3170 - acc: 0.1328 - val_loss: 2.3877 - val_acc: 0.0712\nEpoch 2\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.2628 - acc: 0.2500 - val_loss: 2.3850 - val_acc: 0.0712\nEpoch 3\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.2169 - acc: 0.2656 - val_loss: 2.3838 - val_acc: 0.0712\nEpoch 4\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.1743 - acc: 0.3359 - val_loss: 2.3830 - val_acc: 0.0590\nEpoch 5\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.1343 - acc: 0.3594 - val_loss: 2.3838 - val_acc: 0.0590\nEpoch 6\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.0959 - acc: 0.3516 - val_loss: 2.3858 - val_acc: 0.0590\nEpoch 7\/10\n4\/4 [==============================] - 0s 4ms\/step - loss: 2.0583 - acc: 0.3750 - val_loss: 2.3887 - val_acc: 0.0590\nEpoch 8\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.0223 - acc: 0.4453 - val_loss: 2.3918 - val_acc: 0.0747\nEpoch 9\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 1.9870 - acc: 0.4609 - val_loss: 2.3954 - val_acc: 0.1059\nEpoch 10\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 1.9523 - acc: 0.4609 - val_loss: 2.3995 - val_acc: 0.1059\n<\/code><\/pre>\n\n<h2>Callbacks<\/h2>\n\n<p>Adding callbacks also works, <\/p>\n\n<pre><code>callbacks = [ tf.keras.callbacks.ModelCheckpoint(\"model_{epoch:04d}_{val_acc:.4f}.h5\",\n                    monitor='val_acc',\n                    verbose=1,\n                    save_best_only=True,\n                    mode='max'),\n    tf.keras.callbacks.TensorBoard('.\/logs'),\n    tf.keras.callbacks.EarlyStopping(monitor='val_acc', patience=5, min_delta=0, mode='max')\n]\n\n\nmodel.fit(train_dataset, epochs=10, steps_per_epoch=int(100\/32)+1, validation_data=val_dataset,\n          validation_steps=2, callbacks=callbacks)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Epoch 1\/10\n4\/4\n [==============================] - 0s 59ms\/step - loss: 2.3274 - acc: 0.1094 - val_loss: 2.3143 - val_acc: 0.0833\n\nEpoch 00001: val_acc improved from -inf to 0.08333, saving model to model_0001_0.0833.h5\nEpoch 2\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.2655 - acc: 0.1094 - val_loss: 2.3204 - val_acc: 0.1389\n\nEpoch 00002: val_acc improved from 0.08333 to 0.13889, saving model to model_0002_0.1389.h5\nEpoch 3\/10\n4\/4 [==============================] - 0s 5ms\/step - loss: 2.2122 - acc: 0.1250 - val_loss: 2.3289 - val_acc: 0.1111\n\nEpoch 00003: val_acc did not improve from 0.13889\nEpoch 4\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.1644 - acc: 0.1953 - val_loss: 2.3388 - val_acc: 0.0556\n\nEpoch 00004: val_acc did not improve from 0.13889\nEpoch 5\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.1211 - acc: 0.2734 - val_loss: 2.3495 - val_acc: 0.0556\n\nEpoch 00005: val_acc did not improve from 0.13889\nEpoch 6\/10\n4\/4 [==============================] - 0s 4ms\/step - loss: 2.0808 - acc: 0.2969 - val_loss: 2.3616 - val_acc: 0.0556\n\nEpoch 00006: val_acc did not improve from 0.13889\nEpoch 7\/10\n4\/4 [==============================] - 0s 2ms\/step - loss: 2.0431 - acc: 0.2969 - val_loss: 2.3749 - val_acc: 0.0712\n\nEpoch 00007: val_acc did not improve from 0.13889\n<\/code><\/pre>\n"}]}
+{"QuestionId":52730778,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-09T23:53:27.583","AcceptedAnswerId":52731473.0,"OwnerUserId":1585523.0,"Title":"Keras Metric for bucketizing a regression output","Body":"<p>How do I define a custom keras metric for computing accuracy like so,<\/p>\n\n<pre><code>y_true = [12.5, 45.5]\ny_predicted = [14.5, 29]\nsplits = [-float(\"inf\"), 10, 20, 30, float(\"inf\")]\n\n\"\"\"\nSplits to Classes translation =&gt;\nClass 0: -inf to 9\nClass 1: 10 to 19\nClass 2: 20 to 29\nClass 3: 30 to inf\n\"\"\"\n# using the above translation, \ny_true_classes = [1, 3]\ny_predicted_classes = [1, 2]\n\naccuracy = K.equal( y_true_classes, y_predicted_classes ) # =&gt; 0.5 here\n\nreturn accuracy\n<\/code><\/pre>\n","answers":[{"AnswerId":"52731473","CreationDate":"2018-10-10T01:44:44.863","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>Here is an idea on how you might you around implementing this (although probably not the best one).<\/p>\n\n<pre><code>def convert_to_classes(vals, splits):\n    out = tf.zeros_like(vals, dtype=tf.int32)\n\n    for split in splits:\n        out = tf.where(vals &gt; split, out + 1, out)\n\n    return out\n\n\ndef my_acc(splits):\n    def custom_acc(y_true, y_pred):\n        y_true = convert_to_classes(y_true, splits)\n        y_pred = convert_to_classes(y_pred, splits)\n\n        return K.mean(K.equal(y_true, y_pred))\n    return custom_acc\n<\/code><\/pre>\n\n<p>The function <code>convert_to_classes<\/code> converts the floats to bucks, assuming the bounds are always +-inf. <\/p>\n\n<p>The closure <code>my_acc<\/code> lets you define the splits (without +-inf) at compile time (added statically to the graph), and then returns a metric function as expected with keras.<\/p>\n\n<p>Testing using tensorflow: <\/p>\n\n<pre><code>y_true = tf.constant([12.5, 45.5])\ny_pred = tf.constant([14.5, 29])\n\nwith tf.Session() as sess:\n    print(sess.run(my_acc((10, 20, 30))(y_true, y_pred)))\n<\/code><\/pre>\n\n<p>gives the expected 0.5 accuracy. <\/p>\n\n<p>And quick test with Keras:<\/p>\n\n<pre><code>x = np.random.randn(100, 10)*100\ny = np.random.randn(100)*100\n\nmodel = Sequential([Dense(20, activation='relu'),\n                    Dense(1, activation=None)])\nmodel.compile(optimizer='Adam',\n              loss='mse',\n              metrics=[my_acc(splits=(10, 20, 30))])\n\nmodel.fit(x, y, batch_size=32, epochs=10)\n<\/code><\/pre>\n\n<p>Given the metric (named as the inner function in the closure <code>custom_acc<\/code>)<\/p>\n\n<pre><code>100\/100 [==============================] - 0s 2ms\/step - loss: 10242.2591 - custom_acc: 0.4300\nEpoch 2\/10\n100\/100 [==============================] - 0s 53us\/step - loss: 10101.9658 - custom_acc: 0.4200\nEpoch 3\/10\n100\/100 [==============================] - 0s 53us\/step - loss: 10011.4662 - custom_acc: 0.4300\nEpoch 4\/10\n100\/100 [==============================] - 0s 51us\/step - loss: 9899.7181 - custom_acc: 0.4300\nEpoch 5\/10\n100\/100 [==============================] - 0s 50us\/step - loss: 9815.1607 - custom_acc: 0.4200\nEpoch 6\/10\n100\/100 [==============================] - 0s 74us\/step - loss: 9736.5554 - custom_acc: 0.4300\nEpoch 7\/10\n100\/100 [==============================] - 0s 50us\/step - loss: 9667.0845 - custom_acc: 0.4400\nEpoch 8\/10\n100\/100 [==============================] - 0s 58us\/step - loss: 9589.5439 - custom_acc: 0.4400\nEpoch 9\/10\n100\/100 [==============================] - 0s 61us\/step - loss: 9511.8003 - custom_acc: 0.4400\nEpoch 10\/10\n100\/100 [==============================] - 0s 51us\/step - loss: 9443.9730 - custom_acc: 0.4400\n<\/code><\/pre>\n"}]}
+{"QuestionId":52730922,"AnswerCount":0,"Tags":"<python><numpy><machine-learning><probability><pytorch>","CreationDate":"2018-10-10T00:13:39.293","AcceptedAnswerId":null,"OwnerUserId":3658307.0,"Title":"Pytorch Vectorize Pairwise Gaussian\/Normal Density Evaluations","Body":"<p>In Pytorch, I have three tensors <code>A<\/code>, <code>mu<\/code>, and <code>sigma<\/code>, all with shape <code>B x N x M<\/code>.<\/p>\n\n<p>I want a new tensor <code>D<\/code> of shape <code>B x N x M x B<\/code>, where every entry is the normal density of an entry of <code>A<\/code> with respect to a pair of parameters in <code>mu<\/code> and <code>sigma<\/code>. More concretely:<\/p>\n\n<pre><code>D[i,j,k,l] = univariate_normal_density(A[i,j,k] | mu[l,j,k], sigma[l,j,k])\n<\/code><\/pre>\n\n<p>In other words, each entry of mu and sigma is a mean and variance of a distribution. It doesn't matter whether <code>D<\/code> holds probabilities or log-probabilities, incidentally. Also, note that one can unfold the second and third dimensions, I think I am missing something there.<\/p>\n\n<blockquote>\n  <p>What is the most efficient way to do this?<\/p>\n<\/blockquote>\n\n<p>The most straightforward way seems to be:<\/p>\n\n<ul>\n<li><p>Create <code>B*N*M<\/code> objects of type <code>torch.distributions.normal<\/code>, so that \n<code>N[l,j,k]=N(mu[l,j,k],sigma[l,j,k])<\/code><\/p><\/li>\n<li><p>For each one, evaluate <code>log_prob<\/code> on <code>A<\/code> like <code>D[i,j,k,l] = N[l,j,k].log_prob(A[i,j,k])<\/code><\/p><\/li>\n<\/ul>\n\n<p>I can speed this up via: <code>D[:,j,k,l] = N[l,j,k].log_prob(A[:,j,k])<\/code>, but this still seems undervectorized. I cannot seem to find any built in methods that would do better though.<\/p>\n","answers":[]}
+{"QuestionId":52731151,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T00:50:39.053","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Tesnorflow: How to provide your own `sampled_values` for tf.nn.sampled_softmax_loss?","Body":"<p>In tf.nn.sampled_softmax_loss, one of the optional inputs is to put your own samples values. I would like to provide my own samples values so that I can use float16 (half precision) variables. If <code>sampled_values<\/code> is left blank, Tensorflow will use <code>log_uniform_candidate_sampler<\/code> to get values, which can only return float32. <\/p>\n\n<p>Here are all the inputs. <\/p>\n\n<pre><code>tf.nn.sampled_softmax_loss(\n    weights,\n    biases,\n    labels,\n    inputs,\n    num_sampled,\n    num_classes,\n    num_true=1,\n    sampled_values=None,\n    remove_accidental_hits=True,\n    partition_strategy='mod',\n    name='sampled_softmax_loss',\n    seed=None\n)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sampled_softmax_loss\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sampled_softmax_loss<\/a><\/p>\n\n<p>This is the information they give for the sampled_values arg :<\/p>\n\n<blockquote>\n  <p>sampled_values: a tuple of (sampled_candidates, true_expected_count,\n  sampled_expected_count) returned by a *_candidate_sampler function.\n  (if None, we default to log_uniform_candidate_sampler)<\/p>\n<\/blockquote>\n\n<p>I'm trying to figure out how to provide this tuple. What exactly are the <code>sampled_candidates<\/code>, <code>true_expected_count<\/code>, <code>sampled_expected_count<\/code> ? <\/p>\n\n<p>I know that it's sampling the weights and corresponding biases, so do I put them together in it's own tuple for <code>sampled_candidates<\/code> ? Also, am I putting the int for the place of the weight in the matrix, or am I putting the whole embedding itself?<\/p>\n\n<p>I've also looked at Tensorflow's math supplimental on negative sampling but I couldn't find any information for my issue <a href=\"https:\/\/www.tensorflow.org\/extras\/candidate_sampling.pdf\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extras\/candidate_sampling.pdf<\/a><\/p>\n\n<p>In my search, I found this very similar question on a google forum<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/6IDJ-XAIb9M\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/discuss\/6IDJ-XAIb9M<\/a><\/p>\n\n<p>The Answer given is <\/p>\n\n<blockquote>\n  <p><code>sampled_values<\/code> is the tuple returned by our *candidate_sampler\n  classes. These classes implement methods that sample contrastive\n  labels (not observed, but used during training) according to some\n  distribution Q for use in approximate training methods like\n  noise-contrastive estimation (NCE) and Sampled Softmax. An example is\n  the log_uniform_candidate_sampler, which samples labels according to\n  the log-uniform distribution. <\/p>\n  \n  <p>You almost never need to provide these yourself. You would simply pass\n  in the result of a call to a *candidate_sampler function in the tf.nn\n  module (where * can be \"uniform\", \"log_uniform\", \"zipfian_binned\",\n  etc), e.g.<\/p>\n  \n  <p>sampled_values = tf.nn.zipfian_binned_candidate_sampler(...)<\/p>\n  \n  <p>If you just want to get it to work, just leave it to None, and it\n  would default to the log_uniform_candidate_sampler (often a good\n  choice).<\/p>\n  \n  <p>If you are interested in the math behind this, see this document:\n  <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.8\/extras\/candidate_sampling.pdf\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.8\/extras\/candidate_sampling.pdf<\/a>.<\/p>\n  \n  <p>But to answer your question: For each batch of observed labels L, and\n  a candidate sampling distribution Q, the tuple consists of:<\/p>\n  \n  <ul>\n  <li>the tensor with the actual sampled contrastive labels N, <\/li>\n  <li>the tensor with the log-expected-values of the observed labels L under Q, i.e. log Q(L), and <\/li>\n  <li>the tensor with the log-expected values of the contrastive labels under Q, i.e. log Q(N).<\/li>\n  <\/ul>\n  \n  <p>The latter are required for the math to go through (see above\n  document). So sampled_values contains(with a hopefully clear abuse of\n  notation): <\/p>\n  \n  <p>sampled_values = (N, log Q(L), log Q(N)).<\/p>\n<\/blockquote>\n\n<p>However, I still don't know how to input a value. I'm not sure what the datatypes should be, and if N is the int place in the embedding matrix, or the embedding itself. Also, I'm guessing N should be a list of values itself, the size of the number of negative labels we have to sample. <\/p>\n\n<p>I was wondering if I could get a example with some values. For example, for a negative sampling of 3, do I do something like this? <\/p>\n\n<p>sampled_values = ([4,29, 12], [1, 1, 1], [0, 0, 0])<\/p>\n\n<p>Also, the documentation says that the tuple should be \" returned by a *_candidate_sampler function\"<\/p>\n\n<p>Does that mean I need to provide a function that returns the tuple, instead of the tuple itself?<\/p>\n","answers":[]}
+{"QuestionId":52731302,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-10T01:17:15.600","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"How to convert Variable to keras Layer","Body":"<p>The output is <\/p>\n\n<pre><code>tf.Variable 'Variable_1:0' shape=(10, 15, 9488) dtype=float32_ref\n<\/code><\/pre>\n\n<p>I want to convert it to keras layer. <\/p>\n\n<pre><code>input_output = Input(shape=(10,512))\nmodel = Model(input_output, output) \n<\/code><\/pre>\n\n<p>Is there any way to do that?<\/p>\n\n<p>currently, the error is <\/p>\n\n<pre><code>Output tensors to a Model must be the output of a Keras Layer (thus holding past layer metadata).\nFound:\ntf.Variable 'Variable_1:0' shape=(10, 15, 9488) dtype=float32_ref\n<\/code><\/pre>\n\n<p>=============================================<br>\nI tried to use Lambda, but got the error below: <br>\nValueError: Output tensors to a Model must be the output of a Keras <code>Layer<\/code><br> \n(thus holding past layer metadata). <br>\nFound: keras.layers.core.Lambda object at 0x7fad52faf550<br><\/p>\n\n<p>Here is what I did:<br><\/p>\n\n<pre><code>def convert_tensor(self, outputs):\n    return outputs\n\ninput_result = Input(shape=(int(outputs.shape[0]), int(outputs.shape[1]),int(outputs.shape[2])))\noutputs = Lambda(self.convert_tensor(outputs), output_shape=(10, 15, 9488))\nmodel = Model(input_result, outputs)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52734395","CreationDate":"2018-10-10T06:58:50.477","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You aren't using the Lambda layer correctly, it should be like this:<\/p>\n\n<pre><code>input_result = Input(shape=(int(outputs.shape[0]), int(outputs.shape[1]),int(outputs.shape[2])))\noutputs = Lambda(self.convert_tensor, output_shape=(10, 15, 9488))(input_result)\nmodel = Model(input_result, outputs)\n<\/code><\/pre>\n\n<p>You have to pass the function to the lambda, not a function call, and then you \"call\" the lambda with the appropriate input.<\/p>\n"}]}
+{"QuestionId":52731624,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T02:10:18.777","AcceptedAnswerId":53805395.0,"OwnerUserId":679688.0,"Title":"\"A nested structure of tf.Tensor objects\" for Iterator.get_next() result type","Body":"<p>When I write Tensorflow code, I try to keep in mind the type of different things, e.g. tuple of two Tensors, or list of Tensors. This is important because when the types\/shapes don't match, Tensorflow emits an error.<\/p>\n\n<p>The text of the title of this question shows up a lot in the documentation especially when describing the result of some function, e.g. for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Iterator#get_next\" rel=\"nofollow noreferrer\"><code>Iterator.get_next()<\/code><\/a>, but I find it too vague. It doesn't tell me exactly what to expect, a list of tuples? a tuple of tuples? What exactly is this 'nested structure'? Right now, the only way I can track this is to print the result after Session.run(). Is there a cleaner and more definitive way?<\/p>\n\n<p>Also, it seems that the value of <code>Iterator.get_next()<\/code> is always a list of one element; I haven't been able to make it return a non-list, an empty list or a list with multiple elements. When does <code>Iterator.get_next()<\/code> return something that's not a list of one element? If never, then the wrapping of the content in a list seems superfluous -- why was <code>Iterator.get_next()<\/code> designed this way?<\/p>\n\n<p>This is example code showing I mean:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nds = tf.data.Dataset.from_tensor_slices(np.array(range(0, 8)).reshape(4,2))\nit = ds.make_one_shot_iterator()\n\nwith tf.Session() as sess:\n    for i in range(0, 4):\n        x = sess.run([it.get_next()])\n        print(x)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[array([0, 1])]\n[array([2, 3])]\n[array([4, 5])]\n[array([6, 7])]\n<\/code><\/pre>\n\n<p>Why not just the following?<\/p>\n\n<pre><code>array([0, 1])\narray([2, 3])\narray([4, 5])\narray([6, 7])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53805395","CreationDate":"2018-12-16T18:49:20.907","ParentId":null,"OwnerUserId":"7211137","Title":null,"Body":"<p>Your specific issue of \"Output\" vs. \"Why not just the following?\" results from passing in a list to <code>sess.run<\/code>.  If you instead use <code>sess.run(it.get_next())<\/code> you will get the desired behavior.<\/p>\n\n<p>Note that TensorFlow allows you to pass in several different structures (e.g., lists, dictionaries, namedtuples, etc.), including nested structures (see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/framework\/nest\" rel=\"nofollow noreferrer\">here<\/a>).  It will return the data following the same structure you used to pass it in.<\/p>\n\n<p>For example, with a dictionary:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nds = tf.data.Dataset.from_tensor_slices(np.array(range(0, 8)).reshape(4,2))\nit = ds.make_one_shot_iterator()\n\nwith tf.Session() as sess:\n    for i in range(0, 4):\n        x = sess.run({'x': it.get_next()}])\n        print(x)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>{'x': array([0, 1])}\n{'x': array([2, 3])}\n{'x': array([4, 5])}\n{'x': array([6, 7])}\n<\/code><\/pre>\n"}]}
+{"QuestionId":52731628,"AnswerCount":0,"Tags":"<python><tensorflow><keras><kernel>","CreationDate":"2018-10-10T02:11:02.493","AcceptedAnswerId":null,"OwnerUserId":8729804.0,"Title":"iPython Kernel dies when importing tensorflow and keras","Body":"<p>I am using spyder 3.3.1 within Anaconda build and iPython console version is 7.0.1\nWindows 10, Asus laptop.<\/p>\n\n<p>I have installed tensorflow via anaconda prompt:<\/p>\n\n<p>pip install tensorflow==1.5<\/p>\n\n<p>I have gone through all threads with similar options and have found that none of the solutions have worked for me. I am a novice with tensorflow and just starting to learn it and keras. <\/p>\n\n<p>In ipython:\nimport tensorflow as tf<\/p>\n\n<p>kernel died, restarting<\/p>\n\n<p>I have tried uninstalling all tensorflow packages using the anaconda prompt and then reinstalled as indicated above. \nThanks for the help.<\/p>\n","answers":[]}
+{"QuestionId":52731782,"AnswerCount":1,"Tags":"<python><deep-learning><anaconda><pytorch>","CreationDate":"2018-10-10T02:32:55.840","AcceptedAnswerId":null,"OwnerUserId":9900888.0,"Title":"Get CUDA_HOME environment path PYTORCH","Body":"<p>I have cuda installed via anaconda on my system which has 2 GPUs which is getting recognized by my python. <\/p>\n\n<pre><code>import torch\ntorch.cuda.is_available()\ntrue\n<\/code><\/pre>\n\n<p>However when I try to run a model via its C API, I m getting following error:<\/p>\n\n<pre><code>~\/anaconda3\/lib\/python3.6\/site-packages\/torch\/utils\/cpp_extension.py in _join_cuda_home(*paths)\n722     '''\n723     if CUDA_HOME is None:\n--&gt; 724         raise EnvironmentError('CUDA_HOME environment variable is not set. '\n725                                'Please set it to your CUDA install root.')\n726     return os.path.join(CUDA_HOME, *paths)\n\nOSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/lfd.readthedocs.io\/en\/latest\/install_gpu.html\" rel=\"nofollow noreferrer\">https:\/\/lfd.readthedocs.io\/en\/latest\/install_gpu.html<\/a> page gives instruction to set up CUDA_HOME path if cuda is installed via their method. Since I have installed cuda via anaconda I don't know which path to set.<\/p>\n\n<p>I tried find method but it is returning me too many paths for cuda. Can somebody help me with the path for CUDA. Thanks in advance.<\/p>\n","answers":[{"AnswerId":"58890075","CreationDate":"2019-11-16T11:10:48.690","ParentId":null,"OwnerUserId":"12382444","Title":null,"Body":"<p>Solution to above issue!<\/p>\n\n<p>As cuda installed through anaconda is not the entire package. Please install cuda drivers manually from Nvidia Website[ <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-downloads<\/a> ]<\/p>\n\n<p>After installation of drivers, pytorch would be able to access the cuda path. You can test the cuda path using below sample code. Problem resolved!!!<\/p>\n\n<p>CHECK INSTALLATION:<\/p>\n\n<pre><code>import os\nprint(os.environ.get('CUDA_PATH'))\n<\/code><\/pre>\n\n<p>OUTPUT:<\/p>\n\n<pre><code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.1\n<\/code><\/pre>\n"}]}
+{"QuestionId":52732145,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-10T03:24:01.083","AcceptedAnswerId":null,"OwnerUserId":10481792.0,"Title":"Sampled softmax in tensorflow keras","Body":"<p>I want to do sampled softmax loss in tf keras. I defined my own model by subclassing keras Model. In init, I specify the layers I need including the last Dense projection layer. But this Dense layer shouldn't be called in training as I want to do sampled softmax and only to use it's weights and biases. Then I define the loss function like this:<\/p>\n\n<pre><code>class SampledSoftmax:\n    def init( self,\n              num_sampled,\n              num_classes,\n              projection,\n              bias,\n              hidden_size):\n        self.weights = tf.transpose(projection)\n        self.bias = bias\n        self.num_classes = num_classes\n        self.num_sampled = num_sampled\n        self.hidden_size = hidden_size\n\n    def call(self, y_true, input):\n        \"\"\" reshaping of y_true and input to make them fit each other \"\"\"\n        input = tf.reshape(input, (-1,self.hidden_size))\n        y_true = tf.reshape(y_true, (-1,1))\n\n        return tf.nn.sampled_softmax_loss(\n                   weights=self.weights,\n                   biases=self.bias,\n                   labels=y_true,\n                   inputs=input,\n                   num_sampled=self.num_sampled,\n                   num_classes=self.num_classes,\n                   partition_strategy='div')\n<\/code><\/pre>\n\n<p>It takes in the necessary parameters to initialize and the class call will be the needed sampled softmax loss function. The catch is that to add loss to model compile I need the weights etc of the last Dense. But 1) in training Dense is not included in the model, and 2) even if it does, the Dense layer would only be hooked up with input and thus get its input dimensions etc in call of my custom model. In short, the weights etc won't be available before compiling model. Can anyone offer some help to point me to the right direction?<\/p>\n\n<p>Now the code that caused it to fail. I first subclassed model as follows:<\/p>\n\n<pre><code>class LanguageModel(tf.keras.Model):\n    def __init__(self, \n                 vocal_size=15003, \n                 embedding_size=512\n                 input_len=64)\n       self.embedding = Embedding(vocal_size, embedding_size, \n                                  input_length=input_len)\n       self.lstm = LSTM(hidden_size, return_sequences=True)\n       self.dense = Dense(vocal_size, activation='softmax')\n\n   def call(self, inputs, training=False):\n       emb_out = self.embedding(inputs)\n       lstm_out = self.lstm(embrace_out)\n       res = self.dense(lstm_out)\n       if (training)\n           ''' shouldn't use the last dense as we want to do sampling'''\n           return lstm_out\n       return res\n<\/code><\/pre>\n\n<p>Then the part to train the model as below<\/p>\n\n<pre><code>sampled_loss = SampledSoftmax(num_sampled, vocal_size, \n                   model.dense.kernel, model.dense.bias,\n                   hidden_size)\n\nmodel.compile(optimizer=tf.train.RMSPropOptimizer(lr),\n              loss=sampled_loss)\n<\/code><\/pre>\n\n<p>It would fail however I play around with it, because model.dense.kernel is not accessible as by the time of compiling the model, dense layer has not been initialized in call method. Error message as below:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/lib\/python3.5\/runpy.py\", line 184, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/usr\/lib\/python3.5\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/home\/wuxinyu\/workspace\/nlu\/lm\/main.py\", line 72, in &lt;module&gt;\n    train_main()\n  File \"\/home\/wuxinyu\/workspace\/nlu\/lm\/main.py\", line 64, in train_main\n    train_model.build_lm_model()\n  File \"\/home\/wuxinyu\/workspace\/nlu\/lm\/main.py\", line 26, in build_lm_model\nself.model.dense.kernel,\nAttributeError: 'Dense' object has no attribute 'kernel'\n<\/code><\/pre>\n\n<p>BTW, the loss defined above would work in small test cases like the following.<\/p>\n\n<pre><code>x = Input(shape=(10,), name='input_x')\nemb_out = Embedding(10000,200,input_length=10)(x)\nlstm_out = LSTM(200, return_sequences=True)(emb_out)\n\ndense = Dense(10000, activation='sigmoid')\noutput = dense(lstm_out)\n\nsl = SampledSoftmax(10, 10000, dense.kernel, dense.bias)\n\nmodel = Model(inputs=x, outputs=lstm_out)\nmodel.compile(optimizer='adam', loss=sl)\nmodel.summary()\nmodel.fit(dataset, epochs=20, steps_per_epoch=5)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52732267,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-10T03:39:49.480","AcceptedAnswerId":52733251.0,"OwnerUserId":9283443.0,"Title":"Tensorflow checkpoints are being overwritten","Body":"<p>I'm training a model (a generative adversarial network) over an input-set using Tensorflow, and I would like to save model's parameters every 50 epochs.<\/p>\n\n<p>Let say that I want to train the model for 1000 epochs, and save the model's parameters every 50 epoch, which would end up having 20 different checkpoint files.<\/p>\n\n<p>By having a Session, and a Saver object, I simply use the following code to do so.<\/p>\n\n<pre><code>if num_epoch % 50 == 0:\n    saver.save(sess=sess, path='RGAN-1\/sv\/' + type_exp, global_step=num_epoch)\n<\/code><\/pre>\n\n<p>The problem is, that checkpoints are getting overwritten, and at the end of the experiment, I only have the last 6 checkpoints, while I should have 20 checkpoints.<\/p>\n\n<p>I have no idea why this is happening.<\/p>\n","answers":[{"AnswerId":"52733251","CreationDate":"2018-10-10T05:36:56.197","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p><code>tf.train.Saver<\/code> has a <code>max_to_keep<\/code> argument that is set to 5 by default. You can pass 0 to keep all checkpoints:<\/p>\n\n<pre><code>saver = tf.train.Saver(..., max_to_keep=0)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">the docs<\/a> for a full argument list.<\/p>\n"}]}
+{"QuestionId":52732523,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-10T04:12:01.127","AcceptedAnswerId":52733219.0,"OwnerUserId":7701224.0,"Title":"Input to tensorflow in_top_k should be rank 1 or rank 2?","Body":"<p>I try to experiment with in_top_k function to see what exactly this function is doing. But I found some really confusing behavior. <\/p>\n\n<p>First I coded as follows<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ntarget = tf.constant(np.random.randint(2, size=30).reshape(30,-1), dtype=tf.int32, name=\"target\")\npred = tf.constant(np.random.rand(30,1), dtype=tf.float32, name=\"pred\")\nresult = tf.nn.in_top_k(pred, target, 1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    targetVal = target.eval()\n    predVal = pred.eval()\n    resultVal = result.eval()\n<\/code><\/pre>\n\n<p>Then it generates the following error:<\/p>\n\n<pre><code>ValueError: Shape must be rank 1 but is rank 2 for 'in_top_k\/InTopKV2' (op: 'InTopKV2') with input shapes: [30,1], [30,1], [].\n<\/code><\/pre>\n\n<p>Then I changed my code to <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\ntarget = tf.constant(np.random.randint(2, size=30), dtype=tf.int32, name=\"target\")\npred = tf.constant(np.random.rand(30,1).reshape(-1), dtype=tf.float32, name=\"pred\")\nresult = tf.nn.in_top_k(pred, target, 1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    targetVal = target.eval()\n    predVal = pred.eval()\n    resultVal = result.eval()\n<\/code><\/pre>\n\n<p>But now the error becomes<\/p>\n\n<pre><code>ValueError: Shape must be rank 2 but is rank 1 for 'in_top_k\/InTopKV2' (op: 'InTopKV2') with input shapes: [30], [30], [].\n<\/code><\/pre>\n\n<p>So should the input be rank 1 or rank 2?<\/p>\n","answers":[{"AnswerId":"52733219","CreationDate":"2018-10-10T05:33:43.967","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>For <code>in_top_k<\/code>, the <code>targets<\/code> need to be rank 1 (class indices) and the <code>predictions<\/code> rank 2 (scores for each class). This can be seen <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/in_top_k\" rel=\"nofollow noreferrer\">from the docs<\/a> easily.<br>\nThis means that the two error messages actually complain about <em>different<\/em> inputs each time (targets the first time and predictions the second time), which funnily enough isn't mentioned in the messages at all... Either way, the following snippet should be more like it:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ntarget = tf.constant(np.random.randint(2, size=30), dtype=tf.int32, name=\"target\")\npred = tf.constant(np.random.rand(30,1), dtype=tf.float32, name=\"pred\")\nresult = tf.nn.in_top_k(pred, target, 1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    targetVal = target.eval()\n    predVal = pred.eval()\n    resultVal = result.eval()\n<\/code><\/pre>\n\n<p>Here, we basically combine the \"best of both snippets\": Predictions from the first one and targets from the second one. However, the way I understand the docs, even for binary classification we need <em>two<\/em> values for the predictions, one for each class. So something like<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ntarget = tf.constant(np.random.randint(2, size=30), dtype=tf.int32, name=\"target\")\npred = tf.constant(np.random.rand(30,1), dtype=tf.float32, name=\"pred\")\npred = tf.concat((1-pred, pred), axis=1)\nresult = tf.nn.in_top_k(pred, target, 1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    targetVal = target.eval()\n    predVal = pred.eval()\n    resultVal = result.eval()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52733031,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras><gpu>","CreationDate":"2018-10-10T05:13:52.400","AcceptedAnswerId":null,"OwnerUserId":3444294.0,"Title":"Tensorflow GPU Memory Allocation Affects Model Accuracy","Body":"<p>I have a neural net in Keras 2.2.2, Tensorflow 1.10.0, Ubuntu 18, GTX1080Ti. I trained the model using a) all of the GPU memory and b) approximately half (0.47) of the memory. I get dramatically different results between these cases. Full memory - good result. Half memory - bad. I have repeated these tests and the results are consistent. The model code is too elaborate to show here but I am wondering if this is a known general effect. I set the memory allocation with this code:<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.per_process_gpu_memory_fraction = 0.47\nsession = tf.Session(config=config)\nK.set_session(session)\n<\/code><\/pre>\n\n<p>Batch size is calculated in the code at runtime and is not affected by GPU memory allocation.<\/p>\n","answers":[]}
+{"QuestionId":52733216,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-10T05:33:24.423","AcceptedAnswerId":null,"OwnerUserId":10472174.0,"Title":"How to get the probability percentage in keras predicting model CNN","Body":"<p>Here, I am getting the data as [0 1 0 0] or [0 0 0 1],---\nI get it that it is telling me that [0 1 0 0] is label2,[0 0 0 1] is label4, [1 0 0 0] is label1, [0 0 1 0] is label3.<\/p>\n\n<pre><code>import pickle\nfrom keras.preprocessing.sequence import pad_sequences\n\nMAX_SEQUENCE_LENGTH = 1000\nMAX_NB_WORDS = 20000\n\nwith open ('textsdata', 'rb') as fp:\n    texts = pickle.load(fp)\n\ntokenizer = Tokenizer(num_words=MAX_NB_WORDS)\ntokenizer.fit_on_texts(texts)\nsequences = tokenizer.texts_to_sequences(texts)\ninputquery = [\"Play some music will ya\"]\nsequences = tokenizer.texts_to_sequences(inputquery)\nmodel = load_model('my_model.h5')\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['acc'])\nprint(\"sequences\", sequences)\n\ndata = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)\nclasses = model.predict(data)\ny_classes = classes.argmax(axis=-1)\nprint(y_classes)\n<\/code><\/pre>\n\n<p>I need it in percentages like it is confident that it is label1 as 0.67\nThe value before softmax or the value of that which is confident enough to tell it is label1 or label2 or label3 or label4-- <\/p>\n\n<p>I need the percentage of it being any one of them or all of those percentages like...<\/p>\n\n<p>If a input is given output be like<\/p>\n\n<p>Class1 - 0.87<\/p>\n\n<p>Class2 - 0.3<\/p>\n\n<p>Class3 - 0.5<\/p>\n\n<p>Class4 - 0.5\nHow can I get this kind of output rather than just [1 0 0 0]\nWhat should i add next to the code above please do tell<\/p>\n","answers":[{"AnswerId":"55135601","CreationDate":"2019-03-13T06:21:38.010","ParentId":null,"OwnerUserId":"8956168","Title":null,"Body":"<pre><code>from keras.models import load_model\n\nfrom keras.preprocessing import image\n\nmodel=load_model(\"\/blah\/blah\/blah\")\n\nimg = image.load_img(path, color_mode = \"grayscale\", target_size=(128, 128, 1))\n\ny = image.img_to_array(img)\n\ny = np.expand_dims(y, axis=0)\n\nimages = np.vstack([y])\n\nclasses = model.predict(images\/255.0, batch_size=8, verbose=0)\n<\/code><\/pre>\n"},{"AnswerId":"55135769","CreationDate":"2019-03-13T06:35:39.633","ParentId":null,"OwnerUserId":"5681574","Title":null,"Body":"<p>There is method named <code>predict_proba<\/code> that returns probabilities of individual class instead of class prediction. This can be used as <\/p>\n\n<pre><code>probabilities = model.predict_proba(data)\n<\/code><\/pre>\n\n<p>Find more information in this <a href=\"https:\/\/machinelearningmastery.com\/how-to-make-classification-and-regression-predictions-for-deep-learning-models-in-keras\/\" rel=\"nofollow noreferrer\">blog<\/a>.<\/p>\n"}]}
+{"QuestionId":52733263,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T05:38:09.797","AcceptedAnswerId":null,"OwnerUserId":10309582.0,"Title":"The Way to Connect Multiple Neural Networks in a Series(Not Parallel)","Body":"<p>I wonder there is any way to connect multiple NN as a series in tensorflow.\nFor example, input features to DNN structure, and get the result values for input data of RNN structure.<\/p>\n\n<p>Example code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na = 50 #batch_size\nb = 60 #sequence in RNN\nc = 40 #features\nd = 6  #label classes\nrnn_size = b\nx_data = np.random.rand(a,b,c)\ny_data = np.random.randint(0,high=d,size=[a,1])\n\ntf.reset_default_graph()\nX = tf.placeholder(tf.float32, shape=[None,b,c])\nY = tf.placeholder(tf.float32, shape=[None,d])\nX = tf.transpose(X, (1,0,2))\nX = tf.reshape(X, (-1,c))\nX = tf.split(X, b)\nhidden_units = [40,20,10]\n\n#DNN Structure\ndnn = []\nfor i in range(len(hidden_units)):\n    if i == 0:\n        T = X\n    else:\n        T = dnn[-1]\n    dnn.append(tf.layers.dense(T, hidden_units[i], activation=tf.nn.relu, kernel_initializer=tf.contrib.layers.xavier_initializer()))\n\n# RNN Structure\nrnn = {'w': tf.Variable(tf.random_normal([rnn_size, d], stddev = 0.01), dtype=tf.float32),\n       'b': tf.Variable(tf.random_normal([d], stddev = 0.01), dtype=tf.float32)}\ncell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size)\noutputs, states = tf.contrib.rnn.static_rnn(cell, dnn[-1], dtype=tf.float32)\noutput = tf.matmul(outputs[-1], rnn['w'])+rnn['b']\n\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y,logits=output))\noptimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)\ncorrect = tf.equal(tf.argmax(output,1),tf.argmax(cost,1))\nacc = tf.reduce_mean(tf.cast(correct, tf.float32))\n\n# Run Session\nsess = tf.Session()\nsess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])\n_, c = sess.run([optimizer, cost],feed_dict={X: x_data, Y: tf.Session().run(tf.one_hot(y_data), d)})\nprint('Accuracy: ', sess.run(acc, feed_dict={X: x_data, Y: tf.Session().run(tf.one_hot(y_data), d)}))\n<\/code><\/pre>\n\n<p>When I run this code, there is an error raised:<\/p>\n\n<pre><code>File \"C:\\Anaconda3\\Lib\\site-packages\\tensorflow\\python\\layers\\core.py\", line 250, in dense\ndtype=inputs.dtype.base_dtype,\n\nAttributeError: 'list' object has no attribute 'dtype'\n<\/code><\/pre>\n\n<p>it seems to be related with type of 'dnn[-1]'<\/p>\n\n<p>Is there a connective function or data type controller for the connection of the neural networks?<\/p>\n","answers":[{"AnswerId":"52737247","CreationDate":"2018-10-10T09:44:14.987","ParentId":null,"OwnerUserId":"10309582","Title":null,"Body":"<p>I've solved the problem, finally.\nThe reason of error was little bit ambiguous but X was recognized by 'list' after running 'tf.split', finitely...\nAfter I generate a list of DNN Structures which as a length of the sequence, as following:\n    seq = []\n    for i in range(b):\n        ###dnn structure for i-th array of split###\n        seq.append(dnn structure)<\/p>\n\n<p>and tuned some codes, then the whole code worked well.\nThanks for an attention :)<\/p>\n"}]}
+{"QuestionId":52733624,"AnswerCount":0,"Tags":"<android><firebase><tensorflow><android-bitmap><firebase-mlkit>","CreationDate":"2018-10-10T06:07:42.757","AcceptedAnswerId":null,"OwnerUserId":7153699.0,"Title":"Convert a Bitmap to a N*N*N*N Byte Array for Firebase ML Kit","Body":"<p>I am trying to use the android camera to feed an image as input to my Convolutional Neural Network Model using Firebase ML Kit. <\/p>\n\n<p>I have already uploaded the model to Firebase (as a tf lite model). <\/p>\n\n<p>My model takes a (56, 75, 3) pixel image as input and attempts to classify the image into 4 classes. <\/p>\n\n<p>I am struggling to resize the bitmap properly to fit inside a byte array which is {1, 56, 75, 3} <\/p>\n\n<pre><code>bitmap = android.provider.MediaStore.Images.Media.getBitmap(cr, selectedImage); \/\/getting bitmap from camera\n\n\/\/Boilerplate code to set up Firebase\n\nFirebaseModelInputOutputOptions inputOutputOptions = new FirebaseModelInputOutputOptions.Builder().setInputFormat(0, FirebaseModelDataType.BYTE, new int[]{1, 56, 75, 3}).setOutputFormat(0, FirebaseModelDataType.BYTE, new int[]{1, 4}).build();\n\nBitmap scaled_bitmap Bitmap.createScaledBitmap(bitmap, 56, 75, true)\n<\/code><\/pre>\n\n<p>How do I convert the Bitmap to a [1][56][75][3] byte array so I can feed in the bitmap to the model?<\/p>\n\n<pre><code>byte[][][][] input = new byte[1][56][75][3];\ninput = ?? need help here to convert bitmap to byte array\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52733761,"AnswerCount":1,"Tags":"<python><tensorflow><keras><anaconda>","CreationDate":"2018-10-10T06:16:51.970","AcceptedAnswerId":null,"OwnerUserId":8560600.0,"Title":"Tensorflow installed by Keras overrides Tensorflow-GPU Installation","Body":"<p>I am having a slight issue while trying to make use of <strong>keras.preprocessing.text.Tokenizer<\/strong> class. I've recently installed Tensorflow-GPU and everything worked perfectly fine until I tried to install Keras.<\/p>\n\n<p>When I installed Keras using Anaconda, I saw that it installed Tensorflow 1.3 as well. Thing is, it did not DOWNGRADE previously installed Tensorflow-GPU version. Now that I try to run my python code, nothing happens, here is screenshot (Note: there should be some printing and processing happening):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HnRox.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HnRox.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When I try to run python shell and check for Tensorflow version, I get the one that Keras installed. So I tried to install Tensorflow-GPU again, but it saids that all requirements have been satisfied.<\/p>\n\n<p>How do I get the functionality I need from Keras, but still keep my Tensorflow-GPU installed? And why is my code not even trying to run even on Tensorflow 1.3 that Keras installed with it?<\/p>\n","answers":[{"AnswerId":"52765915","CreationDate":"2018-10-11T17:32:19.137","ParentId":null,"OwnerUserId":"8560600","Title":null,"Body":"<p>I have fixed the issue by using <code>pip install keras<\/code> instead of <code>conda install -c anaconda keras<\/code>. By using pip I am only installing keras, and not dependencies that Anaconda includes with it.<\/p>\n"}]}
+{"QuestionId":52734397,"AnswerCount":2,"Tags":"<python><tensorflow><google-compute-engine><google-colaboratory>","CreationDate":"2018-10-10T06:59:07.063","AcceptedAnswerId":null,"OwnerUserId":8315360.0,"Title":"Python script runs slower on google compute instance as compared with colaboratory","Body":"<p>Why is my machine learning python\/tensorflow script runs faster on colab as compared with a 24-vCPU google compute instance?<\/p>\n\n<p>Invocation on colab: <code>!\/content\/myscript.py<\/code> <\/p>\n\n<p>Invocation on google compute instance: <code>!\/home\/prj1\/myscript.py<\/code><\/p>\n\n<p>epoch time on colab: 0.8 s<\/p>\n\n<p>epoch time on google compute instance: 2.0 s<\/p>\n\n<p>In both cases, I am using tensorflow 1.11 and python 2.7, myscript.py is the only program running, and GPU is not being used. The script displays training progress (text-only, no graphics) on the screen every 10 epochs.<\/p>\n","answers":[{"AnswerId":"52803482","CreationDate":"2018-10-14T14:07:14.467","ParentId":null,"OwnerUserId":"8315360","Title":null,"Body":"<p>I was able to bring the epoch time down to 1.1 s by following the optimization recommended at <a href=\"https:\/\/software.intel.com\/en-us\/articles\/tips-to-improve-performance-for-popular-deep-learning-frameworks-on-multi-core-cpus\" rel=\"nofollow noreferrer\">Tips to Improve Performance for Popular Deep Learning Frameworks on CPUs<\/a><\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import os\nN_CORES                       = int(os.cpu_count()\/2)\nos.environ[\"OMP_NUM_THREADS\"] = str(N_CORES)\nos.environ[\"KMP_BLOCKTIME\"]   = \"30\"\nos.environ[\"KMP_SETTINGS\"]    = \"1\"\nos.environ[\"KMP_AFFINITY\"]    = \"granularity=fine,verbose,compact,1,0\"\nconfig_sess                   = tf.ConfigProto(intra_op_parallelism_threads=N_CORES, inter_op_parallelism_threads=2, allow_soft_placement=True, device_count = {'CPU': N_CORES})\n\nwith tf.Session(config=config_sess) as sess:\n<\/code><\/pre>\n"},{"AnswerId":"52748051","CreationDate":"2018-10-10T20:10:27.090","ParentId":null,"OwnerUserId":"9932473","Title":null,"Body":"<p>Colaboratory is optimized for Tensorflow, while a GCE instance is \na regular machine. <a href=\"https:\/\/medium.com\/tensorflow\/colab-an-easy-way-to-learn-and-use-tensorflow-d74d1686e309\" rel=\"nofollow noreferrer\">\"When you create a new notebook on colab.research.google.com, TensorFlow is already pre-installed and optimized for the hardware being used.\"<\/a> This is likely why you are seeing a performance difference between the two. <\/p>\n"}]}
+{"QuestionId":52734659,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp><lstm>","CreationDate":"2018-10-10T07:16:29.143","AcceptedAnswerId":52735372.0,"OwnerUserId":7115193.0,"Title":"Meaning of the hidden state in Keras LSTM","Body":"<p>Since I am new to deep learning this question may be funny to you. but I couldn't visualize it in the mind. That's why I am asking about it. <\/p>\n\n<p>I am giving a sentence as the vector to the LSTM, Think I have a sentence contains 10 words. Then I change those sentences to the vectors and giving it to the LSTM.<\/p>\n\n<p>The length of the LSTM cells should be 10. But in most of the tutorials, I have seen they have added 128 hidden states. I couldn't understand and visualize it. What's that the word means by LSTM layer with <code>\"128-dimensional hidden state\"<\/code> <\/p>\n\n<p>for example:<\/p>\n\n<pre><code>X = LSTM(128, return_sequences=True)(embeddings)\n<\/code><\/pre>\n\n<p>The summery of this looks<\/p>\n\n<pre><code>lstm_1 (LSTM)                (None, 10, 128)           91648    \n<\/code><\/pre>\n\n<p>Here It looks like 10 LSTM cells are added but why are that 128 hidden states there? Hope you may understand what I am expecting. <\/p>\n","answers":[{"AnswerId":"52735372","CreationDate":"2018-10-10T08:00:40.753","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p><strong>Short Answer:<\/strong>\nIf you are more familiar with Convolutional Networks, you can thick of the size of the LSTM layer (128) is the equivalent to the size of a Convolutional layer. The 10 only means that the size of your input (lenght of your sequence is 10)<\/p>\n\n<p><strong>Longer Answer:<\/strong>\nYou can check this article for more detail <a href=\"https:\/\/towardsdatascience.com\/recurrent-neural-networks-rnns-3f06d7653a85\" rel=\"nofollow noreferrer\">article about RNNs<\/a>.<\/p>\n\n<p>In the left image, a LSTM layer is represented with (<code>xt<\/code>) as the input with output (<code>ht<\/code>). The feedback arrow shows that there is some kind of memory inside the cell.<\/p>\n\n<p>In practice in Keras (right image), this model is \"unrolled\" to give the whole input <code>xt<\/code> in parallel to our layer.<\/p>\n\n<p>So when your summary is:\n    lstm_1 (LSTM)                (None, 10, 128)           91648<br>\nIt means that your input sequence is 10 (<code>x0,x1,x2,...,x9<\/code>), and that the size of your LSTM is 128 (128 will be the dimension of your output <code>ht<\/code>)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/psVGS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/psVGS.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":52734928,"AnswerCount":1,"Tags":"<python><keras><keras-rl>","CreationDate":"2018-10-10T07:33:13.517","AcceptedAnswerId":52735507.0,"OwnerUserId":2253003.0,"Title":"How to print Tensor value, Keras\/Tensorflow","Body":"<p>I am trying to understand some code from a reinforcement learning algorithm. In order to do that I am trying to print the value of a tensor.<\/p>\n\n<p>I made a simple piece of code to show what I mean.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n\nx = K.abs(-2.0)\ntf.Print(x,[x], 'x')\n<\/code><\/pre>\n\n<p>The goal is to have the value '2' printed(the absolute value of -2). But I only get back the following:<\/p>\n\n<pre><code>Using TensorFlow backend.\n\nProcess finished with exit code 0\n<\/code><\/pre>\n\n<p>Nothing, how can I print the value '2' just like a print('...') statement would do?<\/p>\n","answers":[{"AnswerId":"52735507","CreationDate":"2018-10-10T08:09:45.330","ParentId":null,"OwnerUserId":"5055862","Title":null,"Body":"<p>If you are using Jupyter Notebook, then <code>tf.Print()<\/code> so far isn't compatible and would be printing the output to Notebook's server output as is described in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\">docs<\/a><\/p>\n\n<p>In the tensorflow documentation, here is how Tensors are described:<\/p>\n\n<blockquote>\n  <p>When writing a TensorFlow program, the main object you manipulate and pass around is the tf.Tensor. <strong>A tf.Tensor object represents a partially defined computation that will eventually produce a value.<\/strong><\/p>\n<\/blockquote>\n\n<p>Hence, you would have to initialize them with a <code>tf.Session()<\/code> to get their value. To print the value, you <code>eval()<\/code> <\/p>\n\n<p>Here is the code you want:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n\nx= K.abs(-2.0)\nwith tf.Session() as sess:\n    init = tf.global_variables_initializer()\n    sess.run(init)\n    print(x.eval())\n<\/code><\/pre>\n\n<p>The initializer is important to actually initialize x.<\/p>\n"}]}
+{"QuestionId":52735523,"AnswerCount":0,"Tags":"<neural-network><keras><deep-learning>","CreationDate":"2018-10-10T08:10:56.753","AcceptedAnswerId":null,"OwnerUserId":8688652.0,"Title":"what is val_sample_weight in keras?","Body":"<p>What does the val_sample_weight option in model.fit do?\nI understand that it enables calculating weighted matrices (weighted accuracy for example) for validation set, but does (and if so, how?) it effect the calculation of the loss for the validation set?\nBecause I get different results for the loss of the validation set when using it vs not using it.<\/p>\n\n<p>I use class_weight in both scenarios.<\/p>\n\n<p>In the Keras website under model.fit it doesn't really say.<\/p>\n","answers":[]}
+{"QuestionId":52735620,"AnswerCount":0,"Tags":"<ubuntu><tensorflow><bazel><arm64><nvidia-jetson>","CreationDate":"2018-10-10T08:17:25.537","AcceptedAnswerId":null,"OwnerUserId":10478418.0,"Title":"Build TensorFlow on Jetson TX2","Body":"<p>I'm trying <a href=\"https:\/\/www.jetsonhacks.com\/2018\/03\/26\/build-tensorflow-on-nvidia-jetson-tx-development-kits\/\" rel=\"nofollow noreferrer\">this tutorial<\/a> but I had some compile error.\nCan anyone give me some advice, please?<\/p>\n\n<p>My environment is...<\/p>\n\n<ul>\n<li>Ubuntu 16.04<\/li>\n<li>JetsonTX2<\/li>\n<li>Architecture aarch64<\/li>\n<li>JetPack 3.2<\/li>\n<li>CUDA 9.0<\/li>\n<li>cuDNN 7.0.5<\/li>\n<\/ul>\n\n<p>I did the following command:<\/p>\n\n<pre><code>$ sudo -H bash BuildTensorflow.sh -b r1.6\n<\/code><\/pre>\n\n<p>and the error is...<\/p>\n\n<pre><code>    swapon: \/home\/nvidia\/JetsonTFBuild\/swapfile.swap: swapon failed: Device or resource busy\n    Looks like Swap not desired or is already in use\n    dirname: missing operand\n    Try 'dirname --help' for more information.\n    dirname: missing operand\n    Try 'dirname --help' for more information.\n    PYTHON_BIN_PATH=\/usr\/bin\/python2\n    GCC_HOST_COMPILER_PATH=\/usr\/bin\/gcc\n    CUDA_TOOLKIT_PATH=\n    TF_CUDA_VERSION=9.0\n    TF_CUDA_COMPUTE_CAPABILITIES=5.3,6.2\n    CUDNN_INSTALL_PATH=\/usr\/lib\/aarch64-linux-gnu\n    TF_CUDNN_VERSION=7.0.5\n    .\/tf_build.sh: line 20: \/home\/nvidia\/JetsonTFBuild: Is a directory\n    \/usr\/bin\/python2: can't open file 'configure.py': [Errno 2] No such file or directory\n    error: patch failed: tensorflow\/contrib\/lite\/kernels\/internal\/BUILD:21\n    error: tensorflow\/contrib\/lite\/kernels\/internal\/BUILD: patch does not apply\n    INFO: Options provided by the client:\n            Inherited 'common' options: --isatty=1 --terminal_columns=80\n    INFO: Reading rc options for 'build' from \/home\/nvidia\/JetsonTFBuild\/tensorflow\/tools\/bazel.rc:\n            'build' options: --define framework_shared_object=true --define=use_fast_cpp_protos=true --define=allow_oversize_protos=true --define=grpc_no_ares=true --spawn_strategy=standalone --genrule_strategy=standalone -c opt\n    ERROR: Config value opt is not defined in any .rc file\n    Tue Oct 9 20:34:03 JST 2018 : === Using tmpdir: \/tmp\/tmp.swhwrsSoaA\n    cp: cannot stat 'bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package.runfiles\/org_tensorflow\/tensorflow': No such file or directory\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52735895,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-lite>","CreationDate":"2018-10-10T08:31:49.363","AcceptedAnswerId":null,"OwnerUserId":6695892.0,"Title":"Problem converting Keras model to Tensorflow-Lite using TocoConverter.from_keras_model_file","Body":"<p>I'm facing an issue using TOCO to convert a Keras model to TfLite.  <\/p>\n\n<p>Followed the guide of: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/lite\/TocoConverter\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/lite\/TocoConverter<\/a><\/p>\n\n<p>How I use TOCO:<\/p>\n\n<pre><code>def create_lite_model(keras_model_file):\n    lite_model_name = 'lite_model_file.tflite'\n    tf_lite_graph = os.path.join(WEIGHTS_DIRECTORY, lite_model_name)\n    converter = tf.contrib.lite.TocoConverter.from_keras_model_file(keras_model_file)\n    tf_lite_model = converter.convert()\n    open(tf_lite_graph, \"wb\").write(tf_lite_model)\n<\/code><\/pre>\n\n<p>Getting the following error:<\/p>\n\n<pre><code>File \"\/tensorflow\/contrib\/lite\/python\/lite.py\", line 356, in from_keras_model_file\nkeras_model = _keras.models.load_model(model_file)\nFile \"\/tensorflow\/python\/keras\/engine\/saving.py\", line 251, in load_model\ntraining_config['weighted_metrics'])\nKeyError: 'weighted_metrics'\n<\/code><\/pre>\n\n<p>Does anybody has a solution for this problem?<\/p>\n","answers":[{"AnswerId":"52753568","CreationDate":"2018-10-11T06:25:56.420","ParentId":null,"OwnerUserId":"6695892","Title":null,"Body":"<p>Until now I didn't found a solution, but I'm using a workaround.<\/p>\n\n<p>Converting the Keras model to a tf Graph using a SaveBuilder to store the tf Graph and finally using TocoConverter.from_saved_model(...).<\/p>\n\n<pre><code>import os\n\nimport tensorflow as tf\nfrom keras import backend as K\nfrom keras.models import load_model\n\nK.set_learning_phase(False)\n\n\ndef freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):\n    graph = session.graph\n    with graph.as_default():\n        freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))\n        output_names = output_names or []\n        output_names += [v.op.name for v in tf.global_variables()]\n        input_graph_def = graph.as_graph_def()\n        if clear_devices:\n            for node in input_graph_def.node:\n                node.device = \"\"\n        frozen_graph = tf.graph_util.convert_variables_to_constants(session, input_graph_def, output_names,                                                               freeze_var_names)\n        return frozen_graph\n\n\ndef create_lite_model_from_saved_model(saved_model_dir, tf_lite_path):\n    converter = tf.contrib.lite.TocoConverter.from_saved_model(saved_model_dir)\n    tf_lite_model = converter.convert()\n    open(tf_lite_path, \"wb\").write(tf_lite_model)\n\n\ndef save_model(keras_model, session, pb_model_path):\n    x = keras_model.input\n    y = keras_model.output\n    prediction_signature = tf.saved_model.signature_def_utils.predict_signature_def({\"inputs\": x}, {\"prediction\": y})\n    builder = tf.saved_model.builder.SavedModelBuilder(pb_model_path)\n    legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')\n    signature = {tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature, }\n    builder.add_meta_graph_and_variables(session, [tf.saved_model.tag_constants.SERVING], signature_def_map=signature,                                         legacy_init_op=legacy_init_op)\n    builder.save()\n\n\ndef run():\n    sess = K.get_session()\n    keras_model_name = 'keras_model.h5'\n    lite_model_name = 'lite_model_file.tflite'\n    keras_model_file_path = os.path.join('.\/weights', keras_model_name)\n    lite_model_file_path = os.path.join('.\/weights', lite_model_name)\n    pb_model_path = os.path.join('.\/weights', 'saveBuilder')\n    model = load_model(keras_model_file_path)\n    output_names = [node.op.name for node in model.outputs]\n    _ = freeze_session(sess, output_names=output_names)\n    save_model(keras_model=model, session=sess, pb_model_path=pb_model_path)\n    create_lite_model_from_saved_model(saved_model_dir=pb_model_path, tf_lite_path=lite_model_file_path)\n\n\nif __name__ == \"__main__\":\n    run()\n<\/code><\/pre>\n\n<p>Maybe it's helpful for someone.<\/p>\n\n<p>The feeze_session(...) function I used  is from: <a href=\"https:\/\/stackoverflow.com\/questions\/45466020\/how-to-export-keras-h5-to-tensorflow-pb\">How to export Keras .h5 to tensorflow .pb?<\/a><\/p>\n"},{"AnswerId":"52971747","CreationDate":"2018-10-24T14:38:32.000","ParentId":null,"OwnerUserId":"2344535","Title":null,"Body":"<p>If this is a model you've written, make sure you define weighted_metrics in the model compilation:<\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy', optimizer=&lt;some_optimizer&gt;, metrics=['accuracy'], weighted_metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52735986,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-10T08:37:56.720","AcceptedAnswerId":null,"OwnerUserId":10483011.0,"Title":"Hands on TensorFlow","Body":"<p>I am new to deep learning. I want to test my data on TesonFlow. I have installed TensorFlow through anaconda. Please guide me how to use it for testing my data. I have both windows and linux os. Kindly share some handson for my problem.<\/p>\n","answers":[{"AnswerId":"52780209","CreationDate":"2018-10-12T13:08:21.153","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>You can build a simple feed forward network using Keras ,a high level API build on top of Tensorflow. \nLearn more from <a href=\"https:\/\/heartbeat.fritz.ai\/introduction-to-deep-learning-with-keras-c7c3d14e1527\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52736517,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-10-10T09:05:49.220","AcceptedAnswerId":null,"OwnerUserId":10481792.0,"Title":"Tensorflow keras with tf dataset input","Body":"<p>I'm new to tensorflow keras and dataset. Can anyone help me understand why the following code doesn't work?<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.keras as keras\nimport numpy as np\nfrom tensorflow.python.data.ops import dataset_ops\nfrom tensorflow.python.data.ops import iterator_ops\nfrom tensorflow.python.keras.utils import multi_gpu_model\nfrom tensorflow.python.keras import backend as K\n\n\ndata = np.random.random((1000,32))\nlabels = np.random.random((1000,10))\ndataset = tf.data.Dataset.from_tensor_slices((data,labels))\nprint( dataset)\nprint( dataset.output_types)\nprint( dataset.output_shapes)\ndataset.batch(10)\ndataset.repeat(100)\n\ninputs = keras.Input(shape=(32,))  # Returns a placeholder tensor\n\n# A layer instance is callable on a tensor, and returns a tensor.\nx = keras.layers.Dense(64, activation='relu')(inputs)\nx = keras.layers.Dense(64, activation='relu')(x)\npredictions = keras.layers.Dense(10, activation='softmax')(x)\n\n# Instantiate the model given inputs and outputs.\nmodel = keras.Model(inputs=inputs, outputs=predictions)\n\n# The compile step specifies the training configuration.\nmodel.compile(optimizer=tf.train.RMSPropOptimizer(0.001),\n          loss='categorical_crossentropy',\n          metrics=['accuracy'])\n\n# Trains for 5 epochs\nmodel.fit(dataset, epochs=5, steps_per_epoch=100)\n<\/code><\/pre>\n\n<p>It failed with the following error:<\/p>\n\n<pre><code>model.fit(x=dataset, y=None, epochs=5, steps_per_epoch=100)\nFile \"\/home\/wuxinyu\/pyEnv\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1510, in fit\nvalidation_split=validation_split)\nFile \"\/home\/wuxinyu\/pyEnv\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 994, in _standardize_user_data\nclass_weight, batch_size)\nFile \"\/home\/wuxinyu\/pyEnv\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1113, in _standardize_weights\nexception_prefix='input')\nFile \"\/home\/wuxinyu\/pyEnv\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 325, in standardize_input_data\n'with shape ' + str(data_shape))\nValueError: Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (32,)\n<\/code><\/pre>\n\n<p>According to tf.keras guide, I should be able to directly pass the dataset to model.fit, as this example shows:<\/p>\n\n<blockquote>\n  <h1>Input tf.data datasets<\/h1>\n  \n  <p>Use the Datasets API to scale to large datasets or multi-device training. Pass a tf.data.Dataset instance to the fit method:<\/p>\n<\/blockquote>\n\n<pre><code># Instantiates a toy dataset instance:\ndataset = tf.data.Dataset.from_tensor_slices((data, labels))\ndataset = dataset.batch(32)\ndataset = dataset.repeat()\n<\/code><\/pre>\n\n<blockquote>\n  <h1>Don't forget to specify <code>steps_per_epoch<\/code> when calling <code>fit<\/code> on a dataset.<\/h1>\n  \n  <p>model.fit(dataset, epochs=10, steps_per_epoch=30)\n  Here, the fit method uses the steps_per_epoch argument\u2014this is the number of training steps the model runs before it moves to the next epoch. Since the Dataset yields batches of data, this snippet does not require a batch_size.<\/p>\n  \n  <p>Datasets can also be used for validation:<\/p>\n<\/blockquote>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((data, labels))\ndataset = dataset.batch(32).repeat()\n\nval_dataset = tf.data.Dataset.from_tensor_slices((val_data, val_labels))\nval_dataset = val_dataset.batch(32).repeat()\n\nmodel.fit(dataset, epochs=10, steps_per_epoch=30,\n      validation_data=val_dataset,\n      validation_steps=3)\n<\/code><\/pre>\n\n<p>What's the problem with my code, and what's the correct way of doing it?<\/p>\n","answers":[{"AnswerId":"52741907","CreationDate":"2018-10-10T13:55:30.387","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>You are missing defining an iterator which is the reason why there is an error.<\/p>\n\n<pre><code>data = np.random.random((1000,32))\nlabels = np.random.random((1000,10))\ndataset = tf.data.Dataset.from_tensor_slices((data,labels))\ndataset = dataset.batch(10).repeat()\ninputs = Input(shape=(32,))  # Returns a placeholder tensor\n\n# A layer instance is callable on a tensor, and returns a tensor.\nx = Dense(64, activation='relu')(inputs)\nx = Dense(64, activation='relu')(x)\npredictions = Dense(10, activation='softmax')(x)\n\n# Instantiate the model given inputs and outputs.\nmodel = keras.Model(inputs=inputs, outputs=predictions)\n\n# The compile step specifies the training configuration.\nmodel.compile(optimizer=tf.train.RMSPropOptimizer(0.001),\n          loss='categorical_crossentropy',\n          metrics=['accuracy'])\n\n# Trains for 5 epochs\nmodel.fit(dataset.make_one_shot_iterator(), epochs=5, steps_per_epoch=100)\n<\/code><\/pre>\n\n<p>Epoch 1\/5\n100\/100 [==============================] - 1s 8ms\/step - loss: 11.5787 - acc: 0.1010<\/p>\n\n<p>Epoch 2\/5\n100\/100 [==============================] - 0s 4ms\/step - loss: 11.4846 - acc: 0.0990<\/p>\n\n<p>Epoch 3\/5\n100\/100 [==============================] - 0s 4ms\/step - loss: 11.4690 - acc: 0.1270<\/p>\n\n<p>Epoch 4\/5\n100\/100 [==============================] - 0s 4ms\/step - loss: 11.4611 - acc: 0.1300<\/p>\n\n<p>Epoch 5\/5\n100\/100 [==============================] - 0s 4ms\/step - loss: 11.4546 - acc: 0.1360<\/p>\n\n<p>This is the result on my system.<\/p>\n"},{"AnswerId":"54775259","CreationDate":"2019-02-19T21:34:05.130","ParentId":null,"OwnerUserId":"5339857","Title":null,"Body":"<p>To your original question as to why you're getting the error:<\/p>\n\n<pre><code>Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (32,)\n<\/code><\/pre>\n\n<p>The reason your code breaks is because you haven't applied the <code>.batch()<\/code> back to the <code>dataset<\/code> variable, like so:<\/p>\n\n<pre><code>dataset = dataset.batch(10)\n<\/code><\/pre>\n\n<p>You simply called <code>dataset.batch()<\/code>.<\/p>\n\n<p>This breaks because without the <code>batch()<\/code> the output tensors are not batched, i.e. you get shape <code>(32,)<\/code> instead of <code>(1,32)<\/code>.<\/p>\n"}]}
+{"QuestionId":52737412,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2018-10-10T09:52:26.957","AcceptedAnswerId":null,"OwnerUserId":9261771.0,"Title":"Tensorboard loss plots go back in time (Keras)","Body":"<p>I have a similar problem to the one described <a href=\"https:\/\/stackoverflow.com\/questions\/48496479\/tensorflow-1-4-1-tensorboard-visualization-plot-goes-back-in-time\">here<\/a>, only I am using Keras with Tensorflow backend rather then Tensorflow.\nI have trained 18 MLP models for time series forecasting with different meta parameters, all created with the same basic architecture. The 3 meta parameters I am scanning are the lookahead that the model uses for prediction, the depth of the model, and whether I am using L2 regularization. <\/p>\n\n<pre><code>    model = Sequential()\n    # input_shape should be a 3D tensor with shape (batch_size, timesteps ,input_dim)\n    model.add(Flatten())\n    # hidden layer sizes should drop gradually from 256 to 2*lookahead\n    hidden_layer_sizes = [int(256 - i * (256 - 2 * lookahead) \/ depth) for i in range(depth)]\n    for hidden_layer_size in hidden_layer_sizes:\n        if regularization:\n            model.add(Dense(hidden_layer_size, kernel_initializer=\"he_normal\",\n                        kernel_regularizer = regularizers.l2(0.01), activation=activations))\n        else:\n            model.add(Dense(hidden_layer_size, kernel_initializer=\"he_normal\",\n                        activation=activations))\n    model.add(Dense(2 * lookahead))\n    loss = losses.mean_squared_error\n    model.compile(loss=loss, optimizer=self.kwargs[\"optimizer\"], metrics=['mae'])\n<\/code><\/pre>\n\n<p>Each model's tensorboard data is saved in a seperate folder through the relevant Keras callback<\/p>\n\n<pre><code>callback_tensorboard = TensorBoard(log_dir=log_dir,\n                                   histogram_freq=5,\n                                   write_graph=False,\n                                   write_grads=True,\n                                   write_images=False)  \n<\/code><\/pre>\n\n<p>but for some reason, 3 out of the 18 models have two tensorboard files saved instead of one, and the resulting graphs show this weird phenomena of progressing backwards through time\n<a href=\"https:\/\/i.stack.imgur.com\/Cnnuv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Cnnuv.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Why does this happen? and other then deleting the second tensorboard file, what can I do to prevent this? <\/p>\n","answers":[{"AnswerId":"54364853","CreationDate":"2019-01-25T11:54:09.690","ParentId":null,"OwnerUserId":"6583936","Title":null,"Body":"<p>It happens because every time you call <code>mode.fit()<\/code> you reinitialize the internal epoch counter in keras. However you could set it up manually when you compile the model like:<\/p>\n\n<pre><code>model.compile(optimizer=tf.train.AdamOptimizer(),\n              loss=tf.losses.mean_squared_error)\ninit_epoch = 0\n<\/code><\/pre>\n\n<p>And then when ypu call fit, you can pass additional argument to keras, which describes current epoch: <\/p>\n\n<pre><code>epoch_count = 200\ninit_epoch += epoch_count\nhistory = model.fit(x_train, y_train,\n                batch_size=256,\n                epochs=init_epoch,\n                validation_data=(x_test, y_test),\n                verbose=1,\n                initial_epoch=init_epoch-epoch_count,\n                )  \n<\/code><\/pre>\n"}]}
+{"QuestionId":52737844,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><machine-learning><neural-network>","CreationDate":"2018-10-10T10:13:54.677","AcceptedAnswerId":52757629.0,"OwnerUserId":6288172.0,"Title":"ValueError: setting an array element with a sequence. in TensorFlow sess.run()","Body":"<p>When feeding the iteration to the session <code>ValueError: setting an array element with a sequence.<\/code> occurs at <code>_, err = sess.run([train, loss], feed_dict={image: epoch_x, label: epoch_y})<\/code><\/p>\n\n<p>Here <code>data<\/code> is 2d matrix with first column as images and second column as one hot labels.<\/p>\n\n<p>code fragment:<\/p>\n\n<pre><code>for epoch in range(epochs):\n    epoch_loss = 0\n\n    # data = np.random.shuffle(data)\n    data = shuffle(data) #sklearn.utils.shuffle\n\n    ptr = 0\n    for iter in range(int(n_examples\/batch_size)): #batch_size = 10\n        epoch_data = data[ptr : ptr + batch_size] # to get the batch\n        epoch_x = epoch_data[:, 0] # fist col - images\n        epoch_y = epoch_data[:, 1] # second col - labels\n        ptr += batch_size\n\n        _, err = sess.run([train, loss], feed_dict={image: epoch_x, label: epoch_y}) # error occurs here.\n        epoch_loss += err\n    print(\"Epoch: \", epoch, \" - Completed out of: \", epochs, \" - Loss: \", epoch_loss)\n<\/code><\/pre>\n\n<p>Here <code>epoch_x<\/code> is a batch of coloured images:<\/p>\n\n<pre><code>[array([[[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.68235294, 0.6745098 , 0.67058824],\n        [0.68627451, 0.68235294, 0.66666667],\n        [0.69019608, 0.68627451, 0.67058824]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.68235294, 0.6745098 , 0.67058824],\n        [0.67843137, 0.67058824, 0.66666667],\n        [0.68235294, 0.67843137, 0.6627451 ]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.6745098 , 0.66666667, 0.6627451 ],\n        [0.68235294, 0.6745098 , 0.67058824],\n        [0.68235294, 0.67843137, 0.6627451 ]],\n\n       ...,\n\n       [[0.39607843, 0.30980392, 0.61568627],\n        [0.4       , 0.30588235, 0.61176471],\n        [0.39215686, 0.29411765, 0.60784314],\n        ...,\n        [0.24705882, 0.18431373, 0.49019608],\n        [0.23921569, 0.17254902, 0.49019608],\n        [0.23137255, 0.15294118, 0.47843137]],\n\n       [[0.39215686, 0.29019608, 0.61568627],\n        [0.38039216, 0.27843137, 0.60392157],\n        [0.38431373, 0.28235294, 0.61176471],\n        ...,\n        [0.28235294, 0.21960784, 0.5254902 ],\n        [0.27058824, 0.19607843, 0.51372549],\n        [0.2745098 , 0.2       , 0.51764706]],\n\n       [[0.39215686, 0.28235294, 0.63137255],\n        [0.39607843, 0.28627451, 0.63529412],\n        [0.39215686, 0.28235294, 0.63529412],\n        ...,\n        [0.29411765, 0.22745098, 0.54509804],\n        [0.3254902 , 0.25098039, 0.56078431],\n        [0.30588235, 0.23137255, 0.54117647]]])\n array([[[0.85098039, 0.89803922, 0.91372549],\n        [0.85490196, 0.90196078, 0.91764706],\n        [0.85490196, 0.90196078, 0.91764706],\n        ...,\n        [0.09411765, 0.10588235, 0.09803922],\n        [0.10196078, 0.11372549, 0.10588235],\n        [0.11372549, 0.1254902 , 0.11764706]],\n\n       [[0.85882353, 0.90588235, 0.92156863],\n        [0.85882353, 0.90588235, 0.92156863],\n        [0.8627451 , 0.90980392, 0.9254902 ],\n        ...,\n        [0.10980392, 0.11764706, 0.11764706],\n        [0.11764706, 0.1254902 , 0.1254902 ],\n        [0.11372549, 0.12156863, 0.12156863]],\n\n       [[0.85882353, 0.90588235, 0.92156863],\n        [0.85882353, 0.90588235, 0.92156863],\n        [0.8627451 , 0.90980392, 0.9254902 ],\n        ...,\n        [0.14901961, 0.14509804, 0.15294118],\n        [0.15294118, 0.14901961, 0.16470588],\n        [0.14117647, 0.1372549 , 0.15294118]],\n\n       ...,\n\n       [[0.55294118, 0.52941176, 0.47843137],\n        [0.56470588, 0.54117647, 0.49019608],\n        [0.58823529, 0.55686275, 0.50588235],\n        ...,\n        [0.12941176, 0.09803922, 0.10196078],\n        [0.12941176, 0.09803922, 0.10196078],\n        [0.1254902 , 0.09411765, 0.09803922]],\n\n       [[0.60392157, 0.57647059, 0.51764706],\n        [0.58823529, 0.56078431, 0.50196078],\n        [0.60784314, 0.58039216, 0.52156863],\n        ...,\n        [0.13333333, 0.10196078, 0.10588235],\n        [0.13333333, 0.10196078, 0.10588235],\n        [0.13333333, 0.10196078, 0.10588235]],\n\n       [[0.60784314, 0.58039216, 0.52156863],\n        [0.60392157, 0.57647059, 0.50980392],\n        [0.60392157, 0.57647059, 0.50980392],\n        ...,\n        [0.12941176, 0.09803922, 0.10196078],\n        [0.12941176, 0.09803922, 0.10196078],\n        [0.1254902 , 0.09411765, 0.09803922]]])\n array([[[0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        ...,\n        [0.6627451 , 0.67058824, 0.6745098 ],\n        [0.68627451, 0.69411765, 0.69803922],\n        [0.68627451, 0.69411765, 0.69803922]],\n\n       [[0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        ...,\n        [0.67058824, 0.67843137, 0.67843137],\n        [0.67058824, 0.67843137, 0.67843137],\n        [0.69019608, 0.69803922, 0.70196078]],\n\n       [[0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        [0.68627451, 0.69411765, 0.67058824],\n        ...,\n        [0.69019608, 0.70196078, 0.69411765],\n        [0.69019608, 0.69803922, 0.69803922],\n        [0.69411765, 0.70196078, 0.70196078]],\n\n       ...,\n\n       [[0.18431373, 0.1254902 , 0.41568627],\n        [0.19607843, 0.1372549 , 0.42745098],\n        [0.18431373, 0.1254902 , 0.41568627],\n        ...,\n        [0.03529412, 0.04313725, 0.17647059],\n        [0.05098039, 0.05490196, 0.20392157],\n        [0.08235294, 0.08235294, 0.23921569]],\n\n       [[0.20392157, 0.14509804, 0.43529412],\n        [0.23137255, 0.17647059, 0.45882353],\n        [0.21960784, 0.15294118, 0.43921569],\n        ...,\n        [0.04705882, 0.05098039, 0.18431373],\n        [0.08235294, 0.07843137, 0.23529412],\n        [0.12156863, 0.11372549, 0.27843137]],\n\n       [[0.23137255, 0.17647059, 0.45882353],\n        [0.23921569, 0.18431373, 0.46666667],\n        [0.19607843, 0.12941176, 0.41568627],\n        ...,\n        [0.0745098 , 0.0745098 , 0.21568627],\n        [0.1254902 , 0.12156863, 0.27843137],\n        [0.14509804, 0.13333333, 0.30980392]]])\n array([[[0.8627451 , 0.80392157, 0.41176471],\n        [0.8627451 , 0.80392157, 0.41176471],\n        [0.8627451 , 0.80392157, 0.41176471],\n        ...,\n        [0.77647059, 0.70980392, 0.32941176],\n        [0.80784314, 0.72156863, 0.32156863],\n        [0.82352941, 0.7372549 , 0.3254902 ]],\n\n       [[0.85490196, 0.79607843, 0.40392157],\n        [0.85490196, 0.79607843, 0.40392157],\n        [0.85490196, 0.79607843, 0.40392157],\n        ...,\n        [0.77647059, 0.70588235, 0.33333333],\n        [0.79215686, 0.71372549, 0.31764706],\n        [0.82352941, 0.7372549 , 0.3254902 ]],\n\n       [[0.85490196, 0.79607843, 0.40392157],\n        [0.85490196, 0.79607843, 0.40392157],\n        [0.85490196, 0.79607843, 0.40392157],\n        ...,\n        [0.76862745, 0.70196078, 0.3372549 ],\n        [0.78431373, 0.70196078, 0.31372549],\n        [0.82745098, 0.7372549 , 0.3372549 ]],\n\n       ...,\n\n       [[0.68627451, 0.61960784, 0.05098039],\n        [0.68235294, 0.61568627, 0.04705882],\n        [0.68235294, 0.61568627, 0.04705882],\n        ...,\n        [0.30588235, 0.31372549, 0.35294118],\n        [0.40392157, 0.43529412, 0.48627451],\n        [0.49411765, 0.5254902 , 0.59215686]],\n\n       [[0.68627451, 0.61960784, 0.05098039],\n        [0.68627451, 0.61960784, 0.05098039],\n        [0.68235294, 0.61568627, 0.04705882],\n        ...,\n        [0.33333333, 0.34901961, 0.39215686],\n        [0.42352941, 0.45490196, 0.52156863],\n        [0.51764706, 0.56470588, 0.63529412]],\n\n       [[0.68627451, 0.61960784, 0.05098039],\n        [0.68627451, 0.61960784, 0.05098039],\n        [0.68235294, 0.61568627, 0.04705882],\n        ...,\n        [0.39607843, 0.41960784, 0.4627451 ],\n        [0.45490196, 0.49411765, 0.56078431],\n        [0.54509804, 0.58823529, 0.66666667]]])\n array([[[0.71372549, 0.71764706, 0.70980392],\n        [0.71372549, 0.71764706, 0.70980392],\n        [0.71372549, 0.71764706, 0.70980392],\n        ...,\n        [0.36078431, 0.35686275, 0.34117647],\n        [0.41960784, 0.41960784, 0.39607843],\n        [0.46666667, 0.46666667, 0.44313725]],\n\n       [[0.70980392, 0.71372549, 0.70588235],\n        [0.71372549, 0.71764706, 0.70980392],\n        [0.70980392, 0.71372549, 0.70588235],\n        ...,\n        [0.42745098, 0.42745098, 0.40392157],\n        [0.48235294, 0.48235294, 0.45882353],\n        [0.51764706, 0.51764706, 0.49411765]],\n\n       [[0.70980392, 0.71372549, 0.70588235],\n        [0.70980392, 0.71372549, 0.70588235],\n        [0.70980392, 0.71372549, 0.70588235],\n        ...,\n        [0.45882353, 0.46666667, 0.44313725],\n        [0.51764706, 0.5254902 , 0.50196078],\n        [0.53333333, 0.54117647, 0.51764706]],\n\n       ...,\n\n       [[0.44313725, 0.34509804, 0.69803922],\n        [0.44313725, 0.34509804, 0.69803922],\n        [0.43921569, 0.34117647, 0.69411765],\n        ...,\n        [0.48235294, 0.36078431, 0.74901961],\n        [0.48235294, 0.36078431, 0.74901961],\n        [0.49019608, 0.36862745, 0.75686275]],\n\n       [[0.44313725, 0.34117647, 0.69411765],\n        [0.43921569, 0.3372549 , 0.69019608],\n        [0.43529412, 0.33333333, 0.68627451],\n        ...,\n        [0.48235294, 0.36078431, 0.74901961],\n        [0.47843137, 0.35686275, 0.74509804],\n        [0.47843137, 0.35686275, 0.74509804]],\n\n       [[0.43137255, 0.32941176, 0.68235294],\n        [0.42745098, 0.3254902 , 0.67843137],\n        [0.42745098, 0.31372549, 0.67843137],\n        ...,\n        [0.48627451, 0.36470588, 0.75294118],\n        [0.48627451, 0.36470588, 0.75294118],\n        [0.4745098 , 0.34901961, 0.74509804]]])\n array([[[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69803922, 0.69019608, 0.68627451],\n        [0.70588235, 0.69803922, 0.69411765],\n        [0.70980392, 0.70196078, 0.69803922]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69803922, 0.69019608, 0.68627451],\n        [0.70196078, 0.69411765, 0.69019608],\n        [0.70588235, 0.69803922, 0.69411765]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69803922, 0.69019608, 0.68627451],\n        [0.69803922, 0.69019608, 0.68627451],\n        [0.69803922, 0.69019608, 0.68627451]],\n\n       ...,\n\n       [[0.06666667, 0.04313725, 0.22745098],\n        [0.09019608, 0.06666667, 0.25098039],\n        [0.0745098 , 0.05098039, 0.24313725],\n        ...,\n        [0.39607843, 0.28627451, 0.63529412],\n        [0.39215686, 0.29019608, 0.61960784],\n        [0.38823529, 0.29411765, 0.6       ]],\n\n       [[0.04705882, 0.03529412, 0.21176471],\n        [0.0627451 , 0.04313725, 0.21960784],\n        [0.07058824, 0.05098039, 0.22745098],\n        ...,\n        [0.39215686, 0.28235294, 0.63529412],\n        [0.38431373, 0.27843137, 0.61568627],\n        [0.38431373, 0.28235294, 0.61176471]],\n\n       [[0.03529412, 0.02745098, 0.19215686],\n        [0.03921569, 0.03137255, 0.19607843],\n        [0.06666667, 0.04705882, 0.22352941],\n        ...,\n        [0.36078431, 0.25098039, 0.60392157],\n        [0.36470588, 0.25490196, 0.60784314],\n        [0.39215686, 0.28235294, 0.63529412]]])\n array([[[0.75294118, 0.75294118, 0.75294118],\n        [0.74509804, 0.74509804, 0.74509804],\n        [0.74509804, 0.74509804, 0.74509804],\n        ...,\n        [0.80784314, 0.82745098, 0.82352941],\n        [0.80392157, 0.82352941, 0.81960784],\n        [0.81960784, 0.83921569, 0.83529412]],\n\n       [[0.74901961, 0.74901961, 0.74901961],\n        [0.74509804, 0.74509804, 0.74509804],\n        [0.74901961, 0.74901961, 0.74901961],\n        ...,\n        [0.82745098, 0.84705882, 0.84313725],\n        [0.82352941, 0.84313725, 0.83921569],\n        [0.81960784, 0.83921569, 0.83529412]],\n\n       [[0.76078431, 0.76862745, 0.76862745],\n        [0.76862745, 0.76862745, 0.76862745],\n        [0.76862745, 0.76862745, 0.76862745],\n        ...,\n        [0.82352941, 0.84313725, 0.83921569],\n        [0.81960784, 0.83921569, 0.83529412],\n        [0.82352941, 0.84313725, 0.83921569]],\n\n       ...,\n\n       [[0.47058824, 0.45098039, 0.29803922],\n        [0.47058824, 0.45098039, 0.29803922],\n        [0.48235294, 0.45882353, 0.29803922],\n        ...,\n        [0.48235294, 0.47058824, 0.41176471],\n        [0.44705882, 0.43529412, 0.37647059],\n        [0.4627451 , 0.45098039, 0.39215686]],\n\n       [[0.44705882, 0.43137255, 0.27058824],\n        [0.45098039, 0.43529412, 0.2745098 ],\n        [0.4627451 , 0.43921569, 0.27843137],\n        ...,\n        [0.4627451 , 0.45490196, 0.38431373],\n        [0.45098039, 0.44313725, 0.37254902],\n        [0.46666667, 0.45882353, 0.38823529]],\n\n       [[0.45882353, 0.44313725, 0.28235294],\n        [0.42745098, 0.41176471, 0.25098039],\n        [0.42352941, 0.4       , 0.23921569],\n        ...,\n        [0.45098039, 0.44313725, 0.37254902],\n        [0.46666667, 0.45882353, 0.38823529],\n        [0.45098039, 0.44313725, 0.37254902]]])\n array([[[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69411765, 0.69803922, 0.68235294],\n        [0.70196078, 0.70588235, 0.69019608],\n        [0.69019608, 0.69411765, 0.67843137]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.67843137, 0.68235294, 0.66666667],\n        [0.69411765, 0.69803922, 0.68235294],\n        [0.69803922, 0.70196078, 0.68627451]],\n\n       [[0.72156863, 0.7254902 , 0.70980392],\n        [0.72156863, 0.7254902 , 0.70980392],\n        [0.72156863, 0.7254902 , 0.70980392],\n        ...,\n        [0.66666667, 0.67058824, 0.6627451 ],\n        [0.68235294, 0.68627451, 0.67058824],\n        [0.69803922, 0.70196078, 0.68627451]],\n\n       ...,\n\n       [[0.20392157, 0.14117647, 0.39215686],\n        [0.19607843, 0.12941176, 0.39215686],\n        [0.09803922, 0.02745098, 0.29803922],\n        ...,\n        [0.27843137, 0.21960784, 0.50980392],\n        [0.29411765, 0.21960784, 0.5372549 ],\n        [0.29411765, 0.21568627, 0.54117647]],\n\n       [[0.31764706, 0.23921569, 0.53333333],\n        [0.30588235, 0.22745098, 0.52941176],\n        [0.31764706, 0.23529412, 0.54509804],\n        ...,\n        [0.28627451, 0.22352941, 0.52941176],\n        [0.30980392, 0.23137255, 0.55686275],\n        [0.30980392, 0.22745098, 0.56078431]],\n\n       [[0.37647059, 0.28235294, 0.62745098],\n        [0.37254902, 0.27843137, 0.62352941],\n        [0.35686275, 0.25882353, 0.61176471],\n        ...,\n        [0.30588235, 0.23921569, 0.55686275],\n        [0.32156863, 0.23921569, 0.57254902],\n        [0.34117647, 0.25490196, 0.6       ]]])\n array([[[0.80784314, 0.84313725, 0.85490196],\n        [0.78431373, 0.83137255, 0.83921569],\n        [0.79215686, 0.83529412, 0.85098039],\n        ...,\n        [0.57647059, 0.59607843, 0.59215686],\n        [0.60784314, 0.63137255, 0.62745098],\n        [0.65882353, 0.68235294, 0.67843137]],\n\n       [[0.80784314, 0.84313725, 0.85490196],\n        [0.8       , 0.83529412, 0.84705882],\n        [0.8       , 0.84313725, 0.85882353],\n        ...,\n        [0.52156863, 0.54117647, 0.5372549 ],\n        [0.61176471, 0.63137255, 0.62745098],\n        [0.62352941, 0.64705882, 0.64313725]],\n\n       [[0.80784314, 0.84313725, 0.85490196],\n        [0.80784314, 0.84313725, 0.85882353],\n        [0.80392157, 0.83921569, 0.85490196],\n        ...,\n        [0.5254902 , 0.53333333, 0.53333333],\n        [0.54509804, 0.56470588, 0.56078431],\n        [0.63529412, 0.65490196, 0.65098039]],\n\n       ...,\n\n       [[0.20784314, 0.15294118, 0.16862745],\n        [0.22352941, 0.16862745, 0.18431373],\n        [0.21176471, 0.16078431, 0.16862745],\n        ...,\n        [0.76470588, 0.76470588, 0.76470588],\n        [0.76862745, 0.76862745, 0.76862745],\n        [0.76078431, 0.76862745, 0.76862745]],\n\n       [[0.21568627, 0.16078431, 0.17647059],\n        [0.22352941, 0.16862745, 0.18431373],\n        [0.21568627, 0.16470588, 0.17254902],\n        ...,\n        [0.75686275, 0.76470588, 0.76470588],\n        [0.76078431, 0.76862745, 0.76862745],\n        [0.76470588, 0.77254902, 0.77254902]],\n\n       [[0.20392157, 0.14901961, 0.16470588],\n        [0.21176471, 0.15686275, 0.17254902],\n        [0.22745098, 0.17647059, 0.18431373],\n        ...,\n        [0.76078431, 0.76862745, 0.76862745],\n        [0.76470588, 0.77254902, 0.77254902],\n        [0.76862745, 0.77647059, 0.77647059]]])\n array([[[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69019608, 0.68235294, 0.67843137],\n        [0.69803922, 0.69019608, 0.68627451],\n        [0.70196078, 0.69411765, 0.69019608]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69019608, 0.69411765, 0.68627451],\n        [0.69019608, 0.69411765, 0.68627451],\n        [0.70196078, 0.69411765, 0.69019608]],\n\n       [[0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        [0.7254902 , 0.72941176, 0.71372549],\n        ...,\n        [0.69411765, 0.69803922, 0.69019608],\n        [0.69411765, 0.69803922, 0.69019608],\n        [0.69411765, 0.69803922, 0.69019608]],\n\n       ...,\n\n       [[0.40392157, 0.29411765, 0.64705882],\n        [0.40392157, 0.30588235, 0.65098039],\n        [0.40784314, 0.30980392, 0.64705882],\n        ...,\n        [0.06666667, 0.03529412, 0.27058824],\n        [0.09411765, 0.0627451 , 0.30196078],\n        [0.21568627, 0.18039216, 0.42745098]],\n\n       [[0.42745098, 0.32156863, 0.65882353],\n        [0.39607843, 0.29019608, 0.62745098],\n        [0.42745098, 0.3254902 , 0.65490196],\n        ...,\n        [0.10980392, 0.0745098 , 0.32156863],\n        [0.20392157, 0.16078431, 0.43529412],\n        [0.3254902 , 0.27843137, 0.56078431]],\n\n       [[0.39215686, 0.27843137, 0.61176471],\n        [0.40392157, 0.30196078, 0.63137255],\n        [0.39215686, 0.29019608, 0.61568627],\n        ...,\n        [0.18823529, 0.14901961, 0.40784314],\n        [0.29803922, 0.25098039, 0.53333333],\n        [0.31764706, 0.2627451 , 0.56862745]]])]\n<\/code><\/pre>\n\n<p>and epoch_y is batch of labels with 3 classes:<\/p>\n\n<pre><code>[list([0, 1, 0]) list([0, 0, 1]) list([0, 1, 0]) list([1, 0, 0])\n list([0, 1, 0]) list([0, 1, 0]) list([0, 0, 1]) list([0, 1, 0])\n list([0, 0, 1]) list([0, 1, 0])]\n<\/code><\/pre>\n\n<p>When feeding these batches of image and label I get <code>ValueError: setting an array element with a sequence.<\/code><\/p>\n","answers":[{"AnswerId":"52757629","CreationDate":"2018-10-11T10:18:02.730","ParentId":null,"OwnerUserId":"6288172","Title":null,"Body":"<p>Okay I solved it. The reason this happened was because I pre-processed my data in another file and then saved it to a pickle format. My labels should be [] and not list().<\/p>\n\n<p>I had loaded the data from pickle format to the model file and then converted it to a numpy array using <code>data = np.asarray(data)<\/code>. I don't know what it did to cause the problem, but removing this solved the problem. My <code>epoch_y<\/code> now comes in <code>[]<\/code> and not <code>list()<\/code>.<\/p>\n"}]}
+{"QuestionId":52737847,"AnswerCount":3,"Tags":"<keras><coremltools><relu>","CreationDate":"2018-10-10T10:13:59.077","AcceptedAnswerId":52740505.0,"OwnerUserId":10396961.0,"Title":"Core ML coremltools AttributeError: module 'keras.applications.mobilenet' has no attribute 'relu6'","Body":"<p>We are trying to convert a .h5 Keras model into a .mlmodel model, my code is as follows:<\/p>\n\n<pre><code>from keras.models import load_model\nimport keras\nfrom keras.applications import MobileNet\nfrom keras.layers import DepthwiseConv2D\n\nfrom keras.utils.generic_utils import CustomObjectScope\n\nwith CustomObjectScope({'relu6': keras.applications.mobilenet.relu6,'DepthwiseConv2D': keras.applications.mobilenet.DepthwiseConv2D}):\n    model = load_model('CNN_tourist_11.h5', custom_objects={'relu6': MobileNet})\n\noutput_labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\n\ncoreml_model = coremltools.converters.keras.convert(\"CNN_tourist_11.h5\",\n                                                    input_names=\"image\",\n                                                    image_input_names=\"image\",\n                                                    class_labels= output_labels,)\n\ncoremltools.utils.save_spec(coreml_model, 'place10.mlmodel')\n<\/code><\/pre>\n\n<p>We look up for the similiar question posed 6 days ago, and we also imported the MobileNet, but it still showing this error:<\/p>\n\n<pre><code>AttributeError: module 'keras.applications.mobilenet' has no attribute 'relu6'\n<\/code><\/pre>\n\n<p>My Tensorflow version is 1.10.0\nand Keras version is 2.2.2<\/p>\n\n<p>We will really appreciate if anyone can give us advice about why it keeps showing this error, thank you very much.<\/p>\n","answers":[{"AnswerId":"52740505","CreationDate":"2018-10-10T12:39:35.153","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>In latest Keras versions, MobileNet's components have been integrated with the rest of Keras' layers, so they are not available anymore as parts of the mobilenet package. Then you need to change the code to:<\/p>\n\n<pre><code>from keras.models import load_model\nimport keras\nfrom keras.applications import MobileNet\nfrom keras.layers import DepthwiseConv2D\n\nmodel = load_model('CNN_tourist_11.h5')\n\noutput_labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\n\ncoreml_model = coremltools.converters.keras.convert(\"CNN_tourist_11.h5\",\n                                                    input_names=\"image\",\n                                                    image_input_names=\"image\",\n                                                    class_labels= output_labels,)\n\ncoremltools.utils.save_spec(coreml_model, 'place10.mlmodel')\n<\/code><\/pre>\n"},{"AnswerId":"52740522","CreationDate":"2018-10-10T12:40:24.097","ParentId":null,"OwnerUserId":"10396961","Title":null,"Body":"<p>I amended my code:<\/p>\n\n<pre><code>from keras.models import load_model\nimport keras\nfrom keras.applications import MobileNet\nfrom keras.layers import DepthwiseConv2D\n\n#from keras.utils.generic_utils import CustomObjectScope\n\n#with CustomObjectScope({'relu6': keras.applications.mobilenet.relu6,'DepthwiseConv2D': keras.applications.mobilenet.DepthwiseConv2D}):\n    #model = load_model('CNN_tourist_11.h5', custom_objects={'relu6': MobileNet})\n\noutput_labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\n\ncoreml_model = coremltools.converters.keras.convert(\"CNN_tourist_11.h5\",\n                                                    input_names=\"image\",\n                                                    image_input_names=\"image\",\n                                                    class_labels= output_labels,)\n\ncoremltools.utils.save_spec(coreml_model, 'place10.mlmodel')\n<\/code><\/pre>\n\n<p>Some parts of the printout:<\/p>\n\n<pre><code>\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/coremltools\/converters\/keras\/_layers2.py in &lt;module&gt;()\n      8 if _keras.__version__ &gt;= _StrictVersion('2.2.0'):\n      9     from keras.layers import DepthwiseConv2D\n---&gt; 10     from keras_applications.mobilenet import relu6\n     11 else:\n     12     from keras.applications.mobilenet import DepthwiseConv2D, relu6\n\nImportError: cannot import name 'relu6'\n<\/code><\/pre>\n"},{"AnswerId":"52827477","CreationDate":"2018-10-16T03:16:38.693","ParentId":null,"OwnerUserId":"6929141","Title":null,"Body":"<p>Actually the question is answered by @Rex himself in the comments. I just want to make it clear and short as an \"Answer\" for people encountering the same issue.<\/p>\n\n<p>This problem is not about keras, it's about coremltools. You need to find a<code>_layers2.py<\/code> in coremltools files and comment out <code>from keras_application.mobilenet import relu6<\/code>:<\/p>\n\n<ul>\n<li><p>Find <code>_layer2.py<\/code> at <code>[YOUR_PYTHON_INSTALL_DIR]\/lib\/python3.6\/site-packages\/coremltools\/converters\/keras\/_layers2.py<\/code><\/p><\/li>\n<li><p>Comment out the line like this:<\/p>\n\n<pre><code>if _keras.__version__ &gt;= _StrictVersion('2.2.0'):\n    from keras.layers import DepthwiseConv2D\n# Modified by KF 10\/16\/2018\n#    from keras_applications.mobilenet import relu6\nelse:\n   from keras.applications.mobilenet import DepthwiseConv2D, relu6\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>Then, in your code, delete ALL import related to Keras, they are not relavant;<\/p>\n\n<p>For the <code>'Sequential' object has no attribute 'SerializeToString'<\/code> error, use <code>coreml_model.save<\/code> instead of <code>tools.utils.save_spec()<\/code>.<\/p>\n\n<pre><code># from keras.models import load_model\n# import keras\n# from keras.applications import MobileNet\n# from keras.layers import DepthwiseConv2D\n\noutput_labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\n\ncoreml_model = coremltools.converters.keras.convert(\"CNN_tourist_11.h5\",\n                                                input_names=\"image\",\n                                                image_input_names=\"image\",\n                                                class_labels= output_labels,)\n\n#coremltools.utils.save_spec(coreml_model, 'place10.mlmodel')\ncoreml_model.save('place10.mlmodel')\n<\/code><\/pre>\n\n<p>Problem solved.<\/p>\n"}]}
+{"QuestionId":52738146,"AnswerCount":1,"Tags":"<python><neural-network><pytorch><xor><backpropagation>","CreationDate":"2018-10-10T10:29:13.743","AcceptedAnswerId":52744932.0,"OwnerUserId":1019455.0,"Title":"XOR neural network does not learn","Body":"<p>I am trying to solve the very simple non-linear problem. It is <code>XOR<\/code> gate.\nI my school knowledge. <code>XOR<\/code> can be solve by using 2 input nodes, 2 hidden layer nodes. And 1 output. It is binary classification problem.<\/p>\n\n<p>I generate the <code>1000<\/code> of random integer number it is <code>0<\/code> or <code>1<\/code> and then do backpropagation. But for some unknown reason my network has not learned anything. The training accuracy is constant at <code>50<\/code>.<\/p>\n\n<pre><code># coding: utf-8\nimport matplotlib\nimport torch\nimport torch.nn as nn\nfrom torch.autograd import Variable\n\nmatplotlib.use('TkAgg')  # My buggy OSX 10.13.6 requires this\nimport matplotlib.pyplot as plt\nfrom torch.utils.data import Dataset\nfrom tqdm import tqdm\nimport random\n\nN = 1000\nbatch_size = 10\nepochs = 40\nhidden_size = 2\noutput_size = 1\nlr = 0.1\n\n\ndef return_xor(N):\n    tmp_x = []\n    tmp_y = []\n    for i in range(N):\n        a = (random.randint(0, 1) == 1)\n        b = (random.randint(0, 1) == 1)\n        if (a and not b) or (not a and b):\n            q = True\n        else:\n            q = False\n        input_features = (a, b)\n        output_class = q\n        tmp_x.append(input_features)\n        tmp_y.append(output_class)\n    return tmp_x, tmp_y\n\n\n# In[495]:\n\n\n# Training set\nx, y = return_xor(N)\nx = torch.tensor(x, dtype=torch.float, requires_grad=True)\ny = torch.tensor(y, dtype=torch.float, requires_grad=True)\n# Test dataset\nx_test, y_test = return_xor(100)\nx_test = torch.tensor(x_test)\ny_test = torch.tensor(y_test)\n\n\nclass MyDataset(Dataset):\n    \"\"\"Define my own `Dataset` in order to use `Variable` with `autograd`\"\"\"\n\n    def __init__(self, x, y):\n        self.x = x\n        self.y = y\n\n    def __getitem__(self, index):\n        return self.x[index], self.y[index]\n\n    def __len__(self):\n        return len(self.x)\n\n\ndataset = MyDataset(x, y)\ntest_dataset = MyDataset(x_test, y_test)\n\nprint(dataset.x.shape)\nprint(dataset.y.shape)\n\n# Make data iterable by loading to a loader. Shuffle, batch_size kwargs put them here in order to remind I myself\ntrain_loader = torch.utils.data.DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True)\ntest_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)\n\nprint(f\"They are {len(train_loader)} batches in the dataset\")\nshown = 0\nfor (x, y) in train_loader:\n    if shown == 1:\n        break\n    print(f\"{x.shape} {x.dtype}\")\n    print(f\"{y.shape} {y.dtype}\")\n    shown += 1\n\n\nclass MyModel(nn.Module):\n    \"\"\"\n    Binary classification\n    2 input nodes\n    2 hidden nodes\n    1 output node\n    \"\"\"\n\n    def __init__(self, input_size, hidden_size, output_size):\n        super().__init__()\n        self.fc1 = torch.nn.Linear(input_size, hidden_size)\n        self.fc2 = torch.nn.Linear(hidden_size, output_size)\n        self.sigmoid = torch.nn.Sigmoid()\n\n    def forward(self, out):\n        out = self.fc1(out)\n        out = self.fc2(out)\n        out = self.sigmoid(out)\n        return out\n\n\n# Create my network\nnet = MyModel(dataset.x.shape[1], hidden_size, output_size)\nCUDA = torch.cuda.is_available()\nif CUDA:\n    net = net.cuda()\ncriterion = torch.nn.BCELoss(reduction='elementwise_mean')\noptimizer = torch.optim.SGD(net.parameters(), lr=lr)\n\n# Train the network\ncorrect_train = 0\ntotal_train = 0\nfor epoch in range(epochs):\n    for i, (batches, labels) in enumerate(train_loader):\n        batcesh = Variable(batches.float())\n        labels = Variable(labels.float())\n        output = net(batches)  # Forward pass\n        optimizer.zero_grad()\n\n        loss = criterion(output, labels.view(10, 1))\n        loss.backward()\n        optimizer.step()\n        total_train += labels.size(0)\n        correct_train += (predicted == labels.long()).sum()\n        if (i + 1) % 10 == 0:\n            print(f\"\"\"\n                Epoch {epoch+1}\/{epochs}, \n                Iteration {i+1}\/{len(dataset)\/\/batch_size}, \n                Training Loss: {loss.item()},\n                Training Accuracy: {100*correct_train\/total_train}\n              \"\"\")\n<\/code><\/pre>\n\n<p><strong>Solution:<\/strong><br>\nI did initialized weight, Adaptive learning rate\n<a href=\"https:\/\/github.com\/elcolie\/nnbootcamp\/blob\/master\/Study-XOR.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/elcolie\/nnbootcamp\/blob\/master\/Study-XOR.ipynb<\/a><\/p>\n","answers":[{"AnswerId":"52744932","CreationDate":"2018-10-10T16:33:13.393","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>I am not sure what results you are getting, as the code you have posted in the question doesn't work (It gives errors with pytorch 0.4.1 like predicted not defined etc). But syntax issues apart, there are other problems. <\/p>\n\n<p>Your model is not actually two layer as it does not use non-linearity after the first output. Effectively this is one layer network and to fix that you can modify your model's <code>forward<\/code> as follows:<\/p>\n\n<pre><code>def forward(self, out):\n    out = torch.nn.functional.relu(self.fc1(out))\n    out = self.fc2(out)\n    out = self.sigmoid(out)\n    return out\n<\/code><\/pre>\n\n<p>You can try sigmoid or tanh non-linearity as well... but the non-linearity is a must. This should fix the problem. <\/p>\n\n<p>I also see that you are using only 2 hidden units. This might be restrictive and you might want to increase that to something like 5 or 10. <\/p>\n"}]}
+{"QuestionId":52738335,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-10T10:39:12.207","AcceptedAnswerId":null,"OwnerUserId":8640748.0,"Title":"TensorFlow: Correct way of using steps in Stochastic Gradient Descent","Body":"<p>I am currently using TensorFlow tutorial's <code>first_steps_with_tensor_flow.ipynb<\/code>  notebook to learn TF for implementing ML models. In the notebook, they have used Stochastic Gradient Descent (SGD) to optimize the loss function. Below is the snippet of the <code>my_input_function<\/code>:<\/p>\n\n<p><code>def my_input_fn(features, targets, batch_size=1, shuffle=True, num_epochs=None):<\/code><\/p>\n\n<p>Here, it can be seen that the batch_size is 1. The notebook uses a housing data set containing 17000 labeled examples for training. This means for SGD, I will be having 17000 batches.<\/p>\n\n<p><code>LRmodel = linear_regressor.train(input_fn = lambda:my_input_fn(my_feature, \n           targets), steps=100)<\/code><\/p>\n\n<p>I have three questions - <\/p>\n\n<ol>\n<li><p>Why is <code>steps=100<\/code> in <code>linear_regressor.train<\/code> method above? Since we have 17000 batches and <code>steps<\/code> in ML means the count for evaluating one batch, in <code>linear_regressor.train<\/code> method <code>steps = 17000<\/code> should be initialized, right?<\/p><\/li>\n<li><p>Is number of batches equal to the number of steps\/iterations in ML?<\/p><\/li>\n<li><p>With my 17000 examples, if I keep my <code>batch_size=100<\/code>, <code>steps=500<\/code>, and <code>num_epochs=5<\/code>, what does this initialization mean and how does it correlate to 170 batches?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"52904530","CreationDate":"2018-10-20T10:11:05.663","ParentId":null,"OwnerUserId":"7323976","Title":null,"Body":"<p>step is the literal meaning: means you refresh the parameters in your batch size; so for linear_regessor.train, it will train 100 times for this batch_size 1.<br>\nepoch means to refresh the whole data, which is 17,000 in your set.<br><\/p>\n"}]}
+{"QuestionId":52738570,"AnswerCount":2,"Tags":"<python><tensorflow><lstm><tensorboard><recurrent-neural-network>","CreationDate":"2018-10-10T10:52:29.400","AcceptedAnswerId":52738746.0,"OwnerUserId":7292899.0,"Title":"You must feed a value for placeholder tensor","Body":"<p>I am trying to implement an lstm nn using tesorboarb and I am receiving this error message: You must feed a value for placeholder tensor 'performance_1\/loss_summary'.<\/p>\n\n<p>I have already searched for an answer in many question without result.<\/p>\n\n<pre><code>with tf.name_scope('performance'):\nloss = tf.placeholder(tf.float32,shape=None,name='loss_summary') \n    tf_loss_summary = tf.summary.scalar('loss', loss)\n    tf_accuracy_ph = tf.placeholder(tf.float32,shape=None, name='accuracy_summary') \n    tf_accuracy_summary = tf.summary.scalar('accuracy', tf_accuracy_ph)\n\n# Gradient norm summary\nfor g in gradients:\nfor var in v:\n    if 'hidden3' in var.name and 'w' in var.name:\n        with tf.name_scope('Gradients'):\n            tf_last_grad_norm = tf.sqrt(tf.reduce_mean(g**2))\n            tf_gradnorm_summary = tf.summary.scalar('grad_norm', tf_last_grad_norm)\n            break   \n\n# Merge all summaries together\nperformance_summaries = tf.summary.merge([tf_loss_summary,tf_accuracy_summary])\n<\/code><\/pre>\n\n<p>The other part of the code in which I get the error is:<\/p>\n\n<pre><code>for ep in range(epochs):       \n\nfor step in range(train_seq_length\/\/batch_size):\n\n    u_data, u_labels = data_gen.unroll_batches()\n\n    feed_dict = {}\n    for ui,(dat,lbl) in enumerate(zip(u_data,u_labels)):            \n        feed_dict[train_inputs[ui]] = dat.reshape(-1,1)\n        feed_dict[train_outputs[ui]] = lbl.reshape(-1,1)\n\n    feed_dict.update({tf_learning_rate: 0.0001, tf_min_learning_rate:0.000001})\n\n    _, l = session.run([optimizer, loss], feed_dict=feed_dict)\n\n    average_loss += l\n\n\nif (ep+1) % valid_summary == 0:\n\n    average_loss = average_loss\/(valid_summary*(train_seq_length\/\/batch_size))\n\n  # The average loss\n    if (ep+1)%valid_summary==0:\n        print('Average loss at step %d: %f' % (ep+1, average_loss))\n\n    train_mse_ot.append(average_loss)\n\n    average_loss = 0 # reset loss\n\n    predictions_seq = []\n\n    mse_test_loss_seq = []\n<\/code><\/pre>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"52738696","CreationDate":"2018-10-10T10:58:41.257","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p><code>loss<\/code> is a placeholder, so you have to give a value to it. You probably didn't notice it and overrode your actual loss function. Usually summaries are not placeholders, so you have a misunderstanding of your variables and the flow in you code. <\/p>\n"},{"AnswerId":"52738746","CreationDate":"2018-10-10T11:01:41.867","ParentId":null,"OwnerUserId":"8140184","Title":null,"Body":"<p>Initiliaze <code>loss<\/code> as a variable. When you define something as a placeholder, you have to feed it's value when running the graph dependent on it.<\/p>\n"}]}
+{"QuestionId":52739028,"AnswerCount":0,"Tags":"<java><arrays><hadoop><tensorflow><tensorflow-serving>","CreationDate":"2018-10-10T11:17:58.400","AcceptedAnswerId":null,"OwnerUserId":1349516.0,"Title":"Java - Reading Tensorflow model (pb file) from HDFS","Body":"<p>I see that <code>SavedModelBundle.load<\/code> does not read from HDFS.\nI am doing the following to read file from HDFS as <code>Byte array<\/code> and import it as graph.<\/p>\n\n<pre><code>hdfs = FileSystem.get(new URI(configuration.get(\"fs.defaultFS\")), configuration);\n    if (hdfs.exists(location)) {\n        FSDataInputStream hdfsInputStream = hdfs.open(location);\n        int length = (int) hdfs.getFileStatus(location).getLen();\n        byte[] model = new byte[hdfsInputStream.available()];\n        hdfsInputStream.read(model);\n\n        hdfsInputStream.close();\n        final Graph g = new Graph();\n        g.importGraphDef(model);\n        Session session = new Session(g);\n}\n<\/code><\/pre>\n\n<p>I have also tried with <code>hdfsInputStream.readfully(model)<\/code>. It gives same following error:<\/p>\n\n<pre><code>java.lang.IllegalArgumentException: Invalid GraphDef\nat org.tensorflow.Graph.importGraphDef(Native Method)\nat org.tensorflow.Graph.importGraphDef(Graph.java:130)\nat org.tensorflow.Graph.importGraphDef(Graph.java:114)\n<\/code><\/pre>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>Is there a Java equivalent of <code>SavedModelBundle.load<\/code> for HDFS?<\/li>\n<li>What is the alternative? How can I read <code>.pb<\/code> file as <code>byte array<\/code> and import as tensorflow graph correctly?<\/li>\n<\/ol>\n\n<p>Note: All of the above works fine if I read model directory from local file system using <code>SavedModelBundle.load(path)<\/code> function. From local, I am not able to correctly read byte array and convert to graph. However, reading <code>bundle<\/code> works fine and give correct outputs. <\/p>\n","answers":[]}
+{"QuestionId":52740327,"AnswerCount":2,"Tags":"<c++><caffe><xtensor>","CreationDate":"2018-10-10T12:29:51.873","AcceptedAnswerId":null,"OwnerUserId":1446689.0,"Title":"how to fill xtensor array with data using a pointer","Body":"<p>I am trying to create a <code>xtensor<\/code> array from a blob data in <code>caffe<\/code> library. A pointer to data is returned using the function <code>mutable_cpu_data()<\/code> in <code>caffe<\/code> for example by <code>float* data = output-&gt;mutable_cpu_data();<\/code>. Is this possible with <code>xtensor<\/code>? If yes, could you please provide an example. I've found examples which use OpenCV <code>Mat<\/code> but <code>xtensor<\/code> is a lot like <code>numpy<\/code> which makes the operating on the matrix like data much easier.<\/p>\n","answers":[{"AnswerId":"52745886","CreationDate":"2018-10-10T17:36:10.890","ParentId":null,"OwnerUserId":"6866309","Title":null,"Body":"<p>You can do that, but you need the size of your data....<\/p>\n\n<p>Try as said below.<\/p>\n\n<pre><code>float* data = output-&gt;mutable_cpu_data();\n\n\/\/CONVERT YOUR DATA TO FLOAT VECTOR\n\/\/I assume the size of your array could be 4.\n\/\/Replace 4 with intended size of array.\nstd::vector&lt;float&gt; fData(data, data+4); \n\n\/\/INITIALIZE XARRAY\nxt::xarray&lt;float&gt; a(fData);\n<\/code><\/pre>\n"},{"AnswerId":"52849926","CreationDate":"2018-10-17T07:57:22.090","ParentId":null,"OwnerUserId":"7241714","Title":null,"Body":"<p>You can use the xt::adapt function from xadapt.hpp, but you need to provide a shape:<\/p>\n\n<pre><code>float* data = output-&gt;mutable_cpu_data();\nsize_t size = size_of_data;\n\/\/ For a 1D tensor for instance\nxt::static_shape&lt;std::size_t, 1&gt; sh = { size_of_data};\n\/\/ Parameters of adapt are:\n\/\/ - the 1D buffer to adapt\n\/\/ - the size of the buffer\n\/\/ - the ownership flag (should the adaptor destroy your buffer upon deletion, here\n\/\/   probably not)\n\/\/ - the shape\nauto a = xt::adapt(data, size_of_data, false sh);\n<\/code><\/pre>\n\n<p>The advantage compared to the solution provided by Naidu is that you don't copy the data buffer, it is adapted \"in place\".<\/p>\n"}]}
+{"QuestionId":52740436,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-10T12:36:03.250","AcceptedAnswerId":52743441.0,"OwnerUserId":10082348.0,"Title":"Optimizing this stacking of N weight-sharing Keras models","Body":"<p>I have two Keras (Tensorflow backend) models, which are stacked to make a combined model:<\/p>\n\n<ul>\n<li><code>small_model<\/code> with <code>In: (None,K), Out: (None,K)<\/code><\/li>\n<li><code>large_model<\/code> with <code>In: (None,N,K), Out: (None,1)<\/code><\/li>\n<li><code>combined_model<\/code> (<code>N x small_model -&gt; large_model<\/code>) with <code>In: (None,N,K), Out: (None,1)<\/code><\/li>\n<\/ul>\n\n<p><code>large_model<\/code> needs <code>N<\/code> stacked outputs from <code>small_model<\/code> as input. <\/p>\n\n<p>I can define <code>N<\/code> <code>small_model<\/code>s, which share weights, then concatenate their outputs (technically, I need to stack them), and then send that to <code>large_model<\/code>, as in the code below. <\/p>\n\n<p><strong>My problem<\/strong> is that I need to be able to do this for very large <code>N<\/code> (<code>&gt; 10**6<\/code>), and that my current solution uses a lot of memory and time when creating the models, even for <code>N ~ 10**2<\/code>.<\/p>\n\n<p>I'm hoping that there is a solution which sends the <code>N<\/code> data points through <code>small_model<\/code> in parallel (like what is done when giving a batch to a model), collects those points (with the Keras history, so that backprop is possible) and sends that to <code>large_model<\/code>, without having to define the <code>N<\/code> instances of <code>small_model<\/code>. The listed input and output shapes for the three models should not change, but other intermediate models can of course be defined. <\/p>\n\n<p>Thank you.<\/p>\n\n<p><strong>Current unsatisfactory solution<\/strong> (assume that <code>small_model<\/code> and <code>large_model<\/code> already exist, and that <code>N,K<\/code> are defined):<\/p>\n\n<pre><code>from keras.layers import Input, Lambda\nfrom keras.models import Model\nfrom keras import backend as K\n\ndef build_small_model_on_batch():\n\n    def distribute_inputs_to_small_model(input):\n        return [small_model(input[:,i]) for i in range(N)]\n\n    def stacker(list_of_tensors):\n        return K.stack(list_of_tensors, axis=1)\n\n    input = Input(shape=(N,K,))\n    small_model_outputs = Lambda(distribute_inputs_to_small_model)(input)\n    stacked_small_model_outputs = Lambda(stacker)(small_model_outputs)\n\n    return Model(input, stacked_small_model_outputs)\n\ndef build_combined():\n\n    input = Input(shape=(N,K,))\n    stacked_small_model_outputs = small_model_on_batch(input)\n    output = large_model(stacked_small_model_outputs)\n\n    return Model(input, output)\n\nsmall_model_on_batch = build_small_model_on_batch()\ncombined = build_combined()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52743441","CreationDate":"2018-10-10T15:12:01.467","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do that with a <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"nofollow noreferrer\"><code>TimeDistributed<\/code><\/a> layer wrapper:<\/p>\n\n<pre><code>from keras.layers import Input, Dense, TimeDistributed\nfrom keras.models import Sequential, Model\n\nN = None  # Use fixed value if you do not want variable input size\nK = 20\n\ndef small_model():\n    inputs = Input(shape=(K,))\n    # Define the small model\n    # Here it is just a single dense layer\n    outputs = Dense(K, activation='relu')(inputs)\n    return Model(inputs=inputs, outputs=outputs)\n\ndef large_model():\n    inputs = Input(shape=(N, K))\n    # Define the large model\n    # Just a single neuron here\n    outputs = Dense(1, activation='relu')(inputs)\n    return Model(inputs=inputs, outputs=outputs)\n\ndef combined_model():\n    inputs = Input(shape=(N, K))\n    # The TimeDistributed layer applies the given model\n    # to every input across dimension 1 (N)\n    small_model_out = TimeDistributed(small_model())(inputs)\n    # Apply large model\n    outputs = large_model()(small_model_out)\n    return Model(inputs=inputs, outputs=outputs)\n\nmodel = combined_model()\nmodel.compile(loss='mean_squared_error', optimizer='sgd')\nmodel.summary()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, None, 20)          0\n_________________________________________________________________\ntime_distributed_1 (TimeDist (None, None, 20)          420\n_________________________________________________________________\nmodel_2 (Model)              (None, None, 1)           21\n=================================================================\nTotal params: 441\nTrainable params: 441\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n"}]}
+{"QuestionId":52740744,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-10T12:52:25.047","AcceptedAnswerId":null,"OwnerUserId":6457578.0,"Title":"Set weights by name for a keras TimeDistributed layer","Body":"<p>I have a model in Keras as follows:<\/p>\n\n<pre><code>data = Input(shape=input_shape)\n\n# 512 x 640 x 3\npad1 = TimeDistributed(ZeroPadding2D(padding=(100, 100)))(data)\n\n# 712 x 840 x 3\nconv1_1 = TimeDistributed(Conv2D(8, (3,3), padding=\"valid\", activation=\"relu\", name=\"block1_conv1\", data_format=\"channels_last\"))(pad1)\nconv1_2 = TimeDistributed(Conv2D(8, (3,3), padding=\"same\", activation=\"relu\", name=\"block1_conv2\", data_format=\"channels_last\"))(conv1_1)\npool1   = TimeDistributed(MaxPooling2D((2,2), strides=(2,2), padding=\"same\", name=\"block1_pool\", data_format=\"channels_last\"))(conv1_2)\n<\/code><\/pre>\n\n<p>I want to be able to set the trainable weight parameters of conv1_1 and conv1_2 to a pre-trained value for every time step.  Can I do this?  Keras seems to treat these layers as their own entities with their own trainable parameters and not as a collection of Conv2D functions with the same shared trainable weights.  Is there a way to change this? How can I access the trainable weights for a single time slice and distribute that to all time slices?<\/p>\n","answers":[{"AnswerId":"52741409","CreationDate":"2018-10-10T13:30:41.090","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You can just do it like this:<\/p>\n\n<pre><code>data = Input(shape=input_shape)\n\n# 512 x 640 x 3\npad1 = TimeDistributed(ZeroPadding2D(padding=(100, 100)))(data)\n\n# 712 x 840 x 3\nnd_conv1_1 = Conv2D(8, (3,3), padding=\"valid\", activation=\"relu\", name=\"block1_conv1\", data_format=\"channels_last\")\nnd_conv1_2 = Conv2D(8, (3,3), padding=\"same\", activation=\"relu\", name=\"block1_conv2\", data_format=\"channels_last\")\nconv1_1 = TimeDistributed(nd_conv1_1)(pad1)\nconv1_2 = TimeDistributed(nd_conv1_2)(conv1_1)\npool1   = TimeDistributed(MaxPooling2D((2,2), strides=(2,2), padding=\"same\", name=\"block1_pool\", data_format=\"channels_last\"))(conv1_2)\n\nnd_conv1_1.trainable = True\/False\nnd_conv1_2.trainable = True\/False\n<\/code><\/pre>\n"}]}
+{"QuestionId":52741291,"AnswerCount":1,"Tags":"<keras><deep-learning><conv-neural-network><keras-layer>","CreationDate":"2018-10-10T13:24:11.620","AcceptedAnswerId":52741708.0,"OwnerUserId":7396176.0,"Title":"Creating a CNN Model in Keras with feature maps from each of the previous filtered images","Body":"<p>I am trying to implement the artificial convolutional neural network in order to perform a two-class pixel-wise classification as seen in the figure attached (from Chen et al. Nature 2017).\n<img src=\"https:\/\/i.stack.imgur.com\/fvXjE.jpg\" alt=\"\"><\/p>\n\n<p>Can you give me a hint on what the third and fourth layers should look like?<\/p>\n\n<p>This is how far I've got already:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D\n\nmodel = Sequential()\nmodel.add(Conv2D(40, (15, 15), activation='relu', \n          padding='same', input_shape = (64, 64, 1)))  # first layer\nmodel.add(MaxPooling2D((2, 2), padding='same'))  # second layer\n# model.add(...)  # third layer  &lt;-- how to implement this?\n# model.add(...)  # fourth layer &lt;-- how to implement this?\nprint(model.summary())\n<\/code><\/pre>\n\n<p>How many kernels did they use for the remaining layers and how should I interpret the summation symbols in the image?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"52741708","CreationDate":"2018-10-10T13:45:45.520","ParentId":null,"OwnerUserId":"3703716","Title":null,"Body":"<p><strike>The actual question is rather ambiguous.\nI am guessing correctly, that you want someone to implement the missing two lines of code for the network?<\/strike><\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(40, (15, 15), activation='relu', \n          padding='same', input_shape=(64, 64, 1)))\nmodel.add(MaxPooling2D((2, 2), padding='same'))\nmodel.add(Conv2D(40, (15, 15), activation='relu', padding='same'))  # layer 3\nmodel.add(Conv2D(1, (15, 15), activation='linear', padding='same'))  # layer 4\nprint(model.summary())\n<\/code><\/pre>\n\n<p>To get 40 feature maps after layer 3, we just convolve with 40 different kernels.\nAfter layer 4, there should be only one feature map \/ channel, so 1 kernel is enough here.<\/p>\n\n<p>By the way, the figure seems to be from <a href=\"https:\/\/www.nature.com\/articles\/nmeth.4405\" rel=\"nofollow noreferrer\">Convolutional neural networks for automated annotation of cellular cryo-electron tomograms<\/a> (<a href=\"https:\/\/arxiv.org\/pdf\/1701.05567.pdf\" rel=\"nofollow noreferrer\">PDF<\/a>) by Chen et al., a Nature article from 2017.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<blockquote>\n  <p>Comment: [...] why the authors say 1600 kernels in total and there is a summation?<\/p>\n<\/blockquote>\n\n<p>Actually, the authors seem to follow a rather strange notation here. They have an (imho) incorrect way to count kernels. What they rather mean is weights (if given 1x1 kernels...).<\/p>\n\n<p>Maybe they did not understand that the shape of the kernels are in fact 3-D, due to the last dimension equal to the number of feature maps.<\/p>\n\n<p>When we break it down there are<\/p>\n\n<ul>\n<li>40 kernels of size 15x15x1 for the 1st layer (which makes 40 * 15 ** 2 trainable weights)<\/li>\n<li>No kernels in the 2nd layer<\/li>\n<li>40 kernels of size 15x15x40 in the 3rd layer (which makes 1600 * 15 ** 2 trainable weights)<\/li>\n<li>1 kernel of size 15x15x40 for the 4th layer (which makes 40 * 15 ** 2 trainable weights)<\/li>\n<\/ul>\n"}]}
+{"QuestionId":52741872,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T13:53:48.333","AcceptedAnswerId":null,"OwnerUserId":8568787.0,"Title":"Tensorflow save\/load frozen tf.graph and run classification on loaded graph","Body":"<p>as I have mentioned in the topic I would like to save my tf.graph into a frozen_graph.pb file. This should save space later I will try to run it an a jetson tx2. I have made a short MNIST example describing my problem. I run tf 1.7 on python 3.5. <\/p>\n\n<p>Question1: As far as I understood my freeze_graph method takes a checkpoint file transfers all variables to constants except the ones i define with the second parameter. When I try to get the correct tensorname I wrote loggits.name but I get an error no Tensor with that name found in graph.<\/p>\n\n<p>Question2: After that I would be able to extract a frozen graph, how can i load it back and run an classification on that. <\/p>\n\n<p>My Code is attached and should work in a single py file. \nThank you very much in advance<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom tensorflow.examples.tutorials.mnist import input_data\nimport os\nimport time\n\nimport tensorflow as tf\nimport os\nimport argparse\n\n\n#METHODS I WANT TO TEST\n#TAKE THE CHECKPOINT FILE AND DELETE ALL NOTES THAT ARE NOT USEFUL\ndef freeze_graph(checkpoint_directory,output_node_names):\n    #checkpoint = tf.train.get_checkpoint_state(checkpoint_directory)\n    print(checkpoint_directory)\n    checkpoint = tf.train.get_checkpoint_state(checkpoint_directory)\n    input_checkpoint = checkpoint.model_checkpoint_path\n    absolute_model_dir = str(os.sep).join(input_checkpoint.split(os.sep)[:-1])\n    output_graph = absolute_model_dir + \"\/frozen_model.pb\"\n    clear_devices = True\n\n    with tf.Session(graph = tf.Graph()) as sess:\n        #import the metagraph in default graph\n        saver = tf.train.import_meta_graph(input_checkpoint + '.meta',clear_devices=clear_devices)\n\n        #restore the weights\n        saver.restore(sess,input_checkpoint)\n\n        #wrap variables to constants\n        [print(n.name) for n in tf.get_default_graph().as_graph_def().node]\n        output_graph_def = tf.graph_util.convert_variables_to_constants(sess, tf.get_default_graph().as_graph_def(),output_node_names.split(\",\"))\n\n        with tf.gfile.GFile(output_graph, \"wb\") as f:\n            f.write(output_graph_def.SerializeToString())\n        print(\"%d ops in the final graph.\" %len(output_graph_def.node))\n\n    return output_graph_def\n\n\n\n#HERE IS THE METHOD THAT ALLOWS ME TO LOAD MY FROZEN GRAPH AS GRAPH\ndef load_graph(frozen_graph_filename):\n    with tf.gfile.GFile(frozen_graph_filename,\"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    with tf.Graph().as_default() as graph:\n        tf.import_graph_def(graph_def, name = \"prefix\")\n    return graph\n\n\n#get the data\nmnist = input_data.read_data_sets(\"MNIST_data\/\",one_hot=True)\n\n#NETWORK PARAMETERS\nlearning_rate = 0.01\n\ndropout = 0.75\ndisplay_step = 1\nfilter_height = 5\nfilter_width = 5\ndepth_in = 1\ndepth_out1 = 64\ndepth_out2 = 128\n\n#PARAMETERS OF THE DATASET\ninput_height = 28\ninput_width = 28\nn_classes = 10\n\n#TRAINING PARAMETERS\nepochs = 1\nbatch_size = 256\nnum_batches = int(mnist.train.num_examples\/batch_size)\n\nx = tf.placeholder(tf.float32,[None,28*28],name = \"input\")\ny = tf.placeholder(tf.float32,[None,n_classes])\nkeep_prob = tf.placeholder(tf.float32)\n\nweights = {'wc1': tf.Variable(tf.random_normal([filter_height,filter_width,depth_in,depth_out1])),\n           'wc2': tf.Variable(tf.random_normal([filter_height, filter_width, depth_out1, depth_out2])),\n           'wd1': tf.Variable(tf.random_normal([int(input_height\/4)*int(input_height\/4)*depth_out2,1024])),\n           'out': tf.Variable(tf.random_normal([1024,n_classes]))}\n\nbiases = {'bc1': tf.Variable(tf.random_normal([depth_out1])),\n          'bc2': tf.Variable(tf.random_normal([depth_out2])),\n          'bd1': tf.Variable(tf.random_normal([1024])),\n          'out': tf.Variable(tf.random_normal([n_classes]))}\n\n\n#DEFINE YOUR NEURAL NETWORKS LAYER OPERATIONS\ndef ops_conv2d(x,W,b,strides = 1, add_bias = True, activation = tf.nn.relu, use_activation = True):\n\n    x = tf.nn.conv2d(x,W,strides = [1,strides,strides,1],padding = 'SAME')\n    x = tf.nn.bias_add(x,b)\n    if use_activation:\n        return activation(x)\n    else:\n        return x\n\ndef ops_maxpool2d(x,stride=2):\n    return tf.nn.max_pool(x,ksize=[1,stride,stride,1],strides = [1,stride,stride,1], padding = 'SAME' )\n\ndef ops_dropout(input_fully_connected,dropout):\n    return tf.nn.dropout(input_fully_connected,dropout)\n\ndef ops_fullyconnected(input, activation = tf.nn.relu, use_activation = True):\n    fc = tf.reshape(input,[-1,weights['wd1'].get_shape().as_list()[0]])\n    fc = tf.add(tf.matmul(fc,weights['wd1']),biases['bd1'])\n    if use_activation:\n        return activation(fc)\n    else:\n        return fc\n\n#DEFINE NETWORK ARCHTEKTURE (FORWARDPASS)\n\ndef build_network(x,weights,biases,dropout):\n    x = tf.reshape(x,shape=(-1,28,28,1))\n\n    conv_layer_1 = ops_conv2d(x,weights['wc1'],biases['bc1'],activation=tf.nn.relu, use_activation=True)\n    conv_layer_1 = ops_maxpool2d(conv_layer_1,2)\n\n    conv_layer_2 = ops_conv2d(conv_layer_1,weights['wc2'],biases['bc2'],activation=tf.nn.relu, use_activation=True)\n    conv_layer_2 = ops_maxpool2d(conv_layer_2,2)\n\n    fc1 = ops_fullyconnected(conv_layer_2, activation=tf.nn.relu, use_activation=True)\n    fc1 = ops_dropout(fc1,dropout)\n\n    logits = tf.add(tf.matmul(fc1,weights['out']),biases['out'],name = \"logits\")\n\n    return logits\n\n#DEFINE TENSORFLOW BACKPROPAGATION OBJECTS (BACKWARDPASS)\n\nlogits = build_network(x,weights,biases,keep_prob)\nloss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = logits,labels = y))\n\n#CHOSE AN OPTIMIZER\noptimizer = tf.train.AdadeltaOptimizer(learning_rate=learning_rate).minimize(loss=loss)\npredicted_labels = tf.equal(tf.argmax(logits,1),tf.argmax(y,1))\n\n#EVALUATION PARAMETERS\nacc = tf.reduce_mean(tf.cast(predicted_labels,tf.float32))\n\n#NOW INITIALIZE ALL TF VARIABLES\ninit = tf.global_variables_initializer()\n\nsaver = tf.train.Saver(max_to_keep=10)\n\n\n\n#NOW START THE SESSION AND EXECUTE THE GRAPH\nwith tf.Session() as sess:\n    sess.run(init)\n\n    for i in range(epochs):\n        save_path = saver.save(sess, os.curdir + \"checkpoints\/MNIST_TEST.ckpt\")\n        for j in range(num_batches):\n            batch_x, batch_y = mnist.train.next_batch(batch_size)\n            sess.run(optimizer, feed_dict={x:batch_x,y:batch_y,keep_prob:dropout})\n\n            losses,accs = sess.run([loss,acc],feed_dict={x:batch_x,y:batch_y,keep_prob:1.})\n\n            if epochs % display_step == 0:\n                print(\"EPOCH:\",'%04d' % (i+1),\n                      \"loss =\", \"{:.9f}\".format(losses),\n                      \"acc =\", \"{:.5f}\".format(accs))\n    print(\"TRAINING COMPLETED\")\n    #START PREDICTIONS\n    predicted_label = sess.run(logits,feed_dict={x:mnist.test.images[:256],keep_prob:1.})\n    test_classes = np.argmax(predicted_label,1)\n    print(\"TEST ACCURACY:\",sess.run(acc,feed_dict={x:mnist.test.images[:256], y:mnist.test.labels[:256],keep_prob:1.}))\n    f,a = plt.subplots(1,10,figsize = (10,2))\n\n    for i in range(10):\n        a[i].imshow(np.reshape(mnist.test.images[i],(28,28)))\n        print( test_classes[i])\n\n    print(\"TOTAL EXAMPLE FINNISHED\")\n\n    freeze_graph(os.curdir + \"checkpoints\" + os.sep, logits.name)\n\ngraph = load_graph(os.curdir + os.sep + \"checkpoints\" + os.sep + \"frozen_model.pb\")\nwith tf.Session(graph) as sess:\n    sess.run(init)\n    predicted_label = sess.run(logits, feed_dict={x: mnist.test.images[:256], keep_prob: 1.})\n    print(predicted_label)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52757193","CreationDate":"2018-10-11T09:54:35.393","ParentId":null,"OwnerUserId":"8568787","Title":null,"Body":"<p>if anybody has the same problem here is a description how i solved it.\nSaving and Loading the data:<\/p>\n\n<p>First of all note that I have now a different pipeline. First of all I save the session in a saver (ckpt files). Afterwards I construct a metagaph (graph.pb). This graph is then transfered into a frozen graph (frozen.pb). To load the frozen graph I use the load_frozen_graph_from_session method. Inside that method I also test a forward pass through my network. <\/p>\n\n<p>Running an inference on the loaded graph:\nFirst I name my tensors x (name = \"input\") this will result in a tensorname (\"input:0\")\nso when you try to fill this placeholder in the new session you need         <code>predicted_label = sess.run(\"output:0\", feed_dict={\"input:0\":mnist.test.images[:256], \"keep_prob:0\": 1.})<\/code><\/p>\n\n<p>The output is the logit and not the prediction inside my network. This is because if you run the session it will run until it hits the variable you want to fetch. would I take the prediction I need also the placeholder for my y (name=label). <\/p>\n\n<p>Here is the full code: <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom tensorflow.examples.tutorials.mnist import input_data\nimport os\nimport time\n\nimport tensorflow as tf\nimport os\nimport argparse\nfrom tensorflow.python.platform import gfile\nfrom tensorflow.python.framework.graph_util import convert_variables_to_constants\n\n\n\n#METHODS I WANT TO TEST\ndef freeze_graph_from_Session(sess,saver):\n    # convert_variables_to_constants(sess, input_graph_def, output_node_names, variable_names_whitelist=None)\n    save_graph(sess,saver)\n\n    with gfile.FastGFile(\".\/tmp\/\" + \"graph.pb\", 'rb') as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    frozen_graph_def = convert_variables_to_constants(sess, graph_def, [\"output\"])\n\n    with tf.gfile.GFile(\".\/tmp\/\" + \"frozen.pb\", \"wb\") as f:\n        f.write(frozen_graph_def.SerializeToString())\n\ndef save_graph(sess, saver):\n    saver.save(sess, \".\/tmp\/model\", write_meta_graph=True, global_step=1)\n\n    with open(\".\/tmp\/\" + \"graph.pb\", 'wb') as f:\n        f.write(sess.graph_def.SerializeToString())\n    #sess.close()\n\ndef load_frozen_graph_from_session():\n    filename = \".\/tmp\/\" + \"frozen.pb\"\n    print(\"LOADING GRAPH\")\n    with tf.gfile.GFile(filename, \"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    print(\"OPEN GRAPH\")\n    with tf.Graph().as_default() as graph:\n        print(\"DEFINE INPUT\")\n        new_input = tf.placeholder(tf.float32, [None, 28 * 28], name=\"new_input\")\n        print(\"DEFINE INPUT MAP\")\n        tf.import_graph_def(\n            graph_def,\n            # usually, during training you use queues, but at inference time use placeholders\n            # this turns into \"input\n            input_map={\"input:0\": new_input},\n            return_elements=None,\n            # if input_map is not None, needs a name\n            name=\"bla\",\n            op_dict=None,\n            producer_op_list=None\n        )\n\n    checkpoint_path = tf.train.latest_checkpoint(\".\/tmp\/\")\n\n    with tf.Session(graph=graph) as sess:\n        saver = tf.train.import_meta_graph(checkpoint_path + \".meta\", import_scope=None)\n        saver.restore(sess, checkpoint_path)\n        print(\"TRY FORWARD RUN THROUGH LOADED GRAPH\")\n\n        predicted_label = sess.run(\"output:0\", feed_dict={\"input:0\":mnist.test.images[:256], \"keep_prob:0\": 1.})\n        print(\"output\", predicted_label)\n        f, a = plt.subplots(1, 10, figsize=(10, 2))\n        test_classes = np.argmax(predicted_label, 1)\n        for i in range(10):\n            a[i].imshow(np.reshape(mnist.test.images[i], (28, 28)))\n            print(test_classes[i])\n        print (\"output:\", test_classes)\n\n\n#TAKE THE CHECKPOINT FILE AND DELETE ALL NOTES THAT ARE NOT USEFUL\ndef freeze_graph(checkpoint_directory,output_node_names):\n    #checkpoint = tf.train.get_checkpoint_state(checkpoint_directory)\n    print(checkpoint_directory)\n    checkpoint = tf.train.get_checkpoint_state(checkpoint_directory)\n    input_checkpoint = checkpoint.model_checkpoint_path\n    absolute_model_dir = str(os.sep).join(input_checkpoint.split(os.sep)[:-1])\n    output_graph = absolute_model_dir + \"\/frozen_model.pb\"\n    clear_devices = True\n\n    with tf.Session(graph = tf.Graph()) as sess:\n        #import the metagraph in default graph\n        saver = tf.train.import_meta_graph(input_checkpoint + '.meta',clear_devices=clear_devices)\n\n        #restore the weights\n        saver.restore(sess,input_checkpoint)\n\n        #wrap variables to constants\n        [print(n.name) for n in tf.get_default_graph().as_graph_def().node]\n        output_graph_def = tf.graph_util.convert_variables_to_constants(sess, tf.get_default_graph().as_graph_def(),output_node_names.split(\",\"))\n\n        with tf.gfile.GFile(output_graph, \"wb\") as f:\n            f.write(output_graph_def.SerializeToString())\n        print(\"%d ops in the final graph.\" %len(output_graph_def.node))\n\n    return output_graph_def\n\n#HERE IS THE METHOD THAT ALLOWS ME TO LOAD MY FROZEN GRAPH AS GRAPH\ndef load_graph(frozen_graph_filename):\n    with tf.gfile.GFile(frozen_graph_filename,\"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    with tf.Graph().as_default() as graph:\n        tf.import_graph_def(graph_def, name = \"prefix\")\n    return graph\n\n\n#get the data\nmnist = input_data.read_data_sets(\"MNIST_data\/\",one_hot=True)\nprint(mnist.test.labels[:256])\n\nprint(\"load_freeze_graph_from_session: STARTED\")\nload_frozen_graph_from_session()\nprint(\"load_freeze_graph_from_session: ENDED\")\nexit()\n\n#NETWORK PARAMETERS\nlearning_rate = 0.01\n\ndropout = 0.75\ndisplay_step = 1\nfilter_height = 5\nfilter_width = 5\ndepth_in = 1\ndepth_out1 = 64\ndepth_out2 = 128\n\n#PARAMETERS OF THE DATASET\ninput_height = 28\ninput_width = 28\nn_classes = 10\n\n#TRAINING PARAMETERS\nepochs = 1\nbatch_size = 256\nnum_batches = int(mnist.train.num_examples\/batch_size)\n\nx = tf.placeholder(tf.float32,[None,28*28],name=\"input\")\n\ny = tf.placeholder(tf.float32,[None,n_classes],name = \"label\")\nkeep_prob = tf.placeholder(tf.float32,name = \"keep_prob\")\n\nweights = {'wc1': tf.Variable(tf.random_normal([filter_height,filter_width,depth_in,depth_out1])),\n           'wc2': tf.Variable(tf.random_normal([filter_height, filter_width, depth_out1, depth_out2])),\n           'wd1': tf.Variable(tf.random_normal([int(input_height\/4)*int(input_height\/4)*depth_out2,1024])),\n           'out': tf.Variable(tf.random_normal([1024,n_classes]))}\n\nbiases = {'bc1': tf.Variable(tf.random_normal([depth_out1])),\n          'bc2': tf.Variable(tf.random_normal([depth_out2])),\n          'bd1': tf.Variable(tf.random_normal([1024])),\n          'out': tf.Variable(tf.random_normal([n_classes]))}\n\n\n#DEFINE YOUR NEURAL NETWORKS LAYER OPERATIONS\ndef ops_conv2d(x,W,b,strides = 1, add_bias = True, activation = tf.nn.relu, use_activation = True):\n\n    x = tf.nn.conv2d(x,W,strides = [1,strides,strides,1],padding = 'SAME')\n    x = tf.nn.bias_add(x,b)\n    if use_activation:\n        return activation(x)\n    else:\n        return x\n\ndef ops_maxpool2d(x,stride=2):\n    return tf.nn.max_pool(x,ksize=[1,stride,stride,1],strides = [1,stride,stride,1], padding = 'SAME' )\n\ndef ops_dropout(input_fully_connected,dropout):\n    return tf.nn.dropout(input_fully_connected,dropout)\n\ndef ops_fullyconnected(input, activation = tf.nn.relu, use_activation = True):\n    fc = tf.reshape(input,[-1,weights['wd1'].get_shape().as_list()[0]])\n    fc = tf.add(tf.matmul(fc,weights['wd1']),biases['bd1'])\n    if use_activation:\n        return activation(fc)\n    else:\n        return fc\n\n#DEFINE NETWORK ARCHTEKTURE (FORWARDPASS)\n\ndef build_network(x,weights,biases,dropout):\n    x = tf.reshape(x,shape=(-1,28,28,1))\n\n    conv_layer_1 = ops_conv2d(x,weights['wc1'],biases['bc1'],activation=tf.nn.relu, use_activation=True)\n    conv_layer_1 = ops_maxpool2d(conv_layer_1,2)\n\n    conv_layer_2 = ops_conv2d(conv_layer_1,weights['wc2'],biases['bc2'],activation=tf.nn.relu, use_activation=True)\n    conv_layer_2 = ops_maxpool2d(conv_layer_2,2)\n\n    fc1 = ops_fullyconnected(conv_layer_2, activation=tf.nn.relu, use_activation=True)\n    fc1 = ops_dropout(fc1,dropout)\n\n    logits = tf.add(tf.matmul(fc1,weights['out']),biases['out'],name = \"output\")\n\n    return logits\n\n#DEFINE TENSORFLOW BACKPROPAGATION OBJECTS (BACKWARDPASS)\n\nlogits = build_network(x,weights,biases,keep_prob)\n#freeze_graph(os.curdir + \"checkpoints\" + os.sep, logits.name)\nloss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = logits,labels = y))\n\n#CHOSE AN OPTIMIZER\noptimizer = tf.train.AdadeltaOptimizer(learning_rate=learning_rate).minimize(loss=loss)\npredicted_labels = tf.equal(tf.argmax(logits,1),tf.argmax(y,1))\n\n#EVALUATION PARAMETERS\nacc = tf.reduce_mean(tf.cast(predicted_labels,tf.float32))\n\n#NOW INITIALIZE ALL TF VARIABLES\ninit = tf.global_variables_initializer()\n\nsaver = tf.train.Saver(max_to_keep=10)\n\n\n\n#NOW START THE SESSION AND EXECUTE THE GRAPH\nwith tf.Session() as sess:\n    sess.run(init)\n\n    for i in range(epochs):\n        save_path = saver.save(sess, os.curdir + \"checkpoints\/MNIST_TEST.ckpt\")\n        for j in range(num_batches):\n            batch_x, batch_y = mnist.train.next_batch(batch_size)\n            sess.run(optimizer, feed_dict={x:batch_x,y:batch_y,keep_prob:dropout})\n\n            losses,accs = sess.run([loss,acc],feed_dict={x:batch_x,y:batch_y,keep_prob:1.})\n\n            if epochs % display_step == 0:\n                print(\"EPOCH:\",'%04d' % (i+1),\n                      \"loss =\", \"{:.9f}\".format(losses),\n                      \"acc =\", \"{:.5f}\".format(accs))\n    print(\"TRAINING COMPLETED\")\n    #START PREDICTIONS\n    predicted_label = sess.run(logits,feed_dict={x:mnist.test.images[:256],keep_prob:1.})\n    test_classes = np.argmax(predicted_label,1)\n    print(\"TEST ACCURACY:\",sess.run(acc,feed_dict={x:mnist.test.images[:256], y:mnist.test.labels[:256],keep_prob:1.}))\n    f,a = plt.subplots(1,10,figsize = (10,2))\n\n    for i in range(10):\n        a[i].imshow(np.reshape(mnist.test.images[i],(28,28)))\n        print( test_classes[i])\n\n    print(\"TOTAL EXAMPLE FINNISHED\")\n\n    #freeze_graph(os.curdir + \"checkpoints\"+os.sep,logits)\n    print(\"freeze_graph_from_session: STARTED\")\n    freeze_graph_from_Session(sess,saver)\n    print(\"freeze_graph_from_session: ENDED\")\n\nprint(\"load_freeze_graph_from_session: STARTED\")\nload_frozen_graph_from_session()\nprint(\"load_freeze_graph_from_session: ENDED\")\n\n#with tf.Session() as sess:\n#\n#    sess.run(init)\n#    graph = load_graph(os.curdir + os.sep + \"checkpoints\" + os.sep + \"frozen_model.pb\")\n#    predicted_label = sess.run(logits, feed_dict={x: mnist.test.images[:256], keep_prob: 1.})\n#    print(predicted_label)\n<\/code><\/pre>\n\n<p>Thanks goes out to my self. :)<\/p>\n"}]}
+{"QuestionId":52742019,"AnswerCount":0,"Tags":"<tensorflow><object-detection>","CreationDate":"2018-10-10T14:01:11.043","AcceptedAnswerId":null,"OwnerUserId":5380914.0,"Title":"Tensorflow object detection API: weird result","Body":"<p>I've use pre-trained ssd mobilenet v2 model and fine tune on my Drone dataset. I followed their tutorial. I have 300 images for train and 90 images for test. After 190k global step, I test my model but result is so weird. It can detect the drone but it will always detect something at the bottom of every images.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/knyhy.png\" rel=\"nofollow noreferrer\">image1<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Bc3lm.png\" rel=\"nofollow noreferrer\">image2<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fplfB.png\" rel=\"nofollow noreferrer\">image3<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mjpJP.png\" rel=\"nofollow noreferrer\">image4<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UUzgI.png\" rel=\"nofollow noreferrer\">image5<\/a><\/p>\n\n<p>Why is it so?\nThank you<\/p>\n","answers":[]}
+{"QuestionId":52742111,"AnswerCount":2,"Tags":"<c++><tensorflow><cmake>","CreationDate":"2018-10-10T14:06:07.130","AcceptedAnswerId":null,"OwnerUserId":9634525.0,"Title":"Error building tensorflow C++ shared library on windows","Body":"<p>I am trying to build tensorflow as a standalone project and have been following this tutorial<\/p>\n\n<p><a href=\"http:\/\/www.stefanseibert.com\/2017\/10\/tensorflow-as-dll-into-your-windows-c-project-with-gpu-support-and-cmake-v1-3\/\" rel=\"nofollow noreferrer\">http:\/\/www.stefanseibert.com\/2017\/10\/tensorflow-as-dll-into-your-windows-c-project-with-gpu-support-and-cmake-v1-3\/<\/a><\/p>\n\n<p>but alternatively with cpu support<\/p>\n\n<p>My environment setup versions<\/p>\n\n<p><strong>protobuf<\/strong>            3.6.1<br>\n<strong>tensorflow<\/strong>          1.10.0<br>\n<strong>tf.GIT_VERSION<\/strong> = b'v1.10.0-rc1-19-g656e7a2b34'<\/p>\n\n<p>Here are the steps I used to generate the shared lib<\/p>\n\n<ol>\n<li><p>Acquired source code from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow.git\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow.git<\/a><\/p><\/li>\n<li><p>Have installed the dependencies since I do not use the python bindings, there is no need for SWIG, so I installed Git (version 2.15.1.windows.2) and cmake 3.11.1<\/p><\/li>\n<li><p>I used the 64bit tools from Visual Studio 2015 since VS2015 is necessary to build the DLL. I should be able to open the \u201cVS2015 x64 Native Tools Command Prompt\u201d. This is needed so VS uses the 64 bit toolset.<\/p><\/li>\n<li><p>Navigated in the commandline to the \u201ctensorflow\/contrib\/cmake\u201d subfolder of the source code and create a directory with \u201cmkdir build\u201d. Afterwards navigate to the fresh build folder with \u201ccd build\u201d.<\/p><\/li>\n<li><p>Create a build solution: cmake .. -A x64 -DCMAKE_BUILD_TYPE=RelWithDebInfo   -Dtensorflow_BUILD_CC_EXAMPLE=OFF -Dtensorflow_ENABLE_GRPC_SUPPORT=OFF -Dtensorflow_BUILD_CC_TESTS=OFF -Dtensorflow_BUILD_PYTHON_TESTS=OFF -Dtensorflow_ENABLE_GPU=OFF  -Dtensorflow_WIN_CPU_SIMD_OPTIONS=\/arch:AVX -Dtensorflow_BUILD_SHARED_LIB=ON<\/p><\/li>\n<li><p>Everything went fine till this. To build the tensorflow.dll, I issued the following command: MSBuild \/p:Configuration=RelWithDebInfo tensorflow.vcxproj<\/p><\/li>\n<li><p>This throws an error: <strong>D:\\work\\tensorflow\\tensorflow\/core\/lib\/core\/stringpiece.h(34): fatal error C1083: Cannot open include file: 'absl\/strings\/string_view.h': No such file or directory (\ncompiling source file D:\\work\\tensorflow\\tensorflow\\core\\lib\\core\\coding.cc) [D:\\work\\tensorflow\\tensorflow\\contrib\\cmake\\build\\tf_core_lib.vcxproj]<\/strong>.<\/p><\/li>\n<li><p>I fixed the above error with this: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22007#issuecomment-424553600\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22007#issuecomment-424553600<\/a>.<\/p><\/li>\n<li><p>Doing the above I ended up with this error: <strong>path.obj : error LNK2019: unresolved external symbol \"void __cdecl absl::base_internal::ThrowStdOutOfRange(char const *)\" (?ThrowStdOutOfRange@base_internal@absl@@YA\nXPEBD@Z) referenced in function \"class std::basic_string,class std::allocator > __cdecl tensorflow::io::internal::JoinPathIm<\/strong><\/p><\/li>\n<\/ol>\n\n<p>I am not able to proceed further. Any workaround for this? Thanks!<\/p>\n","answers":[{"AnswerId":"55681340","CreationDate":"2019-04-15T00:44:55.857","ParentId":null,"OwnerUserId":"11360825","Title":null,"Body":"<p>I met same issue, I think tensorflow new version doesn't support CMake, but we can solve the issues.\n1. Seems the absl version in project folder is out dated, so I cloned the latest version of abseil-cpp from: <a href=\"https:\/\/github.com\/abseil\/abseil-cpp\" rel=\"nofollow noreferrer\">https:\/\/github.com\/abseil\/abseil-cpp<\/a>\n2. Use cmake to build the abseil-cpp, it will be fast.\n3. Add lib path to tensorflow dependency, the needed one will be D:\\git\\abseil-cpp\\abseil-cpp\\build\\absl\\base\\Release\\absl_absl_throw_delegate.lib\n4. If you meet other linking error, you can find the function name in absl sources and find the library contain it.\nHope this can help you and people who may met this issue in future.<\/p>\n"},{"AnswerId":"54672363","CreationDate":"2019-02-13T14:18:33.353","ParentId":null,"OwnerUserId":"7959545","Title":null,"Body":"<p>lnk2019 error occurs when your directly you are using in your source code are not linked properly. Please add additional dependencies to your project.<\/p>\n\n<ul>\n<li>Going to project properties<\/li>\n<li>select C\/C++ option<\/li>\n<li>Add aditional dependencies<\/li>\n<li>Go to Linker Option Below C\/C++<\/li>\n<li>Add additional Dependencies here.<\/li>\n<\/ul>\n\n<p>It might be help full for you from getting out to LNK2019 problem<\/p>\n\n<p><a href=\"https:\/\/docs.microsoft.com\/en-us\/previous-versions\/799kze2z(v=vs.140)\" rel=\"nofollow noreferrer\">view this to understand LNK2019 error.<\/a><\/p>\n"}]}
+{"QuestionId":52742162,"AnswerCount":1,"Tags":"<python><merge><keras><keras-layer>","CreationDate":"2018-10-10T14:08:37.230","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"More input and one output issue in Keras","Body":"<pre><code>self.embed = Sequential([Embedding(9488, output_dim=512,input_length=14),\n                                Activation('relu'),\n                                    Dropout(0.5)], name='embed.0')\n\nself.fc_embed = Sequential([Dense(512, input_shape=(10,2048)),\n                                    Activation('relu'),\n                                    Dropout(0.5)], name='fc_embed.0')\n\ninputs_bedding = Input(shape=(10,))\nxt = self.embed(inputs_bedding)\n\ninput_feats = Input(shape=(10,2048))\nfc_feats = self.fc_embed(input_feats)\n\nfc_feats_new = K.reshape(fc_feats, [fc_feats.shape[1], fc_feats.shape[2]])\nxt_new = K.reshape(xt, [xt.shape[1], xt.shape[2]])\n\n prev_h = state[0][-1] (shape is (10,512))\n att_lstm_input = Concatenate([prev_h, fc_feats_new, xt_new], axis=1)\n lstm, h_att, c_att = LSTM(units=512, name='core.att_lstm', return_state=True)(att_lstm_input)\n model = Model([input_feats, inputs_att, inputs_bedding], lstm)\n model.summary()\n<\/code><\/pre>\n\n<p>This is the error I get:<\/p>\n\n<pre><code>File \"copy_eval.py\", line 165, in &lt;module&gt;\nmodel1 = TopDownModel.forward(fc_feats, att_feats, seq, att_masks)\n<\/code><\/pre>\n\n<p>File \"\/home\/ubuntu\/misc\/customize_keras.py\", line 127, in forward\n    lstm, h_att, c_att = LSTM(units=512, name='core.att_lstm', return_state=True)(att_lstm_input)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/layers\/recurrent.py\", line 500, in <strong>call<\/strong>\n    return super(RNN, self).<strong>call<\/strong>(inputs, **kwargs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/topology.py\", line 575, in <strong>call<\/strong>\n    self.assert_input_compatibility(inputs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/topology.py\", line 448, in assert_input_compatibility\n    str(inputs) + '. All inputs to the layer '\nValueError: Layer core.att_lstm was called with an input that isn't a symbolic tensor. Received type: . Full input: []. All inputs to the layer should be tensors.<\/p>\n\n<p>For more input, how to merge them into one output?<\/p>\n","answers":[{"AnswerId":"52800157","CreationDate":"2018-10-14T06:41:36.657","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Concatenate should be used as a layer, like this:<\/p>\n\n<pre><code>att_lstm_input = Concatenate(axis=1)([prev_h, fc_feats_new, xt_new])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52742238,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><conv-neural-network><pytorch><rnn>","CreationDate":"2018-10-10T14:12:44.670","AcceptedAnswerId":null,"OwnerUserId":2945859.0,"Title":"How to pretrain\/select CNN for Biomedical Video Analysis","Body":"<p><strong>Data<\/strong>\nI am trying to train a model in the biomedical domain with a rather specialized task (flow prediction).\nMy input consists of video clips and I would like to predict either a single image or a video. I have pixel wise ground truth for every frame of the video but the number of videos is very limited.\nI do however have a lot of unlabeled data of related scenes, which is why I considered transfer learning\/pretraining of some sort.<\/p>\n\n<p><strong>Architecture<\/strong>\nArchitecture wise I am considering a CNN, RNN combination where the CNN provides a representation of the input frames for the RNN to learn about the temporal relationship between input frames.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Vvwwg.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Vvwwg.jpg\" alt=\"Network Architecture\"><\/a>\nNow my question is: What kind of CNN do I use and on what do I pretrain it? Since I am working with biomedical data I would assume image net as well as most other image data sets do not really help as the image content is very different. Are there any data sets\/tasks\/networks I could use for this purpose?<\/p>\n","answers":[{"AnswerId":"52779729","CreationDate":"2018-10-12T12:38:31.060","ParentId":null,"OwnerUserId":"10494960","Title":null,"Body":"<p>Some people recommend shallow autoencoders to pretrain networks if there is not much labeled data, but lots of unlabeled data. A reference would be G\u00e9ron's book \"Hands-On Machine Learning with Scikit-Learn and TensorFlow\", but there are also many articles and tutorials online.<\/p>\n"}]}
+{"QuestionId":52742283,"AnswerCount":1,"Tags":"<python><scikit-learn><keras><grid-search>","CreationDate":"2018-10-10T14:14:58.097","AcceptedAnswerId":null,"OwnerUserId":10478172.0,"Title":"fit: passing multiple inputs with sklearn.model_selection.GridSearchCV","Body":"<p>as explained in <a href=\"https:\/\/keras.io\/models\/model\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/model<\/a>, Keras' <code>fit<\/code> function <code>x<\/code> can be<\/p>\n\n<blockquote>\n  <p>Numpy array of training data (if the model has a single input), or\n  list of Numpy arrays (if the model has multiple inputs). If input\n  layers in the model are named, you can also pass a dictionary mapping\n  input names to Numpy arrays.<\/p>\n<\/blockquote>\n\n<p>but it doesn't seem to be possible with sklearn.model_selection.GridSearchCV <code>fit<\/code>. Is that true? Is there any workaround?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52742732","CreationDate":"2018-10-10T14:37:25.247","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>If you want to just use grid search as in sci-kit learn, you can use the KerasClassifier wrapper. You just create your model in your function.<\/p>\n\n<pre><code>def create_model(dropout_rate):\n    #...\nmodel = KerasClassifier(build_fn=create_model)\n\nparam_grid = dict(dropout_rate=[0.2,0.3,0.5])\ngrid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1)\ngrid_result = grid.fit(X, Y)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/machinelearningmastery.com\/grid-search-hyperparameters-deep-learning-models-python-keras\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/grid-search-hyperparameters-deep-learning-models-python-keras\/<\/a><\/p>\n\n<p>Unfortunately, this approach is kind of limited, as you don\u00b4t have a validation after each epoch. To solve this, you need to implement your own KerasClassifier.<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4278\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4278<\/a><\/p>\n"}]}
+{"QuestionId":52742509,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><computer-vision><tensorflow-lite>","CreationDate":"2018-10-10T14:26:14.090","AcceptedAnswerId":null,"OwnerUserId":8834241.0,"Title":"Can I convert all the tensorflow slim models to tflite?","Body":"<p>I'm training tensorflow slim based models for image classification on a custom dataset. Before I invest a lot of time training such huge a dataset, I wanted to know whether or not can I convert all the models available in the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\" rel=\"nofollow noreferrer\">slim model zoo<\/a> to tflite format.<\/p>\n\n<p>Also, I know that I can convert my custom slim-model to a frozen graph. It is the step after this which I'm worried about i.e, conversion to .tflite from my custom trained .pb model.<\/p>\n\n<p>Is this supported ? or is there anyone who is facing conversion problems that has not yet been resolved ? <\/p>\n\n<p>Thanks.     <\/p>\n","answers":[{"AnswerId":"54584044","CreationDate":"2019-02-07T23:42:49.673","ParentId":null,"OwnerUserId":"11031154","Title":null,"Body":"<p>Many Slim models can be converted to TFLite, but it isn't a guarantee since some models might have ops not supported by TFLite.<\/p>\n\n<p>What you could do, is try and convert your model to TensorFlow Lite using <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/python_api\" rel=\"nofollow noreferrer\">TFLiteConverter<\/a> in Python before training. If the conversion succeeds, then you can train your TF model and convert it once again.<\/p>\n"}]}
+{"QuestionId":52742554,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-10T14:28:25.080","AcceptedAnswerId":52758146.0,"OwnerUserId":10484743.0,"Title":"TensorFlow Object Detection API (Different interest)","Body":"<p>I have one question.\nI trained 2 times one and the same model. For training I used a ready-made model with 90 classes. And I get different percentages of recognition. Why? As far as I understand, if you train your model on an already finished one, then the initial weights will always be the same and therefore the results should always be the same.<\/p>\n","answers":[{"AnswerId":"52758146","CreationDate":"2018-10-11T10:46:37.063","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>If I understand correctly, you took some pre-trained model, and fine-tuned it two different times, but with the same configuration.\nIf so, not only the two resulting models do no have to be the same, moreover - they most probably won't be.<\/p>\n\n<p>The reason for that is that the training procedure has some random options in it, e.g. data shuffling, augmentation, dropout, etc.\nFor example, a common augmentation option is to randomly flip an image horizontally. This means that in one training session an image can be taken as is, while in the other - its mirror image.<\/p>\n\n<p>While this is only one example, quite a few random options are made with different probabilities, and therefore you can't achieve the <strong>exact same<\/strong> model unless you document each and every choice your training session makes.<\/p>\n"}]}
+{"QuestionId":52742586,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T14:30:11.630","AcceptedAnswerId":null,"OwnerUserId":5863503.0,"Title":"tf.layers.dense underperforms low-level api?","Body":"<p>I've built a simple NN using TF that supposes to learn a quadratic equation. Surprisingly, <code>tf.layers.dense<\/code> underperforms actually writing the NN myself, but I can't seem to understand why. I would very much appreciate it someone can point to what it is I'm missing - here's the code:<\/p>\n\n<p><strong>Data generation<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\nfrom sklearn.utils import shuffle\n\ndef f(x):\n    return x**2-6*x+9 \n\ndef data_generator_const_sigma(x,sigma,samples):\n    return np.random.normal(f(x),sigma,samples)\n\nx_vals = np.arange(1,5.2,0.2)\nx_arr = np.array([])\ny_arr = np.array([])\nsamples = 50\nfor x in x_vals:\n    x_arr = np.append(x_arr, np.full(samples,x))\n    y_arr = np.append(y_arr, data_generator_rel_sigma(x,0.1,samples))\nx_arr, y_arr = shuffle(x_arr, y_arr)\nx_test = np.arange(1.1,5.1,0.2)\n\nplt.figure(figsize=(7,7))\nplt.grid(True)\nplt.xlabel('x')\nplt.ylabel('f(x)')\nplt.scatter(x_arr,y_arr)\nplt.plot(x_vals,map(f,x_vals),c='m')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NiIpR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NiIpR.png\" alt=\"data\"><\/a><\/p>\n\n<p><strong>Low level API:<\/strong><\/p>\n\n<pre><code>def nn_layer(prev_layer, weights, bias, activation=None):\n    layer = tf.add(tf.matmul(prev_layer,weights),bias)\n    if activation == 'tanh':\n        layer = tf.nn.tanh(layer)\n    elif activation == 'relu':\n        layer = tf.nn.relu(layer)\n    return layer\n\nhidden_layers_size = [12,12]\nepochs=1000\nbatch_size=50\nlearning_rate=0.0003\ndisplay_step=50\n\nw = dict()\nb = dict()\ntf.reset_default_graph()\nL = len(hidden_layers_size) + 2\n_ls = [1]+hidden_layers_size+[1]\nfor i in range(1,L):\n    w[i] = tf.get_variable(name='w'+str(i), shape=[_ls[i-1],_ls[i]],\n                           initializer=tf.contrib.layers.xavier_initializer())\n    b[i] = tf.get_variable(name='b'+str(i), shape=_ls[i],\n                           initializer=tf.zeros_initializer())\nx = tf.placeholder(name='x',shape=(None,1),dtype=tf.float32)\ny = tf.placeholder(name='y',shape=(None,1),dtype=tf.float32)\n\nlayer = x\nfor i in range(1,L):\n    if i != L-1:\n        layer = nn_layer(layer,w[i],b[i],'tanh')\n    else:\n        layer = nn_layer(layer,w[i],b[i])\noutput = layer\ncost = tf.reduce_mean(tf.losses.mean_squared_error(labels=y,predictions=output))\noptimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)\nbatch_num = int(len(x_arr) \/ batch_size)\nx_batches = np.array_split(x_arr, batch_num)\ny_batches = np.array_split(y_arr, batch_num)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for epoch in range(epochs):\n        avg_cost = 0\n        for i in range(batch_num):\n            x_batch = np.expand_dims(x_batches[i],axis=1)\n            y_batch = np.expand_dims(y_batches[i],axis=1)\n            _, c = sess.run([optimizer,cost], feed_dict={x:x_batch, y:y_batch})\n            avg_cost += c\/batch_num\n        if epoch % display_step == 0:\n            print('Epoch {0} | cost = {1:.4f}'.format(epoch,avg_cost))\n    y_pred = sess.run(output,feed_dict={x:np.expand_dims(x_test,axis=1)})\n\nplt.figure(figsize=(7,7))\nplt.grid(True)\nplt.xlabel('x')\nplt.ylabel('f(x)')\nplt.scatter(x_arr,y_arr,c='b')\nplt.scatter(x_test,y_pred,c='r')\nplt.plot(x_vals,map(f,x_vals),c='m')\nplt.show()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Epoch 0 | cost = 4.5745\nEpoch 50 | cost = 1.7379\nEpoch 100 | cost = 1.1629\nEpoch 150 | cost = 0.4951\nEpoch 200 | cost = 0.2626\nEpoch 250 | cost = 0.1826\nEpoch 300 | cost = 0.1479\nEpoch 350 | cost = 0.1325\nEpoch 400 | cost = 0.1233\nEpoch 450 | cost = 0.1171\nEpoch 500 | cost = 0.1128\nEpoch 550 | cost = 0.1100\nEpoch 600 | cost = 0.1080\nEpoch 650 | cost = 0.1067\nEpoch 700 | cost = 0.1058\nEpoch 750 | cost = 0.1052\nEpoch 800 | cost = 0.1048\nEpoch 850 | cost = 0.1044\nEpoch 900 | cost = 0.1042\nEpoch 950 | cost = 0.1040\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/25GGh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/25GGh.png\" alt=\"lowlevel\"><\/a><\/p>\n\n<p><strong>Using <code>tf.layers.dense<\/code>:<\/strong><\/p>\n\n<pre><code>hidden_layers_size = [12,12]\nepochs=1000\nbatch_size=50\nlearning_rate=0.0003\ndisplay_step=50\n\ntf.reset_default_graph()\nx = tf.placeholder(name='x',shape=(None,1),dtype=tf.float32)\ny = tf.placeholder(name='y',shape=(None,1),dtype=tf.float32)\n\nlayer = tf.layers.dense(inputs=x, units=1, activation=tf.nn.tanh,\n                        kernel_initializer=tf.contrib.layers.xavier_initializer())\nfor l in hidden_layers_size:\n    layer = tf.layers.dense(inputs=layer, units=l, activation=tf.nn.tanh,\n                            kernel_initializer=tf.contrib.layers.xavier_initializer())\noutput = tf.layers.dense(inputs=layer, units=1, \n                         kernel_initializer=tf.contrib.layers.xavier_initializer())\n\ncost = tf.reduce_mean(tf.losses.mean_squared_error(labels=y,predictions=output))\noptimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)\nbatch_num = int(len(x_arr) \/ batch_size)\nx_batches = np.array_split(x_arr, batch_num)\ny_batches = np.array_split(y_arr, batch_num)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for epoch in range(epochs):\n        avg_cost = 0\n        for i in range(batch_num):\n            x_batch = np.expand_dims(x_batches[i],axis=1)\n            y_batch = np.expand_dims(y_batches[i],axis=1)\n            _, c = sess.run([optimizer,cost], feed_dict={x:x_batch, y:y_batch})\n            avg_cost += c\/batch_num\n        if epoch % display_step == 0:\n            print('Epoch {0} | cost = {1:.4f}'.format(epoch,avg_cost))\n    y_pred = sess.run(output,feed_dict={x:np.expand_dims(x_test,axis=1)})\n\nplt.figure(figsize=(7,7))\nplt.grid(True)\nplt.xlabel('x')\nplt.ylabel('f(x)')\nplt.scatter(x_arr,y_arr,c='b')\nplt.scatter(x_test,y_pred,c='r')\nplt.plot(x_vals,map(f,x_vals),c='m')\nplt.show()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Epoch 0 | cost = 3.3261\nEpoch 50 | cost = 1.9012\nEpoch 100 | cost = 1.8725\nEpoch 150 | cost = 1.8572\nEpoch 200 | cost = 1.7526\nEpoch 250 | cost = 1.4249\nEpoch 300 | cost = 1.3330\nEpoch 350 | cost = 1.3043\nEpoch 400 | cost = 1.2935\nEpoch 450 | cost = 1.2893\nEpoch 500 | cost = 1.2877\nEpoch 550 | cost = 1.2868\nEpoch 600 | cost = 1.2861\nEpoch 650 | cost = 1.2854\nEpoch 700 | cost = 1.2847\nEpoch 750 | cost = 1.2838\nEpoch 800 | cost = 1.2828\nEpoch 850 | cost = 1.2814\nEpoch 900 | cost = 1.2798\nEpoch 950 | cost = 1.2780\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Ulxab.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ulxab.png\" alt=\"dense\"><\/a><\/p>\n","answers":[{"AnswerId":"52755149","CreationDate":"2018-10-11T08:04:46.403","ParentId":null,"OwnerUserId":"5863503","Title":null,"Body":"<p>Found the issue - I've accidentally added an additional 1-unit layer with tanh after the input in my second attempt.<br>\nThis:<\/p>\n\n<pre><code>layer = tf.layers.dense(inputs=x, units=1, activation=tf.nn.tanh,\n                        kernel_initializer=tf.contrib.layers.xavier_initializer())\n<\/code><\/pre>\n\n<p>should be replaced with this:<\/p>\n\n<pre><code>layer = x\n<\/code><\/pre>\n"}]}
+{"QuestionId":52742811,"AnswerCount":1,"Tags":"<windows><tensorflow><deep-learning><anaconda>","CreationDate":"2018-10-10T14:41:30.613","AcceptedAnswerId":null,"OwnerUserId":1934948.0,"Title":"Optimize Tensorflow-GPU","Body":"<p>I am learning deep learning on Udemy. I have installed Tensorflow-GPU and it's working fine. In the tutorial, he uses Tensorflow on his CPU and it seems to work 10x faster then mine working on a GPU. I know i cannot compare the 2 but it seems weird. I have following questions:<\/p>\n\n<ol>\n<li>Am i doing anything wrong?<\/li>\n<li>If not, is there a way to improve the performance? If yes, please guide me for the same.<\/li>\n<\/ol>\n\n<p>My system specs:<\/p>\n\n<ul>\n<li>Intel Core i7-8550U<\/li>\n<li>NVIDIA GeFotce MX150<\/li>\n<li>Windows 10 64-Bit<\/li>\n<li>Anaconda environment<\/li>\n<\/ul>\n\n<p>Thanks.<\/p>\n\n<hr>\n\n<p>EDIT: I Used this command while running tensorflow <code>nvidia-smi -q -g 0 -d UTILIZATION -l<\/code> and got the following:\n<a href=\"https:\/\/i.stack.imgur.com\/siOPL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/siOPL.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52744609","CreationDate":"2018-10-10T16:12:50.320","ParentId":null,"OwnerUserId":"9475497","Title":null,"Body":"<p>For inference time, it might happens. Feeding data to GPU will cost some extra time. But for training time, it's weird. You may run <code>nvidia-smi<\/code> to see whether your code is actually running on GPU.<\/p>\n"}]}
+{"QuestionId":52743580,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-10T15:19:11.053","AcceptedAnswerId":52745373.0,"OwnerUserId":7867665.0,"Title":"Definition of sigmoid function in tensorflow","Body":"<p>I'm looking for the exact definition of the sigmoid function in Keras with Tensorflow backend. <\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/activations.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/activations.py<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/math_ops.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/math_ops.py<\/a><\/p>\n\n<p>Keras uses the tensorflow backend and tensorflow calls <\/p>\n\n<pre><code>from tensorflow.python.ops import gen_math_ops\n<\/code><\/pre>\n\n<p>which is machine generated. How can I find out the exact definition of the function if I've installed it in an anaconda environment?<\/p>\n","answers":[{"AnswerId":"52745373","CreationDate":"2018-10-10T17:02:34.557","ParentId":null,"OwnerUserId":"1322179","Title":null,"Body":"<p>As far as I can tell, it's defined in an external library, Eigen<\/p>\n\n<p><a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/src\/6ede590db55c9427bf35a115c011eb545438c84a\/Eigen\/src\/Core\/functors\/UnaryFunctors.h?at=default&amp;fileviewer=file-view-default#UnaryFunctors.h-831\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/eigen\/eigen\/src\/6ede590db55c9427bf35a115c011eb545438c84a\/Eigen\/src\/Core\/functors\/UnaryFunctors.h?at=default&amp;fileviewer=file-view-default#UnaryFunctors.h-831<\/a><\/p>\n\n<p>(the op looks like it was recently renamed: <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/diff\/Eigen\/src\/Core\/functors\/UnaryFunctors.h?diff2=a18cf733769b&amp;at=default\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/eigen\/eigen\/diff\/Eigen\/src\/Core\/functors\/UnaryFunctors.h?diff2=a18cf733769b&amp;at=default<\/a>)<\/p>\n\n<p>I got to that from here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/cwise_ops.h#L659\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/cwise_ops.h#L659<\/a><\/p>\n\n<p>They definitely didn't make it easy to find...<\/p>\n\n<p>A nice trick for tracing the breadcrumb trail through machine-generated python files is to load the directory in ipython and use the <code>??<\/code> lookup. i.e.<\/p>\n\n<pre><code>In [1]: from tensorflow.python.ops import gen_math_ops\n\nIn [2]: gen_math_ops??\n<\/code><\/pre>\n"}]}
+{"QuestionId":52743613,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-10T15:20:51.340","AcceptedAnswerId":52807228.0,"OwnerUserId":10484963.0,"Title":"Tensorflow data : apply function TO batch","Body":"<p>I'm using tf.data to iterate batch from large text corpus.<\/p>\n\n<p>I want to apply a function to only subset of data(or to subset of batch), not one by one element.\nSpecifically, my data iterator yields\n<code>query, reply<\/code> with batch. They are all positive pairs, so I just want to shuffle only subset of next batch(in this case, only \"reply\" batch\") to generate random negatives.<\/p>\n\n<p>For example, \ninputs:<\/p>\n\n<p><code>query1 reply1<\/code><\/p>\n\n<p><code>query2 reply2<\/code><\/p>\n\n<p><code>query3 reply3<\/code><\/p>\n\n<p>...<\/p>\n\n<p>outputs:<\/p>\n\n<ul>\n<li>positive pairs: <code>query1 reply1<\/code> (same as inputs)<\/li>\n<li>negative pairs: <code>query1 replyN<\/code> (replies are random shuffled)<\/li>\n<\/ul>\n\n<p>Of course it is possible to just shuffle text using python, but I want to make it efficient using tf.data because data size is too big.<\/p>\n","answers":[{"AnswerId":"52807228","CreationDate":"2018-10-14T21:24:30.203","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>Assuming you have <code>queries<\/code> and <code>replies<\/code> as two Tensors. What you need is I think something like below what you can concatenate then with the original batch.<\/p>\n\n<pre><code>batch_size = 10\ndef reply_shuffle(queries, replies):\n   shuffled_indices = tf.random_uniform(minval=0, maxval=batch_size+1, shape=[batch_size], dtype=tf.int32)\n   shuffled_replies = tf.gather_nd(replies, shuffled_indices) \n   return queries, shuffled_replies\n<\/code><\/pre>\n"}]}
+{"QuestionId":52743995,"AnswerCount":1,"Tags":"<python><tensorflow><pre-trained-model><transfer-learning>","CreationDate":"2018-10-10T15:40:19.083","AcceptedAnswerId":53121278.0,"OwnerUserId":2864250.0,"Title":"Moblenet V2: Importing graph with lower producer version into a graph with higher producer version","Body":"<p>I developing a transfer learning application where I am retraining MobileNetV2 for my data stream.<\/p>\n\n<p>I am retraining the model with <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">retrain.py<\/a> from tensorflow-hub and have made no modifications.  <\/p>\n\n<p>When I run the script from a terminal, I get this warning directly after the model is downloaded to a temporary directory in my user profile.<\/p>\n\n<pre><code>Importing a graph with a lower producer version 26 into an existing graph with producer \nversion 27. Shape inference will have run different parts of the graph with different \nproducer versions.\n<\/code><\/pre>\n\n<p>During my debugging efforts, I created a <code>test.py<\/code> script to find out where the warning is coming from:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_hub as hub\n\ndef create_module_graph(module_spec):\n  \"\"\"Creates a graph and loads Hub Module into it.\n\n  Args:\n    module_spec: the hub.ModuleSpec for the image module being used.\n\n  Returns:\n    graph: the tf.Graph that was created.\n    bottleneck_tensor: the bottleneck values output by the module.\n    resized_input_tensor: the input images, resized as expected by the module.\n    wants_quantization: a boolean, whether the module has been instrumented\n      with fake quantization ops.\n  \"\"\"\n  FAKE_QUANT_OPS = ('FakeQuantWithMinMaxVars',\n                     'FakeQuantWithMinMaxVarsPerChannel')\n  height, width = hub.get_expected_image_size(module_spec)\n  with tf.Graph().as_default() as graph:\n    resized_input_tensor = tf.placeholder(tf.float32, [None, height, width, 3])\n    m = hub.Module(module_spec)\n    bottleneck_tensor = m(resized_input_tensor)\n    wants_quantization = any(node.op in FAKE_QUANT_OPS\n                             for node in graph.as_graph_def().node)\n  return graph, bottleneck_tensor, resized_input_tensor, wants_quantization\n\n\ndef main():\n\n\n\n   module_spec = hub.load_module_spec('https:\/\/tfhub.dev\/google\/imagenet\/mobilenet_v2_100_96\/classification\/2')\n\n   graph, bottleneck_tensor, resized_input_tensor, wants_quantization = create_module_graph(module_spec)\n\n\n\nif __name__ =='__main__':\n   main()\n<\/code><\/pre>\n\n<p>And found it originated from the <code>create_module_graph<\/code> function in <code>retrain.py<\/code>.  When I run the script from the terminal using <code>python test.py<\/code>, I get producer warning from above.  However, when I run <code>main()<\/code> from an ipython console I do not get the producer version warnings.  <\/p>\n\n<p>I am unsure why I this is happening when all I am doing is creating a graph from the tensorflow-hub repo. I have looked over the <a href=\"https:\/\/www.tensorflow.org\/guide\/version_compat\" rel=\"nofollow noreferrer\">version compatibility docs<\/a> and did not see anything particularly relevant to the error.  Looking through the source <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/graph\/graph_constructor.cc#L1255\" rel=\"nofollow noreferrer\">code<\/a>, it seems to indicate that my graph is reduced to the lowest version before constructing it.<\/p>\n\n<ol>\n<li>Is this something to worry about?  <\/li>\n<li>Does it change how you will load the graph to make predictions?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53121278","CreationDate":"2018-11-02T15:16:32.517","ParentId":null,"OwnerUserId":"2864250","Title":null,"Body":"<p>From my tensorflow-hub <a href=\"https:\/\/github.com\/tensorflow\/hub\/issues\/180#issuecomment-435374790\" rel=\"nofollow noreferrer\">issue<\/a>:<\/p>\n\n<p>TensorFlow Hub modules contain <code>tf.GraphDefs<\/code> at the core, and these have a format version number to help with correctly importing graphs into newer versions of TensorFlow. It so happened that this format version was bumped from 26 to 27 between the upload of TF-Hub modules for the public launch on 31 March 2018 and the current TensorFlow release.<\/p>\n\n<p>However, we are currently not aware of any visible effect on module users from the reported change in shape inference, so our current recommendation is to ignore these warnings. They will go away with the\nnext refresh of modules.<\/p>\n"}]}
+{"QuestionId":52744467,"AnswerCount":1,"Tags":"<python><tensorflow><keras><word2vec><word-embedding>","CreationDate":"2018-10-10T16:05:02.407","AcceptedAnswerId":52745554.0,"OwnerUserId":9650492.0,"Title":"Keras - Translation from Sequential to Functional API","Body":"<p>I've been following Towards Data Science's tutorial about word2vec and skip-gram models, but I stumbled upon a problem that I cannot solve, despite searching about it a lot and trying multiple unsuccessful solutions.<\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/understanding-feature-engineering-part-4-deep-learning-methods-for-text-data-96c44370bbfa\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/understanding-feature-engineering-part-4-deep-learning-methods-for-text-data-96c44370bbfa<\/a><\/p>\n\n<p>The step that it shows you how to build the skip-gram model architecture seems deprecated because of the use of the Merge layer from keras.layers.<\/p>\n\n<p>What I tried to do was translate his piece of code - which is implemented in the Sequential API of Keras - to the Functional API to solve the deprecation of the Merge layer, by replacing it with the keras.layers.Dot layer. However, I'm still stuck in this step of merging the two models (word and context) into the final model, whose architecture must be like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eMt8A.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eMt8A.png\" alt=\"Skip-gram model summary and architecture\"><\/a><\/p>\n\n<p>Here's the code that the author used:<\/p>\n\n<pre><code>from keras.layers import Merge\nfrom keras.layers.core import Dense, Reshape\nfrom keras.layers.embeddings import Embedding\nfrom keras.models import Sequential\n\n# build skip-gram architecture\nword_model = Sequential()\nword_model.add(Embedding(vocab_size, embed_size,\n                         embeddings_initializer=\"glorot_uniform\",\n                         input_length=1))\nword_model.add(Reshape((embed_size, )))\n\ncontext_model = Sequential()\ncontext_model.add(Embedding(vocab_size, embed_size,\n                  embeddings_initializer=\"glorot_uniform\",\n                  input_length=1))\ncontext_model.add(Reshape((embed_size,)))\n\nmodel = Sequential()\nmodel.add(Merge([word_model, context_model], mode=\"dot\"))\nmodel.add(Dense(1, kernel_initializer=\"glorot_uniform\", activation=\"sigmoid\"))\nmodel.compile(loss=\"mean_squared_error\", optimizer=\"rmsprop\")\n<\/code><\/pre>\n\n<p>And here is my attempt to translate the Sequential code implementation into the Functional one:<\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\nfrom keras import Input, Model\n\nword_input = Input(shape=(1,))\nword_x = layers.Embedding(vocab_size, embed_size, embeddings_initializer='glorot_uniform')(word_input)\nword_reshape = layers.Reshape((embed_size,))(word_x)\n\nword_model = Model(word_input, word_reshape)    \n\ncontext_input = Input(shape=(1,))\ncontext_x = layers.Embedding(vocab_size, embed_size, embeddings_initializer='glorot_uniform')(context_input)\ncontext_reshape = layers.Reshape((embed_size,))(context_x)\n\ncontext_model = Model(context_input, context_reshape)\n\nmodel_input = layers.dot([word_model, context_model], axes=1, normalize=False)\nmodel_output = layers.Dense(1, kernel_initializer='glorot_uniform', activation='sigmoid')\n\nmodel = Model(model_input, model_output)\n<\/code><\/pre>\n\n<p>However, when executed, the following error is returned:<\/p>\n\n<blockquote>\n  <p>ValueError: Layer dot_5 was called with an input that isn't a symbolic\n  tensor. Received type: . Full\n  input: [,\n  ]. All inputs to\n  the layer should be tensors.<\/p>\n<\/blockquote>\n\n<p>I'm a total beginner to the Functional API of Keras, I will be grateful if you could give me some guidance in this situation on how could I input the context and word models into the dot layer to achieve the architecture in the image.              <\/p>\n","answers":[{"AnswerId":"52745554","CreationDate":"2018-10-10T17:14:37.937","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You are passing <code>Model<\/code> instances to the layer, however as the error suggests you need to pass Keras Tensors (i.e. outputs of layers or models) to layers in Keras. You have two option here. One is to use the <code>.output<\/code> attribute of the <code>Model<\/code> instance like this:<\/p>\n\n<pre><code>dot_output = layers.dot([word_model.output, context_model.output], axes=1, normalize=False)\n<\/code><\/pre>\n\n<p>or equivalently, you can use the output tensors directly:<\/p>\n\n<pre><code>dot_output = layers.dot([word_reshape, context_reshape], axes=1, normalize=False)\n<\/code><\/pre>\n\n<p>Further, you need to apply the <code>Dense<\/code> layer which is followed on the <code>dot_output<\/code> and pass instances of <code>Input<\/code> layer as inputs of <code>Model<\/code>. Therefore:<\/p>\n\n<pre><code>model_output = layers.Dense(1, kernel_initializer='glorot_uniform',\n                            activation='sigmoid')(dot_output)\n\nmodel = Model([word_input, context_input], model_output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52744559,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-10T16:10:11.260","AcceptedAnswerId":null,"OwnerUserId":9475497.0,"Title":"How to freeze some part of the variable in tensorflow","Body":"<p>I've noticed that a variable can be frozen by setting <code>trainable=False<\/code>. I'm wondering how to freeze only part of it instead of splitting it into two variables.<\/p>\n\n<p>For example, I only want to freeze the first element in <code>v = tf.Variable(tf.zeros([2])<\/code>. It seems that <code>grad_new = tf.multiply(grad, np.array([0,1]))<\/code> doesn't work.<\/p>\n","answers":[{"AnswerId":"52861719","CreationDate":"2018-10-17T18:50:31.250","ParentId":null,"OwnerUserId":"9475497","Title":null,"Body":"<p>Problem solved.<\/p>\n\n<p><code>grad_new = tf.multiply(grad, np.array([0,1]))<\/code> actually works quite well. It's the saver who mess up the code.<\/p>\n"}]}
+{"QuestionId":52745754,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-10T17:27:33.130","AcceptedAnswerId":null,"OwnerUserId":1965914.0,"Title":"Keras Functional API: Pooling input does not give correct output shape","Body":"<p>I am learning Keras with Tensorflow and I build as very simple model which takes input image of size<code>224x224x3<\/code> and applies <code>MaxPooling<\/code>, flatten it into 1D vector and this 1D vector is the output.<\/p>\n\n<p>Please note that, this model is just purely for tutorial purpose. However, when I see the summary of the model i.e., <code>model.summary()<\/code> the output of the pooling layer is of same size as the input.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>    in_shape = (224, 224, 3)\n    in_feats = Input(shape=in_shape)\n    pool = MaxPooling2D(pool_size=4, strides=1, padding='same')(in_feats)\n    flat = Flatten(name='flat')(pool)\n    model = Model(inputs=in_feats, outputs=flat)\n    # print summary\n    print (model.summary())\n<\/code><\/pre>\n\n<p>The summary is as follows:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, 224, 224, 3)       0         \n_________________________________________________________________\npool (MaxPooling2D)          (None, 224, 224, 3)       0         \n_________________________________________________________________\nflat (Flatten)               (None, 150528)            0         \n=================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>As you will notice the output from the pooling layer is same size as the input. What may be causing this?<\/p>\n","answers":[{"AnswerId":"52745917","CreationDate":"2018-10-10T17:38:18.420","ParentId":null,"OwnerUserId":"1965914","Title":null,"Body":"<p>It was a silly mistake. It result depends on the padding type used with the pooling layer.\nThe following gives the correct result. <\/p>\n\n<pre><code>pool = MaxPooling2D(pool_size=4, strides=1, padding='valid')(in_feats)\n<\/code><\/pre>\n\n<p>You can read more about the padding <a href=\"https:\/\/stackoverflow.com\/questions\/37674306\/what-is-the-difference-between-same-and-valid-padding-in-tf-nn-max-pool-of-t\">here<\/a><\/p>\n"}]}
+{"QuestionId":52746315,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-10T18:06:55.690","AcceptedAnswerId":null,"OwnerUserId":6947972.0,"Title":"Error converting .Pb filte to tflite using toco(tensorflow)","Body":"<p>I am following this tutorial <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets-2-tflite\/#2\" rel=\"nofollow noreferrer\">tensorflow for poets<\/a> and i created custom training data and converted it to .pb but when i try to convert .pb to .tflite using toco I am getting an error<\/p>\n\n<blockquote>\n  <p>2018-10-10 13:56:40.442130: I tensorflow\/contrib\/lite\/toco\/import_tensorflow.cc:1233] Converting unsupported operation: DecodeJpeg\n  2018-10-10 13:56:40.471545: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before Removing unused ops: 706 operators, 1102 arrays (0 quantized)\n  2018-10-10 13:56:40.485896: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] After Removing unused ops pass 1: 607 operators, 1003 arrays (0 quantized)\n  2018-10-10 13:56:40.502109: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before general graph transformations: 607 operators, 1003 arrays (0 quantized)\n  2018-10-10 13:56:40.581854: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] After general graph transformations pass 1: 129 operators, 327 arrays (0 quantized)\n  2018-10-10 13:56:40.584067: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] After general graph transformations pass 2: 126 operators, 321 arrays (0 quantized)\n  2018-10-10 13:56:40.586222: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] After general graph transformations pass 3: 125 operators, 319 arrays (0 quantized)\n  2018-10-10 13:56:40.588344: I tensorflow\/contrib\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before dequantization graph transformations: 125 operators, 319 arrays (0 quantized)\n  2018-10-10 13:56:40.590472: I tensorflow\/contrib\/lite\/toco\/allocate_transient_arrays.cc:311] Total transient array allocated size: 0 bytes, theoretical optimal value: 0 bytes.\n  2018-10-10 13:56:40.591866: F tensorflow\/contrib\/lite\/toco\/tflite\/export.cc:304] Some of the operators in the model are not supported by the standard TensorFlow Lite runtime. If you have a custom implementation for them you can disable this error with --allow_custom_ops. Here is a list of operators for which you will need custom implementations: DecodeJpeg, ExpandDims.\n  Aborted (core dumped)<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57936834","CreationDate":"2019-09-14T15:21:35.727","ParentId":null,"OwnerUserId":"10784736","Title":null,"Body":"<pre><code>toco \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --graph_def_file=tf_files\/object_graph.pb \\\n  --output_format=TFLITE \\\n  --output_file=tf_files\/qobject.lite \\\n  --inference_type=QUANTIZED_UINT8 \\\n  --inference_input_type=QUANTIZED_UINT8 \\\n  --input_arrays=input \\\n  --output_arrays=final_result \\\n  --input_shapes=1,224,224,3\\\n  --mean_values=128 \\\n  --std_dev_values=128 \\\n  --default_ranges_min=0 \\\n  --default_ranges_max=6\n<\/code><\/pre>\n\n<p>just use this this code convert .pb file in to lite and also quantized<\/p>\n"}]}
+{"QuestionId":52747020,"AnswerCount":1,"Tags":"<keras><coreml><coremltools>","CreationDate":"2018-10-10T18:56:03.780","AcceptedAnswerId":53409202.0,"OwnerUserId":9442808.0,"Title":"Keras to CoreML value arror: list.remove(x): x not in list","Body":"<p>Im trying to convert my Keras model (mobilenet + dence layers). The problem is that when I want to use coremltools for conversion I faced with the following problem:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"\/usr\/local\/lib\/python3.6\/dist-\npackages\/IPython\/core\/interactiveshell.py\", line 3265, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-4-7905693382e5&gt;\", line 1, in &lt;module&gt;\n    coreml_model = coremltools.converters.keras.convert(loaded_model)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_keras_converter.py\", line 752, in convert\n    custom_conversion_functions=custom_conversion_functions)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_keras_converter.py\", line 550, in convertToSpec\n    custom_objects=custom_objects)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_keras2_converter.py\", line 206, in _convert\n    graph.build()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_topology2.py\", line 687, in build\n    self._remove_old_edges(layer)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_topology2.py\", line 429, in _remove_old_edges\n    self._remove_edge(layer, succ)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/coremltools\/converters\/keras\/_topology2.py\", line 365, in _remove_edge\n    self.edge_map[src].remove(snk)\nValueError: list.remove(x): x not in list\n<\/code><\/pre>\n\n<p>Im trying to do this conversion by the following code:<\/p>\n\n<pre><code>js_file = open(args.ddir + args.mdl +'.json','r')\nloaded_json_model = js_file.read()\njs_file.close()\n\nfrom keras.applications import mobilenet\nfrom keras.utils.generic_utils import CustomObjectScope\nfrom keras.models import model_from_json\n\nwith CustomObjectScope({'relu6': mobilenet.mobilenet.relu6}):\n    loaded_model = model_from_json(loaded_json_model)\n    loaded_model.load_weights(args.ddir + args.mdl + '.h5')\n\ncoreml_model = coremltools.converters.keras.convert(loaded_model,\n                                                    input_names=\"image\",\n                                                    image_input_names=\"image\"\n                                                    )\n<\/code><\/pre>\n","answers":[{"AnswerId":"53409202","CreationDate":"2018-11-21T09:45:29.593","ParentId":null,"OwnerUserId":"9442808","Title":null,"Body":"<p>I solved the problem by using proper version of the Keras which includes both Mobilenet (feature extractor) and at the same time \"relu6\". The only version (up to now) that worked for me is version \"2.1.6\". By this version I successfully did the conversion.\ncoremltools does not supporting some layers for the moment (including relu6). This issue can be handled by \"CustomObjectScope\" an it was shown in the provided code.\nNote that, the network should be trained at this version (2.1.6) again.     <\/p>\n"}]}
+{"QuestionId":52747193,"AnswerCount":1,"Tags":"<python><machine-learning><conv-neural-network><pytorch><autoencoder>","CreationDate":"2018-10-10T19:07:48.983","AcceptedAnswerId":null,"OwnerUserId":7150642.0,"Title":"Convolution Autoencoder Image Dimension Error","Body":"<p>I have the following Convolutional Autoencoder setup:<\/p>\n\n<pre><code>class autoencoder(nn.Module):\ndef  __init__(self):\n    super(autoencoder, self).__init__()\n    self.encoder = nn.Sequential(\n        nn.Conv2d(1, 16, 3, stride=3, padding=1),  # b, 16, 10, 10\n        nn.ReLU(True),\n        nn.MaxPool2d(2, stride=2),  # b, 16, 5, 5\n        nn.Conv2d(16, 8, 3, stride=2, padding=1),  # b, 8, 3, 3\n        nn.ReLU(True),\n        nn.MaxPool2d(2, stride=1)  # b, 8, 2, 2                     \n                                )\n\n    self.decoder = nn.Sequential(\n        nn.ConvTranspose2d(8, 16, 3, stride=2),  # b, 16, 5, 5\n        nn.ReLU(True),\n        nn.ConvTranspose2d(16, 8, 5, stride=3, padding=1),  # b, 8, 15, 15\n        nn.ReLU(True),\n        nn.ConvTranspose2d(8, 1, 2, stride=2, padding=1),  # b, 1, 28, 28\n        nn.Tanh()          \n                                )\n<\/code><\/pre>\n\n<p>This is the main Loop:<\/p>\n\n<pre><code>for epoch in range(epochs):\nrunning_loss = 0\nfor data in (train_loader):\n    image,_=data\n    inputs = image.view(image.size(0),-1)\n    optimizer.zero_grad()\n    #image = np.expand_dims(img, axis=0)\n\n    outputs = net(inputs)\n    loss = criterion(outputs,inputs)\n    loss.backward()\n    optimizer.step()\n    running_loss += loss.data[0]\nprint('At Iteration : %d   ;  Mean-Squared Error : %f'%(epoch + 1,running_loss\/(train_set.train_data.size(0)\/batch_size)))\n<\/code><\/pre>\n\n<p>This is the Error:<\/p>\n\n<pre><code>RuntimeError: Expected 4-dimensional input for 4-dimensional weight [16, 1, 3, 3], but got input of size [1000, 784] instead\n<\/code><\/pre>\n\n<p>This has something to do with the flattening of the image but Im not exactly sure how to deflatten it. <\/p>\n","answers":[{"AnswerId":"52752818","CreationDate":"2018-10-11T05:29:16.120","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Why are you \"flattening\" your input image (2nd line of main loop):<\/p>\n\n<pre><code>inputs = image.view(image.size(0),-1)\n<\/code><\/pre>\n\n<p>This line turns your 4 dimensional <code>image<\/code> (batch - channels - height - width) to a two dimensional \"flat\" vector (batch - c * h * w).<br>\nYou <code>autoencoder<\/code> expects its inputs to be 4D and not \"flat\". just remove this line and you should be okay.<\/p>\n"}]}
+{"QuestionId":52747405,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T19:22:29.750","AcceptedAnswerId":null,"OwnerUserId":10485943.0,"Title":"Segmentation fault running layers.conv2d","Body":"<p>I have a simple test script to illustrate the problem I am having. I am trying to implement a CNN using tensorflow, but I started running into segmentation fault when changing the size of the input. In the test script, I can successfully run it if n_H=3000. But I got a segmentation fault when I set n_H=4000. In addition, if I run it without layars.conv2d by setting with_conv=False, the script runs successfully. Does anyone has any idea what my problem is?<\/p>\n\n<p>I am running this on a host with 12 CPUs. I don't really understand the messages from tensorflow about \"Creating new thread pool with default inter op setting: 2\". I don't know if this has anything to do with my problem.<\/p>\n\n<p>Here is the output when I get a segmentation fault:<\/p>\n\n<pre><code>$ python test.py\n(100, 4000, 100, 1) (100, 8)\n2018-10-10 11:57:23.825704: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2\n2018-10-10 11:57:23.827653: I tensorflow\/core\/common_runtime\/process_util.cc:69] Creating new thread pool with default inter op setting: 2. Tune using inter_op_parallelism_threads for best performance.\nFatal Python error: Segmentation fault\n\nThread 0x00007fea9af1a740 (most recent call first):\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1350 in _call_tf_sessionrun\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1263 in _run_fn\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1278 in _do_call\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1272 in _do_run\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1100 in _run\n  File \"\/home\/seng\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 877 in run\n  File \"test.py\", line 43 in &lt;module&gt;\nSegmentation fault (core dumped)\n<\/code><\/pre>\n\n<p>Here is test script:<\/p>\n\n<pre><code>import faulthandler; faulthandler.enable()\nimport numpy as np\nimport tensorflow as tf\n\nseed = 42\ntf.reset_default_graph()\ntf.set_random_seed(seed)\nnp.random.seed(seed)\n\nnum_examples = 100\nn_H = 4000\nwith_conv = True\n\n# generate training data\nX_gen = np.random.randn(num_examples*n_H*100).reshape(num_examples, n_H, 100, 1)\nY_gen = np.random.randn(num_examples*8).reshape(num_examples, 8)    \nX_train = X_gen[0:num_examples, 0:n_H, ...]\nY_train = Y_gen[0:num_examples, ...]\nprint(X_train.shape, Y_train.shape)\n\n# create placeholders\nX = tf.placeholder(tf.float32, shape=(None, n_H, 100, 1))\nY = tf.placeholder(tf.float32, shape=(None, 8))\n\n# build graph\nif (with_conv):\n    conv1 = tf.layers.conv2d(X, filters=64, kernel_size=[5, 5],strides = 1, padding='valid',activation = tf.nn.relu)    \n    pool1 = tf.layers.max_pooling2d(conv1, pool_size=[2, 2], strides=2, padding='valid')\nelse:\n    pool1 = tf.layers.max_pooling2d(X, pool_size=[2, 2], strides=2, padding='valid')    \npool1_flat = tf.layers.flatten(pool1)\ndense2 = tf.layers.dense(pool1_flat, units=256, activation=tf.nn.relu)\nH = tf.layers.dense(dense2, units=8, activation=tf.nn.relu)\n\n# compute cost\ncost = tf.reduce_mean(tf.square(Y - H))\n\n# initialize variables\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    a = sess.run(cost, feed_dict = {X: X_train, Y:Y_train})\n    print(a)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52757427","CreationDate":"2018-10-11T10:07:19.520","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>Usually, Segmentation fault means that your host doesn't have enough RAM to run the script. \nWhen you change the n_H value, even though you don't add any parameters to the convolutional layer, you add many parameters to the dense2 layer. Also, you add a lot of operations to the conv1 layer because the input is that much bigger.<\/p>\n\n<p>It is perfectly possible that the ops and the 1600 parameters contained in the conv1 layer just finish to saturate you RAM and render your script unrunable. Try tracking your RAM usage using 'htop' or any other tracker. <\/p>\n"}]}
+{"QuestionId":52747643,"AnswerCount":0,"Tags":"<python><keras><image-segmentation><data-generation>","CreationDate":"2018-10-10T19:39:33.993","AcceptedAnswerId":null,"OwnerUserId":7297511.0,"Title":"ValueError: `validation_data` should be a tuple","Body":"<p>I'm am working on some code and was wondering how I can resolve the error below? I am using class generator(keras.utils.Sequence) for more data generator. I am working with dicom images and trying to implement a keras resnet50_unet model. Any assistance would be greatly appreciated. Thank you. <\/p>\n\n<pre><code>history = model.fit_generator(train_gen, validation_data=valid_gen, \n                          epochs=25, steps_per_epoch = 200,shuffle=True, \nverbose=2)\n\n\n\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-18-8e6b34e90593&gt; in &lt;module&gt;()\n  1 history = model.fit_generator(train_gen, validation_data=valid_gen, \n----&gt; 2                               epochs=25, steps_per_epoch = \n200,shuffle=True, verbose=2)\n  3 \n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py in \nwrapper(*args, **kwargs)\n 89                 warnings.warn('Update your `' + object_name +\n 90                               '` call to the Keras 2 API: ' + signature, \n stacklevel=2)\n ---&gt; 91             return func(*args, **kwargs)\n 92         wrapper._original_function = func\n 93         return wrapper\n\n \/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in \n fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, \n validation_data, validation_steps, class_weight, max_queue_size, workers, \n use_multiprocessing, shuffle, initial_epoch)\n 1413             use_multiprocessing=use_multiprocessing,\n 1414             shuffle=shuffle,\n -&gt; 1415             initial_epoch=initial_epoch)\n 1416 \n 1417     @interfaces.legacy_generator_methods_support\n\n \/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training_generator.py \n in fit_generator(model, generator, steps_per_epoch, epochs, verbose, \n callbacks, validation_data, validation_steps, class_weight, max_queue_size, \n workers, use_multiprocessing, shuffle, initial_epoch)\n136                                      '`(val_x, val_y, \nval_sample_weight)` '\n137                                      'or `(val_x, val_y)`. Found: ' +\n--&gt; 138                                      str(validation_data))\n139                 val_x, val_y, val_sample_weights = \nmodel._standardize_user_data(\n140                     val_x, val_y, val_sample_weight)\n\nValueError: `validation_data` should be a tuple `(val_x, val_y, \nval_sample_weight)` or `(val_x, val_y)`. Found: &lt;__main__.generator object \nat 0x7f378513bac8&gt;\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52747655,"AnswerCount":2,"Tags":"<python><debugging><tensorflow><ncurses>","CreationDate":"2018-10-10T19:40:00.270","AcceptedAnswerId":52756287.0,"OwnerUserId":9067015.0,"Title":"Why do I get \"_curses.error: cbreak() returned ERR\" when using TensorFlow CLI Debugger?","Body":"<p>I am trying to use the TensorFlow CLI debugger in order to identify the operation which is causing a NaN during training of a network, but when I try to run the code I get an error:<\/p>\n\n<p><code>_curses.error: cbreak() returned ERR<\/code><\/p>\n\n<p>I'm running the code on an Ubuntu server, which I'm connecting to via SSH, and have tried to follow <a href=\"https:\/\/www.tensorflow.org\/guide\/debugger\" rel=\"nofollow noreferrer\">this tutorial<\/a>.<\/p>\n\n<p>I have tried using <code>tf.add_check_numerics_ops()<\/code>, but the layers in the network include while loops so are not compatible. This is the section of code where the error is being raised:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python import debug as tf_debug\n...\n#Prepare data\ntrain_data, val_data, test_data = dataset.prepare_datasets(model_config)\n\nsess = tf.Session()\nsess = tf_debug.LocalCLIDebugWrapperSession(sess)\n\n# Create iterators\nhandle = tf.placeholder(tf.string, shape=[])\niterator = tf.data.Iterator.from_string_handle(handle, train_data.output_types, train_data.output_shapes)\nmixed_spec, voice_spec, mixed_audio, voice_audio = iterator.get_next()\n\ntraining_iterator = train_data.make_initializable_iterator()\nvalidation_iterator = val_data.make_initializable_iterator()\ntesting_iterator = test_data.make_initializable_iterator()\n\ntraining_handle = sess.run(training_iterator.string_handle())\n...\n<\/code><\/pre>\n\n<p>and the full error is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"main.py\", line 64, in &lt;module&gt;\n    @ex.automain\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/sacred\/experiment.py\", line 137, in automain\n    self.run_commandline()\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/sacred\/experiment.py\", line 260, in run_commandline\n    return self.run(cmd_name, config_updates, named_configs, {}, args)\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/sacred\/experiment.py\", line 209, in run\n    run()\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/sacred\/run.py\", line 221, in __call__\n    self.result = self.main_function(*args)\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/sacred\/config\/captured_function.py\", line 46, in captured_function\n    result = wrapped(*args, **kwargs)\n  File \"main.py\", line 95, in do_experiment\n    training_handle = sess.run(training_iterator.string_handle())\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/debug\/wrappers\/framework.py\", line 455, in run\n    is_callable_runner=bool(callable_runner)))\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/debug\/wrappers\/local_cli_wrapper.py\", line 255, in on_run_start\n    self._run_start_response = self._launch_cli()\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/debug\/wrappers\/local_cli_wrapper.py\", line 431, in _launch_cli\n    title_color=self._title_color)\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/debug\/cli\/curses_ui.py\", line 492, in run_ui\n    self._screen_launch(enable_mouse_on_start=enable_mouse_on_start)\n  File \"\/home\/enterprise.internal.city.ac.uk\/acvn728\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/debug\/cli\/curses_ui.py\", line 445, in _screen_launch\n    curses.cbreak()\n_curses.error: cbreak() returned ERR\n<\/code><\/pre>\n\n<p>I'm pretty new to using Ubuntu (and TensorFlow), but as far as I can tell the server does have ncurses installed, which should allow the required curses based interface:<\/p>\n\n<pre><code>acvn728@america:~\/MScFinalProject$ dpkg -l '*ncurses*' | grep '^ii'\nii  libncurses5:amd64  6.0+20160213-1ubuntu1 amd64        shared libraries for terminal handling\nii  libncursesw5:amd64 6.0+20160213-1ubuntu1 amd64        shared libraries for terminal handling (wide character support)\nii  ncurses-base       6.0+20160213-1ubuntu1 all          basic terminal type definitions\nii  ncurses-bin        6.0+20160213-1ubuntu1 amd64        terminal-related programs and man pages\nii  ncurses-term       6.0+20160213-1ubuntu1 all          additional terminal type definitions\n<\/code><\/pre>\n","answers":[{"AnswerId":"52749488","CreationDate":"2018-10-10T22:14:21.550","ParentId":null,"OwnerUserId":"4518274","Title":null,"Body":"<p><code>cbreak<\/code> would return <strong><code>ERR<\/code><\/strong> if you run a curses application that is not on a <em>real terminal<\/em> (i.e., something that works with <a href=\"http:\/\/pubs.opengroup.org\/onlinepubs\/9699919799\/basedefs\/V1_chap11.html#tag_11\" rel=\"nofollow noreferrer\">POSIX termios calls<\/a>).<\/p>\n\n<p>From the description,<\/p>\n\n<blockquote>\n  <p>but the layers in the network include while loops so are not compatible<\/p>\n<\/blockquote>\n\n<p>it does not seem you are running in a terminal.<\/p>\n"},{"AnswerId":"52756287","CreationDate":"2018-10-11T09:07:41.953","ParentId":null,"OwnerUserId":"9067015","Title":null,"Body":"<p>Problem solved! The solution was to change<\/p>\n\n<pre><code>sess = tf_debug.LocalCLIDebugWrapperSession(sess)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>sess = tf_debug.LocalCLIDebugWrapperSession(sess, ui_type=\"readline\")\n<\/code><\/pre>\n\n<p>This is similar to the solution to <a href=\"https:\/\/stackoverflow.com\/questions\/47833697\/how-to-use-tensorflow-debugging-tool-tfdbg-on-tf-estimator-in-tensorflow\">this question<\/a>, but I I think it is important to note that they are different because a) it refers to a different function and a different API and b) I wasn't trying to run from an IDE, as mentioned in that solution.<\/p>\n"}]}
+{"QuestionId":52747947,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-10T20:02:26.230","AcceptedAnswerId":null,"OwnerUserId":10486207.0,"Title":"Tensorflow read_file() does nothing","Body":"<p>I am trying to read and decode an image file using Tensorflow. I have the following code:<\/p>\n\n<pre><code>dir_path = os.path.dirname(os.path.realpath(__file__))\nfilename = dir_path + '\/images\/cat\/cat1.jpg'\nimage_file = tf.read_file(filename)\nimage_decoded = tf.image.decode_jpeg(image_file, channels=3)\n\nprint(image_file)\nprint(image_decoded)\n<\/code><\/pre>\n\n<p>This results in the following output:<\/p>\n\n<p><code>Tensor(\"ReadFile:0\", shape=(), dtype=string)\nTensor(\"DecodeJpeg:0\", shape=(?, ?, 3), dtype=uint8)<\/code><\/p>\n\n<p>It appears as if the file is not read at all by Tensorflow. However I could not find any error messages indicating that something went wrong. I don't know how I can resolve this, any help would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"52748056","CreationDate":"2018-10-10T20:10:35.213","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>Tensorflow creates a computation graph which should then be evaluated. What you see over there in the result is the op that is created. You need to define a Session object to get the results of your operations.<\/p>\n\n<pre><code>dir_path = os.path.dirname(os.path.realpath(__file__))\nfilename = dir_path + '\/images\/cat\/cat1.jpg'\nimage_file = tf.read_file(filename)\nimage_decoded = tf.image.decode_jpeg(image_file, channels=3)\nwith tf.Session() as sess:\n     f, img = sess.run([image_file, image_decoded])\n     print(f)\n     print(img)\n<\/code><\/pre>\n\n<p>Check out this <a href=\"https:\/\/www.tensorflow.org\/guide\/#low_level_apis\" rel=\"nofollow noreferrer\">tensorflow<\/a> resource to help you further understand!<\/p>\n"},{"AnswerId":"55769602","CreationDate":"2019-04-20T01:56:05.610","ParentId":null,"OwnerUserId":"11381722","Title":null,"Body":"<p>That was one big obstacle when we first tried to use Tensorflow.<\/p>\n\n<p>Now, Tensorflow team has made a solution.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n<\/code><\/pre>\n\n<p>Run the above two lines at the very beginning of your program.<\/p>\n\n<p>Then your print functions will produce something like the followings.<\/p>\n\n<pre><code>&lt;tf.Tensor: id=15, shape=(), dtype=string, numpy=b'\\xff\\xd8\\xff\\xe0\\x00\\x10JFIF\\x00\\x01\\x01\\x00\\x00\\\n&lt;tf.Tensor: id=17, shape=(747, 1024, 3), dtype=uint8, numpy=\narray([[[ 0,  0,  0],\n        [ 0,  0, \n<\/code><\/pre>\n\n<p>Tensorflow team calls it eager execution. Now we don't need to run the tf.Session to see what are inside tensors.<\/p>\n"}]}
+{"QuestionId":52748020,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2018-10-10T20:08:27.880","AcceptedAnswerId":null,"OwnerUserId":9361060.0,"Title":"How can batch include both positive and negative labels?","Body":"<p>I have pedestrian classification problem which I want to solve with VGG-16. In order to do that, I prepared train and test sets. My train set has 2038 images and my test set has 252 images. My batch size is 64. How can I tell Keras that I want these 64 images to include both positive and negative labels while training? I don't want it to learn only on positive or only on negative labels.<\/p>\n","answers":[{"AnswerId":"54429720","CreationDate":"2019-01-29T21:17:37.513","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>If you shuffle your training data (which is strongly recommended), there is a very small chance to get all the 64 samples belonging to the same class. For the two-class task there should be no problem with the data.<\/p>\n\n<p>However, if you wish to guarantee balanced training batches, you can use third-party code like <a href=\"https:\/\/imbalanced-learn.readthedocs.io\/en\/latest\/generated\/imblearn.keras.BalancedBatchGenerator.html#imblearn.keras.BalancedBatchGenerator\" rel=\"nofollow noreferrer\">BalancedBatchGenerator<\/a>. See this <a href=\"https:\/\/imbalanced-learn.readthedocs.io\/en\/latest\/auto_examples\/applications\/porto_seguro_keras_under_sampling.html\" rel=\"nofollow noreferrer\">tutorial<\/a> for further information.<\/p>\n"}]}
+{"QuestionId":52749091,"AnswerCount":1,"Tags":"<python><tensorflow><keras><dataset>","CreationDate":"2018-10-10T21:33:08.427","AcceptedAnswerId":null,"OwnerUserId":9401308.0,"Title":"Build my own dataset to use with Tensorflow or Keras","Body":"<p>Currently I load my dataset this way:<\/p>\n\n<pre><code>from keras.datasets import mnist\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n<\/code><\/pre>\n\n<p>But I would like to use my own dataset that is labeled in folders this way:<\/p>\n\n<pre><code>flower_phoyos\/daisy \nflower_phoyos\/dandelion\nflower_phoyos\/roses\nflower_phoyos\/sunflowers\nflower_phoyos\/tulips\n<\/code><\/pre>\n\n<p>Each subfolder is a class to be classified, the classifier must sort daisy, dandelion, roses, sunflowers ou tulips.<\/p>\n\n<p>The solution can be with tensorflow or keras<\/p>\n","answers":[{"AnswerId":"52749384","CreationDate":"2018-10-10T22:03:02.237","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>Not the most elegant method, but for a small dataset this should be sufficient. \nFirst get a list of all of the files, and note their classes:<\/p>\n\n<pre><code>import os\nimport numpy as np\nfrom matplotlib import pyplot as plt\n\nfile_list = []\nclass_list = []\n\nclassnames = ['daisy', 'dandelion', 'roses', 'sunflowers', 'tulips']   \nfor class_name in classnames:\n    dir_list = os.listdir(os.path.join('flower_phoyos', class_name))\n    file_list.extend(dir_list)\n    class_list.extend([classnames] * len(dir_list))\n<\/code><\/pre>\n\n<p>Now you know how many examples you have (<code>len(file_list)<\/code>) and can read them in, assigning them to a numpy array.<\/p>\n\n<pre><code>img_data = np.zeros(len(file_list), height, width, n_bands)\n# Of course I've assumed that you want all of the images to be the same size.\nfor ii, file_ in enumerate(file_list):\n    img = plt.imread(file_)\n    # Resize to height, width if your images are not all the same size.\n    img_data[ii, :, :, :] = img\nclass_labels = np.array(class_list)\n<\/code><\/pre>\n\n<p>You can now reorder your img_data and class_labels (making sure you reorder them in the same way) and split them into train and test.<\/p>\n"}]}
+{"QuestionId":52749448,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tf-slim>","CreationDate":"2018-10-10T22:09:24.567","AcceptedAnswerId":null,"OwnerUserId":2529125.0,"Title":"How to retrain the last fully connected layer for a ssd mobilenet v2 model","Body":"<p>I currently retrained an ssd mobile net v2 model using the <a href=\"https:\/\/youtube\/JR8CmWyh2E8\" rel=\"nofollow noreferrer\">tutorial<\/a>. I ran the following in command line to execute the retraining. It retrained the entire model.<\/p>\n\n<pre><code>(tensorflow) c:\\models-master\\research&gt;python object_detection\/legacy\/train.py --logtostderr --train_dir=training\/model2\/ --pipeline_config_path=training\/ssd_mobilenet_v2_2.config\n<\/code><\/pre>\n\n<p>I set the config file as such:<\/p>\n\n<pre><code># SSD with Mobilenet v2 configuration for MSCOCO Dataset.\n# Users should configure the fine_tune_checkpoint field in the train config as\n# well as the label_map_path and input_path fields in the train_input_reader and\n# eval_input_reader. Search for \"PATH_TO_BE_CONFIGURED\" to find the fields that\n# should be configured.\n\nmodel {\n  ssd {\n    num_classes: 1\n    box_coder {\n      faster_rcnn_box_coder {\n        y_scale: 10.0\n        x_scale: 10.0\n        height_scale: 5.0\n        width_scale: 5.0\n      }\n    }\n    matcher {\n      argmax_matcher {\n        matched_threshold: 0.5\n        unmatched_threshold: 0.5\n        ignore_thresholds: false\n        negatives_lower_than_unmatched: true\n        force_match_for_each_row: true\n      }\n    }\n    similarity_calculator {\n      iou_similarity {\n      }\n    }\n    anchor_generator {\n      ssd_anchor_generator {\n        num_layers: 6\n        min_scale: 0.2\n        max_scale: 0.95\n        aspect_ratios: 1.0\n        aspect_ratios: 2.0\n        aspect_ratios: 0.5\n        aspect_ratios: 3.0\n        aspect_ratios: 0.3333\n      }\n    }\n    image_resizer {\n      fixed_shape_resizer {\n        height: 300\n        width: 300\n      }\n    }\n    box_predictor {\n      convolutional_box_predictor {\n        min_depth: 0\n        max_depth: 0\n        num_layers_before_predictor: 0\n        use_dropout: false\n        dropout_keep_probability: 0.8\n        kernel_size: 3\n        box_code_size: 4\n        apply_sigmoid_to_scores: false\n        conv_hyperparams {\n          activation: RELU_6,\n          regularizer {\n            l2_regularizer {\n              weight: 0.00004\n            }\n          }\n          initializer {\n            truncated_normal_initializer {\n              stddev: 0.03\n              mean: 0.0\n            }\n          }\n          batch_norm {\n            train: true,\n            scale: true,\n            center: true,\n            decay: 0.9997,\n            epsilon: 0.001,\n          }\n        }\n      }\n    }\n    feature_extractor {\n      type: 'ssd_mobilenet_v2'\n      min_depth: 16\n      depth_multiplier: 1.0\n      use_depthwise: true\n      conv_hyperparams {\n        activation: RELU_6,\n        regularizer {\n          l2_regularizer {\n            weight: 0.00004\n          }\n        }\n        initializer {\n          truncated_normal_initializer {\n            stddev: 0.03\n            mean: 0.0\n          }\n        }\n        batch_norm {\n          train: true,\n          scale: true,\n          center: true,\n          decay: 0.9997,\n          epsilon: 0.001,\n        }\n      }\n    }\n    loss {\n      classification_loss {\n        weighted_sigmoid {\n        }\n      }\n      localization_loss {\n        weighted_smooth_l1 {\n        }\n      }\n      hard_example_miner {\n        num_hard_examples: 3000\n        iou_threshold: 0.99\n        loss_type: CLASSIFICATION\n        max_negatives_per_positive: 3\n        min_negatives_per_image: 3\n      }\n      classification_weight: 1.0\n      localization_weight: 1.0\n    }\n    normalize_loss_by_num_matches: true\n    post_processing {\n      batch_non_max_suppression {\n        score_threshold: 1e-8\n        iou_threshold: 0.6\n        max_detections_per_class: 100\n        max_total_detections: 100\n      }\n      score_converter: SIGMOID\n    }\n  }\n}\n\ntrain_config: {\n  batch_size: 24\n  optimizer {\n    rms_prop_optimizer: {\n      learning_rate: {\n        exponential_decay_learning_rate {\n          initial_learning_rate: 0.004\n          decay_steps: 800720\n          decay_factor: 0.95\n        }\n      }\n      momentum_optimizer_value: 0.9\n      decay: 0.9\n      epsilon: 1.0\n    }\n  }\n  fine_tune_checkpoint: \"D:\/Databases\/Coco\/cctv\/tf\/ssd_mobilenet_v2_coco_2018_03_29\/model.ckpt\"\n  fine_tune_checkpoint_type:  \"detection\"\n  # Note: The below line limits the training process to 200K steps, which we\n  # empirically found to be sufficient enough to train the pets dataset. This\n  # effectively bypasses the learning rate schedule (the learning rate will\n  # never decay). Remove the below line to train indefinitely.\n  num_steps: 200000\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n  data_augmentation_options {\n    ssd_random_crop {\n    }\n  }\n}\n\ntrain_input_reader: {\n  tf_record_input_reader {\n    input_path: \"D:\/Code\/Image\/cctvmodel\/tfrecordfinalALL\/train2.record\"\n  }\n  label_map_path: \"D:\/Code\/Image\/cctvmodel\/label_map.pbtxt\"\n}\n\neval_config: {\n  num_examples: 8000\n  # Note: The below line limits the evaluation process to 10 evaluations.\n  # Remove the below line to evaluate indefinitely.\n  max_evals: 10\n}\n\neval_input_reader: {\n  tf_record_input_reader {\n    input_path: \"D:\/Code\/Image\/cctvmodel\/tfrecordfinalALL\/test2.record\"\n  }\n  label_map_path: \"D:\/Code\/Image\/cctvmodel\/label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n<\/code><\/pre>\n\n<p>How can one retrain only the last fully connected layer? Or does one freeze the upper layers\/weight and retrain?<\/p>\n\n<p>Note that in the pretrained model folder there is .pb file and a single checkpoint file as shown below:\n<a href=\"https:\/\/i.stack.imgur.com\/gAn3W.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gAn3W.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have found a link for <a href=\"https:\/\/datascience.stackexchange.com\/questions\/17522\/fine-tuning-a-model-from-an-existing-checkpoint-with-tensorflow-slim\">fine tuning at existing checkpoint<\/a>. More details can be found here<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\" rel=\"nofollow noreferrer\">link<\/a>. The  example code is :<\/p>\n\n<pre><code>$ DATASET_DIR=\/tmp\/flowers\n$ TRAIN_DIR=\/tmp\/flowers-models\/inception_v3\n$ CHECKPOINT_PATH=\/tmp\/my_checkpoints\/inception_v3.ckpt\n$ python train_image_classifier.py \\\n    --train_dir=${TRAIN_DIR} \\\n    --dataset_dir=${DATASET_DIR} \\\n    --dataset_name=flowers \\\n    --dataset_split_name=train \\\n    --model_name=inception_v3 \\\n    --checkpoint_path=${CHECKPOINT_PATH} \\\n    --checkpoint_exclude_scopes=InceptionV3\/Logits,InceptionV3\/AuxLogits \\\n    --trainable_scopes=InceptionV3\/Logits,InceptionV3\/AuxLogits\n<\/code><\/pre>\n\n<p>How should I set the checkpoints to ensure it trains the last fully connected layer?<\/p>\n","answers":[]}
+{"QuestionId":52749816,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><keras>","CreationDate":"2018-10-10T22:49:42.377","AcceptedAnswerId":null,"OwnerUserId":1009508.0,"Title":"How to use a image set in RNN model","Body":"<p>Hello I am trying to do my first RNN using Keras and Tensorflow, but I am getting stuck on an issue or reshaping my images to fit into the model. <\/p>\n\n<p>I have looked at this post but could not figure out about the reshaping:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47671732\/keras-input-a-3-channel-image-into-lstm\">Keras - Input a 3 channel image into LSTM<\/a><\/p>\n\n<p>What I have is a bunch of images that are taken at every frame in a video. I saved all the frames outside of python so I have a very large folder of images.I separated the frames into 21 frames for a segment so 21 images per motion that I want to capture.  I want to read in these 21 images as one sequence. I have the same sequence captured from multiple cameras\/angles which I want to us in this model.  What I want to try is to model a movement and see if a person is doing this movement or not, so it is a binary model yes or no basically. Not the most sophisticated but its a learning process to use this model and keras. <\/p>\n\n<p>I need help figuring out how to use these images inside the keras model. I have looked at a few tutorials on MINST data set but that didnt help me figure this out.\nAny help will be appreciated.   <\/p>\n\n<p>This is the error that is given to me when I try to train the model<\/p>\n\n<pre><code>ValueError: Error when checking input: expected lstm_1_input to have 3 dimensions, but got array with shape (2026, 200, 200, 1)\n<\/code><\/pre>\n\n<p>My code is this:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation\nfrom keras.layers import LSTM\nfrom tqdm import tqdm\nimport cv2\nimport os\nimport numpy as np\n\nimageSize = 200\n\n#create lables for each image\ndef labelImage(img):\n    wordLabel = img.split('.')[-3]\n    #Conversion to one hot array [lat,not]\n    if wordLabel == \"FWAC\":\n        return[1,0]\n    else:\n        return[0,1]\n\n#Process images and add lables\n#Convert data into an array and add its lable\ndef makeTrainingData():\n    print(\"Creating Training Data\")\n    trainingData = []\n    for img in tqdm(os.listdir(trainDir)):\n        label = labelImage(img)\n        path = os.path.join(trainDir,img)\n        img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (imageSize,imageSize))\n        trainingData.append([np.array(img),np.array(label)])\n\n    #Save the array file to load it into other models if needed\n    np.save(\"trainingData.npy\", trainingData)\n    print(\"Training Data Saved\")\n    return trainingData\n\n#process the testing data in the same manner\ndef processTestData():\n    print(\"Creating Testing Data\")\n    testData = []\n    for img in tqdm(os.listdir(testDri)):\n        print(\"image\", img)\n        path = os.path.join(testDri, img)\n        imgNum = img.split(\".\")[0]\n        img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (imageSize, imageSize))\n        testData.append([np.array(img), imgNum])\n\n    np.save(\"testingData.npy\", testData)\n    print(\"Testing Data Saved\")\n    return testData\n\n\n\nrnnSize = 512\n\nmodel = Sequential()\nmodel.add(LSTM(rnnSize, input_shape=(imageSize, imageSize)))\nmodel.add(Dense(1024))\nmodel.add(Activation('relu'))\nmodel.add(Dense(50))\nmodel.add(Activation('sigmoid'))\nmodel.add(Dense(3))  \nmodel.add(Activation('softmax'))\n\n\nmodel.compile(loss='mean_squared_error', optimizer='adam',metrics=['accuracy'])\n\n#Data\ntrainDir = \"D:\/TrainingDataSets\/TrainingSet\/\"\ntestDri = \"D:\/TrainingDataSets\/TestingSet\/\"\n\n#trainData = makeTrainingData()\n#testData = processTestData()\ntrainData = np.load('trainingData.npy')\ntestData = np.load(\"testingData.npy\")\n#resize the image to this See above\ntrain = trainData[:-500]\ntest = trainData[-200:]\n\nx = []\ny = []\nfor xi in trainData:\n    x.append(xi[0].reshape((-1, imageSize, imageSize)))\n    y.append(xi[1])\n\nx_train = np.array([i[0] for i in train]).reshape(-1,imageSize, imageSize,1)\ny_train = [i[1] for i in train]\n\n\n\ntest_x = np.array([i[0] for i in test]).reshape(-1,imageSize , imageSize,1)\ntest_y = [i[1] for i in test]\n\n\nepoch = 5\nbatchSize = 100\n\nmodel.fit(x_train, y_train, epochs=epoch, batch_size= batchSize, verbose=1, shuffle=False)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52749836,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><lstm>","CreationDate":"2018-10-10T22:51:28.423","AcceptedAnswerId":null,"OwnerUserId":3511901.0,"Title":"Possible to add a bidirectional LSTM before a CNN in Keras?","Body":"<p>I am currently working on a system that classifies whether two sentences share the same content or not. For this purpose I use pretrained word vectors, so there is an array with the word vectors of sentence one (s1) and an array with the word vectors of sentence 2 (s2). In order to classify whether they are similar or not I create a matrix by comparing all vectors in s1 pairwise with the vectors in s2. This matrix is then fed into a CNN classifier and trained on the data. This all pretty straight forward.<\/p>\n\n<p>Now I would like to enhance this system by making using bidirectional LSTMs on s1 and s2. The bidirectional LSTM should be used in order to get the hidden state of each vector in s1 and s2 and these hidden states should then be compared in the same way by pairwise cosine similarity as the vectors of s1 and s2 are compared before. This is in order to capture the information of the sentence context of each word in s1 and s2.<\/p>\n\n<p>Now the question is how to do this in Keras. Currently I am using numpy\/sklearn to create the matrices which are then fed as training data into Keras. \nI found one implementation of what I want to do in plain tensorflow (<a href=\"https:\/\/github.com\/LiuHuiwen\/Pairwise-Word-Interaction-Modeling-by-Tensorflow-1.0\/blob\/master\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/LiuHuiwen\/Pairwise-Word-Interaction-Modeling-by-Tensorflow-1.0\/blob\/master\/model.py<\/a>). <\/p>\n\n<p>I assume that I will have to change the input data to consist of just the two arrays of vectors of s1 and s2. Then I have to run the biLSTM first, get the hidden states, convert everything into matrices and feed this into the CNN. The example in plain tensorflow seems to be quite clear to me, but I cannot come up with an idea of how to do this in Keras. Is it possible at all in Keras or does one have to resort to tensorflow directly in order to do the necessary calculations on the output of the biLSTM? <\/p>\n","answers":[{"AnswerId":"52751219","CreationDate":"2018-10-11T02:19:55.467","ParentId":null,"OwnerUserId":"10487091","Title":null,"Body":"<p>Keras RNN layer including LSTM can return not only the last output in the output sequence but also the full sequence from all hidden layers using <code>return_sequences=True<\/code> option.<\/p>\n\n<p><a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/recurrent\/<\/a><\/p>\n\n<p>When you want to connect Bi-Directional LSTM layer before CNN layer, the following code is an example:<\/p>\n\n<pre><code>from keras.layers import Input, LSTM, Bidirectional, Conv1D\n\ninput = Input(shape=(50, 200))\nseq = Bidirectional(LSTM(16, return_sequences=True))(input)\ncnn = Conv1D(32, 3, padding=\"same\", activation=\"relu\")(seq)\n<\/code><\/pre>\n\n<p>Please note: If you want to use Conv2D layer after Bi-Directional LSTM layer, reshaping to <code>ndim=4<\/code> is required for input of Conv2D, like the following code:<\/p>\n\n<pre><code>from keras.layers import Input, LSTM, Bidirectional, Conv2D, Reshape\n\ninput = Input(shape=(50, 200))\nseq = Bidirectional(LSTM(16, return_sequences=True))(input)\nseq = Reshape((50, 32, 1))(seq)\ncnn = Conv2D(32, (3, 3), padding=\"same\", activation=\"relu\")(seq)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52750622,"AnswerCount":1,"Tags":"<pytorch><powerpc>","CreationDate":"2018-10-11T00:49:29.460","AcceptedAnswerId":null,"OwnerUserId":4595807.0,"Title":"How to install pytorch on Power 8 or PPC64 machine?","Body":"<p>I am trying to install <strong>pytorch<\/strong> using <strong>conda<\/strong> on <strong>power 8 IBM machine<\/strong>. Although, I read articles from <a href=\"https:\/\/developer.ibm.com\/tutorials\/install-pytorch-on-power\/\" rel=\"nofollow noreferrer\">IBM blog<\/a>, I couldn't install with success because I got stuck with compiling <strong>magma<\/strong>.<\/p>\n","answers":[{"AnswerId":"52750623","CreationDate":"2018-10-11T00:49:29.460","ParentId":null,"OwnerUserId":"4595807","Title":null,"Body":"<p>Assuming that <strong>conda<\/strong> is already installed.\nSimply run the following command<\/p>\n\n<pre><code>conda install -c engility pytorch\nconda install -c engility torchvision\n<\/code><\/pre>\n\n<p>Note:<br>\n1. Goto this <a href=\"https:\/\/anaconda.org\/\" rel=\"nofollow noreferrer\">anaconda<\/a> page<br>\n2. Search for <strong>pytorch<\/strong><br>\n3. Scroll down to see which one has <strong>linux-ppc64le<\/strong> as <strong>platform<\/strong><br>\n4. Click into that specific package<br>\n5. You will get the command to install pytorch<\/p>\n"}]}
+{"QuestionId":52750690,"AnswerCount":1,"Tags":"<python><image><optimization><scipy><keras>","CreationDate":"2018-10-11T00:59:35.637","AcceptedAnswerId":null,"OwnerUserId":9581211.0,"Title":"How to use scipy to optimize a function","Body":"<p>I'm learning image style migration with scipy and keras. I'm trying to run a sample code but get an error with scipy.optimize.fmin_l_bfgs_b:<\/p>\n\n<pre><code>x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),fprime=evaluator.grads, maxfun=20, epsilon=1e-7)\n<\/code><\/pre>\n\n<p>But I get an error:<\/p>\n\n<pre><code>ValueError: failed to initialize intent(inout) array -- expected elsize=8 but got 4\n<\/code><\/pre>\n\n<p>Where evaluator is a object of class Evaluator<\/p>\n\n<pre><code>class Evaluator():\n    def __init__(self):\n        self.loss_value = None\n        self.grad_values = None\n\n    def loss(self, x):\n        assert self.loss_value is None\n        loss_value, grad_values = evalurateLossAndGrads(x)\n        self.loss_value = loss_value\n        self.grad_values = grad_values\n        return self.loss_value\n\n    def grads(self,x ):\n        assert self.loss_value is not None\n        grad_values = np.copy(self.grad_values)\n        self.loss_value = None\n        self.grad_values = None\n        return grad_values\n<\/code><\/pre>\n\n<p>The method loss returns a scalar while grads returns a vector which length equals to the number of pixels in the image. Name <code>x<\/code> is a image of size 512*512*3.\nHow can I fix this bug? Thank you very much.<\/p>\n","answers":[{"AnswerId":"58533942","CreationDate":"2019-10-24T04:00:02.633","ParentId":null,"OwnerUserId":"10076057","Title":null,"Body":"<p>This error is because the fortran implementation of scipy's L-BFGS solver requires 8-byte instead of 4-byte floats. It requires you to cast your initial x and your gradients to <code>float64<\/code> datatype.<\/p>\n"}]}
+{"QuestionId":52751040,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><scikit-learn><keras>","CreationDate":"2018-10-11T01:55:51.053","AcceptedAnswerId":52752228.0,"OwnerUserId":1511274.0,"Title":"Different exceptions happened when running Keras and scikit-learn","Body":"<p>I try to pass a Keras model (as a function) to KerasClassifier wrapper from scikit_learn, and then use GridSearchCV to create some settings, and finally fit the train and test datasets (both are numpy array)<\/p>\n\n<p>I then, with the same python script, got <strong>different exceptions<\/strong>, some of them are:<\/p>\n\n<p>_1.<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"mnist_flat_imac.py\", line\n  63, in \n      grid_result = validator.fit(train_images, train_labels)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/model_selection\/_search.py\",\n  line 626, in fit\n      base_estimator = clone(self.estimator)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 62, in clone\n      new_object_params[name] = clone(param, safe=False)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 53, in clone<\/p>\n  \n  <p><strong><em>snipped here<\/em><\/strong><\/p>\n  \n  <p>in _deepcopy_dict\n      y[deepcopy(key, memo)] = deepcopy(value, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 174,\n  in deepcopy\n      rv = reductor(4) TypeError: can't pickle SwigPyObject objects Exception ignored in: > Traceback (most recent call last):   File\n  \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/c_api_util.py\",\n  line 52, in __del__\n      c_api.TF_DeleteGraph(self.graph) AttributeError: 'ScopedTFGraph' object has no attribute 'graph'<\/p>\n<\/blockquote>\n\n<p>_2. <\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"mnist_flat_imac.py\", line\n  63, in \n      grid_result = validator.fit(train_images, train_labels)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/model_selection\/_search.py\",\n  line 626, in fit\n      base_estimator = clone(self.estimator)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 62, in clone\n      new_object_params[name] = clone(param, safe=False)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 53, in clone\n      return copy.deepcopy(estimator)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 182,\n  in deepcopy\n      y = _reconstruct(x, rv, 1, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 297,\n  in _reconstruct<\/p>\n  \n  <p><strong><em>snipped here<\/em><\/strong><\/p>\n  \n  <p>in deepcopy\n      y = _reconstruct(x, rv, 1, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 297,\n  in _reconstruct\n      state = deepcopy(state, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 155,\n  in deepcopy\n      y = copier(x, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 243,\n  in _deepcopy_dict\n      y[deepcopy(key, memo)] = deepcopy(value, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 174,\n  in deepcopy\n      rv = reductor(4) TypeError: can't pickle SwigPyObject objects<\/p>\n<\/blockquote>\n\n<p>_3. <\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"mnist_flat_imac.py\", line\n  63, in \n      grid_result = validator.fit(train_images, train_labels)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/model_selection\/_search.py\",\n  line 626, in fit\n      base_estimator = clone(self.estimator)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 62, in clone\n      new_object_params[name] = clone(param, safe=False)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/site-packages\/sklearn\/base.py\",\n  line 53, in clone<\/p>\n  \n  <p><strong><em>snipped here<\/em><\/strong><\/p>\n  \n  <p>in _deepcopy_dict\n      y[deepcopy(key, memo)] = deepcopy(value, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 182,\n  in deepcopy\n      y = _reconstruct(x, rv, 1, memo)   File \"\/home\/longnv\/PYTHON_ENV\/DataScience\/lib\/python3.5\/copy.py\", line 306,\n  in _reconstruct\n      y.<strong>dict<\/strong>.update(state) AttributeError: 'NoneType' object has no attribute 'update'<\/p>\n<\/blockquote>\n\n<p>Why did it output different errors with the same python script? \nAnd how can I fix this?<\/p>\n\n<p>Thank you so much!<\/p>\n\n<p>P.S.<\/p>\n\n<ul>\n<li>python: 3.5 <\/li>\n<li>tensorflow: 1.10.1<\/li>\n<li>pandas: 0.23.4<\/li>\n<li>Ubuntu: 4.4.0-124-generic<\/li>\n<\/ul>\n","answers":[{"AnswerId":"52752228","CreationDate":"2018-10-11T04:28:21.250","ParentId":null,"OwnerUserId":"1511274","Title":null,"Body":"<p>Found it.<\/p>\n\n<p>It should be:\nclf = KerasClassifier(build_fn=get_model)<\/p>\n\n<p>Instead of:\nclf = KerasClassifier(build_fn=get_model())<\/p>\n"}]}
+{"QuestionId":52752456,"AnswerCount":0,"Tags":"<windows><caffe>","CreationDate":"2018-10-11T04:52:33.580","AcceptedAnswerId":null,"OwnerUserId":5561494.0,"Title":"caffe installation on windows","Body":"<p>I'm trying to install caffe on windows 10, with GTX1070, CUDA10.0, cudnn7.3.1, and VS2017<\/p>\n\n<p>Following this guide\n<a href=\"https:\/\/binvec.me\/archives\/build_caffe_with_vs2017_in_windows\" rel=\"nofollow noreferrer\">https:\/\/binvec.me\/archives\/build_caffe_with_vs2017_in_windows<\/a><\/p>\n\n<p>I have modified all the configurations<\/p>\n\n<p>But when I try to run <code>build_win.cmd<\/code>\uff0c I have got a CMAKE ERROR<\/p>\n\n<pre><code>[vcvarsall.bat] Environment initialized for: 'x64'\n-- The C compiler identification is MSVC 19.15.26730.0\n-- The CXX compiler identification is MSVC 19.15.26730.0\n-- Check for working C compiler: C:\/Program Files (x86)\/Microsoft Visual Studio\/2017\/Community\/VC\/Tools\/MSVC\/14.15.26726\/bin\/Hostx86\/x64\/cl.exe\n-- Check for working C compiler: C:\/Program Files (x86)\/Microsoft Visual Studio\/2017\/Community\/VC\/Tools\/MSVC\/14.15.26726\/bin\/Hostx86\/x64\/cl.exe -- works\n-- Detecting C compiler ABI info\n-- Detecting C compiler ABI info - done\n-- Check for working CXX compiler: C:\/Program Files (x86)\/Microsoft Visual Studio\/2017\/Community\/VC\/Tools\/MSVC\/14.15.26726\/bin\/Hostx86\/x64\/cl.exe\n-- Check for working CXX compiler: C:\/Program Files (x86)\/Microsoft Visual Studio\/2017\/Community\/VC\/Tools\/MSVC\/14.15.26726\/bin\/Hostx86\/x64\/cl.exe -- works\n-- Detecting CXX compiler ABI info\n-- Detecting CXX compiler ABI info - done\n-- Detecting CXX compile features\n-- Detecting CXX compile features - done\n-- Found PythonInterp: D:\/Users\/Ezio\/Anaconda2\/python.exe (found suitable version \"2.7.15\", minimum required is \"2.7\")\n-- Looking for pthread.h\n-- Looking for pthread.h - not found\n-- Found Threads: TRUE\n-- Boost version: 1.61.0\n-- Found the following Boost libraries:\n--   system\n--   thread\n--   filesystem\n--   chrono\n--   date_time\n--   atomic\n-- Found GFlags: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found gflags  (include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include, library: gflags_shared)\n-- Found Glog: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found glog    (include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include, library: glog)\n-- Found Protobuf: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/bin\/protoc.exe (found version \"3.1.0\")\n-- Found PROTOBUF Compiler: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/bin\/protoc.exe\n-- Found LMDB: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found lmdb    (include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include, library: lmdb)\n-- Found LevelDB: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found LevelDB (include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include, library: leveldb)\n-- Found ZLIB: optimized;C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/lib\/caffezlib.lib;debug;C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/lib\/caffezlibd.lib (found version \"1.2.8\")\n-- Found Snappy: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found Snappy  (include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include, library: snappy_static;optimized;C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/lib\/caffezlib.lib;debug;C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/lib\/caffezlibd.lib)\n-- CUDA detected: 10.0\n-- Found cuDNN: ver. 7.3.1 found (include: C:\/Program Files\/NVIDIA GPU Computing Toolkit\/CUDA\/v10.0\/include, library: C:\/Program Files\/NVIDIA GPU Computing Toolkit\/CUDA\/v10.0\/lib\/x64\/cudnn.lib)\n-- Added CUDA NVCC flags for: sm_61\n-- OpenCV found (C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries)\n-- Found OpenBLAS libraries: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/lib\/libopenblas.dll.a\n-- Found OpenBLAS include: C:\/Users\/Ezio\/.caffe\/dependencies\/libraries_v140_x64_py27_1.1.0\/libraries\/include\n-- Found PythonLibs: D:\/Users\/Ezio\/Anaconda2\/libs\/python27.lib (found suitable version \"2.7.15\", minimum required is \"2.7\")\n-- Found NumPy: D:\/Users\/Ezio\/Anaconda2\/lib\/site-packages\/numpy\/core\/include (found suitable version \"1.15.1\", minimum required is \"1.7.1\")\n-- NumPy ver. 1.15.1 found (include: D:\/Users\/Ezio\/Anaconda2\/lib\/site-packages\/numpy\/core\/include)\n-- Boost version: 1.61.0\n-- Found the following Boost libraries:\n--   python\n-- Found Git: D:\/Program Files\/Git\/cmd\/git.exe (found version \"2.19.0.windows.1\")\n--\n-- ******************* Caffe Configuration Summary *******************\n-- General:\n--   Version           :   1.0.0\n--   Git               :   1.0-111-g509dae05-dirty\n--   System            :   Windows\n--   C++ compiler      :   C:\/Program Files (x86)\/Microsoft Visual Studio\/2017\/Community\/VC\/Tools\/MSVC\/14.15.26726\/bin\/Hostx86\/x64\/cl.exe\n--   Release CXX flags :   \/MD \/O2 \/Ob2 \/DNDEBUG \/DWIN32 \/D_WINDOWS \/W3 \/GR \/EHsc\n--   Debug CXX flags   :   \/MDd \/Zi \/Ob0 \/Od \/RTC1 \/DWIN32 \/D_WINDOWS \/W3 \/GR \/EHsc\n--   Build type        :   Release\n--\n--   BUILD_SHARED_LIBS :   0\n--   BUILD_python      :   1\n--   BUILD_matlab      :   0\n--   BUILD_docs        :\n--   CPU_ONLY          :   0\n--   USE_OPENCV        :   ON\n--   USE_LEVELDB       :   ON\n--   USE_LMDB          :   ON\n--   USE_NCCL          :   0\n--   ALLOW_LMDB_NOLOCK :   OFF\n--\n-- Dependencies:\n--   BLAS              :   Yes (Open)\n--   Boost             :   Yes (ver. 1.61)\n--   glog              :   Yes\n--   gflags            :   Yes\n--   protobuf          :   Yes (ver. 3.1.0)\n--   lmdb              :   Yes (ver. 0.9.70)\n--   LevelDB           :   Yes (ver. 1.18)\n--   Snappy            :   Yes (ver. 1.1.1)\n--   OpenCV            :   Yes (ver. 3.1.0)\n--   CUDA              :   Yes (ver. 10.0)\n--\n-- NVIDIA CUDA:\n--   Target GPU(s)     :   Auto\n--   GPU arch(s)       :   sm_61\n--   cuDNN             :   Yes (ver. 7.3.1)\n--\n-- Python:\n--   Interpreter       :   D:\/Users\/Ezio\/Anaconda2\/python.exe (ver. 2.7.15)\n--   Libraries         :   D:\/Users\/Ezio\/Anaconda2\/libs\/python27.lib (ver 2.7.15)\n--   NumPy             :   D:\/Users\/Ezio\/Anaconda2\/lib\/site-packages\/numpy\/core\/include (ver 1.15.1)\n--\n-- Install:\n--   Install path      :   E:\/caffe\/scripts\/build\/install\n--\nCMake Error: The following variables are used in this project, but they are set to NOTFOUND.\nPlease set them or make sure they are set and tested correctly in the CMake files:\nCUDA_cublas_device_LIBRARY (ADVANCED)\n    linked by target \"caffe\" in directory E:\/caffe\/src\/caffe\n\n-- Configuring incomplete, errors occurred!\nSee also \"E:\/caffe\/scripts\/build\/CMakeFiles\/CMakeOutput.log\".\nSee also \"E:\/caffe\/scripts\/build\/CMakeFiles\/CMakeError.log\".\nERROR: Configure failed\n<\/code><\/pre>\n\n<p>While errorlog tells me only \"missing <code>pthread.h<\/code>\" <\/p>\n\n<p>I'm sure that all dependencies have been downloaded and can be find by the script<\/p>\n\n<p>Can you help me\uff1f<\/p>\n","answers":[]}
+{"QuestionId":52752836,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><scikit-learn><k-means>","CreationDate":"2018-10-11T05:30:31.190","AcceptedAnswerId":null,"OwnerUserId":10487710.0,"Title":"Organizing Clusters in K-means clustering","Body":"<p>I am using python for k-means clustering for Mnist database(<a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a>). I am able to successfully cluster the data but unable to label the clusters. Meaning, I am unable to see that what cluster number holds  what digit. For example cluster 5 can hold digit 7. <\/p>\n\n<p>I need to write a code to correctly label the clusters after the k-means clustering has been done. Also need to add a legend to the code. \n<a href=\"https:\/\/i.stack.imgur.com\/HBHaa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HBHaa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>from __future__ import division, print_function, absolute_import\n\nimport tensorflow as tf \nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom mpl_toolkits.mplot3d import Axes3D  #only needed for 3D plots \n#scikit learn\nfrom sklearn.cluster import KMeans\n\n#pandas to read excel file\nimport pandas\nimport xlrd\n# Import MNIST data\nfrom tensorflow.examples.tutorials.mnist import input_data\n\n\nLinks:\n[MNIST Dataset] http:\/\/yann.lecun.com\/exdb\/mnist\/\n\n\n\n\ndf = pandas.read_csv('test_encoded_with_label.csv',header=None, \ndelim_whitespace=True)\n#df = pandas.read_excel('test_encoded_with_label.xls')\n#print column names\nprint(df.columns)\n\n\ndf1 = df.iloc[:,0:2] #0 and 1, the last index is not used for iloc\nlabels = df.iloc[:,2]\nlabels = labels.values\n\n\ndataset = df1.values\n#train indices - depends how many samples\ntrainidx = np.arange(0,9999)\ntestidx = np.arange(0,9999)\ntrain_data = dataset[trainidx,:]\ntest_data = dataset[testidx,:]\ntrain_labels = labels[trainidx] #just 1D, no :\ntpredct_labels = labels[testidx]\n\n\nkmeans = KMeans(n_clusters=10, random_state=0).fit(train_data)\nkmeans.labels_ \n#print(kmeans.labels_.shape)\n\n\nplt.scatter(train_data[:,0],train_data[:,1], c=kmeans.labels_)\n\n\npredct_labels = kmeans.predict(train_data)\n\n\nprint(predct_labels)\nprint('actual label', tpredct_labels)\n\n\ncenters = kmeans.cluster_centers_\nprint(centers)\n\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52770392","CreationDate":"2018-10-11T23:56:59.167","ParentId":null,"OwnerUserId":"9167289","Title":null,"Body":"<p>To add the legend, <a href=\"https:\/\/matplotlib.org\/gallery\/lines_bars_and_markers\/scatter_with_legend.html\" rel=\"nofollow noreferrer\">try<\/a>: <code>\nplt.scatter(train_data[:,0], train_data[:,1], c=kmeans.labels_, label=kmeans.labels_)\nplt.legend()<\/code><\/p>\n"},{"AnswerId":"52770776","CreationDate":"2018-10-12T00:59:58.680","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>To create markers to find cluster of labelled points, you can use the <a href=\"https:\/\/matplotlib.org\/3.0.0\/tutorials\/text\/annotations.html\" rel=\"nofollow noreferrer\">annotate<\/a> method<\/p>\n\n<p>Here is a sample code run on sklearn digits dataset where I try to mark the centroids of the resultant clustering. Note that I just label the clusters from 0-9 just for illustrative purpose:<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\n\nfrom sklearn.cluster import KMeans\nfrom sklearn.datasets import load_digits\nfrom sklearn.decomposition import PCA\nfrom sklearn.preprocessing import scale\n\nnp.random.seed(42)\n\ndigits = load_digits()\ndata = scale(digits.data)\n\nn_samples, n_features = data.shape\nn_digits = len(np.unique(digits.target))\nlabels = digits.target\nh = .02\nreduced_data = PCA(n_components=2).fit_transform(data)\nkmeans = KMeans(init='k-means++', n_clusters=n_digits, n_init=10)\nkmeans.fit(reduced_data)\n\ncentroids = kmeans.cluster_centers_\nplt_data = plt.scatter(reduced_data[:, 0], reduced_data[:, 1], c=kmeans.labels_, cmap=plt.cm.get_cmap('Spectral', 10))\nplt.colorbar()\nplt.scatter(centroids[:, 0], centroids[:, 1],\n            marker='x')\nplt.title('K-means clustering on the digits dataset (PCA-reduced data)\\n'\n          'Centroids are marked with white cross')\nplt.xlabel('component 1')\nplt.ylabel('component 2')\nlabels = ['{0}'.format(i) for i in range(10)]\nfor i in range (10):\n    xy=(centroids[i, 0],centroids[i, 1])\n    plt.annotate(labels[i],xy, horizontalalignment='right', verticalalignment='top')\nplt.show()\n<\/code><\/pre>\n\n<p>This is the result you get:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/stvJo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/stvJo.png\" alt=\"Result\"><\/a><\/p>\n"}]}
+{"QuestionId":52753370,"AnswerCount":1,"Tags":"<python><keras><lstm>","CreationDate":"2018-10-11T06:12:31.590","AcceptedAnswerId":null,"OwnerUserId":1210178.0,"Title":"Keras Sequence to Sequence Prediction, issues with shapes","Body":"<p>I'm trying to do sequence to sequence learning with Keras. My data looks like the following. <\/p>\n\n<pre><code>h h h l l l   i r\nh h l l l l   i r\nh l l l l l   i r\nh m h m h h   c u\n...\n<\/code><\/pre>\n\n<p>What I've done is to one-hot encode the <code>X<\/code>'s and <code>y<\/code>'s so my data gets transformed as follows<\/p>\n\n<pre><code>[[1. 0. 0. 1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 1. 0.]\n [1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 1. 0. 0. 1. 0.]\n ... \n [0. 1. 0. 0. 1. 0. 0. 1. 0. 1. 0. 0. 1. 0. 0. 0. 1. 0.]\n [1. 0. 0. 0. 1. 0. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 1. 0.]]\n<\/code><\/pre>\n\n<p>and my <code>y<\/code>'s look like<\/p>\n\n<pre><code>[[1. 0. 0. 0. 0. 1. 0. 0.]\n [1. 0. 0. 0. 0. 1. 0. 0.]\n...\n [0. 0. 0. 1. 0. 0. 1. 0.]\n [1. 0. 0. 0. 0. 1. 0. 0.]]\n<\/code><\/pre>\n\n<p>The function that does the one-hot encoding for the training set is the following<\/p>\n\n<pre><code>def one_hot_encode_xy(seq_x, seq_y, n_uniq_x, n_uniq_y):\n encoding_x = list()\n encoding_y = list()\n nsize_x = len(n_uniq_x)\n nsize_y = len(n_uniq_y)\n for val in seq_x:\n     vector = [0 for _ in xrange(nsize_x)]\n     index = n_uniq_x.index(val)\n     vector[index] = 1\n     encoding_x += vector\n\n for val in seq_y:\n     vector = [0 for _ in xrange(nsize_y)]\n     index = n_uniq_y.index(val)\n     vector[index] = 1\n     encoding_y += vector\n X0 = array(encoding_x)\n y0 = array(encoding_y)\n return X0, y0\n<\/code><\/pre>\n\n<p>The part that I'm stuck on is I keep getting shape incompatibility errors when I try to fit the model <code>X<\/code> and <code>y<\/code>s. The model I'm using is the following<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(150, input_shape=(1, 18)))\nmodel.add(RepeatVector(1))\nmodel.add(LSTM(150, return_sequences=True))\nmodel.add(TimeDistributed(Dense(18, activation='softmax')))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])\n\nmodel.fit(X, y, epochs=100, verbose=2)\n<\/code><\/pre>\n\n<p>How should I reshape my <code>X<\/code>'s <code>y<\/code>'s and\/or the model so that I can do a simple seq2seq learning on this? Can someone help out please? Any help appreciated.<\/p>\n\n<p>An example of the error I'm getting is the following<\/p>\n\n<blockquote>\n<pre><code>model.fit(X, y, epochs=100, verbose=2)\n<\/code><\/pre>\n  \n  <p>File \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training.py\", line 958, in fit\n     File \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training.py\", line 757, in _standardize_user_data\n    File \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training_utils.py\", line 128, in standardize_input_data\n  ValueError: Error when checking input: expected lstm_1_input to have 3 dimensions, but got array with shape (18, 18)<\/p>\n<\/blockquote>\n\n<p>This is how I initialize my <code>X<\/code>, <code>y<\/code> and <code>X_test<\/code>. I've a toy data set of 18 rows (as shown above) and want to test it on 2 cases.<\/p>\n\n<pre><code>X = np.zeros((18, 18), dtype='float32')\ny = np.zeros((18, 18), dtype='float32')\nX_test = np.zeros((2, 18), dtype='float32')\n<\/code><\/pre>\n","answers":[{"AnswerId":"52764017","CreationDate":"2018-10-11T15:39:59.503","ParentId":null,"OwnerUserId":"8208006","Title":null,"Body":"<p>the Train input shape should be 3 dimensions (Samples,Timesteps,Features)<\/p>\n\n<pre><code>**Samples**. One sequence is one sample. A batch is comprised of one or more samples.\n**Time Steps**. One time step is one point of observation in the sample.\n**Features**. One feature is one observation at a time step.\n<\/code><\/pre>\n\n<p>In your case the input data , if you dont have many features just add one more dimension to input data or you can use look back to create short term variations of sequence, it will work<\/p>\n\n<pre><code>X = X.reshape(X.shape[0],X.shape[1],1)\ny = y.reshape(y.shape[0],1,y.shape[1])\n<\/code><\/pre>\n\n<p>While giving input shape give X[0].shape as input\n.   <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">example<\/a><\/p>\n"}]}
+{"QuestionId":52753780,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><keras><lstm><prediction>","CreationDate":"2018-10-11T06:39:27.040","AcceptedAnswerId":null,"OwnerUserId":10487990.0,"Title":"Keras.prediction() is getting worse with more prediction","Body":"<p>I am new using Keras ans I was experimenting with it.\nI have built the following network, as I am trying to understand the \"stateful\" option in Keras LSTM. :<\/p>\n\n<pre><code>rdrop = 0.3\nrdrop2 = 0\nmodel = Sequential()\nmodel.add(LSTM(1600, return_sequences=False, batch_input_shape=(n_win, L, N), stateful=True, dropout=rdrop))\nmodel.add(Dense(1600, activation='linear'))\nmodel.add(Dropout(0.4))\nmodel.add(RepeatVector(L))\nmodel.add(LSTM(1600, return_sequences=True, stateful=False, dropout=rdrop))\nmodel.add(TimeDistributed(Dense(N, activation='softmax')))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])\n<\/code><\/pre>\n\n<p>with these fit() options :<\/p>\n\n<pre><code>for i in range(30):\n   model.fit(data_input3, data_input3,\n                    shuffle=False,\n                    epochs=1,\n                    batch_size=n_win))\n   model.reset_states()\n<\/code><\/pre>\n\n<p>I have a result which is not bad, but I observed that more I predict() worst is the result using the same input.<\/p>\n\n<p>For example, all the results of the code below are different and the more it runs, the worst it gets.<\/p>\n\n<pre><code>t = np.zeros(100)\nfor i in range(100):\ndata_out2 = model.predict(data_input3[0:n_win, :, :], batch_size=n_win)\nt = np.sum(np.abs(data_out2 - data_input3[0:n_win, :, :]))\n<\/code><\/pre>\n\n<p>It is not the good way to compute the distance but when I am visually checking the results, after calling twice <code>model.predict()<\/code> it starts to be awefull (the first time is nice).<\/p>\n\n<p>Is there any change of the weights during the prediction phase ? I have checked the dropout and it seems to be used just in the training.\nI have tested with Keras 2.2.3 and 2.2.4, both running using TF 1.10<\/p>\n\n<p>In advance, thank you for your help!<\/p>\n","answers":[]}
+{"QuestionId":52753846,"AnswerCount":3,"Tags":"<keras><hdf5><tensorflow-lite>","CreationDate":"2018-10-11T06:43:59.267","AcceptedAnswerId":53048893.0,"OwnerUserId":10411044.0,"Title":"Error in load a model saved by callbakcs.ModelCheckpoint() in Keras","Body":"<p>I saved my model automatically by <code>callbacks.ModelCheckpoint()<\/code> with a HDF5 file.<\/p>\n\n<pre><code># Checkpoint In the \/output folder\nfilepath = \".\/model\/mnist-cnn-best.hd5\"\n\n# Keep only a single checkpoint, the best over test accuracy.\ncheckpoint = keras.callbacks.ModelCheckpoint(filepath, monitor='val_acc', \n                                             verbose=1, save_best_only=True,\n                                             mode='max')\n\n# Train\nmodel.fit(x_train, y_train,\n          batch_size=batch_size,\n          epochs=epochs,\n          verbose=1,\n          validation_data=(x_test, y_test),\n          callbacks=[checkpoint])\n<\/code><\/pre>\n\n<p>When I load a model, an error occured.<\/p>\n\n<pre><code>  model = keras.models.load_model(\".\/mnist-cnn-best.hd5\")\n\n  File \"D:\\Program Files\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\saving.py\", line 251, in load_model\n    training_config['weighted_metrics'])\nKeyError: 'weighted_metrics'\n<\/code><\/pre>\n\n<p>If I load model with param '<em>compile=False<\/em>', it works correctly.<\/p>\n\n<p>I know the normal way to save model in keras is:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'\ndel model  # deletes the existing model\n\n# returns a compiled model\n# identical to the previous one\nmodel = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>By the way, this error also happened when me convert this model by Tensorflow Lite.\nBut I don't know what's wrong with my model.\nDoes anyone has an idea?<\/p>\n","answers":[{"AnswerId":"55907947","CreationDate":"2019-04-29T17:09:12.047","ParentId":null,"OwnerUserId":"10896697","Title":null,"Body":"<p>The way I usually use this is as follows:<\/p>\n\n<pre><code>def create_model():\n    &lt;my model&gt;\n    &lt;model.compile&gt;\n    return model\ncheckpoint = keras.callbacks.ModelCheckpoint(filepath, verbose=&lt;val&gt;, monitor=&lt;val&gt;, save_best_only=True, save_weights_only=True)\n\nclassifier = create_model()\nclassifier.fit(&lt;your parameters&gt;)\nclassifier.evaluate(&lt;your parameters&gt;)\n\nloaded_model = create_model()\nloaded_model.load_weights(filepath)\ny_pred = loaded.model.&lt;predict_method&gt;(test_set,verbose=&lt;val&gt;)\n'''\n<\/code><\/pre>\n"},{"AnswerId":"54140305","CreationDate":"2019-01-11T04:15:52.463","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>If you haven't figured out the answer to this yet, I think I've got it.<\/p>\n\n<p>I haven't dug into the code to exactly figure out why, but basically the model checkpoint callback can only be loaded with the <code>load_weights()<\/code> function, which is then used for evaluation. <\/p>\n\n<p>If you want to save a model that you can load to train again later you need to use <code>model.save<\/code> and <code>model.load_model<\/code>. Hopefully helpful to someone who wanders upon this.<\/p>\n"},{"AnswerId":"53048893","CreationDate":"2018-10-29T15:33:23.207","ParentId":null,"OwnerUserId":"5952028","Title":null,"Body":"<p>I hit a similar problem that yields the same error message, but the cause might be different than yours: <\/p>\n\n<p><strong>Code:<\/strong> (Tensorflow 1.11 and tf.keras.<strong>version<\/strong>: 2.1.6-tf)<\/p>\n\n<pre><code> if load_model_path.endswith('.h5'):\n        model = tf.keras.models.load_model(load_model_path)\n<\/code><\/pre>\n\n<p><strong>Error message:<\/strong><\/p>\n\n<pre><code>  File \".....\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 251, in load_model\n    training_config['weighted_metrics'])\nKeyError: 'weighted_metrics'\n<\/code><\/pre>\n\n<p>And I found out it's because the model was saved in an older Keras version.\nI had to comment out the code related to <code>weighted_metrics<\/code> to be able to load the model. However, it's just a workaround before I can find a sustainable solution to the mismatching problem. Interestingly, <code>@fchollet<\/code> just added <code>weighted_metrics<\/code> to the latest Keras version recently (Oct 2018).<br>\n<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/saving.py#L136\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/saving.py#L136<\/a>\nI hope this will help the people who hit the same problem as I did.<\/p>\n"}]}
+{"QuestionId":52754453,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2018-10-11T07:20:39.910","AcceptedAnswerId":52772339.0,"OwnerUserId":1935883.0,"Title":"Tensorflow Keras - AttributeError: Layer features has no inbound nodes","Body":"<p>Tensorflow version : 1.11.0<\/p>\n\n<p>I am trying to use TensorBoard with Tensorflow keras model for projector visualisation. \nI am getting AttributeError: Layer features has no inbound nodes.\nI am not sure why I get this error in below simple code. I indeed google the error but I could not find right solution to fix it. <\/p>\n\n<pre><code>from os import makedirs\nfrom os.path import exists, join\nimport tensorflow as tf\nmnist = tf.keras.datasets.mnist\n\nimport numpy as np\n\n\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(512, activation=tf.nn.relu),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(10, activation=tf.nn.relu, name='features'),\n  tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nlog_dir = \".\/logs\"\nwith open(join(log_dir, 'metadata.tsv'), 'w') as f:\n    np.savetxt(f, y_test)\n\nfrom tensorflow.keras.callbacks import TensorBoard\n\ntf_board_callback = TensorBoard(\n                            log_dir=log_dir,\n                            batch_size=32,\n                            embeddings_freq=1,\n                            embeddings_layer_names=['features'],\n                            embeddings_metadata='metadata.tsv',\n                            embeddings_data=x_test\n                        )\n\nmodel.fit(x_train, y_train, epochs=5, callbacks=[tf_board_callback])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52772339","CreationDate":"2018-10-12T04:35:01.237","ParentId":null,"OwnerUserId":"1454135","Title":null,"Body":"<p>When defining a network in Keras, the first layer added needs to have input_shape added.<\/p>\n\n<p>See the docs here: <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/#specifying-the-input-shape\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/sequential-model-guide\/#specifying-the-input-shape<\/a><\/p>\n\n<p>So for MNIST, you should have something like input_shape=(28,28,1)<\/p>\n\n<p>There's a nice example here: <a href=\"https:\/\/www.kaggle.com\/adityaecdrid\/mnist-with-keras-for-beginners-99457\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/adityaecdrid\/mnist-with-keras-for-beginners-99457<\/a><\/p>\n"},{"AnswerId":"52755710","CreationDate":"2018-10-11T08:35:22.287","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>I guess you should specify input shape for the first layer of the sequential model<\/p>\n"}]}
+{"QuestionId":52754463,"AnswerCount":1,"Tags":"<python><keras><deep-learning><recurrent-neural-network>","CreationDate":"2018-10-11T07:21:18.513","AcceptedAnswerId":52758162.0,"OwnerUserId":4025749.0,"Title":"batch_size must be divisible by strategy.num_towers (17 vs 8) error when using Google Colab TPU in keras cross validation k-fold","Body":"<p>I wanna run my Keras code on TPU to classify text. when my model wants to evaluate Val_acc, this error is shown:<\/p>\n\n<pre><code>batch_size must be divisible by strategy.num_towers (17 vs 8)\n<\/code><\/pre>\n\n<p>this is my code :<\/p>\n\n<pre><code>from __future__ import unicode_literals\nimport pandas as pd\nimport openpyxl \nfrom hazm import *\nimport tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import LabelEncoder\nfrom sklearn.feature_extraction.text import TfidfVectorizer\nfrom keras.preprocessing.text import one_hot\nfrom keras.preprocessing.sequence import pad_sequences\nfrom tensorflow.keras import layers,models\n\nfrom keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing import sequence\nfrom keras import optimizers\nfrom keras.callbacks import *\nfrom keras.utils import to_categorical\nfrom keras.callbacks import EarlyStopping\nfrom sklearn.model_selection import StratifiedKFold\n%matplotlib inline\n<\/code><\/pre>\n\n<p>my Data:<\/p>\n\n<pre><code>X: texts\nY: labels\n<\/code><\/pre>\n\n<p>k-fold:<\/p>\n\n<pre><code>kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=7)\n<\/code><\/pre>\n\n<p>training step:<\/p>\n\n<pre><code>epochs = 10\nbatch_size = 64\nscores=[]\nfor train, test in kfold.split(X, Y):\n     model = RNN2()\n\n     tpu_model = tf.contrib.tpu.keras_to_tpu_model(\n                 model,\n                 strategy=tf.contrib.tpu.TPUDistributionStrategy(\n                tf.contrib.cluster_resolver.TPUClusterResolver(tpu='grpc:\/\/' \n                + os.environ['COLAB_TPU_ADDR'])\n                   )\n                 )\n     tpu_model.compile(\n                optimizer='adam',\n                loss='binary_crossentropy',\n                metrics=['accuracy']\n               )\n\n     tpu_model.fit(X_sequences_matrix[train],Y[train],\n                   batch_size=batch_size,\n                   epochs=epochs,validation_split=0.15)\n<\/code><\/pre>\n\n<p>these models log in the training step:<\/p>\n\n<pre><code>Epoch 1\/10\nINFO:tensorflow:New input shapes; (re-)compiling: mode=train, [TensorSpec(shape=(4, 920), dtype=tf.float32, name='input_110'), TensorSpec(shape=(4, 1), dtype=tf.float32, name='activation_34_target0')]\nINFO:tensorflow:Overriding default placeholder.\nINFO:tensorflow:Remapping placeholder for input\nINFO:tensorflow:Cloning Adam {'lr': 0.0010000000474974513, 'beta_1': 0.8999999761581421, 'beta_2': 0.9990000128746033, 'decay': 0.0, 'epsilon': 1e-07, 'amsgrad': False}\nINFO:tensorflow:Get updates: Tensor(\"loss_3\/mul:0\", shape=(), dtype=float32)\nINFO:tensorflow:Started compiling\nINFO:tensorflow:Finished compiling. Time elapsed: 11.963297128677368 secs\nINFO:tensorflow:Setting weights on TPU model.\n3648\/3697 [============================&gt;.] - ETA: 1s - loss: 0.3725 - acc: 0.8728 \nAssertionError: batch_size must be divisible by strategy.num_towers (17 vs 8)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52758162","CreationDate":"2018-10-11T10:47:03.650","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>The problem is that you have 8 TPU cores and <code>batch_size<\/code> 17. As shown <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/100f9dee3115d892be1a39ae023bd80e24bf70eb\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py#L636\" rel=\"nofollow noreferrer\">here<\/a> your <code>batch_size<\/code> must be divisible with number of TPU cores. <\/p>\n"}]}
+{"QuestionId":52754492,"AnswerCount":1,"Tags":"<python><keras><generator>","CreationDate":"2018-10-11T07:22:56.837","AcceptedAnswerId":null,"OwnerUserId":8352482.0,"Title":"Write custom Data Generator for Keras","Body":"<p>I have each datapoint stored in a .npy file, with <code>shape=(1024,7,8)<\/code>. I want to load them to a Keras model by a manner similar to <code>ImageDataGenerator<\/code>, so I wrote and tried different custom generators but none of them work, here is one I adapted from <a href=\"https:\/\/stackoverflow.com\/questions\/46493419\/use-a-generator-for-keras-model-fit-generator\">this<\/a><\/p>\n\n<pre><code>def find(dirpath, prefix=None, suffix=None, recursive=True):\n    \"\"\"Function to find recursively all files with specific prefix and suffix in a directory\n    Return a list of paths\n    \"\"\"\n    l = []\n    if not prefix:\n        prefix = ''\n    if not suffix:\n        suffix = ''\n    for (folders, subfolders, files) in os.walk(dirpath):\n        for filename in [f for f in files if f.startswith(prefix) and f.endswith(suffix)]:\n            l.append(os.path.join(folders, filename))\n        if not recursive:\n            break\n    l\n    return l\n\ndef generate_data(directory, batch_size):\n    i = 0\n    file_list = find(directory)\n    while True:\n        array_batch = []\n        for b in range(batch_size):\n            if i == len(file_list):\n                i = 0\n                random.shuffle(file_list)\n            sample = file_list[i]\n            i += 1\n\n            array = np.load(sample)\n            array_batch.append(array)\n\n        yield array_batch\n<\/code><\/pre>\n\n<p>I found this lacks of the label, so it won't be fit into the model using <code>fit_generator<\/code> . How can I add the label into this generator, given that I can store them in a numpy array?<\/p>\n","answers":[{"AnswerId":"52755399","CreationDate":"2018-10-11T08:19:27.063","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<pre><code>from tensorflow.python.keras.utils import Sequence\nimport numpy as np   \n\nclass Mygenerator(Sequence):\n    def __init__(self, x_set, y_set, batch_size):\n        self.x, self.y = x_set, y_set\n        self.batch_size = batch_size\n\n    def __len__(self):\n        return int(np.ceil(len(self.x) \/ float(self.batch_size)))\n\n    def __getitem__(self, idx):\n        batch_x = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]\n        batch_y = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]\n\n        # read your data here using the batch lists, batch_x and batch_y\n        x = [my_readfunction(filename) for filename in batch_x] \n        y = [my_readfunction(filename) for filename in batch_y]\n        return np.array(x), np.array(y)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52755805,"AnswerCount":1,"Tags":"<python><pytorch><autograd>","CreationDate":"2018-10-11T08:40:56.280","AcceptedAnswerId":null,"OwnerUserId":7483494.0,"Title":"PyTorch - Do overwritten variables stay in the graph?","Body":"<p>I am wondering if PyTorch Tensors where the Python Variables are overwritten are still being kept in the computational graph of PyTorch. \n<hr>\n<em>So here is a small example, where I have an RNN Model where hidden states (and some other variables) are reset after every iteration, <code>backward()<\/code> is called later.<\/em><\/p>\n\n<p><strong>Example:<\/strong><\/p>\n\n<pre><code>for i in range(5):\n   output = rnn_model(inputs[i])\n   loss += criterion(output, target[i])\n   ## hidden states are overwritten with a zero vector\n   rnn_model.reset_hidden_states() \nloss.backward()\n<\/code><\/pre>\n\n<hr>\n\n<p><strong><em>So my question is:<\/em><\/strong><\/p>\n\n<ul>\n<li><p>Is there a problem in overwriting the hidden states before calling <code>backward()<\/code>? <\/p><\/li>\n<li><p>Or does the computational graph keep the necessary information of the hidden states of previous iterations in memory to compute the gradients?<\/p><\/li>\n<li><p><em>Edit: It would be great to have a statement of an official source for this. e.g. stating that all variables relevant for the CG are kept - no matter if there are still other python references for this variables. I assume that there is a reference in the graph itself preventing the garbage collector from deleting it. But I'd like to know if this is truly the case.<\/em><\/p><\/li>\n<\/ul>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"52770261","CreationDate":"2018-10-11T23:37:23.257","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>I think it is ok to reset before backward. The graph preserves the required information. <\/p>\n\n<pre><code>class A (torch.nn.Module):\n     def __init__(self):\n         super().__init__()\n         self.f1 = torch.nn.Linear(10,1)\n     def forward(self, x):\n         self.x = x \n         return torch.nn.functional.sigmoid (self.f1(self.x))\n     def reset_x (self):\n        self.x = torch.zeros(self.x.shape) \nnet = A()\nnet.zero_grad()\nX = torch.rand(10,10) \nloss = torch.nn.functional.binary_cross_entropy(net(X), torch.ones(10,1))\nloss.backward()\nparams = list(net.parameters())\nfor i in params: \n    print(i.grad)\nnet.zero_grad() \n\nloss = torch.nn.functional.binary_cross_entropy(net(X), torch.ones(10,1))\nnet.reset_x()\nprint (net.x is X)\ndel X\nloss.backward()     \nparams = list(net.parameters())\nfor i in params:\n    print(i.grad)\n<\/code><\/pre>\n\n<p>In the above code, I print the grads with\/ without resetting input x. Gradient depends on x for sure and resetting it doesn't matter. Therefore, I think graph preserves information to do the backward op.   <\/p>\n"}]}
+{"QuestionId":52756152,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-11T09:00:17.673","AcceptedAnswerId":52756780.0,"OwnerUserId":8909870.0,"Title":"Tensorboard extract scalar by a script","Body":"<p>I want to extract my scalars by a script, because I have a lot of test runs. <\/p>\n\n<p>Based on this <a href=\"https:\/\/stackoverflow.com\/questions\/47232779\/how-to-extract-and-save-images-from-tensorboard-event-summary\">answer<\/a> I can get all tf summaries of one board. I can even separate the tag for the loss: <\/p>\n\n<pre><code>&lt;class 'tensorflow.core.framework.summary_pb2.Value'&gt;\ntag: \"training_loss\"\nsimple_value: 0.0590251199901104\n<\/code><\/pre>\n\n<p>But it seems that every loss value that is saved as on <code>summary_pb2.Value<\/code>. I could extract every loss single value, but I don't find an information about the step number or time of these single values, so that I can order them (They have the same tag as well). Unfortunately this is not well documented, does someone know how I can get this information? <\/p>\n","answers":[{"AnswerId":"52756780","CreationDate":"2018-10-11T09:33:48.367","ParentId":null,"OwnerUserId":"863543","Title":null,"Body":"<p>I would use the <code>EventAccumulator<\/code>:\nYou can pass the model directory to the <code>_load_run()<\/code> function.<\/p>\n\n<pre><code>from tensorboard.backend.event_processing import event_accumulator\nimport numpy as np\n\ndef _load_run(path):\n  event_acc = event_accumulator.EventAccumulator(path)\n  event_acc.Reload()\n  data = {}\n\n  for tag in sorted(event_acc.Tags()[\"scalars\"]):\n    x, y = [], []\n\n    for scalar_event in event_acc.Scalars(tag):\n      x.append(scalar_event.step)\n      y.append(scalar_event.value)\n\n    data[tag] = (np.asarray(x), np.asarray(y))\n  return data\n\nprint(_load_run(\"\/models\/vae\/run_1\"))\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":52756283,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning><dropout>","CreationDate":"2018-10-11T09:07:25.553","AcceptedAnswerId":52760112.0,"OwnerUserId":3711985.0,"Title":"Adding Dropout to testing\/inference phase","Body":"<p>I've trained the following model for some timeseries in Keras:<\/p>\n\n<pre><code>    input_layer = Input(batch_shape=(56, 3864))\n    first_layer = Dense(24, input_dim=28, activation='relu',\n                        activity_regularizer=None,\n                        kernel_regularizer=None)(input_layer)\n    first_layer = Dropout(0.3)(first_layer)\n    second_layer = Dense(12, activation='relu')(first_layer)\n    second_layer = Dropout(0.3)(second_layer)\n    out = Dense(56)(second_layer)\n    model_1 = Model(input_layer, out)\n<\/code><\/pre>\n\n<p>Then I defined a new model with the trained layers of <code>model_1<\/code> and added dropout layers with a different rate, <code>drp<\/code>, to it:<\/p>\n\n<pre><code>    input_2 = Input(batch_shape=(56, 3864))\n    first_dense_layer = model_1.layers[1](input_2)\n    first_dropout_layer = model_1.layers[2](first_dense_layer)\n    new_dropout = Dropout(drp)(first_dropout_layer)\n    snd_dense_layer = model_1.layers[3](new_dropout)\n    snd_dropout_layer = model_1.layers[4](snd_dense_layer)\n    new_dropout_2 = Dropout(drp)(snd_dropout_layer)\n    output = model_1.layers[5](new_dropout_2)\n    model_2 = Model(input_2, output)\n<\/code><\/pre>\n\n<p>Then I'm getting the prediction results of these two models as follow:<\/p>\n\n<pre><code>result_1 = model_1.predict(test_data, batch_size=56)\nresult_2 = model_2.predict(test_data, batch_size=56)\n<\/code><\/pre>\n\n<p>I was expecting to get completely different results because the second model has new dropout layers and theses two models are different (IMO), but that's not the case. Both are generating the same result. Why is that happening? <\/p>\n","answers":[{"AnswerId":"52760112","CreationDate":"2018-10-11T12:29:13.033","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As I mentioned in the comments, the <code>Dropout<\/code> layer is turned off in inference phase (i.e. test mode), so when you use <code>model.predict()<\/code> the <code>Dropout<\/code> layers are not active. However, if you would like to have a model that uses <code>Dropout<\/code> both in training and inference phase, you can pass <code>training<\/code> argument when calling it, <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9412#issuecomment-366487249\" rel=\"nofollow noreferrer\">as suggested by Fran\u00e7ois Chollet<\/a>:<\/p>\n\n<pre><code># ...\nnew_dropout = Dropout(drp)(first_dropout_layer, training=True)\n# ...\n<\/code><\/pre>\n\n<p>Alternatively, If you have already trained your model and now want to use it in inference mode and keep the <code>Dropout<\/code> layers (and possibly other layers which have different behavior in training\/inference phase such as <code>BatchNormalization<\/code>) active, you can define a backend function that takes the model's inputs as well as Keras learning phase:<\/p>\n\n<pre><code>from keras import backend as K\n\nfunc = K.function(model.inputs + [K.learning_phase()], model.outputs)\n\n# to use it pass 1 to set the learning phase to training mode\noutputs = func([input_arrays] + [1.]) \n<\/code><\/pre>\n"}]}
+{"QuestionId":52756374,"AnswerCount":1,"Tags":"<python-3.x><variables><keras><autoencoder><dimensionality-reduction>","CreationDate":"2018-10-11T09:12:11.093","AcceptedAnswerId":52769229.0,"OwnerUserId":5760497.0,"Title":"Saving an Auto-encoder trained to reduce dimensions","Body":"<p>I made an autoencoder for dimensionality reduction, I wanted to save it to be used in the reduction of the test dataset. Here is my code<\/p>\n\n<pre><code>dom_state = seed(123)\nprint('Rescaling Data')\ny = minmax_scale(X, axis=0)\nncol = y.shape[1] #here ncol = 19\nprint('Encoding Dimensions')\nencoding_dim = 3\ninput_dim = Input(shape = (ncol,))\n\nwith tf.Session(config=tf.ConfigProto(intra_op_parallelism_threads=24)) as sess:\n    K.set_session(sess)\n    print('Initiating Encoder Layer')\n    encoded1 = Dense(20, activation = 'relu')(input_dim)\n    encoded2 = Dense(10, activation = 'relu')(encoded1)\n    encoded3 = Dense(5, activation = 'relu')(encoded2)\n    encoded4 = Dense(encoding_dim, activation = 'relu')(encoded3)\n    print('Initiating Decoder Layer')\n    decoded1 = Dense(5, activation = 'relu')(encoded4)\n    decoded2 = Dense(10, activation = 'relu')(decoded1)\n    decoded3 = Dense(20, activation = 'relu')(decoded2)\n    decoded4 = Dense(ncol, activation = 'sigmoid')(decoded3)\n\n    print('Combine Encoder and Decoder layers')\n    autoencoder = Model(input = input_dim, output = decoded4)\n    print('Compiling Mode')\n    autoencoder.compile(optimizer = 'Nadam', loss ='mse')\n    autoencoder.fit(y, y, nb_epoch = 300, batch_size = 20, shuffle = True)\n    encoder = Model(input = input_dim, output = decoded4)\n    encoder.save('reduction_param.h5')\n\n    print('Initiating Dimension Reduction')\n    model = load_model('reduction_param.h5')\n    encoded_input = Input(shape = (encoding_dim, ))\n    encoded_out = model.predict(y)\n<\/code><\/pre>\n\n<p>However, even if I have restricted the dimensions, on the model.predict(y) part, I still get the full 19 columns instead of 3. Furthermore, I am also receiving the error:<\/p>\n\n<pre><code>UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.\n  warnings.warn('No training configuration found in save file:\n<\/code><\/pre>\n\n<p>which I understand, because, the <code>encoder.save('reduction_param.h5')<\/code> is actually not compiled with an optimiser. Am I missing something?<\/p>\n\n<p><strong>EDIT<\/strong>:<\/p>\n\n<p>I do not know if this is the correct way to address the issue, basically I train the MinMAXScaler() to the training dataset, saved the features as a pickle, then re-use it instead while maintaining the auto-encoder, as per code:<\/p>\n\n<pre><code>dom_state = seed(123)\nprint('Rescaling Data')\nfeature_space= MinMaxScaler()\nfeature_pkl = feature_space.fit(X)\nfilename = 'lc_feature_space.sav'\npickle.dump(feature_pkl, open(filename, 'wb'))\nloaded_model = pickle.load(open(filename, 'rb'))\ny = loaded_model.transform(X)\nncol = y.shape[1]\nprint(ncol)\nprint('Encoding Dimensions')\nencoding_dim = 3\ninput_dim = Input(shape = (ncol,))\n\nwith tf.Session(config=tf.ConfigProto(intra_op_parallelism_threads=24)) as sess:\n    K.set_session(sess)\n    print('Initiating Encoder Layer')\n    encoded1 = Dense(20, activation = 'relu')(input_dim)\n    encoded2 = Dense(10, activation = 'relu')(encoded1)\n    encoded3 = Dense(5, activation = 'relu')(encoded2)\n    encoded4 = Dense(encoding_dim, activation = 'relu')(encoded3)\n    print('Initiating Decoder Layer')\n    decoded1 = Dense(5, activation = 'relu')(encoded4)\n    decoded2 = Dense(10, activation = 'relu')(decoded1)\n    decoded3 = Dense(20, activation = 'relu')(decoded2)\n    decoded4 = Dense(ncol, activation = 'sigmoid')(decoded3)\n\n    print('Combine Encoder and Deocoder layers')\n    autoencoder = Model(input = input_dim, output = decoded4)\n    print('Compiling Mode')\n    autoencoder.compile(optimizer = 'Nadam', loss ='mse')\n    autoencoder.fit(y, y, nb_epoch = 300, batch_size = 20, shuffle = True)\n\n    print('Initiating Dimension Reduction')\n    encoder = Model(input = input_dim, output = decoded4)\n    encoded_input = Input(shape = (encoding_dim, ))\n    encoded_out = encoder.predict(y)\n    result = encoded_out[0:2]\n<\/code><\/pre>\n\n<p>my argument here is to save the features of the training dataset at the MINMAXScaler() level, transform the test dataset based on these features, then just reduce using the auto-encoder. Still I do not know if this is correct.<\/p>\n","answers":[{"AnswerId":"52769229","CreationDate":"2018-10-11T21:31:43.880","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>I think the reason why you didn't see <code>encoder<\/code> to work properly, i.e. reducing an input tensor's dimensionality, is because you defined and saved a wrong model. You should use <\/p>\n\n<pre><code>encoder = Model(input = input_dim, output = encoded4 )\n<\/code><\/pre>\n\n<p>whose output node is <code>encoded4<\/code> instead of <code>decoded4<\/code>.<\/p>\n"}]}
+{"QuestionId":52756389,"AnswerCount":0,"Tags":"<python><python-3.x><keras><google-colaboratory><pydot>","CreationDate":"2018-10-11T09:13:08.403","AcceptedAnswerId":null,"OwnerUserId":1031417.0,"Title":"How to run Kera's plot_model in a Google Colaboratory session?","Body":"<p>The following setup commands seem to be working (the notebook is configured to run Python 3):<\/p>\n\n<pre><code>!pip3 install pydotplus\n!pip3 install pydot\n!apt-get -qq install -y graphviz &amp;&amp; pip3 install -q pydot\n!pip3 install pydot-ng\n!pip3 install graphviz\n!apt-get install python-pydot python-pydot-ng graphviz \nimport pydot\n<\/code><\/pre>\n\n<p>As the output is:<\/p>\n\n<pre><code>Requirement already satisfied: pydotplus in \/usr\/local\/lib\/python3.6\/dist-packages (2.0.2)\nRequirement already satisfied: pyparsing&gt;=2.0.1 in \/usr\/local\/lib\/python3.6\/dist-packages (from pydotplus) (2.2.2)\nRequirement already satisfied: pydot in \/usr\/local\/lib\/python3.6\/dist-packages (1.2.4)\nRequirement already satisfied: pyparsing&gt;=2.1.4 in \/usr\/local\/lib\/python3.6\/dist-packages (from pydot) (2.2.2)\nRequirement already satisfied: pydot-ng in \/usr\/local\/lib\/python3.6\/dist-packages (1.0.0)\nRequirement already satisfied: pyparsing&gt;=2.0.1 in \/usr\/local\/lib\/python3.6\/dist-packages (from pydot-ng) (2.2.2)\nRequirement already satisfied: graphviz in \/usr\/local\/lib\/python3.6\/dist-packages (0.9)\nReading package lists... Done\nBuilding dependency tree       \nReading state information... Done\ngraphviz is already the newest version (2.40.1-2).\npython-pydot is already the newest version (1.2.3-1).\npython-pydot-ng is already the newest version (1.0.0-3).\n0 upgraded, 0 newly installed, 0 to remove and 4 not upgraded.\n<\/code><\/pre>\n\n<p>But when I try to plot a compiled model:<\/p>\n\n<pre><code>from keras.utils import plot_model\nplot_model(model)\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-47-4cf84b52a91d&gt; in &lt;module&gt;()\n      9 \n     10 from keras.utils import plot_model\n---&gt; 11 plot_model(model)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/utils\/vis_utils.py in plot_model(model, to_file, show_shapes, show_layer_names, rankdir)\n    132             'LR' creates a horizontal plot.\n    133     \"\"\"\n--&gt; 134     dot = model_to_dot(model, show_shapes, show_layer_names, rankdir)\n    135     _, extension = os.path.splitext(to_file)\n    136     if not extension:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/utils\/vis_utils.py in model_to_dot(model, show_shapes, show_layer_names, rankdir)\n     53     from ..models import Sequential\n     54 \n---&gt; 55     _check_pydot()\n     56     dot = pydot.Dot()\n     57     dot.set('rankdir', rankdir)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/utils\/vis_utils.py in _check_pydot()\n     18     if pydot is None:\n     19         raise ImportError(\n---&gt; 20             'Failed to import `pydot`. '\n     21             'Please install `pydot`. '\n     22             'For example with `pip install pydot`.')\n\nImportError: Failed to import `pydot`. Please install `pydot`. For example with `pip install pydot`.\n\n---------------------------------------------------------------------------\nNOTE: If your import is failing due to a missing package, you can\nmanually install dependencies using either !pip or !apt.\n\nTo view examples of installing some common dependencies, click the\n\"Open Examples\" button below.\n---------------------------------------------------------------------------\n<\/code><\/pre>\n\n<p>Not sure what is missing.<\/p>\n","answers":[]}
+{"QuestionId":52756789,"AnswerCount":0,"Tags":"<docker><tensorflow><kubernetes>","CreationDate":"2018-10-11T09:34:15.283","AcceptedAnswerId":null,"OwnerUserId":7591522.0,"Title":"docker image leads to containers with different tensorflow versions","Body":"<p>I have a Dockerfile in which I create a python environment using miniconda. This environment implicity installs a tensorflow version through specifying a rasa_core version 0.11.11. So I build an image from this Dockerfile. Now if I create a container locally with docker, I find the tensorflow version to be 1.10. However if I use the image in an deployment on kubernetes, the container crashes and the logs show that code from tensorflow version 1.11 is the cause. How is this possible?   <\/p>\n","answers":[]}
+{"QuestionId":52757411,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-11T10:06:21.953","AcceptedAnswerId":null,"OwnerUserId":7813115.0,"Title":"Tensorflow build is failing for GPU '__NV_EXTRA_INITIALIZATION' was not declared in this scope","Body":"<p>This is my first time working on a GPU. I have followed the Tensorflow installation guide <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a> and am trying to build it using bezel. But I am getting the following error<\/p>\n\n<pre><code>INFO: From Compiling external\/nccl_archive\/src\/collectives\/device\/min_all_reduce.cu.cc:\nptxas warning : Too big maxrregcount value specified 96, will be ignored\nINFO: From Compiling external\/nccl_archive\/src\/collectives\/device\/max_all_reduce.cu.cc:\nptxas warning : Too big maxrregcount value specified 96, will be ignored\nERROR: \/home\/ubuntu\/.cache\/bazel\/_bazel_ubuntu\/3be25c7517fcab2b83999fe3669ecd1f\/external\/nccl_archive\/BUILD.bazel:139:1: C++ compilation of rule '@nccl_archive\/\/:device_code' failed (Exit 1)\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc:19:2: warning: #warning \"crt\/link.stub is an internal header file and must not be used directly.  Please use cuda_runtime_api.h or cuda_runtime.h instead.\" [-Wcpp]\n #warning \"crt\/link.stub is an internal header file and must not be used directly.  Please use cuda_runtime_api.h or cuda_runtime.h instead.\"\n  ^\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc: In function 'void __cudaRegisterLinkedBinary(const __fatBinC_Wrapper_t*, void (*)(void**), void*)':\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc:140:5: error: '__NV_EXTRA_INITIALIZATION' was not declared in this scope\n     __NV_EXTRA_INITIALIZATION\n     ^\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc:144:5: error: '__NV_EXTRA_FINALIZATION' was not declared in this scope\n     __NV_EXTRA_FINALIZATION\n     ^\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc:145:23: warning: statement has no effect [-Wunused-value]\n     for (__i = 0; __i &lt; NUM_PRELINKED_OBJECTS; ++__i) {\n                       ^\nbazel-out\/k8-opt\/bin\/external\/nccl_archive\/device_code.cc:145:53: error: expected ';' before ')' token\n     for (__i = 0; __i &lt; NUM_PRELINKED_OBJECTS; ++__i) {\n<\/code><\/pre>\n\n<p><strong>Additional info:<\/strong><br>\nCuda Version 10.0.130<br>\nCuDNN 7.3<br>\nUbuntu 16.04.5 LTS<br>\nGPU:<br>\n00:03.0 VGA compatible controller [0300]: NVIDIA Corporation GK104GL [GRID K520] [10de:118a] (rev a1)\n    Subsystem: NVIDIA Corporation GK104GL [GRID K520] [10de:1014]\n    Kernel driver in use: nvidia\n    Kernel modules: nvidia_410_drm, nvidia_410<\/p>\n","answers":[]}
+{"QuestionId":52757484,"AnswerCount":1,"Tags":"<python><function><keras><keras-layer>","CreationDate":"2018-10-11T10:10:19.457","AcceptedAnswerId":52758489.0,"OwnerUserId":9476597.0,"Title":"How to use a layer the same way you use function in Keras?","Body":"<p>I'm a student and beginner with Keras. I wonder how to use a layer such in the same way of a python function.\nI want to define a layers, a Dense one, for example, and then be able to compute the image of an x given by the layer.<\/p>\n\n<p>Thanks and have a good day.<\/p>\n","answers":[{"AnswerId":"52758489","CreationDate":"2018-10-11T11:04:30.200","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>So first, take a look at <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Keras Functional API<\/a>. That way you can do something like this:<\/p>\n\n<pre><code>from keras.layers import *\n\n# load some data into x\nx = load()\n\ndense_output = Dense(4096, activation='relu')(x)\ndropout_output = Dropout(0.3)(dense_output)   \n...\n<\/code><\/pre>\n\n<p>You can do basically everything you need with Functional API. I hope this clarifies the situation for you :)<\/p>\n"}]}
+{"QuestionId":52758051,"AnswerCount":2,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-10-11T10:41:39.643","AcceptedAnswerId":52799649.0,"OwnerUserId":1714410.0,"Title":"How to save LambdaLR scheduler in pytorch with lambda function?","Body":"<p>Running pytorch 0.4.1 with python 3.6 I encountered this problem:<br>\nI cannot <code>torch.save<\/code> my learning rate scheduler because python won't pickle a lambda function:<\/p>\n\n<pre><code>lambda1 = lambda epoch: epoch \/\/ 30\nscheduler = LambdaLR(optimizer, lr_lambda=lambda1)\ntorch.save(scheduler.state_dict(), 'scheduler.pth.tar')\n<\/code><\/pre>\n\n<p>results with an error<\/p>\n\n<blockquote>\n<pre><code>PicklingError: Can't pickle &lt;function &lt;lambda&gt; at 0x7f7583fe92f0&gt;:\nattribute lookup &lt;lambda&gt; on __main__ failed\n<\/code><\/pre>\n<\/blockquote>\n\n<p>How can I save my scheduler?<\/p>\n\n<hr>\n\n<p>I know that using a proper function instead of a lambda function for <code>lambda1<\/code> can be saved, but I need a lambda function because I want to be able to control this function when it is defined (for instance I want to be able to change the fixed 30 in the denominator).<br>\nHow can this be done and still allows me to save the scheduler?<\/p>\n","answers":[{"AnswerId":"52783566","CreationDate":"2018-10-12T16:26:35.810","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>You can use a <code>dill<\/code> as the pickle module instead of default pickle. <\/p>\n\n<pre><code>import dill\ntorch.save(scheduler.state_dict(), 'scheduler.pth.tar', pickle_module=dill)\n<\/code><\/pre>\n\n<p>and to load <\/p>\n\n<pre><code>import dill\nobject = torch.load('scheduler.pth.tar', pickle_module=dill)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=save#torch.save\" rel=\"nofollow noreferrer\">documentation<\/a> of save for more options. <\/p>\n"},{"AnswerId":"52799649","CreationDate":"2018-10-14T05:08:16.583","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If one wishes to stay with default behavior of <code>torch.save<\/code> and <code>torch.load<\/code>, the lambda function can be replaced with a class, for example:<\/p>\n\n<pre><code>class LRPolicy(object):\n    def __init__(self, rate=30):\n        self.rate = rate\n\n    def __call__(self, epoch):\n        return epoch \/\/ self.rate\n<\/code><\/pre>\n\n<p>The scheduler is now<\/p>\n\n<pre><code>scheduler = LambdaLR(optimizer, lr_lambda=LRPolicy(rate=30))\n<\/code><\/pre>\n\n<p>Now the scheduler can be <code>torch.save<\/code>ed and <code>torch.load<\/code> without alternating the pickling module.<\/p>\n"}]}
+{"QuestionId":52758204,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-11T10:49:22.090","AcceptedAnswerId":null,"OwnerUserId":8532515.0,"Title":"A function contains tf.Session().run() gets slower and slower","Body":"<p>I wrote a function to process some image with tensorflow, and the code looks like this:<\/p>\n\n<pre><code>def function:\n    sess=tf.interactiveSession()\n    for im in some_image_array\n        new_list.append(sess.run(tf.cast(im,tf.float32)))\n    sess.close()\n<\/code><\/pre>\n\n<p>I found that the using time of this function became longer with the repeated using, why and how can I improve it? <\/p>\n","answers":[]}
+{"QuestionId":52758654,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-11T11:13:23.287","AcceptedAnswerId":null,"OwnerUserId":9594598.0,"Title":"Cannot import tf-sentencepiece","Body":"<p>I have installed tf-sentencepiece with pip, and when I run the line:<\/p>\n\n<pre><code>import tf_sentencepiece as tfs\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.NotFoundError: \/home\/ubuntu\/.local\/lib\/python2.7\/site-packages\/tf_sentencepiece\/_sentencepiece_processor_ops.so: undefined symbol: _ZN10tensorflow20OpKernelConstruction21CtxFailureWithWarningEPKciRKNS_6StatusE\n<\/code><\/pre>\n\n<p>Any idea why I can't import the library? Thanks.<\/p>\n","answers":[{"AnswerId":"52772535","CreationDate":"2018-10-12T04:59:44.223","ParentId":null,"OwnerUserId":"9594598","Title":null,"Body":"<p>Solved this by changing tensorflow version to 1.10.0. <\/p>\n"}]}
+{"QuestionId":52758723,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2018-10-11T11:16:58.960","AcceptedAnswerId":null,"OwnerUserId":9460641.0,"Title":"InvalidArgumentError : assertion failed: [maximum box coordinate value is larger than 1.100000: ] [1.15277779]","Body":"<p>I am doing object detection in tensorflow and using the <code>faster_rcnn_resnet101_coco_2018_01_28 model<\/code> for training locally but I am getting the follwing error: <\/p>\n\n<pre><code>    InvalidArgumentError (see above for traceback): assertion failed: [maximum box coordinate value is larger than 1.100000: ] [1.15277779]\n    [[Node: ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert = Assert[T=[DT_STRING, DT_FLOAT], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert\/Switch, Loss\/ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert\/data_0, ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert\/Switch_1)]]\n<\/code><\/pre>\n\n<p>The traceback is :<\/p>\n\n<pre><code>    \/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/ops.py, line 1717, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n\n<p>I tried training using <code>ssd_mobilenet_v1_pets<\/code> but training was successful.\nMy training images are a mixture of different sizes. I don't believe any of my bounding boxes are outside the image coordinates. Even if that's the case why would the ssd_mobilenet_v1_pets model work but not the resnet model?<\/p>\n","answers":[]}
+{"QuestionId":52758806,"AnswerCount":0,"Tags":"<tensorflow><keras><amd-rocm>","CreationDate":"2018-10-11T11:21:47.207","AcceptedAnswerId":null,"OwnerUserId":9468014.0,"Title":"ROCm and CUDA GPUs on one model?","Body":"<p>I want to compile a model on multiple GPU's. Is it possible to use AMD GPUs with ROCm together with Nvidia CUDA GPUs on the same model?\nI guess it's not possible since you need to install a special version of tensorflow in order to enable ROCm tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":52759193,"AnswerCount":2,"Tags":"<python><machine-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2018-10-11T11:41:47.997","AcceptedAnswerId":52759784.0,"OwnerUserId":3482314.0,"Title":"LSTM in Keras: Number of parameters differs between sequential and functional API","Body":"<p><strong>With the Sequential API<\/strong><\/p>\n\n<p>If I create a LSTM with the Sequential API of Keras with the following code:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import LSTM\n\nmodel = Sequential()\nmodel.add(LSTM(2, input_dim=3))\n<\/code><\/pre>\n\n<p>then <\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n\n<p>returns <strong>48<\/strong> parameters, which is OK as indicated in <a href=\"https:\/\/stackoverflow.com\/questions\/38080035\/how-to-calculate-the-number-of-parameters-of-an-lstm-network\">this Stack Overflow question<\/a>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7k9Pr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7k9Pr.png\" alt=\"model.summary() with the Sequential API\"><\/a><\/p>\n\n<p><em>Quick details:<\/em><\/p>\n\n<pre><code>input_dim = 3, output_dim = 2\nn_params = 4 * output_dim * (output_dim + input_dim + 1) = 4 * 2 * (2 + 3 + 1) = 48\n<\/code><\/pre>\n\n<p><strong>With the Functional API<\/strong><\/p>\n\n<p>But if I do the same with the functional API with the following code:<\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Input\nfrom keras.layers import LSTM\n\ninputs = Input(shape=(3, 1))\nlstm = LSTM(2)(inputs)\nmodel = Model(input=inputs, output=lstm)\n<\/code><\/pre>\n\n<p>then<\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n\n<p>returns <strong>32<\/strong> parameters.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xBA9n.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xBA9n.png\" alt=\"model.summary() with the Functional API\"><\/a><\/p>\n\n<p>Why there is such a difference?<\/p>\n","answers":[{"AnswerId":"54274348","CreationDate":"2019-01-20T07:12:30.120","ParentId":null,"OwnerUserId":"97651","Title":null,"Body":"<p>I solved it like the following:<\/p>\n\n<pre><code>Case 1:\nm (input) = 3\nn (output) = 2\n\nparams = 4 * ( (input * output) + (output ^ 2) + output)\n       = 4 * (3*2 + 2^2 + 2)\n       = 4 * (6 + 4 + 2)\n       = 4 * 12\n       = 48\n\n\n\nCase 2:\nm (input) = 1\nn (output) = 2\n\nparams = 4 * ( (input * output) + (output ^ 2) + output)\n       = 4 * (1*2 + 2^2 + 2)\n       = 4 * (2 + 4 + 2)\n       = 4 * 8\n       = 32\n<\/code><\/pre>\n"},{"AnswerId":"52759784","CreationDate":"2018-10-11T12:11:19.197","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The difference is that when you pass <code>input_dim=x<\/code> to a RNN layer, including LSTM layers, it means that the input shape is <code>(None, x)<\/code> i.e. there are varying number of timesteps where each one is a vector of length <code>x<\/code>. However, in the functional API example, you are specifying <code>shape=(3, 1)<\/code> as input shape and it means there are 3 timesteps where each has one feature. Therefore the number of parameters would be: <code>4 * output_dim * (output_dim + input_dim + 1) = 4 * 2 * (2 + 1 + 1) = 32<\/code> which is the number shown in the model summary.<\/p>\n\n<p>Further, if you use Keras 2.x.x, you would get a warning in case of using <code>input_dim<\/code> argument for a RNN layer:<\/p>\n\n<blockquote>\n  <p>UserWarning: The <code>input_dim<\/code> and <code>input_length<\/code> arguments in recurrent\n  layers are deprecated. Use <code>input_shape<\/code> instead.<\/p>\n  \n  <p>UserWarning: Update your <code>LSTM<\/code> call to the Keras 2 API: <code>LSTM(2, input_shape=(None, 3))<\/code><\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":52759259,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2018-10-11T11:44:44.250","AcceptedAnswerId":52761437.0,"OwnerUserId":2996578.0,"Title":"Defining grad of a custom Op theano","Body":"<p>I am trying to define a custom theano Op with a gradient to use it with pymc3 but I don't understand how to define the <code>grad<\/code> method.<\/p>\n\n<p>The code below is where I'm stuck. The function <code>phi()<\/code> is a mock function (in practice, it is an external program); for a scalar input <code>x<\/code> it returns a vector <code>(phi_0(x), phi_1(x), ...)<\/code>. The function <code>phi_diff()<\/code> (also a mock function) returns the vector <code>(dphi_0\/dx, dphi_1\/dx, ...)<\/code>.<\/p>\n\n<p>I wrapped <code>phi()<\/code> and <code>phi_diff()<\/code> in a <code>theano.Op<\/code> object but my implementation of the <code>grad<\/code> function does not work. The documentation of theano contains simpler examples, I don't understand how to adapt them in this case. Any help would be greatly appreciated.<\/p>\n\n<pre><code>import numpy as np\nimport theano.tensor as T\nimport theano\n\ntheano.config.optimizer = \"None\"\ntheano.config.exception_verbosity = \"high\"\n\n\ndef phi(x):\n    return np.arange(n) * x\n\n\ndef phi_diff(x):\n    return np.arange(n)\n\n\nclass PhiOp(theano.Op):\n    itypes = [theano.tensor.dscalar]\n    otypes = [theano.tensor.dvector]\n\n    def perform(self, node, inputs, output_storage):\n        x = inputs[0]\n        output_storage[0][0] = phi(x)\n\n    def grad(self, inputs, output_grads):\n        x = inputs[0]\n        # ???\n        return [PhiDiffOp()(x) * output_grads[0]]\n\n\nclass PhiDiffOp(theano.Op):\n    itypes = [theano.tensor.dscalar]\n    otypes = [theano.tensor.dvector]\n\n    def perform(self, node, inputs, output_storage):\n        x = inputs[0]\n        output_storage[0][0] = phi_diff(x)\n\n\nn = 5\nx = 777.\n\nphi_op = PhiOp()\nx_tensor = T.dscalar(\"x_tensor\")\nphi_func = theano.function([x_tensor], phi_op(x_tensor))\nnp.testing.assert_allclose(phi_func(x), phi(x))\n\nT.jacobian(phi_op(x_tensor), x_tensor)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52761437","CreationDate":"2018-10-11T13:34:58.507","ParentId":null,"OwnerUserId":"2996578","Title":null,"Body":"<p>Found the solution, changes below:<\/p>\n\n<pre><code>def phi_diff(x):\n    return np.arange(n, dtype=np.float_)\n\nclass PhiOp(theano.Op):\n    def grad(self, inputs, output_grads):\n        x = inputs[0]\n        gg = (PhiDiffOp()(x) * output_grads[0]).sum()\n        return [gg]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52759587,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><image-processing><anaconda><object-detection>","CreationDate":"2018-10-11T12:01:22.263","AcceptedAnswerId":null,"OwnerUserId":8778978.0,"Title":"How can I use tensorflow-gpu version if I have AMD graphics card?","Body":"<p>I'm using Anaconda in Windows 10 and I cannot install the current version of <code>tensorflow-gpu<\/code>. I have 4 GB AMD Radeon graphics card. So how can I run <code>Tensorflow-gpu<\/code> version in my laptop?<\/p>\n","answers":[{"AnswerId":"53088315","CreationDate":"2018-10-31T16:42:10.573","ParentId":null,"OwnerUserId":"3316997","Title":null,"Body":"<p>As far as I know you need Linux and ROCm to have AMD GPUs supporting Tensorflow.\nIn my case with ubuntu and ROCm 1.9.1, in Python3:<\/p>\n\n<pre><code>import tensorflow as tf; print(tf.__version__)\n&gt;&gt;1.11.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":52759776,"AnswerCount":0,"Tags":"<python><c++><opencv><tensorflow>","CreationDate":"2018-10-11T12:11:05.400","AcceptedAnswerId":null,"OwnerUserId":10489518.0,"Title":"OpenCV 3.4.1 error readNetFromTensorflow Can't open .pb in cv::dnn::ReadProtoFromBinaryFile","Body":"<p>I have a problem with opening protobuf file using opencv C++.\nI use this code:<\/p>\n\n<pre><code>cv::String weights = \"frozen_inference_graph_face.pb\";\ncv::String pbtxt = \"prototxt.pbtxt\";\nauto graph = cv::dnn::readNetFromTensorflow(weights, pbtxt);\n<\/code><\/pre>\n\n<p>I have this error:<\/p>\n\n<p>OpenCV(3.4.1) Error: Unspecified error (FAILED: fs.is_open(). Can't open \"frozen_inference_graph_face.pb\") in cv::dnn::ReadProtoFromBinaryFile, file C:.hunter_Base\\acbf4b9\\93b3222\\8eb84a0\\Build\\OpenCV\\Source\\modules\\dnn\\src\\caffe\\caffe_io.cpp, line 1126<\/p>\n\n<p>It works well when I open it with Python code like this and detect image correctly:<\/p>\n\n<pre><code>cvNet = \ncv.dnn.readNetFromTensorflow('frozen_inference_graph.pb','prototxt.pbtxt')\n<\/code><\/pre>\n\n<p>I have trained ssd_mobilenet_v1_pets. Cannot understand why I cannot open it with my C++ code and the error is refers to cafe, when I use tensorflow. Maybe the configuration of builded OpenCV is wrong? I set WITH_PROTOBUF=ON and BUILD_opencv_dnn=ON.<\/p>\n","answers":[]}
+{"QuestionId":52760163,"AnswerCount":1,"Tags":"<tensorflow><graph><replace><editor>","CreationDate":"2018-10-11T12:31:37.817","AcceptedAnswerId":null,"OwnerUserId":3049992.0,"Title":"Tensorflow: Replace one op with another(maybe even 2 ops)","Body":"<p>My goal is to build a script to change an operation into another one using TF's graph editor. So far I tried making a script that just changes the input kernel weights of a Conv2D, but to no avail, as the interface is pretty confusing.<\/p>\n\n<pre><code>with tf.Session() as sess:\n\n    model_filename = sys.argv[1]\n\n    with gfile.FastGFile(model_filename, 'r') as f:\n\n        graph_def = graph_pb2.GraphDef()\n        text_format.Merge(f.read(), graph_def)\n        importer.import_graph_def(graph_def)\n\n        #my_sgv = ge.sgv(\"Conv2D\", graph=tf.get_default_graph())\n        #print my_sgv\n\n        convs = find_conv2d_ops(tf.get_default_graph())\n        print convs\n\n        my_sgv = ge.sgv(convs)\n        print my_sgv\n\n        conv_tensor = tf.get_default_graph().get_tensor_by_name(convs[0].name + ':0')\n        conv_weights_input = tf.get_default_graph().get_tensor_by_name(convs[0].inputs[1].name)\n\n        weights_new = tf.Variable(tf.truncated_normal([1, 1, 1, 8], stddev=0.03),\n                                  name='Wnew')\n\n        ge.graph_replace(conv_tensor, {conv_weights_input: weights_new})\n<\/code><\/pre>\n\n<p>The error is \"input needs to be a Tensor: \". Can someone please provide some insights?<\/p>\n","answers":[{"AnswerId":"52761894","CreationDate":"2018-10-11T13:55:27.823","ParentId":null,"OwnerUserId":"1097517","Title":null,"Body":"<p>Since you are dealing with a <code>tf.Variable<\/code> you don't need to use graph editor. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\"><code>tf.assign<\/code><\/a> will be sufficient.<\/p>\n\n<p>You can use it like the following:<\/p>\n\n<pre><code>assign_op = tf.assign(conv_weights_input, weights_new)\nwith tf.Session() as sess:\n    sess.run(assign_op)\n<\/code><\/pre>\n\n<p>If you are looking to sub out operations and not weights. Consider the following example (modified from <a href=\"https:\/\/gist.github.com\/ByungSunBae\/393071e46409737cd341360a69957906\" rel=\"nofollow noreferrer\">this<\/a> example):<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.contrib.graph_editor as ge\n\ndef build():\n    a_pl = tf.placeholder(dtype=tf.float32, name=\"a\")\n    b_pl = tf.placeholder(dtype=tf.float32, name=\"b\")\n    c = tf.add(a_pl, b_pl, name=\"c\")\n\nbuild() #or load graph from disc\n\na = tf.constant(1.0, shape=[2, 3], name=\"a_const\")\nb = tf.constant(2.0, shape=[2, 3], name=\"b_const\")\n\na_pl = tf.get_default_graph().get_tensor_by_name(\"a:0\")\nb_pl = tf.get_default_graph().get_tensor_by_name(\"b:0\")\nc = tf.get_default_graph().get_tensor_by_name(\"c:0\")\n\nc_ = ge.graph_replace(c, {a_pl: a, b_pl: b})\n\nwith tf.Session() as sess:\n    #no need for placeholders\n    print(sess.run(c_))\n    #will give error since a_pl and b_pl have no value\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>The issue with your code is that you're dealing with wights, and not tensors. The crux of the above example is that the first argument is the target tensor (output tensor) that have the to be replaced tensors as dependencies. The second argument are the actual tensors you want to replace.<\/p>\n\n<p>It's also worth noting that <code>conv_weights_input<\/code> is actually a tensor, where <code>weights_new<\/code> is a <code>tf.Variable<\/code>. I believe what you want is to replace <code>weights_new<\/code> with a new conv operation with random weight initialisation.<\/p>\n"}]}
+{"QuestionId":52761157,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-10-11T13:22:33.773","AcceptedAnswerId":null,"OwnerUserId":8043141.0,"Title":"tensorflow-gpu in Anaconda 3.6.6","Body":"<p>stack trace when i run 'import tensorflow as tf' in Anaconda Prompt:<\/p>\n\n<pre><code>import tensorflow as tf\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-1-64156d691fe5&gt;\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\n\nImportError: Traceback (most recent call last):\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions. Include the entire stack trace above this error message when asking for help.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52761338,"AnswerCount":1,"Tags":"<python><tensorflow><chatbot><google-colaboratory>","CreationDate":"2018-10-11T13:30:33.953","AcceptedAnswerId":null,"OwnerUserId":10474916.0,"Title":"No tensorflow in Google Colaboratory local host","Body":"<p>There is no tensorflow in google colab local host machine and there is no way to install it. When I run train.py it shows no module named 'tensorflow'.I would really appreciate some help.<\/p>\n","answers":[{"AnswerId":"52789498","CreationDate":"2018-10-13T04:22:35.143","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>Unfortunately, there's no one-size-fits-all way to install TensorFlow on your local machine. (Part of the value of Colab's backends is dealing with all of that.)<\/p>\n\n<p>Fortunately, TensorFlow does have a fairly comprehensive set of install steps for most common environments; see: <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/<\/a><\/p>\n\n<p>My recommendation is to start there, try to work through the process, and, if you hit a snag, ask a follow-up question that describes your OS, install type, what you tried, and the error you encountered.<\/p>\n"}]}
+{"QuestionId":52761533,"AnswerCount":0,"Tags":"<optimization><neural-network><pytorch><loss-function>","CreationDate":"2018-10-11T13:38:38.000","AcceptedAnswerId":null,"OwnerUserId":7580944.0,"Title":"Loss function for ordinal multi classification in pytorch","Body":"<p>I am a beginner with DNN and pytorch.\nI am dealing with a multi-classification problem where my label are encoded into a one-hotted vector, say of dimension <code>D<\/code>.\nTo this end, I am using the CrossEntropyLoss. However now I want to modify or change such criterion in order to penalize value distant from the actual one, say classify 4 instead of 5 is better than 2 instead of 5.<\/p>\n\n<p>Is there a function already built-in in Pytorch that implement this behavior? Otherwise how can I modify the CrossEntropyLoss to achieve it?<\/p>\n","answers":[]}
+{"QuestionId":52761666,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><conv-neural-network><pytorch>","CreationDate":"2018-10-11T13:44:45.627","AcceptedAnswerId":52771444.0,"OwnerUserId":9049053.0,"Title":"PyTorch RuntimeError Invalid argument 2 of size","Body":"<p>I am experimenting with a neural network (PyTorch) and I get this error.<\/p>\n\n<blockquote>\n  <p>RuntimeError: invalid argument 2: size '[32 x 9216]' is invalid for input with 8192 elements at \/pytorch\/aten\/src\/TH\/THStorage.cpp:84<\/p>\n<\/blockquote>\n\n<p>My task is about image classification with AlexNet and I have backtracked the error to be the size of the images supplied to the neural network. My question is, given the network architecture with its parameters, how does one determine the correct image size required by the network?<\/p>\n\n<p>As per my code below, I first transform the training images before feeding into the neural network. But I noticed the neural network can only accept the size of <code>224<\/code> and or else it gives the error above. For instance, my instinct was to apply <code>transforms.RandomResizedCrop<\/code> of size 64 but apparently this is wrong. Is there a formula to determine the size required?<\/p>\n\n<p><strong>Code<\/strong><\/p>\n\n<pre><code># transformation to be done on images\ntransform_train = transforms.Compose([\n    transforms.RandomResizedCrop(64),\n    transforms.RandomHorizontalFlip(),\n    transforms.ToTensor(),\n    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n])\n\nclass AlexNet(nn.Module):\n\n    def __init__(self, num_classes=1000):\n        super(AlexNet, self).__init__()\n        self.features = nn.Sequential(\n            nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),\n            nn.ReLU(inplace=True),\n            nn.MaxPool2d(kernel_size=3, stride=2),\n            nn.Conv2d(64, 192, kernel_size=5, padding=2),\n            nn.ReLU(inplace=True),\n            nn.MaxPool2d(kernel_size=3, stride=2),\n            nn.Conv2d(192, 384, kernel_size=3, padding=1),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(384, 256, kernel_size=3, padding=1),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(256, 256, kernel_size=3, padding=1),\n            nn.ReLU(inplace=True),\n            nn.MaxPool2d(kernel_size=3, stride=2),\n        )\n        self.classifier = nn.Sequential(\n            nn.Dropout(),\n            nn.Linear(256 * 6 * 6, 4096),\n            nn.ReLU(inplace=True),\n            nn.Dropout(),\n            nn.Linear(4096, 4096),\n            nn.ReLU(inplace=True),\n            nn.Linear(4096, num_classes),\n        )\n\n    def forward(self, x):\n        x = self.features(x)\n        x = x.view(x.size(0), 256 * 6 * 6)\n        x = self.classifier(x)\n        return x\n<\/code><\/pre>\n","answers":[{"AnswerId":"52771444","CreationDate":"2018-10-12T02:43:34.600","ParentId":null,"OwnerUserId":"9049053","Title":null,"Body":"<p>I have figured out the algorithm of getting the right input size.<\/p>\n\n<pre><code>Out = float(((W\u2212F+2P)\/S)+1)\n<\/code><\/pre>\n\n<p>where<\/p>\n\n<ul>\n<li>Out = Output shape<\/li>\n<li>W = Image volume size (image size)<\/li>\n<li>F = Receptive field (filter size)<\/li>\n<li>P = Padding<\/li>\n<li>S = Stride<\/li>\n<\/ul>\n\n<p>Factoring in the given network hyperparameters,<\/p>\n\n<p>The require Image size I need would be<\/p>\n\n<pre><code>W = (55 - 1) * 4 - 2(2) + 11\n  =  223\n  \u2a70  224\n<\/code><\/pre>\n"},{"AnswerId":"52772853","CreationDate":"2018-10-12T05:30:46.177","ParentId":null,"OwnerUserId":"5349822","Title":null,"Body":"<p>The actual formula to calculate the output shape after convolution layer is:<\/p>\n\n<p><code>out_size= floor((in_size + 2p -f)\/s + 1)<\/code><\/p>\n"}]}
+{"QuestionId":52761799,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-11T13:50:42.660","AcceptedAnswerId":52833993.0,"OwnerUserId":5484087.0,"Title":"Manually Get Next Batch or Use Identical Batch with TensorFlow Data API","Body":"<p>I am trying to use the tf.Data API to accelerate my code and prevent GPU data starvation but there is one thing that is stopping me for being comfortable with it and it's the ability to use the same batch when calling the training op multiple times.<\/p>\n\n<p>Suppose I have my dataset set up as<\/p>\n\n<pre><code>dataset = tf.data.TextLineDataset(\"textfile.txt\")\ndataset = dataset.shuffle(dataset_size)\ndataset = dataset.padded_batch(batch_size, ...)\ndataset = dataset.repeat()\niterator = dataset.make_one_shot_iterator()\nx_batch = iterator.get_next()\n\nloss1 = someFunctionOf(x_batch)\nloss2 = someOtherFunctionOf(x_batch)\ntrain_op1 = someOptimizerOf(loss1)\ntrain_op2 = someOtherOptimizerOf(loss2)\n<\/code><\/pre>\n\n<p>but now whenever I call <code>train_op1<\/code>, <code>iterator.get_next()<\/code> is called and so when calling <code>train_op2<\/code>, I am training on the next batch. <\/p>\n\n<p>From <a href=\"https:\/\/stackoverflow.com\/questions\/49358750\/how-can-i-return-the-same-batch-twice-from-a-tensorflow-dataset-iterator\">this<\/a> question, I am aware that I can use a combination of <code>flat_map<\/code> and <code>repeat(n)<\/code> where <code>n<\/code> is the number of times I wanna repeat the same batch but this <code>n<\/code> would depend on the number of <code>train_ops<\/code> that I call which I have to count manually. Also, I need these two <code>train_ops<\/code> because they optimize different parts of my graph.<\/p>\n\n<p>Thank you for your help!<\/p>\n","answers":[{"AnswerId":"52833993","CreationDate":"2018-10-16T11:03:57.983","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>Try the code below. It creates a copy of input and target so hopefully they will not change when you switch optimizer\/loss_op. They are persistent between <code>sess.run<\/code> calls as long as you do not pass <code>is_new:True<\/code> flag.<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\nimport tensorflow as tf\n\n\ndef ds_train(batch_size, num_epochs):  \n    ds = (tf.data.Dataset.from_tensor_slices(([1.0,2.0,3.0,4.0,5.0], [-1,-2,-3,-4,-5]))\n            .batch(batch_size)\n            .repeat(num_epochs)        \n            )\n    return ds\n\n\nbatch_size = 1\ninput_size = 1\nnum_epochs = 2\n\nwith tf.variable_scope(\"dataset\"):       \n    ds_t = ds_train(batch_size, num_epochs)\n\nwith tf.variable_scope(\"iterator\"):\n    iterator_t = ds_t.make_initializable_iterator()\n    iterator_handle = tf.placeholder(tf.string, shape=[], name=\"iterator_handle\")\n    iterator = tf.data.Iterator.from_string_handle(iterator_handle, \n                                                iterator_t.output_types,\n                                                iterator_t.output_shapes)\n\n    def next_item():\n        next_elem = iterator.get_next(name=\"next_element\")\n        x, y = tf.cast(next_elem[0], tf.float32), next_elem[1]# tf.cast(next_elem[1], tf.int32)\n        return x, y        \n\n\ninputs = tf.Variable(tf.zeros(shape=[batch_size,input_size]), dtype=tf.float32, name=\"inputs\", trainable=False, use_resource=True)\ntarget = tf.Variable(tf.zeros(shape=[batch_size], dtype=tf.int32), dtype=tf.int32, name=\"target\", trainable=False,use_resource=True)\nis_new = tf.placeholder_with_default(tf.constant(False), shape=[], name=\"new_item_flag\")\n\ndef new_data(batch_size, input_size):\n    # run the data layer to generate a new batch\n    next_inputs, next_target = next_item()\n    next_inputs = tf.reshape(next_inputs, shape=[batch_size, input_size])\n    with tf.control_dependencies([tf.assign(inputs, next_inputs), tf.assign(target, next_target)]):\n        return tf.identity(inputs), tf.identity(target)\n\ndef old_data():\n    # just forward the existing batch\n    return inputs, target\n\nnext_inputs, next_target = next_item()\n\ninputs, target =  tf.cond(is_new, lambda:new_data(batch_size, input_size), old_data)\n\nwith tf.Session() as sess:\n    sess.run([tf.global_variables_initializer(),tf.local_variables_initializer()])\n    handle_t = sess.run(iterator_t.string_handle())\n    sess.run(iterator_t.initializer)\n    while True:\n        try:\n            print(sess.run([inputs, target], feed_dict={iterator_handle:handle_t, is_new: False}))\n            print(sess.run([inputs, target], feed_dict={iterator_handle:handle_t, is_new: False}))\n            print(sess.run([inputs, target], feed_dict={iterator_handle:handle_t, is_new: True}))\n        except tf.errors.OutOfRangeError:\n            print(\"End of training dataset.\")\n            break        \n<\/code><\/pre>\n"}]}
+{"QuestionId":52761829,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory>","CreationDate":"2018-10-11T13:51:59.953","AcceptedAnswerId":null,"OwnerUserId":10490123.0,"Title":"How do I upload my own data for tensor flow using google colab?","Body":"<p>I've been looking through the tensorflow tutorials online (specifically the housing prices tutorial: <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/samples\/core\/tutorials\/keras\/basic_regression.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/samples\/core\/tutorials\/keras\/basic_regression.ipynb<\/a> )<\/p>\n\n<p>I've been trying to upload my own csv file for a similar project using google colab. But I can't seem to get the format right - I'm very new to this, so I can't find a solution which I can understand. <\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function\n\nimport tensorflow as tf\nfrom tensorflow import keras\n\nimport numpy as np\nimport pandas as pd\n\nprint(tf.__version__)\n\n#Import the csv files\n\nfrom google.colab import files\nuploaded = files.upload()\n\nfor fn in uploaded.keys():\n  print('User uploaded file \"{name}\" with length {length} bytes'.format(\n  name = fn, length = len(uploaded[fn])))\n\n# This is where I upload my csv file\n\nimport io\n\ndf = pd.read_csv(io.StringIO(uploaded[ 'data.csv'].decode('utf-8')))\ndf.head()\n\n(train_data, train_labels), (test_data, test_labels) = uploaded.load_data()\n\n# Shuffle the training set\norder = np.argsort(np.random.random(train_labels.shape))\ntrain_data = train_data[order]\ntrain_labels = train_labels[order]\nprint(boston_housing)\n<\/code><\/pre>\n\n<p>This is where the problem is - I can't seem to separate my data into my training and test data.<\/p>\n\n<p>My data.csv is just 5 columns. Col 1-2 contain two sets of inputs, col3 contains the label, col 3-4 contain test input data. <\/p>\n\n<p>Again, massive newbie, any help would be amazing! I'm so confused<\/p>\n","answers":[{"AnswerId":"52789534","CreationDate":"2018-10-13T04:30:08.070","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>I'm guessing this line is the problem:<\/p>\n\n<pre><code>(train_data, train_labels), (test_data, test_labels) = uploaded.load_data()\n<\/code><\/pre>\n\n<p><code>uploaded<\/code> is the result of the <code>files.upload<\/code> command, and that doesn't include a <code>load_data<\/code> method. Instead, it puts a copy of the file on the local filesystem and returns a dict containing the bytes of each uploaded file indexed by filename keys. For example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vX9n5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vX9n5.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You've already got data as a <code>DataFrame<\/code> in <code>df<\/code>. So, to split into test and training, perhaps do something like the recipes suggested here: <a href=\"https:\/\/stackoverflow.com\/questions\/24147278\/how-do-i-create-test-and-train-samples-from-one-dataframe-with-pandas\">How do I create test and train samples from one dataframe with pandas?<\/a><\/p>\n"}]}
+{"QuestionId":52762173,"AnswerCount":1,"Tags":"<python><neural-network><gpu><conv-neural-network><pytorch>","CreationDate":"2018-10-11T14:08:13.877","AcceptedAnswerId":null,"OwnerUserId":5011934.0,"Title":"Pytorch convolutional network memory usage details","Body":"<p>I am trying to train a neural network for a very large input (5*100,000,000) and it requires much more memory than expected.\nHere is some minimal example:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nimport time\n\nclass Net(nn.Module):\n\n    def __init__(self):\n        super(Net, self).__init__()\n        self.conv1 = nn.Conv1d(in_channels=5, out_channels=1, kernel_size=100000000, stride=10)\n\n    def forward(self, x):\n        x = self.conv1(x)\n        x = torch.sigmoid(x)\n        return x\n\nmodel = Net().cuda()\n\noptimizer = optim.Adam(model.parameters(), lr=0.001)\ncriterion = torch.nn.BCELoss()\n\ndata = torch.normal(torch.zeros(1,5,100000000),torch.ones(1,5,100000000))\ndata = data.cuda()\nlabel = torch.ones(1,1,1)\nlabel = label.cuda()\n\nfor epoch in range(10):\n    output = model(data)\n    loss = criterion(output, label)\n\n    optimizer.zero_grad()\n    loss.backward()\n    optimizer.step()\n\n    print(\"Epoch :\", epoch)\n<\/code><\/pre>\n\n<p>The input is some random data, it uses approximately 2Gb, as expected (32 bit * 5 * 100,000,000= 1.86Gb).This variable has no gradient.\nThe network consists of a single convolutional layer with one filter of the same size as an input, so it has 500M weights, that is another 2Gb.\nAfter the forward pass another 2Gb get used.\nAfter <code>loss.backprop()<\/code> 8Gb are used, after <code>optimizer.step()<\/code> 12 Gb are used, that is all the available memory.<\/p>\n\n<p>During the second epoch forward pass runs ok, but during backpropagation I get RuntimeError: CUDA error: out of memory.<\/p>\n\n<p>What exactly is saved in GPU memory during the epoch? Why the memory is not released after the optimization step is finished? How to reduce memory usage in this case?<\/p>\n\n<p>UPD: Looks like my problems are similar to this issue <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-free-gpu-memory-and-delete-memory-allocated-variables\/20856\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-to-free-gpu-memory-and-delete-memory-allocated-variables\/20856<\/a><\/p>\n","answers":[{"AnswerId":"52779463","CreationDate":"2018-10-12T12:22:15.757","ParentId":null,"OwnerUserId":"10494960","Title":null,"Body":"<p>Since you want to call <code>loss.backprop()<\/code>, PyTorch has to calculate the gradients, and this contributes to the large memory allocation.\nIf you want to drop gradients call <code>.detach()<\/code> on a variable.<\/p>\n\n<p>To release unused memory, you can call <code>torch.cuda.empty_cache()<\/code>\nIf you want to dive into the details, \n<a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/cuda.html\" rel=\"nofollow noreferrer\">the CUDA semantics page<\/a>\nmay be a starting point.<\/p>\n"}]}
+{"QuestionId":52762764,"AnswerCount":1,"Tags":"<tensorflow><face-detection><object-detection-api>","CreationDate":"2018-10-11T14:36:45.383","AcceptedAnswerId":null,"OwnerUserId":7844769.0,"Title":"CNN Object detection: How to reduce high false positive rates","Body":"<p>I am currently using Faster RCNN with inception v2 on Tensorflow object detection API on WIDER-FACE data, i have a lot of false positives with high scores (>0.98, so setting a higher score threshold won't help).\nI have already assigned Hard Example Mining in my code but it doesn't help to much.\nFor image preprocessing, I randomly crop an area from the original image using <strong>tf.image.sample_distorted_bounding_box<\/strong> and resize it to 300*300. The resized image will be randomly flipped by a probability of 0.5<\/p>\n\n<p>I set batch size to 32 and for each image the ratio of <strong>positive:negative<\/strong> in hard negative mining operation is <strong>32:32<\/strong>. I set <strong>IoU > 0.5<\/strong> as positive and <strong>IoU &lt; 0.3<\/strong> as negative. For samples that IoU between 0.5 and 0.3 are ignored   <\/p>\n\n<p>Can anyone help me with this? Many thanks!<\/p>\n","answers":[{"AnswerId":"52763181","CreationDate":"2018-10-11T14:55:53.277","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Usually the ratio between positive and negative examples is set to be 1:3. It could be that because in your model the ratio is set to 1:1, the model doesn't see enough negative examples.<\/p>\n\n<p>You can also try to do an error analysis and check what are the false positives which have high confidence. Maybe they are of a specific type, and if so - add negative examples of this type to your training data.<\/p>\n"}]}
+{"QuestionId":52762825,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-11T14:39:18.307","AcceptedAnswerId":52801261.0,"OwnerUserId":7732470.0,"Title":"How to train model created with tf.Keras model using tf.train?","Body":"<p>I am creating a GAN using Tensorflow. I decided to make Generator and Discriminator using Keras as layers became complex (since defining layers in Keras is easier). I have loss and training code for GAN in Tensorflow but I don't know how to train Keras model using that.<\/p>\n\n<pre><code>D_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_logit_real, labels=tf.ones_like(D_logit_real)))\nD_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_logit_fake, labels=tf.zeros_like(D_logit_fake)))\nD_loss = D_loss_real + D_loss_fake\nG_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_logit_fake, labels=tf.ones_like(D_logit_fake)))\n\nD_solver = tf.train.AdamOptimizer().minimize(D_loss,var_list = ?)\nG_solver = tf.train.AdamOptimizer().minimize(G_loss,var_list = ?)\n<\/code><\/pre>\n\n<p>I was passing <code>tf.Variables<\/code> (that were used in GAN when making with tf) in var_list. However, I don't know how to minimize the keras model since keras takes care of weights on its own.<br>\nI cannot use Keras's compile and fit function because Generator and Discriminator are too interdependent while training.<\/p>\n\n<p>How do I train my keras with loss as shown above?<\/p>\n","answers":[{"AnswerId":"52801261","CreationDate":"2018-10-14T09:25:18.873","ParentId":null,"OwnerUserId":"7732470","Title":null,"Body":"<p><a href=\"https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"nofollow noreferrer\">This article will help.<\/a> Basically, you just need to set your Keras session same as your Tensorflow.<\/p>\n"}]}
+{"QuestionId":52763072,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-11T14:51:04.557","AcceptedAnswerId":null,"OwnerUserId":9242453.0,"Title":"Initializing new object of operator TensorFlow","Body":"<p>I have graph something like this, operators are not important:<\/p>\n\n<pre><code>a = ...\nb = ...\nc = ...\nd = ...\n<\/code><\/pre>\n\n<p>And if I call run method like this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    for i in range(100):\n        if i != 99:\n            aa, bb = sess.run([a, b], feed_dict={ ... })\n        else:\n            aa, bb, cc = sess.run([a, b, c], feed_dict={ ... })\n<\/code><\/pre>\n\n<p>In if block run use same object of operators a and b, but in else branch\ncreates new objects, and that happens only if I have another node in list in run function. I want to stop that, but I don't know how to do that.<\/p>\n","answers":[]}
+{"QuestionId":52763225,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-11T14:57:44.760","AcceptedAnswerId":null,"OwnerUserId":5316628.0,"Title":"Keras predict_generator corrupted images","Body":"<p>I am trying to predict several million images with my trained model using a predict_generator in python 3 with keras and tensorflow as backend. The generator and the model predictions work, however, some images in the directory are broken or corrupted and cause the predict_generator to stop and throw an error. Once the image is removed it works again until the next corrupted\/broken image gets fed through the function. <\/p>\n\n<p>Since there are so many images it is not feasible to run a script to open every image and delete the ones that are throwing an error. Is there a way to incorporate a \"skip image if broken\" argument into the generator or flow from directory function?<br>\nAny help is greatly appreciated!<\/p>\n","answers":[{"AnswerId":"52763460","CreationDate":"2018-10-11T15:09:01.267","ParentId":null,"OwnerUserId":"8093922","Title":null,"Body":"<p>There's no such argument in <code>ImageDataGenerator<\/code> and neither in <code>flow_from_directory<\/code> method as you can see int the Keras docs for both (<a href=\"https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-class\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/keras.io\/preprocessing\/image\/#flow_from_directory\" rel=\"nofollow noreferrer\">here<\/a>). One workaround would be to extend the <code>ImageDataGenerator<\/code> class and overload the <code>flow_from_directory<\/code> method to check wether the image is corrupted or not before yeld it in the generator. <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/master\/keras_preprocessing\/image.py#L918\" rel=\"nofollow noreferrer\">Here<\/a> you can find it's source code.<\/p>\n"}]}
+{"QuestionId":52763506,"AnswerCount":0,"Tags":"<c++><tensorflow><conv-neural-network><convolution>","CreationDate":"2018-10-11T15:11:54.833","AcceptedAnswerId":null,"OwnerUserId":8477210.0,"Title":"Whether tensorflow defines USE_GEMM_FOR_CONV","Body":"<p>The implement of convolutional operation in <code>conv_ops.cc<\/code> involves <code>USE_GEMM_FOR_CONV<\/code><\/p>\n\n<pre><code>#if !defined(USE_GEMM_FOR_CONV)\nTF_CALL_half(REGISTER_CPU);\nTF_CALL_float(REGISTER_CPU);\nTF_CALL_double(REGISTER_CPU);\n#endif  \/\/ USE_GEMM_FOR_CONV\n<\/code><\/pre>\n\n<p>How can I determine whether the program defines <code>USE_GEMM_FOR_CONV<\/code>?<\/p>\n","answers":[]}
+{"QuestionId":52763539,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-11T15:13:47.133","AcceptedAnswerId":null,"OwnerUserId":9854132.0,"Title":"Compute the mean for each row from a tf.SparseTensor in TensorFlow","Body":"<p>I want to compute the mean for axis=0 for a <code>tf.SparseTensor<\/code>. I want something like <code>tf.sparse_reduce_sum<\/code>. TensorFlow doesn't provide a similar function for the mean calculation. Is there any way to count the values in each row in order to divide them with the sum?<\/p>\n\n<pre><code>indices = np.array([[0, 0], [0, 1], [0, 2], [0, 3], [0, 4], [0, 5],\n               [1, 0], [1, 1], [1, 3], [1, 4], [1, 5],\n               [2, 1], [2, 2], [2, 3], [2, 4],\n               [3, 0], [3, 1], [3, 2], [3, 3], [3, 4], [3, 5],\n               [4, 0], [4, 2], [4, 3], [4, 4], [4, 5]], dtype=np.int64)\n\nvalues = np.array([7, 6, 7, 4, 5, 4,\n                   6, 7, 4, 3, 4,\n                   3, 3, 1, 1,\n                   1, 2, 2, 3, 3, 4,\n                   1, 1, 2, 3, 3], dtype=np.float64)\n\ndense_shape = np.array([5, 6], dtype=np.int64)\n\ntRatings = tf.SparseTensor(indices, values, dense_shape)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52825181","CreationDate":"2018-10-15T21:45:17.200","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>You can compute the reduced mean from the reduced sum by dividing by the size of the 0th dimension:<\/p>\n\n<pre><code>tRatings = tf.SparseTensor(indices, values, dense_shape)\nreduced_sum = tf.sparse_reduce_sum(tRatings, 0)  # Sum of each row\nreduced_mean = reduced_sum \/ tf.cast(tRatings.dense_shape[0], tf.float64)  # Mean of each row\n<\/code><\/pre>\n"},{"AnswerId":"52763633","CreationDate":"2018-10-11T15:19:08.777","ParentId":null,"OwnerUserId":"9540764","Title":null,"Body":"<p>Try to use <code>get_shape()<\/code> and then multiply <code>shape[0] * shape[1]<\/code> this is the total number of elements<\/p>\n"}]}
+{"QuestionId":52763883,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><deep-learning>","CreationDate":"2018-10-11T15:32:05.343","AcceptedAnswerId":null,"OwnerUserId":8859758.0,"Title":"The output dimension of dense layer in tensorflow","Body":"<p>I am writing a Alexnet of Cifar10 Implementation on tensorflow. The following is the code for the last few layers<\/p>\n\n<pre><code>   ## pool5\n    pool5 = tf.nn.max_pool(conv5,\n                         ksize=[1, 3, 3, 1],\n                         strides=[1, 2, 2, 1],\n                         padding=\"SAME\",\n                         name='pool5')\n    print_activations(pool5)\n\n    # ## Flatten\n    # pool5=tf.contrib.layers.flatten(pool5, outputs_collections=None, scope=None)\n\n    ## FC1\n    fc1=tf.layers.dense(pool5, 4096, activation=tf.nn.relu,  trainable=True)\n    print_activations(fc1)\n\n    ## FC2\n    fc2=tf.layers.dense(fc1, 4096, activation=tf.nn.relu,  trainable=True)\n    print_activations(fc2)\n\n    ## Ouput\n    out1=tf.layers.dense(fc2, 10, activation=None,  trainable=True)\n<\/code><\/pre>\n\n<p>TO my understanding, the output of the last layer should be tensor with length 10. Considering batch size, it should be [None, 10]. However, when I return 'out1.shape.dims', I got [None, None, None, 10]. Anyone knows what I am missing? Thanks<\/p>\n","answers":[]}
+{"QuestionId":52764489,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><computer-vision><classification>","CreationDate":"2018-10-11T16:06:44.970","AcceptedAnswerId":null,"OwnerUserId":8867390.0,"Title":"Conversion of a conv layer to fc layer and vice vers sa?","Body":"<p>I just wanna know if it's possible to convert a conv layer to a fully connected one and then return back to the conv layer ?<\/p>\n","answers":[{"AnswerId":"52764637","CreationDate":"2018-10-11T16:14:12.120","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>It is just a matter of ensuring that the input is the correct shape. I assume you are using keras.<\/p>\n\n<pre><code>from tensorflow.keras.layers import Dense, Flatten, Conv2D, Reshape\n\n# Add a convolution to the network (previous layer called some_input)\n\nc1 = Conv2D(32, (3, 3), activation='relu', name='first_conv')(some_input)\n\n# Now reshape using 'Flatten'\nf1 = Flatten(name='flat_c1')(c1)\n\n# Now add a dense layer with 10 nodes\ndense1 = Dense(10, activation='relu', name='dense1')(f1)\n\n# Now add a dense layer, making sure it has the right number of nodes for my next conreshape8v layer.\ndense2 = Dense(784, activation='relu', name='dense2')(dense1)\nreshape2 = Reshape((7, 7, 16), name='reshape2')(dense2)\n\n#Now back to convolutions (up or down)\nc2 = Conv2D(16, kernel_size=(3, 3), activation='relu', \n                      name='conv2')(reshape2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52765201,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-11T16:47:47.830","AcceptedAnswerId":null,"OwnerUserId":6834335.0,"Title":"Keras IMDB Sentiment Analysis","Body":"<p>I am new to ML, and I am trying to use Keras for sentiment analysis on the IMDB dataset, based on a tutorial I found. The code below runs and gives an accuracy of around 90% on the test data. However, when I try to predict two easy sentences (one positive, one negative), it gives a value of around 0.50 for the positive and 0.73 for the negative, where it should give 0.71 for the positive and less than 0.1 for the negative, which are the results shown in the tutorial. <\/p>\n\n<p>Any ideas what the problem is?<\/p>\n\n<pre><code>from keras.datasets import imdb\nfrom keras.preprocessing import sequence\nfrom keras.models import *\nfrom keras.layers import *\nimport numpy as np\n\ntop_words = 5000  # 5000\n# first tuple is data and sentiment lists,\n# the second is testing data with sentiment\n# https:\/\/keras.io\/datasets\/\n(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=top_words)\n\n# reverse lookup\nword_to_id = imdb.get_word_index()\n'''word_to_id = {k: (v + INDEX_FROM) for k, v in word_to_id.items()}'''\nword_to_id[\"&lt;PAD&gt;\"] = 0\nword_to_id[\"&lt;START&gt;\"] = 1\nword_to_id[\"&lt;UNK&gt;\"] = 2\n\n# Truncate and pad the review sequences, to take only the first 500 words\nmax_review_length = 500\nx_train = sequence.pad_sequences(x_train, maxlen=max_review_length)\nx_test = sequence.pad_sequences(x_test, maxlen=max_review_length)\n\n# Build the model\n\n# embedding translates the words in a n dimensional\n# space so \"hi\" becomes (0.2,0.1,0.5) in a 3 dimensional space\n# it is the first layer of the network\nembedding_vector_length = 32  # dimensions\n\n# https:\/\/keras.io\/getting-started\/sequential-model-guide\/\nmodel = Sequential()  # sequential is a linear stack of layers\n\n# layer of 500 x 32\nmodel.add(\n    Embedding(\n        top_words,  # how many words to consider based on count\n        embedding_vector_length,  # dimensions\n        input_length=max_review_length))  # input vector\nmodel.add(LSTM(100))  # the parameter are the memory units of the LSTM\n# If you want you can replace LSTM by a flatten layer\n# model.add(LSTM(100))\n# model.add(Flatten())\nmodel.add(Dense(1, activation='sigmoid'))  # output 0&lt;y&lt;1 for every x\nmodel.compile(\n    loss='binary_crossentropy',\n    optimizer='adam',\n    metrics=['accuracy'])\nprint(model.summary())\n\n\n# Train the model\nmodel.fit(\n    x_train,\n    y_train,\n    validation_data=(x_test, y_test),\n    epochs=1)  # original epochs = 3, batch-size = 64\n\n# Evaluate the model\nscores = model.evaluate(x_test, y_test, verbose=0)\nprint(\"Accuracy: %.2f%%\" % (scores[1] * 100))\n\n# predict sentiment from reviews\nbad = \"this movie was terrible and bad\"\ngood = \"i really liked the movie and had fun\"\nfor review in [good, bad]:\n    tmp = []\n    for word in review.split(\" \"):\n        tmp.append(word_to_id[word])\n    tmp_padded = sequence.pad_sequences([tmp], maxlen=max_review_length)\n    print(\"%s. Sentiment: %s\" % (\n        review, model.predict(np.array([tmp_padded[0]]))))\n# i really liked the movie and had fun. Sentiment: 0.715537\n# this movie was terrible and bad. Sentiment: 0.0353295\n<\/code><\/pre>\n","answers":[{"AnswerId":"52772723","CreationDate":"2018-10-12T05:18:25.957","ParentId":null,"OwnerUserId":"9415043","Title":null,"Body":"<p>\"Any ideas what the problem is?\" There may not be a problem, per se. I have a few thoughts, in order of likely influence:<\/p>\n\n<ol>\n<li><p>If your two sentences are not representative of an IMDB review, then one can expect the model to predict poorly and erratically.<\/p><\/li>\n<li><p>Your model only has a single epoch, and the model may not have had enough opportunity to learn a robust mapping from review to sentiment (assuming such a mapping is possible given the data).<\/p><\/li>\n<li><p>There is a random element to neural networks, and as a result, the model you have developed will probably not predict identically to the model in the tutorial.<\/p><\/li>\n<li><p>With \"an accuracy of around 90%\", one would expect (depending on the class distribution) around 1 in 10 predictions to be incorrect. A small number of instances (two, in your case) are generally not a good way to evaluate the performance of a model.<\/p><\/li>\n<\/ol>\n\n<p>When I ran your code, I got a training accuracy of around 80% and test accuracy of around 85%, and \"i really liked the movie and had fun. Sentiment: [[0.75149596]]\" and \"this movie was terrible and bad. Sentiment: [[0.93544275]]\".<\/p>\n"}]}
+{"QuestionId":52765294,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-11T16:54:00.387","AcceptedAnswerId":53229018.0,"OwnerUserId":577905.0,"Title":"Unable to compile Tensorflow from source on Ubuntu 18.04","Body":"<p>When I attempt to compile TensorFlow from source I get the following error. I'm using the GPU Docker image to run the build. Which in theory has all of the proper dependencies set up. <\/p>\n\n<p>The host machine(s) is Ubuntu 18.04. I'm receiving this issue on two different machines that both have the latest Nvidia drivers. One has a 1080 TI and the other has a 1060 nvidia card.<\/p>\n\n<p>Building a docker image based on: tensorflow\/tensorflow:1.11.0-devel-gpu. <\/p>\n\n<p>I run it in interactive mode and run .\/configure. <\/p>\n\n<p>This is the configuration:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>.\/configure\r\nWARNING: --batch mode is deprecated. Please instead explicitly shut down your Bazel server using the command \"bazel shutdown\".\r\nYou have bazel 0.15.0 installed.\r\nPlease specify the location of python. [Default is \/usr\/bin\/python]: \r\n\r\n\r\nFound possible Python library paths:\r\n  \/usr\/local\/lib\/python2.7\/dist-packages\r\n  \/usr\/lib\/python2.7\/dist-packages\r\nPlease input the desired Python library path to use.  Default is [\/usr\/local\/lib\/python2.7\/dist-packages]\r\n\r\nDo you wish to build TensorFlow with jemalloc as malloc support? [Y\/n]: y\r\njemalloc as malloc support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with Google Cloud Platform support? [Y\/n]: y\r\nGoogle Cloud Platform support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with Hadoop File System support? [Y\/n]: y\r\nHadoop File System support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with Amazon AWS Platform support? [Y\/n]: y\r\nAmazon AWS Platform support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with Apache Kafka Platform support? [Y\/n]: y\r\nApache Kafka Platform support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with XLA JIT support? [y\/N]: y\r\nXLA JIT support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with GDR support? [y\/N]: y\r\nGDR support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with VERBS support? [y\/N]: y\r\nVERBS support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with nGraph support? [y\/N]: y\r\nnGraph support will be enabled for TensorFlow.\r\n\r\nDo you wish to build TensorFlow with OpenCL SYCL support? [y\/N]: n\r\nNo OpenCL SYCL support will be enabled for TensorFlow.\r\n\r\nPlease specify the location where CUDA 9.0 toolkit is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda]: \r\n\r\n\r\nPlease specify the location where cuDNN 7 library is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda]: \r\n\r\n\r\nPlease specify the location where TensorRT is installed. [Default is \/usr\/lib\/x86_64-linux-gnu]:\r\n\r\n\r\nPlease specify the location where NCCL 2 library is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda]:\r\n\r\n\r\nDo you want to use clang as CUDA compiler? [y\/N]: n\r\nnvcc will be used as CUDA compiler.\r\n\r\nPlease specify which gcc should be used by nvcc as the host compiler. [Default is \/usr\/bin\/gcc]: \r\n\r\n\r\nDo you wish to build TensorFlow with MPI support? [y\/N]: n\r\nNo MPI support will be enabled for TensorFlow.\r\n\r\nPlease specify optimization flags to use during compilation when bazel option \"--config=opt\" is specified [Default is -march=native]: \r\n\r\n\r\nWould you like to interactively configure .\/WORKSPACE for Android builds? [y\/N]: n\r\nNot configuring the WORKSPACE for Android builds.\r\n\r\nPreconfigured Bazel build configs. You can use any of the below by adding \"--config=&lt;&gt;\" to your build command. See tools\/bazel.rc for more details.\r\n        --config=mkl            # Build with MKL support.\r\n        --config=monolithic     # Config for mostly static monolithic build.\r\nConfiguration finished\r\nroot@17ed0ddbe2a4:\/tensorflow# <\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>This is the error I get.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>ERROR: \/root\/.cache\/bazel\/_bazel_root\/68a62076e91007a7908bc42a32e4cff9\/external\/ngraph_tf\/BUILD.bazel:19:1: C++ compilation of rule '@ngraph_tf\/\/:ngraph_tf' failed (Exit 1)\r\nIn file included from external\/org_tensorflow\/tensorflow\/core\/framework\/common_shape_fns.h:22:0,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/framework\/resource_mgr.h:24,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/device.h:43,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/device_set.h:23,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/optimization_registry.h:25,\r\n                 from external\/ngraph_tf\/src\/ngraph_encapsulate_pass.cc:23:\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h: In function 'tensorflow::TensorShape tensorflow::ShapeFromFormat(tensorflow::TensorFormat, tensorflow::int64, tensorflow::gtl::ArraySlice&lt;long long int&gt;, tensorflow::int64)':\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:501:45: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]\r\n     if (format == FORMAT_NHWC_VECT_W &amp;&amp; dim == spatial.size() - 1) {\r\n                                             ^\r\nexternal\/ngraph_tf\/src\/ngraph_encapsulate_pass.cc: In member function 'tensorflow::Status ngraph_bridge::NGraphEncapsulatePass::EncapsulateFunctions(tensorflow::Graph*)':\r\nexternal\/ngraph_tf\/src\/ngraph_encapsulate_pass.cc:393:17: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]\r\n           if (i &lt; node-&gt;requested_inputs().size()) {\r\n                 ^\r\nexternal\/ngraph_tf\/src\/ngraph_encapsulate_pass.cc:414:42: error: 'class absl::string_view' has no member named 'ToString'\r\n             cluster_idx, tensor_id.first.ToString(), tensor_id.second));\r\n                                          ^\r\nIn file included from external\/org_tensorflow\/tensorflow\/core\/platform\/default\/logging.h:24:0,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/platform\/logging.h:25,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/lib\/core\/refcount.h:22,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/platform\/tensor_coding.h:21,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/framework\/resource_handle.h:19,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/framework\/allocator.h:24,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/device.h:35,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/device_set.h:23,\r\n                 from external\/org_tensorflow\/tensorflow\/core\/common_runtime\/optimization_registry.h:25,\r\n                 from external\/ngraph_tf\/src\/ngraph_encapsulate_pass.cc:23:\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h: In instantiation of 'T tensorflow::GetTensorDim(tensorflow::gtl::ArraySlice&lt;T&gt;, tensorflow::TensorFormat, char) [with T = long long int; tensorflow::gtl::ArraySlice&lt;T&gt; = absl::Span&lt;const long long int&gt;]':\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:452:47:   required from here\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:420:29: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]\r\n   CHECK(index &gt;= 0 &amp;&amp; index &lt; dimension_attributes.size())\r\n                             ^\r\nexternal\/org_tensorflow\/tensorflow\/core\/platform\/macros.h:87:47: note: in definition of macro 'TF_PREDICT_FALSE'\r\n #define TF_PREDICT_FALSE(x) (__builtin_expect(x, 0))\r\n                                               ^\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:420:3: note: in expansion of macro 'CHECK'\r\n   CHECK(index &gt;= 0 &amp;&amp; index &lt; dimension_attributes.size())\r\n   ^\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h: In instantiation of 'T tensorflow::GetFilterDim(tensorflow::gtl::ArraySlice&lt;T&gt;, tensorflow::FilterTensorFormat, char) [with T = long long int; tensorflow::gtl::ArraySlice&lt;T&gt; = absl::Span&lt;const long long int&gt;]':\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:461:54:   required from here\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:435:29: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]\r\n   CHECK(index &gt;= 0 &amp;&amp; index &lt; dimension_attribute.size())\r\n                             ^\r\nexternal\/org_tensorflow\/tensorflow\/core\/platform\/macros.h:87:47: note: in definition of macro 'TF_PREDICT_FALSE'\r\n #define TF_PREDICT_FALSE(x) (__builtin_expect(x, 0))\r\n                                               ^\r\nexternal\/org_tensorflow\/tensorflow\/core\/util\/tensor_format.h:435:3: note: in expansion of macro 'CHECK'\r\n   CHECK(index &gt;= 0 &amp;&amp; index &lt; dimension_attribute.size())\r\n   ^\r\nTarget \/\/tensorflow\/tools\/pip_package:build_pip_package failed to build<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I get the same problem with a 1.10 image too. Or trying to build either 1.10 or 1.11 from source directly on the host.<\/p>\n","answers":[{"AnswerId":"53229018","CreationDate":"2018-11-09T15:49:25.200","ParentId":null,"OwnerUserId":"3411080","Title":null,"Body":"<p>it seems it was solved on this commit <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/08af8cac22af4cc430e092b6218ca77736efb82c\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Check also the thread where it was mentioned in the first place <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22583\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22583<\/a><\/p>\n\n<p>Just pull from master and try to rebuild again.<\/p>\n"}]}
+{"QuestionId":52765373,"AnswerCount":2,"Tags":"<python><performance><tensorflow><keras><runtime>","CreationDate":"2018-10-11T16:59:25.650","AcceptedAnswerId":52765695.0,"OwnerUserId":6240174.0,"Title":"Why is the per sample prediction time on Tensorflow (and Keras) lower when predicting on batches than on individual samples?","Body":"<p>I am using my trained model to make predictions (CPU only). I observe that both on Tensorflow and Keras with Tensorflow backend, the prediction time per sample is much lower when a batch of samples is used as compared to an individual sample. Moreover, the time per sample seems to go down with increasing batch size up to the limits imposed by memory. As an example, on pure Tensorflow, prediction of a single sample takes  ~ 1.5 seconds , on 100 samples it is ~ 17 seconds (per sample time ~ 0.17s) on 1000 samples it is ~ 93 seconds (the per sample time  ~ 0.093s). <\/p>\n\n<p>Is this normal behavior?  If so, is there an intuitive explanation for this?  I guess it might have something to do with initializing the graph, but I need some clarification. Also, why does the per sample time go down as we increase the number of samples for prediction? In my use case, I have to predict on individual samples as and when they become available. So, obviously, I would be losing quite a bit in terms of speed if this is the way things work. <\/p>\n\n<p>Thanks in advance for your help. <\/p>\n\n<p>Edit: I am adding a minimal working example. I have one image input and 4 vector inputs to my model, which produces 4 outputs. I am initializing all inputs to 0 for speed test (I guess the actual values don't matter much for speed?). The initialization time and inference time are calculated separately. I find that the initialization time is a fraction of the inference time (~0.1s for 100 samples). <\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\nimport time\nimport numpy as np\nimport tensorflow as tf\n\nt00=time.time()\ngraph = tf.Graph()\ngraph_def = tf.GraphDef()\n\nwith open(\"output_graph.pb\", \"rb\") as f:\n    graph_def.ParseFromString(f.read())\nwith graph.as_default():\n    tf.import_graph_def(graph_def)\n\n# One image and 4 auxiliary scalar inputs\nimg_input_layer =\"input\"\nqp4_input_layer =\"qp4\"\nqp3_input_layer =\"qp3\"\nqp2_input_layer =\"qp2\"\nqp1_input_layer =\"qp1\"\ninput_name = \"import\/\" + img_input_layer\nqp4_input_name = \"import\/\" + qp4_input_layer\nqp3_input_name = \"import\/\" + qp3_input_layer\nqp2_input_name = \"import\/\" + qp2_input_layer\nqp1_input_name = \"import\/\" + qp1_input_layer\n\n\n\ninput_operation_img = graph.get_operation_by_name(input_name)\ninput_operation_qp4 = graph.get_operation_by_name(qp4_input_name)\ninput_operation_qp3 = graph.get_operation_by_name(qp3_input_name)\ninput_operation_qp2 = graph.get_operation_by_name(qp2_input_name)\ninput_operation_qp1 = graph.get_operation_by_name(qp1_input_name)\n\noutput_operation=[]\n\nfor i in range(4):\n    output_operation.append(graph.get_operation_by_name(\"import\/\" + \"output_\"+str(i)).outputs)\n\n#Initializing dummy inputs\nn=100 # Number of samples for inference\nimg=np.zeros([n,64, 64,1])\nqp4=np.zeros([n,1, 1,1])\nqp3=np.zeros([n,2, 2,1])\nqp2=np.zeros([n,4, 4,1])\nqp1=np.zeros([n,8, 8,1])\nt01=time.time()\nprint(\"Iniialization time\",t01-t00)\n\nt0=time.time()\nwith tf.Session(graph=graph) as sess:\n    results = sess.run(output_operation,\n                       {input_operation_img.outputs[0]: img, input_operation_qp4.outputs[0]: qp4, input_operation_qp3.outputs[0]: qp3,  input_operation_qp2.outputs[0]: qp2,  input_operation_qp1.outputs[0]: qp1})\n\n    # print(results)\nt1 = time.time()\nprint(\"Inference time\", t1-t0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52765677","CreationDate":"2018-10-11T17:16:05.493","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, its completely normal. This happens because as you do inference on GPU (or even multicore CPUs), increasing the batch size allows better use of the GPUs parallel computation resources, decreasing time per sample in the batch. If you use a small batch size, then you are wasting computational resources that are available in the GPU.<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1605.07678\" rel=\"nofollow noreferrer\">This paper<\/a> describes the same effect, and one of the Figures contains a plot that shows inference time per image versus batch size. It shows the same effect as you are seeing.<\/p>\n"},{"AnswerId":"52765695","CreationDate":"2018-10-11T17:16:55.310","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>This depends very much on the model instrumentation, deployment method, and interface -- none of which you've provided, or even described.  In my practice, the common reasons include:<\/p>\n\n<ul>\n<li>Model initialization time: do you \"wake up\" the model in some way for each batch?  If, as you've suggested, you reinitialize the model for each request, then I'm somewhat surprised that the overhead isn't a larger proportion of your time.<\/li>\n<li>Interface overhead: how do the samples get to\/from the model?  Is this in an HTTP request, where you suffer communication costs per request, rather than per sample?<\/li>\n<li>Simple model I\/O time: If your model reads the entire batch at once, then the lag time to open and access the input channel could be the main factor in the lag.<\/li>\n<\/ul>\n\n<p>You have some work to do to diagnose the root causes.  You have only three data points so far; I recommend gathering a few more. What can you fit to a graph of the times?  Are there any jumps that suggest a system limitation, such as input buffer size?  Can you instrument your model with some profiling code to find out what lags are in your model, and which are in the system overhead?<\/p>\n\n<p>Try deploying your model as a service; what happens to the times when it's already loaded into memory, initialized, and merely waiting for the next input?  What lag time do you have in your request interface?<\/p>\n\n<p>The results of these investigations will show you where you might gain from a design change in your usage model.<\/p>\n"}]}
+{"QuestionId":52765841,"AnswerCount":1,"Tags":"<python><python-3.x><macos><tensorflow><conv-neural-network>","CreationDate":"2018-10-11T17:27:22.520","AcceptedAnswerId":52766643.0,"OwnerUserId":6063708.0,"Title":"How do I optimize Tensorflow CNN?","Body":"<p>I'm very new to Tensorflow, so I apologize if my question comes off as ignorant.<\/p>\n\n<p>I have a very simple CNN Tensorflow that takes images and outputs another image. With just a batchsize of 5, it takes minutes to run between epochs and often crashes after 5 epochs.(I'm using python 3.6.5 on my mac with 16 gbs of RAM)<\/p>\n\n<p>This is a snippet of my program<\/p>\n\n<pre><code>learning_rate = 0.01\ninputs_ = tf.placeholder(tf.float32, (None, 224, 224, 3), name='inputs')\ntargets_ = tf.placeholder(tf.float32, (None, 224, 224, 1), name='targets')\n### Encoder\nconv1 = tf.layers.conv2d(inputs=inputs_, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 224x224x32\nmaxpool1 = tf.layers.max_pooling2d(conv1, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 112x112x32\n\nconv2 = tf.layers.conv2d(inputs=maxpool1, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 112x112x32\nmaxpool2 = tf.layers.max_pooling2d(conv2, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 56x56x32\n\nconv3 = tf.layers.conv2d(inputs=maxpool2, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 56x56x32\nmaxpool3 = tf.layers.max_pooling2d(conv3, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 28x28x32\n\nconv4 = tf.layers.conv2d(inputs=maxpool3, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 28x28x32\nmaxpool4 = tf.layers.max_pooling2d(conv4, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 14x14x32\nconv5 = tf.layers.conv2d(inputs=maxpool4, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 14x14x32\nmaxpool5 = tf.layers.max_pooling2d(conv5, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 7x7x32\nconv6 = tf.layers.conv2d(inputs=maxpool5, filters=16, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 7x7x16\nencoded = tf.layers.max_pooling2d(conv6, pool_size=(2,2), strides=(2,2), padding='same')\n# Now 4x4x16\n\n### Decoder\nupsample1 = tf.image.resize_images(encoded, size=(7,7), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 7x7x16\nconv7 = tf.layers.conv2d(inputs=upsample1, filters=16, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 7x7x16\nupsample2 = tf.image.resize_images(conv7, size=(14,14), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 14x14x16\nconv8 = tf.layers.conv2d(inputs=upsample2, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 14x14x32\nupsample3 = tf.image.resize_images(conv8, size=(28,28), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 28x28x32\nconv9 = tf.layers.conv2d(inputs=upsample3, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 28x28x32\n\nupsample4 = tf.image.resize_images(conv9, size=(56,56), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 56x56x32\nconv10 = tf.layers.conv2d(inputs=upsample3, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 56x56x32\n\nupsample5 = tf.image.resize_images(conv10, size=(112,112), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 112x112x32\nconv11 = tf.layers.conv2d(inputs=upsample5, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 112x112x32\n\nupsample6 = tf.image.resize_images(conv11, size=(224,224), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)\n# Now 224x224x32\nconv12 = tf.layers.conv2d(inputs=upsample6, filters=32, kernel_size=(3,3), padding='same', activation=tf.nn.relu)\n# Now 224x224x32\n\nlogits = tf.layers.conv2d(inputs=conv12, filters=1, kernel_size=(3,3), padding='same', activation=None)\n#Now 224x224x1\n# Pass logits through sigmoid and calculate the cross-entropy loss\nloss = tf.nn.sigmoid_cross_entropy_with_logits(labels=targets_, logits=logits)\n\n# Get cost and define the optimizer\ncost = tf.reduce_mean(loss)\nopt = tf.train.AdamOptimizer(learning_rate).minimize(cost)\n\nimagelist = ... #array of all images with 3 channels\nimagelabellist = ... #array of all images with 1 channel\nepochs = 15\n\nfor e in range(epochs):\n            imgs_large = imagelist\n            imgs_target_large = imagelabellist\n            shaped_imgs = tf.image.resize_images(imgs_large, [224, 224])\n            shaped_imgs_target = tf.image.resize_images(imgs_target_large, [224, 224])\n            # Get images from the batch\n            imgs = sess.run(shaped_imgs)\n            imgs_target = sess.run(shaped_imgs_target)\n            batch_cost, _ = sess.run([cost, opt], feed_dict={inputs_: imgs, targets_: imgs_target})\n<\/code><\/pre>\n\n<p>This is the output of the CNN<\/p>\n\n<pre><code>epoch: #1\n0 minutes between epoch\nepoch: #2\n3 minutes between epoch\nepoch: #3\n3 minutes between epoch\nepoch: #4\n12 minutes between epoch\nepoch: #5\n<\/code><\/pre>\n\n<p>...<\/p>\n\n<p>I'm open to any suggestions on how to fix this issue. Thank you.<\/p>\n","answers":[{"AnswerId":"52766643","CreationDate":"2018-10-11T18:21:10.753","ParentId":null,"OwnerUserId":"1322179","Title":null,"Body":"<p><code>tf.image.resize_images<\/code> is a graph op, so you're appending more nodes to the graph (that explains the increasing run time). Before your training loop add <code>sess.graph.finalize()<\/code> if nodes are being added it will throw an error to check this.<\/p>\n\n<p>If you move <code>resize_images<\/code> outside of the loop, that should fixed the issue.<\/p>\n"}]}
+{"QuestionId":52766175,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2018-10-11T17:49:04.070","AcceptedAnswerId":null,"OwnerUserId":8584587.0,"Title":"CNNs generalizing to same class","Body":"<p>So I'am using tensorflow 1.10, and training on subset of retinopathy dataset. The problem is that it always predicts the class that occurs the most. In my case it is the class 0. so, I did a little digging and came across something called under-sampling. I left out all the 0(just to see what happens) it only predicts the class 2 after 0. apparently class 2 has highest frequency. \nHere is the code for optimization:<\/p>\n\n<pre><code>def data_pipe_line(data,checkpoint_path,i_data=None,epoch=5):\n\nplace_X=tf.placeholder(tf.float32,[None,500,400,3],name='p1')\nplace_Y=tf.placeholder(tf.int32,[None],name='p2')\ninfer_data=tf.data.Dataset.from_tensor_slices((place_X,place_Y))\ninfer_data=infer_data.batch(100)\niterator=tf.data.Iterator.from_structure(data.output_types,data.output_shapes)\nnext_image,next_label=iterator.get_next()\nY=tf.one_hot(next_label,5)\nY=tf.cast(Y,tf.float32)\nlogits=model(next_image,0.7)\nprint(logits)\nprint(Y)\ntrain_iterator=iterator.make_initializer(data,name='train_op')\ninference_iterator_op=iterator.make_initializer(infer_data,name='inference_op')\n\n\n\nwith tf.name_scope(\"loss\"):\n    loss=tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y,logits=logits),name='cost')\n    #the learning rate is so low because the batch-size is very small and has a lot of noise\n    optimizer=tf.train.AdamOptimizer(learning_rate=0.0005).minimize(loss)\n    #getting the accuracy\n    prediction=tf.argmax(logits,1,name='pred')\n    equality=tf.equal(prediction,tf.argmax(Y,1))\n    accuracy=tf.reduce_mean(tf.cast(equality,tf.float32))\n    init_op=tf.global_variables_initializer()\n    tf.summary.scalar(\"loss\",loss)\n    tf.summary.scalar(\"accuracy\",accuracy)\n    merged=tf.summary.merge_all()\n    saver=tf.train.Saver()\n\n\n\nj=0\nwith tf.Session() as sess:\n    writer=tf.summary.FileWriter(\".\/nn_logs\",sess.graph)\n    sess.run(init_op)\n    for _ in range(epoch):\n        sess.run(train_iterator)\n        while True:\n            try:\n                #print(sess.run(logits))\n                j=j+1 \n                summary = sess.run(merged)\n                _,acc,l=sess.run([optimizer,accuracy,loss]) \n                if(j%20==0 or j==1):\n                    print(\"iters: {}, loss: {:.10f}, training accuracy: {:.2f}\".format(j, l, acc*100))\n                writer.add_summary(summary,j)\n            except tf.errors.OutOfRangeError:\n\n                break\n    saver.save(sess,checkpoint_path)\n<\/code><\/pre>\n\n<p>The model is training well, loss goes down for sometime and then it just fluctuates there(in a range of 5). the accuracy of course fluctuates a lot as it is only predicting 1 class. <\/p>\n","answers":[{"AnswerId":"52766474","CreationDate":"2018-10-11T18:09:56.067","ParentId":null,"OwnerUserId":"8093922","Title":null,"Body":"<p>By what you said, I can conclude that your dataset is highly unbalanced and then the model that you trained didn't generalized your data at all. The model then predicts the class with most samples, because at a random guess there's more chance that it would be the correct label. To try to solve this, one approach is to balance your data. <\/p>\n\n<h1>Under Sampling<\/h1>\n\n<p>One way to do that is using the technique called under sampling, as you have finded out. In your training set you verify the class that has less samples, and then you pass to the training phase that same number of samples of each class. The selection is done at random. The test set remains intact.<\/p>\n\n<h1> Over Sampling<\/h1>\n\n<p>Here you take the contrary way: take the class with more samples and make all other classes have that same amount of samples. The simplest way to do that is repeating at random your samples. Another way would be augmenting your data.<\/p>\n\n<p>You can also can weight the loss function. For this, G. Anderson already linked a good <a href=\"https:\/\/datascience.stackexchange.com\/questions\/17219\/how-to-do-imbalanced-classification-in-deep-learning-tensorflow-rnn\">question<\/a> about.<\/p>\n\n<h1>The intuiton behind weighted the loss<\/h1>\n\n<p>When the classes are unbalanced, the model understands that if it get the wrong label for the class with less samples it will be poorly penalized for that mistake. Once they have less samples of that class, that mistake will occur less often and when it occur it will have the same impact of a mistake of one class with more samples. So what the weighted loss do is give more importance for a mistake in the class with less samples and less importance for a mistake in the class with more samples. <\/p>\n\n<p><strong>Answering the question in the commentary<\/strong><\/p>\n\n<p>Remember: the loss function is used to correct the learned model weights and, when you add another dense layer in your model, you are doing changes that will affect the learning step. So your model will still attributing the same importance level to the mistakes in every class.<\/p>\n"}]}
+{"QuestionId":52767007,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-11T18:47:31.767","AcceptedAnswerId":null,"OwnerUserId":10491536.0,"Title":"Importing tensorflow not working when upgraded","Body":"<p><code>Tensorflow<\/code> was working fine when I had <code>1.4<\/code> but when I upgraded it, it stopped working.<\/p>\n\n<p>The version that I installed is <code>1.11<\/code> with <code>CUDA 9<\/code> and <code>cuDNN 7<\/code>. <\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\",\n  line 243, in load_module\n      return load_dynamic(name, filename, file)   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\",\n  line 343, in load_dynamic\n      return _load(spec) ImportError: DLL load failed: The specified module could not be found.<\/p>\n<\/blockquote>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"\", line 1, in\n     File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow__init__.py\",\n  line 22, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import   File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python__init__.py\",\n  line 49, in \n      from tensorflow.python import pywrap_tensorflow   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 74, in \n      raise ImportError(msg) ImportError: Traceback (most recent call last):   File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File\n  \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\",\n  line 243, in load_module\n      return load_dynamic(name, filename, file)   File \"C:\\Users\\anime\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\",\n  line 343, in load_dynamic\n      return _load(spec) ImportError: DLL load failed: The specified module could not be found.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58498554","CreationDate":"2019-10-22T06:50:29.177","ParentId":null,"OwnerUserId":"11759240","Title":null,"Body":"<p>you can uninstall the tensorflow and re-install it.<\/p>\n"}]}
+{"QuestionId":52767342,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2018-10-11T19:12:35.907","AcceptedAnswerId":null,"OwnerUserId":10246793.0,"Title":"Error in function 'fwd_prop_enc' :attempt to call global 'forward' (a nil value)","Body":"<p>In Machine Translation Dataset I have successfully pre-trained and trained my model in Lua. Now I move to predict my model.<\/p>\n\n<p>But I get the error in a Lua file in the function encoder_clones:forward(encoder_input)\nError:  attempt to call global 'forward' (a nil value)<\/p>\n\n<p>This is that specific function :<\/p>\n\n<pre><code>function fwd_prop_enc(source, source_l, batch_l, train)\nlocal rnn_state_enc = reset_state(init_fwd_enc, batch_l, 0)\n--print(rnn_state_enc)\nlocal context = context_proto[{{1, batch_l}, {1, source_l}}]\nfor t = 1, source_l do\nif train then\n  encoder_clones[t]:training()\nelse\n  encoder_clones[t]:evaluate()\nend\nlocal encoder_input = {source[t], unpack(rnn_state_enc[t-1])}\nlocal out = encoder_clones:forward(encoder_input)\nprint(\"yes\")\nrnn_state_enc[t] = out\ncontext[{{},t}]:copy(out[#out]) -- copy final layer for t'th timestep    (for entire batch)\n end\n  return rnn_state_enc, context\nend\n<\/code><\/pre>\n\n<p>This is how it is called in the main function<\/p>\n\n<pre><code>local rnn_state_enc, context = fwd_prop_enc(source, source_l, 1, false) -- train=false\n<\/code><\/pre>\n","answers":[{"AnswerId":"52893141","CreationDate":"2018-10-19T13:16:42.180","ParentId":null,"OwnerUserId":"8076767","Title":null,"Body":"<p>That error means that the function <code>encoder_clones.forward<\/code> doesn't actually exist.<\/p>\n\n<p>I do not know what framework you are using, but a <a href=\"https:\/\/www.google.com\/search?rlz=1C1CHBF_enUS819US819&amp;ei=W9jJW_6QObHL0PEPvMmZmAM&amp;q=lua%20encoder_clones&amp;oq=lua%20encoder_clones&amp;gs_l=psy-ab.3...9471.9901..10373...0.0..0.120.458.0j4......0....1..gws-wiz.iqv7vkjBHto\" rel=\"nofollow noreferrer\">quick search<\/a> revealed that <code>encoder_clones<\/code> is probably an array. Looking at your code, I noticed that your reference to <code>encoder_clones<\/code> is inside a for loop <code>for t = 1, source_l do<\/code>. What happens if you change <code>encoder_clones:forward<\/code> to <code>encoder_clones[t]:forward<\/code>? This will probably solve your problem.<\/p>\n\n<p>If this does not solve your problem, can you tell us what framework you are using? Is it OpenNMT? We can't help you much further unless we know what <code>encoder_clones<\/code> is and where it came from.<\/p>\n"}]}
+{"QuestionId":52767380,"AnswerCount":0,"Tags":"<python><python-2.7><tensorflow><convolution>","CreationDate":"2018-10-11T19:15:01.407","AcceptedAnswerId":null,"OwnerUserId":4355878.0,"Title":"Can I use `tf.layers.conv3d` with different image sizes? Use 'None' instead of a fix integer?","Body":"<p>The following script does not work: <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np \ninput = np.ones((1,1, 3,50,50))\nt_ph = tf.placeholder(dtype=tf.float32, shape=(1,1,3,None,50), name='input')\n# t_ph = tf.placeholder(dtype=tf.float32, shape=(1,1,3,50,50), name='input')\nt_conv3d = tf.layers.conv3d(\n    inputs=t_ph,\n    filters=48,\n    kernel_size=[3,3,3],\n    data_format='channels_first',\n    activation=tf.nn.relu\n)\n\nwith tf.Session() as sess: \n    sess.run([tf.local_variables_initializer(), tf.global_variables_initializer()])\n\n    should_fail = t_conv3d.eval(feed_dict={t_ph: input})\n    print \"should_fail \", should_fail.shape\n<\/code><\/pre>\n\n<p>Error: <code>TypeError: unsupported operand type(s) for *: 'int' and 'NoneType'\n<\/code><\/p>\n\n<p>Is it possible to have a conv3d with the parameter 'None' instead of an integer?<\/p>\n","answers":[]}
+{"QuestionId":52767628,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-11T19:32:46.963","AcceptedAnswerId":null,"OwnerUserId":1738210.0,"Title":"how can I build my database for learning a network in keras?","Body":"<p>I have some images that have been attacked with different kind of attacks like JPEG\u060c salt &amp; pepper,... now, I want to design a network that is fed with these images and their label and said what kind of attacks is done. but I do not know how can I make my database and prepare my training data? could you please help me with this issue? thanks in advance.\n    I used the below code but It produces the following error and I do not know how can I solve this problem. please guide me about this issue.\nI add this code but when I implement it, it produces this error: ValueError: Error when checking input: expected conv2d_37_input to have shape (512, 512, 1) but got array with shape (512, 512, 3)\n    import numpy as np\n    import matplotlib.pyplot as plt\n    from keras.preprocessing.image import ImageDataGenerator\n    from sklearn.utils import shuffle\n    from sklearn.cross_validation import train_test_split\n    import keras\n    from keras import backend as K\n    #K.set_image_dim_ordering('th')<\/p>\n\n<pre><code>from keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom keras.optimizers import SGD,RMSprop,adam\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (3,3),border_mode='same', input_shape = (512, 512, 1), activation = 'relu'))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(32, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Conv2D(64, 3, 3))\nmodel.add(Activation('relu'))\n\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(7))\nmodel.add(Activation('softmax'))\n\nmodel.summary()\n\nmodel.compile(loss = 'categorical_crossentropy',\n              optimizer = 'rmsprop',\n              metrics = ['accuracy'])\ntrain_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(\n        r'E:\\PhD\\thesis\\deepwatermark\\databasetest\\train',\n        target_size=(512, 512),\n        batch_size=32,\n        class_mode='binary')\nvalidation_generator = test_datagen.flow_from_directory(\n        r'E:\\PhD\\thesis\\deepwatermark\\databasetest\\validation',\n        target_size=(512, 512),\n        batch_size=32,\n        class_mode='binary')\nmodel.fit_generator(\n        train_generator,\n        steps_per_epoch=20,\n        epochs=50,\n        validation_data=validation_generator,\n        validation_steps=8)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52774813","CreationDate":"2018-10-12T07:52:34.057","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>There are multiple ways of handling datasets, in keras you can feed the network with a numpy array or a generator. <\/p>\n\n<p>Detailed documentation on generators : <\/p>\n\n<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-class\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-class<\/a>\n<a href=\"https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-methods\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-methods<\/a><\/p>\n\n<p>If you are going to use generators, before you do anything, separate your images by their labels into different directories. File hierarchy should look like<\/p>\n\n<ul>\n<li>TrainingData<\/li>\n<li><ul>\n<li>Label1<\/li>\n<\/ul><\/li>\n<li><ul>\n<li><ul>\n<li>image1.jpg<\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<li><ul>\n<li><ul>\n<li>image2.jpg<\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<li><ul>\n<li>Label2<\/li>\n<\/ul><\/li>\n<li><ul>\n<li><ul>\n<li>image1.jpg<\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<li><ul>\n<li><ul>\n<li>image2.jpg<\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>If you are going to store all your data into a numpy array, I suggest putting all your data into a csv file and their labels should be the first or last column. I don't recommend this for big datasets since you may have not enough memory for all of your data.<\/p>\n\n<p>Sample:\n    model = \n    # Create preprocessor for train(map 0-255 rgb values to 0-1 , randomly zoom, \n    shear and flip to artificially increase data-size )\n    train_datagen = ImageDataGenerator(\n    rescale=1.\/255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)<\/p>\n\n<pre><code>#Create validation preprocessor(map 0-255 rgb values to 0-1)\nvalidation_datagen = ImageDataGenerator(rescale=1.\/255)\n\n#Create train generator\ntrain_generator = train_datagen.flow_from_directory(\n    'data\/train',\n    target_size=(150, 150),\n    batch_size=32,\n    class_mode='categorical')\n#Create validation generator\nvalidation_generator = \n    validation_datagen.flow_from_directory('data\/validation',\n    target_size=(150, 150),\n    batch_size=32,\n    class_mode='categorical')\n#fit your model using the data from generators.\nmodel.fit_generator(\n    train_generator,\n    steps_per_epoch=2000,\n    epochs=50,\n    validation_data=validation_generator,\n    validation_steps=800)\n<\/code><\/pre>\n"},{"AnswerId":"52775274","CreationDate":"2018-10-12T08:21:59.047","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I would recommend using generators because they are easy to write and are safe for running your code. As for organizing the data, the answer from @Mete is ok, but consider also making a csv with image names in one column and type of attack in other. Then you just go through the csv file row by row, doing whatever you need to do with it then. <\/p>\n"}]}
+{"QuestionId":52767807,"AnswerCount":1,"Tags":"<tensorflow><cluster-analysis>","CreationDate":"2018-10-11T19:45:16.990","AcceptedAnswerId":null,"OwnerUserId":2927842.0,"Title":"Normalized Mutual Information in Tensorflow","Body":"<p>Is that possible to implement normalized mutual information in Tensorflow? I was wondering if I can do that and if I will be able to differentiate it. Let's say that I have predictions P and labels Y in two different tensors. Is there an easy way to use normalized mutual information?<\/p>\n\n<p>I want to do something similar to this:<\/p>\n\n<p><a href=\"https:\/\/course.ccs.neu.edu\/cs6140sp15\/7_locality_cluster\/Assignment-6\/NMI.pdf\" rel=\"nofollow noreferrer\">https:\/\/course.ccs.neu.edu\/cs6140sp15\/7_locality_cluster\/Assignment-6\/NMI.pdf<\/a><\/p>\n","answers":[{"AnswerId":"52809295","CreationDate":"2018-10-15T03:15:30.707","ParentId":null,"OwnerUserId":"7818309","Title":null,"Body":"<p>Assume your clustering method gives probability predictions\/membership functions <code>p(c|x)<\/code>, e.g., <code>p(c=1|x)<\/code> is the probability of <code>x<\/code> in the first cluster. Assume y is the ground truth class label for <code>x<\/code>.<\/p>\n\n<p>The normalized mutual information is <a href=\"https:\/\/i.stack.imgur.com\/O1sTH.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O1sTH.gif\" alt=\"enter image description here\"><\/a>.<\/p>\n\n<p>The entropy H(Y) can be estimated following this thread: <a href=\"https:\/\/stats.stackexchange.com\/questions\/338719\/calculating-clusters-entropy-python\">https:\/\/stats.stackexchange.com\/questions\/338719\/calculating-clusters-entropy-python<\/a><\/p>\n\n<p>By definition, the entropy H(C) is <a href=\"https:\/\/i.stack.imgur.com\/mN1bd.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mN1bd.gif\" alt=\"enter image description here\"><\/a>, where <a href=\"https:\/\/i.stack.imgur.com\/HHwor.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HHwor.gif\" alt=\"enter image description here\"><\/a>.<\/p>\n\n<p>The conditional mutual information <a href=\"https:\/\/i.stack.imgur.com\/FfxrR.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FfxrR.gif\" alt=\"enter image description here\"><\/a> where <a href=\"https:\/\/i.stack.imgur.com\/9wmBH.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9wmBH.gif\" alt=\"enter image description here\"><\/a>, and <a href=\"https:\/\/i.stack.imgur.com\/acj7E.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/acj7E.gif\" alt=\"enter image description here\"><\/a>.<\/p>\n\n<p>All terms involving integral can be estimated using sampling, i.e., average over training samples. The overall NMI is differentiable. <\/p>\n\n<p>I did not misunderstand your question. I was assuming you used a neural network model which outputs logits as you did not provide any info. Then you need to normalise the logits to get <code>p(c|x)<\/code>.<\/p>\n\n<p>There may be other ways to estimate NMI, but if you discretize the output of whatever model you use, you cannot differentiate them. <\/p>\n\n<hr>\n\n<p>TensorFlow code<\/p>\n\n<p>Assume we have label matrix <code>p_y_on_x<\/code> and cluster predictions <code>p_c_on_x<\/code>. Each row of them corresponds to an observation <code>x<\/code>; each column corresponds to the probability of <code>x<\/code> in each class and cluster (so each row sums up to one). Further assume uniform probability for <code>p(x)<\/code> and <code>p(x|y)<\/code>.<\/p>\n\n<p>Then NMI can then be estimated as below:<\/p>\n\n<pre><code>p_y = tf.reduce_sum(p_y_on_x, axis=0, keepdim=True) \/ num_x  # 1-by-num_y\nh_y = -tf.reduce_sum(p_y * tf.math.log(p_y))\np_c = tf.reduce_sum(p_c_on_x, axis=0) \/ num_x  # 1-by-num_c\nh_c = -tf.reduce_sum(p_c * tf.math.log(p_c))\np_x_on_y = p_y_on_x \/ num_x \/ p_y  # num_x-by-num_y\np_c_on_y = tf.matmul(p_c_on_x, p_x_on_y, transpose_a=True)  # num_c-by-num_y\nh_c_on_y = -tf.reduce_sum(tf.reduce_sum(p_c_on_y * tf.math.log(p_c_on_y), axis=0) * p_y)\ni_y_c = h_c - h_c_on_y\nnmi = 2 * i_y_c \/ (h_y + h_c)\n<\/code><\/pre>\n\n<p>In practice, please be very careful on the probabilities as they should be positive to avoid numeric overflow in <code>tf.math.log<\/code>.<\/p>\n\n<p>Please comment if you find any mistakes.<\/p>\n"}]}
+{"QuestionId":52768071,"AnswerCount":0,"Tags":"<python><tensorflow><predict><tensorflow-serving>","CreationDate":"2018-10-11T20:02:12.067","AcceptedAnswerId":null,"OwnerUserId":4258559.0,"Title":"Passing dictionary for receiver_tensor in tf.estimator.export.ServingInputReceiver","Body":"<p>I am trying to use the following json_serving_fn for predict on my tf model. <\/p>\n\n<pre><code>def json_serving_input_fn():\n  \"\"\"Build the serving inputs.\"\"\"\n  feature_placeholders = {'a': tf.placeholder(shape=[None], dtype='string'),\n            'b': tf.placeholder(shape=[None], dtype='string'), \n            'c': tf.placeholder(shape=[1,10], dtype='string')}\n  return tf.estimator.export.ServingInputReceiver(\n    features=feature_placeholders,\n      receiver_tensors=feature_placeholders)\n<\/code><\/pre>\n\n<p>The train and evaluation processes complete as intended (and orthogonal to the issue), but my serving function fails with the following error code: <\/p>\n\n<pre><code>INFO:tensorflow:Signatures EXCLUDED from export because they cannot be be served via TensorFlow Serving APIs:                                                                         \nINFO:tensorflow:'serving_default' : Classification input must be a single string Tensor; got {'symptoms': &lt;tf.Tensor 'Placeholder_2:0' shape=(1,10) dtype=string&gt;, 'age': &lt;tf.Tensor 'Placeholder_1:0' shape=(?,) dtype=string&gt;, 'sex': &lt;tf.Tensor 'Placeholder:0' shape=(?,) dtype=string&gt;}     \n<\/code><\/pre>\n\n<p>I intend to pass the following json for prediction:<\/p>\n\n<pre><code>{\"a\":\"x\",\"b\":\"y\",\"c\":[\"p1\",\"p2\",\"\",\"\",\"\",\"\",\"\",\"\",\"\",\"\"]}\n<\/code><\/pre>\n\n<p>How do I rewrite my json_serving_input_fun to accomplish that?<\/p>\n","answers":[]}
+{"QuestionId":52768408,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><protocol-buffers><object-detection>","CreationDate":"2018-10-11T20:26:11.390","AcceptedAnswerId":52771416.0,"OwnerUserId":10272643.0,"Title":"Tensorflow object Detection with OpenCV export Text of objects detected","Body":"<p>I am working with Object Detection using Tensorflow and openCV on Raspberry pi 3B+. I am pretty beginner and don't know much inside the actual working of it. So I saw a tutorial <a href=\"https:\/\/www.youtube.com\/watch?v=npZ-8Nj1YwY&amp;feature=youtu.be\" rel=\"nofollow noreferrer\">this<\/a>. \nSo basically it is detecting objects and making a box on object in window and labeling it. I want raspberry pi to speak out the objects that he detects with more than 60% chances of that object. So how do I extract the string of detected object so that another script can speak that out.\nThe script for detecting object is as follows:\nThank you<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>######## Picamera Object Detection Using Tensorflow Classifier #########\r\n#\r\n# Author: Evan Juras\r\n# Date: 4\/15\/18\r\n# Description: \r\n# This program uses a TensorFlow classifier to perform object detection.\r\n# It loads the classifier uses it to perform object detection on a Picamera feed.\r\n# It draws boxes and scores around the objects of interest in each frame from\r\n# the Picamera. It also can be used with a webcam by adding \"--usbcam\"\r\n# when executing this script from the terminal.\r\n\r\n## Some of the code is copied from Google's example at\r\n## https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\r\n\r\n## and some is copied from Dat Tran's example at\r\n## https:\/\/github.com\/datitran\/object_detector_app\/blob\/master\/object_detection_app.py\r\n\r\n## but I changed it to make it more understandable to me.\r\n\r\n\r\n# Import packages\r\nimport os\r\nimport cv2\r\nimport numpy as np\r\nfrom picamera.array import PiRGBArray\r\nfrom picamera import PiCamera\r\nimport tensorflow as tf\r\nimport argparse\r\nimport sys\r\n\r\n# Set up camera constants\r\nIM_WIDTH = 1280\r\nIM_HEIGHT = 720\r\n#IM_WIDTH = 640    Use smaller resolution for\r\n#IM_HEIGHT = 480   slightly faster framerate\r\n\r\n# Select camera type (if user enters --usbcam when calling this script,\r\n# a USB webcam will be used)\r\ncamera_type = 'picamera'\r\nparser = argparse.ArgumentParser()\r\nparser.add_argument('--usbcam', help='Use a USB webcam instead of picamera',\r\n                    action='store_true')\r\nargs = parser.parse_args()\r\nif args.usbcam:\r\n    camera_type = 'usb'\r\n\r\n# This is needed since the working directory is the object_detection folder.\r\nsys.path.append('..')\r\n\r\n# Import utilites\r\nfrom utils import label_map_util\r\nfrom utils import visualization_utils as vis_util\r\n\r\n# Name of the directory containing the object detection module we're using\r\nMODEL_NAME = 'ssdlite_mobilenet_v2_coco_2018_05_09'\r\n\r\n# Grab path to current working directory\r\nCWD_PATH = os.getcwd()\r\n\r\n# Path to frozen detection graph .pb file, which contains the model that is used\r\n# for object detection.\r\nPATH_TO_CKPT = os.path.join(CWD_PATH,MODEL_NAME,'frozen_inference_graph.pb')\r\n\r\n# Path to label map file\r\nPATH_TO_LABELS = os.path.join(CWD_PATH,'data','mscoco_label_map.pbtxt')\r\n\r\n# Number of classes the object detector can identify\r\nNUM_CLASSES = 90\r\n\r\n## Load the label map.\r\n# Label maps map indices to category names, so that when the convolution\r\n# network predicts `5`, we know that this corresponds to `airplane`.\r\n# Here we use internal utility functions, but anything that returns a\r\n# dictionary mapping integers to appropriate string labels would be fine\r\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\r\ncategories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)\r\ncategory_index = label_map_util.create_category_index(categories)\r\n\r\n# Load the Tensorflow model into memory.\r\ndetection_graph = tf.Graph()\r\nwith detection_graph.as_default():\r\n    od_graph_def = tf.GraphDef()\r\n    with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:\r\n        serialized_graph = fid.read()\r\n        od_graph_def.ParseFromString(serialized_graph)\r\n        tf.import_graph_def(od_graph_def, name='')\r\n\r\n    sess = tf.Session(graph=detection_graph)\r\n\r\n\r\n# Define input and output tensors (i.e. data) for the object detection classifier\r\n\r\n# Input tensor is the image\r\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\r\n\r\n# Output tensors are the detection boxes, scores, and classes\r\n# Each box represents a part of the image where a particular object was detected\r\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\r\n\r\n# Each score represents level of confidence for each of the objects.\r\n# The score is shown on the result image, together with the class label.\r\ndetection_scores = detection_graph.get_tensor_by_name('detection_scores:0')\r\ndetection_classes = detection_graph.get_tensor_by_name('detection_classes:0')\r\n\r\n# Number of objects detected\r\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\r\n\r\n# Initialize frame rate calculation\r\nframe_rate_calc = 1\r\nfreq = cv2.getTickFrequency()\r\nfont = cv2.FONT_HERSHEY_SIMPLEX\r\n\r\n# Initialize camera and perform object detection.\r\n# The camera has to be set up and used differently depending on if it's a\r\n# Picamera or USB webcam.\r\n\r\n# I know this is ugly, but I basically copy+pasted the code for the object\r\n# detection loop twice, and made one work for Picamera and the other work\r\n# for USB.\r\n\r\n### Picamera ###\r\nif camera_type == 'picamera':\r\n    # Initialize Picamera and grab reference to the raw capture\r\n    camera = PiCamera()\r\n    camera.resolution = (IM_WIDTH,IM_HEIGHT)\r\n    camera.framerate = 10\r\n    rawCapture = PiRGBArray(camera, size=(IM_WIDTH,IM_HEIGHT))\r\n    rawCapture.truncate(0)\r\n\r\n    for frame1 in camera.capture_continuous(rawCapture, format=\"bgr\",use_video_port=True):\r\n\r\n        t1 = cv2.getTickCount()\r\n        \r\n        # Acquire frame and expand frame dimensions to have shape: [1, None, None, 3]\r\n        # i.e. a single-column array, where each item in the column has the pixel RGB value\r\n        frame = frame1.array\r\n        frame.setflags(write=1)\r\n        frame_expanded = np.expand_dims(frame, axis=0)\r\n\r\n        # Perform the actual detection by running the model with the image as input\r\n        (boxes, scores, classes, num) = sess.run(\r\n            [detection_boxes, detection_scores, detection_classes, num_detections],\r\n            feed_dict={image_tensor: frame_expanded})\r\n\r\n        # Draw the results of the detection (aka 'visulaize the results')\r\n        vis_util.visualize_boxes_and_labels_on_image_array(\r\n            frame,\r\n            np.squeeze(boxes),\r\n            np.squeeze(classes).astype(np.int32),\r\n            np.squeeze(scores),\r\n            category_index,\r\n            use_normalized_coordinates=True,\r\n            line_thickness=8,\r\n            min_score_thresh=0.40)\r\n\r\n        cv2.putText(frame,\"FPS: {0:.2f}\".format(frame_rate_calc),(30,50),font,1,(255,255,0),2,cv2.LINE_AA)\r\n\r\n        # All the results have been drawn on the frame, so it's time to display it.\r\n        cv2.imshow('Object detector', frame)\r\n\r\n        t2 = cv2.getTickCount()\r\n        time1 = (t2-t1)\/freq\r\n        frame_rate_calc = 1\/time1\r\n\r\n        # Press 'q' to quit\r\n        if cv2.waitKey(1) == ord('q'):\r\n            break\r\n\r\n        rawCapture.truncate(0)\r\n\r\n    camera.close()\r\n\r\n### USB webcam ###\r\nelif camera_type == 'usb':\r\n    # Initialize USB webcam feed\r\n    camera = cv2.VideoCapture(0)\r\n    ret = camera.set(3,IM_WIDTH)\r\n    ret = camera.set(4,IM_HEIGHT)\r\n\r\n    while(True):\r\n\r\n        t1 = cv2.getTickCount()\r\n\r\n        # Acquire frame and expand frame dimensions to have shape: [1, None, None, 3]\r\n        # i.e. a single-column array, where each item in the column has the pixel RGB value\r\n        ret, frame = camera.read()\r\n        frame_expanded = np.expand_dims(frame, axis=0)\r\n\r\n        # Perform the actual detection by running the model with the image as input\r\n        (boxes, scores, classes, num) = sess.run(\r\n            [detection_boxes, detection_scores, detection_classes, num_detections],\r\n            feed_dict={image_tensor: frame_expanded})\r\n\r\n        # Draw the results of the detection (aka 'visulaize the results')\r\n        vis_util.visualize_boxes_and_labels_on_image_array(\r\n            frame,\r\n            np.squeeze(boxes),\r\n            np.squeeze(classes).astype(np.int32),\r\n            np.squeeze(scores),\r\n            category_index,\r\n            use_normalized_coordinates=True,\r\n            line_thickness=8,\r\n            min_score_thresh=0.85)\r\n\r\n        cv2.putText(frame,\"FPS: {0:.2f}\".format(frame_rate_calc),(30,50),font,1,(255,255,0),2,cv2.LINE_AA)\r\n        \r\n        # All the results have been drawn on the frame, so it's time to display it.\r\n        cv2.imshow('Object detector', frame)\r\n\r\n        t2 = cv2.getTickCount()\r\n        time1 = (t2-t1)\/freq\r\n        frame_rate_calc = 1\/time1\r\n\r\n        # Press 'q' to quit\r\n        if cv2.waitKey(1) == ord('q'):\r\n            break\r\n\r\n    camera.release()\r\n\r\ncv2.destroyAllWindows()<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"52771416","CreationDate":"2018-10-12T02:38:30.210","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Not sure what your model looks like but here is where the prediction is made (classes)<\/p>\n\n<pre><code>(boxes, scores, classes, num) = sess.run(\n            [detection_boxes, detection_scores, detection_classes, num_detections],\n            feed_dict={image_tensor: frame_expanded})\n<\/code><\/pre>\n\n<p>Here you can see where the labels are being loaded from<\/p>\n\n<pre><code>PATH_TO_LABELS = os.path.join(CWD_PATH,'data','mscoco_label_map.pbtxt')\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\n<\/code><\/pre>\n\n<p>Also, this is the function that is drawing the labels and boxes<\/p>\n\n<pre><code>vis_util.visualize_boxes_and_labels_on_image_array(\n            frame,\n            np.squeeze(boxes),\n            np.squeeze(classes).astype(np.int32),\n            np.squeeze(scores),\n            category_index,\n            use_normalized_coordinates=True,\n            line_thickness=8,\n            min_score_thresh=0.40)\n<\/code><\/pre>\n\n<p>With a little investigation you should be able to see what classes correspond to what label and find the objects with scores about 0.6<\/p>\n"}]}
+{"QuestionId":52768419,"AnswerCount":1,"Tags":"<python><tensorflow><nvidia>","CreationDate":"2018-10-11T20:26:44.467","AcceptedAnswerId":null,"OwnerUserId":8073716.0,"Title":"Failed to load the native TensorFlow runtime. Python 3.6 on Windows 10","Body":"<p>I am installing CUDA GPU Tool (version 9.2) for Python 3.6 on Windows 10. I get the following error:<\/p>\n\n<p>Traceback (most recent call last):<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n    import tensorflow<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow__init__.py\", line 22, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)<\/p>\n\n<p>File \"D:\\Python\\Python36\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n","answers":[{"AnswerId":"53011376","CreationDate":"2018-10-26T14:55:09.937","ParentId":null,"OwnerUserId":"3172047","Title":null,"Body":"<p>Last time I tried CUDA 9.1 few months back I ran into similar issues. I suggest you to install CUDA 9.0 (Reason- Tensorflow pip release may be based on older version of CUDA).\nYou also need to install cuDNN of matching version. You can find the detailed steps guide <a href=\"https:\/\/mlguy.org\/2018\/06\/03\/getting-started-with-tensorflow-gpu-on-windows-10\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52769255,"AnswerCount":3,"Tags":"<path><pip><pytorch>","CreationDate":"2018-10-11T21:34:45.410","AcceptedAnswerId":null,"OwnerUserId":5503981.0,"Title":"Fastai on google colab","Body":"<p>Two days back I ran my model using fastai 0.7.0 on google colab.\nAnd for two days I got busy and now if I am trying to run it , its throwing me an error, on the execution of the line\n*\"from fastai.transforms import \u2019 .\nthe error is AttributeError: module \u2018torch\u2019 has no attribute \u2018float32\u2019.<a href=\"https:\/\/i.stack.imgur.com\/3z5MN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3z5MN.png\" alt=\"Image showing the error\"><\/a><\/p>\n","answers":[{"AnswerId":"54228656","CreationDate":"2019-01-17T03:37:13.650","ParentId":null,"OwnerUserId":"10925701","Title":null,"Body":"<p>add below code. i dont really know which dependency call this. but it worked<\/p>\n\n<pre><code>!pip install -q pandas==0.23.4 \n!pip install -q joblib==0.12.5 \n!pip install -q plotly==3.0.0\n<\/code><\/pre>\n"},{"AnswerId":"52775244","CreationDate":"2018-10-12T08:19:59.510","ParentId":null,"OwnerUserId":"5503981","Title":null,"Body":"<p>At-last found what was wrong, there was a compatibility issue. \nThe pytorch version was 0.3.1 but <strong>torchtext<\/strong> version was not compatible with it, has to be \u201ctorchtext==0.2.3.\u201d<\/p>\n"},{"AnswerId":"53893656","CreationDate":"2018-12-22T06:50:16.800","ParentId":null,"OwnerUserId":"236106","Title":null,"Body":"<p>The following should get you up and running with the 0.7.0 version of fast.ai (used by v2 of the course) on Google Colab:<\/p>\n\n<pre><code>!pip install -q \"fastai==0.7.0\" Pillow==4.1.1 torchtext==0.2.3\n!apt-get -qq install -y libsm6 libxext6 &amp;&amp; pip install -q -U opencv-python\n\nfrom wheel.pep425tags import get_abbr_impl, get_impl_ver, get_abi_tag\nplatform = '{}{}-{}'.format(get_abbr_impl(), get_impl_ver(), get_abi_tag())\n# !apt update -q\n!apt install -y libsm6 libxext6\n\nfrom os import path\naccelerator = 'cu80' if path.exists('\/opt\/bin\/nvidia-smi') else 'cpu'\ntorch_whl = f\"http:\/\/download.pytorch.org\/whl\/{accelerator}\/torch-0.3.0.post4-{platform}-linux_x86_64.whl\"\n!pip install -q {torch_whl} torchvision image\n<\/code><\/pre>\n"}]}
+{"QuestionId":52769339,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-11T21:43:45.437","AcceptedAnswerId":52769583.0,"OwnerUserId":6703017.0,"Title":"Keras Convolution Neural Network Dimensions Question","Body":"<p>I am interested in going from a 3-D matrix to a 1-D matrix.\n1000 rows, 4 columns, 16 sample to 16 sample with 1of4 diseases. <\/p>\n\n<p>I recall matrix algebra: AX +B = Y\nIn this case say B=0, Y= 16x1\nThen how does the dimension of AX work out in this fashion? What layer can I use for Y to eventually get a dimension of 16x1 with 4 categories?<\/p>\n\n<p>And would I start the convolution network with a shape like this:<\/p>\n\n<p>conv2D(filter,kernel, input_shape=(16,1000,4))<\/p>\n\n<p>How would I end it with a layer that gives me 4 categories ie dense(4)?\nPlease give an example.<\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"52769583","CreationDate":"2018-10-11T22:08:47.153","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>It is really difficult for me to understand your question. Do you mean that your input is a feature vector of 1000 dimensions, and in total you have 16 of them, and your network output, i.e. target, is one of four possible diseases? <\/p>\n\n<p>If this is the case, then you misunderstand CNN's input shape, whose batch size is not accounted (you may use different batch sizes, and thus we want to decouple this param from a model). <\/p>\n\n<p>Depending on your understanding of the 1000-dim feature, you want to apply some feature normalization before you feed it into a model. <\/p>\n\n<pre><code>from keras.layers import *\nfrom keras.models import Model\n\n# 1. define model parameters\n# 1.1 dropout rate, you may want to use 0 first\ndropout_rate = 0.2 \n# 1.2 mlp structure\nfc_nodes = [128, 32, 8] \n# activation, e.g. relu or tanh\nactivation = 'relu' \n\n# 2. define model\n# 2.1 define input\n# note: batch_size=16 is NOT included in shape\nfeat_in = Input(shape=(1000,), name='feat_in')\n# 2.2 define body\nx = feat_in\nfor idx, nb_nodes in enumerate( fc_nodes ) :\n    x = Dense( nb_nodes, activation=activation, name='fc-{}'.format(idx) )(x)\n    if ( dropout_rate &gt; 0 ) :\n        x = Dropout( dropout_rate, name='dropout-{}'.format(idx))(x)\n# 2.3 define output\npred_out = Dense( 4, activation='softmax', name='pred_out')(x)\n# 2.4 define model\nmodel = Model( inputs=feat_in, outputs=pred_out, name='my_model')\n\n# 3. display model architecture\nprint model.summary()\n<\/code><\/pre>\n\n<p>This model will take in a 2D tensor of shape <code>(batch_size, 1000)<\/code> and predict a 2D tensor of shape <code>(batch_size, 4)<\/code>.<\/p>\n\n<p>BTW, <code>Conv2D<\/code> should be only used if your input is image-like, which is not your case.<\/p>\n"}]}
+{"QuestionId":52769482,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-11T21:59:21.717","AcceptedAnswerId":null,"OwnerUserId":2659352.0,"Title":"tensorflow: vocabulary lookup with tf.estimator and keras","Body":"<p>I have the following dataset<\/p>\n\n<pre><code>username,itemname,value\n\"carl\",\"socks\",12.50\n\"john doe\",\"shirts\",30.00\n...\n<\/code><\/pre>\n\n<p>I also have the following vocabulary lookup files<\/p>\n\n<p>usernames.txt<\/p>\n\n<pre><code>carl\njohn doe\nbob smith\n...\n<\/code><\/pre>\n\n<p>itemnames.txt<\/p>\n\n<pre><code>socks\nshirts\nshoes\n...\n<\/code><\/pre>\n\n<p>I will be receiving the strings at prediction time. There is no way around that. In order to make training similar I am using tf.contrib.lookup<\/p>\n\n<pre><code>import tf.contrib.lookup\n\nuser_lookup = tf.contrib.lookup.index_table_from_file(\n    vocabulary_file='usernames.txt'\n)\n\nitem_lookup = tf.contrib.lookup.index_table_from_file(\n    vocabulary_file='itemnames.txt'\n)    \n<\/code><\/pre>\n\n<p>Now I have the following model defined using the keras api<\/p>\n\n<pre><code>import tensorflow as tf\n\nuser_input = tf.keras.layers.Input(shape=(1,), dtype=tf.int32)\nitem_input = tf.keras.layers.Input(shape=(1,), dtype=tf.int32)\n\nuser_embedding = tf.keras.layers.Embedding(input_dim=num_users, output_dim=10)(user_input)\nitem_embedding = tf.keras.layers.Embedding(input_dim=num_items, output_dim=10)(item_input)\n\n...\noutput = ...\nmodel = tf.keras.Model([user_input, item_input], output)\nmodel.compile(...)\n<\/code><\/pre>\n\n<p>I am using tf.estimator for training and prediction. So my first instinct is to do the following:<\/p>\n\n<pre><code>my_estimator = tf.keras.estimator.model_to_estimator(keras_model=model)\n\ntf.tables_initializer()\n\ndef train_fn(dataset_iterator):\n     (username, itemname), value = dataset_iterator.get_next()\n     userid = user_lookup.lookup(username)\n     itemid = item_lookup.lookup(itemname)\n     return (username, itemname), value\n\nmy_train_spec = tf.estimator.TrainSpec(\n   input_fn=train_fn(train_data)\n)\n\nmy_eval_spec = tf.estimator.EvalSpec(\n   input_fn=train_fn(validation_data)\n)\n\ntf.estimator.train_and_evaluate(\n    estimator=my_estimator,\n    train_spec=my_train_spec,\n    eval_spec=my_eval_spec\n)\n<\/code><\/pre>\n\n<p>When I run this I get the follow error:<\/p>\n\n<pre><code>ValueError: Tensor(\"Cast_2:0\", shape=(), dtype=int32) must be from the same graph as Tensor(\"Item-Embedding-LMF\/embeddings\/Read\/ReadVariableOp:0\", shape=(429099, 10), dtype=float32, device=\/job:ps\/task:1).\n<\/code><\/pre>\n\n<p>Can anyone recommend a solution to this problem? Or maybe even a different approach to handling this lookup?<\/p>\n","answers":[{"AnswerId":"57616052","CreationDate":"2019-08-22T19:36:47.070","ParentId":null,"OwnerUserId":"1582366","Title":null,"Body":"<p>Mostly, no issue with the lookup. The all the variaables should be asscociated with the same graphs \nMay be write the model in the scope e.g. <\/p>\n\n<pre><code>def model_fn: \n    with tf.variable_scope('my_model', reuse=tf.AUTO_REUSE):\n    ....\n    ..\n     return estimator \n<\/code><\/pre>\n"}]}
+{"QuestionId":52769607,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-11T22:12:07.737","AcceptedAnswerId":null,"OwnerUserId":4334452.0,"Title":"Retrain Tensorflow model","Body":"<p>I have a tensorflow model which I trained using object detection SSD mobile net .\nThe training is finished now and I exported the model inference for testing .My inquiry is if I want to retrain the model later with new dataset of images, what should i do now at the stage to make the weights permeant in the model so that I can retrain it from that point on .I know there is a freeze script ,do I have to use that ? <\/p>\n\n<p>Thanks<\/p>\n\n<p>Ayad<\/p>\n","answers":[{"AnswerId":"52848490","CreationDate":"2018-10-17T06:26:44.597","ParentId":null,"OwnerUserId":"6918487","Title":null,"Body":"<p>You cannot retrain it from that point on. Every time you retrain it should be for all objects that you are interested in.<\/p>\n"}]}
+{"QuestionId":52769986,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-11T22:57:23.650","AcceptedAnswerId":null,"OwnerUserId":8869019.0,"Title":"How to sample efficiently from a list like [[0,1],[0,6],[0,7],[1,4],[1,8]], ideally in pytorch (so this will be a tensor)","Body":"<p>I have a tensor that looks like<\/p>\n\n<pre><code>[[0,1],\n [0,6],\n [0,7],\n [1,4],\n [1,8],\n [2,5],\n [2,7]]\n<\/code><\/pre>\n\n<p>so the first column is always a non-decreasing sequence from 0 to N (N is fixed), and the second column is some integer (within a fixed range). For the sampling, I want to obtain one sample from the second column for each distinct number in the first column. In this example, I want to get one sample from [1,6,7] for 0,  one sample from [4,8] for 1, and one sample from [5,7] for 2, and valid examples of samples can be [1,8,7], [7,8,7], etc. <\/p>\n","answers":[{"AnswerId":"52770044","CreationDate":"2018-10-11T23:06:29.397","ParentId":null,"OwnerUserId":"1322401","Title":null,"Body":"<pre><code>from itertools import groupby\nfrom operator import itemgetter\nfrom random import choice\n\ntnsr = [[0, 1], [0, 6], [0, 7], [1, 4], [1, 8], [2, 5], [2, 7]]\n\nchoices = [choice(list(grp))[1] for key, grp in groupby(tnsr, key=itemgetter(0))]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52770595,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-12T00:32:50.647","AcceptedAnswerId":52770731.0,"OwnerUserId":1727392.0,"Title":"Keras Sequence: Specify input shape with three parameters","Body":"<p>I have a dataframe with the following setup:<\/p>\n\n<pre><code>import numpy as np\n\nX = np.random.rand(100, 20, 3)\n<\/code><\/pre>\n\n<p>Here there are 100 time slices, 20 observations, and 3 attributes per observation.<\/p>\n\n<p>I am trying to figure out how to pass the data above to the following Keras sequence:<\/p>\n\n<pre><code>from keras.models import Sequential, Model\nfrom keras.layers import Dense, LSTM, Dropout, Activation\nimport keras\n\n# config\nstateful = False\nlook_back = 3\nlstm_cells = 1024\ndropout_rate = 0.5\nn_features = int(X.shape[1]*3)\ninput_shape = (look_back, n_features, 3)\noutput_shape = n_features\n\ndef loss(y_true, y_pred):\n  return keras.losses.mean_squared_error(y_true, y_pred)\n\nmodel = Sequential()\nmodel.add(LSTM(lstm_cells, stateful=stateful, return_sequences=True, input_shape=input_shape))\nmodel.add(Dense(output_shape, activation='relu'))\nmodel.compile(loss=loss, optimizer='sgd')\n<\/code><\/pre>\n\n<p>Running this throws:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_23: expected\n  ndim=3, found ndim=4<\/p>\n<\/blockquote>\n\n<p>Does anyone know how I can reshape <code>X<\/code> to pass it into the model? Any suggestions would be helpful!<\/p>\n","answers":[{"AnswerId":"52770731","CreationDate":"2018-10-12T00:52:26.983","ParentId":null,"OwnerUserId":"1727392","Title":null,"Body":"<p>This seems to get things rolling:<\/p>\n\n<pre><code>from keras.models import Sequential, Model\nfrom keras.layers import Dense, LSTM, Dropout, Activation\nimport keras\n\n# config\nstateful = False\nlook_back = 3\nlstm_cells = 1024\ndropout_rate = 0.5\nn_features = int(X.shape[1]) * 3\ninput_shape = (look_back, n_features)\noutput_shape = n_features\n\ndef loss(y_true, y_pred):\n  return keras.losses.mean_squared_error(y_true, y_pred)\n\nmodel = Sequential()\nmodel.add(LSTM(lstm_cells, stateful=stateful, return_sequences=True, input_shape=input_shape))\nmodel.add(LSTM(lstm_cells, stateful=stateful, return_sequences=True))\nmodel.add(LSTM(lstm_cells, stateful=stateful))\nmodel.add(Dense(output_shape, activation='relu'))\nmodel.compile(loss=loss, optimizer='sgd')\n<\/code><\/pre>\n\n<p>Then one can partition the training data as follows:<\/p>\n\n<pre><code># build training data\ntrain_x = []\ntrain_y = []\nn_time = int(X.shape[0])\nn_obs = int(X.shape[1])\nn_attrs = int(X.shape[2])\n\n# note we flatten the last dimension\nfor i in range(look_back, n_time-1, 1):\n  train_x.append( X[i-look_back:i].reshape(look_back, n_obs * n_attrs ) )\n  train_y.append( X[i+1].ravel() )\n\ntrain_x = np.array(train_x)\ntrain_y = np.array(train_y)\n<\/code><\/pre>\n\n<p>Then one can train the toy model:<\/p>\n\n<pre><code>model.fit(train_x, train_y, epochs=10, batch_size=10)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52770780,"AnswerCount":2,"Tags":"<python><tensorflow><reinforcement-learning><q-learning>","CreationDate":"2018-10-12T01:00:40.840","AcceptedAnswerId":null,"OwnerUserId":7867475.0,"Title":"Why is my Deep Q Net and Double Deep Q Net unstable?","Body":"<p>I am trying to implement DQN and DDQN(both with experience reply) to solve OpenAI AI-Gym Cartpole Environment. Both of the approaches are able to learn and solve this problem sometimes, but not always.<\/p>\n\n<p>My network is simply a feed forward network(I've tried using 1 and 2 hidden layers). In DDQN I created one network in DQN, and two networks in DDQN, a target network to evaluate the Q value and a primary network to choose the best action, train the primary network, and copy it to target network after some episodes.<\/p>\n\n<p>The problem in DQN is: <\/p>\n\n<ul>\n<li>Sometimes it can achieve the perfect 200 score within 100 episodes, but sometimes it gets stuck and only achieves 10 score no matter how long it is trained. <\/li>\n<li>Also, in case of successful learning, the learning speed differ.<\/li>\n<\/ul>\n\n<p>The problem in DDQN is:<\/p>\n\n<ul>\n<li>It can learn to achieve 200 score, but then it seems to forget what's learned and the score drops dramatically.<\/li>\n<\/ul>\n\n<p>I've tried tuning batch size, learning rate, number of neurons in the hidden layer, the number of hidden layers, exploration rate, but instability persists. <\/p>\n\n<p>Are there any rule of thumb on the size of network and batch size? I think reasonably larger network and larger batch size will increase stability. <\/p>\n\n<p>Is it possible to make the learning stable? Any comments or references are appreciated!<\/p>\n","answers":[{"AnswerId":"52842749","CreationDate":"2018-10-16T19:28:20.430","ParentId":null,"OwnerUserId":"6387541","Title":null,"Body":"<p>Maybe this can help you with your problem on this environment.<\/p>\n\n<p><a href=\"https:\/\/github.com\/udacity\/deep-learning\/blob\/master\/reinforcement\/Q-learning-cart.ipynb\" rel=\"nofollow noreferrer\">Cartpole problem with DQN algorithm from Udacity<\/a><\/p>\n"},{"AnswerId":"52892468","CreationDate":"2018-10-19T12:36:00.437","ParentId":null,"OwnerUserId":"5912280","Title":null,"Body":"<p>These kind of problems happen pretty often and you shouldn't give up. First, of course, you should do another one or two checks if the code is all right - try to compare your code to other implementations, see how the loss function behave etc. If you are pretty sure your code is all fine - and, as you say that model can learn the task from time to time, it probably is - you should start experimenting with the hyper-parameters.<\/p>\n\n<p>Your problems seem to be connected to hyper-parameters like exploration technique, learning rate, the way you are updating the target networks and to the experience replay memory. I would not play around with the hidden layer sizes - find the values for which the model learned once and keep them fixed. <\/p>\n\n<ul>\n<li><strong>Exploration technique:<\/strong> I assume you use epsilon-greedy strategy. My advice would be to start with a high epsilon value (I usually start with 1.0) and decay it after each step or episode, but define an epsilon_min too. Starting with a low epsilon value may be the problem of different learning speeds and success rates - if you go full random, you always populate your memory with similar kind of transitions at the beginning. With lower epsilon rates at the start, there is a bigger chance for your model to not explore enough before the exploitation phase begins. <\/li>\n<li><strong>Learning rate:<\/strong> Make sure it is not too big. Smaller rate may lower the learning speed, but helps a learned model to not escape back from global minima to some local, worse ones. Also, adaptive learning rates such as these calculated with <a href=\"https:\/\/arxiv.org\/abs\/1412.6980\" rel=\"nofollow noreferrer\">Adam<\/a> might help you. Of course the batch size have an impact as well, but I would keep it fixed and worry about it only if the other hyper-parameter changes won't work. <\/li>\n<li><strong>Target network update (rate and value):<\/strong> This is an important one as well. You have to experiment a bit - not only how often do you perform the update, but also how much of the primary values you copy into the target ones. People often do a hard update each episode or so, but try doing soft updates instead if the first technique does not work. <\/li>\n<li><strong>Experience replay<\/strong>: Do you use it? You should. How big is your memory size? This is very important factor and the memory size can influence the stability and success rate (<a href=\"https:\/\/arxiv.org\/abs\/1712.01275\" rel=\"nofollow noreferrer\">A Deeper Look at Experience Replay<\/a>). Basically, if you notice instability of your algorithm, try a bigger memory size, and if it affects your learning curve a lot, try out the technique proposed in the mentioned paper.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":52771791,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-12T03:25:33.103","AcceptedAnswerId":52773020.0,"OwnerUserId":2596672.0,"Title":"How to actually predict ONE parameter with Keras GRU on multi parameter weather time series","Body":"<p>After training the GRU architecture in the Keras book (F. Chollet) on the Jena weather dataset (chapter #6), I am having difficulties understanding the prediction phase:<\/p>\n\n<p>The last layer - Dense with no activation - outputs as expected, a stream of numbers:\nDimensions: Num of rows X 1.\nI guess these are supposed to be the predictions, but it looks like a probability of ALL the 14 parameters. <strong>Why only ONE prediction on the 14 parameters ?<\/strong><\/p>\n\n<p>Problem is the input is num rows X num cols (14 parameters) and the predictions output <strong>cannot be reshaped<\/strong> into a 14 columns array - such as the weather data set contains.<\/p>\n\n<p>Aren't the predictions supposed to have the <strong>same number of columns<\/strong> as the input, in the Jena Weather time sequence problem ?<\/p>\n\n<p>If the predictions are on ALL 14 parameters and they have been strained thru a Dense layer with ONE unit - then this predictions' \"stream\" should be able to be reshaped back to num rows X num cols ?<\/p>\n","answers":[{"AnswerId":"52787328","CreationDate":"2018-10-12T21:40:11.787","ParentId":null,"OwnerUserId":"2596672","Title":null,"Body":"<p>Thanks NormanZhu for your clarification.\nYou pointed out correctly that the <code>[1]<\/code> in the generator <code>targets[j] = data[rows[j] + delay][1]<\/code> is the detail I've missed.\nIt is the only place in the generator that identifies the column to be used for fit\/predict.<br>\nIt makes sense and it allowed me to finally view the Predictions (blue) <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kMbdc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kMbdc.png\" alt=\"enter image description here\"><\/a>vs Reality (red) charts.\n<a href=\"https:\/\/i.stack.imgur.com\/OegHO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OegHO.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"52773020","CreationDate":"2018-10-12T05:48:44.807","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>Please refer to the define of generator function:<\/p>\n\n<pre><code>samples[j] = data[indices]\ntargets[j] = data[rows[j] + delay][1]\n<\/code><\/pre>\n\n<p>The target is the second column in the data, which should be the '\"T (degC)\"'. So this model predicts the temperature in Celsius degree.  <\/p>\n\n<p>Regarding to the last Dense layer, it has not activation function. So it just works as the simple linear regression model. The output is just summary of the output of previous layer multiple layer weights.<\/p>\n"}]}
+{"QuestionId":52771860,"AnswerCount":1,"Tags":"<python><numpy><keras><autoencoder>","CreationDate":"2018-10-12T03:33:11.853","AcceptedAnswerId":null,"OwnerUserId":907773.0,"Title":"Numpy: Increasing size of 4th dimension","Body":"<p>I have a 4D numpy array: (#images, height, width, channels) of RGB images. The exact dimensions are (46, 224, 224, 3).<\/p>\n\n<p>I also have the grayscale versions of the same images. The exact dimensions are (46, 224, 224, 1).<\/p>\n\n<p>I'm playing with Keras, trying to make an autoencoder to learn the image values from the grayscale ones, but the model complains about the dimensions, so I'd like to add the 2 extra channels to the grayscale array, but I can't seem to figure out how.<\/p>\n\n<pre><code>temp_gray_images = np.zeros(original_images.shape, dtype=np.float32)\ntemp_gray_images[:,:,:,0] = gray_images\nValueError: could not broadcast input array from shape (46,224,224,1) into shape (46,224,224)\n<\/code><\/pre>\n\n<p>So I then learned about numpy.hstack, and I opened a new terminal and tried to play with some dummy data and a number of dimensions I could handle, and was able to get the expected results. However, that doesn't seem to work with 4 dimensions.<\/p>\n\n<pre><code>temp_gray_images = np.hstack([gray_images, np.zeros([original_images.shape[0], original_images.shape[1], original_images.shape[2], 2])])\nValueError: all the input array dimensions except for the concatenation axis must match exactly\n<\/code><\/pre>\n\n<p>Which doesn't make sense, because I kept the first 3 dimensions the same size.<\/p>\n\n<p>But ultimately the validation data doesn't match the train data because of the 2 extra channels in RGB vs. grayscale.<\/p>\n\n<pre><code>x_train shape: (46, 224, 224, 3)\nxtest_shape shape: (46, 224, 224, 1)\nautoencoder.fit(x_train, x_train, epochs=50, batch_size=256, shuffle=True, validation_data=(x_test, x_test))\n<\/code><\/pre>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"52771980","CreationDate":"2018-10-12T03:50:02.670","ParentId":null,"OwnerUserId":"8858832","Title":null,"Body":"<p>Could this be what you are looking for ?<\/p>\n\n<pre><code>import numpy as np\nx_train=np.random.randn(46, 224, 224, 1)\ny=np.repeat(x_train,3,axis=3)\nprint(y.shape) #(46, 224, 224, 3)\n<\/code><\/pre>\n\n<p>repeat will duplicate the gray scale images into the required number of channels<\/p>\n"}]}
+{"QuestionId":52772534,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-12T04:59:42.090","AcceptedAnswerId":null,"OwnerUserId":10333852.0,"Title":"What does the -1 mean in tensor.size(-1)?","Body":"<p>I have seen something like this in the Pytorch documentation,<\/p>\n\n<pre><code>import torch\n\na = torch.tensor([1, 2])\na.size() # torch.Size([2])\na.size(-1) # 2\n<\/code><\/pre>\n\n<p>How does this work? I didn't find a description. Thanks,<\/p>\n","answers":[{"AnswerId":"52772768","CreationDate":"2018-10-12T05:22:02.407","ParentId":null,"OwnerUserId":"5349822","Title":null,"Body":"<p><code>a.size(-1)<\/code> refers to the last dimension. For example, if the shape of x were (10,20) then x.size(-1) refers to second dimension i.e. 20.\nTake a look at the following examples:<\/p>\n\n<pre><code>import torch\na= torch.zeros((2,5)) # a is matrix of 2 rows and 5 columns all elements are 0\n#size gives a 1d tensor containing the shapes\na.size(-1)# refers to the last element in the tensor\n<\/code><\/pre>\n\n<p>This is equivalent to:<\/p>\n\n<pre><code>a_size= a.size()\na_size(-1)\n<\/code><\/pre>\n\n<p>Hope this helps you.<\/p>\n"}]}
+{"QuestionId":52772605,"AnswerCount":1,"Tags":"<docker><tensorflow><tensorflow-serving>","CreationDate":"2018-10-12T05:06:59.387","AcceptedAnswerId":null,"OwnerUserId":9594598.0,"Title":"How to serve a tensorflow model using docker image tensorflow\/serving when there are custom ops?","Body":"<p>I'm trying to use the tf-sentencepiece operation in my model found here <a href=\"https:\/\/github.com\/google\/sentencepiece\/tree\/master\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/sentencepiece\/tree\/master\/tensorflow<\/a><\/p>\n\n<p>There is no issue building the model and getting a saved_model.pb file with variables and assets. However, if I try to use the docker image for tensorflow\/serving, it says <\/p>\n\n<pre><code>Loading servable: {name: model version: 1} failed: \nNot found: Op type not registered 'SentencepieceEncodeSparse' in binary running on 0ccbcd3998d1. \nMake sure the Op and Kernel are registered in the binary running in this process. \nNote that if you are loading a saved graph which used ops from tf.contrib, accessing \n(e.g.) `tf.contrib.resampler` should be done before importing the graph, \nas contrib ops are lazily registered when the module is first accessed.\n<\/code><\/pre>\n\n<p>I am unfamiliar with how to build anything manually, and was hoping that I could do this without many changes.<\/p>\n","answers":[{"AnswerId":"52863970","CreationDate":"2018-10-17T21:36:03.397","ParentId":null,"OwnerUserId":"10163556","Title":null,"Body":"<p>One approach would be to:<\/p>\n\n<ol>\n<li><p>Pull a docker development image<\/p>\n\n<p>$ docker pull tensorflow\/serving:latest-devel<\/p><\/li>\n<li><p>In the container, make your code changes<\/p>\n\n<p>$ docker run -it tensorflow\/serving:latest-devel <\/p><\/li>\n<\/ol>\n\n<p>Modify the code to add the op dependency <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/model_servers\/BUILD#L299\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<ol start=\"3\">\n<li><p>In the container, build TensorFlow Serving<\/p>\n\n<p>container:$ tensorflow_serving\/model_servers:tensorflow_model_server &amp;&amp; cp bazel-bin\/tensorflow_serving\/model_servers\/tensorflow_model_server \/usr\/local\/bin\/<\/p><\/li>\n<li><p>Use the exit command to exit the container<\/p><\/li>\n<li><p>Look up the container ID:<\/p>\n\n<p>$ docker ps<\/p><\/li>\n<li><p>Use that container ID to commit the development image:<\/p>\n\n<p>$ docker commit  $USER\/tf-serving-devel-custom-op<\/p><\/li>\n<li><p>Now build a serving container using the development container as the source<\/p>\n\n<p>$ mkdir \/tmp\/tfserving<\/p>\n\n<p>$ cd \/tmp\/tfserving<\/p>\n\n<p>$ git clone <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving<\/a> .<\/p>\n\n<p>$ docker build -t $USER\/tensorflow-serving --build-arg TF_SERVING_BUILD_IMAGE=$USER\/tf-serving-devel-custom-op -f tensorflow_serving\/tools\/docker\/Dockerfile .<\/p><\/li>\n<li><p>You can now use $USER\/tensorflow-serving to serve your image following the <a href=\"https:\/\/www.tensorflow.org\/serving\/docker#running_a_serving_image\" rel=\"nofollow noreferrer\">Docker instructions<\/a><\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":52772704,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-12T05:16:57.883","AcceptedAnswerId":null,"OwnerUserId":7694905.0,"Title":"How to process input data in Keras ConvLSTM layer","Body":"<p>I have a grid dataset with 14512 days, grid size is 25 * 53. The prediction time is 14 days. I try to put them into Keras ConvLSTM layer, but the result seem to be incorrect. How can I process the input data?<\/p>\n\n<p>preprocess the data: \ndata's shape is (14512, 25, 53)<\/p>\n\n<pre><code>fh = Dataset('slp.nc', mode='r')\nlength = len(fh.variables['time'][:])\nlat = len(fh.variables['lat'][:])\nlon = len(fh.variables['lon'][:])\ndata = [[[0 for _ in range(0, lon)] for _ in range(0, lat)] for _ in range(0, length)]\nfor i in range(0, length):\n    item = fh.variables['slp'][i]\n    for j in range(0, lat):\n        for k in range(0, lon):\n            data[i][j][k] = item[j][k]\ndata = pd.DataFrame(data)\ndata.drop([8665], inplace=True)\n<\/code><\/pre>\n\n<p>Then define the model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(ConvLSTM2D(filters=40, kernel_size=(3, 3), input_shape=(None, 25, 53, 1), padding='same', return_sequences=True))\nmodel.add(BatchNormalization())\nmodel.add(ConvLSTM2D(filters=40, kernel_size=(3, 3), padding='same',   return_sequences=True))\nmodel.add(BatchNormalization())\nmodel.add(ConvLSTM2D(filters=40, kernel_size=(3, 3), padding='same', return_sequences=True))\nmodel.add(BatchNormalization())\nmodel.add(ConvLSTM2D(filters=40, kernel_size=(3, 3), padding='same', return_sequences=True))\nmodel.add(BatchNormalization())\nmodel.add(Conv3D(filters=1, kernel_size=(3, 3, 3), activation='sigmoid', padding='same', data_format='channels_last'))\nmodel.compile(loss='binary_crossentropy', optimizer='adadelta')\n<\/code><\/pre>\n\n<p>Then generate input data and train:<\/p>\n\n<pre><code>train_x = np.zeros((t_day, 14, 25, 53, 1),dtype = np.float)\ntrain_y = np.zeros((t_day, 14, 25, 53, 1),dtype = np.float)\nfor i in range(t_day):\n    for j in range(0, 14):\n        for a in range(25):\n            for b in range(53):\n                if i + j &gt; 13:\n                    train_x[i][j][a][b][0] = data[i-14+j][a][b]\n                if i + j &lt; 13:\n                    train_x[i][j][a][b][0] = 0\n                train_y[i][j][a][b][0] = data[i+14+j][a][b]\nmodel.fit(train_x, train_y, epochs=100, batch_size=128, validation_split=0.05, verbose=1, shuffle=False)\n<\/code><\/pre>\n\n<p>But the loss function seem to be incorrect:<\/p>\n\n<pre><code>Epoch 1\/100\n\n10\/11400 [..............................] - ETA: 2:56:36 - loss: -3.7733\n20\/11400 [..............................] - ETA: 2:18:04 - loss: -110.7391\n30\/11400 [..............................] - ETA: 2:04:56 - loss: -322.3076\n40\/11400 [..............................] - ETA: 1:58:30 - loss: -554.6822\n50\/11400 [..............................] - ETA: 1:54:23 - loss: -746.9508\n60\/11400 [..............................] - ETA: 1:51:40 - loss: -932.8895\n70\/11400 [..............................] - ETA: 1:49:53 - loss: -1141.2996\n80\/11400 [..............................] - ETA: 1:48:23 - loss: -1368.9223\n90\/11400 [..............................] - ETA: 1:47:08 - loss: -1579.1898\n100\/11400 [..............................] - ETA: 1:46:08 - loss: -1773.8168\n110\/11400 [..............................] - ETA: 1:45:18 - loss: -1990.5375\n120\/11400 [..............................] - ETA: 1:44:37 - loss: -2189.3257\n130\/11400 [..............................] - ETA: 1:44:01 - loss: -2364.1422\n140\/11400 [..............................] - ETA: 1:43:36 - loss: -2579.9297\n150\/11400 [..............................] - ETA: 1:43:05 - loss: -2818.1180\n160\/11400 [..............................] - ETA: 1:42:40 - loss: -3001.0507\n170\/11400 [..............................] - ETA: 1:42:17 - loss: -3173.8332\n180\/11400 [..............................] - ETA: 1:41:55 - loss: -3358.1284\n190\/11400 [..............................] - ETA: 1:41:38 - loss: -3541.8308\n200\/11400 [..............................] - ETA: 1:41:20 - loss: -3733.1575\n210\/11400 [..............................] - ETA: 1:41:09 - loss: -3941.8186\n220\/11400 [..............................] - ETA: 1:40:54 - loss: -4164.8925\n230\/11400 [..............................] - ETA: 1:40:41 - loss: -4422.5382\n<\/code><\/pre>\n","answers":[{"AnswerId":"52775143","CreationDate":"2018-10-12T08:14:18.323","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>You did explain what purpose of your model. From the code, I think that it is a regression task. So you should not use the 'binary_crossentropy' loss function, which is for the classification. You should use the loss function for regression, such as mean_squared_error or mean_absolute_error. You can refer to the page of <a href=\"https:\/\/keras.io\/losses\/\" rel=\"nofollow noreferrer\">Keras document<\/a>.<\/p>\n"}]}
+{"QuestionId":52773410,"AnswerCount":2,"Tags":"<tensorflow><neural-network><keras>","CreationDate":"2018-10-12T06:18:28.153","AcceptedAnswerId":null,"OwnerUserId":1247575.0,"Title":"How to average input samples by group in Keras?","Body":"<p>I want to implement a neural network in Keras of this architecture: say if I have some inputs and they belong to some groups.  Then the neural network is like this:<\/p>\n\n<pre><code>input -&gt; some layers -&gt; separate inputs by groups -&gt; average inputs by groups -&gt; output\n<\/code><\/pre>\n\n<p>In brief, I want to separate inputs by groups then take the average of inputs by groups.<\/p>\n\n<p>For example, if I have some inputs tensor [1, 2, 3, 4, 5, 6] and they are belonging to two groups [0, 1, 1, 0, 0, 1]. Then I want to the output tensor is like this: [3.333, 3.666, 3.666, 3.333, 3.333, 3.666]. Here 3.333 is the average of group 0 [1, 4, 5] and 3.666 is the average of group 1 [2, 3, 6].<\/p>\n","answers":[{"AnswerId":"52774155","CreationDate":"2018-10-12T07:09:51.523","ParentId":null,"OwnerUserId":"5055862","Title":null,"Body":"<p>I am not sure if you can separate the inputs as you described above directly in Keras or Tensorflow. Here is what I could come up with:<\/p>\n\n<ol>\n<li>Create a mask corresponding to each class where 1 is for the element at the index being in the class and 0 for any element of another class. So in your example, you would do <code>[0,1,1,0,0,1]<\/code> for one class and <code>[1,0,0,1,1,0]<\/code> for the other. ( if you have more classes, you will correspondingly have more masks )<\/li>\n<li>Stack those vectors to get a 3-D tensor and do 1D convolution with 0 stride. Use <code>tf.nn.conv1d()<\/code>. Think of those masks as filters of a Convolution operation and it's separating the classes. Be sure to reshape your Tensors to match the operation requirements. <\/li>\n<li>After the convolution, you will have a 3-D Tensor where each vector would contain a classes elements. For your example you should get a Tensor with two vectors as <code>[0,2,3,0,0,6]<\/code> and <code>[1,0,0,4,5,0]<\/code>. Use <code>tf.reduce_mean()<\/code> on the correct axis to get the average of each class. <\/li>\n<li>Multiply the Tensor of the mean : <code>[[3.333], [3.666]]<\/code> with the masks using <code>tf.multiply()<\/code> and add the vectors using <code>tf.reduce_sum()<\/code> on the correct axis. And it should result in the vector you desire.  <\/li>\n<\/ol>\n"},{"AnswerId":"52840478","CreationDate":"2018-10-16T16:55:02.823","ParentId":null,"OwnerUserId":"1247575","Title":null,"Body":"<p>I have figured out a method. It can be archived by matrix manipulation. First turn the cluster vector to a categorical matrix, for example, if the batch size is 6, the categorical matrix (cluster) is like:<\/p>\n\n<pre><code>1, 0\n1, 0\n0, 1\n0, 1\n1, 0\n0, 1\n<\/code><\/pre>\n\n<p>then we generate a cluster_mean matrix:<\/p>\n\n<pre><code>1\/3, 0\n1\/3, 0\n0, 1\/3\n0, 1\/3\n1\/3, 0\n0, 1\/3\n<\/code><\/pre>\n\n<p>If we have an input matrix n*b (n is the number of features and b is the batch), then we can get average by cluster by using<\/p>\n\n<pre><code>cluster * t(cluster_mean) * input\n<\/code><\/pre>\n\n<p>Transpose, average and dot product can be archived by using tensorflow functions.<\/p>\n"}]}
+{"QuestionId":52774026,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><classification>","CreationDate":"2018-10-12T07:02:17.970","AcceptedAnswerId":null,"OwnerUserId":10333833.0,"Title":"Does tf.nn.sigmoid_cross_entropy_with_logits covert data into probabilities?","Body":"<p>In classification problems we prefer to apply <code>tf.nn.softmax<\/code> function because this function gives output as probabilities or we can also implement <code>tf.nn.softmax_cross_entropy_with_logits<\/code> which apply softmax activation function on model's output. I want to classify my data into two classses. I came across a code where they used <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code> at the last Layer. I want to know that can we use sigmoid as an activation function at the last layer? Shouldn't we use softmax? Without applying softmax how can I know probabilities? Because at the end of the model when we check how accurate is our model we use following command<\/p>\n\n<pre><code>tf.equal(tf.argmax(y,axis=1),tf.argmax(y_,axis=1))\n<\/code><\/pre>\n\n<p>Can after using sigmoid at the last fully connected layer we can use above command for checking correct predictions?<\/p>\n","answers":[{"AnswerId":"52774293","CreationDate":"2018-10-12T07:18:12.723","ParentId":null,"OwnerUserId":"8394915","Title":null,"Body":"<p>Softmax is used to convert the raw output values of a number of output neurons into probabilities for multi-class classification problems. For binary classification problems, instead of using two output neurons, you use a single output neuron and apply a logistic <a href=\"https:\/\/en.wikipedia.org\/wiki\/Logistic_function\" rel=\"nofollow noreferrer\">sigmoid function<\/a> to the output value. As the sigmoid projects values into the range [0, 1], you get a probability <code>p<\/code>. This is the probability of the current input belonging to one class. The probability of the input belonging to the other class is then <code>1 - p<\/code>.<\/p>\n\n<p>For your particular case, note that the doc of <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code> says:<\/p>\n\n<blockquote>\n  <p>Measures the probability error in discrete classification tasks in\n  which each class is independent and not mutually exclusive. For\n  instance, one could perform multilabel classification where a picture\n  can contain both an elephant and a dog at the same time.<\/p>\n<\/blockquote>\n\n<p>From the code you posted I assume that the network has an output dimension larger than one (otherwise the <code>argmax<\/code> makes no sense), probably two. This means that the code actually performs multiple binary classifications, which should be independent -- but you said that it should be a single binary classification.<\/p>\n\n<p>If my assumptions are correct then the code is wrong because two independent binary classifications are quite different from a single classification problem. In fact, the <code>y<\/code> should have a only single output and you would check whether the probability is <code>&gt; 0.5<\/code> to compare it to the targets.<\/p>\n"},{"AnswerId":"52775667","CreationDate":"2018-10-12T08:46:34.190","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The code that uses <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code> is not doing multi-class classification, it is doing <strong>multi-label classific<\/strong>ation, where you have multiple classes, but these are not exclusive and multiple classes might be predicted at the same time.<\/p>\n\n<p>For the multi-label case, sigmoid is preferable as then you get an independent probability for each class. Evaluation in this case is different, you can't use plain accuracy and precision\/recall for each class is used instead.<\/p>\n"}]}
+{"QuestionId":52774036,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-12T07:02:59.087","AcceptedAnswerId":null,"OwnerUserId":3925023.0,"Title":"Tensorflow - Custom model training in addition to existing one in","Body":"<p>I did follow this nice tutorial\n<a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10<\/a><\/p>\n\n<p>It works pretty fine, but my goal is to ADD new class to existing inception Coco model, so for instance I will still be capable to detect people, cars, etc + custom models.\nIs there a specif section in the class? or setting I should apply before?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52848604","CreationDate":"2018-10-17T06:35:25.603","ParentId":null,"OwnerUserId":"6918487","Title":null,"Body":"<p>you have to retrain with training data consisting of all 90 classes + the new classes you are interested in. As retraining overwrites previous trained classes knowledge\/ability on the final layers. Or just have two models one for the 90 classes and one for the new object that you trained on<\/p>\n"}]}
+{"QuestionId":52774120,"AnswerCount":1,"Tags":"<deep-learning><caffe><batch-normalization>","CreationDate":"2018-10-12T07:07:55.327","AcceptedAnswerId":null,"OwnerUserId":9714242.0,"Title":"Why Batch-normalization Layer follows scale layer in caffe?","Body":"<p>I noticed that Batch Normalization layer follows Scale layer in mobile net. It seems BN layer and scale layer are a pair.<br>\n And Convolution layer + BN layer + Scale layer + ReLU layer works well.<br>\nSo what scale layer do?<br>\nIt seems caffe can't learn parameters in BN layer, so Scale layer is useful, but why?<br>\nIn tensorflow doc, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/batch_norm\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/batch_norm<\/a>  <\/p>\n\n<blockquote>\n  <p>When the next layer is linear (also e.g. nn.relu), this can be\n  disabled since the scaling can be done by the next layer.<\/p>\n<\/blockquote>\n\n<p>It makes me more confuse.\nPlease help me, thanks!<\/p>\n","answers":[{"AnswerId":"52775512","CreationDate":"2018-10-12T08:37:11.343","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Batch Normalization does two things: First normalize with the mean and standard deviation of activations in a batch, and then perform scaling and bias to restore an appropriate range of the activations.<\/p>\n\n<p>Caffe implemented this with two layers, the Batch Normalization layer only does the normalization part, without the scaling and bias, which can be done with the scaling layer, or might not even be needed if the next layer can also do scaling (this is what TF docs mention).<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":52774351,"AnswerCount":2,"Tags":"<python><tensorflow><parallel-processing>","CreationDate":"2018-10-12T07:21:23.320","AcceptedAnswerId":null,"OwnerUserId":1080358.0,"Title":"How to run parallel map_fn when eager execution enabled","Body":"<p>Consider the following tensorflow code snippet:<\/p>\n\n<pre><code>import time\nimport numpy as np\nimport tensorflow as tf\n\ndef fn(i):\n    # do some junk work\n    for _ in range(100):\n        i ** 2\n    return i\n\nn = 1000\nn_jobs = 8\nstuff = np.arange(1, n + 1)\neager = False\nt0 = time.time()\nif eager:\n    tf.enable_eager_execution()\nres = tf.map_fn(fn, stuff, parallel_iterations=n_jobs)\nif not eager:\n    with tf.Session() as sess:\n        res = sess.run(res)\n        print(sum(res))\nelse:\n    print(sum(res))\ndt = time.time() - t0\nprint(\"(eager=%s) Took %ims\" % (eager, dt * 1000))\n<\/code><\/pre>\n\n<p>If run with <code>eager = True<\/code> it is 10x slower than when run with <code>eager = False<\/code>. I did some prints, and found out that in <code>eager = True<\/code> mode, the <code>map_fn<\/code> call is running sequential, instead of spawning 8 parallel threads.<\/p>\n\n<h1>Question<\/h1>\n\n<p>So my question is how to use <code>map_fn<\/code> (with parallel_iterations > 1) in eager execution mode ?<\/p>\n","answers":[{"AnswerId":"56253248","CreationDate":"2019-05-22T09:12:37.873","ParentId":null,"OwnerUserId":"10890778","Title":null,"Body":"<p>Update here for TF2.0 users. \nYou can parallelize the call to tf.map_fn inner operator by wrapping it into a tf.function decorator: <\/p>\n\n<pre><code>import tensorflow as tf\nimport time\n\nx = tf.ones(shape=(10,))\n\ndef op(x):\n    time.sleep(0.1)\n    return 1.\n\n_ = tf.map_fn(op, x, parallel_iterations=10) # will take 1 sec along with the \n                                             # warning message.\n\n# Now wrap tf.map_fn inside tf.function\n@tf.function\ndef my_map(*args, **kwargs):\n    return tf.map_fn(*args, **kwargs)\n\n_ = my_map(op, x, parallel_iterations=10) # will take 0.1 sec along with no \n                                          # warning message.\n<\/code><\/pre>\n"},{"AnswerId":"52785025","CreationDate":"2018-10-12T18:13:22.053","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>Crudely speaking, <code>tf.map_fn(fn, data)<\/code> is essentially shorthand for:<\/p>\n\n<pre><code>for e in data:\n  fn(e)\n<\/code><\/pre>\n\n<p>When eager execution is enabled, operations are executed as the Python interpreter encounters them, and thus there are no opportunities for \"whole program optimizations\".<\/p>\n\n<p>When executing TensorFlow graphs, the TensorFlow runtime sees the complete computation to be executed and can thus apply optimizations such as \"execute operations in <code>fn<\/code> from multiple iterations in the loop in parallel\". This is one of the benefits of expressing the computation as a graph.<\/p>\n\n<p>When eager execution in TensorFlow is enabled, you can still selectively apply graph optimizations to portions of your program using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/eager\/defun\" rel=\"nofollow noreferrer\"><code>tf.contrib.eager.defun<\/code><\/a>.<\/p>\n\n<p>For example (where most of the code is the same as yours above, and then a one line change to use <code>tf.contrib.eager.defun<\/code> to get graph optimization benefits):<\/p>\n\n<pre><code>import time\nimport numpy as np\nimport tensorflow as tf\n\ntf.enable_eager_execution()\n\ndef fn(i):\n    # do some junk work\n    for _ in range(100):\n        i ** 2\n    return i\n\nn = 1000\nn_jobs = 8\nstuff = np.arange(1, n + 1)\n\ndef my_computation(x):\n  return tf.map_fn(fn, x, parallel_iterations=n_jobs)\n\nt0 = time.time()\nmy_computation(stuff)\ndt = time.time() - t0\nprint(\"my_computation took %ims\" % (dt * 1000))\n\nmy_computation = tf.contrib.eager.defun(my_computation)\n# On the very first call, a graph is constructed, so let's discount\n# graph construction time\n_ = my_computation(stuff)\n\n# And then time it\nt0 = time.time()\nmy_computation(stuff)\ndt = time.time() - t0\nprint(\"my_computation took %ims\" % (dt * 1000))\n<\/code><\/pre>\n\n<p>Some additional things of note:<\/p>\n\n<ul>\n<li><p>In the particular example you've provided above, the TensorFlow runtime would probably also detect that <code>fn(i)<\/code> reduces to <code>return i<\/code> and can optimize away the unnecessary loop of <code>range(100)<\/code> since that does not affect the output. So the contrast in performance is quite a bit (as when executing <code>fn(i)<\/code> eagerly, the Python interpreter has no way of knowing that the <code>for<\/code> loop in there is useless, so it will execute that).<\/p><\/li>\n<li><p>If you change the computation in <code>fn()<\/code> to be something more meaningful, say:<\/p>\n\n<p>def fn(i):\n  for _ in range(2):\n    i = i ** 2\n  return i<\/p>\n\n<p>then you'll see a less stark difference.<\/p><\/li>\n<li><p>Note that not everything that can be expressed in Python can be \"defun\"ed. See documentation for <code>tf.contrib.eager.defun<\/code> for some detail and a more detailed spec and implementation is proposed for TensorFlow 2.0 (see <a href=\"https:\/\/github.com\/tensorflow\/community\/pull\/20\" rel=\"nofollow noreferrer\">RFC<\/a>)<\/p><\/li>\n<\/ul>\n\n<p>Hope that helps.<\/p>\n"}]}
+{"QuestionId":52775197,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><scientific-notation>","CreationDate":"2018-10-12T08:17:37.547","AcceptedAnswerId":null,"OwnerUserId":10451078.0,"Title":"Tensorflow model outputs probability values larger than 1","Body":"<p>I'm working on this classification program where i'm training my model to predict whether the object is a nut or a screw. I created my own dataset since i did not get any. I trained my model but i'm not getting correct predictions. Probability of values go beyond 1, basically i get garbage values. \nI get this predicted value: [[9.990779e-01 9.220659e-04]] <\/p>\n\n<p>#Training code<\/p>\n\n<pre><code>import dataset\nimport tensorflow as tf\nimport time\nfrom datetime import timedelta\nimport math\nimport random\nimport numpy as np\nimport os\n<\/code><\/pre>\n\n<p># Adding Seed so that random initialization is consistent<br>\n    from numpy.random import seed<\/p>\n\n<pre><code>seed(1)\nfrom tensorflow import set_random_seed\n\nset_random_seed(2)\n\nbatch_size = 20\n\n# Prepare input data\nclasses = os.listdir('training_set')\nnum_classes = len(classes)\n\n# 20% of the data will automatically be used for validation\nvalidation_size = 0.2\nimg_size = 128\nnum_channels = 3\ntrain_path = 'training_set'\n\n# We shall load all the training and validation images and labels into \n memory using openCV and use that during training\ndata = dataset.read_train_sets(train_path, img_size, classes, \nvalidation_size=validation_size)\n\nprint(\"Complete reading input data. Will Now print a snippet of it\")\nprint(\"Number of files in Training- \nset:\\t\\t{}\".format(len(data.train.labels)))\nprint(\"Number of files in Validation- \nset:\\t{}\".format(len(data.valid.labels)))\n\nsession = tf.Session()\nx = tf.placeholder(tf.float32, shape=[None, img_size, img_size, \n    num_channels], name='x')\n\n## labels\ny_true = tf.placeholder(tf.float32, shape=[None, num_classes], \n         name='y_true')\ny_true_cls = tf.argmax(y_true, dimension=1)\n\n##Network graph params\nfilter_size_conv1 = 3\nnum_filters_conv1 = 32\n\nfilter_size_conv2 = 3\nnum_filters_conv2 = 32\n\nfilter_size_conv3 = 3\nnum_filters_conv3 = 32\n\nfc_layer_size = 128\n\n\ndef create_weights(shape):\n    return tf.Variable(tf.truncated_normal(shape, stddev=0.05))\n\n\ndef create_biases(size):\n    return tf.Variable(tf.constant(0.05, shape=[size]))\n\n\ndef create_convolutional_layer(input,\n                           num_input_channels,\n                           conv_filter_size,\n                           num_filters):\n## We shall define the weights that will be trained using create_weights function.\nweights = create_weights(shape=[conv_filter_size, conv_filter_size, num_input_channels, num_filters])\n## We create biases using the create_biases function. These are also trained.\nbiases = create_biases(num_filters)\n\n## Creating the convolutional layer\nlayer = tf.nn.conv2d(input=input,\n                     filter=weights,\n                     strides=[1, 1, 1, 1],\n                     padding='SAME')\n\nlayer += biases\n\n## We shall be using max-pooling.\nlayer = tf.nn.max_pool(value=layer,\n                       ksize=[1, 2, 2, 1],\n                       strides=[1, 2, 2, 1],\n                       padding='SAME')\n## Output of pooling is fed to Relu which is the activation function for us.\nlayer = tf.nn.relu(layer)\n\nreturn layer\n\n\ndef create_flatten_layer(layer):\n# We know that the shape of the layer will be [batch_size img_size img_size num_channels]\n# But let's get it from the previous layer.\n    layer_shape = layer.get_shape()\n\n## Number of features will be img_height * img_width* num_channels. But we shall calculate it in place of hard-coding it.\n    num_features = layer_shape[1:4].num_elements()\n\n## Now, we Flatten the layer so we shall have to reshape to num_features\n    layer = tf.reshape(layer, [-1, num_features])\n\nreturn layer\n\n\ndef create_fc_layer(input,\n                num_inputs,\n                num_outputs,\n                use_relu=True):\n# Let's define trainable weights and biases.\n     weights = create_weights(shape=[num_inputs, num_outputs])\n     biases = create_biases(num_outputs)\n\n# Fully connected layer takes input x and produces wx+b.Since, these are matrices, we use matmul function in Tensorflow\nlayer = tf.matmul(input, weights) + biases\nif use_relu:\n    layer = tf.nn.relu(layer)\n\nreturn layer\n\n\nlayer_conv1 = create_convolutional_layer(input=x,\n                                     num_input_channels=num_channels,\n                                     conv_filter_size=filter_size_conv1,\n                                     num_filters=num_filters_conv1)\nlayer_conv2 = create_convolutional_layer(input=layer_conv1,\n                                     num_input_channels=num_filters_conv1,\n                                     conv_filter_size=filter_size_conv2,\n                                     num_filters=num_filters_conv2)\n\nlayer_conv3 = create_convolutional_layer(input=layer_conv2,\n                                     num_input_channels=num_filters_conv2,\n                                     conv_filter_size=filter_size_conv3,\n                                     num_filters=num_filters_conv3)\n\nlayer_flat = create_flatten_layer(layer_conv3)\n\nlayer_fc1 = create_fc_layer(input=layer_flat,\n                        num_inputs=layer_flat.get_shape() \n[1:4].num_elements(),\n                        num_outputs=fc_layer_size,\n                        use_relu=True)\n\nlayer_fc2 = create_fc_layer(input=layer_fc1,\n                        num_inputs=fc_layer_size,\n                        num_outputs=num_classes,\n                        use_relu=False)\n\ny_pred = tf.nn.softmax(layer_fc2, name='y_pred')\n\ny_pred_cls = tf.argmax(y_pred, dimension=1)\nsession.run(tf.global_variables_initializer())\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2,\n                                                    labels=y_true)\ncost = tf.reduce_mean(cross_entropy)\noptimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cost)\ncorrect_prediction = tf.equal(y_pred_cls, y_true_cls)\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\nsession.run(tf.global_variables_initializer())\n\n\ndef show_progress(epoch, feed_dict_train, feed_dict_validate, val_loss):\n    acc = session.run(accuracy, feed_dict=feed_dict_train)\n    val_acc = session.run(accuracy, feed_dict=feed_dict_validate)\n    msg = \"Training Epoch {0} --- Training Accuracy: {1:&gt;6.1%}, Validation \nAccuracy: {2:&gt;6.1%},  Validation Loss: {3:.3f}\"\n    print(msg.format(epoch + 1, acc, val_acc, val_loss))\n\n\ntotal_iterations = 0\n\nsaver = tf.train.Saver()\n\n\ndef train(num_iteration):\n    global total_iterations\n\n    for i in range(total_iterations,\n                   total_iterations + num_iteration):\n\n        x_batch, y_true_batch, _, cls_batch = \n            data.train.next_batch(batch_size)\n        x_valid_batch, y_valid_batch, _, valid_cls_batch = \n             data.valid.next_batch(batch_size)\n\n        feed_dict_tr = {x: x_batch,\n                    y_true: y_true_batch}\n        feed_dict_val = {x: x_valid_batch,\n                     y_true: y_valid_batch}\n\n        session.run(optimizer, feed_dict=feed_dict_tr)\n\n        if i % int(data.train.num_examples \/ batch_size) == 0:\n            val_loss = session.run(cost, feed_dict=feed_dict_val)\n            epoch = int(i \/ int(data.train.num_examples \/ batch_size))\n\n            show_progress(epoch, feed_dict_tr, feed_dict_val, val_loss)\n            saver.save(session, 'C:\\\\Nutsbolts\\\\nuts-screws-model')\n\n    total_iterations += num_iteration\n\n\ntrain(num_iteration=3000)\n\n\n\n\n\n#Prediction code\n\nimport tensorflow as tf\nimport numpy as np\nimport os,glob,cv2\nimport sys,argparse\n\n\n# First, pass the path of the image\ndir_path = 'C:\\\\nutsbolts\\\\testing_set\\\\nuts'\nimage_path= 'nuts11.jpg'\nfilename = dir_path +'\/' +image_path\nimage_size=128\nnum_channels=3\nimages = []\n# Reading the image using OpenCV\nimage = cv2.imread(filename)\n# Resizing the image to our desired size and preprocessing will be done \nexactly as done during training\nimage = cv2.resize(image, (image_size, image_size),0,0, cv2.INTER_LINEAR)\nimages.append(image)\nimages = np.array(images, dtype=np.uint8)\nimages = images.astype('float32')\nimages = np.multiply(images, 1.0\/255.0) \n#The input to the network is of shape [None image_size image_size \nnum_channels]. Hence we reshape.\nx_batch = images.reshape(1, image_size,image_size,num_channels)\n\n## Let us restore the saved model \nsess = tf.Session()\n# Step-1: Recreate the network graph. At this step only graph is created.\nsaver = tf.train.import_meta_graph('nuts-screws-model.meta')\n# Step-2: Now let's load the weights saved using the restore method.\nsaver.restore(sess, tf.train.latest_checkpoint('.\/'))\n\n# Accessing the default graph which we have restored\ngraph = tf.get_default_graph()\n\n# Now, let's get hold of the op that we can be processed to get the output.\n# In the original network y_pred is the tensor that is the prediction of the \nnetwork\ny_pred = graph.get_tensor_by_name(\"y_pred:0\")\n\n## Let's feed the images to the input placeholders\nx= graph.get_tensor_by_name(\"x:0\") \ny_true = graph.get_tensor_by_name(\"y_true:0\") \ny_test_images = np.zeros((1, len(os.listdir('testing_set'))))\n\n\n### Creating the feed_dict that is required to be fed to calculate y_pred \nfeed_dict_testing = {x: x_batch, y_true: y_test_images}\nresult=sess.run(y_pred, feed_dict=feed_dict_testing)\n# result is of this format [probabiliy_of_nuts probability_of_screws]\nprint(result)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52775396","CreationDate":"2018-10-12T08:30:45.693","ParentId":null,"OwnerUserId":"10401887","Title":null,"Body":"<p>9.990779e-01 actually is below 1.  You could see it as: 9.990779 * (the exponential of -01).<\/p>\n"}]}
+{"QuestionId":52776332,"AnswerCount":1,"Tags":"<python><tensorflow><nvidia>","CreationDate":"2018-10-12T09:21:24.027","AcceptedAnswerId":52776496.0,"OwnerUserId":10437375.0,"Title":"Tensorflow-GPU without NVIDIA\uff0cpossible\uff1f","Body":"<p>I don't have a NVIDIA graphics card, but I need to use tensorflow-gpu. Is this feasible? what should I do?<\/p>\n","answers":[{"AnswerId":"52776496","CreationDate":"2018-10-12T09:30:17.137","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>You can use Google's colab for free directly in your browser:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/colab.research.google.com\/notebooks\/welcome.ipynb#recent=true\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/notebooks\/welcome.ipynb#recent=true<\/a>\n  <a href=\"https:\/\/colab.research.google.com\/notebooks\/gpu.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/notebooks\/gpu.ipynb<\/a><\/p>\n<\/blockquote>\n\n<p>You can create easily a notebook to run tensorflow (wich is already installed) from the file menu.<\/p>\n\n<p>To activate the GPU simply select \"GPU\" in the Accelerator drop-down in Notebook Settings (either through the Edit menu or the command palette at cmd\/ctrl-shift-P).<\/p>\n"}]}
+{"QuestionId":52776469,"AnswerCount":1,"Tags":"<tensorflow><model><serving>","CreationDate":"2018-10-12T09:28:41.220","AcceptedAnswerId":null,"OwnerUserId":10476806.0,"Title":"How inference an inception pb model by restful API with tensorflow serving?","Body":"<p>I have got an incepiton pb Model by github:\"serving\/tensorflow_serving\/example\/inception_saved_model.py\".I want to got predict by restful API ,but found  an error message.I want to know the argument \"images\" that how to do or what is correctly format data ?<a href=\"https:\/\/i.stack.imgur.com\/dU3Pc.png\" rel=\"nofollow noreferrer\">enter image description here<\/a> <\/p>\n","answers":[{"AnswerId":"52792470","CreationDate":"2018-10-13T11:32:41.987","ParentId":null,"OwnerUserId":"3337758","Title":null,"Body":"<p>Make sure to encode the string so that it is websafe. Web-safe means that the encoder uses - and _ instead of + and \/.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/io\/encode_base64\" rel=\"nofollow noreferrer\">the documentation<\/a> for more info.<\/p>\n"}]}
+{"QuestionId":52776622,"AnswerCount":2,"Tags":"<python-3.x><keras><checkpointing>","CreationDate":"2018-10-12T09:36:43.803","AcceptedAnswerId":52776963.0,"OwnerUserId":4399454.0,"Title":"Keras callbacks keep skip saving checkpoints, claiming val_acc is missing","Body":"<p>I'll run some larger models and want to try intermediate results.<\/p>\n\n<p>Therefore, I try to use checkpoints to save the best model after each epoch.<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(700, input_shape=(X_modified.shape[1], X_modified.shape[2]), return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(700, return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(700))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(Y_modified.shape[1], activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# Save the checkpoint in the \/output folder\nfilepath = \"output\/text-gen-best.hdf5\"\n\n# Keep only a single checkpoint, the best over test accuracy.\ncheckpoint = ModelCheckpoint(filepath,\n                            monitor='val_acc',\n                            verbose=1,\n                            save_best_only=True,\n                            mode='max')\nmodel.fit(X_modified, Y_modified, epochs=100, batch_size=50, callbacks=[checkpoint])\n<\/code><\/pre>\n\n<p>But I am still getting the warning after the first epoch:<\/p>\n\n<pre><code>\/usr\/local\/lib\/python3.6\/site-packages\/keras\/callbacks.py:432: RuntimeWarning: Can save best model only with val_acc available, skipping.\n  'skipping.' % (self.monitor), RuntimeWarning)\n<\/code><\/pre>\n\n<p>To add <code>metrics=['accuracy']<\/code> to the model was in other SO questions (e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/46438247\/unable-to-save-weights-while-using-pre-trained-vgg16-model?rq=1\">Unable to save weights while using pre-trained VGG16 model<\/a>) the solution, but here the error still remains.<\/p>\n","answers":[{"AnswerId":"52776963","CreationDate":"2018-10-12T09:54:28.583","ParentId":null,"OwnerUserId":"7896849","Title":null,"Body":"<p>You are trying to checkpoint the model using the following code<\/p>\n\n<pre><code># Save the checkpoint in the \/output folder\nfilepath = \"output\/text-gen-best.hdf5\"\n\n# Keep only a single checkpoint, the best over test accuracy.\ncheckpoint = ModelCheckpoint(filepath,\n                            monitor='val_acc',\n                            verbose=1,\n                            save_best_only=True,\n                            mode='max')\n<\/code><\/pre>\n\n<p><code>ModelCheckpoint<\/code> will consider the argument <code>monitor<\/code> to take the decision of saving the model or not. In your code it is <code>val_acc<\/code>. So it will save the weights if there is a increase in the <code>val_acc<\/code>.<\/p>\n\n<p>Now in your fit code,<\/p>\n\n<pre><code>model.fit(X_modified, Y_modified, epochs=100, batch_size=50, callbacks=[checkpoint])\n<\/code><\/pre>\n\n<p>you haven't provided any validation data. <code>ModelCheckpoint<\/code> can't save the weights because it don't have the <code>monitor<\/code> argument to check.<\/p>\n\n<p>In order to do check pointing based on <code>val_acc<\/code> you must provide some validation data like this.<\/p>\n\n<pre><code>model.fit(X_modified, Y_modified, validation_data=(X_valid, y_valid), epochs=100, batch_size=50, callbacks=[checkpoint])\n<\/code><\/pre>\n\n<p>If you don't want to use validation data for whatever be the reason and implement check pointing, you have to change the <code>ModelCheckpoint<\/code> to work based on <code>acc<\/code> or <code>loss<\/code> like this<\/p>\n\n<pre><code># Save the checkpoint in the \/output folder\nfilepath = \"output\/text-gen-best.hdf5\"\n\n# Keep only a single checkpoint, the best over test accuracy.\ncheckpoint = ModelCheckpoint(filepath,\n                            monitor='acc',\n                            verbose=1,\n                            save_best_only=True,\n                            mode='max')\n<\/code><\/pre>\n\n<p>Keep in mind that you have to change <code>mode<\/code> to <code>min<\/code> if you are going to <code>monitor<\/code> the <code>loss<\/code><\/p>\n"},{"AnswerId":"52776789","CreationDate":"2018-10-12T09:44:59.483","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>It is missing, not because the metric is missing, but because you have no validation data. Add some through the <code>validation_data<\/code> parameter to <code>fit<\/code>, or use <code>validation_split<\/code>.<\/p>\n"}]}
+{"QuestionId":52777644,"AnswerCount":1,"Tags":"<python><keras><pre-trained-model>","CreationDate":"2018-10-12T10:32:17.930","AcceptedAnswerId":null,"OwnerUserId":5811273.0,"Title":"Addition of data to Pre-Trained model","Body":"<p>For the data-set I'm working with, I have trained and saved a h5py model already using Keras. Now I have to add new data to the pre-trained model and use this new data in the training set. But I do not want to re-train the whole data-set as it took about 7 hours to train and save the model already. What are the methods that are currently available to add any new data to already trained model?<\/p>\n\n<p>I do not want to retrain the whole model is because I do not have large data-set to add. I want to include the new data without training the model from the scratch because of time constraints.<\/p>\n\n<p>How to add this new data?<\/p>\n","answers":[{"AnswerId":"52777750","CreationDate":"2018-10-12T10:38:26.070","ParentId":null,"OwnerUserId":"8858832","Title":null,"Body":"<p>I am hoping that you used model.save(), If you did then you can<\/p>\n\n<pre><code>from keras.models import load_model\nmodel=load_model(&lt;your path&gt;)\n<\/code><\/pre>\n\n<p>and it is just your regular model, you can train it, predict with it whatever you want<\/p>\n\n<pre><code>model.compile(optimizer='adam',loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(train_data, train_labels,epochs=epochs,batch_size=batch_size,validation_data=(validation_data, validation_labels))\n<\/code><\/pre>\n\n<p>and whatever you want to do <\/p>\n"}]}
+{"QuestionId":52778922,"AnswerCount":2,"Tags":"<machine-learning><keras><time-series><lstm><gated-recurrent-unit>","CreationDate":"2018-10-12T11:50:22.460","AcceptedAnswerId":null,"OwnerUserId":7405050.0,"Title":"stock prediction : GRU model predicting same given values instead of future stock price","Body":"<p>i was just testing this model from kaggle <a href=\"https:\/\/www.kaggle.com\/thebrownviking20\/intro-to-recurrent-neural-networks-lstm-gru\" rel=\"nofollow noreferrer\">post<\/a> <strong>this model suppose to predict 1 day ahead from given set of last stocks<\/strong>. After tweaking few parameters i got surprisingly good result, as you can see.\n<a href=\"https:\/\/i.stack.imgur.com\/aR4JT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aR4JT.png\" alt=\"enter image description here\"><\/a>\nmean squared error was 5.193.so overall it looks good at predicting future stocks right? well it turned out to be horrible when i take a look closely on the results.<br><\/p>\n\n<p>as you can see that this model is predicting last value of the given stocks which is our current last stock.<br>\nso i did adjusted predictions to one step back..\n<a href=\"https:\/\/i.stack.imgur.com\/mDACD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mDACD.png\" alt=\"enter image description here\"><\/a> \nso now you can clearly see that model is predicting one step backward or last stock prise instead of future stock predictions.<\/p>\n\n<h2>This is my training data<\/h2>\n\n<pre><code># So for each element of training set, we have 30 previous training set elements \nX_train = []\ny_train = []\n\nprevious = 30\n\nfor i in range(previous,len(training_set_scaled)):\n    X_train.append(training_set_scaled[i-previous:i,0])\n    y_train.append(training_set_scaled[i,0])\nX_train, y_train = np.array(X_train), np.array(y_train)\n\n\nprint(X_train[-1],y_train[-1])\n<\/code><\/pre>\n\n<h2>This is my model<\/h2>\n\n<pre><code># The GRU architecture\nregressorGRU = Sequential()\n# First GRU layer with Dropout regularisation\nregressorGRU.add(GRU(units=50, return_sequences=True, input_shape=(X_train.shape[1],1)))\nregressorGRU.add(Dropout(0.2))\n# Second GRU layer\nregressorGRU.add(GRU(units=50, return_sequences=True))\nregressorGRU.add(Dropout(0.2))\n# Third GRU layer\nregressorGRU.add(GRU(units=50, return_sequences=True))\nregressorGRU.add(Dropout(0.2))\n# Fourth GRU layer\nregressorGRU.add(GRU(units=50))\nregressorGRU.add(Dropout(0.2))\n# The output layer\nregressorGRU.add(Dense(units=1))\n\n# Compiling the RNN\nregressorGRU.compile(optimizer='adam',loss='mean_squared_error')\n# Fitting to the training set\nregressorGRU.fit(X_train,y_train,epochs=50,batch_size=32)\n<\/code><\/pre>\n\n<p><strong>And <a href=\"https:\/\/colab.research.google.com\/drive\/1mhsbBTXEwUv9pb1N7MUpFIQ8r73_6-YB\" rel=\"nofollow noreferrer\">here<\/a> is my full code, you also able to run this code at <code>google colab<\/code>.<\/strong><\/p>\n\n<p>so my question is what is the reason behind it? what am i doing wrong any suggestions?<\/p>\n","answers":[{"AnswerId":"52786399","CreationDate":"2018-10-12T20:11:43.850","ParentId":null,"OwnerUserId":"10491777","Title":null,"Body":"<p>It is a well-known issue with regression actually. Since the task of the regressor is to minimize error, it secures it task by choosing the closest value from the features you input to the regressor.   It becomes the case especially in the time-series problems.<\/p>\n\n<p>1) Never give unprocessed closing value that you want your model to predict, especially in the time-series regression problems. More generally, never give a feature that gives some direct numerical intuition to a regressor about what the label might be.<\/p>\n\n<p>2)If you are not sure whether the model just replicates like your case, be sure to plot the original test set and your prediction all together to visually analize the situation. Moreover, if you can, do a simulation of your model on the real-time data to observe whether your model predicts with the same performance.<\/p>\n\n<p>3)I\u2019d recommend you to apply binary classification rather than regression. <\/p>\n\n<p>I\u2019ve been intensely working on financial signal prediction for nearly a year, do not hesitate to ask more.<\/p>\n\n<p>Have fun.<\/p>\n"},{"AnswerId":"54836042","CreationDate":"2019-02-22T22:26:24.473","ParentId":null,"OwnerUserId":"8679402","Title":null,"Body":"<p>In my opinion this is happening because your loss function here is mean squared error. The best way to minimize this error is by using the previous value as the estimate for the current step (especially if the steps are nearby). <\/p>\n\n<p>You will start seeing the limitations of your model clearly if you try to predict the delta between samples.<\/p>\n\n<p>I am currently experiencing similar issues trying to predict using time series. I haven't found a solution yet, but am working on extracting features as others pointed out.<\/p>\n"}]}
+{"QuestionId":52779602,"AnswerCount":1,"Tags":"<python><keras><autoencoder>","CreationDate":"2018-10-12T12:30:21.317","AcceptedAnswerId":52780095.0,"OwnerUserId":9422346.0,"Title":"Error- AttributeError: 'DirectoryIterator' object has no attribute 'ndim in autoencoder design in keras","Body":"<p>I am a newbie in Python 3.5. I am trying to program a simple auto-encoder which will train on a data-set of 60 images of apple and try to reconstruct the image given in the root. I have used the following codes : <\/p>\n\n<pre><code>from keras.layers import Input, Dense\nfrom keras.models import Model\nimport numpy as np\nfrom PIL import Image \nfrom keras.preprocessing.image import ImageDataGenerator\nimport matplotlib.pyplot as plt\nimage = Image.open('C:\\Python35\\Scripts\\apple.jpg')\nencoding_dim = 32\ninput_img = Input(shape=(65536,))\nencoded = Dense(encoding_dim, activation='relu')(input_img)\ndecoded = Dense(65536, activation='sigmoid')(encoded)\nautoencoder = Model(input_img, decoded)\nencoder = Model(input_img, encoded)\nencoded_input = Input(shape=(encoding_dim,))\ndecoder_layer = autoencoder.layers[-1]\ndecoder = Model(encoded_input, decoder_layer(encoded_input))\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\ntrain_datagen=ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(\n    directory=r\"C:\\Users\\vlsi\\Desktop\\train\",\n    batch_size=32,\n    class_mode=\"categorical\",\n    shuffle=True,\n    seed=42\n)\nautoencoder.fit(train_generator,\n                epochs=2,\n                batch_size=256,\n                shuffle=True)\nencoded_img = encoder.predict(np.array(image))\ndecoded_img = decoder.predict(encoded_img)\nplt.imshow(decoded_img)\n<\/code><\/pre>\n\n<p>It gives an error <\/p>\n\n<blockquote>\n  <p>AttributeError: 'DirectoryIterator' object has no attribute 'ndim'<\/p>\n<\/blockquote>\n\n<p>Any idea what went wrong?<\/p>\n","answers":[{"AnswerId":"52780095","CreationDate":"2018-10-12T13:01:49.900","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The Keras <code>fit<\/code> function takes arrays of data, numpy arrays, not generators. The function you need is <code>fit_generator<\/code>. Note that  <code>fit_generator<\/code> takes slightly different parameters, such as <code>steps_per_epoch<\/code> instead of <code>batch_size<\/code>.<\/p>\n"}]}
+{"QuestionId":52779632,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-12T12:32:20.010","AcceptedAnswerId":null,"OwnerUserId":10333833.0,"Title":"How to improve model's accuracy?","Body":"<p>My model has one hidden layer having 100 neurons, I have used sigmoid as an activation function and in last layer I used softmax. I want to classify my data into two classes. Output layer has two neurons. Can someone please guide me what should I do to improve my model's accuracy?<\/p>\n\n<pre><code>`step: 9000, train_acc: 0.82,test_acc: 0.765\nstep: 9100, train_acc: 0.67,test_acc: 0.72\nstep: 9200, train_acc: 0.66,test_acc: 0.674\nstep: 9300, train_acc: 0.75,test_acc: 0.737\nstep: 9400, train_acc: 0.61,test_acc: 0.613\nstep: 9500, train_acc: 0.82,test_acc: 0.712\nstep: 9600, train_acc: 0.81,test_acc: 0.765\nstep: 9700, train_acc: 0.77,test_acc: 0.773\nstep: 9800, train_acc: 0.76,test_acc: 0.758\nstep: 9900, train_acc: 0.72,test_acc: 0.75`\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52779736,"AnswerCount":3,"Tags":"<python><tensorflow><protocol-buffers><tensorflow-data-validation>","CreationDate":"2018-10-12T12:39:05.730","AcceptedAnswerId":52788456.0,"OwnerUserId":2175173.0,"Title":"TFDV Tensorflow Data Validation: how can I save\/load the protobuf schema to\/from a file","Body":"<p>TFDV generates schema as a Schema protocol buffer. However it seems that there is no helper function to write\/read schema to\/from a file.<\/p>\n\n<pre><code>schema = tfdv.infer_schema(stats)\n<\/code><\/pre>\n\n<p>How can I save it\/load it ?<\/p>\n","answers":[{"AnswerId":"57669620","CreationDate":"2019-08-27T07:25:11.783","ParentId":null,"OwnerUserId":"2708182","Title":null,"Body":"<p>tensorflow_data_validation itself provides you util functions for this:<\/p>\n\n<pre><code>from tensorflow_data_validation.utils.schema_util import write_schema_text, load_schema_text\nwrite_schema_text(schema, \".\/my_schema\")\nschema = load_schema_text(\".\/my_schema\")\n<\/code><\/pre>\n"},{"AnswerId":"55514622","CreationDate":"2019-04-04T11:17:55.460","ParentId":null,"OwnerUserId":"8515812","Title":null,"Body":"<p>If you will be using it with <strong>Tensorflow Transform<\/strong> then I would suggest the following functions:<\/p>\n\n<pre><code>import tensorflow_data_validation as tfdv\nfrom tensorflow.python.lib.io import file_io\nfrom tensorflow_transform.tf_metadata import metadata_io\n\n# Define file path\nfile_io.recursive_create_dir(OUTPUT_DIR)\nschema_file = os.path.join(OUTPUT_DIR, 'schema.pbtxt')\n\n# Write schema\ntfdv.write_schema_text(schema, schema_file)\n\n# Read schema with tfdv\nschema = tfdv.load_schema_text(schema_file)\n\n# Read schema with tensorflow_transform\nschema = metadata_io.read_metadata(OUTPUT_DIR)\n<\/code><\/pre>\n\n<p>The output is human-readable - similar to JSON. But if you prefer to save it in plain <strong>JSON format<\/strong> then you can use the following:<\/p>\n\n<pre><code>from google.protobuf import json_format\nfrom tensorflow.python.lib.io import file_io\nfrom tensorflow_metadata.proto.v0 import schema_pb2\n\ndef write_schema(schema, output_path):\n    schema_text = json_format.MessageToJson(schema)\n    file_io.write_string_to_file(output_path, schema_text)\n\ndef load_schema(input_path):\n    schema_text = file_io.read_file_to_string(input_path)\n    schema = json_format.Parse(schema_text, schema_pb2.Schema())\n    return schema   \n<\/code><\/pre>\n\n<p>Or if you don't need it to be in human-readable format you can use SerializeToString() and ParseFromString(data) for de\/serialization like described <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/pythontutorial#parsing-and-serialization\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"52788456","CreationDate":"2018-10-13T00:21:35.647","ParentId":null,"OwnerUserId":"6447747","Title":null,"Body":"<p>You can use the following methods to write\/load the schema to\/from a file. <\/p>\n\n<pre><code>from google.protobuf import text_format\nfrom tensorflow.python.lib.io import file_io\nfrom tensorflow_metadata.proto.v0 import schema_pb2\n\ndef write_schema(schema, output_path):\n  schema_text = text_format.MessageToString(schema)\n  file_io.write_string_to_file(output_path, schema_text)\n\ndef load_schema(input_path):\n  schema = schema_pb2.Schema()\n  schema_text = file_io.read_file_to_string(input_path)\n  text_format.Parse(schema_text, schema)\n  return schema      \n<\/code><\/pre>\n"}]}
+{"QuestionId":52780012,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-12T12:57:19.373","AcceptedAnswerId":null,"OwnerUserId":6840039.0,"Title":"Tensorflow: logistic regression to mnist","Body":"<p>I try to use logistic regression to mnist dataset, but I have some problem with realization<\/p>\n\n<pre><code>from sklearn.datasets import load_digits\nfrom sklearn.metrics import roc_auc_score\n\ns = tf.InteractiveSession(config=tf.ConfigProto(gpu_options=gpu_options))\n\nmnist = load_digits(2)\n\nX,y = mnist.data, mnist.target\n\n# inputs and shareds\nweights = tf.Variable(np.zeros((64, 1), dtype='float32'))\ninput_X = tf.placeholder(tf.float32, shape=(None, 64))\ninput_y = tf.placeholder(tf.float32, shape=(None, 1))\n\npredicted_y = tf.add(tf.matmul(input_X, weights), input_y)\nloss = tf.losses.log_loss(input_y, predicted_y)\n\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=1e-2).minimize(loss)\ntrain_function = s.run(optimizer, feed_dict={input_X: X, input_y: y.reshape(-1, 1)})\npredict_function = s.run(predicted_y, feed_dict={input_X: X})\n\nX_train, X_test, y_train, y_test = train_test_split(X, y)\n\nfor i in range(5):\n    s.run(optimizer)\n    loss_i = s.run(loss, feed_dict={input_X: X_test, input_y: y_test})\n\n    print(\"loss at iter %i:%.4f\" % (i, loss_i))\n\n    print(\"train auc:\",roc_auc_score(y_train, predict_function(X_train)))\n    print(\"test auc:\",roc_auc_score(y_test, predict_function(X_test)))\n\n\nprint (\"resulting weights:\")\nplt.imshow(weights.eval().reshape([8,8]))\nplt.colorbar()\n<\/code><\/pre>\n\n<p>The problem appears when I try to run <code>train_prediction<\/code>. It returns<\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value Variable_16\n [[Node: Variable_16\/read = Identity[T=DT_FLOAT, _class=[\"loc:@Variable_16\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Variable_16)]]\n<\/code><\/pre>\n\n<p>I try to change shape of the inputs, but I can't understand, what's wrong.<\/p>\n","answers":[{"AnswerId":"52782034","CreationDate":"2018-10-12T14:49:17.990","ParentId":null,"OwnerUserId":"7943474","Title":null,"Body":"<p>TensorFlow documentation explicitly says:<\/p>\n\n<blockquote>\n  <p>When you launch the graph, variables have to be explicitly initialized before you can run Ops that use their value. You can initialize a variable by running its initializer op, restoring the variable from a save file, or simply running an assign Op that assigns a value to the variable. In fact, the variable initializer op is just an assign Op that assigns the variable's initial value to the variable itself.<\/p>\n<\/blockquote>\n\n<p>Therefore, after opening a TF session, you must initialize your <code>weight<\/code> variable running one of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/initializers\" rel=\"nofollow noreferrer\">Initializer Op<\/a>.<\/p>\n\n<p>I suggest you to define new variables using <code>tf.get_variable()<\/code> which create a new variable or retrieve an existing one given the name you provide as argument. In its parameter list, there's the possibility to pass a <code>tf.initializer<\/code>, but if you do not do that, it will load the default initializer.<\/p>\n"}]}
+{"QuestionId":52780110,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-12T13:03:05.403","AcceptedAnswerId":null,"OwnerUserId":10362955.0,"Title":"Keras load_model and get_default_graph.finalize()","Body":"<p>I have a trouble when I'm trying finalize loaded graph. My pipeline is next:\n<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.models import load_model\n\nmodel = load_model(\"model\")\nmodel.load_weights(\"weights\")\ngraph = tf.get_default_graph()\ngraph.finalize()\n<\/code><\/pre>\n\n<p>But I have error like this:<\/p>\n\n<pre><code>RuntimeError: Graph is finalized and cannot be modified.\n<\/code><\/pre>\n\n<p>The question is why it is happened? I'm trying finalize graph AFTER loading model and weights and it should work correct, so I can't get this behaviour.<\/p>\n","answers":[{"AnswerId":"54516480","CreationDate":"2019-02-04T12:48:40.393","ParentId":null,"OwnerUserId":"8165927","Title":null,"Body":"<p>If you are trying to load your model inside within multithreading you need to clear session<\/p>\n\n<pre><code>from keras import backend as K\nK.clear_session()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52780559,"AnswerCount":1,"Tags":"<python><optimization><pytorch><torch>","CreationDate":"2018-10-12T13:26:59.540","AcceptedAnswerId":52898702.0,"OwnerUserId":3719101.0,"Title":"Outer sum, etc. in pytorch","Body":"<p>Numpy offers optimized outer operations for any <code>RxR -&gt; R<\/code> function, like <code>np.multiply.outer<\/code> or <code>np.subtract.outer<\/code>, with the behaviour:<\/p>\n\n<pre><code>&gt;&gt;&gt; np.subtract.outer([6, 5, 4], [3, 2, 1])\narray([[3, 4, 5],\n       [2, 3, 4],\n       [1, 2, 3]])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/index.html\" rel=\"nofollow noreferrer\">Pytorch<\/a>  does not seem to offer such a feature (or I have missed it).<br>\nWhat is the best \/ usual \/ fastest \/ cleanest way to do so with torch tensors?<\/p>\n","answers":[{"AnswerId":"52898702","CreationDate":"2018-10-19T19:20:26.457","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p>Per the <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/broadcasting.html\" rel=\"nofollow noreferrer\">documenation<\/a>:<\/p>\n\n<blockquote>\n  <p>Many PyTorch operations support NumPy Broadcasting Semantics.<\/p>\n<\/blockquote>\n\n<p>An outer subtraction is a broadcasted subtraction from a 2d array to a 1d array, so essentially you can reshape the first array to (3, 1) and then subtract the second array from it:<\/p>\n\n<pre><code>x = torch.Tensor([6, 5, 4])\ny = torch.Tensor([3, 2, 1])\n\nx.reshape(-1, 1) - y\n#tensor([[3., 4., 5.],\n#        [2., 3., 4.],\n#        [1., 2., 3.]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52780606,"AnswerCount":0,"Tags":"<python><image><tensorflow><kaggle>","CreationDate":"2018-10-12T13:29:37.403","AcceptedAnswerId":null,"OwnerUserId":8375248.0,"Title":"KERAS: AttributeError: 'BatchDataset' object has no attribute 'ndim' , Tensorflow image segmentation demo does not work on my anaconda","Body":"<p>I know there is an issue with the line<\/p>\n\n<p><code>history = model.fit(train_ds,\n                   steps_per_epoch=int(np.ceil(num_train_examples \/ float(batch_size))),\n                   epochs=epochs,\n                   validation_data=val_ds,\n                   validation_steps=int(np.ceil(num_val_examples \/ float(batch_size))),\n                   callbacks=[cp])\n<\/code>\nI get the error \nAttributeError: 'BatchDataset' object has no attribute 'ndim' and am using tensorflows demo : image segmentation.\nI think it's to do with the input dimensions but can't find a way out of the problem<\/p>\n\n<p>information for database:<\/p>\n\n<pre><code>train_ds = get_baseline_dataset(x_train_filenames,\n                            y_train_filenames,\n                            preproc_fn=tr_preprocessing_fn,\n                            batch_size=batch_size)\n<\/code><\/pre>\n\n<p>val_ds = get_baseline_dataset(x_val_filenames,\n                              y_val_filenames, \n                              preproc_fn=val_preprocessing_fn,\n                              batch_size=batch_size)<\/p>\n\n<p>def get_baseline_dataset(filenames, \n                         labels,\n                         preproc_fn=functools.partial(_augment),\n                         threads=5, \n                         batch_size=batch_size,\n                         shuffle=True):    <\/p>\n\n<p>thanks in advance  (ps. it works in google colab but just not on anaconda\/spyder)<\/p>\n","answers":[]}
+{"QuestionId":52781277,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><machine-learning><keras>","CreationDate":"2018-10-12T14:05:38.003","AcceptedAnswerId":52783689.0,"OwnerUserId":10494287.0,"Title":"Variation in computation of gradient between Keras's backend and Tensorflow","Body":"<p>Note: <code>keras.backend()<\/code> returns tensorflow. Python 3.5 used.<\/p>\n\n<p>I have encountered a bug in the computation of gradient. I have replicated the bug in a simple Keras model and Tensorflow model shown below.<\/p>\n\n<pre><code>from keras.layers import Dense, Input\nfrom keras.models import Model\nfrom keras.optimizers import Adam\nimport keras\n\nimport keras.backend as K\nimport numpy as np\nimport tensorflow as tf\n\nclass KerasModel(object):\n    def __init__(self, seed, dim_size, optimizer, loss_func):\n\n\n        self.sess=tf.Session()\n\n        I = Input(\n                shape=[dim_size], \n                name='i'\n                 )\n\n        O = Dense(\n                1,\n                activation='relu',\n                kernel_initializer=keras.initializers.glorot_uniform(seed=seed),\n                name='o'\n                 )(I)\n\n\n        self.model = Model(inputs=I,outputs=O)\n        self.model.compile(loss=loss_func, optimizer=optimizer)\n\n\n        self.action_grads = tf.gradients(self.model.output, self.model.input)  \n        self.grad_func= K.function(self.model.inputs, self.action_grads)\n\n        self.sess.run(tf.global_variables_initializer())\n\n    def tf_grad(self, X):\n        return self.sess.run(self.action_grads, feed_dict={self.model.input: X,})[0]\n\n\n    def keras_grad(self, X):        \n        return self.grad_func(X)[0]\n\n\n\nclass TFModel(object):\n    def __init__(self, seed, dim_size, optimizer, loss_func):\n\n        self.graph= tf.Graph()\n\n        with self.graph.as_default():\n            glorot_uniform= tf.glorot_uniform_initializer(seed=seed)\n\n            O= {\n                    'weights': tf.Variable(glorot_uniform([dim_size, 1])),\n                    'bias': tf.Variable( tf.zeros(1) )\n                }\n\n\n            w_list= [ \n                O['weights'], O['bias']\n                    ] \n\n            w_list_placeholder= []\n            w_list_update= []\n            for i in range(0, len(w_list)):\n                w_list_placeholder.append( tf.placeholder(tf.float32) )\n                w_list_update.append( w_list[i].assign( w_list_placeholder[i] ) )\n\n\n            I= tf.placeholder(tf.float32, shape= (None, dim_size))\n\n            output= tf.nn.relu( tf.add( tf.matmul( I, O['weights']), O['bias'] ) )\n\n            gradient= tf.gradients(output, I)\n\n            y= tf.placeholder(tf.float32)\n\n            loss= tf.reduce_mean( loss_func(y, output) )\n\n            train = optimizer.minimize(loss)\n\n            self.tensors= {\n                'output': output, 'I': I, 'y': y, 'grad':gradient,\n                'loss': loss, 'train-op': train, 'w': w_list,\n                'w-placeholder': w_list_placeholder, 'w-update': w_list_update\n                          }\n\n            self.sess= tf.Session(graph=self.graph)\n\n            self.sess.run( tf.variables_initializer( self.graph.get_collection('variables') ) )\n\n\n    def train_on_batch(self, X, y):\n        _, l=self.sess.run( \n            [self.tensors['train-op'], self.tensors['loss']],\n            feed_dict={ \n                self.tensors['I']: X,\n                self.tensors['y']: y\n                       }\n                           )\n        return l\n\n    def predict(self, X):\n        return self.sess.run(self.tensors['output'], feed_dict={self.tensors['I']: X})\n\n    def get_weights(self):\n        return self.sess.run(self.tensors['w'])\n\n    def set_weights(self, new_weights):\n        self.sess.run( \n            self.tensors['w-update'], \n            feed_dict={ x:y for x,y in zip(self.tensors['w-placeholder'], new_weights) } \n                     )\n\n    def grad(self, X):\n        return self.sess.run(self.tensors['grad'], feed_dict={self.tensors['I']:X})[0]\n<\/code><\/pre>\n\n<p>The two models are identical in layers, initialization, optimizer, loss function, weights, etc.<\/p>\n\n<p>I use the two models to compute gradient from the same set of inputs. For the Tensorflow model, this is done with the <code>grad()<\/code> function:<\/p>\n\n<pre><code>def grad(self, X):\n    return self.sess.run(self.tensors['grad'], feed_dict={self.tensors['I']:X})[0]\n<\/code><\/pre>\n\n<p>For the Keras model, this is done with <code>keras_grad()<\/code> and <code>tf_grad()<\/code>.<\/p>\n\n<pre><code>self.action_grads = tf.gradients(self.model.output, self.model.input)\nself.grad_func= K.function(self.model.inputs, self.action_grads)\n\ndef tf_grad(self, X):\n    return self.sess.run(self.action_grads, feed_dict={self.model.input: X,})[0]\n\n\ndef keras_grad(self, X):        \n    return self.grad_func(X)[0]\n<\/code><\/pre>\n\n<p><code>keras_grad()<\/code> utilizes <code>keras.backend.function()<\/code> to do so, while <code>tf_grad()<\/code> utilizes <code>tf.Session()<\/code>.<\/p>\n\n<p>They are then trained with a different set of identical inputs. Again, this is shown below:<\/p>\n\n<pre><code>seed=1\ndim_size=3\nlearning_rate=0.01\ne=1e-8\n\n\nk_adam= KerasModel(\n    seed, dim_size, tf.train.AdamOptimizer(learning_rate=learning_rate,epsilon= e),\n    keras.losses.mean_squared_error\n                  )\n\ntf_model= TFModel(\n    seed, dim_size, tf.train.AdamOptimizer(learning_rate=learning_rate,epsilon= e),\n    keras.losses.mean_squared_error\n                  )\n\n\nX=np.array([[0.25175066, 0.53507285, 0.3210762 ]])\n#X= np.random.random([1,dim_size])\ny= np.random.random([1])\n\nprint(k_adam.keras_grad([X]))\nprint(k_adam.tf_grad(X))\nprint(tf_model.grad(X))\n\nprint()\nX= np.array([[0.47194079, 0.85071664, 0.25451934]])\n#X= np.random.random([1,dim_size])\ny= np.random.random([1])\nk_adam.model.train_on_batch(X,y)\ntf_model.train_on_batch(X,y)\n\nprint(k_adam.keras_grad([X]))\nprint(k_adam.tf_grad(X))\nprint(tf_model.grad(X))\n\nprint()\nX= np.random.random([1,dim_size])\ny= np.random.random([1])\nk_adam.model.train_on_batch(X,y)\ntf_model.train_on_batch(X,y)\n\nprint(k_adam.keras_grad([X]))\nprint(k_adam.tf_grad(X))\nprint(tf_model.grad(X))\n<\/code><\/pre>\n\n<p>Running the code should give you the following output:<\/p>\n\n<pre><code>[[-0.63922524  1.0297645  -1.1010152 ]]\n[[-0.63922524  1.0297645  -1.1010152 ]]\n[[-0.63922524  1.0297645  -1.1010152 ]]\n\n[[-0.62922525  1.0397645  -1.0910152 ]]\n[[-0.63922524  1.0297645  -1.1010152 ]]\n[[-0.62922525  1.0397645  -1.0910152 ]]\n\n[[ ... ]]\n....\n<\/code><\/pre>\n\n<p>And some other results. <\/p>\n\n<p>For the first block of output, I expected the array of gradients computed to be the same among the two models, since they are the same model. And this is true.<\/p>\n\n<p>For the second block of output, I have trained the two models with a different set of identical inputs. So the array of gradients should differ from the first block of output, but they should be the same among themselves (the second block of output). This is true to a degree. The array of gradients from the Tensorflow model and <code>keras_grad()<\/code> is the same between themselves but different from the first block.<\/p>\n\n<p><strong>However, the output of <code>tf_grad()<\/code> did not change from the first block of output.<\/strong> <\/p>\n\n<p>From my own testing, I found that the output from <code>tf_grad()<\/code> varies between either 0 or the same set of numbers after initialization. This behavior repeats even when training with different batches.<\/p>\n\n<p>As mentioned already, the only difference in <code>tf_grad()<\/code> and <code>keras_grad()<\/code>, which are both from the Keras model, is that one (<code>tf_grad<\/code>) is ran with <code>tf.session()<\/code> while the other (<code>keras_grad<\/code>) is ran with <code>keras.backend.function()<\/code>. <\/p>\n\n<p>Why is it that one is able to update accordingly and the other cannot do so?<\/p>\n","answers":[{"AnswerId":"52783689","CreationDate":"2018-10-12T16:37:20.510","ParentId":null,"OwnerUserId":"1372667","Title":null,"Body":"<p>You need to set the session to the keras TF backend<\/p>\n\n<pre><code>K.set_session(self.sess)\n<\/code><\/pre>\n\n<p>Like this:<\/p>\n\n<pre><code>class KerasModel(object):\ndef __init__(self, seed, dim_size, optimizer, loss_func):\n    self.sess=tf.Session()\n\n    K.set_session(self.sess)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52781514,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><keras><deep-learning>","CreationDate":"2018-10-12T14:18:50.597","AcceptedAnswerId":null,"OwnerUserId":4722605.0,"Title":"UnboundLocalError: local variable 'outs' referenced before assignment - durnig training the model","Body":"<p>I'm keras beginner<\/p>\n\n<pre><code>Epoch 9\/100\n99\/100 [============================&gt;.] - ETA: 0s - loss: 0.4229 - acc: 0.8649Epoch 00008: val_loss did not improve\n100\/100 [==============================] - 72s - loss: 0.4222 - acc: 0.8650 - val_loss: 1.2091 - val_acc: 0.5000\nTraceback (most recent call last):\nFile \"Train.py\", line 382, in &lt;module&gt;\ntrain(args)\nFile \"Train.py\", line 351, in train\nscores = model.evaluate_generator(validation_generator, workers=5, steps = None)\nFile \"\/home\/Grace\/kerasenv\/local\/lib\/python2.7\/site-packages\/keras\/legacy\/interfaces.py\", line 87, in wrapper\nreturn func(*args, **kwargs)\nFile \"\/home\/Grace\/kerasenv\/local\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1991, in evaluate_generator\nif not isinstance(outs, list):\nUnboundLocalError: local variable 'outs' referenced before assignment\n<\/code><\/pre>\n\n<p>my code is here.<\/p>\n\n<pre><code>    scores = model.evaluate_generator(validation_generator, workers=5, steps = None)\n\n    history_ft = model.fit_generator(\n    train_generator,\n    epochs=nb_epoch,\n    steps_per_epoch=steps_per_epoch,\n    validation_steps = 3,\n    validation_data=validation_generator,\n    callbacks = [tensorboard, checkpointer, EarlyStopping(monitor='val_loss', min_delta=0.01, patience=5)],\n    class_weight = 'auto',\n    workers = 5,\n    use_multiprocessing = True)\n<\/code><\/pre>\n\n<h1>parameter<\/h1>\n\n<ul>\n<li>steps_per_epoch = 100 <\/li>\n<li>nb_epoch =100<\/li>\n<li><p>Batch Size: 8<\/p>\n\n<ul>\n<li>Found 1134 images belonging to 4 classes.(Train set)<\/li>\n<li>Found 174 images belonging to 4 classes.(validation set)<\/li>\n<\/ul><\/li>\n<li><p>my version of keras 2.0.6, <\/p><\/li>\n<li>tensorflow: 1.3.0<\/li>\n<\/ul>\n\n<p>Somebody knows why? Thanks<\/p>\n","answers":[]}
+{"QuestionId":52781685,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-12T14:28:12.590","AcceptedAnswerId":null,"OwnerUserId":9749306.0,"Title":"Custom Keras generator much slower compared to Keras' bult in generator","Body":"<p>I have a multi label classification problem. I wrote this custom generator. It reads images and output labels from the disk, and returns them in batches of size 32. <\/p>\n\n<pre><code>def get_input(img_name):\n    path = os.path.join(\"images\", img_name)\n    img = image.load_img(path, target_size=(224, 224))\n\n    return img\n\n\ndef get_output(img_name, file_path):\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n\n    img_id = img_name.split(\".\")[0]\n    img_id = img_id.lstrip(\"0\")\n    img_id = int(img_id)\n\n    labels = data.loc[img_id - 1].values\n    labels = labels[1:]\n\n    labels = list(labels)\n    label_arrays = []\n    for i in range(20):\n        val = np.zeros((1))\n        val[0] = labels[i]\n        label_arrays.append(val)\n\n    return label_arrays\n\n\ndef preprocess_input(img_name):\n    img = get_input(img_name)\n    x = image.img_to_array(img)\n    x = np.expand_dims(x, axis=0)    \n    return x\n\ndef train_generator(batch_size):\n    file_path = \"train.txt\"\n    data = pd.read_csv(file_path, delim_whitespace=True, header=None)\n\n    while True:\n        for i in range(math.floor(8000\/batch_size)):\n            x_batch = np.zeros(shape=(32, 224, 224, 3))\n            y_batch = np.zeros(shape=(32, 20))\n            for j in range(batch_size):\n                img_name = data.loc[i * batch_size + j].values\n                img_name = img_name[0]\n                x = preprocess_input(img_name)\n                y = get_output(img_name, file_path)\n                x_batch[j, :, :, :] = x\n                y_batch[j] = y\n\n            ys = []\n            for i in range(20):\n              ys.append(y_batch[:,i])\n\n            yield(x_batch, ys)\n<\/code><\/pre>\n\n<p>Had a little problem with labels returned to the model, and got it solved in this question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/52691538\/training-a-multi-output-keras-model?noredirect=1#comment92379651_52691538\">training a multi-output keras model\n<\/a><\/p>\n\n<p>I tested this generator on a single output problem. This custom generator is very slow. The ETA for a single epoch by using this custom generator is around 27 hours, while the builtin generator(using flow_from_directory) takes 25 minutes for a single epoch. What am I doing wrong? <\/p>\n\n<p>The training process for both tests is identical, except for the generator used. Validation generator is similar to training generator. I know I will not reach the efficiency of Keras' built in generator, but this difference in speed is too much.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Some guides I read for creating custom generators.<\/p>\n\n<p><a href=\"https:\/\/medium.com\/@ensembledme\/writing-custom-keras-generators-fe815d992c5a\" rel=\"nofollow noreferrer\">Writing Custom Keras Generators<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8078\" rel=\"nofollow noreferrer\">custom generator for fit_generator() that yields multiple inputs with different shapes<\/a><\/p>\n","answers":[{"AnswerId":"52782788","CreationDate":"2018-10-12T15:34:58.283","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Maybe the built in generator processes the data on your gpu while your custom generator runs on the cpu, making is significantly slower.<\/p>\n\n<p>Another guess is because Keras is using <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">Dataset<\/a> in the background. Your implementation probably uses <code>feed-dict<\/code> which is the slowest possible way to pass information to TensorFlow. The best way to feed data into the  models is to use an input pipeline to ensure that the GPU never has to wait for new stuff to come in.<\/p>\n"}]}
+{"QuestionId":52782114,"AnswerCount":0,"Tags":"<python><numpy><matrix><optimization><pytorch>","CreationDate":"2018-10-12T14:53:56.800","AcceptedAnswerId":null,"OwnerUserId":3719101.0,"Title":"Map function along arbitrary dimension with pytorch? numpy?","Body":"<p>I wonder why I find no utility to map custom <a href=\"https:\/\/pytorch.org\/docs\/stable\/index.html\" rel=\"nofollow noreferrer\">pytorch<\/a> or <a href=\"https:\/\/docs.scipy.org\/doc\/\" rel=\"nofollow noreferrer\">numpy<\/a> transformations along any dimensions of complicated tensors\/arrays\/matrices.<br>\nI think I remember that such a thing was available in R. With this fantasy <code>tch.map<\/code> utility you could do:<\/p>\n\n<pre><code>&gt;&gt;&gt; import torch as tch # or numpy\n\n&gt;&gt;&gt; # one torch tensor\n&gt;&gt;&gt; a = tch.tensor([0, 1, 2, 3, 4])\n&gt;&gt;&gt; # one torch function (dummy) returning 2 values\n&gt;&gt;&gt; f = lambda x: tch.tensor((x + 1, x * 2))\n&gt;&gt;&gt; # map f along dimension 0 of a, expecting 2 outputs\n&gt;&gt;&gt; res = tch.map(f, a, 0, 2) # fantasy, optimized on CPU\/GPU..\n&gt;&gt;&gt; res\ntensor([[1, 0],\n        [2, 2],\n        [3, 4],\n        [4, 6],\n        [5, 8]])\n&gt;&gt;&gt; res.shape\ntorch.Size([5, 2])\n\n&gt;&gt;&gt; # another tensor\n&gt;&gt;&gt; a = tch.tensor(list(range(24))).reshape(2, 3, 4).type(tch.double)\n&gt;&gt;&gt; # another function (dummy) returning 2 values\n&gt;&gt;&gt; f = lambda x: tch.tensor((tch.mean(x), tch.std(x)))\n&gt;&gt;&gt; # map f along dimension 2 of a, expecting 2 outputs\n&gt;&gt;&gt; res = tch.map(f, a, 2, 2) # fantasy, optimized on CPU\/GPU..\ntensor([[[ 1.5000,  1.2910],\n         [ 5.5000,  1.2910],\n         [ 9.5000,  1.2910]],\n\n        [[13.5000,  1.2910],\n         [17.5000,  1.2910],\n         [21.5000,  1.2910]]])\n&gt;&gt;&gt; res.shape\ntorch.Size([2, 3, 2])\n\n&gt;&gt;&gt; # yet another tensor\n&gt;&gt;&gt; a = tch.tensor(list(range(12))).reshape(3, 4)\n&gt;&gt;&gt; # another function (dummy) returning 2x2 values\n&gt;&gt;&gt; f = lambda x: x + tch.rand(2, 2)\n&gt;&gt;&gt; # map f along all values of a, expecting 2x2 outputs\n&gt;&gt;&gt; res = tch.map(f, a, -1, (2, 2)) # fantasy, optimized on CPU\/GPU..\n&gt;&gt;&gt; print(res)\ntensor([[[[ 0.4827,  0.3043],\n          [ 0.8619,  0.0505]],\n\n         [[ 1.4670,  1.5715],\n          [ 1.1270,  1.7752]],\n\n         [[ 2.9364,  2.0268],\n          [ 2.2420,  2.1239]],\n\n         [[ 3.9343,  3.6059],\n          [ 3.3736,  3.5178]]],\n\n\n        [[[ 4.2063,  4.9981],\n          [ 4.3817,  4.4109]],\n\n         [[ 5.3864,  5.3826],\n          [ 5.3614,  5.1666]],\n\n         [[ 6.6926,  6.2469],\n          [ 6.7888,  6.6803]],\n\n         [[ 7.2493,  7.5727],\n          [ 7.6129,  7.1039]]],\n\n\n        [[[ 8.3171,  8.9037],\n          [ 8.0520,  8.9587]],\n\n         [[ 9.5006,  9.1297],\n          [ 9.2620,  9.8371]],\n\n         [[10.4955, 10.5853],\n          [10.9939, 10.0271]],\n\n         [[11.3905, 11.9326],\n          [11.9376, 11.6408]]]])\n&gt;&gt;&gt; res.shape\ntorch.Size([3, 4, 2, 2])\n<\/code><\/pre>\n\n<p>Instead, I keep finding myself messing around with complicated <code>tch.stack<\/code>, <code>tch.squeeze<\/code>, <code>tch.reshape<\/code>, <code>tch.permute<\/code>, <em>etc.<\/em>, counting dimensions on my fingers not to get lost.<\/p>\n\n<p>Does such a utility exist and I have missed it for some reason?<br>\nIs such a utility impossible to implement for some reason?  <\/p>\n","answers":[]}
+{"QuestionId":52782864,"AnswerCount":1,"Tags":"<python><docker><flask><pip><pytorch>","CreationDate":"2018-10-12T15:40:17.597","AcceptedAnswerId":52845257.0,"OwnerUserId":2175052.0,"Title":"MemoryError when attempting to create a docker image with Torch\/PyTorch","Body":"<p>I am following <a href=\"http:\/\/containertutorials.com\/docker-compose\/flask-simple-app.html\" rel=\"nofollow noreferrer\">this tutorial<\/a> to create a Docker image for a flask application. The application depends on Torch\/PyTorch. As such, my <code>requirements.txt<\/code> file looks like the following.<\/p>\n\n<pre>\nflask\nflask-cors\npytorch\ntorchvision\npandas\n<\/pre>\n\n<p>My Dockerfile then looks like the following.<\/p>\n\n<pre>\nFROM ubuntu:latest\nLABEL My Company \"info@mycompany.com\"\nRUN apt-get update -y\nRUN apt-get install -y python-pip python-dev build-essential\nCOPY . \/app\nWORKDIR \/app\nRUN pip install -vvv --no-cache-dir -r requirements.txt\nENTRYPOINT [\"python3\"]\nCMD [\"app.py\"]\n<\/pre>\n\n<p>When I type in the command <code>docker build -t flask-sample-one:latest .<\/code> I get the following error message (which happens when pip is trying to install torch) on my Mac.<\/p>\n\n<pre>\nException:\nTraceback (most recent call last):\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/basecommand.py\", line 215, in main\n    status = self.run(options, args)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/commands\/install.py\", line 353, in run\n    wb.build(autobuilding=True)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/wheel.py\", line 749, in build\n    self.requirement_set.prepare_files(self.finder)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/req\/req_set.py\", line 380, in prepare_files\n    ignore_dependencies=self.ignore_dependencies))\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/req\/req_set.py\", line 620, in _prepare_file\n    session=self.session, hashes=hashes)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 821, in unpack_url\n    hashes=hashes\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 659, in unpack_http_url\n    hashes)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 882, in _download_http_url\n    _download_url(resp, link, content_file, hashes)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 603, in _download_url\n    hashes.check_against_chunks(downloaded_chunks)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/utils\/hashes.py\", line 46, in check_against_chunks\n    for chunk in chunks:\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 571, in written_chunks\n    for chunk in chunks:\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/utils\/ui.py\", line 139, in iter\n    for x in it:\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/download.py\", line 560, in resp_read\n    decode_content=False):\n  File \"\/usr\/share\/python-wheels\/urllib3-1.22-py2.py3-none-any.whl\/urllib3\/response.py\", line 436, in stream\n    data = self.read(amt=amt, decode_content=decode_content)\n  File \"\/usr\/share\/python-wheels\/urllib3-1.22-py2.py3-none-any.whl\/urllib3\/response.py\", line 384, in read\n    data = self._fp.read(amt)\n  File \"\/usr\/share\/python-wheels\/CacheControl-0.11.7-py2.py3-none-any.whl\/cachecontrol\/filewrapper.py\", line 63, in read\n    self._close()\n  File \"\/usr\/share\/python-wheels\/CacheControl-0.11.7-py2.py3-none-any.whl\/cachecontrol\/filewrapper.py\", line 50, in _close\n    self.__callback(self.__buf.getvalue())\n  File \"\/usr\/share\/python-wheels\/CacheControl-0.11.7-py2.py3-none-any.whl\/cachecontrol\/controller.py\", line 275, in cache_response\n    self.serializer.dumps(request, response, body=body),\n  File \"\/usr\/share\/python-wheels\/CacheControl-0.11.7-py2.py3-none-any.whl\/cachecontrol\/serialize.py\", line 55, in dumps\n    \"body\": _b64_encode_bytes(body),\n  File \"\/usr\/share\/python-wheels\/CacheControl-0.11.7-py2.py3-none-any.whl\/cachecontrol\/serialize.py\", line 12, in _b64_encode_bytes\n    return base64.b64encode(b).decode(\"ascii\")\nMemoryError\nThe command '\/bin\/sh -c pip install -vvv -r requirements.txt' returned a non-zero code: 2\n<\/pre>\n\n<p>I am not sure what the <code>MemoryError<\/code> is coming from. I tried to build the docker image on a Macbook Pro with 16 GB of RAM. Any ideas on what's going on or how to create a docker image with Python? <\/p>\n\n<p>On an AWS EC2 instance, I get the following error.<\/p>\n\n<pre>\nRunning setup.py install for pytorch: started\n    Running command \/usr\/bin\/python -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-build-SeBh33\/pytorch\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-5jxyuS-record\/install-record.txt --single-version-externally-managed --compile\n    Traceback (most recent call last):\n      File \"\", line 1, in \n      File \"\/tmp\/pip-build-SeBh33\/pytorch\/setup.py\", line 13, in \n        raise Exception(message)\n    Exception: You should install pytorch from http:\/\/pytorch.org\n    Running setup.py install for pytorch: finished with status 'error'\nCleaning up...\n  Removing source in \/tmp\/pip-build-SeBh33\/pytorch\nCommand \"\/usr\/bin\/python -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-build-SeBh33\/pytorch\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-5jxyuS-record\/install-record.txt --single-version-externally-managed --compile\" failed with error code 1 in \/tmp\/pip-build-SeBh33\/pytorch\/\nException information:\nTraceback (most recent call last):\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/basecommand.py\", line 215, in main\n    status = self.run(options, args)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/commands\/install.py\", line 360, in run\n    prefix=options.prefix_path,\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/req\/req_set.py\", line 784, in install\n    **kwargs\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/req\/req_install.py\", line 878, in install\n    spinner=spinner,\n  File \"\/usr\/lib\/python2.7\/dist-packages\/pip\/utils\/__init__.py\", line 725, in call_subprocess\n    % (command_desc, proc.returncode, cwd))\nInstallationError: Command \"\/usr\/bin\/python -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-build-SeBh33\/pytorch\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-5jxyuS-record\/install-record.txt --single-version-externally-managed --compile\" failed with error code 1 in \/tmp\/pip-build-SeBh33\/pytorch\/\nThe command '\/bin\/sh -c pip install -vvv --no-cache-dir -r requirements.txt' returned a non-zero code: 1\n<\/pre>\n\n<p>Any ideas on what's going on? I am simply trying to build a docker image with those few dependencies so I can serve a PyTorch model from a flask application. I'm not sure if there's a tutorial out there on doing this (couldn't find one yet).<\/p>\n","answers":[{"AnswerId":"52845257","CreationDate":"2018-10-16T23:23:34.253","ParentId":null,"OwnerUserId":"2175052","Title":null,"Body":"<p>If anyone is interested, I was able to get <code>PyTorch<\/code> installed into a docker container as follows.<\/p>\n\n<p>I modified the <code>requirements.txt<\/code> to look like the following.<\/p>\n\n<pre><code>flask\nflask-cors\npandas\n<\/code><\/pre>\n\n<p>I then modified <code>Dockerfile<\/code> to look like the following. It's quirky and does not follow the conventional way of install Python packages with pip, but it works. Note that I have to install Python 3, followed by PyTorch wheel, and then finally with a pip install of <code>torchvision<\/code>. <\/p>\n\n<pre><code>FROM ubuntu:latest\nLABEL My Company \"info@mycompany.com\"\nRUN apt-get update -y\nRUN apt-get install -y python-pip python-dev build-essential\nRUN apt-get install -y python3.6\nRUN apt-get install -y python3-pip\nCOPY . \/app\nWORKDIR \/app\nRUN pip3 install -vvv --no-cache-dir -r requirements.txt\nRUN pip3 install http:\/\/download.pytorch.org\/whl\/cpu\/torch-0.4.1-cp36-cp36m-linux_x86_64.whl\nRUN pip3 install torchvision\nENTRYPOINT [\"python3\"]\nCMD [\"app.py\"]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52783615,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><computer-vision>","CreationDate":"2018-10-12T16:30:25.423","AcceptedAnswerId":52787752.0,"OwnerUserId":4107990.0,"Title":"Tensorflow Image Classifier Accuracy Fails to Change","Body":"<p>I am new to tensorflow. I'm creating a simple fully connected neural network for image classification. The image is (-1, 224, 224, 3), and label is (-1, 2). However, the result of my code is that the accuracy does not improve at all; it stays at 47% and does not change - even if changed learning rate, optimizer, and different test set.. Any help will be greatly appreciated! Thank you!<\/p>\n\n<pre><code>import matplotlib.pyplot as plt \nfrom util.MacOSFile import MacOSFile\nimport numpy as np\nimport _pickle as pickle\nimport tensorflow as tf\n\ndef pickle_load(file_path):\n    with open(file_path, \"rb\") as f:\n        return pickle.load(MacOSFile(f))\n\n###hyperparameters###\nbatch_size = 32\niterations = 10\n\n###loading training data start###\ndata = pickle_load('training.pickle')\nx_train = []\ny_train = []\n\nfor features, labels in data:\n    x_train.append(features)\n    y_train.append(labels)\n\nx_train = np.array(x_train)\ny_train = np.array(y_train)\n\n###################################\n\n###loading test data start###\ndata = pickle_load('testing.pickle')\nx_test = []\ny_test = []\n\nfor features, labels in data:\n    x_test.append(features)\n    y_test.append(labels)\n\nx_test = np.array(x_test)\ny_test = np.array(y_test)\n\n###################################\n\n\n###neural network###\n\nx_s = tf.placeholder(tf.float32, [None, 224, 224, 3])\ny_s = tf.placeholder(tf.float32, [None, 2])\nx_image = tf.reshape(x_s, [-1, 150528])\n\nW_1 = tf.Variable(tf.truncated_normal([150528, 8224]))\nb_1 = tf.Variable(tf.zeros([8224]))\nh_fc1 = tf.nn.relu(tf.matmul(x_image, W_1) + b_1)\n\nW_2 = tf.Variable(tf.truncated_normal([8224, 1028]))\nb_2 = tf.Variable(tf.zeros([1028]))\nh_fc2 = tf.nn.relu(tf.matmul(h_fc1, W_2) + b_2)\n\nW_3 = tf.Variable(tf.truncated_normal([1028, 2]))\nb_3 = tf.Variable(tf.zeros([2]))\nprediction = tf.nn.softmax(tf.matmul(h_fc2, W_3) + b_3)\n\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_s, logits=prediction)\ntrain_step = tf.train.GradientDescentOptimizer(0.1).minimize(cross_entropy)\ninit = tf.global_variables_initializer()\n\n###neural network end###\n\n\nwith tf.Session() as sess:\n    sess.run(init)\n\n    train_sample_size = len(data) #how many data points?\n    max_batches_in_data = int(train_sample_size\/batch_size) #max number of batches possible; 623 \n\n    for iteration in range(iterations):\n            print('Iteration ', iteration)\n            epoch = int(iteration\/max_batches_in_data)\n            start_idx = (iteration-epoch*max_batches_in_data)*batch_size\n            end_idx = (iteration+1 - epoch*max_batches_in_data)*batch_size\n            mini_x_train = x_train[start_idx: end_idx] \n            mini_y_train = y_train[start_idx: end_idx]\n\n            ##actual training is here\n            sess.run(train_step, feed_dict={x_s: mini_x_train, y_s: mini_y_train})\n\n            #test accuracy#\n            y_pre = sess.run(prediction, feed_dict={x_s: x_train[:100]})\n            correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(y_train[:100], 1))\n            accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n            result = sess.run(accuracy, feed_dict={x_s: x_train[:100], y_s: y_train[:100]})\n            print(\"Result: {0}\".format(result))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52787752","CreationDate":"2018-10-12T22:28:22.297","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>I've made a few observations, firstly your code is a bit outdated, you don't have to manually set up fully connected layers, there is something for that:<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/dense\" rel=\"nofollow noreferrer\">dense layers<\/a>.\nIf you load in images, why don't you use convolutional layers as well?\nI also recommend the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">adam optimizer<\/a> just leave the parameters to their default value. I hope I helped a bit :)<\/p>\n"}]}
+{"QuestionId":52783747,"AnswerCount":1,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2018-10-12T16:41:13.597","AcceptedAnswerId":53125376.0,"OwnerUserId":4513311.0,"Title":"How to do batching with TensorFlow Lite?","Body":"<p>I have a custom CNN model, and I have converted it to .tflite format and deployed it on my Android app. However, I can't figure out how to do batching while inference with tensorflow lite. <\/p>\n\n<p>From this Google <a href=\"https:\/\/firebase.google.com\/docs\/ml-kit\/android\/use-custom-models\" rel=\"nofollow noreferrer\">doc<\/a>, it seems you have to set the input format of your model. However, this doc is using a code example with Firebase API, which I'm not planning on using. <\/p>\n\n<p>To be more specific:<\/p>\n\n<p>I want to inference multiple 100x100x3 images at once, so the input size is <strong>N<\/strong>x100x100x3.<\/p>\n\n<p>Question: <\/p>\n\n<p>How to do this with TF lite? <\/p>\n","answers":[{"AnswerId":"53125376","CreationDate":"2018-11-02T20:23:51.797","ParentId":null,"OwnerUserId":"606314","Title":null,"Body":"<p>You can just call the <code>resizeInput<\/code> API (Java) or <code>ResizeInputTensor<\/code> API (if you're using C++).<\/p>\n\n<p>For example, in Java:<\/p>\n\n<pre><code>interpreter.resizeInput(tensor_index, [num_batch, 100, 100, 3]);\n<\/code><\/pre>\n\n<p>Let us know if you have any problem batching in TensorFlow lite. <\/p>\n"}]}
+{"QuestionId":52783781,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-12T16:43:22.893","AcceptedAnswerId":null,"OwnerUserId":2641587.0,"Title":"Get Keras model input from inside a custom callback","Body":"<p>I have a very simple question. I have a Keras model (TF backend) defined for classification. I want to dump the training images fed into my model during training for debugging purposes. I am trying to create a custom callback that writes Tensorboard image summaries for this.<\/p>\n\n<p>But how can I obtain the real training data inside the callback? <\/p>\n\n<p>Currently I am trying this:<\/p>\n\n<pre><code>class TensorboardKeras(Callback):                                                                                                                                                                                                                                     \n    def __init__(self, model, log_dir, write_graph=True):                                                                                                                                                                                                             \n        self.model = model                                                                                                                                                                                                                                            \n        self.log_dir = log_dir                                                                                                                                                                                                                                        \n        self.session = K.get_session()                                                                                                                                                                                                                                \n\n        tf.summary.image('input_image', self.model.input)                                                                                                                                                                                                             \n        self.merged = tf.summary.merge_all()                                                                                                                                                                                                                          \n\n        if write_graph:                                                                                                                                                                                                                                               \n            self.writer = tf.summary.FileWriter(self.log_dir, K.get_session().graph)                                                                                                                                                                                  \n        else:                                                                                                                                                                                                                                                         \n            self.writer = tf.summary.FileWriter(self.log_dir)\n\n    def on_batch_end(self, batch, logs=None):\n        summary = self.session.run(self.merged, feed_dict={})                                                                                                                                                                                                         \n        self.writer.add_summary(summary, batch)                                                                                                                                                                                                                       \n        self.writer.flush()\n<\/code><\/pre>\n\n<p>But I am getting the error:\nInvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_1' with dtype float and shape [?,224,224,3]<\/p>\n\n<p>There must be a way to see what models, get as an input, right?<\/p>\n\n<p>Or maybe I should try another way to debug it?<\/p>\n","answers":[{"AnswerId":"52888216","CreationDate":"2018-10-19T08:05:14.433","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>You don't need callbacks for this. All you need to do is implementing a function that yields an image and its label as a tuple. <code>flow_from_directory<\/code> function has a parameter called <code>save_to_dir<\/code> which could satisfy all of your needs, in case it doesn't, here is what you can do:<\/p>\n\n<pre><code>def trainGenerator(batch_size,train_path, image_size)\n    #preprocessing see https:\/\/keras.io\/preprocessing\/image\/ for details\n    image_datagen = ImageDataGenerator(horizontal_flip=True)\n    #create image generator see https:\/\/keras.io\/preprocessing\/image\/#flow_from_directory for details\n    train_generator = image_datagen.flow_from_directory(\n        train_path,\n        class_mode = \"categorical\",\n        target_size = image_size,\n        batch_size = batch_size,\n        save_prefix  = \"augmented_train\",\n        seed = seed)\n\n    for (batch_imgs, batch_labels) in train_generator: \n        #do other stuff such as dumping images or further augmenting images\n    yield (batch_imgs,batch_labels)\n\n\nt_generator = trainGenerator(32, \".\/train_data\", (224,224,3))\nmodel.fit_generator(t_generator,steps_per_epoch=10,epochs=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52784204,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-12T17:13:19.650","AcceptedAnswerId":52784359.0,"OwnerUserId":3392589.0,"Title":"how to create iterator.get_next() for validation set","Body":"<p>I am working on project to classify medical images using the CNN model, for my project I use tensorflow, after doing some search, at last, I could use new tensorflow input pipeline to prepare the train, validation and test set, here is the code:<\/p>\n\n<pre><code>train_data = tf.data.Dataset.from_tensor_slices(train_images)\ntrain_labels = tf.data.Dataset.from_tensor_slices(train_labels)\ntrain_set = tf.data.Dataset.zip((train_data,train_labels)).shuffle(500).batch(30)\n\nvalid_data = tf.data.Dataset.from_tensor_slices(valid_images)\nvalid_labels = tf.data.Dataset.from_tensor_slices(valid_labels)\nvalid_set = tf.data.Dataset.zip((valid_data,valid_labels)).shuffle(200).batch(20)\n\ntest_data = tf.data.Dataset.from_tensor_slices(test_images)\ntest_labels = tf.data.Dataset.from_tensor_slices(test_labels)\ntest_set = tf.data.Dataset.zip((test_data, test_labels)).shuffle(200).batch(20)\n\n# create general iterator\niterator = tf.data.Iterator.from_structure(train_set.output_types, train_set.output_shapes)\nnext_element = iterator.get_next()\ntrain_init_op = iterator.make_initializer(train_set)\nvalid_init_op = iterator.make_initializer(valid_set)\ntest_init_op  = iterator.make_initializer(test_set)\n<\/code><\/pre>\n\n<p>I can use <code>next_element<\/code> to iterate over the train set (<code>next_element[0]<\/code> for images and <code>next_element[1]<\/code> for labels), now I want to do is doing same thing for the validation set(creating an iterator for validation set), anyone can give me an idea how to do it?<\/p>\n","answers":[{"AnswerId":"52784359","CreationDate":"2018-10-12T17:24:51.797","ParentId":null,"OwnerUserId":"3954958","Title":null,"Body":"<p>You should be able to use the same <code>next_element<\/code> to get validation and test set. <\/p>\n\n<p>For example, initialize the dataset by <code>sess.run(valid_init_op)<\/code> and then <code>next_element<\/code> generates data in the validation set. <\/p>\n\n<pre><code>with tf.Session as sess:\n  sess.run(train_init_op)\n  image_train, label_train = next_element\n\n  sess.run(valid_init_op)\n  image_val, label_val = next_element\n\n  sess.run(test_init_op)\n  image_test, label_test = next_element\n<\/code><\/pre>\n"}]}
+{"QuestionId":52784333,"AnswerCount":1,"Tags":"<python><tensorflow><cuda>","CreationDate":"2018-10-12T17:22:48.047","AcceptedAnswerId":null,"OwnerUserId":2022518.0,"Title":"CUDA library actually used by TensorFlow","Body":"<p>I have installed (I think) TF with CUDA support using command <code>pip3.6 install tensorflow-gpu<\/code>, per TF installation page.<\/p>\n\n<p>My local CUDA installation is CUDA 9.0 and CUDNN 7.3.1 in <code>\/usr\/local\/cuda-9.0<\/code>.<\/p>\n\n<p>Per tip <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10827\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10827<\/a> I am checking libraries used by TF (in virtualenv):<\/p>\n\n<pre><code>% python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib() + \"\/python\/_pywrap_tensorflow_internal.so\")' | xargs ldd\nlinux-vdso.so.1 (0x00007fff57eb8000)\nlibtensorflow_framework.so =&gt; \/home\/mark\/projects\/bench\/venvs\/ve_tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/..\/libtensorflow_framework.so (0x00007ff29fa25000)\nlibcublas.so.9.0 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcublas.so.9.0 (0x00007ff29bda8000)\nlibcusolver.so.9.0 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcusolver.so.9.0 (0x00007ff2971ad000)\nlibcudart.so.9.0 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcudart.so.9.0 (0x00007ff296f40000)\nlibdl.so.2 =&gt; \/lib\/x86_64-linux-gnu\/libdl.so.2 (0x00007ff296d3c000)\nlibpthread.so.0 =&gt; \/lib\/x86_64-linux-gnu\/libpthread.so.0 (0x00007ff296b1f000)\nlibgomp.so.1 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libgomp.so.1 (0x00007ff2968f2000)\nlibm.so.6 =&gt; \/lib\/x86_64-linux-gnu\/libm.so.6 (0x00007ff2965ee000)\nlibrt.so.1 =&gt; \/lib\/x86_64-linux-gnu\/librt.so.1 (0x00007ff2963e6000)\nlibstdc++.so.6 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libstdc++.so.6 (0x00007ff296064000)\nlibgcc_s.so.1 =&gt; \/lib\/x86_64-linux-gnu\/libgcc_s.so.1 (0x00007ff295e4d000)\nlibc.so.6 =&gt; \/lib\/x86_64-linux-gnu\/libc.so.6 (0x00007ff295aae000)\n\/lib64\/ld-linux-x86-64.so.2 (0x00007ff2cb7c7000)\nlibcuda.so.1 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libcuda.so.1 (0x00007ff294f0e000)\nlibcudnn.so.7 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcudnn.so.7 (0x00007ff282bd5000)\nlibcufft.so.9.0 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcufft.so.9.0 (0x00007ff27ab34000)\nlibcurand.so.9.0 =&gt; \/usr\/local\/cuda-9.0\/lib64\/libcurand.so.9.0 (0x00007ff276bd0000)\nlibnvidia-fatbinaryloader.so.390.77 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libnvidia-fatbinaryloader.so.390.77 (0x00007ff276984000)\n<\/code><\/pre>\n\n<p><code>libcudnn.so<\/code> seem to point to the right library, however, for <code>libcuda.so<\/code> I have doubts:<\/p>\n\n<pre><code>libcuda.so.1 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libcuda.so.1 (0x00007ff294f0e000)\n\n% readlink -f \/usr\/lib\/x86_64-linux-gnu\/libcuda.so.1 \n\/usr\/lib\/x86_64-linux-gnu\/libcuda.so.390.77\n<\/code><\/pre>\n\n<p>OK, so it apparently leads to CUDA library used\/provided by NVIDIA device driver... <\/p>\n\n<p>Is this normal? Should it not use <code>libcuda.so<\/code> in <code>\/usr\/local\/cuda-9.0<\/code>?<\/p>\n\n<p>I do have one: <code>\/usr\/local\/cuda-9.0\/lib64\/stubs\/libcuda.so<\/code>.<\/p>\n","answers":[{"AnswerId":"52784819","CreationDate":"2018-10-12T17:59:20.210","ParentId":null,"OwnerUserId":"1695960","Title":null,"Body":"<p>Yes, it's normal.  <\/p>\n\n<p>The <code>libcuda<\/code> used should definitely be the one provided (installed) by the GPU driver.  It definitely should not be the one in the <code>stubs<\/code> directory.  <\/p>\n\n<p>The one in the <code>stubs<\/code> directory (or anything in the <code>\/usr\/local\/cuda...<\/code> path) is there for a different purpose, basically having to do with application building in certain scenarios, not for running any applications.<\/p>\n\n<p>For running applications (like Tensorflow), it's necessary to use the shared object provided by the driver, for the <code>libcuda<\/code> library.<\/p>\n\n<p>(The <code>libcuda.so<\/code> in the <code>stubs<\/code> directory is provided for the scenario where you have a CUDA toolkit installed but no GPU driver installed, and you want to build GPU applications, but not run them, of course.  Such a scenario could exist on a head node\/login node in a compute cluster for example.  In that scenario, the login node\/build node may not have a GPU installed, but you may still want to build CUDA <em>driver API<\/em> applications.  Such applications need to build i.e. link against the driver API library, and that library is provided by <code>libcuda.so<\/code>.  Therefore, for this scenario, a \"stub\" library is provided.  The \"stub\" library has everything needed for the API linking process, but is otherwise not functional.)<\/p>\n"}]}
+{"QuestionId":52785019,"AnswerCount":3,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-12T18:12:32.650","AcceptedAnswerId":52976824.0,"OwnerUserId":2510391.0,"Title":"Specify shape for categorical feature columns?","Body":"<p>I know that I can use a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/categorical_column_with_identity\" rel=\"nofollow noreferrer\"><code>categorical_column_with_identity<\/code><\/a> to turn a categorical feature into a series of one-hot features.<\/p>\n\n<p>For instance, if my vocabulary is <code>[\"ON\", \"OFF\", \"UNKNOWN\"]<\/code>:<br>\n<code>\"OFF\"<\/code> -> <code>[0, 1, 0]<\/code><\/p>\n\n<pre><code>categorical_column = tf.feature_column.categorical_column_with_identity('column_name', num_buckets=3)\nfeature_column = tf.feature_column.indicator_column(categorical_column))\n<\/code><\/pre>\n\n<hr>\n\n<p>However, I actually have an 1-dimensional array of categorical features. I would like to turn that into a 2-dimensional series of one-hot features:<\/p>\n\n<p><code>[\"OFF\", \"ON\", \"OFF\", \"UNKNOWN\", \"ON\"]<\/code><br>\n-><br>\n<code>[[0, 1, 0], [1, 0, 0], [0, 1, 0], [0, 0, 1], [1, 0, 0]]<\/code><\/p>\n\n<p>Unlike every other feature column, it doesn't seem like there's a <code>shape<\/code> attribute on <code>categorical_column_with_identity<\/code> and I didn't find any help through Google or the docs.<\/p>\n\n<p>Do I have to give up on <code>categorical_column_with_identity<\/code> and create the 2D array myself through a <code>numerical_column<\/code>?<\/p>\n","answers":[{"AnswerId":"52977510","CreationDate":"2018-10-24T20:32:01.003","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>A couple options for binary encoding<\/p>\n\n<pre><code>import tensorflow as tf\ntest = [\"OFF\", \"ON\", \"OFF\", \"UNKNOWN\", \"ON\"]\nencoding = {x:idx for idx, x in enumerate(sorted(set(test)))}\ntest = [encoding[x] for x in test]\nprint(tf.keras.utils.to_categorical(test, num_classes=len(encoding)))\n\n&gt;&gt;&gt;[[1. 0. 0.]\n    [0. 1. 0.]\n    [1. 0. 0.]\n    [0. 0. 1.]\n    [0. 1. 0.]]\n<\/code><\/pre>\n\n<p>Or from scikit as the other answer stated<\/p>\n\n<pre><code>from sklearn.preprocessing import LabelBinarizer\nencoder = LabelBinarizer()\ntransfomed_label = encoder.fit_transform([\"OFF\", \"ON\", \"OFF\", \"UNKNOWN\", \"ON\"])\nprint(transfomed_label)\n\n&gt;&gt;&gt;[[1 0 0]\n    [0 1 0]\n    [1 0 0]\n    [0 0 1]\n    [0 1 0]]\n<\/code><\/pre>\n"},{"AnswerId":"52998078","CreationDate":"2018-10-25T21:03:56.337","ParentId":null,"OwnerUserId":"3148225","Title":null,"Body":"<p>You can use a dict as a map like this:<\/p>\n\n<pre><code>categorical_features = [\"OFF\", \"ON\", \"OFF\", \"UNKNOWN\", \"ON\"]\none_hot_features = []\n\nmap = {\"ON\": [1, 0, 0], \"OFF\": [0, 1, 0], \"UNKNOWN\": [0, 0, 1]}\n\nfor val in categorical_features:\n    one_hot_features.append(map[val])\n<\/code><\/pre>\n\n<p>or with list comprehension:\n    categorical_features = [\"OFF\", \"ON\", \"OFF\", \"UNKNOWN\", \"ON\"]<\/p>\n\n<pre><code>map = {\"ON\": [1, 0, 0], \"OFF\": [0, 1, 0], \"UNKNOWN\": [0, 0, 1]}\none_hot_features = [map[f] for f in categorical_features]\n<\/code><\/pre>\n\n<p>This should give you what you want.<\/p>\n"},{"AnswerId":"52976824","CreationDate":"2018-10-24T19:44:21.290","ParentId":null,"OwnerUserId":"9209546","Title":null,"Body":"<p>As per the comments, I'm not sure this functionality is possible with <code>tensorflow<\/code>. But with Pandas you have a trivial solution via <a href=\"https:\/\/pandas.pydata.org\/pandas-docs\/stable\/generated\/pandas.get_dummies.html\" rel=\"nofollow noreferrer\"><code>pd.get_dummies<\/code><\/a>:<\/p>\n\n<pre><code>import pandas as pd\n\nL = ['OFF', 'ON', 'OFF', 'UNKNOWN', 'ON']\n\nres = pd.get_dummies(L)\n\nprint(res)\n\n   OFF  ON  UNKNOWN\n0    1   0        0\n1    0   1        0\n2    1   0        0\n3    0   0        1\n4    0   1        0\n<\/code><\/pre>\n\n<p>For performance, or if you need only a NumPy array, you can use <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.LabelBinarizer.html\" rel=\"nofollow noreferrer\"><code>LabelBinarizer<\/code><\/a> from <code>sklearn.preprocessing<\/code>:<\/p>\n\n<pre><code>from sklearn.preprocessing import LabelBinarizer\n\nLB = LabelBinarizer()\n\nres = LB.fit_transform(L)\n\nprint(res)\n\narray([[1, 0, 0],\n       [0, 1, 0],\n       [1, 0, 0],\n       [0, 0, 1],\n       [0, 1, 0]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52785599,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-10-12T19:00:51.603","AcceptedAnswerId":null,"OwnerUserId":470184.0,"Title":"Un-normalizing PyTorch data","Body":"<p>Below code : <\/p>\n\n<pre><code>ux = torch.tensor(np.array([[255,1,255],[255,1,255]])).float()\n\nprint(ux)\n\nux = F.normalize(ux, p=2, dim=1)\n\nprint(ux)\n<\/code><\/pre>\n\n<p>prints : <\/p>\n\n<pre><code>tensor([[ 255.,    1.,  255.],\n        [ 255.,    1.,  255.]])\ntensor([[ 0.7071,  0.0028,  0.7071],\n        [ 0.7071,  0.0028,  0.7071]])\n<\/code><\/pre>\n\n<p>How can I un-normalize the <code>ux<\/code> in order to return to values <\/p>\n\n<pre><code>tensor([[ 255.,    1.,  255.],\n        [ 255.,    1.,  255.]])\n<\/code><\/pre>\n\n<p>from<\/p>\n\n<pre><code>tensor([[ 0.7071,  0.0028,  0.7071],\n        [ 0.7071,  0.0028,  0.7071]])\n<\/code><\/pre>\n\n<p>There are various resources that detail this process such as <a href=\"https:\/\/discuss.pytorch.org\/t\/simple-way-to-inverse-normalize-a-batch-of-input-variable\/12385\/3\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/simple-way-to-inverse-normalize-a-batch-of-input-variable\/12385\/3<\/a> but do not detail unnormalizing result of <code>F.normalize<\/code><\/p>\n","answers":[{"AnswerId":"54783867","CreationDate":"2019-02-20T10:17:19.127","ParentId":null,"OwnerUserId":"6793148","Title":null,"Body":"<p><code>F.normalize<\/code> simply divides by the norm according to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=normalize#torch.nn.functional.normalize\" rel=\"nofollow noreferrer\">documentation<\/a>, so you simply need to multiply it by its magnitude.<\/p>\n\n<p>This means <strong>you still need access to the magnitude of the original vector <code>ux<\/code><\/strong>, otherwise, this is not possible, since <em>the information about the magnitude cannot be recovered from the normalized vector<\/em>.<\/p>\n\n<p>Here's how this can be done:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># I modified the input to make it more interesting, but you can use any other value\nux = torch.tensor(np.array([[255,1,255],[101,10,123]])).float()\n\nmagnitude = ux.norm(p=2, dim=1, keepdim=True)    # NEW\nux = F.normalize(ux, p=2, dim=1)\n\nux_orig = ux * magnitude                         # NEW\n\nprint(ux_orig)\n# Outputs:\n# tensor([[255.,   1., 255.],\n#         [101.,  10., 123.]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52785827,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-10-12T19:23:02.377","AcceptedAnswerId":null,"OwnerUserId":8115171.0,"Title":"Run 1-D Conv using tensorflow","Body":"<p>I'm a beginner in TensorFlow. I want to train a 1-D conv model. \nI have one-rowed csv files for each row of my original data.<\/p>\n\n<p>csv files look like this<\/p>\n\n<pre><code>csv_file1: 1.1, 1.3, 1.5, 1.5, 1\ncsv_file2: 2.1, 2.3, 2.7, 2.9, 0\n<\/code><\/pre>\n\n<p>The last column(containing 1 &amp; 0) are the labels for the single-row csv files<\/p>\n\n<p>Following the <a href=\"https:\/\/stackoverflow.com\/questions\/45427637\/numpy-to-tfrecords-is-there-a-more-simple-way-to-handle-batch-inputs-from-tfrec\/45428167#45428167\">link<\/a> I wrote the following pieces of code. <\/p>\n\n<p>I converted the csv files to TFRecord using the following code<\/p>\n\n<pre><code>    with tf.python_io.TFRecordWriter(filename) as writer:\n        features, label = df_values[:, 1:-1], df_values[:, -1:]\n        example = tf.train.Example()\n        example.features.feature[\"features\"].float_list.value.extend(features[0])\n        example.features.feature[\"label\"].int64_list.value.append(label[0])\n        writer.write(example.SerializeToString())\n<\/code><\/pre>\n\n<p>I want to now read the files and this is the code I'm using. <\/p>\n\n<pre><code>def _parse_function(data_record):\n    features = {\n        'label': tf.FixedLenSequenceFeature([], tf.int64, allow_missing = True),\n        'features': tf.FixedLenSequenceFeature([], tf.float32, allow_missing = True),\n    }\n    sample = tf.parse_single_example(data_record, features)\n    return sample['features'], sample['label']\n\nfilenames = glob.glob(\"*.tfrecords\")\ndataset = tf.data.TFRecordDataset(filenames)\ndataset = dataset.map(_parse_function)  \ndataset = dataset.shuffle(buffer_size=10000)\ndataset = dataset.batch(batch_len)\n# Create a one-shot iterator\niterator = dataset.make_one_shot_iterator()\n\nX,y = iterator.get_next()\n<\/code><\/pre>\n\n<p>From here the problem starts,\nFrom the documentation I understand what session does but failing to put it into code.\nAssuming I'll later figure out how to use<\/p>\n\n<blockquote>\n  <p>tf.seesion.run()<\/p>\n<\/blockquote>\n\n<p>I wrote the below code but don't know how to actually include it into my main script and further use it to train my model.<\/p>\n\n<pre><code>x_train_batch, y_train_batch = tf.train.shuffle_batch(\ntensors=[X_train, y_train],\nbatch_size=batch_size,\ncapacity=capacity,\nmin_after_dequeue=min_after_dequeue,\nenqueue_many=True,\nnum_threads=8)\n\nx_train_batch = tf.cast(x_train_batch, tf.float32)\nx_train_batch = tf.reshape(x_train_batch, shape=(batch_size, 1,65281))\n\ny_train_batch = tf.cast(y_train_batch, tf.int64)\ny_train_batch = tf.one_hot(y_train_batch, num_classes)\n<\/code><\/pre>\n\n<p>Any help regarding how to proceed further will help. <\/p>\n\n<p>PS: Assuming my data was loaded into an np.array\nthe dimension would be, (6571, 65281).\nSince it's astronomical data, each star has 65781 points.<\/p>\n","answers":[]}
+{"QuestionId":52786046,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-12T19:40:36.743","AcceptedAnswerId":52786518.0,"OwnerUserId":6410450.0,"Title":"Get Keras input tensor from model","Body":"<p>I'm working through the Keras example for a sequence to sequence model.<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/lstm_seq2seq.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/lstm_seq2seq.py<\/a><\/p>\n\n<p>In this example, they build a model from an input tensor.<\/p>\n\n<pre><code>encoder_inputs = Input(shape=(None, num_encoder_tokens))\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)\nprint(encoder_inputs)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>Tensor(\"input_1:0\", shape=(?, ?, 71), dtype=float32)\n<\/code><\/pre>\n\n<p>Once a model is built, is there a way to retrieve the input tensor from the model?  Something along the lines of <\/p>\n\n<pre><code>encoder_inputs = model.layers[0].??????\n<\/code><\/pre>\n","answers":[{"AnswerId":"52786518","CreationDate":"2018-10-12T20:22:00.500","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, this is simply <code>model.input<\/code> or <code>model.inputs<\/code><\/p>\n"}]}
+{"QuestionId":52786333,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-12T20:05:39.207","AcceptedAnswerId":52786845.0,"OwnerUserId":8960244.0,"Title":"Why am I getting an Improper dimensions error in Keras?","Body":"<p>I am trying to create a simple ANN model in keras. However, I am getting the following error.<\/p>\n\n<p><code>Error when checking target: expected dense_8 to have 4 dimensions, but got array with shape (36069, 1)<\/code><\/p>\n\n<p>Here is my model and input data.<\/p>\n\n<pre><code>x = x.reshape(48074,1,18,1)\nx_train = x[0:36069]\nx_val = x[36069:38472]\nx_test = x[38472:48074]\ny_train = y[0:36069]#36069\ny_val = y[36069:38472]\ny_test = y[38472:48074]\nmodel = Sequential()\nmodel.add(Dense(50),input_shape=(1,18,1))\nmodel.add(Dense(25))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\n\nmodel.fit(x_train,y_train, epochs=200, batch_size=10, verbose=1,\n          validation_data=(x_val, y_val))\n<\/code><\/pre>\n\n<p>I tried <code>model.summary()<\/code> but could not find my error.<\/p>\n\n<pre><code>dense_9 (Dense)              (None, 1, 18, 50)         100       \n_________________________________________________________________\ndense_10 (Dense)             (None, 1, 18, 25)         1275      \n_________________________________________________________________\ndense_11 (Dense)             (None, 1, 18, 1)          26 \n<\/code><\/pre>\n","answers":[{"AnswerId":"52786845","CreationDate":"2018-10-12T20:50:34.483","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You should flatten your data, because by default a Dense layer will only operate on the last dimension of the data.<\/p>\n\n<pre><code>x = x.reshape(48074,1 * 18 * 1)\nx_train = x[0:36069]\nx_val = x[36069:38472]\nx_test = x[38472:48074]\ny_train = y[0:36069]#36069\ny_val = y[36069:38472]\ny_test = y[38472:48074]\n\nmodel = Sequential()\nmodel.add(Dense(50),input_shape=(1 * 18 * 1,))\nmodel.add(Dense(25))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\n\nmodel.fit(x_train,y_train, epochs=200, batch_size=10, verbose=1,\n          validation_data=(x_val, y_val))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52786431,"AnswerCount":1,"Tags":"<python><docker><tensorflow><machine-learning><classification>","CreationDate":"2018-10-12T20:13:43.650","AcceptedAnswerId":null,"OwnerUserId":8244141.0,"Title":"DecodeJpeg \/ Content: 0 'refers to a tensor that does not exist","Body":"<p>After retraining my model on tensorflow by following method in the tutorial video by Siraj Raval \n<a href=\"https:\/\/www.youtube.com\/watch?v=QfNvhPx5Px8\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=QfNvhPx5Px8<\/a><\/p>\n\n<p>I encountered the below error when i finally tested my test image but it generated two errors as seen in screenshot\n<a href=\"https:\/\/i.stack.imgur.com\/Qj31W.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Qj31W.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>There are two errors ,a Type and a Key error and both of their root cause is probably DecodeJpeg\/Contents: 0 \nIf anyone can explain me the errors and give its resolution then it will be really helpful.<\/p>\n","answers":[{"AnswerId":"52787682","CreationDate":"2018-10-12T22:19:32.840","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>DecodeJpeg\/Contents:0 is supposed to be a tensor, and you want to feed data to it, so you consider it as an input. Problem is that it doesn't exist, this probably means that you made a small mistake in the naming.\nrun this before the sess.run(something, {\"DecodeJpeg\/Contents:0\": something})<\/p>\n\n<pre><code>tf.summary.FileWriter(\"name_of_a_folder\", sess.graph)\n<\/code><\/pre>\n\n<p>this will generate a log file in that folder. then run in cli:<\/p>\n\n<pre><code>tensorboard --log_dir \/name\/to\/that\/folder\/\n<\/code><\/pre>\n\n<p>and open your browser on the link provided in the cli, now you can see the graph and check the real name of the tensor. If you still have problems, feel free to share the graph image, or ask away.<\/p>\n"}]}
+{"QuestionId":52787214,"AnswerCount":1,"Tags":"<r><keras><lstm>","CreationDate":"2018-10-12T21:27:47.163","AcceptedAnswerId":null,"OwnerUserId":8053817.0,"Title":"Keras LSTM and multiple input feature: how to define parameters","Body":"<p>I am discoveting Keras in R and the LSTM. Following this <a href=\"http:\/\/rwanjohi.rbind.io\/2018\/04\/05\/time-series-forecasting-using-lstm-in-r\/\" rel=\"nofollow noreferrer\" title=\"blog\">blog post<\/a>, I want to predict time series, and I would like to use various past time point (t-1, t-2) to predict the t point.\nHere is what I tried so far:<\/p>\n\n<pre><code>library(data.table)\nlibrary(tensorflow)\nlibrary(keras)\n\nSerie &lt;- c(5.66333333333333, 5.51916666666667, 5.43416666666667, 5.33833333333333, \n           5.44916666666667, 6.2025, 6.57916666666667, 6.70666666666667, \n           6.95083333333333, 8.1775, 8.55083333333333, 8.42166666666667, \n           8.01333333333333, 8.99833333333333, 11.0025, 10.3116666666667, \n           10.51, 10.9916666666667, 10.6116666666667, 10.8475, 13.7841666666667, \n           16.2916666666667, 15.9975, 14.3683333333333, 13.4041666666667, \n           11.8666666666667, 9.11916666666667, 9.47862416666667, 9.08404666666667, \n           8.79606166666667, 9.93211091666667, 9.03834041666667, 8.58787275, \n           6.77499383333333, 7.21377583333333, 7.53497175, 6.31212966666667, \n           5.5825105, 4.64021041666667, 4.608787, 5.39446983333333, 4.93945983333333, \n           4.8612215, 4.13088808333333, 4.09916575, 3.40943183333333, 3.79573258333333, \n           4.30319966666667, 4.23431266666667, 3.64880758333333, 3.11700716666667, \n           3.321058, 2.53599408333333, 2.20433991666667, 1.66643905833333, \n           0.84187275, 0.467880658333333, 0.810507858333333, 0.795)\n\nNpoints &lt;- 2 # number of previous point to take into account\n<\/code><\/pre>\n\n<p>I then create a data frame with the lagged time series, and create a test and train set:<\/p>\n\n<pre><code>supervised &lt;- data.table(x = diff(Serie, differences = 1))\nsupervised[,c(paste0(\"x-\",1:Npoints)) := lapply(1:Npoints,function(i){c(rep(NA,i),x[1:(.N-i)])})] # create shifted versions\n\n# take the non NA\nsupervised &lt;- supervised[!is.na(get(paste0(\"x-\",Npoints)))]\nhead(supervised)\n\n# Split dataset into training and testing sets\nN = nrow(supervised)\nn = round(N *0.7, digits = 0)\ntrain = supervised[1:n, ]\ntest  = supervised[(n+1):N,  ]\n<\/code><\/pre>\n\n<p>I rescale the data<\/p>\n\n<pre><code>scale_data = function(train, test, feature_range = c(0, 1)) {\n  x = train\n  fr_min = feature_range[1]\n  fr_max = feature_range[2]\n  std_train = ((x - min(x,na.rm = T) ) \/ (max(x,na.rm = T) - min(x,na.rm = T)  ))\n  std_test  = ((test - min(x,na.rm = T) ) \/ (max(x,na.rm = T) - min(x,na.rm = T)  ))\n  scaled_train = std_train *(fr_max -fr_min) + fr_min\n  scaled_test = std_test *(fr_max -fr_min) + fr_min\n  return( list(scaled_train = as.vector(scaled_train), scaled_test = as.vector(scaled_test) ,scaler= c(min =min(x,na.rm = T), max = max(x,na.rm = T))) )\n}\n\nScaled = scale_data(train, test, c(-1, 1))\n\n# define x and y train\ny_train = as.vector(Scaled$scaled_train[, 1]) \nx_train = Scaled$scaled_train[, -1] \n<\/code><\/pre>\n\n<p>And following <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">this post<\/a> I reshape the data in 3D<\/p>\n\n<pre><code>x_train_reshaped &lt;- array(NA,dim= c(1,dim(x_train)))\nx_train_reshaped[1,,] &lt;- as.matrix(x_train)\n<\/code><\/pre>\n\n<p>I do the following model and try to start the learning :<\/p>\n\n<pre><code>model &lt;- keras_model_sequential() \nmodel%&gt;%\n  layer_lstm(units = 1, batch_size = 1, input_shape = dim(x_train), stateful= TRUE)%&gt;%\n  layer_dense(units = 1)\n\n# compile model ####\nmodel %&gt;% compile(\n  loss = 'mean_squared_error',\n  optimizer = optimizer_adam( lr= 0.02, decay = 1e-6 ),  \n  metrics = c('accuracy')\n)\n\n# make a test\nmodel %&gt;% fit(x_train_reshaped, y_train, epochs=1, batch_size=1, verbose=1, shuffle=FALSE)\n<\/code><\/pre>\n\n<p>but I get the following error:<\/p>\n\n<blockquote>\n  <p>Error in py_call_impl(callable, dots$args, dots$keywords) : \n    ValueError: No data provided for \"dense_11\". Need data for each key in: ['dense_11'] <\/p>\n<\/blockquote>\n\n<p>Trying to reshape the data differently didn't help. \nWhat I am doing wrong ?<\/p>\n","answers":[{"AnswerId":"52788034","CreationDate":"2018-10-12T23:06:47.383","ParentId":null,"OwnerUserId":"100129","Title":null,"Body":"<p>Keras and tensorflow in R cannot recognise the size of your input\/target data when they are data frames.<\/p>\n\n<p>y_train is both a data.table and a data.frame:<\/p>\n\n<pre><code>class(y_train)\n[1] \"data.table\" \"data.frame\"\n<\/code><\/pre>\n\n<p>The keras fit documentation states: \"y: Vector, matrix, or array of target (label) data (or list if the model has multiple outputs).\"  Similarly, for x.<\/p>\n\n<p>Unfortunately, there still appears to be an input and\/or target dimensionality mismatch when y_train is cast to a matrix:<\/p>\n\n<pre><code>model %&gt;% \n fit(x_train_reshaped, as.matrix(y_train), epochs=1, batch_size=1, verbose=1, shuffle=FALSE)\nError in py_call_impl(callable, dots$args, dots$keywords) :\n  ValueError: Input arrays should have the same number of samples as target arrays. \nFound 1 input samples and 39 target samples.\n<\/code><\/pre>\n\n<p>Hope this answer helps you, or someone else, make further progress.<\/p>\n"}]}
+{"QuestionId":52787372,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-12T21:45:50.977","AcceptedAnswerId":52787701.0,"OwnerUserId":155137.0,"Title":"TensorFlow: Add dimension (column) with constant value","Body":"<p>I have a Tensor of shape <code>(-1,)<\/code> which represents a list of indices. I want to create a Tensor of shape <code>(-1,2)<\/code>. The first column should be the same as the list of indices, while the second column should be filled with a constant.<\/p>\n\n<p>Eg (let's say the constant is 6):<\/p>\n\n<pre><code>indices = [4, 35, 230, 235]\nmy_goal = [[2, 6], [35, 6], [230, 6], [235, 6]]\n<\/code><\/pre>\n\n<p>What is the best way to do this? I was hoping a broadcasting-<code>concat<\/code> would work, but <code>concat<\/code> doesn't seem to support broadcasting.<\/p>\n","answers":[{"AnswerId":"52787701","CreationDate":"2018-10-12T22:21:23.153","ParentId":null,"OwnerUserId":"2172751","Title":null,"Body":"<p>You can use <code>tensorflow.pad<\/code>. But first you have to make it a two dimensional tensor.<\/p>\n\n<pre><code>indices = tf.constant([1,2,3,4])\nindices = tf.expand_dims(tf, 1)  # now you have a (4,1) tensor\npadding = [[0,0],[0,1]] # no padding before or after the first dimension\n                        # no padding before second dimension. Single-element padding\n                        # after the second dimension\nmy_goal = tf.pad(indices, padding, constant_values=6)\n<\/code><\/pre>\n\n<p>Padding has to be [n,2] tensor. For more information, look into: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad<\/a>.<\/p>\n"}]}
+{"QuestionId":52787643,"AnswerCount":1,"Tags":"<tensorflow><https><tensorflow-serving>","CreationDate":"2018-10-12T22:14:48.240","AcceptedAnswerId":null,"OwnerUserId":3226016.0,"Title":"Tensorflow serving over https","Body":"<p>I need to configure tensorflow serving to work over https\/ssl for the REST API.<\/p>\n\n<p>Currently I am running the model as<\/p>\n\n<p>.\/tensorflow_model_server --rest_api_port=9000 --model_name=t2v.pb --model_base_path=\/apps\/tfmdl_out\/mdl1\/<\/p>\n\n<p>It works over http locally, but due to VPN rules remote requests have to go through SSL.<\/p>\n\n<p>How can I serve the REST API over SSL?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Ilya<\/p>\n","answers":[{"AnswerId":"52787697","CreationDate":"2018-10-12T22:21:04.143","ParentId":null,"OwnerUserId":"748102","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.nginx.com\" rel=\"nofollow noreferrer\">Nginx<\/a> (a web server) as a proxy to listen on HTTPS, and internally ask Nginx to do the HTTP traffic in your network. See <a href=\"https:\/\/docs.nginx.com\/nginx\/admin-guide\/security-controls\/securing-http-traffic-upstream\/\" rel=\"nofollow noreferrer\">here<\/a> for more information on how to do this.<\/p>\n"}]}
+{"QuestionId":52787988,"AnswerCount":1,"Tags":"<c++><tensorflow><imagenet>","CreationDate":"2018-10-12T23:01:10.520","AcceptedAnswerId":null,"OwnerUserId":10497433.0,"Title":"Run tflite accuracy tool on official tensorflow resnet50 model","Body":"<p>I have downloaded the official resnet50 model provided here: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/official\/resnet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/official\/resnet<\/a>. I needed a tflite quantized version of this model and hence I converted the model to a tflite format as follows :<\/p>\n\n<pre><code>toco --output_file \/tmp\/resnet50_quant.tflite --saved_model_dir &lt;path\/to\/saved_model_dir&gt; --output_format TFLITE  --quantize_weights QUANTIZE_WEIGHTS\n<\/code><\/pre>\n\n<p>After this, I thought I'd run the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\" rel=\"nofollow noreferrer\">tflite accuracy tool<\/a> to verify the accuracy of this model is still reasonable. Although it looks like I run into the following issue:<\/p>\n\n<pre><code>bazel run -c opt --copt=-march=native --cxxopt='--std=c++11'   --   \/\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc:imagenet_accuracy_eval --model_file=\/tmp\/resnet50_quant.tflite --ground_truth_images_path=&lt;path\/to\/images&gt; --ground_truth_labels=\/tmp\/validation_labels.txt --model_output_labels=\/tmp\/tf_labels.txt --output_file_path=\/tmp\/accuracy_output.txt --num_images=0\nINFO: Analysed target \/\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc:imagenet_accuracy_eval (0 packages loaded).\nINFO: Found 1 target...\nTarget \/\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc:imagenet_accuracy_eval up-to-date:\n  bazel-bin\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\/imagenet_accuracy_eval\nINFO: Elapsed time: 14.589s, Critical Path: 14.28s\nINFO: 3 processes: 3 local.\nINFO: Build completed successfully, 4 total actions\nINFO: Running command line: bazel-bin\/tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\/imagenet_accuracy_eval '--model_file=\/tmp\/resnet50_quant.tflite' '--ground_truth_images_path=&lt;path\/to\/images&gt;' '--ground_truth_labels=\/tmp\/validation_labels.txt' '--model_output_labels=\/tmp\/tf_labels.txt' '--output_file_path=\/tmp\/accuracy_output.txt' 'INFO: Build completed successfully, 4 total actions\n2018-10-12 15:30:06.237058: E tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\/imagenet_accuracy_eval.cc:155] Starting evaluation with: 4 threads.\n2018-10-12 15:30:06.536802: E tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\/imagenet_accuracy_eval.cc:98] Starting model evaluation: 50000\n2018-10-12 15:30:06.565334: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at run_tflite_model_op.cc:89 : Invalid argument: Data shapes mismatch for tensors: 0 expected: [64,224,224,3] got: [1,224,224,3]\n2018-10-12 15:30:06.565453: F tensorflow\/contrib\/lite\/tools\/accuracy\/ilsvrc\/imagenet_model_evaluator.cc:222] Non-OK-status: eval_pipeline-&gt;Run(CreateStringTensor(image_label.image), CreateStringTensor(image_label.label)) status: Invalid argument: Data shapes mismatch for tensors: 0 expected: [64,224,224,3] got: [1,224,224,3]\n     [[{{node stage_run_tfl_model_output}} = RunTFLiteModel[input_type=[DT_FLOAT], model_file_path=\"\/tmp\/resnet50_quant.tflite\", output_type=[DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](stage_inception_preprocess_output)]]\n<\/code><\/pre>\n\n<p>It looks like the issue is that the official resnet model has an input tensor of [64, 224, 224, 3] whereas the accuracy tool provides an input of [1, 224, 224, 3]. So, the official model seems to expect a batch of 64 images and hence the accuracy tool fails.<\/p>\n\n<p>I was wondering what I need to do to get the accuracy tool to run on the official resnet50 model? I'm guessing that although the input tensor for resnet 50 is [64, 224, 224, 3], there should be a way to still run a single image through the model.<\/p>\n","answers":[{"AnswerId":"54584115","CreationDate":"2019-02-07T23:51:37.430","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>There are two ways to go about it:<\/p>\n\n<ol>\n<li><p>Resize the input of your model to [1, 224, 224, 3] and run the tool. \nYou could try looking at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16216\" rel=\"nofollow noreferrer\">this<\/a> and then modifying <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/tools\/accuracy\/run_tflite_model_op.cc\" rel=\"nofollow noreferrer\">this file<\/a> accordingly.<\/p><\/li>\n<li><p>Alternatively modify the same tool so that it feeds in 64 images at a time instead of 1. You can look at the same code file I point to above and feed 64 at a time instead of 1.<\/p><\/li>\n<\/ol>\n\n<p>If you're looking for long-term support, consider filing a feature request on Github where we can support batching.<\/p>\n"}]}
+{"QuestionId":52788174,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2018-10-12T23:26:31.127","AcceptedAnswerId":null,"OwnerUserId":5758533.0,"Title":"How to index a list of LSTMStateTuple with a tensor?","Body":"<p>I have a python list of <code>LSTMStateTuple<\/code> objects and I have to use a tensor as index to retrieve them. For example:<\/p>\n\n<pre><code>index = tf.constant(0)\nlstm = tf.nn.rnn_cell.LSTMCell(128)\nstates = [lstm.zero_state(10, tf.float32), lstm.zero_state(10, tf.float32)]\n<\/code><\/pre>\n\n<p>If I try <code>state = states[index]<\/code> I get an error and <code>state = tf.gather(states, index)<\/code> converts <code>states<\/code> in a tensor and returns a tensor of shape <code>[10, 2, 128]<\/code>. <\/p>\n\n<p>How can I get a <code>LSTMStateTuple<\/code> instead of a tensor? I would like to avoid the conversions from list of <code>LSTMStateTuple<\/code> to tensor and from tensor to <code>LSTMStateTuple<\/code> when I pass to the lstm the state.<\/p>\n","answers":[{"AnswerId":"53053514","CreationDate":"2018-10-29T20:43:37.247","ParentId":null,"OwnerUserId":"1757224","Title":null,"Body":"<p>You create two states and put them in a <code>LSTMStateTuple<\/code>.<\/p>\n\n<pre><code>cell_state = tf.placeholder(tf.float32, [batch_size, state_size])\nhidden_state = tf.placeholder(tf.float32, [batch_size, state_size])\ninit_state = tf.nn.rnn_cell.LSTMStateTuple(cell_state, hidden_state)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52788635,"AnswerCount":2,"Tags":"<python><keras><prediction><cross-validation>","CreationDate":"2018-10-13T01:01:52.590","AcceptedAnswerId":null,"OwnerUserId":9415043.0,"Title":"Early Stopping with a Cross-Validated Metric in Keras","Body":"<p>Is there a way in Keras to cross-validate the early stopping metric being monitored <code>EarlyStopping(monitor = 'val_acc', patience = 5)<\/code>? Before allowing training to proceed to the next epoch, could the model be cross-validated to get a more robust estimate of the test error? What I have found is that the early stopping metric, say the accuracy on a validation set, can suffer from high variance. Early-stopped models often do not perform nearly as well on unseen data, and I suspect this is because of the high variance associated with the validation set approach.<\/p>\n\n<p>To minimize the variance in the early stopping metric, I would like to k-fold cross-validate the early stopping metric as the model trains from epoch <code>i<\/code> to epoch <code>i + 1<\/code>. I would like to take the model at epoch <code>i<\/code>, divide the training data into 10 parts, learn on 9 parts, estimate the error on the remaining part, repeat so that all 10 parts have had a chance to be the validation set, and then proceed with training to epoch <code>i + 1<\/code> with the full training data as usual. The average of the 10 error estimates will hopefully be a more robust metric that can be used for early stopping.<\/p>\n\n<p>I have tried to write a custom metric function that includes k-fold cross-validation but I can't get it to work. Is there a way to cross-validate the early stopping metric being monitored, perhaps through a custom function inside the Keras model or a loop outside the Keras model? <\/p>\n\n<p>Thanks!!<\/p>\n","answers":[{"AnswerId":"52843229","CreationDate":"2018-10-16T20:02:11.633","ParentId":null,"OwnerUserId":"9415043","Title":null,"Body":"<p>I imagine that using a callback as suggested by @VincentPakson would be cleaner and more efficient, but the level of programming required is beyond me. I was able to create a for loop to do what I wanted by:<\/p>\n\n<ol>\n<li><p>Training a model for a single epoch and saving it using <code>model.save()<\/code>.<\/p><\/li>\n<li><p>Loading the saved model and training the model for a single epoch for each of the 10 folds (i.e. 10 models), then averaging the 10 validation set errors.<\/p><\/li>\n<li><p>Loading the saved model and training for a single epoch using all of the training data, and the overwriting the saved model with this model.<\/p><\/li>\n<li><p>Repeating steps 1-3 until the estimate from 2 stops improving for a given patience.<\/p><\/li>\n<\/ol>\n\n<p>I'd love a better answer but this seems to work. Slowly.<\/p>\n"},{"AnswerId":"52809789","CreationDate":"2018-10-15T04:30:42.913","ParentId":null,"OwnerUserId":"8262615","Title":null,"Body":"<p>Keras really does have a greatly customizable callback functionality as can be seen \n<a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>If you are not satisfied with keras current <code>EarlyStopping<\/code> function, which for me does the work of checking the validation loss during training, you could create a custom callback function. Also custom callback functions can be chained. <\/p>\n\n<p>If your issue is accessing the model inside the callback, then <code>self<\/code> is the variable you want to access as can be seen in this <a href=\"https:\/\/stackoverflow.com\/questions\/47079111\/create-keras-callback-to-save-model-predictions-and-targets-for-each-batch-durin\">answer<\/a>. I don't completely understand why you want to \"redevelop\" the model during test phase. BUT then with that, you could still use callbacks, after the <code>EarlyStopping<\/code> callback, you could create another callback function that could \"redevelop\" the model. <\/p>\n\n<p>If you want to access the models deeper variables you could use, Keras <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">backend<\/a>. <\/p>\n\n<p>I hope I helped.<\/p>\n"}]}
+{"QuestionId":52788890,"AnswerCount":1,"Tags":"<tensorflow><image-processing><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-13T02:07:37.147","AcceptedAnswerId":null,"OwnerUserId":3977329.0,"Title":"Matterport Mask R-CNN - Unpredictable loss values and strange detection results for larger images","Body":"<p>I'm having trouble achieving viable results with Mask R-CNN and I can't seem to pinpoint why. I am using a fairly limited dataset (13 images) of large greyscale images (2560 x 2160) where the detection target is very small (mean area of 26 pixels). I have run inspect_nucleus_data.ipynb across my data and verified that the masks and images are being interpreted correctly. I've also followed the wiki guide (<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\/wiki\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN\/wiki<\/a>) to have my images read and dealt with as greyscale images rather than just converting them to RGB. Here is one of the images with the detection targets labelled.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QdqXQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QdqXQ.png\" alt=\"annotated image\"><\/a><\/p>\n\n<p>During training, the loss values are pretty unpredictable, bouncing between around 1 and 2 without ever reaching a steady decline where it seems like it's converging at all. I'm using these config values at the moment; they're the best I've been able to come up with while fighting off OOM errors:<\/p>\n\n<pre><code>Configurations:\nBACKBONE                       resnet101\nBACKBONE_STRIDES               [4, 8, 16, 32, 64]\nBATCH_SIZE                     1\nBBOX_STD_DEV                   [0.1 0.1 0.2 0.2]\nCOMPUTE_BACKBONE_SHAPE         None\nDETECTION_MAX_INSTANCES        450\nDETECTION_MIN_CONFIDENCE       0\nDETECTION_NMS_THRESHOLD        0.3\nFPN_CLASSIF_FC_LAYERS_SIZE     1024\nGPU_COUNT                      1\nGRADIENT_CLIP_NORM             5.0\nIMAGES_PER_GPU                 1\nIMAGE_CHANNEL_COUNT            1\nIMAGE_MAX_DIM                  1024\nIMAGE_META_SIZE                14\nIMAGE_MIN_DIM                  1024\nIMAGE_MIN_SCALE                0\nIMAGE_RESIZE_MODE              square\nIMAGE_SHAPE                    [1024 1024    1]\nLEARNING_MOMENTUM              0.9\nLEARNING_RATE                  0.001\nLOSS_WEIGHTS                   {'mrcnn_class_loss': 1.0, 'mrcnn_bbox_loss': 1.0,     'rpn_bbox_loss': 1.0, 'mrcnn_mask_loss': 1.0, 'rpn_class_loss': 1.0}\nMASK_POOL_SIZE                 14\nMASK_SHAPE                     [28, 28]\nMAX_GT_INSTANCES               450\nMEAN_PIXEL                     [16.49]\nMINI_MASK_SHAPE                (56, 56)\nNAME                           nucleus\nNUM_CLASSES                    2\nPOOL_SIZE                      7\nPOST_NMS_ROIS_INFERENCE        1000\nPOST_NMS_ROIS_TRAINING         2000\nPRE_NMS_LIMIT                  6000\nROI_POSITIVE_RATIO             0.33\nRPN_ANCHOR_RATIOS              [0.5, 1, 2]\nRPN_ANCHOR_SCALES              (2, 4, 8, 16, 32)\nRPN_ANCHOR_STRIDE              1\nRPN_BBOX_STD_DEV               [0.1 0.1 0.2 0.2]\nRPN_NMS_THRESHOLD              0.9\nRPN_TRAIN_ANCHORS_PER_IMAGE    512\nSTEPS_PER_EPOCH                11\nTOP_DOWN_PYRAMID_SIZE          256\nTRAIN_BN                       False\nTRAIN_ROIS_PER_IMAGE           256\nUSE_MINI_MASK                  True\nUSE_RPN_ROIS                   True\nVALIDATION_STEPS               1\nWEIGHT_DECAY                   0.0001\n<\/code><\/pre>\n\n<p>I'm training on all layers. The output I'm getting generally looks like this, with grid-like detections found in weird spots without ever seeming to accurately identify a nucleus. I've added the red square just to highlight a very obvious cluster of nuclei that have been missed:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/49Ser.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/49Ser.png\" alt=\"output image\"><\/a><\/p>\n\n<p>Here is a binary mask of these same detections so you can see their shape:\n<a href=\"https:\/\/i.stack.imgur.com\/AGkAM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AGkAM.png\" alt=\"output mask\"><\/a><\/p>\n\n<p>Could anyone shed some light on what might be going wrong here?<\/p>\n","answers":[{"AnswerId":"57969234","CreationDate":"2019-09-17T07:21:57.743","ParentId":null,"OwnerUserId":"5303350","Title":null,"Body":"<p>Some information is lost when you resize the training image dimension by half. \nSquare also uses a lot of memory. So you might want to use crop instead. Instead of 1024*1024 in square mode, you will have 512*512. You might encounter NAN due to bounding box out of range, in that case you need to make some adjustment to your data feed. <\/p>\n\n<p>You want to turn off mini mask because that will influence with your accuracy. Using crop mode should help out a ton with memory. So you shouldn't worry.<\/p>\n"}]}
+{"QuestionId":52789086,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><overwrite><tensor>","CreationDate":"2018-10-13T02:49:42.060","AcceptedAnswerId":52789342.0,"OwnerUserId":2956066.0,"Title":"Odd behaviour of \"out\" argument in torch.empty()","Body":"<p>I have a sample tensor like this:<\/p>\n\n<pre><code>In [137]: x = x.new_ones((5, 3), dtype=torch.double)    \nIn [138]: x\nOut[138]: \ntensor([[1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.]], dtype=torch.float64)\n<\/code><\/pre>\n\n<p>Now, I want to free the memory of this tensor by overwriting the contents using <code>torch.empty()<\/code> which takes an <code>out<\/code> argument.<\/p>\n\n<pre><code>In [139]: torch.empty((5, 3), out=x)\nOut[139]: \ntensor([[1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.]], dtype=torch.float64)\n<\/code><\/pre>\n\n<p>However, the values in the original tensor <code>x<\/code> still remains the same. If this is the case, then what is the purpose of this keyword argument <code>out<\/code> in <code>torch.empty<\/code>? What am I missing here?<\/p>\n","answers":[{"AnswerId":"52789342","CreationDate":"2018-10-13T03:46:17.920","ParentId":null,"OwnerUserId":"4686625","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/aten\/src\/ATen\/native\/TensorFactories.cpp#L112\" rel=\"nofollow noreferrer\">Here's<\/a> the C++ implementation of <code>empty<\/code> with an out param from the source code. <\/p>\n\n<pre><code>Tensor&amp; empty_out(Tensor&amp; result, IntList size) {\n  if (result.is_sparse()) {\n    result.sparse_resize_and_clear_(size, size.size(), 0);\n  } else {\n    result.resize_(size);\n  }\n  return result;\n}\n<\/code><\/pre>\n\n<p>So for dense tensors all it does is resize the tensor appropriately - in your case the size is the same. <\/p>\n\n<pre><code>In [21]: x = torch.ones((5, 3), dtype=torch.double)                                                                                                                                        \n\nIn [22]: torch.empty((2, 3), out=x)                                                                                                                                                        \nOut[22]: \ntensor([[1., 1., 1.],\n        [1., 1., 1.]], dtype=torch.float64)\n\nIn [23]: torch.empty((2, 8), out=x)                                                                                                                                                        \nOut[23]: \ntensor([[ 1.0000e+00,  1.0000e+00,  1.0000e+00,  1.0000e+00,  1.0000e+00,\n          1.0000e+00,  1.0000e+00,  1.0000e+00],\n        [ 1.0000e+00,  1.0000e+00,  1.0000e+00,  1.0000e+00,  1.0000e+00,\n          1.0000e+00,  1.0000e+00, 4.6631e-310]], dtype=torch.float64)\n<\/code><\/pre>\n\n<p>First of all, <code>empty<\/code> doesn't free memory - it only cares about allocation of a tensor of an appropriate size. In your case such a tensor has already been allocated, so <code>empty<\/code> has nothing to do.. it is not going to go allocate a new empty tensor somewhere else in memory. In the second <code>empty<\/code> example above we are forced to allocate for a tensor with a larger size (2 * 8 = 16 compared to 5 * 3 = 15), and we can see the last element in this empty array is garbage, since it is beyond the contiguous memory block that had previously been initialized. <code>empty<\/code> won't force-clear your whole tensor to 0 or something like that because again, it is uninitialized data. <\/p>\n"}]}
+{"QuestionId":52789285,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-hub>","CreationDate":"2018-10-13T03:35:29.053","AcceptedAnswerId":null,"OwnerUserId":7874360.0,"Title":"Can't convert 'images': Expected float32, got Ellipsis of type 'ellipsis' instead","Body":"<p>hub.<strong>version<\/strong> == 0.1.1\n'''\n    import tensorflow_hub as hub<\/p>\n\n<pre><code>module_spec = hub.load_module_spec(\"\/home\/neo\/learn_tf\/inception_v3\")\nheight, width = hub.get_expected_image_size(module_spec)\nimages = ...  # A batch of images with shape [batch_size, height, \nwidth, 3].\nmodule = hub.Module(module_spec)\nfeatures = module(images)   # A batch with shape [batch_size, \nnum_features].\n<\/code><\/pre>\n\n<p>'''<\/p>\n\n<p>I just copy it c'<a href=\"https:\/\/tfhub.dev\/google\/imagenet\/inception_v3\/feature_vector\/1\" rel=\"nofollow noreferrer\">https:\/\/tfhub.dev\/google\/imagenet\/inception_v3\/feature_vector\/1<\/a>', when I run it but got <\/p>\n\n<h2>Can't convert 'images': Expected float32, got Ellipsis of type 'ellipsis' instead.<\/h2>\n\n<p>HELP ME<\/p>\n","answers":[{"AnswerId":"53119334","CreationDate":"2018-11-02T13:19:33.043","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>You are asked to supply some actual code where the example snippets displays the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Ellipsis\" rel=\"nofollow noreferrer\">Ellipsis<\/a>.<\/p>\n\n<p>By coincidence, three dots are valid Python3 syntax for use in slice notation, and they are parsed into an object of its own helper type <code>Ellipsis<\/code> (see <a href=\"https:\/\/stackoverflow.com\/questions\/772124\/what-does-the-python-ellipsis-object-do\">this question<\/a> for more). This explains the error message you see.<\/p>\n"}]}
+{"QuestionId":52789325,"AnswerCount":1,"Tags":"<python><keras><lstm>","CreationDate":"2018-10-13T03:42:34.980","AcceptedAnswerId":null,"OwnerUserId":9266500.0,"Title":"How to set size of hidden state vector in LSTM, keras?","Body":"<p>I am currently setting the vector size by using <code>model.add(LSTM(50))<\/code> i.e setting the value in <code>units<\/code> attribute but I highly doubt its correctness(In keras documentation, units is explained as dimensionality of the output space). Anyone who can help me here?<\/p>\n","answers":[{"AnswerId":"52846778","CreationDate":"2018-10-17T03:14:06.837","ParentId":null,"OwnerUserId":"8075709","Title":null,"Body":"<p>If by vector size you mean the number of nodes in a layer, then yes you are doing it right. The output dimensionality of your layer \nis the same as the number of nodes. The same thing applies to convolutional layers, number of filters and output dimensionality along the last axis, aka number of color channels is the same.<\/p>\n"}]}
+{"QuestionId":52789329,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-13T03:43:04.603","AcceptedAnswerId":null,"OwnerUserId":10497857.0,"Title":"How to merge multiple files on an irregular index with forward fill in Tensorflow","Body":"<p><strong>Background:<\/strong><\/p>\n\n<p>I have time sequence data across several csv files that each contain different kinds of sensor features. I am trying to merge the features in these files together on a common timestamp index.<\/p>\n\n<p>The complication I've been having trouble with is that the timestamp a row of data was logged with does not correspond with some regular interval, but instead appears to correspond to the time that a sensor updated its value.<\/p>\n\n<p><strong>Question:<\/strong><\/p>\n\n<p>I am looking for a way to merge these files together in a \"forward-fill\"-like way where if a dataset does not have an an entry for a given timestamp, it would fill with the previous entry's values.<\/p>\n\n<p>For example, if my two data files were<\/p>\n\n<pre><code>A.csv\ntime  X1  X2\n  00   1   2\n  01   3   4\n  02   5   6\n  03   7   8\n\nB.csv\ntime  X3  X4\n  00  10  11\n  02  12  13\n<\/code><\/pre>\n\n<p>I would like a final dataset with values<\/p>\n\n<pre><code>time  X1  X2  X3  X4\n  00   1   2  10  11\n  01   3   4  10  11\n  02   5   6  12  13\n  03   7   8  12  13\n<\/code><\/pre>\n\n<p>that I could pipe into my tensorflow estimator.<\/p>\n\n<p>I have attempted merging different CsvDatasets together, but it seems like it is not possible to \"pause\" a dataset iterator on a value to propagate it and merge it with the other dataset iterators in the way that I want.<\/p>\n\n<p>(I know that I could do this with pandas by outer-joining on the time index and then doing pd.ffill, but I believe that I would then need the whole dataset in memory, which I am trying to avoid. I don't think loading the data in chunks is an option, either, due to the irregular time index.)<\/p>\n\n<p>Does anyone know the best way to create a pipeline like this?<\/p>\n","answers":[{"AnswerId":"52789386","CreationDate":"2018-10-13T03:54:58.903","ParentId":null,"OwnerUserId":"8008776","Title":null,"Body":"<p>You can do the merge in pandas by taking advantage of the fact that <code>pd.concat([A, B], axis=1)<\/code> defaults to merging on the indexes of <code>A<\/code> and <code>B<\/code>, and then <code>ffill<\/code> the result.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>A = pd.read_csv('A.csv')\nB = pd.read_csv('B.csv')\n\nA = A.set_index('time')\nB = B.set_index('time')\n\nres = pd.concat([A, B], axis=1).ffill()\nres\n<\/code><\/pre>\n\n<p>The print of <code>res<\/code> is:<\/p>\n\n<pre><code>      X1  X2    X3    X4\ntime                    \n0      1   2  10.0  11.0\n1      3   4  10.0  11.0\n2      5   6  12.0  13.0\n3      7   8  12.0  13.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":52789457,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><tensor>","CreationDate":"2018-10-13T04:12:57.807","AcceptedAnswerId":52822361.0,"OwnerUserId":4496896.0,"Title":"How to perform np.append() type operation on tensors in tensorflow?","Body":"<p>I want to add a new tensor in an existing tensor as last column. With numpy I can use <code>np.append()<\/code>, but I'm not sure how to do this on tensors in tensorflow. Any suggestion?<\/p>\n\n<pre><code>&gt;&gt; a = np.array([[1,2], [3,4], [5,6]])\n&gt;&gt; b = np.array([[9], [99], [999]])\n&gt;&gt; np.append(a, b, axis=1)\narray([[  1,   2,   9],\n       [  3,   4,  99],\n       [  5,   6, 999]])\n<\/code><\/pre>\n\n<p>in tensorflow..?<\/p>\n\n<pre><code>&gt;&gt; x = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]], tf.float32)\n&gt;&gt; y = tf.constant([[9.0], [99.0], [999.0]], tf.float32)\n&gt;&gt; ???\n<\/code><\/pre>\n","answers":[{"AnswerId":"52822361","CreationDate":"2018-10-15T18:01:11.867","ParentId":null,"OwnerUserId":"4496896","Title":null,"Body":"<p>solved: tf.concat((x, y), axis=-1)<\/p>\n"}]}
+{"QuestionId":52789474,"AnswerCount":0,"Tags":"<python><linux><deep-learning><load><pytorch>","CreationDate":"2018-10-13T04:15:59.883","AcceptedAnswerId":null,"OwnerUserId":9623065.0,"Title":"_pickle.UnpicklingError: unpickling stack underflow","Body":"<p>I am new to Deep-Learning, and I try to recurrence the code in <a href=\"https:\/\/github.com\/swamiviv\/LSD-seg\" rel=\"nofollow noreferrer\">https:\/\/github.com\/swamiviv\/LSD-seg<\/a>, which is used for image semantic segmentation, but a problem happened in Pytorch, I really need your help.<\/p>\n\n<p>Thanks very much.<\/p>\n\n<p>Here's the environment:<\/p>\n\n<pre><code>Ubuntu 14.04.1\nPython 3.6.4\nPytorch 0.4.0\n<\/code><\/pre>\n\n<p>The code is for downloading the pretrained model but my server can't visit googledrive, so I downloaded the model from other method and updated the model file to the server, and changed the related code, but it didn't work.<\/p>\n\n<p>The original code:<\/p>\n\n<pre><code>import os.path as osp\nimport fcn\nimport torch\nimport torchvision\n\ndef VGG16(pretrained=False):\n    model = torchvision.models.vgg16(pretrained=False)\n    if not pretrained:\n        return model\n\n    model_file = _get_vgg16_pretrained_model()\n    state_dict = torch.load(model_file)\n    model.load_state_dict(state_dict)\n    return model\n\ndef _get_vgg16_pretrained_model():\n    return fcn.data.cached_download(url='http:\/\/drive.google.com\/uc?id=0B9P1L--7Wd2vLTJZMXpIRkVVRFk',path=osp.expanduser('\/home\/sunyuze\/LSD\/data\/vgg16_from_caffe.pth'),md5='aa75b158f4181e7f6230029eb96c1b13',)\n<\/code><\/pre>\n\n<p>After I downloaded the model file by other method and put it in the specific position, I modified the code:<\/p>\n\n<pre><code>import os.path as osp\nimport fcn\nimport torch\nimport torchvision\n\ndef VGG16(pretrained=False):\n    model = torchvision.models.vgg16(pretrained=False)\n    if not pretrained:\n        return model\n\n    state_dict = torch.load('\/home\/sunyuze\/LSD\/data\/vgg16_from_caffe.pth')\n    model.load_state_dict(state_dict)\n    return model\n<\/code><\/pre>\n\n<p>The error alert is :<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 157, in &lt;module&gt;\n    main()\n  File \"train.py\", line 76, in main\n    vgg16 = torchfcn.models.VGG16(pretrained=True)\n  File \"\/home\/sunyuze\/LSD\/code\/torchfcn\/models\/vgg.py\", line 17, in VGG16\n    state_dict = torch.load('\/home\/sunyuze\/LSD\/data\/vgg16_from_caffe.pth')\n  File \"\/home\/sunyuze\/anaconda3\/lib\/python3.6\/site-packages\/torch\/serialization.py\", line 303, in load\n    return _load(f, map_location, pickle_module)\n  File \"\/home\/sunyuze\/anaconda3\/lib\/python3.6\/site-packages\/torch\/serialization.py\", line 459, in _load\n    magic_number = pickle_module.load(f)\n_pickle.UnpicklingError: unpickling stack underflow\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52789489,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><recurrent-neural-network><preprocessor>","CreationDate":"2018-10-13T04:18:50.020","AcceptedAnswerId":null,"OwnerUserId":9947487.0,"Title":"How to map words to numbers for input into Tensorflow Neural Network","Body":"<p>I am trying to build a chatbot with a seq2seq neural network implementation with Tensorflow in Python. I've never done seq2seq before, and most of my research has been rather unhelpful. <\/p>\n\n<p>I'm not going to flat out ask for the code for a Sequence to Sequence chatbot. Instead, my question is how to best go about preparing a list of words as input to a network. I haven't got very far into the code, but I've made a script that will load training data from a file and tokenize it. <\/p>\n\n<p>However, obviously a Tensorflow neural network can't accept strings as input. I need to convert these strings into data the network knows how to deal with; namely, numbers.<\/p>\n\n<p>This is my code so far; hopefully the comments give you an idea of where I'm at:<\/p>\n\n<pre><code>#Import dependencies\nimport tensorflow as tf\n\n#Fetch and preprocess data\n\n#Define a tokenizer function\ndef tokenize(string):\n    tokenized_list = []\n    tmp_indx = 0\n    for i in range(len(string)):\n        if string[i] in \"?.,!;\":\n            tokenized_list.append(string[tmp_indx:i])\n            tokenized_list.append(string[i])\n            tmp_indx = i+1\n        elif string[i] == \" \":\n            tokenized_list.append(string[tmp_indx:i])\n            tmp_indx = i+1\n\n    #A quick and dirty way out :\/\n    tokenized_list = [x for x in tokenized_list if x!=\"\"]\n\n\n    return tokenized_list\n\n\nraw_file_data = \"\"\nwith open(\"training_dialogue.txt\") as file:\n    raw_file_data = file.read()\nraw_file_data = raw_file_data.split(\"\\n\")\n\n#Train data as list of values like so: [query, target_response]\ntrain_data = []\n\nfor i in range(0,len(raw_file_data)):\n    if i%2!=0:\n        #Perform the most basic tokenization algorithm\n        query = tokenize(raw_file_data[i-1])\n        target_response = tokenize(raw_file_data[i])\n        train_data.append([query, target_response])\n\n#Now that I have a list of tokens in the form of strings, I need to map these to numbers somehow\n\n\n#Load encoder and decoder networks\n\n\n\n#Define hyperparameters\n\n\n#Train them on the data\n<\/code><\/pre>\n\n<p>It would be great if somebody could show me how to process these words into numbers somehow. I also need to be able to turn them from numbers back to words.<\/p>\n","answers":[{"AnswerId":"52789570","CreationDate":"2018-10-13T04:39:03.963","ParentId":null,"OwnerUserId":"10457393","Title":null,"Body":"<p>I believe the best way would be to create a dictionary \/ index of words mapping to numbers. This would help in converting the numbers back to words as well. The same problem is discussed at <a href=\"https:\/\/stackoverflow.com\/questions\/2489595\/mapping-words-to-numbers-with-respect-to-definition\">this<\/a> thread as well in the NLP context.<\/p>\n\n<p>Code based on this suggestion -<\/p>\n\n<pre><code>wordList = []\nwordMap = {}\ndef getNumber(word):\n  if word in wordMap:\n    return wordMap[word];\n\n  wordIndex = len(wordList)\n  wordList.append(word)\n  wordMap[word] = wordIndex\n  return wordIndex\n\ndef getWord(number):\n  if number &lt; len(wordList):\n    return wordList[number]\n\n  raise Exception(\"Unable to find Word for [{}]\".format(number))\n\n# Main\nprint(getNumber('Hello'))\nprint(getNumber(\"World\"))\nprint(getWord(getNumber(\"World\")))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52789631,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-13T04:47:32.033","AcceptedAnswerId":null,"OwnerUserId":10497980.0,"Title":"\"blind\" Keras CNN , 100 % accuracy, no prediction power","Body":"<p>I am new to Keras and trying to build an image detection based on a dataset of images of all images (150 x 150) with and without marks.\nMy code is based on the Deep Learning book by Francois Chollet.\nJupyter NB, Keras  2.2.0, tensorflow<\/p>\n\n<p>I make two classes (\"clean wall\" and \"marked wall\"), to be able to show the activation (softmax instead of sigmoid), as described in the book showing the last conv layer activation, and to use more classes later.<\/p>\n\n<p>My problem is (with ~30000 clean images and ~40 marked), all 150x150, the model runs at once into 100% accuracy with no loss - but without identifying any marked image as marked if I predict one later. (the same model worked perfectly on 100 cat images, manually having marked same pictures with an additional mini-image as the second class, and resizing all from bigger pics to 150x150).<\/p>\n\n<p>total training images (1 and 0): 28980 and 47\ntotal validation images (1 and 0): 28980 and 6<\/p>\n\n<p>Prediction is always \"clean\" with a high percentage, even on all original marked images, changed marked images (via generator), anything.<\/p>\n\n<p>Any suggestions, please? Am I making a false assumption\/did I miss something somewhere?<\/p>\n\n<p>Thnx in advance!<\/p>\n\n<p>by Andy<\/p>\n\n<pre><code>from keras import layers\nfrom keras import models\nfrom keras import optimizers\n\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu',\n                    input_shape=(150, 150, 3)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Dense(512, activation='relu'))\nmodel.add(layers.Dense(2, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n          optimizer=optimizers.RMSprop(lr=1e-4),\n          metrics=['acc'])\n\n# Preprocess images\nfrom keras.preprocessing.image import ImageDataGenerator\n\n# All images will be rescaled by 1.\/255\ntrain_datagen = ImageDataGenerator(\n  rescale=1.\/255,\n  rotation_range=20, \n  width_shift_range=0.2, \n  height_shift_range=0.2, \n  shear_range=0.2, \n  zoom_range=0.2, \n  vertical_flip=True,\n  horizontal_flip=True,\n  fill_mode='reflect')\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    # This is the target directory\n    train_dir,\n    # All images will be resized to 150x150\n    target_size=(150, 150),\n    batch_size=16,\n    # Since we use binary_crossentropy loss, we need binary labels\n    class_mode='categorical')\n\nvalidation_generator = test_datagen.flow_from_directory(\n    validation_dir,\n    target_size=(150, 150),\n    batch_size=16,\n    class_mode='categorical')\n\n# train the model\n\nfor data_batch, labels_batch in train_generator:\n  print('data batch shape:', data_batch.shape)\n  print('labels batch shape:', labels_batch.shape)\n  break\n\nhistory = model.fit_generator(\n  train_generator,\n  steps_per_epoch=10,\n  epochs=5,\n  validation_data=validation_generator,\n  validation_steps=10)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>data batch shape: (16, 150, 150, 3)\nlabels batch shape: (16, 2)\nEpoch 1\/5\n5\/5 [==============================] - 17s 3s\/step - loss: 0.5646 - acc: 0.8625 - val_loss: 0.3196 - val_acc: 1.0000\nEpoch 2\/5\n5\/5 [==============================] - 18s 4s\/step - loss: 0.2201 - acc: 1.0000 - val_loss: 0.0823 - val_acc: 1.0000 ... \n<\/code><\/pre>\n","answers":[{"AnswerId":"52792719","CreationDate":"2018-10-13T12:02:57.070","ParentId":null,"OwnerUserId":"10497980","Title":null,"Body":"<p>seems, it starts to see. <em>g<\/em><\/p>\n\n<p>Main change: same number of negative pictures, more similar to positive ones.<\/p>\n\n<p>One more dense layer for complexity.<\/p>\n\n<p>Thnx to anyone having dropped a thought on the solution <em>g<\/em><\/p>\n\n<p>BR Andy<\/p>\n"}]}
+{"QuestionId":52790068,"AnswerCount":1,"Tags":"<python><keras><deep-learning><python-3.5><autoencoder>","CreationDate":"2018-10-13T06:13:33.310","AcceptedAnswerId":52791429.0,"OwnerUserId":9422346.0,"Title":"Value error with dimensions in designing a simple autoencoder","Body":"<p>Hi I am trying out a simple autoencoder in Python 3.5 using Keras library. The issue I face is - ValueError: Error when checking input: expected input_40 to have 2 dimensions, but got array with shape (32, 256, 256, 3). My dataset is very small (60 RGB images with dimension - 256*256 and one same type of image to validate). I am a bit new to Python. Please help.<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\n\n#Declaring the model\nencoding_dim = 32\ninput_img = Input(shape=(65536,))\nencoded = Dense(encoding_dim, activation='relu')(input_img)\ndecoded = Dense(65536, activation='sigmoid')(encoded)\nautoencoder = Model(input_img, decoded)\nencoder = Model(input_img, encoded)\nencoded_input = Input(shape=(encoding_dim,))\ndecoder_layer = autoencoder.layers[-1]\ndecoder = Model(encoded_input, decoder_layer(encoded_input))\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\n\n\n#Constructing a data generator iterator\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\nshear_range = 0.2,\nzoom_range = 0.2,\nhorizontal_flip = True)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\ntraining_set=\ntrain_datagen.flow_from_directory('C:\\\\Users\\\\vlsi\\\\Desktop\\\\train',\nbatch_size = 32,\nclass_mode = 'binary')\ntest_set =     \ntest_datagen.flow_from_directory('C:\\\\Users\\\\vlsi\\\\Desktop\\\\validation',\nbatch_size = 32,\nclass_mode = 'binary')\n\n\n#fitting data\nautoencoder.fit_generator(training_set,\nsteps_per_epoch = 80,\nepochs = 25,\nvalidation_data = test_set,\nvalidation_steps = 20)\n\nimport numpy as np from keras.preprocessing import image\ntest_image =            \nimage.load_img('C:\\\\Users\\\\vlsi\\\\Desktop\\\\validation\\\\validate\\\\apple1.jpg')\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis = 0)\n\n#Displaying output\nencoded_imgs = encoder.predict(test_image)\ndecoded_imgs = decoder.predict(encoded_imgs)\nplt.imshow(decoded_imgs)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52791429","CreationDate":"2018-10-13T09:19:24.520","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The problem is here:<\/p>\n\n<pre><code>input_img = Input(shape=(65536,))\n<\/code><\/pre>\n\n<p>You told Keras the input to the network will have 65K dimensions, meaning a vector of shape (samples, 65536), but your actual inputs have shape(samples, 256, 256, 3). Any easy solution would be to use the real input shape and for the network to perform the necessary reshaping:<\/p>\n\n<pre><code>input_img = Input(shape=((256, 256, 3))\nflattened = Flatten()(input_img)\nencoded = Dense(encoding_dim, activation='relu')(flattened)\ndecoded = Dense(256 * 256 * 3, activation='sigmoid')(encoded)\ndecoded = Reshape((256, 256, 3))(decoded)\nautoencoder = Model(input_img, decoded)\nencoder = Model(input_img, encoded)\nencoded_input = Input(shape=(encoding_dim,))\ndecoder_layer = autoencoder.layers[-1]\ndecoder = Model(encoded_input, decoder_layer(encoded_input))\n<\/code><\/pre>\n\n<p>Note that I added a Flatten and a Reshape layer first to flatten the image, and then to take the flattened image back to the shape (256, 256, 3).<\/p>\n"}]}
+{"QuestionId":52790175,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><computer-vision><regression><caffe>","CreationDate":"2018-10-13T06:29:44.697","AcceptedAnswerId":52792953.0,"OwnerUserId":5539391.0,"Title":"Caffe: Check failed: outer_num_ * inner_num_ == bottom[1]->count() (10 vs. 60) Number of labels must match number of predictions","Body":"<p>I am trying to fine tune Alexnet for a multi-label regression task. For this  I have replaced the last layer producing 1000-label output (for image classification task) to 6 label output which provides me with 6 floats. I replaced the last layers <a href=\"https:\/\/gist.github.com\/denoza\/1683e16828b16449fd3919ba5f9873ef\" rel=\"nofollow noreferrer\">as mentioned here<\/a>.<\/p>\n\n<p>My training data is prepared in <code>h5<\/code> format and is shaped as (11000, 3, 544, 1024) for <code>data<\/code> and (11000, 1, 6) for <code>labels<\/code>. While retraining the weights of Alexnet in Caffe library, I get the following error:<\/p>\n\n<pre><code>I1013 10:50:49.759560  3107 net.cpp:139] Memory required for data: 950676640\n\nI1013 10:50:49.759562  3107 layer_factory.hpp:77] Creating layer accuracy_retrain\n\nI1013 10:50:49.759567  3107 net.cpp:86] Creating Layer accuracy_retrain\n\nI1013 10:50:49.759568  3107 net.cpp:408] accuracy_retrain &lt;- fc8_fc8_retrain_0_split_0\n\nI1013 10:50:49.759572  3107 net.cpp:408] accuracy_retrain &lt;- label_data_1_split_0\n\nI1013 10:50:49.759575  3107 net.cpp:382] accuracy_retrain -&gt; accuracy\n\nF1013 10:50:49.759587  3107 accuracy_layer.cpp:31] Check failed: outer_num_ * inner_num_ == bottom[1]-&gt;count() (10 vs. 60) Number of labels must match number of predictions; e.g., if label axis == 1 and prediction shape is (N, C, H, W), label count (number of labels) must be N*H*W, with integer values in {0, 1, ..., C-1}.\n<\/code><\/pre>\n\n<p>My Batchsize for both training and testing phases is <code>10<\/code>. The error arises in the testing phase, possibly in the <code>accuracy<\/code> layer <a href=\"https:\/\/gist.github.com\/denoza\/35e8904986451171fc8c5bf3817ab76e\" rel=\"nofollow noreferrer\">Complete Error Log here<\/a>. I am not sure why this problem arises, might be my <code>label<\/code> is misshaped. Any help in this regard will be highly appreciated.<\/p>\n","answers":[{"AnswerId":"52792953","CreationDate":"2018-10-13T12:31:23.837","ParentId":null,"OwnerUserId":"5539391","Title":null,"Body":"<p>I solved this problem. Seems like the <code>accuracy<\/code> layer is only used for classification tasks along with <code>SoftmaxWithLoss<\/code> layer. As stated in <a href=\"https:\/\/stackoverflow.com\/questions\/39756886\/testing-a-regression-network-in-caffe\">this answer<\/a>, the <code>EuclideanLoss<\/code> can be used to test the regression network.<\/p>\n"}]}
+{"QuestionId":52790444,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-13T07:10:58.690","AcceptedAnswerId":52816765.0,"OwnerUserId":9292428.0,"Title":"How to create two graphs for train and validation?","Body":"<p>When I read tensorflow guidance about graph and session(<a href=\"https:\/\/www.tensorflow.org\/guide\/graphs#programming_with_multiple_graphs\" rel=\"nofollow noreferrer\">Graphs and Sessions<\/a>), I found they suggest to create two graphs for train and validation. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9YJAP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9YJAP.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I think this reasonable and I want to use this because my train and validation models are different (for encoder-decoder mode or dropout). However, i don't know how to make variables in trained graph available for test graph without using tf.saver().<br>\nWhen I create two graphs and create variables inside each graph, I found these two variables are totally different as they belong to different graphs. \nI have googled a lot and I know there are questions about this problems, such as <a href=\"https:\/\/stackoverflow.com\/questions\/49155206\/why-should-i-build-separated-graph-for-training-and-validation-in-tensorflow\">question1<\/a>. But there is still no useful answer. If there is any code example or anyone know how to create two graphs for train and validation separately, such as:<\/p>\n\n<pre><code>def train_model():\n    g_train = tf.graph()\n    with g_train.as_default():\n        train_models\ndef validation_model():\n    g_test = tf.graph()\n    with g_test.as_default():\n         test_models\n<\/code><\/pre>\n","answers":[{"AnswerId":"52816765","CreationDate":"2018-10-15T12:26:12.550","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>One easy way of doing that is to create a 'forward function' that defines the model and change behaviour based on extra parameters.<\/p>\n\n<p>Here is an example:<\/p>\n\n<pre><code>def forward_pass(x, is_training, reuse=tf.AUTO_REUSE, name='model_forward_pass'):\n\n  # Note the reuse attribute as it tells the getter to either create the graph or get the weights\n\n  with tf.variable_scope(name=name, reuse=reuse):\n     x = tf.layers.conv(x, ...)\n     ...\n     x = tf.layers.dense(x, ...)\n     x = tf.layers.dropout(x, rate, training=is_training) # Note the is_training attribute\n\n     ...\n     return x\n<\/code><\/pre>\n\n<p>Now you can call the 'forward_pass' function anywhere in your code. You simply need to provide the is_training attribute to use the correct mode for dropout for example. The 'reuse' argument will automatically get the correct values for your weights as long as the 'name' of the 'variable_scope' is the same.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>train_logits_model1 = forward_pass(x_train, is_training=True, name='model1')\n# Graph is defined and dropout is used in training mode\n\ntest_logits_model1 = forward_pass(x_test, is_training=False, name='model1')\n# Graph is reused but the dropout behaviour change to inference mode\n\ntrain_logits_model2 = forward_pass(x_train2, is_training=True, name='model2')\n# Name changed, model2 is added to the graph and dropout is used in training mode\n<\/code><\/pre>\n\n<p>To add to this answer as you stated that you want to have 2 separated graph, you could to that using an assign function:<\/p>\n\n<pre><code>train_graph = forward_pass(x, is_training=True, reuse=False, name='train_graph')\n\n...\ntest_graph = forward_pass(x, is_training=False, reuse=False, name='test_graph')\n\n...\ntrain_vars = tf.get_collection('variables', 'train_graph\/.*')\ntest_vars = tf.get_collection('variables','test_graph\/.*')\ntest_assign_ops = []\nfor test, train in zip(test_vars, train_vars):\n    test_assign_ops += [tf.assign(test, train)]\nassign_op = tf.group(*test_assign_ops)\n\nsess.run(assign_op) # Replace vars in the test_graph by the one in train_graph\n<\/code><\/pre>\n\n<p>I'm a big advocate of method 1 as it is way cleaner and reduce memory usage.<\/p>\n"}]}
+{"QuestionId":52790720,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-13T07:47:09.857","AcceptedAnswerId":null,"OwnerUserId":9250190.0,"Title":"Tensorflow-Custom Object Detection Api-Helmet Detection","Body":"<p>I am trying to detect person not wearing helmet using Tensorflow Object Detection Api. I am using 3000 images(1000 for persons not wearing helmet,1000 for persons wearing helmet,1000 for persons).I used SSD_mobilenet_v1_pets.config for training.\nplease find the PBTXT file:<\/p>\n\n<pre><code>item {\n    id: 1\n    name: 'withouthelmet'\n}\nitem {\n    id: 2\n    name: 'withhelmet'\n}\nitem {\n    id: 3\n    name: 'person'\n}\n<\/code><\/pre>\n\n<p>In SSD_mobilenet_v1_pets.config file,(id:1 is for person but i have set id:1 to 'withouthelmet',id:2 to 'withhelmet',id:3 to 'person').<\/p>\n\n<p>After training the model,it detects every person as'without helmet' and every car as 'person'.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/djaVh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/djaVh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xPYYJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xPYYJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Can you please let me know where the issue is.<\/p>\n","answers":[]}
+{"QuestionId":52790775,"AnswerCount":2,"Tags":"<python><conv-neural-network><pytorch><torch>","CreationDate":"2018-10-13T07:53:07.063","AcceptedAnswerId":52793614.0,"OwnerUserId":4982729.0,"Title":"Setting custom kernel for CNN in pytorch","Body":"<p>Is there a way to specify our own custom kernel values for a convolution neural network in pytorch? Something like <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/tf\/layers\/conv2d_transpose\" rel=\"nofollow noreferrer\">kernel_initialiser<\/a> in tensorflow? Eg. I want a 3x3 kernel in <code>nn.Conv2d<\/code> with initialization so that it acts as a identity kernel -<\/p>\n\n<p><code>0 0 0<\/code><br>\n<code>0 1 0<\/code><br>\n<code>0 0 0<\/code>  <\/p>\n\n<p>(this will effectively return the same output as my input in the very first iteration)<\/p>\n\n<p>My non-exhaustive research on the subject - <\/p>\n\n<p>I could use <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/init.html\" rel=\"nofollow noreferrer\">nn.init<\/a> but it only has some pre-defined kernel initialisaition values.<\/p>\n\n<p>I tried to follow the discussion on their <a href=\"https:\/\/discuss.pytorch.org\/t\/weight-initilzation\/157\/2\" rel=\"nofollow noreferrer\">official thread<\/a> but it doesn't suit my needs.<\/p>\n\n<p>I might have missed something in my research please feel free to point out.<\/p>\n","answers":[{"AnswerId":"54753586","CreationDate":"2019-02-18T18:47:12.443","ParentId":null,"OwnerUserId":"8327356","Title":null,"Body":"<p>I think an easier solution is to :<\/p>\n\n<pre><code>    deconv = nn.ConvTranspose2d(\n        in_channels=channel_dim, out_channels=channel_dim,\n        kernel_size=kernel_size, stride=stride,\n        bias=False, padding=1, output_padding=1\n    )\n    deconv.weight.data.copy_(\n        get_upsampling_weight(channel_dim, channel_dim, kernel_size)\n    )\n<\/code><\/pre>\n\n<p>in other words use copy_<\/p>\n"},{"AnswerId":"52793614","CreationDate":"2018-10-13T13:48:54.643","ParentId":null,"OwnerUserId":"4982729","Title":null,"Body":"<p>Thanks to <a href=\"https:\/\/discuss.pytorch.org\/t\/setting-custom-kernel-for-cnn-in-pytorch\/27176\/4?u=markroxor\" rel=\"nofollow noreferrer\">ptrblck<\/a> I was able to solve it.\nI can define a new convolution layer as <code>conv<\/code> and as per the example I can set the identity kernel using - <\/p>\n\n<pre><code>weights = ch.Tensor([[0, 0, 0], [0, 1, 0], [0, 0, 0]]).unsqueeze(0).unsqueeze(0)\nweights.requires_grad = True\n\n\nconv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False)\n\nwith ch.no_grad():\n    conv.weight = nn.Parameter(weights)\n<\/code><\/pre>\n\n<p>I can then continue to use <code>conv<\/code> as my regular <code>nn.Conv2d<\/code> layer.<\/p>\n"}]}
+{"QuestionId":52791296,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2018-10-13T09:02:18.527","AcceptedAnswerId":55640850.0,"OwnerUserId":10498304.0,"Title":"with tf.device() does not work in TensorFlow","Body":"<p>I have two GPUs and want to try some distributed training(model-parallelism) in TensorFlow.<\/p>\n\n<p>The two GPUs are:<\/p>\n\n<pre><code>\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: TITAN Xp COLLECTORS EDITION, pci bus id: 0000:04:00.0, compute capability: 6.1\n\/job:localhost\/replica:0\/task:0\/device:GPU:1 -&gt; device: 1, name: TITAN X (Pascal), pci bus id: 0000:82:00.0, compute capability: 6.1\n<\/code><\/pre>\n\n<p>My plan is to divide LeNet into two parts, assign each part to one GPU. <\/p>\n\n<p>LeNet has 5 layers, I use <code>with tf.device('\/gpu:0'):<\/code>  to assign layer 1 to GPU 0, <code>with tf.device('\/gpu:1'):<\/code>to assign layer2-layer5 to GPU 1.<\/p>\n\n<p>I know there is no need to do model-parallelism in this model, but I just want to try model-parallelism in small models. <\/p>\n\n<p>The log for device mapping shows that, all the ops have been assigned to the device as I wish:<\/p>\n\n<pre><code>layer5\/fc3_b: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/fc3_b\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/fc3_b\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/fc3_w: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/fc3_w\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/truncated_normal\/TruncatedNormal: (TruncatedNormal): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/truncated_normal\/mul: (Mul): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/truncated_normal: (Add): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer5\/fc3_w\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_b: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_b\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_b\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_w: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_w\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/truncated_normal\/TruncatedNormal: (TruncatedNormal): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/truncated_normal\/mul: (Mul): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/truncated_normal: (Add): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer4\/fc2_w\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_b: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_b\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_b\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_w: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_w\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/truncated_normal\/TruncatedNormal: (TruncatedNormal): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/truncated_normal\/mul: (Mul): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/truncated_normal: (Add): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer3\/fc1_w\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_b: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_b\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_b\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_w: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_w\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/truncated_normal\/TruncatedNormal: (TruncatedNormal): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/truncated_normal\/mul: (Mul): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/truncated_normal: (Add): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer2\/conv2_w\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:1\ninit\/NoOp_1: (NoOp): \/job:localhost\/replica:0\/task:0\/device:GPU:1\nlayer1\/conv1_b: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/conv1_b\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/conv1_b\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/conv1_w: (VariableV2): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/conv1_w\/read: (Identity): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/truncated_normal\/TruncatedNormal: (TruncatedNormal): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/truncated_normal\/mul: (Mul): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/truncated_normal: (Add): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nlayer1\/conv1_w\/Assign: (Assign): \/job:localhost\/replica:0\/task:0\/device:GPU:0\n<\/code><\/pre>\n\n<p>But I got a different result in <code>timeline.json<\/code> as shown in the figure below.\n<a href=\"https:\/\/i.stack.imgur.com\/iNlvu.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iNlvu.jpg\" alt=\"timeline.json\"><\/a><\/p>\n\n<p>The timeline shows that, it seems ops of layer2-layer5 are launched in GPU1 but run on GPU0. I don't think this is what I want by using <code>with tf.device('\/gpu:1'):<\/code>. <\/p>\n\n<p>Is this expected in TensorFlow? <\/p>\n\n<p>This is the first time I asked question on stack overflow, if any other information is needed, please inform me, thanks.<\/p>\n","answers":[{"AnswerId":"55640850","CreationDate":"2019-04-11T20:42:35.317","ParentId":null,"OwnerUserId":"9296815","Title":null,"Body":"<p>This is just an artifact of Chrome Trace Event Format.<\/p>\n\n<p>Stream, <code>\"\/job:localhost\/replica:0\/task:0\/devuce:GPU:0 Compute\"<\/code> shows the time to launch\/queue CUDA kernels for ops being executed on GPU:0.<\/p>\n\n<p>Stream, <code>\"\/job:localhost\/replica:0\/task:0\/devuce:GPU:1 Compute\"<\/code> shows the time to launch\/queue CUDA kernels for ops being executed on GPU:1.<\/p>\n\n<p>All streams matching, <code>\"\/device:GPU:0\/stream.*<\/code> Compute\" shows the time to actually execute Operations on <strong>all<\/strong> the GPUs. To find out, on which GPU the operation is actually executed, you need to look in the streams <code>\"\/job:localhost\/replica:0\/task:0\/devuce:GPU:.* Compute\"<\/code><\/p>\n\n<p>Hope this answers your question<\/p>\n"}]}
+{"QuestionId":52791322,"AnswerCount":2,"Tags":"<tensorflow><sliding-window><tfrecord>","CreationDate":"2018-10-13T09:05:58.280","AcceptedAnswerId":52824659.0,"OwnerUserId":1171204.0,"Title":"tensorflow - Input pipeline with multiple TFRecord files + tf.contrib.data.sliding_window_batch()","Body":"<p>I have multiple <code>TFRecord<\/code> files, which all hold a specific timeframe of my data. Containing data points are consecutive inside each file but are not consecutive across files. As part of my input pipeline, i am using <code>tf.contrib.data.sliding_window_batch<\/code> in order to process a window of data points as following:<\/p>\n\n<pre><code>filenames = [os.path.join(data_dir, f) for f in os.listdir(data_dir)]\ndataset = tf.data.TFRecordDataset(filenames)\n\ndataset = dataset.map(parser_fn, num_parallel_calls=6)\ndataset = dataset.map(preprocessing_fn, num_parallel_calls=6)\ndataset = dataset.apply(tf.contrib.data.sliding_window_batch(window_size=y + z)) # sliding window\ndataset = dataset.map(lambda x: prepare_fn(x, y, z))\ndataset = dataset.shuffle(buffer_size=100000)\ndataset = dataset.batch(32)\ndataset = dataset.repeat()\ndataset = dataset.prefetch(2)\n<\/code><\/pre>\n\n<p>How can i prevent the window to be spanned across data points from different files? <\/p>\n","answers":[{"AnswerId":"52795874","CreationDate":"2018-10-13T18:11:12.083","ParentId":null,"OwnerUserId":"1171204","Title":null,"Body":"<p>Solved it by using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#filter\" rel=\"nofollow noreferrer\"><code>tf.Dataset.filter(predicate)<\/code><\/a>.<\/p>\n\n<pre><code>filenames = [os.path.join(data_dir, f) for f in os.listdir(data_dir)]\ndataset = tf.data.TFRecordDataset(filenames)\n\ndataset = dataset.map(parser_fn, num_parallel_calls=6)\ndataset = dataset.map(preprocessing_fn, num_parallel_calls=6)\ndataset = dataset.apply(tf.contrib.data.sliding_window_batch(window_size=y + z)) # sliding window\ndataset = dataset.filter(lambda x: tf.equal(x['timeframe'][0], x['timeframe'][-1]))\ndataset = dataset.map(lambda x: prepare_fn(x, y, z))\ndataset = dataset.shuffle(buffer_size=100000)\ndataset = dataset.batch(32)\ndataset = dataset.repeat()\ndataset = dataset.prefetch(2)\n<\/code><\/pre>\n"},{"AnswerId":"52824659","CreationDate":"2018-10-15T20:56:32.740","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>An alternative would be to create the batches on each file independently, and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#interleave\" rel=\"nofollow noreferrer\">interleave<\/a> the results:<\/p>\n\n<pre><code>def interleave_fn(filename):\n  dataset = dataset.map(parser_fn, num_parallel_calls=6)\n  dataset = dataset.map(preprocessing_fn, num_parallel_calls=6)\n  dataset = dataset.apply(tf.contrib.data.sliding_window_batch(window_size=y + z)) # sliding window\n\nfilenames = [os.path.join(data_dir, f) for f in os.listdir(data_dir)]\ndataset = tf.data.Dataset.from_tensor_slices(filenames)\ndataset = dataset.interleave(interleave_fn, num_parallel_calls=...)\ndataset = dataset.map(lambda x: prepare_fn(x, y, z))\ndataset = dataset.shuffle(buffer_size=1000000)\ndataset = dataset.batch(32)\ndataset = dataset.repeat()\ndataset = dataset.prefetch(2)\n<\/code><\/pre>\n\n<p>This is probably more performant as it bypasses the filter step.<\/p>\n"}]}
+{"QuestionId":52791367,"AnswerCount":1,"Tags":"<python><django><tensorflow><machine-learning>","CreationDate":"2018-10-13T09:10:55.397","AcceptedAnswerId":null,"OwnerUserId":10235664.0,"Title":"Value Error Tensorflow While Loading the Machine Learning project in Django","Body":"<p>While loading our machine learning project into a Django server we got following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/django\/core\/handlers\/exception.py\",\n  line 34, in inner\n      response = get_response(request)   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/django\/core\/handlers\/base.py\",\n  line 126, in _get_response\n      response = self.process_exception_by_middleware(e, request)   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/django\/core\/handlers\/base.py\",\n  line 124, in _get_response\n      response = wrapped_callback(request, *callback_args, **callback_kwargs)   File \"\/home\/akhil\/tocoblo\/msg\/views.py\", line 6, in index\n      a=fuctioncall.show()   File \"\/home\/akhil\/tocoblo\/msg\/fuctioncall.py\", line 6, in show\n      a=Loadmodel.predict_string()   File \"\/home\/akhil\/tocoblo\/msg\/Loadmodel.py\", line 69, in predict_string\n      b=loaded_model.predict(y)   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\",\n  line 1164, in predict\n      self._make_predict_function()   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\",\n  line 554, in _make_predict_function\n      **kwargs)   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\",\n  line 2744, in function\n      return Function(inputs, outputs, updates=updates, **kwargs)   File \"\/home\/akhil\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\",\n  line 2546, in <strong>init<\/strong>\n      with tf.control_dependencies(self.outputs):   File \"\/home\/akhil\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 5002, in control_dependencies\n      return get_default_graph().control_dependencies(control_inputs)   File\n  \"\/home\/akhil\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 4541, in control_dependencies\n      c = self.as_graph_element(c)   File \"\/home\/akhil\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 3488, in as_graph_element\n      return self._as_graph_element_locked(obj, allow_tensor, allow_operation)   File\n  \"\/home\/akhil\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 3567, in _as_graph_element_locked\n      raise ValueError(\"Tensor %s is not an element of this graph.\" % obj) ValueError: Tensor Tensor(\"dense_4\/Sigmoid:0\", shape=(?, 6),\n  dtype=float32) is not an element of this graph.<\/p>\n<\/blockquote>\n\n<p>The Loaded code is Loadmodel.py:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport os\nimport sys  \nimport gzip\nimport keras\nimport sys\nimport pickle\nfrom sklearn.model_selection import train_test_split\nfrom keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras.layers import Dense, Input, LSTM, Embedding, Dropout, Activation\nfrom keras.layers import Bidirectional, GlobalMaxPool1D\nfrom keras.models import Model, Sequential\nfrom keras import initializers, regularizers, constraints, optimizers, layers\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.models import model_from_json\n\nimport numpy\nimport os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n\nimport json\nfrom pprint import pprint\n\njson_file = open('msg\/model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n\nloaded_model.load_weights(\"msg\/model.h5\")\nprint(\"Loaded model from disk\")\n\npickle_in = open(\"msg\/dict.pickle\",\"rb\")\n#pickle_in.encoding = 'latin1'\ntokenizer = pickle.load(pickle_in, encoding='latin1') \n#tokenizer = pickle.load(pickle_in)\n\nwith open('msg\/data.json') as f:\n    data = json.load(f)\n\ndef predict_string():    \n    maxlen=200\n    string=\"\"\n    for j in range(0,120):\n        flag=0\n        s=(data[\"maps\"][j][\"comment\"],)\n        x=tokenizer.texts_to_sequences(s)\n        y=pad_sequences(x,maxlen=maxlen)\n\n        b=loaded_model.predict(y)\n        for i in range(0,6):\n            if(b[0][i]&gt;=0.3):\n                flag=1\n\n        cnt=0\n        if(flag==1):\n            for i in range(0,6):\n                if(b[0][i]&gt;0.3):\n                    cnt=cnt+1\n\n        flag=cnt\n        string=string+str(flag)\n\n    return string\n`\nfuctioncall.py\n\n    from . import Loadmodel\nfrom django.http import HttpResponse, JsonResponse\n\n\ndef show():\n    a=Loadmodel.predict_string()\n    return (\"GOT\"+a);\n<\/code><\/pre>\n\n<p><strong>urls.py:<\/strong><\/p>\n\n<pre><code>from django.urls import path\nfrom . import views\nfrom . import fuctioncall\nurlpatterns = [\n    path('', views.index, name='index'),\n    path('&lt;str:com&gt;', views.com, name='com'),\n]\n<\/code><\/pre>\n\n<p>How do I solve this error? Also how do I load the machine learning project in the Django server and call it?<\/p>\n","answers":[{"AnswerId":"56524944","CreationDate":"2019-06-10T10:48:44.853","ParentId":null,"OwnerUserId":"8896060","Title":null,"Body":"<p>Just after the imports, add the following 2 lines:<\/p>\n\n<pre><code>global graph\ngraph = tf.get_default_graph()\n<\/code><\/pre>\n\n<p>Then whenever you try to run inference on the model, use:<\/p>\n\n<pre><code>with graph.as_default():\n        prediction = mode.predict(...)\n<\/code><\/pre>\n\n<p>Hope it helps :)<\/p>\n"}]}
+{"QuestionId":52791415,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-13T09:17:26.320","AcceptedAnswerId":52791493.0,"OwnerUserId":7509978.0,"Title":"In Tensorflow How can I add a Tensor to another Tensor with different shape?","Body":"<p>I use Tensorflow.\nI want to add a tensor A whose shape is [64,64] (=[Batch size,values])\nto another tensor B whose shape is [64,7,7,64].\nI reshaped the tensor A, but it should have same number of elements as tensor B.\nSo, how can I reshape or expand tensor A. Or is there any other way to add A to B? \nSpecifically, I want to add 64 values of A to all 64 values of B 7*7 times.\nI am sorry to my poor English. I cannot explain well but want some of you to understand what I want to say. Thank you.<\/p>\n","answers":[{"AnswerId":"52791493","CreationDate":"2018-10-13T09:28:38.853","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/broadcasting\" rel=\"nofollow noreferrer\">broadcasting<\/a>. Here you have an example:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nA = tf.constant(np.arange(64*64), shape=(64, 64), dtype=tf.int32)\nB = tf.ones(shape=(64, 7, 7, 64), dtype=tf.int32)\nA_ = A[:, None, None, :]  # Shape=(64, 1, 1, 64)\nresult = A_ + B\n\nwith tf.Session() as sess:\n    print(sess.run(result))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52791531,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-13T09:33:20.977","AcceptedAnswerId":52817320.0,"OwnerUserId":9608800.0,"Title":"Forward function in tensorflow","Body":"<p>I'm learning about neural network and reading a piece of code to train the neural network with mnist data set.\nIn the test part of it, it defines the output variable: <code>y = mnist_forward.forward(x, None)<\/code> in the test function: <\/p>\n\n<pre><code>def test(mnist):\n    with tf.Graph().as_default() as g:\n        x = tf.placeholder(tf.float32, [None, mnist_forward.INPUT_NODE])\n        y_ = tf.placeholder(tf.float32, [None, mnist_forward.OUTPUT_NODE])\n        y = mnist_forward.forward(x, None)\n\n        saver = tf.train.Saver()\n        correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\n        while True:\n            with tf.Session() as sess:\n                tf.initialize_all_variables().run()\n                ckpt = tf.train.get_checkpoint_state(mnist_backward.MODEL_SAVE_PATH)\n                if ckpt and ckpt.model_checkpoint_path:\n                    saver.restore(sess, ckpt.model_checkpoint_path)\n                    global_step = ckpt.model_checkpoint_path.split('\/')[-1].split('-')[-1]\n                    accuracy_score = sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels})\n                    print(\"After %s train steps, test accuracy = %g\" % (global_step, accuracy_score))\n                else:\n                    print('No checkpoint file found')\n                    return\n            time.sleep(TEST_INTERVAL_SECS)\n<\/code><\/pre>\n\n<p>However, the function <code>mnist_forward.forward<\/code> is defined as:<\/p>\n\n<pre><code>def forward(x, regularizer):\n    w1 = get_weight([INPUT_NODE, LAYER1_NODE], regularizer)\n    b1 = get_bias([LAYER1_NODE])\n    y1 = tf.nn.relu(tf.matmul(x, w1) + b1)\n\n    w2 = get_weight([LAYER1_NODE, OUTPUT_NODE], regularizer)\n    b2 = get_bias([OUTPUT_NODE])\n    y = tf.matmul(y1, w2) + b2\n    return y\n<\/code><\/pre>\n\n<p>The function <code>get_weight<\/code> and <code>get_bias<\/code> are used to generate the parameters randomly. It seems that each time the forward function is called(to predict), the parameters will be regenerated.<br><br>\nI just cannot understand why the function call to generate parameters should be written inside the forward function(rather than outside) and how it truly works in the prediction process of the model?<\/p>\n","answers":[{"AnswerId":"52817320","CreationDate":"2018-10-15T12:56:05.247","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<blockquote>\n  <p>It seems that each time the forward function is called(to predict), the parameters will be regenerated.<\/p>\n<\/blockquote>\n\n<p>That is the big difference between Tensorflow and standard programming and can be quite puzzling at first: Tensorflow is graph based, so in this case the 'forward' function is only called ONCE to define the graph. <\/p>\n\n<p>Then Tensorflow automatically figure out which node in the graph are needed to compute some other nodes, but it NEVER call the forward function again. <\/p>\n\n<p>The important lines in your code to really understand what's happening inside are the following ones:<\/p>\n\n<pre><code># Define the graph: create all variables, how to initialize them\n# and how to compute the output depending on the input\n#\u00a0CAUTION: VARIABLE TENSORS CURRENTLY HOLD NO VALUES THEY ARE ONLY DEFINED\ny = mnist_forward.forward(x, None)\n\n...\n\n# Add operations to the graph to compute the accuracy metric\ncorrect_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\n...\n\n# Initialize all variables using the initializers defined previously\ntf.initialize_all_variables().run()\n\n...\n\n# Restore all variables to saved values\n# CAUTION: REPLACES ALL INITIALIZED VALUES TO THESE STORED ONES\nsaver.restore(sess, ckpt.model_checkpoint_path)\n\n...\n\n# Runs the graph\naccuracy_score = sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels})\n<\/code><\/pre>\n\n<blockquote>\n  <p>I just cannot understand why the function call to generate parameters ...<\/p>\n<\/blockquote>\n\n<p>As you can see on the comments, what generates parameters is the 'tf.initialize_all_variables().run()' call, the 'forward' function simply defines HOW to initialize. <\/p>\n"}]}
+{"QuestionId":52791686,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><lstm>","CreationDate":"2018-10-13T09:52:38.390","AcceptedAnswerId":52801262.0,"OwnerUserId":10278110.0,"Title":"Get output from last layer in each epoch in LSTM, Keras","Body":"<p>We are already aware that output of each layer can be obtained from following code: <\/p>\n\n<pre><code>def get_layer(model,x):\n    from keras import backend as K\n\n    get_3rd_layer_output = K.function([model.layers[0].input],\n                                      [model.layers[2].output])\n    layer_output = get_3rd_layer_output([x])[0]\n    print(layer_output.shape)\n    return layer_output\n<\/code><\/pre>\n\n<p>LSTM model is fit as:<\/p>\n\n<pre><code>history = model.fit(X_train, y_train, batch_size=batch_size,verbose=1, nb_epoch=10,validation_data=(X_test,y_test))\n<\/code><\/pre>\n\n<p>But how to obtain output of the last layer in the model for each epoch out of 10 epochs?<\/p>\n","answers":[{"AnswerId":"52801262","CreationDate":"2018-10-14T09:25:19.907","ParentId":null,"OwnerUserId":"4995541","Title":null,"Body":"<p>You can create a custom callback (<a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">see Documentation<\/a>) and pass it to the callback list in the <code>fit<\/code> method.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>from keras.callbacks import Callback\n\nclass LogThirdLayerOutput(Callback):\n    def on_epoch_end(self, epoch, logs=None):\n        layer_output = get_3rd_layer_output(self.validation_data)[0]\n        print(layer_output.shape)\n<\/code><\/pre>\n\n<p>And when fitting the model:<\/p>\n\n<pre><code>history = model.fit(X_train, y_train, batch_size=batch_size, verbose=1, nb_epoch=10, validation_data=(X_test,y_test), callbacks=[LogThirdLayerOutput()])\n<\/code><\/pre>\n\n<p>If you don't want to create the callback in a separate class, you should also be able to use <a href=\"https:\/\/keras.io\/callbacks\/#lambdacallback\" rel=\"nofollow noreferrer\">Lambda callbacks<\/a>.<\/p>\n"}]}
+{"QuestionId":52792579,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-10-13T11:46:48.937","AcceptedAnswerId":null,"OwnerUserId":7844769.0,"Title":"How to modify the protocol of matcher in faster-RCNN","Body":"<p>I am currently using Faster-RCNN from the Tensorflow object detection API to train on my own data.<\/p>\n\n<p>I would like to change the positive IoU value and negative IoU value of \"<strong>argmax_matcher<\/strong>\". What fields should I modify in the .config file?<\/p>\n\n<p>Many thanks! <\/p>\n","answers":[{"AnswerId":"52802649","CreationDate":"2018-10-14T12:29:20.617","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Oddly, argmax_matcher is not a valid configuration for Faster R-CNN, but only of SSD.\nInstead, you can change the values that are given to the function call in function <code>create_target_assigner()<\/code> of \/core\/target_assigner.py. \nYou can change both the first stage:<\/p>\n\n<p><code>reference == 'FasterRCNN' and stage == 'proposal'<\/code><\/p>\n\n<p>and the second:<\/p>\n\n<p><code>reference == 'FasterRCNN' and stage == 'detection'<\/code>.<\/p>\n\n<p>The relevant parameters are of course <code>matched_threshold<\/code> and <code>unmatched_threshold<\/code>. You should probably set <code>force_match_for_each_row<\/code> to <code>True<\/code>, and note <code>negatives_lower_than_unmatched<\/code> is <code>True<\/code> by default.<\/p>\n"}]}
+{"QuestionId":52792714,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-13T12:02:14.590","AcceptedAnswerId":null,"OwnerUserId":10499111.0,"Title":"Tensorflow \"Hello World\" example does not work in PyCharm","Body":"<p>I entered these code in PyCharm and included Tensorflow in interpreter:<\/p>\n\n<pre><code>import tensorflow\nhello = tf.constant('Hello, TensorFlow!')\nsess = tf.Session()\nprint(sess.run(hello))\n<\/code><\/pre>\n\n<p>However the terminal didn't print anything and output was \"Process finished with exit code -1073741819 (0xC0000005)\"<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Pw5hf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Pw5hf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6sSYD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6sSYD.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52792769","CreationDate":"2018-10-13T12:07:54.957","ParentId":null,"OwnerUserId":"6278661","Title":null,"Body":"<p>You are refering to <code>tf<\/code> in your code, therefore you have to <code>import tensorflow as tf<\/code>. <\/p>\n\n<p>In addition you should rename your filename. For example change <code>tf.py<\/code> to <code>main.py<\/code>.<\/p>\n"},{"AnswerId":"52792746","CreationDate":"2018-10-13T12:05:30.033","ParentId":null,"OwnerUserId":"7602551","Title":null,"Body":"<p>Replace <code>import tensorflow<\/code> with <code>import tensorflow as tf<\/code>, also it is probably good to rename your script other than tf.py.<\/p>\n"}]}
+{"QuestionId":52793016,"AnswerCount":0,"Tags":"<python><tensorflow><mnist>","CreationDate":"2018-10-13T12:38:22.743","AcceptedAnswerId":null,"OwnerUserId":8621405.0,"Title":"Split mnist dataset with tf.data","Body":"<p>I am trying to use tf.data to split the training mnist dataset with 60000 images to 55000 training images and 5000 validation images.<\/p>\n\n<p>when i run<\/p>\n\n<pre><code>session_config = tf.ConfigProto(log_device_placement=False)\nconfig = tf.estimator.RunConfig(tf_random_seed=230,\n                                model_dir=chpt_dir_path,\n                                save_checkpoints_steps=params[\"save_checkpoints_steps\"],\n                                keep_checkpoint_max=params[\"keep_checkpoint_max\"],\n                                session_config=session_config)\nestimator = tf.estimator.Estimator(model_fn=model_fn, params=params, config=config)\ntrain_dataset, valid_dataset = train_input_fn(args.DATA_DIR_PATH, params)\nestimator.train(lambda: train_dataset)\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>Tensor(\"Iterator:0\", shape=(), dtype=resource) must be from the same graph as Tensor(\"PrefetchDataset:0\", shape=(), dtype=variant).<\/p>\n<\/blockquote>\n\n<p>The problem comes from this function<\/p>\n\n<pre><code>def train_input_fn(data_dir_path, params):\n    \"\"\"Train input function for the MNIST dataset.\n\n    Args:\n        data_dir: (string) path to the data directory\n        params: (Params) contains hyperparameters of the model (ex: `params.num_epochs`)\n    \"\"\"\n    dataset = train(data_dir_path)\n    dataset = dataset.shuffle(params[\"train_size\"] + params[\"valid_size\"], seed=416)  # whole dataset into the buffer\n\n    train_dataset = dataset.take(params[\"train_size\"])\n    valid_dataset = dataset.skip(params[\"train_size\"])\n\n    train_dataset = train_dataset.batch(params[\"batch_size\"])\n    train_dataset = train_dataset.shuffle(params[\"train_size\"])\n    train_dataset = train_dataset.prefetch(1)  # make sure you always have one batch ready to serve\n\n    valid_dataset = valid_dataset.batch(params[\"batch_size\"])\n    valid_dataset = valid_dataset.shuffle(params[\"valid_size\"])\n    valid_dataset = valid_dataset.prefetch(1)  # make sure you always have one batch ready to serve\n    return train_dataset, valid_dataset\n<\/code><\/pre>\n\n<p>I do not know how to fix this problem. Have anyone know how? or have a better method for splitting data with td.data? <\/p>\n\n<p>The following code is for loading mnist dataset and creating the data pipeline.<\/p>\n\n<pre><code>def download(data_dir_path, filename):\n    \"\"\"Download (and unzip) a file from the MNIST dataset if not already done.\"\"\"\n    filepath = os.path.join(data_dir_path, filename)\n    if tf.gfile.Exists(filepath):\n        return filepath\n    if not tf.gfile.Exists(data_dir_path):\n        tf.gfile.MakeDirs(data_dir_path)\n\n    # CVDF mirror of http:\/\/yann.lecun.com\/exdb\/mnist\/\n    url = \"https:\/\/storage.googleapis.com\/cvdf-datasets\/mnist\/\" + filename + \".gz\"\n    zipped_filepath = filepath + \".gz\"\n    print(\"Downloading %s to %s\" % (url, zipped_filepath))\n    urllib.request.urlretrieve(url, zipped_filepath)\n    with gzip.open(zipped_filepath, \"rb\") as f_in, open(filepath, \"wb\") as f_out:\n        shutil.copyfileobj(f_in, f_out)\n    os.remove(zipped_filepath)\n    return filepath\n\n\ndef dataset(data_dir_path, images_file, labels_file):\n    images_file_path = download(data_dir_path, images_file)\n    labels_file_path = download(data_dir_path, labels_file)\n\n    def decode_image(image):\n        # Normalize from [0, 255] to [0.0, 1.0]\n        image = tf.decode_raw(image, tf.uint8)\n        image = tf.cast(image, tf.float32)\n        image = tf.reshape(image, [784])\n        return image \/ 255.0\n\n    def decode_label(label):\n        label = tf.decode_raw(label, tf.uint8)  # tf.string -&gt; [tf.uint8]\n        label = tf.reshape(label, [])  # label is a scalar\n        return tf.to_int32(label)\n\n    images = tf.data.FixedLengthRecordDataset(images_file_path, 28 * 28, header_bytes=16)\n    images = images.map(decode_image)\n    labels = tf.data.FixedLengthRecordDataset(labels_file_path, 1, header_bytes=8)\n    labels = labels.map(decode_label)\n    return tf.data.Dataset.zip((images, labels))\n\ndef train(data_dir_path):\n    \"\"\"tf.data.Dataset object for MNIST training data.\"\"\"\n    return dataset(data_dir_path, \"train-images-idx3-ubyte\", \"train-labels-idx1-ubyte\")\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52793541,"AnswerCount":1,"Tags":"<conv-neural-network><caffe><convolution><pycaffe>","CreationDate":"2018-10-13T13:39:21.510","AcceptedAnswerId":53625746.0,"OwnerUserId":8477210.0,"Title":"Which file does caffe execute when doing convolutional operation?","Body":"<p>There are three files related to convolution in <code>caffe\\src\\layers<\/code>:<code>conv_layer.cpp<\/code>,<code>conv_layer.cu<\/code>,<code>cudnn_conv_layer.cpp<\/code>.Which file does caffe excute when executing convolutional operation? <\/p>\n","answers":[{"AnswerId":"53625746","CreationDate":"2018-12-05T05:33:12.720","ParentId":null,"OwnerUserId":"10389307","Title":null,"Body":"<p>For CPU implementation: <\/p>\n\n<pre><code>.\/src\/caffe\/layers\/conv_layer.cpp\n<\/code><\/pre>\n\n<p>For CUDA GPU implementation: <\/p>\n\n<p><code>.\/src\/caffe\/layers\/conv_layer.cu<\/code><\/p>\n"}]}
+{"QuestionId":52793709,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning>","CreationDate":"2018-10-13T14:00:23.607","AcceptedAnswerId":52793791.0,"OwnerUserId":3204706.0,"Title":"Issues running my Tensorflow model on GPU","Body":"<p>I'm new to Tensorflow. I have a Tensorflow code which is running on CPU, but I'm trying to change it to run on GPU. I added the <code>tf.device('\/device:GPU:0')<\/code> statement somewhere, but apparently it doesn't work, and I don't see a log.<\/p>\n\n<p>How to change it?<\/p>\n\n<pre><code>#...some stuff\ng_in = tf.Graph()\nwith gfile.FastGFile(path + name + '.pb', 'rb') as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n    tf.import_graph_def(graph_def, name=\"\")\n    g_in = tf.get_default_graph()\n    print(g_in)\n\nwith g_in.as_default(), tf.device('\/device:GPU:0'):\n    sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n    input = sess.graph.get_tensor_by_name(\"cnn\/input_1:0\")\n    output = sess.graph.get_tensor_by_name(\"mymodel:0\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"52793791","CreationDate":"2018-10-13T14:09:37.063","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Tensorflow will run on the GPU by default if you have <code>tensorflow-gpu<\/code> installed.<br>\n <code>with tf.device('\/device:GPU:0')<\/code> is used to choose which GPU to use if you do not want to run it on the default.\n<code>pip freeze<\/code> should show you which installation you have. If it only says <code>tensorflow<\/code> and not <code>tensorflow-gpu<\/code> then you need to uninstall it, install CUDA and reinstall <code>tensorflow-gpu<\/code><\/p>\n"}]}
+{"QuestionId":52793794,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-10-13T14:09:50.973","AcceptedAnswerId":null,"OwnerUserId":10499484.0,"Title":"Keras predictions collecting in one class","Body":"<p>\u0130'm using keras and python 2.x and trying to classify dataset. Problem is predictions always collecting in one class.<\/p>\n\n<p>i'm new at keras what must i do?<\/p>\n\n<p>Here is my confusion matris;<\/p>\n\n<pre><code>    Confusion Matrix\n    [[  0  37   0]\n     [  0 823   0]\n     [  0  26   0]]\n<\/code><\/pre>\n\n<p>Here is my report.<\/p>\n\n<pre><code> precision    recall  f1-score   support\n\n     -1       0.00      0.00      0.00        37\n      0       0.93      1.00      0.96       823\n      1       0.00      0.00      0.00        26\n<\/code><\/pre>\n\n<p>avg \/ total       0.86      0.93      0.89       886<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>dataset= pd.read_csv(\"kartnEgitim.csv\"  )\n\ngiris=dataset.iloc [40:,1:20].get_values()\ncikis=dataset.iloc [40:,20].get_values()\n\nfrom sklearn.model_selection import train_test_split\negitimData, testData, egitimEtiket, testEtiket = train_test_split(giris, cikis, test_size = 0.2)\n\n#from sklearn.preprocessing import StandardScaler\n#sc = StandardScaler()\n#egitimData = sc.fit_transform(egitimData)\n#testData = sc.transform(testData)\n\negitimData=keras.utils.normalize(egitimData,axis=1)\ntestData=keras.utils.normalize(testData,axis=1)\n\nmodel=Sequential()\nmodel.add(Dense(6,input_dim=19,activation='relu'))\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(6,activation='relu'))\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(3))\nmodel.add(Activation('softmax'))\n\n\nmodel.compile(optimizer='adam' ,loss='sparse_categorical_crossentropy' , metrics=['accuracy'])\nmodel.fit(egitimData, egitimEtiket,epochs=10,batch_size=64,class_weight='auto') # ,batch_size=64\n\nval_loss, val_acc=model.evaluate(testData,testEtiket)\nprint(val_loss,val_acc)\n\n# Predicting the Test set results\npredictions = model.predict_classes(testData)\n\nfrom sklearn.metrics import confusion_matrix,classification_report\nprint('Confusion Matrix')\ncm = confusion_matrix(testEtiket, predictions, labels=[-1, 0, 1])\nprint (cm)\n\nprint('Classification Report')\nprint(classification_report(testEtiket, predictions,target_names = ['-1', '0', '1']))\n<\/code><\/pre>\n\n<p>i'm new at keras what must i do?<\/p>\n","answers":[]}
+{"QuestionId":52794053,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api><mscoco>","CreationDate":"2018-10-13T14:40:01.143","AcceptedAnswerId":52848660.0,"OwnerUserId":9265440.0,"Title":"Custom object detection using Tensorflow Object Detection API","Body":"<p>I want to detect helmet and person not wearing helmet.. I use ssd_mobilenet_v1_coco_11_06_2017 model with 1000 helmet images and 1000 without helmet.<\/p>\n\n<p>1) Is that ok to use ssd_mobilenet_v1_coco_11_06_2017 model for training? If not then what model I can use? (because COCO have only 90 classes and helmet not in that list)<\/p>\n\n<p>can somebody help me to clear my doubt...<\/p>\n","answers":[{"AnswerId":"52848660","CreationDate":"2018-10-17T06:39:26.490","ParentId":null,"OwnerUserId":"6918487","Title":null,"Body":"<p>Refer following links  to achieve this<\/p>\n\n<p><a href=\"https:\/\/pythonprogramming.net\/training-custom-objects-tensorflow-object-detection-api-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/pythonprogramming.net\/training-custom-objects-tensorflow-object-detection-api-tutorial\/<\/a><\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/building-a-toy-detector-with-tensorflow-object-detection-api-63c0fdf2ac95\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/building-a-toy-detector-with-tensorflow-object-detection-api-63c0fdf2ac95<\/a><\/p>\n\n<p><a href=\"https:\/\/medium.com\/@dana.yu\/training-a-custom-object-detection-model-41093ddc5797\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@dana.yu\/training-a-custom-object-detection-model-41093ddc5797<\/a><\/p>\n"}]}
+{"QuestionId":52794441,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><google-cloud-ml>","CreationDate":"2018-10-13T15:20:55.663","AcceptedAnswerId":null,"OwnerUserId":4086968.0,"Title":"Encountering preemption OS Error when trying to run distributed GCMLE job","Body":"<p>I am trying to run a distributed GCMLE training job and I keep getting the following error:<\/p>\n\n<pre><code>An error was raised. This may be due to a preemption in a connected worker or parameter server. The current session will be closed and a new session will be created. Error: OS Error\n<\/code><\/pre>\n\n<p>The trainer package is a custom estimator modeled much in the same was as the <code>cloudml-samples<\/code> census custom estimator: <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census\/customestimator\/trainer\" rel=\"noreferrer\">https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census\/customestimator\/trainer<\/a>. It's safe to say the <code>task.py<\/code> files are pretty much identical and within the <code>model.py<\/code> file the <code>input_fn()<\/code> and <code>parse_csv()<\/code> functions are the same and the only real differences are within the specifics of my <code>model_fn()<\/code>.<\/p>\n\n<p>If I configure my model to run with a single <code>standard_p100<\/code> GPU I can train at ~15 steps\/sec. However, if I update my configuration to a distributed setting with 4 workers and 3 parameter servers (see config below) then the preemption error pops up and 10 steps will take ~600 seconds...<\/p>\n\n<p>config-distributed.yaml:<\/p>\n\n<pre><code>trainingInput:\n  scaleTier: CUSTOM\n  masterType: standard_p100\n  workerType: standard_p100\n  parameterServerType: large_model\n  workerCount: 3\n  parameterServerCount: 3\n<\/code><\/pre>\n\n<p>If I use this same configuration with the census custom estimator sample then the model trains faster as expected and doesn't run into the preemption errors. I've tried modifying the census example to more closely mimick my exact code, but still haven't been able to reproduce the error. <\/p>\n\n<p>Has anybody encountered similar preemption issues when trying to train a distributed ml engine job? Any advice on how I could better debug the issue? The only advice I have found online was suggesting to have the number of parameter servers be at least half the number of workers (which is why I upped to 3 parameter servers) but I'm still having no luck.<\/p>\n\n<p>To add more context from the logs, this is a typical (repeated) pattern that happens when I try to train in a distributed setting:<\/p>\n\n<pre><code>master-replica-0 loss = 16.5019, step = 53 (124.505 sec)\nmaster-replica-0 An error was raised. This may be due to a preemption in a connected worker or parameter server. The current session will be closed and a new session will be created. Error: OS Error\nmaster-replica-0 Graph was finalized.\nmaster-replica-0 Restoring parameters from gs:\/\/...\/model.ckpt-0\nmaster-replica-0 Running local_init_op.\nmaster-replica-0 Done running local_init_op.\nmaster-replica-0 Saving checkpoints for 0 into gs:\/\/...\nmaster-replica-0 Skip the current checkpoint eval due to throttle secs (600 secs).\nmaster-replica-0 An error was raised. This may be due to a preemption in a connected worker or parameter server. The current session will be closed and a new session will be created. Error: OS Error\n<\/code><\/pre>\n\n<p>And then this cycle repeats...<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I increased the number of parameter servers to 10 and the steps\/sec juped between 5-10 (still less than 15 with only a single GPU) and the error still happened but a little bit more sporadically. What would this suggest if more parameter servers helped? The CPU and Memory utlization is very low regardless (&lt; 5-10%) so it doesn't seem like the PS's are overloaded, but the model does have a lot of variables to update (50k word embeddings w\/ a large number of dimensions). Could this somehow be contributing to the problems?<\/p>\n","answers":[{"AnswerId":"52856611","CreationDate":"2018-10-17T13:54:24.083","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>The bottleneck in distributed training with lots of parameters is often the network bandwidth. If you saturate the network too much, packets get lost and TensorFlow thinks the parameter server is down. By adding more parameter servers you are able to distribute the network load.<\/p>\n\n<p>Keep in mind that if your model is amenable to using GPUs, you will generally get much better throughput on a single machine with 8 GPUs rather than 8 machines with a single GPU because you won't have any networking overhead.<\/p>\n"}]}
+{"QuestionId":52794599,"AnswerCount":0,"Tags":"<python><tensorflow><crash>","CreationDate":"2018-10-13T15:39:04.023","AcceptedAnswerId":null,"OwnerUserId":9152984.0,"Title":"Python crashed, Process finished with exit code -1073741819 (0xC0000005)","Body":"<p>I'm trying to run tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/layers\/cnn_mnist.py\" rel=\"nofollow noreferrer\">example<\/a> from their github but at some point I'm getting window \"python crashed\" and stopped process in PyCharm:<\/p>\n\n<pre><code>Process finished with exit code -1073741819 (0xC0000005)\n<\/code><\/pre>\n\n<p>Execution is interrupted at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/layers\/cnn_mnist.py#L143\" rel=\"nofollow noreferrer\">143rd line<\/a>. I found several solutions but <strong>none helped<\/strong>:<\/p>\n\n<ol>\n<li><p>To Environment variables in Run\/Debug Configurations dialog add this pair: <code>=C:<\/code> as NAME and <code>c:\\''<\/code> as VALUE<\/p><\/li>\n<li><p>Uncheck PyQt compatible ckeckbox in PyCharm project Settings<\/p><\/li>\n<\/ol>\n\n<p>Problem is still remain<\/p>\n\n<p>Windows 7 x64, PyCharm 2018.2.4 (Community Edition), Python 3.6.5<\/p>\n\n<p><strong>UPD:<\/strong> script is finished correctly if I exclude all pooling layers from NN topology<\/p>\n\n<p><strong>UPD2:<\/strong> script works fine even with pooling layers if it run from anaconda prompt so problem is hidden in pycharm?<\/p>\n","answers":[]}
+{"QuestionId":52794949,"AnswerCount":1,"Tags":"<matrix><lambda><keras><layer>","CreationDate":"2018-10-13T16:23:03.713","AcceptedAnswerId":52801569.0,"OwnerUserId":1269281.0,"Title":"Generate matrix in Keras based on row and column","Body":"<p>How I could create a layer in Keras that ouputs matrix given dimensions (e.g. m, n) with cells having a value based on the row and column?\nHere is the forumula:<\/p>\n\n<pre><code>A[i, 2j] = i \/ (10**(2*j))\nA[i, 2j+1] = i \/ (10**(2*j))\n<\/code><\/pre>\n\n<p>I tried to look on the lamba function but it seems Keras passed only the cell value and not the indices! Any other options (not a loop)<\/p>\n","answers":[{"AnswerId":"52801569","CreationDate":"2018-10-14T10:04:49.293","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You could do the following:<\/p>\n\n<pre><code>from keras.layers import Input\nimport keras.backend as K\nimport numpy as np\n\n\ndef CustomConstantInput(m, n):\n    x = np.arange(m)\n    y = 10 ** (2 * (np.arange(n) \/\/ 2))\n    matrix = x[:, None] \/ y[None, :]\n    print(matrix)\n    fixed_input = K.constant(matrix)\n    return Input(tensor=fixed_input)\n\n\nt = CustomConstantInput(3, 4)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52795440,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-13T17:20:37.880","AcceptedAnswerId":null,"OwnerUserId":8075709.0,"Title":"memory saving gradients or memory check pointing in keras","Body":"<p>I recently found a github repo: <a href=\"https:\/\/github.com\/openai\/gradient-checkpointing\" rel=\"nofollow noreferrer\">https:\/\/github.com\/openai\/gradient-checkpointing<\/a>\nThe main purpose is to reduce gpu memory consumption. And the usage seems pretty straight forward:<\/p>\n\n<pre><code>from tensorflow.python.keras._impl.keras import backend as K\nK.__dict__[\"gradients\"] = memory_saving_gradients.gradients_memory\n<\/code><\/pre>\n\n<p>How can I do the same thing but with keras installed separately, not as a part of tensorflow? Since this didn't work:<\/p>\n\n<pre><code>from keras import backend as K\nK.__dict__[\"gradients\"] = memory_saving_gradients.gradients_memory\n<\/code><\/pre>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"55396550","CreationDate":"2019-03-28T11:32:29.670","ParentId":null,"OwnerUserId":"6906319","Title":null,"Body":"<p>I know I am a bit late, but I recently ran into the same problem, and I was able to solve it. <\/p>\n\n<p>The problem (I think) is that <code>memory_saving_gradients.gradients_memory<\/code> uses a heuristic approach which does not work well for many scenarios. Fortunately, there is an alternative function: <code>memory_saving_gradients.gradients_collection<\/code>, which works perfectly fine, but it requires you to specify at which points in the network the gradient must be checkpointed.<\/p>\n\n<p>As an example on how this can be accomplished, suppose that we want to checkpoint all the Keras layers whose name contains the word 'add' (for instance, to make a resnet memory effcient). Then, you could include something like this after building your model, but before training it:<\/p>\n\n<pre><code>layer_names= [layer.name for layer in self.model.layers]\n[tf.add_to_collection(\"checkpoints\", self.model.get_layer(l).get_output_at(0)) \n        for l in [i for i in layer_names if 'add' in i]]\nK.__dict__[\"gradients\"] = memory_saving_gradients.gradients_collection\n<\/code><\/pre>\n\n<p>I hope it helps!<\/p>\n"}]}
+{"QuestionId":52795454,"AnswerCount":0,"Tags":"<probability><pytorch><bayesian><sampling><pyro.ai>","CreationDate":"2018-10-13T17:21:51.487","AcceptedAnswerId":null,"OwnerUserId":6476080.0,"Title":"Sampling from a joint distribution in Pyro","Body":"<p>I understand how to sample from multidimensional categorical, or multivariate normal (with dependence within each column). For example, for a multivariate categorical, this can be done as below:<\/p>\n\n<pre><code>import pyro as p\nimport pyro.distributions as d\nimport torch as t\np.sample(\"obs1\", d.Categorical(logits=logit_pobs1).independent(1), obs=t.t(obs1))\n<\/code><\/pre>\n\n<p>My question is, how can we do the same, if there are multiple distributions? For example, the following is not what I want as <code>obs1<\/code>, <code>obs2<\/code> and <code>obs3<\/code> are independent to each other.<\/p>\n\n<pre><code>p.sample(\"obs1\", d.Categorical(logits=logit_pobs1).independent(1), obs=t.t(obs1))\np.sample(\"obs2\", d.Normal(loc=mu_obs2, scale=t.ones(mu_obs2.shape)).independent(1), obs=t.t(obs2))\np.sample(\"obs3\", d.Bernoulli(logits=logit_pobs3).independent(1),obs3)\n<\/code><\/pre>\n\n<p>I would like to do something like<\/p>\n\n<pre><code>p.sample(\"obs\", d.joint(d.Bernoulli(...), d.Normal(...), d.Bernoulli(...)).independent(1),obs)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52795551,"AnswerCount":1,"Tags":"<tensorflow><object-detection><object-detection-api>","CreationDate":"2018-10-13T17:32:48.960","AcceptedAnswerId":54533184.0,"OwnerUserId":9632288.0,"Title":"Predictions in recorded video using object detection tensorflow API","Body":"<p>I am trying to read a video file (using opencv), loop over all frames using tensorflow's object-detection API to do the predictions and bounding boxes, and writing the predicted frames (with boxes) to a new video file. I used the object_detection_tutorial.ipynb with some modifications to capture the video frames and process it in faster-rcnn-inception-resnet-v2 loaded from a frozen graph (after trained).<\/p>\n\n<p>I am using a tesla P100 gpu in a cloud machine with windows 10 and 56GB ram. Also using tensorflow-gpu.<\/p>\n\n<p>When I run the code, it takes 0,5 second per frame. Is it a normal speed for a tesla P100 or I am doing something wrong in the code to make it slower?<\/p>\n\n<p>This code is just a test, as later I will have to use it in a real time video prediction task. If 0,5 second per frame is an expected speed using tensorflow API, I think I will cannot use it in my task :(<\/p>\n\n<p>So, after running it, i get the following running times<\/p>\n\n<p>processing frame number 1.0<\/p>\n\n<p>time to capture video frame 0.0<\/p>\n\n<p>time to predict 0.49225664138793945<\/p>\n\n<p>time to generate boxes in a frame 0.14833950996398926<\/p>\n\n<p>time to write a frame in video file 0.04687023162841797<\/p>\n\n<p>total time in the loop 0.6874663829803467<\/p>\n\n<p>As you guys can see, the code using the CPU (opencv) goes fast. But when I use the GPU, it takes almost 0,5 seconds just in prediction task (used in sess.run).<\/p>\n\n<p>Any advices? Thank you in advance. Bellow follows my code<\/p>\n\n<p>from distutils.version import StrictVersion\n    import numpy as np\n    import os\n    import six.moves.urllib as urllib\n    import sys\n    import tarfile\n    import tensorflow as tf\n    import zipfile\n    import time<\/p>\n\n<pre><code>from collections import defaultdict\nfrom io import StringIO\n#from matplotlib import pyplot as plt\nfrom PIL import Image\n\nimport cv2\nfrom imutils import paths\n\nimport re\n\n#This is needed since the code is stored in the object_detection    folder.\nsys.path.append(\"..\")\nfrom object_detection.utils import ops as utils_ops\n\nif StrictVersion(tf.__version__) &lt; StrictVersion('1.9.0'):\n  raise ImportError('Please upgrade your TensorFlow installation to v1.9.* or later!')\n\n\nfrom utils import label_map_util\n\nfrom utils import visualization_utils as vis_util\n\n#Detection using tensorflow inside write_video function\n\ndef write_video():\n\n    filename = 'output\/teste_v2.avi'\n    codec = cv2.VideoWriter_fourcc('W', 'M', 'V', '2')\n    cap = cv2.VideoCapture('pneu_trim2.mp4')\n    framerate = round(cap.get(5),2)\n    w = int(cap.get(3))\n    h = int(cap.get(4))\n    resolution = (w, h)\n\n    VideoFileOutput = cv2.VideoWriter(filename, codec, framerate, resolution)    \n\n    ################################\n    # # Model preparation \n\n    # ## Variables\n    # \n    # Any model exported using the `export_inference_graph.py` tool can be loaded here simply by changing `PATH_TO_FROZEN_GRAPH` to point to a new .pb file.  \n    # \n\n\n    # What model to download.\n    MODEL_NAME = 'training\/pneu_incep_step_24887'\n    print(\"loading model from \" + MODEL_NAME)\n\n    # Path to frozen detection graph. This is the actual model that is used for the object detection.\n    PATH_TO_FROZEN_GRAPH = MODEL_NAME + '\/frozen_inference_graph.pb'\n\n    # List of the strings that is used to add correct label for each box.\n    PATH_TO_LABELS = os.path.join('data', 'object-detection.pbtxt')\n\n    NUM_CLASSES = 5\n\n\n    # ## Load a (frozen) Tensorflow model into memory.\n\n    time_graph = time.time()\n    print('loading graphs')\n    detection_graph = tf.Graph()\n    with detection_graph.as_default():\n      od_graph_def = tf.GraphDef()\n      with tf.gfile.GFile(PATH_TO_FROZEN_GRAPH, 'rb') as fid:\n        serialized_graph = fid.read()\n        od_graph_def.ParseFromString(serialized_graph)\n        tf.import_graph_def(od_graph_def, name='')\n    print(\"tempo build graph = \" + str(time.time() - time_graph))\n\n    # ## Loading label map\n\n    label_map = label_map_util.load_labelmap(PATH_TO_LABELS)\n    categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)\n    category_index = label_map_util.create_category_index(categories)\n\n    ################################\n\n    with tf.Session(graph=detection_graph) as sess:\n        with detection_graph.as_default():\n            while (cap.isOpened()):\n              time_loop = time.time()\n              print('processing frame number: ' + str(cap.get(1)))\n              time_captureframe = time.time()\n              ret, image_np = cap.read()\n              print(\"time to capture video frame = \" + str(time.time() - time_captureframe))\n              if (ret != True):\n                  break\n              # the array based representation of the image will be used later in order to prepare the\n              # result image with boxes and labels on it.\n              #image_np = load_image_into_numpy_array(image)\n              # Expand dimensions since the model expects images to have shape: [1, None, None, 3]\n              image_np_expanded = np.expand_dims(image_np, axis=0)\n              image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n              # Each box represents a part of the image where a particular object was detected.\n              boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n              # Each score represent how level of confidence for each of the objects.\n              # Score is shown on the result image, together with the class label.\n              scores = detection_graph.get_tensor_by_name('detection_scores:0')\n              classes = detection_graph.get_tensor_by_name('detection_classes:0')\n              num_detections = detection_graph.get_tensor_by_name('num_detections:0')\n              # Actual detection.\n              time_prediction = time.time()\n              (boxes, scores, classes, num_detections) = sess.run(\n                  [boxes, scores, classes, num_detections],\n                  feed_dict={image_tensor: image_np_expanded})\n              print(\"time to predict = \" + str(time.time() - time_prediction))\n              # Visualization of the results of a detection.\n              time_visualizeboxes = time.time()\n              vis_util.visualize_boxes_and_labels_on_image_array(\n                  image_np,\n                  np.squeeze(boxes),\n                  np.squeeze(classes).astype(np.int32),\n                  np.squeeze(scores),\n                  category_index,\n                  use_normalized_coordinates=True,\n                  line_thickness=8)\n              print(\"time to generate boxes in a frame = \" + str(time.time() - time_visualizeboxes))\n\n\n              time_writeframe = time.time()\n              VideoFileOutput.write(image_np)\n              print(\"time to write a frame in video file = \" + str(time.time() - time_writeframe))\n\n              print(\"total time in the loop = \" + str(time.time() - time_loop))\n\n    cap.release()\n    VideoFileOutput.release()\n    print('done')\n<\/code><\/pre>\n","answers":[{"AnswerId":"54533184","CreationDate":"2019-02-05T11:14:05.553","ParentId":null,"OwnerUserId":"10159144","Title":null,"Body":"<p>Actually the problem is with the model you were using.\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md<\/a>\nBasically the model Faster-rcnn-inception-resnet-v2 will take more time.\nYou can refer the link to know the speed for the model<\/p>\n"}]}
+{"QuestionId":52796121,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-13T18:40:42.417","AcceptedAnswerId":52796122.0,"OwnerUserId":1165181.0,"Title":"How to get the output from a specific layer from a PyTorch model?","Body":"<p>How to extract the features from a specific layer from a pre-trained PyTorch model (such as ResNet or VGG), without doing a forward pass again? <\/p>\n","answers":[{"AnswerId":"52796122","CreationDate":"2018-10-13T18:40:42.417","ParentId":null,"OwnerUserId":"1165181","Title":null,"Body":"<p>You can register a <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.register_forward_hook\" rel=\"nofollow noreferrer\">forward hook<\/a> on the specific layer you want. Something like:<\/p>\n\n<pre><code>def some_specific_layer_hook(module, input_, output):\n    pass  # the value is in 'output'\n\nmodel.some_specific_layer.register_forward_hook(some_specific_layer_hook)\n\nmodel(some_input)\n<\/code><\/pre>\n\n<p>For example, to obtain <em>res5c<\/em> output in ResNet, you may want to use a <code>nonlocal<\/code> variable (or <code>global<\/code> in Python 2):<\/p>\n\n<pre><code>res5c_output = None\n\ndef res5c_hook(module, input_, output):\n    nonlocal res5c_output\n    res5c_output = output\n\nresnet.layer4.register_forward_hook(res5c_hook)\n\nresnet(some_input)\n\n# Then, use `res5c_output`.\n<\/code><\/pre>\n"}]}
+{"QuestionId":52796629,"AnswerCount":1,"Tags":"<tensorflow><keras><pytorch>","CreationDate":"2018-10-13T19:38:24.880","AcceptedAnswerId":52815227.0,"OwnerUserId":3094105.0,"Title":"How to save and use a trained neural network developed in PyTorch \/ TensorFlow \/ Keras?","Body":"<p>Are there ways to save a model after training and sharing just the model with others? Like a regular script? Since the network is a collection of float matrices, is it possible to just extract these trained weights and run it on new data to make predictions, instead of requiring the users to install these frameworks too? I am new to these frameworks and will make any clarifications as needed.<\/p>\n","answers":[{"AnswerId":"52815227","CreationDate":"2018-10-15T11:00:40.610","ParentId":null,"OwnerUserId":"10494960","Title":null,"Body":"<p>PyTorch: As explained in <a href=\"https:\/\/stackoverflow.com\/questions\/42703500\/best-way-to-save-a-trained-model-in-pytorch#43819235\">this post<\/a>, you can save a model's parameters as a dictionary, or load a dictionary to set your model's parameters. \nYou can also save\/load a PyTorch model as an object.\nBoth procedures require the user to have at least one tensor computation framework installed, e.g. for efficient matrix multiplication.<\/p>\n"}]}
+{"QuestionId":52796727,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><google-cloud-platform><conv-neural-network>","CreationDate":"2018-10-13T19:48:27.460","AcceptedAnswerId":null,"OwnerUserId":10307998.0,"Title":"Alternative to tf.records? To train a CNN in tensorflow","Body":"<p>Im training a CNN in tensorflow using google cloud, I have a lot of images, and  when I do a tf.record the size of that record is huge, So exists another alternative to feed the data to my model??? Thankyou<\/p>\n","answers":[]}
+{"QuestionId":52796751,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><tensorflow.js>","CreationDate":"2018-10-13T19:52:00.533","AcceptedAnswerId":52799798.0,"OwnerUserId":1437245.0,"Title":"Error when checking input: expected dense_Dense5_input to have 4 dimension(s). but got array with shape 5,2,5","Body":"<p>I am learning tensorflow.js and I was trying to create a model that would predict the winner of a random match\/game between 2 \"teams\" based on their \"players\".<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const rawMatches = [\r\n  {\r\n    t1: [2, 99, 3, 5, 7],\r\n    t2: [4, 75, 48, 23, 6],\r\n    winner: 0\r\n  },\r\n  {\r\n    t1: [2, 99, 48, 5, 7],\r\n    t2: [4, 75, 3, 23, 6],\r\n    winner: 1\r\n  },\r\n  {\r\n    t1: [2, 83, 3, 4, 23],\r\n    t2: [4, 75, 58, 25, 78],\r\n    winner: 0\r\n  },\r\n  {\r\n    t1: [26, 77, 11, 5, 7],\r\n    t2: [3, 43, 48, 23, 9],\r\n    winner: 1\r\n  },\r\n  {\r\n    t1: [2, 99, 3, 5, 7],\r\n    t2: [6, 65, 28, 23, 6],\r\n    winner: 0\r\n  }\r\n];\r\n\r\nconst train = async () =&gt; {\r\n  \/\/   [\r\n  \/\/     [[2, 99, 3, 5, 7], [4, 75, 48, 23, 6]],\r\n  \/\/     [[2, 99, 48, 5, 7], [4, 75, 3, 23, 6]],\r\n  \/\/     [[2, 99, 3, 5, 7], [4, 75, 48, 23, 6]]\r\n  \/\/   ];\r\n  const xs = tf.tensor3d(\r\n    rawMatches.map((match, index) =&gt; [match.t1, match.t2])\r\n  );\r\n\r\n  \/\/   [[1, 0], [0, 1], [1, 0]];\r\n  const labelsTensor = tf.tensor1d(\r\n    rawMatches.map(match =&gt; (match.winner === 1 ? 1 : 0)),\r\n    \"int32\"\r\n  );\r\n\r\n  const ys = tf.oneHot(labelsTensor, 2);\r\n\r\n  xs.print();\r\n  ys.print();\r\n\r\n  let model = tf.sequential();\r\n  const hiddenLayer = tf.layers.dense({\r\n    units: 15,\r\n    activation: \"sigmoid\",\r\n    inputShape: [5, 2, 5]\r\n  });\r\n  const outputLayer = tf.layers.dense({\r\n    units: 2,\r\n    activation: \"softmax\"\r\n  });\r\n  model.add(hiddenLayer);\r\n  model.add(outputLayer);\r\n\r\n  const optimizer = tf.train.sgd(0.2);\r\n\r\n  model.compile({\r\n    optimizer,\r\n    loss: \"categoricalCrossentropy\"\r\n  });\r\n\r\n  model.fit(xs, ys, { epochs: 1 });\r\n};\r\n\r\ntrain();<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>After trying to fit the model this errors occurs:<\/p>\n\n<p><code>Error when checking input: expected dense_Dense11_input to have 4 dimension(s). but got array with shape 5,2,5<\/code><\/p>\n\n<p>A code sandbox with the complete code: <a href=\"https:\/\/codesandbox.io\/s\/kr37m63w7\" rel=\"nofollow noreferrer\">https:\/\/codesandbox.io\/s\/kr37m63w7<\/a><\/p>\n","answers":[{"AnswerId":"52799798","CreationDate":"2018-10-14T05:37:36.290","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>There is two issues with this model:<\/p>\n\n<p>First the dimension of the input x passed to the method <code>fit<\/code>. <code>xs<\/code> should be one dimension higher than the first <code>inputShape<\/code>. Because <code>xs<\/code> is an array containing data of shape <code>inputShape<\/code>, the inputShape should be <code>[2, 5]<\/code>.<\/p>\n\n<p>Second, since the dimension of the input and the output do not match, you need to use <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.0\/#layers.flatten\" rel=\"nofollow noreferrer\">tf.flatten<\/a> to change the dimension of the data. Both dimensions don't match because the input data shape is <code>[2, 5] (size = 2)<\/code> whereas the output data shape is <code>[2] (size = 1)<\/code><\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const rawMatches = [\r\n  {\r\n    t1: [2, 99, 3, 5, 7],\r\n    t2: [4, 75, 48, 23, 6],\r\n    winner: 0\r\n  },\r\n  {\r\n    t1: [2, 99, 48, 5, 7],\r\n    t2: [4, 75, 3, 23, 6],\r\n    winner: 1\r\n  },\r\n  {\r\n    t1: [2, 83, 3, 4, 23],\r\n    t2: [4, 75, 58, 25, 78],\r\n    winner: 0\r\n  },\r\n  {\r\n    t1: [26, 77, 11, 5, 7],\r\n    t2: [3, 43, 48, 23, 9],\r\n    winner: 1\r\n  },\r\n  {\r\n    t1: [2, 99, 3, 5, 7],\r\n    t2: [6, 65, 28, 23, 6],\r\n    winner: 0\r\n  }\r\n];\r\n\r\nconst train = () =&gt; {\r\n  const xs = tf.tensor3d(\r\n    rawMatches.map((match, index) =&gt; [match.t1, match.t2])\r\n  );\r\n  const labelsTensor = tf.tensor1d(\r\n    rawMatches.map(match =&gt; (match.winner === 1 ? 1 : 0)),\r\n    \"int32\"\r\n  );\r\n\r\n  const ys = tf.oneHot(labelsTensor, 2);\r\n\r\n  xs.print();\r\n  ys.print();\r\n\r\n  let model = tf.sequential();\r\n  const hiddenLayer = tf.layers.dense({\r\n    units: 15,\r\n    activation: \"sigmoid\",\r\n    inputShape: [2, 5]\r\n  });\r\n  const outputLayer = tf.layers.dense({\r\n    units: 2,\r\n    activation: \"softmax\"\r\n  });\r\n  model.add(hiddenLayer);\r\n  model.add(tf.layers.flatten())\r\n  model.add(outputLayer);\r\n\r\n  const optimizer = tf.train.sgd(0.2);\r\n\r\n  model.compile({\r\n    optimizer,\r\n    loss: \"categoricalCrossentropy\"\r\n  });\r\n\r\n  model.fit(xs, ys, { epochs: 1 });\r\n};\r\n\r\ntrain();<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":52796947,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-datasets><supervised-learning>","CreationDate":"2018-10-13T20:18:06.487","AcceptedAnswerId":52797138.0,"OwnerUserId":7455441.0,"Title":"Tensorflow: Matrix size-incompatible error on Tensors","Body":"<p>I am attempting to do binary classification on a univariate numerical dataset with Tensorflow. My dataset contains 6 features\/variables including the label with about 90 instances. Here is a preview of my data:<\/p>\n\n<pre><code>sex,age,Time,Number_of_Warts,Type,Area,Result_of_Treatment\n1,35,12,5,1,100,0\n1,29,7,5,1,96,1\n1,50,8,1,3,132,0\n1,32,11.75,7,3,750,0\n1,67,9.25,1,1,42,0\n<\/code><\/pre>\n\n<p>I am splitting my data with sklearn's train_test_split function like so:<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(data, y, test_size=0.33, random_state=42)\n<\/code><\/pre>\n\n<p>I then convert my data in to Tensors with the following code:<\/p>\n\n<pre><code>X_train=tf.convert_to_tensor(X_train)\nX_test = tf.convert_to_tensor(X_test)\n\ny_train=tf.convert_to_tensor(y_train)\ny_test = tf.convert_to_tensor(y_test)\n<\/code><\/pre>\n\n<p>After this I begin to construct a simple sequential model. <\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\n\nfrom keras import models\nfrom keras import layers\n\nmodel = models.Sequential()\nmodel.add(layers.Dense(16, activation='relu', input_shape=(60,)))\nmodel.add(layers.Dense(16, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))\n\nmodel.compile(optimizer=optimizers.RMSprop(lr=0.001),\n          loss='binary_crossentropy',\n          metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>The error occurs when I call the fit function <\/p>\n\n<pre><code> history = model.fit(X_train,y_train,epochs=10,steps_per_epoch=200)\n\n InvalidArgumentError: Matrix size-incompatible: In[0]: [60,6], In[1]: [60,16]\n [[{{node dense_43\/MatMul}} = MatMul[T=DT_FLOAT, _class=[\"loc:@training_8\/RMSprop\/gradients\/dense_43\/MatMul_grad\/MatMul_1\"], transpose_a=false, transpose_b=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_identity_dense_43_input_0, dense_43\/kernel\/read)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52797138","CreationDate":"2018-10-13T20:42:57.427","ParentId":null,"OwnerUserId":"5707705","Title":null,"Body":"<p>I think it should be<\/p>\n\n<pre><code>model.add(layers.Dense(16, activation='relu', input_shape=(6,)))\n<\/code><\/pre>\n\n<p>You should refer to the columns and not the rows<\/p>\n"}]}
+{"QuestionId":52797062,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-13T20:33:06.123","AcceptedAnswerId":52797708.0,"OwnerUserId":4747755.0,"Title":"'Tensor' object has no attribute '_keras_history' Keras with no Tensorflow tensor","Body":"<p>This code:<\/p>\n\n<pre><code>a = Input(ish)\nfor i in range(a.shape[1]):\n  x=Conv2D(filters=50, kernel_size=3, padding='same', activation=rl)(a[:,i])\n  x=MaxPooling2D(pool_size=2)(x)\n  x=Dropout(0.5)(x)\n  x=Conv2D(filters=100, kernel_size=5, padding='same', activation=rl)(x)\n  x=MaxPooling2D(pool_size=2)(x)\n  x=Dropout(0.5)(x)\n  x=Conv2D(filters=200, kernel_size=7, padding='same', activation=rl)(x)\n  x=MaxPooling2D(pool_size=2)(x)\nt=Flatten()(x)\nt=Dropout(0.7)(t)\nb=Dense(num_classes, activation='softmax')(t)\n\nmodel = Model(inputs=a, outputs=b)\n<\/code><\/pre>\n\n<p>in the last line, gives this error:<\/p>\n\n<pre><code>AttributeError: 'Tensor' object has no attribute '_keras_history'\n<\/code><\/pre>\n\n<p>Any idea what cause the problem and how to solve it?<\/p>\n","answers":[{"AnswerId":"52797708","CreationDate":"2018-10-13T21:59:23.220","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You have to do the indexing inside a Lambda layer, in order to keep the Keras metadata:<\/p>\n\n<pre><code>a = Input(ish)\nfor i in range(a.shape[1]):\n  x=Lambda(lambda x: x[:, i])(a)\n  x=Conv2D(filters=50, kernel_size=3, padding='same', activation=rl)(x)\n  x=MaxPooling2D(pool_size=2)(x)\n  x=Dropout(0.5)(x)\n  x=Conv2D(filters=100, kernel_size=5, padding='same', activation=rl)(x)\n  x=MaxPooling2D(pool_size=2)(x)\n  x=Dropout(0.5)(x)\n  x=Conv2D(filters=200, kernel_size=7, padding='same', activation=rl)(x)\n  x=MaxPooling2D(pool_size=2)(x)\nt=Flatten()(x)\nt=Dropout(0.7)(t)\nb=Dense(num_classes, activation='softmax')(t)\n\nmodel = Model(inputs=a, outputs=b)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52797747,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning>","CreationDate":"2018-10-13T22:07:08.093","AcceptedAnswerId":null,"OwnerUserId":10500621.0,"Title":"Why do I fail to predict a linear equation (Y=2*x) with Keras?","Body":"<p>I tried to predict a linear equation (Y=2*x) with Keras, but there it failed.<\/p>\n\n<p>With a sigmoid activation function I get rectangular predictions, with ReLu I get NaN\u00b4s.<\/p>\n\n<p>What is the cause? How could I change the code to predict y=2*x.<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Dense, Activation\nfrom keras.models import Sequential\nimport matplotlib.pyplot as plt\nimport math\nimport time\n\n\nx = np.arange(-100, 100, 0.5)\ny = x*2\n\nmodel = Sequential()\n\nmodel.add(Dense(10, input_shape=(1,)))\nmodel.add(Activation('sigmoid'))\n\nmodel.add(Dense(20) )\nmodel.add(Activation('sigmoid'))\n\n\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='SGD', metrics=['mean_squared_error'])\n\nt1 = time.clock()\nfor i in range(40):\n    model.fit(x, y, epochs=1000, batch_size=len(x), verbose=0)\n    predictions = model.predict(x)\n    print (i,\" \", np.mean(np.square(predictions - y)),\" t: \", time.clock()-t1)\n\n    plt.hold(False)\n    plt.plot(x, y, 'b', x, predictions, 'r--')\n    plt.hold(True)\n    plt.ylabel('Y \/ Predicted Value')\n    plt.xlabel('X Value')\n    plt.title([str(i),\" Loss: \",np.mean(np.square(predictions - y)),\" t: \", str(time.clock()-t1)])\n    plt.pause(0.001)\n#plt.savefig(\"fig2.png\")\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52798972","CreationDate":"2018-10-14T02:27:53.593","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>Although it seems that the default learning rate may be inappropriate at the first glance, the real problem here is that <strong><code>sigmoid<\/code> activation is inappropriate<\/strong>. <\/p>\n\n<p>Why? Because your desired output should <strong>NOT<\/strong> be bounded, but using <code>sigmoid<\/code> implies a bounded output. To be more precise, your last layer computes an output <code>y<\/code> as<\/p>\n\n<pre><code>y=\\sum_i{w_i*x_i} + b\n<\/code><\/pre>\n\n<p>while <code>x_i<\/code> here is the output of the second last layer, which is activated by <code>sigmoid<\/code>, indicating that <code>x_i \\in [0,1]<\/code>. Because of this reason, your output <code>y<\/code> is bounded as <code>y \\in [-V+b,+V+b]<\/code>, where <code>V=|w_0|+|w_1|+...+|w_19|<\/code>, also known as the L1norm of the weight matrix, i.e. <code>V=L1norm(W)<\/code>.<\/p>\n\n<p>Since the weight matrix <code>W<\/code> will be learned based on your training data, it is safe to conclude that your model will <strong>NOT<\/strong> be generalizable to those testing data, whose value is outside of the range <code>( min(x_train), max(x_train) )<\/code>. <\/p>\n\n<p><strong>How to fix?<\/strong><\/p>\n\n<p><strong>Thought 1<\/strong>: for this simple problem, you actually don't need any nonlinearity. Simply use a linear MLP as follows.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1, input_shape=(1,)))\nmodel.compile(loss='mse', optimizer='adam')\n<\/code><\/pre>\n\n<p>I tested it, and it should converge in 200 epochs with an MSE around 1e-5.<\/p>\n\n<p><strong>Thought 2<\/strong>: use a different activation function that does not suffer the bounded output issue, e.g. <code>relu<\/code> (note: <code>tanh<\/code> is also inappropriate for the same reason).<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(10, input_shape=(1,)))\nmodel.add(Activation('relu'))\nmodel.add(Dense(20) )\nmodel.add(Activation('relu'))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam')\n<\/code><\/pre>\n\n<p>I also test this model, and it should converge even faster with a comparable MSE. <\/p>\n"}]}
+{"QuestionId":52798089,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-13T23:03:48.377","AcceptedAnswerId":null,"OwnerUserId":9633432.0,"Title":"Concatenating features of two pooling layers","Body":"<p>I am trying to design a Bi-Directional LSTM model and I want to concatenate features after Max pooling and Average pooling layers.\nI have this for my model:<\/p>\n\n<pre><code>from keras.layers import Dense, Embedding\nfrom keras.layers.recurrent import LSTM\nfrom keras.layers import Bidirectional\nfrom keras.models import Sequential\nfrom keras.layers.core import Dropout\nfrom features import train,embedding_matrix,words\nfrom keras.layers import concatenate,AveragePooling1D,GlobalMaxPooling1D\n\nmodel=Sequential()\n\nmodel.add(Embedding(words,300,input_length=train.shape[1],weights=[embedding_matrix]))\nmodel.add(Bidirectional(LSTM(20,activation='tanh',kernel_initializer='glorot_uniform',recurrent_dropout = 0.2, dropout = 0.2,return_sequences=True)))\n\nmodel.add(concatenate([GlobalMaxPooling1D(),AveragePooling1D()]))\n\nmodel.add(Dropout(0.2))\n\nmodel.add(Dense(2, activation='softmax'))\n\nprint model.summary()\n<\/code><\/pre>\n\n<p>But I am having:\n<code>ValueError: Layer concatenate_1 was called with an input that isn't a symbolic tensor<\/code> which is because I believe the concatenating layer. As I am not adding the pooling in the model. <\/p>\n\n<p>Can I add two layers in the same model? or Should I define two separate models and then add pooling layers in each of them?<\/p>\n","answers":[{"AnswerId":"52799729","CreationDate":"2018-10-14T05:22:28.953","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>The trick here is to use a graph model instead of a sequential model. <\/p>\n\n<p>Before we get started, I assume <\/p>\n\n<ol>\n<li>your network expects a 2D input tensor of shape <code>(B=batch_size, N=num_of_words)<\/code>, where N is the longest sample length of your training data. (In case you have unequal length samples, you should use <code>keras.preprocessing.sequence.pad_sequences<\/code> to achieve equal length samples)<\/li>\n<li>your vocabulary size is V (probably is 300 if I understand correctly)<\/li>\n<li><p>your embedding layer encodes each word to a feature of F dimension, i.e. your embedding layer's weight matrix is <code>VxF<\/code>.<\/p>\n\n<p>from keras.layers import Dense, Embedding, Input, Concatenate, Lambda\nfrom keras.layers.recurrent import LSTM\nfrom keras.layers import Bidirectional\nfrom keras.models import Model\nfrom keras.layers.core import Dropout\nfrom keras import backend as BKN<\/p>\n\n<p>from keras.layers import concatenate,AveragePooling1D,GlobalMaxPooling1D<\/p>\n\n<p>words = Input( shape=(N,))<\/p>\n\n<p>f = Embedding(input_dim=V,output_dim=F)( words )\nf = Bidirectional(LSTM(20,activation='tanh',\n                       kernel_initializer='glorot_uniform',\n                       recurrent_dropout = 0.2, \n                       dropout = 0.2,return_sequences=True))(f)\ngpf = GlobalMaxPooling1D()(f) \ngpf = Lambda( lambda t : BKN.expand_dims(t, axis=1) )(gpf)\napf = AveragePooling1D( pool_size=2 )(f)\npf  = Concatenate(axis=1)([gpf, apf])\npf  = Dropout(0.2)( pf )\npred = Dense(2, activation='softmax')(pf) # &lt;-- make sure this is correct<\/p>\n\n<p>model = Model( input=words, output=pred )<\/p><\/li>\n<\/ol>\n\n<p>Finally, I fail to find that keras <code>Embedding<\/code> layer supports syntax like <code>weights=[embedding_matrix]<\/code>.<\/p>\n"}]}
+{"QuestionId":52798329,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><google-cloud-storage>","CreationDate":"2018-10-13T23:51:09.560","AcceptedAnswerId":52800039.0,"OwnerUserId":10307998.0,"Title":"How train in google cloud without uploading the data?","Body":"<p>Im training a CNN in google cloud with tensorflow, and my tf.record is huge, and It's going to take a long time to upload, Can I train without uploading my data to the cloud?<\/p>\n","answers":[{"AnswerId":"52800039","CreationDate":"2018-10-14T06:22:18.313","ParentId":null,"OwnerUserId":"81491","Title":null,"Body":"<p>Sadly, no. Training needs to be done with data, so you'll either need to do the training where the data is or put the data where the training will be. Having the data somewhere else won't be helpful, except for validating a model that's already been trained.<\/p>\n"}]}
+{"QuestionId":52798399,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml><tensorflow-estimator>","CreationDate":"2018-10-14T00:07:54.657","AcceptedAnswerId":53487796.0,"OwnerUserId":260826.0,"Title":"EXCLUDED from export because they cannot be be served via TensorFlow Serving APIs","Body":"<p>Tensorflow version 1.10<\/p>\n\n<p>Using: <code>DNNClassifier<\/code> and <code>tf.estimator.FinalExporter<\/code><\/p>\n\n<p>I'm using the Iris example from TF <a href=\"https:\/\/developers.googleblog.com\/2017\/09\/introducing-tensorflow-datasets.html\" rel=\"nofollow noreferrer\">blog<\/a>.\nI defined the following code:<\/p>\n\n<pre><code># The CSV features in our training &amp; test data.\nCOLUMN_NAMES = ['SepalLength',\n                'SepalWidth',\n                'PetalLength',\n                'PetalWidth',\n                'Species']\nFEATURE_COLUMNS = COLUMN_NAMES[:4]\nINPUT_COLUMNS = [\n    tf.feature_column.numeric_column(column) for column in COLUMN_NAMES\n]\n\ndef serving_input_receiver_fn():\n  \"\"\"Build the serving inputs.\"\"\"\n  inputs = {}\n  for feat in INPUT_COLUMNS:\n    inputs[feat.name] = tf.placeholder(shape=[None], dtype=feat.dtype)\n  return tf.estimator.export.ServingInputReceiver(inputs, inputs)\n<\/code><\/pre>\n\n<p>This is how I call my functions:<\/p>\n\n<pre><code>  train_spec = tf.estimator.TrainSpec(\n      train_input, max_steps=hparams.train_steps)\n  exporter = tf.estimator.FinalExporter(\n      'iris', serving_input_receiver_fn)\n  eval_spec = tf.estimator.EvalSpec(\n      eval_input,\n      steps=hparams.eval_steps,\n      exporters=[exporter],\n      name='iris-eval')\n  run_config = tf.estimator.RunConfig(\n      session_config=_get_session_config_from_env_var())\n  run_config = run_config.replace(model_dir=hparams.job_dir)\n  print('Model dir: %s', run_config.model_dir)\n  estimator = model.build_estimator(\n      # Construct layers sizes.\n      config=run_config,\n      hidden_units=[10, 20, 10],\n      n_classes=3)\n  tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>I get the following messages:<\/p>\n\n<pre><code>INFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Signatures INCLUDED in export for Eval: None\nINFO:tensorflow:Signatures INCLUDED in export for Classify: None\nINFO:tensorflow:Signatures INCLUDED in export for Regress: None\nINFO:tensorflow:Signatures INCLUDED in export for Predict: ['predict']\nINFO:tensorflow:Signatures INCLUDED in export for Train: None\nINFO:tensorflow:Signatures EXCLUDED from export because they cannot be be served via TensorFlow Serving APIs:\nINFO:tensorflow:'serving_default' : Classification input must be a single string Tensor; got {'SepalLength': &lt;tf.Tensor 'Placeholder:0' shape=(?,) dtype=float32&gt;, 'PetalLength': &lt;tf.Tensor 'Placeholder_2:0' shape=(?,) dtype=float32&gt;, 'PetalWidth': &lt;tf.Tensor 'Placeholder_3:0' shape=(?,) dtype=float32&gt;, 'SepalWidth': &lt;tf.Tensor 'Placeholder_1:0' shape=(?,) dtype=float32&gt;, 'Species': &lt;tf.Tensor 'Placeholder_4:0' shape=(?,) dtype=float32&gt;}\nINFO:tensorflow:'classification' : Classification input must be a single string Tensor; got {'SepalLength': &lt;tf.Tensor 'Placeholder:0' shape=(?,) dtype=float32&gt;, 'PetalLength': &lt;tf.Tensor 'Placeholder_2:0' shape=(?,) dtype=float32&gt;, 'PetalWidth': &lt;tf.Tensor 'Placeholder_3:0' shape=(?,) dtype=float32&gt;, 'SepalWidth': &lt;tf.Tensor 'Placeholder_1:0' shape=(?,) dtype=float32&gt;, 'Species': &lt;tf.Tensor 'Placeholder_4:0' shape=(?,) dtype=float32&gt;}\nWARNING:tensorflow:Export includes no default signature!\n<\/code><\/pre>\n\n<p>When I print <code>serving_input_receiver_fn<\/code> I get:<\/p>\n\n<pre><code>ServingInputReceiver(features={'sepal_width': &lt;tf.Tensor 'Placeholder_1:0' shape=(?, 1) dtype=float32&gt;, 'petal_width': &lt;tf.Tensor 'Placeholder_3:0' shape=(?, 1) dtype=float32&gt;, 'sepal_length': &lt;tf.Tensor 'Placeholder:0' shape=(?, 1) dtype=float32&gt;, 'petal_length': &lt;tf.Tensor 'Placeholder_2:0' shape=(?, 1) dtype=float32&gt;}, receiver_tensors={'sepal_width': &lt;tf.Tensor 'Placeholder_1:0' shape=(?, 1) dtype=float32&gt;, 'petal_width': &lt;tf.Tensor 'Placeholder_3:0' shape=(?, 1) dtype=float32&gt;, 'sepal_length': &lt;tf.Tensor 'Placeholder:0' shape=(?, 1) dtype=float32&gt;, 'petal_length': &lt;tf.Tensor 'Placeholder_2:0' shape=(?, 1) dtype=float32&gt;}, receiver_tensors_alternatives=None)\n<\/code><\/pre>\n\n<p>In the export folder there is nothing (CSV, JSON, etc.):<\/p>\n\n<pre><code>gs:\/\/&lt;my-bucket&gt;\/iris\/iris_20181014_214916\/export\/:\ngs:\/\/&lt;my-bucket&gt;\/iris\/iris_20181014_214916\/export\/\n<\/code><\/pre>\n","answers":[{"AnswerId":"53487796","CreationDate":"2018-11-26T19:30:10.863","ParentId":null,"OwnerUserId":"260826","Title":null,"Body":"<p>I found a solution <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/iris\/tensorflow\/estimator\/trainer\/model.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<pre><code>def _make_input_parser(with_target=True):\n  \"\"\"Returns a parser func according to file_type, task_type and target.\n  Need to set record_default for last column to integer instead of float in\n  case of classification tasks.\n  Args:\n    with_target (boolean): Pass label or not.\n  Returns:\n    It returns a parser.\n  \"\"\"\n\n  def _decode_csv(line):\n    \"\"\"Takes the string input tensor and parses it to feature dict and target.\n    All the columns except the first one are treated as feature column. The\n    first column is expected to be the target.\n    Only returns target for if with_target is True.\n    Args:\n      line: csv rows in tensor format.\n    Returns:\n      features: A dictionary of features with key as \"column_names\" from\n        self._column_header.\n      target: tensor of target values which is the first column of the file.\n        This will only be returned if with_target==True.\n    \"\"\"\n    column_header = column_names if with_target else column_names[:4]\n    record_defaults = [[0.] for _ in xrange(len(column_names) - 1)]\n    # Pass label as integer.\n    if with_target:\n      record_defaults.append([0])\n    columns = tf.decode_csv(line, record_defaults=record_defaults)\n    features = dict(zip(column_header, columns))\n    target = features.pop(column_names[4]) if with_target else None\n    return features, target\n\n  return _decode_csv\n\n\ndef serving_input_receiver_fn():\n  \"\"\"This is used to define inputs to serve the model.\n  Returns:\n    A ServingInputReciever object.\n  \"\"\"\n  csv_row = tf.placeholder(shape=[None], dtype=tf.string)\n  features, _ = _make_input_parser(with_target=False)(csv_row)\n  return tf.estimator.export.ServingInputReceiver(features,\n                                                  {'csv_row': csv_row})\n<\/code><\/pre>\n"}]}
+{"QuestionId":52798540,"AnswerCount":3,"Tags":"<python><image><tensorflow><mse><ssim>","CreationDate":"2018-10-14T00:40:53.830","AcceptedAnswerId":53361860.0,"OwnerUserId":4526565.0,"Title":"Working with SSIM loss function in tensorflow for RGB images","Body":"<p>I want to use <strong>SSIM metric<\/strong> as my loss function for the model I'm working on in <strong>tensorflow<\/strong>. SSIM should measure the similarity between my reconstructed output image of my denoising autoencoder and the input uncorrupted image <strong>(RGB)<\/strong>. <\/p>\n\n<p>As of what I understood, for using the SSIM metric in tensorflow, the images should be <strong>normalized<\/strong> to [0,1] or [0,255] and not [-1,1]. After converting my tensors to [0,1] and implementing SSIM as my loss function, the reconstructed image is black and white instead of a colorful RGB image.<\/p>\n\n<pre><code>tf.reduce_mean(tf.image.ssim(reconstructed, truth, 1.0))\n<\/code><\/pre>\n\n<p>My model is working fine with <strong>MSE<\/strong> (mean squared error), <strong>the reconstructed images are colorful (RGB)<\/strong>.<\/p>\n\n<p>using <code>tf.losses.mean_squared_error(truth, reconstructed)<\/code> the reconstructed image would be RGB image, while using SSIM would give me a one dimensional image. <\/p>\n\n<p>Why using <strong>SSIM as loss function gives me different result than MSE<\/strong> (in terms of reconstructed image channels) in tensorflow?<\/p>\n","answers":[{"AnswerId":"56779629","CreationDate":"2019-06-26T19:12:27.060","ParentId":null,"OwnerUserId":"11687201","Title":null,"Body":"<p>The TensorFlow documentation says that no colorspace conversion is applied. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/ssim\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/ssim<\/a><\/p>\n\n<blockquote>\n  <p>\"Note: The true SSIM is only defined on grayscale. This function does not perform any colorspace transform. (If input is already YUV, then it will compute YUV SSIM average.)\"<\/p>\n<\/blockquote>\n"},{"AnswerId":"53343206","CreationDate":"2018-11-16T18:06:33.770","ParentId":null,"OwnerUserId":"10664339","Title":null,"Body":"<p>SSIM is designed to only measure the difference between two luminance signals.  The RGB images are converted to greyscale before measuring similarity.  If that was fed back into the loss function, it wouldn't know if the image was losing color saturation because it wouldn't show up in the error metric.  That's just a theory.<\/p>\n"},{"AnswerId":"53361860","CreationDate":"2018-11-18T14:17:52.193","ParentId":null,"OwnerUserId":"4526565","Title":null,"Body":"<p>I was capable of solving the issue by changing the dynamic range of the images to <strong>2.0<\/strong>, since I have images scaled between <strong>[-1, 1]<\/strong> by:<\/p>\n\n<p><code>loss_rec = tf.reduce_mean(tf.image.ssim(truth, reconstructed, 2.0))<\/code><\/p>\n\n<p>And since a better image quality is shown by a higher <strong>SSIM<\/strong> value, I had to <strong>minimize the negative<\/strong> of my loss function (SSIM) to optimize my model:<\/p>\n\n<p><code>optimizer = tf.train.AdamOptimizer(learning_rate).minimize(-1 * loss_rec)<\/code><\/p>\n"}]}
+{"QuestionId":52800025,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><keras-layer>","CreationDate":"2018-10-14T06:19:35.397","AcceptedAnswerId":52814386.0,"OwnerUserId":5220257.0,"Title":"Keras give input to intermediate layer and get final output","Body":"<p>My model is a simple fully connected network like this:<\/p>\n\n<pre><code>inp=Input(shape=(10,))\nd=Dense(64, activation='relu')(inp)\nd=Dense(128,activation='relu')(d)\nd=Dense(256,activation='relu')(d)     #want to give input here, layer3\nd=Dense(512,activation='relu')(d)\nd=Dense(1024,activation='relu')(d)\nd=Dense(128,activation='linear')(d)\n<\/code><\/pre>\n\n<p>So, after saving the model I want to give input to layer 3. What I am doing right now is this:<\/p>\n\n<pre><code>model=load_model('blah.h5')    #above described network\nprint(temp_input.shape)        #(16,256), which is equal to what I want to give\n\nindex=3\nintermediate_layer_model = Model(inputs=temp_input,\n                                 outputs=model.output)\n\nEnd_output = intermediate_layer_model.predict(temp_input)\n<\/code><\/pre>\n\n<p>But it isn't working, i.e. I am getting errors like incompatible input, inputs should be tuple etc. The error message is:<\/p>\n\n<pre><code>raise TypeError('`inputs` should be a list or tuple.') \nTypeError: `inputs` should be a list or tuple.\n<\/code><\/pre>\n\n<p>Is there any way I can pass my own inputs in middle of network and get the output instead of giving an input at the start and getting output from the end? Any help will be highly appreciated.<\/p>\n","answers":[{"AnswerId":"52814386","CreationDate":"2018-10-15T10:12:28.430","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First you must learn that in Keras when you apply a layer on an input, <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#the-concept-of-layer-node\" rel=\"noreferrer\">a new node is created inside this layer<\/a> which connects the input and output tensors. Each layer may have multiple nodes connecting different input tensors to their corresponding output tensors. To build a model, these nodes are traversed and a new graph of the model is created which consists all the nodes needed to reach output tensors from input tensors (i.e. which you specify when creating a model: <code>model = Model(inputs=[...], outputs=[...])<\/code>.<\/p>\n\n<p>Now you would like to feed an intermediate layer of a model and get the output of the model. Since this is a new data-flow path, we need to create new nodes for each layer corresponding to this new computational graph. We can do it like this:<\/p>\n\n<pre><code>idx = 3  # index of desired layer\ninput_shape = model.layers[idx].get_input_shape_at(0) # get the input shape of desired layer\nlayer_input = Input(shape=input_shape) # a new input tensor to be able to feed the desired layer\n\n# create the new nodes for each layer in the path\nx = layer_input\nfor layer in model.layers[idx:]:\n    x = layer(x)\n\n# create the model\nnew_model = Model(layer_input, x)\n<\/code><\/pre>\n\n<p>Fortunately, your model consists of one-branch and we could simply use a <code>for<\/code> loop to construct the new model. However, for more complex models it may not be easy to do so and you may need to write more codes to construct the new model.<\/p>\n"},{"AnswerId":"58110488","CreationDate":"2019-09-26T06:05:27.297","ParentId":null,"OwnerUserId":"3177661","Title":null,"Body":"<p>Here is another method for achieving the same result. Initially create a new input layer and then connect it to the lower layers(with weights).<\/p>\n\n<p>For this purpose, first <strong>re-initialize<\/strong> these layers(with <strong>same name<\/strong>) and <strong>reload<\/strong> the corresponding <strong>weights<\/strong> from the <strong>parent model<\/strong> using <\/p>\n\n<blockquote>\n  <p>new_model.load_weights(\"parent_model.hdf5\", <strong>by_name=True<\/strong>)<\/p>\n<\/blockquote>\n\n<p>This will load the required weights from the parent model.Just make sure you name your layers properly beforehand.<\/p>\n\n<pre><code>idx = 3  \ninput_shape = model.layers[idx].get_input_shape_at(0) layer\n\nnew_input = Input(shape=input_shape)\n\nd=Dense(256,activation='relu', name='layer_3')(new_input)\nd=Dense(512,activation='relu', name='layer_4'))(d)\nd=Dense(1024,activation='relu', name='layer_5'))(d)\nd=Dense(128,activation='linear', name='layer_6'))(d)\n\nnew_model = Model(new_input, d)\nnew_model.load_weights(\"parent_model.hdf5\", by_name=True)\n<\/code><\/pre>\n\n<p>This method will work for complex models with multiple inputs or branches.You just need to copy the same code for required layers, connect the new inputs and finally load the corresponding weights.<\/p>\n"}]}
+{"QuestionId":52800113,"AnswerCount":1,"Tags":"<android><android-permissions><torch><flashlight>","CreationDate":"2018-10-14T06:35:54.900","AcceptedAnswerId":52800206.0,"OwnerUserId":4620217.0,"Title":"How to turn ON flashlight when device screen is off?","Body":"<p>In my app I need to start the flashlight in an activity that is started by a BroadcastReceiver.\nIt works fine if the app is in foreground or background and the screen is ON but if the screen is OFF (with or without keyguard lock), the app crashes.<\/p>\n\n<p>Analysis: \nWhen screen is OFF, creating the camera object still succeeds, but\nCamera.startPreview() throws RuntimeException.<\/p>\n\n<p>I can understand the motiviation to forbid camera activation when screen is off, but how can I just turn on\/off the flashlight?<\/p>\n\n<pre><code>cam = Camera.open();\n\/\/...\ntry {\n     cam.startPreview();\n}catch(RuntimeException e){\n    Log.w(TAG,\"Failed opening camera preview. Maybe screen locked?: \" + e.getMessage());\n    return;\n}\n<\/code><\/pre>\n\n<p>Permissions:<\/p>\n\n<pre><code> &lt;uses-permission android:name=\"android.permission.CAMERA\" \/&gt;\n<\/code><\/pre>\n\n<p>I tried using the FLASHLIGHT (undocumented) permission, but it is ignored during runtime.<\/p>\n\n<p>Tested on Android 8.1, Xiaomi Mi A1\nminSdkVersion 22<\/p>\n\n<p>Links I searched:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/13342210\/flash-light-is-not-turned-on\">flash light is not turned on<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/38541837\/flashlight-not-working-when-device-screen-is-off-sleep-android-studio\">flashlight not working,when device screen is off\/sleep (android studio)<\/a><\/p>\n","answers":[{"AnswerId":"52800206","CreationDate":"2018-10-14T06:48:54.277","ParentId":null,"OwnerUserId":"1631193","Title":null,"Body":"<p>If the screen is off, the Activity won't actually be launched until its turned on.  That's why the camera won't turn on.  Instead, start a service to do it which can actually run immediately.  <\/p>\n"}]}
+{"QuestionId":52800261,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><neural-network><deep-learning><kaggle>","CreationDate":"2018-10-14T06:57:07.277","AcceptedAnswerId":null,"OwnerUserId":9067169.0,"Title":"How to fetch data to train my neural network model in Kaggle?","Body":"<p>I am trying to build a convolution neural network which classifies between cats and dogs,to train that network I am using Kaggle as a tool ,I have uploaded all the data sets consisting of 8000 training set and 2000 testing set.<\/p>\n\n<p>Problem is how to fetch the data from the input folder? \nScreenshot is in the below link.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vJxVk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vJxVk.png\" alt=\"Click here to view image\"><\/a><\/p>\n\n<p>This is the error which I am getting.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GSAYV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GSAYV.png\" alt=\"Click here to view image\"><\/a><\/p>\n\n<p>Thanks for the help!<\/p>\n","answers":[{"AnswerId":"52800883","CreationDate":"2018-10-14T08:31:52.477","ParentId":null,"OwnerUserId":"7398624","Title":null,"Body":"<p>It seems you may have misspelled your actual folder name \"predictive\" and instead have spelled it \"predective\" in the code.<\/p>\n\n<p>-Lando<\/p>\n"}]}
+{"QuestionId":52800456,"AnswerCount":1,"Tags":"<tensorflow><tfrecord>","CreationDate":"2018-10-14T07:29:01.580","AcceptedAnswerId":null,"OwnerUserId":1270741.0,"Title":"Prevent mixing training examples from different TFRecords files","Body":"<p>I'm training a model on video frame sequences. I have videos stored as tfrecords files, (each video is a separate tfrecords file), with each frame encoded as a single tf.train.Example. During training I read batches of frames and run them through the model. However, I see that it is possible for frames from different videos to end up in the same batch. For the model I'm training it doesn't make sense and is actually harming performance. Is there a way to make sure that a batch of frames contains frames only from the same video?<\/p>\n","answers":[{"AnswerId":"52802617","CreationDate":"2018-10-14T12:25:34.100","ParentId":null,"OwnerUserId":"1270741","Title":null,"Body":"<p>I ended up using <code>tf.data.Dataset.filter<\/code> method. The predicate compares video ids of the first and last frames. If they're different, then the whole batch is dropped:<\/p>\n\n<pre><code>def ensure_same_video_in_batch(video_ids, *args):\n    first_frame_video_id, last_frame_video_id = video_ids[0], video_ids[-1]\n    return tf.equal(first_frame_video_id, last_frame_video_id)\n\n\ndataset = dataset.filter(ensure_same_video_in_batch)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52800582,"AnswerCount":3,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-14T07:48:22.817","AcceptedAnswerId":null,"OwnerUserId":10500621.0,"Title":"Why do I fail to predict y=x**4 with Keras? (y=x**3 works)","Body":"<p>I manage to predict <code>y=x**2<\/code> and <code>y=x**3<\/code>, but equations like <code>y=x**4<\/code> or <code>y=x**5<\/code> or <code>y=x**7<\/code> converge only to inaccurate lines?<\/p>\n\n<p>What do I do wrong? What could I improve?<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Dense, Activation\nfrom keras.models import Sequential\nimport matplotlib.pyplot as plt\nimport math\nimport time\n\nx = np.arange(-100, 100, 0.5)\ny = x**4\n\nmodel = Sequential()\nmodel.add(Dense(50, input_shape=(1,)))\nmodel.add(Activation('sigmoid'))\nmodel.add(Dense(50) )\nmodel.add(Activation('elu'))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam')\n\nt1 = time.clock()\nfor i in range(100):\n    model.fit(x, y, epochs=1000, batch_size=len(x), verbose=0)\n    predictions = model.predict(x)\n    print (i,\" \", np.mean(np.square(predictions - y)),\" t: \", time.clock()-t1)\n\n    plt.hold(False)\n    plt.plot(x, y, 'b', x, predictions, 'r--')\n    plt.hold(True)\n    plt.ylabel('Y \/ Predicted Value')\n    plt.xlabel('X Value')\n    plt.title([str(i),\" Loss: \",np.mean(np.square(predictions - y)),\" t: \", str(time.clock()-t1)])\n    plt.pause(0.001)\n\n#plt.savefig(\"fig2.png\")\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52806075","CreationDate":"2018-10-14T18:58:56.507","ParentId":null,"OwnerUserId":"3767239","Title":null,"Body":"<p>The problem is that your input and output variables have too large values and hence are not compatible with the (initial) weights of the network. For <a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\"><code>Dense<\/code> layer<\/a> the default kernel initializer is <a href=\"https:\/\/keras.io\/initializers\/#glorot_uniform\" rel=\"nofollow noreferrer\"><code>glorot_uniform<\/code><\/a>; the documentation states that:<\/p>\n\n<blockquote>\n  <p>It draws samples from a uniform distribution within [-limit, limit] where limit is sqrt(6 \/ (fan_in + fan_out)) where fan_in is the number of input units in the weight tensor and fan_out is the number of output units in the weight tensor.<\/p>\n<\/blockquote>\n\n<p>For your example the weights of the first and last layer are therefore sampled on the interval <code>[0.34, 0.34]<\/code>. Now there are two issues that have to do with the magnitude of weights and input\/outputs:<\/p>\n\n<ul>\n<li>The inputs are in the range <code>[-100, 100]<\/code> and hence the output of the first <code>Dense<\/code> layer will be about <code>58 * 0.2 ~= 10<\/code> (the two numbers are the std. dev. of input and weights respectively); it will be smaller for smaller inputs but larger for larger ones. Because this is fed into a sigmoid activation it is likely to saturate. For the example value it will be <code>(1 + exp(-10))**-1 ~= 0.99995<\/code>. This will cause problems during <a href=\"https:\/\/en.wikipedia.org\/wiki\/Backpropagation#Finding_the_derivative_of_the_error\" rel=\"nofollow noreferrer\">backpropagation<\/a> because the weight updates are proportional to the gradient of the activation function which in this case is very small; i.e. the weights are not updated much.<\/li>\n<li>The other issue has to do with the magnitude of outputs <code>y<\/code>. To see why let's step through the network. The sigmoid activation outputs in the range <code>[0, 1]<\/code> and hence the activation of the next dense layer will be in the same order of magnitude (given the default <code>glorot_uniform<\/code> initializer). The <code>ELU<\/code> activation won't change the order of magnitude and hence the input to the last layer is still on the order of magnitude <code>1<\/code>. It also uses the <code>glorot_uniform<\/code> initializer and hence has weights within the range <code>[-0.34, 0.34]<\/code>. However the outputs are in the range of <code>[-1e8, 1e8]<\/code>. In order to generate such huge outputs this means the optimizer would need to step through about 7 (!) orders of magnitude during the fitting procedure. This will take (almost) forever.<\/li>\n<\/ul>\n\n<p>So what can we do about it? On the one hand we could modify the weight initialization and on the other hand we can scale the inputs and outputs to a more moderate range. The latter is a much better idea since any numerical computation is much more accurate when performed in the order of magnitude <code>1<\/code>. Also MSE loss is going to explode for orders of magnitude difference.<\/p>\n\n<h1>Variable scaling<\/h1>\n\n<p>The <a href=\"http:\/\/scikit-learn.org\/stable\/index.html\" rel=\"nofollow noreferrer\"><code>scikit-learn<\/code><\/a> package provides various routines for <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/classes.html#module-sklearn.preprocessing\" rel=\"nofollow noreferrer\">data preparation<\/a> as for example the <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html\" rel=\"nofollow noreferrer\"><code>StandardScaler<\/code><\/a>. This will subtract the mean from the data and then divide by its standard deviation, i.e. <code>x -&gt; (x - mu) \/ sigma<\/code>.<\/p>\n\n<pre><code>x_scaler = StandardScaler()\ny_scaler = StandardScaler()\n\nx = x_scaler.fit_transform(x[:, None])  # Features are expected as columns vectors.\ny = y_scaler.fit_transform(y[:, None])\n\n... # Model definition and fitting goes here.\n\n# Invert the transformation before plotting.\nx = x_scaler.inverse_transform(x).ravel()\ny = y_scaler.inverse_transform(y).ravel()\npredictions = y_scaler.inverse_transform(predictions).ravel()\n<\/code><\/pre>\n\n<p>After 2000 epochs of training (full batch size):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AxsWm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AxsWm.png\" alt=\"Feature scaling\"><\/a><\/p>\n\n<h1>Weight initialization<\/h1>\n\n<p><strong>Not recommend!<\/strong> Instead feature scaling should be used, I just provide the example for the sake of completeness. So in order to make the weights compatible with the input\/output we can specify custom initializers for the first and last layer of the network:<\/p>\n\n<pre><code>model.add(Dense(50, input_shape=(1,),\n                kernel_initializer=RandomUniform(-0.001, 0.001)))\n\n... # Activations and intermediate layers.\n\nmodel.add(Dense(1, kernel_initializer=RandomUniform(-1e7, 1e7)))\n<\/code><\/pre>\n\n<p>Note the small weights for the first layer (in order to prevent saturation of the sigmoid) and the large weights of the last layer (in order to help the network scaling the outputs by the required 7 orders of magnitude).<\/p>\n\n<p>Again, after 2000 epochs (full batch size):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/n8D8r.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n8D8r.png\" alt=\"Custom weight initialization\"><\/a><\/p>\n\n<p>As you can see, it works as well, but not as well as for the scaled feature approach. Furthermore, the larger the number, the larger the risk to run into numerical instabilities. A good rule of thumb is to try to always stay in the region around <code>1<\/code> (plus\/minus a (very) few orders of magnitude).<\/p>\n"},{"AnswerId":"52800983","CreationDate":"2018-10-14T08:48:55.833","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>That's a cool question!<\/p>\n\n<p>This happens because the data is not properly scaled. As a consequence, some activations (i.e. sigmoid) saturate more easily and gradients get close to zero. The easiest solution is to scale your data as follows:<\/p>\n\n<pre><code>x_orig = x\ny_orig = y\n\nx_mean = np.mean(x)\nx_std = np.std(x)\nx = (x - x_mean)\/x_std\n\ny_mean = np.mean(y)\ny_std = np.std(y)\ny = (y - y_mean)\/y_std\n<\/code><\/pre>\n\n<p>As a result of scaling the data in this manner, the approximation at  the first iteration is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PMdPy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PMdPy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The original range can then be recovered as follows:<\/p>\n\n<pre><code>y_pred = predictions*y_std + y_mean\nplt.plot(x_orig, y_orig, 'b', x_orig, y_pred, 'r--')\n<\/code><\/pre>\n"},{"AnswerId":"52800967","CreationDate":"2018-10-14T08:46:13.627","ParentId":null,"OwnerUserId":"9138892","Title":null,"Body":"<p>I think it is because the range of input data is so large. Adding a batchnorm layer can improve the performance. Here is the result of the model with batchnorm layer. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hPROb.png\" rel=\"nofollow noreferrer\">The figure<\/a><\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import numpy as np\nimport keras\nfrom keras.layers import Dense, Activation\nfrom keras.models import Sequential\nimport matplotlib.pyplot as plt\nimport math\nimport time\n\n\nx = np.arange(-100, 100, 0.5)\ny = x**4\n\n\nmodel = Sequential()\nmodel.add(keras.layers.normalization.BatchNormalization(input_shape=(1,)))\nmodel.add(Dense(200))\nmodel.add(Activation('relu'))\nmodel.add(Dense(50))\nmodel.add(Activation('elu'))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam')\n\n\nt1 = time.clock()\nfor i in range(100):\n    model.fit(x, y, epochs=1000, batch_size=len(x), verbose=0)\n    predictions = model.predict(x)\n    print (i,\" \", np.mean(np.square(predictions - y)),\" t: \", time.clock()-t1)\n\n    plt.hold(False)\n    plt.plot(x, y, 'b', x, predictions, 'r--')\n    plt.hold(True)\n    plt.ylabel('Y \/ Predicted Value')\n    plt.xlabel('X Value')\n    plt.title([str(i),\" Loss: \",np.mean(np.square(predictions - y)),\" t: \", str(time.clock()-t1)])\n    plt.pause(0.001)\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52800849,"AnswerCount":0,"Tags":"<python><deep-learning><pytorch><loss-function>","CreationDate":"2018-10-14T08:28:11.647","AcceptedAnswerId":null,"OwnerUserId":9262321.0,"Title":"A strange loss descent(looks like a 'cliff') appeared in training 3d Unet with weighted cross-entropy loss or dice score loss","Body":"<p><strong>The figure of dice loss:<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/ybuoF.png\" alt=\"The figure of dice loss\"><\/p>\n\n<p><strong>The figure of weighted cross-entropy loss:<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/JvBZS.png\" alt=\"The figure of weighted cross-entropy loss\"><\/p>\n\n<p>The optimizer is  Adam, lr = 0.0002, beta1=0.5, beta2=0.999. Does anyone have the same problem as me? Can you tell me the solution and the reason for this?<\/p>\n\n<p>The 3d-unet models show in following.<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nimport numpy as np\n\nclass UNet3D(nn.Module):\n    def name(self):\n        return self.name\n    def __init__(self, name, in_channels, out_channels, min_filters=2, norm_layer=nn.BatchNorm3d):\n        super(UNet3D, self).__init__()\n        self.name = name\n\n        activation = nn.LeakyReLU(0.2, True) \n\n        self.input = Conv2Block(in_channels, min_filters, activation=activation, norm_layer=norm_layer)\n\n        self.down1 = DownBlock(min_filters, min_filters*2, activation=activation, norm_layer=norm_layer)\n        self.down2 = DownBlock(min_filters*2, min_filters*4, activation=activation, norm_layer=norm_layer)\n        self.down3 = DownBlock(min_filters*4, min_filters*8, activation=activation, norm_layer=norm_layer)\n\n        self.up1 = UpBlock(min_filters*8, min_filters*4, activation=activation, norm_layer=norm_layer)\n        self.up2 = UpBlock(min_filters*4, min_filters*2, activation=activation, norm_layer=norm_layer)\n        self.up3 = UpBlock(min_filters*2, min_filters, activation=activation, norm_layer=norm_layer)\n\n        if self.name == 'netS':\n            self.out = OutBlock(min_filters, out_channels, is_image=False)\n        elif self.name == 'netG':\n            self.out = OutBlock(min_filters, out_channels, is_image=True)\n\n    def forward(self, x):\n        # print('x', np.unique(x))\n        x1 = self.input(x)\n        # print('input', np.unique(x1.detach().numpy()))\n        x2 = self.down1(x1)\n        # print('down1', np.unique(x2.detach().numpy()))\n        x3 = self.down2(x2)\n        # print('down2', np.unique(x3.detach().numpy()))\n        x = self.down3(x3)\n        # print('down3', np.unique(x.detach().numpy()))\n\n        x = self.up1(x, x3)\n        # print('up1', np.unique(x.detach().numpy()))\n        x = self.up2(x, x2)   \n        # print('up2', np.unique(x.detach().numpy()))\n        x = self.up3(x, x1)\n        # print('up3', np.unique(x.detach().numpy()))\n\n        x = self.out(x)\n        # print('out', np.unique(x.detach().numpy()))\n\n        return x\n\nclass Conv2Block(nn.Module):\n    '''\n    Two successive Conv3d, each Conv3d followed by norm_layer and activation.\n        first one| in: in_channels, out: out_channels\n        second one| in: out_channels, out: out_channels\n    '''\n    def __init__(self, in_channels, out_channels, activation, norm_layer=nn.BatchNorm3d, kernel_size=3):\n        super(Conv2Block, self).__init__()\n        self.conv1 = nn.Conv3d(in_channels, out_channels, kernel_size, padding=int(np.ceil((kernel_size-1)\/2)))\n        self.conv2 = nn.Conv3d(out_channels, out_channels, kernel_size, padding=int(np.ceil((kernel_size-1)\/2)))\n\n        self.norm_layer = norm_layer\n        if(norm_layer != None):\n            self.bn1 = norm_layer(out_channels)\n            self.bn2 = norm_layer(out_channels)\n\n        self.activation = activation\n\n    def forward(self, x):\n        if(self.norm_layer != None):\n            x = self.activation(self.bn1(self.conv1(x)))\n            x = self.activation(self.bn2(self.conv2(x)))\n        else:\n            x = self.activation(self.conv1(x))\n            x = self.activation(self.conv2(x))\n        return x\n\nclass DownBlock(nn.Module):\n    '''\n    MaxPool3d + Conv2Block\n    '''\n    def __init__(self, in_channels, out_channels, activation, norm_layer=nn.BatchNorm3d, kernel_size=3):\n        super(DownBlock, self).__init__()\n        self.conv = Conv2Block(in_channels, out_channels, activation, norm_layer=norm_layer, kernel_size=kernel_size)\n\n    def forward(self, x):\n        max_pool = nn.MaxPool3d(2)\n        x = self.conv(max_pool(x))\n        return x\n\nclass UpConv(nn.Module):\n    '''\n    interpolate + conv3d + activation\n    '''\n    def __init__(self, in_channels, out_channels, activation, kernel_size=3):\n        super(UpConv, self).__init__()\n        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, padding=int(np.ceil((kernel_size-1)\/2)))\n        self.activation = activation\n\n    def forward(self, x, sz):\n        x = F.interpolate(x, sz, mode='trilinear', align_corners=True)\n        # x = F.interpolate(x, sz, mode='nearest')\n\n        x = self.activation(self.conv(x))           # TODO \u672a\u4f5cnormalization\n        return x\n\nclass UpBlock(nn.Module):\n    '''\n    UpConv + cat + Conv2Block\n    '''\n    def __init__(self, in_channels, out_channels, activation, norm_layer=nn.BatchNorm3d, kernel_size=3):\n        super(UpBlock, self).__init__()\n        self.upconv = UpConv(in_channels, out_channels, activation, kernel_size=kernel_size)\n        self.conv = Conv2Block(in_channels, out_channels, activation, norm_layer=norm_layer, kernel_size=kernel_size)\n\n    def forward(self, x, x2):\n        x = self.upconv(x, (x2.shape[-3], x2.shape[-2], x2.shape[-1]))\n        x = torch.cat([x, x2], dim=1)\n        x = self.conv(x)\n        return x\n\nclass OutBlock(nn.Module):\n    def __init__(self, in_channels, out_channels, is_image):\n        super(OutBlock, self).__init__()\n        self.conv = nn.Conv3d(in_channels, out_channels, 1)\n\n        self.out = nn.Tanh()\n\n        if is_image:\n            self.out = nn.Tanh()\n        else:\n            self.out = nn.Sigmoid()\n\n    def forward(self, x):\n        x = self.conv(x)\n\n        x = self.out(x)\n        return x\n<\/code><\/pre>\n\n<p>Snippets of weighted cross-entropy loss and dice score.<\/p>\n\n<pre><code>class SoftDiceLoss(nn.Module):\n    def __init__(self):\n        super(SoftDiceLoss, self).__init__()\n\n    def forward(self, preds, labels):\n        preds = F.softmax(preds, dim=1)\n\n        num = labels.size(0)\n        m1 = preds.view(num, -1)\n        m2 = labels.view(num, -1)\n        intersection = (m1 * m2)\n\n        score = 2. * (intersection.sum(1) + 1) \/ (m1.sum(1) + m2.sum(1) + 1)  # m1.sum = cube size\n        score = 1 - score.sum() \/ num\n        return score\n\nclass WeightedCrossEntropyLoss(nn.Module):\n    '''\n    Negative log likelihood loss\n    '''   \n    def __init__(self):\n        super(WeightedCrossEntropyLoss, self).__init__()\n        # self.loss = nn.NLLLoss()\n\n    def forward(self, pred, gt, alpha=3.0):\n        gt = torch.squeeze(gt.argmax(dim=1), dim=1)\n\n        assert gt.dtype == torch.long    \n        import torch.nn.functional as F\n        mtx = F.cross_entropy(pred, gt, reduction='none')\n\n        bg = (gt == 0) + (gt == 5)    # background\n\n        neg = mtx[bg]\n        pos = mtx[1-bg]\n\n        Np, Nn = pos.numel(), neg.numel()\n\n        pos = pos.sum()\n\n        k = min(Np*alpha, Nn)\n        if k &gt; 0:\n            neg, _ = torch.topk(neg, int(k))\n            neg = neg.sum()\n        else:\n            neg = 0.0\n\n\n        loss = (pos + neg)\/(Np + k)\n\n        return loss\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52800897,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2018-10-14T08:34:18.490","AcceptedAnswerId":53285978.0,"OwnerUserId":7890801.0,"Title":"Tensorflow runs \"Running per image evaluation\" indefinitly","Body":"<p>I am running my first tensorflow job (object detection training) right now, using the tensorflow API. I am using the ssd mobilenet network from the modelzoo. I used the >>ssd_mobilenet_v1_0.75_depth_quantized_300x300_coco14_sync.config&lt;&lt; as a config-file and as a fine tune checkpoint the >>ssd_mobilenet_v1_0.75_depth_300x300_coco14_sync_2018_07_03&lt;&lt; checkpoint.<\/p>\n\n<p>I started my training with the following command:<\/p>\n\n<pre><code>PIPELINE_CONFIG_PATH='\/my_path_to_tensorflow\/tensorflow\/models\/research\/object_detection\/models\/model\/ssd_mobilenet_v1_0.75_depth_quantized_300x300_coco14_sync.config'\nMODEL_DIR='\/my_path_to_tensorflow\/tensorflow\/models\/research\/object_detection\/models\/model\/train'\nNUM_TRAIN_STEPS=200000\nSAMPLE_1_OF_N_EVAL_EXAMPLES=1\npython object_detection\/model_main.py \\\n--pipeline_config_path=${PIPELINE_CONFIG_PATH} \\\n--model_dir=${MODEL_DIR} \\\n--num_train_steps=${NUM_TRAIN_STEPS} \\\n--sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \\\n--alsologtostderr\n<\/code><\/pre>\n\n<p>No coming to my problem, I hope the community can help me with. I trained the network over night and it trained for 1400 steps and then started evaluating per image, which was running the entire night. Next morning I saw, that network only evaluated and the training was still at 1400 steps. You can see part of the console output in the image below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8bGwl.png\" rel=\"nofollow noreferrer\">Console output from evaluation<\/a><\/p>\n\n<p>I tried to take control by using the eval config parameter in the config file.<\/p>\n\n<pre><code>eval_config: {\n  metrics_set: \"coco_detection_metrics\"\n  use_moving_averages: false\n  num_examples: 5000\n}\n<\/code><\/pre>\n\n<p>I added max_evals = 1, because the documentation says that I can limit the evaluation like this. I also changend eval_interval_secs = 3600 because I only wanted one eval every hour. Both options had no effect.<\/p>\n\n<p>I also tried other config-files from the modelzoo, with no luck. I searched google for hours, only to find answers which told me to change the parameters I already changed. So I am coming to stackoverflow to find help in this Matter. <\/p>\n\n<p>Can anybody help me, maybe hat the same experience? Thanks in advance for all your help!<\/p>\n\n<h3>Environment information<\/h3>\n\n<pre><code>$ pip freeze | grep tensor\ntensorboard==1.11.0\ntensorflow==1.11.0\ntensorflow-gpu==1.11.0\n\n$ python -V\nPython 2.7.12\n<\/code><\/pre>\n","answers":[{"AnswerId":"52801492","CreationDate":"2018-10-14T09:56:18.600","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Could it be that evaluation takes more than 10 minutes in your case? It could be that since 10 minutes is the default interval for making evaluation, it keeps evaluating.<\/p>\n\n<p>Unfortunately, the current API doesn't easily support altering the time interval for evaluation.\nBy default, evaluation happens after every checkpoint saving, which by default is set to 10 minutes.\nTherefore you can change the time for saving a checkpoint by specifying <code>save_checkpoint_secs<\/code> or <code>save_checkpoint_steps<\/code> as an input to the instance of MonitoredSession (or MonitoredTrainingSession). Unfortunately and best to my knowledge, these parameters are not available to be set as flags to model_main.py or from the config file. Therefore, you can either change their value by hard coding, or exporting them out so that they will be available.<\/p>\n\n<p>An alternative way, without changing the frequency of saving a checkpoint, is modifying the evaluation frequency which is specified as <code>throttle_secs<\/code> to <code>tf.estimator.EvalSpec<\/code>.\nSee my explanation <a href=\"https:\/\/stackoverflow.com\/a\/52425031\/10263280\">here<\/a> as to how to export this parameter to model_main.py.<\/p>\n"},{"AnswerId":"53285978","CreationDate":"2018-11-13T16:57:10.307","ParentId":null,"OwnerUserId":"7890801","Title":null,"Body":"<p>I figured out a solution for the problem. The problem with tensorflow 1.10 and after is, that you can not set checkpoint steps or checkpoint secs in the config file like before. By default tensorflow 1.10 and after saves a checkpoint every 10 min. If your hardware is not fast enough and you need more then 10 min for evaluation, you are stuck in a loop. <\/p>\n\n<p>So to change the time steps or training steps till a new checkpoint is safed (which triggers the evaluation), you have to navigate to the model_main.py in the following folder: <\/p>\n\n<blockquote>\n  <p>tensorflow\/models\/research\/object_detection\/ <\/p>\n<\/blockquote>\n\n<p>Once you opened model_main.py, navigate to line 62. Here you will find <\/p>\n\n<blockquote>\n  <p>config = tf.estimator.RunConfig(model_dir=FLAGS.model_dir) <\/p>\n<\/blockquote>\n\n<p>To trigger the checkpoint save after 2500 steps for example, change the entry to this: <\/p>\n\n<blockquote>\n  <p>config = tf.estimator.RunConfig(model_dir=FLAGS.model_dir,save_checkpoints_steps=2500). <\/p>\n<\/blockquote>\n\n<p>Now the model is saved every 2500 steps and afterwards an evaluation is done.<\/p>\n\n<p>There are multiple parameters you can pass through this option. You can find a documentation here: <\/p>\n\n<blockquote>\n  <p>tensorflow\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/run_config.py. <\/p>\n<\/blockquote>\n\n<p>From Line 231 to 294 you can see the parameters and documentation.<\/p>\n\n<p>I hope I can help you with this and you don't have to look for an answer as long as I did.<\/p>\n"}]}
+{"QuestionId":52801295,"AnswerCount":1,"Tags":"<tensorflow><memory><graph><overflow><out>","CreationDate":"2018-10-14T09:30:37.677","AcceptedAnswerId":null,"OwnerUserId":9821809.0,"Title":"ResourceExhaustedError When running network demo fourth try","Body":"<p>I have 1600 videos and I want to make joint annotation label data about videos.<\/p>\n\n<p>I've already made the open pose network and I put my videos as input of the network and saved the joint data as json file.<\/p>\n\n<p>When I put my first video data as input, there are no errors. And when I put second, third video as input, there are no errors too.<\/p>\n\n<p>But When I put the fourth video data as input, I got these error message.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/T8Rbd.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FghF8.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>these above images are the error message.(OOM)<\/p>\n\n<p>The size of first, second, third, fourth video is the same.<\/p>\n\n<p>When I change name first and fourth video name, I got the same error when putting fourth video.<\/p>\n\n<p>I think this error is about the graph. but I couldn't know why exactly.<\/p>\n\n<p>I think there are many genious on stackoverflow. So please answer my question... :)<\/p>\n","answers":[{"AnswerId":"52803258","CreationDate":"2018-10-14T13:41:16.870","ParentId":null,"OwnerUserId":"9821809","Title":null,"Body":"<p>I solve this problem by using cpu. not using gpu.<\/p>\n\n<p>I use cpu only in tensorflow for solving this problem. and it works!<\/p>\n"}]}
+{"QuestionId":52801498,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-14T09:57:09.430","AcceptedAnswerId":null,"OwnerUserId":4481339.0,"Title":"Multiply a series of matrices","Body":"<p>I have a matrix like below:<\/p>\n\n<pre><code>[0 0 1 1]\n[0 0 1 1]\n[0 0 0 0]\n[0 0 0 0]\n<\/code><\/pre>\n\n<p>I need to divide it into multiple <code>3x3<\/code> matrices starting from top left through right. It's sort of a <code>3x3<\/code> slide across the matrix. In this example, we would have <code>4<\/code> <code>3x3<\/code> matrices like so:<\/p>\n\n<pre><code>    [0 0 1]     [0 1 1]\n1 = [0 0 1] 2 = [0 1 1]\n    [0 0 0]     [0 0 0]\n\n    [0 0 1]     [0 1 1]\n3 = [0 0 0] 4 = [0 0 0]\n    [0 0 0]     [0 0 0]\n<\/code><\/pre>\n\n<p>I've tried this using <code>tf.extract_image_patch<\/code> and got the 4 matrices, but I'm still not sure how I can do a sort-of running product for these matrices in Tensorflow. Or, better could achieve the running product without having to pre-calculate the separate matrices.<\/p>\n\n<p>With running product I mean this: I need to multiple above 1-4 matrices element-wise and need to get <code>1<\/code> <code>3x3<\/code> matrix. For example, <code>1<\/code> <code>&amp;<\/code> <code>2<\/code> matrices would be multiplied, the result would be multiplied with matrix <code>3<\/code>, and the result again would be multiplied with matrix <code>4<\/code>.This operation should give me start of the patch(<code>[[1 1], [1 1]]<\/code>) in my original matrix, a matrix like below:<\/p>\n\n<pre><code>       [0 0 1]   \nres =  [0 0 0] \n       [0 0 0]\n<\/code><\/pre>\n\n<p>Once done, I need to make this operation part of my network, a Tensorflow layer perhaps.<\/p>\n\n<p>I'd appreciate if someone could help me achieve this. Thanks.<\/p>\n\n<p><strong><em>EDIT<\/em><\/strong>\nThis seems to be one way to multiply matrices in a list, but I'm still looking for <code>1)<\/code> slice matrices into multiple parts and multiply them in a better way and <code>2)<\/code> to add this as a layer to a network:<\/p>\n\n<pre><code>tf.scan(lambda a, b: tf.multiply(tf.squeeze(a), tf.squeeze(b)), original)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52802105","CreationDate":"2018-10-14T11:20:02.323","ParentId":null,"OwnerUserId":"200699","Title":null,"Body":"<p>you could use tf.nn.conv2d, manipulating a matrix like this is called a convolution   <\/p>\n\n<p>see tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d <\/p>\n"},{"AnswerId":"52802183","CreationDate":"2018-10-14T11:29:59.533","ParentId":null,"OwnerUserId":"2945357","Title":null,"Body":"<p>You can use numpy array slicing<\/p>\n\n<pre><code>import numpy as np\n\nA = np.array([[0, 0, 1, 1],\n              [0, 0, 1, 1],\n              [0, 0, 0, 0],\n              [0, 0, 0, 0]])\n\nres = A[:-1, :-1] * A[:-1, 1:] * A[1:, :-1] * A[1:, 1:]\n<\/code><\/pre>\n\n<p>and then, perhaps, convert the numpy array to a Tensor object by<\/p>\n\n<pre><code>tf.convert_to_tensor(res)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52801571,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-14T10:05:08.830","AcceptedAnswerId":null,"OwnerUserId":8275126.0,"Title":"LSTM Input 0 of layer basic_lstm_cell_1 in tensorflow","Body":"<p>I am working with lstm using tensor flow when I am running the code it is showing me the error. the code is running fine but when I am running the function <code>tf.nn.dynamic_rnn(lstmCell, data, dtype=tf.float64)<\/code> it is showing Value ERROR<\/p>\n\n<pre><code>import tensorflow as tf\n\nwordsList = np.load('urduwords.npy')\nwordVectors = np.load('urduwordsMatrix.npy')\n\nbatchSize = 24\nlstmUnits = 64\nnumClasses = 2\niterations = 10000\n\ntf.reset_default_graph()\n\n\nlabels = tf.placeholder(tf.float32, [batchSize, numClasses])\ninput_data = tf.placeholder(tf.int32, [batchSize, maxSeqLength])\n\nprint(labels)\n\ndata = tf.Variable(tf.zeros([batchSize, maxSeqLength, numDimensions]),dtype=tf.float32)\nprint(data)\n\n\ndata = tf.nn.embedding_lookup(wordVectors,input_data)\nprint(data)\n\n\nlstmCell = tf.contrib.rnn.BasicLSTMCell(lstmUnits)\nlstmCell = tf.contrib.rnn.DropoutWrapper(cell=lstmCell, output_keep_prob=0.1)\n\nvalue, _ = tf.nn.dynamic_rnn(lstmCell, data, dtype=tf.float64)\n<\/code><\/pre>\n\n<p>How to resolve this error using tensor flow.<\/p>\n\n<pre><code>ValueError: Input 0 of layer basic_lstm_cell_1 is incompatible with the layer: expected ndim=2, found ndim=3. Full shape received: [24, 1, 2]\n<\/code><\/pre>\n\n<p><strong>the shape of the input_data is<\/strong><\/p>\n\n<pre><code>(24, 30, 1, 2)\n<\/code><\/pre>\n\n<p><strong>and the shape of wordVector is<\/strong><\/p>\n\n<pre><code>(24053, 1, 2)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53005953","CreationDate":"2018-10-26T09:45:38.737","ParentId":null,"OwnerUserId":"9992804","Title":null,"Body":"<p>the label shape is 4 dimension because of you feed the wrong type of data to tf,<\/p>\n\n<p>please try to use NumberPy array or List<\/p>\n"},{"AnswerId":"53043442","CreationDate":"2018-10-29T10:25:53.870","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>Since you have not provided a standalone code to reproduce the bug, i have a sample working code as shown below:<\/p>\n\n<pre><code>VOCAB_SIZE = 128\nHIDDEN_SIZE = 200\n\nwordVectors = tf.Variable(tf.random_uniform([VOCAB_SIZE, HIDDEN_SIZE], -1, 1))\n\nlabels = tf.random_normal([batchSize, numClasses])\ninput_data = tf.random_uniform([batchSize, maxSeqLength], maxval=120, dtype=tf.int32)\n\ndata = tf.Variable(tf.zeros([batchSize, maxSeqLength, numDimensions]),dtype=tf.float32)\n\ndata = tf.nn.embedding_lookup(wordVectors,input_data)\n\nlstmCell = tf.contrib.rnn.BasicLSTMCell(lstmUnits)\nlstmCell = tf.contrib.rnn.DropoutWrapper(cell=lstmCell, output_keep_prob=0.1)\n\nvalue, _ = tf.nn.dynamic_rnn(lstmCell, data, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>I have changed the data type of <code>tf.nn.dynamic_rnn<\/code> to <code>tf.float32<\/code> to fix the data type error.<\/p>\n"}]}
+{"QuestionId":52801923,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-14T10:56:42.483","AcceptedAnswerId":52804064.0,"OwnerUserId":2260153.0,"Title":"tf.image.random_flip_left_right() for multiple images together","Body":"<p>I want to randomly flip several input images of different sizes in my input pipeline.\nThe flip needs to be consistent for the different images - either all are flipped or none.\nSince using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/random_flip_left_right\" rel=\"nofollow noreferrer\"><code>tf.image.random_flip_left_right()<\/code><\/a> is inadequate for this, I did this instead:<\/p>\n\n<pre><code>images = [img1, img2]\ndef fliplr(*args):\n    return [tf.image.flip_left_right(t) for t in args]\ndef id(*args):\n    return args\nimg1, img2 = tf.cond(tf.random_uniform([1]) &gt; 0.5, fliplr(images), id(images), name='fliplr')\n<\/code><\/pre>\n\n<p>This is part of a function called by <code>tf.Dataset.map()<\/code> with my input dataset object.<\/p>\n\n<p>However I am getting this error from the <code>flip_left_right<\/code> line:<\/p>\n\n<blockquote>\n  <p>ValueError: Dimension 1 in both shapes must be equal, but are 3 and 1. \n  Shapes are [240,320,3] and [240,320,1].<\/p>\n  \n  <p>From merging shape 0 with other shapes. for 'flip_left_right\/image' (op: 'Pack') with input shapes: [240,320,3], [240,320,1].<\/p>\n<\/blockquote>\n\n<p>I think this is because my images are of different sizes (240x320, either 1 or 3 channels) but I don't understand why this would matter.<\/p>\n\n<p>How can I resolve this error? Alternatively, is there a better way to get what I want that avoids this error?<\/p>\n\n<p>I am using TensorFlow-1.8 (but can upgrade if required).<\/p>\n","answers":[{"AnswerId":"52804064","CreationDate":"2018-10-14T15:10:48.803","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>Just like this?<\/p>\n\n<pre><code>do_flip = tf.random_uniform([]) &gt; 0.5\nimg1 = tf.cond(do_flip, lambda: tf.image.flip_left_right(img1), lambda: img1)\nimg2 = tf.cond(do_flip, lambda: tf.image.flip_left_right(img2), lambda: img2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52802359,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-14T11:54:15.243","AcceptedAnswerId":null,"OwnerUserId":9854132.0,"Title":"Problem with tf.SparseTensor and tf.while_loop","Body":"<p>I face a problem when I try to change the shape of <code>tf.SparseTensor<\/code> inside a <code>tf.while_loop<\/code>. Let's say I have this sparse tensor:<\/p>\n\n<pre><code>indices = np.array([[0, 0], [0, 1], [0, 2], [0, 3], [0, 4], [0, 5],\n               [1, 0], [1, 1], [1, 3], [1, 4], [1, 5],\n               [2, 1], [2, 2], [2, 3], [2, 4],\n               [3, 0], [3, 1], [3, 2], [3, 3], [3, 4], [3, 5],\n               [4, 0], [4, 2], [4, 3], [4, 4], [4, 5]], dtype=np.int64)\n\nvalues = np.array([7, 6, 7, 4, 5, 4,\n              6, 7, 4, 3, 4,\n              3, 3, 1, 1,\n              1, 2, 2, 3, 3, 4,\n              1, 1, 2, 3, 3], dtype=np.float64)\n\ndense_shape = np.array([5, 6], dtype=np.int64)\n\ntRatings = tf.SparseTensor(indices, values, dense_shape)\n<\/code><\/pre>\n\n<p>So, I want to take a slice from the first 3 rows. I know for that purpose I can use <code>tf.sparse_slice<\/code> but this is an example. In my real code, I gather multiple rows from the sparse Tensor which they are not serial. The code I wrote is this:<\/p>\n\n<pre><code>subTensor = tf.sparse_slice(tRatings, [0, 0], [1, 6])\n\ni = tf.constant(1)\ndef condition(i, sub):\n    return tf.less(i, 3)\n\ndef body(i, sub):\n    tempUser = tf.sparse_slice(tRatings, [i, 0], [1, 6])\n    sub = tf.sparse_concat(axis = 0, sp_inputs = [sub, tempUser])\n    return [tf.add(i, 1), sub]\n\nsubTensor = tf.while_loop(condition1, body1, [i, subTensor], shape_invariants=[i.get_shape(), tf.TensorShape([2])])[1] \n<\/code><\/pre>\n\n<p>which does't work for some reason when I run it. I get this:<\/p>\n\n<pre><code>ValueError: Dimensions 1 and 2 are not compatible\n<\/code><\/pre>\n\n<p>According to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop<\/a> it says that:<\/p>\n\n<p>The shape_invariants argument allows the caller to specify a less specific shape invariant for each loop variable, which is needed if the shape varies between iterations. The tf.Tensor.set_shape function may also be used in the body function to indicate that the output loop variable has a particular shape. The shape invariant for SparseTensor and IndexedSlices are treated specially as follows:<\/p>\n\n<p>a) If a loop variable is a SparseTensor, the shape invariant must be TensorShape([r]) where r is the rank of the dense tensor represented by the sparse tensor. It means the shapes of the three tensors of the SparseTensor are ([None], [None, r], [r]). NOTE: The shape invariant here is the shape of the SparseTensor.dense_shape property. It must be the shape of a vector.<\/p>\n\n<p>What am I missing here?<\/p>\n","answers":[{"AnswerId":"52806192","CreationDate":"2018-10-14T19:12:36.453","ParentId":null,"OwnerUserId":"8315956","Title":null,"Body":"<p>There are two problems.\nFirst the problem in Tensorflow code. Change <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py#L593\" rel=\"nofollow noreferrer\">this line<\/a> to:<\/p>\n\n<pre><code>var.indices.set_shape(tensor_shape.TensorShape([None, shape[0]]))\n<\/code><\/pre>\n\n<p>Another small problem in your code. You have to use int64 type for indexing variable:<\/p>\n\n<pre><code>i = tf.constant(1, dtype=tf.int64)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52802370,"AnswerCount":3,"Tags":"<python><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2018-10-14T11:55:22.970","AcceptedAnswerId":52806343.0,"OwnerUserId":4919273.0,"Title":"CNN predicting the same class for all input data","Body":"<p>I am trying the recreate a CNN in Keras to classify point cloud data. The CNN is described in <a href=\"https:\/\/arxiv.org\/pdf\/1710.01408.pdf\" rel=\"nofollow noreferrer\">this<\/a> paper. <\/p>\n\n<p>Network Design\n<a href=\"https:\/\/i.stack.imgur.com\/KApk6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KApk6.png\" alt=\"Image of \"><\/a><\/p>\n\n<p>This is my current implementation:<\/p>\n\n<pre><code>inputs = Input(shape=(None, 3))\n\nx = Conv1D(filters=64, kernel_size=1, activation='relu')(inputs)\nx = BatchNormalization()(x)\nx = Conv1D(filters=64, kernel_size=1, activation='relu')(x)\nx = BatchNormalization()(x)\n\ny = Conv1D(filters=64, kernel_size=1, activation='relu')(x)\ny = BatchNormalization()(y)\ny = Conv1D(filters=128, kernel_size=1, activation='relu')(y)\ny = BatchNormalization()(y)\ny = Conv1D(filters=2048, kernel_size=1, activation='relu')(y)\ny = MaxPooling1D(1)(y)\n\nz = keras.layers.concatenate([x, y], axis=2)\nz = Conv1D(filters=512, kernel_size=1, activation='relu')(z)\nz = BatchNormalization()(z)\nz = Conv1D(filters=512, kernel_size=1, activation='relu')(z)\nz = BatchNormalization()(z)\nz = Conv1D(filters=512, kernel_size=1, activation='relu')(z)\nz = BatchNormalization()(z)\nz = Dense(9, activation='softmax')(z)\n\nmodel = Model(inputs=inputs, outputs=z)\n\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>The problem is that the network predicts the same class for all input data. This may be caused by a mistake in my implementation of the network, overfitting or insufficient training data. Can someone spot a mistake in my implementation?<\/p>\n\n<p><em>Yousefhussien, M., Kelbe, D. J., Ientilucci, E. J., &amp; Salvaggio, C. (2017). A Fully Convolutional Network for Semantic Labeling of 3D Point Clouds. arXiv preprint arXiv:1710.01408.<\/em><\/p>\n","answers":[{"AnswerId":"52806343","CreationDate":"2018-10-14T19:30:24.940","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>The same output class typically indicates a network that has just been initialized, meaning that the training weights are not loaded. Did this same class thing happen during training? Another reason could be bad pre-processing though. Another thing that I noticed is that the paper states \"1D-fully convolutional network\". Your dense layer is a convolutional in the paper.<\/p>\n"},{"AnswerId":"52810859","CreationDate":"2018-10-15T06:29:29.833","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I believe that the mistake is not in the implementation. Most probably the problem is that you have an insufficient amount of data. Also, if network predicts the same class for all input data, it usually means that you lack regularization. Try adding some <code>Dropout<\/code> layers with dropout of <code>0.2<\/code> to <code>0.5<\/code> and see if the results have improved. <\/p>\n\n<p>Also, I don't think that <\/p>\n\n<pre><code>x = Conv1D(filters=64, kernel_size=1, activation='relu')(inputs)\nx = BatchNormalization()(x)\n<\/code><\/pre>\n\n<p>is the same as <\/p>\n\n<pre><code>x = Conv1D(filters=64, kernel_size=1)(inputs)\nx = BatchNormalization()(x)\nx = ReLU(x)\n<\/code><\/pre>\n\n<p>and I think you need the latter.<\/p>\n\n<p>Another thing for you to try is <a href=\"https:\/\/keras.io\/layers\/advanced-activations\/#leakyrelu\" rel=\"nofollow noreferrer\">LeakyReLU<\/a> as it usually gives better results than plain ReLU.<\/p>\n"},{"AnswerId":"52840098","CreationDate":"2018-10-16T16:31:33.207","ParentId":null,"OwnerUserId":"4919273","Title":null,"Body":"<p>The network is fixed as it provides the expected predictions now. Thanks for the help!<\/p>\n\n<p>Based on the answers I changed the following things:<\/p>\n\n<ul>\n<li>The order of the activation and the batch normalization.<\/li>\n<li>The last layer from a dense to a convolutional layer.<\/li>\n<\/ul>\n\n<p>I also added the <code>training=True<\/code> parameter to the batch normalization layer  <\/p>\n\n<p>The code of the correct implementation:<\/p>\n\n<pre><code>inputs = Input(shape=(None, 3))\n\nx = Conv1D(filters=64, kernel_size=1, input_shape=(None, 4))(inputs)\nx = BatchNormalization()(x, training=True)\nx = Activation('relu')(x)\nx = Conv1D(filters=64, kernel_size=1, use_bias=False)(x)\nx = BatchNormalization()(x, training=True)\nx = Activation('relu')(x)\n\ny = Conv1D(filters=64, kernel_size=1)(x)\ny = BatchNormalization()(y, training=True)\ny = Activation('relu')(y)\ny = Conv1D(filters=128, kernel_size=1)(y)\ny = BatchNormalization()(y, training=True)\ny = Activation('relu')(y)\ny = Conv1D(filters=2048, kernel_size=1)(y)\ny = BatchNormalization()(y, training=True)\ny = Activation('relu')(y)\ny = MaxPooling1D(1)(y)\n\nz = keras.layers.concatenate([x, y], axis=2)\nz = Conv1D(filters=512, kernel_size=1)(z)\nz = BatchNormalization()(z, training=True)\nz = Activation('relu')(z)\nz = Conv1D(filters=512, kernel_size=1)(z)\nz = BatchNormalization()(z, training=True)\nz = Activation('relu')(z)\nz = Conv1D(filters=512, kernel_size=1)(z)\nz = BatchNormalization()(z, training=True)\nz = Activation('relu')(z)\nz = Conv1D(filters=2, kernel_size=1, activation='softmax')(z)\n\nmodel = Model(inputs=inputs, outputs=z)\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52803659,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-14T14:26:34.323","AcceptedAnswerId":null,"OwnerUserId":1482046.0,"Title":"Improper accuracy after graph restore","Body":"<p>I've trained a basic CNN on mnist data and has a validation score of around 95% before saving the model after 5 iterations<\/p>\n\n<p>Now my intention is to import saved graph (instead writing again) and restore the parameters from the saved model.<\/p>\n\n<p>All looks good, but when I tried to test the validation accuracy on the restored graph, results doesn't seem to be right. Where am I going wrong?<\/p>\n\n<p><strong>original model<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\",one_hot=True)\n\n\nwith tf.name_scope('inputs'):\n    xs = tf.placeholder(tf.float32, [None, 784], name='xs')\n    ys = tf.placeholder(tf.float32, [None, 10], name='ys')\n    keep_prob = tf.placeholder(tf.float32, name='keep_prob')\n\nx_image = tf.reshape(xs,[-1,28,28,1])\n\nwith tf.variable_scope('conv1',reuse=tf.AUTO_REUSE):\n    W_conv1 = tf.get_variable('weights',dtype=tf.float32, shape=[5,5,1,32], initializer=tf.truncated_normal_initializer(mean=0.0,stddev=0.1))\n    b_conv1 = tf.get_variable('bias', dtype=tf.float32, shape=[32],initializer=tf.constant_initializer(value=0.1))\n    h_conv1 = tf.nn.relu(tf.nn.conv2d(x_image, W_conv1, strides=[1,1,1,1], padding='SAME') + b_conv1,name='conv')\n    h_pool1 = tf.nn.max_pool(h_conv1, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME',name='max_pool')\n\nwith tf.variable_scope('conv2',reuse=tf.AUTO_REUSE):\n    W_conv2 = tf.get_variable('weights',dtype=tf.float32, shape=[5,5,32,64], initializer=tf.truncated_normal_initializer(mean=0.0,stddev=0.1))\n    b_conv2 = tf.get_variable('bias', dtype=tf.float32, shape=[64],initializer=tf.constant_initializer(value=0.1))\n    h_conv2 = tf.nn.relu(tf.nn.conv2d(h_pool1, W_conv2, strides=[1,1,1,1], padding='SAME') + b_conv2,name='conv')\n    h_pool2 = tf.nn.max_pool(h_conv2, ksize=[1,2,2,1], strides=[1,2,2,1], padding = 'SAME',name='max_pool')\n    h_pool2_flatten = tf.reshape(h_pool2, [-1, 7 * 7 * 64])\n\nwith tf.variable_scope('FCLayer_1', reuse=tf.AUTO_REUSE):\n    Weights_1 = tf.get_variable('weights', dtype = tf.float32, shape=[7*7*64, 1024] , initializer=tf.truncated_normal_initializer(mean=0.0,stddev=0.1))\n    Bias_1 = tf.get_variable('bias', dtype=tf.float32, shape=[1024],initializer=tf.constant_initializer(0.1))\n    WxPlusB_1 = tf.nn.relu(tf.add(tf.matmul(h_pool2_flatten,Weights_1), Bias_1),name='WxPlusB')\n    output_1 = tf.nn.dropout(WxPlusB_1, name='dropouts', keep_prob=keep_prob)\n    tf.summary.histogram('FCLayer_1\/weights',Weights_1)\n\n\n\nwith tf.variable_scope('FCLayer_2', reuse=tf.AUTO_REUSE):\n    Weights_2 = tf.get_variable('weights', dtype = tf.float32, shape=[1024, 10] , \n                                initializer=tf.truncated_normal_initializer(mean=0.0,stddev=0.1))\n\n    Bias_2 = tf.get_variable('bias', dtype=tf.float32, shape=[10],initializer=tf.constant_initializer(0.1))\n    predictions = tf.nn.softmax(tf.add(tf.matmul(output_1,Weights_2), Bias_2),name='WxPlusB')\n    tf.summary.histogram('Weights',Weights_2)\n\n\nwith tf.name_scope('accuracy'):\n    accuarcy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(predictions,axis=1), tf.argmax(ys,axis=1)),dtype=tf.float32), name='accuracy')\n    tf.summary.scalar(tensor=accuarcy,name='accuracy')\n\nwith tf.name_scope('loss'):\n    cross_entropy = tf.reduce_mean(-tf.reduce_sum( ys * tf.log(predictions) , axis=1),name='cross_entropy')\n    tf.summary.scalar(tensor=cross_entropy,name='cross_entropy')\n\nwith tf.variable_scope('optimizer',reuse=tf.AUTO_REUSE):\n    optimizer = tf.train.AdamOptimizer(epsilon=1e-4,beta1=0.9,beta2=0.99,name='optimizer')\n\nwith tf.variable_scope('train', reuse=tf.AUTO_REUSE):\n    train = optimizer.minimize(cross_entropy)\n\ninit = tf.global_variables_initializer()\n\nsaver = tf.train.Saver()\n\n\n\nwith tf.Session() as sess:\n\n    try:\n        saver.restore(sess,tf.train.latest_checkpoint('cnn_model'))\n\n    except Exception as err:\n        sess.run(init)\n\n    train_writer = tf.summary.FileWriter('.\/cnnlogs\/train\/',graph=sess.graph)\n    test_writer = tf.summary.FileWriter('.\/cnnlogs\/test', graph=sess.graph)\n    merge_all = tf.summary.merge_all()\n\n    for i in range(50):\n        train_x, train_y = mnist.train.next_batch(100)\n        sess.run(train,feed_dict={xs: train_x, ys : train_y, keep_prob : 0.5})\n        if i % 10 == 0:\n\n            train_result,train_accuarcy,train_loss = sess.run([merge_all, accuarcy, cross_entropy],\n                                                        feed_dict = { xs : train_x, \n                                                                      ys : train_y,\n                                                                      keep_prob : 1.0\n                                                                    })\n            test_result,test_accuarcy,test_loss = sess.run([merge_all, accuarcy, cross_entropy],\n                                                        feed_dict = { xs : mnist.validation.images, \n                                                                      ys : mnist.validation.labels,\n                                                                      keep_prob : 1.0\n                                                                    })\n            train_writer.add_summary(summary=train_result, global_step = i)\n            test_writer.add_summary(summary=test_result, global_step = i)\n\n            print('Train Accuracy: {} Train Loss: {} Test Accuracy: {} Test Loss: {}'.format(train_accuarcy,\n                                                                                             train_loss,\n                                                                                             test_accuarcy,\n                                                                                            test_loss))\n\n    saver.save(sess,'cnn_model\/model.ckpt')\n<\/code><\/pre>\n\n<p><strong>Results<\/strong><\/p>\n\n<pre><code>Train Accuracy: 0.23999999463558197 Train Loss: 7.2988080978393555 **Test Accuracy: 0.1907999962568283** Test Loss: 8.40633773803711\nTrain Accuracy: 0.6299999952316284 Train Loss: 1.0147693157196045 **Test Accuracy: 0.6912000179290771** Test Loss: 1.028536081314087\nTrain Accuracy: 0.8399999737739563 Train Loss: 0.5975856184959412 **Test Accuracy: 0.8307999968528748** Test Loss: 0.6595735549926758\nTrain Accuracy: 0.8700000047683716 Train Loss: 0.36984387040138245 **Test Accuracy: 0.8795999884605408** Test Loss: 0.4149101674556732\nTrain Accuracy: 0.8299999833106995 Train Loss: 0.5599954128265381 **Test Accuracy: 0.9016000032424927** Test Loss: 0.320442616939544\n<\/code><\/pre>\n\n<p><strong>Restored Model<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\",one_hot=True)\n\nwith tf.Session() as sess:\n\n    # resorte the model\n    saver = tf.train.import_meta_graph('cnn_model\/model.ckpt.meta')\n    # resotre the lastest params into model\n    saver.restore(sess,tf.train.latest_checkpoint('cnn_model'))\n\n    graph = tf.get_default_graph()\n\n    xs = graph.get_tensor_by_name('inputs\/xs:0')\n\n    ys = graph.get_tensor_by_name('inputs\/ys:0')\n\n\n    Weights_1 = graph.get_tensor_by_name('FCLayer_1\/weights:0')\n    Weights_2 = graph.get_tensor_by_name('FCLayer_2\/weights:0')\n\n    keep_prob = graph.get_tensor_by_name('inputs\/keep_prob:0')\n\n    accuracy = graph.get_tensor_by_name('accuracy\/accuracy:0')\n    cross_entropy = graph.get_tensor_by_name('loss\/cross_entropy:0')\n\n    test_accuracy , test_loss = sess.run([accuracy, cross_entropy], \n                                        feed_dict={xs : mnist.validation.images,\n                                                   ys : mnist.validation.labels,\n                                                   keep_prob : 1.0})\n\n    print(test_accuracy)\n<\/code><\/pre>\n\n<p><strong>Results<\/strong><\/p>\n\n<pre><code>INFO:tensorflow:Restoring parameters from cnn_model\\model.ckpt\nTest_accuracy: 0.1126   &lt;&lt; Not Right\n<\/code><\/pre>\n\n<p>However if I tried to restore params from the original model itself things seems to work. I'm able to get accuracy > 90% and continue training from where I left last time<\/p>\n\n<p>I've checked the weights and biases in the restored model and seems to be inline to that of the original model. However accuracy, cross_entropies are going for a toss post restoration.<\/p>\n","answers":[]}
+{"QuestionId":52803972,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-14T15:00:59.993","AcceptedAnswerId":52804127.0,"OwnerUserId":10502975.0,"Title":"This issue about keras model and how to compile the model","Body":"<p>I'm trying to create a CNN in pycharm. When I run my code, the console outputs<br>\n<code>RuntimeError: You must compile your model before using it.<\/code><br><\/p>\n\n<p>I write down compile. \nThis is my code: <\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-o\nfrom keras.models import Sequential\nfrom keras.layers import Dense, MaxPool2D, Flatten, Conv2D, Dropout\nfrom keras.preprocessing import image\nfrom keras.optimizers import adadelta\n\ngenerator = image.ImageDataGenerator(\n                rescale=1.\/255,\n                featurewise_center=False,\n                samplewise_center=False,\n                featurewise_std_normalization=False,\n                samplewise_std_normalization=False,\n                zca_whitening=False,\n                rotation_range=10,\n                width_shift_range=0.1,\n                height_shift_range=0.1,\n                horizontal_flip=True,\n                vertical_flip=False,\n               )\n\ndateset = generator.flow_from_directory(\n               shuffle=True,\n               batch_size=100,\n               target_size=(80, 80),\n               directory='\/Users\/Username\/Documents\/Project AI\/Dateset\/blood-cells\/dataset2-master\/images\/TRAIN')\n\n\ndef model():\n    model = Sequential()\n    model.add(Conv2D(80, (3, 3), strides=(1, 1), activation='relu'))\n    model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='relu',     \n              input_shape=(80, 80, 3)))\n    model.add(MaxPool2D(pool_size=(2, 2)))\n    model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='relu'))\n    model.add(Dropout(0.25))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(4, activation='softmax'))\n    model.compile(optimizer=adadelta(lr=0.001),    \n    loss='categorical_crossentropy', metrics=['accuracy'])\n    return model\n\nnn = model()\nnn.fit_generator(dateset,steps_per_epoch=None, epochs=30, verbose=1)\nnn.save('\/Users\/yangzichen\/Documents\/Project AI\/Model.txt')\n<\/code><\/pre>\n","answers":[{"AnswerId":"52804127","CreationDate":"2018-10-14T15:16:24.567","ParentId":null,"OwnerUserId":"7117003","Title":null,"Body":"<p>You seem to overwrite Keras <code>model()<\/code> function with your function. Try this instead:<\/p>\n\n<pre><code>def get_model():  \n    model = Sequential() \n    ...\n    &lt; *rest of your function code here* &gt;\n    ...\n\n    return model\n\nnn = get_model()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52803989,"AnswerCount":2,"Tags":"<pandas><numpy><dataframe><keras>","CreationDate":"2018-10-14T15:02:56.550","AcceptedAnswerId":null,"OwnerUserId":1269281.0,"Title":"How to correct shape of Keras input into a 3D array","Body":"<p>I've a Keras model that when I fit fails with this error<\/p>\n\n<pre><code>&gt; kerasInput = Input(shape=(None, 47))\n&gt; LSTM(..)(kerasInput)\n...\n&gt; model.fit(realInput, ...)\nValueError: Error when checking input: expected input_1 to have 3 dimensions, but got array with shape (10842, 1)\n<\/code><\/pre>\n\n<p>When looking at my input I found it has a shape of <code>(10842, 1)<\/code> but for each row it's actually a list of list. I can verify with<\/p>\n\n<pre><code>&gt; pd.DataFrame(realInput[0]).shape\n(260, 47)\n<\/code><\/pre>\n\n<p>How I could correct my input shape?<\/p>\n\n<p>When trying with keras <code>Reshape<\/code> layer, the creation of the model fails with:<\/p>\n\n<pre><code>Model inputs must come from `keras.layers.Input` (thus holding past layer metadata), they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to your model was not an Input tensor, it was generated by layer reshape_8.\nNote that input tensors are instantiated via `tensor = keras.layers.Input(shape)`.\nThe tensor that caused the issue was: reshape_8\/Reshape:0\n<\/code><\/pre>\n","answers":[{"AnswerId":"52804200","CreationDate":"2018-10-14T15:23:22.210","ParentId":null,"OwnerUserId":"8284987","Title":null,"Body":"<ol>\n<li><p>You can use numpy.expand_dims method to convert the shape to 3D.   <\/p>\n\n<pre><code>import numpy as np\n\nnp.expand_dims(realInput,axis=0)\n<\/code><\/pre><\/li>\n<li><p>Reshape layer keras<\/p>\n\n<p><a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/core\/#reshape<\/a><\/p><\/li>\n<li><p>Use the third parameter as 1<\/p>\n\n<pre><code># Something Similar to this\nX_train = np.reshape(X_train,(X_train.shape[0],X_train.shape[1],1))\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Edit: Added np.reshape method<\/p>\n\n<p>Refer this repository: <a href=\"https:\/\/github.com\/NilanshBansal\/Stock_Price_Prediction\/blob\/master\/Stock_Price_Prediction_20_days_later_4_LSTM.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NilanshBansal\/Stock_Price_Prediction\/blob\/master\/Stock_Price_Prediction_20_days_later_4_LSTM.ipynb<\/a><\/p>\n"},{"AnswerId":"52804900","CreationDate":"2018-10-14T16:45:49.463","ParentId":null,"OwnerUserId":"8113093","Title":null,"Body":"<p>As I said before in the comments. You will need to make sure to reshape your data to match what LSTM expects to receive and also make sure the input_shape is correctly set.<\/p>\n\n<p>I found this post quite helpful when I struggled with inputting to an LSTM layer. I hope it helps you too : <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">Reshape input for LSTM<\/a><\/p>\n"}]}
+{"QuestionId":52804435,"AnswerCount":0,"Tags":"<python><tensorflow><image-processing><filter><convolution>","CreationDate":"2018-10-14T15:51:43.193","AcceptedAnswerId":null,"OwnerUserId":8747033.0,"Title":"Tensorflow multiply broadcasting (convolution before summation)","Body":"<p>I am looking for a tensor flow API which takes a filter and go through a tensor and multiply them and gives me a vector for each element.<\/p>\n\n<p>Let's make it more clear. I have 2D matrix of gray scale image. I have a 2D filter (e.g. 3x3), I am going to roll my filter across the image and get a 9-length vector for each pixel such that each vector's element is corresponding filter element and pixel neighbor multiplication. In other words, I need the result of conv2d before summation. The summation of this vector would be the convolution of filter and corresponding pixel.<\/p>\n\n<p>I can't find any standard API for this operation. I can write a code looping over the elements, but it is not efficient. Does anyone have any idea how to do this?<\/p>\n","answers":[]}
+{"QuestionId":52804669,"AnswerCount":3,"Tags":"<python><tensorflow><raspberry-pi3><tensorflow-lite>","CreationDate":"2018-10-14T16:19:28.310","AcceptedAnswerId":null,"OwnerUserId":993927.0,"Title":"Tensor Flow Lite Raspberry Pi Installation","Body":"<p>I'm looking to install TF Lite on RPi3 for real time image recognition. <\/p>\n\n<p>I've followed: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lite\/g3doc\/rpi.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lite\/g3doc\/rpi.md<\/a><\/p>\n\n<p>All compiled and libtensorflow-lite.a built successfully.<\/p>\n\n<p>Attempting to run the initial example app:<\/p>\n\n<pre><code>python label_image.py\n<\/code><\/pre>\n\n<p>I'm getting:<\/p>\n\n<pre><code>Traceback (most recent call last): \n   File \"label_image.py\", line 23, in\n      &lt;module&gt; import tensorflow as tf\n<\/code><\/pre>\n\n<p>What I'm uncertain about is:<\/p>\n\n<ul>\n<li>What I've missed.<\/li>\n<li>What I need to do to ensure TF uses the lite library and not the full version.<\/li>\n<li>Lastly - doesn't anyone know what FPS TF lite should achieve on a Pi?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"53798797","CreationDate":"2018-12-16T01:49:45.393","ParentId":null,"OwnerUserId":"10794490","Title":null,"Body":"<p>To follow Ishita Shah's review, I am posting here my steps to build\/run the 2 kick-off apps for TfLite running on Raspberry Pi 3 :-)<\/p>\n\n<p>1, Follow <a href=\"https:\/\/www.raspberrypi.org\/learning\/software-guide\/quickstart\/\" rel=\"nofollow noreferrer\">https:\/\/www.raspberrypi.org\/learning\/software-guide\/quickstart\/<\/a>, to install Raspbian onto a Raspberry Pi 3 board. <\/p>\n\n<p><code>#sudo apt install build-essential<\/code>\nFollow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/pi_examples\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/pi_examples<\/a> to install camera. \nVerify if the camera works or not:<\/p>\n\n<p><code>#raspistill -v<\/code><\/p>\n\n<p>2, git clone -b r1.12 <a href=\"https:\/\/github.com\/huaxiaozhong1\/tensorflow.git\" rel=\"nofollow noreferrer\">https:\/\/github.com\/huaxiaozhong1\/tensorflow.git<\/a>. (The best, this is done within the latest docker container of Ubuntu 18.04).\nIt will clone 6 static libraries into \/tensorflow\/tensorflow\/contrib\/lite\/examples\/camera\/libs:\nlibjpeg.a librt.a libv4l1.a libv4l2.a libv4l2rds.a libv4lconvert.a.\nActually they are from Pi environment prepared at step 1:<\/p>\n\n<p>3, In the container of host: <\/p>\n\n<p><code>#apt update<\/code><\/p>\n\n<p><code>#apt upgrade<\/code><\/p>\n\n<p><code>#apt install crossbuild-essential-armhf<\/code><\/p>\n\n<p><code>#cd \/tensorflow<\/code><\/p>\n\n<p><code>#.\/tensorflow\/contrib\/lite\/tools\/make\/download_dependencies.sh<\/code><\/p>\n\n<p><code>#apt install -y libjpeg-dev<\/code><\/p>\n\n<p><code>#apt install libv4l-dev<\/code><\/p>\n\n<p><code>#.\/tensorflow\/contrib\/lite\/tools\/make\/build_rpi_lib.sh<\/code><\/p>\n\n<p>Now, you could find .\/tensorflow\/contrib\/lite\/tools\/make\/gen\/rpi_armv7l\/lib\/libtensorflow-lite.a, \/tensorflow\/tensorflow\/contrib\/lite\/tools\/make\/gen\/rpi_armv7l\/bin\/label_image and \/tensorflow\/tensorflow\/contrib\/lite\/tools\/make\/gen\/rpi_armv7l\/bin\/camera. <\/p>\n\n<p>4, Copy label_image and camera, the Linux executables, to Raspeberry board. \nCopy grace_hopper.bmp from \/tensorflow\/tensorflow\/contrib\/lite\/examples\/label_image\/testdata of host\u2019s container to the board.\nCopy labels.txt from \/tensorflow\/tensorflow\/contrib\/lite\/java\/ovic\/src\/testdata of host\u2019s container to the board.\nDownload Mobilenet_v2_1.0_224_quant.tflite from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/g3doc\/models.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/g3doc\/models.md<\/a> to the board.\nThe 3 data files and 2 Linux executables should be inside the same folder.<\/p>\n\n<p>5, Run the 2 apps one by one. <\/p>\n\n<p>Note: the base branch that I added the 2 apps into is r1.12. Now TfLite has been announced to move out of contrib. But when I ran build_rpi_lib.sh in the latest tensorflow\/tensorflow:nightly-devel, it didn\u2019t get success. So, may I submit the pull-request based at r1.12 right now? I could add the same feature on r1.13+ as soon as it succeeds in building libtensorflow-lite.a.<\/p>\n"},{"AnswerId":"52912891","CreationDate":"2018-10-21T07:06:03.997","ParentId":null,"OwnerUserId":"10535208","Title":null,"Body":"<p>Sorry I did not understand Stack Overflow's rules.<br>\nI modified the comment, but please do delete it if reviewers do not like it.  <\/p>\n\n<p>\"Tensorflow v1.11.0\", which can be introduced by the official pip command, raises problems.  <\/p>\n\n<pre><code>undefined symbol: _ZN6tflite12tensor_utils39NeonMatrixBatchVectorMultiplyAccumulateEPKaiiS2_PKfiPfi\n<\/code><\/pre>\n\n<p>The pip package appears to be damaged. <\/p>\n\n<p>The procedure for building Tensorflow Lite is as follows.<\/p>\n\n<p>1.Execute below.<\/p>\n\n<pre><code>$ sudo apt update;sudo apt upgrade -y\n$ sudo apt-get install -y build-essential openjdk-8-jdk pkg-config \nzip g++ zlib1g-dev unzip\n<\/code><\/pre>\n\n<p>2.Restart Terminal.<\/p>\n\n<p>3.Execute below. (Need to run twice)<\/p>\n\n<pre><code>$ sudo apt-get install -y openjdk-8-jdk\n<\/code><\/pre>\n\n<p>4.Bazel build.<\/p>\n\n<pre><code>$ cd ~\n$ mkdir bazel;cd bazel\n$ wget https:\/\/github.com\/bazelbuild\/bazel\/releases\/download\/0.17.2\/bazel-0.17.2-dist.zip\n$ unzip bazel-0.17.2-dist.zip\n$ nano bazel\/scripts\/bootstrap\/compile.sh\n\nrun \"${JAVAC}\" -classpath \"${classpath}\" -sourcepath \"${sourcepath}\" \\\n-d \"${output}\/classes\" -source \"$JAVA_VERSION\" -target \"$JAVA_VERSION\" \\\n-encoding UTF-8 ${BAZEL_JAVAC_OPTS} \"@${paramfile}\"\n\u2193\nrun \"${JAVAC}\" -classpath \"${classpath}\" -sourcepath \"${sourcepath}\" \\\n-d \"${output}\/classes\" -source \"$JAVA_VERSION\" -target \"$JAVA_VERSION\" \\\n-encoding UTF-8 ${BAZEL_JAVAC_OPTS} \"@${paramfile}\" -J-Xmx500M\n\n$ sudo bash .\/compile.sh\n$ sudo cp output\/bazel \/usr\/local\/bin\n<\/code><\/pre>\n\n<p>5.Prepare for Tensorflow build.<\/p>\n\n<pre><code>$ cd ~\n$ sudo pip2 uninstall tensorflow\n$ sudo pip3 uninstall tensorflow\n$ sudo pip3 install keras_applications==1.0.4 --no-deps\n$ sudo pip3 install keras_preprocessing==1.0.2 --no-deps\n$ sudo pip3 install h5py==2.8.0\n$ git clone -b v1.11.0 https:\/\/github.com\/tensorflow\/tensorflow.git\n$ cd tensorflow\n$ git checkout v1.11.0\n$ .\/tensorflow\/contrib\/lite\/tools\/make\/download_dependencies.sh\n$ .\/tensorflow\/contrib\/lite\/tools\/make\/build_rpi_lib.sh\n$ sudo bazel build tensorflow\/contrib\/lite\/toco:toco\n<\/code><\/pre>\n\n<p>6-1.Tensorflow's configure. (Python2.x)<\/p>\n\n<pre><code>$ .\/configure\nWARNING: Running Bazel server needs to be killed, because the startup options are different.\nWARNING: --batch mode is deprecated. Please instead explicitly shut down your Bazel server using the command \"bazel shutdown\".\nYou have bazel 0.17.2- (@non-git) installed.\nPlease specify the location of python. [Default is \/usr\/bin\/python]: \nFound possible Python library paths:\n\/usr\/local\/lib\/python2.7\/dist-packages\n\/usr\/local\/lib\n\/usr\/lib\/python2.7\/dist-packages\n\/opt\/movidius\/caffe\/python\nPlease input the desired Python library path to use.  Default is [\/usr\/local\/lib\/python2.7\/dist-packages]\n\nDo you wish to build TensorFlow with jemalloc as malloc support? [Y\/n]: n\nNo jemalloc as malloc support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with Google Cloud Platform support? [Y\/n]: n\nNo Google Cloud Platform support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with Hadoop File System support? [Y\/n]: n\nNo Hadoop File System support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with Amazon AWS Platform support? [Y\/n]: n\nNo Amazon AWS Platform support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with Apache Kafka Platform support? [Y\/n]: n\nNo Apache Kafka Platform support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with XLA JIT support? [y\/N]: n\nNo XLA JIT support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with GDR support? [y\/N]: n\nNo GDR support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with VERBS support? [y\/N]: n\nNo VERBS support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with nGraph support? [y\/N]: n\nNo nGraph support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with OpenCL SYCL support? [y\/N]: n\nNo OpenCL SYCL support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with CUDA support? [y\/N]: n\nNo CUDA support will be enabled for TensorFlow.\n\nDo you wish to download a fresh release of clang? (Experimental) [y\/N]: n\nClang will not be downloaded.\n\nDo you wish to build TensorFlow with MPI support? [y\/N]: n\nNo MPI support will be enabled for TensorFlow.\n\nPlease specify optimization flags to use during compilation when bazel option \"--config=opt\" is specified [Default is -march=native]: \n\n\nWould you like to interactively configure .\/WORKSPACE for Android builds? [y\/N]: n\nNot configuring the WORKSPACE for Android builds.\n\nPreconfigured Bazel build configs. You can use any of the below by adding \"--config=&lt;&gt;\" to your build command. See tools\/bazel.rc for more details.\n--config=mkl            # Build with MKL support.\n--config=monolithic     # Config for mostly static monolithic build.\nConfiguration finished\n<\/code><\/pre>\n\n<p>6-2.Tensorflow's configure. (Python3.x)<\/p>\n\n<pre><code>Please specify the location of python. [Default is \/usr\/bin\/python]: \/usr\/bin\/python3\n\n\nFound possible Python library paths:\n  \/usr\/local\/lib\n  \/usr\/lib\/python3\/dist-packages\n  \/usr\/local\/lib\/python3.5\/dist-packages\nPlease input the desired Python library path to use.  Default is [\/usr\/local\/lib] \/usr\/local\/lib\/python3.5\/dist-packages\n<\/code><\/pre>\n\n<p>7.Tensorflow's build.<\/p>\n\n<pre><code>$ sudo bazel build --config opt --local_resources 1024.0,0.5,0.5 \\\n--copt=-mfpu=neon-vfpv4 \\\n--copt=-ftree-vectorize \\\n--copt=-funsafe-math-optimizations \\\n--copt=-ftree-loop-vectorize \\\n--copt=-fomit-frame-pointer \\\n--copt=-DRASPBERRY_PI \\\n--host_copt=-DRASPBERRY_PI \\\n\/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>8.Tensorflow's install.<\/p>\n\n<pre><code>$ sudo .\/bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n$ sudo pip2 install \/tmp\/tensorflow_pkg\/tensorflow-1.11.0-cp27-cp27mu-linux_armv7l.whl\nOr\n$ sudo pip3 install \/tmp\/tensorflow_pkg\/tensorflow-1.11.0-cp35-cp35m-linux_armv7l.whl\n<\/code><\/pre>\n\n<p>9.\u3010Required\u3011 Restart the terminal.<\/p>\n\n<p>10.Execute below.<\/p>\n\n<pre><code>$ sudo apt install -y python-scipy python3-scipy\n$ sudo apt remove openjdk-8-*\n$ sudo apt purge openjdk-8-*\n<\/code><\/pre>\n\n<p>If done so far the problem should have been solved.<br>\nHowever, it takes 27 hours in total.<br>\nIf you do not wait for 27 hours, you can use cross build or drastically reduce waiting time using the proposed wheel file.  <\/p>\n\n<p>Pre-build Tensorflow binary(wheel) is below.<br>\n<a href=\"https:\/\/github.com\/PINTO0309\/Tensorflow-bin\" rel=\"nofollow noreferrer\">https:\/\/github.com\/PINTO0309\/Tensorflow-bin<\/a><\/p>\n"},{"AnswerId":"52905929","CreationDate":"2018-10-20T13:00:27.633","ParentId":null,"OwnerUserId":"2535521","Title":null,"Body":"<blockquote>\n  <p>What I've missed.<\/p>\n<\/blockquote>\n\n<p>To install Tensorflow on Raspberry Pi, You can use official pip package<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>For more detail visit official documentation page -<\/p>\n\n<pre><code>https:\/\/www.tensorflow.org\/install\/pip\n<\/code><\/pre>\n\n<p>Or if you want Cross compiled package, you can follow this official documentation page -<\/p>\n\n<pre><code>https:\/\/www.tensorflow.org\/install\/source_rpi\n<\/code><\/pre>\n\n<blockquote>\n  <p>What I need to do to ensure TF uses the lite library and not the full version.<\/p>\n<\/blockquote>\n\n<p>You just need a trained model and convert it to TFLite model. For complete detail you can follow this tutorial -<\/p>\n\n<pre><code>https:\/\/medium.com\/tensorflow\/training-and-serving-a-realtime-mobile-object-detector-in-30-minutes-with-cloud-tpus-b78971cf1193\n<\/code><\/pre>\n\n<blockquote>\n  <p>Lastly - doesn't anyone know what FPS TF lite should achieve on a Pi?<\/p>\n<\/blockquote>\n\n<p>I am currently working on it, but as there is some issue in tensorflow's current official build for Raspberry pi (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23082\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23082<\/a>), I am not able to calculate the performance on Pi.<\/p>\n"}]}
+{"QuestionId":52805414,"AnswerCount":0,"Tags":"<image><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2018-10-14T17:46:18.377","AcceptedAnswerId":null,"OwnerUserId":9768291.0,"Title":"Super Resolution: convolutions leaving traces on edges","Body":"<p>So I've been working on a Super-Resolution model and I've achieved decent results. However, I have remnants of the network left in the output image: I couldn't feed the whole image as input for memory reasons, which meant I had to chop an image as multiple smaller parts (in my case, <code>128x128<\/code>).<\/p>\n\n<p>The edges of these chops are creating lines for some reason. My question is: <strong>what is the reason and how to prevent that?<\/strong><\/p>\n\n<p>My first thought is that the convolutions on the edge don't have as much information as the rest since padding is required for filters of size exceeding <code>1x1<\/code> so it's normal and unavoidable to have such lines without post-processing. What might be a good post-processing approach?<\/p>\n\n<p>Or would more training eventually solve that issue?<\/p>\n\n<p>Another idea I've had is to feed overlapping regions of the image to the network and to replace the edges of one output with the corresponding non-edge part of the overlap-output.<\/p>\n\n<p>Here is the model I'm using:<\/p>\n\n<pre><code>def setUpModel(x_train, y_train):\n\n    filters = 256\n    kernel_size = 3\n    strides = 1\n    scale_fact = 4\n    res_blocks = 3\n    img_depth = 3\n\n\n    # Head module\n    input = Input(shape=(None, None, img_depth))  # None = can take inputs of different sizes\n    conv0 = Conv2D(filters, kernel_size, strides=strides, padding='same')(input)\n\n    # Body module\n    res = Conv2D(filters, kernel_size, strides=strides, padding='same')(conv0)\n    act = ReLU()(res)\n    res = Conv2D(filters, kernel_size, strides=strides, padding='same')(act)\n    res_rec = Add()([conv0, res])\n\n    for i in range(res_blocks):\n        res1 = Conv2D(filters, kernel_size, strides=strides, padding='same')(res_rec)\n        act  = ReLU()(res1)\n        res2 = Conv2D(filters, kernel_size, strides=strides, padding='same')(act)\n        res_rec = Add()([res_rec, res2])\n\n    conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(res_rec)\n    add  = Add()([conv0, conv])\n\n    # Tail module\n    conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(add)\n    act  = ReLU()(conv)\n    up   = UpSampling2D(size=scale_fact if scale_fact != 4 else 2)(act)  # TODO: try \"Conv2DTranspose\"\n\n    # When it's a 4X factor, we want the upscale split in two procedures\n    if(scale_fact == 4):\n        conv = Conv2D(filters, kernel_size, strides=strides, padding='same')(up)\n        act  = ReLU()(conv)\n        up   = UpSampling2D(size=2)(act)  # TODO: try \"Conv2DTranspose\"\n\n    output = Conv2D(filters=3,\n                    kernel_size=1,\n                    strides=1,\n                    padding='same')(up)\n\n    model = Model(inputs=input, outputs=output)\n\n    return model\n<\/code><\/pre>\n\n<p>Here are two examples:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9bJL3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9bJL3.png\" alt=\"HR1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Y1JNu.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Y1JNu.jpg\" alt=\"traces1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yU3yh.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yU3yh.jpg\" alt=\"HR2\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5T9U1.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5T9U1.jpg\" alt=\"traces2\"><\/a><\/p>\n\n<p><strong>EDIT<\/strong>: As requested, here is the <code>train<\/code> code:<\/p>\n\n<pre><code>def train(model, x_train, y_train):\n    print(\"Training is starting.\")\n\n    if load_model == False:\n        print(\"Compiling the model since it wasn't loaded from memory.\")\n        optimizer = Adadelta(lr=1.0,\n                             rho=0.95,\n                             epsilon=None,\n                             decay=0.0)\n        model.compile(optimizer=optimizer,\n                      loss='mean_squared_error')\n\n    model.fit(x_train,\n              y_train,\n              epochs=6,  # TODO: is it multi-fold testing?\n              verbose=verbosity,\n              shuffle=False,\n              validation_split=0.1,\n              batch_size=7,\n              callbacks=get_callbacks())\n<\/code><\/pre>\n\n<p>I used to have a <code>batch_size<\/code> of 1 and would train on samples of <code>128x128<\/code> (larger would bust my memory) but that gave me very long training times. I eventually settled for a batch size of 7 and training input size of <code>64x64<\/code>.<\/p>\n\n<p>The dataset was so large that I had to split it in multiple sub-parts because I couldn't load it all into memory hence the low epochs value (I manually changed the input's source folder before launching the training session again on the saved model).<\/p>\n\n<p><code>get_callbacks()<\/code> returns <code>[]<\/code> because some callbacks that I wanted to try using were causing memory problems and slowing down the training time too much (TensorBoad, for example).<\/p>\n\n<p>Here is the <a href=\"https:\/\/github.com\/payne911\/SR-ResCNN-Keras-\" rel=\"nofollow noreferrer\">repo<\/a> if you're ever interested in seeing more code to understand.<\/p>\n","answers":[]}
+{"QuestionId":52806467,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><slice>","CreationDate":"2018-10-14T19:43:54.233","AcceptedAnswerId":null,"OwnerUserId":6115719.0,"Title":"Understand parameters of tf.slice in tensorflow and why I cannot change it","Body":"<p>Below is my code:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport matplotlib.image as mp_image\nfilename = \"abc.jpeg\"\ninput_image = mp_image.imread(filename)\nmy_image=tf.placeholder(\"uint8\",[None, None, 3])\nmyimage=tf.placeholder(\"uint8\",[None, None, 3])\nslice1=tf.slice(my_image,[0,100,0],[300,400,-1]) #[x,y,?],[x,y,?]\nwith tf.Session() as sess:\n    result = sess.run(slice1,feed_dict={my_image: input_image})\nprint(result.shape)\nplt.imshow(result)\nplt.show()\n<\/code><\/pre>\n\n<p>In slice 1 what does the parameters passed as list indicate[x,y,?],[x,y,?]. <\/p>\n\n<p>In  tf.slice(image_tensor,[0,0,0],[100,200,-1]).What does 0 and -1 stands for here and why i cannot change them?<\/p>\n","answers":[{"AnswerId":"52807553","CreationDate":"2018-10-14T22:17:16.050","ParentId":null,"OwnerUserId":"7723040","Title":null,"Body":"<p><code>[x,y,?],[x,y,?]<\/code> in your case, the third parameter of the shape is a number of image channels<\/p>\n\n<p>For answer on a second question lets take a look how <code>tf.slice<\/code> working, in case of the image with RGB channels it looks like <code>tf.slice([from_x, from_y, from_channel], [to_x, to_y, to_channel])<\/code> also in shape definition you can use <code>-1<\/code> this way you telling sensor flow slice to maximum available value. From your code sample you trying a slice input image <code>[0,100, 0],[300,400, 3]<\/code>, you can change the third param and it is a valid code, but should remember <a href=\"https:\/\/matplotlib.org\/api\/_as_gen\/matplotlib.pyplot.imshow.html\" rel=\"nofollow noreferrer\">matplotlib<\/a> allows you to show pictures only <code>(M, N), (M, N, 3), (M, N, 4)<\/code> <\/p>\n\n<p>Code explanation:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport tensorflow as tf\nimport numpy as np\nimport matplotlib.image as mp_image\n\nfilename = \"abc.png\"\ninput_image = mp_image.imread(filename)\nmy_image = tf.placeholder(\"float32\",[None, None, 3])\nmyimage = tf.placeholder(\"float32\",[None, None, 3])\n\n# this way you slicing RGB image\nslice1=tf.slice(my_image,[0,100,0],[300, 400, -1])\n\nwith tf.Session() as sess:\n    result = sess.run(slice1, feed_dict={my_image: input_image})\n\nplt.subplot(1, 2, 1)\nplt.imshow(result)\n\n# this way you slicing image and keep just R channel\nslice1=tf.slice(my_image,[0,100,0],[300, 400, 1]) #[x,y,?],[x,y,?]\n\nwith tf.Session() as sess:\n    result = sess.run(slice1, feed_dict={my_image: input_image})\n\nplt.subplot(1, 2, 2)\n# matplotlib imshow show need extra options to render image with single chanel\nplt.imshow(np.reshape(result, (result.shape[0], result.shape[1])), cmap='gray')\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zOuSw.png\" rel=\"nofollow noreferrer\">result image<\/a><\/p>\n"},{"AnswerId":"52824852","CreationDate":"2018-10-15T21:13:44.847","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>Looking at the docstring for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice\" rel=\"nofollow noreferrer\">tf.slice<\/a>, the parameters are <code>input_<\/code>, <code>begin<\/code> and <code>size<\/code> respectively. Your code is doing<\/p>\n\n<pre><code>slice1=tf.slice(my_image,[x_begin,y_begin,channel_begin],[x_size,y_size,channel_size])\n<\/code><\/pre>\n\n<p>Note that the third parameter describes the <code>size<\/code> and not the absolute <code>to<\/code> index (@XMANX is mistaken). The <code>size<\/code> parameter accepts a sentinel value of -1, which means that all remaining elements in the dimension are included in the slice.<\/p>\n\n<p>For example, if you had a tensor <code>t<\/code> with shape <code>[X, Y, Z]<\/code><\/p>\n\n<pre><code>tf.slice(t, [x_begin, y_begin, z_begin], [x_size, y_size, z_size])\n<\/code><\/pre>\n\n<p>is equivalent to doing<\/p>\n\n<pre><code>t[x_begin : x_begin+x_size, y_begin : y_begin+y_size, z_begin : z_begin+z_size]\n<\/code><\/pre>\n\n<p>In order to extract just the R channel from an image, you would do something like:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport matplotlib.image as mp_image\n\nfilename = \"abc.jpeg\"\ninput_image = mp_image.imread(filename)\nmy_image=tf.placeholder(\"uint8\",[None, None, 3])\n# Doesn't slice along the x and y dimensions, but takes only one channel\nsliced = tf.slice(my_image,[0, 0, 0], [-1, -1, 1])\nsqueezed = tf.squeeze(slice)  # Removes last dimension\nwith tf.Session() as sess:\n    result = sess.run(squeezed,feed_dict={my_image: input_image})\nplt.imshow(result)\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52807152,"AnswerCount":0,"Tags":"<tensorflow><neural-network><quantization>","CreationDate":"2018-10-14T21:14:08.463","AcceptedAnswerId":null,"OwnerUserId":7718773.0,"Title":"tf.fake_quant_with_min_max_vars is a differentiable function?","Body":"<p>Quantization schemes are generally non-differentiable because they pass through the threshold, such as round or sign function. It means that we can not get the gradient of trainable variables due to the nature of chain rule. <\/p>\n\n<p>Instead, we can use a trick called <strong>'straight-through-estimator'<\/strong>, which enable us to back-propagating the gradient of individual trainable variables.<\/p>\n\n<p>One such method is <strong>tf.fake_quant_with_min_max_vars<\/strong>, The advantages of this format are that it can represent arbitrary magnitudes of ranges, they don\u2019t have to be symmetrical, it can represent signed and unsigned values, and the linear spread makes doing multiplications straightforward.<a href=\"https:\/\/petewarden.com\/2016\/05\/03\/how-to-quantize-neural-networks-with-tensorflow\/\" rel=\"nofollow noreferrer\">Blog<\/a>, <a href=\"https:\/\/arxiv.org\/abs\/1712.05877\" rel=\"nofollow noreferrer\">Paper<\/a>.<\/p>\n\n<p>So, my question is, can we differentiate the fake_quant function? And if so, does this function apply 'straight-through-estimator'? <\/p>\n\n<p>I did a little bit of this with some snippet code<\/p>\n\n<blockquote>\n  <p>x = tf.cast(np.random.normal(0,1,(10,10), tf.float32)<\/p>\n  \n  <p>x_q = tf.fake_quant_with_min_max_vars(x, min=tf.reduce_min(x), max=tf.reduce_max(x), num_bits=3)<\/p>\n  \n  <p>grad = tf.gradients(x_q, x)<\/p>\n<\/blockquote>\n\n<p>In that case, almost every grad have value 1(i.e, gradient 1), which means it pass through the gradient itself.<\/p>\n\n<p>However, sometimes a few samples have gradient 0, or other constant, such as 2, 3, 4...<\/p>\n\n<p>Am I missing what's going on?<\/p>\n","answers":[]}
+{"QuestionId":52807664,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-14T22:37:45.177","AcceptedAnswerId":null,"OwnerUserId":5252733.0,"Title":"How to run tensorflow optimization with different initial conditions?","Body":"<p>I am relatively new to tensorflow. Is it possible to set up a tensorflow optimization and run the optimization with different initial conditions, without re-running the set-up code?<\/p>\n\n<p>For example, consider the following psuedo code:<\/p>\n\n<pre><code>def foo(N):\n    # do some tensor math\n    return A\n\nN = tf.get_variable(\"N\", initializer = N_init)\nA = foo(N)\n\ncost = tf.losses.mean_squared_error(A, target)\noptimizer = tf.contrib.opt.ScipyOptimizerInterface(cost)\ninit = tf.global_variables_initializer()\nsess.run(init)\noptimizer.minimize(sess)  \n<\/code><\/pre>\n\n<p>My function foo takes a long time to run, and I get different results depending on the initial condition. I would like to set up the problem, and then run the optimization with different initializations for my variable N without re-running the function foo more than once. Can I do this? How?<\/p>\n","answers":[]}
+{"QuestionId":52808148,"AnswerCount":2,"Tags":"<tensorflow><computer-vision><conv-neural-network><object-detection><object-detection-api>","CreationDate":"2018-10-15T00:02:08.207","AcceptedAnswerId":52883047.0,"OwnerUserId":3042444.0,"Title":"How to count objects detected in an image using Tensorflow?","Body":"<p>I have trained my custom object detector using <code>faster_rcnn_inception_v2<\/code>, tested it using <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\" rel=\"nofollow noreferrer\">object_detection_tutorial.ipynb<\/a> and it works perfect, I can find bounding boxes for the objects inside the test image, my problem is how can I actually count the number of those bounding boxes or simply I want to count the number of objects detected for each class.<\/p>\n","answers":[{"AnswerId":"52818627","CreationDate":"2018-10-15T14:08:44.667","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>Because of low reputations I can not comment.<\/p>\n\n<p>As far as I know the object detection API unfortunately has no built-in function for this. <\/p>\n\n<p>You have to write this function by yourself. I assume you run the <code>eval.py<\/code> for evaluation!? To access the individual detected objects for each image you have to follow the following chain of scripts: <\/p>\n\n<p><code>eval.py<\/code> -> <code>evaluator.py<\/code> -><code>object_detection_evaluation.py<\/code> -> <code>per_image_evaluation.py<\/code><\/p>\n\n<p>In the last script you can count the detected objects and bounding boxes per image. You just have to save the numbers and sum them up over your entire dataset.<\/p>\n\n<p>Does this already help you?<\/p>\n"},{"AnswerId":"52883047","CreationDate":"2018-10-18T21:44:40.227","ParentId":null,"OwnerUserId":"3042444","Title":null,"Body":"<p>I solved this using <a href=\"https:\/\/github.com\/ahmetozlu\/tensorflow_object_counting_api\" rel=\"nofollow noreferrer\">Tensorflow Object Counting API<\/a>. We have an example of counting objects in an image using <code>single_image_object_counting.py<\/code>. I just replaced <code>ssd_mobilenet_v1_coco_2017_11_17<\/code> with my own model containing inference graph<\/p>\n\n<pre><code>input_video = \"image.jpg\"\ndetection_graph, category_index = backbone.set_model(MODEL_DIR)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52808654,"AnswerCount":1,"Tags":"<pytorch><pooling>","CreationDate":"2018-10-15T01:31:27.363","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"Pytorch adaptive_avg_pool2d algorithm","Body":"<p>Anyone knows the algorithm for pytorch adaptive_avg_pool2d, for example,<br><\/p>\n\n<pre><code>adaptive_avg_pool2d(image,[14,14])\n<\/code><\/pre>\n\n<p>so question:\n    I want to do the same in keras neural network, for any give inputs, want to get 14*14 output. Any suggestions?<\/p>\n","answers":[{"AnswerId":"55973932","CreationDate":"2019-05-03T16:33:59.923","ParentId":null,"OwnerUserId":"10517687","Title":null,"Body":"<p>I don't think this exists in Keras. You could get the dimension of your input and divide by 14 to get the desired pool_size.<\/p>\n\n<p>For example if your inputs are 28x28 you can use:<\/p>\n\n<p><code>keras.layers.AveragePooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None)<\/code><\/p>\n"}]}
+{"QuestionId":52809336,"AnswerCount":0,"Tags":"<python><python-3.x><keras>","CreationDate":"2018-10-15T03:20:01.583","AcceptedAnswerId":null,"OwnerUserId":9045339.0,"Title":"Possible to use keras \u201cflow_from_directory\u201d with \u201c.mat\u201d files?","Body":"<p>I have built a CNN to predict lymph node positivity (has cancer or not). Right now to  load the data I have a self defined function that loads a batch of data and feeds it to the model for training. <\/p>\n\n<p>Instead of loading batches I would love to use the <code>flow_from_directory<\/code> method. The problem I have is that my data are saved as arrays [#, rows, width, height, PET or CT] not images (that would later be converted to arrays). Example [0,:,:,:,0] is volume sized 48x48x32 from a ct image. <\/p>\n\n<p>If I try to use <code>flow_from_directory<\/code> I get 0 images with 3 classes which I expected since '.mat' is not a recognized file (<a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/362fe9f8daf556151328eb5d02bd5ae638c653b8\/keras_preprocessing\/image.py#L1868\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/362fe9f8daf556151328eb5d02bd5ae638c653b8\/keras_preprocessing\/image.py#L1868<\/a>). Interestingly it doesnt raise any errors but I am indefinitely stuck on 1\/150 epochs.  I am going to see if I can write my own <code>flow_from_directory<\/code>. Not sure if someone has run across this problem and could give my pointers.<\/p>\n\n<p><strong>Illustrating how data is combined<\/strong><\/p>\n\n<pre><code>for fname in fnames:\n     data = scipy.io.loadmat(os.path.join(dir_in_train, fname))['roi_patch']\n     data_PET =  scipy.io.loadmat(os.path.join(dir_in_train_PET, fname))['roi_patch']\n     train_combo[0,:,:,:,0]=data\/4.0950\n     train_combo[0,:,:,:,1]=data_PET\/32.1959\n     train_combo[0,:,:,:,:].shape\n     train_combo = np.zeros((1, 48, 48, 32, 2))\n\n     scipy.io.savemat(fname, {fname: train_combo})\n<\/code><\/pre>\n\n<p>This will create a file ex '1.mat' that has CT data and PET data in one area<\/p>\n\n<p>Then I have code changing it into npy files.<\/p>\n\n<p><strong>Example of data generator I already have<\/strong><\/p>\n\n<pre><code># load training data\n    def load_train_data_batch_generator(self, batch_size=32, rows_in=48, cols_in=48,  zs_in=32, \n                                        channels_in=2, num_classes=3, \n                                        dir_in_train=None, dir_out_train=None):\n\n        # dir_in_train = main_dir + '\/test_CT_PET_combo'\n\n        fnames = ['{}.mat'.format(i) for i in range(1,len(os.listdir(dir_in_train))+1)]\n\n        y_train = np.zeros((batch_size, num_classes))\n        x_train = np.zeros((batch_size, rows_in, cols_in, zs_in, channels_in))\n\n        while True:\n            count = 0\n            for fname in np.random.choice(fnames, batch_size, replace=False):\n\n                data_label = scipy.io.loadmat(os.path.join(dir_out_train, fname))['output']\n\n                # changing one hot encoding to integer\n                integer_label = np.argmax(data_label[0], axis=0)\n                y_train[count,:] = data_label\n\n                # Loading train ct w\/ c and pet\/ct combo that will be saved into new directory\n                train_combo = scipy.io.loadmat(os.path.join(dir_in_train, fname))[fname]\n                x_train[count,:,:,:,:] = train_combo\n\n                count += 1\n\n            yield(x_train, y_train)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52809802,"AnswerCount":0,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2018-10-15T04:32:55.017","AcceptedAnswerId":null,"OwnerUserId":8671854.0,"Title":"Tensorflow custom estimator - Training accuracy & loss remains constant","Body":"<p>I developed a simple tensorflow custom estimator for training dynamic RNN with cloud-ml. While testing the model, I can see that the training\/eval accuracy and loss remains constant. I tried playing around with learning rate, increased data sizes, used a few different optimizers as well tried alternate options for defining model functions. But nothing works. While I am not certain, it appears to me that model is not training at all given that the accuracy and loss don't change at all during the training. It appears to me that the training operation is not executing, although I couldn't find why.<\/p>\n\n<p>Appreciate if someone can point whether I am missing anything in the estimator function. Below is the code I wrote for the estimator:<\/p>\n\n<p>Thanks<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>import tensorflow as tf\r\nfrom util import OUTPUT_FEATURE\r\n\r\ndef metric_fn(probabilities, labels):\r\n    \"\"\"\r\n    \"\"\"\r\n    pred_model = tf.cast(tf.argmax(probabilities, 1), tf.int32)\r\n    actual_model = tf.cast(tf.argmax(labels, 1), tf.int32)\r\n    \r\n    metrics = {}\r\n    metrics['accuracy'] = tf.metrics.accuracy(pred_model, actual_model)\r\n    # metrics['auroc'] = tf.metrics.auc(labels, probabilities)\r\n\r\n    return metrics\r\n\r\n\r\ndef RNN_estimator(hyper_parms, config):\r\n    \"\"\"\r\n    \"\"\"\r\n    num_classes = len(hyper_parms.label_dict)\r\n    hidden_units = hyper_parms.hidden_units\r\n    learning_rate = hyper_parms.learning_rate\r\n    lables = hyper_parms.label_dict\r\n    \r\n    def _RNNModel(features):\r\n        \"\"\" \r\n        \"\"\"\r\n        input_layer = features[OUTPUT_FEATURE]\r\n\r\n        # Create hidden RNN layers given the input layer\r\n        hidden_layers = [tf.nn.rnn_cell.LSTMCell(size) for size in hidden_units]\r\n        multi_rnn_cell = tf.nn.rnn_cell.MultiRNNCell(hidden_layers)\r\n        outputs, state = tf.nn.dynamic_rnn(cell=multi_rnn_cell,\r\n                                           inputs=input_layer,\r\n                                           dtype=tf.float32)\r\n\r\n        logits = tf.layers.dense(outputs[:,-1], num_classes)\r\n\r\n        return logits\r\n\r\n    def _train_op_fn(loss):\r\n        \"\"\"\r\n        \"\"\"\r\n        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\r\n        train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step())\r\n\r\n        return train_op\r\n\r\n    def _model_fn(features, labels, mode):\r\n        \"\"\" \r\n        \"\"\"\r\n        \r\n        logits = _RNNModel(features)\r\n        probabilities = tf.nn.softmax(logits)\r\n        predict_output = tf.cast(tf.argmax(probabilities, 1), tf.int32)\r\n        predict_output = tf.gather(lables, predict_output)\r\n                \r\n        if (mode == tf.estimator.ModeKeys.TRAIN \r\n        or mode == tf.estimator.ModeKeys.EVAL):\r\n            loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(\r\n                                labels=labels, logits=logits)) \r\n            \r\n            evalmetrics = metric_fn(probabilities, labels)     \r\n            \r\n            if mode == tf.estimator.ModeKeys.TRAIN:\r\n                train_op =  _train_op_fn(loss)\r\n            else:\r\n                train_op = None\r\n        else:\r\n            train_op = None\r\n            loss=None\r\n            evalmetrics=None\r\n            \r\n        if (mode == tf.estimator.ModeKeys.PREDICT\r\n            or mode == tf.estimator.ModeKeys.EVAL): \r\n            predictions = {'class': predict_output,\r\n                           'score': tf.reduce_max(probabilities, axis=1)\r\n                          }\r\n            export_outputs = {'prediction': tf.estimator.export.PredictOutput(predictions)\r\n                             }\r\n        else:\r\n            predictions = None\r\n            export_outputs = None\r\n            \r\n        return tf.estimator.EstimatorSpec(mode=mode, \r\n                                          predictions=predictions,\r\n                                          loss=loss,\r\n                                          train_op=train_op,\r\n                                          export_outputs=export_outputs,\r\n                                          eval_metric_ops=evalmetrics)\r\n\r\n    return tf.estimator.Estimator(model_fn=_model_fn,\r\n                                  config=config,\r\n                                  warm_start_from=None)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[]}
+{"QuestionId":52810404,"AnswerCount":1,"Tags":"<tensorflow><categorical-data><continuous>","CreationDate":"2018-10-15T05:49:18.390","AcceptedAnswerId":null,"OwnerUserId":7711685.0,"Title":"How to mix categorical data that requires one-hot encoding with continuous data in tensorflow?","Body":"<p>I have a mix of categorical inputs and continuous inputs. Categorical inputs are not binary, so one-hot encoding is required. How can you set such a situation up for tensorflow?<\/p>\n","answers":[{"AnswerId":"52812520","CreationDate":"2018-10-15T08:24:03.073","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>You can use feature columns. The Tensorflow docs are the best place to start searching for your answer:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/www.tensorflow.org\/guide\/feature_columns\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/feature_columns<\/a><\/p>\n<\/blockquote>\n\n<p>To create feature columns, call functions from the tf.feature_column module. This document explains nine of the functions in that module. As the following figure shows, all nine functions return either a Categorical-Column or a Dense-Column object, except bucketized_column, which inherits from both classes:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/c43au.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/c43au.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To combine the result of your categorical variables with the numerical variables you can use a \"deep and wide\" model combining the two type of features:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/DNNLinearCombinedClassifier\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/DNNLinearCombinedClassifier<\/a><\/p>\n<\/blockquote>\n\n<pre><code>for each column in dnn_feature_columns + linear_feature_columns:\n- if column is a _CategoricalColumn, a feature with key=column.name whose value is a SparseTensor.\n- if column is a _WeightedCategoricalColumn, two features: the first with key the id column name, the second with key the weight column name. Both features' value must be a SparseTensor.\n- if column is a _DenseColumn, a feature with key=column.name whose value is a Tensor.\n<\/code><\/pre>\n"}]}
+{"QuestionId":52810702,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning><conv-neural-network>","CreationDate":"2018-10-15T06:16:25.433","AcceptedAnswerId":52814849.0,"OwnerUserId":4404709.0,"Title":"Modify layer parameters in Keras","Body":"<p>I am interested in updating existing layer parameters in Keras (not removing a layer and inserting a new one instead, rather just modifying existing parameters).<\/p>\n\n<p>I will give an example of a function I'm writing:<\/p>\n\n<pre><code>def add_filters(self, model):\n    conv_indices = [i for i, layer in enumerate(model.layers) if 'convolution' in layer.get_config()['name']]\n    random_conv_index = random.randint(0, len(conv_indices)-1)\n    factor = 2\n    conv_layer = model.layers[random_conv_index]\n    conv_layer.filters = conv_layer.filters * factor\n    print('new conv layer filters after transform is:', conv_layer.filters)\n    print('just to make sure, its:', model.layers[random_conv_index].filters)\n    return model\n<\/code><\/pre>\n\n<p>so what's basically happening here is me taking a random convolutional layer from my network (all my conv layers have 'convolution' in their name) and trying to double the filters. As far as I know this shouldn't cause any 'compilation issues' with input\/output size compatibility in any case.<\/p>\n\n<p>The thing is, my model doesn't change at all. The 2 print-outs I added in the end print the correct number (double the previous amount of filters). But when I compile the model and print model.summary(), I still see the previous filter amount.<\/p>\n\n<p>BTW, I'm not constricted to Keras. If anyone has an idea how to pull this off with PyTorch for example I'll also buy it :D<\/p>\n","answers":[{"AnswerId":"52814849","CreationDate":"2018-10-15T10:38:46.690","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Well, if you would like to create the architecture of a new model based on an existing model, though with some modifications, you can use <code>to_json<\/code> and <code>model_from_json()<\/code> functions. Here is an example:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(10, (3,3), input_shape=(100,100,3)))\nmodel.add(Conv2D(40, (3,3)))\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>Model summary:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_12 (Conv2D)           (None, 98, 98, 10)        280       \n_________________________________________________________________\nconv2d_13 (Conv2D)           (None, 96, 96, 40)        3640      \n=================================================================\nTotal params: 3,920\nTrainable params: 3,920\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Now we modify the number of filters of the first layer and create a new model based on the modified architecture:<\/p>\n\n<pre><code>from keras.models import model_from_json\n\nmodel.layers[0].filters *= 2\nnew_model = model_from_json(model.to_json())\nnew_model.summary()\n<\/code><\/pre>\n\n<p>New model summary:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_12 (Conv2D)           (None, 98, 98, 20)        560       \n_________________________________________________________________\nconv2d_13 (Conv2D)           (None, 96, 96, 40)        7240      \n=================================================================\nTotal params: 7,800\nTrainable params: 7,800\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>You can also modify the output of <code>model.to_json()<\/code> directly without modifying the model instance.<\/p>\n\n<hr>\n\n<p>You can easily use <code>get_weights()<\/code> method to get the current weights of the convolution layer. It would return a list of two numpy arrays. The first one corresponds to filter weights and the second one corresponds to bias parameters. Then you can use <code>set_weights()<\/code> method to set the new weights:<\/p>\n\n<pre><code>conv_layer = model.layers[random_conv_index]\nweights = conv_layer.get_weights()\nweights[0] *= factor  # multiply filter weights by `factor`\nconv_layer.set_weights(weights)\n<\/code><\/pre>\n\n<p>As a side note, the <code>filters<\/code> attribute of a convolution layer which you have used in your code corresponds to the number of filters in this layer and not their weights.<\/p>\n"}]}
+{"QuestionId":52811542,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-15T07:18:38.703","AcceptedAnswerId":52812258.0,"OwnerUserId":3392589.0,"Title":"when trying to run tf.confusion matrix it gives end of sequence error","Body":"<p>I am preparing my dataset using new tensoflow input pipeline, here is my code:<\/p>\n\n<pre><code>train_data = tf.data.Dataset.from_tensor_slices(train_images)\ntrain_labels = tf.data.Dataset.from_tensor_slices(train_labels)\ntrain_set = tf.data.Dataset.zip((train_data,train_labels)).shuffle(500).batch(30)\n\nvalid_data = tf.data.Dataset.from_tensor_slices(valid_images)\nvalid_labels = tf.data.Dataset.from_tensor_slices(valid_labels)\nvalid_set = tf.data.Dataset.zip((valid_data,valid_labels)).shuffle(200).batch(20)\n\ntest_data = tf.data.Dataset.from_tensor_slices(test_images)\ntest_labels = tf.data.Dataset.from_tensor_slices(test_labels)\ntest_set = tf.data.Dataset.zip((test_data, test_labels)).shuffle(200).batch(20)\n\n# create general iterator\niterator = tf.data.Iterator.from_structure(train_set.output_types, train_set.output_shapes)\nnext_element = iterator.get_next()\ntrain_init_op = iterator.make_initializer(train_set)\nvalid_init_op = iterator.make_initializer(valid_set)\ntest_init_op  = iterator.make_initializer(test_set)\n<\/code><\/pre>\n\n<p>Now I wanted to create a confusion matrix for validation set of my CNN model after training, here is what I try to do:<\/p>\n\n<pre><code>sess.run(valid_init_op)\nvalid_img, valid_label = next_element\nfinalprediction = tf.argmax(train_predict, 1)\nactualprediction = tf.argmax(valid_label, 1)\nconfusion_matrix = tf.confusion_matrix(labels=actualprediction,predictions=finalprediction,\n                                       num_classes=num_classes,dtype=tf.int32,name=None, weights=None)\nprint(sess.run(confusion_matrix, feed_dict={keep_prob: 1.0}))\n<\/code><\/pre>\n\n<p>In this way it creates confusion matrix but only for one batch of validation set. for that I tried to collect all validation set batches in list and then use the list for creating confusion matrix:<\/p>\n\n<pre><code>val_label_list = []    \nsess.run(valid_init_op)\nfor i in range(valid_iters):\n    while True:\n          try:\n              elem = sess.run(next_element[1])\n              val_label_list.append(elem)\n          except tf.errors.OutOfRangeError:\n              print(\"End of append.\")\n          break\nval_label_list = np.array(val_label_list)\nval_label_list = val_label_list.reshape(40,2)\n<\/code><\/pre>\n\n<p>and now the <code>val_label_list<\/code> contain the labels for all batches of my validation set and I can use it to create confusion matrix:<\/p>\n\n<pre><code>finalprediction = tf.argmax(train_predict, 1)\nactualprediction = tf.argmax(val_label_list, 1)\nconfusion = tf.confusion_matrix(labels=actualprediction,predictions=finalprediction, \n                    num_classes=num_classes, dtype=tf.int32,name=\"Confusion_Matrix\")\n<\/code><\/pre>\n\n<p>But now when I want to run the confusion matrix and print it:<\/p>\n\n<pre><code>print(sess.run(confusion, feed_dict={keep_prob: 1.0}))\n<\/code><\/pre>\n\n<p>it gives me an error:<\/p>\n\n<pre><code>OutOfRangeError: End of sequence\n     [[Node: IteratorGetNext_5 = IteratorGetNext[output_shapes=[[?,10,32,32], [?,2]], output_types=[DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Iterator_5)]]\n<\/code><\/pre>\n\n<p>anyone can tell me how to deal with this error? or any other solution that solve my original problem?<\/p>\n","answers":[{"AnswerId":"52812258","CreationDate":"2018-10-15T08:07:15.937","ParentId":null,"OwnerUserId":"8380638","Title":null,"Body":"<p>The problem is related with the graph flow execution.\nLook at this line:<\/p>\n\n<pre><code>print(sess.run(confusion, feed_dict={keep_prob: 1.0}))\n<\/code><\/pre>\n\n<p>You are running the graph for getting the 'confusion' value. So all dependent nodes will be also executed. Then:<\/p>\n\n<pre><code>finalprediction = tf.argmax(train_predict, 1)\nactualprediction = tf.argmax(val_label_list, 1)\nconfusion = tf.confusion_matrix(...)\n<\/code><\/pre>\n\n<p>I guess that your call to train_predict will try to obtain a new element from the training iterator which has been already completely iterated and after this the error is triggered.<\/p>\n\n<p>You should compute the confusion matrix directly in the loop and accumulate the results in a variable.<\/p>\n\n<pre><code>sess.run(valid_init_op)\nconfusion_matrix = np.zeros(n_labels,n_labels)\nwhile True:\n      try:\n          conf_matrix = sess.run(confusion)\n          confusion_matrix += conf_matrix\n      except tf.errors.OutOfRangeError:\n          print(\"End of append.\")\n      break\n<\/code><\/pre>\n"}]}
+{"QuestionId":52811553,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-10-15T07:19:15.770","AcceptedAnswerId":52891974.0,"OwnerUserId":5566910.0,"Title":"Encode uint16 png with tfrecord (depth_data)","Body":"<p>In order to <strong>encode the uint16 png with tfrecord<\/strong>, I implemented the following codes. However, it outputs \"uint8\", how to modify it? Thank you~<\/p>\n\n<pre><code>def _bytes_list_feature(values):\n\"\"\"Returns a TF-Feature of bytes.\n\nArgs:\nvalues: A string.\n\nReturns:\nA TF-Feature.\n\"\"\"\n  def norm2bytes(value):\n    if isinstance(value, str) and six.PY3:\n      print \"endcode\"\n      return value.encode()\n    else:\n      # print value\n      return value\n\n  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[norm2bytes(values)]))\n\n\ndef image_to_tfexample(depth_data):\n  \"\"\"Converts depth to tf example.\n\n  Args:\n    depth_data: string of depth data.\n\n  Returns:\n    tf example of depth.\n  \"\"\"\n  return tf.train.Example(features=tf.train.Features(feature={\n    'image\/depth\/encoded': (\n      _bytes_list_feature(depth_data)),\n    'image\/depth\/format': _bytes_list_feature(\n      FLAGS.depth_format),\n  }))\n<\/code><\/pre>\n\n<p>The following Decode codes are used to specify how the TF-Examples are decoded. The decoder's dtype is uint8, while the dtype of source one is uint16. <strong>How to encode and decode uint16 image with tfrecord.<\/strong>   <\/p>\n\n<pre><code>keys_to_features = {\n  'image\/depth\/encoded': tf.FixedLenFeature(\n      (), tf.string, default_value=''),\n  'image\/depth\/format': tf.FixedLenFeature(\n      (), tf.string, default_value='png'),\n}\nitems_to_handlers = {\n  'depth': tfexample_decoder.Image(\n      image_key='image\/depth\/encoded',\n      format_key='image\/depth\/format',\n      channels=1),\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"52888582","CreationDate":"2018-10-19T08:28:20.323","ParentId":null,"OwnerUserId":"1052083","Title":null,"Body":"<p>Not an expert but I had a quick look at the code. It looks like tfexample_decoder.Image only supports uint8 for now. You probably need to update TF code to achieve your goal<\/p>\n\n<p>image.decode_image operator supports both uint8 and uint16 image, however, tfexample_decoder.Image didn't pass it in. <\/p>\n\n<p>It might just work if you simply pass in dtype to decode_image<\/p>\n\n<p>Refs:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/26f66133f86ed6f7371e0cbc2b9bd4f30bf16e93\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py#L420\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/26f66133f86ed6f7371e0cbc2b9bd4f30bf16e93\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py#L420<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/26f66133f86ed6f7371e0cbc2b9bd4f30bf16e93\/tensorflow\/python\/ops\/image_ops_impl.py#L1846\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/26f66133f86ed6f7371e0cbc2b9bd4f30bf16e93\/tensorflow\/python\/ops\/image_ops_impl.py#L1846<\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"52891974","CreationDate":"2018-10-19T12:02:54.933","ParentId":null,"OwnerUserId":"5566910","Title":null,"Body":"<p>In tfexample_decoder.py, I make some changes as follows, and it works.<\/p>\n\n<ol>\n<li>image_ops.decode_image(image_buffer, channels=self._channels, dtype=self._dtype)<\/li>\n<li>replace image = control_flow_ops.case(pred_fn_pairs, default=check_jpeg, exclusive=True) with image = decode_image()<\/li>\n<\/ol>\n\n<p><strong>complete code<\/strong><\/p>\n\n<pre><code>def _decode(self, image_buffer, image_format):\n\"\"\"Decodes the image buffer.\n\nArgs:\n  image_buffer: The tensor representing the encoded image tensor.\n  image_format: The image format for the image in `image_buffer`. If image\n    format is `raw`, all images are expected to be in this format, otherwise\n    this op can decode a mix of `jpg` and `png` formats.\n\nReturns:\n  A tensor that represents decoded image of self._shape, or\n  (?, ?, self._channels) if self._shape is not specified.\n\"\"\"\n\ndef decode_image():\n  \"\"\"Decodes a image based on the headers.\"\"\"\n  return image_ops.decode_image(image_buffer, channels=self._channels, **dtype=self._dtype**)\n\ndef decode_jpeg():\n  \"\"\"Decodes a jpeg image with specified '_dct_method'.\"\"\"\n  return image_ops.decode_jpeg(\n      image_buffer, channels=self._channels, dct_method=self._dct_method)\n\ndef check_jpeg():\n  \"\"\"Checks if an image is jpeg.\"\"\"\n  # For jpeg, we directly use image_ops.decode_jpeg rather than decode_image\n  # in order to feed the jpeg specify parameter 'dct_method'.\n  return control_flow_ops.cond(\n      image_ops.is_jpeg(image_buffer),\n      decode_jpeg,\n      decode_image,\n      name='cond_jpeg')\n\ndef decode_raw():\n  \"\"\"Decodes a raw image.\"\"\"\n  return parsing_ops.decode_raw(image_buffer, out_type=self._dtype)\n\npred_fn_pairs = {\n    math_ops.logical_or(\n        math_ops.equal(image_format, 'raw'),\n        math_ops.equal(image_format, 'RAW')): decode_raw,\n}\n\n**# image = control_flow_ops.case(\n#     pred_fn_pairs, default=check_jpeg, exclusive=True)\n\nimage = decode_image()**\n\nimage.set_shape([None, None, self._channels])\nif self._shape is not None:\n  image = array_ops.reshape(image, self._shape)\n\nreturn image\n<\/code><\/pre>\n"}]}
+{"QuestionId":52811935,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-10-15T07:43:53.853","AcceptedAnswerId":null,"OwnerUserId":258274.0,"Title":"With pytorch module,will computation graph be overrode when forward() was invoke multiple times before backward()?","Body":"<p>I found a line of code:<\/p>\n\n<pre><code>out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)    \n<\/code><\/pre>\n\n<p>in:<a href=\"https:\/\/github.com\/liorshk\/facenet_pytorch\/blob\/master\/train_triplet.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/liorshk\/facenet_pytorch\/blob\/master\/train_triplet.py<\/a><\/p>\n\n<p>as you can see that the \"forward()\" is invoked multiple times before \"backward()\", question are:<\/p>\n\n<p>1, Will next invoking override previous computation graph?<\/p>\n\n<p>2, In my test, the GPU consumption will increase accordingly, so even overriding happened,GPUu memory leaking happens, how to solve this problem?<\/p>\n","answers":[{"AnswerId":"52820207","CreationDate":"2018-10-15T15:37:33.137","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>No, the graph is not over-ridden by different calls to forward.\nEvery time you call forward, computations are added to the graph and hence gpu consumption increases.<\/p>\n\n<p>When you call <code>model.zero_grad()<\/code> graphs are cleared.  <\/p>\n"}]}
+{"QuestionId":52812417,"AnswerCount":0,"Tags":"<python><machine-learning><keras><artificial-intelligence>","CreationDate":"2018-10-15T08:17:28.543","AcceptedAnswerId":null,"OwnerUserId":6342070.0,"Title":"A `Concatenate` layer should be called on a list of at least 2 inputs","Body":"<p>I am trying to create two sequential models in Keras for CNN deep learning and merge(concatenate) these two models before adding it into the dense layer. But I am getting the error:\n    A <code>Concatenate<\/code> layer should be called on a list of at least 2 inputs<\/p>\n\n<pre><code>    model_1 = models.Sequential()\n    model_1.add(layers.Conv1D(num_filters, 7, activation='relu',\n        input_shape=(TEXT_MAX_LENGTH, LENGTH_ALPHABET)))\n    model_1.add(layers.MaxPooling1D(pool_size=3))\n    model_1.add(layers.Flatten())\n\n    model_2 = models.Sequential()\n    model_2.add(layers.Conv1D(num_filters, 7, activation='relu',\n        input_shape=(TEXT_MAX_LENGTH, LENGTH_ALPHABET)))\n    model_2.add(layers.Conv1D(num_filters, 7, activation='relu'))\n    model_2.add(layers.MaxPooling1D(3))\n    model_2.add(layers.Flatten())\n\n    concat = Concatenate([model_1, model_2]) \n\n    merged_model = models.Sequential()      \n\n    model.add(concat)\n\n\n    model.add(layers.Dense(width_hidden, activation='relu'))\n    model.add(layers.Dropout(rate=dropout))\n    model.add(layers.Dense(width_output, activation='softmax'))\n\n    model.compile(loss='categorical_crossentropy',\n           optimizer='adam',\n           metrics=['accuracy'])\n\n    model.fit(x_train, y_train,\n       batch_size=batch_size,\n       epochs=epochs,\n       verbose=1,\n       callbacks=callbacks_list,\n       validation_data=(x_test, y_test)\n       )\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52812479,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-15T08:21:28.593","AcceptedAnswerId":52884799.0,"OwnerUserId":1450609.0,"Title":"Is there any way to use tf.train.Checkpoint with MonitoredTrainingSession?","Body":"<p>Now I am trying to use tf.train.Checkpoint instead of tf.train.Saver in various situations.<\/p>\n\n<p>When I tried to use Checkpoint with MonitoredTrainingSession, I noticed \n that Checkpoint.save changes tf.Graph. Therefore, I cannot use Checkpoint as follows.<\/p>\n\n<pre><code>a = tf.Variable([3], name=\"a\")\ncheckpoint = tf.train.Checkpoint(model=a)\n\nwith tf.train.MonitoredTrainingSession() as sess:\n    checkpoint.save(file_prefix=\"parameter\/ckpt\", session=sess)\n<\/code><\/pre>\n\n<p>I got an error \"RuntimeError: Graph is finalized and cannot be modified\".<\/p>\n\n<p>Can I use Checkpoint with MonitoredTrainingSession?<\/p>\n","answers":[{"AnswerId":"52884799","CreationDate":"2018-10-19T01:48:52.720","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>I found a weird technique, which call checkpoint.write with a temporary session to construct a graph preliminarily.<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    checkpoint.save(\".\/parameter\/sample\")\n\nwith tf.train.MonitoredTrainingSession() as sess:\n    checkpoint.save(file_prefix=\"parameter\/ckpt\", session=sess)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52812739,"AnswerCount":0,"Tags":"<python><keras><package><progress-bar><tqdm>","CreationDate":"2018-10-15T08:39:33.380","AcceptedAnswerId":null,"OwnerUserId":3599147.0,"Title":"Disimporting a package in python","Body":"<p>I'm working with tensorflow.keras and tqdm packages at the same time. But they have some conflict. I have updated both but the problem has not been solved. The problem is that when I import tqdm, the progress bar of Keras does not work properly. So, I want to dis-import(!) tqdm when I'm done with that.<\/p>\n\n<p>Is there any solution?<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p><a href=\"http:\/\/s8.picofile.com\/file\/8340155718\/31.PNG\" rel=\"nofollow noreferrer\">a simple code with this issue<\/a><\/p>\n\n<pre><code>#importing packages and data\n\nimport tensorflow as tf\nfrom tqdm import tqdm as q\n(x_train, y_train), (x_test, y_test)=tf.keras.datasets.mnist.load_data()\n\n# a simple use of tqdm, a package for showing progressbar\n\nfor i in q(range(1000)):\n    pass\n\n# making a simple model for neural network using keras\n\nmodel=tf.keras.Sequential()\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(10,activation='softmax'))\nmodel.compile('adam','sparse_categorical_crossentropy',metrics=['accuracy'])\nmodel.fit(x_train,y_train,validation_split=0.1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52812815,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><mnist>","CreationDate":"2018-10-15T08:44:52.243","AcceptedAnswerId":52813961.0,"OwnerUserId":7490161.0,"Title":"What does keras normalize axis argument does?","Body":"<p>I am a beginner in deep learning and I am working upon the mnist dataset in keras.<\/p>\n\n<p>I used normalization as<\/p>\n\n<pre><code>tf.keras.utils.normalize(x_train, axis = 1)\n<\/code><\/pre>\n\n<p>I don't understand what does the axis argument means. Can you help me out with this?<\/p>\n","answers":[{"AnswerId":"52813961","CreationDate":"2018-10-15T09:47:37.697","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>The normalize function just performs a regular normalization to improve performance:<\/p>\n\n<blockquote>\n  <p>Normalization is a rescaling of the data from the original range so\n  that all values are within the range of 0 and 1.<\/p>\n<\/blockquote>\n\n<p>There is a nice explanation of the axis argument in another post:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/stackoverflow.com\/questions\/47435526\/what-is-the-meaning-of-axis-1-in-keras-argmax\">What is the meaning of axis=-1 in keras.argmax?<\/a><\/p>\n<\/blockquote>\n\n<p>For example:<\/p>\n\n<blockquote>\n  <p>Your data has some shape (19,19,5,80). This means:<\/p>\n  \n  <ul>\n  <li>Axis = 0 - 19 elements <\/li>\n  <li>Axis = 1 - 19 elements <\/li>\n  <li>Axis = 2 - 5 elements<\/li>\n  <li>Axis = 3 - 80 elements<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Also, for those who want to go deeper, there is an explanation from Fran\u00e7ois Chollet - Keras\u2019 author- on GitHub:<\/p>\n\n<blockquote>\n  <ul>\n  <li>For Dense layer, all RNN layers and most other types of layers, the\n  default of axis=-1 is what you should use, <\/li>\n  <li>For Convolution2D layers\n  with dim_ordering=\u201cth\u201d (the default), use axis=1, <\/li>\n  <li>For Convolution2D\n  layers with dim_ordering=\u201ctf\u201d, use axis=-1 (i.e. the default).<\/li>\n  <\/ul>\n  \n  <p><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1921\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/1921<\/a><\/p>\n<\/blockquote>\n"},{"AnswerId":"52814620","CreationDate":"2018-10-15T10:25:40.333","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>keras.utils.normalize()<\/code> function <a href=\"https:\/\/stackoverflow.com\/a\/52571871\/2099607\">calls<\/a> the <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.linalg.norm.html\" rel=\"nofollow noreferrer\"><code>numpy.linalg.norm()<\/code><\/a> to compute the norm and then normalize the input data. The given <code>axis<\/code> argument is therefore passed to <code>norm()<\/code> function to compute the norm along the given axis.<\/p>\n"}]}
+{"QuestionId":52813031,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-10-15T08:56:13.380","AcceptedAnswerId":null,"OwnerUserId":6526922.0,"Title":"Tensorflow object detection (ssd - mobilnet_v1) not converging","Body":"<p>I am trying to train deep-fashion dataset on the tensorflow object detection api. The categories and corresponding number of images in each class are as follows.<\/p>\n\n<pre><code>{'Blouse': 24562, 'Cutoffs': 1669, 'Cardigan': 13311, 'Button-Down': 330, 'Poncho': 791, 'Chinos': 527, 'Top': 10078, 'Anorak': 160, 'Kimono': 2294, 'Blazer': 7497, 'Robe': 150, 'Parka': 676, 'Jodhpurs': 45, 'Sarong': 32, 'Shorts': 19624, 'Caftan': 54, 'Turtleneck': 146, 'Leggings': 5013, 'Jersey': 727, 'Hoodie': 4010, 'Culottes': 486, 'Sweater': 13123, 'Flannel': 324, 'Jeggings': 594, 'Tank': 15418, 'Onesie': 70, 'Coat': 2120, 'Sweatpants': 3021, 'Henley': 716, 'Jacket': 10425, 'Gauchos': 49, 'Sweatshorts': 1098, 'Romper': 7408, 'Joggers': 4405, 'Trunks': 386, 'Halter': 34, 'Dress': 144316, 'Jeans': 14152, 'Tee': 73333, 'Coverup': 34, 'Capris': 154, 'Kaftan': 252, 'Peacoat': 194, 'Bomber': 618, 'Skirt': 29546, 'Jumpsuit': 12306}\n<\/code><\/pre>\n\n<p>There are 46 classes in total.\nMy total loss graph doesn't seem to go to lower values. The best till now has been around 4. <a href=\"https:\/\/i.stack.imgur.com\/09LSc.png\" rel=\"nofollow noreferrer\">This<\/a> is my graph.\nIs this normal for so many classes or should I use another algorithm ?<\/p>\n\n<p>Another doubt I have is that the model has been training for around 3 days now. I don't have a GPU. Should I stop the training process ?<\/p>\n\n<p>Edit:<\/p>\n\n<pre><code>train_config: {\n  batch_size: 24\n  optimizer {\n    rms_prop_optimizer: {\n      learning_rate: {\n        exponential_decay_learning_rate {\n          initial_learning_rate: 0.004\n          decay_steps: 800720\n          decay_factor: 0.95\n        }\n      }\n      momentum_optimizer_value: 0.9\n      decay: 0.9\n      epsilon: 1.0\n    }\n  }\n<\/code><\/pre>\n\n<p>Hardware Information: <\/p>\n\n<pre><code>{\nArchitecture:        x86_64\nCPU op-mode(s):      32-bit, 64-bit\nByte Order:          Little Endian\nCPU(s):              6\nOn-line CPU(s) list: 0-5\nThread(s) per core:  1\nCore(s) per socket:  2\nSocket(s):           3\nNUMA node(s):        1\nVendor ID:           GenuineIntel\nCPU family:          6\nModel:               63\nModel name:          Intel(R) Xeon(R) CPU E5-2683 v3 @ 2.00GHz\nStepping:            2\nCPU MHz:             1997.686\nBogoMIPS:            3995.37\nHypervisor vendor:   VMware\nVirtualization type: full\nL1d cache:           32K\nL1i cache:           32K\nL2 cache:            256K\nL3 cache:            35840K\nNUMA node0 CPU(s):   0-5\nFlags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon nopl xtopology tsc_reliable nonstop_tsc pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm fsgsbase tsc_adjust bmi1 avx2 smep bmi2 invpcid xsaveopt arat\n}\n<\/code><\/pre>\n\n<p>I am new to tensorflow. Any kind of help will be appreciated.\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"52813871","CreationDate":"2018-10-15T09:42:49.577","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<ol>\n<li>The value of the total loss is not very informative unless you give the exact loss you're using, and also then it's not indicative unless you have experimented with this loss. Instead, a more informative values are the performance metrics of your model on your test set.<\/li>\n<li>Time is also not very indicative unless you state the hardware you're using. Instead - state the number of iterations and batch size, and the size of your training data or the number of epochs you trained the model.<\/li>\n<li>Have you used a decaying learning rate? If you use a constant learning rate: (1) with large value - then it will less probable to converge; (2) with small value - it may take a long time to converge.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":52813084,"AnswerCount":1,"Tags":"<python><tensorflow><histogram><tensorboard><summary>","CreationDate":"2018-10-15T08:59:03.710","AcceptedAnswerId":null,"OwnerUserId":5165721.0,"Title":"How to draw weights histogram on tensorboard?","Body":"<p>I have a <code>train.py<\/code> and a <code>model.py<\/code> for 3D classification deep learning model. I define the model in model.py. In train.py, I import model from model.py and start training.<\/p>\n\n<p>I visualized training loss or accuracy, by using <code>tf.summary.scalar<\/code> function in train.py.<\/p>\n\n<p>However, I'm having trouble when I am trying to plot weight's histogram.\nI find <code>tf.summary.histogram<\/code> function, but I can't figure out where to put this function (in train.py or model.py?).<\/p>\n\n<p>In model.py, I defined layers using <code>tf.contrib.layers<\/code> function.\n(sorry for my poor English skills)<\/p>\n\n<p><strong>PS<\/strong>:<br>\nI saved my model as <code>.ckpt<\/code> format. Can I plot weight histogram by loading this saved model? <\/p>\n","answers":[{"AnswerId":"53947247","CreationDate":"2018-12-27T15:20:27.203","ParentId":null,"OwnerUserId":"9437711","Title":null,"Body":"<p>I'll try to answer each of your questions separately. <\/p>\n\n<blockquote>\n  <p>I'm having trouble when I am trying to draw weight's histogram.<\/p>\n<\/blockquote>\n\n<p>Ideally, you define <em>weights<\/em> and <em>biases<\/em> using <code>tf.get_variable<\/code> and add them to summary histograms. <\/p>\n\n<p>For example: <\/p>\n\n<pre><code>with tf.name_scope(\"layer1\"):\n    W1 = tf.get_variable(\"W1\", shape=[input_size, hidden_layer_neurons],\n                         initializer=tf.contrib.layers.xavier_initializer())\n    layer1 = tf.matmul(X, W1)\n    layer1_act = tf.nn.tanh(layer1)\n    tf.summary.histogram(\"weights\", W1)\n<\/code><\/pre>\n\n<p>If needed, we can also add histograms of layer-outputs and activation-outputs:    <\/p>\n\n<pre><code>    tf.summary.histogram(\"layer-outputs\", layer1)\n    tf.summary.histogram(\"activation-outputs\", layer1_act)\n<\/code><\/pre>\n\n<p>But since you're using <code>tf.contrib.layers<\/code>, you don't have such a provision as <code>contrib.layers<\/code> takes care of creating weights and biases for you. In such a case, you can have a look at <code>tf.trainable_variables()<\/code>; This should contain all the trainable variables from your graph, that contains all the weights and biases of the network. <\/p>\n\n<p>You can define a simple function like this:<\/p>\n\n<pre><code>def add_hist(train_vars):\n    for i in train_vars:\n        name = i.name.split(\":\")[0]\n        value = i.value()\n        tf.summary.histogram(name, value)\n<\/code><\/pre>\n\n<blockquote>\n  <p>I can't figure out where to put this function (in train.py? or\n  model.py?)<\/p>\n<\/blockquote>\n\n<p>Scalar summaries like loss and accuracy are obtained during training, hence those are included in <code>train.py<\/code>; Whereas weights and biases are a part of your core-model, and are hence to be included in <code>model.py<\/code><\/p>\n\n<p>So in your <code>model.py<\/code>, include this before you use <code>tf.summary.merge_all()<\/code><\/p>\n\n<pre><code>train_vars = tf.trainable_variables()\nadd_hist(train_vars)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Can I draw weight histogram by load this saved model?<\/p>\n<\/blockquote>\n\n<p>These histograms generally show the distribution of entities (weights or activations etc) during the training. These plots are often mainly used for insights on how these distributions are changing over time, if or not the parameters or their gradients are saturated, any clear steps to be taken to improve etc. <\/p>\n\n<p>Since your saved checkpoint (.ckpt) is only supposed to contain the final weights and biases, only the final distribution is obtained and not the histogram plot over the entire training period.<\/p>\n\n<p>Hope this helps. <\/p>\n"}]}
+{"QuestionId":52813369,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-15T09:14:58.697","AcceptedAnswerId":null,"OwnerUserId":6533075.0,"Title":"ImportError: libcusolver.so.9.0: cannot open shared object file: No such file or directory","Body":"<p>I've a Tensorflow installation (1.7 GPU) and CUDA 9.0. This setup used to work but after an upgrade yesterday things have stopped working. I am not sure what exactly had changed, but I suspect the installation is somehow broken. <\/p>\n\n<p>I did uninstall and reinstall Tensorflow but it does not work anymore, throwing the error of:<\/p>\n\n<pre><code>ImportError: libcusolver.so.9.0: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>My <code>.bashrc<\/code> file has the following:<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0\nexport PATH=${PATH}:\/usr\/local\/cuda-9.0\/bin\nexport CUDA_HOME=${CUDA_HOME}:\/usr\/local\/cuda:\/usr\/local\/cuda-9.0\nexport LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/lib\/\n<\/code><\/pre>\n\n<p>which I am not sure if correct or not. <\/p>\n\n<p>I suspect this is some .bashrc problem but I am not able to solve it.<\/p>\n\n<p>Has anyone faced the same or similar?<\/p>\n","answers":[{"AnswerId":"52813859","CreationDate":"2018-10-15T09:41:35.580","ParentId":null,"OwnerUserId":"7431458","Title":null,"Body":"<p>Probably right with .bashrc problem. <\/p>\n\n<p>Try adding these to your .bashrc:<\/p>\n\n<pre><code>export PATH=\/usr\/local\/cuda-9.0\/bin${PATH:+:{PATH}}\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-9.0\/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}\n<\/code><\/pre>\n"}]}
+{"QuestionId":52814880,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-10-15T10:41:04.593","AcceptedAnswerId":null,"OwnerUserId":6194504.0,"Title":"Neuron freezing in Tensorflow","Body":"<p>I need to implement <strong>neurons freezing<\/strong> in CNN for a deep learning research,\nI tried to find any function in the Tensorflow docs, but I didn't find anything.\nHow can I freeze specific neuron when I implemented the layers with tf.nn.conv2d?<\/p>\n","answers":[{"AnswerId":"52815605","CreationDate":"2018-10-15T11:20:27.867","ParentId":null,"OwnerUserId":"5055862","Title":null,"Body":"<p>As clarified in the comments, you want to freeze Neurons in a tf.nn.conv2d convolution. While there is direct way of doing this in Tensorflow (as per my search), you could try slicing the Tensor and applying <code>tf.stop_gradient()<\/code> to it. Here is a <a href=\"https:\/\/stackoverflow.com\/questions\/33727935\/how-to-use-stop-gradient-in-tensorflow\">stackoverflow answer<\/a> to give you an intuition on how to use <code>tf.stop_gradient()<\/code><\/p>\n\n<p>I haven't tested it, but according to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stop_gradient\" rel=\"nofollow noreferrer\">docs<\/a> I think it should work. <\/p>\n"},{"AnswerId":"52816856","CreationDate":"2018-10-15T12:31:06.637","ParentId":null,"OwnerUserId":"4055338","Title":null,"Body":"<p>A neuron in a dense neural network layer simply corresponds to a column in a weight matrix. You could therefore redefine your weight matrix as a concatenation of 2 parts\/variables, one trainable and one not. Then you could either:<\/p>\n\n<ol>\n<li>selectively pass only the trainable part in the <code>var_list<\/code> argument of the <code>minimize<\/code> function of your optimizer, or<\/li>\n<li>Use <code>tf.stop_gradient<\/code> on the vector\/column corresponding to the neuron you want to freeze.<\/li>\n<\/ol>\n\n<p>The same concept could be used for convolutional layers, although in this case the definition of a \"neuron\" becomes unclear; still, you could freeze any column(s) of a convolutional kernel.<\/p>\n"}]}
+{"QuestionId":52815082,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-15T10:52:09.870","AcceptedAnswerId":null,"OwnerUserId":8380638.0,"Title":"Replace Iterator in loaded graph","Body":"<p>I am recovering a graph from a checkpoint file. The network was feed using a Dataset iterator. I would like to replace this iterator with a new one, since I want to evaluate the results for some new examples.<\/p>\n\n<p>I can't find a valid way to override the old iterator with a new one. <\/p>\n\n<p>This is what I have tried:<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph(\"{}.meta\".format(checkpoint_file), input_map={})\niterator = tf.get_default_graph().get_operation_by_name(\"DataPipe\/Iterator\")\ninit_op = iterator.make_initializer(dataset)\n<\/code><\/pre>\n\n<p>But I get the following error:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'Operation' object has no attribute 'make_initializer'<\/p>\n<\/blockquote>\n\n<p>What could I do? The priority is reuse the network I have already trained. <\/p>\n\n<h2>EDIT:<\/h2>\n\n<p>This is a <a href=\"https:\/\/stackoverflow.com\/questions\/46917588\/restoring-a-tensorflow-model-that-uses-iterators\">related question<\/a>, but my network is already trained so I can't change its definition now.<\/p>\n","answers":[]}
+{"QuestionId":52815171,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-15T10:57:29.323","AcceptedAnswerId":52815622.0,"OwnerUserId":9540764.0,"Title":"Create a tensor from another with a condition","Body":"<p>I have a tensor like this<\/p>\n\n<pre><code>...\n0\n0\n2\n0\n1\n0\n3\n1\n0\n0\n0\n0\n2\n3\n...\n<\/code><\/pre>\n\n<p>and I must find a way to create a tensor of the same shape but with only 0 and 1. The one must be in the same position as a specific number. Here an example<\/p>\n\n<pre><code># for number 2\n...\n0\n0\n1\n0\n0\n0\n0\n0\n0\n0\n0\n0\n1\n0\n...\n<\/code><\/pre>\n\n<p>Is there a built-in function that can do that? I can't find anything online and in the documentation.<\/p>\n\n<p>Then I have to multiply this tensor by a list of numbers like this<\/p>\n\n<pre><code>l = [..., 1.3, 4.3, ...]\n<\/code><\/pre>\n\n<p>to obtain this<\/p>\n\n<pre><code>...\n0\n0\n1.3\n0\n0\n0\n0\n0\n0\n0\n0\n0\n4.3\n0\n...\n<\/code><\/pre>\n\n<p>Is there a way to obtain this result?<\/p>\n\n<p><strong>edit<\/strong><\/p>\n\n<p>I'm having trouble in applying this method in my case. I explain it. My tensor in which to take the indices is like this<\/p>\n\n<pre><code>points = tf.constant([[[0], [1], [0], [2], [1], [0], [2], [2], [0], [1]],\n                      [[1], [2], [0], [0], [1], [0], [1], [2], [0], [2]],\n                      [[0], [2], [1], [0], [2], [0], [1], [2], [0], [1]]], dtype=tf.float32)\n# shape=(3, 10, 1)\n<\/code><\/pre>\n\n<p>I have to take the indices only of the first row so I take them in this way <\/p>\n\n<pre><code>idx = tf.cast(tf.where(tf.equal(tf.slice(points, [0, 0, 0], [1, 10, 1]), tf.constant([2], dtype=tf.float32))), dtype=tf.int32)\n# shape=(3, 3)\nidx = tf.expand_dims(idx, 0)\n# shape=(1, 3, 3)\n<\/code><\/pre>\n\n<p>The values to feed are inside a list called <code>vectors<\/code> and I transformed them to have the same shape like this<\/p>\n\n<pre><code>to_feed = np.expand_dims(np.array(vectors), 0)\n# shape=(1, 3, 3)\n<\/code><\/pre>\n\n<p>and then I apply the method in this way<\/p>\n\n<pre><code>res = tf.scatter_nd(idx, to_feed, tf.shape(tf.slice(points, [0, 0, 0], [1, 10, 3])))\n<\/code><\/pre>\n\n<p>But I get this error<\/p>\n\n<pre><code>ValueError: The inner 0 dimensions of output.shape=[?,?,?] must match the inner 1 dimensions of updates.shape=[1,3,3]: Shapes must be equal rank, but are 0 and 1 for 'ScatterNd' (op: 'ScatterNd') with input shapes: [1,3,3], [1,3,3], [3].\n<\/code><\/pre>\n\n<p>What I need at the end is a tensor like this<\/p>\n\n<pre><code>to_feed = [[[10, 10, 10], [11, 11, 11], [12, 12, 12]]] # shape=(1, 3, 3)\nres = [[[0, 0, 0], [10, 10, 10], [0, 0, 0], [11, 11, 11], [0, 0, 0], [0, 0, 0], [0, 0, 0], [0, 0, 0], [12, 12, 12], [0, 0, 0]]] # shape=(1, 10, 3)\n<\/code><\/pre>\n\n<p>(The <code>[0, 0, 0]<\/code> are put randomly just to have an idea)<\/p>\n","answers":[{"AnswerId":"52815622","CreationDate":"2018-10-15T11:21:36.363","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can just compare with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/equal\" rel=\"nofollow noreferrer\"><code>tf.equal<\/code><\/a> and then convert the boolean result to a number with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cast\" rel=\"nofollow noreferrer\"><code>tf.cast<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nvector = tf.placeholder(tf.int32, [None])\nnum = tf.placeholder(tf.int32, [])\nresult = tf.cast(tf.equal(vector, num), tf.int32)\nwith tf.Session() as sess:\n    print(sess.run(result, feed_dict={vector: [0, 0, 2, 1, 0, 3, 2, 0], num: 2}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[0 0 1 0 0 0 1 0]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>The above solves the simpler first question, but I think what you need to solve your problem is something like the following, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/where\" rel=\"nofollow noreferrer\"><code>tf.where<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_nd\" rel=\"nofollow noreferrer\"><code>tf.scatter_nd<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nvector = tf.placeholder(tf.int32, [None])\nnum = tf.placeholder(tf.int32, [])\nvalues = tf.placeholder(tf.float32, [None])\nidx = tf.where(tf.equal(vector, num))\nresult = tf.scatter_nd(idx, values, tf.cast(tf.shape(vector), idx.dtype))\nwith tf.Session() as sess:\n    print(sess.run(result, feed_dict={vector: [0, 2, 1, 2, 3, 2, 2, 0],\n                                      num: 2,\n                                      values: [3, 4.4, 2, 2.2]}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[0.  3.  0.  4.4 0.  2.  2.2 0. ]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>About your latest example, I have put together a snippet for what I think you are trying to achieve:<\/p>\n\n<pre><code>import tensorflow as tf\n\npoints = tf.constant([[[0], [1], [0], [2], [1], [0], [2], [2], [0], [1]],\n                      [[1], [2], [0], [0], [1], [0], [1], [2], [0], [2]],\n                      [[0], [2], [1], [0], [2], [0], [1], [2], [0], [1]]], dtype=tf.float32)\nvectors = tf.constant([[10, 11, 12], [20, 21, 22], [30, 31, 21]], dtype=tf.float32)\n\npoints_row = points[:1]\nidx = tf.where(tf.equal(points_row, 2))\nidx = tf.cast(idx, tf.int32)\nres_shape = tf.concat([tf.shape(points_row), [tf.shape(vectors)[1]]], axis=0)\nres = tf.scatter_nd(idx, vectors, res_shape)\nres = tf.squeeze(res, 2)\n\nwith tf.Session() as sess:\n    print(sess.run(res))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[ 0.  0.  0.]\n  [ 0.  0.  0.]\n  [ 0.  0.  0.]\n  [10. 11. 12.]\n  [ 0.  0.  0.]\n  [ 0.  0.  0.]\n  [20. 21. 22.]\n  [30. 31. 21.]\n  [ 0.  0.  0.]\n  [ 0.  0.  0.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52815264,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-10-15T11:02:35.230","AcceptedAnswerId":52815644.0,"OwnerUserId":843036.0,"Title":"TypeError: cannot unpack non-iterable int object","Body":"<p>Im trying to make my first CNN using pyTorch and am following online help and code already people wrote. i am trying to reproduce their results. I'm using the Kaggle Dogs Breed Dataset for this and below is the error I get. The trainloader does not return my images and labels and any attempt to get them leads in an error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"E:\\Program Files\\JetBrains\\PyCharm Community Edition 2018.2.4\\helpers\\pydev\\pydevd.py\", line 1664, in &lt;module&gt;\n    main()\n  File \"E:\\Program Files\\JetBrains\\PyCharm Community Edition 2018.2.4\\helpers\\pydev\\pydevd.py\", line 1658, in main\n    globals = debugger.run(setup['file'], None, None, is_module)\n  File \"E:\\Program Files\\JetBrains\\PyCharm Community Edition 2018.2.4\\helpers\\pydev\\pydevd.py\", line 1068, in run\n    pydev_imports.execfile(file, globals, locals)  # execute the script\n  File \"E:\\Program Files\\JetBrains\\PyCharm Community Edition 2018.2.4\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/sbzfk\/PycharmProjects\/my_FCN_attempt\/Kaggle_Dogs_Competition.py\", line 85, in &lt;module&gt;\n    img, label = next(iter(train_loader))\n  File \"C:\\Users\\sbzfk\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 314, in __next__\n    batch = self.collate_fn([self.dataset[i] for i in indices])\n  File \"C:\\Users\\sbzfk\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 314, in &lt;listcomp&gt;\n    batch = self.collate_fn([self.dataset[i] for i in indices])\n  File \"C:\/Users\/sbzfk\/PycharmProjects\/my_FCN_attempt\/Kaggle_Dogs_Competition.py\", line 42, in __getitem__\n    img = self.transform(img)\n  File \"C:\\Users\\sbzfk\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\torchvision\\transforms.py\", line 34, in __call__\n    img = t(img)\n  File \"C:\\Users\\sbzfk\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\torchvision\\transforms.py\", line 187, in __call__\n    w, h = img.size\nTypeError: cannot unpack non-iterable int object\n<\/code><\/pre>\n\n<p>Below is my code:<\/p>\n\n<pre><code>class DogsDataset(Dataset):\n\n    def __init__(self, filenames, labels, root_dir, transform=None):\n        assert len(filenames) == len(labels)        # if the two are not of equal length throw an error\n        self.filenames = filenames\n        self.labels = labels\n        self.root_dir = root_dir\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.filenames)\n\n    def __getitem__(self, idx):\n        this_img = join(self.root_dir, 'train', self.filenames[idx]+'.jpg')\n        print(this_img)\n        img = io.imread(this_img)\n        label = self.labels[idx]\n        print(label)\n\n        if self.transform:\n            img = self.transform(img)\n\n        return [img, label]\n\n\n\nbatch_size = 64\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\ndataset_root = expanduser(join('~', 'Documents', 'kaggle_dogs_dataset'))\n# join will intelligently join directories irrespective of OS, and expanduser will\n# replace with \/home\/ in linux or the username in Windows\n\ncsv_file = pd.read_csv(join(dataset_root, 'labels.csv'))        # csv file has two columns, id which are filenames and breed which are labels\nfilenames = csv_file.id.values              # convert that column to an array, id is the column name and values converty to numpy array\n# le = LabelEncoder()\n# labels = le.fit_transform(csv_file.breed)           # this will just encode the names between 0 to models-1 , basically changing strings to integers\nlabels = csv_file.breed.values\n\nfilenames_train, filenames_eval, labels_train, labels_eval = train_test_split(filenames, labels,\n                                                                              test_size=0.1, stratify=labels)       # this is an import from sklearn as the name implies, it randomly splits data into train and eval, 10% of it to test and rest train\n\ndata_transform = transforms.Compose([transforms.Scale(224),\n                                     transforms.CenterCrop(224),\n                                     transforms.ToTensor(),\n                                     transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])\n\ndog_train = DogsDataset(filenames_train, labels_train, dataset_root, transform=data_transform)\ntrain_loader = DataLoader(dog_train, batch_size, shuffle=True)\n\ndog_eval = DogsDataset(filenames_eval, labels_eval, dataset_root, transform=data_transform)\neval_loader = DataLoader(dog_eval, batch_size, shuffle=True)\n\n\ndef im_show(axis, inp):\n    \"\"\"Denormalize and show\"\"\"\n    inp = inp.numpy().transpose((1, 2, 0))\n    mean = np.array([0.485, 0.456, 0.406])\n    std = np.array([0.229, 0.224, 0.225])\n    inp = std * inp + mean\n    axis.imshow(inp)\n\n\nimg, label = next(iter(train_loader))\nprint(img.size(), label.size())\nfig = plt.figure(1, figsize=(16, 4))\ngrid = ImageGrid(fig, 111, nrows_ncols=(1, 4), axes_pad=0.05)\nfor i in range(img.size()[0]):\n    ax = grid[i]\n    im_show(ax, img[i])\n<\/code><\/pre>\n\n<p>Ive tried debugging it line by line and with <code>transform=none<\/code> I seem to read all the images, only with <code>transform=data_transform<\/code> I seem to get this error.<\/p>\n","answers":[{"AnswerId":"52815644","CreationDate":"2018-10-15T11:22:57.080","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you are using torchvision's image transforms. Some of these transforms are expecting as input a <a href=\"https:\/\/pillow.readthedocs.io\/en\/3.1.x\/reference\/Image.html#image-module\" rel=\"nofollow noreferrer\"><code>PIL.Image<\/code><\/a> object, rather than a tensor or numpy array.<br>\nYou are using <code>io.imread<\/code> to read ths image file, and I suspect this <code>io<\/code> is not <code>PIL.Image<\/code> resulting with a numpy array.<br>\nMake sure you pass <code>PIL.Image<\/code> objects to <code>transforms<\/code> and that your <code>DogsDataset<\/code> returns a 3D tensor for image (C-H-W shaped).<\/p>\n"}]}
+{"QuestionId":52815267,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-15T11:02:49.570","AcceptedAnswerId":null,"OwnerUserId":10506791.0,"Title":"Training multiple Keras models in one script","Body":"<p>I want to train different Keras models (or in some cases just multiple runs of the same model to compare the results) in a queue (using TensorFlow as the backend if that matters). In my current setup I create and fit all of these models in one big python script, e.g. (in a simplified way):<\/p>\n\n<pre><code>for i in range(10):\n    model = create_model(i)\n    model.compile(...)\n    model.fit(...)\n    some_function_to_save_model(model)\n<\/code><\/pre>\n\n<p>The <code>create_model(i)<\/code> function creates the specific model for the <code>i<\/code>'th run. This includes changing the number of inputs \/ labels for example. The compile function can be different (e.g. different optimizer) for each run as well.<\/p>\n\n<p>While this code works for me and I have not found any problems, I am unclear if this is the correct way to do it because all of the models reside in the same TensorFlow Graph (if I understand the way Keras \/ TensorFlow work together correctly). My questions are:<\/p>\n\n<ol>\n<li>is this the correct way to run multiple <strong>independent<\/strong> models. (I do not want any influence of the <code>i<\/code>'th run on the <code>i+1<\/code>'th run)<\/li>\n<li>is running the models from different python scripts (in this example model1.py, model2.py, ... model9.py) in any way better <strong>technically speaking<\/strong> (I am not referring to readability \/ reproducibility here) because each model would then have its own separate TensorFlow Graph \/ Session?<\/li>\n<li>Does clearing the Session \/ deleting the Graph via <code>keras.backend.clear_session()<\/code> have any influence in this case if it is run after the save function (<code>some_function_to_save_model()<\/code> inside the <code>for<\/code> loop)? Is this in some way beneficial compared to the current setup?<\/li>\n<\/ol>\n\n<p>Once again: I am not concerned with the problems that might arise due to creating messy code if all models are cramped together in one script instead of a single script per model only with creating &amp; training models independently.<\/p>\n\n<p>Unfortunately I did not find a concise answer to this (only suggestions using both methods). Maybe someone here can enlighten me?<\/p>\n\n<p>Edit: Maybe I should be more precise. Basically I would like to have a technical explanation regarding the differences (advantages &amp; disadvantages) of the following three cases:<\/p>\n\n<ol>\n<li><p>create_and_train.py:<\/p>\n\n<pre><code>for i in range(10):\n    model = create_model(i)\n    model.compile(...)\n    model.fit(...)\n    some_function_to_save_model(model)\n<\/code><\/pre><\/li>\n<li><p>create_and_train.py:<\/p>\n\n<pre><code>for i in range(10):\n    model = create_model(i)\n    model.compile(...)\n    model.fit(...)\n    some_function_to_save_model(model)\n    # clear session:\n    keras.backend.clear_session()\n<\/code><\/pre><\/li>\n<li><p>create_and_train_i.py with i in [0, 1, ..., 9]:<\/p>\n\n<pre><code>i = 5 # (e.g.)\nmodel = create_model(i)\nmodel.compile(...)\nmodel.fit(...)\nsome_function_to_save_model(model)\n<\/code><\/pre>\n\n<p>and e.g. a bash script that loops through these<\/p><\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":52815708,"AnswerCount":1,"Tags":"<tensorflow><cuda><pytorch>","CreationDate":"2018-10-15T11:26:02.570","AcceptedAnswerId":52816306.0,"OwnerUserId":1565344.0,"Title":"Order of CUDA devices","Body":"<p>I saw <a href=\"https:\/\/stackoverflow.com\/questions\/26123252\/inconsistency-of-ids-between-nvidia-smi-l-and-cudevicegetname\">this<\/a> solution, but it doesn't quite answer my question; it's also quite old so I'm not sure how relevant it is. <\/p>\n\n<p>I keep getting conflicting outputs for the order of GPU units. There are two of them: Tesla K40 and NVS315 (legacy device that is never used). When I run <code>deviceQuery<\/code>, I get <\/p>\n\n<pre><code>Device 0: \"Tesla K40m\"\n...\nDevice PCI Domain ID \/ Bus ID \/ location ID:   0 \/ 4 \/ 0\n\nDevice 1: \"NVS 315\"\n...\nDevice PCI Domain ID \/ Bus ID \/ location ID:   0 \/ 3 \/ 0\n<\/code><\/pre>\n\n<p>On the other hand, <code>nvidia-smi<\/code> produces a different order: <\/p>\n\n<pre><code> 0  NVS 315 \n\n 1  Tesla K40m\n<\/code><\/pre>\n\n<p>Which I find very confusing. The solution I found for Tensorflow (and a similar one for Pytorch) is to use<\/p>\n\n<pre><code>import os\nos.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"  \nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n<\/code><\/pre>\n\n<p>PCI Bus ID is 4 for Tesla and 3 for NVS, so it should set it to 3 (NVS), is that right? <\/p>\n\n<p>In pytorch I set <\/p>\n\n<pre><code>os.environ['CUDA_VISIBLE_DEVICES']='0'\n...\ndevice = torch.cuda.device(0)\nprint torch.cuda.get_device_name(0)\n<\/code><\/pre>\n\n<p>to get <code>Tesla K40m<\/code><\/p>\n\n<p>when I set instead <\/p>\n\n<pre><code>os.environ['CUDA_VISIBLE_DEVICES']='1'\ndevice = torch.cuda.device(1)\nprint torch.cuda.get_device_name(0)\n<\/code><\/pre>\n\n<p>to get <\/p>\n\n<pre><code>UserWarning: \n    Found GPU0 NVS 315 which is of cuda capability 2.1.\n    PyTorch no longer supports this GPU because it is too old.\n\n  warnings.warn(old_gpu_warn % (d, name, major, capability[1]))\nNVS 315\n<\/code><\/pre>\n\n<p>So I'm quite confused: what's the <em>true<\/em> order of GPU devices that tf and pytorch use? <\/p>\n","answers":[{"AnswerId":"52816306","CreationDate":"2018-10-15T11:59:48.793","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>By default, CUDA orders the GPUs by computing power. GPU:0 will be the fastest GPU on your host, in your case the K40m. <\/p>\n\n<p>If you set CUDA_DEVICE_ORDER='PCI_BUS_ID' then CUDA orders your GPU depending on how you set up your machine meaning that GPU:0 will be the GPU on your first PCI-E lane.<\/p>\n\n<p>Both Tensorflow and PyTorch use the CUDA GPU order. That is consistent with what you showed:<\/p>\n\n<pre><code>os.environ['CUDA_VISIBLE_DEVICES']='0'\n...\ndevice = torch.cuda.device(0)\nprint torch.cuda.get_device_name(0)\n<\/code><\/pre>\n\n<p>Default order so GPU:0 is the K40m since it is the most powerful card on your host.<\/p>\n\n<pre><code>os.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"  \nos.environ['CUDA_VISIBLE_DEVICES']='0'\n...\ndevice = torch.cuda.device(0)\nprint torch.cuda.get_device_name(0)\n<\/code><\/pre>\n\n<p>PCI-E lane order, so GPU:0 is the card with the lowest bus-id in your case the NVS.<\/p>\n"}]}
+{"QuestionId":52816688,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-15T12:21:25.703","AcceptedAnswerId":52825367.0,"OwnerUserId":4481339.0,"Title":"Extract multiple submatrices from a Tensor","Body":"<p>I'm sorry that I've to ask this question, as it seems pretty straightforward, but I'm trying to find a way specifically in Tensorflow.<\/p>\n\n<p>I've a Tensor matrix like below:<\/p>\n\n<pre><code>    [0 0 1 1]\nX = [0 0 1 1]\n    [1 1 0 0]\n    [1 1 0 0]\n<\/code><\/pre>\n\n<p>I need to extract both patches:<\/p>\n\n<pre><code>  [1,1]     [1,1]\n  [1,1] &amp;   [1,1]\n<\/code><\/pre>\n\n<p>And I'm also give list of indices that point to the left element of submatrices given right above. E.g. <\/p>\n\n<pre><code>[[0,2]\n [2,0]]\n<\/code><\/pre>\n\n<p>I need to extract the patches in Tensorflow. Thanks.<\/p>\n","answers":[{"AnswerId":"52818044","CreationDate":"2018-10-15T13:37:09.567","ParentId":null,"OwnerUserId":"9178720","Title":null,"Body":"<p>Well, if you know what submatrices you need to extract, tf.slice() is the best choice.<\/p>\n\n<p>The doc is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n\n<p>For the example you presented, the solution with tf.slice() would be:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = [[0, 0, 1, 1],\n     [0, 0, 1, 1],\n     [1, 1, 0, 0],\n     [1, 1, 0, 0]]\nX = tf.Variable(x)\ns1 = tf.slice(X, [2,0], [2,2])\ns1 = tf.slice(X, [0,2], [2,2])\n\n\nwith tf.Session() as sess:\n  init = tf.global_variables_initializer()\n  sess.run(init)\n  print(sess.run([s1, s1]))\n<\/code><\/pre>\n\n<p>This code renders the following result:<\/p>\n\n<pre><code>[array([[1, 1], [1, 1]], dtype=int32), \narray([[1, 1], [1, 1]], dtype=int32)]\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>For a more automatic and less verbose way, you can just use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#__getitem__\" rel=\"nofollow noreferrer\"><strong>getitem<\/strong><\/a> proprety from tensorflow and slice it as you would slice a npArray.<\/p>\n\n<p>The code could be something like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nvar = [[0, 0, 1, 1],\n       [0, 0, 1, 1],\n       [1, 1, 0, 0],\n       [1, 1, 0, 0]]\nX = tf.Variable(var)\nslices = [[0,2], [2,0]]\n\ns = []\nfor sli in slices:\n  y = sli[0]\n  x = sli[1]\n  s.append(X[y:y+2, x:x+2])\n\n\nwith tf.Session() as sess:\n  init = tf.global_variables_initializer()\n  sess.run(init)\n  print(sess.run(s))\n<\/code><\/pre>\n\n<p>This code renders the following result:<\/p>\n\n<pre><code>[array([[1, 1], [1, 1]], dtype=int32), \narray([[1, 1], [1, 1]], dtype=int32)]\n<\/code><\/pre>\n"},{"AnswerId":"52825367","CreationDate":"2018-10-15T22:05:50.143","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>You can do this using tf.gather_nd as well. Below is an example showing all the working bits, and what you can do with gather_nd. You should be able to construct indices so that you only need a single gather_nd op to get all the submatrices you want. I just included the variable indices to show that you can use it to get submatrices from tensors you don't know ahead of time. So for example if you calculate some stuff in the graph and want to get submatrices based on that you could.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# build a tensor\nx = np.arange(25)\nx = np.reshape(x, [5, 5])\ny = x + 4\nthree_d_array = np.stack([x, y], axis=2)\n# just so you can see the shape its made of\nprint(np.all(x == three_d_array[:,:,0]))\nprint(np.all(y == three_d_array[:,:,1]))\n# make it into a tf tensor\nthree_d_tensor = tf.constant(three_d_array)\n\n# create a variable for tensor valued slice indices\nrow_0, col_0 = 0, 0\nrow_1, col_1 = 0, 1\nrow_2, col_2 = 1, 0\nrow_3, col_3 = 1, 1\nslice_tensor = tf.constant([\n    [row_0, col_0],\n    [row_1, col_1],\n    [row_2, col_2],\n    [row_3, col_3]\n])\nslices = tf.Variable(initial_value=slice_tensor)\n\n# op to get the sub matrices\ngather_op = tf.gather_nd(three_d_tensor, slices)\n\nwith tf.Session() as sess:\n  init = tf.global_variables_initializer()\n  sess.run(init)\n\n  submatrices = sess.run(gather_op)\n  print(submatrices[0,:] == three_d_array[row_0, col_0])\n  print(submatrices[1,:] == three_d_array[row_1, col_1])\n  print(submatrices[2,:] == three_d_array[row_2, col_2])\n  print(submatrices[3,:] == three_d_array[row_3, col_3])\n\n  # shift down 2 along 2\n  offset_top_left = tf.constant([2,2])\n  update_variable_op = tf.assign(slices, slices + offset_top_left[None,:])\n  sess.run(update_variable_op)\n  submatrices = sess.run(gather_op)\n  print(submatrices[0, :] == three_d_array[row_0 + 2, col_0 + 2])\n  print(submatrices[1, :] == three_d_array[row_1 + 2, col_1 + 2])\n  print(submatrices[2, :] == three_d_array[row_2 + 2, col_2 + 2])\n  print(submatrices[3, :] == three_d_array[row_3 + 2, col_3 + 2])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52816789,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><nlp>","CreationDate":"2018-10-15T12:27:36.900","AcceptedAnswerId":null,"OwnerUserId":10043103.0,"Title":"Keras error about array shape but shape seems correct","Body":"<p>I'm trying to train a simple model with Keras and python. The text is preprocessed perfectly. But when I try to fit it I get the following error:<\/p>\n\n<pre><code>File \"main.py\", line 47, in &lt;module&gt;\n    model.fit(x_train, y_train, batch_size=32, epochs=3)\n  File \"\/home\/shamildacoder\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 952, in fit\n    batch_size=batch_size)\n  File \"\/home\/shamildacoder\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 789, in _standardize_user_data\n    exception_prefix='target')\n  File \"\/home\/shamildacoder\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\", line 138, in standardize_input_data\n    str(data_shape))\nValueError: Error when checking target: expected dense_2 to have shape (121885,) but got array with shape (1000,)\n<\/code><\/pre>\n\n<p>But <code>print(x_train.shape)<\/code> and <code>print(y_train.shape)<\/code> both return <code>(121885, 1000)<\/code>. I don't see any reason.<\/p>\n\n<p>Code: <a href=\"https:\/\/pastebin.com\/afnzBf6B\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/afnzBf6B<\/a><\/p>\n\n<pre><code>from keras.preprocessing.text import Tokenizer\nfrom keras.layers import Dense\nfrom keras.models import Sequential\n\ndata = open('movie_lines.txt', encoding='ISO-8859-1')\nlines = [line for line in data]\nfiltered_lines = []\n\nfor line in lines:\n    sentence = line.split('+++$+++')[4].strip(' ')\n    filtered_lines.append(sentence)\n\ntrain_size = int(len(filtered_lines) * .8)\ntrain_portion = filtered_lines[:train_size]\ntest_portion = filtered_lines[train_size:]\n\nx_lines = train_portion[::2]\ny_lines = train_portion[1::2]\nx_test = test_portion[::2]\ny_test = test_portion[1::2]\nvocab_size = 1000\nprint('Prepared data')\n\ndef prepare_text(text):\n    tokenizer = Tokenizer(num_words=vocab_size)\n    tokenizer.fit_on_texts(text)\n    matrix = tokenizer.texts_to_matrix(text)\n    return matrix\n\n\nx_train = prepare_text(x_lines)\nprint('matrixed x')\ny_train = prepare_text(y_lines)\nprint('matrixed y')\nprint(f'X shape: {x_train.shape}')\nprint(f'Y shape: {y_train.shape}')\n\nmodel = Sequential()\nmodel.add(Dense(512, input_shape=(vocab_size,), activation='relu'))\nmodel.add(Dense(len(y_lines), activation='softmax'))\n\nmodel.compile(\n        loss='categorical_crossentropy',\n        optimizer='adam',\n        metrics=['accuracy',]\n        )\nprint('Created and compiled model')\n\nmodel.fit(x_train, y_train, epochs=3)\n\nscore = model.evaluate(x_test, y_test, batch_size=32, epochs=3)\nprint('Test Score:'+score[0])\nprint('Test Accuracy:'+score[1])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52817401","CreationDate":"2018-10-15T13:00:13.307","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>In the preprocessing stage you are using <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/2c7ef1d27bff4a24fe8bd2dc3800ccdebfc08a75\/keras_preprocessing\/text.py#L369\" rel=\"nofollow noreferrer\"><code>texts_to_matrix()<\/code><\/a> method which (with default arguments) returns the given sequences as rows of a matrix in one-hot encoded format. Now if you would like to use just Dense layers to go from a one-hot encoded sequence to another one-hot encoded sequence, you need to set the number of units in the last layer to the vocabulary size (i.e. number of columns in the matrix) and use <code>sigmoid<\/code> as the activation function of last layer:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(512, input_shape=(vocab_size,), activation='relu'))\nmodel.add(Dense(vocab_size, activation='sigmoid'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52816938,"AnswerCount":1,"Tags":"<numpy><keras>","CreationDate":"2018-10-15T12:35:51.310","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"how to convert numpy array to keras tensor","Body":"<p>When using the keras model to do predict, I got the error below<br>\nAttributeError: 'Tensor' object has no attribute 'ndim'<br>\nThe reason is that the weights is numpy array, not tensor.<br>\nSo how to convert numpy array to keras tensor?<\/p>\n","answers":[{"AnswerId":"56611154","CreationDate":"2019-06-15T14:16:00.077","ParentId":null,"OwnerUserId":"7906266","Title":null,"Body":"<p>In Tensorflow it can be done the following way:<\/p>\n\n<pre><code>import tensorflow.keras.backend as K\nimport numpy as np\n\na = np.array([1,2,3])\nb = K.constant(a)\nprint(b)\n\n# &lt;tf.Tensor 'Const_1:0' shape=(3,) dtype=float32&gt;\n\nprint(K.eval(b))\n\n# array([1., 2., 3.], dtype=float32)\n<\/code><\/pre>\n\n<p>In raw keras it should be done replacing <code>import tensorflow.keras.backend as K<\/code> with <code>from keras import backend as K<\/code>.<\/p>\n"}]}
+{"QuestionId":52817105,"AnswerCount":1,"Tags":"<pytorch><loss-function>","CreationDate":"2018-10-15T12:44:52.010","AcceptedAnswerId":52817106.0,"OwnerUserId":9270181.0,"Title":"AttributeError: 'Tensor' has no attribute: 'backwards'","Body":"<p>I'm posting this question for those who made the same error I did. I got this error when trying to compute my gradients:<\/p>\n\n<pre><code>criterion = torch.nn.CrossEntropyLoss()\nloss = criterion(y_hat, y_truth)\nloss.backwards()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52817106","CreationDate":"2018-10-15T12:44:52.010","ParentId":null,"OwnerUserId":"9270181","Title":null,"Body":"<p>It's <code>loss.backward()<\/code>, not <code>loss.backwards()<\/code>.<\/p>\n"}]}
+{"QuestionId":52817290,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-15T12:54:36.107","AcceptedAnswerId":null,"OwnerUserId":10507354.0,"Title":"tensorflow.python.framework.errors_impl.InvalidArgumentError: Index out of range using input dim 2; input has only 2 dims","Body":"<p>I write code in Keras using BiLSTM+CRF to do Named Entity Recognition.\nThis is my model\uff1a  <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(1000, output_dim=128, \ninput_length=100))\nmodel.add(Bidirectional(LSTM(200, return_sequences=True)))\nmodel.add(Dropout(0.3))\nmodel.add(TimeDistributed(Dense(2)))\ncrf_layer = CRF(2)\nmodel.add(crf_layer)\nmodel.compile('rmsprop', loss=crf_layer.loss_function, metrics=[crf_layer.accuracy])\nreturn model\n<\/code><\/pre>\n\n<p>And Testdata&amp;label<\/p>\n\n<pre><code>docs = ['Well done!',\n        'Good work',\n        'Great effort',\n        'nice work',\n        'Excellent!',\n        'Weak',\n        'Poor effort!',\n        'not good',\n        'poor work',\n        'Could have done better.']\nlabels = [1,1,1,1,1,0,0,0,0,0]\n#embedding\nvocab_size = 2500\nencoded_docs = [one_hot(d, vocab_size) for d in docs]\nprint(encoded_docs)\n\nmax_length = 100\npadded_docs = pad_sequences(encoded_docs, maxlen=max_length, padding='post')\nprint(padded_docs)\n\nmodel = build_embedding_bilstm2_crf_model()\n\nprint(model.summary())\n\nmodel.fit(padded_docs, labels, epochs=50, verbose=0)\n<\/code><\/pre>\n\n<p>I get the following error.<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Index out of range using input dim 2; input has only 2 dims\n     [[Node: loss\/crf_1_loss\/strided_slice_4 = StridedSlice[Index=DT_INT32, T=DT_FLOAT, begin_mask=7, ellipsis_mask=0, end_mask=5, new_axis_mask=0, shrink_axis_mask=0, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](_arg_crf_1_target_0_1\/_81, metrics\/acc\/strided_slice_9\/stack, metrics\/acc\/strided_slice_2\/stack_1, metrics\/acc\/strided_slice_9\/stack_2)]]\n     [[Node: metrics\/acc\/while_1\/Switch_2\/_181 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_1763_metrics\/acc\/while_1\/Switch_2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](^_cloopmetrics\/acc\/while_1\/ExpandDims\/dim\/_65)]]\n<\/code><\/pre>\n\n<p>Use One_hot to convert words into vectors. Use * to generate a new sequence.\nI guess there is a problem with the structure of the neural network.Why is that?<\/p>\n","answers":[]}
+{"QuestionId":52817771,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-15T13:23:20.333","AcceptedAnswerId":52818310.0,"OwnerUserId":5118879.0,"Title":"Tensorflow ffill implementation","Body":"<p>I'm trying to implement pandas like ffill method for one dimensional tensor in tensorflow framework.<\/p>\n\n<p>I need to propagate last not NaN value in tensor:<\/p>\n\n<pre><code>[1,nan,nan,0,nan,nan,2,nan,3]   =&gt;   [1,1,1,0,0,0,2,2,3]\n<\/code><\/pre>\n\n<p>I tried this implementation with tf.scan, but no luck:<\/p>\n\n<pre><code>def ffill(previous_output, current_input):\n  condition = tf.equal(current_input,np.nan)\n  return tf.cond(condition,lambda: previous_output,lambda : current_input)\n\ninitializer = tf.constant(0.0)\nfilled_tensor = tf.scan(ffill, nan_tensor, initializer)\n<\/code><\/pre>\n\n<p>please help<\/p>\n","answers":[{"AnswerId":"52818310","CreationDate":"2018-10-15T13:51:51.787","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a way to do that:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input data\ninput = tf.placeholder(tf.float32, [None])\n# Find non-NaN values\nmask = ~tf.is_nan(input)\n# Take non-NaN values and precede them with a NaN\nvalues = tf.concat([[math.nan], tf.boolean_mask(input, mask)], axis=0)\n# Use cumsum over mask to find the index of the non-NaN value to pick\nidx = tf.cumsum(tf.cast(mask, tf.int64))\n# Gather values\nresult = tf.gather(values, idx)\n# Test\nwith tf.Session() as sess:\n    input_test = [1, math.nan, math.nan, 0, math.nan, math.nan, 2, math.nan, 3]\n    print(sess.run(result, feed_dict={input: input_test}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[1. 1. 1. 0. 0. 0. 2. 2. 3.]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52817926,"AnswerCount":1,"Tags":"<python><docker><tensorflow>","CreationDate":"2018-10-15T13:30:27.977","AcceptedAnswerId":null,"OwnerUserId":9723527.0,"Title":"docker tensorflow linking with folder","Body":"<p>I am trying docker for tensorflow on windows 10 education, I have installed docker successfully and could run\/ pull\/ import images. I linked my docker container using <\/p>\n\n<pre><code>C:\\User\\xyz_folder&gt; docker run -it tensorflow\/tensorflow:latest-devel\n\nroot@23433215319e:~#cd \/tensorflow\n\nroot@23433215319e:\/tensorflow#git pull\n\nFrom https:\/\/github.com\/tensorflow\/tensorflow\n * [new tag]         v1.11.0    -&gt; v1.11.0\nAlready up-to-date.\n<\/code><\/pre>\n\n<p>Until here it ran fine without error. Following is the problem:<\/p>\n\n<pre><code>root@23433215319e:\/tensorflow# cd abc_folder\nbash: cd: abc_folder: No such file or directory\n<\/code><\/pre>\n\n<p>the abc_folder is there in linked folder but can not be seen in when it list it using 'ls'<\/p>\n\n<pre><code>root@23433215319e:\/tensorflow#ls\nACKNOWLEDGMENTS  CODEOWNERS      LICENSE     WORKSPACE           bazel-out         configure.py  tools ADOPTERS.md      CODE_OF_CONDUCT.md  README.md    arm_compiler.BUILD  bazel-tensorflow  models.BUILD AUTHORS CONTRIBUTING.md     RELEASE.md   bazel-bin     bazel-testlogs    tensorflow BUILD  ISSUE_TEMPLATE.md   SECURITY.md  bazel-genfiles   configure    third_party\n<\/code><\/pre>\n\n<p>Please suggest how to link this properly so that I can see the shared folders content.\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"52818197","CreationDate":"2018-10-15T13:45:53.353","ParentId":null,"OwnerUserId":"9178720","Title":null,"Body":"<p>To make a directory outside the container visible inside the container, you have to use the option -v or --volume as is stated <a href=\"https:\/\/docs.docker.com\/edge\/engine\/reference\/commandline\/run\/#mount-volume--v---read-only\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>So, your command would have to be: <\/p>\n\n<pre><code>docker run -v c:\\local\\directory:container\/directory -it tensorflow\/tensorflow:latest-devel\n<\/code><\/pre>\n\n<p>Whit that, you should be able to see the directory inside the container<\/p>\n"}]}
+{"QuestionId":52818145,"AnswerCount":1,"Tags":"<python><list><numpy><iterator><pytorch>","CreationDate":"2018-10-15T13:42:15.107","AcceptedAnswerId":52858726.0,"OwnerUserId":5918696.0,"Title":"Why pytorch DataLoader behaves differently on numpy array and list?","Body":"<p>The <strong>only<\/strong> difference is one of the parameter passed to DataLoader is in type \"numpy.array\" and the other is in type \"list\", but the DataLoader gives totally different results.<\/p>\n\n<p>You can use the following code to reproduce it:<\/p>\n\n<pre><code>from torch.utils.data import DataLoader,Dataset\nimport numpy as np\n\nclass my_dataset(Dataset):\n    def __init__(self,data,label):\n        self.data=data\n        self.label=label          \n    def __getitem__(self, index):\n        return self.data[index],self.label[index]\n    def __len__(self):\n        return len(self.data)\n\ntrain_data=[[1,2,3],[5,6,7],[11,12,13],[15,16,17]]\ntrain_label=[-1,-2,-11,-12]\n\n########################### Look at here:    \n\ntest=DataLoader(dataset=my_dataset(np.array(train_data),train_label),batch_size=2)\nfor i in test:\n    print (\"numpy data:\")\n    print (i)\n    break\n\n\ntest=DataLoader(dataset=my_dataset(train_data,train_label),batch_size=2)\nfor i in test:\n    print (\"list data:\")\n    print (i)\n    break\n<\/code><\/pre>\n\n<p>The result is:<\/p>\n\n<pre><code>numpy data:\n[tensor([[1, 2, 3],\n        [5, 6, 7]]), tensor([-1, -2])]\nlist data:\n[[tensor([1, 5]), tensor([2, 6]), tensor([3, 7])], tensor([-1, -2])]  \n<\/code><\/pre>\n","answers":[{"AnswerId":"52858726","CreationDate":"2018-10-17T15:39:49.943","ParentId":null,"OwnerUserId":"6212301","Title":null,"Body":"<p>This is because how batching is handled in <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"noreferrer\"><code>torch.utils.data.DataLoader<\/code><\/a>. <code>collate_fn<\/code> argument decides how samples from samples are merged into a single batch. Default for this argument is undocumented <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/7edfe11ba49c8539ee904f06320050916857c8fe\/torch\/utils\/data\/dataloader.py#L196\" rel=\"noreferrer\"><code>torch.utils.data.default_collate<\/code><\/a>.<\/p>\n\n<p>This function handles batching by assuming numbers\/tensors\/ndarrays are primitive data to batch and lists\/tuples\/dicts containing these primitives as structure to be (recursively) preserved. This allow you to have a semantic batching like this:<\/p>\n\n<ol>\n<li><code>(input_tensor, label_tensor) -&gt; (batched_input_tensor, batched_label_tensor)<\/code><\/li>\n<li><code>([input_tensor_1, input_tensor_2], label_tensor) -&gt; ([batched_input_tensor_1, batched_input_tensor_2], batched_label_tensor)<\/code><\/li>\n<li><code>{'input': input_tensor, 'target': target_tensor} -&gt; {'input': batched_input_tensor, 'target': batched_target_tensor}<\/code><\/li>\n<\/ol>\n\n<p>(Left side of <code>-&gt;<\/code> is output of dataset[i], while right side is batched sample from <code>torch.utils.data.DataLoader<\/code>)<\/p>\n\n<p>Your example code is similar to example 2 above: list structure is preserved while <code>int<\/code>s are batched.<\/p>\n"}]}
+{"QuestionId":52819066,"AnswerCount":1,"Tags":"<python><tensorflow><keras><fpga><quantization>","CreationDate":"2018-10-15T14:30:59.223","AcceptedAnswerId":null,"OwnerUserId":1909451.0,"Title":"keras model evaluation with quantized weights post training","Body":"<p>I have a model trained in keras and is saved as a .h5 file. The model is trained with single precision floating point values with tensorflow backend. Now I want to implement an hardware accelerator which performs the convolution operation on an Xilinx FPGA. However, before I decide on the fixed point bit width to be used on the FPGA, I need to evaluate the model accuracy by quantizing the weights to 8 or 16 bit numbers. I came across the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.12\/api_docs\/python\/tf\/quantization\/quantize\" rel=\"nofollow noreferrer\">tensorflow quantise<\/a> but I am not sure how I can go about taking weights from each layer, quantise it and store it in a list of numpy arrays. After all layers are quantised, I want to set the weights of the model to the new formed quantised weights. Could someone help me do this?<\/p>\n\n<p>This is what I have tried so far to reduce precision from float32 to float16. Please let me know if this is the correct approach.<\/p>\n\n<pre><code>for i in range(len(w_orginal)):\ntemp_shape = w_orginal[i].shape\nprint('Shape of index: '+ str(i)+ 'array is :')\nprint(temp_shape)\ntemp_array = w_orginal[i]\ntemp_array_flat = w_orginal[i].flatten()\nfor j in range(len(temp_array)):\n    temp_array_flat[j] = temp_array_flat[j].astype(np.float16)\n\ntemp_array_flat = temp_array_flat.reshape(temp_shape)\nw_fp_16_test.append(temp_array_flat)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52851892","CreationDate":"2018-10-17T09:47:45.420","ParentId":null,"OwnerUserId":"5561494","Title":null,"Body":"<p>Sorry for that I'm not familiar to tensorflow, so I can't give you the code, but maybe my experience with quantizing a caffe model could make sense.<\/p>\n\n<p>If I understand you correctly, you have a tensorflow model(float32) which you want to quantize it to int8 and save it in a <code>numpy.array<\/code>.<\/p>\n\n<p>Firstly, you should read all weights for each layer, which might be python list or <code>numpy.array<\/code> or something else, it does't matter.<\/p>\n\n<p>Then, the quantize algorithm will influence the accuracy significantly, you must choose the best one for your model. However, these algorithms have the same core -- scale. All you need to do is scale all the weights to -127 to 127(int8), like the <code>scale<\/code> layer without <code>bias<\/code>, and record the scale factor.<\/p>\n\n<p>Meanwile, if want to implement it on FPGA, the data should be qiantized too. Here we have a new problem -- the result of int8 * int8 is a int16, which is obvious overflow.<\/p>\n\n<p>To solve this, we create a new parameter -- shift -- to shift int16 result back to int8. Notice, the <code>shift<\/code> parameter won't be constant 8, suppose we have 0 * 0 = 0, we don't need to shift the result at all.<\/p>\n\n<p>The last question we shoud think over is that if the net is too deep, the layer result could overflow because some unreasonable <code>scale<\/code> parameters, so we can't directly quantize each single layer without think about other layers.<\/p>\n\n<p>After all the net finished on FPGA, if you want to dequantize int8 to float32, just use the last scale parameter(of final result) to do some mul\/div(depend on how you define <code>scale<\/code>). <\/p>\n\n<p>This is a basic quantize algorithm, others like <code>tf.quantization<\/code> may have higher accuracy. Now we have the quantized model, you can save it into whatever you like, it's not a hard work.<\/p>\n\n<p>P.S. Why numpy? bin file is the best for FPGA, isn't it?<\/p>\n\n<p>And, do you have some idea about implementing softmax on FPGA? I'm confused about it...<\/p>\n"}]}
+{"QuestionId":52820038,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-15T15:28:17.903","AcceptedAnswerId":52820758.0,"OwnerUserId":1202808.0,"Title":"Tensorflow \/ Keras Tutorial Save \/ Load Model not work","Body":"<p>I am unable to save a simple keras Sequential model as per the tutorial using Tensorflow 1.11.0. I am able to convert the model configuration to a json string and I am able to save the weights, but not the model in its entirety.<\/p>\n\n<p>The error comes from the first line of code in <code>Model.save<\/code> which checks that <code>self._is_graph_network<\/code> is <code>True<\/code>. I feel like there is something basic that I am missing, but I cannot figure it out. I believe my code is analogous to the tutorial code found here: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/save_and_restore_models\" rel=\"nofollow noreferrer\">Keras Save and Restore<\/a> in the Save Entire Model section.<\/p>\n\n<pre><code>Python 3.6.6 (default, Jun 28 2018, 05:43:53)\n[GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.2)] on darwin\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; import tensorflow.keras as tk\n&gt;&gt;&gt; mnist = tk.datasets.mnist\n&gt;&gt;&gt; (xtrain, ytrain), (xtest, ytest) = mnist.load_data()\n&gt;&gt;&gt; xtrain, xtest = xtrain\/255.0, xtest\/255.0\n&gt;&gt;&gt; model = tk.models.Sequential([tk.layers.Flatten(), tk.layers.Dense(64, activation=tf.nn.relu), \\\ntk.layers.Dense(10, activation=tf.nn.softmax)])\n&gt;&gt;&gt; model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n&gt;&gt;&gt; model.fit(xtrain, ytrain, epochs=5)\nEpoch 1\/5\n60000\/60000 [==============================] - 2s 39us\/step - loss: 0.3086 - acc: 0.9126\n...\nEpoch 5\/5\n60000\/60000 [==============================] - 2s 37us\/step - loss: 0.0715 - acc: 0.9787\n&lt;tensorflow.python.keras.callbacks.History object at 0x129819470&gt;\n&gt;&gt;&gt; model.evaluate(xtest, ytest)\n10000\/10000 [==============================] - 0s 17us\/step\n[0.09259423367818817, 0.9711]\n&gt;&gt;&gt; model.save('m0x.h5')\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/Users\/mmelton\/Library\/Python\/3.6\/lib\/python\/site-packages\/tensorflow\/python\/keras\/engine\/n\\\network.py\", line 1358, in save\n    raise NotImplementedError\nNotImplementedError\n&gt;&gt;&gt; model._is_graph_network\nFalse\n&gt;&gt;&gt; model.save_weights('w0.h5', save_format='h5')\n&gt;&gt;&gt;\n\n&gt;&gt;&gt; tf.__version__\n'1.11.0'\n&gt;&gt;&gt; tk.__version__\n'2.1.6-tf'\n&gt;&gt;&gt;\n<\/code><\/pre>\n\n<p>What am I missing?<\/p>\n","answers":[{"AnswerId":"52820758","CreationDate":"2018-10-15T16:10:14.557","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<pre><code>model = tk.models.Sequential([\n    tk.layers.Flatten(), \n    tk.layers.Dense(64, activation=tf.nn.relu), \n    tk.layers.Dense(10, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n\n<p>Your network has no inputs -  <code>Flatten(input_shape=(784,))<\/code>. Also, having <code>Flatten()<\/code> as your first layer is a little funny but go for it if you want.<\/p>\n"}]}
+{"QuestionId":52820076,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-15T15:30:07.883","AcceptedAnswerId":null,"OwnerUserId":7400494.0,"Title":"Tensorflow device only acknowledges CPU even when explicitly specifying tf.device('gpu:0')","Body":"<p>After many attempts to verify that my nvidia quadro M4000M is capable of tensor computations,  Cuda installed correctly,  cuda toolkit 8.0 installed correctly,  deviceQuery and nvidia-smi confirms that the device is functioning properly.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SU2BP.jpg\" rel=\"nofollow noreferrer\">deviceQuery<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/li8Ki.jpg\" rel=\"nofollow noreferrer\">nvidia-smi<\/a><\/p>\n\n<p>When running tensorflow's example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.device(\"\/gpu:0\"):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a', dtype=tf.float32)\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b', dtype=tf.float32)\n    c = tf.matmul(a, b)\n\nsess = tf.Session(config=tf.ConfigProto(allow_soft_placement=False,log_device_placement=True)) \nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>I don't get the device mapping information and the only way I can get the  session to complete is to change the allow_soft_placement=True, which doesn't appear to use the GPU.<\/p>\n\n<p>When allow_soft_placement = False for tf.device(\"\/gpu:0\") or \/cpu:0 the error is:<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>InvalidArgumentError                      Traceback (most recent call last)\n    ~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _do_call(self, fn, *args)    1291     try:\n    -> 1292       return fn(*args)    1293     except errors.OpError as e:<\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata) \n    1274       # Ensure any changes to the graph are reflected in the\n    runtime.\n    -> 1275       self._extend_graph()    1276       return self._call_tf_sessionrun(<\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _extend_graph(self)    1311     with\n    self._graph._session_run_lock():  # pylint: disable=protected-access\n    -> 1312       tf_session.ExtendSession(self._session)    1313 <\/p>\n    \n    <p>InvalidArgumentError: Cannot assign a device for operation 'MatMul': Operation was explicitly assigned to \/device:GPU:0 but available\n    devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make\n    sure the device specification refers to a valid device.    [[{{node\n    MatMul}} = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false,\n    _device=\"\/device:GPU:0\"](a, b)]]<\/p>\n    \n    <p>During handling of the above exception, another exception occurred:<\/p>\n    \n    <p>InvalidArgumentError                      Traceback (most recent call last)  in \n          7 sess =  tf.Session(config=tf.ConfigProto(allow_soft_placement=False,log_device_placement=True))\n          8 \n    ----> 9 print(sess.run(c))\n         10 \n         11 <\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in run(self, fetches, feed_dict, options, run_metadata)\n        885     try:\n        886       result = self._run(None, fetches, feed_dict, options_ptr,\n    --> 887                          run_metadata_ptr)\n        888       if run_metadata:\n        889         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)<\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _run(self, handle, fetches, feed_dict, options, run_metadata)<br>\n    1108     if final_fetches or final_targets or (handle and\n    feed_dict_tensor):    1109       results = self._do_run(handle,\n    final_targets, final_fetches,\n    -> 1110                              feed_dict_tensor, options, run_metadata)    1111     else:    1112       results = []<\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _do_run(self, handle, target_list, fetch_list, feed_dict, options,\n    run_metadata)    1284     if handle is None:    1285       return\n    self._do_call(_run_fn, feeds, fetches, targets, options,\n    -> 1286                            run_metadata)    1287     else:    1288       return self._do_call(_prun_fn, handle, feeds, fetches)<\/p>\n    \n    <p>~\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\client\\session.py\n    in _do_call(self, fn, *args)    1306<br>\n    self._config.experimental.client_handles_error_formatting):    1307<br>\n    message = error_interpolation.interpolate(message, self._graph)\n    -> 1308       raise type(e)(node_def, op, message)    1309     1310   def _extend_graph(self):<\/p>\n    \n    <p>InvalidArgumentError: Cannot assign a device for operation 'MatMul': Operation was explicitly assigned to \/device:GPU:0 but available\n    devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make\n    sure the device specification refers to a valid device.    [[{{node\n    MatMul}} = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false,\n    _device=\"\/device:GPU:0\"](a, b)]]<\/p>\n    \n    <p>Caused by op 'MatMul', defined at:   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\runpy.py\",\n    line 193, in _run_module_as_main\n        \"<strong>main<\/strong>\", mod_spec)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\runpy.py\",\n    line 85, in _run_code\n        exec(code, run_globals)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel_launcher.py\",\n    line 16, in \n        app.launch_new_instance()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\traitlets\\config\\application.py\",\n    line 658, in launch_instance\n        app.start()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\kernelapp.py\",\n    line 499, in start\n        self.io_loop.start()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\platform\\asyncio.py\",\n    line 132, in start\n        self.asyncio_loop.run_forever()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\asyncio\\base_events.py\",\n    line 422, in run_forever\n        self._run_once()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\asyncio\\base_events.py\",\n    line 1434, in _run_once\n        handle._run()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\asyncio\\events.py\",\n    line 145, in _run\n        self._callback(*self._args)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\ioloop.py\",\n    line 758, in _run_callback\n        ret = callback()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\stack_context.py\",\n    line 300, in null_wrapper\n        return fn(*args, **kwargs)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\gen.py\",\n    line 1233, in inner\n        self.run()   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\gen.py\",\n    line 1147, in run\n        yielded = self.gen.send(value)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\kernelbase.py\",\n    line 346, in process_one\n        yield gen.maybe_future(dispatch(*args))   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\gen.py\",\n    line 326, in wrapper\n        yielded = next(result)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\kernelbase.py\",\n    line 259, in dispatch_shell\n        yield gen.maybe_future(handler(stream, idents, msg))   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\gen.py\",\n    line 326, in wrapper\n        yielded = next(result)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\kernelbase.py\",\n    line 513, in execute_request\n        user_expressions, allow_stdin,   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tornado\\gen.py\",\n    line 326, in wrapper\n        yielded = next(result)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\ipkernel.py\",\n    line 294, in do_execute\n        res = shell.run_cell(code, store_history=store_history, silent=silent)   File\n    \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\ipykernel\\zmqshell.py\",\n    line 536, in run_cell\n        return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)   File\n    \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\interactiveshell.py\",\n    line 2817, in run_cell\n        raw_cell, store_history, silent, shell_futures)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\interactiveshell.py\",\n    line 2843, in _run_cell\n        return runner(coro)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\async_helpers.py\",\n    line 67, in _pseudo_sync_runner\n        coro.send(None)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\interactiveshell.py\",\n    line 3018, in run_cell_async\n        interactivity=interactivity, compiler=compiler, result=result)   File\n    \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\interactiveshell.py\",\n    line 3183, in run_ast_nodes\n        if (yield from self.run_code(code, result)):   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\IPython\\core\\interactiveshell.py\",\n    line 3265, in run_code\n        exec(code_obj, self.user_global_ns, self.user_ns)   File \"\", line 5, in \n        c = tf.matmul(a, b)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\ops\\math_ops.py\",\n    line 2053, in matmul\n        a, b, transpose_a=transpose_a, transpose_b=transpose_b, name=name)   File\n    \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\ops\\gen_math_ops.py\",\n    line 4856, in mat_mul\n        name=name)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\",\n    line 787, in _apply_op_helper\n        op_def=op_def)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\",\n    line 488, in new_func\n        return func(*args, **kwargs)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\",\n    line 3272, in create_op\n        op_def=op_def)   File \"C:\\Users\\sellersp\\AppData\\Local\\conda\\conda\\envs\\tensor\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\",\n    line 1768, in <strong>init<\/strong>\n        self._traceback = tf_stack.extract_stack()<\/p>\n    \n    <p>InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'MatMul': Operation was explicitly assigned to\n    \/device:GPU:0 but available devices are [\n    \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make sure the device\n    specification refers to a valid device.    [[{{node MatMul}} =\n    MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false,\n    _device=\"\/device:GPU:0\"](a, b)]]<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>After searching many locations I can't seem to find where my error is.   I've run it in a Jupyter Notebook and as a script at the interpreter.<\/p>\n\n<p>Any help would be greatly appreciated.  Sorry for the lengthy error.<\/p>\n","answers":[{"AnswerId":"52821132","CreationDate":"2018-10-15T16:35:52.543","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Have you installed tensorflow-gpu? Tensorflow has two binary packages (tensorflow-gpu and tensorflow). <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a><\/p>\n"}]}
+{"QuestionId":52820615,"AnswerCount":1,"Tags":"<tensorflow><neural-network><computer-vision><conv-neural-network><object-detection-api>","CreationDate":"2018-10-15T16:01:10.517","AcceptedAnswerId":null,"OwnerUserId":7180063.0,"Title":"Tensorflow object detection api. Detecting 90 default classes + n new classes at a time","Body":"<p>I am using tensorflow object detection api. I was able to detect default 90 classes using pre-trained models. \nI was also able to detect only custom objects (Ex: <code>macncheese<\/code>) by training on new dataset that has only <code>macncheese<\/code>. <br\/>\nI am having a hard time trying to combine 90 (default)+ 1(<code>macncheese<\/code>) object classes for detection. I heard of the solutions like training the 91 classes\nfrom scratch by adding a new dataset to the coco original dataset. I am also trying to go back 1 layer of the pre trained model and train only the new data\nto detect 91 classes. I am new to computer vision so i am not well versed with training a CNN. can you please suggest some solutions.<\/p>\n\n<p>Model used: Single shot detector.\nOS: Windows 10\nPlatform: Anaconda<\/p>\n","answers":[{"AnswerId":"52848420","CreationDate":"2018-10-17T06:20:02.113","ParentId":null,"OwnerUserId":"6918487","Title":null,"Body":"<p>If you want all 91 classes to be recognized, you would have to retrain on training date set which has all 91 classes. Like you did for the one macncheese. Is there any specific question beyond this you are struggling with?<\/p>\n"}]}
+{"QuestionId":52820658,"AnswerCount":0,"Tags":"<python><tensorflow><rnn>","CreationDate":"2018-10-15T16:03:53.933","AcceptedAnswerId":null,"OwnerUserId":5647864.0,"Title":"Sentence classification with tensorflow (Dynamic rnn)","Body":"<p>I'm trying to implement a rnn model for sentence classification.\nPrecisely this is the schema I'm trying to implement<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1T8u9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1T8u9.png\" alt=\"Rnn schema\"><\/a>\nSentences are first encoded with word2vec vectors and then fed into the rnn, they can have different length hence I'm using a dynamic rnn.<\/p>\n\n<p>Problem is that when running it seems the model is not learning at all.<\/p>\n\n<p>Below you find my current implementation:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.reset_default_graph()\n\nonehot_encoding_articles = tf.one_hot(int_article, len(set(articleIds)))\n\nn_articles = len(set(articleIds))\nseqLENGTH = tf.placeholder(tf.int32, [RNN_batch_size])\n\ninputX = tf.placeholder(tf.int32, [RNN_batch_size, None])\ninputs = tf.nn.embedding_lookup(embeddings, inputX)\ninputY = tf.placeholder(tf.int32, [RNN_batch_size])\nlabels = tf.nn.embedding_lookup(onehot_encoding_articles, inputY)\n\n\nrnn_layers = [tf.nn.rnn_cell.LSTMCell(size) for size in [128, 256]]\n\nmulti_rnn_cell = tf.nn.rnn_cell.MultiRNNCell(rnn_layers)\n\ninit_state = multi_rnn_cell.zero_state(RNN_batch_size, tf.float64)\n\noutputs, state = tf.nn.dynamic_rnn(cell=multi_rnn_cell,\n                               inputs=inputs,\n                               initial_state=init_state,\n                               sequence_length=seqLENGTH,\n                               dtype=tf.float64)\n\ndense_layer = tf.layers.dense(state[1].h,32)\n\nlogits = tf.layers.dense(dense_layer,n_articles)\n\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=labels,logits=logits)\nloss = tf.reduce_mean(cross_entropy)\noptimizer = tf.train.AdamOptimizer().minimize(loss)\n<\/code><\/pre>\n\n<p>Then I run it with<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\n\nwith tf.Session() as sess: \n    sess.run(init)\n    for epoch in range(1,5):\n        batches = utils.get_batches_rnn(RNN_batch_size,int_wordsPerInteractionInput[:1000],int_article[:1000], len(embeddings)-1)\n        for x, y, size in batches:\n            feed = {inputX: x,\n                inputY: y,\n                seqLENGTH: size}\n            train_loss, _ = sess.run([loss, optimizer], feed_dict=feed)\n            print(\"Avg. Training loss: {:.4f}\".format(train_loss))\n<\/code><\/pre>\n\n<p>function get_batches_rnn is defined as follow<\/p>\n\n<pre><code>def get_batches_rnn(n_sentences_in_batch, sentences, articleIds, pad_index):  \n\n    for idx in range(0, len(sentences), n_sentences_in_batch):\n        x, y, batch_size = [], [], []\n        x_batch = sentences[idx:idx+n_sentences_in_batch]\n        y_batch = articleIds[idx:idx+n_sentences_in_batch]\n        max_batch = max(len(x) for x in x_batch)\n        for batchIndex in range(len(x_batch)):\n        padding = [word for word in x_batch[batchIndex]]\n           padSize = max_batch - len(x_batch[batchIndex])\n           padding.extend(np.full((padSize), pad_index, dtype=int))\n           x.append(padding)\n           y.append(y_batch[batchIndex])\n           batch_size.append(len(x_batch[batchIndex]))\n        yield np.asarray(x), np.asarray(y), np.asarray(batch_size)\n<\/code><\/pre>\n\n<p>and this is an example of what is returned by the function<\/p>\n\n<pre><code>batches = utils.get_batches_rnn(3,int_wordsPerInteractionInput,int_article,len(embeddings)-1)\nx, y, size = next(batches)\nprint(x)\nprint(y)\nprint(size)\n\nOutput:\n\n[[   20     0 23342]\n [  122     8 23342]\n [  257   243    82]]\n[452 468 521]\n[2 2 3]\n<\/code><\/pre>\n\n<p>Basically the id of the word (that is then converted into word vector), the id of the class (that is then converted into one-hot encoded vector) and the length of the sentence.<\/p>\n\n<p>I am padding with the id of the last word vector embedding (23342) because it is a zero vector.<\/p>\n\n<p>Any hint about what I am doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":52820883,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-model-analysis>","CreationDate":"2018-10-15T16:18:38.023","AcceptedAnswerId":null,"OwnerUserId":8060613.0,"Title":"ImportError: import tensorflow_model_analysis as tfma [ PYTHON on datalab]","Body":"<p>I would like to add model analysis to my model but was unable to import the libraries:<\/p>\n\n<p>I am using datalab environment.<\/p>\n\n<pre><code>import tensorflow as tf\n!pip install tensorflow_model_analysis\nimport tensorflow_model_analysis as tfma\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><code>ImportErrorTraceback (most recent call last)\n&lt;ipython-input-6-f85e4d8fbd99&gt; in &lt;module&gt;()\n----&gt; 1 import tensorflow_model_analysis as tfma\n\n\/usr\/local\/envs\/py2env\/lib\/python2.7\/site-packages\/tensorflow_model_analysis\/__init__.py in &lt;module&gt;()\n     15 \n     16 \n---&gt; 17 from tensorflow_model_analysis import view\n     18 from tensorflow_model_analysis.api import tfma_unit as test\n     19 from tensorflow_model_analysis.api.model_eval_lib import *  # pylint: disable=wildcard-import\n\nImportError: cannot import name view\n<\/code><\/pre>\n\n<p>Please advice what will be the right way to import the library.<\/p>\n\n<p>Thanks,\neilalan<\/p>\n","answers":[{"AnswerId":"53395051","CreationDate":"2018-11-20T14:20:31.193","ParentId":null,"OwnerUserId":"8060613","Title":null,"Body":"<p>Working for me now with the following installation:<\/p>\n\n<pre><code>python 2.7 - to support apache beam\npip install pip==9.0.3  # I am not sure what is the reason, but essential for apache beam pipelines execution\npip install --upgrade tensorflow\npip install tensorflow-model-analysis\nimport tensorflow_model_analysis as tfma\n<\/code><\/pre>\n"}]}
+{"QuestionId":52821066,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving><google-cloud-ml>","CreationDate":"2018-10-15T16:31:28.917","AcceptedAnswerId":53052323.0,"OwnerUserId":2455494.0,"Title":"Cloud-ml unable to deploy model with tables_initializer","Body":"<p>I am looking to deploy a trained model to ml-engine. I can locally run my code with something like:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  sess.run([tf.global_variables_initializer(), tf.tables_initializer()])\n  example_result = sess.run(\n                    my_model, \n                    feed_dict=###snip###\n                  )\n<\/code><\/pre>\n\n<p>I have been trying to export with:<\/p>\n\n<pre><code>export_builder.add_meta_graph_and_variables(\n        sess, \n        [tf.saved_model.tag_constants.SERVING], \n        signature_def_map={\n            tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: sig\n        })\n<\/code><\/pre>\n\n<p>which I can then deploy to ml-engine. However, when I call the model I get an error as the tables are not initialised.<\/p>\n\n<p><strong>How can I deploy a model where the tables are initialized automatically\/by default?<\/strong><\/p>\n\n<hr>\n\n<p><strong>Things I have tried:<\/strong><\/p>\n\n<ol>\n<li>Adding the <code>tf.tables_initializer()<\/code> as a dependency before any of the ops in <code>my_model<\/code><\/li>\n<\/ol>\n\n<p>This works for the <em>first<\/em> time I call my deployed model but fails every subsequent call with an error message telling me the table is already initialized<\/p>\n\n<ol start=\"2\">\n<li>Passing the <code>tf.tables_initializer()<\/code> as the <code>legacy_init_op<\/code> parameter to the <code>add_meta_graph_and_variables<\/code> method.<\/li>\n<\/ol>\n\n<p>this fails to deploy to ml-engine at all with the error message <\/p>\n\n<blockquote>\n  <p>Create Version failed. Bad model detected with error: \"Failed to load\n  model: a bytes-like object is required, not 'str' (Error code: 0)\"<\/p>\n<\/blockquote>\n\n<ol start=\"3\">\n<li>Passing the <code>tf.tables_initializer()<\/code> or <code>tf.saved_model.main_op.main_op()<\/code> as the <code>main_op<\/code> parameter to the <code>add_meta_graph_and_variables<\/code> method.<\/li>\n<\/ol>\n\n<p>In each case, it still fails to deploy the model with the same error message as in (2.):<\/p>\n\n<blockquote>\n  <p>Create Version failed. Bad model detected with error: \"Failed to load\n  model: a bytes-like object is required, not 'str' (Error code: 0)\"<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53052323","CreationDate":"2018-10-29T19:14:42.043","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>We must pass the tables initializer op as the <code>main_op<\/code> but also, crucially, we need to include the <code>assets_collection<\/code> in our call to <code>add_meta_graph_and_variables<\/code>:<\/p>\n\n<pre><code>init_op = tf.tables_initializer()\n\nwith tf.Session() as sess:\n  sess.run([tf.global_variables_initializer(), init_op])\n  example_result = sess.run(\n                my_model, \n                feed_dict=###snip###\n              )\n<\/code><\/pre>\n\n<p>then<\/p>\n\n<pre><code>\n    export_builder.add_meta_graph_and_variables(\n        sess, \n        [tf.saved_model.tag_constants.SERVING],\n        <b>main_op = init_op,<\/b>\n        <b>assets_collection=tf.get_collection(tf.GraphKeys.ASSET_FILEPATHS),<\/b>\n        signature_def_map={\n            tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: sig\n        }\n    )\n\n<\/code><\/pre>\n\n<hr>\n\n<p>As an (arguably simpler) alternative, we could use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\/simple_save\" rel=\"nofollow noreferrer\"><code>simple_save<\/code><\/a> method to save our model<\/p>\n"}]}
+{"QuestionId":52821319,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-15T16:49:14.687","AcceptedAnswerId":null,"OwnerUserId":3946761.0,"Title":"Pre-trained Weight Initialization Problem with Skip Connections in Keras","Body":"<p>I am trying to implement pre-trained VGG-16 for weight initializing along with Skip Connections. Although model's summary shows correct number of parameters, but weights are not initialized properly. My code is given below:<\/p>\n\n<pre><code>vgg16_model = applications.vgg16.VGG16(weights='imagenet', include_top=False,input_shape = (224, 224, 3) )\nL1 = vgg16_model.get_layer('block1_conv1').output\nL3 = vgg16_model.get_layer('block2_conv1').output\nL5 = vgg16_model.get_layer('block3_conv1').output\nL6 = vgg16_model.get_layer('block3_conv2').output\nL8 = vgg16_model.get_layer('block4_conv1').output\nL9 = vgg16_model.get_layer('block4_conv2').output\nL11 = vgg16_model.get_layer('block5_conv1').output\nL12 = vgg16_model.get_layer('block5_conv2').output\nL13= vgg16_model.get_layer('block5_pool').output\nskipConnectionLayers = Concatenate()([L1,L3,L5,L6,L8,L9,L11,L12])\nx = Flatten(name='flatten_1')(L13)\nx = Concatenate()([skipConnectionLayers, x]) #weights are initialized propery if this line is commented\nx = Dense(10, activation='softmax', name='fc4')(x)\nmy_model = Model(input=vgg16_model.input, output=x)\n<\/code><\/pre>\n\n<p>In my opinion, as I am setting include_top = False and making changes in deeper layers, weight initialization should work fine.<\/p>\n\n<p>Thanks,<\/p>\n","answers":[]}
+{"QuestionId":52821355,"AnswerCount":0,"Tags":"<deep-learning><conv-neural-network><pytorch>","CreationDate":"2018-10-15T16:52:09.093","AcceptedAnswerId":null,"OwnerUserId":7245020.0,"Title":"AlexNet paper does not match dimensions they reported?","Body":"<p>I am following the <a href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow noreferrer\">paper where AlexNet was introduced<\/a>, and the dimensions they report just don't match with the figure they attached.<\/p>\n\n<p>The output of the first conv layer (which is 96 11x11x3 convolution) is 55x55x96 (for simple 1GPU case). Now the paper states that the second conv layer is applied to the output of a maxpooling layer. Assuming the MaxPool is a kernel of 3x3 with stride of 2 (as they report with s and z), this means that the input to the second convolution layer should be (55 - 3)\/2 + 1 = 27 but in the picture provided for the AlexNet they write there is a Max Pooling operation there but they don't perform the dimension reduction of the pooling!<\/p>\n\n<p>So the second conv layer should have been applied to a volumen with widht and height = 27 and not 55 right?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hoXIq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hoXIq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Furthermore, I took a look at how PyTorch was implementing it to see if I was missing something and they just changed the configuration, starting with 64 kernels... :<\/p>\n\n<pre><code>AlexNet(\n  (features): Sequential(\n    (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))\n    (1): ReLU(inplace)\n    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))\n    (4): ReLU(inplace)\n    (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (7): ReLU(inplace)\n    (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (9): ReLU(inplace)\n    (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (11): ReLU(inplace)\n    (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n  )\n  (classifier): Sequential(\n    (0): Dropout(p=0.5)\n    (1): Linear(in_features=9216, out_features=4096, bias=True)\n    (2): ReLU(inplace)\n    (3): Dropout(p=0.5)\n    (4): Linear(in_features=4096, out_features=4096, bias=True)\n    (5): ReLU(inplace)\n    (6): Linear(in_features=4096, out_features=1000, bias=True)\n  )\n)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52821476,"AnswerCount":3,"Tags":"<tensorflow><tensorflow-hub>","CreationDate":"2018-10-15T17:00:31.357","AcceptedAnswerId":null,"OwnerUserId":4921197.0,"Title":"how to solve the no 'estimator' in tensorflow?","Body":"<p>I am trying to implement a code from tensorflow: <\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/hub\/blob\/master\/examples\/colab\/semantic_similarity_with_tf_hub_universal_encoder.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/hub\/blob\/master\/examples\/colab\/semantic_similarity_with_tf_hub_universal_encoder.ipynb<\/a><\/p>\n<\/blockquote>\n\n<p>However, I had the error of tensorflow lib as follows, I have tried uninstall and reinstall the tensorflow but no working, is anyone facing this error before?<\/p>\n\n<pre><code>  import tensorflow as tf\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/tensorflow\/python\/__init__.py\", line 52, in &lt;module&gt;\n    from tensorflow.core.framework.graph_pb2 import *\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/tensorflow\/core\/framework\/graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/google\/protobuf\/__init__.py\", line 37, in &lt;module&gt;\n    __import__('pkg_resources').declare_namespace(__name__)\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 3112, in &lt;module&gt;\n    @_call_aside\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 3096, in _call_aside\n    f(*args, **kwargs)\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 3140, in _initialize_master_working_set\n    for dist in working_set\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 3140, in &lt;genexpr&gt;\n    for dist in working_set\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 2647, in activate\n    declare_namespace(pkg)\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 2184, in declare_namespace\n    _handle_ns(packageName, path_item)\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/pkg_resources\/__init__.py\", line 2117, in _handle_ns\n    loader.load_module(packageName)\n  File \"\/data\/d14127800\/dataset10\/google.py\", line 4, in &lt;module&gt;\n    import tensorflow_hub as hub\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/tensorflow_hub\/__init__.py\", line 23, in &lt;module&gt;\n    from tensorflow_hub.estimator import LatestModuleExporter\n  File \"\/data\/d14127800\/py3_170\/lib\/python3.5\/site-packages\/tensorflow_hub\/estimator.py\", line 61, in &lt;module&gt;\n    class LatestModuleExporter(tf.estimator.Exporter):\nAttributeError: module 'tensorflow' has no attribute 'estimator'\n<\/code><\/pre>\n","answers":[{"AnswerId":"54277128","CreationDate":"2019-01-20T13:53:19.000","ParentId":null,"OwnerUserId":"8573240","Title":null,"Body":"<p>I also face similar error but it was in Darkflow \nAfter doing much google nothing worked for me and did many attempt to resolve it\nBut I came out with solution just by doing <\/p>\n\n<p>for Anaconda \n->conda install matplotlib\n->conda install pandas\n(It should be updated Version of Both)\nAfter doing this ReStart Your NoteBook<\/p>\n\n<p>For Normal IDLE\n->pip install pandas\n->pip install matplotlib<\/p>\n"},{"AnswerId":"52834717","CreationDate":"2018-10-16T11:43:42.797","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>Estimator was moved from tf.contrib.learn.Estimator to tf.estimator.Estimator with the TF 1.1.0 release. Make sure that your Tensorflow version is higher using:<\/p>\n\n<pre><code>pip show tensorflow\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>pip show tensorflow-gpu \n<\/code><\/pre>\n\n<p>if you are using the GPU version.<\/p>\n\n<p>If it is lower, you can update Tensorflow with:<\/p>\n\n<pre><code>pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>pip isntall --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p>or use tf.contrib.learn.Estimator instead.<\/p>\n"},{"AnswerId":"56123668","CreationDate":"2019-05-14T05:40:42.770","ParentId":null,"OwnerUserId":"1474204","Title":null,"Body":"<p>I changed the tensorflow-estimator to 1.10.12 by doing <code>pip install tensorflow-estimator==1.10.12<\/code> and it resolved the issue.<\/p>\n"}]}
+{"QuestionId":52821584,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-15T17:06:58.663","AcceptedAnswerId":52834626.0,"OwnerUserId":7447265.0,"Title":"What is going on tf.train.shuffle_batch()","Body":"<p>I am very new to the tensorflow framework and I tried this code to read and explore the CIFAR-10 dataset.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport os\nimport matplotlib.pyplot as plt\n\nsess=tf.Session()\n\nbatch_size = 128\noutput_every = 50\ngenerations = 20000\neval_every = 500\nimage_height = 32\nimage_width = 32\ncrop_height = 24\ncrop_width = 24\nnum_channels = 3\nnum_targets = 10\ndata_dir=\"CIFAR10\"\n\n\nimage_vec_length = image_height * image_width * num_channels\nrecord_length = 1 + image_vec_length\n\ndef read_cifar_files(filename_queue, distort_images = True):\n   reader = tf.FixedLengthRecordReader(record_bytes=record_length*10)\n   key, record_string = reader.read(filename_queue)\n   record_bytes = tf.decode_raw(record_string, tf.uint8)\n\n# Extract label\n   image_label = tf.cast(tf.slice(record_bytes, [image_vec_length-1],[1]),tf.int32)\n\n# Extract image\n   sliced=tf.slice(record_bytes, [0],[image_vec_length])\n   image_extracted = tf.reshape(sliced, [num_channels, image_height,image_width])\n\n# Reshape image\n   image_uint8image = tf.transpose(image_extracted, [1, 2, 0])\n   reshaped_image = tf.cast(image_uint8image, tf.float32)\n\n# Randomly Crop image\n   final_image = tf.image.resize_image_with_crop_or_pad(reshaped_image, crop_width, crop_height)\n   if distort_images:\n\n# Randomly flip the image horizontally, change the brightness and contrast\n     final_image = tf.image.random_flip_left_right(final_image)\n     final_image = tf.image.random_brightness(final_image,max_delta=63)\n     final_image = tf.image.random_contrast(final_image,lower=0.2, upper=1.8)\n\n# standardization\n     final_image = tf.image.per_image_standardization(final_image)\n     return  final_image, image_label\n<\/code><\/pre>\n\n<p>when I run the following input_pipeline() function without tf.train.shuffle_batch() it gives me single image tensor with shape (24,24,3). <\/p>\n\n<pre><code>def input_pipeline(batch_size, train_logical=True):\n    files=[os.path.join(data_dir,\"data_batch_{}.bin\".format(i)) for i in range(1,6)]\n    filename_queue = tf.train.string_input_producer(files)\n    image,label = read_cifar_files(filename_queue)\n    return(image,label)\n\n\nexample_batch,label_batch=input_pipeline(batch_size)\nthreads = tf.train.start_queue_runners(sess=sess)\nimg,label=sess.run([example_batch, label_batch])\n\n#output=(24,24,3) \nprint(img.shape) \n<\/code><\/pre>\n\n<p>But when I run same input_pipeline() function with tf.train.shuffle_batch() function it gives me image tensor which contain 128 images with shape (128, 24, 24, 3).<\/p>\n\n<pre><code>def input_pipeline(batch_size, train_logical=True):\n    files=[os.path.join(data_dir,\"data_batch_{}.bin\".format(i)) for i in range(1,6)]\n    filename_queue = tf.train.string_input_producer(files)\n    image,label = read_cifar_files(filename_queue)\n\n    min_after_dequeue = 1000\n    capacity = min_after_dequeue + 3 * batch_size\n    example_batch, label_batch = tf.train.shuffle_batch([image,label], batch_size, capacity, min_after_dequeue)\n    return(example_batch, label_batch)\n<\/code><\/pre>\n\n<p>how is that possible. it seems that tf.train.shuffle_batch() take single image tensor from read_cifar_files() and return tensor which has 128 images.So what is going on tf.train.shuffle_batch() function.<\/p>\n","answers":[{"AnswerId":"52834626","CreationDate":"2018-10-16T11:38:05.093","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>In Tensorflow, a Tensor is just a node of the graph. The tf.train.shuffle_batch() function takes as input 2 nodes that are connected, thanks to the graph, to data. <\/p>\n\n<p>So it doesn't take 'a single image' as input but a graph that is able to load an image. It then add a new operation to the graph that will execute n = batch_size time the input graph, shuffle the batch and returns an output Tensor of size [bach_size, input_shape].<\/p>\n\n<p>Then, when you run the function in a session, the data will be loaded according to the graph, meaning that every time you call tf.train.shuffle_batch(), you will read n = batch_size images on disk.<\/p>\n"}]}
+{"QuestionId":52821633,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-15T17:10:26.863","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Exporting Keras Model Without Dropout Seemingly Fails","Body":"<p>I have a Keras model with a Tensorflow backend (convolutional neural network) called <code>model<\/code>.  I'm trying to save it without dropout, so it can be used at inference time.<\/p>\n\n<p>I'm doing this:<\/p>\n\n<pre><code>tf.keras.backend.set_learning_phase(0)\nyaml_string = model.to_yaml()\n<\/code><\/pre>\n\n<p>When I inspect <code>yaml_string<\/code>, I see this included in the string:<\/p>\n\n<pre><code>  - class_name: SpatialDropout1D\n    config: {dtype: float32, name: spatial_dropout1d, noise_shape: null, rate: 0.25,\n      seed: null, trainable: true}\n    inbound_nodes:\n    - - - my_embedding_0\n        - 0\n        - 0\n        - {}\n    name: spatial_dropout1d\n<\/code><\/pre>\n\n<p>This makes me think the <code>spatial_dropout1d<\/code> layer is still there.  Is that a correct interpretation?<\/p>\n\n<p>If so, how should I be exporting this model without dropout?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52821996,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><lstm>","CreationDate":"2018-10-15T17:34:33.423","AcceptedAnswerId":null,"OwnerUserId":9023836.0,"Title":"early stopping in lstm with Python","Body":"<p>How to do early stopping in lstm.<\/p>\n\n<p>I am using python tensorflow but not keras.<\/p>\n\n<p>I would appreciate if you can provide a sample python  code.<\/p>\n\n<p>Regards<\/p>\n","answers":[{"AnswerId":"53053455","CreationDate":"2018-10-29T20:39:15.807","ParentId":null,"OwnerUserId":"1757224","Title":null,"Body":"<p>You can find it with a little search \n<a href=\"https:\/\/github.com\/mmuratarat\/handson-ml\/blob\/master\/11_deep_learning.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mmuratarat\/handson-ml\/blob\/master\/11_deep_learning.ipynb<\/a><\/p>\n\n<pre><code>max_checks_without_progress = 20\nchecks_without_progress = 0\nbest_loss = np.infty\n\n....\n\n    if loss_val &lt; best_loss:\n        save_path = saver.save(sess, '.\/my_mnist_model.ckpt')\n        best_loss = loss_val\n        check_without_progress = 0\n    else:\n        check_without_progress +=1\n        if check_without_progress &gt; max_checks_without_progress:\n            print(\"Early stopping!\")\n            break\n\n    print(\"Epoch: {:d} - \".format(epoch), \\\n          \"Training Loss: {:.5f}, \".format(loss_train), \\\n          \"Training Accuracy: {:.2f}%, \".format(accuracy_train*100), \\\n          \"Validation Loss: {:.4f}, \".format(loss_val), \\\n          \"Best Loss: {:.4f}, \".format(best_loss), \\\n          \"Validation Accuracy: {:.2f}%\".format(accuracy_val*100))\n<\/code><\/pre>\n"},{"AnswerId":"52842679","CreationDate":"2018-10-16T19:23:08.247","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>You can do it Using <code>checkpoints<\/code>:<\/p>\n\n<pre><code>from keras.callbacks import EarlyStopping\nearlyStop=EarlyStopping(monitor=\"val_loss\",verbose=2,mode='min',patience=3)\nhistory=model.fit(xTrain,yTrain,epochs=100,batch_size=10,validation_data=(xTest,yTest) ,verbose=2,callbacks=[earlyStop])\n<\/code><\/pre>\n\n<p>Training will stop when \"val_loss\" has not decreased(<code>mode='min'<\/code>)  even after 3 epochs(<code>patience=3<\/code>)<\/p>\n\n<pre><code>#Didn't realize u were note using keras\n<\/code><\/pre>\n"}]}
+{"QuestionId":52822022,"AnswerCount":3,"Tags":"<python><tensorflow><sentiment-analysis><training-data><valueerror>","CreationDate":"2018-10-15T17:35:43.347","AcceptedAnswerId":52822130.0,"OwnerUserId":10018352.0,"Title":"TensorFlow ValueError: Cannot feed value of shape (32, 2) for Tensor 'InputData\/X:0', which has shape '(?, 100)'","Body":"<p>I am new to TensorFlow and machine learning. I'm trying to create a sentiment analysis NN with tensorflow.<\/p>\n\n<p>I've set up my architecture and I'm attempting to train the model but I encounter the error <\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (32, 2) for Tensor 'InputData\/X:0', which has shape '(?, 100)'<\/p>\n<\/blockquote>\n\n<p>I think the error has to do with my input \"layer net = tflearn.input_data([None, 100])\".\nThe tutorial I was following suggested this input shape, batch size as None and the length to be 100 since that's the sequence length. Hence (None, 100), to my understanding this is the dimensions the training data being fed into the network needs to be, correct?<\/p>\n\n<p>Could someone explain why the suggested input shape of batch size was None and also why Tensor flow is attempting to feed the network put shaped (32,2). Where is the sequence length of 2 coming from? <\/p>\n\n<p>If my understanding anywhere in this explanation is wrong feel free to correct me, I'm still trying to learn the theory as well.<\/p>\n\n<p>Thanks in advance<\/p>\n\n<pre><code>In [1]:\n\nimport tflearn\nfrom tflearn.data_utils import to_categorical, pad_sequences\nfrom tflearn.datasets import imdb\n\nIn [2]:\n\n#Loading IMDB dataset\ntrain, test, _ = imdb.load_data(path='imdb.pkl', n_words=10000,\n                                valid_portion=0.1)\ntrainX, trainY = train\ntestX, testY = test\n\nIn [3]:\n\n#Data sequence padding \ntrainX = pad_sequences(trainX, maxlen=100, value=0.)  \ntestX = pad_sequences(testX, maxlen=100, value=0.)\n#converting labels of each review to vectors\ntrainY = to_categorical(trainY, nb_classes=2)\ntrainX = to_categorical(testY, nb_classes=2)\n\n\nIn [4]:\n\n#network building \nnet = tflearn.input_data([None, 100])\nnet = tflearn.embedding(net, input_dim=10000, output_dim=128)\nnet = tflearn.lstm(net, 128, dropout = 0.8)\nnet = tflearn.fully_connected(net, 2, activation='softmax') \nnet = tflearn.regression(net, optimizer = 'adam', learning_rate=0.0001,\n                         loss='categorical_crossentropy')\n\n\nWARNING:tensorflow:From C:\\Users\\Nason\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tflearn\\objectives.py:66: calling reduce_sum (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.\nInstructions for updating:\nkeep_dims is deprecated, use keepdims instead\n\n\nIn [5]:\n\n#Training\nmodel = tflearn.DNN(net, tensorboard_verbose=0)   #train using tensorflow Deep nueral net\nmodel.fit(trainX, trainY, validation_set=(testX, testY), show_metric=True,    #fit launches training process for training and validation data, metric displays data as its training.\n          batch_size=32)\n\n\n---------------------------------\nRun id: U7NONK\nLog directory: \/tmp\/tflearn_logs\/\nINFO:tensorflow:Summary name Accuracy\/ (raw) is illegal; using Accuracy\/__raw_ instead.\n---------------------------------\nTraining samples: 2500\nValidation samples: 2500\n--\n\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-5-7ffd0a8836f9&gt; in &lt;module&gt;()\n      2 model = tflearn.DNN(net, tensorboard_verbose=0)   #train using tensorflow Deep nueral net\n      3 model.fit(trainX, trainY, validation_set=(testX, testY), show_metric=True,    #fit launches training process for training and validation data, metric displays data as its training.\n----&gt; 4           batch_size=32)\n\n~\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tflearn\\models\\dnn.py in fit(self, X_inputs, Y_targets, n_epoch, validation_set, show_metric, batch_size, shuffle, snapshot_epoch, snapshot_step, excl_trainops, validation_batch_size, run_id, callbacks)\n    214                          excl_trainops=excl_trainops,\n    215                          run_id=run_id,\n--&gt; 216                          callbacks=callbacks)\n    217 \n    218     def fit_batch(self, X_inputs, Y_targets):\n\n~\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tflearn\\helpers\\trainer.py in fit(self, feed_dicts, n_epoch, val_feed_dicts, show_metric, snapshot_step, snapshot_epoch, shuffle_all, dprep_dict, daug_dict, excl_trainops, run_id, callbacks)\n    337                                                        (bool(self.best_checkpoint_path) | snapshot_epoch),\n    338                                                        snapshot_step,\n--&gt; 339                                                        show_metric)\n    340 \n    341                             # Update training state\n\n~\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tflearn\\helpers\\trainer.py in _train(self, training_step, snapshot_epoch, snapshot_step, show_metric)\n    816         tflearn.is_training(True, session=self.session)\n    817         _, train_summ_str = self.session.run([self.train, self.summ_op],\n--&gt; 818                                              feed_batch)\n    819 \n    820         # Retrieve loss value from summary string\n\n~\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tensorflow\\python\\client\\session.py in run(self, fetches, feed_dict, options, run_metadata)\n    898     try:\n    899       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 900                          run_metadata_ptr)\n    901       if run_metadata:\n    902         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n~\\Anaconda33\\envs\\TensorFlow1.8CPU\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1109                              'which has shape %r' %\n   1110                              (np_val.shape, subfeed_t.name,\n-&gt; 1111                               str(subfeed_t.get_shape())))\n   1112           if not self.graph.is_feedable(subfeed_t):\n   1113             raise ValueError('Tensor %s may not be fed.' % subfeed_t)\n\nValueError: Cannot feed value of shape (32, 2) for Tensor 'InputData\/X:0', which has shape '(?, 100)'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52822402","CreationDate":"2018-10-15T18:04:03.600","ParentId":null,"OwnerUserId":"9178720","Title":null,"Body":"<p>You left the number of categories for <code>trainX<\/code> as <code>2<\/code>, but your model is expecting <code>100<\/code>. <\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I just notice that you are setting the <code>trainX<\/code> with the <code>testY<\/code> in this bit of code:<\/p>\n\n<pre><code>trainX = to_categorical(testY, nb_classes=2)\n<\/code><\/pre>\n\n<p>Whereas it should be:<\/p>\n\n<pre><code>trainX = to_categorical(trainX, nb_classes=100)\n<\/code><\/pre>\n\n<p>Therefore you need to change your code to:<\/p>\n\n<pre><code>#Data sequence padding\ntrainX = pad_sequences(trainX, maxlen=100, value=0.)  \ntestX = pad_sequences(testX, maxlen=100, value=0.)\n#converting labels of each review to vectors\ntrainY = to_categorical(trainY, nb_classes=2)\n#change the number of Classes\ntrainX = to_categorical(trainX, nb_classes=100) #CHANGE HERE!!\n<\/code><\/pre>\n\n<p>With this change you should be fine. I just tested and it works!<\/p>\n\n<p>It is ok to set the shape of the input with [None, 100] it gives you more flexibility to change the batch size later if you need!<\/p>\n"},{"AnswerId":"52822130","CreationDate":"2018-10-15T17:42:34.457","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>The error comes from <code>trainX = to_categorical(testY, nb_classes=2)<\/code>. This needs to be changed to <code>testY = to_categorical(testY, nb_classes=2)<\/code><\/p>\n\n<p>Also, setting the batch size to <code>None<\/code> means it should expect the batch to be any size. In your case you set the batch size to 32 so you could also set the input shape to <code>[32, 100]<\/code><\/p>\n"},{"AnswerId":"52822411","CreationDate":"2018-10-15T18:04:47.477","ParentId":null,"OwnerUserId":"10467184","Title":null,"Body":"<pre><code>tflearn.input_data([None, 100])\n<\/code><\/pre>\n\n<p>You are expecting the input to be a tensor of any number of instances with 100 features.<\/p>\n\n<pre><code>trainX = pad_sequences(trainX, maxlen=100, value=0.)  \ntestX = pad_sequences(testX, maxlen=100, value=0.)\n#converting labels of each review to vectors\ntrainY = to_categorical(trainY, nb_classes=2)\ntrainX = to_categorical(testY, nb_classes=2) #HEREEEEEE\n<\/code><\/pre>\n\n<p>This is problematic in your code. You are resetting trainX to have another shape instead of the padded one. I think you meant:<\/p>\n\n<pre><code>testY = to_categorical(testY, nb_classes=2)\n<\/code><\/pre>\n\n<p>If this still doesn't work.<\/p>\n\n<p>I suspect that you are missing a reshaping of the data. You are indeed using padding but on the whole trainX,trainY,etc. Try padding each \"row\" separately. Then each instance will have a length of '100' as  you are expecting. <\/p>\n\n<p>Before doing that, print the shapes of the tensors (like <code>print(trainX.shape)<\/code> ) to see if you are really preprocessing the data (I also suggest doing two scripts, one with whole loading, preprocessing, reshaping and padding and the other with tensorFlow logic)<\/p>\n"}]}
+{"QuestionId":52822382,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2018-10-15T18:03:15.850","AcceptedAnswerId":null,"OwnerUserId":2039927.0,"Title":"Unable to train in Google Cloud ML","Body":"<p>I'm not being able to train in ML Engine. The training always stop around iteration 60. I have used Keras to build the model layers, but I train using <code>tf.Session<\/code>.<\/p>\n\n<p>I get this error, but no traceback.<\/p>\n\n<pre><code>ERROR   2018-10-15 10:31:02 -0700   master-replica-0        name: Tesla P100-PCIE-16GB major: 6 minor: 0 memoryClockRate(GHz): 1.3285\nERROR   2018-10-15 10:31:02 -0700   master-replica-0        pciBusID: 0000:00:04.0\nERROR   2018-10-15 10:31:02 -0700   master-replica-0        totalMemory: 15.90GiB freeMemory: 15.61GiB\n<\/code><\/pre>\n\n<p>My config.yaml. I tried different configurations. Same result.<\/p>\n\n<pre><code>trainingInput:\n  scaleTier: CUSTOM\n  masterType: standard_p100\n<\/code><\/pre>\n\n<p>Job submission.<\/p>\n\n<pre><code>gcloud ml-engine jobs submit training $JOB_NAME --labels=\"$LABELS\" --verbosity='debug' --stream-logs --package-path=.\/job --module-name=job.task --staging-bucket=\"$TRAIN_BUCKET\" --region=us-central1 --runtime-version 1.10 --config=job\/config.yaml\n<\/code><\/pre>\n\n<p>Full logs<\/p>\n\n<pre><code>INFO    2018-10-15 10:28:37 -0700   service     Validating job requirements...\nINFO    2018-10-15 10:28:38 -0700   service     Job creation request has been successfully validated.\nINFO    2018-10-15 10:28:38 -0700   service     Job &lt;JOB_NAME&gt; is queued.\nINFO    2018-10-15 10:28:38 -0700   service     Waiting for job to be provisioned.\nINFO    2018-10-15 10:28:41 -0700   service     Waiting for training program to start.\nINFO    2018-10-15 10:30:03 -0700   master-replica-0        Running task with arguments: --cluster={\"master\": [\"127.0.0.1:2222\"]} --task={\"type\": \"master\", \"index\": 0} --job={  \"scale_tier\": \"CUSTOM\",  \"master_type\": \"standard_p100\",  \"package_uris\": [\"gs:\/\/annotator-1286-ml\/&lt;JOB_NAME&gt;\/5b038627d10c914d6309269cefff8d2e0682f87f441bdb8c547a05e8ed1107a7\/job-0.0.0.tar.gz\"],  \"python_module\": \"job.task\",  \"region\": \"us-central1\",  \"runtime_version\": \"1.10\",  \"run_on_raw_vm\": true}\nINFO    2018-10-15 10:30:15 -0700   master-replica-0        Running module job.task.\nINFO    2018-10-15 10:30:15 -0700   master-replica-0        Downloading the package: gs:\/\/annotator-1286-ml\/&lt;JOB_NAME&gt;\/5b038627d10c914d6309269cefff8d2e0682f87f441bdb8c547a05e8ed1107a7\/job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:15 -0700   master-replica-0        Running command: gsutil -q cp gs:\/\/annotator-1286-ml\/&lt;JOB_NAME&gt;\/5b038627d10c914d6309269cefff8d2e0682f87f441bdb8c547a05e8ed1107a7\/job-0.0.0.tar.gz job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:22 -0700   master-replica-0        Installing the package: gs:\/\/annotator-1286-ml\/&lt;JOB_NAME&gt;\/5b038627d10c914d6309269cefff8d2e0682f87f441bdb8c547a05e8ed1107a7\/job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:22 -0700   master-replica-0        Running command: pip install --user --upgrade --force-reinstall --no-deps job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:28 -0700   master-replica-0        Processing .\/job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:29 -0700   master-replica-0        Building wheels for collected packages: job\nINFO    2018-10-15 10:30:29 -0700   master-replica-0          Running setup.py bdist_wheel for job: started\nINFO    2018-10-15 10:30:29 -0700   master-replica-0          Running setup.py bdist_wheel for job: finished with status 'done'\nINFO    2018-10-15 10:30:29 -0700   master-replica-0          Stored in directory: \/root\/.cache\/pip\/wheels\/b8\/10\/df\/bb59eda2baac79b36fbdb8e5305ada7d6bf7779be49c3c5a0d\nINFO    2018-10-15 10:30:29 -0700   master-replica-0        Successfully built job\nINFO    2018-10-15 10:30:29 -0700   master-replica-0        Installing collected packages: job\nINFO    2018-10-15 10:30:29 -0700   master-replica-0        Successfully installed job-0.0.0\nINFO    2018-10-15 10:30:30 -0700   master-replica-0        Running command: pip install --user job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:30 -0700   master-replica-0        Processing .\/job-0.0.0.tar.gz\nINFO    2018-10-15 10:30:30 -0700   master-replica-0        Building wheels for collected packages: job\nINFO    2018-10-15 10:30:30 -0700   master-replica-0          Running setup.py bdist_wheel for job: started\nINFO    2018-10-15 10:30:30 -0700   master-replica-0          Running setup.py bdist_wheel for job: finished with status 'done'\nINFO    2018-10-15 10:30:30 -0700   master-replica-0          Stored in directory: \/root\/.cache\/pip\/wheels\/b8\/10\/df\/bb59eda2baac79b36fbdb8e5305ada7d6bf7779be49c3c5a0d\nINFO    2018-10-15 10:30:30 -0700   master-replica-0        Successfully built job\nINFO    2018-10-15 10:30:31 -0700   master-replica-0        Installing collected packages: job\nINFO    2018-10-15 10:30:31 -0700   master-replica-0          Found existing installation: job 0.0.0\nINFO    2018-10-15 10:30:31 -0700   master-replica-0            Uninstalling job-0.0.0:\nINFO    2018-10-15 10:30:31 -0700   master-replica-0              Successfully uninstalled job-0.0.0\nINFO    2018-10-15 10:30:31 -0700   master-replica-0        Successfully installed job-0.0.0\nINFO    2018-10-15 10:30:31 -0700   master-replica-0        Running command: python -m job.task\nINFO    2018-10-15 10:31:02 -0700   master-replica-0        successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\nINFO    2018-10-15 10:31:02 -0700   master-replica-0        Found device 0 with properties: \nERROR   2018-10-15 10:31:02 -0700   master-replica-0        name: Tesla P100-PCIE-16GB major: 6 minor: 0 memoryClockRate(GHz): 1.3285\nERROR   2018-10-15 10:31:02 -0700   master-replica-0        pciBusID: 0000:00:04.0\nERROR   2018-10-15 10:31:02 -0700   master-replica-0        totalMemory: 15.90GiB freeMemory: 15.61GiB\nINFO    2018-10-15 10:31:02 -0700   master-replica-0        Adding visible gpu devices: 0\nINFO    2018-10-15 10:31:03 -0700   master-replica-0        Device interconnect StreamExecutor with strength 1 edge matrix:\nINFO    2018-10-15 10:31:03 -0700   master-replica-0             0 \nINFO    2018-10-15 10:31:03 -0700   master-replica-0        0:   N \nINFO    2018-10-15 10:31:03 -0700   master-replica-0        Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 15127 MB memory) -&gt; physical GPU (device: 0, name: Tesla P100-PCIE-16GB, pci bus id: 0000:00:04.0, compute capability: 6.0)\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        Mon Oct 15 17:32:06 2018       \nINFO    2018-10-15 10:32:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        | NVIDIA-SMI 396.26                 Driver Version: 396.26                    |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        |-------------------------------+----------------------+----------------------+\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        | Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        |===============================+======================+======================|\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        |   0  Tesla P100-PCIE...  Off  | 00000000:00:04.0 Off |                    0 |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        | N\/A   46C    P0   172W \/ 250W |  15619MiB \/ 16280MiB |     82%      Default |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        +-------------------------------+----------------------+----------------------+\nINFO    2018-10-15 10:32:06 -0700   master-replica-0                                                                                       \nINFO    2018-10-15 10:32:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        | Processes:                                                       GPU Memory |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        |  GPU       PID   Type   Process name                             Usage      |\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        |=============================================================================|\nINFO    2018-10-15 10:32:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        Mon Oct 15 17:37:06 2018       \nINFO    2018-10-15 10:37:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        | NVIDIA-SMI 396.26                 Driver Version: 396.26                    |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        |-------------------------------+----------------------+----------------------+\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        | Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        |===============================+======================+======================|\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        |   0  Tesla P100-PCIE...  Off  | 00000000:00:04.0 Off |                    0 |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        | N\/A   52C    P0    39W \/ 250W |  15619MiB \/ 16280MiB |     33%      Default |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        +-------------------------------+----------------------+----------------------+\nINFO    2018-10-15 10:37:06 -0700   master-replica-0                                                                                       \nINFO    2018-10-15 10:37:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        | Processes:                                                       GPU Memory |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        |  GPU       PID   Type   Process name                             Usage      |\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        |=============================================================================|\nINFO    2018-10-15 10:37:06 -0700   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Setting Parameters\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Create data source\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Parse tfrecords\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Apply transformations\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Parametrize Dataset\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Build Model\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Define input sizes to Keras tensors and assign image tensor\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: Resnet\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"pool1\/MaxPool:0\", shape=(?, 99, 99, 64), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res2a_relu\/Relu:0\", shape=(?, 99, 99, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res2b_relu\/Relu:0\", shape=(?, 99, 99, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res2c_relu\/Relu:0\", shape=(?, 99, 99, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res3a_relu\/Relu:0\", shape=(?, 50, 50, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res3b1_relu\/Relu:0\", shape=(?, 50, 50, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res3b2_relu\/Relu:0\", shape=(?, 50, 50, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4a_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b1_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b2_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b3_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b4_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b5_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b6_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b7_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b8_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b9_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b10_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b11_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b12_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b13_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b14_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b15_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b16_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b17_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b18_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b19_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b20_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b21_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res4b22_relu\/Relu:0\", shape=(?, 25, 25, 1024), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res5a_relu\/Relu:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res5b_relu\/Relu:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"res5c_relu\/Relu:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        get_personlab: PersonLab Head\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        build_personlab_head: Add kp_maps\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"kp_maps\/Sigmoid:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        build_personlab_head: Add short_offsets\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"short_offsets\/BiasAdd:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        build_personlab_head: Add mid_offsets\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"mid_offsets\/BiasAdd:0\", shape=(?, 25, 25, 2048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"kp_maps_tConv1\/BiasAdd:0\", shape=(?, 50, 50, 1048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"kp_maps_tConv2\/BiasAdd:0\", shape=(?, 100, 100, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"kp_maps_tConv3\/BiasAdd:0\", shape=(?, 200, 200, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"kp_maps_tConv3_1\/BiasAdd:0\", shape=(?, 400, 400, 17), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"short_offsets_tConv1\/BiasAdd:0\", shape=(?, 50, 50, 1048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"short_offsets_tConv2\/BiasAdd:0\", shape=(?, 100, 100, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"short_offsets_tConv3\/BiasAdd:0\", shape=(?, 200, 200, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"short_offsets_tConv3_1\/BiasAdd:0\", shape=(?, 400, 400, 34), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"mid_offsets_tConv1\/BiasAdd:0\", shape=(?, 50, 50, 1048), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"mid_offsets_tConv2\/BiasAdd:0\", shape=(?, 100, 100, 512), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"mid_offsets_tConv3\/BiasAdd:0\", shape=(?, 200, 200, 256), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Tensor(\"mid_offsets_tConv3_1\/BiasAdd:0\", shape=(?, 400, 400, 64), dtype=float32)\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Add loss and training operations\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Create Saver object\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Initialize variables\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        Training\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '0')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '10')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '20')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '30')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '40')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '50')\nINFO    2018-10-15 10:38:36 -0700   master-replica-0        ('iteration: ', '60')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/440Si.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/440Si.png\" alt=\"CPU and Memory\"><\/a><\/p>\n\n<p>How can I get more detailed information about the error?<\/p>\n\n<p>Update: I'm fetching data from Google Cloud Storage. I followed <code>https:\/\/www.tensorflow.org\/performance\/datasets_performance<\/code>.<\/p>\n","answers":[{"AnswerId":"52896415","CreationDate":"2018-10-19T16:24:56.153","ParentId":null,"OwnerUserId":"2039927","Title":null,"Body":"<p>I figure out it was very very slow. So it gave me the impression the problem was on the training. I presumed was a problem with distributed training because Keras layers.<\/p>\n\n<p>I changed the config to use <code>complex_model_l_gpu<\/code> and it worked.<\/p>\n"}]}
+{"QuestionId":52822734,"AnswerCount":1,"Tags":"<python><keras><autoencoder>","CreationDate":"2018-10-15T18:27:39.993","AcceptedAnswerId":null,"OwnerUserId":4416395.0,"Title":"merging 2 CNNs into an autoencoder in Keras","Body":"<p>I am trying to merge 2 CNNs and then un-merge them through an autoencoder but I always get the same error message saying \"ValueError: total size of new array must be unchanged\", not sure what is wrong with my network.<\/p>\n\n<p>I am replacing the LSTM and CNN by Dense layers and it works.<\/p>\n\n<p>Thanks for helping!<\/p>\n\n<pre><code>from keras.layers import Input, Dense, Conv1D, MaxPooling1D, UpSampling1D, Embedding, Dropout, Flatten, Concatenate, LSTM, BatchNormalization, Reshape\nfrom keras.models import Model\nfrom keras import backend as K\nfrom keras.callbacks import TensorBoard\n\n\ninput_address = Input(shape=(38,))\nx_address = Reshape((38, 1))(input_address)\nx_address = LSTM(125, activation=\"tanh\", return_sequences=True)(x_address)\nx_address = Dropout(0.4)(x_address)\nx_address = LSTM(125, activation=\"tanh\", return_sequences=True)(x_address)\nx_address = Dropout(0.4)(x_address)\nx_address = Dense(100, activation = 'relu')(x_address)\nx_address = Conv1D(100, 2, activation='relu', padding='same')(x_address)\nx_address = MaxPooling1D(2, padding='same')(x_address)\nx_address = Dropout(0.5)(x_address)\nx_address = Flatten()(x_address)\n\ninput_name = Input(shape=(29,))\nx_name = Reshape((29, 1))(input_name)\nx_name = LSTM(125, activation=\"tanh\", return_sequences=True)(x_name)\nx_name = Dropout(0.4)(x_name)\nx_name = LSTM(125, activation=\"tanh\", return_sequences=True)(x_name)\nx_name = Dropout(0.4)(x_name)\nx_name = Dense(100, activation = 'relu')(x_name)\nx_name = Conv1D(100, 2, activation='relu', padding='same')(x_name)\nx_name = MaxPooling1D(2, padding='same')(x_name)\nx_name = Dropout(0.5)(x_name)\nx_name = Flatten()(x_name)\n\nmerge = Concatenate(name=\"concat\", axis=1)([x_name, x_address])\nencoder = Dense(1, activation = 'relu')(merge)\n\n\nxd_address = Reshape((19, 100))(encoder)\nxd_address = UpSampling1D(2)(xd_address)\nxd_address = Conv1D(100, 2, activation='relu', padding='same')(xd_address)\nxd_address = Dropout(0.4)(xd_address)\nxd_address = LSTM(125, activation=\"tanh\", return_sequences=True)(xd_address)\nxd_address = Dropout(0.4)(xd_address)\nxd_address = LSTM(125, activation=\"tanh\", return_sequences=True)(xd_address)\nxd_address = Flatten()(xd_address)\n\nxd_name = Reshape((15, 100))(encoder)\nxd_name = UpSampling1D(2)(xd_name)\nxd_name = Conv1D(100, 2, activation='relu', padding='same')(xd_name)\nxd_name = Dropout(0.4)(xd_name)\nxd_name = LSTM(125, activation=\"tanh\", return_sequences=True)(xd_name)\nxd_name = Dropout(0.4)(xd_name)\nxd_name = LSTM(125, activation=\"tanh\", return_sequences=True)(xd_name)\nxd_name = Flatten()(xd_name)\n\nautoencoder = Model(inputs=[input_name, input_address], outputs=[xd_name, xd_address])\nautoencoder.compile(optimizer='adam', loss=['mse', 'mse'])\n\nautoencoder.summary()\n\nautoencoder.fit([name_train, address_train], [name_train, address_train]\n                 ,epochs=10\n                 ,batch_size=120\n                 ,shuffle=True\n                 ,validation_data=([name_test, address_test], [name_test, address_test])\n                 ,callbacks=[TensorBoard(log_dir='\/tmp\/autoencoder')])\n<\/code><\/pre>\n\n<p>Here is the error message I get<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-132-19e453d673b6&gt; in &lt;module&gt;()\n     33 \n     34 \n---&gt; 35 xd_address = Reshape((19, 100))(encoder)\n     36 xd_address = UpSampling1D(2)(xd_address)\n     37 xd_address = Conv1D(100, 2, activation='relu', padding='same')(xd_address)\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\base_layer.py in __call__(self, inputs, **kwargs)\n    472             if all([s is not None\n    473                     for s in to_list(input_shape)]):\n--&gt; 474                 output_shape = self.compute_output_shape(input_shape)\n    475             else:\n    476                 if isinstance(input_shape, list):\n\n~\\Anaconda3\\lib\\site-packages\\keras\\layers\\core.py in compute_output_shape(self, input_shape)\n    392             # input shape known? then we can compute the output shape\n    393             return (input_shape[0],) + self._fix_unknown_dimension(\n--&gt; 394                 input_shape[1:], self.target_shape)\n    395 \n    396     def call(self, inputs):\n\n~\\Anaconda3\\lib\\site-packages\\keras\\layers\\core.py in _fix_unknown_dimension(self, input_shape, output_shape)\n    380             output_shape[unknown] = original \/\/ known\n    381         elif original != known:\n--&gt; 382             raise ValueError(msg)\n    383 \n    384         return tuple(output_shape)\n\nValueError: total size of new array must be unchanged\n<\/code><\/pre>\n","answers":[{"AnswerId":"52823109","CreationDate":"2018-10-15T18:55:20.903","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You are reshaping into a size that changes the number elements in the original <code>encoder<\/code> tensor. You need to:<\/p>\n\n<ol>\n<li><code>print(encoder)<\/code> and that will give you a hint on what the original shape is.<\/li>\n<li>Make sure to <code>Reshape<\/code> while preserving the number of elements.<\/li>\n<\/ol>\n\n<p>Here is a short example, <code>(100,)<\/code> cannot be reshaped to <code>(2,45)<\/code> because you omitted 10 elements, i.e. \"total size of new array must be unchanged\".<\/p>\n"}]}
+{"QuestionId":52823162,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-15T18:59:19.937","AcceptedAnswerId":null,"OwnerUserId":7259313.0,"Title":"How to implement a message propagation block for label-relation (Keras)","Body":"<p>I am trying to build a network architecture to classify fashion images I found in this paper <a href=\"https:\/\/arxiv.org\/pdf\/1806.09445.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1806.09445.pdf<\/a> using keras + tensorflow (like the authors in the paper).<\/p>\n\n<p>The paper proposes an interesting method by which not only the visual information of the images is used for classification, but also the relationships among labels are utilized. This is done using a so called \"message propagation block\". Since the authors are unfortunately unable to provide the code, I am trying to build it myself. However, I find myself having trouble implementing the message propagation block of the model.<\/p>\n\n<p>The architecture of this message propagation block is outlined fairly detailed in the paper (Figure 6 + 7).<\/p>\n\n<p>As such, I am trying to use the keras functional API to replicate this model following all the details highlighted by the authors. The paper says the model should have 46.915.690 trainable parameters, however my model has only 39,981,162 parameters.<\/p>\n\n<p>Now I am trying to find my mistake. Since I am fairly new to the usage of the functional API, I wonder if my model implementation is technically correct or if I am missing something obvious (i.e. did some serious mistakes).<\/p>\n\n<p>This is my entire code including the message propagation block (see comment in the code):<\/p>\n\n<pre><code>img_width, img_height = 224, 224\nimg_input = Input(shape=(img_width, img_height, 3))\n\nresnet50 = keras.applications.resnet50.ResNet50(include_top=False, weights='imagenet', input_tensor=img_input, pooling=None)\n\nx = resnet50.output\n\nxcat_input = Dense(1024, activation='relu', name=\"xcat_input\")(x)\nxsubcat_input = Dense(1024, activation='relu', name=\"xsubcat_input\")(x)\nxattribute_input = Dense(1024, activation='relu', name=\"xattribute_input\")(x)\n\nreg = regularizers.l2(0.0005 )\n\n\n#### message propagation block starts here (Figure 6+7 in the paper)\n\n#################\n# Top-down pass #\n#################\n\n# Category-Block\ndense_category_output_blue = Dense(1024, activation='relu',kernel_regularizer=reg)(xcat_input)\n\n# Subcategory-Block\na_minusone = Dense(1024, activation='relu',kernel_regularizer=reg)(dense_category_output_blue)\na = Dense(1024, activation='relu',kernel_regularizer=reg)(xsubcat_input)\ndense_subcategory_output_blue = keras.layers.concatenate([a_minusone, a])\n\n# Attribute-Block\na_minusone = Dense(1024, activation='relu',kernel_regularizer=reg)(dense_category_output_blue)\na = Dense(1024, activation='relu',kernel_regularizer=reg)(xattribute_input)\ndense_attribute_output_blue = keras.layers.concatenate([a_minusone, a])\n\n\n\n##################\n# Bottom-up pass #\n##################\n\n# Attribute-Block\ndense_attribute_output_green = Dense(1024, activation='relu',kernel_regularizer=reg)(xattribute_input)\n\n# Subcategory-Block\ndense_subcategory_output_green = Dense(1024, activation='relu',kernel_regularizer=reg)(xsubcat_input)\n\n# Category-Block\ndense_category_up = Dense(1024, activation='relu',kernel_regularizer=reg)(xcat_input)\nx_category_output_green = Dense(1024, activation='relu',kernel_regularizer=reg)(keras.layers.concatenate([dense_attribute_output_green, dense_subcategory_output_green]))\n\n#### merging the result of each direction from the first part of the message passing\nadd1 = keras.layers.concatenate([dense_category_output_blue, x_category_output_green])\nadd1 = keras.layers.Dropout(0.3, noise_shape=None, seed=None)(add1)\ndense_category_output = Dense(64, activation='softmax', name='dense_category_output')(add1)\n\nadd2 = keras.layers.concatenate([dense_subcategory_output_blue, dense_subcategory_output_green])\nadd2 = keras.layers.Dropout(0.3, noise_shape=None, seed=None)(add2)\ndense_subcategory_output = Dense(95, activation='softmax', name='dense_subcategory_output')(add2)\n\nadd3 = keras.layers.concatenate([dense_attribute_output_blue, dense_attribute_output_green])\nadd3 = keras.layers.Dropout(0.3, noise_shape=None, seed=None)(add3)\ndense_attribute_output = Dense(75, activation='sigmoid', name='dense_attribute_output')(add3)\n\n\nmodel = Model(inputs=img_input, outputs=[dense_category_output, dense_subcategory_output, dense_attribute_output])\n\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy')\n<\/code><\/pre>\n\n<p>Thank you very much in advance!<\/p>\n","answers":[]}
+{"QuestionId":52823569,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-15T19:31:56.400","AcceptedAnswerId":52824408.0,"OwnerUserId":3926152.0,"Title":"How to provide the gradient function GradientDescentOptimizer","Body":"<p>I have seen several example using GradientDescentOptimizer. \nHere is a good one<\/p>\n\n<p><a href=\"https:\/\/learningtensorflow.com\/lesson7\/\" rel=\"nofollow noreferrer\">Tensorflow GradientDescentOptimizer example<\/a><\/p>\n\n<p>The model in the above example is :<\/p>\n\n<pre><code>y_model = tf.multiply(x, w[0]) + w[1]\n<\/code><\/pre>\n\n<p>Do we provide tensorflow the gradient function? How does it know how to calculate the gradient?<\/p>\n","answers":[{"AnswerId":"52824408","CreationDate":"2018-10-15T20:35:28.480","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>No, we do not provide gradient function. This is the beauty of automatic differentiation that TF and many other DL frameworks use.<\/p>\n\n<p>Knowing the function that you need to optimize, the differentiator uses the rules for taking the derivative of each element of a program (when you define any op in core TF, <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op#implement_the_gradient_in_python\" rel=\"nofollow noreferrer\">you need to register a gradient for this op<\/a>). It also uses <a href=\"https:\/\/www.khanacademy.org\/math\/ap-calculus-ab\/ab-differentiation-2-new\/ab-3-1a\/v\/chain-rule-introduction\" rel=\"nofollow noreferrer\">chain rule<\/a> to break complex expressions into simpler ones.<\/p>\n\n<p>For more information check these answers:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/a\/45133884\/1090562\">automatic differentiation<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/a\/44342968\/1090562\">example of auto-diff<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":52823862,"AnswerCount":2,"Tags":"<python><neural-network><keras><sequential>","CreationDate":"2018-10-15T19:54:59.510","AcceptedAnswerId":52823939.0,"OwnerUserId":10397367.0,"Title":"keras sequential().predict(x_test) only returns 1 column for two classes","Body":"<p>I'm having problem with keras <code>sequential().predict(x_test)<\/code>.<\/p>\n\n<p>Btw getting the same output using <code>sequential().predict_proba(x_test)<\/code> as I found that these two are indifferent in sequential now.<\/p>\n\n<p>My data has two classes: 0 or 1, I believe <code>predict(x_test)<\/code> should give two columns, where the first column is the prob for getting 0 and the second is prob of getting 1. However I only have one column with this.<\/p>\n\n<pre><code>    In [85]:y_train.value_counts()\n    Out[85]: \n    0    616751\n    1     11140\n    Name: _merge, dtype: int64\n<\/code><\/pre>\n\n<p>There should be no problem with my data as I used the same x_train, y_train, x_test, y_test for both LogisticRegression model and neural network model, it works perfect in LogisticRegression.<\/p>\n\n<pre><code>In [87]:y_pred_LR\nOut[87]: \narray([[  9.96117151e-01,   3.88284921e-03],\n       [  9.99767583e-01,   2.32417329e-04],\n       [  9.87375774e-01,   1.26242258e-02],\n       ..., \n       [  9.72159138e-01,   2.78408623e-02],\n       [  9.97232916e-01,   2.76708432e-03],\n       [  9.98146985e-01,   1.85301489e-03]])\n<\/code><\/pre>\n\n<p>but I only get 1 column in neural network model.<\/p>\n\n<p>So I guess there is some problem with the NN model setting up? Here is my codes<\/p>\n\n<pre><code>NN = Sequential()\nNN.add(Dense(40, input_dim = 65, kernel_initializer = 'uniform', activation = 'relu'))\nNN.add(Dense(20, kernel_initializer = 'uniform', activation = 'relu'))\nNN.add(Dense(1, kernel_initializer = 'uniform', activation = 'sigmoid'))\nNN.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\nNN.fit(x_train, y_train, batch_size = 50, epochs=5)\ny_pred_NN = NN.predict(x_test)\nprint(y_pred_NN)\n\n    In [86]: print(y_pred_NN)\n    [[ 0.00157279]\n     [ 0.0010451 ]\n     [ 0.03178826]\n     ..., \n     [ 0.01030775]\n     [ 0.00584918]\n     [ 0.00186538]]\n<\/code><\/pre>\n\n<p>Actually it looks like it's the prob of getting 1? \nAny help is appreciated!<\/p>\n\n<p>Btw the shapes of my predictions in both models are as follows<\/p>\n\n<pre><code>In [91]:y_pred_LR.shape\nOut[91]: (300000, 2)\n\nIn [90]:y_pred_NN.shape\nOut[90]: (300000, 1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52823939","CreationDate":"2018-10-15T20:00:42.713","ParentId":null,"OwnerUserId":"7812314","Title":null,"Body":"<p>If you want to output two probabilities, you will have to replace your <code>y_train<\/code> with <code>to_categorical(y_train)<\/code> and then adjust the network accordingly:<\/p>\n\n<pre><code>from keras.utils import to_categorical\n\nNN = Sequential()\nNN.add(Dense(40, input_dim = 65, kernel_initializer = 'uniform', activation = 'relu'))\nNN.add(Dense(20, kernel_initializer = 'uniform', activation = 'relu'))\nNN.add(Dense(2, activation='sigmoid'))\nNN.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\nNN.fit(x_train, to_categorical(y_train), batch_size = 50, epochs=5)\n<\/code><\/pre>\n\n<p>Consult here: <a href=\"https:\/\/keras.io\/utils\/#to_categorical\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/utils\/#to_categorical<\/a><\/p>\n"},{"AnswerId":"52824072","CreationDate":"2018-10-15T20:10:00.950","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>The last line of your model defines the output layer.<\/p>\n\n<pre><code>NN.add(Dense(1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n<\/code><\/pre>\n\n<p>Your last line has one node, with a sigmoid activation. This means that your output will be a single number (per input sample) between 0 and 1, which I believe you can interpret as P(y=1), as per logistic regression.<\/p>\n\n<p>With more than two classes you would have <code>n_classes<\/code> in the output layer:<\/p>\n\n<pre><code>NN.add(Dense(n_classes, kernel_initializer = 'uniform', activation = 'softmax'))\n<\/code><\/pre>\n\n<p>The softmax ensures that the probabilities sum to unity. Of course to do this you would have to one-hot encode your training y values.<\/p>\n\n<p>In your case you can choose between either approach. I favour the second, as it allows you to add more classes in the future.<\/p>\n"}]}
+{"QuestionId":52824051,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-15T20:08:47.937","AcceptedAnswerId":52829621.0,"OwnerUserId":10262090.0,"Title":"Using pre-trained object detection models in tensorflow. Which parameters are used?","Body":"<p>I was wondering how pre-trained models are used in tensorflows object detection repo. I'm not sure if I did understand it correctly.<\/p>\n\n<p>It is kind of an expansion of the question asked in: <a href=\"https:\/\/stackoverflow.com\/questions\/49374984\/tensorflow-object-detection-api-classification-weights-initialization-when-chan\">link<\/a><\/p>\n\n<p><strong>My thoughts<\/strong>:<br>\nTensorflow builds a model based on the config file. Now it takes a look at the pre-trained model and iterates over every layer. If a layer matches between the two by name the parameters\/weights are retained and not retained if the layer does not match. <\/p>\n\n<p>Lets say I want to use the pre-trained model for ssd mobilenet v1 coco. My new model has 100 classes. What tf does is, it initializes the convolutional layer corresponding to the first 90 classes of my new model with the old values while choosing random values for the others at the start of the training. Now if my new model has only 10 classes, those are initialized with the first 10 classes of the pre-trained coco-model.<\/p>\n\n<p><strong>OR<\/strong><\/p>\n\n<p>The specific convolutional layer which corresponds to the classes does not match the shape of the pre-trained model and is thus not retained? Because in SSD the convolutions to determine the scores\/classes\/box coordinate values are contained in multiple convolutional layers with multiple filters each corresponding to a different value. <\/p>\n","answers":[{"AnswerId":"52829621","CreationDate":"2018-10-16T07:00:09.997","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>As you said yourself, if the layer isn't the same - it isn't loaded from the checkpoint\/pre-trained model, and is initialized.\nIf you change the number of classes, then you change the corresponding layer, whether it is fully connected (e.g. number of output neurons) or convolution (number of channels).\nTherefore, if you change the number of classes - all affected layers are initialized.<\/p>\n"}]}
+{"QuestionId":52824127,"AnswerCount":0,"Tags":"<tensorflow><64-bit><python-3.6><opc>","CreationDate":"2018-10-15T20:15:16.830","AcceptedAnswerId":null,"OwnerUserId":4802101.0,"Title":"Python 3.6 x64 OpenOPC","Body":"<p>I am developing a tensorflow model to make a inference of a product property. I want to put this model to run on-line reading data from a OPC Server a writing the result back. To do that I am using the OpenOPC project.<\/p>\n\n<p>Tensorflow just works with Python 3.6 in a 64 bit OS. So I need to make a OPC connection in the same environment.<\/p>\n\n<p>OpenOPC was originally desinged to work with Python 2.7 in a x86 OS. There are some projects that upgraded it to work with Python 3.4+. I followed the procedure in the following link to install it:\n<a href=\"https:\/\/github.com\/joseamaita\/openopc120\" rel=\"nofollow noreferrer\">https:\/\/github.com\/joseamaita\/openopc120<\/a><\/p>\n\n<p>I am developing my model in a VM with Windows 2012 Server. For OPC Server I am using Matrikon OPC for Simulation.<\/p>\n\n<p>The list function from OpenOPC seems to be working fine. When I run the following script<\/p>\n\n<pre><code>import OpenOPC\nopen_host='10.0.2.15'\nopen_port='7766'\n\nopc = OpenOPC.open_client(open_host, open_port)\nopc.connect('Matrikon.OPC.Simulation.1')\n\naux=opc.list('Simulation Items.Random')\nfor item in aux:\n    print(item)\n\nopc.close()\n<\/code><\/pre>\n\n<p>I got the following result<\/p>\n\n<pre><code>Random.ArrayOfReal8\nRandom.ArrayOfString\nRandom.Boolean\nRandom.Int1\nRandom.Int2\nRandom.Int4\nRandom.Money\nRandom.Qualities\nRandom.Real4\nRandom.Real8\nRandom.String\nRandom.Time\nRandom.UInt1\nRandom.UInt2\nRandom.UInt4\n<\/code><\/pre>\n\n<p>Which is promising. But when I tried to run the following code to read a value:<\/p>\n\n<pre><code>import OpenOPC\nopen_host='10.0.2.15'\nopen_port='7766'\n\nopc = OpenOPC.open_client(open_host, open_port)\nopc.connect('Matrikon.OPC.Simulation.1')\n\naux = opc.read(['Random.Int4'])\nfor item in aux:\n    print(item)\n\nopc.close()\n<\/code><\/pre>\n\n<p>the OpenOPC gateway service stops and in the event viewer I have the following<\/p>\n\n<pre><code>Faulting application name: PythonService.exe, version: 0.0.0.0, time stamp: 0x5af24199\nFaulting module name: gbda_aut.dll, version: 1.0.8.304, time stamp: 0x4d92bcca\nException code: 0xc0000005\nFault offset: 0x000000000001b966\nFaulting process id: 0xa90\nFaulting application start time: 0x01d464c2264cd7ec\nFaulting application path: C:\\ProgramData\\Anaconda3\\lib\\site-packages\\win32\\PythonService.exe\nFaulting module path: c:\\Windows\\System32\\gbda_aut.dll\nReport Id: 69629dee-d0b5-11e8-8120-080027bc5cb4\nFaulting package full name: \nFaulting package-relative application ID: ` .\n<\/code><\/pre>\n\n<p>It seems that the problem is related to this gbda_aut.dll. I have download the x64 version of it from <\/p>\n\n<p><a href=\"http:\/\/gray-box.net\/download_daawrapper.php?lang=en\" rel=\"nofollow noreferrer\">http:\/\/gray-box.net\/download_daawrapper.php?lang=en<\/a><\/p>\n\n<p>and registered it in the appropiated way.<\/p>\n\n<p>The problem also happen when I try to use the DCOM mode. When doing that I receive the following.<\/p>\n\n<pre><code>Problem signature:\nProblem Event Name: APPCRASH\nApplication Name:   pythonw.exe\nApplication Version:    3.6.5150.1013\nApplication Timestamp:  5abd3212\nFault Module Name:  gbda_aut.dll\nFault Module Version:   1.0.8.304\nFault Module Timestamp: 4d92bcca\nException Code: c0000005\nException Offset:   000000000001b966\nOS Version: 6.3.9600.2.0.0.272.7\nLocale ID:  1033\nAdditional Information 1:   e0f5\nAdditional Information 2:   e0f51433c52d71c63f2bde60bf8cf401\nAdditional Information 3:   1c6f\nAdditional Information 4:   1c6f1bc49be4e21c20f4309779c02df6\n<\/code><\/pre>\n\n<p>Does anyone have a idea how to make it work or a better way to do a OPC read using Python?<\/p>\n","answers":[]}
+{"QuestionId":52824373,"AnswerCount":2,"Tags":"<python><variables><tensorflow><slice><assign>","CreationDate":"2018-10-15T20:33:24.310","AcceptedAnswerId":null,"OwnerUserId":5112616.0,"Title":"Slice assignment to tensorflow variable","Body":"<p>I am trying to assign values to slice of a variable in tensorflow, but the following error is shown: 'ValueError: Sliced assignment is only supported for variables'. Why is this error showing even though I am trying to do slice assignment to a variable. My code is something like this:<\/p>\n\n<pre><code>var1 = var1[startx:endx,starty:endy].assign(tf.ones([endx-startx,endy-starty],dtype=tf.int32))\n<\/code><\/pre>\n\n<p>where <code>var1<\/code> is a tensorflow variable.<\/p>\n","answers":[{"AnswerId":"52824843","CreationDate":"2018-10-15T21:12:57.340","ParentId":null,"OwnerUserId":"2426955","Title":null,"Body":"<p>Once <code>var1<\/code> is sliced it is no longer a variable.  <\/p>\n\n<p>The numpy style slicing notation (<code>tensor[a:b]<\/code>) is just shorthand for the longer tensorflow notation <code>tf.slice(tensor, a, a+b)<\/code> which outputs a new tensor operation on the graph (see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice<\/a>). <\/p>\n\n<p>The graph you are trying to make looks like (with tensor output types indicated in parentheses):<\/p>\n\n<p><code>Var1<\/code> (tf.Variable) -> <code>tf.slice<\/code> (tf.Tensor) tensor -> <code>tf.assign<\/code> (tf.Variable).<\/p>\n\n<p>Since assign only works on <code>tf.Variable<\/code> objects, it can't work on the output of the slice op.<\/p>\n"},{"AnswerId":"52898022","CreationDate":"2018-10-19T18:23:43.430","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The other answer is correct; doing any operation to a tf variable does not (always) return a tf variable. So if you're chaining assignments, use explicit control dependencies:<\/p>\n\n<pre><code>v = tf.Variable(...)\nwith tf.control_dependencies([v[...].assign(...)]):\n  return v[...].assign(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52824438,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-15T20:38:07.430","AcceptedAnswerId":null,"OwnerUserId":5535186.0,"Title":"Multiple ckpt, meta into single .pb file?","Body":"<p>In tensorflow, the training produced the following files:<\/p>\n\n<pre><code>checkpoint\nmodel.ckpt-10000.meta\nmodel.ckpt-10000.data-00000-of-00001\nmodel.ckpt-10000.index\n\nmodel.ckpt-11000.meta\nmodel.ckpt-11000.data-00000-of-00001\nmodel.ckpt-11000.index\n\nmodel.ckpt-12000.meta\nmodel.ckpt-12000.data-00000-of-00001\nmodel.ckpt-12000.index\n\nmodel.ckpt-8000.meta\nmodel.ckpt-8000.data-00000-of-00001\nmodel.ckpt-8000.index\n\nmodel.ckpt-9000.meta\nmodel.ckpt-9000.data-00000-of-00001\nmodel.ckpt-9000.index\n<\/code><\/pre>\n\n<p>I am interested in creating a .pb file from the output generated training; however, from the examples I have seen, it requires one set of intermediate output files. How do I merge all the output set files into a single .pb? <\/p>\n","answers":[{"AnswerId":"52825218","CreationDate":"2018-10-15T21:49:15.353","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>What you are trying to do does not make sense (at least to me). I recommend you to read about these checkpoint files <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/www.tensorflow.org\/guide\/checkpoints\" rel=\"nofollow noreferrer\">here<\/a> at first.<\/p>\n\n<p>In short, <code>checkpoint<\/code> file just tells you what is the latest model. The <code>.meta<\/code> file stores info about your graph structure, <code>.data<\/code> stores values for variables and <code>.index<\/code> stores key\/value pairs which have info where the values for each parameter can be found in .data files.<\/p>\n\n<p>All your files look like <code>model.ckpt-xxxx<\/code>. This <code>xxxx<\/code> is the step number. So you have snapshots of training at different steps. And this is why it does not make sense to combine the value of the variable at step 9000 with the value at step 11000. Also <code>.meta<\/code> files are probably all the same.<\/p>\n"}]}
+{"QuestionId":52825005,"AnswerCount":1,"Tags":"<tensorflow><generator><mnist>","CreationDate":"2018-10-15T21:29:09.173","AcceptedAnswerId":null,"OwnerUserId":10257663.0,"Title":"MNIST GAN generators white area in middle surrounded by black","Body":"<p>The following code is copied from a GAN MNIST tutorial on UDEMY. When I run the code, it converges towards creating images with a large white area in the center that is black at the sides (picture an empty filled circle against a black background). I have no idea what the problem is as I have only done what the tutorial told me to do word for word. The only difference is that I extract the MNIST data differently. Is there something about tensorflow that has changed recently?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport gzip\nfrom PIL import Image\nimport os.path\n\ndef extract_data(filename, num_images):\n        \"\"\"Extract the images into a 4D tensor [image index, y, x, channels].\n        Values are rescaled from [0, 255] down to [-0.5, 0.5].\n        \"\"\"\n        print('Extracting', filename)\n        with gzip.open(filename) as bytestream:\n            bytestream.read(16)\n            buf = bytestream.read(28 * 28 * num_images)\n            data = np.frombuffer(buf, dtype=np.uint8).astype(np.float32)\n            #data = (data - (PIXEL_DEPTH \/ 2.0)) \/ PIXEL_DEPTH\n            data = data.reshape(num_images, 28, 28, 1)\n            return data\n\nfname_img_train = extract_data('..\/Data\/MNIST\/train-images-idx3-ubyte.gz', 60000)\n\ndef generator(z, reuse=None):\n    with tf.variable_scope('gen',reuse=reuse):\n        hidden1 = tf.layers.dense(inputs=z,units=128)\n        alpha = 0.01\n        hidden1=tf.maximum(alpha*hidden1,hidden1)\n        hidden2=tf.layers.dense(inputs=hidden1,units=128)\n        hidden2 = tf.maximum(alpha*hidden2,hidden2)\n        output=tf.layers.dense(hidden2,units=784, activation=tf.nn.tanh)\n        return output\n\ndef discriminator(X, reuse=None):\n    with tf.variable_scope('dis',reuse=reuse):\n        hidden1=tf.layers.dense(inputs=X,units=128)\n        alpha=0.01\n        hidden1=tf.maximum(alpha*hidden1,hidden1)\n        hidden2=tf.layers.dense(inputs=hidden1,units=128)\n        hidden2=tf.maximum(alpha*hidden2,hidden2)\n        logits=tf.layers.dense(hidden2,units=1)\n        output=tf.sigmoid(logits)\n        return output, logits\n\nreal_images=tf.placeholder(tf.float32,shape=[None,784])\nz=tf.placeholder(tf.float32,shape=[None,100])\nG = generator(z)\nD_output_real, D_logits_real = discriminator(real_images)\nD_output_fake, D_logits_fake = discriminator(G,reuse=True)\n\ndef loss_func(logits_in,labels_in):\n    return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(\n        logits=logits_in,labels=labels_in))\n\nD_real_loss = loss_func(D_logits_real,tf.ones_like(D_logits_real)*0.9)\nD_fake_loss = loss_func(D_logits_fake,tf.zeros_like(D_logits_real))\n\nD_loss = D_real_loss + D_fake_loss\nG_loss = loss_func(D_logits_fake,tf.ones_like(D_logits_fake))\n\nlearning_rate = 0.001\n\ntvars = tf.trainable_variables()\n\nd_vars= [var for var in tvars if 'dis' in var.name]\ng_vars = [var for var in tvars if 'gen' in var.name]\n\nD_trainer = tf.train.AdamOptimizer(learning_rate).minimize(D_loss,var_list=d_vars)\nG_trainer = tf.train.AdamOptimizer(learning_rate).minimize(G_loss,var_list=g_vars)\n\nbatch_size=100\nepochs=30\nset_size=60000\n\ninit = tf.global_variables_initializer()\nsamples=[]\ndef create_image(img, name):\n        img = np.reshape(img, (28, 28))\n        print(\"before\")\n        print(img)\n        img = (np.multiply(np.divide(np.add(img, 1.0), 2.0),255.0).astype(np.int16))\n        print(\"after\")\n        print(img)\n        im = Image.fromarray(img.astype('uint8'))\n        im.save(name)\nwith tf.Session() as sess:\n    sess.run(init)\n    for epoch in range(epochs):\n        np.random.shuffle(fname_img_train)\n        num_batches=int(set_size\/batch_size)\n        for i in range(num_batches):\n            batch = fname_img_train[i*batch_size:((i+1)*batch_size)]\n            batch_images = np.reshape(batch, (batch_size,784))\n            batch_images = batch_images*2.0-1.0\n            batch_z = np.random.uniform(-1,1,size=(batch_size,100))\n            _ = sess.run(D_trainer, feed_dict={real_images:batch_images,z:batch_z})\n            _ = sess.run(G_trainer,feed_dict={z:batch_z})\n\n        print(\"ON EPOCH {}\".format(epoch))\n        sample_z = np.random.uniform(-1,1,size=(batch_size,100))\n        gen_sample = sess.run(G,feed_dict={z:sample_z})\n        create_image(gen_sample[0], \"img\"+str(epoch)+\".png\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"53860914","CreationDate":"2018-12-20T00:18:19.283","ParentId":null,"OwnerUserId":"6200698","Title":null,"Body":"<p>As far as I can see, you are not normalizing the training data. Instead of using your <code>extract_data()<\/code> function, it is much easier to do the following:<\/p>\n\n<pre><code>from tensorflow.keras.datasets.mnist import load_data()\n(train_data, train_labels), _ = load_data()\ntrain_data \/= 255.\n<\/code><\/pre>\n\n<p>Besides, usually people sample twice from the latent space each epoch: once for the discriminator and once for the generator. Still, it did not seem to make a difference.<\/p>\n\n<p>After implementing these changes, using a batch size of 200 and training for 100 epochs, I got the following result: <a href=\"https:\/\/i.stack.imgur.com\/XJewY.png\" rel=\"nofollow noreferrer\">gen_sample<\/a>. The result is pretty bad, but it is definitely better than an \"empty filled circle against a black background\". \nNote that the architecture of the generator and of the discriminator that you are using is very simple. From my experience, stacking some convolutional layers gives perfect results. In addition, I would not use the <code>tf.maximum()<\/code> function, since it creates discontinuities that may negatively impact the flow of the gradients.<\/p>\n\n<p>Finally, instead of your <code>create_image()<\/code> function, I used the following:<\/p>\n\n<pre><code>def plot_mnist(samples, name):\n    fig = plt.figure(figsize=(6,6))\n    gs = gridspec.GridSpec(6,6)\n    gs.update(wspace=0.05, hspace=0.05)\n    for i, sample in enumerate(samples):\n        ax = plt.subplot(gs[i])\n        plt.axis('off')\n        ax.set_xticklabels([])\n        ax.set_yticklabels([])\n        ax.set_aspect('equal')\n    plt.imshow(sample.reshape(28,28), cmap='Greys_r')\n    plt.savefig('{}.png'.format(name))\n    plt.close()\n<\/code><\/pre>\n\n<p>There are many different ways of improving the quality of a GAN model, and the majority of those techniques can be easily found online. Please let me know if you have any specific question.<\/p>\n"}]}
+{"QuestionId":52825381,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-15T22:07:10.953","AcceptedAnswerId":52825851.0,"OwnerUserId":10480709.0,"Title":"File path not read correctly","Body":"<p>I'm new to DL and I'm going by a book to build my first CNN. The thing is (and I've tried various ways) i can't get the files to open a transfer a portion of the images into a smaller dataset. Below is the code and the error i get when i try to run my program. I've tried to change the backlashes in every way i can think of but i can't seem to get the program to check into the subfolder of the catsvdogs folder.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>import os, shutil\n\n\n# Path to the directory where the original DS was uncompressed\n\noriginal_dir = 'datasets\/catsvdogs'\n# Directory to store my smaller dataset\n\nbase_dir = 'datasets\/catsvdogs_small'\nos.mkdir(base_dir)\n\n# Directories for the training, validation, and test splits\n\ntrain_dir = os.path.join(base_dir, 'train')\nos.mkdir(train_dir)\nvalidation_dir = os.path.join(base_dir, 'valid')\nos.mkdir(validation_dir)\ntest_dir = os.path.join(base_dir, 'test')\nos.mkdir(test_dir)\n\n# Directory with training cat pictures\n\ntrain_cats_dir = os.path.join(train_dir, 'cats')\nos.mkdir(train_cats_dir)\n\n# Directory with training dog pictures \n\ntrain_dogs_dir = os.path.join(train_dir, 'dogs')\nos.mkdir(train_dogs_dir)\n\n# Directory with validation cat pictures \n\nvalidation_cats_dir = os.path.join(validation_dir, 'cats')\nos.mkdir(validation_cats_dir)\n\n# Directory with validation dog pictures\n\nvalidation_dogs_dir = os.path.join(validation_dir, 'dogs')\nos.mkdir(validation_dogs_dir)\n\n# Directory with test cat pictures\n\ntest_cats_dir = os.path.join(test_dir, 'cats')\nos.mkdir(test_cats_dir)\n\n# Directory with test dog pictures\n\ntest_dogs_dir = os.path.join(test_dir, 'dogs')\nos.mkdir(test_dogs_dir)\n\n# Copying the first 1000 cat images to train_cats_dir\n\nfnames = ['cat.{}.jpg'.format(i) for i in range(1000)]\nfor fname in fnames:\n    src = os.path.join(original_dir + 'train', fname)\n    dst = os.path.join(train_cats_dir, fname)\n    shutil.copyfile(src, dst)\n<\/code><\/pre>\n\n<p>Error i get:<\/p>\n\n<pre><code>[Errno 2] No such file or directory: 'datasets\/catsvdogstrain\\\\cat.0.jpg'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52825851","CreationDate":"2018-10-15T23:05:53.060","ParentId":null,"OwnerUserId":"6064179","Title":null,"Body":"<p>So I think there are two issues:<\/p>\n\n<ol>\n<li>You use Unix style slashes at first ('\/') but you are on a windows system. So when you call os.path.join later, you also introduce windows style slashes ('\\\\'). That's why in your error message you see both types of slashes. <\/li>\n<\/ol>\n\n<p><strong>To address this<\/strong>: use os.path.join to create all strings that represent directories (see below)<\/p>\n\n<ol start=\"2\">\n<li>In your 3rd to last line <code>src = os.path.join(original_dir + 'train', fname)<\/code>, you directly concatenate two strings rather than joining them with a slash. That's why you get <code>datasets\/catsvdogstrain<\/code> in your error message instead of the expected \n<code>datasets\/catsvdogs\/train<\/code>.<\/li>\n<\/ol>\n\n<p><strong>To address this<\/strong>: Join original_dir and 'train' with os.path.join (see below)<\/p>\n\n<p>Here's how you can fix your code:<\/p>\n\n<pre><code>import os, shutil\n\n# First, use os.path.join to reference your directories.\n# That way you automatically pick the right \"slash\" (i.e. \/ or \\\\)\noriginal_dir = os.path.join('datasets', 'catsvdogs')\nbase_dir = os.path.join('datasets', 'catsvdogs_small')\n\ntrain_dir = os.path.join(base_dir, 'train')\nos.mkdir(train_dir)\n\ntrain_cats_dir = os.path.join(train_dir, 'cats')\nos.mkdir(train_cats_dir)\n\n# Copying the first 1000 cat images to train_cats_dir\nfnames = ['cat.{}.jpg'.format(i) for i in range(1000)]\nfor fname in fnames:\n\n    # This will make a filepath that looks like 'datasets\/catsvdogs\/train'\n    # You directly joined two strings via original_dir + 'train' which\n    # gave you 'datasets\/catsvdogstrain' in your error message\n    src_dir = os.path.join(original_dir, 'train')\n\n    src = os.path.join(src_dir, fname)\n    dst = os.path.join(train_cats_dir, fname)\n\n    shutil.copyfile(src, dst)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52825696,"AnswerCount":2,"Tags":"<angular><typescript><tensorflow><tensorflow.js>","CreationDate":"2018-10-15T22:46:01.927","AcceptedAnswerId":null,"OwnerUserId":5148197.0,"Title":"TensaflowJS: Error: The shape of dict['images'] provided in model.execute(dict) must be [-1,224,224,3], but was [400,600,1]","Body":"<p>I'm using <a href=\"https:\/\/tfhub.dev\/google\/imagenet\/mobilenet_v1_100_224\/classification\/1\" rel=\"nofollow noreferrer\">mobilenet<\/a> neuronal network from google to classify images. I'm using angular 6 + tensorflowJS to build my image classifier app.<\/p>\n\n<p>I'm trying to follow the steps provided by the <a href=\"https:\/\/github.com\/tensorflow\/tfjs-converter\" rel=\"nofollow noreferrer\">tfjs-converter library readme<\/a> and I came up with the following code:<\/p>\n\n<pre><code>import * as tf from '@tensorflow\/tfjs';\nimport { FrozenModel } from '@tensorflow\/tfjs';\n\nexport class NeuronalNetwork {\n    private MODEL_PATH: string = 'http:\/\/localhost:4200\/assets\/model\/tensorflowjs_model.pb';\n    private WEIGHTS_PATH: string = 'http:\/\/localhost:4200\/assets\/model\/weights_manifest.json';\n    constructor(){}\n    async loadModel() {\n        const localModel: FrozenModel = (await tf.loadFrozenModel(this.MODEL_PATH, this.WEIGHTS_PATH));\n        let image: any = document.getElementById('cat');\n\n        let pixels =  tf.fromPixels(image, 1);\n        let result = localModel.predict(pixels);\n\n    }\n    async predict(){\n        let image: any = document.getElementById('cat');\n        debugger;\n        this.model.execute({input: tf.fromPixels(image)});\n    }\n\n}\n<\/code><\/pre>\n\n<p>Image HTML Element:<\/p>\n\n<pre><code>&lt;img id=\"cat\" src=\"http:\/\/localhost:4200\/assets\/images\/cat.jpg\"\/&gt;\n<\/code><\/pre>\n\n<p>When I try to execute the <code>localModel.predict(pixels)<\/code> function, I get the following error:<\/p>\n\n<blockquote>\n  <p>Uncaught (in promise): Error: The shape of dict['images'] provided in model.execute(dict) must be [-1,224,224,3], but was [400,600,1]<\/p>\n<\/blockquote>\n\n<p>I'm a newbie in Tensorflow and TensorflowJS technologies. Any one know what I'm doing wrong?<\/p>\n","answers":[{"AnswerId":"54355945","CreationDate":"2019-01-24T21:59:28.957","ParentId":null,"OwnerUserId":"5087657","Title":null,"Body":"<p>Met the same problem, then found the solution in Github: <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/issues\/193\" rel=\"nofollow noreferrer\">Input appears to be wrong shape<\/a><\/p>\n\n<pre><code>const img = document.getElementById('myimg');\nconst tfImg = tf.fromPixels(img);\nconst smalImg = tf.image.resizeBilinear(tfImg, [368, 432]);\nconst resized = tf.cast(smalImg, 'float32');\nconst t4d = tf.tensor4d(Array.from(resized.dataSync()),[1,368,432,3])\n<\/code><\/pre>\n"},{"AnswerId":"53918617","CreationDate":"2018-12-25T00:56:05.657","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The error indicates that there is a mismatch between the shape defined in the model and the shaped of the input passed to the model. To solve it you have to reshape your input.<\/p>\n\n<p>However using <a href=\"https:\/\/js.tensorflow.org\/api\/0.14.1\/#reshape\" rel=\"nofollow noreferrer\">tf.reshape<\/a> of the input will not work as there is no integer k such as <code>k*224*224*3<\/code> will be equal to <code>400*600*1<\/code>. One can consider slicing of the input image using <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#reshape\" rel=\"nofollow noreferrer\">tf.slice<\/a> so as to keep only the necessary shape for the image.<\/p>\n\n<pre><code>const a = tf.randomNormal([400, 600, 3])\n\/\/ a has shape [400, 600, 3]\n\/\/ a cannot be reshaped into [224, 224, 3]\n\/\/ a can be sliced into [224, 224, 3]\nconst b = a.slice([0, 0, 0], [224, 224, 3])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52826134,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-15T23:46:30.637","AcceptedAnswerId":null,"OwnerUserId":5632454.0,"Title":"keras model subclassing examples","Body":"<p>starting Keras 2.2.0, the 3rd API of model definition is released: Model subclassing.<\/p>\n\n<p>According to the FAQ: <\/p>\n\n<blockquote>\n  <p>However, in subclassed models, the model's topology is defined as\n  Python code (rather than as a static graph of layers). That means the\n  model's topology cannot be inspected or serialized. As a result, the\n  following methods and attributes are not available for subclassed\n  models:<\/p>\n  \n  <p>model.inputs and model.outputs. model.to_yaml() and model.to_json()\n  model.get_config() and model.save().<\/p>\n<\/blockquote>\n\n<p>The only option to save the trained model for inference is to use <strong>model.save_weights<\/strong> method. However, I have not had luck in loading the model back for inference. Encountered error messages include:<\/p>\n\n<blockquote>\n  <p>This model has never been called, thus its weights have not yet been created, so no summary can be displayed. Build the model first (e.g. by calling it on some test data).\n  You are trying to load a weight file containing 4 layers into a model with 0 layers.\n  NotImplementedError<\/p>\n<\/blockquote>\n\n<p>Can anyone give a full toy example for creating a subclassed keras model, train, and save_weights, then load it back for inference?<\/p>\n","answers":[{"AnswerId":"56823828","CreationDate":"2019-06-30T09:41:47.960","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>You need to call the tf.keras.Model.build method before you try to save a subclassed model weights. An alternative to this would be calling tf.keras.Model.fit or tf.keras.Model.fit.call on some inputs before you try to save your model weights. This same applies to load weights into a newly created instance of your subclassed model. you need to call one of the above-mentioned methods before you try to load your weights.\nHere is an example showing both saving and loading weights for a subclassed model<\/p>\n\n<pre><code>import tensorflow as tf\n\nprint('TensorFlow', tf.__version__)\n\nclass ResidualBlock(tf.keras.Model):\n    def __init__(self, block_type=None, n_filters=None):\n        super(ResidualBlock, self).__init__()\n        self.n_filters = n_filters\n        if block_type == 'identity':\n            self.strides = 1\n        elif block_type == 'conv':\n            self.strides = 2\n            self.conv_shorcut = tf.keras.layers.Conv2D(filters=self.n_filters, \n                               kernel_size=1, \n                               padding='same',\n                               strides=self.strides,\n                               kernel_initializer='he_normal')\n            self.bn_shortcut = tf.keras.layers.BatchNormalization(momentum=0.9)\n\n        self.conv_1 = tf.keras.layers.Conv2D(filters=self.n_filters, \n                               kernel_size=3, \n                               padding='same',\n                               strides=self.strides,\n                               kernel_initializer='he_normal')\n        self.bn_1 = tf.keras.layers.BatchNormalization(momentum=0.9)\n        self.relu_1 = tf.keras.layers.ReLU()\n\n        self.conv_2 = tf.keras.layers.Conv2D(filters=self.n_filters, \n                               kernel_size=3, \n                               padding='same', \n                               kernel_initializer='he_normal')\n        self.bn_2 = tf.keras.layers.BatchNormalization(momentum=0.9)\n        self.relu_2 = tf.keras.layers.ReLU()\n\n    def call(self, x, training=False):\n        shortcut = x\n        if self.strides == 2:\n            shortcut = self.conv_shorcut(x)\n            shortcut = self.bn_shortcut(shortcut)\n        y = self.conv_1(x)\n        y = self.bn_1(y)\n        y = self.relu_1(y)\n        y = self.conv_2(y)\n        y = self.bn_2(y)\n        y = tf.add(shortcut, y)\n        y = self.relu_2(y)\n        return y\n\nclass ResNet34(tf.keras.Model):\n    def __init__(self, include_top=True, n_classes=1000):\n        super(ResNet34, self).__init__()\n\n        self.n_classes = n_classes\n        self.include_top = include_top\n        self.conv_1 = tf.keras.layers.Conv2D(filters=64, \n                                               kernel_size=7, \n                                               padding='same', \n                                               strides=2, \n                                               kernel_initializer='he_normal')\n        self.bn_1 = tf.keras.layers.BatchNormalization(momentum=0.9)\n        self.relu_1 = tf.keras.layers.ReLU()\n        self.maxpool = tf.keras.layers.MaxPool2D(3, 2, padding='same')\n        self.residual_blocks = tf.keras.Sequential()\n        for n_filters, reps, downscale in zip([64, 128, 256, 512], \n                                              [3, 4, 6, 3], \n                                              [False, True, True, True]):\n            for i in range(reps):\n                if i == 0 and downscale:\n                    self.residual_blocks.add(ResidualBlock(block_type='conv', \n                                                              n_filters=n_filters))\n                else:\n                    self.residual_blocks.add(ResidualBlock(block_type='identity', \n                                                              n_filters=n_filters))\n        self.GAP = tf.keras.layers.GlobalAveragePooling2D()\n        self.fc = tf.keras.layers.Dense(units=self.n_classes)\n\n    def call(self, x, training=False):\n        y = self.conv_1(x)\n        y = self.bn_1(y)\n        y = self.relu_1(y)\n        y = self.maxpool(y)\n        y = self.residual_blocks(y)\n        if self.include_top:\n            y = self.GAP(y)\n            y = self.fc(y)\n        return y\n\n## saving weights\nmodel = ResNet34()\nmodel.build((1, 224, 224, 3))\nmodel.summary()\nmodel.save_weights('model_weights.h5')\n\n## loading saved weights\nmodel_new = ResNet34()\nmodel_new.build((1, 224, 224, 3))\nmodel_new.load_weights('model_weights.h5')\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":52826223,"AnswerCount":1,"Tags":"<amazon-web-services><ubuntu><tensorflow><multiprocessing>","CreationDate":"2018-10-15T23:59:03.293","AcceptedAnswerId":null,"OwnerUserId":10510044.0,"Title":"CPU usage drops on Linux (Ubuntu)","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/XFEwQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XFEwQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I was using AWS c5.4xlarge instance, which has 16 vCPU, and running a 10 processes python programme. However, the CPU usage of each process gradually dropped down to 10% as showed in the picture in just 10 seconds. The total CPU usage of the 16 vCPU instances was just about 6%.<\/p>\n\n<p>I reduced the number of the processes, but the CPU usage of each process was still quite low. Everything is ok on my own macOS.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8z5s9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8z5s9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What is wrong with this?<\/p>\n","answers":[{"AnswerId":"52844590","CreationDate":"2018-10-16T21:59:32.720","ParentId":null,"OwnerUserId":"10510044","Title":null,"Body":"<p>Ok I find the answer. This is about processor affinity. For Linux beginner:\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Processor_affinity\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Processor_affinity<\/a><\/p>\n\n<p>In Linux you can allocate a cpu to a specific process in the Linux ternimal:<\/p>\n\n<pre><code>$ taskset -cp CPU_ID PID\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>$ taskset -cp 0-1 1000\n<\/code><\/pre>\n\n<p>will allocate CPU 0 and 1 to process with ID 1000.<\/p>\n\n<p>You can find the PID by using<\/p>\n\n<pre><code>$ top\n<\/code><\/pre>\n\n<p>in your terminal.<\/p>\n"}]}
+{"QuestionId":52826311,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-16T00:14:38.443","AcceptedAnswerId":52845549.0,"OwnerUserId":3511901.0,"Title":"Keras Graph disconnects when combining a LSTM with a CNN","Body":"<p>The idea is to train a CNN on a cosine similarity matrix of the hidden states of two bilstms. \nI try to get the following code working, but it is failing giving the error message:<\/p>\n\n<pre><code>Graph disconnected: cannot obtain value for tensor\nTensor(\"bidirectional_4\/concat:0\", shape=(?, ?, 100), dtype=float32)\nat layer \"input_11\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>The code to train the model is the following: <\/p>\n\n<pre><code>def train_model(self, sentences_pair, is_similar, \n    embedding_meta_data_skt, embedding_meta_data_tib , \n    model_save_directory='.\/'):\n    tokenizer_skt = embedding_meta_data_skt['tokenizer']\n    tokenizer_tib = embedding_meta_data_tib['tokenizer']\n    embedding_matrix_skt = embedding_meta_data_skt['embedding_matrix']\n    embedding_matrix_tib = embedding_meta_data_tib['embedding_matrix']\n\n    train_data_x1, train_data_x2, train_labels, leaks_train, \\\n    val_data_x1, val_data_x2, val_labels, leaks_val = create_train_dev_set(tokenizer_skt, sentences_pair,\n                                                                           is_similar, self.max_sequence_length,\n                                                                           self.validation_split_ratio)\n\n    nb_words_skt = len(tokenizer_skt.word_index) + 1\n    nb_words_tib = len(tokenizer_tib.word_index) + 1\n    # Creating word embedding layer\n    embedding_layer_skt = Embedding(nb_words_skt, self.embedding_dim, weights=[embedding_matrix_skt],\n                                input_length=self.max_sequence_length, trainable=False)\n    embedding_layer_tib = Embedding(nb_words_tib, self.embedding_dim, weights=[embedding_matrix_tib],\n                                input_length=self.max_sequence_length, trainable=False)\n\n\n    # Creating LSTM Encoder\n    lstm_layer = Bidirectional(LSTM(self.number_lstm_units, dropout=self.rate_drop_lstm, recurrent_dropout=self.rate_drop_lstm,return_sequences=True))\n\n    # Creating LSTM Encoder layer for First Sentence\n    sequence_1_input = Input(shape=(self.max_sequence_length,), dtype='int32')\n    embedded_sequences_1 = embedding_layer_skt(sequence_1_input)\n    skt_lstm = lstm_layer(embedded_sequences_1)\n\n    # Creating LSTM Encoder layer for Second Sentence\n    sequence_2_input = Input(shape=(self.max_sequence_length,), dtype='int32')\n    embedded_sequences_2 = embedding_layer_tib(sequence_2_input)\n    tib_lstm = lstm_layer(embedded_sequences_2)\n\n    A_input = keras.Input(tensor=skt_lstm)\n    B_input = keras.Input(tensor=tib_lstm)\n    dist_output = keras.layers.Lambda(pairwise_cosine_sim)([skt_lstm,tib_lstm,A_input,B_input])\n    dist_output = Reshape((40,40,1))(dist_output)\n    input_shape = (40,40,1)\n    cnn_model = Conv2D(128, (2, 2), input_shape=input_shape)(dist_output)\n    cnn_model = BatchNormalization(axis=-1)(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = Conv2D(164, (2, 2))(cnn_model)\n    cnn_model = BatchNormalization(axis=-1)(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = Conv2D(192,(3, 3))(cnn_model)\n    cnn_model = BatchNormalization(axis=-1)(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = Conv2D(192, (3, 3))(cnn_model)\n    cnn_model = BatchNormalization(axis=-1)(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = Conv2D(128, (3, 3))(cnn_model)\n    cnn_model = BatchNormalization(axis=-1)(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = MaxPooling2D(pool_size=(2,2))(cnn_model)\n    cnn_model = Dropout(0.40)(cnn_model)\n    cnn_model = Flatten()(cnn_model)\n    # Fully connected layer\n    cnn_model = Dense(256)(cnn_model)\n    cnn_model = BatchNormalization()(cnn_model)\n    cnn_model = Activation('relu')(cnn_model)\n    cnn_model = Dropout(0.5)(cnn_model)\n    cnn_model = Dense(num_classes)(cnn_model)\n    preds = Dense(1, activation='sigmoid')(cnn_model)\n    model = Model(inputs=[sequence_1_input, sequence_2_input], outputs=preds)\n    model.compile(loss=keras.losses.binary_crossentropy,\n                  optimizer=keras.optimizers.Adam(lr=learning_rate),\n                  metrics=['accuracy'])\n\n    #model.compile(loss='binary_crossentropy', optimizer='nadam', metrics=['acc'])\n\n    filepath=\"skt-tib-bs\" + str(batch_size) + \"-\" + \"{epoch:02d}-{val_acc:.2f}.hdf5\"\n    checkpoint = ModelCheckpoint('skt-tib.h5', monitor='val_acc')\n    callbacks_list = [checkpoint]\n\n    model.fit([train_data_x1, train_data_x2, leaks_train], train_labels,validation_data=([val_data_x1, val_data_x2, leaks_val], val_labels),\n              batch_size=batch_size,\n              epochs=epochs,\n              verbose=1,\n              class_weight = class_weight,\n              callbacks = callbacks_list)\n    score = model.evaluate(x_test, y_test, verbose=1)\n    print('Test loss:', score[0])\n    print('Test accuracy:', score[1])\n    model.save(file_name)\n<\/code><\/pre>\n\n<p>The definition of the function calculating the pairwise cosine similarity is the following: <\/p>\n\n<pre><code>def l2_norm(x, axis=None):\n\n    square_sum = K.sum(K.square(x), axis=axis, keepdims=True)\n    norm = K.sqrt(K.maximum(square_sum, K.epsilon()))\n\n    return norm\n\ndef pairwise_cosine_sim(A_B):\n\n    A,B,A_tensor,B_tensor = A_B\n    A_mag = l2_norm(A, axis=2)\n    B_mag = l2_norm(B, axis=2)\n    num = K.batch_dot(A_tensor, K.permute_dimensions(B_tensor, (0,2,1)))\n    den = (A_mag * K.permute_dimensions(B_mag, (0,2,1)))\n    dist_mat =  num \/ den\n    return dist_mat\n<\/code><\/pre>\n\n<p>I Have been trying for a couple of hours to fix it, but it seems to be no good. Somewhere the input and outputs are not connected, but I just can't figure out where the problem lies. Any suggestions on this?<\/p>\n","answers":[{"AnswerId":"52845549","CreationDate":"2018-10-17T00:05:36.283","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Either remove <code>A_input<\/code> and <code>B_input<\/code> entirely as they are not input layers in the first place and use <code>skt_lstm<\/code> and <code>tib_lstm<\/code> directly instead of them, or if you would like to keep them pass them as the inputs of the model as well when you are defining the <code>Model<\/code> since they are actually input layers:<\/p>\n\n<pre><code>model = Model(inputs=[sequence_1_input, sequence_2_input, A_input, B_input], outputs=preds)\n<\/code><\/pre>\n\n<p>However, you don't need to pass any corresponding arrays for them when calling <code>fit<\/code> method as they will be fed using their corresponding tensors <code>skt_lstm<\/code> and <code>tib_lstm<\/code> (i.e. they will act as wrappers around these tensors).<\/p>\n"}]}
+{"QuestionId":52826404,"AnswerCount":1,"Tags":"<python><machine-learning><keras><computer-vision><resnet>","CreationDate":"2018-10-16T00:30:19.110","AcceptedAnswerId":null,"OwnerUserId":4592059.0,"Title":"Why using pretrained ResNet50 results in contradictory losses in training and validation set?","Body":"<p>I'm using ResNet50 pretrained model as building block for a Unet:<\/p>\n\n<pre><code>def ResNet50(include_top=True, weights='imagenet',\n         input_tensor=None, input_shape=None,\n         pooling=None,\n         classes=1000):\nif weights not in {'imagenet', None}:\n    raise ValueError('The `weights` argument should be either '\n                     '`None` (random initialization) or `imagenet` '\n                     '(pre-training on ImageNet).')\n\nif weights == 'imagenet' and include_top and classes != 1000:\n    raise ValueError('If using `weights` as imagenet with `include_top`'\n                     ' as true, `classes` should be 1000')\n\nif input_tensor is None:\n    img_input = Input(shape=input_shape)\nelse:\n    if not K.is_keras_tensor(input_tensor):\n        img_input = Input(tensor=input_tensor, shape=input_shape)\n    else:\n        img_input = input_tensor\nif K.image_data_format() == 'channels_last':\n    bn_axis = 3\nelse:\n    bn_axis = 1\n\nx = Conv2D(64, (7, 7), strides=(2, 2), padding='same', name='conv1')(img_input)\nx = BatchNormalization(axis=bn_axis, name='bn_conv1')(x)\nx = Activation('relu')(x)\nx = MaxPooling2D((3, 3), strides=(2, 2), padding=\"same\")(x)\n\nx = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))\nx = identity_block(x, 3, [64, 64, 256], stage=2, block='b')\nx = identity_block(x, 3, [64, 64, 256], stage=2, block='c')\n\nx = conv_block(x, 3, [128, 128, 512], stage=3, block='a')\nx = identity_block(x, 3, [128, 128, 512], stage=3, block='b')\nx = identity_block(x, 3, [128, 128, 512], stage=3, block='c')\nx = identity_block(x, 3, [128, 128, 512], stage=3, block='d')\n\nx = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')\nx = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')\nx = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')\nx = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')\nx = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')\nx = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')\n\nx = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')\nx = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')\nx = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')\n\n# Ensure that the model takes into account\n# any potential predecessors of `input_tensor`.\nif input_tensor is not None:\n    inputs = get_source_inputs(input_tensor)\nelse:\n    inputs = img_input\n# Create model.\nmodel = Model(inputs, x, name='resnet50')\n\n# load weights\nif weights == 'imagenet':\n    if include_top:\n        weights_path = get_file('resnet50_weights_tf_dim_ordering_tf_kernels.h5',\n                                WEIGHTS_PATH,\n                                cache_subdir='models',\n                                md5_hash='a7b3fe01876f51b976af0dea6bc144eb')\n    else:\n        weights_path = get_file('resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5',\n                                WEIGHTS_PATH_NO_TOP,\n                                cache_subdir='models',\n                                md5_hash='a268eb855778b3df3c7506639542a6af')\n    model.load_weights(weights_path,by_name=True)\nreturn model\n<\/code><\/pre>\n\n<p>Creating the Unet:<\/p>\n\n<pre><code>def conv_block_simple(prevlayer, filters, prefix, strides=(1, 1)):\nconv = Conv2D(filters, (3, 3), padding=\"same\", kernel_initializer=\"he_normal\", strides=strides, name=prefix + \"_conv\")(prevlayer)\nconv = BatchNormalization(name=prefix + \"_bn\")(conv)\nconv = Activation('relu', name=prefix + \"_activation\")(conv)\nreturn conv \n\ndef conv_block_simple_no_bn(prevlayer, filters, prefix, strides=(1, 1)):\nconv = Conv2D(filters, (3, 3), padding=\"same\", kernel_initializer=\"he_normal\", strides=strides, name=prefix + \"_conv\")(prevlayer)\nconv = Activation('relu', name=prefix + \"_activation\")(conv)\nreturn conv\n\nK.clear_session()\n\ndef get_unet_resnet(input_shape):\n    resnet_base = ResNet50(input_shape=input_shape, include_top=False)\n\n    for l in resnet_base.layers:\n        l.trainable = False \n\n\n    conv1 = resnet_base.get_layer(\"activation_1\").output\n    conv2 = resnet_base.get_layer(\"activation_10\").output\n    conv3 = resnet_base.get_layer(\"activation_22\").output\n    conv4 = resnet_base.get_layer(\"activation_40\").output\n    conv5 = resnet_base.get_layer(\"activation_49\").output\n\n    up6 = concatenate([UpSampling2D()(conv5), conv4], axis=-1)\n    conv6 = conv_block_simple(up6, 256, \"conv6_1\")\n    conv6 = conv_block_simple(conv6, 256, \"conv6_2\")\n\n    up7 = concatenate([UpSampling2D()(conv6), conv3], axis=-1)\n    conv7 = conv_block_simple(up7, 192, \"conv7_1\")\n    conv7 = conv_block_simple(conv7, 192, \"conv7_2\")\n\n    up8 = concatenate([UpSampling2D()(conv7), conv2], axis=-1)\n    conv8 = conv_block_simple(up8, 128, \"conv8_1\")\n    conv8 = conv_block_simple(conv8, 128, \"conv8_2\")\n\n    up9 = concatenate([UpSampling2D()(conv8), conv1], axis=-1)\n    conv9 = conv_block_simple(up9, 64, \"conv9_1\")\n    conv9 = conv_block_simple(conv9, 64, \"conv9_2\")\n\n    up10 = UpSampling2D()(conv9)\n    conv10 = conv_block_simple(up10, 32, \"conv10_1\")\n    conv10 = conv_block_simple(conv10, 32, \"conv10_2\")\n    conv10 = SpatialDropout2D(0.2)(conv10)\n    x = Conv2D(1, (1, 1), activation=\"sigmoid\", name=\"prediction\")(conv10)\n\n    model = Model(resnet_base.input, x)\n\n    model.summary()\n\n    return model\n<\/code><\/pre>\n\n<p>I freezed the pretrained ResNet50 layers as proposed in several papers:<\/p>\n\n<pre><code>for l in resnet_base.layers:\n    l.trainable = False \n<\/code><\/pre>\n\n<p>Without the freezing the network works fine, but tends to hugely overfit, which I decreased with higher <code>SpatialDropout2D()<\/code> value. \nWhen I freeze it however the train losses decrease but the validation losses circulate about some weird high value but actually stagnate.\nI can't figure out, why the frozen network do work on training set, meanwhile it doesn't on validation set. I see no logical reason for it (on my current knowledge level).<\/p>\n\n<p>I tried playing with learning rate, but no success.<\/p>\n\n<p>What could be the problem?\nAny help would be highly appreciated.\nThank you.<\/p>\n","answers":[{"AnswerId":"52977740","CreationDate":"2018-10-24T20:51:30.297","ParentId":null,"OwnerUserId":"4684130","Title":null,"Body":"<p>There is certaing problem (Keras creators are keen it is design) with BN layers in Keras. The problem is that during the training, BN layers are actually learning parameters of your new dataset, while during validation phase the parametrs of dataset ResNet was trained with are used (e.g. imagenet or cifar).\nThis results into good training accuracy, but validation accuracy will be just nonexistant. <\/p>\n\n<p>You can find more <a href=\"http:\/\/blog.datumbox.com\/the-batch-normalization-layer-of-keras-is-broken\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":52826906,"AnswerCount":1,"Tags":"<python><image-processing><machine-learning><keras>","CreationDate":"2018-10-16T01:56:54.777","AcceptedAnswerId":null,"OwnerUserId":10144962.0,"Title":"Keras ImageDataGenerator flow_from_dataframe returning KeyError","Body":"<p>I am trying to build an image classifier with keras and the size of my dataset requires me to use the ImageDataGenerator class along with its flow_from_dataframe method. This is the code I am using.<\/p>\n\n<pre><code>train_datagen = keras.preprocessing.image.ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\ntrain_generator = train_datagen.flow_from_dataframe(\n        directory='stage_1_train_images\/',\n        dataframe=box.drop(labels=['patientId'], axis=1).replace(to_replace=float('nan'),value=0),\n        target_size=(1024, 1024))\nmodel = Sequential()\nmodel.add(Conv2D(32, kernel_size=(3, 3),activation='linear',input_shape=(28,28,1),padding='same'))\nmodel.add(LeakyReLU(alpha=0.1))\nmodel.add(MaxPooling2D((2, 2),padding='same'))\nmodel.add(Conv2D(64, (3, 3), activation='linear',padding='same'))\nmodel.add(LeakyReLU(alpha=0.1))\nmodel.add(MaxPooling2D(pool_size=(2, 2),padding='same'))\nmodel.add(Conv2D(128, (3, 3), activation='linear',padding='same'))\nmodel.add(LeakyReLU(alpha=0.1))                  \nmodel.add(MaxPooling2D(pool_size=(2, 2),padding='same'))\nmodel.add(Flatten())\nmodel.add(Dense(128, activation='linear'))\nmodel.add(LeakyReLU(alpha=0.1))                  \nmodel.add(Dense(num_classes, activation='softmax'))\nmodel.compile(loss=keras.losses.categorical_crossentropy,\n              optimizer=keras.optimizers.Adam(lr=1000,decay=.99),\n              metrics=['accuracy'])\nmodel.fit_generator(trainGen, steps_per_epoch=1024\/16, epochs=317)\n<\/code><\/pre>\n\n<p>However, when I run this code, I get the following error<\/p>\n\n<pre><code>KeyError                                  Traceback (most recent call last)\n&lt;ipython-input-7-5a88afda8de5&gt; in &lt;module&gt;\n      7         directory='stage_1_train_images\/',\n      8         dataframe=box.drop(labels=['patientId'], axis=1).replace(to_replace=float('nan'),value=0),\n----&gt; 9         target_size=(1024, 1024))\n     10 model = Sequential()\n     11 model.add(Conv2D(32, kernel_size=(3, 3),activation='linear',input_shape=(28,28,1),padding='same'))\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras_preprocessing\/image.py in flow_from_dataframe(self, dataframe, directory, x_col, y_col, has_ext, target_size, color_mode, classes, class_mode, batch_size, shuffle, seed, save_to_dir, save_prefix, save_format, subset, interpolation)\n   1105                                  save_format=save_format,\n   1106                                  subset=subset,\n-&gt; 1107                                  interpolation=interpolation)\n   1108 \n   1109     def standardize(self, x):\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras_preprocessing\/image.py in __init__(self, dataframe, directory, image_data_generator, x_col, y_col, has_ext, target_size, color_mode, classes, class_mode, batch_size, shuffle, seed, data_format, save_to_dir, save_prefix, save_format, follow_links, subset, interpolation, dtype)\n   2056             raise ValueError(\"has_ext must be either True if filenames in\"\n   2057                              \" x_col has extensions,else False.\")\n-&gt; 2058         self.df = dataframe.drop_duplicates(x_col)\n   2059         self.df[x_col] = self.df[x_col].astype(str)\n   2060         self.directory = directory\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/pandas\/core\/frame.py in drop_duplicates(self, subset, keep, inplace)\n   4329         \"\"\"\n   4330         inplace = validate_bool_kwarg(inplace, 'inplace')\n-&gt; 4331         duplicated = self.duplicated(subset, keep=keep)\n   4332 \n   4333         if inplace:\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/pandas\/core\/frame.py in duplicated(self, subset, keep)\n   4379         diff = Index(subset).difference(self.columns)\n   4380         if not diff.empty:\n-&gt; 4381             raise KeyError(diff)\n   4382 \n   4383         vals = (col.values for name, col in self.iteritems()\n\nKeyError: Index(['filename'], dtype='object')\n<\/code><\/pre>\n\n<p>What is going wrong? I have tried multiple things to fix this but cannot figure out why this is happening.<\/p>\n","answers":[{"AnswerId":"53418815","CreationDate":"2018-11-21T18:54:03.480","ParentId":null,"OwnerUserId":"7213391","Title":null,"Body":"<p>As per the documentation <a href=\"https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator-methods\" rel=\"nofollow noreferrer\">here<\/a>, you need to specify <code>x_col<\/code> and <code>y_col<\/code> as arguments in the <code>flow_from_dataframe<\/code> method. The defaults for <code>x_col<\/code> and <code>y_col<\/code> are 'filename' and 'class', respectively. From the error, I'm guessing that you don't have a column named <code>\"filename\"<\/code> in your DataFrame which is what causes the <code>KeyError<\/code>. To fix this, specify the following two arguments in the <code>flow_from_dataframe<\/code> method.<\/p>\n\n<blockquote>\n  <p><strong>x_col<\/strong>: string, column in the dataframe that contains\n         the filenames of the target images.<\/p>\n  \n  <p><strong>y_col<\/strong>: string or list of strings,columns in\n         the dataframe that will be the target data.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":52827390,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-10-16T03:04:06.637","AcceptedAnswerId":null,"OwnerUserId":9507516.0,"Title":"InvalidArgumentError when running model.fit()","Body":"<p>When I run the below code:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(64,(3,3),input_shape = X.shape[1:],data_format = \"channels_last\"))\nmodel.add(Activation('relu'))\nmodel.add(MaxPool2D(pool_size = (2,2)))\n\nmodel.add(Conv2D(64,(3,3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPool2D(pool_size=(2,2)))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\n\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))    \nmodel.compile(optimizer = 'adam',loss = 'binary_crossentropy',metrics = ['accuracy'])\nmodel.fit(X, y, batch_size=32, epochs=1,\n    validation_split=0.3)\n<\/code><\/pre>\n\n<p>I'm getting:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError                      Traceback (most recent call\n  last)  in ()\n        3           batch_size=32,\n        4           epochs=1,\n  ----> 5           validation_split=0.3)<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\n  in fit(self, x, y, batch_size, epochs, verbose, callbacks,\n  validation_split, validation_data, shuffle, class_weight,\n  sample_weight, initial_epoch, steps_per_epoch, validation_steps,\n  **kwargs)    1361           initial_epoch=initial_epoch,    1362           steps_per_epoch=steps_per_epoch,\n  -> 1363           validation_steps=validation_steps)    1364     1365   def evaluate(self,<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\n  in fit_loop(model, inputs, targets, sample_weights, batch_size,\n  epochs, verbose, callbacks, val_inputs, val_targets,\n  val_sample_weights, shuffle, callback_metrics, initial_epoch,\n  steps_per_epoch, validation_steps)\n      262           ins_batch[i] = ins_batch[i].toarray()\n      263 \n  --> 264         outs = f(ins_batch)\n      265         if not isinstance(outs, list):\n      266           outs = [outs]<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in\n  <strong>call<\/strong>(self, inputs)    2910         feed_symbols != self._feed_symbols or self.fetches != self._fetches or    2911<br>\n  session != self._session):\n  -> 2912       self._make_callable(feed_arrays, feed_symbols, symbol_vals, session)    2913     2914     fetched =\n  self._callable_fn(*array_vals)<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in\n  _make_callable(self, feed_arrays, feed_symbols, symbol_vals, session)    2855     callable_opts.target.append(self.updates_op.name)    2856    <\/p>\n  \n  <h1>Create callable.<\/h1>\n  \n  <p>-> 2857     callable_fn = session._make_callable_from_options(callable_opts)    2858     # Cache\n  parameters corresponding to the generated callable, so that    2859   <\/p>\n  \n  <h1>we can detect future mismatches and refresh the callable.<\/h1>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in\n  _make_callable_from_options(self, callable_options)    1412     \"\"\"    1413     self._extend_graph()\n  -> 1414     return BaseSession._Callable(self, callable_options)    1415     1416 <\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in\n  <strong>init<\/strong>(self, session, callable_options)    1366         with errors.raise_exception_on_not_ok_status() as status:    1367<br>\n  self._handle = tf_session.TF_SessionMakeCallable(\n  -> 1368               session._session, options_ptr, status)    1369       finally:    1370         tf_session.TF_DeleteBuffer(options_ptr)<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\n  in <strong>exit<\/strong>(self, type_arg, value_arg, traceback_arg)\n      517             None, None,\n      518             compat.as_text(c_api.TF_Message(self.status.status)),\n  --> 519             c_api.TF_GetCode(self.status.status))\n      520     # Delete the underlying status object from memory otherwise it stays alive\n      521     # as there is a reference to status from this from the traceback due to<\/p>\n  \n  <p>InvalidArgumentError: Default MaxPoolingOp only supports NHWC on\n  device type CPU    [[Node: max_pooling2d_2\/MaxPool =\n  MaxPoolT=DT_FLOAT,\n  _class=[\"loc:@training_3\/Adam\/gradients\/max_pooling2d_2\/MaxPool_grad\/MaxPoolGrad\"],\n  data_format=\"NCHW\", ksize=[1, 1, 2, 2], padding=\"VALID\", strides=[1,\n  1, 2, 2],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<h1>X.shape ----> (8000, 50, 50, 1)<\/h1>\n\n<p>Also I'm using 4GB RAM CPU. I've run only on one epoch. Does this have anything to do with the memory?<\/p>\n","answers":[]}
+{"QuestionId":52827405,"AnswerCount":0,"Tags":"<tensorflow><keras><convolution><tflearn>","CreationDate":"2018-10-16T03:07:24.417","AcceptedAnswerId":null,"OwnerUserId":9605699.0,"Title":"Changing the shape of convolution kernels in keras or tflearn using tensorflow","Body":"<p>I was wondering if there is a way to change the shape of convolution kernels in keras or tflearn using tensorflow backend.<\/p>\n\n<p>My objective is to test if shape of conv. 2d kernels affects the classification accuracy. That is if using circular, triangular or any other shape other than the default square kernels affect classification accuracy.<\/p>\n\n<p>Keras and tflearn documentation does not help, any resources on this would be greatly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52827483,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T03:17:22.703","AcceptedAnswerId":null,"OwnerUserId":7547975.0,"Title":"Tensorflow: What if my cond is not a scalar in tf.cond in while loop?","Body":"<p>in <code>tf.cond<\/code> of tensorflow, <code>cond<\/code> has to be a scalar, but in my case <code>cond<\/code> need to be rank 1 with shape [batch_size]. Is there any method to solve this problem? Have tensorflow provided a solution to it?<\/p>\n\n<pre><code>import tensorflow as tf\nseq_len = 10\nwhile_length = 10\nbatch_size = 4\n\noutput_ta = tf.TensorArray(\n        dtype=tf.int32,\n        size=seq_len,\n        tensor_array_name='example_1')\n\ncond_tensor = tf.constant([3, 4, 5, 6])\n\n\ndef _step(time, arrays):\n    time_tensor = tf.tile(tf.expand_dims(time, -1), multiples=[batch_size])\n    arrays = arrays\n\n    def _true_function():\n        return tf.constant([1] * seq_len)\n\n    def _false_function():\n        return tf.constant([0] * seq_len)\n\n    bool_cond = tf.less(time_tensor, cond_tensor)\n    arrays_write = tf.cond(bool_cond, true_fn=_true_function, false_fn=_false_function)\n    arrays = arrays.write(time, arrays_write)\n    return time + 1, arrays\n\n\ntrace_time, outputs_tensor_arrays = tf.while_loop(\n        cond=lambda time, *_: time &lt; while_length,\n        body=_step,\n        loop_vars=[0, output_ta],\n        parallel_iterations=32,\n        swap_memory=True)\n\naxes = [1, 0]\noutput = tf.transpose(outputs_tensor_arrays, axes)\n\nwith tf.Session() as sess:\n    sess.run(output)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52828877","CreationDate":"2018-10-16T06:06:15.913","ParentId":null,"OwnerUserId":"7547975","Title":null,"Body":"<p>One potential solution I realized is <code>tf.where<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nseq_len = 10\nbatch_size = 4\n\noutput_ta = tf.TensorArray(\n    dtype=tf.float32,\n    size=seq_len,\n    tensor_array_name='example_1')\n\ncond_tensor = tf.constant([3, 4, 5, 6])\n\nt1 = tf.ones(shape=[batch_size, seq_len])\nt2 = tf.zeros(shape=[batch_size, seq_len])\n\n\ndef _step(time, arrays):\n    time_tensor = tf.tile(tf.expand_dims(time, -1), multiples=[batch_size])\n    # arrays = arrays\n\n    bool_cond = tf.less(time_tensor, cond_tensor)\n    output_array = tf.where(bool_cond, t1, t2)\n    # arrays_write = tf.cond(bool_cond, true_fn=_true_function, false_fn=_false_function)\n    arrays = arrays.write(time, output_array)\n    return time + 1, arrays\n\n\ntrace_time, outputs_tensor_arrays = tf.while_loop(\n    cond=lambda time, *_: time &lt; seq_len,\n    body=_step,\n    loop_vars=[0, output_ta],\n    parallel_iterations=32,\n    swap_memory=True)\n\naxes = [1, 0, 2]\noutput = outputs_tensor_arrays.stack()\noutput = tf.transpose(output, axes)\n\nwith tf.Session() as sess:\n    r_output = sess.run(output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52827601,"AnswerCount":0,"Tags":"<python><c><pycharm><pytorch>","CreationDate":"2018-10-16T03:32:36.380","AcceptedAnswerId":null,"OwnerUserId":1609464.0,"Title":"PyCharm go to implementation of a method - what does this mean?","Body":"<p>I have the following python code:<\/p>\n\n<pre><code>import torch\nt = torch.tensor([1, 2, 3])\nprint(t.dtype)\n<\/code><\/pre>\n\n<p>Within PyCharm, I select the <code>dtype<\/code> in the last line and go to its implementation:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3kN4e.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3kN4e.png\" alt=\"torch version implementation\"><\/a><\/p>\n\n<p>It brings me to this <code>_TensorBase.py<\/code> file, with the beginning that looks like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Q2XPv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Q2XPv.png\" alt=\"_TensorBase.py\"><\/a><\/p>\n\n<p>and the line with <code>dtype<\/code> showing:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OpbB1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OpbB1.png\" alt=\"dtype line\"><\/a><\/p>\n\n<p>Now, this file actually doesn't exist in the pytorch package.  Could someone explain to me what exactly is going when I ask PyCharm to go to the implementation of dtype?  From looking at this, I'm guessing part of the source code is implemented in C and maybe I'm just looking at the interface generated by PyCharm, which explains why the file doesn't actually exist and none of the implementation exists.<\/p>\n","answers":[]}
+{"QuestionId":52827727,"AnswerCount":0,"Tags":"<python-3.x><numpy><pytorch><torch>","CreationDate":"2018-10-16T03:52:34.833","AcceptedAnswerId":null,"OwnerUserId":7282956.0,"Title":"TypeError: join() argument must be str or bytes, not 'BufferedReader'","Body":"<pre><code>(Pdb) import os\n(Pdb) import numpy as np\n(Pdb) from tempfile import mkdtemp\n(Pdb) import os.path as path\n\n(Pdb) tensor\ntensor([     0,      0,  20480,  ...,      0,      0, -16400], dtype=torch.int16)\n(Pdb) tensor.dtype\ntorch.int16\n\n(Pdb) filename = path.join(mkdtemp(), f)\n*** TypeError: join() argument must be str or bytes, not 'BufferedReader' \n\n\n(Pdb) import torch\n\n(Pdb) fp = np.memmp(filename, dtype='int16')\n(Pdb) fp\nmemmap([16, 12, 12, ..., 44, 44, 44], dtype=int16)\n(Pdb) \n<\/code><\/pre>\n\n<p>I am trying to create a memory-map to an array stored in a binary file on disk. But I am getting <code>*** TypeError: join() argument must be str or bytes, not 'BufferedReader'<\/code>  error. Even after the error the memory map is working but I want it free from error. Please help me out. Thanks<\/p>\n","answers":[]}
+{"QuestionId":52828090,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-16T04:48:15.667","AcceptedAnswerId":52828405.0,"OwnerUserId":8292871.0,"Title":"Tensorflow Keras output style","Body":"<p>I have encounter a strange thing while doing a dummy model in Keras. For reasons that are not important now, I decided to try to train a set of weights to become the identity matrix. My code was the following:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport numpy as np\n\ntfe = tf.contrib.eager\ntf.enable_eager_execution()\ni4 = np.eye(4)\ninds = np.random.randint(0,4,size=2000)\ndata = i4[inds]\nmodel = keras.Sequential([keras.layers.Dense(4, kernel_regularizer= \n                         keras.regularizers.l2(.001), kernel_initializer='zeros')])\nmodel.compile(optimizer=tf.train.AdamOptimizer(.001), loss= 'mse',  metrics = ['accuracy'])\nmodel.fit(data,inds, epochs=50)\n<\/code><\/pre>\n\n<p>this did horribly on what should be a <em>very<\/em> simple task.I changed the last line to<\/p>\n\n<pre><code>model.fit(data, data, epochs =50)\n<\/code><\/pre>\n\n<p>which I think essentially means I am feeding the labels as one hot vectors. With this line, the training did exactly what I wanted it to on this very simple task. So, my questions are:<\/p>\n\n<ul>\n<li>Why would this not work with the first line and work with the second?<\/li>\n<li>What do I need to do to be able to feed the output to keras not as one hot vectors? Not that I mind converting. It's just that some of the examples I've seen - even MNIST - don't seem to convert their labels to one hots before feeding them in. What's the issue here? Is keras trying convert the numerical\/other labels I've given it in a way I don't expect? If so, how does it convert such labels so I can predict the response correctly?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"52828405","CreationDate":"2018-10-16T05:21:16.567","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>The model you used is trying to minimize the mean square error. Thus, it is obvious that the second line is the way to go:<\/p>\n\n<pre><code>model.fit(data, data, epochs=50)\n<\/code><\/pre>\n\n<p>because to learn the identity matrix, we should have: <code>x =y<\/code>, and thus data is both the inputs and outputs. <\/p>\n\n<p><strong>Why this does not work:<\/strong><\/p>\n\n<pre><code>model.fit(data, inds, epochs=50)\n<\/code><\/pre>\n\n<p>Well, in this case your network output is of size 4 (dense layer), but you give it outputs of size 1 (inds). You should get an error...<\/p>\n\n<p><strong>How to do it without using one hot vectors for the output vectors<\/strong>:<\/p>\n\n<p>One way is to use the sparse categorical crossentropy loss instead as such: <\/p>\n\n<pre><code>i4 = np.eye(4)\ninds = np.random.randint(0,4,size=32)\ndata = i4[inds]\n\nmodel = keras.Sequential([keras.layers.Dense(4, kernel_initializer='zeros', activation='softmax')])\nmodel.compile(optimizer=tf.train.AdamOptimizer(.001), loss= 'sparse_categorical_crossentropy',  metrics = ['accuracy'])\nmodel.fit(data, inds, epochs=50)\n<\/code><\/pre>\n\n<p>and then you will see that the model will fit the <code>inds<\/code> very accurately: <\/p>\n\n<pre><code>In [4]: np.argmax(model.predict(data), axis=1)\nOut[4]: \narray([3, 1, 1, 3, 0, 3, 2, 0, 2, 1, 0, 2, 0, 0, 1, 2, 3, 2, 3, 0, 3, 2,\n       1, 2, 3, 3, 3, 1, 0, 1, 2, 0])\n\nIn [5]: inds\nOut[5]: \narray([3, 1, 1, 3, 0, 3, 2, 0, 2, 1, 0, 2, 0, 0, 1, 2, 3, 2, 3, 0, 3, 2,\n       1, 2, 3, 3, 3, 1, 0, 1, 2, 0])\n<\/code><\/pre>\n\n<p>and the train accuracy :<\/p>\n\n<pre><code>In [6]: np.mean(np.argmax(model.predict(data), axis=1) == inds)\nOut[6]: 1.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":52828484,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T05:29:46.020","AcceptedAnswerId":null,"OwnerUserId":10510894.0,"Title":"Tensorflow in Windows installation error: No module named '_pywrap_tensorflow_internal'","Body":"<p>after watching few videos on Youtube I though Tensorflow is such a cool kit. Having just a tiny bit of experience with Python, naturally I tried to install it get my hands dirty. Usually when I try to install some add-on for Python I just simply do:\n    python -m pip install <\/p>\n\n<p>(by the way I'm running Win10 on a laptop with i5-8250U &amp; GeForce MX150)\nbut this method gives me no luck with Tensorflow. I then followed suggestion <a href=\"https:\/\/stackoverflow.com\/questions\/38896424\/tensorflow-not-found-using-pip\">here<\/a>, only changing <\/p>\n\n<pre><code>python3 -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>python3 -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-1.11.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p>and it magically installed. But now when I try <\/p>\n\n<pre><code>import tensorflow\n<\/code><\/pre>\n\n<p>I get these nasty errors:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 18, in swig_import_helper\n    fp, pathname, description = imp.find_module('_pywrap_tensorflow_internal', [dirname(__file__)])\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\imp.py\", line 297, in find_module\n    raise ImportError(_ERR_MSG.format(name), name=name)\nImportError: No module named '_pywrap_tensorflow_internal'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 20, in swig_import_helper\n    import _pywrap_tensorflow_internal\nModuleNotFoundError: No module named '_pywrap_tensorflow_internal'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 18, in swig_import_helper\n    fp, pathname, description = imp.find_module('_pywrap_tensorflow_internal', [dirname(__file__)])\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\imp.py\", line 297, in find_module\n    raise ImportError(_ERR_MSG.format(name), name=name)\nImportError: No module named '_pywrap_tensorflow_internal'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\Mee\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 20, in swig_import_helper\n    import _pywrap_tensorflow_internal\nModuleNotFoundError: No module named '_pywrap_tensorflow_internal'\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n\n<p>Would anyone who experienced similar error when importing tf for the first time give me a hand? Much appreciated.  <\/p>\n\n<p>Update: thanks to all who commented, problem solved! I just had to downgrade from Python 3.7 to 3.6. It's all works fine now.<\/p>\n","answers":[]}
+{"QuestionId":52828527,"AnswerCount":0,"Tags":"<deep-learning><probability><pytorch>","CreationDate":"2018-10-16T05:33:44.933","AcceptedAnswerId":null,"OwnerUserId":10510910.0,"Title":"How to get probabilities when using Pytorch's densenet?","Body":"<p>I want to do a binary classification and I used the <a href=\"https:\/\/pytorch.org\/docs\/0.4.0\/torchvision\/models.html#id5\" rel=\"nofollow noreferrer\">DenseNet<\/a> from Pytorch.<\/p>\n\n<p>Here is my predict code:<\/p>\n\n<pre><code>densenet = torch.load(model_path)\ndensenet.eval()\noutput = densenet(input)\nprint(output)\n<\/code><\/pre>\n\n<p>And here is the output:<\/p>\n\n<pre><code>Variable containing:\n54.4869 -54.3721\n[torch.cuda.FloatTensor of size 1x2 (GPU 0)]\n<\/code><\/pre>\n\n<p>I want to get the probabilities of each class. What should I do? <\/p>\n\n<p>I have noticed that <code>torch.nn.Softmax()<\/code> could be used when there are many categories, as discussed <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-extract-probabilities\/2720\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[]}
+{"QuestionId":52828968,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><machine-learning><multi-gpu>","CreationDate":"2018-10-16T06:13:48.147","AcceptedAnswerId":null,"OwnerUserId":8906149.0,"Title":"Multiple gpus (1080Ti) do not speed up training in tensorflow, test on cifar10_estimator code","Body":"<p>I was trying to test the performance of multi GPUs version of <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10_estimator\" rel=\"nofollow noreferrer\">cifar10_estimator<\/a> on 2 or 3 1080Ti, but received no speed-up.<\/p>\n\n<p>I find some useful information about hardware <a href=\"https:\/\/github.com\/rossumai\/keras-multi-gpu\/blob\/master\/blog\/docs\/hardware.md\" rel=\"nofollow noreferrer\">here<\/a>, but still confused how to solve it.<\/p>\n\n<p>My environment:<\/p>\n\n<ul>\n<li>Ubuntu VERSION=16.04.5 LTS (Xenial Xerus)<\/li>\n<li>Python3<\/li>\n<li>CUDA_VERSION=9.0.176<\/li>\n<li>tensorflow-gpu=1.11.0<\/li>\n<\/ul>\n\n<p>GPU info:<\/p>\n\n<pre><code>nvidia-smi topo -m\n\n    GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity\nGPU0     X  PIX PHB PHB SYS SYS SYS SYS 0-7\nGPU1    PIX  X  PHB PHB SYS SYS SYS SYS 0-7\nGPU2    PHB PHB  X  PIX SYS SYS SYS SYS 0-7\nGPU3    PHB PHB PIX  X  SYS SYS SYS SYS 0-7\nGPU4    SYS SYS SYS SYS  X  PIX PHB PHB 8-15\nGPU5    SYS SYS SYS SYS PIX  X  PHB PHB 8-15\nGPU6    SYS SYS SYS SYS PHB PHB  X  PIX 8-15\nGPU7    SYS SYS SYS SYS PHB PHB PIX  X  8-15\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI\/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe switches (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing a single PCIe switch\n  NV#  = Connection traversing a bonded set of # NVLinks\n<\/code><\/pre>\n\n<p>1 gpu bach_size=128<\/p>\n\n<pre><code>INFO:tensorflow:loss = 2.2576141, step = 200 (3.729 sec)\nINFO:tensorflow:learning_rate = 0.1, loss = 2.2576141 (3.729 sec)\nINFO:tensorflow:Average examples\/sec: 2821.06 (2858.65), step = 200\nINFO:tensorflow:Average examples\/sec: 2847.23 (3496.06), step = 210\nINFO:tensorflow:Average examples\/sec: 2857.91 (3102.29), step = 220\nINFO:tensorflow:Average examples\/sec: 2867.04 (3083.62), step = 230\nINFO:tensorflow:Average examples\/sec: 2889.21 (3514.15), step = 240\nINFO:tensorflow:Average examples\/sec: 2913.15 (3636.28), step = 250\nINFO:tensorflow:Average examples\/sec: 2915.99 (2988.94), step = 260\nINFO:tensorflow:Average examples\/sec: 2901.94 (2578.95), step = 270\nINFO:tensorflow:Average examples\/sec: 2888.87 (2575.46), step = 280\nINFO:tensorflow:Average examples\/sec: 2892.13 (2986.66), step = 290\nINFO:tensorflow:global_step\/sec: 24.25\n<\/code><\/pre>\n\n<p>2 gpu bach_size=256 <\/p>\n\n<pre><code>INFO:tensorflow:loss = 2.4630964, step = 200 (5.971 sec)\nINFO:tensorflow:learning_rate = 0.1, loss = 2.4630964 (5.971 sec)\nINFO:tensorflow:Average examples\/sec: 3255.68 (4296.71), step = 200\nINFO:tensorflow:Average examples\/sec: 3297.51 (4437.93), step = 210\nINFO:tensorflow:Average examples\/sec: 3332.15 (4275.33), step = 220\nINFO:tensorflow:Average examples\/sec: 3363.86 (4254.65), step = 230\nINFO:tensorflow:Average examples\/sec: 3395.09 (4316.94), step = 240\nINFO:tensorflow:Average examples\/sec: 3418.44 (4094.23), step = 250\nINFO:tensorflow:Average examples\/sec: 3447.17 (4364.24), step = 260\nINFO:tensorflow:Average examples\/sec: 3474.56 (4379.02), step = 270\nINFO:tensorflow:Average examples\/sec: 3492.73 (4067.13), step = 280\nINFO:tensorflow:Average examples\/sec: 3514.19 (4244.23), step = 290\nINFO:tensorflow:global_step\/sec: 16.6026\n<\/code><\/pre>\n\n<p>3 gpu bach_size=384<\/p>\n\n<pre><code>INFO:tensorflow:loss = 2.0980535, step = 200 (9.329 sec)\nINFO:tensorflow:learning_rate = 0.1, loss = 2.0980535 (9.329 sec)\nINFO:tensorflow:Average examples\/sec: 3214.65 (4165.7), step = 200\nINFO:tensorflow:Average examples\/sec: 3272.85 (5130.99), step = 210\nINFO:tensorflow:Average examples\/sec: 3324.15 (4955.13), step = 220\nINFO:tensorflow:Average examples\/sec: 3376.65 (5174.76), step = 230\nINFO:tensorflow:Average examples\/sec: 3425.48 (5132.15), step = 240\nINFO:tensorflow:Average examples\/sec: 3468.29 (4954.35), step = 250\nINFO:tensorflow:Average examples\/sec: 3509.91 (5014.23), step = 260\nINFO:tensorflow:Average examples\/sec: 3544.29 (4755.56), step = 270\nINFO:tensorflow:Average examples\/sec: 3579.69 (4901.39), step = 280\nINFO:tensorflow:Average examples\/sec: 3617.84 (5156.66), step = 290\nINFO:tensorflow:global_step\/sec: 13.1009\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yqIZX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yqIZX.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52867929","CreationDate":"2018-10-18T06:08:13.853","ParentId":null,"OwnerUserId":"8906149","Title":null,"Body":"<p>I think I can answer my question now. If I want a higher performance for multiple gpus, I should check out <a href=\"https:\/\/github.com\/tensorflow\/benchmarks\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/benchmarks\/<\/a>. Refer to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23032\" rel=\"nofollow noreferrer\">this issue<\/a> for my test results on <a href=\"https:\/\/github.com\/tensorflow\/benchmarks\/tree\/master\/scripts\/tf_cnn_benchmarks\" rel=\"nofollow noreferrer\">tf_cnn_benchmarks<\/a>.  <\/p>\n"}]}
+{"QuestionId":52829787,"AnswerCount":2,"Tags":"<python><tensorflow><keras><generator>","CreationDate":"2018-10-16T07:10:21.780","AcceptedAnswerId":52893337.0,"OwnerUserId":8113093.0,"Title":"Keras Custom generator issue when evaluating the model","Body":"<p>I am training a CNN LSTM model using Keras, and after the training was done, I tried to evaluate the model on the testing data like I did when I fine-tuned my CNN, however an error appears this time.<\/p>\n\n<p>After training was done, I tried to following piece of code to evaluate on my testing set:<\/p>\n\n<pre><code>x, y = zip(*(testgenerator[i] for i in range(len(testgenerator))))\n\nx_test, y_test = np.vstack(x), np.vstack(y)\n\nloss, acc = Bi_LSTM.evaluate(x_test, y_test, batch_size=9)\n\nprint(\"Accuracy: \" ,acc)\nprint(\"Loss: \", loss)\n<\/code><\/pre>\n\n<p>I have used this code before to evaluate my fine tuned model and it had no issue, but now I get the following error:<\/p>\n\n<pre><code>TypeError: object of type 'generator' has no len()\n<\/code><\/pre>\n\n<p>I have tried few solutions online like using len(list(generator)) but it did not work. Is it because I am using a custom generator? How can I do to evaluate model in this case ?<\/p>\n","answers":[{"AnswerId":"52831914","CreationDate":"2018-10-16T09:11:10.977","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I think this line is the problem<\/p>\n\n<pre><code>x, y = zip(*(testgenerator[i] for i in range(len(testgenerator))))\n<\/code><\/pre>\n\n<p>because you call <code>len<\/code> on generator object. \nThe solution may be if you just create some counter, increment it and use it as index in <code>testgenerator[i]<\/code><\/p>\n"},{"AnswerId":"52893337","CreationDate":"2018-10-19T13:27:13.653","ParentId":null,"OwnerUserId":"8113093","Title":null,"Body":"<p>The way I solved this is by using a different method. In this case I do not need to extract values for x,y:<\/p>\n\n<pre><code>loss, acc = Bi_LSTM.evaluate_generator(testgenerator, batch_size=9)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52829803,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-10-16T07:11:30.223","AcceptedAnswerId":null,"OwnerUserId":988709.0,"Title":"learning rate of adam optimizer do not change","Body":"<p>I want to log the learning rate of adam optimizer with estimator of tensorflow like this:<\/p>\n\n<pre><code>def def model_fn(features, labels, mode):\n    ...\n    optimizer = tf.train.AdamOptimizer(learning_rate=0.1)\n    log_hook = tf.train.LoggingTensorHook({\"lr\" : optimizer._lr_t}, every_n_iter=10)\n    return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op, training_hooks=[log_hook]) \n    ...\n<\/code><\/pre>\n\n<p>We know that the learning rate of tf.train.AdamOptimizer decays itself. But my result is always 1.0 like this:<\/p>\n\n<pre><code>INFO:tensorflow:lr = 0.1 (4.537 sec)\nINFO:tensorflow:global_step\/sec: 2.18827\nINFO:tensorflow:loss = 8.285036e-07, step = 16180 (4.570 sec)\nINFO:tensorflow:lr = 0.1 (4.570 sec)\nINFO:tensorflow:global_step\/sec: 2.21156\nINFO:tensorflow:loss = 8.225431e-07, step = 16190 (4.521 sec)\nINFO:tensorflow:lr = 0.1 (4.521 sec)\n<\/code><\/pre>\n\n<p>Am I do the right way for log learning rate of AdamOptimizer?<\/p>\n\n<p><strong>Update:<\/strong>\nI log the optimizer._lr referenced <a href=\"https:\/\/stackoverflow.com\/a\/43953246\/988709\">this answer<\/a>, but got this error:<\/p>\n\n<pre><code>ValueError: Passed 0.1 should have graph attribute that is equal to current graph &lt;tensorflow.python.framework.ops.Graph object at 0x7f96a290a350&gt;.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52830223,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-16T07:36:54.693","AcceptedAnswerId":52830534.0,"OwnerUserId":3392589.0,"Title":"Attempting to use uninitialized value error even if I initialize local and global variables","Body":"<p>I am trying to use tensorflow metrics to find recall and\/or precision, here is my code:<\/p>\n\n<pre><code>    print(\"Starting session...\")\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        sess.run(tf.local_variables_initializer())\n        sess.run(train_init_op)\n        tot_acc = 0\n        print(\"Starting training...\")\n        for i in range(num_epochs):\n            sample_counter = 0\n            l, _, acc = sess.run([loss, optimizer, accuracy], feed_dict={keep_prob: 0.7})\n            tot_acc += acc\n            print(\"Epoch: {}, loss: {:.3f}, training accuracy: {:.2f}%\".format(i, l, acc * 100))\n        print(\"Average training set accuracy over {} epoch is {:.2f}%\".format(num_epochs,\n                                                                                    (tot_acc \/ num_epochs) * 100))\n\n        print(\"Starting Validation...\")\n        sess.run(valid_init_op)\n        tot_acc = 0\n        for i in range(valid_iters):\n            acc = sess.run([accuracy], feed_dict={keep_prob: 1.0})\n            tot_acc += acc[0]\n            print(\"Iter: {}, validation accuracy: {:.2f}%\".format(i, acc[0] * 100))\n        print(\"Average validation set accuracy over {} iterations is {:.2f}%\".format(valid_iters,\n                                                                           (tot_acc \/ valid_iters) * 100))\n        sess.run(valid_init_op)\n        val_img, val_label = next_element\n        finalprediction = tf.argmax(train_predict, 1)\n        actualprediction = tf.argmax(val_label, 1)\n        confusion = tf.confusion_matrix(labels=actualprediction, predictions=finalprediction, \n                                        num_classes=num_classes, dtype=tf.int32, name=\"Confusion_Matrix\")\n        recall = tf.metrics.recall(labels=actualprediction, predictions=finalprediction, name=\"Recall\")\n\n\n        cm = np.zeros([2,2], dtype=int)\n        rc = np.zeros([1,2], dtype=int)\n        for i in range(valid_iters):\n            while True:\n                try:\n                    conf_matrix = sess.run(confusion, feed_dict={keep_prob: 1.0})\n                    rec = sess.run(recall, feed_dict={keep_prob: 1.0})\n                    cm += conf_matrix\n                    rc += rec\n                except tf.errors.OutOfRangeError:\n                    print(\"End of append.\")\n                break\n        print(\"confusion matrix: \", cm)\n        print(\"Recall: \", rc)\n<\/code><\/pre>\n\n<p>but every time I get <code>Attempting to use uninitialized value<\/code>, after doing some googling I found some answers in here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/44624648\/tensorflow-attempting-to-use-uninitialized-value-in-variable-initialization\">TensorFlow: \u201cAttempting to use uninitialized value\u201d in variable initialization<\/a> \nand here: \n<a href=\"https:\/\/stackoverflow.com\/questions\/48558052\/tensorflow-variable-initialization-error-attempting-to-use-uninitialized-value\">tensorflow Variable Initialization error : Attempting to use uninitialized value Variable<\/a>\nI did the same as they suggest but none of that solved my problem? here is the error code: <\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value Recall_3\/true_positives\/count\n     [[Node: Recall_3\/true_positives\/count\/read = Identity[T=DT_FLOAT, _class=[\"loc:@Recall_3\/true_positives\/AssignAdd\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Recall_3\/true_positives\/count)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52830534","CreationDate":"2018-10-16T07:55:39.990","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>Your code is not runnable (it misses some variables\/operations), but as explained in the answers you quote, you need to first define your graph before initializing the variables in it. You might find <a href=\"https:\/\/www.tensorflow.org\/guide\/variables\" rel=\"nofollow noreferrer\">this part of the official doc<\/a> useful.<\/p>\n\n<pre><code>    val_img, val_label = next_element\n    finalprediction = tf.argmax(train_predict, 1)\n    actualprediction = tf.argmax(val_label, 1)\n    confusion = tf.confusion_matrix(labels=actualprediction, predictions=finalprediction, \n                                    num_classes=num_classes, dtype=tf.int32, name=\"Confusion_Matrix\")\n    recall = tf.metrics.recall(labels=actualprediction, predictions=finalprediction, name=\"Recall\")\n\n\nprint(\"Starting session...\")\nwith tf.Session() as sess:\n    # Notice that the call to the variables initializer is made after the call to\n    # to tf.metrics.recall and therefore after the recall variables have been added to the graph.\n\n\n    sess.run(tf.global_variables_initializer())\n    sess.run(tf.local_variables_initializer())\n    sess.run(train_init_op)\n    tot_acc = 0\n    print(\"Starting training...\")\n    for i in range(num_epochs):\n        sample_counter = 0\n        l, _, acc = sess.run([loss, optimizer, accuracy], feed_dict={keep_prob: 0.7})\n        tot_acc += acc\n        print(\"Epoch: {}, loss: {:.3f}, training accuracy: {:.2f}%\".format(i, l, acc * 100))\n    print(\"Average training set accuracy over {} epoch is {:.2f}%\".format(num_epochs,\n                                                                                (tot_acc \/ num_epochs) * 100))\n\n    print(\"Starting Validation...\")\n    sess.run(valid_init_op)\n    tot_acc = 0\n    for i in range(valid_iters):\n        acc = sess.run([accuracy], feed_dict={keep_prob: 1.0})\n        tot_acc += acc[0]\n        print(\"Iter: {}, validation accuracy: {:.2f}%\".format(i, acc[0] * 100))\n    print(\"Average validation set accuracy over {} iterations is {:.2f}%\".format(valid_iters,\n                                                                       (tot_acc \/ valid_iters) * 100))\n    sess.run(valid_init_op)\n\n    cm = np.zeros([2,2], dtype=int)\n    rc = np.zeros([1,2], dtype=int)\n    for i in range(valid_iters):\n        while True:\n            try:\n                conf_matrix = sess.run(confusion, feed_dict={keep_prob: 1.0})\n                rec = sess.run(recall, feed_dict={keep_prob: 1.0})\n                cm += conf_matrix\n                rc += rec\n            except tf.errors.OutOfRangeError:\n                print(\"End of append.\")\n            break\n    print(\"confusion matrix: \", cm)\n    print(\"Recall: \", rc)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52830369,"AnswerCount":0,"Tags":"<c><api><tensorflow>","CreationDate":"2018-10-16T07:46:02.137","AcceptedAnswerId":null,"OwnerUserId":6522386.0,"Title":"Tensorflow C API: SessionRun batch size (how to properly set)","Body":"<p>I am struggling to find the correct way to properly set the batch size for a SessionRun call in the C API.\nWhat I am trying to do is passing multiple images in a single SessionRun to a CNN (VGG16).\nI am wondering which is the correct way to do so and to retrieve the appropriate correlated output values.\nI have also come to believe that a single SessionRun call can only accept a single image as an input, so a proper batched input could only be achieved by having multiple SessionRun calls (for an example inside a \"for\" loop) with a single image each. Am I correct? Is there not a way to batch multiple images inside a single SessionRun call for a single-input CNN?<\/p>\n\n<p>I have searched thoroughly Tensorflow C API documentation but I could not find any evidence of a proper indication on how to do it.<\/p>\n\n<p>I have tried these pseudo-code: (calling these code SISO = Single Input Single Output, and MIMO = Multiple Input Multiple Output, the input is batch_size * 50 * 50 * 3, in SISO batch_size = 1, in MIMO batch_size = 3; remember that I am using VGG16 graph)<\/p>\n\n<p>SISO:<\/p>\n\n<pre><code>loadGraph()\nTF_NewSession()\nTF_AllocateTensor( 1 * 50 * 50 * 3)\n\/\/ copy data into tensor\nnum_bytes = rows() * cols() * channels() * sizeof(float);\nmemcpy(tensor_data, img_data, num_bytes)\nTF_SessionRun()\ndata = static_cast&lt;float*&gt;(TF_TensorData(output_tensor));\n\/\/ data[0] = frame number\n\/\/ data[1] = probabily of prediction\nout = (data[1] &gt; 0.5f) ? true : false;\n<\/code><\/pre>\n\n<p>for SISO, I have always the right output.<\/p>\n\n<p>MIMO:<\/p>\n\n<pre><code>loadGraph()\nTF_NewSession()\nTF_AllocateTensor( 3 * 50 * 50 * 3)\n\/\/ copy data into tensor\nnum_bytes = rows() * cols() * channels() * sizeof(float);\noffset = 0;\nfor(size_t i = 0; i &lt; 3; ++i)\n{\n    memcpy(tensor_data + offset, imgs_data[i], num_bytes);\n    offset += num_bytes;\n}\nTF_SessionRun()\ndata = static_cast&lt;float*&gt;(TF_TensorData(output_tensor));\noffset = 0;\nnum_bytes = 2 * sizeof(float);\n\/\/ data[i] = frame number\n\/\/ data[i+1] = probabily of prediction\nfor(size_t i = 0; i &lt; 3; ++i)\n{\n    out[i] = (*(data + offset + 1) &gt; 0.5f) ? true : false;\n    offset += num_bytes;\n}\n<\/code><\/pre>\n\n<p>for MIMO, only the first classification output value is right, the other ones are random numbers.<\/p>\n\n<p>what is wrong with this code?<\/p>\n","answers":[]}
+{"QuestionId":52831038,"AnswerCount":3,"Tags":"<tensorflow><classification><conv-neural-network>","CreationDate":"2018-10-16T08:24:54.757","AcceptedAnswerId":52831580.0,"OwnerUserId":8128190.0,"Title":"My CNN classifier gives wrong prediction on random images","Body":"<p>I trained my CNN classifier (using tensorflow) with 3 data categories (ID card, passport, bills).<br>\nWhen I test it with images that belong to one of the 3 categories, it gives the right prediction. However, when I test it with a wrong image (a car image for example) it keeps giving me prediction (i.e. it predicts that the car belongs the ID card category).   <\/p>\n\n<p><strong>Is there a way to make it display an error message instead of giving a wrong prediction?<\/strong><\/p>\n","answers":[{"AnswerId":"52831547","CreationDate":"2018-10-16T08:52:17.250","ParentId":null,"OwnerUserId":"9546874","Title":null,"Body":"<p>Easiest way is to simply add a fourth category for anything but the other three and train it with various completely random photos.<\/p>\n"},{"AnswerId":"52831580","CreationDate":"2018-10-16T08:54:20.683","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>This should be tackled differently. This is known as <strong>open set recognition problem<\/strong>. You can google it and find more about it but basically it's this: \nYou cannot train your classifier on every class imaginable. It will always run into some other class that it's not familiar with and that it hasn't already seen before.<\/p>\n\n<p>There are a few solutions from which I will single out the 3 of them:<\/p>\n\n<ol>\n<li><p><strong>Separate binary classifier<\/strong> - You can build separate binary classifier that recognizes images and sorts them in two categories depending on if the bill, passport or ID are in the image or not. If they are, it should let the algorithm you have already build to process the image and classify it into one of the 3 categories. If the first classifier says that some other object is in the image, you can immediately discard the image because it's not the image of bill\/passport\/ID.<\/p><\/li>\n<li><p><strong>Thresholding<\/strong>. In the case when the ID is on the image, probability of the ID is high and probabilities for bill and passport are fairly low. In the case when the image is something else (ex. <em>a car<\/em>), the probabilities are most probably about the same for all 3 classes. In other words, probability for neither of the classes really stand out. That is a situation in which you pick the highest probability of the ones generated and set the output class to be the class of that probability, regardless the value of probability is 0.4 or something like that. To resolve this, you can set a threshold at, let's say 0.7, and say if neither of probabilities is over that threshold, there is something else on the picture (not ID, passport or bill).<\/p><\/li>\n<li><p>Create the fourth class: <strong>Unknown<\/strong>. If you pick this option, you should add few of the other images to the dataset and label them <em>unknown<\/em>. Then train the classifier and see what the result is.<\/p><\/li>\n<\/ol>\n\n<p>I would recommend 1 or 2. Hope it helps :)<\/p>\n"},{"AnswerId":"52831714","CreationDate":"2018-10-16T09:00:28.973","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>This is not really a programming problem, its way more complicated. What you want is called Out of Distribution detection, where the classifier has a way to tell you that the sample is not on the training set.<\/p>\n\n<p>There are recent research papers that deal with this problem, such as <a href=\"https:\/\/arxiv.org\/abs\/1802.04865\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1802.04865<\/a> and <a href=\"https:\/\/arxiv.org\/abs\/1711.09325\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1711.09325<\/a><\/p>\n\n<p>In general you cannot use a model that has not been trained specifically for this, for example, the probabilities produced by a softmax classifier are not calibrated for this purpose, so thresholding these probabilities will not work at all.<\/p>\n"}]}
+{"QuestionId":52831156,"AnswerCount":1,"Tags":"<flask><keras>","CreationDate":"2018-10-16T08:31:03.987","AcceptedAnswerId":52833884.0,"OwnerUserId":9481613.0,"Title":"How do I extract a Keras model from a Jupyter notebook to embed in a Flask web app?","Body":"<p>I want to take my Keras model out of this Jupyter notebook:  <a href=\"https:\/\/github.com\/nyck33\/CNN-image-analyzer\/blob\/master\/dog_app.ipynb\" rel=\"nofollow noreferrer\">my dog breed classifier<\/a><\/p>\n\n<p>and put it in this provided web app:  <a href=\"https:\/\/github.com\/mtobeiyf\/keras-flask-deploy-webapp\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mtobeiyf\/keras-flask-deploy-webapp<\/a><\/p>\n\n<p>I saw this: <a href=\"https:\/\/stackoverflow.com\/questions\/42763094\/how-to-save-final-model-using-keras\">How to save final model using keras?<\/a><\/p>\n\n<p>But I'm wondering if anyone can provide more details, ie. do I save the model and weights separately?  <\/p>\n\n<p>Also if anybody has links to a step-by-step tutorial, etc.  <\/p>\n","answers":[{"AnswerId":"52833884","CreationDate":"2018-10-16T10:57:48.187","ParentId":null,"OwnerUserId":"5164382","Title":null,"Body":"<p>You may follow the following options: <\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.youtube.com\/watch?v=f6Bf3gl4hWY&amp;feature=youtu.be\" rel=\"nofollow noreferrer\">this video<\/a> and <a href=\"https:\/\/github.com\/llSourcell\/how_to_deploy_a_keras_model_to_production\" rel=\"nofollow noreferrer\">the github page<\/a><\/li>\n<li><a href=\"https:\/\/www.pyimagesearch.com\/2018\/02\/05\/deep-learning-production-keras-redis-flask-apache\/\" rel=\"nofollow noreferrer\">this page<\/a><\/li>\n<li><a href=\"https:\/\/www.pyimagesearch.com\/2017\/12\/18\/keras-deep-learning-raspberry-pi\/\" rel=\"nofollow noreferrer\">this page<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":52831235,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-16T08:36:18.500","AcceptedAnswerId":52831346.0,"OwnerUserId":1179925.0,"Title":"How to add constant tensor in Keras?","Body":"<p>What I'm trying to do is to add a constant tensor to output of network:<\/p>\n\n<pre><code>inputs = Input(shape=(config.N_FRAMES_IN_SEQUENCE, config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS))\ncnn = VGG16(include_top=False, weights='imagenet', input_shape=(config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS))\nx = TimeDistributed(cnn)(inputs)\nx = TimeDistributed(Flatten())(x)\nx = LSTM(256)(x)\nx = Dense(config.N_LANDMARKS * 2, activation='linear')(x)\n\nmean_landmarks = np.array(config.MEAN_LANDMARKS, np.float32)\nmean_landmarks = mean_landmarks.flatten()\nmean_landmarks_tf = tf.convert_to_tensor(mean_landmarks)\nx = x + mean_landmarks_tf\n\nmodel = Model(inputs=inputs, outputs=x)\noptimizer = Adadelta()\nmodel.compile(optimizer=optimizer, loss='mae')\n<\/code><\/pre>\n\n<p>But I get error:<\/p>\n\n<pre><code>ValueError: Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata). Found: Tensor(\"add:0\", shape=(?, 136), dtype=float32)\n<\/code><\/pre>\n\n<p>It's trivial in tensorflow, but how to do it in Keras?<\/p>\n","answers":[{"AnswerId":"52831346","CreationDate":"2018-10-16T08:42:25.273","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Seems it can be done with Lamda layer:<\/p>\n\n<pre><code>from keras.layers import Lambda\n\ndef add_mean_landmarks(x):\n    mean_landmarks = np.array(config.MEAN_LANDMARKS, np.float32)\n    mean_landmarks = mean_landmarks.flatten()\n    mean_landmarks_tf = tf.convert_to_tensor(mean_landmarks)\n    x = x + mean_landmarks_tf\n    return x\n\n\nx = Lambda(add_mean_landmarks)(x)\n<\/code><\/pre>\n"},{"AnswerId":"52831446","CreationDate":"2018-10-16T08:46:53.377","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>In addition to your own answer, I would heavily favor a native implementation of addition, as for example <a href=\"https:\/\/keras.io\/layers\/merge\/#add\" rel=\"nofollow noreferrer\">Keras provides <code>Keras.layers.Add<\/code><\/a>.<\/p>\n\n<p>The reason for this is that I am unsure how your own lambda function is pushed down to the lower layers: Essentially, in TensorFlow (or whatever other backend you use), the internal operations make use of the heavily optimized computational graph, whereas custom operations tend to be translated to a heavier-weight (or, in the worst case, bloated) low level execution.<\/p>\n\n<p>The correct way to do it with <code>Keras.layers.Add<\/code> would be simply doing<\/p>\n\n<pre><code>x = keras.layers.Add()([x, add_mean_landmarks])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52831351,"AnswerCount":3,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-16T08:42:42.257","AcceptedAnswerId":52960797.0,"OwnerUserId":1115237.0,"Title":"How to determine an overfitted model based on loss precision and recall","Body":"<p>I've written an LSTM network with Keras (following code):<\/p>\n\n<pre><code>    df = pd.read_csv(\"..\/data\/training_data.csv\")\n\n    # Group by and pivot the data\n    group_index = df.groupby('group').cumcount()\n    data = (df.set_index(['group', group_index])\n            .unstack(fill_value=0).stack())\n\n    # getting np array of the data and labeling\n    # on the label group we take the first label because it is the same for all\n    target = np.array(data['label'].groupby(level=0).apply(lambda x: [x.values[0]]).tolist())\n    data = data.loc[:, data.columns != 'label']\n    data = np.array(data.groupby(level=0).apply(lambda x: x.values.tolist()).tolist())\n\n    # shuffel the training set\n    data, target = shuffle(data, target)\n\n    # spilt data to train and test\n    x_train, x_test, y_train, y_test = train_test_split(data, target, test_size=0.2, random_state=4)\n\n    # ADAM Optimizer with learning rate decay\n    opt = optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)\n\n    # build the model\n    model = Sequential()\n\n    num_features = data.shape[2]\n    num_samples = data.shape[1]\n\n    model.add(LSTM(8, batch_input_shape=(None, num_samples, num_features), return_sequences=True, activation='sigmoid'))\n    model.add(LeakyReLU(alpha=.001))\n    model.add(Dropout(0.2))\n    model.add(LSTM(4, return_sequences=True, activation='sigmoid'))\n    model.add(LeakyReLU(alpha=.001))\n    model.add(Flatten())\n    model.add(Dense(1, activation='sigmoid'))\n\n    model.compile(loss='binary_crossentropy', optimizer=opt,\n                  metrics=['accuracy', keras_metrics.precision(), keras_metrics.recall(),f1])\n\n    model.summary()\n\n\n    # Training, getting the results history for plotting\n    history = model.fit(x_train, y_train, epochs=3000, validation_data=(x_test, y_test))\n<\/code><\/pre>\n\n<p>The monitored metrics are loss, accuracy, precision, recall and f1 score.<\/p>\n\n<p>I've noticed that the validation loss metric start to climb around 300 epochs, so I've figured <strong>overfitting!<\/strong> however, recall is still climbing and precision is slightly improving.  <\/p>\n\n<hr>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AkZ5M.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AkZ5M.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/qNNIz.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qNNIz.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/GYLeO.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GYLeO.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<hr>\n\n<p>Why is that? is my model overfitted?<\/p>\n","answers":[{"AnswerId":"52960797","CreationDate":"2018-10-24T03:46:01.213","ParentId":null,"OwnerUserId":"890242","Title":null,"Body":"<blockquote>\n  <p>the validation loss metric start to climb around 300 epochs (...) recall is still climbing and precision is slightly improving. (...) Why is that?<\/p>\n<\/blockquote>\n\n<p>Precision and recall are measures of how well your classifier performs in terms of the predicted class labels. Model loss on the other hand is a measure of the <a href=\"https:\/\/ml-cheatsheet.readthedocs.io\/en\/latest\/loss_functions.html#cross-entropy\" rel=\"nofollow noreferrer\">cross entropy<\/a>, the error in classification probability : <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PfCRx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PfCRx.png\" alt=\"![![![enter image description here\"><\/a><\/p>\n\n<p>where <\/p>\n\n<pre><code>y = predicted label\np = probability of predicted label\n<\/code><\/pre>\n\n<p>For example, the (softmax) outputs of the model for one observation may look like this for different epochs, say<\/p>\n\n<pre><code># epoch 300\ny = [0.1, 0.9] =&gt; argmax(y) =&gt; 1 (class label 1)\nloss = -(1 * log(0.9)) = 0.10\n\n# epoch 500\ny = [0.4, 0.6] =&gt; argmax(y) =&gt; 1 (class label 1)\nloss = -(1 * log(0.6)) = 0.51\n<\/code><\/pre>\n\n<p>In both cases the precision and recall metrics will stay unchanged (the class label is still predicted correctly), however the model loss has increased. In general terms, the model has become \"less sure\" about it's prediction, but it is still correct.<\/p>\n\n<p><em>Note in your model the loss is calculated for all observations, not just a single one. I limit the discussion for simplicity. The loss formula is trivially expanded to n > 1 observations by taking the average of the loss of all observations.<\/em><\/p>\n\n<blockquote>\n  <p>is my model overfitted?<\/p>\n<\/blockquote>\n\n<p>In order to determine this, you have to compare training loss and validation loss. You cannot tell by validation loss alone. If training loss decreases and validation loss increases, your model is overfitting.  <\/p>\n"},{"AnswerId":"52960410","CreationDate":"2018-10-24T02:50:21.680","ParentId":null,"OwnerUserId":"6347629","Title":null,"Body":"<p>AS @Matthieu mentioned, it could be biased to look at precision and recall of one class alone. May be we have to look at performance on other class as well. <\/p>\n\n<p>Better measure could be concordance (auc of roc), in case of a binary classification. Concordance measure the goodness of the model to rank-order the datapoints based on its likeliness towards to a class. <\/p>\n\n<p>One more option is Macro\/Micro-precision\/recall to get the complete picture of the model performance.<\/p>\n"},{"AnswerId":"52947803","CreationDate":"2018-10-23T11:23:10.127","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>Indeed, if the validation loss starts growing up again, then you may want to stop early. It's a \"standard\" approach, named \"early stopping\" (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Early_stopping\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Early_stopping<\/a>). Clearly, if the loss for your validation and data is increasing, then the model is not doing as great as it could, it is overfitting. <\/p>\n\n<p>Precision and recall are not enough, they can increase if your model is giving more positive results, less negative ones (for instance 9 positives for 1 negative). Then these ratios can seem to be improved, but it's just that you have less true negatives.<\/p>\n\n<p>All these two put together can help shed some light as to what is happening here. The good answers may still be good ones, but with a lower quality (the loss for individual samples increases on average, but still keeps good answers good), and there could be a shift from bad answers to good answers with a bias (true negatives are transformed into false positives).<\/p>\n"}]}
+{"QuestionId":52831366,"AnswerCount":1,"Tags":"<python-3.x><ubuntu><tensorflow><cuda><sudo>","CreationDate":"2018-10-16T08:43:31.477","AcceptedAnswerId":null,"OwnerUserId":8276928.0,"Title":"Ubuntu python import tensorflow error with sudo","Body":"<p>I installed cuda, cudnn and tensorflow-gpu successfully,\nI run <\/p>\n\n<p><code>python3 \nimport tensorflow\n<\/code><\/p>\n\n<p>this worked fine, but I can't import tensorflow with sudo:<\/p>\n\n<p><code>sudo python3\nimport tensorflow\n<\/code>\nI got :\n<code>\nImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory\n<\/code>\nSo what's the problem and how to deal with it? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"52832590","CreationDate":"2018-10-16T09:46:26.073","ParentId":null,"OwnerUserId":"8276928","Title":null,"Body":"<p>I found the environment variables of common command was different with environment variablese of sudo, and I fix this problem by adding CUDA paths to the sudo Environment variables:<\/p>\n\n<pre><code>alias sudo='sudo env PATH=$PATH LD_LIBRARY_PATH=$LD_LIBRARY_PATH'\n<\/code><\/pre>\n\n<p>add this line to the end of <code>~\/.bashrc<\/code>. then <code>source ~\/.bashrc<\/code><\/p>\n"}]}
+{"QuestionId":52831410,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-16T08:45:12.770","AcceptedAnswerId":null,"OwnerUserId":7104060.0,"Title":"Naming layers in keras","Body":"<p>I am using a pre-trained keras model ( Convolutional network) and I am retraining this model again on my dataset.\nNow, I need to get the output of some layers, to visualize the gradient activation. I just found out that every trained model has different naming of layers. for example, the input layer in one model is: <code>input_7 (InputLayer)<\/code> and in another model is <code>input_5 (InputLayer)<\/code>.<\/p>\n\n<p>Do you know how to prevent this bad behavior? How can I keep the same naming without the need to manually name all the layers, as I have more than 53 convolutional layers?<\/p>\n","answers":[]}
+{"QuestionId":52831740,"AnswerCount":1,"Tags":"<tensorflow><proxy><bazel>","CreationDate":"2018-10-16T09:01:46.413","AcceptedAnswerId":null,"OwnerUserId":10511832.0,"Title":"[tensorflow],[bazel] Compile Tensorflow behind a HTTP\/HTTPS Proxy - HTTP 502 Bad gateway","Body":"<p>I know that this is no new questions, but I get lots of troubles when trying to compile tensorflow from the source. It seems like this is a \"bazel behind proxy\" matter. <\/p>\n\n<ul>\n<li>I compile on <em>Fedora28<\/em><\/li>\n<li>I am using <em>bazel 0.17.2<\/em> (installed as a rpm from vbatts-bazel yum repository)<\/li>\n<li>I try to compile the current <code>master<\/code> from github.com\/tensorflow\/tensorflow.git<\/li>\n<\/ul>\n\n<p>I work behind a HTTP\/HTTPS proxy. The environment has http_proxy\/https_proxy set (via <code>\/etc\/environment<\/code>, so everyone get the envvar). For tests I have set envvar in lowercase and uppercase. Using git\/https, yum or wget through this proxy is OK since those envvar are set.<\/p>\n\n<p>When compiling tensorflow (after the <code>.\/configure<\/code> step), I get an explicit HTTP error:<\/p>\n\n<pre><code>[user22@vm0 tensorflow]$ bazel build   --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package\nERROR: error loading package '': Encountered error while reading extension file 'closure\/defs.bzl': no such package '@io_bazel_rules_closure\/\/closure': Error downloading [https:\/\/mirror.bazel.build\/github.com\/bazelbuild\/rules_closure\/archive\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz, https:\/\/github.com\/bazelbuild\/rules_closure\/archive\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz] to \/home\/user22\/.cache\/bazel\/_bazel_user22\/673c993ff8d0b73d5fcaf03ddf761476\/external\/io_bazel_rules_closure\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz: All mirrors are down: ***[Unable to tunnel through proxy. Proxy returns \"HTTP\/1.1 502 Bad Gateway\"]***\nERROR: error loading package '': Encountered error while reading extension file 'closure\/defs.bzl': no such package '@io_bazel_rules_closure\/\/closure': Error downloading [https:\/\/mirror.bazel.build\/github.com\/bazelbuild\/rules_closure\/archive\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz, https:\/\/github.com\/bazelbuild\/rules_closure\/archive\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz] to \/home\/user22\/.cache\/bazel\/_bazel_user22\/673c993ff8d0b73d5fcaf03ddf761476\/external\/io_bazel_rules_closure\/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz: All mirrors are down: ***[Unable to tunnel through proxy. Proxy returns \"HTTP\/1.1 502 Bad Gateway\"]***\nINFO: Elapsed time: 14.841s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)\n<\/code><\/pre>\n\n<p>The issue is clearly this HTTP\/1.1 502 error. I know other people had issues when using bazel behind a proxy, but I did find no reference to this HTTP error. Any idea about how I could fix this ? <\/p>\n","answers":[{"AnswerId":"52833456","CreationDate":"2018-10-16T10:32:08.780","ParentId":null,"OwnerUserId":"2332808","Title":null,"Body":"<p>\"502 Bad gateway\" errors generally mean bazel asked the proxy to do something it was not allowed to do - in particular, bazel does not issue GET requests directly to the proxy, but will ask for a CONNECT command directly (seen with strace):<\/p>\n\n<p><code>\nsendto(220, \"CONNECT mirror.bazel.build:443 HTTP\/1.1\\r\\nUser-Agent: Java\/1.8.0_181\\r\\nHost: mirror.bazel.build\\r\\nAccept: text\/html, image\/gif, image\/jpeg, *; q=.2, *\/*; q=.2\\r\\nProxy-Connection: keep-alive\\r\\n\\r\\n\", 189, 0, NULL, 0 &lt;unfinished ...&gt;\n<\/code><\/p>\n\n<p>Your proxy configuration needs to be updated to allow CONNECT commands.<\/p>\n\n<p>Alternatively, some <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/587\" rel=\"nofollow noreferrer\">bazel github issues<\/a> indicate that you could look at the errors and download the content yourself, but that did not seem to end very well.<\/p>\n"}]}
+{"QuestionId":52831934,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-16T09:12:19.753","AcceptedAnswerId":null,"OwnerUserId":8290827.0,"Title":"Categorical and continuous cross feature column in Tensorflow","Body":"<p>In using Tensorflow's estimators and <code>feature_column<\/code> it is possible to cross a categorical column and a bucketed continuous column <a href=\"https:\/\/www.tensorflow.org\/guide\/feature_columns#crossed_column\" rel=\"nofollow noreferrer\">crossed column<\/a> but not a categorical and numeric cross. Could it be possible to implement this functionality from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/feature_column\/feature_column.py#L704\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/feature_column\/feature_column.py#L704<\/a> ?<\/p>\n\n<p>It would also be great to see any alternate methods for achieving the same outcome within the Tensforflow graph.<\/p>\n\n<pre><code>import numpy as np\n\ncont = np.array([1,2,3])\ncat = np.array(['cat', 'dog', 'cat'])\n\ncross_function(cat, cont) = np.array([[1,0],[0,2],[3,0]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52891782","CreationDate":"2018-10-19T11:50:13.617","ParentId":null,"OwnerUserId":"8290827","Title":null,"Body":"<p>To answer my own question here. The steps involved are: <\/p>\n\n<ol>\n<li>Numerically encoding the categorical feature \n\n<ul>\n<li>Within the graph so it's possible within train and serve<\/li>\n<\/ul><\/li>\n<li>One hot encoding the numerical result<\/li>\n<li>Multiplying this with the continuous variable<\/li>\n<\/ol>\n\n<p>Code:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ncont = np.array([1,2,3])\ncat = np.array(['cat', 'dog', 'cat'])\ncategories = np.unique(cat)\n\ndef categorical_continuous_interaction(categorical_onehot, continuous):\n\n    cont = tf.expand_dims(continuous, 0)\n    return tf.transpose(tf.multiply(tf.transpose(categorical_onehot), cont))\n\ndef transformation_function(feature_dictionary, mapping_table):\n\n    continuous_feature = feature_dictionary['cont']\n\n    categorical_feature = mapping_table.lookup(feature_dictionary['cat'])\n    onehot = tf.one_hot(categorical_feature, categories.shape[0])\n    cross_feature = categorical_continuous_interaction(onehot, continuous_feature)\n\n    return {'feature_name': cross_feature}\n\ndef input_function(dataframe, label_key, ...):\n    # categorical mapping tables, these must be generated outside of the dataset \n    # transformation function but within the input function\n    mapping_table = tf.contrib.lookup.index_table_from_tensor(\n        mapping=tf.constant(categories),\n        num_oov_buckets=0, \n        default_value=-1\n    )\n\n    # Generate the dataset of a dictionary of all of the dataframes columns\n    dataset = tf.data.Dataset.from_tensor_slices(dict(dataframe))\n    #\u00a0Convert to a dataset of tuples of dicts with the labels as one tuple\n    dataset = dataset.map(lambda x: split_label(x, label_key))\n    # Transform the features dict within the dataset\n    dataset = dataset.map(lambda features, labels: (transformation_function(\n        features, mapping_table=mapping_table), labels))\n\n    ...\n\n    return dataset\n\ndef serving_input_fn():\n    # categorical mapping tables, these must be generated outside of the dataset \n    # transformation function but within the input function\n    mapping_table=tf.contrib.lookup.index_table_from_tensor(\n        mapping=tf.constant(categories),\n        num_oov_buckets=0, \n        default_value=-1\n    )\n    numeric_receiver_tensors = {\n        name: tf.placeholder(dtype=tf.float32, shape=[1], name=name+\"_placeholder\")\n        for name in numeric_feature_column_names\n    }\n    categorical_receiver_tensors = {\n        name: tf.placeholder(dtype=tf.string, shape=[1], name=name+\"_placeholder\")\n        for name in categorical_feature_column_names\n    }\n    receiver_tensors = {**numeric_receiver_tensors, **categorical_receiver_tensors}\n\n    features = transformation_function(receiver_tensors, \n        country_mapping_table=country_mapping_table)\n\n    return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52832028,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><type-conversion><fft>","CreationDate":"2018-10-16T09:17:57.197","AcceptedAnswerId":52844522.0,"OwnerUserId":6595663.0,"Title":"calculate stft(Short-time Fourier transform) in tensor flow audio recognition","Body":"<p>I have a hard time running this piece of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/signal\/mfccs_from_log_mel_spectrograms\" rel=\"nofollow noreferrer\">example code<\/a> here to convert the audio signal into stfts. I am using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/speech_commands\/label_wav.py\" rel=\"nofollow noreferrer\">label_wave.py<\/a> and editing the \n\"run graph\" function.<\/p>\n\n<pre><code>def run_graph(wav_data, labels, input_layer_name, output_layer_name,\n              num_top_predictions):\n  \"\"\"Runs the audio data through the graph and prints predictions.\"\"\"\n  data = np.fromstring(wav_data, dtype='int16')[-16000:]\n  # Normalize the data\n  data = data * 1.0 \/ (1 &lt;&lt; 15)   # Range [-1 1]\n  data = tf.reshape(data, (1, 16000))\n  # tf.cast(data, tf.float32)\n  stfts = tf.contrib.signal.stft(data, frame_length=1024, frame_step=256,fft_length=1024)\n<\/code><\/pre>\n\n<p>Error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"mfcc_label_wav.py\", line 149, in &lt;module&gt;\n    tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/platform\/app.py\", line 126, in run\n    _sys.exit(main(argv))\n  File \"mfcc_label_wav.py\", line 121, in main\n    FLAGS.output_name, FLAGS.how_many_labels)\n  File \"mfcc_label_wav.py\", line 115, in label_wav\n    run_graph(wav_data, labels_list, input_name, output_name, how_many_labels)\n  File \"mfcc_label_wav.py\", line 71, in run_graph\n    fft_length=1024)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/signal\/python\/ops\/spectral_ops.py\", line 91, in stft\n    return spectral_ops.rfft(framed_signals, [fft_length])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/ops\/spectral_ops.py\", line 130, in _rfft\n    input_tensor = _ops.convert_to_tensor(input_tensor, _dtypes.float32)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 950, in convert_to_tensor\n    as_ref=False)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 1040, in internal_convert_to_tensor\n    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 883, in _TensorTensorConversionFunction\n    (dtype.name, t.dtype.name, str(t)))\nValueError: Tensor conversion requested dtype float32 for Tensor with dtype float64: 'Tensor(\"stft\/mul:0\", shape=(1, 59, 1024), dtype=float64)'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52844522","CreationDate":"2018-10-16T21:53:53.673","ParentId":null,"OwnerUserId":"6595663","Title":null,"Body":"<p>Performing tf.cast(data, tf.float32) isn't working. So i converted numpy array from float64 to float32 first and then reshaped the data.<\/p>\n\n<pre><code>  data = np.fromstring(wav_data, dtype='int16')[-16000:]\n  # Normalize the data\n  data = data * 1.0 \/ (1 &lt;&lt; 15)   # Range [-1 1]\n  data = np.float32(data)\n  data = tf.reshape(data, (1, 16000))\n  # tf.cast(data, tf.float32)\n  stfts = tf.contrib.signal.stft(data, frame_length=1024, \n                                 frame_step=256,fft_length=1024)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52832207,"AnswerCount":1,"Tags":"<tensorflow><padding><rnn>","CreationDate":"2018-10-16T09:27:15.237","AcceptedAnswerId":null,"OwnerUserId":5647864.0,"Title":"Dynamic RNN: padding word vector","Body":"<p>I am a bit confused on padding, my first question is: <\/p>\n\n<ul>\n<li>Is it possible to pad a shorter sequence with values that are not 0? How do you deal with that then in the RNN?<\/li>\n<li><p>Generally a 0 is used for padding, is there a specific reason why? Does it make it easy in the training because it does not affect the calculation or you still need to mask the loss function? <\/p><\/li>\n<li><p>In case your sentence is composed of vectors embedding from a word2vec model, would padding be applied as a zero vector?<\/p><\/li>\n<\/ul>\n\n<p>Thanks in advance fir any hint!<\/p>\n","answers":[{"AnswerId":"57445265","CreationDate":"2019-08-10T19:52:46.733","ParentId":null,"OwnerUserId":"13969","Title":null,"Body":"<p>Your question is addressed in <a href=\"https:\/\/datascience.stackexchange.com\/questions\/13120\/how-to-overcome-training-examples-different-lengths-when-working-with-word-embe\">How to overcome training example's different lengths when working with Word Embeddings (word2vec)<\/a>.<\/p>\n\n<p>For details on the alternating min\/max padding method, see <a href=\"https:\/\/stats.stackexchange.com\/questions\/221715\/apply-word-embeddings-to-entire-document-to-get-a-feature-vector\">Apply word embeddings to entire document, to get a feature vector<\/a>.<\/p>\n\n<p>See also: <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/#pad_sequences\" rel=\"nofollow noreferrer\">keras.preprocessing.sequence.pad_sequences<\/a>, which can take a value to pad with as an argument.<\/p>\n"}]}
+{"QuestionId":52832497,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T09:42:28.353","AcceptedAnswerId":52838360.0,"OwnerUserId":9540764.0,"Title":"Clarification of the tf.Session() scope TensorFlow","Body":"<p>I have a python program in which I have defined a network and as usual, I'm training it inside a function where I have<\/p>\n\n<pre><code>with tf.Session() as sess:\n...\nfor epoch in xrange(num_epochs):\n...\nfor n in xrange(num_batches):\n       _, c = sess.run([optimizer, loss], feed_dict={....\n<\/code><\/pre>\n\n<p>In the loss function, I have to work a lot to get the loss and in particular, I must take the maximum inside a tensor and use it to do stuff. Here an example<\/p>\n\n<pre><code>values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)\nmax_values = tf.reduce_max(values) # Tensor...\n...\n<\/code><\/pre>\n\n<p>in the <code>max_values<\/code> line if I use the debug it says that it is a tensor and not a value, so if I change my code in this way passing to the function the session created in the previous piece of code<\/p>\n\n<pre><code>values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)\nmax_values = sess.run(tf.reduce_max(values)) # 2.0\n...\n<\/code><\/pre>\n\n<p>it works. But this loss function is already in a scope of a session so my question is why the result is a tensor and not a number? Is there a way to get the value without passing the session to the loss function?<\/p>\n","answers":[{"AnswerId":"52838360","CreationDate":"2018-10-16T14:55:38.580","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Use Tensor.eval() function to convert a Tensor to its value.\nIn the following example, you can get the value of max_values Tensor.<\/p>\n\n<pre><code>def loss():\n  values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)\n  max_values = tf.reduce_max(values)\n  print (max_values.eval())\n\nwith tf.Session() as sess:\n  loss()\n<\/code><\/pre>\n\n<p>If you called loss() outside Session scope, you got an error.<\/p>\n\n<p>Also you can use Eager execution mode. <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/eager_basics\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/eager\/eager_basics<\/a><\/p>\n"},{"AnswerId":"52833236","CreationDate":"2018-10-16T10:19:03.993","ParentId":null,"OwnerUserId":"9178720","Title":null,"Body":"<p>According to the <a href=\"https:\/\/www.tensorflow.org\/guide\/graphs\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>TensorFlow uses the tf.Session class to represent a connection between\n  the client program---typically a Python program, although a similar\n  interface is available in other languages---and the C++ runtime.<\/p>\n<\/blockquote>\n\n<p>It means that when you do <code>values = tf.constant([0, 1, 2, 0, 2], dtype=tf.float32)<\/code> you are just inserting a node to your tensorflow Graph! Since Python is the high level API for a low level C++ runtime, you actually need a Session to evaluate your Python code in this low level runtime.<\/p>\n\n<p>That is why every time you need to calculate or evaluate a Tensorflow variable\/method\/constant\/etc you need to run it in a Session with <code>tf.Session().run(yournode)<\/code><\/p>\n\n<p>I hope it helped<\/p>\n"}]}
+{"QuestionId":52832618,"AnswerCount":1,"Tags":"<tensorflow><evaluation><tensorflow-estimator>","CreationDate":"2018-10-16T09:47:44.900","AcceptedAnswerId":54021286.0,"OwnerUserId":988709.0,"Title":"How to calculate the accuracy for multilabel classification with tf.metrics?","Body":"<p>I want to train multilabel classification model with tensorflow (tf.estimator.Estimator). I need to output the accuracy when do evaluating. But it seem not work with the following code:<\/p>\n\n<pre><code>accuracy = tf.metrics.accuracy(labels=labels, predictions=preds)\nmetrics = {'accuracy': accuracy}\n\nif mode == tf.estimator.ModeKeys.EVAL:\n    return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=metrics)\n<\/code><\/pre>\n\n<p>the <code>tf.metrics.accuracy<\/code> is not for multihot result. So what is the multilabel metrics?<\/p>\n","answers":[{"AnswerId":"54021286","CreationDate":"2019-01-03T11:19:23.990","ParentId":null,"OwnerUserId":"1462770","Title":null,"Body":"<p>Actually <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/metrics\/accuracy\" rel=\"nofollow noreferrer\"><code>tf.metrics.accuracy<\/code><\/a> calculates the accuracy for multilabel classification as well. See the example below:<\/p>\n\n<pre><code>import tensorflow as tf\n\nlabels = tf.constant([[1, 0, 0, 1],\n                      [0, 1, 1, 1],\n                      [1, 1, 0, 0],\n                      [0, 0, 0, 1],\n                      [1, 1, 0, 0]])\n\npreds = tf.constant([[1, 0, 1, 1],\n                     [0, 1, 1, 1],\n                     [1, 1, 0, 0],\n                     [0, 0, 0, 1],\n                     [1, 1, 0, 0]])\n\nacc, acc_op = tf.metrics.accuracy(labels, preds)\n\nwith tf.Session() as sess:\n    sess.run(tf.local_variables_initializer())\n    sess.run(tf.global_variables_initializer())\n    print(sess.run([acc, acc_op]))\n    print(sess.run([acc]))\n<\/code><\/pre>\n\n<p>As you can see we have 20 labels in total that only one entry, in the first row, is mislabeled as a result our accuracy is 0.95%.<\/p>\n"}]}
+{"QuestionId":52832625,"AnswerCount":1,"Tags":"<python><tensorflow><dataset><upgrade>","CreationDate":"2018-10-16T09:48:02.750","AcceptedAnswerId":55746805.0,"OwnerUserId":4507946.0,"Title":"OutOfRangeError logged at each epoch after upgrade from Tensorflow 1.8.0 to 1.11.0","Body":"<p>I have a TfRecordsDataset:<\/p>\n\n<pre><code>...\ndataset = dataset.batch(batch_size)\ndataset = dataset.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)\ndataset_iterator = dataset.make_initializable_iterator()\n(...) = dataset_iterator.get_next()\n<\/code><\/pre>\n\n<p>And then the standard:<\/p>\n\n<pre><code>test_session.run(dataset_iterator.initializer)\nwhile not done:\n    try:\n        ... = test_session.run([...])\n        batch_count += 1\n    except tf.errors.OutOfRangeError:\n        done = True\n<\/code><\/pre>\n\n<p>This worked perfectly with Tensorflow 1.8.0. Today I upgraded Tensorflow to TF 11.1.0, and I now have an error:\n <code>OutOfRangeError: End of sequence [[{{node dataset\/IteratorGetNext}} = IteratorGetNext<\/code><\/p>\n\n<p>at each end of an epoch. The error itself is actually caught, but the error is printed on the console. Can someone point out what is actually happening ?<\/p>\n","answers":[{"AnswerId":"55746805","CreationDate":"2019-04-18T13:13:24.257","ParentId":null,"OwnerUserId":"4507946","Title":null,"Body":"<p>Problem was solved in TF 1.12.0.<\/p>\n"}]}
+{"QuestionId":52833049,"AnswerCount":2,"Tags":"<python><machine-learning><pytorch><loss-function>","CreationDate":"2018-10-16T10:09:20.847","AcceptedAnswerId":null,"OwnerUserId":6866762.0,"Title":"PyTorch: Loss remains constant","Body":"<p>I've written a code in PyTorch with my own implemented loss function <code>focal_loss_fixed<\/code>. But my loss value stays fixed after every epoch. Looks like weights are not being updated. Here is my code snippet:<\/p>\n\n<pre><code>optimizer = optim.SGD(net.parameters(),\n                          lr=lr,\n                          momentum=0.9,\n                          weight_decay=0.0005)\n\n\nfor epoch in T(range(20)):\n    net.train()\n    epoch_loss = 0\n    for n in range(len(x_train)\/\/batch_size):\n        (imgs, true_masks) = data_gen_small(x_train, y_train, iter_num=n, batch_size=batch_size)\n        temp = []\n        for tt in true_masks:\n            temp.append(tt.reshape(128, 128, 1))\n        true_masks = np.copy(np.array(temp))\n        del temp\n        imgs = np.swapaxes(imgs, 1,3)\n        imgs = torch.from_numpy(imgs).float().cuda()\n        true_masks = torch.from_numpy(true_masks).float().cuda()\n        masks_pred = net(imgs)\n        masks_probs = F.sigmoid(masks_pred)\n        masks_probs_flat = masks_probs.view(-1)\n        true_masks_flat = true_masks.view(-1)\n        print((focal_loss_fixed(tf.convert_to_tensor(true_masks_flat.data.cpu().numpy()), tf.convert_to_tensor(masks_probs_flat.data.cpu().numpy()))))\n        loss = torch.from_numpy(np.array(focal_loss_fixed(tf.convert_to_tensor(true_masks_flat.data.cpu().numpy()), tf.convert_to_tensor(masks_probs_flat.data.cpu().numpy())))).float().cuda()\n        loss = Variable(loss.data, requires_grad=True)\n        epoch_loss *= (n\/(n+1))\n        epoch_loss += loss.item()*(1\/(n+1))\n        print('Step: {0:.2f}% --- loss: {1:.6f}'.format(n * batch_size* 100.0 \/ len(x_train), epoch_loss), end='\\r')\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n    print('Epoch finished ! Loss: {}'.format(epoch_loss))\n<\/code><\/pre>\n\n<p>And this is my `focal_loss_fixed' function:<\/p>\n\n<pre><code>def focal_loss_fixed(true_data, pred_data):\n    gamma=2.\n    alpha=.25\n    eps = 1e-7\n    # print(type(y_true), type(y_pred))\n    pred_data = K.clip(pred_data,eps,1-eps)\n    pt_1 = tf.where(tf.equal(true_data, 1), pred_data, tf.ones_like(pred_data))\n    pt_0 = tf.where(tf.equal(true_data, 0), pred_data, tf.zeros_like(pred_data))\n    with tf.Session() as sess:\n        return sess.run(-K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0)))\n<\/code><\/pre>\n\n<p>After each epoch the loss value stays constant(<code>5589.60328<\/code>). What's wrong with it?<\/p>\n","answers":[{"AnswerId":"52849889","CreationDate":"2018-10-17T07:55:15.163","ParentId":null,"OwnerUserId":"6781449","Title":null,"Body":"<p>When computing the loss you call <code>focal_loss_fixed()<\/code> which uses TensorFlow to compute the loss value. <code>focal_loss_fixed()<\/code> creates a graph and runs it in a session to get the value, and by this point PyTorch has no idea of the sequence of operations that led to the loss because they were computed by the TensorFlow backend. It is likely then, that all PyTorch sees in <code>loss<\/code> is a constant, as if you had written <\/p>\n\n<pre><code>loss = 3\n<\/code><\/pre>\n\n<p>So the gradient will be zero, and the parameters will never be updated. I suggest you rewrite your loss function using PyTorch operations so that the gradient with respect to its inputs can be computed.<\/p>\n"},{"AnswerId":"52833233","CreationDate":"2018-10-16T10:18:57.183","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>I think the problem lies in your heavy weight decay.<\/p>\n\n<p>Essentially, you are not <em>reducing the weight by <code>x<\/code><\/em>, but rather you <em>multiply the weights by <code>x<\/code><\/em>, which means that you are instantaneously only doing very small increments, leading to a (seemingly) plateauing loss function.<\/p>\n\n<p>More explanation on this can be found in the PyTorch discussion forum (e.g., <a href=\"https:\/\/discuss.pytorch.org\/t\/how-pytorch-implement-weight-decay\/8436\" rel=\"nofollow noreferrer\">here<\/a>, or <a href=\"https:\/\/discuss.pytorch.org\/t\/weight-decay-implementation\/1796\/2\" rel=\"nofollow noreferrer\">here<\/a>). <br\/>\nUnfortunately, <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/sgd.html#SGD\" rel=\"nofollow noreferrer\">the source for SGD<\/a> alone also does not tell you much about its implementation.\nSimply setting it to a larger value should result in better updates. You can start by leaving it out completely, and then iteratively reducing it (from <code>1.0<\/code>), until you get more decent results.<\/p>\n"}]}
+{"QuestionId":52833079,"AnswerCount":0,"Tags":"<python><neural-network><keras><forecasting>","CreationDate":"2018-10-16T10:10:49.950","AcceptedAnswerId":null,"OwnerUserId":4619760.0,"Title":"Out of sample prediction\/forecasting for uni-variate time series in Keras","Body":"<p>I am using <a href=\"https:\/\/www.kaggle.com\/ternaryrealm\/lstm-time-series-explorations-with-keras\" rel=\"nofollow noreferrer\">this Kaggle guide<\/a> to do time series forecasting (sample data attached).<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><code>def create_dataset(dataset, window_size = 1):\n    data_X, data_Y = [], []\n    for i in range(len(dataset) - window_size - 1):\n        a = dataset[i:(i + window_size), 0]\n        data_X.append(a)\n        data_Y.append(dataset[i + window_size, 0])\n    return(np.array(data_X), np.array(data_Y))   \n\ndef fit_model(train_X, train_Y, window_size = 1):\n    model = Sequential()\n\n    model.add(LSTM(4, \n               input_shape = (1, window_size)))\n    model.add(Dense(1))\n    model.compile(loss = \"mean_squared_error\", \n              optimizer = \"adam\")\n    model.fit(train_X, \n          train_Y, \n          epochs = 100, \n          batch_size = 1, \n          verbose = 0)\n\n    return(model)\n\ndef predict_and_score(model, X, Y):\n# Make predictions on the original scale of the data.\n    pred = MinMaxScaler(feature_range = (0,1)).inverse_transform(model.predict(X))\n# Prepare Y data to also be on the original scale for interpretability.\n    orig_data = MinMaxScaler(feature_range = (0,1)).inverse_transform([Y])\n# Calculate RMSE.\n\n    score = math.sqrt(mean_squared_error(orig_data[0], pred[:, 0]))\n    return(score, pred)\n<\/code><\/pre>\n\n<p>This entire thing is being used in the following function:<\/p>\n\n<pre><code>def nnet(time_series, window_size=1,  ): \n    cmi_total_raw = vstack((time_series.values.astype('float32')))\n    scaler = MinMaxScaler(feature_range = (0,1))\n    cmi_total_scaled = scaler.fit_transform(cmi_total_raw)\n    cmi_train_sc = (cmi_total_scaled[0:int(cmi_split*len(cmi_total_scaled))])\n    cmi_test_sc = cmi_total_scaled[int(cmi_split*len(cmi_total_scaled)) : len(cmi_total_scaled)] \n\n\n    # Create test and training sets for one-step-ahead regression.\n    window_size = 1\n    train_X, train_Y = create_dataset(cmi_train_sc, window_size)\n    test_X, test_Y = create_dataset(cmi_test_sc, window_size)\n\n    # Reshape the input data into appropriate form for Keras.\n    train_X = np.reshape(train_X, (train_X.shape[0], 1, train_X.shape[1]))\n    test_X = np.reshape(test_X, (test_X.shape[0], 1, test_X.shape[1]))\n\n    model = fit_model(train_X, train_Y, window_size)\n\n    rmse_train, train_predict = predict_and_score(nn_model, train_X, train_Y)\n\n    mape_test, test_predict = predict_and_score(model, test_X, test_Y)    \n    return (mape_test, test_predict)   \n<\/code><\/pre>\n\n<p>As far as I understand, it is creating a model based on training data and predicting on in-sample test set and finally calculates the error.<\/p>\n\n<p>The input data has 209 rows and I want to predict the next row(s).<\/p>\n\n<p>Here's what I tried:<\/p>\n\n<p>Since the same thing is done in Auto-Arima using <code>forecast(steps= n_steps)<\/code> method, I looked for something similar in Keras.<\/p>\n\n<p>From Keras <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<p><code>predict(x, batch_size=None, verbose=0, steps=None)<\/code><\/p>\n\n<p>Arguments:<\/p>\n\n<p><code>x: The input data, as a Numpy array (or list of Numpy arrays if the model has multiple inputs).<\/code><\/p>\n\n<p><code>steps: Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None.<\/code><\/p>\n\n<p>I tried changing <code>step<\/code> and it predicted very absurd values of the order of 100,000. Moreover, length of the <code>test_predict<\/code> was no way near the <code>steps<\/code> I gave. So I am assuming <code>step<\/code> means something else here.<\/p>\n\n<p><strong>Question<\/strong>\n- Can Keras even be used to forecast time series data (out of sample)\n- If yes, is there a <code>forecast<\/code> method just as there the aforementioned <code>predict<\/code> method?\n- If no, can the existing <code>predict<\/code> method be used in any way to get out of sample forecast? <\/p>\n\n<p>Sample data (cmi_total):<\/p>\n\n<pre><code>2014-05-25    272.459887\n2014-06-01    272.446022\n2014-06-08    330.301260\n2014-06-15    656.838394\n2014-06-22    670.575110\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52833181,"AnswerCount":0,"Tags":"<python><neural-network><keras><compression><autoencoder>","CreationDate":"2018-10-16T10:15:56.583","AcceptedAnswerId":null,"OwnerUserId":7896479.0,"Title":"Autoencoder in Keras with low Validation loss but reconstruction is just a noise","Body":"<p>I am trying to compress a timewave by using autoencoder. I am using the below-mentioned architecture. <\/p>\n\n<pre><code>#input_wave = Input(shape=(7500,))\nencoded_1 = Dense(1000, activation='tanh',input_shape=(7500,))\nencoded_2 = Dense(500, activation='tanh')\nencoded_3 = Dense(300, activation='tanh')\nencoded_4 = Dense(128, activation='tanh')\n\ndecoded_3 = Dense(300, activation='tanh')\ndecoded_2 = Dense(500, activation='tanh')\ndecoded_1 = Dense(1000, activation='tanh')\ndecoded_0 = Dense(7500, activation='linear')\n<\/code><\/pre>\n\n<p>This is the loss function am using<\/p>\n\n<pre><code>#autoencoder = Model(input_img, decoded)\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\n\nautoencoder.fit(x_train, x_train,\n            epochs=15,\n            batch_size=300,\n            shuffle=True,\n            validation_data=(x_test, x_test),\n            callbacks=[TensorBoard(log_dir='\/tmp\/autoencoder')])\n<\/code><\/pre>\n\n<p>and this is how my validation score looks like. \n<a href=\"https:\/\/i.stack.imgur.com\/b6u3q.png\" rel=\"nofollow noreferrer\">my validation scores<\/a><\/p>\n\n<p>but when I try to use predict and reconstruct the wave it appears to be just a noise. \n<a href=\"https:\/\/i.stack.imgur.com\/1vitx.png\" rel=\"nofollow noreferrer\">Predicted and real timewave<\/a><\/p>\n\n<p>what am I doing wrong? any leads will be highly appreciated.  <\/p>\n\n<p><strong>update:<\/strong>\nafter changing the loss function to MSE and optimizer to adam i saw the network overfitting. Validation loss never goes down significantly. please look at the attached screenshot. \n<a href=\"https:\/\/i.stack.imgur.com\/0uYz4.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p><strong>Update<\/strong>\nI shortend the length of my Data to 2500 and it start working. I think my model was not sophisticated enough to compress the 7500 to 128 or maybe it is just not possible. So i use 3 separate models to compress the data. Thanks everyone. <\/p>\n","answers":[]}
+{"QuestionId":52833220,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-16T10:18:07.373","AcceptedAnswerId":52833889.0,"OwnerUserId":10386005.0,"Title":"Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_3\/MaxPool' (op: 'MaxPool') with input shapes: [?,1,148,32]","Body":"<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=(3, 150, 150),padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(32, (3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(64, (3, 3),padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())  # this converts our 3D feature maps to 1D feature \nvectors\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n    #sgd = optimizers.SGD(lr=0.0001, decay=1e-6, momentum=0.9)\n\n\nmodel.compile(loss='sparse_categorical_crossentropy',  \n`optimizer=Adam(lr=0.001), # Adam optimizer with 1.0e-4 learning rate\nmetrics = ['accuracy']) # Metrics to be evaluated by the model\n<\/code><\/pre>\n\n<p>When I compile the above code i get this error\nNegative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_8\/MaxPool' (op: 'MaxPool') with input shapes: [?,1,75,32].<\/p>\n\n<p>I tried with the same padding and it still doesn't work<\/p>\n","answers":[{"AnswerId":"52833889","CreationDate":"2018-10-16T10:57:55.100","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Pretty sure if you change <\/p>\n\n<pre><code>model.add(Conv2D(32, (3, 3), input_shape=(3, 150, 150),padding='same'))\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>model.add(Conv2D(32, (3, 3), input_shape=(150, 150, 3),padding='same'))\n<\/code><\/pre>\n\n<p>(you may have to change the shape of your data too)<\/p>\n\n<p>it will work as intended.<\/p>\n"}]}
+{"QuestionId":52834134,"AnswerCount":0,"Tags":"<tensorflow><model><restore><tensorflowsharp>","CreationDate":"2018-10-16T11:11:40.603","AcceptedAnswerId":null,"OwnerUserId":1942408.0,"Title":"You must feed a value for placeholder tensor 'X_1' with dtype float and shape [?,3]","Body":"<p>I've wrote a very simple model through the Jupyter Notebook<\/p>\n\n<pre><code>n_input = 3  \nX = tf.placeholder(tf.float32, [None, n_input], name=\"X\")\ndecoder = tf.matmul(X, [[2.0,3.0],[2.0,3.0],[2.0,3.0]], name='decoder')\n\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n\nvalues = sess.run(decoder, feed_dict={X: df_norm[:10]})\n\ngraph = tf.get_default_graph()\ntf.train.write_graph(graph, '.\/model\/','saved_model.pbtxt', as_text=False)\n<\/code><\/pre>\n\n<p>Then I loaded it with the TensorflowSharp<\/p>\n\n<pre><code>using (var graph = new TFGraph())\n            {\n                var bytes = File.ReadAllBytes(@\".\\model\\saved_model.pbtxt\");\n                graph.Import(bytes);\n                var session = new TFSession(graph);\n\n\n                var runner = session.GetRunner();\n                runner.AddInput(graph[\"X\"][0], new float[] { 176.75f, 7.95f, 40397.00f }); \n\n                runner.Fetch(graph[\"decoder\"][0]);\n\n                var output = runner.Run();\n\n                \/\/ Fetch the results from output:\n                TFTensor result = output[0];\n            }\n<\/code><\/pre>\n\n<p>Finally, I got the exception below:<\/p>\n\n<blockquote>\n  <p>TensorFlow.TFException   HResult=0x80131500   \uba54\uc2dc\uc9c0=You must feed a\n  value for placeholder tensor 'X_1' with dtype float and shape [?,3]<br>\n  [[{{node X_1}} = Placeholderdtype=DT_FLOAT, shape=[?,3],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]   \uc18c\uc2a4=TensorFlowSharp   StackTrace:    at\n  TensorFlow.TFStatus.CheckMaybeRaise(TFStatus incomingStatus, Boolean\n  last)    at TensorFlow.TFSession.Run(TFOutput[] inputs, TFTensor[]\n  inputValues, TFOutput[] outputs, TFOperation[] targetOpers, TFBuffer\n  runMetadata, TFBuffer runOptions, TFStatus status)    at\n  TensorFlow.TFSession.Runner.Run(TFStatus status)    at\n  tensorflowsharp_model_restore.Program.Main(String[] args) in\n  G:\\tensorflow\\tensorflowsharp\\tensorflowsharp_model_restore\\tensorflowsharp_model_restore\\Program.cs:line\n  29<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52834197,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><anaconda><jupyter-notebook>","CreationDate":"2018-10-16T11:15:10.577","AcceptedAnswerId":null,"OwnerUserId":8864449.0,"Title":"Jupyter Notebook(Anaconda) and Python IDLE\/Shell and Python on command prompt","Body":"<p>I wanted to know whether all these different and independent of each other? And what does pip list exactly show? Because in my pip list I can see that I have numpy 1.14.3 , and when I import it on command prompt or jupyter notebook, it is successfully importing, but when I tried importing it on Python Shell, it showed that no such module exists. I am using Windows 10. Python IDLE is version 3.6.\n<a href=\"https:\/\/i.stack.imgur.com\/O0bnb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O0bnb.png\" alt=\"This is my python version on command prompt and as there is Anaconda written beside it, maybe both are interlinked\"><\/a><\/p>\n\n<p>Recently, I installed OpenCV using this command-\npip install opencv-contrib-python<\/p>\n\n<p>And after that I was able to import cv2 on command prompt but not on jupyter notebook and neither on Python Shell\/IDLE<\/p>\n\n<p>But when I installed tensorflow using-\npip install tensorflow\nAfter that I was able to import tensorflow both on jupyter notebook and command prompt but not on IDLE.<\/p>\n\n<p>I just wanted to know that how many Pythons do I have on my laptop and when I have to install a package which I do using pip, then what for am I exactly installing it, whether it is for Python Shell or command prompt or jupyter notebook. Please help. Thanks a lot.<\/p>\n","answers":[]}
+{"QuestionId":52835698,"AnswerCount":0,"Tags":"<tensorflow><cuda><keras>","CreationDate":"2018-10-16T12:41:07.627","AcceptedAnswerId":null,"OwnerUserId":9468014.0,"Title":"ROCm and CUDA together in Keras?","Body":"<p>is it possible to use ROCm together with CUDA GPUs on the same model in Keras? I dont think so since you need to have a different tf version for the both? Did anybody try to run it on different GPU's?<\/p>\n","answers":[]}
+{"QuestionId":52836388,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-16T13:18:48.163","AcceptedAnswerId":null,"OwnerUserId":4740456.0,"Title":"Key Error when I was trying to insert image data and label into my pre-trained tensorflow model (shuffleNet)","Body":"<p>I am trying to restore a pre-trained model which is model called sufflenet, and it has been trained by other people. However, when I was trying to retrieve the model and restore tensorflow graph to train a new data set(10000 image data set), I got a key error message in  my terminal:\nHere is my code :<\/p>\n\n<pre><code>meta_path = '.\/model\/model.ckpt-0.meta'\ntf.reset_default_graph()\nsaver = tf.train.import_meta_graph(meta_path)\nrestored_graph = tf.get_default_graph()\n\nfor tensor in restored_graph.get_operations():\n    print (tensor.name)\n\n\nglobal_step_tensor = restored_graph.get_tensor_by_name('Softmax\/prediction:0')\nimage_input_node = restored_graph.get_tensor_by_name('TFRecordIterator\/IteratorGetNext:0')\nlabel_node = restored_graph.get_tensor_by_name('TFRecordIterator\/OneShotIterator:0')\nloss = restored_graph.get_tensor_by_name('sparse_softmax_cross_entropy\/add:0')\n\n# tf.contrib.quantize.create_training_graph(input_graph=restored_graph, quant_delay=2000000)\noptimizer = tf.train.GradientDescentOptimizer(0.01)\ntrain_op = optimizer.minimize(loss)\n# tf.contrib.quantize.create_training_graph(quant_delay=2000000)\n\n\niterations = 100\n\n# run the session\nwith tf.Session() as sess:\n    # restore the saved vairable\n    saver.restore(sess, '.\/model\/model.ckpt-0')\n    # sess.run(optimizer)\n\n    image_files = []\n    labels = []\n\n    with open(\"\/mnt\/ficussweden\/hhzhang\/00_Data\/04_TF_Autobot\/gender\/data\/train.tsv\", 'r') as f:\n        line = f.readline()\n        while line:\n            array = line.rstrip('\\n').split()\n            image_files.append(array[0])\n            labels.append(int(array[1]))\n            line = f.readline()\n\n            if len(image_files)&gt;=10001:\n                break\n\n            # print(image_files[len(image_files)-1])\n\n    data = []\n    data_labels = []\n    for i in range(200):\n        print(\"process batch {}-th data \\n\".format(i))\n        batch = []\n        batch_label = []\n        for j in range(50):\n            img_path = image_files[j+i*50]\n            img = np.array(Image.open(img_path))\n            batch.append(img)\n            batch_label.append(labels[j+i*50])\n\n        batch = np.array(batch)\n        batch_label = np.array(batch_label)\n\n        data.append(batch)\n        data_labels.append(batch_label)\n    data = np.array(data)\n    data_labels = np.array(data_labels)\n    print(data.shape)\n    print(data_labels.shape)\n\n    for i in range(iterations):\n        for j in range(len(data)):\n            batch_data = data[j]\n            batch_label = data_labels[j]\n\n\n            res = sess.run(train_op, feed_dict = {image_input_node: batch_data, label_node: batch_label})\n            print(res)\n<\/code><\/pre>\n\n<p>Notice that I read all of image data into a numpy list, and throw them into my restore graph based on input node. But I got following error message:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"restore_graph_train.py\", line 127, in &lt;module&gt;\n    res = sess.run(train_op, feed_dict = {image_input_node: batch_data, label_node: batch_label})\n  File \"\/mnt\/ficusspain\/cqli\/virtual_env\/quantize_model\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 900, in run\n    run_metadata_ptr)\n  File \"\/mnt\/ficusspain\/cqli\/virtual_env\/quantize_model\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 1088, in _run\n    subfeed_dtype = subfeed_t.dtype.as_numpy_dtype\n  File \"\/mnt\/ficusspain\/cqli\/virtual_env\/quantize_model\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/dtypes.py\", line 128, in as_numpy_dtype\n    return _TF_TO_NP[self._type_enum]\nKeyError: 20\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52836670,"AnswerCount":1,"Tags":"<python><compilation><keras>","CreationDate":"2018-10-16T13:32:06.857","AcceptedAnswerId":52837571.0,"OwnerUserId":10502975.0,"Title":"This issue about compile keras model","Body":"<p>I'm trying to create a CNN in pycharm. When I run my code, the console outputs<\/p>\n\n<p><code>RuntimeError: You must compile your model before using it.<\/code><\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-o\nfrom keras.models import Sequential\nfrom keras.layers import Dense, MaxPool2D, Flatten, Conv2D, Dropout\nfrom keras.preprocessing import image\nfrom keras.optimizers import adadelta\ngenerator = image.ImageDataGenerator(  \n                rescale=1.\/255,\n                featurewise_center=False, \n                samplewise_center=False, \n                featurewise_std_normalization=False, \n                samplewise_std_normalization=False, \n                zca_whitening=False, \n                rotation_range=10,\n                width_shift_range=0.1,\n                height_shift_range=0.1,\n                horizontal_flip=True,\n                vertical_flip=False,\n               )\ndateset = generator.flow_from_directory(  \n               shuffle=True,\n               batch_size=100,\n               target_size=(80, 80),\n               directory='\/Users\/username\/Documents\/Project AI\/Date set\/blood-cells\/dataset2-master\/images\/TRAIN')\n\n\ndef get_model():\n        model = Sequential()\n        model.add(Conv2D(80, (3, 3), strides=(1, 1), activation='relu'))\n        model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='relu', \n                  input_shape=(80, 80, 3)))\n        model.add(MaxPool2D(pool_size=(2, 2)))\n        model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='relu'))\n        model.add(Dropout(0.25))\n        model.add(Flatten())\n        model.add(Dense(128, activation='relu'))\n        model.add(Dropout(0.5))\n        model.add(Dense(4, activation='softmax'))\n        model.compile(optimizer=adadelta(lr=0.001), \n        loss='categorical_crossentropy', metrics=['accuracy'])\n        return model\n\n\nnn = get_model()\nnn.fit_generator(dateset, steps_per_epoch=None, epochs=30, verbose=1)\n<\/code><\/pre>\n\n<p>And this is output of pycharm Python console:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"\/Applications\/PyCharm.app\/Contents\/helpers\/pydev\/_pydev_bundle\/pydev_umd.py\", line 197, in runfile\npydev_imports.execfile(filename, global_vars, local_vars)  # execute the script\n  File \"\/Applications\/PyCharm.app\/Contents\/helpers\/pydev\/_pydev_imps\/_pydev_execfile.py\", line 18, in execfile exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"\/Users\/username\/Documents\/Project AI\/Src\/CNN.py\", line 43, in &lt;module&gt;\nnn.fit_generator(dateset, steps_per_epoch=None, epochs=30, verbose=1)\n  File \"\/Users\/username\/Documents\/Project AI\/venv\/lib\/python3.7\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\nreturn func(*args, **kwargs)\n  File \"\/Users\/username\/Documents\/Project AI\/venv\/lib\/python3.7\/site-packages\/keras\/engine\/training.py\", line 1418, in fit_generator\ninitial_epoch=initial_epoch)\n  File \"\/Users\/username\/Documents\/Project AI\/venv\/lib\/python3.7\/sitepackages\/keras\/engine\/training_generator.py\", line 40, in fit_generator\nmodel._make_train_function()\n  File \"\/Users\/yangzichen\/Documents\/Project AI\/venv\/lib\/python3.7\/site-packages\/keras\/engine\/training.py\", line 496, in _make_train_function\n  raise RuntimeError('You must compile your model before using it.')\nRuntimeError: You must compile your model before using it.\n<\/code><\/pre>\n","answers":[{"AnswerId":"52837571","CreationDate":"2018-10-16T14:15:33.030","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>Seams like you are facing this known Bug:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18287\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18287<\/a><\/p>\n\n<p>Another possible solution is to define input_shape in your first layer, not in the second<\/p>\n"}]}
+{"QuestionId":52836995,"AnswerCount":0,"Tags":"<keras><tensor>","CreationDate":"2018-10-16T13:47:05.343","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"how to convert keras tensor to numpy array","Body":"<pre><code>input_feats = Input(shape=([fc_feats.shape[0]]))\ndef lam_reshape(inputs):\n    return inputs\nfc_feats_new = Lambda(lam_reshape)(input_feats)\nfc_feats_new = K.reshape(fc_feats_new, [10, int(fc_feats_new.shape[1])])\nfc_feats_new = sess.run(fc_feats_new)\n<\/code><\/pre>\n\n<p>got error:\nInvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_54' with dtype float and shape [?,2048]<\/p>\n","answers":[]}
+{"QuestionId":52837189,"AnswerCount":0,"Tags":"<tensorflow><cudnn>","CreationDate":"2018-10-16T13:56:20.930","AcceptedAnswerId":null,"OwnerUserId":10513311.0,"Title":"failed at installing tensorflow gpu","Body":"<h3>System information<\/h3>\n\n<ul>\n<li><strong>Have I written custom code (as opposed to using a stock example script provided in TensorFlow)<\/strong>:  failed at import tensorflow<\/li>\n<li><strong>OS Platform and Distribution<\/strong>: lsb_release -a<\/li>\n<li><strong>TensorFlow installed from<\/strong>: Anaconda<\/li>\n<li><strong>TensorFlow version (use command below)<\/strong>:  1.11.0<\/li>\n<li><strong>Python version<\/strong>: 3.6.6<\/li>\n<li><strong>Bazel version (if compiling from source)<\/strong>: not using<\/li>\n<li><strong>GCC\/Compiler version (if compiling from source)<\/strong>: [GCC 7.3.0] on linux<\/li>\n<li><strong>CUDA\/cuDNN version<\/strong>: 9.1.85 \/  not sure was automatically download<\/li>\n<li><strong>GPU model and memory<\/strong>: GeForce 840m<\/li>\n<\/ul>\n\n<h3>Describe the problem<\/h3>\n\n<p>Basically it doesn't work when i run \"import tensorflow\"<\/p>\n\n<h3>Source code \/ logs<\/h3>\n\n<pre><code>Python 3.6.6 |Anaconda, Inc.| (default, Oct  9 2018, 12:34:16)\n[GCC 7.3.0] on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\nTraceback (most recent call last):\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcuda.so.1: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/herooooooooo\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcuda.so.1: cannot open shared object file: No such file or directory\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52837203,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2018-10-16T13:56:56.820","AcceptedAnswerId":null,"OwnerUserId":6898433.0,"Title":"Combining LSTM along with Embedding layers in Keras","Body":"<p>I'm trying to design a network architecture that looks like this in R Keras:<\/p>\n\n<pre><code>Time Series data (LSTM) ---------------------\\\n                                              Concatenation ---&gt; Output\nCategorical variables (Embedding layers) ----\/\n<\/code><\/pre>\n\n<p>All this to have a time series data augmented by few categorical variables. Here's my code:<\/p>\n\n<pre><code>library(keras)\n\n# First embedding layer\ninput_1 &lt;- layer_input(shape = 1)\ninput_1_embedding &lt;- layer_embedding(input_1, 50, 5)\n\n# Second embedding layer\ninput_2 &lt;- layer_input(shape = 1)\ninput_2_embedding &lt;- layer_embedding(input_2, 100, 10) \n\n# LSTM layer\nlstm_input &lt;- layer_input(batch_shape = c(32, 5, 1))\nlstm &lt;- layer_lstm(lstm_input, 5, stateful = TRUE)\n\n# Concatenate both - the error happens here\noutputs &lt;- layer_concatenate(list(input_1_embedding, input_2_embedding, lstm)) %&gt;% \n  layer_dense(units = 1)\n\nmodel &lt;- keras_model(inputs = list(input_1, input_2, lstm_input), outputs = outputs)\n<\/code><\/pre>\n\n<p>I'm getting the following error <code>A Concatenate layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 1, 5), (None, 1, 10), (32, 5)]<\/code>.<\/p>\n\n<p>As far as I understand, this comes from a fact, that I'm training LSTM on batches of size 32 and the embedding layers are not batched. I tried to introduce <code>batch_shape<\/code> into the embedding layers but this causes errors while training, as not all of the batches contain all embeddings.<\/p>\n\n<p>I also tried to introduce batch shape after the embedding but with no luck so far. Is there any way to achieve mentioned network architecture?<\/p>\n","answers":[]}
+{"QuestionId":52837433,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><conv-neural-network>","CreationDate":"2018-10-16T14:07:49.490","AcceptedAnswerId":52838316.0,"OwnerUserId":4481339.0,"Title":"Extract max-sum submatrices","Body":"<p>I've a <code>2D<\/code> <code>NxN<\/code> matrix that has elements from a set of real numbers. I need to identify top <code>n<\/code> <code>DxD<\/code> sub-matrices from it such that their sum is maximum and return top left index of the sub-matrices. I need to do it in Tensorflow.<\/p>\n\n<p>For example I have following <code>4x4<\/code> matrix:<\/p>\n\n<pre><code>[1 1 4 4]\n[1 1 4 4]\n[3 3 2 2]\n[3 3 2 2]\n<\/code><\/pre>\n\n<p>I need to identify 2 sub-matrices that have the largest sum and return their top left index. In above case, 2 sub-matrices that have the largest and second largest sum are:<\/p>\n\n<pre><code>[[4 4]    [[3 3]\n [4 4]] &amp;  [3 3]]\n<\/code><\/pre>\n\n<p>I need to return <code>[[0,2],[2,0]]<\/code>, the top left indices to both the matrices. Thanks.<\/p>\n","answers":[{"AnswerId":"52838316","CreationDate":"2018-10-16T14:53:00.683","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can get that with the following snippet. The idea is to build a tensor holding the row and column indices of each element of each submatrix, then sum the submatrices and find the largest sums.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input data\ninput = tf.placeholder(tf.int32, [None, None])\n# Submatrix dimension\ndims = tf.placeholder(tf.int32, [2])\n# Number of top submatrices to find\nk = tf.placeholder(tf.int32, [])\n# Sizes\ninput_shape = tf.shape(input)\nrows, cols = input_shape[0], input_shape[1]\nd_rows, d_cols = dims[0], dims[1]\nsubm_rows, subm_cols = rows - d_rows + 1, cols - d_cols + 1\n# Index grids\nii, jj = tf.meshgrid(tf.range(subm_rows), tf.range(subm_cols), indexing='ij')\nd_ii, d_jj = tf.meshgrid(tf.range(d_rows), tf.range(d_cols), indexing='ij')\n# Add indices\nsubm_ii = ii[:, :, tf.newaxis, tf.newaxis] + d_ii\nsubm_jj = jj[:, :, tf.newaxis, tf.newaxis] + d_jj\n# Make submatrices tensor\nsubm = tf.gather_nd(input, tf.stack([subm_ii, subm_jj], axis=-1))\n# Add submatrices\nsubm_sum = tf.reduce_sum(subm, axis=(2, 3))\n# Use TopK to find top submatrices\n_, top_idx = tf.nn.top_k(tf.reshape(subm_sum, [-1]), tf.minimum(k, tf.size(subm_sum)))\n# Get row and column\ntop_row = top_idx \/\/ subm_cols\ntop_col = top_idx % subm_cols\nresult = tf.stack([top_row, top_col], axis=-1)\n\n# Test\nwith tf.Session() as sess:\n    mat = [\n        [1, 1, 4, 4],\n        [1, 1, 4, 4],\n        [3, 3, 2, 2],\n        [3, 3, 2, 2],\n    ]\n    print(sess.run(result, feed_dict={input: mat, dims: [2, 2], k: 2}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[0 2]\n [1 2]]\n<\/code><\/pre>\n\n<p>Note that the output in this case is <code>[0, 2]<\/code> and <code>[1, 2]<\/code>, but not <code>[2, 0]<\/code>. That's because the submatrix starting at <code>[1, 2]<\/code> sums the same amount as the one at <code>[2, 0]<\/code>, and it is before in the matrix, if you iterate it by rows. If you pass <code>k: 3<\/code> in the test you would get <code>[2, 0]<\/code> too in the result.<\/p>\n"}]}
+{"QuestionId":52837867,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-16T14:31:18.893","AcceptedAnswerId":null,"OwnerUserId":7221127.0,"Title":"Reuse of computed Constant Values in Tensorflow","Body":"<p>I have encountered the following behavior and I'm not quite sure if this is a bug or if I'm missing out on something.<\/p>\n\n<h2>Minimal Example<\/h2>\n\n<pre><code># Create constant Matrix A\nA1 = tf.convert_to_tensor(A, dtype=tf.float64)\n\n# Compute the cholesky decomposition\nsqrtA1 = tf.linalg.cholesky(tfA1)\n\nxi = tf.placeholder(tf.float64, shape=[10000])\n\n# Matrix multiplication with the chompsky matrix\nRessqrtA1 = tf.tensordot(tfsqrtA1, tfxi, [[1],[0]])\n\n# Regular Matrix multiplication\nResA1 = tf.tensordot(tfA1, tfxi, [[1],[0]])\n<\/code><\/pre>\n\n<p>Upon Computing something like this: <\/p>\n\n<pre><code>with tf.Session() as sess:\n    _ = sess.run(ResA1, feed_dict={xi: np.random.randn(10000)})\n    _ = sess.run(ResA1, feed_dict={xi: np.random.randn(10000)})\n<\/code><\/pre>\n\n<p>One could assume that the first <code>sess.run()<\/code> will take more time since the Graph has to be build and optimized. And that is what happens, here are some timings just like expected:<\/p>\n\n<pre><code>normal matmul\nWall time: 36.1 s -- first run \n\n41.1 ms \u00b1 2.15 ms per loop (mean \u00b1 std. dev. of 7 runs, 10 loops each) \n--second run with timit\n<\/code><\/pre>\n\n<h2>Long Calculation times despite a constant Operation<\/h2>\n\n<p>Now one might think that the same holds when multiplying the <code>sqrtA1<\/code>, but <strong>unfortunately its not<\/strong>.<br>\nComputing the following:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    _ = sess.run(RessqrtA1, feed_dict={xi: np.random.randn(10000)})\n    _ = sess.run(RessqrtA1, feed_dict={xi: np.random.randn(10000)})\n<\/code><\/pre>\n\n<p>Here are some timings: <\/p>\n\n<pre><code>cholesky\nWall time: 2min 38s\n24.7 s \u00b1 750 ms per loop (mean \u00b1 std. dev. of 7 runs, 1 loop each)\n<\/code><\/pre>\n\n<p>I ran the TensorFlow Profiler and obtained this information: <\/p>\n\n<pre><code>Profile:\nnode name | requested bytes | total execution time | accelerator execution time | cpu execution time\nCholesky                     800.00MB (100.00%, 50.00%),     24.30sec (100.00%, 99.81%),             0us (0.00%, 0.00%),     24.30sec (100.00%, 99.81%)\nMatMul                          80.00KB (50.00%, 0.00%),         46.45ms (0.19%, 0.19%),             0us (0.00%, 0.00%),         46.45ms (0.19%, 0.19%)\nReshape                               0B (0.00%, 0.00%),            15us (0.00%, 0.00%),             0us (0.00%, 0.00%),            15us (0.00%, 0.00%)\nConst                         800.00MB (50.00%, 50.00%),            10us (0.00%, 0.00%),             0us (0.00%, 0.00%),            10us (0.00%, 0.00%)\n<\/code><\/pre>\n\n<h2>Difference between TF 1.7 and TF 1.6<\/h2>\n\n<p>Now for the <strong>interesting Part<\/strong>: This is <strong>only<\/strong> when you use <strong>Version 1.7.0 or higher<\/strong><br>\nI tried this with Version 1.6.0 and the Matmuls with the cholesky matrix took not longer than a normal matmul: <\/p>\n\n<pre><code>Wall time: 1min 18s -- first run\n49.7 ms \u00b1 1.02 ms per loop (mean \u00b1 std. dev. of 7 runs, 10 loops each)\n-- timeit on the second sess.run()\n<\/code><\/pre>\n\n<p>When looking into the Computation in Version 1.6 with the Profiler you get: <\/p>\n\n<pre><code>Profile:node name | requested bytes | total execution time | accelerator execution time | cpu execution time\nCholesky                    800.00MB (100.00%, 100.00%),       891us (100.00%, 100.00%),             0us (0.00%, 0.00%),       891us (100.00%, 100.00%)\n======================End of Report==========================\n<\/code><\/pre>\n\n<p>Lastely, using this code, while replacing <code>RessqrtA1 = tf.tensordot(chol, tfxi, [[1],[0]])<\/code>,where <code>chol<\/code> is another <strong>Placeholder<\/strong>, gives us the desired outcome again.<\/p>\n\n<pre><code>%time temp = sess.run(tfsqrtA1)\n%timeit _ = sess.run(tfRessqrtA1, feed_dict={tfxi: np.random.randn(10000), chol: temp})\n\ncholesky\nWall time: 1min 40s\n42.6 ms \u00b1 3.86 ms per loop (mean \u00b1 std. dev. of 7 runs, 1 loop each)\n<\/code><\/pre>\n\n<h2>My Thoughts:<\/h2>\n\n<p>Somehow since Version 1.7 for every session run the cholesky calculation happens despite being from constant origin and thus it would suffice to calculate it once and reuse the outcome.  <\/p>\n\n<p>I think that such a computation happens in version 1.6 during the construction of the Graph as I measured that the <strong>calculation of the Cholesky Decomposition<\/strong> takes about <strong>26sec in V1.6<\/strong> (used a Placeholder so that it needs to be calculated during runtime), yet during the run it only took 900us as shown above.  <\/p>\n\n<p>In v1.7 they somehow changed this behavior. <\/p>\n\n<h2>The Question<\/h2>\n\n<p>Do you think this behavior change is normal or desired? If so, what would be a better approach when you have a similar situation? <\/p>\n\n<p>I'm thankful for every information, maybe someone else noticed some related problems using the linear algebra class. <\/p>\n\n<h2>What i have tried so far:<\/h2>\n\n<ul>\n<li>tinkering around with the float size (i.e. <code>tf.float32<\/code> instead of <code>tf.float64<\/code>)<\/li>\n<li>used different Versions in seperate environments.<\/li>\n<li>used the profiler to investigate the times of different Ops<\/li>\n<li>ran on a different pc <\/li>\n<li>timing in Python (e.g. <code>timeit<\/code> or <code>pycallgraph<\/code>) <\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":52838304,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T14:52:26.710","AcceptedAnswerId":null,"OwnerUserId":9965058.0,"Title":"ValueError: ('%s is not decorated with @add_arg_scope', ('__main__', 'bottleneck'))","Body":"<p>Here is the code<\/p>\n\n<pre><code>with tf.variable_scope(scope, 'resnet_v2', [inputs], reuse=reuse) as sc:\n    end_points_collection = sc.original_name_scope + '_end_points'\n    with slim.arg_scope([slim.conv2d, bottleneck, stack_blocks_dense],\n                        outputs_collections=[end_points_collection]):\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: ('%s is not decorated with @add_arg_scope', ('<strong>main<\/strong>', 'bottleneck'))\n  d:\\resnet\\main_resnet.py(219)resnet_v2() \n  -> outputs_collections=[end_points_collection])<\/p>\n<\/blockquote>\n\n<p>So what is wrong with the code?<\/p>\n","answers":[{"AnswerId":"53038290","CreationDate":"2018-10-29T03:18:41.107","ParentId":null,"OwnerUserId":"9965058","Title":null,"Body":"<p>forgive my stupid question, just decorates the function with @add_arg_scope before the definition of bottleneck<\/p>\n\n<pre><code>@add_arg_scope\ndef bottleneck():\n<\/code><\/pre>\n"}]}
+{"QuestionId":52838355,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2018-10-16T14:55:23.080","AcceptedAnswerId":null,"OwnerUserId":7091449.0,"Title":"framework caffe does not work when testing","Body":"<p>I am trying to use caffe test in order to check out the accuracy of caffennet. Pretty simple, isn't it?<\/p>\n\n<p>Really NOT. It doesn't work at all.<\/p>\n\n<p>Below I describe the steps I am following (To my mind everythink is OK):<\/p>\n\n<ol>\n<li>I have downloaded imagenet training and validation set.<\/li>\n<li><p>I have used the app \"create_imagenet.sh\" to create a LMDB database. Here I depict my \"create_imagenet.sh\" (I have modified some things since I only need validation database):<\/p>\n\n<pre><code>    RESIZE=true\n    if $RESIZE; then\n      RESIZE_HEIGHT=256\n      RESIZE_WIDTH=256\n    else\n      RESIZE_HEIGHT=0\n      RESIZE_WIDTH=0\n    fi\n\nif [ ! -d \"$VAL_DATA_ROOT\" ]; then\n  echo \"Error: VAL_DATA_ROOT is not a path to a directory: $VAL_DATA_ROOT\"\n  echo \"Set the VAL_DATA_ROOT variable in create_imagenet.sh to the path\" \\\n       \"where the ImageNet validation data is stored.\"\n  exit 1\nfi\n\necho \"Creating train lmdb...\"\n\nGLOG_logtostderr=1 $TOOLS\/convert_imageset \\\n    -resize_height=$RESIZE_HEIGHT \\\n    -resize_width=$RESIZE_WIDTH \\\n    --shuffle \\\n    $VAL_DATA_ROOT \\\n    $DATA\/val.txt \\\n    $EXAMPLE\/ilsvrc12_val_leveldb\n\necho \"Done.\"\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>in $VAL_DATA_ROOT I storage the path to the validation images, in $DATA the path tu val.txt file and finally example the path when I want to save my database. I can MAKE SURE that the paths are totally correct.<\/p>\n\n<ol start=\"3\">\n<li>Once I have my database I have downloaded the weights file bvlc_reference_caffenet.caffemodel from official website and  it's OK.<\/li>\n<li><p>I have modified train_val.prototxt to link the database. Next I depict the first lines of the file. The rest does not change at all (respect to the original).<\/p>\n\n<pre><code># mean pixel \/ channel-wise mean instead of mean image\n#  transform_param {\n#    crop_size: 227\n#    mean_value: 104\n#    mean_value: 117\n#    mean_value: 123\n#    mirror: true\n#  }\n  data_param {\n    source: \"\/home\/paco\/Caffe_Original\/caffe\/examples\/imagenet\/ilsvrc12_train_lmdb\"\n    batch_size: 256\n    backend: LMDB\n  }\n}\nlayer {\n  name: \"data\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TEST\n  }\n#  transform_param {\n#    mirror: false\n#    crop_size: 227\n#    mean_file: \"\/home\/paco\/Caffe\/caffe\/data\/ilsvrc12\/imagenet_mean.binaryproto\"\n#  }\n# mean pixel \/ channel-wise mean instead of mean image\n  transform_param {\n    crop_size: 227\n    mean_value: 104\n    mean_value: 117\n    mean_value: 123\n    mirror: false\n  }\n  data_param {\n    source: \"\/home\/paco\/imagenet_noformat\/ilsvrc12_val_lmdb\"\n    batch_size: 50\n    backend: LMDB\n  }\n}\n<\/code><\/pre><\/li>\n<li><p>I just execute  caffe tool as follows:<\/p>\n\n<p>caffe test -model=\/home\/paco\/Caffe\/caffe\/models\/bvlc_reference_caffenet\/train_val.prototxt -weights=\/home\/paco\/Caffe\/caffe\/models\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel -iterations=2000<\/p><\/li>\n<\/ol>\n\n<p>As you all can see everything looks PERFECT and I don't obtain any error at all!!<\/p>\n\n<p>However, it does not work since I obtain this on my shell!!:<\/p>\n\n<p>**<\/p>\n\n<pre><code>I1016 16:43:32.526727 22178 net.cpp:255] Network initialization done.\nI1016 16:43:32.675029 22178 upgrade_proto.cpp:46] Attempting to upgrade input file specified using deprecated transformation parameters: \/home\/paco\/Caffe\/caffe\/models\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel\nI1016 16:43:32.675051 22178 upgrade_proto.cpp:49] Successfully upgraded file specified using deprecated data transformation parameters.\nW1016 16:43:32.675071 22178 upgrade_proto.cpp:51] Note that future Caffe releases will only support transform_param messages for transformation fields.\nI1016 16:43:32.675122 22178 upgrade_proto.cpp:55] Attempting to upgrade input file specified using deprecated V1LayerParameter: \/home\/paco\/Caffe\/caffe\/models\/bvlc_reference_caffenet\/bvlc_reference_caffenet.caffemodel\nI1016 16:43:32.882395 22178 upgrade_proto.cpp:63] Successfully upgraded file specified using deprecated V1LayerParameter\nI1016 16:43:32.943102 22178 caffe.cpp:281] Running for 2000 iterations.\nI1016 16:43:34.434036 22178 caffe.cpp:304] Batch 0, accuracy = 1\nI1016 16:43:34.434161 22178 caffe.cpp:304] Batch 0, loss = -nan\nI1016 16:43:35.862941 22178 caffe.cpp:304] Batch 1, accuracy = 1\nI1016 16:43:35.862982 22178 caffe.cpp:304] Batch 1, loss = -nan\nI1016 16:43:37.310729 22178 caffe.cpp:304] Batch 2, accuracy = 1\nI1016 16:43:37.310768 22178 caffe.cpp:304] Batch 2, loss = -nan\nI1016 16:43:38.719712 22178 caffe.cpp:304] Batch 3, accuracy = 1\nI1016 16:43:38.719833 22178 caffe.cpp:304] Batch 3, loss = -nan\nI1016 16:43:40.171242 22178 caffe.cpp:304] Batch 4, accuracy = 1\nI1016 16:43:40.171273 22178 caffe.cpp:304] Batch 4, loss = -nan\nI1016 16:43:41.674332 22178 caffe.cpp:304] Batch 5, accuracy = 1\nI1016 16:43:41.674373 22178 caffe.cpp:304] Batch 5, loss = -nan\nI1016 16:43:43.160789 22178 caffe.cpp:304] Batch 6, accuracy = 1\nI1016 16:43:43.160826 22178 caffe.cpp:304] Batch 6, loss = -nan\n<\/code><\/pre>\n\n<p>**<\/p>\n\n<p>This is even worse with alexnet and Squeezenet where following the same steps I ALWAYS obtain 0 of accuracy (really times to times I get 0.001 or something like that).<\/p>\n\n<p>WHAT IS HAPPENENING? Maybe I have to preprocess the images before dump it into the database? ANYBODY CAN HELP ME? I WOULD REALLY APPRECIATE IT.<\/p>\n\n<p>Thank you so much for your attention and help.<\/p>\n","answers":[{"AnswerId":"52870378","CreationDate":"2018-10-18T08:53:30.617","ParentId":null,"OwnerUserId":"7091449","Title":null,"Body":"<p>The problem was MKL library. It was not working yet I was not given any information at all. I have fixed the problem by compiling caffe all over again using OpenBlas library. <\/p>\n\n<p>Thank you to myself.<\/p>\n"}]}
+{"QuestionId":52838888,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T15:21:51.567","AcceptedAnswerId":52838979.0,"OwnerUserId":9540764.0,"Title":"Add a tensor only to a part of another tensor","Body":"<p>I have to add two tensors, one with a shape multiple of the other in the depth direction. Here an example<\/p>\n\n<pre><code>t1 = tf.constant(3, shape=[2, 2, 2], dtype=tf.float32)\nt2 = tf.constant(1, shape=[2, 2, 1], dtype=tf.float32)\n<\/code><\/pre>\n\n<p>I want to use something like <code>tf.add<\/code> to add the second tensor to the first but only in the first layer of the third component of the shape. With numbers<\/p>\n\n<pre><code>t1 = [[[3, 3], [3, 3]],\n      [[3, 3], [3, 3]]]\nt2 = [[[1, 1], [1, 1]]]\n\noutput = [[[4, 4], [4, 4]],\n          [[3, 3], [3, 3]]]\n<\/code><\/pre>\n\n<p>Is there a built-in function to do that?<\/p>\n","answers":[{"AnswerId":"52838979","CreationDate":"2018-10-16T15:26:40.843","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p>Add the first '<em>column<\/em>' of <code>t1<\/code> with <code>t2<\/code> and then concat it with the rest columns of <code>t1<\/code>:<\/p>\n\n<pre><code>t1 = tf.constant(3, shape=[2, 2, 2], dtype=tf.float32)\nt2 = tf.constant(1, shape=[2, 2, 1], dtype=tf.float32)\ntf.InteractiveSession()\n\ntf.concat((t1[...,0:1] + t2, t1[...,1:]), axis=2).eval()\n\n#array([[[4., 3.],\n#        [4., 3.]],\n\n#       [[4., 3.],\n#        [4., 3.]]], dtype=float32)\n<\/code><\/pre>\n\n<p>Notice your second example <code>t2<\/code> has a different shape, i.e. <code>(1,2,2)<\/code> instead of <code>(2,2,1)<\/code>, in which case, slice and concat by the first axis:<\/p>\n\n<pre><code>tf.concat((t1[0:1] + t2, t1[1:]), axis=0).eval()\n\n#array([[[4., 4.],\n#        [4., 4.]],\n\n#       [[3., 3.],\n#        [3., 3.]]], dtype=float32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52839368,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api><mscoco>","CreationDate":"2018-10-16T15:47:26.417","AcceptedAnswerId":53121043.0,"OwnerUserId":1836859.0,"Title":"Understanding COCO evaluation \"maximum detections\"","Body":"<p>I started using the <a href=\"https:\/\/github.com\/cocodataset\/cocoapi\" rel=\"noreferrer\">cocoapi<\/a> to evaluate a model trained using the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"noreferrer\">Object Detection API<\/a>.\nAfter reading various sources that explain mean average precision (mAP) and recall, I am confused with the \"maximum detections\" paramter used in the cocoapi.<\/p>\n\n<p>From what I understood (e.g. <a href=\"http:\/\/tarangshah.com\/blog\/2018-01-27\/what-is-map-understanding-the-statistic-of-choice-for-comparing-object-detection-models\/\" rel=\"noreferrer\">here<\/a>, <a href=\"https:\/\/deshanadesai.github.io\/notes\/Evaluation-of-Results-using-Mean-Avg-Precision\" rel=\"noreferrer\">here<\/a> or <a href=\"https:\/\/medium.com\/@jonathan_hui\/map-mean-average-precision-for-object-detection-45c121a31173\" rel=\"noreferrer\">here<\/a>), one calculates mAP by calculating precision and recall for various model score thresholds. This gives the precision-recall curve and mAP is calculated as an approximation to the area under this curve. Or, expressed differently, as the average of the maximum precision in defined recall ranges (0:0.1:1).<\/p>\n\n<p>However, the cocoapi seems to calculate precision and recall for a given number of maximum detections (<code>maxDet<\/code>) with the highest scores. And from there get the precision-recall curve for <code>maxDets = 1, 10, 100<\/code>. Why is this a good metric since it is clearly not the same as the above method (it potentially excludes datapoints)?<\/p>\n\n<p>In my example, I have ~ 3000 objects per image. Evaluating the result using the cocoapi gives terrible recall because it limits the number of detected objects to 100.<\/p>\n\n<p>For testing purposes, I feed the evaluation dataset as the ground truth and the detected objects (with some artificial scores). I would expect precision and recall pretty good, which is actually happening. But as soon as I feed in more than 100 objects, precision and recall go down with increasing number of \"detected objects\". Even though they are all \"correct\"! How does that make sense?<\/p>\n","answers":[{"AnswerId":"53121043","CreationDate":"2018-11-02T15:02:52.040","ParentId":null,"OwnerUserId":"1836859","Title":null,"Body":"<p>I came to the conclusion, that this is just the way that the cocoapi defines its metric. It probably makes sense in their context, but I can as well define my own (which is what I did), based on the articles I read and linked above.<\/p>\n"},{"AnswerId":"56124891","CreationDate":"2019-05-14T07:15:56.780","ParentId":null,"OwnerUserId":"3134408","Title":null,"Body":"<p>You can change the <code>maxDets<\/code> parameter and define a new <code>summarize()<\/code> instance method. <\/p>\n\n<p>Let's create a <code>COCOeval<\/code> object:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>cocoEval = COCOeval(cocoGt,cocoDt,annType)\ncocoEval.params.maxDets = [200]\ncocoEval.params.imgIds  = imgIdsDt\ncocoEval.evaluate()\ncocoEval.accumulate()\ncocoEval.summarize_2() # instead of calling cocoEval.summarize()\n<\/code><\/pre>\n\n<p>Now, define <code>summarize_2()<\/code> method in <code>cocoeval.py<\/code> module in the following way:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def summarize_2(self):\n    # Copy everything from `summarize` method here except\n    # the function `_summarizeDets()`.\n    def _summarizeDets():\n        stats = np.zeros((12,))\n        stats[0] = _summarize(1, maxDets=self.params.maxDets[0])\n        stats[1] = _summarize(1, iouThr=.5, maxDets=self.params.maxDets[0])\n        stats[2] = _summarize(1, iouThr=.75, maxDets=self.params.maxDets[0])\n        stats[3] = _summarize(1, areaRng='small', maxDets=self.params.maxDets[0])\n        stats[4] = _summarize(1, areaRng='medium', maxDets=self.params.maxDets[0])\n        stats[5] = _summarize(1, areaRng='large', maxDets=self.params.maxDets[0])\n        stats[6] = _summarize(0, maxDets=self.params.maxDets[0])\n        stats[9] = _summarize(0, areaRng='small', maxDets=self.params.maxDets[0])\n        stats[10] = _summarize(0, areaRng='medium', maxDets=self.params.maxDets[0])\n        stats[11] = _summarize(0, areaRng='large', maxDets=self.params.maxDets[0])\n        return stats\n    # Copy other things which are left from `summarize()` here.\n<\/code><\/pre>\n\n<p>If you run the above method over your dataset, you will get an output similar to this:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code> Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=200 ] = 0.507\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=200 ] = 0.699\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=200 ] = 0.575\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=200 ] = 0.586\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=200 ] = 0.519\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=200 ] = 0.501\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=200 ] = 0.598\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=200 ] = 0.640\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=200 ] = 0.566\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=200 ] = 0.564\n<\/code><\/pre>\n"}]}
+{"QuestionId":52839911,"AnswerCount":1,"Tags":"<python><keras><anaconda><travis-ci><pytest>","CreationDate":"2018-10-16T16:20:48.500","AcceptedAnswerId":53104854.0,"OwnerUserId":497600.0,"Title":"how to run pytest for keras sourcecode","Body":"<p>I'm developing a custom layer for the Keras framework and would like to extend their existing testsuite.<\/p>\n\n<p>Since I'm working on the keras sourcecode, I have installed it like this:<\/p>\n\n<ul>\n<li><code>pip install keras-preprocessing<\/code><\/li>\n<li><code>pip install keras-applications<\/code><\/li>\n<li><code>git clone https:\/\/github.com\/keras-team\/keras<\/code><\/li>\n<li><code>cd keras<\/code><\/li>\n<li><code>export PYTHONPATH=$PWD:$PYTHONPATH<\/code><\/li>\n<\/ul>\n\n<p>After these commands, you are in the cloned keras repository, this is the working directory for the code samples below.<\/p>\n\n<p>To begin with, I would like to run the existing tests and see what they do. <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/590aa832680641218b233b22fa8b3bd7fdc0287e\/tests\/keras\/layers\/wrappers_test.py#L636\" rel=\"nofollow noreferrer\">It looks like<\/a> they can simply be run as a python file:<\/p>\n\n<pre><code>if __name__ == '__main__':\n    pytest.main([__file__])\n<\/code><\/pre>\n\n<p>But this: <\/p>\n\n<pre><code>python tests\/keras\/layers\/wrappers_test.py\n<\/code><\/pre>\n\n<p>produces the following output, and doesn't run any tests<\/p>\n\n<pre><code>Using TensorFlow backend.\nusage: wrappers_test.py [options] [file_or_dir] [file_or_dir] [...]\nwrappers_test.py: error: unrecognized arguments: -n tests\/keras\/layers\/wrappers_test.py\n  inifile: \/home\/lhk\/programming\/keras\/pytest.ini\n  rootdir: \/home\/lhk\/programming\/keras\n<\/code><\/pre>\n\n<p>So I tried explicitly calling pytest next: <\/p>\n\n<pre><code>pytest tests\/keras\/layers\/wrappers_test.py\n<\/code><\/pre>\n\n<p>Exactly the same response (without the tensorflow log):<\/p>\n\n<pre><code>usage: pytest [options] [file_or_dir] [file_or_dir] [...]\npytest: error: unrecognized arguments: -n tests\/keras\/layers\/wrappers_test.py\n  inifile: \/home\/lhk\/programming\/keras\/pytest.ini\n  rootdir: \/home\/lhk\/programming\/keras\n<\/code><\/pre>\n\n<p>I'm using PyCharm, and if I run the tests from within pycharm (which I admittedly tried first), it only produces the messages above.<\/p>\n\n<p>Apparently, I'm not configuring it correctly. Pytest is not able to pick up the test suite. To find a reference configuration, I looked at the Keras CI setup. They use Travis and the config is opensource: <a href=\"https:\/\/travis-ci.org\/keras-team\/keras\/jobs\/442252422\/config\" rel=\"nofollow noreferrer\">https:\/\/travis-ci.org\/keras-team\/keras\/jobs\/442252422\/config<\/a><\/p>\n\n<p>It looks like I have all the dependencies installed. And the actual test command is basically what I have already tried:<\/p>\n\n<pre><code>PYTHONPATH=$PWD:$PYTHONPATH py.test tests\/ --ignore=tests\/integration_tests --ignore=tests\/test_documentation.py --ignore=tests\/keras\/legacy\/layers_test.py --cov-config .coveragerc --cov=keras tests\/\n<\/code><\/pre>\n\n<p>It produces the exact same output as above.<\/p>\n\n<p>I assume that the problem lies with pytest. But the installation section of their automated testing simply shows <code>pip install pytest pytest-pip8<\/code>. I ran this, but sure enough, <code>requirement already satisfied<\/code>.<\/p>\n\n<p>How can I execute the keras pytests.\nI'm running ubuntu 18.04.1, python 3.6.5 with anaconda 64bit.<\/p>\n","answers":[{"AnswerId":"53104854","CreationDate":"2018-11-01T15:55:53.973","ParentId":null,"OwnerUserId":"2650249","Title":null,"Body":"<p>Running the <code>keras<\/code> tests is mentioned on the <a href=\"https:\/\/keras.io\/contributing\/\" rel=\"nofollow noreferrer\">Contributing<\/a> page:<\/p>\n\n<blockquote>\n  <p>Run our test suite locally. It's easy: from the Keras folder, simply run: <code>py.test tests\/<\/code>.<\/p>\n  \n  <p><sup>You will need to install the test requirements as well: <code>pip install -e .[tests]<\/code>.<\/sup><\/p>\n<\/blockquote>\n\n<p>If you don't want to install the package in editable mode, just install all of the test dependencies explicitly. <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/2.2.4\/setup.py#L44\" rel=\"nofollow noreferrer\">Looking at <code>keras<\/code>' setup script<\/a>, the command for that would be:<\/p>\n\n<pre><code>$ pip install pytest pytest-pep8 pytest-xdist pytest-cov pytest-timeout pandas requests\n<\/code><\/pre>\n"}]}
+{"QuestionId":52840347,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><deep-learning><convergence>","CreationDate":"2018-10-16T16:47:10.460","AcceptedAnswerId":null,"OwnerUserId":4694075.0,"Title":"DDPG - Actor Critic Network does not converge with continous Actionspace","Body":"<p>I'm currently a bit confused. I implemented an Actor-Critic Network and depending on the setup ist either begins to converge a little but produces values far from right. Or it produces the nearly the same loss values over and over again but right from the start produces values which are kind of correct.\nI really have no clue how that is possible.\nThis is my current Model which produces values but does not converge:<\/p>\n\n<pre><code>def create_actor_model(self):\n    state_input = Input(shape=self.observation_space_shape)\n    h1 = Dense(18, activation='linear')(state_input)\n    h1l = LeakyReLU(alpha=0.01)(h1)\n    h3 = Dense(18, activation='tanh')(h1l)\n    h3n = BatchNormalization()(h3)\n    output = Dense(self.action_space_shape[0], activation='tanh')(h3n)\n    model = Model(input=state_input, output=output)\n    adam = Adam(lr=self.action_space_shape)\n    model.compile(loss=\"mse\", optimizer=adam)\n    return state_input, model\n\ndef create_critic_model(self):\n    state_input = Input(shape=self.observation_space_shape)\n    state_h1 = Dense(18, activation='relu')(state_input)\n    state_h2 = Dense(36)(state_h1)\n\n    action_input = Input(shape=self.action_space_shape)\n    action_h1 = Dense(36)(action_input)\n\n    merged = Add()([state_h2, action_h1])\n    l_h1 = LeakyReLU(alpha=0.01)(merged)\n    merged_h1 = Dense(18, activation='tanh')(l_h1)\n    h1n = BatchNormalization()(merged_h1)\n    output = Dense(1, activation='tanh')(h1n)\n    model = Model(input=[state_input, action_input], output=output)\n\n    adam = Adam(lr=self.action_space_shape)\n    model.compile(loss=\"mse\", optimizer=adam, metrics=['mae', 'mse', 'msle'])\n    return state_input, action_input, model\n\n\n def _train_actor_batch(self, batch_size, s_batch, a_batch, r_batch, s2_batch):\n    predicted_action = self.actor_model.predict_on_batch(s_batch)\n    grads = self.sess.run(self.critic_grads, feed_dict={\n        self.critic_state_input: s_batch,\n        self.critic_action_input: predicted_action\n    })\n\n    self.sess.run(self.optimize, feed_dict={\n        self.actor_state_input: s_batch,\n        self.actor_critic_grad: grads[0]\n    })\n\ndef _train_critic_batch(self, batch_size, s_batch, a_batch, r_batch, s2_batch):\n    target_action = self.target_actor_model.predict_on_batch(s2_batch)\n    future_reward = self.target_critic_model.predict_on_batch([s2_batch, target_action])\n\n    rewards = []\n    for k in range(batch_size):\n        this_future_reward = future_reward[k] if batch_size &gt; 1 else future_reward\n        rewards.append(r_batch[k] + self.gamma * this_future_reward)\n\n    return self.critic_model.train_on_batch([s_batch, a_batch], np.reshape(rewards, batch_size))\n\ndef replay(self, batch_size):\n    memory_length = len(self.memory)\n    if memory_length &lt; batch_size:\n        samples = random.sample(self.memory, memory_length)\n    else:\n        samples = random.sample(self.memory, batch_size)\n\n    s_batch = np.array([cur_state[0] for cur_state, _, _, _ in samples])\n    a_batch = np.array([float(action[0]) for _, action, _, _ in samples])\n    r_batch = np.array([reward[0] for _, _, reward, _ in samples])\n    s2_batch = np.array([new_state[0] for _, _, _, new_state in samples])\n\n    critic_loss = self._train_critic_batch(len(s_batch), s_batch, a_batch, r_batch, s2_batch)\n    self._train_actor_batch(len(s_batch), s_batch, a_batch, r_batch, s2_batch)\n\n    self.update_target()\n    return critic_loss\n\ndef _update_actor_target(self):\n    actor_model_weights = self.actor_model.get_weights()\n    actor_target_weights = self.target_actor_model.get_weights()\n\n    for i in range(len(actor_target_weights)):\n        actor_target_weights[i] = actor_model_weights[i] * self.tau + actor_target_weights[i] * (1 - self.tau)\n    self.target_actor_model.set_weights(actor_target_weights)\n\ndef _update_critic_target(self):\n    critic_model_weights = self.critic_model.get_weights()\n    critic_target_weights = self.target_critic_model.get_weights()\n\n    for i in range(len(critic_target_weights)):\n        critic_target_weights[i] = critic_model_weights[i] * self.tau + critic_target_weights[i] * (1 - self.tau)\n    self.target_critic_model.set_weights(critic_target_weights)\n\ndef update_target(self):\n    self._update_actor_target()\n    self._update_critic_target()\n\n\n\ndef __init__(self):\n    self.memory = deque(maxlen=2000)\n    self.actor_state_input, self.actor_model = self.create_actor_model()\n    _, self.target_actor_model = self.create_actor_model()\n\n    self.actor_critic_grad = tf.placeholder(tf.float32, [None, self.action_space_shape[0]])\n\n    actor_model_weights = self.actor_model.trainable_weights\n    self.actor_grads = tf.gradients(self.actor_model.output,\n                                    actor_model_weights, -self.actor_critic_grad)\n    grads = zip(self.actor_grads, actor_model_weights)\n    self.optimize = tf.train.AdamOptimizer(self.learning_rate).apply_gradients(grads)\n\n    self.critic_state_input, self.critic_action_input, \\\n    self.critic_model = self.create_critic_model()\n    _, _, self.target_critic_model = self.create_critic_model()\n\n    self.critic_grads = tf.gradients(self.critic_model.output,\n                                     self.critic_action_input)\n\n    self.sess.run(tf.initialize_all_variables())\n<\/code><\/pre>\n\n<p>And then i am training with the following method which is called for each epoch (at the end, the memory which is getting cleared is the experience-replay memory):<\/p>\n\n<pre><code>def train(self, states, epoch, env, is_new_epoch):\n    train_size = int(len(states) * 0.70)\n\n    train = dict(list(states.items())[0:train_size])\n    test = dict(list(states.items())[train_size:len(states)])\n    with warnings.catch_warnings():\n        warnings.simplefilter(\"ignore\")\n\n        working_states = copy(train)\n\n        critic_eval = list()\n        rewards = dict()\n        for last_day, (last_state_vec, _, last_action) in working_states.items():\n            this_day = last_day + timedelta(days=1)\n            if this_day in working_states:\n                (new_state_vec, _, _) = working_states.get(this_day)\n                rewards[last_day] = env.get_reward_by_states(last_state_vec, new_state_vec)\n\n        amt = len(working_states)\n        i = 0\n        for last_day, (last_state_vec, _, last_action) in working_states.items():\n            i+= 1\n            this_day = last_day + timedelta(days=1)\n            if this_day in working_states:\n                (new_state_vec, _, _) = working_states.get(this_day)\n                reward = np.reshape(rewards[last_day], [1, ])\n                self.remember(last_state_vec, [last_action], reward, new_state_vec)\n                new_eval = self.replay(env.batch_size)\n                critic_eval.append(new_eval)\n         self.memory.clear()\n<\/code><\/pre>\n\n<p>These are the loss values i got over 15 epochs:\n<a href=\"https:\/\/i.stack.imgur.com\/lEsKY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lEsKY.png\" alt=\"Actor Critic loss values over 15 epochs\"><\/a><\/p>\n\n<p>One Sample as it comes from the memory:<\/p>\n\n<pre><code>state\n  [8 79 48246 53607 29 34 37 Decimal('1.0000000000') 6]\naction\n  0.85\nreward\n  0.2703302\nnext state\n  [9 79 48074 57869 27 28 32 Decimal('1.0000000000') 0]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52840367,"AnswerCount":1,"Tags":"<python><machine-learning><keras><keras-layer><vgg-net>","CreationDate":"2018-10-16T16:48:04.960","AcceptedAnswerId":52840623.0,"OwnerUserId":10186519.0,"Title":"How to add a Lambda layer as an input layer to an existing model in Keras?","Body":"<p>I have a task to add a image preprocessing layer to a Keras model, so after I loaded a Keras model, I want to add a new input layer for this model.<\/p>\n\n<p>I found I can use <code>Lambda<\/code> layer to preprocess the image data. The layer code is:<\/p>\n\n<pre><code>def vgg16preprocessing(x):\n    mean_tensor = K.backend.variable([125.307, 122.95, 113.865], name=\"mean\")\n    std_tensor = K.backend.constant([62.9932, 62.0887, 66.7048], name=\"std_tensor\")\n    result = (x - mean_tensor) \/ (std_tensor)\n    return K.backend.reshape(result, (-1, 32, 32, 3))\npreproc_layer = K.layers.Lambda(vgg16preprocessing, output_shape=(32, 32, 3), input_shape=(32, 32, 3))\n<\/code><\/pre>\n\n<p>But I don't know how to add this layer in the front of my model. I found <a href=\"https:\/\/stackoverflow.com\/questions\/49546922\/keras-replacing-input-layer\">this answer<\/a>, but I can't add the layer in the <code>keras.layers.Input()<\/code>. <\/p>\n\n<p>Is there any ways to set the <code>Lambda<\/code> layer as a new input layer?<\/p>\n","answers":[{"AnswerId":"52840623","CreationDate":"2018-10-16T17:05:33.403","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can use the VGG16 model and apply it on the output of <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>vgg = VGG16(...)\n\ninput_img = Input(shape=...)\npreproc_img = Lambda(vgg16preprocessing)(input_img)\noutput = vgg(preproc_img)\n\nmodel = Model(input_img, output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52840399,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp><summarization>","CreationDate":"2018-10-16T16:50:30.287","AcceptedAnswerId":null,"OwnerUserId":9954360.0,"Title":"GloVe Text Summarization returning stop word gibberish","Body":"<p>I am trying to train a Keras test summarization model to generate a new headline for a news article that I can then compare to the published headline. I am training on GloVe 6B, then predicting against the article, which I have cleaned by removing tags, parsing, removing stop words, lemmatized, and then rejoined. My results tend to look like this:<\/p>\n\n<p>Original Headline:  Ford Traveled To Maryland In August Despite Allegedly Fear Of Flying\nGenerated Headline:  opinion: the the the the to to<\/p>\n\n<p>Article text after cleaning: Brett Kavanaugh accuser Christine Blasey Ford took polygraph test Maryland far home California despite alleged fear flying Documents released Wednesday polygraph test administered Ford Aug. 7 Hilton Hotel Linthicum Heights Maryland far Baltimore Washington International Airport A friend Kate DeVarney   Blasey Ford enjoy flying hard time place \u2019s escape route Christine Blasey Ford professor accusing Supreme Court nominee Brett Kavanaugh having sexually assaulted high school previously told friend alleged encounter 30 year ago lasting effect life Two longtime friend Ford \u2019s told CNN week previously described feeling uncomfortable struggling enclosed space escape route exit door suggested discomfort stemmed alleged encounter Kavanaugh This reason Ford enjoy flying DeVarney said airplane ultimate closed space away The fear flying   Ford able testify timely manner Senate Judiciary   In letter California Democratic Sen. Dianne Feinstein dated July 30 2018 Ford said vacation Mid Atlantic Aug. 7 day polygraph given Ford testify Senate 10 a.m. EST Thursday   <\/p>\n\n<p>Here is my training code:<\/p>\n\n<pre><code>from __future__ import print_function\n\nimport pandas as pd\nfrom sklearn.model_selection import train_test_split\nfrom keras_text_summarization.library.utility.plot_utils import plot_and_save_history\nfrom keras_text_summarization.library.seq2seq import Seq2SeqGloVeSummarizer\nfrom keras_text_summarization.library.applications.fake_news_loader import fit_text\nimport numpy as np\n\nLOAD_EXISTING_WEIGHTS = False\n\n\ndef main():\n    np.random.seed(42)\n    data_dir_path = '.\/data'\n    very_large_data_dir_path = '.\/very_large_data'\n    report_dir_path = '.\/reports'\n    model_dir_path = '.\/models'\n\n    print('loading csv file ...')\n    df = pd.read_csv(\"dcr Man_Cleaned.csv\")\n\n    print('extract configuration from input texts ...')\n    Y = df.Title\n    X = df['Joined']\n    config = fit_text(X, Y)\n\n    print('configuration extracted from input texts ...')\n\n    summarizer = Seq2SeqGloVeSummarizer(config)\n    summarizer.load_glove(very_large_data_dir_path)\n\n    if LOAD_EXISTING_WEIGHTS:\n        summarizer.load_weights(weight_file_path=Seq2SeqGloVeSummarizer.get_weight_file_path(model_dir_path=model_dir_path))\n\n    Xtrain, Xtest, Ytrain, Ytest = train_test_split(X, Y, test_size=0.2, random_state=42)\n\n    print('training size: ', len(Xtrain))\n    print('testing size: ', len(Xtest))\n\n    print('start fitting ...')\n    history = summarizer.fit(Xtrain, Ytrain, Xtest, Ytest, epochs=20, batch_size=16)\n\n    history_plot_file_path = report_dir_path + '\/' + Seq2SeqGloVeSummarizer.model_name + '-history.png'\n    if LOAD_EXISTING_WEIGHTS:\n        history_plot_file_path = report_dir_path + '\/' + Seq2SeqGloVeSummarizer.model_name + '-history-v' + str(summarizer.version) + '.png'\n    plot_and_save_history(history, summarizer.model_name, history_plot_file_path, metrics={'loss', 'acc'})\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>Any thoughts as to what is going wrong here is appreciated. <\/p>\n","answers":[{"AnswerId":"52895782","CreationDate":"2018-10-19T15:43:18.643","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Ok, so this is a very broad question and there are a lot of things that can go wrong. The problem here is that your model is stuck in the classes due to several reasons:<\/p>\n\n<ol>\n<li>Most probably you have an insufficient amount of data and the model is not able to learn from that so it just predicts the most probable output.<\/li>\n<li>Your model is either too big or to small. If it's too big, it's not able to learn enough from the data to predict the correct words so it just predicts the most probable ones so that it minimizes the error. If it's too small, it doesn't have the capacity to learn such a complex structure as natural language.<\/li>\n<\/ol>\n\n<p>Either way, you should try and add some layers, delete some layers, add some regularization, try roulette wheel approach to generate the next word based on words probability. Hope this helps :)<\/p>\n"}]}
+{"QuestionId":52840801,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-16T17:17:36.130","AcceptedAnswerId":null,"OwnerUserId":3877654.0,"Title":"What do the different color bubbles in Tensorboard mean?","Body":"<p>I am using Tensorboard to find why my tensorflow model is not training properly.  The below graph shows three weights (W, W_1 and W_2).  W_1 and W_2 are initialized identically (as random_normal) with the exception of shape.  W_2 trains properly but W_1 does not train at all.  One is orange and the other is pink.  Does this suggest the problem and if so, can anyone tell me how to fix it?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bRFig.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bRFig.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"52852261","CreationDate":"2018-10-17T10:06:01.277","ParentId":null,"OwnerUserId":"8707287","Title":null,"Body":"<p>Can you share the code for the initialization of your weights and biases?<\/p>\n\n<p>The color codes in tensorboard are a useful place to start (since no code snippets have been provided).<\/p>\n\n<p>Note that W and W_1 have one color and W_2, B, B_1 and B_2 have another color. This means they have been initialized differently. This can explain why you get different results for W_2 and W, W_1.<\/p>\n\n<p>A recommendation would be to go back to the initialization of all weights and spot the difference in initialization. Good luck!<\/p>\n"}]}
+{"QuestionId":52840888,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network><cntk>","CreationDate":"2018-10-16T17:23:06.400","AcceptedAnswerId":52858353.0,"OwnerUserId":1427022.0,"Title":"Keras accuracy not increasing over 50% on binary CNN problem","Body":"<p>I am using keras to process the following subset of my data:<\/p>\n\n<pre><code>5000 images of class A\n5000 images of class B\n<\/code><\/pre>\n\n<p>With 1000 of these images for each class being used as validation. Scaling the images to 96x96x3 channels and normalised to be within the range 0-1. I am using the following model:<\/p>\n\n<pre><code>model.add(Conv2D(32, (3, 3), activation=\"relu\", input_shape=inputshape))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(256, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>And then training the model in the following way:<\/p>\n\n<pre><code>sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss=\"binary_crossentropy\", optimizer=sgd, metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>However the accuracy rarely (just by chance) increases over 50% accuracy:<\/p>\n\n<pre><code>Epoch 1\/100\n8000\/8000 [==============================] - 23s 3ms\/step - loss: 0.6939 - acc: 0.5011 - val_loss: 0.6932 - val_acc: 0.5060\nEpoch 2\/100\n8000\/8000 [==============================] - 22s 3ms\/step - loss: 0.6938 - acc: 0.4941 - val_loss: 0.6941 - val_acc: 0.4940\nEpoch 3\/100\n8000\/8000 [==============================] - 22s 3ms\/step - loss: 0.6937     - acc: 0.4981 - val_loss: 0.6932 - val_acc: 0.4915\nEpoch 4\/100\n8000\/8000 [==============================] - 22s 3ms\/step - loss: 0.6933 - acc: 0.5056 - val_loss: 0.6931 - val_acc: 0.5060\nEpoch 5\/100\n8000\/8000 [==============================] - 22s 3ms\/step - loss: 0.6935 - acc: 0.4970 - val_loss: 0.6932 - val_acc: 0.4940\n<\/code><\/pre>\n\n<p>I don't think the problem is the data itself, as I have used an alternative machine learning method and got over 94% accuracy with the exact same images (except using just 5 training images for each class, but that's beside the point).<\/p>\n\n<p>Any help would be greatly appreciated.<\/p>\n\n<p>Oh! In case it matters: I'm using the CNTK backend.<\/p>\n\n<p>Edit: Here is the code I use to read in the images, which also normalises the pixel values to the 0-1 range:<\/p>\n\n<pre><code>import cv2\nimport numpy as np\nfrom keras.preprocessing.image import img_to_array\n\nhealthy_files = sorted(os.listdir(\"..\/..\/uninfected\/\"))\nhealthy_imgs = [cv2.imread(\"..\/..\/uninfected\/\" + x) for x in healthy_files]\ndata = []\nlabels = []\nfor img in healthy_imgs[:5000]:\n    resized = cv2.resize(img, (96, 96)).astype(numpy.float32) \/ 255.0 # normalise data to 0..1 range\n    arr = img_to_array(resized) \n    data += [arr]\n    labels += [0]\n# The for loop above is then repeated over the other half of the dataset, with the labels line using the label [1] instead\ndata = np.array(data, numpy.float32)\n<\/code><\/pre>\n\n<p>Edit 2: Here is the output of model.summary():<\/p>\n\n<pre><code>Model built:\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\nconv2d_1 (Conv2D)            (None, 94, 94, 32)        896\n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 92, 92, 32)        9248\n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 46, 46, 32)        0\n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 44, 44, 64)        18496\n_________________________________________________________________\nconv2d_4 (Conv2D)            (None, 42, 42, 64)        36928\n_________________________________________________________________\nmax_pooling2d_2 (MaxPooling2 (None, 21, 21, 64)        0\n_________________________________________________________________\nflatten_1 (Flatten)          (None, 28224)             0\n_________________________________________________________________\ndense_1 (Dense)              (None, 256)               7225600\n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 257\n=================================================================\nTotal params: 7,291,425\nTrainable params: 7,291,425\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>I noticed that there were no activation layers explicitly listed in this summary so I changed the model to this:<\/p>\n\n<pre><code>model.add(Conv2D(32, (3, 3), input_shape=inputshape))\nmodel.add(Activation(\"relu\"))\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n#model.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation(\"relu\"))\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n#model.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation(\"relu\"))\n#model.add(Dropout(0.5))\n#model.add(Dense(10, activation=\"relu\"))\nmodel.add(Dense(1))\nmodel.add(Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>Which gave a summary output of this:<\/p>\n\n<pre><code>Model built:\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\nconv2d_1 (Conv2D)            (None, 94, 94, 32)        896\n_________________________________________________________________\nactivation_1 (Activation)    (None, 94, 94, 32)        0\n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 92, 92, 32)        9248\n_________________________________________________________________\nactivation_2 (Activation)    (None, 92, 92, 32)        0\n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 46, 46, 32)        0\n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 44, 44, 64)        18496\n_________________________________________________________________\nactivation_3 (Activation)    (None, 44, 44, 64)        0\n_________________________________________________________________\nconv2d_4 (Conv2D)            (None, 42, 42, 64)        36928\n_________________________________________________________________\nactivation_4 (Activation)    (None, 42, 42, 64)        0\n_________________________________________________________________\nmax_pooling2d_2 (MaxPooling2 (None, 21, 21, 64)        0\n_________________________________________________________________\nflatten_1 (Flatten)          (None, 28224)             0\n_________________________________________________________________\ndense_1 (Dense)              (None, 64)                1806400\n_________________________________________________________________\nactivation_5 (Activation)    (None, 64)                0\n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 65\n_________________________________________________________________\nactivation_6 (Activation)    (None, 1)                 0\n=================================================================\nTotal params: 1,872,033\nTrainable params: 1,872,033\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Needless to say, the results remain the same...<\/p>\n","answers":[{"AnswerId":"52843117","CreationDate":"2018-10-16T19:54:00.823","ParentId":null,"OwnerUserId":"3741442","Title":null,"Body":"<p>also it is usually a bad idea to use dropout in convolutional layer, use batch normalization instead.<\/p>\n"},{"AnswerId":"52858353","CreationDate":"2018-10-17T15:20:32.847","ParentId":null,"OwnerUserId":"1427022","Title":null,"Body":"<p>So, after trying all the suggestions the wonderful people had in the comments, I had no luck. I decided to go back to the drawing board, or in this case, try it on an alternative computer. My original code worked! <\/p>\n\n<p>In the end I narrowed it down to the backend - I was using CNTK on the first computer, and Tensorflow on the second. I tried CNTK on the second computer, and it worked perfectly... So I decided to reinstall CNTK on the first computer. This time, the code worked perfectly. So I have no idea what was broken initially, but it had <em>something<\/em> to do with my install of CNTK. I guess in the end, this whole Q&amp;A doesn't really help anyone.. but if anyone experiences a similar issue - try the suggestions in the comments on the question - some really good advice there. And if that doesn't work... try changing your backend!<\/p>\n\n<p>Cheers<\/p>\n"}]}
+{"QuestionId":52841015,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-10-16T17:32:14.773","AcceptedAnswerId":52878050.0,"OwnerUserId":3876830.0,"Title":"Create a Logical AND with Tensorflow using a Deep Neural Network classifier","Body":"<p>I am trying to make a basic Tensorflow model to predict the output of a logical AND:<\/p>\n\n<pre><code>+---------+---------+--------+\n| Input A | Input B | Output |\n+---------+---------+--------+\n|    0    |    0    |   0    |\n|    0    |    1    |   0    |\n|    1    |    0    |   0    |\n|    1    |    1    |   1    |\n+---------+---------+--------+\n<\/code><\/pre>\n\n<p>I know other estimators would be better for this specific task, but I am sticking to <strong>DNNClassifier<\/strong>, because I am going to need it on future projects. This is my code for training and predicting the outputs:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from matplotlib import pyplot as plt\nimport numpy as np\nimport pandas as pd\nfrom sklearn import metrics\nimport tensorflow as tf\nfrom tensorflow.python.data import Dataset\nimport seaborn as sns\nimport glob\nimport os\n\ndef input(features, targets, batch_size=1, shuffle=True, num_epochs=None):\n    # Convert pandas data into a dict of np arrays.\n    features = {key: np.array(value) for key, value in dict(features).items()}\n\n    # Construct a dataset, and configure batching\/repeating.\n    ds = Dataset.from_tensor_slices((features, targets))  # warning: 2GB limit\n    ds = ds.batch(batch_size).repeat(num_epochs)\n\n    # Shuffle the data, if specified.\n    if shuffle:\n        ds = ds.shuffle(10000)\n\n    # Return the next batch of data.\n    features, labels = ds.make_one_shot_iterator().get_next()\n    return features, labels\n\n\ndef train_model(learning_rate, steps, batch_size, training_examples, training_targets, validation_examples, validation_targets):\n    periods = 10\n\n    steps_per_period = steps \/ periods\n    training_input_fn = lambda: input(training_examples, training_targets, batch_size=batch_size)\n    predict_training_input_fn = lambda: input(training_examples, training_targets, num_epochs=1, shuffle=False)\n    predict_validation_input_fn = lambda: input(validation_examples, validation_targets, num_epochs=1, shuffle=False)\n\n    tensorflow_features = set([tf.feature_column.numeric_column(my_feature) for my_feature in training_examples])\n    optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate)\n    optimizer = tf.contrib.estimator.clip_gradients_by_norm(optimizer, 5.0)\n    classifier = tf.estimator.DNNClassifier(\n        feature_columns=tensorflow_features,\n        n_classes=2,\n        hidden_units=[1],\n        optimizer=optimizer,\n        activation_fn=tf.nn.sigmoid,\n        config=tf.contrib.learn.RunConfig(keep_checkpoint_max=1)\n    )\n\n    # Train the model, but do so inside a loop so that we can periodically assess\n    # loss metrics.\n    print(\"Training model...\")\n    print(\"LogLoss error (on validation data):\")\n    training_errors = []\n    validation_errors = []\n    # The model will be trained in 10 periods:\n    for period in range(0, periods):\n        # Train the model, starting from the prior state.\n        classifier.train(\n            input_fn=training_input_fn,\n            steps=steps_per_period\n        )\n\n        # Take a break and compute probabilities.\n        training_predictions = list(classifier.predict(input_fn=predict_training_input_fn))\n        training_probabilities = np.array([item['probabilities'] for item in training_predictions])\n        training_pred_class_id = np.array([item['class_ids'][0] for item in training_predictions])\n        training_pred_one_hot = tf.keras.utils.to_categorical(training_pred_class_id, num_classes=2)\n\n        validation_predictions = list(classifier.predict(input_fn=predict_validation_input_fn))\n        validation_probabilities = np.array([item['probabilities'] for item in validation_predictions])\n        validation_pred_class_id = np.array([item['class_ids'][0] for item in validation_predictions])\n        validation_pred_one_hot = tf.keras.utils.to_categorical(validation_pred_class_id, num_classes=2)\n\n        # Compute training and validation errors.\n        training_log_loss = metrics.log_loss(training_targets, training_pred_one_hot)\n        validation_log_loss = metrics.log_loss(validation_targets, validation_pred_one_hot)\n        # Occasionally print the current loss.\n        print(\"  period %02d : %0.2f\" % (period, validation_log_loss))\n        # Add the loss metrics from this period to our list.\n        training_errors.append(training_log_loss)\n        validation_errors.append(validation_log_loss)\n    print(\"Model training finished.\")\n    # Remove event files to save disk space.\n    _ = map(os.remove, glob.glob(os.path.join(classifier.model_dir, 'events.out.tfevents*')))\n\n    # Calculate final predictions (not probabilities, as above).\n    final_predictions = classifier.predict(input_fn=predict_validation_input_fn)\n    final_predictions = np.array([item['class_ids'][0] for item in final_predictions])\n\n    accuracy = metrics.accuracy_score(validation_targets, final_predictions)\n    print(\"Final accuracy (on validation data): %0.2f\" % accuracy)\n    # Output a graph of loss metrics over periods.\n    plt.ylabel(\"LogLoss\")\n    plt.xlabel(\"Periods\")\n    plt.title(\"LogLoss vs. Periods\")\n    plt.plot(training_errors, label=\"training\")\n    plt.plot(validation_errors, label=\"validation\")\n    plt.legend()\n    plt.show()\n\n    # Plot the confusion matrix.\n    cm = metrics.confusion_matrix(validation_targets, final_predictions)\n    # Normalize the confusion matrix by row (i.e by the number of samples\n    # in each class).\n    cm_normalized = cm.astype(\"float\") \/ cm.sum(axis=1)[:, np.newaxis]\n    ax = sns.heatmap(cm_normalized, cmap=\"bone_r\")\n    ax.set_aspect(1)\n    plt.title(\"Confusion matrix\")\n    plt.ylabel(\"True label\")\n    plt.xlabel(\"Predicted label\")\n    plt.show()\n<\/code><\/pre>\n\n<p>To train and predict the output I just run the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>training_examples = pd.DataFrame([[0,0],[0,1],[1,0],[1,1]], columns=['input_a', 'input_b'])\ntraining_targets = pd.DataFrame([[0],[0],[0],[1]], columns=['output'])\nvalidation_examples = pd.DataFrame([[1,0],[0,1],[1,1],[0,0],[0,1],[0,0],[1,1]], columns=['input_a', 'input_b'])\nvalidation_targets = pd.DataFrame([[0],[0],[1],[0],[0],[0],[1]], columns=['output'])\n\ntrain_model(\n  learning_rate=0.001,\n  steps=10,\n  batch_size=1,\n  training_examples=training_examples,\n  training_targets=training_targets,\n  validation_examples=validation_examples,\n  validation_targets=validation_targets,\n)\n<\/code><\/pre>\n\n<p>After several times tweaking the hyperparameters and running the code I always get pretty bad accuracy and a clearly wrong confusion matrix:\n<a href=\"https:\/\/i.stack.imgur.com\/aXrjo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aXrjo.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there some Tensorflow expert who can bring light on this?<\/p>\n","answers":[{"AnswerId":"52848638","CreationDate":"2018-10-17T06:37:56.587","ParentId":null,"OwnerUserId":"366553","Title":null,"Body":"<p>Try sigmoid activation.<\/p>\n\n<p>The classic example of single-neuron AND has a solution with sigmoid activation function but does not have a single-neuron solution with ReLU, there are no weights possible that would fit the function.<\/p>\n\n<p>If you want to use ReLU, then you will need an extra layer and more neurons to get the behavior of this function.<\/p>\n"},{"AnswerId":"52878050","CreationDate":"2018-10-18T15:57:13.797","ParentId":null,"OwnerUserId":"3876830","Title":null,"Body":"<p>I finally got it working! <\/p>\n\n<p>I have removed the <strong>hidden_units<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate)\noptimizer = tf.contrib.estimator.clip_gradients_by_norm(optimizer, 5.0)\nclassifier = tf.estimator.DNNClassifier(\n    feature_columns=tensorflow_features,\n    n_classes=2,\n    hidden_units=[],\n    optimizer=optimizer,\n    activation_fn=tf.nn.sigmoid,\n    config=tf.contrib.learn.RunConfig(keep_checkpoint_max=1)\n)\n<\/code><\/pre>\n\n<p>and updated the hyperparameters as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>classifier = train_model(\n  learning_rate=1,\n  steps=100,\n  batch_size=1,\n  training_examples=training_examples,\n  training_targets=training_targets,\n  validation_examples=validation_examples,\n  validation_targets=validation_targets,\n)\n<\/code><\/pre>\n\n<p>Final output:<\/p>\n\n<pre><code>Training model...\nLogLoss error (on validation data):\n  period 00 : 9.87\n  period 01 : 0.00\n  period 02 : 0.00\n  period 03 : 0.00\n  period 04 : 0.00\n  period 05 : 0.00\n  period 06 : 0.00\n  period 07 : 0.00\n  period 08 : 0.00\n  period 09 : 0.00\nModel training finished.\nFinal accuracy (on validation data): 100%\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/EaGmf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/EaGmf.png\" alt=\"Confusion matrix\"><\/a><\/p>\n"}]}
+{"QuestionId":52841134,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><download>","CreationDate":"2018-10-16T17:40:35.857","AcceptedAnswerId":null,"OwnerUserId":7657660.0,"Title":"Load images in tarball from a link to tensorflow for training CNN","Body":"<p>I am a newbie in Tensorflow and have hard time figuring out a very simple task. Most of the examples for Tensorflow comes from MNIST and cifar10 which are built into the library, so we don't need to download *.jpg files and convert them to arrays for Tensorflow. I have the following dataset: <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/data\/flowers\/17\/\" rel=\"nofollow noreferrer\">17 flower categories<\/a>. it is a tarball that contains 1360 images. There is no labels for images. We can divide it into batches of 80 images that represent each category starting from the first image, so we get 80x17=1360 images. \nI was not able load the dataset from the URL to python. But, I downloaded it to local folder and then loaded using the line:<\/p>\n\n<pre><code>import tarfile\nf = tarfile.open(\"..\/input\/17flowers.tgz\")\nfor member in f.getmembers():\n   g = f.extractfile(member)\n<\/code><\/pre>\n\n<p>However, this is not helpful in loading images to training my model.\nCan you, please, show me a snippet of code that loads the dataset from the link and then loads images to tensorflow for training. Labels should be from 1 to 17 for each 80 consecutive images.\nThanks!  <\/p>\n","answers":[{"AnswerId":"52863241","CreationDate":"2018-10-17T20:39:19.087","ParentId":null,"OwnerUserId":"7657660","Title":null,"Body":"<p>I was able to solve the problem by referring to the code and modifying it: <a href=\"https:\/\/github.com\/changyq12\/17_flower_cnn_tensorflow\/blob\/master\/convolutional_network.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/changyq12\/17_flower_cnn_tensorflow\/blob\/master\/convolutional_network.py<\/a><\/p>\n"}]}
+{"QuestionId":52842069,"AnswerCount":0,"Tags":"<c++><boost><cmake><anaconda><caffe>","CreationDate":"2018-10-16T18:45:05.253","AcceptedAnswerId":null,"OwnerUserId":3303546.0,"Title":"Finding boost-python3 with Anaconda cmake prefix","Body":"<p><strong>DLDR<\/strong> How do I point the cmake at boost-python3 library? It is not automatically detected by cmake.<\/p>\n\n<hr>\n\n<p>I'm trying to build <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">caffe<\/a> for Python 3.6 using the provided cmake.<\/p>\n\n<p>My system specs:<\/p>\n\n<ul>\n<li>Python 3.6.5, Anaconda custom (64-bit)<\/li>\n<li>Mac OS 10.13.6<\/li>\n<li>No CUDA<\/li>\n<\/ul>\n\n<p>I've installing boost with brew, e.g. <\/p>\n\n<pre><code>brew install boost boost-python3\n<\/code><\/pre>\n\n<p>I can see the boost libraries using <code>find \/ -name libboost* 2&gt;\/dev\/null<\/code>. They occur in three directories<\/p>\n\n<ol>\n<li><code>\/usr\/local\/Cellar\/boost\/1.67.0_1\/lib\/<\/code> <\/li>\n<li><code>\/usr\/local\/lib\/<\/code> -> symlink to above<\/li>\n<li>boost-python3 is in <code>\/usr\/local\/Cellar\/boost-python\/1.67.0\/lib\/<\/code><\/li>\n<\/ol>\n\n<p>If I run <code>cmake -DCMAKE_PREFIX_PATH=&lt;anaconda_env_path&gt; -D python_version=3<\/code>, I get this at the top of the output<\/p>\n\n<pre><code>-- Boost version: 1.67.0\n-- Found the following Boost libraries:\n--   system\n--   thread\n--   filesystem\n--   chrono\n--   date_time\n--   atomic\n<\/code><\/pre>\n\n<p>But further down, I also get <\/p>\n\n<pre><code>CMake Warning at \/Users\/Mauceri\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1723 (message):\n  No header defined for python-py365; skipping header check\nCall Stack (most recent call first):\n  cmake\/Dependencies.cmake:157 (find_package)\n  CMakeLists.txt:49 (include)\n\n\n-- Could NOT find Boost\nCMake Warning at \/Users\/Mauceri\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1723 (message):\n  No header defined for python-py36; skipping header check\nCall Stack (most recent call first):\n  cmake\/Dependencies.cmake:164 (find_package)\n  CMakeLists.txt:49 (include)\n\n\n-- Could NOT find Boost\nCMake Warning at \/Users\/Mauceri\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1723 (message):\n  No header defined for python-py3; skipping header check\nCall Stack (most recent call first):\n  cmake\/Dependencies.cmake:164 (find_package)\n  CMakeLists.txt:49 (include)\n\n-- Could NOT find Boost\n-- Could NOT find Boost\n-- Python interface is disabled or not all required dependencies found. Building without it...\n<\/code><\/pre>\n\n<p>Similar to <a href=\"https:\/\/stackoverflow.com\/questions\/3808775\/cmake-doesnt-find-boost\">Cmake doesn&#39;t find Boost<\/a>, I have tried adding <\/p>\n\n<pre><code>-DBOOST_ROOT=\/usr\/local\/Cellar\/boost\/1.67.0_1\/lib\/ -DBoost_DEBUG=1\n<\/code><\/pre>\n\n<p>to the cmake command. In the resulting output, the following repeats three times with boost_python-py365, boost_python-py36, boost_python-py3, and boost_python:<\/p>\n\n<pre><code>[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1121 ] _boost_TEST_VERSIONS = 1.67.0;1.67;1.66.0;1.66;1.65.1;1.65.0;1.65;1.64.0;1.64;1.63.0;1.63;1.62.0;1.62;1.61.0;1.61;1.60.0;1.60;1.59.0;1.59;1.58.0;1.58;1.57.0;1.57;1.56.0;1.56;1.55.0;1.55;1.54.0;1.54;1.53.0;1.53;1.52.0;1.52;1.51.0;1.51;1.50.0;1.50;1.49.0;1.49;1.48.0;1.48;1.47.0;1.47;1.46.1;1.46.0;1.46\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1123 ] Boost_USE_MULTITHREADED = TRUE\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1125 ] Boost_USE_STATIC_LIBS = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1127 ] Boost_USE_STATIC_RUNTIME = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1129 ] Boost_ADDITIONAL_VERSIONS = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1131 ] Boost_NO_SYSTEM_PATHS = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1199 ] Declared as CMake or Environmental Variables:\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1201 ]   BOOST_ROOT = \/usr\/local\/Cellar\/boost\/1.67.0_1\/include\/\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1203 ]   BOOST_INCLUDEDIR = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1205 ]   BOOST_LIBRARYDIR = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1207 ] _boost_TEST_VERSIONS = 1.67.0;1.67;1.66.0;1.66;1.65.1;1.65.0;1.65;1.64.0;1.64;1.63.0;1.63;1.62.0;1.62;1.61.0;1.61;1.60.0;1.60;1.59.0;1.59;1.58.0;1.58;1.57.0;1.57;1.56.0;1.56;1.55.0;1.55;1.54.0;1.54;1.53.0;1.53;1.52.0;1.52;1.51.0;1.51;1.50.0;1.50;1.49.0;1.49;1.48.0;1.48;1.47.0;1.47;1.46.1;1.46.0;1.46\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1306 ] location of version.hpp: \/usr\/local\/Cellar\/boost\/1.67.0_1\/include\/boost\/version.hpp\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1330 ] version.hpp reveals boost 1.67.0\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1416 ] guessed _boost_COMPILER = \n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1426 ] _boost_MULTITHREADED = -mt\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1502 ] _boost_RELEASE_ABI_TAG = -\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1504 ] _boost_DEBUG_ABI_TAG = -d\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1567 ] _boost_LIBRARY_SEARCH_DIRS_RELEASE = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib;NO_DEFAULT_PATH;NO_CMAKE_FIND_ROOT_PATH_boost_LIBRARY_SEARCH_DIRS_DEBUG   = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib;NO_DEFAULT_PATH;NO_CMAKE_FIND_ROOT_PATH\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1756 ] Searching for PYTHON_LIBRARY_RELEASE: boost_python-mt-1_67;boost_python-mt;boost_python\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:400 ]  Boost_LIBRARY_DIR_RELEASE = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib _boost_LIBRARY_SEARCH_DIRS_RELEASE = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib;NO_DEFAULT_PATH;NO_CMAKE_FIND_ROOT_PATH\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1809 ] Searching for PYTHON_LIBRARY_DEBUG: boost_python-mt-d-1_67;boost_python-mt-d;boost_python-mt;boost_python\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:400 ]  Boost_LIBRARY_DIR_DEBUG = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib _boost_LIBRARY_SEARCH_DIRS_DEBUG = \/usr\/local\/Cellar\/boost\/1.67.0_1\/lib;NO_DEFAULT_PATH;NO_CMAKE_FIND_ROOT_PATH\n[ \/Users\/me\/anaconda\/share\/cmake-3.11\/Modules\/FindBoost.cmake:1883 ] Boost_FOUND = 1\nCould NOT find Boost\nBoost version: 1.67.0\nBoost include path: \/usr\/local\/Cellar\/boost\/1.67.0_1\/include\nCould not find the following Boost libraries:\n        boost_python\nNo Boost libraries were found. You may need to set BOOST_LIBRARYDIR to the directory containing Boost libraries or BOOST_ROOT to the location of Boost.\n<\/code><\/pre>\n\n<p>Therefore, I think the boost-python path is what is missing. I also tried adding <code>-DBOOST_LIBRARYDIR<\/code>, but that didn't seem to change anything. <\/p>\n\n<p>The <a href=\"https:\/\/cmake.org\/cmake\/help\/latest\/module\/FindBoost.html\" rel=\"nofollow noreferrer\">FindBoost documentation<\/a> contains this comment about boost-python<\/p>\n\n<blockquote>\n  <p>Note that Boost Python components require a Python version suffix\n  (Boost 1.67 and later), e.g. python36 or python27 for the versions\n  built against Python 3.6 and 2.7, respectively. This also applies to\n  additional components using Python including mpi_python and numpy.\n  Earlier Boost releases may use distribution-specific suffixes such as\n  2, 3 or 2.7. These may also be used as suffixes, but note that they\n  are not portable.<\/p>\n<\/blockquote>\n\n<p>I noticed that the boost-python3 libraries had the suffix 37 (libboost_python37.a), so I also tried using an anaconda environment with python 3.7, but the same errors persisted. <\/p>\n\n<p>How can I get my cmake to find the boost-python3 libraries?<\/p>\n","answers":[]}
+{"QuestionId":52842426,"AnswerCount":1,"Tags":"<numpy><tensorflow><prediction>","CreationDate":"2018-10-16T19:08:11.987","AcceptedAnswerId":null,"OwnerUserId":4405235.0,"Title":"how to get an array of predictions from tensor flow classification model","Body":"<p>I have the following classification model. \nI would like to get a numpy array similar to y_t which is the test labels one hot encoded. However I keep getting variable error.<\/p>\n\n<pre><code># Construct placeholders\nwith graph.as_default():\n    inputs_ = tf.placeholder(tf.float32, [None, seq_len, n_channels], name = 'inputs')\n    labels_ = tf.placeholder(tf.float32, [None, n_classes], name = 'labels')\n    keep_prob_ = tf.placeholder(tf.float32, name = 'keep')\n    learning_rate_ = tf.placeholder(tf.float32, name = 'learning_rate')\n\nwith graph.as_default():\n    # (batch, 100, 3) --&gt; (batch, 50, 6)\n    conv1 = tf.layers.conv1d(inputs=inputs_, filters=6, kernel_size=2, strides=1, \n                             padding='same', activation = tf.nn.relu)\n    max_pool_1 = tf.layers.max_pooling1d(inputs=conv1, pool_size=2, strides=2, padding='same')\n\nwith graph.as_default():\n    # Flatten and add dropout\n    flat = tf.reshape(max_pool_1, (-1, 6*6))\n    flat = tf.nn.dropout(flat, keep_prob=keep_prob_)\n\n    # Predictions\n    logits = tf.layers.dense(flat, n_classes)\n\n    # Cost function and optimizer\n    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels_))\n    optimizer = tf.train.AdamOptimizer(learning_rate_).minimize(cost)\n\n    # Accuracy\n    correct_pred = tf.equal(tf.argmax(logits, 1), tf.argmax(labels_, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32), name='accuracy')\n<\/code><\/pre>\n\n<p>Then I use the test set <\/p>\n\n<pre><code>with tf.Session(graph=graph) as sess:\n    # Restore\n    saver.restore(sess, tf.train.latest_checkpoint('bschkpnt-cnn'))\n\n    for x_t, y_t in get_batches(X_test, y_test, batch_size):\n        feed = {inputs_: x_t,\n                labels_: y_t,\n                keep_prob_: 1}\n\n        batch_acc = sess.run(accuracy, feed_dict=feed)\n        test_acc.append(batch_acc)\n    print(\"Test accuracy: {:.6f}\".format(np.mean(test_acc)))\n<\/code><\/pre>\n\n<p>y_t is a nX3 bumpy array.\nI want to get a y_pred in similar format<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52844162","CreationDate":"2018-10-16T21:17:34.113","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<blockquote>\n  <p>soft = tf.nn.softmax(logits)<\/p>\n<\/blockquote>\n\n<p>this will be your probability distribution such that <code>sum(soft)<\/code> = 1. Every value in this array will indicate how sure the model is about the class.<\/p>\n\n<blockquote>\n  <p>pred = sess.run(soft, feed_dict=feed)<\/p>\n  \n  <p>print(pred)<\/p>\n<\/blockquote>\n\n<p>So basically all I do is place an additional <code>softmax<\/code>, since you have it inbuilt in the loss you calculate, you've to place it again to predict. Then I ask for the output prediction, and just feed the <code>feed_dict<\/code> again. <\/p>\n\n<p>Hope this helped!<\/p>\n"}]}
+{"QuestionId":52842939,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-16T19:41:35.270","AcceptedAnswerId":null,"OwnerUserId":9481164.0,"Title":"Multivariate Poisson loss function in Keras","Body":"<p>I want to perform a multivariate Poisson regression with Keras. In other words, I want to maximise <code>Product[P(y_true[i] | y_pred[i])]<\/code> for each element in the output vector (<code>i<\/code>). In this case, P will the the Poisson pmf.<\/p>\n\n<p>I believe that I can do this with a custom loss function. However, working out how to implement this loss function using the available functions in <code>keras.backend<\/code> is proving to be a challenge. <\/p>\n\n<p>In particular, I am unable to find a way to implement elementwise factorial on a 1D tensor, something required to calculate the Poisson pmf.<\/p>\n\n<p>Does anyone have any pointers for how this might be done? Or even if it could be done?<\/p>\n","answers":[{"AnswerId":"57169449","CreationDate":"2019-07-23T17:33:48.287","ParentId":null,"OwnerUserId":"6736176","Title":null,"Body":"<p>Keras has a built-in <a href=\"https:\/\/keras.io\/losses\/\" rel=\"nofollow noreferrer\">Poisson loss function<\/a>! If you have multiple outcomes, Keras will just apply the loss function to predictions for each variable, and sum them (you'd have to create a custom loss function if you want to weigh the loss for one variable more than for another). <\/p>\n\n<p>As an aside, you don't need to worry about calculating the factorial in the PMF in order to maximize the likelihood, since <code>factorial(y_true[i])<\/code> isn't a function of the model parameters. There's a quick derivation of the objective function (loss function is just the negative of this) on the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Poisson_regression#Maximum_likelihood-based_parameter_estimation\" rel=\"nofollow noreferrer\">Poisson Regression wiki<\/a>. It's in the context of a linear model, but just replace <code>theta_prime * x[i]<\/code> with <code>y_pred[i]<\/code> and it matches the context here.<\/p>\n"}]}
+{"QuestionId":52843410,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><computer-vision><caffe><lmdb>","CreationDate":"2018-10-16T20:14:51.423","AcceptedAnswerId":null,"OwnerUserId":4849120.0,"Title":"How should LMDB record data be organized so that Caffe's data layer can read them?","Body":"<p>I need to create a fast, efficient, low-overhead routine for storing key \/ value pairs in LMDB for subsequent consumption by Caffe's data layer (i.e., no linking to a bunch of external libraries).<\/p>\n\n<p>I've reviewed the caffe.proto, caffe.pb.h and caffe.pb.cc files and a handful of others pertaining to Google's protocol buffers to gain an understanding of the Datum class, which is the 'value' in LMDB records.<\/p>\n\n<p>The best bet for me appears to be an audit of the datum.SerializeToString() method, which takes all the data structures and nested structures comprising Datum and converts them to some sort of string value.  However, after plumbing the depths of Google's protobuf, I haven't been able to find where this method is defined.<\/p>\n\n<p>Can someone point me in the right direction?  And obviously if there's a faster \/ better \/ cheaper way of understanding how the serialized Datum value should be structured, then I'd definitely be open to it.  Thanks.<\/p>\n","answers":[{"AnswerId":"52847954","CreationDate":"2018-10-17T05:43:20.517","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I think you are looking for <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py#L66-L81\" rel=\"nofollow noreferrer\"><code>caffe.io.array_to_datum<\/code><\/a> method: this is a python wrapper to caffe protobuf interface converting a numpy array (and an optional integer label) into a <code>Datum<\/code> object.<br>\nThere is a more comprehensive example on how to read\/write LMDB for caffe using python interface <a href=\"http:\/\/deepdish.io\/2015\/04\/28\/creating-lmdb-in-python\/\" rel=\"nofollow noreferrer\">here<\/a>.<br>\nIf you are just into converting a list of labeled images into LMDB, you can use <a href=\"https:\/\/stackoverflow.com\/q\/31427094\/1714410\"><code>convert_imageset<\/code><\/a> tool that ships with caffe.<\/p>\n"}]}
+{"QuestionId":52843778,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T20:43:22.140","AcceptedAnswerId":null,"OwnerUserId":8378120.0,"Title":"Tensorflow tf.nn.conv2d giving different results for variable vs constant","Body":"<p>While converting a model from tensorflow to Pytorch, I ran into a strange issue, where tf.nn.conv2d gives a different result when convolving with the same filters, depending whether a <strong>variable<\/strong> or a <strong>constant<\/strong> is used. <\/p>\n\n<p>I was hoping this to be a bug on my code, but I reduced the problem to a small snippet that gives widely different results on tf.nn.conv2d, that I cannot explain:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nrng = np.random.RandomState(42)\n\nsess = tf.InteractiveSession()\nsome_input = rng.normal(size=(10, 10, 10, 32)).astype(np.float32)\nw_value = rng.normal(size=(5, 5, 32, 64)).astype(np.float32)\n\nvariable = tf.Variable(w_value)\nsess.run(variable.initializer)\n\nconstant = tf.constant(w_value)\n\nwith_var = sess.run(tf.nn.conv2d(some_input,\n                                 variable,\n                                 strides=[1, 1, 1, 1],\n                                 padding='SAME'))\n\n\nwith_const = sess.run(tf.nn.conv2d(some_input,\n                                   constant,\n                                   strides=[1, 1, 1, 1],\n                                   padding='SAME'))\n\nprint('Norm between results: {}'.format(np.linalg.norm(with_var - with_const)))\nprint('Max diff: {}'.format(np.max(np.abs(with_var - with_const))))\n\nprint('Norm between variable and constant: {}'.format(\n    np.linalg.norm(sess.run(variable) - sess.run(constant))))\n<\/code><\/pre>\n\n<p>On tensorflow 1.11, it gives me a norm of 1.18 between the outputs, which is way too high (it should be 0). I tested both on GPU (GTX 1080) and CPU, and saw differences in both cases. <\/p>\n\n<p>Any clues to this problem? Using constants should be a very normal operation (e.g. when freezing a graph we convert variables to constants, which should not change the result).<\/p>\n","answers":[]}
+{"QuestionId":52844431,"AnswerCount":1,"Tags":"<python><theano><azure-machine-learning-studio>","CreationDate":"2018-10-16T21:43:21.673","AcceptedAnswerId":52850290.0,"OwnerUserId":1403955.0,"Title":"Azure Machine Learning Studio execute python script, Theano unable to execute optimized C-implementations (for both CPU and GPU)","Body":"<p>I am execute a python script in Azure machine learning studio. I am including other python scripts and python library, Theano. I can see the Theano get loaded and I got the proper result after script executed. But I saw the error message:<\/p>\n\n<blockquote>\n  <p>WARNING (theano.configdefaults): g++ not detected ! Theano will be unable to execute optimized C-implementations (for both CPU and GPU) and will default to Python implementations. Performance will be severely degraded. To remove this warning, set Theano flags cxx to an empty string.<\/p>\n<\/blockquote>\n\n<p>Did anyone know how to solve this problem? Thanks!<\/p>\n","answers":[{"AnswerId":"52850290","CreationDate":"2018-10-17T08:19:37.230","ParentId":null,"OwnerUserId":"155697","Title":null,"Body":"<p>I don't think you can fix that - the Python script environment in Azure ML Studio is rather locked down, you can't really configure it (except for choosing from a small selection of Anaconda\/Python versions). <\/p>\n\n<p>You might be better off using the new Azure ML service, which allows you considerably more configuration options (including using GPUs and the like). <\/p>\n"}]}
+{"QuestionId":52845114,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-16T23:05:51.337","AcceptedAnswerId":null,"OwnerUserId":10515667.0,"Title":"Repeat element of a tensor and form a new tensor in tensorflow","Body":"<p>Say I have a tensor A=[a1,a2,...], I want to to repeat the element of the tensor and form a new tensor. The number of repetition of each element is indicated in another tensor B. For example if B=[1,3,2,2,..], the result should be [a1,a2,a2,a2,a3,a3,a4,a4,...]. Is there an efficient way to perform this in tensorflow without using loop?<\/p>\n","answers":[{"AnswerId":"52853117","CreationDate":"2018-10-17T10:49:36.237","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>This is copied from this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8246\" rel=\"nofollow noreferrer\">issue<\/a>. I am interpreting it because it is interesting and it is without loops. Don't know if it is efficient or not.<\/p>\n\n<p>If we assume that we want to repeat <code>[1,2,3]<\/code> with <code>[3,4,5]<\/code> repetitions then gist of it is that they create a sparse tensor like this.<\/p>\n\n<blockquote>\n  <p>[1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 0. 0.]<\/p>\n<\/blockquote>\n\n<p>and fill it up like this using <code>tf.cumsum<\/code> cleverly.<\/p>\n\n<blockquote>\n  <p>[1. 1. 1. 2. 2. 2. 2. 3. 3. 3. 3. 3.]<\/p>\n<\/blockquote>\n\n<p>I am interpreting the procedure as much as possible here.<\/p>\n\n<p><code>tf.cumsum([3,4,5]<\/code> gives <code>[ 3  7 12]<\/code><\/p>\n\n<p><code>tf.cumsum([3, 4, 5][:-1])<\/code> gives <code>[3 7]<\/code> after removing the last element 12.<\/p>\n\n<p><code>tf.concat([tf.constant([0], dtype=tf.int32), tf.cumsum([3,4,5][:-1])], axis=0)<\/code> gives <code>[0 3 7]<\/code>\nwhich are the indices where we see <code>1<\/code>.<\/p>\n\n<blockquote>\n  <p>[1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 0. 0.]<\/p>\n<\/blockquote>\n\n<p>And <code>1<\/code> is obtained using <code>[1,2,3] - tf.concat([tf.constant([0], dtype=tf.float64), [1,2,3][:-1]], axis=0)<\/code><\/p>\n\n<p>which is equivalent to subtracting like this. <code>[1, 2, 3]<\/code> - <code>[0, 1, 2]<\/code>. This gives <code>[1 1 1]<\/code> for the values for the sparse indices.<\/p>\n\n<p><code>12<\/code> as the output_shape is the total placeholders required in the output which is the reduced cumulative sum of our repeat tensor <code>[3, 4, 5]<\/code><\/p>\n\n<p>This gives the final output.<\/p>\n\n<pre><code>print(sess.run(\n    tf.cumsum(\n    tf.sparse_to_dense(\n    sparse_indices=[0, 3, 7],\n    output_shape=(12,),\n    sparse_values=[1, 1, 1]))))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52845273,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><dropout>","CreationDate":"2018-10-16T23:25:27.550","AcceptedAnswerId":52845699.0,"OwnerUserId":975238.0,"Title":"Keras weight file load exception: loading 2 layers into a model with 0 layers","Body":"<p>Exception happened when I add dropout to the input layer.<\/p>\n\n<p>The exception was mentioned in other threads as well related to another issues and most common suggested solution is to downgrade the Keras version. Is there a workaround for this exception?<\/p>\n\n<pre><code>def baseline_model() :\n    model = Sequential()\n    model.add(Dropout(0.35)) #THIS LINE CAUSES THE EXCEPTION\n    model.add(Dense(200, input_dim=1200, kernel_initializer='normal', activation='relu'))\n    model.add(Dropout(0.8))\n    rms = RMSprop(lr = 0.00050)\n    model.add(Dense(1, kernel_initializer='normal', activation='sigmoid'))\n    model.compile(loss='binary_crossentropy', optimizer=rms, metrics=['accuracy'])\n    return model\n<\/code><\/pre>\n\n<p>Model throws the following exception during weight file loading:<\/p>\n\n<pre><code>ValueError: You are trying to load a weight file containing 2 layers into a model with 0 layers.\n<\/code><\/pre>\n","answers":[{"AnswerId":"52845699","CreationDate":"2018-10-17T00:29:11.183","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you have not assigned the input shape for the first layer:<\/p>\n\n<pre><code>model.add(Dropout(0.35, input_shape=(1200,)))\n<\/code><\/pre>\n\n<p>And then remove the <code>input_dim<\/code> argument of second layer as it is redundant. <\/p>\n"}]}
+{"QuestionId":52845573,"AnswerCount":0,"Tags":"<python><visual-studio><anaconda><pytorch>","CreationDate":"2018-10-17T00:08:58.033","AcceptedAnswerId":null,"OwnerUserId":6728598.0,"Title":"Using conda to install PyTorch within a Visual Studio python project","Body":"<p>I've created a Python project within Visual Studio 2017 Community Edition.<\/p>\n\n<p>I've changed the environment within the Visual Studio Python project to be Anaconda 5.2.<\/p>\n\n<p>If I just use the \"Install Python Package...\" menu in Visual Studio and type PyTorch, it's says the package is not found.<\/p>\n\n<p>So, I thought next I should just run this command...<\/p>\n\n<pre><code>conda install pytorch -c pytorch\n<\/code><\/pre>\n\n<p>If I open the Python Environments within Visual Studio, click on Anaconda 5.2, and click Open in Powershell, and execute the command above in the Powershell terminal, it eventually fails because of a permission error.  Guessing this is due to not having Administrative Privileges.<\/p>\n\n<p>If I just open the Run menu in Windows and open a Powershell with Administrative Privileges then \"conda\" is not found typing the command above.<\/p>\n\n<p>How should I be installing PyTorch in the Visual Studio Anaconda environment?<\/p>\n","answers":[]}
+{"QuestionId":52845711,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2018-10-17T00:31:06.687","AcceptedAnswerId":52845764.0,"OwnerUserId":1179925.0,"Title":"How to apply TimeDistributed layer on a CNN block?","Body":"<p>Here is my attempt:<\/p>\n\n<pre><code>inputs = Input(shape=(config.N_FRAMES_IN_SEQUENCE, config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS))\n\ndef cnn_model(inputs):\n    x = Conv2D(filters=32, kernel_size=(3,3), padding='same', activation='relu')(inputs)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=32, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=64, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=64, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=128, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    return x\n\nx = TimeDistributed(cnn_model)(inputs)\n<\/code><\/pre>\n\n<p>Which gives the following error:<\/p>\n\n<pre><code>AttributeError: 'function' object has no attribute 'built'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52845764","CreationDate":"2018-10-17T00:38:16.617","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to use <code>Lambda<\/code> layer and wrap your function inside it:<\/p>\n\n<pre><code># cnn_model function the same way as you defined it ...\n\nx = TimeDistributed(Lambda(cnn_model))(inputs)\n<\/code><\/pre>\n\n<p>Alternatively, you can define that block as a model and then apply <code>TimeDistributed<\/code> layer on it:<\/p>\n\n<pre><code>def cnn_model():\n    input_frame = Input(shape=(config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS))\n\n    x = Conv2D(filters=32, kernel_size=(3,3), padding='same', activation='relu')(input_frame)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=32, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=64, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=64, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    x = Conv2D(filters=128, kernel_size=(3,3), padding='same', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n\n    model = Model(input_frame, x)\n    return model\n\ninputs = Input(shape=(config.N_FRAMES_IN_SEQUENCE, config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS))\n\nx = TimeDistributed(cnn_model())(inputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52845785,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><keras-layer>","CreationDate":"2018-10-17T00:41:02.183","AcceptedAnswerId":52845882.0,"OwnerUserId":1179925.0,"Title":"load_model and Lamda layer in Keras","Body":"<p>How to load model that have lambda layer?<\/p>\n\n<p>Here is the code to reproduce behaviour:<\/p>\n\n<pre><code>MEAN_LANDMARKS = np.load('data\/mean_shape_68.npy')\n\ndef add_mean_landmarks(x):\n    mean_landmarks = np.array(MEAN_LANDMARKS, np.float32)\n    mean_landmarks = mean_landmarks.flatten()\n    mean_landmarks_tf = tf.convert_to_tensor(mean_landmarks)\n    x = x + mean_landmarks_tf\n    return x\n\ndef get_model():\n    inputs = Input(shape=(8, 128, 128, 3))\n    cnn = VGG16(include_top=False, weights='imagenet', input_shape=(128, 128, 3))\n    x = TimeDistributed(cnn)(inputs)\n    x = TimeDistributed(Flatten())(x)\n    x = LSTM(256)(x)\n    x = Dense(68 * 2, activation='linear')(x)\n\n    x = Lambda(add_mean_landmarks)(x)\n\n    model = Model(inputs=inputs, outputs=x)\n    optimizer = Adadelta()\n    model.compile(optimizer=optimizer, loss='mae')\n\n    return model\n<\/code><\/pre>\n\n<p>Model compiles and I can save it, but when I tried to load it with <code>load_model<\/code> function I get an error:<\/p>\n\n<pre><code>in add_mean_landmarks\n    mean_landmarks = np.array(MEAN_LANDMARKS, np.float32)\nNameError: name 'MEAN_LANDMARKS' is not defined\n<\/code><\/pre>\n\n<p>\u0410s I understand <code>MEAN_LANDMARKS<\/code> is not incorporated in graph as constant tensor. Also it's related to this question: <a href=\"https:\/\/stackoverflow.com\/questions\/52831235\/how-to-add-constant-tensor-in-keras\/52831446#comment92584381_52831446\">How to add constant tensor in Keras?<\/a><\/p>\n","answers":[{"AnswerId":"52845882","CreationDate":"2018-10-17T00:57:02.080","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to pass <code>custom_objects<\/code> argument to <code>load_model<\/code> function:<\/p>\n\n<pre><code>model = load_model('model_file_name.h5', custom_objects={'MEAN_LANDMARKS': MEAN_LANDMARKS})\n<\/code><\/pre>\n\n<p>Look for more info in Keras docs: <a href=\"https:\/\/keras.io\/getting-started\/faq\/#handling-custom-layers-or-other-custom-objects-in-saved-models\" rel=\"nofollow noreferrer\">Handling custom layers (or other custom objects) in saved models\n<\/a>.<\/p>\n"}]}
+{"QuestionId":52847959,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><scikit-learn><normalizing>","CreationDate":"2018-10-17T05:43:39.947","AcceptedAnswerId":52848051.0,"OwnerUserId":7866526.0,"Title":"Manual normalization function taking too long to execute","Body":"<p>I am trying to implement a normalization function manually rather than using the scikit learn's one. The reason is that, I need to define the maximum and minimum parameters manually and scikit learn doesn't allow that alteration. <\/p>\n\n<p>I successfully implemented this to normalize the values between 0 and 1. But it is taking a very long time to run.<\/p>\n\n<p><strong>Question:<\/strong> <strong>Is there another efficient way I can do this? How can I make this execute faster.<\/strong> <\/p>\n\n<p>Shown below is my code:<\/p>\n\n<pre><code>scaled_train_data = scale(train_data)\n\ndef scale(data):\n    for index, row in data.iterrows():\n        X_std = (data.loc[index, \"Close\"] - 10) \/ (2000 - 10)\n        data.loc[index, \"Close\"] = X_std\n\n    return data\n<\/code><\/pre>\n\n<p>2000 and 10 are the attributes that i defined manually rather than taking the minimum and the maximum value of the dataset.<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"52848051","CreationDate":"2018-10-17T05:50:56.643","ParentId":null,"OwnerUserId":"9553594","Title":null,"Body":"<p>Why loop? You can just use<\/p>\n\n<pre><code>train_data['close'] = (train_data['close'] - 10)\/(2000 - 10) \n<\/code><\/pre>\n\n<p>to make use of vectorized numpy functions. Of course, you could also put this in a function, if you prefer.<\/p>\n\n<p>Alternatively, if you want to rescale to a linear range, you could use <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html<\/a>. The advantage of this is that you can save it and then rescale the test data in the same manner.<\/p>\n"},{"AnswerId":"52848174","CreationDate":"2018-10-17T06:01:16.060","ParentId":null,"OwnerUserId":"10336496","Title":null,"Body":"<p>Use numpy's matrix.you can also set your min and max mannually.<\/p>\n\n<pre><code>import numpy as np\ndata = np.array(df)\n_min = np.min(data, axis=0)\n_max = np.max(data, axis=0)\nnormed_data = (data - _min) \/ (_max - _min)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52848427,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2018-10-17T06:20:41.817","AcceptedAnswerId":52849183.0,"OwnerUserId":9364266.0,"Title":"Merge multiple CNNs","Body":"<p>I am trying to perform <code>Conv1D<\/code> on multiple inputs in my model. So I have 15 inputs of size 1x1500 each, where each one is an input to a series of layers. So I have 15 convolutional models which I want to merge before Fully Connected Layer. I have defined the convolutional model in a function, but I cannot understand how to call the function and then merge them.<\/p>\n\n<pre><code>def defineModel(nkernels, nstrides, dropout, input_shape):\n    model = Sequential()\n    model.add(Conv1D(nkernels, nstrides, activation='relu', input_shape=input_shape))\n    model.add(Conv1D(nkernels*2, nstrides, activation='relu'))\n    model.add(BatchNormalization())\n    model.add(MaxPooling1D(nstrides))\n    model.add(Dropout(dropout))\n    return model\n\n\nmodels = {}\nfor i in range(15):\n    models[i] = defineModel(64,2,0.75,(64,1))\n<\/code><\/pre>\n\n<p>I have successfully concatenated 4 models as follows:<\/p>\n\n<pre><code>merged = Concatenate()([ model1.output, model2.output, model3.output, model4.output])\n\nmerged = Dense(512, activation='relu')(merged)\nmerged = Dropout(0.75)(merged)\nmerged = Dense(1024, activation='relu')(merged)\nmerged = Dropout(0.75)(merged)\nmerged = Dense(40, activation='softmax')(merged)\nmodel = Model(inputs=[model1.input, model2.input, model3.input, model4.input], outputs=merged)\n<\/code><\/pre>\n\n<p>How do I do it for 15 layers in the for loop as writing 15 layers separately isn't efficient?<\/p>\n","answers":[{"AnswerId":"52849183","CreationDate":"2018-10-17T07:13:08.050","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>I think the best you can do is use the functional API everywhere:<\/p>\n\n<pre><code>def defineModel(nkernels, nstrides, dropout, input_shape):\n    l_input = Input( shape=input_shape )\n    model = Conv1D(nkernels, nstrides, activation='relu')(l_input)\n    model = Conv1D(nkernels*2, nstrides, activation='relu')(model)\n    model = BatchNormalization()(model)\n    model = MaxPooling1D(nstrides)(model)\n    model = Dropout(dropout)(model)\n    return model, l_input\n\n\nmodels = []\ninputs = []\nfor i in range(15):\n    model, input = defineModel(64,2,0.75,(64,1))\n    models.append( model )\n    inputs.append( input )\n<\/code><\/pre>\n\n<p>Then it is easy to recover the lists of inputs and the outputs of the submodels and merge them<\/p>\n\n<pre><code>merged = Concatenate()(models)\n\nmerged = Dense(512, activation='relu')(merged)\nmerged = Dropout(0.75)(merged)\nmerged = Dense(1024, activation='relu')(merged)\nmerged = Dropout(0.75)(merged)\nmerged = Dense(40, activation='softmax')(merged)\nmodel = Model(inputs=inputs, outputs=merged)\n<\/code><\/pre>\n\n<p>Normally, these operations are not a bottle neck. None of this should have a major impact during training or inference <\/p>\n"},{"AnswerId":"52851934","CreationDate":"2018-10-17T09:49:47.087","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Of course, as @GabrielM suggested, using functional API is the best way to do this, however if you don't want to modify the <code>define_model<\/code> function you can also do it like this:<\/p>\n\n<pre><code>models = []\ninputs = []\noutputs = []\nfor i in range(15):\n    model = defineModel(64,2,0.75,(64,1))\n    models.append(model)\n    inputs.append(model.input)\n    outputs.append(model.output)\n\n\nmerged = Concatenate()(outputs) # this should be output tensors and not models\n\n# the rest is the same ...\n\nmodel = Model(inputs=inputs, outputs=merged)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52848973,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T06:58:22.413","AcceptedAnswerId":52933548.0,"OwnerUserId":10516807.0,"Title":"TensorFlow use dataset to replace function feed_dict","Body":"<p>when I learn a tensorflow project,find one line code:<\/p>\n\n<pre><code>cls_prob, box_pred = sess.run([output_cls_prob, output_box_pred], feed_dict={input_img: blob})\n<\/code><\/pre>\n\n<p>But, this line code It took a lot of time. (use CPU need 15 seconds...\u252d\u252e\ufe4f\u252d\u252e)<\/p>\n\n<p>By consulting information, I find use function 'dataset' could solve this problem which took a lot of time, How should I use it?<\/p>\n\n<p>source of 'blob':<\/p>\n\n<pre><code>img = cv2.imread('.\/imgs\/001.jpg')\nimg_scale = float(600) \/ min(img_data.shape[0], img_data.shape[1])\nif np.round(img_scale * max(img_data.shape[0], img_data.shape[1])) &gt; 1200:\n    img_scale = float(1200) \/ max(img_data.shape[0], img_data.shape[1])\nimg_data = cv2.resize(img_data, None, None, fx=img_scale, fy=img_scale, interpolation=cv2.INTER_LINEAR)\nimg_orig = img_data.astype(np.float32, copy=True)\nblob = np.zeros((1, img_data.shape[0], img_data.shape[1], 3),dtype=np.float32)\nblob[0, 0:img_data.shape[0], 0:img_data.shape[1], :] = img_orig\n<\/code><\/pre>\n\n<p>source of 'output_cls_prob'&amp;'output_box_pred'&amp;'input_img':<\/p>\n\n<pre><code># Actually,read PB model...\ninput_img = sess.graph.get_tensor_by_name('Placeholder:0')\noutput_cls_prob = sess.graph.get_tensor_by_name('Reshape_2:0')\noutput_box_pred = sess.graph.get_tensor_by_name('rpn_bbox_pred\/Reshape_1:0')\n<\/code><\/pre>\n\n<p>Parameter type\uff1a<\/p>\n\n<pre><code>blob\uff1atype 'numpy.ndarray'\n\noutput_cls_prob\uff1aclass 'tensorflow.python.framework.ops.Tensor'\n\noutput_box_pred\uff1aclass 'tensorflow.python.framework.ops.Tensor'\n\ninput_img\uff1aclass 'tensorflow.python.framework.ops.Tensor'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52851460","CreationDate":"2018-10-17T09:25:01.907","ParentId":null,"OwnerUserId":"8380638","Title":null,"Body":"<p>First of all you should know that the use of Dataset API has a great impact in performance when multiples GPUs are used... Otherwise is almost identical to feed_dict. I recommend you to read <a href=\"https:\/\/stackoverflow.com\/a\/50889965\/8380638\">this other answer<\/a> from a TF developer, it has almost everything one needs to know to create a mental image of the benefits of this new API.<\/p>\n"},{"AnswerId":"52933548","CreationDate":"2018-10-22T16:10:09.590","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p><code>tf.data<\/code> is the recommended API for tensorflow input pipelines. Here is a tutorial on <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">tensorflow.org<\/a>. For your example, the section <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#decoding_image_data_and_resizing_it\" rel=\"nofollow noreferrer\">\"Decoding image data and resizing it\"<\/a> could be most useful. For example, you could do something like:<\/p>\n\n<pre><code># Reads an image from a file, decodes it into a dense tensor, and resizes it\n# to a fixed shape.\ndef _parse_function(filename):\n  image_string = tf.read_file(filename)\n  image_decoded = tf.image.decode_jpeg(image_string)\n  image_resized = tf.image.resize_images(image_decoded, [new_width, new_height])\n  image_resized = tf.expand_dims(image_resized, 0)  # Adds size 1 dimension\n  return image_resized\n\n# A vector of filenames.\nfilenames = tf.constant([\".\/imgs\/001.jpg\", ...])\n\ndataset = tf.data.Dataset.from_tensor_slices(filenames)\ndataset = dataset.map(_parse_function)\n<\/code><\/pre>\n\n<p>And instead of having <code>input_img<\/code> be a placeholder, change:<\/p>\n\n<pre><code>input_img = tf.placeholder(tf.float32)\noutput_class_prob, output_class_pred = (... use input_img ...)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>iterator = dataset.make_one_shot_iterator()\ninput_img = iterator.get_next()\noutput_class_prob, output_class_pred = (... use input_img ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52849514,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><tensorboard><object-detection-api><loss>","CreationDate":"2018-10-17T07:33:02.913","AcceptedAnswerId":52849776.0,"OwnerUserId":10517026.0,"Title":"Plot validation loss in Tensorflow Object Detection API","Body":"<p>I'm using Tensorflow Object Detection API for detection and localization of one class object in images. For these purposes I use pre-trained faster_rcnn_resnet50_coco_2018_01_28 model. <\/p>\n\n<p>I want to detect under\/overfitting after training of the model. I see training loss, but after evaluating Tensorboard only shows mAP and Precision metrics and no loss.<\/p>\n\n<p>Is this possible to plot a validation loss on Tensorboard too?<\/p>\n","answers":[{"AnswerId":"57106424","CreationDate":"2019-07-19T06:06:31.880","ParentId":null,"OwnerUserId":"4979679","Title":null,"Body":"<p>To see the validation curve you should change faster_rcnn_resnet50_coco.config:    <\/p>\n\n<p>1- comment <code>max_evals<\/code> line<br>\n2- set <code>eval_interval_secs<\/code>: 60 .<br>\n3- num_examples should be equal or less than the number of \"files\" that you have in \"val.record\" . <\/p>\n\n<pre><code>eval_config: { . \n  num_examples: 600 . \n  eval_interval_secs: 60 . \n  # Note: The below line limits the evaluation process to 10 evaluations.  \n  # Remove the below line to evaluate indefinitely.  \n  # max_evals: 10 .\n}\n<\/code><\/pre>\n"},{"AnswerId":"52849776","CreationDate":"2018-10-17T07:49:32.993","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>There is validation loss.\nAssuming you're using the latest API, the curve under \"loss\" is validation loss while \"loss_1\/2\" is the training loss.<\/p>\n"}]}
+{"QuestionId":52849935,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T07:57:41.303","AcceptedAnswerId":52850712.0,"OwnerUserId":8380638.0,"Title":"Tensorflow performance: Numpy matrix o TF matrix?","Body":"<p>I have the following code: <\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(init_vars)\n    cols = sess.run(tf.shape(descriptors)[1])\n    descriptor_matrix = np.zeros((n_batches*batch_size, cols))\n    while True:\n        batch_descriptor = sess.run(descriptors, feed_dict={dropout_prob: 1})\n        descriptor_matrix[i:i+elements_in_batch] = np.array(batch_descriptor)\n<\/code><\/pre>\n\n<p>I am mixing tensors and numpy vectors. Does this have an important impact on performance? Why is it? Should I just use tensors instead?<\/p>\n","answers":[{"AnswerId":"52850712","CreationDate":"2018-10-17T08:44:43.020","ParentId":null,"OwnerUserId":"7328893","Title":null,"Body":"<p>It is no problem mixing tensors and numpy.\nBut it is not good to perform processing outside \"graph\"(sess.run()).<\/p>\n\n<p>Tensorflow\/XLA compiler that makes \"graph\" in order to optimize performance.\ngraph is using when sess.run() is called.<\/p>\n\n<p>But, Processing outside graph(sess.run()) is normal python code.\nIt does not benefit from XLA compiler.<\/p>\n"}]}
+{"QuestionId":52850039,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-17T08:04:09.453","AcceptedAnswerId":52971645.0,"OwnerUserId":4513311.0,"Title":"TensorFlow Model is still floating point after Post-training quantization","Body":"<p>After applying post-training quantization, my custom CNN model was shrinked to 1\/4 of its original size (from 56.1MB to 14MB). I put the image(100x100x3) that is to be predicted into ByteBuffer as 100x100x3=30,000 bytes. However, I got the following error during inference:<\/p>\n\n<pre><code>java.lang.IllegalArgumentException: Cannot convert between a TensorFlowLite buffer with 120000 bytes and a ByteBuffer with 30000 bytes.**\n        at org.tensorflow.lite.Tensor.throwExceptionIfTypeIsIncompatible(Tensor.java:221)\n        at org.tensorflow.lite.Tensor.setTo(Tensor.java:93)\n        at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:136)\n        at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:216)\n        at org.tensorflow.lite.Interpreter.run(Interpreter.java:195)\n        at gov.nih.nlm.malaria_screener.imageProcessing.TFClassifier_Lite.recongnize(TFClassifier_Lite.java:102)\n        at gov.nih.nlm.malaria_screener.imageProcessing.TFClassifier_Lite.process_by_batch(TFClassifier_Lite.java:145)\n        at gov.nih.nlm.malaria_screener.Cells.runCells(Cells.java:269)\n        at gov.nih.nlm.malaria_screener.CameraActivity.ProcessThinSmearImage(CameraActivity.java:1020)\n        at gov.nih.nlm.malaria_screener.CameraActivity.access$600(CameraActivity.java:75)\n        at gov.nih.nlm.malaria_screener.CameraActivity$8.run(CameraActivity.java:810)\n        at java.lang.Thread.run(Thread.java:762) \n<\/code><\/pre>\n\n<p>The imput image size to the model is: 100x100x3. I'm currently predicting one image at a time. So, if I'm making the Bytebuffer: 100x100x3 = 30,000 bytes. However, the log info above says the TensorFlowLite buffer has 120,000 bytes. This makes me suspect that the converted tflite model is still in float format. Is this expected behavior? How can I get a quantized model that take input image in 8 pit precision like it does in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/307c83106445ab2c52847f08d35a66c51aff19d9\/tensorflow\/contrib\/lite\/examples\/android\/app\/src\/main\/java\/org\/tensorflow\/demo\/TFLiteObjectDetectionAPIModel.java#L163\" rel=\"nofollow noreferrer\">example<\/a> from TensorFlow official repository ?<\/p>\n\n<p>In the example code, the ByteBuffer used as input for tflite.run() is in 8 bit precision for the quantized model. <\/p>\n\n<p>But I also read from the google doc saying, \"At inference, weights are converted from 8-bits of precision to floating-point and computed using floating point kernels.\" This two instances seems to contradict each other.<\/p>\n\n<pre><code>private static final int BATCH_SIZE = 1;\n\nprivate static final int DIM_IMG_SIZE = 100;\n\nprivate static final int DIM_PIXEL_SIZE = 3;\n\nprivate static final int BYTE_NUM = 1;\n\nimgData = ByteBuffer.allocateDirect(BYTE_NUM * BATCH_SIZE * DIM_IMG_SIZE * DIM_IMG_SIZE * DIM_PIXEL_SIZE);\nimgData.order(ByteOrder.nativeOrder());\n\n... ...\n\nint pixel = 0;\n\n        for (int i = 0; i &lt; DIM_IMG_SIZE; ++i) {\n            for (int j = 0; j &lt; DIM_IMG_SIZE; ++j) {\n\n                final int val = intValues[pixel++];\n\n                imgData.put((byte)((val &gt;&gt; 16) &amp; 0xFF));\n                imgData.put((byte)((val &gt;&gt; 8) &amp; 0xFF));\n                imgData.put((byte)(val &amp; 0xFF));\n\n\/\/                imgData.putFloat(((val &gt;&gt; 16) &amp; 0xFF) \/ 255.0f);\n\/\/                imgData.putFloat(((val &gt;&gt; 8) &amp; 0xFF) \/ 255.0f);\n\/\/                imgData.putFloat((val &amp; 0xFF) \/ 255.0f);\n\n            }\n        } \n\n... ...\n\ntfLite.run(imgData, labelProb);\n<\/code><\/pre>\n\n<p>Post-training quantization code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport sys\nimport os\n\nsaved_model_dir = '\/home\/yuh5\/Downloads\/malaria_thinsmear.h5.pb'\n\ninput_arrays = [\"input_2\"]\n\noutput_arrays = [\"output_node0\"]\n\nconverter = tf.contrib.lite.TocoConverter.from_frozen_graph(saved_model_dir, input_arrays, output_arrays)\n\nconverter.post_training_quantize = True\n\ntflite_model = converter.convert()\nopen(\"thinSmear_100.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52971645","CreationDate":"2018-10-24T14:33:46.417","ParentId":null,"OwnerUserId":"2344535","Title":null,"Body":"<p>Post-training quantization does not change the format of the input or output layers. You can run your model with data in the same format as used for training.<\/p>\n\n<p>You may look into quantization-aware training to generate fully-quantized models, but I have no experience with it.<\/p>\n\n<p>As for the sentence \"At inference, weights are converted from 8-bits of precision to floating-point and computed using floating point kernels.\" This means that the weights are \"de-quantized\" to floating point values in memory, and computed with FP instructions, instead of performing integer operations.<\/p>\n"}]}
+{"QuestionId":52850105,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-17T08:07:45.040","AcceptedAnswerId":52851738.0,"OwnerUserId":3494669.0,"Title":"keras Concatenate mulitple layers cause AttributeError: 'NoneType' object has no attribute '_inbound_nodes'","Body":"<p>I am trying to add some fixed <a href=\"https:\/\/en.wikipedia.org\/wiki\/Kernel_(image_processing)\" rel=\"nofollow noreferrer\">kernels<\/a> in my CNN, please see my codes below.<\/p>\n\n<p>This is how I create my kernels:<\/p>\n\n<pre><code># Kernels\ndef create_kernel(x):\n    t = pipe(\n        x, \n        lambda x: tf.constant(x, dtype=tf.float32), \n        lambda x: tf.reshape(x, [3, 3, 1, 1]))\n    return t\n\nk_edge1 = create_kernel([1, 0, -1, 0, 0, 0, -1, 0, 1])\nk_edge2 = create_kernel([0, 1, 0, 1, -4, 1, 0, 1, 0])\nk_edge3 = create_kernel([-1, -1, -1, -1, 8, -1, -1, -1, -1])\n<\/code><\/pre>\n\n<p>and my convolution network is like:<\/p>\n\n<pre><code># Convolution network\n# Input layer\nl_input = Input(shape=(28**2, ))\n# Reshape layer\nl_reshape = Reshape(target_shape=(28, 28, 1))(l_input)\n# Convolution layers\nl_conv1 = Conv2D(filters=20, kernel_size=(3, 3), padding='valid')(l_reshape)\nl_edge1 = tf.nn.conv2d(l_reshape, k_edge1, strides=[1, 1, 1, 1], padding='VALID')\nl_edge2 = tf.nn.conv2d(l_reshape, k_edge2, strides=[1, 1, 1, 1], padding='VALID')\nl_edge3 = tf.nn.conv2d(l_reshape, k_edge3, strides=[1, 1, 1, 1], padding='VALID')\nl_conv1a = Concatenate(axis=3)([l_conv1, l_edge1, l_edge2, l_edge3])  # &lt;- The error should be caused by this line.\nl_conv2 = Conv2D(filters=20, kernel_size=(3, 3), padding='valid')(l_conv1a)\nl_pool1 = MaxPooling2D(pool_size=(2, 2), border_mode='valid')(l_conv2)\n# Flatten layer\nl_flat = Flatten()(l_pool1)\n# Fully connected layers\nl_fc1 = Dense(50, kernel_initializer='he_normal')(l_flat)\nl_act1 = PReLU()(l_fc1)\nl_fc3 = Dense(10, kernel_initializer='he_normal')(l_act1)\nl_output = Activation('softmax')(l_fc1)\n\n# Model\ncnn_model = Model(l_input, l_output)\n<\/code><\/pre>\n\n<p>However, I got the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"&lt;stdin&gt;\", line 2, in &lt;module&gt;\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 93, in __init__\n    self._init_graph_network(*args, **kwargs)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 237, in _init_graph_network\n    self.inputs, self.outputs)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 1353, in _map_graph_network\n    tensor_index=tensor_index)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 1340, in build_map\n    node_index, tensor_index)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 1340, in build_map\n    node_index, tensor_index)\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 1340, in build_map\n    node_index, tensor_index)\n[Previous line repeated 2 more times]\nFile \"C:\\Users\\Perry Cheng\\AppData\\Local\\conda\\conda\\envs\\ml_py_3_6\\lib\\site-packages\\keras\\engine\\network.py\", line 1312, in build_map\n    node = layer._inbound_nodes[node_index]\nAttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n<\/code><\/pre>\n\n<p>After some testing, I think the error comes from:<\/p>\n\n<pre><code>l_conv1a = Concatenate(axis=3)([l_conv1, l_edge1, l_edge2, l_edge3])\n<\/code><\/pre>\n\n<p>Is there any way to solve it?<\/p>\n","answers":[{"AnswerId":"52851738","CreationDate":"2018-10-17T09:39:20.600","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Keras layers accepts Keras Tensors and not Tensors as their input. So if you would like to use <code>tf.nn.conv2d<\/code> instead of <code>Conv2D<\/code> layers in Keras, you need to wrap them inside a <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>l_edge1 = Lambda(lambda x: tf.nn.conv2d(x, k_edge1, strides=[1, 1, 1, 1], padding='VALID'))(l_reshape)\nl_edge2 = Lambda(lambda x: tf.nn.conv2d(x, k_edge2, strides=[1, 1, 1, 1], padding='VALID'))(l_reshape)\nl_edge3 = Lambda(lambda x: tf.nn.conv2d(x, k_edge3, strides=[1, 1, 1, 1], padding='VALID'))(l_reshape)\n<\/code><\/pre>\n"},{"AnswerId":"52853319","CreationDate":"2018-10-17T11:00:49.290","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You cannot use TF functions directly on Keras tensors as you are doing here:<\/p>\n\n<pre><code>l_edge1 = tf.nn.conv2d(l_reshape, k_edge1, strides=[1, 1, 1, 1], padding='VALID')\nl_edge2 = tf.nn.conv2d(l_reshape, k_edge2, strides=[1, 1, 1, 1], padding='VALID')\nl_edge3 = tf.nn.conv2d(l_reshape, k_edge3, strides=[1, 1, 1, 1], padding='VALID')\n<\/code><\/pre>\n\n<p>What you should do is to just use the Conv2D layer and then set the weights manually using <code>layer.set_weights(array)<\/code>. To keep weights non-trainable, just set <code>layer.trainable = False<\/code>, like:<\/p>\n\n<pre><code>conv = Conv2D(filters=1, kernel_size(3, 3), padding='valid')\nconv.set_weights(your_weight_array)\nconv.trainable = False\n\nl_edge1 = conv(l_reshape)\n<\/code><\/pre>\n\n<p>And similarly for the other two Conv2D layers.<\/p>\n"}]}
+{"QuestionId":52850196,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T08:13:21.587","AcceptedAnswerId":52855928.0,"OwnerUserId":9540764.0,"Title":"Take an element of a tensor which is inside also in another tensor","Body":"<p>I have two tensors and I have to iterate on the first to take only the element that is inside the other tensor. There is only one element in <code>t2<\/code> that it is also inside <code>t1<\/code>. Here an example<\/p>\n\n<pre><code>t1 = tf.where(values &gt; 0) # I get some indices example [6, 0], [3, 0]\nt2 = tf.where(values2 &gt; 0) # I get [4, 0], [3, 0]\n\nt3 = .... # [3, 0]\n<\/code><\/pre>\n\n<p>I've tried to evaluate and iterate over them using <code>.eval()<\/code> and checked if an element of <code>t2<\/code> is in <code>t1<\/code> using the operator <code>in<\/code>, but doesn't work. Is there a function from TensorFlow that can do that?<\/p>\n\n<p><strong>edit<\/strong><\/p>\n\n<pre><code>for index in xrange(max_indices):\n    indices = tf.where(tf.equal(values, (index + 1))).eval() # indices: [[1 0]\\n [4 0]\\n [9 0]]\n    cent_indices = tf.where(centers &gt; 0).eval() # cent_indices: [[6 0]\\n [9 0]]\n    indices_list.append(indices)\n    for cent in cent_indices:\n        if cent in indices:\n           centers_list.append(cent)\n           break\n<\/code><\/pre>\n\n<p>The first iteration <code>cent<\/code> has the value <code>[6 0]<\/code> but it enters the <code>if<\/code> condition.<\/p>\n\n<p><strong>answer<\/strong><\/p>\n\n<pre><code>for index in xrange(max_indices):\n    indices = tf.where(tf.equal(values, (index + 1))).eval()\n    cent_indices = tf.where(centers &gt; 0).eval()\n    indices_list.append(indices)\n    for cent in cent_indices:\n        # batch_item is an iterator from an outer loop\n        if values[batch_item, cent[0]].eval() == (index + 1):\n           centers_list.append(tf.constant(cent))\n           break\n<\/code><\/pre>\n\n<p>The solution is related to my task, but if you are looking for a solution in 1D tensor I suggest to have a look on <code>tf.sets.set_intersection<\/code><\/p>\n","answers":[{"AnswerId":"52855928","CreationDate":"2018-10-17T13:20:25.447","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>Is that what you wanted ? I used just these two test cases.<\/p>\n\n<pre><code>x = tf.constant([[1, 2, 3, 4, 5, 6], [1, 2, 3, 4, 5, 1]])\ny = tf.constant([[1, 2, 3, 4, 3, 6], [1, 2, 3, 4, 5, 1]])\n# x = tf.constant([[1, 2], [4, 5], [7, 7]])\n# y = tf.constant([[7, 7], [3, 5]])\n\ndef match(xiterations, yiterations, yvalues, xvalues ):\n    for i in range(xiterations):\n        for j in range(yiterations):\n            if (np.array_equal(yvalues[j], xvalues[i])):\n                print( yvalues[j])\n\nwith tf.Session() as sess:\n    xindex = tf.where( x &gt; 4 )\n    yindex = tf.where( y &gt; 4 )\n\n    xvalues = xindex.eval()\n    yvalues = yindex.eval()\n\n    xiterations =  tf.shape(xvalues)[0].eval()\n    yiterations =  tf.shape(yvalues)[0].eval()\n\n    print(tf.shape(xvalues)[0].eval())\n    print(tf.shape(yvalues)[0].eval())\n\n    if tf.shape(xvalues)[0].eval() &gt;= tf.shape(yvalues)[0].eval():\n        match( xiterations, yiterations, yvalues, xvalues)\n    else:\n        match( yiterations, xiterations, xvalues, yvalues)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52850268,"AnswerCount":0,"Tags":"<lua><scipy><pytorch><torch><mat>","CreationDate":"2018-10-17T08:17:59.517","AcceptedAnswerId":null,"OwnerUserId":7282956.0,"Title":"Why there is diiference between the matrix after loading .mat file from Lua torch and pytorch using scipy?","Body":"<p>I have the output matrix from a .mat file using Lua torch 7 and pytorch with scipy both. I am getting the single bit different matrix in using pytorch . Please some one explain me why ? In pytorch matric is in float point, but in Lua torch it has fixe point. How do I fix the point in pytorch.<\/p>\n\n<p>Here is the matrix.<\/p>\n\n<p>From Scipy and pytorch Result:<\/p>\n\n<pre><code>tensor([[[112.,  -1.,  -1.,  -1.,  -1.,  -1.],\n     [ -1.,  -1., 130., 138.,  -1.,  -1.]],\n\n    [[ 57.,  -1.,  -1.,  -1.,  -1.,  -1.],\n     [ -1.,  -1.,  11.,  53.,  -1.,  -1.]],\n\n    [[ 58.,  -1.,  -1.,  -1.,  -1.,  -1.],\n     [ -1.,  -1.,  31.,  28.,  -1.,  -1.]],\n\n    [[ 46.,  -1.,  -1.,  -1.,  -1.,  -1.],\n     [ -1.,  -1.,  24.,  28.,  -1.,  -1.]]])\n<\/code><\/pre>\n\n<p>From Lua torch7 result <\/p>\n\n<pre><code>[[[113   -1   -1   -1   -1   -1]\n   [-1   -1  131  139   -1   -1]],\n   [[57   -1   -1   -1   -1   -1],\n   [-1   -1   11   53   -1   -1]],\n   [[58   -1   -1   -1   -1   -1],\n   [-1   -1   31   29   -1   -1]],\n   [[47   -1   -1   -1   -1   -1],\n   -[1   -1   24   28   -1   -1]]]\n<\/code><\/pre>\n\n<p>you can see the first  element <code>112<\/code> in pytorch but <code>113<\/code> in lua torch. sSimilary <code>130 &amp; 131<\/code> , <code>138 &amp; 139<\/code> , <code>28 &amp; 29<\/code> and <code>46 and 47<\/code>. I want the result similar to Lua torch which is the second matrix. Is there any method to get the fix point in pytorch also.  Please help me out. Thanks<\/p>\n","answers":[]}
+{"QuestionId":52851080,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-17T09:04:08.203","AcceptedAnswerId":null,"OwnerUserId":6407393.0,"Title":"tensorflow profile explanation","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/I2RlE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/I2RlE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I use tensorflow profile to test the inference of my model and here is the profile details. I find that there are <code>0,1,2,3<\/code>, four numbers where <code>1<\/code> and <code>2<\/code> are filled with blank. So what is the meaning of 0-4 and why there are blanks in <code>1<\/code> and <code>2<\/code>. \nThe machine has 80 cores and does it mean that the inference course only occupy 4 cores of them ?\nThanks.<\/p>\n","answers":[{"AnswerId":"52855065","CreationDate":"2018-10-17T12:38:57.550","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>I suppose that each row corresponds to each worker thread to run operators.\n So your inference processing only occupies 4 cores as you say. \nTensorflow uses multi-threads when <\/p>\n\n<ol>\n<li>There are some independent graph parts.<\/li>\n<li>There is a operator using multi-threads.<\/li>\n<\/ol>\n\n<p>So you can use multi-core effectively, if your graph have many independent graph parts.<\/p>\n\n<p>In the following code, the graph has many independent graph parts. Therefore the number of the rows in profiler matches to \"inter_op_parallelism_threads\".<\/p>\n\n<pre><code>config = tf.ConfigProto(inter_op_parallelism_threads=5, intra_op_parallelism_threads=1)\n\nwith tf.device(\"\/cpu:0\"):\n    list_r = []\n\n    for i in range(80):\n        r = tf.random_normal(shape=[100, 100])\n        list_r.append(r)\n\n    v = tf.add_n(list_r)\n\n\nglobal_step = tf.train.create_global_step()\nhook = tf.train.ProfilerHook(save_steps=1)\nincrement_global = global_step.assign_add(1)\n\n\nwith tf.train.SingularMonitoredSession(hooks=[hook], config=config) as sess:\n    sess.run([v, increment_global])\n<\/code><\/pre>\n\n<p>If you want to know the detail of ConfigProto, you can get information from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto<\/a><\/p>\n"}]}
+{"QuestionId":52851474,"AnswerCount":1,"Tags":"<python><tensorflow><keras><visualization>","CreationDate":"2018-10-17T09:25:34.947","AcceptedAnswerId":null,"OwnerUserId":7649256.0,"Title":"Visualizing intermediate layers in a custom model from keras","Body":"<p>I am currently running a ConvLSTM model from the official Keras repository. One of the more difficult parts about understanding the model is the difficulty to conceive what is happening at each stage of the process. The official code can be found <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/conv_lstm.py\" rel=\"nofollow noreferrer\">here<\/a>.\nLooking at the sequence model, the multiple occurances of the batch norm and convolutional LSTM layers are not obvious choices and is pretty confusing to understand what they actually do. <\/p>\n\n<pre><code>seq = Sequential()\nseq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),\n               input_shape=(None, 40, 40, 1),\n               padding='same', return_sequences=True))\nseq.add(BatchNormalization())\n\nseq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),\n               padding='same', return_sequences=True))\nseq.add(BatchNormalization())\n\nseq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),\n               padding='same', return_sequences=True))\nseq.add(BatchNormalization())\n\nseq.add(ConvLSTM2D(filters=40, kernel_size=(3, 3),\n               padding='same', return_sequences=True))\nseq.add(BatchNormalization())\n<\/code><\/pre>\n\n<p>A bleak understanding of the same is that it helps to extract the frame's features leveraging the usefulness of convolutions when it comes to images. [The input is the frames of a video].\nIt would be really helpful if I can visualize the output from each layer,it can help me better understand what the convolutionalLSTM layers are contributing to, at each stage. Any suggestions on how to go about it or pointers in that direction is appreciated. Thank you for your time.<\/p>\n","answers":[{"AnswerId":"52854381","CreationDate":"2018-10-17T11:58:47.713","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>If you are familiar only with Keras style graph construction, I recommend that you make another Sequential which shares some layers with the origianl Sequential like<\/p>\n\n<pre><code>seq = tf.keras.Sequential()\nd1 = layers.Dense(units=1, use_bias=False, \n                  kernel_initializer=tf.initializers.constant(2.0))\nd2 = layers.Dense(units=1, use_bias=False, \n                  kernel_initializer=tf.initializers.constant(3.0))\nseq.add(d1)\nseq.add(d2)\n\nseq2 = tf.keras.Sequential()\nseq2.add(d1)\n\nprint (seq.predict(np.ones(shape=[1,1])))\nprint (seq2.predict(np.ones(shape=[1,1])))\n<\/code><\/pre>\n\n<p>In the above case, you can get the value of intermediate layer d1. <\/p>\n\n<p>If you often use raw tf.Session(), you can make customized call() method for your model<\/p>\n\n<pre><code>class MultiOut(tf.keras.Model):\n    def __init__(self, name=\"original\"):\n        super().__init__(name=name)\n        self.d1 = layers.Dense(units=1, use_bias=False, \n                               kernel_initializer=tf.initializers.constant(2.0))\n        self.d2 = layers.Dense(units=1, use_bias=False, \n                               kernel_initializer=tf.initializers.constant(3.0))\n\n    def call(self, inputs, multiout=False):\n        d1 = self.d1(inputs)\n        d2 = self.d2(d1)\n\n        if not multiout:\n            return d2\n        else:\n            return d1, d2\n\nmodel = MultiOut()\n\ninput = np.ones(shape=[1,1])\nprint (model.predict(input))\n\nsess = tf.keras.backend.get_session()\nts_input = tf.constant(input, dtype=tf.float32)\n\nprint (sess.run(model(ts_input, multiout=True)))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52851618,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-10-17T09:32:46.990","AcceptedAnswerId":null,"OwnerUserId":6595663.0,"Title":"Input Tensor into tensorflow graph","Body":"<p>I am following tutorial on <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">simple audio recognition<\/a> and currently editing the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/speech_commands\/label_wav.py\" rel=\"nofollow noreferrer\">label_wav.py<\/a>.  In original case we input wave file and the graph predicts the label (in between it calculates spectrum, mfcc's inside the graph). Now i am looking to input mfcc's directly rather than inputting the wave file. <a href=\"https:\/\/github.com\/BhanuKiranChaluvadi\/tensorflow_exericse\/blob\/master\/speech_recognition_test_nn.png\" rel=\"nofollow noreferrer\">Run the graph<\/a> by inputting the mfcc tensor. <\/p>\n\n<pre><code># mfccs:  Tensor(\"strided_slice:0\", shape=(1, 98, 40), dtype=float32)\nmfcc_input_layer_name = 'Reshape:0'\nwith tf.Session() as sess:\n    predictions, = sess.run(softmax_tensor, {mfcc_input_layer_name: mfcc})\n<\/code><\/pre>\n\n<p>After a bit of googling, i found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3389\" rel=\"nofollow noreferrer\">some discussion in git<\/a> and created a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_session_handle\" rel=\"nofollow noreferrer\">session_handle<\/a>.<\/p>\n\n<pre><code># mfccs:  Tensor(\"strided_slice:0\", shape=(1, 98, 40), dtype=float32)\nmfcc_input_layer_name = 'Reshape:0'\nwith tf.Session() as sess:\n      h = tf.get_session_handle(mfccs)\n      h = sess.run(h)\n      predictions, = sess.run(softmax_tensor, {mfcc_input_layer_name: h})\n<\/code><\/pre>\n\n<p><strong>The code is working as expected<\/strong> but I am wondering if there could be a better way of dealing with the tensor rather than creating the handle and then passing it?<\/p>\n","answers":[{"AnswerId":"52856047","CreationDate":"2018-10-17T13:25:45.470","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>I suppose you want to replace an intermediate Tensor with a value by feed_dict. If you have a Tensor object, you can replace it by feed_dict as the following<\/p>\n\n<pre><code>a = tf.constant(3, name=\"a\")\nb = tf.constant(4, name=\"b\")\nc = tf.add(a, b, name=\"c\")\nd = c * 3\n\nwith tf.Session() as sess:\n    print(sess.run(d))    \n    print(sess.run(d, feed_dict={c: 2}))\n<\/code><\/pre>\n\n<p>Even though you don't have the Tensor object, you can get it by get_tensor_by_name <\/p>\n\n<pre><code>a = tf.constant(3, name=\"a\")\nb = tf.constant(4, name=\"b\")\nc = tf.add(a, b, name=\"c\")\nd = c * 3\n\nwith tf.Session() as sess:\n    c_tensor = tf.get_default_graph().get_tensor_by_name(\"c:0\")\n    print(sess.run(d, feed_dict={c_tensor: 2}))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52851866,"AnswerCount":2,"Tags":"<python><machine-learning><keras><conv-neural-network><visual-recognition>","CreationDate":"2018-10-17T09:46:32.697","AcceptedAnswerId":52852075.0,"OwnerUserId":2423198.0,"Title":"How to deal with thousands of images for CNN training Keras","Body":"<p>I have ~10000k images that cannot fit in memory. So for now I can only read 1000 images and train on it... <\/p>\n\n<p>My code is here :<\/p>\n\n<pre><code>img_dir = \"TrainingSet\" # Enter Directory of all images \nimage_path = os.path.join(img_dir+\"\/images\",'*.bmp')\nfiles = glob.glob(image_path)\nimages = []\nmasks = []\ncontours = []\nindexes = []\nfiles_names = []\n\nfor f1 in np.sort(files):\n  img = cv2.imread(f1)\n  result = re.search('original_cropped_(.*).bmp', str(f1))\n  idx = result.group(1)\n  mask_path = img_dir+\"\/masks\/mask_cropped_\"+str(idx)+\".bmp\"\n  mask = cv2.imread(mask_path,0)\n  contour_path = img_dir+\"\/contours\/contour_cropped_\"+str(idx)+\".bmp\"\n  contour = cv2.imread(contour_path,0)\n\n  indexes.append(idx)\n  images.append(img)\n  masks.append(mask)\n  contours.append(contour)\n\ntrain_df = pd.DataFrame({\"id\":indexes,\"masks\": masks, \"images\": images,\"contours\": contours })\ntrain_df.sort_values(by=\"id\",ascending=True,inplace=True)\nprint(train_df.shape)\n\nimg_size_target = (256,256)\n\nids_train, ids_valid, x_train, x_valid, y_train, y_valid, c_train, c_valid = train_test_split(\n    train_df.index.values,\n    np.array(train_df.images.apply(lambda x: cv2.resize(x,img_size_target).reshape(img_size_target[0],img_size_target[1],3))), \n    np.array(train_df.masks.apply(lambda x: cv2.resize(x,img_size_target).reshape(img_size_target[0],img_size_target[1],1))), \n    np.array(train_df.contours.apply(lambda x: cv2.resize(x,img_size_target).reshape(img_size_target[0],img_size_target[1],1))), \n    test_size=0.2, random_state=1337)\n\n#Here we define the model architecture... \n#.....\n#End of model definition\n\n# Training \noptimizer = Adam(lr=1e-3,decay=1e-10)    \nmodel.compile(loss=\"binary_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n\nearly_stopping = EarlyStopping(patience=10, verbose=1)\nmodel_checkpoint = ModelCheckpoint(\".\/keras.model\", save_best_only=True, verbose=1)\nreduce_lr = ReduceLROnPlateau(factor=0.5, patience=5, min_lr=0.00001, verbose=1)\n\nepochs = 200\nbatch_size = 32\n\nhistory = model.fit(x_train, y_train,\n                validation_data=[x_valid, y_valid], \n                epochs=epochs,\n                batch_size=batch_size,\n                callbacks=[early_stopping, model_checkpoint, reduce_lr])\n<\/code><\/pre>\n\n<p>What I would like to know is how can I modify my code in order to do batches of a small set of images without loading all the other 10000 into memory ? which means that the algorithm will read X images each epoch from directory and train on it and after that goes for the next X until the last one.  <\/p>\n\n<p>X here would be a reasonable amount of images that can fit into memory. <\/p>\n","answers":[{"AnswerId":"52852075","CreationDate":"2018-10-17T09:56:50.870","ParentId":null,"OwnerUserId":"7408462","Title":null,"Body":"<p>use fit_generator instead of fit<\/p>\n\n<pre><code>def generate_batch_data(num):\n    #load X images here\n    return images\n\nmodel.fit_generator(generate_batch_data(X),\n        samples_per_epoch=10000, nb_epoch=10)\n<\/code><\/pre>\n\n<p>Alternative you could use train_on_batch instead of fit<\/p>\n\n<p>Discussion on GitHub about this topic: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2708\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/2708<\/a><\/p>\n"},{"AnswerId":"52852098","CreationDate":"2018-10-17T09:57:45.683","ParentId":null,"OwnerUserId":"1122645","Title":null,"Body":"<p><code>np.array(train_df.images.apply(lambda x:cv2.resize(x,img_size_target).reshape(img_size_target[0],img_size_target[1],3)))<\/code><\/p>\n\n<p>You can first apply this filter (and the 2 others) to each individual file and save them to a special folder (images_prepoc, masks_preproc,etc.. ) in a separate script, then load them back already ready for use in the current script.<\/p>\n\n<p>Assuming that the actual images dimensions are greater than 256x256, you will have a faster algorithm, using less memory at the cost of a single preparation phase.<\/p>\n"}]}
+{"QuestionId":52852742,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-17T10:29:37.003","AcceptedAnswerId":52852960.0,"OwnerUserId":3392589.0,"Title":"how to read tensorflow confusion matrix rows and columns","Body":"<p>For my 2 classes (<code>1 = [0, 1]<\/code> and <code>0 = [1, 0]<\/code>) CNN model I use <code>tf.confusion_matrix<\/code> to finding a confusion matrix for the model. one of my results is like below for validation set:<\/p>\n\n<pre><code>[ [1800  17] \n  [283  600] ]\n<\/code><\/pre>\n\n<p>after doing some search I see more than one type of reading, some of them say <code>[[TN  FP][FN  TP]]<\/code>, but some others read it in this way <code>[[TP  FP][FN  TN]]<\/code>, I am confused which one is right for my case? please give me an answer that depends on scientific research if you can.<\/p>\n","answers":[{"AnswerId":"52852960","CreationDate":"2018-10-17T10:40:15.323","ParentId":null,"OwnerUserId":"8380638","Title":null,"Body":"<p>The truth is behind the code ;)\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/confusion_matrix.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/confusion_matrix.py<\/a><\/p>\n\n<blockquote>\n  <p>Class labels are expected to   start at 0. For example, if\n  <code>num_classes<\/code> is 3, then the possible labels   would be <code>[0, 1, 2]<\/code>.\n   Note that the possible labels are assumed to be <code>[0, 1, 2, 3, 4]<\/code>,\n    resulting in a 5x5 confusion matrix.<\/p>\n<\/blockquote>\n\n<p>So better don't pass one hot tensors to the function ;) (tf.argmax might be a good friend here)<\/p>\n\n<p>This means that the first element (row 0 col 0) corresponds with the number of elements that have been properly classified for class 0.<\/p>\n\n<p>Row 0 col 1 will correspond with the missclassified elements of the class 0 and so on.<\/p>\n"}]}
+{"QuestionId":52852885,"AnswerCount":0,"Tags":"<python><machine-learning><keras><keras-2>","CreationDate":"2018-10-17T10:36:28.250","AcceptedAnswerId":null,"OwnerUserId":139150.0,"Title":"Unexpected Keras predictions","Body":"<p>I tried the keras tutorial that I found here...<\/p>\n\n<p><a href=\"https:\/\/github.com\/eijaz1\/Building-a-CNN-in-Keras-Tutorial\/blob\/master\/cnn_tutorial.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/eijaz1\/Building-a-CNN-in-Keras-Tutorial\/blob\/master\/cnn_tutorial.ipynb<\/a><\/p>\n\n<p>Everything worked fine till line 10. \nBut I am not able to predict correctly. I get the results like this...<\/p>\n\n<pre><code>model.predict(X_test[:4])\n\narray([[0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],\n       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],\n       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],\n       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)\n<\/code><\/pre>\n\n<p>The expected results as per the tutorial are:<\/p>\n\n<pre><code>array([[1.6117248e-09, 8.6684462e-16, 6.8095707e-10, 1.5486043e-08,\n        6.2878847e-14, 1.2934288e-15, 1.1453808e-16, 9.9999928e-01,\n        1.0626109e-08, 6.9729606e-07],\n       [1.3555871e-07, 2.6465393e-06, 9.9999511e-01, 2.0351818e-08,\n        1.9796262e-11, 1.6996018e-12, 2.1163373e-06, 1.2008194e-17,\n        4.8792381e-10, 2.6086671e-13],\n       [6.7238901e-08, 9.9785548e-01, 1.9031411e-04, 3.9194603e-08,\n        1.2894072e-04, 1.5791730e-06, 1.2754040e-06, 4.1349044e-09,\n        1.8221687e-03, 5.5910935e-08],\n       [9.9999356e-01, 1.6909821e-12, 8.2496926e-10, 1.7359107e-11,\n        1.7359230e-12, 1.8865266e-13, 6.4659162e-06, 2.3738855e-11,\n        1.1319052e-08, 2.6948474e-08]], dtype=float32)\n<\/code><\/pre>\n\n<p>I am using keras version 2.2.2 if that matters. <\/p>\n\n<hr>\n\n<p>Update:<\/p>\n\n<p>While training the model, I am getting pretty low accuracy compared to that tutorial. <\/p>\n\n<pre><code>model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=3)\n\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/3\n60000\/60000 [==============================] - 75s 1ms\/step - loss: 14.3141 - acc: 0.1118 - val_loss: 14.2677 - val_acc: 0.1148\nEpoch 2\/3\n60000\/60000 [==============================] - 74s 1ms\/step - loss: 14.3741 - acc: 0.1082 - val_loss: 14.4692 - val_acc: 0.1023\nEpoch 3\/3\n60000\/60000 [==============================] - 134s 2ms\/step - loss: 14.3691 - acc: 0.1085 - val_loss: 14.3483 - val_acc: 0.1098\n<\/code><\/pre>\n\n<p>How do I improve the accuracy? I am using exactly same code as shown in the tutorial.<\/p>\n\n<p>Here is the output that I get even if I use the exactly same code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/shantanuo\/Building-a-CNN-in-Keras-Tutorial\/blob\/master\/cnn_tutorial_mismatch.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/shantanuo\/Building-a-CNN-in-Keras-Tutorial\/blob\/master\/cnn_tutorial_mismatch.ipynb<\/a><\/p>\n","answers":[]}
+{"QuestionId":52853080,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><deep-learning>","CreationDate":"2018-10-17T10:47:26.773","AcceptedAnswerId":52854184.0,"OwnerUserId":6349472.0,"Title":"Use of L2-normalization in style transfer - no weights involved?","Body":"<p>I've been going through Chollet's Deep Learning with Python, where he briefly covers L2-normalization with regards to Keras. I understand that it prevents overfitting by adding a penalty proportionate to the sum of the square of the weights to the cost function of the layer, helping to keep weights small.<\/p>\n\n<p>However, in the section covering artistic style transfer, the content loss as a measure is described as:<\/p>\n\n<blockquote>\n  <p>the L2 norm between the activations of\n  an upper layer in a pretrained convnet, computed over the target image, and the activations\n  of the same layer computed over the generated image. This guarantees that, as\n  seen from the upper layer, the generated image will look similar.<\/p>\n<\/blockquote>\n\n<p>The style loss is also related to the L2-norm, but let's focus on the content loss for now.<\/p>\n\n<p>So, the relevant code snippet (p.292):<\/p>\n\n<pre><code>def content_loss(base, combination):\n    return K.sum(K.square(combination - base))\n\n\n\noutputs_dict = dict([(layer.name, layer.output) for layer in model.layers])\ncontent_layer = 'block5_conv2'\nstyle_layers = ['block1_conv1',\n            'block2_conv1',\n            'block3_conv1',\n            'block4_conv1',\n            'block5_conv1']\ntotal_variation_weight = 1e-4\nstyle_weight = 1.\ncontent_weight = 0.025\n#K here refers to the keras backend\nloss = K.variable(0.)\n\nlayer_features = outputs_dict[content_layer]\ntarget_image_features = layer_features[0, :, :, :]\ncombination_features = layer_features[2, :, :, :]\nloss += content_weight * content_loss(target_image_features,\n                         combination_features)\n<\/code><\/pre>\n\n<p>I don't understand why we use the outputs of each layer, which are image feature maps, as opposed to Keras's <code>get_weights()<\/code> method to fetch the weights to perform normalization. I do not follow how using L2-normalization on these feature maps penalizes during training, or moreover what exactly is it penalizing?<\/p>\n","answers":[{"AnswerId":"52854184","CreationDate":"2018-10-17T11:48:49.103","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<blockquote>\n  <p>I understand that it prevents overfitting by adding a penalty proportionate to the sum of the square of the weights to the cost function of the layer, helping to keep weights small.<\/p>\n<\/blockquote>\n\n<p>What you are referring to is <strong>(weight) regularization<\/strong> and in this case, it is L2-regularization. The L2-norm of a vector is the sum of squared of its elements and therefore when you apply L2-regularization on the weights (i.e. parameters) of a layer it would be considered (i.e. added) in the loss function. Since we are minimizing the loss function the side effect is that the L2-norm of the weights will be reduced as well which in turn means that the value of weights has been reduced (i.e. small weights).<\/p>\n\n<p>However, in the style transfer example the content loss is defined as the L2-norm (or L2-loss in this case) of the difference of between activation (and not weights) of a specific layer (i.e. <code>content_layer<\/code>) when applied on the target image and the combination image (i.e. target image + style):<\/p>\n\n<pre><code>return K.sum(K.square(combination - base)) # that's exactly the definition of L2-norm\n<\/code><\/pre>\n\n<p>So no weight regularization is involved here. Rather, the <strong>loss function<\/strong> used is the L2-norm and it is used as a measure of similarity of two arrays (i.e. activations of the content layer). The smaller the L2-norm, the more similar the activations.<\/p>\n\n<p>Why activations of the layer and not its weights? Because we want to make sure that the <strong>contents<\/strong> (i.e. representations given by the <code>content_layer<\/code>) of the target image and the combination image are similar. Note that weights of a layer are fixed and does not change (after training, of course) with respect to an input image; rather, they are used to <strong>describe<\/strong> or <strong>represent<\/strong> an specific input image, and that representation is called <strong>activations<\/strong> of that layer for that specific image.<\/p>\n"}]}
+{"QuestionId":52853955,"AnswerCount":1,"Tags":"<python><json><keras><save>","CreationDate":"2018-10-17T11:36:13.950","AcceptedAnswerId":52872839.0,"OwnerUserId":9805238.0,"Title":"Saving and retrieving model name","Body":"<p>I'm using Keras to train my model and this is how I specify its name:<\/p>\n\n<pre><code>model = Sequential()\nmodel.name = 'NameOfMyModel'\n[...]\n<\/code><\/pre>\n\n<p>Afterwards I save it to the disk as follows:<\/p>\n\n<pre><code>model_json = model.to_json()\nwith open(pathToModel+\".json\",\"w\") as json_file:\n    json_file.write(model_json)\n\nmodel.save_weights(pathToModel+'.h5')\n<\/code><\/pre>\n\n<p>Later, I load it again with the following code:<\/p>\n\n<pre><code>json_file = open(path+'.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\n\nloaded_model = model_from_json(loaded_model_json)\nloaded_model.load_weights(path+\".h5\")\n<\/code><\/pre>\n\n<p>I would like to retrieve the name of my model again from the loaded model by doing:<\/p>\n\n<pre><code>modelName = loaded_model.name\n<\/code><\/pre>\n\n<p>However, instead of the string <code>'NameOfMyModel'<\/code>, the string <code>'sequential_1'<\/code> is returned. What am I missing here?<\/p>\n","answers":[{"AnswerId":"52872839","CreationDate":"2018-10-18T11:20:53.083","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Apparently, by upgrading Keras (from version 2.2.0 to 2.2.4) the issue was resolved.<\/p>\n"}]}
+{"QuestionId":52854159,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><google-cloud-ml><object-detection-api>","CreationDate":"2018-10-17T11:47:20.237","AcceptedAnswerId":null,"OwnerUserId":4578900.0,"Title":"How to build uint8 numpy array input tensor for object detection prediction in ml engine","Body":"<p>I want to make online object detection predictions (or inference) from an already existing model in Google ML engine. But i am not able to build the json request.  <\/p>\n\n<p>The model is  faster_rcnn_inception_resnet_v2_atrous_coco_2017_11_08 from the TF model zoo. The inputs are images the outputs class, bb, score etc...<\/p>\n\n<p>The needed inputs are (from saved_model_cli show )<\/p>\n\n<pre><code>inputs['inputs'] tensor_info:\ndtype: DT_UINT8\nshape: (-1, -1, -1, 3)\nname: image_tensor:0\n<\/code><\/pre>\n\n<p>As it expects an uint8 array I load the image into a numpy array <\/p>\n\n<pre><code>encoded_contents = np.array(image.getdata()).reshape(\n        (im_height, im_width, 3)).astype(np.uint8)\n<\/code><\/pre>\n\n<p>Resize the image\n    image_np_expanded = np.expand_dims(encoded_contents, axis=0)<\/p>\n\n<p>Tried to build the json request  <\/p>\n\n<pre><code>instance = {\"input\":encoded_contents}\nrow = json.dumps(instance,sort_keys=True)\n<\/code><\/pre>\n\n<p>But I am not able to build it because <\/p>\n\n<pre><code>TypeError(repr(o) + \" is not JSON serializable\")\nTypeError: array([[[164, 191, 220],\n[190, 157, 114],\n[190, 157, 114]]], dtype=uint8) is not JSON serializable\n<\/code><\/pre>\n\n<p>If I convert the numpy array to a list with the tolist() method the json file takes 3 megas and the ML-engine refuses it \"message\": \"Request payload size exceeds the limit: 1572864 bytes.\",<\/p>\n\n<p>I will send this json to ml-engine predict as a json file.<\/p>\n\n<pre><code>gcloud ml-engine predict --model=pellaires --version=pellaires14 --json- \ninstances=request.json &gt; response.yaml\n<\/code><\/pre>\n","answers":[{"AnswerId":"52856769","CreationDate":"2018-10-17T14:01:10.600","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>Sending an array of ints that big is not generally efficient (you'll spend non-trivial time encoding and decoding those arrays, latency on the network, etc.). <a href=\"https:\/\/stackoverflow.com\/a\/46222990\/1399222\">This post<\/a> provides a few options (including the <code>tolist<\/code> you tried).<\/p>\n\n<p>I would recommend either \"Tensors Packed as Byte Strings\" or \"Compressed Image Data\".<\/p>\n\n<p><strong>UPDATE 10\/19\/2018<\/strong><\/p>\n\n<p>In order to use these approaches, you will need to modify your graph. If you're able to re-export the model, that is simplest. Instead of:<\/p>\n\n<pre><code>images = tf.placeholder(dtype=tf.uint8, shape=[None, None, None, 3])\n<\/code><\/pre>\n\n<p>You would use:<\/p>\n\n<pre><code>raw_byte_strings = tf.placeholder(dtype=tf.string, shape=[None])\ndecode = lambda raw_byte_str: tf.decode_raw(raw_byte_str, tf.uint8)\nimages = tf.map_fn(decode, raw_byte_strings, dtype=tf.uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52854179,"AnswerCount":1,"Tags":"<if-statement><tensorflow><keras><switch-statement><backend>","CreationDate":"2018-10-17T11:48:40.420","AcceptedAnswerId":52854474.0,"OwnerUserId":7242982.0,"Title":"keras (tensorflow backend) conditional assignment with K.switch()","Body":"<p>I'm trying to implement something like <\/p>\n\n<pre><code>if np.max(subgrid) == np.min(subgrid):\n    middle_middle = cur_subgrid + 1\nelse:\n    middle_middle = cur_subgrid\n<\/code><\/pre>\n\n<p>Since the condition can only be determined at run-time, I'm using keras syntax as following<\/p>\n\n<pre><code>middle_middle = K.switch(K.max(subgrid) == K.min(subgrid), lambda: tf.add(cur_subgrid,1), lambda: cur_subgrid)\n<\/code><\/pre>\n\n<p>But I'm getting this error:<\/p>\n\n<blockquote>\n  <p> in col_loop(j, gray_map, mask_A)\n       56 \n       57 \n  ---> 58             middle_middle = K.switch(K.max(subgrid) == K.min(subgrid), lambda: tf.add(cur_subgrid,1), lambda: cur_subgrid)\n       59 \n       60             print ('ml',middle_left.shape)<\/p>\n  \n  <p>\/nfs\/isicvlnas01\/share\/anaconda3\/lib\/python3.5\/site-packages\/keras\/backend\/tensorflow_backend.py\n  in switch(condition, then_expression, else_expression)    2561<br>\n  The selected tensor.    2562     \"\"\"\n  -> 2563     if condition.dtype != tf.bool:    2564         condition = tf.cast(condition, 'bool')    2565     if not\n  callable(then_expression):<\/p>\n  \n  <p>AttributeError: 'bool' object has no attribute 'dtype'<\/p>\n<\/blockquote>\n\n<p>middle_middle, cur_subgrid and subgrid are all NxN tensors. Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"52854474","CreationDate":"2018-10-17T12:04:18.883","ParentId":null,"OwnerUserId":"7456923","Title":null,"Body":"<p>I think the problem is that with <code>K.max(subgrid) == K.min(subgrid)<\/code> you're creating a <em>python boolean<\/em> comparing two tensor objects, not a <em>tensorflow boolean tensor<\/em> containing the value of the comparison of the <em>values<\/em> of the two input tensors. <\/p>\n\n<p>In other words, what you have written will be evaluated as<\/p>\n\n<pre><code>K.switch(True, lambda: tf.add(cur_subgrid,1), lambda: cur_subgrid)\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>comparison = ... # Some tensor, that at runtime will contain True if min and max are the same, False otherwise. \nK.switch(comparison , lambda: tf.add(cur_subgrid,1), lambda: cur_subgrid)\n<\/code><\/pre>\n\n<p>So what you need to do is to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/equal\" rel=\"nofollow noreferrer\">keras.backend.equal()<\/a> instead of <code>==<\/code>:<\/p>\n\n<pre><code>K.switch(K.equal(K.max(subgrid),K.min(subgrid)), lambda: tf.add(cur_subgrid,1), lambda: cur_subgrid)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52854289,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T11:53:20.937","AcceptedAnswerId":52855528.0,"OwnerUserId":3781647.0,"Title":"How to accumulate 1 to 10 in tensorflow?","Body":"<p>I want to accumulate 1 to 10 in tensorflow. But it does not work.\nCan anyone help me in fixing this.<\/p>\n\n<pre><code>one = tf.constant(1)\nvalue = tf.Variable(0,name=\"value\")\nincreasing_op = tf.add(value,one)\nassign_op = tf.assign(value,increasing_op)\n\n#sum\nsumvalue = tf.Variable(0,name = \"sumvalue\")\nsum_op = tf.add(sumvalue,assign_op)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as session:\n    session.run(init)\n    for _ in range(10):\n        session.run(sum_op)\n    print(session.run(sumvalue))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52855528","CreationDate":"2018-10-17T13:02:16.633","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>The <code>sum_op<\/code> returns the value after the computation. Also you could take advantage of <code>tf.assign_add<\/code>:<\/p>\n\n<pre><code>sumvalue = tf.Variable(0,name = \"sumvalue\")\nsum_op = tf.assign_add(sumvalue, 1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as session:\n    session.run(init)\n    for _ in range(10):\n        sum_value = session.run(sum_op)\n    print(sum_value)\n<\/code><\/pre>\n"},{"AnswerId":"52866090","CreationDate":"2018-10-18T02:27:53.127","ParentId":null,"OwnerUserId":"3781647","Title":null,"Body":"<p>Thanks for your help.\nI revised the code like below<\/p>\n\n<pre><code>n = int(input(\"Enter an integer: \"))\n\none=tf.constant(1)\n#increase\nincreasing_value=tf.Variable(0,name=\"increasing_value\")\nincreasing_op=tf.assign_add(increasing_value,one)\n\n#sum\nsumvalue=tf.Variable(0,name=\"sumvalue\")\nsum_op=tf.assign_add(sumvalue,increasing_value)\n\ninit=tf.global_variables_initializer()\n\nwith tf.Session() as session:\n    session.run(init)\n    for _ in range (n):\n        session.run(increasing_op)\n        session.run(sum_op)\n    print(session.run(sumvalue))\n<\/code><\/pre>\n\n<p>is it possible to use placeholder for n<\/p>\n"}]}
+{"QuestionId":52854618,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T12:11:38.690","AcceptedAnswerId":null,"OwnerUserId":4391025.0,"Title":"How should we deal with the name_scope's auto increment?","Body":"<pre><code>def foo():\n    with tf.Session() as sess:\n        with tf.name_scope(\"foo\") as abso:\n            print(abso)\n\nfor i in range(10):\n    foo()\n<\/code><\/pre>\n\n<p>This is a simple code to test tf's <code>name_scope<\/code>.\nThe output of this code piece is <\/p>\n\n<pre><code>foo\/\nfoo_1\/\nfoo_2\/\nfoo_3\/\nfoo_4\/\nfoo_5\/\nfoo_6\/\nfoo_7\/\nfoo_8\/\nfoo_9\/\n<\/code><\/pre>\n\n<p>By which way, the piece code's output are all <code>foo\/<\/code>s?<\/p>\n\n<p>This is a question when I want to make my tf model a webapp.\nThe app can deal with the first request correctly.\nBut when it comes to the second or later request, it tries to load the graph as well as all variables with another <code>name_scope<\/code>(for example, <code>foo_1<\/code>) and it causes a fatal error.<\/p>\n","answers":[{"AnswerId":"52857059","CreationDate":"2018-10-17T14:15:34.710","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>You could store the scope created by the <code>tf.name_scope<\/code> method and reuse it. For example:<\/p>\n\n<pre><code>scope = \"foo\"\ndef foo():\n    global scope\n    with tf.Session() as sess:\n        with tf.name_scope(scope) as scope:\n            print(scope)\nfor i in range(10):\n    foo()\n<\/code><\/pre>\n\n<p>will print<\/p>\n\n<pre><code>foo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\nfoo\/\n<\/code><\/pre>\n"},{"AnswerId":"52855286","CreationDate":"2018-10-17T12:50:13.547","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6007\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6007<\/a>. \nAs mentioned in the above issue, you can use absolute name_scope like tf.name_scope(\"foo\/\") to prevent auto-increment.<\/p>\n"},{"AnswerId":"52855106","CreationDate":"2018-10-17T12:40:56.967","ParentId":null,"OwnerUserId":"4391025","Title":null,"Body":"<p>I try to add tf.reset_default_graph() before we establish the session and it works<\/p>\n"}]}
+{"QuestionId":52854712,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><multilabel-classification>","CreationDate":"2018-10-17T12:17:40.077","AcceptedAnswerId":null,"OwnerUserId":988709.0,"Title":"Can I minimize tf.nn.sigmoid_cross_entropy_with_logits without reduce_mean?","Body":"<p>I'm training my multi label model with tensorflow. The loss is calculated with <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code>. Can I just minimize the loss without <code>reduce_sum<\/code> or <code>reduce_mean<\/code> like this:<\/p>\n\n<pre><code>...\n#loss = tf.reduce_mean(tf.losses.sigmoid_cross_entropy(multi_class_labels=labels, logits=logits)) ## not do this\nloss = tf.losses.sigmoid_cross_entropy(multi_class_labels=labels, logits=logits)\noptimizer = tf.train.AdamOptimizer(learning_rate=0.1) # do this\ntrain_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"52855556","CreationDate":"2018-10-17T13:03:46.673","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>You can use minimize() with a non-scalar Tensor. But minimize() calculates reduce_sum for a  non-scalar Tensor internally, so I recommend that you use reduce_sum for readability.<\/p>\n"}]}
+{"QuestionId":52854983,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T12:33:52.407","AcceptedAnswerId":52855692.0,"OwnerUserId":4792229.0,"Title":"Why are different libraries searched for in tensorflow, even though both were installed the same way?","Body":"<p>I built tensorflow from source and got a *.whl file that I could install on my pc with pip install *.whl. Now in the virtualenv where I installed it I can open python and do import tensorflow without a problem and also use tf. Now I tried to install this same wheel on an other pc in a virtualenv and it worked successfully, but when I try to use import tensorflow in python I get: <\/p>\n\n<p><code>ImportError: libnvidia-fatbinaryloader.so.390.48: cannot open shared object file: No such file or directory<\/code><\/p>\n\n<p>Now I actually do not have that file on the other pc, but after checking my own pc I also don't have it here. I have on both pcs libnvidia-fatbinaryloader.so.390.87. On both pcs the LD_LIBRARY_PATH points to the directory with that version. <\/p>\n\n<p>How can it be that tensorflow searches for version 48 on the remote pc while searching for 87 and finding it on my pc, even though they are both installed with the same whl file? Is there a config that I need to adjust what version it should search for?<\/p>\n","answers":[{"AnswerId":"52855249","CreationDate":"2018-10-17T12:48:13.057","ParentId":null,"OwnerUserId":"8477210","Title":null,"Body":"<p>The building process is related to the computer environment.Could building tensorflow in the same machine and installing it on the same machine help?Building on one machine and generating the *.whl,but installing on other machines may cause problem.<\/p>\n"},{"AnswerId":"52855692","CreationDate":"2018-10-17T13:09:51.010","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I would say you have a broken CUDA installation somewhere in the library path. It is <code>libcuda.so<\/code> that has a dependency on <code>libnvidia-fatbinaryloader.so<\/code>, so maybe the symbolic links point to a library that no longer exists but was installed before.<\/p>\n\n<p>You can find this information by running the <code>ldd<\/code> command on the <code>libcuda.so<\/code> file.<\/p>\n"}]}
+{"QuestionId":52855843,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-10-17T13:16:39.647","AcceptedAnswerId":52859411.0,"OwnerUserId":10511793.0,"Title":"Multi label classification in pytorch","Body":"<p>I have a multi-label classification problem. I have 11 classes, around 4k examples. Each example can have from 1 to 4-5 label. At the moment, i'm training a classifier separately for each class with log_loss. As you can expect, it is taking quite some time to train 11 classifier, and i would like to try another approach and to train only 1 classifier. The idea is that the last layer of this classifer would have 11 nodes, and would output a real number by classes which would be converted to a proba by a sigmoid. The loss I want to optimize is the mean of the log_loss on all classes. <\/p>\n\n<p>Unfortunately, i'm some kind of noob with pytorch, and even by reading the source code of the losses, i can't figure out if one of the already existing losses does exactly what i want, or if I should create a new loss, and if that's the case, i don't really know how to do it.<\/p>\n\n<p>To be very specific, i want to give for each element of the batch one vector of size 11(which contains a real number for each label (the closer to infinity, the closer this class is predicted to be 1), and 1 vector of size 11 (which contains a 1 at every true label), and be able to compute the mean log_loss on all 11 labels, and optimize my classifier based on that loss.<\/p>\n\n<p>Any help would be greatly appreciated :) <\/p>\n","answers":[{"AnswerId":"52859411","CreationDate":"2018-10-17T16:18:47.343","ParentId":null,"OwnerUserId":"6212301","Title":null,"Body":"<p>You are looking for <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#bceloss\" rel=\"noreferrer\"><code>torch.nn.BCELoss<\/code><\/a>. Here's example code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\nbatch_size = 2\nnum_classes = 11\n\nloss_fn = torch.nn.BCELoss()\n\noutputs_before_sigmoid = torch.randn(batch_size, num_classes)\nsigmoid_outputs = torch.sigmoid(outputs_before_sigmoid)\ntarget_classes = torch.randint(0, 2, (batch_size, num_classes))  # randints in [0, 2).\n\nloss = loss_fn(sigmoid_outputs, target_classes)\n\n# alternatively, use BCE with logits, on outputs before sigmoid.\nloss_fn_2 = torch.nn.BCEWithLogitsLoss()\nloss2 = loss_fn_2(outputs_before_sigmoid, target_classes)\nassert loss == loss2\n<\/code><\/pre>\n"}]}
+{"QuestionId":52855870,"AnswerCount":0,"Tags":"<python><tensorflow><reinforcement-learning><tensorflow-datasets>","CreationDate":"2018-10-17T13:17:46.473","AcceptedAnswerId":null,"OwnerUserId":9486786.0,"Title":"QueueRunner going towards deprecation, but tf.data does not replace all usecases?","Body":"<p>In my code I get QueueRunner deprecation warning:<\/p>\n\n<p><code>QueueRunner.__init__ (from tensorflow.python.training.queue_runner_impl) is deprecated and will be removed in a future version.\nInstructions for updating:\nTo construct input pipelines, use the `tf.data` module.\n<\/code><\/p>\n\n<p>Did a bit of investigation <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">here<\/a> and it seems that tf.data is a new thing that is faster and better than the old queue stuff. But, as I understand, it it not replacing queues because data is expected to be static, (as far as I understood and of all the examples I saw), modification \/ appending new training data to the dataset is not supported, so I don't know how this change applies to the reinforcement learning, which learns on new data environment returns all the time. So, as far as I see, deprecation of queue runner is coming and I don't see tf.data replacing it for this usecase?<\/p>\n\n<p>Any kind of explanation and example is welcome! Thanks!<\/p>\n","answers":[]}
+{"QuestionId":52856155,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-17T13:30:50.773","AcceptedAnswerId":55174412.0,"OwnerUserId":1727392.0,"Title":"Keras: Adding MDN Layer to LSTM Network","Body":"<p>My question in brief: Is the Long Short Term Memory Network detailed below appropriately designed to generate new dance sequences, given dance sequence training data?<\/p>\n\n<p>Context: I am working with a dancer who wishes to use a neural network to generate new dance sequences. She sent me the <a href=\"https:\/\/arxiv.org\/abs\/1605.06921\" rel=\"nofollow noreferrer\">2016 chor-rnn paper<\/a> that accomplished this task using an LSTM network with a Mixture Density Network layer at the end. After adding an MDN layer to my LSTM network, however, my loss goes negative and the results seem chaotic. This may be due to the very small training data, but I'd like to validate the model fundamentals before scaling up training data size. If anyone can advise whether the model below is overlooking something fundamental (which is highly likely), I would be entirely grateful for their feedback.<\/p>\n\n<p>The sample data I'm feeding into the network (<code>X<\/code> below) has shape (626, 55, 3), which corresponds to 626 time snapshots of 55 body positions, each with 3 coordinates (x, y, then z). So X<a href=\"https:\/\/i.stack.imgur.com\/UOV3K.gif\" rel=\"nofollow noreferrer\">1<\/a>[11][2] is the z position of the 11th body part at time 1:<\/p>\n\n<pre><code>import requests\nimport numpy as np\n\n# download the data\nrequests.get('https:\/\/s3.amazonaws.com\/duhaime\/blog\/dancing-with-robots\/dance.npy')\n\n# X.shape = time_intervals, n_body_parts, 3\nX = np.load('dance.npy')\n<\/code><\/pre>\n\n<p>To make sure the data was extracted correctly, I visualize the first few frames of <code>X<\/code>:<\/p>\n\n<pre><code>import mpl_toolkits.mplot3d.axes3d as p3\nimport matplotlib.pyplot as plt\nfrom IPython.display import HTML\nfrom matplotlib import animation\nimport matplotlib\n\nmatplotlib.rcParams['animation.embed_limit'] = 2**128\n\ndef update_points(time, points, X):\n  arr = np.array([[ X[time][i][0], X[time][i][1] ] for i in range(int(X.shape[1]))])\n  points.set_offsets(arr) # set x, y values\n  points.set_3d_properties(X[time][:,2][:], zdir='z') # set z value\n\ndef get_plot(X, lim=2, frames=200, duration=45):\n  fig = plt.figure()\n  ax = p3.Axes3D(fig)\n  ax.set_xlim(-lim, lim)\n  ax.set_ylim(-lim, lim)\n  ax.set_zlim(-lim, lim)\n  points = ax.scatter(X[0][:,0][:], X[0][:,1][:], X[0][:,2][:], depthshade=False) # x,y,z vals\n  return animation.FuncAnimation(fig,\n    update_points,\n    frames,\n    interval=duration,\n    fargs=(points, X),\n    blit=False  \n  ).to_jshtml()\n\nHTML(get_plot(X, frames=int(X.shape[0])))\n<\/code><\/pre>\n\n<p>That produces a little dancing sequence like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UOV3K.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UOV3K.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So far so good. Next I center the features of the x, y, and z dimensions:<\/p>\n\n<pre><code>X -= np.amin(X, axis=(0, 1))\nX \/= np.amax(X, axis=(0, 1))\n<\/code><\/pre>\n\n<p>Visualizing the resulting <code>X<\/code> with <code>HTML(get_plot(X, frames=int(X.shape[0])))<\/code> shows these lines center the data just fine. Next I build the model itself using the Sequential API in Keras:<\/p>\n\n<pre><code>from keras.models import Sequential, Model\nfrom keras.layers import Dense, LSTM, Dropout, Activation\nfrom keras.layers.advanced_activations import LeakyReLU\nfrom keras.losses import mean_squared_error\nfrom keras.optimizers import Adam\nimport keras, os\n\n# config\nlook_back = 32 # number of previous time frames to use to predict the positions at time i\nlstm_cells = 256 # number of cells in each LSTM \"layer\"\nn_features = int(X.shape[1]) * int(X.shape[2]) # number of coordinate values to be predicted by each of `m` models\ninput_shape = (look_back, n_features) # shape of inputs\nm = 32 # number of gaussian models to build\n\n# set boolean controlling whether we use MDN or not\nuse_mdn = True\n\nmodel = Sequential()\nmodel.add(LSTM(lstm_cells, return_sequences=True, input_shape=input_shape))\nmodel.add(LSTM(lstm_cells, return_sequences=True))\nmodel.add(LSTM(lstm_cells))\n\nif use_mdn:\n  model.add(MDN(n_features, m))\n  model.compile(loss=get_mixture_loss_func(n_features, m), optimizer=Adam(lr=0.000001))\nelse:\n  model.add(Dense(n_features, activation='tanh'))\n  model.compile(loss=mean_squared_error, optimizer='sgd')\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>Once the model is built, I arrange the data in <code>X<\/code> to prepare for training. Here we want to predict the x, y, z positions of of the 55 body parts at some time by examining the positions of each body part at the previous <code>look_back<\/code> time slices:<\/p>\n\n<pre><code># get training data in right shape\ntrain_x = []\ntrain_y = []\n\nn_time, n_obs, n_attrs = [int(i) for i in X.shape]\n\nfor i in range(look_back, n_time-1, 1):\n  train_x.append( X[i-look_back:i].reshape(look_back, n_obs * n_attrs) )\n  train_y.append( X[i+1].ravel() )\n\ntrain_x = np.array(train_x)\ntrain_y = np.array(train_y)\n<\/code><\/pre>\n\n<p>And finally I train the model:<\/p>\n\n<pre><code>from livelossplot import PlotLossesKeras\n\n# fit the model\nmodel.fit(train_x, train_y, epochs=1024, batch_size=1, callbacks=[PlotLossesKeras()])\n<\/code><\/pre>\n\n<p>After training, I visualize the new time slices created by the model:<\/p>\n\n<pre><code># generate `n_frames` of new output time slices\nn_frames = 3000\n\n# seed the data to plot with the first `look_back` animation frames\ndata = X[0:look_back]\n\nx0, x1, x2 = [int(i) for i in train_x.shape]\nd0, d1, d2 = [int(i) for i in data.shape]\n\nfor i in range(look_back, n_frames, 1):\n  # get the model's prediction for the next position of points at time `i`\n  result = model.predict(train_x[i].reshape(1, x1, x2))\n  # if using the mixed density network, pull out vals that describe vertex positions\n  if use_mdn:\n    result = np.apply_along_axis(sample_from_output, 1, result, n_features, m, temp=1.0)\n  # reshape the result into the form of rows in `X`\n  result = result.reshape(1, d1, d2)\n  # push the result into the shape of `train_x` observations\n  stacked = np.vstack((data[i-look_back+1:i], result)).reshape(1, x1, x2)\n  # add the result to the `train_x` observations\n  train_x = np.vstack((train_x, stacked))\n  # add the result to the dataset for plotting\n  data = np.vstack((data[:i], result))\n<\/code><\/pre>\n\n<p>If I set <code>use_mdn<\/code> to <code>False<\/code> above and instead use a simple Sum of Squared Errors Loss (L2 Loss), then the resulting visualization seems a little creepy but still has a generally human shape.<\/p>\n\n<p>If I set <code>use_mdn<\/code> to <code>True<\/code>, however, and use the custom MDN loss function, the results are quite odd. I recognize that the MDN layer adds a huge number of parameters that need to be trained, and likely requires orders of magnitude more training data to achieve output that's as human-shaped as the L2 loss function output.<\/p>\n\n<p>That said, I wanted to ask if others who have worked with neural network models more extensively than myself see anything fundamentally wrong with the approach above. Any insights on this question would be tremendously helpful.<\/p>\n","answers":[{"AnswerId":"55174412","CreationDate":"2019-03-15T01:42:58.283","ParentId":null,"OwnerUserId":"1727392","Title":null,"Body":"<p>Good God I got it going [<a href=\"https:\/\/gist.github.com\/duhaime\/0219ac0f638068b3d7ea04144c6918d9#file-fresher-ipynb\" rel=\"nofollow noreferrer\">gist<\/a>]! Here's the MDN class:<\/p>\n\n<pre><code>from keras.layers.advanced_activations import LeakyReLU\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dense, Input, merge, concatenate, Dense, LSTM, CuDNNLSTM\nfrom keras.engine.topology import Layer\nfrom keras import backend as K\nimport tensorflow_probability as tfp\nimport tensorflow as tf\n\n# check tfp version, as tfp causes cryptic error if out of date\nassert float(tfp.__version__.split('.')[1]) &gt;= 5\n\nclass MDN(Layer):\n  '''Mixture Density Network with unigaussian kernel'''\n  def __init__(self, n_mixes, output_dim, **kwargs):\n    self.n_mixes = n_mixes\n    self.output_dim = output_dim\n\n    with tf.name_scope('MDN'):\n      self.mdn_mus    = Dense(self.n_mixes * self.output_dim, name='mdn_mus')\n      self.mdn_sigmas = Dense(self.n_mixes, activation=K.exp, name='mdn_sigmas')\n      self.mdn_alphas = Dense(self.n_mixes, activation=K.softmax, name='mdn_alphas')\n    super(MDN, self).__init__(**kwargs)\n\n  def build(self, input_shape):\n    self.mdn_mus.build(input_shape)\n    self.mdn_sigmas.build(input_shape)\n    self.mdn_alphas.build(input_shape)\n    self.trainable_weights = self.mdn_mus.trainable_weights + \\\n      self.mdn_sigmas.trainable_weights + \\\n      self.mdn_alphas.trainable_weights\n    self.non_trainable_weights = self.mdn_mus.non_trainable_weights + \\\n      self.mdn_sigmas.non_trainable_weights + \\\n      self.mdn_alphas.non_trainable_weights\n    self.built = True\n\n  def call(self, x, mask=None):\n    with tf.name_scope('MDN'):\n      mdn_out = concatenate([\n        self.mdn_mus(x),\n        self.mdn_sigmas(x),\n        self.mdn_alphas(x)\n      ], name='mdn_outputs')\n    return mdn_out\n\n  def get_output_shape_for(self, input_shape):\n    return (input_shape[0], self.output_dim)\n\n  def get_config(self):\n    config = {\n      'output_dim': self.output_dim,\n      'n_mixes': self.n_mixes,\n    }\n    base_config = super(MDN, self).get_config()\n    return dict(list(base_config.items()) + list(config.items()))\n\n  def get_loss_func(self):\n    def unigaussian_loss(y_true, y_pred):\n      mix = tf.range(start = 0, limit = self.n_mixes)\n      out_mu, out_sigma, out_alphas = tf.split(y_pred, num_or_size_splits=[\n        self.n_mixes * self.output_dim,\n        self.n_mixes,\n        self.n_mixes,\n      ], axis=-1, name='mdn_coef_split')\n\n      def loss_i(i):\n        batch_size = tf.shape(out_sigma)[0]\n        sigma_i = tf.slice(out_sigma, [0, i], [batch_size, 1], name='mdn_sigma_slice')\n        alpha_i = tf.slice(out_alphas, [0, i], [batch_size, 1], name='mdn_alpha_slice')\n        mu_i = tf.slice(out_mu, [0, i * self.output_dim], [batch_size, self.output_dim], name='mdn_mu_slice')\n        dist = tfp.distributions.Normal(loc=mu_i, scale=sigma_i)\n        loss = dist.prob(y_true) # find the pdf around each value in y_true\n        loss = alpha_i * loss\n        return loss\n\n      result = tf.map_fn(lambda  m: loss_i(m), mix, dtype=tf.float32, name='mix_map_fn')\n      result = tf.reduce_sum(result, axis=0, keepdims=False)\n      result = -tf.log(result)\n      result = tf.reduce_mean(result)\n      return result\n\n    with tf.name_scope('MDNLayer'):\n      return unigaussian_loss\n<\/code><\/pre>\n\n<p>And the LSTM class:<\/p>\n\n<pre><code>class LSTM_MDN:\n  def __init__(self, n_verts=15, n_dims=3, n_mixes=2, look_back=1, cells=[32,32,32,32], use_mdn=True):\n    self.n_verts = n_verts\n    self.n_dims = n_dims\n    self.n_mixes = n_mixes\n    self.look_back = look_back\n    self.cells = cells\n    self.use_mdn = use_mdn\n    self.LSTM = CuDNNLSTM if len(gpus) &gt; 0 else LSTM\n    self.model = self.build_model()\n    if use_mdn:\n      self.model.compile(loss=MDN(n_mixes, n_verts*n_dims).get_loss_func(), optimizer='adam', metrics=['accuracy'])\n    else:\n      self.model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])\n\n  def build_model(self):\n    i = Input((self.look_back, self.n_verts*self.n_dims))\n    h = self.LSTM(self.cells[0], return_sequences=True)(i) # return sequences, stateful\n    h = self.LSTM(self.cells[1], return_sequences=True)(h)\n    h = self.LSTM(self.cells[2])(h)\n    h = Dense(self.cells[3])(h)\n    if self.use_mdn:\n      o = MDN(self.n_mixes, self.n_verts*self.n_dims)(h)\n    else:\n      o = Dense(self.n_verts*self.n_dims)(h)\n    return Model(inputs=[i], outputs=[o])\n\n  def prepare_inputs(self, X, look_back=2):\n    '''\n    Prepare inputs in shape expected by LSTM\n    @returns:\n      numpy.ndarray train_X: has shape: n_samples, lookback, verts * dims\n      numpy.ndarray train_Y: has shape: n_samples, verts * dims\n    '''\n    # prepare data for the LSTM_MDN\n    X = X.swapaxes(0, 1) # reshape to time, vert, dim\n    n_time, n_verts, n_dims = X.shape\n\n    # validate shape attributes\n    if n_verts != self.n_verts: raise Exception(' ! got', n_verts, 'vertices, expected', self.n_verts)\n    if n_dims != self.n_dims: raise Exception(' ! got', n_dims, 'dims, expected', self.n_dims)\n    if look_back != self.look_back: raise Exception(' ! got', look_back, 'for look_back, expected', self.look_back)\n\n    # lstm expects data in shape [samples_in_batch, timestamps, values]\n    train_X = []\n    train_Y = []\n    for i in range(look_back, n_time, 1):\n      train_X.append( X[i-look_back:i,:,:].reshape(look_back, n_verts * n_dims) ) # look_back, verts * dims\n      train_Y.append( X[i,:,:].reshape(n_verts * n_dims) ) # verts * dims\n    train_X = np.array(train_X) # n_samples, lookback, verts * dims\n    train_Y = np.array(train_Y) # n_samples, verts * dims\n    return [train_X, train_Y]\n\n  def predict_positions(self, input_X):\n    '''\n    Predict the output for a series of input frames. Each prediction has shape (1, y), where y contains:\n      mus = y[:n_mixes*n_verts*n_dims]\n      sigs = y[n_mixes*n_verts*n_dims:-n_mixes]\n      alphas = softmax(y[-n_mixes:])\n    @param numpy.ndarray input_X: has shape: n_samples, look_back, n_verts * n_dims\n    @returns:\n      numpy.ndarray X: has shape: verts, time, dims\n    '''\n    predictions = []\n    for i in range(input_X.shape[0]):\n      y = self.model.predict( train_X[i:i+1] ).squeeze()\n      mus = y[:n_mixes*n_verts*n_dims]\n      sigs = y[n_mixes*n_verts*n_dims:-n_mixes]\n      alphas = self.softmax(y[-n_mixes:])\n\n      # find the most likely distribution then pull out the mus that correspond to that selected index\n      alpha_idx = np.argmax(alphas) # 0\n      alpha_idx = 0\n      predictions.append( mus[alpha_idx*self.n_verts*self.n_dims:(alpha_idx+1)*self.n_verts*self.n_dims] )\n    predictions = np.array(predictions).reshape(train_X.shape[0], self.n_verts, self.n_dims).swapaxes(0, 1)\n    return predictions # shape = n_verts, n_time, n_dims\n\n  def softmax(self, x):\n    ''''Compute softmax values for vector `x`'''\n    r = np.exp(x - np.max(x))\n    return r \/ r.sum()\n<\/code><\/pre>\n\n<p>Then setting up the class:<\/p>\n\n<pre><code>X = data.selected.X\nn_verts, n_time, n_dims = X.shape\nn_mixes = 3\nlook_back = 2\n\nlstm_mdn = LSTM_MDN(n_verts=n_verts, n_dims=n_dims, n_mixes=n_mixes, look_back=look_back)\ntrain_X, train_Y = lstm_mdn.prepare_inputs(X, look_back=look_back)\n<\/code><\/pre>\n\n<p>The gist linked above has the full gory details in case anyone wants to reproduce this and take it apart to better understand the mechanics... <\/p>\n"}]}
+{"QuestionId":52856178,"AnswerCount":1,"Tags":"<tensorflow><keras><jupyter-notebook><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-17T13:31:58.433","AcceptedAnswerId":53327757.0,"OwnerUserId":8463061.0,"Title":"Error processing Timeseries tensorflow notebook on TPU","Body":"<p>Here's a Timeseries notebook I used from the good work by <a href=\"http:\/\/www.hvass-labs.org\/\" rel=\"nofollow noreferrer\">Magnus Erik Hvass Pedersen<\/a> - thanks for that:<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1F6CuGVWN5TNgIjqxdu5glFeGBEr71TgO\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1F6CuGVWN5TNgIjqxdu5glFeGBEr71TgO<\/a><\/p>\n\n<p>I have had success running a version of this notebook via Google Colab on a GPU but when I do the same (after some modifications to make the code compatible on TPUs) I get this error:<\/p>\n\n<p><code>ValueError: Error when checking input: expected input to have shape (299776, 20) but got array with shape (33309, 20)<\/code><\/p>\n\n<p>The full stack trace can be found on the cell location <a href=\"https:\/\/colab.research.google.com\/drive\/1F6CuGVWN5TNgIjqxdu5glFeGBEr71TgO#scrollTo=wdSmXdvDw5HL\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1F6CuGVWN5TNgIjqxdu5glFeGBEr71TgO#scrollTo=wdSmXdvDw5HL<\/a>.<\/p>\n\n<p>It has been a bit of a tug-of-war to get the input\/output data shapes in order but as we kept solving I\/O shape issues other shape related issues started proping up.<\/p>\n\n<p>The notebook is available for sharing and commenting.<\/p>\n\n<p>Any thoughts will be appreciated.<\/p>\n","answers":[{"AnswerId":"53327757","CreationDate":"2018-11-15T20:54:39.920","ParentId":null,"OwnerUserId":"3870291","Title":null,"Body":"<p>To solve the error <code>ValueError: Operation 'tpu_140099307695464\/VarIsInitializedOp'<\/code>.\nTry using <code>tf.train.RMSPropOptimizer<\/code> instead of using RMSProp from <code>tensorflow.keras.optimizers<\/code>.<\/p>\n"}]}
+{"QuestionId":52856423,"AnswerCount":0,"Tags":"<neural-network><deep-learning><conv-neural-network><caffe><pycaffe>","CreationDate":"2018-10-17T13:44:46.763","AcceptedAnswerId":null,"OwnerUserId":9284728.0,"Title":"Caffe CNN Slice layer: 2nd Slice layer produces unknown bottom blob","Body":"<p>Caffe CNN Slice layer: 2nd Slice layer produces unknown bottom blob<\/p>\n\n<p>I have 2 Slice layers (see proto file). It seems the 1st one is working well; whereas the 2nd one's bottom gives \"unknown bottom blob error\" as following:<\/p>\n\n<p>In fact I am not sure the error is related to Slice or Flatten!? \n<strong>please note that the the 2nd Slice does not even printed in the log!!!<\/strong><\/p>\n\n<p>this is the Proto file:<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label_b4_noise\"\n  include {\n    phase: TEST\n  }\n  hdf5_data_param {\n    source: \"data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_val_list.txt\"\n    batch_size: 25\n    shuffle: true\n  }\n}\nlayer {\n  name: \"data\"\n  type: \"HDF5Data\"\n  top: \"data\"\n  top: \"label_b4_noise\"\n  include {\n    phase: TRAIN\n  }\n  hdf5_data_param {\n    source: \"data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_train_list.txt\"\n    batch_size: 25\n    shuffle: true\n  }\n}\nlayer {\n  name: \"slic0\"\n  type: \"Slice\"\n  bottom: \"data\"\n  top: \"data1\"\n  top: \"data2\"\n  slice_param {\n    axis: 1\n    slice_point: 1\n  }\n}\nlayer {\n  name: \"conv_u0d-score_New\"\n  type: \"Convolution\"\n  bottom: \"data1\"\n  top: \"conv_last\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  convolution_param {\n    num_output: 1\n    pad: 0\n    kernel_size: 1\n    weight_filler {\n      type: \"msra\"\n    }\n   }\n}\nlayer {\n  name: \"flat\"\n  type: \"Flatten\"\n  bottom: \"conv_last\"\n  top: \"ampl\"\n}\nlayer {\n  name: \"slic1\"\n  type: \"Slice\"\n  bottom: \"label_b4_noise\"\n  top: \"label_b4_noise1\"\n  top: \"label_b4_noise2\"\n  slice_param {\n    axis: 1\n    slice_point: 1\n  }\n}\nlayer {\n  name: \"flatdata\"\n  type: \"Flatten\"\n  bottom: \"label_b4_noise1\"\n  top: \"flatdata\"\n}\n\nlayer {\n  name: \"loss\"\n  type: \"EuclideanLoss\"\n  bottom: \"ampl\"\n  bottom: \"flatdata\"\n  top: \"loss\"\n  softmax_param {engine: CAFFE}\n}\n<\/code><\/pre>\n\n<p>this is the log file:<\/p>\n\n<pre><code>GL  ----------------------------------------------------------------\nres\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s .\nres\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s .\nWARNING: Logging before InitGoogleLogging() is written to STDERR\nI1018 11:40:22.930601 104201 upgrade_proto.cpp:67] Attempting to upgrade input file specified using deprecated input fields: res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_tmp\/unet_deploy.txt\nI1018 11:40:22.930654 104201 upgrade_proto.cpp:70] Successfully upgraded file specified using deprecated input fields.\nW1018 11:40:22.930658 104201 upgrade_proto.cpp:72] Note that future Caffe releases will only support input layers and not input fields.\nI1018 11:40:23.237383 104201 net.cpp:51] Initializing net from parameters:\nname: \"unet\"\nstate {\n  phase: TEST\n  level: 0\n}\nlayer {\n  name: \"input\"\n  type: \"Input\"\n  top: \"data\"\n  input_param {\n    shape {\n      dim: 1\n      dim: 2\n      dim: 1\n      dim: 2048\n    }\n  }\n}\nlayer {\n  name: \"slic0\"\n  type: \"Slice\"\n  bottom: \"data\"\n  top: \"data1\"\n  top: \"data2\"\n  slice_param {\n    slice_point: 1\n    axis: 1\n  }\n}\nlayer {\n  name: \"conv_u0d-score_New\"\n  type: \"Convolution\"\n  bottom: \"data1\"\n  top: \"conv_last\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  convolution_param {\n    num_output: 1\n    pad: 0\n    kernel_size: 1\n    weight_filler {\n      type: \"msra\"\n    }\n  }\n}\nlayer {\n  name: \"flat\"\n  type: \"Flatten\"\n  bottom: \"conv_last\"\n  top: \"ampl\"\n}\nlayer {\n  name: \"flatdata\"\n  type: \"Flatten\"\n  bottom: \"label_b4_noise1\"\n  top: \"flatdata\"\n}\nF1018 11:40:23.237546 104201 insert_splits.cpp:29] Unknown bottom blob 'label_b4_noise1' (layer 'flatdata', bottom index 0)\n*** Check failure stack trace: ***\n\/pbs\/home\/n\/nhatami\/sps\/spectro\/trainAndTest_4removal: line 101: 104201 Aborted                 $pydir\/dumpLayersSize.py ${tmp_root}_deploy.txt ${oroot}\n\n\/pbs\/home\/n\/nhatami\/sps\/spectro\/trainAndTest_4removal: line 101: 104201 Aborted                 $pydir\/dumpLayersSize.py ${tmp_root}_deploy.txt ${oroot}\nThu Oct 18 11:40:23 CEST 2018\n\/usr\/bin\/time -v caffe -gpu 0 --log_dir=res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_tmp train -solver res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_tmp\/unet_solver.txt\nThu Oct 18 11:41:26 CEST 2018\n\/pbs\/home\/n\/nhatami\/sps\/spectro\/trainAndTest_4removal: line 206: gnuplot: command not found\n\/pbs\/home\/n\/nhatami\/sps\/spectro\/trainAndTest_4removal: line 225: gnuplot: command not found\n\/usr\/bin\/time -v -o res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_time_test.txt python \/pbs\/home\/n\/nhatami\/sps\/spectro\/python\/test_4removal.py -eg hist -l label -mf=data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_met.txt res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_deploy.txt res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet.caffemodel data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_test.h5 res\/4removal_nAmp3nData2_2e5\/unet_bs10\/restest\/unet_test\nWARNING: Logging before InitGoogleLogging() is written to STDERR\nW1018 11:41:42.595289 105177 _caffe.cpp:139] DEPRECATION WARNING - deprecated use of Python interface\nW1018 11:41:42.595335 105177 _caffe.cpp:140] Use this instead (with the named \"weights\" parameter):\nW1018 11:41:42.595338 105177 _caffe.cpp:142] Net('res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_deploy.txt', 1, weights='res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet.caffemodel')\nI1018 11:41:42.597472 105177 upgrade_proto.cpp:67] Attempting to upgrade input file specified using deprecated input fields: res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_deploy.txt\nI1018 11:41:42.597497 105177 upgrade_proto.cpp:70] Successfully upgraded file specified using deprecated input fields.\nW1018 11:41:42.597501 105177 upgrade_proto.cpp:72] Note that future Caffe releases will only support input layers and not input fields.\nI1018 11:41:42.597535 105177 net.cpp:51] Initializing net from parameters:\nname: \"unet\"\nstate {\n  phase: TEST\n  level: 0\n}\nlayer {\n  name: \"input\"\n  type: \"Input\"\n  top: \"data\"\n  input_param {\n    shape {\n      dim: 1\n      dim: 2\n      dim: 1\n      dim: 2048\n    }\n  }\n}\nlayer {\n  name: \"slic0\"\n  type: \"Slice\"\n  bottom: \"data\"\n  top: \"data1\"\n  top: \"data2\"\n  slice_param {\n    slice_point: 1\n    axis: 1\n  }\n}\nlayer {\n  name: \"conv_u0d-score_New\"\n  type: \"Convolution\"\n  bottom: \"data1\"\n  top: \"conv_last\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  convolution_param {\n    num_output: 1\n    pad: 0\n    kernel_size: 1\n    weight_filler {\n      type: \"msra\"\n    }\n  }\n}\nlayer {\n  name: \"flat\"\n  type: \"Flatten\"\n  bottom: \"conv_last\"\n  top: \"ampl\"\n}\nlayer {\n  name: \"flatdata\"\n  type: \"Flatten\"\n  bottom: \"label_b4_noise1\"\n  top: \"flatdata\"\n}\nF1018 11:41:42.597617 105177 insert_splits.cpp:29] Unknown bottom blob 'label_b4_noise1' (layer 'flatdata', bottom index 0)\n*** Check failure stack trace: ***\n('res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet_deploy.txt', 'res\/4removal_nAmp3nData2_2e5\/unet_bs10\/unet.caffemodel', 'data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_test.h5', 'data\/4removal_nAmp3nData2_2e5\/2048_2e5_0.01_s_met.txt')\n<\/code><\/pre>\n\n<p>here is h5disp:<\/p>\n\n<pre><code>Dataset 'label_b4_noise' \n        Size:  2048x1x2x10000\n        MaxSize:  2048x1x2xInf\n        Datatype:   H5T_IEEE_F32LE (single)\n        ChunkSize:  2048x1x2x100\n        Filters:  none\n        FillValue:  0.000000\n\nDataset 'data' \n        Size:  2048x1x2x10000\n        MaxSize:  2048x1x2xInf\n        Datatype:   H5T_IEEE_F32LE (single)\n        ChunkSize:  2048x1x2x100\n        Filters:  none\n        FillValue:  0.000000\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52856441,"AnswerCount":4,"Tags":"<python><python-3.x><pycharm><python-3.6><pytorch>","CreationDate":"2018-10-17T13:45:47.123","AcceptedAnswerId":52906407.0,"OwnerUserId":10424445.0,"Title":"pytorch module not found in pycharm (windows)","Body":"<p>Hello I wanted to install Pytorch via anaconda and it worked but PyCharm cant find the module (<\/p>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'torch'<\/p>\n<\/blockquote>\n\n<p>I also have CUDA installed but when I looked up to add a the package with pycharm it also gives an error. When I added the anaconda interpreter I cant run the code. I use Python 3.6<\/p>\n","answers":[{"AnswerId":"52906407","CreationDate":"2018-10-20T13:53:05.613","ParentId":null,"OwnerUserId":"10424445","Title":null,"Body":"<p>Ok I solved this problem: First install anaconda and open the prompt then type <code>conda install pytorch -c pytorch<\/code> and <code>pip3 install torchvision<\/code>. Then go to PyCharm and create an Project and set the Project Interpreter to the Anaconda one (there is in the path: \\Anaconda.x.x\\python.exe ). Then you go to the Run settings and click <code>Run...<\/code> and then you go to <code>Edit Configurations<\/code> and then you select the <code>Project Default<\/code> interpreter and apply and you should be done! Thanks to the ppl who helped me =)<\/p>\n"},{"AnswerId":"52857497","CreationDate":"2018-10-17T14:37:44.820","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>It sounds like the problem is PyCharm not anaconda. Can you check what shows up in PyCharm's interpreter and select the most appropriate one for your project? <\/p>\n\n<pre><code>File -&gt; Settings -&gt; Project (Your Project Name) -&gt; Project Interpreter\n<\/code><\/pre>\n\n<p>By default, for all projects, you can set it up by: <\/p>\n\n<pre><code>File -&gt; Default Settings -&gt; Project Interpreter\n<\/code><\/pre>\n\n<p>This may have defaulted to a different version of python, or sometimes just <code>&lt;no interpreter&gt;<\/code>. You should then be able to see which packages are installed and select interpreter accordingly.<\/p>\n"},{"AnswerId":"54523742","CreationDate":"2019-02-04T20:18:42.683","ParentId":null,"OwnerUserId":"4924700","Title":null,"Body":"<p>In you are on ubuntu and have installed pytorch using anaconda then in the interpreter of the pycharm which is under \n <em>File -> Settings -> Interpreter<\/em> you have to do following:<\/p>\n\n<p>search for the conda where it is installed by doing <code>where conda<\/code> then in that you will find a python.exe in your environment folder provide the path of that in the python interpreter.\nTry with this it should work as I was facing the similar issue on windows solved it using this procedure.\nAnd probably <code>where conda<\/code> won't work if you don't have support for ubuntu commands in windows than you can go to the folder where Anaconda is there which probably will be in your Users -> 'Your User' folder in that search for you environment and in that go to your environment and find python.exe give this path in pycharm.<\/p>\n"},{"AnswerId":"57046592","CreationDate":"2019-07-15T20:13:58.247","ParentId":null,"OwnerUserId":"6110353","Title":null,"Body":"<p>Installing 64bit version is fixing the problem. <\/p>\n"}]}
+{"QuestionId":52857073,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network><object-detection>","CreationDate":"2018-10-17T14:16:20.667","AcceptedAnswerId":null,"OwnerUserId":9751975.0,"Title":"Keras Conv2D output dimension. Multiclass instance segmentation","Body":"<p>Im trying to create a keras instance segmentation CNN model with an Unet architecture. \nThe Keras CNN model i want to use\/modify is:<a href=\"https:\/\/ufile.io\/32zp4\" rel=\"nofollow noreferrer\">model.py<\/a>\nThe CNN model should be able to detect 3 different object classes:(Main root, secondary root, stem). <\/p>\n\n<p>I've converted the annotation-polygons(3 classes) to bitmap masks with a modified version of the balloon.py file in the samples folder in:<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\" rel=\"nofollow noreferrer\">GitHub: Mask RCNN Instance segmentation package<\/a><\/p>\n\n<p>My annotation->bitmap mask file:<a href=\"https:\/\/ufile.io\/65k2x\" rel=\"nofollow noreferrer\">annotation.py<\/a>\nVisualization of bitmaps of the 3 classes:\n<a href=\"https:\/\/i.stack.imgur.com\/acsqN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/acsqN.png\" alt=\"Bitmap masks\"><\/a><\/p>\n\n<p>I want a last Conv2D output of 3 (my classes) feature maps, a final softmax activation and an categorical_crossentropy. I don't know how to \"compare\" my annotations to the models predictions. Do i have to use some kind of keras.losses.categorical_crossentropy(y_true, y_pred) function after the last Conv2D layer(conv9) in my <a href=\"https:\/\/ufile.io\/32zp4\" rel=\"nofollow noreferrer\">model.py<\/a> file? <\/p>\n","answers":[]}
+{"QuestionId":52857100,"AnswerCount":0,"Tags":"<c++><cmake><caffe><lmdb>","CreationDate":"2018-10-17T14:17:23.667","AcceptedAnswerId":null,"OwnerUserId":6129428.0,"Title":"Cmake fails in compiling Caffe with lmdb.h","Body":"<p>Windows 10 Visual Studio 14<\/p>\n\n<p>I am trying to build a c++ code based on <a href=\"https:\/\/i.stack.imgur.com\/rweq0.png\" rel=\"nofollow noreferrer\">this github page<\/a>. However, when I try to generate lib files in Cmake I get this error:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rweq0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rweq0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I don't know what's wrong with lmdb.h file. Should I build lmdb from scratch?<\/p>\n\n<p>There are 2 error messages:<\/p>\n\n<ol>\n<li><blockquote>\n  <p>CMake Warning (dev) at cmake\/Dependencies.cmake:46 (find_package):<br>\n  Policy CMP0074 is not set: find_package uses _ROOT\n  variables.   Run \"cmake --help-policy CMP0074\" for policy details. \n  Use the cmake_policy   command to set the policy and suppress this\n  warning.<\/p>\n  \n  <p>Environment variable HDF5_ROOT is set to:<\/p>\n\n<pre><code>C:\\Program Files\\HDF_Group\\HDF5\\1.10.3\n<\/code><\/pre>\n  \n  <p>For compatibility, CMake is ignoring the variable. Call Stack (most\n  recent call first):   CMakeLists.txt:49 (include) This warning is for\n  project developers.  Use -Wno-dev to suppress it.<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>CMake Error at cmake\/Utils.cmake:105 (if):   Syntax error in cmake\n  code at<\/p>\n\n<pre><code>C:\/Project\/RCNN\/Faster RCNN\/caffe-faster-rcnn-dev\/cmake\/Utils.cmake:105\n<\/code><\/pre>\n  \n  <p>when parsing string<\/p>\n\n<pre><code>C:\\Project\\RCNN\\Faster RCNN\\Packages\\lmdb-mdb.master\\libraries\\liblmdb\/lmdb.h\n<\/code><\/pre>\n  \n  <p>Invalid escape sequence \\P Call Stack (most recent call first):<br>\n  cmake\/Modules\/FindLMDB.cmake:25 (caffe_parse_header)<br>\n  cmake\/Dependencies.cmake:52 (find_package)   CMakeLists.txt:49\n  (include)<\/p>\n  \n  <p><a href=\"https:\/\/i.stack.imgur.com\/rweq0.png\" rel=\"nofollow noreferrer\">1<\/a>: <a href=\"https:\/\/github.com\/D-X-Y\/caffe-faster-rcnn\/tree\/dev\" rel=\"nofollow noreferrer\">https:\/\/github.com\/D-X-Y\/caffe-faster-rcnn\/tree\/dev<\/a>   [2]:\n  <a href=\"https:\/\/i.stack.imgur.com\/hc30J.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/hc30J.png<\/a><\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>Also the paths for GLOG, GFLAGS, and PROTOBUF are not found.<\/p>\n","answers":[]}
+{"QuestionId":52857213,"AnswerCount":1,"Tags":"<python><pytorch><recurrent-neural-network>","CreationDate":"2018-10-17T14:23:05.107","AcceptedAnswerId":52973933.0,"OwnerUserId":2253564.0,"Title":"Recurrent network (RNN) won't learn a very simple function (plots shown in the question)","Body":"<p>So I am trying to train a simple recurrent network to detect a \"burst\" in an input signal. The following figure shows the input signal (blue) and the desired (classification) output of the RNN, shown in red.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/c7Sor.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/c7Sor.png\" alt=\"Then end of the sine-shaped input signal burst should be detected.\"><\/a> <\/p>\n\n<p>So the output of the network should switch from 1 to 0 whenever the burst is detected and stay like with that output. The only thing that changes between the input sequences used to train the RNN is at which time step the burst occurs.<\/p>\n\n<p>Following the Tutorial on <a href=\"https:\/\/github.com\/MorvanZhou\/PyTorch-Tutorial\/blob\/master\/tutorial-contents\/403_RNN_regressor.py\" rel=\"noreferrer\">https:\/\/github.com\/MorvanZhou\/PyTorch-Tutorial\/blob\/master\/tutorial-contents\/403_RNN_regressor.py<\/a>, I cannot get a RNN to learn. The learned RNN always operates in a \"memoryless\" way, i.e., does not use memory to make its predictions, as shown in the following example behavior:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/24wDJ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/24wDJ.png\" alt=\"The same plot as before, but this time with the output behavior of the network.\"><\/a><\/p>\n\n<p>The green line shows the predicted output of the network. <strong>What do I do wrong in this example so that the network cannot be learned correctly?<\/strong> Isn't the network task quite simple?<\/p>\n\n<p>I'm using:<\/p>\n\n<ol>\n<li>torch.nn.CrossEntropyLoss as loss function<\/li>\n<li>The Adam Optimizer for learning<\/li>\n<li>A RNN with 16 internal\/hidden nodes and 2 output nodes. They use the default activation function of the torch.RNN class.<\/li>\n<\/ol>\n\n<p>The experiment has been repeated a couple of times with different random seeds, but there is little difference in the outcomes. I've used the following code:<\/p>\n\n<pre><code>import torch\nimport numpy, math\nimport matplotlib.pyplot as plt\n\nnofSequences = 5\nmaxLength = 130\n\n# Generate training data\nx_np = numpy.zeros((nofSequences,maxLength,1))\ny_np = numpy.zeros((nofSequences,maxLength))\nnumpy.random.seed(1)\nfor i in range(0,nofSequences):\n    startPos = numpy.random.random()*50\n    for j in range(0,maxLength):\n        if j&gt;=startPos and j&lt;startPos+10:\n            x_np[i,j,0] = math.sin((j-startPos)*math.pi\/10)\n        else:\n            x_np[i,j,0] = 0.0\n        if j&lt;startPos+10:\n            y_np[i,j] = 1\n        else:\n            y_np[i,j] = 0\n\n\n# Define the neural network\nINPUT_SIZE = 1\nclass RNN(torch.nn.Module):\n    def __init__(self):\n        super(RNN, self).__init__()\n\n        self.rnn = torch.nn.RNN(\n            input_size=INPUT_SIZE,\n            hidden_size=16,     # rnn hidden unit\n            num_layers=1,       # number of rnn layer\n            batch_first=True,\n        )\n        self.out = torch.nn.Linear(16, 2)\n\n    def forward(self, x, h_state):\n        r_out, h_state = self.rnn(x, h_state)\n\n        outs = []    # save all predictions\n        for time_step in range(r_out.size(1)):    # calculate output for each time step\n            outs.append(self.out(r_out[:, time_step, :]))\n        return torch.stack(outs, dim=1), h_state\n\n# Learn the network\nrnn = RNN()\noptimizer = torch.optim.Adam(rnn.parameters(), lr=0.01)\nh_state = None      # for initial hidden state\n\nx = torch.Tensor(x_np)    # shape (batch, time_step, input_size)\ny = torch.Tensor(y_np).long()\n\ntorch.manual_seed(2)\nnumpy.random.seed(2)\n\nfor step in range(100):\n\n    prediction, h_state = rnn(x, h_state)   # rnn output\n\n    # !! next step is important !!\n    h_state = h_state.data        # repack the hidden state, break the connection from last iteration\n\n    loss = torch.nn.CrossEntropyLoss()(prediction.reshape((-1,2)),torch.autograd.Variable(y.reshape((-1,))))         # calculate loss\n    optimizer.zero_grad()                   # clear gradients for this training step\n    loss.backward()                         # backpropagation, compute gradients\n    optimizer.step()                        # apply gradients\n\n    errTrain = (prediction.max(2)[1].data != y).float().mean()\n    print(\"Error Training:\",errTrain.item())\n<\/code><\/pre>\n\n<p>For those who want to reproduce the experiment, the plot is drawn using the following code (using Jupyter Notebook):<\/p>\n\n<pre><code>steps = range(0,maxLength)\nplotChoice = 3\n\nplt.figure(1, figsize=(12, 5))\nplt.ion()           # continuously plot\n\nplt.plot(steps, y_np[plotChoice,:].flatten(), 'r-')\nplt.plot(steps, numpy.argmax(prediction.detach().numpy()[plotChoice,:,:],axis=1), 'g-')\nplt.plot(steps, x_np[plotChoice,:,0].flatten(), 'b-')\n\nplt.ioff()\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52973933","CreationDate":"2018-10-24T16:30:37.737","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>From the documentation of <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.RNN\" rel=\"noreferrer\">tourch.nn.RNN<\/a>, the RNN is actually an Elman network, and have the following properties seen <a href=\"https:\/\/en.wikipedia.org\/wiki\/Recurrent_neural_network#Elman_networks_and_Jordan_networks\" rel=\"noreferrer\">here<\/a>.\nThe output of an Elman network is only dependent on the hidden state, while the hidden state is dependent on the last input and the previous hidden state.<\/p>\n\n<p>Since we have set \u201ch_state = h_state.data\u201d, we actually use the hidden state of the last sequence to predict the first state of the new sequence, which will result in an output heavily dependent on the last output of the previous sequence (which was 0). The Elman network can\u2019t separate if we are in the beginning of the sequence or at the end, it only \"sees\" the state and last input.<\/p>\n\n<p>To fix this we can insted set \u201ch_state = None\u201d.\nNow every new sequence start with an empty state. This result in the following prediction (where green line again shows the prediction).<a href=\"https:\/\/i.stack.imgur.com\/GI2mp.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GI2mp.jpg\" alt=\"enter image description here\"><\/a>\nNow we start off at 1, but quickly dips down to 0 before the puls push it back up again.\nThe Elman network can account for some time dependency, but it is not good at remembering long term dependencies, and converge towards an \"most common output\" for that input.<\/p>\n\n<p>So to fix this problem, I suggest using a network which is well known for handling long term dependencies well, namely the Long short-term memory (LSTM) rnn, for more information see <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.LSTM\" rel=\"noreferrer\">torch.nn.LSTM<\/a>. Keep \"h_state = None\" and change torch.nn.RNN to torch.nn.LSTM.<\/p>\n\n<p>for complete code and plot see below<\/p>\n\n<pre><code>import torch\nimport numpy, math\nimport matplotlib.pyplot as plt\n\nnofSequences = 5\nmaxLength = 130\n\n# Generate training data\nx_np = numpy.zeros((nofSequences,maxLength,1))\ny_np = numpy.zeros((nofSequences,maxLength))\nnumpy.random.seed(1)\nfor i in range(0,nofSequences):\n    startPos = numpy.random.random()*50\n    for j in range(0,maxLength):\n        if j&gt;=startPos and j&lt;startPos+10:\n            x_np[i,j,0] = math.sin((j-startPos)*math.pi\/10)\n        else:\n            x_np[i,j,0] = 0.0\n        if j&lt;startPos+10:\n            y_np[i,j] = 1\n        else:\n            y_np[i,j] = 0\n\n\n# Define the neural network\nINPUT_SIZE = 1\nclass RNN(torch.nn.Module):\n    def __init__(self):\n        super(RNN, self).__init__()\n\n        self.rnn = torch.nn.LSTM(\n            input_size=INPUT_SIZE,\n            hidden_size=16,     # rnn hidden unit\n            num_layers=1,       # number of rnn layer\n            batch_first=True,\n        )\n        self.out = torch.nn.Linear(16, 2)\n\n    def forward(self, x, h_state):\n        r_out, h_state = self.rnn(x, h_state)\n\n        outs = []    # save all predictions\n        for time_step in range(r_out.size(1)):    # calculate output for each time step\n            outs.append(self.out(r_out[:, time_step, :]))\n        return torch.stack(outs, dim=1), h_state\n\n# Learn the network\nrnn = RNN()\noptimizer = torch.optim.Adam(rnn.parameters(), lr=0.01)\nh_state = None      # for initial hidden state\n\nx = torch.Tensor(x_np)    # shape (batch, time_step, input_size)\ny = torch.Tensor(y_np).long()\n\ntorch.manual_seed(2)\nnumpy.random.seed(2)\n\nfor step in range(100):\n\n    prediction, h_state = rnn(x, h_state)   # rnn output\n\n    # !! next step is important !!\n    h_state = None        \n\n    loss = torch.nn.CrossEntropyLoss()(prediction.reshape((-1,2)),torch.autograd.Variable(y.reshape((-1,))))         # calculate loss\n    optimizer.zero_grad()                   # clear gradients for this training step\n    loss.backward()                         # backpropagation, compute gradients\n    optimizer.step()                        # apply gradients\n\n    errTrain = (prediction.max(2)[1].data != y).float().mean()\n    print(\"Error Training:\",errTrain.item())\n\n\n###############################################################################\nsteps = range(0,maxLength)\nplotChoice = 3\n\nplt.figure(1, figsize=(12, 5))\nplt.ion()           # continuously plot\n\nplt.plot(steps, y_np[plotChoice,:].flatten(), 'r-')\nplt.plot(steps, numpy.argmax(prediction.detach().numpy()[plotChoice,:,:],axis=1), 'g-')\nplt.plot(steps, x_np[plotChoice,:,0].flatten(), 'b-')\n\nplt.ioff()\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8RkC9.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8RkC9.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":52857557,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><ensemble-learning>","CreationDate":"2018-10-17T14:41:09.187","AcceptedAnswerId":52857819.0,"OwnerUserId":7575552.0,"Title":"Ensemble of custom and pre-trained model gives run-time error","Body":"<p>I'm attempting to create an ensemble of a custom CNN and pre-trained VGG16 for a medical image classification task using Keras with Tensorflow backend (TF: 1.9.0 and Keras: 2.1.6). The code is as given below:<\/p>\n\n<pre><code>#load libraries\nfrom keras.models import Model, Input\nfrom keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D, Activation, Average, Dense\nfrom load_data import load_resized_training_data, load_resized_validation_data\nfrom load_data import load_resized_test_data\nfrom keras.losses import categorical_crossentropy\nfrom keras.callbacks import ModelCheckpoint, TensorBoard\nfrom keras.optimizers import Adam\nfrom keras.applications.vgg16 import VGG16\nimport numpy as np\n################################################################################\n#load data\nbatch_size = 8\nnum_epochs = 1\nimg_rows= 224\nimg_cols = 224\nnum_channels = 3\nnum_classes = 2\n\nX_train, Y_train = load_resized_training_data(img_rows, img_cols)\nX_valid, Y_valid = load_resized_validation_data(img_rows, img_cols)\nX_test, Y_test = load_resized_test_data(img_rows, img_cols)\nprint(X_train.shape, Y_train.shape, X_valid.shape, Y_valid.shape,X_test.shape, Y_test.shape)\nX_train = X_train.astype('float32')\nX_valid = X_valid.astype('float32')\nX_test = X_test.astype('float32')\nX_train \/= 255\nX_valid \/= 255\nX_test \/= 255 \n###############################################################################  \n'''Since the two models work with the data of the same shape, \nit makes sense to define a single input layer that will be used by every model.'''\ninput_shape = X_train[0,:,:,:].shape\nprint(input_shape) # 224, 224, 3\nmodel_input = Input(shape=input_shape)\nprint(model_input) # Tensor(\"input_1:0\", shape=(?, 224, 224, 3), dtype=float32)\n\n###############################################################################\n'''define the first model: a simple sequential model in the form of functional api'''\n\nx = Conv2D(16, kernel_size=(3, 3), activation='relu')(model_input)\nx = MaxPooling2D((2, 2))(x)\nx = Conv2D(32, (3, 3), activation='relu')(x)\nx = MaxPooling2D((2, 2))(x)\nx = Conv2D(64, (3, 3), activation='relu')(x)\nx = MaxPooling2D((2, 2))(x)\nx = Conv2D(128, (3, 3), activation='relu')(x)\nx = MaxPooling2D((2, 2))(x)\nx = Conv2D(256, (3, 3), activation='relu')(x)\nx = MaxPooling2D((2, 2))(x)\nx = Conv2D(512, (3, 3), activation='relu')(x)\nx = Conv2D(2, (1, 1))(x)\nx = GlobalAveragePooling2D()(x)\nx = Activation(activation='softmax')(x)\n\ncustom_model = Model(inputs=model_input, outputs=x, name='custom_cnn')\n###############################################################################\ndef compile_and_train(model, num_epochs): \n\n    model.compile(loss=categorical_crossentropy, optimizer=Adam(), metrics=['acc']) \n    filepath = 'weights\/' + model.name + '.{epoch:02d}-{val_acc:.2f}.hdf5'\n    checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_weights_only=True, save_best_only=True, mode='auto', period=1)\n    tensor_board = TensorBoard(log_dir='logs\/', histogram_freq=0, batch_size=batch_size)\n    history = model.fit(X_train, Y_train, batch_size=batch_size, epochs=num_epochs, verbose=1, callbacks=[checkpoint, tensor_board], validation_data=(X_valid, Y_valid))\n    return history\n\n#compile and train the model\n_ = compile_and_train(custom_model, num_epochs=num_epochs)\n###############################################################################\ndef evaluate_error(model):\n    pred = model.predict(X_test, batch_size = batch_size)\n    pred = np.argmax(pred, axis=1)\n    pred = np.expand_dims(pred, axis=1) # make same shape as y_test\n    error = np.sum(np.not_equal(pred, Y_test)) \/ Y_test.shape[0]  \n\n    return error\n\nevaluate_error(custom_model)\n###############################################################################\n'''second model is a pretrained vgg16 model initialized with imagenet weights and going to be trained from the first layer'''\n\nvgg16_model = VGG16(weights='imagenet', include_top=False, input_shape=(img_rows, img_cols, 3))\nx = vgg16_model.output\nx = GlobalAveragePooling2D()(x)\npredictions = Dense(num_classes, activation='softmax')(x)\nvgg16_custom_model = Model(inputs=vgg16_model.input, outputs=predictions, name='vgg16_cnn')\n\n#compile and train the model\n_ = compile_and_train(vgg16_custom_model, num_epochs=num_epochs)\n\n#Evaluate the model by calculating the error on the test set\nevaluate_error(vgg16_custom_model)\n###############################################################################\ncustom_model.load_weights('weights\/custom_cnn.01-0.60.hdf5')\nvgg16_custom_model.load_weights('weights\/vgg16_cnn.01-0.50.hdf5')\n\nmodels = [custom_model, vgg16_custom_model]\n###############################################################################\ndef ensemble(models, model_input):\n\n    outputs = [model.outputs[0] for model in models]\n    y = Average()(outputs)\n\n    model = Model(inputs=model_input, outputs=y, name='ensemble')\n\n    return model\n\nensemble_model = ensemble(models, model_input)\nevaluate_error(ensemble_model)\n###############################################################################\n<\/code><\/pre>\n\n<p>The code works all fine until the creation of the ensemble definition. When defining the ensemble I get the following error:<\/p>\n\n<pre><code>RuntimeError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_7:0\", shape=(?, 224, 224, 3), dtype=float32) at layer \"input_7\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>I am not sure whether I can pass <code>model_input<\/code> to the pre-trained VGG16 in this manner. How the ensemble definition needs to be modified?<\/p>\n","answers":[{"AnswerId":"52857819","CreationDate":"2018-10-17T14:54:23.083","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that the input of VGG model is not being fed by the input layer of <code>ensemble_model<\/code>. To resolve this issue, you need to modify the definition of <code>ensemble_model<\/code> and create a new input layer which is then passed to the two models:<\/p>\n\n<pre><code>def ensemble(models):\n    input_img = Input(shape=input_shape)\n\n    outputs = [model(input_img) for model in models] # get the output of model given the input image\n    y = Average()(outputs)\n\n    model = Model(inputs=input_img, outputs=y, name='ensemble')\n    return model\n\nensemble_model = ensemble(models)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52857675,"AnswerCount":1,"Tags":"<windows><tensorflow><machine-learning><keras>","CreationDate":"2018-10-17T14:46:31.417","AcceptedAnswerId":null,"OwnerUserId":10456659.0,"Title":"How to install Tensorflow & keras in windows without cuda","Body":"<p>Can anyone guide me how to install keras and tensorflow without cuda in windows as i have only AMD graphics card..How to make it run in CPU?I have already installed python 3.6 and anaconda environment..<\/p>\n","answers":[{"AnswerId":"52858021","CreationDate":"2018-10-17T15:04:26.147","ParentId":null,"OwnerUserId":"9316652","Title":null,"Body":"<p>You can find more information on this link:\n<a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a><\/p>\n\n<p>Follow the instructions of:\n<strong>2. Create a virtual environment (recommended)<\/strong>, and select <strong>Conda<\/strong>.<\/p>\n\n<p>For the following line of code: <code>pip install --ignore-installed --upgrade packageURL<\/code><br>\nyou can use: <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.11.0-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.11.0-cp36-cp36m-win_amd64.whl<\/a> for <code>packageURL<\/code><\/p>\n\n<p>This is the TensorFlow Windows package, for Python 3.6 and CPU only (since you don't have a Nvidia GPU).<\/p>\n"}]}
+{"QuestionId":52857959,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><classification>","CreationDate":"2018-10-17T15:01:31.207","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Same output of the Keras model","Body":"<p>I have a Keras model for predicting moves in the game. I have an input shape of <code>(160,120 ,1)<\/code>. I have the following model with an output of 9 nodes:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D, ZeroPadding2D\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.optimizers import Adam\nfrom keras.regularizers import l2\nfrom keras import optimizers\ndef alexnet_model(n_classes=9, l2_reg=0.,\n    weights=None):\n\n    # Initialize model\n    alexnet = Sequential()\n    alexnet.add(Conv2D(24, (11, 11), input_shape=(160,120,1), activation ='relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Conv2D(36, (5, 5), activation ='relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Conv2D(48, (3, 3),  activation ='relu'))\n    alexnet.add(Conv2D(54, (3, 3),  activation ='relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Flatten())\n    alexnet.add(Dense(300,   activation ='tanh'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(200,   activation ='tanh'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(100,   activation ='tanh'))\n    alexnet.add(Dropout(0.5))\n\n\n    alexnet.add(Dense(n_classes , activation = 'softmax'))\n\n    optimizer = Adam(lr=1e-3)\n\n    alexnet.compile(loss='categorical_crossentropy', optimizer=optimizer)\n\n\n    alexnet.summary()\n\n\n    return alexnet\n<\/code><\/pre>\n\n<p>Then, I run a training script. My <code>X<\/code> has a shape of <code>(12862, 160, 120, 1)<\/code> and <code>y<\/code> of <code>(1000,9)<\/code>.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom random import shuffle\nimport pandas as pd\nfrom tensorflow.keras import layers,models\nfrom keras.preprocessing.image import ImageDataGenerator\nimport tensorflow as tf\n# what to start at\nSTART_NUMBER = 60\n\n# what to end at\nhm_data = 111\n\n# use a previous model to begin?\nSTART_FRESH = False\nWIDTH = 160\nHEIGHT = 120\nLR = 1e-3\nEPOCHS = 1\n\nMODEL_NAME = 'model_new.h5'\nEXISTING_MODEL_NAME = ''\n\nmodel = alexnet_model()\n\nX=[]\n\nY=[]\nfor i in range(EPOCHS):\n    train_data = np.load('training_data_1.npy')\n    print(len(train_data))\n    train = train_data[0:12862]\n    test = train_data[-1000:]\n\n    X = np.array([i[0] for i in train]).reshape(-1,WIDTH,HEIGHT,1)\n    Y = np.array([i[1] for i in train])\n\n    test_x = np.array([i[0] for i in test]).reshape(-1,WIDTH,HEIGHT,1)\n    test_y = np.array([i[1] for i in test])\n    print(X.shape)\n    model.fit(X, Y , batch_size = 16, epochs = 10 , validation_data = (test_x, test_y), verbose=1)\n    model.save(MODEL_NAME)\n\n# tensorboard --logdir=foo:C:\/Users\/H\/Desktop\/ai-gaming-phase5\/log\n<\/code><\/pre>\n\n<p>After testing the model I get an output:<\/p>\n\n<pre><code>array([[2.8518048e-01, 5.5075828e-03, 7.3730588e-02, 5.3255934e-02,\n        1.0635615e-01, 6.4690344e-02, 9.1519929e-08, 7.0413840e-08,\n        4.1127869e-01]], dtype=float32)\n<\/code><\/pre>\n\n<p>with this line of code:<\/p>\n\n<pre><code>model.predict(X[100].reshape(-1,160,120,1)) \n<\/code><\/pre>\n\n<p>I know that it is not good to test model on <code>X<\/code> but it doesn't matter which picture I use but I get the same output. Just for reference (my <code>Y<\/code> values):<\/p>\n\n<pre><code>w = [1,0,0,0,0,0,0,0,0]\ns = [0,1,0,0,0,0,0,0,0]\na = [0,0,1,0,0,0,0,0,0]\nd = [0,0,0,1,0,0,0,0,0]\nwa = [0,0,0,0,1,0,0,0,0]\nwd = [0,0,0,0,0,1,0,0,0]\nsa = [0,0,0,0,0,0,1,0,0]\nsd = [0,0,0,0,0,0,0,1,0]\nnk = [0,0,0,0,0,0,0,0,1]\n<\/code><\/pre>\n\n<p>I tried another model but it still doesn't work. Here is the amount of training data for each class:<\/p>\n\n<pre><code>Counter({'[1, 0, 0, 0, 0, 0, 0, 0, 0]': 5000,\n         '[0, 0, 0, 0, 0, 0, 0, 0, 1]': 5000,\n         '[0, 0, 0, 0, 1, 0, 0, 0, 0]': 1183,\n         '[0, 0, 0, 0, 0, 1, 0, 0, 0]': 982,\n         '[0, 0, 1, 0, 0, 0, 0, 0, 0]': 832,\n         '[0, 0, 0, 1, 0, 0, 0, 0, 0]': 764,\n         '[0, 1, 0, 0, 0, 0, 0, 0, 0]': 101})\n<\/code><\/pre>\n\n<p>I think that the problem is in the model but I don't know how to change it. Could it be the problem of small training data? The loss valus is also not going down: <code>loss: 1.7416 - val_loss: 1.4639<\/code>. It only decreases by a few decimals and sometimes even goes back up.<\/p>\n","answers":[{"AnswerId":"52859705","CreationDate":"2018-10-17T16:36:44.443","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>From what it appears in your code, and since you mentioned that the loss is decreasing very slowly, the best guess is that the input data (which I think are images) is not normalized and therefore this prevents a smooth gradient flow. Try normalizing them. One simple way of doing it is like this:<\/p>\n\n<pre><code>X = X.astype('float32') \/ 255.0\ntest_x = test_x.astype('float32') \/ 255.0\n<\/code><\/pre>\n\n<p>Further, you may need to account for the class imbalance in the training data and counter it by using <code>class_weights<\/code> argument in <a href=\"https:\/\/keras.io\/models\/sequential\/#fit\" rel=\"nofollow noreferrer\"><code>fit<\/code><\/a> method (look at the doc to find out how it can be used).<\/p>\n"},{"AnswerId":"52879651","CreationDate":"2018-10-18T17:37:47.620","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Solved!\nJust normalizing training data didn't work. I decreased amount of nodes and layers and everything worked fine. I guess it was an overfitting problem. <\/p>\n"}]}
+{"QuestionId":52858309,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-17T15:18:09.563","AcceptedAnswerId":null,"OwnerUserId":9897331.0,"Title":"Keras custom layer how to have multiple returns","Body":"<p>I'm writing a custom layer in Keras with Tensorflow. The call function has multiple steps and I want to know if it's possible to skip some if a condition is met. \nFor example, let's take this function:<\/p>\n\n<pre><code>def test_fn(x):\n    if x &lt; 10:\n        return 0\n    else:\n        return 1\n<\/code><\/pre>\n\n<p>Is there a way to do something like that in Keras?\nI have something like this (simplified):<\/p>\n\n<pre><code>class Test_layer(layers.Layer):\n\n    def call(self, inputs, **kwargs):\n        test = get_test_fn(inputs)\n        if test &lt; 10:\n            return 0\n        else:\n            return 1\n<\/code><\/pre>\n\n<p>I know the syntax there is wrong, but is there a way to do something equivalent?<\/p>\n","answers":[]}
+{"QuestionId":52858508,"AnswerCount":0,"Tags":"<python><theano><curve-fitting><bayesian><pymc3>","CreationDate":"2018-10-17T15:28:12.193","AcceptedAnswerId":null,"OwnerUserId":379989.0,"Title":"How do I round and make comparisons with PyMC3 prior parameters in a model?","Body":"<p>I apologize in advance if it seems more complex than needed. Tried to make it more minimal than this, but then some of the questions were more difficult to formulate. This way it is as close as possible to my situation. What I have is some data that I want to fit with a function similar to what is described below. But when implementing the pyMC3 fit, as in the second code section, I run into the errors relating to the two questions.<\/p>\n\n<pre><code># some constants and fixed things\nrin = 1.\nrout = 1000.\ntmin = 10.\nx = np.arange(0.1, 10, 0.2)\n\n# parameters to be fitted\na = 4.0\nb = 0.5\nc = 10\n\n# define r_struct, depends on b, but is one value\nr_struct = 0.1**(1.\/b)\n\n# nrad below, the number of radial points,  \n# depends on r_struct and the constant rout\n# QUESTION 1 here\nnrad = round(r_struct, -1) + round((rout - r_struct)\/10., -1) * 2\n\n# define the radial bins, logarithmically spaced\nrbins = np.logspace(np.log10(rin), np.log10(rout), num=nrad+1, endpoint=True)\nrad = 10**(( np.log10(rbins[1:]) + np.log10(rbins[:-1]))\/2.)\nt = rad**(-b)\n\n# QUESTION 2 here\n# t cannot go below tmin, so I do this\nt[t &lt; tmin] = tmin\n\n# then I want to know where the limit of r_struct is here\niout = rad &gt; r_struct\niin = rad &lt;= r_struct\n\n# depending on if rad is past r_struct or not\n# I do two different things.\nc1  = c_struct * (rad[iin]\/50.)**(-1.*a)\ntaper = np.exp( -(rad[iout]\/rad[iout][0])**(2.-a) )\nc2 = c_taper * taper\/taper[0]\ntau = np.append(c1, c2) \n\ny_true = J0( x[:, np.newaxis] * rad) * (t * (1.0 - np.exp(-1.*tau) ) )\ny_true = y_true.sum(axis=1)\n\ny_obs = y_true + np.random.normal(size=len(x))*50\ny_error = np.ones_like(y_true)*0.1*y_true.max()\n\nplt.errorbar(x,y_obs, yerr= y_error, color='r', marker='.', ls='None', label='Observed')\nplt.plot(x,y_true, 'b', marker='None', ls='-', lw=1, label='True')\nplt.xlabel('x'); plt.ylabel('y'); plt.legend()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NAYMv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NAYMv.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>OK so now I want to fit this in PyMC3. I started with the following.<\/p>\n\n<pre><code>with pm.Model() as basic_model:\n    # parameters\n    a = pm.Uniform(\"a\", 0, 10)\n    b = pm.Uniform(\"b\", 0, 10)\n    c = pm.Uniform(\"c\", 0,100)\n\n    r_struct = 0.1**(1.\/b)\n    nrad = round(r_struct, -1) + round((rout - r_struct)\/10., -1) * 2\n\n    rbins = np.logspace(np.log10(rin), np.log10(rout), num=nrad+1, endpoint=True)\n    rad = 10**(( np.log10(rbins[1:]) + np.log10(rbins[:-1]))\/2.)\n    t = rad**(-b)\n\n    t[t &lt; tmin] = tmin\n    iout = rad &gt; r_struct\n    iin = rad &lt;= r_struct\n\n    c1  = c_struct * (rad[iin]\/50.)**(-1.*a)\n    taper = np.exp( -(rad[iout]\/rad[iout][0])**(2.-a) )\n    c2 = c_taper * taper\/taper[0]\n\n    tau = np.append(c1, c2) \n    y_true = J0( x[:, np.newaxis] * rad) * (t * (1.0 - np.exp(-1.*tau) ) )\n    y_true = y_true.sum(axis=1)\n\n    func=pm.Deterministic('gauss',y_true)\n\n    y = pm.Normal('y', mu=func, tau=1.0\/y_error**2, observed=y_obs)\n\n    trace = pm.sample(5000)\n<\/code><\/pre>\n\n<p>However this fails already at the first \"round()\" call. See questions below.<\/p>\n\n<p><strong>QUESTION 1.<\/strong> <em>\"nrad\" needs to be an integer, it depends on the pymc3 prior \"b\" (through r_struct), how do I round it to and integer when running PyMC3?<\/em>\nFor reference, I get this error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-11-fcf897a21b89&gt; in &lt;module&gt;()\n     15     r_struct = 0.1**(1.\/b)\n     16\n---&gt; 17     nrad = round(r_struct, -1) + round((rout - r_struct)\/10., -1) * 2\n     18     rbins = np.logspace(np.log10(rin), np.log10(rout), num=nrad+1, endpoint=True)\n\nTypeError: type TensorVariable doesn't define __round__ method\n<\/code><\/pre>\n\n<p><strong>QUESTION 2:<\/strong> <em>How do I make comparisons with an array created with a PyMC3 prior?<\/em> The code fails at the comparison.<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAssertionError                            Traceback (most recent call last)\n&lt;ipython-input-12-7f77e70d147a&gt; in &lt;module&gt;()\n     26     t = rad**(-b)\n     27\n---&gt; 28     t[t &lt; tmin] = tmin\n     29     iout = rad &gt; r_struct\n     30     iin = rad &lt;= r_struct\n<\/code><\/pre>\n\n<p>Any suggestions on how to clarify the question, including the title are all welcome!<\/p>\n","answers":[]}
+{"QuestionId":52858515,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><classification>","CreationDate":"2018-10-17T15:28:34.250","AcceptedAnswerId":52858921.0,"OwnerUserId":7924106.0,"Title":"Error passing multiple inputs into Keras model","Body":"<p>I want to train a binary classifier using Keras and my training data is of shape <code>(2000,2,128)<\/code> and labels of shape <code>(2000,)<\/code> as Numpy arrays.<\/p>\n\n<p>The idea is to train such that embeddings together in a single array means they are either same or different, labelled using 0 or 1 respectively.<\/p>\n\n<p>The training data looks like:\n<code>[[[0 1 2 ....128][129.....256]][[1 2 3 ...128][9 9 3 5...]].....]<\/code>\n and the labels looks like <code>[1 1 0 0 1 1 0 0..]<\/code>.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import keras\nfrom keras.layers import Input, Dense\n\nfrom keras.models import Model\n\nfrst_input = Input(shape=(128,), name='frst_input')\nscnd_input = Input(shape=(128,),name='scnd_input')\nx = keras.layers.concatenate([frst_input, scnd_input])\nx = Dense(128, activation='relu')(x)\nx=(Dense(1, activation='softmax'))(x)\nmodel=Model(inputs=[frst_input, scnd_input], outputs=[x])\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy',\n              loss_weights=[ 0.2],metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I am getting the following error while running this code:<\/p>\n\n<pre><code>ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[[ 0.07124118, -0.02316936, -0.12737238, ...,  0.15822273,\n      0.00129827, -0.02457245],\n    [ 0.15869428, -0.0570458 , -0.10459555, ...,  0.0968155 ,\n      0.0183982 , -0.077924...\n<\/code><\/pre>\n\n<p>How can I resolve this issue? Is my code correct to train the classifier using two inputs to classify?<\/p>\n","answers":[{"AnswerId":"52858921","CreationDate":"2018-10-17T15:50:42.243","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Well, you have two options here:<\/p>\n\n<p>1) Reshape the training data to <code>(2000, 128*2)<\/code> and define only one input layer:<\/p>\n\n<pre><code>X_train = X_train.reshape(-1, 128*2)\n\ninp = Input(shape=(128*2,))\nx = Dense(128, activation='relu')(inp)\nx = Dense(1, activation='sigmoid'))(x)\nmodel=Model(inputs=[inp], outputs=[x])\n<\/code><\/pre>\n\n<p>2) Define two input layers, as you have already done, and pass a <strong>list<\/strong> of two input arrays when calling <code>fit<\/code> method:<\/p>\n\n<pre><code># assuming X_train have a shape of `(2000, 2, 128)` as you suggested\nmodel.fit([X_train[:,0], X_train[:,1]], y_train, ...)\n<\/code><\/pre>\n\n<p>Further, since you are doing binary classification here, you need to use <code>sigmoid<\/code> as the activation of last layer (i.e. using <code>softmax<\/code> in this case would always outputs 1, since <code>softmax<\/code> normalizes the outputs such that their sum equals to one).<\/p>\n"}]}
+{"QuestionId":52859359,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-17T16:15:53.557","AcceptedAnswerId":null,"OwnerUserId":1019129.0,"Title":"Train NN with time series?","Body":"<p>As far as I understand Keras NN shuffles the data during training.<\/p>\n\n<p>How do you train NN with time series data where input and output data has to be in sync during training i.e. train at the same sequence as the data is ordered ? (Normal NN, not LSTM or something).<\/p>\n","answers":[]}
+{"QuestionId":52859477,"AnswerCount":1,"Tags":"<tensorflow><object-detection>","CreationDate":"2018-10-17T16:22:33.207","AcceptedAnswerId":null,"OwnerUserId":10519342.0,"Title":"Custom Object detection in tensorflow along with existing detection","Body":"<p>I have setup Tensorflow Object Detection model(e.g.SSD with coco or any other model) to detect a 'car' in a picture, which works just fine. Now I want to train the same SSD model with my own dataset using pre-trained SSD model on coco, to detect a custom object, say a 'door'. \nQ: After training the model, will the trained SSD model still be able to detect a 'car' and 'door'? Or the trained model will detect only 'door' as that's what the model is trained for.\nIf the trained model does not detect 'car' and 'door', how do I achieve to detect a 'car'(which is a standard object) and a 'door' which I want the model to be trained to detect a custom object?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"52868621","CreationDate":"2018-10-18T07:02:07.220","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>When fine-tuning a pre-trained model on a different set of classes (doesn't matter if the new set includes the old one, or included in it, etc.), the weights of each of the detection heads are initialized, since it isn't the same layer anymore.\nFor example, if you have an SSD detector the detection head is a convolutional head with #anchors*(4+#classes) channels. If you change #classes, the number of channels is changed, and it's not longer the same layer.\nTherefore, when you fine-tune a model to detect 'car' and 'door', you'll have to train the model on both 'car' and 'door' even though the pre-trained model already knew to detect 'car'.\nIt could be however that since all other layers are restored, that 'car' will be detected more easily and that 'door' will need more training.<\/p>\n"}]}
+{"QuestionId":52859780,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2018-10-17T16:41:41.990","AcceptedAnswerId":null,"OwnerUserId":7781194.0,"Title":"how different between create ANN using a matrix multiplication and tf.layers.dense() in Tensorflow","Body":"<p>I tried to train the ANN model using the matrix multiplication and tf.layers.dense(). But I got diferent result , ANN model using the matrix multiplication it can not optimize the loss function (loss increase). how different between two method ?<\/p>\n\n<p>ANN model using the matrix multiplication<\/p>\n\n<pre><code>W1 = tf.Variable(tf.zeros([4,64]))\nb1 = tf.Variable(tf.zeros([64]))\ny1 = tf.nn.relu(tf.matmul(x, W1) + b1)\n\nW2 = tf.Variable(tf.zeros([64,64]))\nb2 = tf.Variable(tf.zeros([64]))\ny2 = tf.nn.relu(tf.matmul(y1, W2) + b2)\n\nW3 = tf.Variable(tf.zeros([64,64]))\nb3 = tf.Variable(tf.zeros([64]))\ny3 = tf.nn.relu(tf.matmul(y2, W3) + b3)\n\nW4 = tf.Variable(tf.zeros([64,3]))\nb4 = tf.Variable(tf.zeros([3]))\ny_out = tf.nn.softmax(tf.matmul(y3, W4) + b4)\n<\/code><\/pre>\n\n<p>ANN model using tf.layers.dense()<\/p>\n\n<pre><code>layer1 = tf.layers.dense(x, 64, activation=tf.nn.relu)\nlayer2 = tf.layers.dense(layer1, 64, activation=tf.nn.relu)\nlayer3 = tf.layers.dense(layer2, 64, activation=tf.nn.relu)\nlayer4 = tf.layers.dense(layer3, 64, activation=tf.nn.relu)\nlayer5 = tf.layers.dense(layer4, 64, activation=tf.nn.relu)\nlayer6 = tf.layers.dense(layer5, 64, activation=tf.nn.relu)\ny_out = tf.layers.dense(layer6, 3 , activation = tf.nn.softmax)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52859931","CreationDate":"2018-10-17T16:51:13.863","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You are initializing the weights with zeros, which effectively prevents the network from learning anything as the network always outputs zero, and the gradient is always zero.<\/p>\n\n<p>Initialize your weights with random values, like uniform or gaussian distribution with a small range (less than 0.1).<\/p>\n"}]}
+{"QuestionId":52860103,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-17T17:01:19.817","AcceptedAnswerId":null,"OwnerUserId":7447265.0,"Title":"Why does start_queue_runners() function use with shuffle_batch() function","Body":"<p>I am new to TensorFlow and trying to understand the shuffle_batch() function.When I use the shuffle_batch() with following code it does not printing anything. <\/p>\n\n<pre><code> import tensorflow as tf\n\n sess=tf.Session()\n random=tf.random_normal([5],mean=0.0, stddev=1.0)\n shu=tf.train.shuffle_batch([sliced], 20, 100, 10)\n print(sess.run(shu)) \n<\/code><\/pre>\n\n<p>But after adding the <code>start_queue_runners()<\/code> it gives me the expected output. So what is the relationship between these <code>start_queue_runners()<\/code> and <code>shuffle_batch()<\/code> ?<\/p>\n\n<pre><code> import tensorflow as tf\n\n sess=tf.Session()\n random=tf.random_normal([5],mean=0.0, stddev=1.0)\n shu=tf.train.shuffle_batch([sliced], 20, 100, 10)\n threads = tf.train.start_queue_runners(sess=sess)\n print(sess.run(shu)) \n<\/code><\/pre>\n","answers":[{"AnswerId":"52860614","CreationDate":"2018-10-17T17:34:39.317","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>The queue pipeline has been replaced by tf.dataset. You should have a look at this instead. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\" rel=\"nofollow noreferrer\">tf datasets<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">guide<\/a><\/p>\n\n<p>It is much simpler to use the dataset api. <\/p>\n"}]}
+{"QuestionId":52861386,"AnswerCount":1,"Tags":"<python><tensorflow><keras><theano>","CreationDate":"2018-10-17T18:26:11.840","AcceptedAnswerId":52867393.0,"OwnerUserId":5748861.0,"Title":"Keras with Tensorflow backend and Theano backend make different predictions with same model and same input","Body":"<p>I tried to make predictions from a pre-trained model saved as json strings and h5 weights separately, however it seems different backend(Tensorflow and Theano) would give me different output even if the input and model are exactly the same. I found even at the very first layer, which is a 1D convolution, the activations are different, here are the code to print part of the activation from 5th filter of the convolution1D layer:<\/p>\n\n<p><strong>Theano version:<\/strong><\/p>\n\n<pre><code>from keras.models import model_from_json\nimport numpy as np\nimport os\nos.environ['KERAS_BACKEND'] = 'theano'\n\nmodel_file = 'model.h5'\nx_file = 'x.csv'\n\nmodel_json = '{\"class_name\": \"Model\", \"keras_version\": \"1.2.2\", \"config\": {\"layers\": [{\"class_name\": \"InputLayer\", \"config\": {\"batch_input_shape\": [null, 1002, 6], \"input_dtype\": \"float32\", \"sparse\": false, \"name\": \"input_1\"}, \"inbound_nodes\": [], \"name\": \"input_1\"}, {\"class_name\": \"Convolution1D\", \"config\": {\"batch_input_shape\": [null, null, 6], \"W_constraint\": null, \"b_constraint\": null, \"name\": \"convolution1d_1\", \"activity_regularizer\": null, \"trainable\": true, \"filter_length\": 34, \"init\": \"glorot_uniform\", \"bias\": true, \"nb_filter\": 128, \"input_dtype\": \"float32\", \"subsample_length\": 1, \"border_mode\": \"valid\", \"input_dim\": 6, \"b_regularizer\": null, \"W_regularizer\": null, \"activation\": \"relu\", \"input_length\": null}, \"inbound_nodes\": [[[\"input_1\", 0, 0]]], \"name\": \"convolution1d_1\"}], \"input_layers\": [[\"input_1\", 0, 0]], \"output_layers\": [[\"convolution1d_1\", 0, 0]], \"name\": \"model_1\"}}'\nmodel = model_from_json(model_json)\nmodel.load_weights(model_file)\n\nx=np.loadtxt(x_file)\nx = np.reshape(x,(1,x.shape[0],x.shape[1]))\ny = model.predict(x)\ny[0,range(230),4] \n<\/code><\/pre>\n\n<p>The input and output look like:\n<a href=\"https:\/\/i.stack.imgur.com\/x5AyX.png\" rel=\"nofollow noreferrer\">Theano version<\/a><\/p>\n\n<p><strong>Tensorflow version:<\/strong><\/p>\n\n<pre><code>from keras.models import model_from_json\nimport numpy as np\nimport os\nos.environ['KERAS_BACKEND'] = 'tensorflow'\n\nmodel_file = 'model.h5'\nx_file = 'x.csv'\n\nmodel_json = '{\"class_name\": \"Model\", \"keras_version\": \"1.2.2\", \"config\": {\"layers\": [{\"class_name\": \"InputLayer\", \"config\": {\"batch_input_shape\": [null, 1002, 6], \"input_dtype\": \"float32\", \"sparse\": false, \"name\": \"input_1\"}, \"inbound_nodes\": [], \"name\": \"input_1\"}, {\"class_name\": \"Convolution1D\", \"config\": {\"batch_input_shape\": [null, null, 6], \"W_constraint\": null, \"b_constraint\": null, \"name\": \"convolution1d_1\", \"activity_regularizer\": null, \"trainable\": true, \"filter_length\": 34, \"init\": \"glorot_uniform\", \"bias\": true, \"nb_filter\": 128, \"input_dtype\": \"float32\", \"subsample_length\": 1, \"border_mode\": \"valid\", \"input_dim\": 6, \"b_regularizer\": null, \"W_regularizer\": null, \"activation\": \"relu\", \"input_length\": null}, \"inbound_nodes\": [[[\"input_1\", 0, 0]]], \"name\": \"convolution1d_1\"}], \"input_layers\": [[\"input_1\", 0, 0]], \"output_layers\": [[\"convolution1d_1\", 0, 0]], \"name\": \"model_1\"}}'\nmodel = model_from_json(model_json)\nmodel.load_weights(model_file)\n\nx=np.loadtxt(x_file)\nx = np.reshape(x,(1,x.shape[0],x.shape[1]))\ny = model.predict(x)\ny[0,range(230),4]\n<\/code><\/pre>\n\n<p>The input and output look like:\n<a href=\"https:\/\/i.stack.imgur.com\/8d3ne.png\" rel=\"nofollow noreferrer\">Tensorflow version<\/a><\/p>\n\n<p>After a few experiments, I found Theano tend to have the \"wrong\" answer, here is an example to calculate the first window of the 5th filter(bias is zero in this model and I have checked that):<\/p>\n\n<pre><code>l=model.get_layer(index=1)\nw1 = l.get_weights()[0]\nw2 = l.get_weights()[1]\n\n\ndata1 = w1[:,0,:,4]\ndata2 = x[0,range(34),:]\nans=0\nfor i in range(6):\n    ans += np.sum(np.multiply(data1[:,i],data2[:,i]))\n<\/code><\/pre>\n\n<p>Ans equals to 0.08544020017143339\nTensorflow gives 0.08544022, identical to what should be from my calculation, however Theano gives 0.0518605. Could anyone comes up with an explanation about that?<\/p>\n","answers":[{"AnswerId":"52867393","CreationDate":"2018-10-18T05:15:00.700","ParentId":null,"OwnerUserId":"5748861","Title":null,"Body":"<p>It seems the kernel weights in Theano are flipped during calcuation. Code below clearly  showed the difference of Tensorflow and Theano:<\/p>\n\n<pre><code>import numpy as np\nimport os\nos.environ['KERAS_BACKEND'] = 'theano'\n\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation,Convolution1D\nfrom keras.optimizers import SGD\n\nmodel = Sequential()\nmodel.add(Convolution1D(1, 2, border_mode='valid', input_shape=(6, 1),bias=True,activation='relu'))\n\nl=model.get_layer(index=1)\nw1 = l.get_weights()[0]\nw2 = l.get_weights()[1]\nnp.random.seed(0)\nx = np.random.random((1,6,1))\ny = model.predict(x)\n\n\na_tf = np.sum(np.multiply(w1[:,0,0,0],x[0,range(2),0])) # y[0,0] would equal to this with Tensorflow backend\na_th = np.sum(np.multiply(np.flip(w1[:,0,0,0]),x[0,range(2),0])) # y[0,0] would equal to this with Theano backend\n<\/code><\/pre>\n"}]}
+{"QuestionId":52861413,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-17T18:28:11.573","AcceptedAnswerId":null,"OwnerUserId":10520111.0,"Title":"Tensorflow CNN test data splitting and array sizing problems","Body":"<p>I've tried to figure things out myself and not fallback to actually creating an account here but as a self-taught beginner I've reached a wall with this code.<\/p>\n\n<p>I'm having two major issues besides optimizing the net architecture when everything is working:<\/p>\n\n<ul>\n<li><p>Everytime I've tried to create a new dataset for a test batch I've ran into 'xTensor is not a Tensor' error and could run a session through it, unlike with the iterator which works just fine. I'm loading custom data with dir names as labels with no manually created train and test directories. I'm probably missing a proper method for tf.<\/p><\/li>\n<li><p>I can't work around the current first error I get which is:\n'ValueError: Cannot feed value of shape (100,) for Tensor 'Placeholder_1:0', which has shape '(?, 1)' while feed_dict {y=batch_y}. I've tried some of solutions posted on SO but couldn't get it to work.<\/p><\/li>\n<\/ul>\n\n<p>I'm pasting the whole thing, ########### are the problem triggering zones at the very bottom in the session.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport os\n\n# load custom imageset directory\ndata_path = r\"..\\datasets\\images\\flowers\"\n\n# setup hypervariables for labels and images format\nn_classes = 5\nimg_width = 64\nimg_length = 64\nchannels = 3\n\n# setup hypervariables for network\nlearning_rate = 0.0001\nepochs = 2\nbatch_size = 100\ndrop_rate = 0.6\n\nimagepaths = list()\nlabels = list()\nlabel = 0\nclasses = sorted(os.walk(data_path).__next__()[1])\n# List each sub-directory (the classes)\nfor c in classes:\n    c_dir = os.path.join(data_path, c)\n    walk = os.walk(c_dir).__next__()\n\n    # Add each image to the training set\n    for sample in walk[2]:\n        imagepaths.append(os.path.join(c_dir, sample))\n        labels.append(label)\n    label += 1\n\ntotal_input = len(labels)\n# Convert to Tensor\nimagepaths = tf.convert_to_tensor(imagepaths, dtype=tf.string)\nlabels = tf.convert_to_tensor(labels, dtype=tf.int32)\n# Build a TF Queue, shuffle data\ndataset = tf.data.Dataset.from_tensor_slices((imagepaths, labels))\n\n# read, decode, resize and normalize images on RGB range\ndef parse(imagepath, label):\n    image = tf.read_file(imagepath)\n    image = tf.image.decode_jpeg(image, channels=channels)\n    image = tf.image.resize_images(image, [img_length, img_width])\n    image = image * 1.0\/255\n    return image, label\n\ndataset = dataset.map(parse)\ndataset = dataset.shuffle(buffer_size=batch_size*10)\ndataset = dataset.batch(batch_size)\niterator = dataset.make_one_shot_iterator()\nnext_batch = iterator.get_next()\n\n# hypervariables for layers' output size\nK = 16\nL = 32\nM = 200\n\nx = tf.placeholder(tf.float32, [None, 4326])\nx_shaped = tf.reshape(x, [-1, img_length, img_width, 3])\ny = tf.placeholder(tf.float32, [None, 1])\n\n# weight, bias with stride size and activation method after convolution for layer 1\nW1 = tf.Variable(tf.truncated_normal([5, 5, 3, K], stddev=0.03))\nb1 = tf.Variable(tf.truncated_normal([K], stddev=0.01))\nstride = 1\ny1 = tf.nn.relu(tf.nn.conv2d(x_shaped, W1, strides=[1, stride, stride, 1], padding='SAME') + b1)\n\n# weight, bias with stride size and activation method after convolution for layer 2\nW2 = tf.Variable(tf.truncated_normal([5, 5, K, L], stddev=0.03))\nb2 = tf.Variable(tf.truncated_normal([L], stddev=0.01))\nstride = 2  # output is 14x14\ny2 = tf.nn.relu(tf.nn.conv2d(y1, W2, strides=[1, stride, stride, 1], padding='SAME') + b2)\n\nyflat = tf.reshape(y2, [-1, 7 * 7 * L])\n\nW3 = tf.Variable(tf.truncated_normal([7 * 7 * L, M], stddev=0.1))\nb3 = tf.Variable(tf.truncated_normal([M], stddev=0.01))\ny3 = tf.nn.relu(tf.matmul(yflat, W3) + b3)\n\nW4 = tf.Variable(tf.truncated_normal([M, 10], stddev=0.1))\nb4 = tf.Variable(tf.truncated_normal([10], stddev=0.01))\nylogits = tf.matmul(y3, W4) + b4\ny_ = tf.nn.softmax(ylogits)\n\n# add cross entropy for back prop\ncross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=ylogits, labels=y_))\n\n# add an optimiser for back prop\noptimiser = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cross_entropy)\n\n# define an accuracy assessment operation\ncorrect_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\nwith tf.Session() as sess:\n\n    sess.run(tf.global_variables_initializer())\n    ########## temporary solution for test_x, test_y\n    test_x, test_y = sess.run(next_batch)\n    total_batch = int(total_input \/ batch_size)\n    # define the iterator for the network\n    for epoch in range(epochs):\n        avg_cost = 0\n        for i in range(total_batch):\n            batch_x, batch_y = sess.run(next_batch)\n            ########## ValueError: Cannot feed value of shape (100,) for Tensor 'Placeholder_1:0' -&gt; y: batch_y\n            _, c = sess.run([optimiser, cross_entropy], feed_dict={x_shaped: batch_x, y: batch_y}) \n            avg_cost += c \/ total_batch\n\n        test_acc = sess.run(accuracy,feed_dict={x: test_x, y: test_y})\n        print(\"Epoch:\", (epoch + 1), \"cost =\", \"{:.3f}\".format(avg_cost), \" test accuracy: {:.3f}\".format(test_acc))\n        summary = sess.run(merged, feed_dict={x: test_x, y: test_y})\n\n    print(\"\\nTraining complete!\")\n    print(sess.run(accuracy, feed_dict={x: test_x, y: test_y}))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52863219","CreationDate":"2018-10-17T20:37:25.980","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>are you sure that this part:<\/p>\n\n<pre><code>_, c = sess.run([optimiser, cross_entropy], feed_dict={x_shaped: batch_x, y: batch_y}) \n<\/code><\/pre>\n\n<p>doesn't have to be:<\/p>\n\n<pre><code>_, c = sess.run([optimiser, cross_entropy], feed_dict={x: batch_x, y: batch_y}) \n<\/code><\/pre>\n\n<p>furthermore you've a batchsize of 100, the data is right of the array, the shape is not complete.\nWhat you have (dummy example)<\/p>\n\n<pre><code>np.zeros((100,)).shape\n&gt;&gt;&gt; (100,)\n<\/code><\/pre>\n\n<p>here 100 matches the '?' of the required shape: shape '(?, 1)', the one can be easily added, and it often occurs that numpy does not do this. See the following code:<\/p>\n\n<pre><code> np.expand_dims(np.zeros((100,)), axis=-1).shape\n &gt;&gt;&gt; (100, 1)\n<\/code><\/pre>\n\n<p>axis -1 stand for the last axis, you basically tell numpy to add a dimension in the end. This does not affect the data itself, but the shape of the array. So your code should be:<\/p>\n\n<pre><code>_, c = sess.run([optimiser, cross_entropy], feed_dict={x_shaped: batch_x, y:np.expand_dims(batch_y, axis=-1)}) \n<\/code><\/pre>\n"}]}
+{"QuestionId":52861549,"AnswerCount":2,"Tags":"<python><tensorflow><3d>","CreationDate":"2018-10-17T18:38:18.733","AcceptedAnswerId":null,"OwnerUserId":5449789.0,"Title":"3D Point Cloud Generation using Tensorflow and PointNet","Body":"<p>So I have a corpus of data that consist of a set of specific 3D point Clouds. I\u2019d like to use that data to train a NN so that it can create its own unique 3D point cloud data for mesh conversion later down the road.<\/p>\n\n<p>Here is what my research has revealed so far: that there is a library called PointNet:  <a href=\"https:\/\/github.com\/charlesq34\/pointnet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/charlesq34\/pointnet<\/a>\nIt\u2019s for deep learning on point sets or 3D point clouds and it uses Tensorflow. <\/p>\n\n<p>What I don\u2019t understand is, after training a NN, how do you use the NN to then create a 3D Point Cloud based on it\u2019s training?<\/p>\n","answers":[{"AnswerId":"56507355","CreationDate":"2019-06-08T14:38:08.100","ParentId":null,"OwnerUserId":"1906456","Title":null,"Body":"<p>This model is for identifying things rather than creating things.  To generate 3D objects, you need to implement the conv2d_transpose by yourself.  See how the point_cluods is converted to a 2D image here: <a href=\"https:\/\/github.com\/charlesq34\/pointnet\/blob\/master\/models\/pointnet_cls.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/charlesq34\/pointnet\/blob\/master\/models\/pointnet_cls.py<\/a>.<\/p>\n\n<p>You can kind of do that reverse to get your point cloud back from the out of the conv2d_transpose.<\/p>\n\n<p>The idea is here.  It's not that easy.  I am also searching for some codes available online.<\/p>\n"},{"AnswerId":"54086921","CreationDate":"2019-01-08T07:18:25.700","ParentId":null,"OwnerUserId":"242388","Title":null,"Body":"<p>The input to the network is Nx3 matrix, where N is number of the points. \n<a href=\"https:\/\/i.stack.imgur.com\/4znX7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4znX7.png\" alt=\"Pointnet architecture\"><\/a><\/p>\n\n<p>As far as I understood this method has 3 different use cases. One is classification: given point cloud of a single object estimate the class of the given object. The second is to make part segmentation: for each point in the input mesh model decide the point's \"label\". And the third case given the whole room (indoor) segment it by objects (it can be trained to segment such objects like chair etc.). <\/p>\n\n<p>Given the image here you can see that the input is Nx3 (but in the paper they use Nx9: 3 xyz coordinates, then rgb coordinated (if the rgb image is also available) and then normalized xyz coordinates (in range [0,1]).<\/p>\n\n<p>The output in the third case is a matrix nxm where m is number of classes. <\/p>\n\n<p>Im not sure what is your aim, but it seems like you want to create another point cloud given the original point cloud. But this algorithm is mainly for segmentation or classification.<\/p>\n"}]}
+{"QuestionId":52861686,"AnswerCount":1,"Tags":"<tensorflow><toco>","CreationDate":"2018-10-17T18:47:43.190","AcceptedAnswerId":null,"OwnerUserId":1871689.0,"Title":"tensorflow toco command for .tflite","Body":"<p>I am following the steps at <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets-2-tflite\/#0\" rel=\"nofollow noreferrer\">TFLite Android<\/a><\/p>\n\n<p>I am not able to use the toco command and so cannot generate: \"optimized_graph.lite\"<\/p>\n\n<p>Here is the command that I use :<\/p>\n\n<pre><code>    IMAGE_SIZE=224\ntoco \\\n  --graph_def_file=tf_files\/retrained_graph.pb \\\n  --output_file=tf_files\/optimized_graph.lite \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --output_format=TFLITE \\\n  --input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \\\n  --input_array=input \\\n  --output_array=final_result \\\n  --inference_type=FLOAT \\\n  --input_data_type=FLOAT\n<\/code><\/pre>\n\n<p>And the error is: <\/p>\n\n<p>F tensorflow\/contrib\/lite\/toco\/toco.cc:46] Check failed: parsed_toco_flags.input_file.specified() Missing required flag: input_file<\/p>\n\n<p>I have followed the instructions carefully, on Mac OS, and tensorflow version is 1.7 <\/p>\n","answers":[{"AnswerId":"52868009","CreationDate":"2018-10-18T06:15:28.250","ParentId":null,"OwnerUserId":"1709287","Title":null,"Body":"<p>When I ran <\/p>\n\n<pre><code>toco --help\n<\/code><\/pre>\n\n<p>I saw no <code>--graph_def_file<\/code> in the options of this command. \nInstead I found <code>--input_file<\/code>, and its meaning is\n<em>Input file (model of any supported format). For Protobuf formats, both text and binary are supported regardless of file extension.<\/em><\/p>\n\n<p>After I replaced <code>--graph_def_file<\/code> with <code>--input_file<\/code>, the command worked and I got an optimized_graph.lite file.<\/p>\n\n<p>Below is the complete command.<\/p>\n\n<pre><code>toco \\\n  --input_file=tf_files\/retrained_graph.pb \\\n  --output_file=tf_files\/optimized_graph.lite \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --output_format=TFLITE \\\n  --input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \\\n  --input_array=input \\\n  --output_array=final_result \\\n  --inference_type=FLOAT \\\n  --input_data_type=FLOAT\n<\/code><\/pre>\n"}]}
+{"QuestionId":52862358,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><machine-learning><numpy-ndarray>","CreationDate":"2018-10-17T19:34:09.433","AcceptedAnswerId":null,"OwnerUserId":7775784.0,"Title":"ValueError: all input arrays must have the same shape (Machine Learning)","Body":"<p>I am having a problem with one of my Machine Learning homework assignments. We are asked to use machine learning in order to estimate the age of an individual given an image. We are given a dataset of 28360 train images and 7090 test images. <\/p>\n\n<p>The issue I'm having is in snippet #3 of my code that can be found <a href=\"https:\/\/pastebin.com\/HVL6SqqJ\" rel=\"nofollow noreferrer\"><strong>HERE :)<\/strong><\/a>\nIn particular, I'm getting <code>ValueError: all input arrays must have the same shape<\/code>. Like I mentioned above, my train data set has a size of 28360. After running the program, from <code>temp.append(img)<\/code> i can the place <code>print(len(temp))<\/code> which gives me again, 28360. Maybe I don't have a comprehensive understanding of what exactly <code>np.stack(temp)<\/code> is doing or how it works, but the sizes of my initial and final array seem to be the same so why the issue.?<\/p>\n\n<p>Asked around and am being told that the issue isn't necessarily with the length of my <code>temp<\/code> list, but rather that I may be running into an image which isn't of the same size as the others. That one or more of the contained resized images (which are arrays) have a different shape, maybe because the resize failed or the image is missing or something. If this is the case, how would finding which are different and why in a dataset of 28360 images so I could fix them or drop them.? Is this really the issue, or may it be something else.?<\/p>\n\n<p>Anyone, please help. What am I doing wrong? Any comments\/suggestions would be greatly appreciated. Thank you in advance. Cheers and good day.! :D<\/p>\n","answers":[{"AnswerId":"52862841","CreationDate":"2018-10-17T20:08:31.123","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>YOu know the problem, some images are being a pain, you'll have to debug I guess. Try to preform np.stack() everytime and catch where it goes wrong. The shape of all images will be printed when the stack no longer works, this might give you an idea what is going wrong. Can't help further since I can't reproduce it. np.stack concatenates all images on top of eachother, but therefore every image.shape has to be the same. <\/p>\n\n<pre><code>for image_name in train.Id:\n    img_path = os.path.join(data_dir, 'train', img_name)\n    img = imageio.imread(img_path)\n    img = skimage.transform.resize(img, (32, 32), mode='constant')\n    img = img.astype('float32') # this will help us in later stage\n    temp.append(img)\n    try:\n        train_x = np.stack(temp)\n    except ValueError:\n        [print(im.shape) for im in temp]\n        break\n<\/code><\/pre>\n\n<p>To give you a better feeling with np.stack, consider the following code:<\/p>\n\n<pre><code>tmp = [np.zeros((400,400,3)), np.zeros((400,400,3))]\nprint(np.stack(tmp).shape)\n&gt;&gt;&gt; (2, 400, 400, 3)\n\ntmp = [np.zeros((400,400,3)), np.zeros((400,400))]\nprint(np.stack(tmp).shape)\n&gt;&gt;&gt; ValueError: all input arrays must have the same shape\n<\/code><\/pre>\n\n<p>In the first example I have a list with two 3D arrays, with exactly the same shape. Numpy stacks them together and creates a new dimension, often called the batch size. In the second one I have a 3D and a 2D array in the list, and when you try to stack two arrays with a different shape (400, 400, 3) vs (400, 400) you get the ValueError you mentioned.<\/p>\n\n<p>I also took the liberty to look at the skimage docs right <a href=\"http:\/\/scikit-image.org\/docs\/dev\/api\/skimage.io.html#skimage.io.imread\" rel=\"nofollow noreferrer\">here<\/a><br>\nAt the return you can see the following:<\/p>\n\n<blockquote>\n  <p>img_array : ndarray\n  The different color bands\/channels are stored in the third dimension, such that a gray-image is MxN, an RGB-image MxNx3 and an RGBA-image MxNx4.<\/p>\n<\/blockquote>\n\n<p>I've the feeling some of your images are read as an MxNx4 array instead of MxNx3<\/p>\n\n<p>Hope I helped a bit, feel free to ask me more. <\/p>\n\n<p>EDIT<\/p>\n\n<pre><code>for image_name in train.Id:\n    img_path = os.path.join(data_dir, 'train', img_name)\n    img = imageio.imread(img_path)\n    img = skimage.transform.resize(img, (32, 32), mode='constant')\n    img = img.astype('float32') # this will help us in later stage\n    if len(img.shape) == 3 and img.shape[-1] == 3  # only append if 3D and last dimension is a 3, standing for RGB\n        temp.append(img)\ntrain_x = np.stack(temp)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52864399,"AnswerCount":1,"Tags":"<tensorflow><opencl><amd-gpu><sycl>","CreationDate":"2018-10-17T22:18:35.620","AcceptedAnswerId":null,"OwnerUserId":2510839.0,"Title":"Multiple AMD GPUs with Tensorflow and OpenCL on Ubuntu 16.04","Body":"<p>After a lot of struggles:<\/p>\n\n<ol>\n<li><p><strong>Built successfully Tensorflow with OpenCL<\/strong> on fresh <strong>Ubuntu 16.04<\/strong> with <strong>amdgpu 17.50<\/strong>.<\/p><\/li>\n<li><p>Have <strong>5 identical GPUs (rx580) installed<\/strong> and all of them are reported by clinfo and computecpp_info as expected.<\/p><\/li>\n<li><p>Running the MNIST convnet example, <strong>TF works but emplyes just GPU0 without seeing other GPUs<\/strong>.<\/p><\/li>\n<\/ol>\n\n<p>There are <strong>no errors reported<\/strong> in dmesg about the card, they seem all to be ready at the lowest layer, don't know why <strong>SYCL seems to be ignoring some cards<\/strong>.<\/p>\n\n<p>Here is <strong>computecpp_info<\/strong> output:<\/p>\n\n<pre><code>********************************************************************************\n\nComputeCpp Info (CE 1.0.1)\n\nSYCL 1.2.1 revision 3\n\n********************************************************************************\n\nToolchain information:\n\nGLIBC version: 2.23\nGLIBCXX: 20160609\nThis version of libstdc++ is supported.\n\n********************************************************************************\n\n\nDevice Info:\n\nDiscovered 5 devices matching:\n  platform  : &lt;any&gt;\n  device type : &lt;any&gt;\n\n--------------------------------------------------------------------------------\nDevice 0:\n\n  Device is supported                   : UNTESTED - Vendor not tested on this OS\n  CL_DEVICE_NAME                        : Ellesmere\n  CL_DEVICE_VENDOR                      : Advanced Micro Devices, Inc.\n  CL_DRIVER_VERSION                     : 2527.3\n  CL_DEVICE_TYPE                        : CL_DEVICE_TYPE_GPU\n--------------------------------------------------------------------------------\nDevice 1:\n\n  Device is supported                   : UNTESTED - Vendor not tested on this OS\n  CL_DEVICE_NAME                        : Ellesmere\n  CL_DEVICE_VENDOR                      : Advanced Micro Devices, Inc.\n  CL_DRIVER_VERSION                     : 2527.3\n  CL_DEVICE_TYPE                        : CL_DEVICE_TYPE_GPU\n--------------------------------------------------------------------------------\nDevice 2:\n\n  Device is supported                   : UNTESTED - Vendor not tested on this OS\n  CL_DEVICE_NAME                        : Ellesmere\n  CL_DEVICE_VENDOR                      : Advanced Micro Devices, Inc.\n  CL_DRIVER_VERSION                     : 2527.3\n  CL_DEVICE_TYPE                        : CL_DEVICE_TYPE_GPU\n--------------------------------------------------------------------------------\nDevice 3:\n\n  Device is supported                   : UNTESTED - Vendor not tested on this OS\n  CL_DEVICE_NAME                        : Ellesmere\n  CL_DEVICE_VENDOR                      : Advanced Micro Devices, Inc.\n  CL_DRIVER_VERSION                     : 2527.3\n  CL_DEVICE_TYPE                        : CL_DEVICE_TYPE_GPU\n--------------------------------------------------------------------------------\nDevice 4:\n\n  Device is supported                   : UNTESTED - Vendor not tested on this OS\n  CL_DEVICE_NAME                        : Ellesmere\n  CL_DEVICE_VENDOR                      : Advanced Micro Devices, Inc.\n  CL_DRIVER_VERSION                     : 2527.3\n  CL_DEVICE_TYPE                        : CL_DEVICE_TYPE_GPU\n\nIf you encounter problems when using any of these OpenCL devices, please consult\nthis website for known issues:\nhttps:\/\/computecpp.codeplay.com\/releases\/v1.0.1\/platform-support-notes\n\n********************************************************************************\n<\/code><\/pre>\n\n<p>Here the list from <strong>tensorflow<\/strong>:<\/p>\n\n<pre><code>$ python3 list_gpus.py\n2018-10-17 23:52:44.268968: I .\/tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2018-10-17 23:52:44.385308: I .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:70] Found following OpenCL devices:\n2018-10-17 23:52:44.385342: I .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:72] id: 0, type: GPU, name: Ellesmere, vendor: Advanced Micro Devices, Inc., profile: FULL_PROFILE\n[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 5429869323017416982\n, name: \"\/device:SYCL:0\"\ndevice_type: \"SYCL\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 7347791393919061653\nphysical_device_desc: \"id: 0, type: GPU, name: Ellesmere, vendor: Advanced Micro Devices, Inc., profile: FULL_PROFILE\"\n]\n<\/code><\/pre>\n\n<p><strong>EDIT: After a reboot<\/strong><\/p>\n\n<p>I don't really do know if these warnings are relevant because they go away after the first run.<\/p>\n\n<pre><code>$ python3 list_gpus.py\n2018-10-18 00:47:13.943021: I .\/tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2018-10-18 00:47:13.952909: W .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:45] No OpenCL accelerator nor GPU found that is supported by ComputeCpp\/triSYCL trying OpenCL CPU\n2018-10-18 00:47:13.952930: W .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:52] No OpenCL CPU found that is supported by ComputeCpp\/triSYCL, checking for host sycl device\n2018-10-18 00:47:13.952936: W .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:59] Found SYCL host device\n2018-10-18 00:47:13.953004: I .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:70] Found following OpenCL devices:\n2018-10-18 00:47:13.953014: I .\/tensorflow\/core\/common_runtime\/sycl\/sycl_device.h:72] id: 0, type: Host, name: Host Device, vendor: Codeplay Software Ltd., profile: FULL_PROFILE\n<\/code><\/pre>\n\n<p><strong>EDIT: dmesg details<\/strong><\/p>\n\n<pre><code>[    0.000000] Linux version 4.15.0-36-generic (buildd@lcy01-amd64-017) (gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.10)) #39~16.04.1-Ubuntu SMP Tue Sep 25 08:59:23 UTC 2018 (Ubuntu 4.15.0-36.39~16.04.1-generic 4.15.18)\n[    0.688885] pcie_mp2_amd: AMD(R) PCI-E MP2 Communication Driver Version: 1.0\n[    1.143085] [drm] amdgpu kernel modesetting enabled.\n[    1.173931] amdgpu 0000:03:00.0: enabling device (0000 -&gt; 0003)\n[    1.564757] amdgpu 0000:03:00.0: vgaarb: changed VGA decodes: olddecodes=io+mem,decodes=none:owns=none\n[    2.280211] amdgpu 0000:03:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[    2.280212] amdgpu 0000:03:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[    2.280322] [drm] amdgpu: 4096M of VRAM memory ready\n[    2.280323] [drm] amdgpu: 4096M of GTT memory ready.\n[    2.280427] amdgpu 0000:03:00.0: amdgpu: using MSI.\n[    2.280439] [drm] amdgpu: irq initialized.\n[    2.280452] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[    2.280690] amdgpu 0000:03:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[    2.280758] amdgpu 0000:03:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[    2.280784] amdgpu 0000:03:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[    2.280842] amdgpu 0000:03:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[    2.280903] amdgpu 0000:03:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[    2.280965] amdgpu 0000:03:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[    2.280985] amdgpu 0000:03:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[    2.281001] amdgpu 0000:03:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[    2.281015] amdgpu 0000:03:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[    2.281028] amdgpu 0000:03:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[    2.281332] amdgpu 0000:03:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[    2.281348] amdgpu 0000:03:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[    2.285039] amdgpu 0000:03:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[    2.285056] amdgpu 0000:03:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[    2.285069] amdgpu 0000:03:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[    2.285578] amdgpu 0000:03:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[    2.285594] amdgpu 0000:03:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[    2.980155] amdgpu 0000:03:00.0: kfd not supported on this ASIC\n[    2.980163] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:03:00.0 on minor 0\n[    2.980215] amdgpu 0000:06:00.0: enabling device (0000 -&gt; 0003)\n[    4.068205] amdgpu 0000:06:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[    4.068206] amdgpu 0000:06:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[    4.068220] [drm] amdgpu: 4096M of VRAM memory ready\n[    4.068221] [drm] amdgpu: 4096M of GTT memory ready.\n[    4.068331] amdgpu 0000:06:00.0: amdgpu: using MSI.\n[    4.068344] [drm] amdgpu: irq initialized.\n[    4.068357] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[    4.068444] amdgpu 0000:06:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[    4.068509] amdgpu 0000:06:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[    4.068571] amdgpu 0000:06:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[    4.068639] amdgpu 0000:06:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[    4.068665] amdgpu 0000:06:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[    4.068718] amdgpu 0000:06:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[    4.068740] amdgpu 0000:06:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[    4.068759] amdgpu 0000:06:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[    4.068774] amdgpu 0000:06:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[    4.068787] amdgpu 0000:06:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[    4.069074] amdgpu 0000:06:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[    4.069094] amdgpu 0000:06:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[    4.072854] amdgpu 0000:06:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[    4.072868] amdgpu 0000:06:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[    4.072881] amdgpu 0000:06:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[    4.073362] amdgpu 0000:06:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[    4.073376] amdgpu 0000:06:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[    4.771466] amdgpu 0000:06:00.0: kfd not supported on this ASIC\n[    4.771476] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:06:00.0 on minor 2\n[    4.771515] amdgpu 0000:07:00.0: enabling device (0000 -&gt; 0003)\n[    5.856168] amdgpu 0000:07:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[    5.856169] amdgpu 0000:07:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[    5.856178] [drm] amdgpu: 4096M of VRAM memory ready\n[    5.856179] [drm] amdgpu: 4096M of GTT memory ready.\n[    5.856284] amdgpu 0000:07:00.0: amdgpu: using MSI.\n[    5.856297] [drm] amdgpu: irq initialized.\n[    5.856311] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[    5.856402] amdgpu 0000:07:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[    5.856441] amdgpu 0000:07:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[    5.856464] amdgpu 0000:07:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[    5.856541] amdgpu 0000:07:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[    5.856569] amdgpu 0000:07:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[    5.856641] amdgpu 0000:07:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[    5.856668] amdgpu 0000:07:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[    5.856690] amdgpu 0000:07:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[    5.856707] amdgpu 0000:07:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[    5.856722] amdgpu 0000:07:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[    5.857007] amdgpu 0000:07:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[    5.857027] amdgpu 0000:07:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[    5.860789] amdgpu 0000:07:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[    5.860803] amdgpu 0000:07:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[    5.860817] amdgpu 0000:07:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[    5.861298] amdgpu 0000:07:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[    5.861313] amdgpu 0000:07:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[    6.563837] amdgpu 0000:07:00.0: kfd not supported on this ASIC\n[    6.563845] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:07:00.0 on minor 3\n[    6.563887] amdgpu 0000:08:00.0: enabling device (0000 -&gt; 0003)\n[    7.648177] amdgpu 0000:08:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[    7.648178] amdgpu 0000:08:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[    7.648188] [drm] amdgpu: 4096M of VRAM memory ready\n[    7.648188] [drm] amdgpu: 4096M of GTT memory ready.\n[    7.648292] amdgpu 0000:08:00.0: amdgpu: using MSI.\n[    7.648306] [drm] amdgpu: irq initialized.\n[    7.648322] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[    7.648406] amdgpu 0000:08:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[    7.648470] amdgpu 0000:08:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[    7.648530] amdgpu 0000:08:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[    7.648593] amdgpu 0000:08:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[    7.648649] amdgpu 0000:08:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[    7.648707] amdgpu 0000:08:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[    7.648733] amdgpu 0000:08:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[    7.648751] amdgpu 0000:08:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[    7.648769] amdgpu 0000:08:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[    7.648782] amdgpu 0000:08:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[    7.649069] amdgpu 0000:08:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[    7.649087] amdgpu 0000:08:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[    7.652849] amdgpu 0000:08:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[    7.652862] amdgpu 0000:08:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[    7.652874] amdgpu 0000:08:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[    7.653353] amdgpu 0000:08:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[    7.653366] amdgpu 0000:08:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[    8.355909] amdgpu 0000:08:00.0: kfd not supported on this ASIC\n[    8.355916] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:08:00.0 on minor 4\n[    8.355957] amdgpu 0000:09:00.0: enabling device (0000 -&gt; 0003)\n[    9.440257] amdgpu 0000:09:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[    9.440258] amdgpu 0000:09:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[    9.440268] [drm] amdgpu: 4096M of VRAM memory ready\n[    9.440268] [drm] amdgpu: 4096M of GTT memory ready.\n[    9.440376] amdgpu 0000:09:00.0: amdgpu: using MSI.\n[    9.440390] [drm] amdgpu: irq initialized.\n[    9.440406] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[    9.440499] amdgpu 0000:09:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[    9.440563] amdgpu 0000:09:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[    9.440625] amdgpu 0000:09:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[    9.440690] amdgpu 0000:09:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[    9.440753] amdgpu 0000:09:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[    9.440808] amdgpu 0000:09:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[    9.440831] amdgpu 0000:09:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[    9.440849] amdgpu 0000:09:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[    9.440865] amdgpu 0000:09:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[    9.440880] amdgpu 0000:09:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[    9.441167] amdgpu 0000:09:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[    9.441184] amdgpu 0000:09:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[    9.444946] amdgpu 0000:09:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[    9.444964] amdgpu 0000:09:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[    9.444976] amdgpu 0000:09:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[    9.445456] amdgpu 0000:09:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[    9.445469] amdgpu 0000:09:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[   10.147558] amdgpu 0000:09:00.0: kfd not supported on this ASIC\n[   10.147564] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:09:00.0 on minor 5\n[   10.147606] amdgpu 0000:0a:00.0: enabling device (0000 -&gt; 0003)\n[   11.232197] amdgpu 0000:0a:00.0: VRAM: 4096M 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)\n[   11.232198] amdgpu 0000:0a:00.0: GTT: 256M 0x0000000000000000 - 0x000000000FFFFFFF\n[   11.232207] [drm] amdgpu: 4096M of VRAM memory ready\n[   11.232207] [drm] amdgpu: 4096M of GTT memory ready.\n[   11.232309] amdgpu 0000:0a:00.0: amdgpu: using MSI.\n[   11.232322] [drm] amdgpu: irq initialized.\n[   11.232337] amdgpu: [powerplay] amdgpu: powerplay sw initialized\n[   11.232427] amdgpu 0000:0a:00.0: fence driver on ring 0 use gpu addr 0x0000000000400040, cpu addr 0x        (ptrval)\n[   11.232488] amdgpu 0000:0a:00.0: fence driver on ring 1 use gpu addr 0x00000000004000c0, cpu addr 0x        (ptrval)\n[   11.232551] amdgpu 0000:0a:00.0: fence driver on ring 2 use gpu addr 0x0000000000400140, cpu addr 0x        (ptrval)\n[   11.232615] amdgpu 0000:0a:00.0: fence driver on ring 3 use gpu addr 0x00000000004001c0, cpu addr 0x        (ptrval)\n[   11.232675] amdgpu 0000:0a:00.0: fence driver on ring 4 use gpu addr 0x0000000000400240, cpu addr 0x        (ptrval)\n[   11.232699] amdgpu 0000:0a:00.0: fence driver on ring 5 use gpu addr 0x00000000004002c0, cpu addr 0x        (ptrval)\n[   11.232717] amdgpu 0000:0a:00.0: fence driver on ring 6 use gpu addr 0x0000000000400340, cpu addr 0x        (ptrval)\n[   11.232735] amdgpu 0000:0a:00.0: fence driver on ring 7 use gpu addr 0x00000000004003c0, cpu addr 0x        (ptrval)\n[   11.232749] amdgpu 0000:0a:00.0: fence driver on ring 8 use gpu addr 0x0000000000400440, cpu addr 0x        (ptrval)\n[   11.232763] amdgpu 0000:0a:00.0: fence driver on ring 9 use gpu addr 0x00000000004004e0, cpu addr 0x        (ptrval)\n[   11.233048] amdgpu 0000:0a:00.0: fence driver on ring 10 use gpu addr 0x0000000000400560, cpu addr 0x        (ptrval)\n[   11.233067] amdgpu 0000:0a:00.0: fence driver on ring 11 use gpu addr 0x00000000004005e0, cpu addr 0x        (ptrval)\n[   11.236830] amdgpu 0000:0a:00.0: fence driver on ring 12 use gpu addr 0x000000f4001e6420, cpu addr 0x        (ptrval)\n[   11.236848] amdgpu 0000:0a:00.0: fence driver on ring 13 use gpu addr 0x00000000004006e0, cpu addr 0x        (ptrval)\n[   11.236860] amdgpu 0000:0a:00.0: fence driver on ring 14 use gpu addr 0x0000000000400760, cpu addr 0x        (ptrval)\n[   11.237341] amdgpu 0000:0a:00.0: fence driver on ring 15 use gpu addr 0x00000000004007e0, cpu addr 0x        (ptrval)\n[   11.237355] amdgpu 0000:0a:00.0: fence driver on ring 16 use gpu addr 0x0000000000400860, cpu addr 0x        (ptrval)\n[   11.939330] amdgpu 0000:0a:00.0: kfd not supported on this ASIC\n[   11.939336] [drm] Initialized amdgpu 3.23.0 20150101 for 0000:0a:00.0 on minor 6\n<\/code><\/pre>\n\n<p><strong>EDIT: It is not related to any specific card, just the first available in bus order.<\/strong><\/p>\n\n<p>I tried disconnecting some cards and after all the tests it seems clear that SYCL always lists just the first GPU, no matter which one, always the minimum available bus number. <\/p>\n\n<p>This also confirms that there are no differences among the cards and that all of them can be used (at least individually), so the OS I think is fine and i would guess the problem is in SYCL.<\/p>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"52872795","CreationDate":"2018-10-18T11:18:22.297","ParentId":null,"OwnerUserId":"2510839","Title":null,"Body":"<p>As today's date <strong>multiple GPUs with Tensorflow and OpenCL are currenly unsupported<\/strong> even if not clearly stated in the documentation.<\/p>\n\n<p>You can track the details of the problem here, i opened an issue on Github: <a href=\"https:\/\/github.com\/codeplaysoftware\/tensorflow\/issues\/16\" rel=\"nofollow noreferrer\">https:\/\/github.com\/codeplaysoftware\/tensorflow\/issues\/16<\/a><\/p>\n\n<p>I'll update this answer if something changes but as the developer said this is not a priority for them!<\/p>\n"}]}
+{"QuestionId":52864435,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-17T22:22:58.417","AcceptedAnswerId":null,"OwnerUserId":5991102.0,"Title":"Measuring GPU memory usage with TensorFlow profiler","Body":"<p>Is there a way to properly measure GPU usage of tf.Estimator model using TensorFlow profiler? I've followed the documentation:<\/p>\n\n<pre><code>g = tf.Graph()\nsess = tf.Session(graph=g)\nrun_meta = tf.RunMetadata()\n\ntime_and_memory_args = tf.profiler.ProfileOptionBuilder.time_and_memory()\n\nwith g.as_default():\n    data_shape = [BATCH_SIZE] + [3, 224, 224]\n    in_plh = tf.placeholder(tf.float32, data_shape)\n\n    model = some_model(in_plh, args=model_args, training=True)\n    images = np.random.rand(BATCH_SIZE, 3, 224, 224)\n\n    sess.run(tf.global_variables_initializer())\n    sess.run(model, feed_dict={in_plh: images})\n\n    time_and_memory = tf.profiler.profile(g, run_meta=run_meta, cmd='op',\n                                          options=time_and_memory_args) \n\n    if time_and_memory is not None:\n        print('Total requested bytes:', time_and_memory.total_requested_bytes)\n<\/code><\/pre>\n\n<p>But the printed result is always 0.<\/p>\n","answers":[]}
+{"QuestionId":52864641,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-17T22:47:56.267","AcceptedAnswerId":null,"OwnerUserId":10521067.0,"Title":"AttributeError: module '_pywrap_tensorflow_internal' has no attribute 'TFE_DEVICE_PLACEMENT_EXPLICIT_swigconstant'","Body":"<p>I'm using Python 3.5 64 bit, Cuda 9.2, TensorFlow 1.11, Keras 2.2.4<\/p>\n\n<p>I'm getting this error:<\/p>\n\n<blockquote>\n  <p>AttributeError: module '_pywrap_tensorflow_internal' has no attribute 'TFE_DEVICE_PLACEMENT_EXPLICIT_swigconstant'<\/p>\n<\/blockquote>\n\n<p>Uninstalling and reinstalling TensorFlow stopped the message for <code>import tensorflow<\/code>, but the problem persists for <code>import keras<\/code>.<\/p>\n\n<p>I have seen that <code>dumpbin.exe<\/code> has been useful for some people, but I've been unable to get it to work.<\/p>\n\n<p>The problems started this morning while trying to get Tensor Board to work. I was updating python modules with pip and I'm not sure what set it off.<\/p>\n","answers":[{"AnswerId":"53296077","CreationDate":"2018-11-14T08:43:31.047","ParentId":null,"OwnerUserId":"5072351","Title":null,"Body":"<p>I encountered the same issue and this has been answered <a href=\"https:\/\/stackoverflow.com\/q\/49218685\/5072351\">here<\/a>. Follow the steps to resolve this issue. You can also refer <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5747\" rel=\"nofollow noreferrer\">here<\/a> to see the solution.<\/p>\n"}]}
+{"QuestionId":52864686,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-17T22:54:13.543","AcceptedAnswerId":52881244.0,"OwnerUserId":10484131.0,"Title":"Tensorflow scan multiple matrix rows with offset","Body":"<h3>Question<\/h3>\n\n<p>I want to scan a matrix analogous to Tensorflow's <code>tf.scan()<\/code>, but using multiple rows at a time. So given a <code>[n, m]<\/code> matrix, I want to be able to iterate the <code>m<\/code> rows (with <code>n<\/code> elements) from <code>i + j<\/code> to <code>m<\/code> giving <code>m - j<\/code> slices of shape <code>[i - j, n]<\/code>.<\/p>\n\n<p>How can this be achieved?<\/p>\n\n<p>I know how <code>tf.scan<\/code> does something like this, returning the accumulated value of each iteration. But I don't think shifting the matrix as multiple inputs solves this, since the values that have an offset cannot be precomputed.<\/p>\n\n<h3>Example<\/h3>\n\n<p>To give an example for <code>n = 3<\/code> and <code>m = 5<\/code>, let's say I have a matrix that looks like the following:<\/p>\n\n<pre><code># [[1 0 0]\n#  [1 1 0]\n#  [0 0 0]   row 3\n#  [0 0 0]   row 4\n#  [0 0 0]]  row 5\n\nmatrix_shape = [5, 3]\nmatrix_idx = tf.constant([[0, 0], [1, 0], [1, 1]])\nmatrix = tf.scatter_nd(matrix_idx,\n                       tf.ones(tf.shape(matrix_idx)[0],\n                               dtype=tf.int32),\n                       matrix_shape)\n<\/code><\/pre>\n\n<p>I want to apply the following function from row 3 to row 5:<\/p>\n\n<pre><code># [[ 1  0  0]    \u250c a\n#  [ 1  1  0]    \u251c b\n#  [ 6  4  2]  &lt;\u2500\u2534 output \/ current line\n#  [16 12  6]\n#  [46 34 18]]\n\ndef compute(x):\n    a = x[0]\n    b = x[1]\n    return (a + b + 1) * 2\n<\/code><\/pre>\n\n<p>Does Tensorflow have a function specific to this problem?<\/p>\n","answers":[{"AnswerId":"52881244","CreationDate":"2018-10-18T19:31:53.107","ParentId":null,"OwnerUserId":"10484131","Title":null,"Body":"<p>The following code I wrote does exactly what I wanted.<\/p>\n\n<p>The important part here is the <code>return<\/code> of the function used by <code>tf.scan<\/code>, which not only gives back the current computation <code>c<\/code>, but also the row from the previous step <code>b<\/code>. It is therefore important to later cut off this excess from <code>computation<\/code> by only selecting the later tensor in this list with <code>[1]<\/code>.<\/p>\n\n<pre><code>#!\/usr\/bin\/env python3\nimport tensorflow as tf\n\ndef compute(x, _):\n    a = x[0]\n    b = x[1]\n    c = (a + b + 1) * 2\n    return (b, c)\n\nmatrix_shape = tf.constant([3, 3])\ninit_data = [[1, 0, 0], [1, 1, 0]]\ninitializer = (\n    tf.constant(init_data[0]),\n    tf.constant(init_data[1]),\n)\nmatrix = tf.zeros(matrix_shape, dtype=tf.int32)\ncomputation = tf.scan(compute, matrix, initializer)[1]\nresult = tf.concat((tf.constant(init_data), computation), axis=0)\n\nwith tf.Session() as sess:\n    sess.run(result)\n    print(result.eval())\n<\/code><\/pre>\n\n<p>Since I'm yet lacking experience: May this solution be bad for performance, because the function is returning a tuple and therefore not using Tensorflow's speed optimizations?<\/p>\n"}]}
+{"QuestionId":52864864,"AnswerCount":1,"Tags":"<python><tensorflow><placeholder>","CreationDate":"2018-10-17T23:18:30.023","AcceptedAnswerId":52868471.0,"OwnerUserId":4827407.0,"Title":"How to print weights and loss at the same time?","Body":"<p>I just started studying TensorFlow. Now I am confused about how to print weights and loss at the same time.<\/p>\n\n<p>In this code:<\/p>\n\n<pre><code>x_data=np.random.rand(100)\ny_data=x_data*0.1+0.2\nb=tf.Variable(0.)  #must be 0.  not 0\nk=tf.Variable(0.)\ny=k*x_data+b\nloss=tf.reduce_mean(tf.square(y_data-y))\noptimizer=tf.train.GradientDescentOptimizer(0.2)\ntrain=optimizer.minimize(loss)\ninit=tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    for step in range(501):\n        sess.run(train)\n        if step%20==0:\n            print(step,sess.run([k,b,loss]))\n<\/code><\/pre>\n\n<p>I can print the parameters k, b, and loss at the same time. The result looks like:<\/p>\n\n<pre><code>0 [0.056439728, 0.10114789, 0.014995859]\n20 [0.105686955, 0.19681107, 2.4290184e-06]\n40 [0.10358144, 0.19799174, 9.633536e-07]\n60 [0.10225546, 0.19873528, 3.8206556e-07]\n80 [0.10142042, 0.19920352, 1.515299e-07]\n......\n<\/code><\/pre>\n\n<p>But in another code I have tried:<\/p>\n\n<pre><code>x_data = np.linspace(-0.5, 0.5, 200)[:, np.newaxis]  \nnoise = np.random.normal(0, 0.02, x_data.shape)\ny_data = np.square(x_data) + noise\nx = tf.placeholder(tf.float32, [None, 1])\ny = tf.placeholder(tf.float32, [None, 1])\nWeights_L1 = tf.Variable(tf.random_normal([1, 10]))  # 1 row 10 column\nbiases_L1 = tf.Variable(tf.zeros([1, 10]))\nWx_plus_b_L1 = tf.matmul(x, Weights_L1) + biases_L1\nL1 = tf.nn.tanh(Wx_plus_b_L1)\nWeights_L2 = tf.Variable(tf.random_normal([10, 1]))  # 10 row 1 column\nbiases_L2 = tf.Variable(tf.zeros([1, 1]))\nWx_plus_b_L2 = tf.matmul(L1, Weights_L2) + biases_L2\nprediction = tf.nn.tanh(Wx_plus_b_L2)\nloss = tf.reduce_mean(tf.square(y - prediction))\ntrain_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for step in range(2000):\n        sess.run(train_step, feed_dict={x: x_data, y: y_data})\n        if step % 200 == 0:\n            print(step, sess.run([Weights_L1]))\n            print(step,sess.run([loss]))\n    prediction_value=sess.run(prediction,feed_dict={x:x_data})\n<\/code><\/pre>\n\n<p>Here, I want to use two prints to show weights and loss respectively. But I have the error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder_4' with dtype float and shape [?,1]\n [[Node: Placeholder_4 = Placeholder[dtype=DT_FLOAT, shape=[?,1], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>So, here is my question:<\/p>\n\n<p>How to print weights and loss at the same time? What if I want to print more parameters?<\/p>\n\n<p>Why do I have this error?<\/p>\n","answers":[{"AnswerId":"52868471","CreationDate":"2018-10-18T06:51:16.460","ParentId":null,"OwnerUserId":"2301712","Title":null,"Body":"<p>the <code>loss<\/code> depends on input placeholder <code>x<\/code> and <code>y<\/code>, so <code>print(step,sess.run([loss]))<\/code> should be modified as <code>print(step, sess.run([loss], feed_dict={x: x_data, y: y_data}))<\/code><\/p>\n\n<p>By the way, you can try <code>tf.summary<\/code> to visualize the weights during training, that's very intuitive.<\/p>\n"}]}
+{"QuestionId":52866049,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><keras><pytorch>","CreationDate":"2018-10-18T02:22:45.783","AcceptedAnswerId":null,"OwnerUserId":8292871.0,"Title":"Creating and train only specified weights in TensorFlow or PyTorch","Body":"<p>I am wondering if there is a way in TensorFlow, PyTorch or some other library to selectively connect neurons. I want to make a network with a very large number of neurons in each layer, but that has very few connections between layers. <\/p>\n\n<p>Note that I do not think this is a duplicate of this answer: <a href=\"https:\/\/stackoverflow.com\/questions\/38278965\/selectively-zero-weights-in-tensorflow\">Selectively zero weights in TensorFlow?<\/a>. I implemented a custom keras layer using essentially the same method that appears in that question - essentially by creating a dense layer where all but the specified weights are ignored in training and evaluation. This fulfills part of what I want to do by not training specified weights, and not using them for prediction. But, the problems is that I still waste memory saving the untrained weights, and I waste time calculating the gradients of the zeroed weights. What I would like is for the computation of the gradient matrices to involve only sparse matrices, so that I do not waste time and memory.<\/p>\n\n<p>Is there a way to selectively create and train weights without wasting memory? If my question is unclear or there is more information that it would be helpful for me to provide, please let me know. I would like to be helpful as a question-asker. <\/p>\n","answers":[{"AnswerId":"52887371","CreationDate":"2018-10-19T07:08:53.777","ParentId":null,"OwnerUserId":"8394915","Title":null,"Body":"<p>The usual, simple solution is to initialize your weight matrices to have zeros where there should be no connection. You store a mask of the location of these zeros, and set the weights at these positions to zero after each weight update. You need to do this as the gradient for zero weights may be nonzero, and this would introduce nonzero weights (i.e. connectios) where you don't want any.<\/p>\n\n<p>Pseudocode:<\/p>\n\n<pre><code># setup network\nweights = sparse_init()  # only nonzero for existing connections\nzero_mask = where(weights == 0)\n\n# train\nfor e in range(num_epochs):\n    train_operation()  # may lead to introduction of new connections\n    weights[zero_mask] = 0  # so we set them to zero again\n<\/code><\/pre>\n"},{"AnswerId":"54573050","CreationDate":"2019-02-07T12:10:06.880","ParentId":null,"OwnerUserId":"9665421","Title":null,"Body":"<p>Both tensorflow and pytorch support sparse tensors (<a href=\"https:\/\/pytorch.org\/docs\/stable\/sparse.html\" rel=\"nofollow noreferrer\">torch.sparse<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse\" rel=\"nofollow noreferrer\">tf.sparse<\/a>). <\/p>\n\n<p>My intuitive understanding would be that if you were willing to write your network using the respective low level APIs (e.g. actually implementing the forward-pass yourself), you could cast your weight matrices as sparse tensors. That would in turn result in sparse connectivity, since the weight matrix of layer [L] defines the connectivity between neurons of the previous layer [L-1] with neurons of layer [L]. <\/p>\n"}]}
+{"QuestionId":52866131,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2018-10-18T02:33:05.720","AcceptedAnswerId":null,"OwnerUserId":3422206.0,"Title":"Segfault triggered by multiple GPUs","Body":"<p>I am running a training script with <code>caffe<\/code> on a 8 GPU (1080Ti) server.<\/p>\n\n<p>If I train on 6 or fewer gpus (using <code>CUDA_VISIBLE_DEVICES<\/code>), everything is fine.<\/p>\n\n<p>(I set <code>export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5<\/code> and specify these GPUs in training script.)<\/p>\n\n<p>But if I train on 7 or 8 GPUs, I see this error at the start of training consistently:<\/p>\n\n<pre><code>Error: (unix time) try if you are using GNU date\nSIGSEGV (@0x70) received by PID 17206 (TID 0x7fc678ffd700) from PID 112; stack trace:\n@     0x7fc86186b4b0 (unknown)\n@     0x7fc861983f75 (unknown)\n@     0x7fc863c4b4c7 std::__cxx11::basic_string&lt;&gt;::_M_construct&lt;&gt;()\n@     0x7fc863c4c60b _ZN5caffe2db10LMDBCursor5valueB5cxx11Ev\n@     0x7fc863ace3e7 caffe::AnnotatedDataLayer&lt;&gt;::DataLayerSetUp()\n@     0x7fc863a6e4d5 caffe::BasePrefetchingDataLayer&lt;&gt;::LayerSetUp()\n@     0x7fc863cbf2b4 caffe::Net&lt;&gt;::Init()\n@     0x7fc863cc11ae caffe::Net&lt;&gt;::Net()\n@     0x7fc863bb9c9a caffe::Solver&lt;&gt;::InitTestNets()\n@     0x7fc863bbb84d caffe::Solver&lt;&gt;::Init()\n@     0x7fc863bbbb3f caffe::Solver&lt;&gt;::Solver()\n@     0x7fc863ba7d61 caffe::Creator_SGDSolver&lt;&gt;()\n@     0x7fc863ccc1c2 caffe::Worker&lt;&gt;::InternalThreadEntry()\n@     0x7fc863cf94c5 caffe::InternalThread::entry()\n@     0x7fc863cfa38e boost::detail::thread_data&lt;&gt;::run()\n@     0x7fc85350d5d5 (unknown)\n@     0x7fc83fee56ba start_thread\n@     0x7fc86193d41d clone\n@                0x0 (unknown)```\n<\/code><\/pre>\n\n<p>The <code>Error: (unix time) ...<\/code> at the start of the trace is apparently <a href=\"https:\/\/hpc.nih.gov\/development\/glog.html\" rel=\"nofollow noreferrer\">thrown by <code>glog<\/code><\/a>. <\/p>\n\n<p>It appears to be thrown when a general failure happens.<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1993\" rel=\"nofollow noreferrer\">This thread<\/a> show many different issues triggering <code>Error: (unix time)...<\/code> and similar trace.<\/p>\n\n<p>In the thread, it is noted that multiple GPUs may trigger this error. <\/p>\n\n<p>That is what appears to be the root cause in my case.<\/p>\n\n<p>Are there things I can further look into to understand what is happening?<\/p>\n","answers":[]}
+{"QuestionId":52866333,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-10-18T02:58:14.757","AcceptedAnswerId":null,"OwnerUserId":4497662.0,"Title":"PyTorch: new_ones vs ones","Body":"<p>In PyTorch what is the difference between <code>new_ones()<\/code> vs <code>ones()<\/code>. For example,<\/p>\n\n<pre><code>x2.new_ones(3,2, dtype=torch.double)\n<\/code><\/pre>\n\n<p>vs <\/p>\n\n<pre><code>torch.ones(3,2, dtype=torch.double)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52870290","CreationDate":"2018-10-18T08:48:24.010","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>For the sake of this answer, I am assuming that your <code>x2<\/code> is a previously defined <code>torch.Tensor<\/code>. If we then head over to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.new_ones\" rel=\"noreferrer\">PyTorch documentation<\/a>, we can read the following on <code>new_ones()<\/code>:<\/p>\n\n<blockquote>\n  <p>Returns a Tensor of size <code>size<\/code> filled with <code>1.<\/code> By default, the\n  returned Tensor has the same <code>torch.dtype<\/code> and <code>torch.device<\/code> as this\n  tensor.<\/p>\n<\/blockquote>\n\n<p>Whereas <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.ones\" rel=\"noreferrer\"><code>ones()<\/code><\/a><\/p>\n\n<blockquote>\n  <p>Returns a tensor filled with the scalar value 1, with the shape\n  defined by the variable argument sizes.<\/p>\n<\/blockquote>\n\n<p>So, essentially, <code>new_ones<\/code> allows you to quickly create a new <code>torch.Tensor<\/code> on the same device and data type as a <em>previously existing<\/em> tensor (with ones), whereas <code>ones()<\/code> serves the purpose of creating a <code>torch.Tensor<\/code> from scratch (filled with ones).<\/p>\n"},{"AnswerId":"53380973","CreationDate":"2018-11-19T18:55:24.477","ParentId":null,"OwnerUserId":"5720196","Title":null,"Body":"<blockquote>\n  <p>new_ones()<\/p>\n<\/blockquote>\n\n<pre><code># defining the tensor along with device to run on. (Assuming CUDA hardware is available)\n\nx = torch.rand(5, 3, device=\"cuda\")\n<\/code><\/pre>\n\n<p><code>new_ones()<\/code> works with existing tensor. <code>y<\/code> will inherit the <code>datatype<\/code> from <code>x<\/code> and it will run on same <code>device<\/code> as defined in <code>x<\/code><\/p>\n\n<pre><code>y = x.new_ones(2, 2)\nprint(y)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor([[1., 1.],\n        [1., 1.]], device='cuda:0')\n<\/code><\/pre>\n\n<blockquote>\n  <p>ones()<\/p>\n<\/blockquote>\n\n<pre><code># defining tensor. By default it will run on CPU.\nx = torch.ones(5, 3)\nprint(x)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor([[1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.],\n        [1., 1., 1.]])\n<\/code><\/pre>\n\n<p><code>ones()<\/code> is used to define tensor with <code>1.<\/code> (as shown in example) of given <code>size<\/code> and is not dependent on the existing tensor, whereas <code>new_ones()<\/code> works with existing tensor which inherits properties like <code>datatype<\/code> and <code>device<\/code> from existing tensor and define the <code>tensor<\/code> with given <code>size<\/code>.<\/p>\n"}]}
+{"QuestionId":52866449,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><rnn>","CreationDate":"2018-10-18T03:11:47.080","AcceptedAnswerId":52866760.0,"OwnerUserId":6750211.0,"Title":"Tensorflow RNN how to create zero state with various batch size?","Body":"<p>In this question <a href=\"https:\/\/stackoverflow.com\/questions\/39112622\/how-do-i-set-tensorflow-rnn-state-when-state-is-tuple-true\">How do I set TensorFlow RNN state when state_is_tuple=True?<\/a>: the accepted answer initialize the initial state like this:<\/p>\n\n<pre><code>state_placeholder = tf.placeholder(tf.float32, [num_layers, 2, batch_size, state_size])\n<\/code><\/pre>\n\n<p>I assume this requires a specific batch size, while what I have now is:<\/p>\n\n<pre><code>inputSeq = tf.placeholder(tf.float32, [None, seqLength, observationDim], name='input_seq')\noutputs, final_state = tf.nn.dynamic_rnn(cell, inputSeq, initial_state=initialState)\n<\/code><\/pre>\n\n<p>And I want this <code>initialState<\/code> to be a zero state, and can be configurable as the batch size of <code>inputSeq<\/code> could vary. However, <code>cell.zero_state<\/code> does not accept None as batch size. Is there any workaround?<\/p>\n","answers":[{"AnswerId":"52866760","CreationDate":"2018-10-18T03:51:29.173","ParentId":null,"OwnerUserId":"10323808","Title":null,"Body":"<p><code>cell.zero_state<\/code> accepts a scalar tensor.<\/p>\n\n<p>Get the batch size of the place holder via <code>tf.shape<\/code>, then it is done:\n<code>\nB = tf.shape(state_placeholder)[0]  # the batch size scalar tensor\ninitial_state = cell.zero_state(B)\n<\/code><\/p>\n"}]}
+{"QuestionId":52866886,"AnswerCount":1,"Tags":"<python><tensorflow><grpc><tensorflow-serving><grpc-python>","CreationDate":"2018-10-18T04:08:04.803","AcceptedAnswerId":53557174.0,"OwnerUserId":4391025.0,"Title":"Rendezvous of RPC that terminated with (StatusCode.UNAVAILABLE, Socket closed)>","Body":"<ul>\n<li>tensorflow-gpu 1.10.0<\/li>\n<li>tensorflow-server 1.10.0<\/li>\n<\/ul>\n\n<p>I have deployed a tensorflow server which serves several models.\nThe client code is like <code>client.py<\/code> this and I call the predict function.<\/p>\n\n<pre><code>channel = implementations.insecure_channel(host, port)\nstub = prediction_service_pb2.beta_create_PredictionService_stub(channel)\nrequest = predict_pb2.PredictRequest()\n\ndef predict(data, shape, model_name, signature_name=\"predict\"):\n    request.model_spec.name = model_name\n    request.model_spec.signature_name = signature_name\n    request.inputs['image'].CopyFrom(tf.contrib.util.make_tensor_proto(data, shape=shape))\n    result = stub.Predict(request, 10.0)\n    return result.outputs['prediction'].float_val[0]\n<\/code><\/pre>\n\n<p>I have about 100 clients with the same configuration.\nAnd here is a sample code to call the <code>predict<\/code> function:<\/p>\n\n<pre><code>from client import predict\nwhile True:\n    print(predict(data, shape, model_name))\n    # time.sleep some while\n<\/code><\/pre>\n\n<p>At first, when I run the client code, I can receieve the reponse <strong>correctly<\/strong>.\nBut after several hours, the client crashed with the error<\/p>\n\n<pre><code>_Rendezvous of RPC that terminated with (StatusCode.UNAVAILABLE, Socket closed)\n<\/code><\/pre>\n\n<p>I have try to modify my client code to<\/p>\n\n<pre><code>def predict(data, shape, model_name, signature_name=\"predict\"):\n    channel = implementations.insecure_channel(host, port)\n    stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)\n    request = predict_pb2.PredictRequest()\n    request.model_spec.name = model_name\n    request.model_spec.signature_name = signature_name\n    request.inputs['image'].CopyFrom(tf.contrib.util.make_tensor_proto(data, shape=shape))\n    result = stub.Predict(request, 10.0)\n    return result.outputs['prediction'].float_val[0]\n<\/code><\/pre>\n\n<p>that means I try to establish the connection with the tfs server each time the <code>predict<\/code> function is called. But this code also failed just like before.<\/p>\n\n<p>So what should I do to deal with this situation?<\/p>\n","answers":[{"AnswerId":"53557174","CreationDate":"2018-11-30T12:01:11.007","ParentId":null,"OwnerUserId":"4391025","Title":null,"Body":"<p>Finally I added a <code>channel.close()<\/code> before <code>return<\/code> and it works fine.<\/p>\n"}]}
+{"QuestionId":52867069,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-18T04:31:01.953","AcceptedAnswerId":null,"OwnerUserId":9633432.0,"Title":"Bi-LSTM Attention model in Keras","Body":"<p>I am trying to make an attention model with Bi-LSTM using word embeddings. I came across <a href=\"https:\/\/stackoverflow.com\/questions\/42918446\/how-to-add-an-attention-mechanism-in-keras\">How to add an attention mechanism in keras?<\/a>, <a href=\"https:\/\/github.com\/philipperemy\/keras-attention-mechanism\/blob\/master\/attention_lstm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/philipperemy\/keras-attention-mechanism\/blob\/master\/attention_lstm.py<\/a> and <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4962\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4962<\/a>. <\/p>\n\n<p>However, I am confused about the implementation of <code>Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification<\/code>. So,<\/p>\n\n<pre><code>_input = Input(shape=[max_length], dtype='int32')\n\n# get the embedding layer\nembedded = Embedding(\n        input_dim=30000,\n        output_dim=300,\n        input_length=100,\n        trainable=False,\n        mask_zero=False\n    )(_input)\n\nactivations = Bidirectional(LSTM(20, return_sequences=True))(embedded)\n\n# compute importance for each step\nattention = Dense(1, activation='tanh')(activations)\n<\/code><\/pre>\n\n<p>I am confused here as to which equation is to what from the paper.<\/p>\n\n<pre><code>attention = Flatten()(attention)\nattention = Activation('softmax')(attention)\n<\/code><\/pre>\n\n<p>What will RepeatVector do? <\/p>\n\n<pre><code>attention = RepeatVector(20)(attention)\nattention = Permute([2, 1])(attention)\n\n\nsent_representation = merge([activations, attention], mode='mul')\n<\/code><\/pre>\n\n<p>Now, again I am not sure why this line is here.<\/p>\n\n<pre><code>sent_representation = Lambda(lambda xin: K.sum(xin, axis=-2), output_shape=(units,))(sent_representation)\n<\/code><\/pre>\n\n<p>Since I have two classes, I will have the final softmax as:<\/p>\n\n<pre><code>probabilities = Dense(2, activation='softmax')(sent_representation)\n<\/code><\/pre>\n","answers":[{"AnswerId":"54507633","CreationDate":"2019-02-03T21:13:07.920","ParentId":null,"OwnerUserId":"9461915","Title":null,"Body":"<pre><code>attention = Flatten()(attention)  \n<\/code><\/pre>\n\n<p>transform your tensor of attention weights in a vector (of size max_length if your sequence size is max_length). <\/p>\n\n<pre><code>attention = Activation('softmax')(attention)\n<\/code><\/pre>\n\n<p>allows having all the attention weights between 0 and 1, the sum of all the weights equal to one. <\/p>\n\n<pre><code>attention = RepeatVector(20)(attention)\nattention = Permute([2, 1])(attention)\n\n\nsent_representation = merge([activations, attention], mode='mul')\n<\/code><\/pre>\n\n<p>RepeatVector repeat the attention weights vector (which is of size max_len) with the size of the hidden state (20) in order to multiply the activations and the hidden states element-wise. The size of the tensor variable <strong>activations<\/strong> is max_len*20. <\/p>\n\n<pre><code>sent_representation = Lambda(lambda xin: K.sum(xin, axis=-2), output_shape=(units,))(sent_representation)\n<\/code><\/pre>\n\n<p>This Lambda layer sum the weighted hidden states vectors in order to obtain the vector that will be used at the end. <\/p>\n\n<p>Hope this helped! <\/p>\n"}]}
+{"QuestionId":52867244,"AnswerCount":1,"Tags":"<python><csv><tensorflow><input>","CreationDate":"2018-10-18T04:54:51.897","AcceptedAnswerId":52867374.0,"OwnerUserId":4827407.0,"Title":"How to read csv in TensorFlow?","Body":"<p>I just start using TensorFlow. I am trying to read csv file in TensorFlow. This is an example I found online:<\/p>\n\n<pre><code>filename_queue = tf.train.string_input_producer([\"d:\/Feng\/LP\/tensorflowtrydata.csv\"])\nreader = tf.TextLineReader()\nkey, value = reader.read(filename_queue)\nrecord_defaults = [[1.0], [1.0], [1.0], [1.0], [\"Null\"]] \ncol1, col2, col3, col4, col5 = tf.decode_csv(value,record_defaults=record_defaults) \nfeatures = tf.stack([col1, col2, col3, col4])\nwith tf.Session() as sess:\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n    for i in range(200):\n        example, label = sess.run([features, col5])\n        print (example,col5)\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>But I have the error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Field 0 in record 0 is not a valid float: Sepal.Length\n [[Node: DecodeCSV_5 = DecodeCSV[OUT_TYPE=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_STRING], field_delim=\",\", na_value=\"\", use_quote_delim=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](ReaderReadV2_3:1, DecodeCSV_5\/record_defaults_0, DecodeCSV_5\/record_defaults_0, DecodeCSV_5\/record_defaults_0, DecodeCSV_5\/record_defaults_0, DecodeCSV_5\/record_defaults_4)]]\n<\/code><\/pre>\n\n<p>The data is iris dataset. It looks like:<\/p>\n\n<pre><code>iris.head()\n   Sepal.Length  Sepal.Width  Petal.Length  Petal.Width Species\n0           5.1          3.5           1.4          0.2  setosa\n1           4.9          3.0           1.4          0.2  setosa\n2           4.7          3.2           1.3          0.2  setosa\n3           4.6          3.1           1.5          0.2  setosa\n4           5.0          3.6           1.4          0.2  setosa\niris.dtypes\nSepal.Length    float64\nSepal.Width     float64\nPetal.Length    float64\nPetal.Width     float64\nSpecies          object\n<\/code><\/pre>\n\n<p>You can see that the error info said that it is not a valid float. But all the data are float64.<\/p>\n\n<p>I am not even sure where to start with this. <\/p>\n","answers":[{"AnswerId":"52867374","CreationDate":"2018-10-18T05:12:33.850","ParentId":null,"OwnerUserId":"7093741","Title":null,"Body":"<p>The confusion is that <code>iris.head()<\/code> provides first 5 lines of df <strong>AFTER<\/strong> header has been handled. Whereas in tensorflow, as you can see in the error line:\n<code>Field 0 in record 0 is not a valid float: \ufeffSepal.Length<\/code>, header is not automatically handled and thus <em>Sepal.Length<\/em>, the string, is causing the problem.<\/p>\n\n<p>You can either use <code>pandas.read_csv<\/code> to first import file then transform into whatever tensorflow wants, or use some options in like:<\/p>\n\n<pre><code>reader = tf.TextLineReader(skip_header_line=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52867472,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><keras><lstm>","CreationDate":"2018-10-18T05:24:45.687","AcceptedAnswerId":52872616.0,"OwnerUserId":7444723.0,"Title":"Keras fit model: TypeError: unhashable type: 'numpy.ndarray'","Body":"<p>I implement following code. It successfully works in previous version of Keras:<\/p>\n\n<pre><code>max_sequence = 56\ninput_dim = 26    \n\nprint(\"Build model..1\")\nfirst_input = Input(shape=(max_sequence,input_dim))\nfirst_lstm = LSTM(5, return_sequences=True)(first_input)\nfirst_bn = BatchNormalization()(first_lstm)\nfirst_activation = Activation('tanh')(first_bn)\nfirst_flat = Flatten()(first_activation)\n\nprint(\"Build model..2\")\nsecond_input = Input(shape=(max_sequence,input_dim))\nsecond_lstm = LSTM(5, return_sequences=True)(second_input)\nsecond_bn = BatchNormalization()(second_lstm)\nsecond_activation = Activation('tanh')(second_bn)\nsecond_flat = Flatten()(second_activation)\n\nmerge=concatenate([first_flat, second_flat])\nmerge_dense=Dense(3)(merge)\nmerge_bn = BatchNormalization()(merge_dense)\nmerge_activation = Activation('tanh')(merge_bn)\nmerge_dense2=Dense(1)(merge_activation)\nmerge_activation2 = Activation('tanh')(merge_dense2)\n\ntrain_x_1 = np.reshape(np.array(train_x_1), [2999, 56, 26])\ntrain_x_2 = np.reshape(np.array(train_x_2), [2999, 56, 26])\n\n\nmodel=Model(inputs=[train_x_1,train_x_2], outputs=train_y_class)\n\noptimizer = RMSprop(lr=0.5)\nmodel.compile(optimizer=optimizer, loss='binary_crossentropy', metrics=['accuracy'])\n\n\n\nhistory = model.fit([train_x_1, train_x_2], train_y_class, nb_epoch=300, batch_size=128,\n                    validation_data=([val_x_1, val_x_2], val_y_class))\n<\/code><\/pre>\n\n<p>When running:<\/p>\n\n<pre><code>history = model.fit([train_x_1, train_x_2], train_y_class, nb_epoch=300, batch_size=128,\n                    validation_data=([val_x_1, val_x_2], val_y_class))\n<\/code><\/pre>\n\n<p>the following error occurs:<\/p>\n\n<pre><code>TypeError: unhashable type: 'numpy.ndarray' accours.\n<\/code><\/pre>\n\n<p>So I checked <code>train_x_1<\/code>, <code>train_x_2<\/code>, <code>train_y_class<\/code>. Their type is <code>&lt;class 'numpy.ndarray'&gt;<\/code>. I have searched for a solution, so I tried to change type to tuple but It didn't work.<\/p>\n\n<p>If <code>numpy.ndarray<\/code> is unhashable, what type of input does <code>model.fit<\/code> receive?<\/p>\n\n<p>The shape of train data is as follows:<\/p>\n\n<pre><code>train_x_1.shape\n(2999, 56, 26)\ntrain_x_2.shape\n(2999, 56, 26)\ntrain_y_class.shape\n(2999, 1)\n<\/code><\/pre>\n\n<p>A sample of <code>train_x_1<\/code> is like below:<\/p>\n\n<pre><code>array([[[ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00],\n        [ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00],\n        [ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00],\n        ...,\n        [ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00],\n        [ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00],\n        [ 1.62601626e-02,  2.26890756e-01,  1.17764920e-02, ...,\n          0.00000000e+00,  0.00000000e+00,  0.00000000e+00]],\n<\/code><\/pre>\n","answers":[{"AnswerId":"52872616","CreationDate":"2018-10-18T11:07:38.580","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are directly passing the input and output arrays (and not the input and output tensors) to <code>Model<\/code> class when constructing your model:<\/p>\n\n<pre><code>model = Model(inputs=[train_x_1,train_x_2], outputs=train_y_class)\n<\/code><\/pre>\n\n<p>Instead, you need to pass the corresponding input and output tensors like this:<\/p>\n\n<pre><code>model = Model(inputs=[first_input,second_input], outputs=merge_activation2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52868094,"AnswerCount":0,"Tags":"<python><matrix><distance><pytorch><mahalanobis>","CreationDate":"2018-10-18T06:23:03.713","AcceptedAnswerId":null,"OwnerUserId":6789651.0,"Title":"Mahalanobis distance returns NaN in Pytorch","Body":"<p>I'm testing new metrics to measure distance between weight matrices in Pytorch, right now I'm trying to use Mahalanobis. For that I reshape every matrix into a vector and concat then into one matrix and then use this matrix to calculate the mahalanobis distance between any two rows of this matrix. Problem is some of those are getting me negative results and the square root of a negative throws me NaN.<\/p>\n\n<p>I know the covariance matrix has to be Positive Defined, I guess I'm messing up there, or maybe my ideia of using mahalanobis in this case is not possible? <\/p>\n\n<p>Here's the code I'm using, where I'm passing to it a X with shape (64,121), that being 64 (11x11) matrices<\/p>\n\n<pre><code>def _mahalanobis(X):\n    VI = torch.inverse(cov(X)) #covariance matrix\n    total_dist = 0\n    for i,v in enumerate(X):\n        dist = 0\n        for j,u in enumerate(X):\n            if i == j:\n                continue\n            x = (v-u).unsqueeze(0).t()\n            y = (v - u).unsqueeze(0)\n\n            dist = torch.sqrt(torch.mm(torch.mm(y,VI),x))\n\n            total_dist +=dist\n            print(dist)\n    return total_dist\n\n\n## Returns the covariance matrix of m\n    def cov(m, rowvar=False):\n\n        if m.dim() &gt; 2:\n            raise ValueError('m has more than 2 dimensions')\n        if m.dim() &lt; 2:\n            m = m.view(1, -1)\n        if not rowvar and m.size(0) != 1:\n            m = m.t()\n        # m = m.type(torch.double)  # uncomment this line if desired\n        fact = 1.0 \/ (m.size(1) - 1)\n        m -= torch.mean(m, dim=1, keepdim=True)\n        mt = m.t()  # if complex: mt = m.t().conj()\n        return fact * m.matmul(mt).squeeze()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52868102,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-10-18T06:23:39.963","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"What does the \"greater than\" operator \">\" mean for PyTorch tensors?","Body":"<p>I have a tensor <code>it<\/code> defined as:<\/p>\n\n<pre><code>import torch\nit = torch.tensor([0,  0,  0,  0,  0,  0,  0,  0,  0,  0], device='cuda:0')\n<\/code><\/pre>\n\n<p>Given this definition, what does <code>it &gt; 0<\/code> then mean?<\/p>\n","answers":[{"AnswerId":"52868567","CreationDate":"2018-10-18T06:58:29.407","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Using the <code>&gt;<\/code> operator is the same as using the <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.gt\" rel=\"nofollow noreferrer\"><code>torch.gt()<\/code><\/a> function.  <\/p>\n\n<p>In other words, <\/p>\n\n<pre><code>it &gt; 0\n<\/code><\/pre>\n\n<p>is the same as <\/p>\n\n<pre><code>torch.gt(it, 0)\n<\/code><\/pre>\n\n<p>and it returns a <code>ByteTensor<\/code> (a Boolean tensor) of the same <code>shape<\/code> as <code>it<\/code> where <code>out[i]<\/code> is True if <code>it[i] &gt; 0<\/code> and False otherwise. <\/p>\n"},{"AnswerId":"52868891","CreationDate":"2018-10-18T07:19:56.080","ParentId":null,"OwnerUserId":"5349822","Title":null,"Body":"<p>As seen from the question <code>it<\/code> is a 1-d tensor consisting of 10 elements.\nWhen we write <code>it &gt; 0<\/code> each element of the tensor <code>it<\/code> is compared with 0 and the output is set as <code>True<\/code> or <code>False<\/code> depending upon whether the number is greater than 0 or not. The result is also a 1-d boolean tensor consisting of <code>True<\/code> and <code>False<\/code> values.\nIn your case, you will get a 1-d tensor like this: <code>[False, False, False, False, False, False, False, False, False, False,]<\/code> as none of the elements in <code>it<\/code> is equal to 0.<\/p>\n\n<p>In more simpler terms,<\/p>\n\n<p>If <code>result<\/code> is a variable (actually a 1-d tenosr) to store the output having the same shape as <code>it<\/code> then the equation <code>result= it &gt; 0<\/code> can be written as:<\/p>\n\n<pre><code>if it[i]&gt;0:\n    result[i]= True\nelse:\n    result[i]= False \n<\/code><\/pre>\n\n<p>But when it's performed as <code>result= it &gt; 0<\/code>, the execution is much faster than writing our own <code>for\/while<\/code> loop.<\/p>\n\n<p>Hope this helps you.<\/p>\n"}]}
+{"QuestionId":52868152,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-18T06:28:42.280","AcceptedAnswerId":null,"OwnerUserId":10522104.0,"Title":"Tensorflow and Keras show a little different result even though I build exactly same models using same layer modules","Body":"<p>I'm using both Tensorflow and Keras, and I figure out they show a different result. There are already similar questions but I am a little different from them.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/52088807\/different-learning-curve-of-adam-between-tensorflow-and-keras#comment91140678_52088807\">Different learning curve of Adam between Tensorflow and Keras<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/51546176\/different-behaviour-between-same-implementations-of-tensorflow-and-keras\">Different behaviour between same implementations of TensorFlow and Keras<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/51475745\/different-training-result-using-tensorflow-and-keras\">Different training result using tensorflow and keras<\/a><\/li>\n<\/ul>\n\n<p>In my case, there is only a little difference in loss and accuracy, and I use exactly the same 'tf.keras.layers' modules. I think the only difference is AdamOptimizer and how to train the methods.<\/p>\n\n<ol>\n<li>tf.train.AdamOptimizer vs. tf.keras.optimizers.Adam<\/li>\n<li>tf.keras.models.fit vs. sess.run(train_optimizer)<\/li>\n<\/ol>\n\n<p>I checked that defaults of adam optimizers are same.\nI think that the difference is not caused by its randomness because I got similar results when I run the keras model a couple of times.  <\/p>\n\n<p>Here is my code<\/p>\n\n<blockquote>\n  <p>Keras Model<\/p>\n<\/blockquote>\n\n<pre><code># Build model and train\nX = tf.keras.layers.Input(shape=(sentence_size,), name='X')\n\nembedded_X = tf.keras.layers.Embedding(voca_size,\n                                       embedding_dim,\n                                       weights = [embedding_matrix],\n                                       input_length = sentence_size,\n                                       trainable=True)(X)\n\nhidden_states = tf.keras.layers.Bidirectional(tf.keras.layers.GRU(256, return_sequences=True))(embedded_X)\nl_pool = tf.keras.layers.GlobalMaxPooling1D()(hidden_states)\npreds = tf.keras.layers.Dense(1, activation = 'sigmoid')(l_pool)\n\nmodel = tf.keras.models.Model(X, preds)\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit( tokenized_train, y_train, shuffle=False, epochs=3, batch_size=32, validation_data= (tokenized_val, y_val))\n<\/code><\/pre>\n\n<blockquote>\n  <p>Tensorflow Model<\/p>\n<\/blockquote>\n\n<pre><code># Build model  \nX = tf.placeholder(tf.float32, [None, sentence_size])\n\nembedded_X = tf.keras.layers.Embedding(voca_size,\n                                       embedding_dim,\n                                       weights = [embedding_matrix],\n                                       input_length = sentence_size,\n                                       trainable=True)(X)\n\nhidden_states = tf.keras.layers.Bidirectional(tf.keras.layers.GRU(256, return_sequences=True))(embedded_X)\nl_pool = tf.keras.layers.GlobalMaxPooling1D()(hidden_states)\n_preds = tf.keras.layers.Dense(1, activation = 'sigmoid')(l_pool)\n\nlabels = tf.placeholder(tf.float32, [None, 1])\n_loss = tf.reduce_mean( tf.keras.losses.binary_crossentropy(labels, _preds) )\n_acc = tf.reduce_mean( tf.cast(tf.equal(labels, tf.round(_preds)), tf.float32) )\n_train_op = tf.train.AdamOptimizer().minimize(_loss)\n\n# Hyper parameters and loss_acc print function\nfrom math import ceil\n\nepochs = 3\nbatch_size = 32\nsteps_per_epoch = ceil( len(tokenized_train) \/ batch_size)\n\ndef loss_acc(sess, _loss, _preds, inputs, targets):\n    batch_size = len(inputs)\/\/100\n    steps_per_epoch = ceil( len(inputs) \/ batch_size )\n\n    data = tf.data.Dataset.from_tensor_slices((inputs, targets)).batch(batch_size).make_one_shot_iterator()\n    next_batch = data.get_next()\n\n    acc = 0\n    loss = 0\n\n    for batch in range(steps_per_epoch):\n        x, y = sess.run(next_batch)\n        l, a = sess.run([_loss, _acc], feed_dict={X:x, labels:y})\n\n        acc += a\/100\n        loss += l\/100\n\n    return loss, acc\n\n# Train model\ndata = tf.data.Dataset.from_tensor_slices((tokenized_train, y_train)).batch(batch_size).repeat().make_one_shot_iterator()\nnext_batch = data.get_next()\nsess.run(tf.global_variables_initializer())\n\nfor epoch in range(epochs):\n    for step in range(steps_per_epoch):\n        x, y = sess.run(next_batch)\n        batch_loss, batch_acc, _ = sess.run([_loss, _acc, _train_op], feed_dict={X:x, labels:y})\n        if step%125 == 0:\n            print('\\nBatch: %d' %step)\n            print(batch_loss, batch_acc)\n\n    train_loss, train_acc = loss_acc(sess, _loss, _preds, tokenized_train, y_train)\n    val_loss, val_acc = loss_acc(sess, _loss, _preds, tokenized_val, y_val)\n    print(\"\\nTrain loss: %.4f\" %train_loss)\n    print(\"Train acc: %.4f\" %train_acc)\n    print(\"Val loss: %.4f\" %val_loss)\n    print(\"Val acc: %.4f\" %val_acc)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Results<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5dih1.jpg\" rel=\"nofollow noreferrer\">Keras result<\/a><br>\n<a href=\"https:\/\/i.stack.imgur.com\/GOTA1.jpg\" rel=\"nofollow noreferrer\">Tensorflow result<\/a><\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"52868731","CreationDate":"2018-10-18T07:09:41.683","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>As long as you don\u00b4t have exactly the same initialized weights, you can not eliminate the problem of being non-deterministic. Your results are not varying very much and as you can see in your loss values, the starting points are quite different. Moreover 3 epochs are not very much, try to train more epochs and then compare the results. If your models overfit add some regularization.<\/p>\n"}]}
+{"QuestionId":52868422,"AnswerCount":1,"Tags":"<python><linux><tensorflow><anaconda>","CreationDate":"2018-10-18T06:48:08.410","AcceptedAnswerId":null,"OwnerUserId":9276708.0,"Title":"How to use anaconda packages in python environment on a linux server?","Body":"<p>I have installed <code>anaconda<\/code> , and use <code>conda install -c anaconda tensorflow-gpu<\/code> to install <code>tensorflow<\/code> .<\/p>\n\n<p>I activate <code>tensorflow<\/code> by using <code>source activate tensorflow<\/code> sucessfully, but I find that I can't use other <code>anaconda packages<\/code> , such as <code>pandas<\/code> .<\/p>\n\n<p>How to use <code>tensorflow<\/code> and <code>anaconda packages<\/code> sucessfully?<\/p>\n\n<p>Here is my operations and the versions of <code>anaconda<\/code> and <code>tensorflow<\/code>.<\/p>\n\n<p><strong>anaconda version:<\/strong><\/p>\n\n<pre><code># Name                    Version                   Build  Channel\nanaconda                  5.3.0                    py37_0  \n<\/code><\/pre>\n\n<p><strong>tensorflow version:<\/strong><\/p>\n\n<pre><code>(base) anonym2@amax:\/data\/anonym2\/code$ source activate tensorflow \n(tensorflow) anonym2@amax:\/data\/anonym2\/code$  python -c 'import tensorflow as tf; print(tf.__version__)'\n1.5.0\n<\/code><\/pre>\n\n<p><strong>import pandas in tensorflow environment:<\/strong><\/p>\n\n<pre><code>(tensorflow) anonym2@amax:\/data\/anonym2\/code$ import pandas \nimport: unable to open X server `localhost:10.0' @ error\/import.c\/ImportImageCommand\/364.\n(tensorflow) anonym2@amax:\/data\/anonym2\/code$ python test.py \nTraceback (most recent call last):\n  File \"test.py\", line 9, in &lt;module&gt;\n    import pandas as pd \nModuleNotFoundError: No module named 'pandas'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52871081","CreationDate":"2018-10-18T09:31:54.227","ParentId":null,"OwnerUserId":"2301712","Title":null,"Body":"<p>Please read <a href=\"https:\/\/conda.io\/docs\/user-guide\/tasks\/manage-environments.html#creating-an-environment-with-commands\" rel=\"nofollow noreferrer\">conda usage<\/a> first.<\/p>\n\n<p><code>conda install -n your_env_name pandas<\/code> may solve import pandas error, and replace <code>your_env_name<\/code> into your own environment name.<\/p>\n\n<p>You have to install packages as many as you need. To avoid this, you could create an environment with a specific version of anaconda python before installing tensorflow by <code>conda create -n your_env_name python=3.7 anaconda<\/code>. The python version can be changed as you want.<\/p>\n"}]}
+{"QuestionId":52868425,"AnswerCount":1,"Tags":"<python><tensorflow><merge><keras><lstm>","CreationDate":"2018-10-18T06:48:15.840","AcceptedAnswerId":null,"OwnerUserId":9571259.0,"Title":"\"Merge\" replacement Keras\/Tensorflow\/Python3","Body":"<p>Merge is not running with Keras\/Tensorflow\/Python3 version. With previous versions, Merge was running.\nBut now, it is not running. So I think I should convert this code with replacement \"Merge\". \"Merge\" is missing in new Keras\/Tensorflow\/Python3 version. <\/p>\n\n<p>How can I rewrite Merge(..., mode='ave') and Merge(...,mode='concat') ?<\/p>\n\n<p>My code part is below :<\/p>\n\n<pre><code>...\n\nfrom keras.layers import Merge\n...\nleft_branch = Sequential()\nleft_branch.add(Convolution3D...\n...\nright_branch = Sequential()\nright_branch.add(Convolution3D...\n...\nmerged = Merge([left_branch, right_branch], mode='ave')\nlstm_branch = Sequential()\nlstm_branch.add(Embedding(input_dim=7, output_dim=8, input_length=7))\nlstm_branch.add(LSTM(output_dim=8, input_shape=(7, ) ))\nlstm_branch.add(Dense(32, activation='relu'))\n\nmodel = Sequential()\nmodel.add(merged)\nmerged_2 = Merge([model, lstm_branch], mode='concat')\nmodel_2 = Sequential()\nmodel_2.add(merged_2)\nmodel_2.add(Dense(nb_classes,init='normal'))\nmodel_2.add(Activation('softmax'))\nmodel_2.compile(loss='categorical_crossentropy', optimizer='RMSprop', metrics=['accuracy'])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52868697","CreationDate":"2018-10-18T07:07:21.770","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>Merge has been replaced with Concatenate and Average,\nIt is a change in the new version of Keras. Now you should implement them like this<\/p>\n\n<pre><code>from keras.layers.merge import Concatenate, Average   \n\nlayer_merge = Concatenate()( [layer1, layer2] )\nlayer_avg   = Average()( [layer1, layer2] )\n<\/code><\/pre>\n\n<p>The full documentation can be found <a href=\"https:\/\/keras.io\/layers\/merge\/#average\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":52869004,"AnswerCount":1,"Tags":"<python><tensorflow><github><flask><websocket>","CreationDate":"2018-10-18T07:27:08.853","AcceptedAnswerId":null,"OwnerUserId":10192915.0,"Title":"getting this error while loading a tensorflow project","Body":"<p>getting a error while a installing a tensorflow python project from <a href=\"https:\/\/github.com\/victordibia\/skyfall\" rel=\"nofollow noreferrer\">https:\/\/github.com\/victordibia\/skyfall<\/a>\nfirst i tried to install the requiremnt.txt file i got <strong>Could not detect requirement name for 'git+<a href=\"https:\/\/github.com\/dpallot\/simple-websocket-server.git\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dpallot\/simple-websocket-server.git<\/a>', please specify one with #egg=your_package_name<\/strong> error,<\/p>\n\n<p>then i removed the last line of txt file, then getting this<\/p>\n\n<p><strong>Traceback (most recent call last):\n  File \"app.py\", line 18, in \n    from utils import web_socket_server\n  File \"C:\\Users\\anurag\\Downloads\\skyfall-master\\utils\\web_socket_server.py\", line 4, in \n    from SimpleWebSocketServer import SimpleWebSocketServer, WebSocket\nModuleNotFoundError: No module named 'SimpleWebSocketServer'<\/strong><\/p>\n\n<p>after that i installed git , now getting this<\/p>\n\n<p><strong>Traceback (most recent call last):\n  File \"app.py\", line 19, in \n    from utils import web_socket_client\n  File \"C:\\Users\\anurag\\Downloads\\skyfall-master\\utils\\web_socket_client.py\", line 8, in \n    from websocket import WebSocketException, WebSocketConnectionClosedException\nImportError: cannot import name 'WebSocketException'<\/strong><\/p>\n","answers":[{"AnswerId":"52869071","CreationDate":"2018-10-18T07:31:54.883","ParentId":null,"OwnerUserId":"9871094","Title":null,"Body":"<p>The error means that you have not installed this package: <a href=\"https:\/\/github.com\/dpallot\/simple-websocket-server\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dpallot\/simple-websocket-server<\/a><\/p>\n\n<p>Try this to install it(you must have <a href=\"https:\/\/pip.pypa.io\/en\/stable\/installing\/\" rel=\"nofollow noreferrer\">pip<\/a> installed and also <a href=\"https:\/\/git-scm.com\/book\/en\/v2\/Getting-Started-Installing-Git\" rel=\"nofollow noreferrer\">git<\/a>) and then run the project again:<\/p>\n\n<pre><code>python -m pip install git+https:\/\/github.com\/dpallot\/simple-websocket-server.git\n<\/code><\/pre>\n"}]}
+{"QuestionId":52869327,"AnswerCount":2,"Tags":"<python><keras><time-series><classification><lstm>","CreationDate":"2018-10-18T07:48:49.817","AcceptedAnswerId":52869915.0,"OwnerUserId":3687439.0,"Title":"Keras multi-step LSTM batch train classification at each step","Body":"<p><strong>Question<\/strong><\/p>\n\n<p>How to batch train a multi-step LSTM in Keras for single-label multi-class classificaiton, at each time-step for > 2 classes?<\/p>\n\n<p><strong>Current Error<\/strong><\/p>\n\n<p>Each target batch is a 3-dimensional array with shape <code>(batch_size, n_time_steps, n_classes)<\/code> but Keras expects a 2-dimensional array.<\/p>\n\n<p><strong>Example\/Context<\/strong><\/p>\n\n<p>Suppose we have daily closing prices for N stocks and for each day and stock: m features and one of three actions \"bought\", \"held\", \"sold\".\nIf there are 30 days' worth of data per stock we may train an LSTM to predict each action (on each day, for each stock) as follows.<\/p>\n\n<p>For each batch of samples of size n &lt;&lt; N, <code>X_train<\/code> will have a shape of <code>(n, 30, m)<\/code> i.e. <code>n<\/code> samples, 30 time-steps, and <code>m<\/code> features. After one-hot encoding \"bought\", \"held\", and \"sold\" <code>Y_train<\/code> will have a shape of <code>(n, 30, 3)<\/code>, which is an array of 3-dimensions.<\/p>\n\n<p>The problem is that Keras is giving an error due to expecting <code>Y_train<\/code> to be 2-dimensional.<\/p>\n\n<p>Here is a code snippet:<\/p>\n\n<pre><code>n_time_steps = 30\nn_ftrs = 700\nn_neurons = 100\nn_classes = 3\nbatch_size = 256\nn_epochs = 500\nmodel = Sequential()\nmodel.add(LSTM(n_neurons, input_shape=(n_time_steps, n_ftrs)))\nmodel.add(Dense(n_classes, activation='sigmoid'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', \n    metrics=['accuracy'])\n\nfor e in range(n_epochs):\n  X_train, Y_train = BatchGenerator()\n  # Y_train.shape = (256, 30, 3)\n  model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=1)\n<\/code><\/pre>\n\n<p><strong>Error<\/strong><\/p>\n\n<pre><code>Error when checking target: expected dense_20 to have 2 dimensions, \nbut got array with shape (256, 30, 3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52872475","CreationDate":"2018-10-18T10:59:13.300","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you take a look at the <code>model.summary()<\/code> output you would realize what the problem is:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nlstm_1 (LSTM)                (None, 100)               320400    \n_________________________________________________________________\ndense_74 (Dense)             (None, 3)                 303       \n=================================================================\nTotal params: 320,703\nTrainable params: 320,703\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>As you can see the output shape of <code>LSTM<\/code> layer is <code>(None, 100)<\/code> which means that <strong>only the output of last timestep is returned<\/strong>. And as a result, the output shape of <code>Dense<\/code> layer is <code>(None, 3)<\/code> which means that it would classify the whole input timeseries (i.e. the whole 30-day data of a stock) in to one of 3 classes. This is not what you want. Rather you want to classify each timestep of the input timeseries. To make this happen, as @VegardKT suggested, you can pass <code>return_sequences=True<\/code> to the <code>LSTM<\/code> layer to have its output at each timestep. Let's look at the <code>model.summary()<\/code> output after this change:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nlstm_2 (LSTM)                (None, 30, 100)           320400    \n_________________________________________________________________\ndense_75 (Dense)             (None, 30, 3)             303       \n=================================================================\nTotal params: 320,703\nTrainable params: 320,703\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>\u200bAs you can see, now the <code>LSTM<\/code> layer gives the output of each timestep and therefore the <code>Dense<\/code> layer which acts a classifier would be able to classify each of those timesteps into one of 3 classes as desired.<\/p>\n"},{"AnswerId":"52869915","CreationDate":"2018-10-18T08:26:28.847","ParentId":null,"OwnerUserId":"5143695","Title":null,"Body":"<p>You need to add change your LSTM layer like so:<\/p>\n\n<pre><code>model.add(LSTM(n_neurons, input_shape=(n_time_steps, n_ftrs), return_sequences=True))\n<\/code><\/pre>\n\n<p>This argument does the following:<\/p>\n\n<blockquote>\n  <p>return_sequences: Boolean. Whether to return the last output in the output sequence, or the full sequence.<\/p>\n<\/blockquote>\n\n<p>I'm going to be completely honest and say im not sure why it is like this, my LSTM is a bit rusty, as i would believe it should be the other way around, but im able to get your code running like this. If someone would like to clarify why this works that would be great.<\/p>\n"}]}
+{"QuestionId":52869912,"AnswerCount":0,"Tags":"<python><variables><tensorflow><data-structures><while-loop>","CreationDate":"2018-10-18T08:26:21.807","AcceptedAnswerId":null,"OwnerUserId":1099913.0,"Title":"Passing combination of python data structure and tf tensors as tf.while_loop variables","Body":"<p>I need to use <code>tf.while_loop<\/code> for my project. Is it possible to pass a list of both python dictionary and tensorflow tensors as a list of variables for body and cond functions? <\/p>\n\n<pre><code> a = {'state': 1 , 'path': 2}\n b = tf.get_variable(name = 'my_variable' , shape=[batch_size, None])\n state = {\n    'x' : a ,\n    'y' : b } \nshape = { \n    'y' : tf.TensorShape([batch_size, None])\n   }\n tf.while_loop(\n        cond, body, loop_vars=[state],\n        shape_invariants=[shape], parallel_iterations=1, back_prop=False)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52869987,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-serving>","CreationDate":"2018-10-18T08:31:48.660","AcceptedAnswerId":null,"OwnerUserId":10522743.0,"Title":"tensorflow serving produces wrong answer when I export keras model","Body":"<p>I tried export my keras model to tensorflow serving and all works well. What I'm trying to do is to accept a b64 encoded input image string from client and output a True\/False value. My keras model outputs 3 values and the first value indicates the degree predicted from model, and I will compare it to another fixed value and export the whole algorithm from taking image string to outputing True\/False value to tensorflow serving using RESTful API. However, I did not get correct output from my client program. Long words short, let me show the code<\/p>\n\n<p>My program to export saved model:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.saved_model import builder as saved_model_builder\nfrom tensorflow.python.saved_model import tag_constants, signature_constants, signature_def_utils_impl\nfrom keras.models import load_model\nfrom keras.layers import Input\nimport os\n\ntf.app.flags.DEFINE_string('model_dir', '.\/keras_models',\n                           '''Directory which contains keras models''')\ntf.app.flags.DEFINE_string('output_dir', '.\/model_output',\n                           '''Directory where to export the model''')\ntf.app.flags.DEFINE_string('model_version', '1',\n                           '''version number of the model''')\ntf.app.flags.DEFINE_string('model_file', 'pointer_model.json',\n                           '''json file which contains model architecture''')\ntf.app.flags.DEFINE_string('weights_file', 'pointer_model.h5',\n                           '''h5 file that contains model weights''')\n\nFLAGS = tf.app.flags.FLAGS\n\n\ndef preprocess_image(image_buffer):\n    '''\n    Preprocess JPEG encoded bytes to 3D floate tensor\n\n    :param image_buffer:\n    :return: 4D image tensor (1, width, height, channels)\n    '''\n\n    image = tf.image.decode_jpeg(image_buffer, channels=3)\n    image = tf.image.convert_image_dtype(image, dtype=tf.float32)\n\n    return image\n\n\ndef main(_):\n    with tf.Graph().as_default():\n        serialized_tf_example = tf.placeholder(tf.string, name='input_image')\n        feature_configs = {\n            'image\/encoded': tf.FixedLenFeature(\n                shape=[], dtype=tf.string),\n        }\n        tf_example = tf.parse_example(serialized_tf_example, feature_configs)\n        jpegs = tf_example['image\/encoded']\n        images = tf.map_fn(preprocess_image, jpegs, dtype=tf.float32)\n\n        images = tf.squeeze(images, [0])\n        images = tf.expand_dims(images, axis=0)\n        # now the image shape is [1, ?, ?, 3]\n        images = tf.image.resize_images(images, tf.constant([224, 224]))\n\n        model = load_model('.\/keras_models\/my_model.h5')\n\n        x = Input(tensor=images)\n        y = model(x)\n\n        model.summary()\n        compare_value = tf.Variable(100.0)\n        bool_out = tf.math.greater(y, compare_value)\n\n        bool_out = bool_out[:,0]\n\n        bool_out = tf.cast(bool_out, tf.float32)\n        bool_out = tf.expand_dims(bool_out, axis=0)\n        final_out = tf.concat([tf.transpose(y), bool_out], axis=0)\n        init = tf.global_variables_initializer()\n        with tf.Session() as sess:\n            sess.run(init)\n\n            # predict_tensor_input_info = tf.saved_model.utils.build_tensor_info(jpegs)\n            # predict_tensor_score_info = tf.saved_model.utils.build_tensor_info(bool_out)\n            prediction_signature = \\\n                (tf.saved_model.signature_def_utils.predict_signature_def(\n                    inputs={'images': jpegs},\n                    outputs={'scores': final_out}\n                )\n            )\n\n            export_path = os.path.join(\n                tf.compat.as_bytes(FLAGS.output_dir),\n                tf.compat.as_bytes(FLAGS.model_version)\n            )\n\n            builder = saved_model_builder.SavedModelBuilder(export_path)\n\n            legacy_init_op = tf.group(tf.tables_initializer(),\n                                      name = 'legacy_init_op')\n\n            builder.add_meta_graph_and_variables(\n                sess, [tag_constants.SERVING],\n                signature_def_map={\n                    signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:prediction_signature,\n                },\n                legacy_init_op = legacy_init_op\n            )\n\n            builder.save()\n\nif __name__ ==\"__main__\":\n    tf.app.run()\n<\/code><\/pre>\n\n<p>and this is my client program:<\/p>\n\n<pre><code>import base64\nimport requests\nimport json\nimport argparse\nimport time\nfrom glob import glob\n\nimage_path = glob('.\/segmented_image\/*.jpg')\n\n\nfor i in range(len(image_path)):\n\n    input_image = open(image_path[i], 'rb').read()\n\n    encoded_input_string = base64.b64encode(input_image)\n    input_string = encoded_input_string.decode('utf-8')\n    # input_image_recover = base64.b64decode(input_string)\n    # with open('recovered_image.jpg', 'wb') as output_file:\n    #     output_file.write(input_image_recover)\n    #\n    # print('Base64 encoded string: ' + input_string[:10] + '...' + input_string[-10:])\n\n    instance = [{\"b64\": input_string}]\n    data = json.dumps({\"instances\": instance})\n    print(data[:30]+ '...' + data[-10:])\n\n    json_response = requests.post('http:\/\/localhost:8501\/v1\/models\/pointer_model:predict',\n                                  data=data)\n\n    print(json_response.text)\n    end_time = time.time()\n<\/code><\/pre>\n\n<p>The output from json_response.text is like:<\/p>\n\n<pre><code>{\"instances\": [{\"b64\": \"\/9j\/4A...Y\/\/9k=\"}]}\n{\n    \"predictions\": [[-0.00015692], [-0.000967527], [0.000567942], [0.0]\n    ]\n}\n\n{\"instances\": [{\"b64\": \"\/9j\/4A...if\/9k=\"}]}\n{\n    \"predictions\": [[-0.000157582], [-0.000998327], [0.000598866], [0.0]\n    ]\n}\n<\/code><\/pre>\n\n<p>......<\/p>\n\n<p>The first 3 values in prediction key is supposed to be degree, and x,y coordinates in an image which should be hundreds value... the last value is the True\/False value casted to float32 comparing with 100.0<\/p>\n\n<p>Ok.. and last, I have also tested my model using model.predict, which gives correct answer...<\/p>\n\n<p>Now I'm completely confused. Can someone tell me where is wrong with my code?<\/p>\n","answers":[{"AnswerId":"53335914","CreationDate":"2018-11-16T10:26:23.277","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>use my script to export in tensorflow serving format<\/p>\n\n<pre><code>import sys\nfrom keras.models import load_model\nimport tensorflow as tf\nfrom keras import backend as K\nfrom tensorflow.python.framework import graph_util\nfrom tensorflow.python.framework import graph_io\nfrom tensorflow.python.saved_model import signature_constants\nfrom tensorflow.python.saved_model import tag_constants\n\n\nK.set_learning_phase(0)\nK.set_image_data_format('channels_last')\n\nINPUT_MODEL = sys.argv[1]\nNUMBER_OF_OUTPUTS = 1\nOUTPUT_NODE_PREFIX = 'output_node'\nOUTPUT_FOLDER= 'frozen'\nOUTPUT_GRAPH = 'frozen_model.pb'\nOUTPUT_SERVABLE_FOLDER = sys.argv[2]\nINPUT_TENSOR = sys.argv[3]\n\n\ntry:\n    model = load_model(INPUT_MODEL)\nexcept ValueError as err:\n    print('Please check the input saved model file')\n    raise err\n\noutput = [None]*NUMBER_OF_OUTPUTS\noutput_node_names = [None]*NUMBER_OF_OUTPUTS\nfor i in range(NUMBER_OF_OUTPUTS):\n    output_node_names[i] = OUTPUT_NODE_PREFIX+str(i)\n    output[i] = tf.identity(model.outputs[i], name=output_node_names[i])\nprint('Output Tensor names: ', output_node_names)\n\n\nsess = K.get_session()\ntry:\n    frozen_graph = graph_util.convert_variables_to_constants(sess,             sess.graph.as_graph_def(), output_node_names)    \n    graph_io.write_graph(frozen_graph, OUTPUT_FOLDER, OUTPUT_GRAPH, as_text=False)\n    print(f'Frozen graph ready for inference\/serving at     {OUTPUT_FOLDER}\/{OUTPUT_GRAPH}')\nexcept:\n    print('Error Occured')\n\n\n\nbuilder = tf.saved_model.builder.SavedModelBuilder(OUTPUT_SERVABLE_FOLDER)\n\nwith tf.gfile.GFile(f'{OUTPUT_FOLDER}\/{OUTPUT_GRAPH}', \"rb\") as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n\nsigs = {}\nOUTPUT_TENSOR = output_node_names\nwith tf.Session(graph=tf.Graph()) as sess:\n    tf.import_graph_def(graph_def, name=\"\")\n    g = tf.get_default_graph()\n    inp = g.get_tensor_by_name(INPUT_TENSOR)\n    out = g.get_tensor_by_name(OUTPUT_TENSOR[0] + ':0')\n\n    sigs[signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY] = \\\n        tf.saved_model.signature_def_utils.predict_signature_def(\n            {\"input\": inp}, {\"outout\": out})\n\n    builder.add_meta_graph_and_variables(sess,\n                                         [tag_constants.SERVING],\n                                         signature_def_map=sigs)\n    try:\n        builder.save()\n        print(f'Model ready for deployment at     {OUTPUT_SERVABLE_FOLDER}\/saved_model.pb')\n        print('Prediction signature : ')\n        print(sigs['serving_default'])\n    except:\n        print('Error Occured, please checked frozen graph')\n<\/code><\/pre>\n"}]}
+{"QuestionId":52870315,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-18T08:49:37.237","AcceptedAnswerId":null,"OwnerUserId":10443985.0,"Title":"Label encoding, one_hot encoding to labels","Body":"<p>My use case is about application of machine learning in wireless communication in a physical layer. \nMy input features are complex valued, and I converted it into equivalent real valued features as follows. <\/p>\n\n<pre><code> Y = np.concatenate((np.real(S_bar), np.imag(S_bar)))\nY = array([[-1.,  1.,  3.,  1., -3.,  1., -3., -3.,  1., -1.],\n     [-1.,  3.,  3., -1.,  1.,  3., -3.,  1., -1.,  3.],\n     [ 1., -1.,  3.,  3.,  3., -3., -1., -1., -1., -1.],\n     [ 3.,  1., -3.,  3., -3.,  1.,  1.,  3.,  3., -1.]])\n<\/code><\/pre>\n\n<p>Y is a 4 x 10 training matrix with the 4 rows are features, and the 10 columns are samples or training examples. \nI want to represent this as one_hot encoded training labels of 10 x 4 to fit it in to keras environment like this:<\/p>\n\n<pre><code>      model.fit(X_train, Y_labels, epochs=5, batch_size=100) \n<\/code><\/pre>\n\n<p>I tried it the following way and it does't work. <\/p>\n\n<p>i.  Label encoding: <\/p>\n\n<pre><code>     le = preprocessing.LabelEncoder()\n     y = []\nfor i in range(0,Y.shape[0]):\n    le.fit(Y[i,:])\n   temp = le.transform(Y[i,:]) \n   y.append(temp)\ny = np.array(y)\n<\/code><\/pre>\n\n<p>Y became Labeled:<\/p>\n\n<pre><code> array([[1, 2, 3, 2, 0, 2, 0, 0, 2, 1],\n   [1, 3, 3, 1, 2, 3, 0, 2, 1, 3],\n   [2, 1, 3, 3, 3, 0, 1, 1, 1, 1],\n   [3, 2, 0, 3, 0, 2, 2, 3, 3, 1]], dtype=int64) \n<\/code><\/pre>\n\n<p>ii. one_hot encoding: <\/p>\n\n<pre><code>encoder=OneHotEncoder(sparse=False)\nencoder.fit(y.T)\nlabels = encoder.transform( y.T) \n<\/code><\/pre>\n\n<p>* My problem is,the 'labels' came out with a shape of (10,16) while the model is expecting (10,4). How to fit the size of the one_hot encoder to the true shape of Y_train and Y_test please ? or is there any other way better than this to do it with the above sample. I really appreciate any help on this. <\/p>\n","answers":[{"AnswerId":"52882049","CreationDate":"2018-10-18T20:28:31.250","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>I would use <code>keras.utils.to_categorical<\/code> to get what you are after:<\/p>\n\n<pre><code>from keras.utils import to_categorical\nlabels = to_categorical(y[0])\n# labels.shape == (10, 4)\n<\/code><\/pre>\n\n<p>which basically one-hot encodes the target labels for you.<\/p>\n"}]}
+{"QuestionId":52870674,"AnswerCount":2,"Tags":"<python><tensorflow><windows-10><tensorflow-datasets>","CreationDate":"2018-10-18T09:10:03.950","AcceptedAnswerId":52924512.0,"OwnerUserId":9933958.0,"Title":"How to fix missing files or folder error: File \"generate_tfrecord.py\", line 110, in tf.app.run()?","Body":"<p>When I execute the following command I get the below error from Tensorflow \"missing file or folder\". I've checked all online solutions for this error, but nothing is resolving my error.<\/p>\n\n<p><code>python generate_tfrecord.py --csv_input=images\\train_labels.csv --image_dir=images\\train --output_path=train.record<\/code><\/p>\n\n<p><strong>The error:<\/strong><\/p>\n\n<blockquote>\n<pre><code>    File \"generate_tfrecord.py\", line 110, in \n    tf.app.run()\n    File \"C:\\anaconda3\\envs\\tensorflowc\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n    File \"generate_tfrecord.py\", line 101, in main\n    tf_example = create_tf_example(group, path)\n    File \"generate_tfrecord.py\", line 56, in create_tf_example\n    encoded_jpg = fid.read()\n    File \"C:\\anaconda3\\envs\\tensorflowc\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 125, in read\n    self._preread_check()\n    File \"C:\\anaconda3\\envs\\tensorflowc\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 85, in _preread_check\n    compat.as_bytes(self.__name), 1024 * 512, status)\n    File \"C:\\anaconda3\\envs\\tensorflowc\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 519, in exit\n    c_api.TF_GetCode(self.status.status))\n    tensorflow.python.framework.errors_impl.NotFoundError: NewRandomAccessFile failed to Create\/Open: C:\\tensorflowc\\models\\research\\object_detection\\images\\train\\tr1138a1a1_3_lar : The system cannot find the file specified.\n    ; No such file or directory\n<\/code><\/pre>\n<\/blockquote>\n","answers":[{"AnswerId":"52924512","CreationDate":"2018-10-22T07:44:48.203","ParentId":null,"OwnerUserId":"9933958","Title":null,"Body":"<p>I resolved the problem <\/p>\n\n<p>If you are making <code>.CSV file<\/code> using a <code>xml_to_csv file.py<\/code>,<\/p>\n\n<p>you have to check the file extension such as .jpg, .png, .jpeg in   <code>train_labels.csv<\/code> file.<\/p>\n\n<p>In my case, the xtension names won't be there !<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JtuNs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JtuNs.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Solution:<\/strong> <\/p>\n\n<p>add the extensions like below example and run following command:<\/p>\n\n<pre><code>python generate_tfrecord.py \n    --csv_input=images\\train_labels.csv \n    --image_dir=images\\train \n    --output_path=train.record\n<\/code><\/pre>\n\n<p>It will work !<\/p>\n"},{"AnswerId":"55997205","CreationDate":"2019-05-05T23:13:57.087","ParentId":null,"OwnerUserId":"10780044","Title":null,"Body":"<p>My csv-file contained imagenames with jpg extension and I still had this error OP posted. I tried solving it with:<\/p>\n\n<pre><code>python3 generate_tf_record.py --csv_input=data\/train_labels.csv --output_path=train.record\n\npython3 generate_tf_record.py --csv_input=data\/test_labels.csv --output_path=test.record\n<\/code><\/pre>\n\n<p>All images were in one folder and flags as shown below:<\/p>\n\n<pre><code>flags.DEFINE_string('csv_input', '','data\/train_labels.csv')\nflags.DEFINE_string('output_path','', 'train.record')\nflags.DEFINE_string('image_dir', '', 'images')\n<\/code><\/pre>\n\n<p>The problem was resolved when I copied the saved record file from the main folder to the data folder.<\/p>\n"}]}
+{"QuestionId":52870866,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network><lstm>","CreationDate":"2018-10-18T09:20:53.457","AcceptedAnswerId":52872242.0,"OwnerUserId":8196204.0,"Title":"Error 'Input 0 is incompatible with layer conv1d_48: expected ndim=3, found ndim=2' when adding Conv1D layer","Body":"<p>I am trying to construct the following model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(input_dim = num_top_words, output_dim = 64, input_length = input_length))\nmodel.add(LSTM(100, activation = 'relu'))\nmodel.add(Conv1D(64, kernel_size = 5, activation = 'relu'))\nmodel.add(MaxPooling1D())\nmodel.add(Dense(5, activation = 'softmax'))\nmodel.compile(loss = 'categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>But I get the following error when running it:<\/p>\n\n<pre><code>Input 0 is incompatible with layer conv1d_48: expected ndim=3, found ndim=2\n<\/code><\/pre>\n\n<p>which points out that there is an error at the following line:<\/p>\n\n<pre><code>model.add(Conv1D(64, kernel_size = 5, activation = 'relu'))\n<\/code><\/pre>\n\n<p>What might be the problem?<\/p>\n","answers":[{"AnswerId":"52872242","CreationDate":"2018-10-18T10:43:36.287","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that currently the output shape of <code>LSTM<\/code> layer is <code>(None, 100)<\/code>, however, as the error suggests, <code>Conv1D<\/code> layer like <code>LSTM<\/code> layer expects a 3D input of shape <code>(None, n_steps, n_features)<\/code>. So one way to resolve this is to pass <code>return_sequences=True<\/code> to LSTM layer to have the output of each timestep and therefore its output would be 3D:<\/p>\n\n<pre><code>model.add(LSTM(100, activation = 'relu', return_sequences=True))\n<\/code><\/pre>\n\n<p>Alternatively, you can put the <code>Conv1D<\/code> and <code>MaxPooling1D<\/code> layers before the <code>LSTM<\/code> layer (which may be even better than the current architecture, since one usage of <code>Conv1D<\/code> plus pooling layers is to reduce the dimension of LSTM layer's input and hence reduce the computational complexity):<\/p>\n\n<pre><code>model.add(Conv1D(64, kernel_size = 5, activation = 'relu'))\nmodel.add(MaxPooling1D())\nmodel.add(LSTM(100, activation = 'relu'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52871224,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-10-18T09:39:45.070","AcceptedAnswerId":null,"OwnerUserId":2481799.0,"Title":"RuntimeError: cuda runtime error (8) in pytorch","Body":"<p>when I try to run a pytorch faster rcnn code (from <a href=\"https:\/\/github.com\/rowanz\/neural-motifs\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rowanz\/neural-motifs<\/a>), I get the issuse as follow<\/p>\n\n<pre><code>THCudaCheck FAIL file=\/opt\/conda\/conda-bld\/pytorch_1518238409320\/work\/torch\/lib\/THC\/generic\/THCTensorMathPairwise.cu \nline=21 error=8 : invalid device function\n<\/code><\/pre>\n\n<p>It arises when operate\n    keep.append(keep_im + s)\nwhere keep is a list, s is a int number and keep_im  is a torch.cuda.Longtensor<\/p>\n\n<p>It's strange that, when I modify the code as <\/p>\n\n<pre><code>try:\n    keep.append(keep_im + s)\nexcept BaseException:\n    keep.append(keep_im + s)\n<\/code><\/pre>\n\n<p>It return error in the try and then operates again in the except, and successes....<\/p>\n\n<p>Anyone know what happen here?\nI use python 2.7 + pytorch 0.3 + Cuda 8 + cudnn7.1, Titan XP in Ubuntu 16. Thanks<\/p>\n","answers":[]}
+{"QuestionId":52871561,"AnswerCount":2,"Tags":"<macos><tensorflow><deep-learning><macos-sierra>","CreationDate":"2018-10-18T10:01:11.243","AcceptedAnswerId":null,"OwnerUserId":10522815.0,"Title":"How to install AMD rocm in Apple imac pro 2017 (to further install tensorflow for Deep Learning)","Body":"<p>I didn't find any way after thoroughly searched in the web to install AMD rocm (libraries for AMD graphics card for GPU computing to run deep learning - like tensorflow) for iMac pro 2017 desktop.<\/p>\n\n<p>As per rocm (<a href=\"https:\/\/rocm.github.io\/ROCmInstall.html\" rel=\"nofollow noreferrer\">https:\/\/rocm.github.io\/ROCmInstall.html<\/a>) the only operating system <\/p>\n\n<ul>\n<li>Ubuntu 16.04 &amp;. 18.04 (Version 16.04.3 and newer or kernels 4.13 and\nnewer) <\/li>\n<li>CentOS 7.4 &amp;. 7.5 (Using devetoolset-7 runtime support)<\/li>\n<li>RHEL 7.4. &amp;. 7.5 (Using devetoolset-7 runtime support)<\/li>\n<\/ul>\n\n<p>Also, Apple iMac pro 2017 provided very powerful AMD graphics Card (<code>Radeon Pro Vega 56 8176 MB<\/code>), where the benchmark shows its equivalent to  <\/p>\n\n<p>Does this mean, we can't install rocm on mac (as it's not supported \/ version is not available for mac)?<\/p>\n\n<p>OR<\/p>\n\n<p>Has anyone successfully installed the rocm on mac have AMD graphics card built it and installed &amp; run the tensorflow for Rocm Software Platform (<a href=\"https:\/\/github.com\/ROCmSoftwarePlatform\/tensorflow-upstream\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ROCmSoftwarePlatform\/tensorflow-upstream<\/a>)<\/p>\n","answers":[{"AnswerId":"53789474","CreationDate":"2018-12-15T04:13:49.530","ParentId":null,"OwnerUserId":"377160","Title":null,"Body":"<p>Based on there being an open bug about adding mac support since about a year now (<a href=\"https:\/\/github.com\/RadeonOpenCompute\/ROCm\/issues\/262\" rel=\"nofollow noreferrer\">https:\/\/github.com\/RadeonOpenCompute\/ROCm\/issues\/262<\/a>) I'd venture to say the answer is currently no.<\/p>\n"},{"AnswerId":"57319963","CreationDate":"2019-08-02T04:28:43.123","ParentId":null,"OwnerUserId":"7038317","Title":null,"Body":"<p>According to this issue(<a href=\"https:\/\/github.com\/RadeonOpenCompute\/ROCm\/issues\/262\" rel=\"nofollow noreferrer\">https:\/\/github.com\/RadeonOpenCompute\/ROCm\/issues\/262<\/a>), there is no clear plan for developing macos-compatible rocm.\"ROCm is based on customized linux kernel. so there is no compatibility with Unix kernel.\"<\/p>\n\n<p>and I found a post about how to use do deel learning on mac with amd gpu(<a href=\"https:\/\/medium.com\/@danbrice.datascience\/deep-learning-on-a-mac-with-amd-gpu-4be1f18944a),hope\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@danbrice.datascience\/deep-learning-on-a-mac-with-amd-gpu-4be1f18944a),hope<\/a> it helps.<\/p>\n"}]}
+{"QuestionId":52871934,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><semantic-segmentation>","CreationDate":"2018-10-18T10:23:54.547","AcceptedAnswerId":52879084.0,"OwnerUserId":10523094.0,"Title":"Multi-class image classification, how to load mask","Body":"<p>I'm totally new to Keras, I've only been working for a few days, so I'm pretty inexperienced.<\/p>\n\n<p>I was able to train a U-Net network that works with a class, then input an RGB image and a grayscale mask for training, with the following code:<\/p>\n\n<pre><code>def train_generator():\n\n    while True:\n        for start in range(0, len(ids_train_split), batch_size):\n            x_batch = []\n            y_batch = []\n            end = min(start + batch_size, len(ids_train_split))\n            ids_train_batch = ids_train_split[start:end]\n            for id in ids_train_batch.values:\n\n                img_name = 'IMG_'+str(id).split('_')[2]\n                image_path = os.path.join(\"input\", \"train\", \"{}.JPG\".format(str(img_name)))\n                mca_mask_path = os.path.join(\"input\", \"train_mask\", \"{}.png\".format(id))\n\n                img = cv2.imread(image_path)\n                img = cv2.resize(img, (input_size, input_size))\n\n                mask_mca = cv2.imread(mca_mask_path, cv2.IMREAD_GRAYSCALE)\n                mask_mca = cv2.resize(mask_mca, (input_size, input_size))\n\n                img = randomHueSaturationValue(img,\n                                               hue_shift_limit=(-50, 50),\n                                               sat_shift_limit=(-5, 5),\n                                               val_shift_limit=(-15, 15))\n                img, mask = randomShiftScaleRotate(img, mask,\n                                                   shift_limit=(-0.0625, 0.0625),\n                                                   scale_limit=(-0.1, 0.1),\n                                                   rotate_limit=(-0, 0))\n                img, mask = randomHorizontalFlip(img, mask)\n                mask = np.expand_dims(mask, axis=2)\n                x_batch.append(img)\n                y_batch.append(mask)\n            x_batch = np.array(x_batch, np.float32) \/ 255\n            y_batch = np.array(y_batch, np.float32) \/ 255\n            yield x_batch, y_batch\n<\/code><\/pre>\n\n<p>And this is my U-Net model:<\/p>\n\n<pre><code>def get_unet_1(pretrained_weights=None, input_shape=(1024, 1024, 3), num_classes=1, learning_rate=0.0001):\n    inputs = Input(shape=input_shape)\n    # 1024\n\n    down0b = Conv2D(8, (3, 3), padding='same')(inputs)\n    down0b = BatchNormalization()(down0b)\n    down0b = Activation('relu')(down0b)\n    down0b = Conv2D(8, (3, 3), padding='same')(down0b)\n    down0b = BatchNormalization()(down0b)\n    down0b = Activation('relu')(down0b)\n    down0b_pool = MaxPooling2D((2, 2), strides=(2, 2))(down0b)\n    # 512\n\n    down0a = Conv2D(16, (3, 3), padding='same')(down0b_pool)\n    down0a = BatchNormalization()(down0a)\n    down0a = Activation('relu')(down0a)\n    down0a = Conv2D(16, (3, 3), padding='same')(down0a)\n    down0a = BatchNormalization()(down0a)\n    down0a = Activation('relu')(down0a)\n    down0a_pool = MaxPooling2D((2, 2), strides=(2, 2))(down0a)\n    # 256\n\n    down0 = Conv2D(32, (3, 3), padding='same')(down0a_pool)\n    down0 = BatchNormalization()(down0)\n    down0 = Activation('relu')(down0)\n    down0 = Conv2D(32, (3, 3), padding='same')(down0)\n    down0 = BatchNormalization()(down0)\n    down0 = Activation('relu')(down0)\n    down0_pool = MaxPooling2D((2, 2), strides=(2, 2))(down0)\n    # 128\n\n    down1 = Conv2D(64, (3, 3), padding='same')(down0_pool)\n    down1 = BatchNormalization()(down1)\n    down1 = Activation('relu')(down1)\n    down1 = Conv2D(64, (3, 3), padding='same')(down1)\n    down1 = BatchNormalization()(down1)\n    down1 = Activation('relu')(down1)\n    down1_pool = MaxPooling2D((2, 2), strides=(2, 2))(down1)\n    # 64\n\n    down2 = Conv2D(128, (3, 3), padding='same')(down1_pool)\n    down2 = BatchNormalization()(down2)\n    down2 = Activation('relu')(down2)\n    down2 = Conv2D(128, (3, 3), padding='same')(down2)\n    down2 = BatchNormalization()(down2)\n    down2 = Activation('relu')(down2)\n    down2_pool = MaxPooling2D((2, 2), strides=(2, 2))(down2)\n    # 32\n\n    down3 = Conv2D(256, (3, 3), padding='same')(down2_pool)\n    down3 = BatchNormalization()(down3)\n    down3 = Activation('relu')(down3)\n    down3 = Conv2D(256, (3, 3), padding='same')(down3)\n    down3 = BatchNormalization()(down3)\n    down3 = Activation('relu')(down3)\n    down3_pool = MaxPooling2D((2, 2), strides=(2, 2))(down3)\n    # 16\n\n    down4 = Conv2D(512, (3, 3), padding='same')(down3_pool)\n    down4 = BatchNormalization()(down4)\n    down4 = Activation('relu')(down4)\n    down4 = Conv2D(512, (3, 3), padding='same')(down4)\n    down4 = BatchNormalization()(down4)\n    down4 = Activation('relu')(down4)\n    down4_pool = MaxPooling2D((2, 2), strides=(2, 2))(down4)\n    # 8\n\n    center = Conv2D(1024, (3, 3), padding='same')(down4_pool)\n    center = BatchNormalization()(center)\n    center = Activation('relu')(center)\n    center = Conv2D(1024, (3, 3), padding='same')(center)\n    center = BatchNormalization()(center)\n    center = Activation('relu')(center)\n    # center\n\n    up4 = UpSampling2D((2, 2))(center)\n    up4 = concatenate([down4, up4], axis=3)\n    up4 = Conv2D(512, (3, 3), padding='same')(up4)\n    up4 = BatchNormalization()(up4)\n    up4 = Activation('relu')(up4)\n    up4 = Conv2D(512, (3, 3), padding='same')(up4)\n    up4 = BatchNormalization()(up4)\n    up4 = Activation('relu')(up4)\n    up4 = Conv2D(512, (3, 3), padding='same')(up4)\n    up4 = BatchNormalization()(up4)\n    up4 = Activation('relu')(up4)\n    # 16\n\n    up3 = UpSampling2D((2, 2))(up4)\n    up3 = concatenate([down3, up3], axis=3)\n    up3 = Conv2D(256, (3, 3), padding='same')(up3)\n    up3 = BatchNormalization()(up3)\n    up3 = Activation('relu')(up3)\n    up3 = Conv2D(256, (3, 3), padding='same')(up3)\n    up3 = BatchNormalization()(up3)\n    up3 = Activation('relu')(up3)\n    up3 = Conv2D(256, (3, 3), padding='same')(up3)\n    up3 = BatchNormalization()(up3)\n    up3 = Activation('relu')(up3)\n    # 32\n\n    up2 = UpSampling2D((2, 2))(up3)\n    up2 = concatenate([down2, up2], axis=3)\n    up2 = Conv2D(128, (3, 3), padding='same')(up2)\n    up2 = BatchNormalization()(up2)\n    up2 = Activation('relu')(up2)\n    up2 = Conv2D(128, (3, 3), padding='same')(up2)\n    up2 = BatchNormalization()(up2)\n    up2 = Activation('relu')(up2)\n    up2 = Conv2D(128, (3, 3), padding='same')(up2)\n    up2 = BatchNormalization()(up2)\n    up2 = Activation('relu')(up2)\n    # 64\n\n    up1 = UpSampling2D((2, 2))(up2)\n    up1 = concatenate([down1, up1], axis=3)\n    up1 = Conv2D(64, (3, 3), padding='same')(up1)\n    up1 = BatchNormalization()(up1)\n    up1 = Activation('relu')(up1)\n    up1 = Conv2D(64, (3, 3), padding='same')(up1)\n    up1 = BatchNormalization()(up1)\n    up1 = Activation('relu')(up1)\n    up1 = Conv2D(64, (3, 3), padding='same')(up1)\n    up1 = BatchNormalization()(up1)\n    up1 = Activation('relu')(up1)\n    # 128\n\n    up0 = UpSampling2D((2, 2))(up1)\n    up0 = concatenate([down0, up0], axis=3)\n    up0 = Conv2D(32, (3, 3), padding='same')(up0)\n    up0 = BatchNormalization()(up0)\n    up0 = Activation('relu')(up0)\n    up0 = Conv2D(32, (3, 3), padding='same')(up0)\n    up0 = BatchNormalization()(up0)\n    up0 = Activation('relu')(up0)\n    up0 = Conv2D(32, (3, 3), padding='same')(up0)\n    up0 = BatchNormalization()(up0)\n    up0 = Activation('relu')(up0)\n    # 256\n\n    up0a = UpSampling2D((2, 2))(up0)\n    up0a = concatenate([down0a, up0a], axis=3)\n    up0a = Conv2D(16, (3, 3), padding='same')(up0a)\n    up0a = BatchNormalization()(up0a)\n    up0a = Activation('relu')(up0a)\n    up0a = Conv2D(16, (3, 3), padding='same')(up0a)\n    up0a = BatchNormalization()(up0a)\n    up0a = Activation('relu')(up0a)\n    up0a = Conv2D(16, (3, 3), padding='same')(up0a)\n    up0a = BatchNormalization()(up0a)\n    up0a = Activation('relu')(up0a)\n    # 512\n\n    up0b = UpSampling2D((2, 2))(up0a)\n    up0b = concatenate([down0b, up0b], axis=3)\n    up0b = Conv2D(8, (3, 3), padding='same')(up0b)\n    up0b = BatchNormalization()(up0b)\n    up0b = Activation('relu')(up0b)\n    up0b = Conv2D(8, (3, 3), padding='same')(up0b)\n    up0b = BatchNormalization()(up0b)\n    up0b = Activation('relu')(up0b)\n    up0b = Conv2D(8, (3, 3), padding='same')(up0b)\n    up0b = BatchNormalization()(up0b)\n    up0b = Activation('relu')(up0b)\n    # 1024\n\n    classify = Conv2D(num_classes, (1, 1), activation='sigmoid')(up0b)\n\n    model = Model(inputs=inputs, outputs=classify)\n\n    model.compile(optimizer=RMSprop(lr=learning_rate), loss=make_loss('bce_dice'), metrics=[dice_coef, 'accuracy'])\n\n    if pretrained_weights:\n        model.load_weights(pretrained_weights)\n\n    return model\n<\/code><\/pre>\n\n<p>Now I've to modify the problem and make it a multi-class classifier, so I do not work with a mask anymore but two. So I have two types of grasycale masks (<code>Mca_mask<\/code> and <code>NotMca_mask<\/code> of the same train img), in this case what is the standard practice? Merge the two masks into one?<\/p>\n","answers":[{"AnswerId":"52879084","CreationDate":"2018-10-18T17:01:02.150","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>On this line we can see that your output layer is applying sigmoid:<\/p>\n\n<pre><code>classify = Conv2D(num_classes, (1, 1), activation='sigmoid')(up0b)\n<\/code><\/pre>\n\n<p>Which means all your outputs are transformed to be between <code>[0,1]<\/code> without any dependence between them. This is what you want for multi-class classification. As a side note another common way to transform your output layer to the <code>[0,1]<\/code> range is to apply softmax - this is not good for multiclass because as one class grows in confidence the others must necessarily decrease.<\/p>\n\n<p>Your loss function is defined as Binary Cross Entropy on this line:<\/p>\n\n<pre><code>model.compile(optimizer=RMSprop(lr=learning_rate), loss=make_loss('bce_dice'), metrics=[dice_coef, 'accuracy'])\n<\/code><\/pre>\n\n<p>Which is appropriate for all types of classification (single class or multi-class), and requires outputs in the <code>[0,1]<\/code> range.<\/p>\n\n<p>So basically you're all set to do multi-class classification as you are configured now. All you need to do is create multi-class labels. For example, if your classes are dog, cat, bird, horse, goat and an image has a dog &amp; cat in it, your label would be <code>[1, 1, 0, 0, 0]<\/code>, and you can train the network on that as-is.<\/p>\n"}]}
+{"QuestionId":52871974,"AnswerCount":0,"Tags":"<python><tensorflow><keras><multiprocessing>","CreationDate":"2018-10-18T10:25:46.313","AcceptedAnswerId":null,"OwnerUserId":10478684.0,"Title":"Using model.predict (Keras + TF) in multiprocessing","Body":"<p>I have following problem. I'm using a Tensorflow Keras model to evaluate continuous sensor data. My input for my model consists of 15 sensor data frames. Because the function model.predict() takes near 1 second I wanted to execute this function asynchronous so that I can collect the next data frames in this time period.\nTo accomplish this I created a Pool with the multiprocessing libary and a function to for model.predict. My code looks something like this:<\/p>\n\n<pre><code>def predictData(data): \n   return model.predict(data)\n\nglobal model\nmodel = tf.keras.models.load_model(\"Network.h5\")\nmodel._make_predict_function()\n\np = Pool(processes = 4)\n...\nres = p.apply_async(predictData, ([[iinput]],))\nprint(res.get(timeout = 10))\n<\/code><\/pre>\n\n<p>Now I always get a timeout-error when calling predictData(). It seems like model.predict() is not working right. What am I making wrong?<\/p>\n","answers":[]}
+{"QuestionId":52872239,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T10:43:25.353","AcceptedAnswerId":52872757.0,"OwnerUserId":9936406.0,"Title":"Can tf.scatter_update or tf.scatter_nd_update be used to update column slices of a tensor?","Body":"<p>I want to implement a function that takes in a variable as input, mutate some of its rows or columns and replaces them back in the original variable. I am able to implement it for row slices using tf.gather and tf.scatter_update but unable to do so for column slices since apparently tf.scatter_update only updates the row slices and does not have an axis feature. I am not an expert in tensorflow therefore I may be missing something. Can someone help? <\/p>\n\n<pre><code>def matrix_reg(t, percent_t, beta):\n\n    ''' Takes a variable tensor t as input and regularizes some of its rows.\n    The number of rows to be regularized are specified by the percent_t. Reuturns the original tensor by updating its rows indexed by row_ind.\n\n    Arguements:\n        t -- input tensor\n        percent_t -- percentage of the total rows\n        beta -- the regularization factor\n    Output:\n        the regularized tensor\n        '''\n    row_ind = np.random.choice(int(t.shape[0]), int(percent_t*int(t.shape[0])), replace = False)\n    t_ = tf.gather(t,row_ind)\n    t_reg = (1+beta)*t_-beta*(tf.matmul(tf.matmul(t_,tf.transpose(t_)),t_))\n    return tf.scatter_update(t, row_ind, t_reg)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52872757","CreationDate":"2018-10-18T11:16:37.117","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a small demonstration of how to update rows or columns. The idea is that you specify the row and column indices of the variables where you want each element in the update to end up. That is easy to do with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/meshgrid\" rel=\"nofollow noreferrer\"><code>tf.meshgrid<\/code><\/a>.<\/p>\n\n<pre><code>import tensorflow as tf\n\nvar = tf.get_variable('var', [4, 3], tf.float32, initializer=tf.zeros_initializer())\nupdates = tf.placeholder(tf.float32, [None, None])\nindices = tf.placeholder(tf.int32, [None])\n# Update rows\nvar_update_rows = tf.scatter_update(var, indices, updates)\n# Update columns\ncol_indices_nd = tf.stack(tf.meshgrid(tf.range(tf.shape(var)[0]), indices, indexing='ij'), axis=-1)\nvar_update_cols = tf.scatter_nd_update(var, col_indices_nd, updates)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    print('Rows updated:')\n    print(sess.run(var_update_rows, feed_dict={updates: [[1, 2, 3], [4, 5, 6]], indices: [3, 1]}))\n    print('Columns updated:')\n    print(sess.run(var_update_cols, feed_dict={updates: [[1, 5], [2, 6], [3, 7], [4, 8]], indices: [0, 2]}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Rows updated:\n[[0. 0. 0.]\n [4. 5. 6.]\n [0. 0. 0.]\n [1. 2. 3.]]\nColumns updated:\n[[1. 0. 5.]\n [2. 5. 6.]\n [3. 0. 7.]\n [4. 2. 8.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52872377,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-slim><vgg-net>","CreationDate":"2018-10-18T10:52:21.363","AcceptedAnswerId":null,"OwnerUserId":2645853.0,"Title":"Why Tensorflow's non-standard VGG implementation?","Body":"<p>Tensorflow  .slim library implements vgg, but replaces the fully connected layers with convolutional ones. The code is at: <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/nets\/vgg.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/nets\/vgg.py<\/a><\/p>\n\n<p>So for example, fc8 layer is implemented as:<\/p>\n\n<pre><code>net = layers.conv2d(\n      net,\n      num_classes, [1, 1],\n      activation_fn=None,\n      normalizer_fn=None,\n      scope='fc8')\n<\/code><\/pre>\n\n<p>which looks like a very different operation, with much fewer weights. \nWhat is the reason for this? <\/p>\n","answers":[]}
+{"QuestionId":52872569,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><scikit-learn><supervised-learning>","CreationDate":"2018-10-18T11:05:23.050","AcceptedAnswerId":null,"OwnerUserId":9653537.0,"Title":"How to train model that compares similarities of phone contacts list?","Body":"<p>Let's say I have data of phone numbers:<\/p>\n\n<p>(i.e. a collection that each of its indexes contains a list of phone numbers. \n (<em>array of arrays<\/em>))<\/p>\n\n<p>What I want to is a method to recognize \"<em>similar contacts<\/em>\".\ni.e. how the data in the \"i\" index is similar to the data in the \"j\" index.<\/p>\n\n<p>I want to answer this question using a trained ml model.<\/p>\n\n<p>I guess one way of doing it is using \"<em>supervised learning<\/em>\" that uses a classifier that compares \"<em>vectors<\/em>\" that represent each contact list.<\/p>\n\n<p>I'm new to ml and python also, so can you please advise how it can be done? <\/p>\n","answers":[{"AnswerId":"52872660","CreationDate":"2018-10-18T11:10:51.060","ParentId":null,"OwnerUserId":"4583734","Title":null,"Body":"<p>You don't need to use ML for this problem. You can use string distance which measures how similar two string are (obviously convert the phone numbers to strings before you use the algorithm). There is more info <a href=\"https:\/\/stackabuse.com\/levenshtein-distance-and-text-similarity-in-python\/\" rel=\"nofollow noreferrer\">here<\/a> and a pretty good python package to help <a href=\"https:\/\/pypi.org\/project\/python-Levenshtein\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52873243,"AnswerCount":0,"Tags":"<python><tensorflow><websocket><libwebsockets>","CreationDate":"2018-10-18T11:45:37.130","AcceptedAnswerId":null,"OwnerUserId":10192915.0,"Title":"getting web_socket error while running python project","Body":"<p>I getting this error while running this project <a href=\"https:\/\/github.com\/victordibia\/skyfall\" rel=\"nofollow noreferrer\">https:\/\/github.com\/victordibia\/skyfall<\/a> <\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"app.py\", line 19, in\n  \n      from utils import web_socket_client   File \"C:\\Users\\anurag\\Downloads\\skyfall-master\\utils\\web_socket_client.py\",\n  line 8, in \n      from websocket import WebSocketException, WebSocketConnectionClosedException ImportError: cannot import name\n  'WebSocketException'<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52873329,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-18T11:50:21.600","AcceptedAnswerId":52881462.0,"OwnerUserId":10419781.0,"Title":"Does Tensorflows tf.layers.dense flatten input dimensions?","Body":"<p>I'm searching for a data leak in my model. I'm using <code>tf.layers.dense<\/code> before a masking operation and am concerned that the model could just learn to switch positions in the middle dimension of my input tensor.<\/p>\n\n<p>When I have an input tensor <code>x = tf.ones((2,3,4))<\/code> would <code>tf.layers.dense(x,8)<\/code> flatten <code>x<\/code> to a fully connected layer with 2*3*4=24 input neurons and  2*3*8=48 output neurons then reshape it again to <code>[2,3,8]<\/code>, or would it create 2*3=6 fully connected layers with 4 input and 8 output neurons then concatenate them?<\/p>\n","answers":[{"AnswerId":"52881462","CreationDate":"2018-10-18T19:48:18.827","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As for the Keras Dense layer, it has been already mentioned in <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">another answer<\/a> that its input is not flattened and instead, it is applied on the last axis of its input.<\/p>\n\n<p>As for the TensorFlow Dense layer, it is actually <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c19e29306ce1777456b2dbb3a14f511edf7883a8\/tensorflow\/python\/layers\/core.py#L32\" rel=\"nofollow noreferrer\">inherited from Keras Dense layer<\/a> and as a result, same as Keras Dense layer, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c19e29306ce1777456b2dbb3a14f511edf7883a8\/tensorflow\/python\/keras\/layers\/core.py#L916\" rel=\"nofollow noreferrer\">it is applied on the last axis<\/a> of its input.<\/p>\n"}]}
+{"QuestionId":52874113,"AnswerCount":0,"Tags":"<python><neural-network><keras><deep-learning>","CreationDate":"2018-10-18T12:32:23.830","AcceptedAnswerId":null,"OwnerUserId":10125598.0,"Title":"how to print confusion matrix of keras Multi output?","Body":"<p>I have one question.\nI want to print a confusion matrix.\nmy model is functional api of keras.\nand \nmodel = Model(inputs=[data_input], outputs=[output_1, output_2])<\/p>\n\n<p>output_1 = 9 classes\noutput_2 = 5 classes<\/p>\n\n<p><strong>My multi-classification model<\/strong><\/p>\n\n<pre><code>data_input = Input(shape=(trainX.shape[1], trainX.shape[2]))\n\nConv1 = Conv1D(filters=50, kernel_size=4, padding='valid',  activation='relu', strides=1)(data_input)\nConv1 = MaxPooling1D(pool_size=2)(Conv1)\n\nConv2 = Conv1D(filters=50, kernel_size=4, padding='valid', activation='relu', strides=1)(Conv1)\nConv2 = MaxPooling1D(pool_size=2)(Conv2)\n\nConv3 = Conv1D(filters=50, kernel_size=4, padding='valid', activation='relu', strides=1)(Conv2)\nConv3 = MaxPooling1D(pool_size=2)(Conv3)\n\nClassification1 = LSTM(128, input_shape=(47, 50), return_sequences=False)(Conv3)\nClassification2 = GRU(128, input_shape=(47, 50), return_sequences=False)(Conv3)\n\nactivity = Dense(9)(Classification1)\nactivity = Activation('softmax')(activity)\n\nspeed = Dense(5)(Classification2)\nspeed = Activation('softmax')(speed)\n\nmodel = Model(inputs=[data_input], outputs=[activity, speed])\n\nmodel.compile(loss= 'categorical_crossentropy' , optimizer='adam', metrics=[ 'accuracy' ])\nprint(model.summary())\n\nhistory = model.fit(trainX, {'activation_1': trainY_Activity, 'activation_2': trainY_Speed},\n          validation_data=(testX, {'activation_1': testY_Activity, 'activation_2': testY_Speed}),\n          epochs=epochs, batch_size=batch_size, verbose=1, shuffle=False)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52874153,"AnswerCount":3,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2018-10-18T12:34:01.370","AcceptedAnswerId":null,"OwnerUserId":10523917.0,"Title":"Human Gender Classification- Train and Val accuracy not moving","Body":"<p>I am having 0.3 million image in my Train set - Male\/Female and around ~50K image in the test set - Male\/Female . I am using below to work , also tried to add few more layers and more units . Also, I am doing data augmentation and others provided from keras docs. <\/p>\n\n<pre><code>targetSize =64\nclassifier.add(Conv2D(filters = 32,kernel_size =(3,3),input_shape=(targetSize,targetSize,3),activation ='relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(filters = 32,kernel_size =(3,3),activation ='relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(filters = 32,kernel_size =(3,3),activation ='relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(filters = 32,kernel_size =(3,3),activation ='relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Flatten())\nclassifier.add(Dropout(rate = 0.6))\nclassifier.add(Dense(units = 64, activation='relu'))\nclassifier.add(Dropout(rate = 0.5))\nclassifier.add(Dense(units = 64, activation='relu'))\nclassifier.add(Dropout(rate = 0.2))\nclassifier.add(Dense(units = 1,activation='sigmoid')\n\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n# Part 2 - Fitting the CNN to the images\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\n                     shear_range = 0.2,\n                     zoom_range = 0.2,\n                     height_shift_range = 0.2,\n                     width_shift_range = 0.2,\n                     horizontal_flip = True)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ntraining_set = train_datagen.flow_from_directory('&lt;train_folder_loc&gt;',\n                                                 target_size = (img_size, img_size),\n                                                 batch_size = batch_size_train,\n                                                 class_mode = 'binary')\ntest_set = test_datagen.flow_from_directory('&lt;test_folder_loc&gt;',\n                                            target_size = (img_size, img_size),\n                                            batch_size = batch_size_test,\n                                            class_mode = 'binary')\nclassifier.fit_generator(training_set,\n                         steps_per_epoch = &lt;train_image_count&gt;\/batch_size_train,\n                         epochs = n_epoch,\n                         validation_data = test_set,\n                         validation_steps = &lt;test_image_count&gt;\/batch_size_test,\n                         use_multiprocessing = True,\n                         workers=&lt;mycpu&gt;)\n<\/code><\/pre>\n\n<p>But with many combinations tried I am getting result like below , train acc and val acc is not moving ahead . I tried till 100 epoch and its almost like same.<\/p>\n\n<pre><code>11112\/11111 [==============================] - 156s 14ms\/step - loss: 0.5628 - acc: 0.7403 - val_loss: 0.6001 - val_acc: 0.6967\nEpoch 2\/25\n11112\/11111 [==============================] - 156s 14ms\/step - loss: 0.5516 - acc: 0.7403 - val_loss: 0.6096 - val_acc: 0.6968\nEpoch 3\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5472 - acc: 0.7404 - val_loss: 0.5837 - val_acc: 0.6967\nEpoch 4\/25\n11112\/11111 [==============================] - 155s 14ms\/step - loss: 0.5437 - acc: 0.7408 - val_loss: 0.5850 - val_acc: 0.6978\nEpoch 5\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5409 - acc: 0.7417 - val_loss: 0.5844 - val_acc: 0.6991\nEpoch 6\/25\n11112\/11111 [==============================] - 155s 14ms\/step - loss: 0.5386 - acc: 0.7420 - val_loss: 0.5828 - val_acc: 0.7011\nEpoch 7\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5372 - acc: 0.7427 - val_loss: 0.5856 - val_acc: 0.6984\nEpoch 8\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5347 - acc: 0.7437 - val_loss: 0.5847 - val_acc: 0.7017\nEpoch 9\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5331 - acc: 0.7444 - val_loss: 0.5770 - val_acc: 0.7017\nEpoch 10\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5323 - acc: 0.7443 - val_loss: 0.5803 - val_acc: 0.7037\nEpoch 11\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5309 - acc: 0.7453 - val_loss: 0.5877 - val_acc: 0.7018\nEpoch 12\/25\n11112\/11111 [==============================] - 155s 14ms\/step - loss: 0.5294 - acc: 0.7454 - val_loss: 0.5774 - val_acc: 0.7037\nEpoch 13\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5282 - acc: 0.7464 - val_loss: 0.5807 - val_acc: 0.7024\nEpoch 14\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5276 - acc: 0.7467 - val_loss: 0.5815 - val_acc: 0.7033\nEpoch 15\/25\n11112\/11111 [==============================] - 156s 14ms\/step - loss: 0.5269 - acc: 0.7474 - val_loss: 0.5753 - val_acc: 0.7038\nEpoch 16\/25\n11112\/11111 [==============================] - 154s 14ms\/step - loss: 0.5263 - acc: 0.7477 - val_loss: 0.5825 - val_acc: 0.7039\nEpoch 17\/25\n11112\/11111 [==============================] - 155s 14ms\/step - loss: 0.5249 - acc: 0.7485 - val_loss: 0.5821 - val_acc: 0.7037\n<\/code><\/pre>\n\n<p>I need your suggestion on this or any snippet to try .<\/p>\n","answers":[{"AnswerId":"52886634","CreationDate":"2018-10-19T06:00:13.993","ParentId":null,"OwnerUserId":"1710566","Title":null,"Body":"<p>Try and understand some of the basic principles of CNNs and how they are constructed; implement a simple one which works before arbitrarily putting in your own parameters.<\/p>\n\n<p>For example, typically the number of filters in successive convolutions increases in powers of two (e.g. 32, 64, 128 etc). Your use of dropout also is questionable, 0.6 is very high, not to mention stacking the three dropouts like you have doesn't make any sense.<\/p>\n"},{"AnswerId":"52877216","CreationDate":"2018-10-18T15:12:39.330","ParentId":null,"OwnerUserId":"7812314","Title":null,"Body":"<ol>\n<li><p>Make sure you are overfitting on a small sample before trying to extend the network. <\/p><\/li>\n<li><p>I would remove some\/all of the <code>Dropout<\/code> layers and see if it improves performance. I think 3 Dropout layers is quite high.<\/p><\/li>\n<li><p>Try reducing the learning rate.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"52875024","CreationDate":"2018-10-18T13:19:56.387","ParentId":null,"OwnerUserId":"8262615","Title":null,"Body":"<p>Hmm if you look at it closely, its not that its not moviing. it is moving a bit. There are times when models only get better at a certain point no matter how long you train it, or even how much more layers you add. When that happens, it all boils down to the data. I think it would be best to determine what is hindering your model to improve. Also, my friend, training a good model doesn't happen overnight specially with real world data, much more with complex data such as images of humans. <\/p>\n\n<p>I guess, if you are just following a tutorial which has achieved a better score than yours, you could check the version of packages their using, the data that you have, the steps they took and much more importantly the re run the model. There are instances where models could get different scores on different instances of training. <\/p>\n\n<p>I suggest you should try playing with the layers more, or even use a different type of Neural Network. If not, you should try playing with your data more. 300k images are a lot but when it comes to image classification, it could be really hard. <\/p>\n\n<p>Finally, I guess you could look into <a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">transfer learning<\/a> by tensorflow. You can read about it there. It works by retraining pre-made image recognition models. Keras has a <a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"nofollow noreferrer\">tutorial<\/a> on Transfer learning too. <\/p>\n"}]}
+{"QuestionId":52874480,"AnswerCount":1,"Tags":"<python><gpu><pytorch>","CreationDate":"2018-10-18T12:51:16.410","AcceptedAnswerId":null,"OwnerUserId":4313064.0,"Title":"is it possible make cuda deterministic?","Body":"<p>I am trying to refactor an application and testing new against old version by comparing outputs. At the first pytorch conv layer using identical input I get identical output when using CPU. However using cuda on GPU the output tensors differ in the 5th decimal place on some cells. Is there any way to eliminate this difference? I am already setting:<\/p>\n\n<pre><code>random.seed(seed)\nnp.random.seed(seed)\ntorch.manual_seed(seed)\ntorch.cuda.manual_seed(seed)\ntorch.backends.cudnn.deterministic = True\n<\/code><\/pre>\n","answers":[{"AnswerId":"56383826","CreationDate":"2019-05-30T18:34:38.187","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>Use <code>torch.backends.cudnn.benchmark = False<\/code><\/p>\n"}]}
+{"QuestionId":52874647,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T13:00:04.570","AcceptedAnswerId":52876786.0,"OwnerUserId":5623899.0,"Title":"Tensorflow v1.10+ why is an input serving receiver function needed when checkpoints are made without it?","Body":"<p>I'm in the process of adapting my model to TensorFlow's estimator API. <\/p>\n\n<p>I recently asked a question regarding <a href=\"https:\/\/stackoverflow.com\/questions\/52641737\/tensorflow-1-10-custom-estimator-early-stopping-with-train-and-evaluate\/52642619#52642619\">early stopping based on validation data<\/a> where in addition to early stopping, the best model at this point should be exported. <\/p>\n\n<p>It seems that my understanding of what a model export is and what a checkpoint is is not complete. <\/p>\n\n<p>Checkpoints are made automatically. From my understanding, the checkpoints are sufficient for the estimator to start \"warm\" - either using so per-trained weights or weights prior to an error (e.g. if you experienced a power outage). <\/p>\n\n<p>What is nice about checkpoints is that I do not have to write any code besides what is necessary for a custom estimator (namely, <code>input_fn<\/code> and <code>model_fn<\/code>). <\/p>\n\n<p>While, given an initialized estimator, one can just call its <code>train<\/code> method to train the model, in practice this method is rather lackluster. Often one would like to do several things:<\/p>\n\n<ol>\n<li>compare the network periodically to a validation dataset to ensure you are not over-fitting<\/li>\n<li>stop the training early if over-fitting occurs <\/li>\n<li>save the best model whenever the network finishes (either by hitting the specified number of training steps or by the early stopping criteria). <\/li>\n<\/ol>\n\n<p>To someone new to the \"high level\" estimator API, a lot of low level expertise seems to be required (e.g. for the <code>input_fn<\/code>) as how one could get the estimator to do this is not straight forward.<\/p>\n\n<p>By some light code reworking <strong>#1<\/strong> can be achieved by using <code>tf.estimator.TrainSpec<\/code> and <code>tf.estimator.EvalSpec<\/code> with <code>tf.estimator.train_and_evaluate<\/code>.<\/p>\n\n<p>In the <a href=\"https:\/\/stackoverflow.com\/questions\/52641737\/tensorflow-1-10-custom-estimator-early-stopping-with-train-and-evaluate\/52642619#52642619\">previous question<\/a> user <strong>@GPhilo<\/strong> clarifies how <strong>#2<\/strong> can be achieved by using a semi-unintuitive function from the <code>tf.contrib<\/code>:<\/p>\n\n<pre><code>tf.contrib.estimator.stop_if_no_decrease_hook(my_estimator,'my_metric_to_monitor', 10000)\n<\/code><\/pre>\n\n<p>(unintuitive as \"the early stopping is not triggered according to the number of non-improving evaluations, but to the number of non-improving evals in a certain step range\"). <\/p>\n\n<p><strong>@GPhilo<\/strong> - noting that it is unrelated to <strong>#2<\/strong> - also answered how to do <strong>#3<\/strong> (as requested in the original post). Yet, I do not understand what an <code>input_serving_fn<\/code> is, why it is needed, or how to make it.<\/p>\n\n<p>This is further confusing to me as no such function is needed to make checkpoints, or for the estimator to start \"warm\" from the checkpoint.<\/p>\n\n<p>So my questions are:<\/p>\n\n<ul>\n<li>what is the difference between a checkpoint and an exported best model?<\/li>\n<li>what exactly is a serving input receiver function and how to write one? (I have spent a bit of time reading over the tensorflow docs and do not find it sufficient to understand how I should write one, and why I even have to). <\/li>\n<li>how can I train my estimator, save the best model, and then later load it. <\/li>\n<\/ul>\n\n<p>To aid in answering my question I am providing this <a href=\"https:\/\/colab.research.google.com\/drive\/16G9jEwSCoLJ1MvkLzZGRY8YAfalV5qOT\" rel=\"noreferrer\">Colab<\/a> document.<\/p>\n\n<p>This self contained notebook produces some dummy data, saves it in TF Records, has a very simple custom estimator via <code>model_fn<\/code> and trains this model with an <code>input_fn<\/code> that uses the TF Record files. Thus it should be sufficient for someone to explain to me what placeholders I need to make for the input serving receiver function and and how I can accomplish <strong>#3<\/strong>.<\/p>\n\n<h1>Update<\/h1>\n\n<p><strong>@GPhilo<\/strong> foremost I can not understate my appreciation for you thoughtful consideration and care in aiding me (and hopefully others) understand this matter.<\/p>\n\n<p>My \u201cgoal\u201d (motivating me to ask this question) is to try and build a reusable framework for training networks so I can just pass a different <code>build_fn<\/code> and go (plus have the quality of life features of exported model, early stopping, etc).  <\/p>\n\n<p>An updated (based off your answers) <a href=\"https:\/\/colab.research.google.com\/drive\/10VKcE1sZh8F2-HYASxx7OgV8Mf4DQKFy\" rel=\"noreferrer\">Colab<\/a>  can be found <a href=\"https:\/\/colab.research.google.com\/drive\/10VKcE1sZh8F2-HYASxx7OgV8Mf4DQKFy\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>After several readings of your answer, I have found now some more confusion:<\/p>\n\n<p>1.<\/p>\n\n<blockquote>\n  <p>the way you provide input to the inference model is different than the one you use for the training<\/p>\n<\/blockquote>\n\n<p>Why? To my understanding the data input pipeline is not:<\/p>\n\n<pre><code>load raw \u2014&gt; process \u2014&gt; feed to model\n<\/code><\/pre>\n\n<p>But rather:<\/p>\n\n<pre><code>Load raw \u2014&gt; pre process \u2014&gt; store (perhaps as tf records)\n# data processing has nothing to do with feeding data to the model?\nLoad processed \u2014&gt; feed to model\n<\/code><\/pre>\n\n<p>In other words, it is my understanding (perhaps wrongly) that the point of a tf <code>Example<\/code> \/ <code>SequenceExample<\/code> is to store a complete singular datum entity ready to go - no other processing needed other than reading from the <code>TFRecord<\/code> file.<\/p>\n\n<p>\u2028Thus there can be a difference between the training \/ evaluation <code>input_fn<\/code> and the inference one (e.g. reading from file vs eager \/ interactive evaluation of in memory), but the data format is the same (except for inference you might want to feed only 1 example rather than a batch\u2026)<\/p>\n\n<p>I agree that the \u201c<em>input pipeline is not part of the model itself<\/em>\u201d. However, in my mind, and I am apparently wrong in thinking so, with the estimator I should be able to feed it a batch for training and a single example (or batch) for inference. <\/p>\n\n<p>An aside: \u201c<em>When evaluating, you don't need the gradients and you need a different input function.<\/em> \u201c, the only difference (at least in my case) is the files from which you reading?<\/p>\n\n<ol start=\"2\">\n<li>I am familiar with that TF Guide, but I have not found it useful because it is unclear to me what placeholders I need to add and what additional ops needed to be added to convert the data.<\/li>\n<\/ol>\n\n<p>What if I train my model with records and want to inference with just the dense tensors?<\/p>\n\n<p>Tangentially, I find the <em>example<\/em> in the linked guide subpar, given the tf record interface requires the user to define multiple times how to write to \/ extract features from a tf record file in different contexts.  Further, given that the TF team has explicitly stated they have little interest in documenting tf records, any documentation built on top of it, to me, is therefore equally unenlightening.<\/p>\n\n<ol start=\"3\">\n<li><p>Regarding <code>tf.estimator.export.build_raw_serving_input_receiver_fn<\/code>.\nWhat is the placeholder called? Input?  Could you perhaps show the analog of <code>tf.estimator.export.build_raw_serving_input_receiver_fn<\/code> by writing the equivalent <code>serving_input_receiver_fn<\/code><\/p><\/li>\n<li><p>Regarding your example <code>serving_input_receiver_fn<\/code> with the input images. How do you know to call features \u2018images\u2019 and the receiver tensor \u2018input_data\u2019 ? Is that (the latter) standard?<\/p><\/li>\n<li><p>How to name an export with <code>signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY<\/code>.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"52876786","CreationDate":"2018-10-18T14:51:04.147","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<h2>What is the difference between a checkpoint and an exported best model?<\/h2>\n\n<p>A checkpoint is, at its minimum, a file containing the values of all the variables <em>of a specific graph<\/em> taken at a <em>specific time point<\/em>.\nBy specific graph I mean that when loading back your checkpoint, what TensorFlow does is loop through all the variables defined in your graph (the one in the <code>session<\/code> you're running) and search for a variable in the checkpoint file <strong>that has the same name as the one in the graph<\/strong>. For resuming training, this is ideal because your graph will always look the same between restarts.<\/p>\n\n<p>An exported model serves a different purpose. The idea of an exported model is that, once you're done training, you want to get something you can use for inference that doesn't contain all the (heavy) parts that are specific to training (some examples: gradient computation, global step variable, input pipeline, ...).\nMoreover, and his is the key point, typically <strong>the way you provide input to the inference model is different than the one you use for the training.<\/strong> For training, you have an input pipeline that loads, preprocess and feeds data to your network. This input pipeline is not part of the model itself and may have to be altered for inference. This is a key point when operating with <code>Estimator<\/code>s.<\/p>\n\n<h2>Why do I need a serving input receiver function?<\/h2>\n\n<p>To answer this I'll take first a step back. <strong>Why do we need input functions at all ad what are they?<\/strong> TF's <code>Estimator<\/code>s, while perhaps not as intuitive as other ways to model networks, have a great advantage: they clearly separate between <em>model<\/em> logic and <em>input processing<\/em> logic by means of input functions and model functions.<\/p>\n\n<p>A model lives in 3 different phases: Training, Evaluation and Inference. For the most common use-cases (or at least, all I can think of at the moment), the <em>graph<\/em> running in TF will be different in all these phases. The graph is the combination of input preprocessing, model and all the machinery necessary to run the model in the current phase.<\/p>\n\n<p>A few examples to hopefully clarify further: When training, you need gradients to update the weights, an optimizer that runs the training step, metrics of all kinds to monitor how things are going, an input pipeline that grabs data from the training set, etc. When evaluating, you don't need the gradients and you need a different input function. When you are inferencing, all you need is the <em>forward<\/em> part of the model and again the input function will be different (no <code>tf.data.*<\/code> stuff but typically just a placeholder).<\/p>\n\n<p>Each of these phases in <code>Estimator<\/code>s has its own input function. You're familiar with the training and evaluation ones, the inference one is simply your <code>serving input receiver<\/code> function. In TF lingo, \"serving\" is the process of packing a trained model and using it for inference (there's a whole TensorFlow serving system for large-scale operation but that's beyond this question and you most likely won't need it anyhow).<\/p>\n\n<p>Time to quote a <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#using_savedmodel_with_estimators\" rel=\"noreferrer\">TF guide on the topic<\/a>:<\/p>\n\n<blockquote>\n  <p>During training, an input_fn() ingests data and prepares it for use by\n  the model. At serving time, similarly, a <code>serving_input_receiver_fn()<\/code>\n  accepts inference requests and prepares them for the model. This\n  function has the following purposes:<\/p>\n  \n  <ul>\n  <li>To add placeholders to the graph that the serving system will feed\n  with inference requests. <\/li>\n  <li>To add any additional ops needed to convert\n  data from the input format into the feature Tensors expected by the\n  model.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Now, the serving input function specification depends on how you plan of sending input to your graph.<\/p>\n\n<p>If you're going to pack the data in a (serialized) <strong><code>tf.Example<\/code><\/strong> (which is similar to one of the records in your TFRecord files), your serving input function will have a string placeholder (that's for the serialized bytes for the example) and will need a specification of how to interpret the example in order to extract its data. If this is the way you want to go I invite you to have a look at the example in the linked guide above, it essentially shows how you setup the specification of how to interpret the example and parse it to obtain the input data.<\/p>\n\n<p>If, instead, you're planning on <strong>directly feeding input to the first layer of your network<\/strong> you <em>still<\/em> need to define a serving input function, but this time it will only contain a placeholder that will be plugged directly into the network. TF offers a function that does just that: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/build_raw_serving_input_receiver_fn\" rel=\"noreferrer\"><code>tf.estimator.export.build_raw_serving_input_receiver_fn<\/code><\/a>. <\/p>\n\n<p>So, <strong>do you actually need to write your own input function?<\/strong> IF al you need is a placeholder, no. Just use <code>build_raw_serving_input_receiver_fn<\/code> with the appropriate parameters. IF you need fancier preprocessing, then yes, you might need to write your own. In that case, it would look something like this:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  \"\"\"For the sake of the example, let's assume your input to the network will be a 28x28 grayscale image that you'll then preprocess as needed\"\"\"\n  input_images = tf.placeholder(dtype=tf.uint8,\n                                         shape=[None, 28, 28, 1],\n                                         name='input_images')\n  # here you do all the operations you need on the images before they can be fed to the net (e.g., normalizing, reshaping, etc). Let's assume \"images\" is the resulting tensor.\n\n  features = {'input_data' : images} # this is the dict that is then passed as \"features\" parameter to your model_fn\n  receiver_tensors = {'input_data': input_images} # As far as I understand this is needed to map the input to a name you can retrieve later\n  return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n<\/code><\/pre>\n\n<h2>How can I train my estimator, save the best model, and then later load it?<\/h2>\n\n<p>Your <code>model_fn<\/code> takes the <code>mode<\/code> parameter in order for you to build conditionally the model. In your colab, you always have a optimizer, for example. This is wrong ,as it should only be there for <code>mode == tf.estimator.ModeKeys.TRAIN<\/code>.<\/p>\n\n<p>Secondly, your <code>build_fn<\/code> has an \"outputs\" parameter that is meaningless. This function should represent your inference graph, take as input only the tensors you'll fed to it in the inference and return the logits\/predictions. \nI'll thus assume the <code>outputs<\/code> parameters is not there as the <code>build_fn<\/code> signature should be <code>def build_fn(inputs, params)<\/code>.<\/p>\n\n<p>Moreover, you define your <code>model_fn<\/code> to take <code>features<\/code> as a tensor. While this can be done, it both limits you to having exactly one input and complicates things for the serving_fn (you can't use the canned <code>build_raw_...<\/code> but need to write your own and return a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/TensorServingInputReceiver\" rel=\"noreferrer\"><code>TensorServingInputReceiver<\/code><\/a> instead). I'll choose the more generic solution and assume your <code>model_fn<\/code> is as follows (I omit the variable scope for brevity, add it as necessary):<\/p>\n\n<pre><code>def model_fn(features, labels, mode, params): \n  my_input = features[\"input_data\"]\n  my_input.set_shape(I_SHAPE(params['batch_size']))\n\n  # output of the network\n  onet = build_fn(features, params)\n  predicted_labels = tf.nn.sigmoid(onet)\n  predictions = {'labels': predicted_labels, 'logits': onet}\n  export_outputs = { # see EstimatorSpec's docs to understand what this is and why it's necessary.\n       'labels': tf.estimator.export.PredictOutput(predicted_labels),\n       'logits': tf.estimator.export.PredictOutput(onet) \n  } \n  # NOTE: export_outputs can also be used to save models as \"SavedModel\"s during evaluation.\n\n  # HERE is where the common part of the graph between training, inference and evaluation stops.\n  if mode == tf.estimator.ModeKeys.PREDICT:\n    # return early and avoid adding the rest of the graph that has nothing to do with inference.\n    return  tf.estimator.EstimatorSpec(mode=mode, \n                                       predictions=predictions, \n                                       export_outputs=export_outputs)\n\n  labels.set_shape(O_SHAPE(params['batch_size']))      \n\n  # calculate loss \n  loss = loss_fn(onet, labels)\n\n  # add optimizer only if we're training\n  if mode == tf.estimator.ModeKeys.TRAIN:\n    optimizer = tf.train.AdagradOptimizer(learning_rate=params['learning_rate'])\n  # some metrics used both in training and eval\n  mae = tf.metrics.mean_absolute_error(labels=labels, predictions=predicted_labels, name='mea_op')\n  mse = tf.metrics.mean_squared_error(labels=labels, predictions=predicted_labels, name='mse_op')\n  metrics = {'mae': mae, 'mse': mse}\n  tf.summary.scalar('mae', mae[1])\n  tf.summary.scalar('mse', mse[1])\n\n  if mode == tf.estimator.ModeKeys.EVAL:\n    return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=metrics, predictions=predictions, export_outputs=export_outputs)\n\n  if mode == tf.estimator.ModeKeys.TRAIN:\n    train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n    return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op, eval_metric_ops=metrics, predictions=predictions, export_outputs=export_outputs)\n<\/code><\/pre>\n\n<p>Now, to set up the exporting part, after your call to <code>train_and_evaluate<\/code> finished:<\/p>\n\n<p>1) Define your serving input function:<\/p>\n\n<pre><code>serving_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(\n                                       {'input_data':tf.placeholder(tf.float32, [None,#YOUR_INPUT_SHAPE_HERE (without batch size)#])})\n<\/code><\/pre>\n\n<p>2) Export the model to some folder<\/p>\n\n<pre><code>est.export_savedmodel('my_directory_for_saved_models', serving_fn)\n<\/code><\/pre>\n\n<p>This will save the <strong>current<\/strong> state of the estimator to wherever you specified. If you want a specifc checkpoint, load it before calling <code>export_savedmodel<\/code>.\nThis will save in \"my_directory_for_saved_models\" a <strong>prediction<\/strong> graph with the trained parameters that the estimator had when you called the export function.<\/p>\n\n<p>Finally, you might want t freeze the graph (look up <code>freeze_graph.py<\/code>) and optimize it for inference (look up <code>optimize_for_inference.py<\/code> and\/or <code>transform_graph<\/code>) obtaining a frozen <code>*.pb<\/code> file you can then load and use for inference as you wish.<\/p>\n\n<hr>\n\n<h2>Edit: Adding answers to the new questions in the update<\/h2>\n\n<h3>Sidenote:<\/h3>\n\n<blockquote>\n  <p>My \u201cgoal\u201d (motivating me to ask this question) is to try and build a\n  reusable framework for training networks so I can just pass a\n  different build_fn and go (plus have the quality of life features of\n  exported model, early stopping, etc).<\/p>\n<\/blockquote>\n\n<p>By all means, if you manage, please post it on GitHub somewhere and link it to me. I've been trying to get just the same thing up and running for a while now and the results are not quite as good as I'd like them to be.<\/p>\n\n<h3>Question 1:<\/h3>\n\n<blockquote>\n  <p>In other words, it is my understanding (perhaps wrongly) that the\n  point of a tf Example \/ SequenceExample is to store a complete\n  singular datum entity ready to go - no other processing needed other\n  than reading from the TFRecord file.<\/p>\n<\/blockquote>\n\n<p>Actually, this is typically not the case (although, your way is <em>in theory<\/em> perfectly fine too).\nYou can see TFRecords as a (awfully documented) way to store a dataset in a compact way. For image datasets for example, a record typically contains the <strong>compressed<\/strong> image data (as in, the bytes composing a jpeg\/png file), its label and some meta information. Then the input pipeline reads a record, decodes it, preprocesses it as needed and feeds it to the network. Of course, you can move the decoding and preprocessing before the generation of the TFRecord dataset and store in the examples the ready-to-feed data, but the size blowup of your dataset will be huge.<\/p>\n\n<p>The specific preprocessing pipeline is one example what changes between phases (for example, you might have data augmentation in the training pipeline, but not in the others). Of course, there are cases in which these pipelines are the same, but in general this is not true.<\/p>\n\n<p>About the aside:<\/p>\n\n<blockquote>\n  <p>\u201cWhen evaluating, you don't need the gradients and you need a\n  different input function. \u201c, the only difference (at least in my case)\n  is the files from which you reading?<\/p>\n<\/blockquote>\n\n<p>In your case that may be. But again, assume you're using data augmentation: You need to disable it (or, better, don't have it at all) during eval and this alters your pipeline.<\/p>\n\n<h3>Question 2: What if I train my model with records and want to inference with just the dense tensors?<\/h3>\n\n<p>This is precisely why you separate the pipeline from the model.\nThe model takes as input a tensor and operates on it. Whether that tensor is a placeholder or is the output of a subgraph that converts it from an Example to a tensor, that's a detail that belongs to the framework, not to the model itself.<\/p>\n\n<p>The splitting point is the model input. The model expects a tensor (or, in the more generic case, a dict of <code>name:tensor<\/code> items) as input and uses that to build its computation graph. Where that input comes from is decided by the input functions, but as long as the output of all input functions has the same interface, one can swap inputs as needed and the model will simply take whatever it gets and use it.<\/p>\n\n<p>So, to recap, assuming you train\/eval with Examples and predict with dense tensors, your train and eval input functions will set up a pipeline that reads examples from somewhere, decodes them into tensors and returns those to the model to use as inputs. Your predict input function, on the other hand, just sets up one placeholder per input of your model and returns them to the model, because it assumes you'll put in the placeholders the data ready to be fed to the network.<\/p>\n\n<h3>Question 3:<\/h3>\n\n<p>You pass the placeholder as a parameter of <code>build_raw_serving_input_receiver_fn<\/code>, so you choose its name:<\/p>\n\n<pre><code>tf.estimator.export.build_raw_serving_input_receiver_fn(                                               \n    {'images':tf.placeholder(tf.float32, [None,28,28,1], name='input_images')})\n<\/code><\/pre>\n\n<h3>Question 4:<\/h3>\n\n<p>There was a mistake in the code (I had mixed up two lines), the dict's key should have been <code>input_data<\/code> (I amended the code above).\nThe key in the dict has to be <strong>the key you use to retrieve the tensor from <code>features<\/code> in your <code>model_fn<\/code><\/strong>. In <code>model_fn<\/code> the first line is: <\/p>\n\n<pre><code>my_input = features[\"input_data\"]\n<\/code><\/pre>\n\n<p>hence the key is <code>'input_data'<\/code>.\nAs per the key in <code>receiver_tensor<\/code>, I'm still not quite sure what role that one has, so my suggestion is try setting a different name than the key in <code>features<\/code> and check where the name shows up.<\/p>\n\n<h3>Question 5:<\/h3>\n\n<p>I'm not sure I understand, I'll edit this after some clarification<\/p>\n"}]}
+{"QuestionId":52874908,"AnswerCount":1,"Tags":"<python><web><tensorflow><flask><websocket>","CreationDate":"2018-10-18T13:14:16.193","AcceptedAnswerId":52875552.0,"OwnerUserId":10192915.0,"Title":"im trying to run app after installing requirement.txt from?","Body":"<p>I'm trying to run app after installing <code>requirement.txt<\/code> from : <a href=\"https:\/\/github.com\/victordibia\/skyfall\" rel=\"nofollow noreferrer\">https:\/\/github.com\/victordibia\/skyfall<\/a><\/p>\n\n<p>I am getting the following error:<\/p>\n\n<pre><code>File \"app.py\", line 19, in `&lt;module&gt;`\n\nfrom utils import web_socket_client\n\nFile \"C:\\Users\\anurag\\Downloads\\skyfall-master\\utils\\web_socket_client.py\", line 8, in &lt;module&gt;\n\nfrom websocket import WebSocketException, WebSocketConnectionClosedException ImportError: cannot import name 'WebSocketException'`\n<\/code><\/pre>\n\n<p>Is there another way of running the app..if needed make changes in the file and send the github link <\/p>\n","answers":[{"AnswerId":"52875552","CreationDate":"2018-10-18T13:48:51.080","ParentId":null,"OwnerUserId":"1692060","Title":null,"Body":"<p>I have no problems running the app on Debian Linux and Python 3.6.5. Perhaps you need to run it in the python virtual environment to make sure some installed package does not interfere. Instruction for Linux:<\/p>\n\n<ol>\n<li>Clone the git to the local folder:<\/li>\n<\/ol>\n\n<p><code>git clone https:\/\/github.com\/victordibia\/skyfall<\/code><\/p>\n\n<ol start=\"2\">\n<li>Create python3 virtual environment in the folder nearby:<\/li>\n<\/ol>\n\n<p>Install <code>python3-venv<\/code> package using package manager. Run:<\/p>\n\n<p><code>python3 -m venv skyfall_venv<\/code><\/p>\n\n<p>then\n<code>source skyfall_venv\/bin\/activate<\/code><\/p>\n\n<p>Now you are inside the virtual environment, which is indicated in your shell. Do<\/p>\n\n<p><code>(skyfall_venv): pip install --upgrade pip<\/code><\/p>\n\n<p>if it does not work, do <\/p>\n\n<p><code>(skyfall_venv): pip3 install --upgrade pip<\/code><\/p>\n\n<p>to have the latest pip in your venv.<\/p>\n\n<p>3.Then edit <code>skyfall\/requirements.txt<\/code> removing extra <code>-e<\/code> in front of git, it must look like the follownig:<\/p>\n\n<pre><code>Flask==0.12.2\nopencv_python==3.4.0.12\nnumpy==1.14.2\ntensorflow==1.7.0\nprotobuf==3.5.2.post1\nwebsocket_client==0.47.0\ngit+https:\/\/github.com\/dpallot\/simple-websocket-server.git\n<\/code><\/pre>\n\n<p>Run inside venv:<\/p>\n\n<p><code>(skyfall_venv): pip install -r requirements.txt<\/code> (or <code>pip3<\/code> if the former does not work)<\/p>\n\n<p>4.You should see no errors. The run the app in venv:<\/p>\n\n<p><code>(skyfall_venv): cd skyfall<\/code><\/p>\n\n<p><code>(skyfall_venv): python app.py<\/code>  (or python3 if the former does not work)<\/p>\n\n<p>I see the window with gesture recognition, and if I open web address in browser which the app suggests, I can play the game.<\/p>\n\n<p><code>cvcam.py<\/code>:<\/p>\n\n<pre><code>import cv2\n\ncap = cv2.VideoCapture(0)\n\n# Check if the webcam is opened correctly\nif not cap.isOpened():\n    raise IOError(\"Cannot open webcam\")\n\nwhile True:\n    ret, frame = cap.read()\n    frame = cv2.resize(frame, None, fx=0.5, fy=0.5, interpolation=cv2.INTER_AREA)\n    cv2.imshow('Input', frame)\n\n    c = cv2.waitKey(1)\n    if c == 27:\n        break\n\ncap.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52875630,"AnswerCount":3,"Tags":"<python><tensorflow><cv2>","CreationDate":"2018-10-18T13:52:45.503","AcceptedAnswerId":52875926.0,"OwnerUserId":10230588.0,"Title":"Resizing image with cv2","Body":"<p>I'm trying resize images retrieved from cifar10 in the original 32x32 to 96x96 for use with MobileNetV2, howevery I'm running into this error. Tried a variety of solutions but nothing seems to work.<\/p>\n\n<p>My code:<\/p>\n\n<pre><code>for a in range(len(train_images)):\n    train_images[a] = cv2.resize(train_images[a], dsize=(minSize, minSize), interpolation=cv2.INTER_CUBIC)\n<\/code><\/pre>\n\n<p>Error I'm getting:<\/p>\n\n<pre><code>----&gt; 8     train_images[a] = cv2.resize(train_images[a], dsize=(minSize, minSize), interpolation=cv2.INTER_CUBIC)\nValueError: could not broadcast input array from shape (96,96,3) into shape (32,32,3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52875926","CreationDate":"2018-10-18T14:09:10.370","ParentId":null,"OwnerUserId":"7994236","Title":null,"Body":"<p>This is simply because you are reading the 32x32 image from train_images and trying to save the reshaped image (96x96) in the same array which is impossible! \nTry something like:<\/p>\n\n<pre><code>train_images_reshaped = np.array((num_images, 96, 96, 3))\nfor a in range(len(train_images)):\n    train_images_reshaped[a] = cv2.resize(train_images[a], dsize=(minSize, minSize), interpolation=cv2.INTER_CUBIC)\n<\/code><\/pre>\n"},{"AnswerId":"52876087","CreationDate":"2018-10-18T14:16:41.557","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Sometimes you have to convert the image from RGB to grayscale. If that is the problem, the only thing you should do is <code>gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)<\/code>, <em>resize the image<\/em> and then again <code>resized_image = cv2.cvtColor(gray_image, cv2.COLOR_GRAY2RGB)<\/code><\/p>\n\n<p>I have never run into this error but if the first option doesn't work, you can try and resize image with pillow like this:<\/p>\n\n<pre><code>from PIL import Image\n\nim = Image.fromarray(cv2_image)\nnx, ny = im.size\nim2 = im.resize((nx*2, ny*2), Image.LANCZOS)\ncv2_image = cv2.cvtColor(numpy.array(im2), cv2.COLOR_RGB2BGR)\n<\/code><\/pre>\n\n<p>You can make this into a function and call it in the list comprehension. I hope this solves your problem :)<\/p>\n"},{"AnswerId":"55875808","CreationDate":"2019-04-26T22:56:22.577","ParentId":null,"OwnerUserId":"5079616","Title":null,"Body":"<p>There are some interpolation algorithms in OpenCV. Such as-<\/p>\n\n<ul>\n<li>INTER_NEAREST \u2013 a nearest-neighbor interpolation<\/li>\n<li>INTER_LINEAR \u2013 a bilinear interpolation (used by default)<\/li>\n<li>INTER_AREA \u2013 resampling using pixel area relation. It may be a\npreferred method for image decimation, as it gives moire\u2019-free\nresults. But when the image is zoomed, it is similar to the\nINTER_NEAREST method.<\/li>\n<li>INTER_CUBIC \u2013 a bicubic interpolation over 4\u00d74 pixel  neighborhood<\/li>\n<li>INTER_LANCZOS4 \u2013 a Lanczos interpolation over 8\u00d78 pixel  neighborhood<\/li>\n<\/ul>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre><code>image_scaled=cv2.resize(image,None,fx=.75,fy=.75,interpolation = cv2.INTER_LINEAR)\nimg_double=cv2.resize(image,None,fx=2,fy=2,interpolation=cv2.INTER_CUBIC)\nimage_resize=cv2.resize(image,(200,300),interpolation=cv2.INTER_AREA)\nimage_resize=cv2.resize(image,(500,400),interpolation=cv2.INTER_LANCZOS4)\n<\/code><\/pre>\n\n<p>You can find the details about <strong>python implementation<\/strong> here as well: <a href=\"https:\/\/www.life2coding.com\/how-to-resize-images-in-opencv-python-using-different-interpolation-methods\/\" rel=\"nofollow noreferrer\">How to resize images in OpenCV python<\/a><\/p>\n"}]}
+{"QuestionId":52875989,"AnswerCount":1,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-10-18T14:11:33.383","AcceptedAnswerId":null,"OwnerUserId":9913405.0,"Title":"Pytorch : significance of grad_h_relu.clone() in Backprop","Body":"<p>I am learning Pytorch, while I'm looking to the tutorial on their site I can't understand the significance of <code>grad_h = grad_h_relu.clone()<\/code>.<\/p>\n\n<pre><code>h = x.mm(w1)\nh_relu = h.clamp(min=0)\ny_pred = h_relu.mm(w2)\n\n# Compute and print loss\nloss = (y_pred - y).pow(2).sum().item()\nprint(t, loss)\n\n# Backprop to compute gradients of w1 and w2 with respect to loss\ngrad_y_pred = 2.0 * (y_pred - y)\ngrad_w2 = h_relu.t().mm(grad_y_pred)\ngrad_h_relu = grad_y_pred.mm(w2.t())\n\n\ngrad_h = grad_h_relu.clone() # what is the signifigance of this line?\n\n\ngrad_h[h &lt; 0] = 0\ngrad_w1 = x.t().mm(grad_h)\n\n# Update weights using gradient descent\nw1 -= learning_rate * grad_w1\nw2 -= learning_rate * grad_w2\n<\/code><\/pre>\n","answers":[{"AnswerId":"52897550","CreationDate":"2018-10-19T17:45:48.030","ParentId":null,"OwnerUserId":"5056575","Title":null,"Body":"<pre><code>grad_h = grad_h_relu.clone()\n<\/code><\/pre>\n\n<p>It means that you are making a copy of the gradient of relu such that it does not share the memory with the original <code>grad_h_relu<\/code>. Then you perform some operation on it. As they are stored in two different locations, changing the value of <code>grad_h<\/code> by the following operation will not affect the <code>grad_h_relu<\/code>.<\/p>\n\n<pre><code>grad_h[h &lt; 0] = 0\ngrad_w1 = x.t().mm(grad_h)\n<\/code><\/pre>\n\n<p>This <code>grad_w1<\/code> is needed to update your parameters of your network. <\/p>\n"}]}
+{"QuestionId":52876128,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T14:18:43.123","AcceptedAnswerId":52878510.0,"OwnerUserId":3853339.0,"Title":"transfer parameters between models in tensorflow slows down training time","Body":"<p>I've developed a model that requires me to have two versions of a model, one before the training step and one after. I thought I could simply do this using a tf.assign() method call but it seems that this has massively slowed down the training. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/37966924\/why-does-tf-assign-slow-the-execution-time\">Why does tf.assign() slow the execution time?<\/a><\/p>\n\n<p>This post asks a similar question however the author is simply trying to update the learning rate and can do so by just adding a feed_dict. However, calling tf.assign can't really be avoided in my case? The other solution involved separating the graph definition and the graph run but since I require both to be in the session since they need access to the parameters of other models I'm unsure how to do this.<\/p>\n\n<p>Any help is appreciated.<\/p>\n\n<p>Code is as simple as:<\/p>\n\n<pre><code>tf.assign(var[0], var[2])\ntf.assign(var[1], var[3])\n\nQ_agent.train(...)\n<\/code><\/pre>\n\n<p>and <code>var[0]<\/code> and <code>var[1]<\/code> are the params of the Q_agent. <\/p>\n\n<p>The training time is quite long in this case. I've adapted the code to try and use a tf.placeholder. The code is as follows:<\/p>\n\n<pre><code>var = tf.trainable_variables()\nparams = [var[4], var[5]]\nupdate_hidden = tf.placeholder(params[0].dtype, shape=params[0].get_shape())\nupdate_value = tf.placeholder(params[1].dtype, shape=params[1].get_shape())\n\nfor loop: \n\n    var = tf.trainable_variables()\n    old_hidden = var[0] \n    old_value = var[1]\n\n    new_hidden = var[2]\n    new_value = var[3]\n    update_h = old_hidden.assign(update_hidden)\n    update_v = old_value.assign(update_value)\n\n    sess.run([update_h, update_v], feed_dict={update_hidden:   new_hidden.eval(), update_value: new_value.eval()})\n<\/code><\/pre>\n\n<p>Though the train function now runs quickly, this hasn't improved the efficiency of the code because there's continuous slow down in performance when running <code>update_h<\/code> and <code>update_v<\/code>. Any ideas?<\/p>\n","answers":[{"AnswerId":"52878510","CreationDate":"2018-10-18T16:23:25.830","ParentId":null,"OwnerUserId":"3853339","Title":null,"Body":"<p>Solved. Key is to define tf.assign() so that it is called once and NOT in the training loop. Otherwise if you call it every time then this adds a new node to the graph and it means you have to do additional computation after every iteration. <\/p>\n\n<pre><code>var = tf.trainable_variables()\nold_hidden = var[0] \nold_value = var[1]\nupdate_h = old_hidden.assign(update_hidden)\nupdate_v = old_value.assign(update_value)\n\nfor loop:\n\n    # overwrite old_hidden and old_value here.\n    var = tf.trainable_variables()\n    old_hidden = var[0] \n    old_value = var[1]\n\n    new_hidden = var[2]\n    new_value = var[3]\n    # update_h = old_hidden.assign(update_hidden)\n    # update_v = old_value.assign(update_value)\n\n    sess.run([update_h, update_v], feed_dict={update_hidden:   new_hidden.eval(), update_value: new_value.eval()})\n<\/code><\/pre>\n\n<p>I'm 100% sure there's a tidier way of doing this but this is what I have!<\/p>\n"}]}
+{"QuestionId":52876496,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-18T14:37:12.290","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"ImportError: cannot import name 'ReLU'. in Keras 2.1.3","Body":"<p>Using Keras 2.1.3 I get <code>ImportError: cannot import name 'ReLU'<\/code> on <code>from keras.layers import ReLU<\/code>, what is the proper way to import <code>ReLU<\/code> in Keras 2.1.3?<\/p>\n\n<p>I can import <code>ReLU<\/code> in Keras 2.2.2:<\/p>\n\n<pre><code>   import keras as k\n   print(k.__version__)\n   from keras.layers import ReLU\n<\/code><\/pre>\n\n<p>Yes, I know about <code>Activation('relu')<\/code>, but I'm looking for <code>ReLU<\/code> function used in <a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/mobilenet.py#L386\" rel=\"nofollow noreferrer\">MobileNet implementation<\/a> <\/p>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>Looking at <code>mobilenet.py<\/code> of Keras 2.1.3, seems they use:<\/p>\n\n<pre><code>def relu6(x):\n    return K.relu(x, max_value=6)\n<\/code><\/pre>\n\n<p>According to this thread <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/7431\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/7431<\/a> in Keras 2.1.3 custom layers of mobilenet can be imported like:<\/p>\n\n<pre><code>from keras.applications.mobilenet import relu6\nfrom keras.applications.mobilenet import DepthwiseConv2D\n<\/code><\/pre>\n","answers":[{"AnswerId":"52877253","CreationDate":"2018-10-18T15:14:48.363","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>There is no <code>keras.layers.ReLU<\/code>, hence your inport error. Relu is an activation <code>keras.activations.relu(x, alpha=0.0, max_value=None, threshold=0.0)<\/code><\/p>\n\n<p>The code you referenced defines layers as:<\/p>\n\n<pre><code>global backend, layers, models, keras_utils\nbackend, layers, models, keras_utils = get_submodules_from_kwargs(kwargs)\n<\/code><\/pre>\n\n<p>I'm not quite sure what this does (it's defined <a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/61de21815728aa66377ebb2a7b4c2f5492a34dd2\/keras_applications\/__init__.py\" rel=\"nofollow noreferrer\">here<\/a>).<\/p>\n\n<p>I find this code confusing, because I often import layers using <code>from keras import layers<\/code>.<\/p>\n"}]}
+{"QuestionId":52877859,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T15:46:20.243","AcceptedAnswerId":null,"OwnerUserId":5577396.0,"Title":"Tensorflow - assertion failed: [predictions must be in [0, 1]]","Body":"<p>I am using the <code>Estimator<\/code> API of Tensorflow and I am running into the following problem. I want to check the f1 score other than the accuracy and when I am evaluating after training there are no problems at all, when I am testing it asks for normalized values, which I have already normalized. <\/p>\n\n<p>This is the model of my of the network (first part is omitted):<\/p>\n\n<pre><code>#### architecture omitted #####\n\npredictions = {\n        \"classes\": tf.argmax(input=logits, axis=1),\n        \"probabilities\": tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.cast(labels, tf.float32), logits=tf.cast(logits, tf.float32), name=\"sigmoid_tensor\")\n}\n\n\nif mode == tf.estimator.ModeKeys.PREDICT:\n    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)\n\nloss = tf.losses.sigmoid_cross_entropy(multi_class_labels=labels, logits=logits)\n\nif mode == tf.estimator.ModeKeys.TRAIN:\n    optimizer = tf.train.AdamOptimizer(learning_rate=0.001)\n    #optimizer = tf.train.MomentumOptimizer(learning_rate=0.01, momentum=0.96)\n    train_op = optimizer.minimize(\n                              loss=loss,\n                              global_step=tf.train.get_global_step())\n    logging_hook = tf.train.LoggingTensorHook({\"loss\" : loss}, every_n_iter=10)\n    return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op, training_hooks = [logging_hook])\n\n\neval_metric_ops = {\n    \"accuracy\": tf.metrics.accuracy(\n    labels=tf.argmax(input=labels, axis=1),\n    predictions=predictions[\"classes\"]),\n\n    \"f1 score\" : tf.contrib.metrics.f1_score(\n    labels = tf.argmax(input=labels, axis=1),\n    predictions = tf.cast(predictions[\"classes\"],tf.float32)\/tf.norm(tf.cast(predictions[\"classes\"], tf.float32)))\n}\n\nreturn tf.estimator.EstimatorSpec(mode=mode, loss=loss, eval_metric_ops=eval_metric_ops) \n<\/code><\/pre>\n\n<p>This is the training script I am using, I don't have any problems when evaluating (no errors here).<\/p>\n\n<pre><code>classifier = tf.estimator.Estimator(model_fn=instrument_recognition_model, model_dir=saved_model_path)\ntrain_input_fn = tf.estimator.inputs.numpy_input_fn(x=X_train, y=y_train, batch_size=16, num_epochs=30, shuffle=True)\nclassifier.train(input_fn=train_input_fn)\n\n# Evaluate results on the training set\neval_input_fn = tf.estimator.inputs.numpy_input_fn(x=X_eval,y=y_eval,num_epochs=1,shuffle=False)\neval_results = classifier.evaluate(input_fn=eval_input_fn)\nprint(eval_results)\n<\/code><\/pre>\n\n<p>This is my testing script, here the program fails:<\/p>\n\n<pre><code>classifier = tf.estimator.Estimator(model_fn=instrument_recognition_model, model_dir=saved_model_path)\n\n# Keep only certain samples\nkey_indices = [np.where(instruments == x)[0][0] for x in keys]\nexample_indices = np.array([])\nfor ind in key_indices:\n    tmp = np.argwhere(labels[:,ind] == True).flatten()\n    example_indices = np.union1d(example_indices, tmp).astype(np.int32)\n\nfeatures = features[example_indices].astype(np.float32)\nexample_indices = [[x for i in key_indices] for x in example_indices]\nlabels = labels[example_indices, key_indices].astype(np.int)\n\n# Evaluate results on the test set\nprint(features)\nprint(labels)\neval_input_fn = tf.estimator.inputs.numpy_input_fn(x=features, y=labels, batch_size=1, num_epochs=1, shuffle=False)\neval_results = classifier.evaluate(input_fn=eval_input_fn)\nprint(eval_results)\n<\/code><\/pre>\n\n<p>I have really not idea of what's wrong since I've follwed the same process for evaluation and testing. Without the f1 metric (only accuracy) everything works well, when I add the f1 metric, it fails in the test script.<\/p>\n\n<p>The fragment of the error is the following:<\/p>\n\n<pre><code>### trace error omitted ###\nFile \"..\/models.py\", line 207, in instrument_recognition_model\n    predictions = tf.cast(predictions[\"classes\"],tf.float32)\/tf.norm(tf.cast(predictions[\"classes\"], tf.float32)))\n### trace error ommitted ###\n\n    TheInvalidArgumentError (see above for traceback): assertion failed: [predictions must be in [0, 1]] [Condition x &lt;= y did not hold element-wise:x (div:0) = ] [nan] [y (f1\/Cast_1\/x:0) = ] [1]\n         [[Node: f1\/assert_less_equal\/Assert\/AssertGuard\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_FLOAT, DT_STRING, DT_FLOAT], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/Switch, f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/data_0, f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/data_1, f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/Switch_1, f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/data_3, f1\/assert_less_equal\/Assert\/AssertGuard\/Assert\/Switch_2)]]\n<\/code><\/pre>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"54934373","CreationDate":"2019-02-28T21:15:05.717","ParentId":null,"OwnerUserId":"1430342","Title":null,"Body":"<p>Because your data is probably multi-class. If you check official docs of f1-score of Tensorflow <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/metrics\/f1_score\" rel=\"nofollow noreferrer\">here<\/a>, you can see that it is implemented for binary classification only.<br>\nYou can do something like <a href=\"https:\/\/stackoverflow.com\/a\/43960730\/1430342\">this<\/a> if you really want f1-score.<\/p>\n"}]}
+{"QuestionId":52878027,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><regression>","CreationDate":"2018-10-18T15:56:17.893","AcceptedAnswerId":52878805.0,"OwnerUserId":5208232.0,"Title":"Keras regression prediction is not same dimension as output dimension","Body":"<p>Hello I'm trying to do Energy Disaggregation (predict the energy use of appliances while given the total energy consumption of a certain household.)<\/p>\n\n<p>Now I have an input dimension of 2 because of 2 main energy measurements.\nThe output dimension of the Keras Sequential model should be 18 because I have 18 appliances I would like to make a prediction for.\nI have enough data using the REDD dataset (this is no problem).<\/p>\n\n<p>I have trained the model and gained reasonable loss and accuracy. \nBut when I want to make a prediction for some test data, the prediction consists of values in a 1-dimensional array. Meanwhile the outputs are 18-dimensional?<\/p>\n\n<p>How is this possible or am I trying something that isn't really viable?<\/p>\n\n<p>Some code:<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Dense(HIDDEN_LAYER_NEURONS,input_dim=2))\nmodel.add(Activation('relu'))\nmodel.add(Dense(18))\n\nmodel.compile(loss=LOSS,\n          optimizer=OPTIMIZER,\n          metrics=['accuracy'])\n\nmodel.fit(X_train, y_train, epochs=EPOCHS, batch_size=BATCH_SIZE, \nverbose=1, validation_split=VALIDATION_SPLIT)\n\npred = model.predict(X_test).reshape(-1)\npred.shape # prints the following 1 dimensional array: (xxxxx,) dimensional\n<\/code><\/pre>\n\n<p>The ALL_CAPS variables are constants.\nX_train is 2-dim\ny_train is 18-dim<\/p>\n\n<p>Any help is appreciated!<\/p>\n","answers":[{"AnswerId":"52878805","CreationDate":"2018-10-18T16:42:30.373","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Well you are reshaping the predictions and flattening them here:<\/p>\n\n<pre><code>pred = model.predict(X_test).reshape(-1)\n<\/code><\/pre>\n\n<p>The <code>reshape(-1)<\/code> effectively makes the array one-dimensional. Just take the predictions directly:<\/p>\n\n<pre><code>pred = model.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52878198,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T16:04:30.843","AcceptedAnswerId":52882023.0,"OwnerUserId":826983.0,"Title":"Can I have hooks \/ callbacks outside a session or between the creation of one?","Body":"<p>With <code>train_and_evaluate()<\/code> it is possible to execute a schedule which trains and evaluates a model according to the specifications I am passing down. There are hooks I can register with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EvalSpec\" rel=\"nofollow noreferrer\">EvalSpec<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/TrainSpec\" rel=\"nofollow noreferrer\">TrainSpec<\/a> but there are limits.<\/p>\n\n<p>The problem is that I can <em>only<\/em> have a <code>SessionRunHook<\/code>, which <em>will<\/em> work as a callback, but always only <em>within<\/em> a session.<\/p>\n\n<p>My problem is that I am having a more complicated schedule. During evaluation I also want to quantize the model and further evaluate that quantized model. The problem here is, as mentioned, I am always <em>in<\/em> a session if I am using <code>SessionRunHook<\/code>-like objects.<\/p>\n\n<p>So the question is whether there is  a way so use <code>train_and_evaluate()<\/code> and register some callbacks in between:<\/p>\n\n<pre><code>train_and_evaluate(..):\n\n  # .. deep down ..\n\n  while &lt;condition&gt;:\n    with tf.Session() as train_sess:\n      # Do training ..\n\n    if the_callback_i_want:\n      the_callback_i_want()\n\n    with tf.Session() as eval_sess:\n      # Do evaluation ..\n<\/code><\/pre>\n\n<p>Is this possible?<\/p>\n","answers":[{"AnswerId":"52882023","CreationDate":"2018-10-18T20:26:48.223","ParentId":null,"OwnerUserId":"4282745","Title":null,"Body":"<p>I guess you could implement the <code>begin<\/code> method of your own <code>SessionHook<\/code> subclass.<\/p>\n\n<p>For the sake of the example I used the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/samples\/core\/get_started\/premade_estimator.py\" rel=\"nofollow noreferrer\">iris code<\/a> (see <a href=\"https:\/\/www.tensorflow.org\/guide\/premade_estimators\" rel=\"nofollow noreferrer\">this doc<\/a>).<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef the_callback_i_want():\n    # You need to work in a new graph so let's create a new one\n    g = tf.Graph()\n    with g.as_default():\n        x = tf.get_variable(\"x\", ())\n        x = tf.assign_add(x, 1)\n        init = tf.global_variables_initializer()\n        with tf.Session() as sess:   \n            sess.run(init)\n            print(\"I'm here !\", sess.run(x))\n\n\nclass MyHook(tf.train.SessionRunHook):\n\n  def begin(self):\n    \"\"\"Called once before using the session.\n\n    When called, the default graph is the one that will be launched in the\n    session.  The hook can modify the graph by adding new operations to it.\n    After the `begin()` call the graph will be finalized and the other callbacks\n    can not modify the graph anymore. Second call of `begin()` on the same\n    graph, should not change the graph.\n    \"\"\"\n    the_callback_i_want()\n\n\nimport iris_data\n# Fetch the data\n(train_x, train_y), (test_x, test_y) = iris_data.load_data()\n\n# Feature columns describe how to use the input.\nmy_feature_columns = []\nfor key in train_x.keys():\n    my_feature_columns.append(tf.feature_column.numeric_column(key=key))\n\n# Build 2 hidden layer DNN with 10, 10 units respectively.\nclassifier = tf.estimator.DNNClassifier(\n    feature_columns=my_feature_columns, hidden_units=[10, 10],  n_classes=3)\n\n# Fetch the data\n(train_x, train_y), (test_x, test_y) = iris_data.load_data()\n\n# Feature columns describe how to use the input.\nmy_feature_columns = []\nfor key in train_x.keys():\n    my_feature_columns.append(tf.feature_column.numeric_column(key=key))\n\ntrain_spec = tf.estimator.TrainSpec(input_fn=lambda:iris_data.train_input_fn(train_x, train_y,\n                                                 10), max_steps=100)\neval_spec = tf.estimator.EvalSpec(input_fn=lambda:iris_data.eval_input_fn(test_x, test_y,\n                                                10), hooks=[MyHook()])\ntf.estimator.train_and_evaluate(classifier, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>And it prints:<\/p>\n\n<pre><code>INFO:tensorflow:Saving checkpoints for 100 into \/var\/folders\/***\/model.ckpt.\nINFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Starting evaluation at 2018-10-18-20:19:28\nI'm here ! 1.9386581\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from \/var\/folders\/***\/model.ckpt-100\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Finished evaluation at 2018-10-18-20:19:28\n<\/code><\/pre>\n"}]}
+{"QuestionId":52878311,"AnswerCount":1,"Tags":"<python-3.x><numpy><tensorflow>","CreationDate":"2018-10-18T16:11:39.037","AcceptedAnswerId":52878532.0,"OwnerUserId":4082304.0,"Title":"How to extract rows and columns from a 3D array in Tensorflow","Body":"<p>I wanted to do the following indexing operation on a TensorFlow tensor.\nWhat should be the equivalent operations in TensorFlow to get <code>b<\/code> and <code>c<\/code> as output? Although <code>tf.gather_nd<\/code> documentation has several examples but I could not generate equivalent <code>indices<\/code> tensor to get these results.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na=np.arange(18).reshape((2,3,3))\n\nidx=[2,0,1] #it can be any validing re-ordering index list\n\n#These are the two numpy operations that I want to do in Tensorflow\nb=a[:,idx,:]\nc=a[:,:,idx] \n\n# TensorFlow operations\n\naT=tf.constant(a)\nidxT=tf.constant(idx)\n\n# what should be these two indices  \nidx1T=tf.reshape(idxT, (3,1)) \nidx2T=tf.reshape(idxT, (1,1,3))\n\nbT=tf.gather_nd(aT, idx1T ) #does not work\ncT=tf.gather_nd(aT, idx2T)  #does not work\n\nwith tf.Session() as sess:\n    b1,c1=sess.run([bT,cT])\n\nprint(np.allclose(b,b1))\nprint(np.allclose(c,c1))\n<\/code><\/pre>\n\n<p>I am not restricted to <code>tf.gather_nd<\/code> Any other suggestion to achieve the same operations on GPU will be helpful.<\/p>\n\n<h1>Edit: I have updated the question for a typo:<\/h1>\n\n<p>old statement: <code>c=a[:,idx]<\/code>, <\/p>\n\n<p>New statement: <code>c=a[:,:,idx]<\/code>\nWhat I wanted to achieve was re-ordering of columns as well. <\/p>\n","answers":[{"AnswerId":"52878532","CreationDate":"2018-10-18T16:24:48.940","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>That can be done with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a>, using the <code>axis<\/code> parameter:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na = np.arange(18).reshape((2,3,3))\nidx = [2,0,1]\nb = a[:, idx, :]\nc = a[:, :, idx]\n\naT = tf.constant(a)\nidxT = tf.constant(idx)\nbT = tf.gather(aT, idxT, axis=1)\ncT = tf.gather(aT, idxT, axis=2)\n\nwith tf.Session() as sess:\n    b1, c1=sess.run([bT, cT])\n\nprint(np.allclose(b, b1))\nprint(np.allclose(c, c1))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>True\nTrue\n<\/code><\/pre>\n"}]}
+{"QuestionId":52878460,"AnswerCount":1,"Tags":"<tensorflow><prediction><multilabel-classification><multiclass-classification>","CreationDate":"2018-10-18T16:20:49.147","AcceptedAnswerId":52879403.0,"OwnerUserId":7886651.0,"Title":"How to calculate the accuracy when dealing with multi-class mutlilabel classification in tensorflow?","Body":"<p>I am working with FER2013Plus dataset from <a href=\"https:\/\/github.com\/Microsoft\/FERPlus\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Microsoft\/FERPlus<\/a> which contains the fer2013new.csv file. This file contains labels for each image in the dataset. An example on labels could be:<\/p>\n\n<p>(4, 0, 0, 2, 1, 0, 0, 3) <\/p>\n\n<p>where each dimension is a different emotion. Finally, in their paper <a href=\"https:\/\/arxiv.org\/pdf\/1608.01041.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1608.01041.pdf<\/a>, they converted the labels distribution into probabilities => the new label would become <\/p>\n\n<p>(0.5, 0, 0, 0.25, 0.125, 0, 0, 0.375)<\/p>\n\n<p>In other words, the person in the image is happy with probability of 0.5, sad with probability of 0.25 and so on... And the sum of of the probabilities is 1.<\/p>\n\n<p>Now while training I used <code>tf.nn.softmax_cross_entropy_with_logits_v2<\/code> to calculate the loss between my predictions and the labels. Now how to compute the accuracy?<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[{"AnswerId":"52879403","CreationDate":"2018-10-18T17:22:03.053","ParentId":null,"OwnerUserId":"10458815","Title":null,"Body":"<p>Here is an excerpt from the paper:<\/p>\n\n<p>\"We take the majority emotion as the single\nemotion label, and we measure prediction accuracy against\nthe majority emotion.\"<\/p>\n\n<p>They are using a discrete classification task. So you just need to take the <code>tf.argmax()<\/code> on your logits to get the highest probability, and then compare that with the <code>tf.argmax()<\/code> of the labels.<\/p>\n\n<p>For example, if your label is <code>(0.5, 0, 0, 0.25, 0.125, 0, 0, 0.375)<\/code>, then happy is the majority emotion, so you would check if your logits had happy as the majority emotion as well.<\/p>\n"}]}
+{"QuestionId":52878998,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-18T16:56:34.203","AcceptedAnswerId":52879116.0,"OwnerUserId":10455742.0,"Title":"What is the difference between keras.backend.max vs keras.backend.argmax?","Body":"<p>I am a beginner in Deep Learning and while performing a practical assignment, came across the Keras documentation on <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">keras.backend<\/a>.<\/p>\n\n<p>I went through the explanation a number of times. however, i cannot exactly understand the difference between max and argmax function.<\/p>\n","answers":[{"AnswerId":"52879116","CreationDate":"2018-10-18T17:03:06.907","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>I will explain this using <code>max<\/code> and <code>argmax<\/code> from the <code>numpy<\/code> package, but the two functions are identical to the ones in the Keras backend:<\/p>\n\n<pre><code>import numpy as np\nvector = np.array([1, 2, 3, 2, 1])\n<\/code><\/pre>\n\n<p>Now, <code>np.max(vector)<\/code> returns the number <code>3<\/code>, as this is the maximal <em>value<\/em> in the vector. <code>np.argmax(vector)<\/code> however returns <code>2<\/code>, as this is the <em>index<\/em> of the maximal value in the vector.<\/p>\n\n<p>The <code>argmax<\/code> function is often used to post-process the output of a softmax layer. Say the output layer of your classifier (which classifies some image into one of four classes) is<\/p>\n\n<pre><code>output = Dense(4, activation='softmax')(...)\n<\/code><\/pre>\n\n<p>and the output of <code>predict(some_random_image)<\/code> is <code>[0.02, 0.90, 0.06, 0.02]<\/code>. Then, <code>argmax([0.02, 0.90, 0.06, 0.02])<\/code> immediately gives you the class (<code>1<\/code>).<\/p>\n"}]}
+{"QuestionId":52879126,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T17:03:41.390","AcceptedAnswerId":52879296.0,"OwnerUserId":3926152.0,"Title":"How are tensors immutable in TensorFlow?","Body":"<p>I read the following sentence in the TensorFlow documentation:<\/p>\n\n<blockquote>\n  <p>With the exception of tf.Variable, the value of a tensor is immutable,\n  which means that in the context of a single execution tensors only\n  have a single value. However, evaluating the same tensor twice can\n  return different values; for example that tensor can be the result of\n  reading data from disk, or generating a random number.<\/p>\n<\/blockquote>\n\n<p>Can someone elaborate a little bit on the \"immutable\" aspect of a Tensor?<\/p>\n\n<ol>\n<li>What is the \"scope of the immutability\" since evaluating a tensor twice could return different results? <\/li>\n<li>What does it mean \"the context of a single execution\"?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"52879296","CreationDate":"2018-10-18T17:14:51.583","ParentId":null,"OwnerUserId":"4903846","Title":null,"Body":"<p>Tensors, differently from variables, can be compared to a math equation.<\/p>\n\n<p>When you say a tensor equals 2+2, it's value is not actually 4, it's the computing instructions that leads to the value of 2+2 and when you start a session an execute it, TensorFlow runs the computations needed to return the value of 2+2 and gives you the output. And because of the tensor beeing the computations, rather than the the result, a tensor is immutable<\/p>\n\n<p>Now for your questions:<\/p>\n\n<ol>\n<li><p>By saying the tensor can be evaluated with different values it means that if you for example say that a tensor equals to a random number, when you run it different times, you will have different values (as the equation itself is a random one), but the value of the tensor itself as mentioned before, is not the value, is the steps that leads to it (in this case a random formula)<\/p><\/li>\n<li><p>The context of a single execution means that when you run a tensor, it will only output you one value. Think executing a tensor like applying the equation i mentioned. If i say a tensor equals random + 1, when you execute the tensor a single time, it will return you a random value +1, nothing else. But since the tensor contains a randomic output, if you run it multiple times, you will most likely get different values<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":52879230,"AnswerCount":2,"Tags":"<python><caffe><pycaffe>","CreationDate":"2018-10-18T17:09:23.613","AcceptedAnswerId":null,"OwnerUserId":9183291.0,"Title":"module 'caffe' has no attribute 'set_mode_gpu'","Body":"<p>I have installed caffe from source. I have used the Cmake for installation. I have updated the respective paths as well.\nMy caffe root directory is: \/home\/ashj\/caffe\nI have updated the PYTHON path as: \n<code>export PYTHONPATH=&lt;caffe-home&gt;\/python:$PYTHONPATH<\/code>\nwhich is by using \n<code>**export PYTHONPATH=\/home\/ashj\/caffe\/python:$PYTHONPATH<\/code>**<\/p>\n\n<p>I could load import the module caffe. However I am not able to access any methods or any layers inside the caffe like set_mode_gpu(), set_mode_cpu() or layers or params. I am getting errors like:<\/p>\n\n<p>When I used<\/p>\n\n<p> <\/p>\n\n<blockquote>\n  <p>import caffe<\/p>\n  \n  <p>caffe.set_mode_gpu()<\/p>\n<\/blockquote>\n\n<p>I am getting following error:<\/p>\n\n<blockquote>\n  <p><code>Traceback (most recent call last):<\/code><\/p>\n  \n  <p><code>File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;<\/code><\/p>\n  \n  <p><code>AttributeError: module 'caffe' has no attribute 'set_mode_gpu'<\/code><\/p>\n<\/blockquote>\n\n<p>PS: I have also tried using caffe.__caffe.set_mode_gpu() as mentioned in <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/issues\/373#issuecomment-186678800\" rel=\"nofollow noreferrer\">this link.<\/a> but it is not working for me.\nMy system specs: Ubuntu 18.04<\/p>\n\n<p>TIA<\/p>\n","answers":[{"AnswerId":"52879554","CreationDate":"2018-10-18T17:32:23.153","ParentId":null,"OwnerUserId":"10508893","Title":null,"Body":"<p>Try these steps and then set your python PATH:\nYou might've already done steps 1 and 3.<\/p>\n\n<pre><code>make all \nmake pycaffe\nmake distribute\nmkdir ~\/python \nmv distribute\/python\/caffe ~\/python\n<\/code><\/pre>\n\n<p>Set your PYTHONPATH after this \u2014 this should be some dir like <code>caffe\/python\/caffe<\/code><\/p>\n"},{"AnswerId":"54860437","CreationDate":"2019-02-25T06:14:29.317","ParentId":null,"OwnerUserId":"6759403","Title":null,"Body":"<p>Though may be late, but I encountered this same problem and found a way work around:<\/p>\n\n<pre><code>sys.path.insert(0, '\/path\/to\/caffe\/python')\nimport caffe\ncaffe.set_mode_gpu()\n<\/code><\/pre>\n\n<p>namely, add the <code>caffe\/pathon<\/code> path to you <code>sys.path<\/code> before <code>import caffe<\/code>.<br>\nWrote a <a href=\"https:\/\/www.oukohou.wang\/2019\/01\/16\/reocrds_of_installation_errors_and_operations\/#32-solution\" rel=\"nofollow noreferrer\">post here<\/a> for detailed analysis, hope helpful.<\/p>\n\n<hr>\n\n<p>This problem may be the results of package <code>caffe<\/code>'s path.  <\/p>\n\n<p>For me, if I do the following from the Ubuntu Terminal, everything goes fine:\n<a href=\"https:\/\/i.stack.imgur.com\/nJqq2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nJqq2.png\" alt=\"from_terminal\"><\/a>  <\/p>\n\n<p>but if I do from the Pycharm IDE, errors occur:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/b0wLB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/b0wLB.png\" alt=\"from_pycharm\"><\/a>  <\/p>\n\n<p>note that I tested the package <code>caffe<\/code>'s path in both ways, and got different results:<br>\n- in the Ubuntu terminal, namely the way which goes fine, I got <\/p>\n\n<pre><code>'\/home\/CVAR-B\/softwares\/caffe\/caffe\/python\/caffe\/__init__.pyc'\n<\/code><\/pre>\n\n<p>which is the expected result;\n- in the Pycharm IDE way, namely the way error occurs, I got <\/p>\n\n<pre><code>'\/usr\/local\/lib\/python2.7\/dist-packages\/caffe\/__init__.pyc'\n<\/code><\/pre>\n\n<p>which is not the expected result.  <\/p>\n\n<p>In view of this discovery, I did this one more thing to handle the error:<\/p>\n\n<pre><code>sys.path.insert(0, '\/path\/to\/caffe\/python')\nimport caffe\ncaffe.set_mode_gpu()\n<\/code><\/pre>\n\n<p>namely, add the <code>caffe\/pathon<\/code> path to you <code>sys.path<\/code> before <code>import caffe<\/code>.  <\/p>\n\n<p>and the result shows this can be a workaround:\n<a href=\"https:\/\/i.stack.imgur.com\/lTkEK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lTkEK.png\" alt=\"pycharm_IDE_success\"><\/a><br>\n<sub>(source: <a href=\"https:\/\/s1.ax2x.com\/2019\/02\/25\/5jquKK.png\" rel=\"nofollow noreferrer\">ax2x.com<\/a>)<\/sub>  <\/p>\n\n<p>See the <code>caffe.__file__<\/code>'s result, now returns the expected path.   <\/p>\n"}]}
+{"QuestionId":52879519,"AnswerCount":2,"Tags":"<python><keras><coreml><coremltools>","CreationDate":"2018-10-18T17:30:14.130","AcceptedAnswerId":52879954.0,"OwnerUserId":1179925.0,"Title":"Convert MobileNet from Keras to CoreML","Body":"<p>I'm using Keras 2.1.3 and I want to convert MobileNet to CoreML:<\/p>\n\n<pre><code>from keras.applications import MobileNet\nfrom keras.applications.mobilenet import relu6\nfrom keras.applications.mobilenet import DepthwiseConv2D\n\nimport coremltools.converters.keras as k\n\ndef save_model():\n    model = MobileNet(input_shape=(128,128,3), include_top=False)\n    model.save('temp.h5')\n\ndef convert():\n    model = k.convert('temp.h5',\n                      input_names=['input'],\n                      output_names=['output'],\n                      model_precision='float16',\n                      custom_conversion_functions={'relu6': relu6, 'DepthwiseConv2D': DepthwiseConv2D})\n    model.save('temp.model')\n\nsave_model()\nconvert()\n<\/code><\/pre>\n\n<p>This gives an error: <code>ValueError: Unknown activation function:relu6<\/code><\/p>\n","answers":[{"AnswerId":"54466679","CreationDate":"2019-01-31T18:02:03.677","ParentId":null,"OwnerUserId":"7064856","Title":null,"Body":"<p>for Keras 2.2.4 and Tensorflow 1.12.0 I found a solution.<\/p>\n\n<p>Save model weights &amp; architecture like:<\/p>\n\n<pre><code>model_json = model.to_json()\nopen('architecture.json', 'w').write(model_json)\nmodel.save_weights('weights.h5', overwrite=True)\n<\/code><\/pre>\n\n<p>And for converting a model to CoreML .mlmodel I use:<\/p>\n\n<pre><code>import coremltools\n\nfrom keras.layers import DepthwiseConv2D, ReLU\nfrom pathlib import Path\nfrom keras.models import model_from_json\nfrom tensorflow.python.keras.utils.generic_utils import CustomObjectScope\n\nmodel_architecture = '.\/Networks\/architecture.json'\nmodel_weights = '.\/Networks\/weights.h5'\n\nmodel_structure = Path(model_architecture).read_text()\n\nwith CustomObjectScope({'relu6': ReLU ,'DepthwiseConv2D': DepthwiseConv2D}):\n    model = model_from_json(model_structure)\n    model.load_weights(model_weights)\n\n    output_labels = ['0', '1', '2', '3', '4', '5', '6']\n    coreml_model = coremltools.converters.keras.convert(\n        model, input_names=['image'], output_names=['output'],\n        class_labels=output_labels, image_input_names='image')\n\n    coreml_model.save('ModelX.mlmodel')\n<\/code><\/pre>\n"},{"AnswerId":"52879954","CreationDate":"2018-10-18T17:59:40.907","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Here is solution based on <a href=\"https:\/\/github.com\/apple\/coremltools\/issues\/38\" rel=\"nofollow noreferrer\">https:\/\/github.com\/apple\/coremltools\/issues\/38<\/a><\/p>\n\n<pre><code>from keras.applications import MobileNet\nimport keras\n\nimport coremltools.converters.keras as k\nfrom keras.utils.generic_utils import CustomObjectScope\n\ndef save_model():\n    model = MobileNet(input_shape=(128,128,3), include_top=False)\n    model.save('temp.h5')\n\ndef convert():\n    with CustomObjectScope({'relu6': keras.applications.mobilenet.relu6,\n                            'DepthwiseConv2D': keras.applications.mobilenet.DepthwiseConv2D}):\n        model = k.convert(\"temp.h5\",\n                          input_names=['input'],\n                          output_names=['output'],\n                          model_precision='float16')\n\n    model.save('temp.mlmodel')\n\nsave_model()\nconvert()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52880384,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-10-18T18:29:29.947","AcceptedAnswerId":null,"OwnerUserId":3276237.0,"Title":"Keras loss becomes nan only at epoch end","Body":"<p>I am observing some strange behavior from Keras. I am training a small model where the training loss becomes nan <em>only<\/em> at the end of the first epoch.<\/p>\n\n<p>So if I have 100 batches, and I terminate training at batch 99, then resume for another 99 it trains fine. Otherwise, once it reaches the end of an epoch it always returns nan. <\/p>\n\n<p>I am using a custom loss function:<\/p>\n\n<pre><code>def corr(x, y):\n    xc = x - K.mean(x)\n    yc = y - K.mean(y)\n    r_num = K.mean(xc*yc) \n    r_den = K.std(x)*K.std(y)\n    return r_num\/r_den\n<\/code><\/pre>\n\n<p>And I have tried all of the standard tricks like dropping my learning rate, clipping the norm and value of my gradient, and increasing batch size. Only in the event of increasing my batch size to something unrealistic like 100,000 (I have 1 million data points) does it actually continue past an epoch, but I would like to understand what is going on at the end that is causing this strange behavior. I also tried different optimizers (currently using Adam), and tried this on different systems to make sure it wasn't a problem on my one computer. <\/p>\n\n<p>My input and output is one dimensional and my model is summarized below. <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_7 (InputLayer)         (None, 1)                 0         \n_________________________________________________________________\ndense_7 (Dense)              (None, 100)               200       \n_________________________________________________________________\ndense_8 (Dense)              (None, 100)               10100     \n_________________________________________________________________\ndense_9 (Dense)              (None, 1)                 101       \n=================================================================\nTotal params: 10,401\nTrainable params: 10,401\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Does Keras so something special at the end of an epoch? I couldn't find anything other than the standard logger callback. I also wrote a custom callback which evaluates my model each batch and stores the output, and when I plot it over time it does not appear to blow up or do anything strange. It just looks like it's slowly improving, then the training dies.<\/p>\n","answers":[{"AnswerId":"52882417","CreationDate":"2018-10-18T20:54:03.217","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Probably it is caused by a division by zero in the loss function. Make sure the denominator is always positive by adding a small constant to it. You can use <code>K.epsilon()<\/code> for this purpose:<\/p>\n\n<pre><code>    return r_num \/ (r_den + K.epsilon())\n<\/code><\/pre>\n"}]}
+{"QuestionId":52880636,"AnswerCount":0,"Tags":"<python><pandas><scikit-learn><keras>","CreationDate":"2018-10-18T18:47:42.983","AcceptedAnswerId":null,"OwnerUserId":7020965.0,"Title":"pandas data-frame from .csv file into single cell","Body":"<p>I have video files in three directories (Train , Development , Test ) and<\/p>\n\n<p>there is two .csv files ( the features are in one  .csv file , but the labels <\/p>\n\n<p>and file names are  in other .csv file ). I'm trying to make a dataframe by <\/p>\n\n<p>selecting some colunms from the two .csv files to feed it to keras CNN model .<\/p>\n\n<p>Below is what I tried to read some features from the 1st  .csv file , but I <\/p>\n\n<p>don't know if this this is the right way to do it ?<\/p>\n\n<pre><code>import io\nimport numpy as np\ndef read_data(file_path):\n\ndata = pd.read_csv(file_path,sep=',')\ndata =data.iloc[0:1560, [0,680,681,682,683,684]] \nreturn data\n\ndf=read_data('dev_001.csv')\nprint(df.shape)\ndf.head()\n\ndata_frame = pd.DataFrame(df)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QrqNI.png\" rel=\"nofollow noreferrer\"> output features saved in data_frame<\/a><\/p>\n\n<p>I want now to insert this data frame into one cell so I have something like this :<\/p>\n\n<p>Table :<\/p>\n\n<pre><code>Features                                          Subject_id           labels   \n\nExtracted features from .csv file (data_frame )   ID from .csv file   Label from .csv file\n<\/code><\/pre>\n\n<p>Where I need  all features in .csv file in first cell. How this can be done ?<\/p>\n\n<p>I tried this :<\/p>\n\n<pre><code> new_df = pd.DataFrame({'Features': [temp_data_frame]})\n\n print(temp_data_frame.shape)\n\n print(new_df.shape)\n(1560, 35)\n(1, 1)\n<\/code><\/pre>\n\n<p>But it only takes the first row from (temp_data_frame) <\/p>\n","answers":[]}
+{"QuestionId":52881412,"AnswerCount":1,"Tags":"<python><tensorflow><convolution>","CreationDate":"2018-10-18T19:45:16.357","AcceptedAnswerId":52884174.0,"OwnerUserId":4185012.0,"Title":"Tensorflow Convolution return Nan","Body":"<p>I'm trying to read image from folder and perform convolution.<\/p>\n\n<p>First, I input these images and package them as <code>.tfrecords<\/code>, and decode the tfrecords by <code>tf.train.batch<\/code>.\nNext, I put all data(image,label) into convolution (as code).\nIn this step,the <code>bias<\/code>(b_conv1) and <code>weight<\/code>(w_conv1) will be <em>Nan<\/em>, and the model won't work any more.<\/p>\n\n<pre><code>image_batch, label_batch = decodeTFRecord(dirPath,batch_size)\nimage_batch = tf.reshape(image_batch,[-1,128*128])\nlabel_batch = tf.one_hot(label_batch,Label_size)\n\nx = tf.placeholder(tf.float32,[None,128*128])\ny = tf.placeholder(tf.float32,[None,10])\nx_image = tf.convert_to_tensor(tf.reshape(x,[-1,128,128,1]))\n#conv1 layer\nw_conv1 = weight_variable([5,5,1,32])\nb_conv1 = bias_variable([32])\nh_conv1 = tf.nn.relu(conv2d(x_image,w_conv1)+b_conv1) #outsize = 128*128*32\nh_pool1 = max_pool_2x2(h_conv1) # outsize = 64*64*32\n<\/code><\/pre>\n\n<p><strong>conv2d function:<\/strong><\/p>\n\n<pre><code>def conv2d(x,W):\n   return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding ='SAME')\n<\/code><\/pre>\n\n<p><strong>max_pool_2x2 function:<\/strong><\/p>\n\n<pre><code>def max_pool_2x2(x):\n    return tf.nn.max_pool(x,ksize=[1,2,2,1], strides=[1,2,2,1],padding='SAME')\n<\/code><\/pre>\n\n<p>Full Code: \n<a href=\"https:\/\/codeshare.io\/5O7ddj\" rel=\"nofollow noreferrer\">https:\/\/codeshare.io\/5O7ddj<\/a><\/p>\n","answers":[{"AnswerId":"52884174","CreationDate":"2018-10-19T00:06:51.890","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>Try a smaller learning rate to start with <code>1e-5<\/code> and make your initial weights smaller <code>tf.truncated_normal(shape, stddev = 0.0001)<\/code> and see if either of those common issues fix your problem.<\/p>\n\n<p>Based on comments it sounds like one of these two common issues caused the <code>NaN<\/code> problem (please comment if I misread your comment).<\/p>\n\n<p>This issue often occurs when the weights are randomly initialized because large weights will have a long way to travel to improve, which can result in a very steep gradient step, e.g. the exploding gradient problem. Small weights\/learning rates will ameliorate this issue. <\/p>\n\n<p>Also notable is that BatchNorm will tend to ameliorate the problem as well. You can generally get away with much larger learning rates precisely because BatchNorm keeps things from getting really out of whack as the signal travels through the network.<\/p>\n"}]}
+{"QuestionId":52882948,"AnswerCount":1,"Tags":"<javascript><tensorflow><artificial-intelligence><game-development><tensorflow.js>","CreationDate":"2018-10-18T21:37:01.557","AcceptedAnswerId":null,"OwnerUserId":10526249.0,"Title":"Develop game AI with Tensorflow.js","Body":"<p>I've made a game in my github repository (<a href=\"https:\/\/diteix.github.io\/lil-square-of-eight\/\" rel=\"nofollow noreferrer\">\"Lil Square of Eight\"<\/a>) and created some code to play against machine. Now, for study and joy purposes, I would like to create some AI with Tensorflow.js that can play the game and even win.<\/p>\n\n<p>The game is simple, each turn a player click a square side and it's marked with player color. The player who closes the square do a score and can keep going with his turn.<\/p>\n\n<p>I've done some tests with Tensorflow.js and I didn't understood it at all. All I could do was verify if one square has a side that can be clicked and don't give the oportunity to the other player score, or do a score itself.<\/p>\n\n<pre><code>const model = tf.sequential();\n\nmodel.add(tf.layers.dense({ units: 16, inputShape: [8] }));\nmodel.add(tf.layers.dense({ units: 16, inputShape: [16], activation: 'sigmoid' }));\nmodel.add(tf.layers.dense({ units: 1, inputShape: [16] }));\nmodel.compile({ optimizer: tf.train.adam(0.1), loss: 'meanSquaredError', lr: 0.3 });\n\nconst xs1 = tf.tensor2d([\n                [0, 1, 1, 1, 2, 0, 3, 0], [0, 1, 1, 0, 2, 1, 3, 0], [0, 1, 1, 0, 2, 0, 3, 1], [0, 0, 1, 1, 2, 1, 3, 0], [0, 0, 1, 1, 2, 0, 3, 1], [0, 0, 1, 0, 2, 1, 3, 1],\n                [0, 1, 1, 1, 2, 1, 3, 1],\n                [0, 0, 1, 0, 2, 0, 3, 0], [0, 1, 1, 0, 2, 0, 3, 0], [0, 0, 1, 1, 2, 0, 3, 0], [0, 0, 1, 0, 2, 1, 3, 0], [0, 0, 1, 0, 2, 0, 3, 1],\n                [0, 0, 1, 1, 2, 1, 3, 1], [0, 1, 1, 0, 2, 1, 3, 1], [0, 1, 1, 1, 2, 0, 3, 1], [0, 1, 1, 1, 2, 1, 3, 0]]);\n\nawait model.fit(xs1, tf.tensor2d([[0], [0], [0], [0], [0], [0], [0], [1], [1], [1], [1], [1], [2], [2], [2], [2]]), {epochs: 550});\n\nconsole.log(model.predict(tf.tensor2d([[0, 1, 1, 0, 2, 0, 3, 0]])));\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>As asked, I would like help in how I can create an AI with Tensorflow.js that can think and play my game in a smart way. It can learn with previuos games and start to play the game.<\/p>\n","answers":[{"AnswerId":"56318691","CreationDate":"2019-05-27T01:34:47.213","ParentId":null,"OwnerUserId":"1250389","Title":null,"Body":"<p>What you're looking for is called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Reinforcement_learning\" rel=\"nofollow noreferrer\">Reinforcement learning<\/a>.   It's not as popular as most Tensorflow methods for ML.  TFJS examples generally orbit around supervised learning methods.<\/p>\n\n<p>In this case, here's my recommendation.   The most popular (to me) method of game play on ML reinforcement learning, is called Q Learning.  Firstly, you can research Q Learning in regular old Tensorflow and Python, then you can convert it over to TFJS once you deeply understanding the concepts.  You'll be one of the first!<\/p>\n\n<p>Here's a great series on Tensorflow AND Q Learning:  <a href=\"https:\/\/medium.com\/emergent-future\/simple-reinforcement-learning-with-tensorflow-part-0-q-learning-with-tables-and-neural-networks-d195264329d0\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/emergent-future\/simple-reinforcement-learning-with-tensorflow-part-0-q-learning-with-tables-and-neural-networks-d195264329d0<\/a><\/p>\n\n<p>I hope this gives you some direction!<\/p>\n"}]}
+{"QuestionId":52883145,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-18T21:54:07.433","AcceptedAnswerId":null,"OwnerUserId":1019129.0,"Title":"Why Keras model on \"bare\" CPU is faster?","Body":"<p>Running a Keras model... the bad thing is that it is faster not to use CPU extentions (it should be the other way around).\nLook at the output below.<\/p>\n\n<p>Is there a config file where I can set inter_op_parallelism option ?<\/p>\n\n<hr>\n\n<pre><code> Using TensorFlow backend.\n 2018-10-18 17:21:32.620461: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n 2018-10-18 17:21:32.621535: I tensorflow\/core\/common_runtime\/process_util.cc:69] Creating new thread pool with default inter op setting: 2. Tune using inter_op_parallelism_threads for best performance.\n Results: -33.20 (23.69) MSE\n\n real   2m55.990s\n user   4m8.784s\n sys    3m50.192s\n<\/code><\/pre>\n\n<hr>\n\n<pre><code> Using TensorFlow backend.\n 2018-10-18 17:25:04.773578: I tensorflow\/core\/platform\/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n Results: -32.57 (23.16) MSE\n\n real   1m48.847s\n user   2m15.792s\n sys    0m13.440s\n<\/code><\/pre>\n","answers":[{"AnswerId":"53252276","CreationDate":"2018-11-11T19:16:59.153","ParentId":null,"OwnerUserId":"530123","Title":null,"Body":"<p>Here you are the code I'm using with keras, just put it on the top of your code. <\/p>\n\n<pre><code>from keras import backend as K\nimport tensorflow as tf\n\nNUM_PARALLEL_EXEC_UNITS = 6\n\nconfig = tf.ConfigProto(intra_op_parallelism_threads = NUM_PARALLEL_EXEC_UNITS, \n         inter_op_parallelism_threads = 1, \n         allow_soft_placement = True, \n         device_count = {'CPU': NUM_PARALLEL_EXEC_UNITS })\n\nsession = tf.Session(config=config)\n\nK.set_session(session)\n\nimport os\n\nos.environ[\"OMP_NUM_THREADS\"] = str(NUM_PARALLEL_EXEC_UNITS)\nos.environ[\"KMP_BLOCKTIME\"] = \"30\"\nos.environ[\"KMP_SETTINGS\"] = \"1\"\nos.environ[\"KMP_AFFINITY\"]= \"granularity=fine,verbose,compact,1,0\"\n<\/code><\/pre>\n\n<p>Note: I'm a little bit disappointed with the results. I could reach maximum 150% speeding up only playing with these parameters.<\/p>\n"}]}
+{"QuestionId":52883177,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2018-10-18T21:57:43.500","AcceptedAnswerId":null,"OwnerUserId":6941347.0,"Title":"GPU compatibility with tensorflow installation","Body":"<p>New in the field of Deep Learning.<\/p>\n\n<p>I need to install Tensorflow with GPU support. Before I purchase GPU I need to know like GPU is compatible with Tensorflow or not. In the Tensorflow installation page, with option Tensorflow with GPU below are the software requirements.<\/p>\n\n<p>The following NVIDIA\u00ae software must be installed on your system:<\/p>\n\n<ol>\n<li>NVIDIA\u00ae GPU drivers \u2014CUDA 9.0 requires 384.x or higher. CUDA\u00ae<\/li>\n<li>Toolkit \u2014TensorFlow supports CUDA 9.0. CUPTI ships with the CUDA<\/li>\n<li>Toolkit. cuDNN SDK (>= 7.2) <\/li>\n<li>(Optional) NCCL 2.2 for multiple GPU support. <\/li>\n<li>(Optional) TensorRT 4.0 to improve latency and throughput for inference on \nsome models.<\/li>\n<\/ol>\n\n<p>If I go with brands available in market other than Nvidia, like Gigabyte, Asus etc lets say <em>AORUS GeForce\u00ae GTX 1080 Ti Xtreme Edition 11G<\/em> , will it allow me to proceed with Tensorflow or not ?<\/p>\n","answers":[{"AnswerId":"53205700","CreationDate":"2018-11-08T10:20:31.677","ParentId":null,"OwnerUserId":"8271309","Title":null,"Body":"<p>For compatibility with Nvidia drivers\nVisit website: <a href=\"https:\/\/www.nvidia.com\/object\/unix.html\" rel=\"nofollow noreferrer\">https:\/\/www.nvidia.com\/object\/unix.html<\/a>\nClick on driver, which you are interested in. In your case it is 390.87 and 410.73 click on SUPPORTED PRODUCTS. If there is your GPU you are ok.<\/p>\n\n<p>In your case: yes you can use latest version of tensorflow. <\/p>\n\n<p>Installation can be little tricky, but there are lots of how on the internet to install Nvidia driver, Cuda, cuDNN, NCCL and so on<\/p>\n"}]}
+{"QuestionId":52883624,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-18T22:49:07.247","AcceptedAnswerId":52886649.0,"OwnerUserId":8764887.0,"Title":"Remove last rows and columns of image with tensorflow","Body":"<p>I want to remove last rows and columns of my image based on given desired height and width.<\/p>\n\n<p>For example consider I have a image of size 351x572x3.<\/p>\n\n<p>I would like to have image of size 350x550x3 i.e I want to remove last row and last 22 columns of my image . <\/p>\n\n<p>How can i achieve this using tensorflow ?<\/p>\n\n<p>I know the desired_width and desired_height of image in advance<\/p>\n\n<p>code:<\/p>\n\n<pre><code>def parse_fn(filename,desired_height,desired_width):\n    \"\"\"Decode the bmp image from the filename and convert to [0, 1].\"\"\"\n    image_string = tf.read_file(filename)\n    image_decoded = tf.image.decode_bmp(image_string)\n    # This will convert to float values in [0, 1]\n    image = tf.image.convert_image_dtype(image_decoded, tf.float32)\n\n    # I want to do convert the image to desired width and height here\n\n    return image\n<\/code><\/pre>\n\n<p>Thanks in Advance<\/p>\n","answers":[{"AnswerId":"52886649","CreationDate":"2018-10-19T06:01:31.960","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>You can just index into tensors like in numpy. So e.g.<\/p>\n\n<pre><code>image_cutoff = image[:desired_height, :desired_width]\n<\/code><\/pre>\n"},{"AnswerId":"52887480","CreationDate":"2018-10-19T07:18:13.063","ParentId":null,"OwnerUserId":"2702509","Title":null,"Body":"<p>There is<\/p>\n\n<pre><code>tf.image.crop_to_bounding_box(\n    image,\n    offset_height,\n    offset_width,\n    target_height,\n    target_width\n)\n<\/code><\/pre>\n\n<p>For your case,<\/p>\n\n<p>image = tf.image.crop_to_bounding_box(image,0,0,desired_height,desired_width)<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/crop_to_bounding_box\" rel=\"nofollow noreferrer\">link to the doc<\/a><\/p>\n"}]}
+{"QuestionId":52884843,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2018-10-19T01:57:47.157","AcceptedAnswerId":null,"OwnerUserId":10526884.0,"Title":"Tensorflow SSD300 for Android","Body":"<p>I'm a college student who studies machine leraning in Japan.\nI'm not good at using English, but I will make efforts to convey my situation in English.<\/p>\n\n<p>I'm now trying to use the object detection model for android.\nI used SSD300_mobile_net for training and then I got .hdf5 file which has model's weights.<\/p>\n\n<p>Next, I convert this .hdf5file into .pb file which is fit for tensorflow.\nAt last, I want to use this .pb file for tensorflow android .(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android<\/a>)<\/p>\n\n<p>But I got an error on Android Studio as below while debugging the application by using my device<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qB6Eo.png\" rel=\"nofollow noreferrer\">This is the error capture<\/a>\nI want to know how to solve this problem.<\/p>\n","answers":[{"AnswerId":"54473710","CreationDate":"2019-02-01T06:01:45.883","ParentId":null,"OwnerUserId":"9441041","Title":null,"Body":"<p>I sense that it's an issue with your hardware. Generally the hdf5 files are very big and even bigger is the computation that is being done on it. Since your system is giving OOM exception, I believe you are getting this error while loading weights into the model.<\/p>\n"}]}
+{"QuestionId":52885948,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-19T04:33:17.943","AcceptedAnswerId":null,"OwnerUserId":10527246.0,"Title":"error of imoprt tensorflow as tf in python 3.5.2","Body":"<p>I have installed TensorFlow in virtual environment on Ubunut 16.04. when I enter in virtualen by using command \"<strong><em>source ~\/tensorflow\/bin\/activate<\/em><\/strong>\" it enters in virtualen. but after that when I enter the command \" <strong>import tensorflow as tf<\/strong>\" it gives me the following error \n\"<strong><em>import: not authorized 'tf' @ error\/constitute.c\/WriteImages\/1028.<\/em><\/strong>\"\nhow to solve this..<\/p>\n","answers":[{"AnswerId":"53217774","CreationDate":"2018-11-08T23:38:08.923","ParentId":null,"OwnerUserId":"6396954","Title":null,"Body":"<p>Maybe you forgot about telling which interpreter to use. Two variants:<\/p>\n\n<ol>\n<li>Add shebang <code>#!\/usr\/bin\/env python3<\/code> at the beginning of you script<\/li>\n<\/ol>\n\n<p><strong>OR<\/strong><\/p>\n\n<ol start=\"2\">\n<li>Run script like <code>python3 my_scripy.py<\/code><\/li>\n<\/ol>\n"}]}
+{"QuestionId":52886117,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><keras><computer-vision><keras-layer>","CreationDate":"2018-10-19T04:59:18.227","AcceptedAnswerId":null,"OwnerUserId":6060127.0,"Title":"Keras Init Sequential model layers by Model layers","Body":"<p>I'm trying to build some app using Transfer Learning. I want to use Vgg16 so I've done sth like this:\n<code>vgg16_model = keras.applications.vgg16.VGG16()<\/code> but I want to transfer layers from Vgg16 to my model.\n<code>model = Sequential(layers=vgg16_model.layers)<\/code> (I've seen this <a href=\"https:\/\/www.youtube.com\/watch?v=oDHpqu52soI&amp;list=PLZbbT5o_s2xrwRnXk_yCPtnqqo4_u2YGL&amp;index=13\" rel=\"nofollow noreferrer\">here<\/a>)\nbut it leads me to error <\/p>\n\n<blockquote>\n  <p>TypeError: The added layer must be an instance of class Layer. Found:\n   tensorflow.python.keras.engine.input_layer.InputLayer<\/p>\n<\/blockquote>\n\n<p>How can I init my Sequential model by vgg16 layers?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"52887427","CreationDate":"2018-10-19T07:14:36.280","ParentId":null,"OwnerUserId":"7374610","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>vgg = VGG16()\n\nmodel = Sequential()\nmodel.add(vgg)\nmodel.add(...) # add additional layers\n<\/code><\/pre>\n"}]}
+{"QuestionId":52886591,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-19T05:55:25.743","AcceptedAnswerId":null,"OwnerUserId":2137798.0,"Title":"How to manage memory while feeding a model","Body":"<p>I am using Keras with a tensorflow backend to solve an image classification problem. I had initially taken a na\u00efve approach of loading all of the images from the files into an array which I passed to the Keras fit() function. This worked as long as I used thumbnails of the images, but I have lately decided to use the full size images. <\/p>\n\n<p>No sooner had I tried this when I got the following error:<\/p>\n\n<p><em>2018-10-18 20:39:14.486023: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:275] Allocator (mklcpu) ran out of memory trying to allocate 63.50GiB.  Current allocation summary follows.\n2018-10-18 20:39:14.490332: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:630] Bin (256):   Total Chunks: 24, Chunks in use: 22. 6.0KiB allocated for chunks. 5.5KiB in use in bin. 2.8KiB client-requested in use in bin.<\/em><\/p>\n\n<p>This of course was hardly a surprise, because the images are over 1 MB apiece, and I have about 5000 of them.The question is what to do about this.<\/p>\n\n<p>One possible na\u00efve solution is simply to load the images in small numbers, and repeatedly call the fit () function with about, say, 10 images at a time. What I want to know is if Keras and Tensorflow can offer a better and more elegant solution to this problem. I am thinking essentially of a callback type function that would execute every time a batch is loaded, in which I would manually read the files into bitmaps to feed into the optimizer.<\/p>\n\n<p>I don't need a detailed answer-if someone could just point me in a general way to the correct area of the documentation, I would appreciate that.<\/p>\n","answers":[{"AnswerId":"52888472","CreationDate":"2018-10-19T08:21:16.813","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a><\/p>\n\n<p>This page has almost all of the information you need. I recommend reading all of it.<\/p>\n"}]}
+{"QuestionId":52886684,"AnswerCount":1,"Tags":"<python><keras><pytorch>","CreationDate":"2018-10-19T06:06:01.757","AcceptedAnswerId":null,"OwnerUserId":10282036.0,"Title":"How do the same function for pytorch expand_as in keras","Body":"<p>pytorch code:<\/p>\n\n<pre><code>att_h = att_h.unsqueeze(1).expand_as(att)\n<\/code><\/pre>\n\n<p>att_h shape is (10,512)<\/p>\n\n<p>att shape is (10,196,512)<\/p>\n\n<p>Keras code:<\/p>\n\n<pre><code>K.expand_dims(att_h, 1).expand_as(att)\n<\/code><\/pre>\n\n<p>Got error:\n    'Tensor' object has no attribute 'expand_as'<\/p>\n\n<p>Not sure how to do the same thing in keras.<\/p>\n","answers":[{"AnswerId":"52886872","CreationDate":"2018-10-19T06:23:34.017","ParentId":null,"OwnerUserId":"7699859","Title":null,"Body":"<p>There isn't any inbuilt function in Keras as such. But you can achieve the same result using : <code>np.reshape<\/code>. <\/p>\n\n<p>So you can :<\/p>\n\n<pre><code>K.expand_dims(att_h,1).reshape(att.shape)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52887109,"AnswerCount":1,"Tags":"<tensorflow><node-gyp><tensorflow.js>","CreationDate":"2018-10-19T06:47:21.297","AcceptedAnswerId":null,"OwnerUserId":1656749.0,"Title":"How to use tfjs-node with libtensorflow that is built from source","Body":"<p>I see that it is possible to use <code>libtensorflow<\/code> that is built from source, as mentioned in the README\n<a href=\"https:\/\/github.com\/tensorflow\/tfjs-node#optional-build-libtensorflow-from-tensorflow-source\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tfjs-node#optional-build-libtensorflow-from-tensorflow-source<\/a><\/p>\n\n<p>I have successfully built from source, but I don't know how to let <code>tfjs-node<\/code> use this custom built version, instead of the pre-built one.<\/p>\n","answers":[{"AnswerId":"52890939","CreationDate":"2018-10-19T10:52:28.080","ParentId":null,"OwnerUserId":"1656749","Title":null,"Body":"<p>I found that the following workaround seems to work:<\/p>\n\n<p>Copy the file <code>bazel-bin\/tensorflow\/libtensorflow.so<\/code> from the tensorflow source directory to replace <code>node_modules\/@tensorflow\/tfjs-node\/deps\/lib\/libtensorflow.so<\/code> inside your project.<\/p>\n\n<p>Hopefully there is a better way to do <code>npm install<\/code> such that there is no need to download the pre-compiled version.<\/p>\n"}]}
+{"QuestionId":52887141,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api><tensorflow-slim>","CreationDate":"2018-10-19T06:49:38.780","AcceptedAnswerId":52947248.0,"OwnerUserId":10527330.0,"Title":"Tensorflow Object Detection got different results in two evaluations on the same model checkpoint","Body":"<p>I'm a rookie to tensorflow and currently working on object detection API.<\/p>\n\n<p>I've chosen ssd_resnet50_fpn to get started and downloaded the pretrained model from tensorflow model zoo to do transfer learning with my own dataset with only 1 class (person). The training configurations was defined in the pipeline.config which was revised from the one in the same package of the pretrained model, and I trained the model with legacy train.py script. <\/p>\n\n<p>The training process was fine and the loss decreased as expected, and I've exported my evaluation images with bounded boxes when doing the evaluation (via legacy eval.py). The inference was fine and those exported images worked as expected.<\/p>\n\n<p>However, I found something weird that I got different evaluation results on the same model checkpoint. When I perform eval.py twice with the same parameters, I found that the bounded boxes from inference were different on the same image.<\/p>\n\n<p>Here's the evaluation result made by eval.py (coco_detection_metrics)<\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/gOqQkyy.png\" rel=\"nofollow noreferrer\">First time<\/a><\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/AN7ROSj.png\" rel=\"nofollow noreferrer\">Second time<\/a><\/p>\n\n<p>Since I'm not really understand about the mAP, but the two results were slightly different.<\/p>\n\n<p>And here's one of the exported image during the evaluation, the left one is the first evaluation and the right is the second.<\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/IGFksXy.jpg\" rel=\"nofollow noreferrer\">Exported Image during evaluation<\/a> <\/p>\n\n<p>Seems like the model weight changes during inference, \nhow can I find out the problem? Is there any configuration that I missed? <\/p>\n\n<p>I'm using tensorflow 1.10.1 with python 3.5.2 and cloned object detection API from <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models<\/a> without change.<\/p>\n\n<p>Here's my pipeline.config:<\/p>\n\n<pre><code>model {\n  ssd {\n    num_classes: 1\n    image_resizer {\n      fixed_shape_resizer {\n        height: 640\n        width: 640\n      }\n    }\n    feature_extractor {\n      type: \"ssd_resnet50_v1_fpn\"\n      depth_multiplier: 1.0\n      min_depth: 16\n      conv_hyperparams {\n        regularizer {\n          l2_regularizer {\n            weight: 0.000399999989895\n          }\n        }\n        initializer {\n          truncated_normal_initializer {\n            mean: 0.0\n            stddev: 0.0299999993294\n          }\n        }\n        activation: RELU_6\n        batch_norm {\n          decay: 0.996999979019\n          scale: true\n          epsilon: 0.0010000000475\n        }\n      }\n      override_base_feature_extractor_hyperparams: true\n    }\n    box_coder {\n      faster_rcnn_box_coder {\n        y_scale: 10.0\n        x_scale: 10.0\n        height_scale: 5.0\n        width_scale: 5.0\n      }\n    }\n    matcher {\n      argmax_matcher {\n        matched_threshold: 0.5\n        unmatched_threshold: 0.5\n        ignore_thresholds: false\n        negatives_lower_than_unmatched: true\n        force_match_for_each_row: true\n        use_matmul_gather: true\n      }\n    }\n    similarity_calculator {\n      iou_similarity {\n      }\n    }\n    box_predictor {\n      weight_shared_convolutional_box_predictor {\n        conv_hyperparams {\n          regularizer {\n            l2_regularizer {\n              weight: 0.000399999989895\n            }\n          }\n          initializer {\n            random_normal_initializer {\n              mean: 0.0\n              stddev: 0.00999999977648\n            }\n          }\n          activation: RELU_6\n          batch_norm {\n            decay: 0.996999979019\n            scale: true\n            epsilon: 0.0010000000475\n          }\n        }\n        use_dropout: true\n        dropout_keep_probability: 0.7\n        depth: 256\n        num_layers_before_predictor: 4\n        kernel_size: 3\n        class_prediction_bias_init: -4.59999990463\n      }\n    }\n    anchor_generator {\n      multiscale_anchor_generator {\n        min_level: 3\n        max_level: 7\n        anchor_scale: 4.0\n        aspect_ratios: 1.0\n        aspect_ratios: 2.0\n        aspect_ratios: 0.5\n        scales_per_octave: 2\n      }\n    }\n    post_processing {\n      batch_non_max_suppression {\n        score_threshold: 0.300000011921\n        iou_threshold: 0.600000023842\n        max_detections_per_class: 100\n        max_total_detections: 100\n      }\n      score_converter: SIGMOID\n    }\n    normalize_loss_by_num_matches: true\n    loss {\n      localization_loss {\n        weighted_smooth_l1 {\n        }\n      }\n      classification_loss {\n        weighted_sigmoid_focal {\n          gamma: 2.0\n          alpha: 0.25\n        }\n      }\n      classification_weight: 1.0\n      localization_weight: 1.0\n    }\n    encode_background_as_zeros: true\n    normalize_loc_loss_by_codesize: true\n    inplace_batchnorm_update: true\n    freeze_batchnorm: false\n  }\n}\ntrain_config {\n  batch_size: 8\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n  data_augmentation_options {\n    random_crop_image {\n      min_object_covered: 0.0\n      min_aspect_ratio: 0.75\n      max_aspect_ratio: 3.0\n      min_area: 0.75\n      max_area: 1.0\n      overlap_thresh: 0.0\n    }\n  }\n  sync_replicas: false\n  optimizer {\n    adam_optimizer: {\n      learning_rate: {\n        exponential_decay_learning_rate {\n          initial_learning_rate: 0.0001\n          decay_steps: 5000\n          decay_factor: 0.9\n        }\n      }\n    }\n    use_moving_average: false\n  }\n  fine_tune_checkpoint: \"\/tf-object-detection-training\/models\/ssd_resnet50\/saved\/model.ckpt-652123\"\n  num_steps: 2000000\n  from_detection_checkpoint: true\n  load_all_detection_checkpoint_vars: true\n  startup_delay_steps: 0.0\n  replicas_to_aggregate: 8\n  max_number_of_boxes: 100\n  unpad_groundtruth_tensors: false\n}\ntrain_input_reader {\n  label_map_path: \"\/tf-object-detection-training\/dataset_VOC\/label.pbtxt\"\n  tf_record_input_reader {\n    input_path: \"\/tf-object-detection-training\/dataset_VOC\/person_train.record-?????-of-00010\"\n  }\n}\neval_config {\n  num_examples: 10000\n  num_visualizations: 100\n  eval_interval_secs: 60\n  metrics_set: \"coco_detection_metrics\"\n  use_moving_averages: false\n  min_score_threshold: 0.5\n  retain_original_images: false\n  keep_image_id_for_visualization_export: true\n  visualization_export_dir: \"\/tf-object-detection-training\/models\/ssd_resnet50\/eval_detections\/\"\n}\neval_input_reader {\n  label_map_path: \"\/tf-object-detection-training\/dataset_VOC\/label.pbtxt\"\n  shuffle: false\n  num_readers: 1\n  tf_record_input_reader {\n    input_path: \"\/tf-object-detection-training\/dataset_VOC\/person_val.record-?????-of-00010\"\n  }\n}\n<\/code><\/pre>\n\n<p>Thanks for any advice<\/p>\n","answers":[{"AnswerId":"52947248","CreationDate":"2018-10-23T10:53:21.520","ParentId":null,"OwnerUserId":"10527330","Title":null,"Body":"<p>After tracing code for a long time, I've found the answer is that the 'use_dropout' flag set in the pipeline.config.\nSeems like the dropout function is not removed while doing inference so the eval.py and frozen_inference_graph were all applied dropout function and making random inference. <\/p>\n\n<p>To solve this, simply remove 'use_dropout' from pipeline.config fixes this.<\/p>\n"}]}
+{"QuestionId":52887147,"AnswerCount":0,"Tags":"<tensorflow><keras><seq><tensorflowjs-converter>","CreationDate":"2018-10-19T06:49:59.657","AcceptedAnswerId":null,"OwnerUserId":9798214.0,"Title":"Orthogonal initializer is being called on a matrix with more elements: Slowness may result","Body":"<p>tensorflowjs version: 0.11.7<br>\nKeras version: 2.0.4<\/p>\n\n<p>I am trying to run Keras converted model on browser. For conversion I have used tensorflowjs converter, the conversion went fine. However, at load time,<\/p>\n\n<blockquote>\n  <p>Orthogonal initializer is being called on a matrix with more than 2000\n  (2560000) elements: Slowness may result.<\/p>\n<\/blockquote>\n\n<p>This message pops up and the memory blowup happens and lastly the browser stops working.\nFollowing is my keras model architecture:<img src=\"https:\/\/i.stack.imgur.com\/NhKbK.png\" alt=\"\">\nAs we run the tensorflowjs code to load the aforementioned model, following warning pops up:<img src=\"https:\/\/i.stack.imgur.com\/dMhPh.png\" alt=\"\"><\/p>\n\n<p>Is there a way to speed up the Orthogonal initialization process, so the model gets loaded in less time and with less effort?\nAny help will be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52887179,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-19T06:52:43.383","AcceptedAnswerId":52890312.0,"OwnerUserId":9666574.0,"Title":"TensorFlow outer product of two 2-rank tensors","Body":"<p>Suppose I have a rank two tensor, <code>[[a,b],[c,d]]<\/code>(in general a m-by-n matrix). <\/p>\n\n<p>I want to expand every element by a 2-2 identity matrix(outer product) and result in <\/p>\n\n<pre><code>[[a, 0, b, 0],[0,a,0,b],[c,0,d,0],[0,c,0,d]]. \n<\/code><\/pre>\n\n<p>What is the most efficient way to implement it in tensorflow? <\/p>\n\n<p>This operation appears a lot in the framework. <\/p>\n","answers":[{"AnswerId":"52890312","CreationDate":"2018-10-19T10:13:15.400","ParentId":null,"OwnerUserId":"10356594","Title":null,"Body":"<p>I would like to do this in 2-step process.\nIf I have a m-by-n matrix, and a 2-2 identity matrix.\nFirst I will enlarge (repeat) the matrix to \"2m-2n matrix\"<br>\nand then multiply by the enlarged identity matrix (2m-2n matrix). As following shows.<\/p>\n\n<pre><code>import tensorflow as tf\n#Process 1  repeat the tensor in 2D.\n#e.g. [1,2,3,4]  --&gt; [[1,1,2,2],[1,1,2,2],[3,3,4,4],[3,3,4,4]]\n\n# assuming  m x n idenity matrix e.g. [1,2,3],[4,5,6]] , m=2,n=3\nid_matrix_size=2  # size of identity matrix (e.g. 2x2 3x3 ...)\nm=2\nn=3\nmytensor=tf.constant([[1,2,3],[4,5,6] ],dtype = tf.float32)\n\n# similar to np.repeat in x-dimension.\nflatten_mytensor=tf.reshape(mytensor,[-1,1])  \na=tf.tile(flatten_mytensor,  [1, id_matrix_size])\nb=tf.reshape( a, [m,-1])\n\n# tile in y-dimension\nc=tf.tile(b,[1,id_matrix_size])\nd=tf.reshape(c,[id_matrix_size*m,id_matrix_size*n])\n\n#Process 2  tile identity matrix in 2D. \nidentity_matrix=tf.eye(id_matrix_size) # identity matrix \nidentity_matrix_2D= tf.tile(identity_matrix,[m,n])\n\n#Process 3  elementwise multiply\noutput = tf.multiply(d,identity_matrix_2D )\n\nwith tf.Session() as sess:\n    print(sess.run(output) )\n#output :\n#[[1. 0. 2. 0. 3. 0.]\n# [0. 1. 0. 2. 0. 3.]\n# [4. 0. 5. 0. 6. 0.]\n# [0. 4. 0. 5. 0. 6.]]\n<\/code><\/pre>\n\n<p>And also, a def is more convenient if lots of implements are needed.<\/p>\n\n<pre><code>def Expand_tensor(mytensor,id_matrix_size):\n    m=mytensor.shape[0]    \n    n=mytensor.shape[1]\n    # similar to np.repeat in x-dimension.\n    flatten_mytensor=tf.reshape(mytensor,[-1,1])  \n    a=tf.tile(flatten_mytensor,  [1, id_matrix_size])\n    b=tf.reshape( a, [m,-1])\n\n    # tile in y-dimension\n    c=tf.tile(b,[1,id_matrix_size])\n    d=tf.reshape(c,[id_matrix_size*m,id_matrix_size*n])\n\n\n    # tile identity matrix in 2D. \n    identity_matrix=tf.eye(id_matrix_size) # identity matrix \n    identity_matrix_2D= tf.tile(identity_matrix,[m,n])\n\n    #elementwise multiply\n    output = tf.multiply(d,identity_matrix_2D )\n    return output\n\nmytensor=tf.constant([[1,2,3],[4,5,6] ],dtype = tf.float32)\nwith tf.Session() as sess:\n    print(sess.run(Expand_tensor(mytensor,2)) )\n<\/code><\/pre>\n"}]}
+{"QuestionId":52887227,"AnswerCount":1,"Tags":"<image><keras><generator>","CreationDate":"2018-10-19T06:56:21.073","AcceptedAnswerId":52925920.0,"OwnerUserId":10484156.0,"Title":"Understanding 'follow_links' argument in Keras's ImageDataGenerator?","Body":"<p>The Keras documentations says <\/p>\n\n<blockquote>\n  <p>'follow_links: Whether to follow symlinks inside class subdirectories\n  (default: False)'<\/p>\n<\/blockquote>\n\n<p>, but what is the exact role of this argument?<\/p>\n","answers":[{"AnswerId":"52925920","CreationDate":"2018-10-22T09:15:20.840","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>Symlinks are used to link a file or directory. For example you have a symlink, which points to the latest version of any installed version of python.exe (or any other file or directory), while you have multiple version installed. <\/p>\n\n<p>Following this symlink meens you don\u00b4t see the symlink as a symlink, but resolve the target file\/directory, where it points to. In the python example you would open the latest python console. In the keras example keras would also consider symlinks as image input data. A long you don\u00b4t have symlinks in your data directory, you can just let it be False.<\/p>\n\n<p>See <a href=\"https:\/\/en.wikipedia.org\/wiki\/Symbolic_link\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Symbolic_link<\/a><\/p>\n"}]}
+{"QuestionId":52887336,"AnswerCount":1,"Tags":"<keras><package><python-import><google-colaboratory>","CreationDate":"2018-10-19T07:05:41.777","AcceptedAnswerId":52887544.0,"OwnerUserId":4058766.0,"Title":"Unable to import the latest version of a package: Google Colab","Body":"<p>I am trying to use version 2.2.2 of keras in Google Colab in order to get around a compatibility issue with keras_contrib. I first ran<\/p>\n\n<pre><code> !pip uninstall keras %to get rid of the existing version\n !pip install keras==2.2.2\n<\/code><\/pre>\n\n<p>then <\/p>\n\n<pre><code> import keras\n print(keras.__version__)  \n<\/code><\/pre>\n\n<p>this gave the output<\/p>\n\n<pre><code> 2.1.6\n<\/code><\/pre>\n\n<p>But, when I run<\/p>\n\n<pre><code> !pip show keras\n<\/code><\/pre>\n\n<p>it shows version 2.2.2 is installed.\nIn any case, importing keras_contrib generates an error which leads me to believe that the older version (2.1.6) is being used.<\/p>\n\n<p>Why isn't version 2.2.2 being installed? <\/p>\n","answers":[{"AnswerId":"52887544","CreationDate":"2018-10-19T07:21:25.960","ParentId":null,"OwnerUserId":"7545460","Title":null,"Body":"<p>What version of Python are you using?\nIf you're using Python 3, then you should use the command pip3 to install keras.<\/p>\n"}]}
+{"QuestionId":52887454,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-19T07:16:21.827","AcceptedAnswerId":null,"OwnerUserId":3781647.0,"Title":"how to use placeholder in tensorflow in below case","Body":"<p>For the below code, can I use placeholder for the n, and pass the n by feeding?\nI am quite new to tensorflow.<\/p>\n\n<pre><code>n = int(input(\"Enter an integer: \"))\n\none=tf.constant(1)\n#increase\nincreasing_value=tf.Variable(0,name=\"increasing_value\")\nincreasing_op=tf.assign_add(increasing_value,one)\n\n#sum\nsumvalue=tf.Variable(0,name=\"sumvalue\")\nsum_op=tf.assign_add(sumvalue,increasing_value)\n\ninit=tf.global_variables_initializer()\n\nwith tf.Session() as session:\n    session.run(init)\n    for _ in range (n):\n        session.run(increasing_op)\n        session.run(sum_op)\n    print(session.run(sumvalue))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52889655","CreationDate":"2018-10-19T09:33:45.533","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>Can you try this ?<\/p>\n\n<pre><code>n = tf.placeholder(tf.int32, name='n')\n\nfedvalue = session.run( n , feed_dict = { n : 10 })\nfor _ in range ( fedvalue ):\n    session.run(increasing_op)\n    session.run(sum_op)\n<\/code><\/pre>\n"},{"AnswerId":"52889739","CreationDate":"2018-10-19T09:39:21.890","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You could replicate your code in TensorFlow with a <code>tf.while_loop<\/code>.:<\/p>\n\n<pre><code>import tensorflow as tf\n\nn = tf.placeholder(tf.int32, [])\nincreasing_value = tf.constant(0, dtype=tf.int32)\nsum_value = tf.constant(0, dtype=tf.int32)\n\ndef loop_body(i, increasing_value, sum_value):\n    increased_value = increasing_value + 1\n    return i + 1, increased_value, sum_value + increased_value\n\ni = tf.constant(0, dtype=tf.int32)\n_, increasing_value, sum_value = tf.while_loop(\n    lambda i, _, __: i &lt; n,\n    loop_body,\n    [i, increasing_value, sum_value])\n\nwith tf.Session() as session:\n    for x in range(10):\n        print(session.run(sum_value, feed_dict={n: x}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>0\n1\n3\n6\n10\n15\n21\n28\n36\n45\n<\/code><\/pre>\n\n<p>However, TensorFlow loops are generally slow, and you should try to find a vectorial way of doing your computation (that is, express it as operations over arrays of values and reductions). In your particular case, your code is just computing <a href=\"https:\/\/en.wikipedia.org\/wiki\/1_%2B_2_%2B_3_%2B_4_%2B_%E2%8B%AF\" rel=\"nofollow noreferrer\">1 + 2 + 3 + 4 + \u22ef<\/a>, which is simply <code>(n * (n + 1)) \/ 2<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nn = tf.placeholder(tf.int32, [])\nsum_value = (n * n + n) \/\/ 2\n\nwith tf.Session() as session:\n    for x in range(10):\n        print(session.run(sum_value, feed_dict={n: x}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>0\n1\n3\n6\n10\n15\n21\n28\n36\n45\n<\/code><\/pre>\n"}]}
+{"QuestionId":52887606,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-19T07:26:22.107","AcceptedAnswerId":52894038.0,"OwnerUserId":10523748.0,"Title":"L1-regularization to all weights (not weights and bias) in neural network in tensorflow","Body":"<p>I am using TensorFlow's eager execution.    <\/p>\n\n<p>For L1-regularization I am using <code>tf.contrib.layers.l1_regularizer<\/code>. I do not know how I can apply the regularizer just to the weights. <code>tf.contrib.layers.apply_regularization(l1_regularizer, model.trainable_weights)<\/code> also apply the regularization to the bias because <code>model.trainable_weights<\/code> also returns bias. But I do not know how to just get the weights from the model.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.contrib.eager as tfe\nimport numpy as np\n\ntf.enable_eager_execution()\n\n#create loss function with regularizer\ndef loss(model, x, y, l1_regularizer):\n   y_ = model(x)\n   weights = model.trainable_weights\n   return tf.losses.mean_squared_error(labels=y, predictions=y_)+tf.contrib.layers.apply_regularization(l1_regularizer, weights)\n\n #function for gradient calculation\n  def grad(model, inputs, targets, l1_regularizer):\n  with tf.GradientTape() as tape:\n    loss_value = loss(model, inputs, targets, l1_regularizer)\n  return loss_value, tape.gradient(loss_value, model.trainable_variables)\n\n\n  optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.5)\n  global_step = tf.train.get_or_create_global_step()\n\n  model = tf.keras.Sequential([\n  tf.keras.layers.Dense(2, activation=tf.sigmoid, input_shape=(2,)),  # input shape required\n  tf.keras.layers.Dense(2, activation=tf.sigmoid)\n  ])    \n\n #set weights\n  weights=[np.array([[0.15, 0.25],[0.2,0.3]]),np.array([0.35,0.35]),np.array([[0.4,0.5],[0.45, 0.55]]),np.array([0.6,0.6])]\n\nmodel.set_weights(weights)\n\nmodel.get_weights()\n\n features = tf.convert_to_tensor([[0.05,0.10 ]])\n labels =  tf.convert_to_tensor([[0.01,0.99 ]])  \n\n\n model(features) \n #calculate the loss\n loss(model, features, labels,l1_regularizer)\n  #calculate the gradients\n loss, grads = grad(model, features, labels,l1_regularizer)\n #optimization step\n optimizer.apply_gradients(zip(grads, model.variables),\n                                      global_step)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52894038","CreationDate":"2018-10-19T14:06:28.773","ParentId":null,"OwnerUserId":"6832556","Title":null,"Body":"<p>Since you're using Keras layers, the weights usually will have 'kernel' in their name. Use this to subset out the weights from all the trainables.<\/p>\n\n<pre><code>weights = [x for x in model.trainable_weights if 'kernel' in x.name]\ntf.contrib.layers.apply_regularization(l1_regularizer, weights)\n<\/code><\/pre>\n\n<p>Make sure you use the right case for the name, if it's 'Kernel' instead of 'kernel' it won't work.<\/p>\n\n<p>If you don't want to apply L1 to specifically biases<\/p>\n\n<pre><code>weights = [x for x in model.trainable_weights if 'bias' not in x.name]\ntf.contrib.layers.apply_regularization(l1_regularizer, weights)\n<\/code><\/pre>\n\n<hr>\n\n<p>Alternatively, Keras layers also have an option of adding a regularizer for weights, bias or activation. From the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Dense\" rel=\"nofollow noreferrer\">Dense<\/a>: <\/p>\n\n<blockquote>\n  <p><strong>kernel_regularizer<\/strong>: Regularizer function applied to the kernel weights matrix.<\/p>\n  \n  <p><strong>bias_regularizer<\/strong>: Regularizer function applied to the bias vector. <\/p>\n  \n  <p><strong>activity_regularizer<\/strong>: Regularizer function applied to the output of the layer (its \"activation\")..<\/p>\n<\/blockquote>\n\n<p>So you need to import the regularizers from keras and add them as an option to your layers.<\/p>\n\n<pre><code>from tf.keras import regularizers\n    ....\n    ....\n\n model = tf.keras.Sequential([\n  tf.keras.layers.Dense(2, activation=tf.sigmoid, input_shape=(2,), kernel_regularizer=regularizers.l1()),  # input shape required\n  tf.keras.layers.Dense(2, activation=tf.sigmoid, kernel_regularizer=regularizers.l1())\n  ])    \n<\/code><\/pre>\n"}]}
+{"QuestionId":52888026,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-19T07:53:28.303","AcceptedAnswerId":null,"OwnerUserId":1703586.0,"Title":"Complicated LSTM in Keras","Body":"<p>I followed keras lstm examples and can understand those. I understand the single lstm layer networks. <\/p>\n\n<p>Now, I am trying to have something similar to the following,\n<a href=\"https:\/\/i.stack.imgur.com\/WDKIh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WDKIh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am confused as to how to get the output of the last lstm layer (out of 3) as input to the first layer in the next time (l2 in the diagram below)<\/p>\n\n<pre><code>model.add( keras.layers.Embedding( vocab_size, output_dim=64, input_length=t_steps ))\nmodel.add(keras.layers.LSTM(512, return_sequences=True ) )\nmodel.add(keras.layers.LSTM(512, return_sequences=True ) )\nmodel.add(keras.layers.LSTM(512, return_sequences=True ) )\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52888146,"AnswerCount":1,"Tags":"<numpy><tensorflow><keras><deep-learning><pytorch>","CreationDate":"2018-10-19T08:01:42.450","AcceptedAnswerId":null,"OwnerUserId":10258375.0,"Title":"What is the equivalent of torch.nn.functional.grid_sample in Tensorflow \/ Numpy?","Body":"<p>I am new to pytorch and have been trying to convert some code . Can't find this particular functionality . Does it exist in tensorflow ?<\/p>\n","answers":[{"AnswerId":"52896615","CreationDate":"2018-10-19T16:38:21.183","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I do not think there is anything like that provided in TensorFlow. Here is a possible implementation for the 2D case (I have not considered padding, but the code should behave like the <code>border<\/code> mode). Note that, unlike the PyTorch version, I am assuming the input dimension order is <code>(batch_size, height, width, channels)<\/code> (as is common in TensorFlow).<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\ndef grid_sample_2d(inp, grid):\n    in_shape = tf.shape(inp)\n    in_h = in_shape[1]\n    in_w = in_shape[2]\n\n    # Find interpolation sides\n    i, j = grid[..., 0], grid[..., 1]\n    i = tf.cast(in_h - 1, grid.dtype) * (i + 1) \/ 2\n    j = tf.cast(in_w - 1, grid.dtype) * (j + 1) \/ 2\n    i_1 = tf.maximum(tf.cast(tf.floor(i), tf.int32), 0)\n    i_2 = tf.minimum(i_1 + 1, in_h - 1)\n    j_1 = tf.maximum(tf.cast(tf.floor(j), tf.int32), 0)\n    j_2 = tf.minimum(j_1 + 1, in_w - 1)\n\n    # Gather pixel values\n    n_idx = tf.tile(tf.range(in_shape[0])[:, tf.newaxis, tf.newaxis], tf.concat([[1], tf.shape(i)[1:]], axis=0))\n    q_11 = tf.gather_nd(inp, tf.stack([n_idx, i_1, j_1], axis=-1))\n    q_12 = tf.gather_nd(inp, tf.stack([n_idx, i_1, j_2], axis=-1))\n    q_21 = tf.gather_nd(inp, tf.stack([n_idx, i_2, j_1], axis=-1))\n    q_22 = tf.gather_nd(inp, tf.stack([n_idx, i_2, j_2], axis=-1))\n\n    # Interpolation coefficients\n    di = tf.cast(i, inp.dtype) - tf.cast(i_1, inp.dtype)\n    di = tf.expand_dims(di, -1)\n    dj = tf.cast(j, inp.dtype) - tf.cast(j_1, inp.dtype)\n    dj = tf.expand_dims(dj, -1)\n\n    # Compute interpolations\n    q_i1 = q_11 * (1 - di) + q_21 * di\n    q_i2 = q_12 * (1 - di) + q_22 * di\n    q_ij = q_i1 * (1 - dj) + q_i2 * dj\n\n    return q_ij\n\n# Test it\ninp = tf.placeholder(tf.float32, [None, None, None, None])\ngrid = tf.placeholder(tf.float32, [None, None, None, 2])\nres = grid_sample_2d(inp, grid)\nwith tf.Session() as sess:\n    # Make test image\n    im_grid_i, im_grid_j = np.meshgrid(np.arange(6), np.arange(10), indexing='ij')\n    im = im_grid_i + im_grid_j\n    im = im \/ im.max()\n    im = np.stack([im] * 3, axis=-1)\n    # Test grid 1: complete image\n    grid1 = np.stack(np.meshgrid(np.linspace(-1, 1, 15), np.linspace(-1, 1, 18), indexing='ij'), axis=-1)\n    # Test grid 2: lower right corner\n    grid2 = np.stack(np.meshgrid(np.linspace(0, 1, 15), np.linspace(.5, 1, 18), indexing='ij'), axis=-1)\n    # Run\n    res1, res2 = sess.run(res, feed_dict={inp: [im, im], grid: [grid1, grid2]})\n    # Plot image and sampled grids\n    plt.figure()\n    plt.imshow(im)\n    plt.figure()\n    plt.imshow(res1)\n    plt.figure()\n    plt.imshow(res2)\n<\/code><\/pre>\n\n<p>Here are the resulting images, first the input:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WJyYG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WJyYG.png\" alt=\"Input image\"><\/a><\/p>\n\n<p>First grid result, which is the first image but with different shape:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bzGwl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bzGwl.png\" alt=\"First grid result\"><\/a><\/p>\n\n<p>Second grid result, which spans a region in the lower right corner:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/j5Iij.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/j5Iij.png\" alt=\"Second grid result\"><\/a><\/p>\n"}]}
+{"QuestionId":52888624,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-19T08:32:02.593","AcceptedAnswerId":null,"OwnerUserId":9589731.0,"Title":"What dose zero_debias mean in tf.train.ExponentialMovingAverage?","Body":"<p>tf.train.ExponentialMovingAverage has a parameter 'zero_debias', but I don't know what will happen if I set it to True. So what's this parameter for? <\/p>\n\n<p>Great thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":52888657,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-19T08:34:02.960","AcceptedAnswerId":null,"OwnerUserId":9829878.0,"Title":"Is Fid score script correct if I used batchsize as 1","Body":"<p>Is the FID score works correctly in <code>tensorflow &gt; 1.5<\/code> i.e if I run it with batch size 1?\nI am using implementation of FID from <a href=\"https:\/\/github.com\/bioinf-jku\/TTUR\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bioinf-jku\/TTUR<\/a> where there's an issue that it only runs with a batch size of 1.<\/p>\n\n<p>I am running out of time, so I cannot write a script for implementing FID, therefore, can anyone confirm whether it gives correct value or not if I run with a batch size of 1.<\/p>\n","answers":[{"AnswerId":"52898806","CreationDate":"2018-10-19T19:29:24.533","ParentId":null,"OwnerUserId":"10530728","Title":null,"Body":"<p>Yes, the script from <a href=\"https:\/\/github.com\/bioinf-jku\/TTUR\/tree\/16150b2f1ce217c7def60b9b7c1cd180fb6d476c#provided-code\" rel=\"nofollow noreferrer\">TTUR implementation<\/a> that calculates the FID score will work with any batch size value.\nBut note that they load all the images into the memory for this.<\/p>\n\n<p>The mean and covariance matrix calculated for the FID score are defined over the activations of all the images. But to get the activations values we need to load the inception model and get those activations considering the amount of images defined by the batch size. <\/p>\n\n<p>Those activations are calculated and accumulated by the function <code>get_activations<\/code> (<a href=\"https:\/\/github.com\/bioinf-jku\/TTUR\/blob\/16150b2f1ce217c7def60b9b7c1cd180fb6d476c\/fid.py#L67\" rel=\"nofollow noreferrer\">link<\/a>)<\/p>\n\n<pre><code>def get_activations(images, sess, batch_size=50, verbose=False):\n   \"\"\"Calculates the activations of the pool_3 layer for all images.\n   ...\n<\/code><\/pre>\n\n<p>which are used to calculate the activation statistics (mean and covariance) with the function <code>calculate_activation_statistics<\/code> (<a href=\"https:\/\/github.com\/bioinf-jku\/TTUR\/blob\/16150b2f1ce217c7def60b9b7c1cd180fb6d476c\/fid.py#L159\" rel=\"nofollow noreferrer\">link<\/a>). <\/p>\n"}]}
+{"QuestionId":52888824,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-10-19T08:45:04.320","AcceptedAnswerId":52895083.0,"OwnerUserId":9540764.0,"Title":"Define a function to iterate over a tensor before the running phase","Body":"<p>I'm defining a loss function in which I must iterate over the values of a tensor. Of course, this is a python function coded before the training phase and I'm struggling on how to define a for loop on the tensor. In this tensor, there are some zeros and other numbers but I don't know how many, it depends on the training file of the current batch. Could be 2, 5, 10, ... I don't know so I can't use a fixed value. Here an example<\/p>\n\n<pre><code>points = tf.constant([[[0], [1], [0], [2], [1], [0], [2], [2], [0], [1]],\n                      [[1], [2], [0], [0], [1], [0], [1], [2], [0], [2]],\n                      [[0], [2], [1], [0], [2], [0], [1], [2], [0], [1]]], dtype=tf.float32)\nmax_indices = tf.reduce_max(points[:1])\nfor index in xrange(max_indices): # error here\n    # do stuff\n<\/code><\/pre>\n\n<p>This is the error<\/p>\n\n<pre><code>TypeError: an integer is required\n<\/code><\/pre>\n\n<p>So I've tried in another way using this<\/p>\n\n<pre><code>points = tf.constant([[[0], [1], [0], [2], [1], [0], [2], [2], [0], [1]],\n                      [[1], [2], [0], [0], [1], [0], [1], [2], [0], [2]],\n                      [[0], [2], [1], [0], [2], [0], [1], [2], [0], [1]]], dtype=tf.float32)\nitems, _ = tf.unique(tf.reshape(points[:1], [-1]))\nfor item in tf.unstack(items): # error here\n    # do stuff\n<\/code><\/pre>\n\n<p>And the error is<\/p>\n\n<pre><code>ValueError: Cannot infer num from shape (?,)\n<\/code><\/pre>\n\n<p>Of course, these errors are given because in the definition phase I don't have values but I don't know how to fix it.<\/p>\n","answers":[{"AnswerId":"52895083","CreationDate":"2018-10-19T15:02:37.750","ParentId":null,"OwnerUserId":"10458815","Title":null,"Body":"<p>I am assuming based on \"in the definition phase I don't have values\" that you are in a session when you need to iterate on your dataset.<\/p>\n\n<p>You can use a <code>tf.while_loop(cond, body, loop_vars)<\/code> and add an end condition to the end of every tensor. This should be some value that would not show up in your dataset otherwise. For example, if your dataset contains only positive real numbers, you can use <code>-1<\/code> as your end condition.<\/p>\n\n<p>Basically, <code>tf.while_loop(cond, body, loop_vars)<\/code> first calls <code>cond<\/code> which should return a boolean tensor. Meanwhile, if <code>cond<\/code> does not return false, <code>body<\/code> will be called with <code>loop_vars<\/code>. Note that everything you want to use in your loop should be contained in <code>loop_vars<\/code>. Global variables and variables outside of the <code>loop_vars<\/code> context won't work.<\/p>\n\n<p>So you can use <code>t = tf.constant(0)<\/code> as a timestep, and increment it each pass through <code>body<\/code>. Then, in <code>cond<\/code>, you can check if your dataset <code>x<\/code> equals your end condition by using <code>x[t]<\/code> to get the value at every iteration. If it equals this end condition, you have fully iterated through your tensor and <code>cond<\/code> should return false. Otherwise, <code>cond<\/code> should return true, and you can then use <code>t<\/code> as an iterator for some dataset <code>x<\/code> by using <code>x[t]<\/code> in your <code>body<\/code>.<\/p>\n\n<p>This is what I used to solve iterating over variable length sentences in an RNN. I used <code>&lt;end&gt;<\/code> as my end condition, and it works well enough. I hope this also works for you!<\/p>\n"}]}
+{"QuestionId":52888929,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><tensorflow-estimator>","CreationDate":"2018-10-19T08:52:25.770","AcceptedAnswerId":null,"OwnerUserId":8084623.0,"Title":"Stopping criteria for pre-made estimators in TensorFlow","Body":"<p>I have a question about TensorFlow's estimators in <code>tf.estimator<\/code>, in particular <code>DNNClassifier<\/code>. It says in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.7\/api_docs\/python\/tf\/estimator\/DNNClassifier\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><code>max_steps<\/code>: Number of total steps for which to train model. If <code>None<\/code>, train forever or train until <code>input_fn<\/code> generates the <code>OutOfRange<\/code> error or <code>StopIteration<\/code> exception<\/p>\n<\/blockquote>\n\n<p>In the doc on <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets_for_estimators\" rel=\"nofollow noreferrer\">datasets for estimators<\/a> it mentions that for training you need to use the <code>shuffle()<\/code>, <code>repeat()<\/code>, and <code>batch_size<\/code> methods, so that the iterator on the data set does stop after it's gone through the data once.<\/p>\n\n<p>Does this mean that the pre-made estimators such as <code>DNNClassifier<\/code> have no stopping criterion based on the learning rate or changes in the loss? Is it really the case you can only have these models stop training based on how you specify the input function or by giving a maximum number of steps?<\/p>\n","answers":[{"AnswerId":"52896970","CreationDate":"2018-10-19T17:04:16.620","ParentId":null,"OwnerUserId":"9576876","Title":null,"Body":"<p>TensforFlow will not presume to know upon what learning rate or loss it should stop at.  This is reasonable because they're problem-dependent.  You could reasonably argue that it could infer a sensible limit based on round-off error for the given data types (if they're consistent, e.g. float32) but then many problems should be stopped earlier.  So there is no sensible, broadly applicable default.<\/p>\n\n<p>However, you can control this behaviour yourself using callbacks.  TensorFlow includes the <code>EarlyStopping<\/code> callback.  You can find the (python) documentation for it <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/EarlyStopping\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52889099,"AnswerCount":1,"Tags":"<android><tensorflow><xamarin><tensorflow-lite>","CreationDate":"2018-10-19T09:02:38.360","AcceptedAnswerId":null,"OwnerUserId":8511928.0,"Title":"Xamarin tf.lite input objects","Body":"<p>Im trying to reproduce tensorflow object detection on xamarin. <\/p>\n\n<pre><code>    private MappedByteBuffer LoadModelFile()\n    {\n        AssetFileDescriptor fileDescriptor = Assets.OpenFd(\"detect.tflite\");           \n        FileInputStream inputStream = new FileInputStream(fileDescriptor.FileDescriptor);\n        FileChannel fileChannel = inputStream.Channel;\n        long startOffset = fileDescriptor.StartOffset;\n        long declaredLength = fileDescriptor.DeclaredLength;\n        return fileChannel.Map(FileChannel.MapMode.ReadOnly, startOffset, declaredLength);\n    }\n\n        View view = (View) sender;\n\n        MappedByteBuffer buffer = LoadModelFile();\n\n        Interpreter interpreter = new Interpreter(buffer);\n\n        var sr = new StreamReader(Assets.Open(\"labels.txt\"));\n\n        var labels = sr.ReadToEnd()\n            .Split('\\n')\n            .Select(s =&gt; s.Trim())\n            .Where(s =&gt; !string.IsNullOrEmpty(s))\n            .ToList();\n\n        var bitmap =  BitmapFactory.DecodeResource(Resources, 2130837608);\n        var resizedBitmap = Bitmap.CreateScaledBitmap(bitmap, 1000, 750, false)\n            .Copy(Bitmap.Config.Argb8888, false);\n\n        float[][][][] imgData = null;\n\n        imgData  = new float[1][][][];\n        imgData[0] = new float[1000][][];\n\n        for (int i = 0; i &lt; imgData[0].Length; i++)\n        {\n            imgData[0][i] = new float[750][];\n            for (int j = 0; j &lt; imgData[0][i].Length; j++)\n            {\n                imgData[0][i][j] = new float[3];\n            }\n        }\n        var intValuess = new int[1000 * 750];\n        resizedBitmap.GetPixels(intValuess, 0, 1000, 0, 0, 1000, 750);\n\n        int pixels = 0;\n        for (int i = 0; i &lt; imgData[0].Length; i++)\n        {\n            for (int j = 0; j &lt; imgData[0][i].Length; j++)\n            {\n                var val = intValuess[pixels++];\n                imgData[0][i][j][0] = (float)((val &gt;&gt; 16) &amp; 0xFF);\n                imgData[0][i][j][1] = (float)((val &gt;&gt; 8) &amp; 0xFF);\n                imgData[0][i][j][2] = (float)(val &amp; 0xFF);\n            }\n        }\n\n        var outputs = new float[labels.Count];\n\n        interpreter.Run(imgData, outputs);\n<\/code><\/pre>\n\n<p>but i have error \"cannot convert float[][][][] to Java.Lang.Object in line interpreter.Run(imgData, outputs);\nHow i can convert float[][][][] to Java.Lang.Object or where i can find tensorflow lite with xamarin examples. <\/p>\n","answers":[{"AnswerId":"57430061","CreationDate":"2019-08-09T12:26:56.277","ParentId":null,"OwnerUserId":"11872776","Title":null,"Body":"<p>I know it has been a while since you asked this question but maybe my response can be useful to someone.\nI am also trying to use Xamarin with tflite, to run a simple CNN.\nHere is my code:<\/p>\n\n<pre><code>private MappedByteBuffer LoadModelFile()\n{\n    var assets = Application.Context.Assets;\n    AssetFileDescriptor fileDescriptor = assets.OpenFd(\"seed_model_no_qt.tflite\");\n    FileInputStream inputStream = new FileInputStream(fileDescriptor.FileDescriptor);\n    FileChannel fileChannel = inputStream.Channel;\n    long startOffset = fileDescriptor.StartOffset;\n    long declaredLength = fileDescriptor.DeclaredLength;\n    return fileChannel.Map(FileChannel.MapMode.ReadOnly, startOffset, declaredLength);\n}\n\nprivate string Classify(MediaFile mediaFile)\n{\n    var assets = Application.Context.Assets;\n\n    Bitmap bp = BitmapFactory.DecodeStream(mediaFile.GetStream());\n    var resizedBitmap = Bitmap.CreateScaledBitmap(bp, 1280, 1280, false).Copy(Bitmap.Config.Argb8888, false);\n\n    var bufint = new int[1280 * 1280];\n    resizedBitmap.GetPixels(bufint, 0, 1280, 0, 0, 1280, 1280);\n    int pixels = 0;\n    var input_buffer = new byte[4 * 1280 * 1280 * 3];            \n    for(int i = 0; i &lt; 1280; i++)\n    {\n        for(int k = 0; k &lt; 1280; k++)\n        {\n            int val = bufint[pixels++];\n            Array.Copy(BitConverter.GetBytes(((val &gt;&gt; 16) &amp; 0xFF) * (1f \/ 255f)), 0, input_buffer, (i * 1280 + k) * 12, 4);\n            Array.Copy(BitConverter.GetBytes(((val &gt;&gt; 8) &amp; 0xFF) * (1f \/ 255f)), 0, input_buffer, (i * 1280 + k) * 12 + 4, 4);\n            Array.Copy(BitConverter.GetBytes((val &amp; 0xFF) * (1f \/ 255f)), 0, input_buffer, (i * 1280 + k) * 12 + 8, 4);\n        }\n    }\n    var bytebuffer = Java.Nio.ByteBuffer.Wrap(input_buffer);\n    var output = Java.Nio.ByteBuffer.AllocateDirect(4*160*160);\n    interpreter.Run(bytebuffer, output);\n\n    var buffer = new byte[4 * 160 * 160];\n\n    Marshal.Copy(output.GetDirectBufferAddress(), buffer, 0, 4 * 160 * 160);\n\n    float sum = 0.0f;\n    for(int i = 0; i &lt; 160*160; i++)\n    {\n        sum += BitConverter.ToSingle(buffer, i * 4);\n    }\n\n    return \"Count : \" + ((int)(sum\/255)).ToString();\n}\n<\/code><\/pre>\n\n<p>I reused your LoadModelFile() function as it is. The code takes an image from a mediaFile (coming from the phone camera), then resizes it to 1280x1280 rgb image before feeding it to a CNN as an array of float32 values.\nYour float[][][][] to Java.Lang.Object issue came from the interpreter.Run() method expecting a Java Object. Some people online solve it by giving a Java.Nio.ByteBuffer as a parameter, instead of an array. It implies some bitwise manipulations but the Run method does accept the ByteBuffer object.\nWhen filling the ByteBuffer, I advise you not to use its methods such as PutFloat(), but to fill a byte[] buffer and then use the Java.Nio.ByteBuffer.Wrap() method as I did. Using ByteBuffer's methods seemed to imply large performance issues in my case.\nSame thing happens when manipulating the output of my CNN (a 160x160 heatmap of float32 values). Using ByteBuffer.Get() method to access the values was very slow. Instead, use Marshal.Copy to store the values into a byte array, then get back the float values with BitConverter.ToSingle.<\/p>\n"}]}
+{"QuestionId":52889385,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-19T09:17:56.713","AcceptedAnswerId":52889791.0,"OwnerUserId":7857121.0,"Title":"You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [2,2]","Body":"<p>I am having <code>You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [2,2]<\/code>error.I have already fed the placeholder<code>aa<\/code><\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\naa = tf.placeholder(dtype=tf.float32, shape=(2, 2))\nbb = tf.Variable(aa)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n  sess.run(init)\n  print(sess.run(bb,feed_dict={aa:np.random.rand(2,2)}))\n<\/code><\/pre>\n","answers":[{"AnswerId":"52889791","CreationDate":"2018-10-19T09:42:36.213","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The problem is in <code>sess.run(init)<\/code>; you need a value for <code>aa<\/code> in order to initialize <code>bb<\/code>. <code>aa<\/code> is not needed to retrieve <code>bb<\/code> later, though, since it already has been assigned a value.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\naa = tf.placeholder(dtype=tf.float32, shape=(2, 2))\nbb = tf.Variable(aa)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n  sess.run(init, feed_dict={aa: np.random.rand(2, 2)})\n  print(sess.run(bb))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52889973,"AnswerCount":1,"Tags":"<python><subprocess><pytorch>","CreationDate":"2018-10-19T09:52:28.023","AcceptedAnswerId":null,"OwnerUserId":6631334.0,"Title":"CUDA error: out of memory when using subprocess to launch multiple training process in Pytorch","Body":"<p>I am training some neural networks, and I have a script to search for hyperparameters. <\/p>\n\n<pre><code>\"\"\"Peform hyperparemeters search\"\"\"\n\nimport argparse\nimport os\nfrom subprocess import check_call\nimport sys\nimport numpy as np\n\nimport utils\n\n\nPYTHON = sys.executable\nparser = argparse.ArgumentParser()\nparser.add_argument('--parent_dir', default='experiments\/learning_rate',\n                    help='Directory containing params.json')\nparser.add_argument('--data_dir', default='data', help=\"Directory containing the dataset\")\n\n\ndef launch_training_job(parent_dir, data_dir, job_name, params):\n    \"\"\"Launch training of the model with a set of hyperparameters in parent_dir\/job_name\n\n    Args:\n        model_dir: (string) directory containing config, weights and log\n        data_dir: (string) directory containing the dataset\n        params: (dict) containing hyperparameters\n    \"\"\"\n    # Create a new folder in parent_dir with unique_name \"job_name\"\n    model_dir = os.path.join(parent_dir, job_name)\n    if not os.path.exists(model_dir):\n        os.makedirs(model_dir)\n\n    # Write parameters in json file\n    json_path = os.path.join(model_dir, 'params.json')\n    params.save(json_path)\n\n    # Launch training with this config\n    cmd = \"{python} -W ignore train.py --model_dir={model_dir} --data_dir {data_dir}\".format(python=PYTHON, \n                                                                                             model_dir=model_dir, \n                                                                                             data_dir=data_dir)\n    print(cmd)\n    check_call(cmd, shell=True)\n\n\nif __name__ == \"__main__\":\n    # Load the \"reference\" parameters from parent_dir json file\n    args = parser.parse_args()\n    json_path = os.path.join(args.parent_dir, 'params.json')\n    assert os.path.isfile(json_path), \"No json configuration file found at {}\".format(json_path)\n    params = utils.Params(json_path)\n\n    # Perform hypersearch over one parameter\n    learning_rates = [1e-5]\n\n    for learning_rate in learning_rates:\n        # Modify the relevant parameter in params\n        params.learning_rate = learning_rate\n\n        # Launch job (name has to be unique)\n        job_name = \"learning_rate_{}\".format(learning_rate)\n        launch_training_job(args.parent_dir, args.data_dir, job_name, params)\n<\/code><\/pre>\n\n<p>If I try calling train.py by itself, it does not run out of memory - it only uses around 2500MB of the 12000MB available on the GPU. <\/p>\n\n<p>But if I call this script to search hyperparameters, it will run out of memory EVEN if I call it with a single subprocess, specifically just testing ONE learning rate. <\/p>\n\n<p>This is the error message: <code>RuntimeError: CUDA error: out of memory<\/code>. <\/p>\n\n<p>Note that this does not happen for another similar model that I have that's smaller. <\/p>\n\n<p>Why does this happen? And how can I solve it? <\/p>\n","answers":[{"AnswerId":"52891229","CreationDate":"2018-10-19T11:10:13.477","ParentId":null,"OwnerUserId":"8231100","Title":null,"Body":"<p>I don't have enough reputation to comment,so I will write an answer.You should try looking into this link ifyou are using a Unix system:\n<a href=\"https:\/\/stackoverflow.com\/questions\/5306075\/python-memory-allocation-error-using-subprocess-popen\">Python memory allocation error using subprocess.Popen<\/a> \nIt seems that Unix has an issue with subprocesses...\nHope this can helps.<\/p>\n"}]}
+{"QuestionId":52890108,"AnswerCount":1,"Tags":"<tensorflow><profiling><tensorflow-xla>","CreationDate":"2018-10-19T10:00:17.893","AcceptedAnswerId":null,"OwnerUserId":4484867.0,"Title":"How to open tensorflow xla","Body":"<p>I tried the example under tensorflow\/examples\/tutorials\/mnist &amp; try to apply xla for speedup.\nHowever, I cannot see XlaLaunch as <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/jit\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/xla\/jit<\/a> said.<\/p>\n\n<p>Also, I try to profile execution time by:<\/p>\n\n<pre><code>  train_loops = 100000\n  t_start = time.time()\n  for i in range(train_loops):\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n\n    # Create a timeline for the last loop and export to json to view with\n    # chrome:\/\/tracing\/.\n    if i == train_loops - 1:\n      sess.run(train_step,\n               feed_dict={x: batch_xs,\n                          y_: batch_ys},\n               options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE),\n               run_metadata=run_metadata)\n      trace = timeline.Timeline(step_stats=run_metadata.step_stats)\n      with open('timeline.ctf.json', 'w') as trace_file:\n        trace_file.write(trace.generate_chrome_trace_format())\n    else:\n      sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})\n\n  tdiff = time.time() - t_start\n  print(\"tdiff\", tdiff, \" i = \", i)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1wtnd.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1wtnd.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>It seems to be no difference between with &amp; without xla.<\/strong><\/p>\n\n<p><strong>I saw some articles said I should \"rebuild\" tensorflow source to open xla? should I?<\/strong> <\/p>\n\n<p><strong>Is there any other way to open it? Or it has been opened by default but I am in wrong way to use it.<\/strong><\/p>\n\n<p><strong>Is there benchmark to profile the speedup once xla has been opened?<\/strong><\/p>\n\n<p>Thanks~<\/p>\n","answers":[{"AnswerId":"54875939","CreationDate":"2019-02-25T22:53:37.603","ParentId":null,"OwnerUserId":"7891960","Title":null,"Body":"<p>The latest TensorFlow binaries available in pip may include XLA support built in already. <code>pip install tensorflow --upgrade --force-reinstall<\/code> or <code>pip install tensorflow-gpu --upgrade --force-reinstall<\/code> will provide this support for you (as of TF 1.12). This works for me on Ubuntu with or without NVIDIA GPU support, but not on my MacBook. If your binary doesn't, you'll know below when trying to use the benchmark. <\/p>\n\n<p>If you are properly calling the XLA JIT and you do not have support built into the binary, you would get an exception. Something like <\/p>\n\n<p><code>ValueError: Op type not registered 'XlaClusterOutput' in binary running on localhost. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib, accessing (e.g.) tf.contrib.resampler should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed. while building NodeDef 'tower_0\/v\/output0'\n<\/code><\/p>\n\n<p>If you get this, try installing the latest tensorflow, or even tf-nightly. If that doesn't work, time to cross your fingers and <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">build from source<\/a>.<\/p>\n\n<p>Where in your code did you enable XLA?<\/p>\n\n<p>For a test benchmark, you can try to the official TensorFlow Benchmarks repository <a href=\"https:\/\/github.com\/tensorflow\/benchmarks\/tree\/master\/scripts\/tf_cnn_benchmarks\" rel=\"nofollow noreferrer\">here<\/a>. You can use synthetic data if you don't have ImageNet downloaded. You can enable or disable XLA with the flag <code>--xla_compile=True<\/code> or <code>False<\/code>.<\/p>\n\n<p>For instance:<\/p>\n\n<pre><code>python tf_cnn_benchmarks.py --num_batches=2000 --xla_compile=True\n<\/code><\/pre>\n"}]}
+{"QuestionId":52890394,"AnswerCount":1,"Tags":"<python><tensorflow><slice><lstm>","CreationDate":"2018-10-19T10:17:28.987","AcceptedAnswerId":52890674.0,"OwnerUserId":2886527.0,"Title":"Slice a 3D tensor, based on the given sequence length array in tensorflow","Body":"<p>I want a tensorflow function, which accepts a 3D matrix and an array ( shape of the array is similar to the first dimension of a 3D matrix ) and I want to slice the elements from each 2D matrix inside the 3D matrix based on the given array. The equivalent numpy looks like as follows. The basic idea is to picking all hidden states of each input in a batch ( avoid the padded ) in a dynamic rnn    <\/p>\n\n<pre><code>import numpy as np\na = np.random.uniform(-1,1,(3,5,7))\na_length = np.random.randint(5,size=(3))\n\na_tf = tf.convert_to_tensor(a)\na_length_tf = tf.convert_to_tensor(a_length)\n\nres = []\nfor index, length_ in enumerate(a_length):\n    res.extend(a[index,:length_,:])\nres = np.array(res)\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<pre><code>print(a_length)\narray([1, 4, 4])\n\n\nprint(res)\n\narray([[-0.060161  ,  0.36000953,  0.46160677, -0.66576281,  0.28562044,\n    -0.60026872,  0.08034777],\n   [ 0.04776443,  0.38018207, -0.73352382,  0.61847258, -0.89731857,\n     0.57264147, -0.88192537],\n   [ 0.92657628,  0.6236141 ,  0.41977008,  0.88720247,  0.44639323,\n     0.26165976,  0.2678753 ],\n   [-0.78125831,  0.76756136, -0.05716537, -0.64696257,  0.48918477,\n     0.15376225, -0.41974593],\n   [-0.625326  ,  0.3509537 , -0.7884495 ,  0.11773297,  0.23713942,\n     0.30296786,  0.12932378],\n   [ 0.88413986, -0.10958306,  0.9745586 ,  0.8975006 ,  0.23023047,\n    -0.89991669, -0.60032688],\n   [ 0.33462775,  0.62883724, -0.81839566, -0.70312966, -0.00246936,\n    -0.95542994, -0.33035891],\n   [-0.26355579, -0.58104982, -0.54748412, -0.30236209, -0.74270132,\n     0.46329941,  0.34277915],\n   [ 0.92837516, -0.06748299,  0.32837354, -0.62863672,  0.86226447,\n     0.63604586,  0.0905248 ]])\n\nprint(a)\narray([[[-0.060161  ,  0.36000953,  0.46160677, -0.66576281,\n      0.28562044, -0.60026872,  0.08034777],\n    [ 0.26379226,  0.67066755, -0.90139221, -0.86862163,\n      0.36405595,  0.71342926, -0.1265208 ],\n    [ 0.15007877,  0.82065234,  0.03984378, -0.20038364,\n     -0.09945102,  0.71605241, -0.55865999],\n    [ 0.27132257, -0.84289149, -0.15493576,  0.74683429,\n     -0.71159896,  0.50397217, -0.99025404],\n    [ 0.51546368,  0.45460343,  0.87519031,  0.0332339 ,\n     -0.53474897, -0.01733648, -0.02886814]],\n\n   [[ 0.04776443,  0.38018207, -0.73352382,  0.61847258,\n     -0.89731857,  0.57264147, -0.88192537],\n    [ 0.92657628,  0.6236141 ,  0.41977008,  0.88720247,\n      0.44639323,  0.26165976,  0.2678753 ],\n    [-0.78125831,  0.76756136, -0.05716537, -0.64696257,\n      0.48918477,  0.15376225, -0.41974593],\n    [-0.625326  ,  0.3509537 , -0.7884495 ,  0.11773297,\n      0.23713942,  0.30296786,  0.12932378],\n    [ 0.44550219, -0.38828221,  0.35684203,  0.789946  ,\n     -0.8763921 ,  0.90155917, -0.75549455]],\n\n   [[ 0.88413986, -0.10958306,  0.9745586 ,  0.8975006 ,\n      0.23023047, -0.89991669, -0.60032688],\n    [ 0.33462775,  0.62883724, -0.81839566, -0.70312966,\n     -0.00246936, -0.95542994, -0.33035891],\n    [-0.26355579, -0.58104982, -0.54748412, -0.30236209,\n     -0.74270132,  0.46329941,  0.34277915],\n    [ 0.92837516, -0.06748299,  0.32837354, -0.62863672,\n      0.86226447,  0.63604586,  0.0905248 ],\n    [ 0.70272633,  0.17122912, -0.58209965,  0.55557024,\n     -0.46295566, -0.33845157, -0.62254313]]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52890674","CreationDate":"2018-10-19T10:36:40.017","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a way to do that using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/boolean_mask\" rel=\"nofollow noreferrer\"><code>tf.boolean_mask<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# NumPy\/Python implementation\na = np.random.uniform(-1,1,(3,5,7)).astype(np.float32)\na_length = np.random.randint(5,size=(3)).astype(np.int32)\nres = []\nfor index, length_ in enumerate(a_length):\n    res.extend(a[index,:length_,:])\nres = np.array(res)\n\n# TensorFlow implementation\na_tf = tf.convert_to_tensor(a)\na_length_tf = tf.convert_to_tensor(a_length)\n# Make a mask for all wanted elements\nmask = tf.range(tf.shape(a)[1]) &lt; a_length_tf[:, tf.newaxis]\n# Apply mask\nres_tf = tf.boolean_mask(a_tf, mask)\n# Test\nwith tf.Session() as sess:\n    print(np.allclose(sess.run(res_tf), res))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>True\n<\/code><\/pre>\n"}]}
+{"QuestionId":52890944,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-19T10:52:36.260","AcceptedAnswerId":null,"OwnerUserId":10528024.0,"Title":"How to use properly Tensorflow Dataset with batch?","Body":"<p>I am new to Tensorflow and deep learning, and I am struggling with the Dataset class. I tried a lot of things and I can\u2019t find a good solution.<\/p>\n\n<h2>What I am trying<\/h2>\n\n<p>I have a large amount of images (500k+) to train my DNN with. This is a denoising autoencoder so I have a pair of each image. I am using the dataset class of TF to manage the data, but I think I use it really badly.<\/p>\n\n<p>Here is how I load the filenames in a dataset:<\/p>\n\n<pre><code>class Data:\ndef __init__(self, in_path, out_path):\n    self.nb_images = 512\n    self.test_ratio = 0.2\n    self.batch_size = 8\n\n    # load filenames in input and outputs\n    inputs, outputs, self.nb_images = self._load_data_pair_paths(in_path, out_path, self.nb_images)\n\n    self.size_training = self.nb_images - int(self.nb_images * self.test_ratio)\n    self.size_test = int(self.nb_images * self.test_ratio)\n\n    # split arrays in training \/ validation\n    test_data_in, training_data_in = self._split_test_data(inputs, self.test_ratio)\n    test_data_out, training_data_out = self._split_test_data(outputs, self.test_ratio)\n\n    # transform array to tf.data.Dataset\n    self.train_dataset = tf.data.Dataset.from_tensor_slices((training_data_in, training_data_out))\n    self.test_dataset = tf.data.Dataset.from_tensor_slices((test_data_in, test_data_out))\n<\/code><\/pre>\n\n<p>I have a function to call at each epoch that will prepare the dataset. It shuffles the filenames, and transforms filenames to images and batch data.<\/p>\n\n<pre><code>def get_batched_data(self, seed, batch_size):\n    nb_batch = int(self.size_training \/ batch_size)\n\n    def img_to_tensor(path_in, path_out):\n        img_string_in = tf.read_file(path_in)\n        img_string_out = tf.read_file(path_out)\n        im_in = tf.image.decode_jpeg(img_string_in, channels=1)\n        im_out = tf.image.decode_jpeg(img_string_out, channels=1)\n        return im_in, im_out\n\n    t_datas = self.train_dataset.shuffle(self.size_training, seed=seed)\n    t_datas = t_datas.map(img_to_tensor)\n    t_datas = t_datas.batch(batch_size)\n    return t_datas\n<\/code><\/pre>\n\n<p>Now during the training, at each epoch we call the <code>get_batched_data<\/code> function, make an iterator, and run it for each batch, then feed the array to the optimizer operation.<\/p>\n\n<pre><code>for epoch in range(nb_epoch):\n    sess_iter_in = tf.Session()\n    sess_iter_out = tf.Session()\n\n    batched_train = data.get_batched_data(epoch)\n    iterator_train = batched_train.make_one_shot_iterator()\n    in_data, out_data = iterator_train.get_next()\n\n    total_batch = int(data.size_training \/ batch_size)\n    for batch in range(total_batch):\n        print(f\"{batch + 1} \/ {total_batch}\")\n        in_images = sess_iter_in.run(in_data).reshape((-1, 64, 64, 1))\n        out_images = sess_iter_out.run(out_data).reshape((-1, 64, 64, 1))\n        sess.run(optimizer, feed_dict={inputs: in_images,\n                                       outputs: out_images})\n<\/code><\/pre>\n\n<h2>What do I need ?<\/h2>\n\n<p>I need to have a pipeline that loads only the images of the current batch (otherwise it will not fit in memory) and I want to shuffle the dataset in a different way for each epoch.<\/p>\n\n<h2>Questions and problems<\/h2>\n\n<p>First question, am I using the Dataset class in a good way? I saw very different things on the internet, for example in <a href=\"https:\/\/towardsdatascience.com\/how-to-use-dataset-in-tensorflow-c758ef9e4428\" rel=\"nofollow noreferrer\">this<\/a> blog post the dataset is used with a placeholder and fed during the learning with the datas. It seems strange because the data are all in an array, so loaded in memory. I don't see the point of using <code>tf.data.dataset<\/code> in this case.<\/p>\n\n<p>I found solution by using <code>repeat(epoch)<\/code> on the dataset, like <a href=\"https:\/\/stackoverflow.com\/a\/47217160\/10528024\">this<\/a>, but the shuffle will not be different for each epoch in this case.<\/p>\n\n<p>The second problem with my implementation is that I have an <code>OutOfRangeError<\/code> in some cases. With a small amount of data (512 like in the exemple) it works fine, but with a bigger amount of data, the error occurs. I thought it was because of a bad calculation of the number of batch due to bad rounding, or when the last batch has a smaller amount of data, but it happens in batch 32 out of 115... Is there any way to know the number of batch created after a <code>batch(n)<\/code> call on dataset?<\/p>\n\n<p>Sorry for this loooonng question, but I've been struggling with this for a few days.<\/p>\n","answers":[{"AnswerId":"52895475","CreationDate":"2018-10-19T15:24:12.223","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>As far as I know, <a href=\"https:\/\/www.tensorflow.org\/performance\/datasets_performance\" rel=\"nofollow noreferrer\">Official Performance Guideline<\/a> is the best teaching material to make input pipelines.<\/p>\n\n<blockquote>\n  <p>I want to shuffle the dataset in a different way for each epoch.<\/p>\n<\/blockquote>\n\n<p>Using shuffle() and repeat(), you can get different shuffle pattern for each epochs. You can confirm it with the following code<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices([1,2,3,4])\ndataset = dataset.shuffle(4)\ndataset = dataset.repeat(3)\n\niterator = dataset.make_one_shot_iterator()\nx = iterator.get_next()\n\nwith tf.Session() as sess:\n    for i in range(10):\n        print(sess.run(x))\n<\/code><\/pre>\n\n<p>You can also use tf.contrib.data.shuffle_and_repeat as the mentioned by the above official page.<\/p>\n\n<p>There are some problems in your code outside of creating data pipelines. You confuse graph construction with graph execution. You are repeating to create data input pipeline, so there are many redundant input pipelines as many as epochs. You can observe the redundant pipelines by Tensorboard.<\/p>\n\n<p>You should place your graph construction code outside of loop as the following code (pseudo code)<\/p>\n\n<pre><code>batched_train = data.get_batched_data()\niterator = batched_train.make_initializable_iterator()\nin_data, out_data = iterator_train.get_next()\n\nfor epoch in range(nb_epoch):\n    # reset iterator's state\n    sess.run(iterator.initializer)\n\n    try:\n        while True:\n            in_images = sess.run(in_data).reshape((-1, 64, 64, 1))\n            out_images = sess.run(out_data).reshape((-1, 64, 64, 1))\n            sess.run(optimizer, feed_dict={inputs: in_images,\n                                           outputs: out_images})\n    except tf.errors.OutOfRangeError:\n        pass\n<\/code><\/pre>\n\n<p>Moreover there are some unimportant inefficient code. You loaded a list of file path with from_tensor_slices(), so the list was embedded in your graph. (See <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#consuming_numpy_arrays\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/datasets#consuming_numpy_arrays<\/a> for detail)<\/p>\n\n<p>You would be better off using prefetch, and decreasing sess.run call by combining your graph. <\/p>\n"}]}
+{"QuestionId":52891649,"AnswerCount":2,"Tags":"<python-2.7><cython><caffe2>","CreationDate":"2018-10-19T11:40:47.030","AcceptedAnswerId":null,"OwnerUserId":10396692.0,"Title":"ImportError: ...\/detectron\/utils\/cython_nms.so: undefined symbol: PyFPE_jbuf, Detectron, Caffe2, Anaconda, Cython","Body":"<p>I've got import errors when trying to reuse the <a href=\"https:\/\/github.com\/facebookresearch\/Detectron\" rel=\"nofollow noreferrer\">detectron<\/a><\/p>\n\n<h3>Expected Results:<\/h3>\n\n<p>Successful inference<\/p>\n\n<h3>Actual Results<\/h3>\n\n<pre><code>Traceback (most recent call last):\n\nFile \"infer_simple.py\", line 44, in &lt;module&gt;\n     import detectron.core.test_engine as infer_engine\n File \"...\/detectron\/core\/test_engine.py\", line 35, in &lt;module&gt;\n     from detectron.core.rpn_generator import generate_rpn_on_dataset\n File \"...\/detectron\/core\/rpn_generator.py\", line 42, in &lt;module&gt;\n     from detectron.datasets import task_evaluation\n File \"...\/detectron\/datasets\/task_evaluation.py\", line 47, in &lt;module&gt;\n     import detectron.datasets.json_dataset_evaluator as json_dataset_evaluator\n File \"...\/detectron\/datasets\/json_dataset_evaluator.py\", line 34, in &lt;module&gt;\n     import detectron.utils.boxes as box_utils\n\n File \"...\/detectron\/utils\/boxes.py\", line 52, in &lt;module&gt;\n     import detectron.utils.cython_nms as cython_nms\n     ImportError: ...\/detectron\/utils\/cython_nms.so: undefined symbol: PyFPE_jbuf\n<\/code><\/pre>\n\n<h3>Detailed steps to reproduce<\/h3>\n\n<pre><code>&lt;code&gt;conda install pytorch-nightly -c pytorch&lt;\/code&gt;&lt;br&gt;\nDETECTRON=\/path\/to\/clone\/detectron&lt;br&gt;\n&lt;code&gt;git clone https:\/\/github.com\/facebookresearch\/detectron $DETECTRON&lt;\/code&gt;&lt;br&gt;\n&lt;code&gt;pip install -r $DETECTRON\/requirements.txt&lt;\/code&gt;&lt;br&gt;\n&lt;code&gt;cd $DETECTRON &amp;&amp; make &lt;code&gt;\n&lt;code&gt;python $DETECTRON\/detectron\/tools\/infer_simple.py&lt;\/code&gt;\n<\/code><\/pre>\n\n<h3>System information<br><\/h3>\n\n<pre><code>Operating system: Ubuntu 16.04\nCompiler version: Cmake 3.12.0\nCUDA version: 9.2\ncuDNN version: 9.0\nNVIDIA driver version: 396.37\nGPU model: Nvidia K80\nPYTHONPATH: empty\n\npython --version\n&gt; Python 2.7.15 :: Anaconda, Inc.\n<\/code><\/pre>\n\n<p>Anything else that seems relevant:<br>\n<code>Caffe2<\/code> is working just fine, <code>test_spatial_narrow_as_op.py<\/code> returns 'OK'<\/p>\n\n<p>From what I found on the web, the problem is that I have different versions of cython, numpy or opencv on my different versions of <code>python<\/code> and <code>anaconda<\/code>. But if I remove <code>numpy<\/code> or <code>opencv<\/code> from <code>anaconda<\/code>, then <code>caffe<\/code> doesn't work anymore, and if I remove <code>cython<\/code> from my original <code>python<\/code> packages, then the make command in cd detectron &amp;&amp; make fails.<\/p>\n\n<p>I feel like if I could tell make to use Cython from anaconda and not from the default python it should work.<\/p>\n","answers":[{"AnswerId":"52891700","CreationDate":"2018-10-19T11:44:30.463","ParentId":null,"OwnerUserId":"10396692","Title":null,"Body":"<p>Ok, I found the solution while posting :-p<\/p>\n\n<p>The solution is to modify the MakeFile in <code>\/dectron<\/code> as follows:<\/p>\n\n<pre><code># Don't use the --user flag for setup.py develop mode with virtualenv.\nDEV_USER_FLAG=$(shell python -c \"import sys; print('' if hasattr(sys, 'real_prefix') else '--user')\")\nPYTHON_EXE=\/home\/[username]\/anaconda3\/envs\/[myenv]\/bin\/python2.7\n\n.PHONY: default\ndefault: dev\n\n.PHONY: install\ninstall:\n    $(PYTHON_EXE) setup.py install\n\n.PHONY: ops\nops:\n    mkdir -p build &amp;&amp; cd build &amp;&amp; cmake .. &amp;&amp; make -j$(shell nproc)\n\n.PHONY: dev\ndev:\n    $(PYTHON_EXE) setup.py develop $(DEV_USER_FLAG)\n\n.PHONY: clean\nclean:\n    $(PYTHON_EXE) setup.py develop --uninstall $(DEV_USER_FLAG)\n    rm -rf build\n<\/code><\/pre>\n\n<p>It will allow the make command to use <code>python<\/code> from the desired anaconda env and the corresponding <code>Cython<\/code> installation<\/p>\n"},{"AnswerId":"55181122","CreationDate":"2019-03-15T11:00:47.267","ParentId":null,"OwnerUserId":"11208213","Title":null,"Body":"<p>This is my trackback:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"tools\/tracking\/greedy_tracking_from_raw_dets.py\", line 8, in &lt;module&gt;\n    from vdetlib.vdet.track import greedily_track_from_raw_dets, fcn_tracker\n  File \".\/vdetlib\/vdet\/track.py\", line 13, in &lt;module&gt;\n    from ..utils.cython_nms import track_det_nms\nImportError: .\/vdetlib\/utils\/cython_nms.so: undefined symbol: PyFPE_jbuf\n<\/code><\/pre>\n\n<p>You should:<\/p>\n\n<ol>\n<li><code>pip install cython kivy<\/code><\/li>\n<li>re-make detectron<\/li>\n<\/ol>\n\n<p>Then it can work!<\/p>\n"}]}
+{"QuestionId":52892159,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-19T12:16:05.920","AcceptedAnswerId":null,"OwnerUserId":4397273.0,"Title":"How to train single example in keras MLP","Body":"<p>So I am doing agent based modelling where each agent has MLP kind of network in it but I have limitation agent can have only single example at time hence agent can train only single example <\/p>\n\n<p>so when I tried to train using keras fit function I got chain of errors <\/p>\n\n<pre><code>model = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(128,activation=tf.nn.tanh))\nmodel.add(tf.keras.layers.Dense(128,activation=tf.nn.tanh))\nmodel.add(tf.keras.layers.Dense(1,activation=tf.nn.tanh))\nmodel.compile(optimizer='SGD',\n            loss='mean_squared_error',validation_split=0)\n\nmodel.fit(x_train[0][np.newaxis,:,:],np.array([y_train[0]]),epochs=3,batch_size=1)\n<\/code><\/pre>\n\n<p>I just wanted to train single example at a time \nbut got following error<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-81-688643bc66bc&gt; in &lt;module&gt;()\n----&gt; 1 model.fit(x_train[0][np.newaxis,:,:],np.array([y_train[0]]),epochs=3,batch_size=1)\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1261         steps_name='steps_per_epoch',\n   1262         steps=steps_per_epoch,\n-&gt; 1263         validation_split=validation_split)\n   1264 \n   1265     # Prepare validation data.\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split)\n    905           feed_output_shapes,\n    906           check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 907           exception_prefix='target')\n    908 \n    909       # Generate sample-wise weight values given the `sample_weight` and\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    189                 'Error when checking ' + exception_prefix + ': expected ' +\n    190                 names[i] + ' to have shape ' + str(shape) +\n--&gt; 191                 ' but got array with shape ' + str(data_shape))\n    192   return data\n    193 \n\nValueError: Error when checking target: expected dense_5 to have shape (10,) but got array with shape (1,)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52892576,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-19T12:42:49.357","AcceptedAnswerId":null,"OwnerUserId":2666182.0,"Title":"TF Eager Pretrained Models","Body":"<p>With <code>tfhub<\/code> unable to support Eager execution and most (if not all) released trained models working with the default execution mode in TensorFlow, are there any sources \/ repositories of pretrained models like <code>VGGNet<\/code>, <code>InceptionV3<\/code> .etc that can be used with TF-Eager?<\/p>\n","answers":[]}
+{"QuestionId":52893369,"AnswerCount":0,"Tags":"<python><tensorflow><artificial-intelligence><genetic-algorithm><neat>","CreationDate":"2018-10-19T13:29:19.593","AcceptedAnswerId":null,"OwnerUserId":10528459.0,"Title":"Evolving Neural Network using NEAT python on Forex data","Body":"<p>I am trying to implement an evolving neural network on time series Forex data where the model will receive as inputs 3 different exchange rates on a particular timeframe and the base currency will be the same in all 3 inputs (e.g. USD\/CHF, USD\/JPY and USDZAR all have the same base currency namely USD when used as inputs). The network would then have to predict the direction of the base currency of the next weeks or days (based on timeframe used in data) using based on the inputs. It has the following options for prediction to choose from:<\/p>\n\n<ul>\n<li>Sell<\/li>\n<li>Strong Sell<\/li>\n<li>Buy<\/li>\n<li>Strong Buy<\/li>\n<\/ul>\n\n<p>I have tried implementing this in the NEAT python framework based on the simplest example at <a href=\"https:\/\/neat-python.readthedocs.io\/en\/latest\/xor_example.html\" rel=\"nofollow noreferrer\">https:\/\/neat-python.readthedocs.io\/en\/latest\/xor_example.html<\/a> using a very similar config file, representing \"Sell\" as \"0\", \"Strong Sell\" as \"1\", \"Buy\" as \"2\" and \"Strong Buy\" as \"3\" in the code. However I haven't been successful since the maximum fitness level I could get on the training set of 84 observations about 75%, this is after I decide to change the NEAT to predict only \"Sell\" and \"Buy\" represented by \"0\" and \"1\" respectively.\nI have tried implementing this using a recurrent as well as feedforward type network, not much improvement took place. I have also tried modifying the mutation rates. Kindly assist me with increasing the models fitness level as well as prediction accuracy on the test set. See code below and Thank you.<\/p>\n\n<p>The data used below is data used here runs over <strong>3 years<\/strong>.<\/p>\n\n<pre><code>from __future__ import print_function\nimport neat\nimport visualize\nimport random\n\n# 3-input, inputs and expected outputs(training set).\nxor_inputs = [(1.1241,1.2976,1.02923013585838), \n(1.1201,1.2434,1.0231225700839),(1.0971,1.2187,1.00999899000101), \n(1.0884,1.2235,1.00644122383253),(1.0987,1.2185,1.01142914938809), \n(1.1138,1.2519,1.03241792277514),(1.0856,1.2602,1.01163378856854),             \n(1.0588,1.2346,0.99009900990099),(1.0587,1.2476,0.986679822397632), \n(1.0673,1.2729,0.988630746416214),(1.0562,1.2576,0.982318271119843), \n(1.0451,1.2497,0.974563882662508),(1.0456,1.2292,0.973994350832765), \n(1.0516,1.2338,0.981932443047918),(1.0533,1.2287,0.982028871648826), \n(1.0644,1.2179,0.990982063224656),(1.0703,1.2375,0.998203234178479), \n(1.0699,1.2555,1.000100010001),(1.0784,1.2488,1.00745516824501), \n(1.0641,1.2491,0.997506234413965),(1.0614,1.2414,0.997207818109294), \n(1.0563,1.2468,0.992457324335054),(1.0623,1.2295,0.992457324335053), \n(1.0672,1.2171,0.989119683481701),(1.0738,1.2394,1.00130169219986), \n(1.0799,1.247,1.00867460157353),(1.0652,1.255,0.996909580301067), \n(1.059,1.2373,0.991080277502478),(1.0612,1.2524,0.994629003381739), \n(1.0728,1.2813,1.00411687920474),(1.0897,1.2951,1.0048231511254), \n(1.0998,1.2981,1.01317122593718),(1.0931,1.2891,0.99930048965724), \n(1.1207,1.3035,1.02838338132456),(1.1183,1.2804,1.02616726526424), \n(1.1282,1.2887,1.0391769718383),(1.1196,1.2745,1.03135313531353), \n(1.1198,1.2776,1.02690490860546),(1.1194,1.2718,1.03145951521403), \n(1.1426,1.3027,1.04351455702807),(1.1401,1.2892,1.03734439834025), \n(1.1469,1.3096,1.03788271925272),(1.1663,1.2995,1.05741778576716), \n(1.1752,1.3135,1.03220478943022),(1.1772,1.3036,1.02785486689279), \n(1.1821,1.3012,1.03960910697578),(1.1762,1.2876,1.03626943005181), \n(1.1926,1.2887,1.04515050167224),(1.1859,1.2951,1.03659168653467), \n(1.2035,1.3198,1.05876124933827),(1.1943,1.3593,1.0417751849151), \n(1.1954,1.3493,1.03177878662815),(1.1814,1.3397,1.03284445362528), \n(1.1734,1.3066,1.02249488752556),(1.1823,1.3287,1.02616726526424), \n(1.1785,1.319,1.01595042161943),(1.161,1.3129,1.00230530219505), \n(1.1609,1.3076,0.99930048965724),(1.1665,1.3191,1.00421771440048), \n(1.1795,1.3214,1.01153145862836),(1.193,1.3337,1.02061645233721), \n(1.1891,1.3473,1.02396067990989),(1.1764,1.3392,1.00704934541793), \n(1.1754,1.3321,1.00959111559818),(1.1859,1.3362,1.01081572829273), \n(1.1998,1.3515,1.02627257799672),(1.203,1.3571,1.02543068088597), \n(1.2187,1.3729,1.03284445362528),(1.2222,1.3852,1.03842159916926), \n(1.2421,1.417,1.07100781835707),(1.2462,1.4123,1.07376785139053), \n(1.2235,1.3838,1.06371662589086),(1.2406,1.4042,1.07781849536538), \n(1.2293,1.3967,1.06746370623399),(1.2317,1.3805,1.06587081645705), \n(1.2307,1.3849,1.05130361648444),(1.2289,1.3941,1.0501995379122), \n(1.2353,1.4134,1.05574324324324),(1.2323,1.4018,1.0481081647626), \n(1.2283,1.4092,1.04242676951944),(1.2331,1.4241,1.03906899418121), \n(1.2288,1.4004,1.02606197414324),(1.213,1.378,1.01224820325944), \n(1.196,1.3533,1)]\nxor_outputs = [(3,),(3,),(2,),(0,),(1,),(2,),(3,),(2,),(1,),(3,),(3,), \n(2,),(1,),(0,),(0,),(1,),(0,),(0,),(2,),(3,),(2,),(2,),(2,),(1,),(1,), \n(2,),(3,),(1,),(1,),(1,),(1,),(3,),(1,),(3,),(1,),(3,),(0,),(2,),(1,), \n(3,),(1,),(0,),(0,),(2,),(0,),(3,),(1,),(2,),(1,),(2,),(2,),(2,),(3,), \n(1,),(3,),(3,),(3,),(1,),(1,),(1,),(0,),(3,),(2,),(1,),(1,),(0,),(1,), \n(1,),(1,),(0,),(3,),(1,),(3,),(2,),(2,),(2,),(1,),(3,),(2,),(0,),(3,), \n(3,),(3,),(2,)\n]\n# 3-input, inputs and expected outputs(test set).\nxor_inputs2 = [(1.1944,1.3543,0.99940035978413), \n(1.1778,1.3474,1.00220485067148),(1.1652,1.3309,1.01030511214387), \n(1.1661,1.3348,1.0120433154539),(1.1768,1.3412,1.01502233049127), \n(1.1609,1.3285,1.00240577385726),(1.1657,1.327,1.01214574898785), \n(1.1685,1.3209,1.00938730190774),(1.1747,1.3285,1.01020305081321), \n(1.1685,1.3235,0.998302885095338),(1.172,1.3134,1.00745516824501), \n(1.1658,1.3104,1.00553041729512),(1.1567,1.3008,1.00573267625465), \n(1.1411,1.2769,1.00462125778582),(1.1439,1.2752,1.00411687920474), \n(1.1623,1.2845,1.01698362656361),(1.1601,1.2963,1.03220478943022), \n(1.1553,1.2924,1.03167234086454),(1.163,1.3068,1.03359173126615), \n(1.175,1.3077,1.04307916970898),(1.1782,1.3159,1.03626943005181)\n]\nxor_outputs2 = [(2,),(2,),(1,),(1,),(3,),(0,),(2,),(1,),(3,),(0,), \n(3,),(2,),(3,),(2,),(1,),(0,),(3,),(1,),(1,),(0,),(3,)\n]\n\n\ndef eval_genomes(genomes, config): #function Used for training model \nusing the training set\n   for genome_id, genome in genomes:\n       genome.fitness = 84.0\n       net = neat.nn.RecurrentNetwork.create(genome, config)\n       for xi, xo in zip(xor_inputs, xor_outputs):\n           output = net.activate(xi)\n           genome.fitness -= (output[0] - xo[0]) ** 2 #Distance from \nthe correct output summed for all 84 inputs patterns\n\n\n# Load configuration.\nconfig = neat.Config(neat.DefaultGenome, neat.DefaultReproduction,\n                 neat.DefaultSpeciesSet, neat.DefaultStagnation,\n                 'config-feedforward-research_nn')\n\n# Create the population, which is the top-level object for a NEAT run.\np = neat.Population(config)\n\n# Add a stdout reporter to show progress in the terminal.\np.add_reporter(neat.StdOutReporter(True))\nstats = neat.StatisticsReporter()\np.add_reporter(stats)\n#p.add_reporter(neat.Checkpointer(100))\n\n# Run until a solution is found.\nwinner = p.run(eval_genomes, 30000)  # run for 30000 to test \ngenerations\n\n# Display the winning genome.\nprint('\\nBest genome:\\n{!s}'.format(winner))\n\n# Make and show prediction on unseen data (test set) using winner NN's \ngenome.\nprint('\\nOutput:')\nwinner_net = neat.nn.RecurrentNetwork.create(winner, config)\nfor xi, xo in zip(xor_inputs2, xor_outputs2):\n  output = winner_net.activate(xi)\n  print(\"  input {!r}, expected output {!r}, got {!r}\".format(\n    xi, xo, output))\n\nnode_names = {-1: 'Input 1', -2: 'Input 2', -3: 'Input 3', 0: \n'Prediction'}\nvisualize.draw_net(config, winner, True, node_names=node_names,\n               filename=\"winner-feedforward-research_nn.gv\")\n<\/code><\/pre>\n\n<p>The config file:<\/p>\n\n<pre><code>[NEAT]\n# changed\nfitness_criterion     = max\nfitness_threshold     = 75 \npop_size              = 500 \nreset_on_extinction   = False\n\n[DefaultGenome]\n# node activation options \n# changed\nactivation_default      = sigmoid \nactivation_mutate_rate  = 0.0\nactivation_options      = sigmoid, square \n\n# node aggregation options\naggregation_default     = sum\naggregation_mutate_rate = 0.0\naggregation_options     = sum\n\n# node bias options\nbias_init_mean          = 0.0\nbias_init_stdev         = 1.0\nbias_max_value          = 30.0\nbias_min_value          = -30.0\nbias_mutate_power       = 0.5\nbias_mutate_rate        = 0.7\nbias_replace_rate       = 0.1\n\n# genome compatibility options\ncompatibility_disjoint_coefficient = 1.0\ncompatibility_weight_coefficient   = 0.5\n\n# connection add\/remove rates\nconn_add_prob           = 0.5\nconn_delete_prob        = 0.5\n\n# connection enable options\n# changed\nenabled_default         = True\nenabled_mutate_rate     = 0.04 \n\nfeed_forward            = false\ninitial_connection      = full_direct\n\n# node add\/remove rates\n# changed\nnode_add_prob           = 0.5 \nnode_delete_prob        = 0.5 \n\n# network parameters\n# changed\nnum_hidden              = 0\nnum_inputs              = 3 \nnum_outputs             = 1\n\n# node response options\nresponse_init_mean      = 1.0\nresponse_init_stdev     = 0.0\nresponse_max_value      = 30.0\nresponse_min_value      = -30.0\nresponse_mutate_power   = 0.0\nresponse_mutate_rate    = 0.0\nresponse_replace_rate   = 0.0\n\n# connection weight options\n# changed\nweight_init_mean        = 0.0\nweight_init_stdev       = 1.0\nweight_max_value        = 30\nweight_min_value        = -30\nweight_mutate_power     = 0.5\nweight_mutate_rate      = 0.9\nweight_replace_rate     = 0.1\n\n[DefaultSpeciesSet]\ncompatibility_threshold = 3.0\n\n[DefaultStagnation]\nspecies_fitness_func = max\nmax_stagnation       = 20\nspecies_elitism      = 2\n\n[DefaultReproduction]\nelitism            = 2\nsurvival_threshold = 0.2\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52893684,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-19T13:47:50.140","AcceptedAnswerId":52894897.0,"OwnerUserId":913098.0,"Title":"Unexpected keyword argument passed to optimizer: amsgrad","Body":"<p>When trying to follow the <a href=\"https:\/\/keras.io\/optimizers\/\" rel=\"nofollow noreferrer\">Keras doc<\/a> on Adam, I copy this line from the doc:<\/p>\n\n<pre><code>keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\n<\/code><\/pre>\n\n<p>and get this error<\/p>\n\n<blockquote>\n  <p>Unexpected keyword argument passed to optimizer: amsgrad<\/p>\n<\/blockquote>\n\n<hr>\n\n<p><strong>EDIT 1<\/strong><\/p>\n\n<p>Omitting the <code>amsgrad<\/code> parameter agrees to interprt the line <\/p>\n\n<p><code>keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0)<\/code><\/p>\n\n<p>but then when trying to train the model using <\/p>\n\n<pre><code>happyModel.fit(x = X_train, y = Y_train, epochs = 50, batch_size = 600)\n<\/code><\/pre>\n\n<p>gives the following error:<\/p>\n\n<blockquote>\n  <p>None values not supported.<\/p>\n<\/blockquote>\n\n<p>Full error:<\/p>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- ValueError                                Traceback (most recent call\n  last)  in ()\n        1 ### START CODE HERE ### (1 line)\n  ----> 2 happyModel.fit(x = X_train, y = Y_train, epochs = 50, batch_size = 100)\n        3 ### END CODE HERE ###<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in\n  fit(self, x, y, batch_size, epochs, verbose, callbacks,\n  validation_split, validation_data, shuffle, class_weight,\n  sample_weight, initial_epoch, steps_per_epoch, validation_steps,\n  **kwargs)    1574         else:    1575             ins = x + y + sample_weights\n  -> 1576         self._make_train_function()    1577         f = self.train_function    1578 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in\n  _make_train_function(self)\n      958                     training_updates = self.optimizer.get_updates(\n      959                         params=self._collected_trainable_weights,\n  --> 960                         loss=self.total_loss)\n      961                 updates = self.updates + training_updates\n      962                 # Gets loss and metrics. Updates weights at each call.<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py in\n  wrapper(*args, **kwargs)\n       85                 warnings.warn('Update your <code>' + object_name +\n       86                               '<\/code> call to the Keras 2 API: ' + signature, stacklevel=2)\n  ---> 87             return func(*args, **kwargs)\n       88         wrapper._original_function = func\n       89         return wrapper<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/optimizers.py in\n  get_updates(self, loss, params)\n      432             m_t = (self.beta_1 * m) + (1. - self.beta_1) * g\n      433             v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)\n  --> 434             p_t = p - lr_t * m_t \/ (K.sqrt(v_t) + self.epsilon)\n      435 \n      436             self.updates.append(K.update(m, m_t))<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py\n  in binary_op_wrapper(x, y)\n      827       if not isinstance(y, sparse_tensor.SparseTensor):\n      828         try:\n  --> 829           y = ops.convert_to_tensor(y, dtype=x.dtype.base_dtype, name=\"y\")\n      830         except TypeError:\n      831           # If the RHS is not a tensor, it might be a tensor aware object<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\n  in convert_to_tensor(value, dtype, name, preferred_dtype)\n      674       name=name,\n      675       preferred_dtype=preferred_dtype,\n  --> 676       as_ref=False)\n      677 \n      678 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\n  in internal_convert_to_tensor(value, dtype, name, as_ref,\n  preferred_dtype)\n      739 \n      740         if ret is None:\n  --> 741           ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n      742 \n      743         if ret is NotImplemented:<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/constant_op.py\n  in _constant_tensor_conversion_function(v, dtype, name, as_ref)\n      111                                          as_ref=False):\n      112   _ = as_ref\n  --> 113   return constant(v, dtype=dtype, name=name)\n      114 \n      115 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/constant_op.py\n  in constant(value, dtype, shape, name, verify_shape)\n      100   tensor_value = attr_value_pb2.AttrValue()\n      101   tensor_value.tensor.CopyFrom(\n  --> 102       tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape, verify_shape=verify_shape))\n      103   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)\n      104   const_tensor = g.create_op(<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_util.py\n  in make_tensor_proto(values, dtype, shape, verify_shape)\n      362   else:\n      363     if values is None:\n  --> 364       raise ValueError(\"None values not supported.\")\n      365     # if dtype is provided, forces numpy array to be the type\n      366     # provided if possible.<\/p>\n  \n  <p>ValueError: None values not supported.<\/p>\n<\/blockquote>\n\n<p>Thus simply omitting the parameter doesn't do the trick<\/p>\n\n<p>How to get the adam optimizer to work?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"52894897","CreationDate":"2018-10-19T14:52:49.123","ParentId":null,"OwnerUserId":"913098","Title":null,"Body":"<ol>\n<li>This is probably due to an old version of keras which does not support the <code>amsgrad<\/code> parameter<\/li>\n<li>removing the parameter allows the interpreter to understand the line.<\/li>\n<li>The <code>None values not supported.<\/code> problem comes from None in the <code>epsilon<\/code> parameter. need to specify a value.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":52894349,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-10-19T14:23:57.220","AcceptedAnswerId":null,"OwnerUserId":5577396.0,"Title":"Tensorflow - Disaccording TF METRICS (Accuracy, Precision, Recall)","Body":"<p>I am new with Tensorflow and I am working with a CNN for an instrument recognition problem, using <strong>IRMAS<\/strong> dataset. <\/p>\n\n<p>The <strong>IRMAS<\/strong> dataset has a clear <strong>training<\/strong> set, based only on singular instruments, therefore each instrument (guitar, violin, etc) has its own directory with multiple samples (<em>multi-classification<\/em>). The <strong>testing<\/strong> set, differently, contains polyphonic music, therefore multiple classes of instruments may appear in the same song (<em>multi-classification<\/em>, <em>multi-label<\/em>).<\/p>\n\n<p>This is the final part of my network with the prediction and loss calculation, I am using the <code>Estimator API<\/code>.<\/p>\n\n<pre><code>### previous part of network is omitted\n\n# logits layer\nlogits = tf.layers.dense(inputs=dropout, units=labels.shape[1])\n\nprint('Shape Logits:', logits.shape)\n\n\npredictions = {\n    # \"classes\": tf.equal(tf.argmax(tf.cast(logits, tf.float32), 1), tf.argmax(tf.cast(labels, tf.float32), 1)),\n    \"classes\": tf.argmax(input=logits, axis=1),\n    \"probabilities\": tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.cast(labels, tf.float32), logits=tf.cast(logits, tf.float32), name=\"sigmoid_tensor\")\n}\n\n# correct_prediction = tf.equal(tf.argmax(tf.cast(logits, tf.float32), 1), tf.argmax(tf.cast(labels, tf.float32), 1))\n\nif mode == tf.estimator.ModeKeys.PREDICT:\n    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)\n\nloss = tf.losses.sigmoid_cross_entropy(multi_class_labels=labels, logits=logits)\n\nif mode == tf.estimator.ModeKeys.TRAIN:\n    optimizer = tf.train.AdamOptimizer(learning_rate=0.001)\n    #optimizer = tf.train.MomentumOptimizer(learning_rate=0.01, momentum=0.96)\n    train_op = optimizer.minimize(\n                              loss=loss,\n                              global_step=tf.train.get_global_step())\n    logging_hook = tf.train.LoggingTensorHook({\"loss\" : loss}, every_n_iter=10)\n    return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op, training_hooks = [logging_hook])\n\n\neval_metric_ops = {\n    \"accuracy\": tf.metrics.accuracy(\n    labels=tf.argmax(input=labels, axis=1),\n    predictions=predictions[\"classes\"]),\n\n    \"recall\": tf.metrics.recall(\n    labels=tf.argmax(input=labels, axis=1),\n    predictions=predictions[\"classes\"]),\n\n    \"precision\": tf.metrics.precision(\n    labels=tf.argmax(input=labels, axis=1),\n    predictions=predictions[\"classes\"])\n}\n<\/code><\/pre>\n\n<p>After few epochs, the training process seems to work since the loss goes down quite smoothly. However, the tensorflow metrics needed to evaluate the model does not seem to agree.<\/p>\n\n<p>These are my results on the evaluation set (15% of the training set):<\/p>\n\n<pre><code>{'recall': 0.9989496, 'accuracy': 0.40656063, 'global_step': 1428, 'precision': 0.95004994, 'loss': 0.2432195}\n<\/code><\/pre>\n\n<p>These are the results obtained from the testing set:<\/p>\n\n<pre><code>{'recall': 0.9981618, 'accuracy': 0.26530612, 'global_step': 1428, 'precision': 0.96533334, 'loss': 0.46097097}\n<\/code><\/pre>\n\n<p>In my opinion, the accuracy value seems to be the only right, since decreases in the testing set, which is composed by polyphonic music rather than singular instruments. I don't understand why I get these results for precision and recall, they should be lower compared to the accuracy I get. <\/p>\n\n<p>I've tried to display the FP, FN, TN and TP and they seem to be in accord with precision and recall. In fact, by computing the accuracy with them I got around 0.97 (too high for this classification problem).<\/p>\n\n<p>I have not idea of what I am doing wrong, really appreciate some help.<\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[]}
+{"QuestionId":52895986,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><normalization>","CreationDate":"2018-10-19T15:55:50.840","AcceptedAnswerId":null,"OwnerUserId":2621303.0,"Title":"Why normalize when all features are on the same scale?","Body":"<p>So I'm doing the tensorflow tutorial found here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a><\/p>\n\n<p>Basically, my input is a [28x28] matrix (image) that I flatten to a [1x784] vector.<\/p>\n\n<p>The tutorial then says:<\/p>\n\n<blockquote>\n  <p>We scale these values to a range of 0 to 1 before feeding to the neural network model. For this, cast the datatype of the image components from an integer to a float, and divide by 255.<\/p>\n<\/blockquote>\n\n<p>My question is why do we need to normalize in this case? My understanding is that when we have features that are on different scales, we need normalization if not the output of the model is distorted. But in this case all pixel ranges go from 0 to 255 (all features are the same scale)<\/p>\n\n<p>I went ahead and ran it with normalization, and get an accuracy of over 85%, whereas no normalization, my accuracy falls to 10%.<\/p>\n\n<p>Any ideas?<\/p>\n","answers":[]}
+{"QuestionId":52897168,"AnswerCount":3,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-19T17:20:02.007","AcceptedAnswerId":52904470.0,"OwnerUserId":3821472.0,"Title":"Tensorflow, how to concatenate multiple datasets with varying batch sizes","Body":"<p>Imagine I have:<\/p>\n\n<ul>\n<li>dataset 1 with data [5, 5, 5, 5, 5]<\/li>\n<li>dataset 2 with data [4, 4]<\/li>\n<\/ul>\n\n<p>I want to have take batches from both datasets and concatenate them so that I get batches of size 3 where:<\/p>\n\n<ul>\n<li>I read dataset 1 with batch size 2<\/li>\n<li>I read dataset 2 with batch size 1.<\/li>\n<\/ul>\n\n<p>I also want to read the final batch if some datasets get emptied first.\nIn this instance, I would get [5, 5, 4], [5, 5, 4], [5] as my final result.<\/p>\n\n<p>How can I do this?\nI've seen the answer here: <a href=\"https:\/\/stackoverflow.com\/questions\/48272035\/tensorflow-how-to-generate-unbalanced-combined-data-sets\">Tensorflow how to generate unbalanced combined data sets<\/a><\/p>\n\n<p>It is a good try, but it does not work if one of the datasets gets emptied before the others (because then <code>tf.errors.OutOfRangeError<\/code> gets outputted pre-emptively when you try to fetch elements from the dataset that gets emptied first and I do not get the final batch). Therefore I only get [5, 5, 4], [5, 5, 4]<\/p>\n\n<p>I thought of using <code>tf.contrib.data.choose_from_datasets<\/code>:<\/p>\n\n<pre><code>ds1 = tf.data.Dataset.from_tensor_slices([5, 5, 5, 5, 5]).batch(2)\nds2 = tf.data.Dataset.from_tensor_slices([4, 4, 4, 4]).batch(1)\nchoice_dataset = [1, 2, 1, 2, 1]\nds = tf.contrib.data.choose_from_datasets([ds1, ds2], choice_dataset)\nds = ds.apply(tf.contrib.data.unbatch())\nds = ds.batch(3, drop_remainder=False)\n<\/code><\/pre>\n\n<p>This kind of works but is rather inelegant (there is unbatch and batch); also, I don't really have a great control over exactly what goes into a batch. (for instance if ds1 was [7] * 7 with batch size 2, and ds2 was [2, 2] with batch size 1, I would get [7, 7, 1], [7, 7, 1], [7, 7, 7]. But what if I actually want to have [7, 7, 1], [7, 7, 1], [7, 7], [7]? i.e. keep the number of elements from each dataset fixed.<\/p>\n\n<p>Is there another better solution? <\/p>\n\n<p>Another idea I had was to try to use <code>tf.data.Dataset.flat_map<\/code>:<\/p>\n\n<pre><code>ds1 = tf.data.Dataset.from_tensor_slices([5, 5, 5, 5, 5])\nds2 = tf.data.Dataset.from_tensor_slices([4, 4, 4, 4])\nbatch_sizes = [2, 1]\ndef concat(*inputs):\n  concat = partial(functools.reduce, lambda x, y: x.concatenate(y))\n  datasets = [tf.data.Dataset.from_tensors(input) for input in inputs]\n  datasets = [dataset.batch(batch_size) for batch_size, dataset in zip(batch_sizes, datasets)]\n  return concat(datasets)\ndataset = (tf.data.Dataset\n           .zip((ds1, ds2))\n           .flat_map(_concat_and_batch)\n           .batch(sum(batch_sizes)))\n<\/code><\/pre>\n\n<p>but it does not seem to work..<\/p>\n","answers":[{"AnswerId":"52938239","CreationDate":"2018-10-22T21:51:24.780","ParentId":null,"OwnerUserId":"5737630","Title":null,"Body":"<p>Here is a solution that requires you to use a \"control input\", to choose which batch to use, and you decide on this according to which dataset was consumed first. This can be detected using the thrown exception.<\/p>\n\n<p>To explain this solution, I will first present an attempt that does not work.<\/p>\n\n<h3>Attempted solution #1<\/h3>\n\n<pre><code>import tensorflow as tf\n\nds1 = tf.data.Dataset.from_tensor_slices([5,5,5,5,5])\nds2 = tf.data.Dataset.from_tensor_slices([4,4])\n\nds1 = ds1.batch(2)\nds2 = ds2.batch(1)\n\niter1 = ds1.make_one_shot_iterator()\niter2 = ds2.make_one_shot_iterator()\n\nbatch1 = iter1.get_next(name='batch1')\nbatch2 = iter2.get_next(name='batch2')\nbatch12 = tf.concat((batch1, batch2), 0)\n\n# this is a \"control\" placeholder. Its value determines whether to use `batch1`,`batch2` or `batch12`\nwhich_batch = tf.placeholder(tf.int32)\n\nbatch = tf.cond(\n               tf.equal(which_batch,0), # if `which_batch`==0, use `batch12`\n                       lambda:batch12,\n        lambda:tf.cond(tf.equal(which_batch,1), # elif `which_batch`==1, use `batch1`\n                       lambda:batch1,\n        lambda:batch2)) # else, use `batch2`\n\nsess = tf.Session()\n\nwhich = 0 # this value will be fed into the control placeholder `which_batch`\nwhile True:\n    try:\n        print(sess.run(batch,feed_dict={which_batch:which}))\n    except tf.errors.OutOfRangeError as e:\n        # use the error to detect which dataset was consumed, and update `which` accordingly\n        if which==0:\n            if 'batch2' in e.op.name:\n                which = 1\n            else:\n                which = 2\n        else:\n            break\n<\/code><\/pre>\n\n<p>This solution does not work, since for any value of <code>which_batch<\/code>, the <code>tf.cond()<\/code> command will evaluate all the predecessors of its branches (see <a href=\"https:\/\/stackoverflow.com\/a\/37064128\/5737630\">this answer<\/a>) . Therefore, even when <code>which_batch<\/code> has the value 1, <code>batch2<\/code> will be evaluated and an <code>OutOfRangeError<\/code> will be thrown.<\/p>\n\n<h3>Attempted solution #2<\/h3>\n\n<p>This problem can be fixed by moving the definitions of <code>batch1<\/code>, <code>batch2<\/code> and <code>batch12<\/code> into functions.<\/p>\n\n<pre><code>import tensorflow as tf\n\nds1 = tf.data.Dataset.from_tensor_slices([5,5,5,5,5])\nds2 = tf.data.Dataset.from_tensor_slices([4,4])\n\nds1 = ds1.batch(2)\nds2 = ds2.batch(1)\n\niter1 = ds1.make_one_shot_iterator()\niter2 = ds2.make_one_shot_iterator()\n\ndef get_batch1():\n    batch1 = iter1.get_next(name='batch1')\n    return batch1\n\ndef get_batch2():\n    batch2 = iter2.get_next(name='batch2')\n    return batch2\n\ndef get_batch12():\n    batch1 = iter1.get_next(name='batch1_')\n    batch2 = iter2.get_next(name='batch2_')\n    batch12 = tf.concat((batch1, batch2), 0)\n    return batch12\n\n# this is a \"control\" placeholder. It's value determines whether to ues `batch1`,`batch2` or `batch12`\nwhich_batch = tf.placeholder(tf.int32)\n\nbatch = tf.cond(\n               tf.equal(which_batch,0), # if `which_batch`==0, use `batch12`\n                       get_batch12,\n        lambda:tf.cond(tf.equal(which_batch,1), # elif `which_batch`==1, use `batch1`\n                       get_batch1,\n        get_batch2)) # elif `which_batch`==2, use `batch2`\n\nsess = tf.Session()\n\nwhich = 0 # this value will be fed into the control placeholder `which_batch`\nwhile True:\n    try:\n        print(sess.run(batch,feed_dict={which_batch:which}))\n    except tf.errors.OutOfRangeError as e:\n        # use the error to detect which dataset was consumed, and update `which` accordingly\n        if which==0:\n            if 'batch2' in e.op.name:\n                which = 1\n            else:\n                which = 2\n        else:\n            break\n<\/code><\/pre>\n\n<p>However, this doesn't work either. The reason is that at the step when <code>batch12<\/code> was formed and the dataset <code>ds2<\/code> was consumed, then we took the batch from dataset <code>ds1<\/code> and \"dropped\" it without using it.<\/p>\n\n<h3>Solution<\/h3>\n\n<p>We need a mechanism to make sure that we don't \"drop\" any batch in the event that the other dataset is consumed. We can do this by defining a variable which will be assigned the current batch of <code>ds1<\/code>, but <em>only immediately before trying to obtain<\/em> <code>batch12<\/code>. Otherwise, this variable will persist its previous value. Then, if <code>batch12<\/code> fails due to <code>ds1<\/code> being consumed, then this assignment will fail and <code>batch2<\/code> was not dropped and we can use it next time. Otherwise, if <code>batch12<\/code> fails due to <code>ds2<\/code> being consumed, then we have a backup of <code>batch1<\/code> in the variable that we have defined, and after using this backup we can proceed with taking <code>batch1<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\n\nds1 = tf.data.Dataset.from_tensor_slices([5,5,5,5,5])\nds2 = tf.data.Dataset.from_tensor_slices([4,4])\n\nds1 = ds1.batch(2)\nds2 = ds2.batch(1)\n\niter1 = ds1.make_one_shot_iterator()\niter2 = ds2.make_one_shot_iterator()\n\n# this variable will store a backup of `batch1`, in case it is dropped\nbatch1_backup = tf.Variable(0, trainable=False, validate_shape=False)\n\ndef get_batch12():\n    batch1 = iter1.get_next(name='batch1')\n\n    # form the combined batch `batch12` only after backing-up `batch1`\n    with tf.control_dependencies([tf.assign(batch1_backup, batch1, validate_shape=False)]):\n        batch2 = iter2.get_next(name='batch2')\n        batch12 = tf.concat((batch1, batch2), 0)\n    return batch12\n\ndef get_batch1():\n    batch1 = iter1.get_next()\n    return batch1\n\ndef get_batch2():\n    batch2 = iter2.get_next()\n    return batch2\n\n# this is a \"control\" placeholder. Its value determines whether to use `batch12`, `batch1_backup`, `batch1`, or `batch2`\nwhich_batch = tf.Variable(0,trainable=False)\n\nbatch = tf.cond(\n               tf.equal(which_batch,0), # if `which_batch`==0, use `batch12`\n                       get_batch12,\n        lambda:tf.cond(tf.equal(which_batch,1), # elif `which_batch`==1, use `batch1_backup`\n                       lambda:batch1_backup,\n        lambda:tf.cond(tf.equal(which_batch,2), # elif `which_batch`==2, use `batch1`\n                       get_batch1,\n       get_batch2))) # else, use `batch2`\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nwhich = 0  # this value will be fed into the control placeholder\nwhile True:\n    try:\n        print(sess.run(batch,feed_dict={which_batch:which}))\n\n        # if just used `batch1_backup`, proceed with `batch1`\n        if which==1:\n            which = 2\n    except tf.errors.OutOfRangeError as e:\n        # use the error to detect which dataset was consumed, and update `which` accordingly\n        if which == 0:\n            if 'batch2' in e.op.name:\n                which = 1\n            else:\n                which = 3\n        else:\n            break\n<\/code><\/pre>\n"},{"AnswerId":"52901018","CreationDate":"2018-10-19T23:19:35.933","ParentId":null,"OwnerUserId":"5737630","Title":null,"Body":"<p>Here is a solution. It has some problems, but I hope it satisfies your needs.<\/p>\n\n<p>The idea is as follows: You batch each of the two datasets, zip them together, and do a map function to combine each zipped tuple into one batch (so far, this is similar to what is suggested in <a href=\"https:\/\/stackoverflow.com\/a\/47344405\/5737630\">this<\/a> and <a href=\"https:\/\/stackoverflow.com\/a\/49621448\/5737630\">this<\/a> answers.)<\/p>\n\n<p>The problem, as you noticed, is that the zipping only works well for two datasets that are of the same length. Otherwise, one dataset is consumed before the other, and the remaining unconsumed elements are not used.<\/p>\n\n<p>My (kind of hacky) solution to this is to concatenate to both of the datasets another infinite dummy dataset. This dummy dataset consists only of values which you know are not going to appear in your real dataset. This eliminates the problem with zipping. However, you need to get rid of all the dummy elements. This can be easily be done by filtering and a mapping.<\/p>\n\n<pre><code>import tensorflow as tf\n\nds1 = tf.data.Dataset.from_tensor_slices([5,5,5,5,5])\nds2 = tf.data.Dataset.from_tensor_slices([4,4])\n\n# we assume that this value will never occur in `ds1` and `ds2`:\nUNUSED_VALUE = -1 \n\n# an infinite dummy dataset:\ndummy_ds = tf.data.Dataset.from_tensors(UNUSED_VALUE).repeat() \n\n# make `ds1` and `ds2` infinite:\nds1 = ds1.concatenate(dummy_ds)\nds2 = ds2.concatenate(dummy_ds)\n\nds1 = ds1.batch(2)\nds2 = ds2.batch(1)\n\n# this is the solution mentioned in the links above\nds = tf.data.Dataset.zip((ds1,ds2))\nds = ds.map(lambda x1, x2: tf.concat((x1,x2),0))\n\n# filter the infinite dummy tail:\nds = ds.filter(lambda x: tf.reduce_any(tf.not_equal(x,UNUSED_VALUE)))\n\n# filter from batches the dummy elements:\nds = ds.map(lambda x: tf.boolean_mask(x,tf.not_equal(x,UNUSED_VALUE)))\n<\/code><\/pre>\n\n<p>There are two major problems with this solution:<\/p>\n\n<p><strong>(1)<\/strong> We need to have a value for <code>UNUSED_VALUE<\/code> which we are certain will not appear in the datasets. I suspect that there is a workaround this, maybe by making the dummy dataset consist of empty tensors (instead of tensors with a constant value ), but I couldn't figure out how to do this yet.<\/p>\n\n<p><strong>(2)<\/strong> Although this dataset has a finite number of elements, the following loop will never terminate:<\/p>\n\n<pre><code>iter = ds.make_one_shot_iterator()\nbatch = iter.get_next()\nsess = tf.Session()\nwhile True:\n    print(sess.run(batch))\n<\/code><\/pre>\n\n<p>The reason is that the iterator keeps filtering out dummy examples, without knowing when to stop. This can be addressed by changing the <code>repeat()<\/code> call above to <code>repeat(n)<\/code>, where <code>n<\/code> is a number that you know is longer than the difference between the lengths of the two datasets. <\/p>\n"},{"AnswerId":"52904470","CreationDate":"2018-10-20T10:00:57.677","ParentId":null,"OwnerUserId":"5737630","Title":null,"Body":"<p>If you don't mind running a session during the construction of the new dataset, you can do the following:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nds1 = tf.data.Dataset.from_tensor_slices([5,5,5,5,5])\nds2 = tf.data.Dataset.from_tensor_slices([4,4])\n\nds1 = ds1.batch(2)\nds2 = ds2.batch(1)\n\niter1 = ds1.make_one_shot_iterator()\niter2 = ds2.make_one_shot_iterator()\n\nbatch1 = iter1.get_next()\nbatch2 = iter2.get_next()\n\nsess = tf.Session()\n\n# define a generator that will sess.run both datasets, and will return the concatenation of both\ndef GetBatch():\n    while True:\n        try:\n            b1 = sess.run(batch1)\n        except tf.errors.OutOfRangeError:\n            b1 = None\n        try:\n            b2 = sess.run(batch2)\n        except tf.errors.OutOfRangeError:\n            b2 = None\n        if (b1 is None) and (b2 is None):\n            break\n        elif b1 is None:\n            yield b2\n        elif b2 is None:\n            yield b1\n        else:\n            yield np.concatenate((b1,b2))\n\n# create a dataset from the above generator\nds = tf.data.Dataset.from_generator(GetBatch,tf.int32)\n<\/code><\/pre>\n\n<p>Notice that the above session can be hidden\\encapsulated if you wish (for example, inside a function), for example:<\/p>\n\n<pre><code>iter = ds.make_one_shot_iterator()\nbatch = iter.get_next()\n\nsess2 = tf.Session()\n\nwhile True:\n    print(sess2.run(batch))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52897402,"AnswerCount":2,"Tags":"<arrays><numpy><keras>","CreationDate":"2018-10-19T17:35:38.970","AcceptedAnswerId":null,"OwnerUserId":9277688.0,"Title":"How to iterate through two numpy arrays of different dimensions","Body":"<p>I am working with the MNIST dataset, <code>x_test<\/code> has dimension of (10000,784) and <code>y_test<\/code> has a dimension of (10000,10). I need to iterate through each sample of these two numpy arrays at the same time as I need to pass them individually to <code>score.evaluate()<\/code><\/p>\n\n<p>I tried <code>nditer<\/code>, but it throws an error saying operands could not be broadcasted together since hey have different shape. <\/p>\n\n<pre><code>    score=[]\n    for x_sample, y_sample in np.nditer ([x_test,y_test]):\n         a=x_sample.reshape(784,1)\n         a=np.transpose(a)\n         b=y_sample.reshape(10,1)\n         b=np.transpose(b)\n         s=model.evaluate(a,b,verbose=0)\n         score.append(s)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52933771","CreationDate":"2018-10-22T16:23:48.597","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p><em>Assuming<\/em> that what you are actually trying to do here is to get the individual loss per sample in your test set, here is a way to do it (in your approach, even if you get past the iteration part, you will have issues with <code>model.evaluate<\/code>, which was not designed for <em>single<\/em> sample pairs)...<\/p>\n\n<p>To make the example reproducible, here I also assume we have first run the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"nofollow noreferrer\">Keras MNIST CNN example<\/a>  for only 2 epochs; so, the shape of our data is:<\/p>\n\n<pre><code>x_test.shape\n# (10000, 28, 28, 1)\n\ny_test.shape\n# (10000, 10)\n<\/code><\/pre>\n\n<p>Given that, here is a way to get the individual loss per sample:<\/p>\n\n<pre><code>from keras import backend as K\n\ny_pred = model.predict(x_test)\ny_test = y_test.astype('float32') # necessary, as y_pred.dtype is 'float32'\ny_test_tensor = K.constant(y_test)\ny_pred_tensor = K.constant(y_pred)\ng = K.categorical_crossentropy(target=y_test_tensor, output=y_pred_tensor)\n\nce = K.eval(g)  # 'ce' for cross-entropy\nce\n# array([1.1563368e-05, 2.0206178e-05, 5.4946734e-04, ..., 1.7662416e-04,\n#        2.4232995e-03, 1.8954457e-05], dtype=float32)\n\nce.shape\n# (10000,)\n<\/code><\/pre>\n\n<p>i.e. <code>ce<\/code> now contains what the <code>score<\/code> list in your question was supposed to contain.<\/p>\n\n<p>For <strong>confirmation<\/strong>, let's calculate the loss for all test samples using <code>model.evaluate<\/code>:<\/p>\n\n<pre><code>score = model.evaluate(x_test, y_test, verbose=0)\nprint('Test loss:', score[0])\n# Test loss: 0.050856544668227435\n<\/code><\/pre>\n\n<p>and again manually, averaging the values of the <code>ce<\/code> we have just calculated:<\/p>\n\n<pre><code>import numpy as np\n\nlog_loss = np.sum(ce)\/ce.shape[0]\nlog_loss\n# 0.05085654296875\n<\/code><\/pre>\n\n<p>which, although not exactly equal (due to different numeric precision involved in the two ways of calculation), they are <em>practically<\/em> equal indeed:<\/p>\n\n<pre><code>log_loss == score[0]\n# False\n\nnp.isclose(log_loss, score[0])\n# True\n<\/code><\/pre>\n\n<p>Now, the adaptation of this to your own case, where the shape of <code>x_test<\/code> is <code>(10000, 784)<\/code>, is arguably straighforward...<\/p>\n"},{"AnswerId":"52897545","CreationDate":"2018-10-19T17:45:31.770","ParentId":null,"OwnerUserId":"4313195","Title":null,"Body":"<p>You are mixing between training features and testing labels. The training set has 60,000 samples and the test set has 10,000 samples (that is, your <code>x_test<\/code> should be of dimension <code>(10000,784)<\/code>). Make sure you have download all the correct data, and don't mix up training data with testing data.<\/p>\n\n<p><a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a><\/p>\n"}]}
+{"QuestionId":52897874,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-19T18:10:29.403","AcceptedAnswerId":52902412.0,"OwnerUserId":5483457.0,"Title":"Difference between tensorflow flattening methods","Body":"<p>I came across 2 different methods of flattening the convolution pool but I do not understand why both method works. The first method, is quite straight forward to me, given from the tensorflow <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn\" rel=\"nofollow noreferrer\">example<\/a> uses: <\/p>\n\n<pre><code>pool2 = tf.reshape(pool2, [-1, 7 * 7 * 64])\n<\/code><\/pre>\n\n<p>However, I came across another code using the following:<\/p>\n\n<p>Here, <code>conv2<\/code> is taken from the convolution layer.<\/p>\n\n<pre><code>pool2 = tf.squeeze(tf.reduce_max(conv2, 1), squeeze_dims=[1])\n<\/code><\/pre>\n\n<p>both are then passed to <\/p>\n\n<pre><code>logits = tf.layers.dense(inputs=pool2, units=MAX_LABEL, activation=None)\n<\/code><\/pre>\n\n<p>which <code>logits<\/code> is finally used to calculate the <code>entropy<\/code><\/p>\n\n<pre><code>entropy = tf.reduce_mean(\n    tf.nn.softmax_cross_entropy_with_logits_v2(labels=tf.one_hot(y_, MAX_LABEL), logits=logits)) \n<\/code><\/pre>\n","answers":[{"AnswerId":"52902412","CreationDate":"2018-10-20T04:25:37.040","ParentId":null,"OwnerUserId":"10356594","Title":null,"Body":"<p>The shapes of <code>conv2<\/code> tensors are the same in both examples?\nI think the second one has reshaped the tensor in advanced.<\/p>\n\n<p>The shape of <code>conv2<\/code> in tensorflow example is <code>(batch, y, x, filters)<\/code>, while the second one would be  <code>(batch, y*x, 1, filters)<\/code> so the following code can work normally.<\/p>\n\n<p>And these two approaches have the same output, if the size of <code>conv2<\/code> is the same as <code>pool_size<\/code> (2x2 in this case), which will only output one value per filter and this is \ntotally the same idea as <code>reduce_max<\/code>.<\/p>\n\n<p>For example: <\/p>\n\n<pre><code>import tensorflow as tf\n'The same input, but different shape' \nrandom_tensor = np.random.random([16,2,2,64])\nmethod1_input=tf.constant(random_tensor)             # shape = (16,2,2,64)\nmethod2_input=tf.reshape(method1_input,[16,4,1,64])  # shape = (16,4,1,64)\n'method 1 : maxpool'\nmaxpool      = tf.layers.max_pooling2d(inputs=method1_input, pool_size=[2, 2], strides=2)\nmaxpool_flat = tf.reshape(maxpool, [-1,64])\n\n'method 2 : reduce_max and squeeze'\nred_max = tf.reduce_max(method2_input, 1)   # shape = (16,1,64)\npool2   = tf.squeeze(red_max, squeeze_dims=[1])  # shape = (16,64)  ,literally squeeze dim [1]\n\nwith tf.Session() as sess :\n    method1_result=(sess.run(maxpool_flat) )\n    method2_result=(sess.run(pool2) )\n    Is_true = sess.run(tf.equal(method1_result,method2_result)  )\n    print(Is_true)\n    # output\n    #[[ True  True  True ...  True  True  True]\n    # [ True  True  True ...  True  True  True]\n    # [ True  True  True ...  True  True  True]\n    # ...\n    # [ True  True  True ...  True  True  True]\n    # [ True  True  True ...  True  True  True]\n    # [ True  True  True ...  True  True  True]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52897897,"AnswerCount":2,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-10-19T18:12:58.237","AcceptedAnswerId":null,"OwnerUserId":10530445.0,"Title":"How do I import a trained deep CNN model?","Body":"<p>I have trained a deep CNN model on GPU on a cluster using Tensorflow. Now I want to load that model in my laptop. But when I want to use the following command I receive an error:<\/p>\n\n<pre><code>new_saver = tf.train.import_meta_graph('.\/3\/Model_Arch3\/Deep_CNN_Color_Arch8.ckpt-178000.meta')\n<\/code><\/pre>\n\n<p>The error that I receive is:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"dev_test.py\", line 28, in\n  \n      new_saver = tf.train.import_meta_graph('.\/3\/Model_Arch3\/Deep_CNN_Color_Arch8.ckpt-178000.meta')\n  File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\",\n  line 1810, in import_meta_graph\n      **kwargs)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/meta_graph.py\",\n  line 660, in import_scoped_meta_graph\n      producer_op_list=producer_op_list)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/importer.py\",\n  line 285, in import_graph_def\n      raise ValueError('No op named %s in defined operations.' % node.op) ValueError: No op named ParseSingleExample in defined\n  operations.<\/p>\n<\/blockquote>\n\n<p>I really appreciate if someone help me with this problem.<\/p>\n","answers":[{"AnswerId":"52898354","CreationDate":"2018-10-19T18:52:26.183","ParentId":null,"OwnerUserId":"10508893","Title":null,"Body":"<p>This may also be a lazy loading issue with the operation you are trying to perform.\nTry adding <code>dir(tf.contrib)<\/code> a line before you load your graph.<\/p>\n\n<p>For reference \u2014 a similar issue was raised in an old TF GitHub issue \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10130\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10130<\/a><\/p>\n\n<hr>\n\n<p>UPDATE:<\/p>\n\n<p>I think I see the issue now. Since you trained on a GPU and youre running on a regular machine \/ CPU now. You need to clear the devices from the loaded object. Try adding clear_devices as a param:<\/p>\n\n<pre><code>dir(tf.contrib)\nwith tf.Session() as sess:\n    new_saver = tf.train.import_meta_graph(full_file_path, clear_devices=True)\n<\/code><\/pre>\n\n<p>Lastly, your TF version between your local and GPU machine may be out of sync. Try to update your local version.<\/p>\n"},{"AnswerId":"52900124","CreationDate":"2018-10-19T21:27:32.233","ParentId":null,"OwnerUserId":"10530445","Title":null,"Body":"<p>I found what was the problem.\nI was using python 3.5.4 for training procedure. Then I used python 2.7 for reconstructing the graph and that was the problem.\nThanks from those who tried to help<\/p>\n"}]}
+{"QuestionId":52897995,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-19T18:21:39.287","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"tf.transform: add preprocessing to Keras model?","Body":"<p>I have a <code>keras<\/code> model for text classification using a <code>tensorflow<\/code> backend.  It currently assumes the input is a numpy array of integers.<\/p>\n\n<p>I'd like to modify this so that I can train and predict on raw text.  From what I've gathered, this involves using <code>tf.transform<\/code> to convert a tensor of strings into a tensor of integers.<\/p>\n\n<p>I've done this using <code>tf.transform<\/code> but now am unsure how to add this preprocessing step to my model as the very first layer \/ step.  To be clear, my input data looks like this:<\/p>\n\n<p><code>[{\"review\":\"movie is great},\n{\"review\":\"awful film\"}]<\/code><\/p>\n\n<p>and the output is:<\/p>\n\n<p><code>[{\"review_out\": array([-1, -1, 1, 0, 2])},\n{\"review_out\": array([-1, -1, -1, 3, 4])]<\/code><\/p>\n\n<p>The function that does this is called <code>preprocess<\/code>.  So I just want to include running <code>preprocess<\/code> as the first step in my DAG.<\/p>\n\n<p>How might I do this?<\/p>\n\n<p>For reference, this is important because I want to do live prediction on ML Engine.<\/p>\n","answers":[{"AnswerId":"52899970","CreationDate":"2018-10-19T21:11:36.510","ParentId":null,"OwnerUserId":"10458815","Title":null,"Body":"<p>If you use a <code>tf.data.Dataset()<\/code>, there exists a <code>tf.data.Dataset().map(map_func)<\/code> function that allows you to apply <code>map_func<\/code> to all elements of your dataset. This could be used to add your preprocessing step.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset<\/a><\/p>\n\n<p>For example:<\/p>\n\n<p><code>dataset = tf.data.Dataset.from_tensor_slices((x, y))<\/code><\/p>\n\n<p><code>dataset = dataset.map(preprocess)<\/code><\/p>\n\n<p><code>model.fit(dataset, ....)<\/code><\/p>\n\n<p>There are other advantages to using <code>tf.data.Dataset()<\/code> over a numpy array as well.<\/p>\n"}]}
+{"QuestionId":52898784,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-10-19T19:27:26.950","AcceptedAnswerId":null,"OwnerUserId":10302242.0,"Title":"Freezing weight of neural network such that its output takes a particular value at a particular point (tensorflow)","Body":"<p>Let's say I have a neural network that looks like this<\/p>\n\n<pre><code>def neural_net(x):\n    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])\n    layer_1 = tf.nn.sigmoid(layer_1)\n\n    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])\n    layer_2 = tf.nn.sigmoid(layer_2)\n\n    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']\n    return out_layer\n<\/code><\/pre>\n\n<p>Is there a way in tensorflow to fix the weights in such a way that <code>neural_net(a)<\/code> always returns <code>b<\/code> (where <code>a,b<\/code> are real numbers), e.g., <code>f(1) = 0<\/code>?<\/p>\n","answers":[{"AnswerId":"52902870","CreationDate":"2018-10-20T05:59:05.303","ParentId":null,"OwnerUserId":"5707705","Title":null,"Body":"<p>Sure, however, the answer depends a bit on the purpose. <\/p>\n\n<p>The easiest solution is to just scale the output. For example by running the result through a linear regressor. While this gives the desired result it is probably not what you want.<\/p>\n\n<p>However, probably the better way is to integrate this additional objective in the loss function during training. This way you can trade off between your additional requirement and fitting the weights of your neural network. A generic description how to adapt the loss you can find at <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.losses\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.losses<\/a>    <\/p>\n\n<pre><code>images, labels = LoadData(...)\npredictions = MyModelPredictions(images)\n\nweight = MyComplicatedWeightingFunction(labels)\nweight = tf.div(weight, tf.size(weight))\nloss = tf.contrib.losses.mean_squared_error(predictions, depths, weight)\n<\/code><\/pre>\n\n<p>The weight for your special case need to be extremely high.\nThat way your critiria is not full guarantee but very likely.<\/p>\n\n<p>In addition you need to rewrite the mini-batching mechanism to inject in each batch your <code>(x,y) = (1,0)<\/code> example<\/p>\n"}]}
+{"QuestionId":52899550,"AnswerCount":0,"Tags":"<keras><autoencoder>","CreationDate":"2018-10-19T20:33:37.640","AcceptedAnswerId":null,"OwnerUserId":10530826.0,"Title":"Keras autoencoder optimizer and loss function","Body":"<p>I've been implementing an autoencoder which receives as inputs vectors that consist only of 0 and 1, such as [1, 0, 1, 0, 1, 0, ...]. <\/p>\n\n<p>Likewise, another autoencoder that receives as inputs vectors that consist in values between 0 and 1, such as [0.123, 1, 0.9, 0.01, 0.9, ...]. In both cases each vector element is the input value of a node. The activation function of the hidden layers is relu and for the output layer is sigmoid.<\/p>\n\n<p>I've seen some examples of autoencoders where <code>adam\/adadelta<\/code> are used as optimizer and binary_crossentropy is used as a loss function. For that reason I implemented in both <code>adadelta<\/code> and <code>binary_crossentropy<\/code>, but I'm not sure if for both cases it's the correct configuration.<\/p>\n","answers":[]}
+{"QuestionId":52899988,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-19T21:12:53.537","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras Lambda Layer Before Embedding: Use to Convert Text to Integers","Body":"<p>I currently have a <code>keras<\/code> model which uses an <code>Embedding<\/code> layer.  Something like this:<\/p>\n\n<pre><code>input = tf.keras.layers.Input(shape=(20,) dtype='int32')\nx = tf.keras.layers.Embedding(input_dim=1000,\n                              output_dim=50,\n                              input_length=20,\n                              trainable=True,\n                              embeddings_initializer='glorot_uniform',\n                              mask_zero=False)(input)\n<\/code><\/pre>\n\n<p>This is great and works as expected.  However, I want to be able to send text to my model, have it preprocess the text into integers, and continue normally.<\/p>\n\n<p>Two issues:<\/p>\n\n<p>1) The Keras docs say that Embedding layers can only be used as the first layer in a model: <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/embeddings\/<\/a><\/p>\n\n<p>2) Even if I could add a <code>Lambda<\/code> layer before the <code>Embedding<\/code>, I'd need it to keep track of certain state (like a dictionary mapping specific words to integers).  How might I go about this stateful preprocessing?<\/p>\n\n<p>In short, I need to modify the underlying Tensorflow DAG, so when I save my model and upload to ML Engine, it'll be able to handle my sending it raw text.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53993288","CreationDate":"2019-01-01T05:51:19.700","ParentId":null,"OwnerUserId":"8304839","Title":null,"Body":"<p>Here are the first few layers of a model which uses a string input:<\/p>\n\n<pre><code>input = keras.layers.Input(shape=(1,), dtype=\"string\", name='input_1')\nlookup_table_op = tf.contrib.lookup.index_table_from_tensor(\n mapping=vocab_list,\n num_oov_buckets=num_oov_buckets,\n default_value=-1,\n)\nlambda_output = Lambda(lookup_table_op.lookup)(input)\nemb_layer = Embedding(int(number_of_categories),int(number_of_categories**0.25))(lambda_output)\n<\/code><\/pre>\n\n<p>Then you can continue the model as you normally would after an embedding layer. This is working for me and the model trains fine from string inputs.<\/p>\n\n<p>It is recommended that you do the string -> int conversion in some preprocessing step to speed up the training process. Then after the model is trained you create a second keras model that just converts string -> int and then combine the two models to get the full string -> target model.<\/p>\n"}]}
+{"QuestionId":52900017,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-19T21:15:31.820","AcceptedAnswerId":53829337.0,"OwnerUserId":1727392.0,"Title":"Keras: Loading model built with CuDNNLSTM on host without GPU","Body":"<p>I trained a keras model that uses <code>CuDNNLSTM<\/code> cells, and now wish to load the model on a host device that lacks a GPU. Because <code>CuDNNLSTM<\/code> cells require a GPU, though, the loading process bombs out, throwing:<\/p>\n\n<blockquote>\n  <p>No OpKernel was registered to support Op 'CudnnRNN' with these attrs.<\/p>\n<\/blockquote>\n\n<p>Is there some backdoor that will allow me to load the model on a host without a GPU? Any suggestions would be very helpful!<\/p>\n","answers":[{"AnswerId":"53829337","CreationDate":"2018-12-18T08:52:19.817","ParentId":null,"OwnerUserId":"10630288","Title":null,"Body":"<p>Note: I am using Keras 2.2.4 and TensorFlow 1.12.0.\nI was able to solve the issue with the following steps:<\/p>\n\n<p>1) Train the model with CudnnLSTM and save the model (model_GPU.json) and the weights (*.h5).<\/p>\n\n<p>2) Define the same model changing CudnnLSTM for LSTM, this must be done in a system\/computer with no GPU, and then you can save the model (model_CPU.json).<\/p>\n\n<p>2*) In the LSTM cell set activation='tanh',recurrent_activation='sigmoid'. Since these are the default ones in CudnnLSTM.<\/p>\n\n<p>3) Then you can load model_CPU.json with the weights trained with CudnnLSTM.<\/p>\n\n<p>Specifically, I used the following<\/p>\n\n<p><strong>CPU:<\/strong><\/p>\n\n<pre><code>from keras.layers import LSTM\n\nBidirectional(LSTM(hidden_units_LSTM, return_sequences=True,activation='tanh',recurrent_activation='sigmoid'))(output)\n<\/code><\/pre>\n\n<p><strong>GPU:<\/strong><\/p>\n\n<pre><code>from keras.layers import CuDNNLSTM\n\nBidirectional(CuDNNLSTM(hidden_units_LSTM, return_sequences=True))(output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52901546,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-20T00:59:11.413","AcceptedAnswerId":52901778.0,"OwnerUserId":10531469.0,"Title":"And operator on two rank-1 tensorflow tensors","Body":"<p>I have two rank-1 tensors, one of them contains X floats, the other is one-hot, also with X entries. <\/p>\n\n<p>I want to create a new rank-1 tensor, with a single element: the float in the first tensor at the index where the one-hot vector is equal to 1.<\/p>\n\n<p>Any help would be much appreciated.<\/p>\n","answers":[{"AnswerId":"52901778","CreationDate":"2018-10-20T01:53:13.847","ParentId":null,"OwnerUserId":"6353827","Title":null,"Body":"<p>It sounds like tf.boolean_mask is what you're looking for: \n  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/boolean_mask\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/boolean_mask<\/a><\/p>\n"}]}
+{"QuestionId":52901611,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-lite>","CreationDate":"2018-10-20T01:14:44.953","AcceptedAnswerId":null,"OwnerUserId":10530640.0,"Title":"Changing input batch size of provided model","Body":"<p>Sorry for the newbie question. I've tried looking for examples but most of the time it's pseudo-code\/snippets I can't run and it's hard to tell what's what, especially with all the different TF workflows.<\/p>\n\n<p>I am trying to convert a pre-trained TensorFlow Saved Model, ResNet-50 v2 (fp32), to a quantized TensorFlow Lite file and have two issues:<\/p>\n\n<ol>\n<li>The batch size appears to be fixed at 64. i.e. the input tensor is 64x224x224x3. I would like it to be 1x224x224x3<\/li>\n<li>There are error messages when loading\/converting the Saved Model, both when using tflite_convert\/toco and when using resnet_v2_50()<\/li>\n<\/ol>\n\n<p>The model appears correct at a glance, even with the error messages, so I'm mostly concerned about the batch size modification.<\/p>\n\n<p>The Saved Model I tried converting:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/official\/resnet#pre-trained-model\" rel=\"nofollow noreferrer\">ResNet-50 v2 (fp32) Model<\/a><\/p>\n\n<ul>\n<li><a href=\"http:\/\/download.tensorflow.org\/models\/official\/20181001_resnet\/savedmodels\/resnet_v2_fp32_savedmodel_NHWC.tar.gz\" rel=\"nofollow noreferrer\">SavedModel (NHWC)<\/a><\/li>\n<\/ul>\n\n<p>The Checkpoint data and .py model I tried using to generate a new Saved Model:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim#pre-trained-models\" rel=\"nofollow noreferrer\">ResNet V2 50^<\/a><\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/slim\/nets\/resnet_v2.py\" rel=\"nofollow noreferrer\">resnet_v2.py model<\/a><\/li>\n<li><a href=\"http:\/\/download.tensorflow.org\/models\/resnet_v2_50_2017_04_14.tar.gz\" rel=\"nofollow noreferrer\">resnet_v2_50_2017_04_14.tar.gz checkpoint<\/a><\/li>\n<\/ul>\n\n<p>How I converted it to .tflite in bash:<\/p>\n\n<pre><code>tflite_convert --output_file resnet_imagenet_v2_uint8_20181001.tflite --saved_model_dir . --post_training_quantize\n<\/code><\/pre>\n\n<p>This results in a reasonable model for 64x224x224x3. This might run okay for Android\/iOS (haven't tried) despite the errors, but I'm trying to use it on a custom platform for experimenting.<\/p>\n\n<p>The script I tried to generated a Saved Model with my target input shape:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nfrom tensorflow.contrib.slim.nets import resnet_v2\n\ndef main():\n    # Directory containing resnet_v2_50.ckpt\n    ckpt_dir = \"\/media\/resnet_v2_50_2017_04_14\/\"\n    with tf.contrib.slim.arg_scope(resnet_v2.resnet_arg_scope()):\n        input_tensor = tf.placeholder(tf.float32, shape=[1,224,224,3], name=\"input_tensor\")\n        output_tensor = tf.placeholder(tf.float32, shape=[1,1000])\n\n        # Create model\n        # Generates errors for all Conv2D nodes like:\n        # 2018-10-19 16:41:41.393976: E tensorflow\/core\/framework\/node_def_util.cc:110] Error in the node: {{node resnet_v2_50\/conv1\/Conv2D}} = Conv2D[T=DT_FLOAT, data_format=\"NHWC\", dilations=[1, 1, 1, 1], padding=\"VALID\", strides=[1, 2, 2, 1], use_cudnn_on_gpu=true](resnet_v2_50\/Pad, resnet_v2_50\/conv1\/weights\/read)\n        net, end_points = resnet_v2.resnet_v2_50(input_tensor, 1000)\n\n        # Load checkpoint data\n        sv = tf.train.Supervisor(logdir=ckpt_dir)\n        with sv.managed_session() as sess:\n            # Allows saving, but unexpected results:\n            # sess.graph._unsafe_unfinalize() \n\n            # Below call fails with:\n            # RuntimeError: Graph is finalized and cannot be modified.\n            tf.saved_model.simple_save(\n                sess,\n                \".\/export\",\n                inputs={\"input_tensor\": input_tensor},\n                outputs={\"resnet_v2_50\/predictions\/Softmax\": output_tensor}\n            )\n\nmain()\n<\/code><\/pre>\n\n<p>I think I'm doing something quite wrong, but it looks close to examples I've found online so I'm a bit lost. Using <code>sess.graph._unsafe_unfinalize()<\/code> allows the simple_save() to run, creating a .pb and variables\/ directory under export\/, but when I view it in Netron, a model viewer I found, there's many more nodes than the provided Saved Model (batch size 64) I downloaded from the tensorflow models repo. 3395 vs 1930.<\/p>\n\n<p>Trying to convert this model anyways results in this error:<\/p>\n\n<pre><code>$ tflite_convert --output_file export.tflite --saved_model_dir . --post_training_quantize\n.\n.\n.\nTraceback (most recent call last):\n  File \"\/home\/tfuser\/venv\/bin\/tflite_convert\", line 11, in &lt;module&gt;\n    sys.exit(main())\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 412, in main\n    app.run(main=run_main, argv=sys.argv[:1])\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 408, in run_main\n    _convert_model(tflite_flags)\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 162, in _convert_model\n    output_data = converter.convert()\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", line 453, in convert\n    **converter_kwargs)\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 370, in toco_convert_impl\n    input_data.SerializeToString())\n  File \"\/home\/tfuser\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 149, in toco_convert_protos\n    \"TOCO failed see console for info.\\n%s\\n%s\\n\" % (stdout, stderr))\nRuntimeError: TOCO failed see console for info.\nb'2018-10-19 17:00:58.673690: F tensorflow\/contrib\/lite\/toco\/tooling_util.cc:886] Check failed: GetOpWithInput(model, input_array.name()) Specified input array \"input_tensor\" is not consumed by any op in this graph. Is it a typo? To silence this message, pass this flag:  allow_nonexistent_arrays\\n'\nNone\n<\/code><\/pre>\n\n<p>So it looks like my use of input_tensor was wrong and is creating a lot of extra nodes?<\/p>\n\n<p>My setup:<\/p>\n\n<ul>\n<li>CentOS 7<\/li>\n<li>Python 3.6.6<\/li>\n<li>TensorFlow Nightly 1.13.0.dev20181018 (CPU)<\/li>\n<\/ul>\n\n<p>Some pages I looked at for help:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/40586471\/what-is-the-best-way-to-run-saved-model-with-different-batch-size-in-tensorflow\">This one<\/a> seems to suggest changing the model to take a placeholder batch size and re-train. I don't think this is suitable if I'm using a pre-trained model.<\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/49872844\/how-to-load-trained-tensorflow-model-to-predict-with-different-batch-size\">This one<\/a> regenerates the frozen model with a different input placeholder.<\/li>\n<li><a href=\"https:\/\/www.reddit.com\/r\/tensorflow\/comments\/9aijuz\/changing_the_batch_size_of_a_frozen_graph_pb\/e4wlp8m\" rel=\"nofollow noreferrer\">This<\/a> sounds good, but I'm not sure how to do it.<\/li>\n<\/ul>\n\n<p><strong>Any ideas on how to achieve this? A better strategy or a fix to my script would be very appreciated and welcome.<\/strong><\/p>\n\n<p>Secondary questions:<\/p>\n\n<ul>\n<li>The documentation for simple_save() says inputs and outputs should be a map of strings to tensors.\n\n<ul>\n<li>By tensor does it mean the tf.placeholder? Or a numpy array? Or a tf.Variable?<\/li>\n<li>For the string, does it expect the name of the input and output tensors? So \"input_tensor\" and \"resnet_v2_50\/predictions\/Softmax\" are correct?<\/li>\n<\/ul><\/li>\n<li>Do I need to call <code>sess.run(tf.global_variables_initializer())<\/code> at some point? If I try to call it in the above script, it says the graph is finalized.<\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":52901812,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning>","CreationDate":"2018-10-20T02:01:16.447","AcceptedAnswerId":null,"OwnerUserId":1924507.0,"Title":"how to change this model output from float to string?","Body":"<pre><code>from __future__ import print_function\n\nimport numpy as np\nfrom keras.callbacks import EarlyStopping\nfrom keras.layers.core import Dense, Activation\nfrom keras.models import Sequential\nfrom keras.optimizers import SGD\nfrom keras.layers import Layer\nimport keras.backend as K\n\nnp.random.seed(1337)  # for reproducibility\n\nclass Round(Layer):\n    def __init__(self, **kwargs):\n        super(Round, self).__init__(**kwargs)\n\n    def get_output(self, train=False):\n        X = self.get_input(train)\n        return K.round(X)\n\n    def get_config(self):\n        config = {\"name\": self.__class__.__name__}\n        base_config = super(Round, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\n\ndef build_and_train_mlp_network(X_train, y_train, X_test, y_test):\n\n    nb_epoch = 1000\n    batch_size = 4\n\n    model = Sequential()\n    model.add(Dense(2, input_shape=(X_train.shape[1],)))\n    model.add(Activation('sigmoid'))\n    model.add(Dense(1))\n    model.add(Activation('sigmoid'))\n    model.add(Round()) # return K.round(X)\n\n    sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\n    model.compile(loss='mean_squared_error', optimizer=sgd)  # or binary_crossentropy\n\n    model.fit(X_train,\n              y_train,\n              batch_size=batch_size,\n              nb_epoch=nb_epoch,\n              verbose=0,\n              validation_data=(X_test, y_test))\n    return model\n\n\nif __name__ == \"__main__\":\n\n    X_test = X_train = np.array([[1, 1], [1, 0], [0, 1], [0, 0]])\n\n    y_train = y_test = np.array([1, 2, 3, 4])\n\n    model = build_and_train_mlp_network(X_train, y_train, X_test, y_test)\n    print(model.predict(X_test))\n<\/code><\/pre>\n\n<p>how to change this model output from float to string ?<\/p>\n\n<p>output is <code>[[0.9999342 ] [0.99991477] [0.99989116] [0.9998405 ]]<\/code><\/p>\n\n<p>now i change <\/p>\n\n<pre><code>y_train = y_test = np.array(['a', 'b', 'c', 'd'])\n<\/code><\/pre>\n\n<p>how to make output ( model.predict(X_test) ) to be string ?<\/p>\n","answers":[{"AnswerId":"52913273","CreationDate":"2018-10-21T08:06:48.413","ParentId":null,"OwnerUserId":"5707705","Title":null,"Body":"<p>Here is a bit of guessing involved: i assume your \"round\" layers purpose is to get an output which is a integer number. However, that way you are most likely killing the gradient in the SGD.<\/p>\n\n<p>To avoid the problem in the gradient you should do it as a post processing step. Hence, I'd recommend to alter the neural network the following way:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    model = Sequential()\n    model.add(Dense(2, input_shape=(X_train.shape[1],)))\n    model.add(Activation('sigmoid'))\n    model.add(Dense(1))\n#    model.add(Activation('sigmoid'))\n    model.add(Activation('linear'))\n<\/code><\/pre>\n\n<p>The linear activation to achieve numbers larger than <code>1<\/code>. this function is the reason you got <code>[[0.9999342 ] [0.99991477] [0.99989116] [0.9998405 ]]<\/code>\n Afterwards you can post process the outputs<\/p>\n\n<pre><code>print(np.round(model.predict(X_test)))\n<\/code><\/pre>\n\n<p>or to map to characters you can use<\/p>\n\n<pre><code>maps = ['foo','a','b','c','d','bar'] # ugly what about -1...6,... etc\n[maps[int(round(i))] for i in  model.predict(X_test).flatten()]\n<\/code><\/pre>\n\n<p>With the functional API you can actually push the rounding into the NN. I had to re-write your layer a bit<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Round(Layer):\n    def __init__(self, **kwargs):\n        super(Round, self).__init__(**kwargs)\n\n    def call(self,X):\n        print(X)\n        return K.round(X)\n\n    def get_config(self):\n        config = {\"name\": self.__class__.__name__}\n        base_config = super(Round, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\n\nr = Round()(outputs)\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodelr = Model(inputs=inputs, outputs=r)\nmodelr.compile(loss='mean_squared_error', optimizer=sgd)\nmodelr.predict(X_test)\n<\/code><\/pre>\n\n<p>Having said that I think you have a fundamental flaw in your model. The model assumes that there is a ordering in the characters. Which means that <code>b<\/code> is something in between <code>a<\/code> and <code>c<\/code>. This is true from the perspective of the alphabet, however, wrong in the sense of semantic.<\/p>\n\n<p>Therefore, you should use the standard one-hot encoding approach. Also the problem which unmappel characters, e.g., <code>foo<\/code>,<code>bar<\/code> wouldn't happen. Check the standard TF\/Keras Mnist example<\/p>\n\n<p>You can than build your model the following way <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.utils import np_utils\n\nnum_classes=5 # sorry but we need zero\n\ndef build_and_train_mlp_network(X_train, y_train, X_test, y_test):\n\n    nb_epoch = 1000\n    batch_size = 4\n\n    inputs = Input(shape=(X_train.shape[1],))\n    d1 = Dense(2)(inputs)\n    a1 = Activation('sigmoid')(d1)\n    d2 = Dense(num_classes)(a1)\n    a2 = Activation(activation='softmax')(d2)\n\n\n    model = Model(inputs=inputs, outputs=a2)\n\n    sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\n    model.compile(loss='mean_squared_error', optimizer=sgd)  # or binary_crossentropy\n\n    model.fit(X_train,\n              y_train,\n              batch_size=batch_size,\n              nb_epoch=nb_epoch,\n              verbose=0,\n              validation_data=(X_test, y_test))\n    return model,inputs,a2\n\n\n\n\nX_test = X_train = np.array([[1, 1], [1, 0], [0, 1], [0, 0],[1, 1]])\ny_train = y_test = np_utils.to_categorical(np.array([0, 1, 2, 3,0]),num_classes=5)\nmodel,inputs,outputs = build_and_train_mlp_network(X_train, y_train, X_test, y_test)\nprint(model.summary())\nmodel.predict(X_test)\n<\/code><\/pre>\n\n<p>and if you would like to get max class again from the compute graph<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class ArgMax(Layer):\n    def __init__(self, **kwargs):\n        super(ArgMax, self).__init__(**kwargs)\n\n    def call(self,X):\n        print(X)\n        return K.argmax(X, axis=None)\n\n    def get_config(self):\n        config = {\"name\": self.__class__.__name__}\n        base_config = super(Round, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\nam = ArgMax()(outputs)\nmodelam=Model(inputs=inputs,outputs=[am,outputs])\nclazzes,confidence = modelam.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52902319,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><precision-recall>","CreationDate":"2018-10-20T04:03:53.903","AcceptedAnswerId":null,"OwnerUserId":9836402.0,"Title":"Tensorflow Object-detection api","Body":"<p>How can we draw ground truth boundary box with predicted boundary box at the time of inference by making use of tensorflow object detection api?<\/p>\n\n<p>How to calculate precision,recall &amp; mAP for object detection using SSD model with KITTI like dataset?<\/p>\n","answers":[]}
+{"QuestionId":52902937,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning><keras>","CreationDate":"2018-10-20T06:12:23.750","AcceptedAnswerId":null,"OwnerUserId":7536495.0,"Title":"How to predict input probability of image using trained model in Keras?","Body":"<p>I'm beginning ML and have come to a huge road bump, looked at this thing for hours. I want to get the predict() probability as output but the predictions are only outputting [[1.]] for every test image. Using larger numbers of training data and more epochs, the acc and validation acc are up to about 90% each. This is only binary classification but I don't want predict_classes. I have no idea why it's printing [[1.]]<\/p>\n\n<p>Here is the code I'm using:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras.layers import Activation, Dropout, Flatten, Dense\nfrom keras import backend as K\n\nimg_width, img_height = 150, 150\n\ntrain_data_dir = 'D:\\Machine_Learning\\\\train'\nvalidation_data_dir = 'D:\\Machine_Learning\\\\test'\nnb_train_samples = 20000\nnb_validation_samples = 7000\nepochs = 50\nbatch_size = 40\n\nif K.image_data_format() == 'channels_first':\n    input_shape = (3, img_width, img_height)\nelse:\n    input_shape = (img_width, img_height, 3)\n\n# Build model structure\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=input_shape))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\n\n\nmodel.compile(loss='binary_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n\n# Image augmentation\ntrain_datagen = ImageDataGenerator(\n    rescale=1. \/ 255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1. \/ 255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode='binary')\n\nvalidation_generator = test_datagen.flow_from_directory(\n    validation_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode='binary')\n\nmodel.fit_generator(\n    train_generator,\n    steps_per_epoch=nb_train_samples \/\/ batch_size,\n    epochs=epochs,\n    validation_data=validation_generator,\n    validation_steps=nb_validation_samples \/\/ batch_size)\n\nmodel.save('first_try.h5')\nmodel.save_weights('my_weights.model')\n<\/code><\/pre>\n\n<p>And for predictions:<\/p>\n\n<pre><code>from keras.models import load_model\nfrom keras.preprocessing import image\nmodel = load_model('first_try.h5')\n\ndef prepare(image):\n    imageSize = 150\n    #imageArray = cv2.imread(filePath)\n    newImageArray = cv2.resize(image, (imageSize, imageSize))\n    return newImageArray.reshape(-1, imageSize, imageSize, 3)\n\nfor i in range(len(test_images)):\n    print(test_images[i])\n    im = prepare(test_images[i])\n    prediction = model.predict([im])\n    print(prediction)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52903419,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-10-20T07:29:56.893","AcceptedAnswerId":null,"OwnerUserId":7604220.0,"Title":"[Keras]Some goes wrong when i create my Model","Body":"<p>When i use kears create my model,this error is occurring.\nScale3 Scale2 Scale1 are same operate.\nThe code is:<\/p>\n\n<pre><code>Scale3 = myconv2d(Scale3,255,1,1,'valid')\nmodel = Model(inputs = X_input, outputs = [Scale3,Scale2,Scale1], name='YOLOV3')\n<\/code><\/pre>\n\n<p>Error is:<a href=\"https:\/\/i.stack.imgur.com\/0McCh.png\" rel=\"nofollow noreferrer\">error_image<\/a><\/p>\n\n<p>myconv2d function is that<\/p>\n\n<pre><code>def myconv2d(X,f,size,s,stage,p = 'same'):\n    X = Conv2D(f,kernel_size = (size,size),strides = (s,s),padding = 'same',name = 'conv2d_'+ str(stage))(X)\n    X = BatchNormalization(axis=3,name ='bn_' + str(stage))(X)\n    X = K.relu(X,0.1)\n    return X\n<\/code><\/pre>\n\n<p>How can i fix it?<\/p>\n","answers":[{"AnswerId":"52907080","CreationDate":"2018-10-20T15:11:52.503","ParentId":null,"OwnerUserId":"8005524","Title":null,"Body":"<p>You can use<code>layers.advanced_activations.LeakyReLU(0.1)(X)<\/code> to replace the code:<code>K.relu(X,0.1)<\/code><\/p>\n\n<p>Matias Valdenegro is also correct<\/p>\n"},{"AnswerId":"52904279","CreationDate":"2018-10-20T09:36:14.960","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You can't apply backend functions directly on Keras tensors, you need to use a Lambda layer for that:<\/p>\n\n<pre><code>def myconv2d(X,f,size,s,stage,p = 'same'):\n    X = Conv2D(f,kernel_size = (size,size),strides = (s,s),padding = 'same',name = 'conv2d_'+ str(stage))(X)\n    X = BatchNormalization(axis=3,name ='bn_' + str(stage))(X)\n    X = Lambda(lambda x: K.relu(x,0.1)(X)\n    return X\n<\/code><\/pre>\n"}]}
+{"QuestionId":52903493,"AnswerCount":1,"Tags":"<keras><google-colaboratory>","CreationDate":"2018-10-20T07:40:08.363","AcceptedAnswerId":52904537.0,"OwnerUserId":10386005.0,"Title":"Error in colab in finding directory and file","Body":"<pre><code>img = load_img('train\/cat\/cat.12499.jpg')\n<\/code><\/pre>\n\n<p>I'm using colab with the same code that I used in Jupyter, but now I get the below error, although the directory and file there are in:<\/p>\n\n<p>No such file or directory: 'train\/cat\/cat.12499.jpg'<\/p>\n\n<p>I copied the directory in google drive!<\/p>\n","answers":[{"AnswerId":"52904537","CreationDate":"2018-10-20T10:12:14.340","ParentId":null,"OwnerUserId":"10386005","Title":null,"Body":"<p>I solved with these commands<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/drive')\n<\/code><\/pre>\n\n<p>and I've modified the code in this way:<\/p>\n\n<pre><code>img = load_img(\"drive\/My Drive\/app\/train\/cat\/cat.12499.jpg\")  # t\n<\/code><\/pre>\n"}]}
+{"QuestionId":52903988,"AnswerCount":2,"Tags":"<tensorflow><keras><gpu>","CreationDate":"2018-10-20T08:56:40.510","AcceptedAnswerId":null,"OwnerUserId":9328846.0,"Title":"tensorflow gpu is only running on CPU","Body":"<p>I installed Anaconda-Navigatoron Windows 10 and all necessary Nvidia\/Cuda packages, created a new environment called tensorflow-gpu-env, updated PATH information, etc. When I run a model (build by using <code>tensorflow.keras<\/code>), I see that CPU utilization increases significantly, GPU utilization is 0%, and the model just does not train. <\/p>\n\n<p>I run a couple of tests to make sure how things look:<\/p>\n\n<pre><code>print(tf.test.is_built_with_cuda())\nTrue\n<\/code><\/pre>\n\n<p>The above output ('True') looks correct.<\/p>\n\n<p>Another try:<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 1634313269296444741\n, name: \"\/device:GPU:0\"\ndevice_type: \"GPU\"\nmemory_limit: 1478485606\nlocality {\n  bus_id: 1\n  links {\n  }\n}\nincarnation: 16493618810057409699\nphysical_device_desc: \"device: 0, name: GeForce 940MX, pci bus id: 0000:01:00.0, compute capability: 5.0\"\n]\n<\/code><\/pre>\n\n<p>So far so good... Later in my code, I start the training with the following code:<\/p>\n\n<pre><code>history = merged_model.fit_generator(generator=train_generator,\n                                     epochs=60,\n                                     verbose=2,\n                                     callbacks=[reduce_lr_on_plateau],\n                                     validation_data=val_generator,\n                                     use_multiprocessing=True,\n                                     max_queue_size=50,\n                                     workers=3)\n<\/code><\/pre>\n\n<p>I also tried to run the training as following:<\/p>\n\n<pre><code>with tf.device('\/gpu:0'):\n    history = merged_model.fit_generator(generator=train_generator,\n                                         epochs=60,\n                                         verbose=2,\n                                         callbacks=[reduce_lr_on_plateau],\n                                         validation_data=val_generator,\n                                         use_multiprocessing=True,\n                                         max_queue_size=50,\n                                         workers=3)\n<\/code><\/pre>\n\n<p>No matter how I start the training, it never starts the training, I keep seeing increased CPU utilization with 0% GPU utilization.<\/p>\n\n<p>Why is my tensorflow-gpu installation is only using the CPU? Spent HOURS with literally no progress.<\/p>\n\n<p><strong>ADDENDUM<\/strong><\/p>\n\n<p>When I run <code>conda list<\/code> on the console, I see the following regarding tensorflow:<\/p>\n\n<pre><code>tensorflow-base           1.11.0          gpu_py36h6e53903_0\ntensorflow-gpu            1.11.0                    &lt;pip&gt;\n<\/code><\/pre>\n\n<p>What is this tensorflow-base? Can it cause a problem? Before installing tensorflow-gpu, I made sure that I uninstalled tensorflow and tensorflow-gpu by using both conda and pip; and then installed tensorflow-gpu by using <code>pip<\/code>. I am not sure if this tensorflow-base came with my <code>tensorflow-gpu<\/code> installation. <\/p>\n\n<p><strong>ADDENDUM 2<\/strong>\nIt looks like tensorflow-base was a part of conda because I could uninstall it with <code>conda uninstall tensorflow-base<\/code>. I still have tensorflow-gpu installation in place but I now cannot import tensorflow anymore. It says \"No module named tensorflow\". It looks like my conda environment is not seeing my tensorflor-gpu installation. I am quite confused at the moment.<\/p>\n","answers":[{"AnswerId":"52912035","CreationDate":"2018-10-21T04:06:15.747","ParentId":null,"OwnerUserId":"9328846","Title":null,"Body":"<p>@Smokrow, thank you for your answers above.\nIt appears to be the case that Keras seems to have problems with multiprocessing in Windows platforms.<\/p>\n\n<pre><code>history = merged_model.fit_generator(generator=train_generator,\n                                     epochs=60,\n                                     verbose=2,\n                                     callbacks=[reduce_lr_on_plateau],\n                                     validation_data=val_generator,\n                                     use_multiprocessing=True,\n                                     max_queue_size=50,\n                                     workers=3)\n<\/code><\/pre>\n\n<p>The piece of code above causes the Keras to hang and literally no progress is seen. If the user is runing his\/her code on Windows, use_multiprocessor needs to be set to False! Otherwise, it does not work.\nInterestingly, workers can still be set to a number that is greater than one and it still gives performance benefits. I am having difficulties to understand what really is happening in the background but it does give performance improvement. So the following piece of code made it work.<\/p>\n\n<pre><code>history = merged_model.fit_generator(generator=train_generator,\n                                     epochs=60,\n                                     verbose=2,\n                                     callbacks=[reduce_lr_on_plateau],\n                                     validation_data=val_generator,\n                                     use_multiprocessing=False,  # CHANGED\n                                     max_queue_size=50,\n                                     workers=3)\n<\/code><\/pre>\n"},{"AnswerId":"52905362","CreationDate":"2018-10-20T11:52:48.033","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>Depending on the size of your network it could be, that your CPU is loading data most of the time. <\/p>\n\n<p>Since you are using Python Generators most of your time will be spend in Python Code opening your files. The Generator is probably bottlenecking your pipeline.<\/p>\n\n<p>Once the data is loaded it is probably evaluated instantly on the GPU resulting in almost 0% GPU utilization since your Gpu keeps waiting for new data. You could try using TensorFlows  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">dataset<\/a> API. tfrecords are extremely fast in load times.  Take a look at this <a href=\"https:\/\/medium.com\/@moritzkrger\/speeding-up-keras-with-tfrecord-datasets-5464f9836c36?source=linkShare-b7d1e0266b3-1540036306\" rel=\"nofollow noreferrer\">article <\/a><\/p>\n"}]}
+{"QuestionId":52904098,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-20T09:14:00.567","AcceptedAnswerId":null,"OwnerUserId":8047895.0,"Title":"Tensorflow predicting for specific ids","Body":"<p>im using the Tensorflow dataset api and try to predict for a multilabel classification. Its working fine but the resulting predictions have no correspondending id, so I don't know to what the prediciton belongs to.<\/p>\n\n<p>I'm using the following code to create the dataset and predict:<\/p>\n\n<pre><code>def test_input_fn():\n  filenames = tf.train.match_filenames_once(\"..\/input\/test\/*_green.png\")\n  dataset = tf.data.Dataset.from_tensor_slices(filenames)\n\n  def _parse_image_data(filename):\n      image_string = tf.read_file(filename)\n      image_decoded = tf.image.decode_png(image_string, channels=1)\n      image_reshape = tf.reshape(image_decoded, [512*512*1])\n      return image_reshape\n\n  return dataset.map(_parse_image_data).batch(8)\n\npred_result = estimator.predict(test_input)\n<\/code><\/pre>\n\n<p>Is there an easy way to somehow append the ids to the prediction.<\/p>\n","answers":[{"AnswerId":"53077747","CreationDate":"2018-10-31T06:46:03.997","ParentId":null,"OwnerUserId":"10253771","Title":null,"Body":"<p>You can <code>return image_reshape,filename<\/code> in your <code>_parse_image_data(filename)<\/code> function, and modify your model function accordingly, so that for predictions it will return not only the predicted labels but also the filenames. <\/p>\n\n<p>More detailedly, suppose your model function is <code>model_fn(features, labels, mode, params)<\/code>, now your <code>features<\/code> is no longer just <code>image_reshape<\/code>, but <code>image_shape,filename<\/code>, so basically you should change wherever you use <code>features<\/code> to <code>features[0]<\/code>; you may also need to append <code>features[1]<\/code> to your <code>predictions<\/code> (I'm assuming you name your variables like this, I can give more specific codes if you can show me your model function).<\/p>\n\n<p>Or, if you would like a separate function that simply prints the <code>filenames<\/code> content :<\/p>\n\n<pre><code>def print_filenames(): \n    filenames = tf.train.match_filenames_once(\"..\/input\/test\/*_green.png\")\n    with tf.Session() as sess:\n        sess.run(tf.local_variables_initializer())\n        print(sess.run(filenames))\n<\/code><\/pre>\n\n<p>In the above codes, the output is in the as-is order, so it should be the same order of your predictions.<\/p>\n"}]}
+{"QuestionId":52904531,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-10-20T10:11:18.070","AcceptedAnswerId":null,"OwnerUserId":2117367.0,"Title":"Tensorflow Object Detection API: how to find out false positives, false negatives, true positives?","Body":"<p>I am using <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">Tensorflow Object Detection API<\/a> to finetune a pretrained model from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">the model zoo<\/a> for custom object detection. Once my model is converged I use <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/eval_util.py\" rel=\"nofollow noreferrer\"><code>eval_util.py<\/code><\/a> with <code>EvalConfig.metrics_set='open_images_V2_detection_metrics'<\/code> to obtain the <code>mAP<\/code>(and class-specific <code>AP<\/code>s) which lets me measure the quality of my model. <\/p>\n\n<p>But just <code>mAP<\/code> is not enough for my purposes. For better analysis, I want to know the exact breakdown of my model's results into false positives, false negatives and true positives. I wish to be able to see this breakdown in terms of actual test images - that is, I want to see my test images being physically divided into those three groups, automatically.<\/p>\n\n<p>How can I do that?<\/p>\n\n<p>I tried searching through Tensorflow's offical documentation and, to some extent, through the relevant python files on github, but I haven't found a way yet.<\/p>\n","answers":[{"AnswerId":"53259941","CreationDate":"2018-11-12T10:11:44.630","ParentId":null,"OwnerUserId":"10639413","Title":null,"Body":"<p>We also meet this problem. Now we find some clues in object_detection\/utils\/metrics.py. Maybe you can have a try. Hope you can share your solutions!<\/p>\n"},{"AnswerId":"52905266","CreationDate":"2018-10-20T11:41:03.670","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>I think what you are looking for is a confusion matrix. \nTake a look at this link:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/confusion_matrix\" rel=\"nofollow noreferrer\">Tensorflow Confusion Matrix<\/a><\/p>\n\n<p>You can basically evaluate your predictions with this function. <\/p>\n"}]}
+{"QuestionId":52906028,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-20T13:11:40.020","AcceptedAnswerId":52906490.0,"OwnerUserId":10005190.0,"Title":"Tensorflow body part recognition","Body":"<p>I am a developer who is really interested in ML and using Tensorflow but despite all the tutorials I have gone through, I am really struggling to start building my own model and was hoping you guys can help answer some questions that might be pretty simple.<\/p>\n\n<p>I am trying to build an algorithm that will detect specific points on a body (shoulders, wrists and others) from an image, however I am not sure what the output of the model should be, My training data is a set of images and a list of x,y values for the points and I am hoping to return a prediction for the location of the points on the image (similar to facial recognition).<\/p>\n\n<p>But I have a few questions when building this:<\/p>\n\n<p>1- What should the output of my application be? I want the algorithm to return the most likely x,y positions for each point but not sure how to represent that in tensorflow, do i need a large number of tensors in my output for each pixel and the correct pixels have a value of 1?<\/p>\n\n<p>2- How can I deal with different size images, since the input tensor will have a predetermined number of features how can my algorithm accept any size image?<\/p>\n\n<p>3- How do I determine if there are no valid points on the image? Is it just by including empty images in my training set?<\/p>\n\n<p>I hope these questions make sense and I am thankful for whoever takes the time to read it! This would be a really big help for me understanding Tensorflow and making good progress in pursing my interest in ML!<\/p>\n","answers":[{"AnswerId":"52906490","CreationDate":"2018-10-20T14:02:03.790","ParentId":null,"OwnerUserId":"10484131","Title":null,"Body":"<p>This can be done in various different ways and depends how you want to encode your data's semantics. I try to give a very simple example that may not have the best outcome:<\/p>\n\n<p>1 - You can represent a human body pose as a list of vertices, which is a vector. This could be fore example:<\/p>\n\n<pre><code>[left_shoulder,\n right_shoulder,\n torso,\n ...]\n<\/code><\/pre>\n\n<p>Each vertex is a x, y tuple so your output dimension would be <code>[vertex_count, 2]<\/code> or <code>vertex_count * 2<\/code> (flattened version). The output would therefore be these x,y positions for each vertex. This is a regression problem, because your output tries to <a href=\"https:\/\/math.stackexchange.com\/a\/141386\/606315\">generate continuous values<\/a>. You could also add a value for confidence of the recognized body pose. This would have the advantage to be able to not display\/handle a pose if none is recognized and can work with a simple cutoff when interpreting your model's output.<\/p>\n\n<p>2 - The easiest way to handle images of different size is an added preprocessing step: Simply resize the image to fit your trained model.<\/p>\n\n<p>3 - Yes this could be a good starting point. Maybe also bodies that don't fit in the frame where it's hard to label.<\/p>\n\n<p>The article on Medium shared by @EighteenthVariable is also a very good start to get a better understanding.<\/p>\n"}]}
+{"QuestionId":52906186,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-20T13:28:59.087","AcceptedAnswerId":52981472.0,"OwnerUserId":6599915.0,"Title":"Keras TPU. Compilation failure: Detected unsupported operations","Body":"<p>I try to run my keras UNet model using Google Colab TPU and I faced this problem with <code>UpSampling2D<\/code>. Any solutions or workaround?<\/p>\n\n<p>Code to run:<\/p>\n\n<pre><code>import os\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import UpSampling2D\n\nmodel = Sequential()\nmodel.add(UpSampling2D((2, 2), input_shape=(16, 16, 1)))\nmodel.compile(optimizer=tf.train.RMSPropOptimizer(learning_rate=0.01), \n              loss='binary_crossentropy', metrics=['acc'])\n\nTPU_WORKER = 'grpc:\/\/' + os.environ['COLAB_TPU_ADDR']\ntf.logging.set_verbosity(tf.logging.INFO)\n\n\nmodel = tf.contrib.tpu.keras_to_tpu_model(\n    model,strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))\n\nX = np.zeros((1024, 16, 16, 1))\nY = np.zeros((1024, 32, 32, 1))\n\nmodel.fit(X, Y, batch_size=1024)\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<blockquote>\n  <p>RuntimeError: Compilation failed: Compilation failure: Detected\n  unsupported operations when trying to compile graph\n  cluster_3_5095732716396540171[] on XLA_TPU_JIT: ResizeNearestNeighbor\n  (No registered 'ResizeNearestNeighbor' OpKernel for XLA_TPU_JIT\n  devices compatible with node {{node\n  tpu_140211339657168\/up_sampling2d_1\/ResizeNearestNeighbor}} =\n  ResizeNearestNeighbor[T=DT_FLOAT, align_corners=false,\n  _device=\"\/device:TPU_REPLICATED_CORE\"](infeed-train_1:1, tpu_140211339657168\/up_sampling2d_1\/mul)\n        .  Registered:  device='CPU'; T in [DT_DOUBLE]\n        device='CPU'; T in [DT_FLOAT]\n        device='CPU'; T in [DT_BFLOAT16]\n        device='CPU'; T in [DT_HALF]\n        device='CPU'; T in [DT_INT8]\n        device='CPU'; T in [DT_UINT8]\n        device='CPU'; T in [DT_INT16]\n        device='CPU'; T in [DT_UINT16]\n        device='CPU'; T in [DT_INT32]\n        device='CPU'; T in [DT_INT64]\n      ){{node tpu_140211339657168\/up_sampling2d_1\/ResizeNearestNeighbor}}<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52981472","CreationDate":"2018-10-25T04:41:18.247","ParentId":null,"OwnerUserId":"10199571","Title":null,"Body":"<p>From the error it looks like one of the operations in your Tensorflow backend (<code>ResizeNearestNeighbor<\/code>) graph for Keras is currently not compatible with TPUs. There are a small number of Tensorflow ops that are currently not available for TPUs (<a href=\"https:\/\/cloud.google.com\/tpu\/docs\/faq\" rel=\"nofollow noreferrer\">Cloud TPU FAQs<\/a>). <\/p>\n\n<p>You can check out the current list of TPU compatible Tensorflow ops <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tensorflow-ops\" rel=\"nofollow noreferrer\">here<\/a>. You can also use Tensorboard to see <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/cloud-tpu-tools#graphs\" rel=\"nofollow noreferrer\">TPU Compatibility Graphs<\/a>.<\/p>\n\n<p>As a workaround, you can try combine the TPU compatible Tensorflow ops to replicate the behavior of <code>ResizeNearestNeighbor<\/code>. In particular, you may be interested in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/resize-bilinear\" rel=\"nofollow noreferrer\">ResizeBilinear<\/a> Op, which is TPU compatible.<\/p>\n"}]}
+{"QuestionId":52906404,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-20T13:52:40.787","AcceptedAnswerId":null,"OwnerUserId":5483457.0,"Title":"Implementation of Embedded Layer for CNN","Body":"<p>Hi there I have been trying to follow the 2 examples:<\/p>\n\n<p>1)<a href=\"https:\/\/towardsdatascience.com\/how-to-do-text-classification-using-tensorflow-word-embeddings-and-cnn-edae13b3e575\" rel=\"nofollow noreferrer\"> How to do text classification with CNNs, TensorFlow and word embedding<\/a><\/p>\n\n<p>2) <a href=\"https:\/\/stackoverflow.com\/questions\/52168247\/how-to-build-an-embedding-layer-in-tensorflow-rnn\">How to build an embedding layer in Tensorflow RNN?<\/a><\/p>\n\n<p>However, I have been running into following error for the example 1,<\/p>\n\n<pre><code>Input 0 of layer conv2d_1 is incompatible with the layer: expected ndim=4, found ndim=3. Full shape received: [None, 100, 20]\n<\/code><\/pre>\n\n<p>and the following error for example 2. <\/p>\n\n<pre><code>ValueError: Layer conv2d_1 expects 1 inputs, but it received 100 input tensors. Inputs received: [&lt;tf.Tensor 'unstack:0' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:1' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:2' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:3' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:4' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:5' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:6' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:7' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:8' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:9' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:10' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:11' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:12' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:13' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:14' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:15' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:16' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:17' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:18' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:19' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:20' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:21' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:22' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:23' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:24' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:25' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:26' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:27' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:28' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:29' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:30' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:31' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:32' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:33' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:34' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:35' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:36' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:37' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:38' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:39' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:40' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:41' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:42' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:43' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:44' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:45' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:46' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:47' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:48' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:49' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:50' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:51' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:52' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:53' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:54' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:55' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:56' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:57' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:58' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:59' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:60' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:61' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:62' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:63' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:64' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:65' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:66' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:67' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:68' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:69' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:70' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:71' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:72' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:73' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:74' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:75' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:76' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:77' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:78' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:79' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:80' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:81' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:82' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:83' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:84' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:85' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:86' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:87' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:88' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:89' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:90' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:91' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:92' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:93' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:94' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:95' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:96' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:97' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:98' shape=(?, 20) dtype=float32&gt;, &lt;tf.Tensor 'unstack:99' shape=(?, 20) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>I have change the code instead of using <code>word_list = tf.unstack(word_vectors, axis=1)<\/code> I have changed to the following implementation to get around the errors but I do not know if the implementation is correct.<\/p>\n\n<pre><code>word_vectors = tf.contrib.layers.embed_sequence(\n    x, vocab_size=n_words, embed_dim=EMBEDDING_SIZE)\n\nword_list = tf.reshape( word_vectors, [-1, MAX_DOCUMENT_LENGTH, EMBEDDING_SIZE, 1])\n\nconv1, pool1 = cnn(word_list, FILTER_SHAPE1, 'CNN_Layer1')\n<\/code><\/pre>\n\n<p>The function <code>cnn<\/code> and the other parameters are as defined<\/p>\n\n<pre><code>MAX_DOCUMENT_LENGTH = 100\nN_FILTERS = 10\nFILTER_SHAPE1 = [20, 20]\nPOOLING_WINDOW = 4\nPOOLING_STRIDE = 2\nMAX_LABEL = 15\nEMBEDDING_SIZE = 20\n\ndef cnn(x, filter_shape, name):\n    with tf.variable_scope(name):\n        conv1 = tf.layers.conv2d(\n            x,\n            filters=N_FILTERS,\n            kernel_size=filter_shape,\n            padding='VALID',\n            activation=tf.nn.relu)\n        pool1 = tf.layers.max_pooling2d(\n            conv1,\n            pool_size=POOLING_WINDOW,\n            strides=POOLING_STRIDE,\n            padding='SAME')\n\n    return conv1, pool1\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52906669,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><tensorflow-estimator>","CreationDate":"2018-10-20T14:23:56.540","AcceptedAnswerId":52907489.0,"OwnerUserId":10307998.0,"Title":"Why softmax cross entropy loss never gives a value of zero in tensorflow?","Body":"<p>Im doing a neural network in tensorflow and Im using softmax_cross_entropy to calculate the loss, I'm doing tests and note that it never gives a value of zero, even if I compare the same values, this is my code<\/p>\n\n<pre><code>labels=[1,0,1,1]\n\n\nwith tf.Session() as sess:\n    onehot_labels=tf.one_hot(indices=labels,depth=2)\n    logits=[[0.,1.],[1.,0.],[0.,1.],[0.,1.]]\n    print(sess.run(onehot_labels))\n    loss=tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,logits=logits)\n    print(sess.run(loss))\n<\/code><\/pre>\n\n<p>I obtain this<\/p>\n\n<pre><code>[[0. 1.]\n [1. 0.]\n [0. 1.]\n [0. 1.]]\n0.31326166\n<\/code><\/pre>\n\n<p>Why is not zero??<\/p>\n","answers":[{"AnswerId":"52907489","CreationDate":"2018-10-20T15:57:01.293","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Matias's post is correct. The following code gives the same result as your code<\/p>\n\n<pre><code>labels=[1,0,1,1]\n\nwith tf.Session() as sess:\n    onehot_labels=tf.one_hot(indices=labels,depth=2)\n    logits=[[0.,1.],[1.,0.],[0.,1.],[0.,1.]]\n    print(sess.run(onehot_labels))\n\n    probabilities = tf.nn.softmax(logits=logits)\n    # cross entropy\n    loss = -tf.reduce_sum(onehot_labels * tf.log(probabilities)) \/ len(labels)\n\n    print(sess.run(loss))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52907059,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-20T15:09:33.903","AcceptedAnswerId":52907572.0,"OwnerUserId":10533365.0,"Title":"'Model' object has no attribute '_name'","Body":"<p>I'm making a CNN in Keras. But I have a problem in making Keras model. Here's my code:<\/p>\n\n<pre><code>x = Input(shape=(256,256,1))\nfor i in range(16):\n    u = int(16 * 2 ** (i\/\/4))\n    x = BatchNormalization()(x)\n    x1 = Conv2D(u, kernel_size=(1,1), strides=(1,1), activation='relu')(x)\n    x1 = MaxPooling2D(pool_size=(3,3), strides=(1,1))(x1)\n    x2 = Conv2D(u, kernel_size=(2,2), strides=(1,1), activation='relu')(x)\n    x2 = MaxPooling2D(pool_size=(2,2), strides=(1,1))(x2)\n    x3 = Conv2D(u, kernel_size=(3,3), strides=(1,1), activation='relu')(x)\n    x3 = MaxPooling2D(pool_size=(1,1), strides=(1,1))(x3)\n    x = multiply([x1, x2, x3])\n    #x = Dropout(0.45)(x)\n    x = MaxPooling2D(pool_size=(3,3), strides=(1,1))(x)\nout = BatchNormalization()(x)\nmodel = tf.keras.models.Model(inputs=x, outputs=out)\n<\/code><\/pre>\n\n<p>and I got the following error:<\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-99-630b3ef0b15f&gt; in &lt;module&gt;()\n     13     x = MaxPooling2D(pool_size=(3,3), strides=(1,1))(x)\n     14 out = BatchNormalization()(x)\n---&gt; 15 model = tf.keras.models.Model(inputs=x, outputs=out)\n...\n\nAttributeError: 'Model' object has no attribute '_name'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52907572","CreationDate":"2018-10-20T16:08:14.133","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are assigning other Tensors to <code>x<\/code> after defining it as Input tensor. Therefore, it can't be used as the input of the model, i.e. <code>inputs=x<\/code>. To resolve this with minimal modifications simply store <code>x<\/code> in another variable after defining it as Input tensor:<\/p>\n\n<pre><code>x = Input(shape=(256,256,1))\ninp = x\n\n# the rest is the same...\n\nmodel = tf.keras.models.Model(inputs=inp, outputs=out) # pass `inp` as inputs\n<\/code><\/pre>\n"}]}
+{"QuestionId":52907362,"AnswerCount":0,"Tags":"<tensorflow><out-of-memory><tensorflow-estimator>","CreationDate":"2018-10-20T15:41:41.713","AcceptedAnswerId":null,"OwnerUserId":10277386.0,"Title":"tf.train.Saver instance does not close checkpoint files","Body":"<p>In my effort to obtaining the best checkpoint possible following the fitting stage for my model using the Estimator API, I ended up using a tf.train.SessionRunHook where I reference the tf.train.Saver instance created by the estimator to manually save the best checkpoint found so far in the fitting process, to a separate directory. My intention is to let the estimator instance follow it's course when it comes to doing the checkpoints, but if I find a better set of weights then checkpoint the model to a specific directory.<br>\nHere is a code snippet to illustrate what i have in mind:  <\/p>\n\n<pre><code>class SessRunHOOK(tf.train.SessionRunHook):\n     ...  \n     def end(self, session):\n         ...\n         # after each call to estimator.evaluate() if\n         # the metrics for this model are the best so far\n         # save the checkpoint to a specific directory like so\n         saver = session.graph.get_collection('savers')[0]\n         saver.save(session,\n                    join(best_ckpt_dir, 'ckpt'),\n                    global_step = global_step,\n                    latest_filename = None,\n                    meta_graph_suffix = 'meta',\n                    write_meta_graph = True,\n                    write_state = True,\n                    strip_default_attrs = False)\n<\/code><\/pre>\n\n<p>In the model function for the estimator, the hook instance is passed to both train and evaluation hook arguments for the EstimatorSpec.<\/p>\n\n<p>The Estimator instance is created and used like so:<\/p>\n\n<pre><code>RUN_CONFIG = tf.estimator.RunConfig(\n\n                    save_summary_steps     = None,\n                    log_step_count_steps   = None,\n                    keep_checkpoint_max    = 3,\n                    save_checkpoints_steps = TRAIN_STEPS,\n                    session_config         = None\n\n                )\nestimator = tf.estimator.Estimator(\n\n   model_fn  = model_fn,\n   params    = ESTIMATOR_PARAMS,\n   model_dir = DUMP_DIR,\n   config    = RUN_CONFIG \n\n)\n\nwhile True:\n\n   estimator.train(train_input_fn, train_steps = TRAIN_STEPS)\n   estimator.evaluate(eval_input_fn)\n<\/code><\/pre>\n\n<p>What ends up happening is that even tho the estimator instance keeps at most 3  checkpoints in the model_dir as specified in the run config, alongside them it will pile up temporary metafiles. Here is a screenshot of the model_dir with these temp files <a href=\"https:\/\/i.stack.imgur.com\/nKhlc.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nKhlc.jpg\" alt=\"\"><\/a><\/p>\n\n<p>The only way to get rid of them is to reboot the system, refreshing the file explorer GUI will not make any difference, leading me to think that there is something wrong with the way I am doing the checkpoints and I would appreciate if someone could point that out.<\/p>\n\n<p>Eventually it will end up using all the memory on the system yielding the following stack trace:  <\/p>\n\n<p>2018-10-21 04:07:10.571642: W tensorflow\/core\/framework\/op_kernel.cc:1275] OP_REQUIRES failed at save_restore_v2_ops.cc:184 : Resource exhausted: \/home\/m232\/catalinh\/models\/orange_juice\/dtt\/keurig_new_logo_160x160x1_with_negatives\/ckpts\/model.ckpt-19494.data-00000-of-00001; Too many open files\nTraceback (most recent call last):\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1278, in _do_call\n    return fn(*args)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1263, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1350, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.ResourceExhaustedError: \/home\/m232\/catalinh\/models\/orange_juice\/dtt\/keurig_new_logo_160x160x1_with_negatives\/ckpts\/model.ckpt-19494.data-00000-of-00001; Too many open files\n     [[Node: save\/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_save\/Const_0_0, save\/RestoreV2\/tensor_names, save\/RestoreV2\/shape_and_slices)]]\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n\n<pre><code> [[Node: save\/RestoreV2\/_149 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_154_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\/home\/m232\/.virtualenvs\/ml\/bin\/tensor_dyve\", line 11, in \n    sys.exit(main())\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/TensorDyve\/tensor_dyve.py\", line 352, in main\n    estimator.train( input_fn = train_input_pipe, steps = TRAIN_STEPS )\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 376, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1145, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1173, in _train_model_default\n    saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1448, in _train_with_estimator_spec\n    log_step_count_steps=self._config.log_step_count_steps) as mon_sess:\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 421, in MonitoredTrainingSession\n    stop_grace_period_secs=stop_grace_period_secs)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 832, in <strong>init<\/strong>\n    stop_grace_period_secs=stop_grace_period_secs)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 555, in <strong>init<\/strong>\n    self._sess = _RecoverableSession(self._coordinated_creator)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1018, in <strong>init<\/strong>\n    _WrappedSession.<strong>init<\/strong>(self, self._create_session())\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1023, in _create_session\n    return self._sess_creator.create_session()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 712, in create_session\n    self.tf_sess = self._session_creator.create_session()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 483, in create_session\n    init_fn=self._scaffold.init_fn)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/session_manager.py\", line 281, in prepare_session\n    config=config)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/session_manager.py\", line 211, in _restore_checkpoint\n    saver.restore(sess, ckpt.model_checkpoint_path)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1725, in restore\n    {self.saver_def.filename_tensor_name: save_path})\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 877, in run\n    run_metadata_ptr)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1100, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1272, in _do_run\n    run_metadata)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1291, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.ResourceExhaustedError: \/home\/m232\/catalinh\/models\/orange_juice\/dtt\/keurig_new_logo_160x160x1_with_negatives\/ckpts\/model.ckpt-19494.data-00000-of-00001; Too many open files\n     [[Node: save\/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_save\/Const_0_0, save\/RestoreV2\/tensor_names, save\/RestoreV2\/shape_and_slices)]]\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n\n<pre><code> [[Node: save\/RestoreV2\/_149 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_154_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n\n<p>Caused by op 'save\/RestoreV2', defined at:\n  File \"\/home\/m232\/.virtualenvs\/ml\/bin\/tensor_dyve\", line 11, in \n    sys.exit(main())\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/TensorDyve\/tensor_dyve.py\", line 352, in main\n    estimator.train( input_fn = train_input_pipe, steps = TRAIN_STEPS )\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 376, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1145, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1173, in _train_model_default\n    saving_listeners)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1448, in _train_with_estimator_spec\n    log_step_count_steps=self._config.log_step_count_steps) as mon_sess:\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 421, in MonitoredTrainingSession\n    stop_grace_period_secs=stop_grace_period_secs)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 832, in <strong>init<\/strong>\n    stop_grace_period_secs=stop_grace_period_secs)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 555, in <strong>init<\/strong>\n    self._sess = _RecoverableSession(self._coordinated_creator)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1018, in <strong>init<\/strong>\n    _WrappedSession.<strong>init<\/strong>(self, self._create_session())\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1023, in _create_session\n    return self._sess_creator.create_session()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 712, in create_session\n    self.tf_sess = self._session_creator.create_session()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 474, in create_session\n    self._scaffold.finalize()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 214, in finalize\n    self._saver.build()\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1293, in build\n    self._build(self._filename, build_save=True, build_restore=True)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1330, in _build\n    build_save=build_save, build_restore=build_restore)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 772, in _build_internal\n    restore_sequentially, reshape)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 450, in _AddShardedRestoreOps\n    name=\"restore_shard\"))\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 397, in _AddRestoreOps\n    restore_sequentially)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 829, in bulk_restore\n    return io_ops.restore_v2(filename_tensor, names, slices, dtypes)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_io_ops.py\", line 1463, in restore_v2\n    shape_and_slices=shape_and_slices, dtypes=dtypes, name=name)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 454, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3155, in create_op\n    op_def=op_def)\n  File \"\/home\/m232\/.virtualenvs\/ml\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1717, in <strong>init<\/strong>\n    self._traceback = tf_stack.extract_stack()<\/p>\n\n<p>ResourceExhaustedError (see above for traceback): \/home\/m232\/catalinh\/models\/orange_juice\/dtt\/keurig_new_logo_160x160x1_with_negatives\/ckpts\/model.ckpt-19494.data-00000-of-00001; Too many open files\n     [[Node: save\/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_save\/Const_0_0, save\/RestoreV2\/tensor_names, save\/RestoreV2\/shape_and_slices)]]\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n\n<pre><code> [[Node: save\/RestoreV2\/_149 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_154_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.<\/p>\n","answers":[]}
+{"QuestionId":52907593,"AnswerCount":1,"Tags":"<python><tensorflow><typeerror>","CreationDate":"2018-10-20T16:10:11.023","AcceptedAnswerId":52957355.0,"OwnerUserId":5865579.0,"Title":"TypeError: unsupported operand type(s) for \/: 'Dimension' and 'float' in TensorFlow","Body":"<p>I have an code for implementing in deep learning in <code>Tensorflow<\/code>. I use the <code>Keras<\/code> module<\/p>\n\n<pre><code>self.n_clusters = 10\nself.alpha = 0.01\nclustering_layer = ClusteringLayer(self.n_clusters, alpha=self.alpha, name='clustering')(hidden)\n<\/code><\/pre>\n\n<p>my error is mainly from above, so I on attach it.<\/p>\n\n<p>It give me the following error:<\/p>\n\n<pre><code>--&gt; 118         clustering_layer = ClusteringLayer(self.n_clusters, alpha=self.alpha, name='clustering')(hidden)\n    119         self.model = Model(inputs=self.autoencoder.input, outputs=clustering_layer)\n    120 \n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.pyc in __call__(self, inputs, *args, **kwargs)\n    694         if all(hasattr(x, 'get_shape') for x in input_list):\n    695           input_shapes = nest.map_structure(lambda x: x.get_shape(), inputs)\n--&gt; 696         self.build(input_shapes)\n    697 \n    698       # Check input assumptions set after layer building, e.g. input shape.\n\n&lt;ipython-input-13-8890754cc8a3&gt; in build(self, input_shape)\n     73         input_dim = input_shape[1]\n     74         self.input_spec = InputSpec(dtype=K.floatx(), shape=(None, input_dim))\n---&gt; 75         self.clusters = self.add_weight(shape=(self.n_clusters, input_dim), initializer='glorot_uniform', name='clusters')\n     76         if self.initial_weights is not None:\n     77             self.set_weights(self.initial_weights)\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.pyc in add_weight(self, name, shape, dtype, initializer, regularizer, trainable, constraint, partitioner, use_resource, getter)\n    532         trainable=trainable and self.trainable,\n    533         partitioner=partitioner,\n--&gt; 534         use_resource=use_resource)\n    535 \n    536     if regularizer is not None:\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/checkpointable\/base.pyc in _add_variable_with_custom_getter(self, name, shape, dtype, initializer, getter, overwrite, **kwargs_for_getter)\n    495     new_variable = getter(\n    496         name=name, shape=shape, dtype=dtype, initializer=initializer,\n--&gt; 497         **kwargs_for_getter)\n    498 \n    499     # If we set an initializer and the variable processed it, tracking will not\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.pyc in make_variable(name, shape, dtype, initializer, partition_info, trainable, caching_device, validate_shape, constraint, use_resource, partitioner)\n   1871       validate_shape=validate_shape,\n   1872       constraint=constraint,\n-&gt; 1873       use_resource=use_resource)\n   1874   return v\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.pyc in variable(initial_value, trainable, collections, validate_shape, caching_device, name, dtype, constraint, use_resource)\n   2232                          name=name, dtype=dtype,\n   2233                          constraint=constraint,\n-&gt; 2234                          use_resource=use_resource)\n   2235 \n   2236 \n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.pyc in &lt;lambda&gt;(**kwargs)\n   2222              constraint=None,\n   2223              use_resource=None):\n-&gt; 2224   previous_getter = lambda **kwargs: default_variable_creator(None, **kwargs)\n   2225   for getter in ops.get_default_graph()._variable_creator_stack:  # pylint: disable=protected-access\n   2226     previous_getter = _make_getter(getter, previous_getter)\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.pyc in default_variable_creator(next_creator, **kwargs)\n   2194         collections=collections, validate_shape=validate_shape,\n   2195         caching_device=caching_device, name=name, dtype=dtype,\n-&gt; 2196         constraint=constraint)\n   2197   elif not use_resource and context.executing_eagerly():\n   2198     raise RuntimeError(\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/resource_variable_ops.pyc in __init__(self, initial_value, trainable, collections, validate_shape, caching_device, name, dtype, variable_def, import_scope, constraint)\n    310           name=name,\n    311           dtype=dtype,\n--&gt; 312           constraint=constraint)\n    313 \n    314   # pylint: disable=unused-argument\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/resource_variable_ops.pyc in _init_from_args(self, initial_value, trainable, collections, validate_shape, caching_device, name, dtype, constraint)\n    415               with ops.name_scope(\"Initializer\"), ops.device(None):\n    416                 initial_value = ops.convert_to_tensor(\n--&gt; 417                     initial_value(), name=\"initial_value\", dtype=dtype)\n    418               self._handle = _eager_safe_variable_handle(\n    419                   shape=initial_value.get_shape(),\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.pyc in &lt;lambda&gt;()\n   1858         initializer = initializer(dtype=dtype)\n   1859       init_val = lambda: initializer(  # pylint: disable=g-long-lambda\n-&gt; 1860           shape, dtype=dtype, partition_info=partition_info)\n   1861       variable_dtype = dtype.base_dtype\n   1862   if use_resource is None:\n\n\/usr\/local\/python\/2.7-conda5.2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/init_ops.pyc in __call__(self, shape, dtype, partition_info)\n    466       scale \/= max(1., fan_out)\n    467     else:\n--&gt; 468       scale \/= max(1., (fan_in + fan_out) \/ 2.)\n    469     if self.distribution == \"normal\":\n    470       stddev = math.sqrt(scale)\n\nTypeError: unsupported operand type(s) for \/: 'Dimension' and 'float'\n<\/code><\/pre>\n\n<p>line 118 is my code location. The error seems to occur in tensorflow package. it give me the <code>TypeError: unsupported operand type(s) for \/: 'Dimension' and 'float'<\/code>. I try both <code>python 2.7<\/code> and <code>python 3.6<\/code> but with same problem. <\/p>\n\n<p>How to deal with this situation? <\/p>\n\n<p>An very similar situation is in <a href=\"https:\/\/github.com\/DrSleep\/tensorflow-deeplab-resnet\/issues\/83\" rel=\"nofollow noreferrer\">github<\/a>, its error can be addressed in its code, but my error seems to be happen in  <code>init_ops.pyc<\/code><\/p>\n","answers":[{"AnswerId":"52957355","CreationDate":"2018-10-23T20:26:44.950","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The issue here seems to be that ClusteringLayer is passing a tf.Dimension object as the number of clusters, which is used to initialize the weights. Do int(self.n_clusters) instead to bypass the dimension object issues.<\/p>\n"}]}
+{"QuestionId":52907605,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-10-20T16:11:40.553","AcceptedAnswerId":52907652.0,"OwnerUserId":8821185.0,"Title":"Putting sess.run output into a list","Body":"<p>So my code has the line in it and it works:<\/p>\n\n<pre><code>    _, loss_value0, loss_value1, loss_value2, loss_value3 = sess.run([train_op, loss0, loss1, loss2, loss3])\n<\/code><\/pre>\n\n<p>I wanted to make this a list of losses so I ran:<\/p>\n\n<pre><code>    loss_value_list = []\n    _, loss_value_list = sess.run([train_op] + loss_list)\n<\/code><\/pre>\n\n<p>But that doesn't work. Probably I can't assign a list like that? Is there someway to do that?<\/p>\n\n<p>Thanks for any help!<\/p>\n","answers":[{"AnswerId":"52907652","CreationDate":"2018-10-20T16:15:51.783","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p>No need to pre define the <code>loss_value_list<\/code>; Use <code>*<\/code> operator to unpack the result into a list variable instead:<\/p>\n\n<pre><code>lst = ['a', 1, 2, 3, 4]\n_, *loss_value_list = lst\n\nloss_value_list\n# [1, 2, 3, 4]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52907741,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><deep-learning><tensorboard>","CreationDate":"2018-10-20T16:26:57.137","AcceptedAnswerId":null,"OwnerUserId":8954691.0,"Title":"Can't see dropout in Tensorboard graph","Body":"<p>I am applying dropout to the first layer in the following way : <\/p>\n\n<pre><code>out = layers.fully_connected(X, num_outputs=hidden[0], activation_fn=None)\nout1 = tf.nn.relu(out)\ndrop_out = tf.nn.dropout(out1, keep_prob) #keep_prob = 0.5\n\nout2 = layers.fully_connected(drop_out, num_outputs=hidden[1], activation_fn=None)\nout2 = tf.nn.relu(out2)\n\nout3 = layers.fully_connected(out4, num_outputs=num_actions, activation_fn=None)\n<\/code><\/pre>\n\n<p>But I cannot see dropout in tensorboard computational graph. Do I need to explicitly write a command in the summary to include Dropout layer ? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eBwPf.png\" rel=\"nofollow noreferrer\">Tensorboard graph for the above model<\/a><\/p>\n","answers":[]}
+{"QuestionId":52908188,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><neural-network><quantization>","CreationDate":"2018-10-20T17:18:16.317","AcceptedAnswerId":null,"OwnerUserId":9104473.0,"Title":"Build and Train Quantized Neural Network using numpy","Body":"<p>I am new to quantization and would like to learn how to build and train quantized neural network using numpy only.<\/p>\n\n<p>Would be really grateful if anyone could point me in the right direction.<\/p>\n\n<p>I would like to learn it using numpy as I will have a better understanding of the how the network behaves and can play with intermediate values if necessary.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":52908388,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-20T17:39:09.400","AcceptedAnswerId":null,"OwnerUserId":10533804.0,"Title":"How to train a convolution neural network with my own dataset for image clasification and save it as .pb.Everything in tensorflow and python","Body":"<p>I want to learn how to train a convolution neural network in tensorflow.May you write a link with some code.For example a github project.After I train the convolution neural network with my own dataset ,how to save it as .pb or .pbtxt.May you discribe the procedure to do that?<\/p>\n","answers":[{"AnswerId":"52910887","CreationDate":"2018-10-20T23:21:32.950","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>Well not sure if you want code you can understand or just a model. If you want just the model check out this <a href=\"https:\/\/github.com\/tensorflow\/hub\/tree\/master\/examples\/image_retraining\" rel=\"nofollow noreferrer\">retraining code of tensorflow<\/a>. For full understanding, this <a href=\"https:\/\/www.kaggle.com\/pouryaayria\/convolutional-neural-networks-tutorial-tensorflow\" rel=\"nofollow noreferrer\">link<\/a> might be useful. The pb file you're referring to is also called a saved model, see this <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#simple_save\" rel=\"nofollow noreferrer\">link<\/a>.<\/p>\n"}]}
+{"QuestionId":52908537,"AnswerCount":1,"Tags":"<python><machine-learning><keras><time-series><lstm>","CreationDate":"2018-10-20T17:56:05.767","AcceptedAnswerId":null,"OwnerUserId":10533802.0,"Title":"Error when checking target: expected dense_2 to have 2 dimensions, but got array with shape (1, 1226, 2)","Body":"<p>Here is the code that I am trying to run: <\/p>\n\n<pre><code>y = Df[['label']]\n\ny_train = np.column_stack((y_train['label']))\n\ny_test = np.column_stack((y_test['label']))\n\ndata_dim = 18\ntimesteps = 1\nnum_classes = 2\n\nmodel = Sequential()\n\nmodel.add(LSTM(19, return_sequences=True,\n               input_shape=(timesteps, data_dim)))  \nmodel.add(LSTM(19, return_sequences=True))  \nmodel.add(LSTM(19))  # return a single vector of dimension 30\nmodel.add(Dense(1, activation='softmax'))\n\nmodel.compile(loss='binary_crossentropy',\n              optimizer='rmsprop',\n              metrics=['accuracy'])\n\nmodel.summary()\n\nmodel.fit(X_train, y_train, batch_size = 400, epochs = 20, verbose = \n\naccuracy_train = accuracy_score(y_train, model.predict(X_train))\naccuracy_test = accuracy_score(y_test, model.predict(X_test))\nprint('\\nTrain Accuracy:{: .2f}%'.format(accuracy_train*100))\nprint('Test Accuracy:{: .2f}%'.format(accuracy_test*100))\n<\/code><\/pre>\n\n<p>I am getting the following error: <\/p>\n\n<pre><code>Error when checking target: expected dense_1 to have shape (1,) but got array with shape (1226,)\n<\/code><\/pre>\n\n<p>Model Summary:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlstm_37 (LSTM)               (None, 1, 19)             2888      \n_________________________________________________________________\nlstm_38 (LSTM)               (None, 1, 19)             2964      \n_________________________________________________________________\nlstm_39 (LSTM)               (None, 19)                2964      \n_________________________________________________________________\ndense_13 (Dense)             (None, 1)                 20        \n=================================================================\nTotal params: 8,836\nTrainable params: 8,836\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p><code>X_train<\/code>: <\/p>\n\n<pre><code>[[[-0.02444971  0.02444971  1.          1.          1.\n    1.         -1.          2.          3.          4.\n    3.          1.          0.          0.06938705 -0.04329106\n    0.02458854  0.06025883  0.01439807]]]\n[[[ 0.00733477  0.02033006 -1.          1.          1.\n    1.         -1.          0.          1.          2.\n    1.          0.          0.          0.03837079 -0.00829683\n   -0.00734757  0.00985466 -0.04543226]]]\n<\/code><\/pre>\n\n<p><code>y_train<\/code>: <\/p>\n\n<pre><code>[[ 1  1 -1 ...  1  1 -1]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"52908763","CreationDate":"2018-10-20T18:21:37.293","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>It seems you are doing binary classification. Therefore, there are a few things that need to be fixed:<\/p>\n\n<ol>\n<li><p>The <code>y_train<\/code> should consists of zeros and ones. The following would convert all the <code>-1<\/code> labels to zeros:<\/p>\n\n<pre><code>y_train = (y_train == 1).astype('float32')\n<\/code><\/pre><\/li>\n<li><p>Reshape the labels to have a shape of <code>(n_samples, 1)<\/code>:<\/p>\n\n<pre><code>y_train = y_train.reshape(-1, 1)\n<\/code><\/pre><\/li>\n<li><p>Use <code>'sigmoid'<\/code> as the activation of last layer:<\/p>\n\n<pre><code>model.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
+{"QuestionId":52908567,"AnswerCount":1,"Tags":"<keras><deep-learning><h5py>","CreationDate":"2018-10-20T17:59:15.173","AcceptedAnswerId":52911466.0,"OwnerUserId":9326936.0,"Title":"Saving Inception_Resnet_v2 Features in h5py","Body":"<p>I am currently trying to use a number of pre-trained ImageNet networks in keras to extract features from images. To do this I am removing the top layer of the network, pre-processing the input according to each of the networks requirements and then saving the outputs in an hdf5 file. I have used several other pre-trained networks using this exact same approach and code (only switching out the network) and it seems to work perfectly. However, the network I am struggling to get to work is 'InceptionResNetV2'. I don't believe I am getting any problems with the network, only with the saving - I've attached a slightly simplified version of the code. Changing the model in the feature extractor and the model in pre-processing means it works perfectly - for vgg16, vgg19, resnet, inception etc - all good. <\/p>\n\n<pre><code>db = h5py.File(hdf5_path, mode=\"w\")\n\nfeaturesDB = db.create_dataset(\"features\", shape=features_shape, dtype=\"float\")\n\nimages = [cv2.imread(path, 1) for path in image_paths[start:end]]\n\nimages = inception_resnet_v2.preprocess_input(images)\n\nfeatures = feature_extractor.extract(images)\n\nfeaturesDB[start:end] = features\n<\/code><\/pre>\n\n<p>However, this produces the following error. I have tried to change the dtype of the data going into featuresDB to an int, but that had no effect. Any suggestions appreciated!  <\/p>\n\n<pre><code>File \"h5py\/utils.pyx\", line 101, in h5py.utils.convert_tuple\nTypeError: an integer is required\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"Extract_Features.py\", line 111, in &lt;module&gt;\n    extract_features(image_paths, hdf5_path=args[\"features_db\"], \nfeature_extractor=feature_extractor)\n\nFile \"Extract_Features.py\", line 83, in extract_features\n    featuresDB = db.create_dataset(\"features\", shape=features_shape, dtype=\"float\")\n\nFile \"\/home\/will\/tensorflow\/lib\/python3.5\/site-packages\/h5py\/_hl\/group.py\", line \n106, in create_dataset\n    dsid = dataset.make_new_dset(self, shape, dtype, data, **kwds)\n\nFile \"\/home\/will\/tensorflow\/lib\/python3.5\/site-packages\/h5py\/_hl\/dataset.py\", line 137, in make_new_dset\n    sid = h5s.create_simple(shape, maxshape)\n\nFile \"h5py\/_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\n\nFile \"h5py\/_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\n\nFile \"h5py\/h5s.pyx\", line 95, in h5py.h5s.create_simple\n\nFile \"h5py\/utils.pyx\", line 103, in h5py.utils.convert_tuple\nTypeError: Can't convert element 1 (None) to hsize_t\n<\/code><\/pre>\n","answers":[{"AnswerId":"52911466","CreationDate":"2018-10-21T01:36:44.740","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<pre><code>In [201]: f = h5py.File('test.h5','w')\n<\/code><\/pre>\n\n<p>I can recreate your error with this expression:<\/p>\n\n<pre><code>In [203]: ds = f.create_dataset('features', shape=(None,3), dtype=float)\n---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\nh5py\/utils.pyx in h5py.utils.convert_tuple()\n\nTypeError: an integer is required\n\nDuring handling of the above exception, another exception occurred:\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-203-a5733d841c5c&gt; in &lt;module&gt;()\n----&gt; 1 ds = f.create_dataset('features', shape=(None,3), dtype=float)\n\n~\/.local\/lib\/python3.6\/site-packages\/h5py\/_hl\/group.py in create_dataset(self, name, shape, dtype, data, **kwds)\n    104         \"\"\"\n    105         with phil:\n--&gt; 106             dsid = dataset.make_new_dset(self, shape, dtype, data, **kwds)\n    107             dset = dataset.Dataset(dsid)\n    108             if name is not None:\n\n~\/.local\/lib\/python3.6\/site-packages\/h5py\/_hl\/dataset.py in make_new_dset(parent, shape, dtype, data, chunks, compression, shuffle, fletcher32, maxshape, compression_opts, fillvalue, scaleoffset, track_times)\n    135         sid = h5s.create(h5s.NULL)\n    136     else:\n--&gt; 137         sid = h5s.create_simple(shape, maxshape)\n    138 \n    139 \n\nh5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\nh5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\nh5py\/h5s.pyx in h5py.h5s.create_simple()\n\nh5py\/utils.pyx in h5py.utils.convert_tuple()\n\nTypeError: Can't convert element 0 (None) to hsize_t\n<\/code><\/pre>\n\n<p>A <code>None<\/code> is ok in the <code>maxshape<\/code> parameter, but not the <code>shape<\/code> - this is for a resizeable dataset:<\/p>\n\n<pre><code>In [204]: ds = f.create_dataset('features', shape=(10,3), maxshape=(None,3), dty\n     ...: pe=float)\nIn [205]: ds\nOut[205]: &lt;HDF5 dataset \"features\": shape (10, 3), type \"&lt;f8\"&gt;\nIn [206]: ds.resize((20,3))\nIn [207]: ds\nOut[207]: &lt;HDF5 dataset \"features\": shape (20, 3), type \"&lt;f8\"&gt;\n<\/code><\/pre>\n\n<p>I haven't used <code>keras<\/code>, but from other SO questions, and documentation, <code>None<\/code> is allowed in its <code>shape<\/code>.  <code>None<\/code> is not allowed in <code>numpy<\/code> array shape.<\/p>\n"}]}
+{"QuestionId":52908789,"AnswerCount":0,"Tags":"<deep-learning><pytorch><loss-function>","CreationDate":"2018-10-20T18:24:09.960","AcceptedAnswerId":null,"OwnerUserId":10480830.0,"Title":"How to create a custom loss function in Pytorch","Body":"<p>I\u2019m trying to train a network by my own loss function.\nI can train a network with loss functions are included in the PyTorch.<\/p>\n\n<p>But, I meet the challenge when I am trying to define my own loss function.\nIndeed, I need to a correct example to train a network by custom loss function in details<\/p>\n\n<p>My loss function is:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/OIRtX.jpg\" alt=\"enter image description here\"><\/p>\n","answers":[]}
+{"QuestionId":52908944,"AnswerCount":3,"Tags":"<python><tensorflow><keras><deep-learning><google-colaboratory>","CreationDate":"2018-10-20T18:40:52.690","AcceptedAnswerId":52957781.0,"OwnerUserId":10533846.0,"Title":"\"KeyError: 0\" when trying to load a sequential model in Keras","Body":"<p>I have trained a Sequential model in Keras using Google Colab to work on the Google Doodle data set. I am doing some simple image classification over here.<\/p>\n\n<p>The following function defines the architecture of my model:<\/p>\n\n<pre><code>def create_model(input_shape):\n   model = keras.Sequential()  \n   model.add(layers.Conv2D(16, (3, 3), padding = 'same', input_shape = input_shape, activation = 'relu'))\n   model.add(layers.BatchNormalization(axis = 3)) \n   model.add(layers.MaxPooling2D(pool_size = (2, 2)))\n\n   model.add(layers.Conv2D(32, (3, 3), padding = 'same', activation = 'relu'))\n   model.add(layers.BatchNormalization(axis = 3))\n   model.add(layers.MaxPooling2D(pool_size = (2, 2)))\n\n   model.add(layers.Conv2D(64, (3, 3), padding = 'same', activation = 'relu'))\n   model.add(layers.BatchNormalization(axis = 3))\n   model.add(layers.MaxPooling2D(pool_size = (2,2)))\n\n   model.add(layers.Flatten())\n   model.add(layers.Dense(128, activation = 'relu'))\n   model.add(layers.Dense(28, activation = 'softmax')) \n\n   return model\n<\/code><\/pre>\n\n<p>and the following piece of code creates, compiles, and fits it:<\/p>\n\n<pre><code>doodle_model = create_model((image_size, image_size, 1)) #image_size = 28\ndoodle_model.compile (optimizer = \"Adam\", loss = \"categorical_crossentropy\", metrics = [\"accuracy\"])\ndoodle_model.fit (x = X_train, y = Y_train, epochs = 4, batch_size = 256)\n<\/code><\/pre>\n\n<p>I was getting good results with this model and so I decided to save it to deploy it later on in a web app. I have used:<\/p>\n\n<pre><code>doodle_model.save(\"my_model.h5\")\n<\/code><\/pre>\n\n<p>to save the model<\/p>\n\n<p>However, when I tried to load the model again with a:<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model (\"my_model.h5\")\n<\/code><\/pre>\n\n<p>to make sure if it is saved correctly, I get the following error...<\/p>\n\n<pre><code>KeyError                                  Traceback (most recent call last)\n&lt;ipython-input-61-6cde554a8add&gt; in &lt;module&gt;()\n      1 from keras.models import load_model\n      2 \n----&gt; 3 model = load_model (\"my_model.h5\")\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/models.py in load_model(filepath, custom_objects, compile)\n    268             raise ValueError('No model found in config file.')\n    269         model_config = json.loads(model_config.decode('utf-8'))\n--&gt; 270         model = model_from_config(model_config, custom_objects=custom_objects)\n    271 \n    272         # set weights\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/models.py in model_from_config(config, custom_objects)\n    345                         'Maybe you meant to use '\n    346                         '`Sequential.from_config(config)`?')\n--&gt; 347     return layer_module.deserialize(config, custom_objects=custom_objects)\n    348 \n    349 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/layers\/__init__.py in deserialize(config, custom_objects)\n     53                                     module_objects=globs,\n     54                                     custom_objects=custom_objects,\n---&gt; 55                                     printable_module_name='layer')\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/utils\/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    142                 return cls.from_config(config['config'],\n    143                                        custom_objects=dict(list(_GLOBAL_CUSTOM_OBJECTS.items()) +\n--&gt; 144                                                            list(custom_objects.items())))\n    145             with CustomObjectScope(custom_objects):\n    146                 return cls.from_config(config['config'])\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/models.py in from_config(cls, config, custom_objects)\n   1404     @classmethod\n   1405     def from_config(cls, config, custom_objects=None):\n-&gt; 1406         if 'class_name' not in config[0] or config[0]['class_name'] == 'Merge':\n   1407             return cls.legacy_from_config(config)\n   1408 \n\nKeyError: 0\n<\/code><\/pre>\n\n<p>As far as I know, KeyErrors are related to python dictionaries but I am not sure why I am encountering a KeyError over here. Any help on why I am getting this error and how can it be resolved would be greatly appreciated.<\/p>\n\n<p>Every piece of code, after the line <\/p>\n\n<pre><code>----&gt; 3 model = load_model (\"my_model.h5\")\n<\/code><\/pre>\n\n<p>of this error message is some internal code of Keras and is not written by me.<\/p>\n","answers":[{"AnswerId":"52957781","CreationDate":"2018-10-23T21:01:00.890","ParentId":null,"OwnerUserId":"7649152","Title":null,"Body":"<p>It sounds like you may be trying to load a saved model using a version of Keras incompatible with the version you used for saving the model.<\/p>\n\n<p>What versions are you using in each case? You can check using:<\/p>\n\n<pre><code>import keras\nprint(keras.__version__)\n<\/code><\/pre>\n\n<p>It may be that the fix for you would be to upgrade you version of Keras.<\/p>\n"},{"AnswerId":"57428908","CreationDate":"2019-08-09T11:15:25.833","ParentId":null,"OwnerUserId":"7888603","Title":null,"Body":"<p>It is because of keras version mismatch.Check the versions before proceeding. Build a virtual env with all compatible version. Try using keras 2.2.4<\/p>\n"},{"AnswerId":"55781873","CreationDate":"2019-04-21T10:21:19.790","ParentId":null,"OwnerUserId":"3103059","Title":null,"Body":"<p>What Keras version are you using? If it is 2.1.6 then try Keras 2.2.4 or newer and your problem will be gone. It happened to me.<\/p>\n"}]}
+{"QuestionId":52909037,"AnswerCount":0,"Tags":"<cmake><pytorch>","CreationDate":"2018-10-20T18:52:22.250","AcceptedAnswerId":null,"OwnerUserId":2507197.0,"Title":"CMake not finding files - Can't build detectron with pytorch","Body":"<p>make ops for <a href=\"https:\/\/github.com\/facebookresearch\/Detectron\" rel=\"nofollow noreferrer\">detectron<\/a> isn't working <\/p>\n\n<pre><code>export Caffe2_DIR=&lt;path to my pytorch&gt;\/torch\nmake ops\n<\/code><\/pre>\n\n<h3>Error Message<\/h3>\n\n<pre><code>[ 40%] Building CXX object CMakeFiles\/caffe2_detectron_custom_ops.dir\/detectron\/ops\/zero_even_op.cc.o\nIn file included from \/home\/ubtuntu\/detectron\/detectron\/ops\/zero_even_op.cc:17:0:\n\/home\/ubtuntu\/detectron\/detectron\/ops\/zero_even_op.h:20:33: fatal error: caffe2\/core\/context.h: No such file or directory\ncompilation terminated.\nCMakeFiles\/caffe2_detectron_custom_ops.dir\/build.make:62: recipe for target 'CMakeFiles\/caffe2_detectron_custom_ops.dir\/detectron\/ops\/zero_even_op.cc.o' failed\nmake[3]: *** [CMakeFiles\/caffe2_detectron_custom_ops.dir\/detectron\/ops\/zero_even_op.cc.o] Error 1\nmake[3]: Leaving directory '\/home\/ubtuntu\/detectron\/build'\nCMakeFiles\/Makefile2:72: recipe for target 'CMakeFiles\/caffe2_detectron_custom_ops.dir\/all' failed\nmake[2]: *** [CMakeFiles\/caffe2_detectron_custom_ops.dir\/all] Error 2\nmake[2]: *** Waiting for unfinished jobs....\n\/home\/ubtuntu\/detectron\/detectron\/ops\/zero_even_op.cu:17:37: fatal error: caffe2\/core\/context_gpu.h: No such file or directory\ncompilation terminated.\nCMake Error at caffe2_detectron_custom_ops_gpu_generated_zero_even_op.cu.o.cmake:219 (message):\n  Error generating\n  \/home\/ubtuntu\/detectron\/build\/CMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/detectron\/ops\/.\/caffe2_detectron_custom_ops_gpu_generated_zero_even_op.cu.o\n\n\nCMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/build.make:63: recipe for target 'CMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/detectron\/ops\/caffe2_detectron_custom_ops_gpu_generated_zero_even_op.cu.o' failed\nmake[3]: *** [CMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/detectron\/ops\/caffe2_detectron_custom_ops_gpu_generated_zero_even_op.cu.o] Error 1\nmake[3]: Leaving directory '\/home\/ubtuntu\/detectron\/build'\nCMakeFiles\/Makefile2:109: recipe for target 'CMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/all' failed\nmake[2]: *** [CMakeFiles\/caffe2_detectron_custom_ops_gpu.dir\/all] Error 2\nmake[2]: Leaving directory '\/home\/ubtuntu\/detectron\/build'\nMakefile:129: recipe for target 'all' failed\nmake[1]: *** [all] Error 2\nmake[1]: Leaving directory '\/home\/ubtuntu\/detectron\/build'\nMakefile:13: recipe for target 'ops' failed\nmake: *** [ops] Error 2\n<\/code><\/pre>\n\n<h3>What I've Tried<\/h3>\n\n<p>In the error message, it says it can't find <code>caffe2\/core\/context.h<\/code> however the file exists in <code>pytorch\/caffe2\/core\/context.h<\/code>. So I've tried exporting the caffe2 folder instead of the torch folder, but then the build fails earlier\n- I've also tried setting it to pytorch\/cmake and pytorch\/build folders<\/p>\n\n<p>I then tried creating a caffe2 link to pytorch\/caffe2 inside of pytorch\/torch, however that had no effect. I'm out of ideas :( <\/p>\n","answers":[]}
+{"QuestionId":52909563,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-20T19:51:49.797","AcceptedAnswerId":null,"OwnerUserId":3508847.0,"Title":"Freezing Graph in Tensorflow","Body":"<p>I have trained a neural network and have generated a model folder with <code>.ckpt<\/code>, <code>.index<\/code>, etc files from it. The following is code I am using to generate a <code>.pb<\/code> file; however, I receive the following error:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'NoneType' object has no attribute 'model_checkpoint_path'<\/p>\n<\/blockquote>\n\n<p>Does anyone assist me with solving the problem?<\/p>\n\n<pre><code>dir = os.path.dirname(os.path.realpath(__file__))\n\ndef freeze_graph(model_dir, output_node_names):\n\"\"\"Extract the sub graph defined by the output nodes and convert \nall its variables into constant \nArgs:\n    model_dir: the root folder containing the checkpoint state file\n    output_node_names: a string, containing all the output node's names, \n                        comma separated\n\"\"\"\nif not tf.gfile.Exists(model_dir):\n    raise AssertionError(\n        \"Export directory doesn't exists. Please specify an export \"\n        \"directory: %s\" % model_dir)\n\nif not output_node_names:\n    print(\"You need to supply the name of a node to --output_node_names.\")\n    return -1\n\n# We retrieve our checkpoint fullpath\ncheckpoint = tf.train.get_checkpoint_state(model_dir)\ninput_checkpoint = checkpoint.model_checkpoint_path\n\n# We precise the file fullname of our freezed graph\nabsolute_model_dir = \"\/\".join(input_checkpoint.split('\/')[:-1])\noutput_graph = absolute_model_dir + \"\/frozen_model.pb\"\n\n# We clear devices to allow TensorFlow to control on which device it will load operations\nclear_devices = True\n\n# We start a session using a temporary fresh Graph\nwith tf.Session(graph=tf.Graph()) as sess:\n    # We import the meta graph in the current default Graph\n    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=clear_devices)\n\n    # We restore the weights\n    saver.restore(sess, input_checkpoint)\n\n    # We use a built-in TF helper to export variables to constants\n    output_graph_def = tf.graph_util.convert_variables_to_constants(\n        sess, # The session is used to retrieve the weights\n        tf.get_default_graph().as_graph_def(), # The graph_def is used to retrieve the nodes \n        output_node_names.split(\",\") # The output node names are used to select the usefull nodes\n    ) \n\n    # Finally we serialize and dump the output graph to the filesystem\n    with tf.gfile.GFile(output_graph, \"wb\") as f:\n        f.write(output_graph_def.SerializeToString())\n    print(\"%d ops in the final graph.\" % len(output_graph_def.node))\n\nreturn output_graph_def\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser()\n    parser.add_argument(\"--model_dir\", type=str, default=\"\", help=\"Model folder to export\")\n    parser.add_argument(\"--output_node_names\", type=str, default=\"\", help=\"The name of the output nodes, comma separated.\")\n   args = parser.parse_args()\n\n   freeze_graph(args.model_dir, args.output_node_names)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52909618,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2018-10-20T20:00:17.807","AcceptedAnswerId":null,"OwnerUserId":8618419.0,"Title":"error while importing keras ModuleNotFoundError: No module named 'tensorflow.examples'; 'tensorflow' is not a package","Body":"<p>i tried import keras but it says:<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import keras\n          Using TensorFlow backend.\n          Traceback (most recent call last):\n            File \"\", line 1, in \n            File \n          \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site- \n         packages\/keras\/<strong>init<\/strong>.py\", line 3, in  \n              from . import utils\n            File \n          \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site- \n          packages\/keras\/utils\/<strong>init<\/strong>.py\", line 6, in \n              from . import conv_utils\n            File \n          \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site- \n         packages\/keras\/utils\/conv_utils.py\", line 9, in \n              from .. import backend as K\n            File \n          \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site- \n         packages\/keras\/backend\/<strong>init<\/strong>.py\", line 89, in  \n              from .tensorflow_backend import *\n            File \n          \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site- \n         packages\/keras\/backend\/tensorflow_backend.py\", line 5, in \n              import tensorflow as tf\n            File \"\/Users\/elanchezhian\/tensorflow.py\", line 2, in \n              from tensorflow.examples.tutorials.mnist import input_data\n          ModuleNotFoundError: No module named 'tensorflow.examples'; 'tensorflow' is not a package<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>i am using mac os high sierra and i am running keras in python3 and i installed keras with pip<\/p>\n\n<p>edit: ok ,i already had a file named tensorflow.py ,i removed it and new error came <\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/keras\/<strong>init<\/strong>.py\", line 3, in \n    from . import utils\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/keras\/utils\/<strong>init<\/strong>.py\", line 6, in \n    from . import conv_utils\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in \n    from .. import backend as K\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/keras\/backend\/<strong>init<\/strong>.py\", line 89, in \n    from .tensorflow_backend import *\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 5, in \n    import tensorflow as tf\nImportError: bad magic number in 'tensorflow': b'\\x03\\xf3\\r\\n'<\/p>\n\n<p>edit: i have removed tensorflow.pyc now this error came :<\/p>\n\n<pre><code>RuntimeError: module compiled against API version 0xc but this version \nof \nnumpy is 0xb\nImportError: numpy.core.multiarray failed to import\nImportError: numpy.core.umath failed to import\nImportError: numpy.core.umath failed to import\n2018-10-21 10:53:26.723451: F \ntensorflow\/python\/lib\/core\/bfloat16.cc:675] \nCheck failed: PyBfloat16_Type.tp_base != nullptr \nAbort trap: 6\n<\/code><\/pre>\n","answers":[{"AnswerId":"52909700","CreationDate":"2018-10-20T20:09:22.410","ParentId":null,"OwnerUserId":"494134","Title":null,"Body":"<p>Your script is named <code>tensorflow.py<\/code>, which conflicts with the name of the real tensorflow package. Name your script something else.<\/p>\n"}]}
+{"QuestionId":52909886,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2018-10-20T20:34:16.280","AcceptedAnswerId":null,"OwnerUserId":228841.0,"Title":"Keras: trainable placeholders or layer weights per sample","Body":"<p>I'm using Keras. My goal is to build an autoencoder that learns an additional per-sample latent code. Here is some math to explain my problem: let <code>z\u1d62 = e(x\u1d62;a)<\/code> be the encoder and <code>y\u1d62 = d(z\u1d62;b)<\/code> be the decoder, where <code>x\u1d62<\/code> are the inputs, <code>y\u1d62<\/code> are the outputs, <code>z\u1d62<\/code> are the latent codes, and <code>a<\/code>, <code>b<\/code> are the parameters of the encoder and decoder, respectively. A normal autoencoder training task is to minimize the loss <code>L(.,.)<\/code> with respect to the weights <code>a<\/code> and <code>b<\/code>, i.e. <code>min_{a,b} L(x\u1d62, d(e(x\u1d62;a);b)) \u2200i<\/code>.<\/p>\n\n<p>I want to add an additional per-sample parameter <code>k\u1d62<\/code> in the latent dimension which should be minimized as well, i.e. <code>min_{a,b,k\u1d62} L(x\u1d62, d(e(x\u1d62;a)+k\u1d62);b)<\/code>. However, <code>a<\/code> and <code>b<\/code> are normal model parameters that are shared across all samples of <code>x\u1d62<\/code>, but <code>k\u1d62<\/code> shall be a per sample parameter or weight.<\/p>\n\n<p>I want to realize this model in keras. The easiest way would be to use the functional API and specify <code>k\u1d62<\/code> as Input (placeholder). My main problem is that this input\/placeholder would not be trainable.<\/p>\n\n<p>Here is a minimal example program:<\/p>\n\n<pre><code>import numpy as np\nimport keras\nimport keras.layers as L\nfrom keras.models import Model\n\nx = L.Input([2])\nenc = L.Dense(1, activation='tanh')\ndec = L.Dense(2, activation='tanh')\n\nk = L.Input([1])\nz = enc(x)\ny = dec(L.add([z,k]))\nm = Model([x,k],y)\n\nX = np.random.randn(10,2)  # my samples\nK = np.random.randn(10,1)  # the latent code that should be learned\n\nm.compile('sgd', 'mse')\nm.fit([X,K],X)   # K won't update, obviously\n<\/code><\/pre>\n\n<p>As the comment says, K won't be trained. Is there a way to achieve what I want, i.e. make <code>K<\/code> trainable but keep the data outside the model? I know that this could be done using the low-level backend API. But this would require me to essentially re-implement all the high-level tools (optimizers, layers, etc.) which is the main reason why I want to use keras.<\/p>\n","answers":[]}
+{"QuestionId":52910106,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory>","CreationDate":"2018-10-20T21:05:15.463","AcceptedAnswerId":52910801.0,"OwnerUserId":10499813.0,"Title":"Tensorflow error Fetch argument <built-in function sum> must be a string or Tensor","Body":"<p>I keep getting this tensorflow error and I cannot figure out why. \nmy code:<\/p>\n\n<pre><code>__future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\nimport tensorflow as tf\nimport os\n\na = tf.constant(2.5)\nb = tf.constant(4.5)\ntotal = a + b;\n\ntf.summary.scalar(\"a\",a)\ntf.summary.scalar(\"b\",b)\ntf.summary.scalar(\"total\",total)\n\nmerged_op = tf.summary.merge_all()\n\nwriter = tf.summary.FileWriter(\"summary\")\n\n# Launch Session\nwith tf.Session() as sess:\n  _, summary = sess.run([sum,merged_op])\n  writer.add_summary(summary)\n  writer.close()\n\nAnd this is the full error:\nTypeErrorTraceback (most recent call last)\n&lt;ipython-input-82-155dabddd08a&gt; in &lt;module&gt;()\n      1 with tf.Session() as sess:\n----&gt; 2   _, summary = sess.run([sum,merged_op])\n      3   writer.add_summary(summary)\n      4   writer.close()\n      5 \n\n\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.pyc in run(self, fetches, feed_dict, options, run_metadata)\n    927     try:\n    928       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 929                          run_metadata_ptr)\n    930       if run_metadata:\n    931         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n\nTypeError: Fetch argument &lt;built-in function sum&gt; has invalid type &lt;type 'builtin_function_or_method'&gt;, must be a string or Tensor. (Can not convert a builtin_function_or_method into a Tensor or Operation.)\n<\/code><\/pre>\n\n<p>I don't know what to do I am confused why I am getting this error because basically copied the code from a book I am reading called tensorflow for dummies. <\/p>\n","answers":[{"AnswerId":"52910801","CreationDate":"2018-10-20T23:04:09.903","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>you've not declared 'sum' before doing sess.run. sum is a built-in function in python.\nYou might want to run 'total'<\/p>\n\n<pre><code>sess.run([total,merged_op])\n<\/code><\/pre>\n\n<p>to proof that sum is built-in:<\/p>\n\n<pre><code>sum([5,4])\n&gt;&gt;&gt; 9\n<\/code><\/pre>\n"}]}
+{"QuestionId":52910319,"AnswerCount":0,"Tags":"<python><tensorflow><keras><image-segmentation><tensorflow-datasets>","CreationDate":"2018-10-20T21:34:09.507","AcceptedAnswerId":null,"OwnerUserId":10532410.0,"Title":"Can't use Tensorflow Dataset with Keras model for Image Segmentation","Body":"<p>I am trying to do image segmentation in Keras using a tensorflow dataset, but I am getting a <code>AttributeError: 'MapDataSet' has no attribute 'ndim'<\/code> error when I try to call <code>model.fit()<\/code>.  I am using a Jupyter Notebook in google colab.  The final line of code produces the error.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow import keras \nimport sys\nfrom json import load\nimport math\nimport tensorflow as tf\nfrom tensorflow.contrib import layers\nfrom tensorflow.contrib.data import map_and_batch\n\nexport_file = 'export.json'\nIMG_WIDTH = 256\nIMG_HEIGHT = 256\nwith open(export_file) as f:\n  export_json = load(f)\nlegend = export_json['legend']\ntfrecord_paths = export_json['tfrecord_paths']\n\n# Fairly certain nothing wrong here\n# -----------------------------------------------\ndef _parse_tfrecord(serialized_example):\n    example = tf.parse_single_example(\n        serialized_example,\n        features={\n            'image\/encoded': tf.FixedLenFeature([], tf.string),\n            'image\/filename': tf.FixedLenFeature([], tf.string),\n            'image\/ID': tf.FixedLenFeature([], tf.string),\n            'image\/format': tf.FixedLenFeature([], tf.string),\n            'image\/height': tf.FixedLenFeature([], tf.int64),\n            'image\/width': tf.FixedLenFeature([], tf.int64),\n            'image\/channels': tf.FixedLenFeature([], tf.int64),\n            'image\/colorspace': tf.FixedLenFeature([], tf.string),\n            'image\/segmentation\/class\/encoded': tf.FixedLenFeature([], tf.string),\n            'image\/segmentation\/class\/format': tf.FixedLenFeature([], tf.string),\n            })\n    image = tf.image.decode_image(example['image\/encoded'])\n    image.set_shape([IMG_WIDTH, IMG_HEIGHT, 3])\n    label = tf.image.decode_image(example['image\/segmentation\/class\/encoded'])\n    label.set_shape([IMG_WIDTH, IMG_HEIGHT, 1])\n    image_float = tf.to_float(image)\n    label_float = tf.to_float(label)\n    return (image_float, label_float)\n# -----------------------------------------------\n\n# Create training and testing datasets\ntest_set_size = math.floor(0.20 * len(tfrecord_paths))\ntraining_dataset = tf.data.TFRecordDataset(tfrecord_paths)\ntraining_dataset = training_dataset.skip(test_set_size)\ntraining_dataset = training_dataset.map(_parse_tfrecord)\n\ntest_dataset = tf.data.TFRecordDataset(tfrecord_paths)\ntest_dataset = test_dataset.take(test_set_size)\ntest_dataset = test_dataset.map(_parse_tfrecord)\n\n# Printing training_dataset yields: '&lt;MapDataset shapes: ((256, 256, 3), (256, 256, 1)), types: (tf.float32, tf.float32)&gt;'\n\n# Load Inception v3\nfrom keras.applications.inception_v3 import InceptionV3\nmodel = InceptionV3(input_shape = (IMG_HEIGHT, IMG_WIDTH, 3), include_top=False, weights='imagenet')\nmodel.compile(optimizer='RMSProp', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(training_dataset, epochs=10, steps_per_epoch=30) # -=CAUSES ERROR=-\n<\/code><\/pre>\n\n<p>From <a href=\"https:\/\/stackoverflow.com\/questions\/46135499\/how-to-properly-combine-tensorflows-dataset-api-and-keras\">How to Properly Combine TensorFlow&#39;s Dataset API and Keras?<\/a>, I tried to use the second solution but I got the same error, even when changing <code>model.fit()<\/code> to accept a <code>one_shot_iterator<\/code> instead of a <code>tf.data.Dataset<\/code>.  <\/p>\n\n<p>It seems like this might be a more common issue, as I found at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20698\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20698<\/a>.  However, every solution posted here resulted in the same error that I got before.  <\/p>\n","answers":[]}
+{"QuestionId":52910630,"AnswerCount":0,"Tags":"<python><tensorflow><keras><attributeerror>","CreationDate":"2018-10-20T22:28:41.477","AcceptedAnswerId":null,"OwnerUserId":10534379.0,"Title":"How to use tf.keras.layers.DepthwiseConv2D in Keras model? Got AttributeError: 'Node' object has no attribute 'output_masks'","Body":"<p>I have difficulty in using DepthwiseConv2D layer in my Keras model. I tried to look it up but many of the accepted answer regarding the problem is not to mix tf.keras.layers with pure keras.layers. <\/p>\n\n<p>Is there any way I can use the tf.keras.layers.DepthwiseConv2D in my keras.layers model? Or is there anyway I can perform DepthwiseConv2D with Keras? <\/p>\n\n<p>My Code:<\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Input,Conv2D\nfrom tensorflow.keras.layers import DepthwiseConv2D\n\ninputs = Input((224, 224,3))\nconv1 = Conv2D(16, 3, activation = 'relu', padding = 'same')(inputs)\nconv1 = DepthwiseConv2D((3, 3), padding='valid',depth_multiplier=1, strides=(1,1), use_bias=False)(conv1)\nmodel= Model(input=inputs,output=conv1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52911483,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><mini-batch>","CreationDate":"2018-10-21T01:39:41.313","AcceptedAnswerId":52912767.0,"OwnerUserId":10534776.0,"Title":"Custom CNN mini-batch (Keras, TF) to avoid repeated measurements in training\/testing","Body":"<p>I am currently building a 1D-CNN for classification. The predictors are spectra (<strong>X<\/strong>-matrix with 779 features), and the dependent variable contains two classes.<\/p>\n\n<p>However, the <strong>X<\/strong>-matrix contains repeated measurements (series of 15-20 replicates). It is crucial that during training repeated measurements are not included both in the sets for training and evaluation of the loss function. Is there a way to build \"custom\" mini-batches which would avoid this?<\/p>\n","answers":[{"AnswerId":"52912767","CreationDate":"2018-10-21T06:43:44.113","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>You should try using data generators.<\/p>\n\n<p>A DataGenerator is an object that takes as input the X_train and y_train matrices and put the samples into batches following some criterion. It can also be used to handle large volumes of data that cannot be loaded at once on the virtual memory.<\/p>\n\n<p>Here is an example on how to implement one !<\/p>\n\n<p>Basically get_item will give you your next batch so that's the place to implement all the conditions you might need.<\/p>\n\n<pre><code>import numpy as np\nimport keras\n\nclass DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, X, labels, batch_size=32, dim=(32,32,32), n_channels=1,\n                 n_classes=10, shuffle=True):\n        'Initialization'\n        self.dim = dim\n        self.batch_size = batch_size\n        self.labels = labels\n        self.X = X\n        self.n_channels = n_channels\n        self.n_classes = n_classes\n        self.shuffle = shuffle\n        self.on_epoch_end()\n\n    def __len__(self):\n        'Denotes the number of batches per epoch'\n        return int(np.floor(len(self.X) \/ self.batch_size))\n\n    def __getitem__(self, index):\n        'Generate one batch of data'\n        # Generate indexes of the batch to make sure samples dont repeat\n        list_IDs_temp = ... your code\n\n        # Generate data\n        X, y = self.__data_generation(list_IDs_temp)\n\n        return X, y\n\n    def on_epoch_end(self):\n        'Updates indexes after each epoch'\n        self.indexes = np.arange(len(self.X))\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n\n    def __data_generation(self, list_IDs_temp):\n        'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n        # Initialization\n        X = np.empty((self.batch_size, *self.dim, self.n_channels))\n        y = np.empty((self.batch_size), dtype=int)\n\n        # Generate data\n        for i, ID in enumerate(list_IDs_temp):\n            # Store sample\n            X[i,] = self.X[ID,]\n\n            # Store class\n            y[i] = self.labels[ID]\n\n        return X, keras.utils.to_categorical(y, num_classes=self.n_classes)\n<\/code><\/pre>\n\n<p>Source: <a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">This<\/a> <\/p>\n"}]}
+{"QuestionId":52912143,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-21T04:33:42.480","AcceptedAnswerId":null,"OwnerUserId":10534934.0,"Title":"Why is tensorflow installation failing","Body":"<p>I'm attempting to install tensorflow using the conda install tensoflow command.<\/p>\n\n<p>However, the installation appears to fail when attempting to uninstall numpy.<\/p>\n\n<p>Preparing transaction: done\nVerifying transaction: done\nExecuting transaction: failed\nERROR conda.core.link:_execute(481): An error occurred while uninstalling package 'defaults::numpy-1.14.0-py36h4a99626_1'.\nPermissionError(13, 'Access is denied')\nAttempting to roll back.<\/p>\n\n<p>Rolling back transaction: done<\/p>\n\n<p>PermissionError(13, 'Access is denied')<\/p>\n\n<p>does anyone know how to fix it?<\/p>\n","answers":[{"AnswerId":"52912168","CreationDate":"2018-10-21T04:41:32.077","ParentId":null,"OwnerUserId":"10535020","Title":null,"Body":"<p>You should try <code>sudo chown -R user anaconda3<\/code> and try again.<\/p>\n\n<ul>\n<li>Check out : <a href=\"https:\/\/github.com\/conda\/conda\/issues\/3448\" rel=\"nofollow noreferrer\">conda update permission error<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":52912214,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-21T04:52:34.087","AcceptedAnswerId":52912656.0,"OwnerUserId":4958349.0,"Title":"Attempting to use uninitialized value output_softmax_w","Body":"<p>Here's my LSTM <code>model.py<\/code><\/p>\n\n<pre><code>class Model(object):\n\n    def __init__(self, is_training, batch_size, seq_max_length, hidden_size, vocab_size, \n                 num_layers, learning_rate, dropout=0.5, init_scale=0.05):\n        self.is_training = is_training\n        self.batch_size = batch_size\n        self.seq_max_length = seq_max_length\n        self.hidden_size = hidden_size\n        self.learning_rate = learning_rate\n\n        self.inputs = tf.placeholder(name='inputs', shape=[batch_size, seq_max_length], dtype=tf.int32)\n        self.targets = tf.placeholder(name='targets', shape=[batch_size, seq_max_length], dtype=tf.int32)\n\n        embedding = tf.get_variable('embedding', shape=[vocab_size, self.hidden_size], initializer=tf.contrib.layers.variance_scaling_initializer())\n        inputs = tf.nn.embedding_lookup(embedding, self.inputs)\n\n        cell = tf.contrib.rnn.LSTMCell(self.hidden_size)\n\n        output, self.state = tf.nn.dynamic_rnn(cell, inputs, dtype=tf.float32)\n        output = tf.reshape(output, [-1, hidden_size])\n\n        self.softmax_w = tf.get_variable('output_softmax_w', shape=[self.hidden_size, vocab_size], \n                                    dtype=tf.float32)\n        self.softmax_b = tf.get_variable('output_softmax_b', shape=[vocab_size], \n                                    dtype=tf.float32)\n        logits = tf.nn.xw_plus_b(output, self.softmax_w, self.softmax_b)\n        logits = tf.reshape(logits, [self.batch_size, self.seq_max_length, vocab_size])\n\n        loss = tf.contrib.seq2seq.sequence_loss(\n            logits,\n            self.targets,\n            tf.ones([self.batch_size, self.seq_max_length], dtype=tf.float32),\n            average_across_timesteps=False,\n            average_across_batch=True\n        )\n        self.cost = tf.reduce_sum(loss)\n\n        if not is_training:\n            return\n\n        tvars = tf.trainable_variables()\n        grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars), 5)\n        optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.learning_rate)\n        self.train_op = optimizer.apply_gradients(\n            zip(grads, tvars),\n            global_step=tf.contrib.framework.get_or_create_global_step())\n<\/code><\/pre>\n\n<p>when i run the <code>train<\/code> code, <\/p>\n\n<pre><code>tf.reset_default_graph()\n\ninit_op = tf.global_variables_initializer()\nmodel = Model(is_training=True, batch_size=2, seq_max_length=10, hidden_size=100, vocab_size=72, num_layers=1, learning_rate=1)\n\nwith tf.Session() as sess:\n\n    sess.run(init_op)\n\n    batch_x = np.array(train_X[0:2])\n    batch_y = np.array(train_y[0:2])\n    inputs_value = sess.run([model.cost, model.train_op], feed_dict={model.inputs: batch_x, model.targets: batch_y})\n<\/code><\/pre>\n\n<p>The error show:<\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value output_softmax_w\n     [[Node: output_softmax_w\/read = Identity[T=DT_FLOAT, _class=[\"loc:@output_softmax_w\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](output_softmax_w)]]\n     [[Node: GradientDescent\/update\/_18 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_384_GradientDescent\/update\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>I don't know why i wrote everything the same as others did, but i got the error.<\/p>\n","answers":[{"AnswerId":"52912656","CreationDate":"2018-10-21T06:22:11.103","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Move your init op after the graph construction code like<\/p>\n\n<pre><code>model = Model(is_training=True, batch_size=2, seq_max_length=10, hidden_size=100, vocab_size=72, num_layers=1, learning_rate=1)\ninit_op = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>global_variable_initializer() makes an initializer for the variables decalared before it.<\/p>\n"}]}
+{"QuestionId":52912424,"AnswerCount":2,"Tags":"<deep-learning><caffe><multilabel-classification>","CreationDate":"2018-10-21T05:44:09.353","AcceptedAnswerId":52912707.0,"OwnerUserId":4749728.0,"Title":"Caffe - multi-class and multi-label image classification","Body":"<p>I'm trying to create a single multi-class and multi-label net configuration in caffe.<\/p>\n\n<p>Let's say classification of dogs:\nIs the dog small or large? (class)\nWhat color is it? (class)\nis it have a collar? (label)<\/p>\n\n<p>Is this thing possible using caffe?\nWhat is the proper way to do so?\nWhat is the right way to build the lmdb file?<\/p>\n\n<p>All the publications about multi-label classification are from around 2015, something in this subject changed since then?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53029555","CreationDate":"2018-10-28T08:12:46.543","ParentId":null,"OwnerUserId":"4749728","Title":null,"Body":"<p>Thanks <a href=\"https:\/\/stackoverflow.com\/a\/52912707\/1714410\">Shai<\/a>,<\/p>\n\n<p>Just trying to understand the practical way.. \nAfter creating 2 .text files (one for training and one for validation) containing all the tags of the images, for example:<\/p>\n\n<pre><code>\/train\/img\/1.png 0 4 18\n\/train\/img\/2.png 1 7 17 33\n\/train\/img\/3.png 0 4 17\n<\/code><\/pre>\n\n<p>Running the py script:<\/p>\n\n<pre><code>import h5py, os\nimport caffe\nimport numpy as np\n\nSIZE = 227 # fixed size to all images\nwith open( 'train.txt', 'r' ) as T :\n    lines = T.readlines()\n# If you do not have enough memory split data into\n# multiple batches and generate multiple separate h5 files\nX = np.zeros( (len(lines), 3, SIZE, SIZE), dtype='f4' ) \ny = np.zeros( (len(lines),1), dtype='f4' )\nfor i,l in enumerate(lines):\n    sp = l.split(' ')\n    img = caffe.io.load_image( sp[0] )\n    img = caffe.io.resize( img, (SIZE, SIZE, 3) ) # resize to fixed size\n    # you may apply other input transformations here...\n    # Note that the transformation should take img from size-by-size-by-3 and transpose it to 3-by-size-by-size\n    # for example\n    transposed_img = img.transpose((2,0,1))[::-1,:,:] # RGB-&gt;BGR\n    X[i] = transposed_img\n    y[i] = float(sp[1])\nwith h5py.File('train.h5','w') as H:\n    H.create_dataset( 'X', data=X ) # note the name X given to the dataset!\n    H.create_dataset( 'y', data=y ) # note the name y given to the dataset!\nwith open('train_h5_list.txt','w') as L:\n    L.write( 'train.h5' ) # list all h5 files you are going to use\n<\/code><\/pre>\n\n<p>And creating train.h5 and val.h5 (is X data set containing the images and Y contain the labels?).<\/p>\n\n<p>Replace my network input layers from:<\/p>\n\n<pre><code>layers { \n name: \"data\" \n type: DATA \n top:  \"data\" \n top:  \"label\" \n data_param { \n   source: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/train_db\" \n   backend: LMDB \n   batch_size: 64 \n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/mean.binaryproto\" \n    mirror: true \n  } \n  include: { phase: TRAIN } \n} \nlayers { \n name: \"data\" \n type: DATA \n top:  \"data\" \n top:  \"label\" \n data_param { \n   source: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/val_db\"  \n   backend: LMDB \n   batch_size: 64\n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/mean.binaryproto\" \n    mirror: true \n  } \n  include: { phase: TEST } \n} \n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  top: \"X\" # same name as given in create_dataset!\n  top: \"y\"\n  hdf5_data_param {\n    source: \"train_h5_list.txt\" # do not give the h5 files directly, but the list.\n    batch_size: 32\n  }\n  include { phase:TRAIN }\n}\n\nlayer {\n  type: \"HDF5Data\"\n  top: \"X\" # same name as given in create_dataset!\n  top: \"y\"\n  hdf5_data_param {\n    source: \"val_h5_list.txt\" # do not give the h5 files directly, but the list.\n    batch_size: 32\n  }\n  include { phase:TEST }\n}\n<\/code><\/pre>\n\n<p>I guess HDF5 doesn't need a mean.binaryproto?<\/p>\n\n<p>Next, how the output layer should change in order to output multiple label probabilities?\nI guess I need cross- entropy layer instead of softmax?\nThis is the current output layers:<\/p>\n\n<pre><code>layers {\n  bottom: \"prob\"\n  bottom: \"label\"\n  top: \"loss\"\n  name: \"loss\"\n  type: SOFTMAX_LOSS\n  loss_weight: 1\n}\nlayers {\n  name: \"accuracy\"\n  type: ACCURACY\n  bottom: \"prob\"\n  bottom: \"label\"\n  top: \"accuracy\"\n  include: { phase: TEST }\n}\n<\/code><\/pre>\n"},{"AnswerId":"52912707","CreationDate":"2018-10-21T06:31:55.517","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The problem with Caffe's LMDB interface is that it only allows for a <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">single int label per image<\/a>.<br>\nIf you want multiple labels per image you'll have to use a different input layer.<br>\nI suggest using <a href=\"http:\/\/caffe.help\/manual\/layers\/hdf5data.html\" rel=\"nofollow noreferrer\"><code>\"HDF5Data\"<\/code><\/a> layer:<br>\nThis allows for more flexibility setting the input data, you may have as many <code>\"top\"<\/code>s as you want for this layer. You may have multiple labels per input image and have multiple losses for your net to train on.<\/p>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">this post<\/a> on how to create hdf5 data for caffe.<\/p>\n"}]}
+{"QuestionId":52912553,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-21T06:05:43.363","AcceptedAnswerId":null,"OwnerUserId":9653537.0,"Title":"How To Detect English Language Words Using Machine Learning From Data","Body":"<p>I have data that contains English text messages.<\/p>\n\n<p>I want to detect messages that are \"written in English letters\", but aren't English words. <em>(For example with codes based rules, but I don't want to hard coded the rules).<\/em><\/p>\n\n<p>Please note that the computer being used does not have an active internet connection <em>(so I cannot check against online dictionary)<\/em>.<\/p>\n\n<p><strong>Example Data<\/strong><\/p>\n\n<ol>\n<li>\"hello how are you\"<\/li>\n<li>\"fjrio kjfdelf ejfe\" <em>&lt;-- code (let's say is means \"how are you\" in spanish)<\/em><\/li>\n<li>\"i am fine thanks\"<\/li>\n<li>\"10x man\"<\/li>\n<li>\"jfrojf feoif\" <em>&lt;-- code (let's say it means \"hello world\" in japanish)<\/em><\/li>\n<\/ol>\n\n<p>I'm new to machine learning, so for my understanding, maybe one approach could \nbe using nlp?<\/p>\n","answers":[{"AnswerId":"52912811","CreationDate":"2018-10-21T06:51:13.307","ParentId":null,"OwnerUserId":"7233155","Title":null,"Body":"<p>Character frequency scanning is one way to do this.<\/p>\n\n<p>For example for each language obtain a list of character frequencies,\nA: 3%\nB: 1%\nC: 0.5%\nD: 0.7%\nE: 4%\netc..<\/p>\n\n<p>Then evaluate your string's character frequency against your static map. You can obtain a probabilistic model of the likelihood of the string being one of your languages.<\/p>\n\n<p>Of course this works best for longer strings where there is enough statistical data to capture the true frequency. You would also need to train your frequencies on samples from your target source, e.g. English tweets likely have a different letter frequency to works of Shakespeare. <\/p>\n\n<p>Another option is to find the most likely n-grams in a language, e.g, 'we' is a common 2-gram in english. If you scan your code for how often these most likely n-grams occur you can generally detect if something is in a specific language or not.<\/p>\n\n<p>I'm sure there are also other ideas or combinations of classifiers, but this gives you a start. Don't underestimate the power of an ensemble of classifiers either. For example suppose you came up with 3 different models that were all different and uncorrelated, and say each model could detect english correctly 3 times out of 4 (75%). If you then used all 3 models with an equally weighted vote, so if 3 of 3 or 2 of 3 voted english it was classed as english then your error improves to about 3.4 times correct from 4 (85%) (=0.75^3 + 3*0.75^2*0.25)<\/p>\n"}]}
+{"QuestionId":52912731,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><neural-network><text-classification>","CreationDate":"2018-10-21T06:37:30.510","AcceptedAnswerId":null,"OwnerUserId":6451727.0,"Title":"Using sparse_softmax_cross_entropy() with unequal number of examples in each class","Body":"<p>I am working on convolutional neural network based text classifier. I am using the <code>sparse_softmax_cross_entropy()<\/code> on my <code>logits<\/code> tensor to get the loss but it is producing the following error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Received a label value\nof -9223372036854775808 which is outside the valid range of [0, 9).\n<\/code><\/pre>\n\n<p>Based on what I could gather this is due to unequal number of examples across my classes(I tried running the classifier with equal number of examples for each class, but this gives a final accuracy of just 0.11). The difference of examples is not that profound given that I have 16542 examples spread across 9 classes and with 6 classes consisting of 2001, and three with 1452, 1300 and 1787 examples. How can I get the classifier to work without reducing the number of examples I have for training.<\/p>\n","answers":[]}
+{"QuestionId":52913022,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-21T07:27:32.517","AcceptedAnswerId":null,"OwnerUserId":5254777.0,"Title":"How to implement stochastic pooling in Keras","Body":"<p>Keras still doesn't support any built-in function for stochastic pooling. But I need to use it in my model. How can I implement stochastic pooling and use it in my CNN model in Keras? I am using Keras with Tensorflow backend. <\/p>\n","answers":[]}
+{"QuestionId":52913307,"AnswerCount":1,"Tags":"<python><machine-learning><lstm><pytorch><rnn>","CreationDate":"2018-10-21T08:10:52.547","AcceptedAnswerId":null,"OwnerUserId":10535429.0,"Title":"My LSTM code give flat prediction and don't know what is wrong..Anyone can help take a look?","Body":"<p>Here is the code. I think the class <code>mylstm<\/code> has problem but I can not find it... The input is simple, which is just 7 columns data.<\/p>\n\n<p>I tried to print out all the tensors but did not find what was wrong. Thanks for help!<\/p>\n\n<pre><code>class mylstm(nn.Module):\n    def __init__(self, input_size, hidden_size, T, logger):\n        super(mylstm, self).__init__()\n\n        self.T = T\n        self.input_size = input_size\n        self.hidden_size = hidden_size\n        self.logger = logger\n\n\n        self.lstm_layer = nn.LSTM(input_size = 7, hidden_size = hidden_size)\n        self.fc = nn.Linear(hidden_size, 1)\n        #self.fc.weight.data.normal_()\n\n    def forward(self, input_data):\n\n        hidden = self.init_hidden(input_data)\n        cell = self.init_hidden(input_data)\n        for t in range(self.T - 1):\n            if t &lt; self.T - 1:\n                self.lstm_layer.flatten_parameters()\n                _, lstm_output = self.lstm_layer(input_data[:,t,:].unsqueeze(0), (hidden, cell))\n                hidden = lstm_output[0] \n                cell = lstm_output[1] \n        y_pred = self.fc(hidden[0])\n        return y_pred\n\n\n    def init_hidden(self, x):\n        return Variable(x.data.new(1, x.size(0), self.hidden_size).zero_())\n\n\n# Train the model\nclass rnn:\n    def __init__(self, file_data, logger, input_size = 7, hidden_size = 64, T = 10,\n                 learning_rate = 0.01, batch_size = 128, parallel = True, debug = False):\n        self.T = T\n        dat = pd.read_csv(file_data, nrows = 100 if debug else None)\n        self.logger = logger\n        self.logger.info(\"Shape of data: %s.\\nMissing in data: %s.\", dat.shape, dat.isnull().sum().sum())\n\n        self.X = dat.loc[:, [x for x in dat.columns.tolist()]].values\n        self.y = np.array(dat.rtm_spp)\n        self.batch_size = batch_size\n\n        self.lstm1 = mylstm(input_size = input_size,\n                               hidden_size = hidden_size,\n                               T = T, logger = logger)\n\n        if parallel:\n            self.lstm1 = nn.DataParallel(self.lstm1)\n\n        self.lstm1_optimizer = optim.Adam(params = filter(lambda p: p.requires_grad, self.lstm1.parameters()),\n                                           lr = learning_rate)\n\n        self.train_size = 20000\n        self.y = self.y - np.mean(self.y[:self.train_size]) # Question: why Adam requires data to be normalized?\n        self.logger.info(\"Training size: %d.\", self.train_size)\n\n    def train(self, n_epochs = 10):\n        iter_per_epoch = int(np.ceil(self.train_size * 1. \/ self.batch_size))\n        logger.info(\"Iterations per epoch: %3.3f ~ %d.\", self.train_size * 1. \/ self.batch_size, iter_per_epoch)\n        self.iter_losses = np.zeros(n_epochs * iter_per_epoch)\n        self.epoch_losses = np.zeros(n_epochs)\n\n        self.loss_func = nn.MSELoss()\n\n        n_iter = 0\n\n        learning_rate = 1.\n\n        for i in range(n_epochs):\n            perm_idx = np.random.permutation(self.train_size - self.T-1)\n            j = 0\n            while j &lt; self.train_size:\n                batch_idx = perm_idx[j:(j + self.batch_size)]\n                X = np.zeros((len(batch_idx), self.T - 1, self.X.shape[1]))\n                #y_history = np.zeros((len(batch_idx), self.T - 1))\n                y_target = self.y[batch_idx + self.T]\n\n                for k in range(len(batch_idx)):\n                    X[k, :, :] = self.X[batch_idx[k] : (batch_idx[k] + self.T - 1), :]\n\n                loss = self.train_iteration(X, y_target)\n                self.iter_losses[i * iter_per_epoch + j \/\/ self.batch_size] = loss\n                #if (j \/ self.batch_size) % 50 == 0:\n\n                j += self.batch_size\n                n_iter += 1\n\n                if n_iter % 10000 == 0 and n_iter &gt; 0:\n                    for param_group in self.lstm1_optimizer.param_groups:\n                        param_group['lr'] = param_group['lr'] * 0.9\n\n\n            self.epoch_losses[i] = np.mean(self.iter_losses[range(i * iter_per_epoch, (i + 1) * iter_per_epoch)])\n            if i % 10 == 0:\n                self.logger.info(\"Epoch %d, loss: %3.3f.\", i, self.epoch_losses[i])\n                y_train_pred = self.predict(on_train = True)\n                y_test_pred = self.predict(on_train = False)\n\n\n\n    def train_iteration(self, X,y_target):\n        self.lstm1_optimizer.zero_grad()\n        y_pred = self.lstm1(Variable(torch.from_numpy(X).type(torch.FloatTensor)))\n        y_true = Variable(torch.from_numpy(y_target).type(torch.FloatTensor))\n        y_true = y_true.view(y_true.shape[0],1)\n        y_pred=y_pred.squeeze(0)\n        print(y_pred)\n        loss = self.loss_func(y_pred, y_true)\n        loss.backward()\n        self.lstm1_optimizer.step()\n        return loss.data[0]\n\n    def predict(self, on_train = False):\n        if on_train:\n             y_pred = np.zeros(self.train_size - self.T +1)\n        else:\n             y_pred = np.zeros(self.X.shape[0] - self.train_size)\n\n        i = 0\n        while i &lt; len(y_pred):\n            batch_idx = np.array(range(len(y_pred)))[i : (i + self.batch_size)]\n            X = np.zeros((len(batch_idx), self.T - 1, self.X.shape[1]))\n            #y_history = np.zeros((len(batch_idx), self.T - 1))\n            for j in range(len(batch_idx)):\n                if on_train:\n                    X[j, :, :] = self.X[range(batch_idx[j], batch_idx[j] + self.T - 1), :]\n\n                else:\n                    X[j, :, :] = self.X[range(batch_idx[j] + self.train_size - self.T, batch_idx[j] + self.train_size - 1), :]\n\n            input_data = Variable(torch.from_numpy(X).type(torch.FloatTensor))\n           # print(self.lstm1(torch.randn(128,9,7)))\n            #print(self.lstm1(X).data.numpy())\n            y_pred[i:(i + self.batch_size)] = self.lstm1(input_data).data.numpy()[:,0]\n            i += self.batch_size\n        return y_pred\n\n\nmodel = rnn(file_data = 'L.csv', logger = logger, parallel = False,\n              learning_rate = .001)\n\nmodel.train(n_epochs = 1000)\n\ny_pred = model.predict()\n<\/code><\/pre>\n","answers":[{"AnswerId":"52919484","CreationDate":"2018-10-21T20:16:35.423","ParentId":null,"OwnerUserId":"2904579","Title":null,"Body":"<p>It might be good if you can reduce your codes into the simplest form which still reproduce your problem. Asking people to debug over 200 lines of codes may be too big an ask. If you can give a small example of your problem, using a very simple NN model instead of the current one, many others will be willing to look into your codes and help identify the issue.<\/p>\n"}]}
+{"QuestionId":52913777,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-21T09:20:26.650","AcceptedAnswerId":null,"OwnerUserId":1500887.0,"Title":"Tensorflow cond - ValueError: Operation 'case_3\/cond\/Merge' has been marked as not fetchable","Body":"<p>I'm trying to run the simple code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.random_uniform([], minval=-1, maxval=1, dtype=tf.float32)\ny = tf.random_uniform([], minval=-1, maxval=1, dtype=tf.float32)\nz = tf.Variable(1.0)\ndef f1(): \n    return x + y + z\ndef f2(): return x - y\ndef f3(): return tf.constant(0, dtype=tf.float32)\nout = tf.case({tf.less(x, y): f1, tf.greater(x, y): f2}, default=f3, exclusive=True)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer()) #Edited\n    print(sess.run(out)) \n<\/code><\/pre>\n\n<p>And I receive the error:<\/p>\n\n<pre><code> ValueError: Operation 'case_3\/cond\/Merge' has been marked as not fetchable.\n<\/code><\/pre>\n\n<p>I have read this question: <a href=\"https:\/\/stackoverflow.com\/questions\/52190877\/tensorflow-tf-case-giving-errors\">Tensorflow: tf.case giving errors<\/a>, which is similar, but they do not have a <code>Variable<\/code> which seems to fix the error<\/p>\n\n<p>And I have looked at this issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4094#issue-173787623\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4094#issue-173787623<\/a>\nand understood vaguely how this error occurs.. however I'm wondering if there is any way to solve this. <\/p>\n\n<p>What I'm actually trying to understand is how (and if) Tensorflow will calculate gradients through an if statement.<\/p>\n","answers":[]}
+{"QuestionId":52914205,"AnswerCount":2,"Tags":"<python><scikit-learn><theano><google-colaboratory>","CreationDate":"2018-10-21T10:12:18.333","AcceptedAnswerId":52917662.0,"OwnerUserId":7819329.0,"Title":"Google collab does not find Python package","Body":"<p>I want to use <code>sklearn_theano<\/code> package in Google colab.<\/p>\n\n<p>However, I simply do that :<\/p>\n\n<pre><code>import sklearn_theano\n<\/code><\/pre>\n\n<p>and get <\/p>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'sklearn_theano'<\/p>\n<\/blockquote>\n\n<p>How to make the library available ??<\/p>\n","answers":[{"AnswerId":"52914237","CreationDate":"2018-10-21T10:15:22.480","ParentId":null,"OwnerUserId":"8852013","Title":null,"Body":"<pre><code>!pip install -q library\n<\/code><\/pre>\n\n<p>where library is what you want to install.<\/p>\n"},{"AnswerId":"52917662","CreationDate":"2018-10-21T16:49:19.617","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>Here's an example notebook demonstrating the install:<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1gYny47qg3I3iiI-i5oMKuu4ZSsIzUiIJ\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1gYny47qg3I3iiI-i5oMKuu4ZSsIzUiIJ<\/a><\/p>\n\n<p>It mirrors the <a href=\"http:\/\/sklearn-theano.github.io\/install.html\" rel=\"nofollow noreferrer\">published installation instructions<\/a>.<\/p>\n"}]}
+{"QuestionId":52914302,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-10-21T10:23:58.187","AcceptedAnswerId":null,"OwnerUserId":9211303.0,"Title":"How to process Estimator categorical_column when request tensorflow-model-server?","Body":"<p>As you known, request Tensorflow-serving server need organize request data format like this structure: <\/p>\n\n<pre><code>Example(\n    features=Features(\n        feature={\n            'SepalLength': Feature(float_list=FloatList(value=[5.1])),\n            'SepalWidth': Feature(float_list=FloatList(value=[3.3])),\n            'PetalLength': Feature(float_list=FloatList(value=[1.7])),\n            'PetalWidth': Feature(float_list=FloatList(value=[0.5])),\n        }\n    )\n)\n<\/code><\/pre>\n\n<p>if my data only <code>tf.feature_column.numeric_column<\/code>, that's easy, i can process like this:<\/p>\n\n<pre><code>tf.train.Feature(float_list=tf.train.FloatList(value=[value]))\n<\/code><\/pre>\n\n<p>but my training data like is : <\/p>\n\n<pre><code>categorical_column = tf.feature_column.categorical_column_with_identity(\n            key = name, num_buckets = 3 )\nindicator_column = tf.feature_column.indicator_column( categorical_column )\n<\/code><\/pre>\n\n<p>it's <code>categorical_column<\/code>, like some kind of one hot structure, in this situation, i don't know how to handle this, and i can't find any advanced tutorials to handle this,<\/p>\n\n<p>How should i do if i want to process my data to adapt <code>tensorflow-model-server\n<\/code> request requirement ? thanks.<\/p>\n\n<p>PS: \ni use components below:\ntraining framework:\n<code>tensorflow estimator<\/code>\nclient:\n<code>tensorflow-seving-api<\/code>\nserver:\n<code>tensorflow-model-server<\/code><\/p>\n","answers":[]}
+{"QuestionId":52914390,"AnswerCount":0,"Tags":"<python><tensorflow><lstm>","CreationDate":"2018-10-21T10:35:45.323","AcceptedAnswerId":null,"OwnerUserId":9798353.0,"Title":"Tensorflow sequence to sequence with LSTM","Body":"<p>I am new to python. When I try to run the below program I get the following error.<\/p>\n\n<blockquote>\n  <p>ValueError: Variable model\/embedding\/seq2seq\/seq2seq_encode\/rnn\/multi_rnn_cell\/cell_0\/lstm_cell\/kernel does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?<\/p>\n<\/blockquote>\n\n<p>How do I solve the error?\nBelow is the code which creates the lstm model:<\/p>\n\n<pre><code>def create_model(encode_seqs, decode_seqs, src_vocab_size,emb_dim,is_train=True, reuse=False):\n\n    with tf.variable_scope(\"model\", reuse=reuse):\n        with tf.variable_scope(\"embedding\") as vs:\n\n              tl.layers.set_name_reuse(True)\n              net_encode = EmbeddingInputlayer(\n                inputs = encode_seqs,\n                vocabulary_size = src_vocab_size,\n                embedding_size = emb_dim,\n                name = 'seq_embedding')\n\n              vs.reuse_variables()\n\n\n              net_decode = EmbeddingInputlayer(\n                inputs = decode_seqs,\n                vocabulary_size = src_vocab_size,\n                embedding_size = emb_dim,\n                name = 'seq_embedding')\n\n              net_rnn = Seq2Seq(net_encode, net_decode,\n                cell_fn = tf.nn.rnn_cell.LSTMCell,\n                n_hidden = emb_dim,\n                initializer = tf.random_uniform_initializer(-0.1, 0.1),\n                encode_sequence_length = retrieve_seq_length_op2(encode_seqs),\n                decode_sequence_length = retrieve_seq_length_op2(decode_seqs),\n                initial_state_encode = None,\n                dropout = (0.5 if is_train else None),\n                n_layer = 3,\n                return_seq_2d = True,\n                name = 'seq2seq')\n\n         net_out = DenseLayer(net_rnn, n_units=src_vocab_size, act=tf.identity, name='output')\n\n    return net_out, net_rnn\n<\/code><\/pre>\n\n<p>I have looked at simliar questions but none of the solutions seem to solve my problem.<\/p>\n","answers":[]}
+{"QuestionId":52914542,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><rnn>","CreationDate":"2018-10-21T10:55:40.420","AcceptedAnswerId":52915108.0,"OwnerUserId":2374138.0,"Title":"Tensorflow 1.11.0 throws an error when trying to reuse cudnnGRU variables with scope.reusevariables()","Body":"<p>I'm implementing an encoder-decoder rnn by using <code>tf.contrib.cudnn_rnn.CudnnGRU()<\/code> as the encoder and I've found a problem:<\/p>\n\n<p>I want to reuse the variables so I can create the same model but use it with other data and to put it simple this would be the code to reproduce my problem:<\/p>\n\n<pre><code>tf.reset_default_graph()\n\ndef create_model():\n    return tf.contrib.cudnn_rnn.CudnnGRU(num_layers=1, num_units=100,\n                         direction='unidirectional')\n\n# (time, batch_size, num_inputs)\nx = tf.random_normal((100, 16, 100))\n\nwith tf.variable_scope('model') as scope:\n    model_1 = create_model()\n    rnn_out_1, rnn_state_1 = model_1(x)\n    scope.reuse_variables()\n    model_2 = create_model()\n    rnn_out_2, rnn_state_2 = model_2(x)\n<\/code><\/pre>\n\n<p>This throws the following error: <\/p>\n\n<blockquote>\n  <p>Variable model\/cudnn_gru_1\/opaque_kernel does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=tf.AUTO_REUSE in VarScope?<\/p>\n<\/blockquote>\n\n<p>So the second model is trying to find the <code>model\/cudnn_gru_1\/opaque_kernel<\/code> variable but can not find it because it should be looking for <code>model\/cudnn_gru\/opaque_kernel:0<\/code>.<\/p>\n\n<p>The thing is I don't know why this is happening since by following the Variables tensorflow reference it seems to be ok. On the other hand, I also tried to write it differently as tensorflow doc states that my above implementation and the one I'm showing next do actually the same:<\/p>\n\n<pre><code>tf.reset_default_graph()\n\ndef create_model():\n    return tf.contrib.cudnn_rnn.CudnnGRU(num_layers=1, num_units=100,\n                         direction='unidirectional')\n\n# (time, batch_size, num_inputs)\nx = tf.random_normal((100, 16, 100))\n\nwith tf.variable_scope('model'):\n    model_1 = create_model()\n    rnn_out_1, rnn_state_1 = model_1(x)\nwith tf.variable_scope('model', reuse=True):\n    model_2 = create_model()\n    rnn_out_2, rnn_state_2 = model_2(x)\n<\/code><\/pre>\n\n<p>This second way is actually working (or at least I think it is). So I don't really know what I'm doing wrong in the first implementation, I'm also not sure on wether both implementations should be doing the same (which I think they should).\nSo does anyone please can help me figure out what I'm doing wrong or the things that I'm not able to understand properly?<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"52915108","CreationDate":"2018-10-21T12:02:17.227","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>CudnnGRU looks like keras-style Model object. So you should reuse the object to share parameters among layers like <\/p>\n\n<pre><code>def create_model():\n    return tf.contrib.cudnn_rnn.CudnnGRU(num_layers=1, num_units=100,\n                                  direction='unidirectional')\n\n\n# (time, batch_size, num_inputs)\nx = tf.random_normal((100, 16, 100))\n\nmodel = create_model()\nrnn_out_1, rnn_state_1 = model(x)\nrnn_out_2, rnn_state_2 = model(x)\n<\/code><\/pre>\n\n<p>I don't know why only the second way is running correctly.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I found CudnnGRU makes variable names for its variables uniquely in its current variable scope. <\/p>\n\n<p>In the first way model_2 makes an new name like cudnn_gru_1 to make its name unique. On the other hand, in the second way you made a new variable scope, so the unique variable names of model_2 matches those of model_1.<\/p>\n\n<p>You can found why CudnnGRU makes unique variable name in Layer._set_scope() (tensorflow\\python\\layers\\base.py#L150). Layer class makes a new variable scope for its variable with default_name argument (scope is None in this case), so its name become unique.<\/p>\n"}]}
+{"QuestionId":52914571,"AnswerCount":0,"Tags":"<keras><embedding><seq2seq>","CreationDate":"2018-10-21T10:58:28.400","AcceptedAnswerId":null,"OwnerUserId":7641859.0,"Title":"Embedding version seq2seq model (Keras)","Body":"<p>I want to build up an embedding version seq2seq model by modifying the example on keras github.\n<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/lstm_seq2seq.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/lstm_seq2seq.py<\/a> <\/p>\n\n<p>I've tried the np.reshape but it won't work.<\/p>\n\n<pre><code>from keras.layers.embeddings import Embedding\n\nembedding = 100\nvocab_size = 10000\n\nencoder_inputs = Input(shape=(None,), name=\"Encoder_input\")\nencoder = LSTM(latent_dim, return_state=True, name='Encoder_lstm') \nShared_Embedding = Embedding(output_dim=embedding, input_dim=vocab_size, name=\"Embedding\") \nword_embedding_context = Shared_Embedding(encoder_inputs) \nencoder_outputs, state_h, state_c = encoder(word_embedding_context) \nencoder_states = [state_h, state_c] \n\ndecoder_inputs = Input(shape=(None,))\ndecoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True, name=\"Decoder_lstm\") \nword_embedding_answer = Shared_Embedding(decoder_inputs) \ndecoder_outputs, _, _ = decoder_lstm(word_embedding_answer, initial_state=encoder_states) \n\n\ndecoder_dense = Dense(vocab_size, activation='softmax', name=\"Dense_layer\") \ndecoder_outputs = decoder_dense(decoder_outputs) \n\n\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs) \n\n\n\nmodel.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy')\n\n# Note that `decoder_target_data` needs to be one-hot encoded,\n\n# rather than sequences of integers like `decoder_input_data`!\n\nmodel.fit([encoder_input_data, decoder_input_data], decoder_target_data,\n\n          batch_size=batch_size,\n\n          epochs=epochs,\n\n          validation_split=0.2)\nmodel.save('s2s.h5')\n<\/code><\/pre>\n\n<p>However, I got the following error messages, can anyone help me ? \nThank a lot!<\/p>\n\n<p>ValueError: Error when checking input: expected Encoder_input to have 2 dimensions, but got array with shape (4999, 53, 3132)<\/p>\n","answers":[]}
+{"QuestionId":52914942,"AnswerCount":0,"Tags":"<python><tensorflow><keras><ensemble-learning>","CreationDate":"2018-10-21T11:42:15.363","AcceptedAnswerId":null,"OwnerUserId":9969431.0,"Title":"Using Neural networks as a base learner in SuperLearner ensemble model","Body":"<p>I would like to use Neural Network from Keras as an estimator for mlens superlearner but I get the below error:<\/p>\n\n<pre><code>&lt;keras.engine.sequential.Sequential object at 0x7f395d6e2e48&gt;' does not \nappear to be a valid estimator as it does not implement a 'get_params' \nmethod. Type: &lt;class 'keras.engine.sequential.Sequential'&gt;\n<\/code><\/pre>\n\n<p>Anyone knows how to solve this issue?<\/p>\n\n<p>Below is my code:<\/p>\n\n<pre><code>from mlens.ensemble import SuperLearner\nsl = SuperLearner(\nfolds=10,\nrandom_state=1,\nverbose=2,\nbackend=\"multiprocessing\"\n )\n  # Neural Network: \n  model=Sequential()\n  model.add(Dense(85,activation='relu',input_shape=(49,)))\n  model.add(Dense(85, activation='relu')) \n  model.add(Dense(85, activation='relu')) \n  model.add(Dense(1,activation='sigmoid'))\n\n  models={\n\n \"nn\":model\n  }\n sl.fit(X_train, Y_train, batch_size=64, epochs=1000, verbose=0)'\n\n meta_learner = Ridge(\n solver='auto', fit_intercept=True, alpha=1.0,\n max_iter=100, normalize=False, tol=0.05, random_state=1,\n )\n\nsl.add(nn)\nsl.add_meta(meta_learner)\nsl.fit(X_train, Y_train, batch_size=64, epochs=1000, verbose=0)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52915150,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-21T12:07:25.063","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Keras strange crash","Body":"<p>Training model in Keras I get exception at some epoch:<\/p>\n\n<pre><code>Exception ignored in: &lt;bound method BaseSession.__del__ of &lt;tensorflow.python.client.session.Session object at 0x7f036d74b7f0&gt;&gt;\nTraceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 686, in __del__\nTypeError: 'NoneType' object is not callable\n<\/code><\/pre>\n\n<p>What could be reason of such behaviour?<\/p>\n","answers":[{"AnswerId":"55014865","CreationDate":"2019-03-06T03:01:19.373","ParentId":null,"OwnerUserId":"11152597","Title":null,"Body":"<p>This might be caused by the Cache file, you can try deleting the <strong>pycache<\/strong> folder. This works for me.\nI am using the Tensorflow version.<\/p>\n"}]}
+{"QuestionId":52915515,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-21T12:53:06.820","AcceptedAnswerId":null,"OwnerUserId":10536144.0,"Title":"Keras loss value not changing","Body":"<p>I am trying to apply a deep learning network on the Loan status dataset to check if I can get a better result than conventional machine learning algorithms.<\/p>\n\n<p>Accuracy seems to be very low (even lower than using normal logistic regression). How can I improve it?<\/p>\n\n<p>Things I've tried:\n- Changing the learning rate\n- increasing number of layers\n- increase\/decreasing number of nodes**<\/p>\n\n<pre><code>X = df_dummies.drop('Loan_Status', axis=1).values\ny = df_dummies['Loan_Status'].values\nmodel = Sequential()\n\nmodel.add(Dense(50, input_dim = 17, activation = 'relu'))\nmodel.add(Dense(100, activation = 'relu'))\nmodel.add(Dense(100, activation = 'relu'))\nmodel.add(Dense(100, activation = 'relu'))\nmodel.add(Dense(100, activation = 'relu'))\nmodel.add(Dense(1, activation = 'sigmoid'))\n\nsgd = optimizers.SGD(lr = 0.00001)\n\nmodel.compile(optimizer=sgd, loss='binary_crossentropy', metrics=`['accuracy'])`\n\nmodel.fit(X, y, epochs = 50, shuffle=True, verbose=2)\nmodel.summary()\n<\/code><\/pre>\n\n<blockquote>\n  <p>Epoch 1\/50\n       - 1s - loss: 4.9835 - acc: 0.6873\n      Epoch 2\/50\n       - 0s - loss: 4.9830 - acc: 0.6873\n      Epoch 3\/50\n       - 0s - loss: 4.9821 - acc: 0.6873\n      Epoch 4\/50\n       - 0s - loss: 4.9815 - acc: 0.6873\n      Epoch 5\/50\n       - 0s - loss: 4.9807 - acc: 0.6873\n      Epoch 6\/50\n       - 0s - loss: 4.9800 - acc: 0.6873\n      Epoch 7\/50\n       - 0s - loss: 4.9713 - acc: 0.6873\n      Epoch 8\/50\n       - 0s - loss: 8.5354 - acc: 0.4397\n      Epoch 9\/50\n       - 0s - loss: 4.8322 - acc: 0.6743\n      Epoch 10\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 11\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 12\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 13\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 14\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 15\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 16\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 17\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 18\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 19\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 20\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 21\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 22\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 23\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 24\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 25\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 26\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 27\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 28\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 29\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 30\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 31\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 32\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 33\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 34\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 35\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 36\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 37\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 38\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 39\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 40\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 41\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 42\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 43\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 44\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 45\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 46\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 47\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 48\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 49\/50\n       - 0s - loss: 4.9852 - acc: 0.6873\n      Epoch 50\/50\n       - 0s - loss: 4.9852 - acc: 0.6873<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ndense_19 (Dense)             (None, 50)                900       \n_________________________________________________________________\ndense_20 (Dense)             (None, 100)               5100      \n_________________________________________________________________\ndense_21 (Dense)             (None, 100)               10100     \n_________________________________________________________________\ndense_22 (Dense)             (None, 100)               10100     \n_________________________________________________________________\ndense_23 (Dense)             (None, 100)               10100     \n_________________________________________________________________\ndense_24 (Dense)             (None, 1)                 101       \n=================================================================\nTotal params: 36,401\nTrainable params: 36,401\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n<\/blockquote>\n","answers":[{"AnswerId":"52916332","CreationDate":"2018-10-21T14:27:11.020","ParentId":null,"OwnerUserId":"10536144","Title":null,"Body":"<p>**I was able to get a slight improvement by making the network deeper &amp; adding dropouts, but I still think this can be improved further as using normal logistic regression gives a much better accuracy (80%+).<\/p>\n\n<p>Does anyone know a way to achieve further improvement?**<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Dense(1000, input_dim = 17, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1000, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1000, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1000, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1000, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1000, activation = 'relu'))\nmodel.add(Dense(1, activation = 'sigmoid'))\n\nsgd = optimizers.SGD(lr = 0.0001)\n\nmodel.compile(optimizer=sgd, loss='binary_crossentropy', metrics=['accuracy'])\n\n\n\nmodel.fit(X_train, y_train, epochs = 20, shuffle=True, verbose=2, batch_size=30)\n\n\n\nEpoch 1\/20\n - 2s - loss: 4.8965 - acc: 0.6807\nEpoch 2\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 3\/20\n - 1s - loss: 4.6091 - acc: 0.7040\nEpoch 4\/20\n - 1s - loss: 4.5642 - acc: 0.7040\nEpoch 5\/20\n - 1s - loss: 4.6937 - acc: 0.7040\nEpoch 6\/20\n - 1s - loss: 4.6830 - acc: 0.7063\nEpoch 7\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 8\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 9\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 10\/20\n - 1s - loss: 4.6452 - acc: 0.7086\nEpoch 11\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 12\/20\n - 1s - loss: 4.6824 - acc: 0.7063\nEpoch 13\/20\n - 1s - loss: 4.7200 - acc: 0.7040\nEpoch 14\/20\n - 1s - loss: 4.6608 - acc: 0.7063\nEpoch 15\/20\n - 1s - loss: 4.6940 - acc: 0.7040\nEpoch 16\/20\n - 1s - loss: 4.7136 - acc: 0.7040\nEpoch 17\/20\n - 1s - loss: 4.6056 - acc: 0.7063\nEpoch 18\/20\n - 1s - loss: 4.5640 - acc: 0.7016\nEpoch 19\/20\n - 1s - loss: 4.7009 - acc: 0.7040\nEpoch 20\/20\n - 1s - loss: 4.6892 - acc: 0.7040\n<\/code><\/pre>\n"}]}
+{"QuestionId":52915913,"AnswerCount":1,"Tags":"<python><sqlite><keras>","CreationDate":"2018-10-21T13:41:01.370","AcceptedAnswerId":52917812.0,"OwnerUserId":2722607.0,"Title":"Using sqlite with keras generator","Body":"<p>I'm trying to batch train cnn with keras generator.\nMy data is in SQLITE db.\nSQLITE can't work with multithreaded code so each time i need to import batch i open connection to db, next i try to execute simple sql query (The query is executed with out error in simple script)  and i'm getting this error:<\/p>\n\n<pre><code>File \"&lt;ipython-input-4-6c84648166ec&gt;\", line 1, in &lt;module&gt;\n    db_cursor.execute(sql_query)\nTypeError: convert_array() takes 1 positional argument but 2 were given\n<\/code><\/pre>\n\n<p>My code:<\/p>\n\n<pre><code>import numpy as np\nimport keras\nimport sqlite3\nimport io\n\nclass DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, path, num_features, end_index, start_index=1,    augmentation_ratio=2, batch_size=250,\n                 shuffle=True, exclude = []):\n        'Initialization'\n        sqlite3.register_adapter(np.ndarray, self.adapt_array)\n        sqlite3.register_converter(\"array\", self.convert_array)\n        self.path = path\n        self.db = sqlite3.connect(self.path, detect_types=sqlite3.PARSE_DECLTYPES, check_same_thread=False)\n        self.db_cursor = self.db.cursor()\n        self.exclude = exclude\n        self.N = num_features\n        self.start_index = start_index\n        self.end_index = end_index\n        self.augmentation_ratio = augmentation_ratio\n        self.batch_size = int(batch_size)\n        self.shuffle = shuffle\n        self.import_size = int(np.floor(self.batch_size \/ self.augmentation_ratio))\n        if exclude.__len__() == 0:\n            self.sample_index = np.arange(self.start_index, self.end_index)\n        else:\n            query = 'SELECT ind FROM TABLE WHERE ind &gt; ? AND ind &lt; ? AND class NOT IN ({})'. \\\n                format(','.join(str(label) for label in self.exclude))\n            self.db_cursor.execute(query, [self.start_index, self.end_index])\n            self.sample_index = np.asarray(self.db_cursor.fetchall())\n        self.n_samples = self.sample_index.shape[0]\n        self.on_epoch_end()\n        self.db.close()\n\n\n\n\n    def __len__(self):\n        'Denotes the number of batches per epoch'\n        return int(np.floor(self.n_samples \/ self.import_size))\n\n    def __getitem__(self, index):\n        'Generate one batch of data'\n        # Generate indexes of the batch\n        samples_batch = np.arange((index - 1) * self.import_size, index * self.import_size)\n\n        # Generate data\n        x, y = self.__data_generation(samples_batch)\n\n        augmented_x = self.augment_data(x)\n\n        samples = np.concatenate((x, augmented_x))\n        y_mat = np.concatenate((y, y))\n        # return X_reshaped, xData_complete\n        return  samples, y_mat\n\n    def on_epoch_end(self):\n        'Updates indexes after each epoch'\n        if self.shuffle:\n            np.random.shuffle(self.sample_index)\n\n    def __data_generation(self, samples_batch):\n        'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n        # Initialization\n        x = np.empty((self.import_size, self.N, 1))\n        y = np.empty((self.import_size, 1))\n        inds = self.sample_index[samples_batch]\n        sqlite3.register_adapter(np.ndarray, self.adapt_array)\n        sqlite3.register_converter(\"array\", self.convert_array)\n        db = sqlite3.connect(self.path, detect_types=sqlite3.PARSE_DECLTYPES)\n        db_cursor = db.cursor()\n        sql_query = \"SELECT class,features FROM TABLE WHERE ind in ({})\".\\\n            format(','.join(str(ind[0]) for ind in inds))\n        db_cursor.execute(sql_query)\n        for i in range(self.import_size):\n            line = db_cursor.fetchone()\n            y[i] = line[0]\n            x[i, :] = line[1]\n        y = keras.utils.to_categorical(y)\n        return x, y\n\n    def adapt_array(arr):\n        out = io.BytesIO()\n        np.save(out, arr)\n        out.seek(0)\n        return sqlite3.Binary(out.read())\n\n    def convert_array(text):\n        out = io.BytesIO(text)\n        out.seek(0)\n        return np.load(out)\n<\/code><\/pre>\n\n<p>What am i doing wrong?<\/p>\n","answers":[{"AnswerId":"52917812","CreationDate":"2018-10-21T17:07:10.733","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>Your <code>adapt_array<\/code> and <code>convert_array<\/code> methods are missing the <code>self<\/code> argument:<\/p>\n\n<pre><code>def convert_array(self, text):\n    out = io.BytesIO(text)\n    out.seek(0)\n    return np.load(out)\n<\/code><\/pre>\n\n<p>Alternatively, you can use the <code>@staticmethod<\/code> decorator:<\/p>\n\n<pre><code>@staticmethod\ndef convert_array(text):\n    out = io.BytesIO(text)\n    out.seek(0)\n    return np.load(out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52916351,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-21T14:29:26.310","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Keras TensorBoard showing incorrect Recall\/Precision metrics","Body":"<p>The idea is to set up TensorBoard with good metrics to track my model training and evaluate model performance. However, most of the metrics seem incorrect (especially precision\/recall\/false_negative_count) I am using tf v1.8.0.<\/p>\n\n<p>To <strong>create the metrics<\/strong>, I use the following peace of code:<\/p>\n\n<pre><code>def as_keras_metric(method, **kwargs):\n    @functools.wraps(method)\n    def wrapper(self, args):\n        \"\"\" Wrapper for turning tensorflow metrics into keras metrics \"\"\"\n        value, update_op = method(self, args, **kwargs)\n        tf.keras.backend.get_session().run(tf.local_variables_initializer())\n        with tf.control_dependencies([update_op]):\n            value = tf.identity(value)\n        return value\n    return wrapper\n\n\ndef create_metrics():\n    auc_roc = as_keras_metric(tf.metrics.auc)\n    recall = as_keras_metric(tf.metrics.recall)\n    precision = as_keras_metric(tf.metrics.precision)\n\n    fn = as_keras_metric(tf.metrics.false_negatives)\n    fp = as_keras_metric(tf.metrics.false_positives)\n\n    kwargs = {'specificity': 0.95}\n    sensitivity_at_specificity = as_keras_metric(tf.metrics.sensitivity_at_specificity, **kwargs)\n\n    mean_per_class_accuracy = as_keras_metric(tf.metrics.mean_per_class_accuracy, **{'num_classes': 2})\n\n    return [fn, fp, auc_roc, recall, precision, sensitivity_at_specificity, mean_per_class_accuracy, tf.keras.metrics.binary_accuracy]\n<\/code><\/pre>\n\n<p>I compile the model using:<\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy', optimizer='adadelta', metrics=create_metrics())\n<\/code><\/pre>\n\n<p>and fit using <\/p>\n\n<p><code>model.fit_generator(generator=KerasSequence)<\/code><\/p>\n\n<h2>Now here is the problem:<\/h2>\n\n<p>My data is imbalanced, I have around 2% of observations belonging to the +ve class. Let's look at the last few epochs.<\/p>\n\n<blockquote>\n  <p>Epoch 16\/50 6\/7 [========================>.....] - ETA: 0s - loss:\n      5.9874 - auc: 0.9171 - recall: 1.0000 - precision: 0.0068 - sensitivity_at_specificity: 0.8050 - mean_per_class_accuracy: 0.5000 -\n  false_negatives: 0.0000e+00 - false_positives: 13570.6667 -\n  binary_accuracy: 1.0000<\/p>\n  \n  <p>Epoch 50\/50 6\/7 [========================>.....] - ETA: 0s - loss:\n  0.0205 - auc:\n      0.9207 - recall: 1.0000 - precision: 0.0069 - sensitivity_at_specificity: 0.8494 - mean_per_class_accuracy: 0.5000 -\n  false_negatives: 0.0000e+00 - false_positives: 43821.6667 -\n  binary_accuracy: 1.0000<\/p>\n<\/blockquote>\n\n<p>I know that 6 out of 500 samples were predicted as +ve class. So I don't understand how it's possible to have these metrics. They are mutually exclusive. Recall at 1 and precision so low. Which suggests that I actually predict 1 for all observations, when in reality only 5\/500 are predicted to be +ve class. Additionaly, how is it possible to get <code>false_positive<\/code> count at 43000... when I only have 500 training samples. And it's growing with every epoch<\/p>\n\n<h2>Screenshots<\/h2>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jSx3e.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jSx3e.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The loss goes to 0 (overfitting), but the false positive count increases per epoch... why? Interestingly, after the 1st epoch, false positive count is at 200. This makes me thing that the metrics don't reset after epoch end and keep adding.<\/p>\n","answers":[]}
+{"QuestionId":52916736,"AnswerCount":0,"Tags":"<opencv><tensorflow><keras><conv-neural-network><cv2>","CreationDate":"2018-10-21T15:09:36.693","AcceptedAnswerId":null,"OwnerUserId":7865134.0,"Title":"error: (-215:Assertion failed) !ssize.empty() in function 'cv::resize'","Body":"<p>This code is for classify 5 different classes of images. I have tried to use ImageDataGenerator() for image augmentation but when i run this code, I am havinh this error \"error: (-215:Assertion failed) !ssize.empty() in function 'cv::resize'\". <\/p>\n\n<p>How can i fix this error?\nAlso I will appreciate if anyone can make other suitable changes to the code.<\/p>\n\n<pre><code>import cv2                \nimport numpy as np         \nimport os                \nfrom random import shuffle \nfrom tqdm import tqdm      \n#from tensorflow import keras\n#from tensorflow.contrib import lite\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.models import Sequential\n\nTRAIN_DIR = 'train'\nVALID_DIR = 'validate'\nTEST_DIR = 'test'\nIMG_SIZE = 28\nLR = 1e-3\ntrain_samples = 250\nvaldate_samples = 250\nepochs = 5\nbatch_size = 10\n\nMODEL_NAME = 'snakes-{}-{}.model'.format(LR, '2conv-basic')\n\ndef label_img(img):\n    print(\"\\nImage = \",img)\n    print(\"\\n\",img.split('.')[-2])\n    temp_name= img.split('.')[-2]\n    #print(\"\\n\",temp_name[0:3])\n    #temp_name=temp_name[0:3]\n    print(\"\\n\",temp_name[:1])\n    temp_name=temp_name[:1]\n    #word_label = img.split('.')[-3]\n    word_label = temp_name\n\n   # word_label = img[0]\n\n    if word_label == 'A': return [0,0,0,0,1]    \n    elif word_label == 'B': return [0,0,0,1,0]\n    elif word_label == 'C': return [0,0,1,0,0]\n    elif word_label == 'D': return [0,1,0,0,0]\n    elif word_label == 'E' : return [1,0,0,0,0]   \n\ndef create_train_data():\n    training_data = []\n    for img in tqdm(os.listdir(TRAIN_DIR)):\n        label = label_img(img)\n        path = os.path.join(TRAIN_DIR,img)\n        img = cv2.imread(path,cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))\n        training_data.append([np.array(img),np.array(label)])\n    shuffle(training_data)\n    np.save('train_data.npy', training_data)\n    return training_data\n\n\ndef create_validate_data():\n    validating_data = []\n    for img in tqdm(os.listdir(TRAIN_DIR)):\n        label = label_img(img)\n        path = os.path.join(VALID_DIR,img)\n        img = cv2.imread(path,cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))\n        validating_data.append([np.array(img),np.array(label)])\n    shuffle(validating_data)\n    np.save('validate_data.npy', validating_data)\n    return validating_data\n\n\ndef process_test_data():\n    testing_data = []\n    for img in tqdm(os.listdir(TEST_DIR)):\n        path = os.path.join(TEST_DIR,img)\n        img_num = img.split('.')[0]\n        img = cv2.imread(path,cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))\n        testing_data.append([np.array(img), img_num])\n    shuffle(testing_data)\n    np.save('test_data.npy', testing_data)\n    return testing_data\n\ntrain_data = create_train_data()\nvalidate_data = create_validate_data()\n\n\nimport tflearn\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\n\n\n\n\nimport tensorflow as tf\ntf.reset_default_graph()\n\n\nconvnet = input_data(shape=[None, IMG_SIZE, IMG_SIZE, 1], name='input')\n\nconvnet = conv_2d(convnet, 32, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\n\nconvnet = conv_2d(convnet, 64, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\n\nconvnet = conv_2d(convnet, 128, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\n\nconvnet = conv_2d(convnet, 64, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\n\nconvnet = conv_2d(convnet, 32, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\n\nconvnet = fully_connected(convnet, 1024, activation='relu')\nconvnet = dropout(convnet, 0.8)\n\nconvnet = fully_connected(convnet, 5, activation='softmax')\nconvnet = regression(convnet, optimizer='adam', learning_rate=LR, loss='categorical_crossentropy', name='targets')\n\nmodel = tflearn.DNN(convnet, tensorboard_dir='log')\n\n\n\nif os.path.exists('{}.meta'.format(MODEL_NAME)):\n    model.load(MODEL_NAME)\n    print('model loaded!')\n\n#train = train_data[:-500]\n#test = train_data[-500:]\n\n#train = train_data[:-200]\n#test = train_data[-200:]\ntrain = train_data[0:]\nvalidate = validate_data[0:]\n\nX = np.array([i[0] for i in train]).reshape(-1,IMG_SIZE,IMG_SIZE,1)\nY = [i[1] for i in train]\n\n#test_x = np.array([i[0] for i in test]).reshape(-1,IMG_SIZE,IMG_SIZE,1)\n#test_y = [i[1] for i in test]\nvalidate_x = np.array([i[0] for i in validate]).reshape(-1,IMG_SIZE,IMG_SIZE,1)\nvalidate_y = [i[1] for i in validate]\n\nmodel.fit({'input': X}, {'targets': Y}, n_epoch=epochs, validation_set=({'input': validate_x}, {'targets': validate_y}), \n    snapshot_step=500, show_metric=True, run_id=MODEL_NAME)\n\n\ntrain_datagen = ImageDataGenerator(\n    rescale=1. \/ 255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)\n\nvalidaton_datagen = ImageDataGenerator(rescale=1. \/ 255)\n\ntrain_generator = train_datagen.flow_from_directory('data\/train',\n                                                    target_size=(IMG_SIZE),\n                                                    batch_size=batch_size,\n                                                    class_mode='categorical')\n\nvalidation_generator = validaton_datagen.flow_from_directory('data\/validation',\n                                                        target_size=(IMG_SIZE),\n                                                        batch_size=batch_size,\n                                                        class_mode='categorical')\n\nmodel.fit_generator(train_generator,\n                    steps_per_epoch=25,\n                    epochs=epochs,\n                    validation_data=validation_generator,\n                    validation_steps=25)\n\n\nmodel.save(MODEL_NAME)\n\n#keras_file = \"linear.h5\"\n#keras.models.save_model(model,keras_file)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52916992,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-21T15:38:32.623","AcceptedAnswerId":null,"OwnerUserId":10536694.0,"Title":"Word frequencies in Python instead of listing","Body":"<p>I have the following code:<\/p>\n\n<pre><code>from keras.preprocessing import text\n\nwith open('engl_bible.txt', 'r') as file:\nnorm_bible = file\n\ntokenizer = text.Tokenizer()\ntokenizer.fit_on_texts(norm_bible)\n\nword2id = tokenizer.word_index\nid2word = {v:k for k, v in word2id.items()}\n\nvocab_size = len(word2id) + 1\nembed_size = 100\n\nwids = [[word2id[w] for w in text.text_to_word_sequence(doc)] for doc in norm_bible]\n\nprint('Vocabulary Size:', vocab_size)\n<\/code><\/pre>\n\n<p>print('Vocabulary Sample:', list(word2id.items())[:10])<\/p>\n\n<p>This leads to following output:<\/p>\n\n<pre><code>Vocabulary Size: 3847\nVocabulary Sample: [('and', 1), ('the', 2), ('to', 3), ('of', 4), ('you', 5), ('he', 6), ('in', 7), ('a', 8), ('is', 9), ('him', 10)]\n<\/code><\/pre>\n\n<p>But it should create something comparable to (so not a ranking of 1 to 10, but with word frequencies) :<\/p>\n\n<pre><code>Vocabulary Size: 12425\nVocabulary Sample: [('perceived', 1460), ('flagon', 7287), ('gardener', 11641), ('named', 973), ('remain', 732), ('sticketh', 10622), ('abstinence', 11848), ('rufus', 8190), ('adversary', 2018), ('jehoiachin', 3189)]\n<\/code><\/pre>\n\n<p>I really can't figure out what went wrong, hope you can help me! Thanks so much!<\/p>\n","answers":[{"AnswerId":"52918271","CreationDate":"2018-10-21T17:59:14.263","ParentId":null,"OwnerUserId":"4404709","Title":null,"Body":"<p>if you want the word frequencies you need to use <code>tokenizer.word_counts<\/code> and not <code>tokenizer.word_index<\/code>.\nSo the code will be:<\/p>\n\n<pre><code>from keras.preprocessing import text\n\nwith open('engl_bible.txt', 'r') as file:\n    norm_bible = file\n\n    tokenizer = text.Tokenizer()\n    tokenizer.fit_on_texts(norm_bible)\n\n    word2id = tokenizer.word_counts\n    id2word = {v:k for k, v in word2id.items()}\n\n    vocab_size = len(word2id) + 1\n    embed_size = 100\n\n    wids = [[word2id[w] for w in text.text_to_word_sequence(doc)] for doc in norm_bible]\n\n    print('Vocabulary Size:', vocab_size)\n    print('Vocabulary Sample:', list(word2id.items())[:10])\n<\/code><\/pre>\n\n<p>pay attention though that now <code>word2id<\/code> isn't really word2id but rather <code>word2frequency<\/code>...<\/p>\n"}]}
+{"QuestionId":52917486,"AnswerCount":0,"Tags":"<python><tensorflow><rnn>","CreationDate":"2018-10-21T16:30:06.503","AcceptedAnswerId":null,"OwnerUserId":10536775.0,"Title":"InvalidArgumentError: Input to reshape is a tensor with 12000 values, but the requested shape requires a multiple of 140","Body":"<p>I've tried to calculate the dimensions of each step in detail, trying to find the problem, but failed, and the run network is a challenge for beginners like me.<\/p>\n\n<p>In all, there are 5,800 rows, each row of data has seven dimensions and each row of data corresponds to a one-dimensional label.These data are text time series data.<\/p>\n\n<pre><code>normalized_train_data.shape\n#(5800, 8)\n<\/code><\/pre>\n\n<p>I chose the step length of each training is 20 and the batch size is 60,and the input size is 7,output size is 1,The code is shown below.<\/p>\n\n<pre><code>batch_index = []\ntrain_x,train_y = [],[]\ntime_step = 20\nbatch_size = 60\nfor i in range(len(normalized_train_data)-time_step):\n    if i%batch_size == 0:\n        batch_index.append(i)\n    x = normalized_train_data[i:i+time_step,:7]\n    y = normalized_train_data[i:i+time_step,7,np.newaxis]\n    train_x.append(x.tolist())\n    train_y.append(y.tolist())\nbatch_index.append(len(normalized_train_data)-time_step)\ninput_size = 7\nrnn_unit = 10\noutput_size = 1\ntime_step = 20\nbatch_size = 60\nlearning_rate = 1e-4\n\nweights = {\n'in':tf.Variable(tf.random_normal([input_size,rnn_unit])),\n'out':tf.Variable(tf.random_normal([rnn_unit,output_size]))\n}\nbiases = {\n'in':tf.Variable(tf.constant(value=0.1,shape=[rnn_unit,])),\n'out':tf.Variable(tf.constant(value=0.1,shape=[output_size,]))\n}\n\ndef RNN(inputs):\n    w_in = weights['in']\n    b_in = biases['in']\n    batch_size = tf.shape(inputs)[0]\n    time_step = tf.shape(inputs)[1]\n    input_x = tf.reshape(inputs,shape=[-1,input_size])\n    input_rnn = tf.matmul(input_x,w_in) + b_in\n    input_rnn = tf.reshape(input_rnn,[-1,time_step,input_size])\n    cell = tf.nn.rnn_cell.BasicRNNCell(rnn_unit)\n    init_state = cell.zero_state(batch_size,dtype=tf.float32)\n    output_rnn,final_states = tf.nn.dynamic_rnn(cell=cell,inputs=input_rnn,initial_state=init_state,dtype=tf.float32)\n    output = tf.reshape(output_rnn,[-1,rnn_unit])\n    w_out = weights['out']\n    b_out = biases['out']\n    pred = tf.matmul(output,w_out)+b_out\n    return pred,final_states\n\n\nX_holder = tf.placeholder(tf.float32,[None,time_step,input_size])\nY_holder = tf.placeholder(tf.float32,[None,time_step,output_size])\n\npred,_ = RNN(X_holder)\nloss = tf.reduce_mean(tf.square(tf.reshape(pred,[-1])-tf.reshape(Y_holder,\n[-1])))\ntrain = tf.train.AdamOptimizer(learning_rate).minimize(loss)\n# saver = tf.trian.saver(tf.global_variables,max_to_keep=15)\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n<\/code><\/pre>\n\n<p>The training process code is as follows:<\/p>\n\n<pre><code>total_loss = []\nfor i in range(1000):\n    for step in range(len(batch_index)-1):\n        _,loss_ = sess.run([train,loss],feed_dict={X_holder:train_x[batch_index[step]:batch_index[step+1]],Y_holder:train_y[batch_index[step]:batch_index[step+1]]})\n        total_loss.append(loss_)\n<\/code><\/pre>\n\n<p>The final error is as follows:<\/p>\n\n<pre><code>    InvalidArgumentError      Traceback (most recent call last)  \nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_call(self, fn, *args)\n   1360     try:\n-&gt; 1361       return fn(*args)\n   1362     except errors.OpError as e:\n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run_fn(session, feed_dict, fetch_list, target_list, options, run_metadata)\n   1339           return tf_session.TF_Run(session, options, feed_dict, fetch_list,\n-&gt; 1340                                    target_list, status, run_metadata)\n   1341 \n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)\n    515             compat.as_text(c_api.TF_Message(self.status.status)),\n--&gt; 516             c_api.TF_GetCode(self.status.status))\n    517     # Delete the underlying status object from memory otherwise it stays alive\n\nInvalidArgumentError: Input to reshape is a tensor with 12000 values, but the requested shape requires a multiple of 140\n     [[Node: Reshape_2 = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](add, Reshape_1\/shape)]]\n     [[Node: Mean\/_11 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_505_Mean\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nDuring handling of the above exception, another exception occurred:\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-34-c6007dcf9016&gt; in &lt;module&gt;()\n      2 for i in range(1000):\n      3     for step in range(len(batch_index)-1):\n----&gt; 4         _,loss_ = sess.run([train,loss],feed_dict={X_holder:train_x[batch_index[step]:batch_index[step+1]],Y_holder:train_y[batch_index[step]:batch_index[step+1]]})\n      5         total_loss.append(loss_)\n      6         if i % 10==0:\n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in run(self, fetches, feed_dict, options, run_metadata)\n    903     try:\n    904       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 905                          run_metadata_ptr)\n    906       if run_metadata:\n    907         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1135     if final_fetches or final_targets or (handle and feed_dict_tensor):\n   1136       results = self._do_run(handle, final_targets, final_fetches,\n-&gt; 1137                              feed_dict_tensor, options, run_metadata)\n   1138     else:\n   1139       results = []\n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)\n   1353     if handle is None:\n   1354       return self._do_call(_run_fn, self._session, feeds, fetches, targets,\n-&gt; 1355                            options, run_metadata)\n   1356     else:\n   1357       return self._do_call(_prun_fn, self._session, handle, feeds, fetches)\n\nJ:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_call(self, fn, *args)\n   1372         except KeyError:\n   1373           pass\n-&gt; 1374       raise type(e)(node_def, op, message)\n   1375 \n   1376   def _extend_graph(self):\n\nInvalidArgumentError: Input to reshape is a tensor with 12000 values, but the requested shape requires a multiple of 140\n     [[Node: Reshape_2 = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](add, Reshape_1\/shape)]]\n     [[Node: Mean\/_11 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_505_Mean\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nCaused by op 'Reshape_2', defined at:\n  File \"J:\\Anaconda3\\lib\\runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"J:\\Anaconda3\\lib\\runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel_launcher.py\", line 16, in &lt;module&gt;\n    app.launch_new_instance()\n  File \"J:\\Anaconda3\\lib\\site-packages\\traitlets\\config\\application.py\", line 658, in launch_instance\n    app.start()\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\kernelapp.py\", line 486, in start\n    self.io_loop.start()\n  File \"J:\\Anaconda3\\lib\\site-packages\\tornado\\platform\\asyncio.py\", line 127, in start\n    self.asyncio_loop.run_forever()\n  File \"J:\\Anaconda3\\lib\\asyncio\\base_events.py\", line 422, in run_forever\n    self._run_once()\n  File \"J:\\Anaconda3\\lib\\asyncio\\base_events.py\", line 1432, in _run_once\n    handle._run()\n  File \"J:\\Anaconda3\\lib\\asyncio\\events.py\", line 145, in _run\n    self._callback(*self._args)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tornado\\platform\\asyncio.py\", line 117, in _handle_events\n    handler_func(fileobj, events)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tornado\\stack_context.py\", line 276, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"J:\\Anaconda3\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 450, in _handle_events\n    self._handle_recv()\n  File \"J:\\Anaconda3\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 480, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"J:\\Anaconda3\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 432, in _run_callback\n    callback(*args, **kwargs)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tornado\\stack_context.py\", line 276, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 283, in dispatcher\n    return self.dispatch_shell(stream, msg)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 233, in dispatch_shell\n    handler(stream, idents, msg)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 399, in execute_request\n    user_expressions, allow_stdin)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\ipkernel.py\", line 208, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"J:\\Anaconda3\\lib\\site-packages\\ipykernel\\zmqshell.py\", line 537, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"J:\\Anaconda3\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2662, in run_cell\n    raw_cell, store_history, silent, shell_futures)\n  File \"J:\\Anaconda3\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2785, in _run_cell\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"J:\\Anaconda3\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2903, in run_ast_nodes\n    if self.run_code(code, result):\n  File \"J:\\Anaconda3\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2963, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-17-ed57b1fae759&gt;\", line 1, in &lt;module&gt;\n    input_rnn = tf.reshape(input_rnn,[-1,time_step,input_size])\n  File \"J:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 5101, in reshape\n    \"Reshape\", tensor=tensor, shape=shape, name=name)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3271, in create_op\n    op_def=op_def)\n  File \"J:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1650, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): Input to reshape is a tensor with 12000 values, but the requested shape requires a multiple of 140\n     [[Node: Reshape_2 = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](add, Reshape_1\/shape)]]\n     [[Node: Mean\/_11 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_505_Mean\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>This problem has been bothering me for a long time. No matter how hard I try, I can never solve it.Please help to solve it. Thank you very much.<\/p>\n","answers":[]}
+{"QuestionId":52918051,"AnswerCount":10,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-21T17:34:00.130","AcceptedAnswerId":null,"OwnerUserId":9479655.0,"Title":"How to convert .pb to TFLite format?","Body":"<p>I downloaded a <code>retrained_graph.pb<\/code> and <code>retrained_labels.txt<\/code> file of a model I trained in Azure cognitive service. Now I want to make an Android app using that model and to do so I have to convert it to TFLite format. I used toco and I am getting the following error:<\/p>\n\n<pre><code>ValueError: Invalid tensors 'input' were found.\n<\/code><\/pre>\n\n<p>I am basically following this tutorial and have problem on step 4 and direcly \ncopy pasted the terminal code:\n<a href=\"https:\/\/heartbeat.fritz.ai\/neural-networks-on-mobile-devices-with-tensorflow-lite-a-tutorial-85b41f53230c\" rel=\"noreferrer\">https:\/\/heartbeat.fritz.ai\/neural-networks-on-mobile-devices-with-tensorflow-lite-a-tutorial-85b41f53230c<\/a><\/p>\n","answers":[{"AnswerId":"57733601","CreationDate":"2019-08-30T21:54:57.720","ParentId":null,"OwnerUserId":"6114891","Title":null,"Body":"<p>Substituting <code>Mul<\/code> for <code>input<\/code> fixed it for me.<\/p>\n\n<pre><code>IMAGE_SIZE=299\ntflite_convert \\\n  --graph_def_file=tf_files\/retrained_graph.pb \\\n  --output_file=tf_files\/optimized_graph.lite \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --output_format=TFLITE \\\n  --input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \\\n  --input_array=Mul \\\n  --output_array=final_result \\\n  --inference_type=FLOAT \\\n  --input_data_type=FLOAT\n<\/code><\/pre>\n"},{"AnswerId":"53527198","CreationDate":"2018-11-28T19:58:35.383","ParentId":null,"OwnerUserId":"724072","Title":null,"Body":"<p>Most likely it's because during the retraining process the input and output tensors were renamed. If this is a retrained inceptionv3 graph, try using <em>Mul<\/em> as the input tensor name and <em>final_result<\/em> as the output tensor name.<\/p>\n\n<pre><code>bazel run --config=opt \/\/tensorflow\/contrib\/lite\/toco:toco -- \\\n    ... other options ...\n    --input_shape=1,299,299,3 \\\n    --input_array=Mul \\\n    --output_array=final_result\n<\/code><\/pre>\n\n<p>Similar adjustment if you use tflife_convert as Aleksandr suggest.<\/p>\n"},{"AnswerId":"55949942","CreationDate":"2019-05-02T09:58:22.590","ParentId":null,"OwnerUserId":"8069206","Title":null,"Body":"<p>without bazel you can try the following code<\/p>\n\n<pre><code>pip uninstall tensorflow\npip install tf-nightly\npip show protobuf\n<\/code><\/pre>\n\n<p>If protobuf is version 3.6.1, then proceed to installing the pre-release version of 3.7.0.<\/p>\n\n<pre><code>pip uninstall protobuf\npip install protobuf==3.7.0rc2 \n<\/code><\/pre>\n\n<p>I still couldn\u2019t get the command line version to work. It kept returning the error: \u201ctflite_convert: error: \u2013input_arrays and \u2013output_arrays are required with \u2013graph_def_file\u201d although both parameters were supplied. It worked in Python, however.<\/p>\n\n<pre><code>import tensorflow as tf\n\ngraph_def_file = \"model.pb\"\ninput_arrays = [\"model_inputs\"]\noutput_arrays = [\"model_outputs\"]\n\nconverter = tf.lite.TFLiteConverter.from_frozen_graph(\n        graph_def_file, input_arrays, output_arrays)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n"},{"AnswerId":"56521325","CreationDate":"2019-06-10T06:14:26.840","ParentId":null,"OwnerUserId":"11167926","Title":null,"Body":"<pre><code>import tensorflow as tf\ngf = tf.GraphDef()\nm_file = open('frozen_inference_graph.pb','rb')\nfor n in gf.node:\n    print( n.name )\n<\/code><\/pre>\n\n<p>first one is input_arrays\nlast names are output_arrays (could be more than one depends on your number of output of the model)<\/p>\n\n<p>my output<\/p>\n\n<ul>\n<li>image_tensor     &lt;--- input_array <\/li>\n<li>Cast <\/li>\n<li>Preprocessor\/map\/Shape Preprocessor\/map\/strided_slice\/stack<\/li>\n<li>Preprocessor\/map\/strided_slice\/stack_1<\/li>\n<li>. <\/li>\n<li>. <\/li>\n<li>. <\/li>\n<li>Postprocessor\/BatchMultiClassNonMaxSuppression\/map\/ <\/li>\n<li>TensorArrayStack_5\/TensorArrayGatherV3<\/li>\n<li>Postprocessor\/Cast_3 <\/li>\n<li>Postprocessor\/Squeeze <\/li>\n<li>add\/y <\/li>\n<li>add<\/li>\n<li>detection_boxes          &lt;---output_array<\/li>\n<li>detection_scores             &lt;---output_array<\/li>\n<li>detection_multiclass_scores <\/li>\n<li>detection_classes           &lt;---output_array<\/li>\n<li>num_detections               &lt;---output_array<\/li>\n<li>raw_detection_boxes<\/li>\n<li>raw_detection_scores<\/li>\n<\/ul>\n"},{"AnswerId":"53678271","CreationDate":"2018-12-08T00:06:12.640","ParentId":null,"OwnerUserId":"10361081","Title":null,"Body":"<p>The error hints that you have not entered the correct<\/p>\n\n<pre><code>--input_arrays\n<\/code><\/pre>\n\n<p>From <a href=\"https:\/\/www.tensorflow.org\/lite\/devguide\" rel=\"nofollow noreferrer\">TF Lite Developer Guide<\/a> \nI quote :<\/p>\n\n<p><em>\"Setting the <strong>input_array<\/strong> and <strong>output_array<\/strong> arguments is not straightforward. The easiest way to find these values is to explore the graph using <strong>TensorBoard<\/strong>.\"<\/em><\/p>\n\n<p>Using the Tensorboard isn't hard either, by simply running this command<\/p>\n\n<pre><code>tensorboard --logdir=path\/to\/log-directory\n<\/code><\/pre>\n\n<p>View the TensorBoard at <\/p>\n\n<pre><code>localhost:6006   \n<\/code><\/pre>\n"},{"AnswerId":"57179125","CreationDate":"2019-07-24T09:03:52.510","ParentId":null,"OwnerUserId":"11829279","Title":null,"Body":"<p>import tensorflow as tf<\/p>\n\n<pre><code>!tflite_convert \\\n--output_file \"random.tflite\" \\\n--graph_def_file \"pb file path\" \\\n--input_arrays \"input tensor name\" \\\n--output_arrays \"output tensor name\"    \n<\/code><\/pre>\n"},{"AnswerId":"58583419","CreationDate":"2019-10-27T21:13:03.337","ParentId":null,"OwnerUserId":"11517841","Title":null,"Body":"<p>To run the tflite converter on your local machine, you will need bazel and toco.<\/p>\n\n<p>And if you read some issues in GitHub, in some versions of Tensrflow tflite causes a lot of trouble. To overcome this trouble, some recommend using tf-nightly!<\/p>\n\n<p>To avoid all this, simply use Google Colab to convert your .pb into .lite or .tflite.<\/p>\n\n<p>Since Colab started having the \"upload\" option for uploading your files into the current kernel, this I think is the most simple way without having to worry about other packages and their dependencies.<\/p>\n\n<p>Here is the Colab notebook:<\/p>\n\n<p><a href=\"https:\/\/drive.google.com\/file\/d\/1lDcttsmZC0Y6dXxwe0EVZUsyVoR8HuR-\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/1lDcttsmZC0Y6dXxwe0EVZUsyVoR8HuR-\/view?usp=sharing<\/a><\/p>\n\n<p>There are two ways in which you can upload your .pb file to the current session:<\/p>\n\n<p>i) (The easy way) After running the first cell in the above notebook, the drive will be mounted. So on your left part of the screen go to the files column and right click on the folder you want to upload your .pb file and choose upload.\nThen use \"ls\" and \"cd\" commands to work your way into the folder and run the tflite converter cell.<\/p>\n\n<p>ii) Run the cell with files.upload() command and click on browse and choose the .pb file from your local machine.<\/p>\n\n<p>Once the file is uploaded, give its path to the variable \"localpb\" and also the name of the .lite model. Then simply run the cell having the \"TFLiteConverter\" comamnd.<\/p>\n\n<p>And voila. You should have a tflite model appear in your drive. Simply right-click on it and download to your local machine to run inferences.<\/p>\n"},{"AnswerId":"53081748","CreationDate":"2018-10-31T10:57:42.180","ParentId":null,"OwnerUserId":"8049751","Title":null,"Body":"<p>You can use utility <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/cmdline_examples\" rel=\"noreferrer\">tflite_convert<\/a> which is the part of tensorflow 1.10 (or higher) package.<\/p>\n\n<p>The simple use for float inference is something like:<\/p>\n\n<pre><code>tflite_convert \\\n    --output_file=\/tmp\/retrained_graph.tflite \\\n    --graph_def_file=\/tmp\/retrained_graph.pb \\\n    --input_arrays=input \\\n    --output_arrays=output\n<\/code><\/pre>\n\n<p>Where input and output - are input and ouput tensors of your tensorflow graph<\/p>\n"},{"AnswerId":"55794573","CreationDate":"2019-04-22T12:39:32.837","ParentId":null,"OwnerUserId":"4663377","Title":null,"Body":"<p>I am making a wild guess here, maybe you entered <code>input_arrays=input<\/code>.\nWhich may not be true. Use this script to find the name of the input and output arrays of the frozen inference graph<\/p>\n\n<pre><code>import tensorflow as tf\ngf = tf.GraphDef()   \nm_file = open('frozen_inference_graph.pb','rb')\ngf.ParseFromString(m_file.read())\n\nwith open('somefile.txt', 'a') as the_file:\n    for n in gf.node:\n        the_file.write(n.name+'\\n')\n\nfile = open('somefile.txt','r')\ndata = file.readlines()\nprint \"output name = \"\nprint data[len(data)-1]\n\nprint \"Input name = \"\nfile.seek ( 0 )\nprint file.readline()\n<\/code><\/pre>\n\n<p>In my case they are:<\/p>\n\n<pre><code>output name: SemanticPredictions\ninput name: ImageTensor\n<\/code><\/pre>\n"},{"AnswerId":"58024283","CreationDate":"2019-09-20T08:03:01.840","ParentId":null,"OwnerUserId":"4663377","Title":null,"Body":"<p>I am following up from my previous answer you can use the following script to convert your trained model on ssd mobilenet to tflte using <\/p>\n\n<pre><code>python object_detection\/export_tflite_ssd_graph \\\n    --pipeline_config_path ssd_0.75_export\/pipeline.config \\\n    --trained_checkpoint_prefix ssd_0.75_export\/model.ckpt \\\n    --output_directory ssd_to_tflite_output\n<\/code><\/pre>\n\n<p>To do this you will first need to be present in research folder of tensorflow object detection API, and change the dile path\/name as per your names.\nIf this dosent work try running this script from research folder and rerun:<\/p>\n\n<pre><code>protoc object_detection\/protos\/*.proto --python_out=.\nexport PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`\/slim\n<\/code><\/pre>\n"}]}
+{"QuestionId":52918090,"AnswerCount":2,"Tags":"<python><tensorflow><mnist>","CreationDate":"2018-10-21T17:38:22.237","AcceptedAnswerId":52957689.0,"OwnerUserId":9548638.0,"Title":"How to get two tf.dataset from tf.data.Dataset.zip((images, labels))","Body":"<p>I am working on the Python\/tensorflow\/mnist tutorial.<\/p>\n\n<p>Since a few weeks using the orignal code from tensorflow web site i get the warning that the image dataset would soon be deprecated abd that i should use the following one :\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/official\/mnist\/dataset.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/official\/mnist\/dataset.py<\/a><\/p>\n\n<p>I load it it my code using :<\/p>\n\n<pre><code>from tensorflow.models.official.mnist import dataset\ntrainfile = dataset.train(data_dir)\n<\/code><\/pre>\n\n<p>Which returns :<\/p>\n\n<pre><code>tf.data.Dataset.zip((images, labels))\n<\/code><\/pre>\n\n<p>The issue is that I cannot find a,way to separate them in the following way for example :<\/p>\n\n<pre><code>  trainfile = dataset.train(data_dir)\n  train_data= trainfile.images\n  train_label= trainfile.label\n<\/code><\/pre>\n\n<p>But this clearly doesnot work because the attributrs images and label do not exist. trainfile is a tf.dataset.<\/p>\n\n<p>Knowing that tf.dataset is made of int32 and float32 i tried :<\/p>\n\n<pre><code>  train_data = trainfile.map(lambda x,y : x.dtype == tf.float32)\n<\/code><\/pre>\n\n<p>But it returns and empty dataset.<\/p>\n\n<p>I insist (but will be open mimded) in doing it this way (two complete batches of image and label) because this is how the tutorial works :<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn<\/a><\/p>\n\n<p>I saw a lot of solution to get elements from datasets but nothing to go back from the zip operations that is done in the following code <\/p>\n\n<pre><code>tf.data.Dataset.zip((images, labels))\n<\/code><\/pre>\n\n<p>Thanks you in advance for your help.<\/p>\n","answers":[{"AnswerId":"52957689","CreationDate":"2018-10-23T20:53:08.410","ParentId":null,"OwnerUserId":"3670165","Title":null,"Body":"<p>I hope this helps:<\/p>\n\n<pre><code>inputs = tf.placeholder(tf.float32, shape=(None, 784), name='inputs')\noutputs = tf.placeholder(tf.float32, shape=(None,), name='outputs')\n\n#Prepare a tensorflow dataset\nds = tf.data.Dataset.from_tensor_slices((x_train, y_train))\n\nds = ds.shuffle(buffer_size=10, reshuffle_each_iteration=True).batch(batch_size=batch_size, drop_remainder=True).repeat()\niter = ds.make_one_shot_iterator()\nnext = iter.get_next()\n\ninputs = next[0]\noutputs = next[1]\n<\/code><\/pre>\n"},{"AnswerId":"52957329","CreationDate":"2018-10-23T20:24:06.950","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Instead of separating into two datasets, one for images and another for labels, it's best to make a single iterator which returns both the image and the label.<\/p>\n\n<p>The reason why this is preferred is that it's a lot easier to ensure that you match each example with its label even after a complicated series of shuffles, reorderings, filterings, etc, as you might have in a nontrivial input pipeline.<\/p>\n"}]}
+{"QuestionId":52918107,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-21T17:40:28.280","AcceptedAnswerId":52918108.0,"OwnerUserId":913098.0,"Title":"tensorflow tf.maximum(0, x) returns error","Body":"<p>When trying to use <code>tf.maximum<\/code> as one would expect:<\/p>\n\n<pre><code>loss = tf.maximum(0, basic_loss)\n<\/code><\/pre>\n\n<p>this error is obtained<\/p>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- ValueError                                Traceback (most recent call\n  last)\n  \/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\n  in apply_op(self, op_type_name, name, **keywords)\n      489                 as_ref=input_arg.is_ref,\n  --> 490                 preferred_dtype=default_dtype)\n      491           except TypeError as err:<\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\n  in internal_convert_to_tensor(value, dtype, name, as_ref,\n  preferred_dtype)\n      740         if ret is None:\n  --> 741           ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n      742 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\n  in _TensorTensorConversionFunction(t, dtype, name, as_ref)\n      613         \"Tensor conversion requested dtype %s for Tensor with dtype %s: %r\"\n  --> 614         % (dtype.name, t.dtype.name, str(t)))\n      615   return t<\/p>\n  \n  <p>ValueError: Tensor conversion requested dtype int32 for Tensor with\n  dtype float32: 'Tensor(\"add_13:0\", shape=(), dtype=float32)'<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>TypeError                                 Traceback (most recent call\n  last)  in ()\n        5               tf.random_normal([3, 128], mean=1, stddev=1, seed = 1),\n        6               tf.random_normal([3, 128], mean=3, stddev=4, seed = 1))\n  ----> 7     loss = triplet_loss(y_true, y_pred)\n        8 \n        9     print(\"loss = \" + str(loss.eval()))<\/p>\n  \n  <p> in triplet_loss(y_true, y_pred, alpha)\n       26     basic_loss = pos_dist - neg_dist + alpha\n       27     # Step 4: Take the maximum of basic_loss and 0.0. Sum over the training examples.\n  ---> 28     loss = tf.maximum(0, basic_loss)\n       29     ### END CODE HERE ###\n       30 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\n  in maximum(x, y, name)    1261     A <code>Tensor<\/code>. Has the same type as\n  <code>x<\/code>.    1262   \"\"\"\n  -> 1263   result = _op_def_lib.apply_op(\"Maximum\", x=x, y=y, name=name)    1264   return result    1265 <\/p>\n  \n  <p>\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\n  in apply_op(self, op_type_name, name, **keywords)\n      524                   \"%s type %s of argument '%s'.\" %\n      525                   (prefix, dtypes.as_dtype(attrs[input_arg.type_attr]).name,\n  --> 526                    inferred_from[input_arg.type_attr]))\n      527 \n      528           types = [values.dtype]<\/p>\n  \n  <p>TypeError: Input 'y' of 'Maximum' Op has type float32 that does not\n  match type int32 of argument 'x'.<\/p>\n<\/blockquote>\n\n<p>What seems to be the problem?<\/p>\n","answers":[{"AnswerId":"52918108","CreationDate":"2018-10-21T17:40:28.280","ParentId":null,"OwnerUserId":"913098","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/maximum\" rel=\"nofollow noreferrer\">The tensor flow doc<\/a> does not state that the <code>maximum<\/code> function is <strong>non-commutative<\/strong>. <\/p>\n\n<p>It works only if the type of the 1st argument is a Tensor, but not if its type is int.<\/p>\n\n<p>Need to call this function with replaced positions of the arguments for constants:<\/p>\n\n<p><code>tf.maximum(basic_loss, 0)<\/code><\/p>\n\n<p>instead of<\/p>\n\n<p><code>tf.maximum(0, basic_loss)<\/code><\/p>\n"}]}
+{"QuestionId":52918197,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-21T17:50:03.270","AcceptedAnswerId":null,"OwnerUserId":2867237.0,"Title":"TimeDistributed in Keras\/Tensorflow","Body":"<p>I am trying to implement a simple many to many LSTM for Sequence Prediction. The problem is very easy. The input is a sequence of 0s and 1s. The output at each time step is the count of ones in the sequence until that time step. For example assume the input is [0 1 0 1]. The output of the given input would be time0=0, time1=1, time2=1, time3=2. I should note that I use One hot encoding to represent the output. <\/p>\n\n<p>Assumptions: the length of the input sequence is 20 (so at most I can have 20 ones in the sequence). Therefore, I consider 21 classes for output (one hot encoding). Class 0 means there is no one in the sequence. Class 21 shows that we have 20 ones in the sequence.<\/p>\n\n<p>So far, I use the following model:<\/p>\n\n<pre><code># create LSTM\n\nmodel = tf.keras.models.Sequential()\n\nmodel.add(tf.keras.layers.LSTM(30, input_shape=(20, 1), return_sequences=True ))\n#model.add(tf.keras.layers.LSTM(30, input_shape=(20, 1)))\nprint (model.input_shape)\nprint (model.output_shape)\nmodel.add(tf.keras.layers.Dropout(0.2))\n\n#model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(21, activation='softmax')))\nmodel.add(tf.keras.layers.Dense(21, activation='softmax'))\n\n\nprint(model.summary())\n\n\nmodel.compile(loss='categorical_crossentropy',\n             optimizer='adam',\n             metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I evaluated it by adding and removing \"tf.keras.layers.TimeDistributed\". Both of them reach the same accuracy of 99%! I am wondering why is that? So when we need to use \"TimeDistributed\"? What is it for then?<\/p>\n","answers":[{"AnswerId":"52919096","CreationDate":"2018-10-21T19:32:35.470","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>For <code>Dense<\/code> layer you don't have to use <code>TimeDistributed<\/code> because the kernel gets <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/broadcasting\" rel=\"nofollow noreferrer\">broadcasted<\/a>. For example you have <code>(30, 21)<\/code> as your W and <code>(batch, 20, 30)<\/code> as your x, so when you multiply the kernal gets broadcasted multiplied with every minibatch entry and you end up with (batch, 20, 30) times (30, 21) gives you <code>(batch, 20, 21)<\/code>. The equation is <code>Wx<\/code> here.<\/p>\n\n<p>You use <code>TimeDistributed<\/code> when you have more complicated layer or even a model. Imagine a CNN model which you want to apply to every frame of the video. Then you could <code>TimeDistributed<\/code> to it's full potential.<\/p>\n"}]}
+{"QuestionId":52918337,"AnswerCount":1,"Tags":"<r><tensorflow><keras>","CreationDate":"2018-10-21T18:05:34.653","AcceptedAnswerId":52919027.0,"OwnerUserId":5541868.0,"Title":"R Keras: Convert tensorflow tensor to R array","Body":"<p>I am using R Keras. I can obtain the output of an intermediate layer by for example executing:<\/p>\n\n<pre><code>layer_output &lt;- get_layer(mymodel, index=1)$output\n<\/code><\/pre>\n\n<p>where mymodel is a Keras model.\nThe problem is that layer_output is a tensor.<\/p>\n\n<pre><code>class(layer_output)\n\n    [1] \"tensorflow.tensor\" \"tensorflow.python.framework.ops.Tensor\"     \n    [3] \"tensorflow.python.framework.ops._TensorLike\" \"python.builtin.object\"  \n<\/code><\/pre>\n\n<p>I want to convert layer_output to an R object such as an array or matrix, but I can't find a way to do it. Many thanks if you can advise me.<\/p>\n","answers":[{"AnswerId":"52919027","CreationDate":"2018-10-21T19:25:32.410","ParentId":null,"OwnerUserId":"100129","Title":null,"Body":"<p>One option is to create a new model which will only output your layer of interest.<\/p>\n\n<p>First create the original model:<\/p>\n\n<pre><code>model &lt;- ...  # create original model\n<\/code><\/pre>\n\n<p>Then create the new model and use predict to get the output:<\/p>\n\n<pre><code>layer_name &lt;- 'my_layer'\nintermediate_layer_model &lt;- keras_model(inputs = model$input,\n                                        outputs = get_layer(model, layer_name)$output)\nintermediate_output &lt;- predict(intermediate_layer_model, data)\n<\/code><\/pre>\n\n<p>Depending on your model the R keras predict function should return a vector, matrix or array - instead of the layer instance returned by get_layer.<\/p>\n"}]}
+{"QuestionId":52918613,"AnswerCount":2,"Tags":"<machine-learning><keras><deep-learning><conv-neural-network><activation-function>","CreationDate":"2018-10-21T18:38:32.577","AcceptedAnswerId":52920547.0,"OwnerUserId":9652400.0,"Title":"Can I use a Sigmoid activation for my output layer, even if my CNN model is doing a regression?","Body":"<p>Final objective: Object Midpoint calculation. <\/p>\n\n<p>I have a small dataset (around 120 images), which has an object (the same in all cases), and the labels are the normalized x,y coordinates of the midpoint of the object in the image (always between 0 and 1)<\/p>\n\n<p>e.g. x = image_005 ; y = (0.1, 0.15) for an image with the object placed near the bottom left corner<\/p>\n\n<p>I am trying to use a ResNet architecture but customized for my image-size (all are identical images). Since the output values are always between 0 and 1, for both coordinates, I was wondering if it is possible to use Sigmoid activation in my last layer: <\/p>\n\n<pre><code> X = Dense(2, activation='sigmoid', name='fc', kernel_initializer = glorot_uniform(seed=0))(X)\n<\/code><\/pre>\n\n<p>instead of a linear activation (as is advised often when you are trying to achieve a regression result)<\/p>\n\n<p>For the loss function, I use MSE, with 'rmsprop' optimizer and in addition to accuracy and MSE, I have written a custom metric for telling me if the predicted points are off from the labels by more than 5%<\/p>\n\n<pre><code>model.compile(optimizer='rmsprop', loss='mean_squared_error', metrics=['mse','acc',perc_midpoint_err])\n<\/code><\/pre>\n\n<p>I am not getting good results, after training the model on around 150 epochs (I experimented with different batch sizes too)<\/p>\n\n<p>Should I change the activation layer to linear? Or is there a different modification I can do to my model? Or is ResNet completely unsuitable for this task?<\/p>\n","answers":[{"AnswerId":"52920547","CreationDate":"2018-10-21T22:41:53.467","ParentId":null,"OwnerUserId":"290864","Title":null,"Body":"<p>Your task is related to object detection. The difference is, that you seem to have only one object in each of your images, whereas in detection there may be multiple objects or no object present. For object detection, there are networks such as YOLOv3 (<a href=\"https:\/\/pjreddie.com\/media\/files\/papers\/YOLOv3.pdf\" rel=\"nofollow noreferrer\">https:\/\/pjreddie.com\/media\/files\/papers\/YOLOv3.pdf<\/a>) or Single Shot Multibox Detector - SSD (<a href=\"https:\/\/arxiv.org\/pdf\/1512.02325.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1512.02325.pdf<\/a>) but also ResNet can be trained as an object detection network (as in this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1506.01497.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1506.01497.pdf<\/a>)<\/p>\n\n<p>I will shortly describe how YOLO solves the regression problem for bounding box x,y coordinates: <\/p>\n\n<ul>\n<li>YOLO uses a sigmoid activation function for x,y<\/li>\n<li>It devides the image into grid cells and predicts offsets for a potential object in each grid-cell. This may be helpful in case you have large images or objects at multiple locations.<\/li>\n<li>The original paper uses MSE as a loss function, but in my favorite keras-reimplementation they use crossentropy loss with the Adam optimizer.<\/li>\n<\/ul>\n\n<p>In principle your setup looks fine to me. But there are many things which could result in poor performance, since you don't tell about the domain of your dataset: Are you using a pretrained network or are you training from scratch? Is it a new category which you are to learning or an object category the network has seen before? etc. <\/p>\n\n<p>Here are some ideas which you could try:<\/p>\n\n<ul>\n<li>change the optimizer (to SGD or Adam)<\/li>\n<li>change the learning rate (better smaller than too large) <\/li>\n<li>increase your dataset size. For retraining the network for a new object category my rule of thumb is to use about 500-1000 images. For retraining from scratch you need orders of magnitude more.<\/li>\n<li>you may want to check out YOLO or SSD and modify those networks for your case<\/li>\n<\/ul>\n\n<p>I hope you find some inspiration for your solution.<\/p>\n"},{"AnswerId":"57897211","CreationDate":"2019-09-11T21:47:06.907","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>Apart from what you have done, there are lots of other things that you can do:<\/p>\n\n<ol>\n<li>Use <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\"><strong>ImageAugmentation<\/strong><\/a> technique to generate more data. Also, normalize the images.<\/li>\n<li>Make a deeper model with a few more convolution layers.<\/li>\n<li>Use a proper <strong>weights initializer<\/strong> maybe He-normal for the convolution layers.<\/li>\n<li>Use <a href=\"https:\/\/keras.io\/layers\/normalization\/\" rel=\"nofollow noreferrer\"><strong>BatchNormalization<\/strong><\/a> between layers to make the <strong>mean<\/strong> and <strong>std<\/strong> of your filter values equal to 0 and 1 respectively. <\/li>\n<li>Trying using <strong>crossentropy loss<\/strong> as it helps in better calculation of your gradients. In MSE the gradients become very small over time although it preferred for regression kind of problem. You can also try <a href=\"https:\/\/keras.io\/losses\/\" rel=\"nofollow noreferrer\"><strong>MeanSquaredLogarithmicError<\/strong><\/a>.<\/li>\n<li>Try to change the optimizer to <strong>Adam<\/strong> or <strong>Stochastic Gradient Descent with nestrov momentum<\/strong> (performs better than Adam on validation set).<\/li>\n<li>In case, you have a few more classes in your dataset, and you have class imbalance problem, you can use <strong>Focal loss<\/strong>, a variant of crossentropy loss which penalizes the misclassified labels more than the correctly classified labels. Also, reducing the <strong>batch size<\/strong> and <strong>Upsampling<\/strong> should help.<\/li>\n<li>Use <strong>Bayesian Optimization<\/strong> techniques for <strong>hyperparameter<\/strong> tuning of your model.<\/li>\n<\/ol>\n\n<p>Here in a simple implementation of <strong>Resnet<\/strong>:<\/p>\n\n<pre><code>def unit(x, filters, pool=False):\n    res = x\n    if pool:\n        x = MaxPooling2D(pool_size=(2, 2))(x)\n        res = Conv2D(filters=filters, kernel_size=(1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(res)\n    out = BatchNormalization()(x)\n    out = Activation('relu')(out)\n    out = Conv2D(filters=filters, kernel_size=(3, 3), strides=(1, 1), padding='same', kernel_initializer='he_normal')(out)\n\n    out = BatchNormalization()(out)\n    out = Activation('relu')(out)\n    out = Conv2D(filters=filters, kernel_size=(3, 3), strides=(1, 1), padding='same', kernel_initializer='he_normal')(out)\n\n    x = keras.layers.add([res, out])\n    return x\n\ndef model(inputs):\n    inp = Input(inputs)\n    x = Conv2D(32, (3, 3), padding='same', kernel_initializer='he_uniform')(inp)\n    x = unit(x, 32)\n    x = unit(x, 32)\n    x = unit(x, 32)\n\n    x = unit(x, 64, pool=True)\n    x = unit(x, 64)\n    x = unit(x, 64)\n\n    x = unit(x, 128, pool=True)\n    x = unit(x, 128)\n    x = unit(x, 128)\n\n    x = unit(x, 256, pool=True)\n    x = unit(x, 256)\n    x = unit(x, 256)\n\n    x = BatchNormalization()(x)\n    x = Activation('relu')(x)\n    x = Dropout(0.25)(x)\n\n    x = AveragePooling2D((3, 3))(x)\n    x = Flatten()(x)\n    x = Dense(2, activation='sigmoid')(x)\n\n    model = Model(inputs=inp, outputs=x)\n    optimizer = Adam(lr=0.001)\n    # model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])\n    model.compile(optimizer=optimizer, loss=keras.losses.MeanSquaredLogarithmicError(), metrics=['accuracy'])\n    return model    \n<\/code><\/pre>\n"}]}
+{"QuestionId":52918681,"AnswerCount":2,"Tags":"<python><tensorflow><raspberry-pi>","CreationDate":"2018-10-21T18:45:50.610","AcceptedAnswerId":null,"OwnerUserId":8614420.0,"Title":"Traceback (most recent call last) error on Tensorflow Rasperry PI zero W","Body":"<p>I have Rasppery PI zero Wireless and I am trying to use Tensorflow on it.<\/p>\n\n<p>I did all steps on link below.\n<a href=\"https:\/\/www.youtube.com\/watch?v=npZ-8Nj1YwY\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=npZ-8Nj1YwY<\/a><\/p>\n\n<p>github link:\n<a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-on-the-Raspberry-Pi\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-on-the-Raspberry-Pi<\/a><\/p>\n\n<p>Object_detection_picamera.py:<\/p>\n\n<pre><code>######## Picamera Object Detection Using Tensorflow Classifier #########\n#\n# Author: Evan Juras\n# Date: 4\/15\/18\n# Description: \n# This program uses a TensorFlow classifier to perform object detection.\n# It loads the classifier uses it to perform object detection on a Picamera feed.\n# It draws boxes and scores around the objects of interest in each frame from\n# the Picamera. It also can be used with a webcam by adding \"--usbcam\"\n# when executing this script from the terminal.\n\n## Some of the code is copied from Google's example at\n## https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\n\n## and some is copied from Dat Tran's example at\n## https:\/\/github.com\/datitran\/object_detector_app\/blob\/master\/object_detection_app.py\n\n## but I changed it to make it more understandable to me.\n\n\n# Import packages\nimport os\nimport cv2\nimport numpy as np\nfrom picamera.array import PiRGBArray\nfrom picamera import PiCamera\nimport tensorflow as tf\nimport argparse\nimport sys\n\n# Set up camera constants\nIM_WIDTH = 1280\nIM_HEIGHT = 720\n#IM_WIDTH = 640    Use smaller resolution for\n#IM_HEIGHT = 480   slightly faster framerate\n\n# Select camera type (if user enters --usbcam when calling this script,\n# a USB webcam will be used)\ncamera_type = 'picamera'\nparser = argparse.ArgumentParser()\nparser.add_argument('--usbcam', help='Use a USB webcam instead of picamera',\n                    action='store_true')\nargs = parser.parse_args()\nif args.usbcam:\n    camera_type = 'usb'\n\n# This is needed since the working directory is the object_detection folder.\nsys.path.append('..')\n\n# Import utilites\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util\n\n# Name of the directory containing the object detection module we're using\nMODEL_NAME = 'ssdlite_mobilenet_v2_coco_2018_05_09'\n\n# Grab path to current working directory\nCWD_PATH = os.getcwd()\n\n# Path to frozen detection graph .pb file, which contains the model that is used\n# for object detection.\nPATH_TO_CKPT = os.path.join(CWD_PATH,MODEL_NAME,'frozen_inference_graph.pb')\n\n# Path to label map file\nPATH_TO_LABELS = os.path.join(CWD_PATH,'data','mscoco_label_map.pbtxt')\n\n# Number of classes the object detector can identify\nNUM_CLASSES = 90\n\n## Load the label map.\n# Label maps map indices to category names, so that when the convolution\n# network predicts `5`, we know that this corresponds to `airplane`.\n# Here we use internal utility functions, but anything that returns a\n# dictionary mapping integers to appropriate string labels would be fine\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\ncategories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)\ncategory_index = label_map_util.create_category_index(categories)\n\n# Load the Tensorflow model into memory.\ndetection_graph = tf.Graph()\nwith detection_graph.as_default():\n    od_graph_def = tf.GraphDef()\n    with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:\n        serialized_graph = fid.read()\n        od_graph_def.ParseFromString(serialized_graph)\n        tf.import_graph_def(od_graph_def, name='')\n\n    sess = tf.Session(graph=detection_graph)\n\n\n# Define input and output tensors (i.e. data) for the object detection classifier\n\n# Input tensor is the image\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n\n# Output tensors are the detection boxes, scores, and classes\n# Each box represents a part of the image where a particular object was detected\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n\n# Each score represents level of confidence for each of the objects.\n# The score is shown on the result image, together with the class label.\ndetection_scores = detection_graph.get_tensor_by_name('detection_scores:0')\ndetection_classes = detection_graph.get_tensor_by_name('detection_classes:0')\n\n# Number of objects detected\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\n# Initialize frame rate calculation\nframe_rate_calc = 1\nfreq = cv2.getTickFrequency()\nfont = cv2.FONT_HERSHEY_SIMPLEX\n\n# Initialize camera and perform object detection.\n# The camera has to be set up and used differently depending on if it's a\n# Picamera or USB webcam.\n\n# I know this is ugly, but I basically copy+pasted the code for the object\n# detection loop twice, and made one work for Picamera and the other work\n# for USB.\n\n### Picamera ###\nif camera_type == 'picamera':\n    # Initialize Picamera and grab reference to the raw capture\n    camera = PiCamera()\n    camera.resolution = (IM_WIDTH,IM_HEIGHT)\n    camera.framerate = 10\n    rawCapture = PiRGBArray(camera, size=(IM_WIDTH,IM_HEIGHT))\n    rawCapture.truncate(0)\n\n    for frame1 in camera.capture_continuous(rawCapture, format=\"bgr\",use_video_port=True):\n\n        t1 = cv2.getTickCount()\n\n        # Acquire frame and expand frame dimensions to have shape: [1, None, None, 3]\n        # i.e. a single-column array, where each item in the column has the pixel RGB value\n        frame = frame1.array\n        frame.setflags(write=1)\n        frame_expanded = np.expand_dims(frame, axis=0)\n\n        # Perform the actual detection by running the model with the image as input\n        (boxes, scores, classes, num) = sess.run(\n            [detection_boxes, detection_scores, detection_classes, num_detections],\n            feed_dict={image_tensor: frame_expanded})\n\n        # Draw the results of the detection (aka 'visulaize the results')\n        vis_util.visualize_boxes_and_labels_on_image_array(\n            frame,\n            np.squeeze(boxes),\n            np.squeeze(classes).astype(np.int32),\n            np.squeeze(scores),\n            category_index,\n            use_normalized_coordinates=True,\n            line_thickness=8,\n            min_score_thresh=0.40)\n\n        cv2.putText(frame,\"FPS: {0:.2f}\".format(frame_rate_calc),(30,50),font,1,(255,255,0),2,cv2.LINE_AA)\n\n        # All the results have been drawn on the frame, so it's time to display it.\n        cv2.imshow('Object detector', frame)\n\n        t2 = cv2.getTickCount()\n        time1 = (t2-t1)\/freq\n        frame_rate_calc = 1\/time1\n\n        # Press 'q' to quit\n        if cv2.waitKey(1) == ord('q'):\n            break\n\n        rawCapture.truncate(0)\n\n    camera.close()\n\n### USB webcam ###\nelif camera_type == 'usb':\n    # Initialize USB webcam feed\n    camera = cv2.VideoCapture(0)\n    ret = camera.set(3,IM_WIDTH)\n    ret = camera.set(4,IM_HEIGHT)\n\n    while(True):\n\n        t1 = cv2.getTickCount()\n\n        # Acquire frame and expand frame dimensions to have shape: [1, None, None, 3]\n        # i.e. a single-column array, where each item in the column has the pixel RGB value\n        ret, frame = camera.read()\n        frame_expanded = np.expand_dims(frame, axis=0)\n\n        # Perform the actual detection by running the model with the image as input\n        (boxes, scores, classes, num) = sess.run(\n            [detection_boxes, detection_scores, detection_classes, num_detections],\n            feed_dict={image_tensor: frame_expanded})\n\n        # Draw the results of the detection (aka 'visulaize the results')\n        vis_util.visualize_boxes_and_labels_on_image_array(\n            frame,\n            np.squeeze(boxes),\n            np.squeeze(classes).astype(np.int32),\n            np.squeeze(scores),\n            category_index,\n            use_normalized_coordinates=True,\n            line_thickness=8,\n            min_score_thresh=0.85)\n\n        cv2.putText(frame,\"FPS: {0:.2f}\".format(frame_rate_calc),(30,50),font,1,(255,255,0),2,cv2.LINE_AA)\n\n        # All the results have been drawn on the frame, so it's time to display it.\n        cv2.imshow('Object detector', frame)\n\n        t2 = cv2.getTickCount()\n        time1 = (t2-t1)\/freq\n        frame_rate_calc = 1\/time1\n\n        # Press 'q' to quit\n        if cv2.waitKey(1) == ord('q'):\n            break\n\n    camera.release()\n\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p>When I came the time 15:32 in the video I had an error as you can see below.<\/p>\n\n<pre><code>pi@raspberrypi:~\/tensorflow1\/models\/research\/object_detection $ python3 Object_detection_picamera.py\nTraceback (most recent call last):\n  File \"Object_detection_picamera.py\", line 84, in &lt;module&gt;\n    serialized_graph = fid.read()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 120, in read\n    self._preread_check()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 80, in _preread_check\n    compat.as_bytes(self.__name), 1024 * 512, status)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/errors_impl.py\", line 519, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.NotFoundError: \/home\/pi\/tensorflow1\/models\/research\/object_detection\/ssdlite_mobilenet_v2_coco_2018_05_09\/frozen_inference_graph.pb; No such file or directory\n<\/code><\/pre>\n\n<p>I search on net and tried to open it via python2.7 and then I got another error which is couldn't find cv2 error.<\/p>\n\n<p>So, could anyone tell me an idea about it. <\/p>\n","answers":[{"AnswerId":"52941369","CreationDate":"2018-10-23T04:49:03.347","ParentId":null,"OwnerUserId":"8614420","Title":null,"Body":"<p>Fixed it. We have to look always at the path again and again. <\/p>\n"},{"AnswerId":"52937248","CreationDate":"2018-10-22T20:23:47.677","ParentId":null,"OwnerUserId":"8185230","Title":null,"Body":"<p>You need to provide a tensorflow frozen model to run. We provide some pretrained models in our <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">model zoo<\/a>.<\/p>\n"}]}
+{"QuestionId":52919239,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-21T19:49:16.717","AcceptedAnswerId":null,"OwnerUserId":7732751.0,"Title":"ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.(importing tensorflow error)","Body":"<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import tensorflow as tf\n      Traceback (most recent call last):\n        File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n          from tensorflow.python.pywrap_tensorflow_internal import *\n        File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n          _pywrap_tensorflow_internal = swig_import_helper()\n        File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n          _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n        File \"E:\\path\\envs\\py36_tf_cpu\\lib\\imp.py\", line 243, in load_module\n          return load_dynamic(name, filename, file)\n        File \"E:\\path\\envs\\py36_tf_cpu\\lib\\imp.py\", line 343, in load_dynamic\n          return _load(spec)\n      ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow__init__.py\", line 22, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"E:\\path\\envs\\py36_tf_cpu\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.<\/p>\n","answers":[{"AnswerId":"55648155","CreationDate":"2019-04-12T09:08:19.760","ParentId":null,"OwnerUserId":"9524424","Title":null,"Body":"<p>As of April 2019, I solved the \"DLL load failed\" problem under Windows 10 \/ Python 3.6.x \/ GPU RTX 20xx by installing <strong>CUDA 10.0<\/strong> (not 10.1 or 9.x!!!) with <strong>cuDNN 7.5.0<\/strong>. I also have Visual Studio 2015 installed. <\/p>\n\n<p>If you have CUDA 10.1, uninstall it, install 10.0, and reboot. Place the files from cuDNN in the respective dirs of your CUDA installation. Don't forget to add the location of your cuDNN <code>*.dll<\/code> files (the <code>\/bin\/<\/code> folder of your CUDA installation) to your PATH.<\/p>\n\n<p>Tensorflow can be installed using <code>pip install tensorflow-gpu<\/code> (currently version 1.13.1).<\/p>\n\n<p>You can test if the GPU is registered, using:<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras\n\nconfig = tf.ConfigProto( device_count = {'GPU': 1 , 'CPU': 56} ) \nsess = tf.Session(config=config) \nkeras.backend.set_session(sess)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52919458,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-21T20:13:42.940","AcceptedAnswerId":null,"OwnerUserId":9938059.0,"Title":"Cannot create file and use filepath in Modelcheckpoint","Body":"<p>I am following this super nice tutorial here <a href=\"https:\/\/dashee87.github.io\/data%20science\/deep%20learning\/python\/another-keras-tutorial-for-neural-network-beginners\/\" rel=\"nofollow noreferrer\">https:\/\/dashee87.github.io\/data%20science\/deep%20learning\/python\/another-keras-tutorial-for-neural-network-beginners\/<\/a>. <\/p>\n\n<p>However, I can't use the filepath in Modelcheckpoint correctly. <\/p>\n\n<pre><code>filepath=\"C:\/Users\/Thomas01\/Documents\/pimaindiansdiabetescsv\/nnweights- \n{epoch:02d}.hdf5\"\ncheckpoint = keras.callbacks.ModelCheckpoint(filepath, monitor='val_acc', \nverbose=0, save_weights_only=False, save_best_only=False, mode='max') \n<\/code><\/pre>\n\n<p>[...]<\/p>\n\n<pre><code>temp_test_model.load_weights(\"C:\/Users\/Thomas01\/Documents\/pimaindiansdiabetescsv\/nnweights-%02d.hdf5\" % i)\n<\/code><\/pre>\n\n<p>[...]<\/p>\n\n<p>I get :<\/p>\n\n<p>OSError: Unable to open file (unable to open file: name = 'C:\/Users\/Thomas01\/Documents\/pimaindiansdiabetescsv\/nnweights-00.hdf5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)<\/p>\n\n<p>Tried to use a raw string (but I think I lose the argument specifier ?) , or \\, but none work.<\/p>\n","answers":[{"AnswerId":"52924573","CreationDate":"2018-10-22T07:49:30.403","ParentId":null,"OwnerUserId":"9590820","Title":null,"Body":"<p>The error message says (<code>No such file or directory<\/code>) that the file path to the model checkpoint cannot be found so the directory will not be opened. You need to create a folder like <code>pimaindiansdiabetescsv<\/code> in this location:<code>C:\/Users\/Thomas01\/Documents\/<\/code>.<\/p>\n"}]}
+{"QuestionId":52919464,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-21T20:14:19.333","AcceptedAnswerId":null,"OwnerUserId":1738210.0,"Title":"ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (454, 512, 512)","Body":"<p>I used this code to produce my dataset in keras. but when I implement my code it produces this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_1_input to\n  have 4 dimensions, but got array with shape (454, 512, 512)<\/p>\n<\/blockquote>\n\n<p>and I can not solve it.  could you please tell me what is the problem? I expand the dimension before using in network but it does not work! could you please answer me fast, due to I search for several days but I could not find the solution and I do not have enough time:<\/p>\n\n<pre><code>import os,cv2\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nfrom sklearn.utils import shuffle\nfrom sklearn.cross_validation import train_test_split\n\nfrom keras import backend as K\n#K.set_image_dim_ordering('th')\n\nfrom keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Convolution2D, MaxPooling2D\nfrom keras.optimizers import SGD,RMSprop,adam\n\n#%%\n\nPATH = os.getcwd()\n# Define data path\ndata_path = r\"E:\\PhD\\thesis\\deepwatermark\\databasetest\\train\"\ndata_dir_list = os.listdir(data_path)\n\nimg_rows=512\nimg_cols=512\nnum_channel=1\nnum_epoch=20\n\n# Define the number of classes\nnum_classes = 7\n\nlabels_name={'CRP':0,'GF':1,'GN':2,'JPG':3,'MED':4,'ROT':5,'SP':6}\n\nimg_data_list=[]\nlabels_list = []\n\nfor dataset in data_dir_list:\n    img_list=os.listdir(data_path+'\/'+ dataset)\n    print ('Loading the images of dataset-'+'{}\\n'.format(dataset))\n    label = labels_name[dataset]\n    for img in img_list:\n        input_img=cv2.imread(data_path + '\/'+ dataset + '\/'+ img )\n        input_img=cv2.cvtColor(input_img, cv2.COLOR_BGR2GRAY)\n        input_img_resize=cv2.resize(input_img,(512,512))\n        img_data_list.append(input_img_resize)\n        labels_list.append(label)\n\nimg_data = np.array(img_data_list)\nimg_data = img_data.astype('float32')\nimg_data \/= 255\nprint (img_data.shape)\n\nlabels = np.array(labels_list)\n# print the count of number of samples for different classes\nprint(np.unique(labels,return_counts=True))\n# convert class labels to on-hot encoding\nY = np_utils.to_categorical(labels, num_classes)\n\n#Shuffle the dataset\nx,y = shuffle(img_data,Y, random_state=2)\n# Split the dataset\nX_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=2)\nimg_data= np.expand_dims(img_data, axis=4)** \nprint (img_data.shape)\n\n\n# Defining the model\ninput_shape=img_data[0].shape\n\nmodel = Sequential()\n\nmodel.add(Convolution2D(32, 3,3,border_mode='same',input_shape=input_shape))\nmodel.add(Activation('relu'))\nmodel.add(Convolution2D(32, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Convolution2D(64, 3, 3))\nmodel.add(Activation('relu'))\n#model.add(Convolution2D(64, 3, 3))\n#model.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes))\nmodel.add(Activation('softmax'))\n\n#sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\n#model.compile(loss='categorical_crossentropy', optimizer=sgd,metrics=[\"accuracy\"])\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop',metrics=[\"accuracy\"])\n\n# Viewing model_configuration\n\nmodel.summary()\nmodel.get_config()\nmodel.layers[0].get_config()\nmodel.layers[0].input_shape         \nmodel.layers[0].output_shape            \nmodel.layers[0].get_weights()\nnp.shape(model.layers[0].get_weights()[0])\nmodel.layers[0].trainable\n\n#%%\n# Training\nhist = model.fit(X_train, y_train, batch_size=16, nb_epoch=num_epoch, verbose=1, validation_data=(X_test, y_test))\n<\/code><\/pre>\n\n<p>my new code with generator is here, did you see any problem? my dataset is the same as before.<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom sklearn.utils import shuffle\nfrom sklearn.cross_validation import train_test_split\nfrom keras import backend as K\n#K.set_image_dim_ordering('th')\n\nfrom keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom keras.optimizers import SGD,RMSprop,adam\n\n\ntrain_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n#\nvalid_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n#\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n#\ntrain_generator = train_datagen.flow_from_directory(\n    directory=r\"E:\\databasetest\\train\",\n    target_size=(512, 512),\n    color_mode=\"grayscale\",\n    batch_size=32,\n    class_mode=\"categorical\",\n    shuffle=True,\n    seed=42\n)\n#\nvalid_generator = valid_datagen.flow_from_directory(\n   directory=r\"E:\\databasetest\\validation\",\n    target_size=(512, 512),\n    color_mode=\"grayscale\",\n    batch_size=32,\n    class_mode=\"categorical\",\n    shuffle=True,\n    seed=42\n)\n#\ntest_generator = test_datagen.flow_from_directory(\n    directory=r\"E:\\databasetest\\test\",\n    target_size=(512, 512),\n    color_mode=\"grayscale\",\n    batch_size=16,\n    class_mode=None,\n    shuffle=False,\n    seed=42\n)\n#\n## neural network model\nmodel = Sequential()\nmodel.add(Conv2D(32, (3,3),border_mode='same', input_shape = (512, 512, 1), activation = 'relu'))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(32, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Conv2D(64, 3, 3))\nmodel.add(Activation('relu'))\n#model.add(Convolution2D(64, 3, 3))\n#model.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(7))\nmodel.add(Activation('softmax'))\n\nmodel.summary()\n\nmodel.compile(loss = 'categorical_crossentropy',\n              optimizer = 'rmsprop',\n              metrics = ['accuracy'])\nSTEP_SIZE_TRAIN=train_generator.n\/\/train_generator.batch_size\nSTEP_SIZE_VALID=valid_generator.n\/\/valid_generator.batch_size\nmodel.fit_generator(generator=train_generator,\n                    steps_per_epoch=STEP_SIZE_TRAIN,\n                    validation_data=valid_generator,\n                    validation_steps=STEP_SIZE_VALID,\n                    epochs=10\n)\n<\/code><\/pre>\n\n<p>but when I implement it I received this error again:<\/p>\n\n<p>ResourceExhaustedError: OOM when allocating tensor with shape[32,32,512,512] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n     [[Node: conv2d_1\/convolution = Conv2D[T=DT_FLOAT, data_format=\"NCHW\", dilations=[1, 1, 1, 1], padding=\"SAME\", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](conv2d_1\/convolution-0-TransposeNHWCToNCHW-LayoutOptimizer, conv2d_1\/kernel\/read)]]<\/p>\n","answers":[{"AnswerId":"52920648","CreationDate":"2018-10-21T23:02:17.640","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<pre><code>    for img in img_list:\n        input_img=cv2.imread(data_path + '\/'+ dataset + '\/'+ img )\n        input_img=cv2.cvtColor(input_img, cv2.COLOR_BGR2GRAY)\n        input_img_resize=cv2.resize(input_img,(512,512))\n    ---&gt;input_img_resize = np.expand_dims(input_img_resize, axis=-1)\n        img_data_list.append(input_img_resize)\n        labels_list.append(label)\n<\/code><\/pre>\n\n<p>this will make all your arrays 512x512x1, which should do the trick and ends up zith an array of shape  (454, 512, 512, 1). You sure you want to use grayscaled images though?<\/p>\n\n<p>Another thing is this snippet of code<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, `random_state=2)`\nimg_data= np.expand_dims(img_data, axis=4)**\nprint (img_data.shape)\n<\/code><\/pre>\n\n<p>You apply another dimension to your img_data, after you've already declared x_train, etc. And in the end you feed x_train, which is not extended, hence the error. If you do it in the beginning, and remove the expanding in the end, then your code should work. <\/p>\n\n<p>EDIT OOM<\/p>\n\n<p>I recommend creating a separate question for the OOM problem, so more people see it. Possible problems are the size of the images and the batch size. Reduce the image size to 64 x 64 and change the batch size to 5. If that still raises the error, try also kicking out this dense layer.<\/p>\n\n<pre><code>model.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\n<\/code><\/pre>\n\n<p>If these reductions still cause the error then I have the following questions: are you running on GPU\/CPU, and which one?<\/p>\n\n<p>Just to repeat myself: the code is fine, just needs a few changes perhaps.<\/p>\n"}]}
+{"QuestionId":52919506,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-21T20:19:35.513","AcceptedAnswerId":52943237.0,"OwnerUserId":10537497.0,"Title":"Is there a way to do matrix-by-tensor multiply?","Body":"<p>Is there any way to get tensor (with batch dimension) multiplication behavior similar to tf.matmul between 2D-matrix where the batch dimension equal to one ? <\/p>\n\n<p>Specifically, I want to do 2D-matrix (6,255) and Tensor (2,255, 255,1) (with batch dimension equal to 2), where :<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nim = np.random.rand(255, 255)\nA = np.random.rand(6, 255)\nB = np.array([im,im]).reshape([-1,255, 255,1])\n\nbatch_size = 2\na = tf.placeholder(tf.float64,shape=(6, 255))\nb = tf.placeholder(tf.float64,shape=(batch_size,255, 255,1))\nout_mat = tf.matmul(a,b) #Didn't work \nwith tf.Session() as sess:\n    sess.run(out_mat, feed_dict={a: A, b: B})\n<\/code><\/pre>\n\n<p>and the result should have (2, 6, 255, 1) (Thanks you @rvinas ) shape.<\/p>\n\n<p>Note: In tensorflow, matmul can only handle 2D-matrices and batch_matmul can\nonly do (...,m,n) by (...,n,p) where ... is the same in both A,B.<\/p>\n","answers":[{"AnswerId":"52943237","CreationDate":"2018-10-23T07:18:09.380","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Here's one way to do it using implicit broadcasting and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_sum\" rel=\"nofollow noreferrer\">tf.reduce_sum<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nbatch_size = 2\ndim_1 = 3\ndim_2 = 4\ndim_3 = 5\ndim_4 = 32\n\nim = np.arange(dim_2 * dim_3).reshape(dim_2, dim_3)\nA = np.arange(dim_1 * dim_2).reshape(dim_1, dim_2)\nB = im\nC = np.matmul(A, B)\nprint('NumPy result (for batch_size=1):\\n {}'.format(C))\n\nB = np.repeat(B[None, ..., None], batch_size, axis=0)\nB = np.repeat(B, dim_4, axis=3)\nprint(B.shape)  # B shape=(batch_size, dim_2, dim_3, dim_4)\n\na = tf.placeholder(tf.float64, shape=(dim_1, dim_2))\nb = tf.placeholder(tf.float64, shape=(batch_size, dim_2, dim_3, dim_4))\na_ = a[None, :, :, None, None]  # Shape=(1, dim_1, dim_2, 1, 1)\nb_ = b[:, None, :, :, :]  # Shape=(batch_size, 1, dim_2, dim_3, dim_4)\nout_mat = tf.reduce_sum(a_ * b_, axis=2)\n\nwith tf.Session() as sess:\n    c = sess.run(out_mat, feed_dict={a: A, b: B})\n    print('TF result (for batch_size={}):\\n {}'.format(batch_size, c))\n    assert c.shape == (batch_size, dim_1, dim_3, dim_4)\n<\/code><\/pre>\n\n<p>And an alternative way using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"nofollow noreferrer\">tf.matmul<\/a>, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/manip\/reshape\" rel=\"nofollow noreferrer\">tf.reshape<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\">tf.transpose<\/a>:<\/p>\n\n<pre><code>b_ = tf.transpose(b, [1, 0, 2, 3])  # Shape=(dim_2, batch_size, dim_3, dim_4)\nb_ = tf.reshape(b_, [dim_2, -1])  # Shape=(dim_2, batch_size * dim_3 * dim_4)\nmatmul = a @ b_  # Shape=(dim_1, batch_size * dim_3 * dim_4)\nmatmul_ = tf.reshape(matmul, [dim_1, batch_size, dim_3, dim_4])\nout_mat = tf.transpose(matmul_, [1, 0, 2, 3])  # Shape=(batch_size, dim_1, dim_3, dim_4)\n<\/code><\/pre>\n\n<p>For your particular example, you would set <code>batch_size=2<\/code>, <code>dim_1=6<\/code>, <code>dim_2=255<\/code>, <code>dim_3=255<\/code> and <code>dim_4=1<\/code>.<\/p>\n"}]}
+{"QuestionId":52919559,"AnswerCount":1,"Tags":"<python><tensorflow><keras><training-data><loss>","CreationDate":"2018-10-21T20:27:00.470","AcceptedAnswerId":null,"OwnerUserId":7094421.0,"Title":"Are those keras loss and accuracy weird?","Body":"<p>I have a relatively small mri dataset and I'm trying to do a binary segmentation. I have built an ordinary U-Net structure and trained it. <\/p>\n\n<p>But the output seems a bit weird to me. Both train and validation accuracies stucked at a value first, but then both accuracies made a sudden big jump at 27th or 28th epoch. <\/p>\n\n<p>Loss graph looks more acceptable, next is the graphs:<\/p>\n\n<p><strong>Accuracy Graph:<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/5UfH8.jpg\" alt=\"Accuracy Graph\"><\/p>\n\n<p><strong>Loss Graph:<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/ueT83.jpg\" alt=\"Loss Graph\"><\/p>\n\n<p>I have another issue that even if I have an <code>%97-98<\/code> accuracy on training data, when I tested it on some images from training data, results converted to binary mask were not that good. <\/p>\n\n<p>Then I have decreased the threshold from 0.5 to 0.35 while retrieving output images and the results were almost perfect.<\/p>\n\n<p>What do you think about that? thanks in advance.<\/p>\n","answers":[{"AnswerId":"52919714","CreationDate":"2018-10-21T20:45:30.490","ParentId":null,"OwnerUserId":"10467184","Title":null,"Body":"<p>They seem a little off with those stuck epochs, it really means the model isn't learning (weights are not changing, new cases are not providing useful information) but that is totally plausible.<\/p>\n\n<p>Just to be sure. What optimizer are you using? and did you try with another one?<\/p>\n"}]}
+{"QuestionId":52919731,"AnswerCount":2,"Tags":"<python><tensorflow><keras><reinforcement-learning><openai-gym>","CreationDate":"2018-10-21T20:47:02.660","AcceptedAnswerId":null,"OwnerUserId":6725958.0,"Title":"DQN with prioritized experience replay and target network does not improve","Body":"<p>I tried to code a neural network to solve OpenAI's CartPole environment with Tensorflow and Keras. The network uses prioritized experience replay and a seperate target network which is updated every tenth episode. Here's the code:<\/p>\n\n<pre><code>import numpy as np\nimport gym\nimport matplotlib.pyplot as plt\nfrom collections import deque\n\nfrom tensorflow import keras\n\nclass agent:\n    def __init__(self,inputs,outputs,gamma):\n        self.gamma = gamma\n        self.epsilon = 1.0\n        self.epsilon_decay = 0.999\n        self.epsilon_min = 0.01\n        self.inputs = inputs\n        self.outputs = outputs\n        self.network = self._build_net()\n        self.target_network = self._build_net()\n        self.replay_memory = deque(maxlen=100000)\n\n        self.target_network.set_weights(self.network.get_weights())\n\n    def _build_net(self):\n        model = keras.models.Sequential()\n\n        model.add(keras.layers.Dense(10,activation='relu',input_dim=self.inputs))\n        model.add(keras.layers.Dense(10,activation='relu'))\n        model.add(keras.layers.Dense(10,activation='relu'))\n        model.add(keras.layers.Dense(self.outputs,activation='sigmoid'))\n        model.compile(optimizer='adam',loss='categorical_crossentropy')\n\n        return model\n\n    def act(self,state,testing=False):\n        if not testing:\n            if self.epsilon &gt; np.random.rand():\n                return np.random.randint(self.outputs)\n            else:\n                return np.argmax(self.network.predict(np.array([state])))\n\n        else:\n            return np.argmax(self.network.predict(np.array([state])))\n\n    def remember(self,state,action,next_state,reward,win):\n        self.replay_memory.append((state,action,next_state,reward,win))\n\n    def get_batch(self):\n        batch = []\n        losses = []\n        targets = []\n\n        if len(self.replay_memory) &gt;= 32:\n            for state,action,next_state,reward,done in self.replay_memory:\n                target_f = np.zeros([1,self.outputs])\n                if done != False:\n                    target = (reward + self.gamma * np.amax(self.target_network.predict(np.array([next_state]))[0]))\n                    target_f[0][action] = target\n\n                targets.append(target_f)\n                loss = np.mean(self.network.predict(np.array([state]))-target_f)**2\n                losses.append(loss)\n\n            indexes = np.argsort(losses)[:32]\n\n            for indx in indexes:\n                batch.append((self.replay_memory[indx][0], targets[indx]))\n\n            return batch\n\n    def replay(self,batch):\n        for state,target in batch:\n            self.network.fit(np.array([state]), target, epochs=1, verbose=0)\n        self.epsilon = max(self.epsilon_min, self.epsilon*self.epsilon_decay)\n\n    def update_target_network(self):\n        self.target_network.set_weights(self.network.get_weights())\n\n    def save(self):\n        self.network.save_weights('.\/DQN.model')\n\n    def load(self):\n        self.network.load_weights('.\/DQN.model')\n\nenv = gym.make('CartPole-v0')\nepisodes = 1000\nagent = agent(env.observation_space.shape[0],env.action_space.n,0.99)\nrewards = []\n\nstate = env.reset()\nfor _ in range(500):\n    action = agent.act(state,1.0)\n    next_state, reward, win, _ = env.step(action)\n    agent.remember(state,action,next_state,reward,win)\n    state = next_state\n\n    if win:\n        state = env.reset()\n\nfor e in range(episodes):\n    print('episode:',e+1)\n    batch = agent.get_batch()\n    state = env.reset()\n    for t in range(400):\n        action = agent.act(state)\n        next_state, reward, win, _ = env.step(action)\n        agent.remember(state,action,next_state,reward,win)\n        state = next_state\n\n        if win:\n            break\n\n        agent.replay(batch)\n\n    print('score:',t)\n    print('epsilon:',agent.epsilon)\n    print('')\n\n    if e%10 == 0:\n        agent.update_target_network()\n        agent.save()\n\n    rewards.append(t)\n    plt.plot(list(range(e+1)),rewards)\n    plt.savefig('.\/reward.png')\n<\/code><\/pre>\n\n<p>The problem is that the agent gets worse as epsilon decreases. And when epsilon is a its lowest value, the agent only gets through 7-9 steps before the Pole falls, as seen in the Image below. Can someone tell me why my agent isn't learning anything and how to fix it?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xFKQi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xFKQi.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"56403919","CreationDate":"2019-06-01T04:44:40.027","ParentId":null,"OwnerUserId":"8878627","Title":null,"Body":"<p>Try making the reward = -reward if done. It helps<\/p>\n"},{"AnswerId":"56471417","CreationDate":"2019-06-06T05:34:52.233","ParentId":null,"OwnerUserId":"4625843","Title":null,"Body":"<p>Your network output should be using <strong>linear<\/strong> as activation as opposed to <strong>sigmoid<\/strong>, since you are bootstrapping summation of all future rewards.  The loss should be changed to <strong>mse<\/strong> as well, according to the derivation for the td error from the original paper.<\/p>\n\n<p>Also, if the experience ended in a terminal state (done==True), the td target for the action taken should set equal to the reward (you are currently setting it as 0).<\/p>\n"}]}
+{"QuestionId":52920098,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-10-21T21:33:11.840","AcceptedAnswerId":52928417.0,"OwnerUserId":458274.0,"Title":"What does the underscore suffix in PyTorch functions mean?","Body":"<p>In PyTorch, many methods of a tensor exist in two versions - one with a underscore suffix, and one without. If I try them out, they seem to do the same thing:<\/p>\n\n<pre><code>In [1]: import torch\n\nIn [2]: a = torch.tensor([2, 4, 6])\n\nIn [3]: a.add(10)\nOut[3]: tensor([12, 14, 16])\n\nIn [4]: a.add_(10)\nOut[4]: tensor([12, 14, 16])\n<\/code><\/pre>\n\n<p>What is the difference between<\/p>\n\n<ul>\n<li><code>torch.add<\/code> and <code>torch.add_<\/code><\/li>\n<li><code>torch.sub<\/code> and <code>torch.sub_<\/code><\/li>\n<li>...and so on?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"52920099","CreationDate":"2018-10-21T21:33:11.840","ParentId":null,"OwnerUserId":"458274","Title":null,"Body":"<p>According to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html?highlight=underscore\" rel=\"nofollow noreferrer\">documentation<\/a>, Methods which end in an underscore change the tensor <strong>in-place<\/strong>. That means that no new memory is being allocated by doing the operation, which <strong>in general<\/strong> <a href=\"https:\/\/stackoverflow.com\/q\/470683\/458274\">increase performance<\/a>, but <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd\" rel=\"nofollow noreferrer\">can lead to problems and <strong>worse<\/strong> performance in PyTorch<\/a>.<\/p>\n\n<pre><code>In [2]: a = torch.tensor([2, 4, 6])\n<\/code><\/pre>\n\n<p><strong>tensor.add()<\/strong>:<\/p>\n\n<pre><code>In [3]: b = a.add(10)\n\nIn [4]: a is b\nOut[4]: False # b is a new tensor, new memory was allocated\n<\/code><\/pre>\n\n<p><strong>tensor._add()<\/strong>:<\/p>\n\n<pre><code>In [3]: b = a.add_(10)\n\nIn [4]: a is b\nOut[4]: True # Same object, no new memory was allocated\n<\/code><\/pre>\n\n<p>Notice, that the operators <code>+<\/code> and <code>+=<\/code> are also <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/720c7b1e2c9580fa897110abca3a41ae676174da\/torch\/tensor.py#L256\" rel=\"nofollow noreferrer\">two different implementations<\/a>. <code>+<\/code> creates a new tensor by using <code>.add()<\/code>, while <code>+=<\/code> modifies the tensor by using <code>.add_()<\/code><\/p>\n\n<pre><code>In [2]: a = torch.tensor([2, 4, 6])\n\nIn [3]: id(a)\nOut[3]: 140250660654104\n\nIn [4]: a += 10\n\nIn [5]: id(a)\nOut[5]: 140250660654104 # Still the same object, no memory allocation was required\n\nIn [6]: a = a + 10\n\nIn [7]: id(a)\nOut[7]: 140250649668272 # New object was created\n<\/code><\/pre>\n"},{"AnswerId":"52928417","CreationDate":"2018-10-22T11:35:52.063","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p><em>You have already answered your own question that the underscore indicates in-place operations in PyTorch. However I want to point out briefly why in-place operations can be problematic:<\/em><\/p>\n\n<ul>\n<li><p>First of all on the PyTorch sites it is <strong>recommended to not use in-place operations<\/strong> in most cases. Unless working under heavy memory pressure it is <strong>more efficient<\/strong> in most cases to <strong>not use in-place operations<\/strong>.<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd\" rel=\"noreferrer\">https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd<\/a><\/p><\/li>\n<li><p>Secondly there can be problems calculating the gradients when using in-place operations:<\/p>\n\n<blockquote>\n  <p>Every tensor keeps a version counter, that is incremented every time\n  it is marked dirty in any operation. When a Function saves any tensors\n  for backward, a version counter of their containing Tensor is saved as\n  well. Once you access <code>self.saved_tensors<\/code> it is checked, and if it is\n  greater than the saved value an error is raised. This ensures that if\n  you\u2019re using in-place functions and not seeing any errors, you can be\n  sure that the computed gradients are correct.\n  <em><a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd\" rel=\"noreferrer\">Same source as above.<\/a><\/em><\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p>Here is a shot and slightly modified example taken from the answer you've posted:<\/p>\n\n<p><strong><em>First the in-place version:<\/em><\/strong><\/p>\n\n<pre><code>import torch\na = torch.tensor([2, 4, 6], requires_grad=True, dtype=torch.float)\nadding_tensor = torch.rand(3)\nb = a.add_(adding_tensor)\nc = torch.sum(b)\nc.backward()\nprint(c.grad_fn)\n<\/code><\/pre>\n\n<p><strong><em>Which leads to this error:<\/em><\/strong><\/p>\n\n<pre><code>---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-27-c38b252ffe5f&gt; in &lt;module&gt;\n      2 a = torch.tensor([2, 4, 6], requires_grad=True, dtype=torch.float)\n      3 adding_tensor = torch.rand(3)\n----&gt; 4 b = a.add_(adding_tensor)\n      5 c = torch.sum(b)\n      6 c.backward()\n\nRuntimeError: a leaf Variable that requires grad has been used in an in-place operation.\n<\/code><\/pre>\n\n<hr>\n\n<p><strong><em>Secondly the non in-place version:<\/em><\/strong><\/p>\n\n<pre><code>import torch\na = torch.tensor([2, 4, 6], requires_grad=True, dtype=torch.float)\nadding_tensor = torch.rand(3)\nb = a.add(adding_tensor)\nc = torch.sum(b)\nc.backward()\nprint(c.grad_fn)\n<\/code><\/pre>\n\n<p><strong><em>Which works just fine - output:<\/em><\/strong><\/p>\n\n<pre><code>&lt;SumBackward0 object at 0x7f06b27a1da0&gt;\n<\/code><\/pre>\n\n<hr>\n\n<p>So as a take-away I just wanted to point out to carefully use in-place operations in PyTorch.<\/p>\n"}]}
+{"QuestionId":52920222,"AnswerCount":2,"Tags":"<python><windows><tensorflow><object-detection><pycocotools>","CreationDate":"2018-10-21T21:50:34.033","AcceptedAnswerId":53030222.0,"OwnerUserId":9136356.0,"Title":"Could not install pycocotools in windows: fatal error C1083: Cannot open include file: 'io.h': No such file or directory error:","Body":"<p>I'm new to machine learning and have started with a windows 8.1 pc having GeForce GTX 540M. I followed this <a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\" rel=\"nofollow noreferrer\">tutorial<\/a> to get started with the object detection models. I built my own dataset and tried to train it as per the tutorial but with \"ssd_mobilenet_v1_coco_2017_11_17\" model. But could not perform it successfully as i had troubles with the \"train.py\" file given in the tutorial.  <\/p>\n\n<p>So I googled and found that we have to use \"model_main.py\" to train the model. While trying to train using \"model_main.py\" I got this error:<br>\n<code>\"ImportModuleError: No module named 'pycocotools' was found\"<\/code>  <\/p>\n\n<p>So I tried to install the cocoapi from this <a href=\"https:\/\/github.com\/philferriere\/cocoapi\" rel=\"nofollow noreferrer\">link<\/a> . I followed the instructions as given, but got stuck with the error shown below, I've tried a lot of solutions but nothing worked. Any help will be much appreciated.<\/p>\n\n<pre><code>(tensorflow1) C:\\tensorflow1\\models\\research\\object_detection&gt;pip install git+https:\/\/github.com\/philferriere\/cocoapi.git#subdirectory=PythonAPI\nCollecting git+https:\/\/github.com\/philferriere\/cocoapi.git#subdirectory=PythonAPI\n  Cloning https:\/\/github.com\/philferriere\/cocoapi.git to c:\\users\\home\\appdata\\l\nocal\\temp\\pip-req-build-76t4e6ys\nBuilding wheels for collected packages: pycocotools\n  Running setup.py bdist_wheel for pycocotools ... error\n  Complete output from command c:\\users\\home\\anaconda3\\envs\\tensorflow1\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Home\\\\AppData\\\\Local\\\\Temp\\\\pip-req-build-76t4e6ys\\\\PythonAPI\\\\setup.py';f=getattr(tokenize, 'open',open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" bdist_wheel -d C:\\Users\\Home\\AppData\\Local\\Temp\\pip-wheel-4_r37hm6 --python-tag cp36:\n  running bdist_wheel\n  running build\n  running build_py\n  creating build\n  creating build\\lib.win-amd64-3.6\n  creating build\\lib.win-amd64-3.6\\pycocotools\n  copying pycocotools\\coco.py -&gt; build\\lib.win-amd64-3.6\\pycocotools\n  copying pycocotools\\cocoeval.py -&gt; build\\lib.win-amd64-3.6\\pycocotools\n  copying pycocotools\\mask.py -&gt; build\\lib.win-amd64-3.6\\pycocotools\n  copying pycocotools\\__init__.py -&gt; build\\lib.win-amd64-3.6\\pycocotools\n  running build_ext\n  building 'pycocotools._mask' extension\n  creating build\\temp.win-amd64-3.6\n  creating build\\temp.win-amd64-3.6\\Release\n  creating build\\temp.win-amd64-3.6\\Release\\pycocotools\n  creating build\\temp.win-amd64-3.6\\common\n  C:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\bin\\HostX86\\x64\\cl.exe \/c\/nologo \/Ox \/W3 \/GL \/DNDEBUG \/MD -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\lib\\site-packages\\numpy\\core\\include -I..\/common -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\include -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\include \"-IC:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\ATLMFC\\include\" \"-IC:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\include\" \/Tcpycocotools\/_mask.c \/Fobuild\\temp.win-amd64-3.6\\Release\\pycocotools\/_mask.obj\n  _mask.c\n  c:\\users\\home\\anaconda3\\envs\\tensorflow1\\include\\pyconfig.h(59): fatal error C1083: Cannot open include file: 'io.h': No such file or directory\n  error: command 'C:\\\\Program Files (x86)\\\\Microsoft Visual Studio\\\\2017\\\\Community\\\\VC\\\\Tools\\\\MSVC\\\\14.15.26726\\\\bin\\\\HostX86\\\\x64\\\\cl.exe' failed with exit status 2\n\n  ----------------------------------------\n  Failed building wheel for pycocotools\n  Running setup.py clean for pycocotools\n  Complete output from command c:\\users\\home\\anaconda3\\envs\\tensorflow1\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Home\\\\AppData\\\\Local\\\\Temp\\\\pip-req-build-76t4e6ys\\\\PythonAPI\\\\setup.py';f=getattr(tokenize, 'open',open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" clean --all:\n  Traceback (most recent call last):\n    File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    File \"C:\\Users\\Home\\AppData\\Local\\Temp\\pip-req-build-76t4e6ys\\PythonAPI\\setup.py\", line 25, in &lt;module&gt;\n      cythonize(ext_modules)\n    File \"c:\\users\\home\\anaconda3\\envs\\tensorflow1\\lib\\site-packages\\Cython\\Build\\Dependencies.py\", line 956, in cythonize\n      aliases=aliases)\n    File \"c:\\users\\home\\anaconda3\\envs\\tensorflow1\\lib\\site-packages\\Cython\\Buil\nd\\Dependencies.py\", line 801, in create_extension_list\n      for file in nonempty(sorted(extended_iglob(filepattern)), \"'%s' doesn't ma\ntch any files\" % filepattern):\n    File \"c:\\users\\home\\anaconda3\\envs\\tensorflow1\\lib\\site-packages\\Cython\\Buil\nd\\Dependencies.py\", line 111, in nonempty\n      raise ValueError(error_msg)\n  ValueError: 'pycocotools\/_mask.pyx' doesn't match any files\n\n  ----------------------------------------\n  Failed cleaning build dir for pycocotools\nFailed to build pycocotools\nInstalling collected packages: pycocotools\n  Running setup.py install for pycocotools ... error\n    Complete output from command c:\\users\\home\\anaconda3\\envs\\tensorflow1\\python.exe -u -c \"import setuptools, tokenize;__file__='C:\\\\Users\\\\Home\\\\AppData\\\\Local\\\\Temp\\\\pip-req-build-76t4e6ys\\\\PythonAPI\\\\setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" build_ext \"-L\/C:\/Program Files (x86)\/Windows Kits\/10\/Include\/10.0.17134.0\/ucrt\" install --record C:\\Users\\Home\\AppData\\Local\\Temp\\pip-record-ykypadav\\install-record.txt --single-version-externally-managed --compile:\n    running build_ext\n    building 'pycocotools._mask' extension\n    C:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\bin\\HostX86\\x64\\cl.exe \/c\/nologo \/Ox \/W3 \/GL \/DNDEBUG \/MD -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\lib\\site-packages\\numpy\\core\\include -I..\/common -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\include -Ic:\\users\\home\\anaconda3\\envs\\tensorflow1\\include \"-IC:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\ATLMFC\\include\" \"-IC:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Tools\\MSVC\\14.15.26726\\include\" \/Tcpycocotools\/_mask.c \/Fobuild\\temp.win-amd64-3.6\\Release\\pycocotools\/_mask.obj\n    _mask.c\n    c:\\users\\home\\anaconda3\\envs\\tensorflow1\\include\\pyconfig.h(59): fatal error\n C1083: Cannot open include file: 'io.h': No such file or directory\n    error: command 'C:\\\\Program Files (x86)\\\\Microsoft Visual Studio\\\\2017\\\\Community\\\\VC\\\\Tools\\\\MSVC\\\\14.15.26726\\\\bin\\\\HostX86\\\\x64\\\\cl.exe' failed with exit\n status 2\n\n    ----------------------------------------\nCommand \"c:\\users\\home\\anaconda3\\envs\\tensorflow1\\python.exe -u -c \"import setup\ntools, tokenize;__file__='C:\\\\Users\\\\Home\\\\AppData\\\\Local\\\\Temp\\\\pip-req-build-76t4e6ys\\\\PythonAPI\\\\setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" build_ext \"-L\/C:\/Program Files (x86)\/Windows Kits\/10\/Include\/10.0.17134.0\/ucrt\" install --record C:\\Users\\Home\\AppData\\Local\\Temp\\pip-record-ykypadav\\install-record.txt --single-version-externally-managed --compile\" failed with error code 1\nin C:\\Users\\Home\\AppData\\Local\\Temp\\pip-req-build-76t4e6ys\\PythonAPI\n\n(tensorflow1) C:\\tensorflow1\\models\\research\\object_detection&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"55493756","CreationDate":"2019-04-03T11:11:03.723","ParentId":null,"OwnerUserId":"7454706","Title":null,"Body":"<p>First install Visual C++ 2015 Build Tools from <a href=\"https:\/\/go.microsoft.com\/fwlink\/?LinkId=691126\" rel=\"nofollow noreferrer\">https:\/\/go.microsoft.com\/fwlink\/?LinkId=691126<\/a> with default selection.\nThen,<br>\n<code>pip install git+https:\/\/github.com\/philferriere\/cocoapi.git#egg=pycocotools^&amp;subdirectory=PythonAPI<\/code><\/p>\n"},{"AnswerId":"53030222","CreationDate":"2018-10-28T09:51:07.540","ParentId":null,"OwnerUserId":"9236858","Title":null,"Body":"<p>I was getting the same error<\/p>\n\n<pre><code> ValueError: 'pycocotools\/_mask.pyx' doesn't match any files\n<\/code><\/pre>\n\n<p>So I cloned <a href=\"https:\/\/github.com\/waleedka\/coco\" rel=\"nofollow noreferrer\">https:\/\/github.com\/waleedka\/coco<\/a> to my local device. Then when I was running make I was getting the above error.\nthe exact command I was running was--<\/p>\n\n<pre><code>C:\\Users\\Admin\\Desktop\\coco-master&gt;python3 PythonAPI\/setup.py build_ext install\n<\/code><\/pre>\n\n<p>I was when I was running above command I got the problem.\nThen I found out a solution on the issues page of the same repository. It happens that you have to cd into PythonAPI then run--<\/p>\n\n<pre><code>python setup.py build_ext --inplace\n<\/code><\/pre>\n\n<p>That solved some problems for me.\nMoreover some people were saying they still got problems but when they used python3 instead of python their problem was solved.<\/p>\n\n<p>Hope you find this answer useful.<\/p>\n"}]}
+{"QuestionId":52920476,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-21T22:28:50.497","AcceptedAnswerId":52920947.0,"OwnerUserId":3801449.0,"Title":"Tensor - define function while evaluating other one","Body":"<p>I need to define an error function looking like this: <code>e = (y(0) - 1)^2<\/code>.<\/p>\n\n<p>My <code>y<\/code> function\/tensor looks like this:<\/p>\n\n<pre><code>n = 10\n\nw = tf.Variable(np.random.random((1, n)))\nb = tf.Variable(np.random.random(n))\na = tf.Variable(np.random.random(n))\ns = tf.placeholder(\"float64\")\nf = tf.nn.sigmoid(-s * w + b)\n\ny_1 = tf.reduce_sum(a * f, axis=1)\n<\/code><\/pre>\n\n<p>I have read, that all the variables have to be declared before the session start.<\/p>\n\n<p>I'd define the error function\/tensor like this:<\/p>\n\n<pre><code>e = tf.reduce_sum((y_1[0].eval(feed_dict={s: 0}) - 1) ** 2)\n<\/code><\/pre>\n\n<p>But, I'm not able to use <code>eval()<\/code> method outside the session block.<\/p>\n\n<p>So, how should I define the tensor containing <code>y_1(0)<\/code>?<\/p>\n","answers":[{"AnswerId":"52920947","CreationDate":"2018-10-21T23:59:11.770","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>You shouldn't evaluate y_1[0] if you want to use it inside the graph.  Why not just use<\/p>\n\n<pre><code>e = tf.reduce_sum((y_1[0]) - 1) ** 2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52920618,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-21T22:57:17.767","AcceptedAnswerId":null,"OwnerUserId":3241790.0,"Title":"Make custom loss function for Keras","Body":"<p>I would implement custom Keras loss function for focal loss which explained in this <a href=\"https:\/\/arxiv.org\/pdf\/1708.02002.pdf\" rel=\"nofollow noreferrer\">paper<\/a>; however, I have to get the weight of the trainable layers. So, is there any way to read the weight of the layer during the training inside the loss function?<\/p>\n\n<pre><code>def focal_coef(model, y_true, y_pred, content, label_remap, gamma_focal=2, w_d=1e-4):\n    w = np.empty(len(content))\n    f = np.empty(len(content))\n    for key in content:\n        e = 0.001\n        f_c = content[key]\n        f[label_remap[key]] = f_c\n        w[label_remap[key]] = 1 \/ (f_c + e)\n    median_freq = np.median(f)\n    print(\"\\nFrequencies of classes:\\n\", f)\n    print(\"\\nMedian freq:\\n\", median_freq)\n    print(\"\\nWeights for loss function (1\/frec(c)):\\n\", w)\n    w = median_freq * w\n    print(\"\\nWeights for loss function (median frec\/frec(c)):\\n\", w)\n    w_mask = w.astype(np.bool).astype(np.float32)\n    print(\"w_mask\", w_mask)\n    softmax = k.softmax(y_pred)\n    softmax_mat = k.reshape(softmax, (-1, len(content)))\n    zerohot_softmax_mat = 1 - softmax_mat\n    # make the labels one-hot for the cross-entropy\n    print(\"datatype\", y_true.type)\n    onehot_mat = k.reshape(k.one_hot(y_true, len(content)),\n                       (-1, len(content)))\n    # make the zero hot to punish the false negatives, but ignore the\n    # zero-weight classes\n    masked_sum = k.reduce_sum(onehot_mat * w_mask, axis=1)\n    zeros = onehot_mat * 0.0\n    zerohot_mat = k.where(k.less(masked_sum, 1e-5),\n                      x=zeros,\n                      y=1 - onehot_mat)\n    # focal loss p and gamma\n    loss_epsilon = 1e-10\n    gamma = np.full(onehot_mat.get_shape().as_list(), fill_value=gamma_focal)\n    gamma_tf = k.constant(gamma, dtype=k.float32)\n    focal_softmax = k.pow(1 - softmax_mat, gamma_tf) * \\\n                k.log(softmax_mat + loss_epsilon)\n    zerohot_focal_softmax = k.pow(1 - zerohot_softmax_mat, gamma_tf) * \\\n                        k.log(zerohot_softmax_mat + loss_epsilon)\n\n    # calculate xentropy\n    cross_entropy = - k.reduce_sum(k.multiply(focal_softmax * onehot_mat +\n                                          zerohot_focal_softmax * zerohot_mat, w),\n                               axis=[1])\n    loss = k.reduce_mean(cross_entropy, name='xentropy_mean')\n    for layer in model.layers:\n        if layer.trainable == True:\n            loss += w_d * (k.sum(layer ** 2) \/ 2)\n    return loss\n\ndef focal_loss(model, content, label_remap, gamma_focal=2, w_d=1e-4):\n    def focal(y_true, y_pred):\n        return -focal_coef(model, y_true, y_pred, content, label_remap, gamma_focal, w_d)\n\n    return focal\n<\/code><\/pre>\n","answers":[{"AnswerId":"52926680","CreationDate":"2018-10-22T09:59:15.983","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Your custom loss function must have two arguments these are y_true and y_pred, if you want more parameters you can call a function inside of your loss function to get the parameters, since you want weights from layers, your custom loss function must have access to your model<\/p>\n\n<pre><code>model = ...\n\ndef custom_loss(y_true, y_pred):\n    w1 = model.get_layers[0].get_weights()[0] # weights of the first layer\n    b1 = model.get_layers[0].get_weights()[1] # bias of the first layer\n    w2 = model.get_layers[1].get_weights()[0] # weights of the second layer\n    .\n    .\n    .\n    loss = ...\n    return loss\nmodel.compile(loss=custom_loss, optimizer = ...)\nmodel.fit(...)\n<\/code><\/pre>\n\n<p>If you want all of the weights you can do<code>model.get_weights()<\/code><\/p>\n"}]}
+{"QuestionId":52920798,"AnswerCount":1,"Tags":"<python><keras><mnist>","CreationDate":"2018-10-21T23:35:43.003","AcceptedAnswerId":52920851.0,"OwnerUserId":1556227.0,"Title":"MNIST dataset missing classes","Body":"<p>I'm training a ConvNet using Keras and Theano, but before doing that I decided to take a peek into the dataset, its data samples and classes... And I don't like what I'm seeing.<\/p>\n\n<p>I'm using the following code to load both training and test datasets and count how many data samples are labeled for each one:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.datasets import mnist\n\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\n\ntrain_classes = [0,0,0,0,0,0,0,0,0,0]\ntest_classes = [0,0,0,0,0,0,0,0,0,0]\n\nfor i in y_train:\n    train_classes[y_train[i]] = train_classes[y_train[i]] + 1\n\nfor i in y_test:\n    test_classes[y_test[i]] = test_classes[y_test[i]] + 1\n\nprint('Training classes: ', train_classes)\nprint('\\nTesting classes: ', test_classes)\n<\/code><\/pre>\n\n<p>... And the results are worrying:<\/p>\n\n<pre><code>(ann) C:\\Users\\shado\\mnist&gt;python statistics.py\nUsing Theano backend.\nTraining classes:  [6742, 17900, 5421, 6265, 11907, 5923, 0, 0, 0, 5842]\n\nTesting classes:  [1010, 1924, 1135, 0, 1940, 974, 0, 980, 0, 2037]\n<\/code><\/pre>\n\n<p>So as you can see from the label counts, the training dataset is missing the '6', '7' and '8' classes, while the testing dataset is missing the '3', '6' and '8' classes. And of course, the class distribution is all over the place, specially on the training dataset.<\/p>\n\n<p>Am I downloading the wrong dataset? Am I missing something here?<\/p>\n","answers":[{"AnswerId":"52920851","CreationDate":"2018-10-21T23:44:03.913","ParentId":null,"OwnerUserId":"4565947","Title":null,"Body":"<p>The logic you need is:<\/p>\n\n<pre><code>for i in y_train:\n    train_classes[i] += 1\n<\/code><\/pre>\n\n<p>since <code>i<\/code> is already the label.<\/p>\n\n<p>Or equivalently:<\/p>\n\n<pre><code>for i in range(len(y_train)):\n    train_classes[y_train[i]] += 1\n<\/code><\/pre>\n\n<p>Your current code is essentially randomly sampling the first 10 labels of each sets...<\/p>\n\n<p>Note: you can also simply use: <code>np.unique(y_train, return_counts=True)<\/code>.<\/p>\n"}]}
+{"QuestionId":52920988,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><lstm><rnn>","CreationDate":"2018-10-22T00:07:10.147","AcceptedAnswerId":null,"OwnerUserId":6750211.0,"Title":"tensorflow dynamic_rnn how to save and feed hidden states","Body":"<p>I've got a tensorflow multiple layer rnn cell like this:<\/p>\n\n<pre><code>def MakeLSTMCell(self):\n    cells = []\n    for n in self.numUnits:\n        cell = tf.nn.rnn_cell.LSTMCell(n)\n        dropout = tf.nn.rnn_cell.DropoutWrapper(cell,\n                                                input_keep_prob=self.keep_prob,\n                                                output_keep_prob=self.keep_prob)\n        cells.append(dropout)\n    stackedRNNCell = tf.nn.rnn_cell.MultiRNNCell(cells)\n    return stackedRNNCell\n\ndef BuildGraph(self):\n    \"\"\"\n    Build the Graph of the recurrent reinforcement neural network.\n    \"\"\"\n    with self.graph.as_default():\n        with tf.variable_scope(self.scope):\n            self.inputSeq = tf.placeholder(tf.float32, [None, None, self.observationDim], name='input_seq')\n            self.batch_size = tf.shape(self.inputSeq)[0]\n            self.seqLength = tf.shape(self.inputSeq)[1]\n            self.cell = self.MakeLSTMCell()\n\n            with tf.name_scope(\"LSTM_layers\"):\n                self.zeroState = self.cell.zero_state(self.batch_size, tf.float32)\n                self.cellState = self.zeroState\n\n                self.outputs, self.outputState = tf.nn.dynamic_rnn(self.cell,\n                                                         self.inputSeq,\n                                                         initial_state=self.cellState,\n                                                         swap_memory=True)\n<\/code><\/pre>\n\n<p>However, this <code>self.cellState<\/code> is not configurable. I would like to know how could I save the lstm hidden state (keeps the same form so that I could feed it back to the rnn at any time) and reuse it at any time as <code>initial_state<\/code>?<\/p>\n\n<p>I've tried the accepted answer in this question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/37969065\/tensorflow-best-way-to-save-state-in-rnns\">Tensorflow, best way to save state in RNNs?<\/a>\nHowever, dynamic batch size is not allowed when creating tf Variable.<\/p>\n\n<p>Any help will be appreciated<\/p>\n","answers":[]}
+{"QuestionId":52921033,"AnswerCount":2,"Tags":"<javascript><node.js><tensorflow><tensorflow.js>","CreationDate":"2018-10-22T00:19:16.420","AcceptedAnswerId":52934927.0,"OwnerUserId":10538017.0,"Title":"how to produce a tensor containing the indices of nonzero elements?","Body":"<p>I have a tensor like <\/p>\n\n<pre><code> [[0, 0, 0, 1, 0, 0],\n  [0, 3, 0, 0, 0, 1],\n  [0, 0, 0, 0, 0, 0],\n  [0, 0, 0, 0, 0, 0],\n  [0, 0, 0, 2, 0, 0],\n  [0, 0, 0, 0, 0, 0]]\n<\/code><\/pre>\n\n<p>How do i produce a tensor containing only the non-zero indices of the former?<\/p>\n\n<pre><code>[[0,3],[1,1],[1,5],[4,3]]\n<\/code><\/pre>\n\n<p>I'm on tfjs so it doesn't appear to have boolean_mask or some other helpful functions others seem to use in indexing. Is this possible without using the <code>.data()<\/code> escape hatch and map\/filtering over the array? (i'm just using tensorflow for its fast linear algebra methods so that wouldn't be the worst, right?)<\/p>\n","answers":[{"AnswerId":"52934927","CreationDate":"2018-10-22T17:44:27.987","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>True that currently there are no operators one can use to define a partition using a mask. You can consider this <a href=\"https:\/\/stackoverflow.com\/questions\/51218420\/partition-or-mask-or-filter-a-tensor-in-tensorflow-js\/51902433#51902433\">answer<\/a> as a workaround. But even using a mask, the issue here cannot be solved easily because creating a partition supposes that one knows beforehand the indexes to retrieve.<\/p>\n\n<p>To solve the issue, one might use the <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.0\/#logicalAnd\" rel=\"nofollow noreferrer\">logicalAnd<\/a> between the input tensor and a tensor of the same shape with all the elements equal to 1. It is usefull because it will return a tensor of the same shape than the input shape with values 0 if the initial element was 0 and 1 if not.Using <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.0\/#argMax\" rel=\"nofollow noreferrer\">argMax<\/a>, one can retrieve the index of the elements that are equal to 1<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>let t = tf.tensor2d([[0, 0, 0, 1, 0, 0], [0, 3, 0, 0, 0, 1], [0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0], [0, 0, 0, 2, 0, 0], [0, 0, 0, 0, 0, 0]])\r\nlet a = t.as1D()\r\n\r\nconst b = tf.cast(a, 'bool')\r\nlet helper = b.logicalAnd(tf.cast(tf.ones([a.size]), 'bool'))\r\nconst n = helper.sum().dataSync()[0]\r\nconst noNull = []\r\nfor(let i = 0; i &lt; n ; i++) {\r\n  let ind = tf.argMax(helper)\r\n  let s = ind.dataSync()[0]\r\n  noNull.push(tf.argMax(helper).dataSync()[0])\r\n  if (s === 0) {\r\n     const [x, y] = helper.split([1, helper.size - 1])\r\n     helper = tf.concat(tf.tensor1d([0]), y)\r\n  } else if (s === helper.size) {\r\n    const [x, y] = helper.split([helper.size -1, 1])\r\n    helper = tf.concat(x, tf.tensor1d([0]))\r\n  } else {\r\n    const [x, _, y] = helper.split([s, 1, helper.size - s - 1])\r\n    helper = tf.concat([x,tf.tensor1d([0]), y])\r\n  }\r\n}\r\n\r\nconst indexToCoords = (index, shape) =&gt; {\r\n  const pseudoShape = shape.map((a, b, c) =&gt; c.slice(b + 1).reduce((a, b) =&gt; a * b, 1))\r\n  let coords = []\r\n  let ind = index\r\n  for (let i = 0; i &lt; shape.length; i++) {\r\n    coords.push(Math.floor(ind \/ pseudoShape[i]))\r\n    ind = ind % pseudoShape[i]\r\n  }\r\n  return coords\r\n}\r\n\r\nconst coords = noNull.map(e =&gt; indexToCoords(e, t.shape))\r\n\r\nconsole.log(coords)<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Following this <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/?utm_medium=email&amp;utm_source=footer#!msg\/tfjs\/rStLUO5o--A\/12NCVV2DBgAJ\" rel=\"nofollow noreferrer\">thread<\/a>, there is a much simpler way to achieve the same thing<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>(async function() {\r\n  const x = tf.tensor2d(\r\n    [[0, 0, 0, 1, 0, 0],\r\n    [0, 3, 0, 0, 0, 1],\r\n    [0, 0, 0, 0, 0, 0],\r\n    [0, 0, 0, 0, 0, 0],\r\n    [0, 0, 0, 2, 0, 0],\r\n    [0, 0, 0, 0, 0, 0]]);\r\n\r\n  const mask = x.greater([0]).asType('bool');\r\n  const coords = await tf.whereAsync(mask);\r\n\r\n  coords.print();\r\n}());<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"},{"AnswerId":"57973842","CreationDate":"2019-09-17T12:05:00.373","ParentId":null,"OwnerUserId":"375050","Title":null,"Body":"<p>You can use <code>tf.whereAsync(condition)<\/code> to return the indices. See the <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#whereAsync\" rel=\"nofollow noreferrer\">docs<\/a><\/p>\n\n<p>Example:<\/p>\n\n<pre><code>(await tf.whereAsync(tf.tensor1d([0,1,2,0,1]).notEqual(0))).print();\n==&gt; Tensor\n    [[1],\n     [2],\n     [4]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52921102,"AnswerCount":1,"Tags":"<boost><cmake><caffe><pycaffe>","CreationDate":"2018-10-22T00:31:50.063","AcceptedAnswerId":53168620.0,"OwnerUserId":6323596.0,"Title":"Error installing caffe due to boost-python","Body":"<p>I'm trying to install pycaffe, but am failing the command <code>make pycaffe<\/code> with the error:<\/p>\n\n<pre><code>ld: library not found for -lboost_python\nclang: error: linker command failed with exit code 1 (use -v to see invocation)\nmake: *** [python\/caffe\/_caffe.so] Error 1\n<\/code><\/pre>\n\n<p>I installed <code>boost-python<\/code> with the command <code>brew install --build-from-source -vd boost boost-python<\/code>. In my <code>\/usr\/local\/Cellar\/boost-python<\/code> folder, I have v1.67 with files <code>libboost_python27.dylib<\/code> and <code>libboost_python27-mt.dylib<\/code>, as well as their respective <code>.a<\/code> files.<\/p>\n\n<p>Since caffe is looking for <code>lboost_python<\/code> and not <code>libboost_python<\/code>, should I just simlink one to the other? How would I go about this?<\/p>\n","answers":[{"AnswerId":"53168620","CreationDate":"2018-11-06T08:59:37.837","ParentId":null,"OwnerUserId":"9427281","Title":null,"Body":"<p>If you are using python3<\/p>\n\n<pre><code>brew install boost --with-python3\nbrew install boost-python\n<\/code><\/pre>\n\n<p>Confirm boost and boost-python is installed<\/p>\n\n<pre><code>    brew list | grep 'boost'\n<\/code><\/pre>\n\n<p>Then create symlink<\/p>\n\n<pre><code> Example : sudo ln -s libboost_python37.dylib libboost_python3.dylib\n<\/code><\/pre>\n\n<p>This should solve your issue<\/p>\n"}]}
+{"QuestionId":52921819,"AnswerCount":2,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-10-22T02:52:14.047","AcceptedAnswerId":null,"OwnerUserId":10243093.0,"Title":"Why in ConvLSTM when setting return_sequence = False you get an error?","Body":"<p>I have tried to model by attached three layer of ConvLSTM but when I set in the first ConvLSTM return_sequence = False program won't run.<\/p>\n\n<p>See the model summary\n<a href=\"https:\/\/i.stack.imgur.com\/bQFOx.png\" rel=\"nofollow noreferrer\">Model summary<\/a><\/p>\n\n<p>The model run after I set return_sequence = True in first ConvLSTM layer but if I set return_sequence = False program won't run<\/p>\n\n<p>The second part in last layer of ConvLSTM when return_sequence=False the output dimension change from 5 to 4 with timestep dimension disappear, why this happened?<\/p>\n","answers":[{"AnswerId":"52925617","CreationDate":"2018-10-22T08:57:03.897","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>If return_sequence is true, it means the LSTM layer will return the full sequence of the output not only the final output. So the input of next layer is still time sequence, it also means the next layer MUST be RNN to handle time sequence.<\/p>\n\n<p>If return_sequence is false, it means the LSTM layer will ONLY return the final output, which is not a time sequence any more. So the dimension  will be reduced one. For your example, it will be changed from 5 to 4. And because the input for next layer is not time sequence anymore, the next layer MUST NOT be RNN any more. <\/p>\n"},{"AnswerId":"52924404","CreationDate":"2018-10-22T07:38:10.040","ParentId":null,"OwnerUserId":"9590820","Title":null,"Body":"<p>When you set the parameter <code>return_sequence = False<\/code>, only one vector is returned after that, which does not satisfy the input dimension requirements of the next layer. This causes the program not to run. When you select true, a sequence is returned so that your timestep dimension does not disappear.<\/p>\n"}]}
+{"QuestionId":52921884,"AnswerCount":0,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-10-22T03:03:34.267","AcceptedAnswerId":null,"OwnerUserId":4192366.0,"Title":"Tensorflow result inconsistent across batch size","Body":"<p>I was playing around with batch MLP evaluation. Since GPU memory size is limited, I have to use small batches (e.g. 1000x43) of a very large matrix (e.g. 1000000x43). I found the result of the computation inconsistent across batch size. Here is a minimal example:<\/p>\n\n<pre><code>with tf.Session():\n    x = tf.Variable(np.random.rand(43, 60000).astype('f2'))\n    w = tf.Variable(np.ones((58624, 43), 'f2'))\n    b = tf.Variable(np.ones((58624, 1), 'f2'))\n    v1 = w @ x + b\n    w = tf.Variable(np.ones((5376, 43), 'f2'))\n    b = tf.Variable(np.ones((5376, 1), 'f2'))\n    v2 = w @ x + b\n    tf.global_variables_initializer().run()\n    print(v1.eval(), v2.eval())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]\n [ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]\n [ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]\n ..., \n [ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]\n [ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]\n [ 21.375     18.578125  22.765625 ...,  23.828125  24.203125  22.0625  ]] [[ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]\n [ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]\n [ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]\n ..., \n [ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]\n [ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]\n [ 22.375     19.578125  23.765625 ...,  24.828125  25.203125  23.0625  ]]\n<\/code><\/pre>\n\n<p>Apparently the result of two batch sizes (58624 and 5376) is off by 1, while it should not happen since the first axis of the matrices <code>w<\/code> and <code>b<\/code> should be broadcast-able.<\/p>\n\n<p>I'm using <code>tensorflow-gpu 1.11.0<\/code> from <code>conda<\/code>, platform is <code>Ubuntu Server 18.04.01<\/code>, architecture <code>amd64<\/code>, graphics card <code>Nvidia Titan Xp<\/code>. Is there really a bug, or is my logic wrong somewhere?<\/p>\n\n<p>BTW:<\/p>\n\n<ul>\n<li><p>In this example the result seems to be not related with the value of batch size, but in my actual (much bigger) example, which utilizes regular and batch matrix multiplication (both <code>@<\/code>), broadcast-ed <code>tanh<\/code>, tensor addition (<code>+<\/code>), and <code>tf.linalg.norm<\/code> over last two axis to compute matrix norm in batch, the result seems to be proportional to batch size.<\/p><\/li>\n<li><p>If I remove the <code>+b<\/code> part from <code>v1<\/code> and <code>v2<\/code>, the result is equivalent. But if I replace the matrix multiplication with a random tensor, I can't reproduce the problem.<\/p><\/li>\n<\/ul>\n\n<p>The actual bigger example is over here:<\/p>\n\n<pre><code>x = tf.Variable(np.ones((43, 60000), 'f2'))\ny = tf.Variable(np.ones((8, 60000), 'f2'))\nw = tf.Variable(np.ones((916* 64, 43), 'f2'))\nb = tf.Variable(np.ones((916* 64, 1), 'f2'))\nv = tf.Variable(np.ones((916, 8, 64), 'f2'))\nc = tf.Variable(np.ones((916, 8, 1), 'f2'))\nr = tf.linalg.norm(tf.cast(v@tf.reshape(tf.tanh(w@x+b),(916,64,60000))+c,'float32'),axis=(1,2))#prevent overflow\nw = tf.Variable(np.ones((84* 64, 43), 'f2'))\nb = tf.Variable(np.ones((84* 64, 1), 'f2'))\nv = tf.Variable(np.ones((84, 8, 64), 'f2'))\nc = tf.Variable(np.ones((84, 8, 1), 'f2'))\nR = tf.linalg.norm(tf.cast(v@tf.reshape(tf.tanh(w@x+b),(84,64,60000))+c,'float32'),axis=(1,2))#prevent overflow\ntf.global_variables_initializer().run()\nprint(r.eval()[::100], R.eval()[::10])#summary\n<\/code><\/pre>\n\n<p>which produces:<\/p>\n\n<pre><code>[ 1906641.625  1906641.625  1906641.625  1906641.625  1906641.625\n  1906641.625  1906641.625  1906641.625  1906641.625  1906641.625]\n[ 45033.32421875  45033.32421875  45033.32421875  45033.32421875\n  45033.32421875  45033.32421875  45033.32421875  45033.32421875\n  45033.32421875]\n<\/code><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":52922401,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-22T04:23:36.667","AcceptedAnswerId":52925263.0,"OwnerUserId":2530674.0,"Title":"pytorch modify array with list of indices","Body":"<p>Suppose I have a list of indices and wish to modify an existing array with this list. Currently the only way I can do this is by using a for loop as follows. Just wondering if there is a faster\/ efficient way.<\/p>\n\n<pre><code>torch.manual_seed(0)\na = torch.randn(5,3)\nidx = torch.Tensor([[1,2], [3,2]], dtype=torch.long)\nfor i,j in idx:\n    a[i,j] = 1\n<\/code><\/pre>\n\n<p>I initially assumed that <code>gather<\/code> or <code>index_select<\/code> would go some way in answering this question, but looking at <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.gather\" rel=\"nofollow noreferrer\">documentation<\/a> this doesn't seem to be the answer.<\/p>\n\n<p>In my particular case, a is a 5 dimensional vector and idx is a Nx5 vector. So the output (after subscripting with something like <code>a[idx]<\/code>) I'd expect is a <code>(N,)<\/code> shaped vector.<\/p>\n\n<h1>Answer<\/h1>\n\n<p>Thanks to @shai below, the answer that I was seeking was:\n<code>a[idx.t().chunk(chunks=2,dim=0)]<\/code>. Taken from this <a href=\"https:\/\/stackoverflow.com\/a\/52092603\/2530674\">SO answer<\/a>.<\/p>\n","answers":[{"AnswerId":"52925263","CreationDate":"2018-10-22T08:36:16.840","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It's quite simple<\/p>\n\n<pre><code>a[idx[:,0], idx[:,1]] = 1\n<\/code><\/pre>\n\n<p>You can find a more general solution in <a href=\"https:\/\/stackoverflow.com\/q\/52092230\/1714410\">this thread<\/a>.<\/p>\n"}]}
+{"QuestionId":52922445,"AnswerCount":3,"Tags":"<pytorch><tensor><attention-model>","CreationDate":"2018-10-22T04:32:00.977","AcceptedAnswerId":52923871.0,"OwnerUserId":5833067.0,"Title":"RuntimeError: \"exp\" not implemented for 'torch.LongTensor'","Body":"<p>I am following this tutorial: <a href=\"http:\/\/nlp.seas.harvard.edu\/2018\/04\/03\/attention.html\" rel=\"noreferrer\">http:\/\/nlp.seas.harvard.edu\/2018\/04\/03\/attention.html<\/a>\nto implement the Transformer model from the \"Attention Is All You Need\" paper. <\/p>\n\n<p>However I am getting the following error : \nRuntimeError: \"exp\" not implemented for 'torch.LongTensor'<\/p>\n\n<p>This is the line, in the PositionalEnconding class, that is causing the error:<\/p>\n\n<pre><code>div_term = torch.exp(torch.arange(0, d_model, 2) * -(math.log(10000.0) \/ d_model))\n<\/code><\/pre>\n\n<p>When it is being constructed here: <\/p>\n\n<pre><code>pe = PositionalEncoding(20, 0)\n<\/code><\/pre>\n\n<p>Any ideas?? I've already tried converting this to perhaps a Tensor Float type, but this has not worked.<\/p>\n\n<p>I've even downloaded the whole notebook with accompanying files and the error seems to persist in the original tutorial.<\/p>\n\n<p>Any ideas what may be causing this error? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52923871","CreationDate":"2018-10-22T07:02:10.420","ParentId":null,"OwnerUserId":"8498686","Title":null,"Body":"<p>I happened to follow this tutorial too. <\/p>\n\n<p>For me I just got the <code>torch.arange<\/code> to generate float type tensor<\/p>\n\n<p>from <\/p>\n\n<pre><code>position = torch.arange(0, max_len).unsqueeze(1)\ndiv_term = torch.exp(torch.arange(0, d_model, 2) * -(math.log(10000.0) \/ d_model))\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>position = torch.arange(0., max_len).unsqueeze(1)\ndiv_term = torch.exp(torch.arange(0., d_model, 2) * -(math.log(10000.0) \/ d_model))\n<\/code><\/pre>\n\n<p>Just a simple fix. But now it works for me. It is possible that the torch <code>exp<\/code> and <code>sin<\/code> previously support <code>LongTensor<\/code> but not anymore (not very sure about it).<\/p>\n"},{"AnswerId":"54042795","CreationDate":"2019-01-04T16:31:48.233","ParentId":null,"OwnerUserId":"10868475","Title":null,"Body":"<p>The suggestion given by @shai worked for me.  I modified the <code>init<\/code> method of <code>PositionalEncoding<\/code> by using <code>0.0<\/code> in two spots:<\/p>\n\n<pre><code>position = torch.arange(0.0, max_len).unsqueeze(1)\ndiv_term = torch.exp(torch.arange(0.0, d_model, 2) * -(math.log(10000.0) \/ d_model))\n<\/code><\/pre>\n"},{"AnswerId":"52922568","CreationDate":"2018-10-22T04:47:47.057","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like <code>torch.arange<\/code> returns a <code>LongTensor<\/code>, try <code>torch.arange(0.0, d_model, 2)<\/code> to force torch to return a <code>FloatTensor<\/code> instead.<\/p>\n"}]}
+{"QuestionId":52922746,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp><lstm>","CreationDate":"2018-10-22T05:10:45.837","AcceptedAnswerId":52946975.0,"OwnerUserId":9633432.0,"Title":"Multiplying custom trainable parameter to hidden units","Body":"<p>I want to take square of the difference between the tensors\/output after the LSTM layer and multiply it with the trainable parameter.<\/p>\n\n<p>As pointed out by @rvinas, I tried to write my own layer for the purpose,<\/p>\n\n<pre><code>class MyLayer(Layer):\n    def __init__(self,W_regularizer=None,W_constraint=None, **kwargs):\n        self.init = initializers.get('glorot_uniform')\n        self.W_regularizer = regularizers.get(W_regularizer)\n        self.W_constraint = constraints.get(W_constraint)\n        super(MyLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        assert len(input_shape) == 3\n        # Create a trainable weight variable for this layer.\n        self.W = self.add_weight((input_shape[-1],),\n                                 initializer=self.init,\n                                 name='{}_W'.format(self.name),\n                                 regularizer=self.W_regularizer,\n                                 constraint=self.W_constraint,\n                                trainable=True)\n        super(MyLayer, self).build(input_shape)  \n<\/code><\/pre>\n\n<p>The call function is multiplying only the tensors and Weight which I have initialized. Still I need to find how to take the pairwise difference and square them.<\/p>\n\n<pre><code>    def call(self, x):\n        uit = K.dot(x, self.W)\n        return uit\n\n    def compute_output_shape(self, input_shape):\n        return input_shape[0], input_shape[-1]\n<\/code><\/pre>\n\n<p>But then I am getting <code>AssertionError<\/code>at <code>assert len(input_shape) &gt;= 3<\/code>.<\/p>\n\n<p>I want to perform: <\/p>\n\n<pre><code>from keras.layers import Input, Lambda, LSTM\nfrom keras.models import Model\nimport keras.backend as K\nfrom keras.layers import Lambda\n\nlstm=LSTM(128, return_sequences=True)(input)\nsomething=MyLayer()(lstm)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52946975","CreationDate":"2018-10-23T10:39:41.040","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Please correct me if I am wrong, but I believe this is what you intend to do:<\/p>\n\n<pre><code>def call(self, x):\n    diff = K.square(x[:, 1:, :] - x[:, :-1, :])  # Shape=(batch_size, nb_timesteps-1, hidden_dim)\n    uit = K.dot(diff, self.W[:, None])  # Shape=(batch_size, nb_timesteps-1, 1)\n    uit = K.squeeze(uit, axis=-1)  # Shape=(batch_size, nb_timesteps-1)\n    return uit\n<\/code><\/pre>\n\n<p><strong>NOTE<\/strong>: Not tested.<\/p>\n"}]}
+{"QuestionId":52922816,"AnswerCount":1,"Tags":"<python><keras><typeerror>","CreationDate":"2018-10-22T05:18:26.967","AcceptedAnswerId":null,"OwnerUserId":7374610.0,"Title":"TypeError when loading in Keras model","Body":"<p>I've defined a model using the MobileNet architecture from keras using the code below:<\/p>\n\n<pre><code>model = MobileNet(input_shape=(size, size, 1), alpha=1.0, weights=None, classes=NCATS)\nmodel.compile(optimizer=Adam(lr=0.002), loss='categorical_crossentropy',\n              metrics=[categorical_crossentropy, categorical_accuracy, top_3_accuracy])\n<\/code><\/pre>\n\n<p>After training, I saved the model the using<\/p>\n\n<pre><code>model.save(MOD1_PATH)\n<\/code><\/pre>\n\n<p>I am now trying to load in the model using<\/p>\n\n<pre><code>model = load_model(MOD1_PATH)\n<\/code><\/pre>\n\n<p>but I am getting the following TypeError when trying to load it:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-26-2166dee60e7c&gt; in &lt;module&gt;\n----&gt; 1 load_model(MOD1_PATH)\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/saving.py in load_model(filepath, custom_objects, compile)\n    418     f = H5Dict(filepath, 'r')\n    419     try:\n--&gt; 420         model = _deserialize_model(f, custom_objects, compile)\n    421     finally:\n    422         if opened_new_file:\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/saving.py in _deserialize_model(f, custom_objects, compile)\n    224         raise ValueError('No model found in config.')\n    225     model_config = json.loads(model_config.decode('utf-8'))\n--&gt; 226     model = model_from_config(model_config, custom_objects=custom_objects)\n    227     model_weights_group = f['model_weights']\n    228 \n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/saving.py in model_from_config(config, custom_objects)\n    457                         '`Sequential.from_config(config)`?')\n    458     from ..layers import deserialize\n--&gt; 459     return deserialize(config, custom_objects=custom_objects)\n    460 \n    461 \n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/layers\/__init__.py in deserialize(config, custom_objects)\n     53                                     module_objects=globs,\n     54                                     custom_objects=custom_objects,\n---&gt; 55                                     printable_module_name='layer')\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/utils\/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    143                     config['config'],\n    144                     custom_objects=dict(list(_GLOBAL_CUSTOM_OBJECTS.items()) +\n--&gt; 145                                         list(custom_objects.items())))\n    146             with CustomObjectScope(custom_objects):\n    147                 return cls.from_config(config['config'])\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/network.py in from_config(cls, config, custom_objects)\n   1020         # First, we create all layers and enqueue nodes to be processed\n   1021         for layer_data in config['layers']:\n-&gt; 1022             process_layer(layer_data)\n   1023         # Then we process nodes in order of layer depth.\n   1024         # Nodes that cannot yet be processed (if the inbound node\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/network.py in process_layer(layer_data)\n   1006 \n   1007             layer = deserialize_layer(layer_data,\n-&gt; 1008                                       custom_objects=custom_objects)\n   1009             created_layers[layer_name] = layer\n   1010 \n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/layers\/__init__.py in deserialize(config, custom_objects)\n     53                                     module_objects=globs,\n     54                                     custom_objects=custom_objects,\n---&gt; 55                                     printable_module_name='layer')\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/utils\/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    145                                         list(custom_objects.items())))\n    146             with CustomObjectScope(custom_objects):\n--&gt; 147                 return cls.from_config(config['config'])\n    148         else:\n    149             # Then `cls` may be a function returning a class.\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/engine\/base_layer.py in from_config(cls, config)\n   1107             A layer instance.\n   1108         \"\"\"\n-&gt; 1109         return cls(**config)\n   1110 \n   1111     def count_params(self):\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/Keras-2.2.4-py3.6.egg\/keras\/layers\/advanced_activations.py in __init__(self, max_value, negative_slope, threshold, **kwargs)\n    290                  threshold=0., **kwargs):\n    291         super(ReLU, self).__init__(**kwargs)\n--&gt; 292         if max_value is not None and max_value &lt; 0.:\n    293             raise ValueError('max_value of ReLU layer '\n    294                              'cannot be negative value: %s' % str(max_value))\n\nTypeError: '&lt;' not supported between instances of 'dict' and 'float'\n<\/code><\/pre>\n\n<p>Has anyone run into this kind of problem loading a trained model before? Any help would be appreciated. Thanks.<\/p>\n","answers":[{"AnswerId":"52928512","CreationDate":"2018-10-22T11:41:51.130","ParentId":null,"OwnerUserId":"10540199","Title":null,"Body":"<p>I had the same issue with:<\/p>\n\n<p><em>Tensorflow<\/em> <strong>1.11.0<\/strong>,<\/p>\n\n<p><em>Keras<\/em> <strong>2.1.6-tf<\/strong><\/p>\n\n<p>I created a notebook in Google Colaboratory and the result was the same.\nGoogle colaboratory configuration: <\/p>\n\n<p><em>Tensorflow<\/em> <strong>1.12.0-rc1<\/strong>,<\/p>\n\n<p><em>Keras<\/em> <strong>2.1.6-tf<\/strong><\/p>\n\n<p>Then I tryied to <code>load_model<\/code> with:<\/p>\n\n<p><em>Tensorflow<\/em> <strong>1.9.0<\/strong>,<\/p>\n\n<p><em>Keras<\/em> <strong>2.1.6-tf<\/strong><\/p>\n\n<p>and everything began to work.\nMay be this is a bug with new versions of <em>Tensorflow<\/em><\/p>\n"}]}
+{"QuestionId":52922980,"AnswerCount":1,"Tags":"<python><tensorflow><quantization>","CreationDate":"2018-10-22T05:37:54.760","AcceptedAnswerId":null,"OwnerUserId":9104473.0,"Title":"Error when using Quantized Conv2D with tf.qint8 inputs","Body":"<p>I am new to quantization and I was learning usage of QuantizedConv2D operation in tensorflow. The code is as below<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.constant([1.5,1.2],shape = [1,1,1,2], name ='a')\na1 = tf.constant([[2.0],[1.0]],shape = [1,1,2,1], name ='b')\nmin_a= tf.reduce_min(a)\nmax_a= tf.reduce_max(a)\nx = tf.quantize(a,min_a,max_a,tf.qint8)\nmin_a1= tf.reduce_min(a1)\nmax_a1= tf.reduce_max(a1)\nx1 = tf.quantize(a1,min_a1,max_a1,tf.qint8)\ny1 = t  f.nn.quantized_conv2d(x.output,x1.output,x.output_min,x.output_max,x1.output_min,x1.output_max,[1,1,1,1],\"SAME\")\nwith tf.Session() as sess:\n  print (sess.run(y1))\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): No OpKernel was\n  registered to support Op 'QuantizedConv2D' with these attrs. \n  Registered devices: [CPU], Registered kernels:   device='CPU'; Tinput\n  in [DT_QUINT8]; Tfilter in [DT_QUINT8]; out_type in [DT_QINT32]\n           [[Node: QuantizedConv2D = QuantizedConv2D[Tfilter=DT_QINT8, Tinput=DT_QINT8, dilations=[1, 1, 1, 1], out_type=DT_QINT32,\n  padding=\"SAME\", strides=[1, 1, 1, 1]](QuantizeV2, QuantizeV2_1,\n  QuantizeV2:1, QuantizeV2:2, QuantizeV2_1:1, QuantizeV2_1:2)]]<\/p>\n<\/blockquote>\n\n<p>I am not sure if my code is right or this is a bug.<\/p>\n\n<p>Any help will be appreciated.<\/p>\n\n<p>Thanks and Regards,<\/p>\n\n<p>Abhinav<\/p>\n","answers":[{"AnswerId":"52924864","CreationDate":"2018-10-22T08:10:59.717","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>I believe the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/quantized_conv2d\" rel=\"nofollow noreferrer\">tf.nn.quantized_conv2d<\/a> is confusing, because this operation is only registered for the input datatype <code>tf.quint8<\/code>. You could do the following:<\/p>\n\n<pre><code>x = tf.quantize(a, min_a, max_a, tf.quint8)\nx1 = tf.quantize(a1, min_a1, max_a1, tf.quint8)\n<\/code><\/pre>\n\n<p>I opened an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23157\" rel=\"nofollow noreferrer\">issue<\/a> on GitHub.<\/p>\n"}]}
+{"QuestionId":52923062,"AnswerCount":1,"Tags":"<python><tensorflow><convolution>","CreationDate":"2018-10-22T05:49:08.247","AcceptedAnswerId":52924158.0,"OwnerUserId":3715778.0,"Title":"tensorflow compute tf.nn.conv2d","Body":"<p>I have calculated a convolution between a 3x3 image and two 2x2 filters in Excel manually:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tKcQW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tKcQW.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>I want to reproduce the same result using tensorflow <code>tf.nn.conv2d:<\/code><\/p>\n\n<pre><code>x_raw = np.array([\n    [2,5,3],\n    [3,4,2],\n    [4,1,1]\n])\n\nf_raw = np.array(\n[[\n    [2,1],\n    [3,4]\n],[\n    [4,1],\n    [1,2]   \n]])\n\nf = tf.constant(f_raw, dtype=tf.float32)\nx = tf.constant(x_raw, dtype=tf.float32)\n\nfilter = tf.reshape(f, [2, 2, 1, 2])\nimage  = tf.reshape(x, [1, 3, 3, 1])\n\ntf.nn.conv2d(image, filter, [1, 1, 1, 1], \"VALID\").eval()\n<\/code><\/pre>\n\n<p>But the output I have from tensorflow is off:<\/p>\n\n<blockquote>\n  <p>array([[[[35.,33.],[37.,25.]],[[35.,25.],[19.,15.]]]], dtype=float32)<\/p>\n<\/blockquote>\n\n<p>What I am doing wrong ?<\/p>\n","answers":[{"AnswerId":"52924158","CreationDate":"2018-10-22T07:20:48.590","ParentId":null,"OwnerUserId":"6278661","Title":null,"Body":"<p>To get identical results as in your excel example, you need to make the following changes: <\/p>\n\n<ol>\n<li>create two seperate weights<\/li>\n<li>calculate convolutions for each weight separately <\/li>\n<\/ol>\n\n<p>Code example: <\/p>\n\n<pre><code>x_raw = np.array([\n    [2,5,3],\n    [3,4,2],\n    [4,1,1]\n])\n#created two seperate weights \nweight1 = np.array(\n[[\n    [2,1],\n    [3,4]\n]])\n\nweight2 = np.array(\n[[\n    [4,1],\n    [1,2]\n]]\n)\nweight1 = tf.constant(weight1, dtype=tf.float32)\nweight2 = tf.constant(weight2, dtype=tf.float32)\nx = tf.constant(x_raw, dtype=tf.float32)\n\n#change out_channels to 1 \nfilter1 = tf.reshape(weight1, [2, 2, 1, 1])\nfilter2 = tf.reshape(weight2, [2, 2, 1, 1])\nimage = tf.reshape(x, [1, 3, 3, 1])\n\nwith tf.Session() as sess:\n  print(tf.nn.conv2d(image, filter1, [1, 1, 1, 1], \"VALID\").eval())\n  print(tf.nn.conv2d(image, filter2, [1, 1, 1, 1], \"VALID\").eval())\n<\/code><\/pre>\n"}]}
+{"QuestionId":52923348,"AnswerCount":1,"Tags":"<python><tensorflow><nlp><data-science>","CreationDate":"2018-10-22T06:18:57.000","AcceptedAnswerId":52947673.0,"OwnerUserId":7625035.0,"Title":"How to get entities with direction in relation extraction?","Body":"<p>I have been working with relation extraction for a week. But what I need is direction between two entities, such as Company_x got bought by Company_y. So the model should predict the entities like Company_y->bought-> Company_X. Any models you guys think will be helpful for this?<\/p>\n","answers":[{"AnswerId":"52947673","CreationDate":"2018-10-23T11:16:33.367","ParentId":null,"OwnerUserId":"4324153","Title":null,"Body":"<p>The passive voice is usually a good indicator of the direction of the relationship.<\/p>\n\n<p>You can extract, from the context between the two entities, a pattern which starts by a verb, and then detecting the presence or absence of the passive voice.<\/p>\n\n<p>Some simple proof-of-concept code (this could actually be much more simpler using RegexpParser from NLTK)<\/p>\n\n<pre><code>from nltk import pos_tag\nfrom nltk import word_tokenize\nfrom nltk.stem.wordnet import WordNetLemmatizer\n\nlmtzr = WordNetLemmatizer()\naux_verbs = ['be']\n\ndef detect_passive_voice(pattern):\n    passive_voice = False\n\n    if len(pattern) &gt;= 3:\n        if pattern[0][1].startswith('V'):\n            verb = lmtzr.lemmatize(pattern[0][0], 'v')\n            if verb in aux_verbs:\n                if (pattern[1][1] == 'VBN' or pattern[1][1] == 'VBD') and pattern[-1][0] == 'by':\n                    passive_voice = True\n\n            # past verb + by\n            elif (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':\n                passive_voice = True\n\n        # past verb + by\n        elif (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':\n                passive_voice = True\n\n    # past verb + by\n    elif len(pattern) &gt;= 2:\n        if (pattern[-2][1] == 'VBN' or pattern[-2][1] == 'VBD') and pattern[-1][0] == 'by':\n            passive_voice = True\n\nreturn passive_voice\n<\/code><\/pre>\n\n<p>Running some examples:<\/p>\n\n<pre><code>In [4]: tokens = word_tokenize(\"was bought by\")\n   ...: tags = pos_tag(tokens)\n   ...: detect_passive_voice(tags)\nOut[4]: True\n\nIn [5]: tokens = word_tokenize(\"mailed the letter\")\n   ...: tags = pos_tag(tokens)\n   ...: detect_passive_voice(tags)\nOut[5]: False\n\nIn [7]: tokens = word_tokenize(\"was mailed by\")\n   ...: tags = pos_tag(tokens)\n   ...: detect_passive_voice(tags)\nOut[7]: True\n<\/code><\/pre>\n\n<p>You can add more auxiliary verbs and also allow for the existence of adverbs or adjectives in-between.<\/p>\n"}]}
+{"QuestionId":52923486,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-22T06:30:44.867","AcceptedAnswerId":null,"OwnerUserId":260826.0,"Title":"Tensorflow\/Keras LinearRegression Warning Error","Body":"<p>I'm trying to convert this model to a Keras Linear regressor.\nCurrent <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_regression\" rel=\"nofollow noreferrer\">code<\/a>:<\/p>\n\n<pre><code>run_config = tf.estimator.RunConfig(save_checkpoints_steps=500)\n\ndef build_model():\n  model = keras.Sequential([\n    keras.layers.Dense(64, activation=tf.nn.relu,\n                       input_shape=(train_data.shape[1],)),\n    keras.layers.Dense(64, activation=tf.nn.relu),\n    keras.layers.Dense(1)\n  ])\n  optimizer = tf.train.RMSPropOptimizer(0.001)\n  model.compile(loss='mse',\n                optimizer=optimizer,\n                metrics=['mae'])\n  return model        \n<\/code><\/pre>\n\n<p>I convert it to:<\/p>\n\n<pre><code>  model = models.Sequential()\n  model.add(Dense(1, activation='linear', input_shape=(num_features,)))  \n  optimizer = keras.optimizers.SGD(lr=learning_rate)\n  model.compile(optimizer=optimizer, loss='mse', metrics=['mae'])\n  return tf.keras.estimator.model_to_estimator(\n      keras_model=model, model_dir=model_dir, config=config)\n<\/code><\/pre>\n\n<p>There is an infinite loop showing this:<\/p>\n\n<pre><code>WARNING:tensorflow:It seems that global step (tf.train.get_global_step) has not been increased. Current value (could be stable): 0 vs previous value: 0. You could increase the global step by passing tf.train.get_global_step() to Optimizer.apply_gradients or Optimizer.minimize.\n<\/code><\/pre>\n\n<p>My learning rate is set to .001: (Tried with .01 and .1) same results.<\/p>\n","answers":[]}
+{"QuestionId":52923585,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2018-10-22T06:40:32.667","AcceptedAnswerId":52932013.0,"OwnerUserId":10538939.0,"Title":"using a trained cnn for a test data","Body":"<p>I'm training a CNN in Tensorflow and I save it with the code: <\/p>\n\n<pre><code>saver = tf.train.Saver() \n<\/code><\/pre>\n\n<p>end of each Epoch I save it using this, \n<code>saver.save(sess,Save_Dir)<\/code>. <\/p>\n\n<p>After my CNN has been trained, I want to use it on a test data. I wonder, How I can restore my Trained CNN and How I can see How my trained CNN works on test data that I have. I've searched a lot in this regard, but I didn't find something that I can Learn about the process. Moreover, I do not have any sense about the 4 types of files that the saver generate. (.meta,....).\nThank You. <\/p>\n","answers":[{"AnswerId":"52932013","CreationDate":"2018-10-22T14:43:36.930","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<blockquote>\n  <p>restore my Trained CNN<\/p>\n<\/blockquote>\n\n<p>To load your model, you do <code>saver.restore(session, load_path)<\/code> which should load the graph back into the session.<\/p>\n\n<blockquote>\n  <p>How I can see How my trained CNN works on test data that I have<\/p>\n<\/blockquote>\n\n<p>You can use <code>tf.metrics<\/code> or define your own custom metric<\/p>\n\n<blockquote>\n  <p>Moreover, I do not have any sense about the 4 types of files that the saver generate. (.meta,....). Thank You. <\/p>\n<\/blockquote>\n\n<p>I found <a href=\"https:\/\/jacobbuckman.com\/post\/tensorflow-the-confusing-parts-2\/\" rel=\"nofollow noreferrer\">Tensorflow: The confusing Parts 2<\/a> to be extremely useful.<\/p>\n"}]}
+{"QuestionId":52924644,"AnswerCount":1,"Tags":"<python><keras><deep-learning><keras-layer>","CreationDate":"2018-10-22T07:55:54.337","AcceptedAnswerId":52926643.0,"OwnerUserId":10032132.0,"Title":"Dealing with inserting noise to train data in Keras (Deep learning)","Body":"<p>I am using Keras for Deep learning.<\/p>\n\n<p>I want to put noise into train data at each epoch during training.<\/p>\n\n<p>So, at every epoch, the train data should be different from before epoch, because of random noise insertion.<\/p>\n\n<p>This is my code: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(GaussianNoise(SNR_std))\nmodel.add(Dense(neuron,input_dim=1920,\n                kernel_initializer=initializers.he_normal(seed=seed_num),\n                use_bias=False)\nmodel.add(BatchNormalization())\nmodel.add(Activation('relu'))\n<\/code><\/pre>\n\n<p>Did I do it in a right way for my intention?<\/p>\n","answers":[{"AnswerId":"52926643","CreationDate":"2018-10-22T09:57:10.340","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>It seems correct to me. <\/p>\n\n<p>One thing to note is that if you change the images using noise like this you should visualize those images at least once before starting your training, so that you actually know what you are learning. So getting a handle to the output of that layer is key. Answers on how to do this can be found all over the internet: (<a href=\"https:\/\/datascience.stackexchange.com\/questions\/20469\/keras-visualizing-the-output-of-an-intermediate-layer\">https:\/\/datascience.stackexchange.com\/questions\/20469\/keras-visualizing-the-output-of-an-intermediate-layer<\/a>)<\/p>\n"}]}
+{"QuestionId":52924669,"AnswerCount":1,"Tags":"<python><pandas><tensorflow><keras>","CreationDate":"2018-10-22T07:57:21.253","AcceptedAnswerId":null,"OwnerUserId":8853214.0,"Title":"ValueError in DL","Body":"<p>I get this error , post choosing the epochs: ValueError: Input arrays should have the same number of samples as target arrays. Found 5516 input samples and 12870 target samples. Any suggestions are welcome. Thanks in advance<\/p>\n\n<p>Im using a dataset with a lot of categorical variables and they add up to 95 after creating the dummy variables, the code until I choose the number of epochs runs flawlessly and then I get this error, what is the reason for this error, its important, one, I could use it in the future and 2, Im unable to proceed :)<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport pandas as pd\n\ndataset = pd.read_csv('zrpl_data.csv')\nX = dataset.iloc[:, 0:6].values\ny = dataset.iloc[:, 6].values\n\nfrom sklearn.preprocessing import LabelEncoder, OneHotEncoder\nlabelencoder_X_1 = LabelEncoder()\nX[:, 0] = labelencoder_X_1.fit_transform(X[:, 0])\n\nlabelencoder_X_2 = LabelEncoder()\nX[:, 1] = labelencoder_X_2.fit_transform(X[:, 1])\nlabelencoder_X_3 = LabelEncoder()\nX[:, 2] = labelencoder_X_1.fit_transform(X[:, 2])\n\nlabelencoder_X_4 = LabelEncoder()\nX[:, 3] = labelencoder_X_1.fit_transform(X[:, 3])\n\nlabelencoder_X_5 = LabelEncoder()\nX[:, 4] = labelencoder_X_1.fit_transform(X[:, 4])\n\nonehotencoder = OneHotEncoder(categorical_features = [0,1,2,3,4])\nX = onehotencoder.fit_transform(X).toarray()\nX = X[:, 1:]\n\nfrom sklearn.model_selection import train_test_split\nX_train, X_test, y_train, y_test = train_test_split(X,y,test_size=0.3, \nrandom_state = 0)\n\nfrom sklearn.preprocessing import StandardScaler\nsc = StandardScaler()\nX_train = sc.fit_transform(X_train)\nX_test = sc.transform(X_test)\n\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\nclassifier = Sequential()\n\nclassifier.add(Dense(output_dim=47, \ninit='uniform',activation='relu',input_dim=95)) \n\nclassifier.add(Dense(output_dim=47, init='uniform',activation='relu'))\nclassifier.add(Dense(output_dim=1, \ninit='uniform',activation='sigmoid'))\n\nclassifier.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n\nclassifier.fit (X_train, y_train, batch_size=10,nb_epoch=100)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52924963","CreationDate":"2018-10-22T08:16:21.943","ParentId":null,"OwnerUserId":"10026409","Title":null,"Body":"<p>You have 5516 feature samples and 12870 target samples (you should have equal), before train the model double check their dimensions.<\/p>\n"}]}
+{"QuestionId":52925003,"AnswerCount":0,"Tags":"<python><tensorflow><keras><keras-layer>","CreationDate":"2018-10-22T08:19:32.313","AcceptedAnswerId":null,"OwnerUserId":6495871.0,"Title":"Custom Layer Keras: transform output at certain positions","Body":"<p>I am trying to write a custom layer in Keras (with tensorflow backend) that makes certain positions binary. <\/p>\n\n<p>For example suppose that I have [0.6,0.8,0.9,0.2] and that position 1 and 3 must be binary, I would like to have a layer that outputs [0.6,1,0.9,0]<\/p>\n\n<p>e.g. output[pos] > 0.5 then output[pos] = 1, else output[pos] = 0<\/p>\n\n<p>I wrote this but it is not working at all...<\/p>\n\n<pre><code>...Layers of the net...\nx = Lambda(self.adjust_positions)(x)\n<\/code><\/pre>\n\n<p>here the functions I wrote<\/p>\n\n<pre><code>    def update_1(self, x, pos):\n        with tf.control_dependencies([tf.assign(x[pos],[1])]):\n                return tf.identity(x)\n\n    def update_0(self, x, pos):\n        with tf.control_dependencies([tf.assign(x[pos],[0])]):\n                return tf.identity(x)\n\n    def adjust_positions(self, x):\n        for pos in indexes:\n            tf.cond(tf.gather(x, pos)&lt;[0.5], self.update_0(x, pos), self.update_1(x,pos))\n        return x\n<\/code><\/pre>\n\n<p>The error I get is:<\/p>\n\n<pre><code>ValueError: Sliced assignment is only supported for variables\n     55     def update_0(self, x, pos):\n---&gt; 56         with tf.control_dependencies([tf.assign(x[pos],[0])]):\n     57                 return tf.identity(x)\n<\/code><\/pre>\n\n<p>How can I implement this functionality? What I have done is reasonable?<\/p>\n","answers":[]}
+{"QuestionId":52925048,"AnswerCount":1,"Tags":"<python><tensorflow><rnn>","CreationDate":"2018-10-22T08:22:22.883","AcceptedAnswerId":null,"OwnerUserId":3474338.0,"Title":"Extremely Slow Building Bidirection TF Lstm Cell","Body":"<p>I am building a bidirectional lstm layer by Tensorflow, but the building process is very slow, it takes likes hours to finish building for a fairly small network. \nmy test hidden layer size is 16; input embedding vector size is 32; input sequence length is 20, vocabulary size is 20000<\/p>\n\n<p>is the performance of tf as expected?<\/p>\n\n<p>attached the part of I used to build the network for more information:<\/p>\n\n<pre><code>cell_fw = tf.contrib.rnn.LSTMCell(hidden_layer_size)\ncell_bw = tf.contrib.rnn.LSTMCell(hidden_layer_size)\n(emb_encoder_inputs, fw_state, _) = tf.contrib.rnn.static_bidirectional_rnn(\n          cell_fw, cell_bw, emb_encoder_inputs, dtype=tf.float32,\n          sequence_length=article_lens)\n<\/code><\/pre>\n","answers":[{"AnswerId":"55397382","CreationDate":"2019-03-28T12:14:19.033","ParentId":null,"OwnerUserId":"5101045","Title":null,"Body":"<p>bidirectional_dynamic_rnn is faster than static_bidirectional_rnn.<\/p>\n"}]}
+{"QuestionId":52925767,"AnswerCount":1,"Tags":"<python><tensorflow><dataset>","CreationDate":"2018-10-22T09:06:44.360","AcceptedAnswerId":52931411.0,"OwnerUserId":4186749.0,"Title":"Cannot iterate over tf.data.Dataset","Body":"<p>I'm using the tensorflow.data.Dataset api by tensorflow. However I need to create datasets on the fly filtering out elements other dataset. While training goes well and I can iterate over the training set and the dev set, when I reinitialize the iterator with a new dataset that I just created with a filter, i receive the following exception: <\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.NotFoundError: Function tf_predicate_5HKZIzWZBv8 is not defined.\n<\/code><\/pre>\n\n<p>I'm using the following function to create an initialiser out of a dataset:<\/p>\n\n<pre><code>self.iterator.make_initializer(dataset)\n<\/code><\/pre>\n\n<p>where self.iterator is defined as follow:<\/p>\n\n<pre><code>self.iterator = tf.data.Iterator.from_structure(ds_types, ds_shapes)\n<\/code><\/pre>\n\n<p>Do you guys have any idea about why this is happening? Note that it happens if I call make_initializer after I have created a session, run a dataset, and then create a new initializer. If after the creation I also recreate the Session everything works (except the fact that all the variables have to be reinitialized)<\/p>\n","answers":[{"AnswerId":"52931411","CreationDate":"2018-10-22T14:11:16.287","ParentId":null,"OwnerUserId":"4186749","Title":null,"Body":"<p>I found the solution and I'm sharing in case somebody will run into this problem. The thing is that, as I'm defining a new dataset after the session has been initialised it, it doesn't have the new operation I'm adding for the new dataset (In this case I'm using a new filter everytime I create a new dataset) and that's why the session can't find the operation. To overcome the problem I defined all the datasets I needed to use before the session is initialised and I used a filter that takes as input a placeholder so that I always use the same filter feeded everytime at iterator init time with the right value.<\/p>\n"}]}
+{"QuestionId":52925908,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-22T09:14:56.007","AcceptedAnswerId":52946070.0,"OwnerUserId":10539582.0,"Title":"how to use If statement in tensorflow?","Body":"<p>So I need to calculate one tensor that kind of needs an 'if' statement. Specifically, I need to calculate a tensor and compare it with a constant, then implement different functions. <\/p>\n\n<p>Here is how I did it(failed):<\/p>\n\n<hr>\n\n<pre><code>num_instances = tf.size(unique_labels)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>l_var=tf.cond(tf.less(num_instances,1),true_fn=return_zero(),false_fn=fn())\n<\/code><\/pre>\n\n<hr>\n\n<p><em>num_instances<\/em> the tensor is changing its value as diffrent data is fed into the model during training<\/p>\n\n<p>Sadly, it didn't work,and the terminal reports<\/p>\n\n<blockquote>\n  <p><strong>'TypeError:<\/strong> true_fn must be callable. '<\/p>\n<\/blockquote>\n\n<p>you can find my <strong>output<\/strong> here :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BoioV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BoioV.png\" alt=\"sample output\"><\/a><\/p>\n","answers":[{"AnswerId":"52946070","CreationDate":"2018-10-23T09:50:14.270","ParentId":null,"OwnerUserId":"10535783","Title":null,"Body":"<p>'if' statement example for:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python2\n# -*- coding: utf-8 -*-\n\"\"\"\n@author: myhaspl@myhaspl.com\n\"\"\"\n\nimport tensorflow as tf\nx = tf.constant(11)\ny = tf.constant(22)\n\n\nz = tf.multiply(x, y)\nresult1 = tf.cond(x &gt; y, lambda: tf.add(x, y), lambda: tf.square(y))\nresult2 = tf.cond(x &lt; y, lambda: tf.add(x, y), lambda: tf.square(y))\n\ninit_op = tf.global_variables_initializer()\nsess=tf.Session()\nwith sess: \n    sess.run(init_op)\n    print sess.run(result1)\n    print sess.run(result2)\n<\/code><\/pre>\n\n<p>result1=11*22<\/p>\n\n<p>result2=11+22<\/p>\n\n<p>running results:<\/p>\n\n<p>484<\/p>\n\n<p>33<\/p>\n"}]}
+{"QuestionId":52926240,"AnswerCount":1,"Tags":"<python><neural-network><keras><recurrent-neural-network>","CreationDate":"2018-10-22T09:34:33.763","AcceptedAnswerId":null,"OwnerUserId":10441470.0,"Title":"Issue with Keras when calling Dense","Body":"<p>Ive been trying to run this code below and it keeps throwing out an error.<\/p>\n\n<p>This is the code im trying to run on ipython. Can someone tell me what the issue is?<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nfrom util import getKaggleMNIST\nfrom keras.models import Model\nfrom keras.layers import Dense, Activation, Input\n\nxtrain, ytrain, xtest, ytest = getKaggleMNIST()\n\nN, D = xtrain.shape\nk = len(set(ytrain))\n\ni = Input(shape=(D,))\nx = Dense(500, activation='relu')(i)\nx = Dense(300, activation='relu')(x)\nx = Dense(k, activation='softmax')(k)\n\nmodel = Model(inputs=i, outputs=x)\n\nmodel.compile( loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'] )\n\nr = model.fit(Xtrain, Ytrain, validation_data=(Xtest, Ytest), epochs=15, batch_size=32)\n\nprint(\"Returned:\", r)\n\nplt.plot(r.history['loss'], label='loss') \nplt.plot(r.history['val_loss'], label='val_loss') \nplt.legend() \nplt.show()\n<\/code><\/pre>\n\n<p>This is the error:<\/p>\n\n<pre><code>ValueError: Layer dense_3 was called with an input that isn't a symbolic tensor. Received type: &lt;type 'int'&gt;. Full input: [10]. All inputs to the layer should be tensors.\n<\/code><\/pre>\n\n<p>Thanks is advance.<\/p>\n","answers":[{"AnswerId":"52926846","CreationDate":"2018-10-22T10:09:17.983","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>replace <code>x = Dense(k, activation='softmax')(k)<\/code> with <code>x = Dense(k, activation='softmax')(x)<\/code><\/p>\n"}]}
+{"QuestionId":52926847,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-22T10:09:30.313","AcceptedAnswerId":52926956.0,"OwnerUserId":7824576.0,"Title":"pytorch how to compute grad after clone a tensor","Body":"<p>My simple code:<\/p>\n\n<pre><code>import torch\n\nx = torch.randn(4, requires_grad=True).cuda()\ny = torch.randn(4, requires_grad=True).cuda()\nz = torch.zeros(4)\nz = torch.clone(x)\nz.retain_grad()\nh = (z + y) * z\nl = torch.randn(4).cuda()\nloss = (l - h).pow(2).sum()\nloss.backward()\nprint('x.grad=', x.grad)\nprint('z.grad=', z.grad)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>x.grad= None\nz.grad= tensor([-15.3401,  -3.2623,  -2.1670,   0.1410], device='cuda:0')\n<\/code><\/pre>\n\n<p>Why x.grad is None but not same as z.grad ?\nWhat should I do if I want they are same?<\/p>\n","answers":[{"AnswerId":"52926956","CreationDate":"2018-10-22T10:16:29.003","ParentId":null,"OwnerUserId":"7602551","Title":null,"Body":"<p>You need to call <code>x.retain_grad()<\/code> after declaring x if you want to keep the grad of tensor x.<\/p>\n"}]}
+{"QuestionId":52927499,"AnswerCount":1,"Tags":"<python><plot><keras>","CreationDate":"2018-10-22T10:46:36.953","AcceptedAnswerId":null,"OwnerUserId":5078831.0,"Title":"plot loss and accuracy of a model in keras","Body":"<p>i have a function to build a model in keras such as below:<\/p>\n\n<pre><code>def build_model(lr = 0.0):\n    inp = Input(shape = (max_len,))\n    x = Embedding_layer\n    y = LSTM_layer(x)\n    y = Convolution_layer(y)\n    x = GlobalMaxPooling1D(y)\n\n    x = Dense(3, activation = \"sigmoid\")(x)\n    model = Model(inputs = inp, outputs = x)\n    model.compile(loss = \"binary_crossentropy\", optimizer = Adam(), metrics = [\"accuracy\"])\n    history = model.fit(X_train, Y_train, batch_size = 256, epochs = 3, \n                        verbose = 1, callbacks = [ra_val, check_point, early_stop])\n    model = load_model(file_path)\n    return model\n\nmodel = build_model(lr = 1e-3)\n<\/code><\/pre>\n\n<p>and now i want to plot history loss and accuracy after training phase, but model doesn't have history option.<\/p>\n\n<p>how can i plot loss and accuracy ?<\/p>\n","answers":[{"AnswerId":"52928281","CreationDate":"2018-10-22T11:27:47.280","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>The <a href=\"https:\/\/keras.io\/models\/sequential\/#fit\" rel=\"nofollow noreferrer\">fit<\/a> function will return a <code>History<\/code> object that contains the loss values and the metrics of the training.<\/p>\n"}]}
+{"QuestionId":52928828,"AnswerCount":0,"Tags":"<python><opencv><tensorflow>","CreationDate":"2018-10-22T12:00:16.453","AcceptedAnswerId":null,"OwnerUserId":9600714.0,"Title":"\"Failed to load the native TensorFlow runtime\" and \"ImportError: DLL load failed: The specified module could not be found.\"","Body":"<p>I'm trying to run some instances of some ImageAI code and some separate OpenCV codes. The code that gives out the error uses is the following:<\/p>\n\n<pre><code>    from imageai.Detection import ObjectDetection\nimport os\n\nexecution_path = os.getcwd()\n\ndetector = ObjectDetection()\ndetector.setModelTypeAsYOLOv3()\ndetector.setModelPath(os.path.join(execution_path, \"yolo.h5\"))\ndetector.loadModel()\ndetections = detector.detectObjectsFromImage(input_image=os.path.join(execution_path, \"image2.jpg\"),\n                                             output_image_path=os.path.join(execution_path, \"image2new.jpg\"),\n                                             minimum_percentage_probability=30)\n\nfor eachObject in detections:\n    print(eachObject[\"name\"], \" : \", eachObject[\"percentage_probability\"], \" : \", eachObject[\"box_points\"])\nprint(\"--------------------------------\")\n<\/code><\/pre>\n\n<p>and the error it spits out is:<\/p>\n\n<pre><code>    Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"C:\/Users\/Okan\/Desktop\/face-detect\/ImageAI-LiveVideoDedection.py\", line 1, in &lt;module&gt;\n    from imageai.Detection import ObjectDetection\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\imageai\\Detection\\__init__.py\", line 3, in &lt;module&gt;\n    from imageai.Detection.keras_retinanet.models.resnet import resnet50_retinanet\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\imageai\\Detection\\keras_retinanet\\models\\resnet.py\", line 19, in &lt;module&gt;\n    import keras\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\backend\\__init__.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\Okan\\AppData\\Local\\Programs\\Python\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>The weird thing is, on my desktop the code worked perfectly for one or two times. Then, without changing anything, the same error started to show up just like my laptop, which never got to run the code properly. This made no sense to me so I am here asking for help. I would appreciate any kind of support over this. The \"solutions\" on the internet were mostly uninstalling and reinstalling both tensorflow and tensorflow-gpu (In same or older versions with pip), and I don't think it has worked. Again, any help would be very much appreciated. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":52929161,"AnswerCount":4,"Tags":"<python><tensorflow><object-detection><protoc>","CreationDate":"2018-10-22T12:16:44.277","AcceptedAnswerId":null,"OwnerUserId":10521738.0,"Title":"Cannot find `protoc` command","Body":"<p>When I run this:<\/p>\n\n<pre><code>protoc --python_out=. .\\object_detection\\protos\\anchor_generator.proto\n<\/code><\/pre>\n\n<p>on a tensorflow object detection file on the command line, I get:<\/p>\n\n<pre><code>'protoc' is not recognized as an internal or external command, operable program or batch file.\n<\/code><\/pre>\n\n<p>I followed all instructions in <a href=\"https:\/\/stackoverflow.com\/questions\/46491112\/protoc-object-detection-protos-proto-no-such-file-or-directory\">protoc object_detection\/protos\/*.proto: No such file or directory<\/a>, but they never worked for me <\/p>\n\n<p>In addition to the searches I made, I tried this command:<\/p>\n\n<pre><code>py-3.6 -m pip install protobuf\n<\/code><\/pre>\n\n<p>and it works for me.<\/p>\n\n<p>I typed  the command again <\/p>\n\n<pre><code>python_out=. .\\object_detection\\protos\\anchor_generator.proto \n<\/code><\/pre>\n\n<p>But it still fails and this:<\/p>\n\n<pre><code>C:\\Users\\laptop market\\Downloads\\bin\\protoc object_detection\/protos\/*.proto --py_out=.\n<\/code><\/pre>\n\n<p>also fails.<\/p>\n","answers":[{"AnswerId":"52937214","CreationDate":"2018-10-22T20:21:04.300","ParentId":null,"OwnerUserId":"8185230","Title":null,"Body":"<p>Depending on your platform, you might have to install protobuf locally. See <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md#manual-protobuf-compiler-installation-and-usage\" rel=\"nofollow noreferrer\">the docs<\/a> for more detailed instructions.<\/p>\n"},{"AnswerId":"52938217","CreationDate":"2018-10-22T21:49:04.613","ParentId":null,"OwnerUserId":"8995868","Title":null,"Body":"<p>First of all consider installing <code>Protobuf Compiler<\/code> inside the <em>C:\\Program Files<\/em> folder. Then you need to compile <code>protobuf<\/code> every time you want to use Tensorflow Object Detection API. To do this, open <strong>cmd<\/strong> in the <code>..\/models\/research<\/code> directory and run <\/p>\n\n<p><code>\u201cC:\\Program Files\\protoc-&lt;PROTO_VERSION&gt;\\bin\\protoc.exe\u201d object_detection\/protos\/*.proto --python_out=.<\/code><\/p>\n\n<p>where <code>&lt;PROTO_VERSION&gt;<\/code> is the version of protobuf compiler, you're using (for example <code>protoc-3.5.1-win64<\/code> or the name of the folder you've picked for the protobuf). <\/p>\n\n<p>Be advised, that you should compile protobuf in the every <strong>cmd<\/strong> window that is launched to work with Object Detection.<\/p>\n"},{"AnswerId":"54429608","CreationDate":"2019-01-29T21:09:21.077","ParentId":null,"OwnerUserId":"10987045","Title":null,"Body":"<p>Hey I got the solution you have to save the protoc.exe file which is located in the bin file to  C:\\Users\\Satyarth\\AppData\\Local\\Programs\\Python\\Python36\\Scripts   and it will solve the problem and if it still not give the above path to the enviroment variables and it will surely help that.\n  Change the \"Satyarth\" from the path and use yours<\/p>\n"},{"AnswerId":"55260211","CreationDate":"2019-03-20T11:56:30.837","ParentId":null,"OwnerUserId":"9360135","Title":null,"Body":"<p>Better you just go to this directory:\n<a href=\"https:\/\/www.github.com\/google\/protobuf\/releases\" rel=\"nofollow noreferrer\">https:\/\/www.github.com\/google\/protobuf\/releases<\/a> <\/p>\n\n<p>then just download the protoc.py file which suits your OS virsion\nand extract that one. you will get a protoc.exe file, then place it on your local machine<\/p>\n"}]}
+{"QuestionId":52929471,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-22T12:32:01.540","AcceptedAnswerId":null,"OwnerUserId":10042439.0,"Title":"Hooks in TensorFlow and distributed training","Body":"<p>I have a training job with a hook that sets variables during training:<\/p>\n\n<pre><code>class MyHook(tf.train.SessionRunHook):\n   ...\n   def begin(self):\n       self._assign_op = tf.assign(...)\n       self._assign_placeholder = ...\n\n   def after_run(self, run_context, run_values):\n       ...\n       run_context.session.run(self._assign_op, feed_dict={self._assign_placeholder: ...})\n<\/code><\/pre>\n\n<p>Training happens as usual witn estimators:<\/p>\n\n<pre><code>my_model.train(..., hooks=[MyHook()])\n<\/code><\/pre>\n\n<p>Now I'm trying to understand what happens when if I move over to distributed training. Will the hook run on all machines? If I currently find the tensor that I'm assigning to inside my hook using <code>session.graph.get_tensor_by_name()<\/code>, then will this work or will the tensor have a different name on the different replicas?<\/p>\n\n<p>In other words, what do I need to consider to make something like this work with multiple GPUs or multinode training?<\/p>\n","answers":[]}
+{"QuestionId":52929823,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras><mnist>","CreationDate":"2018-10-22T12:50:58.663","AcceptedAnswerId":null,"OwnerUserId":10502762.0,"Title":"how to add more data with mnist data to train keras model","Body":"<p>I tried to recognize using mnist data as train data but failed so I want to add my train data with that so how can I train my model with mnist dataset and my own data. I tried training with my own data also but failed any suggestions are welcome .....\nfor example first I trained with mnist dataset followed this approach <\/p>\n\n<p><a href=\"https:\/\/medium.com\/coinmonks\/handwritten-digit-prediction-using-convolutional-neural-networks-in-tensorflow-with-keras-and-live-5ebddf46dc8\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/coinmonks\/handwritten-digit-prediction-using-convolutional-neural-networks-in-tensorflow-with-keras-and-live-5ebddf46dc8<\/a><\/p>\n\n<p>but the results are not satisfying so I tried to train that model by my data but looking at the results maybe I don't have enough data so now I want to train my model using mnist data + my data . so how can I do that<\/p>\n","answers":[{"AnswerId":"52943843","CreationDate":"2018-10-23T07:52:17.647","ParentId":null,"OwnerUserId":"9883438","Title":null,"Body":"<p>There are several ways how to concatenate two <code>numpy<\/code> arrays. The most obvious one is probably <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.0\/reference\/generated\/numpy.concatenate.html\" rel=\"nofollow noreferrer\">np.concatenate()<\/a><\/p>\n\n<pre><code>import numpy as np\n\nall_train_x = np.concatenate((mnist_train_x, my_train_x))\nall_train_y = np.concatenate((mnist_train_y, my_train_y))\n<\/code><\/pre>\n\n<p>Note the double parentheses. <\/p>\n"},{"AnswerId":"52944254","CreationDate":"2018-10-23T08:16:31.307","ParentId":null,"OwnerUserId":"5349822","Title":null,"Body":"<p>You can also try <code>data augmentation<\/code> technique. This will increase you size of the dataset and will also improve the performance of your model. You can some useful information <a href=\"https:\/\/medium.com\/nanonets\/how-to-use-deep-learning-when-you-have-limited-data-part-2-data-augmentation-c26971dc8ced\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52929868,"AnswerCount":1,"Tags":"<python><tensorflow><keras><jupyter-notebook><google-colaboratory>","CreationDate":"2018-10-22T12:52:59.353","AcceptedAnswerId":52930654.0,"OwnerUserId":10539725.0,"Title":"Simple binary classification by CNN with Keras, But got only 50% acc","Body":"<p>Today I try to use fit_generator function for binary classifying plain black and white image but it gives me only 50% accuracy<\/p>\n\n<p>This is just my coding exercise but I think accuracy should reach 100%. So I am curious what is my mistake.<\/p>\n\n<p>I do all the code in Google-colaboratory.<\/p>\n\n<p>here is my code.<\/p>\n\n<p><strong>Set up<\/strong> <\/p>\n\n<pre><code>import numpy as np\nimport random\nfrom matplotlib import pyplot as plt\n\nimg_height = 150\nimg_width = 150\nbatch_size = 8\n\nclass MyDataset(object):\n\n    def __init__(self):\n        placeholder = 0\n\n    def generator(self):\n        is_black = True\n        X, y = [], []\n        while True:\n            if is_black:\n                img = np.full((img_height, img_width, 3), 255)\n            else:\n                img = np.zeros((img_height, img_width, 3))\n            img = img \/ 255.\n            X.append(img)\n            y.append(is_black)\n            is_black = not is_black\n\n            if len(X) &gt;= batch_size:\n                c = list(zip(X, y))\n                random.shuffle(c)\n                X, y = zip(*c)\n                yield np.asarray(X, dtype=np.float32), np.asarray(y, dtype=np.float32)\n                X, y = [], []\n\ndataset = MyDataset()\nsample_gen = dataset.generator()\n<\/code><\/pre>\n\n<p><strong>Visualize data<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Edc0C.png\" rel=\"nofollow noreferrer\">Example of inputs for model<\/a><\/p>\n\n<pre><code>X, y = next(sample_gen)\n\nlabel_dict = {0:'black', 1:'white'}\n\nsample_size = len(X)\n\nfig = plt.figure(figsize=(16, 8))\n\nfor sample in range(sample_size):\n    img = X[sample]\n    lbl = label_dict[y[sample]]\n\n    fig.add_subplot(2, sample_size\/\/2, sample + 1)\n    f = plt.imshow(img)\n    f.axes.get_xaxis().set_visible(False)\n    f.axes.get_yaxis().set_visible(False)\n    plt.title(lbl)\n\nplt.show()\n<\/code><\/pre>\n\n<p><strong>Create model<\/strong><\/p>\n\n<p>I create a small size model. It has only 9 parameters.<\/p>\n\n<pre><code>import tensorflow as tf\n\nmodel = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Conv2D(filters=1, kernel_size=(1,1), padding='same', \n                                 activation='relu', input_shape=(img_height, img_width, 3)))\nmodel.add(tf.keras.layers.MaxPooling2D(pool_size=(img_height\/\/2,img_height\/\/2)))\n\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(1, activation='softmax'))\n\nmodel.summary()\n<\/code><\/pre>\n\n<p><strong>Train model<\/strong><\/p>\n\n<pre><code>model.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit_generator(\n    sample_gen, \n    steps_per_epoch = 100\/\/batch_size , \n    epochs=300)\n<\/code><\/pre>\n\n<p><strong>Result<\/strong><\/p>\n\n<p>After 200+ epochs, accuracy still be 0.5.<\/p>\n\n<pre><code>Epoch 218\/300\n12\/12 [==============================] - 0s 8ms\/step - loss: 7.9712 - acc: 0.5000\nEpoch 219\/300\n12\/12 [==============================] - 0s 8ms\/step - loss: 7.9712 - acc: 0.5000\nEpoch 220\/300\n12\/12 [==============================] - 0s 8ms\/step - loss: 7.9712 - acc: 0.5000\nEpoch 221\/300\n12\/12 [==============================] - 0s 9ms\/step - loss: 7.9712 - acc: 0.5000\nEpoch 222\/300\n12\/12 [==============================] - 0s 8ms\/step - loss: 7.9712 - acc: 0.5000\n<\/code><\/pre>\n\n<p>I already studied a bit about CNN and I am beginner at Keras.<\/p>\n","answers":[{"AnswerId":"52930654","CreationDate":"2018-10-22T13:30:43.603","ParentId":null,"OwnerUserId":"9703830","Title":null,"Body":"<p>The problem is at the end of your model definition, specificaly here:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(1, activation='softmax'))\n<\/code><\/pre>\n\n<p>By applying softmax you -- by definition -- enforce it's outputs to sum to one. The only way how a single value can comply, is to become 1 itself. Therefore, no information is propagated through.<\/p>\n\n<p>To fix it, turn the softmax into a logistic sigmoid, e.g.:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>This way, you also can interpret the output of your model as the posterior probability that the data comes from class <code>1<\/code>.<\/p>\n"}]}
+{"QuestionId":52930129,"AnswerCount":0,"Tags":"<python><apache-spark><tensorflow><keras><multiprocessing>","CreationDate":"2018-10-22T13:05:10.453","AcceptedAnswerId":null,"OwnerUserId":9024698.0,"Title":"Keras\/Tensorflow reproducible results on Spark","Body":"<p>I am running a python program with Keras (where I am performing grid search on a deep neural network) on a Spark cluster and I want to secure that my results are reproducible.<\/p>\n\n<p>My Spark cluster is ignited by running the following <code>.sh<\/code> script:<\/p>\n\n<pre><code>SPARK_HOME=\/u\/users\/******\/spark-2.3.0 \\\nQ_CORE_LOC=\/u\/users\/******\/q-core \\\nENV=local \\\nHIVE_HOME=\/usr\/hdp\/current\/hive-client \\\nSPARK2_HOME=\/u\/users\/******\/spark-2.3.0 \\\nHADOOP_CONF_DIR=\/etc\/hadoop\/conf \\\nHIVE_CONF_DIR=\/etc\/hive\/conf \\\nHDFS_PREFIX=hdfs:\/\/ \\\nPYTHONPATH=\/u\/users\/******\/q-core\/python-lib:\/u\/users\/******\/three-queues\/python-lib:\/u\/users\/******\/pyenv\/prod_python_libs\/lib\/python2.7\/site-packages\/:$PYTHON_PATH \\\nYARN_HOME=\/usr\/hdp\/current\/hadoop-yarn-client \\\nSPARK_DIST_CLASSPATH=$(hadoop classpath):$(yarn classpath):\/etc\/hive\/conf\/hive-site.xml \\\nPYSPARK_PYTHON=\/usr\/bin\/python2.7 \\\nQQQ_LOC=\/u\/users\/******\/three-queues \\\nspark-submit \\\n--master yarn 'dnn_grid_search.py' \\\n--executor-memory 10g \\\n--num-executors 8 \\\n--executor-cores 10 \\\n--conf spark.port.maxRetries=80 \\\n--conf spark.dynamicAllocation.enabled=False \\\n--conf spark.default.parallelism=6000 \\\n--conf spark.sql.shuffle.partitions=6000 \\\n--principal ************************ \\\n--queue default \\\n--name lets_get_starting \\\n--keytab \/u\/users\/******\/.******.keytab \\\n--driver-memory 10g\n<\/code><\/pre>\n\n<p>By running <code>nohup .\/spark_python_shell.sh &gt; output.log &amp;<\/code> at my <code>bash<\/code> shell I ignite the Spark cluster and I also get my python-Keras script running (see above <code>spark-submit \\ --master yarn 'dnn_grid_search.py'<\/code>).<\/p>\n\n<p>To secure the reproducibility of results I tried to do what I successfully did on my laptop's CPU to get reproducible results (see also my answers on StackOverflow: <a href=\"https:\/\/stackoverflow.com\/a\/52897289\/9024698\">answer_1<\/a>, <a href=\"https:\/\/stackoverflow.com\/a\/52897216\/9024698\">answer_2<\/a>):<\/p>\n\n<pre><code># Seed value\n# Apparently you may use different seed values at each stage\nseed_value= 0\n\n# 1. Set `PYTHONHASHSEED` environment variable at a fixed value\nimport os\nos.environ['PYTHONHASHSEED']=str(seed_value)\n\n# 2. Set `python` built-in pseudo-random generator at a fixed value\nimport random\nrandom.seed(seed_value)\n\n# 3. Set `numpy` pseudo-random generator at a fixed value\nimport numpy as np\nnp.random.seed(seed_value)\n\n# 4. Set `tensorflow` pseudo-random generator at a fixed value\nimport tensorflow as tf\ntf.set_random_seed(seed_value)\n\n# 5. Configure a new global `tensorflow` session\nfrom keras import backend as K\nsession_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)\nsess = tf.Session(graph=tf.get_default_graph(), config=session_conf)\nK.set_session(sess)\n<\/code><\/pre>\n\n<p>However, by including part <code>(5)<\/code> then I am pretty sure that the program is not running properly because some <code>print<\/code> statements which I have in it do not print anything in the <code>output.log<\/code> file.<\/p>\n\n<p>Specifically, I created a very simple custom scorer function to insert this print statement to it so that I can see if the grid search is really running:<\/p>\n\n<pre><code>def precision_recall(y_true, y_predict):\n\n    # Print time\n    from datetime import datetime\n    time_now = datetime.utcnow()\n    print('Grid Search time now:', time_now)\n\n    from sklearn.metrics import precision_recall_fscore_support\n    _, recall, _, _ = precision_recall_fscore_support(y_true, y_predict, average='binary')\n\n    return  recall\n\nfrom sklearn.metrics import make_scorer\ncustom_scorer = make_scorer(score_func=precision_recall, greater_is_better=True, needs_proba=False)\n\n# Execute grid search\n# Notice 'n_jobs=-1' for parallelizing the jobs (across the cluster)\nfrom sklearn.model_selection import GridSearchCV\nclassifiers_grid = GridSearchCV(estimator=classifier, param_grid=parameters, scoring=custom_scorer, cv=5, n_jobs=-1)\nclassifiers_grid.fit(X, y)\n<\/code><\/pre>\n\n<p>But while the print statement within the custom score function (<code>precision_recall<\/code>) was properly printing the time when I was running other ML algorithms (random forest etc) then when I tried that with Keras\/Tensorflow with its seed system and session then nothing is printed.<\/p>\n\n<p><strong>Therefore, my question is how can I get reproducible results with Keras\/Tensorflow on Spark?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":52930708,"AnswerCount":2,"Tags":"<python><file><keras>","CreationDate":"2018-10-22T13:33:40.603","AcceptedAnswerId":null,"OwnerUserId":10536694.0,"Title":"error list index out of range instead of proper listing in python","Body":"<p>Following code:<\/p>\n\n<pre><code>from keras.preprocessing.sequence import skipgrams\nfrom keras.preprocessing import text\n\nwith open('engl_bible.txt', 'r') as file:\nnorm_bible = file\n\ntokenizer = text.Tokenizer()\ntokenizer.fit_on_texts(norm_bible)\n\nword2id = tokenizer.word_counts\nid2word = {v:k for k, v in word2id.items()}\n\nvocab_size = len(word2id) + 1\nembed_size = 100\n\nwids = [[word2id[w] for w in text.text_to_word_sequence(doc)] for doc in norm_bible]\n\n#print('Vocabulary Size:', vocab_size)\n#print('Vocabulary Sample:', list(word2id.items())[:10])\n\n\n\n# generate skip-grams\n#skip_grams = keras.preprocessing.sequence.skipgrams(sequence, vocabulary_size, window_size=4, negative_samples=1.0, shuffle=True,\n                                      # categorical=False, sampling_table=None, seed=None)\n\nskip_grams = [skipgrams(wid, vocabulary_size=vocab_size, window_size=10) for wid in wids]\n\n# view sample skip-grams\npairs = skip_grams[0]\nlabels = skip_grams[1]\nfor i in range(10):\n    print(\"({:s} ({:d}), {:s} ({:d})) -&gt; {:d}\".format(id2word[pairs[i][0]], pairs[i][0],id2word[pairs[i][1]], pairs[i][1],labels[i]))\n<\/code><\/pre>\n\n<p>this should lead to an output like this:<\/p>\n\n<pre><code>(james (1154), king (13)) -&gt; 1\n(king (13), james (1154)) -&gt; 1\n(james (1154), perform (1249)) -&gt; 0\n(bible (5766), dismissed (6274)) -&gt; 0\n<\/code><\/pre>\n\n<p>but the output I get is the following:<\/p>\n\n<pre><code>pairs = skip_grams[0]\nIndexError: list index out of range\n<\/code><\/pre>\n","answers":[{"AnswerId":"52930995","CreationDate":"2018-10-22T13:47:59.550","ParentId":null,"OwnerUserId":"10540131","Title":null,"Body":"<p>It means that <code>skip_grams<\/code> is holding an empty list. That's why when you are doing <code>skip_grams[0]<\/code>, you are getting the error as \"IndexError: list index out of range\" because there is nothing at 0th index in your list.<\/p>\n\n<p>On seeing you code, I see these relevant lines:<\/p>\n\n<pre><code>with open('engl_bible.txt', 'r') as file:\n    norm_bible = file\n...\n\nwids = [[word2id[w] for w in text.text_to_word_sequence(doc)] for doc in norm_bible]\n\n...\n\nskip_grams = [skipgrams(wid, vocabulary_size=vocab_size, window_size=10) for wid in wids]\n<\/code><\/pre>\n\n<p>which makes me think that your file 'engl_bible.txt' is empty.<\/p>\n\n<p><strong>Empty file <code>'engl_bible.txt'<\/code>-> empty list <code>wids<\/code> -> empty list <code>skip_grams<\/code><\/strong><\/p>\n"},{"AnswerId":"52931608","CreationDate":"2018-10-22T14:20:54.987","ParentId":null,"OwnerUserId":"3288092","Title":null,"Body":"<p>I highly suspect you are not opening the file right and you might want to fix your indentations. Try opening the file with:<\/p>\n\n<pre><code>with open('engl_bible.txt', 'r') as file:\n    norm_bible = file.read()\n    #norm_bible = file.readlines()\n\nprint(len(norm_bible))\n<\/code><\/pre>\n\n<p>Your initial code does not actually read the file. <code>read()<\/code> returns the file contents as a string and <code>read_lines()<\/code> returns each newline as an element in a list. Seeing that you are applying a tokenizer, <code>read()<\/code> might be the way to go.<\/p>\n\n<p>Then check if your <code>norm_bible<\/code> variable is not empty.<\/p>\n"}]}
+{"QuestionId":52930826,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-22T13:39:39.573","AcceptedAnswerId":52930943.0,"OwnerUserId":8954691.0,"Title":"Restoring Tensorflow model from a previous checkpoint","Body":"<p>I am tuning my model for 200 episodes by training it 2 times each for 100 episodes. The first run for 100 eps is finished and right now 2nd run is going on. My problem is I want to restore the model after it finished the first run that is from the 100th episode. And now that I have already started the 2nd run, it has overwritten the checkpoint files. <\/p>\n\n<p>I came across <a href=\"https:\/\/stackoverflow.com\/questions\/43386016\/tensorflow-saver-seems-to-overwrite-existing-saved-variable-files\">this post<\/a> which says tensorflow keeps last 5 models. How can I retrieve the model parameters from the previous checkpoint which was created after the first run? <\/p>\n","answers":[{"AnswerId":"52930943","CreationDate":"2018-10-22T13:45:44.707","ParentId":null,"OwnerUserId":"10288309","Title":null,"Body":"<p>Tensorflow saves the best 5 models in a single run, you have started another new run, i am afraid your models are lost, you will have to retrain them, rule of thumb always keep the naming of your models dynamic so they aren't replaced, for example:<\/p>\n\n<pre><code>model_name = 'Model_{:d}'.format(episode_number) + '.ckpt'\n<\/code><\/pre>\n"}]}
+{"QuestionId":52931007,"AnswerCount":0,"Tags":"<android><python><tensorflow><keras><image-recognition>","CreationDate":"2018-10-22T13:49:03.907","AcceptedAnswerId":null,"OwnerUserId":9299873.0,"Title":"Same label value for all digit in handwritten digit recognition from photo (not drawing canvas) on android using keras, tensorflow","Body":"<p>I am working on an Android app that recognizes handwritten Bengali digit from both photo and canvas. I have trained the dataset in python using keras and tensorflow as backend. The app works for digit drawn in canvas and detects all 10 digits perfectly. \nFor photo chosen from gallery, I followed the same .pb file and code to detect from photos containing digit (written on white paper with black marker). I took the bitmap value of image by <code>MediaStore.Images.Media.getBitmap(getApplicationContext().getContentResolver(),uri)<\/code> \nand scaled it to 32*32 (input size). Then got the pixel values and converted color to use in tensorflow. \nWhen I run the app it detects 5 for all digits with a low confidence. I think trained model is okay since it gives accurate answer when I draw any digit on the canvas. Please help to find why this is happening :( TIA.<\/p>\n\n<pre><code>private void onRecognize() {\n\n    Bitmap scaledImage = Bitmap.createScaledBitmap(bitmap, 32, 32, false);\n    image.setImageBitmap(scaledImage);\n\n    int pixels[] = new int[height*width];\n    scaledImage.getPixels(pixels,0,height, 0,0,width,height);\n    float[] normalizedPixels = normalizePixels(pixels);\n\n    int[] previewPixels = pixelsForPreview(pixels, normalizedPixels);\n\n    Bitmap preview = Bitmap.createBitmap(previewPixels, width, height, Bitmap.Config.ARGB_8888);\n    image.setImageBitmap(preview);\n\n    classify(normalizedPixels);\n}\n\nprivate void classify(float[] normalizedPixels) {\n    DigitClassification dc = digitClassifier.recognize(normalizedPixels);\n    String result = String.format(\"Digit %s with confidence: %f\",dc.getLabel(),dc.getConfidence());\n    Toast.makeText(this, result, Toast.LENGTH_SHORT).show();\n}\n\npublic DigitClassification recognize(float[] pixels){\n\n    tf.feed(input_name, pixels, 1, input_size, input_size, 1);\n    tf.run(output_names);\n    tf.fetch(output_name,output);\n\n    DigitClassification digit = new DigitClassification();\n\n    for(int i=0;i&lt;output.length;i++){\n        Log.d(TAG, String.format(\"Class: %s Confidence: %f\", labels.get(i), output[i]));\n        if (output[i] &gt; threshold &amp;&amp; output[i] &gt; digit.getConfidence()) {\n            digit.update(labels.get(i),output[i]);\n        }\n    }\n\n    return digit;\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52931047,"AnswerCount":0,"Tags":"<python><api><tensorflow><keras><layer>","CreationDate":"2018-10-22T13:51:06.970","AcceptedAnswerId":null,"OwnerUserId":6509922.0,"Title":"ValueError: Output tensors to a Model must be the output of a TensorFlow `Layer` (thus holding past layer metadata).","Body":"<p>Encountering an error using the functional API of keras from within TensorFlow. The layers compile fine until I attempt to initiate the Model object (tf.keras.Model) at which point I receive the below error and traceback.<\/p>\n\n<p>The last layer wrapper is below:<\/p>\n\n<pre><code>def dense(self,kwargs):\n    print('simple old Dense layer')\n    self.layer_outputs[kwargs['LayerID']] = tf.keras.layers.Lambda(tf.layers.Dense(name=self._namer(kwargs['LayerID']), **kwargs['LayerKwargs']))(self.layer_outputs[kwargs['LayerInput']])\n    if kwargs['args'].setdefault('norm',True):\n        self.layer_outputs[kwargs['LayerID']] = self.batchnorm_relu(self.layer_outputs[kwargs['LayerID']],kwargs['LayerID'])\n    return self.layer_outputs[kwargs['LayerID']]\n<\/code><\/pre>\n\n<p>I've reviewed other questions regarding similar errors, <a href=\"https:\/\/stackoverflow.com\/questions\/50715928\/valueerror-output-tensors-to-a-model-must-be-the-output-of-a-tensorflow-layer\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/51506691\/valueerror-output-tensors-to-a-model-must-be-the-output-of-a-tensorflow-layer-w?rq=1\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/43633883\/output-tensors-to-a-model-must-be-keras-tensors\">here<\/a>. All of these issues were resolved either by correcting the input or wrapping the output in a keras.Lambda layer wrapper, neither of which seem to be working for me. <\/p>\n\n<p>My input is below:<\/p>\n\n<pre><code>self.input_proxy = tf.keras.layers.Input(shape=self.batch_input_shape[1:],dtype=tf.float32)\n<\/code><\/pre>\n\n<p>One answer which I haven't yet tried is <a href=\"https:\/\/stackoverflow.com\/questions\/43633883\/output-tensors-to-a-model-must-be-keras-tensors\">here<\/a>, wherein I would assign the missing requisite meta_data manually. This may work, but seems awfully hacky and I'd like a more straightforward solution that works with the API if possible.<\/p>\n\n<p>Why is this happening and how can I resolve it?<\/p>\n\n<pre><code>   C:\\Users\\asus\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in __init__(self, *args, **kwargs)\n        115\n        116   def __init__(self, *args, **kwargs):\n    --&gt; 117     super(Model, self).__init__(*args, **kwargs)\n        118     # Create a cache for iterator get_next op.\n        119     self._iterator_get_next = weakref.WeakKeyDictionary()\n\n    C:\\Users\\asus\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in __init__(self, *args, **kwargs)\n         78         'inputs' in kwargs and 'outputs' in kwargs):\n         79       # Graph network\n    ---&gt; 80       self._init_graph_network(*args, **kwargs)\n         81     else:\n         82       # Subclassed network\n\n    C:\\Users\\asus\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\training\\checkpointable\\base.py in _method_wrapper(self, *args, **kwargs)\n        424     self._setattr_tracking = False  # pylint: disable=protected-access\n        425     try:\n    --&gt; 426       method(self, *args, **kwargs)\n        427     finally:\n        428       self._setattr_tracking = previous_value  # pylint: disable=protected-access\n\n    C:\\Users\\asus\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in _init_graph_network(self, inputs, outputs, name)\n        222         raise ValueError('Output tensors to a ' + cls_name + ' must be '\n        223                          'the output of a TensorFlow `Layer` '\n    --&gt; 224                          '(thus holding past layer metadata). Found: ' + str(x))\n        225\n        226     self._base_init(name=name)\n\n    ValueError: Output tensors to a Model must be the output of a TensorFlow `Layer` (thus holding past layer metadata). Found: Tensor(\"lambda\/Last__0\/BiasAdd:0\",\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52931446,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-22T14:12:55.937","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"How to add dynamic dimension to tensor in tensorflow?","Body":"<p>I am having a <code>input_node<\/code> with shape <code>(None, 80, 3)<\/code> meaning <code>(sample_length, resolution, channel)<\/code>. However, this tensor is missing a fourth dimension for the batches. Is there a way to reshape <code>input_node<\/code> to <code>(None, None, 80, 3)<\/code>?<\/p>\n\n<hr>\n\n<p>Using <code>tf.expand_dims<\/code> I won't be able to work with batches:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>wave_input = tf.placeholder(dtype=tf.float32, shape=[None], name='wave_input')\nencoded_dict = problem.preprocess_example({'waveforms': wave_input, 'targets': [0]}, mode, hparams)\n\n# This is where a dynamic shape would be required\nencoded_dict['inputs'] = tf.expand_dims(encoded_dict['inputs'], 0)\nencoded_dict['targets'] = tf.expand_dims(encoded_dict['targets'], 0)\n\n# Get input\/output nodes\n\ninferred = model.infer(encoded_dict)\n\ntf.identity(encoded_dict['inputs'], name=input_name)\ntf.identity(inferred['outputs'], name=output_name)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52932406,"AnswerCount":1,"Tags":"<python><tensorflow><keras><generator>","CreationDate":"2018-10-22T15:02:33.817","AcceptedAnswerId":53617602.0,"OwnerUserId":9328846.0,"Title":"Is the class generator (inheriting Sequence) thread safe in Keras\/Tensorflow?","Body":"<p>For making the training of a model faster, it seems to be a good practice to populate\/generate batches on CPU and run the training of the model on GPU in parallel. For this purpose, a generator class can be written in Python that inherits the <code>Sequence<\/code> class.<\/p>\n\n<p>Here is the link to the documentation:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/utils\/Sequence\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/utils\/Sequence<\/a><\/p>\n\n<p>The important thing that the document states is:  <\/p>\n\n<blockquote>\n  <p><code>Sequence<\/code> are a safer way to do multiprocessing. This structure\n  guarantees that the network will only train once on each sample per\n  epoch which is not the case with generators.<\/p>\n<\/blockquote>\n\n<p>And it gives a simple code example as following:<\/p>\n\n<pre><code>from skimage.io import imread\nfrom skimage.transform import resize\nimport numpy as np\nimport math\n\n# Here, `x_set` is list of path to the images\n# and `y_set` are the associated classes.\n\nclass CIFAR10Sequence(Sequence):\n\n    def __init__(self, x_set, y_set, batch_size):\n        self.x, self.y = x_set, y_set\n        self.batch_size = batch_size\n\n    def __len__(self):\n        return math.ceil(len(self.x) \/ self.batch_size)\n\n    def __getitem__(self, idx):\n        batch_x = self.x[idx * self.batch_size:(idx + 1) *\n        self.batch_size]\n        batch_y = self.y[idx * self.batch_size:(idx + 1) *\n        self.batch_size]\n\n        return np.array([\n            resize(imread(file_name), (200, 200))\n               for file_name in batch_x]), np.array(batch_y)\n<\/code><\/pre>\n\n<p>What - to my understanding - ideally needs to be done in the model is to create an instance of this generator class and give it to the <code>fit_generator(...)<\/code> function.<\/p>\n\n<pre><code>gen = CIFAR10Sequence(x_set, y_set, batch_size)\n# Train the model\nmodel.fit_generator(generator=gen,\n                    use_multiprocessing=True,\n                    workers=6)\n<\/code><\/pre>\n\n<p>Here is a quote from Keras documentation:<\/p>\n\n<blockquote>\n  <p>The use of <code>keras.utils.Sequence<\/code> guarantees the ordering and guarantees\n  the single use of every input per epoch <strong>when using\n  <code>use_multiprocessing=True<\/code><\/strong>.<\/p>\n<\/blockquote>\n\n<p>In this shape, I assume that this setup is thread safe. \n<strong>Question 1)<\/strong> Is my assumption correct?<\/p>\n\n<p>One confusing thing though is that the parameter <code>use_multiprocessing<\/code> may not be set to True on Windows 10. Keras does not allow it; seemingly it only can be set to True on Linux. (I don't know how it is in other platforms.) But the <code>workers<\/code> parameter can still be set to a value that is greater than 0.<\/p>\n\n<p>Let's have a look at the definition of these 2 parameters:<\/p>\n\n<blockquote>\n  <p><strong><code>workers:<\/code><\/strong> Integer. Maximum number of processes to spin up when using\n  process-based threading. If unspecified, workers will default to 1. If\n  0, will execute the generator on the main thread. <\/p>\n  \n  <p><strong><code>use_multiprocessing:<\/code><\/strong> Boolean. If True, use process-based threading. If\n  unspecified, use_multiprocessing will default to False. Note that\n  because this implementation relies on multiprocessing, you should not\n  pass non-picklable arguments to the generator as they can't be passed\n  easily to children processes.<\/p>\n<\/blockquote>\n\n<p>So, by using the <code>workers<\/code> parameter, it seems to be possible to create multiple processes to speed up the training independent from whether <code>use_multiprocessing<\/code> is True or not.<\/p>\n\n<p>If one wants to use the generator class inheriting <code>Sequence<\/code> (<strong>on Windows 10<\/strong>), s\/he has to set the <code>use_multiprocessing<\/code> to False as following:<\/p>\n\n<pre><code>gen = CIFAR10Sequence(x_set, y_set, batch_size)\n# Train the model\nmodel.fit_generator(generator=gen,\n                    use_multiprocessing=False,  # CHANGED\n                    workers=6)\n<\/code><\/pre>\n\n<p>And there are still multiple processes running here because workers = 6.<\/p>\n\n<p><strong>Question 2)<\/strong> Is this setup still thread safe or is the thread safe characteristic lost now after setting the <code>use_multiprocessing<\/code> parameter to False? I cannot make it clear based on documentation.<\/p>\n\n<p><strong>Question 3)<\/strong>  Still related to this topic... When training is done in this way where data is generated by the CPU and training on GPU, if the model that is being trained is shallow, the GPU utilization ends up being very low and CPU utilization becomes significantly higher because the GPU keeps waiting for data that will come from CPU. In such cases, is there a way to utilize some GPU resources as well for data generation?<\/p>\n","answers":[{"AnswerId":"53617602","CreationDate":"2018-12-04T16:36:04.183","ParentId":null,"OwnerUserId":"9328846","Title":null,"Body":"<p>Among those who have seen this post, no one seems to have the ultimate answer so that I wanted to give my answer that worked out for me. Because of lack of documentation in the domain, my answer might be missing some relevant details. Please feel free to add more information that I do not mention down here.<\/p>\n\n<p>Seemingly, writing a generator class in Python that inherits the <code>Sequence<\/code> class is just not supported in <strong>Windows<\/strong>. (You can seemingly make it work on Linux.) To be able to make it work, you need to set the parameter <code>use_multiprocessing=True<\/code> (with the class approach). But it is not working on Windows as mentioned so that you have to set <code>use_multiprocessing<\/code> to <code>False<\/code> (on Windows).\nNevertheless, that does not mean that multiprocessing does not work on Windows. Even if you set <code>use_multiprocessing=False<\/code>, multiprocessing can still be supported when the code is run with the following setup where you just set the <code>workers<\/code> parameter to any value that is bigger than 1.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>history = \\\n   merged_model.fit_generator(generator=train_generator,\n                              steps_per_epoch=trainset_steps_per_epoch,\n                              epochs=300,\n                              verbose=1,\n                              use_multiprocessing=False,\n                              workers=3,\n                              max_queue_size=4)\n<\/code><\/pre>\n\n<p>At this point, let's remember the Keras documentation again:<\/p>\n\n<blockquote>\n  <p>The use of keras.utils.Sequence guarantees the ordering and guarantees\n  the single use of every input per epoch when using\n  use_multiprocessing=True.<\/p>\n<\/blockquote>\n\n<p>To my understanding, if <code>use_multiprocessing=False<\/code>, then the generator is not thread safe anymore, which makes it difficult to write a <strong>generator class<\/strong> that inherits <code>Sequence<\/code>.<\/p>\n\n<p>To come around this problem, I have written a generator myself which I have made thread safe manually. Here is an example pseudocode:<\/p>\n\n<pre><code>import tensorflow as tf\nimport threading\n\nclass threadsafe_iter:\n    \"\"\"Takes an iterator\/generator and makes it thread-safe by\n    serializing call to the `next` method of given iterator\/generator.\n    \"\"\"\n    def __init__(self, it):\n        self.it = it\n        self.lock = threading.Lock()\n\n    def __iter__(self):\n        return self\n\n    def __next__(self): # Py3\n        return next(self.it)\n\n    #def next(self):     # Python2 only\n    #    with self.lock:\n    #        return self.it.next()\n\ndef threadsafe_generator(f):\n    \"\"\"A decorator that takes a generator function and makes it thread-safe.\n    \"\"\"\n    def g(*a, **kw):\n        return threadsafe_iter(f(*a, **kw))\n    return g\n\n\n@threadsafe_generator\ndef generate_data(tfrecord_file_path_list, ...):\n\n    dataset = tf.data.TFRecordDataset(tfrecord_file_path_list)\n\n    # example proto decode\n    def _parse_function(example_proto):\n      ...\n      return batch_data\n\n    # Parse the record into tensors.\n    dataset = dataset.map(_parse_function)  \n\n    dataset = dataset.shuffle(buffer_size=100000)\n\n    # Repeat the input indefinitly\n    dataset = dataset.repeat()  \n\n    # Generate batches\n    dataset = dataset.batch(batch_size)\n\n    # Create an initializable iterator\n    iterator = dataset.make_initializable_iterator()\n\n    # Get batch data\n    batch_data = iterator.get_next()\n\n    iterator_init_op = iterator.make_initializer(dataset)\n\n    with tf.Session() as sess:\n\n        sess.run(iterator_init_op)\n\n        while True:            \n            try:\n                batch_data = sess.run(batch_data)\n            except tf.errors.OutOfRangeError:\n                break\n            yield batch_data\n<\/code><\/pre>\n\n<p>Well, it can be discussed if it is really elegant to do it in this way but it seems to be working pretty well.<\/p>\n\n<p>To summarize:<\/p>\n\n<ul>\n<li>If writing your program on Windows, set <code>use_multiprocessing<\/code> to <code>False<\/code>.<\/li>\n<li>(As of today, to my knowledge) it is not supported to write a generator class that inherits <code>Sequence<\/code> when writing code on Windows. (It is a Tensorflow\/Keras problem I guess).<\/li>\n<li>To come around the problem, write an ordinary generator, make your generator thread safe, and set <code>workers<\/code> to a number that is greater than 1.<\/li>\n<\/ul>\n\n<p><strong>Important note:<\/strong> In this setup, the generator is being run on CPU and the training is being done on GPU. One problem I could observe is that if the model you are training is shallow enough, the utilization of GPU remains very low while CPU utilization gets high. If the model is shallow and the dataset is small enough, it can be a good option to store all the data in the memory and run everything on GPU. It should speed up the training significantly. If, for any reason, you would like to use CPU and GPU simultaneously, my modest recommendation is to try to use Tensorflow's tf.data API which significantly speeds up the data preprocessing and batch preparation. If the generator is only written in Python, GPU keeps waiting for data to continue with the training. One can say everything about the Tensorflow\/Keras documentation, but it is really efficient code!<\/p>\n\n<p>Anyone having more complete knowledge on the API and seeing this post, please feel free to correct me here in case I misunerstand anything or the API is updated to solve the problems even on Windows.<\/p>\n"}]}
+{"QuestionId":52933070,"AnswerCount":1,"Tags":"<json><tensorflow><tensorflow-serving>","CreationDate":"2018-10-22T15:41:18.790","AcceptedAnswerId":52938419.0,"OwnerUserId":2621303.0,"Title":"Inference on served tensorflow model fails with float input","Body":"<p>I'm trying to run inference on a served tensorflow model. More specifically, sending a JSON request through REST to get a prediction. <\/p>\n\n<p>I'm actually just doing the basic example outlined here:\n<a href=\"https:\/\/www.tensorflow.org\/serving\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/serving\/<\/a><\/p>\n\n<p>Everything works properly, except when I change the \"instances\" input to floats that don't end in '.0', it fails.<\/p>\n\n<p><code>curl -d '{\"instances\": [1.1, 2.2, 5.3]}' \\\n   -X POST http:\/\/localhost:8501\/v1\/models\/half_plus_two:predict<\/code><\/p>\n\n<p><strong>{ \"error\": \"Failed to process element: 0 of \\'instances\\' list. Error: Invalid argument: Cannot convert JSON value: 1.1 to float without loss of precision.\" }<\/strong><\/p>\n\n<p>For my eventual model though, I will need to pass floats through the JSON request. I also tried to send it using python, using simplejson to serialize the Decimal when sending the request.<\/p>\n\n<p>I'm assuming it can't parse the float properly when trying to load the JSON. Is this something that has to be coded in the graph of the model? (note that I didn't code the graph\/model so I'm not too familiar with it)<\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"52938419","CreationDate":"2018-10-22T22:09:11.537","ParentId":null,"OwnerUserId":"3301620","Title":null,"Body":"<p>I just came across this error too. It's a <a href=\"https:\/\/github.com\/tensorflow\/serving\/commit\/01c2fb804161125bebda0b5880657ad046ab2da7\" rel=\"nofollow noreferrer\">bug<\/a> in Tensorflow Serving <code>1.11<\/code>, a fix has already been made just not released yet. <\/p>\n\n<p>To work around this for now, you can install the previous version, tag <code>1.10.0<\/code>, of the Tensorflow Serving image instead of the latest. <\/p>\n\n<pre><code>docker pull tensorflow\/serving:1.10.0\n<\/code><\/pre>\n\n<p>Just, make sure that have removed the latest image and are running the correct version tag like below.<\/p>\n\n<pre><code>\u279c docker images\nREPOSITORY           TAG                 IMAGE ID            CREATED             SIZE\ntensorflow\/serving   1.10.0              a7b9844ed1ae        2 months ago        215MB\n<\/code><\/pre>\n\n<p>Your curl request with payload containing <code>float<\/code>s should work now. <\/p>\n"}]}
+{"QuestionId":52933222,"AnswerCount":1,"Tags":"<python><keras><deep-learning><conv-neural-network><lstm>","CreationDate":"2018-10-22T15:49:54.293","AcceptedAnswerId":null,"OwnerUserId":10083414.0,"Title":"How to combine RNN with CNN?","Body":"<p>I'm trying to combine LSTM with CNN but I got stuck because of an error.\nHere is the model I'm trying to implement:<\/p>\n\n<pre><code>model=Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=(28, 28,3), activation='relu'))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='relu'))\nmodel.add(LSTM(128, return_sequences=True,input_shape=(1,32), activation='relu'))\nmodel.add(LSTM(256))\nmodel.add(Dropout(0.25))\nmodel.add(Dense(37))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam')\n<\/code><\/pre>\n\n<p>and error happens in the first LSTM layer:<\/p>\n\n<pre><code>ERROR: Input 0 is incompatible with layer lstm_12: expected ndim=3, found ndim=2\n<\/code><\/pre>\n","answers":[{"AnswerId":"52935152","CreationDate":"2018-10-22T17:59:41.750","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The input of LSTM layer should be a 3D array which represents a sequence or a timeseries (this is what the error is trying to say: <code>expected ndim=3<\/code>). However, in your model the input of LSTM layer, which is actually the output of the Dense layer before it, is a 2D array (i.e. <code>found ndim=2<\/code>). To make it into a 3D array of shape <code>(n_samples, n_timesteps, n_features)<\/code>, one solution is to use a <code>RepeatVector<\/code> layer to repeat it as much as the number of timesteps (which you need to specify in your code):<\/p>\n\n<pre><code>model.add(Dense(32, activation='relu'))\nmodel.add(RepeatVector(n_timesteps))\nmodel.add(LSTM(128, return_sequences=True, input_shape=(n_timesteps,32), activation='relu'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52933327,"AnswerCount":0,"Tags":"<tensorflow><machine-learning>","CreationDate":"2018-10-22T15:55:40.357","AcceptedAnswerId":null,"OwnerUserId":7571462.0,"Title":"TensorFlow Output: One-Hot Encoding vs. Index","Body":"<p>I want to ask a conceptual question about when to use one-hot encoding and when to use index to represent the labels in multi-class classification problems in tensorflow. I encountered the dimension problems about these, because I am not sure when to use which.<\/p>\n\n<p>For example, in this fully connected NN example, one-hot encoding is proper. (<a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/neural_network_raw.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/neural_network_raw.py<\/a>)\nBut in this CNN example, index is proper. (<a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/convolutional_network.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/convolutional_network.py<\/a>).<\/p>\n\n<p>When I used one-hot encoding for the labels in the CNN example code, I got the error: \"ValueError: Rank mismatch: Rank of labels (received 2) should equal rank of logits minus 1 (received 2)\". But when I used index for labels, no problem.<\/p>\n\n<p>Could someone explain when to use one-hot encoding and when to use index in tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":52933410,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-22T16:00:49.640","AcceptedAnswerId":null,"OwnerUserId":10541543.0,"Title":"Pre-trained \"CNN\" model for \"Head Pose\" detection","Body":"<p>Dears,<\/p>\n\n<p>If you please, I am trying to implement a \"Head Pose\" (Front, Up, Down, Left, Right) detection project using \"CNN\" and I need to use a pre-trained network as I have small amount of training images.<\/p>\n\n<p>My questions are:<\/p>\n\n<ul>\n<li>What is the best architecture for such a task (LeNet, VGG, ...)?<\/li>\n<li>How to decide which layer\/layers to train with my data?<\/li>\n<\/ul>\n\n<p>Note: I am using \"Python\/Tensorflow\/Keras\" and I have a mid-range Nvidia GPU.<\/p>\n\n<p>Thank you very much.<\/p>\n","answers":[]}
+{"QuestionId":52933947,"AnswerCount":2,"Tags":"<python><tensorflow><keras><gpu><cpu>","CreationDate":"2018-10-22T16:35:13.980","AcceptedAnswerId":null,"OwnerUserId":10541724.0,"Title":"keras not using gpu but tensorflow is","Body":"<p>Keras is not using the my GPU, even though tensorflow seems to run fine with it. I followed other folks suggestion to check tensorflow with:<\/p>\n\n<pre><code>import tensorflow\nfrom tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n<\/code><\/pre>\n\n<p>Which gives<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 13541243483275802230\n, name: \"\/device:GPU:0\"\ndevice_type: \"GPU\"\nmemory_limit: 6694395576\nlocality {\n  bus_id: 1\n  links {\n  }\n}\nincarnation: 17715053295272939021\nphysical_device_desc: \"device: 0, name: GeForce GTX 1070, pci bus id: 0000:08:00.0, compute capability: 6.1\"\n]\n<\/code><\/pre>\n\n<p>So far so good, but when I specify a classifier in Keras and train it, it runs at glacial pace. No sign of GPU acceleration:<\/p>\n\n<pre><code>classifier.fit(X_train, y_train, batch_size = 10, epochs = 100, verbose=1)\n<\/code><\/pre>\n\n<p>I tried this:<\/p>\n\n<pre><code>with tensorflow.device('\/gpu:0'):\n    classifier.fit(X_train, y_train, batch_size = 10, epochs = 100)\n<\/code><\/pre>\n\n<p>With the same result. \nI don't know how to tell if Keras is using GPU except by speed and obvious CPU usage.<\/p>\n\n<p>I also ran this example from the tensorflow documentation and in my terminal I can clearly see that it uses the GPU. It runs much quicker than the keras example above.\n    import tensorflow\n    # Creates a graph.\n    a = tensorflow.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n    b = tensorflow.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n    c = tensorflow.matmul(a, b)\n    # Creates a session with log_device_placement set to True.\n    sess = tensorflow.Session(config=tensorflow.ConfigProto(log_device_placement=True))\n    # Runs the op.\n    print(sess.run(c))<\/p>\n\n<p>I would greatly appreciate your kind help in finding out why Keras can't see my GPU<\/p>\n\n<p>I use python 3.6.5, tensorflow-gpu 1.11.0 (tensorflow not installed), keras 2.2.4.\nI need to mention that I had to fiddle a while to get tensorflow to use the GPU and I still don't know why it suddenly did, but it does so consistently now. My assumption was that Keras would automatically inherit this.<\/p>\n\n<p>A.<\/p>\n","answers":[{"AnswerId":"52936896","CreationDate":"2018-10-22T19:56:54.580","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You could try removing your keras and install keras-gpu instead (available in anaconda, maybe in pip too)<\/p>\n\n<p>If you want to be sure to use <code>with tensorflow.device('\/gpu:0'):<\/code>, use it \"when defining the model\":<\/p>\n\n<pre><code>with tensorflow.device('\/gpu:0'):\n\n    #blablablabla - layers for functional API model\n\n    classifier = Model(inputs, outputs) #or classifier = Sequential()\n\n    #blablabla - layers for sequential model \n<\/code><\/pre>\n"},{"AnswerId":"52938699","CreationDate":"2018-10-22T22:43:58.557","ParentId":null,"OwnerUserId":"10541724","Title":null,"Body":"<p>I am not entirely sure anymore of my originally stated problem. I think Keras was indeed using the GPU, but that I had a significant bottleneck between CPU and GPU. When I increased the batch size, things ran significantly faster (for each epoch), which doesn't make much sense but seems to indicate I have a bottleneck elsewhere.\nI have no idea how to debug this though<\/p>\n"}]}
+{"QuestionId":52934204,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><google-colaboratory>","CreationDate":"2018-10-22T16:53:33.983","AcceptedAnswerId":null,"OwnerUserId":7458186.0,"Title":"Google Colab not making use of its GPU","Body":"<p>I'm using keras with tensorflow on google colab with hardware accelerator set to 'GPU'. I checked and GPU is configured.<\/p>\n\n<pre><code>Found GPU at: \/device:GPU:0\n<\/code><\/pre>\n\n<p>But still, when I run the code, it's taking a lot of time and GPU usage is also close to none and CPU is being used most of the time.\nCan anyone tell what am I doing wrong here?<\/p>\n","answers":[]}
+{"QuestionId":52934764,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-10-22T17:32:39.283","AcceptedAnswerId":null,"OwnerUserId":2968016.0,"Title":"Keras: Adaptive Max Pooling","Body":"<p>I'm working on the CNN, I have the images with different sizes and aspect ratio, resizing and center cropping the images in pre-processing step may loose some valuable information (specially from edges - right,left,up and down).\nI think, the better way is to apply the adaptive max pooling (giving the targeted output dimensions) in order to resize the images before we pass it to our final layer. \nIn keras, I can add a simple max pooling layer, but is there a way in keras to tell\/bound the max pooling layer the output size? e.g: if an image is of 800x1520 but the required output is 200x200, the max pooling applies and the image resized to 200x200<\/p>\n","answers":[{"AnswerId":"52935730","CreationDate":"2018-10-22T18:36:20.837","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>There is no \"adaptive pooling layer\" in Keras, but there is the family of <a href=\"https:\/\/keras.io\/layers\/pooling\/#globalmaxpooling1d\" rel=\"noreferrer\">GlobalMaxPooling<\/a> layers. They can deal with undefined input shapes (i.e. one dimension can be <code>None<\/code>), but always have the same output shape.<\/p>\n\n<p>However, note that within a single batch, all inputs need to have exactly the same dimension. If all your images are of different size, that means that you are restricted to a batch size of 1, which will make training and inference painfully slow.<\/p>\n\n<p>I suggest that you first try to see how well your model works when you resize all images to the same size, or if you pad the smaller images such that they reach the same dimensions as your largest image. Which of the two options works better obviously depends a lot on the images and your feature detection architecture, but due to the restriction mentioned above, working with variable input sizes should be the last option.<\/p>\n"}]}
+{"QuestionId":52934795,"AnswerCount":2,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-10-22T17:34:42.967","AcceptedAnswerId":52937580.0,"OwnerUserId":1185242.0,"Title":"What is difference frozen_inference_graph.pb and saved_model.pb?","Body":"<p>I have a trained model (Faster R-CNN) which I exported using <code>export_inference_graph.py<\/code> to use for inference. I'm trying to understand the difference between the created <code>frozen_inference_graph.pb<\/code> and <code>saved_model.pb<\/code> and also <code>model.ckpt*<\/code> files. I've also seen <code>.pbtxt<\/code> representations.<\/p>\n\n<p>I tried reading through this but couldn't really find the answers: <a href=\"https:\/\/www.tensorflow.org\/extend\/tool_developers\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/extend\/tool_developers\/<\/a><\/p>\n\n<p>What do each of these files contain?\nWhich ones can be converted to which other ones?\nWhat is the ideal purpose of each?<\/p>\n","answers":[{"AnswerId":"52937580","CreationDate":"2018-10-22T20:50:02.403","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>frozen_inference_graph.pb, is a frozen graph that cannot be trained anymore, it defines the graphdef and is actually a serialized graph and can be loaded with this code:<\/p>\n\n<pre><code>def load_graph(frozen_graph_filename):\n    with tf.gfile.GFile(frozen_graph_filename, \"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n        return graph_def\ntf.import_graph_def(load_graph(\"frozen_inference_graph.pb\"))\n<\/code><\/pre>\n\n<p>the saved model is a model generated by tf.saved_model.builder and is has to be imported into a session, this file contains the full graph with all training weights (just like the frozen graph) but here can be trained upon, and this one is not serialized and needs to be loaded by this snippet. The [] are tagconstants which can be read by the <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#cli_to_inspect_and_execute_savedmodel\" rel=\"noreferrer\">saved_model_cli<\/a>. This model is also often served to predict on, like google ml engine par example:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    tf.saved_model.loader.load(sess, [], \"foldername to saved_model.pb, only folder\")\n<\/code><\/pre>\n\n<p>model.ckpt files are checkpoints, generated during training, this is used to resume training or to have a back up when something goes wrong after along training. If you have a saved model and a frozen graph, then you can ignore this.<\/p>\n\n<p>.pbtxt files are basically the same as previous discussed models, but then human readable, not binary. These can be ignored as well.<\/p>\n\n<p>To answer your conversion question:\nsaved models can be transformed into a frozen graph and vice versa, although a saved_model extracted from a frozen graph is also no trainable, but the way it is stored is in saved model format. Checkpoints can be read in and loaded into a session, and there you can build a saved model from them.<\/p>\n\n<p>Hope I helped, any questions, ask away!<\/p>\n"},{"AnswerId":"54924205","CreationDate":"2019-02-28T11:04:12.213","ParentId":null,"OwnerUserId":"1147398","Title":null,"Body":"<p>Like to add, frozen_graph.pb includes two things:\n1. Graph definition\n2. Trained parameters<\/p>\n\n<p>Whereas save_model.pb, just have graph definition.<\/p>\n\n<p>That's why if you check the size of both the .pb files, frozen_graph.pb always be larger in size.<\/p>\n"}]}
+{"QuestionId":52936132,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-22T19:03:02.517","AcceptedAnswerId":52936626.0,"OwnerUserId":10542348.0,"Title":"4D input in LSTM layer in Keras","Body":"<p>I have data with a shape of <code>(10000, 20, 15, 4)<\/code> where <code>num samples = 10000<\/code>, <code>num series in time = 20<\/code>, <code>height = 15<\/code>, <code>weight = 4<\/code>. So I have table <code>15x4<\/code> which is distributed over time. Here is the model I want to train it over this data:<\/p>\n\n<pre><code>...\nmodel.add((LSTM(nums-1,return_sequences=True,input_shape=(20,15,4), activation='relu')))\nmodel.add((LSTM(nums-1,return_sequences=False,input_shape=(20,15,4), activation='tanh')))\nmodel.add(Dense(15,activation='relu'))\n...\n<\/code><\/pre>\n\n<p>However, I get the following error:<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, \nfound ndim=4\n<\/code><\/pre>\n\n<p>How do I define a LSTM layer with 4D input shape?<\/p>\n","answers":[{"AnswerId":"52936626","CreationDate":"2018-10-22T19:37:47.127","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>LSTM layer accepts a 3D array as input which has a shape of <code>(n_sample, n_timesteps, n_features)<\/code>. Since the features of each timestep in your data is a <code>(15,4)<\/code> array, you need to first flatten them to a feature vector of length 60 and then pass it to your model:<\/p>\n\n<pre><code>X_train = X_train.reshape(10000, 20, -1)\n\n# ...\nmodel.add(LSTM(...,input_shape=(20,15*4), ...)) # modify input_shape accordingly\n<\/code><\/pre>\n\n<p>Alternatively, you can use a <code>Flatten<\/code> layer wrapped in a <code>TimeDistributed<\/code> layer <strong>as the first layer<\/strong> of your model to flatten each timestep:<\/p>\n\n<pre><code>model.add(TimeDistributed(Flatten(input_shape=(15,4))))\n<\/code><\/pre>\n\n<p>Further, note that if each timestep (i.e. array <code>(15, 4)<\/code>) is a feature map where there is a local spatial relationship between its elements, say like an image patch, you can also use <a href=\"https:\/\/keras.io\/layers\/recurrent\/#convlstm2d\" rel=\"nofollow noreferrer\"><code>ConvLSTM2D<\/code><\/a> instead of <code>LSTM<\/code> layer. Otherwise, flattening the timesteps and using LSTM would be fine.<\/p>\n\n<hr>\n\n<p><strong>As a side note:<\/strong> you only need to specify <code>input_shape<\/code> argument on the first layer of the model. Specifying it on other layers would be redundant and will be ignored since their input shape is automatically inferred by Keras.<\/p>\n"}]}
+{"QuestionId":52937055,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network><lstm>","CreationDate":"2018-10-22T20:09:12.180","AcceptedAnswerId":null,"OwnerUserId":10083414.0,"Title":"how to implement ctc loss function for Sequential model in keras","Body":"<p>Hi guys i have cnn+lstm keras model and i want to use ctc loss function,there are bunch off solutions around but all of them are very confusing any help will be appreciated.\nhere is my model:<\/p>\n\n<pre><code>model=Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=(28, 28,3), activation='relu'))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='relu'))\nmodel.add(RepeatVector(25))\nmodel.add(LSTM(128, return_sequences=True, input_shape=(25,32), activation='relu'))\nmodel.add(LSTM(256))\nmodel.add(Dropout(0.25))\nmodel.add(Dense(37))\nmodel.compile(loss='ctc_loss', optimizer='rmsprop')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52937195,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-22T20:19:55.587","AcceptedAnswerId":null,"OwnerUserId":1544149.0,"Title":"Loading images and eval() in tensorflow are super slow","Body":"<pre><code>X = []\nfilelist = gfile.ListDirectory(path_imgs)\n\nfor filename in filelist:\n    path_filename = path_imgs + filename\n        image_file = file_io.FileIO(path_filename,'rb')\n        image_raw = image_file.read()\n\n        img = tf.image.decode_image(image_raw)\n        img = tf.image.convert_image_dtype(img, tf.float32)\n        img = tf.image.resize_image_with_pad(img, img_size, img_size, method=1).eval(session=tf.Session())\n        X.append(img)\n\n\nimgs = np.array(X) \n<\/code><\/pre>\n\n<p><br>\nTried some things with session, but didn't work. Probably it should be handled differently, but I don't know how to do it. Any ideas?<\/p>\n\n<p><br>\nEDIT:<\/p>\n\n<p>Yes, I want to train ANN to segment objects in images.<\/p>\n\n<p>There are folders with images and their masks. Size is 1000s, and could be 10s of 1000s.<\/p>\n\n<p>I need single numpy array of images, which will be saved, and used later as dataset for model training.<\/p>\n","answers":[{"AnswerId":"52940256","CreationDate":"2018-10-23T02:21:54.877","ParentId":null,"OwnerUserId":"10535783","Title":null,"Body":"<pre><code>def getImageData(fileNameList):\n    imageData=[]\n    for fn in fileNameList:        \n        testImage = Image.open(fn)\n        testImage.show() \n        imageData.append(np.array(testImage))\n    return np.array(imageData,dtype=np.float32)\n\nimageFn=(\"dog.png\",)\nimageData=getImageData(imageFn)\n<\/code><\/pre>\n\n<p>you must import something:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom PIL import Image    \nimport numpy as np\n<\/code><\/pre>\n"}]}
+{"QuestionId":52937256,"AnswerCount":1,"Tags":"<python><node.js><keras>","CreationDate":"2018-10-22T20:24:18.627","AcceptedAnswerId":null,"OwnerUserId":6273108.0,"Title":"How to keep a Keras model loaded into memory and use it when needed?","Body":"<p>I was reading a <a href=\"https:\/\/blog.keras.io\/building-a-simple-keras-deep-learning-rest-api.html\" rel=\"nofollow noreferrer\">Keras blog<\/a> teaching how to create a simple image classifier restful API with Flask. I was wondering how to achieve the same approach of loading model in other web frameworks that do not use python.\nIn the code below the model is loaded into memory just before the server starts and it runs until the server is alive:<\/p>\n\n<pre><code># if this is the main thread of execution first load the model and\n# then start the server\nif __name__ == \"__main__\":\n    print((\"* Loading Keras model and Flask starting server...\"\n        \"please wait until server has fully started\"))\n    load_model()\n    app.run()\n<\/code><\/pre>\n\n<p>I'm familiar with <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/model_persistence.html\" rel=\"nofollow noreferrer\">Pickle<\/a> and I know how to run python code in other web frameworks (such as <a href=\"https:\/\/www.npmjs.com\/package\/python-shell\" rel=\"nofollow noreferrer\">python-shell<\/a> of Node.js). pickled models are built once and can be loaded each time they're needed. but I'm looking to achieve the same thing as the tutorial suggests which is loading it only once, and using it multiple times. Is creating a separate python server app that serves the loaded model to Node.js requests a good idea?<\/p>\n","answers":[{"AnswerId":"52937314","CreationDate":"2018-10-22T20:29:13.547","ParentId":null,"OwnerUserId":"2887760","Title":null,"Body":"<p>You can load a model in Keras using <code>load_model<\/code> and pass in a path:<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('model.hd5')\n<\/code><\/pre>\n\n<p>I have created a Flask API that loads a Keras model, you can take a look here if it helps:\n<a href=\"https:\/\/github.com\/Ares513\/DetectingTrollsApi\/blob\/master\/api.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Ares513\/DetectingTrollsApi\/blob\/master\/api.py<\/a><\/p>\n"}]}
+{"QuestionId":52937480,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-10-22T20:41:51.933","AcceptedAnswerId":null,"OwnerUserId":8307101.0,"Title":"How to drop the nodes in specified positions of the layer in PyTorch","Body":"<p>I\u2019m trying to implement a fully connected network to do a \n    classification task, while my training data contains some missing \n    values and the positions of missingness are not fixed. <\/p>\n\n<p>For example:<\/p>\n\n<pre><code>[x1, x2, NaN, x4, x5]  [y]\n[x1, x2, x3, x4, NaN]  [y]\n[NaN, x2, x3, NaN, x5] [y]\n<\/code><\/pre>\n\n<p>I have tried to implement some methods to fill in the missing values \nbut it will introduce bias after all.<\/p>\n\n<p>To solve the bias problem, I\u2019m planning to keep the missingness but \ndrop the corresponding nodes that the missing values appear in the \ninput layer so that the <code>NaNs<\/code> will not be fed into the network and cause \nerrors, and the following layers of the network stay the same. I wonder \nif there are specific methods to achieve it in PyTorch, either by:<\/p>\n\n<ol>\n<li>Drop the node in a specified position of the layer, or<\/li>\n<li>Enable variable input size of the neural network.<\/li>\n<\/ol>\n\n<p>However, these are just two possible solutions I can come up with, I \nwould be appreciated if you could share with me other useful ways.<\/p>\n","answers":[]}
+{"QuestionId":52937643,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><backpropagation>","CreationDate":"2018-10-22T20:55:36.767","AcceptedAnswerId":52939752.0,"OwnerUserId":4789373.0,"Title":"Why `tf.train.Optimizer().compute_gradients(loss)` also returns variables not in the subgraph of `loss`?","Body":"<p>I'm manually collecting the gradients statistics of the <strong>Multi-Tasking<\/strong> model, the <strong>graph<\/strong> of which looks schematically like this:<\/p>\n\n<pre><code>input -&gt; [body_var1 ... body_varN] --&gt; [task1_var1 ... task1_varM] &lt;-- loss_1\n                                   \\-&gt; [task2_var1 ... task2_varM] &lt;-- loss_2\n<\/code><\/pre>\n\n<p>I'm defining a separate optimizer for each loss as follows (the actual code is much complicated, the following is simplified for this question):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code># for simplicity, just demonstrate the case with the 1st task\ntask_index = 1\n\n# here we define the optimizer (create an instance in graph)\nloss = losses[task_index]\noptimizer = tf.train.GradientDescentOptimizer()\ngrads_and_vars = optimizer.compute_gradients(loss)\n\n# now let's see what it returns\nfor g, v in grads_and_vars:\n    print('  grad:', g, ', var:', v)\n<\/code><\/pre>\n\n<p>So, the code above clearly creates a separated optimizer only for the branch of task 1, then we create the gradient computation ops with <code>optimizer.compute_gradients(loss)<\/code> and print the vars to which we apply the gradients to.<\/p>\n\n<p><strong>Expected results<\/strong>:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>grad: body_var1_grad, var: body_var1    # \\\n...                                     # --&gt; body vars and gradients\ngrad: body_varN_grad, var: body_varN    # \/\ngrad: task1_var1_grad, var: task1_var1  # \\\n...                                     # --&gt; task 1 vars and gradients\ngrad: task1_var1_grad, var: task1_var1  # \/\n<\/code><\/pre>\n\n<p>So I'm expecting that the optimizer only contains gradient computing ops for the branch it was applied on (i.e. the branch for 1st task)<\/p>\n\n<p><strong>Actual results<\/strong><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>grad: body_var1_grad, var: body_var1    # \\\n...                                     # --&gt; body vars and gradients\ngrad: body_varN_grad, var: body_varN    # \/\ngrad: task1_var1_grad, var: task1_var1  # \\\n...                                     # --&gt; task 1 vars and gradients\ngrad: task1_var1_grad, var: task1_var1  # \/\ngrad: None, var: task2_var1             # \\\n...                                     # --&gt; task 2 vars, with None gradients\ngrad: None, var: task2_var1             # \/\n<\/code><\/pre>\n\n<p>So it looks like <code>optimizer.compute_gradients(loss)<\/code> captures not only the sub-graph that outputs to <code>loss<\/code> (which can be extracted using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/graph_util\/extract_sub_graph\" rel=\"nofollow noreferrer\"><code>tf.graph_util.extract_sub_graph<\/code><\/a>), but also all trainable variables that are connected to <code>loss<\/code> without creating a gradient variable for them (so the returned gradient variables are <code>None<\/code>).<\/p>\n\n<blockquote>\n  <p><strong>Question:<\/strong> is such behavior normal?<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"52939752","CreationDate":"2018-10-23T01:14:40.837","ParentId":null,"OwnerUserId":"9535747","Title":null,"Body":"<p>Yes, it is, because <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/GradientDescentOptimizer#compute_gradients\" rel=\"nofollow noreferrer\">compute_gradients()<\/a> computes gradients of <code>loss<\/code> with respect to a list of <code>tf.Variable<\/code> objects which is passed to the <code>var_list<\/code> parameter. If <code>var_list<\/code> is not provided, the function calculates gradients with respect to all variables from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GraphKeys\" rel=\"nofollow noreferrer\">GraphKeys.TRAINABLE_VARIABLES<\/a> collection. Also, if <code>loss<\/code> does not depend on certain variables, the gradients of <code>loss<\/code> with respect to those variables are not defined, i.e. <code>None<\/code> is returned. Based on the code you provided, this seems to be the case.<\/p>\n\n<p>If you want the <code>optimizer<\/code> to calculate gradients with respect to certain variables only, you should make a list of such variables and pass it to the <code>var_list<\/code> parameter of <code>compute_gradients()<\/code>.<\/p>\n"}]}
+{"QuestionId":52937685,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection><google-colaboratory><object-detection-api>","CreationDate":"2018-10-22T20:59:11.593","AcceptedAnswerId":null,"OwnerUserId":4794648.0,"Title":"google colab setting a '^C' in the proccess","Body":"<p>I'm running this <a href=\"https:\/\/drive.google.com\/open?id=1gZTADeRnAX4li-yK-\" rel=\"nofollow noreferrer\">code<\/a> that i got from this <a href=\"https:\/\/www.youtube.com\/watch?v=COlbP62-B-U&amp;list=PLQVvvaa0QuDcNK5GeCQnxYnSSaar2tpku\" rel=\"nofollow noreferrer\">tutorial<\/a>\nI'm trying running the tensorflow object detection api, all code work well, if you run all calls, all cell will works well, and in the end, my images are classified.<\/p>\n\n<p>Buuut have 1 cell that dont work well, it's work, but doesn't like it must work.<\/p>\n\n<p>When i will train my model with <code>!python legacy\/train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_mobilenet_v1_pets.config<\/code>\nit start the tensorflow and start the training, buuut it only run 3 steps, 4 steps, some times 20,21,23 steps and in the end, the google colab set a <code>^C<\/code> in the process<\/p>\n\n<p>I never can finish my training because the google colab close my process, some one know whatsap happening?<\/p>\n\n<p>I already try use GPU and TPU instances.<\/p>\n\n<pre><code>[...]\nINFO:tensorflow:Restoring parameters from training\/model.ckpt-0\nI1022 20:41:48.368024 139794549495680 tf_logging.py:115] Restoring parameters from training\/model.ckpt-0\nINFO:tensorflow:Running local_init_op.\nI1022 20:41:52.779153 139794549495680 tf_logging.py:115] Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nI1022 20:41:52.997912 139794549495680 tf_logging.py:115] Done running local_init_op.\nINFO:tensorflow:Starting Session.\nI1022 20:41:59.072830 139794549495680 tf_logging.py:115] Starting Session.\nINFO:tensorflow:Saving checkpoint to path training\/model.ckpt\nI1022 20:41:59.245162 139793493063424 tf_logging.py:115] Saving checkpoint to path training\/model.ckpt\nINFO:tensorflow:Starting Queues.\nI1022 20:41:59.252097 139794549495680 tf_logging.py:115] Starting Queues.\nINFO:tensorflow:global_step\/sec: 0\nI1022 20:42:10.151180 139793484670720 tf_logging.py:159] global_step\/sec: 0\nINFO:tensorflow:Recording summary at step 0.\nI1022 20:42:16.119055 139793476278016 tf_logging.py:115] Recording summary at step 0.\nINFO:tensorflow:global step 1: loss = 14.0911 (28.770 sec\/step)\nI1022 20:42:28.496783 139794549495680 tf_logging.py:115] global step 1: loss = 14.0911 (28.770 sec\/step)\nINFO:tensorflow:global step 2: loss = 12.4958 (10.529 sec\/step)\nI1022 20:42:39.334129 139794549495680 tf_logging.py:115] global step 2: loss = 12.4958 (10.529 sec\/step)\nINFO:tensorflow:global step 3: loss = 11.6073 (8.267 sec\/step)\nI1022 20:42:47.601801 139794549495680 tf_logging.py:115] global step 3: loss = 11.6073 (8.267 sec\/step)\n^C\n<\/code><\/pre>\n","answers":[{"AnswerId":"57263727","CreationDate":"2019-07-30T02:56:03.210","ParentId":null,"OwnerUserId":"10037689","Title":null,"Body":"<p>You can use the following GitHub <a href=\"https:\/\/github.com\/jiteshm17\/tensorflow_object_detection_colab\" rel=\"nofollow noreferrer\">repo<\/a> for training a tensorflow object detection model on Google Colab. It has a readme, a .ipynb file, a model config file and a sample label_map file. Please do let me know if you face any issues. Hope this helps<\/p>\n"}]}
+{"QuestionId":52937848,"AnswerCount":1,"Tags":"<python><python-3.x><deep-learning><pytorch><torchvision>","CreationDate":"2018-10-22T21:14:10.993","AcceptedAnswerId":null,"OwnerUserId":10537618.0,"Title":"How to use checkpoint model file in pytorch to test the CIFAR-10 dataset?","Body":"<pre><code>model = SqueezeNext()\nmodel = model.to(device)\n\ndef load_checkpoint(model, optimizer, losslogger, filename='SqNxt_23_1x_Cifar.ckpt'):\n# Note: Input model &amp; optimizer should be pre-defined.  This routine only updates their states.\nstart_epoch = 0\nif os.path.isfile(filename):\n    print(\"=&gt; loading checkpoint '{}'\".format(filename))\n    checkpoint = torch.load(filename)\n    start_epoch = checkpoint['epoch']\n    model.load_state_dict(checkpoint['state_dict'])\n    optimizer.load_state_dict(checkpoint['optimizer'])\n    losslogger = checkpoint['losslogger']\n    print(\"=&gt; loaded checkpoint '{}' (epoch {})\"\n              .format(filename, checkpoint['epoch']))\nelse:\n    print(\"=&gt; no checkpoint found at '{}'\".format(filename))\n\n\nreturn model, optimizer, start_epoch, losslogger\n\nmodel, optimizer, start_epoch, losslogger = load_checkpoint(model, optimizer, losslogger)\n<\/code><\/pre>\n\n<hr>\n\n<blockquote>\n  <p>TypeError: Traceback (most recent call last)\n   in ()\n       41 test_loader   = torch.utils.data.DataLoader(test_dataset, batch_size=80, num_workers=8, shuffle=False)\n       42 \n  ---> 43 model = SqueezeNext()\n       44 model = model.to(device)\n       45 def load_checkpoint(model, optimizer, losslogger, filename='SqNxt_23_1x_Cifar.ckpt'): TypeError: <strong>init<\/strong>() missing 3\n  required positional arguments: 'width_x', 'blocks', and 'num_classes'<\/p>\n<\/blockquote>\n\n<p>I think I am not implementing this in right manner!!    <\/p>\n","answers":[{"AnswerId":"58840112","CreationDate":"2019-11-13T15:13:15.910","ParentId":null,"OwnerUserId":"9717778","Title":null,"Body":"<p>Your error isn't from your checkpoint function. If we look at the traceback:<\/p>\n\n<pre><code>&gt; TypeError: Traceback (most recent call last)\n&gt; &lt;ipython-input-51-94c8be648862&gt; in &lt;module&gt;()\n&gt;      41 test_loader   = torch.utils.data.DataLoader(test_dataset, batch_size=80, num_workers=8, shuffle=False)\n&gt;      42 \n&gt; ---&gt; 43 model = SqueezeNext()\n&gt;      44 model = model.to(device)\n&gt;      45 def load_checkpoint(model, optimizer, losslogger, filename='SqNxt_23_1x_Cifar.ckpt'): TypeError: __init__() missing 3\n&gt; required positional arguments: 'width_x', 'blocks', and 'num_classes'\n<\/code><\/pre>\n\n<p>The line that we're being told is breaking the line 43:<\/p>\n\n<pre><code>&gt; ---&gt; 43 model = SqueezeNext()\n<\/code><\/pre>\n\n<p>And the error is:<\/p>\n\n<pre><code>&gt; required positional arguments: 'width_x', 'blocks', and 'num_classes'\n<\/code><\/pre>\n\n<p>I'm assuming you're using <a href=\"https:\/\/github.com\/luuuyi\/SqueezeNext.PyTorch\/blob\/master\/SqueezeNext.ipynb\" rel=\"nofollow noreferrer\">this implementation<\/a> of SqueezeNext, but whichever implementation you're using, you're not passing all the arguments needed to initialise the model. You'll want to change your code to something like:<\/p>\n\n<pre><code>model = SqueezeNext(width_x=1.0, blocks=[6, 6, 8, 1], num_classes=10)\n<\/code><\/pre>\n\n<p>If you're not using that implementation, you'll need to find the source code for the <code>SqueezeNext<\/code> model, and see what arguments the <code>__init__<\/code> function requires. You can try this:<\/p>\n\n<pre><code>import inspect\n\ninspect.signature(SqueezeNext.__init__)\n<\/code><\/pre>\n\n<p>Which should give you the signature. <\/p>\n"}]}
+{"QuestionId":52938457,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-22T22:13:22.940","AcceptedAnswerId":null,"OwnerUserId":3129625.0,"Title":"Tensorflow debugging","Body":"<pre><code>        y1 = tf.stack([55, 0, 55])\n        y2 = tf.stack([11 22 44])\n        y3 = tf.constant([0., 0., 1.])\n        x = tf.stack([y1, y2, y3])\n        x1 = tf.reshape(x, [3, 3])\n<\/code><\/pre>\n\n<p>Is there a way to print or obtain the above tensor <code>x1<\/code> or <code>x<\/code> in tensorflow. The <code>tf.print<\/code> and <code>sess.eval(x1)<\/code>? It seems to not print the stacked frame.<\/p>\n\n<p>Tried the <code>tfdg<\/code> too<\/p>\n\n<pre><code>sess = tf.Session()\n        sess = tf_debug.TensorBoardDebugWrapperSession(sess, \"Host:7000\")\n        sess.run(x1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52940303","CreationDate":"2018-10-23T02:28:05.003","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>When doing <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntf.reset_default_graph()\nwith tf.Session() as sess:\n    y1 = tf.stack([55, 0, 55])\n    y2 = tf.stack([11, 22, 44])\n    y3 = tf.constant([0, 0, 1])\n\n    x_tmp = tf.stack([y1, y2, y3])\n    x1 = tf.reshape(x, [3, 3])\n\n    x_res, x1_res = sess.run([x, x1])\n\n\n    print(x_res)\n    print(x1_res)\n<\/code><\/pre>\n\n<p>It seems to print out <\/p>\n\n<pre><code>[[55  0 55]\n [11 22 44]\n [ 0  0  1]]\n[[55  0 55]\n [11 22 44]\n [ 0  0  1]]\n<\/code><\/pre>\n\n<p>just fine? Perhaps I'm misunderstanding the question<\/p>\n"}]}
+{"QuestionId":52938714,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><deep-learning>","CreationDate":"2018-10-22T22:45:56.810","AcceptedAnswerId":null,"OwnerUserId":8075683.0,"Title":"tensorflow feed list feature (multi-hot) to tf.estimator","Body":"<p>Some of the feature columns' data type is <code>list<\/code>. And their length can be different. I want to encode this column as a multi-hot categorical feature and feed it to <code>tf.estimator<\/code>. I tried the following but the error <code>Unable to get element as bytes<\/code> showed. I think this is a common practice in deep learning especially recommendation systems e.g. Deep &amp; Wide model. I found a relevant question <a href=\"https:\/\/stackoverflow.com\/questions\/48697799\/tensorflow-feature-column-for-variable-list-of-values\">here<\/a> but it doesn't show how to feed to the estimator.<\/p>\n\n<pre><code>import pandas as pd\nimport tensorflow as tf\n\nOUTDIR = \".\/data\"\n\ndata = {\"x\": [[\"a\", \"c\"], [\"a\", \"b\"], [\"b\", \"c\"]], \"y\": [\"x\", \"y\", \"z\"]}\ndf = pd.DataFrame(data)\n\nY = df[\"y\"]\nX = df.drop(\"y\", axis=1)\n\nindicator_features = [\n    tf.feature_column.indicator_column(\n        categorical_column=tf.feature_column.categorical_column_with_vocabulary_list(\n            key=\"x\", vocabulary_list=[\"a\", \"b\", \"c\"]\n        )\n    )\n]\n\nmodel = tf.estimator.LinearClassifier(\n    feature_columns=indicator_features, model_dir=OUTDIR\n)\n\ntraining_input_fn = tf.estimator.inputs.pandas_input_fn(\n    x=X, y=Y, batch_size=64, shuffle=True, num_epochs=None\n)\n\nmodel.train(input_fn=training_input_fn)\n<\/code><\/pre>\n\n<p>The following error:<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Using default config. INFO:tensorflow:Using config:\n  {'_model_dir': 'testalg', '_tf_random_seed': None,\n  '_save_summary_steps': 100, '_save_checkpoints_steps': None,\n  '_save_checkpoints_secs': 600, '_session_config': None,\n  '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000,\n  '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn':\n  None, '_service': None, '_cluster_spec':\n  , '_task_type': 'worker', '_task_id': 0,\n  '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '',\n  '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}\n  INFO:tensorflow:Calling model_fn. INFO:tensorflow:Done calling\n  model_fn. INFO:tensorflow:Create CheckpointSaverHook.\n  INFO:tensorflow:Graph was finalized. INFO:tensorflow:Running\n  local_init_op. INFO:tensorflow:Done running local_init_op.\n  INFO:tensorflow:Error reported to Coordinator: , Unable to\n  get element as bytes. INFO:tensorflow:Saving checkpoints for 0 into\n  testalg\/model.ckpt.\n  ------------------------------------------------------- InternalError         Traceback (most recent call last)\n  \/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _do_call(self, fn, *args)    1321     try:\n  -> 1322       return fn(*args)    1323     except errors.OpError as e:<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata) \n  1306       return self._call_tf_sessionrun(\n  -> 1307           options, feed_dict, fetch_list, target_list, run_metadata)    1308 <\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _call_tf_sessionrun(self, options, feed_dict, fetch_list,\n  target_list, run_metadata)    1408           self._session, options,\n  feed_dict, fetch_list, target_list,\n  -> 1409           run_metadata)    1410     else:<\/p>\n  \n  <p>InternalError: Unable to get element as bytes.<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>InternalError         Traceback (most recent call last)\n   in ()\n       44 \n       45 \n  ---> 46 model.train(input_fn=training_input_fn)<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\n  in train(self, input_fn, hooks, steps, max_steps, saving_listeners)\n      364 \n      365       saving_listeners = _check_listeners_type(saving_listeners)\n  --> 366       loss = self._train_model(input_fn, hooks, saving_listeners)\n      367       logging.info('Loss for final step: %s.', loss)\n      368       return self<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\n  in _train_model(self, input_fn, hooks, saving_listeners)    1117<br>\n  return self._train_model_distributed(input_fn, hooks,\n  saving_listeners)    1118     else:\n  -> 1119       return self._train_model_default(input_fn, hooks, saving_listeners)    1120     1121   def _train_model_default(self,\n  input_fn, hooks, saving_listeners):<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\n  in _train_model_default(self, input_fn, hooks, saving_listeners)<br>\n  1133       return self._train_with_estimator_spec(estimator_spec,\n  worker_hooks,    1134<br>\n  hooks, global_step_tensor,\n  -> 1135                                              saving_listeners)    1136     1137   def _train_model_distributed(self, input_fn, hooks,\n  saving_listeners):<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\n  in _train_with_estimator_spec(self, estimator_spec, worker_hooks,\n  hooks, global_step_tensor, saving_listeners)    1334       loss = None\n  1335       while not mon_sess.should_stop():\n  -> 1336         _, loss = mon_sess.run([estimator_spec.train_op, estimator_spec.loss])    1337     return loss    1338 <\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\n  in <strong>exit<\/strong>(self, exception_type, exception_value, traceback)\n      687     if exception_type in [errors.OutOfRangeError, StopIteration]:\n      688       exception_type = None\n  --> 689     self._close_internal(exception_type)\n      690     # <strong>exit<\/strong> should return True to suppress an exception.\n      691     return exception_type is None<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\n  in _close_internal(self, exception_type)\n      724         if self._sess is None:\n      725           raise RuntimeError('Session is already closed.')\n  --> 726         self._sess.close()\n      727       finally:\n      728         self._sess = None<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\n  in close(self)\n      972     if self._sess:\n      973       try:\n  --> 974         self._sess.close()\n      975       except _PREEMPTION_ERRORS:\n      976         pass<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py\n  in close(self)    1116       self._coord.join(    1117<br>\n  stop_grace_period_secs=self._stop_grace_period_secs,\n  -> 1118           ignore_live_threads=True)    1119     finally:    1120       try:<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/coordinator.py\n  in join(self, threads, stop_grace_period_secs, ignore_live_threads)\n      387       self._registered_threads = set()\n      388       if self._exc_info_to_raise:\n  --> 389         six.reraise(*self._exc_info_to_raise)\n      390       elif stragglers:\n      391         if ignore_live_threads:<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/six.py in\n  reraise(tp, value, tb)\n      683             value = tp()\n      684         if value.<strong>traceback<\/strong> is not tb:\n  --> 685             raise value.with_traceback(tb)\n      686         raise value\n      687 <\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/inputs\/queues\/feeding_queue_runner.py\n  in _run(self, sess, enqueue_op, feed_fn, coord)\n       92         try:\n       93           feed_dict = None if feed_fn is None else feed_fn()\n  ---> 94           sess.run(enqueue_op, feed_dict=feed_dict)\n       95         except (errors.OutOfRangeError, errors.CancelledError):\n       96           # This exception indicates that a queue was closed.<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in run(self, fetches, feed_dict, options, run_metadata)\n      898     try:\n      899       result = self._run(None, fetches, feed_dict, options_ptr,\n  --> 900                          run_metadata_ptr)\n      901       if run_metadata:\n      902         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _run(self, handle, fetches, feed_dict, options, run_metadata)<br>\n  1133     if final_fetches or final_targets or (handle and\n  feed_dict_tensor):    1134       results = self._do_run(handle,\n  final_targets, final_fetches,\n  -> 1135                              feed_dict_tensor, options, run_metadata)    1136     else:    1137       results = []<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _do_run(self, handle, target_list, fetch_list, feed_dict, options,\n  run_metadata)    1314     if handle is None:    1315       return\n  self._do_call(_run_fn, feeds, fetches, targets, options,\n  -> 1316                            run_metadata)    1317     else:    1318       return self._do_call(_prun_fn, handle, feeds, fetches)<\/p>\n  \n  <p>\/home\/yinan.li1\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\n  in _do_call(self, fn, *args)    1333         except KeyError:    1334 \n  pass\n  -> 1335       raise type(e)(node_def, op, message)    1336     1337   def _extend_graph(self):<\/p>\n  \n  <p>InternalError: Unable to get element as bytes.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"55173339","CreationDate":"2019-03-14T23:07:48.533","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>I think one of the problems in your case is that the type of column in pandas is actually object not string. If you converted that to separate string columns you would get rid of this error. Keep in mind that <code>The basic TensorFlow tf.string dtype allows you to build tensors of byte strings.<\/code> and as you are storing object in this column not a string you get an error.<\/p>\n\n<p>The code below will get overcome the error you get above, but it will not solve your problem completely. The variable lenght of lists would have to be dealt with by padding or the list or something similar as again <code>indicator_column<\/code> might have problems dealing with missing values. <\/p>\n\n<pre><code>X2= pd.DataFrame(X['x'].values.tolist(), columns=['x1','x2'])\n\nfeat1 = tf.feature_column.categorical_column_with_vocabulary_list(\n            key=\"x1\", vocabulary_list=[\"a\", \"b\", \"c\"]\n        )\nfeat2 = tf.feature_column.categorical_column_with_vocabulary_list(\n            key=\"x2\", vocabulary_list=[\"a\", \"b\", \"c\"]\n        )\nindicator_features = [\n    tf.feature_column.indicator_column(\n        categorical_column=feat1\n    ),tf.feature_column.indicator_column(\n        categorical_column=feat2\n    )\n]\n\ntraining_input_fn = tf.estimator.inputs.pandas_input_fn(\n    x=X2, y=Y, batch_size=64, shuffle=True, num_epochs=None\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52938836,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-22T23:04:28.640","AcceptedAnswerId":null,"OwnerUserId":8262057.0,"Title":"Shearing image in tensorflow","Body":"<p>I am using <code>tf.keras<\/code> to build my network. And I am  doing all the augmentation in tensor_wise level since my data in <code>tfrecords<\/code> file. Then I needed to do <strong>shearing<\/strong> and <strong>zca<\/strong> for augmentation but couldn't find a proper implementation in tensor flow. And I can't use the <code>DataImageGenerator<\/code> that did both operation I needed because as I said my data doesn't fit in memory and it is in <code>tfrecord<\/code> format. So all my augmentations process should be tesnorwise.<\/p>\n\n<p>@fchollet <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/68\" rel=\"nofollow noreferrer\">here<\/a> suggested a way to use <code>ImgaeDataGenerator<\/code> with large dataset.<\/p>\n\n<p>My first questino is \nif I use @fchollet way, which is basically using <code>X-sample<\/code> of the large data to run the <code>ImageDataGenerator<\/code> then using <code>train_on_batch<\/code> to train the network , how I can feed my validation data to the network.<\/p>\n\n<p>My Second question is there any tensor-wise implementation for <strong>shear<\/strong> and <strong>zca<\/strong> operations. Some people like <a href=\"https:\/\/stackoverflow.com\/questions\/52214953\/tensorflow-is-there-a-way-to-implement-tensor-wise-image-shear-rotation-transl\">here<\/a> suggested using <code>tf.contrib.image.transform<\/code> but couldn't understand how. If some one have the idea on how to do it, I will appreciate that. <\/p>\n\n<p><strong>Update<\/strong>:<\/p>\n\n<p>This is my trial to construct the transformation matrix through ski_image<\/p>\n\n<pre><code>from skimage import io\nfrom skimage import transform as trans\nimport tensor flow as tf \n\ndef augment()\n  afine_tf = trans.AffineTransform(shear=0.2)\n  transform = tf.contrib.image.matrices_to_flat_transforms(tf.linalg.inv(afine_tf.params))\n  transform= tf.cast(transform, tf.float32)\n  image = tf.contrib.image.transform(image, transform)  # Image here is a tensor \n  return image\n\n\ndataset_train = tf.data.TFRecordDataset(training_files, num_parallel_reads=calls)\n\ndataset_train = dataset_train.apply(tf.contrib.data.shuffle_and_repeat(buffer_size=1000+ 4 * batch_size))\ndataset_train = dataset_train.map(decode_train, num_parallel_calls= calls)  \ndataset_train = dataset_train.map(augment,num_parallel_calls=calls )  \ndataset_train = dataset_train.batch(batch_size)    \ndataset_train = dataset_train.prefetch(tf.contrib.data.AUTOTUNE)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52940460","CreationDate":"2018-10-23T02:48:02.190","ParentId":null,"OwnerUserId":"10253771","Title":null,"Body":"<p>I will answer the second question. <\/p>\n\n<p>Today one of my old questions was commented by a user, but the comments have been deleted when I was adding more details on how to use <code>tf.contrib.image.transform<\/code>. I guess it's you, right?<\/p>\n\n<p>So, I have edited my question and added an example, check it <a href=\"https:\/\/stackoverflow.com\/questions\/52214953\/tensorflow-is-there-a-way-to-implement-tensor-wise-image-shear-rotation-transl\/52218068\">here<\/a>.<\/p>\n\n<p><strong>TL;DR:<\/strong><\/p>\n\n<pre><code>def transformImg(imgIn,forward_transform):\n    t = tf.contrib.image.matrices_to_flat_transforms(tf.linalg.inv(forward_transform))\n    # please notice that forward_transform must be a float matrix,\n    # e.g. [[2.0,0,0],[0,1.0,0],[0,0,1]] will work\n    # but [[2,0,0],[0,1,0],[0,0,1]] will not\n    imgOut = tf.contrib.image.transform(imgIn, t, interpolation=\"BILINEAR\",name=None)\n    return imgOut\ndef shear_transform_example(filename,shear_lambda):\n    image_string = tf.read_file(filename)\n    image_decoded = tf.image.decode_jpeg(image_string, channels=3)\n    img = transformImg(image_decoded, [[1.0,shear_lambda,0],[0,1.0,0],[0,0,1.0]])\n    # Notice that this is a shear transformation parallel to the x axis\n    # If you want a y axis version, use this:\n    # img = transformImg(image_decoded, [[1.0,0,0],[shear_lambda,1.0,0],[0,0,1.0]])\n    return img\nimg = shear_transform_example(\"white_square.jpg\",0.1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52939042,"AnswerCount":1,"Tags":"<tensorflow><tensorflowsharp>","CreationDate":"2018-10-22T23:29:53.160","AcceptedAnswerId":54983992.0,"OwnerUserId":973987.0,"Title":"Tensorflowsharp and Retinanet -- How to determine what to Fetch when graph is run?","Body":"<p>I've been using TensorflowSharp with Faster RCNN successfully for a while now; however, I recently trained a Retinanet model, verified it works in python, and have created a frozen pb file for use with Tensorflow.  For FRCNN, there is an example in the TensorflowSharp GitHub repo that shows how to run\/fetch this model.  For Retinanet, I tried modifying the code but nothing seems to work.  I have a model summary for Retinanet that I've tried to work from, but it's not obvious to me what should be used.<\/p>\n\n<p>For FRCNN, the graph is run in this way:<\/p>\n\n<pre><code>    var runner = m_session.GetRunner();\n\n    runner\n        .AddInput(m_graph[\"image_tensor\"][0], tensor)\n        .Fetch(\n        m_graph[\"detection_boxes\"][0],\n        m_graph[\"detection_scores\"][0],\n        m_graph[\"detection_classes\"][0],\n        m_graph[\"num_detections\"][0]);\n\n       var output = runner.Run();\n\n        var boxes = (float[,,])output[0].GetValue(jagged: false);\n        var scores = (float[,])output[1].GetValue(jagged: false);\n        var classes = (float[,])output[2].GetValue(jagged: false);\n        var num = (float[])output[3].GetValue(jagged: false);\n<\/code><\/pre>\n\n<p>From the model summary for FRCNN, it is obvious what the input (\"image_tensor\") and outputs (\"detection_boxes\", \"detection_scores\", \"detection_classes\", and \"num_detections\") are.  They are not the same for Retinanet (I've tried), and I can't figure out what they should be.  The \"Fetch\" part of the code above is causing a crash, and I'm guessing its because I'm not getting the node names right.<\/p>\n\n<p>I won't paste the entire Retinanet summary here, but here is the first few nodes:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, None, None, 3 0                                            \n__________________________________________________________________________________________________\npadding_conv1 (ZeroPadding2D)   (None, None, None, 3 0           input_1[0][0]                    \n__________________________________________________________________________________________________\nconv1 (Conv2D)                  (None, None, None, 6 9408        padding_conv1[0][0]              \n__________________________________________________________________________________________________\nbn_conv1 (BatchNormalization)   (None, None, None, 6 256         conv1[0][0]                      \n__________________________________________________________________________________________________\nconv1_relu (Activation)         (None, None, None, 6 0           bn_conv1[0][0]                   \n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>And here are the last several nodes:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nanchors_0 (Anchors)             (None, None, 4)      0           P3[0][0]                         \n__________________________________________________________________________________________________\nanchors_1 (Anchors)             (None, None, 4)      0           P4[0][0]                         \n__________________________________________________________________________________________________\nanchors_2 (Anchors)             (None, None, 4)      0           P5[0][0]                         \n__________________________________________________________________________________________________\nanchors_3 (Anchors)             (None, None, 4)      0           P6[0][0]                         \n__________________________________________________________________________________________________\nanchors_4 (Anchors)             (None, None, 4)      0           P7[0][0]                         \n__________________________________________________________________________________________________\nregression_submodel (Model)     (None, None, 4)      2443300     P3[0][0]                         \n                                                                 P4[0][0]                         \n                                                                 P5[0][0]                         \n                                                                 P6[0][0]                         \n                                                                 P7[0][0]                         \n__________________________________________________________________________________________________\nanchors (Concatenate)           (None, None, 4)      0           anchors_0[0][0]                  \n                                                                 anchors_1[0][0]                  \n                                                                 anchors_2[0][0]                  \n                                                                 anchors_3[0][0]                  \n                                                                 anchors_4[0][0]                  \n__________________________________________________________________________________________________\nregression (Concatenate)        (None, None, 4)      0           regression_submodel[1][0]        \n                                                                 regression_submodel[2][0]        \n                                                                 regression_submodel[3][0]        \n                                                                 regression_submodel[4][0]        \n                                                                 regression_submodel[5][0]        \n__________________________________________________________________________________________________\nboxes (RegressBoxes)            (None, None, 4)      0           anchors[0][0]                    \n                                                                 regression[0][0]                 \n__________________________________________________________________________________________________\nclassification_submodel (Model) (None, None, 1)      2381065     P3[0][0]                         \n                                                                 P4[0][0]                         \n                                                                 P5[0][0]                         \n                                                                 P6[0][0]                         \n                                                                 P7[0][0]                         \n__________________________________________________________________________________________________\nclipped_boxes (ClipBoxes)       (None, None, 4)      0           input_1[0][0]                    \n                                                                 boxes[0][0]                      \n__________________________________________________________________________________________________\nclassification (Concatenate)    (None, None, 1)      0           classification_submodel[1][0]    \n                                                                 classification_submodel[2][0]    \n                                                                 classification_submodel[3][0]    \n                                                                 classification_submodel[4][0]    \n                                                                 classification_submodel[5][0]    \n__________________________________________________________________________________________________\nfiltered_detections (FilterDete [(None, 300, 4), (No 0           clipped_boxes[0][0]              \n                                                                 classification[0][0]             \n==================================================================================================\nTotal params: 36,382,957\nTrainable params: 36,276,717\nNon-trainable params: 106,240\n<\/code><\/pre>\n\n<p>Any help with figure out how to fix the \"Fetch\" part of this would be greatly appreciated.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>To dig a little further into this, I found a python function to print the operation names from a .pb file.  When doing this for the FRCNN .pb file, it clearly gave the output node names, as can be seen below (only posting the last several lines from the output of the python function).<\/p>\n\n<pre><code>import\/SecondStagePostprocessor\/BatchMultiClassNonMaxSuppression\/map\/TensorArrayStack_4\/TensorArrayGatherV3\nimport\/SecondStagePostprocessor\/ToFloat_1\nimport\/add\/y\nimport\/add\nimport\/detection_boxes\nimport\/detection_scores\nimport\/detection_classes\nimport\/num_detections\n<\/code><\/pre>\n\n<p>If I do the same thing for the Retinanet .pb file, it's not obvious what the outputs are.  Here's the last several lines from the python function.<\/p>\n\n<pre><code>import\/filtered_detections\/map\/while\/NextIteration_4\nimport\/filtered_detections\/map\/while\/Exit_2\nimport\/filtered_detections\/map\/while\/Exit_3\nimport\/filtered_detections\/map\/while\/Exit_4\nimport\/filtered_detections\/map\/TensorArrayStack\/TensorArraySizeV3\nimport\/filtered_detections\/map\/TensorArrayStack\/range\/start\nimport\/filtered_detections\/map\/TensorArrayStack\/range\/delta\nimport\/filtered_detections\/map\/TensorArrayStack\/range\nimport\/filtered_detections\/map\/TensorArrayStack\/TensorArrayGatherV3\nimport\/filtered_detections\/map\/TensorArrayStack_1\/TensorArraySizeV3\nimport\/filtered_detections\/map\/TensorArrayStack_1\/range\/start\nimport\/filtered_detections\/map\/TensorArrayStack_1\/range\/delta\nimport\/filtered_detections\/map\/TensorArrayStack_1\/range\nimport\/filtered_detections\/map\/TensorArrayStack_1\/TensorArrayGatherV3\nimport\/filtered_detections\/map\/TensorArrayStack_2\/TensorArraySizeV3\nimport\/filtered_detections\/map\/TensorArrayStack_2\/range\/start\nimport\/filtered_detections\/map\/TensorArrayStack_2\/range\/delta\nimport\/filtered_detections\/map\/TensorArrayStack_2\/range\nimport\/filtered_detections\/map\/TensorArrayStack_2\/TensorArrayGatherV3\n<\/code><\/pre>\n\n<p>For reference, here's the python function that I used:<\/p>\n\n<pre><code>def printTensors(pb_file):\n\n    # read pb into graph_def\n    with tf.gfile.GFile(pb_file, \"rb\") as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    # import graph_def\n    with tf.Graph().as_default() as graph:\n        tf.import_graph_def(graph_def)\n\n    # print operations\n    for op in graph.get_operations():\n        print(op.name)\n<\/code><\/pre>\n\n<p>Hope this helps.<\/p>\n","answers":[{"AnswerId":"54983992","CreationDate":"2019-03-04T13:09:16.670","ParentId":null,"OwnerUserId":"429476","Title":null,"Body":"<p>I am not sure exactly the problem you are facing; You can get the ouputs from TF Serving output, Actually in retinanet Ipython\/Jupyter notebook they have mentioned the output format as well<\/p>\n\n<p>Querying the save model gives<\/p>\n\n<pre><code>  \"\"\"  The given SavedModel SignatureDef contains the following output(s):\n    outputs['filtered_detections\/map\/TensorArrayStack\/TensorArrayGatherV3:0'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 300, 4)\n        name: filtered_detections\/map\/TensorArrayStack\/TensorArrayGatherV3:0\n    outputs['filtered_detections\/map\/TensorArrayStack_1\/TensorArrayGatherV3:0'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 300)\n        name: filtered_detections\/map\/TensorArrayStack_1\/TensorArrayGatherV3:0\n    outputs['filtered_detections\/map\/TensorArrayStack_2\/TensorArrayGatherV3:0'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 300)\n        name: filtered_detections\/map\/TensorArrayStack_2\/TensorArrayGatherV3:0\n  Method name is: tensorflow\/serving\/predict\n  ---\n  From retina-net\n  In general, inference of the network works as follows:\n  boxes, scores, labels = model.predict_on_batch(inputs)\n  Where `boxes` are shaped `(None, None, 4)` (for `(x1, y1, x2, y2)`), scores is shaped `(None, None)` (classification score) and labels is shaped `(None, None)` (label corresponding to the score). In all three outputs, the first dimension represents the shape and the second dimension indexes the list of detections.\n\"\"\"\n<\/code><\/pre>\n"}]}
+{"QuestionId":52939732,"AnswerCount":1,"Tags":"<amazon-ec2><torch><amazon-sagemaker>","CreationDate":"2018-10-23T01:10:24.407","AcceptedAnswerId":null,"OwnerUserId":3772726.0,"Title":"Installing Torch on Amazon Linux","Body":"<p>This took me an excruciatingly long time for me to figure out so I decided to save people the trouble.<\/p>\n\n<p>Trying to install torch on an ec2 instance with Amazon Linux(redhat) failed with the following command:<\/p>\n\n<pre><code>git clone https:\/\/github.com\/torch\/distro.git ~\/torch --recursive\ncd ~\/torch; bash install-deps;\n.\/install.sh\n<\/code><\/pre>\n\n<p>It gave me the error:<\/p>\n\n<pre><code>OpenBLAS Failed to compile\n<\/code><\/pre>\n","answers":[{"AnswerId":"52939744","CreationDate":"2018-10-23T01:13:20.007","ParentId":null,"OwnerUserId":"3772726","Title":null,"Body":"<p>Turns out that you just run the install-deps script and ignore the failure, then run the following line of code:<\/p>\n\n<pre><code>export TORCH_NVCC_FLAGS=\"-D__CUDA_NO_HALF_OPERATORS__\"\n<\/code><\/pre>\n\n<p>and follow with the install.sh script as usual:<\/p>\n\n<pre><code>.\/install.sh\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/torch\/torch7\/issues\/1086\" rel=\"nofollow noreferrer\">Source Here<\/a><\/p>\n"}]}
+{"QuestionId":52939737,"AnswerCount":1,"Tags":"<tensorflow><keras><google-cloud-platform><google-cloud-ml>","CreationDate":"2018-10-23T01:11:21.077","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Google Cloud ML Engine: Trouble With Local Prediction Given saved_model.pb","Body":"<p>I've trained a Keras model using the <code>tf.data.Dataset<\/code> API and am trying to see if I've saved it (as <code>saved_model.pb<\/code>) correctly, so I can use it on ML Engine.  Here's what I've done:<\/p>\n\n<pre><code>estimator = tf.keras.estimator.model_to_estimator(my_model)\n# create serving function...\nestimator.export_savedmodel('.\/export', serving_fn)\n<\/code><\/pre>\n\n<p>So now I'm trying to use <code>gcloud ml-engine local predict<\/code> to see if I can get a prediction back.  I'm doing:<\/p>\n\n<pre><code>gcloud ml-engine local predict --model-dir=~\/path\/to\/folder --json-instances=instances.json\n<\/code><\/pre>\n\n<p>Unfortunately, I get:<\/p>\n\n<pre><code>cloud.ml.prediction.prediction_utils.PredictionError: Failed to load model: Cloud ML only supports TF 1.0 or above and models saved in SavedModel format. (Error code: 0)\n<\/code><\/pre>\n\n<p>So then I try adding <code>--runtime-version=1.2<\/code> to my command like this:<\/p>\n\n<pre><code>gcloud ml-engine local predict --model-dir=~\/path\/to\/folder --json-instances=instances.json --runtime-version=1.2\n<\/code><\/pre>\n\n<p>and I get back:<\/p>\n\n<pre><code>ERROR: (gcloud.ml-engine.local.predict) unrecognized arguments: --runtime-version=1.2\n<\/code><\/pre>\n\n<p>Any idea what I'm doing incorrectly \/ how to fix?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53105935","CreationDate":"2018-11-01T17:00:58.220","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>For posterity: the problem turned out to be an incorrect path. If anybody else encounters this issue try using the full absolute path and ensure you are point to the directory containing the <code>saved_model.pb<\/code> file.<\/p>\n"}]}
+{"QuestionId":52939792,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-23T01:20:53.353","AcceptedAnswerId":52940252.0,"OwnerUserId":7885866.0,"Title":"Tensorflow, check which values in tensor are whole numbers","Body":"<p>I have a tensor:<\/p>\n\n<pre><code>numbers = tf.constant([[4.00, 3.33], [2.34, 7.00]])\n<\/code><\/pre>\n\n<p>What I want to do is get a tensor that has the same dimensions as 'numbers', but has 1's at the indices where numbers is a whole number, and 0's at the indices where it is not, like so:<\/p>\n\n<pre><code>ans = [[1, 0],[0, 1]]\n<\/code><\/pre>\n\n<p>I'm guessing I would have to use tf.where() maybe? I'm really unsure how to do something like this using tensorflow. Thank you<\/p>\n","answers":[{"AnswerId":"52940252","CreationDate":"2018-10-23T02:20:42.237","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<pre><code>import tensorflow as tf\nimport numpy as np\n\ntf.reset_default_graph()\nwith tf.Session() as sess:\n    fake_data = np.asarray([[1,2.4, 3.5], [3.4, 2.00, 10.001], [105.1, 100, 10]])\n    a = tf.constant(data)\n\n    # find where floor == actual value (thus, is a whole number)\n    mask = tf.equal(tf.floor(data), data)\n\n    # Get the indices\n    idx = tf.where(mask)\n\n\n    print(sess.run(a))\n    print(sess.run(idx))\n<\/code><\/pre>\n\n<p>This should do the trick :) I tried commenting it so that what I was doing was clear and I think it's easy to understand. What I wrote is based on <a href=\"https:\/\/stackoverflow.com\/a\/48377032\/3532564\">this comment<\/a><\/p>\n"}]}
+{"QuestionId":52940184,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook><tensorflow-hub>","CreationDate":"2018-10-23T02:14:02.170","AcceptedAnswerId":52940185.0,"OwnerUserId":9494790.0,"Title":"Tensorflow Hub : Stuck while importing a model","Body":"<p>Trying to import some models with Tensorflow Hub with this code :<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_hub as hub\n\nelmo_model = hub.Module('https:\/\/tfhub.dev\/google\/elmo\/2', trainable=True)\n<\/code><\/pre>\n\n<p>Makes my notebook stuck. The only log line appearing before getting stuck is :<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Using \/tmp\/tfhub_modules to cache modules.<\/p>\n<\/blockquote>\n\n<p>How to unstuck it and allow me to import models from Tensorflow Hub ?<\/p>\n","answers":[{"AnswerId":"52940185","CreationDate":"2018-10-23T02:14:02.170","ParentId":null,"OwnerUserId":"9494790","Title":null,"Body":"<p>It was simply about privileges : I couldn't access the default directory where Tensorflow Hub store the models (<code>\/tmp\/tfhub_modules<\/code>).<\/p>\n\n<p>To solve it, I just choose a directory to store the models which I can access :<\/p>\n\n<pre><code>import os\nimport tensorflow as tf\nimport tensorflow_hub as hub\n\nos.environ['TFHUB_CACHE_DIR'] = '\/home\/user\/workspace\/tf_cache'\nelmo_model = hub.Module('https:\/\/tfhub.dev\/google\/elmo\/2', trainable=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52940399,"AnswerCount":3,"Tags":"<caffe><lmdb>","CreationDate":"2018-10-23T02:40:26.653","AcceptedAnswerId":null,"OwnerUserId":4849120.0,"Title":"LMDB: How to interpret output from mdb_stat and mdb_dump utilities","Body":"<p>I have a functional LMDB that, for test purposes, currently contains only 21 key \/ value records.  I've successfully tested inserting and reading records, and I'm comfortable with the database working as intended.<\/p>\n\n<p>However, when I use the <code>mdb_stat<\/code> and <code>mdb_dump<\/code> utilities, I see the following output, respectively:<\/p>\n\n<pre><code>Status of Main DB\n   Tree depth: 1\n   Branch pages: 0\n   Leaf pages: 1\n   Overflow pages: 0\n   Entries: 1\n\nVERSION=3\nformat=bytevalue\ntype=btree\nmapsize=1073741824\nmaxreaders=126\ndb_pagesize=4096\nHEADER=END\n4d65737361676573\n<\/code><\/pre>\n\n<p><code>000000000000010000000000000000000100000000000000d81e0000000000001500000000000000ba1d000000000000<\/code><\/p>\n\n<pre><code>DATA=END\n<\/code><\/pre>\n\n<p>In particular, why would <code>mdb_stat<\/code> indicate only one entry when I have 21?  Moreover, each entry comprises 1024 x 300 values of five bytes per value.  <code>mdb_dump<\/code> obviously doesn't show anywhere near the 1,536,000 bytes I'd expect to see, yet the values I <code>mdb_put()<\/code> and <code>mdb_get()<\/code> on the fly are correct.  Anyone know what's going on?<\/p>\n","answers":[{"AnswerId":"53055614","CreationDate":"2018-10-30T00:09:44.013","ParentId":null,"OwnerUserId":"4849120","Title":null,"Body":"<p>The relationship between an operating system's directory and an LMDB environment's data.mdb and lock.mdb files is one-to-one.<\/p>\n\n<p>If the LMDB environment (in the OS directory) has more than one database, then the environment also contains a separate LMDB database containing all of its <em>named<\/em> databases.<\/p>\n\n<p>The <code>mdb_stat<\/code> and <code>mdb_dump<\/code> utilities appear to contain minimal logic, so when they are fed a given directory via the command line, they appear to produce results only for the database storing database names and not the database(s) storing the actual data of interest.<\/p>\n"},{"AnswerId":"58245403","CreationDate":"2019-10-05T04:41:51.127","ParentId":null,"OwnerUserId":"8094572","Title":null,"Body":"<p>Since you are using a sub-database, the number of entries in the main database corresponds to the number of sub-databases you've created (ie just 1).<\/p>\n\n<p>Try using <code>mdb_stat -a<\/code>. This will show you a break-down of all the sub-databases (as well as the main DB). In this breakdown it will list the number of entries for each sub-database. Here you should see your 21 entries.<\/p>\n"},{"AnswerId":"57425193","CreationDate":"2019-08-09T07:19:21.683","ParentId":null,"OwnerUserId":"4898880","Title":null,"Body":"<p><code>4d65737361676573<\/code> is the Ascii for \"Messages\", which is the name of table (\"sub-db\" in lmdb terminology) storing the actual data in your case.<\/p>\n\n<p>The <code>mdb_dump<\/code> command only dumps the main db by default. You can use the -s option to dump that sub-db, i.e. <\/p>\n\n<pre><code>mdb_dump -s Messages\n<\/code><\/pre>\n\n<p>or you can use the -a option to dump all the sub-dbs.<\/p>\n"}]}
+{"QuestionId":52940528,"AnswerCount":0,"Tags":"<python><machine-learning><scikit-learn><keras><deep-learning>","CreationDate":"2018-10-23T03:00:29.230","AcceptedAnswerId":null,"OwnerUserId":10186035.0,"Title":"Incremental training of Keras image classification model","Body":"<p>I used the smaller VGG model and modified the training script of the following tutorial for training a previously trained model. \nOriginal source of model and script:\n<a href=\"https:\/\/www.pyimagesearch.com\/2018\/04\/16\/keras-and-convolutional-neural-networks-cnns\/\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2018\/04\/16\/keras-and-convolutional-neural-networks-cnns\/<\/a><\/p>\n\n<p><strong>This is what I did:<\/strong><\/p>\n\n<p><strong>1st training session<\/strong>:<\/p>\n\n<ul>\n<li>train the model with image dataset of 2 classes A and B with the original training script from the tutorial<\/li>\n<\/ul>\n\n<p><strong>2nd training session:<\/strong><\/p>\n\n<ul>\n<li>load the trained keras model and train with image dataset of class C without any new data of class A and B with the following modified training script \n(the loading and saving model method is referenced from the following stackoverflow thread: <a href=\"https:\/\/stackoverflow.com\/questions\/42666046\/loading-a-trained-keras-model-and-continue-training\">Loading a trained Keras model and continue training<\/a>)<\/li>\n<li>load the pickled array of labels of the 1st session, combine it with the new label array in the 2nd session and save it in lb.pickle<\/li>\n<\/ul>\n\n<p><strong>Result:<\/strong><\/p>\n\n<p>The trained model after the 2nd session can only recognize the new class in the 2nd session. It seems other classes trained in the 1st session are lost. It just doesn't work.<\/p>\n\n<p><strong>My question:<\/strong>\nHow to fix the following script to make incremental training work?\nOr any other suggestion or reference of incremental training that is similar to my case?<\/p>\n\n<p><strong>My modified training script:<\/strong><\/p>\n\n<pre><code>from keras.utils import np_utils\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.optimizers import Adam\nfrom keras.preprocessing.image import img_to_array\nfrom keras.models import load_model\n\nfrom sklearn.preprocessing import LabelBinarizer\nfrom sklearn.model_selection import train_test_split\nfrom smallervggnet import SmallerVGGNet\nfrom imutils import paths\nimport numpy as np\nimport argparse, os, sys\nimport random\nimport pickle\nimport cv2\n\nap = argparse.ArgumentParser()\nap.add_argument(\"-d\", \"--dataset\", required=True,\nhelp=\"path to input dataset (i.e., directory of images)\")\nap.add_argument(\"-im\", \"--loadmodel\", required=True,\nhelp=\"path to model to be loaded\")\nap.add_argument(\"-m\", \"--model\", required=True,\nhelp=\"path to output model\")\nap.add_argument(\"-l\", \"--labelbin\", required=True,\nhelp=\"path to output label binarizer\")\nap.add_argument(\"-p\", \"--plot\", type=str, default=\"plot.png\",\nhelp=\"path to output accuracy\/loss plot\")\nargs = vars(ap.parse_args())\n\n\nEPOCHS = 100\nINIT_LR = 1e-3\nBS = 10\nIMAGE_DIMS = (256, 256, 3)\n\ndata = []\nlabels = []\n\n\nprint(\"[INFO] loading images...\")\nimagePaths = sorted(list(paths.list_images(args[\"dataset\"])))\nrandom.seed(42)\nrandom.shuffle(imagePaths)\n\nfor imagePath in imagePaths:\n    image = cv2.imread(imagePath)\n    image = cv2.resize(image, (IMAGE_DIMS[1], IMAGE_DIMS[0]))\n    image = img_to_array(image)\n    data.append(image)\n\n\nlabel = imagePath.split(os.path.sep)[-2]\nlabels.append(label)\n\n\ndata = np.array(data, dtype=\"float\") \/ 255.0\nlabels = np.array(labels)\nprint(\"[INFO] data matrix: {:.2f}MB\".format(\ndata.nbytes \/ (1024 * 1000.0)))\n\n\nlb = LabelBinarizer()\nbLabels = lb.fit_transform(labels)\n\n\n(trainX, testX, trainY, testY) = train_test_split(data,\nbLabels, test_size=0.2, random_state=42)\n\n\n#add these 2 lines to avoid error\ntrainY = np_utils.to_categorical(trainY, 2)\ntestY = np_utils.to_categorical(testY, 2)\n\naug = ImageDataGenerator(rotation_range=25, width_shift_range=0.1,\nheight_shift_range=0.1, shear_range=0.2, zoom_range=0.2,\nhorizontal_flip=True, fill_mode=\"nearest\")\n\n\nprint(\"[INFO] load previously trained model\")\nmodelPath = args[\"loadmodel\"]\nmodel = load_model(modelPath)\n\n\nprint(\"[INFO] training network...\")\nH = model.fit_generator(\naug.flow(trainX, trainY, batch_size=BS),\nvalidation_data=(testX, testY),\nsteps_per_epoch=len(trainX) \/\/ BS,\nepochs=EPOCHS, verbose=1)\n\n\nprint(\"[INFO] serializing network...\")\nmodel.save(args[\"model\"])\n\n\n# my attempt to keep the labels of all the training session in label binarizer\nprevArray = '.\/train_output\/previous_data_array.pickle'\n\narrPickle = labels\n\nif os.path.getsize(prevArray) &gt; 0:  \n    prev = pickle.loads(open(prevArray, 'rb').read())\n    arrPickle = np.concatenate((prev,labels), axis=0)\n\nlb = LabelBinarizer()\nlb.fit_transform(arrPickle)\n\nprint(\"[INFO] serializing combined label array...\")\nf = open(prevArray, \"wb\")\nf.write(pickle.dumps(arrPickle))\nf.close()\n\nprint(\"[INFO] serializing label binarizer...\")\nf = open(args[\"labelbin\"], \"wb\")\nf.write(pickle.dumps(lb))\nf.close()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52940552,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-23T03:05:05.047","AcceptedAnswerId":52966635.0,"OwnerUserId":8100655.0,"Title":"ValueError: Operation u'tpu_140462710602256\/VarIsInitializedOp' has been marked as not fetchable","Body":"<p>The code works fine on GPU and CPU.But when I use keras_to_tpu_model function to make the model able to run on TPU, the error occurred.     <\/p>\n\n<p>This is the full output on colab:<a href=\"https:\/\/colab.research.google.com\/gist\/WangHexie\/2252beb26f16354cb6e9ba2639970e5b\/tpu-error.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/gist\/WangHexie\/2252beb26f16354cb6e9ba2639970e5b\/tpu-error.ipynb<\/a><br>\nChange runtype to TPU,I think this can be reproduced.         <\/p>\n\n<p>Code on github:<a href=\"https:\/\/github.com\/WangHexie\/DHNE\/blob\/master\/src\/hypergraph_embedding.py#L60\" rel=\"nofollow noreferrer\">https:\/\/github.com\/WangHexie\/DHNE\/blob\/master\/src\/hypergraph_embedding.py#L60<\/a>     <\/p>\n\n<p>You can test the code on GPU by changing to the gpu branch.           <\/p>\n\n<p>Traceback<\/p>\n\n<pre><code>    Traceback (most recent call last):\n  File \"src\/hypergraph_embedding.py\", line 158, in &lt;module&gt;\n    h.train(dataset)\n  File \"src\/hypergraph_embedding.py\", line 75, in train\n    epochs=self.options.epochs_to_train, verbose=1)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 2177, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/engine\/training_generator.py\", line 176, in fit_generator\n    x, y, sample_weight=sample_weight, class_weight=class_weight)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1940, in train_on_batch\n    outputs = self.train_function(ins)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 1238, in __call__\n    infeed_manager)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 1143, in _tpu_model_ops_for_input_specs\n    infeed_manager)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 1053, in _specialize_model\n    _model_fn, inputs=[[]] * self._tpu_assignment.num_towers)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/tpu.py\", line 687, in split_compile_and_replicate\n    outputs = computation(*computation_inputs)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 959, in _model_fn\n    self.model.cpu_optimizer)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 378, in _clone_optimizer\n    config = optimizer.get_config()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/optimizers.py\", line 275, in get_config\n    'lr': float(K.get_value(self.lr)),\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/backend.py\", line 2709, in get_value\n    return x.eval(session=get_session())\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/backend.py\", line 469, in get_session\n    _initialize_variables(session)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/keras\/backend.py\", line 731, in _initialize_variables\n    [variables_module.is_variable_initialized(v) for v in candidate_vars])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1137, in _run\n    self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 484, in __init__\n    self._assert_fetchable(graph, fetch.op)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 497, in _assert_fetchable\n    'Operation %r has been marked as not fetchable.' % op.name)\nValueError: Operation u'tpu_140276544043536\/VarIsInitializedOp' has been marked as not fetchable.\n<\/code><\/pre>\n","answers":[{"AnswerId":"52966635","CreationDate":"2018-10-24T10:27:06.837","ParentId":null,"OwnerUserId":"7288564","Title":null,"Body":"<p>I have a same issue which confuses me two days. I find a solution is that just switch to using <code>tf.train.RMSPropOptimizer<\/code> instead of using RMSProp from <code>tensorflow.keras.optimizers<\/code>.<\/p>\n"}]}
+{"QuestionId":52940635,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-23T03:16:49.433","AcceptedAnswerId":null,"OwnerUserId":6750211.0,"Title":"tensorflow how can I solve dtype conflict between one_hot and sign","Body":"<p>My neural network has the following output:<\/p>\n\n<ul>\n<li><code>logits<\/code> are outputs from tanh nodes, so the value is a float within (-1,1)<\/li>\n<li><code>action<\/code> are the sign of <code>logits<\/code><\/li>\n<li><code>one_hot<\/code> are the one_hot version of <code>action<\/code>, with dimension 3 represents -1, 0, and +1<\/li>\n<\/ul>\n\n<p>The problem is, my loss function is related to the <code>one_hot<\/code> values, so I build the neural network output part like this:<\/p>\n\n<pre><code>logits = tf.contrib.layers.fully_connected(outputs, 1, activation_fn=tf.tanh)\naction = tf.sign(logits)\none_hot = tf.one_hot(action+1, depth=3)\n<\/code><\/pre>\n\n<p>and this gives me a TypeError of <\/p>\n\n<blockquote>\n  <p>TypeError: Value passed to parameter 'indices' has DataType float32\n  not in list of allowed values: uint8, int32, int64<\/p>\n<\/blockquote>\n\n<p>Then I tried to change <code>one_hot<\/code> to:<\/p>\n\n<pre><code>one_hot = tf.one_hot(tf.cast(action, tf.int32)+1, depth=3)\n<\/code><\/pre>\n\n<p>And I got another error of no gradients:<\/p>\n\n<blockquote>\n  <p>ValueError: No gradients provided for any variable, check your graph\n  for ops that do not support gradients, between variables [...]<\/p>\n<\/blockquote>\n\n<p>Is there any workarounds that I could use to avoid both errors. Any help is appreciated.<\/p>\n","answers":[]}
+{"QuestionId":52940677,"AnswerCount":1,"Tags":"<pandas><tensorflow><amazon-sagemaker>","CreationDate":"2018-10-23T03:20:56.293","AcceptedAnswerId":52944464.0,"OwnerUserId":1007922.0,"Title":"AWS Sagemaker: AttributeError: module 'pandas' has no attribute 'core'","Body":"<p>Let me prefix this by saying I'm very new to tensorflow and even newer to AWS Sagemaker.<\/p>\n\n<p>I have some tensorflow\/keras code that I wrote and tested on a local dockerized Jupyter notebook and it runs fine. In it, I import a csv file as my input.<\/p>\n\n<p>I use Sagemaker to spin up a jupyter notebook instance with conda_tensorflow_p36. I modified the pandas.read_csv() code to point to my input file, now hosted on a S3 bucket.<\/p>\n\n<p>So I changed this line of code from<\/p>\n\n<pre><code>import pandas as pd\n\ndata = pd.read_csv(\"\/input.csv\", encoding=\"latin1\")\n<\/code><\/pre>\n\n<p>to this<\/p>\n\n<pre><code>import pandas as pd\n\ndata = pd.read_csv(\"https:\/\/s3.amazonaws.com\/my-sagemaker-bucket\/input.csv\", encoding=\"latin1\")\n<\/code><\/pre>\n\n<p>and I get this error<\/p>\n\n<pre><code>AttributeError: module 'pandas' has no attribute 'core'\n<\/code><\/pre>\n\n<p>I'm not sure if it's a permissions issue. I read that as long as I name my bucket with the string \"sagemaker\" it should have access to it.<\/p>\n","answers":[{"AnswerId":"52944464","CreationDate":"2018-10-23T08:27:49.857","ParentId":null,"OwnerUserId":"10535783","Title":null,"Body":"<p>Pull our data from S3 for example:<\/p>\n\n<pre><code>import boto3\nimport io\nimport pandas as pd\n\n\n# Set below parameters\nbucket = '&lt;bucket name&gt;'\nkey = 'data\/training\/iris.csv'\nendpointName = 'decision-trees'\n\n# Pull our data from S3\ns3 = boto3.client('s3')\nf = s3.get_object(Bucket=bucket, Key=key)\n\n# Make a dataframe\nshape = pd.read_csv(io.BytesIO(f['Body'].read()), header=None)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52940691,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><neural-network><keras><rnn>","CreationDate":"2018-10-23T03:22:26.113","AcceptedAnswerId":null,"OwnerUserId":7934206.0,"Title":"Why RNN is throwing batch input shape error?","Body":"<p>My <code>x_train<\/code> shape is (798,3) and <code>y_train<\/code> input shape is (798, 1). I am creating a RNN like this <\/p>\n\n<pre><code>def create_rnn_model():\n    model = Sequential()\n    model.add(SimpleRNN(20,return_sequences=False,stateful=stateful,activation='relu',batch_input_shape=(1,3,1)))\n    model.add(Activation('relu'))\n    adam = optimizers.Adam(lr=0.001)\n    model.compile(loss='mean_squared_error', optimizer=adam, metrics=[root_mean_squared_error])\n\n    return model\n<\/code><\/pre>\n\n<p>But this returns the error<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected simple_rnn_1_input to\n  have 3 dimensions, but got array with shape (798, 3)<\/p>\n<\/blockquote>\n\n<p>My batch size =1 and my timestep is 3 and <code>dat_dim=1<\/code> .Then where am I doing it wrong?\nAny help is appreciated.<\/p>\n\n<blockquote>\n  <p>EDIT<\/p>\n<\/blockquote>\n\n<p>I changed my x_train to shape (798,3,1) and y_train shape to (798,) and ran the model but it threw an error <\/p>\n\n<pre><code>ValueError: Error when checking target: expected activation_1 to have shape (20,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>But I can run the model with 1 unit. How do I specify the model to run with 20 units instead of one.<\/p>\n\n<p>My Network Model is like this and my current L=3.\n<a href=\"https:\/\/i.stack.imgur.com\/I0gRF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/I0gRF.png\" alt=\"Network\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":52940928,"AnswerCount":1,"Tags":"<python><neural-network><keras><lstm><dimensions>","CreationDate":"2018-10-23T03:55:48.633","AcceptedAnswerId":52961010.0,"OwnerUserId":680074.0,"Title":"LSTM keras - Value error how to resolve input dimension","Body":"<p>I have made a function for training data which pick random files either positive or negative. it is a binary classification problem.<\/p>\n\n<pre><code>    model=Sequential()\nInputBatch = np.expand_dims(InputBatch, 0)\nprint(InputBatch.shape)\nmodel.add(LSTM(100,input_shape=(1,6,30),return_sequences=True))\nmodel.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])\nmodel.fit(InputBatch,PositiveOrNegativeLabel,batch_size=6,nb_epoch=10,verbose=1,validation_split=0.05)\n<\/code><\/pre>\n\n<p>the shape of InputBatch variable is (1,6,30) <\/p>\n\n<p>for example my input data is <\/p>\n\n<pre><code>[[   nan  1520.  1295.    nan  8396.  9322. 12715.    nan  5172.  7232.\n  11266.    nan 11266.  2757.  4416. 12020. 12111.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  3045. 11480.   900.  5842. 11496.  4463.    nan 11956.   900.\n  10400.  8022.  2504. 12106.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  9307. 12003.  2879.  6398.  9372.  4614.  5222.    nan    nan\n   2879. 10364.  6923.  4709.  4860. 11871.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  6689.  2818. 12003.  6480.    nan     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]]\n<\/code><\/pre>\n\n<p>I have set the shape of data which is (6,30)<\/p>\n\n<p>I am getting value error <\/p>\n\n<pre><code>ValueError: Error when checking input: expected lstm_16_input to have 3 dimensions, but got array with shape (1,6, 30)\n<\/code><\/pre>\n\n<p><strong>it is getting three dimension input i don't understand how and why<\/strong><\/p>\n","answers":[{"AnswerId":"52961010","CreationDate":"2018-10-24T04:14:20.983","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>LSTM inputs have to be in <code>3 dimensional(samples,timesteps,features)<\/code>. And your data seems to be in 2D. You can use the <code>numpys reshape()<\/code> function to convert your data to 3D.<\/p>\n\n<p>For example if you are using 1 time-step then you will have to reshape it as <code>array.reshape(6,1,30)<\/code> or if you are using 6 timestep then <code>array.reshape(1,6,30)<\/code> <\/p>\n\n<p>For more information on reshaping input for LSTM you can check this <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">site<\/a><\/p>\n\n<p>[[update]]\nThere are so many problems with your code<\/p>\n\n<pre><code> model=Sequential()\nInputBatch = np.expand_dims(InputBatch, 0)\nprint(InputBatch.shape)\nmodel.add(LSTM(100,input_shape=(1,6,30),return_sequences=True))\nmodel.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])\nmodel.fit(InputBatch,PositiveOrNegativeLabel,batch_size=6,nb_epoch=10,verbose=1,validation_split=0.05)\n<\/code><\/pre>\n\n<p>When  you convert your data to (1,6,30) you are basically saying that you have just one sample(only 1) ,batch_size is 6 but you have only 1 sample,you have only one sample but you are doing validation split.And since you have only one X value it will have only one Y ( PositiveOrNegativeLabel),So I have asisgneed just one value ie 1.<\/p>\n\n<p>I have ran your program with the changes few changes with code and data that u have shown in the question (i changed NA to 0)  :<\/p>\n\n<pre><code>    a=np.array([ \n             [0,1520,1295,0,8396,9322,12715,0,5172,7232,11266,0,11266,2757,4416,12020,12111,0,0,0,0,0,0,0,0,0,0,0,0,0],\n             [0,3045,11480,900,5842,11496,4463,0,11956,900,10400,8022,2504,12106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],\n             [0,9307,12003,2879,6398,9372,4614,5222,0,0,2879,10364,6923,4709,4860,11871,0,0,0,0,0,0,0,0,0,0,0,0,0,0],\n             [0,6689,2818,12003,6480,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],\n             [0,3395,1087,11904,7232,8840,10115,4494,11516,7441,8535,12106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],\n             [0,1287,420,4070,11087,7410,12186,2387,12111,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],\n           ]\n          )\n\nPositiveOrNegativeLabel=np.array([[1]])\nPositiveOrNegativeLabel=PositiveOrNegativeLabel.reshape(1,-1)\nPositiveOrNegativeLabel.shape\nInputBatch =InputBatch.reshape(1,6,30)\nInputBatch.shape\nmodel=Sequential()\nmodel.add(LSTM(1,input_shape=(6,30)))\nmodel.add(Dense(1,activation=\"sigmoid\"))\nmodel.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])\nmodel.fit(InputBatch,PositiveOrNegativeLabel,batch_size=1,verbose=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52941071,"AnswerCount":1,"Tags":"<tensorflow><julia>","CreationDate":"2018-10-23T04:12:20.737","AcceptedAnswerId":53017157.0,"OwnerUserId":10521726.0,"Title":"Is Tensor Flow compatible with Julia 0.6.4?","Body":"<p>So I have a project in my machine learning class and we are using Julia as our programming language. We can use any packages we want to build neural networks but I can't seem to get Tensor Flow to test correctly. <code>Pkg.add(\"TensorFlow\")<\/code> works seemingly fine but here is the output for <code>Pkg.test(\"TensorFlow\")<\/code><\/p>\n\n<pre><code>julia&gt; Pkg.test(\"TensorFlow\")\nINFO: Testing TensorFlow\nERROR: LoadError: LoadError: could not load library \"C:\\Users\\Ryan .LAPTOP-                \nKJUJGIC7\\.julia\\v0.6\\TensorFlow\\src\\..\\deps\\usr\\bin\\libtensorflow\"\nThe specified module could not be found.\n\nStacktrace:\n [1] dlopen(::String, ::UInt32) at .\\libdl.jl:97\n [2] TensorFlow.Graph() at C:\\Users\\Ryan .LAPTOP-   KJUJGIC7\\.julia\\v0.6\\TensorFlow\\src\\core.jl:21\n [3] include_from_node1(::String) at .\\loading.jl:576\n [4] include(::String) at .\\sysimg.jl:14\n [5] include_from_node1(::String) at .\\loading.jl:576\n [6] include(::String) at .\\sysimg.jl:14\n [7] process_options(::Base.JLOptions) at .\\client.jl:305\n [8] _start() at .\\client.jl:371\nwhile loading C:\\Users\\Ryan .LAPTOP-KJUJGIC7\\.julia\\v0.6\\TensorFlow\\test\\..\\examples\\logistic.jl, in expression     starting on line 22\nwhile loading C:\\Users\\Ryan .LAPTOP-KJUJGIC7\\.julia\\v0.6\\TensorFlow\\test\\runtests.jl, in expression starting on line 6\n=================================================[ ERROR: TensorFlow ]==================================================\n\nfailed process: Process(`'C:\\Users\\Ryan .LAPTOP-KJUJGIC7\\AppData\\Local\\Julia-0.6.4\\bin\\julia.exe' -Cgeneric '-JC:\\Users\\Ryan .LAPTOP-KJUJGIC7\\AppData\\Local\\Julia-0.6.4\\lib\\julia\\sys.dll' --compile=yes --depwarn=yes --check-bounds=yes --code-coverage=none --color=yes --compilecache=yes 'C:\\Users\\Ryan .LAPTOP-KJUJGIC7\\.julia\\v0.6\\TensorFlow\\test\\runtests.jl'`, ProcessExited(1)) [1]\n\n========================================================================================================================\nERROR: TensorFlow had test errors\n<\/code><\/pre>\n\n<p>I'm running Julia Version 0.6.4 on Windows 10; if there's a way to resolve this error or a workaround I'd love some suggestions.<\/p>\n","answers":[{"AnswerId":"53017157","CreationDate":"2018-10-26T22:46:34.820","ParentId":null,"OwnerUserId":"9957710","Title":null,"Body":"<p><code>TensorFlow.jl<\/code> does not support Windows.\nYou have two options:<\/p>\n\n<p>(1) Try using TensorFlow via <code>PyCall.jl<\/code>:<\/p>\n\n<pre><code>using Conda\nConda.runconda(\"install -c conda-forge tensorflow\")\n<\/code><\/pre>\n\n<p>(2) Use <code>Flux.jl<\/code> instead <\/p>\n"}]}
+{"QuestionId":52941192,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-23T04:30:23.047","AcceptedAnswerId":52943475.0,"OwnerUserId":9660268.0,"Title":"Multiply multiple tensors pairwise keras","Body":"<p>I want to ask if it is possible to multiply two tensors pairwise. So for example, I have tensor output from LSTM layer,<\/p>\n\n<pre><code>lstm=LSTM(128,return_sequences=True)(input)\n\noutput=some_function()(lstm)\n<\/code><\/pre>\n\n<p><code>some_function()<\/code> should do <code>h1*h2,h2*h3....hn-1*hn<\/code>\nI found <a href=\"https:\/\/stackoverflow.com\/questions\/44294913\/how-do-i-take-the-squared-difference-of-two-keras-tensors\">How do I take the squared difference of two Keras tensors?<\/a> little helpful but since, I will have trainable paramter, I will have to make my own layer. Also, will <code>some_function<\/code> layer interpret input dimension automatically as it will be <code>hn-1<\/code><\/p>\n\n<p>I am confused on how to deal with <code>call()<\/code><\/p>\n","answers":[{"AnswerId":"52943475","CreationDate":"2018-10-23T07:31:42.237","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>One possibility is to do two crop operations and then a multiplication.\nThis does the trick!<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Input, Lambda, Multiply, LSTM\nfrom keras.models import Model\nfrom keras.layers import add\n\n\nbatch_size   = 1\nnb_timesteps = 4\nnb_features  = 2\nhidden_layer = 2\n\nin1 = Input(shape=(nb_timesteps,nb_features))\n\nlstm=LSTM(hidden_layer,return_sequences=True)(in1)\n\n# Make two slices\nfactor1 = Lambda(lambda x: x[:, 0:nb_timesteps-1, :])(lstm)\nfactor2 = Lambda(lambda x: x[:, 1:nb_timesteps, :])(lstm)\n\n# Multiply them\nout = Multiply()([factor1,factor2])\n\n# set the two outputs so we can see both results\nmodel = Model(in1,[out,lstm])\n\na = np.arange(batch_size*nb_timesteps*nb_features).reshape([batch_size,nb_timesteps,nb_features])\n\nprediction = model.predict(a)\nout_, lstm_ = prediction[0], prediction[1]\n\n\nfor x in range(nb_timesteps-1):\n    assert all( out_[0,x] == lstm_[0,x]*lstm_[0,x+1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52941277,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-23T04:39:37.343","AcceptedAnswerId":null,"OwnerUserId":2414957.0,"Title":"with model_dict['graph'].as_default(), tf.Session() as sess: ^ SyntaxError: invalid syntax","Body":"<p>I am not sure why I get this syntax error, because I use the same train method for another thing and I get no error. How can I fix it?<\/p>\n\n<pre><code>import copy\nimport os\nimport math\nimport numpy as np\nimport scipy\nimport scipy.io\n\nfrom six.moves import range\n\nimport read_data\n\n@read_data.restartable\ndef svhn_dataset_generator(dataset_name, batch_size):\n    assert dataset_name in ['train', 'test']\n    assert batch_size &gt; 0 or batch_size == -1  # -1 for entire dataset\n\n    path = '.\/svhn_mat\/' # path to the SVHN dataset you will download in Q1.1\n    file_name = '%s_32x32.mat' % dataset_name\n    file_dict = scipy.io.loadmat(os.path.join(path, file_name))\n    X_all = file_dict['X'].transpose((3, 0, 1, 2))\n    y_all = file_dict['y']\n    data_len = X_all.shape[0]\n    batch_size = batch_size if batch_size &gt; 0 else data_len\n\n    X_all_padded = np.concatenate([X_all, X_all[:batch_size]], axis=0)\n    y_all_padded = np.concatenate([y_all, y_all[:batch_size]], axis=0)\n    y_all_padded[y_all_padded == 10] = 0\n\n    for slice_i in range(int(math.ceil(data_len \/ batch_size))):\n        idx = slice_i * batch_size\n        X_batch = X_all_padded[idx:idx + batch_size]\n        y_batch = np.ravel(y_all_padded[idx:idx + batch_size])\n        yield X_batch, y_batch\n\n\n\ndef apply_classification_loss(model_function):\n    with tf.Graph().as_default() as g:\n        #with tf.device(\"\/gpu:0\"):  # use gpu:0 if on GPU\n        x_ = tf.placeholder(tf.float32, [None, 32, 32, 3])\n        y_ = tf.placeholder(tf.int32, [None])\n        y_logits = model_function(x_)\n\n        y_dict = dict(labels=y_, logits=y_logits)\n        losses = tf.nn.sparse_softmax_cross_entropy_with_logits(**y_dict)\n        cross_entropy_loss = tf.reduce_mean(losses)\n        trainer = tf.train.AdamOptimizer(learning_rate=0.001)\n        train_op = trainer.minimize(cross_entropy_loss)\n\n        y_pred = tf.argmax(tf.nn.softmax(y_logits), axis=1)\n        correct_prediction = tf.equal(tf.cast(y_pred, tf.int32), y_)\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\n    model_dict = {'graph': g, 'inputs': [x_, y_], 'train_op': train_op,\n                  'accuracy': accuracy, 'loss': cross_entropy_loss}\n\n    return model_dict\n\n\n\n\ndef train_model(model_dict, dataset_generators, epoch_n, print_every):\n    with model_dict['graph'].as_default(), tf.Session() as sess:\n        with tf.device(\"\/gpu:0\"): \n            sess.run(tf.global_variables_initializer())\n\n            for epoch_i in range(epoch_n):\n                for iter_i, data_batch in enumerate(dataset_generators['train']):\n                    train_feed_dict = dict(zip(model_dict['inputs'], data_batch))\n                    sess.run(model_dict['train_op'], feed_dict=train_feed_dict)\n\n                    if iter_i % print_every == 0:\n                        collect_arr = []\n                        for test_batch in dataset_generators['test']:\n                            test_feed_dict = dict(zip(model_dict['inputs'], test_batch))\n                            to_compute = [model_dict['loss'], model_dict['accuracy']]\n                            collect_arr.append(sess.run(to_compute, test_feed_dict))\n                        averages = np.mean(collect_arr, axis=0)\n                        fmt = (epoch_i, iter_i, ) + tuple(averages)\n                        print('epoch {:d} iter {:d}, loss: {:.3f}, '\n                              'accuracy: {:.3f}'.format(*fmt))\n\n\ndataset_generators = {\n        'train': svhn_dataset_generator('train', 256),\n        'test': svhn_dataset_generator('test', 256)\n}\n\n\n\n\n\n\ndef cnn_expanded(x_):\n    conv1 = tf.layers.conv2d(\n            inputs=x_,\n            filters=32,  # number of filters\n            kernel_size=[5, 5],\n            padding=\"same\",\n            activation=tf.nn.relu)\n\n    pool1 = tf.layers.max_pooling2d(inputs=conv1, \n                                    pool_size=[2, 2], \n                                    strides=2)  # convolution stride\n\n    conv2 = tf.layers.conv2d(\n            inputs=pool1,\n            filters=32, # number of filters\n            kernel_size=[9, 9],\n            padding=\"same\",\n            activation=tf.nn.relu)\n    pool2 = tf.layers.max_pooling2d(inputs=conv2, \n                                    pool_size=[2, 2], \n                                    strides=2)  # convolution stride\n\n\n    conv3 = tf.layers.conv2d(\n            inputs=pool1,\n            filters=32,\n            kernel_size=[9,9],\n            padding=\"same\",\n            activation=tf.nn.relu)\n    pool3 = tf.layers.max_pooling2d(inputs=conv2,\n                                   pool_size=[2,2],\n                                   strides=2) #convolution stride\n    pool_flat = tf.contrib.layers.flatten(pool3, scope='pool3flat')\n\n\n    #pool_flat = tf.contrib.layers.flatten(pool2, scope='pool2flat')\n    dense = tf.layers.dense(inputs=pool_flat, units=500, activation=tf.nn.relu)\n    logits = tf.layers.dense(inputs=dense, units=10)\n    return logits\n\n\n\n\nmodel_dict = apply_classification_loss(cnn_expanded)\ntrain_model(model_dict, dataset_generators, epoch_n=50, print_every=20)\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><code>$ python cnn_expansion.py \n  File \"cnn_expansion.py\", line 63\n    with model_dict['graph'].as_default(), tf.Session() as sess:\n                                         ^\nSyntaxError: invalid syntax\n<\/code><\/pre>\n","answers":[{"AnswerId":"52951813","CreationDate":"2018-10-23T14:41:18.187","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>If I understand what you're doing, you need to do <\/p>\n\n<pre><code>with model_dict['graph'].as_default() as g:\n    with tf.Session() as sess:\n<\/code><\/pre>\n\n<p>instead of what you're currently doing. <\/p>\n\n<p>The code you're using is essentially trying to say <\/p>\n\n<pre><code>with (model_dict['graph'].as_default(), tf.Session()) as sess\n<\/code><\/pre>\n\n<p>which <\/p>\n\n<ol>\n<li>python doesn't know how to parse, <\/li>\n<li><p>shouldn't work even if it knew how to parse it as the resulting sess is now a tuple, so things like <\/p>\n\n<p>sess.run() <\/p><\/li>\n<\/ol>\n\n<p>would fail<\/p>\n"}]}
+{"QuestionId":52941888,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2018-10-23T05:44:11.680","AcceptedAnswerId":null,"OwnerUserId":5256563.0,"Title":"Multiple Inputs \/ Outputs for Dataset class in PyTorch","Body":"<p>I've created a Dataset class for PyTorch that can read\/generate batches for problems with <code>Nin<\/code> images as inputs and <code>Nout<\/code> images as outputs. Now I have to pass this class to the DataLoader, but I don't know how to format such inputs \/ outputs in order to be supported by the DataLoader. For example I don't think it has to be <code>BatchSize x Nin x Channels x Height x Width<\/code> because it could not support images with different channels (or dimensions in general).<\/p>\n","answers":[]}
+{"QuestionId":52941982,"AnswerCount":0,"Tags":"<json><tensorflow><tfrecord>","CreationDate":"2018-10-23T05:53:53.647","AcceptedAnswerId":null,"OwnerUserId":5404411.0,"Title":"I am getting an error while converting JSON file to TFRecord","Body":"<p>I am using LabelBox for my upcoming project for image annotation and labeling. I am following <strong>LabelBox<\/strong> tutorial for converting <strong>export.json<\/strong> to TFRecord.  <\/p>\n\n<p>My code snippet: <br>\n<code>import json\ntfrecord_paths = json.load('.\/export.json')['\"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7pxl6r3hx0d3351zn38ae.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6luh8que50d33x3wps35m.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7pahar3b80d333c1trilh.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7ohv2r3280d33df8c1npo.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7ngcmr2rc0d33azbifckb.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7mwlpr2mt0d33oseorjrb.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7m2slszc6081346kdhjuv.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7lk8yr2dy0d338b9c5sir.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7kw9wr28n0d33u7k8s508.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7kc1cr21c0d33blomz3rr.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7joowr1v60d33zdj992nk.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7invdr1ln0d33jialkrca.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7iaw0r1gl0d33l8qtuvbx.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7hr01r1a90d33mh4ey6c4.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7h5vlr15a0d33cspzs8ec.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7gmb2r11c0d3366mddgv7.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7ftcir0qs0d335n8qtgro.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7cz3pqzrr0d33d58ka0jh.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7cfqwqzk30d33xxmd2qly.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7a10pqyxy0d335byj15yz.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng798iwqymf0d332uxr6yn0.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng78gkvqyag0d33k6oewyl2.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng77jf4qxxa0d33ij4ysiv9.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng76ja1qxpt0d33i8qjdhuh.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6x7b5qwfs0d33y5rm5dvx.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6wzy6qwdj0d3369tcjy66.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6wv8cqwc80d33skmptn56.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6vvryqw4m0d33lthmmmc4.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6vdioqw1x0d33g5mowuhs.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6usezqvym0d33fpjaj8ns.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6u81xqvwp0d33g8hw09bj.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6t2ewqvs90d33yy28e8dd.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6st7aqvr50d33zpvxnk95.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6rrn1qvl50d339zzy290r.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6pngyqv8k0d333d6unjav.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6mt5vquma0d33ot8ikfij.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng7qpk4r3ra0d33jw2qth9l.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6ky96qu640d33gk7w6m95.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6k3fcqtor0d33ywbsadge.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6jaa4qtbu0d33erynqglq.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6j6ipqtay0d33d9m0s170.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6j34bqt8o0d33paf4t28j.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6iyjxqt690d3399qi19y4.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6it3vqt4q0d336i0lp0i8.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6hujrqsr80d335huc63do.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6h459qsk40d332dx3q5t1.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6e1ipqrob0d333j6onvvx.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6d88hqrji0d33f48zbazj.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6cecdqred0d33ut1eh0do.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6bmilqrb10d33lo4xs7r9.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6atqtqr6k0d33pnakui4u.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng6a440qr3f0d33hw84l66j.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng68u24qqy50d33wtju9u4q.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng67du8qqt80d33dlhmnush.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng66ywyqqrp0d337zfyywxq.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng65lvjqqgw0d33k2x6cvr6.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng64tt4qqbj0d33bo5itc3x.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4k6x3qj7l0d33rpo02nsy.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4im5tqit60d33c7cts2s4.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4haumsr1808130g9bpftc.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4gnm6qic80d33zpsjxu0j.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4fxyuqi3o0d33tu4fhwf7.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4fa8vqhwp0d33asgvrcic.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4f2ebqhv00d33fom2jsam.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4e08jqhkz0d3336240xbs.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4cc9oqh7k0d33qnphuhyz.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4bewnqh0o0d33ueltn54c.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4admjqgu60d33v2wm1g31.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng4a6spqgt30d3314dwll6y.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng49hfsqgoc0d33dj4ey9b7.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng47yrxqghn0d33nr4ttxk4.tfrecord\", \"gs:\/\/tfrecord-exports\/cjng3yz76qf4v0d335kcjoj0r-2018-10-20T14:24:41-c934d9bf6b92a4a2a6486ab1\/cjng45s6wqg4b0d3334s7ph0s.tfrecord\"']\ntest_set_size = math.ceil(0.2 * len(tfrecord_paths))\ntraining_dataset = (tf.data.TFRecordDataset(tfrecord_paths).skip(test_set_size).map(_parse_tfrecord).apply(tf.contrib.data.shuffle_and_repeat(50)).apply(tf.contrib.data.map_and_batch(_resize(512), 8)))<\/code> <br> <br><\/p>\n\n<p>I am getting an error as :\n<strong>AttributeError: 'str' object has no attribute 'read'<\/strong> <br><br><\/p>\n\n<p>I didn't find a proper documentation for this. Kindly help me.<\/p>\n","answers":[]}
+{"QuestionId":52942038,"AnswerCount":1,"Tags":"<python><machine-learning><keras><recurrent-neural-network><sentiment-analysis>","CreationDate":"2018-10-23T05:58:43.497","AcceptedAnswerId":52942210.0,"OwnerUserId":7942047.0,"Title":"How to pass a string value to a Sentiment Analysis RNN Sequential Model and get back a prediction","Body":"<p>I recreated a sentiment analysis machine learning project using my own data set along with some minor modifications to improve its completion time, I can create good model, compile it, fit it and test it without issues, the problem comes however on how to pass the model a new string\/ article and it in return pass a prediction on whether the string comments are positive or negative and was hoping someone could help me.<\/p>\n\n<p>I posted my code below for your review.<\/p>\n\n<pre><code>class tensor_rnn():\ndef __init__(self, corp_paths, hidden_layers=3, loadfile=True):\n    self.h_layers = hidden_layers\n    self.num_words = []\n    if loadfile == False:\n        data_set = pd.DataFrame(columns=['Article', 'Polarity'])\n        craptopass = []\n        for files in os.listdir(corp_paths[0]):\n            with open(corp_paths[0] + '\\\\' + files, 'r', errors='replace') as text_file:\n                line = text_file.readline().replace('|', '')\n                text_file.close()\n            if len(line.split(' ')) &gt; 3:\n                line = ''.join([i if ord(i) &lt; 128 else ' ' for i in line])\n                craptopass.append([line, 1])\n        good = data_set.append(pd.DataFrame(craptopass, columns=['Article', 'Polarity']), ignore_index=True)\n        data_set = pd.DataFrame(columns=['Article', 'Polarity'])\n        craptopass = []\n        for files in os.listdir(corp_paths[1]):\n            with open(corp_paths[1] + '\\\\' + files, 'r', errors='replace') as text_file:\n                line = text_file.readline().replace('|', '')\n                text_file.close()\n            if len(line.split(' ')) &gt; 3:\n                line = ''.join([i if ord(i) &lt; 128 else ' ' for i in line])\n                craptopass.append([line, -1])\n        bad = data_set .append(pd.DataFrame(craptopass, columns=['Article', 'Polarity']), ignore_index=True)\n        for line in good['Article'].tolist():\n            counter = len(line.split())\n            self.num_words.append(counter)\n\n        for line in bad['Article'].tolist():\n            counter = len(line.split())\n            self.num_words.append(counter)\n        self.features = pd.concat([good, bad]).reset_index(drop=True)\n        # self.features = self.features.str.replace(',', '')\n        self.features.to_csv('Headlines.csv', sep='|')\n    else:\n        self.features = pd.read_csv('Headlines.csv', sep='|')\n        self.features['totalwords'] = self.features['Article'].str.count(' ') + 1\n        self.num_words.extend(self.features['totalwords'].tolist())\n\n    self.features = shuffle(self.features)\n    self.max_len = len(max(self.features['Article'].tolist()))\n    tokenizer = self.tok = preprocessing.text.Tokenizer(num_words=len(self.num_words), split=' ')\n    self.tok.fit_on_texts(self.features['Article'].values)\n    X = tokenizer.texts_to_sequences(self.features['Article'].values)\n    self.X = preprocessing.sequence.pad_sequences(X)\n    self.Y = pd.get_dummies(self.features['Polarity']).values\n    self.X_train, self.X_test, self.Y_train, self.Y_test = train_test_split(self.X, self.Y,\n                                                                            test_size=0.20, random_state=36)\n\ndef RNN(self):\n    embed_dim = 128\n    lstm_out = 128\n    model = Sequential()\n    model.add(Embedding(len(self.num_words), embed_dim, input_length=self.X.shape[1]))\n    model.add(Bidirectional(CuDNNLSTM(lstm_out)))\n    model.add(Dropout(0.2))\n    model.add(Dense(2, activation='softmax'))\n    opt = Adam(lr=0.0001, decay=1e-4)   #1e-3\n    model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])\n    return model\n\ndef model_train(self):\n    self.model = self.RNN()\n\ndef model_test(self):\n    batch_size = 128\n    self.model.fit(self.X_train, self.Y_train, epochs=4, batch_size=batch_size, verbose=2,\n                                callbacks=[EarlyStopping(monitor='val_loss', min_delta=0.0001,\n                                                         patience=5, verbose=2, mode='auto')], validation_split=0.2)\n\n\nif __name__ == \"__main__\":\n    paths = 'PATHS TO ARTICLES'\n    a = tensor_rnn([paths + '\\\\pos', paths + '\\\\neg'])\n    a.model_train()\n    a.model_test()\n    a.model.save('RNNModelArticles.h5', include_optimizer=True)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52942210","CreationDate":"2018-10-23T06:11:23.040","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>All you need to do is preprocess the new text that you want to feed to the model the same way you preprocessed text for the training. After that, you should have a predict method that will output it's prediction in the same way the model outputs prediction in the training. So, in the predict method you should write something like:<\/p>\n\n<pre><code>def predict(self, sequence):\n  presprocessed = preprocess(sequence)\n  prediction = self.model.predict(preprocessed, batch_size=None, verbose=0, steps=None)\n<\/code><\/pre>\n\n<p>Does this clarify things for you?<\/p>\n"}]}
+{"QuestionId":52943489,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-10-23T07:32:41.827","AcceptedAnswerId":null,"OwnerUserId":988709.0,"Title":"what is XLA_GPU and XLA_CPU for tensorflow","Body":"<p>I can list gpu devices sing the following tensorflow code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.client import device_lib\nprint device_lib.list_local_devices()\n<\/code><\/pre>\n\n<p>The result is:<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\n device_type: \"CPU\"\n memory_limit: 268435456\n locality {\n }\n incarnation: 17897160860519880862, name: \"\/device:XLA_GPU:0\"\n device_type: \"XLA_GPU\"\n memory_limit: 17179869184\n locality {\n }\n incarnation: 9751861134541508701\n physical_device_desc: \"device: XLA_GPU device\", name: \"\/device:XLA_CPU:0\"\n device_type: \"XLA_CPU\"\n memory_limit: 17179869184\n locality {\n }\n incarnation: 5368380567397471193\n physical_device_desc: \"device: XLA_CPU device\", name: \"\/device:GPU:0\"\n device_type: \"GPU\"\n memory_limit: 21366299034\n locality {\n   bus_id: 1\n   links {\n     link {\n       device_id: 1\n       type: \"StreamExecutor\"\n       strength: 1\n     }\n   }\n }\n incarnation: 7110958745101815531\n physical_device_desc: \"device: 0, name: Tesla P40, pci bus id: 0000:02:00.0, compute capability: 6.1\", name: \"\/device:GPU:1\"\n device_type: \"GPU\"\n memory_limit: 17336821351\n locality {\n   bus_id: 1\n   links {\n     link {\n       type: \"StreamExecutor\"\n       strength: 1\n     }\n   }\n }\n incarnation: 3366465227705362600\n physical_device_desc: \"device: 1, name: Tesla P40, pci bus id: 0000:03:00.0, compute capability: 6.1\", name: \"\/device:GPU:2\"\n device_type: \"GPU\"\n memory_limit: 22590563943\n locality {\n   bus_id: 2\n   numa_node: 1\n   links {\n     link {\n       device_id: 3\n       type: \"StreamExecutor\"\n       strength: 1\n     }\n   }\n }\n incarnation: 8774017944003495680\n physical_device_desc: \"device: 2, name: Tesla P40, pci bus id: 0000:83:00.0, compute capability: 6.1\", name: \"\/device:GPU:3\"\n device_type: \"GPU\"\n memory_limit: 22590563943\n locality {\n   bus_id: 2\n   numa_node: 1\n   links {\n     link {\n       device_id: 2\n       type: \"StreamExecutor\"\n       strength: 1\n     }\n   }\n }\n incarnation: 2007348906807258050\n physical_device_desc: \"device: 3, name: Tesla P40, pci bus id: 0000:84:00.0, compute capability: 6.1\"]\n<\/code><\/pre>\n\n<p>I want to know what is <code>XLA_GPU<\/code> and <code>XLA_CPU<\/code>?<\/p>\n","answers":[{"AnswerId":"58487325","CreationDate":"2019-10-21T13:25:53.007","ParentId":null,"OwnerUserId":"1631855","Title":null,"Body":"<p>As mentioned in the docs, XLA stands for \"accelerated linear algebra\". It's Tensorflow's relatively new optimizing compiler that can further speed up your ML models' GPU operations by combining what used to be multiple CUDA kernels into one (simplifying because this isn't that important for your question).<\/p>\n\n<p>To your question, my understanding is that XLA is separate enough from the default Tensorflow compiler that they separately register GPU devices and have slightly different constraints on which GPUs they treat as visible (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/30748\" rel=\"nofollow noreferrer\">here<\/a> for more on this). Looking at the output of the command you ran, it looks like XLA is registering 1 GPU and normal TF is registering 3. <\/p>\n\n<p>I'm not sure if you're having issues or are just curious, but if it's the former, I recommend taking a look at the issue I linked above and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/30388\" rel=\"nofollow noreferrer\">this one<\/a>. Tensorflow's finicky about which CUDA\/cuDNN versions with which it works flawlessly and it's possible you're using incompatible versions. (If you're not having issues, then hopefully the first part of my answer is sufficient.) <\/p>\n"}]}
+{"QuestionId":52943866,"AnswerCount":1,"Tags":"<python-3.x><keras><google-cloud-platform><google-cloud-pubsub>","CreationDate":"2018-10-23T07:54:33.777","AcceptedAnswerId":null,"OwnerUserId":10544741.0,"Title":"How to call a global variable from a pubsub callback function in google-cloud-platform vminstance?","Body":"<p>I'm running a neural network model that supposed to run in callback function every time asynchronously when a pubsub event is triggered. I loaded the model as global variable, during the initialization of main program, so that it can access in any function and to avoid model loading every time in callback function itself.\nTypical code can be:  <\/p>\n\n<pre><code>global cv_model\ncv_model = load_model('my_model.h5')\n\ndef callback():\n    prediction = cv_model.predict(image)\n<\/code><\/pre>\n\n<p>This method works perfectly in my local machine. But when I tried with pubsub event call back I'm getting some errors. After some search I figured out 'pubsub is a child-process of the main program and we cannot share global variables between processes'. I tried nested-function approach and OOPs approach but none of them worked. <\/p>\n\n<p>How can I avoid model loading everytime when callback function being called?\nAny alternative methods to resolve this conflict?<\/p>\n\n<p>Thanks for the suggestions!!<\/p>\n","answers":[{"AnswerId":"52955356","CreationDate":"2018-10-23T18:11:54.147","ParentId":null,"OwnerUserId":"328036","Title":null,"Body":"<p>If you can't achieve this with the Pub\/Sub callback, you could try writing a <a href=\"https:\/\/cloud.google.com\/functions\/docs\/calling\/pubsub#functions-deploy-command-python\" rel=\"nofollow noreferrer\">Google Cloud Function with a Pub\/Sub Trigger<\/a> instead, which doesn't have this problem:<\/p>\n\n<pre><code>from keras.models import load_model\ncv_model = load_model('my_model.h5')\n\ndef handle_pubsub(data, context):\n    image = ...  # do something with the `data` arg here\n    prediction = cv_model.predict(image)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52944529,"AnswerCount":1,"Tags":"<python-3.x><pytorch><onnx>","CreationDate":"2018-10-23T08:30:30.610","AcceptedAnswerId":null,"OwnerUserId":1322801.0,"Title":"Problem loading Pytourch 3.0 model unexpected key \"module.features.0.weight\" in state_dict","Body":"<p>I am trying to load a model I have trained using Pytorch,\nbut I keep getting the following error:<\/p>\n\n<blockquote>\n  <p>File \"convert.py\", line 12, in \n      model.load_state_dict(torch.load('model\/model_vgg2d_2.pth'))   File\n  \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\",\n  line 490, in load_state_dict\n      .format(name)) KeyError: 'unexpected key \"module.features.0.weight\" in state_dict'<\/p>\n<\/blockquote>\n\n<p>Below is my code:<\/p>\n\n<pre><code>import torch.onnx\nimport torch.nn as nn\n\nclass TempModel(nn.Module):\n    def __init__(self):\n        super(TempModel, self).__init__()\n        self.conv1 = nn.Conv2d(3, 5, (3, 3))\n    def forward(self, inp):\n        return self.conv1(inp)\n\nmodel = nn.DataParallel(TempModel())\nmodel.load_state_dict(torch.load('model\/model_vgg2d_2.pth'))\ndummy_input = Variable(torch.randn(1, 3, 224, 224))\ntorch.onnx.export(model, dummy_input, \"model_onnx\/model_vgg2d_0.onnx\")\n<\/code><\/pre>\n\n<p>I am working on the same machine that I have used to train the model(which has multiple GPUs).\nAny ideas what am I doing wrong? <\/p>\n","answers":[{"AnswerId":"52945177","CreationDate":"2018-10-23T09:05:19.913","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When loading <code>state_dict<\/code> you need it to be a <code>state_dict<\/code> of the <strong>same<\/strong> model: you cannot load a <code>state_dict<\/code> of a VGG model into a completely different <code>BasicModel<\/code>. <\/p>\n\n<hr>\n\n<p><strong>old answer<\/strong><br>\nYou saved the model without <code>nn.DataParallel<\/code> applied to the model and now you are trying to load after adding this. Try<\/p>\n\n<pre><code>model = TempModel()\nmodel.load_state_dict(torch.load('model\/model_vgg2d_2.pth'))\nmodel = nn.DataParallel(model)  # parallel AFTER load\n<\/code><\/pre>\n"}]}
+{"QuestionId":52945010,"AnswerCount":0,"Tags":"<python><keras><deep-learning><classification>","CreationDate":"2018-10-23T08:56:36.863","AcceptedAnswerId":null,"OwnerUserId":10341369.0,"Title":"Keras binary image classifier returns 1 class even with training set","Body":"<p>Im currently building a binary image classifier using keras with Python.\nThe data are split evenly for both training and test.<\/p>\n\n<p>I've been having problem with it with an accuracy of 50%, which means that it was guessing one class during training.<\/p>\n\n<p>I introduced gradient clipping, BatchNormalization with momentum, and Leaky Relu to solve some of the issues. With these, I got a classifier with high training accuracy (Realms of 95%) and low loss(Around 0.05).<\/p>\n\n<p>The val_acc is usually 0.1 lower than training accuracy and a val_loss of around 0.3. This suggests overfitting.<\/p>\n\n<p>I decided to just try to predict with the training images, where they should at least predict those correctly in any case. However, it only predicts 1 class even on training set. <\/p>\n\n<p>Below is my model:<\/p>\n\n<pre><code>    classifier = Sequential()\n\n\n    classifier.add(Conv2D(8, (5, 5), input_shape = (256, 204, 3), activation = 'linear', padding = 'same'))\n    classifier.add(LeakyReLU(alpha=0.001))\n    classifier.add(BatchNormalization(momentum = 0.1))\n    classifier.add(MaxPooling2D(pool_size = (2, 2)))\n    classifier.add(Dropout(0.5))\n\n    classifier.add(Conv2D(8, (5, 5), activation = 'linear', padding = 'same'))\n    classifier.add(LeakyReLU(alpha=0.001))\n    classifier.add(BatchNormalization(momentum=0.1))\n    classifier.add(MaxPooling2D(pool_size = (2, 2)))\n    classifier.add(Dropout(0.5))\n\n    classifier.add(Flatten())\n\n\n   classifier.add(Dense(units = 128, activation = 'linear'))\n   classifier.add(LeakyReLU(alpha=0.001))\n   classifier.add(Dense(units = 1, activation = 'sigmoid'))\n\n\n   adam = optimizers.Adam(lr = 0.001, decay = 1e-2, clipvalue = 1, clipnorm = 1)\n  classifier.compile(optimizer = adam, loss = 'binary_crossentropy', metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>I've made the model as simple as possible to prevent overfitting. \nMy images are to classifier is there are droplets of water on a piece of cloth. Which looks a bit like blood stains on a wall.<\/p>\n\n<p>Does anyone have any idea why this is the case, or what should I try next?<\/p>\n","answers":[]}
+{"QuestionId":52945355,"AnswerCount":0,"Tags":"<python><tensorflow><keras><keras-2>","CreationDate":"2018-10-23T09:14:16.307","AcceptedAnswerId":null,"OwnerUserId":10541533.0,"Title":"Tensorflow Keras Model working on CPU but doesn't work with GPU","Body":"<p>I am beginner in <code>Tensorflow<\/code> and <code>Keras<\/code>.<\/p>\n\n<p>I ran the code on my CPU with the following configuration and it works completely fine:<\/p>\n\n<pre><code>Tensorflow version: 1.10.0\ntf.keras version 2.1.6\n<\/code><\/pre>\n\n<p>But when I try to run the code on a GPU server with the following configuration:<\/p>\n\n<pre><code>Tensorflow(GPU) version: 1.8.0\ntf.keras version 2.1.5\n<\/code><\/pre>\n\n<p>I receive this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"Keras3 tut.py\", line 127, in &lt;module&gt;\n    history = model.fit(ds_train, steps_per_epoch = 1500, epochs = 15)\n  File \"\/usr\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training.py\", line 1143, in fit\n    batch_size=batch_size)\n  File \"\/usr\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training.py\", line 765, in _standardize_user_data\n    exception_prefix='input')\n  File \"\/usr\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training_utils.py\", line 150, in standardize_input_data\n    data = [standardize_single_array(x) for x in data]\n  File \"\/usr\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training_utils.py\", line 150, in &lt;listcomp&gt;\n    data = [standardize_single_array(x) for x in data]\n  File \"\/usr\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/_impl\/keras\/engine\/training_utils.py\", line 88, in standardize_single_array\n    elif x.ndim == 1:\nAttributeError: 'BatchDataset' object has no attribute 'ndim'\n<\/code><\/pre>\n\n<p>The code is as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nfrom tensorflow import FixedLenFeature\n\nfeatdef = {\n    'train\/image': FixedLenFeature(shape=[], dtype= tf.string),\n    'train\/label': FixedLenFeature(shape = [], dtype = tf.int64)\n}\n\ndef _parse_record (example_proto, clip = False):\n    ex = tf.parse_single_example(example_proto, featdef)\n\n    im = tf.decode_raw(ex['train\/image'], tf.uint8)\n    im = tf.reshape(im, (28,28,3))\n    im = tf.cast(im, tf.float32)*(1.\/255)\n\n    #label = (ex['train\/label'])\n    label = tf.one_hot(ex['train\/label'], 2)\n    print(label)\n    print(im)\n    return im, label\n\nbatch_size = 50\nds_train = tf.data.TFRecordDataset('\/home\/mudit\/AI\/TFrecord_Datasets\/Melanoma_training_uint8.tfrecords').map(_parse_record)\nds_train = ds_train.repeat().shuffle(1000).batch(batch_size)\n\nIM_SIZE = 28\n\nimage_input = tf.keras.Input(shape = (IM_SIZE, IM_SIZE, 3), name = 'input_layer')\n\nconv_1 = tf.keras.layers.Conv2D(32, kernel_size = (3,3),\n                              padding = 'same',\n                              activation = 'relu')(image_input)\n\nconv1 = tf.keras.layers.MaxPooling2D(padding = 'same')(conv_1)\n\n\nconv_2 = tf.keras.layers.Conv2D(32, kernel_size = (3,3), padding = 'same',\n                               activation = 'relu')(conv_1)\n\nconv_flat = tf.keras.layers.Flatten()(conv_2)\n\nfc_1 = tf.keras.layers.Dense(28,\n                             activation='relu')(conv_flat)\nfc_1 = tf.keras.layers.Dropout(0.2)(fc_1)\n\nfc_2 = tf.keras.layers.Dense(28,\n                             activation='relu')(fc_1)\n\nfc_2 = tf.keras.layers.Dropout(0.4)(fc_2)\n\nlabel_output = tf.keras.layers.Dense(2,\n                                       activation='softmax',\n                                       name='label')(fc_2)\n\nmodel = tf.keras.Model(inputs = image_input, outputs = [label_output])\n\nprint(model.summary())\n\nmodel.compile(optimizer = 'adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\nhistory = model.fit(ds_train, steps_per_epoch = 500, epochs = 8)\n<\/code><\/pre>\n\n<p>On the GPU, it gives error at <code>model.fit()<\/code>.<\/p>\n\n<p>The code is working completely fine on my CPU but not on the GPU server. \nIs it possible that the problem is because of the version difference? I tried finding the solutions for it but to no avail.<\/p>\n\n<p>EDIT 1:\nI tried running the same code on my PC using the same configuration as GPU server (Tensorflow 1.8.0 Keras 2.1.5) and it gives me exactly the same error. So, it's definitely a version problem. I reckon, there must've been some modifications in the tf.data API but can't find out what.<\/p>\n","answers":[]}
+{"QuestionId":52945427,"AnswerCount":1,"Tags":"<python><lstm><pytorch><rnn>","CreationDate":"2018-10-23T09:17:55.817","AcceptedAnswerId":53195257.0,"OwnerUserId":639973.0,"Title":"PyTorch: manually setting weight parameters with numpy array for GRU \/ LSTM","Body":"<p>I'm trying to fill up GRU\/LSTM with manually defined parameters in pytorch.<\/p>\n\n<p>I have numpy arrays for parameters with shapes as defined in their documentation (<a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.GRU\" rel=\"noreferrer\">https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.GRU<\/a>).<\/p>\n\n<p>It seems to work but I'm not sure whether the returned values are correct.<\/p>\n\n<p>Is this a right way to fill up GRU\/LSTM with numpy parameters?<\/p>\n\n<pre><code>gru = nn.GRU(input_size, hidden_size, num_layers,\n              bias=True, batch_first=False, dropout=dropout, bidirectional=bidirectional)\n\ndef set_nn_wih(layer, parameter_name, w, l0=True):\n    param = getattr(layer, parameter_name)\n    if l0:\n        for i in range(3*hidden_size):\n            param.data[i] = w[i*input_size:(i+1)*input_size]\n    else:\n        for i in range(3*hidden_size):\n            param.data[i] = w[i*num_directions*hidden_size:(i+1)*num_directions*hidden_size]\n\ndef set_nn_whh(layer, parameter_name, w):\n    param = getattr(layer, parameter_name)\n    for i in range(3*hidden_size):\n        param.data[i] = w[i*hidden_size:(i+1)*hidden_size]\n\nl0=True\n\nfor i in range(num_directions):\n    for j in range(num_layers):\n        if j == 0:\n            wih = w0[i, :, :3*input_size]\n            whh = w0[i, :, 3*input_size:]  # check\n            l0=True\n        else:\n            wih = w[j-1, i, :, :num_directions*3*hidden_size]\n            whh = w[j-1, i, :, num_directions*3*hidden_size:]\n            l0=False\n\n        if i == 0:\n            set_nn_wih(\n                gru, \"weight_ih_l{}\".format(j), torch.from_numpy(wih.flatten()),l0)\n            set_nn_whh(\n                gru, \"weight_hh_l{}\".format(j), torch.from_numpy(whh.flatten()))\n        else:\n            set_nn_wih(\n                gru, \"weight_ih_l{}_reverse\".format(j), torch.from_numpy(wih.flatten()),l0)\n            set_nn_whh(\n                gru, \"weight_hh_l{}_reverse\".format(j), torch.from_numpy(whh.flatten()))\n\ny, hn = gru(x_t, h_t)\n<\/code><\/pre>\n\n<p>numpy arrays are defined as following:<\/p>\n\n<pre><code>rng = np.random.RandomState(313)\nw0 = rng.randn(num_directions, hidden_size, 3*(input_size +\n               hidden_size)).astype(np.float32)\nw = rng.randn(max(1, num_layers-1), num_directions, hidden_size,\n              3*(num_directions*hidden_size + hidden_size)).astype(np.float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53195257","CreationDate":"2018-11-07T18:06:06.333","ParentId":null,"OwnerUserId":"4385912","Title":null,"Body":"<p>\nThat is a good question, and you already give a decent answer. However, it reinvents the wheel - there is a very elegant Pytorch internal routine that will allow you to do the same without as much effort - and one that is applicable for any network.<\/p>\n\n<p>The core concept here is PyTorch's <code>state_dict<\/code>. The state dictionary effectively contains the <code>parameters<\/code> organized by the tree-structure given by the relationship of the <code>nn.Modules<\/code> and their submodules, etc.<\/p>\n\n<h3>The short answer<\/h3>\n\n<p>If you only want the code to load a value into a tensor using the <code>state_dict<\/code>, then try this line (where the <code>dict<\/code> contains a valid <code>state_dict<\/code>):<\/p>\n\n<pre><code>`model.load_state_dict(dict, strict=False)`\n<\/code><\/pre>\n\n<p>where <code>strict=False<\/code> is crucial if you want to load <em>only some parameter values<\/em>.<\/p>\n\n<h3>The long answer - including an introduction to PyTorch's <code>state_dict<\/code><\/h3>\n\n<p>Here's an example of how a state dict looks for a GRU (I chose <code>input_size = hidden_size = 2<\/code> so that I can print the entire state dict):<\/p>\n\n<pre><code>rnn = torch.nn.GRU(2, 2, 1)\nrnn.state_dict()\n# Out[10]: \n#     OrderedDict([('weight_ih_l0', tensor([[-0.0023, -0.0460],\n#                         [ 0.3373,  0.0070],\n#                         [ 0.0745, -0.5345],\n#                         [ 0.5347, -0.2373],\n#                         [-0.2217, -0.2824],\n#                         [-0.2983,  0.4771]])),\n#                 ('weight_hh_l0', tensor([[-0.2837, -0.0571],\n#                         [-0.1820,  0.6963],\n#                         [ 0.4978, -0.6342],\n#                         [ 0.0366,  0.2156],\n#                         [ 0.5009,  0.4382],\n#                         [-0.7012, -0.5157]])),\n#                 ('bias_ih_l0',\n#                 tensor([-0.2158, -0.6643, -0.3505, -0.0959, -0.5332, -0.6209])),\n#                 ('bias_hh_l0',\n#                 tensor([-0.1845,  0.4075, -0.1721, -0.4893, -0.2427,  0.3973]))])\n<\/code><\/pre>\n\n<p>So the <code>state_dict<\/code> all the parameters of the network. If we have \"nested\" <code>nn.Modules<\/code>, we get the tree represented by the parameter names:<\/p>\n\n<pre><code>class MLP(torch.nn.Module):      \n    def __init__(self):\n        torch.nn.Module.__init__(self)\n        self.lin_a = torch.nn.Linear(2, 2)\n        self.lin_b = torch.nn.Linear(2, 2)\n\n\nmlp = MLP()\nmlp.state_dict()\n#    Out[23]: \n#        OrderedDict([('lin_a.weight', tensor([[-0.2914,  0.0791],\n#                            [-0.1167,  0.6591]])),\n#                    ('lin_a.bias', tensor([-0.2745, -0.1614])),\n#                    ('lin_b.weight', tensor([[-0.4634, -0.2649],\n#                            [ 0.4552,  0.3812]])),\n#                    ('lin_b.bias', tensor([ 0.0273, -0.1283]))])\n\n\nclass NestedMLP(torch.nn.Module):\n    def __init__(self):\n        torch.nn.Module.__init__(self)\n        self.mlp_a = MLP()\n        self.mlp_b = MLP()\n\n\nn_mlp = NestedMLP()\nn_mlp.state_dict()\n#   Out[26]: \n#        OrderedDict([('mlp_a.lin_a.weight', tensor([[ 0.2543,  0.3412],\n#                            [-0.1984, -0.3235]])),\n#                    ('mlp_a.lin_a.bias', tensor([ 0.2480, -0.0631])),\n#                    ('mlp_a.lin_b.weight', tensor([[-0.4575, -0.6072],\n#                            [-0.0100,  0.5887]])),\n#                    ('mlp_a.lin_b.bias', tensor([-0.3116,  0.5603])),\n#                    ('mlp_b.lin_a.weight', tensor([[ 0.3722,  0.6940],\n#                            [-0.5120,  0.5414]])),\n#                    ('mlp_b.lin_a.bias', tensor([0.3604, 0.0316])),\n#                    ('mlp_b.lin_b.weight', tensor([[-0.5571,  0.0830],\n#                            [ 0.5230, -0.1020]])),\n#                    ('mlp_b.lin_b.bias', tensor([ 0.2156, -0.2930]))])\n<\/code><\/pre>\n\n<p>So - what if you want to not extract the state dict, but change it - and thereby the network's parameters? Use <code>nn.Module.load_state_dict(state_dict, strict=True)<\/code> (<a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.load_state_dict\" rel=\"nofollow noreferrer\">link to the docs<\/a>)\nThis method allows you to load an entire state_dict with arbitrary values <em>into an instantiated model of the same kind<\/em> as long as the keys (i.e. the parameter names) are correct and the values (i.e. the parameters) are <code>torch.tensors<\/code> of the right shape.\nIf the <code>strict<\/code> kwarg is set to <code>True<\/code> (the default), the dict you load has to exactly match the original state dict, except for the values of the parameters. That is, there has to be one new value for each parameter.<\/p>\n\n<p>For the GRU example above, we need a tensor of the correct size (and the correct device, btw) for each of <code>'weight_ih_l0', 'weight_hh_l0', 'bias_ih_l0', 'bias_hh_l0'<\/code>. As we sometimes only want to load <em>some<\/em> values (as I think you want to do), we can set the <code>strict<\/code> kwarg to <code>False<\/code> - and we can then load only partial state dicts, as e.g. one that only contains parameter values for <code>'weight_ih_l0'<\/code>.<\/p>\n\n<p>As a practical advice, I'd simply create the model you want to load values into, and then print the state dict (or at least a list of the keys and the respective tensor sizes)<\/p>\n\n<pre><code>print([k, v.shape for k, v in model.state_dict().items()])\n<\/code><\/pre>\n\n<p>That tells you what the exact name of the parameter is you want to change. You then simply create a state dict with the respective parameter name and tensor, and load it:<\/p>\n\n<pre><code>from dollections import OrderedDict\nnew_state_dict = OrderedDict({'tensor_name_retrieved_from_original_dict': new_tensor_value})\nmodel.load_state_dict(new_state_dict, strict=False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52946030,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-23T09:46:24.803","AcceptedAnswerId":52947744.0,"OwnerUserId":7934206.0,"Title":"Where is RNN is taking batch size from?","Body":"<p>I am training a RNN by by below :<\/p>\n\n<pre><code>def create_rnn_model(stateful,length):\n    model = Sequential()\n    model.add(SimpleRNN(20,return_sequences=False,stateful=stateful,batch_input_shape=(1,length,1)))\n    adam = optimizers.Adam(lr=0.001)\n    model.add(Dense(1))\n    model.compile(loss='mean_absolute_error', optimizer=adam, metrics=[root_mean_squared_error])\n    print(model.summary())\n    return model\n<\/code><\/pre>\n\n<p>and the fit  by <\/p>\n\n<pre><code>model_info = model_rnn_stateful.fit(x=x_train, y=y_train, validation_data=(x_test, y_test), batch_size=1, epochs=10,verbose=1)\n<\/code><\/pre>\n\n<p>and predict by<\/p>\n\n<pre><code>predicted_rnn_stateful = model_rnn_stateful.predict(x_test)\n<\/code><\/pre>\n\n<p>But when I predict it throws an error <\/p>\n\n<blockquote>\n  <p>ValueError: In a stateful network, you should only pass inputs with a\n  number of samples that can be divided by the batch size. Found: 200\n  samples. Batch size: 32.<\/p>\n<\/blockquote>\n\n<p>There is no where I specify 32. I don't know where it is coming from.My batch size is 1 only.Any help is appreciated.<\/p>\n\n<p><strong>EDIT<\/strong>\n I have no breakpoints used in my script\/IDE.Thank you <\/p>\n","answers":[{"AnswerId":"52947744","CreationDate":"2018-10-23T11:20:15.933","ParentId":null,"OwnerUserId":"10494794","Title":null,"Body":"<p>From <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">Keras documentation<\/a><\/p>\n\n<blockquote>\n  <ul>\n  <li><strong>batch_size<\/strong>: Integer or None. Number of samples per gradient update. If unspecified, batch_size will default to 32.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>1 is probably an incorrect value for batch_size, then it took the default value which is 32. Try with 2 or 20 as batch_size<\/p>\n"}]}
+{"QuestionId":52946110,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><image-segmentation>","CreationDate":"2018-10-23T09:51:51.043","AcceptedAnswerId":54031459.0,"OwnerUserId":8375248.0,"Title":"U-net low contrast test images, predict output is grey box","Body":"<p>I am running the unet from <a href=\"https:\/\/github.com\/zhixuhao\/unet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zhixuhao\/unet<\/a> but when I run the unet the predicted images are all grey. I get an error saying low contrast image for my test data, any one had or resolved this problem?<\/p>\n\n<p>I am training with 50 ultrasound images and get around 2000\/3000 after augmentation, on 5 epochs with 300 steps per epoch and batch size of 2.<\/p>\n\n<p>Many thanks in advance\nHelena<\/p>\n","answers":[{"AnswerId":"54031459","CreationDate":"2019-01-04T00:00:22.090","ParentId":null,"OwnerUserId":"5763590","Title":null,"Body":"<blockquote>\n  <p>After you made sure that your data pipeline is correct. There are a few things to consider here, I hope one of the bellow mentioned helps:<\/p>\n<\/blockquote>\n\n<p><strong>1. Choose the right loss function<\/strong>\nBinary crossentropy might lead your network in the direction of optimizing for all labels, now if you have an unbalanced amount of labels in your image, it might draw your network to just give back either white, gray or black image predictions. Try using the dice coefficient loss<\/p>\n\n<p><strong>2. Change the line in testGenerator<\/strong>\nA thing that seems to be an issue in <code>data.py<\/code> and the <code>testGenerator<\/code> method is    the following line:<\/p>\n\n<pre><code>img = img \/ 255\n<\/code><\/pre>\n\n<p>Change it to:<\/p>\n\n<pre><code>img \/=255. \n<\/code><\/pre>\n\n<p><strong>3. Reduce learning rate<\/strong>\nif your learning rate is too high you might converge in non-sufficient optima, which also tend to optimize for gray, black or white predictions only.\nTry a learning rate around <code>Adam(lr = 3e-5)<\/code> and train for a sufficient amount of epochs, you should print dice loss and not accuracy to check your convergence.<\/p>\n\n<p><strong>4. Do not use activation functions for the last set of convolutions<\/strong>\nFor the last set of convolutions, that is 128-> 64 -> 64 -> 1, the activation function should not be used! The activation function causes the values to vanish!<\/p>\n\n<p><strong>5. Your saving method could have a \"bug\"<\/strong> make sure you scale your image to values between 0 and 255 before saving. Skimage usually warns you with a low contrast image warning.<\/p>\n\n<pre><code>from skimage import img_as_uint\n\nio.imsave(os.path.join(save_path,\"%d_predict.tif\"%(i)),img_as_uint(img))\n<\/code><\/pre>\n\n<p><strong>6. Your saving format could have a \"bug\"<\/strong> make sure you save your image in a proper format. I experienced that saving as .png gives only black or gray images, whereas .tif works like a charm.<\/p>\n\n<p><strong>7. You might just not train enough<\/strong> often you'll just freak out when your network does not do as you would like it to and abort the training. Chance is, additional training epochs is exactly what it would have needed.<\/p>\n"}]}
+{"QuestionId":52946349,"AnswerCount":1,"Tags":"<python><conv-neural-network><pytorch>","CreationDate":"2018-10-23T10:04:57.680","AcceptedAnswerId":null,"OwnerUserId":9316652.0,"Title":"How to efficiently normalize group of elements in a tensor","Body":"<h3>Context:<\/h3>\n\n<p>I am trying to replicate Hinton's \"Matrix capsules with EM routing\" (<a href=\"https:\/\/openreview.net\/forum?id=HJWLfGWRb\" rel=\"nofollow noreferrer\">https:\/\/openreview.net\/forum?id=HJWLfGWRb<\/a>).<\/p>\n\n<p>At some point, there are convolutional operations that are performed (in the sense that an output tensor is connected to an input tensor, and each element in the output tensor is influenced only by the input elements contained in a 2D mask of size K).<\/p>\n\n<p><strong>Input<\/strong>\nTensor <code>x<\/code> of shape <code>w_in,w_in<\/code>\nwhere<\/p>\n\n<ul>\n<li><code>w_in=14<\/code><\/li>\n<\/ul>\n\n<p><strong>Intermediary tensor mapped to the input<\/strong>\nTensor <code>x_mapped<\/code> of shape <code>w_out,w_out,K,K<\/code>\nwhere<\/p>\n\n<ul>\n<li><code>K=3<\/code> is the convolution kernel size<\/li>\n<li><code>w_out=6<\/code>, resulting from convolution with <code>stride=2<\/code><\/li>\n<\/ul>\n\n<p>Summing on the dimensions 2 and 3 (both of size <code>K<\/code>) means summing on the input elements connected to an output element whose location is given by dimensions 0 and 1.<\/p>\n\n<h3>Question:<\/h3>\n\n<p>How can I efficiently normalize (to 1) groups of elements in <code>x_mapped<\/code>, based on their location in the input tensor <code>x<\/code>?<\/p>\n\n<p>For example:<br>\n<code>x_mapped(0,0,2,2)<\/code><br>\n<code>x_mapped(1,0,0,2)<\/code><br>\n<code>x_mapped(0,1,2,0)<\/code><br>\n<code>x_mapped(1,1,0,0)<\/code><br>\nare all connected to <code>x(2,2)<\/code> (the formula is <code>i_out*stride + K_index = i_in<\/code>). For that reason, I would like the sum of those 4 elements to be 1. <\/p>\n\n<p>And I would like to do that for all the groups of elements in <code>x_mapped<\/code> that are \"connected\" to the same element in <code>x<\/code>.<\/p>\n\n<p>I can figure out how to do it by:<\/p>\n\n<ol>\n<li>Building a dictionary with input location as key and list of output elements as value<\/li>\n<li>Looping on the dictionary, summing the elements in the list for a given input location and dividing them by that sum<\/li>\n<\/ol>\n\n<p>but that seems really inefficient to me.<\/p>\n","answers":[{"AnswerId":"52963736","CreationDate":"2018-10-24T07:59:13.810","ParentId":null,"OwnerUserId":"9316652","Title":null,"Body":"<p>I solved this in the following way:<\/p>\n\n<ol>\n<li>Creating a dictionary with a 2-tuple as key (coordinates in <code>x<\/code>) and a list of elements of <code>x_mapped<\/code> as values.   <\/li>\n<li>One loop over the dictionary, zipping all the elements of one dictionary item, then normalizing.  <\/li>\n<\/ol>\n\n<p>Here is the code:<\/p>\n\n<pre><code>from collections import defaultdict\nimport torch\n\nho = 6\nwo = 6\nstride = 2\nK = 3\n\nd = defaultdict(list)\n\nx_mapped = torch.arange(0,ho*wo*K*K).view(ho,wo,K,K).type(dtype = torch.DoubleTensor)\n\nfor i_out in range(0,ho):\n    for j_out in range(0,wo):\n        for K_i in range(0,K):\n            for K_j in range(0, K):\n                i_in = i_out * stride + K_i\n                j_in = j_out * stride + K_j\n\n                d[(i_in, j_in)].append((i_out, j_out, K_i, K_j))\n\nfor _ , value in d.items():\n    ho_list, wo_list, K_i_list, K_j_list = zip(*value)\n    x_mapped[ho_list, wo_list, K_i_list, K_j_list] = x_mapped[ho_list, wo_list, K_i_list, K_j_list] \/ torch.sum(\n        x_mapped[ho_list, wo_list, K_i_list, K_j_list])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52946435,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-10-23T10:09:10.160","AcceptedAnswerId":53375322.0,"OwnerUserId":7644710.0,"Title":"Pytorch (0.4.1) can not find GPU (Nvidia V100)","Body":"<p>I have installed Pytorch(0.4.1) with CUDA 9.0 via anaconda.<br>\nI also have installed Nvidia driver (390.30).<br>\nThrough nvidia-smi, I can see all GPUs(V100) with their situation.<br>\nBut I always get False when calling torch.cuda.is_avaliable().<br>\nI also tried to downgrade the CUDA version from 9.0 to 8.0, but the situation was just the same.<br>\nTried to install Pytorch again but nothing changed. Could anyone give me a hand, please?<\/p>\n","answers":[{"AnswerId":"53375322","CreationDate":"2018-11-19T13:06:26.323","ParentId":null,"OwnerUserId":"7644710","Title":null,"Body":"<p>It seems that it's a Nvidia container issue.<br>\nAfter downgrading the Nvidia container from 18.08 to 18.07, all code runs well.<\/p>\n"},{"AnswerId":"52960614","CreationDate":"2018-10-24T03:18:54.177","ParentId":null,"OwnerUserId":"2846062","Title":null,"Body":"<p>It's better to let anaconda handle everything by itself. For the time being, I have pytorch version 0.2, cudatoolkit version 7.5, cudnn version 5.1. It works fine. You can try <\/p>\n\n<pre><code>conda update --all\n<\/code><\/pre>\n\n<p>It should solve all the conflicts.<\/p>\n"}]}
+{"QuestionId":52946920,"AnswerCount":1,"Tags":"<python><torch>","CreationDate":"2018-10-23T10:36:31.473","AcceptedAnswerId":52948967.0,"OwnerUserId":10159479.0,"Title":"Bool value of Tensor with more than one value is ambiguous in Pytorch","Body":"<p>I want to create a model in pytorch, but I can't \ncompute the loss. \nIt's always return <strong>Bool value of Tensor with more \nthan one value is ambiguous<\/strong>\nActually, I run example code, it work.<\/p>\n\n<pre><code>loss = CrossEntropyLoss()\ninput = torch.randn(8, 5)\ninput\ntarget = torch.empty(8,dtype=torch.long).random_(5)\ntarget\noutput = loss(input, target)\n<\/code><\/pre>\n\n<p>Here is my code,<\/p>\n\n<pre><code>################################################################################\n##\n##\nimport torch\nfrom torch.nn import Conv2d, MaxPool2d, Linear, CrossEntropyLoss, MultiLabelSoftMarginLoss\nfrom torch.nn.functional import relu, conv2d, max_pool2d, linear, softmax\nfrom torch.optim import adadelta\n##\n##\n##  Train\nTrain = {}\nTrain[\"Image\"]    = torch.rand(2000, 3, 76, 76)\nTrain[\"Variable\"] = torch.rand(2000, 6)\nTrain[\"Label\"] = torch.empty(2000, dtype=torch.long).random_(2)\n##\n##\n##  Valid\nValid = {}\nValid[\"Image\"]    = torch.rand(150, 3, 76, 76)\nValid[\"Variable\"] = torch.rand(150, 6)\nValid[\"Label\"]    = torch.empty(150, dtype=torch.long).random_(2)\n################################################################################\n##\n##\n##  Model\nImageTerm    = Train[\"Image\"]\nVariableTerm = Train[\"Variable\"]\nPip = Conv2d(in_channels=3, out_channels=32, kernel_size=(3,3), stride=1, padding=0)(ImageTerm)\nPip = MaxPool2d(kernel_size=(2,2), stride=None, padding=0)(Pip)\nPip = Conv2d(in_channels=32, out_channels=64, kernel_size=(3,3), stride=1, padding=0)(Pip)\nPip = MaxPool2d(kernel_size=(2,2), stride=None, padding=0)(Pip)\nPip = Pip.view(2000, -1)\nPip = torch.cat([Pip, VariableTerm], 1)\nPip = Linear(in_features=18502, out_features=1000 , bias=True)(Pip)\nPip = Linear(in_features=1000, out_features=2 , bias=True)(Pip)\n##\n##\n##  Loss\nLoss = CrossEntropyLoss(Pip, Train[\"Label\"])\n<\/code><\/pre>\n\n<p>The error is on  <strong>Loss = CrossEntropyLoss(Pip, Train[\"Label\"])<\/strong>,\nthanks.<\/p>\n","answers":[{"AnswerId":"52948967","CreationDate":"2018-10-23T12:21:36.893","ParentId":null,"OwnerUserId":"10494960","Title":null,"Body":"<p>In your minimal example, you create an object \"loss\" of the class \"CrossEntropyLoss\". This object is able to compute your loss as<\/p>\n\n<pre><code>loss(input, target)\n<\/code><\/pre>\n\n<p>However, in your actual code, you try to create the object \"Loss\", while passing Pip and the labels to the \"CrossEntropyLoss\" class constructor.\nInstead, try the following:<\/p>\n\n<pre><code>loss = CrossEntropyLoss()\nloss(Pip, Train[\"Label\"])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52946993,"AnswerCount":1,"Tags":"<python><tensorflow><cosine-similarity>","CreationDate":"2018-10-23T10:39:42.170","AcceptedAnswerId":52947534.0,"OwnerUserId":984308.0,"Title":"Computing the Cosine Similarity of two sets of vectors in Tensorflow","Body":"<p>Hi fellow Stackoverflow users,<\/p>\n\n<p>I am currently struggling with a solution to this problem:<\/p>\n\n<p>I've got 2 2d tensors:<\/p>\n\n<pre><code>a = Tensor(shape=[600,52]) # 600 vectors of length 52\nb = Tensor(shape=[16000,52]) # 1600 vectors of length 52\n<\/code><\/pre>\n\n<p>I am trying to compute the cosine similarity of all vector combinations and store them in a 3rd tensor.<\/p>\n\n<pre><code>similarity = Tensor(shape=[600, 16000])\n<\/code><\/pre>\n\n<p>My questions now are the following<\/p>\n\n<p>a) I am not quite sure on how to implement this in a non-iterative way, I thought about using the broadcasting semantics in combination with tf.losses.cosine_distance, but I can not quite wrap my head around what that would actually look like.<\/p>\n\n<p>b) Depending on the implementation (if tf.losses.cosine_distance is used, which requires matching dimensions for both input tensors) the memory footprint might become quite large, as it would require to create two tensors of shape [600, 1600, 52] in order to compute the distances for all combinations of vectors. Could you think of any possibility to work around that?<\/p>\n\n<p>I hope that I was able to convey my thoughts in an understandable manner, thank you for your help<\/p>\n\n<p>Best,<\/p>\n","answers":[{"AnswerId":"52947534","CreationDate":"2018-10-23T11:08:41.853","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can compute that simply like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Vectors\na = tf.placeholder(tf.float32, shape=[600, 52])\nb = tf.placeholder(tf.float32, shape=[16000, 52])\n# Cosine similarity\nsimilarity = tf.reduce_sum(a[:, tf.newaxis] * b, axis=-1)\n# Only necessary if vectors are not normalized\nsimilarity \/= tf.norm(a[:, tf.newaxis], axis=-1) * tf.norm(b, axis=-1)\n# If you prefer the distance measure\ndistance = 1 - similarity\n<\/code><\/pre>\n"}]}
+{"QuestionId":52947103,"AnswerCount":1,"Tags":"<python><tensorflow><rnn><softmax>","CreationDate":"2018-10-23T10:46:15.450","AcceptedAnswerId":52951782.0,"OwnerUserId":5647864.0,"Title":"Sampled Softmax Loss in Tensorflow with RNN - Dimensions mismatch problem","Body":"<p>I have an implementation of a Many-to-one RNN with variable sequence length (a sentence classification problem)<\/p>\n\n<p>I am trying to implement a sampled softmax loss since I have 500 classes and want to speed up the training.<\/p>\n\n<p>The following are my input parameters shapes<\/p>\n\n<pre><code>WLast.shape\nTensorShape([Dimension(500), Dimension(500)])\n\nbLast.shape \nTensorShape([Dimension(500)])\n\nlabels.shape\nTensorShape([Dimension(None), Dimension(500)])\n\npred_out.shape\nTensorShape([Dimension(None), Dimension(500)])\n<\/code><\/pre>\n\n<p>Pred_out is the last prediction from the RNN<\/p>\n\n<p>Problem is that when I run:<\/p>\n\n<pre><code>cost = tf.nn.sampled_softmax_loss(WLast,bLast,labels,pred_out,10,500)\n<\/code><\/pre>\n\n<p>it gives me this error:<\/p>\n\n<pre><code>InvalidArgumentError: Dimension must be 1 but is 500 for 'sampled_softmax_loss\/ComputeAccidentalHits' (op: 'ComputeAccidentalHits') with input shapes: [?,500], [10].\n<\/code><\/pre>\n\n<p>I don't understand, the shapes matches the arguments of the function, does someone know what could I be doing wrong?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"52951782","CreationDate":"2018-10-23T14:39:45.173","ParentId":null,"OwnerUserId":"5647864","Title":null,"Body":"<p>I found this implementation: <a href=\"https:\/\/github.com\/olirice\/sampled_softmax_loss\" rel=\"nofollow noreferrer\">https:\/\/github.com\/olirice\/sampled_softmax_loss<\/a>\nand solved the problem by reshaping the labels<\/p>\n\n<pre><code>labels = tf.reshape(tf.argmax(labels, 1), [-1,1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52947339,"AnswerCount":1,"Tags":"<c++><protocol-buffers><caffe>","CreationDate":"2018-10-23T10:56:47.267","AcceptedAnswerId":null,"OwnerUserId":10285454.0,"Title":"How to write a caffe network to prototxt in C++ using google protobuf","Body":"<p>I am using protobuf and caffe. I want to create a network using the caffe API in C++ and then write this to a protobuf file. <\/p>\n\n<p>name: \"AlexNet\"\ninput: \"data\"\ninput_shape {\n  dim: 1\n  dim: 3\n  dim: 227\n  dim: 227\n}<\/p>\n\n<p>I have a C++ program that is creating a <code>NetParameter<\/code> object <code>param<\/code>. My question is which function should I call to write this to a prototxt file. <\/p>\n\n<p>I think I should be using <code>WriteProtoToTextFile(const Message&amp; proto, const char* filename)<\/code> located here in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/fc77ef3d23423e57d83e16c32844bbbe8d4d9f0c\/src\/caffe\/util\/io.cpp#L32\" rel=\"nofollow noreferrer\">src\/caffe\/util\/io.cpp:<\/a><\/p>\n\n<pre><code>#include &lt;fcntl.h&gt;\n#include &lt;google\/protobuf\/io\/coded_stream.h&gt;\n#include &lt;google\/protobuf\/io\/zero_copy_stream_impl.h&gt;\n#include &lt;google\/protobuf\/text_format.h&gt;\n#include &lt;stdint.h&gt;\n#include &lt;algorithm&gt;\n#include &lt;fstream&gt; \/\/ NOLINT(readability\/streams)\n#include &lt;string&gt;\n#include &lt;vector&gt;\n#include &lt;string&gt;\n#include \"caffe.pb.h\"\n#include &lt;iostream&gt;\n#include &lt;caffe\/caffe.hpp&gt;\n\nusing namespace std;\nusing google::protobuf::Message;\nusing google::protobuf::io::CodedInputStream;\nusing google::protobuf::io::CodedOutputStream;\nusing google::protobuf::io::FileInputStream;\nusing google::protobuf::io::FileOutputStream;\nusing google::protobuf::io::ZeroCopyInputStream;\nusing google::protobuf::io::ZeroCopyOutputStream;\n\nint main()\n{\n    caffe::NetParameter param;\n\n    const char *filename = \"\/test.prototxt\";\n\n    int fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC, 0644);\n    if (fd == -1)\n        cout &lt;&lt; \"File not found: \" &lt;&lt; filename;\n\n    FileOutputStream *output = new FileOutputStream(fd);\n\n    param.set_name(\"AlexNet\");\n\n    \/\/How to convert param to a proto message \n\n    \/\/ Call WriteProtoToTextFile(const Message&amp; proto, const char* filename) ??\n\n    delete outputput;\n\n    close(fd);\n    return 0;\n}\n<\/code><\/pre>\n\n<p>Is this correct? This function take a proto message as the first argument. How do I convert the <code>param<\/code> object to a proto message?<\/p>\n","answers":[{"AnswerId":"52951757","CreationDate":"2018-10-23T14:38:12.013","ParentId":null,"OwnerUserId":"2071932","Title":null,"Body":"<p>After you've created the NetParameter, you can write it to a prototxt by using:<\/p>\n\n<pre><code>caffe::NetParameter param;\n\n\/\/ ... param initialization ...\n\nstd::ofstream ofs; \nofs.open (\"test.prototxt\", std::ofstream::out | std::ofstream::app);\nofs &lt;&lt; param.Utf8DebugString(); \nofs.close();\n<\/code><\/pre>\n"}]}
+{"QuestionId":52948351,"AnswerCount":1,"Tags":"<python><tensorflow><amazon-sagemaker>","CreationDate":"2018-10-23T11:49:56.750","AcceptedAnswerId":null,"OwnerUserId":10545958.0,"Title":"Restore a specific checkpoint for deploying with Sagemaker and TensorFlow","Body":"<p>I'm using SageMaker for training some custom TF model I realized. During training I naturally evaluate the model multiple times in order to understand when the NN actually starts overfitting. After training I'd like to restore the model that works best (i.e. which presents minimum validation loss) and deploy it on an endpoint. However, if I use the classic Tensorflow.attach() the model that is restored corresponds with the one stored in output\/model.tar.gz, which, if I got it correctly, would be the one corresponding with the last training iteration (thus it may overfits). <\/p>\n\n<p>Is there a way for specifying to SageMaker which checkpoint restoring without necessarily retraining the model with early stopping? Even forcing SM to save in model.tar.gz the model that presents minimum validation loss and not the last one would work for me, unfortunately I didn't find any immediate way to do so...<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"53018423","CreationDate":"2018-10-27T03:15:37.170","ParentId":null,"OwnerUserId":"3361753","Title":null,"Body":"<p>You can use the checkpoint_path to restore the Model from a previous checkpoint:<\/p>\n\n<pre><code>previous_checkpoint_path = 's3:\/\/location\/of\/my\/previous\/generated\/checkpoints'\n\ntf_estimator = TensorFlow('tf-train.py', role='SageMakerRole',\n                          checkpoint_path=previous_checkpoint_path\n                          training_steps=10000, evaluation_steps=100,\n                          train_instance_count=1, train_instance_type='ml.p2.xlarge',\n                          framework_version='1.10.0')\ntf_estimator.fit('s3:\/\/bucket\/path\/to\/training\/data')\n<\/code><\/pre>\n\n<p>For more information:<\/p>\n\n<p><a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/README.rst#restoring-from-checkpoints\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/README.rst#restoring-from-checkpoints<\/a><\/p>\n\n<p><a href=\"https:\/\/sagemaker.readthedocs.io\/en\/latest\/sagemaker.tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/sagemaker.readthedocs.io\/en\/latest\/sagemaker.tensorflow.html<\/a><\/p>\n"}]}
+{"QuestionId":52948410,"AnswerCount":1,"Tags":"<python-3.x><deep-learning><pytorch>","CreationDate":"2018-10-23T11:51:53.587","AcceptedAnswerId":null,"OwnerUserId":9974686.0,"Title":"Frequently getting CUDA error out of memory?","Body":"<p>I have written a code for training a deep learning model where I am deleting cuda tensors after every batch and doing <code>torch.cuda.empty_cache()<\/code> after that. I am pretty sure that batch size is not large enough to cause this error. What could be the possible reason for this?<\/p>\n\n<pre><code>    for epoch in range(1+last_epoch, self.num_epochs+1):\n        for phase in ['train', 'val']:\n            loss_arr = []\n            if phase == 'train':\n                model.train()                    \n                scheduler.step()\n                was_training = True\n            else:\n                model.eval()\n                was_training = False\n            for i_batch, sample_batched in enumerate(dataloaders[phase]):\n                X = sample_batched[0]\n                y = sample_batched[1].type(torch.LongTensor)\n                w = sample_batched[2]\n\n                if model.is_cuda:\n                    X, y, w = X.cuda(non_blocking=True), y.cuda(non_blocking=True),  w.cuda(non_blocking=True)\n\n                output = model(X)\n                loss = self.loss_func(output, y, w)\n                if phase == 'train':\n                    curr_iteration+=1\n                    optim.zero_grad()                        \n                    loss.backward()\n                    optim.step()\n                    if (curr_iteration % log_nth == 0):\n                        self.logWriter.loss_per_iter(loss.item(), curr_iteration)\n\n                loss_arr.append(loss.item())\n\n                with torch.no_grad():\n                    self.logWriter.update_cm_per_iter(output, y, self.labels, phase)\n\n                del X, y, w, output, loss\n                torch.cuda.empty_cache()\n\n            self.logWriter.loss_per_epoch(loss_arr, phase, epoch)\n\n            epoch_output, epoch_labels = model.predict(dataloaders[phase].dataset.X), dataloaders[phase].dataset.y\n            self.logWriter.dice_score_per_epoch(epoch_output, epoch_labels, phase, epoch)\n            index = np.random.choice(len(dataloaders[phase].dataset), 3, replace=False)\n            self.logWriter.image_per_epoch(epoch_output[index], epoch_labels[index], phase, epoch)\n            self.logWriter.cm_per_epoch(self.labels, phase, epoch, i_batch)\n            del epoch_output, epoch_labels\n\n        print(\"==== Epoch [\"+str(epoch)+\" \/ \"+str(self.num_epochs)+\"] done ====\")        \n        model.save('models\/' + self.exp_dir_name + '\/quicknat_epoch' + str(epoch) + '.model')\n<\/code><\/pre>\n\n<p>And the perdict function inside the model is as follows<\/p>\n\n<pre><code>def predict(self, X, enable_dropout = False):\n    \"\"\"\n    Predicts the outout after the model is trained.\n    Inputs:\n    - X: Volume to be predicted\n    \"\"\"        \n    self.eval()\n\n    if type(X) is np.ndarray:\n        X = torch.tensor(X, requires_grad = False).cuda(non_blocking=True)\n    elif type(X) is torch.Tensor and not X.is_cuda:\n        X = X.cuda(non_blocking=True)\n\n    if enable_dropout:\n        self.enable_test_dropout()\n\n    with torch.no_grad():         \n        out = self.forward(X)\n\n    max_val, idx = torch.max(out,1)\n    idx = idx.data.cpu().numpy()\n    prediction = np.squeeze(idx)\n    del X, out, idx, max_val\n    return prediction\n<\/code><\/pre>\n","answers":[{"AnswerId":"52953192","CreationDate":"2018-10-23T15:51:48.030","ParentId":null,"OwnerUserId":"9974686","Title":null,"Body":"<p>I realized that I am feeding the entire dataset for the prediction after each epoch. Doing that batch-wise solved the issue.<\/p>\n"}]}
+{"QuestionId":52948432,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><loss-function>","CreationDate":"2018-10-23T11:53:02.603","AcceptedAnswerId":null,"OwnerUserId":9805951.0,"Title":"How to change tensorflow loss function from single number to an array","Body":"<p>I am working with Tensorflow's DNNRegressor module to train a 2 hidden layered neural network with 100 nodes each in order to map an input layer containing 9 parameters and an output layer containing 3980 parameters.<\/p>\n\n<p>The size of my training set is approximately 500,000 data points. So far, I have achieved some success in training this neural network. I work with the default loss function, which is the Mean Squared Error. This is a single floating point variable, and I would like to have a loss function that has the same size as the output layer - 3980 floating values (because the output layer is a spectrum of light).<\/p>\n\n<p>Has anyone considered\/dealt with such a situation?<\/p>\n","answers":[]}
+{"QuestionId":52950245,"AnswerCount":0,"Tags":"<python><apache-spark><tensorflow><keras><grid-search>","CreationDate":"2018-10-23T13:23:14.803","AcceptedAnswerId":null,"OwnerUserId":9024698.0,"Title":"Keras\/Tensorflow GridSearchCV on Spark cluster","Body":"<p>I am trying to execute a Grid Search on a Spark cluster. For this reason, I am running <code>nohup .\/spark_python_shell.sh &gt; output.log &amp;<\/code> at my <code>bash<\/code> shell to ignite the Spark cluster and I also get my python script running (see below <code>spark-submit \\ --master yarn 'dnn_grid_search.py'<\/code>):<\/p>\n\n<pre><code>    SPARK_HOME=\/u\/users\/******\/spark-2.3.0 \\\n    Q_CORE_LOC=\/u\/users\/******\/q-core \\\n    ENV=local \\\n    HIVE_HOME=\/usr\/hdp\/current\/hive-client \\\n    SPARK2_HOME=\/u\/users\/******\/spark-2.3.0 \\\n    HADOOP_CONF_DIR=\/etc\/hadoop\/conf \\\n    HIVE_CONF_DIR=\/etc\/hive\/conf \\\n    HDFS_PREFIX=hdfs:\/\/ \\\n    PYTHONPATH=\/u\/users\/******\/q-core\/python-lib:\/u\/users\/******\/three-queues\/python-lib:\/u\/users\/******\/pyenv\/prod_python_libs\/lib\/python2.7\/site-packages\/:$PYTHON_PATH \\\n    YARN_HOME=\/usr\/hdp\/current\/hadoop-yarn-client \\\n    SPARK_DIST_CLASSPATH=$(hadoop classpath):$(yarn classpath):\/etc\/hive\/conf\/hive-site.xml \\\n    PYSPARK_PYTHON=\/usr\/bin\/python2.7 \\\n    QQQ_LOC=\/u\/users\/******\/three-queues \\\n    spark-submit \\\n    --master yarn 'dnn_grid_search.py' \\\n    --executor-memory 10g \\\n    --num-executors 8 \\\n    --executor-cores 10 \\\n    --conf spark.port.maxRetries=80 \\\n    --conf spark.dynamicAllocation.enabled=False \\\n    --conf spark.default.parallelism=6000 \\\n    --conf spark.sql.shuffle.partitions=6000 \\\n    --principal ************************ \\\n    --queue default \\\n    --name lets_get_starting \\\n    --keytab \/u\/users\/******\/.******.keytab \\\n    --driver-memory 10g\n<\/code><\/pre>\n\n<p>In this <code>dnn_grid_search.py<\/code> python script there is the following source code which tries to connect the Grid Search with the Spark cluster:<\/p>\n\n<pre><code># Create the 'Keras' classifier\n# 'keras_model' is a function which returns the 'Keras' model which I built\nfrom sklearn.model_selection import GridSearchCV\nfrom keras.wrappers.scikit_learn import KerasClassifier\nclassifier = KerasClassifier(build_fn=keras_model, verbose=0)\n\n# Spark configuration\nfrom pyspark import SparkContext, SparkConf\nconf = SparkConf()\nsc = SparkContext(conf=conf)\n\n# Execute grid search - using spark_sklearn library\nfrom spark_sklearn import GridSearchCV\nclassifiers_grid = GridSearchCV(sc, estimator=classifier, param_grid=parameters, scoring=custom_scorer, cv=train_valid_predefined_split,n_jobs=-1)\nclassifiers_grid.fit(X, y)\n<\/code><\/pre>\n\n<p>When I am running the python script then I am getting an error at the line <code>classifiers_grid.fit(X, y)<\/code> which is the following:<\/p>\n\n<pre><code>ImportError: No module named keras.wrappers.scikit_learn\n<\/code><\/pre>\n\n<p>or in quite greater detail (but without including everything because it is too long) is the following:<\/p>\n\n<pre><code>18\/10\/23 12:53:38 INFO scheduler.TaskSetManager: Starting task 2.0 in stage 0.0 (TID 2, *******.*******.com, executor 2, partition 2, PROCESS_LOCAL, 42479 bytes)\n18\/10\/23 12:53:38 WARN scheduler.TaskSetManager: Lost task 1.0 in stage 0.0 (TID 1, *******.*******.com, executor 2): org.apache.spark.api.python.PythonException: Traceback (most recent call last):\n  File \"\/u\/applic\/data\/hdfs12\/hadoop\/yarn\/local\/usercache\/*******\/appcache\/application_1539785180345_30553\/container_e126_1539785180345_30553_01_000003\/pyspark.zip\/pyspark\/worker.py\", line 216, in main\n    func, profiler, deserializer, serializer = read_command(pickleSer, infile)\n  File \"\/u\/applic\/data\/hdfs12\/hadoop\/yarn\/local\/usercache\/*******\/appcache\/application_1539785180345_30553\/container_e126_1539785180345_30553_01_000003\/pyspark.zip\/pyspark\/worker.py\", line 60, in read_command\n    command = serializer.loads(command.value)\n  File \"\/u\/applic\/data\/hdfs12\/hadoop\/yarn\/local\/usercache\/*******\/appcache\/application_1539785180345_30553\/container_e126_1539785180345_30553_01_000003\/pyspark.zip\/pyspark\/serializers.py\", line 562, in loads\n    return pickle.loads(obj)\nImportError: No module named keras.wrappers.scikit_learn\n\n        at org.apache.spark.api.python.BasePythonRunner$ReaderIterator.handlePythonException(PythonRunner.scala:298)\n        at org.apache.spark.api.python.PythonRunner$$anon$1.read(PythonRunner.scala:438)\n        at org.apache.spark.api.python.PythonRunner$$anon$1.read(PythonRunner.scala:421)\n        at org.apache.spark.api.python.BasePythonRunner$ReaderIterator.hasNext(PythonRunner.scala:252)\n        at org.apache.spark.InterruptibleIterator.hasNext(InterruptibleIterator.scala:37)\n        at scala.collection.Iterator$class.foreach(Iterator.scala:893)\n        at org.apache.spark.InterruptibleIterator.foreach(InterruptibleIterator.scala:28)\n        at scala.collection.generic.Growable$class.$plus$plus$eq(Growable.scala:59)\n        at scala.collection.mutable.ArrayBuffer.$plus$plus$eq(ArrayBuffer.scala:104)\n        at scala.collection.mutable.ArrayBuffer.$plus$plus$eq(ArrayBuffer.scala:48)\n        at scala.collection.TraversableOnce$class.to(TraversableOnce.scala:310)\n        at org.apache.spark.InterruptibleIterator.to(InterruptibleIterator.scala:28)\n        at scala.collection.TraversableOnce$class.toBuffer(TraversableOnce.scala:302)\n        at org.apache.spark.InterruptibleIterator.toBuffer(InterruptibleIterator.scala:28)\n        at scala.collection.TraversableOnce$class.toArray(TraversableOnce.scala:289)\n        at org.apache.spark.InterruptibleIterator.toArray(InterruptibleIterator.scala:28)\n        at org.apache.spark.rdd.RDD$$anonfun$collect$1$$anonfun$12.apply(RDD.scala:939)\n        at org.apache.spark.rdd.RDD$$anonfun$collect$1$$anonfun$12.apply(RDD.scala:939)\n        at org.apache.spark.SparkContext$$anonfun$runJob$5.apply(SparkContext.scala:2067)\n        at org.apache.spark.SparkContext$$anonfun$runJob$5.apply(SparkContext.scala:2067)\n        at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:87)\n        at org.apache.spark.scheduler.Task.run(Task.scala:109)\n        at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:345)\n        at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)\n        at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)\n        at java.lang.Thread.run(Thread.java:745)\n<\/code><\/pre>\n\n<p><strong>How can I use GridSearchCV with Keras models on a Spark cluster<\/strong> (in general or specifically with the <code>spark_sklearn<\/code> library)<strong>?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":52950449,"AnswerCount":5,"Tags":"<python><tensorflow><keras><google-cloud-platform><gpu>","CreationDate":"2018-10-23T13:34:01.440","AcceptedAnswerId":null,"OwnerUserId":8898218.0,"Title":"valueError when using multi_gpu_model in keras","Body":"<p>I am using google cloud VM with 4 Tesla K80 GPU's. <\/p>\n\n<p>I am running a keras model using multi_gpu_model with gpus=4(since i have 4 gpu's). But, i am getting the following error<\/p>\n\n<blockquote>\n  <p>ValueError: To call <code>multi_gpu_model<\/code> with <code>gpus=4<\/code>, we expect the\n  following devices to be available: ['\/cpu:0', '\/gpu:0', '\/gpu:1',\n  '\/gpu:2', '\/gpu:3']. However this machine only has: ['\/cpu:0',\n  '\/xla_cpu:0', '\/xla_gpu:0', '\/gpu:0']. Try reducing <code>gpus<\/code>.<\/p>\n<\/blockquote>\n\n<p>I can see that there are only two gpu's here namely <code>'\/xla_gpu:0', '\/gpu:0'<\/code>. so, i tried with <code>gpus = 2<\/code> and again got the following error<\/p>\n\n<blockquote>\n  <p>ValueError: To call <code>multi_gpu_model<\/code> with <code>gpus=2<\/code>, we expect the\n  following devices to be available: ['\/cpu:0', '\/gpu:0', '\/gpu:1'].\n  However this machine only has: ['\/cpu:0', '\/xla_cpu:0', '\/xla_gpu:0',\n  '\/gpu:0']. Try reducing <code>gpus<\/code>.<\/p>\n<\/blockquote>\n\n<p>can anyone help me out with the error. Thanks!<\/p>\n","answers":[{"AnswerId":"55759571","CreationDate":"2019-04-19T09:25:35.493","ParentId":null,"OwnerUserId":"11383035","Title":null,"Body":"<p>You can check all device list using following code:<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\ndevice_lib.list_local_devices()\n<\/code><\/pre>\n"},{"AnswerId":"58273653","CreationDate":"2019-10-07T16:31:49.220","ParentId":null,"OwnerUserId":"10760691","Title":null,"Body":"<p>I had the same issue. <strong>Tensorflow-gpu 1.14<\/strong> installed, <strong>CUDA 10.0<\/strong>, and <strong>4 XLA_GPUs<\/strong> were displayed with <strong>device_lib.list_local_devices()<\/strong>.<\/p>\n\n<p>I have another conda environement and there is just <strong>Tensorflow 1.14<\/strong> installed and no tensorflow-gpu, and i don't know why, but i can run my multi_gpu model on all gpus with that environment.<\/p>\n"},{"AnswerId":"56194403","CreationDate":"2019-05-17T22:44:27.437","ParentId":null,"OwnerUserId":"6543020","Title":null,"Body":"<p>This can be caused by using <code>tensorflow<\/code> instead of <code>tensorflow-gpu<\/code>.<\/p>\n\n<p>One way to fix this is the following:<\/p>\n\n<pre><code>$ pip uninstall tensorflow\n$ pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>More information can be found here: <a href=\"https:\/\/stackoverflow.com\/a\/42652258\/6543020\">https:\/\/stackoverflow.com\/a\/42652258\/6543020<\/a><\/p>\n"},{"AnswerId":"52957261","CreationDate":"2018-10-23T20:18:31.417","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>TensorFlow is only seeing one GPU (the gpu and xla_gpu devices are two backends over the same physical device). Are you setting CUDA_VISIBLE_DEVICES? Does nvidia-smi show all GPUs?<\/p>\n"},{"AnswerId":"52951437","CreationDate":"2018-10-23T14:22:32.770","ParentId":null,"OwnerUserId":"7812314","Title":null,"Body":"<p>It looks like Keras only sees one of the GPUs.<\/p>\n\n<p>Make sure that all 4 GPUs are accessible, you can use <code>device_lib<\/code> with TensorFlow.<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\n\ndef get_available_gpus():\n    local_device_protos = device_lib.list_local_devices()\n    return [x.name for x in local_device_protos if x.device_type == 'GPU']\n<\/code><\/pre>\n\n<p>You might need to manually install or update GPU drivers on your instance. Consult <a href=\"https:\/\/cloud.google.com\/compute\/docs\/gpus\/add-gpus#install-driver-script\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52950838,"AnswerCount":1,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2018-10-23T13:53:12.430","AcceptedAnswerId":null,"OwnerUserId":10103615.0,"Title":"Is it possible to loop when size of a tensor is not known?","Body":"<p>I would like to read time series sequence saved in tfrecord format. Each time series has different length. What I want to achieve is to split a long tensor into batch of smaller tensors of requested length. It is very easy to do with numpy arrays and it would look like this:<\/p>\n\n<pre><code>length = 200\nfor begin in range(tensor_size-length):\n     tensor_slice = tf.slice(my_tensor, begin, length)\n     my_slices.append(tensor_slice)\n<\/code><\/pre>\n\n<p>In such function my problem is: how to get size of a tensor, so that using a loop could be possible?\nBelow is part of code where example is readed and decoded.<\/p>\n\n<pre><code>file_queue = tf.train.string_input_producer(tf_files, num_epochs=num_epochs)\n        reader = tf.TFRecordReader()\n        _, serialized_records = reader.read(file_queue)\n\n        feature_map = {\n            \"speed\":tf.FixedLenSequenceFeature([], tf.float32, allow_missing=True),\n            \"battery\":tf.FixedLenSequenceFeature([], tf.float32, allow_missing=True)\n        }\n\n        features = tf.parse_single_example(serialized_records, feature_map)\n        speed = tf.cast(features['speed'], tf.float32)\n        battery = tf.cast(features['battery'], tf.float32)\n        speeds = []\n        batteries = []\n\n        #SPLIT TENSOR INTO SMALLER TENSORS\n        features = tf.train.shuffle_batch([speeds, batteries],\n                                           batch_size=batch_size,\n                                           capacity=5000,\n                                           num_threads=4,\n                                           min_after_dequeue=1)\n\n        return features\n<\/code><\/pre>\n","answers":[{"AnswerId":"52951390","CreationDate":"2018-10-23T14:20:11.760","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You cannot loop through a tensor like that in Python. You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop<\/code><\/a>, although it is generally avoided unless it is really the only way to achieve what you want, since it tends to be slow. In your case, you can get the result you want without looping, for example using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a>:<\/p>\n\n<pre><code>length = 200\nfeatures = ...\n# Number of elements\nn = tf.shape(features)[0]\n# Index from zero to number of subtensors\nsplit_idx = tf.range(n - length + 1)\n# Index from zero to subtensor length\nlength_idx = tf.range(length)\n# Indices for gather; each row advances one position, like a \"rolling window\"\ngather_idx = split_idx[:, tf.newaxis] + length_idx\n# Gather result\nfeatures_split = tf.gather(features, gather_idx)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52951131,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><autoencoder>","CreationDate":"2018-10-23T14:08:09.460","AcceptedAnswerId":52951580.0,"OwnerUserId":8889727.0,"Title":"Split a trained neural network to 2 parts and use the latter part","Body":"<p>I have a simple autoencoder DNN with 4 convolution layers followed by 4 deconvolution layers.<\/p>\n\n<pre><code>import tensorflow.contrib.layers as lays\n\nconv1 = lays.conv2d(inputs, 64, [5, 5], stride=2, padding='SAME')\nconv2 = lays.conv2d(conv1, 32, [5, 5], stride=2, padding='SAME')\nconv3 = lays.conv2d(conv2, 16, [5, 5], stride=4, padding='SAME')\nconv4 = lays.conv2d(conv3, 100, [5, 5], stride=4, padding='SAME')  #shape (None,1,1,100)\n\ndconv1 = lays.conv2d_transpose(conv4, 16, [5, 5], stride=4, padding='SAME')\ndconv2 = lays.conv2d_transpose(dconv1, 32, [5, 5], stride=4, padding='SAME')\ndconv3 = lays.conv2d_transpose(dconv2, 64, [5, 5], stride=2, padding='SAME')\ndconv4 = lays.conv2d_transpose(dconv3, 1, [5, 5], stride=2, padding='SAME', activation_fn=tf.nn.tanh) #shape (64,64,1)\n<\/code><\/pre>\n\n<p>Here the inputs are (64,64) gray-scaled images and the autoencoder reconstructs the same input image at the output layer(dconv4). <\/p>\n\n<p>Please note that the objective of this autoencoder is not to do segmentation, but to represent a (64x64) image as a (1,1,100) shaped unique tensor with a minimum loss.<\/p>\n\n<p>As you can see, the shape of the conv4 tensor is (None,1,1,100), which is what I want.<\/p>\n\n<p>After training, this autoencoder works fine for test images.<\/p>\n\n<p>However, I am interested in splitting the trained DNN to convolution and deconvolution parts, take the deconvolution part, input my own (None,1,1,100) tensor to the dconv1 layer and observe the output at dconv4.<\/p>\n\n<p>How do I feed my own input to a tensor in the middle of a neural network (conv4)?<\/p>\n","answers":[{"AnswerId":"52951580","CreationDate":"2018-10-23T14:29:54.113","ParentId":null,"OwnerUserId":"3344139","Title":null,"Body":"<p>You could consider replacing<\/p>\n\n<pre><code>dconv1 = lays.conv2d_transpose(conv4, 16, [5, 5], stride=4, padding='SAME')\n<\/code><\/pre>\n\n<p>with a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder_with_default\" rel=\"nofollow noreferrer\">placeholder_with_default<\/a> like this:<\/p>\n\n<pre><code>latent_ph = tf.placeholder_with_default(conv4, [None, 1, 1, 100], name=\"latent_ph\")\ndconv1 = lays.conv2d_transpose(latent_ph, 16, [5, 5], stride=4, padding='SAME')\n<\/code><\/pre>\n\n<p>This way, you will automatically use conv4, but you can switch it to a different input when you want by passing a feed dict to the <code>session.run<\/code> call when you need it to be something else:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    my_latent = np.random.random((10, 1, 1, 100))\n    result = sess.run(dconv4, feed_dict={latent_ph: my_latent})\n<\/code><\/pre>\n\n<p>If you do not want to retrain, you have to load the weights from your checkpoint and assign them to the new graph with the placeholder. Something similar is done <a href=\"https:\/\/innerpeace-wu.github.io\/2017\/12\/13\/Tensorflow-Restore-partial-weights\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":52951350,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-10-23T14:18:26.130","AcceptedAnswerId":52951524.0,"OwnerUserId":4296244.0,"Title":"Is Eager Execution meant to replace the tensorflow session approach?","Body":"<p>My last project with tensorflow was in 2016, since then Eager Execution was introduced. After reading further into it, it seems as if it can be used instead of the original approach with a session object.<\/p>\n\n<p>Since Eager Execution gives better feedback and a more straight forward approach to designing and debugging neural networks, should we always use Eager Execution for new projects, or are there advantages of the original session approach?<\/p>\n","answers":[{"AnswerId":"52951524","CreationDate":"2018-10-23T14:27:07.090","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>EDIT:<\/p>\n\n<p>The answer below is based on the announcement post of Eager move and the current available documentation. However l, <a href=\"https:\/\/stackoverflow.com\/a\/52994183\">rachelim's answer<\/a> gives further insight about future plans for Eager mode and its more important role in TensorFlow 2.x.<\/p>\n\n<hr>\n\n<p>No, eager execution was introduced as a \"simple mode\" for learning and experimentation. The strength of TensorFlow is in its computation graph structure and its independence from the execution environment (session); this is what allows you to run the same model distributed on different machines or export it to use across different languages, for example.<\/p>\n\n<p>Of course, you can very much use eager mode in \"production\", if that fulfils all your needs, but note that you will not have access to many advanced features, and, even in the cases where these can be \"ported\" to eager execution, doing so may not be a priority for the developers, since it is not supposed to supersede the \"traditional\" execution mode.<\/p>\n\n<p>EDIT: For further reference, see <a href=\"https:\/\/www.tensorflow.org\/guide\/eager#work_with_graphs\" rel=\"nofollow noreferrer\">Work with graphs<\/a> in the Eager Execution guide, which points out some of the benefits of graph-based vs eager code and how you would go about working with both.<\/p>\n"},{"AnswerId":"52994183","CreationDate":"2018-10-25T16:35:48.820","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>In tensorflow 2.0, eager execution will be enabled by default. It is intended to be able to completely replace graph\/session mode, and is a priority for tensorflow developers. For more details, and to join in the discussion on the topic, check out the this RFC on the tensorflow github: <a href=\"https:\/\/github.com\/tensorflow\/community\/pull\/20\" rel=\"nofollow noreferrer\">RFC: Functions, not sessions in 2.0<\/a>. The <a href=\"https:\/\/www.tensorflow.org\/community\/roadmap\" rel=\"nofollow noreferrer\">tensorflow roadmap<\/a> also alludes to this.<\/p>\n\n<p>Using eager mode for new projects is fine. As of now, there are still advantages to the original session approach (as jdehesa mentioned, some \"advanced features\" are not eager-compatible yet), but tensorflow developers are actively working on improving eager (and graph) mode. If you encounter any issues or missing features, we appreciate bug reports on github.<\/p>\n"}]}
+{"QuestionId":52952057,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-23T14:53:36.227","AcceptedAnswerId":null,"OwnerUserId":3711219.0,"Title":"Can't get the right shape for tf.matmul or tf.squared_difference","Body":"<p>From a DNA sequence, I one hot encoded every base (A,C,G,T,N) to <\/p>\n\n<pre><code>{'A': [1, 0, 0, 0, 0],\n 'C': [0, 1, 0, 0, 0],\n 'G': [0, 0, 1, 0, 0],\n 'T': [0, 0, 0, 1, 0],\n 'N': [0, 0, 0, 0, 1]}\n<\/code><\/pre>\n\n<p>Every DNA sequence has 400 bases. So my final training data has a shape <code>X_train.shape = (111453, 400, 5)<\/code> (111453 rows, 400 letters, every letter encoded in a 5 element vector)<\/p>\n\n<p>My label data is a simple yes\/no, so if the DNA sequence has an error it is [1], if not, it is [0], so <code>Y_train.shape is (111453,1)<\/code><\/p>\n\n<p>I'm trying to build a small NN with tensorflow.<\/p>\n\n<pre><code>layer_1_nodes = 5\nlayer_2_nodes = 10\nlayer_3_nodes = 5\nlearning_rate = 0.001\ntraining_epochs = 5\nnumber_of_outputs = 1\n\n# Input Layer\nwith tf.variable_scope('input'):\n    X = tf.placeholder(tf.float32, shape=(None, 400, 5), name=\"X\")\n\n# Layer 1\nwith tf.variable_scope('layer_1'):\n    weights = tf.get_variable(\"weights1\", shape=[400, 5, layer_1_nodes], initializer=tf.contrib.layers.xavier_initializer())\n    biases = tf.get_variable(name=\"biases1\", shape=[layer_1_nodes], initializer=tf.zeros_initializer())\n    layer_1_output = tf.nn.relu(tf.matmul(X, weights) + biases)\n\n\n# Layer 2\nwith tf.variable_scope('layer_2'):\n    weights = tf.get_variable(\"weights2\", shape=[400, layer_1_nodes, layer_2_nodes], initializer=tf.contrib.layers.xavier_initializer())\n    biases = tf.get_variable(name=\"biases2\", shape=[layer_2_nodes], initializer=tf.zeros_initializer())\n    layer_2_output = tf.nn.relu(tf.matmul(layer_1_output, weights) + biases)\n\n# Layer 3\nwith tf.variable_scope('layer_3'):\n    weights = tf.get_variable(\"weights3\", shape=[400, layer_2_nodes, layer_3_nodes], initializer=tf.contrib.layers.xavier_initializer())\n    biases = tf.get_variable(name=\"biases3\", shape=[layer_3_nodes], initializer=tf.zeros_initializer())\n    layer_3_output = tf.nn.relu(tf.matmul(layer_2_output, weights) + biases)\n\nwith tf.variable_scope('layer_drop'):    \n  dropout = tf.layers.dropout(\n      inputs=layer_3_output, rate=0.4)\n\n# Output Layer\nwith tf.variable_scope('output'):\n    weights = tf.get_variable(\"weights4\", shape=[400, layer_3_nodes, number_of_outputs], initializer=tf.contrib.layers.xavier_initializer())\n    biases = tf.get_variable(name=\"biases4\", shape=[number_of_outputs], initializer=tf.zeros_initializer())\n    prediction = tf.matmul(dropout, weights) + biases\n\nwith tf.variable_scope('cost'):\n    Y = tf.placeholder(tf.float32, shape=(None, 1), name=\"Y\")\n    cost = tf.reduce_mean(tf.squared_difference(prediction, Y))  \n<\/code><\/pre>\n\n<p>But I'm always getting errors regarding to the shape of the tensors. Either the first matmul in layer 1, or the cost squared_difference() <code>(Incompatible shapes: [400,400,1] &lt;- the prediction tensor vs. [111453,1])<\/code><\/p>\n\n<p>Tried a model with keras:<\/p>\n\n<pre><code>input_shape = (400, 5, 1)\n\nmodel = Sequential()\nmodel.add(Conv2D(32, kernel_size=(3, 3),\n                 activation='relu',\n                 input_shape=input_shape))\n<\/code><\/pre>\n\n<p>But I can't get my input shape right.<\/p>\n","answers":[]}
+{"QuestionId":52952428,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-23T15:11:29.797","AcceptedAnswerId":null,"OwnerUserId":8954691.0,"Title":"Return tensorflow layer outputs from a function","Body":"<p>I have written the following model in a separate file : <\/p>\n\n<pre><code>def fun(X,num_actions,hiddens,layer_norm,keep_prob = 0.5):\n\n\n\n    out = X\n\n    out = layers.fully_connected(out, num_outputs=hidden[0], activation_fn=None)\n    out1 = tf.nn.relu(out)\n\n    drop_out = tf.nn.dropout(out1, keep_prob)\n\n    out2 = layers.fully_connected(drop_out, num_outputs=hidden[1], activation_fn=None)\n    out2 = tf.nn.relu(out2)\n\n    out3 = layers.fully_connected(out2, num_outputs=hidden[2], activation_fn=None)\n    out3 = tf.nn.relu(out3)\n\n\n    out4 = layers.fully_connected(out3, num_outputs=num_actions, activation_fn=None)\n\n   return out4\n<\/code><\/pre>\n\n<p>I am calling this function in a class defined in different file like this : <\/p>\n\n<pre><code>class a(): \n    def __init__(scope):\n     with tf.variable_scope(scope):\n        # Build the graph\n        self._build_model(...) #this calls the function fun()\n<\/code><\/pre>\n\n<p>I want to get values of <code>out1,drop_out,out2,out3,out4<\/code> returned from <code>fun()<\/code> function but tensorflow gives me errors. <\/p>\n\n<p>I read some posts to get the any tensor using <code>tf.get_variable()<\/code>. But I don't know how to get these layer outputs as I didnt name them &amp; searching them in the the global variable space is a big task. <\/p>\n\n<p>Do you think these tensors will be available under the scope name I am using in <code>class a()<\/code> \n ? My final aim is to write these values to tensorboard. <\/p>\n","answers":[{"AnswerId":"52953909","CreationDate":"2018-10-23T16:34:57.737","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<blockquote>\n  <p>I want to get values of out1,drop_out,out2,out3,out4 returned from fun() function but tensorflow gives me errors. <\/p>\n<\/blockquote>\n\n<p>You'll need to <code>return<\/code> those values if you want to get them.<\/p>\n\n<blockquote>\n  <p>but tensorflow gives me errors. <\/p>\n<\/blockquote>\n\n<p>can't help you if we don't know what the errors are<\/p>\n"}]}
+{"QuestionId":52953219,"AnswerCount":0,"Tags":"<python><tensorflow><cuda><virtualenv>","CreationDate":"2018-10-23T15:52:51.557","AcceptedAnswerId":null,"OwnerUserId":8599554.0,"Title":"I have cuda installed with version 8.0. I installed tensorflow-gpu using pip in virtualenvironment","Body":"<p>I am having the following problem: <\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n_pywrap_tensorflow_internal = swig_import_helper()\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/imp.py\", line 242, in load_module\nreturn load_dynamic(name, filename, file)\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/imp.py\", line 342, in load_dynamic\nreturn _load(spec)\nImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\nFile \"tensorflowtest.py\", line 3, in &lt;module&gt;\nimport tensorflow as tf\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 74, in &lt;module&gt;\nraise ImportError(msg)\nImportError: Traceback (most recent call last):\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n_pywrap_tensorflow_internal = swig_import_helper()\n File \"\/home\/soroy\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/imp.py\", line 242, in load_module\nreturn load_dynamic(name, filename, file)\nFile \"\/home\/soroy\/venv\/lib\/python3.5\/imp.py\", line 342, in load_dynamic\nreturn _load(spec)\nImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n\n<p>I am trying to run the code which is in the following link <a href=\"https:\/\/learningtensorflow.com\/lesson10\/\" rel=\"nofollow noreferrer\">https:\/\/learningtensorflow.com\/lesson10\/<\/a> <\/p>\n\n<p>I can't get to thing why is it showing this error.\nI have tried different versions of tensorflow-gpu the same error I get when I try to run a code. I have installed the tensorflow-gpu using pip manager in virtualenvironment.<\/p>\n","answers":[]}
+{"QuestionId":52953363,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-10-23T15:59:12.497","AcceptedAnswerId":null,"OwnerUserId":7571462.0,"Title":"TypeError: Input 'serialized' of 'ParseExample' Op has type float32 that does not match expected type of string","Body":"<p>I was developing a CNN neural network for image classification by using estimators in Tensorflow. This error was encountered in the saving process. The following is my code for the saving part:<\/p>\n\n<pre><code>        def serving_input_rec_fn():\n            serialized_tf_example = tf.placeholder(dtype=tf.float32,\n                                                   # shape=[-1, self.num_input[0], self.num_input[1], 1],\n                                                   name='input')\n            receiver_tensors = {\"myInput\": serialized_tf_example}\n            feature_spec = {\"examples\": tf.FixedLenFeature(shape=[self.num_input[0], self.num_input[1]], dtype=tf.float32)}\n            features = tf.parse_example(serialized_tf_example, feature_spec)\n            return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n\n        model.export_savedmodel(export_dir, serving_input_rec_fn,\n                                     strip_default_attrs=True)\n<\/code><\/pre>\n\n<p>The error seems to be that in the serialized_tf_example argument of the tf.parse_example function, it expects the data type to be string. But my input image type is actually float32. How can I solve the problem? In my understanding of the placeholder here, it will serve as the input subgraph with the main graph when the model is exported for serving. Therefore, why is the string type expected for the serialized_tf_example?<\/p>\n","answers":[]}
+{"QuestionId":52953988,"AnswerCount":0,"Tags":"<python><tensorflow><multi-gpu>","CreationDate":"2018-10-23T16:40:05.630","AcceptedAnswerId":null,"OwnerUserId":8079099.0,"Title":"What is the purpose of \"tower loss\" in TensorFlow multi-GPU code?","Body":"<p>I am trying to write multi-GPU TensorFlow code for my project and came across the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow noreferrer\">CIFAR10 code<\/a> in the TensorFlow repo. I don't quite understand the use of <code>tower_loss<\/code> in the code. Why can't we just use <code>loss = cifar10.loss(logits, labels)<\/code> in place of line 178 and take the average of all losses?<\/p>\n","answers":[]}
+{"QuestionId":52954212,"AnswerCount":0,"Tags":"<python><string><tensorflow><tensor><tensorflow-serving>","CreationDate":"2018-10-23T16:54:47.553","AcceptedAnswerId":null,"OwnerUserId":5687831.0,"Title":"Tensorflow: Error when operate on tensor of string","Body":"<p>I am using TensorFlow Serving. As an input, I have a tensor of string. First I want to do some transformations with the input tensor and then I want to do the operations of the model. My code looks like this:<\/p>\n\n<pre><code>split_value = person_mail.split(\"@\")\n\nplaceholder_person_mail= 'person_mail'    \nperson_mail = tf.placeholder(tf.string, 1, name=placeholder_person_mail)\n\nprint(\"Test {}\".format(sess.run(split_value, feed_dict={clients_name:[\"fran@gmail.com\"])))\n<\/code><\/pre>\n\n<p>The output should be: <code>[\"fran\", \"gmail.com\"]<\/code><\/p>\n\n<p>But I am getting the following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n       File \"\/tmp\/zeppelin_pyspark-7422666405996158092.py\", line 367, in \n       raise Exception(traceback.format_exc())\n       Exception: Traceback (most recent call last):\n       File \"\/tmp\/zeppelin_pyspark-7422666405996158092.py\", line 355, in \n       exec(code, _zcUserQueryNameSpace)\n       File \"\", line 8, in \n       AttributeError: 'Tensor' object has no attribute 'split'<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":52954349,"AnswerCount":0,"Tags":"<tensorflow><logistic-regression>","CreationDate":"2018-10-23T17:03:18.553","AcceptedAnswerId":null,"OwnerUserId":260826.0,"Title":"LogisticRegression with Tensorflow","Body":"<p>I'm using TF 1.10, and I want to use the banking notes <a href=\"https:\/\/archive.ics.uci.edu\/ml\/datasets\/banknote+authentication\" rel=\"nofollow noreferrer\">dataset<\/a> to Predict if a Bank Note is forged or not:<\/p>\n\n<pre><code>df_dataset = pd.read_csv(banknote_dataset, header=None, names = CSV_COLUMNS)\ndf_train, df_valid, df_test = np.split(df_dataset.sample(frac=1), [int(.6*len(df_dataset)), int(.8*len(df_dataset))])\n\ndef train_input_fn(df, num_epochs):\n  return tf.estimator.inputs.pandas_input_fn(\n    x = df,\n    y = df[LABEL],\n    batch_size = 128,\n    num_epochs = num_epochs,\n    shuffle = True    \n  )\n\ndef eval_input_fn(df):\n  return tf.estimator.inputs.pandas_input_fn(\n    x = df,\n    y = df[LABEL],\n    batch_size = 128,\n    shuffle = False\n  )\n\ndef prediction_input_fn(df):\n  return tf.estimator.inputs.pandas_input_fn(\n    x = df,\n    y = None,\n    batch_size = 128,\n    shuffle = False,\n  )\n\ndef get_feature_cols():\n  input_columns = [tf.feature_column.numeric_column(k) for k in FEATURES]\n  return input_columns\n\nOUTDIR = 'banknote_trained'\nshutil.rmtree(OUTDIR, ignore_errors = True) # start fresh each time\n\nmodel = tf.estimator.LinearRegressor(\n      feature_columns = get_feature_cols(), \n      optimizer=tf.train.FtrlOptimizer(learning_rate=0.1),\n      model_dir = OUTDIR)\n\nmodel.train(input_fn = train_input_fn(df_train, num_epochs = 100))\n\npredictions = model.predict(input_fn = prediction_input_fn(df_test))\nfor items in predictions:\n  print(items)\n<\/code><\/pre>\n\n<p>My results are:<\/p>\n\n<pre><code>{'predictions': array([-0.28320795], dtype=float32)}\n{'predictions': array([0.8572771], dtype=float32)}\n{'predictions': array([0.68809825], dtype=float32)}\n{'predictions': array([0.9708319], dtype=float32)}\n{'predictions': array([0.0971362], dtype=float32)}\n{'predictions': array([0.98395026], dtype=float32)\n<\/code><\/pre>\n\n<p>Which according to Logistic Regression, should be between 0 and 1.\nUse of Logistic Regression is confusing in TF. Check <a href=\"https:\/\/datascience.stackexchange.com\/questions\/40071\/logistic-regression-with-tensorflow\">this<\/a>.<\/p>\n","answers":[]}
+{"QuestionId":52954647,"AnswerCount":1,"Tags":"<python><neural-network><keras><nlp><embedding>","CreationDate":"2018-10-23T17:23:24.437","AcceptedAnswerId":null,"OwnerUserId":9464528.0,"Title":"Combine text with categorical features using entity embeddings and Keras functional API","Body":"<p>I'm trying to combine text and categorical variables using Keras functional API.<\/p>\n\n<p>The model compiles but when I try to train the model using \"fit\", it gives me an error.\nLooks like the way I put the data is wrong.\nHas anyone ever done something similar and knows how to make it work?<\/p>\n\n<p>Code for building the model:<\/p>\n\n<pre><code>all_inputs = []\ncat_embeddings = []\n\ntext_input = Input(shape=(MAX_SEQUENCE_LENGTH, ),name='text_input')\nall_inputs.append(text_input)\n\nx = Embedding(embedding_matrix.shape[0], # or len(word_index) + 1\n              embedding_matrix.shape[1], # or EMBEDDING_DIM,\n              weights=[embedding_matrix],\n              input_length=MAX_SEQUENCE_LENGTH,\n              trainable=True)(text_input)\n\nx = SpatialDropout1D(0.2)(x)\nx = Bidirectional(GRU(128, return_sequences=True,dropout=0.1,recurrent_dropout=0.1))(x)\nx = Conv1D(64, kernel_size = 3, padding = \"valid\", kernel_initializer = \"glorot_uniform\")(x)\nx = GlobalMaxPooling1D()(x)\nx = Dense(128, activation='relu')(x)\n\nfor cat in Categorical_Features:\n    cat_input = Input(shape=(1,), name=cat)\n    no_of_unique_cat  = X_train[cat].nunique()\n    embedding_size = np.ceil((no_of_unique_cat)\/2)\n    embedding_size = int(embedding_size)\n    cat_embedding = Embedding(no_of_unique_cat+1, embedding_size, input_length = 1)(cat_input)\n    cat_embedding = Reshape(target_shape=(embedding_size,))(cat_embedding)\n\n    all_inputs.append(cat_input)\n    cat_embeddings.append(cat_embedding)\n\nconc = Concatenate()(cat_embeddings)\nx = concatenate([conc, x])\n\nx = Dense(128, activation='relu')(x)\nx = Dropout(0.1)(x)\nx = Dense(128, activation='relu')(x)\nx = Dropout(0.1)(x)\nx = Dense(128, activation='relu')(x)\n\npreds = Dense(93, activation=\"sigmoid\")(x)\n\nmodel = Model(inputs=all_inputs, outputs=preds)\nmodel.compile(loss='categorical_crossentropy', optimizer=Adam(lr=1e-3), metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Then I try to put the data in this way:<\/p>\n\n<pre><code>model.fit([X_train_seq,\n           X_train['Categorical_Features_1'],\n           X_train['Categorical_Features_2'],\n           X_train['Categorical_Features_3'],\n           X_train['Categorical_Features_4'],\n           X_train['Categorical_Features_5'],\n           X_train['Categorical_Features_6']]\n          ,\n          y_train, \n          validation_split=0.2, \n          class_weight = d_class_weights,\n          epochs=5, \n          batch_size=512)\n<\/code><\/pre>\n\n<p>Then I get this error:<\/p>\n\n<pre><code>InvalidArgumentError: indices[460,0] = 421 is not in [0, 406)\n     [[{{node embedding_18\/GatherV2}} = GatherV2[Taxis=DT_INT32, Tindices=DT_INT32, Tparams=DT_FLOAT, _class=[\"loc:@training_8\/Adam\/gradients\/embedding_18\/GatherV2_grad\/Reshape\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](embedding_18\/embeddings\/read, embedding_18\/Cast, embedding_17\/GatherV2\/axis)]]\n<\/code><\/pre>\n\n<p>This code is inspired by the code in this Kaggle competition:<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/aquatic\/entity-embedding-neural-net\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/aquatic\/entity-embedding-neural-net<\/a><\/p>\n\n<p>and the paper \"Entity Embeddings of Categorical Variables\":<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1604.06737\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1604.06737<\/a><\/p>\n","answers":[{"AnswerId":"52956641","CreationDate":"2018-10-23T19:37:05.050","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>The problem is in the <code>Embedding<\/code> layer whereby your <code>len(word_index) + 1<\/code> or <code>embedding_matrix.shape[0]<\/code> doesn't actually correspond to the maximum index. In this case <code>len(word_index) + 1 = 407<\/code> but your first input integer index is actually <code>421<\/code> which is out of bounds.<\/p>\n\n<p>You need to double check the <code>X_train_seq<\/code> indices and whether they are within the range provided to the <code>Embedding<\/code> layer.<\/p>\n"}]}
+{"QuestionId":52955071,"AnswerCount":0,"Tags":"<javascript><html><tensorflow><html5-video><object-detection>","CreationDate":"2018-10-23T17:53:21.313","AcceptedAnswerId":null,"OwnerUserId":3689253.0,"Title":"object detection javascript - How to draw detection boundries on video","Body":"<p>I have html video tag with recording, and array of json objects with detections of people in every frame.<\/p>\n\n<p>when i play the video ,i'm already drawing the boxes with html and css by setInterval function, but the boxes aren't fit the position of the people in the video although the detections seems correct.\nis there more exact way to put the boxes on the video?\nHere is a sample json:\n<a href=\"https:\/\/i.stack.imgur.com\/tiqJR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tiqJR.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":52955118,"AnswerCount":1,"Tags":"<python><tensorflow><entropy><loss-function>","CreationDate":"2018-10-23T17:56:42.713","AcceptedAnswerId":52957787.0,"OwnerUserId":3935315.0,"Title":"Tensorflow pairwise custom loss for mutual information","Body":"<p>I'm learning how to use tensorflow and have run into a problem in implementing a custom loss function. Specifically, I'm trying to compute the average mutual information between all pairs of variables (the idea being to determine what predictions for one class are tightly correlated with another).<\/p>\n\n<p>For example, if I have an array<\/p>\n\n<pre><code># In simple case, 2 entries of data showing predictions for non-exclusive\n# properties A, B, C, and D in that order.\ndata = np.array([[0.99, 0.05, 0.85, 0.2], [0.97, 0.57, 0.88, 0.1]])\n<\/code><\/pre>\n\n<p>I'd like to get back a tensor showing the mutual information between A and B, A and C, A and D, B and A, etc. where A is the 1st element of each row vector, B is the 2nd, etc. I would also be ok with just getting the average pairwise mutual information for each variable (e.g. average of MI(A, B), MI(A, C), and MI(A, D))<\/p>\n\n<p>The way I would do this is by calculating the entropy across rows of every pair of variables and then subtracting off the entropy for each variable alone.<\/p>\n\n<p>As a starting point, I looked at existing code for computing the covariance of two variables:<\/p>\n\n<pre><code>def tf_cov(x):\n    mean_x = tf.reduce_mean(x, axis=0, keepdims=True)\n    mx = tf.matmul(tf.transpose(mean_x), mean_x)\n    vx = tf.matmul(tf.transpose(x), x)\/tf.cast(tf.shape(x)[0], tf.float32)\n    cov_xx = vx - mx\n    return cov_xx\n<\/code><\/pre>\n\n<p>This is a nice example of how to get pair-wise statistics, but it doesn't get me quite the metrics I want.<\/p>\n\n<p>I'm able to compute the entropy for a single variable as well:<\/p>\n\n<pre><code>def tf_entropy(prob_a):\n    # Calculates the entropy along each column\n    col_entropy = tf.reduce_sum(prob_a * tf.log(prob_a), axis=0)\n\n    return col_entropy\n<\/code><\/pre>\n\n<p>Does anyone know of a good way to compute the pairwise entropies? I imagine it will look a lot like matmul, but instead of summing the element-wise products, I would compute the entropy. Of course, if you know of existing tensorflow functions that already do what I want, that would be great. I've been reading up on various entropy-related functions, but they never seem to be quite what I want.<\/p>\n","answers":[{"AnswerId":"52957787","CreationDate":"2018-10-23T21:01:36.177","ParentId":null,"OwnerUserId":"3670165","Title":null,"Body":"<p>If you want to calculate the Mutual information between say X and Y, it depends on the underlying assumptions you can make. If you have very high dimensional data and complicated distributions, I suggest binning, which is non parametric. Also there are some more sopisticated methods that I am using.\nSee <a href=\"https:\/\/github.com\/BiuBiuBiLL\/NPEET_LNC\" rel=\"nofollow noreferrer\">here<\/a>, <a href=\"https:\/\/github.com\/gregversteeg\/NPEET\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/arxiv.org\/pdf\/1801.04062.pdf\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The first two don't really scale well and the last one involves some hyperparameter tuning that can get your numbers totally off (or I am doing smth. wrong), but scales relatively well.<\/p>\n"}]}
+{"QuestionId":52955320,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving><tensorflow-hub>","CreationDate":"2018-10-23T18:09:53.093","AcceptedAnswerId":null,"OwnerUserId":9478102.0,"Title":"Trying to use Universal Sentence Encoder Lite\/2 via Tensorflow Serving","Body":"<p>I created a SavedModel using the Universal Sentence Encoder Lite version. If I load the SavedModel using tf.saved_model.loader.load, it works perfectly fine.<\/p>\n\n<p>However, if I try to serve the model using Tensorflow Serving, I'm getting the following error:<\/p>\n\n<blockquote>\n  <p>\"error\": \"indices[3] = 1 is not in [0, 1)\\n\\t [[Node:\n  lite_module_apply_default\/Encoder_en\/KonaTransformer\/ClipToMaxLength\/GatherV2_1\n  = GatherV2[Taxis=DT_INT32, Tindices=DT_INT64, Tparams=DT_INT64, _output_shapes=[[?]], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_qsph_v_0_7,\n  lite_module_apply_default\/Encoder_en\/KonaTransformer\/ClipToMaxLength\/Reshape,\n  lite_module_apply_default_1\/Encoder_en\/KonaTransformer\/SequenceMask\/Const)]]\"<\/p>\n<\/blockquote>\n\n<p>Any reasons why that would be occurring?<\/p>\n\n<p>python version: 3.6<\/p>\n\n<p>tensorflow version: 1.9.0<\/p>\n\n<p>tensorflow_hub version: 0.1.1<\/p>\n\n<p>Using tensorflow\/serving docker 1.11<\/p>\n","answers":[{"AnswerId":"56628797","CreationDate":"2019-06-17T09:49:27.590","ParentId":null,"OwnerUserId":"9478102","Title":null,"Body":"<p>I was giving the input tensors in <strong>row format<\/strong>. By changing the format of the input tensors to <strong>columnar format<\/strong>, I was able to rectify the issue. A detailed description of row and columnar formats can be found <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/api_rest#request_format_2\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":52955529,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-23T18:22:47.677","AcceptedAnswerId":null,"OwnerUserId":9705936.0,"Title":"how I can dynamically create an list in Tensorflow","Body":"<p>how I can dynamically create a loss list from a list of tasks (self.prediction) without having to create the variables:<\/p>\n\n<p><strong>Current:<\/strong><\/p>\n\n<pre><code>loss0 = tf.losses.softmax_cross_entropy( logits = self.prediction[0], onehot_labels = self.Y[0] ) # task 0\nloss1 = tf.losses.softmax_cross_entropy( logits = self.prediction[1], onehot_labels = self.Y[1] ) # task 1\nloss2 = tf.losses.softmax_cross_entropy( logits = self.prediction[2], onehot_labels = self.Y[2] ) # task 2\n\nself.losses = tf.reduce_sum( [ loss0, loss1, loss2 ] )\n<\/code><\/pre>\n\n<p><strong>Goal:<\/strong><\/p>\n\n<pre><code>list_loss = ?\nself.losses = tf.reduce_sum( list_loss )\n<\/code><\/pre>\n","answers":[{"AnswerId":"52956512","CreationDate":"2018-10-23T19:26:38.023","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>If I understand what you mean, you're asking for<\/p>\n\n<pre><code>def calculate_loss(prediction, label, idx):\n    return tf.losses.softmax_cross_entropy(logits = prediction[idx], \n                                           onehot_labels = label[idx])\n\nlosses = []\nfor i in range(3):\n    losses.append(calculate_loss(self.prediction, self.Y, i)\nself.losses = tf.reduce_sum(losses)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52955535,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-23T18:23:01.300","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"tf.data.Dataset: Split and Turn String into Array of Integers","Body":"<p>I'm using the tf.data.Dataset API and am starting with an <code>x<\/code> numpy array and a <code>y<\/code> numpy array for my labels.  My <code>x<\/code> just consists of strings:<\/p>\n\n<p><code>array([[\"good movie\"], [\"terrible film\"]]<\/code><\/p>\n\n<p>So I split into train and test and create a <code>tf.data.Dataset<\/code>:<\/p>\n\n<pre><code>x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.1, random_state=42)\n\ndataset = tf.data.Dataset.from_tensor_slices(({\"reviews\": x_train}, y_train))\n<\/code><\/pre>\n\n<p>But now, I'd like to do this:<\/p>\n\n<pre><code>dataset[\"reviews\"] = tf.string_split(dataset[\"sequence\"])\n<\/code><\/pre>\n\n<p>But this syntax is incorrect.  I've seen <code>map<\/code> but I don't want to do it to all the columns in my <code>tf.data.Dataset<\/code>.  <\/p>\n\n<p>Also, what is the best way to turn strings into arrays of integers (each word gets a unique integer).  So:<\/p>\n\n<p><code>[\"good movie\"]<\/code> might turn into <code>[2, 4]<\/code><\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53251180","CreationDate":"2018-11-11T17:12:29.373","ParentId":null,"OwnerUserId":"1462770","Title":null,"Body":"<p>You can find a good example about <code>tf.data<\/code> in Tensorflow repository <a href=\"https:\/\/github.com\/tensorflow\/nmt\/blob\/master\/nmt\/utils\/iterator_utils.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>In the repository, in addition to its parallel dataset, it includes two dictionary files, named <code>src_vocab_file<\/code> and <code>trg_vocab_file<\/code> which contained all the vocabulary in our corpus.<\/p>\n\n<p><strong>First<\/strong>, it converts these two files into <code>src_vocab_table<\/code> and <code>trg_vocab_table<\/code> with <a href=\"https:\/\/github.com\/tensorflow\/nmt\/blob\/master\/nmt\/utils\/vocab_utils.py\" rel=\"nofollow noreferrer\">vocab_utils.py<\/a>:<\/p>\n\n<pre><code>from tensorflow.python.ops import lookup_ops\n\nsrc_vocab_table = lookup_ops.index_table_from_file(\n      src_vocab_file, default_value=UNK_ID)\n<\/code><\/pre>\n\n<p><strong>Second<\/strong>, it split, shorten, and convert the string words in the dataset into IDs with <a href=\"https:\/\/github.com\/tensorflow\/nmt\/blob\/master\/nmt\/utils\/iterator_utils.py\" rel=\"nofollow noreferrer\"><code>iterator_utils.py<\/code><\/a>:<\/p>\n\n<pre><code>src_dataset = src_dataset.map(lambda src: tf.string_split([src]).values)\nsrc_dataset = src_dataset.map(lambda src: src[:src_max_len])\nsrc_dataset = src_dataset.map(lambda src: tf.cast(src_vocab_table.lookup(src), tf.int32))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52956033,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-23T18:52:06.337","AcceptedAnswerId":null,"OwnerUserId":1610428.0,"Title":"Tensorflow: why doesn't tensorflow print each value in this loop?","Body":"<p>I was going through some Tensorflow code that runs a <code>tf.while_loop()<\/code> and I had a question. The code calculates the roots of a 4th degree polynomial as part of a lab for learning <code>tensorflow.<\/code> My question is why a specific print statement will not output all of the intermediate values. <\/p>\n\n<p>The code is below:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef f(x, w):\n    return (w[0] + w[1] * x + w[2] * tf.pow(x,2) + w[3] * tf.pow(x,3) + w[4] * tf.pow(x,4) )\n\ndef f1(x, w):\n    return (w[1] + 2. * w[2] * x + 3. * w[3] * tf.pow(x,2) + 4. * w[4] * tf.pow(x,3) )\n\ndef f2(x, w):\n    return (2. * w[2] + 6. * w[3] * x + 12. * w[4] * tf.pow(x,2) )\n\ndef fxn_plus_1(xn, w):\n    return (xn - (2. * f(xn, w) * f1(xn, w) \/ (2. * tf.square(f1(xn, w)) - f(xn, w) * f2(xn, w))))\n\ndef c(x, weights):\n    return tf.abs(x - fxn_plus_1(x, weights)) &gt; 0.001\n\ndef b(x, weights):\n    x = fxn_plus_1(x, weights)\n    return x, weights\n\nweights = tf.constant( [-1000., 1., 1. , 5. , 0.1])\n\nx = fxn_plus_1(-10., weights)\n\nout = tf.while_loop(c, b, [x, weights])\n\nwith tf.Session() as sess:\n    x, weights = sess.run(out)\n    print(x)\n<\/code><\/pre>\n\n<p>The output is correct with a values of <code>5.575055<\/code>. Now, I wanted to see what the intermediate values were of the loop body <code>b()<\/code> as the algorithm proceeded. I changed the function <code>b()<\/code> to the following:<\/p>\n\n<pre><code>def b(x, weights):\n    x = fxn_plus_1(x, weights)\n    print(x) ## ADDED PRINT STATEMENT\n    print(weights)\n    return x, weights\n<\/code><\/pre>\n\n<p>What I get in return is:<\/p>\n\n<pre><code>Tensor(\"while_4\/sub_4:0\", shape=(), dtype=float32)\nTensor(\"while_4\/Identity_1:0\", shape=(5,), dtype=float32)\n5.575055\n<\/code><\/pre>\n\n<p>this seems to give the debug output or the graph information for the values of <code>x,weights<\/code> instead of the actual values. I was not sure how to get the loop to actually print the values at each step. <\/p>\n\n<p>Any suggestions?\nThanks.<\/p>\n\n<p>UPDATE on things attempted and suggested:<\/p>\n\n<p>One of the commentors @user49593, suggested I try using <code>tf.Print()<\/code>. <\/p>\n\n<p>Here is the code, and the output. I am still just getting the graph info instead of the actual vector of values. <\/p>\n\n<pre><code>def b(x, weights):\n    x = fxn_plus_1(x, weights)\n    x = tf.Print(x, [x], message=\"here: \") #CHANGED TO tf.Print STATEMENT\n    return x, weights\n<\/code><\/pre>\n\n<p>The output is still just <code>5.575055<\/code>. No vector of intermediate values. <\/p>\n","answers":[{"AnswerId":"52977889","CreationDate":"2018-10-24T21:04:13.483","ParentId":null,"OwnerUserId":"1610428","Title":null,"Body":"<p>The answer to this question, was really found by @user49593. The problem is that the Jupyter Notebooks don't show the output of the print statements. So in order to see the output, you need to put the code into a file and then run it from the command line. <\/p>\n\n<p>So if I wrap up the code above into a file named <code>equation.py<\/code>. Then I would go to the command line and type.<\/p>\n\n<pre><code>python equation.py\n<\/code><\/pre>\n\n<p>This will generate the expected output. Looks like there is another SE post related to this, which might have some additional insights. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/37898478\/is-there-a-way-to-get-tensorflow-tf-print-output-to-appear-in-jupyter-notebook-o\">Is there a way to get tensorflow tf.Print output to appear in Jupyter Notebook output<\/a><\/p>\n"}]}
+{"QuestionId":52956268,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-23T19:07:45.697","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"tf.data.Dataset: Map fails to split string","Body":"<p>I have a <code>tf.data.Dataset<\/code> that I've created like this:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({\"reviews\": x_train}, y_train))\n<\/code><\/pre>\n\n<p>I want to split just the reviews (strings) on whitespace.  When I do this:<\/p>\n\n<pre><code>dataset = dataset.map(lambda string: tf.string_split([string]))\n<\/code><\/pre>\n\n<p>Python complains, telling me:<\/p>\n\n<pre><code>TypeError: &lt;lambda&gt;() takes exactly 1 argument (2 given)\n<\/code><\/pre>\n\n<p>I've looked at the docs and it's not obvious why Python thinks I've given two arguments...any ideas?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52956540","CreationDate":"2018-10-23T19:28:57.630","ParentId":null,"OwnerUserId":"6064179","Title":null,"Body":"<p>Looks like it has to do with the way <code>map()<\/code> is defined by tensorflow. Take a look at the docs here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#map\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#map<\/a><\/p>\n\n<p><code>map()<\/code>'s signature:<\/p>\n\n<blockquote>\n  <p>map(\n      map_func,\n      num_parallel_calls=None\n  )<\/p>\n<\/blockquote>\n\n<p>The important bit is this:<\/p>\n\n<blockquote>\n  <p>The input signature of map_func is determined by the structure of each element in this dataset.<\/p>\n<\/blockquote>\n\n<p>So your <code>dataset<\/code> must somehow be arranged as elements of tuples of size 2 which makes <code>map<\/code> pass 2 arguments into <code>map_func<\/code>. However, you define your lambda function like:<\/p>\n\n<pre><code>lambda string: tf.string_split([string])\n<\/code><\/pre>\n\n<p>which means it expects 1 input, namely <code>string<\/code>. <\/p>\n"}]}
+{"QuestionId":52956553,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><keras><deep-learning><lstm>","CreationDate":"2018-10-23T19:30:05.493","AcceptedAnswerId":null,"OwnerUserId":6709477.0,"Title":"Emotion detection on text","Body":"<p>I am a newbie in ML and was experimenting with emotion detection on the text.\nSo I have an ISEAR dataset which contains tweets with their emotion labeled.\nSo my current accuracy is 63% and I want to increase to at least 70% or even more maybe.<\/p>\n\n<p>Heres the code :<\/p>\n\n<pre><code>inputs = Input(shape=(MAX_LENGTH, ))\n\nembedding_layer = Embedding(vocab_size,\n                            64,\n                            input_length=MAX_LENGTH)(inputs)\n\n# x = Flatten()(embedding_layer)\nx = LSTM(32, input_shape=(32, 32))(embedding_layer)\nx = Dense(10, activation='relu')(x)\n\npredictions = Dense(num_class, activation='softmax')(x)\nmodel = Model(inputs=[inputs], outputs=predictions)\nmodel.compile(optimizer='adam',\n              loss='categorical_crossentropy',\n              metrics=['acc'])\n\nmodel.summary()\nfilepath=\"weights-simple.hdf5\"\ncheckpointer = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\nhistory = model.fit([X_train], batch_size=64, y=to_categorical(y_train), verbose=1, validation_split=0.1, \n          shuffle=True, epochs=10, callbacks=[checkpointer])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52957559","CreationDate":"2018-10-23T20:42:25.120","ParentId":null,"OwnerUserId":"6431410","Title":null,"Body":"<p>That's a pretty general question, optimizing the performance of a neural network may require tuning many factors.\nFor instance:<\/p>\n\n<ol>\n<li>The optimizer chosen: in NLP tasks rmsprop is also a popular\noptimizer<\/li>\n<li>Tweaking the learning rate<\/li>\n<li>Regularization - e.g dropout, recurrent_dropout, batch norm. This may help the model to generalize better<\/li>\n<li>More units in the LSTM <\/li>\n<li>More dimensions in the embedding <\/li>\n<\/ol>\n\n<p>You can try grid search, e.g. using different optimizers and evaluate on a validation set.\nThe data may also need some tweaking, such as:<\/p>\n\n<ol start=\"6\">\n<li><p>Text normalization - better representation of the tweets - remove unnecessary tokens (@, #)<\/p><\/li>\n<li><p>Shuffle the data before the fit - keras validation_split creates a validation set using the last data records<\/p><\/li>\n<\/ol>\n\n<p>There is no simple answer to your question.<\/p>\n"}]}
+{"QuestionId":52956601,"AnswerCount":1,"Tags":"<python><tensorflow><mnist>","CreationDate":"2018-10-23T19:34:07.610","AcceptedAnswerId":52957471.0,"OwnerUserId":3562088.0,"Title":"Where are the images in the MNIST data array?","Body":"<p>I am trying to have a look at the MNIST data set for machine learning. In Tensorflow the MNIST data set can be imported with <\/p>\n\n<pre><code>mnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\nfull_data_x = mnist.train.images\n<\/code><\/pre>\n\n<p>However when I try to visualize an 80x80 array of the data using<\/p>\n\n<pre><code>test_x, test_y = mnist.test.images, mnist.test.labels\nplt.gray()\nplt.imshow(1-test_x[80:160,80:160])\n<\/code><\/pre>\n\n<p>it looks really strange like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GDTDM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GDTDM.png\" alt=\"Image of the dataset array\"><\/a><\/p>\n\n<p>How can I extract an image of the actual hand-written digits, like they are shown in the internet:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zvKDg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zvKDg.png\" alt=\"Internet image of the MNIST dataset\"><\/a><\/p>\n\n<p>I saw the similar questions like <a href=\"https:\/\/stackoverflow.com\/questions\/38308378\/tensorflow-show-image-from-mnist-dataset\">this<\/a>. However I would especially be interested <em>where<\/em> in the training data array the images are actually hidden. I know that tensor flow module provides a function to display the images.<\/p>\n","answers":[{"AnswerId":"52957471","CreationDate":"2018-10-23T20:35:37.983","ParentId":null,"OwnerUserId":"10508893","Title":null,"Body":"<p>I think I understand your question now, and it is a bit different than the one I thought was duplicate.<\/p>\n\n<p>The images are not necessarily hidden. Each index of that list is an image in itself:<\/p>\n\n<pre><code>num_test_images, num_train_images = len(mnist.test.images), len(mnist.train.images)\n\nsize_of_first_test_image, size_of_first_train_image =  len(mnist.test.images[0]), len(mnist.train.images[0])\n\nprint num_test_images, num_train_images\nprint size_of_first_test_image, size_of_first_train_image\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>10000 55000\n784 784\n<\/code><\/pre>\n\n<p>You can see that the number of training and testing images is the length of each mnist list. Each image is a flat array of size 784. You will have to reshape it yourself to display it using numpy or something of the sort.<\/p>\n\n<p>Try this:<\/p>\n\n<pre><code>first_test_image = np.array(mnist.test.images[0], dtype='float')\nreshaped_image = first_image.reshape((28, 28))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52957339,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning>","CreationDate":"2018-10-23T20:25:38.313","AcceptedAnswerId":null,"OwnerUserId":10548459.0,"Title":"tensorflow model.load problems","Body":"<p>I was working on tensorflow and tried to save and load a model. model resides in below file<\/p>\n\n<pre><code>model = keras.Sequential()\nmodel.add(keras.layers.Dense(785, activation ='sigmoid' ))\nmodel.add(keras.layers.Dense(25, activation = 'sigmoid'))\nmodel.add(keras.layers.Dense(10, activation = 'sigmoid'))\n\nmodel.compile(optimizer=tf.train.GradientDescentOptimizer(0.01),\n          loss='mse',\n          metrics=['mae'])\n\nmodel.fit(X,Y,epochs = 20, callbacks=[history])\nf = h5py.File(r'C:\\Users\\akash\\Desktop\\Big Data\\Model\\model1', \"w\")\ntf.keras.models.save_model(\n    model,\n    f,\n    overwrite=True,\n    include_optimizer=True\n)\n<\/code><\/pre>\n\n<p>and my load file is as below<\/p>\n\n<pre><code>model1 = tf.keras.models.load_model(\nr'C:\\Users\\akash\\Desktop\\Big Data\\Model\\model1',\ncustom_objects=None,\ncompile=True\n)\nmodel1.compile(optimizer=tf.train.GradientDescentOptimizer(0.01),\n          loss='mse',\n          metrics=['mae'])\n<\/code><\/pre>\n\n<p>I had to compile my model again as tensorflow requires you to do so and does not allow optimisers to be saved<\/p>\n\n<p>and due to this im getting the below error<\/p>\n\n<pre><code>Using TensorFlow backend.\nWARNING:tensorflow:No training configuration found in save file: the model was *not* compiled. Compile it manually.\nTraceback (most recent call last):\n  File \"C:\/Users\/akash\/Desktop\/Big Data\/scripts\/load_model.py\", line 21, in &lt;module&gt;\nmetrics=['mae'])\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\training\\checkpointable\\base.py\", line 426, in _method_wrapper\nmethod(self, *args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 525, in compile\nmetrics, self.output_names)\nAttributeError: 'Sequential' object has no attribute 'output_names'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52957632","CreationDate":"2018-10-23T20:49:26.200","ParentId":null,"OwnerUserId":"3670165","Title":null,"Body":"<p>Maybe this helps you:<\/p>\n\n<pre><code># MLP for Pima Indians Dataset Serialize to JSON and HDF5\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.models import model_from_json\nimport numpy\nimport os\n# fix random seed for reproducibility\nnumpy.random.seed(7)\n# load pima indians dataset\ndataset = numpy.loadtxt(\"pima-indians-diabetes.csv\", delimiter=\",\")\n# split into input (X) and output (Y) variables\nX = dataset[:,0:8]\nY = dataset[:,8]\n# create model\nmodel = Sequential()\nmodel.add(Dense(12, input_dim=8, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dense(8, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dense(1, kernel_initializer='uniform', activation='sigmoid'))\n# Compile model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n# Fit the model\nmodel.fit(X, Y, epochs=150, batch_size=10, verbose=0)\n# evaluate the model\nscores = model.evaluate(X, Y, verbose=0)\nprint(\"%s: %.2f%%\" % (model.metrics_names[1], scores[1]*100))\n\n# serialize model to JSON\nmodel_json = model.to_json()\nwith open(\"model.json\", \"w\") as json_file:\n    json_file.write(model_json)\n# serialize weights to HDF5\nmodel.save_weights(\"model.h5\")\nprint(\"Saved model to disk\")\n\n# later...\n\n# load json and create model\njson_file = open('model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n# load weights into new model\nloaded_model.load_weights(\"model.h5\")\nprint(\"Loaded model from disk\")\n\n# evaluate loaded model on test data\nloaded_model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\nscore = loaded_model.evaluate(X, Y, verbose=0)\nprint(\"%s: %.2f%%\" % (loaded_model.metrics_names[1], score[1]*100))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52957566,"AnswerCount":4,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-23T20:43:07.090","AcceptedAnswerId":null,"OwnerUserId":7571462.0,"Title":"tf.estimator.export.build_raw_serving_input_receiver_fn problem","Body":"<p>I have a problem about exporting a savedmodel from an estimator of Tensorflow. My tensorflow program is using an estimator to do the CNN function, where the input is a 2D image. This is my code for the saving part.<\/p>\n\n<pre><code>            def serving_input_rec_fn():\n                serving_features = {'images': tf.placeholder(shape=[None, self.num_input[0], self.num_input[1]], dtype=tf.float32)}\n                return tf.estimator.export.build_raw_serving_input_receiver_fn(features=serving_features)\n\n            self.model.export_savedmodel(export_dir, serving_input_rec_fn,\n                                         strip_default_attrs=True)\n<\/code><\/pre>\n\n<p>But when I ran export_savedmodel function, it produced the following error:<\/p>\n\n<pre><code>AttributeError: 'function' object has no attribute 'features'\n<\/code><\/pre>\n\n<p>When I checked the code, I actually provided the serving_features here. Could any one help me solve this problem?<\/p>\n","answers":[{"AnswerId":"56347082","CreationDate":"2019-05-28T17:07:01.143","ParentId":null,"OwnerUserId":"5118339","Title":null,"Body":"<p>In serving_input_rec_fn() change:<\/p>\n\n<pre><code>def serving_input_rec_fn():\n   ...\n   return tf.estimator.export.build_raw_serving_input_receiver_fn(features=serving_features)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>def serving_input_rec_fn():\n   ...\n   return tf.estimator.export.build_raw_serving_input_receiver_fn(serving_features)\n<\/code><\/pre>\n\n<p>As of documentation (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/build_raw_serving_input_receiver_fn\" rel=\"nofollow noreferrer\">here<\/a>), <strong><em>build_raw_serving_input_receiver_fn<\/em><\/strong> function does not have a <strong>features<\/strong> argument<\/p>\n"},{"AnswerId":"55998687","CreationDate":"2019-05-06T04:09:52.353","ParentId":null,"OwnerUserId":"5642366","Title":null,"Body":"<p>In method <code>export_savedmodel()<\/code>, <\/p>\n\n<p>change<\/p>\n\n<p><strong>serving_input_rec_fn<\/strong><\/p>\n\n<p>to<\/p>\n\n<p><strong>serving_input_rec_fn()<\/strong><\/p>\n"},{"AnswerId":"54899438","CreationDate":"2019-02-27T06:37:22.027","ParentId":null,"OwnerUserId":"8110337","Title":null,"Body":"<p>you miss a bracket in the argument 'serving_input_rec_fn' passed to export_savedmodel, the right way should be like:<\/p>\n\n<pre><code>self.model.export_savedmodel(export_dir, **serving_input_rec_fn()**, strip_default_attrs=True)\n<\/code><\/pre>\n\n<p>Because the export_savedmodel api require a 'serving_input_receiver_fn' function, while the value of 'serving_input_rec_fn' is a function: 'serving_input_rec_fn'. <br\/>You need to call 'serving_input_rec_fn()' which returns tf.estimator.export.build_raw_serving_input_receiver_fn(features=serving_features) which then returns a function: 'serving_input_receiver_fn'.<\/p>\n"},{"AnswerId":"58937999","CreationDate":"2019-11-19T15:52:10.063","ParentId":null,"OwnerUserId":"3068214","Title":null,"Body":"<p>it should be like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def serving_input_rec_fn():\n    ...\n    tf.estimator.export.build_raw_serving_input_receiver_fn(serving_features)()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52957809,"AnswerCount":2,"Tags":"<python><autocomplete><sublimetext3><pytorch><code-completion>","CreationDate":"2018-10-23T21:02:39.183","AcceptedAnswerId":null,"OwnerUserId":2956066.0,"Title":"PyTorch autocomplete (code completion) in Sublime Text 3","Body":"<p>I'm a fan of Sublime Text 3 and would like to get code autocompletion for PyTorch. However, I can't get this working yet. Any suggestions or starting points where I can begin to get this working?<\/p>\n\n<p>I have searched in the packages repository of Sublime Text but unfortunately there's none.<\/p>\n\n<p>Note: I have looked at a related question here <a href=\"https:\/\/stackoverflow.com\/questions\/43287476\/ide-autocomplete-for-pytorch\">IDE autocomplete for pytorch<\/a> but that's only for VS Code.<\/p>\n","answers":[{"AnswerId":"54489613","CreationDate":"2019-02-02T03:07:51.503","ParentId":null,"OwnerUserId":"11003216","Title":null,"Body":"<p>Even though pytorch is pretty popular, I doubt that there will be any autocompletion plug-ins for it as most people use a more general possibility.<\/p>\n\n<p>Back in the days I used <a href=\"https:\/\/packagecontrol.io\/packages\/All%20Autocomplete\" rel=\"nofollow noreferrer\">All Autocomplete<\/a> for a very similar purpose.\nIt basically offers autocompletion based on all files that are currently opened. That means you can not only use it for pytorch, but virtually any library or file.\nThe only real downside is that you have to keep the file open, but I look up things there all the time anyways so it doesn't really bother me that much.<\/p>\n\n<p>I hope this is what you were looking for.\nKind regards<\/p>\n"},{"AnswerId":"54492308","CreationDate":"2019-02-02T10:50:04.077","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>Not a sublime user, but <a href=\"https:\/\/github.com\/davidhalter\/jedi\" rel=\"nofollow noreferrer\">Jedi autocompletion<\/a> handles PyTorch suggestions just fine.<\/p>\n\n<p>Jedi is an autocompletion engine and <strong>should<\/strong> work for any Python package, including the ones in virtual environment if you set it up and including PyTorch\/Tensorflow.<\/p>\n\n<p>Anyway, <a href=\"https:\/\/packagecontrol.io\/packages\/Jedi%20-%20Python%20autocompletion\" rel=\"nofollow noreferrer\">sublime version<\/a> seems to be missing a few things like results caching, hence you may have to wait a few seconds after issuing <code>torch.<\/code> (library is quite heavy, that's why you most probably will experience some lag). If you'd like to speed it up, I see no other possibility than changing it manually (vim plugin for jedi I am using has this option implemented, you may check how they've done it <a href=\"https:\/\/github.com\/zchee\/deoplete-jedi\" rel=\"nofollow noreferrer\">here<\/a>, both  jedi plugins seem to be written in Python hence should be tunable\/fixable).<\/p>\n\n<p>Oh, and if something isn't working (or maybe you would like to submit a PR request to jedi team\/sublime jedi team), the community around it is quite vibrant and you should get some help (definitely better and more in-depth than here on StackOverflow).<\/p>\n"}]}
+{"QuestionId":52958946,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2018-10-23T23:06:10.947","AcceptedAnswerId":52968363.0,"OwnerUserId":10090361.0,"Title":"How to load a data set into Jupyter Notebook","Body":"<p>When loading a dataset into Jupyter, I know it requires lines of code to load it in:<\/p>\n\n<pre><code>from tensorflow.contrib.learn.python.learn.datasets   import base\n\n # Data files\nIRIS_TRAINING = \"iris_training.csv\"\nIRIS_TEST = \"iris_test.csv\"\n\n# Load datasets.\ntraining_set =  base.load_csv_with_header(filename=IRIS_TRAINING,\n                                         features_dtype=np.float32,\n                                         target_dtype=np.int)\ntest_set = base.load_csv_with_header(filename=IRIS_TEST,\n                                    features_dtype=np.float32,\n                                    target_dtype=np.int)\n<\/code><\/pre>\n\n<p>So why is ther error <code>NotFoundError: iris_training.csv<\/code>\nstill thrown? I feel as though there is more to loading data sets on to jupyter, and would be grateful on any help on this topic<\/p>\n\n<p>I'm following a course through AI adventures, and dont know how to add in the .csv file; the video mentions nothing about how to add it on.<\/p>\n\n<p>Here is the link: <a href=\"https:\/\/www.youtube.com\/watch?v=G7oolm0jU8I&amp;list=PLIivdWyY5sqJxnwJhe3etaK7utrBiPBQ2&amp;index=3\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=G7oolm0jU8I&amp;list=PLIivdWyY5sqJxnwJhe3etaK7utrBiPBQ2&amp;index=3<\/a><\/p>\n","answers":[{"AnswerId":"52968363","CreationDate":"2018-10-24T11:54:07.207","ParentId":null,"OwnerUserId":"2955102","Title":null,"Body":"<p>The issue is that you either need to use file's absolute path i.e <strong>C:\\path_to_csv\\iris_training.csv<\/strong> for windows and for UNIX\/Linux <strong>\/path_to_csv\/iris_training.csv<\/strong> or you will need to place the file in your notebook workspace i.e directory that is being listed in your Jupyter UI which can be found at <strong><a href=\"http:\/\/localhost:8888\/tree\" rel=\"nofollow noreferrer\">http:\/\/localhost:8888\/tree<\/a><\/strong> Web UI. If you are having trouble finding the directory then just execute below python code and place the file in the printed location<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>import os\r\ncwd = os.getcwd()\r\nprint(cwd)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":52959039,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-23T23:20:30.180","AcceptedAnswerId":53200675.0,"OwnerUserId":9912116.0,"Title":"Using tf.layer.conv proper usage - activation and activity_regularizer","Body":"<p>In tensorflow's tf.layer.conv1d webpage <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv1d\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv1d<\/a>, they provide the option to set the \"activation\" and \"activity_regularizer\" for your model. I already have a model built, so it would have been nice if I could have just set activity_regularizer = tf.layer.batch_normalization(training=True) or activation=tf.nn.leaky_relu(alpha = 0.001).However, if i try to put any inputs into tf.layer.batch_normalization or tf.nn.leaky_relu they give errors saying there are arguments missing (\"inputs\" for former and \"features\" for latter).<\/p>\n\n<p>TLDR:<\/p>\n\n<pre><code>          conv1 = tf.layers.conv1d(inputs=_inputs, filters=6, kernel_size=4, strides=1,\n                                  padding='same', activation=tf.nn.leaky_relu, activity_regularizer = tf.layers.batch_normalization)\n<\/code><\/pre>\n\n<p>seems to \"work\" (it runs at least), but this:<\/p>\n\n<pre><code>          conv1 = tf.layers.conv1d(inputs=_inputs, filters=6, kernel_size=4, strides=1,\n                                  padding='same', activation=tf.nn.leaky_relu(alpha = 0.001), activity_regularizer = tf.layers.batch_normalization(training = True, trainable = True))\n<\/code><\/pre>\n\n<p>says there are missing arguments for activation and activity_regularizer.<\/p>\n\n<p>I may be using activation and activity_regularizer completely wrong so feel free to correct me. I am hoping there is an easy fix to this, otherwise the only option for me is to write extra lines of code to separate the activation and batch normalizations from conv1d. Although, I do not see the purpose of having activation and activity_regularizer built-in tf.layers.conv1d if I can't change their parameters. <\/p>\n\n<p>Last point: I am particularly worried about tf.layer.batch_normalization because I am assuming it is set at the default training=False and trainable = True which should not always be the case in the block of code that \"worked\".<\/p>\n","answers":[{"AnswerId":"53200675","CreationDate":"2018-11-08T02:24:19.587","ParentId":null,"OwnerUserId":"6780025","Title":null,"Body":"<p>The argument to <code>activation<\/code> should be a function that takes a tensor and returns a tensor. <code>tf.nn.leaky_relu<\/code> is such a function.<\/p>\n\n<p><code>tf.nn.leaky_relu(alpha = 0.001)<\/code> is not valid Python. You are not specifying the only required argument <code>features<\/code>. Python does not know how to run a function without a required argument. What you want is something like this:<\/p>\n\n<pre><code>def my_relu(features):\n    return tf.nn.leaky_relu(features, alpha=0.001)\n\ntf.layers.conv1d(..., activation=my_relu, ...)\n<\/code><\/pre>\n\n<p>The argument to <code>activity_regularizer<\/code> is a function taking your layer's activity (i.e. output) and computing a penalty for it. Usually you give high penalty for large activations. This function should output a scalar. I don't know why <code>activity_regularizer = tf.layers.batch_normalization<\/code> does not complain but it is probably doing not what you expect. Typical regularizers are <code>l1<\/code> and <code>l2<\/code> norms (see <a href=\"https:\/\/keras.io\/regularizers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/regularizers\/<\/a>). Batch normalization is a layer, not a regularizer.<\/p>\n"}]}
+{"QuestionId":52959685,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-24T00:55:20.523","AcceptedAnswerId":52960991.0,"OwnerUserId":8879559.0,"Title":"How to get the prediction of new data by LSTM in python","Body":"<p>This is a univariate time series prediction problem. As the following code shows, I divide the initial data into a train dataset (<code>trainX<\/code>) and a test dataset(<code>testX<\/code>), then I create a LSTM network by keras. Next, I train the model by the train dataset. However, when I want to get the prediction, I need to know the test value, so my problem is: why do I have to predict since I have known the true value which is test dataset in this problem. What I want to get is the prediction value of future time? If I have some misunderstandings about LSTM network, please tell me.    <\/p>\n\n<p>Thank you!  <\/p>\n\n<pre><code># create and fit the LSTM network\nmodel = Sequential()\nmodel.add(LSTM(4, input_shape=(1, look_back)))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.fit(trainX, trainY, epochs=100, batch_size=1, verbose=2)\n# make predictions\ntrainPredict = model.predict(trainX)\ntestPredict = model.predict(testX)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52960937","CreationDate":"2018-10-24T04:04:57.870","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>Since we don't have the future value with us while training the model, we just divide the data into train and test sets. Then we just imagine that test sets are the future values. We train our model using train set (and also usually a validation set). And after our model is trained, we test it using the test set to check our models performance.<\/p>\n"},{"AnswerId":"52960991","CreationDate":"2018-10-24T04:12:39.410","ParentId":null,"OwnerUserId":"1698143","Title":null,"Body":"<blockquote>\n  <p>why do I have to predict since I have known the true value which is test dataset in this problem. What I want to get is the prediction value of future time?<\/p>\n<\/blockquote>\n\n<p>In ML, we give test data X and it returns us Y. In the case of time-series, it may mislead a beginner a bit as we use the X and output is apparently X as well: The difference here is that we are inputting old values of time-series as X and the output Y is value of same time-series but we are predicting in future (can be applied for present or even past as well) as you have identified it correctly. <\/p>\n\n<p><em>(P.S: I would recommend you to begin with simple regression and then come to LSTMs etc. if all you want is to learn the Machine Learning.)<\/em><\/p>\n"}]}
+{"QuestionId":52960188,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-24T02:15:32.133","AcceptedAnswerId":null,"OwnerUserId":8262535.0,"Title":"How to calculate mean of each slice in tensorflow tensor?","Body":"<p>I have a stack of images, a tensor shape=(?, 64, 64). I am trying to apply a mask to each image, get the mean value of pixels within the mask, and compare to mean value calculated using a different mask. I would use this for comparing mean squared error. Normally, I would calculate this by looping through slices but tensorflow doesn't allow looping. Any suggestions how to do this? <\/p>\n\n<pre><code>for l in range(NumberOfExams): \nI = images[l,:,:]\n    meansRef[l] = I[referenceMasks[l,:,:]]\n    meansOut[l] = I[calculatedMasks[l,:,:]]\n\nMSE = np.mean((meansRef - meansOut)**2\/meansRef**2)\n<\/code><\/pre>\n\n<p>I have tried using tf.boolean_mask which can give me mean value based on all referenceMasks and for calculatedMasks but so far it's not clear to me how to get a per image mean value using this or another tensorflow method. <\/p>\n\n<p>Many thanks. <\/p>\n","answers":[{"AnswerId":"53565800","CreationDate":"2018-11-30T22:31:17.407","ParentId":null,"OwnerUserId":"8262535","Title":null,"Body":"<p>The comparison of masked means for each image can be achieved using the <code>tf.map_fn<\/code>. This returns the result of an arbitrary operation on each image in a batch, essentially allowing looping: <\/p>\n\n<p>So: <\/p>\n\n<pre><code>for l in range(NumberOfExams): \nI = images[l,:,:]\n    meansRef[l] = I[referenceMasks[l,:,:]]\n    meansOut[l] = I[calculatedMasks[l,:,:]]\n\nMSE = np.mean((meansRef - meansOut)**2) \n<\/code><\/pre>\n\n<p>Can be done as (can maybe use cleanup): <\/p>\n\n<pre><code>calculatedMasksInf = tf.where(calc_mask &gt; 0.5,Ex,-np.inf * tf.ones_like(calc_mask ))\nreferenceMasksInf = tf.where(ref_mask &gt; 0.5, Ex, -np.inf * tf.ones_like(ref_mask ))\n\nstiff_means_calculated = tf.map_fn(lambda frame: tf.reduce_mean(tf.boolean_mask(frame,tf.logical_not(tf.is_inf(frame)))), calculatedMasksInf)\nstiff_means_reference = tf.map_fn(lambda frame: tf.reduce_mean(tf.boolean_mask(frame,tf.logical_not(tf.is_inf(frame)))), referenceMasksInf)\n\nSqErr = (stiff_means_calculated - stiff_means_reference) ** 2\nMSE = tf.reduce_mean(tf.boolean_mask(SqErr,tf.logical_not(tf.is_nan(SqErr))))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52960338,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2018-10-24T02:40:10.973","AcceptedAnswerId":null,"OwnerUserId":10549449.0,"Title":"Why it still says that \"FailedPreconditionError: Attempting to use uninitialized value\"","Body":"<p>I'd like to create some layers when doing tensorboard. I defined all the variables and put the name into each of them. Of course I used <code>init = tf.global_variables_initializer()<\/code> and <code>sess.run('init')<\/code> to make sure that every variables being defined. But the weird thing is, the error <\/p>\n\n<pre><code>Attempting to use uninitialized value Layers\/output_layer\/weight_out\/W_out\" still happened in W_out, so as W1_Hidden_1 and W2_Hidden_2...\n<\/code><\/pre>\n\n<p>Does anyone can help me with it? Here is my code anyway<\/p>\n\n<pre><code>with tf.name_scope('Layers'):\n# Initializers\nweight_initializer = tf.contrib.layers.variance_scaling_initializer(factor=1.0, mode=\"FAN_AVG\", uniform=True)\nbias_initializer = tf.zeros_initializer()\n\nwith tf.name_scope('Hidden_layer_1'):\n    with tf.name_scope('weight_1'):\n        W_hidden_1 = tf.Variable(weight_initializer([n_stocks, n_neurons_1]), name = 'W_Hidden_1')\n    with tf.name_scope('bias_1'):\n        bias_hidden_1 = tf.Variable(bias_initializer([n_neurons_1]), name = 'Bias_Hidden_1')\n    with tf.name_scope('hidden_1_output'):\n        hidden_1 = tf.nn.relu(tf.add(tf.matmul(X, W_hidden_1), bias_hidden_1))\n\nwith tf.name_scope('Hidden_layer_2'):\n    with tf.name_scope('weight_2'):\n        W_hidden_2 = tf.Variable(weight_initializer([n_neurons_1, n_neurons_2]), name = 'W_Hidden_2')\n    with tf.name_scope('bias_2'):\n        bias_hidden_2 = tf.Variable(bias_initializer([n_neurons_2]), name = 'Bias_Hidden_2')\n    with tf.name_scope('hidden_2_output'):\n        hidden_2 = tf.nn.relu(tf.add(tf.matmul(hidden_1, W_hidden_2), bias_hidden_2))\n\nwith tf.name_scope('output_layer'):\n    with tf.name_scope('weight_out'):\n        W_out = tf.Variable(weight_initializer([n_neurons_2, 1]), name = 'W_out')\n    with tf.name_scope('bias_out'):\n        bias_out = tf.Variable(bias_initializer([1]), name = 'Bias_out')\n    with tf.name_scope('output'):\n        output = tf.transpose(tf.add(tf.matmul(hidden_2, W_out), bias_out))\n\nwith tf.name_scope('loss'):\n    # Cost function\n    loss = tf.reduce_mean(tf.square(output - Y))\n    tf.summary.scalar('loss', loss)\n\nwith tf.name_scope('train'):\n    # Optimizer\n    train_step = tf.train.AdamOptimizer().minimize(loss)\n\n# Init\ninit = tf.global_variables_initializer()\n\nmerged = tf.summary.merge_all()\n\nwith tf.Session() as sess:\n    sess.run(init)\n    writer = tf.summary.FileWriter('logs\/', sess.graph)    \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52960555,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-24T03:11:12.583","AcceptedAnswerId":null,"OwnerUserId":10547105.0,"Title":"My CNN model predicts only the first class no matter what image I put","Body":"<p>I am beginner in Machine Learning and followed a template in one of the ML courses to train images of Cats and Dogs to classify them.<\/p>\n\n<p>If I load  an image to be predicted in my model, no matter what, the prediction comes to be the first class I have defined in the list platetype at the end.<\/p>\n\n<p>I am doing this to classify other type of images however I was getting the same error to that dataset so I thought of using the classic cats and dogs dataset.<\/p>\n\n<p>I want to predict that if the image of text I give to the trained model has a fancy text or a standard text font.<\/p>\n\n<p>Here is my code.<\/p>\n\n<pre><code>#create classifier \nclassifier = Sequential()\n\n#adding convolution layer\nclassifier.add(Convolution2D(32, 3, 3, input_shape = (64, 64, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2), strides = (2,2)))\nclassifier.add(Convolution2D(32, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2), strides = (2,2)))\nclassifier.add(Flatten())\nclassifier.add(Dense(output_dim = 128, activation = 'relu'))\nclassifier.add(Dense(output_dim = 1, activation = 'sigmoid'))\n\n\n# In[54]:\n\n#compiling\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\n# In[55]:\n#making Image size same\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntraining_set = train_datagen.flow_from_directory(\n        'D:\/Third Year\/kaggle\/cats\/New Data\/Convolutional_Neural_Networks\/dataset\/training_set',\n        target_size=(64, 64),\n        batch_size=32,\n        class_mode='binary')\n\ntest_set = test_datagen.flow_from_directory(\n        'D:\/Third Year\/kaggle\/cats\/New Data\/Convolutional_Neural_Networks\/dataset\/test_set',\n        target_size=(64, 64),\n        batch_size=32,\n        class_mode='binary')\nprint('TRAINING:',training_set)\nprint('TEST: ',test_set)\n\n\n# In[56]:\n\n#checking if already a weight file exists. if it does loads it into the model\nif os.path.isfile(\"modelCNN_CD.h5\") :\n        classifier.load_weights(\"modelCNN_CD.h5\")\n\n#checkpoint saves the model.\nfilepath=\"modelCNN_CD.h5\"       \ncheckpoint1 = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n\n\nclassifier.summary()\nclassifier.fit_generator(\n        training_set,\n        steps_per_epoch=8000,\n        epochs=25,\n        validation_data=test_set,\n        validation_steps=2000,callbacks=[checkpoint1,])\n\n\n# In[67]:\n# load the model\n#model = VGG16()\n# load an image from file\nimage = load_img('D:\/Third Year\/kaggle\/cats\/New Data\/Convolutional_Neural_Networks\/dog.4029.jpg', target_size=(64, 64))\n# convert the image pixels to a numpy array\nimage = img_to_array(image)\n# reshape data for the model\nimage = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))\n\nyhat = classifier.predict(image)\nprint(yhat)\n\nimport numpy as np\nprint(platetype[np.argmax(yhat)])\n\n# In[57]:\n\nplatetype = ['Cat','Dog']\n\n# In[9]:\n\nfrom keras.models import load_model\nclassifier = load_model('modelCNN_LP.h5')\n<\/code><\/pre>\n","answers":[{"AnswerId":"52965161","CreationDate":"2018-10-24T09:12:21.270","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>It seems that the reason why your model is unable to make an accurate prediction of your new image is because you forgot to rescale it. <\/p>\n\n<p>You used ImageDataGenerator() during training with a rescaling factor of 1.\/255.<\/p>\n\n<p>Just add : <\/p>\n\n<pre><code>...\nimage = load_img('D:\/Third Year\/kaggle\/cats\/New Data\/Convolutional_Neural_Networks\/dog.4029.jpg', target_size=(64, 64))\n\nimage = img_to_array(image)\n\nimage = image\/255. \u00a0#\u00a0Add this line \n\nimage = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))\n\nyhat = classifier.predict(image)\n...\n<\/code><\/pre>\n\n<p>and it should work.<\/p>\n"},{"AnswerId":"54811584","CreationDate":"2019-02-21T16:14:06.043","ParentId":null,"OwnerUserId":"5208239","Title":null,"Body":"<p>Is your predictor always returning 0 as the class?<\/p>\n\n<p>The reason why I ask is that I was having the same problem and the issue is with <code>\"platetype[np.argmax(yhat)]\"<\/code> and the fact you are using binary class mode classification. <\/p>\n\n<p>argmax would be returning the index position of the result but as you are using binary classes and in your final layer you have 1 dense. It will only return a single value so it will always return the first class (0 as the index position). As the network is only set, to return one class. <\/p>\n\n<p>There are 2 solutions and it depends which you prefer:<\/p>\n\n<ol>\n<li>Is to change the class_mode to 'categorical' for the train and test generators, change the final dense layer from 1 to 2 so this will return scores\/probabilities for both classes. So when you use argmax, it will return the index position of the top score indicating which class it has predicted.<\/li>\n<li>The other way would be to stick with what you have got but you would have to change how to determine the class. You would use the score so yhat will be a list. You would need to access the score and based on that determine which class the model has predicted. Maybe someone can clarify this as I have not use this method and I am not sure.<\/li>\n<\/ol>\n\n<p>Hope this helps!. I had the same issue as you and this fixed it for me (I went with option 1).<\/p>\n\n<p>Let me know if it worked for you.<\/p>\n"}]}
+{"QuestionId":52960707,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2018-10-24T03:32:28.923","AcceptedAnswerId":52960805.0,"OwnerUserId":6750211.0,"Title":"How to suppress small elements in a tensor","Body":"<p>I have an LSTM network that outputs a decision via <code>tf.tanh<\/code>. Then I use a <code>tf.sign<\/code> to get the binary action of either 1, -1 or 0. However, I do not want the model to make a non zero action when output of <code>tf.tanh<\/code> is small, so I want to implement a threshold <code>T<\/code> before <code>tf.sign<\/code> so that only output larger than T or smaller than -T will be transformed to +1 and -1.<\/p>\n\n<p>Example tensor output from <code>tf.tanh<\/code>:<\/p>\n\n<pre><code>logits = [0.6,0.4,-0.6,-0.4]\nthreshold = 0.5\n<\/code><\/pre>\n\n<p>Desired tensor output while preserving the gradient:<\/p>\n\n<pre><code>action = [1,0,-1,0]\n<\/code><\/pre>\n\n<p>It's sort of like a double side ReLU with a threshold, but I don't know how to do it inside tensorflow graph. Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"52960805","CreationDate":"2018-10-24T03:46:41.270","ParentId":null,"OwnerUserId":"3033441","Title":null,"Body":"<p>Here is something that works for me. The code is pretty self-explanatory if you look at the variable names. Let me know if something is unclear.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef suppress_range(x, a, b):\n    \"\"\"Sets all the elements in the range [a, b) to 0\"\"\"\n    assert (a &lt; b), \"a must be less than b\"\n    significant = tf.logical_or(x &lt;= a, x &gt; b)\n    zero = tf.zeros_like(x)\n    return tf.where(significant, x, zero)\n\ndef main():\n    logits = tf.placeholder(tf.float32)\n    output = tf.sign(suppress_range(logits, -0.5, 0.5))\n\n    with tf.Session() as sess:\n        x = [[0.6, 0.4, -0.6, -0.4], [0.5, 0.1, -0.7, -0.2]]\n        print(sess.run(output, feed_dict={logits: x}))\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n"}]}
+{"QuestionId":52960780,"AnswerCount":0,"Tags":"<python><keras><resnet>","CreationDate":"2018-10-24T03:43:34.833","AcceptedAnswerId":null,"OwnerUserId":7330786.0,"Title":"Keras ValueError: A merge layer should be called on a list of inputs","Body":"<p>When I have already finish my function model , I can't use it the predict_classes(), and I try Sequential(layers=model.layers) to let me use predict_classes() it tell me the wrong, It's very strange     <\/p>\n\n<h3>This is my function model<\/h3>\n\n<pre><code>def resnet_v1(input_shape, depth, num_classes):\n\n    if (depth - 2) % 6 != 0:\n        raise ValueError('depth should be 6n+2 (eg: 20, 32, 44 in [a])')\n    # Start model definition.\n    num_filters = 16\n    num_res_blocks = int((depth - 2) \/ 6)\n\n    inputs = Input(shape=input_shape)\n    x = resnet_layer(inputs=inputs)\n    # Instantiate the stack of residual units\n    for stack in range(3):\n        for res_block in range(num_res_blocks):\n            strides = 1\n            if stack &gt; 0 and res_block == 0:  # first layer but not first stack\n                strides = 2  # downsample\n            y = resnet_layer(inputs=x,\n                             num_filters=num_filters,\n                             strides=strides)\n            y = resnet_layer(inputs=y,\n                             num_filters=num_filters,\n                             activation=None)\n            if stack &gt; 0 and res_block == 0:  # first layer but not first stack\n                # linear projection residual shortcut connection to match\n                # changed dims\n                x = resnet_layer(inputs=x,\n                                 num_filters=num_filters,\n                                 kernel_size=1,\n                                 strides=strides,\n                                 activation=None,\n                                 batch_normalization=False)\n            x = keras.layers.add([x, y])\n            x = Activation('relu')(x)\n            #x = LeakyReLU()(x) #\u6539relu -&gt; LeakyReLU\n        num_filters *= 2\n\n    # Add classifier on top.\n    # v1 does not use BN after last shortcut connection-ReLU\n    x = AveragePooling2D(pool_size=8)(x)\n    y = Flatten()(x)\n    outputs = Dense(num_classes,\n                    activation='softmax',\n                    kernel_initializer='he_normal')(y)\n\n    # Instantiate model.\n    model = Model(inputs=inputs, outputs=outputs)\n    return model\n\nmodel = resnet_v1(input_shape=input_shape, depth=depth)\nmodel = resnet_v1((32,32,1), 26, 26)\n<\/code><\/pre>\n\n<p>It give me the Error, when I want to change function model to Sequential model\nKeras ValueError: A merge layer should be called on a list of inputs<\/p>\n","answers":[]}
+{"QuestionId":52961507,"AnswerCount":1,"Tags":"<python><tensorflow><computer-vision><pose-estimation>","CreationDate":"2018-10-24T05:16:05.683","AcceptedAnswerId":53138543.0,"OwnerUserId":9088536.0,"Title":"Facing \"No gradients for any variable\" Error while training a SIAMESE NETWORK","Body":"<p>I'm currently building a model on Tensorflow( ver:1.8 os:Ubuntu MATE16.04) platform.\nThe model's purpose is to detect\/match Keypoints of human body.\nWhile training, the error \"No gradients for any variable\"  occurred, and I have difficulties to fix it.<\/p>\n\n<p><strong>Background of the model<\/strong>:\nIts basic ideas came from these two papers:<\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.iis.sinica.edu.tw\/~kevinlin311.tw\/cvprw15.pdf\" rel=\"nofollow noreferrer\">Deep Learning of Binary Hash Codes for fast Image Retrieval<\/a><\/li>\n<li><a href=\"http:\/\/www.iis.sinica.edu.tw\/~kevinlin311.tw\/cvpr16-deepbit.pdf\" rel=\"nofollow noreferrer\">Learning Compact Binary Descriptors with Unsupervised Deep Neural Networks<\/a><\/li>\n<\/ol>\n\n<p>They showed it's possible to match images according to Hash codes generated from a convolutional network.\nThe similarity of two pictures is determined by the Hamming distance between their corresponding hash codes.<\/p>\n\n<p>I think it's possible to develop a extremely light weight model to perform real-time human pose estimation on a video with \"constant human subject\" and \"fixed background\".<\/p>\n\n<hr>\n\n<p><strong>Model Structure<\/strong><\/p>\n\n<p><strong>01.Data source:<\/strong><\/p>\n\n<p>3 images from one video with the same human subject and similar background.\n               Every human keypoints in each image are well labeled.\n               2 of the images will be used as the \"hint sources\" and the last image will be the target for keypoint detection\/matching.<\/p>\n\n<p><strong>02.Hints:<\/strong><\/p>\n\n<p>23x23pixel ROIs will be cropped from the \"hint source\" images according to the location of human keypoints.\n          The center of these ROIs are the keypoints.<\/p>\n\n<p><strong>03.convolutional network \"for Hints\":<\/strong><\/p>\n\n<p>A simple 3-layered structure.\nThe first two layers are  convolution by [2,2] stride with a 3x3 filter.\n          The last layer is a 5x5 convolution on a 5x5 input with no padding(equals to a fully connected layer)<\/p>\n\n<p>This will turn a 23x23pixel Hint ROI into one 32 bit Hash codes.\n          One hint souce image will generate a set of 16 Hash codes.<\/p>\n\n<p><strong>04.Convolutional network \"for target image\":<\/strong>\nThe network share the smae weights with the hint network.\nBut in this case, each convolution layer have paddings.\nThe 301x301pixel image will be turned into a 76x76 \"Hash map\"<\/p>\n\n<p><strong>05.Hash matching:<\/strong><\/p>\n\n<p>I made a function called \" locateMin_and_get_loss \" to calculate the Hamming distance between \"hint hash\" and the hash codes on each point of the hash map.\nThis function will create a \"distance map\".\nhe location of the point with lowest distance value will be treated as the location of the keypoint.<\/p>\n\n<p><strong>06.Loss calculation:<\/strong><\/p>\n\n<p>I made a function \"get_total_loss_and_result\" to calculate the total loss of 16 keypoints.\nThe loss are normalized euclidean distance between ground truth label points and the points located by the model.<\/p>\n\n<p><strong>07.proposed work flow:<\/strong><\/p>\n\n<p>Before initializing this model, the user will take two pictures of the target human subject from different angles.\nThe pictures will be labeled by the state of art models like OpenPose or DeepPose and generate Hint Hashs from them with convolution network mentioned in 03.<\/p>\n\n<p>Finally the video stream will be started and processd by the model.<\/p>\n\n<p><strong>08.Why \"Two\" sets of hints?<\/strong><\/p>\n\n<p>One human joint\/keypoint observed from different angles will have very diferent appearance.\nInstead of increasing dimetionality of the neural networ, I want to \"cheat the game\" by gathering two hints instead of one.\nI want to know whether it can increase the precision and generalizational capacity of the model or not.<\/p>\n\n<hr>\n\n<p><strong>The problems I faced:<\/strong><\/p>\n\n<p>01.The \"No gradients for any variable \" error\n<strong>(My main question of this post)<\/strong>:<\/p>\n\n<p>As mentioned above, I'm facing this error while training the model.\n    I tried to learn from posts like <a href=\"https:\/\/stackoverflow.com\/questions\/38778760\/tensorflow-no-gradients-provided-for-any-variable\">this<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/41689451\/valueerror-no-gradients-provided-for-any-variable\">this<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/14894\" rel=\"nofollow noreferrer\">this<\/a>.\n    But currently I have no clue even though I checked the computational graph.<\/p>\n\n<p>02.The \"Batch\" problem:<\/p>\n\n<p>Due to its unique structure, it's hard to use conventional placeholder to contain the input data of multiple batch.\n    I fixed it by setting the batch number to 3 and manually combine the value of loss functions.<\/p>\n\n<blockquote>\n  <p>2018.10.28 Edit:<\/p>\n<\/blockquote>\n\n<p>The simplified version with only one hint set:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport time\nfrom imageLoader import getPaddedROI,training_data_feeder\nimport math\n'''\ncreated by Cid Zhang \na sub-model for human pose estimation\n'''\ndef truncated_normal_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.truncated_normal_initializer(stddev=0.01)))\ndef zero_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.constant_initializer(0.0)))\n\nroi_size = 23\nimage_input_size = 301\n\n#input placeholders\n#batch1 hints\ninputs_b1h1 = tf.placeholder(tf.float32, ( 16, roi_size, roi_size, 3), name='inputs_b1h1')\n\ninputs_s = tf.placeholder(tf.float32, (None, image_input_size, image_input_size, 3), name='inputs_s')\nlabels = tf.placeholder(tf.float32,(16,76,76), name='labels')\n\n#define the model\ndef paraNet(input):\n    out_l1 = tf.layers.conv2d(input, 8, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_1')\n    out_l1 = tf.nn.relu6(out_l1)\n    out_l2 = tf.layers.conv2d(out_l1, 16, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_2')\n    out_l2 = tf.nn.relu6(out_l2)\n    out_l3 = tf.layers.conv2d(out_l2, 32, [5, 5],strides=(1, 1), padding ='valid' ,name='para_conv_3')\n    return out_l3\n\n#network pipeline to create the first Hint Hash Sets (Three batches)\nwith tf.variable_scope('conv'):\n    out_b1h1_l3 = paraNet(inputs_b1h1)\n    #flatten and binerize the hashs\n    out_b1h1_l3 =tf.squeeze(  tf.round(tf.nn.sigmoid(out_b1h1_l3)) )\n\n\nwith tf.variable_scope('conv', reuse=True):\n    out_2_l1 = tf.layers.conv2d(inputs_s,  8, [3, 3],strides=(2, 2),     padding ='same' ,name='para_conv_1')\n    out_2_l1 = tf.nn.relu6(out_2_l1)\n    out_2_l2 = tf.layers.conv2d(out_2_l1, 16, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_2')\n    out_2_l2 = tf.nn.relu6(out_2_l2)\n    out_2_l3 = tf.layers.conv2d(out_2_l2, 32, [5, 5],strides=(1, 1), padding ='same' ,name='para_conv_3')\n    #binerize the value into Hash code\n\n    out_2_l3 = tf.round(tf.nn.sigmoid(out_2_l3))\n\n    orig_feature_map_size = tf.shape(out_2_l3)[1]\n\n    #calculate Hamming distance maps\n    map0 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[0] , out_2_l3 ) ) , axis=3 )  \n    map1 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[1] , out_2_l3 ) ) , axis=3 )  \n    map2 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[2] , out_2_l3 ) ) , axis=3 )  \n    map3 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[3] , out_2_l3 ) ) , axis=3 )  \n    map4 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[4] , out_2_l3 ) ) , axis=3 )  \n    map5 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[5] , out_2_l3 ) ) , axis=3 )  \n    map6 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[6] , out_2_l3 ) ) , axis=3 )  \n    map7 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[7] , out_2_l3 ) ) , axis=3 )  \n    map8 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[8] , out_2_l3 ) ) , axis=3 )  \n    map9 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[9] , out_2_l3 ) ) , axis=3 )  \n    map10 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[10] , out_2_l3 ) ) , axis=3 )  \n    map11 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[11] , out_2_l3 ) ) , axis=3 )  \n    map12 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[12] , out_2_l3 ) ) , axis=3 )  \n    map13 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[13] , out_2_l3 ) ) , axis=3 )  \n    map14 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[14] , out_2_l3 ) ) , axis=3 )  \n    map15 = tf.reduce_sum ( tf.abs (tf.subtract( out_b1h1_l3[15] , out_2_l3 ) ) , axis=3 )  \n\n    totoal_map =tf.div( tf.concat([map0, map1, map2, map3, map4, map5, map6, map7,\n                               map8, map9, map10,map11,map12, map13, map14, map15], 0) , 32)\n    loss = tf.nn.l2_loss(totoal_map - labels  , name = 'loss'  )\n\n#ValueError: No gradients provided for any variable, check your graph     for ops that do not support gradients, between variables \n    train_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss )\n\n\ninit =  tf.global_variables_initializer()\nbatchsize = 3\n\nwith tf.Session() as sess:\n#writer = tf.summary.FileWriter(\".\/variable_graph\",graph = sess.graph)\nsess.run(init)\n\n#load image from dataset(train set)\njoint_data_path = \".\/custom_data.json\"\ntrain_val_path = \".\/train_val_indices.json\"\nimgpath = \".\/000\/\"\ninput_size = 301\nhint_roi_size = 23\n\nhintSet01_norm_batch = []\nhintSet02_norm_batch = []\nt_img_batch = []\nt_label_norm_batch = []\n#load data\nhintSet01,hintSet02,t_img,t_label_norm = training_data_feeder(joint_data_path, train_val_path, imgpath, input_size, hint_roi_size )\n#Normalize the image pixel values to 0~1\nhintSet01_norm = []\nhintSet02_norm = []\n\nt_img = np.float32(t_img \/255.0)\n\nfor rois in hintSet01:\n    tmp = np.float32(rois \/ 255.0)\n    hintSet01_norm.append(tmp.tolist())\nfor rois in hintSet02:\n    tmp = np.float32(rois \/ 255.0)\n    hintSet02_norm.append(tmp.tolist())\n\nprint(tf.trainable_variables())\n\ntemp = sess.run(totoal_map , feed_dict={inputs_s:  [t_img]  , \n                                    inputs_b1h1: hintSet01_norm, \n                                    labels: t_label_norm \n                                                   })\nprint(temp)\nprint(np.shape(temp))\n<\/code><\/pre>\n\n<p>The code: <a href=\"https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/paraposeNetworkV3.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/paraposeNetworkV3.py<\/a><\/p>\n\n<p>The Tensorflow graph: <a href=\"https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/variable_graph\/events.out.tfevents.1540296979.pharmboy-K30AD-M31AD-M51AD\" rel=\"nofollow noreferrer\">https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/variable_graph\/events.out.tfevents.1540296979.pharmboy-K30AD-M31AD-M51AD<\/a><\/p>\n\n<p>The Dataset: <\/p>\n\n<p>It's a custom dataset generated from mpii dataset.\n             It have 223 clusters of images.\n             Each cluster have one constant human subject in various poses and the background remains the same.\nOne cluster have at least 3 pictures. It's about 627MB and I'll try to pack it and upload it later.<\/p>\n\n<blockquote>\n  <p>2018.10.26 Edit:<\/p>\n<\/blockquote>\n\n<p>You can download it on GoogleDrive, the whole data set was divided into 9 parts.( I can't post more than 8 links in this article. The links are in this file:\n<a href=\"https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/000\/readme.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/000\/readme.md<\/a><\/p>\n","answers":[{"AnswerId":"53138543","CreationDate":"2018-11-04T07:06:00.493","ParentId":null,"OwnerUserId":"9088536","Title":null,"Body":"<p>I used \"eager execution\" described in <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/eager<\/a> to check the gradient.<\/p>\n\n<p>In the end I found \"tf.round\" and \"tf.nn.relu6\" will erase or set the gradient to zero.<\/p>\n\n<p>I made some modification to the code and now I can enter the training phase:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport time\nfrom imageLoader import getPaddedROI,training_data_feeder\nimport math\nimport cv2\n'''\ncreated by Cid Zhang \na sub-model for human pose estimation\n'''\ntf.reset_default_graph()\n\ndef truncated_normal_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.truncated_normal_initializer(stddev=0.01)))\ndef zero_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.constant_initializer(0.0)))\n\nroi_size = 23\nimage_input_size = 301\n\n#input placeholders\n#batch1 hints\ninputs_b1h1 = tf.placeholder(tf.float32, ( 16, roi_size, roi_size, 3), name='inputs_b1h1')\n#inputs_b1h2 = tf.placeholder(tf.float32, ( 16, roi_size, roi_size, 3), name='inputs_b1h2')\n\n\ninputs_s = tf.placeholder(tf.float32, (None, image_input_size, image_input_size, 3), name='inputs_s')\nlabels = tf.placeholder(tf.float32,(16,76,76), name='labels')\n\n#define the model\n\ndef paraNet(inputs, inputs_s):\n    with tf.variable_scope('conv'):\n        out_l1 = tf.layers.conv2d(inputs, 16, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_1')\n        out_l1r = tf.nn.relu(out_l1)\n        out_l2 = tf.layers.conv2d(out_l1r, 48, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_2')\n        out_l2r = tf.nn.relu(out_l2)\n        out_l3 = tf.layers.conv2d(out_l2r, 96, [5, 5],strides=(1, 1), padding ='valid' ,name='para_conv_3')\n        out_l3r = tf.nn.relu(out_l3)\n        out_l4 = tf.layers.conv2d(out_l3r, 32, [1, 1],strides=(1, 1), padding ='valid' ,name='para_conv_4')\n        out_l4r = tf.squeeze(  tf.sign( tf.sigmoid(out_l4) ) )\n\n    with tf.variable_scope('conv', reuse=True):\n        out_2_l1 = tf.layers.conv2d(inputs_s,  16, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_1')\n        out_2_l1r = tf.nn.relu(out_2_l1)\n        out_2_l2 = tf.layers.conv2d(out_2_l1r, 48, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_2')\n        out_2_l2r = tf.nn.relu(out_2_l2)\n        out_2_l3 = tf.layers.conv2d(out_2_l2r, 96, [5, 5],strides=(1, 1), padding ='same' ,name='para_conv_3')\n        out_2_l3r = tf.nn.relu(out_2_l3)\n        out_2_l4 = tf.layers.conv2d(out_2_l3r, 32, [1, 1],strides=(1, 1), padding ='same' ,name='para_conv_4')\n        out_2_l4r =tf.sign( tf.sigmoid(out_2_l4))\n    return out_l4r , out_2_l4r  \n\ndef lossFunc(inputs_hint, inputs_sample, labels):    \n    hint, sample = paraNet(inputs_hint, inputs_sample)\n\n    map0 = tf.reduce_sum ( tf.abs (tf.subtract( hint[0] , sample ) ) , axis=3 )  \n    map1 = tf.reduce_sum ( tf.abs (tf.subtract( hint[1] , sample ) ) , axis=3 )  \n    map2 = tf.reduce_sum ( tf.abs (tf.subtract( hint[2] , sample ) ) , axis=3 )  \n    map3 = tf.reduce_sum ( tf.abs (tf.subtract( hint[3] , sample ) ) , axis=3 )  \n    map4 = tf.reduce_sum ( tf.abs (tf.subtract( hint[4] , sample ) ) , axis=3 )  \n    map5 = tf.reduce_sum ( tf.abs (tf.subtract( hint[5] , sample ) ) , axis=3 )  \n    map6 = tf.reduce_sum ( tf.abs (tf.subtract( hint[6] , sample ) ) , axis=3 )  \n    map7 = tf.reduce_sum ( tf.abs (tf.subtract( hint[7] , sample ) ) , axis=3 )  \n    map8 = tf.reduce_sum ( tf.abs (tf.subtract( hint[8] , sample ) ) , axis=3 )  \n    map9 = tf.reduce_sum ( tf.abs (tf.subtract( hint[9] , sample ) ) , axis=3 )  \n    map10 = tf.reduce_sum ( tf.abs (tf.subtract( hint[10] , sample ) ) , axis=3 )  \n    map11 = tf.reduce_sum ( tf.abs (tf.subtract( hint[11] , sample ) ) , axis=3 )  \n    map12 = tf.reduce_sum ( tf.abs (tf.subtract( hint[12] , sample ) ) , axis=3 )  \n    map13 = tf.reduce_sum ( tf.abs (tf.subtract( hint[13] , sample ) ) , axis=3 )  \n    map14 = tf.reduce_sum ( tf.abs (tf.subtract( hint[14] , sample ) ) , axis=3 )  \n    map15 = tf.reduce_sum ( tf.abs (tf.subtract( hint[15] , sample ) ) , axis=3 )  \n\n    totoal_map =tf.div( tf.concat([map0, map1, map2, map3, map4, map5, map6, map7,\n                               map8, map9, map10,map11,map12, map13, map14, map15], 0) , 64)\n    loss = tf.nn.l2_loss( totoal_map -  labels , name = 'loss'  )\n    return loss, totoal_map\n\nloss, totoal_map = lossFunc(inputs_b1h1, inputs_s, labels)\ntrain_step = tf.train.GradientDescentOptimizer(2.0).minimize(loss)\n\n#init =  tf.global_variables_initializer()\n\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess:\n    #writer = tf.summary.FileWriter(\".\/variable_graph\",graph = sess.graph)\n    #sess.run(init)\n\n    #load image from dataset(train set)\n    joint_data_path = \".\/custom_data.json\"\n    train_val_path = \".\/train_val_indices.json\"\n    imgpath = \".\/000\/\"\n    input_size = 301\n    hint_roi_size = 23\n    '''\n    #load data\n    hintSet01,hintSet02,t_img,t_label_norm = training_data_feeder(joint_data_path,     train_val_path, imgpath, input_size, hint_roi_size )\n    #Normalize the image pixel values to 0~1\n    hintSet01_norm = []\n    hintSet02_norm = []\n\n    t_img =[ np.float32(t_img \/255.0) ]\n    #print(type(t_img))\n    #print(np.shape(t_img))\n    #print(type(t_label_norm))\n    for rois in hintSet01:\n        tmp = np.float32(rois \/ 255.0)\n        hintSet01_norm.append(tmp.tolist())\n    for rois in hintSet02:\n        tmp = np.float32(rois \/ 255.0)\n        hintSet02_norm.append(tmp.tolist())\n\n    loss_value , total_map_value = sess.run ([loss, totoal_map], feed_dict = {inputs_s:  t_img, \n                                                                                                  inputs_b1h1: hintSet01_norm, \n                                                                          labels:     t_label_norm\n                                                                          })\n    print(\"-----loss value:\",loss_value)\n    print(\"-----total_map_value:\", total_map_value[0,0] )\n    print(\"-----label_value\", t_label_norm[0,0] )\n    #cv2.imshow(\"t_img\",t_img[0])\n    #for img in t_label_norm:\n    #    print(img)\n    #    cv2.imshow(\"hint\", img)\n    #    cv2.waitKey(0)\n\n    #print(tf.trainable_variables())\n    #print(hash_set01)\n    #print(out_2_l3)\n    '''\n    saver.restore(sess, \".\/temp_model\/model4.ckpt\")\n\n\n    for i in range(1000):\n\n        #load data\n        hintSet01,hintSet02,t_img,t_label_norm = training_data_feeder(joint_data_path, train_val_path, imgpath, input_size, hint_roi_size )\n        #Normalize the image pixel values to 0~1\n        hintSet01_norm = []\n        hintSet02_norm = []\n\n        t_img =[ np.float32(t_img \/255.0) ]\n        #print(type(t_img))\n        #print(np.shape(t_img))\n        #print(type(t_label_norm))\n        for rois in hintSet01:\n            tmp = np.float32(rois \/ 255.0)\n            hintSet01_norm.append(tmp.tolist())\n        for rois in hintSet02:\n            tmp = np.float32(rois \/ 255.0)\n            hintSet02_norm.append(tmp.tolist())\n        loss_val, _ = sess.run([loss, train_step] , \n                      feed_dict = {inputs_s:  t_img, \n                                   inputs_b1h1: hintSet01_norm, \n                                   labels: t_label_norm })\n        if i % 50 == 0:\n            print(loss_val)\n\n    save_path = saver.save(sess, \".\/temp_model\/model\" + '5' + \".ckpt\")\n    #print(temp)\n    #print(np.shape(temp))\n<\/code><\/pre>\n\n<p>But unfortunately the loss value was not decreasing during the training.<\/p>\n\n<p>I think there are still some bugs in the code.\nThe saved check point file are always named \"XXXX.ckpt.data-00000-of-00001\" no matter how long the iteration I set.<\/p>\n\n<p>I'll make another post about it since the main problem of this post is solved.<\/p>\n"}]}
+{"QuestionId":52961541,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-24T05:19:41.287","AcceptedAnswerId":null,"OwnerUserId":10461729.0,"Title":"Tensorflow model detecting wrong objects?","Body":"<p>I am new to Tensorflow. I trained a model with 900 images of shoes. I put 20%(180 images) into test folder and 80%(720 images) into train folder. But after training, my trained model is detecting other objects also as shoes. Attached below is the screen shot of the prediction by my model.<\/p>\n\n<p><strong>Question1:<\/strong>  Can anyone can help me please, where am i going wrong?<\/p>\n\n<p>I am training this model on MAC Machine, and using faster_rcnn_inception_v2_coco_2018_01_28 model to train my model.<\/p>\n\n<p>I follow the link to train model:\n<a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\/tree\/d1c5b59803543e48362c27c48d704d4b0d92d135\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\/tree\/d1c5b59803543e48362c27c48d704d4b0d92d135<\/a><\/p>\n\n<p><strong>Question 2:<\/strong> And more thing when i run this model on webcam it is very slow why?<\/p>\n\n<p>Thanks in advance...<\/p>\n\n<p><strong>Screenshots<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UVHKX.png\" rel=\"nofollow noreferrer\">Please check this Image detecting wrong object as shoe<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IXsLp.png\" rel=\"nofollow noreferrer\">Please check this image detecting shoe<\/a><\/p>\n","answers":[]}
+{"QuestionId":52961581,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-24T05:23:38.807","AcceptedAnswerId":52961701.0,"OwnerUserId":9422637.0,"Title":"Keras, get output of a layer at each epochs","Body":"<p><strong>What I have done?<\/strong><\/p>\n\n<p>I implemented a <code>keras<\/code> model as follow:<\/p>\n\n<pre><code>train_X, test_X, train_Y, test_Y = train_test_split(X, Y, test_size=0.2, random_state=np.random.seed(7), shuffle=True)\n\ntrain_X = np.reshape(train_X, (train_X.shape[0], 1, train_X.shape[1]))\ntest_X = np.reshape(test_X, (test_X.shape[0], 1, test_X.shape[1]))\n\nmodel = Sequential()\nmodel.add(LSTM(100, return_sequences=False, input_shape=(train_X.shape[1], train_X.shape[2])))\nmodel.add(Dense(train_Y.shape[1], activation='softmax'))\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\n\nmodel.fit(train_X, train_Y, validation_split=.20,\n                        epochs=1000, batch_size=50)\n<\/code><\/pre>\n\n<p><strong>What i want?<\/strong><\/p>\n\n<p>I want to give <code>support vector machine(SVM)<\/code> the output of the penultimate layer (LSTM), in any <code>epoch<\/code>(that is 1000) to <code>svm<\/code> also be trained.<\/p>\n\n<p>But I do not know how to do this?<\/p>\n\n<p>Any idea?<\/p>\n\n<p><strong>UPDATED:<\/strong><\/p>\n\n<p>I use from <code>ModelCheckpoint<\/code> as follow:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(100, return_sequences=False, input_shape=(train_X.shape[1], train_X.shape[2])))\nmodel.add(Dense(train_Y.shape[1], activation='softmax'))\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\n\n# checkpoint\nfilepath=\"weights-{epoch:02d}-{val_acc:.2f}.hdf5\"\ncheckpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\ncallbacks_list = [checkpoint]\n\nmodel.fit(train_X, train_Y, validation_split=.20,\n                    epochs=1000, batch_size=50, callbacks=callbacks_list, verbose=0)\n<\/code><\/pre>\n\n<p><strong>OUTPUT:<\/strong><\/p>\n\n<pre><code>Epoch 00991: val_acc did not improve\nEpoch 00992: val_acc improved from 0.93465 to 0.93900, saving model to weights-992-0.94.hdf5\nEpoch 00993: val_acc did not improve\nEpoch 00994: val_acc did not improve\nEpoch 00995: val_acc did not improve\nEpoch 00996: val_acc did not improve\nEpoch 00997: val_acc did not improve\nEpoch 00998: val_acc improved from 0.93900 to 0.94543, saving model to weights-998-0.94.hdf5\nEpoch 00999: val_acc did not improve\n<\/code><\/pre>\n\n<p><strong>PROBLEM:<\/strong><\/p>\n\n<p>How to load all these models to obtain the output of the LSTM layer in each epochs as @IonicSolutions said?<\/p>\n","answers":[{"AnswerId":"52961701","CreationDate":"2018-10-24T05:38:29.407","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>What works best in your situation depends on how exactly you set up and train your SVM, but there are at least two options using callbacks:<\/p>\n\n<p>You could use the <a href=\"https:\/\/keras.io\/callbacks\/#modelcheckpoint\" rel=\"nofollow noreferrer\"><code>ModelCheckpoint<\/code> callback<\/a> to save a copy of the model you are training at each epoch and then later load all these models to obtain the output of the LSTM layer.<\/p>\n\n<p>You can also create your own callback by implementing the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L148\" rel=\"nofollow noreferrer\"><code>Callback<\/code> base class<\/a>. Within the callback, the model can be accessed and you can use <code>on_epoch_end<\/code> to extract the LSTM output at the end of each epoch.<\/p>\n\n<p><strong>Edit:<\/strong> To get convenient access to the penultimate layer, you can do the following:<\/p>\n\n<pre><code># Create the model with the functional API\ninp = Input((train_X.shape[1], train_X.shape[2],))\nlstm = LSTM(100, return_sequences=False)(inp)\ndense = Dense(train_Y.shape[1], activation='softmax')(lstm)\n\n# Create the full model\nmodel = Model(inputs=inp, outputs=dense)\n\n# Create the model for access to the LSTM layer\naccess = Model(inputs=inp, outputs=lstm)\n<\/code><\/pre>\n\n<p>Then, you can pass <code>access<\/code> to your callback when you instantiate it. The key thing to note here is that <code>model<\/code> and <code>access<\/code> share the very same LSTM layer, whose weights will change when training <code>model<\/code>.<\/p>\n"}]}
+{"QuestionId":52961627,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-24T05:29:38.893","AcceptedAnswerId":null,"OwnerUserId":274627.0,"Title":"How to prepare an array for an LSTM?","Body":"<p>The input array has 4 columns: the first 3 are real numbers between -1 and 1 and the 4th is the desired output that can be either -1 or 0 or +1 (this is a classification problem, -1 and 1 represent useful categories and 0 means the sample doesn't fall to any of them). Obviously the network type is chosen to be LSTM because the category depends not only on the values of the first 3 columns but also  on the properties of the preceding samples.<\/p>\n\n<p>With a simple non-LSTM network I would do (but I'm not sure even this is correct as I'm totally new to Keras and using Python for neural networks, I've only used Encog in Java and C# before and it didn't have LSTMs)<\/p>\n\n<pre><code># dataset = the array I've described above\nx = dataset[:, 0:3]\ny = dataset[:, 3]\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit(x, y, epochs=epochs, batch_size=10)\n<\/code><\/pre>\n\n<p>But with an LSTM this raises ValueError as an LSTM requires a 3D input array and I don't understand how to reshape the array I've described correctly.<\/p>\n","answers":[{"AnswerId":"52961840","CreationDate":"2018-10-24T05:50:45.857","ParentId":null,"OwnerUserId":"4283749","Title":null,"Body":"<p>if you're using Keras and have access to a GPU I'd recommend using a <code>CuDNNLSTM<\/code> instead of a regular LSTM for faster training. Back to your question though.<\/p>\n\n<p>The LSTM requires a 3D input so here's a quick example to get you started. Let's say you have the following two columns of tabular data:<\/p>\n\n<pre><code>series 1: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0\nseries 2: 1.0, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1\n<\/code><\/pre>\n\n<p>So the data is present as follows (10, 2):<\/p>\n\n<pre><code>from numpy import array\ndata = array([\n    [0.1, 1.0],\n    [0.2, 0.9],\n    [0.3, 0.8],\n    [0.4, 0.7],\n    [0.5, 0.6],\n    [0.6, 0.5],\n    [0.7, 0.4],\n    [0.8, 0.3],\n    [0.9, 0.2],\n    [1.0, 0.1]])\n<\/code><\/pre>\n\n<p>In order to reshape this to a 3D input, you'd do the following:<\/p>\n\n<pre><code>data = data.reshape(1, 10, 2)\n<\/code><\/pre>\n\n<p>where you're reshaping the data into a 1 sample of 10 time steps and 2 features.<\/p>\n\n<p>Source:\n<a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/<\/a><\/p>\n\n<p>If you need a more direct answer let me know.<\/p>\n"}]}
+{"QuestionId":52961962,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-24T06:02:26.330","AcceptedAnswerId":52963207.0,"OwnerUserId":3259896.0,"Title":"How to convert a tensorflow Graph into using the Estimator API","Body":"<p>I'm a little stuck on trying to convert my tensorflow code into using the Estimator API<\/p>\n\n<p>My graph is below<\/p>\n\n<pre><code>batch_size = 1024\n\nembedding_size = 500 # 2^8 Dimension of the embedding vector. Crashed at 158 for Embed size 2656016. So possible values are 154-157. Possible choices 154, 156\nnum_inputs =5\n\nnum_sampled = 128 # Number of negative examples to sample.\n\ngraph = tf.Graph()\n\nwith graph.as_default(): \n\n    train_dataset = tf.placeholder(tf.int32, shape=[batch_size, num_inputs ])\n    train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])\n\n    epochCount = tf.get_variable( 'epochCount', initializer= 0) #to store epoch count to total # of epochs are known\n    update_epoch = tf.assign(epochCount, epochCount + 1)\n\n    embeddings = tf.get_variable( 'embeddings', dtype=tf.float32,\n        initializer= tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0, dtype=tf.float32) )\n\n    softmax_weights = tf.get_variable( 'softmax_weights', dtype=tf.float32,\n        initializer= tf.truncated_normal([vocabulary_size, embedding_size],\n                             stddev=1.0 \/ math.sqrt(embedding_size), dtype=tf.float32 ) )\n\n    softmax_biases = tf.get_variable('softmax_biases', dtype=tf.float32,\n        initializer= tf.zeros([vocabulary_size], dtype=tf.float32),  trainable=False )\n\n    embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n    embed_reshaped = tf.reshape( embed, [batch_size*num_inputs, embedding_size] )\n\n    segments= np.arange(batch_size).repeat(num_inputs)\n\n    averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n    loss = tf.reduce_mean(\n        tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds, \n                                   sampled_values=tf.nn.uniform_candidate_sampler(true_classes=tf.cast(train_labels, tf.int64), num_sampled=num_sampled, num_true=1, unique=True, range_max=vocabulary_size, seed=None),\n                                   labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)) \n\n    optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) \n\n    saver = tf.train.Saver()\n<\/code><\/pre>\n\n<p>I'm reading the official Estimator guide here<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator<\/a><\/p>\n\n<p>My best interpretation on what to do is that I have to convert my graph into a function? And set that to <code>model_fn<\/code>?<\/p>\n\n<p>'params' are the my embeddings and softmax weights. So do I explicitly name these variables here?<\/p>\n\n<p>Edit: <\/p>\n\n<p>Thanks to answers from Sorin, this is my attempt at converting this to using TF.Estimator<\/p>\n\n<p>This is the function I used to generate batches of data. <\/p>\n\n<p>data_index = 0\nepoch_index = 0\nrecEpoch_indexA = 0 #Used to help keep store of the total number of epoches with the models<\/p>\n\n<pre><code>def generate_batch(batch_size, inputCount): #batch size = number of labels\n  #inputCount = number of inputs per label\n    global data_index, epoch_index\n\n    batch = np.ndarray(shape=(batch_size, inputCount), dtype=np.int32) \n    labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)\n\n    #Begin New stuff\n\n    n=0\n    while n &lt; batch_size:\n      if len(    set(my_data[data_index, 1])   ) &gt;= inputCount:\n        labels[n,0] = my_data[data_index, 0]\n        batch[n] = random.sample( set(my_data[data_index, 1]),  inputCount)\n        n = n+1\n        data_index = (data_index + 1) % len(my_data) #may have to do something like len my_data[:]\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n      else:\n        data_index = (data_index + 1) % len(my_data)\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n\n    return batch, labels     \n<\/code><\/pre>\n\n<p>This is where I define my model function. I commented out the place holders and put in the train data and train labels as the input<\/p>\n\n<pre><code>def my_model( train_dataset, train_labels):\n\n#     train_dataset = tf.placeholder(tf.int32, shape=[batch_size, num_inputs ])\n#     train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])\n\n    epochCount = tf.get_variable( 'epochCount', initializer= 0) #to store epoch count to total # of epochs are known\n    update_epoch = tf.assign(epochCount, epochCount + 1)\n\n    embeddings = tf.get_variable( 'embeddings', dtype=tf.float32,\n        initializer= tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0, dtype=tf.float32) )\n\n    softmax_weights = tf.get_variable( 'softmax_weights', dtype=tf.float32,\n        initializer= tf.truncated_normal([vocabulary_size, embedding_size],\n                             stddev=1.0 \/ math.sqrt(embedding_size), dtype=tf.float32 ) )\n\n    softmax_biases = tf.get_variable('softmax_biases', dtype=tf.float32,\n        initializer= tf.zeros([vocabulary_size], dtype=tf.float32),  trainable=False )\n\n    embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n    embed_reshaped = tf.reshape( embed, [batch_size*num_inputs, embedding_size] )\n\n    segments= np.arange(batch_size).repeat(num_inputs)\n\n    averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n    loss = tf.reduce_mean(\n        tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds, \n                                   sampled_values=tf.nn.uniform_candidate_sampler(true_classes=tf.cast(train_labels, tf.int64), num_sampled=num_sampled, num_true=1, unique=True, range_max=vocabulary_size, seed=None),\n                                   labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)) \n\n    optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) \n\n    saver = tf.train.Saver()\n<\/code><\/pre>\n\n<p>This is where I define the Estimator<\/p>\n\n<pre><code>#Define the estimator\nword2vecEstimator = tf.estimator.Estimator(\n        model_fn=my_model,\n        params={\n            'batch_size': 1024,\n            'embedding_size': 50,\n            'num_inputs': 5,\n            'num_sampled':128\n        })\n<\/code><\/pre>\n\n<p>This is where I tell the Estimator to train<\/p>\n\n<pre><code>#Train with Estimator\nword2vecEstimator.train(\n    input_fn=lambda:iris_data.generate_batch(batch_size, num_inputs),\n    steps=1000)\n<\/code><\/pre>\n\n<p>Does this seem right?<\/p>\n\n<p>Also, after a certain amount of steps, I would like to save the checkpoint and upload it to my google drive. I was wondering how\/where I would place the code to do that. <\/p>\n","answers":[{"AnswerId":"52963207","CreationDate":"2018-10-24T07:28:03.337","ParentId":null,"OwnerUserId":"241013","Title":null,"Body":"<p>You are right, model_fn is the way to go.<\/p>\n\n<p>Params is more about hyperparameters than anything else. Any variable you have you should keep in the model(as it is).<\/p>\n\n<p>You shouldn't have any placeholders in your new function. They come via the features and labels.<\/p>\n\n<p>One thing to watch for is that the same function is callled for predictions, training and evaluation. Use the mode parameter to distinguish. This also means label is None when you do predictions, so make sure the code works correctly in that case.<\/p>\n"}]}
+{"QuestionId":52962019,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-24T06:07:12.723","AcceptedAnswerId":52962099.0,"OwnerUserId":8270077.0,"Title":"TensorFlow exception: module 'tensorflow.tools.api.generator.api.math' has no attribute 'log'","Body":"<p>I run version 1.8 of TensorFlow on Python 3.<\/p>\n\n<p>I am getting the following exception:<\/p>\n\n<pre><code>import tensorflow as tf\n\nMatrix_one = tf.constant([[2,3],[3,4]])\n\nwith tf.Session() as session:\n\n    print(session.run(tf.math.log(Matrix_one)))\n\nAttributeError: module 'tensorflow.tools.api.generator.api.math' has no attribute 'log'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52962099","CreationDate":"2018-10-24T06:15:00.037","ParentId":null,"OwnerUserId":"10535783","Title":null,"Body":"<pre><code>import tensorflow as tf\n\nMatrix_one = tf.constant([[2,3],[3,4]],dtype=tf.float32)\n\nwith tf.Session() as session:\n\n    print(session.run(tf.log(Matrix_one)))\n<\/code><\/pre>\n\n<p>The result is as follows:\n[[0.6931472 1.0986123]\n [1.0986123 1.3862944]]<\/p>\n\n<pre><code>    tf.log(\n        x,\n        name=None\n    )\nThe parameter x must be one of the following types:\n\nbfloat16, half, float32, float64, complex64, complex128\n<\/code><\/pre>\n\n<p>At version 1.11 and above, the tf.math.log function is available because tensorflow moves the base arithmetic operator into tf.math.<\/p>\n"}]}
+{"QuestionId":52962523,"AnswerCount":1,"Tags":"<keras><deep-learning><aws-lambda>","CreationDate":"2018-10-24T06:45:29.020","AcceptedAnswerId":null,"OwnerUserId":3156383.0,"Title":"AWS Lambda is getting timed out while downloading the VGG16 imagenet model data","Body":"<p>We have a lambda function where we will use Keras to extract features from images. So, when the lambda function is trying to download the VGG16 imagenet model, there it is getting timed out. <\/p>\n\n<p>Code where lambda is getting timed out:-<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nmodel = VGG16(weights='imagenet', include_top=True)\n<\/code><\/pre>\n\n<p>How to solve this issue ? Can we login to the backend container of a lambda function and download the model ?<\/p>\n\n<p>Lambda code:-<\/p>\n\n<pre><code>import os\nimport shutil\nimport stat\nimport zipfile\nimport boto3\nfrom six.moves import urllib\n\n\ns3 = boto3.client('s3')\n\ndef download(url, local_fpath):\n    print('downloading other files........')\n    s3.download_file('HARDCODED_BUCKET', url,local_fpath)\n\ndef make_gcc_executable():\n    for fpath in os.listdir(\"\/tmp\/gcc\/bin\"):\n        fpath = os.path.join(\"\/tmp\/gcc\/bin\", fpath)\n        st = os.stat(fpath)\n        os.chmod(fpath, st.st_mode | stat.S_IXOTH | stat.S_IXGRP | stat.S_IXUSR)\n\n    for fpath in os.listdir(\"\/tmp\/gcc\/libexec\/gcc\/x86_64-linux-gnu\/4.6.4\"):\n        fpath = os.path.join(\"\/tmp\/gcc\/libexec\/gcc\/x86_64-linux-gnu\/4.6.4\", fpath)\n        st = os.stat(fpath)\n        os.chmod(fpath, st.st_mode | stat.S_IXOTH | stat.S_IXGRP | stat.S_IXUSR)\n\n# Download GCC and uncompress it.\ndownload('test\/imglib-new\/gcc.zip', \"\/tmp\/gcc.zip\")\nzipfile.ZipFile(\"\/tmp\/gcc.zip\").extractall(\"\/tmp\/gcc\")\n\nmake_gcc_executable()\n\n\n\nfrom keras.preprocessing import image\nfrom keras.applications.vgg16 import VGG16\nfrom keras.applications.vgg16 import preprocess_input\nimport numpy as np\n\ndef _get_model2():\n    print('downloadin our model....')\n    s3.download_file('BUCKET','test\/model\/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5','\/tmp\/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5')\n\n    #model = VGG16(weights='imagenet', include_top=False)\n    print('model download is done')\n    model = VGG16(weights='\/tmp\/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5')\n    return model\n\nmodel = _get_model2()\n\n\n\ndef handler(event, context):\n    print('entering the function..downloading the images')\n    download_loc='\/tmp\/dog2.jpg'\n    s3.download_file('BUCKET', 'test\/images\/dog2.jpg',download_loc)\n    print('image is downloaded')\n\n\n\n    print('running the model now..')\n\n    img = image.load_img(download_loc, target_size=(224, 224))\n    img_data = image.img_to_array(img)\n    img_data = np.expand_dims(img_data, axis=0)\n    img_data = preprocess_input(img_data)\n    print('extracting the feature...')\n    vgg16_feature = model.predict(img_data)\n    print('over..')\n    print(vgg16_feature.shape)\n<\/code><\/pre>\n\n<p>Error message:-<\/p>\n\n<pre><code>START RequestId: 00827d13-d7a4-11e8-9ea2-c5365d4fbfcb Version: $LATEST\nmodule initialization error: Compilation failed (return status=1): \/tmp\/.theano\/compiledir_Linux-4.14-amzn1.x86_64-x86_64-with-glibc2.2.5-x86_64-3.6.1-64\/lazylinker_ext\/mod.cpp:1:20: fatal error: Python.h: No such file or directory. compilation terminated.. \n\nEND RequestId: 00827d13-d7a4-11e8-9ea2-c5365d4fbfcb\nREPORT RequestId: 00827d13-d7a4-11e8-9ea2-c5365d4fbfcb  Duration: 153.15 ms Billed Duration: 200 ms     Memory Size: 832 MB Max Memory Used: 225 MB \nmodule initialization error\nCompilation failed (return status=1): \/tmp\/.theano\/compiledir_Linux-4.14-amzn1.x86_64-x86_64-with-glibc2.2.5-x86_64-3.6.1-64\/lazylinker_ext\/mod.cpp:1:20: fatal error: Python.h: No such file or directory. compilation terminated.. \n<\/code><\/pre>\n","answers":[{"AnswerId":"52964719","CreationDate":"2018-10-24T08:52:23.963","ParentId":null,"OwnerUserId":"2531132","Title":null,"Body":"<p>Lambda does not give you access to log-in to the the container. You haven't mentioned the size of the model, but you have some options:<\/p>\n\n<ul>\n<li>A lambda deployment can be up to 50mb, so you may well be able to upload the model as part of your code package and save yourself downloading it.<\/li>\n<li>If you <em>have<\/em> to download the model, up your timeout to the full 15 minutes. It's unlikely that any file that takes more time than that to download would even fit in a lambda's RAM in any case.<\/li>\n<li>Consider the source of your download. S3 to Lambda downloads at 50mb\/s, so uploading your model to S3 would give you better speed on larger downloads.<\/li>\n<li>Make sure you do the download out of the scope of the handler function- that way you'll only spend the time download the model when your lambda's container loads for the first time.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":52962608,"AnswerCount":1,"Tags":"<keras><crf>","CreationDate":"2018-10-24T06:51:50.623","AcceptedAnswerId":null,"OwnerUserId":10169660.0,"Title":"Why does my keras-contrib's crf.accuracy increase so fast?","Body":"<p>I use BiLSTM+CRF to do a sequence tagging model for postagging. But I find my crf.accuracy increases so fast and reach 1.0 just in the first epoch.\nI use CRF layer from keras_contrib lib, and write code referring to the CRF example. My backend is Tensorflow.\n<code>\n        output = TimeDistributed(Dense(100,activation='relu'),name='Dense_1')(lstm_output)\n        output = Dropout(rate=0.3)(output)\n        if self.use_crf:\n            crf_layer = CRF(self.label_num,sparse_target=True)\n            output = crf_layer(output)\n            model = Model(inputs, output)\n            model.compile('adam', loss=crf_layer.loss_function, metrics=[crf_layer.accuracy])\n        else:\n            output = TimeDistributed(Dense(self.label_num,activation='softmax'), name='Dense_2')(output)\n            model = Model(inputs, output)\n            model.compile('adam', loss='categorical_crossentropy', metrics=['accuracy'])<\/code> <\/p>\n\n<p>but the crf_layer.accuracy seems to increase so fast. In the first epoch, it shows as<\/p>\n\n<p>1888\/2000 [===========================>..] - ETA: 4s - loss: 17.2046 - acc: 0.9666\n 1920\/2000 [===========================>..] - ETA: 3s - loss: 17.1906 - acc: 0.9671\n 1952\/2000 [============================>.] - ETA: 2s - loss: 17.2806 - acc: 0.9677\n 1984\/2000 [============================>.] - ETA: 0s - loss: 17.2222 - acc: 0.9682<\/p>\n\n<p>But the accuracyis is just 0.4 if I don't use CRF, just like the code in else part, and the training looks more normal. <\/p>\n\n<p>So whether I ignore some details in keras_contrib's CRF<\/p>\n","answers":[{"AnswerId":"52963045","CreationDate":"2018-10-24T07:18:50.027","ParentId":null,"OwnerUserId":"10169660","Title":null,"Body":"<p>I find the bug. Because my target is a one-hot vector, I can't set sparse_target=True. After I fix this, it looks normal now.<\/p>\n"}]}
+{"QuestionId":52962853,"AnswerCount":1,"Tags":"<java><android><tensorflow><tensorflow-lite><firebase-mlkit>","CreationDate":"2018-10-24T07:06:54.013","AcceptedAnswerId":52963575.0,"OwnerUserId":10550294.0,"Title":"How do I convert a .png image to a Byte Array - Android, TFLITE application","Body":"<p>I have made a <strong>TFLITE<\/strong> model and am using <strong>Google's ML KIT<\/strong> to access it from my mobile application. I have run into trouble with trying to get my data into the <code>byte[1][299][299][3]<\/code> format that I need for feeding to my classifier. <\/p>\n\n<p>I was trying to arrange a <code>bytestream<\/code> into that format but I don't know if it's feeding <code>left-right<\/code>, <code>top-bottom<\/code>, R-G-B, etc.<\/p>\n\n<p>Can anyone point me to some documentation I can read about parsing <code>.png<\/code> files?<\/p>\n","answers":[{"AnswerId":"52963575","CreationDate":"2018-10-24T07:51:43.153","ParentId":null,"OwnerUserId":"10295689","Title":null,"Body":"<p>You can convert your .png file to an Android <code>Bitmap<\/code> using a <code>BitmapFactory<\/code> and an <code>InputStream<\/code>.<\/p>\n\n<pre><code>ByteArrayOutputStream stream = new ByteArrayOutputStream();\n\nBitmap bitmap = BitmapFactory.decodeFile(filePath);\nbitmap.compress(Bitmap.CompressFormat.PNG, 100, stream);\nbyte[] image = stream.toByteArray();\n<\/code><\/pre>\n\n<p>With your Bitmap, you can also get the color of a pixel at a specified position. Unfortunately, the color is returned as <code>int<\/code>, which you have to decompose to your RGBA values. For that, see: <a href=\"https:\/\/developer.android.com\/reference\/android\/graphics\/Color\" rel=\"nofollow noreferrer\">https:\/\/developer.android.com\/reference\/android\/graphics\/Color<\/a><\/p>\n"}]}
+{"QuestionId":52963909,"AnswerCount":1,"Tags":"<machine-learning><conv-neural-network><pytorch>","CreationDate":"2018-10-24T08:08:55.863","AcceptedAnswerId":52981440.0,"OwnerUserId":4777141.0,"Title":"Pytorch: Getting the correct dimensions for final layer","Body":"<p>Pytorch newbie here! I am trying to fine-tune a VGG16 model to predict 3 different classes. Part of my work involves converting FC layers to CONV layers. However, the values of my predictions don't fall between 0 to 2 (the 3 classes). <\/p>\n\n<p>Can someone point me to a good resource on how to compute the correct dimensions for the final layer? <\/p>\n\n<p>Here are the original fC layers of VGG16:<\/p>\n\n<pre><code>(classifier): Sequential(\n    (0): Linear(in_features=25088, out_features=4096, bias=True)\n    (1): ReLU(inplace)\n    (2): Dropout(p=0.5)\n    (3): Linear(in_features=4096, out_features=4096, bias=True)\n    (4): ReLU(inplace)\n    (5): Dropout(p=0.5)\n    (6): Linear(in_features=4096, out_features=1000, bias=True)\n  )\n<\/code><\/pre>\n\n<p>My code for converting FC layers to CONV:<\/p>\n\n<pre><code> def convert_fc_to_conv(self, fc_layers):\n        # Replace first FC layer with CONV layer\n        fc = fc_layers[0].state_dict()\n        in_ch = 512\n        out_ch = fc[\"weight\"].size(0)\n        first_conv = nn.Conv2d(512, out_ch, kernel_size=(1, 1), stride=(1, 1))\n\n        conv_list = [first_conv]\n        for idx, layer in enumerate(fc_layers[1:]):\n            if isinstance(layer, nn.Linear):\n                fc = layer.state_dict()\n                in_ch = fc[\"weight\"].size(1)\n                out_ch = fc[\"weight\"].size(0)\n                if idx == len(fc_layers)-4:\n                    in_ch = 3\n                conv = nn.Conv2d(out_ch, in_ch, kernel_size=(1, 1), stride=(1, 1))\n                conv_list += [conv]\n            else:\n                conv_list += [layer]\n            gc.collect()\n\n        avg_pool = nn.AvgPool2d(kernel_size=2, stride=1, ceil_mode=False)\n        conv_list += [avg_pool, nn.Softmax()]\n        top_layers = nn.Sequential(*conv_list)  \n        return top_layers\n<\/code><\/pre>\n\n<p>Final model architecture:<\/p>\n\n<pre><code>    Model(\n    (features): Sequential(\n    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (1): ReLU(inplace)\n    (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (3): ReLU(inplace)\n    (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (6): ReLU(inplace)\n    (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (8): ReLU(inplace)\n    (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (11): ReLU(inplace)\n    (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (13): ReLU(inplace)\n    (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (15): ReLU(inplace)\n    (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (18): ReLU(inplace)\n    (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (20): ReLU(inplace)\n    (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (22): ReLU(inplace)\n    (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (25): ReLU(inplace)\n    (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (27): ReLU(inplace)\n    (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (29): ReLU(inplace)\n    (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))\n\n    (classifier): Sequential(\n    (0): Conv2d(512, 4096, kernel_size=(1, 1), stride=(1, 1))\n    (1): ReLU(inplace)\n    (2): Dropout(p=0.5)\n    (3): Conv2d(4096, 3, kernel_size=(1, 1), stride=(1, 1))\n    (4): ReLU(inplace)\n    (5): Dropout(p=0.5)\n    (6): AvgPool2d(kernel_size=2, stride=1, padding=0)\n    (7): Softmax()\n  )\n)\n<\/code><\/pre>\n\n<p>Summary of the model: <\/p>\n\n<pre><code>            Conv2d-1         [-1, 64, 224, 224]           1,792\n              ReLU-2         [-1, 64, 224, 224]               0\n            Conv2d-3         [-1, 64, 224, 224]          36,928\n              ReLU-4         [-1, 64, 224, 224]               0\n         MaxPool2d-5         [-1, 64, 112, 112]               0\n            Conv2d-6        [-1, 128, 112, 112]          73,856\n              ReLU-7        [-1, 128, 112, 112]               0\n            Conv2d-8        [-1, 128, 112, 112]         147,584\n              ReLU-9        [-1, 128, 112, 112]               0\n        MaxPool2d-10          [-1, 128, 56, 56]               0\n           Conv2d-11          [-1, 256, 56, 56]         295,168\n             ReLU-12          [-1, 256, 56, 56]               0\n           Conv2d-13          [-1, 256, 56, 56]         590,080\n             ReLU-14          [-1, 256, 56, 56]               0\n           Conv2d-15          [-1, 256, 56, 56]         590,080\n             ReLU-16          [-1, 256, 56, 56]               0\n        MaxPool2d-17          [-1, 256, 28, 28]               0\n           Conv2d-18          [-1, 512, 28, 28]       1,180,160\n             ReLU-19          [-1, 512, 28, 28]               0\n           Conv2d-20          [-1, 512, 28, 28]       2,359,808\n             ReLU-21          [-1, 512, 28, 28]               0\n           Conv2d-22          [-1, 512, 28, 28]       2,359,808\n             ReLU-23          [-1, 512, 28, 28]               0\n        MaxPool2d-24          [-1, 512, 14, 14]               0\n           Conv2d-25          [-1, 512, 14, 14]       2,359,808\n             ReLU-26          [-1, 512, 14, 14]               0\n           Conv2d-27          [-1, 512, 14, 14]       2,359,808\n             ReLU-28          [-1, 512, 14, 14]               0\n           Conv2d-29          [-1, 512, 14, 14]       2,359,808\n             ReLU-30          [-1, 512, 14, 14]               0\n        MaxPool2d-31            [-1, 512, 7, 7]               0\n           Conv2d-32           [-1, 4096, 7, 7]       2,101,248\n             ReLU-33           [-1, 4096, 7, 7]               0\n          Dropout-34           [-1, 4096, 7, 7]               0\n           Conv2d-35              [-1, 3, 7, 7]          12,291\n             ReLU-36              [-1, 3, 7, 7]               0\n          Dropout-37              [-1, 3, 7, 7]               0\n        AvgPool2d-38              [-1, 3, 6, 6]               0\n          Softmax-39              [-1, 3, 6, 6]               0\n<\/code><\/pre>\n","answers":[{"AnswerId":"52981440","CreationDate":"2018-10-25T04:37:33.237","ParentId":null,"OwnerUserId":"5373738","Title":null,"Body":"<p>I wrote a function that takes a Pytorch model as input and converts the classification layer to convolution layer. It works for VGG and Alexnet for now, but you can extend it for other models as well.<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nfrom torchvision.models import alexnet, vgg16\n\ndef convolutionize(model, num_classes, input_size=(3, 224, 224)):\n    '''Converts the classification layers of VGG &amp; Alexnet to convolutions\n\n    Input:\n        model: torch.models\n        num_classes: number of output classes\n        input_size: size of input tensor to the model\n\n    Returns:\n        model: converted model with convolutions\n    '''\n    features = model.features\n    classifier = model.classifier\n\n    # create a dummy input tensor and add a dim for batch-size\n    x = torch.zeros(input_size).unsqueeze_(dim=0)\n\n    # change the last layer output to the num_classes\n    classifier[-1] = nn.Linear(in_features=classifier[-1].in_features,\n                               out_features=num_classes)\n\n    # pass the dummy input tensor through the features layer to compute the output size\n    for layer in features:\n        x = layer(x)\n\n    conv_classifier = []\n    for layer in classifier:\n        if isinstance(layer, nn.Linear):\n            # create a convolution equivalent of linear layer\n            conv_layer = nn.Conv2d(in_channels=x.size(1),\n                                   out_channels=layer.weight.size(0),\n                                   kernel_size=(x.size(2), x.size(3)))\n\n            # transfer the weights\n            conv_layer.weight.data.view(-1).copy_(layer.weight.data.view(-1))\n            conv_layer.bias.data.view(-1).copy_(layer.bias.data.view(-1))\n            layer = conv_layer\n\n        x = layer(x)\n        conv_classifier.append(layer)\n\n    # replace the model.classifier with newly created convolution layers\n    model.classifier = nn.Sequential(*conv_classifier)\n\n    return model\n\ndef visualize(model, input_size=(3, 224, 224)):\n    '''Visualize the input size though the layers of the model'''\n    x = torch.zeros(input_size).unsqueeze_(dim=0)\n    print(x.size())\n    for layer in list(model.features) + list(model.classifier):\n        x = layer(x)\n        print(x.size())\n<\/code><\/pre>\n\n<p>This is how the input looks when passed through the model<\/p>\n\n<pre><code>_vgg = vgg16()\nvgg = convolutionize(_vgg, 100)\nprint('\\n\\nVGG')\nvisualize(vgg)\n<\/code><\/pre>\n\n<p>...<\/p>\n\n<pre><code>VGG\ntorch.Size([1, 3, 224, 224])\ntorch.Size([1, 64, 224, 224])\ntorch.Size([1, 64, 224, 224])\ntorch.Size([1, 64, 224, 224])\ntorch.Size([1, 64, 224, 224])\ntorch.Size([1, 64, 112, 112])\ntorch.Size([1, 128, 112, 112])\ntorch.Size([1, 128, 112, 112])\ntorch.Size([1, 128, 112, 112])\ntorch.Size([1, 128, 112, 112])\ntorch.Size([1, 128, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 56, 56])\ntorch.Size([1, 256, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 28, 28])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 14, 14])\ntorch.Size([1, 512, 7, 7])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 4096, 1, 1])\ntorch.Size([1, 100, 1, 1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":52964192,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-24T08:24:54.187","AcceptedAnswerId":null,"OwnerUserId":10550643.0,"Title":"Implementing a generator for keras results in worse results","Body":"<p>I am doing multi-label image classification on a dataset of around 3000 images. Because this is the limit of my working memory and the dataset will increase, I was trying to implement my own generator because I am also parsing the images from an online source. The network reached an accuracy of 25% where the three labels with the highest accuracy gave a pretty good representation of the images.<\/p>\n\n<p>A normal batch would be of shape (32, 64, 64, 3) with labels of shape (32, 57).\nMy model looks like:<\/p>\n\n<pre><code>def createModel(shape, classes):\n    x = shape[0]\n    y = shape[1]\n    z = shape[2]\n    model = Sequential()\n    model.add(Conv2D(32, (2, 2), padding='same',input_shape=(x,y,z)))\n    model.add(BatchNormalization())\n    model.add(Activation('relu'))\n    model.add(Conv2D(32, (2, 1), padding='same'))\n    model.add(BatchNormalization())\n    model.add(Activation('relu'))\n    model.add(Conv2D(32, (1, 2), padding='same'))\n    model.add(BatchNormalization())\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides = 2))\n\n    model.add(Conv2D(48, (2, 2), padding='same'))\n    model.add(BatchNormalization())\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides = 2))\n\n    model.add(Conv2D(80, (2, 2), padding='same'))\n    model.add(BatchNormalization())\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides = 2))\n\n    model.add(Flatten())\n    model.add(Dense(classes, activation = 'sigmoid'))\n<\/code><\/pre>\n\n<p>where classes would be 57 and the x,y,z would be 64,64,3.<\/p>\n\n<p>My generator looks like:<\/p>\n\n<pre><code>def generator(data, urls, labels, batch_size):\n    counter = 0\n    X_train = []\n    Y_train = []\n\n    while 1:\n        for i in range(len(data)):\n            if counter == batch_size:\n                yield (np.array(X_train), np.array(Y_train))\n                X_train = []\n                Y_train = []\n                counter = 0\n            try:\n                ID = data[i][0]\n                if random.uniform(0,1) &gt; 0.5:\n                    X_train.append(getImage(64, urls[ID]))\n                else:\n                    X_train.append(np.flip(getImage(64, urls[ID]),1))\n                Y_train.append(labels[i])\n                counter+=1\n            except:\n                continue\n<\/code><\/pre>\n\n<p>where data is an list with image ID and labels, urls is a list with image ID and the url to find the image, labels is the labels converted by MultiLabelBinarizer() (.fit_transform) and the batch_size. the getImage() function results in a np.array() where 64 gives the shape.<\/p>\n\n<p>The main calls:<\/p>\n\n<pre><code>epochs = 60\nlr = 1e-6\n\nmlc = model.createModel((64,64,3), 57)\n\nopt = Adam(lr=lr, decay=lr \/ epochs)\n\ntrainGenerator = data.generator(structuredData, urls, mlb_labels, 32)\nvalidationGenerator = data.generator(structuredData, urls, mlb_labels, 32)\n\nmlc.compile(loss=\"categorical_crossentropy\", optimizer=opt, metrics= \n[\"accuracy\"])\n\nmlc.fit_generator(trainGenerator, steps_per_epoch = 10, epochs=epochs, \nvalidation_steps = 1, validation_data=validationGenerator)\n\nmlc.save(\"datagenerator_test.h5\")\n<\/code><\/pre>\n\n<p>Furthermore the network thus already works and trains if I do not use the generator, with the generator it seems to get a random accuracy between 1 and 3%. I hope this provides enough information.<\/p>\n\n<p>EDIT: I takes about 90 seconds to prepare one batch of 32 images. Does the training wait for a batch to be ready?<\/p>\n","answers":[]}
+{"QuestionId":52964273,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><artificial-intelligence>","CreationDate":"2018-10-24T08:29:05.857","AcceptedAnswerId":52983190.0,"OwnerUserId":10550693.0,"Title":"Classifier to Predict activities taking place","Body":"<p>I have multiple datasets here that i took from Kaggle. There are multiple csv files and each csv file is made specifically for sit, stand, walking, running etc. The data is taken from sensors like accelerometers and gyroscopes. The values in datasets are of axis like x, y and z. <\/p>\n\n<p>Sample Data<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GCq2S.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GCq2S.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here is a sample dataset of jogging. Now i need to make classifiers in my program so that my program can detect itself whether the data is of jogging, sitting, standing etc. I want to mix all the datasets in a single csv file and then upload it into my webapge and then i want the javascript code to start detecting whether a particular row is of sitting, standing, jogging etc. I don't want any code help but instead i just need a little explanation or a way to start coding it. How can i started making such classifier? I know it is kind of broad question but i think i have tried to explain myself in best way possible. Once my program has detected every row with specific activity it will count all the activities separately and then show it in a table format in webpage.<\/p>\n","answers":[{"AnswerId":"52983190","CreationDate":"2018-10-25T06:59:16.423","ParentId":null,"OwnerUserId":"9461340","Title":null,"Body":"<p>In order to answer properly to your question, it would be very helpful to know which is your level of understanding and experience with  machine learning. <\/p>\n\n<p>If you are a beginner I would suggest to try to run and understand a couple of tutorials that can be easily found on the web. <\/p>\n\n<p>If you need an idea of which is the \"standard\" approach for machine learning development, I will try to give you a general idea of the process.<\/p>\n\n<p>You can summarize the process in these main steps:<\/p>\n\n<p>Data pre-processing-> Data splitting -> Feature selection -> Model Training -> Validation -> Deployment <\/p>\n\n<p>Data pre-processing is meant to clean and format the data: removing NA values, decision about categorical variables, outlier analysis,.... This is a complex step that depends on the application. In your case I would start checking that the data in the different data-sets are homogeneous, i.e. the features have the same meaning across csv and corresponding features respect the same distribution. While the meaning of each feature should be explained in the description of your csv, the check of the distributions could be easily done plotting the box-plots for each feature and csv. If distributions of the same feature across different csv files don't overlap you should investigate further the issue.   <\/p>\n\n<p>An important step in the design of a good model is the splitting of the data. You should split your data in training\/validation set (training\/validation\/test for a more comprehensive approach). This step allows you to train your model on the training set and test the model on the validation set computing unbiased performance of your model. I suggest here to become familiar with concepts as: Cross Validation, stratified-cross-validation, nested-cross-validation for hyper-parameter tuning, overfitting, bias,.... The Validation of the model will give you an idea of the expected performance that it will have on unseen data. If you are considering the use of more than one model, you can use the validation results to choose the \"best\" one. I suggest here a comparison using the confidence interval or if possible a significance test (e.g t-test, anova,...). Before the deployment the model is trained on all the available data.<\/p>\n\n<p>The choice of the model depends on the data that you are using: number of samples, number of features, type of variable (numerical, categorical),.... <\/p>\n\n<p>I'm not an expert of javascript, but I believe (just a feeling) that python and R are more common choices for developing Machine learning applications. Both have libraries specifically developed for the task and you can find a lot of materials and tutorial around. <\/p>\n\n<p>With a bit of more context I think that I could be more specific.<\/p>\n\n<p>I hope it helps<\/p>\n"}]}
+{"QuestionId":52964395,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><computer-vision><conv-neural-network><pytorch>","CreationDate":"2018-10-24T08:34:55.600","AcceptedAnswerId":52964887.0,"OwnerUserId":7305303.0,"Title":"What are the default values of kernel size and padding in pytorch?","Body":"<p>For an instance, take this piece of code :<\/p>\n\n<pre><code>conv = conv2d(in_channels = 3, out_channels = 64)\n<\/code><\/pre>\n\n<p>What can I expect the padding and kernel size to be, by default?<\/p>\n","answers":[{"AnswerId":"52964887","CreationDate":"2018-10-24T08:59:32.950","ParentId":null,"OwnerUserId":"2846062","Title":null,"Body":"<blockquote>\n  <p>class torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)<\/p>\n<\/blockquote>\n\n<p>So to answer your question, by default <code>padding=0<\/code>, there is no default value for <code>kernel_size<\/code>. If you leave it empty, you will get an error<\/p>\n"}]}
+{"QuestionId":52964532,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><forecasting>","CreationDate":"2018-10-24T08:42:40.613","AcceptedAnswerId":null,"OwnerUserId":10245126.0,"Title":"Financial Time Series Forecasting with Keras\/Tensorflow: Three forecasting methods tried, three poor results had, what am I doing wrong?","Body":"<p>I'm working with a dataset of a bunch of different features for stock forecasting for a project to help teach myself programming and to get into machine learning. I am running into some issues with predicting future data (or time series forecasting) and I was hoping someone out there could give  me some advice! Any advice or criticism you could provide will be greatly appreciated.<\/p>\n\n<p>Below I've listed detailed examples of the three implementations I have tried for forecasting time series data. I could definitely be wrong on this but I don't believe this is mechanical code issue because all of the results are consistent despite me re-coding it a few times (the only thing I can really think of here is not using MinMaxScaler correctly. See closing thoughts). It could definitely, however, be a macrocoding mistake. <\/p>\n\n<p>I didn't post any code for the project here because it was starting to turn into a wall of words and I had three separate examples, but if you have any questions or think it would benefit your assistance to see the code used for any of the below examples or the data used for all of them feel free to let me know and I'll link whatever's needed.<\/p>\n\n<p>The three forecasting implementations I have tried:<\/p>\n\n<p><strong>1)<\/strong> - A sliding window implementation. Input data is shifted backwards in timesteps (x-1, x-2...), target data is current timestep (x). Data used for forecasting the first prediction is n-rows of test data shifted in same manner as input data. For every subsequent prediction the oldest timestep is removed and the prediction is appended to front of prediction row, maintaining the same number of total timesteps but progressing forward in time.<\/p>\n\n<p><strong>2)<\/strong> - Input data is just x, target data is shifted 30 timesteps forward for prediction (y+1, y+2 ... y+30). Attempting to forecast future by taking the first sample of x in test data and predicting 30 steps into the future with it.<\/p>\n\n<p><strong>3)<\/strong> - Combination of both methods, input data is shifted backward and in the example shown below, 101 timesteps including the present timestep (x-100, x-99 ... x) were used. Target data, similar to implementation 2, is shifted 30 timesteps into the future (y+1, y+2... y+30). With this, I am attempting to forecast the future by taking 101 timesteps of the first n-rows of test data and predicting 30 steps into the future from there.<\/p>\n\n<p>For all tests, I cut off the end of my dataset at an arbitrary amount (last ~10% of total dataset), split everything before the cutoff into training\/validation (80\/20) and saved everything after the cutoff for testing &amp; forecasting purposes. <\/p>\n\n<p>As for network architectures, I've tried a bunch of different ones, from bidirectional LSTM to multi-input CNN \/ GRU, to a wavenet like CNN implementation and all produce prediction results that are bad in a similar enough way that I feel like this is either a data manipulation problem or a problem of me not understanding how model.predict(), or my model's output works.<\/p>\n\n<p>The architectures I will be using for each implementation below are: <\/p>\n\n<p><strong>1)<\/strong> causal dilation CNN<\/p>\n\n<p><strong>2)<\/strong> two layers LSTM<\/p>\n\n<p>neural network architecture diagrams here: <a href=\"https:\/\/imgur.com\/a\/cY2RWNG\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/cY2RWNG<\/a><\/p>\n\n<p>For every example below the model's weights were tuned by the model training on the training data (first 80% of dataset) and attempting to achieve the lowest validation loss possible using the validation data (last 20% of dataset).<\/p>\n\n<p><strong>--- First Implementation ---<\/strong><\/p>\n\n<p>(unfortunately, there's an image limit on stack overflow for my current rating or whatever so I've put each implementation into its own album)<\/p>\n\n<p>Implementation 1 - Graphs for both CNN\/LSTM: Images 1-7 <a href=\"https:\/\/imgur.com\/a\/36DZCIf\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/36DZCIf<\/a><\/p>\n\n<p>In every training\/validation test graph, black represents the actual data and red is the predicted data, in the forecasting predictions blue represents the predictions made and orange represents the actual close price on a longer time scale than the prediction for better scale, all forecast predictions are 30 days into the future.<\/p>\n\n<p>Using this methodology, and displaying the actual close price against the predicted close price in every instance:<\/p>\n\n<p><strong>Image 1<\/strong> - sliding window set up for this implementation using one and two features and a range of numbers for viewing ease<\/p>\n\n<p>CNN:<\/p>\n\n<p>(images 2 &amp; 3 description in album)<\/p>\n\n<p><strong>Image 4<\/strong> - Sliding window approach of forecasting every feature in the data, with the prediction for close price plotted against the actual close price. The timesteps start at the first row of the cutoff data.<\/p>\n\n<p>When the first prediction is made I append the prediction to the end of this row and remove the first timestep, repeating for every future timestep I wish to predict.<\/p>\n\n<p>I really don't even know what to say about this prediction, it's really bad...<\/p>\n\n<p>LSTM:<\/p>\n\n<p>(images 5 &amp; 6 description in album)<\/p>\n\n<p><strong>Image 7<\/strong> - Sliding window prediction: <a href=\"https:\/\/i.imgur.com\/Ywf6xvr.png\" rel=\"nofollow noreferrer\">https:\/\/i.imgur.com\/Ywf6xvr.png<\/a><\/p>\n\n<p>This prediction seems to be getting the trend somewhat accurately I guess.. But the starting point is really nowhere near the last known data point which is confusing.<\/p>\n\n<p><strong>--- Second Implementation ---<\/strong><\/p>\n\n<p>Implementation 2 - Graphs for both CNN\/LSTM: Images 1-7\n<a href=\"https:\/\/imgur.com\/a\/3CAk1xc\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/3CAk1xc<\/a><\/p>\n\n<p>For this attempt, I made the target prediction many timesteps into the future. With this implementation, the model takes in the current timestep(x) of features and attempts to predict the closing price at y+1, y+2,y+3 etc. There is only one prediction here -- a sequence of time steps into the future.<\/p>\n\n<p>The same graphing and network conventions as implementation 1 had applied to this too.<\/p>\n\n<p><strong>Image 1<\/strong> - Set up of input and target data, using a range and only one or two features for viewing ease.<\/p>\n\n<p>CNN:<\/p>\n\n<p>(images 2 &amp; 3 description in album)<\/p>\n\n<p><strong>Image 4<\/strong> - Plotting all 30 predictions made from the first row of data features after the cutoff... this is horrible, why is the start again nowhere near the last known data point? I don't understand how it can predict y+1 being so far away from the closing price of x when in every instance of its training y+1 was almost certainly extremely close to x.<\/p>\n\n<p>LSTM:<\/p>\n\n<p>(images 5 &amp; 6 description in album)<\/p>\n\n<p><strong>Image 7<\/strong> - All 30 predictions into the future made from the first row of cutoff data: Again, both all over the place and the predictions start nowhere near the last actual data point, not sure what else to add.<\/p>\n\n<p>It's starting to appear that either my CNN implementation is poorly done or LSTM is just a better solution here. Regardless, the predictions and actual forecasting are still terrible so I'll hold my judgment on the network architecture until I get something that looks remotely like an actual forecast prediction.<\/p>\n\n<p><strong>--- Third Implementation ---<\/strong><\/p>\n\n<p>Implementation 3 - Graphs for both CNN\/LSTM: Images 1-7\n<a href=\"https:\/\/imgur.com\/a\/clcKFF8\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/clcKFF8<\/a><\/p>\n\n<p>This was the final idea I had for forecasting the future and it's essentially a combination of the first two. For this implementation, I take x-n (x, x-1, x-2, x-3) etc., which is similar to the first implementation, and set the target data to y+1, y+2, y+3, which is similar to the second implementation. My goal for predicting with this was the same strategy as the second implementation where I would predict 30 days into the future, but instead of doing so on one timestep of features, I'd do so on many timesteps into the past. I had hoped that this implementation would give the prediction enough supporting data to accurately forecast the future<\/p>\n\n<p><strong>Image 1<\/strong> - Input data or \"x\" and Target data or \"y\" implementation and set up. I use a range of numbers again. In this example, the input data has 2 features, includes the present timestep (x) and 4 timesteps shifted backward (x-1, x-2, x-3, x-4) and the target data has 5 timesteps into the future (y+1, y+2, y+3, y+4, y+5)<\/p>\n\n<p>CNN:<\/p>\n\n<p>(images 2 &amp; 3 description in album)<\/p>\n\n<p><strong>Image 4<\/strong> - 30 predictions into the future using 101 timesteps of x<\/p>\n\n<p>This is probably the worst result yet and that's despite the prediction having way more timesteps back of data to use.<\/p>\n\n<p>LSTM:<\/p>\n\n<p>(images 5 &amp; 6 description in album)<\/p>\n\n<p><strong>Image 7<\/strong> - 30 predictions on input data row of 101 timesteps.<\/p>\n\n<p>This actually has some action to it I guess, but it's all over the place, doesn't start near the last actual data point and is clearly not accurate at all.<\/p>\n\n<p><strong>closing thoughts<\/strong><\/p>\n\n<p>I've also tried removing the target variable (close price) from the input data but it doesn't seem to change much and the past n-days of closing data should be available to a model anyway I would think.<\/p>\n\n<p>Originally I MinMaxScaled all of my data in my pre-processing page and did not inverse_transform any of the data. The results were basically just as bad as the examples above. For the examples above I have min max scaled the prediction, validation &amp; test datasets separately to be within the range of 0.2 - 0.8. For the actual forecasting predictions, I've inverse_transformed the data before plotting it against the actual closing price which was never transformed.<\/p>\n\n<p>If I am doing something fundamentally wrong in the above examples I would love to know as I'm just starting out and this is my first real programming\/machine learning project.<\/p>\n\n<p>A few other things relating to this that I've come across \/ tried:<\/p>\n\n<p>I've experimented briefly with using a stateful model where I reset_states() after every prediction to some moderate success.<\/p>\n\n<p>I've read that sequence to sequence models could be useful for forecasting time series data but I'm really not sure what that system is designed to do with time series despite reading into it quite a bit and am thus not sure how to implement it or test it out.<\/p>\n\n<p>I tried bidirectional LSTM because of one random StackOverflow post suggesting it for time series forecasting... the results were very mediocre however and it doesn't seem to make much sense to me in this situation from what I understand of how it works. I've only tried it with the first implementation above though, let me know if it's something to look more into.<\/p>\n\n<p>Any tips\/criticism at all that you could provide would be greatly appreciated, I'm really not sure how to progress from here. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"56324098","CreationDate":"2019-05-27T10:18:26.080","ParentId":null,"OwnerUserId":"11134789","Title":null,"Body":"<p>I have been through that,for me the sliding window approach with LSTM, NN worked like magic for small time series But on a bigger time series with data coming in on hourly basis for a few years it failed miserably. <\/p>\n\n<p>Later on I ditched LSTM,GBTs &amp; started implementing algos from statsmodels.tsa, ARIMA, SARIMA most of the time, I'll suggest you to read about them too. Very easy to implement no need to worry about sliding window, moving data few timestamps back, it takes care of all. Just train, tune the parameters &amp; predict for the next timestamps.<\/p>\n\n<p>Sometimes, I also faced issues like my time series had missing timestamps &amp; data, then I had to impute those values, the frequency on which I trained (hourly,weekly,monthly) was different from the frequency on which I wanted to predict, then I had to bring data in right form too. I faced different frequency issue while visualising on a plot as well.<\/p>\n\n<pre><code>model=statsmodels.api.tsa.SARIMAX(train_df,order=(1,0,1),seasonal_order=(1,1,0,24))\nmodel = model.fit()\n<\/code><\/pre>\n\n<p>other than the data pre-processing part, imputing missing data, training on right frequency &amp; some logic for parameters tuning, you will need to use just these two lines, your data_frame will have the index in a date format &amp; columns will have time series data<\/p>\n"}]}
+{"QuestionId":52964807,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-10-24T08:56:28.910","AcceptedAnswerId":52968308.0,"OwnerUserId":5468446.0,"Title":"Padding in Pytorch","Body":"<p>In PyTorch tensor, I would like to obtain the output from the input as follows:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/yiuho.png\" alt=\"enter image description here\"><\/p>\n\n<p>How can I achieve this padding in Pytroch?<\/p>\n","answers":[{"AnswerId":"52968308","CreationDate":"2018-10-24T11:51:56.873","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>One way of doing this is<\/p>\n\n<pre><code>def my_odd_padding(list_of_2d_tensors, pad_value):\n  # get the sizes of the matrices\n  hs = [t_.shape[0] for t_ in list_of_2d_tensors]\n  ws = [t_.shape[1] for t_ in list_of_2d_tensors]\n  # allocate space for output\n  result = torch.zeros(sum(hs), sum(ws))\n  result.add_(pad_value)\n  fh = 0\n  fw = 0\n  for i, t_ in enumerate(list_of_2d_tensors):\n    result[fh:fh+hs[i], fw:fw+ws[i]] = t_\n    fh += hs[i]\n    fw += ws[i]\n  return result \n<\/code><\/pre>\n\n<p>assuming all tensors on <code>list_of_2d_tensors<\/code> are of the same <code>dtype<\/code> and on the same <code>device<\/code> you may explicitly set this dtype and device to <code>result<\/code> when you allocate it using <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.zeros\" rel=\"nofollow noreferrer\"><code>torch.zeros<\/code><\/a><\/p>\n"}]}
+{"QuestionId":52965004,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-24T09:04:59.907","AcceptedAnswerId":52985372.0,"OwnerUserId":10550766.0,"Title":"Tensorflow: Batch TFRecord Dataset with tensors of arbitrary dimensions","Body":"<h2>How can I batch tensors of arbitrary shapes using the TFRecordsDataset?<\/h2>\n\n<p>I am currently working on the input pipeline of an object detection Network and am struggeling with the batching of my labels. The labels consist of Bounding Box coordinates and classes of the Objects in an image. Since there may be multiple objects in an image the label <strong>dimensions are arbitrary<\/strong><\/p>\n\n<hr>\n\n<p>When working with <code>tf.train.batch<\/code> there is the possibility to set <code>dynamic_padding=True<\/code> to fit the shapes to the same dimensions. However there is no such option in the <code>data.TFRecordDataset.batch()<\/code>. <\/p>\n\n<p>The desired shape i would like to batch would be <code>[batch_size, arbitrary , 4]<\/code> for my Boxes and <code>[batch_size, arbitrary, 1]<\/code> for the classes.<\/p>\n\n<pre><code>def decode(serialized_example):\n\"\"\"\nDecodes the information of the TFRecords to image, label_coord, label_classes\nLater on will also contain the Image Sequence!\n\n:param serialized_example: Serialized Example read from the TFRecords\n:return: image, label_coordinates list, label_classes list\n\"\"\"\nfeatures = {'image\/shape': tf.FixedLenFeature([], tf.string),\n            'train\/image': tf.FixedLenFeature([], tf.string),\n            'label\/coordinates': tf.VarLenFeature(tf.float32),\n            'label\/classes': tf.VarLenFeature(tf.string)}\n\nfeatures = tf.parse_single_example(serialized_example, features=features)\n\nimage_shape = tf.decode_raw(features['image\/shape'], tf.int64)\nimage = tf.decode_raw(features['train\/image'], tf.float32)\nimage = tf.reshape(image, image_shape)\n\n# Contains the Bounding Box coordinates in a flattened tensor\nlabel_coord = features['label\/coordinates']\nlabel_coord = label_coord.values\nlabel_coord = tf.reshape(label_coord, [1, -1, 4])\n\n# Contains the Classes of the BBox in a flattened Tensor\nlabel_classes = features['label\/classes']\nlabel_classes = label_classes.values\nlabel_classes = tf.reshape(label_classes, [1, -1, 1])\n\n\nreturn image, label_coord, label_classes\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>    dataset = tf.data.TFRecordDataset(filename)\n    dataset = dataset.map(decode)\n    dataset = dataset.map(augment)\n    dataset = dataset.map(normalize)\n    dataset = dataset.repeat(num_epochs)\n    dataset = dataset.batch(batch_size)\n\n    dataset = dataset.batch(batch_size)\n<\/code><\/pre>\n\n<hr>\n\n<p>The error that is thrown is <code>Cannot batch tensors with different shapes in component 1. First element had shape [1,1,4] and element 1 had shape [1,7,4].<\/code><\/p>\n\n<p>Also currently the <code>augment<\/code> and <code>normalize<\/code> functions are just placeholders.<\/p>\n","answers":[{"AnswerId":"52985372","CreationDate":"2018-10-25T09:01:18.363","ParentId":null,"OwnerUserId":"10550766","Title":null,"Body":"<p>It turns out <code>tf.data.TFRecordDataset<\/code> has an other function called <code>padded_batch<\/code> which is basically doing the thing <code>tf.train.batch(dynamic_pad=True)<\/code> does. This solves the proble rather easily...<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset(filename)\n\ndataset = dataset.map(decode)\ndataset = dataset.map(augment)\ndataset = dataset.map(normalize)\n\ndataset = dataset.shuffle(1000+3*batch_size)\ndataset = dataset.repeat(num_epochs)\ndataset = dataset.padded_batch(batch_size,\n                               drop_remainder=False,\n                               padded_shapes=([None, None, None],\n                                              [None, 4],\n                                              [None, 1])\n                              )\n<\/code><\/pre>\n"}]}
+{"QuestionId":52965101,"AnswerCount":1,"Tags":"<javascript><tensorflow><neural-network><xor><tensorflow.js>","CreationDate":"2018-10-24T09:09:41.663","AcceptedAnswerId":52967162.0,"OwnerUserId":6212453.0,"Title":"XOR Neural network returning rubish","Body":"<p>my XOR \"neural network\" in Tensorflow.js keeps returning <strong>rubbish prediction<\/strong> and the <strong>loss<\/strong> is always stuck at <strong>0.25<\/strong>. I have no idea what did I do wrong. Thank you for your help!<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential();\r\n\r\nmodel.add(tf.layers.dense({units: 2, activation: 'sigmoid', inputShape: [2]}));\r\nmodel.add(tf.layers.dense({units: 1, activation: 'sigmoid'}));\r\n\r\nmodel.compile({loss:'meanSquaredError', optimizer: 'sgd'});\r\ntf.train.sgd(0.5);\r\n\r\nconst xs = tf.tensor2d([[0,0],[0,1],[1,0],[1,1]]);\r\nconst ys = tf.tensor2d([[0],[1],[1],[0]]);\r\n\r\nasync function train() {\r\n    for(let i = 0; i &lt; 200; i++){\r\n        const history = await model.fit(xs, ys, {epochs: 20, shuffle: true});\r\n        console.log(\"loss: \" + history.history.loss[19] + \" on \" + i + \". iteration.\");\r\n    }\r\n}\r\n\r\ntrain().then(() =&gt; {\r\n    console.log(\"trained with \" + tf.memory().numTensors + \"tensors\");\r\n    model.predict(xs).print();\r\n});<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;script src=\"https:\/\/cdnjs.cloudflare.com\/ajax\/libs\/tensorflow\/0.13.0\/tf.min.js\"&gt;&lt;\/script&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"52967162","CreationDate":"2018-10-24T10:53:44.313","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>I have changed the optimizer and it predicts as expected.<\/p>\n\n<pre><code>    [[0.0156993],\n     [0.985333 ],\n     [0.9862437],\n     [0.0150503]]\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential();\r\n\r\nmodel.add(tf.layers.dense({units: 2, activation: 'sigmoid', inputShape: [2]}));\r\nmodel.add(tf.layers.dense({units: 1, activation: 'sigmoid'}));\r\nconst optimizer = tf.train.adam(0.01);\r\nmodel.compile({loss:'meanSquaredError', optimizer: optimizer });\r\n\r\nconst xs = tf.tensor2d([[0,0],[0,1],[1,0],[1,1]]);\r\nconst ys = tf.tensor2d([[0],[1],[1],[0]]);\r\n\r\nasync function train() {\r\n    for(let i = 0; i &lt; 200; i++){\r\n        const history = await model.fit(xs, ys, {epochs: 20, shuffle: true});\r\n        console.log(\"loss: \" + history.history.loss[19] + \" on \" + i + \". iteration.\");\r\n    }\r\n}\r\n\r\ntrain().then(() =&gt; {\r\n    console.log(\"trained with \" + tf.memory().numTensors + \"tensors\");\r\n    model.predict(xs).print();\r\n});<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;script src=\"https:\/\/cdnjs.cloudflare.com\/ajax\/libs\/tensorflow\/0.13.0\/tf.min.js\"&gt;&lt;\/script&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":52965474,"AnswerCount":1,"Tags":"<cuda><pytorch>","CreationDate":"2018-10-24T09:26:32.123","AcceptedAnswerId":52965603.0,"OwnerUserId":10511793.0,"Title":"How to make cuda unavailable in pytorch","Body":"<p>i'm running some code with cudas, and I need to test the same code on CPU to compare running time. To decide between regular pytorch tensor and cuda float tensor, the library I use calls torch.cuda.is_available(). Is there an easy method to make this function return false? I tried changing the Cuda visible devices with <\/p>\n\n<pre><code>os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"\"\n<\/code><\/pre>\n\n<p>but torch.cuda.is_available() still return True. I went into pytorch source code, and in my case, torch.cuda.is_avaible returns <\/p>\n\n<pre><code>torch._C._cuda_getDeviceCount() &gt; 0\n<\/code><\/pre>\n\n<p>I assume I should be able to \"hide\" my GPU at the start of my notebook, so the device count is equal to zero, but i didn't get any success so far. Any help is appreciated :)<\/p>\n","answers":[{"AnswerId":"52965603","CreationDate":"2018-10-24T09:32:00.490","ParentId":null,"OwnerUserId":"240443","Title":null,"Body":"<blockquote>\n  <p>my code<\/p>\n<\/blockquote>\n\n<p>Instead of trying to trick it, why not rewrite your code? For example,<\/p>\n\n<pre><code>use_gpu = torch.cuda.is_available() and not os.environ['USE_CPU']\n<\/code><\/pre>\n\n<p>Then you can start your program as <code>python runme.py<\/code> to run on GPU if available, and <code>USE_CPU=1 python3 runme.py<\/code> to force CPU execution (or make it semi-permanent by <code>export USE_CPU=1<\/code>).<\/p>\n\n<blockquote>\n  <p>I tried changing the Cuda visible devices with<\/p>\n<\/blockquote>\n\n<p>You can also try running your code with <code>CUDA_VISIBLE_DEVICES=\"\" python3 runme.py<\/code>; if you're setting the environment variable inside your code, it might be set later than PyTorch initialisation, and have no effect.<\/p>\n"}]}
+{"QuestionId":52965686,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><computer-vision>","CreationDate":"2018-10-24T09:36:45.123","AcceptedAnswerId":52970511.0,"OwnerUserId":9134476.0,"Title":"difference between categorical and binary cross entropy","Body":"<p>Using keras I have to train a model to predict either the image belongs to class 0 or class 1. I am confused in binary and categorical_cross_entropy. I have searched for that but I am still confused. Some have mentioned that we only use categorical cross entropy when we are trying to predict multi-classes and we should use one-hot-encoder vector for this. So it means we dont need any one-hot-encoded vector labels when we are going to train using binary_cross_entrpoy. Some have suggested to represent one_hot vectors as [0. 1.] (if class is 1) or [1. 0.] (if class is 0) for binary_cross_entropy.\nI am using one hot encoders [0 1] or [1 0] with categorical cross entropy. My last layer is <\/p>\n\n<pre><code>model.add(Dense(num_classes, activation='softmax'))\n\n    # Compile model\n  model.compile(loss='categorical_crossentropy', optimizer='adadelta', metrics=['accuracy'])\n<\/code><\/pre>\n","answers":[{"AnswerId":"52970511","CreationDate":"2018-10-24T13:39:42.160","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>They are mathematically identical for <em>2 classes<\/em> hence binary. In other words, 2 class categorical cross entropy is the same as single output binary cross entropy. To give a more tangible example these are identical:<\/p>\n\n<pre><code>model.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', ...)\n# is the same as\nmodel.add(Dense(2, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', ...)\n<\/code><\/pre>\n\n<p><em>Which one to use?<\/em> To avoid one-hot encoding categorical outputs, if you only have 2 classes it is easier - from  a coding perspective - to use binary cross entropy. The binary case might be computationally more efficient depending on the implementation.<\/p>\n"},{"AnswerId":"52966179","CreationDate":"2018-10-24T10:01:27.230","ParentId":null,"OwnerUserId":"8609330","Title":null,"Body":"<p>Seems, <em>binary cross entropy<\/em> it's just a special case of the <em>categorical cross entropy<\/em>. So, when you have only two classes, you can use <em>binary cross entropy<\/em>, you don't need to do <em>one hot encoding<\/em> - your code will be couple of the lines less.<\/p>\n"}]}
+{"QuestionId":52965773,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-10-24T09:41:09.793","AcceptedAnswerId":52965846.0,"OwnerUserId":7177402.0,"Title":"Using convolution layer trained weights for different image size","Body":"<p>I want to use the first three convolution layers of vgg-16 to generate feature maps.\nBut i want to use it with variable image size,i.e not imagenet size of 224x224 or 256x256. Such as 480x640or any other randome image dimension. <\/p>\n\n<p>As convolution layer are independent of image spatial size, how can I use the weights for varying image sizes?\nSo how do we use the pre-trained weights of vgg-16 upto the first three convolution layers.\nKindly let me know if that is possible.<\/p>\n","answers":[{"AnswerId":"52968053","CreationDate":"2018-10-24T11:39:44.293","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>It is easy to reuse variables of Convolution layer. First, you define a graph containing convolution layer, then you restore values of them. The following is a pseudo code<\/p>\n\n<pre><code>def network(your_inputs):\n    filter1 = tf.get_variable(shape=[filter_size, filter_size, in_channel, out_channel], name=\"vgg16\/layer1\")\n    features = tf.nn.conv2d(your_inputs, filter1, strides=[1,1,1,1])\n    filter2 = tf.get_variable(shape=[filter_size, filter_size, in_channel, out_channel], name=\"vgg16\/layer2\")\n    features = tf.nn.conv2d(features, filter2, strides=[1,1,1,1])\n\n    restore_filters = [filter1, filter2]\n\n    ...\n\n    return logits, restore_filters \n\noutputs, restore_filters = network(inputs)\nsaver = tf.train.Saver(restore_filters)\nsaver.restore(sess, \"vgg-checkpoint.ckpt\")\n<\/code><\/pre>\n\n<p>Of course, you must match your filter sizes with VGG network. When you variable names differ from them of a checkpoint file, you can use Saver with dictionary argument.<\/p>\n"},{"AnswerId":"52965846","CreationDate":"2018-10-24T09:44:52.253","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p><code>As convolution layer are independent of image size<\/code><\/p>\n\n<p>Actually it's more complicated than that. The kernel itself is independent of the image size because we apply it on each pixel. And indeed, the training of these kernels can be reused.<\/p>\n\n<p>But this means that the output size is dependent on the image size, because this is the number of nodes that are fed out of the layer for each input pixel. So the dense layer is not adapted to your image, even if the feature extractors are independent.<\/p>\n\n<p>So you need to preprocess your image to fit into the size of the first layer or you retrain your dense layers from scratch.<\/p>\n\n<p>When people talk about \"transfer-learning\" is what people have done in segmentation for decades. You reuse the best feature extractors and then you train a dedicated model with these features.<\/p>\n"}]}
+{"QuestionId":52965934,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><conv-neural-network>","CreationDate":"2018-10-24T09:49:11.247","AcceptedAnswerId":null,"OwnerUserId":6941347.0,"Title":"Why we are going with more dimension if we can manage with single?","Body":"<p>I observed for images we are creating a model with same number of dimensions but we have option to gray it out that change dimension from 3 to 1. For example :- if we have a color image of dimension (with RGB) 28x28x3, same we feed as input to our model. With the same image if we convert it to gray using rgb2gray method, that converts its dimension from 28x28x3 to 28x28. Why we are not feeding this grayed image always to our model?<\/p>\n","answers":[{"AnswerId":"52980561","CreationDate":"2018-10-25T02:42:12.897","ParentId":null,"OwnerUserId":"2060205","Title":null,"Body":"<p>Of course it depends on what you want to do...<\/p>\n\n<p>If you want to train a model to discriminate between \"Granny Smith\" (green) and \"Pink Lady\" (red) apples, using greyscale images is not very helpful.<\/p>\n\n<p>If you want to train a model to discriminate between apples and bananas, a greyscale image should be enough.<\/p>\n"},{"AnswerId":"52966473","CreationDate":"2018-10-24T10:18:15.930","ParentId":null,"OwnerUserId":"9178720","Title":null,"Body":"<p>That approach depends on what you are trying to do. In some cases you need 3 channels in the image to achieve a better degree of precision. Since color is also information, smashing all the channels into a single channel discarts a lot of sometimes useful information. It is common to keep 3 channels in problems such as recognition of objects, segmentation, classification, etc. On the other hand gray scale is often used for document processing, HTR, OCR, etc.<\/p>\n\n<p>One more thing: in Tensorflow you are going the need that last dimension, so if you want to use gray scaled images they must me shaped 28x28x1. Assuming you are using OpenCv you are gonna have to reshape the image before feeding it to the net.<\/p>\n\n<p>Bare in mind that if the model was trained with gray images it won't work with color images for prediction!<\/p>\n\n<p>I hope it helped<\/p>\n"}]}
+{"QuestionId":52966089,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-24T09:57:02.960","AcceptedAnswerId":null,"OwnerUserId":4624229.0,"Title":"Create a tree structure of Keras layers in python","Body":"<p>I'm trying to obtain a structure of Keras layer. Within a simple network, this is possible by just iterating through the <code>model.layers<\/code>. However, if the network is complex (e.g. concatenates different layer) this approach is not sufficient.\nThis is an example:<\/p>\n\n<pre><code>FEATURES = ['A','B','C','D']\nIMPORTANT_FEATURES = [0, 3]\nNORMAL_FEATURES = [1, 2]\ninputLayer = [Input(shape=(1, )) for i,f in enumerate(FEATURES)]\nimportantInput = keras.layers.Concatenate(axis=-1)([inputLayer[i] for i in IMPORTANT_FEATURES])\nlayer1 = Dense()(importantInput)\nnormalInput  = keras.layers.Concatenate(axis=-1)(layer1 + [inputLayer[i] for i in NORMAL_FEATURES])\nlayer2 = Dense()(normalInput)\nmodel = Model([\n    inputLayer[i]\n    for i in range(len(FEATURES))\n], layer2)\n<\/code><\/pre>\n\n<p>This produces in <code>model.layer<\/code> a list of <strong>Keras<\/strong> layers composed like this:<\/p>\n\n<pre><code>[Input1, Input2, Concatenate1, Dense1, Input3, Input4, Concatenate2, Dense2]\n<\/code><\/pre>\n\n<p>The only way to get which level is in input to the concatenate is to access to <code>Concatenate1.input<\/code>. However, this returns a list of <strong>Tensorflow<\/strong> layers.\nIs it possible to obtain a tree-structure of layers just using <strong>Keras<\/strong> layers?<\/p>\n","answers":[{"AnswerId":"52966479","CreationDate":"2018-10-24T10:18:36.330","ParentId":null,"OwnerUserId":"4624229","Title":null,"Body":"<p>The <strong>Keras<\/strong> list of input layers can be obtained by <code>Concatenate1._inbound_nodes[0].inbound_layers<\/code><\/p>\n"}]}
+{"QuestionId":52966175,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><keras-layer>","CreationDate":"2018-10-24T10:01:16.327","AcceptedAnswerId":52966624.0,"OwnerUserId":8767101.0,"Title":"Keras TimeDistributed layer with multiple inputs","Body":"<p>I'm trying to make the following lines of code working:<\/p>\n\n<pre><code>low_encoder_out = TimeDistributed( AutoregressiveDecoder(...) )([X_tf, embeddings])\n<\/code><\/pre>\n\n<p>Where <code>AutoregressiveDecoder<\/code> is a custom layer that takes two inputs.\nAfter a bit of googling, the problem seems to be that the <code>TimeDistributed<\/code> wrapper doesn't accept multiple inputs. There are solutions that proposes to merge the two inputs before feeding it to the layer, but since their shape is<\/p>\n\n<pre><code>X_tf.shape: (?, 16, 16, 128, 5)\nembeddings.shape: (?, 16, 1024)\n<\/code><\/pre>\n\n<p>I really don't know how to merge them. Is there a way of having the <code>TimeDistributed<\/code> layer to work with more than one input? Or, alternatively, is there any way to merge the two inputs in a nice way?<\/p>\n","answers":[{"AnswerId":"52966624","CreationDate":"2018-10-24T10:26:14.327","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As you mentioned <code>TimeDistributed<\/code> layer does not support multiple inputs. One (not-very-nice) workaround, considering the fact that the number of timesteps (i.e. second axis) must be the same for all the inputs, is to reshape all of them to <code>(None, n_timsteps, n_featsN)<\/code>, concatenate them and then feed them as input of <code>TimeDistributed<\/code> layer:<\/p>\n\n<pre><code>X_tf_r = Reshape((n_timesteps, -1))(X_tf)\nembeddings_r = Reshape((n_timesteps, -1))(embeddings)\n\nconcat = concatenate([X_tf_r, embeddings_r])\nlow_encoder_out = TimeDistributed(AutoregressiveDecoder(...))(concat)\n<\/code><\/pre>\n\n<p>Of course, you might need to modify the definition of your custom layer and separate the inputs back if necessary.<\/p>\n"}]}
+{"QuestionId":52966243,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-24T10:05:08.360","AcceptedAnswerId":null,"OwnerUserId":987397.0,"Title":"Tensorflow: NotImplementedError: The reduce() transformation does not currently support nested datasets as inputs","Body":"<p>In Tensorflow 1.12 the <code>tf.data.Dataset.reduce()<\/code> and <code>tf.data.Dataset.window()<\/code> methods are introduced.<\/p>\n\n<p>From the release notes:<\/p>\n\n<ul>\n<li><p>\"New <code>tf.data.Dataset.reduce()<\/code> API allows users to reduce a finite dataset to a single element using a user-provided reduce function.\"<\/p><\/li>\n<li><p>\"New <code>tf.data.Dataset.window()<\/code> API allows users to create finite windows of input dataset; when combined with the <code>tf.data.Dataset.reduce()<\/code> API, this allows users to implement customized batching.\"<\/p><\/li>\n<\/ul>\n\n<p>However how to use these functions?<\/p>\n\n<pre><code>def reduce_func(old_state, input_element):\n    pdb.set_trace()\n    return new_state\n\ndataset = tf.data.Dataset.from_generator(frame_generator, (tf.string, tf.string))\ndataset = dataset.window(2).reduce(np.int64(0), reduce_func)\n<\/code><\/pre>\n\n<p>This gives a NotImplementedError:<\/p>\n\n<blockquote>\n  <p>NotImplementedError: The reduce() transformation does not currently<br>\n  support nested datasets as inputs. <\/p>\n<\/blockquote>\n\n<p>I use tensorflow version '1.12.0-rc1'<\/p>\n\n<p>EDIT:\nFrom <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.12\/api_docs\/python\/tf\/contrib\/data\/sliding_window_batch\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.12\/api_docs\/python\/tf\/contrib\/data\/sliding_window_batch<\/a><\/p>\n\n<p><em>THIS FUNCTION IS DEPRECATED.<\/em> It will be removed in a future version. Instructions for updating: Use <code>tf.data.Dataset.window(size=window_size, shift=window_shift, stride=window_stride).flat_map(lambda x: x.batch(window.size))<\/code><\/p>\n\n<p>But how to use this if the dataset is generated with<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_generator(frame_generator, (tf.string, tf.string))\n<\/code><\/pre>\n\n<p>So each item in the dataset contains two elements. Then there is a TypeError:<\/p>\n\n<blockquote>\n  <p>TypeError: () takes 1 positional argument but 2 were given<\/p>\n<\/blockquote>\n\n<p>EDIT: \nSolved by using zip<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_generator(frame_generator, (tf.string, tf.string))\nwindow_size = 2\ndataset = dataset.window(window_size).flat_map(lambda x,y: tf.data.Dataset.zip((x,y)).batch(window_size))\ndataset = dataset.map(self.parse_function)\n<\/code><\/pre>\n","answers":[{"AnswerId":"55624542","CreationDate":"2019-04-11T04:07:39.837","ParentId":null,"OwnerUserId":"1387006","Title":null,"Body":"<p>The problem here is that when you call the <code>.reduce()<\/code> function it's applying to the outer dataset.\nThe dataset after calling <code>.window<\/code> is now a dataset where each element is itself a dataset.\nWhat you want to do is to use reduce on those individual datasets created by window.\nYou can do this using map, and then mapping each of the inner datasets to a reduce.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def reduce_func(old_state, input_element):\n    pdb.set_trace()\n    return new_state\n\ndataset = tf.data.Dataset.from_generator(frame_generator, (tf.string, tf.string))\ndataset = dataset.window(2).map(lambda ds: ds.reduce(np.int64(0), reduce_func))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52966617,"AnswerCount":0,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2018-10-24T10:25:49.213","AcceptedAnswerId":null,"OwnerUserId":7438024.0,"Title":"tf.string_input_producer gives out single epoch even after setting epochs greater than one","Body":"<p>My input api uses tf.string_input_producer along with tf.parse_single_sequence_example. When I set num_epochs > 1 in tf.string_input_producer my queue still finishes just after one epochs.<\/p>\n\n<p>Is it the expected behaviour or I am doing some mistake ? \nHere is the relevant code :<\/p>\n\n<pre><code>class TFRecordReader():\n\n\n\n    def __init__(self):\n        #some code....\n\n    def execute_queue(self, tensor_queue, exception_message: str, log_dir_path=None):\n        import os\n        if log_dir_path is None:\n            path = os.path.abspath('..\/..\/audio_log_dir\/')\n        else:\n            path = log_dir_path\n        writer = self._summary_file_writer(path)\n        coord, thread = self._coord_thread()\n        print('should_stop: ', coord.should_stop())\n\n        if not coord.should_stop():\n            try:\n                if self._data_v is None:\n                    self._data_v = self._parse_tensor(tensor_queue)\n\n                 return self._data_v\n            except self.tf.errors.OutOfRangeError:\n                print(exception_message)\n            finally:\n                coord.request_stop()\n                coord.join(thread)\n                writer.close()\n\n\n    def single_sequence_batch(self,\n                              tf_record_path,\n                              feature_map,\n                              parse_function,\n                              num_epochs=None,\n                              tf_record_compression=None,\n                              queue_completion_message='Data Exhausted!',\n                              log_dir_path=None\n                              ):\n        self.feature_map = feature_map\n        self.parse_func = parse_function\n        batch = self._single_sequence_batch(tf_record_path=tf_record_path,\n                                        num_epochs=num_epochs,\n                                        tf_record_compression=tf_record_compression)\n        data_queue = self.execute_queue(batch, queue_completion_message, log_dir_path=log_dir_path)\n        return data_queue\n\ndef _test_single_sequence_batch(num_epochs=1):\n    tfr_path = r'C:\/audio_tfrecord\/audioapi.tfrecord'\n    reader = TFRecordReader()\n    data_q = reader.single_sequence_batch(tf_record_path=tfr_path,\n                                      feature_map=feature_mapping,\n                                      parse_function=parse_func,\n                                      tf_record_compression=True,\n                                      num_epochs=num_epochs)\n    print(len(data_q))\n    c = 0\n    try:\n        for i in range(num_epochs):\n            val = reader.session.run(data_q)\n            print(val)\n            c += 1\n    except tf.errors.OutOfRangeError:\n        print(\"Total Examples :\", c)\n        print('Finished!')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52966677,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-24T10:28:53.420","AcceptedAnswerId":52969128.0,"OwnerUserId":7797315.0,"Title":"Tensorflow: imported ops always result in \"None\" despite valid input & kernel","Body":"<p>I would like to do following:<\/p>\n\n<ul>\n<li>a) define model graph <\/li>\n<li>b) initiate the model and train it for a very\nlimited NUMBER_OF_PRE_ITERATIONS (let's say,\nNUMBER_OF_PRE_ITERATIONS=20). Save the model.  <\/li>\n<li>c) repeat point b)\nNUMBER_OF_PRE_RUNS times (e.g. NUMBER_OF_PRE_RUNS=3). Each time, if a\n\"new\" model gives better results than the \"saved\" one, \"new\" model\ngets saved.<\/li>\n<\/ul>\n\n<p>As a result of c), I should have a model with (probably) best initial values out of NUMBER_OF_PRE_RUNS models.<\/p>\n\n<ul>\n<li>d) import the \"saved\" (probably) best model and proceed with training.<\/li>\n<\/ul>\n\n<p>PROBLEM: every time, all ops of the imported model result in \"None\" value.\nInput figures seem valid, kernel &amp; bias parameters as well. No \"None\" or any other suspicious figures there, just float numbers. So why \"None\" as result of every op?<\/p>\n\n<p>Here is the (simplified) code:<\/p>\n\n<pre><code># imports\n# declarations of variables etc.\n\ntf.reset_default_graph()\n\nwith tf.variable_scope(name_or_scope='Model_', reuse=tf.AUTO_REUSE):\n    x = tf.placeholder(...)\n    y = tf.placeholder(...)\n    input_layer = tf.layers.dense(...)\n    inter_layer = tf.layers.dense(...)\n    output_layer = tf.layers.dense(...)\n\n    loss_funct = tf.reduce_mean(tf.abs(y - output_layer), name='loss_f')\n    optimit = tf.train.GradientDescentOptimizer(learning_rate=LEARNING_RATE, name='optimit')\n    train = optimit.minimize(loss=loss_funct, name='train')\n\n\nsaver = tf.train.Saver()\n\nbest_pre_result = 99999\nfor pre_run in range(NUMBER_OF_PRE_RUNS):\n    with tf.Session() as presess:\n\n        presess.run(tf.global_variables_initializer())\n\n        for pre_iteration in range(NUMBER_OF_PRE_ITERATIONS):\n            for batch in range(NUMBER_OF_BATCHES):\n                x_batch, y_batch = get_batches(...)\n                presess.run(train, feed_dict={x: x_batch, y: y_batch})\n            if pre_iteration % 10 == 0:\n                # Monitor performance here. Training seems to progress in valid way.\n            if pre_iteration == NUMBER_OF_PRE_ITERATIONS - 1:\n                pre_result = presess.run(loss_funct, feed_dict={x: all_features, y: all_targets})\n                if pre_result &lt; best_pre_result:\n                    best_pre_result = pre_result\n                    model_path = saver.save(sess=presess, save_path=SAVE_MODEL_PATH)\n\n# tried without the below line, but result was the same.\ntf.reset_default_graph()\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())        \n\n    restorer = tf.train.import_meta_graph(SAVE_MODEL_PATH + \".meta\")\n    restorer.restore(sess=sess, save_path=model_path)\n\n    x_ = tf.get_default_graph().get_tensor_by_name(name='Model_\/x:0')\n    y_ = tf.get_default_graph().get_tensor_by_name(name='Model_\/y:0')\n    compute_input_ = tf.get_default_graph().get_operation_by_name(name='Model_\/input_layer\/Tanh')\n    # otehr operations &amp; tensors imported here for performance monitoring\n    loss_funct_ = tf.get_default_graph().get_operation_by_name(name='Model_\/loss_f')\n    train_ = tf.get_default_graph().get_operation_by_name(name='Model_\/train')\n\n    for iteration in range(...):\n        for batch in range(...):\n            x_batch, y_batch = get_batches(...)\n            sess.run(train_, feed_dict={x_: x_batch, y_: y_batch})\n            # Monitor performance here. All ops result in \"None\".\n<\/code><\/pre>\n","answers":[{"AnswerId":"52969128","CreationDate":"2018-10-24T12:34:23.793","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>What tensor do you get None from? If you get None from loss_funct, it is the reason that you used get_operation_by_name for loss_funct. We should use get_tensor_by_name to get a result of a operation. <\/p>\n\n<p>In the following code, the last line shows None.<\/p>\n\n<pre><code>with tf.variable_scope(\"model\"):\n    x = tf.get_variable(name=\"x\", shape=[3], initializer=tf.initializers.random_normal())\n    y = tf.reduce_sum(x, name=\"y\")\n\nsaver = tf.train.Saver()\nsave_path = \".\/saver\"\npre_max = -10000\n\nfor i in range(3):\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        cur_y = sess.run(y)\n        print(cur_y)\n\n        if cur_y &gt; pre_max:\n            pre_max = cur_y\n            model_path = saver.save(sess=sess, save_path=save_path)\n\nprint(\"max_y : \", pre_max)\ntf.reset_default_graph()\n\nwith tf.Session() as sess:\n    restorerer = tf.train.import_meta_graph(save_path + \".meta\")\n    restorerer.restore(sess, save_path)\n\n    y_tensor = tf.get_default_graph().get_tensor_by_name(name='model\/y:0')\n    y_op = tf.get_default_graph().get_operation_by_name(name='model\/y')\n\n    print(sess.run(y_tensor))\n    print(sess.run(y_op))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52966765,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-24T10:33:32.837","AcceptedAnswerId":52967811.0,"OwnerUserId":10551333.0,"Title":"Error when checking target: expected softmax_1 to have shape (1,) but got array with shape (2,)', Keras","Body":"<p>I am using Keras for building Conv Net for the first time. My layers are as follows:<\/p>\n\n<pre><code>layers = [\nConv2D(8,kernel_size=(4,4),padding='same',input_shape=( 200, 180,3),kernel_initializer=\"glorot_normal\",data_format=\"channels_first\"),\nActivation(\"relu\"),\nMaxPooling2D(pool_size=(8,8),padding='same',data_format='channels_first'),\nConv2D(16,(2,2),padding='same',kernel_initializer=\"glorot_normal\"),\nActivation(\"relu\"),\nMaxPooling2D(pool_size=(4,4),padding='same',data_format='channels_first'),\nConv2D(4,(3,3),padding='same',kernel_initializer=\"glorot_normal\"),\nActivation(\"relu\"),\nMaxPooling2D(pool_size=(2,2),padding='same',data_format='channels_first'),\nFlatten(),\nDense(2,input_shape=(48,)),\nSoftmax(axis=-1)\n]\n#Edit, here is the part for compiling the model and fitting it\nmodel = Sequential(layers)    \n\nmodel.compile(optimizer=\"adam\",loss=\"sparse_categorical_crossentropy\" \nmetrics=[\"accuracy\"])\ntrainHistory = model.fit(x=X_train,y=Y_train,batch_size=3,epochs=1000)\n<\/code><\/pre>\n\n<p>My labels array is of shape <code>(,2)<\/code>. But when I try to use fit on the model, it gives me the error that <code>softmax_1<\/code> expected to have shape <code>(1,)<\/code>. But I have clearly mentioned units of <code>Dense<\/code> as 2 and softmax returns output of the same dimension as the input.<\/p>\n\n<p>So where did the 1, came from? I tried to use dummy label array of 1 dimension and it runs. So what am I doing wrong? How do I use 2 dimensional array that I have?<\/p>\n","answers":[{"AnswerId":"52967811","CreationDate":"2018-10-24T11:27:42.693","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are using <code>sparse_categorical_crossentropy<\/code> as the loss function. This loss function is used when the given labels (i.e. <code>Y_train<\/code>) are encoded as integers (i.e. 0, 1, 2, ...). However, If the labels are one-hot encoded, which seems to be the case in your code, you need to use <code>categorical_crossentropy<\/code> as the loss function instead.<\/p>\n"}]}
+{"QuestionId":52966950,"AnswerCount":1,"Tags":"<pytorch><rnn><google-colaboratory>","CreationDate":"2018-10-24T10:42:54.643","AcceptedAnswerId":null,"OwnerUserId":8547664.0,"Title":"Error while executing RNN code on pyTorch?","Body":"<p>I am running a code on Binary Addition of two strings using PyTorch. <\/p>\n\n<p>However, while training the model I am getting the following error:<\/p>\n\n<pre><code>can't convert np.ndarray of type numpy.object. \nThe only supported types are: double, float, float16, int64, int32, and uint8.  \n<\/code><\/pre>\n\n<p>Can anyone help me? Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>featDim=2 # two bits each from each of the String\noutputDim=1 # one output node which would output a zero or 1\n\nlstmSize=10\n\nlossFunction = nn.MSELoss()\nmodel =Adder(featDim, lstmSize, outputDim)\nprint ('model initialized')\n#optimizer = optim.SGD(model.parameters(), lr=3e-2, momentum=0.8)\noptimizer=optim.Adam(model.parameters(),lr=0.001)\nepochs=500\n### epochs ##\ntotalLoss= float(\"inf\")\nwhile totalLoss &gt; 1e-5:\n  print(\" Avg. Loss for last 500 samples = %lf\"%(totalLoss))\n  totalLoss=0\n  for i in range(0,epochs): # average the loss over 200 samples\n\n    stringLen=4\n    testFlag=0\n    x,y=getSample(stringLen, testFlag)\n\n    model.zero_grad()\n\n\n    x_var=autograd.Variable(torch.from_numpy(x).unsqueeze(1).float()) #convert to torch tensor and variable\n    # unsqueeze() is used to add the extra dimension since\n    # your input need to be of t*batchsize*featDim; you cant do away with the batch in pytorch\n    seqLen=x_var.size(0)\n    #print (x_var)\n    x_var= x_var.contiguous()\n    y_var=autograd.Variable(torch.from_numpy(y).float()) ##ERROR ON THIS LINE\n    finalScores = model(x_var)\n    #finalScores=finalScores.\n\n    loss=lossFunction(finalScores,y_var)  \n    totalLoss+=loss.data[0]\n    optimizer.zero_grad()\n    loss.backward()\n    optimizer.step()\n\n\n  totalLoss=totalLoss\/epochs\n<\/code><\/pre>\n","answers":[{"AnswerId":"52970436","CreationDate":"2018-10-24T13:36:05.343","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p><em>The main problem here is the type of your <code>y<\/code>. You haven't given any information about this, therefore this here will be more general:<\/em><\/p>\n\n<p>But obviously your <code>ndarray<\/code> doesn't contain numeric data types. You have to use one of these mentioned in your error message:<\/p>\n\n<blockquote>\n  <p>The only supported types are: double, float, float16, int64, int32,\n  and uint8.<\/p>\n<\/blockquote>\n\n<p>So here is a short example to demonstrate the issue:<\/p>\n\n<p>If you use one of the previously mentioned data types it works just fine:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport numpy as np\na = np.ndarray(shape=(2,2), dtype=np.float) # data type np.float\nprint(a)\nprint(torch.autograd.Variable(torch.from_numpy(a).float()))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[[2.16641777e-314 2.16641777e-314]\n [2.16641777e-314 2.16641777e-314]]\nVariable containing:\n 0  0\n 0  0\n[torch.FloatTensor of size 2x2]\n<\/code><\/pre>\n\n<p><hr>\nBut if you use some other numpy data type (like <code>np.object<\/code>) you will get this error message:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport numpy as np\na = np.ndarray(shape=(2,2), dtype=np.object) # data type np.object\nprint(a)\nprint(torch.autograd.Variable(torch.from_numpy(a).float()))\n<\/code><\/pre>\n\n<p>This results in:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[[None None]\n [None None]]\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-27-01e1e4bec020&gt; in &lt;module&gt;()\n      3 a = np.ndarray(shape=(2,2), dtype=np.object)\n      4 print(a)\n----&gt; 5 print(torch.autograd.Variable(torch.from_numpy(a).float()))\n\nRuntimeError: can't convert a given np.ndarray to a tensor - it has an invalid type. The only supported types are: double, float, int64, int32, and uint8.\n<\/code><\/pre>\n\n<p>You probably haven't specified the data type <code>np.object<\/code> directly. I guess this is maybe a result of some nested arrays or so. <\/p>\n\n<p>But your need to bring your numpy array <code>y<\/code> into proper shape with a numeric data type then it should work for you.<\/p>\n"}]}
+{"QuestionId":52966986,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><google-cloud-platform><recommendation-engine><google-cloud-ml>","CreationDate":"2018-10-24T10:44:24.897","AcceptedAnswerId":null,"OwnerUserId":7863752.0,"Title":"Item to Item recommendation engine on Google Cloud","Body":"<p>I found an implementation of the recommendation engine on Google Cloud <a href=\"https:\/\/cloud.google.com\/solutions\/machine-learning\/recommendation-system-tensorflow-overview\" rel=\"nofollow noreferrer\">here<\/a>, which turned out to be the <strong>User to Item<\/strong> recommendation.<\/p>\n\n<p>I was wondering whether there is the similar solution for Google Cloud with <strong>Item to Item<\/strong> recommendation.<\/p>\n","answers":[{"AnswerId":"52967710","CreationDate":"2018-10-24T11:23:14.483","ParentId":null,"OwnerUserId":"7438024","Title":null,"Body":"<p>You can just replace the <strong>User<\/strong> with <strong>Item<\/strong> data. Treat Item as User component. I guess rest of the things remains same in the SVD Matrix.<\/p>\n"}]}
+{"QuestionId":52967355,"AnswerCount":2,"Tags":"<tensorflow><keras><deep-learning><artificial-intelligence>","CreationDate":"2018-10-24T11:03:32.817","AcceptedAnswerId":null,"OwnerUserId":2623001.0,"Title":"Deep Learning architecture for input and output of different size?","Body":"<p>I am quite new to deep learning and I am trying to play around with Tensorflow and Keras. I have worked through some of the tutorials that explain how to use CNN on labeled data. <\/p>\n\n<p>Now I want to try something different. I have input data of size (m,n,o) and output data of size (m,n,p). Thus input and output data have the same number of pixels (m,n) but a different number of channels. Using the o parameters per pixel an output pixel with p channels should be generated.<\/p>\n\n<p>How do I achieve that? Is a CNN a proper architecture or should I use a different architecture? Do you know any tutorials or buzz words regarding this problem? <\/p>\n","answers":[{"AnswerId":"52967644","CreationDate":"2018-10-24T11:20:14.843","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>You could look into <a href=\"http:\/\/blog.qure.ai\/notes\/semantic-segmentation-deep-learning-review\" rel=\"nofollow noreferrer\">Semantic Segmentation<\/a>. That sounds very similar. Usually the input has 3 channels (RGB) and the number of output channels equals the number of classes you want to detect.<\/p>\n\n<p>There are quite a few CNN based architectures for this task, some of which are described in the link above, that could be maybe used for your task.<\/p>\n"},{"AnswerId":"52975232","CreationDate":"2018-10-24T17:53:42.927","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>The applicable phrase in my part of the mind-space is \"image-2-image\".  Whether you want a CNN depends very much on what your translation involves.  If the transformation from <code>o<\/code> channels to <code>p<\/code> channels is independent of context, then CNN is <em>huge<\/em> overkill.<\/p>\n\n<p>For instance, if you're trying to translate from display phosphors (RGB) to printer's separation (CYMK), then you need only a simple matrix multiplication on each individual pixel.<\/p>\n\n<p>If, instead, you're involving spatial proximity and regional features to make the transformation -- using nearby pixels to recognize shapes, then generating the output image from those regional shapes -- then a CNN is warranted.<\/p>\n"}]}
+{"QuestionId":52968575,"AnswerCount":1,"Tags":"<tensorflow><tensorflow.js>","CreationDate":"2018-10-24T12:05:23.850","AcceptedAnswerId":null,"OwnerUserId":1395305.0,"Title":"Realtime Single Object Tracking with TensorFlow.js","Body":"<p>Just like everyone, I'm putting my first steps in Machine Learning, I went through many TensorFlow.js tutorials already and I'm trying to achieve this: \"Realtime Single Object Tracking\/Detection\"<\/p>\n\n<p>Something like this -> input: webcam\/video -> output: object bounding box<\/p>\n\n<p>I know there are SSD and YOLO, and other libraries to predict &amp; locate the objects. But the predicted time is very slow (in browser), I guessed it's because the Neural Network have to predict between so many objects.<\/p>\n\n<p><a href=\"https:\/\/github.com\/ModelDepot\/tfjs-yolo-tiny\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ModelDepot\/tfjs-yolo-tiny<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection<\/a><\/p>\n\n<p>What if I just want to track a single object? Would it be possible? Will the performance be better? Where should I start? <\/p>\n\n<p>I've been thinking about extract the pre-trained class (object) from a SavedModel, then start training more from it. But there doesn't seems to have any instructions around Google :)<\/p>\n\n<p>If someone can give any clues, I would be very appreciated!<\/p>\n\n<p>Thanks in advanced!<\/p>\n","answers":[{"AnswerId":"56318579","CreationDate":"2019-05-27T01:12:50.627","ParentId":null,"OwnerUserId":"1250389","Title":null,"Body":"<p>I found some fantastic code by IBM, which I used in the video in this tweet:  <a href=\"https:\/\/twitter.com\/GantLaborde\/status\/1125735283343921152?s=20\" rel=\"nofollow noreferrer\">https:\/\/twitter.com\/GantLaborde\/status\/1125735283343921152?s=20<\/a><\/p>\n\n<p>I extracted that code to make a ReactJS component for detecting Rock\/Paper\/Scissors here: <a href=\"https:\/\/github.com\/GantMan\/rps_tfjs_demo\/blob\/master\/src\/AdvancedModel.js\" rel=\"nofollow noreferrer\">https:\/\/github.com\/GantMan\/rps_tfjs_demo\/blob\/master\/src\/AdvancedModel.js<\/a><\/p>\n\n<p>If you'd like to play with the demo, it's at the bottom of this page: <a href=\"https:\/\/rps-tfjs.netlify.com\/\" rel=\"nofollow noreferrer\">https:\/\/rps-tfjs.netlify.com\/<\/a><\/p>\n\n<p>All of this is open source and seems to work perfectly fast for detecting a single object in realtime.<a href=\"https:\/\/i.stack.imgur.com\/muL0l.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/muL0l.gif\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":52968865,"AnswerCount":1,"Tags":"<r><indexing><keras><tokenize>","CreationDate":"2018-10-24T12:21:01.040","AcceptedAnswerId":52970391.0,"OwnerUserId":9770403.0,"Title":"InvalidArgumentError: indices[127,7] = 43 is not in [0, 43) in Keras R","Body":"<p>The issue is related \n to: <a href=\"https:\/\/stackoverflow.com\/questions\/42567398\/invalidargumenterror-see-above-for-traceback-indices1-10-is-not-in-0-10\/52310338#52310338\">InvalidArgumentError (see above for traceback): indices[1] = 10 is not in [0, 10)<\/a>\nI need it for R and therefore another solution than given in the link above. <\/p>\n\n<pre><code>maxlen &lt;- 40\nchars &lt;- c(\"'\",  \"-\",  \" \",  \"!\",  \"\\\"\", \"(\",  \")\",  \",\",  \".\",  \":\",  \";\",  \"?\",  \"[\",  \"]\",  \"_\",  \"=\",  \"0\", \"a\",  \"b\",  \"c\",  \"d\",  \"e\", \"f\",  \"g\",  \"h\",  \"i\",  \"j\",  \"k\",  \"l\",  \"m\",  \"n\",  \"o\",  \"p\",  \"q\",  \"r\",  \"s\",  \"t\",  \"u\",  \"v\",  \"w\",  \"x\",  \"y\",  \"z\")\n\n\n\ntokenizer &lt;- text_tokenizer(char_level = T, filters = NULL)\n\ntokenizer %&gt;% fit_text_tokenizer(chars)\nunlist(tokenizer$word_index)\n<\/code><\/pre>\n\n<p>Output is: <\/p>\n\n<pre><code> '  -     !  \"  (  )  ,  .  :  ;  ?  [  ]  _  =  0  a  b  c  d  e  f  g  h  i  j  k  l  m  n  o  p  q  r  s  t  u  v  w  x  y  z \n 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 \n<\/code><\/pre>\n\n<p>How can I change the indexing so it starts from 0 not from 1 in text_tokenizer?<\/p>\n\n<p>The error I get after running fit() is as follows:<\/p>\n\n<pre><code>InvalidArgumentError: indices[127,7] = 43 is not in [0, 43)\n     [[Node: embedding_3\/embedding_lookup = GatherV2[Taxis=DT_INT32, Tindices=DT_INT32, Tparams=DT_FLOAT, _class=[\"loc:@training_1\/RMSprop\/Assign_1\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](embedding_3\/embeddings\/read, embedding_3\/Cast, training_1\/RMSprop\/gradients\/embedding_3\/embedding_lookup_grad\/concat\/axis)]]\n<\/code><\/pre>\n\n<p>But I believe that changing the Indexing will solve my problem.<\/p>\n","answers":[{"AnswerId":"52970391","CreationDate":"2018-10-24T13:33:49.350","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>Index 0 is often reserved for <strong>padding<\/strong> so it is not a wise idea to start your actual character indices from 0 as well. Instead you should venture to the <code>Embedding<\/code> layer and add 1 to the input size as suggested by the <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>input_dim:<\/strong> int > 0. Size of the vocabulary, i.e. maximum integer index + 1.<\/p>\n<\/blockquote>\n\n<p>In your case this would be 43 + 1 = 44.<\/p>\n"}]}
+{"QuestionId":52969630,"AnswerCount":1,"Tags":"<windows><tensorflow><compilation><bazel>","CreationDate":"2018-10-24T12:57:46.110","AcceptedAnswerId":53006033.0,"OwnerUserId":3979919.0,"Title":"Tensorflow with bazel: avx ignored?","Body":"<p>I'm trying to compile tensorflow by hand to take Advantage of avx.\nUnfortunately I it seems to be ignoring specifically these instructions.\nI might add that I'm working on a Windows 10.\nAny help would be appreciated. Thank you.<\/p>\n\n<p>I used the following command for bazel:<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Oefzm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Oefzm.png\" alt=\"Compile with bazel\"><\/a><\/p>\n","answers":[{"AnswerId":"53006033","CreationDate":"2018-10-26T09:51:03.193","ParentId":null,"OwnerUserId":"6521181","Title":null,"Body":"<p><code>-mavx<\/code> and so on are for GCC and Clang compilers only. Since you are using MSVC, you should use <code>--copt=\/arch:AVX --copt=\/arch:AVX2<\/code> and so on instead.<\/p>\n"}]}
+{"QuestionId":52969706,"AnswerCount":0,"Tags":"<c++><caffe>","CreationDate":"2018-10-24T13:00:55.743","AcceptedAnswerId":null,"OwnerUserId":10285454.0,"Title":"How to add a c++ vector of weights to a caffe model in C++","Body":"<p>In the c++ program below how do I add the c++ vector of weights to the convolution layer? <\/p>\n\n<p>I have a <code>caffe::ConvolutionParameter<\/code> obejct <code>convParam<\/code>, how do I add weights to it? Do I need to use a <code>caffe::BlobProto<\/code> object?<\/p>\n\n<p>Do I need to loop and copy the weights in one by one?<\/p>\n\n<p>I've searched online and cannot find any examples of how to do this in C++. <\/p>\n\n<p>When I have this done I will write it to a caffemodel file. I am already writing the network to a prototxt file.<\/p>\n\n<pre><code>#include &lt;fcntl.h&gt;\n#include &lt;google\/protobuf\/io\/coded_stream.h&gt;\n#include &lt;google\/protobuf\/io\/zero_copy_stream_impl.h&gt;\n#include &lt;google\/protobuf\/text_format.h&gt;\n#include &lt;stdint.h&gt;\n#include &lt;algorithm&gt;\n#include &lt;fstream&gt; \/\/ NOLINT(readability\/streams)\n#include &lt;string&gt;\n#include &lt;vector&gt;\n#include &lt;string&gt;\n#include \"caffe.pb.h\"\n#include &lt;iostream&gt;\n#include &lt;caffe\/caffe.hpp&gt;\n\nusing namespace std;\nusing google::protobuf::Message;\nusing google::protobuf::io::CodedInputStream;\nusing google::protobuf::io::CodedOutputStream;\nusing google::protobuf::io::FileInputStream;\nusing google::protobuf::io::FileOutputStream;\nusing google::protobuf::io::ZeroCopyInputStream;\nusing google::protobuf::io::ZeroCopyOutputStream;\n\n\/*Network object*\/\ncaffe::NetParameter netParam; \n\nint main()\n{\n           caffe::LayerParameter *layerParam = netParam.add_layer();\n\n            \/*Set name and type of the layer*\/\n            layerParam-&gt;set_name(\"Test\";\n            layerParam-&gt;set_type(\"Convolution\");\n\n            \/*Get the input operation*\/\n            layerParam-&gt;add_bottom(\"bootom\");\n\n            \/*Get the output operation*\/\n            layerParam-&gt;add_top(\"top\");\n\n            \/*Set layer to have a conv parameter*\/\n            caffe::ConvolutionParameter *convParam = layerParam-&gt;mutable_convolution_param();\n\n            \/*Set stride on ConvolutionParameter object*\/\n            convParam-&gt;add_stride(2);\n\n            \/*Set kernel on ConvolutionParameter object*\/\n            convParam-&gt;add_kernel_size(3);\n\n            \/*Set number of output channels*\/\n            convParam-&gt;set_num_output(64);\n\n             \/*weights*\/\n             vector&lt;int&gt; weights{ 10, 20, 30 }; \n\n            \/\/How to add weights to the layer??\n\n            \/\/add_blobs()\n            caffe::BlobProto *blobProto = layerParam-&gt;add_blobs();\n            blobProto-&gt;has_shape();\n\n\n            \/*create prototxt*\/\n            std::ofstream ofs;\n            ofs.open(\"test.prototxt\", std::ofstream::out | std::ofstream::app);\n            ofs &lt;&lt; netParam.Utf8DebugString();\n            ofs.close();\n\n            return 0;\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52969708,"AnswerCount":1,"Tags":"<python><tensorflow><failed-installation>","CreationDate":"2018-10-24T13:00:59.060","AcceptedAnswerId":null,"OwnerUserId":10552114.0,"Title":"ModuleNotFoundError: No module named 'tensorflow', even when tensorflow is installed","Body":"<p>I am trying to install tensorflow for one of my machine learning projects. However,  even though I have installed it, I still get this error<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>To help illustrate this better, I have created a <code>test.py<\/code> file, with the following content:<\/p>\n\n<pre><code>import tensorflow as tf\n\nprint('Hello world!')\n<\/code><\/pre>\n\n<p>However, still the same error, on line 1.<\/p>\n\n<p><strong>Relevant questions:<\/strong><\/p>\n\n<p>I've tried doing many other answers, but none of them seems to help. Any answers would be appreciated.<\/p>\n\n<p><strong>Here's some debugging outputs that might help:<\/strong><\/p>\n\n<blockquote>\n  <p>pip3 show tensorflow<\/p>\n<\/blockquote>\n\n<pre><code>Name: tensorflow\nVersion: 1.11.0\nSummary: TensorFlow is an open source machine learning framework for everyone.\nHome-page: https:\/\/www.tensorflow.org\/\nAuthor: Google Inc.\nAuthor-email: opensource@google.com\nLicense: Apache 2.0\nLocation: c:\\program files\\anaconda3\\lib\\site-packages\nRequires: absl-py, termcolor, keras-applications, astor, six, tensorboard, keras-preprocessing, wheel, gast, setuptools, grpcio, protobuf, numpy\nRequired-by:\n<\/code><\/pre>\n\n<blockquote>\n  <p>pip3 --version<\/p>\n<\/blockquote>\n\n<pre><code>pip 18.1 from c:\\program files\\anaconda3\\lib\\site-packages\\pip (python 3.6)\n<\/code><\/pre>\n\n<blockquote>\n  <p>python --version<\/p>\n<\/blockquote>\n\n<pre><code>Python 3.6.0 :: Anaconda 4.3.0 (64-bit)\n<\/code><\/pre>\n\n<blockquote>\n  <p>py test.py<\/p>\n<\/blockquote>\n\n<pre><code>Traceback (most recent call last):\n  File \"test.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\nModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n","answers":[{"AnswerId":"52985051","CreationDate":"2018-10-25T08:45:08.333","ParentId":null,"OwnerUserId":"10552114","Title":null,"Body":"<p>I fixed it! Special thanks to the folks at the Tensorflow Talk slack who helped me, especially @akofman.<\/p>\n\n<p>It was a combination of 2 problems:<\/p>\n\n<p><strong>Problem 1<\/strong><\/p>\n\n<p>It seems that one of the reasons it was failing was due to one of tensorflow's dependencies being outdated\/misinstalled\/something. The dependency is <code>h5py<\/code>. I found out about this by attempting to run <code>import tensorflow<\/code> in the python interpreter (type <code>python<\/code>), which gave me a long stack trace, unlike the test file (see problem 2). I fixed this by reinstalling the dependency.<\/p>\n\n<p><strong>Problem 2<\/strong><\/p>\n\n<p>It turns out that I have 3, <em>that's right, 3!<\/em>, versions of python on my computer<\/p>\n\n<p><code>python -V<\/code> ---> 3.6.0<\/p>\n\n<p><code>python -V<\/code> (in an anaconda enviorment) ---> 3.6.7<\/p>\n\n<p><code>py -V<\/code> ---> 3.7.0<\/p>\n\n<p>I was running my test file with <code>py<\/code>, which is apparently 3.7.0 (I thought it was synonymous with <code>python<\/code>), I guess tensorflow doesn't support that version?<\/p>\n"}]}
+{"QuestionId":52969739,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-24T13:02:07.710","AcceptedAnswerId":null,"OwnerUserId":9421400.0,"Title":"Convolutional neural network model not learning anything in keras","Body":"<p>I have this convolutional neural network in Keras: <\/p>\n\n<pre><code>storefeaturesInfo = Input(shape=(21,), name='storefeatures')\nstorefeaturesInfo_x=Dense(64,activation='relu')(storefeaturesInfo)\n\nnext48additionalInfo=Input(shape=(None, 8), name='next48additionalInfo')\nx_next48additionalInfo=GRU(32)(next48additionalInfo)\nnext48additionalInfo_x=Dense(64,activation='relu')(x_next48additionalInfo)\n\ninput1 = Input(shape=(None, 10), name='firstfewtimesteps')    \nx=Conv1D(24, 3, activation='relu', padding=\"causal\")(input1)\nx=BatchNormalization()(x)\nx=Conv1D(32, 3, activation='relu', padding=\"causal\")(x)\nx=BatchNormalization()(x)\nx=Conv1D(64, 3, activation='relu', padding=\"causal\")(x)\nx=BatchNormalization()(x)\nx=Conv1D(128, 3, activation='relu', padding=\"causal\")(x)\nx=BatchNormalization()(x)\nx=Conv1D(256, 3, activation='relu', padding=\"causal\")(x)\nx=BatchNormalization()(x)\nx=Conv1D(512, 3, activation='relu', padding=\"causal\")(x)\nx=BatchNormalization()(x)\nx=GlobalMaxPool1D()(x)\nx=Dropout(0.25)(x)\nx=Dense(512,activation='relu')(x)\nx=Dense(256,activation='relu')(x)\nx=Dense(128,activation='relu')(x)\nx=Dense(64,activation='relu')(x)\nx=Dropout(0.25)(x)\n\noutput_48=[]\nfinalInput = keras.layers.concatenate([x,storefeaturesInfo_x, next48additionalInfo_x])\nfor i in range(48):\n    target=Dense(1, activation=\"relu\", name=str(i))(finalInput)\n    output_48.append(target)\n\nmodel = Model([input1,storefeaturesInfo,next48additionalInfo], output_48)\nmodel.compile(optimizer=adam,loss=mse,loss_weights=z)\n<\/code><\/pre>\n\n<p>I am trying to forecast sales as in this challenge (<a href=\"https:\/\/www.kaggle.com\/c\/rossmann-store-sales\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/rossmann-store-sales<\/a>). The first Input, storefeaturesInfo consists of a 21 valued vector for a total of 1100 stores shaped: (1100,21), second layer describes the additional information for the model for the days that it is trying to forecast for such as whether the store is open or closed. I concatenate all this input and stack 48 dense layers as that is the number of predictions I need for future. I use 48 different layers because I weight them differently. <\/p>\n\n<p>I train the model using a sliding window mechanism, given the past 12 weeks of data predict for the next 48 days. I slide this window for a total time length of 900 days. However, the model fails very badly at capturing any information at all, performing just barely better than predicting the historical average. The learning curves look like this: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UvDiH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UvDiH.png\" alt=\"Trained with different learning rates\"><\/a><\/p>\n\n<p>What these curves show are the validation error for the second last 48 points, the goal is then to predict the 48 after these and submit the solution to Kaggle. \nI post this question to have a reason why this is failing. <\/p>\n","answers":[]}
+{"QuestionId":52969867,"AnswerCount":1,"Tags":"<python><tensorflow><mat-file>","CreationDate":"2018-10-24T13:08:06.067","AcceptedAnswerId":52989446.0,"OwnerUserId":10354279.0,"Title":"make tensorflow dataset from huge number of images(*.jpg) and labels(*.mat)","Body":"<p>I have a huge number of images with their labels (.mat) file (cannot use <code>tf.data.Dataset.from_tensor_slices()<\/code>) and I want to use <code>tf.data<\/code> API to make a tensorflow dataset out of it.<\/p>\n\n<p>As I read in the documentation, I can use <code>tf.data.TextLineDataset<\/code> for large number of data(I have to have a txt file with the address of all the images and send the path of the txt file as <code>tf.data.TextLineDataset<\/code> argument).\nThen, I can use <code>map<\/code> method to read txt file (<code>tf.read_file<\/code>) decode jpg image (<code>tf.image.decode_jpeg<\/code>) and do some basic transformation on the image.<\/p>\n\n<p>However, I cannot use <code>scipy.io.loadmat<\/code> in any part of <code>map<\/code> method because I have no string indicating the path to the mat file. All I have is <code>tf.Tensor<\/code>.<\/p>\n\n<p>I don't think that reading all images and making a TFRecord out of it is that much efficient in this case because then I am basically doing every thing two times. Once, reading the whole images and making TFRecord, and once again, reading TFRecord to make tensorflow dataset.<\/p>\n\n<p>Any idea how I can resolve this issue?<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>dataset = tf.data.TextLineDataset(txt_file).map(read_img_and_mat)\n<\/code><\/pre>\n\n<p>and then:<\/p>\n\n<pre><code>def read_img_and_mat(path):\n    image_string = tf.read_file(path)\n    image_decoded = tf.image.decode_jpeg(image_string, channels=3)\n    label = ... # get label from mat file\n    return image_decoded, label\n<\/code><\/pre>\n","answers":[{"AnswerId":"52989446","CreationDate":"2018-10-25T12:36:29.750","ParentId":null,"OwnerUserId":"10354279","Title":null,"Body":"<p>I found a way to do it using <code>tf.data.from_generator<\/code>\nThe trick I found was to make two separate Dataset (one for mat file and one for the jpg file) and then to combine them using <code>tf.data.Dataset.zip<\/code><\/p>\n\n<p>Here is how it works:<\/p>\n\n<pre><code>mat_dataset = tf.data.Dataset.from_generator(read_mat_file, tf.int64)\n\ndef read_mat_file():\n    while True:\n        with open('mat_addresses.txt', 'r') as input_:\n            for line in input_:\n                # open file and extract info from it as np.array\n                yield tuple(label)  # why tuple? https:\/\/github.com\/tensorflow\/tensorflow\/issues\/13101\n<\/code><\/pre>\n\n<p>in order to get the next batch one just have to do:<\/p>\n\n<pre><code>iter = mat_dataset.make_one_shot_iterator()\nsess.run(iter.get_next())\n<\/code><\/pre>\n\n<p>however, one can make <code>img_dataset<\/code> and combine it with <code>mat_dataset<\/code> like this:<\/p>\n\n<pre><code>img_dataset = tf.data.TextLineDataset('img_addresses.txt').map(read_img)\n\ndef read_img(path):\n    image_string = tf.read_file(path)\n    image_decoded = tf.image.decode_jpeg(image_string, channels=3)\n    return image_decoded\n\ndataset = tf.data.Dataset.zip((mat_dataset, img_dataset))\n<\/code><\/pre>\n\n<p>and now, getting next next batch like mentioned.<\/p>\n\n<p>PS. I have no idea about how efficient the code works in comparison to <code>feed_dict<\/code><\/p>\n"}]}
+{"QuestionId":52969995,"AnswerCount":1,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-10-24T13:15:38.650","AcceptedAnswerId":52989641.0,"OwnerUserId":10552107.0,"Title":"usage about rnn.pack_padded_sequence","Body":"<p>When I use <code>torch.nn.utils.rnn.pack_padded_sequence()<\/code>, an error occurred.\nHere is my code:<\/p>\n\n<pre><code>import torch\nimport numpy as np\nx = torch.from_numpy(np.array([[1,2,3,4,5,6,0,0],[6,7,8,9,0,0,0,0],[12,83,84,0,0,0,0,0]]))\nlength =[6,4,3]\nprint(torch.nn.utils.rnn.pack_padded_sequence(input=x, lengths=length, batch_first=True))\n<\/code><\/pre>\n\n<p>The error is as follows\uff1a<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/pc\/PycharmProjects\/padded\/padded.py\", line 112, in &lt;module&gt;\n    print(torch.nn.utils.rnn.pack_padded_sequence(input=x, lengths=length, batch_first=True))\n  File \"\/home\/pc\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/__init__.py\", line 67, in wrapper\n    if not might_trace(args):\n  File \"\/home\/pc\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/__init__.py\", line 141, in might_trace\n    first_arg = args[0]\nIndexError: tuple index out of range\n<\/code><\/pre>\n\n<p>But when I enter the following code:<\/p>\n\n<pre><code>import torch\nimport numpy as np\nx = torch.from_numpy(np.array([[1,2,3,4,5,6,0,0],[6,7,8,9,0,0,0,0],[12,83,84,0,0,0,0,0]]))\nlength =[6,4,3]\nprint(torch.nn.utils.rnn.pack_padded_sequence(x, lengths=length, batch_first=True))\n<\/code><\/pre>\n\n<p>The result is normal. And I don't know why.<\/p>\n\n<p>Can you help me solve the problem?<\/p>\n","answers":[{"AnswerId":"52989641","CreationDate":"2018-10-25T12:47:05.840","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>So the reason for this sounds incredibly stupid, but I think it is due to the way how the wrapper\/low-level translator function operates on PyTorch functions:<\/p>\n\n<p>From what I understand about the <code>*args, **kwargs<\/code> decorators of python (see more <a href=\"https:\/\/docs.python.org\/dev\/tutorial\/controlflow.html#more-on-defining-functions\" rel=\"nofollow noreferrer\">here<\/a>, the problem is that you only have ordered arguments if you pass them <em>without the <code>=<\/code><\/em>.<\/p>\n\n<p>Meaning, they get stored in a dictionary of keyword\/value pairs. Instead, if we look at the relevant part to our error message (this part: <code>File \"\/home\/pc\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/__init__.py\"<\/code>), we can see the following:<\/p>\n\n<pre><code>first_arg = args[0]\n<\/code><\/pre>\n\n<p>meaning that it wants to address this with an index. Unfortunately, since the keyword is only stored and passed down in the key\/value dictionary, we cannot address this with the index, and subsequently the error is thrown.<\/p>\n\n<p>I have so far not been able to figure out a way around this, because this is very abstract and (potentially) used by many different classes, specifically since it gets passed down through the intermediate layer (in line 67, see stack trace).<\/p>\n"}]}
+{"QuestionId":52970934,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><jupyter-notebook><artificial-intelligence>","CreationDate":"2018-10-24T14:00:03.500","AcceptedAnswerId":52971120.0,"OwnerUserId":10090361.0,"Title":"Value error with tensorflow while loading csv with iris example","Body":"<p>When trying to run a simple get data sequence on Jupyter, for a system to recognise the iris flower tyoes teough fisher's table,  the error:<\/p>\n\n<pre><code>  ValueError                                Traceback (most recent call last)\n &lt;ipython-input-12-269564554b65&gt; in &lt;module&gt;\n 10 training_set = base.load_csv_with_header(filename=IRIS_TRAINING,\n 11                                      features_dtype=np.float32,\n ---&gt; 12                                      target_dtype=np.float32)\n 13 test_set = base.load_csv_with_header(filename=IRIS_TEST,\n 14                                  features_dtype=np.float32,\n\n \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/base.py in load_csv_with_header(filename, target_dtype, features_dtype, target_column)\n 46     data_file = csv.reader(csv_file)\n 47     header = next(data_file)\n ---&gt; 48     n_samples = int(header[0])\n 49     n_features = int(header[1])\n 50     data = np.zeros((n_samples, n_features), dtype=features_dtype)\n\nValueError: invalid literal for int() with base 10: '5.1'\n<\/code><\/pre>\n\n<p>is being shown. The error indicates that it is unable to use the int() function despite the fact that there is no int in the code at all. Here is the code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.contrib.learn.python.learn.datasets import base\n\n# Data files\nIRIS_TRAINING = \"iris_training.csv\"\nIRIS_TEST = \"iris_test.csv\"\n\n# Load datasets.\ntraining_set = base.load_csv_with_header(filename=IRIS_TRAINING,\n                                     features_dtype=np.float32,\n                                     target_dtype=np.float32)\ntest_set = base.load_csv_with_header(filename=IRIS_TEST,\n                                 features_dtype=np.float32,\n                                 target_dtype=np.float32)\n\nprint(training_set.data)\n\nprint(training_set.target)\n<\/code><\/pre>\n\n<p>Why is <code>target_dytype=np.int<\/code> not working, as the error shows?\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"52971120","CreationDate":"2018-10-24T14:08:38.750","ParentId":null,"OwnerUserId":"3288092","Title":null,"Body":"<p>The answer is right there, you are not able to convert a decimal to an integer. Try using the numpy data type <code>float32<\/code> instead.<\/p>\n\n<pre><code>target_dtype=np.float32\n<\/code><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p><code>tensorflow.base<\/code> has several <code>load_csv..<\/code>, you can try either <code>base.load_csv_without_header<\/code> or <code>base.load_csv<\/code>.<\/p>\n\n<p>The specific <code>iris<\/code> tensorflow example is applied on to their own data set, <a href=\"http:\/\/download.tensorflow.org\/data\/iris_test.csv\" rel=\"nofollow noreferrer\">here<\/a> where the first column header stores the number of examples, which caused the error below:<\/p>\n\n<pre><code> 46     data_file = csv.reader(csv_file)\n 47     header = next(data_file)\n ---&gt; 48     n_samples = int(header[0])\n 49     n_features = int(header[1])\n<\/code><\/pre>\n\n<p><code>header<\/code> returns you the first line of the csv file, and <code>n_samples<\/code> is used to store the number of samples from the first column.<\/p>\n"}]}
+{"QuestionId":52971599,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-10-24T14:31:38.160","AcceptedAnswerId":null,"OwnerUserId":1410665.0,"Title":"How to create a custom layer in TensorFlow version 0.12?","Body":"<p>I'm very new in TensorFlow and using a very old version of TensorFlow (Version 0.12.1). The original TensorFlow code was developed in Version 0.12.1 so now when I'm trying to run the code in upgraded version (e.g., 1.10, 1.11), I'm facing some issues. Hence, I'm planning to work on the older version only. The older version doesn't support Keras\/ Eager and some other operations. My aim is to develop a custom layer for thresholding operation. <\/p>\n\n<p>I shall be highly obliged if any of you kindly guide me how can I start in version 0.12.1.\nThe links which I already read <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">Eager<\/a><a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_layers\" rel=\"nofollow noreferrer\">TensorFlow<\/a><\/p>\n","answers":[{"AnswerId":"53129507","CreationDate":"2018-11-03T07:57:03.760","ParentId":null,"OwnerUserId":"1410665","Title":null,"Body":"<p>Some parameters of version 0.12 need to change. Those are dims to axis, mul to multiply and sub to subtract. Rest the model should work<\/p>\n"}]}
+{"QuestionId":52971972,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-24T14:49:15.773","AcceptedAnswerId":null,"OwnerUserId":9421400.0,"Title":"How to implement a weighted mean squared error function in Keras","Body":"<p>I am defining a weighted mean squared error in Keras as follows: <\/p>\n\n<pre><code>def weighted_mse(yTrue,yPred):\n    data_weights = [w0,w1,w2,w3]\n    data_weights_np = np.asarray(data_weights, np.float32)\n    weights = tf.convert_to_tensor(data_weights_np, np.float32)\n    return K.mean(weights*K.square(yTrue-yPred))\n<\/code><\/pre>\n\n<p>I have a list of weights for each prediction. The predictions are of shape for example: (25,4). That is generated via final dense layer with dimension 4. I wish to weights these prediction in the mean squared error, so I generate a tensor and multiply it with the sum of squares error. Is this the correct way to do so? \nBecause, when I print the shape of the tensor, using tf.shape for YTrue and YPred it shows:\n<code>Tensor(\"loss_19\/dense_20_loss\/Shape:0\", shape=(3,), dtype=int32)<\/code><\/p>\n\n<p>and for weights: <\/p>\n\n<pre><code>Tensor(\"loss_19\/dense_20_loss\/Shape_2:0\", shape=(1,), dtype=int32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52972521","CreationDate":"2018-10-24T15:14:50.410","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>The Keras API already provides a mechanism to provide weights, for example the <code>model.fit<\/code> function. From the <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>class_weight:<\/strong> Optional dictionary mapping class indices (integers) to a weight (float) value, used for weighting the loss function (during training only). This can be useful to tell the model to \"pay more attention\" to samples from an under-represented class.<\/p>\n  \n  <p><strong>sample_weight:<\/strong> Optional Numpy array of weights for the training samples, used for weighting the loss function (during training only). You can either pass a flat (1D) Numpy array with the same length as the input samples (1:1 mapping between weights and samples), or in the case of temporal data, you can pass a 2D array with shape  (samples, sequence_length), to apply a different weight to every timestep of every sample. In this case you should make sure to specify sample_weight_mode=\"temporal\" in compile().<\/p>\n<\/blockquote>\n\n<p>If you have a weight for each sample, you can pass the NumPy array as <code>sample_weight<\/code> to achieve the same effect without writing your own loss function.<\/p>\n"}]}
+{"QuestionId":52972386,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-24T15:08:40.807","AcceptedAnswerId":null,"OwnerUserId":9769779.0,"Title":"how to implement the variable array with one and zero in tensorflow","Body":"<p>I'm totally new on tensorflow, and I just want to implement a kind of selection function by using matrices multiplication.<\/p>\n\n<p>example below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#input:\nI = [[9.6, 4.1, 3.2]]\n\n#selection:(single \"1\" value , and the other are \"0s\")\ns = tf.transpose(tf.Variable([[a, b, c]]))\ne.g.  s could be [[0, 1, 0]] or [[0, 0, 1]] or [[1, 0, 0]]\n\n#result:(multiplication)\no = tf.matul(I, s)\n<\/code><\/pre>\n\n<p>sorry for the poor expression, \n    I intend to find the 'solution'  in distribution functions with different means and sigmas. (value range from 0 to 1). \nso now, i have three variable  i, j, index. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>value1 = np.exp(-((index - m1[i]) ** 2.) \/ s1[i]** 2.)\nvalue2 = np.exp(-((index - m2[j]) ** 2.) \/ s2[j]** 2.)\nm1 = [1, 3, 5]  s = [0.2, 0.4, 0.5]. #first graph\nm2 = [3, 5, 7]. s = [0.5, 0.5, 1.0]. #second graph\n<\/code><\/pre>\n\n<p>I want to get the max or optimization of total value<\/p>\n\n<p>e.g. value1 + value2 = 1+1 = 2  and one of the solutions: i = 2, j=1, index=5<\/p>\n\n<p>or I could do this in the other module?<\/p>\n","answers":[]}
+{"QuestionId":52973241,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><bigdata>","CreationDate":"2018-10-24T15:51:42.100","AcceptedAnswerId":null,"OwnerUserId":1786577.0,"Title":"Consuming big data by Tensorflow","Body":"<p>Suppose, I have a huge list of objects, each of which could be a list of numpy arrays, for example. <\/p>\n\n<p>What\u2019s the best way to pass this dataset to tensorflow? <\/p>\n\n<p>I want to be able to randomly shuffle the data and form batches. May be it\u2019s worth to shuffle the dataset and form batches using standard python(numpy) procedures and after that use something like <code>tf.data.Dataset.from_generator()<\/code>? <\/p>\n\n<p>Straightforward approach of transforming full dataset to <code>tf.Tensor<\/code> seems to be useless due to size limit for the <code>tf.GraphDef<\/code> protocol buffer(according to the Tensorflow documentation).<\/p>\n","answers":[{"AnswerId":"53001537","CreationDate":"2018-10-26T04:40:13.293","ParentId":null,"OwnerUserId":"2082009","Title":null,"Body":"<p>Looks like your data is large but still small enough to fit in memory? If so, you are on the right track with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#from_generator\" rel=\"nofollow noreferrer\">tf.data.Dataset.from_generator()<\/a>. You could then shuffle and batch with something like<\/p>\n\n<pre><code>import itertools\n\n# your data\ndata = range(1024)\ndef gen():\n  for item in data:\n    yield data\n\nds = Dataset.from_generator(\n    gen, tf.int64, tf.TensorShape([])).shuffle(buffer_size=128).batch(batch_size=4)\nvalue = ds.make_one_shot_iterator().get_next()\n\nsess.run(value)  # array([0, 1, 2, 3])\n<\/code><\/pre>\n\n<p>Alternatively, you could dump your data to a TFRecord file and read from it using the TFRecordDataset. This <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/bf1efe8d5b4722e5a19ace712301c92d40d7abdb\/tensorflow\/python\/data\/kernel_tests\/reader_dataset_ops_test.py#L625\" rel=\"nofollow noreferrer\">test<\/a> should help you get started.<\/p>\n"}]}
+{"QuestionId":52973758,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><caffe2><onnx>","CreationDate":"2018-10-24T16:19:49.300","AcceptedAnswerId":null,"OwnerUserId":10081913.0,"Title":"Is there any way to convert pretrained model from PyTorch to ONNX?","Body":"<p>I trained StarGAN model on my custom dataset.\nAnd I need to convert this model from .pth(Pytorch) to .pb for using on Android studio.\nI searched a lot and I found some ways for conversion. \nHowever all solutions don't work on my case.<\/p>\n\n<p>I tried on small network that includes only one nn.Linear layer.\nOn this network, solutions work very well!<\/p>\n\n<p>I think, my network includes Conv2D layer and MaxPooling2D layer so conversion processing doesn't work.<\/p>\n\n<p>First, this is my network(StarGAN).<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport numpy as np\n\n\nclass ResidualBlock(nn.Module):\n    def __init__(self, dim_in, dim_out):\n        super(ResidualBlock, self).__init__()\n        self.main = nn.Sequential(\n            nn.Conv2d(dim_in, dim_out, kernel_size=3, stride=1, padding=1, bias=False),\n            nn.InstanceNorm2d(dim_out, affine=True, track_running_stats=True),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(dim_out, dim_out, kernel_size=3, stride=1, padding=1, bias=False),\n            nn.InstanceNorm2d(dim_out, affine=True, track_running_stats=True))\n\n    def forward(self, x):\n        return x + self.main(x)\n\n\nclass Generator(nn.Module):\n    def __init__(self, conv_dim=64, c_dim=5, repeat_num=6):\n        super(Generator, self).__init__()\n\n        layers = []\n        layers.append(nn.Conv2d(3 + c_dim, conv_dim, kernel_size=7, stride=1, padding=3, bias=False))\n        layers.append(nn.InstanceNorm2d(conv_dim, affine=True, track_running_stats=True))\n        layers.append(nn.ReLU(inplace=True))\n\n        curr_dim = conv_dim\n        for _ in range(2):\n            layers.append(nn.Conv2d(curr_dim, curr_dim * 2, kernel_size=4, stride=2, padding=1, bias=False))\n            layers.append(nn.InstanceNorm2d(curr_dim * 2, affine=True, track_running_stats=True))\n            layers.append(nn.ReLU(inplace=True))\n            curr_dim = curr_dim * 2\n\n        for _ in range(repeat_num):\n            layers.append(ResidualBlock(dim_in=curr_dim, dim_out=curr_dim))\n\n        for _ in range(2):\n            layers.append(nn.ConvTranspose2d(curr_dim, curr_dim \/\/ 2, kernel_size=4, stride=2, padding=1, bias=False))\n            layers.append(nn.InstanceNorm2d(curr_dim \/\/ 2, affine=True, track_running_stats=True))\n            layers.append(nn.ReLU(inplace=True))\n            curr_dim = curr_dim \/\/ 2\n\n        layers.append(nn.Conv2d(curr_dim, 3, kernel_size=7, stride=1, padding=3, bias=False))\n        layers.append(nn.Tanh())\n        self.main = nn.Sequential(*layers)\n\n    def forward(self, x, c):\n        c = c.view(c.size(0), c.size(1), 1, 1)\n        c = c.repeat(1, 1, x.size(2), x.size(3))\n        x = torch.cat([x, c], dim=1)\n        return self.main(x)\n\n\nclass Discriminator(nn.Module):\n    def __init__(self, image_size=128, conv_dim=64, c_dim=5, repeat_num=6):\n        super(Discriminator, self).__init__()\n        layers = []\n        layers.append(nn.Conv2d(3, conv_dim, kernel_size=4, stride=2, padding=1))\n        layers.append(nn.LeakyReLU(0.01))\n\n        curr_dim = conv_dim\n        for _ in range(1, repeat_num):\n            layers.append(nn.Conv2d(curr_dim, curr_dim * 2, kernel_size=4, stride=2, padding=1))\n            layers.append(nn.LeakyReLU(0.01))\n            curr_dim = curr_dim * 2\n\n        kernel_size = int(image_size \/ np.power(2, repeat_num))\n        self.main = nn.Sequential(*layers)\n        self.conv1 = nn.Conv2d(curr_dim, 1, kernel_size=3, stride=1, padding=1, bias=False)\n        self.conv2 = nn.Conv2d(curr_dim, c_dim, kernel_size=kernel_size, bias=False)\n\n    def forward(self, x):\n        h = self.main(x)\n        out_src = self.conv1(h)\n        out_cls = self.conv2(h)\n        return out_src, out_cls.view(out_cls.size(0), out_cls.size(1))\n<\/code><\/pre>\n\n<p>And this is the error message.<\/p>\n\n<pre><code>TypeError: object of type 'torch._C.Value' has no len() (occurred when translating repeat)\n<\/code><\/pre>\n\n<p>Is there any way for conversion? Help me.<\/p>\n","answers":[{"AnswerId":"53008279","CreationDate":"2018-10-26T12:00:15.340","ParentId":null,"OwnerUserId":"9316652","Title":null,"Body":"<p>I have the same issue when trying to produce a graph of my model, using TensorboardX.<\/p>\n\n<p>I believe the error comes from what operators <code>torch.onnx<\/code> currently supports. You can check this link:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/onnx.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/onnx.html<\/a><br>\nUnder section <strong>Supported operators<\/strong>, you will see that <code>repeat<\/code> is not listed.<\/p>\n\n<p>To answer your question, it seems that you currently cannot convert a model using <code>repeat<\/code> with <code>torch.onnx<\/code>.<\/p>\n"}]}
+{"QuestionId":52973928,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-24T16:30:21.347","AcceptedAnswerId":null,"OwnerUserId":10531092.0,"Title":"Cannot Import TensorFlow on Windows 7 Professional","Body":"<p>I've tried using different versions of TensorFlow, I've tried setting up different paths but no matter what I do I get the following error.\nI'm using Spyder v3.2.8, Python v3.6.6, and TensorFlow v1.11.0 on Windows 7 Professional. <\/p>\n\n<p>import keras\nUsing TensorFlow backend.\nTraceback (most recent call last):<\/p>\n\n<pre><code>File \"&lt;ipython-input-1-88d96843a926&gt;\", line 1, in &lt;module&gt;\nimport keras\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\nfrom . import utils\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\nfrom . import conv_utils\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\nfrom .. import backend as K\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\keras\\backend\\__init__.py\", line 89, in &lt;module&gt;\nfrom .tensorflow_backend import *\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 5, in &lt;module&gt;\nimport tensorflow as tf\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow\n\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\nraise ImportError(msg)\n\nImportError: Traceback (most recent call last):\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n_pywrap_tensorflow_internal = swig_import_helper()\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\Lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\lib\\imp.py\", line 243, in load_module\nreturn load_dynamic(name, filename, file)\nFile \"C:\\Users\\dchisholm\\AppData\\Local\\Continuum\\anaconda3.3.7\\envs\\Python35\\lib\\imp.py\", line 343, in load_dynamic\nreturn _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n<\/code><\/pre>\n","answers":[{"AnswerId":"52995081","CreationDate":"2018-10-25T17:34:36.673","ParentId":null,"OwnerUserId":"10531092","Title":null,"Body":"<p>I finally figured out what the problem was. The python version needed to be 3.5.x and all of the TensorFlow packages needed to be 1.8.0. Other version variations I tried resulted in import failures.<\/p>\n\n<p>Using these in conjunction, keras modules were able to be imported and I was able to set up an ANN without issue.<\/p>\n"}]}
+{"QuestionId":52973989,"AnswerCount":1,"Tags":"<tensorflow><optimization>","CreationDate":"2018-10-24T16:33:42.043","AcceptedAnswerId":52974609.0,"OwnerUserId":7886651.0,"Title":"Using different optimizers to train the same layer in tensorflow","Body":"<p>I have a model which consists of convolutional layers followed by fully connected layers. I trained this model on the fer dataset. This is considered a classification problem where the number of output is equal to 8. <\/p>\n\n<p>After training this model, I kept the fully connected layer, and replaced only the last layer with a new one that has 3 outputs. Therefore, the purpose was to fine tune the fully connected layers along with training the output layer. <\/p>\n\n<p>Therefore, I have used an optimizer at the beginning to train the whole model. Then I created a new optimizer to fine tune the fully connected layer along with training the last layer. <\/p>\n\n<p>As a result, I got the following error:<\/p>\n\n<pre><code>ValueError: Variable Dense\/dense\/bias\/Adam\/ already exists,\n<\/code><\/pre>\n\n<p>I know the reason for getting error. The second optimizer was trying to create a kernel for updating the weights using the same name; because a kernel with the same name was created by the first optimizer. <\/p>\n\n<p>Hence, I would like to know how to fix this problem. Is there a way to delete the kernels associated with the first optimizer?<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[{"AnswerId":"52974609","CreationDate":"2018-10-24T17:13:21.257","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>This is probably caused by both optimizers using the (same) default name <code>'Adam'<\/code>. To avoid this clash, you can give the second optimizer a different name, e.g.<\/p>\n\n<pre><code>opt_finetune = tf.train.AdamOptimizer(name='Adam_finetune')\n<\/code><\/pre>\n\n<p>This should make <code>opt_finetune<\/code> create its variables under different names. Please let us know whether this works!<\/p>\n"}]}
+{"QuestionId":52974196,"AnswerCount":2,"Tags":"<python><neural-network><caffe><pycaffe>","CreationDate":"2018-10-24T16:45:48.840","AcceptedAnswerId":null,"OwnerUserId":1714692.0,"Title":"How to interpret the file mean.binaryproto when loading a Neural Network?","Body":"<p>I want to load a Neural Network that has been trained with caffe for image classification.<\/p>\n\n<p>The NN contains a file <code>mean.binaryproto<\/code> which has the means to be subtracted before inputting an image to be classified.<\/p>\n\n<p>I am trying to understand what is contained in this file so I used Google Colab to see what is inside it.<\/p>\n\n<p>The code to load it is the following:<\/p>\n\n<pre><code># Load the Drive helper and mount\nfrom google.colab import drive\n\n# This will prompt for authorization.\ndrive.mount('\/content\/drive')\n!ls \"\/content\/drive\/My Drive\"\n\n#install packages\n!apt install -y caffe-cuda\n!apt update\n!apt upgrade\n!apt dist-upgrade\n!ls \"\/content\/drive\/My Drive\/NeuralNetwork\/CNRPark-Trained-Models\/mAlexNet-on-CNRPark\/\"\nimport caffe\nimport numpy as np\nwith open('\/content\/drive\/My Drive\/NeuralNetwork\/CNRPark-Trained-Models\/mAlexNet-on-CNRPark\/mean.binaryproto', 'rb') as f:\n    blob = caffe.proto.caffe_pb2.BlobProto()\n    blob.ParseFromString(f.read())\n    arr = np.array( caffe.io.blobproto_to_array(blob) )\n    print(arr.shape)\n    out = arr[0]\n    data = np.array(blob.data).reshape([blob.channels, blob.height, blob.width])\n    print (data.shape)\n    print(data[0])\n #display the mean image\n from PIL import Image\n from IPython.display import Image as Im, display\n display(Image.fromarray(data[0], 'RGB'))\n<\/code><\/pre>\n\n<p>which outputs:<\/p>\n\n<pre><code>(1, 3, 256, 256)\n(3, 256, 256)\n<\/code><\/pre>\n\n<p>What I have understood is that the file contain the means and the images we are talking about are 3 channel images so there is a mean for each channel.<\/p>\n\n<p>However I was expecting a single value per channel instead I found a 256x256 array: does it mean that the took a mean on each pixel of each channel?<\/p>\n\n<p>Another question is the following: I want to use such NN with OpenCV which instead of RGB uses BGR: How to know if the mean 3x256x256 uses RGB or BGR?<\/p>\n\n<p>The link to the model is <a href=\"http:\/\/www.cnrpark.it\" rel=\"nofollow noreferrer\">this<\/a>. The model I am looking at is contained in the zip file <code>CNRPark-Trained-Models.zip<\/code> within the folder: <code>mAlexNet-on-CNRPark<\/code>.<\/p>\n","answers":[{"AnswerId":"53198053","CreationDate":"2018-11-07T21:25:01.287","ParentId":null,"OwnerUserId":"6327658","Title":null,"Body":"<blockquote>\n  <p>However I was expecting a single value per channel instead I found a\n  256x256 array: does it mean that the took a mean on each pixel of each\n  channel?<\/p>\n<\/blockquote>\n\n<p>Exactly. According to the shape of <code>mean.binaryproto<\/code>, this file is the average image of some dataset, which means that it took the mean of each pixel (feature) for each channel.<\/p>\n\n<p>This should not be confused with the mean pixel, which, as you stated, is a single value for each channel. <\/p>\n\n<p>For example, mean pixel was adoped by <a href=\"https:\/\/arxiv.org\/pdf\/1409.1556.pdf\" rel=\"nofollow noreferrer\">Very Deep Convolutional Networks for Large-Scale Image Recognition<\/a>. According to their paper:<\/p>\n\n<blockquote>\n  <p>The only pre-processing we do is subtracting the mean RGB value,\n  computed on the training set, from each pixel<\/p>\n<\/blockquote>\n\n<p>In other words, if you consider an RGB image to be 3 feature arrays of size N x N, the average image will be the mean of <strong>each<\/strong> feature and the mean pixel will be the mean of <strong>all<\/strong> features.<\/p>\n\n<hr>\n\n<blockquote>\n  <p>Another question is the following: I want to use such NN with OpenCV\n  which instead of RGB uses BGR: How to know if the mean 3x256x256 uses\n  RGB or BGR?<\/p>\n<\/blockquote>\n\n<p>I doubt the binary file you are reading stores any information about its color format, but a practical way to figure out is to plot this image using <code>matplotlib<\/code> and see if the colors make sense. <\/p>\n\n<p>For example, face images. If red and blue channels are swapped the skin tone will look blueish.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8agIV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8agIV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In fact, the image above is an example of average image (face images) :)<\/p>\n\n<p>You could also assume it is BGR since OpenCV uses this color format.<\/p>\n\n<p>However, the correct way to find out how this <code>mean.binaryproto<\/code> was generated is by looking at their repositories or by asking the owner of the model.<\/p>\n"},{"AnswerId":"55376690","CreationDate":"2019-03-27T11:59:47.337","ParentId":null,"OwnerUserId":"5660092","Title":null,"Body":"<pre><code>import os, sys, glob, caffe\nimport numpy as np\nmean_file= \"path\/to\/file\/mean.binaryproto\"\n#convert mean file to image\nblob= caffe.proto.caffe_pb2.BlobProto()\ntry:\n    data = open( mean_file, 'rb' ).read()\nexcept:\n    data = open( mean_file, 'r' ).read()\nblob.ParseFromString(data)\narr = np.uint8(np.array( caffe.io.blobproto_to_array(blob) )[0])\n#a= arr[0];    b= arr[1];    c= arr[2]\nimg= np.zeros([128,200,3])\nimg[:,:,0]= arr[0];    img[:,:,1]= arr[1];    img[:,:,2]= arr[2]\nimport cv2\ncv2.imwrite(mean_file.replace(\".binaryproto\", \".bmp\"), img)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52975301,"AnswerCount":1,"Tags":"<docker><tensorflow><prometheus><tensorflow-serving>","CreationDate":"2018-10-24T17:58:26.060","AcceptedAnswerId":null,"OwnerUserId":10553653.0,"Title":"Prometheus for Tensorflow Serving","Body":"<p>What are the steps to use the Prometheus exporter for Tensorflow serving?\nAccording to 1.11 TF serving supports prometheus metrics:\n<a href=\"https:\/\/github.com\/tensorflow\/serving\/releases\/tag\/1.11.0\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/releases\/tag\/1.11.0<\/a><\/p>\n\n<p>I'm starting a docker from the example <a href=\"https:\/\/www.tensorflow.org\/serving\/docker\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/serving\/docker<\/a> and the following:<\/p>\n\n<p>docker run -p 8501:8501 -p 8500:8500 \\\n--mount type=bind,\\\nsource=\/tmp\/tfserving\/serving\/tensorflow_serving\/servables\/tensorflow\/testdata\/saved_model_half_plus_two_cpu,\\\ntarget=\/models\/half_plus_two \\\n-e MODEL_NAME=half_plus_two -t tensorflow\/serving &amp;<\/p>\n\n<p>Prometheus configuration file:\nglobal:\n  scrape_interval:     10s\n  evaluation_interval: 10s\n  external_labels:\n    monitor: 'tf-serving-monitor'<\/p>\n\n<p>scrape_configs:\n  - job_name: 'tensorflow'\n    scrape_interval: 5s\n    static_configs:\n      - targets: ['localhost:8501']<\/p>\n\n<p>But prometheus fails to find the metrics exposed by tf serving.\nIs there a specific port I should open on docker or some parameter I should pass to TF serving?<\/p>\n","answers":[{"AnswerId":"52984664","CreationDate":"2018-10-25T08:25:27.450","ParentId":null,"OwnerUserId":"8657904","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/commit\/021efbd3281aa815cab0b35eab6d6d25249c12d4\" rel=\"nofollow noreferrer\">According to the release notes you linked to<\/a> TensorFlow exports Prometheus metrics at <code>\/monitoring\/prometheus\/metrics<\/code> (as opposed to Prometheus' default <code>\/metrics<\/code>). So at the very least you need to add a <code>metrics_path<\/code> to your config:<\/p>\n\n<pre><code>scrape_configs:\n  - job_name: 'tensorflow'\n    scrape_interval: 5s\n    metrics_path: '\/monitoring\/prometheus\/metrics'\n    static_configs:\n      - targets: ['localhost:8501']\n<\/code><\/pre>\n\n<p>But first make sure you can see the metrics exported at <a href=\"http:\/\/localhost:8501\/monitoring\/prometheus\/metrics\" rel=\"nofollow noreferrer\">http:\/\/localhost:8501\/monitoring\/prometheus\/metrics<\/a> in your browser. Else poke around (with your browser) until you find the proper URL (and reflect that in your config).<\/p>\n"}]}
+{"QuestionId":52975711,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-24T18:25:18.170","AcceptedAnswerId":null,"OwnerUserId":1324967.0,"Title":"How to load pre-loaded tensorflow constants into a model","Body":"<p>I do all of my tensorflow data generation AND training on the GPU.  This means that my data sets are represented as tensorflow constants.  For example:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Generate x and y on the GPU\nx = tf.constant([[1], [2], [3], [4]], dtype=tf.float32)\ny_true = tf.constant([[0], [-1], [-2], [-3]], dtype=tf.float32)\n<\/code><\/pre>\n\n<p>This means that when I go to train my models, it looks something like the following code.  I have greatly simplified it so that I can recreate my problem on other computers (based on the code at the <a href=\"https:\/\/www.tensorflow.org\/guide\/low_level_intro\" rel=\"nofollow noreferrer\">bottom of the tensorflow low-level API tutorial<\/a> Where it says \"Complete Program\")<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Declare constants\nx = tf.constant([[1], [2], [3], [4]], dtype=tf.float32)\ny_true = tf.constant([[0], [-1], [-2], [-3]], dtype=tf.float32)\n\n# Create model to train\nlinear_model = tf.layers.Dense(units=1)\ny_pred = linear_model(x) # &lt;---- my model\nloss = tf.losses.mean_squared_error(labels=y_true, predictions=y_pred)\noptimizer = tf.train.AdamOptimizer(.2)\ntrain = optimizer.minimize(loss)\n\n# Boilerplate code\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n\n# Train the model to give the correct y given the x\nfor i in range(100):\n  _, loss_value = sess.run((train, loss))\n  print(loss_value, end=\", \")\n\n# Test our output to make sure it looks good\nprint()\nprint(sess.run(y_pred))\n\n# Generate new data to test or train in the model\nnew_x = tf.constant([[0], [1], [2]], dtype=tf.float32)\nnew_y_true = tf.constant([[10], [11], [12]], dtype=tf.float32)\n\n# Test the model on new_x and see if it looks similar to new_y_true and\/or train based on new_x and new_y_true\n# ??\n<\/code><\/pre>\n\n<p>When I finish training my model (where <code>y_pred<\/code> is declared) I want to reuse that linear model and test it on my <code>new_x<\/code> and <code>new_y_true<\/code> data or even train on it.  How would I do this?<\/p>\n\n<p>I tried replacing <code>x<\/code> and <code>y_true<\/code> with placeholders, but you cannot put <code>tf.Tensor<\/code> objects into the <code>feed_dict<\/code> parameter.  <\/p>\n","answers":[{"AnswerId":"52976676","CreationDate":"2018-10-24T19:34:22.650","ParentId":null,"OwnerUserId":"3344139","Title":null,"Body":"<p>I'm not exactly sure why you want to define all your data in graph, and if you want to do this with more data than given in this toy example, you're going to run into problems. You can't feed a placeholder with a tensor (see <a href=\"https:\/\/stackoverflow.com\/questions\/42560209\/tensorflow-how-to-feed-a-placeholder-variable-with-a-tensor\">Tensorflow: How to feed a placeholder variable with a tensor?<\/a>).<\/p>\n\n<p>The only thing that I can think of that would make this work is to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\">tf.cond<\/a> with a boolean placeholder for indicating whether you want to train or test. Keep in mind that <code>tf.cond<\/code> is slow, and you should read it's documentation because it has some unintuitive behavior regarding which operations will be executed. It then looks like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Declare constants\nx = tf.constant([[1], [2], [3], [4]], dtype=tf.float32)\ny_true = tf.constant([[0], [-1], [-2], [-3]], dtype=tf.float32)\n\nnew_x = tf.constant([[0], [1], [2]], dtype=tf.float32)\nnew_y_true = tf.constant([[10], [11], [12]], dtype=tf.float32)\n\n# Define indicator if you want to train or test\nis_train = tf.placeholder(dtype=tf.bool, shape=[], name=\"is_training\")\n\n\n\n# Define which input data to use depending if you are training or not\ninput_data = tf.cond(pred=is_train,\n                 true_fn=lambda: x,\n                 false_fn=lambda: new_x\n)\n\n# Create model to train\nlinear_model = tf.layers.Dense(units=1)\ny_pred = linear_model(input_data)\n\nloss = tf.losses.mean_squared_error(labels=y_true, predictions=y_pred)\noptimizer = tf.train.AdamOptimizer(.2)\ntrain =  optimizer.minimize(loss)\n\n# Boilerplate code\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n\n# Train the model to give the correct y given the x\nfor i in range(100):\n  _, loss_value = sess.run((train, loss), feed_dict={is_train: True})\n  print(loss_value, end=\", \")\n\n# Test our output to make sure it looks good\nprint()\nprint(sess.run(y_pred, feed_dict={is_train: False}))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52975831,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><classification>","CreationDate":"2018-10-24T18:34:18.840","AcceptedAnswerId":null,"OwnerUserId":10553598.0,"Title":"keras training 2 models simultaneously","Body":"<p>I have 2 models: model1 and model2. <\/p>\n\n<p>I need to take model 1 output and manipulate myData manually and set it (manipulated myData) as input of model2.<\/p>\n\n<p>model2's output is classification of the responses of myData (to model1 output manipulation), relative to predefined classification (i.e. supervised).<\/p>\n\n<ol>\n<li>I need to improve model1 output and improve model2 classification simultaneously. However in testing I will work with each model separately. <\/li>\n<li>To my opinion I need to use model 2 cost function as model 1 cost function - how can it be done? <\/li>\n<li>Any other idea how can it be done? <\/li>\n<\/ol>\n\n<p>I emphasize: Concatenate does NOT solve the problem. <\/p>\n\n<p>Please refer to the <a href=\"https:\/\/i.stack.imgur.com\/hwnBv.png\" rel=\"nofollow noreferrer\">attached diagram<\/a><\/p>\n","answers":[{"AnswerId":"52976174","CreationDate":"2018-10-24T18:58:29.703","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Well, the general sketch would be as follows:<\/p>\n\n<pre><code># define model 1 architecture\n...\n\n# define model 2 architecture\n...\n\n# define manipulation logic\nout1 = model1.output  # get the output of model1\nout1 = SomeLayer()(out1)  # apply any number of layers as you wish\n...\n\nout_final = model2(out1) # feed the manipulated output to model2\n\n# define the joint model\nfinal_model = Model(model1.input, out_final)\n\n# compile the model ...\nfinal_model.compile(loss=..., optimizer=...) # loss is computed based on the output of model2\n\n# fit the model\nfinal_model.fit(...)\n<\/code><\/pre>\n\n<p>This way both <code>model1<\/code> and <code>model2<\/code> will be trained simultaneously and also you can use them independently (e.g. use <code>model1.predict()<\/code> or <code>model2.predict()<\/code>).<\/p>\n"}]}
+{"QuestionId":52975843,"AnswerCount":1,"Tags":"<python><tensorflow><pytorch>","CreationDate":"2018-10-24T18:35:36.483","AcceptedAnswerId":52975844.0,"OwnerUserId":1840086.0,"Title":"Comparing Conv2D with padding between Tensorflow and PyTorch","Body":"<p>I am trying to import weights saved from a Tensorflow model to PyTorch. So far the results have been very similar. I ran into a snag when the model calls for <code>conv2d<\/code> with <code>stride=2<\/code>.<\/p>\n\n<p>To verify the mismatch, I set up a very simple comparison between TF and PyTorch. First, I compare <code>conv2d<\/code> with <code>stride=1<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\n\n\nnp.random.seed(0)\nsess = tf.Session()\n\n# Create random weights and input\nweights = torch.empty(3, 3, 3, 8)\ntorch.nn.init.constant_(weights, 5e-2)\nx = np.random.randn(1, 3, 10, 10)\n\nweights_tf = tf.convert_to_tensor(weights.numpy(), dtype=tf.float32)\n# PyTorch adopts [outputC, inputC, kH, kW]\nweights_torch = torch.Tensor(weights.permute((3, 2, 0, 1)))\n\n# Tensorflow defaults to NHWC\nx_tf = tf.convert_to_tensor(x.transpose((0, 2, 3, 1)), dtype=tf.float32)\nx_torch = torch.Tensor(x)\n\n# TF Conv2D\ntf_conv2d = tf.nn.conv2d(x_tf,\n                         weights_tf,\n                         strides=[1, 1, 1, 1],\n                         padding=\"SAME\")\n\n# PyTorch Conv2D\ntorch_conv2d = F.conv2d(x_torch, weights_torch, padding=1, stride=1)\n\nsess.run(tf.global_variables_initializer())\ntf_result = sess.run(tf_conv2d)\n\ndiff = np.mean(np.abs(tf_result.transpose((0, 3, 1, 2)) - torch_conv2d.detach().numpy()))\nprint('Mean of Abs Diff: {0}'.format(diff))\n<\/code><\/pre>\n\n<p>The result of this execution is:<\/p>\n\n<pre><code>Mean of Abs Diff: 2.0443112092038973e-08\n<\/code><\/pre>\n\n<p>When I change <code>stride<\/code> to 2, the results start to vary.<\/p>\n\n<pre><code># TF Conv2D\ntf_conv2d = tf.nn.conv2d(x_tf,\n                         weights_tf,\n                         strides=[1, 2, 2, 1],\n                         padding=\"SAME\")\n\n# PyTorch Conv2D\ntorch_conv2d = F.conv2d(x_torch, weights_torch, padding=1, stride=2)\n<\/code><\/pre>\n\n<p>The result of this execution is:<\/p>\n\n<pre><code>Mean of Abs Diff: 0.2104552686214447\n<\/code><\/pre>\n\n<p>According to PyTorch documentation, <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=conv2d#torch.nn.functional.conv2d\" rel=\"noreferrer\"><code>conv2d<\/code> uses zero-padding<\/a> defined by the <code>padding<\/code> argument. Thus, zeros are added to the left, top, right, and bottom of the input in my example. <\/p>\n\n<p>If PyTorch simply adds padding on both sides based on the input parameter, it should be easy to replicate in Tensorflow.<\/p>\n\n<pre><code># Manually add padding - consistent with PyTorch\npaddings = tf.constant([[0, 0], [1, 1], [1, 1], [0, 0]])\nx_tf = tf.convert_to_tensor(x.transpose((0, 2, 3, 1)), dtype=tf.float32)\nx_tf = tf.pad(x_tf, paddings, \"CONSTANT\")\n\n# TF Conv2D\ntf_conv2d = tf.nn.conv2d(x_tf,\n                         weights_tf,\n                         strides=[1, 2, 2, 1],\n                         padding=\"VALID\")\n<\/code><\/pre>\n\n<p>The result of this comparison is:<\/p>\n\n<pre><code>Mean of Abs Diff: 1.6035047067930464e-08\n<\/code><\/pre>\n\n<p>What this tells me is that if I am somehow able to replicate the default padding behavior from Tensorflow into PyTorch, then my results will be similar.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/44853243\/tensorflow-conv2d-diffrent-start-index-between-even-and-odd-stride?noredirect=1&amp;lq=1\">This question<\/a> inspected the behavior of padding in Tensorflow. <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/nn#Notes_on_SAME_Convolution_Padding\" rel=\"noreferrer\">TF documentation explains how padding is added for \"SAME\" convolutions.<\/a> I discovered these links while writing this question.<\/p>\n\n<p>Now that I know the padding strategy of Tensorflow, I can implement it in PyTorch.<\/p>\n","answers":[{"AnswerId":"52975844","CreationDate":"2018-10-24T18:35:36.483","ParentId":null,"OwnerUserId":"1840086","Title":null,"Body":"<p>To replicate the behavior, padding sizes are calculated as described in the Tensorflow documentation. Here, I test the padding behavior by setting <code>stride=2<\/code> and padding the PyTorch input.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\n\n\nnp.random.seed(0)\nsess = tf.Session()\n\n# Create random weights and input\nweights = torch.empty(3, 3, 3, 8)\ntorch.nn.init.constant_(weights, 5e-2)\nx = np.random.randn(1, 3, 10, 10)\n\nweights_tf = tf.convert_to_tensor(weights.numpy(), dtype=tf.float32)\nweights_torch = torch.Tensor(weights.permute((3, 2, 0, 1)))\n\n# Tensorflow padding behavior. Assuming that kH == kW to keep this simple.\nstride = 2\nif x.shape[2] % stride == 0:\n    pad = max(weights.shape[0] - stride, 0)\nelse:\n    pad = max(weights.shape[0] - (x.shape[2] % stride), 0)\n\nif pad % 2 == 0:\n    pad_val = pad \/\/ 2\n    padding = (pad_val, pad_val, pad_val, pad_val)\nelse:\n    pad_val_start = pad \/\/ 2\n    pad_val_end = pad - pad_val_start\n    padding = (pad_val_start, pad_val_end, pad_val_start, pad_val_end)\n\nx_tf = tf.convert_to_tensor(x.transpose((0, 2, 3, 1)), dtype=tf.float32)\nx_torch = torch.Tensor(x)\nx_torch = F.pad(x_torch, padding, \"constant\", 0)\n\n# TF Conv2D\ntf_conv2d = tf.nn.conv2d(x_tf,\n                         weights_tf,\n                         strides=[1, stride, stride, 1],\n                         padding=\"SAME\")\n\n# PyTorch Conv2D\ntorch_conv2d = F.conv2d(x_torch, weights_torch, padding=0, stride=stride)\n\nsess.run(tf.global_variables_initializer())\ntf_result = sess.run(tf_conv2d)\n\ndiff = np.mean(np.abs(tf_result.transpose((0, 3, 1, 2)) - torch_conv2d.detach().numpy()))\nprint('Mean of Abs Diff: {0}'.format(diff))\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>Mean of Abs Diff: 2.2477470551507395e-08\n<\/code><\/pre>\n\n<p>I wasn't quite sure why this was happening when I started writing this question, but a bit of reading clarified this very quickly. I hope this example can help others.<\/p>\n"}]}
+{"QuestionId":52976606,"AnswerCount":2,"Tags":"<python><tensorflow><batch-normalization>","CreationDate":"2018-10-24T19:29:33.350","AcceptedAnswerId":53216124.0,"OwnerUserId":9226290.0,"Title":"Global step not incrementing with batch norm and custom estimator","Body":"<p>I have a customer estimator that has several layers that look like the following in the model function:<\/p>\n\n<pre><code>natural_layer = tf.layers.dense(inputs = natural_layer, \n                                units = units, \n                                activation = None,\n                                use_bias = False,\n                                kernel_regularizer = params['regularizer'],\n                                name = 'pre_batch_norm_layer_' + str(i + 1))\n\nnatural_layer = tf.layers.batch_normalization(natural_layer,\n                                              axis = 1,\n                                              center = True,\n                                              scale = True,\n                                              training = (mode == tf.estimator.ModeKeys.TRAIN),\n                                              name = 'batch_norm_layer_' + str(i + 1))\n\nnatural_layer = params['natural_layer_activation'](natural_layer, name = 'activation_layer_' + str(i + 1))\n<\/code><\/pre>\n\n<p>Because I'm using batch norm, the training op is set up like this:<\/p>\n\n<pre><code>update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nwith tf.control_dependencies(update_ops):\n    optimizer = tf.contrib.opt.MultitaskOptimizerWrapper(params['optimization_algorithm'](params['training_rate']))\n    train_op = optimizer.minimize(loss, global_step = tf.train.get_global_step())\n<\/code><\/pre>\n\n<p>Where the optimizer is usually tf.train.AdamOptimizer.<\/p>\n\n<p>However, when I go to train the estimator the global step never increments (so training will run forever), and I get this:<\/p>\n\n<p>WARNING:tensorflow:It seems that global step (tf.train.get_global_step) has not been increased. Current value (could be stable): 0 vs previous value: 0. You could increase the global step by passing tf.train.get_global_step() to Optimizer.apply_gradients or Optimizer.minimize.<\/p>\n\n<p>I am passing tf.train.get_global_step() to minimize, so I'm not sure why it never gets updated. My hunch is that it has something to do with the batch normalization because when I remove that or replace it with dropout, everything works fine (even when keeping the update ops lines that are required for batch normalization per the documentation).<\/p>\n\n<p>Anyone know what is going on? Happy to post more code if helpful. <\/p>\n","answers":[{"AnswerId":"54707391","CreationDate":"2019-02-15T10:32:23.897","ParentId":null,"OwnerUserId":"10223381","Title":null,"Body":"<p>I get the same issue even without batch normalization. From what I remember, I was able to get around it by calling <code>update_ops<\/code> in <code>sess.run<\/code> along with <code>train_ops<\/code>, as opposed to in a <code>tf.control_dependencies<\/code> clause. This probably is just a bug with tensorflow.<\/p>\n"},{"AnswerId":"53216124","CreationDate":"2018-11-08T21:05:15.603","ParentId":null,"OwnerUserId":"9226290","Title":null,"Body":"<p>I couldn't figure out why the global step was not incrementing automatically, but manually incrementing the global step as follows by adding it to the train_op with tf.group is a nice work around. <\/p>\n\n<pre><code>update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n    with tf.control_dependencies(update_ops):\n        optimizer = tf.contrib.opt.MultitaskOptimizerWrapper(params['optimization_algorithm'](params['training_rate']))\n        train_op = optimizer.minimize(loss)\n\n        global_step = tf.train.get_global_step()\n        update_global_step = tf.assign(global_step, global_step + 1, name = 'update_global_step')\n\n        return tf.estimator.EstimatorSpec(mode, loss = loss, train_op = tf.group(train_op, update_global_step))\n<\/code><\/pre>\n"}]}
+{"QuestionId":52976958,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><scipy><itertools>","CreationDate":"2018-10-24T19:52:56.323","AcceptedAnswerId":null,"OwnerUserId":10550092.0,"Title":"convert python iterative method to parallel","Body":"<p>I am using python itertools and I have method that creates a product of the elements in an array. It looks like this:<\/p>\n\n<pre><code>import itertools as it\nc=it.product('AB', repeat=10)\nfor elem in c:\n     print (''.join(elem))\n<\/code><\/pre>\n\n<p>The result is:<\/p>\n\n<pre><code>AAAAAAAAAA\nAAAAAAAAAB\nAAAAAAAABA\nAAAAAAAABB\nAAAAAAABAA\nAAAAAAABAB\n.\n.\n.\nBBBBBBBBBB\n<\/code><\/pre>\n\n<p>I am looking for a way to convert this iterative process to a matrix calculation. Would NumPy, or TensorFlow, or some other such method be of use in converting it? If so, how? Is this something better or more quickly done with a GPU?<\/p>\n","answers":[{"AnswerId":"52990996","CreationDate":"2018-10-25T13:50:39.837","ParentId":null,"OwnerUserId":"8317830","Title":null,"Body":"<p>Generators and functions will work a little faster.<\/p>\n\n<p>Generator:<\/p>\n\n<pre><code>import itertools as it\nc=it.product('AB', repeat=10)\na = [''.join(elem) for elem in c]\n<\/code><\/pre>\n\n<p>Function:<\/p>\n\n<pre><code>import itertools as it\nc=it.product('AB', repeat=10)\na = map(lambda x: ''.join(x),c)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52977133,"AnswerCount":1,"Tags":"<python><arrays><tensorflow><keras><training-data>","CreationDate":"2018-10-24T20:03:22.073","AcceptedAnswerId":52977916.0,"OwnerUserId":8506781.0,"Title":"expected ndim = 4 found ndim = 5 and other errors - Keras - GTSRB dataset","Body":"<p>I'm trying to make a CNN model based on the GTSRB dataset (link given below), but I'm faced with the following error:<\/p>\n\n<p>When I set the input_shape = input_shape=(3, IMG_SIZE, IMG_SIZE), I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_34_input to\n  have 4 dimensions, but got array with shape (9030, 1)<\/p>\n<\/blockquote>\n\n<p>When I researched the problem, I found that one solution may be passing the batch_size as a parameter, when I try that, I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer conv2d_40: expected\n  ndim=4, found ndim=5<\/p>\n<\/blockquote>\n\n<p>When I try to reshape the training_images, I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: cannot reshape array of size 9030 into shape (48,48,3)<\/p>\n<\/blockquote>\n\n<p>Snippets of the code:\nLoading the training dataset:<\/p>\n\n<pre><code>import csv\n\n# Training dataset\ndef readTrafficSignsTrain(rootpath):\n    '''Reads traffic sign data for German Traffic Sign Recognition Benchmark.\n\n    Arguments: path to the traffic sign data, for example '.\/GTSRB\/Training'\n    Returns:   list of images, list of corresponding labels'''\n    images = [] # images\n    labels = [] # corresponding labels\n\n    # loop over all 42 classes\n    for c in range(0,43):\n#         prefix = rootpath + '\/' + format(c, '05d') + '\/' # subdirectory for class\n#         annFile = open(prefix + 'GT-'+ format(c, '05d') + '.csv') # annotations file\n        prefix = rootpath + '\/00000' + '\/'\n        annFile = open(prefix + 'GT-00000' + '.csv')\n        annReader = csv.reader(annFile, delimiter=';') # csv parser for annotations file\n        next(annReader, None) # skip header\n\n        # loop over all images in current annotations file\n        for row in annReader:\n            images.append(plt.imread(prefix + row[0])) # the 1st column is the filename\n            labels.append(row[7]) # the 8th column is the label\n\n        annFile.close()\n    return images, labels\n\ntraining_images, training_labels = readTrafficSignsTrain('.\/GTSRB\/Training')\n<\/code><\/pre>\n\n<p>Here's a problem, the images shapes are not the same, for example<\/p>\n\n<pre><code>print(len(training_images))\nprint(len(training_labels))\nprint()\nprint(training_images[0].shape)\nprint(training_images[20].shape)\nprint(training_images[200].shape)\nprint(training_images[2000].shape)\n<\/code><\/pre>\n\n<p>Outputs<\/p>\n\n<blockquote>\n  <p>9030 9030<\/p>\n  \n  <p>(30, 29, 3) (54, 57, 3) (69, 63, 3) (52, 51, 3)<\/p>\n<\/blockquote>\n\n<p>Layer setup (copied and pasted from a Keras documentation linked below):<\/p>\n\n<pre><code>IMG_SIZE = 48\nNUM_CLASSES = 43\nK.set_image_data_format('channels_first')\n\nbatch_size = 32\n\ndef cnn_model():\n    model = Sequential()\n\n    model.add(Conv2D(32, (3, 3), padding='same',\n                     input_shape=(3, IMG_SIZE, IMG_SIZE),\n                     activation='relu',\n                     data_format=\"channels_first\"))\n    model.add(Conv2D(32, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), data_format=\"channels_first\"))\n    model.add(Dropout(0.2))\n\n    model.add(Conv2D(64, (3, 3), padding='same',\n                     activation='relu'))\n    model.add(Conv2D(64, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.2))\n\n    model.add(Conv2D(128, (3, 3), padding='same',\n                     activation='relu'))\n    model.add(Conv2D(128, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.2))\n\n    model.add(Flatten())\n    model.add(Dense(512, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(NUM_CLASSES, activation='softmax'))\n    return model\n\nmodel = cnn_model()\n<\/code><\/pre>\n\n<p>Training the model (just model.fit for the time being<\/p>\n\n<pre><code>import numpy\n\ntrim = numpy.array(training_images)\ntrlb = numpy.array(training_labels)\n\nprint(training_images[0].shape)\nprint(trim.shape)\n\ntrim - trim.reshape(48, 48, 3)\n\nmodel.fit(trim, trlb, epochs = 30, batch_size = 32)\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<blockquote>\n  <p>ValueError: cannot reshape array of size 9030 into shape (48,48,3)<\/p>\n<\/blockquote>\n\n<p>When I remove the reshape<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_41_input to\n  have 4 dimensions, but got array with shape (9030, 1)<\/p>\n<\/blockquote>\n\n<p>When I use this instead<\/p>\n\n<pre><code>model.fit(training_images, training_labels, epochs = 30, batch_size = 32)\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<pre><code>&gt; ValueError: Error when checking model input: the list of Numpy arrays\n&gt; that you are passing to your model is not the size the model expected.\n&gt; Expected to see 1 array(s), but instead got the following list of 9030\n&gt; arrays: [array([[[ 75,  78,  80],\n&gt;             [ 74,  76,  78],\n&gt;             [ 86,  87,  84],\n&gt;             ...,\n&gt;             [ 68,  75,  75],\n&gt;             [ 65,  69,  68],\n&gt;             [ 66,  67,  66]],\n&gt;     \n&gt;            [[ 83,  84,  86],\n&gt;             [...\n<\/code><\/pre>\n\n<p>So, if I do that (not really sure why)<\/p>\n\n<pre><code>for i in range(len(training_images)):\n    model.fit(training_images[i], training_labels[i], epochs = 30, batch_size = 32)\n<\/code><\/pre>\n\n<p>I get<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_41_input to\n  have 4 dimensions, but got array with shape (30, 29, 3)<\/p>\n<\/blockquote>\n\n<p>That is with <\/p>\n\n<pre><code>input_shape=(3, IMG_SIZE, IMG_SIZE)\n<\/code><\/pre>\n\n<p>If I make <\/p>\n\n<pre><code>input_shape=(batch_size, 3, IMG_SIZE, IMG_SIZE)\n<\/code><\/pre>\n\n<p>I get <\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer conv2d_47: expected\n  ndim=4, found ndim=5<\/p>\n<\/blockquote>\n\n<p>Output of model.summary()<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_34 (Conv2D)           (None, 32, 48, 48)        896       \n_________________________________________________________________\nconv2d_35 (Conv2D)           (None, 32, 46, 46)        9248      \n_________________________________________________________________\nmax_pooling2d_14 (MaxPooling (None, 32, 23, 23)        0         \n_________________________________________________________________\ndropout_14 (Dropout)         (None, 32, 23, 23)        0         \n_________________________________________________________________\nconv2d_36 (Conv2D)           (None, 64, 23, 23)        18496     \n_________________________________________________________________\nconv2d_37 (Conv2D)           (None, 64, 21, 21)        36928     \n_________________________________________________________________\nmax_pooling2d_15 (MaxPooling (None, 64, 10, 10)        0         \n_________________________________________________________________\ndropout_15 (Dropout)         (None, 64, 10, 10)        0         \n_________________________________________________________________\nconv2d_38 (Conv2D)           (None, 128, 10, 10)       73856     \n_________________________________________________________________\nconv2d_39 (Conv2D)           (None, 128, 8, 8)         147584    \n_________________________________________________________________\nmax_pooling2d_16 (MaxPooling (None, 128, 4, 4)         0         \n_________________________________________________________________\ndropout_16 (Dropout)         (None, 128, 4, 4)         0         \n_________________________________________________________________\nflatten_4 (Flatten)          (None, 2048)              0         \n_________________________________________________________________\ndense_10 (Dense)             (None, 512)               1049088   \n_________________________________________________________________\ndropout_17 (Dropout)         (None, 512)               0         \n_________________________________________________________________\ndense_11 (Dense)             (None, 43)                22059     \n=================================================================\nTotal params: 1,358,155\nTrainable params: 1,358,155\nNon-trainable params: 0\n_________________________________________________________________\nNone\n<\/code><\/pre>\n\n<p>If anyone can help, that would be mostly appreciated.<\/p>\n\n<p>Links\nGTSRB: <a href=\"http:\/\/benchmark.ini.rub.de\/?section=gtsrb&amp;subsection=news\" rel=\"nofollow noreferrer\">http:\/\/benchmark.ini.rub.de\/?section=gtsrb&amp;subsection=news<\/a>\nKeras documentation I got the model from: <a href=\"https:\/\/chsasank.github.io\/keras-tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/chsasank.github.io\/keras-tutorial.html<\/a><\/p>\n\n<p>Link for the full project on github: <a href=\"https:\/\/github.com\/PavlySz\/TSR-Project\" rel=\"nofollow noreferrer\">https:\/\/github.com\/PavlySz\/TSR-Project<\/a><\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52977916","CreationDate":"2018-10-24T21:06:45.927","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>You cannot reshape an np.array to something the dimesions do not allow. Here is what you can do <\/p>\n\n<pre><code>import numpy as np \nimg_arr = np.array([np.ones((30, 29, 3)), \n                    np.ones((54, 57, 3)), \n                    np.ones((69, 63, 3)), \n                    np.ones((52, 51, 3))])\n\nprint(img_arr.shape)\n\nimport cv2\nimg_arr_conv = np.array([cv2.resize(img, dsize=(48, 48)) for img in img_arr])\nprint(img_arr_conv.shape)\n\n&gt;&gt;&gt;(4,)\n&gt;&gt;&gt;(4, 48, 48, 3)\n<\/code><\/pre>\n\n<p>You were getting the <code>ValueError: cannot reshape array of size 9030 into shape (48,48,3)<\/code> because numpy cannot infer the dimensions of an array if the elements are all different sizes and it cannot reshape an array that the dimensions do not allow. This is also the case for <code>ValueError: Error when checking input: expected conv2d_41_input to have 4 dimensions, but got array with shape (9030, 1)<\/code>. Numpy only knows there are 9030 elements in the array. It cannot do anything more than that because all the dimensions of the elements are different.<br>\nExample<\/p>\n\n<pre><code>img_arr_bad = np.array([np.ones((30, 29, 3)), \n                        np.ones((54, 57, 3)), \n                        np.ones((69, 63, 3)), \n                        np.ones((52, 51, 3))])\n\nimg_arr_good = np.array([np.ones((48, 48, 3)), \n                         np.ones((48, 48, 3)), \n                         np.ones((48, 48, 3)), \n                         np.ones((48, 48, 3))])\n\nprint(img_arr_bad.shape)\nprint(img_arr_good.shape)\n\n&gt;&gt;&gt;(4,)\n&gt;&gt;&gt;(4, 48, 48, 3)\n<\/code><\/pre>\n\n<p>Hope this helps<\/p>\n"}]}
+{"QuestionId":52977150,"AnswerCount":2,"Tags":"<python><tensorflow><tensor><transformer>","CreationDate":"2018-10-24T20:04:28.827","AcceptedAnswerId":null,"OwnerUserId":5833067.0,"Title":"InvalidArgumentError: Mismatch between the current graph and the graph from the checkpoint","Body":"<p>So I am basically using this transformer implementation for my project: <a href=\"https:\/\/github.com\/Kyubyong\/transformer\" rel=\"noreferrer\">https:\/\/github.com\/Kyubyong\/transformer<\/a> . \nIt works great on the German to English translation it was originally written for and I modified the processing python script in order to create vocabulary files for the languages that I want to translate. This seems to work fine.<\/p>\n\n<p>However when it comes to training I get the following error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Restoring from\n  checkpoint failed. This is most likely due to a mismatch between the\n  current graph and the graph from the checkpoint. Please ensure that\n  you have not altered the graph expected based on the checkpoint.\n  Original error:<\/p>\n  \n  <p>Assign requires shapes of both tensors to match. lhs shape= [9796,512]\n  rhs shape= [9786,512]      [[{{node save\/Assign_412}} =\n  Assign[T=DT_FLOAT, _class=[\"loc:@encoder\/enc_embed\/lookup_table\"],\n  use_locking=true, validate_shape=true,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](encoder\/enc_embed\/lookup_table\/Adam_1,\n  save\/RestoreV2:412)]]<\/p>\n<\/blockquote>\n\n<p>Now I have no idea why I am getting the above error. I also reverted to the original code to translate from German to English and now I get the same error (except the lhs and rhs tensor shapes are different of course), when before it was working! <\/p>\n\n<p>Any ideas on why this could be happening?<\/p>\n\n<p>Thanks in advance<\/p>\n\n<p>EDIT: This is the specific file in question here, the train.py when it is run: <a href=\"https:\/\/github.com\/Kyubyong\/transformer\/blob\/master\/train.py\" rel=\"noreferrer\">https:\/\/github.com\/Kyubyong\/transformer\/blob\/master\/train.py<\/a> \nNothing has been modified other than the fact that the vocab loaded for de and en are differently (they're in fact vocab files with single letters as words). However as I mentioned that even when resorting back to the prevous working example I get the same error with different lhs and rhs dimensions. <\/p>\n","answers":[{"AnswerId":"55119829","CreationDate":"2019-03-12T10:57:30.463","ParentId":null,"OwnerUserId":"4940835","Title":null,"Body":"<p>I was getting a similar error, I my case it seems that the output of previous failed jobs was remaining on the output dir and there were some incompatibilities when saving\/restoring the checkpoints of the new job, so I just <strong>cleaned it up the output dir<\/strong> and then the new job worked correctly. <\/p>\n"},{"AnswerId":"53073854","CreationDate":"2018-10-30T22:35:52.430","ParentId":null,"OwnerUserId":"268408","Title":null,"Body":"<p>The large number is almost certainly the size of your vocabulary. The initial matrix will have size [vocab_size, hidden_dim]. So by changing the size of your vocab you are breaking things.<\/p>\n\n<p>Presumably the solution is just to make sure you clean out all your checkpoints so that you are only looking at models trained with the vocabulary you want.<\/p>\n"}]}
+{"QuestionId":52977453,"AnswerCount":1,"Tags":"<python><tensorflow><pip><artificial-intelligence>","CreationDate":"2018-10-24T20:27:11.370","AcceptedAnswerId":53006871.0,"OwnerUserId":10090361.0,"Title":"Why is pip not updating tensorflow correctly, and, if it is, why is the 'attrib error' still thrown?","Body":"<p>I've installed tensorflow over pip3 and python3, and am working on it. While using the colum function, the commonly experienced error <code>AttributeError: module 'tensorflow' has no attribute 'feature_column'<\/code>. <\/p>\n\n<p>It might look like a duplicate question, but I've looked at the other occurrences of the same question, but, after updating the file (<code>pip3 install --upgrade tensorflow<\/code>), I checked the version. The version 0.12.0 is shown. So why does pip still show its completely new. Is 0.12.0 the newest version?<\/p>\n\n<p>When I attempted to uninstall tensorflow and re-install it, it refuses to re-install. I'm using <code>python3 -m pip install tensorflow<\/code>. The error thrown here is <code>Could not find a version that satisfies the requirement tensorflow (from versions: ) No matching distribution found for tensorflow<\/code><\/p>\n\n<p>Thanks in advance for your help<\/p>\n","answers":[{"AnswerId":"53006871","CreationDate":"2018-10-26T10:37:55.713","ParentId":null,"OwnerUserId":"10090361","Title":null,"Body":"<p>As mentioned in the comments, the most probable solution to the attribute error is the update problem. However, if you're encountering the <code>Could not find a version that satisfies the requirement tensorflow (from versions: ) No matching distribution found for tensorflow<\/code>, the easiest solution is to use following code.<\/p>\n\n<p><code>pip install --upgrade  https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-1.1.0-py2-none-any.whl<\/code> (for python 2)<\/p>\n\n<p>or<\/p>\n\n<p><code>pip install --upgrade  https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-1.1.0-py3-none-any.whl<\/code> (for python 3)<\/p>\n"}]}
+{"QuestionId":52977513,"AnswerCount":0,"Tags":"<python><tensorflow><scikit-learn><keras>","CreationDate":"2018-10-24T20:32:08.663","AcceptedAnswerId":null,"OwnerUserId":6341510.0,"Title":"Using classification_report to evaluate a Keras model","Body":"<p><strong>The Problem:<\/strong><\/p>\n\n<p>During training the performance of my model looks quite allright. However, the results of the classification_report from sklearn yields a precision, recall and f1 of zero almost everywhere. What am I doing wrong to get such a missmatch between training performance and inference? (I am using Keras with a TensorFlow backend.)<\/p>\n\n<p><strong>My code:<\/strong><\/p>\n\n<p>I use the <code>valiation_split<\/code> argument to generate two generators (train, validation) like so:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\nrescale=1. \/ 255, validation_split=0.15)\n\ntrain_generator = train_datagen.flow_from_directory(\ntrain_data_dir,\ntarget_size=(img_height, img_width),\nbatch_size=batch_size,\nclass_mode='categorical', subset=\"training\")\n\nvalidation_generator = train_datagen.flow_from_directory(\ntrain_data_dir,\ntarget_size=(img_height, img_width),\nbatch_size=batch_size,\nclass_mode='categorical', subset=\"validation\", shuffle=False)\n<\/code><\/pre>\n\n<p>I set <code>shuffle=False<\/code> in my validation_generator to make sure it does not mix the relationship of images and labels for my evaluation later on.<\/p>\n\n<p>Next, I train my model like so:<\/p>\n\n<pre><code>history = model.fit_generator(\n    train_generator,\n    steps_per_epoch=nb_train_samples \/\/ batch_size,\n    epochs=epochs,\n    validation_data=validation_generator,\n    validation_steps=nb_validation_samples \/\/ batch_size,\n    verbose=1)\n<\/code><\/pre>\n\n<p>Performance is allright:<\/p>\n\n<pre><code>Epoch 1\/5\n187\/187 [==============================] - 44s 233ms\/step - loss: 0.7835 - acc: 0.6744 - val_loss: 1.2918 - val_acc: 0.6079\nEpoch 2\/5\n187\/187 [==============================] - 42s 225ms\/step - loss: 0.7578 - acc: 0.6901 - val_loss: 1.2962 - val_acc: 0.6149\nEpoch 3\/5\n187\/187 [==============================] - 40s 216ms\/step - loss: 0.7535 - acc: 0.6907 - val_loss: 1.3426 - val_acc: 0.6061\nEpoch 4\/5\n187\/187 [==============================] - 41s 217ms\/step - loss: 0.7388 - acc: 0.6977 - val_loss: 1.2866 - val_acc: 0.6149\nEpoch 5\/5\n187\/187 [==============================] - 41s 217ms\/step - loss: 0.7282 - acc: 0.6960 - val_loss: 1.2988 - val_acc: 0.6297\n<\/code><\/pre>\n\n<p>Now, I extract the necessary info for the classification_report following the method suggested here <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2607#issuecomment-302365916\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/2607#issuecomment-302365916<\/a> . This gives me the following:<\/p>\n\n<pre><code>validation_steps_per_epoch = np.math.ceil(validation_generator.samples \/ validation_generator.batch_size)\n\npredictions = model.predict_generator(validation_generator, steps=validation_steps_per_epoch)\n# Get most likely class\npredicted_classes = np.argmax(predictions, axis=1) \n\ntrue_classes = validation_generator.classes\nclass_labels = list(validation_generator.class_indices.keys())  \n<\/code><\/pre>\n\n<p>Finally, I output the classification report using:<\/p>\n\n<pre><code>from sklearn.metrics import classification_report\nreport = classification_report(true_classes, predicted_classes, target_names=class_labels)\nprint(report)    \n<\/code><\/pre>\n\n<p>Which results in zeros all over the place (see avgs. below): <\/p>\n\n<pre><code>                   precision    recall  f1-score   support\n      micro avg       0.01      0.01      0.01      2100\n      macro avg       0.01      0.01      0.01      2100\n   weighted avg       0.01      0.01      0.01      2100\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52977683,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-24T20:46:30.877","AcceptedAnswerId":null,"OwnerUserId":10201498.0,"Title":"custom keras loss function","Body":"<p>I want to write a custom loss function in keras, which also uses the gradient of the output w.r.t the input<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\n\nmodel = Sequential()\nmodel.add(Dense(200, input_dim=20, activation='relu'))\nmodel.add(Dense(64, input_dim=20, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy',\n          optimizer='rmsprop',\n          metrics=['accuracy']) \n<\/code><\/pre>\n\n<p>Now, I want to replace the loss function. I only know how to write using tensorflow as follows:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=[None, 1], name=\"x\") \next_f0 = tf.placeholder(tf.float32, shape=[None, 1], name=\"f\") \nxx0 = tf.concat([[[R_variable['x_start'] * 1.0],[R_variable['x_end'] * 1.0]], x], 0)\nyy0 = univAprox(xx0)\nBoundary_y = yy0[0:2]\ny = yy0[2:]\nGradU = tf.gradients(y,x)  #***KEY***\nGradSum = tf.reduce_sum(tf.square(GradU)) \/ 2 * dx_train\nFSum = tf.reduce_sum(tf.multiply(ext_f0,y)) * dx_train\nloss = GradSum + FSum + Beta * tf.reduce_sum(tf.square(Boundary_y))\n<\/code><\/pre>\n\n<p>The key is I need to get the <code>GradU<\/code>. How should I do this in keras?<\/p>\n","answers":[{"AnswerId":"52982532","CreationDate":"2018-10-25T06:17:33.410","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You can write your custom loss function. All you should know is it's form:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n  # do something to compute loss\n  ...\n  return loss\n<\/code><\/pre>\n\n<p>And that should be it.<\/p>\n"}]}
+{"QuestionId":52977700,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-24T20:48:12.567","AcceptedAnswerId":52977999.0,"OwnerUserId":1141351.0,"Title":"logits and labels must be broadcastable: logits_size=[32,1] labels_size=[16,1]","Body":"<p>I'm trying to learn tensorflow and I'm getting the following error: \nlogits and labels must be broadcastable: logits_size=[32,1] labels_size=[16,1]<\/p>\n\n<p>The code runs fine when I got this as input:     <\/p>\n\n<pre><code>self.input = np.ones((500, 784))\nself.y = np.ones((500, 1))\n<\/code><\/pre>\n\n<p>However, when I add and extra dimension the error is thrown:<\/p>\n\n<pre><code>self.input = np.ones((500, 2, 784))\n    self.y = np.ones((500, 1))\n<\/code><\/pre>\n\n<p>The code to build the graph<\/p>\n\n<pre><code>    self.x = tf.placeholder(tf.float32, shape=[None] + self.config.state_size)\n    self.y = tf.placeholder(tf.float32, shape=[None, 1])\n\n    # network architecture\n    d1 = tf.layers.dense(self.x, 512, activation=tf.nn.relu, name=\"dense1\")\n    d2 = tf.layers.dense(d1, 1, name=\"dense2\")\n\n\n    with tf.name_scope(\"loss\"):\n        self.cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=self.y, logits=d2))\n        self.train_step = tf.train.AdamOptimizer(self.config.learning_rate).minimize(self.cross_entropy,\n                                                                                     global_step=self.global_step_tensor)\n        correct_prediction = tf.equal(tf.argmax(d2, 1), tf.argmax(self.y, 1))\n        self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n<\/code><\/pre>\n\n<p>Could someone explain me why this is happening and how I can fix this?<\/p>\n","answers":[{"AnswerId":"52977999","CreationDate":"2018-10-24T21:14:16.043","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p><code>logits<\/code> is the name typically given to the output of the network, these are your predictions. A size of <code>[32, 10]<\/code> tells me that you have a batch size of 32, and 10 outputs, such as is common with mnist, as you appear to be working with.<\/p>\n\n<p>Your labels are sized <code>[16, 10]<\/code>, which is to say, you're providing 16 labels\/vectors of size 10. The number of labels you're providing is in conflict with the output of the network, they should be the same.<\/p>\n\n<p>I'm not quite clear what you're doing with the extra dimension in the input, but I guess you must be accidentally doubling the samples in some way. Perhaps the <code>[500, 2, 784]<\/code> shape is being reshaped to <code>[1000, 784]<\/code> automatically somewhere along the way, which is then not matching the 500 labels. Also, your <code>self.y<\/code> should be shaped <code>[500, 10]<\/code> not, <code>[500, 1]<\/code>, your labels need to be in one-hot encoding format. E.g. a single label of shape <code>[1, 10]<\/code> for digit 3 would be <code>[[0,0,0,1,0,0,0,0,0,0,0]]<\/code>, not in digit representation, e.g. <code>[3]<\/code> as you seem to have it set up in your sanity-test here.<\/p>\n"}]}
+{"QuestionId":52977830,"AnswerCount":0,"Tags":"<deep-learning><caffe><pycaffe>","CreationDate":"2018-10-24T20:59:27.537","AcceptedAnswerId":null,"OwnerUserId":6327658.0,"Title":"Train Caffe CNN to output multidimensional features","Body":"<p>I'd like to build a feature extractor using Caffe's CNNs and I already have a large sample of input features and desired output features.<\/p>\n\n<p>Now I need to train some convolutional layers to learn how to transform the input features into the output.<\/p>\n\n<p>My question is: How can I achieve this on Caffe?<\/p>\n\n<p>As a minimal example, suppose I wanted to train a CNN that inverts the values of a 2D array.<\/p>\n\n<p>For example, if my input is <\/p>\n\n<pre><code>[[0,1,0],\n [1,1,1],\n [0,1,0]]\n<\/code><\/pre>\n\n<p>the CNN should output <\/p>\n\n<pre><code>[[1,0,1],\n [0,0,0],\n [1,0,1]].\n<\/code><\/pre>\n\n<p>For <\/p>\n\n<pre><code>[[0,0,0],\n [0,1,0],\n [0,0,0]]\n<\/code><\/pre>\n\n<p>the output should be <\/p>\n\n<pre><code>[[1,1,1],\n [1,0,1],\n [1,1,1]]\n<\/code><\/pre>\n\n<p>and so on.<\/p>\n\n<p>Of course this is just a minimal example to share, the actual problem is nearly impossible to tackle without the use of multiple convolutions.<\/p>\n\n<p>I was able to create this code for this problem. I used the Euclidean Loss at the end, but unfortunately the CNN is not learning anything.<\/p>\n\n<pre><code>ROOT_DIR = '\/home'\n\nfrom os.path import join\nimport numpy as np\nimport h5py\nfrom itertools import product\nimport caffe\nfrom caffe import layers\nfrom caffe.proto import caffe_pb2\n\n#%% GENERATE DATA\ndata_in = np.array([np.array(seq).reshape(1,3,3) for seq in product([0,1], repeat=9)])\ndata_out = np.array([-1*array+1 for array in data_in])\n\nwith open(join(ROOT_DIR, 'data.txt'), 'w') as ftxt:\n    with h5py.File(join(ROOT_DIR, 'data.hdf5'), 'w') as fhdf5:\n        fhdf5['data'] = data_in.astype(np.float32)\n        fhdf5['label'] = data_out.astype(np.float32)\n        ftxt.write(join(ROOT_DIR, 'data.hdf5'))\n\n#%%DEFINE NET\nnet = caffe.NetSpec()\nnet.data, net.label = layers.HDF5Data(batch_size=64, source=join(ROOT_DIR, 'data.txt'), ntop=2)\nnet.conv1 = layers.Convolution(net.data, kernel_size=1, num_output=128)\nnet.relu1 = layers.ReLU(net.conv1, in_place=True)\nnet.conv2 = layers.Convolution(net.relu1, kernel_size=1, num_output=1)\nnet.relu2 = layers.ReLU(net.conv2, in_place=True)\nnet.loss = layers.EuclideanLoss(net.relu2, net.label)\nnet.to_proto()\n\nwith open(join(ROOT_DIR, 'invert_net.prototxt'), 'w') as f:\n    f.write(str(net.to_proto()))\n\n#%% DEFINE SOLVER\nsolver = caffe_pb2.SolverParameter()\nsolver.train_net = join(ROOT_DIR, 'invert_net.prototxt')\nsolver.max_iter = 10000\nsolver.base_lr = 0.01\nsolver.lr_policy = 'fixed'\n\nwith open(join(ROOT_DIR, 'solver.prototxt'), 'w') as f:\n    f.write(str(solver))\n\n#%% TRAIN NET\ncaffe.set_mode_cpu()\nsolver = caffe.SGDSolver(join(ROOT_DIR, 'solver.prototxt'))\nsolver.solve()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":52977897,"AnswerCount":2,"Tags":"<tensorflow><gpu><nvidia>","CreationDate":"2018-10-24T21:05:05.950","AcceptedAnswerId":53200554.0,"OwnerUserId":4835204.0,"Title":"TensorFlow GPU compile flags?","Body":"<p>I'm looking for some clarification on compile flag options when using TensorFlow with NVIDIA GPUs on Ubuntu 18.04.  I'm using TensorFlow for both Python (for training) and calling from C++ (for executing in production).<\/p>\n\n<p>Since Ubuntu 18.04 ships with GCC 7.x I have to use CUDA 9.2, so I can't use the Google provided TensorFlow pre-compiled binaries (those currently only work with CUDA 9.0 which is not compliant with GCC 7.x).  Therefore, I have to compile TensorFlow from source to use with both Python and C++<\/p>\n\n<p>Currently I'm using the following compile flags:<\/p>\n\n<p><strong>Python compile:<\/strong><\/p>\n\n<pre><code>bazel build --config=opt \\\n            --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p><strong>C++ compile:<\/strong><\/p>\n\n<pre><code>bazel build -c opt \\\n            --copt=-mavx \\\n            --copt=-mavx2 \\\n            --copt=-mfma \\\n            --copt=-mfpmath=both \\\n            --copt=-msse4.2 \\\n            --config=cuda \/\/tensorflow:libtensorflow_cc.so\n<\/code><\/pre>\n\n<p>This is based mostly on popular vote on the interwebs which makes me uncomfortable.  Here are some of the sites\/posts consulted which lead me to these choices:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/source<\/a>\nhas:<\/p>\n\n<pre><code>bazel build --config=opt \\\\\n            --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p><a href=\"http:\/\/www.bitbionic.com\/2017\/08\/18\/run-your-keras-models-in-c-tensorflow\/\" rel=\"nofollow noreferrer\">http:\/\/www.bitbionic.com\/2017\/08\/18\/run-your-keras-models-in-c-tensorflow\/<\/a>\nhas:<\/p>\n\n<pre><code>bazel build --jobs=6 \\\n            --verbose_failures \\\n            -c opt \\\n            --copt=-mavx \\\n            --copt=-mfpmath=both \\\n            --copt=-msse4.2 \/\/tensorflow:libtensorflow_cc.so\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/41293077\/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions\">How to compile Tensorflow with SSE4.2 and AVX instructions?<\/a>\nhas:<\/p>\n\n<pre><code>bazel build -c opt \\\n            --copt=-mavx \\\n            --copt=-mavx2 \\\n            --copt=-mfma \\\n            --copt=-mfpmath=both \\\n            --config=cuda -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/43311362\/re-build-tensorflow-with-desired-optimization-flags\">Re-build Tensorflow with desired optimization flags<\/a>\nhas:<\/p>\n\n<pre><code>bazel build -c opt \\\n            --copt=-mavx \\\n            --copt=-mavx2 \\\n            --copt=-mfma \\\n            --copt=-mfpmath=both \\\n            --copt=-msse4.2 \\\n            --config=cuda -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>Can somebody enlighten us all on these flag options?  Specifically I have the following questions:<\/p>\n\n<p>1) What are the <code>mavx<\/code>, <code>mavx2<\/code>, <code>mfma<\/code>, and <code>mfpmath<\/code> flags?  Should these be used for both the Python and C++ compile or only for the C++ compile?  The fact that the Google walk-through does not use these for the Python compile inclines me towards the same.<\/p>\n\n<p>2) Clearly the <code>--copt=-msse4.2<\/code> is to enable SSE optimization for Intel CPUs and the <code>--config=cuda<\/code> is for CUDA GPUs, but what is the <code>-k<\/code> option at the end of the CUDA flag?  Note that some of the above examples use the <code>-k<\/code> option and some do not.<\/p>\n\n<p>3) Is there a place where these options are documented?  I wonder if there are other flags that could be beneficial or if some of the above should be omitted.  I checked the TensorFlow and Bazel GitHubs and did not find anything on this topic.<\/p>\n","answers":[{"AnswerId":"53200554","CreationDate":"2018-11-08T02:07:10.473","ParentId":null,"OwnerUserId":"6780025","Title":null,"Body":"<p>1) There flags are passed by bazel to gcc. See GCC documentation for exact meaning. For example, for <code>fpmath<\/code> see something like <a href=\"https:\/\/gcc.gnu.org\/onlinedocs\/gcc-4.4.7\/gcc\/i386-and-x86_002d64-Options.html\" rel=\"nofollow noreferrer\">https:\/\/gcc.gnu.org\/onlinedocs\/gcc-4.4.7\/gcc\/i386-and-x86_002d64-Options.html<\/a>. You should use them when you know that the CPU you will be running the binary on supports these options. It should not hurt to use them everywhere.<\/p>\n\n<p>2) -k is the bazel flag documented here: <a href=\"https:\/\/docs.bazel.build\/versions\/master\/user-manual.html#flag--keep_going\" rel=\"nofollow noreferrer\">https:\/\/docs.bazel.build\/versions\/master\/user-manual.html#flag--keep_going<\/a><\/p>\n\n<p>3) Yes, see above.<\/p>\n"},{"AnswerId":"54032969","CreationDate":"2019-01-04T04:04:02.927","ParentId":null,"OwnerUserId":"131935","Title":null,"Body":"<p>From the gcc documentation it seems that the tensorflow default of -march=native is all that should be necessary to utilize all of the capabilities of your processor.  That's probably better than setting these optimization flags manually, as you could be missing some.<\/p>\n"}]}
+{"QuestionId":52978434,"AnswerCount":1,"Tags":"<python><scikit-learn><neural-network><keras><cross-validation>","CreationDate":"2018-10-24T21:54:00.427","AcceptedAnswerId":52994859.0,"OwnerUserId":6145828.0,"Title":"Is it possible to use StratifiedKFold from sklearn in multi input Neural Networks?","Body":"<p>I have a dataset that can be passed to a multi input neural network in the shape of a python <em>dictionary<\/em> or a <em>list<\/em>:<\/p>\n\n<p><strong>Example<\/strong>:<\/p>\n\n<pre><code>#dict\n{'input1': X1, 'input2': X2, 'input3': X3}, {'output': Y}\n#list\n[ X1, X2, X3], y\n<\/code><\/pre>\n\n<p>Now I would like to use <strong>K-fold cross validation<\/strong> to estimate the performance of my model. Do you think I can use <code>StratifiedKFold<\/code> from <code>sklearn<\/code> as in this mono-input example?<\/p>\n\n<pre><code>for train, test in kfold.split(X, Y):\n  # create model\n    model = Sequential()\n    model.add(Dense(12, input_dim=8, activation='relu'))\n    model.add(Dense(8, activation='relu'))\n    model.add(Dense(1, activation='sigmoid'))\n    # Compile model\n    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n    # Fit the model\n    model.fit(X[train], Y[train], epochs=150, batch_size=10, verbose=0)\n    # evaluate the model\n    scores = model.evaluate(X[test], Y[test], verbose=0)\n    print(\"%s: %.2f%%\" % (model.metrics_names[1], scores[1]*100))\n    cvscores.append(scores[1] * 100)\n<\/code><\/pre>\n\n<p>Otherwise, what would be the best way?<\/p>\n","answers":[{"AnswerId":"52994859","CreationDate":"2018-10-25T17:21:34.597","ParentId":null,"OwnerUserId":"8544925","Title":null,"Body":"<p>skf.split() returns you indices and it depends only on Y:<\/p>\n\n<pre><code>for train_index, test_index in skf.split(X, y):\n...    print(\"TRAIN:\", train_index, \"TEST:\", test_index)\n...    X_train, X_test = X[train_index], X[test_index]\n...    y_train, y_test = y[train_index], y[test_index]\n<\/code><\/pre>\n\n<p>So you can pass any X array here (even X1 for layer 1 converted to dataframe or synthetic X). Then you take your<\/p>\n\n<pre><code>train_index\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>test_index\n<\/code><\/pre>\n\n<p>and filter all your inputs.<\/p>\n\n<p>Again, skf() depends on your Y only. So the goal is to pass proper Y and get indices.<\/p>\n\n<p>Other way: union all you inputs in outputs in one dataframe and keep column names for each layer. In this case you have one 'big' X. Split it firstly train_index and test_index and then split into X1, X2 and X3 using column names you saved above.<\/p>\n"}]}
+{"QuestionId":52979374,"AnswerCount":2,"Tags":"<tensorflow><pca><gensim><tensorboard>","CreationDate":"2018-10-24T23:42:34.947","AcceptedAnswerId":54315313.0,"OwnerUserId":8383446.0,"Title":"Tensorboard projector will compute PCA endlessly","Body":"<p>I have just over 100k word embeddings which I created using gensim, originally each containing 200 dimensions. I've been trying to visualize them within tensorboard's projector but I have only failed so far.\nMy problem is that tensorboard seems to freeze while computing PCA. At first, I left the page open for 16 hours, imagining that it was just too much to be calculated, but nothing happened. At this point, I started to try and test different scenarios just in case all I needed was more time and I was trying to rush things. The following is a list of my testing so far, all of which failed at the same spot, computing PCA:<\/p>\n\n<ul>\n<li>I plotted only 10 points of 200 dimensions;<\/li>\n<li>I retrained my gensim model so that I could reduce its dimensionality to 100;<\/li>\n<li>Then I reduced it to 10;<\/li>\n<li>Then to 2;<\/li>\n<li>Then I tried plotting only 2 points, i.e. 2 two dimensional points;<\/li>\n<\/ul>\n\n<p>I am using Tensorflow 1.11;\nYou can find my last saved tensor flow session <a href=\"https:\/\/drive.google.com\/open?id=10Cnzc2RH9nDFUYrf51r6d8d0jsl0-H6Y\" rel=\"nofollow noreferrer\">here<\/a>, would you mind trying it out?<\/p>\n\n<p>I am still a beginner, therefore I used a couple tutorial to get me started; I used <a href=\"https:\/\/github.com\/sudharsan13296\/visualise-word2vec\/blob\/master\/Word2vec%20Embeddings.ipynb\" rel=\"nofollow noreferrer\">Sud Harsan<\/a> work so far.<\/p>\n\n<p>Any help is much appreciated. Thanks.<\/p>\n\n<hr>\n\n<p>Updates:<\/p>\n\n<p>A) I've found someone else <a href=\"https:\/\/stackoverflow.com\/questions\/44054907\/tensorfboard-embeddings-hangs-with-computing-pca\">dealing with the same problem<\/a>; I tried the solution provided, but it didn't change anything. <\/p>\n\n<p>B) I thought it could have something to do with my installation, therefore I tried uninstalling tensorflow and installing it back; no luck. I then proceeded to create a new environment dedicated to tensorflow and that also didn't work. <\/p>\n\n<p>C) Assuming there was something wrong with my code, I ran <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">tensorflow's basic embedding tutorial<\/a> to check if I could open its projector's results. And guess what?! I still can't go past \"Calculating PCA\"<\/p>\n\n<p>Now, I did visit <a href=\"https:\/\/projector.tensorflow.org\" rel=\"nofollow noreferrer\">the online projector example<\/a> and that loads perfectly. <\/p>\n\n<p>Again, Any help would be more than appreciated. Thanks!<\/p>\n","answers":[{"AnswerId":"54315313","CreationDate":"2019-01-22T19:38:28.390","ParentId":null,"OwnerUserId":"8383446","Title":null,"Body":"<p>As mentioned by Bluedrops, updating tensorboard and tensorflow seems to fix the problem.<\/p>\n\n<p>I created a new environment with conda and installed the newest versions of Tensorflow, Tensorboard and their dependencies and that seems to fix the issue.<\/p>\n"},{"AnswerId":"53140464","CreationDate":"2018-11-04T11:46:00.653","ParentId":null,"OwnerUserId":"7813311","Title":null,"Body":"<p>I have the same problem with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">word2vec_basic.py<\/a><\/p>\n\n<p>My environment: win10, conda, python 3.6.7, tensorflow 1.11, tensorboard 1.11<\/p>\n\n<p>That may not your fault because I roll back tensorflow &amp; tensorboard from 1.11 to 1.7<\/p>\n\n<p>And guess what?! The projector appears just a few seconds!<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/1480\" rel=\"nofollow noreferrer\">reference<\/a><\/p>\n\n<hr>\n\n<p>Update 10\/11<\/p>\n\n<p>tensorboard &amp; tensorflow 1.12 are available in conda today, I take a try and this problem seems to be fixed.<\/p>\n"}]}
+{"QuestionId":52979727,"AnswerCount":1,"Tags":"<tensorflow><keras><artificial-intelligence><gpu><pycuda>","CreationDate":"2018-10-25T00:36:14.260","AcceptedAnswerId":52999957.0,"OwnerUserId":4910036.0,"Title":"Specify Keras GPU without using CUDA_VISIBLE_DEVICES","Body":"<p>I have a system with two GPUs, and am using Keras with Tensorflow backend. Gpu:0 is being allocated to PyCUDA, which is performing a unique operation which is fed forward to Keras, and changes with each batch. As such, I would like to run a Keras model on gpu:1 while leaving gpu:0 allocated to PyCUDA.<\/p>\n\n<p>Is there any way to do this? Looking through prior threads I've found several depreciated solutions.<\/p>\n","answers":[{"AnswerId":"52999957","CreationDate":"2018-10-26T00:51:04.320","ParentId":null,"OwnerUserId":"4910036","Title":null,"Body":"<p>So I don't think that this feature is meaningfully implemented in Keras currently. Found a workaround that I recommend whereby you just create multiple processes using Python's default multiprocessing library.<\/p>\n\n<p>Note: Currently for this setup you need to spawn the new process, rather than fork it, to avoid a weird interaction with one of the PyCUDA backend libraries.<\/p>\n"}]}
+{"QuestionId":52979824,"AnswerCount":1,"Tags":"<python><opencv><keras><deep-learning><computer-vision>","CreationDate":"2018-10-25T00:51:22.067","AcceptedAnswerId":null,"OwnerUserId":2163392.0,"Title":"Is it possible to classify patches of the image of the same object, but with different areas with CNN?","Body":"<p>I have two pictures of a printed object, the first has a printed area of 2.5x2.5 cm^2, and the second is the same object, but the printed area is 5.0x5.0 cm^2. After separating the object from the background and equalizing the histogram of both pictures, I am trying to use small patches (64x64) in a deep learning approach (CNN) to understanding their patterns and classify them. I am trying to use 64x64 patches from the 2.5x2.5cm^2 printed objects to train a deep learning classifier and test them with patches from 5.0x5.0cm^2 objects. The digital images of both objects have approximately the same resolution, as it is defined from the object extractor. Here are examples of the 64x64 patches used to train and test the CNN binary classifier.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JkZnZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JkZnZ.jpg\" alt=\"\"><\/a><br>\n64x64 patch of a 2.5x2.5cm^2 object<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ULDfY.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ULDfY.jpg\" alt=\"\"><\/a><\/p>\n\n<p>64x64 patch of a 5x5cm^2 object<\/p>\n\n<p>The classes I want to predict are the following:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BTHbM.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BTHbM.jpg\" alt=\"Negative Class\"><\/a><\/p>\n\n<p>Negative Class (printed for the first time)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zFf5Q.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zFf5Q.jpg\" alt=\"Positive Class\"><\/a><\/p>\n\n<p>Positive class (copied and reprinted)<\/p>\n\n<p>What I found out:<\/p>\n\n<ol>\n<li>Patches from 2.5x2.5cm^2 objects are easily classified if the CNN is trained with patches from the same size (area) objects<\/li>\n<li>If the CNN is trained with 64x64 patches from 2.5x2.5cm^2 objects and tested with 64x64 patches from 5x5cm^2 objects, the predictions are just for one class (50% accuracy).<\/li>\n<li>Some multiscale and multiple resolution descriptors work perfectly in this scenario, such as using Bag of Visual Words<\/li>\n<li>Other baseline CNNs also fail in this scenario, such as Mobilenet, Densenet and Resnet<\/li>\n<li>I tried to include zooms in my data augmentation procedure (like suggested by one answer). It did not work either :-(<\/li>\n<\/ol>\n\n<p>This is the keras model I have tried so far<\/p>\n\n<pre><code>model = Sequential()\n\n# GROUP1\nmodel.add(Conv2D(filters=32, kernel_size=3, strides=1, padding='same',\n                 input_shape=input_shape))\nmodel.add(LeakyReLU(alpha=0.2))\n\n\n# GROUP2\nmodel.add(Conv2D(filters=32, kernel_size=3, strides=2, padding='same'))\nmodel.add(LeakyReLU(alpha=0.2))\nmodel.add(BatchNormalization(axis=-1, momentum=0.9, epsilon=0.001))\n\n# GROUP3\nmodel.add(Conv2D(filters=64, kernel_size=3, strides=1, padding='same'))\nmodel.add(LeakyReLU(alpha=0.2))\nmodel.add(BatchNormalization(axis=-1, momentum=0.9, epsilon=0.001))\n\n# GROUP4\nmodel.add(Conv2D(filters=64, kernel_size=3, strides=2, padding='same'))\nmodel.add(LeakyReLU(alpha=0.2))\nmodel.add(BatchNormalization(axis=-1, momentum=0.9, epsilon=0.001))\n\n# GROUP5\nmodel.add(Conv2D(filters=96, kernel_size=3, strides=1, padding='same'))\nmodel.add(LeakyReLU(alpha=0.2))\nmodel.add(BatchNormalization(axis=-1, momentum=0.9, epsilon=0.001))\n\n# GROUP6\nmodel.add(Conv2D(filters=96, kernel_size=3, strides=2, padding='same'))\nmodel.add(LeakyReLU(alpha=0.2))\nmodel.add(BatchNormalization(axis=-1, momentum=0.9, epsilon=0.001))\n\nmodel.add(Flatten())\nmodel.add(Dense(1024))\nmodel.add(LeakyReLU(alpha=0.2))\n\nmodel.add(Dense(2, activation='softmax'))\n\nreturn model\n<\/code><\/pre>\n\n<p>and here is the data augmentation I am using<\/p>\n\n<pre><code>datagen = ImageDataGenerator(\n            width_shift_range=0.2,\n            height_shift_range=0.2,\n            horizontal_flip=True,\n        vertical_flip=True,\n        zoom_range=0.2,\n            fill_mode='nearest')\n\n\n    datagen.fit(x_train)\n    datagen.fit(x_validation)\n\n    # Fit the model on the batches generated by datagen.flow().\n    model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),\n                        steps_per_epoch=x_train.shape[0] \/\/ batch_size,\n                        validation_data=datagen.flow(x_validation, y_validation, batch_size=batch_size),\n                        epochs=nb_epoch, verbose=1, max_q_size=100,\n                        validation_steps=x_validation.shape[0]\/\/batch_size,\n                        callbacks=[lr_reducer, early_stopper, csv_logger, model_checkpoint])\n<\/code><\/pre>\n\n<p>So, is there any solution to increase the accuracy in this very difficult scenario for CNNs? I mean, CNN learns features from data and, as you can see, the training and testing data from the same class are different. So, is that possible for me to perform any data augmentation or CNN operation (my CNN has no dropouts and no pooling, as you can see above) that could minimize or simulate the testing data in the training data? <\/p>\n","answers":[{"AnswerId":"52980401","CreationDate":"2018-10-25T02:19:22.977","ParentId":null,"OwnerUserId":"2060205","Title":null,"Body":"<p>If your goal is to predict at multiple zoom levels, you need to train the CNN with multiple zoom levels...\nI think the current augmentation is generating samples which are not what you want. For example, this is one of the images that could be generated when zoom=1.2:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9CEzF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9CEzF.png\" alt=\"augmented_image\"><\/a><\/p>\n\n<p>The simplest solution would be to use a generator like this when training with the 5x5cm^2 patches:<\/p>\n\n<pre><code>ImageDataGenerator(horizontal_flip=True,\n                   vertical_flip=True,\n                   zoom_range=[0.5, 1])\n<\/code><\/pre>\n\n<p>In that case, when zoom=0.5 you will get an image like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/g6aKJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/g6aKJ.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>which is more or less equivalent to a 2.5x2.5cm^2 image.<\/p>\n\n<p>If you have to train it using the 2.5x2.5 patches try:<\/p>\n\n<pre><code>ImageDataGenerator(horizontal_flip=True,\n                   vertical_flip=True,\n                   zoom_range=[1, 2],\n                   fill_mode='constant',\n                   cval=0)\n<\/code><\/pre>\n\n<p>which generate images like this one:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZFX2E.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZFX2E.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>With enough samples and epochs, the CNN should be able to learn that the padding zeros can be ignored.<\/p>\n"}]}
+{"QuestionId":52980291,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><computer-vision>","CreationDate":"2018-10-25T02:03:42.887","AcceptedAnswerId":null,"OwnerUserId":6157515.0,"Title":"Train and Test accuracy not promising on large data set for image classification","Body":"<p>I have around <strong>170K<\/strong> images in training set and <strong>27k<\/strong> in validation set equally  distributed for cats and dogs . I am using below code to classify cats vs dogs . Apart from the below architect, I tried different combinations of epochs\/batch size\/layers<\/p>\n\n<pre><code>img_width,img_height =128,128\nif K.image_data_format() == 'channels_first':\n    input_shape = (3, img_width, img_height)\nelse:\n    input_shape = (img_width, img_height, 3)\n# Initialising the CNN\nclassifier = Sequential()\n\n# Step 1 - Convolution\nclassifier.add(Conv2D(32, (3, 3), input_shape=input_shape, activation='relu'))\n\n# Step 2 - Pooling\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))  # 2x2 is optimal\n\n# Adding a second convolutional layer\nclassifier.add(Conv2D(32, (3, 3), activation='relu'))\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))\n\nclassifier.add(Conv2D(32, (3, 3), activation='relu'))\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))\n\nclassifier.add(Conv2D(64, (3, 3), activation='relu'))\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))\n\n# Step 3 - Flattening\nclassifier.add(Flatten())\n\n# Step 4 - Full connection\n\nclassifier.add(Dense(64, activation='relu'))\nclassifier.add(Dropout(0.4))\nclassifier.add(Dense(1, activation='sigmoid'))\nclassifier.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n\n# Part 2 - Fitting the CNN to the images\nbatch_size =128\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,\n                                   shear_range=0.2,\n                                   zoom_range=0.2,\n                                   horizontal_flip=True,\n                                   fill_mode='nearest')\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntraining_set = train_datagen.flow_from_directory(training_set_path,\n                                                 target_size=input_size,\n                                                 batch_size=batch_size,\n                                                 class_mode='binary',\n                                                 shuffle=True)\n\ntest_set = test_datagen.flow_from_directory(test_set_path,\n                                            target_size=input_size,\n                                            batch_size=batch_size,\n                                            class_mode='binary')\n\n\nclassifier.fit_generator(training_set,\n                         steps_per_epoch=&lt;170K&gt;\/batch_size,\n                         epochs=128,\n                         validation_data=test_set,\n                         validation_steps=&lt;27K&gt;\/batch_size,\n                         use_multiprocessing=True,\n                         workers=12)\n<\/code><\/pre>\n\n<p>The output does not seems promising to me. I need your suggestion\/help to increase training accuracy as well as validation accuracy .<\/p>\n\n<pre><code>Epoch 78: accuracy -&gt; 0.7592, val_accuracy -&gt; 0.6669\nEpoch 79: accuracy -&gt; 0.7602, val_accuracy -&gt; 0.6620\nEpoch 80: accuracy -&gt; 0.7592, val_accuracy -&gt; 0.6579\nEpoch 81: accuracy -&gt; 0.7606, val_accuracy -&gt; 0.6636\nEpoch 82: accuracy -&gt; 0.7583, val_accuracy -&gt; 0.6614\nEpoch 83: accuracy -&gt; 0.7614, val_accuracy -&gt; 0.6455\nEpoch 84: accuracy -&gt; 0.7609, val_accuracy -&gt; 0.6573\nEpoch 85: accuracy -&gt; 0.7611, val_accuracy -&gt; 0.6653\nEpoch 86: accuracy -&gt; 0.7600, val_accuracy -&gt; 0.6631\nEpoch 87: accuracy -&gt; 0.7597, val_accuracy -&gt; 0.6516\nEpoch 88: accuracy -&gt; 0.7606, val_accuracy -&gt; 0.6661\nEpoch 89: accuracy -&gt; 0.7618, val_accuracy -&gt; 0.6639\nEpoch 90: accuracy -&gt; 0.7608, val_accuracy -&gt; 0.6675\nEpoch 91: accuracy -&gt; 0.7622, val_accuracy -&gt; 0.6642\nEpoch 92: accuracy -&gt; 0.7630, val_accuracy -&gt; 0.6621\nEpoch 93: accuracy -&gt; 0.7609, val_accuracy -&gt; 0.6557\nEpoch 94: accuracy -&gt; 0.7630, val_accuracy -&gt; 0.6664\nEpoch 95: accuracy -&gt; 0.7620, val_accuracy -&gt; 0.6664\nEpoch 96: accuracy -&gt; 0.7639, val_accuracy -&gt; 0.6646\nEpoch 97: accuracy -&gt; 0.7638, val_accuracy -&gt; 0.6697\nEpoch 98: accuracy -&gt; 0.7634, val_accuracy -&gt; 0.6633\nEpoch 99: accuracy -&gt; 0.7629, val_accuracy -&gt; 0.6607\nEpoch 100: accuracy -&gt; 0.7639, val_accuracy -&gt; 0.6588\nEpoch 101: accuracy -&gt; 0.7640, val_accuracy -&gt; 0.6616\nEpoch 102: accuracy -&gt; 0.7641, val_accuracy -&gt; 0.6597\nEpoch 103: accuracy -&gt; 0.7656, val_accuracy -&gt; 0.6672\nEpoch 104: accuracy -&gt; 0.7638, val_accuracy -&gt; 0.6650\nEpoch 105: accuracy -&gt; 0.7646, val_accuracy -&gt; 0.6572\nEpoch 106: accuracy -&gt; 0.7648, val_accuracy -&gt; 0.6649\nEpoch 107: accuracy -&gt; 0.7645, val_accuracy -&gt; 0.6630\nEpoch 108: accuracy -&gt; 0.7651, val_accuracy -&gt; 0.6537\nEpoch 109: accuracy -&gt; 0.7668, val_accuracy -&gt; 0.6623\nEpoch 110: accuracy -&gt; 0.7659, val_accuracy -&gt; 0.6620\nEpoch 111: accuracy -&gt; 0.7648, val_accuracy -&gt; 0.6660\nEpoch 112: accuracy -&gt; 0.7660, val_accuracy -&gt; 0.6621\nEpoch 113: accuracy -&gt; 0.7661, val_accuracy -&gt; 0.6714\nEpoch 114: accuracy -&gt; 0.7661, val_accuracy -&gt; 0.6601\nEpoch 115: accuracy -&gt; 0.7639, val_accuracy -&gt; 0.6625\nEpoch 116: accuracy -&gt; 0.7657, val_accuracy -&gt; 0.6676\nEpoch 117: accuracy -&gt; 0.7652, val_accuracy -&gt; 0.6700\nEpoch 118: accuracy -&gt; 0.7669, val_accuracy -&gt; 0.6608\nEpoch 119: accuracy -&gt; 0.7668, val_accuracy -&gt; 0.6711\nEpoch 120: accuracy -&gt; 0.7680, val_accuracy -&gt; 0.6718\nEpoch 121: accuracy -&gt; 0.7681, val_accuracy -&gt; 0.6676\nEpoch 122: accuracy -&gt; 0.7666, val_accuracy -&gt; 0.6565\nEpoch 123: accuracy -&gt; 0.7673, val_accuracy -&gt; 0.6527\nEpoch 124: accuracy -&gt; 0.7670, val_accuracy -&gt; 0.6548\nEpoch 125: accuracy -&gt; 0.7666, val_accuracy -&gt; 0.6602\nEpoch 126: accuracy -&gt; 0.7671, val_accuracy -&gt; 0.6696\nEpoch 127: accuracy -&gt; 0.7680, val_accuracy -&gt; 0.6628\n<\/code><\/pre>\n\n<p>Need your valuable suggestion\/guidance or snippet to try , please.<\/p>\n","answers":[{"AnswerId":"52981862","CreationDate":"2018-10-25T05:26:18.450","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>You have the same filter number for the first, second, and third Convolutional Layers. According to my experience, it is better to increase the filter numbers with the Convolutional Layers, for example 32->64->128 and so on. Then your modal can get more features.<\/p>\n"},{"AnswerId":"52982316","CreationDate":"2018-10-25T06:03:56.477","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>@NormanZhu gave you a good tip. But there are other things you may try. Like adding <code>BatchNormalization<\/code> layer and use <code>LeakyReLU<\/code> instead of ReLU:<\/p>\n\n<pre><code>x = Conv1D(filters=64, kernel_size=1)(inputs)\nx = BatchNormalization()(x)\nx = LeakyReLU(x)\n<\/code><\/pre>\n\n<p>After that you can use <code>MaxPooling<\/code> and that may be the building block for your code. Stack few of those and then add <code>Dense<\/code> layer(s) and you should be good. Maybe even try lowering the dropout.<\/p>\n\n<p>Maybe also try preprocessing images in some way, maybe you can get the accuracy to increase.<\/p>\n"}]}
+{"QuestionId":52980437,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><deep-learning><intel-mkl><google-dl-platform>","CreationDate":"2018-10-25T02:23:34.110","AcceptedAnswerId":null,"OwnerUserId":1238896.0,"Title":"tensorflow performance degradation with intel mkl","Body":"<p>Running wide and linear and deep model inference with plain vanilla tf1.11\nthat comes pre built with deep learning images version M9 for gpu shows much better performance relative to version M10 (for cpu inference)<\/p>\n\n<p>M9 : tf-latest-cu92<\/p>\n\n<p>M10 : tf-latest-cpu<\/p>\n\n<p>In both images tf version is 1.11 and they come prebuilt with intel mkl optimized binaries. \nI turn on the verbosity logging for mkl instructions and on M10 image i see a lot of mkl related settings <\/p>\n\n<pre><code> KMP_AFFINITY=granularity=fine,verbose,compact,1,0\n KMP_BLOCKTIME=0\n KMP_SETTINGS=1\n OMP_NUM_THREADS=32\n<\/code><\/pre>\n\n<p>And logging of mkl instruction with timings. on M9 image i dont observe any such thing even though both images show version info as :<\/p>\n\n<pre><code>MKL_VERBOSE Intel(R) MKL 2019.0 Product build 20180829 for Intel(R) 64 architecture Intel(R) Advanced Vector Extensions 2 (Intel(R) AVX2) enabled processors, Lnx 2.20GHz lp64 intel_thread\nMKL_VERBOSE SDOT(2,0x5622b7736500,1,0x5622b7736500,1) 2.54ms CNR:OFF Dyn:1 FastMM:1 TID:0  NThr:16\n1.11.0\n<\/code><\/pre>\n\n<p>I am seeing 2-4x worse performance when intel mkl instructions are being used as opposed to M9 image.\nNote : even though M9 image is for gpu , i turned of cuda device visibility and benchmarking only cpu inference. Same observation made on another linux box with pip install of tf 1.11 in a clean virtualenv. <\/p>\n\n<p>Any insights on how to debug or get the maximum out of the intel mkl library.<\/p>\n","answers":[{"AnswerId":"54271112","CreationDate":"2019-01-19T20:29:36.933","ParentId":null,"OwnerUserId":"3982224","Title":null,"Body":"<p>This behavior has been fixed in M16+ (that has TF 1.12).<\/p>\n"}]}
+{"QuestionId":52980583,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><tensorboard>","CreationDate":"2018-10-25T02:44:34.287","AcceptedAnswerId":null,"OwnerUserId":5211892.0,"Title":"How to use Tensorflow Keras API","Body":"<p>Well I start learning Tensorflow but I notice there's so much confusion about how to use this thing..\nFirst, some tutorials present models using low level API tf.varibles, scopes...etc, but other tutorials use Keras instead and for example to use tensor board to invoke callbacks.\nSecond, what's the purpose of having ton of duplicate API, really what's the purpose behind  using high level API like Keras when you have low  level to build  model like Lego blocks?\nFinally, what's the true purpose of using eager execution?<\/p>\n","answers":[{"AnswerId":"52981417","CreationDate":"2018-10-25T04:32:58.170","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>You can use these APIs all together. E.g. if you have a regular dense network, but with an special layer you can use higher level API for dense layers (tf.layers and tf.keras) and low level API for your special layer.  Furthermore, it is complex graphs are easier to define in low level APIs, e.g. if you want to share variables, etc. <\/p>\n\n<p>Eager execution helps you for fast debugging, it evaluates tensors directly without a need of invoking a session. <\/p>\n"},{"AnswerId":"52994488","CreationDate":"2018-10-25T16:55:48.613","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>There are different \"levels\" of APIs (high-level APIs such as keras and estimators, and low level APIs such as Variables, etc) to suit different developer needs.<\/p>\n\n<p>For the average industry developer, who already knows approximately what ML model you intend to use, keras is a good fit. For example, if you know you want to implement a sequential model with two dense layers with softmax activation, you need only do something like:<\/p>\n\n<pre><code>model = keras.Sequential([\n  keras.layers.Dense(128, activation=tf.nn.softmax),\n  keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n\n<p>Using keras is generally simpler as you don't have to think about low-level implementation details such as tf.Variables. For more complete examples, check out the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/\" rel=\"nofollow noreferrer\">keras tutorials<\/a> on tensorflow.org.<\/p>\n\n<p>The low-level API allows users finer control over the models you're developing. These APIs are more commonly used by developers and researchers developing novel ML methods; for example, if you need a specialized layer that does something different from canonical ML methods, you can manually define a layer using low level APIs.<\/p>\n\n<p>Finally, eager execution is an imperative programming style. It enables faster debugging, and has a gentler learning curve for those new to tensorflow, since it is more \"pythonic\"\/intuitive. Check out the <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">eager guide<\/a> for more.<\/p>\n"}]}
+{"QuestionId":52980880,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-25T03:22:29.357","AcceptedAnswerId":null,"OwnerUserId":8836663.0,"Title":"Why are the Keras based multiple LSTM model's performance much better than the pure Tensorflow based for Time Series Prediction?","Body":"<p>I trained a 4 keras based layer lstm wrapped with dropout layer stacked model to do time series prediction. And the Result is not too bad, The code is: <\/p>\n\n<pre><code>regressor = keras.models.Sequential()\n\n# input layer\nregressor.add(keras.layers.LSTM(units=50, return_sequences=True, input_shape=(X_train.shape[1], 1)))\nregressor.add(keras.layers.Dropout(0.2))\n\n# second layer\nregressor.add(keras.layers.LSTM(units=50, return_sequences=True))\nregressor.add(keras.layers.Dropout(0.2))\n\n# third layer\nregressor.add(keras.layers.LSTM(units=50, return_sequences=True))\nregressor.add(keras.layers.Dropout(0.2))\n\n# forth layer\nregressor.add(keras.layers.LSTM(units=50))\nregressor.add(keras.layers.Dropout(0.2))\n\n# output layer\nregressor.add(keras.layers.Dense(units=1))\n\nregressor.compile(optimizer='adam', loss='mean_squared_error')\n\nregressor.fit(X_train, y_train, epochs=1, batch_size=32)\n<\/code><\/pre>\n\n<p>The result plot is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nwOrH.jpg\" rel=\"nofollow noreferrer\">Expected Result(Sorry for the link cause my reputation)<\/a>  <\/p>\n\n<p>Than I want to translate the keras based code to pure tensorflow code. <\/p>\n\n<pre><code>X = tf.placeholder(tf.float32, [None, num_time_steps, num_inputs])\ny = tf.placeholder(tf.float32, [None, predict_time_steps, num_outputs])\n\ncell1 = tf.nn.rnn_cell.BasicLSTMCell(num_units=num_units, activation=tf.nn.relu)\ndropout1 = tf.nn.rnn_cell.DropoutWrapper(cell1, input_keep_prob=0.8)\n\ncell2 = tf.nn.rnn_cell.BasicLSTMCell(num_units=num_units, activation=tf.nn.relu)\ndropout2 = tf.nn.rnn_cell.DropoutWrapper(cell2, input_keep_prob=0.8)\n\ncell3 = tf.nn.rnn_cell.BasicLSTMCell(num_units=num_units, activation=tf.nn.relu)\ndropout3 = tf.nn.rnn_cell.DropoutWrapper(cell3, input_keep_prob=0.8)\n\ncell4 = tf.nn.rnn_cell.BasicLSTMCell(num_units=num_units, activation=tf.nn.relu)\ndropout4 = tf.nn.rnn_cell.DropoutWrapper(cell4, input_keep_prob=0.8)\n\nlstm_layers = tf.contrib.rnn.OutputProjectionWrapper(\n    tf.nn.rnn_cell.MultiRNNCell([dropout1, dropout2, dropout3, dropout4]), output_size=num_outputs)\n\noutputs, states = tf.nn.dynamic_rnn(lstm_layers, X, dtype=tf.float32)\n\nloss = tf.reduce_mean(tf.square(outputs - y))\noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\n\ntrain = optimizer.minimize(loss)\n<\/code><\/pre>\n\n<p>But the result is completely wrong. \n<a href=\"https:\/\/i.stack.imgur.com\/Q3WZo.jpg\" rel=\"nofollow noreferrer\">Wrong Result(Sorry for the link cause my reputation)<\/a><\/p>\n\n<p>What's the problem?  Thank you guys.<\/p>\n","answers":[{"AnswerId":"52983374","CreationDate":"2018-10-25T07:11:18.987","ParentId":null,"OwnerUserId":"5324310","Title":null,"Body":"<p>I would recommend you to look into documentations, for example using Keras you have these default settings:<\/p>\n\n<blockquote>\n  <p>keras.layers.LSTM(units, activation='tanh',\n  recurrent_activation='hard_sigmoid', use_bias=True,\n  kernel_initializer='glorot_uniform',\n  recurrent_initializer='orthogonal', bias_initializer='zeros',\n  unit_forget_bias=True, kernel_regularizer=None,\n  recurrent_regularizer=None, bias_regularizer=None,\n  activity_regularizer=None, kernel_constraint=None,\n  recurrent_constraint=None, bias_constraint=None, dropout=0.0,\n  recurrent_dropout=0.0, implementation=1, return_sequences=False,\n  return_state=False, go_backwards=False, stateful=False, unroll=False)<\/p>\n<\/blockquote>\n\n<p>And TensorFlow this: <\/p>\n\n<blockquote>\n  <p><strong>init<\/strong>(\n      num_units,\n      forget_bias=1.0,\n      state_is_tuple=True,\n      activation=None,\n      reuse=None,\n      name=None,\n      dtype=None,\n      **kwargs )<\/p>\n<\/blockquote>\n\n<p>Can you spot the difference? <\/p>\n"}]}
+{"QuestionId":52981044,"AnswerCount":0,"Tags":"<python><ubuntu><tensorflow><magenta>","CreationDate":"2018-10-25T03:46:29.010","AcceptedAnswerId":null,"OwnerUserId":10555444.0,"Title":"Illegal instruction (core dumped) error when running magenta tensorflow","Body":"<p>This is my output from the terminal. Just trying to run a basic example...<\/p>\n\n<pre><code>~$ source activate magenta\n\n(magenta)~$ melody_rnn_generate --config=basic_rnn --bundle_file=~\/basic_rnn.mag --output_dir=~\/melodies --num_outputs=10 --num_steps=128 --primer_melody=\"[60]\"\nIllegal instruction (core dumped)\n\n~$ python --version\nPython 2.7.15 :: Anaconda, Inc.\n\n~$ pip list | grep tensorflow\ntensorflow                         1.8.0      \ntensorflow-tensorboard             1.5.1  \n\n~$ lsb_release -a\nNo LSB modules are available.\nDistributor ID: Ubuntu\nDescription:    Ubuntu 16.04.3 LTS\nRelease:    16.04\nCodename:   xenial\n\n~$ uname -i\nx86_64\n<\/code><\/pre>\n\n<p>Not sure how to proceed... Thanks in advance. let me know if any more info is needed<\/p>\n","answers":[]}
+{"QuestionId":52981219,"AnswerCount":0,"Tags":"<python><tensorflow><optimization><loss-function>","CreationDate":"2018-10-25T04:08:02.207","AcceptedAnswerId":null,"OwnerUserId":10398750.0,"Title":"Two different loss functions with different labels in tensorflow","Body":"<p>I am training a network with two paths, each consisting of several conv layers. One path belongs to video input, and the other path belongs to audio input. At some point, I mix the features and then apply some other conv layers and fully connected layers toward the output of the network.<\/p>\n\n<p>I want to define two separate loss functions: <code>loss1<\/code> just before mixing the audio and video features, and <code>loss2<\/code> at the output. The problem is that the labels are different for each optimization function. \nI have a two class label for the mix layer that indicate whether the video and audio are match, and another two class label at the end of network that indicate certain action in the video.<\/p>\n\n<p>Any suggestions on how I should define my two loss functions, optimizers and train the network in tensorflow? Note that although the labels and loss functions are different, the majority of the network is the same.<\/p>\n\n<p>Thanks,<\/p>\n","answers":[]}
+{"QuestionId":52982574,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-25T06:20:38.927","AcceptedAnswerId":null,"OwnerUserId":2690054.0,"Title":"Calculating the mAP from scores obtained in Tensorflow object detection API","Body":"<p>I am actually working on a server where I run the Tensorflow OD API to train the model for my custom dataset. So, I divide my images into training, validation and test sets and run the train.py on training and validation sets. Next I run the inference using exported model checkpoint and frozen graph on my test images.\nNow, my question is when I run the inference as provided in the example , I get an output dict with detection scores, number of detections, detection classes, detection masks etc for each image, so from these outputs how do I calculate the mAP for my test set?<\/p>\n\n<p>Any guidance in this direction will be really helpful, thanks in advance.<\/p>\n","answers":[{"AnswerId":"52983291","CreationDate":"2018-10-25T07:05:57.627","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>You can use COCO's API for calculating COCO's metrics withing TF OD API. See <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md#coco-api-installation\" rel=\"nofollow noreferrer\">this<\/a>.\nTF feeds COCO's API with your detections and GT, and COCO API will compute COCO's metrics and return it the TF (thus you can display their progress for example in TensorBoard). mAP@0.5 is probably the metric which is most relevant (at it is the standard metric used for PASCAL VOC, Open Images, etc), while mAP@0.5:0.95 is a much more difficult one localization-wise.<\/p>\n"}]}
+{"QuestionId":52982784,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-25T06:34:33.973","AcceptedAnswerId":null,"OwnerUserId":1964720.0,"Title":"Tensorflow classiffier","Body":"<p>I'm developing a classifier with tensorflow, from a series of variables catalog data in various types, the problem is that I have no vision on which variables are significant...<\/p>\n\n<p>It is possible to know from the generated model, which variables have been significant? Is it possible to know their weights?<\/p>\n\n<p>Greetings.<\/p>\n","answers":[]}
+{"QuestionId":52983831,"AnswerCount":2,"Tags":"<python><machine-learning><keras><lstm><rnn>","CreationDate":"2018-10-25T07:36:54.240","AcceptedAnswerId":52984917.0,"OwnerUserId":5182526.0,"Title":"Keras LSTM for continuous input and continuous output","Body":"<p>For example I have binary data, let say: 0, 0, 0, 1, 1, 0, 1, 1. This may continue indefinitely. For each input, there is corresponding output. Let say we use XOR operation. So, output may look like this: 0, 0, 0, 1, 0, 1, 1, 0.<\/p>\n\n<p>How do I shape Keras input shape? How do I set timesteps? If I declare timesteps 1 are for each 1 timestep considered different case or it can still take account of previous input as sequence or learned memory? <\/p>\n\n<p>Keras is using LSTM or GRU for it's hidden layer.<\/p>\n\n<p>I've tried 2 method for this problem, but none seem succeed. Both method stuck at 37.5 acc. In fact, it keep guessing 1.<\/p>\n\n<p>Method 1:<\/p>\n\n<pre><code>data = [[[0], [0], [0], [1], [1], [0], [1], [1]]]\noutput = [[[0], [0], [0], [1], [0], [1], [1], [0]]]\n\nmodel = Sequential()\nmodel.add(GRU(10, input_shape=(8, 1), return_sequences=True))\nmodel.add(GRU(10, return_sequences=True))\nmodel.add(Dense(1, activation='softmax'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['acc'])\nmodel.fit(np.asarray(data), np.asarray(output), epochs=3000)\n<\/code><\/pre>\n\n<p>Method 2:<\/p>\n\n<pre><code>data = [[[0]], [[0]], [[0]], [[1]], [[1]], [[0]], [[1]], [[1]]]\noutput = [[0], [0], [0], [1], [0], [1], [1], [0]]\n\nmodel = Sequential()\nmodel.add(GRU(10, input_shape=(1, 1), return_sequences=True))\nmodel.add(GRU(10))\nmodel.add(Dense(1, activation='softmax'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['acc'])\nmodel.fit(np.asarray(data), np.asarray(output), epochs=300)\n<\/code><\/pre>\n","answers":[{"AnswerId":"52989083","CreationDate":"2018-10-25T12:17:14.870","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>Change <code>activation='softmax'<\/code> to <code>activation='sigmoid'<\/code><\/p>\n"},{"AnswerId":"52984917","CreationDate":"2018-10-25T08:38:40.813","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<blockquote>\n  <p>In fact, it keep guessing 1.<\/p>\n<\/blockquote>\n\n<p>That's because you have used <code>softmax<\/code> as the activation of last layer. Since the last layer have only one unit and the softmax function normalizes its input such that the sum of elements equals to one, it would always outputs 1. Instead, you need to use <code>sigmoid<\/code> as the activation function of last layer to have an output between zero and one.<\/p>\n"}]}
+{"QuestionId":52983880,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-25T07:39:38.727","AcceptedAnswerId":52983881.0,"OwnerUserId":9589731.0,"Title":"tf.case get unexpected results when use lambda","Body":"<p>Look this example.<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.reset_default_graph()\nLENGTH = 25\nM_list = []\nfor i in range(LENGTH):\n    M_list.append(tf.get_variable('M'+str(i), shape=[1], initializer=tf.constant_initializer(i)))\n\nchoose_mat = tf.placeholder(tf.int32, shape=[LENGTH])\ncase_set = [(tf.equal(choose_mat[i], 1), lambda: M_list[i]) for i in range(LENGTH)]\nM = tf.case(case_set)\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    CM1 = [0] * LENGTH\n    CM1[0] = 1\n    CM2 = [0] * LENGTH\n    CM2[1] = 1\n\n    m1 = sess.run(M, feed_dict={choose_mat: CM1})\n    m2 = sess.run(M, feed_dict={choose_mat: CM2})\n    print(m1) # [24.]\n    print(m2) # [24.]\n\n    m1_ = sess.run(M_list[0])\n    m2_ = sess.run(M_list[1])\n    print(m1_) # [0.]\n    print(m2_) # [1.]\n<\/code><\/pre>\n\n<p>We expect m1, m2 is 0, 1\nbut we got 24.\nAnd the result of M_list is right, just like m1_ and m2_, it's strange.<\/p>\n\n<p>Although I have fixed this bug(see my answer), I still have a question, I don't know why this code will cause closure, case_set is not in any function, dose anyone know why this is closure?<\/p>\n","answers":[{"AnswerId":"52983881","CreationDate":"2018-10-25T07:39:38.727","ParentId":null,"OwnerUserId":"9589731","Title":null,"Body":"<p>Infact, this bug is not caused by tensorflow, the real reason is python's closure. <a href=\"https:\/\/stackoverflow.com\/questions\/13355233\/python-lambda-closure-scoping\">see this link<\/a>\nSo this code will get expected result.<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.reset_default_graph()\nLENGTH = 25\nM_list = []\nfor i in range(LENGTH):\n    M_list.append(tf.get_variable('M'+str(i), shape=[1], initializer=tf.constant_initializer(i)))\n\nchoose_mat = tf.placeholder(tf.int32, shape=[LENGTH])\ncase_set = [(tf.equal(choose_mat[i], 1), lambda i=i: M_list[i]) for i in range(LENGTH)]\nM = tf.case(case_set)\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    CM1 = [0] * LENGTH\n    CM1[0] = 1\n    CM2 = [0] * LENGTH\n    CM2[1] = 1\n\n    m1 = sess.run(M, feed_dict={choose_mat: CM1})\n    m2 = sess.run(M, feed_dict={choose_mat: CM2})\n    print(m1) # [0.]\n    print(m2) # [1.]\n<\/code><\/pre>\n\n<p>Although I have fixed this bug, I still don't know why this code will cause closure, case_set is not in any function, dose anyone know why this is closure?<\/p>\n"}]}
+{"QuestionId":52983925,"AnswerCount":0,"Tags":"<linux><tensorflow><pip><avx2><fma>","CreationDate":"2018-10-25T07:42:19.537","AcceptedAnswerId":null,"OwnerUserId":8278545.0,"Title":"tensorflow-1.12.0rc1-cp27-cp27mu-linux_x86_64.whl is not a supported wheel on this platform","Body":"<p>I installed tensor flow on Intel NUC with pip3\npip3 install --upgrade tensor flow\nBut got below error \n2018-10-25 20:14:31.685641: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\ncannot open X server<\/p>\n\n<p>After googling found that i need to build tensor flow from source. So i installed bazel version - 0.18\nGit clone tensor flow and checked out 1.12 version. Further used below commands:\nbazel build -c opt --copt=-mavx --copt=-msse4.1 --copt=-msse4.2 -k \/\/tensorflow\/tools\/pip_package:build_pip_package<\/p>\n\n<p>bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg<\/p>\n\n<p>pip3 install \/tmp\/tensorflow_pkg\/NAME_OF_WHEEL.whl<\/p>\n\n<p>and i got error \ntensorflow-1.12.0rc1-cp27-cp27mu-linux_x86_64.whl is not a supported wheel on this platform.<\/p>\n\n<p>pip3 --version\npip 9.0.3 from \/usr\/local\/lib\/python3.5\/dist-packages (python 3.5)<\/p>\n\n<p>sudo pip3 --version\npip 9.0.3 from \/usr\/local\/lib\/python3.5\/dist-packages (python 3.5)<\/p>\n\n<p>How to solve this issue?<\/p>\n","answers":[]}
+{"QuestionId":52984417,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2018-10-25T08:12:19.363","AcceptedAnswerId":52987508.0,"OwnerUserId":9422637.0,"Title":"Keras - Save output of intermediate layer","Body":"<p><strong>What I have done?<\/strong><\/p>\n\n<p>I'm using from this code to implement my keras'model:<\/p>\n\n<pre><code>X, tx, Y, ty = train_test_split(X, Y, test_size=0.2, random_state=np.random.seed(7), shuffle=True)\n\nX = np.reshape(X, (X.shape[0], 1, X.shape[1]))\ntx = np.reshape(tx, (tx.shape[0], 1, tx.shape[1]))\n\nmodel = Sequential()\nmodel.add(LSTM(100, return_sequences=False, input_shape=(X.shape[1], X.shape[2])))\nmodel.add(Dense(Y.shape[1], activation='softmax'))\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\nfilepath=\"weights-{epoch:02d}-{val_acc:.2f}.hdf5\"\ncheckpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\ncallbacks_list = [checkpoint]\n\nmodel.fit(X, Y, validation_split=.20,\n                    epochs=1000, batch_size=50, callbacks=callbacks_list, verbose=0) \n<\/code><\/pre>\n\n<p><strong>OUTPUT:<\/strong><\/p>\n\n<p>Below is a part of the program's output:<\/p>\n\n<pre><code>Epoch 00993: val_acc did not improve\nEpoch 00994: val_acc did not improve\nEpoch 00995: val_acc did not improve\nEpoch 00996: val_acc did not improve\nEpoch 00997: val_acc did not improve\nEpoch 00998: val_acc improved from 0.93900 to 0.94543, saving model to weights-998-0.94.hdf5\nEpoch 00999: val_acc did not improve\n<\/code><\/pre>\n\n<p><strong>PROBLEM:<\/strong> <\/p>\n\n<p>I need to save the output of LSTM layer in any epoch but i do not know how?<\/p>\n\n<p>Any idea?<\/p>\n","answers":[{"AnswerId":"52987508","CreationDate":"2018-10-25T10:51:34.647","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You can use <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">Keras functional API<\/a>. You will have to rewrite your model creation but it's not much work.  Then when you write something like this:<\/p>\n\n<pre><code>lstm_output = LSTM(128, ...)(x)\n<\/code><\/pre>\n\n<p>the output of the LSTM layer will be in the <code>lstm_output<\/code> variable and you can save it in every iteration of every epoch. <\/p>\n\n<p>I hope this answers your question. <\/p>\n"}]}
+{"QuestionId":52984932,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-10-25T08:39:28.637","AcceptedAnswerId":53047796.0,"OwnerUserId":3941809.0,"Title":"TensorFlow save and restore the trained well CNN model was not working","Body":"<p>I used TensorFlow with python3 to build a CNN(Convolutional Neural Network) model to train and predict the MNIST handwritten digit database.<\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\n<\/code><\/pre>\n\n<p>I train my CNN the MNIST train datasbase and predict the MNIST test datasbase. The accuracy of my model was over 95%. It's ok.<\/p>\n\n<pre><code>import tensorflow as tf\nimport os\n\nfrom tensorflow.examples.tutorials.mnist import input_data\n\ntf.logging.set_verbosity(tf.logging.ERROR)\n\n\ndef compute_accuracy(v_xs, v_ys):\n    global prediction\n    y_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1})\n    correct_prediction = tf.equal(tf.argmax(y_pre, 1), tf.argmax(v_ys, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n    result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})\n    return result\n\n\ndef weight_variable(shape):\n    initial = tf.truncated_normal(shape, stddev=0.1)\n    return tf.Variable(initial)\n\n\ndef bias_variable(shape):\n    initial = tf.constant(0.1, shape=shape)\n    return tf.Variable(initial)\n\n\ndef conv2d(x, W):\n    return tf.nn.conv2d(input=x, filter=W, strides=[1, 1, 1, 1], padding='SAME')\n\n\ndef max_pool_2x2(x):\n    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n\n\nIMAGE_WIDTH = 28\nIMAGE_HEIGHT = 28\nCHANNEL_COUNT = 1\nCONV1_FEATURE_MAP_COUNT = 32\nCONV1_FILTER_HEIGHT = 5\nCONV1_FILTER_WEIGHT = 5\nCONV2_FILTER_HEIGHT = 5\nCONV2_FILTER_WEIGHT = 5\nCONV2_FEATURE_MAP_COUNT = 64\nFULL_CONNECTED_OUTPUT_SIZE = 1024\nOUTPUT_TYPE_COUNT = 10\n\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n\npb_file_dir = \"{path}{sep}pb_modelsaved_model\".format(path=os.getcwd(), sep=os.path.sep)\nckpt_file_dir = \"{path}{sep}ckpt_model{sep}model.ckpt\".format(path=os.getcwd(), sep=os.path.sep)\n\nwith tf.name_scope('input'):\n    xs = tf.placeholder(tf.float32, [None, IMAGE_WIDTH * IMAGE_HEIGHT], name=\"images\") \/ 255.  # 28x28\n    ys = tf.placeholder(tf.float32, [None, 10], name=\"labels\")\n    keep_prob = tf.placeholder(tf.float32, name=\"keep_prob\")\n\nx_image = tf.reshape(xs, [-1, IMAGE_HEIGHT, IMAGE_WIDTH, CHANNEL_COUNT])\n\nwith tf.name_scope(\"conv_layer1\"):\n    with tf.name_scope('weights'):\n        W_conv1 = weight_variable([CONV1_FILTER_HEIGHT, CONV1_FILTER_WEIGHT, CHANNEL_COUNT, CONV1_FEATURE_MAP_COUNT])\n    with tf.name_scope('biases'):\n        b_conv1 = bias_variable([CONV1_FEATURE_MAP_COUNT])\n    with tf.name_scope('conv'):\n        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)\n    with tf.name_scope('pool'):\n        h_pool1 = max_pool_2x2(h_conv1)\n\nwith tf.name_scope(\"conv_layer2\"):\n    with tf.name_scope('weights'):\n        W_conv2 = weight_variable(\n            [CONV2_FILTER_HEIGHT, CONV2_FILTER_WEIGHT, CONV1_FEATURE_MAP_COUNT, CONV2_FEATURE_MAP_COUNT])\n    with tf.name_scope('biases'):\n        b_conv2 = bias_variable([CONV2_FEATURE_MAP_COUNT])\n    with tf.name_scope('conv'):\n        h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)\n    with tf.name_scope('pool'):\n        h_pool2 = max_pool_2x2(h_conv2)\n\nwith tf.name_scope(\"fc_layer1\"):\n    with tf.name_scope('weights'):\n        W_fc1 = weight_variable([7 * 7 * 64, FULL_CONNECTED_OUTPUT_SIZE])\n    with tf.name_scope('biases'):\n        b_fc1 = bias_variable([FULL_CONNECTED_OUTPUT_SIZE])\n    with tf.name_scope('output'):\n        h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])\n        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)\n\nwith tf.name_scope(\"fc_layer2\"):\n    with tf.name_scope('weights'):\n        W_fc2 = weight_variable([FULL_CONNECTED_OUTPUT_SIZE, OUTPUT_TYPE_COUNT])\n    with tf.name_scope('biases'):\n        b_fc2 = bias_variable([OUTPUT_TYPE_COUNT])\n    with tf.name_scope('output'):\n        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob, name=\"drop\")\n        prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2, name=\"prediction\")\n\nwith tf.name_scope('loss'):\n    cross_entropy = tf.reduce_mean(\n        -tf.reduce_sum(ys * tf.log(prediction), reduction_indices=[1]),\n        name=\"loss\")\n\nwith tf.name_scope('train'):\n    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy, name=\"train_step\")\n\nsaver = tf.train.Saver()\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(1000):\n        batch_xs, batch_ys = mnist.train.next_batch(100)\n        sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5})\n        if i % 50 == 0:\n            accuracy = compute_accuracy(mnist.test.images[:1000], mnist.test.labels[:1000])\n    save_path = saver.save(sess, ckpt_file_dir)\n<\/code><\/pre>\n\n<p>However, When I try to save my model and restore my model to predict the MNIST test datasbase. The accuracy of my model was 10%!<\/p>\n\n<pre><code>import tensorflow as tf\nimport os\n\nfrom tensorflow.examples.tutorials.mnist import input_data\n\ntf.logging.set_verbosity(tf.logging.ERROR)\n\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n\nckpt_file_dir = \"{path}{sep}ckpt_model{sep}model.ckpt\".format(path=os.getcwd(), sep=os.path.sep)\n\n\ndef compute_accuracy(v_xs, v_ys):\n    global prediction\n    y_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1})\n    correct_prediction = tf.equal(tf.argmax(y_pre, 1), tf.argmax(v_ys, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n    result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})\n    return result\n\n\ntf.reset_default_graph()\n\nwith tf.Session() as sess:\n    saver = tf.train.import_meta_graph(ckpt_file_dir + \".meta\")\n    saver.restore(sess, ckpt_file_dir)\n    xs = sess.graph.get_tensor_by_name('input\/images:0')\n    ys = sess.graph.get_tensor_by_name('input\/labels:0')\n    keep_prob = sess.graph.get_tensor_by_name('input\/keep_prob:0')\n    prediction = sess.graph.get_tensor_by_name('fc_layer2\/output\/prediction:0')\n\n    cross_entropy = sess.graph.get_tensor_by_name('loss\/loss:0')\n\n    print(compute_accuracy(mnist.test.images[:1000], mnist.test.labels[:1000]))\n<\/code><\/pre>\n\n<p>It's very difficult for me because I'm new to TensorFlow. I tried to print all variable in my saved model and find it's same as the model after my training.<\/p>\n\n<p>And I also try to save my model with \"pb\" file and the accuracy of my model was also 10%!<\/p>\n\n<p>I'm obsessed with this problem and hope you can help me to solve it! Thank you!<\/p>\n\n<hr>\n\n<p>I try to print all weights name in my model:<\/p>\n\n<pre><code>W_conv1 conv_layer1\/weights\/Variable:0\nW_conv2 conv_layer2\/weights\/Variable:0\nW_fc1 fc_layer1\/weights\/Variable:0\nW_fc2 fc_layer2\/weights\/Variable:0\n<\/code><\/pre>\n\n<p>I try to print the result of label value and predict value:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lF8os.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lF8os.png\" alt=\"the result of the CNN model has trained\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/U6P6z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/U6P6z.png\" alt=\"the result of the CNN model restored\"><\/a><\/p>\n","answers":[{"AnswerId":"53047796","CreationDate":"2018-10-29T14:36:08.610","ParentId":null,"OwnerUserId":"3941809","Title":null,"Body":"<p>I find the problem of my code.<\/p>\n\n<p>I change this line:<\/p>\n\n<pre><code>xs = tf.placeholder(tf.float32, [None, IMAGE_WIDTH * IMAGE_HEIGHT], name=\"images\") \/ 255\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>xs = tf.placeholder(tf.float32, [None, IMAGE_WIDTH * IMAGE_HEIGHT], name=\"images\")\n<\/code><\/pre>\n\n<p>That's all right!<\/p>\n"}]}
+{"QuestionId":52985676,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-25T09:16:33.483","AcceptedAnswerId":52986812.0,"OwnerUserId":9589731.0,"Title":"How to use tf.case efficiently?","Body":"<p>I want use tf.case to choose different weight in my network, but it is so inefficiency!\nFor example:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom datetime import datetime\n\ntf.reset_default_graph()\n\nM_list = []\ntime0 = datetime.now()\nLENGTH = 100\n\nfor i in range(LENGTH):\n    M_list.append(tf.get_variable('M'+str(i), shape=[10, 10], initializer=tf.constant_initializer(i)))\n\nwith tf.Session() as sess:\n    Ma = tf.get_variable('Ma', shape=[10, 1000], initializer=tf.constant_initializer(1))\n    choose_mat = tf.placeholder(tf.int32, shape=[LENGTH])\n    case_set = [(tf.equal(choose_mat[i], 1), lambda i=i: tf.matmul(M_list[i], Ma)) for i in range(LENGTH)]\n    Mo = tf.case(case_set)\n    sess.run(tf.global_variables_initializer())\n\n    time1 = datetime.now()\n    create_time = time1 - time0\n    print('create time: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n\n    for i in range(LENGTH):\n        CM = [0] * LENGTH\n        CM[i] = 1\n        mo = sess.run(Mo, feed_dict={choose_mat: CM})\n\n    time2 = datetime.now()\n    run_time = time2 - time1\n    total_time = time2 - time0\n    print('run time: ', str(run_time.seconds) + '.' + str(run_time.microseconds).zfill(6))\n    print('total time: ', str(total_time.seconds) + '.' + str(total_time.microseconds).zfill(6))\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>create time:  23.969327\nrun time:  12.362408\ntotal time:  36.331735\n<\/code><\/pre>\n\n<p>I know tf.case will compute all of the branches in case_set.\nSo I choose weight before compute matmul, like this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom datetime import datetime\n\ntf.reset_default_graph()\n\nM_list = []\ntime0 = datetime.now()\nLENGTH = 100\n\nfor i in range(LENGTH):\n    M_list.append(tf.get_variable('M'+str(i), shape=[10, 10], initializer=tf.constant_initializer(i)))\n\n\nwith tf.Session() as sess:\n    choose_mat = tf.placeholder(tf.int32, shape=[LENGTH])\n    case_set = [(tf.equal(choose_mat[i], 1), lambda i=i: M_list[i]) for i in range(LENGTH)]\n    M = tf.case(case_set)\n\n    Ma = tf.get_variable('Ma', shape=[10, 1000], initializer=tf.constant_initializer(1))\n    Mo = tf.matmul(M, Ma)\n    sess.run(tf.global_variables_initializer())\n\n    time1 = datetime.now()\n    create_time = time1 - time0\n    print('create time2: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n\n    for i in range(LENGTH):\n        CM = [0] * LENGTH\n        CM[i] = 1\n        mo = sess.run(Mo, feed_dict={choose_mat: CM})\n\n    time2 = datetime.now()\n    create_time = time2 - time1\n    total_time = time2 - time0\n    print('run time: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n    print('total time: ', str(total_time.seconds) + '.' + str(total_time.microseconds).zfill(6))\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>create time2:  23.321199\nrun time:  5.747378\ntotal time:  29.068577\n<\/code><\/pre>\n\n<p>It's faster, seem to have some effect, but it's still slow. If we don't use tf.case, like this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom datetime import datetime\n\ntf.reset_default_graph()\n\nM_list = []\ntime0 = datetime.now()\nLENGTH = 100\n\nfor i in range(LENGTH):\n    M_list.append(tf.get_variable('M'+str(i), shape=[10, 10], initializer=tf.constant_initializer(i)))\n\nwith tf.Session() as sess:\n    Ma = tf.get_variable('Ma', shape=[10, 1000], initializer=tf.constant_initializer(1))\n    choose_mat = tf.placeholder(tf.int32, shape=[LENGTH])\n    Mo_list = [tf.matmul(M_list[i], Ma) for i in range(LENGTH)]\n\n    sess.run(tf.global_variables_initializer())\n\n    time1 = datetime.now()\n    create_time = time1 - time0\n    print('create time: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n\n    for i in range(LENGTH):\n        CM = [0] * LENGTH\n        CM[i] = 1\n        mo = sess.run(Mo_list[i], feed_dict={choose_mat: CM})\n\n    time2 = datetime.now()\n    run_time = time2 - time1\n    total_time = time2 - time0\n    print('run time: ', str(run_time.seconds) + '.' + str(run_time.microseconds).zfill(6))\n    print('total time: ', str(total_time.seconds) + '.' + str(total_time.microseconds).zfill(6))\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>create time:  0.547081\nrun time:  0.596932\ntotal time:  1.144013\n<\/code><\/pre>\n\n<p>So the runtime with tf.case is more than ten times of runtime without tf.case!\nBut I want select weight by tensor, so how to use tf.case efficiently? Or dose any more efficient method exist?<\/p>\n\n<p>Great thanks in advance.<\/p>\n","answers":[{"AnswerId":"52986812","CreationDate":"2018-10-25T10:14:10.087","ParentId":null,"OwnerUserId":"9589731","Title":null,"Body":"<p>I found a efficient method, and DON'T USE tf.case! It's stupid.<\/p>\n\n<p>This way can select weights by tensor efficiently:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom datetime import datetime\n\ntf.reset_default_graph()\n\ntime0 = datetime.now()\nLENGTH = 100\nM_list = [tf.get_variable('M'+str(i), shape=[10, 10, 1], initializer=tf.constant_initializer(i)) for i in range(LENGTH)]\nM_concat = tf.concat(M_list, axis=2, name='M_concat')\n\nwith tf.Session() as sess:\n    choose_mat = tf.placeholder(tf.int32, shape=[1])\n    M = M_concat[:, :, choose_mat[0]]\n    M = tf.squeeze(M)\n\n    Ma = tf.get_variable('Ma', shape=[10, 1000], initializer=tf.constant_initializer(1))\n    Mo = tf.matmul(M, Ma)\n\n    sess.run(tf.global_variables_initializer())\n\n    time1 = datetime.now()\n    create_time = time1 - time0\n    print('create time: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n\n    for i in range(LENGTH):\n        mo = sess.run(Mo, feed_dict={choose_mat: [i]})\n\n    time2 = datetime.now()\n    create_time = time2 - time1\n    total_time = time2 - time0\n    print('run time: ', str(create_time.seconds) + '.' + str(create_time.microseconds).zfill(6))\n    print('total time: ', str(total_time.seconds) + '.' + str(total_time.microseconds).zfill(6))\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>create time:  0.540483\nrun time:  0.085812\ntotal time:  0.626295\n<\/code><\/pre>\n\n<p>I don't know what tf.case is use for since it's so slow. But we do need if or switch to control flow, I think it should have been more efficient. <\/p>\n"}]}
+{"QuestionId":52986667,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-25T10:07:43.323","AcceptedAnswerId":null,"OwnerUserId":4829166.0,"Title":"How to reuse rnn cell for inference","Body":"<p>Some of my graph definition is for training. Looks like this<\/p>\n\n<pre><code>with tf.variable_scope('RNN', initializer=tf.contrib.layers.xavier_initializer()):\n     self.rnn_cell = tf.nn.rnn_cell.GRUCell(HID_SZ, name='rnncell')\n     self.init_state = tf.get_variable('init', [1, HID_SZ], tf.float32)\n\n     self.init_state_train = tf.tile(self.init_state, [SZ_BATCH, 1])\n\n     outputs, state = tf.nn.dynamic_rnn(self.rnn_cell, emb, initial_state=self.init_state_train, dtype=tf.float32, time_major=True)\n<\/code><\/pre>\n\n<p>Then I have the portion defined for inference. Currently it looks like<\/p>\n\n<pre><code>with tf.variable_scope(\"\", reuse=True):\n    [...]\n    self.rnn_infer = tf.get_variable('RNN\/rnncell')\n    inputs_single = tf.expand_dims(emb_single, 0)\n    input_state_ = tf.expand_dims(self.input_state, 0)\n    output, hidden = self.rnn_infer(inputs_single, input_state_, name='rnncall')\n<\/code><\/pre>\n\n<p>But the <code>tf.get_variable('RNN\/rnncell')<\/code> results in an error:<\/p>\n\n<pre><code>ValueError: You can only pass an initializer function that expects no arguments to its callable when the shape is not fully defined. The given initializer function expects the following args ['self', 'shape', 'dtype', 'partition_info']\n<\/code><\/pre>\n\n<p>I'm trying to reuse the variable that I assigned to <code>self.rnn_cell<\/code> for inference, how can I do that?<\/p>\n","answers":[{"AnswerId":"52997200","CreationDate":"2018-10-25T20:00:44.477","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>The keypoint is that when you make a cell and put it into an rnn, the weights, and ops, are created on the graph as usual. So you can just restore the weights as you usually would.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport os\n\n\ndef build_and_train():\n    HID_SZ = 1\n    graph = tf.Graph()\n    sess = tf.Session(graph=graph)\n    ones = np.ones([2, 3])\n\n    with graph.as_default():\n        in_ = tf.placeholder(tf.float32, [2, 3])\n        cell = tf.nn.rnn_cell.GRUCell(HID_SZ, name='rnncell')\n        state = tf.zeros([2, HID_SZ])\n        out, state = cell(in_, state)\n        sess.run(tf.global_variables_initializer())\n        saver = tf.train.Saver()\n\n    saver.save(sess, os.getcwd() + '\\\\model.ckpt')\n    print('Cell output after training')\n    print(sess.run(out, feed_dict={in_:ones}))\n\ndef infer():\n    HID_SZ = 1\n    graph = tf.Graph()\n    sess = tf.Session(graph=graph)\n    ones = np.ones([2, 3])\n\n    with graph.as_default():\n        in_ = tf.placeholder(tf.float32, [2, 3])\n        cell = tf.nn.rnn_cell.GRUCell(HID_SZ, name='rnncell')\n        state = tf.zeros([2, HID_SZ])\n        out, state = cell(in_, state)\n        sess.run(tf.global_variables_initializer())\n        saver = tf.train.Saver()\n\n    print('random cell output')\n    print(sess.run(out, feed_dict={in_:ones}))\n\n    saver.restore(sess, 'model.ckpt')\n\n    print('Trained cell output')\n    print(sess.run(out, feed_dict={in_:ones}))\n\n\nbuild_and_train()\ninfer()\n<\/code><\/pre>\n\n<p>This outputs:<\/p>\n\n<pre><code>Cell output after training\n[[0.02710133]\n [0.02710133]]\nrandom cell output\n[[0.2458247]\n [0.2458247]]\nTrained cell output\n[[0.02710133]\n [0.02710133]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":52986850,"AnswerCount":0,"Tags":"<gpu><theano><google-colaboratory><theano-cuda>","CreationDate":"2018-10-25T10:16:04.623","AcceptedAnswerId":null,"OwnerUserId":10540633.0,"Title":"How can I solve theano GPU pygpu error on Google Colab?","Body":"<p>I'm using theano gpu on Google Colab and I get this error:<\/p>\n\n<pre><code>ERROR (theano.gpuarray): pygpu was configured but could not be imported or is too old (version 0.7 or higher required)\n<\/code><\/pre>\n\n<p>I've set:<\/p>\n\n<pre><code>import os\nos.environ[\"THEANO_FLAGS\"]=\"device=cuda, floatX=float32\"\n<\/code><\/pre>\n\n<p>And changed the runtime type to hardware accelerator 'GPU'<\/p>\n\n<p>Could you please help me solving this issue?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":52987835,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-25T11:08:14.153","AcceptedAnswerId":53030878.0,"OwnerUserId":6479506.0,"Title":"Keras flow_from_dataframe wrong data ordering","Body":"<p>I am using keras's data generator with flow_from_dataframe. for training it works just fine, but when using model.predict_generator on the test set, I discovered that the ordering of the generated results is different than the ordering of the \"id\" column in my dataframe.<\/p>\n\n<p>shuffle=False does make the ordering of the generator consistent, but it is a different ordering than the dataframe. I also tried different batch sizes and the corresponding correct steps for the predict_generator function. (for example: batch_Size=1, steps=len(data))<\/p>\n\n<p>how can I make sure the labels predicted for my test set are ordered in the same way of my dataframe \"id\" column?<\/p>\n","answers":[{"AnswerId":"53030878","CreationDate":"2018-10-28T11:14:09.423","ParentId":null,"OwnerUserId":"6479506","Title":null,"Body":"<p>While I haven't found a way to decide the order in which the generator produces data, the order can be obtained with the generator.filenames property.<\/p>\n"}]}
+{"QuestionId":52988311,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-25T11:33:46.093","AcceptedAnswerId":null,"OwnerUserId":6544207.0,"Title":"What is the difference between keras and keras-gpu?","Body":"<p>I am setting up my computer to run DL with a GPU and I couldn't find info on whether one should install keras or keras-gpu. Currently I have it running with conda and keras using tensorflow-gpu as backend. What would be the difference if I switch keras to keras-gpu? <\/p>\n","answers":[{"AnswerId":"53216758","CreationDate":"2018-11-08T21:55:56.383","ParentId":null,"OwnerUserId":"6244155","Title":null,"Body":"<p>this is a paragraph borrowed from Wikipedia:<br>\nKeras was conceived to be an interface rather than a standalone machine-learning framework. It offers a higher-level, more intuitive set of abstractions that make it easy to develop deep learning models regardless of the computational backend used.<br>\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Keras\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Keras<\/a> <br>\nSo as long as all three back ends of Keras (TensorFlow, Microsoft Cognitive Toolkit, or Theano) do not support Gpu versions, using keras-gpu may cause you trouble.<\/p>\n"}]}
+{"QuestionId":52988876,"AnswerCount":1,"Tags":"<pytorch><backpropagation>","CreationDate":"2018-10-25T12:05:15.283","AcceptedAnswerId":null,"OwnerUserId":9316652.0,"Title":"How can I visualize what happens during loss.backward()?","Body":"<p>I am confident in my understanding of the forward pass of my model, how can I control its backward pass?<\/p>\n\n<p>This is not a theoretical question about what back-propagation is. The question is a practical one, about whether or not there are tools suited to visualize\/track\/control what happens during back-propagation.<\/p>\n\n<p>Ideally, this tool would allow to visualize the structure of the computational graph of the model (a graph of the model's operations), its inputs and its trainable parameters.<\/p>\n\n<p>Now, I do:<\/p>\n\n<pre><code>loss.backward()\n<\/code><\/pre>\n\n<p>and I would like to visualize what happens in that step.<\/p>\n","answers":[{"AnswerId":"53024678","CreationDate":"2018-10-27T17:47:59.923","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>There has been been already mention of <code>pytorchviz<\/code> which lets you visualize the graph. <\/p>\n\n<p>Here is a small example that might help you to understand how <code>pytorchviz<\/code> does trace the graph using the <code>grad_fn<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nfrom torch import nn\nd = 5\nx = torch.rand(d, requires_grad=True)\nprint('Tensor x:', x)\ny = torch.ones(d, requires_grad=True)\nprint('Tensor y:', y)\nloss = torch.sum(x*y)*3\n\ndel x\nprint()\nprint('Tracing back tensors:')\ndef getBack(var_grad_fn):\n    print(var_grad_fn)\n    for n in var_grad_fn.next_functions:\n        if n[0]:\n            try:\n                tensor = getattr(n[0], 'variable')\n                print(n[0])\n                print('Tensor with grad found:', tensor)\n                print(' - gradient:', tensor.grad)\n                print()\n            except AttributeError as e:\n                getBack(n[0])\n\nloss.backward()\ngetBack(loss.grad_fn)\n<\/code><\/pre>\n\n<p><em>Output:<\/em><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Tensor x: tensor([0.0042, 0.5376, 0.7436, 0.2737, 0.4848], requires_grad=True)\nTensor y: tensor([1., 1., 1., 1., 1.], requires_grad=True)\n\nTracing back tensors:\n&lt;MulBackward object at 0x1201bada0&gt;\n&lt;SumBackward0 object at 0x1201bacf8&gt;\n&lt;ThMulBackward object at 0x1201bae48&gt;\n&lt;AccumulateGrad object at 0x1201badd8&gt;\nTensor with grad found: tensor([0.0042, 0.5376, 0.7436, 0.2737, 0.4848], requires_grad=True)\n - gradient: tensor([3., 3., 3., 3., 3.])\n\n&lt;AccumulateGrad object at 0x1201bad68&gt;\nTensor with grad found: tensor([1., 1., 1., 1., 1.], requires_grad=True)\n - gradient: tensor([0.0125, 1.6129, 2.2307, 0.8211, 1.4543])\n<\/code><\/pre>\n\n<p>Further you should definately take a look into how <em>autograd functions<\/em> (that are used by the <code>backward()<\/code>-function) are actually work !<br>\nHere is a tutorial from the pytorch site with an easy and short example:<\/p>\n\n<p><strong><em><a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/examples_autograd\/two_layer_net_custom_function.html\" rel=\"nofollow noreferrer\">PyTorch: Defining New autograd Functions<\/a><\/em><\/strong><\/p>\n\n<p>Hope this helps a bit!<\/p>\n"}]}
+{"QuestionId":52989085,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-25T12:17:20.087","AcceptedAnswerId":null,"OwnerUserId":6341510.0,"Title":"Prevent predict_generator from shuffling batches","Body":"<p>I am trying to check the performance of my model on the validation-dataset. As such, I am using <code>predict_generator<\/code> to return predictions from my <code>validation_generator<\/code>. However, I am not able to match the predictions with true labels returned from <code>validation_generator.classes<\/code> since the order of my predictions is mixed up. <\/p>\n\n<p>This is how I initialize my generator:<\/p>\n\n<pre><code>BATCH_SIZE = 64\ndata_generator = ImageDataGenerator(rescale=1.\/255, \n                                    validation_split=0.20)\n\ntrain_generator = data_generator.flow_from_directory(main_path, target_size=(IMAGE_HEIGHT, IMAGE_SIZE), shuffle=False, seed=13,\n                                                     class_mode='categorical', batch_size=BATCH_SIZE, subset=\"training\")\n\nvalidation_generator = data_generator.flow_from_directory(main_path, target_size=(IMAGE_HEIGHT, IMAGE_SIZE), shuffle=False, seed=13,\n                                                     class_mode='categorical', batch_size=BATCH_SIZE, subset=\"validation\")\n\n#Found 4473 images belonging to 3 classes.\n#Found 1116 images belonging to 3 classes.\n<\/code><\/pre>\n\n<p>Now I am using the predict_generator like so:<\/p>\n\n<pre><code>validation_steps_per_epoch = np.math.ceil(validation_generator.samples \/ validation_generator.batch_size)\n\npredictions = model.predict_generator(validation_generator, steps=validation_steps_per_epoch)\n<\/code><\/pre>\n\n<p>I realize that there is a mismatch between my validation-data size (=1116) and validation_steps_per_epoch (=1152). Since these two dont match, I find the output <code>predictions<\/code> is different each time I run model.predict_generator(...).<\/p>\n\n<p>Is there any way to fix this besides changing batch_size to 1 in order to make sure that generator steps through all samples?<\/p>\n\n<p>I found someone with a similar issue here <a href=\"https:\/\/stackoverflow.com\/questions\/51088826\/keras-predict-generator-is-shuffling-its-output-when-using-a-keras-utils-sequenc\">keras predict_generator is shuffling its output when using a keras.utils.Sequence<\/a>, however his solution does not fix my problem since I am not writing any custom functions. <\/p>\n","answers":[{"AnswerId":"52989511","CreationDate":"2018-10-25T12:39:28.630","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>There is no randomization or shuffling going on, what happens is that since the batch size of the validation generator does not exactly divide the number of samples, then the leftover samples spill into the next time the generator is called, which messes up everything.<\/p>\n\n<p>What you could do is set a batch size for the validation generator that divides exactly the number of validation samples, or set the batch size to one.<\/p>\n"}]}
+{"QuestionId":52990569,"AnswerCount":0,"Tags":"<python><keras><time-series><anomaly-detection>","CreationDate":"2018-10-25T13:32:15.160","AcceptedAnswerId":null,"OwnerUserId":10557874.0,"Title":"Multivariate time series, detect anomalies, Keras","Body":"<p>I am currently facing a task in which I need to <strong>recognize the presence of anomalies in instances, each described by multiple time series<\/strong>.\nEach anomaly may be 10 seconds long, or more (typically, less than a couple of minutes).<\/p>\n\n<p>Specifically, this is the situation:<\/p>\n\n<ul>\n<li>each instance is characterized by 10 time series<\/li>\n<li>given an instance, each time series describes the same time interval (typically, 8-10 hours)<\/li>\n<li>all instances are characterized by the same kind of 10 time series<\/li>\n<li>each time series has a sampling rate (number of observations per second)<\/li>\n<li>the class labels are very unbalanced: anomalies may cover around 5% of the entire time interval described by the series, if present at all<\/li>\n<li>finally, for each instance, <strong>I have a hand-made labelling, telling me which time intervals are \"anomalous\"<\/strong><\/li>\n<\/ul>\n\n<p>I would like to train a model on a set of labelled instances, capable of recognizing the presence of anomalous intervals in a new, previously unobserved, instance.<\/p>\n\n<p>The first approach that I have tried is based on windowing: given a training instance, I simply apply a \"moving window\" on each time series (e.g., 10 seconds long), and I calculate some measures on the values belonging to the window (e.g., average, min, max, variance, entropy, energy, ... I use TsFresh library for such a task). Then, I mark each window as \"anomalous\" or not, based on the hand-made labelling (specifically, if there is an anomaly which overlaps the 95% of the window, then I mark it as anomalous).<\/p>\n\n<p>As a result, the training set is made by instances of the form, e.g.:\n[average_ts_1, min_ts_1, max_ts_1, ..., average_ts_10, min_ts_10, max_ts_10, LABEL]<\/p>\n\n<p>Although this approach is giving me some encouraging results, <strong>I would like to deal with the time series data directly<\/strong>, so not to lose any potential information.\nIn such case, I think that the task would become: given 10 time series as input, predict another time series made just by 1s and 0s. Such predicted time series would cover the same time period as the time series given as input, and would tell me in which intervals the anomaly is present or not (e.g., with a 1-second granularity).<\/p>\n\n<p>Now, finally, the question: is there any (deep) learning architecture that I could use for such a task?<\/p>\n\n<p>Currently, I am working with Python, relying on XGBoost with the windowing techique.\nI have also some experience with Keras, but I am open to any suggestion.<\/p>\n","answers":[]}
+{"QuestionId":52991795,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-25T14:30:04.340","AcceptedAnswerId":52991983.0,"OwnerUserId":680074.0,"Title":"Lstm Keras - How to increase the number of epochs","Body":"<p>I am working on binary classification I want to increase the number of epochs in my code this is my data set when i am increasing the value in the dense function i am getting Error when checking target: <code>expected dense_16 to have shape (10,) but got array with shape (1,)<\/code><\/p>\n\n<pre><code>[[   nan  1520.  1295.    nan  8396.  9322. 12715.    nan  5172.  7232.\n  11266.    nan 11266.  2757.  4416. 12020. 12111.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  3045. 11480.   900.  5842. 11496.  4463.    nan 11956.   900.\n  10400.  8022.  2504. 12106.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  9307. 12003.  2879.  6398.  9372.  4614.  5222.    nan    nan\n   2879. 10364.  6923.  4709.  4860. 11871.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  6689.  2818. 12003.  6480.    nan     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [   nan  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]]\n<\/code><\/pre>\n\n<p>I want to increase the number of epochs here<\/p>\n\n<pre><code>PositiveOrNegativeLabel=np.array([[1]])\nPositiveOrNegativeLabel=PositiveOrNegativeLabel.reshape(1,-1)\nPositiveOrNegativeLabel.shape\ninputBatch =inputBatch.reshape(1,6,30)\nprint(PositiveOrNegativeLabel.shape)\nmodel=Sequential()\nmodel.add(LSTM(100,input_shape=(6,30)))\nmodel.add(Dense(1,activation=\"sigmoid\"))\nmodel.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])\nmodel.fit(inputBatch,PositiveOrNegativeLabel,batch_size=24,verbose=1)\n<\/code><\/pre>\n\n<p><strong>this is the value error i am getting ValueError: Error when checking target: expected dense_16 to have shape (10,) but got array with shape (1,)<\/strong><\/p>\n","answers":[{"AnswerId":"52991983","CreationDate":"2018-10-25T14:38:46.680","ParentId":null,"OwnerUserId":"10416582","Title":null,"Body":"<p>I believe this may be a mismatch between the output of the last layer and your expected output dimensions. One easy way to fix this is to change your line<\/p>\n\n<p><code>model.add(Dense(1,activation=\"sigmoid\"))<\/code><\/p>\n\n<p>to:<\/p>\n\n<p><code>model.add(Dense(10,activation=\"sigmoid\"))<\/code><\/p>\n\n<p>Could you list all variables you are using and the dimensions of them, if you need further assistance?<\/p>\n\n<p>Additionally, you have some whitespace issues here not in accordance with PEP8. I suggest you check out: <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0008\/\" rel=\"nofollow noreferrer\">https:\/\/www.python.org\/dev\/peps\/pep-0008\/<\/a><\/p>\n"}]}
+{"QuestionId":52991969,"AnswerCount":2,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2018-10-25T14:38:21.360","AcceptedAnswerId":null,"OwnerUserId":1735729.0,"Title":"How to compare the similarity of two keras models","Body":"<p>I have built a Keras model using the functional API and I created a second model using <code>model_from_json()<\/code> function. I want to see if the model layers (not the weights) of the two models are exactly the same. <\/p>\n\n<p>How can I compare the two Keras models?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Based on the comments below, I could possibly compare each layer. Would something like the one below make sense:<\/p>\n\n<pre><code>for l1, l2 in zip(mdl.layers, mdl2.layers):\n    print (l1.get_config() == l2.get_config())\n<\/code><\/pre>\n","answers":[{"AnswerId":"52992945","CreationDate":"2018-10-25T15:24:38.180","ParentId":null,"OwnerUserId":"2145789","Title":null,"Body":"<p><strong>Update:<\/strong> your approach is correct.<\/p>\n\n<p>You can iterate over the two models layers and compare one by one (since you don't care about the weights or how the model is compiled and optimized).<\/p>\n\n<p>You can do this:<\/p>\n\n<pre><code>for l1, l2 in zip(mdl.layers, mdl2.layers):\n    print(l1.get_config() == l2.get_config())\n<\/code><\/pre>\n\n<p>Or just this:<\/p>\n\n<pre><code>print(mdl.get_config() == mdl2.get_config())\n<\/code><\/pre>\n"},{"AnswerId":"52992477","CreationDate":"2018-10-25T15:02:36.837","ParentId":null,"OwnerUserId":"10541372","Title":null,"Body":"<p>you can use ROC curves to compare classifiers.<\/p>\n\n<p>With binary classifiers, one calculates the true-positive-rate and the false-positive-rate for all possible thresholds and plots the first on the y-axis and the second on the x-axis. The resulting curve per classifier can be integrated and the resulting integral, the so called \"area under the curve\", is equal to the probability that the classifier rankes a randomly chosen positive sample higher than a randomly chosen negative one. This value can be used to compare classifiers because a higher value shows a overall better performance than a lower one. Fawcett also gives a method to apply this to multi-class classification<\/p>\n"}]}
+{"QuestionId":52992595,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-25T15:08:33.740","AcceptedAnswerId":null,"OwnerUserId":10238509.0,"Title":"Fail to import tensorflow on windows","Body":"<p>I am using windows10, python 3.6 without anaconda. I was trying to install tensorflow for my GTX1060.\nAfter installing it, I found that I cannot import tensorflow and the error message as followed:<br>\nTraceback (most recent call last):\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Program Files\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Program Files\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: \u627e\u4e0d\u5230\u6307\u5b9a\u7684\u6a21\u5757\u3002<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow__init__.py\", line 22, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Program Files\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Program Files\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: \u627e\u4e0d\u5230\u6307\u5b9a\u7684\u6a21\u5757\u3002<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.<\/p>\n","answers":[]}
+{"QuestionId":52992805,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><histogram><distribution>","CreationDate":"2018-10-25T15:18:07.807","AcceptedAnswerId":null,"OwnerUserId":3935315.0,"Title":"Tensorflow 2d Histogram","Body":"<p>I'm trying to create a 2D histogram in tensorflow to use in a custom loss function in tensorflow. More generally, I think people could benefit from using co-activations of neurons, which requires a similar structure.<\/p>\n\n<p>Here's specifically what I'm trying to do:<\/p>\n\n<p>Given a Nx2 Tensor, where N is some number of samples, I'd like to create a (binned) histogram of co-activations. For example, in the simple case of input=[[0, 0.01], [0, 0.99], [0.5, 0.5]] and with 10000 bins total, I'd like to generate a 100x100 Tensor with all 0s except for 3 entries at (0, 0.01), (0, 0.99), and (0.5, 0.5), where the value would be 1\/3 (scaling is easy, so I'd be fine with 1 instead).<\/p>\n\n<p>I could do this easily using standard numpy or array operations<\/p>\n\n<pre><code>neuron1 = data[:, 1]\nneuron2 = data[:, 2]\n\nhist_2d = np.zeros((100, 100))\n\nfor neuron1_output in neuron1:\n    for neuron2_output in neuron2:\n        hist_2d[int(100 * neuron1_output), int(100 * neuron2_output)] += 1\n<\/code><\/pre>\n\n<p>If I want to use hist_2d as part of a loss function in Tensorflow, though, it looks like I can't do this kind of iterating.<\/p>\n\n<p>Does anyone know of a good way to generate the 2d histogram I'm looking for? I was happy to find tf.histogram_fixed_width(), but that only generates 1d histograms. I've started looking into tf.while_loop() and tf.map_fn(), but I'm pretty new to Tensorflow, so I'm not sure what avenue is most promising.<\/p>\n","answers":[{"AnswerId":"58849750","CreationDate":"2019-11-14T05:08:54.957","ParentId":null,"OwnerUserId":"6183244","Title":null,"Body":"<p>Maybe <a href=\"http:\/\/%20https:\/\/gist.github.com\/isentropic\/a86effab2c007e86912a50f995cac52b\" rel=\"nofollow noreferrer\">this<\/a> snipped would help you.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n@tf.function\ndef get2dHistogram(x, y,\n                   value_range,\n                   nbins=100,\n                   dtype=tf.dtypes.int32):\n    \"\"\"\n    Bins x, y coordinates of points onto simple square 2d histogram\n\n    Given the tensor x and y:\n    x: x coordinates of points\n    y: y coordinates of points\n    this operation returns a rank 2 `Tensor` \n    representing the indices of a histogram into which each element\n    of `values` would be binned. The bins are equal width and\n    determined by the arguments `value_range` and `nbins`.\n\n\n  Args:\n    x:  Numeric `Tensor`.\n    y: Numeric `Tensor`.\n    value_range[0] lims for x\n    value_range[1] lims for y\n\n    nbins:  Scalar `int32 Tensor`.  Number of histogram bins.\n    dtype:  dtype for returned histogram.\n\n\n\n    \"\"\"\n    x_range = value_range[0]\n    y_range = value_range[1]\n\n    histy_bins = tf.histogram_fixed_width_bins(y, y_range, nbins=nbins, dtype=dtype)\n\n    H = tf.map_fn(lambda i: tf.histogram_fixed_width(x[histy_bins == i], x_range, nbins=nbins),\n              tf.range(nbins))\n    return H # Matrix!\n<\/code><\/pre>\n\n<p>Written in tensorflow 2.0, but you surely can manage it.<\/p>\n"},{"AnswerId":"53010447","CreationDate":"2018-10-26T14:00:44.170","ParentId":null,"OwnerUserId":"3935315","Title":null,"Body":"<p>Posting an \"answer\", more like a workaround, that I figured out.<\/p>\n\n<p>The whole reason I wanted to create the 2D histogram is that I wanted to calculate the entropy of the joint distribution of the activations of two neurons. I'm already discretizing the values of the activations into bins, so it's ok if I shuffle the distribution around as that won't modify the entropy value.<\/p>\n\n<p>Given that, here's what I did: I created a 1D histogram with a squared number of bins and simply slid the values so that the first half of the digits corresponded to activations for neuron1 and the second half for activations for neuron2. In python:<\/p>\n\n<pre><code># Calculate the entropy of a 1D tensor, fuzzing the edges with epsilon to keep numbers\n# clean.\ndef calculate_entropy(y, epsilon):\n    clipped = tf.clip_by_value(y, epsilon, 1 - epsilon)\n    return -tf.cast(tf.reduce_sum(clipped * tf.log(clipped)), dtype=tf.float32)\n\n\n# Sandbox for developing calculating the entropies of y\ndef tf_entropies(y, epsilon, nbins):\n    # Create histograms for the activations in the batch.\n    value_range = [0.0, 1.0]\n    # For prototype, only consider first two features.\n    neuron1 = y[:, 0]\n    neuron2 = y[:, 1]\n    hist1 = tf.histogram_fixed_width(neuron1, value_range, nbins=nbins)\n    hist2 = tf.histogram_fixed_width(neuron2, value_range, nbins=nbins)\n    # Normalize\n    count = tf.cast(tf.count_nonzero(hist1), tf.int32)\n    dist1 = tf.divide(hist1, count)\n    dist2 = tf.divide(hist2, count)\n    neuron1_entropy = calculate_entropy(dist1, epsilon)\n    neuron2_entropy = calculate_entropy(dist2, epsilon)\n\n    # Calculate the joint distribution and then get the entropy\n    recast_n1 = tf.cast(tf.divide(tf.cast(nbins * neuron1, tf.int32), nbins), tf.float32)\n    meshed = recast_n1 + tf.divide(neuron2, nbins)  # Shift over the numbers for neuron2\n    joint_hist = tf.histogram_fixed_width(meshed, value_range, nbins=nbins * nbins)\n    joint_dist = tf.divide(joint_hist, count)\n    joint_entropy = calculate_entropy(joint_dist, epsilon)\n\n    return neuron1_entropy, neuron2_entropy, joint_entropy, joint_dist\n<\/code><\/pre>\n\n<p>Once I've got the joint histogram, I can get the joint entropy using normal procedures. I validated that I got the correct result by implementing the same logic using normal numpy operations. The entropy calculations match.<\/p>\n\n<p>I hope this helps others if they run into similar problems.<\/p>\n"}]}
+{"QuestionId":52992821,"AnswerCount":2,"Tags":"<tensorflow><graph>","CreationDate":"2018-10-25T15:18:55.437","AcceptedAnswerId":53011073.0,"OwnerUserId":6935676.0,"Title":"Tensorflow graph fetch all consts in a scope","Body":"<p>I create a graph and now I want to fetch their ops, how can I do this?<\/p>\n\n<pre><code>g = tf.Graph()\n\nwith g.as_default():\n    # Define inputs\n    with tf.name_scope(\"inputs\"):\n        a = tf.constant(2, tf.int32, name=\"a\")\n        b = tf.constant(3, tf.int32, name=\"b\")\n\n    # Ops\n    with tf.name_scope(\"ops\"):\n        c = tf.multiply(a, b, name=\"c\")\n        d = tf.add(a, b, name=\"d\")\n        e = tf.subtract(c, d, name=\"e\")\n\nsess = tf.InteractiveSession()\n\n_c, _d, _e = ... &lt;-- (I need some code here!)\n<\/code><\/pre>\n\n<p>Can you show me document links about this?<\/p>\n","answers":[{"AnswerId":"53011073","CreationDate":"2018-10-26T14:38:16.337","ParentId":null,"OwnerUserId":"5874587","Title":null,"Body":"<pre><code>sess = tf.Session(graph=g)\n\n_c, _d, _e = sess.run([c, d, e])\nprint(\"c =\", _c)\nprint(\"d =\", _d)\nprint(\"e =\", _e)\n<\/code><\/pre>\n\n<p>This will solve your problem :)<\/p>\n"},{"AnswerId":"52995121","CreationDate":"2018-10-25T17:36:46.077","ParentId":null,"OwnerUserId":"7636396","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#get_operations\" rel=\"nofollow noreferrer\">Get_Operations()<\/a> will do the job<\/p>\n\n<pre><code>g = tf.Graph()\n...\n\nfor ops in g.get_operations():\n    print(ops)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52993008,"AnswerCount":1,"Tags":"<tensorflow><object-detection><object-detection-api><tensorflow-slim>","CreationDate":"2018-10-25T15:27:38.693","AcceptedAnswerId":52993537.0,"OwnerUserId":10558075.0,"Title":"get feature vector from Fast-RCNN","Body":"<p>I am trying to identify and group similar images, I followed this tutorial: <a href=\"https:\/\/douglasduhaime.com\/posts\/identifying-similar-images-with-tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/douglasduhaime.com\/posts\/identifying-similar-images-with-tensorflow.html<\/a>.<\/p>\n\n<p>The problem is that I am using faster_rcnn_resnet_101 and I found out that feature vectors in fast-rcnn get dropped after SecondStageBoxPredictor. I used <a href=\"https:\/\/gist.github.com\/markdtw\/02ece6b90e75832bd44787c03a664e8d\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/markdtw\/02ece6b90e75832bd44787c03a664e8d<\/a> to get the vectores before being dropped.<\/p>\n\n<p><code>feat_avg = graph.get_tensor_by_name('SecondStageBoxPredictor\/AvgPool:0')<\/code> <\/p>\n\n<pre><code>np.savetxt('output1\/' + \"test\" + '.npz',feat_vector,delimiter=',')\n<\/code><\/pre>\n\n<p>However when I try to save the vector I get an error:<\/p>\n\n<pre><code>ValueError: Expected 1D or 2D array, got 4D array instead\n<\/code><\/pre>\n\n<p>I printed the extracted feature vector to see the result:<\/p>\n\n<pre><code>    Tensor(\"SecondStageFeatureExtractor\/resnet_v1_101\/block4\/unit_3\/bottleneck_v1\/Relu:0\", shape=(?, 7, 7, 2048), dtype=float32) \n [[[[0.         0.         0.         ... 0.         0.\n    0.        ]\n   [0.         0.         0.         ... 0.         0.\n    0.        ]\n   [0.         0.         0.         ... 0.         0.\n    0.        ]\n   ...\n   [2.9170244  0.         0.33220196 ... 0.         0.\n    0.        ]\n   [0.         0.         0.         ... 0.         0.\n    0.        ]\n   [0.         0.         0.         ... 0.         0.\n    0.        ]].....\n<\/code><\/pre>\n\n<p>I am a novice in TensorFlow and CV, what I want to do is extract the feature vectors and then use TSNE clusterring.\nWhat exactly is wrong with the feature vector I extracted <\/p>\n","answers":[{"AnswerId":"52993537","CreationDate":"2018-10-25T15:55:17.970","ParentId":null,"OwnerUserId":"8315956","Title":null,"Body":"<p>Feature vector is 4D sensor: [Batch, Height, Width, Channels].\nnp.savetxt waits 1D or 2D array. You can slice feature vector to 2D arrays or use some other functional for save it as 4D. <\/p>\n"}]}
+{"QuestionId":52993397,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><lstm>","CreationDate":"2018-10-25T15:47:22.907","AcceptedAnswerId":null,"OwnerUserId":10558684.0,"Title":"What is the default activation function of cudnnlstm in tensorflow","Body":"<p>What's the default activation function of <code>cudnnlstm<\/code> in TensorFlow? How can I set an activation function such as <code>relu<\/code>? Maybe it's just linear model? I read the document, but I did not find it.<\/p>\n\n<p>For example, the code is below:<\/p>\n\n<pre><code>lstmcell=tf.contrib.cudnn_rnn.CudnnLSTM(1,encoder_size,direction=\"bidirectional\")\n\nhq,_ =lstmcell(query)\n<\/code><\/pre>\n\n<p>And I read the document of TensorFlow <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/cudnn_rnn\/CudnnLSTM?hl=zh-cn\" rel=\"nofollow noreferrer\">From this link<\/a>.<\/p>\n\n<p>The function is below<\/p>\n\n<pre><code>__init__(\n\n    num_layers,\n    num_units,\n    input_mode=CUDNN_INPUT_LINEAR_MODE,\n    direction=CUDNN_RNN_UNIDIRECTION,\n    dropout=0.0,\n    seed=None,\n    dtype=tf.float32,\n    kernel_initializer=None,\n    bias_initializer=None,\n    name=None\n)\n<\/code><\/pre>\n\n<p>And no keyword to set a parameter such as <code>\"activation = \"tanh\"<\/code> just like <code>tf.nn.rnn_cell.LSTMell<\/code>.<\/p>\n\n<p>So what's the default activation function of <code>cudnnlstm<\/code> in TensorFlow, and how to change it to <code>leaky_relu<\/code>.<\/p>\n","answers":[{"AnswerId":"55466938","CreationDate":"2019-04-02T04:24:40.263","ParentId":null,"OwnerUserId":"7941307","Title":null,"Body":"<pre><code>tf.contrib.cudnn_rnn.CudnnLSTM() : Tanh\n<\/code><\/pre>\n\n<p>This was given in the Keras github.<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8510#issuecomment-429255318\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/8510#issuecomment-429255318<\/a> <\/p>\n\n<p>Nvidia documentation.<\/p>\n\n<p><a href=\"https:\/\/devblogs.nvidia.com\/optimizing-recurrent-neural-networks-cudnn-5\/\" rel=\"nofollow noreferrer\">https:\/\/devblogs.nvidia.com\/optimizing-recurrent-neural-networks-cudnn-5\/<\/a>  <\/p>\n"}]}
+{"QuestionId":52993992,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook><tensorboard>","CreationDate":"2018-10-25T16:24:21.860","AcceptedAnswerId":null,"OwnerUserId":6690971.0,"Title":"tensorboard written in the jupyter notebook doesn't work","Body":"<p>I run the code<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.placeholder(tf.float32, [None, 1])\ny_ = tf.placeholder(tf.float32, [None, 1])\nm = tf.Variable(0.)\ny = m*x\nloss = tf.reduce_mean(tf.pow(y - y_, 2))\nupdate = tf.train.GradientDescentOptimizer(0.0001).minimize(loss)\n\nmsum = tf.summary.scalar('m', m)\nlosssum = tf.summary.scalar('loss', loss)\nmerged = tf.summary.merge_all()\n\nsess = tf.Session()\nfile_writer = tf.summary.FileWriter('.\/my_graph', sess.graph)\nsess.run(tf.global_variables_initializer())\ndata_dict = {x:[[7.01], [3.02], [4.99], [8.]], y_:[[14.01], [6.01], [10.], [16.04]]}\nfor i in range(0,1000):\n   [_,curr_sammary] = sess.run([update,merged], feed_dict = data_dict)\n   file_writer.add_summary(curr_sammary, i)\n\nfile_writer.close()\nprint(sess.run(m))\n<\/code><\/pre>\n\n<p>But this code is not running because of the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)\n   1321     try:\n-&gt; 1322       return fn(*args)\n   1323     except errors.OpError as e:\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)\n   1306       return self._call_tf_sessionrun(\n-&gt; 1307           options, feed_dict, fetch_list, target_list, run_metadata)\n   1308 \n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)\n   1408           self._session, options, feed_dict, fetch_list, target_list,\n-&gt; 1409           run_metadata)\n   1410     else:\n\nInvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,1]\n     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[?,1], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nDuring handling of the above exception, another exception occurred:\n\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-42-c57f3c0f3525&gt; in &lt;module&gt;()\n      1 data_dict = {x:[[7.01], [3.02], [4.99], [8.]], y_:[[14.01], [6.01], [10.], [16.04]]}\n      2 for i in range(0,1):\n----&gt; 3    [_,curr_sammary] = sess.run([update,merged], feed_dict = data_dict)\n      4    file_writer.add_summary(curr_sammary, i)\n      5 \n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in run(self, fetches, feed_dict, options, run_metadata)\n    898     try:\n    899       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 900                          run_metadata_ptr)\n    901       if run_metadata:\n    902         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1133     if final_fetches or final_targets or (handle and feed_dict_tensor):\n   1134       results = self._do_run(handle, final_targets, final_fetches,\n-&gt; 1135                              feed_dict_tensor, options, run_metadata)\n   1136     else:\n   1137       results = []\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)\n   1314     if handle is None:\n   1315       return self._do_call(_run_fn, feeds, fetches, targets, options,\n-&gt; 1316                            run_metadata)\n   1317     else:\n   1318       return self._do_call(_prun_fn, handle, feeds, fetches)\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)\n   1333         except KeyError:\n   1334           pass\n-&gt; 1335       raise type(e)(node_def, op, message)\n   1336 \n   1337   def _extend_graph(self):\n\nInvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,1]\n     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[?,1], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nCaused by op 'Placeholder', defined at:\n  File \"\/usr\/lib\/python3.5\/runpy.py\", line 184, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/usr\/lib\/python3.5\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel_launcher.py\", line 16, in &lt;module&gt;\n    app.launch_new_instance()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/traitlets\/config\/application.py\", line 658, in launch_instance\n    app.start()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/kernelapp.py\", line 486, in start\n    self.io_loop.start()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tornado\/platform\/asyncio.py\", line 132, in start\n    self.asyncio_loop.run_forever()\n  File \"\/usr\/lib\/python3.5\/asyncio\/base_events.py\", line 345, in run_forever\n    self._run_once()\n  File \"\/usr\/lib\/python3.5\/asyncio\/base_events.py\", line 1312, in _run_once\n    handle._run()\n  File \"\/usr\/lib\/python3.5\/asyncio\/events.py\", line 125, in _run\n    self._callback(*self._args)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tornado\/platform\/asyncio.py\", line 122, in _handle_events\n    handler_func(fileobj, events)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tornado\/stack_context.py\", line 300, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 450, in _handle_events\n    self._handle_recv()\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 480, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 432, in _run_callback\n    callback(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tornado\/stack_context.py\", line 300, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/kernelbase.py\", line 283, in dispatcher\n    return self.dispatch_shell(stream, msg)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/kernelbase.py\", line 233, in dispatch_shell\n    handler(stream, idents, msg)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/kernelbase.py\", line 399, in execute_request\n    user_expressions, allow_stdin)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/ipkernel.py\", line 208, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/ipykernel\/zmqshell.py\", line 537, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2662, in run_cell\n    raw_cell, store_history, silent, shell_futures)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2785, in _run_cell\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2901, in run_ast_nodes\n    if self.run_code(code, result):\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2961, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-1-8fcfa2faf332&gt;\", line 3, in &lt;module&gt;\n    x = tf.placeholder(tf.float32, [None, 1])\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/ops\/array_ops.py\", line 1734, in placeholder\n    return gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 4924, in placeholder\n    \"Placeholder\", dtype=dtype, shape=shape, name=name)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3414, in create_op\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 1740, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,1]\n     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[?,1], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>After trying to understand the source of the problem I saw that if the line<\/p>\n\n<pre><code>[_,curr_sammary] = sess.run([update,merged], feed_dict = data_dict)\n<\/code><\/pre>\n\n<p>in the loop I change the line   <\/p>\n\n<pre><code>sess.run(update, feed_dict = data_dict)\n<\/code><\/pre>\n\n<p>the problem is solved.<\/p>\n\n<p>I tried to search for understand what the error was and didn't find anything to solve the problem.<\/p>\n\n<p>I need the line to work in its original form (without the deletions)<\/p>\n\n<p>I would be happy if someone can tell me what the problem is and what its solution is.<\/p>\n","answers":[{"AnswerId":"52994645","CreationDate":"2018-10-25T17:06:38.513","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>I'm on <code>tensorflow-gpu (1.7.0)<\/code> and the original, unedited version <\/p>\n\n<p><code>[_,curr_sammary] = sess.run([update,merged], feed_dict = data_dict)<\/code><\/p>\n\n<p>runs successfully runs with no errors on both <code>Jupyter<\/code> and on the terminal.<\/p>\n\n<hr>\n\n<p>I DID, however, notice that when i ran the script a second time on <code>Jupyter<\/code>, that I got the same error as you. This leads me to believe you're keeping state between runs.<\/p>\n\n<p>On the terminal (where no state is kept between consecutive runs), it all worked fine<\/p>\n"}]}
+{"QuestionId":52994435,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2018-10-25T16:52:41.790","AcceptedAnswerId":53009161.0,"OwnerUserId":10558946.0,"Title":"PyTorch: create non-fully-connected layer \/ concatenate output of hidden layers","Body":"<p>In PyTorch, I want to create a hidden layer whose neurons are not fully connected to the output layer. I try to concatenate the output of two linear layers but run into the following error:<\/p>\n\n<blockquote>\n  <p>RuntimeError: size mismatch, m1: [2 x 2], m2: [4 x 4]<\/p>\n<\/blockquote>\n\n<p>my current code:<\/p>\n\n<pre><code>class NeuralNet2(nn.Module):\n    def __init__(self):\n        super(NeuralNet2, self).__init__()\n\n        self.input = nn.Linear(2, 40)\n        self.hiddenLeft = nn.Linear(40, 2)\n        self.hiddenRight = nn.Linear(40, 2)\n        self.out = nn.Linear(4, 4)\n\n    def forward(self, x):\n        x = self.input(x)\n        xLeft, xRight = torch.sigmoid(self.hiddenLeft(x)), torch.sigmoid(self.hiddenRight(x))\n        x = torch.cat((xLeft, xRight))\n        x = self.out(x)\n\n        return x\n<\/code><\/pre>\n\n<p>I don't get why there is a size mismatch? Is there an alternative way to implement non-fully-connected layers in pytorch?<\/p>\n","answers":[{"AnswerId":"53009161","CreationDate":"2018-10-26T12:50:49.130","ParentId":null,"OwnerUserId":"10558946","Title":null,"Body":"<p>It turned out to be a simple comprehension problem with the concatenation function.\nChanging\n    <code>x = torch.cat((xLeft, xRight))<\/code>\nto\n    <code>x = torch.cat((xLeft, xRight), dim=1)<\/code>\ndid the trick.\nThanks @dennlinger<\/p>\n"}]}
+{"QuestionId":52995288,"AnswerCount":1,"Tags":"<python><unity3d><keras><deep-learning><ml-agent>","CreationDate":"2018-10-25T17:48:22.817","AcceptedAnswerId":null,"OwnerUserId":7030761.0,"Title":"Training in Keras with external evaluation function","Body":"<p>Let me first describe the setup:\nWe have an autonomous agent in Unity, whose decisions are based on the perceived environment(level) and some pre-defined parameters for value mapping. Our aim is to pre-train the agents' parameters in a DNN. So the idea is basically to define an error metric to evaluate the performance of the agent in a Unity simulation (run the level, e.g. measure the deviation from the optimal trajectory = ground truth in unity). So based on the input level into the DNN, the network should train to output the params, the simulation is performed and the error is passed back to the network as the error value, like accuracy, so the network could train based on that error\/performance.<\/p>\n\n<p>Is there any way to perform the evaluation(comparison to the ground truth) during the training outside of Keras? Usually, one passes X data to the network, train stuff and compare it to the ground truth Y. This works fine for predictions, but I don't want to predict something. What I do want is to measure the deviation from the ground truth inside the simulation.\nI know there is Unity ML Agents, but as far as I could read, the 'brain' controls the agent on runtime, i.e. update it on every frame and control the movement. What I want is to perform the whole simulation to update the params\/weights of the network.<\/p>\n\n<p>Best wishes.<\/p>\n","answers":[{"AnswerId":"53008270","CreationDate":"2018-10-26T11:59:24.053","ParentId":null,"OwnerUserId":"7030761","Title":null,"Body":"<p>After some talks at my university:\nthe setup won't work this way since I need to split the process.<\/p>\n\n<p>I need the parameters of working agents to train the network based only on the level description(e.g. matrix like video game description language). To obtain the parametrized agents based on the actual level and the ground truth data(e.g. deviation from trajectory), one need to use reinforcement deep learning with a score function to obtain these parameters. Therefore Unity ML Agents might be useful. Afterwards, I can use the parameters settings and the correlating level data to train a network to yield the desired parameters based only on the level description.<\/p>\n"}]}
+{"QuestionId":52995368,"AnswerCount":1,"Tags":"<python><pandas><numpy><keras>","CreationDate":"2018-10-25T17:53:36.693","AcceptedAnswerId":52997210.0,"OwnerUserId":5552705.0,"Title":"Reshape 2D dataframe to include previous data points (for LSTM in Keras)","Body":"<p>I have a dataframe with 100000 samples and 10 features. I want to convert this to a 3D array where [X][0] is the current data point, [X][1] is the previous data point (X-1) etc. So if i want a \"lookback\" of five units, the shape of the data would look like (100000, 5, 10). The reason being that i want my LSTM to consider previous data points when training the model.<\/p>\n\n<p>Is there an easy way to do this?<\/p>\n","answers":[{"AnswerId":"52997210","CreationDate":"2018-10-25T20:01:30.600","ParentId":null,"OwnerUserId":"4995541","Title":null,"Body":"<p>If your dataframe is of shape (100000, 10), you cannot reshape it to a (100000, 5, 10), simply because you don't have 100000x5x10 values in it. So, the first step is to create the lagged features:<\/p>\n\n<pre><code>&gt; df = pd.DataFrame({'a':[i for i in range(5)], 'b':[i for i in range(5, 10)]})\n&gt; df\n&gt;       a   b\n    0   0   5\n    1   1   6\n    2   2   7\n    3   3   8\n    4   4   9\n&gt; columns = [df.shift(i) for i in range()]\n&gt; df = pd.concat(columns, axis=1)\n&gt; df\n&gt;       a   b   a   b   a   b\n    0   0   5   NaN NaN NaN NaN\n    1   1   6   0.0 5.0 NaN NaN\n    2   2   7   1.0 6.0 0.0 5.0\n    3   3   8   2.0 7.0 1.0 6.0\n    4   4   9   3.0 8.0 2.0 7.0\n<\/code><\/pre>\n\n<p>After this, you can reshape your dataframe to suit your needs:<\/p>\n\n<pre><code>&gt; # 5 samples, 3 timesteps, 2 features\n&gt; df.values.reshape(5,3,2)\n&gt; array([[[ 0.,  5.],\n    [nan, nan],\n    [nan, nan]],\n\n   [[ 1.,  6.],\n    [ 0.,  5.],\n    [nan, nan]],\n  ...\n<\/code><\/pre>\n"}]}
+{"QuestionId":52995453,"AnswerCount":1,"Tags":"<python><tensorflow><matplotlib>","CreationDate":"2018-10-25T18:00:34.870","AcceptedAnswerId":null,"OwnerUserId":8547664.0,"Title":"How to plot training error , validation error and prediction accuracy over training progress in tensorflow?","Body":"<p>I am running Vanilla RNN code on tensorflow in google colab. I want to plot training error, validation error and prediction accuracy over training progress without using tensorboard. I am new to tensorflow. Can anyone please guide me. Here is a part of my code <\/p>\n\n<pre><code>       # Model predictions\n    cls_prediction = tf.argmax(output_logits, axis=1, name='predictions')\n\n    # Define the loss function, optimizer, and accuracy\n    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=output_logits), name='loss')\n    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, name='Adam-op').minimize(loss)\n    correct_prediction = tf.equal(tf.argmax(output_logits, 1), tf.argmax(y, 1), name='correct_pred')\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')\ninit = tf.global_variables_initializer()\nsess = tf.InteractiveSession()\nsess.run(init)\nglobal_step = 0\n# Number of training iterations in each epoch\nnum_tr_iter = int(len(y_train) \/ batch_size)\nfor epoch in range(epochs):\n    print('Training epoch: {}'.format(epoch + 1))\n    x_train, y_train = randomize(x_train, y_train)\n    for iteration in range(num_tr_iter):\n        global_step += 1\n        start = iteration * batch_size\n        end = (iteration + 1) * batch_size\n        x_batch, y_batch = get_next_batch(x_train, y_train, start, end)\n        x_batch = x_batch.reshape((batch_size, timesteps, num_input))\n        # Run optimization op (backprop)\n        feed_dict_batch = {x: x_batch, y: y_batch}\n        sess.run(optimizer, feed_dict=feed_dict_batch)\n\n        if iteration % display_freq == 0:\n            # Calculate and display the batch loss and accuracy\n            loss_batch, acc_batch = sess.run([loss, accuracy],\n                                             feed_dict=feed_dict_batch)\n\n            print(\"iter {0:3d}:\\t Loss={1:.2f},\\tTraining Accuracy={2:.01%}\".\n                  format(iteration, loss_batch, acc_batch))\n\n    # Run validation after every epoch\n\n    feed_dict_valid = {x: x_valid[:1000].reshape((-1, timesteps, num_input)), y: y_valid[:1000]}\n    loss_valid, acc_valid = sess.run([loss, accuracy], feed_dict=feed_dict_valid)\n    print('---------------------------------------------------------')\n    print(\"Epoch: {0}, validation loss: {1:.2f}, validation accuracy: {2:.01%}\".\n          format(epoch + 1, loss_valid, acc_valid))\n    print('---------------------------------------------------------')\n<\/code><\/pre>\n\n<p>What changes should I make in the code to get the plots?<\/p>\n","answers":[{"AnswerId":"52996044","CreationDate":"2018-10-25T18:38:41.270","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>One such way would be to store the values in a list, then use something like <a href=\"https:\/\/matplotlib.org\/tutorials\/introductory\/pyplot.html\" rel=\"nofollow noreferrer\">matplotlib<\/a> to plot the values<\/p>\n\n<p>Example code:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nplt.plot([1, 2, 3, 4])\nplt.ylabel('some numbers')\nplt.show()\n<\/code><\/pre>\n\n<p>will plot a straight line. In your case, you'd call <code>plt.plot(list_of_prediction_accuracy)<\/code> or whatever list you want to visualize<\/p>\n"}]}
+{"QuestionId":52995853,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><deep-learning><padding>","CreationDate":"2018-10-25T18:27:23.440","AcceptedAnswerId":57877486.0,"OwnerUserId":1410665.0,"Title":"Custom padding (Float Type) for transpose convolutions in TensorFlow","Body":"<p>I'm working on a TensorFlow transpose convolution model where (batch size, layer height, layer width, output channel) is (100, 25, 25, 64).\nNow I'm applying a transpose operation where the kernel of size 4 x 4 with a stride is 2 and padding is 0.25. My aim is to get output size 50 e.g. (100, 50, 50, 64)\nI checked <a href=\"https:\/\/stackoverflow.com\/questions\/37659538\/custom-padding-for-convolutions-in-tensorflow\">Stack OverFlow<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad\" rel=\"nofollow noreferrer\">tfpad<\/a> where pad type mentioned Init32. \nIt will be a great help if any of you kindly suggest how can I use padding of type float or how can I get size output layer size 50 after transpose operation. <\/p>\n","answers":[{"AnswerId":"57877486","CreationDate":"2019-09-10T20:01:42.140","ParentId":null,"OwnerUserId":"11495777","Title":null,"Body":"<p>Padding carries only zero. So it doesn't matter if it is float or anything else.<\/p>\n"}]}
+{"QuestionId":52996505,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-25T19:11:20.103","AcceptedAnswerId":52998808.0,"OwnerUserId":1937132.0,"Title":"Using scatter_nd with top_k output","Body":"<p>I've been trying to do something seemingly simple, with no success.\nI have a <code>(?,4)<\/code> tensor, where each row will be 4 floats between <code>0<\/code> and <code>1<\/code>.\nI want to replace this with a new tensor where each row has only the top 2 entries and zeros everywhere else.<\/p>\n\n<p>Example with a <code>(2, 4)<\/code>:<\/p>\n\n<pre><code>source = [ [0.1, 0.2, 0.5, 0.6],\n           [0.8, 0.7, 0.2, 0.1] ]\n\nresult = [ [0.0, 0.0, 0.5, 0.6],\n           [0.8, 0.7, 0.0, 0.0] ]\n<\/code><\/pre>\n\n<p>I tried using <strong>top_k<\/strong> on the source and then using <strong>scatter_nd<\/strong> with the indices returned by top_k, but it has literally been 4 hours of mismatched shapes and rank errors in scatter_nd.<\/p>\n\n<p>I'm ready to give up, but I thought I would ask for help here first.\nI've found a couple of questions here closely related, but I'm failing to generalize the info in there for my case.<\/p>\n\n<p>Another approach I just tried is this:<\/p>\n\n<pre><code>tensor = tf.constant( [ [0.1, 0.2, 0.8], [0.1, 0.2, 0.7] ])\nvalues, indices = tf.nn.top_k(tensor, 1)\nelems = (tensor, values)\nmasked_a = tf.map_fn( \n           lambda a : tf.where( tf.greater_equal(a[0], a[1]), a[0], \n           tf.zeros_like(a[0]) ), \n           elems)\n<\/code><\/pre>\n\n<p>but this one gives me the following error:<\/p>\n\n<pre><code>ValueError: The two structures don't have the same number of elements.\nFirst structure (2 elements): (tf.float32, tf.float32)\nSecond structure (1 elements): Tensor(\"map\/while\/Select:0\", shape=(3,), dtype=float32)\n<\/code><\/pre>\n\n<p>I'm relatively new with TensorFlow, so apologies if I'm missing something simple or being unclear.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52998808","CreationDate":"2018-10-25T22:09:51.367","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>You can do it with tf.scatter_nd by appending the row indice to the indices returned by top_k.<\/p>\n\n<pre><code>import tensorflow as tf\n\nsource = tf.constant([\n    [0.1, 0.2, 0.5, 0.6],\n    [0.8, 0.7, 0.2, 0.1]])\n\n# get indices of top k\nk = 2\ntop_k, top_k_inds = tf.nn.top_k(source, k, )\n\n# indices are only columns, we will stack \n# it so the row indice is also there and\n# make tensor of row numbers ie.\n# [[0, 0],\n#  [1, 1],\n#  ...\nnum_rows = tf.shape(source)[0]\nrow_range = tf.range(num_rows)\nrow_tensor = tf.tile(row_range[:,None], (1, k))\n\n# stack along the final dimension, as this is what\n# scatter_nd uses as the indices\ntop_k_row_col_indices = tf.stack([row_tensor, top_k_inds], axis=2)\n\n# to mask off everything, we will multiply the top_k by\n# 1. so all the updates are just 1\nupdates = tf.ones([num_rows, k], dtype=tf.float32)\n\n# build the mask\nzero_mask = tf.scatter_nd(top_k_row_col_indices, updates, [num_rows, 4])\n\nwith tf.Session() as sess:\n    zeroed = source*zero_mask\n    print(zeroed.eval())\n<\/code><\/pre>\n\n<p>This should print<\/p>\n\n<pre><code>[[0.  0.  0.5 0.6]\n[0.8 0.7 0.  0. ]]\n<\/code><\/pre>\n"},{"AnswerId":"55565600","CreationDate":"2019-04-08T02:48:37.690","ParentId":null,"OwnerUserId":"1360131","Title":null,"Body":"<p>Just paste some lines of code :)<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef attach_indice(tensor, top_k = None):\n    flatty = tf.reshape(tensor, [-1])\n    orig_shape = tf.shape(tensor)\n    length = tf.shape(flatty)[0]\n    if top_k is not None:\n        orig_shape = orig_shape[:-1] # dim for top_k\n        length \/\/= top_k\n    indice = tf.unravel_index(tf.range(length), orig_shape)\n    indice = tf.transpose(indice)\n    if indice.dtype != tensor.dtype:\n        indice = tf.cast(indice, tensor.dtype)\n    if top_k is not None:\n        _dims = len(tensor.shape) - 1 # indice of indice\n        shape = [1 for _ in range(_dims)]\n        shape[-1] *= top_k\n        indice = tf.reshape(tf.tile(indice, shape), [-1, _dims])\n    return tf.concat([indice, flatty[:, None]], -1)\n\nimport os\nos.environ['CUDA_VISIBLE_DEVICES'] = '0'\n# tf.enable_eager_execution()\n\nfrom time import time\n\ntop_k = 3\nshape = [50, 40, 100]\n\nq = tf.random_uniform(shape)\n\n# fast: 4.376221179962158 (GPU) \/ 2.483684778213501 (CPU)\nv, k = tf.nn.top_k(q, top_k)\nk = attach_indice(k, top_k)\ns = tf.scatter_nd(k, tf.reshape(v, [-1]), shape)\n\n# very slow: 281.82796931266785 (GPU) \/ 35.163344860076904 (CPU)\n# s = tf.map_fn(lambda v__k__: tf.map_fn(lambda v_k_: tf.scatter_nd(v_k_[1][:, None], v_k_[0], [shape[-1]]), v__k__, q.dtype), tf.nn.top_k(q, top_k), q.dtype)\n\nstart = time()\nwith tf.Session() as sess:\n    for _ in range(1000):\n        sess.run(s)\nprint('time', time() - start)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52996937,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-25T19:42:51.333","AcceptedAnswerId":null,"OwnerUserId":3241790.0,"Title":"Keras Custom method has y_pred shape different from output of model","Body":"<p>I am working on custom loss function and the y_pred I got has shape is different from what is output in the summary<\/p>\n\n<pre><code>conv2d_100 (Conv2D)             (None, 3, 360, 512)  85          leaky_re_lu_27[0][0]             \n==================================================================================================\nTotal params: 4,128,849\nTrainable params: 4,119,761\nNon-trainable params: 9,088\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>This the output I got: <\/p>\n\n<pre><code>Defining loss function\ny_true [None, None, None]\ny_pred [None, 3, None, 512]\n<\/code><\/pre>\n\n<p>The code is this:<\/p>\n\n<pre><code>def focal_loss(content, label_remap, gamma_=2, w_d=1e-4):\n    def focal_loss_fixed(y_true, y_pred):\n        num_classes = len(content.keys())\n        print(\"y_true_b\", y_true.get_shape().as_list())\n\n        cv_eqation = K.constant([0.114, 0.587, 0.299])\n        y_true = tf.multiply(y_true, cv_eqation)\n        y_true = tf.reduce_sum(y_true, axis=3)\n        y_true = tf.cast(y_true, dtype=tf.uint8)\n\n        print(\"Defining loss function\")\n        print(\"y_true\", y_true.get_shape().as_list())\n        print(\"y_pred\", y_pred.get_shape().as_list())\n        .\n        .\n        return loss\nreturn focal_loss_fixed\n<\/code><\/pre>\n\n<p>So, why I have <strong>None<\/strong> instead of the high of the output?<\/p>\n","answers":[{"AnswerId":"52997069","CreationDate":"2018-10-25T19:52:08.870","ParentId":null,"OwnerUserId":"3241790","Title":null,"Body":"<p>I found the answer in different <a href=\"https:\/\/stackoverflow.com\/posts\/46667294\/revisions\">question<\/a> to get the right shape you should use <\/p>\n\n<pre><code>K.int_shape(y_pred)\n<\/code><\/pre>\n"}]}
+{"QuestionId":52997654,"AnswerCount":2,"Tags":"<tensorflow><google-compute-engine><google-cloud-tpu>","CreationDate":"2018-10-25T20:31:22.887","AcceptedAnswerId":null,"OwnerUserId":2341218.0,"Title":"training MNIST with TPU generates errors","Body":"<p>Following the <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tutorials\/mnist\" rel=\"nofollow noreferrer\">Running MNIST on Cloud TPU<\/a> tutorial:<\/p>\n\n<p>I get the following error when I try to train:<\/p>\n\n<pre><code>python \/usr\/share\/models\/official\/mnist\/mnist_tpu.py \\\n  --tpu=$TPU_NAME \\\n  --DATA_DIR=${STORAGE_BUCKET}\/data \\\n  --MODEL_DIR=${STORAGE_BUCKET}\/output \\\n  --use_tpu=True \\\n  --iterations=500 \\\n  --train_steps=2000\n<\/code><\/pre>\n\n<p>=><\/p>\n\n<pre><code>alexryan@alex-tpu:~\/tpu$ .\/train-mnist.sh \nW1025 20:21:39.351166 139745816463104 __init__.py:44] file_cache is unavailable when using oauth2client &gt;= 4.0.0 or google-auth\nTraceback (most recent call last):\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/googleapiclient\/discovery_cache\/__init__.py\", line 41, in autodetect\n    from . import file_cache\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/googleapiclient\/discovery_cache\/file_cache.py\", line 41, in &lt;module&gt;\n    'file_cache is unavailable when using oauth2client &gt;= 4.0.0 or google-auth')\nImportError: file_cache is unavailable when using oauth2client &gt;= 4.0.0 or google-auth\nTraceback (most recent call last):\n  File \"\/usr\/share\/models\/official\/mnist\/mnist_tpu.py\", line 173, in &lt;module&gt;\n    tf.app.run()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"\/usr\/share\/models\/official\/mnist\/mnist_tpu.py\", line 152, in main\n    tpu_config=tf.contrib.tpu.TPUConfig(FLAGS.iterations, FLAGS.num_shards),\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/tpu_config.py\", line 207, in __init__\n    self._master = cluster.master()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/cluster_resolver\/python\/training\/tpu_cluster_resolver.py\", line 223, in master\n    job_tasks = self.cluster_spec().job_tasks(self._job_name)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/cluster_resolver\/python\/training\/tpu_cluster_resolver.py\", line 269, in cluster_spec\n    (compat.as_text(self._tpu), response['health']))\nRuntimeError: TPU \"alex-tpu\" is unhealthy: \"TIMEOUT\"\nalexryan@alex-tpu:~\/tpu$ \n<\/code><\/pre>\n\n<p>The only places where I varied from the instructions were:<\/p>\n\n<p>Instead of running ctpu in the cloud shell, I ran it on the mac.<\/p>\n\n<pre><code>&gt;ctpu version\nctpu version: 1.7\n<\/code><\/pre>\n\n<p>The zone where the TPU resided was different than the default zone of my config, so I specified it as an option like so:<\/p>\n\n<pre><code>&gt;cat ctpu-up.sh \nctpu up --zone us-central1-b --preemptible\n<\/code><\/pre>\n\n<p>I was able to move the MNIST files to the gcs bucket from the vm no problem:<\/p>\n\n<pre><code>alexryan@alex-tpu:~$ gsutil cp -r .\/data ${STORAGE_BUCKET}\nCopying file:\/\/.\/data\/validation.tfrecords [Content-Type=application\/octet-stream]...\nCopying file:\/\/.\/data\/train-images-idx3-ubyte.gz [Content-Type=application\/octet-stream]...\n<\/code><\/pre>\n\n<p>I tried the <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tensorboard-setup\" rel=\"nofollow noreferrer\">(Optional) Set up TensorBoard<\/a> >\n<a href=\"https:\/\/cloud.google.com\/tpu\/docs\/cloud-tpu-tools#install_cloud_tpu_profiler\" rel=\"nofollow noreferrer\">Running cloud_tpu_profiler<\/a><\/p>\n\n<blockquote>\n  <p>Go to the Cloud Console > TPUs > and click on the TPU you created.\n  Locate the service account name for the Cloud TPU and copy it, for\n  example:<\/p>\n  \n  <p>service-11111111118@cloud-tpu.iam.myserviceaccount.com<\/p>\n  \n  <p>In the list of buckets, select the bucket you want to use, select Show\n  Info Panel, and then select Edit bucket permissions. Paste your\n  service account name into the add members field for that bucket and\n  select the following permissions:<\/p>\n<\/blockquote>\n\n<p>\"Cloud Console > TPUs\" does not exist as an option<br>\nso I used the service account associate with the VM<br>\n\"Cloud Console > Compute Engine > alex-tpu\"<\/p>\n\n<p>since the last error message was \"RuntimeError: TPU \"alex-tpu\" is unhealthy: \"TIMEOUT\", I used ctpu to delete the vm and re-create it and ran it again.\nThis time I got more errors:<\/p>\n\n<p>This one seems like it might be just a warning ...<\/p>\n\n<pre><code>ImportError: file_cache is unavailable when using oauth2client &gt;=\n4.0.0 or google-auth\n<\/code><\/pre>\n\n<p>Not sure about this one ...<\/p>\n\n<pre><code>ERROR:tensorflow:Operation of type Placeholder (reshape_input) is not supported on the TPU. Execution will fail if this op is used in the graph. \n<\/code><\/pre>\n\n<p>this one seemed to kill the training ...<\/p>\n\n<pre><code>INFO:tensorflow:Error recorded from training_loop: File system scheme '[local]' not implemented (file: '\/tmp\/tmpaiggRW\/model.ckpt-0_temp_9216e11a1368405795d9b5282775f562')      [[{{node save\/SaveV2}} = SaveV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64],\n_device=\"\/job:worker\/replica:0\/task:0\/device:CPU:0\"](save\/ShardedFilename, save\/SaveV2\/tensor_names, save\/SaveV2\/shape_and_slices, conv2d\/bias\/Read\/ReadVariableOp, conv2d\/kernel\/Read\/ReadVariableOp, conv2d_1\/bias\/Read\/ReadVariableOp, conv2d_1\/kernel\/Read\/ReadVariableOp, dense\/bias\/Read\/ReadVariableOp, dense\/kernel\/Read\/ReadVariableOp, dense_1\/bias\/Read\/ReadVariableOp, dense_1\/kernel\/Read\/ReadVariableOp, global_step\/Read\/ReadVariableOp)]]\n\nCaused by op u'save\/SaveV2', defined at:   File \"\/usr\/share\/models\/official\/mnist\/mnist_tpu.py\", line 173, in &lt;module&gt;\n    tf.app.run()   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))   File \"\/usr\/share\/models\/official\/mnist\/mnist_tpu.py\", line 163, in main\n    estimator.train(input_fn=train_input_fn, max_steps=FLAGS.train_steps)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/tpu_estimator.py\", line 2394, in train\n    saving_listeners=saving_listeners   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/estimator\/estimator.py\", line 356, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1181, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1215, in _train_model_default\n    saving_listeners)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1406, in _train_with_estimator_spec\n    log_step_count_steps=self._config.log_step_count_steps) as mon_sess:   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 504, in MonitoredTrainingSession\n    stop_grace_period_secs=stop_grace_period_secs)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 921, in __init__\n    stop_grace_period_secs=stop_grace_period_secs)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 643, in __init__\n    self._sess = _RecoverableSession(self._coordinated_creator)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1107, in __init__\n    _WrappedSession.__init__(self, self._create_session())   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1112, in _create_session\n    return self._sess_creator.create_session()   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 800, in create_session\n    self.tf_sess = self._session_creator.create_session()   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 557, in create_session\n    self._scaffold.finalize()   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 215, in finalize\n    self._saver.build()   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1106, in build\n    self._build(self._filename, build_save=True, build_restore=True)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1143, in _build\n    build_save=build_save, build_restore=build_restore)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 778, in _build_internal\n    save_tensor = self._AddShardedSaveOps(filename_tensor, per_device)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 369, in _AddShardedSaveOps\n    return self._AddShardedSaveOpsForV2(filename_tensor, per_device)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 343, in _AddShardedSaveOpsForV2\n    sharded_saves.append(self._AddSaveOps(sharded_filename, saveables))   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 284, in _AddSaveOps\n    save = self.save_op(filename_tensor, saveables)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 202, in save_op\n    tensors)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/ops\/gen_io_ops.py\", line 1690, in save_v2\n    shape_and_slices=shape_and_slices, tensors=tensors, name=name)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/util\/deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3272, in create_op\n    op_def=op_def)   File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 1768, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nUnimplementedError (see above for traceback): File system scheme '[local]' not implemented (file: '\/tmp\/tmpaiggRW\/model.ckpt-0_temp_9216e11a1368405795d9b5282775f562')\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I get this error ...<\/p>\n\n<pre><code>INFO:tensorflow:Error recorded from training_loop: File system scheme '[local]' not implemented \n<\/code><\/pre>\n\n<p>... even when --use_tpu=False<\/p>\n\n<pre><code>alexryan@alex-tpu:~\/tpu$ cat train-mnist.sh \npython \/usr\/share\/models\/official\/mnist\/mnist_tpu.py \\\n  --tpu=$TPU_NAME \\\n  --DATA_DIR=${STORAGE_BUCKET}\/data \\\n  --MODEL_DIR=${STORAGE_BUCKET}\/output \\\n  --use_tpu=False \\\n  --iterations=500 \\\n  --train_steps=2000\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/52860914\/2341218\">This stack overflow answer<\/a> suggests that the tpu is trying to write to a non-existent file system instead of the gcs bucket I specified. It is unclear to me why that would be happening.<\/p>\n","answers":[{"AnswerId":"52999485","CreationDate":"2018-10-25T23:35:24.557","ParentId":null,"OwnerUserId":"3870291","Title":null,"Body":"<p>In the first scenario, it seems the TPU you created is not in healthy state. So, deleting and recreating the TPU or the entire VM is the right way to resolve this.<\/p>\n\n<p>I think the error comes in second scenario (where you deleted the vm and re-created it again) is because your ${STORAGE_BUCKET} is either undefined or not a proper GCS bucket. It should be a GCS bucket. Local path won't work and gives the following error. \nMore information on creating a GCS bucket is in the section \"Create a Cloud Storage bucket\" at <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tutorials\/mnist\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/tpu\/docs\/tutorials\/mnist<\/a><\/p>\n\n<p>Hope this answers your question.<\/p>\n"},{"AnswerId":"53466481","CreationDate":"2018-11-25T10:11:14.897","ParentId":null,"OwnerUserId":"2344466","Title":null,"Body":"<p>Ran into the same problem and found that there was a typo in the tutorial. If you check <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/official\/mnist\/mnist_tpu.py\" rel=\"nofollow noreferrer\">mnist_tpu.py<\/a> you'll find that the params need to be lowercase.<\/p>\n\n<p>If you change that, it works fine.<\/p>\n\n<pre><code>python \/usr\/share\/models\/official\/mnist\/mnist_tpu.py \\\n  --tpu=$TPU_NAME \\\n  --data_dir=${STORAGE_BUCKET}\/data \\\n  --model_dir=${STORAGE_BUCKET}\/output \\\n  --use_tpu=True \\\n  --iterations=500 \\\n  --train_steps=2000\n<\/code><\/pre>\n"}]}
+{"QuestionId":52997784,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><mnist>","CreationDate":"2018-10-25T20:41:15.410","AcceptedAnswerId":52998302.0,"OwnerUserId":7490161.0,"Title":"Can't get accuracy in own digit image dataset for mnist","Body":"<p>I am new to machine learning and I tried the mnist dataset and I got an accuracy of around 97% but then I tried working on my image dataset and I got an accuracy of 0%. Please help me out.<\/p>\n\n<p>This is the 97% accuracy model code:<\/p>\n\n<pre><code>from keras.models import Sequential                  \nfrom keras.layers import Dense, Dropout, Conv2D, Flatten  \nfrom keras.callbacks import ModelCheckpoint               \n\n\nx_train = tf.keras.utils.normalize(x_train, axis =1)\nx_test = tf.keras.utils.normalize(x_test, axis = 1)\n\nmodel = Sequential()\nmodel.add(Flatten())\nmodel.add(Dense(128, activation = 'relu')) \nmodel.add(Dense(128, activation = 'relu'))\nmodel.add(Dense(10, activation = 'softmax')) \n\n\nmodel.compile(optimizer = 'adam', loss = 'sparse_categorical_crossentropy', metrics = ['accuracy'])\n\ncheckpointer = ModelCheckpoint(filepath = 'mnist.model.weights.best.hdf5',verbose = 1,save_best_only = True, monitor = 'loss')\n\nmodel.fit(x_train, y_train, epochs = 3, callbacks = [checkpointer], \n          batch_size = 32,verbose = 2,shuffle = True)\n<\/code><\/pre>\n\n<p>Now I tried with my 10 images and none of them were predicted correctly.\nBelow is the code:<\/p>\n\n<pre><code>from skimage import io\nfrom skimage import color\nimport numpy as np\nimport tensorflow as tf\nimport keras\n\nimg_0 = color.rgb2gray(io.imread(\"0.jpg\"))\nimg_2 = color.rgb2gray(io.imread(\"2.jpg\"))\nimg_3 = color.rgb2gray(io.imread(\"3.jpg\"))\nimg_4 = color.rgb2gray(io.imread(\"4.jpg\"))\nimg_5 = color.rgb2gray(io.imread(\"5.jpg\"))\nimg_6 = color.rgb2gray(io.imread(\"6.jpg\"))\nimg_7 = color.rgb2gray(io.imread(\"7.jpg\"))\nimg_8 = color.rgb2gray(io.imread(\"8.jpg\"))\nimg_9 = color.rgb2gray(io.imread(\"9.jpg\"))\narray = [img_0, img_2, img_3, img_4, img_5, img_6, img_7, img_8, img_9]\n\n#normalized the data between 0-1\narray = tf.keras.utils.normalize(array, axis = 1)\n\n#used the loop to increase the dimensions of the input layer as 1,28,28 which will be converted into 1*784\nfor i in array:\n    i = np.expand_dims(i,axis = 0)\n    print(i.shape)\n\n\nnew_model = tf.keras.models.load_model('mnist_save.model')\nnew_model.load_weights('mnist.model.weights.best.hdf5')\npredictions = new_model.predict(array)\n<\/code><\/pre>\n\n<p>Can you please help me out with my problem.<\/p>\n","answers":[{"AnswerId":"52998302","CreationDate":"2018-10-25T21:22:40.160","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>If I were you, I will check the following three things.<\/p>\n\n<p><strong>1. Visualize both training and testing data side by side<\/strong><\/p>\n\n<p>This is the simplest way to see whether the low performance is reasonable. Basically, if the testing data is very different looking than the training data, there is no way for your pretrained model to achieve high performance in this new testing domain. Even this is not the case, visualization should be helpful to decide what simple domain adaptation can be applied to achieve better performance.<\/p>\n\n<p><strong>2. Double check with your L2normalization<\/strong><\/p>\n\n<p>I took a look at the source code of <code>keras.utils.normalize<\/code><\/p>\n\n<pre><code>@tf_export('keras.utils.normalize')\ndef normalize(x, axis=-1, order=2):\n  \"\"\"Normalizes a Numpy array.\n  Arguments:\n      x: Numpy array to normalize.\n      axis: axis along which to normalize.\n      order: Normalization order (e.g. 2 for L2 norm).\n  Returns:\n      A normalized copy of the array.\n  \"\"\"\n  l2 = np.atleast_1d(np.linalg.norm(x, order, axis))\n  l2[l2 == 0] = 1\n  return x \/ np.expand_dims(l2, axis)\n<\/code><\/pre>\n\n<p>Since you are using the tensorflow backend, <code>normalize<\/code> along the 1st axis means what? Normalize each row? This is strange. The right way to do normalization is to (1) vectorize your input image, i.e. each image becomes a vector; and (2) <code>normalize<\/code> the resulting vector (at axis=1).<\/p>\n\n<p>Actually, it is somewhat inappropriate especially when you want to apply a pretrained model in a different domain. This is because L2normalization is more sensitive to nonzero values. In MNIST samples, almost binarized, i.e. either 0s or 1s. However, in a grayscale image, you may meet values in [0,255], which is a completely different distribution. <\/p>\n\n<p>You may try the simple (0,1) normalization, i.e. <\/p>\n\n<pre><code>x_normalized = (x-min(x))\/(max(x)-min(x))\n<\/code><\/pre>\n\n<p>but this requires you to retrain a new model.<\/p>\n\n<p><strong>3. Apply domain adaptation techniques<\/strong><\/p>\n\n<p>This means you want to do the following things before feeding a testing image to your model (even before normalization).<\/p>\n\n<ul>\n<li>binarize your testing image, i.e. convert to 0\/1 images<\/li>\n<li>negate your testing image, i.e. make 0s to 1s and 1s to 0s<\/li>\n<li>centralize your testing image, i.e. shift your image such that its mass center is the image center.<\/li>\n<\/ul>\n\n<p>Of course, what techniques to apply is dependent on the domain differences that you observe in visualization results.<\/p>\n"}]}
+{"QuestionId":52997810,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><convolution>","CreationDate":"2018-10-25T20:43:10.730","AcceptedAnswerId":null,"OwnerUserId":10559900.0,"Title":"Kernel size change in convolutional neural networks","Body":"<p>I have been working on creating a convolutional neural network from scratch, and am a little confused on how to treat kernel size for hidden convolutional layers. For example, say I have an <code>MNIST<\/code> image as <code>input (28 x 28)<\/code> and put it through the following layers.<\/p>\n\n<p>Convolutional layer with kernel_size = (5,5) with 32 output channels<\/p>\n\n<ul>\n<li>new dimension of throughput = (32, 28, 28)<\/li>\n<\/ul>\n\n<p>Max Pooling layer with pool_size (2,2) and step (2,2)<\/p>\n\n<ul>\n<li>new dimension of throughput = (32, 14, 14)<\/li>\n<\/ul>\n\n<p>If I now want to create a second convolutional layer with kernel size = (5x5) and 64 output channels, how do I proceed? Does this mean that I only need two new filters <code>(2 x 32 existing channels)<\/code> or does the kernel size change to be <code>(32 x 5 x 5)<\/code> since there are already 32 input channels? <\/p>\n\n<p>Since the initial input was a 2D image, I do not know how to conduct convolution for the hidden layer since the input is now 3 dimensional <code>(32 x 14 x 14)<\/code>.<\/p>\n","answers":[{"AnswerId":"52998511","CreationDate":"2018-10-25T21:42:39.183","ParentId":null,"OwnerUserId":"10528078","Title":null,"Body":"<p>you need 64 kernel, each with the size of (32,5,5) . <\/p>\n\n<p>depth(#channels) of kernels, 32 in this case, or 3 for a RGB image, 1 for gray scale etc, should always match the input depth, but values are all the same.\ne.g. if you have a 3x3 kernel like this : [-1 0 1; -2 0 2; -1 0 1]  and now you want to convolve it with an input with <strong>N<\/strong> as depth or say channel, you just copy this 3x3 kernel <strong>N<\/strong> times in 3rd dimension, the following math is just like the 1 channel case, you sum all values in all <strong>N<\/strong> channels which your kernel window is currently on them after multiplying the kernel values with them and get the value of just 1 entry or pixel. so what you get as output in the end is a matrix with 1 channel:) how much depth you want your matrix for next layer to have? that's the number of kernels you should apply. hence in your case it would be a kernel with this size (64 x 32 x 5 x 5) which is actually 64 kernels with 32 channels for each and same 5x5 values in all cahnnels.<\/p>\n\n<p>(\"I am not a very confident english speaker hope you get what I said, it would be nice if someone edit this :)\")<\/p>\n"}]}
+{"QuestionId":52998303,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-25T21:22:42.637","AcceptedAnswerId":null,"OwnerUserId":3241790.0,"Title":"Keras get the number of image in batch from y_pred in side custom loss function","Body":"<p>In this loss function, I need to create full metrics depends on the number of images in batch and size of the image. However, I can get the image size from y_pred but no the batch size because it is coming as None while initializing the graph.<\/p>\n\n<pre><code>def focal_loss(content, label_remap, gamma_=2, w_d=1e-4):\ndef focal_loss_fixed(y_true, y_pred):\n    num_classes = len(content.keys())\n    print(\"y_true_b\", y_true.get_shape().as_list())\n\n    cv_eqation = K.constant([0.114, 0.587, 0.299])\n    y_true = tf.multiply(y_true, cv_eqation)\n    y_true = tf.reduce_sum(y_true, axis=3)\n    y_true = tf.cast(y_true, dtype=tf.uint8)\n\n    lbls_resized = y_true\n    logits_train = y_pred\n\n    b, c, w, h = K.int_shape(y_pred)\n    batch = K.constant(b)\n    channel = K.variable(c)\n    width = K.variable(w)\n    high = K.variable(h)\n    with tf.variable_scope(\"loss\"):\n        ......\n        # make the labels one-hot for the cross-entropy\n        onehot_mat = tf.reshape(tf.one_hot(lbls_resized, num_classes), (-1, num_classes))\n\n        # focal loss p and gamma\n        gamma = np.full((high * width * batch, channel), fill_value=gamma_)\n        print(\"gamma\", gamma.shape)\n        .........\n    return loss\nreturn focal_loss_fixed\n<\/code><\/pre>\n\n<p>Also, I tried a different way by using onehot_mat shape but it has none value in its shape.<\/p>\n","answers":[{"AnswerId":"53007842","CreationDate":"2018-10-26T11:33:16.497","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>Please refere to <a href=\"https:\/\/stackoverflow.com\/a\/37096395\/977098\">this answer<\/a>. It sounds like you want to know the dynamic shape of the tensor but are using the static shape instead. You need to use <code>tf.shape<\/code> to get the dynamic shape of your batch during runtime instead of <code>get_shape<\/code> which only returns the static shape that is known while constructing the network.<\/p>\n\n<p><strong>Update:<\/strong>\nFor your particular tasks it looks to me like you are trying to create a tensor that has a shape dependend on the current batch size. I think you could do something like this:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nvector = tf.Variable(tf.random_normal([3], stddev=0.1), name=\"weights\")\nbatch = tf.placeholder(tf.float32, shape=[None,2,2])\n\nvector_times_batchsize = tf.tile(vector, tf.shape(batch)[0:1])\n\ninit_all_op = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init_all_op)\n    print sess.run(vector_times_batchsize, feed_dict={batch: np.zeros((5,2,2), np.float32)}).shape\n<\/code><\/pre>\n\n<p>This creates a new tensor by repeating the <code>gamma_image<\/code> tensor depending on the shape of the first dimension of the tensor <code>batch<\/code>.<\/p>\n\n<p>Please be aware that you can't use numpy functions since the creation of this tensor needs to be part of the tensorflow graph because the sice is only known during run time but not while the graph is created.<\/p>\n"}]}
+{"QuestionId":52998304,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-10-25T21:22:42.827","AcceptedAnswerId":null,"OwnerUserId":2117367.0,"Title":"Adding comments inside tensorflow object detection pipeline configuration file","Body":"<p>How to add comments inside tensorflow object detection API's pipeline configuration file (typically named <code>pipeline.conf<\/code>)?<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/configuring_jobs.md\" rel=\"nofollow noreferrer\">corresponding official documentation<\/a> does not mention any way to do so.<\/p>\n","answers":[{"AnswerId":"53029673","CreationDate":"2018-10-28T08:31:12.383","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Simply use <code># comment here<\/code> wherever you want.\nKeep in mind that the pipeline.config is written into the model_dir. Therefore, if it is the same directory as your pipeline.config, it will be overridden, and the comments will be deleted. You can simply avoid this by naming the pipeline.config file you use differently.<\/p>\n"}]}
+{"QuestionId":52999283,"AnswerCount":1,"Tags":"<tensorflow><tfrecord>","CreationDate":"2018-10-25T23:05:44.980","AcceptedAnswerId":52999606.0,"OwnerUserId":9984948.0,"Title":"tf.data API read the TFRecord files","Body":"<p>I am trying to use the tf.data API read the TFRecord file.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom PIL import Image\nimport numpy as np\nimport os\n\ndef train_input_fn():\n    filenames = [\"mytrain.tfrecords\"]\n    dataset = tf.data.TFRecordDataset(filenames)\n\n    def parser(record):\n        keys_to_features = {\n            \"image_data\": tf.FixedLenFeature((), tf.string, default_value=\"\"),\n            \"date_time\": tf.FixedLenFeature((), tf.int64, default_value=\"\"),\n            \"label\": tf.FixedLenFeature((), tf.int64,\n                                        default_value=tf.zeros([], dtype=tf.int64)),\n        }\n        parsed = tf.parse_single_example(record, keys_to_features)\n\n        image = tf.decode_jpeg(parsed[\"image_data\"])\n        image = tf.reshape(image, [128, 128, 3])\n        label = tf.cast(parsed[\"label\"], tf.int32)\n\n        return {\"image_data\": image, \"date_time\": parsed[\"date_time\"]}, label\n\n    dataset = dataset.map(parser)\n    dataset = dataset.shuffle(buffer_size=10000)\n    dataset = dataset.batch(32)\n    dataset = dataset.repeat(1)\n    iterator = dataset.make_one_shot_iterator()\n\n    features, labels = iterator.get_next()\n    return features, labels\n\noutput = train_input_fn()\n\ninit_op = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init_op)\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord = coord)\n    for i in range(230):\n        image, label = sess.run(output)\n        img = Image.fromarray(image, 'RGB')\n        img.save(cwd+str(i) + '_''Label_'+str(l)+'.jpg')\n        print(image, label)\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>Traceback (most recent call last):\n  File \"E:\/Tensorflow\/Wenshan_Cai_Nanoletters\/tf_data.py\", line 34, in \n    output = train_input_fn()\n  File \"E:\/Tensorflow\/Wenshan_Cai_Nanoletters\/tf_data.py\", line 25, in train_input_fn\nTypeError: Expected int64, got '' of type 'str' instead.<\/p>\n","answers":[{"AnswerId":"52999606","CreationDate":"2018-10-25T23:51:54.950","ParentId":null,"OwnerUserId":"2117367","Title":null,"Body":"<p>Note <code>TypeError: Expected int64, got '' of type 'str' instead<\/code> from your error log. You have a bug in your code. <\/p>\n\n<p><strong>The bug<\/strong><\/p>\n\n<p>In the following line:<\/p>\n\n<p><code>\"date_time\": tf.FixedLenFeature((), tf.int64, default_value=\"\"),<\/code><\/p>\n\n<p>The default value for a <code>tf.int64<\/code> type variable is specified as a string <code>\"\"<\/code>.<\/p>\n\n<p><strong>A fix<\/strong><\/p>\n\n<p>So say your expected default is 0, then you should change line to:<\/p>\n\n<p><code>\"date_time\": tf.FixedLenFeature((), tf.int64, default_value=0),<\/code><\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
+{"QuestionId":53000016,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-10-26T01:02:05.223","AcceptedAnswerId":53000798.0,"OwnerUserId":2117367.0,"Title":"Tensorflow object detection API: How to disable loading from checkpoint","Body":"<p>I have created a custom variation of MobileNetV2 feature extractor architecture, by changing the <code>expansion_size<\/code> from 6 to 4 in <code>research\/slim\/nets\/mobilenet\/mobilenet_v2.py<\/code> of <code>tensorflow\/models<\/code> repo. <\/p>\n\n<p>I want to be able to train the SSD + Mobilenet_v2 (with this change) model with <code>model_main.py<\/code> script as described at Object Detection API's <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/running_locally.md.\" rel=\"nofollow noreferrer\">running_locally tutorial<\/a>. <\/p>\n\n<p>When I do so I see the following error, which makes sense:<\/p>\n\n<pre><code>`InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint.\n<\/code><\/pre>\n\n<p>To address this:<\/p>\n\n<ol>\n<li>I removed the <code>finetune_checkpoint<\/code> specification from my <code>pipeline.config<\/code>.<\/li>\n<li>I changed <code>load_pretrained=True<\/code> to <code>load_pretrained=False<\/code> in <code>object_detection\/model_hparams.py<\/code>.<\/li>\n<li>I added <code>--hparams_overrides='load_pretrained=false'<\/code> as a command line input argument to <code>model_main.py<\/code>.<\/li>\n<\/ol>\n\n<p>Despite of these, I still see the same error.<\/p>\n\n<p>Why is tensorflow still trying to restore a checkpoint. How can I make it not do so?<\/p>\n","answers":[{"AnswerId":"53000798","CreationDate":"2018-10-26T02:58:56.280","ParentId":null,"OwnerUserId":"2117367","Title":null,"Body":"<p>Found the solution myself. It turns out that even though I had made arrangements for it to not restore checkpoint from my pipeline configuration file, it turns out that the internal <code>tf.Estimator<\/code> object can still use a checkpoint from the <code>model_dir<\/code> specified; even though the primary use of <code>model_dir<\/code> is as an output directory, where output checkpoints are written to.<\/p>\n\n<p>I found this information in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\">official documentation for tf.Estimator<\/a>. Here's the relevant excerpt for reference:<\/p>\n\n<blockquote>\n  <p>`model_dir: Directory to save model parameters, graph and etc. <strong>This can also be used to load checkpoints from the directory into an estimator to continue training a previously saved model<\/strong>. If PathLike object, the path will be resolved. If None, the model_dir in config will be used if set. If both are set, they must be same. If both are None, a temporary directory will be used.<\/p>\n<\/blockquote>\n\n<p>I had an old checkpoint sitting in my original <code>model_dir<\/code> which was architecturally incompatible with my custom model. Hence I was seeing the error. To resolve it, I simply changed my <code>model_dir<\/code> to point to a new empty directory and it worked. I hope that helps someone with a similar problem.<\/p>\n"}]}
+{"QuestionId":53000910,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-26T03:16:55.877","AcceptedAnswerId":null,"OwnerUserId":10153461.0,"Title":"how to check the quantized weights in tflite files","Body":"<p>When I quantized the model by lite modules in <code>tensorflow<\/code>, I cann't check the weights values that had <code>quantized<\/code>.Is there any way to  view these values in the <code>.tflite<\/code> files? or Is there any way to parse the <code>.tflite<\/code> files?<\/p>\n","answers":[{"AnswerId":"53054910","CreationDate":"2018-10-29T22:37:47.310","ParentId":null,"OwnerUserId":"10530640","Title":null,"Body":"<p>There are some neural network visualizers that can also provide an interface to inspect the file. I have been using <a href=\"https:\/\/github.com\/lutzroeder\/Netron\" rel=\"nofollow noreferrer\">Netron<\/a>. You can click on the \"weights\" tab of the layer you are interested in to view the data. I haven't tried it yet, but there appears to be a floppy-disk save icon when you view weights\/biases in the right side-bar.<\/p>\n\n<p>The data format is a Google FlatBuffer, defined by the schema file <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lite\/schema\/schema.fbs\" rel=\"nofollow noreferrer\">here<\/a>. You may prefer doing this if you want to do <em>something<\/em> with the data like output it in a different format. I found the output from parsing it myself using the schema.fbs file to match Netron's for the CNN's I passed in. You can check out the FlatBuffers documentation <a href=\"http:\/\/google.github.io\/flatbuffers\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"53096272","CreationDate":"2018-11-01T06:36:35.763","ParentId":null,"OwnerUserId":"8049751","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/50646958\/how-to-know-tensorflow-lite-models-input-output-feature-info\">here<\/a> in first answer is guide, how to create json view of .tflite model. There you can see quantized values<\/p>\n"}]}
+{"QuestionId":53000921,"AnswerCount":0,"Tags":"<python><tensorflow><helper><seq2seq>","CreationDate":"2018-10-26T03:18:29.567","AcceptedAnswerId":null,"OwnerUserId":7644642.0,"Title":"Tensorflow Sequence to Sequence CustomHelper","Body":"<p>There are limited amount of documentation on Sequence to Sequence CustomHelper <\/p>\n\n<p><code>helper = tf.contrib.seq2seq.CustomHelper(initialize_fn = initialize_fn,sample_fn = sample_fn, next_inputs_fn = next_inputs_fn)<\/code><\/p>\n\n<p>in Tensorflow.<\/p>\n\n<p>Would anyone explain the inputs of the Custom-helper, in terms of the input data<\/p>\n\n<p><code>X = tf.placeholder(tf.float32, [batch_size x time_steps x features])<\/code><\/p>\n\n<p>and encoder, <\/p>\n\n<p><code>encoder_cell = tf.contrib.rnn.BasicLSTMCell(hidden_size)<\/code><\/p>\n\n<p><code>initial_state = encoder_cell.zero_state(batch_size, dtype=tf.float32)<\/code><\/p>\n\n<p>and\/or possibly <\/p>\n\n<p><code>rnn_output, rnn_states = tf.nn.dynamic_rnn(encoder_cell, X, dtype=tf.float32)<\/code>\n?<\/p>\n","answers":[]}
+{"QuestionId":53001124,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-26T03:48:25.590","AcceptedAnswerId":53002789.0,"OwnerUserId":10437878.0,"Title":"Neural net inside for loop","Body":"<p>I have smth like this<\/p>\n\n<pre><code>for q in range(10):\n   # generate some samples\n   x = Input(batch_shape=(n_batch, xx.shape[1]))\n   x = Dense(20)(x)\n   x = LeakyReLU(alpha=0.001)(x)\n   y = Dense(1)(x)\n   y = LeakyReLU(alpha=0.001)(y)\n   model = Model(inputs=x, outputs=y) \n   model.compile(loss='mean_squared_error', optimizer='Adam', metrics=['accuracy'])\n   for i in range(10):\n      model.fit(x, y, epochs=1, batch_size=n_batch, verbose=0, shuffle=False)\n      model.reset_states()\n<\/code><\/pre>\n\n<p>I wonder if the neural net is built from scratch for every q or it retains everything from the previous q? If it retains, how do I reset and build, compile and fit neural net separately for every q?<\/p>\n","answers":[{"AnswerId":"53002789","CreationDate":"2018-10-26T06:33:50.920","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>When you make a layer with keras or tensorflow, tensorflow adds a node or more to its graph, every time you add an optimizer, a loss function or an activation function it does the same thing and adds a node for them.<\/p>\n\n<p>When you call <code>model.fit()<\/code> tensorflow executes its graph starting from its root. If you add your nodes in a loop, previous ones will not be deleted. They will take space in memory, and they will lower your performance.<\/p>\n\n<p>What to do instead? Its very simple, re-initialize your weights and re-use the same nodes. Your code won't change much just move the sample generation down with the for loop and define a function to re-initialize.<\/p>\n\n<p>I also took the second for loop down and just increased the epoch number to 10, you can put that for loop back if you have a reason to have it there.<\/p>\n\n<pre><code>def reset_weights(model):\n    session = K.get_session()\n    for layer in model.layers: \n        if hasattr(layer, 'kernel_initializer'):\n            layer.kernel.initializer.run(session=session)\n\nx = Input(batch_shape=(n_batch, xx.shape[1]))\nx = Dense(20)(x)\nx = LeakyReLU(alpha=0.001)(x)\ny = Dense(1)(x)\ny = LeakyReLU(alpha=0.001)(y)\nmodel = Model(inputs=x, outputs=y) \nmodel.compile(loss='mean_squared_error', optimizer='Adam', metrics=['accuracy'])\nfor q in range(10):\n    #generate some samples\n    model.fit(x, y, epochs=10, batch_size=n_batch, verbose=1, shuffle=False)\n    model.reset_states()\n    reset_weights(model)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53001403,"AnswerCount":1,"Tags":"<python><tensorflow><dataset>","CreationDate":"2018-10-26T04:24:18.823","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Tensorflow: Strange behavior with splitting string in tf.data.Dataset","Body":"<p>I'm using the <code>tf.data.Dataset<\/code> API in Tensorflow.  I have 2 numpy arrays, where <code>data<\/code> is 2-d and <code>labels<\/code> is 1-d.  I've created a <code>Dataset<\/code> like this:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((data, labels))\nval_dataset = dataset.map(lambda x, y: ({'reviews': x}, y))\n<\/code><\/pre>\n\n<p>I have a preprocessing function I'd like to use that looks like this:<\/p>\n\n<pre><code>def preprocess(x, y):\n    # split on whitespace\n    x['reviews'] = tf.string_split(x['reviews'])\n    return x, y\n<\/code><\/pre>\n\n<p>I try to use <code>map<\/code> like this:<\/p>\n\n<pre><code>dataset = dataset.map(preprocess)\n<\/code><\/pre>\n\n<p>but I get back:<\/p>\n\n<pre><code>ValueError: Shape must be rank 1 but is rank 0 for 'StringSplit' (op: 'StringSplit') with input shapes: [], [].\n<\/code><\/pre>\n\n<p>I googled around and found that someone suggested this approach in the preprocessing function:<\/p>\n\n<pre><code>x['reviews'] = tf.string_split([x['reviews']])\n<\/code><\/pre>\n\n<p>But it's unclear to me why I would do that.  It doesn't error out as before, but the shape of my data is incorrect.  For instance, this is what I see for the first element in my <code>dataset<\/code>:<\/p>\n\n<pre><code>({'sequence': array([[ 6391,  3352, 10236,   244,  1362,   244,  9350,  7649,  6391,\n         6324,  6063,  3620,   244,  8153,  6542, 10056,  7303,  1955,\n         1362,  6194, 10250,  6391,   550,   244,  7577,   850,  3620,\n         5807, 10325,  1362,  6542,   595,  9060,  9052,  9459,   351,\n         4676,  9354,  7648,  3082,  7694,  8497, 10703,  1610,  9454,\n        10236,   244,  7965,  8018,  9392,  6391,  6063,  2878,  1318,\n         3169,  8198,  9354,  4131,  3620,  3082,  3352,  9052,  8018,\n         7527,  3419,  1907,  8835,   796,   244,  8957,  4325,  8171,\n         9454,  7602,  4435,  7648,  3169,  2083,  9454,  4789,  9620,\n         9261,   556,  3524,  8497,  9174,  8299,  5871,  9052,  2888,\n         9846,  1610,  1362,  4930,  2150,  1362,  8018,  3867,   341,\n         7694,  8497,  6063,  3620,   244,  5807,  6089,  3169,  6350,\n         1174,  7694,   949,  1292,   244,  9052,  9440,  3690,  1362,\n         1907,  9011,  4156,  6081,   145,  1174,  7694,  9986,   949,\n         1292,  3169,  1455,  6372,  9760,  5013,  3169,  1455,  5942,\n         4365,  1362,  1907,   244,  5813,   244,  7994,  3525,  3550,\n         7509,  6372,  9760,  7860,  9052,  2888,  7694,  8497,  1610,\n         1316,   326,  1174,  3039,  3524,  9703,  3620,  6612,  1455,\n          556,  9011,  3169,  1927,  9052,   409,  4059,  9354,   700,\n         5503,  3550,  9052,  2083,  1963,   595,  3169,  7715, 10236,\n         9442,  1174, 10087,  3169,  5312,  7474,  9052,  3525,  3169,\n         5826,  7885,  6944,  7130,  5821,  2878,  7184,   153,  3169,\n         8633,  8574,  1283,   606,  7902,  6110,  3082,  6406,  3169,\n         8316,  6126,   688, 10236,  9440,  3082, 10584,  2143,  5460,\n         5809,  1362,  2878, 10439,  3419,  1907,  4598,  4156, 10239,\n         1450,  5514,  5010,  9350,   244,   651]])}, 0)\n<\/code><\/pre>\n\n<p>So the dictionary value is a 2-d array when it should only be 1-d.  Where am I going wrong?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53088667","CreationDate":"2018-10-31T17:04:20.283","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Not taking scalars seems to be a limitation of <code>tf.string_split<\/code>. Please do file an issue at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues<\/a><\/p>\n\n<p>In terms of workarounds, the suggestion to wrap in a list is a good one, but you'll also need to squeeze it after the split so you have a vector of components rather than a two-dimensional tensor.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\nscalar = tf.constant('ab c de')\nprint(scalar.shape)  # () scalar\nvector = scalar[None]\nprint(vector.shape)  # (1,) vector\noutput = tf.sparse.to_dense(tf.string_split(vector), default_value='')\nprint(output)  # tf.Tensor([[b'ab' b'c' b'de']], shape=(1, 3), dtype=string)\nsqueezed = tf.squeeze(output, axis=0)\nprint(squeezed.shape)  # (3,) vector\nprint(squeezed)  # tf.Tensor([b'ab' b'c' b'de'], shape=(3,), dtype=string)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53002190,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T05:44:01.143","AcceptedAnswerId":53007892.0,"OwnerUserId":679688.0,"Title":"How to get deterministic value of global step variable (strictly after or strictly before increment)","Body":"<p>Consider the following code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nglobal_step = tf.train.create_global_step()\nx = tf.Variable(100.0)\noptimizer = tf.train.AdamOptimizer()\ntrain_op = optimizer.minimize(x, global_step=global_step)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    step, _ = sess.run([global_step, train_op])\n    print(step)\n<\/code><\/pre>\n\n<p>The output I got was '1', but I think nothing stops Tensorflow from giving me '0' instead, i.e. the value of the global step variable before the 'assign' op that increments it in the <code>train_op<\/code>. I actually have another, more complicated, Tensorflow program that exhibits this behavior, where the global step value I get from <code>Session.run([global_step, train_op])<\/code> is off-by-one between two machines where I run it.<\/p>\n\n<p>For the global step variable, how can I definitively get its value from before the <code>train_op<\/code> or definitively get its value from after the <code>train_op<\/code>?<\/p>\n\n<p>I know I can do a <code>sess.run([global_step])<\/code> separately before or after <code>sess.run([train_op])<\/code> but I would like to do as much as possible within a single <code>session.run()<\/code> if doesn't involve overly complicating the code. I know I can get the the pre-increment value by assigning <code>global_step<\/code> to another variable and making a control dependency between the <code>tf.assign<\/code> and <code>train_op<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nglobal_step = tf.train.create_global_step()\nglobal_step2 = tf.get_variable('step-mirror', dtype=global_step.dtype, \nshape=global_step.shape)\nglobal_step2 = tf.assign(global_step2, global_step)\nx = tf.Variable(100.0)\noptimizer = tf.train.AdamOptimizer()\nwith tf.control_dependencies([global_step2]):\n    train_op = optimizer.minimize(x, global_step=global_step)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    step, _ = sess.run([global_step2, train_op])\n    print(step)\n<\/code><\/pre>\n\n<p>But I'm looking for a simpler way, perhaps a Tensorflow feature that I'm missing for specifying evaluating a variable from before any assignment on it.<\/p>\n\n<p>EDIT: In response to this <a href=\"https:\/\/stackoverflow.com\/questions\/53002190\/how-to-get-deterministic-value-of-global-step-variable-strictly-after-or-strict#comment93215051_53007892\">comment<\/a>, this does not work, and it prints '1' instead of '0':<\/p>\n\n<pre><code>import tensorflow as tf\n\nglobal_step = tf.train.create_global_step()\nx = tf.Variable(100.0)\noptimizer = tf.train.AdamOptimizer()\nwith tf.control_dependencies([global_step]):\n    train_op = optimizer.minimize(x, global_step=global_step)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    step, _ = sess.run([global_step, train_op])\n    print(step)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53007892","CreationDate":"2018-10-26T11:36:29.883","ParentId":null,"OwnerUserId":"4255553","Title":null,"Body":"<p>You can use this to read the global step after the train_op:<\/p>\n\n<pre><code>import tensorflow as tf\n\nglobal_step = tf.train.create_global_step()\nx = tf.Variable(100.0)\noptimizer = tf.train.AdamOptimizer()\ntrain_op = optimizer.minimize(x, global_step=global_step)\n\nwith tf.control_dependencies([train_op]):\n    global_step_value = global_step.read_value()\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    step, _ = sess.run([global_step_value, train_op])\n    print(step)\n<\/code><\/pre>\n\n<p>Here <code>global_step_value<\/code> is not a variable anymore. It is a tensor with the value of <code>global_step<\/code> after <code>train_op<\/code> has been calculated. This is described <a href=\"https:\/\/www.tensorflow.org\/guide\/variables\" rel=\"nofollow noreferrer\">here<\/a> under \"Using variables\".<\/p>\n"}]}
+{"QuestionId":53002518,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-26T06:14:22.587","AcceptedAnswerId":null,"OwnerUserId":7089239.0,"Title":"Poor exit code when managing multiple sessions","Body":"<p>I'm having an weird issue when training models using <code>tf.Graph<\/code> and <code>tf.Session<\/code>. And the implementation is somewhat odd, so bear with me. I'd like to explain the application structure.<\/p>\n\n<hr>\n\n<p>The issue has been finally (and somewhat embarassingly) <a href=\"https:\/\/stackoverflow.com\/questions\/53317523\/part-3-switching-between-multiple-contexts-no-error-and-a-bad-exit-code\">resolved<\/a> by updating all packages.<\/p>\n\n<hr>\n\n<h3>Application<\/h3>\n\n<p>The application is a service for handling multiple neural networks: training them and making predictions on them. For this reason a single graph wasn't quite enough. So when creating a new model, I firstly initialise both a <code>Graph<\/code> and a <code>Session<\/code> like so:<\/p>\n\n<pre><code>def __init__(self):\n    self.graph = tf.Graph()\n    with self.graph.as_default():\n        self.session = tf.Session()\n<\/code><\/pre>\n\n<p>These are then used both in the training process and when loading a model from disk.<\/p>\n\n<pre><code>def fit(self, x_train, y_train, n=200, batch=256):\n    with self.graph.as_default():\n        with self.session.as_default():\n            self.model.fit(x_train, y_train, epochs=n, batch_size=batch, verbose=0)\n<\/code><\/pre>\n\n<p>This is where the problem occurs (I've managed to comment everything out one by one, and the fit method is where it's at), but for further context, here is the (stripped down) creation method as well. It uses Keras.<\/p>\n\n<pre><code>def create(self):\n    with self.graph.as_default():\n        with self.session.as_default():\n            self.model = Sequential()\n            self.model.add(Dense(64, input_dim=shape[0], activation='relu',\n                                 kernel_regularizer=reg.l1_l2(0.1, 0.2)))\n            self.model.add(Dropout(0.5))\n            self.model.add(Dense(1, activation='sigmoid', kernel_regularizer=reg.l1_l2(0.1, 0.2)))\n            self.model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n<\/code><\/pre>\n\n<h3>Problem<\/h3>\n\n<p>When initialising the network and fitting it with data, the process exits with a bad code: <code>0xC0000005<\/code>. This doesn't give much information on the problem itself, and the bad exit code is given just then, on exit. <em>Even a <code>print<\/code> statement is executed after the routines successfully.<\/em> This has led me to suspect it is not a problem with the implementation, but something else.<\/p>\n\n<h3>Environment<\/h3>\n\n<p>I'm using Python 3.6.5 on PyCharm, but the problem has occured even when executing from a command line. As I said, multiple models are juggled around, but one training is enough to crash.<\/p>\n\n<hr>\n\n<p>What could possibly be at fault here? I realise it's not such a reproducable problem, but any pointers towards even debugging would be greatly appreciated.<\/p>\n\n<hr>\n\n<h3>Adventures<\/h3>\n\n<p>I tried modifying the <code>fit<\/code> function according to <a href=\"https:\/\/stackoverflow.com\/a\/51448207\/7089239\">this answer<\/a>, but with no luck. Here's the modified version:<\/p>\n\n<pre><code>from keras import backend as K\nimport gc\n\ndef fit(self, x_train, y_train, n=20, batch=256):\n    K.set_session(self.session)\n    with self.graph.as_default():\n        with self.session.as_default():\n            self.model.fit(x_train, y_train, epochs=n*10, batch_size=batch, verbose=0)\n    K.clear_session()\n    gc.collect()\n<\/code><\/pre>\n\n<p>Next I tried to create a new session for each computation (<code>tf.Session(graph=self.graph)<\/code>). It worked when using the <code>cg.collect()<\/code>, but after training the model, I could not make predictions with a new session. <code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value dense_1\/bias<\/code><\/p>\n\n<h3>Update<\/h3>\n\n<p>Currently (Nov 8th) I'm releasing all possible resources before creating and when loading a model. This has had the effect that I can create the model once, but the second time around (I do two training passes to evaluate the model independently) the program crashes like before. Let's try a new question, this is getting out of hand. <a href=\"https:\/\/stackoverflow.com\/questions\/53206900\/sound-way-of-managing-multiple-sessions-and-graphs\">Q v.2<\/a><\/p>\n","answers":[]}
+{"QuestionId":53002534,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T06:16:03.800","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Covert Pandas DF to Tensorflow tf.record file where 1st Pandas column has a single int, 2nd has multiple ints","Body":"<p>I have a Pandas DF where 1 column contains a single int, and the other column contains anywhere from 2 to 50 ints. <\/p>\n\n<p>Here is an example below<\/p>\n\n<pre><code>           EmbedID                          MappedC\n1911    3096611        [610580, 1396024, 1383000, 2480745, 751823, 97...\n1912    3096612        [365607, 917990]\n1913    3096613      [1067171, 638200, 2192752, 1609109, 1984544, 3...\n1914    3096614       [521163, 217279, 347655]\n1915    3096615      [1139429, 1254616, 3034840, 2312074, 68243]\n<\/code><\/pre>\n\n<p>The numbers <code>EmbedID<\/code>serves as the label, and two random numbers chosen from the <code>MappedC<\/code> column serves as the corresponding input numbers .<\/p>\n\n<p>What's the best way to convert this into a tf.record file?<\/p>\n\n<p>I see guides for converting a single numpy column to a tf.record file, such as these<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/swyoon\/8185b3dcf08ec728fb22b99016dd533f\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/swyoon\/8185b3dcf08ec728fb22b99016dd533f<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/45427637\/numpy-to-tfrecords-is-there-a-more-simple-way-to-handle-batch-inputs-from-tfrec\">Numpy to TFrecords: Is there a more simple way to handle batch inputs from tfrecords?<\/a><\/p>\n\n<p><a href=\"http:\/\/www.machinelearninguru.com\/deep_learning\/tensorflow\/basics\/tfrecord\/tfrecord.html\" rel=\"nofollow noreferrer\">http:\/\/www.machinelearninguru.com\/deep_learning\/tensorflow\/basics\/tfrecord\/tfrecord.html<\/a><\/p>\n\n<p>However, they all have trouble when the column \/ array has a varying number of ints. <\/p>\n\n<p>Edit:<\/p>\n\n<p>If this changes anything, here is more details about what exactly I am doing with the data. <\/p>\n\n<p>For training on Tensorflow, the single int column contains an index for a vector in an embedding matrix. That vector will be used as the label. <\/p>\n\n<p>The column with multiple ints have the 'input data'. For each label from the column containing a single int, 2 numbers will be chosen at random from the column containing multiple ints. <\/p>\n\n<p>I am basically doing a word2vec Cbow type of training<\/p>\n","answers":[{"AnswerId":"53028568","CreationDate":"2018-10-28T05:13:58.307","ParentId":null,"OwnerUserId":"4323","Title":null,"Body":"<p>First, shuffle your MappedC values:<\/p>\n\n<pre><code>import random\ndf.MappedC.apply(random.shuffle)\n<\/code><\/pre>\n\n<p>Then take the first and second values:<\/p>\n\n<pre><code>df.MappedC.str[0]\ndf.MappedC.str[1]\n<\/code><\/pre>\n\n<p><code>df.MappedC.str<\/code> looks like something about strings, which may be confusing, but <code>Series.str<\/code> works for lists as well as for strings, so this lets us choose the first and second element of each list, and efficiently construct new Series from those.<\/p>\n\n<p>You can now use the usual methods to put the data into TensorFlow, as you now have two plain Series of integers.<\/p>\n\n<p>Alternatively, this will give you a Series of randomly chosen pairs:<\/p>\n\n<pre><code>df.MappedC.map(lambda row: random.sample(row, 2))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53003155,"AnswerCount":1,"Tags":"<python><tensorflow><floating-point><quantization>","CreationDate":"2018-10-26T06:59:34.017","AcceptedAnswerId":null,"OwnerUserId":9104473.0,"Title":"How to find float output range for quantized matmul\/conv2D operation","Body":"<p>I am new to tensorflow and quantization, am trying to implement quantized matmul operation for two int8 inputs. Was curious to know the math behind the operation. I see in tensorflow they have implemented the same only for uint8 inputs , would like to know how to use that for signed int8 matmul\/conv2D.<\/p>\n\n<p>More precisely I would like to know how to get the float output range for the matmul\/conv2D operation.<\/p>\n\n<p>Any help would be highly appreciated.<\/p>\n","answers":[{"AnswerId":"53060650","CreationDate":"2018-10-30T09:01:47.123","ParentId":null,"OwnerUserId":"10546584","Title":null,"Body":"<p>I have investigated the quantization in tensorflow a bit and applied it to convert float operations into quant oeprations. <\/p>\n\n<p>In my case I have still a float input to the net. The input gets quantized right before entering the quant operations. Tensorflow prefers keeping float values as long as possible in order to be compatible to float operations. \nThis is also the reason, why tensorflow keeps the min and max float ranges after the float input gets quantized into 8bit integer format. \nThe min and max float values as a result from quantization are also inputs to the quant operations.<\/p>\n\n<p>In your case, the Quant_conv2d operation does a convolution with inputs: <\/p>\n\n<ul>\n<li>unsigned 8bit data form qunatization<\/li>\n<li>unsigned 8bit quantized kernel values <\/li>\n<\/ul>\n\n<p>The outputs are:<\/p>\n\n<ul>\n<li>result as 32 bit <\/li>\n<li>the new min and max range as float values<\/li>\n<\/ul>\n\n<p>The new float ranges are calculated from the ranges of the kernel values and the ranges of the input using the QuantizationRangeForMultiplication function stated in:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/quantization_utils.h\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/quantization_utils.h<\/a><\/p>\n\n<p>As stated, the output is a 32 bit with min and max float values to map to the absolute values and possbily convert the 8bit quantized format back to float.<\/p>\n\n<p>Hope this helps to understand Tensorflow quantization algorithms.<\/p>\n"}]}
+{"QuestionId":53003208,"AnswerCount":2,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-26T07:02:30.427","AcceptedAnswerId":53056640.0,"OwnerUserId":3259896.0,"Title":"How to save a Tensorflow Checkpoint file from Google Colaboratory in when using TPU mode?","Body":"<p>When I use <code>saver = tf.train.Saver()<\/code> and <code>save_path = saver.save(session, \"checkpointsFolder\/checkpoint.ckpt\")<\/code><\/p>\n\n<p>I get a <code>UnimplementedError (see above for traceback): File system scheme '[local]' not implemented<\/code> error<\/p>\n\n<p>Here is the full error<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nUnimplementedError                        Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)\n   1333     try:\n-&gt; 1334       return fn(*args)\n   1335     except errors.OpError as e:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)\n   1318       return self._call_tf_sessionrun(\n-&gt; 1319           options, feed_dict, fetch_list, target_list, run_metadata)\n   1320 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)\n   1406         self._session, options, feed_dict, fetch_list, target_list,\n-&gt; 1407         run_metadata)\n   1408 \n\nUnimplementedError: File system scheme '[local]' not implemented (file: 'checkpointsBook2Vec5Inputs')\n     [[{{node save\/SaveV2}} = SaveV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:tpu_worker\/replica:0\/task:0\/device:CPU:0\"](_recv_save\/Const_0, save\/SaveV2\/tensor_names, save\/SaveV2\/shape_and_slices, embeddings, embeddings\/Shampoo, embeddings\/Shampoo_1, embeddings\/Shampoo_2, epochCount, softmax_biases, softmax_weights, softmax_weights\/Shampoo, softmax_weights\/Shampoo_1, softmax_weights\/Shampoo_2)]]\n\nDuring handling of the above exception, another exception occurred:\n\nUnimplementedError                        Traceback (most recent call last)\n&lt;ipython-input-22-ca87cd5e5739&gt; in &lt;module&gt;()\n     48             print('recEpoch_indexA is', recEpoch_indexA)\n     49 \n---&gt; 50             save_path = saver.save(session, \"checkpointsBook2Vec5Inputs\/Research2VecCS4.ckpt\") #Save checkpoint\n     51             print( 'epochCount.eval() is ', epochCount.eval() )\n     52 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py in save(self, sess, save_path, global_step, latest_filename, meta_graph_suffix, write_meta_graph, write_state, strip_default_attrs)\n   1439           model_checkpoint_path = sess.run(\n   1440               self.saver_def.save_tensor_name,\n-&gt; 1441               {self.saver_def.filename_tensor_name: checkpoint_file})\n   1442 \n   1443         model_checkpoint_path = compat.as_str(model_checkpoint_path)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in run(self, fetches, feed_dict, options, run_metadata)\n    927     try:\n    928       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 929                          run_metadata_ptr)\n    930       if run_metadata:\n    931         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1150     if final_fetches or final_targets or (handle and feed_dict_tensor):\n   1151       results = self._do_run(handle, final_targets, final_fetches,\n-&gt; 1152                              feed_dict_tensor, options, run_metadata)\n   1153     else:\n   1154       results = []\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)\n   1326     if handle is None:\n   1327       return self._do_call(_run_fn, feeds, fetches, targets, options,\n-&gt; 1328                            run_metadata)\n   1329     else:\n   1330       return self._do_call(_prun_fn, handle, feeds, fetches)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)\n   1346           pass\n   1347       message = error_interpolation.interpolate(message, self._graph)\n-&gt; 1348       raise type(e)(node_def, op, message)\n   1349 \n   1350   def _extend_graph(self):\n\nUnimplementedError: File system scheme '[local]' not implemented (file: 'checkpointsBook2Vec5Inputs')\n     [[node save\/SaveV2 (defined at &lt;ipython-input-15-c14caac2081d&gt;:45)  = SaveV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:tpu_worker\/replica:0\/task:0\/device:CPU:0\"](_recv_save\/Const_0, save\/SaveV2\/tensor_names, save\/SaveV2\/shape_and_slices, embeddings, embeddings\/Shampoo, embeddings\/Shampoo_1, embeddings\/Shampoo_2, epochCount, softmax_biases, softmax_weights, softmax_weights\/Shampoo, softmax_weights\/Shampoo_1, softmax_weights\/Shampoo_2)]]\n\nCaused by op 'save\/SaveV2', defined at:\n  File \"\/usr\/lib\/python3.6\/runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/usr\/lib\/python3.6\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel_launcher.py\", line 16, in &lt;module&gt;\n    app.launch_new_instance()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/traitlets\/config\/application.py\", line 658, in launch_instance\n    app.start()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/kernelapp.py\", line 477, in start\n    ioloop.IOLoop.instance().start()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/ioloop.py\", line 177, in start\n    super(ZMQIOLoop, self).start()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tornado\/ioloop.py\", line 888, in start\n    handler_func(fd_obj, events)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tornado\/stack_context.py\", line 277, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 440, in _handle_events\n    self._handle_recv()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 472, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 414, in _run_callback\n    callback(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tornado\/stack_context.py\", line 277, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/kernelbase.py\", line 283, in dispatcher\n    return self.dispatch_shell(stream, msg)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/kernelbase.py\", line 235, in dispatch_shell\n    handler(stream, idents, msg)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/kernelbase.py\", line 399, in execute_request\n    user_expressions, allow_stdin)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/ipkernel.py\", line 196, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel\/zmqshell.py\", line 533, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2718, in run_cell\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2822, in run_ast_nodes\n    if self.run_code(code, result):\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/IPython\/core\/interactiveshell.py\", line 2882, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-15-c14caac2081d&gt;\", line 45, in &lt;module&gt;\n    saver = tf.train.Saver()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1102, in __init__\n    self.build()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1114, in build\n    self._build(self._filename, build_save=True, build_restore=True)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1151, in _build\n    build_save=build_save, build_restore=build_restore)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 792, in _build_internal\n    save_tensor = self._AddSaveOps(filename_tensor, saveables)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 284, in _AddSaveOps\n    save = self.save_op(filename_tensor, saveables)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 202, in save_op\n    tensors)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gen_io_ops.py\", line 1690, in save_v2\n    shape_and_slices=shape_and_slices, tensors=tensors, name=name)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3274, in create_op\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 1770, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nUnimplementedError (see above for traceback): File system scheme '[local]' not implemented (file: 'checkpointsBook2Vec5Inputs')\n     [[node save\/SaveV2 (defined at &lt;ipython-input-15-c14caac2081d&gt;:45)  = SaveV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:tpu_worker\/replica:0\/task:0\/device:CPU:0\"](_recv_save\/Const_0, save\/SaveV2\/tensor_names, save\/SaveV2\/shape_and_slices, embeddings, embeddings\/Shampoo, embeddings\/Shampoo_1, embeddings\/Shampoo_2, epochCount, softmax_biases, softmax_weights, softmax_weights\/Shampoo, softmax_weights\/Shampoo_1, softmax_weights\/Shampoo_2)]]\n<\/code><\/pre>\n\n<p>Looking up this error, I found the following:<\/p>\n\n<p>From Google official TPU debugging guide<\/p>\n\n<p><a href=\"https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting<\/a><\/p>\n\n<blockquote>\n  <p>Error Message<\/p>\n  \n  <p>InvalidArgumentError: Unimplemented: File system scheme '[local]' not\n  implemented<\/p>\n  \n  <p>Details<\/p>\n  \n  <p>All input files and the model directory must use a cloud storage\n  bucket path (gs:\/\/bucket-name\/...), and this bucket must be accessible\n  from the TPU server. Note that all data processing and model\n  checkpointing is performed on the TPU server, not the local machine.\n  For information on how to properly configure cloud storage for use\n  with the TPU, see the guide Connecting to Cloud Storage Buckets.<\/p>\n<\/blockquote>\n\n<p>Someone else with a similar issue<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/52703047\/tpu-local-filesystem-doesnt-exist\">TPU local Filesystem doesn&#39;t exist?<\/a><\/p>\n\n<blockquote>\n  <p>The local filesystem is not available on Cloud TPU's. Model\n  directories (checkpoints etc) and input data should be stored in\n  Google Cloud Storage (and prefixed with \"gs:\/\/\").<\/p>\n  \n  <p>More details here<\/p>\n  \n  <p><a href=\"https:\/\/cloud.google.com\/tpu\/docs\/storage-buckets\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/tpu\/docs\/storage-buckets<\/a><\/p>\n<\/blockquote>\n\n<p>However, I do not have a Google Cloud service, I am just using Google Colab. Is there a way to save a Tensorflow checkpoint when in TPU mode? <\/p>\n","answers":[{"AnswerId":"53056640","CreationDate":"2018-10-30T02:42:29.087","ParentId":null,"OwnerUserId":"3870291","Title":null,"Body":"<p>Another way to do this is to rewrite the model using Keras and use tf.contrib.tpu.keras_to_tpu_model(..) with tf.contrib.tpu.TPUDistributionStrategy(...). Here is small code snippet for this: <\/p>\n\n<pre><code>def get_model():\n  return keras.Sequential([\n    keras.layers.Dense(10, input_shape=(4,), activation=tf.nn.relu, name = \"Dense_1\"),\n    keras.layers.Dense(10, activation=tf.nn.relu, name = \"Dense_2\"),\n    keras.layers.Dense(3, activation=None, name = \"logits\"),\n    keras.layers.Dense(3, activation=tf.nn.softmax, name = \"softmax\")\n  ])\n\ndnn_model = get_model()\n\ndnn_model.compile(optimizer=tf.train.AdagradOptimizer(learning_rate=0.1), \n              loss='sparse_categorical_crossentropy',\n              metrics=['sparse_categorical_crossentropy'])\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    dnn_model,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_ADDRESS)))\n\n# Train the model\ntpu_model.fit(\n  train_x, train_y,\n  steps_per_epoch = steps_per_epoch,\n  epochs=epochs,\n)\n\ntpu_model.save_weights('.\/saved_weights.h5', overwrite=True)\n<\/code><\/pre>\n"},{"AnswerId":"53015637","CreationDate":"2018-10-26T20:06:36.160","ParentId":null,"OwnerUserId":"10199571","Title":null,"Body":"<p>You can create a Google Cloud account under the <a href=\"https:\/\/cloud.google.com\/free\/\" rel=\"nofollow noreferrer\">free tier<\/a> and then create a <a href=\"https:\/\/cloud.google.com\/storage\/\" rel=\"nofollow noreferrer\">GCS bucket<\/a>. After doing that you can authenticate yourself in Colab to get write access to your GCS bucket from Colab by doing the following:<\/p>\n\n<pre><code>from google.colab import auth\nauth.authenticate_user()\n<\/code><\/pre>\n\n<p>Here is a <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/shakespeare_with_tpuestimator.ipynb\" rel=\"nofollow noreferrer\">sample Colab notebook<\/a> that uses Cloud TPUs and GCS.<\/p>\n"}]}
+{"QuestionId":53003231,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T07:04:15.230","AcceptedAnswerId":null,"OwnerUserId":5327659.0,"Title":"How to get the dimension of tensors at runtime?","Body":"<p>I can get the dimensions of tensors at graph construction time via manually printing shapes of tensors(tf.shape()) but how to get the shape of these tensors at session runtime? <br>\nThe reason that I want shape of tensors at runtime is because at graph construction time shape of some tensors is coming as (?,8) and I cannot deduce the first dimension then. <\/p>\n","answers":[{"AnswerId":"53004082","CreationDate":"2018-10-26T07:58:32.637","ParentId":null,"OwnerUserId":"10356594","Title":null,"Body":"<p>The reason why some tensors are in shape of (?, ?) in tensorflow is that they are  placeholders. It can change during operation depends on your input data.<\/p>\n\n<p>So you must feed data into the placeholder,so it can tell what is the exact shape of your tensor.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.placeholder(tf.float32, shape=(None, None))\nprint(x.shape)  # ( ?,?)\nwith tf.Session() as sess:\n\n  rand_array = np.random.rand(3, 3)\n  after_sess_x = sess.run(x,feed_dict={x: rand_array})\n  print(after_sess_x.shape) # ( 3,3)\n<\/code><\/pre>\n"},{"AnswerId":"53003804","CreationDate":"2018-10-26T07:42:03.307","ParentId":null,"OwnerUserId":"10494794","Title":null,"Body":"<p>You have to make the tensors an output of the graph. For example, if <code>showme_tensor<\/code> is the tensor you want to print, just run the graph like that : <\/p>\n\n<p><code>_showme_tensor = sess.run(showme_tensor)<\/code> <\/p>\n\n<p>and then you can just print the output as you print a list. If you have different tensors to print, you can just add them like that :<\/p>\n\n<p><code>_showme_tensor_1, _showme_tensor_2 = sess.run([showme_tensor_1, showme_tensor_2])<\/code> <\/p>\n"}]}
+{"QuestionId":53003869,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T07:46:27.523","AcceptedAnswerId":null,"OwnerUserId":10561247.0,"Title":"How can I make MLP with weights from tensorflow","Body":"<p>I build Neural network in tensorflow and trained it.\nI extracted weights and bias from estimator <\/p>\n\n<pre><code>weights1 = self.model.get_variable_value('dense\/kernel')\nbias1 = self.model.get_variable_value('dense\/bias')\nweights2 = self.model.get_variable_value('dense_1\/kernel')\nbias1 = self.model.get_variable_value('dense_1\/bias')\n...\n<\/code><\/pre>\n\n<p>And I build MLP with numpy in python <\/p>\n\n<pre><code>layer1 = np.dot(inputs, weight1)\nlayer1 = np.add(layer1, bias1)\nlayer1 = np.maximum(layer1, layer1 * 0.2, layer1)\n...\n<\/code><\/pre>\n\n<p>I used leaky_relu activation function, so I was implemented it too but the output is completely different with from tensorflow. I don't know what's wrong about it.<\/p>\n\n<p>Edit)<\/p>\n\n<pre><code>def my_dnn_regression_fn(features, labels, mode, params):\n    top = tf.feature_column.input_layer(features, params[\"feature_columns\"])\n\n    for units in params.get(\"hidden_units\", [20]):\n        top = tf.layers.dense(inputs=top, units=units, activation=tf.nn.leaky_relu)\n\n    output_layer = tf.layers.dense(inputs=top, units=1)\n\n    predictions = tf.squeeze(output_layer, 1)\n\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        return tf.estimator.EstimatorSpec(\n            mode=mode, predictions={\"label\": predictions})\n\n    average_loss = tf.losses.mean_squared_error(labels, predictions)\n\n    batch_size = tf.shape(labels)[0]\n    total_loss = tf.to_float(batch_size) * average_loss\n\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        mse = tf.metrics.mean_squared_error(labels, predictions)\n        rmse = tf.metrics.root_mean_squared_error(labels, predictions)\n        absolute_error = tf.metrics.mean_absolute_error(labels, predictions)\n        mre = tf.metrics.mean_relative_error(labels, predictions, labels)\n\n        tf.summary.scalar('mse', mse[1])\n        tf.summary.scalar('mre', mre[1])\n        tf.summary.scalar('rmse', rmse[1])\n        tf.summary.scalar('absolute', absolute_error[1])\n\n        # vars = tf.trainable_variables()\n        # lossL2 = tf.add_n([tf.nn.l2_loss(v) for v in vars]) * 0.001\n\n        l1_regularizer = tf.contrib.layers.l1_regularizer(\n            scale=0.001, scope=None\n        )\n        weights = tf.trainable_variables()  # all vars of your graph\n        lossL1 = tf.contrib.layers.apply_regularization(l1_regularizer, weights)\n\n        # average_loss = tf.add(average_loss, lossL2)\n        average_loss = tf.add(average_loss, lossL1)\n\n        total_loss = tf.to_float(batch_size) * average_loss\n\n        optimizer = params.get(\"optimizer\", tf.train.AdamOptimizer)\n        optimizer = optimizer(params.get(\"learning_rate\", None))\n        train_op = optimizer.minimize(\n            loss=average_loss, global_step=tf.train.get_global_step())\n        # eval_metrics = {\"rmse\": rmse, \"absolute\": absolute_error, \"mre\": mre}\n        eval_metrics = {\"mse\": mse, \"rmse\": rmse, \"absolute\": absolute_error, \"mre\": mre}\n        return tf.estimator.EstimatorSpec(\n            mode=mode, loss=total_loss, train_op=train_op, eval_metric_ops=eval_metrics)\n\n    assert mode == tf.estimator.ModeKeys.EVAL\n\n    mse = tf.metrics.mean_squared_error(labels, predictions)\n    rmse = tf.metrics.root_mean_squared_error(labels, predictions)\n    absolute_error = tf.metrics.mean_absolute_error(labels, predictions)\n    mre = tf.metrics.mean_relative_error(labels, predictions, labels)\n\n    eval_metrics = {\"mse\": mse, \"rmse\": rmse, \"absolute\": absolute_error, \"mre\": mre}\n\n    return tf.estimator.EstimatorSpec(\n        mode=mode,\n        loss=total_loss,\n        eval_metric_ops=eval_metrics)\n<\/code><\/pre>\n\n<p>My dnn regrassion code!! <\/p>\n","answers":[]}
+{"QuestionId":53004212,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-10-26T08:06:22.357","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow v10 reformulate sigmoid cross entropy loss to work given sigmoid?","Body":"<p>Reading the TF guide for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/a> one passes the <code>logits<\/code> and <code>labels<\/code> to the function, after which the <code>sigmoid<\/code> operation is applied to <code>logits<\/code>. To avoid overflow in <code>exp(-x)<\/code> when <code>x &lt; 0<\/code> they reformulate the function to:<\/p>\n\n<pre><code>max(x, 0) - x * z + log(1 + exp(-abs(x)))\n<\/code><\/pre>\n\n<p>This is all well and good, however it may be the case where you want the output of your network to have the sigmoid transformation and use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/a> without having the loss apply the sigmoid again. How would one reformulate this for that while avoiding overflow for <code>exp(-x)<\/code> with <code>x &lt; 0<\/code>?<\/p>\n","answers":[{"AnswerId":"53577325","CreationDate":"2018-12-02T04:11:07.520","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<blockquote>\n  <p>it may be the case where you want the output of your network to have\n  the sigmoid transformation and use\n  tf.nn.sigmoid_cross_entropy_with_logits without having the loss apply\n  the sigmoid again.<\/p>\n<\/blockquote>\n\n<p>If I am not wrong you are trying to give the logits a value processed by a sigmoid activation. A processed value bounded by (0, 1). I don't think the x would ever be &lt; 0. It seems not reasonable put a value between 0 and 1 to the logits since the sigmoid of that value is always between 0.5 and sigmoid(1), which means for your negative labels the model tries to train them as 0 but the minimun value from the model is 0.5. The loss always exists(as log(1+exp(0.5))) for the negative samples and can never be minimized. <\/p>\n"}]}
+{"QuestionId":53005487,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-26T09:19:58.473","AcceptedAnswerId":null,"OwnerUserId":863543.0,"Title":"Eager-Mode very slow (~22x slower than Graph-Mode)","Body":"<p>I read that Tensorflow 2.0 will have some major changes, and a big part will be eager-execution [1], so I tried to play a bit around with the eager-mode of tensorflow.<\/p>\n\n<p>I took a code from a github-repo and tried to run it in eager-mode (however, without usage of Keras-Model\/Layers as proposed).\nIt turned out, that its quite slow. So I tried different modifications and compared it with the original source (graph-mode) of the model. The result is, that <strong>the graph-mode is around 22x times faster<\/strong> than the eager-mode. Its total clear to me, that the graph mode is faster, but by this number?<\/p>\n\n<p>Is this always the case or do I need some special modifications \/ configurations of the variables to get a comparable performance to graph mode?<\/p>\n\n<p>The source code, for both variants, can be found at [2].<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>Eager-Mode:<\/p>\n\n<pre><code># With \n#  with tf.device(\"\/gpu:0\"):\n#    ...\n#\n# Runtime is 0.35395\n# Runtime is 0.12711\n# Runtime is 0.12438\n# Runtime is 0.12428\n# Runtime is 0.12572\n# Runtime is 0.12593\n# Runtime is 0.12505\n# Runtime is 0.12527\n# Runtime is 0.12418\n# Runtime is 0.12340\n<\/code><\/pre>\n\n<p>Graph Mode:<\/p>\n\n<pre><code># Runtime is 0.81241\n# Runtime is 0.00573\n# Runtime is 0.00573\n# Runtime is 0.00570\n# Runtime is 0.00555\n# Runtime is 0.00564\n# Runtime is 0.00545\n# Runtime is 0.00540\n# Runtime is 0.00591\n# Runtime is 0.00574\n<\/code><\/pre>\n\n<p>[1] <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/developers\/JHDpgRyFVUs\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!topic\/developers\/JHDpgRyFVUs<\/a><\/p>\n\n<p>[2] <a href=\"https:\/\/gist.github.com\/lhlmgr\/f6709e5aba4a5314b5221d58232b09bd\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/lhlmgr\/f6709e5aba4a5314b5221d58232b09bd<\/a><\/p>\n","answers":[{"AnswerId":"53163713","CreationDate":"2018-11-05T23:18:49.743","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>Using eager execution may mean undoing some habits developed with TensorFlow graphs since code snippets that used to run once (e.g., Python function that constructs the graph to compute the loss) will run repeatedly (the same Python function will now compute the loss on each iteration).<\/p>\n\n<p>I took a cursory look at code links provided and noticed some easy wins that would probably also be seen by using standard Python profiling tools. You may want use those (cProfile, pyspy etc.)<\/p>\n\n<p>For example, the Keras network is currently implemented as:<\/p>\n\n<pre><code>class NFModel(tf.keras.Model):\n  def __init__(self, *args, **kwargs):\n    super().__init__(*args, **kwargs)\n\n  def call(self, *args, **kwargs):\n    num_layers = 6\n    d, r = 2, 2\n    bijectors = []\n\n    for i in range(num_layers):\n      with tf.variable_scope('bijector_%d' % i):\n        V = tf.get_variable('V', [d, r], dtype=DTYPE)  # factor loading\n        shift = tf.get_variable('shift', [d], dtype=DTYPE)  # affine shift\n        L = tf.get_variable('L', [d * (d + 1) \/ 2], dtype=DTYPE)  # lower triangular\n        bijectors.append(tfb.Affine(\n          scale_tril=tfd.fill_triangular(L),\n          scale_perturb_factor=V,\n          shift=shift,\n        ))\n\n        alpha = tf.get_variable('alpha', [], dtype=DTYPE)\n        abs_alpha = tf.abs(alpha) + .01\n        bijectors.append(LeakyReLU(alpha=abs_alpha))\n\n    base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], DTYPE))\n    mlp_bijector = tfb.Chain(list(reversed(bijectors[:-1])), name='2d_mlp_bijector')\n    dist = tfd.TransformedDistribution(distribution=base_dist, bijector=mlp_bijector)\n<\/code><\/pre>\n\n<p>Instead, if you create the variables in <code>__init__<\/code> once and avoid <code>tf.get_variable<\/code> calls on every call to the network, you should see a big improvement.<\/p>\n\n<pre><code>class NFModel(tf.keras.Model):\n  def __init__(self, *args, **kwargs):\n    super(NFModel, self).__init__(*args, **kwargs)\n    num_layers = 6\n    d, r = 2, 2\n    self.num_layers = num_layers\n    self.V = [tf.get_variable('V', [d, r], dtype=DTYPE)  for _ in range(num_layers)]\n    self.shift = [tf.get_variable('shift', [d], dtype=DTYPE)   for _ in range(num_layers)]\n    self.L = [tf.get_variable('L', [d * (d + 1) \/ 2], dtype=DTYPE)  for _ in range(num_layers)]\n    self.alpha = [tf.get_variable('alpha', [], dtype=DTYPE) for _ in range(num_layers)]\n\n\n  def call(self, *args, **kwargs):\n    bijectors = []\n\n    for i in range(self.num_layers):\n      V = self.V[i]\n      shift = self.shift[i]\n      L = self.L[i]\n      bijectors.append(tfb.Affine(\n        scale_tril=tfd.fill_triangular(L),\n        scale_perturb_factor=V,\n        shift=shift,\n      ))\n\n      alpha = self.alpha[i]\n      abs_alpha = tf.abs(alpha) + .01\n      bijectors.append(LeakyReLU(alpha=abs_alpha))\n\n    base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], DTYPE))\n    mlp_bijector = tfb.Chain(list(reversed(bijectors[:-1])), name='2d_mlp_bijector')\n    dist = tfd.TransformedDistribution(distribution=base_dist, bijector=mlp_bijector)\n\n    return {\"dist\": dist}\n<\/code><\/pre>\n\n<p>There are probably other such easy wins, a profiling tool will nudge you in the right direction.<\/p>\n\n<p>Also, note that, TF 2.0 is less about \"eager execution\" and more about how one interacts with graphs, as per the <a href=\"https:\/\/github.com\/tensorflow\/community\/pull\/20\" rel=\"nofollow noreferrer\">RFC<\/a><\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
+{"QuestionId":53005812,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-26T09:38:41.390","AcceptedAnswerId":53007714.0,"OwnerUserId":622097.0,"Title":"How do I early stop my model training in Keras?","Body":"<p>I want to stop my model training in Keras as soon as the accuracy 'acc' reaches a certain value, in my case 0.9. I wanted to apply the EarlyStopping callback but did not succeed yet. I thought that the following would suffice:<\/p>\n\n<pre><code>    early_stopping = EarlyStopping(\n            monitor = 'acc',\n            verbose = 2,\n            mode = 'max',\n            baseline = 0.9\n    )\n<\/code><\/pre>\n\n<p>Training stops after one epoch at an accuracy of 0.34 without any notice (thought verbose = 2). I tried to add a min_delta = 0.1 in the expectation that this rule would only be applied after the baseline was reached, but that did not help. The manual says that:<\/p>\n\n<blockquote>\n  <p>baseline: Baseline value for the monitored quantity to reach. Training will stop if the model doesn't show improvement over the baseline.<\/p>\n<\/blockquote>\n\n<p>and <\/p>\n\n<blockquote>\n  <p>mode: one of {auto, min, max}. ... in max mode it will stop when the quantity monitored has stopped increasing...<\/p>\n<\/blockquote>\n\n<p>From previous runs I know my model is increasing to 1.00 in about 25 epochs. Why does it stop so early without any notice?<\/p>\n","answers":[{"AnswerId":"53007714","CreationDate":"2018-10-26T11:26:25.457","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>To me your line works like expected. Since you did not provide a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Minimal_working_example\" rel=\"nofollow noreferrer\">minimal workable example<\/a> it is probably the easiest, if you debug the issue yourself. It is actually quite easy.<\/p>\n\n<p>Add this line right before your call to <code>model.fit<\/code> and run your code:<\/p>\n\n<pre><code>import ipdb; ipdb.set_trace()\n<\/code><\/pre>\n\n<p>You will now get the ipdb-prompt. Type this:<\/p>\n\n<pre><code>ipdb&gt; b EarlyStopping.on_epoch_end\nipdb&gt; c\n<\/code><\/pre>\n\n<p>Now your program should run until it reaches the breakpoint. Now you can step through the code using <code>n<\/code> to advance a line and you can use <code>p<\/code> print variables. Here is an example to check, what is the current best accuracy and what is the new value for the current epoch:<\/p>\n\n<pre><code>    543         if self.monitor_op(current - self.min_delta, self.best):\n--&gt; 544             self.best = current\n    545             self.wait = 0\n\nipdb&gt; p self.best\n0.9\nipdb&gt; p current\n0.9196166666348775\n<\/code><\/pre>\n\n<p>This way you can check, which path through the code is taking and where it deviates from what you expect. If that is not sufficient to solve your problem the additional information will definitely be helpful for others do help you solve your problem.<\/p>\n\n<p><strong>Update<\/strong>:\nYou are correct. It does not actually work as expected. It just happen to stop at the right time in my test I did. But using the original class as a template it's easy to create your own class so that you get this behaviour:<\/p>\n\n<pre><code>class CustomEarlyStopping(EarlyStopping):\n    def __init__(self, threshold):\n        self.threshold = threshold\n        super(CustomEarlyStopping, self).__init__(monitor='acc')\n\n    def on_epoch_end(self, epoch, logs=None):\n        current = self.get_monitor_value(logs)\n        if current is None:\n            return\n\n        if current &gt;= self.threshold:\n            self.stopped_epoch = epoch\n            self.model.stop_training = True\n<\/code><\/pre>\n"}]}
+{"QuestionId":53005909,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-10-26T09:43:48.443","AcceptedAnswerId":null,"OwnerUserId":10557479.0,"Title":"CNN model on cifar10 dataset don't seems to be trained","Body":"<p>I just wanted to take a look for a simple tutorial on cnn and I want to train my network composed with 5 filters(3 conv layers and 2 dense layers)\nfollowed each by max_pool layers and dropout layers.<\/p>\n\n<p>My problem is that my network don't seems to be trained at all, and I cant find where is the problem....<\/p>\n\n<p>I'll put all my codes (train.py test.py model.py and main.py) in the following link on github.<\/p>\n\n<p><a href=\"https:\/\/github.com\/mmecoco\/tensorflow_cnn_cifar10\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mmecoco\/tensorflow_cnn_cifar10<\/a><\/p>\n\n<p>Help me...<\/p>\n","answers":[{"AnswerId":"53007024","CreationDate":"2018-10-26T10:47:03.240","ParentId":null,"OwnerUserId":"10557479","Title":null,"Body":"<p>I just found my problem, this was caused from my last fc layer, the problem was  that i tryed to apply a softmax activation on my output, and if I use a softmax cross entropy loss function, the value of my loss will be on the side of -inf or inf.\nBecause of this, the networks stops the learning process and this is why I failed my training.<\/p>\n"}]}
+{"QuestionId":53006188,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-26T10:00:31.130","AcceptedAnswerId":null,"OwnerUserId":7227887.0,"Title":"Tensorflow taking input in the same order as input","Body":"<p>I am using tensorflow to test my trained model on test images. I am feeding the images to tensorflow as below:<\/p>\n\n<pre><code>image_ab, image_aba = sess.run(fetches, feed_dict={self.image_a: image_a,\n                                               self.is_train: False})\n<\/code><\/pre>\n\n<p>I printed the <code>image_a<\/code> and <code>image_ab<\/code> and observed that <code>image_a<\/code> is not in the same order as the input images i give.<\/p>\n\n<p>For some reasons i want the output also to be in the same order as input images.  <\/p>\n\n<p>Does tensorflow usually takes input in the same order as the input given?<\/p>\n","answers":[{"AnswerId":"53006337","CreationDate":"2018-10-26T10:08:36.090","ParentId":null,"OwnerUserId":"497600","Title":null,"Body":"<p>I assume you mean <code>image_ab<\/code> is not in the same order. Because <code>image_a<\/code> is the input that you feed to tensorflow. If this input is not ordered correctly, it will be your preprocessing, not tensorflow.<\/p>\n\n<p>Tensorflow usually works on batches of data. For images, the convention for batch dimensions is:<\/p>\n\n<pre><code>[batch, x, y, colors]\n<\/code><\/pre>\n\n<p>The operations that tensorflow performs are parallelized along the batch. If you simply plug convolutional layers together, the order of the batch should be preserved.<\/p>\n\n<p>However, it is surely possible to reorder things in tensorflow:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nx = tf.placeholder(shape=(2,1), dtype=\"float32\")\ny = tf.concat([x[1], x[0]], axis=0)\n\nsess = tf.Session()\nsess.run([x,y], feed_dict={x:np.random.rand(2,1)})\n<\/code><\/pre>\n\n<p>This code will read in x, change the order of its entries and produce y.\nSo tensorflow can reorder your images. You could search your code for a pattern like the one in my example.<\/p>\n"}]}
+{"QuestionId":53006787,"AnswerCount":2,"Tags":"<python><machine-learning><keras><rnn><gated-recurrent-unit>","CreationDate":"2018-10-26T10:32:47.020","AcceptedAnswerId":null,"OwnerUserId":10560255.0,"Title":"RNN with GRU in Keras","Body":"<p>I want to implement Recurrent Neural network with GRU using Keras in python. I have problem in running code and I change variables more and more but it doesn't work. Do you have an idea for solve it?<\/p>\n\n<pre><code>inputs = 42          #number of columns input  \nnum_hidden =50      #number of neurons in the layer\noutputs = 1           #number of columns output  \nnum_epochs = 50\nbatch_size = 1000\nlearning_rate = 0.05\n#train       (125973, 42)  125973 Rows and 42 Features\n#Labels  (125973,1) is True Results\nmodel = tf.contrib.keras.models.Sequential()\nfv=tf.contrib.keras.layers.GRU\nmodel.add(fv(units=42, activation='tanh', input_shape= (1000,42),return_sequences=True))  #i want to send Batches to train\n\n\n#model.add(tf.keras.layers.Dropout(0.15))  # Dropout overfitting\n\n#model.add(fv((1,42),activation='tanh', return_sequences=True))\n#model.add(Dropout(0.2))  # Dropout overfitting\n\nmodel.add(fv(42, activation='tanh'))\nmodel.add(tf.keras.layers.Dropout(0.15))  # Dropout overfitting\n\nmodel.add(tf.keras.layers.Dense(1000,activation='softsign'))\n#model.add(tf.keras.layers.Activation(\"softsign\"))\n\n\nstart = time.time()\n# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\n# model.compile(loss=\"mse\", optimizer=sgd)\nmodel.compile(loss=\"mse\", optimizer=\"Adam\") \ninp = np.array(train)\noup = np.array(labels)\nX_tr = inp[:batch_size].reshape(-1, batch_size, inputs)\nmodel.fit(X_tr,labels,epochs=20, batch_size=batch_size)\n<\/code><\/pre>\n\n<p>However I get the following error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense to have shape (1000,) but got array with shape (1,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53010118","CreationDate":"2018-10-26T13:42:57.280","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you read the error carefully you would realize there is a shape mismatch between the shapes of labels you provide, which is <code>(None, 1)<\/code>, and the shape of output of model, which is <code>(None, 1)<\/code>:<\/p>\n\n<pre><code>ValueError: Error when checking target:  &lt;--- This means the output shapes\nexpected dense to have shape (1000,)     &lt;--- output shape of model  \nbut got array with shape (1,)            &lt;--- the shape of labels you give when training\n<\/code><\/pre>\n\n<p>Therefore you need to make them consistent. You just need to change the number of units in the last layer to 1 since there is one output per input sample:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(1, activation='softsign')) # 1 unit in the output\n<\/code><\/pre>\n"},{"AnswerId":"53006899","CreationDate":"2018-10-26T10:39:29.950","ParentId":null,"OwnerUserId":"3830485","Title":null,"Body":"<p>Here, you have mentioned input vector shape to be 1000.<\/p>\n\n<p><code>\nmodel.add(fv(units=42, activation='tanh', input_shape= (1000,42),return_sequences=True))  #i want to send Batches to train\n<\/code><\/p>\n\n<p>However, shape of your training data (X_tr) is 1-D\nCheck your X_tr variable and have same dimension for input layer.<\/p>\n"}]}
+{"QuestionId":53007095,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-26T10:50:50.387","AcceptedAnswerId":null,"OwnerUserId":1719901.0,"Title":"WRN Install python-prctl so that processes can be cleaned with guarantee","Body":"<p>I am getting this Warning \"WRN Install python-prctl so that processes can be cleaned with guarantee.\" though I have installed 'python-prctl' from <a href=\"https:\/\/pythonhosted.org\/python-prctl\/\" rel=\"nofollow noreferrer\">this<\/a> site. But still getting this warning, I am using tensorpack to load data quickly by using data_generator.<a href=\"https:\/\/i.stack.imgur.com\/TOGJq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TOGJq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am using python3, I have attached the image above of that warning.<\/p>\n\n<p>I don't know why I am getting this error?\nActually I am using tensorpack specially for this data generator cause it has some special features to load data quickly.<\/p>\n","answers":[]}
+{"QuestionId":53007455,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><matplotlib><keras>","CreationDate":"2018-10-26T11:10:23.210","AcceptedAnswerId":null,"OwnerUserId":8293726.0,"Title":"blank image is saved after not using imshow using matplotlib","Body":"<p>I am implementing GAN (generative adversarial network) in keras from <a href=\"https:\/\/github.com\/Zackory\/Keras-MNIST-GAN\/blob\/master\/mnist_gan.py\" rel=\"nofollow noreferrer\">here<\/a>. In this code after every 20 epoch image is showed and saved. But i donot want to show image and just want to save. so i just commented line number 89 from code, which is following:<\/p>\n\n<p><code>plt.imshow(generatedImages[i], interpolation='nearest', cmap='gray_r')<\/code>. but after commenting this line blank images saved, and also blank images poped up. but when i remove comment and run full code, images saved properly. accoeding to me <code>imshow<\/code> is used to used images, it doesn't part in constructing images. why this happens.<\/p>\n","answers":[]}
+{"QuestionId":53009000,"AnswerCount":1,"Tags":"<python><tensorflow><keras><vgg-net>","CreationDate":"2018-10-26T12:42:57.797","AcceptedAnswerId":null,"OwnerUserId":3261772.0,"Title":"VGG Transfer learning - predict generator shows different result than evaluate generator","Body":"<p>I built a CNN based on VGG pre-trained model.  I fine tuned the last conv block (#5) together with my added Fully connected layers.  My classification problem has 4 classes, my last activation layer is 'softmax' and I am using 'sparse_categorical_crossentropy' as my loss function.  When I am creating my train\/valid generator, I am using class_mode ('sparse').  So I fit my model and I get 92% accuracy on my validation dataset after 2 epochs.  The problem that I have, is when I check the predict_generator to see my predictions on my validation set the accuracy is around 73 True predictions\/ (73 + 161 wrong predictions) = 0.31%, however the model.evaluate_generator(validation_generator) has 91% accuracy.<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(\n        path_data_train,\n        target_size=(img_width, img_height),\n        batch_size=16,\n        class_mode=\"sparse\")\n\nvalidation_generator = test_datagen.flow_from_directory(\n        path_data_valid,\n        target_size=(img_width, img_height),\n        batch_size=16,\n        class_mode=\"sparse\")\n\nnb_train_samples = 1874\nnb_validation_samples = 234\nepochs = 10\nbatch_size = 16\n\nhistory = model.fit_generator(\n        train_generator,\n        steps_per_epoch=nb_train_samples \/\/ batch_size,\n        epochs=epochs,\n        validation_data=validation_generator,\n        validation_steps=nb_validation_samples \/\/ batch_size,\n        callbacks = [checkpointer])\n\nEpoch 1\/10\n117\/117 [==============================] - 42s 356ms\/step - loss: 0.0850 - acc: 0.9690 - val_loss: 0.4173 - val_acc: 0.9062\nEpoch 2\/10\n117\/117 [==============================] - 42s 360ms\/step - loss: 0.0690 - acc: 0.9765 - val_loss: 0.4423 - val_acc: 0.894\n\nprint (model.metrics_names)\nmodel.evaluate_generator(validation_generator)\n['loss', 'acc']\n[0.39189313988909763, 0.9059829049640231]\n\npreds = model.predict_generator(validation_generator)\n<\/code><\/pre>\n\n<p>Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"53009837","CreationDate":"2018-10-26T13:27:35.910","ParentId":null,"OwnerUserId":"3261772","Title":null,"Body":"<p>I post the answer if in case someone in future hit the same issue.  You need to set Shuffle = False in validation_generator.<\/p>\n\n<pre><code>validation_generator = test_datagen.flow_from_directory(\n        path_data_valid,\n        target_size=(img_width, img_height),\n        batch_size=16,\n        class_mode=\"sparse\", shuffle = False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53009708,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2018-10-26T13:20:40.260","AcceptedAnswerId":null,"OwnerUserId":5175180.0,"Title":"Getting StatusCode.FAILED_PRECONDITION from Tensorflow serving server: Serving signature key \"tensorflow\/serving\/predict\" not found","Body":"<p>Getting following error from my python grpc client when trying to connect to the tf-serving server:<\/p>\n\n<pre><code>&lt;_Rendezvous of RPC that terminated with:\nstatus = StatusCode.FAILED_PRECONDITION\ndetails = \"Serving signature key \"tensorflow\/serving\/predict\" not found.\"\ndebug_error_string = \"{\"created\":\"@1540559424.720048798\",\"description\":\"Error received from peer\",\"file\":\"src\/core\/lib\/surface\/call.cc\",\"file_line\":1099,\"grpc_message\":\"Serving signature key \"tensorflow\/serving\/predict\" not found.\",\"grpc_status\":9}\"\n&gt;\n<\/code><\/pre>\n\n<p>I believe this means when I build the servables the method name was wrong. But I see that in my servable model builder script I am using the correct name.<\/p>\n\n<pre><code>prediction_signature = tf.saved_model.signature_def_utils.build_signature_def(\n        inputs=input_signature,\n        outputs=output_signature,\n        method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)\n\n    builder = saved_model_builder.SavedModelBuilder(os.path.join(output_path, str(version_number)))\n<\/code><\/pre>\n\n<p>What can be the reason for this error at client side?<\/p>\n","answers":[]}
+{"QuestionId":53010183,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-26T13:46:45.387","AcceptedAnswerId":null,"OwnerUserId":583464.0,"Title":"classifier after binary classifier","Body":"<p>I have some images which contains some structure (building for example).<\/p>\n\n<p>I have the pixel data and the labels data.<\/p>\n\n<p>The labels data are in the form:<\/p>\n\n<pre><code>np.array([[0, nan, nan],\n          [1, 2, 1],\n          [1, 3, 2]])\n<\/code><\/pre>\n\n<blockquote>\n  <p>The first column means: is there or not a structure ? 1 or 0<\/p>\n  \n  <p>The second: what type is the structure ? types: 1,2,3<\/p>\n  \n  <p>The third: how many structures are there?<\/p>\n<\/blockquote>\n\n<p>When we don't have a structure, all other values are nan.<\/p>\n\n<p>So, for the example above we have:<\/p>\n\n<blockquote>\n  <p>first line: no structure<\/p>\n  \n  <p>second line: we have a structure of type 2 and it is one<\/p>\n  \n  <p>third line: we have a structure of type 3 and there are two of them<\/p>\n<\/blockquote>\n\n<p>So, first I am doing a binary classification to find out if we have a structure or not.<\/p>\n\n<p>I am using vggnet with pretrained weights.<\/p>\n\n<pre><code>imagenet_weights = '.\/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5'\nbase_vgg16 = VGG16(include_top=False, weights=imagenet_weights, input_shape=input_shape)\nlast_layer = base_vgg16.output\n\nx = Flatten()(last_layer)\nx = Dense(512, activation='relu')(x)\nx = Dropout(0.3)(x)\npreds = Dense(1, activation='sigmoid')(x)\n\nbase_vgg16.trainable = False\nmodel_vgg16 = Model(base_vgg16.input, preds)\n<\/code><\/pre>\n\n<p>and for training I am using 3kfold and data augmentation.<\/p>\n\n<p>This gives 97% accuracy.<\/p>\n\n<p>Now, I want to make a classification in order to find out the type, based on the binary classification results.<\/p>\n\n<p>So , I have :<\/p>\n\n<pre><code>pred_val = model.predict(X_val)\n<\/code><\/pre>\n\n<p><strong>Now, I am not sure how to proceed.<\/strong><\/p>\n\n<p>What to use as input for the classifier.<\/p>\n\n<p>I tried:<\/p>\n\n<pre><code>X_train = X_train[np.where(pred_val &gt;= 0.5), :]\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>pred_train = model.predict(X_train)\nX_train = X_train[np.where(pred_train.squeeze() &gt;= 0.5), :]\n<\/code><\/pre>\n\n<p>( I am not sure if it has any meaning to predict on train data)<\/p>\n\n<p>Now, however I run the second network, I am always receiving awful results.<\/p>\n\n<pre><code>Val_loss is almost steady around (10-11)\nTrain_loss is around 10-12\n\nVal_acc is almost steady around (0.25-0.35)\nTrain_acc is 0.28-0.38\n<\/code><\/pre>\n\n<p>For the second network:<\/p>\n\n<pre><code> inputs = Input(shape=input_shape)\n x = Flatten()(inputs)\n x = Dense(512, activation='relu')(x)\n x = Dense(256, activation='relu')(x)\n preds = Dense(5, activation='softmax')(x)      \n model = Model(inputs, preds)\n<\/code><\/pre>\n\n<p>Whatever combinations I tried, more or less units ( I tried only a few like : <code>x = Dense(8, activation='relu')(x)<\/code>)\n or whatever batch size and optimizer and learning rate the result is the same.<\/p>\n","answers":[]}
+{"QuestionId":53010465,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><lstm><pytorch>","CreationDate":"2018-10-26T14:01:42.090","AcceptedAnswerId":53041720.0,"OwnerUserId":1981484.0,"Title":"Bidirectional LSTM output question in PyTorch","Body":"<p>Hi I have a question about how to collect the correct result from a BI-LSTM module\u2019s output.<\/p>\n\n<p>Suppose I have a 10-length sequence feeding into a single-layer LSTM module with 100 hidden units:<\/p>\n\n<pre><code>lstm = nn.LSTM(5, 100, 1, bidirectional=True)\n<\/code><\/pre>\n\n<p><code>output<\/code> will be of shape:<\/p>\n\n<pre><code>[10 (seq_length), 1 (batch),  200 (num_directions * hidden_size)]\n# or according to the doc, can be viewed as\n[10 (seq_length), 1 (batch),  2 (num_directions), 100 (hidden_size)]\n<\/code><\/pre>\n\n<p>If I want to get the 3rd (1-index) input\u2019s output at both directions (two 100-dim vectors), how can I do it correctly?<\/p>\n\n<p>I know <code>output[2, 0]<\/code> will give me a 200-dim vector. <strong>Does this 200 dim vector represent the output of 3rd input at both directions?<\/strong><\/p>\n\n<p>A thing bothering me is that when do reverse feeding, the 3rd (1-index) output vector is calculated from the 8th(1-index) input, right?<\/p>\n\n<p>Will pytorch automatically take care of this and group output considering direction?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53041720","CreationDate":"2018-10-29T08:47:02.437","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>When using a BiLSTM the hidden states of the directions are just concatenated (the second part after the middle is the hidden state for feeding in the reversed sequence). <br>So splitting up in the middle works just fine. <\/p>\n\n<p>As reshaping works from the right to the left dimensions you won't have any problems in separating the two directions.<\/p>\n\n<hr>\n\n<p>Here is a small example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># so these are your original hidden states for each direction\n# in this case hidden size is 5, but this works for any size\ndirection_one_out = torch.tensor(range(5))\ndirection_two_out = torch.tensor(list(reversed(range(5))))\nprint('Direction one:')\nprint(direction_one_out)\nprint('Direction two:')\nprint(direction_two_out)\n\n# before outputting they will be concatinated \n# I'm adding here batch dimension and sequence length, in this case seq length is 1\nhidden = torch.cat((direction_one_out, direction_two_out), dim=0).view(1, 1, -1)\nprint('\\nYour hidden output:')\nprint(hidden, hidden.shape)\n\n# trivial case, reshaping for one hidden state\nhidden_reshaped = hidden.view(1, 1, 2, -1)\nprint('\\nReshaped:')\nprint(hidden_reshaped, hidden_reshaped.shape)\n\n# This works as well for abitrary sequence lengths as you can see here\n# I've set sequence length here to 5, but this will work for any other value as well\nprint('\\nThis also works for more multiple hidden states in a tensor:')\nmulti_hidden = hidden.expand(5, 1, 10)\nprint(multi_hidden, multi_hidden.shape)\nprint('Directions can be split up just like this:')\nmulti_hidden = multi_hidden.view(5, 1, 2, 5)\nprint(multi_hidden, multi_hidden.shape)\n<\/code><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Direction one:\ntensor([0, 1, 2, 3, 4])\nDirection two:\ntensor([4, 3, 2, 1, 0])\n\nYour hidden output:\ntensor([[[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]]]) torch.Size([1, 1, 10])\n\nReshaped:\ntensor([[[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]]]) torch.Size([1, 1, 2, 5])\n\nThis also works for more multiple hidden states in a tensor:\ntensor([[[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]],\n\n        [[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]],\n\n        [[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]],\n\n        [[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]],\n\n        [[0, 1, 2, 3, 4, 4, 3, 2, 1, 0]]]) torch.Size([5, 1, 10])\nDirections can be split up just like this:\ntensor([[[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]],\n\n\n        [[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]],\n\n\n        [[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]],\n\n\n        [[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]],\n\n\n        [[[0, 1, 2, 3, 4],\n          [4, 3, 2, 1, 0]]]]) torch.Size([5, 1, 2, 5])\n<\/code><\/pre>\n\n<p><em>Hope this helps! :)<\/em><\/p>\n"},{"AnswerId":"54938505","CreationDate":"2019-03-01T05:34:32.453","ParentId":null,"OwnerUserId":"4410235","Title":null,"Body":"<blockquote>\n  <p>I know output[2, 0] will give me a 200-dim vector. Does this 200 dim vector represent the output of 3rd input at both directions?<\/p>\n<\/blockquote>\n\n<p>The answer is <strong>YES<\/strong>. <\/p>\n\n<p>The <code>output<\/code> tensor of LSTM module output is the concatenation of forward LSTM output and backward LSTM output at corresponding postion in input sequence.\nAnd <code>h_n<\/code> tensor is the output at last timestamp which is output of the lsat token in forward LSTM but the first token in backward LSTM.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>In [1]: import torch\n   ...: lstm = torch.nn.LSTM(input_size=5, hidden_size=3, bidirectional=True)\n   ...: seq_len, batch, input_size, num_directions = 3, 1, 5, 2\n   ...: in_data = torch.randint(10, (seq_len, batch, input_size))\n   ...: output, (h_n, c_n) = lstm(in_data)\n   ...: \n\nIn [2]: # output of shape (seq_len, batch, num_directions * hidden_size)\n   ...: \n   ...: print(output)\n   ...: \ntensor([[[ 0.0379,  0.0169,  0.2539,  0.2547,  0.0456, -0.1274]],\n\n        [[ 0.7753,  0.0862, -0.0001,  0.3897,  0.0688, -0.0002]],\n\n        [[ 0.7120,  0.2965, -0.3405,  0.0946,  0.0360, -0.0519]]],\n       grad_fn=&lt;CatBackward&gt;)\n\nIn [3]: # h_n of shape (num_layers * num_directions, batch, hidden_size)\n   ...: \n   ...: print(h_n)\n   ...: \ntensor([[[ 0.7120,  0.2965, -0.3405]],\n\n        [[ 0.2547,  0.0456, -0.1274]]], grad_fn=&lt;ViewBackward&gt;)\n\nIn [4]: output = output.view(seq_len, batch, num_directions, lstm.hidden_size)\n   ...: print(output[-1, 0, 0])  # forward LSTM output of last token\n   ...: print(output[0, 0, 1])  # backward LSTM output of first token\n   ...: \ntensor([ 0.7120,  0.2965, -0.3405], grad_fn=&lt;SelectBackward&gt;)\ntensor([ 0.2547,  0.0456, -0.1274], grad_fn=&lt;SelectBackward&gt;)\n\nIn [5]: h_n = h_n.view(lstm.num_layers, num_directions, batch, lstm.hidden_size)\n   ...: print(h_n[0, 0, 0])  # h_n of forward LSTM\n   ...: print(h_n[0, 1, 0])  # h_n of backward LSTM\n   ...: \ntensor([ 0.7120,  0.2965, -0.3405], grad_fn=&lt;SelectBackward&gt;)\ntensor([ 0.2547,  0.0456, -0.1274], grad_fn=&lt;SelectBackward&gt;)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53010789,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-10-26T14:21:49.487","AcceptedAnswerId":null,"OwnerUserId":9606168.0,"Title":"Training in batches but performing single image classification using tf.estimators in Tensorflow","Body":"<p>Building a convolutional neural network for image classification has been my first foray into both Python and Tensorflow. I have a large database of images to train on, and my application is rapid, real-time classification of <strong>single<\/strong> images. Currently, in order to perform inference of single images, I have to make the training batch size equal to 1. If I don't (and the training batch size is 16) I get the following error:<\/p>\n\n<pre><code>ValueError: Cannot feed value of shape (1, 132, 128, 1) for Tensor \n'tf_reshape1:0', which has shape '(16, 132, 128, 1)'\n<\/code><\/pre>\n\n<p>I'd really like to have the flexibility to train on larger batch sizes, whilst still being able to classify single images.<\/p>\n\n<p>My question is very similar to other asked on here e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/46430351\/training-in-batches-but-testing-individual-data-item-in-tensorflow\">Training in batches but testing individual data item in Tensorflow?<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/46193626\/tensorflow-layer-size-dependent-on-batch-size\">Tensorflow: Layer size dependent on batch size?<\/a> but my lack of experience with tf means that I'm unable to implement the suggested solutions into my code.<\/p>\n\n<p>I've written my training and inference codes for image classification by basing them around the example given on the tensorflow website for classification of the MNIST database <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn#train_eval_mnist\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn#train_eval_mnist<\/a>. My codes use tf.estimators <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/estimators<\/a> a high level tensorflow API. The solutions to the similar question asked and answered above suggest modifying the tf.placeholders which I am not (knowingly) using. I've copied my code for the input function and the model function that I use below. I'm sure I'll need to post more but this is also the first SO question I've asked so apologies if I've forgotten many things. <\/p>\n\n<p>Training:<\/p>\n\n<pre><code>bead_classifier = tf.estimator.Estimator(\n                 model_fn=bead_model_fn, \n\nmodel_dir=r'\/tmp\/trained_model')\n\n# Train the model\ntrain_input_fn = tf.estimator.inputs.numpy_input_fn(\n  x={\"x\": training_imgs},\n  y=training_labels,\n  batch_size=16,\n  num_epochs=None,\n  shuffle=True)\nbead_classifier.train(\n  input_fn=train_input_fn,\n  steps = 13000)\n  hooks=[logging_hook])\n<\/code><\/pre>\n\n<p>Model function (first few layers only):<\/p>\n\n<pre><code> def bead_model_fn(features, labels, mode):\n  \"\"\"Model function for CNN.\"\"\"\n  # Input Layer\n  # Reshape X to 4-D tensor: [batch_size, width, height, channels]\n\ninput_layer = tf.reshape(features[\"x\"], [-1, 132, 128, 1], name = \n'tf_reshape1')\n\n  # Convolutional Layer #1\n\n  conv1 = tf.layers.conv2d(\n      inputs=input_layer,\n      filters=4,\n      kernel_size=[2, 2],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n  # Pooling Layer #1\n  pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)\n\n  # Convolutional Layer #2\n  conv2 = tf.layers.conv2d(\n      inputs=pool1,\n      filters=8,\n      kernel_size=[2, 2],\n      padding=\"same\",\n      activation=tf.nn.relu)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53010965,"AnswerCount":1,"Tags":"<python><keras><lstm><stateless><stateful>","CreationDate":"2018-10-26T14:31:49.043","AcceptedAnswerId":53011186.0,"OwnerUserId":10341061.0,"Title":"keras lstm difference stateless and stateful","Body":"<p>I am trying to understand the concept of LSTM layers in Keras.\nI just want to confirm some behavior of LSTM and check if i understand it correctly.<\/p>\n\n<p>Assuming that I have 1000 samples and this samples have 1 time step and i have a batch size of 1 when<\/p>\n\n<pre><code>stateful = True\n<\/code><\/pre>\n\n<p>Is this the same as 1 sample with 1000 time steps and a batch size of 1 with <\/p>\n\n<pre><code>stateful = False\n<\/code><\/pre>\n\n<p>Here I am also assuming that in both cases i have the same information just in different shapes and i reset the state of my LSTM layer after every training epoch.<\/p>\n\n<p>I also think that the batch size in the stateless case only matters for my training sequence, because if i set<\/p>\n\n<pre><code>stateful = False \n<\/code><\/pre>\n\n<p>i can use input_shape instead of batch_input_shape. So my LSTM layer does not need a batch dimension only time steps and feature dimensions. is this correct?<\/p>\n\n<p>i got this conclusions from:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/recurrent.py#L1847\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/recurrent.py#L1847<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/43882796\/when-does-keras-reset-an-lstm-state\/46331227#46331227\">When does keras reset an LSTM state?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/38714959\/understanding-keras-lstms\">Understanding Keras LSTMs<\/a><\/p>\n\n<p>And if i have a multi layer LSTM net if the first LSTM layer is stateful, all other layers should also be stateful right?<\/p>\n\n<p>I hope somebody understands what i mean and can help me. If my questions are not understandable please tell me and i will update this post.<\/p>\n\n<p>Thanks everybody.<\/p>\n","answers":[{"AnswerId":"53011186","CreationDate":"2018-10-26T14:44:54.997","ParentId":null,"OwnerUserId":"10494794","Title":null,"Body":"<p><code>stateful=True<\/code> means that you keep the final state for every batch and pass it as initial state for the next batch. \nSo yes, in this case it's the same if you have 1 batch of 1000 samples or 1000 batches of 1 sample.<\/p>\n"}]}
+{"QuestionId":53010986,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-26T14:32:46.643","AcceptedAnswerId":null,"OwnerUserId":10362405.0,"Title":"Changing batch_size parameter in keras leads to broadcast error","Body":"<p>I am running a simple encoder-decoder setup to train a representation for a one dimensional image. In this sample the input are lines with varying slopes and in the encoded layer we would expect something that resembles the slope. My setup is keras with a tensorflow backend. I am very new to this as well.\nIt all works fine, at least until I move away from steps_per_epoch to batch_size in the model.fit() method. Certain values of the batch_size, such as 1,2,3, 8 and 16 do work, for others I get a value error. My initial guess was 2^n, but that did not work.\nThe error I get for batch_size = 5 \n<code>ValueError: operands could not be broadcast together with shapes (5,50) (3,50) (5,50)<\/code><\/p>\n\n<p>I am trying to understand which relation between batch_size and training data is valid such that it always passes. I assumed that the training set would be simply divided into floor(N\/batch_size) batches and the remainder would be processed as such.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>What is the relation between size of data set and batch_size that are allowed.\n\n<ol start=\"2\">\n<li>What exactly is the keras\/tensorflow trying to do such that the batch_size is important? <\/li>\n<\/ol><\/li>\n<\/ol>\n\n<p>Thank you very much for the help.<\/p>\n\n<p>The code to reproduce this is<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Model\nfrom keras.layers import Input, Dense, Conv1D, Concatenate\nfrom keras.losses import mse\nfrom keras.optimizers import Adam\n\nINPUT_DIM = 50\nINTER_DIM = 15\nLATENT_DIM = 1\n\n# Prepare Sample Data\none_line = np.linspace(1, 30, INPUT_DIM).reshape(1, INPUT_DIM)\ntest_array = np.repeat(one_line, 1000, axis=0)\nslopes = np.linspace(0, 1, 1000).reshape(1000, 1)\ndata = test_array * slopes\n# Train test split\ntrain_mask = np.where(np.random.sample(1000) &lt; 0.8, 1, 0).astype('bool')\nx_train = data[train_mask].reshape(-1, INPUT_DIM, 1)\nx_test = data[~train_mask].reshape(-1, INPUT_DIM, 1)\n\n# Define Model\ninput = Input(shape=(INPUT_DIM, 1), name='input')\nconv_layer_small = Conv1D(filters=1, kernel_size=[3], padding='same')(input)\nconv_layer_medium = Conv1D(filters=1, kernel_size=[5], padding='same')(input)\nmerged_convs = Concatenate()(\n    [conv_layer_small, conv_layer_medium])\n\nlatent = Dense(LATENT_DIM, name='latent_layer',\n               activation='relu')(merged_convs)\nencoder = Model(input, latent)\n\ndecoder_int = Dense(INTER_DIM, name='dec_int_layer', activation='relu')(latent)\noutput = Dense(INPUT_DIM, name='output', activation='linear')(decoder_int)\n\nencoder_decoder = Model(input, output, name='encoder_decoder')\n\n\n# Add Loss\nreconstruction_loss = mse(input, output)\nencoder_decoder.add_loss(reconstruction_loss)\nencoder_decoder.compile(optimizer='adam')\n\n\nif __name__ == '__main__':\n    epochs = 100\n    encoder_decoder.fit(\n        x_train,\n        epochs=epochs,\n        batch_size=4,\n        verbose=2\n    )\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53010997,"AnswerCount":1,"Tags":"<object><tensorflow><identification>","CreationDate":"2018-10-26T14:33:34.077","AcceptedAnswerId":null,"OwnerUserId":7766898.0,"Title":"How can I improve a tensorflow model to recognize vehicle registration plate detection?","Body":"<p>I have struggled for a week to train a model that can give me the bounding numbers for extraction of vehicle registration plates. (It's a law enforcement initiative.) The latest attempt - there were many - went like this:<\/p>\n\n<ol>\n<li>I used the pets pre-trained model. <\/li>\n<li>I took images of roads and overlaid a different registration plate on each road image. The plate is always in the same position to simplify the generation of the xml files.<\/li>\n<li>Upsized the plates to 300x80 - previously much smaller with similar results.<\/li>\n<li>1600 training images and 200 testing images.<\/li>\n<li>Converted the xml to csv and then to tfrecord format.<\/li>\n<li>Trained the model using the ssd_mobilenet_v1_pets.config with minor changes, such as the num_classes and num_examples and paths.<\/li>\n<li>Trained on an AWS GPU instance for a few hours.<\/li>\n<\/ol>\n\n<p>Results are horrible. Windscreens are detected as plates and square fronts of vehicles are detected. I can upload if it will help.<\/p>\n\n<p>I'm about to give up and use a text extraction algorithm, but that brings other complexities such as ignoring other text in the image.<\/p>\n\n<p>I hate giving up, so any advice eagerly accepted.<\/p>\n\n<p>Thanks,\nAlan.<\/p>\n","answers":[{"AnswerId":"53011511","CreationDate":"2018-10-26T15:02:52.047","ParentId":null,"OwnerUserId":"7766898","Title":null,"Body":"<p>Ouch. Down-voted already. Apologies. Marked as answered. Guess I misunderstood how to formulate questions. Plan B it is.<\/p>\n"}]}
+{"QuestionId":53011315,"AnswerCount":1,"Tags":"<keras><google-cloud-storage><large-data><google-cloud-ml>","CreationDate":"2018-10-26T14:51:21.940","AcceptedAnswerId":null,"OwnerUserId":10563517.0,"Title":"How should I handle large video datasets in Google Cloud ML Engine?","Body":"<p>I am experimenting with video classification using Keras in Cloud ML Engine. My dataset consists in video sequences saved as separate images (eg. seq1_frame1.png, seq1.frame2.png...) which I have uploaded to a GCS bucket. <\/p>\n\n<p>I use a csv file referencing the start of end frames of different subclips, and a generator which feeds batch of clips to the model. The generator is responsible for loading frames from the bucket, reading them as images, and concatenating them as numpy arrays.<\/p>\n\n<p>My training is fairly long, and I suspect the generator is my bottleneck due to the numerous reading operations. <\/p>\n\n<p>In the exemples I found online, people usually save pre-formatted clips as tfrecords files directly to GCS. I feel like this solution isn't ideal for very large datasets as it implies duplicating the data, even more so if we decide to extract overlapping subclips.<\/p>\n\n<p>Is there something wrong in my approach ? And more importantly, is there a \"golden-standard\" for using large video datasets for machine learning ?<\/p>\n\n<p>PS : I explained my setup for reference, but my question is not bound to Keras, generators or Cloud ML. <\/p>\n","answers":[{"AnswerId":"53011669","CreationDate":"2018-10-26T15:12:41.603","ParentId":null,"OwnerUserId":"9161094","Title":null,"Body":"<p>In this, you are almost always going to be trading time for space. You just have to work out which is more important.<\/p>\n\n<p>In theory, for every frame, you have height*width*3 bytes. That's assuming 3 colour channels. One possible way you could save space is to use only one channel (probably choose green, or, better still, convert your complete dataset to greyscale). That would reduce your full size video data to one third size. Colour data in video tends to be at a lower resolution than luminance data so it might not affect your training, but it depends on your source files.<\/p>\n\n<p>As you probably know, .png is a lossless image compression. Every time you load one, the generator will have to decompress first, and then concatenate to the clip. You could save even more space by using a different compression codec, but that would mean every clip would need full decompression and probably add to your time. You're right, the repeated decompression will take time. And saving the video uncompressed will take up quite a lot of space. There are places you could save space, though:<\/p>\n\n<ul>\n<li>reduce to greyscale (or green scale as above)<\/li>\n<li>temporally subsample frames (do you need EVERY consecutive frame, or could you sample every second one?)<\/li>\n<li>do you use whole frames or just patches? Can you crop or rescale the video sequences?<\/li>\n<li>are you using optical flow? It's pretty processor intensive, consider it as a pre-processing step, too, so you only have to do it once per clip (again this is trading space for time)<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53012526,"AnswerCount":1,"Tags":"<python><matrix><neural-network><keras><xor>","CreationDate":"2018-10-26T16:05:22.740","AcceptedAnswerId":53037250.0,"OwnerUserId":3737798.0,"Title":"Keras Matrix Multiplication to get Predicted Values","Body":"<p>I'm looking to take the output of a Keras model to manually calculate the predicted values through matrix multiplication. I would like to do this to help understand how Keras is working under the hood. I'll use the simple XOR problem. Here is my code:<\/p>\n\n<pre><code>import numpy as np\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense\nfrom keras.callbacks import LambdaCallback\n\nclass LossHistory(keras.callbacks.Callback):\n    def on_train_begin(self, logs={}):\n        self.losses = []\n\n    def on_batch_end(self, batch, logs={}):\n        self.losses.append(logs.get('loss'))\n\n\nhistory = LossHistory()\n\n# the four different states of the XOR gate\ntraining_data = np.array([[0,0],[0,1],[1,0],[1,1]], \"float32\")\n\n# the four expected results in the same order\ntarget_data = np.array([[0],[1],[1],[0]], \"float32\")\n\nmodel = Sequential()\nmodel.add(Dense(4, input_dim=2, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n\nprint_weights = LambdaCallback(on_epoch_end=lambda batch, logs: print(model.layers[0].get_weights()))\n\nmodel.compile(loss='mean_squared_error',\n              optimizer='adam',\n              metrics=['binary_accuracy'])\n\nhistory2 = model.fit(training_data, target_data, epochs=50, verbose=2, callbacks=[print_weights, history])\n\nprint(model.predict(training_data).round())\n\n\nW1 = model.get_weights()[0]\nX1 = np.matrix([[0,0],[1,1]], \"float32\")\nwx = np.dot(X1,W1)\nb = model.get_weights()[1]\nwx = np.reshape(wx,(4,2))\nb = np.reshape(b, (4,1))\nz = wx + b\nfrom numpy import array, exp\na1 = 1 \/ (1 + exp(-z))\nprint('g =\\n', a1)\n\nW2 = model.get_weights()[2]\nb2 = model.get_weights()[3]\nW2 = np.reshape(W2,(1,4))\na1 = np.reshape(a1, (4,1))\nwa = np.dot(W2,a1)\nz2 = wa + b2\na2 = 1 \/ (1 + exp(-z2))\nprint('g =\\n', a2)\n<\/code><\/pre>\n\n<p>From what I understand, <code>get_weights()[0]<\/code> and <code>get_weights()[1]<\/code> are the weights and biases for the first layer, respectively, and <code>get_weights()[2]<\/code> and <code>get_weights()[3]<\/code> are the weights and biases for the second layer. I believe the issue I'm have is figuring out what x1 and x2 are as they relate to the equation z = Wx + b. The weights are retrieved from the last epoch and are usually weights that achieve 100% accuracy. The output I'm expecting is [0,1,1,0] for the y-hat predictions based on the manual calculation of z = Wx + b and then taking the sigmoid of z. <\/p>\n","answers":[{"AnswerId":"53037250","CreationDate":"2018-10-29T00:17:53.350","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>You are very close!<\/p>\n\n<p>First off, 50 epochs using a training set with only 4 events was not enough to replicate a consisting correct output (0,1,1,0), so I bumped the number of epochs to 1000. \nBelow is the code I used with the decimal and rounded outputs:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense\n\n# Set seed for reproducibility\nnp.random.seed(1)\n\n# the four different states of the XOR gate\ntraining_data = np.array([[0,0],[0,1],[1,0],[1,1]], \"float32\")\n# the four expected results in the same order\ntarget_data = np.array([[0],[1],[1],[0]], \"float32\")\n\nmodel = Sequential()\nmodel.add(Dense(4, input_dim=2, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='mean_squared_error',optimizer='adam',metrics=['binary_accuracy'])\n\nhistory = model.fit(training_data, target_data, epochs=1000, verbose=1)\n\n# decimal output\nprint('decimal output:\\n'+str(model.predict(training_data)))\n# rounded output\nprint('rounded output:\\n'+str(model.predict(training_data).round()))\n# ouputs:\ndecimal output:\n[[ 0.25588933]\n [ 0.82657152]\n [ 0.83840138]\n [ 0.16465074]]\nrounded output:\n[[ 0.]\n [ 1.]\n [ 1.]\n [ 0.]]\n<\/code><\/pre>\n\n<p>The model is giving the correct rounded output, good! The decimal output is nice to use for comparing the manual approach.<\/p>\n\n<p>For the manual approach, X1 is the input to the model, either [0,0], [0,1], [1,0] or [1,1]. X2 is the output of the first layer, and input to the last layer. The weight and biases are exactly as you said (\"get_weights()[0] and get_weights()[1] are the weights and biases for the first layer, respectively, and get_weights()[2] and get_weights()[3] are the weights and biases for the second layer\"). But it may seem like you forgot the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Rectifier_(neural_networks)\" rel=\"nofollow noreferrer\">relu activation function<\/a> in the first layer? Lets look at the solution code:<\/p>\n\n<pre><code># Parameters layer 1\nW1 = model.get_weights()[0]\nb1 = model.get_weights()[1]\n\n# Parameters layer 2\nW2 = model.get_weights()[2]\nb2 = model.get_weights()[3]\n\n# Input\nX1 = np.array([[0,0],[0,1],[1,0],[1,1]], \"float32\")\n# Use the following X1 for single input instead of all at once\n#X1 = np.array([[0,0]])\n\n# First layer calculation\nL1 = np.dot(X1,W1)+b1\n# Relu activation function\nX2 = np.maximum(L1,0)\n# Second layer calculation\nL2 = np.dot(X2,W2)+b2\n# Sigmoid\noutput = 1\/(1+np.exp(-L2))\n\n# decimal output\nprint('decimal output:\\n'+str(output))\n# rounded output\nprint('rounded output:\\n'+str(output.round()))\n# ouputs:\ndecimal output:\n[[ 0.25588933]\n [ 0.82657152]\n [ 0.83840144]\n [ 0.16465074]]\nrounded output:\n[[ 0.]\n [ 1.]\n [ 1.]\n [ 0.]]\n<\/code><\/pre>\n\n<p>You can use all 4 input simultaneously as above, or just a single input as the commented out #X1 suggest. Note that the decimal \"model.predict\" output, and the manual approach gives the exact same output (with a small deviation at the third value, probably due to some keras\/numpy rounding deviation?)<\/p>\n"}]}
+{"QuestionId":53012917,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-26T16:30:48.293","AcceptedAnswerId":null,"OwnerUserId":7232350.0,"Title":"Tensorflow Matrix Manipulation","Body":"<p>I have below matrix<\/p>\n\n<pre><code>x = [x1, x2, x3\n     y1, y2, y3\n     z1, z2, z3\n    ] \n<\/code><\/pre>\n\n<p>and another matrix<\/p>\n\n<pre><code>y = [False, False, True, True, True,  False, False, False\n     False, True,  True, True, False, False, False, False\n     False, False, False,False,True,  True,  True,  False\n    ]\n<\/code><\/pre>\n\n<p>I want to create below the new matrix in which x are filled in z the same way True is filled in y, and 0 are filled otherwise.<\/p>\n\n<pre><code>z = [0, 0,  x1, x2, x3, 0, 0, 0\n     0, x1, x2, x3, 0,  0, 0, 0\n     0, 0,  0,  0,  x1, x2,x3,0\n    ] \n<\/code><\/pre>\n\n<p>I am able to achieve this by<\/p>\n\n<pre><code>index = tf.cast(tf.where(y), tf.int32)\nz = tf.sparse_to_dense(index, tf.shape(y), tf.reshape(x, [-1]))\n<\/code><\/pre>\n\n<p>However using sparse_to_dense is complicated and ends up with other issues. Anyone can help with some other way to do this?<\/p>\n\n<p>thanks!<\/p>\n","answers":[{"AnswerId":"53032114","CreationDate":"2018-10-28T13:38:08.520","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>This produces the result using\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_nd_update\" rel=\"nofollow noreferrer\">tf.scatter_nd_update<\/a>.<\/p>\n\n<p><code>tf.tile<\/code> just replicates <code>100 200 300<\/code> 3 times so that 3 copies can be filled in the 3 locations.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = [100, 200, 300,\n     1, 2, 3,\n     4, 5, 6\n    ]\n\ny = [False, False, True, True, True,  False, False, False,\n     False, True,  True, True, False, False, False, False,\n     False, False, False,False,True,  True,  True,  False\n    ]\n\nx0 = tf.Variable( x )\n\ny0 = tf.Variable( tf.cast( y, tf.int32 ))\n\nwith tf.Session() as sess :\n    sess.run(tf.initialize_all_variables())\n\n    indices = tf.constant( sess.run( tf.where( tf.equal(y0 , True )) ))\n\n     print(sess.run( tf.scatter_nd_update( y0,\n                                           indices,\n                                           tf.tile(x0[0:3], tf.constant([3])))))\n<\/code><\/pre>\n\n<p>output is<\/p>\n\n<pre><code>[  0   0 100 200 300   0   0   0   0 100 200 300   0   0   0   0   0   0\n   0   0 100 200 300   0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53013146,"AnswerCount":1,"Tags":"<keras><lstm><multi-layer>","CreationDate":"2018-10-26T16:45:57.980","AcceptedAnswerId":53020897.0,"OwnerUserId":10341061.0,"Title":"multi layer LSTM net with stateful=True","Body":"<p>My question is does the this code make sense? And if this makes sense what should be the purpose?<\/p>\n\n<pre><code>model.add(LSTM(18, return_sequences=True,batch_input_shape=(batch_size,look_back,dim_x), stateful=True))\nmodel.add(Dropout(0.3))\nmodel.add(LSTM(50,return_sequences=False,stateful=False))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(1, activation='linear'))\n<\/code><\/pre>\n\n<p>Because if my first LSTM layer returns my state from one batch to the next, why shouldn't do my second LSTM layer the same?<\/p>\n\n<p>I'm having a hard time to understand the LSTM mechanics in Keras so I'm very thankful for any kind of help :)<\/p>\n\n<p>And if you down vote this post could you tell me why in the commands? thanks.<\/p>\n","answers":[{"AnswerId":"53020897","CreationDate":"2018-10-27T10:20:17.400","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>Your program is a regression problem where your model consists of 2 lstm layers with 18 and 50 layers each and finally a dense layer to show the regression value.<\/p>\n\n<p>LSTM requires a 3D input.Since the output of your first LSTM layer is going to the input for the second LSTM layer.The input of the Second LSTM layer should also be in 3D. so we set the retrun sequence as true in 1st  as it will return a 3D output which can then be used as an input for the second LSTM.<\/p>\n\n<p>Your second LSTMs value  does not return a sequence because after the second LSTM you have a dense layer which does not need a 3D value as input. <\/p>\n\n<p>[update]<\/p>\n\n<p>In keras by default LSTM states are reset after each batch of training data,so if you  don't want the states to be reset after each batch you can set the stateful=True. If LSTM is made stateful final state of a batch will be used as an initial state for the next batch.\nYou can later reset the states by calling <code>reset_states()<\/code><\/p>\n"}]}
+{"QuestionId":53013242,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-26T16:53:04.667","AcceptedAnswerId":53013837.0,"OwnerUserId":1316501.0,"Title":"Tensorflow: FailedPreconditionError: Table not initialized (using tf.data.Dataset API)","Body":"<p>I'm using the <code>tf.data.Dataset<\/code> API in conjunction with <code>tf.contrib.lookup.index_table_from_tensor<\/code>.  <\/p>\n\n<p>My dataset was created like this:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({'reviews': x}, y)))\n<\/code><\/pre>\n\n<p>Here's what I'm doing:<\/p>\n\n<pre><code>data_table = tf.contrib.lookup.index_table_from_tensor(tf.constant(data_vocab))\nlabels_table = tf.contrib.lookup.index_table_from_tensor(tf.constant(labels_vocab))\n<\/code><\/pre>\n\n<p>Then I map a preprocessing function across my <code>dataset<\/code>:<\/p>\n\n<pre><code>def preprocess(x, y):\n    # split on whitespace\n    x['reviews'] = tf.string_split([x['reviews']])\n    # turn into integers\n    return data_table.lookup(x['reviews']), labels_table.lookup(y)\n<\/code><\/pre>\n\n<p>All good so far.  However, when I try to pass my dataset to my Keras model for training, I get:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Table not initialized.\n<\/code><\/pre>\n\n<p>I googled around and people suggest that I need to include:<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.tables_initializer())\n<\/code><\/pre>\n\n<p>But now I get:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Table not initialized.\n     [[Node: hash_table_Lookup = LookupTableFindV2[Tin=DT_STRING, Tout=DT_INT64](hash_table_lookup_placeholder, StringSplit:1, hash_table_lookup_placeholder_1)]]\n     [[Node: IteratorGetNext_1 = IteratorGetNext[output_shapes=[[?,?], [?,20]], output_types=[DT_INT64, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Iterator_1)]]\n<\/code><\/pre>\n\n<p>Any idea why my lookup table is <em>still<\/em> not initialized \/ how to fix this?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53013837","CreationDate":"2018-10-26T17:42:31.363","ParentId":null,"OwnerUserId":"8593338","Title":null,"Body":"<p>Hi it's quiet weird maybe the following working example will help you:<\/p>\n\n<pre><code>x = ['this is aswesome', 'i dont like it', 'i love it', 'i hate it']\ny = ['positive','negative','positive','negative']\ndata_vocab = list({word for sentence in x for word in sentence.split(' ')})\nlabel_vocab = list(set(y))\n\ndataset = tf.data.Dataset.from_tensor_slices(({'reviews': x}, y))\n\ndata_table=tf.contrib.lookup.index_table_from_tensor(tf.constant(data_vocab))\nlabels_table = tf.contrib.lookup.index_table_from_tensor(tf.constant(label_vocab))\n\ndef preprocess(x, y):\n    # split on whitespace\n    x['reviews'] = tf.string_split([x['reviews']])\n    # turn into integers\n    return data_table.lookup(x['reviews']), labels_table.lookup(y)\n\npreprocessed = dataset.map(preprocess)\n\nit = preprocessed.make_initializable_iterator()\nsess = tf.Session()\nsess.run(it.initializer)\nsess.run(tf.tables_initializer()) \n<\/code><\/pre>\n\n<p>If you call <code>sess.run(it.get_next())<\/code>you get <code>(SparseTensorValue(indices=array([[0, 0],\n        [0, 1],\n        [0, 2]]), values=array([2, 7, 4]), dense_shape=array([1, 3])), 1)<\/code><\/p>\n\n<p>Hope this will help you !<\/p>\n"}]}
+{"QuestionId":53013555,"AnswerCount":1,"Tags":"<python-3.x><keras><conv-neural-network><convolution><keras-layer>","CreationDate":"2018-10-26T17:19:11.513","AcceptedAnswerId":null,"OwnerUserId":5330906.0,"Title":"ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv1d_1\/convolution\/Conv2D","Body":"<p>Binary classification problem: I want to have One input layer(optional), One Conv1D layer then output layer of 1 neuron predicting either 1 or 0.\n<br>Here is my model:<\/p>\n\n<pre><code>x_train = np.expand_dims(x_train,axis=1)\nx_valid = np.expand_dims(x_valid,axis=1)\n#x_train = x_train.reshape(x_train.shape[0], 1, x_train.shape[1])\n#x_valid = x_train.reshape(x_valid.shape[0], 1, x_train.shape[1])\n\nmodel = Sequential()\n\n#hidden layer\nmodel.add(Convolution1D(filters = 1, kernel_size = (3),input_shape=(1,x_train.shape[2])))\n#output layer\nmodel.add(Flatten())\nmodel.add(Dense(1, activation = 'softmax'))\n\nsgd = SGD(lr=0.01, nesterov=True, decay=1e-6, momentum=0.9)\n\nmodel.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n\nprint('model compiled successfully')\nmodel.fit(x_train, y_train, nb_epoch = nb_epochs, validation_data=(x_valid,y_valid), batch_size=100)\n<\/code><\/pre>\n\n<p><strong>Input shape<\/strong>: x_train.shape = (5,1,133906) which is (batch,steps,channels) respectively. Steps added through expand_dims. Actual size (5,133906) which is 5 samples of time series data of length 133906 sampled randomly sometimes at 2 ms and sometimes at 5 ms. <br> <br>\n<strong>Error Message<\/strong>: <strong>ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv1d_1\/convolution\/Conv2D' (op: 'Conv2D') with input shapes: [?,1,1,133906], [1,3,133906,1].<\/strong> <br> <br>\nHow do I resolve this issue? What should the size of x_train and the input_size argument passed inside Conv1D be?    <\/p>\n","answers":[{"AnswerId":"54495047","CreationDate":"2019-02-02T16:30:16.107","ParentId":null,"OwnerUserId":"5968693","Title":null,"Body":"<p>Convolution1D layers takes input in a format of [batch, steps, channels]<\/p>\n\n<p>Your length of convolution window (kernel size) cannot be larger than number of steps.<\/p>\n\n<p>Therefore if you want to use your defined input shape of:<\/p>\n\n<pre><code>x_train.shape = (5,1,133906)\n<\/code><\/pre>\n\n<p>you need to change kernel size to 1<\/p>\n\n<p>i.e. change line 9 to<\/p>\n\n<pre><code>model.add(Convolution1D(filters = 1, kernel_size = 1,input_shape=(1,x_train.shape[2])))\n<\/code><\/pre>\n\n<p>However, this will only enable your example to work. Depending on your goals, data type, etc. you might want to try different combinations of your kernel size and dimensions of input data to obtain best results.<\/p>\n"}]}
+{"QuestionId":53013956,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T17:52:22.423","AcceptedAnswerId":53024803.0,"OwnerUserId":4400904.0,"Title":"Tensorflow hang at compute_gradients","Body":"<p>I am having an issue where my Tensorflow execution is getting stuck at <code>compute_gradients<\/code>. I am initializing my model and then setting up the loss function like so. Notice that at this point I haven't began training so the issue is not my data.:<\/p>\n\n<pre><code># The model for training\ngiven_model = GivenModel(images_input=images_t)\n\nprint(\"Done setting up the model\")\n\nwith tf.device('\/gpu:0'):\n    with tf.variable_scope('prediction_loss'):\n        logits = given_model.prediction\n\n        softmax_loss_per_sample = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels))\n\n\n        total_training_loss = softmax_loss_per_sample\n\n        optimizer = tf.train.AdamOptimizer()\n        gradients, variables = zip(*optimizer.compute_gradients(total_training_loss))\n        gradients, _ = tf.clip_by_global_norm(gradients, gradient_clip_threshold)\n        optimize = optimizer.apply_gradients(zip(gradients, variables))\n\n\n    with tf.control_dependencies([optimize]):\n        train_op = tf.constant(0)\n<\/code><\/pre>\n\n<p>This code just hangs and doesn't do anything. When I ctrl+c out of it (no matter how long of running), it is always stuck at the <code>compute_gradients<\/code>.<\/p>\n\n<p>Does anyone have any idea why this may be happening? I am not doing this inside a loop and my model is not that big. It also seems to be using the CPU to do this (no memory allocated on GPU yet), and I can't force it to use the GPU despite the <code>with tf.device('\/gpu:0'):<\/code> option. <\/p>\n\n<p>Thanks<\/p>\n\n<p>Here is what prints when I do ctrl+c:<\/p>\n\n<pre><code>gradients, variables = zip(*optimizer.compute_gradients(total_training_loss))\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 35$\n, in compute_gradients\n    colocate_gradients_with_ops=colocate_gradients_with_ops)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 48$\n, in gradients\n    in_grads = grad_fn(op, *out_grads)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/nn_grad.py\", line 269, in _$\neluGrad\n    return gen_nn_ops._relu_grad(grad, op.outputs[0])\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_nn_ops.py\", line 2212, $\nn _relu_grad\n    features=features, name=name)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", l$\nne 763, in apply_op\n    op_def=op_def)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2395, i$\n create_op\n    original_op=self._default_original_op, op_def=op_def)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1268, i$\n __init__\n    self._control_flow_context.AddOp(self)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line $\n039, in AddOp\n    self._AddOpInternal(op)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line $\n062, in _AddOpInternal\n    real_x = self.AddValue(x)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line $\n998, in AddValue\n    real_val = grad_ctxt.grad_state.GetRealValue(val)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line $\n001, in GetRealValue\n    history_value = cur_grad_state.AddForwardAccumulator(cur_value)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 8\n92, in AddForwardAccumulator\n    self.forward_index.op._add_control_input(push.op)\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1434, in\n _add_control_input\n    self._add_control_inputs([op])\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1422, in\n _add_control_inputs\n    self._recompute_node_def()\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1442, in\n _recompute_node_def\n    self._control_inputs])\n  File \".local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1317, in\n name\n    return self._node_def.name\nKeyboardInterrupt\n<\/code><\/pre>\n","answers":[{"AnswerId":"54702751","CreationDate":"2019-02-15T04:36:52.430","ParentId":null,"OwnerUserId":"5002686","Title":null,"Body":"<p>I met this problem with three cause:<\/p>\n\n<ol>\n<li><p>model to large, so decrease batch size<\/p><\/li>\n<li><p>there is var with no gradient:<\/p>\n\n<pre><code> clone_gradients = optimizer.compute_gradients(total_clone_loss)\n for grad_and_vars in zip(*clone_grads):\n     tf.logging.info(\"clone_grads\"+str(grad_and_vars))\n<\/code><\/pre>\n\n<p>It prints:<\/p>\n\n<p>INFO:tensorflow:after clone_grads((, ),)\nINFO:tensorflow:after clone_grads((None, ),)<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"53024803","CreationDate":"2018-10-27T18:02:23.690","ParentId":null,"OwnerUserId":"10530728","Title":null,"Body":"<p>If at this point you did not began the training, maybe it is related to the graph construction. Are you sure that the GivenModel is correct?\nBecause I adapted this autoencoder <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/autoencoder.py\" rel=\"nofollow noreferrer\">example<\/a> with your definition of the optimizer as below and I did not found any problems when executing this code:<\/p>\n\n<pre><code>from __future__ import division, print_function, absolute_import\n\nimport tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# Import MNIST data\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n\n# Training Parameters\nlearning_rate = 0.01\nnum_steps = 10\nbatch_size = 8\n\n# Network Parameters\nnum_hidden_1 = 256 # 1st layer num features\n\nnum_hidden_2 = 128 # 2nd layer num features (the latent dim)\nnum_input = 784 # MNIST data input (img shape: 28*28)\n\n# tf Graph input (only pictures)\nX = tf.placeholder(\"float\", [None, num_input])\n\nweights = {\n    'encoder_h1': tf.Variable(tf.random_normal([num_input, num_hidden_1])),\n    'encoder_h2': tf.Variable(tf.random_normal([num_hidden_1, num_hidden_2])),\n    'decoder_h1': tf.Variable(tf.random_normal([num_hidden_2, num_hidden_1])),\n    'decoder_h2': tf.Variable(tf.random_normal([num_hidden_1, num_input])),\n}\nbiases = {\n    'encoder_b1': tf.Variable(tf.random_normal([num_hidden_1])),\n    'encoder_b2': tf.Variable(tf.random_normal([num_hidden_2])),\n    'decoder_b1': tf.Variable(tf.random_normal([num_hidden_1])),\n    'decoder_b2': tf.Variable(tf.random_normal([num_input])),\n}\n\n# Building the encoder\ndef encoder(x):\n    # Encoder Hidden layer with sigmoid activation #1\n    layer_1 = tf.nn.sigmoid(tf.add(tf.matmul(x, weights['encoder_h1']),\n                                   biases['encoder_b1']))\n    # Encoder Hidden layer with sigmoid activation #2\n    layer_2 = tf.nn.sigmoid(tf.add(tf.matmul(layer_1, weights['encoder_h2']),\n                                   biases['encoder_b2']))\n    return layer_2\n\n\n# Building the decoder\ndef decoder(x):\n    # Decoder Hidden layer with sigmoid activation #1\n    layer_1 = tf.nn.sigmoid(tf.add(tf.matmul(x, weights['decoder_h1']),\n                                   biases['decoder_b1']))\n    # Decoder Hidden layer with sigmoid activation #2\n    layer_2 = tf.nn.sigmoid(tf.add(tf.matmul(layer_1, weights['decoder_h2']),\n                                   biases['decoder_b2']))\n    return layer_2\n\n# Construct model\nencoder_op = encoder(X)\ndecoder_op = decoder(encoder_op)\n\n# Prediction\ny_pred = decoder_op\n# Targets (Labels) are the input data.\ny_true = X\n\n# Define loss and optimizer, minimize the squared error\n### your code with a reconstruction loss\nwith tf.device('\/gpu:0'):\n    with tf.variable_scope('prediction_loss'):\n\n        loss = tf.reduce_mean(tf.pow(y_true - y_pred, 2))\n\n        optimizer = tf.train.AdamOptimizer()\n        gradients, variables = zip(*optimizer.compute_gradients(loss))\n        gradients, _ = tf.clip_by_global_norm(gradients, 5.0)\n        optimize = optimizer.apply_gradients(zip(gradients, variables))\n\n    with tf.control_dependencies([optimize]):\n        train_op = tf.constant(0)\n### end of your code\n\n# Initialize the variables (i.e. assign their default value)\ninit = tf.global_variables_initializer()\n\n# Start Training\n# Start a new TF session\nwith tf.Session() as sess:\n\n    # Run the initializer\n    sess.run(init)\n\n    # Training\n    for i in range(1, num_steps+1):\n        # Prepare Data\n        # Get the next batch of MNIST data (only images are needed, not labels)\n        batch_x, _ = mnist.train.next_batch(batch_size)\n\n        # Run optimization op (backprop) and cost op (to get loss value)\n        _, l = sess.run([train_op, loss], feed_dict={X: batch_x})\n        # Display logs per step\n        print('Step %i: Minibatch Loss: %f' % (i, l))\n<\/code><\/pre>\n\n<p>So, I think that maybe the problem is with the rest of the model, but to be sure we need further details of the model.<\/p>\n\n<p>Now, about whether the placement of the model is in cpu or gpu. If you did not define anything being on cpu, the gpu device will be automatically selected for you. So, in theory the model will be allocated on gpu automatically. But, again, maybe the graph construction has problems and it did not reach the point when the model is actually allocated in the gpu memory. <\/p>\n"},{"AnswerId":"53026863","CreationDate":"2018-10-27T22:35:01.603","ParentId":null,"OwnerUserId":"4400904","Title":null,"Body":"<p>The problem for me was that the model was too big. Making it smaller fixed the issue.<\/p>\n"}]}
+{"QuestionId":53014688,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-26T18:47:44.323","AcceptedAnswerId":null,"OwnerUserId":8909870.0,"Title":"Extract histogram values from tensorboard and plot it with matplotlib","Body":"<p>I want plot histograms from tensorboard on my own, to publish it. I wrote this extraction function to get the histogram values:<\/p>\n\n<pre><code>def _load_hist_from_tfboard(path):\n    event_acc = event_accumulator.EventAccumulator(path)\n    event_acc.Reload()\n    vec_dict = {}\n    for tag in sorted(event_acc.Tags()[\"distributions\"]):\n        hist_dict = {}\n\n        for hist_event in event_acc.Histograms(tag):\n            hist_dict.update({hist_event.step: (hist_event.histogram_value.bucket_limit,\n                                                hist_event.histogram_value.bucket)})\n        vec_dict[tag] = hist_dict\n    return vec_dict\n<\/code><\/pre>\n\n<p>The function collects all histograms of a event file. The output of one bucket_limit and bucket is as follows:<\/p>\n\n<pre><code>[0.0, 1e-12, 0.0005418219168477906, 0.0005960041085325697, 0.0020575678275470133, 0.0022633246103017147, 0.004009617609950718, 0.00441057937094579, 0.005336801038844407, 0.005870481142728848, 0.007813610400972098, 0.008594971441069308, 0.022293142370048362, 0.0245224566070532, 0.026974702267758523, 0.035903328718386605, 0.03949366159022527, 0.043443027749247805, 0.04778733052417259, 0.052566063576589855, 0.057822669934248845, 0.06360493692767373, 0.06996543062044111, 0.07696197368248522, 0.24153964213356663, 0.2656936063469233, 0.29226296698161564, 0.3214892636797772, 0.35363819004775493, 0.38900200905253046, 0.42790220995778355, 0.47069243095356195, 0.5177616740489182, 0.56953784145381, 0.6264916255991911, 0.6891407881591103, 0.7580548669750213, 0.8338603536725235, 0.917246389039776, 1.0089710279437536]\n[0.0, 3999936.0, 0.0, 4.0, 0.0, 4.0, 0.0, 4.0, 0.0, 4.0, 0.0, 4.0, 0.0, 4.0, 4.0, 0.0, 8.0, 8.0, 0.0, 4.0, 4.0, 0.0, 8.0, 4.0, 0.0, 9.0, 45.0, 50.0, 48.0, 85.0, 100.0, 109.0, 114.0, 15908.0, 74.0, 15856.0, 11908.0, 3973.0, 42.0, 7951679.0]\n<\/code><\/pre>\n\n<p>Can someone help me to interpret these numbers to a histogram.<\/p>\n","answers":[]}
+{"QuestionId":53015441,"AnswerCount":1,"Tags":"<docker><tensorflow><centos>","CreationDate":"2018-10-26T19:49:21.213","AcceptedAnswerId":null,"OwnerUserId":10564919.0,"Title":"TensorFlow Docker Images on CentOS","Body":"<p>TensorFlow Official Supports a limited set of Operating System. If I use the tensorflow docker image on CentOS would that be a supported configuration?<\/p>\n\n<p>I understand that I can get TensorFlow to work, I am cwondering about being an approved configuration<\/p>\n\n<p>Thanks, <\/p>\n","answers":[{"AnswerId":"53019655","CreationDate":"2018-10-27T07:22:09.007","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>If you use Docker, you can install Ubuntu as base image. You should definitely try Nvidia Cloud GPU. It is a free docker from Nvidia to use tensorflow or any other framework from a container.<\/p>\n"}]}
+{"QuestionId":53015457,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-26T19:51:14.497","AcceptedAnswerId":null,"OwnerUserId":3161192.0,"Title":"How to build a TensorFlow graph with NN layers in a for loop?","Body":"<p>The TF tutorial I'm using builds the graph for forward propagation in the following way:<\/p>\n\n<pre><code>Z1 = tf.add(tf.matmul(W1, X), b1)\nA1 = tf.nn.relu(Z1)\nZ2 = tf.add(tf.matmul(W2, A1), b2)\nA2 = tf.nn.relu(Z2)\nZ3 = tf.add(tf.matmul(W3, A2), b3)\n<\/code><\/pre>\n\n<p>I'm trying to build basically the same thing, but not with a fixed set of layers.  My initial intent was to adapt the structure of my hand-coded python version of the NN. I start by initializing the parameters (in a function):<\/p>\n\n<pre><code>def initialize_parameters_deep(self, layer_dims):\n    \"\"\"\n    Arguments:\n    layer_dims -- python array (list) containing the dimensions of each layer in the network\n\n    Returns:\n    parameters -- python dictionary containing the parameters (as tensors) \"W1\", \"b1\", ..., \"WL\", \"bL\":\n                    Wl -- weight matrix of shape (layer_dims[l], layer_dims[l-1])\n                    bl -- bias vector of shape (layer_dims[l], 1)\n    \"\"\"\n\n    parameters = {}\n    L = len(layer_dims)            # number of layers in the network\n\n    for l in range(1, L):\n        parameters['W' + str(l)] = tf.get_variable('W' + str(l), [layer_dims[l], layer_dims[l-1]], initializer=tf.contrib.layers.xavier_initializer())\n        parameters['b' + str(l)] = tf.get_variable('b' + str(l), [layer_dims[l], 1], initializer=tf.zeros_initializer())\n\n    return parameters\n<\/code><\/pre>\n\n<p>With that as my parameters, I want to do forward propagation this way:<\/p>\n\n<pre><code>def forward_propagation(self, X, parameters):\n    \"\"\"\n    Implement forward propagation for the [LINEAR-&gt;RELU]*(L-1)-&gt;LINEAR-&gt;SOFTMAX computation\n\n    Arguments:\n    X -- data, tensor with shape (input size, number of examples)\n    parameters -- output of initialize_parameters_deep()\n\n    Returns:\n    AL -- last post-activation value\n    \"\"\"\n\n    A = X\n    L = len(parameters) \/\/ 2                  # number of layers in the neural network\n    for l in range(1, L):\n        A_prev = A \n        Z = tf.add(tf.matmul(parameters[\"W\" + str(l)], A_prev), parameters[\"b\" + str(l)])\n        A = tf.nn.relu(Z)\n\n    Z = tf.add(tf.matmul(parameters[\"W\" + str(L), A), parameters[\"b\" + str(L)])\n    AL = tf.nn.softmax(Z)\n\n    return AL, caches\n<\/code><\/pre>\n\n<p>But I feel like there's no way for TF to evaluate that python for loop. Is there a way to do this with a loop that TF can effectively build into the graph?<\/p>\n","answers":[]}
+{"QuestionId":53015550,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-10-26T19:58:52.047","AcceptedAnswerId":53015631.0,"OwnerUserId":3047919.0,"Title":"Why is a learning curve necessary to determine if a neural network has high bias or variance?","Body":"<p>In Andrew Ng's machine learning course it is recommended that you plot the learning curve (training set size vs cost) to determine if your model has a high bias or variance. <\/p>\n\n<p>However, I am training my model using Tensorflow and see that my validation loss is increasing while my training loss is decreasing. It's my understanding that this means my model is overfitting and so I have high variance. Is there still a reason to plot the learning curve?<\/p>\n","answers":[{"AnswerId":"53015631","CreationDate":"2018-10-26T20:05:51.397","ParentId":null,"OwnerUserId":"2161778","Title":null,"Body":"<p>Yes, there is, but it's not for spotting overfitting only. But anyway, plotting is just fancy way to see numbers, and sometimes it gives you insights. If you are monitoring loss on train\/validation simultaneously \u2013 you're looking at same data, obviously.<\/p>\n\n<p>Regarding Andrew's ideas \u2013 I suggest looking into Deep Learning course by him, he clarifies that in modern applications (DL + a lot of data, and I believe, this is your case) bias is not an opposite of variance.<\/p>\n"}]}
+{"QuestionId":53016377,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T21:13:49.320","AcceptedAnswerId":null,"OwnerUserId":4007547.0,"Title":"tf.MonitoredSession accessing raw session without hooks: run_step_fn vs _tf_sess","Body":"<p>I am using tf.MonitoredSession, it happens sometimes that you want to evaluate some nodes but you don't want to proceed in the training step, so you do not want to use the Hooks to log stuffs.\nIt seems that the session is wrapped several times and the raw session is stored in a nested way: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8425\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8425<\/a>\nbut it can also be retrieved with method _tf_sess() <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11971\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11971<\/a><\/p>\n\n<p>To recap I think there are two possible ways to run without hooks:<\/p>\n\n<ol>\n<li>use step_fn <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/MonitoredSession#run_step_fn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/MonitoredSession#run_step_fn<\/a><\/li>\n<li>access the raw session with <code>raw_session = sess._tf_sess()<\/code><\/li>\n<\/ol>\n\n<p>I think way 2. is the more intuitive for me if I want to be free from the hooks, but is it proper practice are there some problems in doing this?\nI wonder why is it private and it is not contemplated to have a method to access the raw session like in <code>SingularMonitoredSession.raw_session()<\/code>? <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/SingularMonitoredSession\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/SingularMonitoredSession<\/a>\n<br\/> <br\/>\nAt line 1010 in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/training\/monitored_session.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/training\/monitored_session.py<\/a> I can see that raw_session is defined exactly returning self._tf_sess(), so I wonder why the explicit choice to not have the user the possibility to access the raw session in MonitoredSession, is it possible to incur in some problems using sess._tf_sess()?\n  I wonder if it has something to do with some problems in the distributed setting..<\/p>\n","answers":[]}
+{"QuestionId":53016619,"AnswerCount":1,"Tags":"<python><tensorflow><pip>","CreationDate":"2018-10-26T21:41:01.320","AcceptedAnswerId":null,"OwnerUserId":10389160.0,"Title":"How to correctly install TensorFlow using pip?","Body":"<p>I'm trying to install TensorFlow via pip3 on my Mac. Whenever I run the command <code>pip3 install tensorflow<\/code> or <code>pip3 install --upgrade tensorflow<\/code> I receive the same error...<\/p>\n\n<pre><code>Could not find a version that satisfies the requirement tensorflow (from versions: )\n<\/code><\/pre>\n\n<p>No matching distribution found for tensorflow<\/p>\n\n<p>Python version: 3.7.0<\/p>\n\n<p>Pip version: 18.1<\/p>\n\n<p>Is there another way I could go about installing?<\/p>\n","answers":[{"AnswerId":"53016740","CreationDate":"2018-10-26T21:54:21.787","ParentId":null,"OwnerUserId":"3811555","Title":null,"Body":"<p>Aren't you using 'notebook'?\n'Anaconda' and 'notebook' are best to work with python and ML packages. You shouldn't have trouble with conda.<\/p>\n\n<p>Try just \"pip install tensorflow\"<\/p>\n"}]}
+{"QuestionId":53016653,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-26T21:45:23.787","AcceptedAnswerId":53020115.0,"OwnerUserId":4500293.0,"Title":"DenseNet in Tensorflow","Body":"<p>I am fairly new to tensorflow and I am interested in developing a DeseNet Architecture. I have found implementations from scratch on Github. I was wondering if the tensorflow API happen to implement the dense blocks. Is tensorflow's tf.layers.dense the same as the dense blocks in DenseNet?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53020115","CreationDate":"2018-10-27T08:36:29.767","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>No, <code>tf.layers.dense<\/code> implements what is more commonly known as a fully-connected layer, i.e. the basic building block of multilayer perceptrons. If you want dense blocks, you will need to to write your own implementation or use one of those you found on Github.<\/p>\n"}]}
+{"QuestionId":53017177,"AnswerCount":1,"Tags":"<python><tensorflow><keras><sequential>","CreationDate":"2018-10-26T22:48:44.470","AcceptedAnswerId":53017742.0,"OwnerUserId":7494282.0,"Title":"Multiple inputs to Keras Sequential model","Body":"<p>I am trying to merge output from two models and give them as input to the third model using keras sequential model. \nModel1 :<\/p>\n\n<pre><code>inputs1 = Input(shape=(750,))\nx = Dense(500, activation='relu')(inputs1)\nx = Dense(100, activation='relu')(x)\n<\/code><\/pre>\n\n<p>Model1 :<\/p>\n\n<pre><code>inputs2 = Input(shape=(750,))\ny = Dense(500, activation='relu')(inputs2)\ny = Dense(100, activation='relu')(y)\n<\/code><\/pre>\n\n<p>Model3 :<\/p>\n\n<pre><code>merged = Concatenate([x, y])\nfinal_model = Sequential()\nfinal_model.add(merged)\nfinal_model.add(Dense(100, activation='relu'))\nfinal_model.add(Dense(3, activation='softmax'))\n<\/code><\/pre>\n\n<p>\nTill here, my understanding is that, output from two models as x and y are merged and given as input to the third model. But when I fit this all like,<\/p>\n\n<pre><code>module3.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\nmodule3.fit([in1, in2], np_res_array)\n<\/code><\/pre>\n\n<p>in1 and in2 are two numpy ndarray of dimention 10000*750 which contains my training data and np_res_array is the corresponding target.<br> <strong>This gives me error as 'list' object has no attribute 'shape'<\/strong> As far as know, this is how we give multiple inputs to a model, but what is this error? How do I resolve it?<\/p>\n","answers":[{"AnswerId":"53017742","CreationDate":"2018-10-27T00:22:02.260","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can't do this using Sequential API. That's because of two reasons:<\/p>\n\n<ol>\n<li><p>Sequential models, as their name suggests, are a sequence of layers where each layer is connected directly to its previous layer and therefore they cannot have branches (e.g. merge layers, multiple input\/output layers, skip connections, etc.).<\/p><\/li>\n<li><p>The <code>add()<\/code> method of Sequential API accepts a <code>Layer<\/code> instance as its argument and not a <code>Tensor<\/code> instance.  In your example <code>merged<\/code> is a Tensor (i.e. concatenation layer's output).<\/p><\/li>\n<\/ol>\n\n<p>Further, the correct way of using <code>Concatenate<\/code> layer is like this:<\/p>\n\n<pre><code>merged = Concatenate()([x, y])\n<\/code><\/pre>\n\n<p>However, you can also use <code>concatenate<\/code> (note the lowercase \"c\"), its equivalent functional interface, like this:<\/p>\n\n<pre><code>merged = concatenate([x, y])\n<\/code><\/pre>\n\n<p>Finally, to be able to construct that third model you also need to use the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">functional API<\/a>.<\/p>\n"}]}
+{"QuestionId":53017373,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><training-data><mini-batch>","CreationDate":"2018-10-26T23:17:12.963","AcceptedAnswerId":53041284.0,"OwnerUserId":9473627.0,"Title":"Does TensorFlow optimizer minimize API implemented mini-batch?","Body":"<p>Does Tensorflow <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/GradientDescentOptimizer\" rel=\"nofollow noreferrer\">minimize<\/a> API for, say GradientDescentOptimizer implement mini-batch already when we feed the input tensor with a minibatch size of data?<\/p>\n\n<p>I was reading this <a href=\"http:\/\/deeplearnphysics.org\/Blog\/minibatch.html\" rel=\"nofollow noreferrer\">bolg<\/a> which indicated that the minibatch is not implemented in the <code>minimize<\/code> method, we have to do <code>compute_gradients<\/code> first and then accumulate the gradients, finally do the <code>apply_gradients<\/code> to finish the minibatch training.<\/p>\n\n<pre><code>def train_standard(opt_type, learning_rate, image_set):\n# Arrays for logging accuracy and loss\nacc_log = np.zeros(len(image_set))\nloss_log = np.zeros(len(image_set))\n# Create optimizer\nopt = opt_type(learning_rate)\n#\n# no-minibatch (standard, simple) operation\n#\nminimize = opt.minimize(loss)\n# Create session to execute ops\nsess = tf.InteractiveSession()\n# Necessary initializations\ntf.set_random_seed(1234)\ntf.global_variables_initializer().run()\n# Train loop\nfor i, batch in enumerate(image_set):\n    sess.run(minimize, feed_dict={x: batch[0], y_: batch[1]})\n\n    acc_log[i] = sess.run(accuracy, \n                          feed_dict={x: mnist.test.images, y_: mnist.test.labels})\n    loss_log[i] = sess.run(loss, \n                           feed_dict={x: mnist.test.images, y_: mnist.test.labels})\n\nreturn acc_log, loss_log\n<\/code><\/pre>\n\n<p>However, when I do the experiments, I found that two approaches generate similar results. I wonder if the <code>minimize<\/code> method will do the minibatch update if we <code>feed_dict<\/code> is a minibatch size matrix instead just one row of the training data.<\/p>\n\n<p>Can anyone help me to clarify this question and correct me if I am wrong?<\/p>\n\n<p>Best regards<\/p>\n","answers":[{"AnswerId":"53041284","CreationDate":"2018-10-29T08:15:21.107","ParentId":null,"OwnerUserId":"8394915","Title":null,"Body":"<p>It depends on your definition of learning with minibatches. One way to do that is to simply sample a minibatch, perform a weight update (i.e., compute forward and backward pass), sample another minibatch, perform another weight update, and so on. This is easily done with <code>Optimizer.minimize()<\/code> when you just feed one minibatch at a time. AFAIK, this is the most commonly used method.<\/p>\n\n<p>The post you're liking aims to do something else: compute the gradients on multiple minibatches (i.e. compute forward and backward pass, but don't change the weights), and then perform a single weight update using all of the the accumulated gradients. That is of course different, and more work to implement (as shown in the post).<\/p>\n"}]}
+{"QuestionId":53017571,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-10-26T23:53:09.470","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Is there a decent workaround to saving checkpoints in local drive when using TPU in Tensorflow?","Body":"<p>A follow up to this question:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/53003208\/how-to-save-a-tensorflow-checkpoint-file-from-google-colaboratory-in-when-using\">How to save a Tensorflow Checkpoint file from Google Colaboratory in when using TPU mode?<\/a><\/p>\n\n<p>Where the official way of saving a checkpoint when using a Tensorflow TPU is to use the Google Cloud Service. <\/p>\n\n<p>I am working if there is a workaround to this for those who do not wish to use GCS. Perhaps for each variable, do a .eval(), save the variable. And then set the save variable to the 'init' value for each variable. <\/p>\n\n<p>A major issue I foresee though is saving and loading the parameters for the optimizers. <\/p>\n\n<p>For Keras, the weights do seem to be saved from TPU to local<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/shakespeare_with_tpu_and_keras.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/shakespeare_with_tpu_and_keras.ipynb<\/a><\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Copying TPU weights to the CPU<\/p>\n<\/blockquote>\n\n<p>So I imagine that there's a general workaround too, without using keras. <\/p>\n","answers":[{"AnswerId":"53036803","CreationDate":"2018-10-28T22:58:30.677","ParentId":null,"OwnerUserId":"1433554","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/bde19b79b3d5e0d4be679f8144acb7f06b1c83ee\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py#L2040-L2072\" rel=\"nofollow noreferrer\">THIS CODE from Keras<\/a> <\/p>\n\n<p>If I understood correctly weights are not saved drectly from TPU, instead weights are synced to CPU and the saved to colab storage.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Also see: <a href=\"https:\/\/stackoverflow.com\/a\/53056640\/1433554\">this answer<\/a>.<\/p>\n"}]}
+{"QuestionId":53017722,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite><quantization>","CreationDate":"2018-10-27T00:17:21.720","AcceptedAnswerId":53069641.0,"OwnerUserId":10360610.0,"Title":"Input image of a fully quantized tensorflow lite model","Body":"<p>I've trained a simple CNN model on Cifar-10 in tensorflow with fake quantization (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/quantize\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/quantize<\/a>). I then generated a .tflite file using toco. Now I want to use a python interpreter to test the tflite model. <br\/><\/p>\n\n<p>Since I used tf.image.per_image_standardization to subtract mean and divide by variance during training. I need to do the same thing to the testing data right? But, the problem is, my model is already fully quantized by tflite, and it only takes uint8 data as inputs. To do image standardization, I need to convert my image to float32. So how do I convert it back to uint8, or is image standardization even necessary for the testing data in this case? Thanks.<\/p>\n","answers":[{"AnswerId":"53069641","CreationDate":"2018-10-30T17:17:57.817","ParentId":null,"OwnerUserId":"10360610","Title":null,"Body":"<p>So, it turns out I need to do standardization on the testing data for a good accuracy.\nTo do it, I directly feed uint8 input images to the tf.image.per_image_standardization function. The function would convert the uint8 data to float32, and then do standardization (subtract mean, divide by std). You can find source code of the function here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/ops\/image_ops_impl.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/ops\/image_ops_impl.py<\/a><\/p>\n\n<p>Now, I have the standardized float32 input images. What I did is writing a quantization function to quantize the float32 images back to uint8. The math comes from this paper: <a href=\"https:\/\/arxiv.org\/abs\/1803.08607\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1803.08607<\/a><\/p>\n\n<p>Now, I have the <strong>standardized uint8<\/strong> input images, I then use tflite interpreter python API to test the model. It works as expected.<\/p>\n"}]}
+{"QuestionId":53018117,"AnswerCount":1,"Tags":"<javascript><tensorflow><google-chrome-extension><computer-vision>","CreationDate":"2018-10-27T01:59:29.203","AcceptedAnswerId":null,"OwnerUserId":2225387.0,"Title":"Variable not defined after including third party library in chrome extension","Body":"<p>I'm trying to add the posenet tensor flow model to a chrome extension. I copied the libraries from the example: <\/p>\n\n<pre><code>&lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.11.7\"&gt;&lt;\/script&gt;\n&lt;!-- Load Posenet --&gt;\n&lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow-models\/posenet@0.1.2\"&gt;&lt;\/script&gt;\n<\/code><\/pre>\n\n<p>to a folder called \"tensorflow\" in my extension's directory.\nI imported the scripts from local in my background.html file <\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code> &lt;html&gt;\r\n    &lt;video autoplay=\"true\" id=\"webcamVideo\" width=\"227\" height=\"227\"&gt;&lt;\/video&gt;\r\n    &lt;script src=\"third_party\/deeplearnjs\/deeplearn.0.5.js\"&gt;&lt;\/script&gt;\r\n    &lt;script src=\"third_party\/deeplearnjs\/deeplearn_knn_image_classifier.0.4.2.js\"&gt; &lt;\/script&gt;\r\n    &lt;script src=\"third_party\/@tensorflow\/tensorflow.js\"&gt;&lt;\/script&gt;\r\n    &lt;script src=\"third_party\/@tensorflow-models\/posenet.js\"&gt;&lt;\/script&gt;\r\n    &lt;script src=\"background.js\"&gt;&lt;\/script&gt;\r\n    &lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>This seems to have no errors and I can successfully use objects from the deeplearn libraries in background.js. But when I try to use posenet in background.js, I get the error that posenet is not defined.<\/p>\n\n<pre><code>posenet.load().then(function(net){\n  return net.estimateSinglePose(imageElement, imageScaleFactor, flipHorizontal, outputStride)\n}).then(function(pose){\n  console.log(pose);\n})\n<\/code><\/pre>\n\n<p>I've used posenet in React successfuly but I'm struggling to get it to work in this extension. Am I missing something in this workflow?<\/p>\n","answers":[{"AnswerId":"53102463","CreationDate":"2018-11-01T13:39:12.570","ParentId":null,"OwnerUserId":"2225387","Title":null,"Body":"<p>It seems that the dependency of tensorflow in posenet required the tensorflow library to be @tensorflow\/tfjs so I just renamed the file to tfjs.js and it worked. <\/p>\n"}]}
+{"QuestionId":53018182,"AnswerCount":1,"Tags":"<numpy><machine-learning><scikit-learn><keras><one-hot-encoding>","CreationDate":"2018-10-27T02:16:41.840","AcceptedAnswerId":null,"OwnerUserId":7455441.0,"Title":"Multi-label classification: decoding one hot vector","Body":"<p>I am currently working on a multi-label classification problem where I am trying to classify images of fruits. Once I convert categories with one hot encoding how would I decode after  I train my model and want to get the appropriate category back ?<\/p>\n\n<pre><code>a = np.array(['a', 'b', 'c', 'a', 'b', 'c'])\nb = pandas.get_dummies(a)\n<\/code><\/pre>\n\n<p><strong>Sample<\/strong><\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(df, b, test_size=0.20, random_state=42)\n<\/code><\/pre>\n\n<p>......<\/p>\n\n<p><strong>Model training<\/strong><\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\n\nmodel = models.Sequential()\nmodel.add(layers.Dense(16, activation='relu', input_shape=(6,)))\nmodel.add(layers.Dense(16, activation='relu'))\nmodel.add(layers.Dense(1, activation='lerelu'))\n<\/code><\/pre>\n\n<p>.....<\/p>\n\n<p><strong>Predict<\/strong><\/p>\n\n<pre><code>model.predict(X_test[0]) ---&gt;&gt; result  ??????\n<\/code><\/pre>\n","answers":[{"AnswerId":"53019069","CreationDate":"2018-10-27T05:43:22.260","ParentId":null,"OwnerUserId":"10546779","Title":null,"Body":"<p>If I'm understanding your question correctly, wouldn't it just be:<\/p>\n\n<pre><code>test_predictions = model.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53018210,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2018-10-27T02:23:16.973","AcceptedAnswerId":53021132.0,"OwnerUserId":7879810.0,"Title":"How to share filter weights in Tensorflow","Body":"<p>For simplicity, assume the image has only 1 channel, we have one filter of size [2,2] which could be written as W=[[a, b],[c,d]], I want to use this set of parameters to do 2 convolutions, first is just using W itself:<\/p>\n\n<p>output = tf.nn.conv2d(input, W)<\/p>\n\n<p>Second convolution I would like to use filter W_2=[[d,c],[b,a]] which is basically some kind of 'transpose' of W. It needs to share parameters with W. How should I write this in tensorflow?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"53021132","CreationDate":"2018-10-27T10:49:56.983","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>You can just reverse the weights:<\/p>\n\n<pre><code>import tensorflow as tf\n\nw = tf.Variable([[1, 2.], [3, 4]])\nw_transposed = w[::-1, ::-1]\ngrad = tf.gradients(w_transposed, w)[0]\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print('Values')\n    print(w.eval())\n    print(w_transposed.eval())\n    print('Gradient')\n    print(grad.eval())\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre><code>Values\n[[1. 2.]\n [3. 4.]]\n[[4. 3.]\n [2. 1.]]\nGradient\n[[1. 1.]\n [1. 1.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53018226,"AnswerCount":1,"Tags":"<python><keras><deep-learning><reinforcement-learning><q-learning>","CreationDate":"2018-10-27T02:26:14.333","AcceptedAnswerId":53085844.0,"OwnerUserId":null,"Title":"Reinforcement Learning with Keras model","Body":"<p>I was trying to implement a q-learning algorithms in Keras. According to the articles i found these lines of code.     <\/p>\n\n<pre><code>for state, action, reward, next_state, done in sample_batch:\n        target = reward\n        if not done:\n            #formula\n          target = reward + self.gamma * np.amax(self.brain.predict(next_state)[0])\n        target_f = self.brain.predict(state)\n        #shape (1,2)\n        target_f[0][action] = target\n        print(target_f.shape)\n        self.brain.fit(state, target_f, epochs=1, verbose=0)\n    if self.exploration_rate &gt; self.exploration_min:\n        self.exploration_rate *= self.exploration_decay\n<\/code><\/pre>\n\n<p>Variable <code>sample_batch<\/code> is the array that contains sample <code>state, action, reward, next_state, done<\/code> from collected data.\nI also found the following q-learning formula <a href=\"https:\/\/i.stack.imgur.com\/XBsvH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XBsvH.png\" alt=\"Formula\"><\/a><\/p>\n\n<p>Why is there no <code>-<\/code> sign in the equation(code)? I found out that <code>np.amax<\/code> returns the maximum of an array or maximum along an axis. When i call <code>self.brain.predict(next_state)<\/code>, I get <code>[[-0.06427538 -0.34116858]]<\/code>. So it plays the role of prediction  in this equation? As we go forward <code>target_f<\/code> is the predicted output for the current state and then we also append to it the reward with this step. Then, we train model on current <code>state<\/code>(<code>X<\/code>) and <code>target_f<\/code>(<code>Y<\/code>). I have a few questions. What is the role of <code>self.brain.predict(next_state)<\/code> and why there is no minus? Why do we predict twice on one model? Ex <code>self.brain.predict(state) and self.brain.predict(next_state)[0]<\/code><\/p>\n","answers":[{"AnswerId":"53085844","CreationDate":"2018-10-31T14:33:27.123","ParentId":null,"OwnerUserId":"9740170","Title":null,"Body":"<blockquote>\n  <p>Why is there no - sign in the equation(code)?<\/p>\n<\/blockquote>\n\n<p>It's because loss calculation is done inside the fit function.  <\/p>\n\n<pre><code>reward + self.gamma * np.amax(self.brain.predict(next_state)[0])\n<\/code><\/pre>\n\n<p>This is the same as the <strong>target<\/strong> component in the loss function. <\/p>\n\n<p>Inside the fit method in keras loss will be calculated as given below.\nFor a single training data point (standard notations of neural networks),<\/p>\n\n<pre><code>x = input state\n\ny = predicted value\n\ny_i = target value\n\nloss(x) = y_i - y\n<\/code><\/pre>\n\n<p>at this step <strong>target - prediction<\/strong> happens internally. <\/p>\n\n<blockquote>\n  <p>Why do we predict twice on one model?<\/p>\n<\/blockquote>\n\n<p>Good question !!!<\/p>\n\n<pre><code> target = reward + self.gamma * np.amax(self.brain.predict(next_state)[0])\n<\/code><\/pre>\n\n<p>In this step we are predicting the value of the next state to calculate target <strong>value for state s if we take specific action a<\/strong> (which is denoted as Q(s,a) )<\/p>\n\n<pre><code> target_f = self.brain.predict(state)\n<\/code><\/pre>\n\n<p>In this step we are calculating all the <strong>Q values for every action<\/strong> which we can take in state s.<\/p>\n\n<pre><code>target = 1.00    \/\/ target is a single value for action a\ntarget_f = (0.25,0.25,0.25,0.25)   \/\/target_f is a list of values for all actions\n<\/code><\/pre>\n\n<p>following step is then executed.<\/p>\n\n<pre><code>target_f[0][action] = target\n<\/code><\/pre>\n\n<p>we only change the value of selected action. ( if we take action 3 )<\/p>\n\n<pre><code>target_f = (0.25,0.25,1.00,0.25)  \/\/ only action 3 value will change\n<\/code><\/pre>\n\n<p>Now target_f will be the <strong>actual target value<\/strong> we are trying to predict with correct shape.<\/p>\n"}]}
+{"QuestionId":53018240,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><modeling>","CreationDate":"2018-10-27T02:30:00.923","AcceptedAnswerId":null,"OwnerUserId":10240511.0,"Title":"Tensorflow: Does tf.estimator.DNNClassifier include a summary for training accuracy?","Body":"<p>I'm currently using tensorflow's built in DNN classifier (<code>tf.estimator.DNNClassifier<\/code>). Got the model to run successfully however when I view the summary scalars on tensorboard, it literally has everything I need except for training accuracy.<\/p>\n\n<p>I searched all over but only found solutions to custom built estimators. I was wondering if anyone here was able to get the training accuracy using <code>tf.estimator.DNNClassifier<\/code>. <\/p>\n\n<p>Any tips or suggestions are welcomed. Thanks in advance. <\/p>\n","answers":[{"AnswerId":"53028162","CreationDate":"2018-10-28T03:34:30.400","ParentId":null,"OwnerUserId":"10414259","Title":null,"Body":"<p>If you have a trained <code>DNNClassifier<\/code> then you can use a a built in method of the same class to get the accuracy. You can use the <code>evaluate()<\/code> method. Just make sure you define your input function according to you evaluation\/testing data. Here's an example of fitting a model on random data, evaluating the model on more random data and then printing the summary of the evaluation metrics, including the accuracy:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport pandas as pd\n\nnum = 100000\n\nfeat1 = np.random.rand(num,1)\nfeat2 = np.random.rand(num,1)\nlabel = (np.round(np.random.rand(num,1),0))\nData = pd.DataFrame(np.concatenate((feat1,feat2,label),axis=1), columns=['feat1','feat2','label'])\nfeature_columns = [\n    tf.feature_column.numeric_column(key=\"feat1\"),\n    tf.feature_column.numeric_column(key=\"feat2\"),\n]\n\nmodel = tf.estimator.DNNClassifier(feature_columns=feature_columns,hidden_units=[8,8,8,8,8,8,8,8,4,2])\n\ninput_train = tf.estimator.inputs.pandas_input_fn(\n    x=Data,\n    y=Data['label'],\n    shuffle=True,\n    num_epochs = 1\n)\n\nmodel.train(input_fn=input_train,steps=10000)\n\nfeat1T = np.random.rand(num,1)\nfeat2T = np.random.rand(num,1)\nlabelT = (np.round(np.random.rand(num,1),0))\nDataT = pd.DataFrame(np.concatenate((feat1T,feat2T,labelT),axis=1), columns=['feat1','feat2','label'])\n\ninput_eval = tf.estimator.inputs.pandas_input_fn(\n    x=DataT,\n    y=DataT['label'],\n    shuffle=True,\n    num_epochs = 1\n)\n\nmetrics = model.evaluate(input_fn=input_eval, steps=100)\n\nprint(metrics)\n<\/code><\/pre>\n\n<p>For my randomly random seed, this printed: <\/p>\n\n<blockquote>\n  <p>{'accuracy': 0.49734375, 'accuracy_baseline': 0.5026562, 'auc': 0.5, 'auc_precision_recall': 0.7486719, 'average_loss': 0.6931635, 'label\/mean': 0.49734375, 'loss': 88.72493, 'precision': 0.49734375, 'prediction\/mean': 0.5012483, 'recall': 1.0, 'global_step': 782}<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53018376,"AnswerCount":1,"Tags":"<python><keras><lstm><mfcc>","CreationDate":"2018-10-27T03:02:20.063","AcceptedAnswerId":53018987.0,"OwnerUserId":10565754.0,"Title":"mfcc in keras lstm error: Expected ndim=3, found ndim=2","Body":"<p>How do i solve this problem? <\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_10: expected\n  ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/zahiruddinnorzain\/keras_lstm_mfcc\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zahiruddinnorzain\/keras_lstm_mfcc<\/a><\/p>\n\n<p>dataset link : <a href=\"https:\/\/drive.google.com\/open?id=14o5djf8L7OjYLnoABg4rDi3kVWyGecRA\" rel=\"nofollow noreferrer\">dataset<\/a><\/p>\n\n<p>I got the above mentioned error when running this code. This code will training the mfcc data of digit, from 0 to 9 with 13 coefficient.<\/p>\n\n<pre><code>from __future__ import print_function\nimport numpy as np\n\nfrom keras.optimizers import SGD\nnp.random.seed(1337)\nfrom keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Activation\nfrom keras.layers.recurrent import LSTM\n#from SpeechResearch import loadData\n\nfrom sklearn.preprocessing import LabelEncoder\nimport pandas\n\n'exception_verbosity = high'\nbatch_size = 5\nhidden_units = 13\nnb_classes = 10\nprint('Loading data...')\n\n\n# load train dataset\ndataframe = pandas.read_csv(\"train.csv\", header=None)\ndataset = dataframe.values\nX_train = dataset[:,0:13] #.astype(float)\nY = dataset[:,13]\n# encode class values as integers\nencoder = LabelEncoder()\nencoder.fit(Y)\nencoded_Y = encoder.transform(Y)\n# convert integers to dummy variables (i.e. one hot encoded)\ny_train = np_utils.to_categorical(encoded_Y)\n\n\n# load test dataset\ndataframe = pandas.read_csv(\"test.csv\", header=None)\ndataset = dataframe.values\nX_test = dataset[:,0:13] #.astype(float)\ny_test = dataset[:,13]\n# encode class values as integers\nencoder2 = LabelEncoder()\nencoder2.fit(y_test)\nencoded_Y2 = encoder.transform(y_test)\n# convert integers to dummy variables (i.e. one hot encoded)\nY_test = np_utils.to_categorical(encoded_Y2)\n\n\n\n#(X_train, y_train), (X_test, y_test) = loadData.load_mfcc(10, 2)\n\nprint(len(X_train), 'train sequences')\nprint(len(X_test), 'test sequences')\nprint('X_train shape:', X_train.shape)\nprint('X_test shape:', X_test.shape)\nprint('y_train shape:', y_train.shape)\nprint('y_test shape:', y_test.shape)\nprint(y_test)\nprint('Build model...')\n\nY_train = np_utils.to_categorical(y_train, nb_classes)\nY_test = np_utils.to_categorical(y_test, nb_classes)\n\nmodel = Sequential()\nmodel.add(LSTM(output_dim=hidden_units, init='uniform', inner_init='uniform',\n               forget_bias_init='one', activation='tanh', inner_activation='sigmoid', input_shape=X_train.shape[1:]))\n\n\nmodel.add(Dense(nb_classes))\nmodel.add(Activation('softmax'))\n\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='categorical_crossentropy', optimizer=sgd)\n\nprint(\"Train...\")\nmodel.fit(X_train, y_train, batch_size=batch_size, nb_epoch=3, validation_data=(X_test, Y_test), show_accuracy=True)\nscore, acc = model.evaluate(X_test, Y_test,\n                            batch_size=batch_size,\n                            show_accuracy=True)\nprint('Test score:', score)\nprint('Test accuracy:', acc)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53018987","CreationDate":"2018-10-27T05:27:46.463","ParentId":null,"OwnerUserId":"10062944","Title":null,"Body":"<p>I worked around your code and here is the working code. There were many errors I had to face and warning were also there as you seemed to use old version of keras.\nBelow code is as per the updated version of keras:<\/p>\n\n<pre><code>from __future__ import print_function\nimport numpy as np\n\nfrom keras.optimizers import SGD\nnp.random.seed(1337)\nfrom keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Activation\nfrom keras.layers.recurrent import LSTM\n#from SpeechResearch import loadData\n\nfrom sklearn.preprocessing import LabelEncoder\nimport pandas\n\n'exception_verbosity = high'\nbatch_size = 5\nhidden_units = 13\nnb_classes = 10\nprint('Loading data...')\n\n\n# load train dataset\ndataframe = pandas.read_csv(\"train.csv\", header=None)\ndataset = dataframe.values\nX_train = dataset[:,0:13] #.astype(float)\nY = dataset[:,13]\n# encode class values as integers\nencoder = LabelEncoder()\nencoder.fit(Y)\nencoded_Y = encoder.transform(Y)\n# convert integers to dummy variables (i.e. one hot encoded)\n# y_train = np_utils.to_categorical(encoded_Y)\n\n\n# load test dataset\ndataframe = pandas.read_csv(\"test.csv\", header=None)\ndataset = dataframe.values\nX_test = dataset[:,0:13] #.astype(float)\ny_test = dataset[:,13]\n# encode class values as integers\nencoder2 = LabelEncoder()\nencoder2.fit(y_test)\nencoded_Y2 = encoder.transform(y_test)\n# convert integers to dummy variables (i.e. one hot encoded)\n# y_test = np_utils.to_categorical(encoded_Y2)\n\n\n\n#(X_train, y_train), (X_test, y_test) = loadData.load_mfcc(10, 2)\n\nprint(len(X_train), 'train sequences')\nprint(len(X_test), 'test sequences')\nprint('X_train shape:', X_train.shape)\nprint('X_test shape:', X_test.shape)\n# print('y_train shape:', y_train.shape)\n# print('y_test shape:', y_test.shape)\n# print(y_test)\nprint('Build model...')\n\nX_train = X_train.reshape(1, X_train.shape[0], X_train.shape[1])\nX_test = X_test.reshape(1, X_test.shape[0], X_test.shape[1])\nprint('X_train shape:', X_train.shape)\nprint('X_test shape:', X_test.shape)\n\nY_train = np_utils.to_categorical(encoded_Y, nb_classes)\nY_test = np_utils.to_categorical(encoded_Y2, nb_classes)\nprint(Y_train.shape)\nprint(Y_test.shape)\nY_train = Y_train.reshape(1, Y_train.shape[0], Y_train.shape[1])\nY_test = Y_test.reshape(1, Y_test.shape[0], Y_test.shape[1])\n\nmodel = Sequential()\nmodel.add(LSTM(units=hidden_units, kernel_initializer='uniform',\n           unit_forget_bias='one', activation='tanh', recurrent_activation='sigmoid', input_shape=(None,X_train.shape[2]),     return_sequences=True))\n\n\nmodel.add(Dense(nb_classes))\nmodel.add(Activation('softmax'))\n\nsgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\nprint(model.summary())\n\nprint(\"Train...\")\nmodel.fit(X_train, Y_train, batch_size=batch_size, epochs=3)\nscore, acc = model.evaluate(X_test, Y_test,\n                        batch_size=batch_size)\nprint('Test score:', score)\nprint('Test accuracy:', acc)\n<\/code><\/pre>\n\n<p>The error you were facing because, you hadn't reshaped the data before inputing in LSTM\nThe second error was faced because, in updated version of keras, There is no term like <code>show_accuracy<\/code>. We just need to define <code>metrics = ['accuracy]<\/code> while compiling model in <code>model.compile<\/code>\nThe third error was the way we define input shape in input layer. \nRest, You can read and let me know if there is any confusion.<\/p>\n"}]}
+{"QuestionId":53018621,"AnswerCount":1,"Tags":"<python><tensorflow><plot><keras>","CreationDate":"2018-10-27T04:07:38.947","AcceptedAnswerId":null,"OwnerUserId":10328263.0,"Title":"Plot model's testing accuracy when loading a model using load_model","Body":"<p>I'm trying to plot a testing accuracy for a model that is called using <code>tf.keras.models.load_model<\/code> (I'd initially saved the model as a .h5 file after getting satisfactory results in training &amp; validation accuracies).<\/p>\n\n<p>During training &amp; validation, I used the following codes to plot training &amp; validation accuracies:<\/p>\n\n<pre><code>epoch_list2 = list(range(1, len(hist2.history['acc']) + 1))\nplt.plot(epoch_list2, hist2.history['acc'], epoch_list2, hist2.history['val_acc'])\nplt.legend((\"Training Accuracy\", \"Validation Accuracy\"))\nplt.show()\n<\/code><\/pre>\n\n<p>However, when I now the call the model to test it on unseen data (for demo purposes), I'm using the following codes to first load the model, then evaluate on the test data &amp; finally calculate the accuracy:<\/p>\n\n<pre><code># Load the model\nfrom keras.models import load_model\nmodel = load_model('\/home\/ubuntu\/CNN_Model.h5')\nprint(model.summary())\n\n# Use model.evaluate:\ntest_score = model.evaluate(testset, labels, verbose = 0)\nprint(\"Accuracy: %.2f%%\" % (test_score[1]*100))\n<\/code><\/pre>\n\n<p>How do I now plot the testing accuracy score &amp; is it possible to plot all three (i.e. training, validation &amp; testing scores) in the same graph <strong>when loading a model<\/strong>?<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>Abhay.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>So after searching stack, i found this <a href=\"https:\/\/stackoverflow.com\/questions\/51004071\/how-can-we-plot-accuracy-and-loss-graphs-from-a-keras-model-saved-earlier\">link<\/a> that helps first save the <em>training history<\/em> and load it back. With that sorted out, how do I now plot the training, validation &amp; test scores in a single plot?<\/p>\n","answers":[{"AnswerId":"55339686","CreationDate":"2019-03-25T14:07:19.360","ParentId":null,"OwnerUserId":"7332429","Title":null,"Body":"<p>I think what you could do is to save the history.history['val_acc'], history.history['train_acc'] into a separate .csv file. And after you load the model and get an accuracy on the test data, you plot them together in a plot where test_accuracy, of course, is just a dot on plot, make sure to use \"maker\" in plot.<\/p>\n"}]}
+{"QuestionId":53019301,"AnswerCount":1,"Tags":"<python><neural-network><keras><svm>","CreationDate":"2018-10-27T06:23:46.400","AcceptedAnswerId":53019843.0,"OwnerUserId":9422637.0,"Title":"Add SVM to last layer","Body":"<p><strong>What I did:<\/strong><\/p>\n\n<p>I implement the following model using of Keras:<\/p>\n\n<pre><code>train_X, test_X, train_Y, test_Y = train_test_split(X, Y, test_size=0.2, random_state=np.random.seed(7), shuffle=True)\n\ntrain_X = np.reshape(train_X, (train_X.shape[0], 1, train_X.shape[1]))\ntest_X = np.reshape(test_X, (test_X.shape[0], 1, test_X.shape[1]))\n\ninp = Input((train_X.shape[1], train_X.shape[2]))\nlstm = LSTM(1, return_sequences=False)(inp)\noutput = Dense(train_Y.shape[1], activation='softmax')(lstm)\n\nmodel = Model(inputs=inp, outputs=output)\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\nmodel.fit(train_X, train_Y, validation_split=.20, epochs=2, batch_size=50)\n<\/code><\/pre>\n\n<p><strong>What I want:<\/strong><\/p>\n\n<p>I want to add SVM to the last layer in my model but i dont know how? Any idea?<\/p>\n","answers":[{"AnswerId":"53019843","CreationDate":"2018-10-27T07:53:09.003","ParentId":null,"OwnerUserId":"10062944","Title":null,"Body":"<p>This should work for adding svm as last layer.<\/p>\n\n<pre><code>inp = Input((train_X.shape[1], train_X.shape[2]))\nlstm = LSTM(1, return_sequences=False)(inp)\noutput = Dense(train_Y.shape[1], activation='softmax', W_regularizer=l2(0.01)))(lstm)\n\nmodel = Model(inputs=inp, outputs=output)\nmodel.compile(loss='hinge', optimizer='adam', metrics=['accuracy'])\nmodel.fit(train_X, train_Y, validation_split=.20, epochs=2, batch_size=50)\n<\/code><\/pre>\n\n<p>Here I have used <code>hinge<\/code> as loss considering binary categorised target. But if it is more than that, then you can consider using <code>categorical_hinge<\/code><\/p>\n"}]}
+{"QuestionId":53019944,"AnswerCount":1,"Tags":"<python><tensorflow><pre-trained-model>","CreationDate":"2018-10-27T08:08:09.623","AcceptedAnswerId":53020118.0,"OwnerUserId":10238777.0,"Title":"Read and preprocess image for tensorflow pretrained model","Body":"<p>I don't have much experience in Tensorflow. I am trying to use a pretrained ResNet152 model to get the activations of the last layer as output. The images I use for input are stored on my harddrive. So I need to load the images, preprocess them and then get the output from the pretrained model. I found examples for that using URLs of images but when I try it with image paths I can't get it to work. This is what I have so far (only one image for now):<\/p>\n\n<pre><code>with tf.Graph().as_default():\n\n    filename_queue = tf.train.string_input_producer(['.\/testimg\/A_008.jpg'])\n    reader = tf.WholeFileReader()\n    key, value = reader.read(filename_queue)\n    image = tf.image.decode_jpeg(value, channels=3)\n    preprocessing = preprocessing_factory.get_preprocessing('resnet_v2_152', is_training=False)\n    processed_image = preprocessing(image, 299,299)\n    processed_images  = tf.expand_dims(processed_image, 0)\n\n    with slim.arg_scope(resnet_v2.resnet_arg_scope()):\n        logits, end_points = resnet_v2.resnet_v2_152(processed_images, is_training=False)\n\n    checkpoints_dir='.\/models\/resnet_v2_152' \n    init_fn = slim.assign_from_checkpoint_fn(\n        os.path.join(checkpoints_dir, 'resnet_v2_152.ckpt'),\n        slim.get_variables_to_restore())\n\n    with tf.Session() as sess:\n        init_fn(sess)        \n        np_image, fv = sess.run([image, logits])\n<\/code><\/pre>\n\n<p>I am doing this in a Jupyter Notebook. When I execute the code I don't get an error message, it just keeps running and running until I restart the kernel.<\/p>\n\n<p>Any ideas what I did wrong? And how would I do it for multiple images?<\/p>\n","answers":[{"AnswerId":"53020118","CreationDate":"2018-10-27T08:37:06.553","ParentId":null,"OwnerUserId":"10238777","Title":null,"Body":"<p>I found the solution by replacing the <code>tf.WholeFileReader()<\/code> with <code>tf.read_file()<\/code>:<\/p>\n\n<pre><code>graph = tf.Graph()\n\nwith graph.as_default():\n    image_path = image = tf.placeholder(tf.string)\n    image = tf.image.decode_jpeg(tf.read_file(image_path), channels=3)\n    preprocessing = preprocessing_factory.get_preprocessing('resnet_v2_152', is_training=False)\n    processed_image = preprocessing(image, image_size, image_size)\n    processed_images  = tf.expand_dims(processed_image, 0)\n\n    with slim.arg_scope(resnet_v2.resnet_arg_scope()):\n        logits, end_points = resnet_v2.resnet_v2_152(processed_images, is_training=False)\n\n    checkpoints_dir='.\/models\/resnet_v2_152' \n    init_fn = slim.assign_from_checkpoint_fn(\n        os.path.join(checkpoints_dir, 'resnet_v2_152.ckpt'),\n        slim.get_variables_to_restore())\n\n\nimages = ['.\/testimg\/A_008.jpg', '.\/testimg\/logo.jpg']\n\nwith tf.Session(graph=graph) as sess:\n    init_fn(sess)  \n\n    for img in images:\n        fv = sess.run(logits, feed_dict={image_path: img})\n        print(fv)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53019950,"AnswerCount":0,"Tags":"<python><tensorflow><keras><mnist>","CreationDate":"2018-10-27T08:09:33.177","AcceptedAnswerId":null,"OwnerUserId":9137971.0,"Title":"mnist dataset. not predicting correctly on real life handwriting","Body":"<p>I am trying to predict handwriting from real data instead of test data that has been given in mnist by using keras and tensorflow ,but not able to predict correctly .On test and train data i am getting 96-97 percent accuracy , therefore it is not overfitting .here is my image <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/q2IYP.png\" rel=\"nofollow noreferrer\">image here<\/a><\/p>\n\n<p>and here is my code to read image using open cv and resizing it using numpy ,please help<\/p>\n\n<pre><code>img = cv2.imread('seven.png')\nimg = cv2.resize(img,(28,28))\nimg = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)\nimg = np.invert(img)\n\nimg = img\/255\nimg = img.astype('float64')\n\nprint(img)\n\nplt.imshow(img)\n\nimg= np.expand_dims(img,axis=0)\nprint(img.shape)\n<\/code><\/pre>\n\n<p>here is the prediction getting 6 as an answer:-<\/p>\n\n<pre><code>pr = model.predict(img)\nprint(np.argmax(pr[0]))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53020144,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-27T08:40:23.187","AcceptedAnswerId":53020202.0,"OwnerUserId":10263275.0,"Title":"How do you solve this problem in python when importing tensorflow","Body":"<p>FutureWarning: Conversion of the second argument of issubdtype from <code>float<\/code> to <code>np.floating<\/code> is deprecated. In future, it will be treated as <code>np.float64 == np.dtype(float).type<\/code>.\n  from ._conv import register_converters as _register_converters<\/p>\n","answers":[{"AnswerId":"53020202","CreationDate":"2018-10-27T08:47:41.743","ParentId":null,"OwnerUserId":"5822776","Title":null,"Body":"<p>It is caused by <code>h5py<\/code> package as <a href=\"https:\/\/github.com\/h5py\/h5py\/issues\/974\" rel=\"nofollow noreferrer\">known issue<\/a>.<\/p>\n\n<p>To overcome, You can do:<\/p>\n\n<pre><code>pip install numpy==1.13.0\nsudo pip3 install h5py==2.8.0rc1\n<\/code><\/pre>\n"}]}
+{"QuestionId":53020172,"AnswerCount":1,"Tags":"<python><keras><loss-function>","CreationDate":"2018-10-27T08:44:03.837","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"Custom loss function in Keras runs ony once","Body":"<p>I have defined a custom loss function in Keras as follows:<\/p>\n\n<pre><code>def loss_function_test(yTrue, yPred):\n    global i\n    i += 1\n    print(\"\\n\")\n    print(\"Loss Function:\", i, \" ----- \", yTrue, yPred)\n    print(\"\\n\")\n    res = k.sum(yTrue - yPred)\n    return res\n<\/code><\/pre>\n\n<p>and set it as follows:<\/p>\n\n<pre><code>model.compile(loss=loss_function_test,\n              optimizer='sgd',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>The problem is checking the <code>global i<\/code> and the <code>print<\/code> function in the loss function code, I realize that the function is only called once in training when I call <code>model.fit<\/code>. Even when I debug it, it is only called once. It seems that maybe the loss function is only called while <code>compile<\/code>. Can anyone explain why?<\/p>\n","answers":[{"AnswerId":"53020483","CreationDate":"2018-10-27T09:26:03.090","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>The function to <em>construct<\/em> the loss function op runs only once.<\/p>\n\n<p>If you want to see how often the op is evaluated by tensorflow you need to add a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\">tf.Print<\/a> operation into the tensorflow graph.<\/p>\n"}]}
+{"QuestionId":53020761,"AnswerCount":0,"Tags":"<tensorflow><go><k-means>","CreationDate":"2018-10-27T10:02:09.517","AcceptedAnswerId":null,"OwnerUserId":1526756.0,"Title":"How to run Tensorflow clustering algorithm model","Body":"<p>I need to run k-means algorithm from Tensorflow in Go, i.e. cluster a graph intro subgraphs according to nodes similarity matrix.<\/p>\n\n<p>I came across <a href=\"https:\/\/www.tonytruong.net\/running-a-keras-tensorflow-model-in-golang\/\" rel=\"nofollow noreferrer\">this article<\/a> which shows an example on how to run a Keras  trained model in Go. In this example the algo is of a supervised learning type. However in clustering algos, as I understand, there will be no model to save and export it to Go implementation.<\/p>\n\n<p>The reason I am interested in Tensorflow, is because I think its code is optimized and will run much faster than <a href=\"https:\/\/github.com\/mpraski\/clusters\" rel=\"nofollow noreferrer\">k-mean implementation<\/a> in Go, even with the scenario I described above.<\/p>\n\n<p>I need an opinion of whether:<\/p>\n\n<ol>\n<li>It is indeed impossible to use a Tensorflow k-mean algorithm in Go, and it is much better just to use k-means implemented in Go for this case.<\/li>\n<li>It is possible to do this, and some sort of example or ideas on how to do this are very much appreciated.<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":53020781,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><ssh><pip>","CreationDate":"2018-10-27T10:04:30.393","AcceptedAnswerId":null,"OwnerUserId":8003861.0,"Title":"Python program not working when run in a SSH session","Body":"<p>So I have a Python program stored on a remote server (which is running Ubuntu Server 18.04) . When I use VNC Viewer to remote control the server and run the python program inside a terminal (in that VNC session), everything work great.<\/p>\n\n<p>But when i created a SSH connection and try to run that same program inside my terminal windows on my local machine it will not work as I expected.<\/p>\n\n<p>So the command was like this:\n<code>python3 main.py --train<\/code><\/p>\n\n<p>And here is the error:<\/p>\n\n<pre><code>Unable to init server: Could not connect: Connection refused\nUnable to init server: Could not connect: Connection refused\n\n(main.py:24109): Gdk-CRITICAL **: 09:45:07.931: gdk_cursor_new_for_display: assertion 'GDK_IS_DISPLAY (display)' failed\n\n(main.py:24109): Gdk-CRITICAL **: 09:45:07.932: gdk_cursor_new_for_display: assertion 'GDK_IS_DISPLAY (display)' failed\ntf.estimator package not installed.\nTraceback (most recent call last):\n  File \"main.py\", line 8, in &lt;module&gt;\n    from models import Model, DecoderType\n  File \"\/home\/kop\/handwriting\/src\/models.py\", line 3, in &lt;module&gt;\n    from md_lstm import multi_dimensional_rnn_while_loop\n  File \"\/home\/kop\/handwriting\/src\/md_lstm.py\", line 2, in &lt;module&gt;\n    from tensorflow.contrib.rnn import RNNCell, LSTMStateTuple\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/__init__.py\", line 39, in &lt;module&gt;\n    from tensorflow.contrib import distribute\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/__init__.py\", line 27, in &lt;module&gt;\n    from tensorflow.contrib.distribute.python.tpu_strategy import TPUStrategy\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/python\/tpu_strategy.py\", line 24, in &lt;module&gt;\n    from tensorflow.contrib import tpu\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/tpu\/__init__.py\", line 64, in &lt;module&gt;\n    from tensorflow.contrib.tpu.python.tpu.keras_support import tpu_model as keras_to_tpu_model\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/tpu\/python\/tpu\/keras_support.py\", line 67, in &lt;module&gt;\n    from tensorflow.python.estimator import model_fn as model_fn_lib\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/__init__.py\", line 25, in &lt;module&gt;\n    import tensorflow.python.estimator.estimator_lib\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator_lib.py\", line 41, in &lt;module&gt;\n    from tensorflow.python.estimator.inputs import inputs\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/inputs\/inputs.py\", line 22, in &lt;module&gt;\n    from tensorflow.python.estimator.inputs.numpy_io import numpy_input_fn\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/inputs\/numpy_io.py\", line 26, in &lt;module&gt;\n    from tensorflow.python.estimator.inputs.queues import feeding_functions\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/inputs\/queues\/feeding_functions.py\", line 40, in &lt;module&gt;\n    import pandas as pd\n  File \"\/home\/kop\/.local\/lib\/python3.6\/site-packages\/pandas\/__init__.py\", line 50, in &lt;module&gt;\n    plot_params = pandas.plotting._style._Options(deprecated=True)\nAttributeError: module 'pandas' has no attribute 'plotting'\n<\/code><\/pre>\n\n<p>I tried to re-run <code>pip install<\/code> some packages inside my SSH session but it does not solve the problem.<\/p>\n\n<p>I also tried to run <code>python3<\/code> and try to import <code>tensorflow<\/code> and <code>pandas<\/code> which I thought it will be the source of the problem and it still can import that packages:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QAKmw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QAKmw.png\" alt=\"lol i can import those packages and it said it can&#39;t find it ???\"><\/a><\/p>\n\n<p>So I think that there are maybe some think wrong about the <code>PATH<\/code> of those packages so that I can't run the program, but I don't know how to fix that and I also can't find anything related to my problem on the internet ... so I really need you guys help here.<\/p>\n\n<p>Keep in mind that the <strong>user<\/strong> that I used when I do <strong>remote control<\/strong> and the <strong>user<\/strong> when I use <strong>SSH connection<\/strong> is the <strong>same user<\/strong>.<\/p>\n\n<p>Thanks you so much for your help.<\/p>\n","answers":[]}
+{"QuestionId":53020929,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-27T10:24:16.393","AcceptedAnswerId":null,"OwnerUserId":10566739.0,"Title":"ValueError: `decode_predictions` expects a batch of predictions (i.e. a 2D array of shape (samples, 1000)). Found array with shape: (1, 3)","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/RxpUn.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>I've been Google for a long time, but they don't ask me the same questions as I do. They use a ready-made model. I'm a self-trained model, but I'm puzzled by the same mistakes.<\/p>\n","answers":[{"AnswerId":"53052091","CreationDate":"2018-10-29T18:56:51.360","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that the <a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/6df456f59069c5daa752cbeef2b4f3a63c8874c1\/keras_applications\/imagenet_utils.py#L198\" rel=\"nofollow noreferrer\"><code>decode_predictions<\/code><\/a>, expect that the number of classes to be 1000 (i.e. the number of classes in ImageNet dataset). Whereas it seems your custom model have only 3 output classes:<\/p>\n\n<blockquote>\n  <p><code>decode_predictions<\/code> expects a batch of predictions (i.e. a 2D array\n  of shape <code>(samples, 1000)<\/code>). Found array with shape: <code>(1, 3)<\/code>.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53021010,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-27T10:34:07.847","AcceptedAnswerId":null,"OwnerUserId":9479875.0,"Title":"While trying to train the model I am getting ValueError: setting an array element with a sequence","Body":"<p>Here is the code.<\/p>\n\n<p>While trying to run this file, I am getting ValueError: setting an array element with a sequence.<\/p>\n\n<pre><code>ValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>I am using tensorflow 1.9<\/p>\n\n<p><a href=\"https:\/\/github.com\/satyenrajpal\/Concrete-Crack-Detection\/blob\/master\/Train_CD.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/satyenrajpal\/Concrete-Crack-Detection\/blob\/master\/Train_CD.py<\/a><\/p>\n\n<p>Here is the code I am trying to run.<\/p>\n","answers":[{"AnswerId":"53022006","CreationDate":"2018-10-27T12:43:52.657","ParentId":null,"OwnerUserId":"4101431","Title":null,"Body":"<p>AFAIK this problem usually arises from having a one or more entries in your array that is not the same length as the rest, so essentially the array isn't seen as a valid matrix, but rather a 'sequence' of arrays that differ in length.<\/p>\n\n<p>Try debugging the code and making sure that all the elements of your array are the same shape, I haven't looked through all your code, so I'm not sure if you intentionally want to feed sequences to your model, but in that case you should look into padding your data before inputting it, so that all the elements in your array are of the same length. <\/p>\n\n<p>There are ample options for doing something like this, can't list them all off the top of my head, but one that comes to mind, that I've personally used before is <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.6\/api_docs\/python\/tf\/keras\/preprocessing\/sequence\/pad_sequences\" rel=\"nofollow noreferrer\">tf.keras.preprocessing.sequence.pad_sequences<\/a> (that links to API v1.6, but it's also available in v1.11, just checked now)<\/p>\n"}]}
+{"QuestionId":53022148,"AnswerCount":1,"Tags":"<python><keras><concatenation><reshape>","CreationDate":"2018-10-27T13:03:20.337","AcceptedAnswerId":53052797.0,"OwnerUserId":3968093.0,"Title":"Concatenate identity matrix to each vector","Body":"<p>I want to modify my input by adding several different suffixes to the input vectors. For example, if the (single) input is <code>[1, 5, 9, 3]<\/code> I want to create three vectors (stored as matrix) like this:<\/p>\n\n<pre><code>[[1, 5, 9, 3, 1, 0, 0],\n [1, 5, 9, 3, 0, 1, 0],\n [1, 5, 9, 3, 0, 0, 1]]\n<\/code><\/pre>\n\n<p>Of course, this is just one observation so the input to the model is <code>(None, 4)<\/code> in this case. The simple way is to prepare the input data somewhere else (numpy most probably) and adjust the shape of input accordingly. That I can do but I would prefer doing it inside TensorFlow\/Keras.<\/p>\n\n<p>I have isolated the problem into this code:<\/p>\n\n<pre><code>import keras.backend as K\nfrom keras import Input, Model\nfrom keras.layers import Lambda\n\n\ndef build_model(dim_input: int, dim_eye: int):\n    input = Input((dim_input,))\n    concat = Lambda(lambda x: concat_eye(x, dim_input, dim_eye))(input)\n    return Model(inputs=[input], outputs=[concat])\n\n\ndef concat_eye(x, dim_input, dim_eye):\n    x = K.reshape(x, (-1, 1, dim_input))\n    x = K.repeat_elements(x, dim_eye, axis=1)\n    eye = K.expand_dims(K.eye(dim_eye), axis=0)\n    eye = K.tile(eye, (-1, 1, 1))\n    out = K.concatenate([x, eye], axis=2)\n    return out\n\n\ndef main():\n    import numpy as np\n\n    n = 100\n    dim_input = 20\n    dim_eye = 3\n\n    model = build_model(dim_input, dim_eye)\n    model.compile(optimizer='sgd', loss='mean_squared_error')\n\n    x_train = np.zeros((n, dim_input))\n    y_train = np.zeros((n, dim_eye, dim_eye + dim_input))\n    model.fit(x_train, y_train)\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>The problem seems to be in the <code>-1<\/code> in <code>shape<\/code> argument in <code>tile<\/code> function. I tried to replace it with <code>1<\/code> and <code>None<\/code>. Each has its own error:<\/p>\n\n<ul>\n<li><p><code>-1<\/code>: error during <code>model.fit<\/code><\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Expected multiples[0] &gt;= 0, but got -1\n<\/code><\/pre><\/li>\n<li><p><code>1<\/code>: error duting <code>model.fit<\/code><\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: ConcatOp : Dimensions of inputs should match: shape[0] = [32,3,20] vs. shape[1] = [1,3,3]\n<\/code><\/pre><\/li>\n<li><p><code>None<\/code>: error during <code>build_model<\/code>:<\/p>\n\n<pre><code>Failed to convert object of type &lt;class 'tuple'&gt; to Tensor. Contents: (None, 1, 1). Consider casting elements to a supported type.\n<\/code><\/pre><\/li>\n<\/ul>\n","answers":[{"AnswerId":"53052797","CreationDate":"2018-10-29T19:50:07.330","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to use <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/aa0f12ce9eacd70932466d248fde746e0b0dfa0d\/keras\/backend\/tensorflow_backend.py#L538\" rel=\"nofollow noreferrer\"><code>K.shape()<\/code><\/a> instead to get the <strong>symbolic shape<\/strong> of input tensor as the batch size is <code>None<\/code> and therefore passing <code>K.int_shape(x)[0]<\/code> or <code>None<\/code> or <code>-1<\/code> as a part of the second argument of <code>K.tile()<\/code> would not work:<\/p>\n\n<pre><code>eye = K.tile(eye, (K.shape(x)[0], 1, 1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53022405,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><normalization>","CreationDate":"2018-10-27T13:35:06.700","AcceptedAnswerId":53022536.0,"OwnerUserId":7490161.0,"Title":"Can anyone explain tf.random_normal and tf.truncated_normal shape argument?","Body":"<p>I am working with tensorflow and until and unless the arguments were 2 dimensional, I didn't faced any issue but now I came across a syntax <\/p>\n\n<pre><code>tf.Variable(tf.random_normal([5, 5, 32, 64])\n\ntf.Variable(tf.random_normal([7*7*64, 1024]))\n<\/code><\/pre>\n\n<p>and one more<\/p>\n\n<pre><code>tf.Variable(tf.random_normal([1024])\n<\/code><\/pre>\n\n<p>And similarly in truncated_normal also.\nCan anyone explain the working of these lines and how this is implemented behind the code? It would be very helpful.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53022536","CreationDate":"2018-10-27T13:50:39.497","ParentId":null,"OwnerUserId":"4101431","Title":null,"Body":"<p>Both of these functions sample for either a normal distribution or a truncated ( cut at the tips) normal distribution respectively, the shape parameter is used to specify the shape of the resulting tensor ie. how many samples you are going to take from the respective distribution and what shape will they be in. <\/p>\n\n<p>Anything up to 3 dimensions is easy to picture,<\/p>\n\n<p><code>tf.Variable(tf.random_normal([7*7*64, 1024]))<\/code> creates a 2D-tensor of shape 3136x1024 of randomly sampled values from a normal distribution.<\/p>\n\n<p>Similarly, <code>tf.Variable(tf.random_normal([1024]))<\/code> will create a vector of size 1024 (or a 2D-tensor of size 1x1024) <\/p>\n\n<p>Finally, <code>tf.Variable(tf.random_normal([5, 5, 32, 64]))<\/code> is a bit harder to picture mentally, this results in a 4D-Tensor of size 5x5x32x64. <\/p>\n\n<p>This could represent a batch of 5 3-D tensors, where the first dimension (the first '5' in this case) will represent the <em>batch size<\/em> of some input data. <\/p>\n\n<p>Not sure how much detail you're looking for, I'm afraid that's as far as my knowledge goes, hope this helps. <\/p>\n"}]}
+{"QuestionId":53022752,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp>","CreationDate":"2018-10-27T14:15:23.163","AcceptedAnswerId":null,"OwnerUserId":4858848.0,"Title":"Dimension Problem in Keras Multilabel Classification with Word Embeddings","Body":"<p>I am currently solving an exercise which involves reading in TED talks, labelling them according to the topics they are about, and training a Feed Forward NN in Keras that can label new talks accordingly, using pre-trained word embeddings.<\/p>\n\n<p>Depending on what the talk is about (technology, education or design or multiple of those topics), it can have one of the following labels:<\/p>\n\n<pre><code>labels_dict = {\n'txx': 0, 'xex': 1, 'xxd': 2, 'tex': 3, 'txd': 4, 'xed': 5, 'ted': 6, 'xxx': 7\n}\n<\/code><\/pre>\n\n<p>I load the talks like this:<\/p>\n\n<pre><code>def load_talks(path):\ntree = et.parse(path)\nroot = tree.getroot()\nfor file in root:\n    label = ''\n    keywords = file.find('head').find('keywords').text.lower()\n    if 'technology' in keywords:\n        label += 't'\n    else:\n        label += 'x'\n\n    if 'education' in keywords:\n        label += 'e'\n    else:\n        label += 'x'\n\n    if 'design' in keywords:\n        label += 'd'\n    else:\n        label += 'x'\n\n    talk = file.find('content').text\n    talk = process_text(talk)\n    texts.append(talk)\n    labels.append(labels_dict[label])\n<\/code><\/pre>\n\n<p>I then calculate TF-IDF scores for the tokens in the texts:<\/p>\n\n<pre><code>tf_idf_vect = TfidfVectorizer()\ntf_idf_vect.fit_transform(texts)\ntf_idf_vectorizer_tokens = tf_idf_vect.get_feature_names()\n<\/code><\/pre>\n\n<p>Then I use a tokenizer to assign the tokens in the texts to indexes:<\/p>\n\n<pre><code>t = Tokenizer()\nt.fit_on_texts(texts)\n\nvocab_size = len(t.word_index) + 1\nencoded_texts = t.texts_to_sequences(texts)\n\nprint('Padding the docs')\n# pad documents to a max length of 4 words\nmax_length = max(len(d) for d in encoded_texts)\npadded_docs = pad_sequences(encoded_texts, maxlen=max_length, padding='post')\n<\/code><\/pre>\n\n<p>Next, I compute the embedding matrix:<\/p>\n\n<pre><code>def compute_embedding_matrix(word_index, embedding_dim):\nembedding_matrix = np.zeros((len(word_index) + 1, embedding_dim))\nfor word, i in word_index.items():\n    embedding_vector = word_embeddings.get(word)\n    if embedding_vector is not None and get_tf_idf_score(word) &gt; TF_IDF_THRESHOLD:\n        # words not found in embedding index and with a too low tf-idf score will be all-zeros.\n        embedding_matrix[i] = embedding_vector\nreturn embedding_matrix\n\nembedding_dim = load_word_embeddings('word_embeddings\/de\/de.tsv') + 1\nembedding_matrix = compute_embedding_matrix(t.word_index, embedding_dim)\n<\/code><\/pre>\n\n<p>I then prepare the labels and split the data in training and testing:<\/p>\n\n<pre><code>labels = to_categorical(np.array(labels))\n\nX_train, X_test, y_train, y_test = train_test_split(padded_docs, labels, test_size=0.1, random_state=0)\n<\/code><\/pre>\n\n<p>The following prints output this:<\/p>\n\n<pre><code>print(X_train.shape)\nprint(X_test.shape)\nprint(y_train.shape)\nprint(y_test.shape)\n\n(1647, 6204)\n(184, 6204)\n(1647, 8)\n(184, 8)\n<\/code><\/pre>\n\n<p>I then prepare my model like this:<\/p>\n\n<pre><code>e = Embedding(input_dim=vocab_size,\n              weights=[embedding_matrix],\n              input_length=max_length,\n              output_dim=embedding_dim,\n              trainable=False)\n\nprint('Preparing the network')\nmodel = models.Sequential()\nmodel.add(e)\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(units=100, input_dim=embedding_dim, activation='relu'))\nmodel.add(layers.Dense(len(labels), activation='softmax'))\n# compile the model\nmodel.compile(loss='binary_crossentropy', metrics=['acc'])\n# summarize the model\nprint(model.summary())\n# fit the model\nprint('Fitting the model')\nmodel.fit(X_train, y_train, epochs=10, verbose=0, batch_size=500)\n# evaluate the model\nloss, accuracy = model.evaluate(X_test, y_test, verbose=0)\nprint('Accuracy: %f' % (accuracy*100))\n<\/code><\/pre>\n\n<p>This outputs the following error:<\/p>\n\n<pre><code>embedding_1 (Embedding)      (None, 6204, 301)         47951106  \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 1867404)           0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 100)               186740500 \n_________________________________________________________________\ndense_2 (Dense)              (None, 1831)              184931    \n=================================================================\nTotal params: 234,876,537\nTrainable params: 186,925,431\nNon-trainable params: 47,951,106\n_________________________________________________________________\nNone\nFitting the model\n    batch_size=batch_size)\n  File \"\/Users\/tim\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 789, in _standardize_user_data\n    exception_prefix='target')\n  File \"\/Users\/tim\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\", line 138, in standardize_input_data\n    str(data_shape))\nValueError: Error when checking target: expected dense_2 to have shape (1831,) but got array with shape (8,)\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>Can someone point me in the right direction about how to go about fitting the dimensions of this model?<\/p>\n","answers":[{"AnswerId":"53023114","CreationDate":"2018-10-27T14:53:08.150","ParentId":null,"OwnerUserId":"4858848","Title":null,"Body":"<p>I found the issue:<\/p>\n\n<p>The last Dense layer should have 8 units, as I have 8 labels.<\/p>\n"}]}
+{"QuestionId":53022933,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-27T14:36:03.080","AcceptedAnswerId":null,"OwnerUserId":10547221.0,"Title":"Prediction on model exported model using tf.estimator on a base64 input in JSON","Body":"<p>i want to built a rest api for model prediction my model accepts base64 string of image in json file as i have tested over google cloud ML. and it is not accepting predict request from android app so i am trying to use rest api. from python script my prediction is giving me so many errors by just testing locally without api yet\n here is my export model code<\/p>\n\n<pre><code>import os\nimport shutil\nimport tensorflow as tf\nfrom tensorflow import keras\n\n\n\nHEIGHT = 48\nWIDTH = 48\nCHANNELS = 1\nversion = 'v1'\n\n\nh5_model_path = os.path.join('model_4layer_2_2_pool  yes.h5')\ntf_model_path = os.path.join('D:\/university\/working\/trying\/Facial- \nExpression-Recognition-master\/tryMore')\nestimator = keras.estimator.model_to_estimator(\nkeras_model_path=h5_model_path,\nmodel_dir=tf_model_path)\n\n\ndef image_preprocessing(image):\n\n image = tf.expand_dims(image, 0)\n image = tf.image.resize_bilinear(image, [HEIGHT, WIDTH], \n align_corners=False)\n image = tf.squeeze(image, axis=[0])\n image = tf.cast(image, dtype=tf.uint8)\n return image\n<\/code><\/pre>\n\n#\n\n<h1>IMAGE AS BASE64 BYTES<\/h1>\n\n<pre><code>def serving_input_receiver_fn():\n  def prepare_image(image_str_tensor):\n    image = tf.image.decode_jpeg(image_str_tensor, channels=CHANNELS)\n    return image_preprocessing(image)\n\n input_ph = tf.placeholder(tf.string, shape=[None])\n images_tensor = tf.map_fn(\n    prepare_image, input_ph, back_prop=False, dtype=tf.uint8)\n images_tensor = tf.image.convert_image_dtype(images_tensor, \n dtype=tf.float32)\n\n return tf.estimator.export.ServingInputReceiver(\n    {'conv2d_1_input': images_tensor},\n    {'image_bytes': input_ph})\n\n export_path = os.path.join('models\\json_b64', version)\n if os.path.exists(export_path):  # clean up old exports with this \n version\n shutil.rmtree(export_path)\n estimator.export_savedmodel(\n export_path,\n serving_input_receiver_fn=serving_input_receiver_fn)\n<\/code><\/pre>\n\n<p><strong>and my code for prediction is as follows<\/strong><\/p>\n\n<pre><code>import tensorflow as tf \nimport pickle\nimport os\nimport imgJSON\n\ndir_path = os.path.dirname(__file__)\nexported_path= os.path.join(dir_path,  \"models\/json_b64\/v1\/1539157474\")\n\nmodel_input=imgJSON.img_bytes\nglobal data\ndef main():\nwith tf.Session() as sess:\n\n Model=tf.saved_model.loader.load(sess, \n               [tf.saved_model.tag_constants.SERVING], exported_path)\n\n predictor= tf.contrib.predictor.from_saved_model(exported_path)\n\nimport json \nwith open(\"D:\/university\/working\/trying\/Facial-Expression-Recognition- \n  master\/tryMore\/test_json_b64.json\") as f:\n  data = json.loads(f.read())\n  print(data[\"image_bytes\"])\n\n\noutput_dict= predictor({\"image_bytes\":[model_input]})\nprint(\" prediction is \" , output_dict['scores'])\n\n\nif __name__ == \"__main__\":\n   main()\n<\/code><\/pre>\n\n<p>my error message is as follows:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Expected image (JPEG, PNG, or GIF), got unknown format starting with '\/9j\/4AAQSkZJRgAB'\n [[Node: map\/while\/DecodeJpeg = DecodeJpeg[acceptable_fraction=1, channels=1, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](map\/while\/TensorArrayReadV3)]]\n<\/code><\/pre>\n\n<p>how to predict passing base64 string to my respective exported model? <\/p>\n","answers":[]}
+{"QuestionId":53023399,"AnswerCount":1,"Tags":"<keras><generative-adversarial-network>","CreationDate":"2018-10-27T15:21:57.750","AcceptedAnswerId":null,"OwnerUserId":10516607.0,"Title":"Keras GAN with multiple output from generator and discriminator respectively","Body":"<p>I am using Keras to model GAN, and I need to combine two losses as I have two outputs. One output is from Discriminator, which is denoted as \"label\" in the following code, and another is from Generator, which is denoted as \"Bloss\". So is it possible to train the combined model of GAN (combine Generator and Discriminator) with two outputs from G and D respectively? <\/p>\n\n<pre><code>    input = Input(shape=self.input_shape)\n    output_G, Bloss = self.G(input)\n\n    # For the combined model we will only train the generator\n    self.D.trainable = False\n\n    label = self.D(output_G)\n\n    self.combined = Model(inputs=input,\n                          outputs=[label, Bloss])\n    self.combined.compile(loss=['categorical_crossentropy', B_loss],\n                          optimizer='RMSprop',\n                          loss_weights=[1,0.01])\n...\ndef B_loss(y_true, y_pred):\n    return K.mean(y_pred - y_true, axis=-1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53026673","CreationDate":"2018-10-27T22:03:32.723","ParentId":null,"OwnerUserId":"6431410","Title":null,"Body":"<p>I don't see why not, as long as you create a proper y that combines both the label (D output) and Bloss (partial G output). <\/p>\n"}]}
+{"QuestionId":53023727,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-27T15:57:12.340","AcceptedAnswerId":53065877.0,"OwnerUserId":6331998.0,"Title":"How to efficiently extract all slices of given length using tensorflow","Body":"<p>I am trying to extract all slices of length 4 along 0th axis of a 2-dim tensor. So far I can do it mixing pure Python with tensorflow.<\/p>\n\n<pre><code>r = test.shape[0] # test should be a tensor\nn = 4\na_list = list(range(r))\nthe_list = np.array([a_list[slice(i, i+n)] for i in range(r - n+1)])\ntest_stacked = tf.stack(tf.gather(test, the_list))\n<\/code><\/pre>\n\n<p>What would be an efficient way of doing that without using pure Python? Note that the \"test\" array is actually supposed to be a tensor, thus its shape isn't known before I execute the first part of the graph.<\/p>\n\n<p>A full vanilla example:<\/p>\n\n<pre><code>array = np.array([[0, 1],[1, 2],[2, 3],[3, 4],[4, 5],[5, 6]])\narray.shape # (6,2)\n\nr = array.shape[0]\nn = 4\na_list = list(range(r))\nthe_list = np.array([a_list[slice(i, i+n)] for i in range(r - n+1)])\n\nresult = array[the_list] # all possible slices of length 4 of the array along 0th axis\nresult.shape # (3, 4, 2)\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>[[[0 1]\n  [1 2]\n  [2 3]\n  [3 4]]\n\n [[1 2]\n  [2 3]\n  [3 4]\n  [4 5]]\n\n [[2 3]\n  [3 4]\n  [4 5]\n  [5 6]]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53065877","CreationDate":"2018-10-30T13:52:37.823","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>You may want to try the more general <code>tf.extract_image_patches<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant([[0, 1],[1, 2],[2, 3],[3, 4],[4, 5],[5, 6]])\n# tf.extract_image_patches requires a [batch, in_rows, in_cols, depth] tensor\na = a[None, :, :, None]\nb = tf.extract_image_patches(a,\n  ksizes=[1, 4, 2, 1],\n  strides=[1, 1, 1, 1],\n  rates=[1, 1, 1, 1],\n  padding='VALID')\nb = tf.reshape(tf.squeeze(b), [-1, 4, 2])\n\nsess = tf.InteractiveSession()\nprint(b.eval())\n<\/code><\/pre>\n"},{"AnswerId":"53054003","CreationDate":"2018-10-29T21:20:35.740","ParentId":null,"OwnerUserId":"7078425","Title":null,"Body":"<p>I believe <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">gather_nd<\/a> is what you are looking for.<\/p>\n\n<pre><code># a is a tensor of size (6, 2)\n\ndef get_indices(l, d):\n    return [[[j] for j in range(i, i + d)] for i in range(l - d + 1)]\n\nb = tf.gather_nd(a, get_indices(6, 4))\n# b is a tensor of shape (3, 4, 2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53023991,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-27T16:26:49.453","AcceptedAnswerId":null,"OwnerUserId":7916098.0,"Title":"Keras Gates in Cell LSTM","Body":"<p>I tried using the lstm layer in keras. I need manipulate the gates at forget gate, input gate and output gate. I don't find a local to manipulate the variables of the gates. Anybody know if is possible to do in keras layers?<\/p>\n","answers":[]}
+{"QuestionId":53024012,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-10-27T16:29:30.663","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Cannot compare tensor to scalar anymore in new pytorch","Body":"<p>I was working with a self compiled version of pytorch 0.4.0 for a while and hence my code has the following line for multidimensional tensors:<\/p>\n\n<pre><code>myTensor3 = torch.arange(torch.numel(myTensor0) # this is a 1 dimensional Tensor\noutTensor = myTensor3[(myTensor1 == 1) &amp; (myTensor2 != 1)] # myTensor1 and myTensor2 are multidimensional\n<\/code><\/pre>\n\n<p>Whenever I try to switch to another (newer) version of pytorch, the above line throws the following error: <\/p>\n\n<pre><code>IndexError: too many indices for tensor of dimension 1\n<\/code><\/pre>\n\n<p>I assume in the newer version of pytorch the index array returned by the comparison operation has dimension 1, whereas before it had the same size as myTensor3.\nHow can I change back to the old behaviour in the newer version of pytorch?<\/p>\n\n<p><strong>Edit:<\/strong>\nA simple fix was to linearize myTensor1 and 2:<\/p>\n\n<pre><code>outTensor = myTensor3[(myTensor1.view(-1) == 1) &amp; (myTensor2.view(-1) != 1)]\n<\/code><\/pre>\n\n<p>for some strange reason this was already working in the self compiled pytorch version<\/p>\n","answers":[]}
+{"QuestionId":53024517,"AnswerCount":1,"Tags":"<neural-network><keras><lstm><recurrent-neural-network>","CreationDate":"2018-10-27T17:28:27.370","AcceptedAnswerId":53041748.0,"OwnerUserId":5117173.0,"Title":"LSTM neural network with two sources of data","Body":"<p>I have the following configuration: One lstm network that receives a text with n-grams with size 2. Below a simple schematic:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xdSgX.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xdSgX.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After some tests, I noticed that for some classes I have an significant incrise on accuracy when I use ngrams with size 3. Now I want to train a new LSTM neural network with both ngram sizes at same time, like the following schematic:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6lRC1.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6lRC1.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>How can I provide the data and build this model, using keras to perform this task?<\/p>\n","answers":[{"AnswerId":"53041748","CreationDate":"2018-10-29T08:48:41.360","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>I assume you already have a function to split words into n-grams, as you already have the 2-grams and 3-grams model working? Therefor I just construct a one-sample example of the word \"cool\" for a working example. I had to use embedding for my example, as an LSTM layer with 26^3=17576 nodes was a little too much for my computer to handle. I expect you did the same in your 3-grams code?<\/p>\n\n<p>Below is a complete working example:<\/p>\n\n<pre><code>from tensorflow.keras.layers import Input, Embedding, LSTM, Dense, concatenate\nfrom tensorflow.keras.models import Model\nimport numpy as np\n\n# c-&gt;2 o-&gt;14 o-&gt;14 l-&gt;11\nnp_2_gram_in = np.array([[26*2+14,26*14+14,26*14+11]])#co,oo,ol\nnp_3_gram_in = np.array([[26**2*2+26*14+14,26**2*14+26*14+26*11]])#coo,ool\n\nnp_output = np.array([[1]])\n\noutput_shape=1\nlstm_2_gram_embedding = 128\nlstm_3_gram_embedding = 192\n\ninputs_2_gram = Input(shape=(None,))\nem_input_2_gram = Embedding(output_dim=lstm_2_gram_embedding, input_dim=26**2)(inputs_2_gram)\nlstm_2_gram = LSTM(lstm_2_gram_embedding)(em_input_2_gram)\ninputs_3_gram = Input(shape=(None,))\nem_input_3_gram = Embedding(output_dim=lstm_3_gram_embedding, input_dim=26**3)(inputs_3_gram)\nlstm_3_gram = LSTM(lstm_3_gram_embedding)(em_input_3_gram)\nconcat = concatenate([lstm_2_gram, lstm_3_gram])\noutput = Dense(output_shape,activation='sigmoid')(concat)\n\nmodel = Model(inputs=[inputs_2_gram, inputs_3_gram], outputs=[output])\nmodel.compile(optimizer='adam', loss='binary_crossentropy')\n\nmodel.fit([np_2_gram_in, np_3_gram_in], [np_output], epochs=5)\nmodel.predict([np_2_gram_in,np_3_gram_in])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53024766,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-27T17:57:45.117","AcceptedAnswerId":53025535.0,"OwnerUserId":2666182.0,"Title":"How to make flags as necessary in tensorflow","Body":"<p>I am using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/app\/flags\" rel=\"nofollow noreferrer\"><code>tf.app.flags<\/code><\/a> to parse command line arguments in my code. Is there any way I can make some flags as necessary (as in <code>argparse<\/code>)? Using default value as <code>None<\/code> doesn't seem to help<\/p>\n","answers":[{"AnswerId":"53025535","CreationDate":"2018-10-27T19:27:04.913","ParentId":null,"OwnerUserId":"2666182","Title":null,"Body":"<p>Thanks to @etarion 's comment, it seems the correct method to use is <code>flags.mark_flag_as_required<\/code>. Note that this won't work if the default value is not <code>None<\/code>.<\/p>\n"}]}
+{"QuestionId":53026292,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-27T21:08:44.333","AcceptedAnswerId":null,"OwnerUserId":1019129.0,"Title":"Training lstm model?","Body":"<p>trying to train LSTM model with (5,xxx) data stream f.e. <\/p>\n\n<pre><code>(5, 17516)\narray([[ 3820.,  2873.,  2369., ..., 18865., 16893., 14242.],\n   [ 4656.,  3820.,  2873., ..., 19967., 18865., 16893.],\n   [ 6210.,  4656.,  3820., ..., 20223., 19967., 18865.],\n   [ 8127.,  6210.,  4656., ..., 20319., 20223., 19967.],\n   [10844.,  8127.,  6210., ..., 17246., 20319., 20223.]])\n<\/code><\/pre>\n\n<p>here is the model :<\/p>\n\n<pre><code>def lstm_model(self, window=5):\n    self.model = Sequential()\n    self.model.add(LSTM(4, input_shape=( window, 1)))\n    self.model.add(Dense(1))\n    self.model.compile(loss='mean_squared_error', optimizer='adam')\n    return self.model\n<\/code><\/pre>\n\n<p>here is the fit :<\/p>\n\n<pre><code>self.history = self.model.fit(\n        windowed_data , self.data.data,\n        validation_split=0.2, nb_epoch=55, batch_size=10, verbose=1)\n<\/code><\/pre>\n\n<p>here is the error I'm getting :<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected lstm_6_input to have 3 dimensions, but got array with shape (5, 17516)<\/p>\n<\/blockquote>\n\n<p>What I'm doing wrong ?<\/p>\n\n<hr>\n\n<p>This seems to solve it.<\/p>\n\n<pre><code>w.reshape(w.shape[0], w.shape[1],1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53026495","CreationDate":"2018-10-27T21:37:00.783","ParentId":null,"OwnerUserId":"6273108","Title":null,"Body":"<p>according to <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">keras docs<\/a> the input data should be a 3d tensor i.e. (nb_samples, timesteps, input_dim). <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">this<\/a> is a good tutorial on how to reshape your data for lstm models.<\/p>\n"}]}
+{"QuestionId":53026480,"AnswerCount":1,"Tags":"<docker><tensorflow><localhost><tensorboard>","CreationDate":"2018-10-27T21:34:42.253","AcceptedAnswerId":null,"OwnerUserId":10083000.0,"Title":"WhatIf Google Tensorboard Tutorial confusion","Body":"<p>I am trying to use WhatIf to interpret a model and followed along this here: <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/tree\/master\/tensorboard\/plugins\/interactive_inference#what-do-i-need-to-use-it\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/tree\/master\/tensorboard\/plugins\/interactive_inference#what-do-i-need-to-use-it<\/a><\/p>\n\n<p>In particular I followed this tutorial: <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/tensorboard\/plugins\/interactive_inference\/WIT_from_scratch.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/tensorboard\/plugins\/interactive_inference\/WIT_from_scratch.ipynb<\/a> <\/p>\n\n<p>However parts of it are poorly written. I finally got my docker to spin up a localhost at port 8005 with this command: <code>docker run -p 8500:8500 --mount type=bind,source=\"$(pwd)\/mod\nels\/iris\",target=\/models\/iris -e MODEL_NAME=iris -t tensorflow\n\/serving<\/code> which output this: <code>I tensorflow_serving\/model_servers\/server.cc:301] Exporting HTTP\/REST API at:localhost:8501 ...\n[evhttp_server.cc : 235] RAW: Entering the event loop ...<\/code><\/p>\n\n<p>I then in my Anaconda command line I ran <code>tensorboard --logdir .<\/code> as the tutorial told me to which outputted this: <code>TensorBoard 1.9.0 at http:\/\/DESKTOP-[serial_code_for_my_computer]:6006 (Press CTRL+C to quit)<\/code><\/p>\n\n<p>The tutorial tells me to \"Navigate to <a href=\"http:\/\/localhost:6006\/#whatif&amp;inferenceAddress=localhost%3A8500&amp;modelName=iris\" rel=\"nofollow noreferrer\">http:\/\/localhost:6006\/#whatif&amp;inferenceAddress=localhost%3A8500&amp;modelName=iris<\/a>\"<\/p>\n\n<p>However nothing loads when I do this. I suspect there is some disconnect between my Docker which is hosting at localhost:8500 and my tensorboard which is at DESKTOP-[computer_serial]:6006? This last URL loads an empty tensorboard with nothing in it but even when I add #whatif&amp;inferenceAddress=localhost%3A8500&amp;modelName=iris to this DESKTOP URL it fails to load anything. <\/p>\n\n<p>Thank you for any help that can be provided. I look forward to being able to use WhatIf!<\/p>\n","answers":[{"AnswerId":"53854983","CreationDate":"2018-12-19T16:03:49.307","ParentId":null,"OwnerUserId":"4490705","Title":null,"Body":"<p>In my case, the model name you enter to both of docker command and localhost address should've been matched. For me, the tensorboard loaded successfully the instances but not other functionalities, and made requestnetworkerror. Hope it helps.<\/p>\n"}]}
+{"QuestionId":53027019,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-27T23:01:58.890","AcceptedAnswerId":null,"OwnerUserId":9655448.0,"Title":"keras Deep Learning Slowness - Example IMDB dataset from Deep Learning with Python Chollet","Body":"<p>I am having an issue with keras leading to my processor seemingly getting bogged down while working through examples.<\/p>\n\n<p>In the IMDB data set for instance (exercise 3.4.1 in Deep Learning with Python by Chollet if anyone knows the book), running the script:<\/p>\n\n<pre><code>import keras\nfrom keras.datasets import imdb    \n(train_data, train_labels), (test_data, test_labels) = \n    imdb.load_data(num_words=10000)\n<\/code><\/pre>\n\n<p>Produces an output looking something like:<\/p>\n\n<pre><code>[=====&gt;...] - ETA: 59s\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u2713\u271316105472\/17464789\n<\/code><\/pre>\n\n<p>That updates increasingly slowly as the numbers get larger and move toward completion.<\/p>\n\n<p>I'm assuming my installation of keras\/Tensorflow\/CUDA\/cuDNN is to blame, but curious if you know of anything obvious that would solve the issue.<\/p>\n\n<p>Running Ubuntu Linux, NVIDIA GTX 1080, Keras\/Tensorflow (GPU)\/CUDA,cuDNN (maybe, assuming I installed everything correctly which is probably not accurate).<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53027243,"AnswerCount":2,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-10-27T23:47:56.187","AcceptedAnswerId":null,"OwnerUserId":10568864.0,"Title":"Tensorflow.js Please make sure the operations that use variables are inside the function f passed to minimize()","Body":"<p>I have started to use javascript and TensorFlow.js to work on some machine learning projects, I am working on creating a linear regression model, however, I cannot figure out what is causing this error<\/p>\n\n<p><strong>Can not find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize().<\/strong>\"<\/p>\n\n<p>I have created two Tensors <\/p>\n\n<pre><code>globalTensorXs = tf.tensor2d(globaArrayXs); \/\/input data\nglobalTensorYs =  tf.tensor1d(globaArrayYs); \/\/y output\n<\/code><\/pre>\n\n<p>I have created the coefficients\/weights as below as an array of tf scalars.<\/p>\n\n<pre><code>function createWeights(_numWeights)\n{       \n  for ( var x = 0; x &lt; _numWeights; x++)\n  {\n  globalWeightsTensorArr.push(tf.variable(tf.scalar(Math.random())));  \n }      \n} \n<\/code><\/pre>\n\n<p>There is A training function which I pass the x and y Tensors into, it is the call to the optimise.minimize that causes the issue. it does not detect the variable for training, which are stored in globalWeightsTensorArr<\/p>\n\n<pre><code>async function train(xsTensor, ysTensor, numIterations)\n{\n\n    \/*\/\/\/\/\/\/OPTIMISER.MINIMISE\/\/\/\/\/\/\/\/\/\/\/\/\/\n    Minimize takes a function that does two things:\n    It predicts y values for all the x values using the predict \n    model function. \n    It returns the mean squared error loss for those predictions \n    using the loss function. Minimize then automatically adjusts any \n    Variables used by thi predict\/loss function in order to minimize \n    the return value (our loss), in this case the variables are in \n    \"globalWeightsTensorArr\" which contains the coefficient values \n     to be altered by the modeld during \"numIterations\" iterations of \n     SGD.\n*\/\n\n  for (let iter = 0; iter &lt; numIterations; iter++) \n  {     \n      optimiser.minimize(function ()\n      {\n        return loss(predict(xsTensor), ysTensor);\n       }, globalWeightsTensorArr);          \n  }\n\n}\n<\/code><\/pre>\n\n<p>\/\/ the predict and loss function are here...<\/p>\n\n<p>\/\/The following code constructs a predict function that takes inputs(X's) \/\/and returns prediction Y: it represents our 'model'. Given an input \/\/'xs' it will try and * predict the appropriate output 'y'.<\/p>\n\n<pre><code>function predict(_Xs) \n{           \n    return tf.tidy(() =&gt; {  \n    for ( var x = 0; x &lt; 8; x++)\n        globalWeightsArr[x] = globalWeightsTensorArr[x].dataSync();\n        const weightTensor =  tf.tensor1d(globalWeightsArr); \n        const prediction = tf.dot(_Xs, weightTensor);\n        return prediction;\n });        \n}\n<\/code><\/pre>\n\n<p>\/\/The loss function takes the predictions from the predict function \n\/\/and the actual lables and adjusts the weights\n\/\/the weights are considered to be any tensor variable that impact the \/\/function We can define a MSE loss function in TensorFlow.js as follows:<\/p>\n\n<pre><code>function loss(_predictedTensor, _labels) \n{\nconst meanSquareError =_predictedTensor.sub(_labels).square().mean();\nreturn meanSquareError ;\n}\n<\/code><\/pre>\n\n<p>can anyone please help explain the problem? <\/p>\n\n<p>Regards\nAideen<\/p>\n","answers":[{"AnswerId":"53084476","CreationDate":"2018-10-31T13:25:50.803","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The issue is related to <code>tf.variable<\/code>. One needs to use <code>tf.variable<\/code> to create the weights that will be updated by the function created by <code>optimiser.minimize()<\/code>.<\/p>\n\n<p>A variable created by <code>tf.variable<\/code> is mutable contrary to <code>tf.tensor<\/code> that is immutable. As a result if one uses <code>tf.tenso<\/code>r to create the weights they could not be updated during the training<\/p>\n"},{"AnswerId":"53080496","CreationDate":"2018-10-31T09:47:32.483","ParentId":null,"OwnerUserId":"10568864","Title":null,"Body":"<p>We resolved the issue by changing the way the weights\/coefficients were created. Now minimize can detect the variables used by predict, and it adjusts them accordingly. later I will post the entire solution to codepen. still learning!<\/p>\n\n<pre><code>function createWeights(_numWeights) {\n    const randomTensor = tf.randomUniform([_numWeights, 1]);\n    globalWeightsTensorVar = tf.variable(randomTensor);\n}\n<\/code><\/pre>\n\n<p>here is the predict function used b<\/p>\n\n<pre><code>function predictLogical(_Xs) {\n\n    return tf.dot(_Xs, globalWeightsTensorVar);\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":53027698,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2018-10-28T01:39:33.880","AcceptedAnswerId":null,"OwnerUserId":5959374.0,"Title":"Error while trying to use tensorflow in jupyter notebook","Body":"<p>I am running jupyter notebook on a server which has tensorflow installed. The server runs ubuntu 16.04 and I activate the tensorflow environment by running<\/p>\n\n<pre><code>source \/usr\/local\/tensorflow\u2212python3\/bin\/activate\n<\/code><\/pre>\n\n<p>I am able to use tensorflow through the ipython3 shell but not when I am in the notebook file. I get an error saying that the tensorflow module is not found when I try to import it.<\/p>\n\n<p>I had thought that this might be due to a mismatch in the python versions being used, but the result of executing the following commands is exactly the same in both the ipython3 shell and the notebook ...<\/p>\n\n<pre><code>import sys\nsys.version\nsys.executable\nsys.path\n<\/code><\/pre>\n\n<p>It is a mystery to me why one of them is able to find the tf module while the other isn't. Can someone please help me figure this out?<\/p>\n","answers":[]}
+{"QuestionId":53027716,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T01:42:30.277","AcceptedAnswerId":null,"OwnerUserId":4207206.0,"Title":"differentiable histogram in tensorflow","Body":"<p>How can one get a differentiable histogram function in tensorflow. I want to use the histogram of images to calculate a particular loss function. But the histogram function provided by tensorflow, tf.histogram_fixed_width() are not differentiable. How can I go about this ?<\/p>\n","answers":[]}
+{"QuestionId":53027819,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><scikit-learn><deep-learning>","CreationDate":"2018-10-28T02:10:23.620","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"tf.losses.log_loss for multiply classes","Body":"<p>I have tested <code>tf.losses.log_loss<\/code> with binary classification problem comparing with numpy implementation and sklearn implementation, it looks good:<\/p>\n\n<pre><code># Array of samples\ny_pred_arr = np.array([0.6,0.3,0.9], np.float32)\ny_true_arr = np.array([1,0,1], np.float32)\n\nprint('y_pred_arr.shape', y_pred_arr.shape)\nprint('y_true_arr.shape', y_true_arr.shape)\n\n# 1 : Implemented in numpy\ndef cross_entropy_loss(y_pred_arr, y_true_arr, eps=1e-6):\n    p_arr = np.clip(y_pred_arr, eps, 1 - eps)\n    n_samples = p_arr.shape[0]\n    loss_arr = np.zeros((n_samples,), np.float32)\n    for i in range(n_samples):        \n        if y_true_arr[i] == 1:\n            loss_arr[i] = -np.log(p_arr[i])\n        else:\n            loss_arr[i] = -np.log(1 - p_arr[i])\n    loss = np.mean(loss_arr)\n    return loss\nloss = cross_entropy_loss(y_pred_arr, y_true_arr)\nprint(loss)\n\n# 2 : Implemented in sklearn\nfrom sklearn.metrics import log_loss\nloss = log_loss(y_true_arr, y_pred_arr, labels=[0,1], eps=1e-6)\nprint(loss)\n\n# 3 : Impemented in tensorflow using tf.losses.log_loss\ny_pred_tf = tf.convert_to_tensor(y_pred_arr, np.float32)\ny_true_tf = tf.convert_to_tensor(y_true_arr, np.float32)\nloss_tf = tf.losses.log_loss(labels=y_true_tf, predictions=y_pred_tf, epsilon=1e-6)\nwith tf.Session() as sess:\n    loss = sess.run(loss_tf)\n    print(loss)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>y_pred_arr.shape (3,)\ny_true_arr.shape (3,)\n0.32428703\n0.3242870296041171\n0.32428563\n<\/code><\/pre>\n\n<p>But for multiclass problem results are different:<\/p>\n\n<pre><code># Array of samples\ny_pred_arr = np.array([[0.4,0.3,0.3], [1.0,0.0,0.0], [0.0,0.0,1.0]], np.float32)\ny_true_arr = np.array([[1,0,0], [1,0,0], [0,0,1]], np.float32)\n\nprint('y_pred_arr.shape', y_pred_arr.shape)\nprint('y_true_arr.shape', y_true_arr.shape)\n\n# 1 : Implemented in numpy\ndef cross_entropy_loss(y_pred_arr, y_true_arr, eps=1e-6):\n    p_arr = np.clip(y_pred_arr, eps, 1 - eps)\n    n_samples = p_arr.shape[0]\n    n_classes = p_arr.shape[1]\n    loss_arr = np.zeros((n_samples,), np.float32)\n    for i in range(n_samples):\n        for c in range(n_classes):\n            loss_arr[i] += -y_true_arr[i][c] * np.log(p_arr[i][c])\n    loss = np.mean(loss_arr)\n    return loss\nloss = cross_entropy_loss(y_pred_arr, y_true_arr)\nprint(loss)\n\n# 2 : Implemented in sklearn\nfrom sklearn.metrics import log_loss\nloss = log_loss(np.argmax(y_true_arr,axis=1), y_pred_arr, labels=[0,1,2], eps=1e-6)\nprint(loss)\n\n# 3 : Impemented in tensorflow using tf.losses.log_loss\ny_pred_tf = tf.convert_to_tensor(y_pred_arr, np.float32)\ny_true_tf = tf.convert_to_tensor(y_true_arr, np.float32)\nloss_tf = tf.losses.log_loss(labels=y_true_tf, predictions=y_pred_tf, epsilon=1e-6)\nwith tf.Session() as sess:\n    loss = sess.run(loss_tf)\n    print(loss)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>y_pred_arr.shape (3, 3)\ny_true_arr.shape (3, 3)\n0.30543092\n0.30543154478209544\n0.18106996\n<\/code><\/pre>\n","answers":[{"AnswerId":"53031293","CreationDate":"2018-10-28T12:01:20.477","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Like @functor mentioned <code>tf.losses.log_loss<\/code> is for binary crossentropy, you can check source code here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/losses\/losses_impl.py#L309\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/losses\/losses_impl.py#L309<\/a><\/p>\n\n<p>Also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2462\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2462<\/a><\/p>\n\n<p>Here is example:<\/p>\n\n<pre><code># Array of samples\ny_pred_arr = np.array([[0.4,0.3,0.3], [1.0,0.0,0.0], [0.0,0.0,1.0]], np.float32)\ny_true_arr = np.array([[1,0,0], [1,0,0], [0,0,1]], np.float32)\n\nprint('y_pred_arr.shape', y_pred_arr.shape)\nprint('y_true_arr.shape', y_true_arr.shape)\n\n# 1 : Implemented in numpy\ndef cross_entropy_loss(y_pred_arr, y_true_arr, eps=1e-6):\n    p_arr = np.clip(y_pred_arr, eps, 1 - eps)\n    n_samples = p_arr.shape[0]\n    n_classes = p_arr.shape[1]\n    loss_arr = np.zeros((n_samples,), np.float32)\n    for i in range(n_samples):\n        for c in range(n_classes):\n            loss_arr[i] += -y_true_arr[i][c] * np.log(p_arr[i][c])\n    loss = np.mean(loss_arr)\n    return loss\nloss = cross_entropy_loss(y_pred_arr, y_true_arr)\nprint(loss)\n\n# 2 : Implemented in sklearn\nfrom sklearn.metrics import log_loss\nloss = log_loss(np.argmax(y_true_arr,axis=1), y_pred_arr, labels=[0,1,2], eps=1e-6)\nprint(loss)\n\n# 3 : Implemented in tensorflow\ny_pred_tf = tf.convert_to_tensor(y_pred_arr, np.float32)\ny_true_tf = tf.convert_to_tensor(y_true_arr, np.float32)\neps = 1e-6\ncliped_y_pref_tf = tf.clip_by_value(y_pred_tf, eps, 1-eps)\nloss_tf = tf.reduce_mean(-tf.reduce_sum(y_true_tf * tf.log(cliped_y_pref_tf), axis=1))\nwith tf.Session() as sess:\n    loss = sess.run(loss_tf)\n    print(loss)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>y_pred_arr.shape (3, 3)\ny_true_arr.shape (3, 3)\n0.30543092\n0.30543154478209544\n0.30543092\n<\/code><\/pre>\n"},{"AnswerId":"53028664","CreationDate":"2018-10-28T05:34:59.290","ParentId":null,"OwnerUserId":"4554481","Title":null,"Body":"<p><code>tf.losses.log_loss<\/code> only works for binary classes. For multiclass loss, you need to compute manually, <\/p>\n\n<p>like:<\/p>\n\n<pre><code>cliped_y_pref_tf = tf.clip_by_value(y_pred_tf, eps, 1-eps)\nloss_tf = tf.reduce_sum(tf.multiply(-y_true_tf,tf.log(cliped_y_pref_tf )))\/len(y_pred_arr)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53028051,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning>","CreationDate":"2018-10-28T03:06:31.837","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"FailedPreconditionError Keras","Body":"<p>I am trying to apply mask to the part of the screen that i capture. <code>screen = grab_screen(region=(0,40,800,640))<\/code>. Then, i apply my function that is contained in  another file. <code>screen = road_lines_image(screen)<\/code>(that is how i import it <code>from linedetect import road_lines_image<\/code>). In another file, i have a following function. <\/p>\n\n<pre><code>def road_lines_image(imageIn):\n    #crop to 720x1280, img[y: y + h, x: x + w], 300:940\n    image = imageIn[230:950, 0:1280]\n    image = imresize(image, (640, 1280, 3))\n\n    # Get image ready for feeding into model\n    small_img = imresize(image, (80, 160, 3))\n    small_img = np.array(small_img)\n\n    small_img = small_img[None,:,:,:]\n\n    # Make prediction with neural network (un-normalize value by multiplying by 255)\n    prediction = model.predict(small_img)[0] * 255\n<\/code><\/pre>\n\n<p>(function is longer but the last line is where i get an error. After my function call from the main file, i get an error <\/p>\n\n<pre><code>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value Final\/kernel\n         [[{{node Final\/kernel\/read}} = Identity[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Final\/kernel)]]\n<\/code><\/pre>\n\n<p>I tried to run following code from the file where road_lines_image is<\/p>\n\n<pre><code>img = cv2.imread(\"road-traffic-car-981035.jpg\")\nimg  = road_lines_image(img)\ncv2.imshow('image',img)\ncv2.waitKey(0)\n<\/code><\/pre>\n\n<p>This code works perfect and i get a desired output. I encounter the problem when i run from the main file. <\/p>\n","answers":[{"AnswerId":"53028085","CreationDate":"2018-10-28T03:13:56.907","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Solved! I declared model outside of the function. I added it inside and now it works perfectly!<\/p>\n"}]}
+{"QuestionId":53028372,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><gpu>","CreationDate":"2018-10-28T04:27:04.027","AcceptedAnswerId":null,"OwnerUserId":5493855.0,"Title":"TensorFlow GPU failed to load\/import","Body":"<p>I installed TensorFlow-GPU 1.11.0 and when I try to import it, I'm getting error. I'm stuck at this point.\nMy system specifications are,<\/p>\n\n<ul>\n<li>Windows 10 64bit<\/li>\n<li>GeForce GTX1080 <\/li>\n<li>Graphic Driver 385.54<\/li>\n<li>Python 3.5.4<\/li>\n<li>Cuda Toolkit 9.0<\/li>\n<li>CuDNN 7.3<\/li>\n<li>Visual Studio 2015<\/li>\n<\/ul>\n\n<p>Don't know what is missing. Kindly help me regarding this issue. I will be thankful. Below is the error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.\n<\/code><\/pre>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\Ahsan\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\Ahsan\\AppData\\Loca\n<\/code><\/pre>\n","answers":[{"AnswerId":"53028413","CreationDate":"2018-10-28T04:42:14.833","ParentId":null,"OwnerUserId":"8274431","Title":null,"Body":"<pre><code> pip install tensorflow --upgrade --force-reinstall\n<\/code><\/pre>\n\n<p>This command worked for me previously in a similar case. Hope it helps you.<\/p>\n"}]}
+{"QuestionId":53028377,"AnswerCount":1,"Tags":"<python><keras><lstm><text-classification><variable-length>","CreationDate":"2018-10-28T04:28:42.850","AcceptedAnswerId":53030229.0,"OwnerUserId":7519863.0,"Title":"\"ValueError: The shape of the input to \"Flatten\" is not fully defined\" with variable length LSTM","Body":"<p>Here's my code:<\/p>\n\n<pre><code>    from keras.layers import LSTM, Bidirectional, Dense, Input, Flatten\n    from keras.models import Model\n\n    input = Input(shape=(None, 100))\n    lstm_out = Bidirectional(LSTM(10, return_sequences=True))(input)\n    something = Flatten()(lstm_out)\n    output = Dense(22, activation='softmax')(something)\n\n    model = Model(inputs=input, outputs=output)\n    model.compile(loss='categorical_crossentropy', optimizer='adadelta', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I'm building an LSTM with variable input through \n<a href=\"https:\/\/stackoverflow.com\/questions\/38189070\/how-do-i-create-a-variable-length-input-lstm-in-keras\">this stackoverflow question<\/a>. But now my model is saying <code>ValueError: The shape of the input to \"Flatten\" is not fully defined (got (None, 20)<\/code>. How can I fix this?<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"53030229","CreationDate":"2018-10-28T09:52:41.100","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You can't <em>fix<\/em> this particular problem because you can pass a variable size vector to a <code>Dense<\/code> layer. <em>Why?<\/em> Because it has a fixed size weights matrix, i.e. the kernel <code>W<\/code>.<\/p>\n\n<p>You should instead look at layers that can handle variable length sequences such as RNNs. For example you can let the LSTM learn a representation over the entire sequence:<\/p>\n\n<pre><code>input = Input(shape=(None, 100))\nlstm_out = Bidirectional(LSTM(10))(input) # the LSTM produces a single fixed size vector\noutput = Dense(22, activation='softmax')(lstm_out) # Dense classifies it\n<\/code><\/pre>\n\n<p>If you want more capacity in your model you can chain RNN layers so long as the last one doesn't return sequences:<\/p>\n\n<pre><code>lstm_out = Bidirectional(LSTM(10, return_sequences=True))(input)\nlstm_out = Bidirectional(LSTM(10))(lstm_out) # this LSTM produces a single vector\n<\/code><\/pre>\n"}]}
+{"QuestionId":53029118,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T06:59:46.030","AcceptedAnswerId":null,"OwnerUserId":10535511.0,"Title":"Tensorflow tf.scatter_update does not update float value","Body":"<p>It seems like tf.scatter_update does not update the variable if the type of the variable is defined as float. Here's a code you can try:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef cond(size, i):\n    return tf.less(i,size)\n\ndef body(size, i):\n    b=2*7.5+c\n\n    with tf.variable_scope(\"a\", reuse=tf.AUTO_REUSE):\n        a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n        a = tf.scatter_update(a,i,b)\n\n        with tf.control_dependencies([a]):\n            return (size, i+1)\n\nwith tf.Session() as sess:\n    c=tf.constant(4.0)\n    i = tf.constant(0)\n    size = tf.constant(6)\n    _,i = tf.while_loop(cond,\n            body,\n            [size, i])\n\n    a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n    init = tf.initialize_all_variables()\n    sess.run(init)\n    print(sess.run([a,i]))\n<\/code><\/pre>\n\n<p>The result would be random ! since I did not initialize the variable on purpose to see how it gets updated, it seems that it never gets updated and the random initialization would be printed out every time. You would see something like this: <\/p>\n\n<p>[array([-0.35466522,  0.44001752,  0.21131486, -0.48532146,  0.3019274 ,\n       -0.19926369], dtype=float32), 6]<\/p>\n\n<p>Is this a bug ? as you can see, I am still using the tf.control_dependencies and it only happens if the type of the variable a is set to float. <\/p>\n","answers":[{"AnswerId":"53032388","CreationDate":"2018-10-28T14:08:56.397","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>Is your expected output this ?<\/p>\n\n<pre><code>[array([19., 19., 19., 19., 19., 19.], dtype=float32), 6]\n<\/code><\/pre>\n\n<p>These two patterns produce that.<\/p>\n\n<h3>Pattern 1<\/h3>\n\n<pre><code>import tensorflow as tf\n\ndef cond(size, i):\n    return tf.less(i,size)\n\ndef body(size, i):\n    b=2*7.5+c\n\n    with tf.variable_scope(\"a\", reuse=tf.AUTO_REUSE):\n        a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n        a = tf.scatter_update(a,i,b)\n        with tf.control_dependencies([a]):\n            return (size, i+1)\n\nwith tf.Session() as sess:\n    c=tf.constant(4.0)\n    i = tf.constant(0)\n    size = tf.constant(6)\n    _,i = tf.while_loop(cond,\n            body,\n            [size, i])\n\n    with tf.variable_scope(\"a\", reuse=tf.AUTO_REUSE):\n        a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n        init = tf.initialize_all_variables()\n        sess.run(init)\n\n        print(sess.run([a,i]))\n<\/code><\/pre>\n\n<h3>Pattern 2<\/h3>\n\n<pre><code>def body(size, i):\n    b=2*7.5+c\n\n    a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n    a = tf.scatter_update(a,i,b)\n    #Reuse variables \n    tf.get_variable_scope().reuse_variables()\n\n    with tf.control_dependencies([a]):\n        return (size, i+1)\n\nwith tf.Session() as sess:\n    c=tf.constant(4.0)\n    i = tf.constant(0)\n    size = tf.constant(6)\n    _,i = tf.while_loop(cond,\n            body,\n            [size, i])\n\n    a = tf.get_variable(\"a\",[6],dtype=tf.float32)\n\n    init = tf.initialize_all_variables()\n    sess.run(init)\n\n    print(sess.run([a,i]))\n<\/code><\/pre>\n"},{"AnswerId":"55192536","CreationDate":"2019-03-16T01:11:50.303","ParentId":null,"OwnerUserId":"4479188","Title":null,"Body":"<p>Your question is answered much better <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23333\" rel=\"nofollow noreferrer\">here<\/a> as a GitHub Tensorflow issue.<\/p>\n\n<p>I'll try to summarize the tread here. There are two main issues:<\/p>\n\n<ol>\n<li>When variable <code>a<\/code> is retrieved before running the session, it must be under the variable scope \"a\", otherwise you're creating a new variable. As @MohanRadhakrishnan mentioned.<\/li>\n<li><code>sess.run([a,i])<\/code> does not define the evaluation order of <code>a<\/code> and <code>i<\/code>. <code>a<\/code> can be evaluated either before, after, or during the loop, and before the loop <code>a<\/code> is randomly initialized. Therefore you can get different results. Thus, in this case the order must be defined by running <code>sess.run(i)<\/code> before running <code>sess.run(a)<\/code>. This can cause more issues on GPU due to more parallelization.<\/li>\n<\/ol>\n\n<p>So the code will be:<\/p>\n\n<pre><code>def cond(size, i):\n    return tf.less(i, size)\n\ndef body(size, i):\n    b = 2 * 7.5 + c\n\n    with tf.variable_scope(\"a\", reuse=tf.AUTO_REUSE):\n        a = tf.get_variable(\"a\", [6], dtype=tf.float32)\n\n        a = tf.scatter_update(a, i, b)\n\n        with tf.control_dependencies([a]):\n            return (size, i + 1)\n\nwith tf.Session() as sess:\n    c = tf.constant(4.0)\n    i = tf.constant(0)\n    size = tf.constant(6)\n    _, i = tf.while_loop(cond,\n                         body,\n                         [size, i])\n\n    with tf.variable_scope(\"a\", reuse=tf.AUTO_REUSE):  # Issue 1. Reuse 'a'.\n        a = tf.get_variable(\"a\", [6], dtype=tf.float32)\n\n    init = tf.initialize_all_variables()\n    sess.run(init)\n    print(sess.run(i))  # issue 2. Define the order of operation.  \n    print(sess.run(a))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53029594,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><rnn>","CreationDate":"2018-10-28T08:18:10.833","AcceptedAnswerId":53030192.0,"OwnerUserId":4841832.0,"Title":"Processing sequences of variable length within a batch in Keras","Body":"<p>Suppose I have some data of shape <code>(?, 10)<\/code>.<\/p>\n\n<p>That is, the features of the data are sequences of variable length and each element of the sequence is expressed by 10 numbers.<\/p>\n\n<p>We want to feed this to a LSTM, so we prepare each batch to have length <code>(32, m, 10)<\/code>, where m is the max sequence length of the examples within the batch.<\/p>\n\n<p>The examples within the batch of sequence length shorter than m are padded with zeroes.<\/p>\n\n<p>Now, we want to feed this to a LSTM and we want the LSTM to stop updating the output on the padding inputs.<\/p>\n\n<p>In Tensorflow, this would be done via the parameter <code>sequence_length<\/code> of its <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"nofollow noreferrer\"><code>dynamic_rnn<\/code><\/a>.<\/p>\n\n<p>How can I achieve the same effect in Keras?<\/p>\n","answers":[{"AnswerId":"53030192","CreationDate":"2018-10-28T09:46:24.043","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You need to use a <a href=\"https:\/\/keras.io\/layers\/core\/#masking\" rel=\"nofollow noreferrer\">Masking<\/a> which will produce a <em>mask<\/em> that allows the LSTM to skip those padded values. From the documentation:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Masking(mask_value=0., input_shape=(None, 10)))\nmodel.add(LSTM(32))\n<\/code><\/pre>\n\n<p>The LSTM above will now skip timesteps where all the 10 features are 0 padded. <strong>Note:<\/strong> if you are returning sequences then the previous hidden state will be returned:<\/p>\n\n<pre><code>x = [2, 0, 1, 0] # for example\n# will produce\ny = [h1, h1, h2, h2] # so 0 pads are skipped\n# but returning sequences repeats the last hidden state for masked values\n<\/code><\/pre>\n"}]}
+{"QuestionId":53029932,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T09:09:18.183","AcceptedAnswerId":53035749.0,"OwnerUserId":8005524.0,"Title":"How can i get some values of arrays in tensorflow?","Body":"<p>I have two tensors(A,B),A's shape is (N,N,12) and B's shape is (N,N).<br\/>\nI want retain some values and delete others in A according to B.B just like a dictionary.<br\/>\nFor example:<\/p>\n\n<pre><code>B[1][1] = 2\nnewArray[1][1] = A[1][1][2*4:2*4+4]\nB[i][j] = n  # n in [0:2]\nnewArray[i][j] = A[i][j][n*4:n*4+4]\nassert(newArray.shape == (N,N,4))\n<\/code><\/pre>\n\n<p>How do i write the code in tensorflow?<br\/>\nThank you very much.<\/p>\n","answers":[{"AnswerId":"53035749","CreationDate":"2018-10-28T20:23:03.217","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think this does what you want:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input data\nA = tf.placeholder(tf.float32, (None, None, 12))\nB = tf.placeholder(tf.int32, (None, None))\n# Reshape 12-vectors into 3x4 inner matrices\nA_shape = tf.shape(A)\nrows, cols = A_shape[0], A_shape[1]\nA_res = tf.reshape(A, (rows, cols, 3, 4))\n# Make indices\nii, jj = tf.meshgrid(tf.range(rows), tf.range(cols), indexing='ij')\nB_idx = tf.stack([ii, jj, B], axis=-1)\n# Gather result\nresult = tf.gather_nd(A_res, B_idx)\n# Test\nwith tf.Session() as sess:\n    A_val = [\n        [\n            [ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11,],\n            [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,],\n        ],\n        [\n            [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35,],\n            [36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,],\n        ],\n    ]\n    B_val = [\n        [2, 0],\n        [1, 2],\n    ]\n    result_val = sess.run(result, feed_dict={ A: A_val, B: B_val })\n    print(result_val)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[ 8.  9. 10. 11.]\n  [12. 13. 14. 15.]]\n\n [[28. 29. 30. 31.]\n  [44. 45. 46. 47.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53030001,"AnswerCount":0,"Tags":"<python><dataset><pytorch>","CreationDate":"2018-10-28T09:19:20.293","AcceptedAnswerId":null,"OwnerUserId":5650267.0,"Title":"How to index specific object in dataloader without enumeration","Body":"<p>I have the following code for testing a network in pytorch (taken from <a href=\"https:\/\/github.com\/junyanz\/pytorch-CycleGAN-and-pix2pix\/blob\/master\/test.py\" rel=\"nofollow noreferrer\">pix2pix<\/a>):<\/p>\n\n<pre><code>for i, data in enumerate(dataset):\n    model.set_input(data)\n    model.test()\n<\/code><\/pre>\n\n<p>I have tried altering this to access only one element (any) over iterating a number of elements. the following:<\/p>\n\n<pre><code>data=dataset[0]\n<\/code><\/pre>\n\n<p>leads to<\/p>\n\n<blockquote>\n  <p>TypeError: 'CustomDatasetDataLoader' object does not support indexing<\/p>\n<\/blockquote>\n\n<p>where <code>CustomDatasetDataLoader<\/code> is a <a href=\"https:\/\/github.com\/junyanz\/pytorch-CycleGAN-and-pix2pix\/blob\/master\/data\/__init__.py\" rel=\"nofollow noreferrer\">custom class by the author<\/a> which (if I am understanding correctly) herniates from <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html\" rel=\"nofollow noreferrer\">torch.utils.data.DataLoader<\/a>. <\/p>\n\n<p>I am fairly new to python, how do I index a single specific element of <code>dataset<\/code> if it is at all possible? <\/p>\n","answers":[]}
+{"QuestionId":53030121,"AnswerCount":1,"Tags":"<python><google-app-engine><keras><gensim><google-app-engine-python>","CreationDate":"2018-10-28T09:37:13.743","AcceptedAnswerId":53032208.0,"OwnerUserId":10117402.0,"Title":"How to load files to Google-App-Engine in standard enviroment","Body":"<p>I am using Google-App-Engine standard (Not flex) Enviroment with Python2.7, and I need to load some pre-trained models (Gensim's Word2vec and Keras's LSTM).<\/p>\n\n<p>I need to load it once (since it very slow - takes around 1.5 seconds) and keep it in faster access for several hours.<\/p>\n\n<p>What is the best &amp; fastest way to do so? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53032208","CreationDate":"2018-10-28T13:48:45.260","ParentId":null,"OwnerUserId":"4495081","Title":null,"Body":"<p>IMHO the best place for read-only data (including imported code!) needed to be accessed at any time by individual requests is the global application variables area.<\/p>\n\n<p>Such variables would typically be loaded exactly once per GAE instance lifetime and available until the instance goes away. <\/p>\n\n<p>Since loading of the data is expensive you need to be aware that it could impact the response time for requests coming in while the instance is starting up (i.e. while the loading request is still active). There are 2 ways to address this:<\/p>\n\n<ul>\n<li><p>one would be to use \"lazy\" loading of the data - effective if just a small percentage of the incoming requests actually need the data. But the requests which actually need the data when it's not available will still be affected, so it'll just reduce the impact of the problem. The method is described in detail in the <a href=\"https:\/\/medium.com\/google-cloud\/app-engine-startup-time-and-the-global-variable-problem-7ab10de1f349\" rel=\"nofollow noreferrer\">App Engine Startup time and the Global Variable problem<\/a> article:<\/p>\n\n<pre><code>from google.appengine.ext import ndb\n# a global variable\ngCDNServer = None\ndef getCDN():\n    global gCDNServer\n    if gCDNServer==None:\n        gCDNServer = Settings.query(Settings.name == \"gCDNServer\").value\n    return gCDNServer\n<\/code><\/pre><\/li>\n<li><p>another approach, which would completely eliminate the problem, is to make your app support <a href=\"https:\/\/cloud.google.com\/appengine\/docs\/standard\/python\/configuring-warmup-requests\" rel=\"nofollow noreferrer\">warmup requests<\/a> (available only if you're using automatic scaling). The data would be loaded by the warmup request handler and will always be available for \"live\" requests (because no \"live\" requests will be routed to the instance until the warmup request handling completes).<\/p><\/li>\n<\/ul>\n\n<p>It <em>might<\/em> be possible to add logic to drop the data from memory (to reduce the app's memory footprint) if\/when you know it'll no longer be needed (i.e. after those several hours you mentioned expired), but that complicates the picture, especially if you configured your app as <code>threadsafe<\/code>. I'd simply separate the code which doesn't need the data from the one which does in different services and leave autoscaling shut down the instances with the global data when no longer needed.<\/p>\n"}]}
+{"QuestionId":53030289,"AnswerCount":1,"Tags":"<java><spring-boot><jpa><caching><caffeine>","CreationDate":"2018-10-28T10:00:10.647","AcceptedAnswerId":null,"OwnerUserId":4464017.0,"Title":"How to add entire table to cache in spring","Body":"<p>I have a very small table which is not updated frequently. \nI want to add this to cache such that it updates every day.\nI am using spring and caffeine to implement this. \nI able to load a startup but don't how to refresh it. \nPlease help. <\/p>\n\n<pre><code>@Bean\npublic CacheManager cacheManager() {\n\n    SimpleCacheManager simpleCacheManager = new SimpleCacheManager();\n\n    Cache stringStringCache = new CaffeineCache(\"name\", Caffeine.newBuilder()\n            .recordStats()\n            .maximumSize(100)\n            .expireAfterWrite(1, TimeUnit.DAYS)\n            .build());\n\n    simpleCacheManager.setCaches(Collections.singleton(stringStringCache));\n    return simpleCacheManager;\n}\n<\/code><\/pre>\n\n<p>I can simply  fetch all records from repository and put that in the cache using \ncache.put().\nBut how i refresh it again from table after specified time interval.<\/p>\n","answers":[{"AnswerId":"53048961","CreationDate":"2018-10-29T15:37:10.403","ParentId":null,"OwnerUserId":"4464017","Title":null,"Body":"<p>found the answer.<br>\nIt loads the table on first call.<br>\nAfter that we just to read from cache from subsequent calls<\/p>\n\n<pre><code>@Cacheable(\"name\")\n@Override\npublic Map&lt;String,String&gt; findNameById() {\n    log.info(\"db call\");\n    return IteratorUtils.toList(bookRepository\n            .findAll()\n            .iterator())\n            .stream()\n            .collect(Collectors.toMap(Book::getId,Book::getName));\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":53030636,"AnswerCount":1,"Tags":"<java><python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-28T10:46:23.187","AcceptedAnswerId":null,"OwnerUserId":1617910.0,"Title":"TensorFlow Model with Fashion MNIST dataset","Body":"<p>I am very new to <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"nofollow noreferrer\">TensorFlow<\/a>. I want to use TensorFlow for my fashion e-commerce site (runs on java) in such a way that all product images should be auto tagged using CNN.<\/p>\n\n<p>I checked <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/main\/java\/org\/tensorflow\/examples\/LabelImage.java\" rel=\"nofollow noreferrer\">LabelImage.java<\/a> example using inception5h pre trained tensorflow model for auto tagging\/label new images.<\/p>\n\n<p>It seems that inception5h is not fit for fashion dataset. I also found one guide explained on train first neural network with basic classification used Fashion MNIST dataset.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a><\/p>\n\n<p>I also want to create my own model, train it on Fashion MNIST dataset and use it in my java code for auto tagging.<\/p>\n\n<p>Question -<\/p>\n\n<ol>\n<li>Can we retrain inception5h model for Fashion MNIST dataset as well ?<\/li>\n<li>Is there any pre-trained tensorflow model available for fashion dataset ?<\/li>\n<li>If I trained the model, can i used the same model in my java code by referring <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/main\/java\/org\/tensorflow\/examples\/LabelImage.java\" rel=\"nofollow noreferrer\">LabelImage.java<\/a> example ? LabelImage example required model and graph label.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"54165635","CreationDate":"2019-01-13T02:30:31.270","ParentId":null,"OwnerUserId":"10906564","Title":null,"Body":"<p>I am just looking into Fashion-MNIST myself. It is said that F-MNIST deliberately imitates all the property that MNIST has and any model trained on MNIST can be used on F-MNIST with just a slight change of dataset path.<\/p>\n\n<p>Although there is a problem of overfitting while transferring my own MNIST model to a Fashion-MNIST dataset. I'm still checking my code and trying to find out what's wrong.<\/p>\n"}]}
+{"QuestionId":53030642,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2018-10-28T10:47:00.353","AcceptedAnswerId":null,"OwnerUserId":10131221.0,"Title":"tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [1568] vs. [32,49]","Body":"<p>my keras encoder-decoder code works fine on cpu. but on gpu I cannot launch it with batch_size other  than 1 ! This is strange. Even when I do not use batch size argument in this fitting I get:<\/p>\n\n<p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [1568] vs. [32,49]<\/p>\n\n<p>(by default keras takes batch_size=32)<\/p>\n\n<p>I can run many other programs with whatever batch_size I want. Wht module is comparing a two dimensional object with one dimensional ?<\/p>\n","answers":[]}
+{"QuestionId":53031162,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-28T11:47:59.563","AcceptedAnswerId":53031294.0,"OwnerUserId":1298461.0,"Title":"Getting dimensions right for a single layer keras LSTM","Body":"<p>I have some hard time to get the dimensions of a LSTM network right.<\/p>\n\n<p>So I have the following data:<\/p>\n\n<pre><code>train_data.shape\n (25391, 3) # to be read as 25391 timesteps and 3 features\n\ntrain_labels.shape\n (25391, 1) # to be read as 25391 timesteps and 1 feature\n<\/code><\/pre>\n\n<p>So I have thought my input dimension is <code>(1, len(train_data), train_data.shape[1])<\/code> as I plan to submit 1 batch. But I get the following error:<\/p>\n\n<pre><code>Error when checking target: expected lstm_10 to have 2 dimensions, but got array with shape (1, 25391, 1)\n<\/code><\/pre>\n\n<p>Here is the model code:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(1, # predict one feature and one timestep\n               batch_input_shape=(1, len(train_data), train_data.shape[1]),\n               activation='tanh',\n               return_sequences=False))\n\nmodel.compile(loss = 'categorical_crossentropy', optimizer='adam', metrics = ['accuracy'])\nprint(model.summary())\n\n# as 1 sample with len(train_data) time steps and train_data.shape[1] features.\nmodel.fit(x=train_data.values.reshape(1, len(train_data), train_data.shape[1]), \n          y=train_labels.values.reshape(1, len(train_labels), train_labels.shape[1]), \n          epochs=1, \n          verbose=1, \n          validation_split=0.8, \n          validation_data=None, \n          shuffle=False)\n<\/code><\/pre>\n\n<p>How should the input dimensions look like?<\/p>\n","answers":[{"AnswerId":"53031294","CreationDate":"2018-10-28T12:01:22.273","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is in the target (i.e. labels) shape you provide (i.e. <code>Error when checking target<\/code>). The output of LSTM layer in your model, which is also the output of the model, has a shape of <code>(None, 1)<\/code> since you are specifying to only the final output to be returned (i.e. <code>return_sequences=False<\/code>). In order to have the output of each timestep you need to set <code>return_sequences=True<\/code>. This way the output shape of LSTM layer would be <code>(None, num_timesteps, num_units)<\/code> which is consistent with the shape of labels array you provide.<\/p>\n"}]}
+{"QuestionId":53031277,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T12:00:04.643","AcceptedAnswerId":null,"OwnerUserId":10570324.0,"Title":"ConcatOp : Dimensions of inputs should match: shape[0] = [1,1024,300,476] vs. shape[1] = [1,1024,76,120]","Body":"<pre><code>with slim.arg_scope(resnet_arg_scope(is_training=(is_training and not_freezed[2]))):\n    C4, _ = resnet_v1.resnet_v1(C3,\n                                blocks[2:3],\n                                global_pool=False,\n                                include_root_block=False,\n                                scope=scope_name)\n# add_heatmap(C4, name='Layer\/C4')\n# C4 = tf.Print(C4, [tf.shape(C4)], summarize=10, message='C4_shape')\nfeature_c2=slim.conv2d_transpose(C2,\n                                 1024,\n                                 [3,3],                                  \n                                 stride=4,\n                                 padding='SAME')\n\nfeature_c3=slim.conv2d_transpose(C3,\n                                 1024,\n                                 [3,3],\n                                 stride=2,\n                                 padding='SAME') \nfeature_to_cropped=tf.concat([**strong text**feature_c2,feature_c3,C4],3)\nfeature_to_cropped=slim.conv2d(feature_to_cropped,1024,[3,3],1)\nprint(C4)\nprint(feature_to_cropped)\nreturn feature_to_cropped\n<\/code><\/pre>\n\n<p>I want to concat the feature map and I checked the shape of feature_c2 and feature_c3 are[1,height,width,1024],c4 is [1,height,width,1024],I concat them and conv ,the result to return is [1,height,width,1024],but why it need channel in the dim 2?<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): ConcatOp : Dimensions\n  of inputs should match: shape[0] = [1,1024,300,476] vs. shape[1] =\n  [1,1024,76,120]\n           [[Node: concat = ConcatV2[N=3, T=DT_FLOAT, Tidx=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](Conv2d_transpose\/Relu,\n  Conv2d_transpose_1\/Relu,\n  resnet_v1_101_3\/block3\/unit_23\/bottleneck_v1\/Relu,\n  postprocess_RPN\/ones\/Const)]]\n           [[Node: postprocess_fastrcnn\/GatherV2_20\/_1405 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\",\n  send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\",\n  send_device_incarnation=1,\n  tensor_name=\"edge_4116_postprocess_fastrcnn\/GatherV2_20\",\n  tensor_type=DT_FLOAT,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53031436,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning><computer-vision>","CreationDate":"2018-10-28T12:19:19.827","AcceptedAnswerId":null,"OwnerUserId":5936685.0,"Title":"Trouble building a custom loss function in tensorflow","Body":"<p>I have been trying to develop the YOLO cost function which I have shown below. This is the first time I have tried to develop my own cost function in Tensorflow and am unsure if I am approaching it correctly or not. For one, my model uses a number of intermediate steps. I not sure if this complicates the computational graph in some meaningfully destructive way? Or, I am using an abs. value step and am unsure whether it would have some negative effect on my backprop? Any assistance would be helpful in regard to whether I am approaching this problem correctly.<\/p>\n\n<p>I can answer any questions about my implementation.<\/p>\n\n<p>Note - Z13 is the prediction, y are the true values. There are 49 cells in my model (7x7) with each cell being represented by a 7x1 vector: [prob of anything in cell, x midpoint, y midpoint, box width, box height, prob dog, prob cat] .Referenced paper: <a href=\"https:\/\/arxiv.org\/pdf\/1506.02640.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1506.02640.pdf<\/a> which explains the cost function in depth.<\/p>\n\n<p>I believe that there is either an issue with my forward prop or my cost function as my model is not learning meaningful representations.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6rQNC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6rQNC.png\" alt=\"YOLO cost function\"><\/a><\/p>\n\n<pre><code>def cost_function(Z13,y,coord=5,noobj=0.5):\n    \"\"\"\n    Z13: shape (None,7,7,7)\n    y: shape (None,7,7,7)\n    \"\"\"\n\n    # Masks are used as classification score for box coords only applies to cell where actual bounding box is\n    c_mask_true = y[:,:,:,0:1] &gt; 0 # Mask which determines which cell has bounding box\n    c_mask_false = y[:,:,:,0:1] &lt; 1 # Mask for cells w\/o bounding boxes\n\n    # Confidence scores\n    ci_guess_t = tf.boolean_mask(Z13[:,:,:,0:1],c_mask_true)\n    ci_guess_f = tf.boolean_mask(Z13[:,:,:,0:1],c_mask_false)\n    ci_act_t = tf.boolean_mask(y[:,:,:,0:1],c_mask_true)\n    ci_act_f = tf.boolean_mask(y[:,:,:,0:1],c_mask_false)\n\n    # Bounding box coordinated for ground truth box prediction\n    xi_guess = tf.boolean_mask(Z13[:,:,:,1:2],c_mask_true) # Midpoint x position\n    xi_act = tf.boolean_mask(y[:,:,:,1:2],c_mask_true)\n    yi_guess = tf.boolean_mask(Z13[:,:,:,2:3],c_mask_true) # Midpoint y position\n    yi_act = tf.boolean_mask(y[:,:,:,2:3],c_mask_true)\n    # Width:\n    wi_guess = tf.boolean_mask(Z13[:,:,:,3:4],c_mask_true) # Midpoint width pos.\n    wi_guess = tf.minimum(tf.sqrt(tf.abs(wi_guess)),wi_guess) # prevent sqrt(neg) and increase cost for neg prediction\n    wi_act = tf.sqrt(tf.boolean_mask(y[:,:,:,3:4],c_mask_true))\n    # Height:\n    hi_guess = tf.boolean_mask(Z13[:,:,:,4:5],c_mask_true) # Midpoint height pos.\n    hi_guess = tf.minimum(tf.sqrt(tf.abs(hi_guess)),hi_guess) # prevent sqrt(neg) and increase cost for neg prediction\n    hi_act = tf.sqrt(tf.boolean_mask(y[:,:,:,4:5],c_mask_true))   \n\n    # Predicted classes: \n    class_g_dog = tf.boolean_mask(Z13[:,:,:,5:6],c_mask_true)\n    class_t_dog = tf.boolean_mask(y[:,:,:,5:6],c_mask_true)\n    class_g_cat = tf.boolean_mask(Z13[:,:,:,6:7],c_mask_true)\n    class_t_cat = tf.boolean_mask(y[:,:,:,6:7],c_mask_true)\n\n    # Parts correspond with the cost function equations above\n    part1 = coord * tf.reduce_sum(tf.square(xi_act - xi_guess)+tf.square(yi_act - yi_guess))\n    part2 = coord * tf.reduce_sum(tf.square(wi_act - wi_guess)+tf.square(hi_act - hi_guess))\n    part3 = tf.reduce_sum(tf.square(ci_act_t - ci_guess_t))\n    part4 = noobj * tf.reduce_sum(tf.square(ci_act_f - ci_guess_f))\n    part5 = tf.reduce_sum(tf.square(class_t_dog - class_g_dog)+tf.square(class_t_cat - class_g_cat))\n    total_cost = part1 + part2 + part3 + part4 + part5\n\n    return total_cost\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53031574,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-28T12:34:15.240","AcceptedAnswerId":null,"OwnerUserId":4841832.0,"Title":"Using a TimeDistributed Conv2D layer wrapper","Body":"<p>I want to create a Keras model consisting of a CNN => RNN => FCN.<\/p>\n\n<p>The model must be able to drop the elements of each sequence that are padded as zero 2D arrays.<\/p>\n\n<pre><code># Create model\nmodel = Sequential()\ninput_shape = (n_tel, img_w, img_h)\n\n# mask the inputs that correspond to padding\nmodel.add(Masking(mask_value=0., input_shape=input_shape))\n\n# Add CNN feature extractor\nmodel.add(TimeDistributed(\n        Conv2D(\n                filters=16, \n                kernel_size=(3,3), \n                padding='same',\n                activation='relu'\n                )\n        ))\nmodel.add(TimeDistributed(\n        Conv2D(\n                filters=32, \n                kernel_size=(3,3), \n                padding='same',\n                activation='relu'\n                )\n        ))\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2,2), strides=None)))\n\n# Add LSTM feature extractor\nmodel.add(LSTM(units=100))\n\nmodel.add(Dense(100))\nmodel.add(Activation('relu'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n<\/code><\/pre>\n\n<p>The above code returns an error that triggers on the first Conv2D layer:<\/p>\n\n<pre><code>IndexError: list index out of range\n<\/code><\/pre>\n\n<p>What is going on?<\/p>\n\n<hr>\n\n<p>EDIT: Full traceback<\/p>\n\n<pre><code>  Traceback (most recent call last):\n\n  File \"&lt;ipython-input-10-7253565e1e8b&gt;\", line 1, in &lt;module&gt;\n    runfile('\/home\/jsevillamol\/Documentos\/ctlearn_keras\/models\/build_cnn_rnn.py', wdir='\/home\/jsevillamol\/Documentos\/ctlearn_keras\/models')\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/spyder_kernels\/customize\/spydercustomize.py\", line 668, in runfile\n    execfile(filename, namespace)\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/spyder_kernels\/customize\/spydercustomize.py\", line 108, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"\/home\/jsevillamol\/Documentos\/ctlearn_keras\/models\/build_cnn_rnn.py\", line 78, in &lt;module&gt;\n    model = build_cnn_rnn(args.n_tel, args.img_w, args.img_h)\n\n  File \"\/home\/jsevillamol\/Documentos\/ctlearn_keras\/models\/build_cnn_rnn.py\", line 27, in build_cnn_rnn\n    activation='relu'\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/checkpointable\/base.py\", line 364, in _method_wrapper\n    method(self, *args, **kwargs)\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/sequential.py\", line 184, in add\n    output_tensor = layer(self.outputs[0])\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 728, in __call__\n    self.build(input_shapes)\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/wrappers.py\", line 213, in build\n    self.layer.build(tuple(child_input_shape))\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/convolutional.py\", line 182, in build\n    self.rank + 2))\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 828, in __init__\n    input_channels_dim = input_shape[num_spatial_dims + 1]\n\n  File \"\/home\/jsevillamol\/anaconda3\/envs\/ctlearn\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 615, in __getitem__\n    return self._dims[key]\n\nIndexError: list index out of range\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53031577,"AnswerCount":1,"Tags":"<python><amazon-web-services><azure><tensorflow><keras>","CreationDate":"2018-10-28T12:35:03.833","AcceptedAnswerId":null,"OwnerUserId":1412431.0,"Title":"Classify New Objects in Images Live","Body":"<p>I have a webcam, a microphone and a python GUI. The user shows the camera an object and using voice command asks \"what is this object?\". The webcam takes a photo of the camera frame and pushes it to a flask endpoint. I have a VGG16 model hosted in the flask app which responds with a object class based on the image from camera.<\/p>\n\n<p>What I want to do now is, if the object is not recognised, a learning cycle should be triggered. In this cycle I will tell the model what the object is (voice to text), which will be the label for the object. I got that working.<\/p>\n\n<p>What's not working is the next time (after the learning is complete), if I show the object to the camera, the model should be able to tell me what this object is.<\/p>\n\n<p>Could someone please advise me on the following:<\/p>\n\n<ol>\n<li><p>Is VGG16 (trained initially on 2 objects using transfer learning) the best model for this type of task? Currently, it is classifying unseen objects as one of the two classes.<\/p><\/li>\n<li><p>How would you go about implementing this solution on the cloud (AWS, Azure etc.)<\/p><\/li>\n<\/ol>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53032153","CreationDate":"2018-10-28T13:42:57.997","ParentId":null,"OwnerUserId":"10288309","Title":null,"Body":"<ol>\n<li><blockquote>\n  <p>Is VGG16 (trained initially on 2 objects using transfer learning) the\n  best model for this type of task? Currently, it is classifying unseen\n  objects as one of the two classes.<\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>Firstly the reason  for \"it is classifying unseen objects as one of the two classes \" is simply that, you only allow it to classify between the two classes, so what happens is that even if you show it an unknown object it is supposed to fit it on either of the classes and post its best prediction, what you should be do is that train it on 3 distinctive classes {object1, object2, unknownObject}, so you may correctly predict unknown objects, now this may be a bit problematic, and would require retraining the model, the other thing that you could do is that you should set a threshold, basically when you give it an unseen object it could give the prediction confidence as follows {75%, 25%} or maybe {51%, 49%} you should set a threshold here, that unless your model is more than 90-95% sure of the prediction you take it to be an unknown object, now i am in no way saying this should be your threshold value, you threshold could either be .9 \/ .95 as mentioned above, or it could even be .75, that is a hyper parameter and you should employ ways to figure that threshold out<\/p>\n\n<ol start=\"2\">\n<li><blockquote>\n  <p>How would you go about implementing this solution on the cloud (AWS,\n  Azure etc.)<\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>You already have a flask server, just deploy the flask server on an AWS machine, and make it directly accessible via a static public external IP, hence you can post\/get a request to that specific IP and run your model from any interface<\/p>\n"}]}
+{"QuestionId":53031887,"AnswerCount":2,"Tags":"<javascript><python><tensorflow><tensorflow.js>","CreationDate":"2018-10-28T13:12:21.920","AcceptedAnswerId":53032944.0,"OwnerUserId":4921713.0,"Title":"Unable to load python trained model in tensorflow.js","Body":"<p>I am getting the following errors when I am trying to load a Model, which I trained in Python, when I use the <code>loadModel()<\/code> function tensorflow.js:<\/p>\n\n<blockquote>\n  <p>Failed to load resource: net::ERR_NAME_NOT_RESOLVED<\/p>\n  \n  <p>Uncaught (in promise) TypeError: Failed to fetch<\/p>\n<\/blockquote>\n\n<p>Below is the <strong><em>predict.js<\/em><\/strong> file<\/p>\n\n<pre><code>console.log (\"hello\");\n\nlet model;\n(async function () {\n    model = await tf.loadModel(\"http:\/\/keras_model\/model.json\");\n    $(\".progress-bar\").hide();\n    console.log(\"it works\");\n})();\n<\/code><\/pre>\n\n<p><strong>The directory structure:<\/strong><\/p>\n\n<pre><code>main\n  -dataset  (contains images for training the model) \n  -training_scripts (python scripts to train the model)\n  -user_interface\n     --server.js (server made using node.js(and express))\n     --static (this folder contains the trained keras model)\n         --index.html (html file to be served)\n         --predict.js \n         --keras_model(this folder contains the model.json file)\n<\/code><\/pre>\n\n<p>Any help will be appreciated!!<\/p>\n","answers":[{"AnswerId":"53032944","CreationDate":"2018-10-28T15:07:58.513","ParentId":null,"OwnerUserId":"9507270","Title":null,"Body":"<p>If you want to load local files in tfjs you need to use the file type <code>file:\/\/\/<\/code> and for this to work you need the node extension of tfjs. You can load it by installing and requiring <code>node-fetch<\/code> into your program.<\/p>\n"},{"AnswerId":"53770027","CreationDate":"2018-12-13T21:02:56.887","ParentId":null,"OwnerUserId":"6457035","Title":null,"Body":"<p>You can also use the <code>fileSystem<\/code> handler exposed in tfjs-node like this:<\/p>\n\n<pre><code>const tf = require(\"@tensorflow\/tfjs\");\nconst tfn = require(\"@tensorflow\/tfjs-node\");\nconst handler = tfn.io.fileSystem(\".\/path\/to\/your\/model.json\");\nconst model = await tf.loadModel(handler);\n<\/code><\/pre>\n"}]}
+{"QuestionId":53031959,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-28T13:20:03.897","AcceptedAnswerId":null,"OwnerUserId":680074.0,"Title":"Keras - LSTM how to reshape input value","Body":"<p>I am using LSTM with keras my testing and training is working fine but when i try to enter the different input i am getting error cannot reshape size 20 to <code>(1,20,30)<\/code><\/p>\n\n<p>this is my code for model.fit()<\/p>\n\n<pre><code>PositiveOrNegativeLabel=np.array([[1]])\nPositiveOrNegativeLabel=PositiveOrNegativeLabel.reshape(1,-1)\nPositiveOrNegativeLabel.shape\ninputBatch =inputBatch.reshape(1,24,30)\ntestBatch =testBatch.reshape(1,24,30)\nmodel=Sequential()\nmodel.add(LSTM(100,input_shape=(24,30)))\nmodel.add(Dense(1,activation=\"relu\"))\nmodel.compile(loss='mean_absolute_error',optimizer='adam')\nmodel.fit(inputBatch,PositiveOrNegativeLabel,batch_size=24,epochs=9,verbose=1)\n<\/code><\/pre>\n\n<p><strong>my input is firstSentence array which is like this<\/strong><\/p>\n\n<pre><code>[  174 11501   420  4242 12111     0     0     0     0     0     0     0\n     0     0     0     0     0     0     0     0]\n<\/code><\/pre>\n\n<p><strong>the shape of firstSentence is (20,)<\/strong><\/p>\n\n<p>the kind of input which i have used to fit model<\/p>\n\n<pre><code>[[    0. 12184.   420.  4636.     0.  8840.     0.     0. 10499. 11508.\n   7511.     0.  5468.  2879.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  6689.  2818. 12003.  6480.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.     0.  3045. 11087.  2710.     0.   494.  1087.   420.  4995.\n  11516.  3637.  5842.     0.  9963.  7015. 11090.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.   494.     0.     0.  6541.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  8744. 11105.  1570.  5842.   174. 11266.  2929. 10438.  2879.\n      0. 10936.  6330.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 11956.  5222.     0.     0. 12106.  6481.     0.  7093. 13756.\n  12152.     0.     0.     0.     0. 10173.     0.  5173. 13756.  9371.\n      0.  9956.     0.     0.  9716.     0.     0.     0.     0.     0.]\n [    0.  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.   420.  5842.  3058. 11875.  2879.  1087. 11105.  4995.  8840.\n      0. 11100. 11875.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  5419.   420.  2250.  1299.  2151. 12111.  6454.     0. 11501.\n   8094.  5842.   942.  7503.  7410.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.   420.  5842.  3058. 11875.  2879.  1087. 11105.  4995.  8840.\n      0. 11100. 11875.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 11501.  1592. 10603. 11102.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.   174.  5842.  2387. 10453. 11090.     0.  7531. 11956.   450.\n    420. 11516.  6693.  2624.  9963. 11992.  9322. 11090. 12106.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  7544.     0.  1709.   420. 10936.  5222.  5842. 10407.  6937.\n  11329.  2937.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  1520.  1295.     0.  8396.  9322. 12715.     0.  5172.  7232.\n  11266.     0. 11266.  2757.  4416. 12020. 12111.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  7544.     0.  1709.   420. 10936.  5222.  5842. 10407.  6937.\n  11329.  2937.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.     0.  9191.  5952.     0.     0. 11516.  9413.  3081.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.     0.     0.     0.     0.     0. 11516.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.\n   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  9371. 10412.  2356.  5412. 11502.     0.  1087.   228.     0.\n   2937. 11480. 10412.  5412.   420.  9435.  2937.   228.  1057.  9435.\n  12111.     0.     0.     0.     0.     0.     0.     0.     0.     0.]]\n<\/code><\/pre>\n\n<p>The value error i am getting is like this<\/p>\n\n<pre><code>cannot reshape array of size 20 into shape (1,20,30)\n<\/code><\/pre>\n\n<p><strong>and I am running this code to predict it<\/strong> <\/p>\n\n<pre><code>predict=model.predict(firstSentence, batch_size=24, verbose=1, steps=None)\n# make a prediction\nynew = model.predict_classes(firstSentence)\n# show the inputs and predicted outputs\nfor i in range(len(predict)):\n    print(\"X=%s, Predicted=%s\" % (predict[i], ynew[i]))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53035927","CreationDate":"2018-10-28T20:47:13.853","ParentId":null,"OwnerUserId":"5755357","Title":null,"Body":"<p>I think you are confusing the <code>batch_size<\/code> with the number of timesteps. The <code>input_shape<\/code> to an LSTM layer in Keras should be (<code>num_timesteps, num_features)<\/code>. The <code>batch_size<\/code> is inferred based on the <code>batch_size<\/code> used in <code>fit<\/code>. From what I can tell, your training data has 24 examples, each with 30 timesteps that each have 1 feature. <\/p>\n\n<p>Moreover, your code only had one label, but from what I can tell you have 24 examples and will therefore need 24 labels. <\/p>\n\n<pre><code>import numpy as np \nfrom keras.models import Sequential\nfrom keras.layers import LSTM, Dense\n\n# Need same number of labels as examples\nPositiveOrNegativeLabel = np.ones(shape = (24, 1))\n# 24 examples, each with 30 timesteps, 1 feature at each timestep\ninputBatch = result.reshape(24, 30, 1)\n\nmodel=Sequential()\n\n# input_shape is (num_timesteps, num_features)\nmodel.add(LSTM(100, input_shape=(30, 1)))\nmodel.add(Dense(1, activation=\"relu\"))\n\nmodel.compile(loss='mean_absolute_error', optimizer='adam')\nmodel.fit(inputBatch, PositiveOrNegativeLabel,\n          batch_size=24, epochs=9, verbose=1)\n<\/code><\/pre>\n\n<p>Your testing sentence only has a single example with 20 timesteps each of which has 1 feature. Therefore, you will need to pad the testing sentence with zeros. <\/p>\n\n<pre><code>from keras.preprocessing.sequence import pad_sequences\n\n# Need 30 timesteps\nfirstSentence = pad_sequences([firstSentence], maxlen = 30)[0]\nfirstSentence = firstSentence.reshape((1, 30, 1))\n\npredict=model.predict(firstSentence, batch_size=1, verbose=1, steps=None)\n\n# make a prediction\nynew = model.predict_classes(firstSentence)\n# show the inputs and predicted outputs\nfor i in range(len(predict)):\n    print(\"X=%s, Predicted=%s\" % (predict[i], ynew[i]))\n<\/code><\/pre>\n\n<p>This code works, but I may have interpreted your set-up incorrectly. Thus, if you only have 1 training example as your training labels suggest, then the shape of the input batch should be <code>(1, 24, 30)<\/code> and your testing sequence will also need 24 timesteps with 30 features at each. <\/p>\n\n<p>For more on shaping inputs to an LSTM in Keras, see <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">this article<\/a>. <\/p>\n"}]}
+{"QuestionId":53031982,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-10-28T13:21:51.690","AcceptedAnswerId":null,"OwnerUserId":10204228.0,"Title":"Is there a way to directly analyze and manipulate tensorboard \"events.*\" log file?","Body":"<p>everyone. Tensorboard is a wonderful tool to visualize learning process, but it has some inconvenience for me. For some reason, sometimes I want to remove part of learning curve (like erase scalars after X-th step). However, all scalars are put together in a single \"events.*\" file, plus Tensorboard only provides high-level api  (only \"adding\" but no \"removing\") to the best of my knowledge. Does anyone have some ideas about this? Thanks!<\/p>\n","answers":[{"AnswerId":"53032873","CreationDate":"2018-10-28T15:00:03.627","ParentId":null,"OwnerUserId":"4101431","Title":null,"Body":"<p>You can view the contents of a checkpoint file using <code>print_tensors_in_checkpoint_file()<\/code>, <\/p>\n\n<pre><code>from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file\n\nprint_tensors_in_checkpoint_file(file_name=\"&lt;your_path_here&gt;\/model.ckpt-&lt;whatever_step&gt;\", tensor_name='', all_tensors=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53032356,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-10-28T14:04:21.660","AcceptedAnswerId":null,"OwnerUserId":8584587.0,"Title":"GAN outputting noise after training","Body":"<p>So, I set out to train a gan on generating cat\/dog photos, from kaggle's cats vs dogs dataset. After hyper-parameter tuning and everything in my power I'm still getting noise as output.\nThis what I did:<\/p>\n\n<p>1)I converted the images into 64*64<\/p>\n\n<p>2)Converted them into protobuffer<\/p>\n\n<p>3)Pushed it to the graph like:<\/p>\n\n<pre><code>def input_pipe(filenames,batch_size=32):\ndataset=tf.data.TFRecordDataset(filenames=filenames)\ndataset=dataset.map(parse,num_parallel_calls=64)\ndataset=dataset.shuffle(25000)\ndataset=dataset.batch(batch_size)\ndataset=dataset.prefetch(16)\nreturn dataset\n<\/code><\/pre>\n\n<p>4) declared the generator function like:<\/p>\n\n<pre><code>def generator(z, out_channel_dim,keep_prob,is_train=True, alpha=0.2):\nwith tf.variable_scope('generator'):\n\n\n\n    fc = tf.layers.dense(z, 7*7*128, use_bias=False)\n\n    fc = tf.reshape(fc, (-1, 7, 7, 128))\n    bn0 = tf.layers.batch_normalization(fc)\n    lrelu0 = tf.maximum(alpha * bn0, bn0)\n\n\n    conv1 =  tf.layers.conv2d_transpose(lrelu0, 512, 2, 1, 'valid', use_bias=False)\n    bn1 = tf.layers.batch_normalization(conv1)\n    lrelu1 = tf.maximum(alpha * bn1, bn1)\n\n\n\n    conv2 = tf.layers.conv2d_transpose(lrelu1, 256, 3, 2, 'same', use_bias=False)\n    bn2 = tf.layers.batch_normalization(conv2)\n    lrelu2 = tf.maximum(alpha * bn2, bn2)\n\n\n\n    conv3 = tf.layers.conv2d_transpose(lrelu2, 128, 3, 2, 'same')\n    bn3 = tf.layers.batch_normalization(conv3)\n    lrelu3 = tf.maximum(alpha * bn3, bn3)\n\n\n\n    logits = tf.layers.conv2d_transpose(lrelu3, 3, 3, 2, 'same')\n\n    out = tf.tanh(logits,name='image')\n\n    return out\n<\/code><\/pre>\n\n<p>5) then the discriminator:<\/p>\n\n<pre><code>def discriminator(images, reuse=False, alpha=0.2):\n\nwith tf.variable_scope('discriminator', reuse=reuse):\n\n\n    conv1 = tf.layers.conv2d(images, 128, 4, 2, padding='same')\n    lrelu1 = tf.maximum(alpha * conv1, conv1)\n\n\n    conv2 = tf.layers.conv2d(lrelu1, 256, 3, 1, 'valid')\n    bn2 = tf.layers.batch_normalization(conv2)\n    lrelu2 = tf.maximum(alpha * bn2, bn2)\n\n\n\n    conv3 = tf.layers.conv2d(lrelu2,256, 2, 1, 'valid')\n    bn3 = tf.layers.batch_normalization(conv3)\n    lrelu3 = tf.maximum(alpha * bn3, bn3)\n\n\n\n    flat = tf.layers.flatten(lrelu3)\n    logits = tf.layers.dense(flat, 1)\n    out = tf.sigmoid(logits)\n    return out,logits\n<\/code><\/pre>\n\n<p>6) The loss function:<\/p>\n\n<pre><code>def model_loss(input_real, input_z, out_channel_dim,keep_prob, alpha=0.2, smooth_factor=0.1):\nwith tf.name_scope(\"loss\"):\n    d_model_real, d_logits_real = discriminator(input_real, alpha=alpha)\n\n    d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real,labels=tf.ones_like(d_model_real) * (1 - smooth_factor)))\n\n    input_fake = generator(input_z, out_channel_dim, keep_prob, alpha=alpha)\n\n    d_model_fake, d_logits_fake = discriminator(input_fake, reuse=True, alpha=alpha)\n\n    d_loss_fake = tf.reduce_mean(\n        tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.zeros_like(d_model_fake)))\n\n    g_loss = tf.reduce_mean(\n        tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.ones_like(d_model_fake)),name='g_loss')\n\n    final_d_loss=tf.add(d_loss_real,d_loss_fake,name='d_loss')\n\n    tf.summary.scalar(\"d_loss\",final_d_loss)\n    tf.summary.scalar(\"g_loss\",g_loss)\n    merged=tf.summary.merge_all()\n    return final_d_loss, g_loss,input_fake,merged\n<\/code><\/pre>\n\n<p>The loss after training looks like this:\n<a href=\"https:\/\/i.stack.imgur.com\/C9div.png\" rel=\"nofollow noreferrer\">Discriminator Loss<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/5GO9t.png\" rel=\"nofollow noreferrer\">Generator Loss<\/a><\/p>\n\n<p>I have already tried batch_norm in input_real_image, drop-out, changing the learning rate of generator and dicriminator nothing meaningful comes out only noise.:<a href=\"https:\/\/i.stack.imgur.com\/KlcRp.png\" rel=\"nofollow noreferrer\">GAN Output<\/a><\/p>\n\n<p>Note: discriminator loss doesn't go to zero it's around 0.5 in case u wanna suggest gradient vanish<\/p>\n","answers":[]}
+{"QuestionId":53032377,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning>","CreationDate":"2018-10-28T14:07:33.443","AcceptedAnswerId":null,"OwnerUserId":8751904.0,"Title":"validation accuracy is not increasing training ResNet50","Body":"<p>I was doing fine-tuning with ResNet50 model for face recognition using data agumentation, but observed that the model accuracy was increasing but validation accuracy from the very starting point is not imporving, i am not getting where is it getting wrong, please review my code.  <\/p>\n\n<p>I have tried manipulating the top layers which i have added but it didn't helped. <\/p>\n\n<pre><code>import os\nos.environ['KERAS_BACKEND'] = 'tensorflow'\nfrom keras.applications import ResNet50\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Flatten, GlobalAveragePooling2D,Input,Dropout\n\nnum_classes = 13\n\nbase = ResNet50(include_top=False, weights='resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5',input_tensor=Input(shape=(100,100,3)))\nfrom keras.models import Model\n\nx = base.output\n\n#x = GlobalAveragePooling2D()(x)\n\nx = Flatten()(x)\n\nx = Dense(1024, activation = 'relu')(x)\n\nx = Dropout(0.5)(x)\n\npredictions = Dense(13, activation='softmax')(x)\n\nmodel = Model(inputs=base.input, outputs=predictions)\n\nfor layers in base.layers:\n    layers.trainable= False\n\nmodel.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_generator = ImageDataGenerator(featurewise_center=True,\n                                rotation_range=20,\n                                rescale=1.\/255,\n                                shear_range=0.2,\n                                zoom_range=0.2,\n                                width_shift_range=0.2,\n                                height_shift_range=0.2,\n                                horizontal_flip=True)\n\ntest_generator = ImageDataGenerator(rescale=1.\/255)\n\nfrom sklearn.model_selection import train_test_split\n\nx_train,x_test,y_train,y_test = train_test_split(image,label,test_size=0.2,shuffle=True,random_state=0)\n\ntrain_generator.fit(x_train)\ntest_generator.fit(x_test)\n\nmodel.fit_generator(train_generator.flow(x_train,y_train,batch_size=32),\n                       steps_per_epoch =10,epochs=50, \n                    validation_data=test_generator.flow(x_test,y_test))\n<\/code><\/pre>\n\n<p>Output :<\/p>\n\n<pre><code>Epoch 19\/50\n10\/10 [==============================] - 105s 10s\/step - loss: 1.9387 - acc: 0.3803 - val_loss: 2.6820 - val_acc: 0.0709\nEpoch 20\/50\n10\/10 [==============================] - 107s 11s\/step - loss: 2.0725 - acc: 0.3230 - val_loss: 2.6689 - val_acc: 0.0709\nEpoch 21\/50\n10\/10 [==============================] - 103s 10s\/step - loss: 1.8884 - acc: 0.3375 - val_loss: 2.6677 - val_acc: 0.0709\nEpoch 22\/50\n10\/10 [==============================] - 95s 10s\/step - loss: 1.8265 - acc: 0.4051 - val_loss: 2.6799 - val_acc: 0.0709\nEpoch 23\/50\n10\/10 [==============================] - 100s 10s\/step - loss: 1.8346 - acc: 0.3812 - val_loss: 2.6929 - val_acc: 0.0709\nEpoch 24\/50\n10\/10 [==============================] - 102s 10s\/step - loss: 1.9547 - acc: 0.3352 - val_loss: 2.6952 - val_acc: 0.0709\nEpoch 25\/50\n10\/10 [==============================] - 104s 10s\/step - loss: 1.9472 - acc: 0.3281 - val_loss: 2.7168 - val_acc: 0.0709\nEpoch 26\/50\n10\/10 [==============================] - 103s 10s\/step - loss: 1.8818 - acc: 0.4063 - val_loss: 2.7071 - val_acc: 0.0709\nEpoch 27\/50\n10\/10 [==============================] - 106s 11s\/step - loss: 1.8053 - acc: 0.4000 - val_loss: 2.7059 - val_acc: 0.0709\nEpoch 28\/50\n10\/10 [==============================] - 104s 10s\/step - loss: 1.9601 - acc: 0.3493 - val_loss: 2.7104 - val_acc: 0.0709\n<\/code><\/pre>\n","answers":[{"AnswerId":"53036787","CreationDate":"2018-10-28T22:56:13.647","ParentId":null,"OwnerUserId":"8751904","Title":null,"Body":"<p>It was happening because i was just directly adding the fully-connected layers without training it first, as mentioned in Blog of keras, \n<a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html<\/a><\/p>\n\n<blockquote>\n  <p>in order to perform fine-tuning, all layers should start with properly trained weights: for instance you should not slap a randomly initialized fully-connected network on top of a pre-trained convolutional base. This is because the large gradient updates triggered by the randomly initialized weights would wreck the learned weights in the convolutional base. In our case this is why we first train the top-level classifier, and only then start fine-tuning convolutional weights alongside it.<\/p>\n<\/blockquote>\n\n<p>so answer is first train the top-model separately, then create a new model having ResNet50 model with its weight, with top-model and its weights on top of resnet model(base-model), then train it first by freezing the base-model(ResNet50) and the with the last layer of the base-model.<\/p>\n"}]}
+{"QuestionId":53032412,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T14:11:19.180","AcceptedAnswerId":54388906.0,"OwnerUserId":10484131.0,"Title":"tf.scan obscures Tensor's shape","Body":"<p>I want to do some computation within a <code>tf.scan<\/code> loop. My problem is, that it seems that the shape of <code>tf.slice<\/code> seems to either be unknown within the <code>tf.scan<\/code> loop, or unknown compared to to outside of the <code>tf.scan<\/code> loop. The MWE I came up with looks like this:<\/p>\n\n<pre><code>def compute(x, _):\n    i = x[0]\n    a = x[1]\n    print(tf.slice(a, [0], [i]).shape) # First line\n    return (i + 1, a)\n\na = tf.constant([1, 2, 3, 4, 5])\nwith tf.Session() as sess:\n    print(tf.slice(a, [0], [tf.constant(1)]).shape) # Second line\n    x = tf.scan(\n        compute,\n        tf.zeros(a.shape),\n        initializer=(tf.constant(1, tf.int32), a)\n    )\n<\/code><\/pre>\n\n<p>Which returns the two following lines:<\/p>\n\n<pre><code>(1,)\n(?,)\n<\/code><\/pre>\n\n<p>Why does <code>tf.slice<\/code> return <code>(?,)<\/code> within the <code>compute<\/code> function? And is there a way to cast the slice into a specific shape?<\/p>\n","answers":[{"AnswerId":"54388906","CreationDate":"2019-01-27T13:57:58.843","ParentId":null,"OwnerUserId":"10974492","Title":null,"Body":"<p>The reason why <code>tf.slice<\/code> returns a Tensor of shape <code>(?,)<\/code> is simple. This method has the following signature:<\/p>\n\n<pre><code>tf.slice(input_, begin, size, name=None)\n<\/code><\/pre>\n\n<p>The pythonic way to rewrite the call <code>tf.slice(a, [0], [i])<\/code> is <code>a[0:i]<\/code>.<\/p>\n\n<p>At each iteration of <code>tf.scan<\/code>, tensor <code>i<\/code> inside <code>compute<\/code> is derived from the results of the previous iteration, may change and have any arbitrary value. During the compile time, TensorFlow is not able to infer the exact shape of the slice and returns <code>(?,)<\/code>.<\/p>\n\n<p>In your case, at each iteration <code>i<\/code> is increased and you have slices of various shapes. However, if you do not update <code>i<\/code> and have slices of the same size, the output will still be <code>(?,)<\/code> because the value for <code>i<\/code> comes from outside and previous iterations.<\/p>\n\n<p>You do not really need to cast the shape of the slice if you are confident in its shape during the runtime. For example, having a varying shape at each iteration as in your case. The main rule with <code>tf.scan<\/code> is to have consistent shapes for the output and initializer tensors, otherwise a <code>ValueError<\/code> is thrown.<\/p>\n"}]}
+{"QuestionId":53032586,"AnswerCount":2,"Tags":"<python><tuples><lstm><pytorch><torch>","CreationDate":"2018-10-28T14:30:39.170","AcceptedAnswerId":53035016.0,"OwnerUserId":2253003.0,"Title":"AttributeError: 'tuple' object has no attribute 'dim', when feeding input to Pytorch LSTM network","Body":"<p>I am trying to run the following code:<\/p>\n\n<pre><code>import matplotlib.pylab as plt\nimport numpy as np\nimport torch\nimport torch.nn as nn\n\nclass LSTM(nn.Module):\n    def __init__(self, input_shape, n_actions):\n        super(LSTM, self).__init__()\n\n        self.lstm = nn.LSTM(input_shape, 12)\n        self.hidden2tag = nn.Linear(12, n_actions)\n\n    def forward(self, x):\n        out = self.lstm(x)\n        out = self.hidden2tag(out)\n        return out\n\n\nstate = [(1,2,3,4,5),(2,3,4,5,6),(3,4,5,6,7),(4,5,6,7,8),(5,6,7,8,9),(6,7,8,9,0)]\n\ndevice = torch.device(\"cuda\")\nnet = LSTM(5, 3).to(device)\n\nstate_v = torch.FloatTensor(state).to(device)\n\nq_vals_v = net(state_v.view(1, state_v.shape[0], state_v.shape[1]))\n_, action = int(torch.max(q_vals_v, dim=1).item())\n<\/code><\/pre>\n\n<p>And that returns this error:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Traceback (most recent call last):\n  File \"\/home\/dikkerj\/Documents\/PycharmProjects\/LSTMReactor\/QuestionStackoverflow.py\", line 26, in &lt;module&gt;\n    q_vals_v = net(state_v.view(1, state_v.shape[0], state_v.shape[1]))\n  File \"\/home\/dikkerj\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/dikkerj\/Documents\/PycharmProjects\/LSTMReactor\/QuestionStackoverflow.py\", line 15, in forward\n    out = self.hidden2tag(out)\n  File \"\/home\/dikkerj\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/dikkerj\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/linear.py\", line 55, in forward\n    return F.linear(input, self.weight, self.bias)\n  File \"\/home\/dikkerj\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/functional.py\", line 1022, in linear\n    if input.dim() == 2 and bias is not None:\nAttributeError: 'tuple' object has no attribute 'dim'\n<\/code><\/pre>\n\n<p>Anyone knows how to fix this? (to get rid of the tensor being a tuple so that it can be fed into the LSTM network)<\/p>\n","answers":[{"AnswerId":"53032985","CreationDate":"2018-10-28T15:12:50.197","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>Transform your state in a numpy array first:<\/p>\n\n<pre><code>state = np.array(state)\n<\/code><\/pre>\n\n<p>PyTorch is probably missing a <code>np.asarray<\/code> in their API.<\/p>\n"},{"AnswerId":"53035016","CreationDate":"2018-10-28T18:55:46.163","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The pytorch LSTM returns a tuple.<br> So you get this error as your linear layer <code>self.hidden2tag<\/code> can not handle this tuple.<\/p>\n\n<p>So change:<\/p>\n\n<pre><code>out = self.lstm(x)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>out, states = self.lstm(x)\n<\/code><\/pre>\n\n<p>This will fix your error, by splitting up the tuple so that <code>out<\/code> is just your output tensor.<\/p>\n\n<p><code>out<\/code> then stores the hidden states, while <code>states<\/code> is another tuple that contains the last hidden and cell state. <\/p>\n\n<p><em>You can also take a look here:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.LSTM\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.LSTM<\/a><\/em><\/p>\n\n<p><em>You will get another error for the last line as <code>max()<\/code> returns a tuple as well. But this should be easy to fix and is yet different error :)<\/em><\/p>\n"}]}
+{"QuestionId":53032922,"AnswerCount":1,"Tags":"<python><tensorflow><while-loop>","CreationDate":"2018-10-28T15:05:05.127","AcceptedAnswerId":53033052.0,"OwnerUserId":2260153.0,"Title":"TensorFlow while loop with condition dependent on body","Body":"<p>I want to have a while loop with the condition dependent on a tensor computed in the loop body, but I don't know how to accomplish this with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop()<\/code><\/a>.<\/p>\n\n<p>My input processing includes random cropping, but some crops can lead to low-quality examples and I want to discard those and try a new random crop until an example of sufficient quality is obtained. The inputs are cropped by<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nIMAGE_SHAPE = [960, 720]\nCROP_SHAPE = [320, 240]\nmax_begin_index = np.array(IMAGE_SHAPE) - np.array(CROP_SHAPE)\ncrop_begin_index = tf.round(tf.random_uniform([2]) * max_begin_index)\nimg_crop = tf.slice(img, crop_begin_index, crop_shape + [-1])\n<\/code><\/pre>\n\n<p>and the condition is<\/p>\n\n<pre><code>cond = tf.count_nonzero(img_crop &gt; 0) &gt; 0.5 * tf.size(img_crop)\n<\/code><\/pre>\n\n<p>Going over the documentation and examples of <code>tf.while_loop(cond, body, loop_vars, ...)<\/code>, what I understand is that both <code>cond<\/code> and <code>body<\/code> should take the same arguments given in <code>loop_vars<\/code>.\nI don't see how I can have <code>cond<\/code> depend on <code>img_crop<\/code> which would be calculated inside <code>body<\/code>, and isn't provided in <code>loop_vars<\/code>.<\/p>\n\n<p>I could equivalently compute <code>cond<\/code> using <code>crop_begin_index<\/code> without actually cropping, but it depends on the random values computed inside the loop, so I have the same problem.<\/p>\n\n<p>Is this indeed a limitation of TF looping? If not, how can I rewrite my code to use <code>tf.while_loop()<\/code>?<\/p>\n","answers":[{"AnswerId":"53033052","CreationDate":"2018-10-28T15:19:04.743","ParentId":null,"OwnerUserId":"4101431","Title":null,"Body":"<p>The arguments that are passed on to the <code>condition<\/code> function are the arguments returned from your <code>body<\/code> function. So you just have to return that value that you want to base your condition on in the <code>body<\/code> function, then carry out the condition on that value in your <code>cond<\/code> function. Something like, <\/p>\n\n<pre><code>def body(image_shape, crop_shape, img_crop):\n    max_begin_index = np.array(IMAGE_SHAPE) - np.array(CROP_SHAPE)\n    crop_begin_index = tf.round(tf.random_uniform([2]) * max_begin_index)\n    img_crop = tf.slice(img, crop_begin_index, crop_shape + [-1])\n    return (image_shape, crop_shape, img_crop)\n\ndef cond(image_shape, crop_shape, img_crop):\n    return tf.count_nonzero(img_crop &gt; 0) &gt; 0.5 * tf.size(img_crop)\n\nimage_shape, crop_shape, img_crop = tf.while_loop(cond=cond, body=body, loop_vars=([960, 720], [320, 240], img_crop))\n<\/code><\/pre>\n\n<p>Don't have access to an interpreter right now, so there might be some syntax problems there, but something like that. <\/p>\n\n<p>Also, if I recall correctly, the body and the condition need to be pure functions, you cannot alter the outer state from within the functions.<\/p>\n\n<p>Also note, you'll need to specify some initial value for <code>img_crop<\/code> in the loop vars.<\/p>\n\n<p>Moreover, by default, <code>tf.while_loop<\/code> expects the shapes of all the <code>loop_vars<\/code> to remain the same across all loop runs. You can modify this through the <code>shape_invariants<\/code> argument. <\/p>\n"}]}
+{"QuestionId":53033738,"AnswerCount":1,"Tags":"<python><keras><lstm><embedding>","CreationDate":"2018-10-28T16:37:04.210","AcceptedAnswerId":53033903.0,"OwnerUserId":6266336.0,"Title":"keras. Initializing a bidirecional LSTM. Passing word embeddings","Body":"<p>In the implementation i am using, the lstm is initialized in the following way:<\/p>\n\n<pre><code>l_lstm = Bidirectional(LSTM(64, return_sequences=True))(embedded_sequences)\n<\/code><\/pre>\n\n<p>What i don't really understand and it might be because of the lack of experience in Python generally: the notation <code>l_lstm= Bidirectional(LSTM(...))(embedded_sequences)<\/code>. \nI don't get what i am passing the <code>embedded_sequences<\/code> to? Because it is not a parameter of <code>LSTM()<\/code> but also does not seem to be an argument for <code>Bidirectional()<\/code> as it stands separately. <\/p>\n\n<p>Here is the documentation for Bidirectional:<\/p>\n\n<pre><code>  def __init__(self, layer, merge_mode='concat', weights=None, **kwargs):\n        if merge_mode not in ['sum', 'mul', 'ave', 'concat', None]:\n            raise ValueError('Invalid merge mode. '\n                             'Merge mode should be one of '\n                             '{\"sum\", \"mul\", \"ave\", \"concat\", None}')\n        self.forward_layer = copy.copy(layer)\n        config = layer.get_config()\n        config['go_backwards'] = not config['go_backwards']\n        self.backward_layer = layer.__class__.from_config(config)\n        self.forward_layer.name = 'forward_' + self.forward_layer.name\n        self.backward_layer.name = 'backward_' + self.backward_layer.name\n        self.merge_mode = merge_mode\n        if weights:\n            nw = len(weights)\n            self.forward_layer.initial_weights = weights[:nw \/\/ 2]\n            self.backward_layer.initial_weights = weights[nw \/\/ 2:]\n        self.stateful = layer.stateful\n        self.return_sequences = layer.return_sequences\n        self.return_state = layer.return_state\n        self.supports_masking = True\n        self._trainable = True\n        super(Bidirectional, self).__init__(layer, **kwargs)\n        self.input_spec = layer.input_spec\n        self._num_constants = None\n<\/code><\/pre>\n","answers":[{"AnswerId":"53033903","CreationDate":"2018-10-28T16:52:36.407","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>Let's try to break down what is going on:<\/p>\n\n<ol>\n<li>You start with <code>LSTM(...)<\/code> which creates an LSTM <em>Layer<\/em>. Now <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">layers in Keras are callable<\/a> which means you can use them like functions. For example <code>lstm = LSTM(...)<\/code> and then <code>lstm(some_input)<\/code> will <em>call<\/em> the LSTM on the given input tensor.<\/li>\n<li>The <code>Bidirectional(...)<\/code> <em>wraps<\/em> any RNN layer and returns you <em>another layer<\/em> that when called applies the wrapped layer in both directions. So <code>l_lstm = Bidirectional(LSTM(...))<\/code> is a layer when called with some input will apply the <code>LSTM<\/code> in both direction. <em>Note:<\/em> Bidirectional creates a copy of passed LSTM layer, so backwards and forwards are different LSTMs.<\/li>\n<li>Finally, when you call <code>Bidirectional(LSTM(...))(embedded_seqences)<\/code> bidirectional layer takes the input sequences, passes it to the wrapped LSTMs in both directions, collects their output and concatenates it.<\/li>\n<\/ol>\n\n<p>To understand more about layers and their callable nature, you can look at the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">functional API guide<\/a> of the documentation.<\/p>\n"}]}
+{"QuestionId":53033767,"AnswerCount":1,"Tags":"<python><tensorflow><printing><while-loop><tensor>","CreationDate":"2018-10-28T16:39:40.150","AcceptedAnswerId":53053554.0,"OwnerUserId":2221231.0,"Title":"How to print a tensor's value inside tf.while_loop without returning it?","Body":"<p>What I want is to print a tensor's value inside tf.while_loop body without returning the tensor but still using a computational graph. Below I have some simple examples to explain what I want to succeed and what I have done so far.<\/p>\n\n<p>Method 1 (works): <\/p>\n\n<p>TF supports an option to print tensors while evaluating a model by introducing tf.Print operation to the graph, but that requires for the tensor to be returned from the body : <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.placeholder(tf.float32, [1])\n\ndef body(x):\n    a = tf.constant( np.array([2]) , dtype=tf.float32)\n    x = a + x\n    x = tf.Print(x,[x], summarize=100) &lt;= print here (works)\n    return x\n\ndef condition(x):\n    return tf.reduce_sum(x) &lt; 10\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    result = tf.while_loop(cond=condition, body=body, loop_vars=[x])\n    result_out = sess.run([result], feed_dict={ x : np.zeros(1)})\n    print(result_out)\n<\/code><\/pre>\n\n<p>Out :<\/p>\n\n<pre><code>[2]\n[4]\n[6]\n[8] \n[10]\n[array([10.], dtype=float32)] \n<\/code><\/pre>\n\n<p>Method 2 (works):  <\/p>\n\n<p>TF supports an option to print tensors without creating a computational graph using Eager mode. Same example below : <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntf.enable_eager_execution()\n\ndef body(x):\n    a = tf.constant(np.array([2]), dtype=tf.int32)\n    x = a + x\n    print(x) &lt;= print here (works)\n    return x\n\ndef cond(x):\n    return tf.reduce_sum(x) &lt; 10 # sum over an axis\n\nx = tf.constant(0, shape=[1])\n\n#result = tf.while_loop(cond=condition, body=body, loop_vars=(x,0))\nresult=tf.while_loop(cond, body, [x])\nprint(result)\n<\/code><\/pre>\n\n<p>Out :<\/p>\n\n<pre><code>tf.Tensor([2], shape=(1,), dtype=int32)\ntf.Tensor([4], shape=(1,), dtype=int32)\ntf.Tensor([6], shape=(1,), dtype=int32)\ntf.Tensor([8], shape=(1,), dtype=int32)\n\ntf.Tensor([10], shape=(1,), dtype=int32)\ntf.Tensor([10], shape=(1,), dtype=int32) \n<\/code><\/pre>\n\n<p>Method 3 (failed):<\/p>\n\n<p>What I want is to print the tensor using eager execution in a graph environment (like described : <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">here<\/a>).  <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntfe = tf.contrib.eager\n\nx = tf.placeholder(tf.int32, [1])\n\ndef my_py_func(x):\n  print(x)  # It's eager!\n\ndef body(x):\n    a = tf.constant( np.array([2]) , dtype=tf.int32)\n    x = a + x\n    tfe.py_func(my_py_func, x, tf.int32) &lt;= print here (does not work)\n    return x\n\ndef condition(x):\n    return tf.reduce_sum(x) &lt; 10\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    result = tf.while_loop(condition, body, [x])\n    result_out = sess.run([result], feed_dict={ x : np.zeros(1)} )\n    print(result_out)\n<\/code><\/pre>\n\n<p>Out :<\/p>\n\n<pre><code>TypeError: Expected list for 'input' argument to 'EagerPyFunc' Op, not Tensor(\"while\/add:0\", shape=(1,), dtype=int32).\n<\/code><\/pre>\n\n<p>Of course in this example I return tensor <code>x<\/code> from the body, but I want to print inside the loop !<\/p>\n","answers":[{"AnswerId":"53053554","CreationDate":"2018-10-29T20:46:33.153","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>In the first method, you can print the value of a tensor without returning it. For example:<\/p>\n\n<pre><code>x = tf.Print(x, [a])\n<\/code><\/pre>\n\n<p>In this case, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\">tf.Print<\/a> is an identity operation with the side effect of printing the value of tensor <code>a<\/code> when evaluating.<\/p>\n"}]}
+{"QuestionId":53033780,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2018-10-28T16:40:41.387","AcceptedAnswerId":53035991.0,"OwnerUserId":868829.0,"Title":"Indexing Ops in TF","Body":"<p>Is there a way to index an operation in tensorflow? In particular, I'm interested in indexing by the iterator variable of <code>tf.while_loop<\/code>.<\/p>\n\n<p>More concretely, let's say I have <code>my_ops = [op1, op2]<\/code>. I would like to have:<\/p>\n\n<pre><code>my_ops = [...]\ni = tf.constant(0)\nc = lambda i: tf.less(i, 10)\nb = lambda i: my_ops[i](...)\nr = tf.while_loop(c, b, [i])\n<\/code><\/pre>\n\n<p>which unfortunately will not work, since a python array supports only integer indexing.<\/p>\n","answers":[{"AnswerId":"53035991","CreationDate":"2018-10-28T20:55:57.970","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>I believe this is not possible. However, you could instead use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"nofollow noreferrer\">tf.stack<\/a> to stack the operations' output tensors, and then use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\">tf.gather<\/a> to obtain the desired output.<\/p>\n\n<p>Here you have an example:<\/p>\n\n<pre><code>import tensorflow as tf\n\n\ndef condition(i, x):\n    return tf.less(i, 10)\n\n\ndef body_1(my_ops):\n    def b(i, x):\n        stacked_results = tf.stack([op(x) for op in my_ops])\n        gather_idx = tf.mod(i, 2)\n        return [i + 1, tf.gather(stacked_results, gather_idx)]\n\n    return b\n\n\ndef body_2(my_ops):\n    def b(i, x):\n        nb_ops = len(my_ops)\n        pred_fn_pairs = [(tf.equal(tf.mod(i, nb_ops), 0), lambda: my_ops[0](x)),\n                         (tf.equal(tf.mod(i, nb_ops), 1), lambda: my_ops[1](x))]\n        result = tf.case(pred_fn_pairs)\n        return [i + 1, result]\n\n    return b\n\n\nmy_ops = [lambda x: tf.Print(x + 1, [x, 1]),\n          lambda x: tf.Print(x + 2, [x, 2])]\ni = tf.constant(0)\nx = tf.constant(0)\nr = tf.while_loop(condition, body_2(my_ops), [i, x])  # See the difference with body_1\n\nwith tf.Session() as sess:\n    i, x = sess.run(r)\n    print(x)  # Prints 15 = 5*2 + 5*1\n<\/code><\/pre>\n"}]}
+{"QuestionId":53033953,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-28T16:58:05.273","AcceptedAnswerId":53035071.0,"OwnerUserId":9524145.0,"Title":"Searching for a CNN Time Series Prediction Tutorial","Body":"<p>I'm looking for instructions on how to make a regression time series prediction using a CNN. I want to implement a multi-step prediction for a univariate time series. I have read a few instructions but found nothing suitable for my dataset: one feature and around 400 observations.<\/p>\n\n<p>Does anyone know an easily understandable and applicable code example for such a time series?<\/p>\n\n<p>I would be very grateful for any help,<\/p>\n\n<p>Leon<\/p>\n","answers":[{"AnswerId":"53035071","CreationDate":"2018-10-28T19:01:28.670","ParentId":null,"OwnerUserId":"9415043","Title":null,"Body":"<p>Using CNNs for sequence data can be a bit tricky to set up. In my experience, CNNs achieve results near RNNs (GRUs and LSTMs) but CNNs are far faster to compute.<\/p>\n\n<p>First, make sure your data is shaped the way <code>Conv1D<\/code> expects: <code>(instances, time steps, predictors)<\/code>.<\/p>\n\n<pre><code>X_cnn = X.reshape(X.shape[0], X.shape[1] \/\/ predictors, predictors)\n<\/code><\/pre>\n\n<p>Then, the syntax is:<\/p>\n\n<pre><code>model_cnn = Sequential()\nmodel_cnn.add(layers.Conv1D(A, B, activation = 'relu', \n        input_shape = (X_cnn.shape[1], X_cnn.shape[2])))\nmodel_cnn.add(layers.Flatten())\nmodel_cnn.add(layers.Dense(1))\n<\/code><\/pre>\n\n<p>Where <code>A<\/code> is the number of neurons, and <code>B<\/code> is the number of time steps to consider. Note the <code>Flatten()<\/code> layer after the <code>Conv1D<\/code> layer. This should hopefully get you started.<\/p>\n"}]}
+{"QuestionId":53034013,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-28T17:03:50.050","AcceptedAnswerId":null,"OwnerUserId":1324967.0,"Title":"Tensorflow - how to duplicate an Estimator","Body":"<p>Using tensorflow, I have created an <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> named <code>my_estimator<\/code><\/p>\n\n<p>When I train this estimator with <code>my_estimator.train(...)<\/code> the weights inside of my estimator change and I get different results for <code>my_estimator.predict(...)<\/code>.<\/p>\n\n<p>Is there a way to save my estimator model just before training into another variable named <code>old_estimator<\/code> so that I can make predictions before training?  I want this operation to be fast, so it would be nice if I could keep this in RAM.  <\/p>\n\n<p>I'm thinking something like this pseudocode should work:<\/p>\n\n<pre><code># Do a prediction with our estimator\nprediction = next(my_estimator.predict(foo))\n\n# Copy the estimator - this is the function I'm looking for\nold_estimator = my_estimator.copy()\n\n# Train our estimator\nmy_estimator.train(...)\n\n# Should be true\nnext(old_estimator.predict(foo)) == prediction\n\n# Should be false\nnext(old_estimator.predict(foo)) == next(my_estimator.predict(foo))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53034757,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-10-28T18:24:57.790","AcceptedAnswerId":53035535.0,"OwnerUserId":3946761.0,"Title":"Keras Input Shape Issue","Body":"<p>I can find many questions and answers related to my question but somehow they did not solve my problem. I have data with shape (10584, 56) and specified input_shape=(10584,56) in the code but getting following error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected dense_1_input to have 3 dimensions, but got array with shape (10584, 56).\n<\/code><\/pre>\n\n<p>I have somehow idea that I have to reshape my input data frame but not sure how. Following is my code:<\/p>\n\n<pre><code>y = df['Target']\nx_train, x_test, y_train, y_test = train_test_split(df, y, test_size=0.2)\nmodel = keras.models.Sequential()\nmodel.add(keras.layers.Dense(64,input_shape(10584,56),activation='relu'))\n<\/code><\/pre>\n\n<p>Any help\/suggestion will be much appreciated.<\/p>\n","answers":[{"AnswerId":"53035535","CreationDate":"2018-10-28T19:56:20.873","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>There is always an additional dimension for the batch size that you need add even if you want to use a batch size of 1.<\/p>\n\n<p>Another possibility: If in fact your samples are not 2d vectors but 1d vectors of size 64 and 10584 is the number of samples you have, than the number of samples is not part of the input shape. You only provide the size of a single sample. Keras will take care of splitting your data into batches and setting the network up the right way.<\/p>\n"}]}
+{"QuestionId":53034884,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-28T18:38:57.947","AcceptedAnswerId":null,"OwnerUserId":1808384.0,"Title":"Logging accuracy when using tf.estimator.Estimator","Body":"<p>I'm following this tutorial - <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn<\/a> - to build a CNN using TensorFlow Estimators with the MNIST data.<\/p>\n\n<p>I would like to visualize training accuracy and loss at each step, but I'm not sure how to do that with <code>tf.train.LoggingTensorHook<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":53035190,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-10-28T19:13:18.457","AcceptedAnswerId":null,"OwnerUserId":3755087.0,"Title":"ValueError: Error when checking target: expected dense_108 to have 2 dimensions, but got array with shape (36020, 10, 2)","Body":"<p>I'm building a convolutional neural network with <code>Convolution1D<\/code> layer. My network model is given below. The input at dense layer seems to produce an array of shape <code>(36020,10,2)<\/code>.<\/p>\n\n<pre><code>#network model\ncnn = Sequential()\ncnn.add(Convolution1D(64, 3, border_mode=\"same\",activation=\"relu\",input_shape=(25,1)))\ncnn.add(MaxPooling1D(pool_length=(2)))\ncnn.add(Flatten())\ncnn.add(Dense(128, activation=\"relu\"))\ncnn.add(Dropout(0.5))\ncnn.add(Dense(2, activation=\"softmax\"))\n<\/code><\/pre>\n\n<p>The data I'm trying to fit the model on is:    <\/p>\n\n<pre><code>X_train=[[[1.0000000e+00]\n  [3.0122564e-08]\n  [1.6120090e-05]\n  ...\n  [0.0000000e+00]\n  [9.4886076e-08]\n  [3.0170717e-08]]\n\n [[1.0000000e+00]\n  [0.0000000e+00]\n  [0.0000000e+00]\n  ...\n  [0.0000000e+00]\n  [0.0000000e+00]\n  [1.2500001e-12]]\n\n [[1.0000000e+00]\n  [0.0000000e+00]\n  [0.0000000e+00]\n  ...\n  [0.0000000e+00]\n  [0.0000000e+00]\n  [3.1249999e-11]]\n\n ...\n\n [[0.0000000e+00]\n  [1.0842798e-05]\n  [1.0943735e-06]\n  ...\n  [0.0000000e+00]\n  [9.6288932e-09]\n  [1.3172292e-10]]\n\n [[0.0000000e+00]\n  [2.8011250e-01]\n  [8.8251436e-01]\n  ...\n  [0.0000000e+00]\n  [4.1974179e-04]\n  [3.6202004e-04]]\n\n [[0.0000000e+00]\n  [8.3799750e-06]\n  [9.5839296e-06]\n  ...\n  [0.0000000e+00]\n  [8.8683461e-09]\n  [1.0194775e-10]]]\n\ny_train = [[[0. 1.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]\n\n [[0. 1.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]\n\n [[0. 1.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]\n\n ...\n\n [[1. 0.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]\n\n [[1. 0.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]\n\n [[1. 0.]\n  [1. 0.]\n  [1. 0.]\n  ...\n  [1. 0.]\n  [1. 0.]\n  [1. 0.]]]\n<\/code><\/pre>\n\n<p>I keep getting the error about the dense layer dimension. I'm really new to neural net programming.<\/p>\n","answers":[{"AnswerId":"53035580","CreationDate":"2018-10-28T20:01:48.057","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that the output shape of model, <code>(None, 2)<\/code>, is inconsistent with the shape of labels array, <code>(36020, 10, 2)<\/code>, you provide to it when training the model. Either you need to change the shape of labels array to <code>(num_samples, 2)<\/code> or just change the layers' parameters and architecture of the model to make it have an output shape of <code>(None, 10, 2)<\/code> (i.e. to be consistent with <code>(36020, 10, 2)<\/code>). I can't comment further as to which one is the correct way to go, since it entirely depends on what the real input and output shapes of the problem you are working on is and you have given no information in this regard.<\/p>\n"}]}
+{"QuestionId":53035555,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-28T19:58:55.200","AcceptedAnswerId":53035607.0,"OwnerUserId":10230859.0,"Title":"index error while trying to fit model keras","Body":"<p>i'm trying to create an autoencoder with keras, here's my code:<\/p>\n\n<pre><code>from keras import models, layers\nfrom numpy import array\nimport random\n\ndata = array(\n    [array([[random.randint(0, 100) for i in range(50)]]) for i in range(500)]\n).reshape((500, 50))\nmodel = models.Sequential()\nmodel.add(layers.Dense(input_dim=50, units=50, activation=\"sigmoid\"))\nmodel.add(layers.Dense(units=40, activation=\"sigmoid\"))\nmodel.add(layers.Dense(units=50, activation=\"sigmoid\"))\nmodel.compile(optimizer=\"adam\", loss=\"mean_squared_error\", metrics=[\"accuracy\"])\nmodel.fit(data, epochs=1)\n<\/code><\/pre>\n\n<p>and my error is :<\/p>\n\n<pre><code>Python\\Python36\\lib\\site-packages\\keras\\engine\\training_arrays.py\", line 139, in fit_loop\n    if issparse(ins[i]) and not K.is_sparse(feed[i]):\nIndexError: list index out of range\n<\/code><\/pre>\n","answers":[{"AnswerId":"53035607","CreationDate":"2018-10-28T20:04:50.503","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>You forgot to provide the target data. In you case its the same as the input data, but you still need to tell keras that. This line should work:<\/p>\n\n<pre><code>model.fit(data, data, epochs=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53035839,"AnswerCount":0,"Tags":"<r><keras><word2vec><word-embedding><glove>","CreationDate":"2018-10-28T20:36:26.100","AcceptedAnswerId":null,"OwnerUserId":7170259.0,"Title":"How to calculate similarity for pre-trained word embeddings","Body":"<p>I want to know the most similar words to another from a pretrained embedding vectors in <strong>R<\/strong>. E.g: words similar to \"beer\". For this, I download the pretrained embedding vectors on <a href=\"http:\/\/nlp.stanford.edu\/data\/glove.twitter.27B.zip\" rel=\"nofollow noreferrer\">http:\/\/nlp.stanford.edu\/data\/glove.twitter.27B.zip<\/a> and applied the code below:<\/p>\n\n<p>Source code:<\/p>\n\n<pre><code>glove_dir = \"~\/Downloads\/glove.6B\"\nlines &lt;- readLines(file.path(glove_dir, \"glove.6B.100d.txt\"))\nembeddings_index &lt;- new.env(hash = TRUE, parent = emptyenv())\nfor (i in 1:length(lines)) {\n    line &lt;- lines[[i]]\n    values &lt;- strsplit(line, \" \")[[1]]\n    word &lt;- values[[1]]\n    embeddings_index[[word]] &lt;- as.double(values[-1])\n}\ncat(\"Found\", length(embeddings_index), \"word vectors.\\n\")\n\nembedding_dim &lt;- 100\nembedding_matrix &lt;- array(0, c(max_words, embedding_dim))\nfor (word in names(word_index)) {\n  index &lt;- word_index[[word]]\n  if (index &lt; max_words) {\n    embedding_vector &lt;- embeddings_index[[word]]\n    if (!is.null(embedding_vector))\n      embedding_matrix[index+1,] &lt;- embedding_vector\n  }\n}\n<\/code><\/pre>\n\n<p>But I do not how to get the words most similar. \nI found the examples but doesn't work because the structure of embeddings vectors are different<\/p>\n\n<pre><code>find_similar_words &lt;- function(word, embedding_matrix, n = 5) {\n    similarities &lt;- embedding_matrix[word, , drop = FALSE] %&gt;%\n    sim2(embedding_matrix, y = ., method = \"cosine\")\n    similarities[,1] %&gt;% sort(decreasing = TRUE) %&gt;% head(n)\n}\nfind_similar_words(\"beer\", embedding_matrix)\n<\/code><\/pre>\n\n<p>How to calculate similarity for pre-trained word embeddings in <strong>R<\/strong>?<\/p>\n","answers":[]}
+{"QuestionId":53036313,"AnswerCount":1,"Tags":"<tensorflow><audio><direction>","CreationDate":"2018-10-28T21:37:04.400","AcceptedAnswerId":58304549.0,"OwnerUserId":4271452.0,"Title":"Tensorflow: how to detect audio direction","Body":"<p>I have a task: <strong>to determine the sound source location<\/strong>.<\/p>\n\n<p>I had some experience working with <code>tensorflow<\/code>, creating predictions on some simple features and datasets. I assume that for this task, there would be necessary to analyze the sound frequences and probably other related data on training and then prediction steps. The sound goes from the headset, so human ear is able to detect the direction.  <\/p>\n\n<p><em>1) Did somebody already perform that?<\/em> (unfortunately couldn't find any similar project)<br>\n<em>2) What kind of caveats could I meet while trying to achieve that?<\/em><br>\n<em>3) Am I able to do that using this technology approach? Are there any other sound processing frameworks \/ technologies \/ open source projects that could help me ?<\/em> <\/p>\n\n<p>I am asking that here, since my research on google, github, stackoverflow didn't show me any relevant results on that specific topic, so any help is highly appreciated!  <\/p>\n","answers":[{"AnswerId":"58304549","CreationDate":"2019-10-09T12:51:07.753","ParentId":null,"OwnerUserId":"2217539","Title":null,"Body":"<p>This is typically done with more traditional DSP with multiple sensors. You might want to look into time difference of arrival(TDOA) and direction of arrival(DOA). Algorithms such as GCC-PHAT and MUSIC will be helpful.<\/p>\n\n<p>Issues that you might encounter are: DOA accuracy is function of the direct to reverberant ratio of the source, i.e. the more reverberant the environment the harder it is to determine the source location.<\/p>\n\n<p>Also you might want to consider the number of location dimensions you want to resolve. A point in 3D space is much more difficult than a direction relative to the sensors<\/p>\n\n<p>Using ML as an approach to this is not entirely without merit but you will have to consider what it is you would be learning, i.e. you probably don't want to learn the test rooms reverberant properties but instead the sensors spatial properties.<\/p>\n"}]}
+{"QuestionId":53036403,"AnswerCount":1,"Tags":"<python><tensorflow><statistics><pytorch><tensorflow-probability>","CreationDate":"2018-10-28T21:50:25.623","AcceptedAnswerId":null,"OwnerUserId":8931942.0,"Title":"KL Divergence of Normal and Laplace isn't Implemented in TensorFlow Probability and PyTorch","Body":"<p>In both TensorFlow Probability (<code>v0.4.0<\/code>) and PyTorch (<code>v0.4.1<\/code>) the KL Divergence of the Normal distribution (<a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/distributions\/Normal\" rel=\"nofollow noreferrer\"><code>tfp<\/code><\/a>, <a href=\"https:\/\/pytorch.org\/docs\/stable\/distributions.html#normal\" rel=\"nofollow noreferrer\">PyTorch<\/a>) and the Laplace distribution (<a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/distributions\/Laplace\" rel=\"nofollow noreferrer\"><code>tfp<\/code><\/a>, <a href=\"https:\/\/pytorch.org\/docs\/stable\/distributions.html#laplace\" rel=\"nofollow noreferrer\">PyTorch<\/a>) isn't implemented resulting in a <code>NotImplementedError<\/code> error being thrown.<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; import tensorflow_probability as tfp\n&gt;&gt;&gt; tfd = tfp.distributions\n&gt;&gt;&gt; import torch\n&gt;&gt;&gt;\n&gt;&gt;&gt; tf.__version__\n'1.11.0'\n&gt;&gt;&gt; tfp.__version__\n'0.4.0'\n&gt;&gt;&gt; torch.__version__\n'0.4.1'\n&gt;&gt;&gt; \n&gt;&gt;&gt; p = tfd.Normal(loc=0., scale=1.)\n&gt;&gt;&gt; q = tfd.Laplace(loc=0., scale=1.)\n&gt;&gt;&gt; tfd.kl_divergence(p, q)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/root\/miniconda\/envs\/example\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/distributions\/kullback_leibler.py\", line 95, in kl_divergence\n    % (type(distribution_a).__name__, type(distribution_b).__name__))\nNotImplementedError: No KL(distribution_a || distribution_b) registered for distribution_a type Normal and distribution_b type Laplace\n&gt;&gt;&gt; \n&gt;&gt;&gt; a = torch.distributions.normal.Normal(loc=0., scale=1.)\n&gt;&gt;&gt; b = torch.distributions.laplace.Laplace(loc=0., scale=1.)\n&gt;&gt;&gt; torch.distributions.kl.kl_divergence(a,b)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/root\/miniconda\/envs\/example\/lib\/python3.6\/site-packages\/torch\/distributions\/kl.py\", line 161, in kl_divergence\n    raise NotImplementedError\nNotImplementedError\n<\/code><\/pre>\n\n<p>I assume as this is missing from both of these libraries that there is some good reason for this and that the user would be expected to implement it themselves with <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/distributions\/RegisterKL\" rel=\"nofollow noreferrer\"><code>tfp.distributions.RegisterKL<\/code><\/a> in TensorFlow Probability and <a href=\"https:\/\/pytorch.org\/docs\/stable\/distributions.html#module-torch.distributions.kl\" rel=\"nofollow noreferrer\"><code>torch.distributions.kl.register_kl<\/code><\/a> in PyTorch.<\/p>\n\n<p>Is this the correct assumption? If so, can someone explain why the KL Divergence wouldn't be implemented for given distribution classes? I think I am missing something very basic about this.<\/p>\n\n<p>If my assumption is wrong, can someone explain how to properly have TensorFlow and PyTorch implement these operations?<\/p>\n\n<p>For additional reference, using for this example an older version of TensorFlow that works with Edward,<\/p>\n\n<pre><code>pip install tensorflow==1.7\npip install edward\n<\/code><\/pre>\n\n<p>In this minimal example above, I'm trying to implement the equivalent of the following <code>edward<\/code> toy example code in <code>tfp<\/code> (or in <code>torch<\/code>).<\/p>\n\n<pre><code>import tensorflow as tf\nimport edward as ed\n\np = ed.models.Normal(loc=0., scale=1.)\ns = tf.Variable(1.)\nq = ed.models.Laplace(loc=0., scale=s)\ninference = ed.KLqp({p: q})\ninference.run(n_iter=5000)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53057552","CreationDate":"2018-10-30T04:42:05.147","ParentId":null,"OwnerUserId":"2281298","Title":null,"Body":"<p>IIRC, Edward's KLqp switches tries to use the Analytic form, and if not switches to using the sample KL.<\/p>\n\n<p>For TFP, and I think PyTorch, kl_divergence only works for distributions \nregistered, and unlike Edward only computes the analytic KL. As you mention, these aren't implemented in TFP, and I would say that's more of because the common cases (such as KL(MultivariateNormal || MultivariateNormal) have been implemented.<\/p>\n\n<p>To register the KL divergence, you would do something like: <a href=\"https:\/\/github.com\/tensorflow\/probability\/blob\/07878168731e0f6d3d0e7c878bdfd5780c16c8d4\/tensorflow_probability\/python\/distributions\/gamma.py#L275\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/blob\/07878168731e0f6d3d0e7c878bdfd5780c16c8d4\/tensorflow_probability\/python\/distributions\/gamma.py#L275<\/a>. (It would be great if you could file a PR at <a href=\"https:\/\/github.com\/tensorflow\/probability\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability<\/a>!). <\/p>\n\n<p>If it turns out that there isn't a suitable analytic form of this (off the top of my head, I don't know if there is one), then one can form the sample KL and do optimization with that. That can be done explicitly in TFP (by sampling and computing the sample KL. Also please file a PR if you would like this to be done more automatically as well. This is something some of us on TFP are interested in.<\/p>\n\n<p>It would be interesting to see for what cases analytic KL's can be automated. For instance, if q and p come from the same exponential family, then there is a nice form for the KL divergence in terms of sufficient statistics and the normalizer. But for KL's that are across exponential families (or even not exponential families), I'm not aware of results on classes of distributions where you can calculate the KL within the class semi-automatically.<\/p>\n"}]}
+{"QuestionId":53036736,"AnswerCount":0,"Tags":"<python><tensorflow><keras><reinforcement-learning><tensorflow-probability>","CreationDate":"2018-10-28T22:47:37.443","AcceptedAnswerId":null,"OwnerUserId":1599247.0,"Title":"TensorFlow: Using Keras with learnable tfp.bijectors\/tfp.distributions","Body":"<p>I'm trying to refactor some of our reinforcement learning methods using tf.keras. The keras models seem to work pretty nicely for feedforward networks that I'm using for value functions and preprocessors, but I'm struggling when trying to implement some of the probabilistic models (e.g. policies) with keras. In particular, combining the tfp.distributions and tfp.bijectors with the tf.keras.Model feels very unintuitive to me, and all the examples I've seen (e.g. [1, 2, 3]) are either too simplistic or rely on \"hacks\" that, to me, seem to undermine many of the benefits of using keras models (e.g. ability to hide the input handling, sessions, and numeric evaluation from the model itself).<\/p>\n\n<p>Suppose I want to implement a latent space policy, as described in [4], that uses a RealNVP flow to transform gaussian samples, conditioned on states, into actions. The policy needs to support at least two operations:\n1. Sampling of actions Y, such that\n    Y = g(X|S)\n    X ~ Normal(0, 1),\n   where g is the RealNVP transformation described in [4, 5], and S is the condition variables (e.g. the state observations in the RL case).\n2. Computing the log probabilities of the sampled Y's.<\/p>\n\n<p>A simple implementation could look something like this:<\/p>\n\n<pre><code>class LearnableConditionalRealNVP(object):\n    def __init__(self, input_shape, output_shape):\n        self._input_shape = input_shape\n        self._output_size = np.prod(output_shape)\n\n        conditions = tf.keras.layers.Input(shape=input_shape)\n\n        batch_size = tf.keras.layers.Lambda(\n            lambda x: tf.shape(x)[0])(conditions)\n\n        def samples_and_log_probs_fn(inputs):\n            conditions, batch_size = inputs\n\n            base_distribution = tfp.distributions.MultivariateNormalDiag(\n                loc=tf.zeros(output_shape),\n                scale_diag=tf.ones(output_shape))\n\n            real_nvp_bijector = tfp.bijectors.RealNVP(\n                num_masked=self._output_size \/\/ 2,\n                shift_and_log_scale_fn=conditioned_real_nvp_template(\n                    hidden_layers=(128, 128),\n                    activation=tf.nn.relu),\n                name='real_nvp')\n\n            distribution = (\n                tfp.distributions.ConditionalTransformedDistribution(\n                    distribution=base_distribution,\n                    bijector=real_nvp_bijector))\n\n            samples = distribution.sample(batch_size)\n            log_probs = distribution.log_prob(samples)\n\n            return [samples, tf.reshape(log_probs, (-1, 1))]\n\n        samples, log_probs = tf.keras.layers.Lambda(\n            samples_and_log_probs_fn)([conditions, batch_size])\n\n        self.samples_and_log_probs_model = tf.keras.Model(\n            conditions, [samples, log_probs])\n\n    def samples_and_log_probs(self, conditions):\n        return self.samples_and_log_probs_model(conditions)\n\n    def samples_and_log_probs_np(self, conditions):\n        return self.samples_and_log_probs_model.predict(conditions)\n<\/code><\/pre>\n\n<p>where <code>conditioned_real_nvp_template<\/code> creates a feedforward network that concatenates the latent samples and condition values along the last axis, and uses them as its input. The full example can be found here: <a href=\"https:\/\/gist.github.com\/hartikainen\/17ac2ec102032e986cb4d31e225f592a\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/hartikainen\/17ac2ec102032e986cb4d31e225f592a<\/a><\/p>\n\n<p>This way of handling the distribution would have two major benefits for me. First, I don't have to handle the reuse of parameters manually. I can call the <code>samples_and_log_probs<\/code> multiple times in my code, and it automatically reuses the parameters of the model. Second, I don't have to know anything about sessions if I want to get numeric outputs. Wrapping the intermediate layers into their own models allow me to call the predict methods that handle the sessions.<\/p>\n\n<p>The problems arise when extending the example even a bit further. Suppose I wanted to modify the <code>LearnableConditionalRealNVP<\/code> such that I could provide the latent samples x as an input to it, and instead of calling <code>distribution.sample()<\/code> in the <code>samples_and_log_probs_fn<\/code>, it would return <code>distribution.forward(x)<\/code>. Or maybe I want to get the samples and log_probs separately out of the model. This would require me to split the <code>samples_and_log_probs_fn<\/code> into two separate lambda functions, but doing that is not trivial if I want to share the parameters of the RealNVP bijector (because I can't pass the bijector as an input\/output to\/from the keras layer).<\/p>\n\n<p>I tried addressing these problems by subclassing the <code>LearnableConditionalRealNVP<\/code> from <code>tf.keras.Model<\/code>, but all my attempts resulted in messy implementations, mainly due to the changing inputs and outputs. Specifically, I was not able to create a <code>call<\/code>-method for the model such that the model would retain the ability to be used with <code>predict<\/code>, and I would have to make some tricks in the <code>__call__<\/code>-method. Neither of these feels horrible, yet they do increase the overhead of using keras models enough that it's easier for me to implement these types of things in plain tensorflow and handle the sessions, numpy-outputs, etc. manually.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>Are tensorflow bijectors\/distributions supposed to be compatible with keras (models)? If they are, does anyone know if there are any non-trivial examples available that I could take a look at? If they are not, are there plans to make them compatible?<\/li>\n<li>How should I use the keras models in cases, where what I consider to be one model has multiple different outputs? E.g., in my example above, the learnable RealNVP distribution intuitively feels like it should be a single model, yet it has multiple, possibly independent, inputs\/outputs, which makes it hard to fit in the keras model framework. I'm willing to accept that my intutition here is wrong, in which case it would be nice to hear what the best practice for building such models is.<\/li>\n<li>Is there a way to pass non-tensor data as input\/output to\/from a keras model as is done in [1], while still keeping the models somehow connected. If instead of <code>tfe.Variable<\/code>s, you use inputs in this example, it breaks because the graph is not connected.<\/li>\n<\/ol>\n\n<p>Edit: after posting this and doing some additional testing with the above implementation, I noticed that this model is not trainable after all, since the variables for the RealNVP bijector are created in a keras lambda layer. This suggests me that the functional way of building these models cannot be used for these types of models at all.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/tensorflow\/probability\/blob\/5f5510201865350b6cce2a0f18fbe0cdf4f15eee\/tensorflow_probability\/examples\/disentangled_vae.py#L186\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/blob\/5f5510201865350b6cce2a0f18fbe0cdf4f15eee\/tensorflow_probability\/examples\/disentangled_vae.py#L186<\/a><\/p>\n\n<p>[2] <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a><\/p>\n\n<p>[3] <a href=\"http:\/\/louistiao.me\/posts\/implementing-variational-autoencoders-in-keras-beyond-the-quickstart-tutorial\/\" rel=\"nofollow noreferrer\">http:\/\/louistiao.me\/posts\/implementing-variational-autoencoders-in-keras-beyond-the-quickstart-tutorial\/<\/a><\/p>\n\n<p>[4] <a href=\"https:\/\/arxiv.org\/pdf\/1804.02808.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1804.02808.pdf<\/a><\/p>\n\n<p>[5] <a href=\"https:\/\/arxiv.org\/abs\/1605.08803\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1605.08803<\/a><\/p>\n","answers":[]}
+{"QuestionId":53037339,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-29T00:34:42.017","AcceptedAnswerId":null,"OwnerUserId":4258928.0,"Title":"Tensorflow. No Loop way of subtract consecutive entry in tensor?","Body":"<p>So for each 1D array <code>x = [1,3,4,7,8,9,2,5]<\/code>. I'm looking for an output that is <code>[2,1,3,1,1,-7,3]<\/code> results from <code>x[i] - x[i-1]<\/code><\/p>\n\n<p>Similarly, for a 2D tensor. <\/p>\n\n<pre><code>[[1,3,14,4,2],\n [7,5,1,5,8],\n [9,2,1,5,7]]\n<\/code><\/pre>\n\n<p>The results would be <\/p>\n\n<pre><code>[[2,11,-10,-2],\n [-2,-4,4,3],\n [-7,-1,4,2]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53037450,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2018-10-29T00:54:02.907","AcceptedAnswerId":null,"OwnerUserId":6934018.0,"Title":"Keras lstm input and output shape","Body":"<p>I am new to python, deep learning and keras. I known many people asked similar questions before and i tried to read through them but my issues is still not solve. could someone please give me a hand<\/p>\n\n<p>I want to build a 6 inputs and 1 output model. below are my codes. your help or hint will be truly appreciated. <\/p>\n\n<p>input and output shape: <\/p>\n\n<pre><code>print(x_train.shape, y_train.shape)\n<\/code><\/pre>\n\n<p>output: <\/p>\n\n<pre><code>(503, 6) (503, 1)\n<\/code><\/pre>\n\n<p>model codes:<\/p>\n\n<pre><code>inputList={}\nlstmList={}\nfor i in range (x_train.shape[1]):\n  inputList[varList[i]]=Input(shape=(x_train.shape[0], 1), name=varList[i])  \n  lstmList[varList[i]]=LSTM(64, activation='relu', return_sequences=None, dropout=0.2)(inputList[varList[i]])\n\nz=concatenate([lstmList[i] for i in varList])\noutput=Dense(next_number_prediction, activation='softmax')(z)\nmodel = Model(inputs=[inputList[i] for i in varList], outputs=[output])\nmodel.compile(optimizer='rmsprop',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.summary()\n<\/code><\/pre>\n\n<p>the output is:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\nopen (InputLayer)               (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\nhigh (InputLayer)               (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\nlow (InputLayer)                (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\nclose (InputLayer)              (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\nchange (InputLayer)             (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\npct (InputLayer)                (None, 503, 1)       0                                            \n__________________________________________________________________________________________________\nlstm_7 (LSTM)                   (None, 64)           16896       open[0][0]                       \n__________________________________________________________________________________________________\nlstm_8 (LSTM)                   (None, 64)           16896       high[0][0]                       \n__________________________________________________________________________________________________\nlstm_9 (LSTM)                   (None, 64)           16896       low[0][0]                        \n__________________________________________________________________________________________________\nlstm_10 (LSTM)                  (None, 64)           16896       close[0][0]                      \n__________________________________________________________________________________________________\nlstm_11 (LSTM)                  (None, 64)           16896       change[0][0]                     \n__________________________________________________________________________________________________\nlstm_12 (LSTM)                  (None, 64)           16896       pct[0][0]                        \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 384)          0           lstm_7[0][0]                     \n                                                                 lstm_8[0][0]                     \n                                                                 lstm_9[0][0]                     \n                                                                 lstm_10[0][0]                    \n                                                                 lstm_11[0][0]                    \n                                                                 lstm_12[0][0]                    \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 1)            385         concatenate_1[0][0]              \n==================================================================================================\nTotal params: 101,761\nTrainable params: 101,761\nNon-trainable params: 0\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>Data treatment and model.fit:<\/p>\n\n<pre><code>Data={}    \nfor i in range (x_train.shape[1]):\n  Data[varList[i]]=np.expand_dims(x_train[:, i], axis=0)\n  Data[varList[i]]=np.reshape(Data[varList[i]], (1,x_train.shape[0],1))\n\nmodel.fit(\n    [Data[i] for i in varList], \n    [y_train], \n    epochs=10)\n<\/code><\/pre>\n\n<p>and the error is <\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-21-392e0052f15a&gt; in &lt;module&gt;()\n      1 model.fit(\n      2     [Data[i] for i in varList],\n----&gt; 3     [y_train], \n            epochs=10)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, max_queue_size, workers, use_multiprocessing, **kwargs)\n   1534         steps_name='steps_per_epoch',\n   1535         steps=steps_per_epoch,\n-&gt; 1536         validation_split=validation_split)\n   1537 \n   1538     # Prepare validation data.\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split)\n    990         x, y, sample_weight = next_element\n    991     x, y, sample_weights = self._standardize_weights(x, y, sample_weight,\n--&gt; 992                                                      class_weight, batch_size)\n    993     return x, y, sample_weights\n    994 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_weights(self, x, y, sample_weight, class_weight, batch_size)\n   1167       # Check that all arrays have the same length.\n   1168       if not self._distribution_strategy:\n-&gt; 1169         training_utils.check_array_lengths(x, y, sample_weights)\n   1170         if self._is_graph_network and not context.executing_eagerly():\n   1171           # Additional checks to avoid users mistakenly using improper loss fns.\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training_utils.py in check_array_lengths(inputs, targets, weights)\n    424                      'the same number of samples as target arrays. '\n    425                      'Found ' + str(list(set_x)[0]) + ' input samples '\n--&gt; 426                      'and ' + str(list(set_y)[0]) + ' target samples.')\n    427   if len(set_w) &gt; 1:\n    428     raise ValueError('All sample_weight arrays should have '\n\nValueError: Input arrays should have the same number of samples as target arrays. Found 1 input samples and 503 target samples.\n<\/code><\/pre>\n\n<p>The feed input and output dimension<\/p>\n\n<pre><code>print (Data[varList[i]].shape)\nprint (np.array([Data[i] for i in varList]).shape)\nprint (y_train.shape)\n<\/code><\/pre>\n\n<p>output: <\/p>\n\n<pre><code>(1, 503, 1)\n(6, 1, 503, 1)\n(503, 1)\n<\/code><\/pre>\n\n<p>tried new codes: <\/p>\n\n<pre><code>input = Input(shape=(x_train.shape))\nlstm = LSTM(64, activation='relu', return_sequences=True, dropout=0.2)(input)\noutput = Dense(1)(lstm)\nmodel2 = Model(inputs=input, outputs=output)\nmodel2.compile(optimizer='rmsprop',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel2.fit(x_train[np.newaxis,:,:], y_train[np.newaxis,:,:])\n<\/code><\/pre>\n\n<p>gives a untrained model: <\/p>\n\n<pre><code>Epoch 1\/10\n1\/1 [==============================] - 4s 4s\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 2\/10\n1\/1 [==============================] - 0s 385ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 3\/10\n1\/1 [==============================] - 0s 387ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 4\/10\n1\/1 [==============================] - 0s 386ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 5\/10\n1\/1 [==============================] - 0s 390ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 6\/10\n1\/1 [==============================] - 0s 390ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 7\/10\n1\/1 [==============================] - 0s 390ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 8\/10\n1\/1 [==============================] - 0s 389ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 9\/10\n1\/1 [==============================] - 0s 387ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\nEpoch 10\/10\n1\/1 [==============================] - 0s 391ms\/step - loss: 0.0000e+00 - acc: 0.0000e+00\n&lt;tensorflow.python.keras.callbacks.History at 0x7f4c97583e80&gt;\n<\/code><\/pre>\n\n<p>where the max and min of the data are: <\/p>\n\n<pre><code>print (max(y_train), x_train.max(axis=0))\nprint (min(y_train), x_train.min(axis=0))\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[0.79951533] [0.79930947 0.79750822 0.79934846 0.79951533 0.72939786 0.99697845]\n[0.19443386] [1.94643871e-01 1.96481512e-01 1.94604099e-01 1.94433856e-01\n 2.52289062e-04 3.70721060e-01]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53043742","CreationDate":"2018-10-29T10:41:10.593","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>Your network expects only a single label for the whole sequence. If I adapt the code like this it runs:<\/p>\n\n<pre><code>model.fit(\n      [Data[i] for i in varList],\n      [y_train[0:1]],\n      epochs=10)\n<\/code><\/pre>\n\n<p>Of course you need to decide whether this reflects your attention or whether you need to restructure your network so it accepts one label for every element in the sequence.<\/p>\n\n<p>By the way: This is how I would have constructed the network. So if you are new to this, maybe this is the architecture that you actually want:<\/p>\n\n<pre><code>input = Input(shape=(x_train.shape))\nlstm = LSTM(64, activation='relu', return_sequences=True, dropout=0.2)(input)\noutput = Dense(1)(lstm)\nmodel2 = Model(inputs=input, outputs=output)\nmodel2.compile(optimizer='rmsprop',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel2.fit(x_train[np.newaxis,:,:], y_train[np.newaxis,:,:])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53037451,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><deep-learning><multilabel-classification>","CreationDate":"2018-10-29T00:54:12.957","AcceptedAnswerId":53050708.0,"OwnerUserId":6456568.0,"Title":"keras custom metrics for multi-label classfication","Body":"<p>I'm using <code>sigmoid<\/code> and <code>binary_crossentropy<\/code> for multi-label classification. <\/p>\n\n<p>For example, the label of <code>y_true<\/code> is like <code>[1,0,1,0,0]<\/code>, and the label of <code>y_pred<\/code> is like <code>[0.8,0.3,0.9,0,0]<\/code>.<\/p>\n\n<p>How can I set a Keras custom metric function so that each element in <code>y_pred<\/code> larger than 0.5 is mapped to 1, each element in <code>y_pred<\/code> lower than 0.5 is mapped to 0, then compare the number of labels in <code>y_pred<\/code> that matches the <code>y_true<\/code>?<\/p>\n","answers":[{"AnswerId":"53050708","CreationDate":"2018-10-29T17:20:21.680","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since you are doing multi-label classification, it seems you want to compare the whole true and predicted labels with each other. For example, for one single sample with a true label of <code>[1, 0, 0]<\/code> and a predicted label of <code>[0, 0, 0]<\/code> you consider the prediction accuracy as zero (despite the fact that the labels for the second and third classes have been predicted correctly). If that's the case, you can compare the labels and then use <code>all()<\/code> method from backend to make sure the labels for all the classes match with each other:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef full_multi_label_metric(y_true, y_pred):\n    comp = K.equal(y_true, K.round(y_pred))\n    return K.cast(K.all(comp, axis=-1), K.floatx())\n<\/code><\/pre>\n"}]}
+{"QuestionId":53037510,"AnswerCount":2,"Tags":"<python><machine-learning><keras><training-data>","CreationDate":"2018-10-29T01:03:19.793","AcceptedAnswerId":53047744.0,"OwnerUserId":3837778.0,"Title":"Can flow_from_directory get train and validation data from the same directory in Keras?","Body":"<p>I got the following example from <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(\n        'data\/train',\n        target_size=(150, 150),\n        batch_size=32,\n        class_mode='binary')\n\nvalidation_generator = test_datagen.flow_from_directory(\n        'data\/validation',\n        target_size=(150, 150),\n        batch_size=32,\n        class_mode='binary')\n<\/code><\/pre>\n\n<p>There are two separate directories for train and validation. Just curious whether I can get train and validation data split from the same directory instead of two separate directories? Any example?<\/p>\n","answers":[{"AnswerId":"53047744","CreationDate":"2018-10-29T14:32:59.423","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can pass <code>validation_split<\/code> argument (a number between 0 and 1) to <code>ImageDataGenerator<\/code> class instance to split the data to train and validation sets:<\/p>\n\n<pre><code>generator = ImagaDataGenerator(..., validation_split=0.3)\n<\/code><\/pre>\n\n<p>And then pass <code>subset<\/code> argument to <code>flow_from_directory<\/code> to specify training and validation generators:<\/p>\n\n<pre><code>train_gen = generator.flow_from_directory(dir_path, ..., subset='training')\nval_gen = generator.flow_from_directory(dir_path, ..., subset='validation')\n<\/code><\/pre>\n"},{"AnswerId":"56755450","CreationDate":"2019-06-25T13:41:00.617","ParentId":null,"OwnerUserId":"6684726","Title":null,"Body":"<p>The above solution requires you to apply the same augmentations to the training and validation set, which might not be desired (You might not want to apply shear,rotation and zoom etc to the validation data). Separate training and validation augmentations from the same folder is not yet available. <\/p>\n\n<p>See <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/5862\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/5862<\/a> for full discussion (and some possible ways to handle this). People have usually resorted to scripts that create a new folder for validation, but that won't be an exact answer to this question.  <\/p>\n"}]}
+{"QuestionId":53038113,"AnswerCount":1,"Tags":"<python><tensorflow><keras><jupyter-notebook>","CreationDate":"2018-10-29T02:49:56.790","AcceptedAnswerId":null,"OwnerUserId":9880344.0,"Title":"Memory Error While Preprocessing Using Keras","Body":"<p>I am working on a project to classify dog breeds using a CNN in Keras. While preprocessing the data, I am getting this memory error: <a href=\"https:\/\/i.stack.imgur.com\/e9TkG.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/e9TkG.png<\/a>. What should I do?<\/p>\n","answers":[{"AnswerId":"53040248","CreationDate":"2018-10-29T06:56:10.237","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You should use <code>data generator<\/code>. <\/p>\n\n<p>You get this error because your machine doesn't have enough RAM to fit all the data on it. If you write the generator then you can get as much data as you need and then you can use <code>model.fit_generator()<\/code>. With this, you grab some amount of data, preprocess it, and forward it to model, grab another batch of data, preprocess it and forward to the model and so on.<\/p>\n\n<p>You can find out more <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":53038166,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-29T02:58:25.313","AcceptedAnswerId":null,"OwnerUserId":6193709.0,"Title":"Generate Balanced Batches from Unbalanced Data in tfrecords","Body":"<p>I have a dataset with 20k positive instances, and 300k negative instances.<\/p>\n\n<p>I would like to create a tfrecord file for each class, and then shuffle the data and generate batches with balanced classes.<\/p>\n","answers":[]}
+{"QuestionId":53038293,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-29T03:18:52.097","AcceptedAnswerId":53038344.0,"OwnerUserId":4987390.0,"Title":"How can I duplicate tensorflow layer?","Body":"<p>Suppose I have a layer (i.e. a collection of ops under the same name scope) in Tensorflow. How can I duplicate it together with input connections?<\/p>\n\n<p>More specifically, suppose I have the following graph:<\/p>\n\n<pre><code>A --&gt; B --&gt; C --&gt; D\n<\/code><\/pre>\n\n<p>now I want to duplicate <code>C<\/code> as <code>C1<\/code>, where <code>C<\/code> is a whole name scope:<\/p>\n\n<pre><code>A --&gt; B --&gt; C --&gt; D\n        \\-&gt; C\n<\/code><\/pre>\n\n<p>How can I do that in TensorFlow?<\/p>\n","answers":[{"AnswerId":"53038344","CreationDate":"2018-10-29T03:25:57.333","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<p>This can be done using <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.graph_editor\" rel=\"nofollow noreferrer\">tf.contrib.graph_editor<\/a>. Let's see how it can be done:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow.contrib.graph_editor as ge\n\n# Get the SubgraphView of given layer\nlayer_sgv = ge.make_view_from_scope(layer_name, tf.get_default_graph())\n\n# Retrieve the incoming tensors to the layer from ops outside.\n# We need these to preserve input hierarchy while duplicating.\nreplacement_ts = {}\nfor op in layer_sgv.inputs:\n    replacement_ts[op] = op\n\n# Duplicate the layer\nduplicate_sgv, info = ge.copy_with_input_replacements(\n    layer_sgv,\n    replacement_ts=replacement_ts,\n    src_scope=layer_name,\n    dst_scope=new_layer_name)\n<\/code><\/pre>\n\n<p>You can read more on <code>SubgraphView<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/graph_editor\/SubGraphView\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"57135777","CreationDate":"2019-07-21T18:09:31.727","ParentId":null,"OwnerUserId":"11767031","Title":null,"Body":"<p>The solution can be divided to 2 parts.<\/p>\n\n<h2>1. Replicate the graph of the layer<\/h2>\n\n<p>This is straightforward: just use the same code that you created that layer to do that. I suggest using Keras instead of raw TensorFlow &mdash; that will give you more flexibility and easiness in doing this step.<\/p>\n\n<h2>2. Copy the weights<\/h2>\n\n<p>The idea is you only need to copy <code>tf.Variables<\/code>, which are basically a group of following ops: <code>initializer<\/code>, <code>kernel<\/code>, and <code>assign<\/code>. <a href=\"https:\/\/stackoverflow.com\/questions\/44167134\/whats-the-difference-between-tensor-and-variable-in-tensorflow\">Here<\/a> is a good explanation. So the code will look as follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>vars = tf.trainable_variables()  # getting the variables\nvars_vals = sess.run(vars)       # getting their weights as numpy arrays\nvars_duplicates = ...            # here, get the weights of your layer,\n                                 # that should be in the same order\nfor var, val in zip(vars_duplicates, vars_vals):\n    var.load(val, sess)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53038635,"AnswerCount":0,"Tags":"<javascript><tensorflow><tensorflowjs-converter>","CreationDate":"2018-10-29T04:10:05.527","AcceptedAnswerId":null,"OwnerUserId":8403946.0,"Title":"Not able to load tensorflow JS model","Body":"<p>I have already converted keras model to tensorflowjs model. But when I am trying to load the same in browser it gives error which says that <\/p>\n\n<blockquote>\n  <p>Uncaught (in promise) Error: Provided weight data has no target variable: block1_conv1_1\/kernel. <\/p>\n<\/blockquote>\n\n<p>I have used transfer learning on vgg16 model in keras.<\/p>\n","answers":[]}
+{"QuestionId":53038804,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-29T04:35:52.607","AcceptedAnswerId":null,"OwnerUserId":7583237.0,"Title":"Selectively training a branch of multi-output model","Body":"<p>so I have a keras model with over 10 outputs, and I want to be able to train each output individually. All the other solutions I've seen say to make multiple models, but that makes it so then I have 10+ models I need to save and loading becomes a bit tricky then as well. I've have tried doing that but it seems very inefficient with the number of outputs I'm using so any other ideas would be appreciated. Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53038930,"AnswerCount":0,"Tags":"<image><tensorflow><multilabel-classification><multiclass-classification><resnet>","CreationDate":"2018-10-29T04:52:42.280","AcceptedAnswerId":null,"OwnerUserId":9915424.0,"Title":"hierarchical image classification in tensorflow","Body":"<p>I want to do two steps classification. for each input I want to go for classify it to class1, 2, or ... and then based on each class, classify my input to specific class (for example is in first step high score belongs to class2, final label should be class2_1, or ... class2_m).  <\/p>\n\n<p>input image --\n              class1\n                      class1_1\n                      class1_2\n              class2\n                      class2_1\n                         .\n                         .\n                .\n                .\n              class n \n                      classn_1\n                         .\n                         .<\/p>\n\n<p>I need tensorflow codes sample. Can any one help?\nThanks!<\/p>\n","answers":[]}
+{"QuestionId":53040241,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-10-29T06:55:44.913","AcceptedAnswerId":null,"OwnerUserId":10573326.0,"Title":"Why TensorFlow Lite need android version 5.1 or higher","Body":"<p>I will start developing with TensorFlowLite from now on. I have a question so please help me.<\/p>\n\n<p>I'm trying to purchase EPSON BT-300 for the development of TensorFlowLite now. EPSON BT-300 supports android 5.1. However, in the TensorflowLite tutorial \"Make sure the Android SDK version is greater than 26\" is written. Why do I need a version higher than 5.1?\nIs it possible to develop with EPSON BT-300 using TensorFlowLite?<\/p>\n","answers":[{"AnswerId":"53082424","CreationDate":"2018-10-31T11:36:10.847","ParentId":null,"OwnerUserId":"8049751","Title":null,"Body":"<p>It's possible to use tflite with android 5.1 (or other). I used <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/lite\/tools\/benchmark\" rel=\"nofollow noreferrer\">this<\/a> benchmark with android 5.1\nSome constraints may be only from using some newer features from SDK>22 in your android application<\/p>\n"}]}
+{"QuestionId":53040359,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><machine-learning><computer-vision>","CreationDate":"2018-10-29T07:05:20.543","AcceptedAnswerId":null,"OwnerUserId":10573076.0,"Title":"How to detect crop row using python OpenCV?","Body":"<p>I am new to machine learning and image recognition. I did quite a lot of research and found that we can detect the line on python using Pattern Hough Transform. With all my research and implementation, I couldn't get to the correct solution. <\/p>\n\n<p>Basically, I need to detect the crop rows, draw the line and get the coordinates of the lines as shown on the image. Can anyone please help me with the code?<\/p>\n\n<p>This is the raw image \n<a href=\"https:\/\/i.stack.imgur.com\/yja31.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yja31.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is how I want to process the image\n<a href=\"https:\/\/i.stack.imgur.com\/lqACa.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lqACa.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This did not detect the correct pattern:<\/p>\n\n<pre><code>#import dependencies\nimport cv2\nimport numpy as np\nimport matplotlib.pyplot as plt\n#load image using cv's imread(nameoffile)\nimg = cv2.imread('crop.png')\n\n#split the image into blue, green, and red channels\nb,g,r = cv2.split(img) \n#here we 'amplify' the color green to stand out, without red\/blue\ngscale = 2*g-r-b #we are going to refer to this as our grayscale img\n#Canny edge detection \ngscale = cv2.Canny(gscale,400,400,apertureSize = 3)\n#checking the results (good practice)\nplt.figure()\nplt.plot(), plt.imshow(gscale)\nplt.title('Canny Edge-Detection Results')\nplt.xticks([]), plt.yticks([])\nplt.show()\n\nsize = np.size(gscale) #returns the product of the array dimensions\nskel = np.zeros(gscale.shape,np.uint8) #array of zeros\nret,gscale = cv2.threshold(gscale,128,255,0) #thresholding the image\nelement = cv2.getStructuringElement(cv2.MORPH_CROSS,(3,3))\ndone = False\nwhile( not done):\n    eroded = cv2.erode(gscale,element)\n    temp = cv2.dilate(eroded,element)\n    temp = cv2.subtract(gscale,temp)\n    skel = cv2.bitwise_or(skel,temp)\n    gscale = eroded.copy()\n    zeros = size - cv2.countNonZero(gscale)\n    if zeros==size:\n        done = True\n\n\nlines = cv2.HoughLines(skel,1,np.pi\/180,130)\na,b,c = lines.shape\nfor i in range(a):\n    rho = lines[i][0][0]\n    theta = lines[i][0][1]    \n    a = np.cos(theta)\n    b = np.sin(theta)\n    x0 = a*rho\n    y0 = b*rho\n    x1 = int(x0 + 1000*(-b))\n    y1 = int(y0 + 1000*(a))\n    x2 = int(x0 - 1000*(-b))\n    y2 = int(y0 - 1000*(a))\n    cv2.line(img,(x1,y1),(x2,y2),(0,0,255),2, cv2.LINE_AA)\n#showing the results:\nplt.subplot(121)\n#OpenCV reads images as BGR, this corrects so it is displayed as RGB\nplt.plot(),plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) \nplt.title('Row Detection'), plt.xticks([]), plt.yticks([])\nplt.subplot(122)\nplt.plot(),plt.imshow(skel,cmap='gray')\nplt.title('Skeletal Image'), plt.xticks([]), plt.yticks([])\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53040412,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-10-29T07:09:55.073","AcceptedAnswerId":null,"OwnerUserId":8186293.0,"Title":"Alternative to global pooling using tf.reduce_mean?","Body":"<pre><code>output = tf.reduce_mean(net, [1, 2], name='global_pool', keep_dims=False) \n<\/code><\/pre>\n\n<p><code>net<\/code> is of shape (batch_size, height, width, #channels)<\/p>\n\n<p>Is there any other way to do global pooling and obtain the same result?<\/p>\n","answers":[{"AnswerId":"53060072","CreationDate":"2018-10-30T08:23:07.527","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>I have one that is computational very complex, but it works:<\/p>\n\n<ol>\n<li>Use <code>tf.split(..., axis=-1)<\/code><\/li>\n<li>Perform <code>tf.layers.average_pooling2d<\/code> for every split and chose the kernel size = tensor size<\/li>\n<\/ol>\n\n<p>The second option is:\n<code>tf.keras.layers.GlobalAveragePooling2D<\/code>, but I would suggest the use <code>tf.reduce_mean<\/code>under the hood, but dint looked in the source.<\/p>\n"}]}
+{"QuestionId":53041331,"AnswerCount":1,"Tags":"<tensorflow><bounding-box><object-detection-api>","CreationDate":"2018-10-29T08:19:33.830","AcceptedAnswerId":null,"OwnerUserId":9740368.0,"Title":"Tensorflow Object Detection Unusually large bounding boxes and wrong results","Body":"<p>I am building an object detector in TensorFlow to detect, motorbike riders with and without helmet, I have 1000 Images each for riders with helmet, withouthelmet and pedestrians(pu together -- 3000 IMAGES),  My last checkpoint was 35267 steps, I have tested using a traffic video, but I see unusally large bounding boxes with wrong results. Can someone please explain  the reason for such detections? Do I need to wait for atleast 50000 steps?? or Do I need to add datasets(Images in the angle to Traffic Cameras)?<\/p>\n\n<p>Model - SSD Mobilenet COCO - Custom Object Detection,\nTraining Platform - Google Colab<\/p>\n\n<p>Please find the Images attached<a href=\"https:\/\/i.stack.imgur.com\/a6zBn.png\" rel=\"nofollow noreferrer\">Video Snapshot 1<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2S6Dr.png\" rel=\"nofollow noreferrer\">Video Snapshot 2<\/a><\/p>\n\n<h2>Day 2 - 10\/30\/2018<\/h2>\n\n<p>I have tested with Images today, I have got different results, seems to be correct,2nd Day if I test with single object in a Image. Please find the results\n<a href=\"https:\/\/i.stack.imgur.com\/a6zBn.png\" rel=\"nofollow noreferrer\">Single Object IMage Test 1<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/2S6Dr.png\" rel=\"nofollow noreferrer\">Single Object Image Test 2<\/a><\/p>\n\n<p>Tested CHeckpoint - 52,000 Steps<\/p>\n\n<p>But, If I test with the Images with multiple objects in a road, the detection is wrong and bounding boxes are weirdly bigger, Is it because of the dataset, as I am training with One Motorbike rider(with or with out helmet) per image.<\/p>\n\n<p>Please find the wrong results<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5rqHj.png\" rel=\"nofollow noreferrer\">Multi Object Image Test<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/bVKMF.png\" rel=\"nofollow noreferrer\">Multi Object Image Test<\/a><\/p>\n\n<p>I had also tested with images like all Motorbikes in the scene, In this case, I did not get any results, Please find the Images<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xhE8f.png\" rel=\"nofollow noreferrer\">No Result Image<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/7l1Ck.png\" rel=\"nofollow noreferrer\">No Result Image<\/a><\/p>\n\n<p>The results are very confusing, Is there anything I am missing?,<\/p>\n","answers":[{"AnswerId":"53046051","CreationDate":"2018-10-29T12:59:51.160","ParentId":null,"OwnerUserId":"3125927","Title":null,"Body":"<p>There is no need to wait till 50000 epocs you should get decent result in 35k or even in 10k. I would suggest <\/p>\n\n<ol>\n<li>go through you data-set again and check all the bounding boxes (data cleaning)<\/li>\n<li>Check your model with inference code for changes like batch normalization etc<\/li>\n<li>Add some more data with different features, angles and color complexities <\/li>\n<\/ol>\n\n<p>I would check these points before going further. <\/p>\n"}]}
+{"QuestionId":53041396,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-29T08:23:40.513","AcceptedAnswerId":null,"OwnerUserId":10573609.0,"Title":"implementing image classification in rnn","Body":"<p>I have implemented an example of classifying cats and dogs using cnn. You can get the code from <a href=\"https:\/\/github.com\/venkateshtata\/cnn_medium.\/tree\/master\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/becominghuman.ai\/building-an-image-classifier-using-deep-learning-in-python-totally-from-a-beginners-perspective-be8dbaf22dd8\" rel=\"nofollow noreferrer\">here<\/a>. I want to do the same but with RNN. <\/p>\n\n<p>How to do that? I want to use my own dataset. <\/p>\n\n<p>Below you find the code I used with my own datast:<\/p>\n\n<pre><code># Importing the Keras libraries and packages\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense\n# Initialising the CNN\nclassifier = Sequential()\n# Step 1 - Convolution\nclassifier.add(Conv2D(32, (3, 3), input_shape = (64, 64, 3), activation = \n'relu'))\n# Step 2 - Pooling\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n# Adding a second convolutional layer\nclassifier.add(Conv2D(32, (3, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n# Step 3 - Flattening\nclassifier.add(Flatten())\n# Step 4 - Full connection\nclassifier.add(Dense(units = 128, activation = 'relu'))\nclassifier.add(Dense(units = 1, activation = 'sigmoid'))\n# Compiling the CNN\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics \n= ['accuracy'])\n# Part 2 - Fitting the CNN to the images\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\nshear_range = 0.2,\nzoom_range = 0.2,\nhorizontal_flip = True)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\ntraining_set = train_datagen.flow_from_directory('dataset\/training_set',\ntarget_size = (64, 64),\nbatch_size = 32,\nclass_mode = 'binary')\ntest_set = test_datagen.flow_from_directory('dataset\/test_set',\ntarget_size = (64, 64),\nbatch_size = 32,\nclass_mode = 'binary')\nclassifier.fit_generator(training_set,\nsteps_per_epoch = 8000,\nepochs = 25,\nvalidation_data = test_set,\nvalidation_steps = 2000)\n# Part 3 - Making new predictions\nimport numpy as np\nfrom keras.preprocessing import image\ntest_image = image.load_img('dataset\/single_prediction\/cat_or_dog_1.jpg', \ntarget_size = (64, 64))\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis = 0)\nresult = classifier.predict(test_image)\ntraining_set.class_indices\nif result[0][0] == 1:\nprediction = 'dog'\nelse:\nprediction = 'cat'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53042944","CreationDate":"2018-10-29T09:58:44.070","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>Keras provides an example of how to classify the <a href=\"https:\/\/en.wikipedia.org\/wiki\/MNIST_database\" rel=\"nofollow noreferrer\">MNIST<\/a> dataset using an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Long_short-term_memory\" rel=\"nofollow noreferrer\">LSTM<\/a> <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_irnn.py\" rel=\"nofollow noreferrer\">here<\/a>. You can replace their dataset with yours, adjust the number of classes and you should be done.<\/p>\n"}]}
+{"QuestionId":53042634,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T09:41:59.410","AcceptedAnswerId":null,"OwnerUserId":10170730.0,"Title":"Are there any tensorflow optimizers which do not require gradient information?","Body":"<p>I want to optimize a function for which no gradient information is available.<\/p>\n\n<p>I was looking for an optimizer which does that but I cannot find one. Do they not exist?<\/p>\n\n<p>Since there are several ways of estimating a gradient I expect that they do.<\/p>\n","answers":[]}
+{"QuestionId":53042946,"AnswerCount":1,"Tags":"<ios><tensorflow><coreml><coremltools>","CreationDate":"2018-10-29T09:58:55.043","AcceptedAnswerId":null,"OwnerUserId":9158564.0,"Title":"Converting a tensorflow model to a coreml model on windows gives issues","Body":"<p>I'm having problems with a conversion of a custom tf model.\nThe model looks like this:\n<img src=\"https:\/\/user-images.githubusercontent.com\/22317343\/47638866-1cb88c00-db60-11e8-9631-61419f9ea2d6.png\" alt=\"tensorflow model\"><\/p>\n\n<p>and after conversion(without errors) it looks like this:<\/p>\n\n<p><img src=\"https:\/\/user-images.githubusercontent.com\/22317343\/47638901-35c13d00-db60-11e8-8502-fe5c2903335f.png\" alt=\"coreml model\"><\/p>\n\n<p>Where is the add operator?<\/p>\n\n<p>Another issue is that the output of the conversion gave me this:<\/p>\n\n<pre><code>Core ML input(s):\n[name: \"x_placeholder__0\"\ntype {\n  multiArrayType {\n    shape: 41\n    dataType: DOUBLE\n  }\n}\n]\nCore ML output(s):\n[name: \"softmax_prediction__0\"\ntype {\n  multiArrayType {\n    shape: 2\n    dataType: DOUBLE\n  }\n}\n]\n<\/code><\/pre>\n\n<p>But my model has only float values? (conversion did actually change it from float32 to float64)<\/p>\n\n<p>Can someone answer my questions and maybe tell me what I'm doing wrong?<\/p>\n\n<p>Thx<\/p>\n","answers":[{"AnswerId":"53045701","CreationDate":"2018-10-29T12:38:52.347","ParentId":null,"OwnerUserId":"7501629","Title":null,"Body":"<p>A MatMul followed by an Add is the same thing as the innerProduct operator. <\/p>\n\n<p>Note that the innerProduct layer has a \"bias\". If you look at those bias values, you'll see that these are the same values as are used in your Add operator. So coremltools \/ tf-coreml has simply combined these two operations into a single layer.<\/p>\n\n<p>The default datatype for MLMultiArray objects is DOUBLE. You can change this to FLOAT, but that may not necessarily be any faster. Here's how you would do this in Python:<\/p>\n\n<pre><code>import coremltools  \nimport sys  \n\ndef update_multiarray_to_float32(feature):  \n    if feature.type.HasField('multiArrayType'):  \n        import coremltools.proto.FeatureTypes_pb2 as _ft  \n        feature.type.multiArrayType.dataType = _ft.ArrayFeatureType.FLOAT32  \n\nif __name__ == \"__main__\":  \n    if len(sys.argv) != 3:  \n        print \"USAGE: %s &lt;input_model_path&gt; &lt;output_model_path&gt;\" % sys.argv[0]  \n        sys.exit(1)  \n\n    input_model_path = sys.argv[1]  \n    output_model_path = sys.argv[2]  \n\n    spec = coremltools.utils.load_spec(input_model_path)  \n\n    for input_feature in spec.description.input:  \n        update_multiarray_to_float32(input_feature)  \n\n    for output_feature in spec.description.output:  \n        update_multiarray_to_float32(output_feature)  \n\n    coremltools.utils.save_spec(spec, output_model_path)  \n<\/code><\/pre>\n\n<p>Script courtesy of a friendly Apple employee (see <a href=\"https:\/\/forums.developer.apple.com\/thread\/84401\" rel=\"nofollow noreferrer\">https:\/\/forums.developer.apple.com\/thread\/84401<\/a>).<\/p>\n"}]}
+{"QuestionId":53043005,"AnswerCount":0,"Tags":"<tensorflow><opencl><armv8><mali><sycl>","CreationDate":"2018-10-29T10:01:40.543","AcceptedAnswerId":null,"OwnerUserId":4154294.0,"Title":"Crashed after running my tensorflow code on an armv8 device with GPU","Body":"<p>I wanna run the deeplab models on my armv8 device with GPU through using the sycl. But I got stuck there. <\/p>\n\n<p>I have already built from source then installed the Tensorflow wheel on my firefly RK3399 device and no errors happened.\nBut when I run the tensorflow model with the line:<\/p>\n\n<pre><code>    with sv.managed_session(FLAGS.master,config=config,start_standard_services=False) as sess:\n<\/code><\/pre>\n\n<p>a crash invoked as follows:<\/p>\n\n<pre><code>    termintate called after throwing an instance of 'cl::sycl::compile_program_error'\n<\/code><\/pre>\n\n<p>then I switched to the traditional method\n<code>with tf.Session() as sess:<\/code> and when I run <code>sess.run()<\/code>\nthe same error came.\nI have also checked the error_code on the\n<a href=\"https:\/\/xilinx.github.io\/triSYCL\/Doxygen\/triSYCL\/html\/group__error__handling.html#classcl_1_2sycl_1_2compile__program__error\" rel=\"nofollow noreferrer\">site<\/a>:<\/p>\n\n<p>But I have no idea to solve this, what's the real reason, how to handle this if I got \"Error while compiling the SYCL kernel to a SYCL device\".<\/p>\n\n<h2><strong>Configuration<\/strong><\/h2>\n\n<ul>\n<li>CPU aarch64<\/li>\n<li>GPU Mali-T860<\/li>\n<li>OS ubuntu 16.04<\/li>\n<li>DEVICE Firefly rk3399<\/li>\n<li>SOFTWARE  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/v1.6.0-rc0\" rel=\"nofollow noreferrer\">tensorflow-1.6.0<\/a> bazel-0.11.1 computeCpp-1.0.0 python2.7.12 gcc-5.4.0<\/li>\n<\/ul>\n\n<h2><strong>Here is my computecpp_info outputs:<\/strong><\/h2>\n\n<p>firefly@firefly:\/usr\/local\/computecpp\/bin$ .\/computecpp_info <\/p>\n\n<p>ComputeCpp Info (CE 1.0.0)<\/p>\n\n<p>SYCL 1.2.1 revision 3<\/p>\n\n<p>Toolchain information:<\/p>\n\n<p>GLIBC version: 2.23\nGLIBCXX: 20160609\nThis version of libstdc++ is supported.<\/p>\n\n<p>Device Info:<\/p>\n\n<p>Discovered 1 devices matching:\n  platform    : \n  device type : <\/p>\n\n<p>Device 0:<\/p>\n\n<p>Device is supported                     : NO - Device does not support SPIR <\/p>\n\n<p>CL_DEVICE_NAME                          : Mali-T860<\/p>\n\n<p>CL_DEVICE_VENDOR                        : ARM<\/p>\n\n<p>CL_DRIVER_VERSION                       : 1.2<\/p>\n\n<p>CL_DEVICE_TYPE                          : CL_DEVICE_TYPE_GPU <\/p>\n\n<p>If you encounter problems when using any of these OpenCL devices, please consult\nthis website for known issues:\n<a href=\"https:\/\/computecpp.codeplay.com\/releases\/v1.0.0\/platform-support-notes\" rel=\"nofollow noreferrer\">https:\/\/computecpp.codeplay.com\/releases\/v1.0.0\/platform-support-notes<\/a><\/p>\n\n<p>I got the highlight warning of <code>Device is supported : NO - Device does not support SPIR<\/code>, but the instructions told me that's ok! So I didn't mind that, Should it be an issue?<\/p>\n","answers":[]}
+{"QuestionId":53043328,"AnswerCount":0,"Tags":"<theano><pymc3>","CreationDate":"2018-10-29T10:19:04.490","AcceptedAnswerId":null,"OwnerUserId":3399944.0,"Title":"Resetting Theano graph","Body":"<p>I need to run a PyMC3 model in a loop to estimate\/make predictions every month. How do you reset the Theano graph? I'm familiar with Tensorflow and I know this can be done, but googling doesn't seem to lead to any solutions. Alternatively, how are you meant to run a PyMC3 model in a loop?<\/p>\n","answers":[]}
+{"QuestionId":53043683,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T10:37:56.640","AcceptedAnswerId":53047322.0,"OwnerUserId":4296244.0,"Title":"Use tf.nn.l2_loss on a collection of differently shaped vectors","Body":"<p>I want to calcualte the l2 loss over all my weights and biases in my neural network. Therefor I add all weights and biases to the 'tf.GraphKeys.REGULARIZATION_LOSSES' and want to calculate the l2 loss with the in tensorflow defined function:<\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([inputDim, outputDim], stddev=0.1), name='W')\nb = tf.Variable(tf.ones([outputDim])\/10, name='b')\n\ntf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, W)\ntf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, b)\n...\n<\/code><\/pre>\n\n<p>and later in the code:<\/p>\n\n<pre><code>...\nvars = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\nl2_loss = tf.nn.l2_loss(vars) * config.L2PENALTY\n<\/code><\/pre>\n\n<p>I get this error using the function on a network with 3 layers and couldn't find a solution to it:<\/p>\n\n<pre><code>ValueError: Tried to convert 't' to a tensor and failed. Error: Shapes must be equal rank, but are 2 and 1\n    From merging shape 4 with other shapes. for 'l2_loss\/L2Loss\/packed' (op: 'Pack') with input shapes: [784,512], [512], [512,256], [256], [256,10], [10].\n<\/code><\/pre>\n","answers":[{"AnswerId":"53047322","CreationDate":"2018-10-29T14:10:16.937","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Although tf.nn.l2_loss receives Tensor as its argument, you passed a list of Tensors to tf.nn.l2_loss. So the error message means l2_loss cannot convert the list to a Tensor.<\/p>\n\n<p>We should calculate L2 loss like<\/p>\n\n<pre><code>vars = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\nlist_l2_loss = []\nfor v in vars:\n    list_l2_loss.append(tf.nn.l2_loss(v))\ntotal_l2_loss = tf.add_n(list_l2_loss)\n<\/code><\/pre>\n\n<p><em>Adding Information<\/em><\/p>\n\n<p>You can add n-D Tensor to a collection by tf.add_to_collection. But tf.GraphKeys.REGULARIZATION_LOSSES is the reserved name for regularizers created by tf.get_variable. So we should you an original name like \"MyReguralizers\".<\/p>\n"}]}
+{"QuestionId":53043713,"AnswerCount":0,"Tags":"<python><gpu><pytorch>","CreationDate":"2018-10-29T10:39:48.510","AcceptedAnswerId":null,"OwnerUserId":4524700.0,"Title":"Pytorch is not using GPU even it detects the GPU","Body":"<p>I made my windows 10 jupyter notebook as a server and running some trains on it.<\/p>\n\n<p>I've installed CUDA 9.0 and cuDNN properly, and python detects the GPU. This is what I've got on the anaconda prompt.<\/p>\n\n<pre><code>&gt;&gt;&gt; torch.cuda.get_device_name(0)\n'GeForce GTX 1070'\n<\/code><\/pre>\n\n<p>And I also placed my model and tensors on cuda by .cuda()<\/p>\n\n<pre><code>model = LogPPredictor(1, 58, 64, 128, 1, 'gsc')\n\nif torch.cuda.is_available():\n    torch.set_default_tensor_type(torch.cuda.DoubleTensor)\n    model.cuda()\nelse:\n    torch.set_default_tensor_type(torch.FloatTensor)\n\nlist_train_loss = list()\nlist_val_loss = list()\nacc = 0\nmse = 0\n\noptimizer = args.optim(model.parameters(),\n                       lr=args.lr,\n                       weight_decay=args.l2_coef)\n\ndata_train = DataLoader(args.dict_partition['train'], \n                        batch_size=args.batch_size,\n                        pin_memory=True,\n                        shuffle=args.shuffle)\n\ndata_val = DataLoader(args.dict_partition['val'],\n                     batch_size=args.batch_size,\n                     pin_memory=True,\n                     shuffle=args.shuffle)\n\nfor epoch in tqdm_notebook(range(args.epoch), desc='Epoch'):\n    model.train()\n    epoch_train_loss = 0\n    for i, batch in enumerate(data_train):\n        list_feature = torch.tensor(batch[0]).cuda()\n        list_adj = torch.tensor(batch[1]).cuda()\n        list_logP = torch.tensor(batch[2]).cuda()\n        list_logP = list_logP.view(-1,1)\n\n        optimizer.zero_grad()\n        list_pred_logP = model(list_feature, list_adj)\n        list_pred_logP.require_grad = False\n        train_loss = args.criterion(list_pred_logP, list_logP)\n        epoch_train_loss += train_loss.item()\n        train_loss.backward()\n        optimizer.step()\n\n    list_train_loss.append(epoch_train_loss\/len(data_train))\n\n    model.eval()\n    epoch_val_loss = 0\n    with torch.no_grad():\n        for i, batch in enumerate(data_val):\n            list_feature = torch.tensor(batch[0]).cuda()\n            list_adj = torch.tensor(batch[1]).cuda()\n            list_logP = torch.tensor(batch[2]).cuda()\n            list_logP = list_logP.view(-1,1)\n\n\n            list_pred_logP = model(list_feature, list_adj)\n            val_loss = args.criterion(list_pred_logP, list_logP)\n            epoch_val_loss += val_loss.item()\n\n    list_val_loss.append(epoch_val_loss\/len(data_val))\n\ndata_test = DataLoader(args.dict_partition['test'],\n                   batch_size=args.batch_size,\n                   pin_memory=True,\n                   shuffle=args.shuffle)\n\nmodel.eval()\nwith torch.no_grad():\n    logP_total = list()\n    pred_logP_total = list()\n    for i, batch in enumerate(data_val):\n        list_feature = torch.tensor(batch[0]).cuda()\n        list_adj = torch.tensor(batch[1]).cuda()\n        list_logP = torch.tensor(batch[2]).cuda()\n        logP_total += list_logP.tolist()\n        list_logP = list_logP.view(-1,1)\n\n\n    list_pred_logP = model(list_feature, list_adj)\n\n    pred_logP_total += list_pred_logP.tolist()\n\nmse = mean_squared_error(logP_total, pred_logP_total)\n<\/code><\/pre>\n\n<p>But on the Process Manager of Windows, whenever I start training, only CPU usage goes up to 25% and GPU usage remains 0. How can I fix this???<\/p>\n","answers":[]}
+{"QuestionId":53043763,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><deployment><tensorflow-serving>","CreationDate":"2018-10-29T10:42:26.763","AcceptedAnswerId":null,"OwnerUserId":4948889.0,"Title":"deploying the Tensorflow model in Python","Body":"<p>Need help in implementing the Tensorflow model in real time. \nWhile I am training everything is working fine but when I move on for a realtime forecast or prediction, the output what I received flunked.\nI do not know why is this happening.\nI used the reference of teh code from here: <a href=\"https:\/\/www.kaggle.com\/raoulma\/ny-stock-price-prediction-rnn-lstm-gru\/notebook\" rel=\"noreferrer\">https:\/\/www.kaggle.com\/raoulma\/ny-stock-price-prediction-rnn-lstm-gru\/notebook<\/a>\nAnd tried to implement or deploy using the same code with few changes.  <\/p>\n\n<p>See the following code:  <\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport sklearn\nimport sklearn.preprocessing\nimport datetime\nimport os\nimport tensorflow as tf\n\ndf = pd.read_csv(\"Realtime_Values.csv\", index_col = 0)\ndf.info()\ndef load_data(stock,seq_len):\n\n    data_raw = stock.as_matrix() # convert to numpy array\n    data = []\n\n    for index in range(len(data_raw) - seq_len): \n        data.append(data_raw[index: index + seq_len])\n    #print(len(data))\n    data = np.array(data);\n\n    x_forecast = data[:,:-1,:]\n    return x_forecast\n\ndef normalize_data(df):\n    cols = list(df.columns.values)\n    min_max_scaler = sklearn.preprocessing.MinMaxScaler()\n    df = pd.DataFrame(min_max_scaler.fit_transform(df.values))\n    df.columns = cols\n    return df\nmodel_path =\"modelsOHLC\"\nseq_len = 9\n# parameters\nn_steps = seq_len-1 \nn_inputs = 4\nn_neurons = 100 \nn_outputs = 4\nn_layers = 4\nlearning_rate = 0.01\nbatch_size = 10\nn_epochs = 1000\ntf.reset_default_graph()\n\nX = tf.placeholder(tf.float32, [None, n_steps, n_inputs])\ny = tf.placeholder(tf.float32, [None, n_outputs])\nlayers = [tf.contrib.rnn.BasicRNNCell(num_units=n_neurons, activation=tf.nn.elu)\n          for layer in range(n_layers)]\nmulti_layer_cell = tf.contrib.rnn.MultiRNNCell(layers)\nrnn_outputs, states = tf.nn.dynamic_rnn(multi_layer_cell, X, dtype=tf.float32)\n\nstacked_rnn_outputs = tf.reshape(rnn_outputs, [-1, n_neurons]) \nstacked_outputs = tf.layers.dense(stacked_rnn_outputs, n_outputs)\noutputs = tf.reshape(stacked_outputs, [-1, n_steps, n_outputs])\noutputs = outputs[:,n_steps-1,:] # keep only last output of sequence\n\nloss = tf.reduce_mean(tf.square(outputs - y)) # loss function = mean squared error \noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) \ntraining_op = optimizer.minimize(loss)\nsaver = tf.train.Saver()\nsess  =tf.Session()\nsess.run(tf.global_variables_initializer())    \nif(tf.train.checkpoint_exists(tf.train.latest_checkpoint(model_path))):\n        saver.restore(sess, tf.train.latest_checkpoint(model_path))\ndf = normalize_data(df)\nx_forecast = load_data(df,seq_len)\ny_forecast_pred = sess.run(outputs, feed_dict={X: x_forecast})\nprint(y_forecast_pred)\n<\/code><\/pre>\n\n<p>Can anyone help me in getting the above code run in real time without any issues?<\/p>\n","answers":[{"AnswerId":"53174321","CreationDate":"2018-11-06T14:51:39.597","ParentId":null,"OwnerUserId":"58001","Title":null,"Body":"<p>There is a possibility that the code failed to find the saved weights when program trains the model; thus the predictions are being generated at an untrained state. Your code for training model is:<\/p>\n\n<pre><code>if (tf.train.checkpoint_exists(tf.train.latest_checkpoint(model_path))):\n    saver.restore(sess, tf.train.latest_checkpoint(model_path))\n<\/code><\/pre>\n\n<p>To fix this problem:<\/p>\n\n<ul>\n<li><p>Add a debugging code such as <code>print(\"checkpoint exists!\")<\/code> <\/p><\/li>\n<li><p>Place breakpoint through a debugger before or after <code>save.restore(...)<\/code> to find a checkpoint to restore from. <\/p><\/li>\n<li><p>Look at the <code>model_path<\/code> to ensure your checkpoints are saved correctly.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53043922,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><conv-neural-network><image-segmentation>","CreationDate":"2018-10-29T10:51:21.163","AcceptedAnswerId":null,"OwnerUserId":4481339.0,"Title":"No Gradients Provided","Body":"<p>My problematic network takes a <code>10x10x1<\/code> image as input and returns <code>4x2x3<\/code> image as output. Like so:\n<strong>Input to the network<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/naxSE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/naxSE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The input image has two patches of size <code>2x2x1<\/code>.<\/p>\n\n<p><strong>Output of the network:<\/strong>\nMy problematic network needs to extract output image of size <code>4x2x3<\/code>. Like so:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PzpKb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PzpKb.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>These above patches should be directly extracted from another image I'd call <em>Original image<\/em>.<\/p>\n\n<p><strong>Original Image<\/strong>\nThis is the original image, where from I need to extract the original patches. This original Image contains patches of size <code>4x2x3<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/byLDe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/byLDe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I extract the left top location of each patch in my last layer from <code>10x10x1<\/code> input image, which has the patches at approximately the same locations as those of the input image. But, these values are predicted, so may not be totally accurate. Once I have the top left location of the patches, I extract them from the <code>10x10x3<\/code> original image and output them. This output is fed to the loss function. It seems my gradients aren't flowing back and I am getting following error:<\/p>\n\n<pre><code>ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables [\"&lt;tf.Variable 'cv\/conv1_W:0' shape=(4, 4, 1, 16) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'cv\/conv1_b:0' shape=(16,) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'fc\/fc_1_W:0' shape=(400, 100) dtype=float32_ref&gt;\", \"&lt;tf.Variable 'fc\/fc_1_b:0' shape=(100,) dtype=float32_ref&gt;\"] and loss Tensor(\"Mean:0\", shape=(), dtype=float32).\n<\/code><\/pre>\n\n<p>However, if I extract the image from <code>10x10x1<\/code> input image, everything works fine. I need to extract the <em>locations<\/em> from <code>10x10x1<\/code> image, and then extract the actual patch contents from <code>10x10x3<\/code> original image. Locations may not be accurate always, so I need to train my network in order to improve the locations given in the input <code>10x10x1<\/code> image. So, I need to be able extract the patches, compute gradient, improve the weights, and try extracting again.<\/p>\n\n<p>I am using Tensorflow and Conv Nets to perform this functionality. I'd appreciate any help. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53044066,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-29T11:00:08.420","AcceptedAnswerId":null,"OwnerUserId":1196757.0,"Title":"TF Dataset API for Image augmentation","Body":"<p>I am using tf Dataset API to read images and its labels. I like to do multiple image augmentations on images and increase my training data size. What i have done now is like below.<\/p>\n\n<pre><code>def flip(self, img, lbl):\n  image = tf.image.flip_left_right(img)\n  return image, lbl\n\ndef transpose(self, img, lbl):\n  image = tf.image.transpose_image(img)\n  return image, lbl\n\n# just read and resize the image.\nprocess_fn = lambda img, lbl: self.read_convert_image(img, lbl, self.args)\nflip_fn = lambda img, lbl: self.flip(img,lbl)\ntranspose_fn = lambda img, lbl: self.transpose(img,lbl)\n\ntrain_set = self.train_set.repeat()\ntrain_set = train_set.shuffle(args.batch_size)\ntrain_set = train_set.map(process_fn)\n\nfliped_data = train_set.map(flip_fn)\ntranspose_data = train_set.map(transpose_fn)\n\ntrain_set = train_set.concatenate(fliped_data)\ntrain_set = train_set.concatenate(transpose_data)\n\ntrain_set = train_set.batch(args.batch_size)\niterator = train_set.make_one_shot_iterator()\n\nimages, labels = iterator.get_next()\n<\/code><\/pre>\n\n<p>Is there a better way to do multiple augmentations. The problem with above approach is if i add more augmentation function , that many map and concatenate is required.<\/p>\n\n<p>Thank You<\/p>\n","answers":[{"AnswerId":"53060488","CreationDate":"2018-10-30T08:51:33.590","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>If you want to do augmentations yourself, without relying on Keras's <code>ImageDataGenerator<\/code> you can create a function like <code>img_aug<\/code> and then use it in your model or in the Dataset API pipeline. The code below is just a pseudocode, but it shows the idea. You define all your transformations, then you have some generic threshold above which you apply a transformation and try to apply them up to X times (in the code below it is 4)<\/p>\n\n<pre><code>def img_aug(image):\n  image = distorted_image\n\n  def h_flip():\n    return tf.image.flip_left_right(distorted_image)                \n  def v_flip():\n    return tf.image.flip_up_down(distorted_image)\n\n  threshold = tf.constant(0.9, dtype=tf.float32)      \n\n  def body(i, distorted_image):\n    p_order = tf.random_uniform(shape=[2], minval=0., maxval=1., dtype=tf.float32)\n    distorted_image = tf.case({                                      \n                               tf.greater(p_order[0], threshold): h_flip,  \n                               tf.greater(p_order[1], threshold): v_flip, \n                              }\n                              ,default=identity, exclusive=False)\n    return (i+1, distorted_image)\n\n  def cond(i, *args):\n    return i &lt; 4 # max number of transformations\n\n  parallel_iterations = 1\n  tf.while_loop(cond, body, [0,distorted_image], \n                parallel_iterations=parallel_iterations)\n  return distorted_image\n<\/code><\/pre>\n"},{"AnswerId":"53060255","CreationDate":"2018-10-30T08:35:52.027","ParentId":null,"OwnerUserId":"4257991","Title":null,"Body":"<p>A simple alternative for Image Augmentation is using <a href=\"https:\/\/www.tensorflow.org\/guide\/keras\" rel=\"nofollow noreferrer\">Tensorflow implemented Keras<\/a> \nwhich contains easy to use <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">api<\/a><\/p>\n\n<p>It's looks like this <\/p>\n\n<pre><code>ImageDataGenerator(rescale=1.\/255, \n    rotation_range=40,\n    width_shift_range=0.2,\n    height_shift_range=0.2,\n    shear_range=0.2,\n    zoom_range = 0.2, \n    horizontal_flip = True)\n<\/code><\/pre>\n\n<p>And you are ready to use the augmented image as much as you want. <\/p>\n\n<p>Here is an working github code example <a href=\"https:\/\/github.com\/uzl\/deep_learning\/blob\/master\/dog_vs_cat\/Con_net_with_augmentation.ipynb\" rel=\"nofollow noreferrer\">Conv_net_with_augmentation<\/a><\/p>\n"}]}
+{"QuestionId":53045116,"AnswerCount":1,"Tags":"<numpy><tensorflow>","CreationDate":"2018-10-29T12:03:23.657","AcceptedAnswerId":53045644.0,"OwnerUserId":10574680.0,"Title":"How to read data from npy file as TensorFlow dataset","Body":"<p>I have npy contains two ndarrays, one is a one-hot and the other one is the label. How can I input the data? <\/p>\n\n<p>How to convert these data into x_train as a tensor? I only have experience on MNIST dataset and i don\u2019t have to handle the input data. <\/p>\n","answers":[{"AnswerId":"53045644","CreationDate":"2018-10-29T12:35:57.997","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>MNIST datsets are loaded in a similar way as far as <em>.npy<\/em> arrays are concerned.<\/p>\n\n<p>This line of code<\/p>\n\n<pre><code>(X_train, y_train), (X_test, y_test) = \n\n keras.datasets.mnist.load_data(path=\"C:\/Users\/476458\/.keras\/datasets\/mnist.npz\")\n<\/code><\/pre>\n\n<p>in turn uses the following code.<\/p>\n\n<pre><code>path = get_file(path,\n                origin='https:\/\/s3.amazonaws.com\/img-datasets\/mnist.npz',\n                file_hash='8a61469f7ea1b51cbae51d4f78837e45')\nf = np.load(path)\nx_train, y_train = f['x_train'], f['y_train']\nx_test, y_test = f['x_test'], f['y_test']\nf.close()\nreturn (x_train, y_train), (x_test, y_test)\n<\/code><\/pre>\n\n<p><em>np.load<\/em>  loads the <em>.npz<\/em> file which is a zipped archive which contains 4 <em>.npy<\/em> files ( x_train.npy, y_train.npy, x_test.npy, y_test ).<\/p>\n\n<p>But you should be able to load a <em>.npy<\/em> file directly using <em>np.load<\/em>. <\/p>\n\n<p>I hope this gives you some ideas.<\/p>\n"}]}
+{"QuestionId":53045129,"AnswerCount":2,"Tags":"<memory-leaks><gpu><pytorch>","CreationDate":"2018-10-29T12:04:14.427","AcceptedAnswerId":null,"OwnerUserId":258274.0,"Title":"How to use a PyTorch model without calculating gradients?","Body":"<p>I found a line of code in <a href=\"https:\/\/github.com\/liorshk\/facenet_pytorch\/blob\/master\/train_triplet.py\" rel=\"nofollow noreferrer\">this file<\/a>:<\/p>\n\n<pre><code>out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)    \n<\/code><\/pre>\n\n<p>As you can see the <code>forward()<\/code> function is invoked multiple times before <code>backward()<\/code>,In my test,the GPU consumption will increase accordingly,so GPU memory leaking happens. The questions here are:<\/p>\n\n<ol>\n<li>How to solve this problem?<\/li>\n<li>Is it possible use model as a pure feature extractor for these three consecutive invoking (with some modification,of course),then in fourth invoking,use it as <code>forward()<\/code>,if \"Yes\",how to implement it?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53079508","CreationDate":"2018-10-31T08:51:11.493","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>As an alternative to the already proposed answer you can also locally disable gradient computation by using <code>torch.no_grad()<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with torch.no_grad():\n    out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)  \n<\/code><\/pre>\n\n<hr>\n\n<p>This also prevents any gradient calculations within the <code>with<\/code> statement and thus reduces memory usage significantly:<\/p>\n\n<blockquote>\n  <p>Context-manager that disabled gradient calculation.<\/p>\n  \n  <p>Disabling gradient calculation is useful for inference, when you are\n  sure that you will not call <code>Tensor.backward()<\/code>. It will reduce memory\n  consumption for computations that would otherwise have\n  <em>requires_grad=True<\/em>. In this mode, the result of every computation will\n  have <em>requires_grad=False<\/em>, even when the inputs have\n  <em>requires_grad=True<\/em>.<\/p>\n<\/blockquote>\n\n<p><em>Source: <a href=\"https:\/\/pytorch.org\/docs\/stable\/autograd.html#torch.autograd.no_grad\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/autograd.html#torch.autograd.no_grad<\/a><\/em><\/p>\n"},{"AnswerId":"53079330","CreationDate":"2018-10-31T08:40:19.163","ParentId":null,"OwnerUserId":"6624101","Title":null,"Body":"<p>If you only want to take the model as a pure feature extractor (i.e., no update during the process), here are something you can do<\/p>\n\n<p>Assuming you have a model, <code>model<\/code>, you'd like to set it to the evaluation mode first.<\/p>\n\n<pre><code>model = FeatureExtractor().cuda()\nmode.eval()\n<\/code><\/pre>\n\n<p>Next, to decrease the comsuption of gpu memory, you need to stop gradient accumulation and computation.<\/p>\n\n<pre><code>for p in model.parameters():\n    p.require_grads = False\nout_a, out_p, out_n = model(data_a), model(data_p), model(data_n)\n<\/code><\/pre>\n\n<p>Then, I guess your problem will be solved. Note that, there is \"no\" gradient you can backprop through <code>out_a<\/code>, <code>out_p<\/code>, and <code>out_n<\/code>.<\/p>\n"}]}
+{"QuestionId":53045466,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-29T12:24:51.103","AcceptedAnswerId":null,"OwnerUserId":9764277.0,"Title":"Initialize weights as identity matrix in Tensorflow NN model","Body":"<p>I am relatively new to the work of Neural Network in one of the models on GitHub I noticed that after initializing the weights and biases as zero, an identity matrix was added to the biases variable. <\/p>\n\n<pre><code> weights = tf.get_variable('weights', [256, K*K],\n                              initializer=tf.constant_initializer(0.0),\n                              dtype=tf.float32)\n biases = tf.get_variable('biases', [K*K],\n                             initializer=tf.constant_initializer(0.0),\n                             dtype=tf.float32)\n biases += tf.constant(np.eye(K).flatten(), dtype=tf.float32)\n transform = tf.matmul(net, weights)\n transform = tf.nn.bias_add(transform, biases)\n<\/code><\/pre>\n\n<p>Is it particular to this model or is this the a common practice? If so, what is the purpose served by initializing biases as identity matrix? <\/p>\n\n<p>Link to the repo: <a href=\"https:\/\/github.com\/charlesq34\/pointnet\" rel=\"nofollow noreferrer\">Here<\/a>\nLink to the paper: <a href=\"https:\/\/arxiv.org\/pdf\/1612.00593.pdf\" rel=\"nofollow noreferrer\">Here<\/a><\/p>\n","answers":[{"AnswerId":"53045762","CreationDate":"2018-10-29T12:42:57.960","ParentId":null,"OwnerUserId":"8532404","Title":null,"Body":"<p>Initializing all weights to zero is actually a really bad idea. If you do this, the output of your neural network is going to be always zero, because no matterwhat the value of your input is, if you are multiplying it by zero, you get zero as output.<\/p>\n\n<p>For neural networks it is a common practice to initialize the weights and biases with random noise between -1 and 1.<\/p>\n\n<p>I personally use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/xavier_initializer\" rel=\"nofollow noreferrer\">Xavier initialization<\/a>, which is designed to keep the scale of the gradients approximately the same in all layers, therefore improving convergence.<\/p>\n"},{"AnswerId":"53047101","CreationDate":"2018-10-29T13:57:48.477","ParentId":null,"OwnerUserId":"10574412","Title":null,"Body":"<p>Initialize your weights to zeroes is not what you want. As fPecc said random noise is a good inizialization, otherwise you could try normal distribution.<\/p>\n"}]}
+{"QuestionId":53045887,"AnswerCount":2,"Tags":"<python><tensorflow><conda>","CreationDate":"2018-10-29T12:48:55.583","AcceptedAnswerId":null,"OwnerUserId":10574824.0,"Title":"Conda does not use package installed but using package from outside","Body":"<p>I got a problem when I using conda environment.<\/p>\n\n<p>I create a environment simply by:<\/p>\n\n<pre><code>conda create -n NAME python==3.5\n<\/code><\/pre>\n\n<p>So there is a python 3.5 virtual environment.\nThen I enter into the virtual environment and install tensorflow-gpu with:<\/p>\n\n<pre><code>conda install tensorflow-gpu\n<\/code><\/pre>\n\n<p>Now I can see the tensorflow-gpu package along with the cudatoolkit and cudnn.\nBut when I run a simple program it just failed to use gpu.I don't think it is a nvidia driver problem or package from different resources because it work in the old conda environment but when I create a new it failed.<\/p>\n\n<p>More interesting is it will using the package(tensorrt and uff) from outside of the enviroment.<\/p>\n\n<p>When I create a new environment with python 3.6.It works as expected using tensorflow-gpu and cannot import package which isnt installed in the environment.<\/p>\n\n<p>Does some one know how to fix the problem?<\/p>\n\n<blockquote>\n  <p>Ubuntu 16.04<\/p>\n  \n  <p>Tensorflow 1.8.0<\/p>\n  \n  <p>Python 3.5 and 3.6<\/p>\n  \n  <p>TensorRT 5.0.0.10<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53057054","CreationDate":"2018-10-30T03:39:45.163","ParentId":null,"OwnerUserId":"10574824","Title":null,"Body":"<p>Well I had fixed it.\nThis problem turn out to be the conda environment added a path from local and imported package from the local path instead of the environment.<\/p>\n\n<p>If you run into the same problem.Try open python in your conda environment by:<\/p>\n\n<pre><code>python\n<\/code><\/pre>\n\n<p>and check the system path:<\/p>\n\n<pre><code>import sys\nprint(sys.path)\n<\/code><\/pre>\n\n<p>if there is a lib from outside just remove it using:<\/p>\n\n<pre><code>rm -r ~\/PATH\n<\/code><\/pre>\n\n<p>This problem might origin from adding path unintentionally in some program.<\/p>\n\n<p>Hope that helps.<\/p>\n"},{"AnswerId":"55993949","CreationDate":"2019-05-05T16:12:34.053","ParentId":null,"OwnerUserId":"3418334","Title":null,"Body":"<p>This also happens if you use jupyter notebook. For example,\nif you have jupyter notebook installed and THEN create your environment, the jupyter notebook is OUTSIDE the environment and IGNORES all the packages you installed in your environment. I needed to install jupyter notebook into every environment I created to keep these errors from happening.<\/p>\n"}]}
+{"QuestionId":53046234,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-10-29T13:11:17.590","AcceptedAnswerId":null,"OwnerUserId":8375248.0,"Title":"Issues with downloading anaconda, and running tensorflow 1.10 on spyder","Body":"<p>I have had issues with anaconda and spyder and tensorflow on windows. I get that in spyder tensorflow is not imported and I have added it to my python path.. any help on the full process would be great help <\/p>\n","answers":[{"AnswerId":"53046435","CreationDate":"2018-10-29T13:22:58.503","ParentId":null,"OwnerUserId":"10349960","Title":null,"Body":"<p>Try to open your command line and write:<\/p>\n\n<pre><code>conda install tensorflow\n<\/code><\/pre>\n\n<p>Installation process should start.<\/p>\n\n<p>Same when other packages you need are not available.<\/p>\n"}]}
+{"QuestionId":53046583,"AnswerCount":1,"Tags":"<python><dataset><pytorch><torchtext>","CreationDate":"2018-10-29T13:31:23.983","AcceptedAnswerId":53100681.0,"OwnerUserId":10472356.0,"Title":"How to create a torchtext.data.TabularDataset directly from a list or dict","Body":"<p><code>torchtext.data.TabularDataset<\/code> can be created from a TSV\/JSON\/CSV file and then it can be used for building the vocabulary from Glove, FastText or any other embeddings. But my requirement is to create a <code>torchtext.data.TabularDataset<\/code> directly, either from a <code>list<\/code> or a <code>dict<\/code>. <\/p>\n\n<p>Current implementation of the code by reading TSV files<\/p>\n\n<pre><code>self.RAW = data.RawField()\nself.TEXT = data.Field(batch_first=True)\nself.LABEL = data.Field(sequential=False, unk_token=None)\n\n\nself.train, self.dev, self.test = data.TabularDataset.splits(\n    path='.data\/quora',\n    train='train.tsv',\n    validation='dev.tsv',\n    test='test.tsv',\n    format='tsv',\n    fields=[('label', self.LABEL),\n            ('q1', self.TEXT),\n            ('q2', self.TEXT),\n            ('id', self.RAW)])\n\n\nself.TEXT.build_vocab(self.train, self.dev, self.test, vectors=GloVe(name='840B', dim=300))\nself.LABEL.build_vocab(self.train)\n\n\nsort_key = lambda x: data.interleave_keys(len(x.q1), len(x.q2))\n\n\nself.train_iter, self.dev_iter, self.test_iter = \\\n    data.BucketIterator.splits((self.train, self.dev, self.test),\n                               batch_sizes=[args.batch_size] * 3,\n                               device=args.gpu,\n                               sort_key=sort_key)\n<\/code><\/pre>\n\n<p>This is the current working code for reading data from a file. So in order to create the dataset directly from a List\/Dict I tried inbuilt functions like <code>Examples.fromDict<\/code> or Examples.fromList but then while coming to the last for loop, it throws an error that <code>AttributeError: 'BucketIterator' object has no attribute 'q1'<\/code><\/p>\n","answers":[{"AnswerId":"53100681","CreationDate":"2018-11-01T11:47:14.913","ParentId":null,"OwnerUserId":"10472356","Title":null,"Body":"<p>It required me to write an own class inheriting the Dataset class and with few modifications in <code>torchtext.data.TabularDataset<\/code> class.<\/p>\n\n<pre><code>class TabularDataset_From_List(data.Dataset):\n\n    def __init__(self, input_list, format, fields, skip_header=False, **kwargs):\n        make_example = {\n            'json': Example.fromJSON, 'dict': Example.fromdict,\n            'tsv': Example.fromTSV, 'csv': Example.fromCSV}[format.lower()]\n\n        examples = [make_example(item, fields) for item in input_list]\n\n        if make_example in (Example.fromdict, Example.fromJSON):\n            fields, field_dict = [], fields\n            for field in field_dict.values():\n                if isinstance(field, list):\n                    fields.extend(field)\n                else:\n                    fields.append(field)\n\n        super(TabularDataset_From_List, self).__init__(examples, fields, **kwargs)\n\n    @classmethod\n    def splits(cls, path=None, root='.data', train=None, validation=None,\n               test=None, **kwargs):\n        if path is None:\n            path = cls.download(root)\n        train_data = None if train is None else cls(\n            train, **kwargs)\n        val_data = None if validation is None else cls(\n            validation, **kwargs)\n        test_data = None if test is None else cls(\n            test, **kwargs)\n        return tuple(d for d in (train_data, val_data, test_data)\n                     if d is not None)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53047003,"AnswerCount":1,"Tags":"<image-processing><machine-learning><computer-vision><caffe><multilabel-classification>","CreationDate":"2018-10-29T13:52:52.120","AcceptedAnswerId":53047591.0,"OwnerUserId":4749728.0,"Title":"Multi-class and multi-label image classification using Caffe","Body":"<p>I'm trying to create a single multi-class and multi-label net configuration in caffe.<\/p>\n\n<p>Let's say classification of dogs: Is the dog small or large? (class) What color is it? (class) is it have a collar? (label)<\/p>\n\n<p>Is this thing possible using caffe? \nWhat is the proper way to do so?<\/p>\n\n<p>Just trying to understand the practical way.. \nAfter creating 2 .text files (one for training and one for validation) containing all the tags of the images, for example:<\/p>\n\n<pre><code>\/train\/img\/1.png 0 4 18\n\/train\/img\/2.png 1 7 17 33\n\/train\/img\/3.png 0 4 17\n<\/code><\/pre>\n\n<p>Running the py script:<\/p>\n\n<pre><code>import h5py, os\nimport caffe\nimport numpy as np\n\nSIZE = 227 # fixed size to all images\nwith open( 'train.txt', 'r' ) as T :\n    lines = T.readlines()\n# If you do not have enough memory split data into\n# multiple batches and generate multiple separate h5 files\nX = np.zeros( (len(lines), 3, SIZE, SIZE), dtype='f4' ) \ny = np.zeros( (len(lines),1), dtype='f4' )\nfor i,l in enumerate(lines):\n    sp = l.split(' ')\n    img = caffe.io.load_image( sp[0] )\n    img = caffe.io.resize( img, (SIZE, SIZE, 3) ) # resize to fixed size\n    # you may apply other input transformations here...\n    # Note that the transformation should take img from size-by-size-by-3 and transpose it to 3-by-size-by-size\n    # for example\n    transposed_img = img.transpose((2,0,1))[::-1,:,:] # RGB-&gt;BGR\n    X[i] = transposed_img\n    y[i] = float(sp[1])\nwith h5py.File('train.h5','w') as H:\n    H.create_dataset( 'X', data=X ) # note the name X given to the dataset!\n    H.create_dataset( 'y', data=y ) # note the name y given to the dataset!\nwith open('train_h5_list.txt','w') as L:\n    L.write( 'train.h5' ) # list all h5 files you are going to use\n<\/code><\/pre>\n\n<p>And creating train.h5 and val.h5 (is X data set containing the images and Y contain the labels?).<\/p>\n\n<p>Replace my network input layers from:<\/p>\n\n<pre><code>layers { \n name: \"data\" \n type: DATA \n top:  \"data\" \n top:  \"label\" \n data_param { \n   source: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/train_db\" \n   backend: LMDB \n   batch_size: 64 \n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/mean.binaryproto\" \n    mirror: true \n  } \n  include: { phase: TRAIN } \n} \nlayers { \n name: \"data\" \n type: DATA \n top:  \"data\" \n top:  \"label\" \n data_param { \n   source: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/val_db\"  \n   backend: LMDB \n   batch_size: 64\n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/digits\/digits\/jobs\/20181010-191058-21ab\/mean.binaryproto\" \n    mirror: true \n  } \n  include: { phase: TEST } \n} \n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  top: \"X\" # same name as given in create_dataset!\n  top: \"y\"\n  hdf5_data_param {\n    source: \"train_h5_list.txt\" # do not give the h5 files directly, but the list.\n    batch_size: 32\n  }\n  include { phase:TRAIN }\n}\n\nlayer {\n  type: \"HDF5Data\"\n  top: \"X\" # same name as given in create_dataset!\n  top: \"y\"\n  hdf5_data_param {\n    source: \"val_h5_list.txt\" # do not give the h5 files directly, but the list.\n    batch_size: 32\n  }\n  include { phase:TEST }\n}\n<\/code><\/pre>\n\n<p>I guess HDF5 doesn't need a mean.binaryproto?<\/p>\n\n<p>Next, how the output layer should change in order to output multiple label probabilities?\nI guess I need cross- entropy layer instead of softmax?\nThis is the current output layers:<\/p>\n\n<pre><code>layers {\n  bottom: \"prob\"\n  bottom: \"label\"\n  top: \"loss\"\n  name: \"loss\"\n  type: SOFTMAX_LOSS\n  loss_weight: 1\n}\nlayers {\n  name: \"accuracy\"\n  type: ACCURACY\n  bottom: \"prob\"\n  bottom: \"label\"\n  top: \"accuracy\"\n  include: { phase: TEST }\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"53047591","CreationDate":"2018-10-29T14:24:26.807","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h3>Mean subtraction<\/h3>\n\n<p>While lmdb input data layer is able to handle various input transformations for you, <code>\"HDF5Data\"<\/code> layer does not support this functionality.<br>\nTherefore, you <strong>must<\/strong> take care of all input transformations (in particular mean subtraction) when you create your hdf5 files.<br>\nSee where your code says<\/p>\n\n<pre><code># you may apply other input transformations here...\n<\/code><\/pre>\n\n<h3>Multiple labels<\/h3>\n\n<p>Although your .txt lists several labels for each image, you only save the first one to hdf5 file. If you want to use these labels you have to feed them to the net.<br>\nAn issue that immediately rise from your example is that you do not have a fixed number of labels for each training image -- why? what does it mean?<br>\nAssuming you have three labels for each image (in .txt files):  <\/p>\n\n<blockquote>\n  <p>&lt; filename > &lt; dog size > &lt; dog color > &lt; has collar ><\/p>\n<\/blockquote>\n\n<p>Then you can have <code>y_size<\/code>, <code>y_color<\/code> and <code>y_collar<\/code> (instead of a single <code>y<\/code>) in your hdf5.<\/p>\n\n<pre><code>y_size[i] = float(spl[1])\ny_color[i] = float(spl[2])\ny_collar[i] = float(spl[3])\n<\/code><\/pre>\n\n<p>Your input data layer will have more <code>\"top\"<\/code>s accordingly:<\/p>\n\n<pre><code>layer {\n  type: \"HDF5Data\"\n  top: \"X\" # same name as given in create_dataset!\n  top: \"y_size\"\n  top: \"y_color\"\n  top: \"y_collar\"\n  hdf5_data_param {\n    source: \"train_h5_list.txt\" # do not give the h5 files directly, but the list.\n    batch_size: 32\n  }\n  include { phase:TRAIN }\n}\n<\/code><\/pre>\n\n<h3>Prediction<\/h3>\n\n<p>Currently your net only predict a single label (layer with <code>top: \"prob\"<\/code>). You need your net to predict all three labels, therefore you need to add layers that compute <code>top: \"prob_size\"<\/code>, <code>top: \"prob_color\"<\/code> and <code>top: \"prob_collar\"<\/code> (different layer for each <code>\"prob_*\"<\/code>).<br>\nOnce you have prediction for each label, you need a loss (again, a loss for each label).<\/p>\n"}]}
+{"QuestionId":53047021,"AnswerCount":1,"Tags":"<python><tensorflow><range><slice>","CreationDate":"2018-10-29T13:53:43.677","AcceptedAnswerId":53050477.0,"OwnerUserId":10170730.0,"Title":"How can I use Tensorflows scatter_nd with slices?","Body":"<p>I am trying to optimize only parts of a variable. I found <a href=\"https:\/\/stackoverflow.com\/a\/49180836\/10170730\">this<\/a> seemingly useful answer.<\/p>\n\n<p>However my variable is an image and I want to change only parts of it so I am trying to extend the code to more dimensions. This seems to work fine:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.contrib.opt as opt\n\nX = tf.Variable([[1.0, 2.0], [3.0, 4.0]])\n\n# the next two lines need to change because\n# manually specifying the values is not feasible\nindexes = tf.constant([[0, 0], [1, 0]])\nupdates = [X[0, 0], X[1, 0]]\n\npart_X = tf.scatter_nd(indexes, updates, [2, 2])\nX_2 = part_X + tf.stop_gradient(-part_X + X)\nY = tf.constant([[2.5, -3.5], [5.5, -7.5]])\nloss = tf.reduce_sum(tf.squared_difference(X_2, Y))\nopt = opt.ScipyOptimizerInterface(loss, [X])\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    opt.minimize(sess)\n    print(\"X: {}\".format(X.eval()))\n<\/code><\/pre>\n\n<p>However, since my image dimensions and the area I would like to select are much bigger manually specifying all the indices is not feasible. I would like to know how to use slices or range assignments to do so.<\/p>\n","answers":[{"AnswerId":"53050477","CreationDate":"2018-10-29T17:05:47.347","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do that like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input with size (50, 100)\nX = tf.Variable([[0] * 100] * 50)\n# Selected slice\nrow_start = 10\nrow_end = 30\ncol_start = 20\ncol_end = 50\n# Make indices from meshgrid\nindexes = tf.meshgrid(tf.range(row_start, row_end),\n                      tf.range(col_start, col_end), indexing='ij')\nindexes = tf.stack(indexes, axis=-1)\n# Take slice\nupdates = X[row_start:row_end, col_start:col_end]\n# Build tensor with \"filtered\" gradient\npart_X = tf.scatter_nd(indexes, updates, tf.shape(X))\nX_2 = part_X + tf.stop_gradient(-part_X + X)\n# Continue as before...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53047191,"AnswerCount":0,"Tags":"<tensorflow><keras><eager>","CreationDate":"2018-10-29T14:01:34.000","AcceptedAnswerId":null,"OwnerUserId":9118978.0,"Title":"Tensorflow eager execution - nested gradient tape","Body":"<p>I have been testing my WGAN-GP algorithm on TensorFlow using the traditional graph implementation. Recently, I happened to get to know TensorFlow Eager Execution and tried to convert my code to run on the eager execution mode.<\/p>\n\n<p>Let me first show you the previous code:<\/p>\n\n<pre><code>self.x_ = self.g_net(self.z)\nself.d  = self.d_net(self.x, reuse=False)\nself.d_ = self.d_net(self.x_)\n\nself.d_loss = tf.reduce_mean(self.d) - tf.reduce_mean(self.d_)\n\nepsilon = tf.random_uniform([], 0.0, 1.0)\n\nx_hat = epsilon * self.x + (1 - epsilon) * self.x_\nd_hat = self.d_net(x_hat)\n\nddx = tf.gradients(d_hat, x_hat)[0]\nddx = tf.sqrt(tf.reduce_sum(tf.square(ddx), axis=1))\nddx = tf.reduce_mean(tf.square(ddx - 1.0) * scale)\n\nself.d_loss = self.d_loss + ddx\n\nself.d_adam = tf.train.AdamOptimizer().minimize(self.d_loss, var_list=self.d_net.vars)\n<\/code><\/pre>\n\n<p>Then it is converted to:<\/p>\n\n<pre><code>self.x_ = self.g_net(self.z)\n\nepsilon = tf.random_uniform([], 0.0, 1.0)\n\nx_hat = epsilon * self.x + (1 - epsilon) * self.x_\n\nwith tf.GradientTape(persistent=True) as temp_tape:\n    temp_tape.watch(x_hat)\n    d_hat = self.d_net(x_hat)\n\nddx = temp_tape.gradient(d_hat, x_hat)[0]\nddx = tf.sqrt(tf.reduce_sum(tf.square(ddx), axis=1))\nddx = tf.reduce_mean(tf.square(ddx - 1.0) * 10)\n\nwith tf.GradientTape() as d_tape:\n    d  = self.d_net(x)\n    d_ = self.d_net(x_)\n\n    loss_d = tf.reduce_mean(d) - tf.reduce_mean(d_) + ddx\n\ngrad_d = d_tape.gradient(loss_d, self.d_net.variables)\n\nself.d_adam.apply_gradients(zip(grad_d, self.d_net.variables))\n<\/code><\/pre>\n\n<p>I tried several alternative ways to implement WGAN-GP loss, but in any way the d_loss is diverging! I hope there is someone who could enlighten me by pointing out my mistake(s).<\/p>\n\n<p>Furthermore, I wonder whether I could use Keras layers with my previous loss and optimizer implementation. Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":53047251,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-29T14:05:30.153","AcceptedAnswerId":53047505.0,"OwnerUserId":2067492.0,"Title":"How to create a tensorflow operation that will be reused","Body":"<p>I have some data that I want to process before feeding\/training a model. For this example I want to do a max pool 2d. I wrote a short function to do that with tensorflow.<\/p>\n\n<pre><code>import tensorflow\nimport tensorflow.nn as nn\n\ndef _tfMaxPool(arr, pool=(4,4), sess=None):\n    op = nn.max_pool(arr, (1, 1, pool[0], 1), (1, 1, pool[0], 1 ), padding=\"VALID\")\n    op = nn.max_pool(op, (1, 1, 1, pool[1]), (1, 1, 1, pool[1]), padding=\"VALID\")\n    if sess is None:\n        sess = tensorflow.Session();\n\n    return sess.run(op)\n<\/code><\/pre>\n\n<p>The problem is this can add nodes to my graph each time, which seems to clutter my session. One alternative way is to create a model.<\/p>\n\n<pre><code>import keras\n\nseq = keras.Sequential([ \n            keras.layers.InputLayer((1, 512, 512)), \n            keras.layers.MaxPool2D((4, 4), (4, 4), data_format=\"channels_first\")\n                     ])\ndef _tfMaxPool2(arr, pool=(4,4), sess=None):\n    swapped = arr.swapaxes(0,1)\n    return seq.predict(swapped).swapaxes(0,1)\n<\/code><\/pre>\n\n<p>The model is nearly exactly like what I want, but I think I am missing something fundamental. <\/p>\n","answers":[{"AnswerId":"53047505","CreationDate":"2018-10-29T14:20:27.107","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>Why don't you reuse your graph with various input? In the following code,tfMaxpool is defined only once.<\/p>\n\n<pre><code>def _tfMaxPool(arr, pool=(4,4)):\n    op = nn.max_pool(arr, (1, 1, pool[0], 1), (1, 1, pool[0], 1 ), padding=\"VALID\")\n    op = nn.max_pool(op, (1, 1, 1, pool[1]), (1, 1, 1, pool[1]), padding=\"VALID\")\n\n    return op\n\ninput = tf.placeholder()\noutput = _tfMaxPool(input)\n\nwith tf.Session() as sess:\n    sess.run(output, feed_dict={input:arr1})\n    sess.run(output, feed_dict={input:arr2})\n<\/code><\/pre>\n"}]}
+{"QuestionId":53047414,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-10-29T14:15:24.453","AcceptedAnswerId":null,"OwnerUserId":6840039.0,"Title":"Tensorflow: execute computations and avoid memory exceeded","Body":"<p>I try to compute tensorflow task for data size <code>(624003, 17424)<\/code>, that was getting from text using <code>CountVectorizer<\/code>.<\/p>\n\n<p>I always get an error <code>tensorflow.python.framework.errors_impl.InternalError: Dst tensor is not initialized<\/code><\/p>\n\n<p>But if I use sample of data like <code>(213556, 11605)<\/code> sample it works well.\nBut after increasing size of dataset it fail.<\/p>\n\n<p>I try to use this code for tensorflow<\/p>\n\n<pre><code>batch_size = 1024\n\nX = tf.placeholder(tf.float32, shape=(None, X_train.shape[1]), name=\"X\")\ny = tf.placeholder(tf.float32, shape=(None, y_train.shape[1]), name=\"y\")\n\n# set model weights\nweights = tf.Variable(tf.random_normal([X_train.shape[1], y_train.shape[1]], stddev=0.5), name=\"weights\")\n\n# construct model\ny_pred = tf.nn.sigmoid(tf.matmul(X, weights))\n\n# minimize error using cross entropy\n# cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(y_pred), reduction_indices=1))\ncost = tf.reduce_mean(-(y*tf.log(y_pred) + (1 - y)*tf.log(1 - y_pred)))\n\noptimizer_01 = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost)\noptimizer_001 = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)\n\n# saving model weights\nsaver = tf.train.Saver({\"weights\": weights})\n\n# variables initializing\ninit = tf.global_variables_initializer()\n\n# starting session\nwith tf.Session(config=tf.ConfigProto(device_count={'GPU': 0})) as sess:\n    sess.run(init)\n<\/code><\/pre>\n\n<p>And in main block I train on train data and get acc of test data.<\/p>\n\n<p>How can I learn on all train data and avoid memory exceeded?<\/p>\n\n<p>For batch I use next function<\/p>\n\n<pre><code>def optimize(session, optimizer, X_train, X_test, y_train, y_test, epoch=1):\n    for epoch in range(epoch):\n        for batch_i, (start, end) in enumerate(split(0, X_train.shape[0], batch_size)):\n            x_batch, y_true_batch, = X_train[start:end].toarray(), y_train[start:end]\n            feed_dict_train = {X: x_batch, y: y_true_batch}\n            session.run(optimizer, feed_dict=feed_dict_train)\n\n        feed_dict_test = {X: X_test.toarray(), y: y_test}\n        cost_step_test = session.run(cost, feed_dict={X: X_test.toarray(), y: y_test})\n<\/code><\/pre>\n","answers":[{"AnswerId":"53047737","CreationDate":"2018-10-29T14:32:36.937","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>(624003, 17424) Tensor is about 40GBytes. So you shouldn't allocate such a big Tensor.\nYou need to give up full-batch training, and switch to mini-batch training.<\/p>\n"}]}
+{"QuestionId":53047663,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-10-29T14:28:22.210","AcceptedAnswerId":null,"OwnerUserId":1378104.0,"Title":"Keras - training loss vs validation loss","Body":"<p>Just for the sake of the argument I am using the same data during training for training and validation, like this:<\/p>\n\n<pre><code>model.fit_generator(\n    generator=train_generator,\n    epochs=EPOCHS,\n    steps_per_epoch=train_generator.n \/\/ BATCH_SIZE,\n    validation_data=train_generator,\n    validation_steps=train_generator.n \/\/ BATCH_SIZE\n)\n<\/code><\/pre>\n\n<p>So I would expect that the loss and the accuracy of training and validation at the end of each epoch would be pretty much the same? Still it looks like this:<\/p>\n\n<pre><code>Epoch 1\/150\n26\/26 [==============================] - 55s 2s\/step - loss: 1.5520 - acc: 0.3171 - val_loss: 1.6646 - val_acc: 0.2796\nEpoch 2\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 1.2924 - acc: 0.4996 - val_loss: 1.5895 - val_acc: 0.3508\nEpoch 3\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 1.1624 - acc: 0.5873 - val_loss: 1.6197 - val_acc: 0.3262\nEpoch 4\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 1.0601 - acc: 0.6265 - val_loss: 1.9420 - val_acc: 0.3150\nEpoch 5\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 0.9790 - acc: 0.6640 - val_loss: 1.9667 - val_acc: 0.2823\nEpoch 6\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 0.9191 - acc: 0.6951 - val_loss: 1.8594 - val_acc: 0.3342\nEpoch 7\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 0.8811 - acc: 0.7087 - val_loss: 2.3223 - val_acc: 0.2869\nEpoch 8\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 0.8148 - acc: 0.7379 - val_loss: 1.9683 - val_acc: 0.3358\nEpoch 9\/150\n26\/26 [==============================] - 46s 2s\/step - loss: 0.8068 - acc: 0.7307 - val_loss: 2.1053 - val_acc: 0.3312\n<\/code><\/pre>\n\n<p>Why does especially the accuracy differ so much although its from the same data source? Is there something about the way how this is calculated that I am missing?<\/p>\n\n<hr>\n\n<p>The generator is created like this:<\/p>\n\n<pre><code>train_images = keras.preprocessing.image.ImageDataGenerator(\n    rescale=1.\/255\n)\n\ntrain_generator = train_images.flow_from_directory(\n    directory=\"data\/superheros\/images\/train\",\n    target_size=(299, 299),\n    batch_size=BATCH_SIZE,\n    shuffle=True\n)\n<\/code><\/pre>\n\n<p>Yes, it shuffles the images, but as it iterates over all images also for validation, shouldn't the accuracy at least be close?<\/p>\n\n<hr>\n\n<p>So the model looks like this:<\/p>\n\n<pre><code>inceptionV3 = keras.applications.inception_v3.InceptionV3(include_top=False)\n\nfeatures = inceptionV3.output\n\nnet = keras.layers.GlobalAveragePooling2D()(features)\npredictions = keras.layers.Dense(units=2, activation=\"softmax\")(net)\n\nfor layer in inceptionV3.layers:\n    layer.trainable = False\n\nmodel = keras.Model(inputs=inceptionV3.input, outputs=predictions)\noptimizer = keras.optimizers.RMSprop()\nmodel.compile(\n    optimizer=optimizer,\n    loss=\"categorical_crossentropy\",\n     metrics=['accuracy']\n)\n<\/code><\/pre>\n\n<p>So no dropout or anything, just the inceptionv3 with a softmax layer on top. I would expect that the accuracy differs a bit, but not in this magnitude.<\/p>\n","answers":[{"AnswerId":"53047881","CreationDate":"2018-10-29T14:40:27.427","ParentId":null,"OwnerUserId":"51685","Title":null,"Body":"<p>Are you sure your <code>train_generator<\/code> returns the same data when Keras retrieves training data and validation data, if it's a generator?<\/p>\n\n<p>The name being <code>generator<\/code>, I'd expect it not to :)<\/p>\n"}]}
+{"QuestionId":53047848,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-29T14:38:53.617","AcceptedAnswerId":53048831.0,"OwnerUserId":10568568.0,"Title":"keras LSTM poor prediction","Body":"<p>i am trying to predict the value based on some sequence (i have 5 values like 1,2,3,4,5 and want preditc the next one - 6). I am using LSTM keras for that.<\/p>\n\n<p>creating training data:<\/p>\n\n<pre><code>import numpy as np \nfrom keras.models import Sequential\nfrom keras.layers import LSTM,Dense\na = [float(i) for i in range(1,100)]\na = np.array(a)\n\ndata_train = a[:int(len(a)*0.9)]\ndata_test = a[int(len(a)*0.9):]\n\nx = 5\ny = 1\nz = 0\n\ntrain_x = []\ntrain_y = []\nfor i in data_train:\n    t = data_train[z:x]\n    r = data_train[x:x+y]\n    if len(r) == 0:\n        break\n    else:\n        train_x.append(t)\n        train_y.append(r)\n        z = z + 1\n        x = x+1\n\ntrain_x = np.array(train_x)\ntrain_y = np.array(train_y)\n\nx = 5\ny = 1\nz = 0\n\ntest_x = []\ntest_y = []\nfor i in data_test:\n    t = data_test[z:x]\n    r = data_test[x:x+y]\n    if len(r) == 0:\n        break\n    else:\n        test_x.append(t)\n        test_y.append(r)\n        z = z + 1\n        x = x+1\n\ntest_x = np.array(test_x)\ntest_y = np.array(test_y)\n\nprint(train_x.shape,train_y.shape)\nprint(test_x.shape,test_y.shape)\n<\/code><\/pre>\n\n<p>transform it into LSTM freandly shape:<\/p>\n\n<pre><code>train_x_1 = train_x.reshape(train_x.shape[0],len(train_x[0]),1)\ntrain_y_1 = train_y.reshape(train_y.shape[0],1)\ntest_x_1 = test_x.reshape(test_x.shape[0],len(test_x[0]),1)\ntest_y_1 = test_y.reshape(test_y.shape[0],1)\n\n\nprint(train_x_1.shape, train_y_1.shape)\nprint(test_x_1.shape, test_y_1.shape)\n<\/code><\/pre>\n\n<p>build and train model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(32,return_sequences = False,input_shape=(trein_x_1.shape[1],1)))\nmodel.add(Dense(1))\n\nmodel.compile(loss='mse',  optimizer='adam', metrics=['accuracy'])\nhistory = model.fit(train_x_1,\n                    train_y_1,\n                    epochs=20,\n                    shuffle=False, \n                    batch_size=1, \n                    verbose=2, \n                    validation_data=(test_x_1,test_y_1))\n<\/code><\/pre>\n\n<p>but I get a realy bad result, can somebody explaine me what I am doing wrong.<\/p>\n\n<pre><code>pred = model.predict(test_x_1)\nfor i,a in enumerate(pred):\n    print(pred[i],test_y_1[i])\n[89.71895] [95.]\n[89.87877] [96.]\n[90.03465] [97.]\n[90.18714] [98.]\n[90.337006] [99.]\n<\/code><\/pre>\n\n<p>Thenks.<\/p>\n","answers":[{"AnswerId":"53048831","CreationDate":"2018-10-29T15:30:15.413","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p>You expect the network to extrapolate from the data you used for training. Neural networks <a href=\"https:\/\/stackoverflow.com\/a\/18310746\/977098\">are not good at this<\/a>. You could try to normalize your data so that you are not extrapolating anymore by, for example, using relative values instead of absolute values. That would make this example of course very trivial.<\/p>\n"}]}
+{"QuestionId":53047885,"AnswerCount":1,"Tags":"<caffe><tensorrt><nvidia-jetson>","CreationDate":"2018-10-29T14:40:31.977","AcceptedAnswerId":53119938.0,"OwnerUserId":5578289.0,"Title":"Why are the layer weights null and TRT not finding cache when trying to run nvidia's tutorial code on Jetson TX2?","Body":"<p>I'm trying to run the tutorial code from nvidia's repo <a href=\"https:\/\/github.com\/dusty-nv\/jetson-inference\" rel=\"nofollow noreferrer\">here.<\/a>\nHere's what happens with the console imagenet program on my Jetson TX2:<\/p>\n\n<pre><code>    nvidia@tegra-ubuntu:~\/jetson-inference\/build\/aarch64\/bin$ .\/imagenet-console orange_0.pjg output_0.jpg\nimagenet-console\n  args (3):  0 [.\/imagenet-console]  1 [orange_0.pjg]  2 [output_0.jpg]  \n\n\nimageNet -- loading classification network model from:\n         -- prototxt     networks\/googlenet.prototxt\n         -- model        networks\/bvlc_googlenet.caffemodel\n         -- class_labels networks\/ilsvrc12_synset_words.txt\n         -- input_blob   'data'\n         -- output_blob  'prob'\n         -- batch_size   2\n\n[TRT]  TensorRT version 4.0.2\n[TRT]  attempting to open cache file networks\/bvlc_googlenet.caffemodel.2.tensorcache\n[TRT]  cache file not found, profiling network model\n[TRT]  platform has FP16 support.\n[TRT]  loading networks\/googlenet.prototxt networks\/bvlc_googlenet.caffemodel\nWeights for layer conv1\/7x7_s2 doesn't exist\n[TRT]  CaffeParser: ERROR: Attempting to access NULL weights\nWeights for layer conv1\/7x7_s2 doesn't exist\n[TRT]  CaffeParser: ERROR: Attempting to access NULL weights\n[TRT]  Parameter check failed at: ..\/builder\/Network.cpp::addConvolution::40, condition: kernelWeights.values != NULL\nerror parsing layer type Convolution index 1\n[TRT]  failed to parse caffe network\nfailed to load networks\/bvlc_googlenet.caffemodel\nfailed to load networks\/bvlc_googlenet.caffemodel\nimageNet -- failed to initialize.\nimagenet-console:   failed to initialize imageNet\n<\/code><\/pre>\n\n<p>I do not have Caffe installed on the Jetson board, as the tutorial specifically states that it is not needed. I'm not sure if the null weights error would be fixed if TRT would properly cache. Any ideas?<\/p>\n\n<ul>\n<li>Python 2.7<\/li>\n<li>Cuda 9.0<\/li>\n<li>TensorRT 4.0<\/li>\n<\/ul>\n","answers":[{"AnswerId":"53119938","CreationDate":"2018-11-02T13:57:51.520","ParentId":null,"OwnerUserId":"5578289","Title":null,"Body":"<p>The corporate firewall was preventing the proper download of the models. Downloading the models manually and putting them in the networks folder solved the problem.<\/p>\n"}]}
+{"QuestionId":53047983,"AnswerCount":1,"Tags":"<python><matplotlib><anaconda><spyder><pytorch>","CreationDate":"2018-10-29T14:46:48.227","AcceptedAnswerId":null,"OwnerUserId":6619902.0,"Title":"Importing pytorch in Spyder crashes kernel after installing matplotlib","Body":"<p>I created an environment in Anaconda3 and installed pytorch and spyder on a Linux machine. Here are the specifications:  <\/p>\n\n<pre><code>spyder                    3.3.1  \nipython                   7.0.1  \npython                    3.7.0  \npytorch                   0.4.1  \ntorchvision               0.2.1  \n<\/code><\/pre>\n\n<p>When I open spyder and import torch, it works. Afterwards I installed matplotlib 3.0.1. Restarting spyder and importing again pytorch results in a message on the ipython window in spyder:<\/p>\n\n<pre><code>An error ocurred while starting the kernel\nterminate called after throwing an instance of 'std::runtime_error'\nwhat(): expected ) but found 'ident' here:\naten::_addmv(Tensor self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1) \u2011&gt; Tensor\n~~~~~~ &lt;\u2011\u2011\u2011 HERE\n<\/code><\/pre>\n\n<p>On the bash terminal, I get the message:  <\/p>\n\n<pre><code>js: Not allowed to load local resource:   \nfile:\/\/\/home\/user\/anaconda3\/envs\/myenv\/lib\/python3.7\/site-packages\/spyder\/utils\/help\/static\/css\/default.css \n<\/code><\/pre>\n\n<p>I have been using all these packages in another environment for months (so they are in an older version), so it must be something with the new versions.<br>\nIf I run ipython or python on the terminal, importing works, so I am concluding it has something to do with spyder.<\/p>\n\n<p>The 'solution' is obvious: install older versions of the packages, but is there any other more sustainable solution?  <\/p>\n","answers":[{"AnswerId":"53088918","CreationDate":"2018-10-31T17:20:36.647","ParentId":null,"OwnerUserId":"10587170","Title":null,"Body":"<p>I have ipython 7.0.1 and matplotlib 2.0.2 and the same problem, it seems like ipython crashes after the following two commands: <code>%matplotlib auto<\/code> followed by <code>import torch<\/code>.<\/p>\n\n<p>This happens both in spyder as in jupyter notebook when the two commands are in seperate blocks.<\/p>\n\n<p>What worked for me was:\nFirst making sure that spyders backend graphics is set to inline: \nTools -> Preferences -> IPython console -> Graphics backed to Inline.\nThen <code>import torch<\/code> followed by switching from inline to external plotting with <code>%matplotlib auto<\/code>.<\/p>\n\n<p>Note that this does not happen any more with\nipython 7.2.0 and matplotlib 3.0.2<\/p>\n"}]}
+{"QuestionId":53048673,"AnswerCount":1,"Tags":"<python><amazon-web-services><tensorflow><amazon-ec2>","CreationDate":"2018-10-29T15:21:41.197","AcceptedAnswerId":null,"OwnerUserId":8734624.0,"Title":"Tensorflow Import Error on Amazon EC2 Instance","Body":"<p>I am trying to run a convolutional neural network in my p3.16xlarge Amazon EC2 instance. I installed tensorflow and all the other requisite libraries for my program using pip install, however, when I run this program, I receive the error message: <\/p>\n\n<p>\"ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory\"\n\"Failed to load the native TensorFlow runtime.\"<\/p>\n\n<p>These are my instance details:\nPython 2.7.14 (default, May  2 2018, 18:31:34) \nAmazon Linux AMI 2018.03.0.20180811 x86_64 HVM GP2<\/p>\n\n<p>Upon looking into the issue, it seems like I need CUDA 9.0 whereas most linux machines use CUDA 9.1. How exactly can I downgrade CUDA\/change\/configure my instance so that tensorflow runs? I do not know how to do this.<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"54268422","CreationDate":"2019-01-19T15:07:47.180","ParentId":null,"OwnerUserId":"4686192","Title":null,"Body":"<p>a lot of customers find that the Deep Learning AMI is the simplest way to run their workloads on EC2. It comes pre-installed with all the popular libraries.<\/p>\n\n<p><a href=\"https:\/\/aws.amazon.com\/machine-learning\/amis\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/machine-learning\/amis\/<\/a><\/p>\n\n<p>Alternatively, you could consider using Amazon SageMaker Store your data in S3, train and deploy on fully-managed infrastructure. Quite easier than managing their own EC2 instances :)<\/p>\n\n<p><a href=\"https:\/\/aws.amazon.com\/sagemaker\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/sagemaker\/<\/a><\/p>\n\n<p>I'd love to hear your feedback and answer any questions.<\/p>\n"}]}
+{"QuestionId":53048921,"AnswerCount":0,"Tags":"<python><machine-learning><keras><deep-learning><rnn>","CreationDate":"2018-10-29T15:34:58.947","AcceptedAnswerId":null,"OwnerUserId":10555987.0,"Title":"Can't apply simple 1 dimensional pooling in Keras","Body":"<p>I am trying to apply 1 dimensional pooling ti reduce the size of my input embeddings to a standard size.<\/p>\n\n<p>I am trying:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(n_vocab,600)) embeddings_initializer=Constant(embedding_matrix), trainable=True))\nmodel.add(AveragePooling1D(pool_size=2, strides=2))\nmodel.add(SimpleRNN(300,return_sequences=True))\nmodel.add(TimeDistributed(Dense(n_classes, activation='softmax')))\nmodel.compile('rmsprop', 'categorical_crossentropy')\n<\/code><\/pre>\n\n<p>But I get the error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [1,18,127] vs. [1,9,127]\n<\/code><\/pre>\n\n<p>How can I resolve the error, I just want the input to be pooled to half its original size.<\/p>\n","answers":[]}
+{"QuestionId":53049449,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2018-10-29T16:02:48.613","AcceptedAnswerId":null,"OwnerUserId":10263280.0,"Title":"How to merge classes post-training in TensorFlow Object Detection API","Body":"<p>TF OD API has a label map in which you tell the model what are your classes, give them ID's and names.\nLet's say I have a trained model on a given dataset, say MS COCO, and I want to merge\/unify some of the classes post-training (i.e. without any fine-tuning).\nFor example, I want to merge bicycle and motorcycle to the class \"two wheeler\", merge the entire food super-category to the single class \"food\", and merge the entire animals super-category to the single class \"animal\".\nCan I do it using the current API by modifying the label map, the label map util, or any other option?<\/p>\n\n<p>One can do it in a \"sophisticated way\" by adding the confidences of the relevant classes into a single confidence (assuming the confidences sum up to 1),\nbut I'm okay with just applying \"any\" to the class with the highest confidence, i.e. if the class with the highest confidence is any of the classes in the merged set, then assign it with the merged set class.<\/p>\n\n<p>I have tried modifying the label map by changing the names of the classes I want to merge to the merged set class (e.g. naming \"bird\", \"cat\", \"dog\", etc. simply as \"animal\"), but when evaluating the model it still treats to the classes in the set differently, by their ID.\nOn the other hand, I can't omit IDs I want to merge, since I still want the to detect them all, only treat them as the same.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53049983,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-29T16:33:26.680","AcceptedAnswerId":null,"OwnerUserId":1138158.0,"Title":"Classifying a new text by a Keras model saved on disk","Body":"<p>I have developed one text classification model using Keras. I also saved the model using the following code:<\/p>\n\n<pre><code>model.save('E:\/New Notes\/Python\/My BOW Model\/my_model.h5')\n<\/code><\/pre>\n\n<p>The file has been successfully saved. Now I want to load the file and classify a new text. <\/p>\n\n<p>Here is my code:   <\/p>\n\n<pre><code>from keras.models import load_model\nfrom string import punctuation\nfrom nltk.corpus import stopwords\nfrom keras.preprocessing.text import Tokenizer\n\nmodel = load_model('E:\/New Notes\/My BOW Model\/my_model.h5')\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n\n# load doc into memory\ndef load_doc(filename):\n    # open the file as read only\n    file = open(filename, 'r')\n    # read all text\n    text = file.read()\n    # close the file\n    file.close()\n    return text\n\n# load the vocabulary\nvocab_filename = 'E:\/New Notes\/My BOW Model\/vocab.txt'\nvocab = load_doc(vocab_filename)\nvocab = vocab.split()\nvocab = set(vocab)\n\n# turn a doc into clean tokens\ndef clean_doc(doc):\n    # split into tokens by white space\n    tokens = doc.split()\n    # remove punctuation from each token\n    table = str.maketrans('', '', punctuation)\n    tokens = [w.translate(table) for w in tokens]\n    # remove remaining tokens that are not alphabetic\n    tokens = [word for word in tokens if word.isalpha()]\n    # filter out stop words\n    stop_words = set(stopwords.words('english'))\n    tokens = [w for w in tokens if not w in stop_words]\n    # filter out short tokens\n    tokens = [word for word in tokens if len(word) &gt; 1]\n    return tokens\n\ndef predict_sentiment(review, vocab, tokenizer, model):\n    # clean\n    tokens = clean_doc(review)\n    # filter by vocab\n    tokens = [w for w in tokens if w in vocab]\n    # convert to line\n    line = ' '.join(tokens)\n    # encode\n    encoded = tokenizer.texts_to_matrix([line], mode='freq')\n    # prediction\n    classname = model.predict(encoded, verbose=0)\n    return round(classname[0,0])\n# classify a review as negative (0) or positive (1)\n# test positive text\ntext = 'The movie has a good appeal.'\ntokenizer = Tokenizer()\nprint(predict_sentiment(text, vocab, tokenizer, model))\n# test negative text\ntext = 'I dont like the movie at all.'\nprint(predict_sentiment(text, vocab, tokenizer, model))\n<\/code><\/pre>\n\n<p>After running the above code, I'm getting this error:<\/p>\n\n<pre><code>AttributeError: 'Tokenizer' object has no attribute 'word_index'. \n<\/code><\/pre>\n\n<p>What am I doing wrong? Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53050234,"AnswerCount":0,"Tags":"<python><keras><deep-learning><theano><ioerror>","CreationDate":"2018-10-29T16:50:04.007","AcceptedAnswerId":null,"OwnerUserId":10300316.0,"Title":"IOError: [Errno 2] No such file or directory when training Keras model","Body":"<p>I have been having some issues with training my sequential model in Keras. I am new to the subject, therefore there is a lot of following tutorials and code snippets involved...<\/p>\n\n<p>I was building a basic CNN to distinguish between linear infrastructure and geomorphological features on the basis of orthophotos. Since the images are very huge, I had to create a data generator which was following the Keras Docs. \nThe compilation of the model works fine. But every time I run the model.fit_generator() command I get the error, that one of my training images is missing (which it is not). I set five epochs to start, the error already occurs in the first one. <\/p>\n\n<p>I am thankful for any ideas what might have gone wrong.<\/p>\n\n<p>I am working on Ubuntu 16.04.5 LTS, with an iypthon notebook, theano backend.<\/p>\n\n<pre><code>train_path = '\/path\/to\/train\/folder'\ntree_top = os.listdir(train_path)\ntraining_filenames = os.listdir('%s\/%s\/' %(train_path, tree_top[0])) + os.listdir('%s\/%s\/' %(train_path, tree_top[1]))\n\nvalid_path = '\/path\/to\/valid\/folder'\ntree_top = os.listdir(valid_path)\nvalid_filenames = os.listdir('%s\/%s\/' %(valid_path, tree_top[0])) + os.listdir('%s\/%s\/' %(valid_path, tree_top[1]))\n<\/code><\/pre>\n\n<p>Data Generator<\/p>\n\n<pre><code>from skimage.io import imread\nfrom skimage.transform import resize\nimport numpy as np\n\nclass MY_Generator(Sequence):    # inherits from Sequence class \n\n    def __init__(self, image_filenames, labels, batch_size):\n        self.image_filenames, self.labels = image_filenames, labels\n        self.batch_size = batch_size\n\n    def __len__(self):    # computes number of batches by dividing sample size by the batch_size\n        return np.ceil(len(self.image_filenames) \/ float(self.batch_size))\n        num_training_samples = len(self.image_filenames)\n        return num_training_samples\n\n    def __getitem__(self, idx):\n        batch_x = self.image_filenames[idx * self.batch_size:(idx + 1) * self.batch_size]\n        batch_y = self.labels[idx * self.batch_size:(idx + 1) * self.batch_size]\n\n        return np.array([\n            resize(imread(file_name), (200, 200))\n               for file_name in batch_x]), np.array(batch_y)\n\n\nmy_training_batch_generator = MY_Generator(training_filenames, tree_top, batch_size)\nmy_validation_batch_generator = MY_Generator(valid_filenames, tree_top, batch_size)\n<\/code><\/pre>\n\n<p>The convnet<\/p>\n\n<pre><code>model = Sequential([\n        Conv2D(3, (3, 3), activation='relu', input_shape=(300,400,3)),\n        Flatten(),\n        Dense(2, activation='softmax'),\n    ])\nmodel.compile(Adam(lr=.0001), loss='categorical_crossentropy', metrics=['accuracy'])\n\nmodel.fit_generator(generator=my_training_batch_generator,\n                                          steps_per_epoch=(len(my_training_batch_generator.image_filenames) \/\/ batch_size),\n                                          epochs=5,\n                                          verbose=1,\n                                          validation_data=my_validation_batch_generator,\n                                          validation_steps=(len(my_validation_batch_generator.image_filenames) \/\/ batch_size)\n                                          )\n<\/code><\/pre>\n\n<p>The Output is as follows:<\/p>\n\n<blockquote>\n  <p>Epoch 1\/5<\/p>\n  \n  <p>--------------------------------------------------------------------------- IOError                                   Traceback (most recent call\n  last)  in ()\n        4                                           verbose=1,\n        5                                           validation_data=my_validation_batch_generator,\n  ----> 6                                           validation_steps=(len(my_validation_batch_generator.image_filenames)\n  \/\/ batch_size)\n        7                                           )<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/legacy\/interfaces.pyc in\n  wrapper(*args, **kwargs)\n       89                 warnings.warn('Update your <code>' + object_name + '<\/code> call to the ' +\n       90                               'Keras 2 API: ' + signature, stacklevel=2)\n  ---> 91             return func(*args, **kwargs)\n       92         wrapper._original_function = func\n       93         return wrapper<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training.pyc in\n  fit_generator(self, generator, steps_per_epoch, epochs, verbose,\n  callbacks, validation_data, validation_steps, class_weight,\n  max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch) \n  1416             use_multiprocessing=use_multiprocessing,    1417<br>\n  shuffle=shuffle,\n  -> 1418             initial_epoch=initial_epoch)    1419     1420     @interfaces.legacy_generator_methods_support<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training_generator.pyc\n  in fit_generator(model, generator, steps_per_epoch, epochs, verbose,\n  callbacks, validation_data, validation_steps, class_weight,\n  max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n      179             batch_index = 0\n      180             while steps_done &lt; steps_per_epoch:\n  --> 181                 generator_output = next(output_generator)\n      182 \n      183                 if not hasattr(generator_output, '<strong>len<\/strong>'):<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/utils\/data_utils.pyc in\n  get(self)\n      599         except Exception as e:\n      600             self.stop()\n  --> 601             six.reraise(*sys.exc_info())\n      602 \n      603 <\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/utils\/data_utils.pyc in\n  get(self)\n      593         try:\n      594             while self.is_running():\n  --> 595                 inputs = self.queue.get(block=True).get()\n      596                 self.queue.task_done()\n      597                 if inputs is not None:<\/p>\n  \n  <p>\/usr\/lib\/python2.7\/multiprocessing\/pool.pyc in get(self, timeout)\n      565             return self._value\n      566         else:\n  --> 567             raise self._value\n      568 \n      569     def _set(self, i, obj):<\/p>\n  \n  <p>IOError: [Errno 2] No such file or directory: 'DJI_0168.JPG'<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53050448,"AnswerCount":1,"Tags":"<python><machine-learning><keras><keras-layer><activation-function>","CreationDate":"2018-10-29T17:04:16.797","AcceptedAnswerId":53051673.0,"OwnerUserId":7867665.0,"Title":"Custom activation with parameter","Body":"<p>I'm trying to create an activation function in Keras that can take in a parameter <code>beta<\/code> like so:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.utils.generic_utils import get_custom_objects\nfrom keras.layers import Activation\n\nclass Swish(Activation):\n\n    def __init__(self, activation, beta, **kwargs):\n        super(Swish, self).__init__(activation, **kwargs)\n        self.__name__ = 'swish'\n        self.beta = beta\n\n\ndef swish(x):\n    return (K.sigmoid(beta*x) * x)\n\nget_custom_objects().update({'swish': Swish(swish, beta=1.)})\n<\/code><\/pre>\n\n<p>It runs fine without the <code>beta<\/code> parameter, but how can I include the parameter in the activation definition? I also want this value to be saved when I do <code>model.to_json()<\/code> like for ELU activation.<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong> I wrote the following code based on @today's answer:<\/p>\n\n<pre><code>from keras.layers import Layer\nfrom keras import backend as K\n\nclass Swish(Layer):\n    def __init__(self, beta, **kwargs):\n        super(Swish, self).__init__(**kwargs)\n        self.beta = K.cast_to_floatx(beta)\n        self.__name__ = 'swish'\n\n    def call(self, inputs):\n        return K.sigmoid(self.beta * inputs) * inputs\n\n    def get_config(self):\n        config = {'beta': float(self.beta)}\n        base_config = super(Swish, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\n    def compute_output_shape(self, input_shape):\n        return input_shape\n\nfrom keras.utils.generic_utils import get_custom_objects\nget_custom_objects().update({'swish': Swish(beta=1.)})\ngnn = keras.models.load_model(\"Model.h5\")\narch = gnn.to_json()\nwith open(directory + 'architecture.json', 'w') as arch_file:\n    arch_file.write(arch)\n<\/code><\/pre>\n\n<p>However, it does not currently save the <code>beta<\/code> value in the .json file. How can I make it save the value?<\/p>\n","answers":[{"AnswerId":"53051673","CreationDate":"2018-10-29T18:27:19.717","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since you want to save the parameters of activation function when serializing the model, I think it is better to define the activation function as a layer like the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/advanced_activations.py\" rel=\"nofollow noreferrer\">advanced activations which have been defined in Keras<\/a>. You can do it like this:<\/p>\n\n<pre><code>from keras.layers import Layer\nfrom keras import backend as K\n\nclass Swish(Layer):\n    def __init__(self, beta, **kwargs):\n        super(Swish, self).__init__(**kwargs)\n        self.beta = K.cast_to_floatx(beta)\n\n    def call(self, inputs):\n        return K.sigmoid(self.beta * inputs) * inputs\n\n    def get_config(self):\n        config = {'beta': float(self.beta)}\n        base_config = super(Swish, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\n    def compute_output_shape(self, input_shape):\n        return input_shape\n<\/code><\/pre>\n\n<p>Then you can use it the same way you use a Keras layer:<\/p>\n\n<pre><code># ...\nmodel.add(Swish(beta=0.3))\n<\/code><\/pre>\n\n<p>Since <code>get_config()<\/code> method has been implemented in its definition, the parameter <code>beta<\/code> would be saved when using methods like <code>to_json()<\/code> or <code>save()<\/code>.<\/p>\n"}]}
+{"QuestionId":53050723,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><keras><object-detection>","CreationDate":"2018-10-29T17:21:12.653","AcceptedAnswerId":null,"OwnerUserId":8596498.0,"Title":"how to train model with batches","Body":"<p>I trying yolo model in python.\nTo process the data and annotation I'm taking the data in batches.<\/p>\n\n<pre><code>batchsize = 50\n\n#boxList= []\n#boxArr = np.empty(shape = (0,26,5))\nfor i in range(0, len(box_list), batchsize):\n    boxList = box_list[i:i+batchsize]\n    imagesList = image_list[i:i+batchsize]\n\n    #to convert the annotation from VOC format\n    convertedBox = np.array([np.array(get_boxes_for_id(box_l)) for box_l in boxList])\n\n    #pre-process on image and annotaion\n    image_data, boxes = process_input_data(imagesList,max_boxes,convertedBox)\n    boxes = np.array(list(itertools.chain.from_iterable(boxes)))\n    detectors_mask, matching_true_boxes = get_detector_mask(boxes, anchors)\n<\/code><\/pre>\n\n<p>after this, I want to pass my data to the model to train.\nwhen I append the list it gives memory error because of array size.\nand when i append array gives dimensionality error because of shape.<\/p>\n\n<p>how can i train the data and what shoud i use model.fit() or model.train_on_batch()<\/p>\n","answers":[{"AnswerId":"53055378","CreationDate":"2018-10-29T23:35:58.177","ParentId":null,"OwnerUserId":"7973057","Title":null,"Body":"<p>If you are using Keras to Train your model with a bunch of Images you can use Train generator and validation generator, all you have to do is put your images in there respective class folders. look at a sample code . also take a look at this link maybe it may help you <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a> . i hope i have answered your question unless i did not understand it<a href=\"https:\/\/i.stack.imgur.com\/4eKur.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4eKur.png\" alt=\"sample keras code\"><\/a><\/p>\n"}]}
+{"QuestionId":53050923,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><tensorflow-lite><toco>","CreationDate":"2018-10-29T17:34:43.760","AcceptedAnswerId":null,"OwnerUserId":2353196.0,"Title":"Wildly different quantization performance on tensorflow-lite conversion of keras-trained DenseNet models","Body":"<p>I have two models that I have trained using Keras. The two models use the same architecture (the DenseNet169 implementation from <code>keras_applications.densenet<\/code> package), however they each have a different number of target classes (80 in one case, 200 in the other case).<\/p>\n\n<ul>\n<li><p>Converting both models to .pb format works just fine (identical performance in inference). I use the <code>keras_to_tensorflow<\/code> utility found at <a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amir-abdi\/keras_to_tensorflow<\/a><\/p><\/li>\n<li><p>Converting both models to .tflite format using TOCO works just fine (again, identical performance in inference).<\/p><\/li>\n<li><p>Converting the 80-class model to .tflite using quantization in TOCO works reasonably well (&lt;1% drop in top 3 accuracy).<\/p><\/li>\n<li><p>Converting the 200-class model to .tflite using quantization in TOCO goes off the rails (~30% drop in top 3 accuracy).<\/p><\/li>\n<\/ul>\n\n<p>I'm using an identical command-line to TOCO for both of the models:<\/p>\n\n<p><code>toco --graph_def_file frozen_graph.pb \\\n     --output_file quantized_graph.tflite \\\n     --inference_type FLOAT \\\n     --inference_input_type FLOAT \\\n     --output_format TFLITE \\\n     --input_arrays input_1 \\\n     --output_arrays output_node0 \\\n     --quantize True<\/code><\/p>\n\n<p>My tensorflow version is 1.11.0 (installed via pip on macOS Mojave, although I have also tried the same command\/environment on the Ubuntu machine I use for training with identical results).<\/p>\n\n<p>I'm at a complete loss as to why the accuracy of inference is so drastically affected for one model and not the other. This holds true for many different trainings of the same two architecture\/target class combinations. I feel like I must be missing something, but I'm baffled.<\/p>\n","answers":[{"AnswerId":"53056589","CreationDate":"2018-10-30T02:34:46.777","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<p><em>This was intended to be just a small sneaky comment since i'm not sure if this can help, but then it got so long that I decided to make it an answer...<\/em><\/p>\n\n<hr>\n\n<p><strong>My wild guess<\/strong> is that the accuracy drop may be caused by the variance of the output of your network. After quantization (btw, tensorflow uses <a href=\"https:\/\/heartbeat.fritz.ai\/8-bit-quantization-and-tensorflow-lite-speeding-up-mobile-inference-with-low-precision-a882dfcafbbd\" rel=\"nofollow noreferrer\">fixed-point quantization<\/a>), you are playing with only <code>256<\/code> points (8 bit) instead of the full dense range of <code>float32<\/code>.<\/p>\n\n<p>On most of the blogs around the web, it is stated that the main assumption of <em>Quantization<\/em> is that weights and activations tends to lie in a small range of values. However, there is an implicit assumption that is less talked about in blogs and literature: <em>the activations of the network on a single sample should be decently spread across the quantized range<\/em>.<\/p>\n\n<p>Consider the following scenario where the <strong>assumption holds place<\/strong> (a histogram of activations on single sample at specific layer, and the vertical lines are quantization points):\n<a href=\"https:\/\/i.stack.imgur.com\/WJDqQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WJDqQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now consider the scenario where the <strong>second assumption is not true<\/strong>, but the first assumption can <strong>still hold place<\/strong> (the blue is overall value distribution, gray is for given sample, vertical strips are quantization points):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/o173X.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/o173X.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In the first scenario, the distribution for the given sample is covered well (with a lot of quant points). In the second, only 2 quant points. <strong>The similar thing can happen to your network as well:<\/strong> maybe for 80 classes it still have enough quantization points to distinguish, but with 200 classes we might not have enough...<\/p>\n\n<blockquote>\n  <p>Hey, but it doesn't affect MobileNet with 1000 classes, and even MobileNetV2, which is residual?<\/p>\n<\/blockquote>\n\n<p>That's why I called it \"a wild guess\". Maybe MobileNet and MobileNetV2 does not have such a wide output variance as DenseNet. The former only have one input at each layer (which is already normalized by BN), while DenseNet have connections all over the places so it can have larger variance as well as sensitivity to small changes, and BN might not help as much.<\/p>\n\n<hr>\n\n<p>Now, try this checklist:<\/p>\n\n<ul>\n<li>Manually collect activation statistics of both 80 and 200 models on <em>TensorFlow<\/em>, not  only the outputs but inner layers as well. Is the values focused in one area or it spreads out widely?<\/li>\n<li>See if single-input activations of the <em>TensorFlow<\/em> model spreads out nicely, or we may have some issues with it concentrating in one place?<\/li>\n<li><strong>Most importantly<\/strong>: see what are the outputs of the <em>Quantized TF-Lite<\/em> model? If there are problems with the variance as described above, here is where it will show itself the most.<\/li>\n<\/ul>\n\n<hr>\n\n<p><strong>PS:<\/strong> please share your results as well, I think many will be interested in troubleshooting quantization issues :)<\/p>\n"}]}
+{"QuestionId":53050961,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><niftynet>","CreationDate":"2018-10-29T17:36:52.077","AcceptedAnswerId":null,"OwnerUserId":9555629.0,"Title":"Implement transfer learning on niftynet","Body":"<p>I want to implement transfer learning using the Dense V-Net architecture. As I was searching on how to do this, I found that this feature is currently being worked on (<a href=\"https:\/\/stackoverflow.com\/questions\/50768864\/how-do-i-implement-transfer-learning-in-niftynet\">How do I implement transfer learning in NiftyNet?<\/a>). <\/p>\n\n<p>Although from that answer it is quite clear that there is not a straight way to implement it, I was trying to:<\/p>\n\n<p>1) Create the Dense V-Net<\/p>\n\n<p>2) Restore weigths from the .ckpt file<\/p>\n\n<p>3) Implement transfer learning on my own<\/p>\n\n<p>To perform step 1, I thought I could use the niftynet.network.dense_vnet module. Therefore, I tried the following:<\/p>\n\n<pre><code>checkpoint = '\/path_to_ckpt\/model.ckpt-3000.index'\nx = tf.placeholder(dtype=tf.float32, shape=[None,1,144,144,144])\narchitecture_parameters = dict(\n    use_bdo=False,\n    use_prior=False,\n    use_dense_connections=True,\n    use_coords=False)\n\nhyperparameters = dict(\n    prior_size=12,\n    n_dense_channels=(4, 8, 16),\n    n_seg_channels=(12, 24, 24),\n    n_input_channels=(24, 24, 24),\n    dilation_rates=([1] * 5, [1] * 10, [1] * 10),\n    final_kernel=3,\n    augmentation_scale=0)\nmodel_instance = DenseVNet(num_classes=9,hyperparameters=hyperparameters,\n                             architecture_parameters=architecture_parameters)\n\nmodel_net = DenseVNet.layer_op(model_instance, x)\n<\/code><\/pre>\n\n<p>However, I get the following error: <\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;type 'list'&gt; to Tensor. Contents: [None, 1, 72, 72, 24]. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>So, the question is:<\/p>\n\n<p>Is there any way to implement this?<\/p>\n","answers":[{"AnswerId":"53441925","CreationDate":"2018-11-23T06:54:16.917","ParentId":null,"OwnerUserId":"6890908","Title":null,"Body":"<p>Transfer learning has been added been added to NiftyNet.<\/p>\n\n<p>You can select which variables you want to restore through the <code>vars_to_restore<\/code> config parameter and which variables to freeze through the <code>vars_to_freeze<\/code> config parameter.<\/p>\n\n<p>See <a href=\"https:\/\/niftynet.readthedocs.io\/en\/dev\/transfer_learning.html\" rel=\"nofollow noreferrer\">here<\/a> for more information.<\/p>\n"},{"AnswerId":"53410142","CreationDate":"2018-11-21T10:32:52.810","ParentId":null,"OwnerUserId":"5157376","Title":null,"Body":"<p>A simple transfer learning can be achieved with restoring weights from existing model in the way that you set the parameter <code>starting_iter<\/code> in <code>[TRAINING]<\/code> section of your config file to the number of pretrained model. In your example <code>starting_iter=3000<\/code>.<\/p>\n\n<p>This will restore the weights from your model and new iterations will start with this initialisation. <\/p>\n\n<p>Here the architecture of your model has to be exactly the same, otherwise you will get an error.<\/p>\n\n<p>For more sophisticated transfer learning or maybe also fine tunning where you can restore only a part of weights, there is a great implementation <a href=\"https:\/\/github.com\/aleks-djuric\/NiftyNet\/tree\/add-transfer-learning\" rel=\"nofollow noreferrer\">here<\/a>. It will be probably merged with official niftynet repository very soon, but you can already use it.<\/p>\n"}]}
+{"QuestionId":53051233,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2018-10-29T17:55:01.197","AcceptedAnswerId":53060979.0,"OwnerUserId":4229637.0,"Title":"tensorflow object-detection api eval on custom dataset","Body":"<p>I have a custom dataset that I trained using tensorflow object-detection api [Faster RCNN]. <\/p>\n\n<p>I saved the model checkpoints. Now, I want to run this model on my eval set. But, I cannot find the script that I use to run on eval set. \nI have tfrecord files for eval set.  <\/p>\n\n<p>How do I run the eval script from the object-detection api for my own saved model checkpoint? I couldn't find the script. <\/p>\n","answers":[{"AnswerId":"53060979","CreationDate":"2018-10-30T09:21:28.050","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>You are looking fot the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/legacy\/eval.py\" rel=\"nofollow noreferrer\">eval.py<\/a>.<\/p>\n\n<p>Run this script with the command:<\/p>\n\n<pre><code>python eval.py \\\n--logtostderr \\\n--pipeline_config_path=PATH_TO_YOUR_CONFIG_FILE \\\n--checkpoint_dir=PATH_TO_CHECKPOINT \\\n--eval_dir=OUTPUT_PATH\n<\/code><\/pre>\n"}]}
+{"QuestionId":53051274,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><google-colaboratory>","CreationDate":"2018-10-29T17:57:27.810","AcceptedAnswerId":null,"OwnerUserId":7429494.0,"Title":"I trained a keras model on google colab. Now not able to load it locally on my system.","Body":"<pre><code>with open('2model.json','r') as f:\njson = f.read()\nmodel = model_from_json(json)\nmodel.load_weights(\"color_tensorflow_real_mode.h5\")\n<\/code><\/pre>\n\n<p>I trained a keras model on google colab. Now not able to load it locally on my system. Getting this error: ValueError: Unknown initializer: GlorotUniform<\/p>\n\n<p>How to solve this??\nEvery time I make a model on colab and try loading it locally I am unable to do so.\nGetting this error message:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-17-c3ed162a8277&gt; in &lt;module&gt;()\n----&gt; 1 model = model_from_json(json)\n      2 model.load_weights(\"color_tensorflow_real_mode.h5\")\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\saving.py in model_from_json(json_string, custom_objects)\n    349   config = json.loads(json_string)\n    350   from tensorflow.python.keras.layers import deserialize  # pylint: disable=g-import-not-at-top\n--&gt; 351   return deserialize(config, custom_objects=custom_objects)\n    352 \n    353 \n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\serialization.py in deserialize(config, custom_objects)\n     62       module_objects=globs,\n     63       custom_objects=custom_objects,\n---&gt; 64       printable_module_name='layer')\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    171             custom_objects=dict(\n    172                 list(_GLOBAL_CUSTOM_OBJECTS.items()) +\n--&gt; 173                 list(custom_objects.items())))\n    174       with CustomObjectScope(custom_objects):\n    175         return cls.from_config(config['config'])\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in from_config(cls, config, custom_objects)\n   1290     # First, we create all layers and enqueue nodes to be processed\n   1291     for layer_data in config['layers']:\n-&gt; 1292       process_layer(layer_data)\n   1293     # Then we process nodes in order of layer depth.\n   1294     # Nodes that cannot yet be processed (if the inbound node\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in process_layer(layer_data)\n   1276       from tensorflow.python.keras.layers import deserialize as deserialize_layer  # pylint: disable=g-import-not-at-top\n   1277 \n-&gt; 1278       layer = deserialize_layer(layer_data, custom_objects=custom_objects)\n   1279       created_layers[layer_name] = layer\n   1280 \n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\serialization.py in deserialize(config, custom_objects)\n     62       module_objects=globs,\n     63       custom_objects=custom_objects,\n---&gt; 64       printable_module_name='layer')\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    173                 list(custom_objects.items())))\n    174       with CustomObjectScope(custom_objects):\n--&gt; 175         return cls.from_config(config['config'])\n    176     else:\n    177       # Then `cls` may be a function returning a class.\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py in from_config(cls, config)\n   1615         A layer instance.\n   1616     \"\"\"\n-&gt; 1617     return cls(**config)\n   1618 \n   1619 \n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\convolutional.py in __init__(self, filters, kernel_size, strides, padding, data_format, dilation_rate, activation, use_bias, kernel_initializer, bias_initializer, kernel_regularizer, bias_regularizer, activity_regularizer, kernel_constraint, bias_constraint, **kwargs)\n    464         activation=activations.get(activation),\n    465         use_bias=use_bias,\n--&gt; 466         kernel_initializer=initializers.get(kernel_initializer),\n    467         bias_initializer=initializers.get(bias_initializer),\n    468         kernel_regularizer=regularizers.get(kernel_regularizer),\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\initializers.py in get(identifier)\n    153     return None\n    154   if isinstance(identifier, dict):\n--&gt; 155     return deserialize(identifier)\n    156   elif isinstance(identifier, six.string_types):\n    157     config = {'class_name': str(identifier), 'config': {}}\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\initializers.py in deserialize(config, custom_objects)\n    145       module_objects=globals(),\n    146       custom_objects=custom_objects,\n--&gt; 147       printable_module_name='initializer')\n    148 \n    149 \n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    161       cls = module_objects.get(class_name)\n    162       if cls is None:\n--&gt; 163         raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)\n    164     if hasattr(cls, 'from_config'):\n    165       arg_spec = tf_inspect.getfullargspec(cls.from_config)\n\nValueError: Unknown initializer: GlorotUniform\n<\/code><\/pre>\n\n<p>Stackoverflow is asking me to add details while I have none to add. Or I am not sure what to add. Please help.<\/p>\n","answers":[{"AnswerId":"53056693","CreationDate":"2018-10-30T02:49:47.957","ParentId":null,"OwnerUserId":"4987390","Title":null,"Body":"<ol>\n<li><p>Make sure you have the newest version of <code>Keras<\/code> and <code>tensorflow<\/code> (which are <code>2.4.4<\/code> and <code>1.11.0<\/code>) by running either <code>pip install keras tensorflow<\/code> or <code>conda install keras tensorflow<\/code>.<\/p><\/li>\n<li><p>In case it is Google Colab that uses deprecated objects, you may need to  use custom objects:<\/p><\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.utils import CustomObjectScope\nfrom keras.initializers import glorot_uniform\n\nwith CustomObjectScope({'GlorotUniform': glorot_uniform()}):\n    model = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>Not sure  if this is your case though.<\/p>\n"}]}
+{"QuestionId":53051519,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-10-29T18:16:14.443","AcceptedAnswerId":null,"OwnerUserId":8767101.0,"Title":"Problem with LSTM __call__ initial_state parameter","Body":"<p>I have this piece of code:<\/p>\n\n<pre><code>s = Input(shape=(self.s_length,), name=\"s\")\nz = Input(shape=(self.z_length,), name=\"z\")\ndecoder_inputs = [s, z]\n\n# latent.shape: (None, self.s_length + self.z_length)\nlatent = Concatenate(name=\"latent_concat\")([s, z])\n\n# get initial state of high decoder\n# init_state.shape: (None, X_high_size)\ninit_state = Dense(X_high_size, activation=\"tanh\", name=\"hidden_state_init\")(latent)\n\n# high decoder produces embeddings\n# h_X.shape: (None, n_embeddings, self.s_length + self.z_length)\nh_X = RepeatVector(n_embeddings, name=\"latent_repeat\")(latent)\nfor l in range(X_high_depth):\n    h_X = LSTM(\n        X_high_size,\n        return_sequences=True,\n        activation=\"relu\",\n        name=f\"high_encoder_{l}\"\n    )(h_X, initial_state=[init_state, init_state])\n<\/code><\/pre>\n\n<p>When I run it, during <em>training<\/em> time I get the error:<\/p>\n\n<pre><code>File \"\/home\/xxx\/anaconda3\/envs\/mygan\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1215, in train_on_batch\n    outputs = self.train_function(ins)\n  File \"\/home\/xxx\/anaconda3\/envs\/mygan\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2666, in __call__\n    return self._call(inputs)\n  File \"\/home\/xxx\/anaconda3\/envs\/mygan\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2636, in _call\n    fetched = self._callable_fn(*array_vals)\n  File \"\/home\/xxx\/anaconda3\/envs\/mygan\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1382, in __call__\n    run_metadata_ptr)\n  File \"\/home\/xxx\/anaconda3\/envs\/mygan\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py\", line 519, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 's' with dtype float and shape [?,16]\n     [[Node: s = Placeholder[dtype=DT_FLOAT, shape=[?,16], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n     [[Node: loss_2\/z_discriminator_loss\/Mean_3\/_1073 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_41285_loss_2\/z_discriminator_loss\/Mean_3\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>(<code>16<\/code> is the input <code>s<\/code> dimension, by the way. )\nHowever I noticed that if I just <em>omit<\/em> the <code>initial_state=[init_state, init_state]<\/code> argument from the LSTM layer call, everything runs smoothly.\nThe code seems correct to me, I don't know what I'm doing wrong...might be a Keras bug?<\/p>\n","answers":[{"AnswerId":"53056064","CreationDate":"2018-10-30T01:18:48.300","ParentId":null,"OwnerUserId":"10574412","Title":null,"Body":"<p>A placeholder is an empty input tensor that is expecting to receive the data when fitting or predicting. \nIf you specify the \"initial state\", the LSTM expects real data, real numbers, while when you call them, the placeholder is still not populated.<\/p>\n"}]}
+{"QuestionId":53051913,"AnswerCount":1,"Tags":"<python><pytorch><gradient>","CreationDate":"2018-10-29T18:44:01.440","AcceptedAnswerId":53052767.0,"OwnerUserId":3008221.0,"Title":"Modifying a pytorch tensor and then getting the gradient lets the gradient not work","Body":"<p>I am a beginner in pytorch and I face the following issue: <\/p>\n\n<p>When I get the gradient of the below tensor (note that I use some variable x in some way as you can see below), I get the gradient:<\/p>\n\n<pre><code>import torch\nmyTensor = torch.randn(2, 2,requires_grad=True)\nwith torch.enable_grad():\n    x=myTensor.sum() *10\nx.backward()\nprint(myTensor.grad)\n<\/code><\/pre>\n\n<p>Now, if I try to modify an element of <code>myTensor<\/code>, I get the error of <code>leaf variable has been moved into the graph interior<\/code>. See this code:<\/p>\n\n<pre><code>import torch\nmyTensor = torch.randn(2, 2,requires_grad=True)\nmyTensor[0,0]*=5\nwith torch.enable_grad():\n    x=myTensor.sum() *10\nx.backward()\nprint(myTensor.grad)\n<\/code><\/pre>\n\n<p>What is wrong with my latter code? And how do I correct it?<\/p>\n\n<p>Any help would be highly appreciated. Thanks a lot!<\/p>\n","answers":[{"AnswerId":"53052767","CreationDate":"2018-10-29T19:48:10.007","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The problem here is that this line represents an <em>in-place<\/em> operation:<\/p>\n\n<pre><code>myTensor[0,0]*=5\n<\/code><\/pre>\n\n<p>And <em>PyTorch<\/em> or more precisely <em>autograd<\/em> is not very good in handling <em>in-place<\/em> operations, especially on those tensors with the <code>requires_grad<\/code> flag set to <code>True<\/code>.<\/p>\n\n<p>You can also take a look here:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd\" rel=\"noreferrer\">https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#in-place-operations-with-autograd<\/a><\/p>\n\n<p>Generally you should avoid <em>in-place<\/em> operations where it is possible, in some cases it can work, but you should always avoid <em>in-place<\/em> operations on tensors where you set <code>requires_grad<\/code> to <code>True<\/code>.<\/p>\n\n<p>Unfortunately there are not many pytorch functions to help out on this problem. So you would have to use a helper tensor to avoid the <code>in-place<\/code> operation in this case:<\/p>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\nmyTensor = torch.randn(2, 2,requires_grad=True)\nhelper_tensor = torch.ones(2, 2)\nhelper_tensor[0, 0] = 5\nnew_myTensor = myTensor * helper_tensor # new tensor, out-of-place operation\nwith torch.enable_grad():\n    x=new_myTensor.sum() *10 # of course you need to use the new tensor\nx.backward()                 # for further calculation and backward\nprint(myTensor.grad)\n<\/code><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<pre><code>tensor([[50., 10.],\n        [10., 10.]])\n<\/code><\/pre>\n\n<p>Unfortunately this is not very nice and I would appreciate if there would be a better or nicer solution out there. <br>But for all I know in the current version <em>(0.4.1)<\/em> you will have to got with this workaround for tensors with gradient resp. <code>requires_grad=True<\/code>. <\/p>\n\n<p>Hopefully for future versions there will be a better solution.<\/p>\n\n<hr>\n\n<p>Btw. if you activate the gradient later you can see that it works just fine:<\/p>\n\n<pre><code>import torch\nmyTensor = torch.randn(2, 2,requires_grad=False) # no gradient so far\nmyTensor[0,0]*=5                                 # in-place op not included in gradient\nmyTensor.requires_grad = True                    # activate gradient here\nwith torch.enable_grad():\n    x=myTensor.sum() *10\nx.backward()                                     # no problem here\nprint(myTensor.grad)\n<\/code><\/pre>\n\n<p>But of course this will yield to a different result:<\/p>\n\n<pre><code>tensor([[10., 10.],\n        [10., 10.]])\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":53051989,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><deep-learning>","CreationDate":"2018-10-29T18:48:58.190","AcceptedAnswerId":54465339.0,"OwnerUserId":3889510.0,"Title":"How to use tf.contrib.rnn.convLSTMCell class in tensorflow","Body":"<p>I would like to use a convolution LSTM in my research but I'm having a difficult time figuring out the exact way to implement this class in tensorflow. Here is what I have so far. I get no errors, but I am seriously doubting my implementation. Can anyone confirm if I am doing this correctly?<\/p>\n\n<pre><code>n_input = 4\nx = tf.placeholder(tf.float32,shape=[None,n_input,HEIGHT,WIDTH,2])\ny = tf.placeholder(tf.float32,shape=[None,HEIGHT,WIDTH,2])\n\nconvLSTM_cell = tf.contrib.rnn.ConvLSTMCell(\n    conv_ndims=2,\n    input_shape = [HEIGHT,WIDTH,DEPTH],\n    output_channels=2,\n    kernel_shape=[3,3]\n    )\n\noutputs, states = tf.nn.dynamic_rnn(convLSTM_cell, x, dtype=tf.float32)\n\nweights =  tf.Variable(tf.random_normal([3,3,2,2]))\nbiases = tf.Variable(tf.random_normal([2]))\n\nconv_out = tf.nn.conv2d(outputs[-1],weights,strides=[1,1,1,1],padding='SAME')\nout = tf.nn.sigmoid(conv_out + biases)\n<\/code><\/pre>\n\n<p>UPDATE:<\/p>\n\n<p>printing the size of outputs gives the shape=(?,4,436,1024,2) but I think I want (?,5,436,1024,2) or (?,1,436,1024,2).<\/p>\n\n<p>UPDATE2:<\/p>\n\n<p>So according to a fellow lab mate, the 4 outputs corresponds to the lstm outputs for each frame and so it is working correctly. Apparently all I have to do is take output #4 and that is the predicted future time frame. <\/p>\n\n<p>A stackoverflow confirmation would put my mind at ease on this whole thing.<\/p>\n","answers":[{"AnswerId":"54465339","CreationDate":"2019-01-31T16:43:36.760","ParentId":null,"OwnerUserId":"9674919","Title":null,"Body":"<p>Yes, you are correct!\nThe output dimension will match the input dimension. If you actually want the <code>(?,5,436,1024,2)<\/code> output, you will have to look at the history, <code>state.h<\/code>. the last four <code>[-4]<\/code> of it will still correspond to the output.<\/p>\n"}]}
+{"QuestionId":53052188,"AnswerCount":1,"Tags":"<tensorflow><cuda><ubuntu-14.04>","CreationDate":"2018-10-29T19:03:26.670","AcceptedAnswerId":null,"OwnerUserId":1269942.0,"Title":"Will CUDA10 + CUDNN + tensorflow work on Ubuntu14.04?","Body":"<p>It is now Oct 29, 2018<\/p>\n\n<p>After much googling, I have not found a definitive answer or any examples of people using the latest cuda10 for tensorflow on ubuntu 14.04.<\/p>\n\n<p>My dilemma is whether to upgrade my OS (currently at 14.04) in order to run cuda9 so I can use the latest tensorflow version or use CUDA10 on my existing 14.04 install.  <\/p>\n\n<p>Note cuda9 does not support 14.04, however, Nvidia has indicated that 14.04 will be supported for cuda10.<\/p>\n\n<p>So, any examples\/experiences of people using tensorflow with cuda10 on ubuntu14.04 are keenly sought after!<\/p>\n\n<p>Also note cuda10 is not specifically supported by tensorflow...yet...they say \"soon\". But TF can be built from source with cuda10.<\/p>\n\n<p>This is a link for cuda10+tensorflow on ubuntu16.04: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22706\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22706<\/a><\/p>\n\n<p>The short answer, I realize, is \"try building it myself\". Before I do that, I thought I'd ask around. Thanks.<\/p>\n","answers":[{"AnswerId":"53395194","CreationDate":"2018-11-20T14:28:29.340","ParentId":null,"OwnerUserId":"10655413","Title":null,"Body":"<p>I don't know whether CUDA 10 can work well on Ubuntu 14.04, but I was managed to build TensorFlow with CUDA 10 on Ubuntu 18.04 with using <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/dgx\/tensorflow-release-notes\/running.html#running\" rel=\"nofollow noreferrer\">NVIDIA released docker image<\/a>. <\/p>\n\n<ul>\n<li>You can pull the 'TensorFlow Release 18.09' and try it on your current system.<\/li>\n<li>If the previous step does not work, consider upgrading your OS to 18.04.<\/li>\n<\/ul>\n\n<p>I wrote down my installation experience on <a href=\"https:\/\/f-exp.gitbook.io\/tensorflow-notebook\/\" rel=\"nofollow noreferrer\">this page<\/a>, you could read it for some detail if you need.<\/p>\n"}]}
+{"QuestionId":53052301,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-10-29T19:13:06.007","AcceptedAnswerId":null,"OwnerUserId":4354355.0,"Title":"Eval after every train epoch","Body":"<p>Trying to get a workflow using tensorflow object detection API.<\/p>\n\n<p>I have created a tfrecords for my training set (4070 images) and my validation set (1080 images).<\/p>\n\n<p>Training goes for 400 iterations then switches to evaluation where it takes a quite a long time (about 10 minutes) then proceeds to repeat the evaluation for ever. I've been confused why I'm getting poor performance but it seems that I train for only 10mins despite running for 24hours.<\/p>\n\n<p>Eval config is<\/p>\n\n<pre><code>   eval_config {\n  num_visualizations: 30\n  num_examples: 30\n  eval_interval_secs: 600\n  max_evals: 1\n  include_metrics_per_category: true\n  batch_size: 1\n}\neval_input_reader {\n  label_map_path: \"D:\/src\/Wind\/WindObjectDetector\/data\/labels.pbtxt\"\n  shuffle: false\n  num_epochs: 1\n  num_readers: 1\n  tf_record_input_reader {\n    input_path: \"D:\/src\/Wind\/WindObjectDetector\/data\/val.record\"\n  }\n  sample_1_of_n_examples: 1\n}\n<\/code><\/pre>\n\n<p>Training config:<\/p>\n\n<pre><code>train_config {\n  batch_size: 8\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n  data_augmentation_options {\n    ssd_random_crop {\n    }\n  }\n  optimizer {\n    rms_prop_optimizer {\n      learning_rate {\n        exponential_decay_learning_rate {\n          initial_learning_rate: 0.0010000000474974513\n          decay_steps: 1017\n          decay_factor: 0.949999988079071\n        }\n      }\n      momentum_optimizer_value: 0.8999999761581421\n      decay: 0.8999999761581421\n      epsilon: 1.0\n    }\n  }\n  fine_tune_checkpoint: \"D:\/src\/Wind\/WindObjectDetector\/experiments\/test2\/tl\/model.ckpt\"\n  from_detection_checkpoint: true\n  num_steps: 25400\n}\n<\/code><\/pre>\n\n<p>What I really want is the workflow I had doing regular image classification and segmentation. That is, train for one epoch (num_images\/\/batch_size iterations). Then evaluate on validation set for one epoch then repeat. So every loop I see every image once and evaluate on every (validation) image once. At this point I'd be happy to just have it train.<\/p>\n","answers":[{"AnswerId":"53060782","CreationDate":"2018-10-30T09:10:24.453","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>Training and evaluation are two seperated steps in the object detection API. You have to run <code>train.py<\/code> for training and <code>eval.py<\/code> for evaluation. If you want to evaluate during training process both scripts have to run parallel. I can not really say why you are training only 10 minutes, without seeing the training part of your config file. How many steps do you train? <\/p>\n\n<p>If you want to evaluate after every epoch, you have to generate a checkpoint after every single one (for changing the checkpoint generation frequency have a look <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5139#issuecomment-418963839\" rel=\"nofollow noreferrer\">here<\/a>). When a new checkpoint is generated the eval script will automatically evaluate this new one. <\/p>\n"}]}
+{"QuestionId":53052605,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T19:36:23.863","AcceptedAnswerId":53053242.0,"OwnerUserId":1410665.0,"Title":"ValueError: Shape of a new variable (local1\/weights) must be fully defined, but instead was (?, 1000)","Body":"<p>I'm new in Tensorflow. After convolution the shape of my layer is <code>shape=(5, 5, 5, 5), dtype=float32<\/code> but when I'm applying deconvolution, getting shape like <code>shape=(?, 25, 25, 640)<\/code>, <code>dtype=float32.<\/code> That means batch size is not showing properly (? sign) after deconvolution. For deconvolution, I used this <a href=\"https:\/\/riptutorial.com\/tensorflow\/example\/29767\/using-tf-nn-conv2d-transpose-for-arbitary-batch-sizes-and-with-automatic-output-shape-calculation-\" rel=\"nofollow noreferrer\">Deconvolution<\/a> function. <\/p>\n\n<blockquote>\n  <p>Error <strong>ValueError: Shape of a new variable (local1\/weights) must be\n  fully defined, but instead was (?, 1000).<\/strong> <\/p>\n<\/blockquote>\n\n<p>I already tried <a href=\"https:\/\/stackoverflow.com\/questions\/40430186\/tensorflow-valueerror-cannot-feed-value-of-shape-64-64-3-for-tensor-uplace\">example1<\/a>but didn't work well<\/p>\n","answers":[{"AnswerId":"53053004","CreationDate":"2018-10-29T20:03:24.193","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>the difference is that the example you send is a tensor getting the wrong data fed to it. Your problem is that the weights of a deconvolutional filter is not fully defined. The weights are not dependent on the batch size, and need to be of fixed size, hence the error. I know you understood the error, just want to make clear that the problem you have and the example has, is quite different.<\/p>\n\n<p>I recommend to use this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d_transpose\" rel=\"nofollow noreferrer\">function<\/a> instead:<\/p>\n\n<pre><code> tf.nn.conv2d_transpose()\n<\/code><\/pre>\n\n<p>it is defined like you'd do with a normal convolutional layer. It's default in TensorFlow, and I wonder why you didn't use it to start with?<\/p>\n"},{"AnswerId":"53053242","CreationDate":"2018-10-29T20:21:14.210","ParentId":null,"OwnerUserId":"2482068","Title":null,"Body":"<p>From the description of used <a href=\"https:\/\/riptutorial.com\/tensorflow\/example\/29767\/using-tf-nn-conv2d-transpose-for-arbitary-batch-sizes-and-with-automatic-output-shape-calculation-\" rel=\"nofollow noreferrer\">Deconvolution<\/a> function<\/p>\n\n<blockquote>\n<pre><code>  #Now output.get_shape() is equal (?,?,?,?) which can become a problem in the \n  #next layers. This can be repaired by reshaping the tensor to its shape:\n  output = tf.reshape(output, output_shape)\n  #now the shape is back to (?, H, W, C) or (?, C, H, W)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Batch size shouldn't be displayed, cause it designed to be unknown. It is made so to preserve an ability to process batches with different size (first dimension size). So that you can run model on batches of different size, for example, train on 5 and predict 20 images in one run.<\/p>\n\n<p>And fully agree with <a href=\"https:\/\/stackoverflow.com\/a\/53053004\/2482068\">T. Kelher<\/a>:<\/p>\n\n<blockquote>\n  <p>I recommend to use this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/conv2d_transpose\" rel=\"nofollow noreferrer\">function<\/a> instead:<\/p>\n\n<pre><code>tf.nn.conv2d_transpose()\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"53072210","CreationDate":"2018-10-30T20:12:50.377","ParentId":null,"OwnerUserId":"1410665","Title":null,"Body":"<p>The issue has been solved and the previous transpose\/deconvolution code is running nicely. Just we have to make some minor changes. We have to define batch size in the output shape. <\/p>\n"}]}
+{"QuestionId":53052674,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><loss-function>","CreationDate":"2018-10-29T19:41:40.830","AcceptedAnswerId":null,"OwnerUserId":7326400.0,"Title":"Custom loss function in keras returns nan loss","Body":"<p>Following is my custom loss function in keras.\n  Both <code>y_true<\/code> and <code>y_pred<\/code> have shape <code>(256,384,3)<\/code>.<\/p>\n\n<pre><code>def myloss(y_true,y_pred):\n    diff = y_pred - y_true\n    diff = diff[96:160 , 144:240]\n    return K.mean(K.square(diff), axis=-1)\n<\/code><\/pre>\n\n<p>Here I'm slicing a portion of the difference and finding the <code>mean_square_loss<\/code> for that. However, this function is returning <code>nan<\/code> as the loss. Why is that?<\/p>\n","answers":[]}
+{"QuestionId":53053138,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><lstm>","CreationDate":"2018-10-29T20:13:19.897","AcceptedAnswerId":null,"OwnerUserId":1757224.0,"Title":"Freezing an LSTM layer","Body":"<p>Is it possible to freeze one LSTM layer and build another LSTM layer on the top of it?<\/p>\n\n<p>The idea is quite simple but mechanism in Tensorflow makes it harder. All I need to do is to build one LSTM layer and save the model.  Then, I restore this model to get kernel (weight) matrix and biases using:<\/p>\n\n<p><code>tf.get_default_graph().get_tensor_by_name(\"rnn\/multi_rnn_cell\/cell_0\/LSTM_cell\/kernel:0\") tf.get_default_graph().get_tensor_by_name(\"rnn\/multi_rnn_cell\/cell_0\/LSTM_cell\/bias:0\")<\/code><\/p>\n\n<p>Then, I want to get these two tensors and keep them untrained and build another LSTM layer on the top of this then fetch the variables of the second layer.<\/p>\n\n<p>However, the only topic I came up related to what I want to do is <a href=\"https:\/\/stackoverflow.com\/questions\/42962281\/can-you-manually-set-tensorflow-lstm-weights\">here<\/a> and it says that it is not possible, because I cannot set weights manually. <\/p>\n\n<p>Doing this is super simple in feed forward neural network. <\/p>\n\n<p>Does anyone have any idea?<\/p>\n","answers":[]}
+{"QuestionId":53053335,"AnswerCount":2,"Tags":"<python><tensorflow><conv-neural-network><tensorflow.js>","CreationDate":"2018-10-29T20:27:49.277","AcceptedAnswerId":null,"OwnerUserId":4921713.0,"Title":"Uncaught Error: Error when checking : expected conv2d_input to have 4 dimension(s), but got array with shape [275,183,3]","Body":"<p>I performed following operations on the images before training my keras model:<\/p>\n\n<pre><code>for img in os.listdir(path):    \n    # convert to array\n    img_array = cv2.imread(os.path.join(path,img) ,cv2.IMREAD_GRAYSCALE) \n    # resize to normalize data size\n    new_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE)) \n    # add this to our training_data list\n    training_data.append([new_array, class_num]) \n\n#shuffle the data \nrandom.shuffle(training_data)\n\n#empty lists (X for features, y for labels)\nX = []\ny = []\n\nfor features,label in tqdm(training_data):\n    X.append(features)\n    y.append(label)\n\n\nX = np.array(X).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\n<\/code><\/pre>\n\n<p>I am then training the model. Here is the starting layer:<\/p>\n\n<pre><code>#start creating model \nmodel = Sequential()\n\nmodel.add(Conv2D(256, (3, 3), input_shape=X.shape[1:])\n<\/code><\/pre>\n\n<p>I am using the using the trained model to get some predictions( <strong>I am training the  model in python and then I am loading the model into Tensorflow.js<\/strong>) <\/p>\n\n<p>Code snippet for prediction<\/p>\n\n<pre><code>let imageTensor = tf.fromPixels(image);\nmodel.predict(imageTensor).print();\n<\/code><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<blockquote>\n  <p>Uncaught Error: Error when checking : expected conv2d_input to have 4\n  dimension(s), but got array with shape [275,183,3]<\/p>\n<\/blockquote>\n\n<p>Changing the above code to<\/p>\n\n<pre><code>let imageTensor = tf.fromPixels(image).resizeNearestNeighbor([50,50]).mean(2).toInt().expandDims(2);\nmodel.predict(imageTensor).print();\n<\/code><\/pre>\n\n<p>gives following error:<\/p>\n\n<blockquote>\n  <p>Uncaught Error: Error when checking : expected conv2d_input to have 4\n  dimension(s), but got array with shape [50,50,1]<\/p>\n<\/blockquote>\n\n<p>lastly, when I do this<\/p>\n\n<pre><code>let imageTensor = tf.fromPixels(image).resizeNearestNeighbor([50,50]).toInt().expandDims();\n model.predict(imageTensor).print();\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<blockquote>\n  <p>Error when checking : expected conv2d_input to have shape\n  [null,50,50,1] but got array with shape [1,50,50,3].<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53072574","CreationDate":"2018-10-30T20:39:54.323","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>It is related to a mismatch of the dimension of the input of the model and the dimension of the image passed in as parameter to the predict method.<\/p>\n\n<p>One might consider reshaping the image the following way:<\/p>\n\n<pre><code>imageTensor.reshape([-1, 50, 50, 3])\n<\/code><\/pre>\n"},{"AnswerId":"53057387","CreationDate":"2018-10-30T04:22:40.180","ParentId":null,"OwnerUserId":"10023182","Title":null,"Body":"<p>In the first example, with [256, (3,3), and the last thing], keras treats this list as having three dimensions or elements, when it is looking for four dimensions. Remove the parentheses to yield:<\/p>\n\n<pre><code>[256, 3, 3, input_shape=X.shape[1:]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53053649,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T20:54:02.923","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Tensorflow: Truncating a Tensor Seems to Have No Effect","Body":"<p>I'm using the <code>tf.data.Dataset<\/code> API and am trying to truncate a bunch of tensors to length 100.  Here's what my <code>dataset<\/code> looks like:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({'reviews': x}, y))\n<\/code><\/pre>\n\n<p>My reviews are just movie reviews (strings), so I perform some preprocessing and map that function on my dataset:<\/p>\n\n<pre><code>def preprocess(x, y):\n    # split on whitespace\n    x['reviews'] = tf.string_split([x['reviews']])\n    # turn into integers\n    x['reviews'], y = data_table.lookup(x['reviews']), labels_table.lookup(y)\n    x['reviews'] = tf.sparse_tensor_to_dense(x['reviews'])\n    # truncate at length 100\n    x['reviews'] = x['reviews'][:100]\n    x['reviews'] = x['reviews'][0]\n    x['reviews'] = tf.pad(x['reviews'],\n           paddings=[[100 - tf.shape(x['reviews'])[0], 0]],\n           mode='CONSTANT',\n           name='pad_input',\n           constant_values=0)\n    return x, y\n\ndataset = dataset.map(preprocess)\n<\/code><\/pre>\n\n<p>However, my code fails with:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Paddings must be non-negative: -140 0\n<\/code><\/pre>\n\n<p>On an input of length 240. So, it seems like my padding step calculates 100 - 240 = -140 and I get this error.<\/p>\n\n<p>Here's my question: how is this possible, given that I truncate to length 100 with:<\/p>\n\n<pre><code>x['reviews'] = x['reviews'][:100]\n<\/code><\/pre>\n\n<p>It seems clear that this line isn't having any effect, so I'm trying to understand why. The docs are very clear that this is acceptable syntactic sugar for <code>tf.slice<\/code>: <\/p>\n\n<pre><code>Note that tf.Tensor.getitem is typically a more pythonic way to \n\nperform slices, as it allows you to write foo[3:7, :-2] instead of \n\ntf.slice(foo, [3, 0], [4, foo.get_shape()[1]-2]).\n<\/code><\/pre>\n\n<p>Any ideas?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53054668,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T22:14:31.177","AcceptedAnswerId":53120602.0,"OwnerUserId":2927842.0,"Title":"Assign values between 0 and N -1 for a vector of length L with N different elements in Tensorflow","Body":"<p>Given a 1-D tensor T of length L which has just N different values, how can I convert it to a tensor T2 of length L with values between 0 and N - 1 corresponding to the values of the original tensor T.\nExample:<\/p>\n\n<pre><code>T  = [45, 58, 72, 33, 45, 58, 58, 33]\nT2 = [ 0,  1,  2,  3,  0,  1,  1,  3]\n<\/code><\/pre>\n\n<p>The ordering is not important for example this also would be OK:<\/p>\n\n<pre><code>T2 = [1, 0, 2, 3, 1, 0 , 0, 3]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53120602","CreationDate":"2018-11-02T14:38:14.053","ParentId":null,"OwnerUserId":"10596820","Title":null,"Body":"<p>Maybe try: <\/p>\n\n<p>tf.unique(T)[1]<\/p>\n\n<p>tf.unique_with_counts(T)[1]<\/p>\n"}]}
+{"QuestionId":53054699,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><random>","CreationDate":"2018-10-29T22:17:45.300","AcceptedAnswerId":null,"OwnerUserId":10577372.0,"Title":"Why is random number generator tf.random_uniform in tensorflow much faster than the numpy equivalent","Body":"<p>The following code is what I used to test the performance:<\/p>\n\n<pre><code>import time\nimport numpy as np\nimport tensorflow as tf\n\nt = time.time()\nfor i in range(400):\n    a = np.random.uniform(0,1,(1000,2000))\nprint(\"np.random.uniform: {} seconds\".format(time.time() - t))\n\nt = time.time()\nfor i in range(400):\n    a = np.random.random((1000,2000))\nprint(\"np.random.random:  {} seconds\".format(time.time() - t))\n\nt = time.time()\nfor i in range(400):\n    a = tf.random_uniform((1000,2000),dtype=tf.float64);\nprint(\"tf.random_uniform: {} seconds\".format(time.time() - t))\n<\/code><\/pre>\n\n<p>All the three segments generate a uniformly random 1000*2000 matrix in double precision 400 times. The timing differences are striking. On my Mac,<\/p>\n\n<pre><code>np.random.uniform: 10.4318959713 seconds\nnp.random.random:  8.76161003113 seconds\ntf.random_uniform: 1.21312117577 seconds\n<\/code><\/pre>\n\n<p>Why is tensorflow much faster than numpy?<\/p>\n","answers":[{"AnswerId":"53054930","CreationDate":"2018-10-29T22:39:45.303","ParentId":null,"OwnerUserId":"4606149","Title":null,"Body":"<p>You've just created a computational graph that generates an operation to output random numbers. For the values to be calculated, you must execute the graph in a <code>tf.Session<\/code>.<\/p>\n\n<pre><code>\/\/ build the graph\na = tf.random_uniform((1000,2000))\n\n\/\/ run the graph\nwith tf.Session() as sess:\n    t = time.time()\n    for i in range(400):\n        computed_rand_values = sess.run(a)\n    \/\/print(...)\n<\/code><\/pre>\n\n<p>I did not tested it, but I'm sure that the calculation time will be longer than the result before<\/p>\n"},{"AnswerId":"53054887","CreationDate":"2018-10-29T22:36:35.190","ParentId":null,"OwnerUserId":"567620","Title":null,"Body":"<p><code>tf.random_uniform<\/code> in this case is returning an <em>unevaluated<\/em> Tensor type, <code>tensorflow.python.framework.ops.Tensor<\/code>, and if you setup a session context in which to evaluate <code>a<\/code> in the <code>tf.random_uniform<\/code> case, you'll see that it takes a while too.<\/p>\n\n<p>For example, here in the <code>tf<\/code> case I added <code>sess.run<\/code> (on a CPU-only machine), and it takes ~16 seconds to evaluate and materialize, which makes sense given some overhead to marshal into the numpy data type on output.<\/p>\n\n<pre><code>In [1]: %cpaste\nPasting code; enter '--' alone on the line to stop or use Ctrl-D.\n:import time\nimport numpy as np\nimport tensorflow as tf\n\nt = time.time()\nfor i in range(400):\n    a = np.random.uniform(0,1,(1000,2000))\nprint(\"np.random.uniform: {} seconds\".format(time.time() - t))\n\nt = time.time()\nfor i in range(400):\n    a = np.random.random((1000,2000))\nprint(\"np.random.random:  {} seconds\".format(time.time() - t))\n\nsess = tf.Session()\nt = time.time()\nfor i in range(400):\n    a = sess.run(tf.random_uniform((1000,2000),dtype=tf.float64))\nprint(\"tf.random_uniform: {} seconds\".format(time.time() - t))::::::::::::::::::\n:--\nnp.random.uniform: 11.066569805145264 seconds\nnp.random.random:  9.299575090408325 seconds\n2018-10-29 18:34:58.612160: W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.\n2018-10-29 18:34:58.612191: W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.\n2018-10-29 18:34:58.612210: W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.\ntf.random_uniform: 16.619441747665405 seconds\n<\/code><\/pre>\n"}]}
+{"QuestionId":53054954,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-29T22:42:29.103","AcceptedAnswerId":null,"OwnerUserId":10577367.0,"Title":"How can I get the indices of the data used in every batch?","Body":"<p>I need to save the indices of the data that are used in every mini-batch.<\/p>\n\n<p>For example if my data is:<\/p>\n\n<pre><code>x = np.array([[1.1], [2.2], [3.3], [4.4]])\n<\/code><\/pre>\n\n<p>and the first mini-batch is <code>[1.1]<\/code> and <code>[3.3]<\/code>, then I want to store <code>0<\/code> and <code>2<\/code> (since <code>[1.1]<\/code> is the 0th observations and <code>[3.3]<\/code> is the 2nd observation).<\/p>\n\n<p>I am using tensorflow in eager execution with the keras.sequential APIs.<\/p>\n\n<p>As far as I can tell from reading the source code, this information is not stored anywhere so I was unable to do this with a callback.<\/p>\n\n<p>I am currently solving my problem by creating an object that stores the indices.<\/p>\n\n<pre><code>class IndexIterator(object):\n    def __init__(self, n, n_epochs, batch_size, shuffle=True):\n        data_ix = np.arange(n)\n        if shuffle:\n            np.random.shuffle(data_ix)\n\n        self.ix_batches = np.array_split(data_ix, np.ceil(n \/ batch_size))\n        self.batch_indices = []\n\n    def generate_arrays(self, x, y):\n        batch_ixs = np.arange(len(self.ix_batches))\n        while 1: \n            np.random.shuffle(batch_ixs)\n            for batch in batch_ixs:\n                self.batch_indices.append(self.ix_batches[batch])\n                yield (x[self.ix_batches[batch], :], y[self.ix_batches[batch], :])\n\ndata_gen = IndexIterator(n=32, n_epochs=100, batch_size=16)\ndnn.fit_generator(data_gen.generate_arrays(x, y), \n                  steps_per_epoch=2, \n                  epochs=100)\n# This is what I am looking for\nprint(data_gen.batch_indices)\n<\/code><\/pre>\n\n<p><strong>Is there no way to do this using a tensorflow callback?<\/strong><\/p>\n","answers":[{"AnswerId":"53074351","CreationDate":"2018-10-30T23:39:50.683","ParentId":null,"OwnerUserId":"2426955","Title":null,"Body":"<p>Not sure if this will be more efficient than your solution, but is certainly more general.<\/p>\n\n<p>If you have training data with <code>n<\/code> indices you can create a secondary <code>Dataset<\/code> that contains only these indices and zip it with the \"real\" dataset.  <\/p>\n\n<p>I.E.<\/p>\n\n<pre><code>real_data = tf.data.Dataset ... \nindices = tf.data.Dataset.from_tensor_slices(tf.range(data_set_length)))\ntotal_dataset = tf.data.Dataset.zip((real_data, indices))\n\n# Perform optional pre-processing ops.\n\niterator = total_dataset.make_one_shot_iterator()\n\n# Next line yields `(original_data_element, index)`\nitem_and_index_tuple = iterator.get_next() \n<\/code><\/pre>\n\n<p>` <\/p>\n"}]}
+{"QuestionId":53054968,"AnswerCount":2,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-10-29T22:43:43.680","AcceptedAnswerId":null,"OwnerUserId":8177201.0,"Title":"How to save a Tensorflow.js model?","Body":"<p>I would like to make a user interface that creates,saves and trains tensorflow.js models. But i can't save a model after creating it. I even copied this code from the tensorflow.js documenation but it does't work: <\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential(\r\n     {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});\r\nconsole.log('Prediction from original model:');\r\nmodel.predict(tf.ones([1, 3])).print();\r\n\r\nconst saveResults = await model.save('localstorage:\/\/my-model-1');\r\n\r\nconst loadedModel = await tf.loadModel('localstorage:\/\/my-model-1');\r\nconsole.log('Prediction from loaded model:');\r\nloadedModel.predict(tf.ones([1, 3])).print();<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I always get the error message \"<em>Uncaught SyntaxError: await is only valid in async function\"<\/em> .How can I fix this? thanks!<\/p>\n","answers":[{"AnswerId":"53055039","CreationDate":"2018-10-29T22:50:50.157","ParentId":null,"OwnerUserId":"9507270","Title":null,"Body":"<p>You need to be in an async environment. Either create an async function (<code>async function name(){...}<\/code>) and call it when you need to or the shortest way would be a self invoking async arrow function:<\/p>\n\n<pre><code>(async ()=&gt;{\n   \/\/you can use await in here\n})()\n<\/code><\/pre>\n"},{"AnswerId":"53055156","CreationDate":"2018-10-29T23:04:02.843","ParentId":null,"OwnerUserId":"6427965","Title":null,"Body":"<p>Create an async function and invoke it:<\/p>\n\n<pre><code>async function main() {\n  const model = tf.sequential({\n    layers: [tf.layers.dense({ units: 1, inputShape: [3] })]\n  });\n  console.log(\"Prediction from original model:\");\n  model.predict(tf.ones([1, 3])).print();\n\n  const saveResults = await model.save(\"localstorage:\/\/my-model-1\");\n\n  const loadedModel = await tf.loadModel(\"localstorage:\/\/my-model-1\");\n  console.log(\"Prediction from loaded model:\");\n  loadedModel.predict(tf.ones([1, 3])).print();\n}\n\nmain();\n<\/code><\/pre>\n"}]}
+{"QuestionId":53055101,"AnswerCount":2,"Tags":"<python><pytorch><tensor>","CreationDate":"2018-10-29T22:57:27.823","AcceptedAnswerId":53060074.0,"OwnerUserId":5196092.0,"Title":"Why does Pytorch expect a DoubleTensor instead of a FloatTensor?","Body":"<p>From everything I see online, <code>FloatTensors<\/code> are Pytorch's default for everything, and when I create a tensor to pass to my generator module it is a <code>FloatTensor<\/code>, but when I try to run it through a linear layer it complains that it wants a <code>DoubleTensor<\/code>. <\/p>\n\n<pre><code>class Generator(nn.Module):\n  def __init__(self):\n    super(Generator, self).__init__()\n    self.fully_connected = nn.Linear(100, 1024*4*4, bias=False)\n\n  def forward(self, zvec):\n    print(zvec.size())\n    fc = self.fully_connected(zvec)\n    return(fc.size())\n\ngen = Generator();\n\ngen(torch.from_numpy(np.random.normal(size=100)))\n<\/code><\/pre>\n\n<p>Which produces <\/p>\n\n<pre><code>RuntimeError: Expected object of type torch.DoubleTensor but found type torch.FloatTensor for argument #2 'mat2'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53055191","CreationDate":"2018-10-29T23:08:25.563","ParentId":null,"OwnerUserId":"5196092","Title":null,"Body":"<p>Numpy returns a 64 bit floating point number, which is cast to <code>DoubleTensor<\/code> when <code>.from_numpy()<\/code> is called. The confusing part is, if you run <code>print(type(zvec))<\/code> it returns <code>torch.Tensor<\/code> whether or not it is a float or a double.<\/p>\n"},{"AnswerId":"53060074","CreationDate":"2018-10-30T08:23:18.077","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The Problem here is that your numpy input uses <code>double<\/code> as data type the same data type is also applied on the resulting tensor. <\/p>\n\n<p>The <code>weights<\/code> of your layer <code>self.fully_connected<\/code> on the other hand are <code>float<\/code>. When feeding data trough the layer a matrix multiplication is applied and this multiplication requires both matrices to be of same data type.<\/p>\n\n<p><strong>So you have two solutions:<\/strong><\/p>\n\n<ul>\n<li><strong>You can convert your input to float:<\/strong><\/li>\n<\/ul>\n\n<p><em>By changing:<\/em><\/p>\n\n<pre><code>gen(torch.from_numpy(np.random.normal(size=100)))\n<\/code><\/pre>\n\n<p><em>To:<\/em><\/p>\n\n<pre><code>gen(torch.from_numpy(np.random.normal(size=100)).float())\n<\/code><\/pre>\n\n<p>Your input which is fed into <code>gen<\/code> will be converted to <code>float<\/code> then.<\/p>\n\n<p><em>Full working code for converting inputs:<\/em><\/p>\n\n<pre><code>from torch import nn\nimport torch\nimport numpy as np\nclass Generator(nn.Module):\n    def __init__(self):\n        super(Generator, self).__init__()\n        self.fully_connected = nn.Linear(100, 1024*4*4, bias=False)\n\n    def forward(self, zvec):\n        print(zvec.size())\n        fc = self.fully_connected(zvec)\n        return(fc.size())\n\ngen = Generator();\ngen(torch.from_numpy(np.random.normal(size=100)).float()) # converting network input to float\n<\/code><\/pre>\n\n<hr>\n\n<ul>\n<li><strong>Or alternatively you can convert your layer weights to double:<\/strong><\/li>\n<\/ul>\n\n<p>If you need the <em>double precision<\/em> you can also convert your <code>weights<\/code> to <code>double<\/code>.<\/p>\n\n<p><em>Change this line:<\/em><\/p>\n\n<pre><code>self.fully_connected = nn.Linear(100, 1024*4*4, bias=False)\n<\/code><\/pre>\n\n<p><em>Just to:<\/em><\/p>\n\n<pre><code>self.fully_connected = nn.Linear(100, 1024*4*4, bias=False).double()\n<\/code><\/pre>\n\n<p><em>Full working code for converting weights:<\/em><\/p>\n\n<pre><code>from torch import nn\nimport torch\nimport numpy as np\nclass Generator(nn.Module):\n    def __init__(self):\n        super(Generator, self).__init__()\n        self.fully_connected = nn.Linear(100, 1024*4*4, bias=False).double() # converting layer weights to double()\n\n    def forward(self, zvec):\n        print(zvec.size())\n        fc = self.fully_connected(zvec)\n        return(fc.size())\n\ngen = Generator();\ngen(torch.from_numpy(np.random.normal(size=100)))\n<\/code><\/pre>\n\n<hr>\n\n<p>So both ways should work for you, but if you don't need the extra precision of <code>double<\/code> you should go with <code>float<\/code> as <code>double<\/code> requires more computational power.<\/p>\n"}]}
+{"QuestionId":53055122,"AnswerCount":0,"Tags":"<python><tensorflow><keras><gpu>","CreationDate":"2018-10-29T23:00:00.750","AcceptedAnswerId":null,"OwnerUserId":9633432.0,"Title":"Specifying the GPU ID using tensorflow backend","Body":"<p>I am trying to train the two different models using GPU ID.\nI have tried the command <code>CUDA_VISIBLE_DEVICES=1 python filename.py<\/code> but it picks up the GPU <code>0<\/code> rather than <code>1<\/code>.\nI also added the os environment variable in my code, but I got the same behavior.\nI am not sure how can I fix this as this is first time to use GPU.<\/p>\n","answers":[]}
+{"QuestionId":53055239,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-29T23:16:00.223","AcceptedAnswerId":null,"OwnerUserId":2058333.0,"Title":"TensorFlow conv2d dataset shape not accepted","Body":"<p>Trying to train a cnn2d with a list of two 8x8 images.\nThis is what I get in return. Do I have to flatten the 2x8x8 array somehow?\nI am not sure what a supported type would be.<\/p>\n\n<p>The images are a 8x8 checker board before and after a move.<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class \n'tensorflow.python.data.ops.dataset_ops.TensorSliceDataset'&gt; to Tensor.\nContents: &lt;TensorSliceDataset shapes: ((2, 8, 8), ()), types: \n(tf.float64, tf.float64)&gt;. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>Sorry, I'm just starting out with <code>TF<\/code>.<\/p>\n","answers":[{"AnswerId":"53055656","CreationDate":"2018-10-30T00:15:36.320","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>tensorflow plays around with Tensors as returned by almost all tensorflow layers and placeholders, etc. Check <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">this link<\/a> and look at the return parameter, same with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\">this one<\/a>. That's also what your error is pointing to, it wants a tensor, not a TensorSliceDataset. I'm not familiar at all with tensorflow datasets, for that I gladly refer to some Google links: <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">1<\/a>. Doing some <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets_for_estimators#slices\" rel=\"nofollow noreferrer\">lookup work<\/a>, I've the feeling you're trying an mnist example, and something went wrong.<\/p>\n\n<p>If you'd want to share some code, then I can look at it tomorrow, perhaps the line where it goes wrong is enough, because I've the feeling you're passing data into the graph, instead of tensors. Data is always fed afterwards through a sess.run() or similar.<\/p>\n\n<p>Hope I helped a bit!<\/p>\n"},{"AnswerId":"53055510","CreationDate":"2018-10-29T23:54:05.290","ParentId":null,"OwnerUserId":"10349960","Title":null,"Body":"<p>Try to use .reshape(2, 64) method. Also, I recommend you to check out some codes dealing with MNIST dataset, which is the basic dataset to train yourself to deal with images (it's a database of handwritten digits). The images in the dataset are reshaped to (784,) from original (28,28).<\/p>\n"}]}
+{"QuestionId":53055601,"AnswerCount":1,"Tags":"<shell><gpu><caffe>","CreationDate":"2018-10-30T00:07:33.583","AcceptedAnswerId":null,"OwnerUserId":10312435.0,"Title":"Shell script problem about caffe framework","Body":"<p>When I run caffe shell, I meet a code:<\/p>\n\n<pre><code>--gpu 2 2&gt;&amp;1|tee .\/scripts_remote\/cuhk_pair\/train.log\n<\/code><\/pre>\n\n<p>I don't know its function is - what does it mean?<\/p>\n","answers":[{"AnswerId":"53098029","CreationDate":"2018-11-01T09:02:53.290","ParentId":null,"OwnerUserId":"9417105","Title":null,"Body":"<p>1) <code>2&gt;&amp;1<\/code>\n   2 -> STDERR - Standard Error\n   1 -> STDOUT - STandard Output<\/p>\n\n<p>We are redirecting error to same location where output will be directed.<\/p>\n\n<p>2) <code>|<\/code> - It's pipe between output of command <code>2&gt;&amp;1<\/code> and tee command i.e. out put of command on left of it (<code>2&gt;&amp;1<\/code>) will serve as input for that one at right (<code>tee<\/code>).<\/p>\n\n<p>3) <code>tee<\/code> - It put what it reads (here STDOUT and STDERR from command at left of pipe) to STDOUT and file provided (.\/scripts_remote\/cuhk_pair\/train.log).<\/p>\n\n<p>Finally we are redirecting all error and output of command to standard output and log file.<\/p>\n"}]}
+{"QuestionId":53056104,"AnswerCount":0,"Tags":"<opencv><caffe><openpose>","CreationDate":"2018-10-30T01:25:17.593","AcceptedAnswerId":null,"OwnerUserId":4642437.0,"Title":"Setup of Openpose: opencv2\/core.hpp not found","Body":"<p>I'm trying to setup Openpose + Pyopenpose following in a container following this gist: <a href=\"https:\/\/github.com\/christian-lanius\/PyOpenPose\" rel=\"nofollow noreferrer\">https:\/\/github.com\/christian-lanius\/PyOpenPose<\/a>.<\/p>\n\n<p>That works fine until the caffe compilation step, during which a <code>opencv2\/core.hpp not found<\/code> exception is thrown (note that the OpenCV installation step terminates successfully). While there are a number of rather similar questions around, I nonetheless couldn't find a viable solution in my context.<\/p>\n\n<p>I'm assuming I have to somehow link OpenCV so it can be found by other applications - do you guys have any idea how to fix or circumvent this error?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53056550,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-10-30T02:29:27.497","AcceptedAnswerId":null,"OwnerUserId":9729859.0,"Title":"How can I test my handwritten digit model using keras to detect new input image?","Body":"<p>I have trained my model for handwritten digit recognition  using keras(python) . Now I want to test my model for a new external input image of a handwritten digit.How should  I preprocessed  my image and how to get actual output?\nplease find this github link.  <a href=\"https:\/\/github.com\/Kundan8296\/Machine-Learning\/blob\/master\/Handwritten%20digit.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Kundan8296\/Machine-Learning\/blob\/master\/Handwritten%20digit.ipynb<\/a><\/p>\n","answers":[{"AnswerId":"53258100","CreationDate":"2018-11-12T08:12:51.220","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You should preprocess the new image the same way you preprocessed images for the training and then call the function <code>model.predict(external_image_preprocessed)<\/code> where <code>external_image_preprocessed<\/code>, as the name suggests, is a image you want to predict but preprocessed the same way you preprocessed the images for training.<\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
+{"QuestionId":53056992,"AnswerCount":1,"Tags":"<python><windows><tensorflow>","CreationDate":"2018-10-30T03:32:38.587","AcceptedAnswerId":null,"OwnerUserId":10577930.0,"Title":"having a problem offline install tensorflow on windows","Body":"<p>I am using Python 3.6 from Anaconda 4.3.1 on Windows 10. I'm trying to offline install TensorFlow\uff0c and i have download a .whl file to install module.\n<a href=\"https:\/\/i.stack.imgur.com\/sRlGY.jpg\" rel=\"nofollow noreferrer\">install tensorflow warning<\/a><\/p>\n\n<p>But I have installed the tensorboard 1.12 moudule on my computer. \n<a href=\"https:\/\/i.stack.imgur.com\/z6IPl.jpg\" rel=\"nofollow noreferrer\">I have installed the tensorboard moudule<\/a><\/p>\n\n<p>I have no way to install tensorflow now. please help me\uff0c thanks\uff01<\/p>\n","answers":[{"AnswerId":"53059561","CreationDate":"2018-10-30T07:46:44.710","ParentId":null,"OwnerUserId":"10577930","Title":null,"Body":"<p>I have solved this problem. Tensorflow 1.11 need tensorboard version 1.11, but I have installed tensorboard 1.12.<\/p>\n"}]}
+{"QuestionId":53057591,"AnswerCount":2,"Tags":"<python><tensorflow><importerror>","CreationDate":"2018-10-30T04:46:41.730","AcceptedAnswerId":null,"OwnerUserId":10578384.0,"Title":"ImportError: DLL load failed with error code -1073741795","Body":"<p><strong>My Computer System :<\/strong><\/p>\n\n<p>windows 7 64 bit\nCPU intel Core 2 Duo 8400\npython version 3.5.2rc1\ntensorflow 1.11\nI was trying to install tensorflow cpu version on windows7 , but always got error when import tensorflow.<\/p>\n\n<p>I read the common_installation_problems , tried many solution I found on github, stackoverflow, etc. I install visual studio 2015, visual studio 2017, Microsoft Visual C++ 2015 Redistributable Update 3 (64 bit), msvcp140.dll can find in both System32 and SysWow64 folder. But still can't import tensorflow.<\/p>\n\n<p>Is there something I missed out?<\/p>\n\n<h3>Log Info<\/h3>\n\n<pre><code>ImportError Traceback (most recent call last)\nc:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in \n57\n---&gt; 58 from tensorflow.python.pywrap_tensorflow_internal import *\n59 from tensorflow.python.pywrap_tensorflow_internal import version\n\nc:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in \n27 return _mod\n---&gt; 28 _pywrap_tensorflow_internal = swig_import_helper()\n29 del swig_import_helper\n\nc:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py in swig_import_helper()\n23 try:\n---&gt; 24 _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n25 finally:\n\nc:\\python\\lib\\imp.py in load_module(name, file, filename, details)\n241 else:\n--&gt; 242 return load_dynamic(name, filename, file)\n243 elif type_ == PKG_DIRECTORY:\n\nc:\\python\\lib\\imp.py in load_dynamic(name, path, file)\n341 name=name, loader=loader, origin=path)\n--&gt; 342 return _load(spec)\n343\n\nImportError: DLL load failed with error code -1073741795\n\nDuring handling of the above exception, another exception occurred:\n\nImportError Traceback (most recent call last)\nin \n----&gt; 1 import tensorflow as tf\n\nc:\\python\\lib\\site-packages\\tensorflow_init_.py in \n20\n21 # pylint: disable=g-bad-import-order\n---&gt; 22 from tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import\n23\n24 try:\n\nc:\\python\\lib\\site-packages\\tensorflow\\python_init_.py in \n47 import numpy as np\n48\n---&gt; 49 from tensorflow.python import pywrap_tensorflow\n50\n51 # Protocol buffers\n\nc:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in \n72 for some common reasons and solutions. Include the entire stack trace\n73 above this error message when asking for help.\"\"\" % traceback.format_exc()\n---&gt; 74 raise ImportError(msg)\n75\n76 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long\n\nImportError: Traceback (most recent call last):\nFile \"c:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"c:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n_pywrap_tensorflow_internal = swig_import_helper()\nFile \"c:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nFile \"c:\\python\\lib\\imp.py\", line 242, in load_module\nreturn load_dynamic(name, filename, file)\nFile \"c:\\python\\lib\\imp.py\", line 342, in load_dynamic\nreturn _load(spec)\nImportError: DLL load failed with error code -1073741795\n<\/code><\/pre>\n","answers":[{"AnswerId":"55935604","CreationDate":"2019-05-01T11:32:57.840","ParentId":null,"OwnerUserId":"10219869","Title":null,"Body":"<p>I have solved this problem by installing <code>tensorflow<\/code> 1.5.0 as I downgraded it from <code>tensorflow<\/code> 1.13.1, which was my previous version.<\/p>\n"},{"AnswerId":"53274495","CreationDate":"2018-11-13T05:38:42.227","ParentId":null,"OwnerUserId":"10389355","Title":null,"Body":"<p>please try to install the tensorflow version 1.5 or use <strong>pip install --upgrade tensorflow<\/strong>.<\/p>\n"}]}
+{"QuestionId":53057728,"AnswerCount":0,"Tags":"<json><python-3.x><callback><keras>","CreationDate":"2018-10-30T05:03:45.330","AcceptedAnswerId":null,"OwnerUserId":2089899.0,"Title":"Keras callback for intermediate dense layer output","Body":"<p>I'm trying to implement a callback that at the end of training saves the output of an intermediate dense layer to a json file. My attempt creates the file and doesn't error, but the file is empty. How should I go about this?<\/p>\n\n<pre><code>import json\njson_log = open('fingerprints.json', mode='wt', buffering=1)\nfingerprint_callback = tf.keras.callbacks.LambdaCallback(\n    on_train_end=lambda _: json_log.write(\n        json.dumps({'fingerprints': model.layers[-5].output})))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53058116,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-30T05:46:15.740","AcceptedAnswerId":null,"OwnerUserId":2089899.0,"Title":"Replicating TF model in Keras from pbtxt and ckpts","Body":"<p>I'm not sure if this is actually possible with the information I have, but please let me know if that's the case.<\/p>\n\n<p>From a previously trained Tensorflow model I have the following files:<\/p>\n\n<p><code>graph.pbtxt<\/code>, <code>checkpoint<\/code>, <code>model.ckpt-10000.data-00000-of-00001<\/code>, <code>model.ckpt-10000.index<\/code>, and <code>model.ckpt-10000.meta<\/code><\/p>\n\n<p>I was told that the input size of this model was a Dense layer of size <code>5000<\/code> and the output was a Dense sigmoid binary classification, but I don't know how many\/what size layers were in between. (I'm also not 100% positive that the input size is correct).<\/p>\n\n<p>From this information and associated files, is there a way to replicate the TF model with trained weights into a Keras functional model? <\/p>\n\n<p>(The idea was that this small dense network was added onto the last FC layer of VGG-16, so that'll be my end goal.)<\/p>\n","answers":[]}
+{"QuestionId":53058311,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-10-30T06:06:35.783","AcceptedAnswerId":null,"OwnerUserId":10056036.0,"Title":"Tensorflow multiple session.run( ) in same iteration","Body":"<p>Following are 2 code patterns for the training part of Tensorflow NN. <\/p>\n\n<p>I find it logical to use the model 1. But I am seeing model 2 in multiple places frequently. I feel Model 2 is wrong. Wont model run the graph in session twice for every iteration for the same data?  Is there something I am missing and people do it for any other reasons?<\/p>\n\n<p>Model 1<\/p>\n\n<pre><code>for epoch in range(epochs):\n    for iteration in range(num_tr_iter):\n        _, loss, accuracy = sess.run([optimizer, loss, accuracy], feed_dict)\n<\/code><\/pre>\n\n<p>Model 2<\/p>\n\n<pre><code>for epoch in range(epochs):\n    for iteration in range(num_tr_iter):\n        sess.run(optimizer, feed_dict)\n        loss, accuracy = sess.run([loss, accuracy],feed_dict)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT :<\/strong> Am expanding the question for more clarity<\/p>\n\n<p>If the below sess.run() execute optimizer node, it will execute all its dependent node. It will run the underlying convnet and loss function as well.<\/p>\n\n<pre><code>    sess.run(optimizer, feed_dict)\n<\/code><\/pre>\n\n<p>Next if the below sess.run() execute the loss node, why doesnt it execute the convnet with its current weights. I am not inferring that it will run the optimization again. Even to arrive at the current loss, wouldnt tensorflow execute the convnet and calculate the loss?<\/p>\n\n<pre><code>        loss, accuracy = sess.run([loss, accuracy],feed_dict)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53058459","CreationDate":"2018-10-30T06:21:43.277","ParentId":null,"OwnerUserId":"8163791","Title":null,"Body":"<p><code>Session.Run<\/code> will only run calculate the graph element(s) given to in the first argument. <\/p>\n\n<p>So in Model 2, <code>sess.run(optimizer, feed_dict)<\/code> just applies the weight updates for the model, and <code>loss, accuracy = sess.run([loss, accuracy],feed_dict)<\/code> just calculates the loss and accuracy on the model after weights have been updated. The whole computation isn't being run twice since these operations are independent of each other.<\/p>\n\n<p>Note that the Session.Run <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\">docs<\/a> mention that:<\/p>\n\n<blockquote>\n  <p>Order in which... operations are evaluated inside the call is undefined.<\/p>\n<\/blockquote>\n\n<p>So Model 1 might not even give an accurate loss and accuracy, since it's unclear if it was measured before or after the weight update. (In practice this might be fine, though)<\/p>\n"},{"AnswerId":"53058422","CreationDate":"2018-10-30T06:17:47.840","ParentId":null,"OwnerUserId":"10395701","Title":null,"Body":"<p>In model 2, the first line of code runs the optimizer on the model, and the second line calculates the loss and accuracy. It is not running the graph twice.<\/p>\n\n<p>Model 1 and Model 2 are essentially the same, but points go to model 1 for brevity.<\/p>\n"}]}
+{"QuestionId":53059814,"AnswerCount":1,"Tags":"<tensorflow><mandelbrot>","CreationDate":"2018-10-30T08:05:16.783","AcceptedAnswerId":null,"OwnerUserId":1306442.0,"Title":"tensorflow variable.assign_add function is mysterious in this example","Body":"<p>I'm trying to learn tensorflow from working examples online but came across the example where i'm literally wondered how it works. Can any explain the maths behind this particular function of tensorflow and how [ns] get its value out of boolean data type.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nY, X = np.mgrid[-2.3:2.3:0.005, -5:5:0.005]\nZ = X+1j*Y\n\nc = tf.constant(Z, np.complex64)#.astype(np.complex64))\nzs = tf.Variable(c)\nns = tf.Variable(tf.zeros_like(c, tf.float32))\n\nsess = tf.InteractiveSession()\n\ntf.global_variables_initializer().run()\n\nzs_ = zs*zs + c\n\nnot_diverged = tf.abs(zs_) &gt; 4\n\nstep = tf.group(zs.assign(zs_),\n ns.assign_add(tf.cast(not_diverged, tf.float32)))\n\nnx = tf.reduce_sum(ns)\nzx = tf.reduce_sum(zs_)\ncx = tf.reduce_sum(c)\nzf = tf.reduce_all(not_diverged)\n\nfor i in range(200): \n    step.run()\n    print(sess.run([nx,zx,cx,zf]))\n\nplt.imshow(ns.eval())\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53072144","CreationDate":"2018-10-30T20:07:12.400","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# this defines the complex plane\nY, X = np.mgrid[-2.3:2.3:0.005, -5:5:0.005]\nZ = X+1j*Y\nc = tf.constant(Z, np.complex64)\n\n# tensors are immutable in tensorflow,\n# but variabels arent, so use variable\n# to update values later on\nzs = tf.Variable(c)\n\n# ns will keep count of what has diverged\nns = tf.Variable(tf.zeros_like(c, tf.float32))\n\nsess = tf.InteractiveSession()\ntf.global_variables_initializer().run()\n\n# mandlebrot set M is defined as\n# c \\in M \\iff |P_c^n(0)| &lt;= 2 \\iff abs(P_c^n(0)) &lt;= 4\n# where P_c(z) = z^2 + c\n# the variable name is confusing, as it is actually\n# the opposite, I renamed it below\nzs_ = zs*zs + c\ndiverged = tf.abs(zs_) &gt; 4\n\n# ns gets its value as a bool casted to a float\n# is given by True \\mapsto 1., False \\mapsto 0.\n# the assign add just says, add tf.cast(diverged, tf.float32)\n# to the variabel ns, and assign that value to the variable\nstep = tf.group(\n    zs.assign(zs_),\n    ns.assign_add(tf.cast(diverged, tf.float32)))\n\n\n# here we iterate n to whatever we like\n# each time we are moving further along the\n# sequence P^n_c(0), which must be bounded\n# in a disk of radius 2 to be in M\nfor i in range(200):\n    step.run()\n\n# anywhere with value &gt; 0 in the plot is not in the Mandlebrot set\n# anywhere with value = 0 MIGHT be in the Mandlebrot set\n# we don't know for sure if it is in the set, \n# because we can only ever take n to be some\n# finite number. But to be in the Mandlebrot set, it has\n# to be bounded for all n!\nplt.imshow(ns.eval())\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53059896,"AnswerCount":1,"Tags":"<deep-learning><conv-neural-network><caffe><caffe2>","CreationDate":"2018-10-30T08:11:30.650","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Check failed: error == cudaSuccess (74 vs. 0) misaligned address","Body":"<p>I have two data set and training using caffe library for CNN.<\/p>\n\n<p>First data set has a lot of training data more 60,000 train images and 16,000 test images. Its solver file can be seen as follow. Batch size is set to 32 in training.<\/p>\n\n<pre><code>train_net: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/NumberPlate\/InceptionNet\/6\/train_0.prototxt\"\ntest_net: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/NumberPlate\/InceptionNet\/6\/test_0.prototxt\"\ntest_iter: 2080\ntest_interval: 4000\nbase_lr: 0.0010000000475\ndisplay: 10\nmax_iter: 16000\nlr_policy: \"multistep\"\ngamma: 0.10000000149\nmomentum: 0.899999976158\nweight_decay: 0.000500000023749\nsnapshot: 2000\nsnapshot_prefix: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/NumberPlate\/InceptionNet\/6\/InceptionNet\"\nsolver_mode: GPU\ndevice_id: 0\ndebug_info: false\nsnapshot_after_train: true\ntest_initialization: false\naverage_loss: 10\nstepvalue: 4000\nstepvalue: 8000\nstepvalue: 12000\niter_size: 1\nmomentum2: 0.999000012875\ntype: \"Adam\"\neval_type: \"detection\"\nap_version: \"11point\"\nnum_total_train_images: 62308\npathtolog: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/NumberPlate\/InceptionNet\/6\"\nbatchsize: 32\nmeanprecision: 0.5\nscratch: 1\n<\/code><\/pre>\n\n<p>I have second data set with fewer number of train images.\nOnly 2883 train images and 709 test images and batch size for training is set 16 as follow.<\/p>\n\n<pre><code>train_net: \"\/home \/Softwares\/Projects\/caffe-ssd-2\/Nextan\/InceptionNet\/0\/train_0.prototxt\"\ntest_net: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/Nextan\/InceptionNet\/0\/test_0.prototxt\"\ntest_iter: 177\ntest_interval: 500\nbase_lr: 0.0010000000475\ndisplay: 10\nmax_iter: 8000\nlr_policy: \"multistep\"\ngamma: 0.10000000149\nmomentum: 0.899999976158\nweight_decay: 0.000500000023749\nsnapshot: 1000\nsnapshot_prefix: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/Nextan\/InceptionNet\/0\/InceptionNet\"\nsolver_mode: GPU\ndevice_id: 0\ndebug_info: false\nsnapshot_after_train: true\ntest_initialization: false\naverage_loss: 10\nstepvalue: 2000\nstepvalue: 4000\nstepvalue: 6000\niter_size: 1\nmomentum2: 0.999000012875\ntype: \"Adam\"\neval_type: \"detection\"\nap_version: \"11point\"\nnum_total_train_images: 2883\npathtolog: \"\/home\/Softwares\/Projects\/caffe-ssd-2\/Nextan\/InceptionNet\/0\"\nbatchsize: 16\nmeanprecision: 0.5\nscratch: 1\n<\/code><\/pre>\n\n<p>I trained on the same PC with same GPU and resources.\nSecond data set gave me <code>\"Check failed: error == cudaSuccess (74 vs. 0) misaligned address\"<\/code>\nBut first dataset is successfully trained.\nWhat could be wrong?<\/p>\n","answers":[{"AnswerId":"56578697","CreationDate":"2019-06-13T10:41:27.207","ParentId":null,"OwnerUserId":"8472265","Title":null,"Body":"<p>It is as internal bug in Caffe due to the fact that in some situations max_workspace is not a multiple of 16 and that causes workspace to be unaligned in memory. The first thing I would try is to change the batch size.<\/p>\n\n<p>Here you can see a pull request with the issue:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/6548\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/pull\/6548<\/a><\/p>\n"}]}
+{"QuestionId":53060473,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-10-30T08:50:45.680","AcceptedAnswerId":null,"OwnerUserId":3215384.0,"Title":"Restore intermediate checkpoint files of Tensorflow","Body":"<p>Tensorflow version =1.8.0<\/p>\n\n<p>I am trying to restore my model using one of the intermediate checkpoint files in Tensorflow. By default Tensorflow will take the last saved checkpoint file. \nFor example, the folder contains files like:<\/p>\n\n<p>checkpoint\nmodel-56000.index  model-56000.data-00000-of-00001  model-56000.meta                 model-57000.index  model-57000.data-00000-of-00001  model-57000.meta<\/p>\n\n<p>By default, Tensorflow loads the last 57K checkpoint, but for reasons, I want to load the weights for the 56K checkpoint. \nFollowing is my code for restoring the model:<\/p>\n\n<pre><code>def load_G(self, checkpoint_dir):\n    print(\" [*] Reading checkpoints of G...\")\n    ckpt = tf.train.get_checkpoint_state(checkpoint_dir)\n    if ckpt and ckpt.model_checkpoint_path:\n        ckpt_name = os.path.basename(ckpt.model_checkpoint_path)\n        self.saver_gen.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))\n        return True\n    else:\n        return False\n<\/code><\/pre>\n\n<p>From Tensorflow's page, I read that for tf.train.get_checkpoint_state(), I can specify tf.train.get_checkpoint_state(checkpoint_dir, latest_filename=None). But I am not able to figure, what should I write for latest_filename. I tried writing latest_filename = model-56000\nBut that did not load the model.<\/p>\n\n<p>I also tried writing latest_filename = model-56000.meta. That also did not work.<\/p>\n\n<p>So, what is the correct way to load some intermediate checkpoint files in Tensorflow.<\/p>\n","answers":[{"AnswerId":"53061373","CreationDate":"2018-10-30T09:42:45.410","ParentId":null,"OwnerUserId":"10288309","Title":null,"Body":"<p>The ckpt file name will be <code>model-56000.ckpt<\/code> <\/p>\n\n<p><code>model-56000.meta<\/code> points to the meta information of the ckpt<\/p>\n\n<p><code>model-56000<\/code> is file name for either ckpt, data files, or meta files<\/p>\n"},{"AnswerId":"53268137","CreationDate":"2018-11-12T18:35:37.043","ParentId":null,"OwnerUserId":"3215384","Title":null,"Body":"<p>Ok,so a hack is to modify the <em>checkpoint<\/em> protobuf file and change the first line of that file from: <em>model_checkpoint_path: \"model-57000\"<\/em> to <em>model_checkpoint_path: \"model-56000\"<\/em> ad now it loads the 56K checkpoint.\nLooking for some better ways to do this.<\/p>\n"}]}
+{"QuestionId":53060501,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-10-30T08:52:14.870","AcceptedAnswerId":null,"OwnerUserId":3219201.0,"Title":"Exporting keras model into tflite","Body":"<p>I am trying to combine this two examples and create the tflite file for my  android app.<\/p>\n\n<p><a href=\"https:\/\/medium.com\/nybles\/create-your-first-image-recognition-classifier-using-cnn-keras-and-tensorflow-backend-6eaab98d14dd\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/nybles\/create-your-first-image-recognition-classifier-using-cnn-keras-and-tensorflow-backend-6eaab98d14dd<\/a><\/p>\n\n<p><a href=\"https:\/\/medium.com\/@xianbao.qian\/convert-keras-model-to-tflite-e2bdf28ee2d2\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@xianbao.qian\/convert-keras-model-to-tflite-e2bdf28ee2d2<\/a><\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code># Part 1 - Building the CNN\n\n# Importing the Keras libraries and packages\nfrom keras.models import Sequential\nfrom keras.layers import Convolution2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense\nimport tensorflow as tf\nfrom keras.models import load_model\n\n\n# Initialising the CNN\nclassifier = Sequential()\n\n# Step 1 - Convolution\nclassifier.add(Convolution2D(32, 3, 3, input_shape = (64, 64, 3), activation = 'relu'))\n\n# Step 2 - Pooling\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n\n# Adding a second convolutional layer\nclassifier.add(Convolution2D(32, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n\n# Step 3 - Flattening\nclassifier.add(Flatten())\n\n# Step 4 - Full connection\nclassifier.add(Dense(output_dim = 128, activation = 'relu'))\nclassifier.add(Dense(output_dim = 1, activation = 'sigmoid'))\n\n# Compiling the CNN\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\n# Part 2 - Fitting the CNN to the images\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   horizontal_flip = True)\n\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ntraining_set = train_datagen.flow_from_directory('dataset\/training_set',\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'binary')\n\ntest_set = test_datagen.flow_from_directory('dataset\/test_set',\n                                            target_size = (64, 64),\n                                            batch_size = 32,\n                                            class_mode = 'binary')\n\nclassifier.fit_generator(training_set,\n                         samples_per_epoch = 80,\n                         nb_epoch = 1,\n                         validation_data = test_set,\n                         nb_val_samples = 20)\n\n\n\noutput_names = [node.op.name for node in classifier.outputs]\nsess = tf.keras.backend.get_session()\nfrozen_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, output_names)    \n\n\ntflite_model = tf.contrib.lite.toco_convert(frozen_def, [inputs], output_names)\nwith tf.gfile.GFile(tflite_graph, 'wb') as f:\n    f.write(tflite_model)                     \n<\/code><\/pre>\n\n<p>And at this line: <\/p>\n\n<pre><code>frozen_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, output_names)\n<\/code><\/pre>\n\n<p>I got an exception:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value conv2d_1\/bias\n[[Node: _retval_conv2d_1\/bias_0_0 = _Retval[T=DT_FLOAT, index=0, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](conv2d_1\/bias)]]\n<\/code><\/pre>\n\n<p>I am a beginner in the machine learning and absolutely have no idea what is this error about :-(<\/p>\n\n<p>Can somebody explain to me what is wrong? \nAll that I need is processed several folders with many pics and make it possible to predict the relations of the new coming pictures to the certain folder.\nThank you.<\/p>\n","answers":[{"AnswerId":"55416848","CreationDate":"2019-03-29T11:50:55.760","ParentId":null,"OwnerUserId":"3033937","Title":null,"Body":"<p>It is possible to directly convert a keras-model to <code>.tflite<\/code> using the <code>tf.lite.TFLiteConverter.from_session<\/code> function. Place the following code after <code>fit_generator<\/code> to export it (tested with tensorflow 1.3.1)<\/p>\n\n<pre><code>with tf.keras.backend.get_session() as sess:\n    sess.run(tf.global_variables_initializer())    \n    converter = tf.lite.TFLiteConverter.from_session(sess, model.inputs, model.outputs)\n    tflite_model = converter.convert()\n    with open(\"model.tflite\", \"wb\") as f:\n        f.write(tflite_model)   \n<\/code><\/pre>\n"}]}
+{"QuestionId":53061117,"AnswerCount":0,"Tags":"<tensorflow><classification><conv-neural-network><multilabel-classification>","CreationDate":"2018-10-30T09:29:37.563","AcceptedAnswerId":null,"OwnerUserId":9276708.0,"Title":"Why the classification results are concentrated in one category in CNN model?","Body":"<p>I want to use <code>CNN<\/code> model in classification and the <code>num_classes<\/code> is <code>22<\/code>. I use the <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/convolutional_network_raw.py\" rel=\"nofollow noreferrer\">example code<\/a> of TensorFlow directly.<\/p>\n\n<p>The number of training samples is <code>16522<\/code>. All the pictures has a same shape <code>(50,52)<\/code>. <\/p>\n\n<pre><code># Create model\ndef conv_net(x, weights, biases, dropout):\n\n    x = tf.reshape(x, shape=[-1, 28, 28, 1])\n\n    # Convolution Layer\n    conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n    conv1 = maxpool2d(conv1, k=2)\n\n    # Convolution Layer\n    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n    # Max Pooling (down-sampling)\n    conv2 = maxpool2d(conv2, k=2)\n\n    # Fully connected layer\n    # Reshape conv2 output to fit fully connected layer input\n    fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])\n    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n    fc1 = tf.nn.relu(fc1)\n    # Apply Dropout\n    fc1 = tf.nn.dropout(fc1, dropout)\n\n    # Output, class prediction\n    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n    return out\n\n# Store layers weight &amp; bias\nweights = {\n    # 5x5 conv, 1 input, 32 outputs\n    'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),\n    # 5x5 conv, 32 inputs, 64 outputs\n    'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),\n    # fully connected, 7*7*64 inputs, 1024 outputs\n    'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),\n    # 1024 inputs, 22 outputs (class prediction)\n    'out': tf.Variable(tf.random_normal([1024, num_classes]))\n}\n\nbiases = {\n    'bc1': tf.Variable(tf.random_normal([32])),\n    'bc2': tf.Variable(tf.random_normal([64])),\n    'bd1': tf.Variable(tf.random_normal([1024])),\n    'out': tf.Variable(tf.random_normal([num_classes]))\n}\n<\/code><\/pre>\n\n<p>The following is my results.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/74IrS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/74IrS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2h9PY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2h9PY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As you can see, my predicted result (the first picture) is bad, and many categories are incorrectly predicted as <code>Fault 4<\/code>.<\/p>\n\n<p>I use <code>softmax_cross_entropy_with_logits<\/code> as my loss function, the losses have been falling, but the accuracy is also decreasing. In my opinion, the loss function is useless in this situation, which leads to overfitting. <\/p>\n\n<p>Any suggestions for this situation? <\/p>\n","answers":[]}
+{"QuestionId":53061284,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-30T09:38:15.503","AcceptedAnswerId":null,"OwnerUserId":635523.0,"Title":"Tensorflow: Assert that values are in set","Body":"<p>I have an one-dimensional <code>tf.uint8<\/code> tensor <code>x<\/code> and want to assert that all values inside that tensor are in set <code>s<\/code> I define. <code>s<\/code> is fixed at graph definition time, so it's not a dynamically computed tensor.<\/p>\n\n<p>In plain Python, I want to do sth. like the following:<\/p>\n\n<pre><code>x = [1, 2, 3, 1, 11, 3, 5]\ns = {1, 2, 3, 11, 12, 13}\nassert all(el in s for el in x), \"This should fail, as 5 is not in s\"\n<\/code><\/pre>\n\n<p>I know that I can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Assert\" rel=\"nofollow noreferrer\"><code>tf.Assert<\/code><\/a> for the assertion part but I'm struggling with defining the condition part (<code>el in s<\/code>). What's the simplest\/most canonical way to do this?<\/p>\n\n<p>The 2.5 year-old answer in <a href=\"https:\/\/stackoverflow.com\/questions\/34618207\/determining-if-a-value-is-in-a-set-in-tensorflow\">Determining if A Value is in a Set in TensorFlow<\/a> is not sufficient to me: First of all it's complex to write down and understand, and second it's using a broadcasted <code>tf.equal<\/code>, which is more expensive computation wise than a proper set-based check.<\/p>\n","answers":[{"AnswerId":"53062467","CreationDate":"2018-10-30T10:39:09.377","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>A simple way could be something like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = [1, 2, 3, 1, 11, 3, 5]\ns = {1, 2, 3, 11, 12, 13}\n\nx_t = tf.constant(x, dtype=tf.uint8)\ns_t = tf.constant(list(s), dtype=tf.uint8)\n# Check every value in x against every value in s\nxs_eq = tf.equal(x_t[:, tf.newaxis], s_t)\n# Check every element in x is equal to at least one element in s\nassert_op = tf.Assert(tf.reduce_all(tf.reduce_any(xs_eq, axis=1)), [x_t])\nwith tf.control_dependencies([assert_op]):\n    # Use x_t...\n<\/code><\/pre>\n\n<p>This creates an intermediate tensor with size <code>(len(x), len(s))<\/code>. If that is problematic, you could also split the problem into independent tensors, for example like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = [1, 2, 3, 1, 11, 3, 5]\ns = {1, 2, 3, 11, 12, 13}\n\nx_t = tf.constant(x, dtype=tf.uint8)\n# Count where each x matches each s\nx_in_s = [tf.cast(tf.equal(x_t, si), tf.int32) for si in s]\n# Add matches and check there is at least one match per x\nassert_op = tf.Assert(tf.reduce_all(tf.add_n(x_in_s) &gt; 0), [x_t])\n<\/code><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>Actually, since you said your values are <code>tf.uint8<\/code>, you can make things even better with boolean arrays:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = [1, 2, 3, 1, 11, 3, 5]\ns = {1, 2, 3, 11, 12, 13}\n\nx_t = tf.constant(x, dtype=tf.uint8)\ns_t = tf.constant(list(s), dtype=tf.uint8)\n# One-hot vectors of values included in x and s\nx_bool = tf.scatter_nd(tf.cast(x_t[:, tf.newaxis], tf.int32),\n                       tf.ones_like(x_t, dtype=tf.bool), [256])\ns_bool = tf.scatter_nd(tf.cast(s_t[:, tf.newaxis], tf.int32),\n                       tf.ones_like(s_t, dtype=tf.bool), [256])\n# Check that all values in x are in s\nassert_op = tf.Assert(tf.reduce_all(tf.equal(x_bool, x_bool &amp; s_bool)), [x_t])\n<\/code><\/pre>\n\n<p>This takes linear time and constant memory.<\/p>\n\n<p>EDIT 2: While the last method is theoretically the best in this case, doing a couple of quick benchmarks I can only see a significant difference in performance  when I go up to hundreds of thousands of elements, and in any case the three are still quite fast with <code>tf.uint8<\/code>.<\/p>\n"}]}
+{"QuestionId":53061418,"AnswerCount":1,"Tags":"<variables><pytorch><multi-gpu>","CreationDate":"2018-10-30T09:44:54.783","AcceptedAnswerId":null,"OwnerUserId":4394807.0,"Title":"Pytorch incorrect value of member variable when using Multi-gpu","Body":"<p>Here is a simple class for running in multi-gpu environment. The member variable <code>self.firstIter<\/code> should be <code>False<\/code> after the first iteration.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Class TestNetwork(nn.Module):\n\n    def __init__(self):\n        super(TestNetwork, self).__init__()\n        self.firstIter = True #indicates whether it's the first iteration\n\n    def forward(self, input):\n        print 'is firstIter: ', self.firstIter #always True!!\n        if self.firstIter is True:\n            self.firstIter = False\n        # do otherthings\n<\/code><\/pre>\n\n<p>The code works as expected when using only one gpu. <\/p>\n\n<p>However when using multi-gpu (i.e. <code>nn.DataParallel<\/code>), the value of <code>self.firstIter<\/code> is always printed as <code>True<\/code>. <\/p>\n\n<p><strong>Why does this happen? What is wrong with the code?<\/strong> <\/p>\n\n<p><em>Using PyTorch version 0.3.1.<\/em><\/p>\n","answers":[{"AnswerId":"55324705","CreationDate":"2019-03-24T14:14:54.800","ParentId":null,"OwnerUserId":"11250789","Title":null,"Body":"<p>Basically,  DataParallel operates on model replicas, and changes made to replicas (during forward) are not visible outside forward\/backward calls if number of divices is large than 1.<\/p>\n\n<p>Plz refer to <a href=\"https:\/\/discuss.pytorch.org\/t\/nonetype-attribute-when-using-dataparallel\/11566\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/nonetype-attribute-when-using-dataparallel\/11566<\/a> for details.<\/p>\n"}]}
+{"QuestionId":53062798,"AnswerCount":1,"Tags":"<python><windows><tensorflow><matplotlib>","CreationDate":"2018-10-30T10:57:26.890","AcceptedAnswerId":null,"OwnerUserId":258483.0,"Title":"Tensorflow importing crashes Python without any error on Windows","Body":"<p>Yesterday I was trying to solve matplotlib problems on Windows, which included solving <code>from matplotlib import ft2font<\/code> error. I have added <code>HDF5_DISABLE_VERSION_CHECK=2<\/code> for that. But today I found, that tensorflow doesn't work anymore:<\/p>\n\n<pre><code>&gt; python                                                                                              \nPython 3.5.5 |Anaconda, Inc.| (default, Apr  7 2018, 04:52:34) [MSC v.1900 64 bit (AMD64)] on win32   \nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.                                \n&gt;&gt;&gt; import tensorflow as tf                                                                           \n\n&gt;\n<\/code><\/pre>\n\n<p>It thinks 30 seconds, then exits Python without any error.<\/p>\n\n<p>Where can I see logs or crashdumps or something?                                                                                                     <\/p>\n","answers":[{"AnswerId":"53683691","CreationDate":"2018-12-08T14:52:44.757","ParentId":null,"OwnerUserId":"9275698","Title":null,"Body":"<p>If you try <code>import h5py<\/code>, and get <strong>ImportError: DLL load failed<\/strong>, you probably have some problem as mine. Try:<\/p>\n\n<blockquote>\n  <ol>\n  <li><p>Uninstall <code>h5py<\/code> via anaconda navigator, some relative libs (<code>tensorflow<\/code>) will be uninstalled as well. <\/p><\/li>\n  <li><p>Go VM path in anaconda installation, <code>D:\\Anaconda\\envs\\YOUR_VM\\Lib\\site-packages<\/code>, delete <code>h5py<\/code> and <code>h5py-2.8.0-py3.6.egg-info<\/code> folders<\/p><\/li>\n  <li><p>Reinstall <code>tensorflow<\/code> or <code>tensorflow-gpu<\/code><\/p><\/li>\n  <\/ol>\n<\/blockquote>\n"}]}
+{"QuestionId":53063116,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-30T11:16:01.557","AcceptedAnswerId":null,"OwnerUserId":10083164.0,"Title":"Neural Networks works worse than RandomForest","Body":"<p>I have a classification problem that target contains 5 classes, 15 features(all continuous)\nand have 1 million for training data, 0.5 million for validation data.\ne.g., <\/p>\n\n<pre><code>shape of X_train = (1000000,15)\nshape of X_validation = (500000,15)\n<\/code><\/pre>\n\n<p>First, I used Random Forest that can get 88% Avg. Accuracy.<\/p>\n\n<p>After that I tried many Neural Network architecture, the best one got ~80% Avg. Accuracy both on training and validation data, which was worse than Random forest.\n(I don't know much about designing Neural Network architecture)<\/p>\n\n<p>Following is the best one of my NN architecture. (~80% Avg.Accuracy)<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1000, input_dim=15, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(900, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(800, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(700, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(600, activation='relu'))\nmodel.add(Dense(5, activation='softmax'))#output layer\nadadelta = Adadelta()\nmodel.compile(loss='categorical_crossentropy', optimizer=adadelta, metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Batch Size = 128 and epochs = 100<\/p>\n\n<p>I have read this <a href=\"https:\/\/stackoverflow.com\/questions\/40459022\/neural-network-in-tensorflow-works-worse-than-random-forest-and-predict-the-same\">question<\/a>. The answer point out that NN needs amount of data and some regulization. I think my data size is good enough and I have also tried higer Dropout rate and L2 regulization but still not working. \nWhat could the problem be?<\/p>\n\n<p>This is biological data that I have no domain knowledge so sorry about that I can't explain it. I've plot the feature distribution as below, all features are between 0 to 3\n<a href=\"https:\/\/i.stack.imgur.com\/fq282.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fq282.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53063170,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-10-30T11:19:59.493","AcceptedAnswerId":null,"OwnerUserId":4829349.0,"Title":"Different results in training and evaluation of the same batch in Keras","Body":"<p>How it is possible to get into the following situation? <\/p>\n\n<pre><code>In[20] model.test_on_batch(x, y)\nOut[20]: [4.861001, 0.0]\nIn[21]: model.train_on_batch(x, y)\nOut[21]: [0.019215763, 0.99609375]\n<\/code><\/pre>\n\n<p>The two output values correspond to the binary crossentropy and the binary accuracy respectively.<\/p>\n\n<p>Namely, the training and testing results of a model on the <em>same batch<\/em> are radically different. The training results are presumably the good ones. The predictions (with <code>model.predict<\/code>) are not correct either.<\/p>\n\n<p>Running Keras 2.2.4 under Tensorflow 1.8.0.<\/p>\n","answers":[{"AnswerId":"53064546","CreationDate":"2018-10-30T12:42:04.477","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>This is completely normal if you use a model that has Dropout or BatchNormalization layers, as these have different behavior during training and testing times. If you add any regularization terms then these will be disabled during testing and it will affect any loss you get from <code>model.evaluate<\/code><\/p>\n"}]}
+{"QuestionId":53063339,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-10-30T11:29:46.380","AcceptedAnswerId":53063373.0,"OwnerUserId":10310255.0,"Title":"AttributeError: 'list' object has no attribute 'set_model'","Body":"<p>I'm trying to use early stopping and Model check points to save the best model while training a deep convolution neural network. However, I get the following error:<\/p>\n\n<pre><code> callback.set_model(model)\n AttributeError: 'list' object has no attribute 'set_model'\n<\/code><\/pre>\n\n<p>My code so far is:<\/p>\n\n<pre><code>model = Sequential()\n\n###First block\nmodel.add(Conv2D(100,kernel_size = (3,3),activation = 'relu',padding = 'same',input_shape=(12,11,1)))\nmodel.add(Conv2D(100,kernel_size = (3,3),activation = 'relu',padding = 'same'))\nmodel.add(MaxPooling2D(pool_size = (2,2)))\nmodel.add(Dropout(0.20))\n\n\n###Second block\n\nmodel.add(Conv2D(128,kernel_size = (3,3),activation = 'relu',padding = 'same'))\nmodel.add(Conv2D(128,kernel_size = (3,3),activation = 'relu',padding = 'same'))\nmodel.add(MaxPooling2D(pool_size = (2,2)))\nmodel.add(Dropout(0.10))\n\n\nmodel.add(Flatten())\n\n#model.add(Dense(100,activation = 'relu',kernel_regularizer=regularizers.l2(0.01)))\nmodel.add(Dense(1000,activation = 'relu',kernel_regularizer=regularizers.l2(0.01)))\nmodel.add(Dropout(0.30))\nmodel.add(Dense(500,activation = 'relu',kernel_regularizer=regularizers.l2(0.01)))\nmodel.add(Dropout(0.10))\n#model.add(Dense(500,activation = 'relu',kernel_regularizer=regularizers.l2(0.01)))\n#model.add(Dropout(0.15))\n\nmodel.add(Dense(5,activation = 'softmax')) \n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n\n\nearlystop = [EarlyStopping(monitor='val_acc', min_delta=0.001, patience=5,\n                          verbose=1, mode='auto')]\n\n\noutputModel = 'outputModel'\n\nmodel_json = model.to_json()\nwith open(outputModel+\".json\", \"w\") as json_file:\n    json_file.write(model_json)\nmodWeightsFilepath=outputModel+\"_weights.hdf5\"\ncheckpoint = ModelCheckpoint(modWeightsFilepath, monitor='val_acc', verbose=1, save_best_only=True, save_weights_only=True, mode='auto')\n\ncallbacks_list = [earlystop,checkpoint]\n\n\nhistory = model.fit(x_train, Y,\n      batch_size=100, ##number of observations per batch\n      epochs=100, ##Number of epochs\n      callbacks = callbacks_list,\n      verbose=1,\n      shuffle = True,\n      validation_split=0.2) ###Data for evaluation\n<\/code><\/pre>\n\n<p>I don't know what I'm doing wrong. I read that ModelCheckPoint and earlystopping should be given as a list which is why I explicitly make it as:<\/p>\n\n<pre><code>  callbacks_list = [earlystop,checkpoint]\n<\/code><\/pre>\n\n<p>Help will be appreciated.<\/p>\n","answers":[{"AnswerId":"53063373","CreationDate":"2018-10-30T11:31:12.803","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You are correct about callbacks but <code>earlystop<\/code> is already a list here. Remove the brackets around <code>[EarlyStopping(..)]<\/code> to fix the problem.<\/p>\n"}]}
+{"QuestionId":53063761,"AnswerCount":0,"Tags":"<amazon-web-services><tensorflow>","CreationDate":"2018-10-30T11:54:47.023","AcceptedAnswerId":null,"OwnerUserId":1125992.0,"Title":"How to setup MirroredStrategy in Tensorflow on AWS EC2 for Objectdetection retraining?","Body":"<p>I built Tensorflow 1.11 from source. I am able tp retrain Mobilenet SSD v2 on a single GPU. I am trying Mirrored Strategy on 4 GPU AWS g2.8xlarge EC2 instace.<\/p>\n\n<p>I added MirroredStrategy in RunConfig in the file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/model_main.py\" rel=\"nofollow noreferrer\">model_main.py<\/a>, using the following code:<\/p>\n\n<pre><code>def main(unused_argv):\n  flags.mark_flag_as_required('model_dir')\n  flags.mark_flag_as_required('pipeline_config_path')\n\n  distribution = tf.contrib.distribute.MirroredStrategy()\n  config = tf.estimator.RunConfig(model_dir=FLAGS.model_dir,train_distribute=distribution)\n<\/code><\/pre>\n\n<p>I am running into an AssertionError. Below is the complete log. Can anyone please help? Thanks.<\/p>\n\n<pre><code>ubuntu@ip-172-31-30-151:~\/tensorflow\/models\/research$ python object_detection\/model_main.py \\\n&gt;     --pipeline_config_path=${PIPELINE_CONFIG_PATH} \\\n&gt;     --model_dir=${MODEL_DIR} \\\n&gt;     --num_train_steps=${NUM_TRAIN_STEPS} \\\n&gt;     --sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \\\n&gt;     --alsologtostderr\n\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/h5py\/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n  from ._conv import register_converters as _register_converters\n\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/utils\/visualization_utils.py:27: UserWarning: \nThis call to matplotlib.use() has no effect because the backend has already\nbeen chosen; matplotlib.use() must be called *before* pylab, matplotlib.pyplot,\nor matplotlib.backends is imported for the first time.\n\nThe backend was *originally* set to 'Qt5Agg' by the following code:\n  File \"object_detection\/model_main.py\", line 26, in &lt;module&gt;\n    from object_detection import model_lib\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/model_lib.py\", line 27, in &lt;module&gt;\n    from object_detection import eval_util\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/eval_util.py\", line 27, in &lt;module&gt;\n    from object_detection.metrics import coco_evaluation\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/metrics\/coco_evaluation.py\", line 20, in &lt;module&gt;\n    from object_detection.metrics import coco_tools\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/metrics\/coco_tools.py\", line 47, in &lt;module&gt;\n    from pycocotools import coco\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/pycocotools\/coco.py\", line 49, in &lt;module&gt;\n    import matplotlib.pyplot as plt\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/matplotlib\/pyplot.py\", line 71, in &lt;module&gt;\n    from matplotlib.backends import pylab_setup\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/matplotlib\/backends\/__init__.py\", line 16, in &lt;module&gt;\n    line for line in traceback.format_stack()\n\n\n  import matplotlib; matplotlib.use('Agg')  # pylint: disable=multiple-statements\n2018-10-24 22:06:33.449647: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-24 22:06:33.450249: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1411] Found device 0 with properties: \nname: GRID K520 major: 3 minor: 0 memoryClockRate(GHz): 0.797\npciBusID: 0000:00:03.0\ntotalMemory: 3.94GiB freeMemory: 3.90GiB\n2018-10-24 22:06:33.479913: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-24 22:06:33.480481: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1411] Found device 1 with properties: \nname: GRID K520 major: 3 minor: 0 memoryClockRate(GHz): 0.797\npciBusID: 0000:00:04.0\ntotalMemory: 3.94GiB freeMemory: 3.90GiB\n2018-10-24 22:06:33.512979: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-24 22:06:33.513529: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1411] Found device 2 with properties: \nname: GRID K520 major: 3 minor: 0 memoryClockRate(GHz): 0.797\npciBusID: 0000:00:05.0\ntotalMemory: 3.94GiB freeMemory: 3.90GiB\n2018-10-24 22:06:33.546408: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-10-24 22:06:33.546944: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1411] Found device 3 with properties: \nname: GRID K520 major: 3 minor: 0 memoryClockRate(GHz): 0.797\npciBusID: 0000:00:06.0\ntotalMemory: 3.94GiB freeMemory: 3.90GiB\n2018-10-24 22:06:33.547089: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1490] Adding visible gpu devices: 0, 1, 2, 3\n2018-10-24 22:06:34.907497: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-24 22:06:34.907580: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:977]      0 1 2 3 \n2018-10-24 22:06:34.907607: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 0:   N N N N \n2018-10-24 22:06:34.907623: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 1:   N N N N \n2018-10-24 22:06:34.907636: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 2:   N N N N \n2018-10-24 22:06:34.907650: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 3:   N N N N \n2018-10-24 22:06:34.908235: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 3644 MB memory) -&gt; physical GPU (device: 0, name: GRID K520, pci bus id: 0000:00:03.0, compute capability: 3.0)\n2018-10-24 22:06:34.946961: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:1 with 3644 MB memory) -&gt; physical GPU (device: 1, name: GRID K520, pci bus id: 0000:00:04.0, compute capability: 3.0)\n2018-10-24 22:06:34.981965: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:2 with 3644 MB memory) -&gt; physical GPU (device: 2, name: GRID K520, pci bus id: 0000:00:05.0, compute capability: 3.0)\n2018-10-24 22:06:35.020808: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:3 with 3644 MB memory) -&gt; physical GPU (device: 3, name: GRID K520, pci bus id: 0000:00:06.0, compute capability: 3.0)\nWARNING:tensorflow:Forced number of epochs for all eval validations to be 1.\nW1024 22:06:35.072655 139904432719616 tf_logging.py:125] Forced number of epochs for all eval validations to be 1.\nWARNING:tensorflow:Expected number of evaluation epochs is 1, but instead encountered `eval_on_train_input_config.num_epochs` = 0. Overwriting `num_epochs` to 1.\nW1024 22:06:35.073021 139904432719616 tf_logging.py:125] Expected number of evaluation epochs is 1, but instead encountered `eval_on_train_input_config.num_epochs` = 0. Overwriting `num_epochs` to 1.\nWARNING:tensorflow:Estimator's model_fn (&lt;function create_model_fn.&lt;locals&gt;.model_fn at 0x7f3d960158c8&gt;) includes params argument, but params are not passed to Estimator.\nW1024 22:06:35.073723 139904432719616 tf_logging.py:125] Estimator's model_fn (&lt;function create_model_fn.&lt;locals&gt;.model_fn at 0x7f3d960158c8&gt;) includes params argument, but params are not passed to Estimator.\n2018-10-24 22:06:35.077395: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1490] Adding visible gpu devices: 0, 1, 2, 3\n2018-10-24 22:06:35.077580: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-10-24 22:06:35.077606: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:977]      0 1 2 3 \n2018-10-24 22:06:35.077626: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 0:   N N N N \n2018-10-24 22:06:35.077649: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 1:   N N N N \n2018-10-24 22:06:35.077661: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 2:   N N N N \n2018-10-24 22:06:35.077676: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 3:   N N N N \n2018-10-24 22:06:35.077982: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/device:GPU:0 with 3644 MB memory) -&gt; physical GPU (device: 0, name: GRID K520, pci bus id: 0000:00:03.0, compute capability: 3.0)\n2018-10-24 22:06:35.078267: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/device:GPU:1 with 3644 MB memory) -&gt; physical GPU (device: 1, name: GRID K520, pci bus id: 0000:00:04.0, compute capability: 3.0)\n2018-10-24 22:06:35.078644: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/device:GPU:2 with 3644 MB memory) -&gt; physical GPU (device: 2, name: GRID K520, pci bus id: 0000:00:05.0, compute capability: 3.0)\n2018-10-24 22:06:35.079453: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1103] Created TensorFlow device (\/device:GPU:3 with 3644 MB memory) -&gt; physical GPU (device: 3, name: GRID K520, pci bus id: 0000:00:06.0, compute capability: 3.0)\nWARNING:tensorflow:num_readers has been reduced to 2 to match input file shards.\nW1024 22:06:35.100583 139904432719616 tf_logging.py:125] num_readers has been reduced to 2 to match input file shards.\nWARNING:tensorflow:From \/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/core\/preprocessor.py:1207: calling squeeze (from tensorflow.python.ops.array_ops) with squeeze_dims is deprecated and will be removed in a future version.\nInstructions for updating:\nUse the `axis` argument instead\nW1024 22:06:35.548344 139904432719616 tf_logging.py:125] From \/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/core\/preprocessor.py:1207: calling squeeze (from tensorflow.python.ops.array_ops) with squeeze_dims is deprecated and will be removed in a future version.\nInstructions for updating:\nUse the `axis` argument instead\nWARNING:tensorflow:From \/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/builders\/dataset_builder.py:148: batch_and_drop_remainder (from tensorflow.contrib.data.python.ops.batching) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.batch(..., drop_remainder=True)`.\nW1024 22:06:37.296217 139904432719616 tf_logging.py:125] From \/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/builders\/dataset_builder.py:148: batch_and_drop_remainder (from tensorflow.contrib.data.python.ops.batching) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.batch(..., drop_remainder=True)`.\nTraceback (most recent call last):\n  File \"object_detection\/model_main.py\", line 111, in &lt;module&gt;\n    tf.app.run()\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"object_detection\/model_main.py\", line 107, in main\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/training.py\", line 471, in train_and_evaluate\n    return executor.run()\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/training.py\", line 610, in run\n    return self.run_local()\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/training.py\", line 711, in run_local\n    saving_listeners=saving_listeners)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 356, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1179, in _train_model\n    return self._train_model_distributed(input_fn, hooks, saving_listeners)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1290, in _train_model_distributed\n    self.config)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/distribute.py\", line 718, in call_for_each_tower\n    return self._call_for_each_tower(fn, *args, **kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/python\/mirrored_strategy.py\", line 552, in _call_for_each_tower\n    return _call_for_each_tower(self, fn, *args, **kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/python\/mirrored_strategy.py\", line 183, in _call_for_each_tower\n    coord.join(threads)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/coordinator.py\", line 389, in join\n    six.reraise(*self._exc_info_to_raise)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/six.py\", line 693, in reraise\n    raise value\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/coordinator.py\", line 297, in stop_on_exception\n    yield\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/python\/mirrored_strategy.py\", line 783, in run\n    self.main_result = self.main_fn(*self.main_args, **self.main_kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/estimator.py\", line 1169, in _call_model_fn\n    model_fn_results = self._model_fn(features=features, **kwargs)\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/model_lib.py\", line 268, in model_fn\n    features[fields.InputDataFields.true_image_shape])\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/meta_architectures\/ssd_meta_arch.py\", line 501, in predict\n    preprocessed_inputs)\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/models\/ssd_mobilenet_v2_feature_extractor.py\", line 134, in extract_features\n    image_features=image_features)\n  File \"\/home\/ubuntu\/tensorflow\/models\/research\/object_detection\/models\/feature_map_generators.py\", line 379, in multi_resolution_feature_maps\n    scope=layer_name + '_depthwise')\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 182, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 2809, in separable_convolution2d\n    collections=weights_collections)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 182, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 350, in model_variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 182, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 277, in variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1484, in get_variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1234, in get_variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 521, in get_variable\n    return custom_getter(**custom_getter_kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1922, in wrapped_custom_getter\n    *args, **kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1749, in layer_variable_getter\n    return _model_variable_getter(getter, *args, **kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1740, in _model_variable_getter\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 182, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 350, in model_variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 182, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 277, in variable\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/distribute.py\", line 461, in disable_partitioned_variables\n    return getter(*args, **kwargs)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 492, in _true_getter\n    aggregation=aggregation)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 938, in _get_single_variable\n    with ops.colocate_with(v):\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/contextlib.py\", line 81, in __enter__\n    return next(self.gen)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 4092, in _colocate_with_for_gradient\n    with self.colocate_with(op, ignore_existing):\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/contextlib.py\", line 81, in __enter__\n    return next(self.gen)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 4144, in colocate_with\n    op = internal_convert_to_tensor_or_indexed_slices(op, as_ref=True).op\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1305, in internal_convert_to_tensor_or_indexed_slices\n    value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1144, in internal_convert_to_tensor\n    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"\/home\/ubuntu\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distribute\/python\/values.py\", line 447, in _tensor_conversion_mirrored\n    assert not as_ref\nAssertionError\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53063850,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-30T11:59:35.970","AcceptedAnswerId":null,"OwnerUserId":7787721.0,"Title":"Keras model.fit_generator raises Stopiteration error","Body":"<p>I am trying to run a port of the SSD on a Win10 system from the ssd300_training.ipynb code in  this github repo <a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pierluigiferrari\/ssd_keras<\/a> .\nI am getting a \"StopIteration\" error when training the model and the solutions out there, with my little research, suggest adding a while True block in the data generator function. However, I am not as proficient enough to do that with the code as it wasn't written like a function format.\nHere's the data generator code:<\/p>\n\n<pre><code># 6: Create the generator handles that will be passed to Keras' `fit_generator()` function.\ntrain_generator = train_dataset.generate(batch_size=batch_size,\n                                     shuffle=True,\n                                     transformations=[ssd_data_augmentation],\n                                     label_encoder=ssd_input_encoder,\n                                     returns={'processed_images',\n                                              'encoded_labels'},\n                                     keep_images_without_gt=False)\n\nval_generator = val_dataset.generate(batch_size=batch_size,\n                                 shuffle=False,\n                                 transformations=[convert_to_3_channels,\n                                                  resize],\n                                 label_encoder=ssd_input_encoder,\n                                 returns={'processed_images',\n                                          'encoded_labels'},\n                                 keep_images_without_gt=False)  \n<\/code><\/pre>\n\n<p>Please how can I add a while True: loop to this so it can stop giving the StopIteration error. Here's the training code and the error it gives:<\/p>\n\n<pre><code>initial_epoch   = 0\nfinal_epoch     = 50\nsteps_per_epoch = 10\n\n\n\nhistory = model.fit_generator(generator=train_generator,\n                          steps_per_epoch=steps_per_epoch,\n                          epochs=final_epoch,\n                          callbacks=callbacks,\n                          validation_data=val_generator,\n                          validation_steps=ceil(val_dataset_size\/batch_size),\n                          initial_epoch=initial_epoch\n                         )\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>    Epoch 1\/50\n 9\/10 [==========================&gt;...] - ETA: 24s - loss: 3.45 - ETA: 13s - loss: 3.45 - ETA: 9s - loss: 3.4506 - ETA: 6s - loss: 3.450 - ETA: 5s - loss: 3.450 - ETA: 3s - loss: 3.450 - ETA: 2s - loss: 3.450 - ETA: 1s - loss: 3.450 - ETA: 0s - loss: 3.4504\n---------------------------------------------------------------------------\nStopIteration                             Traceback (most recent call last)\n&lt;ipython-input-16-484d98ebb6b5&gt; in &lt;module&gt;()\n     11                               #callbacks=callbacks,\n     12                               validation_data=val_generator,\n---&gt; 13                               validation_steps=ceil(val_dataset_size\/batch_size),\n     14                               #initial_epoch=initial_epoch\n     15                              )\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\legacy\\interfaces.py in wrapper(*args, **kwargs)\n     89                 warnings.warn('Update your `' + object_name +\n     90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)\n---&gt; 91             return func(*args, **kwargs)\n     92         wrapper._original_function = func\n     93         return wrapper\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\engine\\training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n   1413             use_multiprocessing=use_multiprocessing,\n   1414             shuffle=shuffle,\n-&gt; 1415             initial_epoch=initial_epoch)\n   1416 \n   1417     @interfaces.legacy_generator_methods_support\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\engine\\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n    228                             val_enqueuer_gen,\n    229                             validation_steps,\n--&gt; 230                             workers=0)\n    231                     else:\n    232                         # No need for try\/except because\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\legacy\\interfaces.py in wrapper(*args, **kwargs)\n     89                 warnings.warn('Update your `' + object_name +\n     90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)\n---&gt; 91             return func(*args, **kwargs)\n     92         wrapper._original_function = func\n     93         return wrapper\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\engine\\training.py in evaluate_generator(self, generator, steps, max_queue_size, workers, use_multiprocessing, verbose)\n   1467             workers=workers,\n   1468             use_multiprocessing=use_multiprocessing,\n-&gt; 1469             verbose=verbose)\n   1470 \n   1471     @interfaces.legacy_generator_methods_support\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\engine\\training_generator.py in evaluate_generator(model, generator, steps, max_queue_size, workers, use_multiprocessing, verbose)\n    325 \n    326         while steps_done &lt; steps:\n--&gt; 327             generator_output = next(output_generator)\n    328             if not hasattr(generator_output, '__len__'):\n    329                 raise ValueError('Output of generator should be a tuple '\n\nc:\\users\\keboc\\anaconda3\\envs\\tensorflow_1.8\\lib\\site-packages\\keras\\utils\\data_utils.py in get(self)\n    783                 all_finished = all([not thread.is_alive() for thread in self._threads])\n    784                 if all_finished and self.queue.empty():\n--&gt; 785                     raise StopIteration()\n    786                 else:\n    787                     time.sleep(self.wait_time)\n\nStopIteration: \n<\/code><\/pre>\n\n<p>Thanks for your help<\/p>\n","answers":[]}
+{"QuestionId":53063860,"AnswerCount":1,"Tags":"<tensorflow><crop><object-detection>","CreationDate":"2018-10-30T12:00:21.263","AcceptedAnswerId":null,"OwnerUserId":9493219.0,"Title":"After detecting objects in a video stream, I want to crop and save these objects","Body":"<p>I had detected objects in a video and i want to crop these objects. I tried tensorflow APIs but none of them worked with me. When trying tf.image.crop_to_bounding_box(\n    image,\n    offset_height,\n    offset_width,\n    target_height,\n    target_width\n) \nit tells me that offset_height didn't defined. <\/p>\n\n<p>So, i need a guide into how to crop object from an image using tensorflow.<\/p>\n","answers":[{"AnswerId":"54035916","CreationDate":"2019-01-04T09:11:37.330","ParentId":null,"OwnerUserId":"10866602","Title":null,"Body":"<p>Try to use :\ntf.image.crop_and_resize(image,boxes,box_ind,crop_size,method='bilinear',extrapolation_value=0,name=None)<\/p>\n\n<p>For example: <\/p>\n\n<pre><code>    #bounding box coordinates\n    ymin =boxes[0][0][0]\n    xmin =boxes[0][0][1]\n    ymax =boxes[0][0][2]\n    xmax =boxes[0][0][3]\n    test = tf.image.crop_and_resize(image=frame_expanded\/255,\n           boxes=[[ymin,xmin,ymax,xmax]],\n           box_ind=[0],\n           crop_size=[100,100])\n<\/code><\/pre>\n\n<p>Worked for me!<\/p>\n\n<p>or if you want to use tf.image.crop_to_bounding_box(...) try this:<\/p>\n\n<pre><code>    #bounding box coordinates\n    ymin =boxes[0][0][0]\n    xmin =boxes[0][0][1]\n    ymax =boxes[0][0][2]\n    xmax =boxes[0][0][3]\n    #image_size\n    (im_width,im_height) = image.size\n    (xminn, xmaxx, yminn, ymaxx) = (xmin * im_width, xmax * im_width,   ymin * im_height, ymax * im_height)\n    test=tf.image.crop_to_bounding_box(frame,int(yminn), int(xminn), int(ymaxx-yminn), int(xmaxx-xminn))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53063944,"AnswerCount":1,"Tags":"<python><tensorflow><computer-vision><object-detection><image-recognition>","CreationDate":"2018-10-30T12:06:09.420","AcceptedAnswerId":53064651.0,"OwnerUserId":3935060.0,"Title":"Racecar image tagging","Body":"<p>I am working on a system to simplify our image library which grows anywhere from 7k to 20k new pictures per week.  The specific application is identifying which race cars are in pictures (all cars are similar shapes with different paint schemes).  I plan to use python and tensorflow for this portion of the project.<\/p>\n\n<p>My initial thought was to use image classification to classify the image by car; however, there is a very high probability of the picture containing multiple cars.  My next thought is to use object detection to detect the car numbers (present in fixed location on all cars [nose, tail, both doors, and roof] and consistent font week to week).  Lastly there is the approach of object recognition of the whole car. This, on the surface, seems to be the most practical; however, the paint schemes change enough that it may not be.<\/p>\n\n<p>Which approach will give me the best results?  I have pulled a large number of images out for training, and obviously the different methods require very different training datasets.<\/p>\n","answers":[{"AnswerId":"53064651","CreationDate":"2018-10-30T12:47:38.233","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>The best approach would be to use all 3 methods as an ensamble. You train all 3 of those models, and pass the input image to all 3 of them. Then, there are several ways how you can evaluate output. <\/p>\n\n<ol>\n<li><p>You can sum up the probabilities for all of the classes for all 3 models and then draw a conclusion based on the highest probability.<\/p><\/li>\n<li><p>You can get prediction from every model and decide based on number of votes: 1. model - class1, 2. model - class2, 3. model - class2 ==> class2<\/p><\/li>\n<li><p>You can do something like weighted decision making. So, let's say that first model is the best and the most robust one but you don't trust it 100% and want to see what other models will say. Than you can weight the output of the first model with 0.6, and output of other two models with weight of 0.2.<\/p><\/li>\n<\/ol>\n\n<p>I hope this helps :)<\/p>\n"}]}
+{"QuestionId":53064834,"AnswerCount":1,"Tags":"<python><tensorflow><keras><resnet>","CreationDate":"2018-10-30T12:57:24.640","AcceptedAnswerId":null,"OwnerUserId":4637381.0,"Title":"Can't reload a trained resnet50 model from a h5 file","Body":"<p>I am very new on neural network, and I am trying to use Keras to build a fish classifier. I got a python code that uses Keras with tensorflow-backend to create a RESNET-50 model, the code use model.save() to output a h5 file.<br>\n<br>\nI write another python file to reload the RESNET-50 model from the h5 file. I use keras.models.load_model() to load h5 file. But python give me the error message like below, and I don't knwo why. Please give me a hand, thanks.<br><\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"keras-cnn-predict.py\", line 75, in &lt;module&gt;\n    model=keras.models.load_model(model_config_filename)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\saving.py\", line 419, in load_model\n    model = _deserialize_model(f, custom_objects, compile)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\saving.py\", line 225, in _deserialize_model\n    model = model_from_config(model_config, custom_objects=custom_objects)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\saving.py\", line 458, in model_from_config\n    return deserialize(config, custom_objects=custom_objects)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 145, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\network.py\", line 1022, in from_config\n    process_layer(layer_data)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\network.py\", line 1008, in process_layer\n    custom_objects=custom_objects)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 147, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 1109, in from_config\n    return cls(**config)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\layers\\convolutional.py\", line 490, in __init__\n    **kwargs)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\layers\\convolutional.py\", line 117, in __init__\n    self.kernel_initializer = initializers.get(kernel_initializer)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\initializers.py\", line 508, in get\n    return deserialize(identifier)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\initializers.py\", line 503, in deserialize\n    printable_module_name='initializer')\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 147, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\initializers.py\", line 30, in from_config\n    return cls(**config)\n  File \"C:\\Python\\Python36\\lib\\site-packages\\keras\\initializers.py\", line 195, in __init__\n    'but got', distribution)\nValueError: ('Invalid `distribution` argument: expected one of {\"normal\", \"uniform\"} but got', 'truncated_normal')\n<\/code><\/pre>\n","answers":[{"AnswerId":"56574210","CreationDate":"2019-06-13T06:02:39.873","ParentId":null,"OwnerUserId":"10693366","Title":null,"Body":"<pre><code>D:\\Anaconda3\\lib\\site-packages\\keras\\initializers.py in __init__(self, scale, mode, distribution, seed)\n    193             raise ValueError('Invalid `distribution` argument: '\n    194                              'expected one of {\"normal\", \"uniform\"} '\n--&gt; 195                              'but got', distribution)\n    196         self.scale = scale\n    197         self.mode = mode\n\nValueError: ('Invalid `distribution` argument: expected one of {\"normal\", \"uniform\"} but got', 'truncated_normal')\n<\/code><\/pre>\n\n<p>In <code>import tensorflow.python.keras<\/code> replace <code>keras<\/code>. Then it's successful.<\/p>\n"}]}
+{"QuestionId":53065066,"AnswerCount":0,"Tags":"<python><keras><project-structure>","CreationDate":"2018-10-30T13:11:07.270","AcceptedAnswerId":null,"OwnerUserId":4841832.0,"Title":"Structuring a Keras project","Body":"<p>I am developing a Keras environment to train DL models, and I am struggling trying to decide the file structure of the project.<\/p>\n\n<p>Currently it looks like this:<\/p>\n\n<pre><code>.\n\u251c\u2500\u2500 build_model.py\n\u251c\u2500\u2500 ctalearn\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 data\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 data_loading.py\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 data_processing.py\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 image_mapping.py\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 __init__.py\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 pixel_pos_files\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 FACT_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 HESS-II_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 HESS-I_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 LST_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 MSTF_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 MSTN_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 MSTS_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 SST1_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u251c\u2500\u2500 SSTA_pos.npy\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0     \u2514\u2500\u2500 SSTC_pos.npy\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 __init__.py\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 predict.py\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 train.py\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 utils.py\n\u251c\u2500\u2500 example_model_schematics.yaml\n\u251c\u2500\u2500 example_train_config.yaml\n\u251c\u2500\u2500 models\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 build_cnn_rnn.py\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 cnn_rnn.yml\n\u251c\u2500\u2500 setup.py\n\u2514\u2500\u2500 test.py\n<\/code><\/pre>\n\n<p><code>train.py<\/code> and <code>predict.py<\/code> are the main scripts. They take a path to a model architecture stored in <code>.yml<\/code> format and a configuration file in <code>.yml<\/code> format as well and run the thing.<\/p>\n\n<p><code>data_loading.py<\/code> is a factory of <code>Sequence<\/code> objects that are fed to <code>model.fit_generator()<\/code> \/ a manager to access the real data files (which are located outside the project structure).<\/p>\n\n<p>The rest of the <code>data<\/code>folder are helper classes for the factory.<\/p>\n\n<p>The Keras architectures are currently build from a <code>.yaml<\/code> schematic by the script <code>build_model.py<\/code> or by a ad hoc script (for example <code>build_cnn_rnn.py<\/code>) and are stored in <code>.yml<\/code> format, to be loaded by <code>train.py<\/code> or <code>predict.py<\/code> and afterwards compiled and fed a data generator.<\/p>\n\n<p>Does my structure make sense?<\/p>\n\n<p>Where should I put <code>build_model.py<\/code>?<\/p>\n\n<p>Should I have separate <code>train<\/code> and <code>predict<\/code> scripts or mix them together and let the user choose if he want to predict or train via the <code>config.yml<\/code> file?<\/p>\n\n<p>Should the model path be specified as a command line argument or as an option inside the <code>config.yaml<\/code> file?<\/p>\n\n<p>Should I remove the <code>__init__.py<\/code> in the inner folder?<\/p>\n","answers":[]}
+{"QuestionId":53065077,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><loss-function>","CreationDate":"2018-10-30T13:11:52.253","AcceptedAnswerId":53070658.0,"OwnerUserId":1298461.0,"Title":"How do I get back the error values of a keras loss function (tensor)","Body":"<p>I would like to plot all the different loss functions available in Keras. Therefore I have created a dataframe and invoke the loss function. But how can I get back the values from the tensor?<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom keras import losses\n\npoints = 100\ndf = pd.DataFrame({\"error\": np.linspace(-3,3,points)})\ndf[\"mean_squared_error\"] = losses.mean_squared_error(np.zeros(points), df[\"error\"])\ndf.plot(x=\"error\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"53070658","CreationDate":"2018-10-30T18:27:37.193","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The loss functions in Keras return a Tensor object. You need to evaluate that Tensor object using the <code>eval()<\/code> function from the backend to get its actual value. Further, if you take a look at the definition of loss functions in Keras, say <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/a37b41d5ae799c338958b3eeeb884074dfba865f\/keras\/losses.py#L13\" rel=\"nofollow noreferrer\"><code>mean_squared_error()<\/code><\/a>, you would realize that there is <code>K.mean()<\/code> operation which takes the average over the last axis which is the output axis (don't confuse this with batch or sample axis). Therefore, you may need to pass the true and predicted values in a shape of <code>(n_samples, n_outputs)<\/code>, hence the reshapes in the following code:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom keras import losses\nfrom keras import backend as K\n\npoints = 100\ndf = pd.DataFrame({\"error\": np.linspace(-3,3,points)})\nmse_loss = losses.mean_squared_error(np.zeros((points,1)), df[\"error\"].values.reshape(-1,1))\ndf[\"mean_squared_error\"] = K.eval(mse_loss)\ndf.plot(x=\"error\")\n<\/code><\/pre>\n\n<p>Here is the output plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FGBHc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FGBHc.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53065420,"AnswerCount":3,"Tags":"<tensorflow><parallel-processing><deep-learning><gpu>","CreationDate":"2018-10-30T13:29:11.457","AcceptedAnswerId":53065806.0,"OwnerUserId":10580626.0,"Title":"By default, does TensorFlow use GPU\/CPU simultaneously for computing or GPU only?","Body":"<p>By default, TensorFlow will use our available GPU devices. That said, does TensorFlow use GPUs <strong>and CPUs<\/strong> simultaneously for <strong>computing<\/strong>, or GPUs for computing and CPUs for job handling (no matter how, CPUs are always active, as I think)?<\/p>\n","answers":[{"AnswerId":"53065806","CreationDate":"2018-10-30T13:48:56.867","ParentId":null,"OwnerUserId":"4095039","Title":null,"Body":"<p>Someone correct me if I'm wrong.<\/p>\n\n<p>But from what I'm aware TensorFlow only uses either GPU or CPU depending on what installation you ran. For example if you used <code>pip install TensorFlow<\/code> for python 2 or <code>python3 -m pip install TensorFlow<\/code> for python 3 you'll <strong>only<\/strong> use the CPU version.<\/p>\n\n<p>Vise versa for GPU.<\/p>\n\n<p>If you still have any questions or if this did not correctly answer your question feel free to ask me more.<\/p>\n"},{"AnswerId":"53065878","CreationDate":"2018-10-30T13:52:39.430","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Generally it uses both, the CPU and the GPU (assuming you are using a GPU-enabled TensorFlow). What actually gets used depends on the actual operations that your code is using.<\/p>\n\n<p>For each operation available in TensorFlow, there are several \"implementations\" of such operation, generally a CPU implementation and a GPU one. Some operations only have CPU implementations as it makes no sense for a GPU implementation, but overall most operations are available for both devices.<\/p>\n\n<p>If you make custom operations then you need to provide implementations that you want.<\/p>\n"},{"AnswerId":"53066142","CreationDate":"2018-10-30T14:06:12.077","ParentId":null,"OwnerUserId":"3299803","Title":null,"Body":"<p>TensorFlow operations come packaged with a list of devices they can execute on and a list of associated priorities. <\/p>\n\n<p>For example, a convolution is very conducive to computation on a GPU; but can still be done on a CPU whereas scalar additions should definitely be done on a CPU. You can override this selection using tf.Device and the key attached to the device of interest. <\/p>\n"}]}
+{"QuestionId":53065659,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-30T13:41:10.153","AcceptedAnswerId":null,"OwnerUserId":5863503.0,"Title":"Designing Deep Q Network Loss Function in Tensorflow","Body":"<p>I'm working on my first Deep Q Learning network, and I'm having difficulties designing the loss function in Tensorflow. As the loss function uses both the current network weights and previous outdated weights at the same time (see second equation <a href=\"https:\/\/storage.googleapis.com\/deepmind-media\/dqn\/DQNNaturePaper.pdf\" rel=\"nofollow noreferrer\">here<\/a>). I see only two ways to accomplish this:<\/p>\n\n<ol>\n<li>predict in advance all possible states of the outdated network when it is being updated and then use these values<\/li>\n<li>use <code>tf.train.Saver()<\/code> and switch between the models for every back-propagation<\/li>\n<\/ol>\n\n<p>Is there something more suitable for this in Tensorflow that Im missing out?<\/p>\n","answers":[{"AnswerId":"53065964","CreationDate":"2018-10-30T13:57:11.237","ParentId":null,"OwnerUserId":"3299803","Title":null,"Body":"<p>You need to store the outdated weights in different tf.Variables to be able to use them later. I don't have access to any of your code, but I'd recommend running your model construction twice, and using one as simple storage. Another solution would be to modify your current graph with two of every variable and creating connections to the copies. <\/p>\n\n<p>That is, if you're creating a TensorFlow variable A and you'd like to store its previous values as B you can do as follows: <\/p>\n\n<pre><code>A = tf.Variable(5)\nB = tf.Variable(0)\n\n# Use A to do something\nA = A * 5\n# Store the value of A in B\nB = A\n\nwith tf.Session() as sess:\n    sess.run(B) # Store A in B\n    sess.run(A) # Run an update on A\n\nprint A, B\n<\/code><\/pre>\n"}]}
+{"QuestionId":53066376,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-30T14:18:47.123","AcceptedAnswerId":null,"OwnerUserId":7571462.0,"Title":"When to use tensorflow estimators?","Body":"<p>I have a general tensorflow question about when to use estimators. I feel sometimes estimators are not convenient to build something, since we need to meet some fixed requirements when building the graph. On the other hand, using lower level api can be tedious sometimes. Therefore, I want to ask when it is proper to use estimators and when it is not. Thanks!<\/p>\n","answers":[{"AnswerId":"53066538","CreationDate":"2018-10-30T14:27:25.560","ParentId":null,"OwnerUserId":"9280994","Title":null,"Body":"<p>This is a very opinionated answer but I will still write it:<\/p>\n\n<p>The Estimator-API was developed to simplify building and sharing models. You could compare it with Keras and in fact Estimators is built with <code>tf.keras.layers<\/code> so one could say it is a simplification of a simplification.<\/p>\n\n<p>This is obviously good for beginners or people who come from other fields (as people working with ML often do) but can also delimit the things you can do.<\/p>\n\n<p>As a general rule of thumb I would use Estimators if you want to work on or share a model with people that do not have a good CS background but want to get going either way.<\/p>\n"}]}
+{"QuestionId":53066762,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2018-10-30T14:38:16.400","AcceptedAnswerId":null,"OwnerUserId":8814881.0,"Title":"Understanding 1D convolution of DNA sequences encoded as a one hot vector","Body":"<p>I am trying to run a classification task with convolution neural network for DNA sequences. The DNA sequences are being converted to a input array encoded as a one hot vector. For example - \"ACTG\" is encoded as [[1,0,0,0], [0,1,0,0], [0,0,0,1], [0,0,1,0]]. I have encoded each of the samples like that. The Dimension of the input would be number_of_samples * length_of_samples*4. I am trying to now understand how a 1D Convolution would work on an input array like this one, but I cannot workout how the output of the 1d convolution would look like. Would really appreciate some help. For reference I am using <a href=\"http:\/\/github.com\/kundajelab\/dfim\" rel=\"nofollow noreferrer\">this<\/a> code by Kundaje Lab, Standord University. I cannot understand how a 1D convolution would work for an input of 3 Dimensions.<\/p>\n","answers":[{"AnswerId":"53072395","CreationDate":"2018-10-30T20:27:08.940","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>Here is the documentation to the Keras <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">Conv1D<\/a> module, where they describe the input to the model as fixed or variable number of sequences of a fixed length (as given in the example, (10,128) : 10 sequences, each of length 128). <\/p>\n\n<p>1D convolution can be thought of as running through a single spatial or temporal dimension of a 2D data. This <a href=\"https:\/\/stackoverflow.com\/a\/44628011\/10111931\">stack overflow<\/a> answer gives a pretty clear explanation about the various types of Conv Layers.<\/p>\n\n<p>Coming to your problem, I have made a toy program with 2 conv layers and random data, which I think you might find useful.<\/p>\n\n<pre><code>data = np.random.random((64,4,4))\nlabels = np.random.random((64,2))\ndataset = tf.data.Dataset.from_tensor_slices((data,labels))\ndataset = dataset.batch(2).repeat()\ninputs = Input(shape=(4,4))\n\nx = Conv1D(32, 3, activation='relu')(inputs)\nx = Flatten()(x)\nx = Dense(32, activation='relu')(x)\npredictions = Dense(2, activation='softmax')(x)\n\nmodel = keras.Model(inputs=inputs, outputs=predictions)\n\nmodel.compile(optimizer='adam',\n          loss='categorical_crossentropy',\n          metrics=['accuracy'])\nmodel.summary()\nmodel.fit(dataset.make_one_shot_iterator(), epochs=5, steps_per_epoch=100)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, 4, 4)              0         \n_________________________________________________________________\nconv1d (Conv1D)              (None, 2, 32)             416       \n_________________________________________________________________\nflatten (Flatten)            (None, 64)                0         \n_________________________________________________________________\ndense (Dense)                (None, 32)                2080      \n_________________________________________________________________\ndense_1 (Dense)              (None, 2)                 66        \n=================================================================\nTotal params: 2,562\nTrainable params: 2,562\nNon-trainable params: 0\n\nEpoch 1\/5\n100\/100 [==============================] - 1s 11ms\/step - loss: 0.7027 - acc: 0.5450\nEpoch 2\/5\n100\/100 [==============================] - 1s 7ms\/step - loss: 0.6874 - acc: 0.6000\nEpoch 3\/5\n100\/100 [==============================] - 1s 7ms\/step - loss: 0.6838 - acc: 0.6200\nEpoch 4\/5\n100\/100 [==============================] - 1s 7ms\/step - loss: 0.6753 - acc: 0.6100\nEpoch 5\/5\n100\/100 [==============================] - 1s 7ms\/step - loss: 0.6656 - acc: 0.6300\n<\/code><\/pre>\n\n<p>now you could replace the 4 there with a shape (no_of_sequences,4) and define your own model this way.\nHowever, if you want to use something like (None,4), in the case where your sequences are variable and there is no fixed length you can choose, you would run in to trouble with the Dense layer when using Tensorflow backend, which requires the last dimension of the input.\nSo you could probably decide on the best shape that matches this requirement.<\/p>\n"}]}
+{"QuestionId":53067003,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network><pytorch>","CreationDate":"2018-10-30T14:49:47.777","AcceptedAnswerId":null,"OwnerUserId":9106489.0,"Title":"How to interpret the gradient norms in convolution neural networks?","Body":"<p>It might be basic thing, but I am unable to understand the interpretation of computing the gradient norms.<\/p>\n\n<p>For example, I am working on binary segmentation using architecture by stacking multiple U-Nets with each net has encoder similar to Resnet34. I trained this network and I obtain the gradient mean and norm values as below:<\/p>\n\n<ul>\n<li><blockquote>\n  <p>conv1 has mean grad of -1.77767194275e-14<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>conv1 has norm grad of 2.57230658463e-05<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>layer1 has mean grad of 1.27381299952e-11<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>layer1 has norm grad of 0.000390226632589<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>layer2 has mean grad of -2.3107595698e-11<\/p>\n<\/blockquote><\/li>\n<li><blockquote>\n  <p>layer2 has norm grad of 0.000265486567514<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p>All the values seems to be very low. I am not able to get shall I consider them as vanishing gradient problem or gradients can be really low depending upon the problem.<\/p>\n\n<p>If the gradients can be really low, what is best way to check for vanishing gradient problem?<\/p>\n\n<p>Any advice is appreciated to analyze the gradients data. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"53068543","CreationDate":"2018-10-30T16:11:17.380","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>The low mean values are not problematic. Of course some problem can cause the low values, but the low values don't mean a problem by itself. Moreover I think the norms of gradients are not too low.<\/p>\n\n<p>We should rethink what causes vanishing gradient problem in this case. When we use ReLU as activation function, gradients don't vanish (but they can be zero).<\/p>\n"}]}
+{"QuestionId":53067160,"AnswerCount":0,"Tags":"<tensorflow><levenberg-marquardt>","CreationDate":"2018-10-30T14:57:33.213","AcceptedAnswerId":null,"OwnerUserId":2729462.0,"Title":"Levenberg-Marquardt in tensorflow","Body":"<p>I wish to find out if anyone has tried to implement the well-known Levenberg-Marquardt algorithm in tensorflow? I have a number of issues on trying to implement it, during parameter updates. The following code snippet shows an implementation of the update function:<\/p>\n\n<pre><code>def func_var_update(cost, parameters):\n  # compute gradients or Jacobians for cost with respect to parameters\n  dloss_dw = tf.gradients(cost, parameters)[0]\n  # Return dimension of gradient vector\n  dim, _ = dloss_dw.get_shape()\n  # Compute hessian matrix using results of gradients\n  hess = []\n  for i in range(dim):\n        # Compute gradient ot Jacobian matrix for loss function\n        dfx_i = tf.slice(dloss_dw, begin=[i,0] , size=[1,1])\n        ddfx_i = tf.gradients(dfx_i, parameters)[0] \n        # Get the actual tensors at the end of tf.gradients\n        hess.append(ddfx_i)\n    hess = tf.squeeze(hess)     \n    dfw_new = tf.diag(dloss_dw)\n    # Update factor consisting of the hessian, product of identity matrix and Jacobian vector\n    JtJ = tf.linalg.inv(tf.ones((parameters.shape[0], parameters.shape[0])) + hess)\n    # product of gradient and damping parameter\n    pdt_JtJ = tf.matmul(JtJ, dloss_dw)\n    # Performing update here\n    new_params = tf.assign(parameters, parameters - pdt_JtJ)\n  return new_params\n<\/code><\/pre>\n\n<p>And the following call:<\/p>\n\n<pre><code>def mainfunc()\n    with tf.Session():\n   .....\n       vec_up = sess.run(func_var_update(), feed_dict=....)\n<\/code><\/pre>\n\n<p>results in the following error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Input is not invertible.\n<\/code><\/pre>\n\n<p>But Both the dimension of the Jacobian\/gradient and hessian are okay when I print them during runtime. The other problem I have is not being able to keep track of the parameters after each update and then adapt them to personal needs before feeding them into the optimizer later. I wanted to fix some parameters, and compute hessian and jacobian for others while performing optimization at the same time. Any help will be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53067348,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><pickle>","CreationDate":"2018-10-30T15:07:18.703","AcceptedAnswerId":null,"OwnerUserId":6547250.0,"Title":"How to load only a batch of images at a time in tensorflow?","Body":"<p>I am reading cifar100 images from <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">here<\/a> and I want to read the images in batches from the pickle file. With this code that I am using this line (Loaded %d examples.\" % num) loads all the images present in the training dataset. Then using tf.data i can read batches. But when it loads all the images my memory gets utilised and it does not even start the training. I am using something similar to <a href=\"https:\/\/github.com\/denocris\/cifar10-CNN-TensorFlow-Estimators\/blob\/master\/cifar10_5cnn.py\" rel=\"nofollow noreferrer\">this<\/a>. This link uses tfrecords and I want to read cifar data using pickle. So does anyone know how I can read only batches of data from cifar100 pickle file so that my memory does not become full ?<\/p>\n\n<pre><code>def read_data():\n    def in_data():\n        all_images = []\n        all_labels = []\n        with open(\"%s%s\" % (\".\/data\/cifar-100-python\/\", \"train\"),\"rb\") as fo:\n            dict = pickle.load(fo, encoding='latin1')\n            images = np.array(dict['data'])\n            labels = np.array(dict['fine_labels'])\n            num = images.shape[0]\n\n            # images = normalize(images)\n            images = images.astype(dtype=np.float32)\n            labels = labels.astype(dtype=np.int32)\n            images = np.reshape(images, [num, 3, 32, 32])\n            images = np.transpose(images, [0, 2, 3, 1])\n\n            print(\"Loaded %d examples.\" % num)\n            #print('BeforeLables: ', labels)\n            labels = one_hot_encode(labels)\n            #print('afterLables: ', labels)\n            all_images.append(images)\n            all_labels.append(labels)\n            # print('SIZE:', len(all_images))\n\n        all_images = np.concatenate(all_images)\n        all_labels = np.concatenate(all_labels)\n        self.size = len(all_images)\n        img_dataset = tf.data.Dataset.from_tensor_slices(all_images).batch(2)\n        label_dataset = tf.data.Dataset.from_tensor_slices(all_labels).batch(2)\n\n        dataset = tf.data.Dataset.zip((img_dataset, label_dataset)).repeat(None)\n\n        images, labels = dataset.make_one_shot_iterator().get_next()\n        print('image Shape:',images.shape)\n        return images, labels\n    return in_data\n<\/code><\/pre>\n","answers":[{"AnswerId":"53068260","CreationDate":"2018-10-30T15:55:32.400","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>I recommend that you read <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#decoding_image_data_and_resizing_it\" rel=\"nofollow noreferrer\" title=\"official tutorial\">tutorial about importing data<\/a>. There is very useful and very similar example. In the example, we don't embed image data to computation graph with from_tensor_slice. Instead we embed file names to the graph.<\/p>\n\n<p>Moreover if your data was a single too big data to load, you had to split preliminarily it to multiple files <\/p>\n"}]}
+{"QuestionId":53067504,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-30T15:14:43.760","AcceptedAnswerId":null,"OwnerUserId":10581141.0,"Title":"How tf.Session.run work when fetch include both Operation and Tensor?","Body":"<p>The value in the red rectangle is confusing:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/qdABZ.png\" alt=\"image1\"><\/p>\n\n<p>The description of <code>Session.run<\/code>:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/YC9hG.png\" alt=\"image2\"><\/p>\n\n<p>Why the value of Variable tmpA is 1.0, it should be 9999.0 since the value of B is 10001.0<\/p>\n\n<p>Can somebody explain how the function tf.Session.run work?<\/p>\n","answers":[{"AnswerId":"53067932","CreationDate":"2018-10-30T15:38:14.393","ParentId":null,"OwnerUserId":"1450609","Title":null,"Body":"<p>I recommend that you check the graph with Tensorboard. The order of calculation cause this result.<\/p>\n\n<p>When you run line 13 sess.run([A, ops, B]), the program <\/p>\n\n<ol>\n<li>fetch the value of A for tmpA (1.0)<\/li>\n<li>assign 9999 to A<\/li>\n<li>calculate B<\/li>\n<\/ol>\n\n<p>So the value of tmpA is 1.0, and tmpBB is 10001.<\/p>\n\n<p>Actually the order of evaluation of [A, ops, B] is not deterministic, so there is a possibility that tmpA become 9999 and tmpBB become 3.<\/p>\n"}]}
+{"QuestionId":53067722,"AnswerCount":0,"Tags":"<tensorflow><gpgpu><cudnn>","CreationDate":"2018-10-30T15:26:14.610","AcceptedAnswerId":null,"OwnerUserId":3667217.0,"Title":"TensorFlow error : checked narrowing failed; values not equal post-conversion Aborted (core dumped)","Body":"<p>I'm using TensorFlow 1.10 on Ubuntu 16.04 system. When I tried to set the <code>per_process_gpu_memory_fraction<\/code> to 10 to allow the GPU to use system memory to train a ResNet on very large RGB images (12000, 12000, 3), I got an error:<\/p>\n\n<pre><code>F:tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:91] Checked failed narrow == wide (-1990967296 vs 2304000000) checked narrowing failed; values not equal post-conversion Aborted (core dumped)\n<\/code><\/pre>\n\n<p>The code block referenced above is as follows:<\/p>\n\n<pre><code>\/\/ Converts (via narrowing) a type T value to a type U, and checks that the\n\/\/ value has no value change due to the conversion.\ntemplate &lt;typename WideT, typename NarrowT&gt;\nNarrowT CheckedNarrowing(const WideT&amp; wide) {\nNarrowT narrow = wide;\nCHECK_EQ(narrow, wide)\n    &lt;&lt; \"checked narrowing failed; values not equal post-conversion\";\nreturn narrow;\n}\n<\/code><\/pre>\n\n<p>A slightly smaller image, say (10000, 10000, 3), doesn't generate this error. I'm wondering what actually caused this problem and how can I solve this? <\/p>\n","answers":[]}
+{"QuestionId":53068437,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-30T16:05:45.427","AcceptedAnswerId":null,"OwnerUserId":10581436.0,"Title":"predict function is returning exponential values","Body":"<p>I have built and trained a sequential binary classification model using keras layers. Everything seems to work fine. Until I start using the predict method. This function starts to give me a weird exponential value rather than a probability.\n<a href=\"https:\/\/i.stack.imgur.com\/Xdpar.png\" rel=\"nofollow noreferrer\">This what I get after training and using predict method on the model<\/a><\/p>\n\n<p>This classification model has two classes lets say a cat or a dog, so I was expecting the result to be something like [99.9999, 0.0001] suggesting that its a cat. I'm not sure how to interpret the value that I'm getting back instead.<\/p>\n\n<p>Here is the code I have:<\/p>\n\n<pre><code># Get the data.\n    (train_texts, train_labels), (val_texts, val_labels) = data\n    train_labels = np.asarray(train_labels).astype('float32')\n    val_labels = np.asarray(val_labels).astype('float32')\n\n    # Vectorizing data\n    train_texts,val_texts, word_index = vectorize_data.sequence_vectorize(\n        train_texts, val_texts)\n\n    # Building the model architecture( adding layers to the model)\n    model = build_model.simple_model_layers(train_texts.shape[1:])\n\n    # Setting and compiling with the features like the optimizer, loss and metrics functions\n    model = build_model.simple_model_compile(model=model)\n\n    # This is when the learning happens\n    history = model.fit(train_texts,\n                        train_labels,\n                        epochs=EPOCHS,\n                        validation_data=(val_texts, val_labels),\n                        verbose=VERBOSE_OFF, batch_size=BATCH_SIZE)\n\n    print('Validation accuracy: {acc}, loss: {loss}'.format(\n        acc=history['val_acc'][-1], loss=history['val_loss'][-1]))\n\n    # loading data to predict on\n    test_text = any\n    with open('text_req.pickle', 'rb') as pickle_file:\n        test_text = pickle.load(pickle_file)\n\n\n    print('Lets make a prediction of this requirement:')\n    prediction = model.predict(test_text, batch_size=None, verbose=0, steps=None)\n    print(prediction)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53068510","CreationDate":"2018-10-30T16:09:30.170","ParentId":null,"OwnerUserId":"8274431","Title":null,"Body":"<p>Usually unskewed data are converted to its log values while training.\nSo <\/p>\n\n<pre><code>np.exp(your_value)\n<\/code><\/pre>\n\n<p>might be something you are looking for.<\/p>\n"}]}
+{"QuestionId":53068443,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><conda>","CreationDate":"2018-10-30T16:06:08.830","AcceptedAnswerId":null,"OwnerUserId":10581490.0,"Title":"I am having trouble installing Tensorflow - gpu into my anaconda virtual enviorment","Body":"<p>Every time I try and re-install Cuda it fails and whenever I try and import Tensorflow with the current set up I have it will pip install but it will not import and will instead return:<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import tensorflow as tf\n      Traceback (most recent call last):\n        File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n          from tensorflow.python.pywrap_tensorflow_internal import *\n        File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n          _pywrap_tensorflow_internal = swig_import_helper()\n        File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n          _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n        File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\imp.py\", line 243, in load_module\n          return load_dynamic(name, filename, file)\n        File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\imp.py\", line 343, in load_dynamic\n          return _load(spec)\n      ImportError: DLL load failed: The specified module could not be found.<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow__init__.py\", line 22, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\MBUSI\\Anaconda3\\pkgs\\conda-4.4.10-py36_0\\Library\\bin\\conda\\envs\\tensorflow2\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n","answers":[{"AnswerId":"53072377","CreationDate":"2018-10-30T20:25:40.753","ParentId":null,"OwnerUserId":"1410665","Title":null,"Body":"<p>This is basically DLL load error. The answer is already given in the past posts. \n<a href=\"https:\/\/stackoverflow.com\/questions\/20201868\/importerror-dll-load-failed-the-specified-module-could-not-be-found\">DLL failed-1<\/a>,<a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/9263\" rel=\"nofollow noreferrer\">DLL load failed during Pytorch<\/a>,<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10033\" rel=\"nofollow noreferrer\">DLL failure 3<\/a><\/p>\n"}]}
+{"QuestionId":53070199,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-10-30T17:56:34.077","AcceptedAnswerId":null,"OwnerUserId":2089899.0,"Title":"Get list of Keras variables","Body":"<p>I'd like to compare variables in a Keras model with those from a TensorFlow checkpoint. I can get the TF variables like this:<\/p>\n\n<p><code>vars_in_checkpoint = tf.train.list_variables(os.path.join(\".\/model.ckpt\"))<\/code><\/p>\n\n<p>How can I get the Keras variables to compare from my <code>model<\/code>?<\/p>\n","answers":[{"AnswerId":"53163169","CreationDate":"2018-11-05T22:28:19.443","ParentId":null,"OwnerUserId":"7649152","Title":null,"Body":"<p>You can get the variables of a Keras model via <code>model.weights<\/code> (list of <code>tf.Variable<\/code> instances).<\/p>\n"}]}
+{"QuestionId":53071237,"AnswerCount":1,"Tags":"<python><neural-network><keras><keras-layer>","CreationDate":"2018-10-30T19:03:52.710","AcceptedAnswerId":53071594.0,"OwnerUserId":5750238.0,"Title":"Can't figure out keras input shape error?","Body":"<p>I am trying to pass my neural network an array with shape with shape (1169909, 10, 10).<\/p>\n\n<p>But no matter what I do...<\/p>\n\n<pre><code>input_shape=(None,10,10)\ninput_shape=x_train.shape[1:]\ninput_shape=x_train.shape\ninput_shape=(1169909, 1)\ninput_shape=(10,10)\ninput_shape=(1169909,10)\ninput_shape=(1169909,10,10)\ninput_shape=(1,10,10)\n<\/code><\/pre>\n\n<p>I still get an error. The error changes:<\/p>\n\n<pre><code>but got array with shape (1169909, 10, 10)\nbut got array with shape (1169909, 1)\n<\/code><\/pre>\n\n<p>Depending on which way I give the input, which only adds to my confusion. <\/p>\n\n<p>The actual input looks like this, it's an array of these smaller 10x10 arrays:<\/p>\n\n<pre><code>array([[ 2, -1, -1, -1, -1, -1, -1, -1, -1,  0],\n       [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],\n       [11, -1, 11,  6,  3,  2, -1, -1, -1, 11],\n       [-1, -1, -1, -1,  5,  7, -1, -1,  2,  7],\n       [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],\n       [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1],\n       [-1, -1, -1, -1, 22, 25, -1, -1, -1, 22],\n       [22, -1, -1, -1, 26, 29, -1, -1, 26, 25],\n       [27, -1, -1, -1, -1, -1, -1, -1, -1, 23],\n       [31, 24, 31, -1, -1, -1, -1, -1, -1, 31]])\n<\/code><\/pre>\n\n<p>I've tried to understand this issue by looking on stack overflow and elsewhere, but I can't figure out the problem and those solutions don't work for me.<\/p>\n\n<p>Here's the model as of now:<\/p>\n\n<pre><code>x_train, x_test, y_train, y_test = train_test_split(xving, ywing, test_size=0.2)\n\nx_train = numpy.array(x_train)\ny_train = numpy.array(y_train)\n\n\nmodel = Sequential()\nmodel.add(keras.layers.Dense(250,activation='tanh', input_shape=(None,10,10)))\nmodel.add(keras.layers.Dense(150,activation='relu'))\nmodel.add(keras.layers.Dense(25,activation='sigmoid'))\nmodel.add(keras.layers.Dense(2,activation='softmax'))\noptimizerr = keras.optimizers.SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False)\nmodel.compile(optimizer=optimizerr, loss='categorical_crossentropy', metrics=['accuracy'])\n\nmodel.fit(x_train, y_train,epochs = 25, batch_size = 32, verbose=1)\n<\/code><\/pre>\n\n<p>EDIT:\nSo when I use:<\/p>\n\n<pre><code>input_shape=x_train.shape[1:]\n<\/code><\/pre>\n\n<p>The error becomes with the <em>target<\/em>:<\/p>\n\n<pre><code>Error when checking target: expected dense_32 to have 3 dimensions, but got array with shape (1169909, 1)\n<\/code><\/pre>\n\n<p>But the target is an array. When I keep it as a list, I instead get the error:<\/p>\n\n<pre><code>ValueError: Error when checking model target: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 1169909 arrays: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...\n<\/code><\/pre>\n\n<p>So now I guess my question concerns why the target is an issue. <\/p>\n\n<pre><code>y_train.shape\n<\/code><\/pre>\n\n<p>yields:<\/p>\n\n<pre><code>(1169909,)\n<\/code><\/pre>\n\n<p>So I'm confused still?<\/p>\n","answers":[{"AnswerId":"53071594","CreationDate":"2018-10-30T19:29:12.590","ParentId":null,"OwnerUserId":"9041710","Title":null,"Body":"<p>Assuming you have 1169909 pictures (or 2D data items) as a dataset you should use a 2D convolutional layer at the front if you want to pass them to your model as matrices, otherwise it would be no different than flattening each picture before passing them all to your model. I suggest using a convolutional schema, but if that's not the case you can flatten the array like this:<\/p>\n\n<pre><code>x = x.reshape((1169909, 100))\nmodel.add(keras.layers.Dense(250, activation='tanh', input_dim=100))\n<\/code><\/pre>\n\n<p>Again, passing 2D data to a dense layer will not take advantage of the structural properties of the data.<\/p>\n"}]}
+{"QuestionId":53071291,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><lstm>","CreationDate":"2018-10-30T19:07:59.673","AcceptedAnswerId":null,"OwnerUserId":814601.0,"Title":"Music generation with LSTM - loss stops decreasing","Body":"<p>I'm using a LSTM to generate pieces in the style of Bach. So far, I had success using only the melodic sequences of a single voice, ie, I'm able to generate a single melodic line with a <strong>fixed duration for each note<\/strong>.<\/p>\n\n<p>Now, I'm trying to model the <strong>four voices<\/strong> found on some Bach chorales <strong>and also<\/strong> the notes duration (see 'Chorales' midis at the bottom of <a href=\"http:\/\/www.bachcentral.com\/midiindexcomplete.html\" rel=\"nofollow noreferrer\">http:\/\/www.bachcentral.com\/midiindexcomplete.html<\/a>).<\/p>\n\n<p>The problem is that, at some point, my LSTM stops learning. Here is my architecture:<\/p>\n\n<p>For each voice, I extract a sequence of L notes as input, and the next note as output. Then I do the same for the notes duration. So I have:<\/p>\n\n<pre><code>voice0_x = sequence of notes, sequence of duration\nvoice0_y = next note, next duration\n\nvoice1_x = ... the same\nvoice1_y = ... the same\n<\/code><\/pre>\n\n<p>Then the nnet code goes like:<\/p>\n\n<pre><code>input_notes = Input(shape=(notes_in.shape[1], notes_in.shape[2]))\nnote_out = LSTM(256, return_sequences=True)(input_notes)\n\ninput_durations = Input(shape=(durations_in.shape[1], durations_in.shape[2]))\nduration_out = LSTM(256, return_sequences=True)(input_durations)\n\nmerge = concatenate([note_out, duration_out])\nmerge = Flatten()(merge)\nmerge = Dense(512, activation='relu')(merge)\nmerge = Dropout(0.3)(merge)\n\nnote_final = Dense(notes_out.shape[1], activation='softmax')(merge)\nduration_final = Dense(durations_out.shape[1], activation='softmax')(merge)\n\nmodel = Model(inputs=[input_notes, input_durations], outputs=[note_final, duration_final])\nmodel.compile(loss=['categorical_crossentropy', 'categorical_crossentropy'], optimizer = 'rmsprop')\n<\/code><\/pre>\n\n<p>So, I just feed the notes and durations sequences to separate LSTM's, concatenate their outputs and send everything to a dense layer. The network is clearly learning, but the loss stops decreasing at ~12.35:<\/p>\n\n<pre><code>loss: 12.3549 - dense_2_loss: 6.3240 dense_3_loss: 6.0309\n<\/code><\/pre>\n\n<p>I tried increasing the number of LSTM cells, without success. Given that the melodic lines of each voice have a relation with each other (the same goes for durations), is that the correct architecture? Should I increase even more the number of cells?<\/p>\n\n<p>To be a little bit more clear, suppose that my goal is to 'overfit' the input data, so in theory the network should converge to a small loss value, but that's not happening.<\/p>\n\n<p>Thanks for your help, and I can provide more details if needed.<\/p>\n","answers":[]}
+{"QuestionId":53071588,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-10-30T19:29:00.867","AcceptedAnswerId":null,"OwnerUserId":10581436.0,"Title":"Can't figure out why I'm only getting one value from predict function?","Body":"<p>I have built and trained a sequential binary classification model using keras layers. Everything seems to work fine until I start using the <code>predict<\/code> method. This function starts to give me a weird exponential value rather than probabilities of the two classes.\n<a href=\"https:\/\/i.stack.imgur.com\/Xdpar.png\" rel=\"nofollow noreferrer\">This what I get after training and using predict method on the model<\/a><\/p>\n\n<p>This classification model has two classes, let's say a cat or a dog, so I was expecting the result to be something like [99.9999, 0.0001] suggesting that it's a cat. I'm not sure how to interpret the value that I'm getting back instead.<\/p>\n\n<p>Here is the code I have:<\/p>\n\n<pre><code># Get the data.\n    (train_texts, train_labels), (val_texts, val_labels) = data\n    train_labels = np.asarray(train_labels).astype('float32')\n    val_labels = np.asarray(val_labels).astype('float32')\n\n    # Vectorizing data\n    train_texts,val_texts, word_index = vectorize_data.sequence_vectorize(\n        train_texts, val_texts)\n\n    # Building the model architecture( adding layers to the model)\n    model = build_model.simple_model_layers(train_texts.shape[1:])\n\n    # Setting and compiling with the features like the optimizer, loss and metrics functions\n    model = build_model.simple_model_compile(model=model)\n\n    # This is when the learning happens\n    history = model.fit(train_texts,\n                        train_labels,\n                        epochs=EPOCHS,\n                        validation_data=(val_texts, val_labels),\n                        verbose=VERBOSE_OFF, batch_size=BATCH_SIZE)\n\n    print('Validation accuracy: {acc}, loss: {loss}'.format(\n        acc=history['val_acc'][-1], loss=history['val_loss'][-1]))\n\n    # loading data to predict on\n    test_text = any\n    with open('text_req.pickle', 'rb') as pickle_file:\n        test_text = pickle.load(pickle_file)\n\n\n    print('Lets make a prediction of this requirement:')\n    prediction = model.predict(test_text, batch_size=None, verbose=0, steps=None)\n    print(prediction)\n<\/code><\/pre>\n\n<p>Here is how the simple model function looks like:<\/p>\n\n<pre><code>model = models.Sequential()\n    model.add(Dense(26, activation='relu', input_shape=input_shape))\n    model.add(Dense(16, activation='relu'))\n    model.add(Dense(10, activation='relu'))\n    model.add(Dense(1, activation='sigmoid'))\n    return model\n<\/code><\/pre>\n\n<p>Gradient desent functions:\n    optimizer='adam', loss='binary_crossentropy'<\/p>\n\n<p>Sample data is of String type which I convert to constant size matrices of 1\u2019s and 0\u2019s using padding and all. The features have two classes, so labels are simply 1 and 0. That's all for data. In my opinion, data doesn\u2019t seem to be the problem it could be something more trivial than that which I\u2018m overlooking and have failed to recognize.<\/p>\n\n<p><strong>Thank you guys, This last problem was resolved, but I need better understanding at this:<\/strong><\/p>\n\n<p>I read that <code>sigmoid<\/code> returns the probability of all possible classes and all the probabilities should add up to 1. The values that I am getting back are:<\/p>\n\n<pre><code>Validation accuracy: 0.792168688343232, loss: 2.8360600299145804\nLet's make a prediction of this requirement:\n\n    [[2.7182817, 1.       ]\n     [2.7182817, 1.       ]\n     [1.,        2.7182817]\n     [1. ,       2.7182817]]\n<\/code><\/pre>\n\n<p>They don't add up to 1 and looking at these values 1 or otherwise is not intuitive enough in what to make of it.<\/p>\n","answers":[{"AnswerId":"53072303","CreationDate":"2018-10-30T20:20:33.823","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>I want to clarify that you don't get a weird exponential value, you just get a weird value. The E is scientific notation for x10, so you basically get 2.7 x 10^-12. I'd love to help but I can't check your data nor your model. I tried to Google some parts of your code, in the hope to find some clarification but I can't seem to find out what's under the hood of these two lines:<\/p>\n\n<pre><code>model = build_model.simple_model_layers(train_texts.shape[1:])\nmodel = build_model.simple_model_compile(model=model)\n<\/code><\/pre>\n\n<p>I've no clue what network has been build, I'd like to know at least the loss function and the full final layer, that'd already be much to go by. Are you also sure that your data is correct?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>sigmoid does not do what you describe, softmax does that. Sigmoid is often used as multilabel classification, since it can detect multiple labels as True. Sigmoid output could look for example like [0.99, 0.3], it has the ability to look at each label separately. Softmax on the other hand doesn't, softmax could look like [0.99, 0.01], and the sum of all probabilities is always 1.<\/p>\n\n<p>That solved that confusion, now about your output, I've no clue what that is, it should be between 1 and 0, unless I'm missing something here.<\/p>\n\n<p>To answer your data question you asked to K. Streutker:\nThe goal of a neural network is to create the labels you feed it, on new data. If you want a probability distribution, then you also need to feed one. Every image should have a label [1, 0] and dog [0, 1], or reversed, whatever you like.  Then once the model is trained it will be able to give you two outputs that make sense. The loss function, most likely cross entropy takes these labels and the output of your model, and tries to minimize the difference over time. So this is sort of what you need:<\/p>\n\n<blockquote>\n  <p>image (dog)--> model --> loss --> optimizer that updates the weights<\/p>\n  \n  <p>labels ([0,1]) ------------------\u2518<\/p>\n<\/blockquote>\n\n<p>then predicting will look like this<\/p>\n\n<blockquote>\n  <p>image --> model --> labels<\/p>\n<\/blockquote>\n\n<p>Hope I helped a bit!<\/p>\n"},{"AnswerId":"53074859","CreationDate":"2018-10-31T01:02:34.583","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Your model only has one output. If your training labels are set to 0 for cat and 1 for dog then that means the network thinks its a cat if the output is <code>[[2.977094e-12]]<\/code>. If you want the probabilities of the two classes like you were expecting then you need to change the output of your model as follows:<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(Dense(26, activation='relu', input_shape=input_shape))\nmodel.add(Dense(16, activation='relu'))\nmodel.add(Dense(10, activation='relu')\nmodel.add(Dense(2, activation='softmax'))\n<\/code><\/pre>\n\n<p>Your labels would also need to change to <code>[1, 0]<\/code> and <code>[0, 1]<\/code> for cat and dog.<\/p>\n"}]}
+{"QuestionId":53072489,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-probability>","CreationDate":"2018-10-30T20:33:53.607","AcceptedAnswerId":null,"OwnerUserId":10582624.0,"Title":"Tensorflow_probability integer type error","Body":"<p>I am trying to use <code>tensorflow_probability<\/code> to construct a mcmc chain. This is my code:<\/p>\n\n<pre><code>    chain_states, kernel_results = tfp.mcmc.sample_chain(\n    num_results=tf.constant(1e3, dtype=tf.int32),\n    num_burnin_steps=tf.constant(1e2, dtype=tf.int32),\n    parallel_iterations=tf.constant(10, dtype=tf.int32),\n    current_state=current_state,\n    kernel=tfp.mcmc.MetropolisHastings(\n        inner_kernel=tfp.mcmc.HamiltonianMonteCarlo(\n        target_log_prob_fn=joint_log_prob,\n        num_leapfrog_steps=tf.constant(2, dtype=tf.int32),\n        step_size=tf.Variable(1.),\n        step_size_update_fn=tfp.mcmc.make_simple_step_size_update_policy()\n        )))\n<\/code><\/pre>\n\n<p>But I got this error::<\/p>\n\n<pre><code>&gt;     InvalidArgumentError                      Traceback (most recent call last) &lt;ipython-input-13-7e972cc65053&gt; in &lt;module&gt;()\n&gt; ----&gt; 1 make_model(well_complex, well_ligand, fi_complex, fi_ligand)\n&gt; \n&gt; ~\/Documents\/GitHub\/assaytools2\/assaytools2\/assaytools2\/inference.py in\n&gt; make_model(well_complex, well_ligand, fi_complex, fi_ligand)\n&gt;     162             num_leapfrog_steps=tf.constant(2, dtype=tf.int32),\n&gt;     163             step_size=tf.Variable(1.),\n&gt; --&gt; 164             step_size_update_fn=tfp.mcmc.make_simple_step_size_update_policy()\n&gt;     165             )))\n&gt;     166 \n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/sample.py\n&gt; in sample_chain(num_results, current_state, previous_kernel_results,\n&gt; kernel, num_burnin_steps, num_steps_between_results,\n&gt; parallel_iterations, name)\n&gt;     238 \n&gt;     239     if previous_kernel_results is None:\n&gt; --&gt; 240       previous_kernel_results = kernel.bootstrap_results(current_state)\n&gt;     241     return tf.scan(\n&gt;     242         fn=_scan_body,\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/metropolis_hastings.py\n&gt; in bootstrap_results(self, init_state)\n&gt;     261         name=mcmc_util.make_name(self.name, 'mh', 'bootstrap_results'),\n&gt;     262         values=[init_state]):\n&gt; --&gt; 263       pkr = self.inner_kernel.bootstrap_results(init_state)\n&gt;     264       if not has_target_log_prob(pkr):\n&gt;     265         raise ValueError(\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/hmc.py\n&gt; in bootstrap_results(self, init_state)\n&gt;     506   def bootstrap_results(self, init_state):\n&gt;     507     \"\"\"Creates initial `previous_kernel_results` using a supplied `state`.\"\"\"\n&gt; --&gt; 508     kernel_results = self._impl.bootstrap_results(init_state)\n&gt;     509     if self.step_size_update_fn is not None:\n&gt;     510       step_size_assign = self.step_size_update_fn(self.step_size, None)  # pylint:\n&gt; disable=not-callable\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/metropolis_hastings.py\n&gt; in bootstrap_results(self, init_state)\n&gt;     261         name=mcmc_util.make_name(self.name, 'mh', 'bootstrap_results'),\n&gt;     262         values=[init_state]):\n&gt; --&gt; 263       pkr = self.inner_kernel.bootstrap_results(init_state)\n&gt;     264       if not has_target_log_prob(pkr):\n&gt;     265         raise ValueError(\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/hmc.py\n&gt; in bootstrap_results(self, init_state)\n&gt;     672           init_target_log_prob,\n&gt;     673           init_grads_target_log_prob,\n&gt; --&gt; 674       ] = mcmc_util.maybe_call_fn_and_grads(self.target_log_prob_fn, init_state)\n&gt;     675       return UncalibratedHamiltonianMonteCarloKernelResults(\n&gt;     676           log_acceptance_correction=tf.zeros_like(init_target_log_prob),\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/util.py\n&gt; in maybe_call_fn_and_grads(fn, fn_arg_list, result, grads,\n&gt; check_non_none_grads, name)\n&gt;     232     fn_arg_list = (list(fn_arg_list) if is_list_like(fn_arg_list)\n&gt;     233                    else [fn_arg_list])\n&gt; --&gt; 234     result, grads = _value_and_gradients(fn, fn_arg_list, result, grads)\n&gt;     235     if not all(r.dtype.is_floating\n&gt;     236                for r in (result if is_list_like(result) else [result])):  # pylint: disable=superfluous-parens\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/mcmc\/util.py\n&gt; in _value_and_gradients(fn, fn_arg_list, result, grads, name)\n&gt;     207         ]\n&gt;     208       else:\n&gt; --&gt; 209         grads = tfe.gradients_function(fn)(*fn_arg_list)\n&gt;     210     else:\n&gt;     211       if is_list_like(result) and len(result) == len(fn_arg_list):\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/backprop.py\n&gt; in decorated(*args, **kwds)\n&gt;     368     \"\"\"Computes the gradient of the decorated function.\"\"\"\n&gt;     369 \n&gt; --&gt; 370     _, grad = val_and_grad_function(f, params=params)(*args, **kwds)\n&gt;     371     return grad\n&gt;     372 \n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/backprop.py\n&gt; in decorated(*args, **kwds)\n&gt;     469                        \"receive keyword arguments.\")\n&gt;     470     val, vjp = make_vjp(f, params)(*args, **kwds)\n&gt; --&gt; 471     return val, vjp(dy=dy)\n&gt;     472 \n&gt;     473   return decorated\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/backprop.py\n&gt; in vjp(dy)\n&gt;     539       return imperative_grad.imperative_grad(\n&gt;     540           _default_vspace, this_tape, nest.flatten(result), sources,\n&gt; --&gt; 541           output_gradients=dy)\n&gt;     542     return result, vjp\n&gt;     543 \n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/imperative_grad.py\n&gt; in imperative_grad(vspace, tape, target, sources, output_gradients)\n&gt;      61   \"\"\"\n&gt;      62   return pywrap_tensorflow.TFE_Py_TapeGradient(\n&gt; ---&gt; 63       tape._tape, vspace, target, sources, output_gradients)  # pylint: disable=protected-access\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/backprop.py\n&gt; in _gradient_function(op_name, attr_tuple, num_inputs, inputs,\n&gt; outputs, out_grads)\n&gt;     115     return [None] * num_inputs\n&gt;     116 \n&gt; --&gt; 117   return grad_fn(mock_op, *out_grads)\n&gt;     118 \n&gt;     119 \n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_grad.py\n&gt; in _ProdGrad(op, grad)\n&gt;     158   with ops.device(\"\/cpu:0\"):\n&gt;     159     rank = array_ops.rank(op.inputs[0])\n&gt; --&gt; 160     reduction_indices = (reduction_indices + rank) % rank\n&gt;     161     reduced = math_ops.cast(reduction_indices, dtypes.int32)\n&gt;     162     idx = math_ops.range(0, rank)\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py\n&gt; in binary_op_wrapper(x, y)\n&gt;     860     with ops.name_scope(None, op_name, [x, y]) as name:\n&gt;     861       if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):\n&gt; --&gt; 862         return func(x, y, name=name)\n&gt;     863       elif not isinstance(y, sparse_tensor.SparseTensor):\n&gt;     864         try:\n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\n&gt; in add(x, y, name)\n&gt;     322       else:\n&gt;     323         message = e.message\n&gt; --&gt; 324       _six.raise_from(_core._status_to_exception(e.code, message), None)\n&gt;     325 \n&gt;     326 \n&gt; \n&gt; ~\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/six.py in\n&gt; raise_from(value, from_value)\n&gt; \n&gt; InvalidArgumentError: cannot compute Add as input #0(zero-based) was\n&gt; expected to be a int32 tensor but is a int64 tensor [Op:Add] name:\n&gt; mcmc_sample_chain\/mh_bootstrap_results\/mh_bootstrap_results\/hmc_kernel_bootstrap_results\/maybe_call_fn_and_grads\/value_and_gradients\/add\/\n\nI doubled checked and none of my initial tensors were of integer type.\n\nI wonder where I did it wrong.\n\nThanks!\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53072772,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing>","CreationDate":"2018-10-30T20:53:47.183","AcceptedAnswerId":53073195.0,"OwnerUserId":7364377.0,"Title":"Issues with Tensor.eval() in session","Body":"<p>I have a really simple script that gets an image in tensorflow, decodes it, then attempts to store that tensor as a numpy array. However, in the current code below, tensorflow hangs indefinitely while evaluating the tensor in the session. <\/p>\n\n<pre><code>import tensorflow as tf\n\nresult = 0\n\n#Get image into appropriately sized tensor\nfilename_queue = \n\ntf.train.string_input_producer(['C:\\Python3\\TensorflowExp\\Pokemon_data\\filename0.jpg'])\nreader = tf.WholeFileReader()\nkey, value = reader.read(filename_queue)\nmy_img = tf.image.decode_jpeg(value, channels=1)\nresized_image = tf.image.resize_images(my_img, [299, 299])\n\n#print out tensor\nprint(resized_image)\n\n# Construct a `Session` to execute the graph.\nsess = tf.Session()\n\n#initialize Session\ninit = tf.global_variables_initializer()\nsess.run(init)\n\n#Print out numpy array in tensor \nwith sess.as_default():\n    result = resized_image.eval()\n<\/code><\/pre>\n\n<p>I tried adding a queue runner, initializing global variables, and running the session in various ways. In those cases, I get an unknown error that I cannot debug nor google what that problem is.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1350, in _do_call\n    return fn(*args)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1329, in _run_fn\n    status, run_metadata)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 473, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: NewRandomAccessFile failed to Create\/Open: C:\\Python3\\TensorflowExp\\Pokemon_datailename0.jpg : The filename, directory name, or volume label syntax is incorrect.\n; Unknown error\n         [[Node: ReaderReadV2 = ReaderReadV2[_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](WholeFileReaderV2, input_producer)]]\n         [[Node: DecodeJpeg\/_3 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_5_DecodeJpeg\", tensor_type=DT_UINT8, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \".\\ex5.py\", line 27, in &lt;module&gt;\n    print(resized_image.eval())\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 648, in eval\n    return _eval_using_default_session(self, feed_dict, self.graph, session)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 4758, in _eval_using_default_session\n    return session.run(tensors, feed_dict)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 895, in run\n    run_metadata_ptr)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1128, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1344, in _do_run\n    options, run_metadata)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1363, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: NewRandomAccessFile failed to Create\/Open: C:\\Python3\\TensorflowExp\\Pokemon_datailename0.jpg : The filename, directory name, or volume label syntax is incorrect.\n; Unknown error\n         [[Node: ReaderReadV2 = ReaderReadV2[_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](WholeFileReaderV2, input_producer)]]\n         [[Node: DecodeJpeg\/_3 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_5_DecodeJpeg\", tensor_type=DT_UINT8, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n\nCaused by op 'ReaderReadV2', defined at:\n  File \".\\ex5.py\", line 8, in &lt;module&gt;\n    key, value = reader.read(filename_queue)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\ops\\io_ops.py\", line 209, in read\n    return gen_io_ops._reader_read_v2(self._reader_ref, queue_ref, name=name)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\ops\\gen_io_ops.py\", line 713, in _reader_read_v2\n    queue_handle=queue_handle, name=name)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3160, in create_op\n    op_def=op_def)\n  File \"C:\\Python3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1625, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): NewRandomAccessFile failed to Create\/Open: C:\\Python3\\TensorflowExp\\Pokemon_datailename0.jpg : The filename, directory name, or volume label syntax is incorrect.\n; Unknown error\n         [[Node: ReaderReadV2 = ReaderReadV2[_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](WholeFileReaderV2, input_producer)]]\n         [[Node: DecodeJpeg\/_3 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_5_DecodeJpeg\", tensor_type=DT_UINT8, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>any guidance would be greatly appreciated, thanks!<\/p>\n","answers":[{"AnswerId":"53073195","CreationDate":"2018-10-30T21:30:47.023","ParentId":null,"OwnerUserId":"6067528","Title":null,"Body":"<p>This behaviour is expected because <code>tf.train.string_input_producer<\/code> is creating a queue pipeline from <code>tf.Queuebase<\/code>. If, as in your case, the queue is empty when the operation executes, it will block until there is an element to dequeue.<\/p>\n\n<p>Change the <code>with<\/code> block to the below to terminate:<\/p>\n\n<pre><code>with sess.as_default():\n\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    result = resized_image.eval()\n\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Coordinator\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Coordinator<\/a><\/p>\n"}]}
+{"QuestionId":53072916,"AnswerCount":1,"Tags":"<keras><autoencoder>","CreationDate":"2018-10-30T21:06:40.310","AcceptedAnswerId":null,"OwnerUserId":4722983.0,"Title":"Implementing variational autoencoder in keras with reconstruction probability","Body":"<p>I'm trying to implement the variational autoencoder in keras and use reconstruction probability instead of reconstruction error for anomaly detection. There is an example in deep learning 4j and someone has already asked the same question here: <a href=\"https:\/\/stackoverflow.com\/questions\/47070393\/variational-autoencoder-and-reconstruction-log-probability-vs-reconstruction-err\">Variational autoencoder and reconstruction Log Probability vs Reconstruction error<\/a><\/p>\n\n<p>Thanks for your help<\/p>\n","answers":[{"AnswerId":"53073216","CreationDate":"2018-10-30T21:33:21.703","ParentId":null,"OwnerUserId":"2061927","Title":null,"Body":"<p>Depends on your use case. In the example below, you can take the trace of the inner product of the reconstruction matrix and the input matrix (provided it makes sense to case the reconstruction matrix as a probability). Then edit your custom loss function to return that value instead (or in addition to) of standard VAE loss. Adam doesn't care what is being optimized, however, the nice benefits of using a VAE if you are not using its loss might vanish.\nFrom <a href=\"https:\/\/github.com\/samsinai\/VAE_protein_function\/blob\/master\/VAE_for_protein_function_prediction.ipynb\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><code>def compute_log_probability(one_hot_inp,pwm_output):\n    prod_mat=np.matmul(one_hot_inp.T,pwm_output)\n    log_prod_mat=np.log(prod_mat)\n    sum_diag=np.trace(log_prod_mat)\n    return sum_diag\n\noutput = x_decoded.reshape(dim1,dim2)\noutput = normalize(output,axis=0, norm='l1') #column-wise normalization in this case\nprob=compute_log_probability(input,output)\n<\/code><\/pre>\n\n<p>In the protein input case it makes sense to normalize per column because each column can only really have one value. In other cases you might want <\/p>\n"}]}
+{"QuestionId":53073085,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-30T21:20:30.733","AcceptedAnswerId":53078710.0,"OwnerUserId":3259896.0,"Title":"Select random value from row in a TF.record array, with limits on what the value can be?","Body":"<p>Say that I have a Tf.record file, and each row in the tf.records contain ints that are 0 or positive, and then padded with -1 so that all the rows are even size. So something like<\/p>\n\n<pre><code>0 3 43 223 23 -1 -1 -1\n4 12 3  11  435 2 4 -1\n9 3 11 32  34 322 9 7\n. \n. \n. \n<\/code><\/pre>\n\n<p>How do I randomly select 3 numbers from each of the rows ? <\/p>\n\n<p>The numbers will act like indexes to look up values in an embedding matrix, and then those embeddings will be averaged (basically word2vec CBOW model). <\/p>\n\n<p>More specifically, how do I avoid selecting the padding values of '-1'. -1 is just what I used to pad my rows so that each row will be the same size in order to use tf.record.(If there is a way to use varying length rows in tfrecords, let me know). <\/p>\n","answers":[{"AnswerId":"53078710","CreationDate":"2018-10-31T07:56:41.887","ParentId":null,"OwnerUserId":"4101431","Title":null,"Body":"<p>I think you're looking for something like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/VarLenFeature\" rel=\"nofollow noreferrer\">tf.VarLenFeature()<\/a>, more specifically, you do not necessarily have to pad your rows prior to creating the tfrecord file. You can create the tf_example, <\/p>\n\n<pre><code>from tensorflow.train import BytesList, Feature, Features, Example, Int64List\n\ntf_example = Example(\n    features=Features(\n        feature={\n            \"my_feature\": Feature(\n                int64_list=Int64List(value=[0,3,43,223,23])\n            )\n        })\n    )\n)\n\nwith TFRecordWriter(tfrecord_file_path) as tf_writer:\n    tf_writer.write(tf_example.SerializeToString())\n<\/code><\/pre>\n\n<p>Do this for all of your rows, that can vary in length. <\/p>\n\n<p>You'll parse the tf_examples with something like, <\/p>\n\n<pre><code>def parse_tf_example(example):\n    feature_spec = {\n        \"my_feature\": tf.VarLenFeature(dtype=tf.int64)\n    }\n    return tf.parse_example([example], features=feature_spec)\n<\/code><\/pre>\n\n<p>Now, this will return your features as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/SparseTensor\" rel=\"nofollow noreferrer\">tf.SparseTensor<\/a>s, if you don't want to deal with that at this stage, and carry on using tensor ops as you would normally, you can simply use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_tensor_to_dense\" rel=\"nofollow noreferrer\">tf.sparse_tensor_to_dense()<\/a> and carry on as you normally would with tensors. <\/p>\n\n<p>The returned dense tensors will be of varying lengths, so you shouldn't have to worry about selecting '-1's, there won't be any. <em>Unless you convert the sparse tensors to dense in batches<\/em>, in that case the batches will be padded to the length of the longest tensor in the batch, and the padding value can be set by the <code>default_value<\/code> parameter.<\/p>\n\n<p>That is in so far as your question about using varying length rows in tfrecords and getting back varying length tensors.<\/p>\n\n<p>With regards to the lookup op, I haven't used it myself, but I think <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/embedding_lookup_sparse\" rel=\"nofollow noreferrer\">tf.nn.embedding_lookup_sparse()<\/a> might help you out here, it offers the ability to lookup the embeddings from the sparse tensor, forgoing the need to convert it to a dense tensor first, and also has a <code>combiner<\/code> parameter to specify a reduction op on those embeddings, which in your case would be 'mean'.<\/p>\n\n<p>I hope this helps in some way, good luck. <\/p>\n"}]}
+{"QuestionId":53073967,"AnswerCount":1,"Tags":"<tensorflow><google-colaboratory>","CreationDate":"2018-10-30T22:49:02.630","AcceptedAnswerId":null,"OwnerUserId":8324477.0,"Title":"Fix no module 'layer' in current Google Colab TF version","Body":"<p>I use Google colab (python3 GPU)<\/p>\n\n<p>I want to run for example <a href=\"https:\/\/github.com\/tomlepaine\/fast-wavenet\" rel=\"nofollow noreferrer\">this repo codes<\/a> but I have an error when running demo ipynb in these lines:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom layers import (_causal_linear, _output_linear, conv1d, dilated_conv1d) \n<\/code><\/pre>\n\n<p>When I run these two lines I have an error \"no module layer\"<\/p>\n\n<p>I don't think this is a bug or something, because this repo has over 1000 stars.\nI think this is rather a tf version problem.<\/p>\n\n<p>Any ideas how to solve this?<\/p>\n","answers":[{"AnswerId":"53074271","CreationDate":"2018-10-30T23:29:13.203","ParentId":null,"OwnerUserId":"2426955","Title":null,"Body":"<p>\"Layers\" is a module in the package that you linked NOT in Tensorflow. See <a href=\"https:\/\/github.com\/tomlepaine\/fast-wavenet\/blob\/master\/wavenet\/layers.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>BTW if you wanted to import from a submodule of tensor flow you would have to do <code>from tensorflow.package import ...<\/code>.<\/p>\n"}]}
+{"QuestionId":53074258,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><mnist><tensor>","CreationDate":"2018-10-30T23:26:40.663","AcceptedAnswerId":53074403.0,"OwnerUserId":4618426.0,"Title":"Why does TensorFlow give me an error that I am feeding the incorrect shape and type into a placeholder?","Body":"<p>I can't figure this out. I've been going back and forth and I understand that I can copy and paste a working tutorial but I want to understand why this doesn't work.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom tensorflow.examples.tutorials.mnist import input_data\n\n\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)   \n\n#simple constants\nlearning_rate = .01\nbatch_size = 100\ntraining_epoch = 10\nt = 0\nl = t\n\n#gather the data\nx_train = mnist.train.images\ny_train = mnist.train.labels\nbatch_count = int(len(x_train)\/batch_size)\n\n#Set the variables\nY_ = tf.placeholder(tf.float32, [None,10], name = 'Labels')\nX = tf.placeholder(tf.float32,[None,784], name = 'Inputs')\nW = tf.Variable(tf.zeros([784,10]))\nb = tf.Variable(tf.zeros([10]))\n\n#Build the graph (Y = WX + b)\nY = tf.nn.softmax(tf.matmul(X,W) + b, name = 'softmax')\n\ncross_entropy = -tf.reduce_mean(Y_ * tf.log(Y)) * 1000.0\n\ncorrect_prediction = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1))\n\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\ntrain_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cross_entropy) \n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n\n    for epoch in range(training_epoch):\n        for i in range(batch_count):\n            t += batch_size\n            print(y_train[l:t].shape)\n            print(x_train[l:t].shape)\n            print(y_train[l:t].dtype)\n            sess.run(train_step,feed_dict={X: x_train[l:t], Y: y_train[l:t]})\n            l = t\n        print('Epoch = ', epoch)\n    print(\"Accuracy: \", accuracy.eval(feed_dict={X: x_test, Y_: y_test})) \n    print('Done') \n<\/code><\/pre>\n\n<p>The error message:<\/p>\n\n<pre><code>InvalidArgumentError: You must feed a value for placeholder tensor 'Labels_2' with dtype float and shape [?,10]\n     [[Node: Labels_2 = Placeholder[dtype=DT_FLOAT, shape=[?,10], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]\n<\/code><\/pre>\n\n<p>I also understand that there is more than I need to add to get this to work but I want to struggle through that on my own for now.  I'm running this on a jupyter notebook. I am positive that the <code>y_train<\/code> has a shape (100, 10) and a type float64.<\/p>\n\n<p>I've been stuck for a few days, so I appreciate the help.<\/p>\n","answers":[{"AnswerId":"53074318","CreationDate":"2018-10-30T23:34:24.937","ParentId":null,"OwnerUserId":"6067528","Title":null,"Body":"<p>You need to feed the placeholder tensor for <code>Y_<\/code> when you call <code>sess.run<\/code>. <\/p>\n\n<p>In <code>feed_dict<\/code>, just change <code>Y: y_train[l:t]<\/code> to <code>Y_: y_train[l:t]<\/code>. This will feed the <code>y_train[l:t]<\/code> into the placeholder.<\/p>\n"},{"AnswerId":"53074403","CreationDate":"2018-10-30T23:47:47.853","ParentId":null,"OwnerUserId":"10349960","Title":null,"Body":"<p>Change Y to Y_ in this line:<\/p>\n\n<pre><code>sess.run(train_step,feed_dict={X: x_train[l:t], Y_: y_train[l:t]})\n<\/code><\/pre>\n"}]}
+{"QuestionId":53074594,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-31T00:15:39.087","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Difficulty Preparing Data for Prediction Using tf.data.Dataset","Body":"<p>I've trained a Keras model using the <code>tf.data.Dataset<\/code> API.  So I have something like this:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((x, y))\ndataset = dataset.map(lambda x, y: ({'reviews': x}, y))\ndataset = dataset.map(preprocess)\nsess = tf.Session()\nsess.run(tf.tables_initializer())\ntf.keras.backend.set_session(sess)\ndataset = dataset.batch(50).repeat()\n<\/code><\/pre>\n\n<p>and then I do:<\/p>\n\n<pre><code>model.fit(dataset, steps_per_epoch=100, epochs=20)\n<\/code><\/pre>\n\n<p>and it all works.<\/p>\n\n<p>However, I then try to evaluate my model on a test <code>tf.data.Dataset<\/code> with:<\/p>\n\n<pre><code>dataset = dataset.map(preprocess)\nx = dataset.map(lambda x, y: x[\"reviews\"])\ny = dataset.map(lambda x, y: y)\npreds = model.predict(x)\n<\/code><\/pre>\n\n<p>and it fails before prediction with this error:<\/p>\n\n<pre><code>TypeError: cannot concatenate 'str' and 'int' objects\n<\/code><\/pre>\n\n<p>from this call:<\/p>\n\n<pre><code>x = dataset.map(lambda x, y: x[\"reviews\"])\n<\/code><\/pre>\n\n<p>I'm not sure what I'm doing incorrectly; when I inspect my <code>dataset<\/code>, I see elements that look like what I'd expect:<\/p>\n\n<pre><code>(array([30000,    28,  4417,    39,    77,    87,    52, 30000, 30000,\n          28,   955,    55, 30000,   707, 30000,   164,   295,   755,\n         183, 30000, 30000,    15,  1322, 30000, 30000,   102,    23,\n         755,   527,    52,   803,   804,  1125, 30000, 30000,    65,\n       30000,    21, 30000,   111,   127,   743,    23, 30000, 30000,\n        3337,    77,  2125,     5,   111,   220,   383,   102,    23,\n       30000,   527,   408,     5, 30000, 30000,   856,  1487,    28,\n          39,    14,    87,    52,    42,    31, 30000, 30000, 30000,\n       30000,    95,   153,    14,    87,   547,    12,   940, 30000,\n       30000,    15,   620,  1109,    55,    14,    87,    93,   408,\n          52, 30000, 30000, 30000,  1066,     6, 30000, 30000,    95,\n         153], dtype=int32), array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.,\n       0., 0., 0.], dtype=float32))\n<\/code><\/pre>\n\n<p>Any ideas what I'm doing incorrectly?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53074607,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><image-resizing><scikit-image>","CreationDate":"2018-10-31T00:18:15.037","AcceptedAnswerId":null,"OwnerUserId":2089899.0,"Title":"Efficiently resize batch of np.array images","Body":"<p>I have a 4D np.array size (10000,32,32,3) that represents a set of 10000 RGB images.<\/p>\n\n<p>How can I use <code>skimage.transform.resize<\/code> or other function to resize all images efficiently so that the (32,32) is interpolated to (224,224)? I'd prefer to do this with skimage, but I'm open to any solutions that don't use <code>tf.image.resize_images<\/code>.<\/p>\n\n<p>My current solution is using <code>tf.image.resize_images<\/code>, but it's causing GPU memory issues later down in my pipeline (won't free up memory after finishing in jupyter notebook) so I'd like to replace it. <\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import tensorflow as tf\nX = tf.image.resize_images(X,[224, 224])\nwith tf.Session() as sess:\n    X = X.eval()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53074650","CreationDate":"2018-10-31T00:27:44.297","ParentId":null,"OwnerUserId":"5067311","Title":null,"Body":"<p>One possibility is <a href=\"https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.ndimage.zoom.html#scipy.ndimage.zoom\" rel=\"nofollow noreferrer\"><code>scipy.ndimage.zoom<\/code><\/a> which can work with your collection of images and use spline interpolation of a given order to upsample your images:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.ndimage as ndi\n\nimgs_in = np.random.rand(100, 32, 32, 3)\nfactor = 224\/imgs_in.shape[1]\nimgs_out = ndi.zoom(imgs_in, (1, factor, factor, 1), order=2)\nprint(imgs_out.shape)\n<\/code><\/pre>\n\n<p>The resulting shape is <code>(100, 224, 224, 3)<\/code> as expected.<\/p>\n\n<p>You'll have to check whether the runtime and result is acceptable for your needs. Twiddling with the order of interpolation will probably affect this: there is a noticeable speed difference between second-order and (the default) third-order splines, at the cost of interpolation quality.<\/p>\n"},{"AnswerId":"53074736","CreationDate":"2018-10-31T00:41:39.573","ParentId":null,"OwnerUserId":"2089899","Title":null,"Body":"<p>I won't likely accept my own answer, but it seems that a simple for loop is actually fairly fast (says ~300% cpu utilization from <code>top<\/code>).<\/p>\n\n<pre><code>from skimage.transform import resize\n\nimgs_in = np.random.rand(100, 32, 32, 3)\nimgs_out = np.zeros((100,224,224,3))\n\nfor n,i in enumerate(imgs_in):\n    imgs_out[n,:,:,:] = resize(imgs_in[n,:,:,:], imgs_out.shape[1:], anti_aliasing=True)\n\nprint(imgs_out.shape)\n<\/code><\/pre>\n\n<p>Seems to be 7-8x faster than <code>ndi.zoom<\/code> on my machine. Trying to parallelize this further with <code>multiprocessing<\/code> would be even better I think.<\/p>\n"}]}
+{"QuestionId":53074703,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-10-31T00:35:15.253","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Failure to specify steps argument","Body":"<p>I have a keras model trained on a <code>tf.data.Dataset<\/code> like this:<\/p>\n\n<pre><code>dataset = dataset.map(preprocess)\nsess = tf.Session()\nsess.run(tf.tables_initializer())\ntf.keras.backend.set_session(sess)\ndataset = dataset.batch(50).repeat() \nmodel.fit(dataset, steps_per_epoch=100, epochs=20)\n<\/code><\/pre>\n\n<p>So I preprocess in the same way:<\/p>\n\n<pre><code>x = tf.data.Dataset.from_tensor_slices(data_test)\nx = x.map(lambda x: ({'reviews': x}))\nx = x.map(preprocess)\n<\/code><\/pre>\n\n<p>and then I do:<\/p>\n\n<pre><code>preds = self.model.predict(x)\n<\/code><\/pre>\n\n<p>but unfortunately get back:<\/p>\n\n<pre><code>ValueError: When using iterators as input to a model, you should specify the `steps` argument.\n<\/code><\/pre>\n\n<p>Adding in <code>x = x.batch(50).repeat()<\/code> before predicting leads to the same error.<\/p>\n\n<p>Any thoughts on what I'm doing incorrectly?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53075152,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning>","CreationDate":"2018-10-31T01:45:04.540","AcceptedAnswerId":null,"OwnerUserId":559379.0,"Title":"struggling to create graph file in Tensorflow","Body":"<p>i am using below command to generate trained model file<\/p>\n\n<pre>\n<code>\n    python tensorflow\/examples\/image_retraining\/retrain.py \\\n --image_dir=\/YourTrainImageFolder \\\n \u2014-output_graph=trainOut\/hero_graph.pb \\\n --output_labels=trainOut\/hero_labels.txt \\\n --summaries_dir=trainOut\/summaries\/ \\\n --bottleneck_dir=trainOut\/bottlenecks \\\n --model_dir=trainOut\/models\/ \\\n --saved_model_dir=trainOut\/saveModelDir\n\n<\/code>\n<\/pre>\n\n<p>It is generating all files except .pb file, i am following Tensorflow poets &amp; \n<a href=\"https:\/\/medium.com\/@elye.project\/applying-tensorflow-in-android-in-4-steps-to-recognize-superhero-f224597eb055\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@elye.project\/applying-tensorflow-in-android-in-4-steps-to-recognize-superhero-f224597eb055<\/a> tutorial <\/p>\n\n<p>kindly help<\/p>\n","answers":[]}
+{"QuestionId":53075203,"AnswerCount":1,"Tags":"<python><scikit-learn><keras><time-series><data-processing>","CreationDate":"2018-10-31T01:51:49.287","AcceptedAnswerId":53078923.0,"OwnerUserId":9841270.0,"Title":"How to do standardization on time series data with Scikit-learn Standard Scaler?","Body":"<p>I am using <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">Keras<\/a>, so the shape of data is (batch_size, timesteps, input_dim). And <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html\" rel=\"nofollow noreferrer\">Standard Scaler<\/a> just fits 2D data. <\/p>\n\n<p>One solution I thought was using partial fit and then transform.<\/p>\n\n<pre><code>scaler = StandardScaler()\nfor sample in range(data.shape[0]):\n    scaler.partial_fit(data[sample])\n\nfor sample in range(data.shape[0]):\n    data[sample] = scaler.transform(data[sample])\n<\/code><\/pre>\n\n<p>Is this a correct\/efficient approach?<\/p>\n","answers":[{"AnswerId":"53078923","CreationDate":"2018-10-31T08:12:25.783","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>You have two possibilities<\/p>\n\n<pre><code>data = np.random.randn(batch_size*time_length*nb_feats).reshape((bsize,time,feats))\n<\/code><\/pre>\n\n<p>Version 1 is doing what you say:<\/p>\n\n<pre><code>scaler = StandardScaler()\nfor sample in range(data.shape[0]):\n    scaler.partial_fit(data[sample])\n\nfor sample in range(data.shape[0]):\n    data[sample] = scaler.transform(data[sample])\n<\/code><\/pre>\n\n<p>Another possibility (Version 2) is to flatten the array, fit and transform and then reshape it <\/p>\n\n<pre><code>scaler = StandardScaler()\ndata   = scaler.fit_transform(data.reshape((bsize*time,feats))).reshape((bsize,time,feats))\n<\/code><\/pre>\n\n<p>In my computer <\/p>\n\n<p>Version 1 takes 0.8759770393371582  seconds<\/p>\n\n<p>Version 2 takes 0.11733722686767578 seconds<\/p>\n"}]}
+{"QuestionId":53075492,"AnswerCount":1,"Tags":"<python-3.x><keras><python-multiprocessing>","CreationDate":"2018-10-31T02:37:10.597","AcceptedAnswerId":null,"OwnerUserId":2089899.0,"Title":"Returning an object with multiprocessing.Process","Body":"<p>Is this possible? I'd like to modify this to return <code>model<\/code><\/p>\n\n<pre><code>from tensorflow.keras.models import load_model\nimport tensorflow as tf\nimport multiprocessing\nimport numpy as np\n\ndef get_model():\n    model = load_model(\".\/model.h5\")\n    optimizer = tf.keras.optimizers.SGD(lr=0.0001, momentum=0.99, nesterov=True)\n    model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\n    return model # can I get this back?\n\n# option 1: execute code with extra process\np = multiprocessing.Process(target=get_model)\np.start()\np.join()\n\n# wait until user presses enter key\ninput()\n<\/code><\/pre>\n\n<p>(I'm trying to do this because I read that it's one solution to not being able to free my GPU memory after loading this Keras model, which ends up causing a ResourceExhaustedError later on.)<\/p>\n","answers":[{"AnswerId":"53076630","CreationDate":"2018-10-31T05:05:34.587","ParentId":null,"OwnerUserId":"10583434","Title":null,"Body":"<p>if you load that to a global variable and keep the Memory dump i think your issue will going to be sorted.<\/p>\n\n<pre><code>`global model\nmodel = init()\ndef get_model():\n    model = load_model(\".\/model.h5\")\n    optimizer = tf.keras.optimizers.SGD(lr=0.0001, momentum=0.99, nesterov=True)\n    model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\nreturnt model\n`\n<\/code><\/pre>\n\n<p>and after that you will be able to call the global variable in yor code. <\/p>\n"}]}
+{"QuestionId":53075531,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-10-31T02:41:42.763","AcceptedAnswerId":null,"OwnerUserId":3367646.0,"Title":"Getting InvalidArgumentError: NewRandomAccessFile failed to Create\/Open while working with tf.data on the custom image dataset","Body":"<p>Code:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))\ndataset = dataset.map(_parseJpeg, num_parallel_calls = nthread)\n\ndataset = dataset.batch(batch_size)     \ndataset = dataset.prefetch(batch_size)\n\niterator = dataset.make_one_shot_iterator()\nimages, labels = iterator.get_next()\n\n with tf.Session() as sess:\n    sess.run([images, labels])\n<\/code><\/pre>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1278, in _do_call\n      return fn(*args)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1263, in _run_fn\n      options, feed_dict, fetch_list, target_list, run_metadata)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1350, in _call_tf_sessionrun\n      run_metadata)\n  tensorflow.python.framework.errors_impl.InvalidArgumentError: NewRandomAccessFile failed to Create\/Open: Tensor(\"arg0:0\", shape=(), dtype=string) : The filename, directory name, or volume label syntax is incorrect.\n  ; Unknown error\n           [[Node: ReadFile = ReadFile]]\n           [[Node: IteratorGetNext = IteratorGetNextoutput_shapes=[[?,224,224,3], [?,1000]], output_types=[DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]\n           [[Node: IteratorGetNext\/_1 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_37_IteratorGetNext\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>Traceback (most recent call last):\n    File \"vgg16_train_batch_processing.py\", line 231, in \n      sess.run(optimizer)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 877, in run\n      run_metadata_ptr)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1100, in _run\n      feed_dict_tensor, options, run_metadata)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1272, in _do_run\n      run_metadata)\n    File \"D:\\Python\\Anaconda\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1291, in _do_call\n      raise type(e)(node_def, op, message)\n  tensorflow.python.framework.errors_impl.InvalidArgumentError: NewRandomAccessFile failed to Create\/Open: Tensor(\"arg0:0\", shape=(), dtype=string) : The filename, directory name, or volume label syntax is incorrect.\n  ; Unknown error\n           [[Node: ReadFile = ReadFile]]\n           [[Node: IteratorGetNext = IteratorGetNextoutput_shapes=[[?,224,224,3], [?,1000]], output_types=[DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]\n           [[Node: IteratorGetNext\/_1 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_37_IteratorGetNext\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53079073","CreationDate":"2018-10-31T08:23:53.160","ParentId":null,"OwnerUserId":"3367646","Title":null,"Body":"<p>Never mind, fixed the problem myself.<\/p>\n\n<p>It was in the parser function for dataset map.<\/p>\n\n<p>Instead of <code>tf.read_file(filename)<\/code>, I put <code>tf.read_file(str(filename))<\/code> by mistake,\nwhich was causing the error because <code>filename<\/code> here is a tensor.<\/p>\n"}]}
+{"QuestionId":53075766,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-10-31T03:16:06.527","AcceptedAnswerId":53094502.0,"OwnerUserId":10438842.0,"Title":"How tensorflow pass through a deep learning model calculation?","Body":"<p>I am new to tensorflow and trying to figure out how tensorflow conduct the following task :<\/p>\n\n<p>(aligned facial images) > (facenet nn model with trained weights) > (embeddings calculation)<\/p>\n\n<p><a href=\"https:\/\/www.python36.com\/face-detection-matching-using-facenet\/\" rel=\"nofollow noreferrer\">https:\/\/www.python36.com\/face-detection-matching-using-facenet\/<\/a><\/p>\n\n<p>Here listed below are codes for the above task if I understand correctly, my question is whether if the calculation was executed via sess.run() or not? <\/p>\n\n<p>Correct me if I am wrong - (1) xxx.pb has everything for the model : layers and weights of the neural network, (2) declare images_placeholder and embeddings for the input and the output tensor of the neural network, (3) assign 'images' to input through feed_dict, and (4) execute sess.run() to pass through the neural network (xxx.pb) from input, 'feed_dict', to calculate output, 'embeddings'. <\/p>\n\n<p>What confuse to me is there is no other expression for 'embeddings' when executing :  <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>sess.run(embeddings, feed_dict=feed_dict)<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>By the way, for output tensor, does it matter if I replace name \"embeddings\" with \"output\" or any other name? It doesn't look to me there is code line for declaring \"embeddings\", same as \"input:0\".<\/p>\n\n<pre><code>    facenet.load_model(\"xxx.pb\")\n\n    # prepare placeholders for input and output tensors\n\n    images_placeholder = tf.get_default_graph().get_tensor_by_name(\"input:0\")\n    embeddings = tf.get_default_graph().get_tensor_by_name(\"embeddings:0\")\n\n    # execute the tensorflow graph with \"images\" input\n\n    feed_dict = { images_placeholder: images, ... }\n    images_embeddings = sess.run(embeddings,feed_dict=feed_dict)\n\n    &lt;facenet.py&gt;\n    ......\n    def load_model(model):\n        model_exp = os.path.expanduser(model)\n        if (os.path.isfile(model_exp)):\n            print('Model filename: %s' % model_exp)\n            with gfile.FastGFile(model_exp,'rb') as f:\n                graph_def = tf.GraphDef()\n                graph_def.ParseFromString(f.read())\n                tf.import_graph_def(graph_def, name='')\n        else        \n            ......\n<\/code><\/pre>\n","answers":[{"AnswerId":"53094502","CreationDate":"2018-11-01T02:36:22.913","ParentId":null,"OwnerUserId":"10438842","Title":null,"Body":"<p>Yes, the execution flow (1)~(4) stated above are proper. <\/p>\n\n<p>\"embeddings\" can be calculated via running the following code<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>python face_detect_demo.py --img=images\/faces.jpg<\/p>\n  <\/blockquote>\n<\/blockquote>\n"}]}
+{"QuestionId":53075868,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-10-31T03:30:54.930","AcceptedAnswerId":53082870.0,"OwnerUserId":6809144.0,"Title":"When to use an iterator in Tensorflow Estimator","Body":"<p>In the Tensorflow guides there are two separate places where the guide describes the input function for the Iris Data example. One input function returns just the dataset itself, while the other returns the dataset with an iterator.<\/p>\n\n<p>From the premade Estimator guide: <a href=\"https:\/\/www.tensorflow.org\/guide\/premade_estimators\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/guide\/premade_estimators<\/a><\/p>\n\n<pre><code>def train_input_fn(features, labels, batch_size):\n\"\"\"An input function for training\"\"\"\n# Convert the inputs to a Dataset.\ndataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))\n\n# Shuffle, repeat, and batch the examples.\nreturn dataset.shuffle(1000).repeat().batch(batch_size)\n<\/code><\/pre>\n\n<p>From the custom estimator guide: <a href=\"https:\/\/www.tensorflow.org\/guide\/custom_estimators\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/guide\/custom_estimators<\/a><\/p>\n\n<pre><code>def train_input_fn(features, labels, batch_size):\n\"\"\"An input function for training\"\"\"\n# Convert the inputs to a Dataset.\ndataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))\n\n# Shuffle, repeat, and batch the examples.\ndataset = dataset.shuffle(1000).repeat().batch(batch_size)\n\n# Return the read end of the pipeline.\nreturn dataset.make_one_shot_iterator().get_next()\n<\/code><\/pre>\n\n<p>I'm confused which one is correct, and if they both are used for different cases, when is it correct to return the dataset using an iterator?<\/p>\n","answers":[{"AnswerId":"53082870","CreationDate":"2018-10-31T11:59:59.623","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>If your input function returns a <code>tf.data.Dataset<\/code>, an iterator is created under the hood and its <code>get_next()<\/code> function is used to supply inputs to the model. This is somewhat hidden in the source code, see <code>parse_input_fn_result<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/estimator\/util.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I believe this was only implemented in a more recent update, so older tutorials still explicitly return <code>get_next()<\/code> in their input function since it was the only option back then.  There should be no difference between using either, but you can save a tiny bit of code by returning the dataset instead of the iterator.<\/p>\n"}]}
+{"QuestionId":53076293,"AnswerCount":1,"Tags":"<python><c++><tensorflow><bazel>","CreationDate":"2018-10-31T04:27:01.790","AcceptedAnswerId":53077848.0,"OwnerUserId":2175052.0,"Title":"When compiling Tensorflow with GPU support, CUDA_TOOLKIT_PATH unbound variable is encountered","Body":"<p>I am following <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">this link<\/a> to compile Tensorflow with GPU support. After <code>.\/configure<\/code>, I then type in <code>bazel build --config=opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package<\/code>. During the compilation, I see the following error.<\/p>\n\n<pre>\nERROR: \/home\/super\/.cache\/bazel\/_bazel_super\/6fad3aac5da8510114d7933cea773172\/external\/nccl_archive\/BUILD.bazel:139:1: Executing genrule @nccl_archive\/\/:device_code_fatbin_h failed (Exit 1)\n\/bin\/bash: CUDA_TOOLKIT_PATH: unbound variable\nTarget \/\/tensorflow\/tools\/pip_package:build_pip_package failed to build\nUse --verbose_failures to see the command lines of failed build steps.\nINFO: Elapsed time: 0.973s, Critical Path: 0.12s\nINFO: 4 processes: 4 local.\nFAILED: Build did NOT complete successfully\n<\/pre>\n\n<p>I checked the <code>.tf_configure.bazelrc<\/code> file and the <code>CUDA_TOOLKIT_PATH<\/code> is defined.<\/p>\n\n<pre>\nbuild --action_env PYTHON_BIN_PATH=\"\/home\/super\/anaconda3\/envs\/tf\/bin\/python\"\nbuild --action_env PYTHON_LIB_PATH=\"\/home\/super\/anaconda3\/envs\/tf\/lib\/python3.6\/site-packages\"\nbuild --python_path=\"\/home\/super\/anaconda3\/envs\/tf\/bin\/python\"\nbuild:xla --define with_xla_support=true\nbuild --config=xla\nbuild --action_env TF_NEED_OPENCL_SYCL=\"0\"\nbuild --action_env TF_NEED_ROCM=\"0\"\nbuild --action_env TF_NEED_CUDA=\"1\"\nbuild --action_env CUDA_TOOLKIT_PATH=\"\/usr\/local\/cuda\"\nbuild --action_env TF_CUDA_VERSION=\"10.0\"\nbuild --action_env CUDNN_INSTALL_PATH=\"\/usr\/local\/cuda-10.0\"\nbuild --action_env TF_CUDNN_VERSION=\"7\"\nbuild --action_env TF_NCCL_VERSION=\"\"\nbuild --action_env TF_CUDA_COMPUTE_CAPABILITIES=\"5.2\"\nbuild --action_env LD_LIBRARY_PATH=\"\/usr\/local\/cuda-10.0\/lib64\"\nbuild --action_env TF_CUDA_CLANG=\"0\"\nbuild --action_env GCC_HOST_COMPILER_PATH=\"\/usr\/bin\/gcc\"\nbuild --config=cuda\ntest --config=cuda\nbuild:opt --copt=-march=native\nbuild:opt --copt=-Wno-sign-compare\nbuild:opt --host_copt=-march=native\nbuild:opt --define with_default_optimizations=true\nbuild:v2 --define=tf_api_version=2\n<\/pre>\n\n<p>My <code>.bashrc<\/code> has the following.<\/p>\n\n<pre>\nexport PATH=\/usr\/local\/cuda-10.0\/bin${PATH:+:${PATH}}\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}\nexport CUDA_HOME=\/usr\/loca\/cuda\nexport CUDA_TOOLKIT_PATH=\/usr\/local\/cuda\n<\/pre>\n\n<p>My environment settings are as follows.<\/p>\n\n<ul>\n<li>bazel 0.18.0<\/li>\n<li>CUDA 10.0<\/li>\n<li>cuDNN 7.3.1<\/li>\n<li>Ubuntu 16.04<\/li>\n<\/ul>\n\n<p>Any ideas on what's going on?<\/p>\n","answers":[{"AnswerId":"53077848","CreationDate":"2018-10-31T06:53:48.683","ParentId":null,"OwnerUserId":"8777621","Title":null,"Body":"<p>My assumption is that you're trying to compile directly from the master branch. I was running into the same issue earlier, but after checking out a published release (v1.12.0-rc2), I was able to compile TensorFlow without anymore issues. Try <code>git checkout v1.12.0-rc2<\/code> then run through the <code>.\/configure<\/code> and bazel compilation again. The master branch updates very frequently. My thought is that the build was broken at the moment we were trying to install.<\/p>\n"}]}
+{"QuestionId":53076989,"AnswerCount":1,"Tags":"<python><keras><cleverhans>","CreationDate":"2018-10-31T05:41:59.100","AcceptedAnswerId":53475655.0,"OwnerUserId":8568673.0,"Title":"'RandomForestClassifier' object has no attribute 'layers'","Body":"<p>I'm trying to attack my random forest classifier.<\/p>\n\n<pre><code>clf = RandomForestClassifier(max_features=\"sqrt\", n_estimators=500, n_jobs=-1, verbose=1, warm_start=True)\nclf.fit(X_train, y_train)\n<\/code><\/pre>\n\n<p>After this definition I do my predictions and after that I did the code below:<\/p>\n\n<pre><code>from keras import backend\nfrom cleverhans.utils_keras import KerasModelWrapper\nfrom cleverhans.attacks import FastGradientMethod\nsess =  backend.get_session()\n\nwrap = KerasModelWrapper(clf)\nfgsm = FastGradientMethod(wrap, sess=sess)\nfgsm_params = {'eps': 0.15,\n               'clip_min': 0.,\n               'clip_max': 1.}\n\nadv_x = fgsm.generate_np(X_test, **fgsm_params) \nadv_x.shape\n<\/code><\/pre>\n\n<p>And at --> 10 <strong>adv_x = fgsm.generate_np(X_test, **fgsm_params)<\/strong> I get an attribute error.<\/p>\n\n<blockquote>\n  <p>AttributeError: 'RandomForestClassifier' object has no attribute\n  'layers'<\/p>\n<\/blockquote>\n\n<p>I mean, my classifier does not have layers but how can I do this fgsm attack? Is there a way to add randomforestclassifier to sequential model to have layers? Or is there another way to attack?<\/p>\n","answers":[{"AnswerId":"53475655","CreationDate":"2018-11-26T06:18:01.347","ParentId":null,"OwnerUserId":"10560104","Title":null,"Body":"<p>You will not be able to run the FGSM attack on a sklearn model because CleverHans would not be able to compute the gradients needed to find the direction in which to perturb the input to find an adversarial example. To compute these gradients, a symbolic definition of the model is needed, which requires the model to be defined using TensorFlow (either directly or through a high-level abstraction like Keras). <\/p>\n\n<p>If you'd like to use a sklearn model, you can however wrap it as a CleverHans model and use <em>only<\/em> <strong>gradient-free<\/strong> attacks such as SPSA. <\/p>\n"}]}
+{"QuestionId":53077159,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-31T05:58:15.543","AcceptedAnswerId":null,"OwnerUserId":1976743.0,"Title":"tensorflow worker process hang in cluster mode if one worker process close the session later","Body":"<p>using the code in this question as example, save the code as trainer.py you can try launch two worker and one ps using command:<\/p>\n\n<pre><code>python trainer.py \\\n     --ps_hosts=localhost:2222 \\\n     --worker_hosts=localhost:2223,localhost:2224 \\\n     --job_name=worker --task_index=0\n\npython trainer.py \\\n     --ps_hosts=localhost:2222 \\\n     --worker_hosts=localhost:2223,localhost:2224 \\\n     --job_name=worker --task_index=1\n\npython trainer.py \\\n     --ps_hosts=localhost:2222 \\\n     --worker_hosts=localhost:2223,localhost:2224 \\\n     --job_name=ps --task_index=0\n<\/code><\/pre>\n\n<p>after the train finished, worker 0 will hang and won't exit, if I remove the following code <\/p>\n\n<pre><code>    if is_chief:\n        import time\n        time.sleep(2)\n<\/code><\/pre>\n\n<p>then all workers will exit at the same time. It seems all worker must close the session at exact same time. Is this expected or I am not using it properly?<\/p>\n\n<pre><code>import argparse\nimport sys\n\nimport tensorflow as tf\nimport numpy as np\n\nFLAGS = None\n\n\ndef main(_):\n  ps_hosts = FLAGS.ps_hosts.split(\",\")\n  worker_hosts = FLAGS.worker_hosts.split(\",\")\n\n  # Create a cluster from the parameter server and worker hosts.\n  cluster = tf.train.ClusterSpec({\"ps\": ps_hosts, \"worker\": worker_hosts})\n\n  # Create and start a server for the local task.\n  server = tf.train.Server(cluster,\n                           job_name=FLAGS.job_name,\n                           task_index=FLAGS.task_index)\n\n  if FLAGS.job_name == \"ps\":\n    server.join()\n  elif FLAGS.job_name == \"worker\":\n\n    train_X = np.linspace(-1.0, 1.0, 100)\n    train_Y = 2.0 * train_X + np.random.randn(*train_X.shape) * 0.33 + 10.0\n\n    X = tf.placeholder(\"float\")\n    Y = tf.placeholder(\"float\")\n    # Assigns ops to the local worker by default.\n    with tf.device(tf.train.replica_device_setter(\n        worker_device=\"\/job:worker\/task:%d\" % FLAGS.task_index,\n        cluster=cluster)):\n\n      # Build model...\n      w = tf.Variable(0.0, name=\"weight\")\n      b = tf.Variable(0.0, name=\"bias\")\n      loss =  tf.square(Y - tf.multiply(X, w) - b)\n      global_step = tf.contrib.framework.get_or_create_global_step()\n\n      train_op = tf.train.AdagradOptimizer(0.01).minimize(\n          loss, global_step=global_step)\n\n    # The StopAtStepHook handles stopping after running given steps.\n    hooks=[tf.train.StopAtStepHook(last_step=10000)]\n\n    is_chief = (FLAGS.task_index == 0)\n    # The MonitoredTrainingSession takes care of session initialization,\n    # restoring from a checkpoint, saving to a checkpoint, and closing when done\n    # or an error occurs.\n    with tf.train.MonitoredTrainingSession(master=server.target,\n                                           is_chief=is_chief,\n                                           checkpoint_dir=\"\/tmp\/train_logs\",\n                                           hooks=hooks) as mon_sess:\n      while not mon_sess.should_stop():\n        # Run a training step asynchronously.\n        # See &lt;a href=\".\/..\/api_docs\/python\/tf\/train\/SyncReplicasOptimizer\"&gt;&lt;code&gt;tf.train.SyncReplicasOptimizer&lt;\/code&gt;&lt;\/a&gt; for additional details on how to\n        # perform *synchronous* training.\n        # mon_sess.run handles AbortedError in case of preempted PS.\n        for (x, y) in zip(train_X, train_Y):\n            if mon_sess.should_stop():\n                break\n            _, step, loss_value = mon_sess.run([train_op, global_step, loss], feed_dict={X: x,\n                                                                       Y: y})\n\n            print(\"Step: {}, loss: {}\".format(step, loss_value))\n        if is_chief:\n            import time\n            time.sleep(2)\n\nif __name__ == \"__main__\":\n  parser = argparse.ArgumentParser()\n  parser.register(\"type\", \"bool\", lambda v: v.lower() == \"true\")\n  # Flags for defining the tf.train.ClusterSpec\n  parser.add_argument(\n      \"--ps_hosts\",\n      type=str,\n      default=\"\",\n      help=\"Comma-separated list of hostname:port pairs\"\n  )\n  parser.add_argument(\n      \"--worker_hosts\",\n      type=str,\n      default=\"\",\n      help=\"Comma-separated list of hostname:port pairs\"\n  )\n  parser.add_argument(\n      \"--job_name\",\n      type=str,\n      default=\"\",\n      help=\"One of 'ps', 'worker'\"\n  )\n  # Flags for defining the tf.train.Server\n  parser.add_argument(\n      \"--task_index\",\n      type=int,\n      default=0,\n      help=\"Index of task within the job\"\n  )\n  FLAGS, unparsed = parser.parse_known_args()\n  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53077164,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-10-31T05:58:59.590","AcceptedAnswerId":null,"OwnerUserId":9255615.0,"Title":"Difference between graph_pb2.GraphDef and tf.GraphDef","Body":"<p>What is the difference between graph_pb2.GraphDef and tf.GraphDef?<\/p>\n\n<p>Are these two equivalent? If not, how are they different, which one should I use?<\/p>\n\n<pre><code>with open(\"input.pb\", \"rb\") as f:\n    graph_def = graph_pb2.GraphDef()\n    graph_def.ParseFromString(f.read())\n\nwith open(\"input.pb\", \"rb\") as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53077429,"AnswerCount":1,"Tags":"<tensorflow><anaconda><python-3.6>","CreationDate":"2018-10-31T06:22:38.860","AcceptedAnswerId":53077859.0,"OwnerUserId":10583636.0,"Title":"Error when importing tensorflow, ImportError: DLL load failed: The specified module could not be found","Body":"<p><strong>System information:<\/strong><\/p>\n\n<p>win10<\/p>\n\n<p>python 3.6.2<\/p>\n\n<p>tensorflow 1.11.0<\/p>\n\n<p>numpy 1.15.3<\/p>\n\n<p>conda 4.3.30<\/p>\n\n<p>cuda 8.0<\/p>\n\n<p>cudnn 6.0<\/p>\n\n<p>pip 18.1<\/p>\n\n<p>I created a new environment with python3.6.2 in anaconda and  installed tensorflow with pip, and met with such a problem, thanks for help.<\/p>\n\n<blockquote>\n  <p>Python 3.6.2 |Continuum Analytics, Inc.| (default, Jul 20 2017, 12:30:02) [MSC v.1900 64 bit (AMD64)] on win32\n  Type \"help\", \"copyright\", \"credits\" or \"license\" for more information.<\/p>\n  \n  <blockquote>\n    <p>import tensorflow as tf<\/p>\n  <\/blockquote>\n  \n  <p>Traceback (most recent call last):\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n    File \"I:\\Anaconda\\envs\\envPyTF\\lib\\imp.py\", line 242, in load_module\n      return load_dynamic(name, filename, file)\n    File \"I:\\Anaconda\\envs\\envPyTF\\lib\\imp.py\", line 342, in load_dynamic\n      return _load(spec)\n  ImportError: DLL load failed: \u627e\u4e0d\u5230\u6307\u5b9a\u7684\u6a21\u5757\u3002<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>Traceback (most recent call last):\n    File \"\", line 1, in \n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow__init__.py\", line 22, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n      from tensorflow.python import pywrap_tensorflow\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n      raise ImportError(msg)\n  ImportError: Traceback (most recent call last):\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()\n    File \"C:\\Users\\SZF\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n    File \"I:\\Anaconda\\envs\\envPyTF\\lib\\imp.py\", line 242, in load_module\n      return load_dynamic(name, filename, file)\n    File \"I:\\Anaconda\\envs\\envPyTF\\lib\\imp.py\", line 342, in load_dynamic\n      return _load(spec)\n  ImportError: DLL load failed: \u627e\u4e0d\u5230\u6307\u5b9a\u7684\u6a21\u5757\u3002<\/p>\n  \n  <p>Failed to load the native TensorFlow runtime.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53077859","CreationDate":"2018-10-31T06:55:04.180","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>Maybe try installing it with conda\n<a href=\"https:\/\/anaconda.org\/conda-forge\/tensorflow\" rel=\"nofollow noreferrer\">tensorflow <\/a>\n<a href=\"https:\/\/anaconda.org\/anaconda\/tensorflow-gpu\" rel=\"nofollow noreferrer\">tensorflow-gpu<\/a><\/p>\n"}]}
+{"QuestionId":53077602,"AnswerCount":1,"Tags":"<python><lstm><pytorch>","CreationDate":"2018-10-31T06:34:58.200","AcceptedAnswerId":null,"OwnerUserId":10584166.0,"Title":"PyTorch: nn.LSTM output different results for the same inputs in the same batch","Body":"<p>I try to implement a two-layer bidirectional LSTM with <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?#torch.nn.LSTM\" rel=\"nofollow noreferrer\">torch.nn.LSTM<\/a>. <\/p>\n\n<p>I made a toy example: a batch of 3 tensors, which are exactly the same (see my code below). And I expected the outputs of the BiLSTM  to be the the same along the batch dimension, i.e. <code>out[:,0,:] == out[:,1,:] == out[:, 2, :]<\/code>.<\/p>\n\n<p>But it seemed not to be the case. According to my experiments, 20%~40% of the time, the output were not the same. So I wonder where I got it wrong.<\/p>\n\n<pre><code># Python 3.6.6, Pytorch 0.4.1\nimport torch\n\ndef test(hidden_size, in_size):\n    seq_len, batch = 4, 3\n    bilstm = torch.nn.LSTM(input_size=in_size, hidden_size=hidden_size, \n                            num_layers=2, bidirectional=True)\n\n    # create a batch with 3 exactly the same tensors\n    a = torch.rand(seq_len, 1, in_size)  # (seq_len, 1, in_size)\n    x = torch.cat((a, a, a), dim=1)\n\n    out, _ = bilstm(x)  # (seq_len, batch, n_direction * hidden_size)\n\n    # expect the output should be the same along the batch dimension\n    assert torch.equal(out[:, 0, :], out[:, 1, :])  \n    assert torch.equal(out[:, 1, :], out[:, 2, :])\n\nif __name__ == '__main__':\n    count, total = 0, 0\n    for h_size in range(1, 51):\n        for in_size in range(1, 51):\n            total += 1\n            try:\n                test(h_size, in_size)\n            except AssertionError:\n                count += 1\n    print('percentage of assertion error:', count \/ total)\n<\/code><\/pre>\n","answers":[{"AnswerId":"54795245","CreationDate":"2019-02-20T21:06:32.917","ParentId":null,"OwnerUserId":"11092171","Title":null,"Body":"<p>What is confusing you is the float precision. Floating point operations are slightly inaccurate and can differ by very small amounts\nUse this instead:  <\/p>\n\n<pre><code>torch.set_default_dtype(torch.float64) \n<\/code><\/pre>\n\n<p>Then you will see they should be the same along the batch dim.<\/p>\n\n<p>Thanks for correcting some English grammar mistakes.<\/p>\n"}]}
+{"QuestionId":53078505,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorboard><google-aiy>","CreationDate":"2018-10-31T07:42:30.107","AcceptedAnswerId":null,"OwnerUserId":6471140.0,"Title":"google aiy vision kit compiler: get output node from tensorboard","Body":"<p>I'm trying to deploy to a google aiy vision kit v2 a custom model trained with Keras API in Tensorflow,when compiling the model from a <code>.pb<\/code> Tensorflow file with the vision kit compiler i get: <\/p>\n\n<pre><code>output node named \"&lt;output node&gt;\" does not appear in the graph.\n<\/code><\/pre>\n\n<p>My question is,  how can i get from tensorboard the output node passed to the compiler in the parameter <code>--output_tensor_names<\/code>?<\/p>\n\n<p>I already tried the values returned by:<\/p>\n\n<pre><code>keras_model.outputs[0].op.name    # returns final_output\/Sigmoid\nkeras_model.outputs[0].name    # returns final_output\/Sigmoid:0\n<\/code><\/pre>\n\n<p>and <code>final_output<\/code>, which is the name given to the last layer in keras.<\/p>\n\n<p>But none of these worked, how can i see the correct expected node in tensorboard?<\/p>\n","answers":[]}
+{"QuestionId":53078632,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><multi-gpu>","CreationDate":"2018-10-31T07:51:51.627","AcceptedAnswerId":null,"OwnerUserId":10411151.0,"Title":"Pytorch 3-GPUs, just can only use 2 of them to train","Body":"<p>I have three 1080TI, but when train I can only use 2 of them..<\/p>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre><code>device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel.cuda()\n\ncriterion = nn.CrossEntropyLoss().cuda()\noptimizer_conv = optim.SGD(model.classifier.parameters(), lr=0.0001, momentum=0.9)\nexp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)\n<\/code><\/pre>\n\n<p><strong>train part:<\/strong><\/p>\n\n<pre><code>outputs = nn.parallel.data_parallel(model,inputs,device_ids=[0,1,2])\n<\/code><\/pre>\n\n<p><strong>With \"CUDA_VISIBLE_DEVICES=\"1,2,3\" python train.py\"\nGot this:<\/strong>  <\/p>\n\n<pre><code>| 22%   35C    P8    10W \/ 250W |     12MiB \/ 11178MiB |      0%  \n| 43%   59C    P2    92W \/ 250W |   1169MiB \/ 11178MiB |     49%   \n| 44%   60C    P2    91W \/ 250W |   1045MiB \/ 11175MiB |     54% \n<\/code><\/pre>\n\n<p><strong>With \"CUDA_VISIBLE_DEVICES=\"0,1,2\" python train.py\"\nGot this:<\/strong>  <\/p>\n\n<pre><code>| 21%   38C    P2    95W \/ 250W |   1169MiB \/ 11178MiB |     78%      Default |  \n| 42%   63C    P2    93W \/ 250W |    777MiB \/ 11178MiB |     76%      Default |  \n| 43%   64C    P0    85W \/ 250W |    282MiB \/ 11175MiB |      0%      Default |\n<\/code><\/pre>\n","answers":[{"AnswerId":"53079059","CreationDate":"2018-10-31T08:22:18.970","ParentId":null,"OwnerUserId":"10411151","Title":null,"Body":"<p>eeeee.. I found the reason:<br>\nmy batchsize = 4 when there are three GPUs<br>\nso Change batchsize bigger can solve this \"weird\" problem  <\/p>\n"}]}
+{"QuestionId":53078725,"AnswerCount":0,"Tags":"<python><tensorflow><google-cloud-platform><object-detection><object-detection-api>","CreationDate":"2018-10-31T07:57:26.617","AcceptedAnswerId":null,"OwnerUserId":9460641.0,"Title":"Google cloud ml-engine prediction failure","Body":"<p>I am new to Tensorflow and followed <a href=\"https:\/\/cloud.google.com\/blog\/products\/gcp\/performing-prediction-with-tensorflow-object-detection-models-on-google-cloud-machine-learning-engine\" rel=\"nofollow noreferrer\">this<\/a> Google blog to deploy my object detection model to GCP and try to get predictions. When I run the <code>$ gcloud ml-engine predict --model=${YOUR_MODEL} --version=${YOUR_VERSION} --json-instances=inputs.json\n<\/code> command in the terminal to run predictions, I get the following error:<\/p>\n\n<pre><code> {\n  \"error\": \"Prediction failed: Error processing input: Expected uint8, got '\\\\xff\\\\xd8\\\\xff\\\\xe0.....\\\\xd9' of type 'str' instead.\"\n }\n<\/code><\/pre>\n\n<p>My <strong>inputs.json<\/strong> file has the base64 encoded form of the input image for the prediction to run on:<\/p>\n\n<pre><code>{\"b64\": \"\/9j\/4AAQSk....QB\/\/9k=\"}\n<\/code><\/pre>\n\n<p>What's going wrong here and how can I fix it?<\/p>\n\n<p><strong>UPDATE<\/strong>: I just found out that the format for the input image in <strong>inputs.json<\/strong> is <\/p>\n\n<p><code>{\"instances\": [{\"b64\": \"ASa8asdf\"}, {\"b64\": \"JLK7ljk3\"}]}\n<\/code>\nfrom <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/v1\/predict-request\" rel=\"nofollow noreferrer\">this<\/a> docs post. But I am still getting the same error.<\/p>\n","answers":[]}
+{"QuestionId":53079108,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-31T08:26:05.960","AcceptedAnswerId":null,"OwnerUserId":5656827.0,"Title":"Error when loading Keras model trained by tensorflow","Body":"<p>I'm trying to load a model that I trained and saved using Tensorflow &amp; Keras, but it's given me an error.<\/p>\n\n<p>Python version: 3.6.6<\/p>\n\n<p>Tensorflow version: 1.11.0<\/p>\n\n<p>Output:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/engine\/saving.py\", line 230, in load_model\n    model = model_from_config(model_config, custom_objects=custom_objects)\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/engine\/saving.py\", line 310, in model_from_config\n    return deserialize(config, custom_objects=custom_objects)\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/layers\/serialization.py\", line 64, in deserialize\n    printable_module_name='layer')\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/utils\/generic_utils.py\", line 173, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/engine\/sequential.py\", line 339, in from_config\n    custom_objects=custom_objects)\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/layers\/serialization.py\", line 64, in deserialize\n    printable_module_name='layer')\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/utils\/generic_utils.py\", line 175, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 1617, in from_config\n    return cls(**config)\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/layers\/advanced_activations.py\", line 310, in __init__\n    if max_value is not None and max_value &lt; 0.:\nTypeError: '&lt;' not supported between instances of 'dict' and 'float'\n<\/code><\/pre>\n\n<p>I've also tried just saving the weights instead of the whole model, but that doesn't seem more successful:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"predict_from_NN.py\", line 44, in <\/p>\n\n<pre><code>    model.load_weights('\/home\/me\/Data\/Out\/finished_model_2_weights.hdf5.index')\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/keras\/engine\/network.py\", line 1526, in load_weights\n    checkpointable_utils.streaming_restore(status=status, session=session)\n  File \"\/packages\/tensorflow\/1.11.0\/Python-3.6.6\/tensorflow\/python\/training\/checkpointable\/util.py\", line 880, in streaming_restore\n    \"Streaming restore not supported from name-based checkpoints. File a \"\nNotImplementedError: Streaming restore not supported from name-based checkpoints. File a feature request if this limitation bothers you.\n<\/code><\/pre>\n\n<p>Although I'm not exactly sure why\/how I'm doing a 'streaming restore', and google is not very useful in both cases.<\/p>\n\n<p>In case it helps, here's the code for my model:<\/p>\n\n<pre><code>from tensorflow.python.keras.layers import Conv2D, MaxPooling2D, ReLU\n<\/code><\/pre>\n\n<p>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Flatten, Activation, Dense<\/p>\n\n<pre><code>def cnn_model(img_rows, img_cols, img_channels):\n    model = Sequential()\n    model.add(Conv2D(64, (3, 3),activation='linear',kernel_initializer='he_uniform',\n                     input_shape=(img_rows, img_cols, img_channels)))\n    model.add(ReLU())   # add an advanced activation\n    model.add(MaxPooling2D(pool_size=(5, 5)))\n    model.add(Conv2D(32, (3, 3),activation='linear',kernel_initializer='he_uniform'))\n    model.add(ReLU())   # add an advanced activation\n    model.add(MaxPooling2D(pool_size=(3, 3)))\n    model.add(Conv2D(16, (3, 3),activation='linear',kernel_initializer='he_uniform'))\n    model.add(ReLU())   # add an advanced activation\n    model.add(MaxPooling2D(pool_size=(3, 3)))\n    model.add(Flatten())\n    model.add(Dense(1024))\n    model.add(Dense(1024))\n    model.add(ReLU())   # add an advanced activation\n    model.add(Dense(4))\n    model.add(Activation('softmax'))\n\n    return model\n<\/code><\/pre>\n\n<p>And I save my model like this:<\/p>\n\n<pre><code>model.save(os.path.join(output_folder, model_name + '_GPU.hdf5'))\n<\/code><\/pre>\n\n<p>And try to load it like this:<\/p>\n\n<pre><code>from tensorflow.python.keras.models import load_model\nmodel = load_model(model_file)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58390541","CreationDate":"2019-10-15T08:39:54.683","ParentId":null,"OwnerUserId":"12199329","Title":null,"Body":"<p>Try updating Keras as I also faced similar error and updating repositories helped. \nUpdate keras using github link in terminal\n<code>git clone https:\/\/github.com\/fchollet\/keras.git<\/code>. In terminal redirect to location and run\n    <code>sudo python3 setup.py install<\/code>. Hope this helps you too.<\/p>\n"},{"AnswerId":"53079453","CreationDate":"2018-10-31T08:47:34.050","ParentId":null,"OwnerUserId":"9805238","Title":null,"Body":"<p>Have you tried saving and loading the model with the \"to_json\" function as described below?<\/p>\n\n<pre><code>from keras.models import model_from_json\n[...]\n# serialize model to JSON\nmodel_json = model.to_json()\nwith open(\"model.json\", \"w\") as json_file:\n    json_file.write(model_json)\n# serialize weights to HDF5\nmodel.save_weights(\"model.h5\")\nprint(\"Saved model to disk\")\n\n# later...\n\n# load json and create model\njson_file = open('model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n# load weights into new model\nloaded_model.load_weights(\"model.h5\")\nprint(\"Loaded model from disk\")\n<\/code><\/pre>\n\n<p>P.S: I borrowed this code from <a href=\"https:\/\/machinelearningmastery.com\/save-load-keras-deep-learning-models\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53079436,"AnswerCount":1,"Tags":"<tensorflow><python-3.6>","CreationDate":"2018-10-31T08:46:24.633","AcceptedAnswerId":53084507.0,"OwnerUserId":1059860.0,"Title":"tensorflow Tf.cond giving unexpected output","Body":"<p>I seem to be having a misunderstanding on how <code>tf.cond<\/code> works. In the tensorflow <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\">documentation<\/a>, it gives the following example:<\/p>\n\n<pre><code>z = tf.multiply(a, b)\nresult = tf.cond(x &lt; y, lambda: tf.add(x, z), lambda: tf.square(y))\n<\/code><\/pre>\n\n<p>The result of the example, if <code>x&lt;y<\/code> is <code>True<\/code> is <code>tf.add(x,z)<\/code> else <code>tf.square(y)<\/code><\/p>\n\n<p>Following this example, I am trying to build a small example with tf.cond and the result doesnt go along the lines mentioned in the documentation.<\/p>\n\n<p>in my example, <code>deterministic_action = 4<\/code>, <code>random_action = 11<\/code>, <code>chose_random=False<\/code>. The <code>stochastic_action<\/code> should be <code>4<\/code>, instead it is <code>1<\/code>.\nWhere did the value 1 come from?<\/p>\n\n<pre><code>#!\/usr\/bin\/env python3\n\nimport tensorflow as tf\nimport numpy as np\n\nwith tf.Graph().as_default():\n    with tf.device('\/cpu:0'):\n        stochastic_ph = tf.placeholder(tf.bool, (), name=\"stochastic\")\n        eps = tf.get_variable(\"eps\", (), initializer=tf.constant_initializer(0))\n        with tf.variable_scope('test_cond') as sc:\n            deterministic_action = tf.random_uniform([], minval=0, maxval=15, dtype=tf.int64, seed=0) # 4\n            random_action = tf.random_uniform([], minval=0, maxval=15, dtype=tf.int64, seed=1) # 11\n            chose_random = tf.random_uniform([], minval=0, maxval=1, dtype=tf.float32) &lt; eps # False because eps = 0\n            stochastic_action = tf.cond(chose_random, lambda: random_action, lambda: deterministic_action) # S_action should be 4 but it is 1\n            #output_action = tf.cond(stochastic_ph, lambda: stochastic_action, lambda: deterministic_action)\n\n\n    init = tf.global_variables_initializer()\n    sess = tf.Session()\n    sess.run(init, feed_dict={stochastic_ph: True})\n    print (\"s_ph = \", stochastic_ph)\n    d_action = sess.run(deterministic_action)\n    print (\"det_action= \", d_action)\n    r_action = sess.run(random_action)\n    print (\"rand_action= \", r_action)\n    e = sess.run(eps)\n    c_action = sess.run(chose_random)\n    print (\"chose_rand= \", c_action)\n    s_action = sess.run(stochastic_action)\n    print (\"s_action= \", s_action)\n    #output = sess.run(output_action)\n<\/code><\/pre>\n\n<p>here is the output:<\/p>\n\n<pre><code>python random_vec.py\n2018-10-31 09:46:15.028376: I tensorflow\/core\/platform\/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\ns_ph =  Tensor(\"stochastic:0\", shape=(), dtype=bool, device=\/device:CPU:0)\ndet_action=  4\nrand_action=  11\nchose_rand=  False\ns_action=  1\n<\/code><\/pre>\n","answers":[{"AnswerId":"53084507","CreationDate":"2018-10-31T13:27:46.243","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>This is because you are evaluating again in a new sess.run. \nSince you are generating a random number for deterministic_action, the result turns out to be the next random number after 4, which is 1.\nHere is the result of your code, when I extract the value of deterministic_action as well in the last step.<\/p>\n\n<p>Modification:<\/p>\n\n<pre><code>print (\"s_ph = \", stochastic_ph)\nd_action = sess.run(deterministic_action)\nprint (\"det_action= \", d_action)\nr_action = sess.run(random_action)\nprint (\"rand_action= \", r_action)\ne = sess.run(eps)\nc_action = sess.run(chose_random)\nprint (\"chose_rand= \", c_action)\ns_action, d_action = sess.run([stochastic_action, deterministic_action])\nprint (\"s_action= \", s_action)\nprint (\"det_action= \", d_action)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>s_ph =  Tensor(\"stochastic:0\", shape=(), dtype=bool, device=\/device:CPU:0)\ndet_action=  4\nrand_action=  11\nchose_rand=  False\ns_action=  1\ndet_action=  1\n<\/code><\/pre>\n\n<p>Now all you need to do is run everything in one sess.run<\/p>\n\n<pre><code>d_action, r_action, e,  c_action, s_action = sess.run([deterministic_action, random_action, eps, chose_random, stochastic_action])\nprint (\"det_action= \", d_action)\nprint (\"rand_action= \", r_action)\nprint (\"chose_rand= \", c_action)\nprint (\"s_action= \", s_action)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>s_ph =  Tensor(\"stochastic:0\", shape=(), dtype=bool, device=\/device:CPU:0)\ndet_action=  4\nrand_action=  11\nchose_rand=  False\ns_action=  4\n<\/code><\/pre>\n\n<p>Update:<\/p>\n\n<p>I was not clear on why the random_uniform generates different values when seed is set. This is because the code is running with the same session object that it initialized the variables with.\nModifying the code with a new session object, this is what happens:<\/p>\n\n<pre><code>print (\"s_ph = \", stochastic_ph)\nd_action = sess.run(deterministic_action)\nprint (\"det_action= \", d_action)\nsess.close()\nsess = tf.Session()\nsess.run(init, feed_dict={stochastic_ph: True})\ns_action = sess.run(stochastic_action)\nprint (\"s_action= \", s_action)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>s_ph =  Tensor(\"stochastic:0\", shape=(), dtype=bool, device=\/device:CPU:0)\ndet_action=  4\ns_action=  4\n<\/code><\/pre>\n"}]}
+{"QuestionId":53079511,"AnswerCount":3,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-10-31T08:51:13.703","AcceptedAnswerId":53080422.0,"OwnerUserId":236106.0,"Title":"How to correctly set the value of placeholder in tensorflow?","Body":"<p>I've written the following quick program in tensorflow to print fibonacci numbers. Initial fib numbers are initialized as placeholders <code>x1<\/code> and <code>x2<\/code> However when I try to feed the value of the placeholders in <code>session.run<\/code> it results in the error : <\/p>\n\n<p><code>InvalidArgumentError: You must feed a value for placeholder tensor 'x2' with dtype int64 and shape [1]<\/code><\/p>\n\n<p>Can you help me understand and fix the problem with my code? <\/p>\n\n<pre><code>import tensorflow as tf\nsession = tf.InteractiveSession()\nimport numpy as np\n\nones = tf.ones((1,))\n\nn = 10\nx1 = tf.placeholder(tf.int64, [1], name='x1')\nx2 = tf.placeholder(tf.int64, [1], name='x2')\ntemp = tf.Variable(ones, name='temp')\n\ntf.initialize_all_variables().run()\nfib_nums = [x1, x2]\nfor i in range(100):\n  temp = x1 + x2\n  x1 = x2\n  x2 = temp\n  fib_nums.append(temp)\n\nseries = tf.stack(fib_nums)\nprint(np.ones(1).astype(np.int64))\nsession.run(series, feed_dict={x1:np.ones(1).astype(np.int64), x2:np.ones(1).astype(np.int64)})\nprint(series.eval())\n<\/code><\/pre>\n","answers":[{"AnswerId":"53080422","CreationDate":"2018-10-31T09:43:31.463","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>There are a couple of errors here. First, since you are reusing the Python names <code>x1<\/code> and <code>x2<\/code>, when you give them in the <code>feed_dict<\/code> they no longer refer to the placeholders, but to the last results of the loop. So you should change your code so the keys that you give in <code>feed_dict<\/code> are truly the placeholders. Second, you first call <code>session.run<\/code> with the <code>feed_dict<\/code>, which is correct, but then you call <code>series.eval()<\/code>, which is essentially the same as the previous line, only you are not providing the <code>feed_dict<\/code> in this case so it is not going to work. You do not really need to call <code>series.eval()<\/code>, you can just take the value returned by <code>session.run<\/code>. Your fixed program could look something like this:<\/p>\n\n<pre><code>import tensorflow as tf\nsession = tf.InteractiveSession()\nimport numpy as np\n\nones = tf.ones((1,))\n\nn = 10\n# Reserve these Python names for the placeholders\nx1_ph = tf.placeholder(tf.int64, [1], name='x1')\nx2_ph = tf.placeholder(tf.int64, [1], name='x2')\ntemp = tf.Variable(ones, name='temp')\n\ntf.initialize_all_variables().run()\n# Use placeholders as initial values in the iterations\nx1, x2 = x1_ph, x2_ph\nfib_nums = [x1, x2]\nfor i in range(100):\n  temp = x1 + x2\n  x1 = x2\n  x2 = temp\n  fib_nums.append(temp)\n\nseries = tf.stack(fib_nums)\nprint(np.ones(1).astype(np.int64))\n# You can just give lists as inputs and TensorFlow will convert their type\nseries_val = sess.run(series, feed_dict={x1_ph: [1], x2_ph: [1]})\nprint(series_val)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[1]\n[[                   1]\n [                   1]\n [                   2]\n [                   3]\n [                   5]\n [                   8]\n [                  13]\n [                  21]\n [                  34]\n [                  55]\n [                  89]\n [                 144]\n [                 233]\n [                 377]\n [                 610]\n [                 987]\n [                1597]\n [                2584]\n...\n<\/code><\/pre>\n"},{"AnswerId":"53080337","CreationDate":"2018-10-31T09:38:38.540","ParentId":null,"OwnerUserId":"648946","Title":null,"Body":"<p>Just change <code>tf.placeholder<\/code> to a tensor and it will work<\/p>\n\n<pre><code>x1 = tf.ones(dtype=tf.int64, shape=1, name='x1') \nx2 = tf.ones(dtype=tf.int64, shape=1, name='x2') \n<\/code><\/pre>\n"},{"AnswerId":"53084044","CreationDate":"2018-10-31T13:04:39.863","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<p>I think another variation is this using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/identity\" rel=\"nofollow noreferrer\">tf.identity<\/a>.<\/p>\n\n<pre><code>y = tf.identity(x1)\ny1 = tf.identity(x2)\nfib_nums = [y, y1]\nfor i in range(100):\n  temp = y + y1\n  y = y1\n  y1 = temp\n  fib_nums.append(temp)\n<\/code><\/pre>\n\n<p>This placeholder modification issue is also discussed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9960\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>And this is one more way to get the series.<\/p>\n\n<pre><code>def cond(i, x_next, x_prev):\n    return x_next &lt;= 100\n\ndef body( i, x_next, x_prev ):\n\n    nextinseries = x_next + x_prev\n\n    next = tf.Print(x_next, [x_next], message=\"Next is : \")\n    prev = tf.Print(x_prev, [x_prev], message=\"Previous is : \")\n\n    with tf.control_dependencies([nextinseries]):\n        prev = tf.identity( next )\n\n\n    return [i + 1, nextinseries , prev ]\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\n\nsess.run(tf.while_loop(cond, body, [1, 1, 1]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53079901,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><machine-learning>","CreationDate":"2018-10-31T09:14:27.970","AcceptedAnswerId":null,"OwnerUserId":10584801.0,"Title":"How to build an image processing model with new dataset using machine learning (Tensorflow)?","Body":"<p>I want to extract the data shown on an instrument panel cluster of a bike using image processing and ML (mostly Tensorflow). The sample of such cluster is <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/IJFDg.jpg\" alt=\"here\">.<\/p>\n\n<p>My input will be such complete image and I want to get the data such as speed shown by Speedometer, status of Neutral indicator(ON\/OFF), Odometer reading etc from the cluster. Can anyone tell me what is the complete process for doing so? Or where are tutorials for the same? I am new to ML and there are lot of tutorials on the net but I dont know which one of them is useful for me.<\/p>\n","answers":[{"AnswerId":"53082445","CreationDate":"2018-10-31T11:37:17.953","ParentId":null,"OwnerUserId":"7904315","Title":null,"Body":"<p>There are a few things to consider here. You will first want to process the image a little before feeding it into a neural network for instance. Maybe crop out the section which contains the speedometer reading and perform some kind of OCR (Optical Character Recognition) on it for the speed reading.<\/p>\n\n<p>The neutral indicator is a little easier actually, since there are only two states you can just look at a few sample images yourself and see which pixels actually change colors and write a quick check yourself, maybe something like so<\/p>\n\n<pre><code>(ON if pixel at (x,y) is green else OFF)  \n<\/code><\/pre>\n\n<p>Do the same for the odometer, just identify the area where the reading is, crop it out and perform OCR on the smaller image. So now really you're just looking for OCR tutorials using TensorFlow, if that's a constraint.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>you can do all of this cropping and editing the image programatically in python using a library called cv2 which you can install with pip<\/p>\n\n<pre><code> pip install opencv-python\n<\/code><\/pre>\n"}]}
+{"QuestionId":53080032,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-10-31T09:21:53.920","AcceptedAnswerId":null,"OwnerUserId":1310104.0,"Title":"Tensorflow 1.11 estimator + dataset training speed low and irregular","Body":"<p><strong>Problem<\/strong><\/p>\n\n<ul>\n<li>training a custom tensorflow 1.11 <code>tf.estimator.Estimator<\/code> with a <code>tf.data.Dataset<\/code> runs much more slowly than when using <code>tf.keras<\/code> with the same model architecture and feeding the data directly<\/li>\n<li>However, it does run quickly at times (in terms of global_step\/sec), but is slow around epoch start and end.<\/li>\n<li>during the \"fast\" batches, GPU util is around 30%. During the slow ones, around 1%<\/li>\n<\/ul>\n\n<p><strong>Probable causes<\/strong><\/p>\n\n<ul>\n<li>my input pipeline isn't well-optimized enough so that GPU is idle, waiting on the CPU data processing. But I don't understand why that's only in the last few minibatches before episode end<\/li>\n<\/ul>\n\n<p><strong>What have I tried<\/strong><\/p>\n\n<p>Followed the steps in the <a href=\"https:\/\/www.tensorflow.org\/performance\/datasets_performance\" rel=\"nofollow noreferrer\">Input Pipeline Performance Guide<\/a>. That sped up the batches that are not near epoch boundaries. I'm not sure how to improve further.<\/p>\n\n<p><strong>Minimal working example<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport tensorflow as tf\n\ntrain_data = pd.DataFrame(np.random.randn(1030255, 1021)).rename(columns={c:str(c) for c in range(1021)})\ntrain_target = pd.DataFrame(np.round(np.random.rand(1030255, 15))).rename(columns={c:str(c) for c in range(15)})\nval_data = pd.DataFrame(np.random.randn(491077, 1021)).rename(columns={c:str(c) for c in range(1021)})\nval_target = pd.DataFrame(np.round(np.random.rand(491077, 15))).rename(columns={c:str(c) for c in range(15)})\n\ndef model_fn(features, labels, mode,params): \n    net = tf.feature_column.input_layer(features, params['feature_columns'])\n    net = tf.layers.dense(net, units=params['hidden_units'], activation=tf.nn.tanh)\n    logits = tf.layers.dense(net, params['outputs'], activation=None)\n    loss = tf.losses.sigmoid_cross_entropy(labels, logits=logits)\n    if mode == tf.estimator.ModeKeys.EVAL:\n        return tf.estimator.EstimatorSpec(mode, loss=loss)\n    elif mode == tf.estimator.ModeKeys.TRAIN:\n        optimizer = tf.train.AdamOptimizer()\n        train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n        return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)\n\ncols = [str(c) for c in np.sort(np.random.choice(1021, size=98, replace=False))]\ntarget_cols = [str(c) for c in np.arange(11)]\nVALIDATION_BATCH_SIZE = int(val_data.shape[0] \/ 4.0)\nBATCH_SIZE = 2**10\n\ndef train_input_fn():\n    features = {k: train_data[k].values for k in cols}\n    dataset = tf.data.Dataset.from_tensor_slices((features, train_target[target_cols].values))\n    dataset = dataset.repeat() \\\n        .shuffle(train_data.shape[0]) \\\n        .batch(BATCH_SIZE) \\\n        .prefetch(BATCH_SIZE)\n    return dataset\n\ndef validation_input_fn():\n    features = {k: val_data[k].values for k in cols}\n    dataset = tf.data.Dataset.from_tensor_slices((features, val_target[target_cols].values))\n    dataset = dataset.repeat() \\\n        .batch(VALIDATION_BATCH_SIZE) \\\n        .prefetch(VALIDATION_BATCH_SIZE)\n    return dataset\n\nfeature_columns = [tf.feature_column.numeric_column(f) for f in cols]\nrun_cfg = tf.estimator.RunConfig(tf_random_seed=1, save_checkpoints_steps=1000, save_checkpoints_secs=None)\nclassifier = tf.estimator.Estimator(\n        model_fn=model_fn,\n        config=run_cfg,\n        params={\n            'feature_columns': feature_columns,\n            'hidden_units': 128,\n            'outputs': len(target_cols)\n        })\ntf.estimator.train_and_evaluate(\n    classifier,\n    train_spec=tf.estimator.TrainSpec(train_input_fn),\n    eval_spec=tf.estimator.EvalSpec(validation_input_fn, steps=4, start_delay_secs=30, throttle_secs=30))\n<\/code><\/pre>\n\n<p><strong>Sample output<\/strong><\/p>\n\n<p>(The time the first step takes is misrepresented as being huge since tensorflow takes the difference with the last training step. However, that is not the case for the last training step before starting eval.)<\/p>\n\n<pre><code>INFO:tensorflow:Finished evaluation at 2018-10-31-08:17:35\nINFO:tensorflow:Saving dict for global step 1000: global_step = 1000, loss = 0.694723\nINFO:tensorflow:Saving 'checkpoint_path' summary for global step 1000: \/tmp\/tmpvQq9DV\/model.ckpt-1000\nINFO:tensorflow:global_step\/sec: 0.930553\nINFO:tensorflow:loss = 0.69407356, step = 1000 (107.463 sec)\nINFO:tensorflow:global_step\/sec: 171.673\nINFO:tensorflow:loss = 0.69456786, step = 1100 (0.583 sec)\nINFO:tensorflow:global_step\/sec: 166.964\nINFO:tensorflow:loss = 0.69411445, step = 1200 (0.599 sec)\nINFO:tensorflow:global_step\/sec: 172.226\nINFO:tensorflow:loss = 0.6940959, step = 1300 (0.579 sec)\nINFO:tensorflow:global_step\/sec: 170.882\nINFO:tensorflow:loss = 0.69440323, step = 1400 (0.586 sec)\nINFO:tensorflow:global_step\/sec: 173.453\nINFO:tensorflow:loss = 0.69332886, step = 1500 (0.577 sec)\nINFO:tensorflow:global_step\/sec: 167.078\nINFO:tensorflow:loss = 0.6950055, step = 1600 (0.598 sec)\nINFO:tensorflow:global_step\/sec: 159.763\nINFO:tensorflow:loss = 0.69460225, step = 1700 (0.626 sec)\nINFO:tensorflow:global_step\/sec: 161.674\nINFO:tensorflow:loss = 0.6940766, step = 1800 (0.617 sec)\nINFO:tensorflow:global_step\/sec: 8.83793\nINFO:tensorflow:loss = 0.6936994, step = 1900 (11.315 sec)\nINFO:tensorflow:Saving checkpoints for 2000 into \/tmp\/tmpvQq9DV\/model.ckpt.\nINFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Starting evaluation at 2018-10-31-08:18:11\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from \/tmp\/tmpvQq9DV\/model.ckpt-2000\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Evaluation [1\/4]\nINFO:tensorflow:Evaluation [2\/4]\nINFO:tensorflow:Evaluation [3\/4]\nINFO:tensorflow:Evaluation [4\/4]\nINFO:tensorflow:Finished evaluation at 2018-10-31-08:19:36\n<\/code><\/pre>\n\n<p><strong>Without eval step<\/strong><\/p>\n\n<p>I thought the eval step (which in itself is freakishly slow) could be the issue. But when I only train, it runs even more slowly:<\/p>\n\n<pre><code>classifier.train(input_fn=train_input_fn, steps=5000)\n\n[...]\nINFO:tensorflow:Saving checkpoints for 1000 into \/tmp\/tmpIJctNp\/model.ckpt.\nINFO:tensorflow:global_step\/sec: 6.95414\nINFO:tensorflow:loss = 0.6948092, step = 1000 (14.381 sec)\nINFO:tensorflow:global_step\/sec: 12.0491\nINFO:tensorflow:loss = 0.6942879, step = 1100 (8.298 sec)\nINFO:tensorflow:global_step\/sec: 8.98388\nINFO:tensorflow:loss = 0.6939402, step = 1200 (11.131 sec)\nINFO:tensorflow:global_step\/sec: 8.86219\nINFO:tensorflow:loss = 0.6946343, step = 1300 (11.284 sec)\nINFO:tensorflow:global_step\/sec: 8.74248\nINFO:tensorflow:loss = 0.694865, step = 1400 (11.439 sec)\n<\/code><\/pre>\n\n<p>Thanks in advance for any help.<\/p>\n","answers":[]}
+{"QuestionId":53080215,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-10-31T09:32:20.360","AcceptedAnswerId":null,"OwnerUserId":7866526.0,"Title":"what is the difference between tf.layers.dense() and tf.contrib.layers.fully_connected()?","Body":"<p>I searched for this question on stackoverlflow but none the <a href=\"https:\/\/stackoverflow.com\/questions\/44912297\/are-tf-layers-dense-and-tf-contrib-layers-fully-connected-interchangeable\">answers<\/a> clarified my issue. <\/p>\n\n<p>I am trying to do a uni-variate forecasting and I have manually written a  dense layer at the end of my LSTM. <\/p>\n\n<pre><code> weight = tf.Variable(tf.truncated_normal([config.lstm_size, config.input_size]))\n bias = tf.Variable(tf.constant(0.1, shape=[config.input_size]))\n\n prediction = tf.matmul(last, weight) + bias\n<\/code><\/pre>\n\n<p>Then I tried to add an activation to the result.<\/p>\n\n<pre><code>weight = tf.Variable(tf.truncated_normal([config.lstm_size, config.input_size]))\nbias = tf.Variable(tf.constant(0.1, shape=[config.input_size]))\n\nprediction = tf.nn.tanh(tf.matmul(last, weight) + bias)\n<\/code><\/pre>\n\n<p><strong>Question:<\/strong> Is this the same as adding a tf.layers.dense() or a  tf.contrib.layers.fully_connected()?<\/p>\n\n<pre><code>hidden = tf.layers.dense(last, units=1, activation=tf.nn.relu)\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>tf.contrib.layers.fully_connected(last, num_outputs=1, activation_fn=tf.nn.relu)\n<\/code><\/pre>\n\n<p><strong>Question:<\/strong> If I do something like:<\/p>\n\n<pre><code>hidden = tf.layers.dense(last, units=1, activation=tf.nn.relu)\n\nprediction = tf.contrib.layers.fully_connected(hidden, num_outputs=1, activation_fn=tf.nn.relu)\n<\/code><\/pre>\n\n<p>Does that mean I have two dense layers?<\/p>\n\n<p>Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":53080289,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><activation-function>","CreationDate":"2018-10-31T09:35:38.990","AcceptedAnswerId":null,"OwnerUserId":1115237.0,"Title":"Keras - Nan in summary histogram LSTM","Body":"<p>I've written an LSTM model using Keras, and using LeakyReLU advance activation:<\/p>\n\n<pre><code>    # ADAM Optimizer with learning rate decay\n    opt = optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)\n\n    # build the model\n    model = Sequential()\n\n    num_features = data.shape[2]\n    num_samples = data.shape[1]\n\n    model.add(\n        LSTM(16, batch_input_shape=(None, num_samples, num_features), return_sequences=True, activation='linear'))\n    model.add(LeakyReLU(alpha=.001))\n    model.add(Dropout(0.1))\n    model.add(LSTM(8, return_sequences=True, activation='linear'))\n    model.add(Dropout(0.1))\n    model.add(LeakyReLU(alpha=.001))\n    model.add(Flatten())\n    model.add(Dense(1, activation='sigmoid'))\n\n    model.compile(loss='binary_crossentropy', optimizer=opt,\n                  metrics=['accuracy', keras_metrics.precision(), keras_metrics.recall(), f1])\n<\/code><\/pre>\n\n<p>My data is a balanced binary labeled set. i.e: 50% labeled 1 50% labeled 0. I've used <code>activation='linear'<\/code> for the LSTM layers preceding the LeakyReLU activation, similar to <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/117\" rel=\"nofollow noreferrer\">this example<\/a> I found on GitHub.<\/p>\n\n<p>The model throws <code>Nan in summary histogram<\/code> error in that configuration. Changing the LSTM activations to <code>activation='sigmoid'<\/code> works well, but seems like the wrong thing to do.<\/p>\n\n<p>Reading <a href=\"https:\/\/stackoverflow.com\/questions\/39854390\/nan-in-summary-histogram\">this StackOverflow<\/a> question suggested <em>\"introducing a small value when computing the loss\"<\/em>, I'm just not sure how to do it on a built-in loss function.<\/p>\n\n<p>Any help\/explanation would be appreciated.<\/p>\n\n<p><strong>Update:<\/strong>\nI can see that the loss is nan on the first epoch<\/p>\n\n<pre><code>260\/260 [==============================] - 6s 23ms\/step - \nloss: nan - acc: 0.5000 - precision: 0.5217 - recall: 0.6512 - f1: nan - val_loss: nan - val_acc: 0.0000e+00 - val_precision: -2147483648.0000 - val_recall: -49941480.1860 - val_f1: nan\n<\/code><\/pre>\n\n<p><strong>Update 2<\/strong>\nI've upgraded both TensorFlow &amp; Keras to versions 1.12.0 &amp; 2.2.4 . There was no effect.<\/p>\n\n<p>I also tried adding a loss to the first LSTM layer as suggested by @Oluwafemi Sule, it looks like a step in the right direction, now the loss is not nan on the first epoch, however, I still get the same error ... probably because of other nan values, like the val_loss \/ val_f1.<\/p>\n\n<pre><code>[==============================] - 7s 26ms\/step - \nloss: 1.9099 - acc: 0.5077 - precision: 0.5235 - recall: 0.6544 - f1: 0.5817 - val_loss: nan - val_acc: 0.5172 - val_precision: 35.0000 - val_recall: 0.9722 - val_f1: nan\n<\/code><\/pre>\n\n<p><strong>Update 3<\/strong>\nI tried to compile the network with just the accuracy metric, with no success:<\/p>\n\n<pre><code>Epoch 1\/300\n260\/260 [==============================] - 8s 29ms\/step - loss: nan - acc: 0.5538 - val_loss: nan - val_acc: 0.0000e+00\n<\/code><\/pre>\n","answers":[{"AnswerId":"53181649","CreationDate":"2018-11-06T23:31:45.917","ParentId":null,"OwnerUserId":"5189811","Title":null,"Body":"<p>This answers starts from the suggestion to <em>introduce a small value when computing the loss<\/em>.<\/p>\n\n<p><code>keras.layers.LSTM<\/code> as with all layers that are direct or indirect subclasses of <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/2.2.4\/keras\/engine\/base_layer.py#L934\" rel=\"nofollow noreferrer\"><code>keras.engine.base_layer.Layer<\/code><\/a> has a <code>add_loss<\/code> method that can be used to set a starting value for the loss.<\/p>\n\n<p>I suggest to do this for the LSTM layer and see if it makes any difference for your results.<\/p>\n\n<pre><code>lstm_layer = LSTM(8, return_sequences=True, activation='linear')\nlstm_layer.add_loss(1.0)\n\nmodel.add(lstm_layer)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53080929,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-10-31T10:11:36.923","AcceptedAnswerId":null,"OwnerUserId":9134476.0,"Title":"Train, Save and Load a Tensorflow Model","Body":"<p>Refer to <a href=\"https:\/\/github.com\/FelixMohr\/Deep-learning-with-Python\/blob\/master\/DCGAN-MNIST.ipynb\" rel=\"nofollow noreferrer\">this<\/a> to train a GAN model for MNIST dataset, I want to save a model and restore it for further prediction. After having some understanding of <a href=\"https:\/\/cv-tricks.com\/tensorflow-tutorial\/save-restore-tensorflow-models-quick-complete-tutorial\/\" rel=\"nofollow noreferrer\">Saving and Importing a Tensorflow Model<\/a> I am able to save and restore some variables of inputs and outputs but for <a href=\"https:\/\/github.com\/FelixMohr\/Deep-learning-with-Python\/blob\/master\/DCGAN-MNIST.ipynb\" rel=\"nofollow noreferrer\">this<\/a> network I am able to save the model only after some specific iterations and not able to predict some output. <\/p>\n","answers":[{"AnswerId":"53081792","CreationDate":"2018-10-31T10:59:27.780","ParentId":null,"OwnerUserId":"3863319","Title":null,"Body":"<p>Did you refer to <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">this<\/a> guide? It explains very clearly how to load and save tensorflow models in all possible formats.<\/p>\n"},{"AnswerId":"53081993","CreationDate":"2018-10-31T11:10:25.547","ParentId":null,"OwnerUserId":"5400620","Title":null,"Body":"<p>If you are new to ML, I'd recommend you give Keras a try first, which is much easier to use. See <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model<\/a>, pretty much you can use:<\/p>\n\n<pre><code>model.save('my_model.h5')\n<\/code><\/pre>\n\n<p>to save your model to disk.<\/p>\n\n<pre><code>model = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>to load your model and make prediction<\/p>\n"}]}
+{"QuestionId":53081134,"AnswerCount":0,"Tags":"<python><machine-learning><deep-learning><pytorch><autoencoder>","CreationDate":"2018-10-31T10:23:22.237","AcceptedAnswerId":null,"OwnerUserId":6677037.0,"Title":"how to use MaxUnpool1d from pytorch library in convolution autoencoder?","Body":"<p>I want to build a Convolution AutoEncoder using Pytorch library in python.\nAll of the examples have no MaxUnpool1d.<\/p>\n\n<p>The architecture I want to build should be like:<\/p>\n\n<pre><code>class autoencoder(nn.Module):\n    def __init__(self):\n            self.encoder = nn.Sequential(\n                                 nn.Conv1d(in_channels=1, out_channels=64 ,kernel_size=16 , stride=3),\n                                 nn.ReLU(True),\n                                 nn.MaxPool1d(4),\n                                 nn.Conv1d(64, 128 ,8, 3),\n                                 nn.ReLU(True),\n                                 nn.MaxPool1d(4), \n                                 )\n\n            self.decoder = nn.Sequential(\n                                 nn.MaxUnpool1d(4),\n                                 nn.ConvTranspose1d(in_channels =128, out_channels =64 , kernel_size =8, stride= 3),\n                                 nn.ReLU(True),\n                                 nn.MaxUnpool1d(4),\n                                 nn.ConvTranspose1d(64, 1 ,16, 3),\n                                 nn.Tanh()\n                                 )\n\n    def forward(self, x):\n        x = self.encoder(x)\n        x = self.decoder(x)\n        return x\n<\/code><\/pre>\n\n<p>To make it work I should provide indices for the MaxUnpool1d layer.\nHow do I use it for the current architecture and also making the input size equals to the output?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53081187,"AnswerCount":0,"Tags":"<api><tensorflow><dataset><pipeline>","CreationDate":"2018-10-31T10:26:30.560","AcceptedAnswerId":null,"OwnerUserId":10584824.0,"Title":"Got a mismatch result between writing and reading tfrecord files","Body":"<p>Here I use this func to write multiple tfrecord files:<\/p>\n\n<pre><code>writer = tf.python_io.TFRecordWriter(save)\nfor pth, lb in tqdm(zip(piece_p, piece_l)):\n    # mind that the path should be read into image data first\n    # to convert the byteslist data format into raw bytes\n    data = Image.open(pth)\n    if resize is not None:\n        data.thumbnail(resize, Image.ANTIALIAS)\n    features = tf.train.Features(feature={\n        'image': tf.train.Feature(\n            bytes_list=tf.train.BytesList(value=[data.tobytes()])),\n        'label': tf.train.Feature(\n            int64_list=tf.train.Int64List(value=[lb]))\n    })\n    example = tf.train.Example(features=features)\n\n    # serialize the constructed data format before writing step\n    writer.write(example.SerializeToString())\n    sys.stdout.flush()\nwriter.close()\n<\/code><\/pre>\n\n<p>And parse the binary file using code as below:<\/p>\n\n<pre><code>def parse_fn(serialized):\n    features = {\n        'image': tf.FixedLenFeature([], tf.string),\n        'label': tf.FixedLenFeature([], tf.int64)\n    }\n    parse_exp = tf.parse_single_example(serialized=serialized,\n                                    features=features)\n    labels = parse_exp['label']\n    data = parse_exp['image']\n    data = tf.decode_raw(data, tf.uint8)\n    data = tf.cast(data, tf.float32)\n    del parse_exp\n    return data, labels\n\ndataset = tf.data.Dataset.list_files(data_list, shuffle=True)\ndataset = dataset.interleave(tf.data.TFRecordDataset,\n                             cycle_length=file_num)\n# dataset = tf.data.TFRecordDataset(data_list[0])\n\ndataset = dataset.map(parse_fn, num_parallel_calls=4)\n<\/code><\/pre>\n\n<p>But why is the number of labels and data always mismatching...?\nevery time when adding the following code to make multiple batches or sth<\/p>\n\n<pre><code>dataset = dataset.batch(12)\ndataset = dataset.repeat(1)\niterator = dataset.make_initializable_iterator()\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsess.run(iterator.initializer)\ndata, labels = iterator.get_next()\n<\/code><\/pre>\n\n<p>and the labels quantity always remains half of data. are there something wrong with my arguments setting? I pretty sure that there are no wrong with my saving part and reading part separately... but there are some problems when combining them together.<\/p>\n","answers":[]}
+{"QuestionId":53081706,"AnswerCount":0,"Tags":"<python><tensorflow><queue>","CreationDate":"2018-10-31T10:55:58.727","AcceptedAnswerId":null,"OwnerUserId":1650630.0,"Title":"How to process data stream with Tensorflow and FIFOQueue","Body":"<p>I am trying to process data that arrives continuously in a neural network. To optimize the entire process, I have:<\/p>\n\n<ul>\n<li>A thread that is responsible for filling the FIFOQueue.<\/li>\n<li>Another thread that should extract as many elements as possible from the queue up to a threshold. That is: min(queue.size(), MAX_BATCH_SIZE). MAX_BATCH_SIZE is the maximum that I am capable of processing at the same time.<\/li>\n<\/ul>\n\n<p>I have managed to make it work, but it is something rudimentary and I wonder if there is a better implementation of this method. I focus on the consumer thread of the queue.<\/p>\n\n<p><strong>Definition of operations:<\/strong><\/p>\n\n<pre><code>self.queue = tf.FIFOQueue ...\nself.input_dequeue_op = self.queue.dequeue_many(\n                                tf.maximum(\n                                    tf.constant(1), \n                                    tf.minimum(\n                                        tf.constant(MAX_BATCH_SIZE), \n                                        self.queue.size()\n                                    )\n                                )\n                            )\n<\/code><\/pre>\n\n<p><strong>Consumer Thread:<\/strong><\/p>\n\n<pre><code>with network.sess.as_default():\n    while True:\n\n            results = network.sess.run(self.input_dequeue_op)\n<\/code><\/pre>\n\n<p>Basically what I intend is to send to a neural network the maximum number of elements that are currently in the queue up to MAX_BATCH_SIZE (which is the maximum that can be processed at the same time).<\/p>\n\n<p>The operation obtains the maximum real batch that I can load.\nIt seems to me a very rudimentary form, I have also tried to use it directly:<\/p>\n\n<pre><code>self.input_dequeue_op = self.queue.dequeue_up_to(tf.constant(MAX_BATCH_SIZE))\n<\/code><\/pre>\n\n<p>But the queue is blocked indefinitely even if elements are added.<\/p>\n\n<p>Is there another better way to do this?<\/p>\n","answers":[]}
+{"QuestionId":53082485,"AnswerCount":2,"Tags":"<python><tensorflow><keras><resuming-training>","CreationDate":"2018-10-31T11:39:28.387","AcceptedAnswerId":null,"OwnerUserId":4841832.0,"Title":"Keras - manage history","Body":"<p>I am training Keras models, saving them with <code>model.save()<\/code> and than later loading them and resuming training.<\/p>\n\n<p>I would like to plot after each training the whole training history, but <code>model.fit_generator()<\/code> only returns the history of the last session of training.<\/p>\n\n<p>I can save the history of the initial session and update it myself, but I wonder if there is a standard way in Keras of managing the training history.<\/p>\n\n<pre><code>history1 = model.fit_generator(my_gen)\nplot_history(history1)\nmodel.save('my_model.h5')\n\n# Some days afterwards...\n\nmodel = load_model('my_model.h5')\nhistory2 = model.fit_generator(my_gen)\n\n# here I would like to reconstruct the full_training history\n# including the info from history1 and history2\nfull_history = ???\n<\/code><\/pre>\n","answers":[{"AnswerId":"53082649","CreationDate":"2018-10-31T11:48:41.053","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Let's say this line<\/p>\n\n<pre><code>print(history.history.keys())\n<\/code><\/pre>\n\n<p>produces the following output:<\/p>\n\n<pre><code>['acc', 'loss', 'val_acc', 'val_loss']\n<\/code><\/pre>\n\n<p>Based on the assumption that the loaded model should have the same performance as saved model, you can try something like concatenating histories. For example, concatenate new accuracy history on the loaded accuracy history of the loaded model. <\/p>\n\n<p>It should start from the same point in the plotting space where the loaded model ended (maybe you will have to add (+) epochs of the previously trained model for the plot so that new values of accuracy don't start from epoch 0, but loaded models' last epoch).<\/p>\n\n<p>I hope that you understand my idea and that it will help you :)<\/p>\n"},{"AnswerId":"53085925","CreationDate":"2018-10-31T14:37:56.167","ParentId":null,"OwnerUserId":"4841832","Title":null,"Body":"<p>Turns out there is no standard way of doing this in Keras yet AFAIK.<\/p>\n\n<p>See <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6697\" rel=\"nofollow noreferrer\">this issue<\/a> for context.<\/p>\n"}]}
+{"QuestionId":53082519,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-10-31T11:41:44.530","AcceptedAnswerId":53084871.0,"OwnerUserId":9540764.0,"Title":"Problem with tf.train.Saver() and GPU - TensorFlow","Body":"<p>My code is structured as follows:<\/p>\n\n<pre><code>with tf.device('\/gpu:1'):\n...\nmodel = get_model(input_pl)\n...\n    with tf.Session() as sess:\n        saver = tf.train.Saver()\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(num_epochs):\n           ...\n           for n in range(num_batches):\n              ...\n              sess.run(...)\n           # eval epoch\n        saver.save(sess, ...)\n<\/code><\/pre>\n\n<p>I want to save the model after the training phase. When I run it gives me this error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'save\/SaveV2': Could not satisfy explicit device specification '\/device:GPU:1' because no supported kernel for GPU devices is available.\n<\/code><\/pre>\n\n<p>Reading <a href=\"https:\/\/stackoverflow.com\/questions\/38173580\/tensorflow-crash-using-tf-train-saver-with-gpu\">this question<\/a> I changed the code in this way:<\/p>\n\n<pre><code>saver = tf.train.Saver()\nwith tf.device('\/gpu:1'):\n...\nmodel = get_model(pointcloud_pl)\n...\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(num_epochs):\n           ...\n           for n in range(num_batches):\n              ...\n              sess.run(...)\n           # eval epoch\n        saver.save(sess, ...)\n<\/code><\/pre>\n\n<p>But now I get this error:<\/p>\n\n<pre><code>ValueError: No variables to save\n<\/code><\/pre>\n\n<p>I've tried also to do this way:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    saver = tf.train.Saver()\n    ...\n    with tf.device('\/gpu:1'):\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(num_epochs):\n        ...\n            for n in range(num_batches):\n               ...\n               sess.run()\n            # eval epoch\n        saver.save(sess, ...)\n<\/code><\/pre>\n\n<p>And I still get the same error. The error is always in the <code>saver = tf.train.Saver()<\/code> line.<\/p>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"53084871","CreationDate":"2018-10-31T13:44:51.000","ParentId":null,"OwnerUserId":"9540764","Title":null,"Body":"<p>Solved doing this:<\/p>\n\n<ol>\n<li><code>tf.Session()<\/code><\/li>\n<li>model<\/li>\n<li><code>saver = tf.train.Saver()<\/code><\/li>\n<li><code>with tf.device():<\/code><\/li>\n<\/ol>\n\n<p>Here an example code<\/p>\n\n<pre><code>with tf.Session() as sess:\n    ...\n    model = get_model(input_pl)\n    saver = tf.train.Saver()\n    ...\n    with tf.device('\/gpu:1'):\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(num_epochs):\n        ...\n            for n in range(num_batches):\n               ...\n               sess.run()\n            # eval epoch\n        saver.save(sess, ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53082740,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><google-colaboratory>","CreationDate":"2018-10-31T11:53:24.873","AcceptedAnswerId":null,"OwnerUserId":8625204.0,"Title":"ERROR:tensorflow:Couldn't understand architecture name '.50_%image_size%'","Body":"<p>Hello every one i tried to build imageclassifer based on TensorFlow for Poets i use colab for running tensorflow but when i run the below code<\/p>\n\n<pre><code>    IMAGE_SIZE=224\n    ARCHITECTURE=\"mobilenet_0.50_${IMAGE_SIZE}\"\n    !python -m scripts.retrain \\\n      --bottleneck_dir=tf_files\/bottlenecks \\\n      --how_many_training_steps=500 \\\n      --model_dir=tf_files\/models\/ \\\n      --summaries_dir=tf_files\/training_summaries\/\"${ARCHITECTURE}\" \\\n      --output_graph=tf_files\/retrained_graph.pb \\\n      --output_labels=tf_files\/retrained_labels.txt \\\n      --architecture=\"${ARCHITECTURE}\" \\\n      --image_dir=tf_files\/flower_photos\n<\/code><\/pre>\n\n<p>I get the following error<\/p>\n\n<pre><code>ERROR:tensorflow:Couldn't understand architecture name '.50_%image_size%'\nTraceback (most recent call last):\n  File \"\/usr\/lib\/python3.6\/runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/usr\/lib\/python3.6\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/content\/tensorflow-for-poets-2\/scripts\/retrain.py\", line 1326, in &lt;module&gt;\n    tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"\/content\/tensorflow-for-poets-2\/scripts\/retrain.py\", line 976, in main\n    model_info = create_model_info(FLAGS.architecture)\n  File \"\/content\/tensorflow-for-poets-2\/scripts\/retrain.py\", line 923, in create_model_info\n    raise ValueError('Unknown architecture', architecture)\nValueError: ('Unknown architecture', '.50_%image_size%')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53090689","CreationDate":"2018-10-31T19:29:22.483","ParentId":null,"OwnerUserId":"1227961","Title":null,"Body":"<p>There are two errors in the code cell. First you need to set the ARC\u0397\u0399TECTURE variable to <code>f\"mobilenet_0.50_${IMAGE_SIZE}<\/code>. And second you should not use <code>$<\/code> in the shell command.\nHere is the code with the corrections.<\/p>\n\n<pre><code>IMAGE_SIZE=224\nARCHITECTURE=f\"mobilenet_0.50_${IMAGE_SIZE}\"\n!python -m scripts.retrain \\\n  --bottleneck_dir=tf_files\/bottlenecks \\\n  --how_many_training_steps=500 \\\n  --model_dir=tf_files\/models\/ \\\n  --summaries_dir=tf_files\/training_summaries\/\"{ARCHITECTURE}\" \\\n  --output_graph=tf_files\/retrained_graph.pb \\\n  --output_labels=tf_files\/retrained_labels.txt \\\n  --architecture=\"{ARCHITECTURE}\" \\\n  --image_dir=tf_files\/flower_photos\n<\/code><\/pre>\n"}]}
+{"QuestionId":53083312,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-31T12:26:37.050","AcceptedAnswerId":null,"OwnerUserId":10585780.0,"Title":"ValueError: ('Some keys in session_kwargs are not supported at this time: %s', dict_keys(['class_mode'])","Body":"<pre><code>from keras.models import Sequential  \nfrom keras.layers.core import Dense, Dropout, Activation  \nfrom keras.optimizers import SGD  \nfrom keras.datasets import mnist  \nimport numpy \n\nmodel = Sequential()\n\nmodel.add(Dense(500,input_shape=(784,))) # 28*28=784  \nmodel.add(Activation('tanh')) # tanh  \nmodel.add(Dropout(0.5)) # 50% dropout\n\nmodel.add(Dense(500)) # 500\u4e2a\nmodel.add(Activation('tanh'))  \nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(10))\nmodel.add(Activation('softmax')) \n\n\nsgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True) \n\nmodel.compile(loss='categorical_crossentropy', optimizer=sgd, class_mode='categorical') \n\n(X_train, y_train), (X_test, y_test) = mnist.load_data() \n\nX_train = X_train.reshape(X_train.shape[0], X_train.shape[1] * X_train.shape[2]) \nX_test = X_test.reshape(X_test.shape[0], X_test.shape[1] * X_test.shape[2])  \nY_train = (numpy.arange(10) == y_train[:, None]).astype(int) \nY_test = (numpy.arange(10) == y_test[:, None]).astype(int)\n\nmodel.fit(X_train,Y_train,batch_size=200,epochs=50,shuffle=True,verbose=0,validation_split=0.3)\nmodel.evaluate(X_test, Y_test, batch_size=200, verbose=0)\n\nprint(\"test set\")\nscores = model.evaluate(X_test,Y_test,batch_size=200,verbose=0)\nprint(\"\")\nprint(\"The test loss is %f\" % scores)\nresult = model.predict(X_test,batch_size=200,verbose=0) \n<\/code><\/pre>\n\n<p>I found this post Error when profiling keras models, which modifies the tensorflow library.<\/p>\n\n<p>So, I checked Keras library code from the link. But could not find anything like ['class_mode'] to modify the keras library. Next, I tried running the code after re-installing keras, but even that didn't work.<\/p>\n\n<p>I used anaconda to import Kreas\uff0c maybe I install wrong?<\/p>\n\n<p>Can anyone suggest a solution for this?<\/p>\n","answers":[{"AnswerId":"53774193","CreationDate":"2018-12-14T05:56:39.917","ParentId":null,"OwnerUserId":"8220456","Title":null,"Body":"<p>remove class_mode='categorical',it runs<\/p>\n"}]}
+{"QuestionId":53083872,"AnswerCount":1,"Tags":"<keras><rnn>","CreationDate":"2018-10-31T12:55:47.823","AcceptedAnswerId":53085682.0,"OwnerUserId":9747668.0,"Title":"confused about the concept of timesteps and output shape in keras.layers.RNN","Body":"<blockquote>\n  <p>keras.layers.RNN<\/p>\n  \n  <p>Input shape\n  3D tensor with shape (batch_size, timesteps, input_dim).<\/p>\n  \n  <p>Output shape<\/p>\n  \n  <p>if return_state: a list of tensors. The first tensor is the output. The remaining tensors are the last states, each with shape  (batch_size, units).<\/p>\n  \n  <p>if return_sequences: 3D tensor with shape  (batch_size, timesteps, units).\n  else, 2D tensor with shape (batch_size, units).<\/p>\n<\/blockquote>\n\n<p>1.I am comfused about the concept of timesteps.<\/p>\n\n<p>2.I am confused about the process that how  it deals with the input of three axes.<\/p>\n\n<h1>Simplified code<\/h1>\n\n<pre><code>import keras\nfrom keras.applications.inception_resnet_v2 import InceptionResNetV2\nfrom keras.applications.mobilenet import MobileNet\nfrom keras.applications.vgg19 import VGG19\nfrom keras.applications.densenet import DenseNet\nfrom keras.preprocessing import image\nfrom keras.engine import Layer\nfrom keras.applications.inception_resnet_v2 import preprocess_input\nfrom keras.layers import Conv2D, UpSampling2D, InputLayer, Conv2DTranspose, Input, Reshape, merge, concatenate\nfrom keras.layers import Activation, Dense, Dropout, Flatten\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.callbacks import TensorBoard \nfrom keras.models import Sequential, Model\nfrom keras.layers.core import RepeatVector, Permute\nfrom keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img\nfrom skimage.color import rgb2lab, lab2rgb, rgb2gray, gray2rgb\nfrom skimage.transform import resize\nfrom skimage.io import imsave\nimport numpy as np\nimport os\nimport random\nimport tensorflow as tf\nimport keras.backend as K\nfrom keras.layers.recurrent import GRU\nfrom keras.layers.merge import add\n\nencoder_input = Input(shape=(32, 32, 1))\n\nrnn_size = 16\nencoder_output = Conv2D(16, (3,3), activation='relu', padding='same')(encoder_input)\nsequence_output = Reshape(target_shape=(32, 512))(encoder_output)  \ngru_1 = GRU(rnn_size, return_sequences=False,kernel_initializer='he_normal', name='gru1')(sequence_output)\ngru_1b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru1_b')(sequence_output)\ngru1_merged = add([gru_1, gru_1b])\ngru_2 = GRU(rnn_size, return_sequences=True,kernel_initializer='he_normal', name='gru2')(gru1_merged)\ngru_2b = GRU(rnn_size, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', name='gru2_b')(gru1_merged)\nsequence_output = concatenate([gru_2, gru_2b])\nsequence_output = K.expand_dims(sequence_output, 3)\nfusion_output = concatenate([encoder_output,sequence_output ], axis=3) \n\nmodel = Model(inputs=encoder_input, outputs=fusion_output)\nmodel.summary()\n<\/code><\/pre>\n\n<h1>output error<\/h1>\n\n<blockquote>\n  <p>ValueError: A <code>Concatenate<\/code> layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 32, 32, 16), (None, None, 32, 1)]<\/p>\n<\/blockquote>\n\n<p>I thought the shape of 'sequence_output' would be (None, 32, 32, 1).But it was (None, None, 32, 1).I didn't know what went wrong so I started to doubt my understanding about RNN.<\/p>\n\n<h1>What I have done<\/h1>\n\n<h1>1.output gru1 and gru_1b<\/h1>\n\n<p>after annotation:<\/p>\n\n<p>'#sequence_output = K.expand_dims(sequence_output, 3)'<\/p>\n\n<p>'#fusion_output = concatenate([encoder_output,sequence_output ], axis=3)'\nThen I got ()<a href=\"https:\/\/i.stack.imgur.com\/UckkM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UckkM.png\" alt=\"output\"><\/a><\/p>\n\n<h3>I am quite confused about the shape of gru1 and gru_1b. Why they are different?<\/h3>\n\n<h1>2.I set return_sequences and return_state True and then got<\/h1>\n\n<p>valueerror:\n<a href=\"https:\/\/i.stack.imgur.com\/2O2Yx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2O2Yx.png\" alt=\" \"><\/a>\nIn fact I have no idea what to do next.<\/p>\n","answers":[{"AnswerId":"53085682","CreationDate":"2018-10-31T14:25:33.387","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>This question is more suitable for Cross Validated forum but ok. <\/p>\n\n<p>So to answer your first question, timesteps is just a number that denotes the length of a sequence. RNNs work in a specific manner because they have a recurrent connection back to itself. Very good explanation of RNNs is given <a href=\"http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\" rel=\"nofollow noreferrer\">here<\/a> on the example of LSTM. There you can see the difference between <code>cell state<\/code> and <code>state<\/code>.<\/p>\n\n<p>To answer the second question, <code>(batch_size, timesteps, units)<\/code> is a output dimension. <code>timesteps<\/code> is again, dimensionality of the input sequence and it's something you should standardize and define on input shape (all sequences on the input have to be of the same length, if they're not, you should pad them to a specified length). <code>units<\/code> is a dimensionality of the output, it's output of every RNN cell you have in your layer. <\/p>\n\n<p>The whole point of those <code>return_state<\/code> and <code>return_sequences<\/code> arguments is what you need for the next layer and for your computations. The first one is there to return cell state as the first element of the output. The second one is the states after every timestep. So after every word (or sequence element) is read, RNN updates the <code>state<\/code> based on the read input element and <code>cell state<\/code>. So, with <code>return_sequences<\/code> you can get sequence after every word is processed inside the RNN and the <code>state<\/code> has been updated.<\/p>\n\n<p>I think that everything will be much clearer after you read that blog post I linked in this answer. I hope this is helpful :)<\/p>\n"}]}
+{"QuestionId":53084056,"AnswerCount":0,"Tags":"<python><keras><multiprocessing><gpu><virtual-memory>","CreationDate":"2018-10-31T13:05:21.947","AcceptedAnswerId":null,"OwnerUserId":9841270.0,"Title":"Keras fit generator slow","Body":"<p>Keras <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\"><code>fit_generator<\/code><\/a> is very slow. The GPU is not used constantly in training, sometimes it's usage drops to 0%. Even on 4 workers and <code>multiproceesing=True<\/code>. <\/p>\n\n<p>Also, the processes of the script are requesting too much virtual memory and are with a D status, <a href=\"https:\/\/www.linuxquestions.org\/questions\/linux-general-1\/process-status-d-44989\/%7D\" rel=\"nofollow noreferrer\"><em>uninterruptible sleep (usually IO)<\/em><\/a>.<\/p>\n\n<p>I already tried different combinations of <code>max_queue_size<\/code> but it didn't work.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HArOL.png\" rel=\"nofollow noreferrer\">Screenshot<\/a> GPU Usage<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XRlvP.png\" rel=\"nofollow noreferrer\">Screenshot<\/a> of Processes Virtual Memory and Status<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OBtRA.png\" rel=\"nofollow noreferrer\">Hardware Info<\/a>\nGPU = Titan XP 12Gb<\/p>\n\n<p>Code of Data Generator Class<\/p>\n\n<pre><code>import numpy as np\nimport keras\nimport conf\n\n\nclass DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, list_IDs, labels, batch_size=32, dim=(conf.max_file, 128),\n                 n_classes=10, shuffle=True):\n        'Initialization'\n        self.dim = dim\n        self.batch_size = batch_size\n        self.labels = labels\n        self.list_IDs = list_IDs\n        self.n_classes = n_classes\n        self.shuffle = shuffle\n        self.on_epoch_end()\n\n    def __len__(self):\n        'Denotes the number of batches per epoch'\n        return int(np.floor(len(self.list_IDs) \/ self.batch_size))\n\n    def __getitem__(self, index):\n        'Generate one batch of data'\n        # Generate indexes of the batch\n        indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]\n\n        # Find list of IDs\n        list_IDs_temp = [self.list_IDs[k] for k in indexes]\n\n        # Generate data\n        X, y = self.__data_generation(list_IDs_temp)\n\n        return X, y\n\n    def on_epoch_end(self):\n        'Updates indexes after each epoch'\n        self.indexes = np.arange(len(self.list_IDs))\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n\n    def __data_generation(self, list_IDs_temp):\n        'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n        # Initialization\n        X = np.empty((self.batch_size, *self.dim))\n        y = np.empty((self.batch_size, conf.max_file, self.n_classes))\n        # Generate data\n        for i, ID in enumerate(list_IDs_temp):\n            # Store sample\n            X[i, ] = np.load(conf.dir_out_data+\"data_by_file\/\" + ID)\n\n            # Store class\n            y[i, ] = np.load(conf.dir_out_data +\n                            'data_by_file\/' + self.labels[ID])\n\n        return X, y\n<\/code><\/pre>\n\n<p>Code of python script<\/p>\n\n<pre><code>training_generator = DataGenerator(partition['train'], labels, **params)\n\nvalidation_generator = DataGenerator(partition['validation'], labels, **params)\n\nmodel.fit_generator(generator = training_generator,\n                    validation_data = validation_generator,\n                    epochs=steps,\n                    callbacks=[tensorboard, checkpoint],\n                    workers=4,\n                    use_multiprocessing=True,\n                    max_queue_size=50)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53084384,"AnswerCount":1,"Tags":"<java><caching><invalidation><caffeine>","CreationDate":"2018-10-31T13:21:17.057","AcceptedAnswerId":53150933.0,"OwnerUserId":2779330.0,"Title":"Caffeine cache - many keys to single value","Body":"<p>I have Caffeine cache with Key->Value mapping. There are multiple implementations of <code>Key<\/code> interface with different <code>equals<\/code> methods. In order to delete value from cache based on <code>someOtherVal<\/code>, I had to use code like <code>cache.asMap().keySet().removeIf(comp::isSame)<\/code> which is super slow.<\/p>\n\n<p>Is there any other solution for this kind of many keys to single value mapping in cache? One thing that comes to my mind is to have 2 Cache instances, one with <code>Cache&lt;Key, String&gt;<\/code> and other with <code>Cache&lt;someOtherVal, Key&gt;<\/code>, and whenever I want to delete a value I locate Key using this other cache.<\/p>\n\n<p>Then only question is how to keep this 2 caches in sync? Are there already solutions for this?<\/p>\n\n<pre><code>import java.time.Duration;\nimport java.util.Objects;\nimport java.util.UUID;\n\nimport com.github.benmanes.caffeine.cache.Cache;\nimport com.github.benmanes.caffeine.cache.Caffeine;\nimport com.google.common.base.Stopwatch;\n\npublic class Removal {\n    private static final int MAX = 1_000_000;\n\n    interface Key{\n        String getSomeOtherVal();\n        default boolean isSame(Key k){\n            return Objects.equals(k.getSomeOtherVal(),getSomeOtherVal());\n        }\n    }\n\n    static class KeyImpl implements Key{\n        int id;\n        String someOtherVal;\n\n        public KeyImpl(int id, String someOtherVal) {\n            this.id = id;\n            this.someOtherVal = someOtherVal;\n        }\n\n        public int getId() {\n            return id;\n        }\n\n        @Override\n        public String getSomeOtherVal() {\n            return someOtherVal;\n        }\n\n        @Override\n        public boolean equals(Object o) {\n            if (this == o)\n                return true;\n            if (o == null || getClass() != o.getClass())\n                return false;\n            KeyImpl key = (KeyImpl)o;\n            return id == key.id;\n        }\n\n        @Override\n        public int hashCode() {\n            return Objects.hash(id);\n        }\n    }\n\n    Cache&lt;Key, String&gt; cache = Caffeine.newBuilder().build();\n\n    public static void main(String[] args) {\n        Removal s = new Removal();\n        s.fill();\n        Duration sRem = s.slowRemovalFirst100();\n        Duration fRem = s.fastRemoval100To200();\n        System.out.println(\"Slow removal in \" + sRem);\n        System.out.println(\"Fast removal in \" + fRem);\n    }\n\n    private Duration slowRemovalFirst100(){\n        Stopwatch sw = Stopwatch.createStarted();\n        for(int i=0; i&lt;100; i++){\n            Key comp = new KeyImpl(i, String.valueOf(i));\n            cache.asMap().keySet().removeIf(comp::isSame);  \/\/Finds a key by some other property and then removes it (SLOW)\n            \/\/System.out.println(\"Removed \" + i);\n        }\n        return sw.stop().elapsed();\n    }\n\n    private Duration fastRemoval100To200(){\n        Stopwatch sw = Stopwatch.createStarted();\n        for(int i=100; i&lt;200; i++){\n            Key comp = new KeyImpl(i, String.valueOf(i));\n            cache.invalidate(comp); \/\/Uses direct access to map by key (FAST)\n            \/\/System.out.println(\"Removed \" + i);\n        }\n        return sw.stop().elapsed();\n    }\n\n    private void fill(){\n        for(int i=0; i&lt;MAX; i++){\n            cache.put(new KeyImpl(i, String.valueOf(i)), UUID.randomUUID().toString());\n        }\n    }\n}\n<\/code><\/pre>\n\n<p>Result of running this code on my machine: <\/p>\n\n<pre><code>Slow removal in PT2.807105177S\nFast removal in PT0.000126183S\n<\/code><\/pre>\n\n<p>where you can see such a big difference...<\/p>\n","answers":[{"AnswerId":"53150933","CreationDate":"2018-11-05T08:42:50.157","ParentId":null,"OwnerUserId":"2779330","Title":null,"Body":"<p>Ok, I managed to solve this:<\/p>\n\n<pre><code>public class IndexedCache&lt;K,V&gt; implements Cache&lt;K,V&gt; {\n\n    @Delegate\n    private Cache&lt;K, V&gt; cache;\n    private Map&lt;Class&lt;?&gt;, Map&lt;Object, Set&lt;K&gt;&gt;&gt; indexes;\n\n    private IndexedCache(Builder&lt;K, V&gt; bldr){\n        this.indexes = bldr.indexes;\n        cache = bldr.caf.build();\n    }\n\n    public &lt;R&gt; void invalidateAllWithIndex(Class&lt;R&gt; clazz, R value) {\n        cache.invalidateAll(indexes.get(clazz).getOrDefault(value, new HashSet&lt;&gt;()));\n    }\n\n    public static class Builder&lt;K, V&gt;{\n        Map&lt;Class&lt;?&gt;, Function&lt;K, ?&gt;&gt; functions = new HashMap&lt;&gt;();\n        Map&lt;Class&lt;?&gt;, Map&lt;Object, Set&lt;K&gt;&gt;&gt; indexes = new ConcurrentHashMap&lt;&gt;();\n        Caffeine&lt;K,V&gt; caf;\n\n        public &lt;R&gt; Builder&lt;K,V&gt; withIndex(Class&lt;R&gt; clazz, Function&lt;K, R&gt; function){\n            functions.put(clazz, function);\n            indexes.put(clazz, new ConcurrentHashMap&lt;&gt;());\n            return this;\n        }\n\n        public IndexedCache&lt;K, V&gt; buildFromCaffeine(Caffeine&lt;Object, Object&gt; caffeine) {\n            caf = caffeine.writer(new CacheWriter&lt;K, V&gt;() {\n\n                @Override\n                public void write( K k, V v) {\n                    for(Map.Entry&lt;Class&lt;?&gt;, Map&lt;Object, Set&lt;K&gt;&gt;&gt; indexesEntry : indexes.entrySet()){\n                        indexesEntry.getValue().computeIfAbsent(functions.get(indexesEntry.getKey()).apply(k), (ky)-&gt; new HashSet&lt;&gt;())\n                        .add(k);\n                    }\n                }\n\n                @Override\n                public void delete( K k,  V v,  RemovalCause removalCause) {\n                    for(Map.Entry&lt;Class&lt;?&gt;, Map&lt;Object, Set&lt;K&gt;&gt;&gt; indexesEntry : indexes.entrySet()){\n                        indexesEntry.getValue().remove(functions.get(indexesEntry.getKey()).apply(k));\n                    }\n                }\n            });\n            return new IndexedCache&lt;&gt;(this);\n        }\n    }\n\n}\n<\/code><\/pre>\n\n<p>and this is use-case:<\/p>\n\n<pre><code>@AllArgsConstructor\n    @Data\n    @EqualsAndHashCode(onlyExplicitlyIncluded = true)\n    static class CompositeKey{\n        @EqualsAndHashCode.Include\n        Integer k1;\n        String k2;\n        Long k3;\n    }\n\n\n    public static void main(String[] args) {\n\n\n        Caffeine&lt;Object, Object&gt; cfein = Caffeine.newBuilder().softValues().maximumSize(200_000);\n\n        IndexedCache&lt;CompositeKey, String&gt; cache = new IndexedCache.Builder&lt;CompositeKey, String&gt;()\n                .withIndex(Long.class, ck -&gt; ck.getK3())\n                .withIndex(String.class, ck -&gt; ck.getK2())\n                .buildFromCaffeine(cfein);\n\n\n        for(int i=0; i&lt;100; i++){\n            cache.put(new CompositeKey(i, String.valueOf(i), Long.valueOf(i)), \"sdfsdf\");\n        }\n\n\n        for(int i=0; i&lt;10; i++){\n            \/\/use equals method of CompositeKey to do equals comp.\n            cache.invalidate(new CompositeKey(i, String.valueOf(i), Long.valueOf(i)));\n        }\n\n        for(int i=10; i&lt;20; i++){\n            \/\/use Long index\n            cache.invalidateAllWithIndex(Long.class, Long.valueOf(i));\n        }\n\n        for(int i=20; i&lt;30; i++){\n            \/\/use String index\n            cache.invalidateAllWithIndex(String.class, String.valueOf(i));\n        }\n\n\n        int y = 4;\n\n    }\n<\/code><\/pre>\n\n<p>here is link to discussion I had: <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/issues\/279\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ben-manes\/caffeine\/issues\/279<\/a><\/p>\n"}]}
+{"QuestionId":53084604,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-10-31T13:32:23.450","AcceptedAnswerId":53084852.0,"OwnerUserId":6286493.0,"Title":"Keras - NN prediction of error in metrological measurement.","Body":"<p>I have example data which contain coordinates of points on the x-y plane (for \nexample 2.0000 , 4.0000), next using monte carlo method, a small random error \nis added to those coordinates to simulate a set of points measured by a \nmetrological machine.<\/p>\n\n<p>This may sound trivial, but Im not really sure what to do next with this data, Im trying to build a model which predicts the error in measurement, but I have problem in visualizing the whole concept, ie. should the input layer of the network have neurons which receive real coordinates of the points and \nsimulated coordinates, or simulated ones only? Or maybe I should estimate \nmeasurement error for each simulated point and and use it with coordinates of \nthose points in the input layer? Also, how many neurons the network should \nhave on the output layer and how should I interpret that data? I know this \nprobably isnt the best description of the problem, but I am a complete \nbegginner in this field, so any theoretical help or practical examples will be greatly appreciated. <\/p>\n","answers":[{"AnswerId":"53084852","CreationDate":"2018-10-31T13:43:38.250","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I'm not even sure if this is possible except the points are always integers in which case you can do this without neural networks. But, nonetheless, here are the answers to your question:<\/p>\n\n<p>You should put the simulated coordinates only to the input of the model.<\/p>\n\n<p>The output should have 2 neurons, 1 will output the estimated error of dimension <code>x<\/code>, and the other of dimension <code>y<\/code>.<\/p>\n\n<p>The process of training is as following: you put the simulated points to the network input layer, let the network predict the error for <code>x<\/code> and <code>y<\/code> axis and then you compare the predicted results with the real (correct) ones. If the predicted ones are correct, you move on to the next sample (pair of coordinates) and if they are not the same, you update weights with backpropagation and SGD. You repeat this process for desired number of epochs, depending on your data (fine-tune number of epochs so it's not too low and not too high).<\/p>\n\n<p>I hope this clears up things for you :)<\/p>\n"}]}
+{"QuestionId":53084743,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2018-10-31T13:38:03.020","AcceptedAnswerId":53085790.0,"OwnerUserId":6329284.0,"Title":"Modifying (keras\/tensorflow) Tensors using numpy methods","Body":"<p>I want to perform a specific operation. Namely, from a matrix:<\/p>\n\n<pre><code>A = np.array([[1,2],\n            [3,4]])\n<\/code><\/pre>\n\n<p>To the following<\/p>\n\n<pre><code>B = np.array([[1, 0, 0, 2, 0, 0],\n              [0, 1, 0, 0, 2, 0],\n              [0, 0, 1, 0, 0, 2],\n              [3, 0, 0, 4, 0, 0],\n              [0, 3, 0, 0, 4, 0],\n              [0, 0, 3, 0, 0, 4]])\n<\/code><\/pre>\n\n<p>Or in words: multiply every entry by the identity matrix and keep the same order.\nNow I have accomplished this by using numpy, using the following code. Here <code>N<\/code> and <code>M<\/code> are the dimensions of the starting matrix, and the dimension of the identity matrix.<\/p>\n\n<pre><code>l_slice = 3\nn_slice = 2\nA = np.reshape(np.arange(1, 1+N ** 2), (N, N))\nB = np.array([i * np.eye(M) for i in A.flatten()])\nC = B.reshape(N, N, M, M).reshape(N, N * M, M).transpose([0, 2, 1]).reshape((N * M, N * M))\n<\/code><\/pre>\n\n<p>where <code>C<\/code> has my desired properties.<\/p>\n\n<p>But now I want do this modification in Keras\/Tensorflow, where the matrix <code>A<\/code> is the outcome of one of my layers.<\/p>\n\n<p>However, I am not sure yet if I will be able to properly create matrix B. Especially when batches are involved, I think I will somehow mess up the dimensions of my problem.\nCan anyone with more Keras\/Tensorflow experience comment on this 'reshape' and how he\/she sees this happening within Keras\/Tensorflow?<\/p>\n","answers":[{"AnswerId":"53085790","CreationDate":"2018-10-31T14:30:41.890","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a way to do that with TensorFlow:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndata = tf.placeholder(tf.float32, [None, None])\nn = tf.placeholder(tf.int32, [])\neye = tf.eye(n)\nmult = data[:, tf.newaxis, :, tf.newaxis] * eye[tf.newaxis, :, tf.newaxis, :]\nresult = tf.reshape(mult, n * tf.shape(data))\nwith tf.Session() as sess:\n    a = sess.run(result, feed_dict={data: [[1, 2], [3, 4]], n: 3})\n    print(a)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[1. 0. 0. 2. 0. 0.]\n [0. 1. 0. 0. 2. 0.]\n [0. 0. 1. 0. 0. 2.]\n [3. 0. 0. 4. 0. 0.]\n [0. 3. 0. 0. 4. 0.]\n [0. 0. 3. 0. 0. 4.]]\n<\/code><\/pre>\n\n<p>By the way, you can do basically the same in NumPy, which should be faster than your current solution:<\/p>\n\n<pre><code>import numpy as np\n\ndata = np.array([[1, 2], [3, 4]])\nn = 3\neye = np.eye(n)\nmult = data[:, np.newaxis, :, np.newaxis] * eye[np.newaxis, :, np.newaxis, :]\nresult = np.reshape(mult, (n * data.shape[0], n * data.shape[1]))\nprint(result)\n# The output is the same as above\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>I'll try to give some intuition about why\/how this works, sorry if it's too long. It is not that hard but I think it's sort of tricky to explain. Maybe it is easier to see how the following multiplication works<\/p>\n\n<pre><code>import numpy as np\n\ndata = np.array([[1, 2], [3, 4]])\nn = 3\neye = np.eye(n)\nmult1 = data[:, :, np.newaxis, np.newaxis] * eye[np.newaxis, np.newaxis, :, :]\n<\/code><\/pre>\n\n<p>Now, <code>mult1<\/code> is a sort of \"matrix of matrices\". If I give two indices, I will get the diagonal matrix for the corresponding element in the original one:<\/p>\n\n<pre><code>print(mult1[0, 0])\n# [[1. 0. 0.]\n#  [0. 1. 0.]\n#  [0. 0. 1.]]\n<\/code><\/pre>\n\n<p>So you could say this matrix could be visualize like this:<\/p>\n\n<pre><code>| 1 0 0 |  | 2 0 0 |\n| 0 1 0 |  | 0 2 0 |\n| 0 0 1 |  | 0 0 2 |\n\n| 3 0 0 |  | 4 0 0 |\n| 0 3 0 |  | 0 4 0 |\n| 0 0 3 |  | 0 0 4 |\n<\/code><\/pre>\n\n<p>However this is deceiving, because if you try to reshape this to the final shape the result is not the right one:<\/p>\n\n<pre><code>print(np.reshape(mult1, (n * data.shape[0], n * data.shape[1])))\n# [[1. 0. 0. 0. 1. 0.]\n#  [0. 0. 1. 2. 0. 0.]\n#  [0. 2. 0. 0. 0. 2.]\n#  [3. 0. 0. 0. 3. 0.]\n#  [0. 0. 3. 4. 0. 0.]\n#  [0. 4. 0. 0. 0. 4.]]\n<\/code><\/pre>\n\n<p>The reason is that reshaping (conceptually) \"flattens\" the array first and then gives the new shape. But the flattened array in this case is not what you need:<\/p>\n\n<pre><code>print(mult1.ravel())\n# [1. 0. 0. 0. 1. 0. 0. 0. 1. 2. 0. 0. 0. 2. 0. ...\n<\/code><\/pre>\n\n<p>You see, it first traverses the first submatrix, then the second, etc. What you want though is for it to traverse first the first row of the first submatrix, then the first row of the second submatrix, then second row of first submatrix, etc. So basically you want something like:<\/p>\n\n<ul>\n<li>Take the first two submatrices (the ones with <code>1<\/code> and <code>2<\/code>)\n\n<ul>\n<li>Take all the first rows (<code>[1, 0, 0]<\/code> and <code>[2, 0, 0]<\/code>).\n\n<ul>\n<li>Take the first of these (<code>[1, 0, 0]<\/code>)\n\n<ul>\n<li>Take each of its elements (<code>1<\/code>, <code>0<\/code> and <code>0<\/code>).<\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>And then continue for the rest. So if you think about it, we traversing first the axis 0 (row of \"matrix of matrices\"), then 2 (rows of each submatrix), then 1 (column of \"matrix of matrices\") and finally 3 (columns of submatrices). So we can just reorder the axis to do that:<\/p>\n\n<pre><code>mult2 = mult1.transpose((0, 2, 1, 3))\nprint(np.reshape(mult2, (n * data.shape[0], n * data.shape[1])))\n# [[1. 0. 0. 2. 0. 0.]\n#  [0. 1. 0. 0. 2. 0.]\n#  [0. 0. 1. 0. 0. 2.]\n#  [3. 0. 0. 4. 0. 0.]\n#  [0. 3. 0. 0. 4. 0.]\n#  [0. 0. 3. 0. 0. 4.]]\n<\/code><\/pre>\n\n<p>And it works! So in the solution I posted, to avoid the tranposing, I just make the multiplication so the order of the axes is exactly that:<\/p>\n\n<pre><code>mult = data[\n        :,           # Matrix-of-matrices rows\n        np.newaxis,  # Submatrix rows\n        :,           # Matrix-of-matrices columns\n        np.newaxis   # Submatrix columns\n    ] * eye[\n        np.newaxis,  # Matrix-of-matrices rows\n        :,           # Submatrix rows\n        np.newaxis,  # Matrix-of-matrices columns\n        :            # Submatrix columns\n    ]\n<\/code><\/pre>\n\n<p>I hope that makes it slightly clearer. To be honest, in this case in particular I could came up with the solution quickly because I had to solve a similar problem not too long ago, and I guess you end up building an intuition of these things.<\/p>\n"},{"AnswerId":"53085842","CreationDate":"2018-10-31T14:33:21.247","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>Another way to achieve the same effect in numpy is to use the following: <\/p>\n\n<pre><code>A = np.array([[1,2],\n            [3,4]])\nB = np.repeat(np.repeat(A, 3, axis=0), 3, axis=1) * np.tile(np.eye(3), (2,2))\n<\/code><\/pre>\n\n<p>Then, to replicate it in tensorflow, we can use <code>tf.tile<\/code>, but there is no <code>tf.repeat<\/code>, however someone has provided this function on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8246\" rel=\"nofollow noreferrer\">tensorflow tracker<\/a>. <\/p>\n\n<pre><code>def tf_repeat(tensor, repeats):\n    \"\"\"\n    Args:\n\n    input: A Tensor. 1-D or higher.\n    repeats: A list. Number of repeat for each dimension, length must be the same as the number of dimensions in input\n\n    Returns:\n\n    A Tensor. Has the same type as input. Has the shape of tensor.shape * repeats\n    \"\"\"\n    with tf.variable_scope(\"repeat\"):\n        expanded_tensor = tf.expand_dims(tensor, -1)\n        multiples = [1] + list(repeats)\n        tiled_tensor = tf.tile(expanded_tensor, multiples=multiples)\n        repeated_tesnor = tf.reshape(tiled_tensor, tf.shape(tensor) * repeats)\n    return repeated_tesnor\n<\/code><\/pre>\n\n<p>and thus the <strong>tensorflow<\/strong> implementation will look like the following. Here I also consider that the first dimension represents batches, and thus we do not operate on it.<\/p>\n\n<pre><code>N = 2\nM = 3\nnbatch = 2\nAin = np.reshape(np.arange(1, 1 + N*N*nbatch), (nbatch, N, N))\n\nA = tf.placeholder(tf.float32, shape=(nbatch, N, N))\nB = tf.tile(tf.eye(M), [N, N]) * tf_repeat(A, [1, M, M])\n\nwith tf.Session() as sess:\n    print(sess.run(C, feed_dict={A: Ain}))\n<\/code><\/pre>\n\n<p>and the result:<\/p>\n\n<pre><code> [[[1. 0. 0. 2. 0. 0.]\n  [0. 1. 0. 0. 2. 0.]\n  [0. 0. 1. 0. 0. 2.]\n  [3. 0. 0. 4. 0. 0.]\n  [0. 3. 0. 0. 4. 0.]\n  [0. 0. 3. 0. 0. 4.]]\n\n [[5. 0. 0. 6. 0. 0.]\n  [0. 5. 0. 0. 6. 0.]\n  [0. 0. 5. 0. 0. 6.]\n  [7. 0. 0. 8. 0. 0.]\n  [0. 7. 0. 0. 8. 0.]\n  [0. 0. 7. 0. 0. 8.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53085007,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2018-10-31T13:51:03.943","AcceptedAnswerId":null,"OwnerUserId":9255615.0,"Title":"Re-train a frozen *.pb model in TensorFlow","Body":"<blockquote>\n  <p>How do I import a frozen protobuf to enable it for re-training? <\/p>\n<\/blockquote>\n\n<p>All the methods i've found online expect checkpoints. Is there a way to read a protobuf such that kernel and bias constants are converted to variables?<\/p>\n\n<hr>\n\n<p><strong>Edit 1<\/strong>:\nThis is similar to the following question: <a href=\"https:\/\/stackoverflow.com\/questions\/51379506\/how-to-retrain-model-in-graph-pb\">How to retrain model in graph (.pb)?<\/a><\/p>\n\n<p>I looked at DeepSpeech, which was recommended in the answers to that question. They seem to have <a href=\"https:\/\/github.com\/mozilla\/DeepSpeech\/commit\/46d1cece4f98d726d88579500734a95686cfe16b#diff-79c8372f97f7f5a6d8660bd265f836d7\" rel=\"noreferrer\">removed support<\/a> for <code>initialize_from_frozen_model<\/code>. I couldn't find the reason.<\/p>\n\n<hr>\n\n<p><strong>Edit 2<\/strong>: I tried creating a new GraphDef object where I replace the kernels and biases with Variables:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>probable_variables = [...] # kernels and biases of Conv2D and MatMul\n\nnew_graph_def = tf.GraphDef()\n\nwith tf.Session(graph=graph) as sess:\n    for n in sess.graph_def.node:\n\n        if n.name in probable_variables:\n            # create variable op\n            nn = new_graph_def.node.add()\n            nn.name = n.name\n            nn.op = 'VariableV2'\n            nn.attr['dtype'].CopyFrom(attr_value_pb2.AttrValue(type=dtype))\n            nn.attr['shape'].CopyFrom(attr_value_pb2.AttrValue(shape=shape))\n\n        else:\n            nn = new_model.node.add()\n            nn.CopyFrom(n)\n<\/code><\/pre>\n\n<p>Not sure if I am on the right path. Don't know how to set <code>trainable=True<\/code> in a <code>NodeDef<\/code> object.<\/p>\n","answers":[{"AnswerId":"58707201","CreationDate":"2019-11-05T08:25:47.530","ParentId":null,"OwnerUserId":"12324493","Title":null,"Body":"<p>Thanks for @FalconUA and @Max wu.\nI added a way to quickly get the variables' names.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>import tensorflow as tf\r\n\r\n\r\n# Load protobuf as graph, given filepath\r\ndef load_pb(path_to_pb):\r\n    with tf.gfile.GFile(path_to_pb, 'rb') as f:\r\n        graph_def = tf.GraphDef()\r\n        graph_def.ParseFromString(f.read())\r\n    with tf.Graph().as_default() as graph:\r\n        tf.import_graph_def(graph_def, name='')\r\n        return graph\r\n\r\n\r\ntf_graph = load_pb('mobilenet_v1_1.0_224_frozen_ccl.pb')\r\nvariables = [op.name for op in tf_graph.get_operations() if op.type == \"Const\"]\r\nprint(variables)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"},{"AnswerId":"57561445","CreationDate":"2019-08-19T17:29:33.687","ParentId":null,"OwnerUserId":"10620814","Title":null,"Body":"<p>I have verified @FalconUA's solution with tested code. Slight modifications were needed (notably, I use the <code>initializer<\/code> option in <code>get_variable<\/code> to properly initialize the Variables). Here it is!<\/p>\n\n<p>Assuming your frozen model is stored in <code>frozen_graph.pb<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>probable_variables = [...] # kernels and biases of Conv2D and MatMul\ntf_graph = load_pb('frozen_graph.pb')\n\nconst_var_name_pairs = []\nwith tf_graph.as_default() as g:\n\n    for name in probable_variables:\n        tensor = g.get_tensor_by_name('{}:0'.format(name))\n        with tf.Session() as sess:\n            tensor_as_numpy_array = sess.run(tensor)\n        var_shape = tensor.get_shape()\n        # Give each variable a name that doesn't already exist in the graph\n        var_name = '{}_turned_var'.format(name)\n        # Create TensorFlow variable initialized by values of original const.\n        var = tf.get_variable(name=var_name, dtype='float32', shape=var_shape, \\  \n                      initializer=tf.constant_initializer(tensor_as_numpy_array))\n        # We want to keep track of our variables names for later.\n        const_var_name_pairs.append((name, var_name))\n\n    # At this point, we added a bunch of tf.Variables to the graph, but they're\n    # not connected to anything.\n\n    # The magic: we use TF Graph Editor to swap the Constant nodes' outputs with\n    # the outputs of our newly created Variables.\n\n    for const_name, var_name in const_var_name_pairs:\n        const_op = g.get_operation_by_name(const_name)\n        var_reader_op = g.get_operation_by_name(var_name + '\/read')\n        ge.swap_outputs(ge.sgv(const_op), ge.sgv(var_reader_op))\n<\/code><\/pre>\n\n<p>Note: if you save the converted model and view it in Tensorboard or Netron, you will see that Variables have taken the Constants' places. You will also see a bunch of dangling Constants, which you can optionally remove.<\/p>\n\n<p>I have verified that the weight values are the same between the frozen and unfrozen versions.<\/p>\n\n<p>Here is the <code>load_pb<\/code> function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n# Load protobuf as graph, given filepath\ndef load_pb(path_to_pb):\n    with tf.gfile.GFile(path_to_pb, 'rb') as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n    with tf.Graph().as_default() as graph:\n        tf.import_graph_def(graph_def, name='')\n        return graph\n<\/code><\/pre>\n"},{"AnswerId":"53094020","CreationDate":"2018-11-01T01:22:45.950","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<p><em>You were actually in the right direction with the snippet you provided :)<\/em><\/p>\n\n<hr>\n\n<h3>Step 1: get the name of previously trainable variables<\/h3>\n\n<p>The most tricky part is to get the <em>names<\/em> of previously trainable variables. Hopefully the model was created with some high-level frameworks, like <a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\"><code>keras<\/code><\/a> or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/slim\" rel=\"nofollow noreferrer\"><code>tf.slim<\/code><\/a> - they wraps their  variables nicely in something like <code>conv2d_1\/kernel<\/code>, <code>dense_1\/bias<\/code>, <code>batch_normalization\/gamma<\/code>, etc.<\/p>\n\n<p>If you're not sure, the most useful thing to do is to visualize the graph...<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># read graph definition\nwith tf.gfile.GFile('frozen.pb', 'rb') as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n\n# now build the graph in the memory and visualize it\nwith tf.Graph().as_default() as graph:\n    tf.import_graph_def(graph_def, name=\"prefix\")\n    writer = tf.summary.FileWriter('out', graph)\n    writer.close()\n<\/code><\/pre>\n\n<p>... with tensorboard:<\/p>\n\n<pre><code>$ tensorboard --logdir out\/\n<\/code><\/pre>\n\n<p>and see for yourself what the graph looks like and what the naming is.<\/p>\n\n<hr>\n\n<h3>Step 2: replace constants with variables (the fun part :D)<\/h3>\n\n<p>All you need is the magical library called <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.graph_editor\" rel=\"nofollow noreferrer\"><code>tf.contrib.graph_editor<\/code><\/a>. Now let's say you've stored the names of <em>previously trainable<\/em> ops (that previously were variables but now they are <code>Const<\/code>) in <code>probable_variables<\/code> (as in your <em>Edit 2<\/em>). <\/p>\n\n<p><em>Note: remember the difference between <code>ops<\/code>, <code>tensors<\/code>, and <code>variables<\/code>. Ops are elements of the graph, tensor is a buffer that contains results of ops, and variables are <a href=\"https:\/\/stackoverflow.com\/questions\/44167134\/whats-the-difference-between-tensor-and-variable-in-tensorflow\/44167844\">wrappers<\/a> around tensors, with 3 ops: <code>assign<\/code> (to be called when you initialize the variable), <code>read<\/code> (called by other ops, e.g. <code>conv2d<\/code>), and <a href=\"https:\/\/stackoverflow.com\/questions\/42783909\/internals-of-variable-in-tensorflow\"><code>ref tensor<\/code><\/a> (which holds the values).<\/em><\/p>\n\n<p><em>Note 2: <code>graph_editor<\/code> can <strong>only<\/strong> be run <strong>outside a session<\/strong> &ndash; you cannot make any graph modification online!<\/em><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import numpy as np\nimport tensorflow.contrib.graph_editor as ge\n\n# load the graphdef into memory, just as in Step 1\ngraph = load_graph('frozen.pb')\n\n# create a variable for each constant, beware the naming\nconst_var_name_pairs = []\nfor name in probable_variables:\n    var_shape = graph.get_tensor_by_name('{}:0'.format(name)).get_shape()\n    var_name = '{}_a'.format(name)\n    var = tf.get_variable(name=var_name, shape=var_shape, dtype='float32')\n    const_var_name_pairs.append((name, var_name))\n\n# from now we're going to work with GraphDef\nname_to_op = dict([(n.name, n) for n in graph.as_graph_def().node])\n\n# magic: now we swap the outputs of const and created variable\nfor const_name, var_name in const_var_name_pairs:\n    const_op = name_to_op[const_name]\n    var_reader_op = name_to_op[var_name + '\/read']\n    ge.swap_outputs(ge.sgv(const_op), ge.sgv(var_reader_op))\n\n# Now we can safely create a session and copy the values\nsess = tf.Session(graph=graph)\nfor const_name, var_name in const_var_name_pairs:\n    ts = graph.get_tensor_by_name('{}:0'.format(const_name))\n    var = tf.get_variable(var_name)\n    var.load(ts.eval(sess))\n\n# All done! Now you can make sure everything is correct by visualizing\n# and calculate outputs for some inputs.\n<\/code><\/pre>\n\n<p><strong><em>PS:<\/strong> this code was not tested; however, i've been using <code>graph_editor<\/code> and performing network surgery quite often lately, so I think it should mostly be correct :)<\/em><\/p>\n"}]}
+{"QuestionId":53085509,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-10-31T14:17:28.897","AcceptedAnswerId":53087861.0,"OwnerUserId":7886651.0,"Title":"Running the model on the development dataset after restoring the model while using tf.data.Dataset api in tensorflow","Body":"<p>I have trained my model in tensorflow while using the tf.data.Dataset api in tensorflow. I would like to demonstrate some parts of my code:<\/p>\n\n<pre><code># This corresponds to loading the data using the tf.data.Dataset api...\nnames_train, detected_train, arousal_train, valence_train, liking_train, is_talking_train, images_train,\\\niterator_train_all = load_train_sewa_tfrecords(filenames_train, train_batch_size)\n\nnames_devel, detected_devel, arousal_devel, valence_devel, liking_devel, is_talking_devel, images_devel, \\\niterator_dev_all = load_devel_sewa_tfrecords(filenames_dev, test_batch_size)\n<\/code><\/pre>\n\n<p>Now, I have seen that training and testing the model should be done by creating the model and then reusing it for testing. So:<\/p>\n\n<pre><code>train_predictions, model_layers_name_shape = simpler_CNN(images_train, my_initializer, phase_train, dropout,\n                                                         dopout_per_call_bool, mseed, reuse=False)\n\ndevel_predictions, _ = simpler_CNN(images_devel, my_initializer, phase_train, dropout, dopout_per_call_bool,\n                                   mseed, reuse=True)\n<\/code><\/pre>\n\n<p>Now, here is my problem: when creating the model, I have passed the training dataset images as inputs to the model. On the other hand, when it comes to extracting certain features from the model given the testing dataset, tensorflow will ask me to initialize the training iterator because it assumes I am feeding the training dataset (given how I created the model <strong>\"i.e. I have passed the training images as input to the model where the reuse was False\"<\/strong>).<\/p>\n\n<p>Now I have tried using a condition as follows:<\/p>\n\n<pre><code># the train_dataset will tell whether we are using the training or the testing dataset\ntrain_dataset = tf.placeholder(dtype=tf.bool, shape=(), name='train_dataset')\nimages = tf.cond(train_dataset, return_images_train, return_images_devel)\n\n# and then passing the images to the simpler_CNN while reuse is False:\ntrain_predictions, model_layers_name_shape = simpler_CNN(images, my_initializer, phase_train, dropout,\n                                                     dopout_per_call_bool, mseed, reuse=False)\n<\/code><\/pre>\n\n<p>Yet, this does not solve the problem and tensorflow will still ask me to initialize the training iterator. <\/p>\n\n<p>Furthermore, if I run initialize both the training and the testing iterators, then run the model on the testing dataset till the end; now, If I try feeding the training dataset, I will get <code>End of sequence<\/code> on the training iterator <em>as if the training iterator was running also while feeding the testing dataset<\/em> (this behavior is weird for me).<\/p>\n\n<p>In conclusion, how can I tell the model that I would like to load the testing dataset? Also, note that I can run <code>devel_prediction<\/code> and tensorflow will know what I am talking about, but running a hidden layer in the model will create the problem.<\/p>\n","answers":[{"AnswerId":"53087861","CreationDate":"2018-10-31T16:17:10.550","ParentId":null,"OwnerUserId":"7886651","Title":null,"Body":"<p>So when creating the model for the first time, let's assume we have the following names for the model layers:<\/p>\n\n<pre><code>layer1\/Relu:0\nlayer2\/Relu:0\nlayer3\/Relu:0\nlayer4\/Relu:0\nlayer5\/Relu:0\nglobal_average_pooling:0\ndense\/BiasAdd:0\n<\/code><\/pre>\n\n<p>Now, after reusing the model for the testing dataset, tensorflow will create another model but the weights will reference the initial weights created when creating the model for the first time. And when printing out the name of the layers of the reused model I got the following:<\/p>\n\n<pre><code>layer1_1\/Relu:0\nlayer2_1\/Relu:0\nlayer3_1\/Relu:0\nlayer4_1\/Relu:0\nlayer5_1\/Relu:0\nglobal_average_pooling_1:0\ndense_1\/BiasAdd:0\n<\/code><\/pre>\n\n<p>Therefore, I figured out that in my case, I was running a hidden layer such as <code>layer5\/Relu:0<\/code>. Thus, this will require us using the training dataset and this is why tensorflow was complaining. On the other hand, I should have ran <code>layer5_1\/Relu:0<\/code> and problem solved!!<\/p>\n"}]}
+{"QuestionId":53086166,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><torch>","CreationDate":"2018-10-31T14:50:16.680","AcceptedAnswerId":53091954.0,"OwnerUserId":5511906.0,"Title":"What's the reason for the weights of my NN model don't change a lot?","Body":"<p>I am training a neural network model, and my model fits the training data well. The training loss decreases stably. Everything works fine. <br>However, when I output the weights of my model, I found that it didn't change too much since random initialization (I didn't use any pretrained weights. All weights are initialized by default in PyTorch). All dimension of the weights only changed about 1%, while the accuracy on training data climbed from 50% to 90%.\nWhat could account for this phenomenon? Is the dimension of weights too high and I need to reduce the size of my model? Or is there any other possible explanations?  <br><br>\nI understand this is a quite broad question, but I think it's impractical for me to show my model and analyze it mathematically here. So I just want to know what could be the general \/ common cause for this problem.<\/p>\n","answers":[{"AnswerId":"53091954","CreationDate":"2018-10-31T21:06:01.083","ParentId":null,"OwnerUserId":"10528078","Title":null,"Body":"<p>There are almost always many local optimal points in a problem so one thing you can't say specially in high dimensional feature spaces is which optimal point your model parameters will fit into. one important point here is that for every set of weights that you are computing for your model to find a optimal point, because of <strong>real<\/strong> value weights,  there are infinite set of weights for that optimal point, the proportion of weights to each other is the only thing that matters, because you are trying to minimize the cost, not finding a unique set of weights with loss of 0 for every sample. every time you train you may get different result based on initial weights. when weights change very closely with almost same ratio to each others this means your features are highly correlated(i.e. redundant) and since you are getting very high accuracy just with a little bit of change in weights, only thing i can think of is that your data set classes are far away from each other. try to remove features one at a time, train and see results if accuracy was good continue to remove another one till you hopefully reach to a 3 or 2 dimensional space which you can plot your data and visualize it to see how data points are distributed and make some sense out of this.<\/p>\n\n<p><strong>EDIT: Better approach is to use PCA for dimensionality reduction instead of removing one by one<\/strong><\/p>\n"}]}
+{"QuestionId":53087036,"AnswerCount":1,"Tags":"<keras><deep-learning><conv-neural-network>","CreationDate":"2018-10-31T15:32:44.033","AcceptedAnswerId":null,"OwnerUserId":10585308.0,"Title":"How to fix specific weight at embedding layer","Body":"<p>I want to implement cnn sentence classification and padding 0 for variation length sentences.<\/p>\n\n<p>But it is not reasonalble for this task because 0 will be treated as a vocabulary.<\/p>\n\n<p>To solve this problem, I want to fix weight all zero for 0 map to word-vector in embedding layer,and other vocabular still can be trained. <\/p>\n\n<p>Reference paper: <a href=\"https:\/\/arxiv.org\/abs\/1408.5882\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1408.5882<\/a><\/p>\n\n<p>Following is my current code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>w2v_weight = np.array(list(data['id_vec'].values()))\n# add zeros in first dim:\nzero = np.zeros((1,w2v_weight.shape[1]))\nw2v_weight = np.concatenate((zero,w2v_weight),axis=0) \n\nembedding_layer = Embedding(len(data[\"word_id\"]) + 1, 300, weights= [w2v_weight],\n                        input_length=data['x_test'].shape[1], trainable=True)\n\nembedding_layer2 = Embedding(len(data[\"word_id\"]) + 1, 300, weights= [w2v_weight],\n                        input_length=data['x_test'].shape[1], trainable=False)\n\nmodel_input = Input(shape=(None,),dtype='int32')\n\nembedded_sequences2 = embedding_layer2(model_input)\nembedded_sequences1 = embedding_layer(model_input)\nebd_cct = Concatenate()([embedded_sequences1,embedded_sequences2])\n\nconv1 = Convolution1D(filters=100,kernel_size = 3,padding=\"same\")(ebd_cct)\nconv2 = Convolution1D(filters=100,kernel_size = 4,padding=\"same\")(ebd_cct)\nconv3 = Convolution1D(filters=100,kernel_size = 5,padding=\"same\")(ebd_cct)\nconv_a = Concatenate()([conv1,conv2,conv3])\nconv_a = Activation(\"relu\")(conv_a)\nconv_add = GlobalMaxPool1D()(conv_a) \nz = Dropout(0.5)(conv_add) \n\nmodel_output = Dense(4, activation=\"softmax\",kernel_constraint = max_norm(3.))(z)\nmodel_1_two = Model(model_input, model_output)\n\nmodel_1_two.summary()\n\nmodel_1_two.compile(loss=\"categorical_crossentropy\", optimizer=\"Adadelta\",         \nmetrics=['acc'])\nhistory_1_two = model_1_two.fit(data[\"x_train\"], data[\"y_train\"],shuffle=True,\n      callbacks = [EarlyStopping(monitor=\"val_acc\",patience=50)],\n      batch_size=50, epochs=20000,validation_data=(data[\"x_test\"], data[\"y_test\"]))    \n<\/code><\/pre>\n","answers":[{"AnswerId":"53087285","CreationDate":"2018-10-31T15:45:50.423","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>From your code I assume you're using Keras. Then, you can define the use of <code>0<\/code> as the out-of-vocabulary (OOV) index by setting <code>mask_zero<\/code> to <code>True<\/code>:<\/p>\n\n<pre><code>embedding_layer = Embedding(len(data[\"word_id\"]) + 1, 300, weights= [w2v_weight],\n                            input_length=data['x_test'].shape[1], trainable=True,\n                            mask_zero=True)\n<\/code><\/pre>\n\n<p>For more information, see the <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">documentation for <code>Embedding()<\/code><\/a>.<\/p>\n"}]}
+{"QuestionId":53087320,"AnswerCount":1,"Tags":"<keras><deep-learning>","CreationDate":"2018-10-31T15:47:40.230","AcceptedAnswerId":53116335.0,"OwnerUserId":6544207.0,"Title":"Keras keyboard interrupt to stop training?","Body":"<p>There seems to be a way to do it in tensorflow: <a href=\"https:\/\/stackoverflow.com\/questions\/47706467\/keyboard-interrupt-tensorflow-run-and-save-at-that-point\">Keyboard interrupt tensorflow run and save at that point<\/a> <\/p>\n\n<p>Is there something like this in Keras?<\/p>\n","answers":[{"AnswerId":"53116335","CreationDate":"2018-11-02T09:58:32.480","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You could catch the <code>KeyboardInterrupt<\/code> exception and save the model within the <code>except<\/code> block:\n<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>save_path = '.\/keras-saves\/_lastest.ckpt'\ntry:\n    model.fit(x_train, y_train,\n              batch_size=batch_size,\n              epochs=epochs)\nexcept KeyboardInterrupt:\n    model.save(save_path)\n    print('Output saved to: \"{}.\/*\"'.format(save_path))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53087669,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-10-31T16:06:27.703","AcceptedAnswerId":null,"OwnerUserId":2554262.0,"Title":"NARX implementation using keras","Body":"<p>I am trying to implement a simple NARX network in keras (backend Tensoflow). I am building the model by subclassing <code>tf.keras.Model<\/code>. A NARX network can be trained as a simple feedforward network (series-parallel architecture), therefore I did not have any problem with this definition. However for prediction I need to close the loop and give back the output as input to the model. How can I modify the network so that it is able to make predictions and forecastings?<\/p>\n\n<p>Here is a minimal version of the code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nnumPreviousSteps = 8\ninputShape = (None, numPreviousSteps + 2)\n\nclass Narx(keras.Model):\n\n    def __init__(self):\n        super(Narx, self).__init__(name='narx')\n        self.dense = keras.layers.Dense(10, input_shape=inputShape,\n                                        activation=keras.activations.tanh)\n        self.outputLayer = keras.layers.Dense(1, activation=keras.activations.linear)\n\n    def call(self, inputs, training = False):\n        if (training):\n            x = self.dense(inputs)\n            return self.outputLayer(x)\n        else: # TODO: what should the network do when used for prediction\n            x = self.dense(inputs)\n            return self.outputLayer(x)\n\n\nmodel = Narx()\nmodel.compile(optimizer=keras.training.RMSPropOptimizer(0.001),\n              loss=tf.losses.mean_squared_error,\n              metric=tf.metrics.mean_absolute_error)\n\n# input data generation\nnumTsSamples = 1000\n\n# time series to learn from\ny = np.random.random((numTsSamples + numPreviousSteps + 1,))\nx = np.random.random((numTsSamples,)) # exogenous input\n\n# creation of tapped delay\ndata = [np.roll(y, -i)[:numTsSamples] for i in range(numPreviousSteps, -1, -1)]\ndata = [x] + data\n\n# training data\ndata = np.stack(data, axis=1)\n\n# expected results\nyNext = y[numPreviousSteps : -1]\n\n# model training\nmodel.fit(data, yNext)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53087772,"AnswerCount":0,"Tags":"<python><tensorflow><encoding><keras>","CreationDate":"2018-10-31T16:12:15.960","AcceptedAnswerId":null,"OwnerUserId":1506850.0,"Title":"changes in training metrics after changing label encoding","Body":"<p>I am training a binary classifier using Keras.<\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy',\n              optimizer='rmsprop',\n              metrics=[auroc,'accuracy'])\n<\/code><\/pre>\n\n<p>I use a custom metric, AUROC.<a href=\"https:\/\/stackoverflow.com\/questions\/41032551\/how-to-compute-receiving-operating-characteristic-roc-and-auc-in-keras\">as in here<\/a><\/p>\n\n<pre><code>import tensorflow as tf\nfrom sklearn.metrics import roc_auc_score  \ndef auroc(y_true, y_pred):\n    return tf.py_func(roc_auc_score, (y_true, y_pred), tf.double)\n<\/code><\/pre>\n\n<p>So far, I have encoded my target as one-hot encoding \nand I had a last layer as <\/p>\n\n<pre><code>from keras.utils import to_categorical\ny = to_categorical(y)\n[...]\nmodel.add(Dense(2, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>I have learned that in principle here <a href=\"https:\/\/stackoverflow.com\/questions\/42081257\/keras-binary-crossentropy-vs-categorical-crossentropy-performance\">Keras binary_crossentropy vs categorical_crossentropy performance?<\/a>I should not perform categorical encoding, and I should predict only one class using <\/p>\n\n<pre><code># y = to_categorical(y)\n[...]\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>However, if I apply this and only this change my training <code>auroc<\/code> changes dramatically, from high 0.90s to 0.50. Even more strangely, <code>val_auroc<\/code> loss seems unaffected<\/p>\n\n<p>How did that happen?<\/p>\n","answers":[]}
+{"QuestionId":53088223,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-10-31T16:37:17.463","AcceptedAnswerId":null,"OwnerUserId":6807211.0,"Title":"Keras image augmentation for images with two channels","Body":"<p>I performed data augmentation for images with two channels. My data set is formatted in the shape of <code>(image_Numbers, image_height, image_weights, image_channels)<\/code>, where <code>image_channels = 2<\/code>. <\/p>\n\n<p>In performing data augmentation using <code>datagen<\/code> (created by <code>ImageDataGenerator<\/code>), a userwarining message is generated:<\/p>\n\n<pre><code>UserWarning: NumpyArrayIterator is set to use the data format convention     \n\"channels_last\" (channels on axis 3), \ni.e. expected either 1, 3 or 4 channels on axis 3. \nHowever, it was passed an array with shape (1, 150, 150, 2) (2 channels).\n<\/code><\/pre>\n\n<p>Does the warning imply the data augmentation was unsuccessful? Was it only performed for one channel images? If so, how to perform data augmentation for two-channels of images (not one channel this time and then concatenation)?<\/p>\n","answers":[{"AnswerId":"53090919","CreationDate":"2018-10-31T19:46:56.187","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It means they don't expect two channel images. It's non standard. <\/p>\n\n<p>The standard images are:<\/p>\n\n<ul>\n<li>1 channel: grayscale   <\/li>\n<li>3 channels: RGB    <\/li>\n<li>4 channels: RGBA<\/li>\n<\/ul>\n\n<p>Since it's a warning, we don't really know what's going on.   <\/p>\n\n<p>Check the outputs of this generator yourself. <\/p>\n\n<pre><code>x, y = theGenerator[someIndex]\n<\/code><\/pre>\n\n<p>Plot <code>x[0]<\/code> and others.<\/p>\n\n<p>In case the generated images aren't good, you can do the augmentations yourself using a python generator or a custom <code>keras.utils.Sequence<\/code>.   <\/p>\n"}]}
+{"QuestionId":53088690,"AnswerCount":2,"Tags":"<keras><rgb><grayscale>","CreationDate":"2018-10-31T17:06:02.060","AcceptedAnswerId":null,"OwnerUserId":10587168.0,"Title":"RGB to GRAY using opencv removes 1 channel","Body":"<p>I'm using these for convertin RGB image to GRAY<\/p>\n\n<pre><code>i = cv2.resize(img, (100, 100))\ni = cv2.cvtColor(i, cv2.COLOR_BGR2GRAY)\n<\/code><\/pre>\n\n<p>This gives me (100,100) but I want (100,100,1) as this is required as input to CNN as follows<\/p>\n\n<pre><code>model.add(Conv2D(16, (2,2), input_shape=(100,100,1 ), activation='relu'))\n<\/code><\/pre>\n\n<p>Show how should I convert RGB to GRAY such that 1 channel factor remains retained. <\/p>\n","answers":[{"AnswerId":"53088755","CreationDate":"2018-10-31T17:10:34.210","ParentId":null,"OwnerUserId":"5741205","Title":null,"Body":"<p>Use numpy reshape:<\/p>\n\n<pre><code>i = i.reshape(i.shape + (1,))\n<\/code><\/pre>\n"},{"AnswerId":"53090413","CreationDate":"2018-10-31T19:08:42.100","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>You can just add the extra dimension<br>\n<code>np.expand_dims(i, axis=0)<\/code><\/p>\n"}]}
+{"QuestionId":53088825,"AnswerCount":0,"Tags":"<python><out-of-memory><gpu><cpu><pytorch>","CreationDate":"2018-10-31T17:14:45.973","AcceptedAnswerId":null,"OwnerUserId":10587121.0,"Title":"PyTorch out of GPU memory just after initialization","Body":"<p>I built my model in PyTorch. <strong>The code works well on CPU<\/strong>. However, it seems to be running out of GPU memory <strong>just after initializing the network and switching it to cuda<\/strong>. <\/p>\n\n<p>The failed code is:<\/p>\n\n<pre><code>model = init_from_scratch(args, train_exs, dev_exs)\nmodel.init_optimizer()\n\nnParams= sum([np.prod(list(p.size())) for p in model.network.parameters()])\nprint('* total number of parameters:',nParams)\n\ndevice = torch.device(\"cuda:\"+str(args.gpu) if args.cuda else \"cpu\")\nmodel.set_device(device) #call last!!\n\n# DATA ITERATORS  (not run to here)\ntrain_dataset = data.ReaderDataset()\ntrain_sampler = torch.utils.data.sampler.RandomSampler()\ntrain_loader = torch.utils.data.DataLoader()\n...\n\n\n# TRAIN (not run to here)\nfor epoch in range(start_epoch, args.num_epochs):\n....\n\ndef set_device(self,device):\n    print(\"device\",device)\n    self.use_cuda = False if str(device) == \"cpu\" else True\n    self.network = self.network.to(device)\n<\/code><\/pre>\n\n<ul>\n<li><p>total number of parameters: 30708906<\/p><\/li>\n<li><p>Torch version is '0.4.0' <\/p><\/li>\n<li><p>TITAN Xp with 12G GPU memory and nvidia-smi shows it free<\/p><\/li>\n<\/ul>\n\n<p>So, what's wrong with my code? <\/p>\n\n<p>The error output is:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"script\/train2.0.py\", line 683, in &lt;module&gt;\n  main(args)\nFile \"script\/train2.0.py\", line 542, in main\n  model.set_device(device)\nFile \".\/model.py\", line 588, in set_device\n  self.network = self.network.to(device)\nFile \"\/home\/username\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 393, in to\n  return self._apply(lambda t: t.to(device))\nFile \"\/home\/username\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 176, in _apply\n  module._apply(fn)\nFile \"\/home\/username\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 182, in _apply\n  param.data = fn(param.data)\nFile \"\/home\/username\/.local\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 393, in &lt;lambda&gt;\n  return self._apply(lambda t: t.to(device))\nFile \"\/home\/username\/.local\/lib\/python3.5\/site-packages\/torch\/cuda\/__init__.py\", line 161, in _lazy_init\n  torch._C._cuda_init()\n<\/code><\/pre>\n\n<p>RuntimeError: cuda runtime error (2) : out of memory at \/pytorch\/aten\/src\/THC\/THCTensorRandom.cu:25<\/p>\n","answers":[]}
+{"QuestionId":53088856,"AnswerCount":2,"Tags":"<tensorflow><neural-network><keras><deep-learning>","CreationDate":"2018-10-31T17:16:34.947","AcceptedAnswerId":null,"OwnerUserId":3241790.0,"Title":"Adam optimizer Keras vs Tensorflow","Body":"<p>I would ask why Keras implementation for Adam optimizer has Decay argument and Tensorflow doesn't? and what idea of this argument?<\/p>\n","answers":[{"AnswerId":"53089330","CreationDate":"2018-10-31T17:50:56.410","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Why is very hard to answer.<\/p>\n\n<p>It's interesting to have a decay, though, for when your train reaches a limit. \nLowering learning rates may improve your model with finer results. But machine learning is all about testing.<\/p>\n\n<p>The idea is simply reduce the value of the learning rate in every batch update. <\/p>\n\n<p>This is the formula Keras uses:<\/p>\n\n<pre><code>lr = self.lr\nif self.initial_decay &gt; 0:\n    lr = lr * (1. \/ (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))\n<\/code><\/pre>\n\n<p>Basically its:<\/p>\n\n<pre><code>lr \/ (1 + decay*currentBatch) #considering currentBatch keeps incresing, not looping    \n<\/code><\/pre>\n"},{"AnswerId":"54079462","CreationDate":"2019-01-07T17:59:36.723","ParentId":null,"OwnerUserId":"5763590","Title":null,"Body":"<blockquote>\n  <p>The differences might somehow reflect the discussion whether learning\n  rate decay is even needed when applying <a href=\"https:\/\/arxiv.org\/abs\/1412.6980\" rel=\"nofollow noreferrer\">Adam<\/a>.<\/p>\n<\/blockquote>\n\n<ol>\n<li>Adam updates any parameter with an individual learning rate. This means that every parameter in the network have a specific learning rate associated.<\/li>\n<li>The single learning rates for parameters is computed using the initial learning rate as upper limit. This means that every single learning rate can vary from 0 (no update) to the initial learning rate.<\/li>\n<li>The learning rates adapt themselves during train steps, but if you want to be sure that every update step do not exceed an upper Limit you can than lower your initial (global) learning rate, using exponential decay. <\/li>\n<\/ol>\n\n<p>So these reasons show why there is a discussion whether learning rate decay with Adam is necessary after all.<\/p>\n"}]}
+{"QuestionId":53089078,"AnswerCount":1,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-10-31T17:31:39.563","AcceptedAnswerId":53089168.0,"OwnerUserId":7517834.0,"Title":"model with a various length of input parameters?","Body":"<p>with the following model, how to ensure that input length is variable ?<\/p>\n\n<pre><code>const input = tf.input({shape: [5]});\nconst denseLayer1 = tf.layers.dense({units: 10, activation: \n'relu'});\nconst denseLayer2 = tf.layers.dense({units: 4, activation: \n'softmax'});\nconst output = \ndenseLayer2.apply(denseLayer1.apply(input));\nconst model = tf.model({inputs: input, outputs: output});\nmodel.predict(tf.ones([2, 5])).print();\n<\/code><\/pre>\n","answers":[{"AnswerId":"53089168","CreationDate":"2018-10-31T17:38:02.877","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>You can specify a variable length by using <code>null<\/code> in the inputShape. It does works only for input of dimension greater than 1.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const input = tf.input({shape: [null, 2]});\r\nconst denseLayer1 = tf.layers.dense({units: 10, activation: \r\n'relu'});\r\nconst denseLayer2 = tf.layers.dense({units: 4, activation: \r\n'softmax'});\r\nconst output = \r\ndenseLayer2.apply(denseLayer1.apply(input));\r\nconst model = tf.model({inputs: input, outputs: output});\r\nmodel.predict(tf.randomNormal([6, 5, 2])).print();\r\nmodel.predict(tf.randomNormal([6, 4, 2])).print();<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.0\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53089102,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2018-10-31T17:33:09.320","AcceptedAnswerId":null,"OwnerUserId":3723031.0,"Title":"How to share weights among subset of input features with keras","Body":"<p>In this neural network, there are 9 input features:<\/p>\n\n<pre><code>  f1,f2,f3,f4,f5,f6,f7,f8,f9\n<\/code><\/pre>\n\n<p>I want some of the input features, but not all, to have identical weights between the input layer and the first hidden layer.  All remaining layers will not have any shared weights.<\/p>\n\n<pre><code>  f1,f2,f3 should share the same weights\n  f4,f5,f6 should share the same weights\n  f7 should not share weights with other features\n  f8 should not share weights with other features\n  f9 should not share weights with other features\n<\/code><\/pre>\n\n<p>I'm pretty sure I need 1D convolutions, but not on the entire layer.  I may have missed it, but haven't seen this use case documented. Any ideas how to do this in Keras?<\/p>\n\n<p>Rephrasing the question, what is the right way to express equal importance among a group of features?<\/p>\n\n<p>Features (<code>f1<\/code>, <code>f2<\/code>, <code>f3<\/code>) have equal importance when predicting the output class.  Features (<code>f4<\/code>, <code>f5<\/code>, <code>f6<\/code>) also have equal importance when predicting the output class.<\/p>\n\n<p>There are three possible predicted classes.  Features <code>f1<\/code> and <code>f4<\/code> are evidence for output classA. Features <code>f2<\/code> and <code>f5<\/code> are evidence for output <code>classB<\/code>. Features <code>f3<\/code> and <code>f6<\/code> are evidence for output <code>classC<\/code>.<\/p>\n\n<p>Is there a way to reduce the number of parameters in the network by sharing parameters among the equal-importance features?<\/p>\n","answers":[{"AnswerId":"53089260","CreationDate":"2018-10-31T17:45:48.447","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Equivalent to summing <code>f1+f2+f3<\/code> and <code>f4+f5+f6<\/code> before a dense layer.<\/p>\n\n<p>Suggestion of layer to sum:<\/p>\n\n<pre><code>from keras.layers import Lambda\nimport keras.backend as K\n\ndef sumFeatures(x):\n    f1_3 = x[:,:3]\n    f4_6 = x[:,3:6]\n\n    return K.concatenate(\n        [\n            K.sum(f1_3, axis=-1, keepdims=True),\n            K.sum(f4_6, axis=-1, keepdims=True),\n            x[:,6:]\n        ], axis=-1)\n<\/code><\/pre>\n\n<p>Sequential model:<\/p>\n\n<pre><code>model.add(Lambda(sumFeatures))\n<\/code><\/pre>\n\n<p>Functional Model:<\/p>\n\n<pre><code>outputTensor = Lambda(sumFeatures)(inputTensor)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53089132,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><nltk><pos-tagger>","CreationDate":"2018-10-31T17:35:17.803","AcceptedAnswerId":null,"OwnerUserId":8414280.0,"Title":"How to calculate pos tagger accuracy by each tag","Body":"<p>I've made a keras model pos tagger using the NLP4Hackers' <a href=\"https:\/\/www.google.com.br\/amp\/s\/nlpforhackers.io\/lstm-pos-tagger-keras\/amp\/\" rel=\"nofollow noreferrer\">article<\/a> as baseline. Currently I can computate the accuracy directly over the keras <em>model.evaluate<\/em> method. Acctualy, I would like to calculate the accuracy by tag as show below:<\/p>\n\n<p>'JJ': 98.56 accuracy,<\/p>\n\n<p>'NNS': 99.01 accuracy <\/p>\n\n<p>'NN': 96.43 accuracy<\/p>\n\n<p>...<\/p>\n\n<p>Any suggestion will be appreciated.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53091690","CreationDate":"2018-10-31T20:44:11.990","ParentId":null,"OwnerUserId":"5121457","Title":null,"Body":"<p>All the evaluation metrics that you can imagine are in Scikit-learn<\/p>\n\n<p>You have two possibilities, either you compute the <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.confusion_matrix.html#sklearn.metrics.confusion_matrix\" rel=\"nofollow noreferrer\">confusion matrix<\/a> \nand you look the diagonal values.<\/p>\n\n<pre><code>y_true = [\"cat\", \"ant\", \"cat\", \"cat\", \"ant\", \"bird\"]\ny_pred = [\"ant\", \"ant\", \"cat\", \"cat\", \"ant\", \"cat\"]\nconfusion_matrix(y_true, y_pred, labels=[\"ant\", \"bird\", \"cat\"])\narray([[2, 0, 0],\n       [0, 0, 1],\n       [1, 0, 2]])\n<\/code><\/pre>\n\n<p>Or you compute the <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score\" rel=\"nofollow noreferrer\">F1 score<\/a> label by label<\/p>\n\n<pre><code>from sklearn.metrics import f1_score\ny_true = [0, 1, 2, 0, 1, 2]\ny_pred = [0, 2, 1, 0, 0, 1]\nf1_score(y_true, y_pred, average=None)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53089366,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-10-31T17:53:39.007","AcceptedAnswerId":53089687.0,"OwnerUserId":10587312.0,"Title":"TensorFlow installation on Win7x64 (no matching distribution \/ not a supported wheel)","Body":"<p>I'm trying to get TensorFlow to work on my Win7x64 following this guide:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a><\/p>\n\n<p>Everything has gone well until <code>3. Install the TensorFlow pip package<\/code><\/p>\n\n<p>When i do \"<code>python<\/code>\" to see the version i get the following:<\/p>\n\n<pre><code>Python 3.7.1 (v3.7.1:260ec2c36a, Oct 20 2018, 14:57:15) [MSC v.1915 64 bit (AMD64)] on win32\n<\/code><\/pre>\n\n<p>When i do \"<code>pip install tensorflow<\/code>\" i get:<\/p>\n\n<pre><code>Could not find a version that satisfies the requirement of tensorflow (from versions: )\nNo matching distribution found for tensorflow\n<\/code><\/pre>\n\n<p>If i try \"<code>python -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.2.1-cp35-cp35m-win_amd64.whl<\/code>\" i get this:<\/p>\n\n<pre><code>tensorflow-1.2.1-cp55-cp35m-win_amd64.whl is not a supported wheel on this platform.\n<\/code><\/pre>\n\n<p>Please help me understand what am i doing wrong or why is it not working? How to get it to work? Thank you very much!<\/p>\n","answers":[{"AnswerId":"53089687","CreationDate":"2018-10-31T18:13:39.007","ParentId":null,"OwnerUserId":"9176689","Title":null,"Body":"<p>You have an unsupported version of Python. Tensorflow currently only supports Python 3.6 and below, not 3.7 and even if you do manage to install it you will run into issues as Python 3.7 made <code>async<\/code> a keyword and many libraries, like Tensorflow, use this for variable names. In order to make this work on your system you need to downgrade to 3.6, there is actually a thread with troubleshooting steps on this issue on GitHub: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17022<\/a><\/p>\n"}]}
+{"QuestionId":53089849,"AnswerCount":1,"Tags":"<spring-cache><caffeine>","CreationDate":"2018-10-31T18:25:45.417","AcceptedAnswerId":53090609.0,"OwnerUserId":9091513.0,"Title":"Spring cache set expire time dinamically - Caffeine","Body":"<p>I use an authentication API to get a token and use other services. This API returns the token and expire time. It's possible to get the expire time it returns and set expire_after_write with these value? Currently this property is in application.properties, but I'm afraid that someday the service provider change expire_time and we get some errors because the token is expired <\/p>\n","answers":[{"AnswerId":"53090609","CreationDate":"2018-10-31T19:23:41.503","ParentId":null,"OwnerUserId":"19450","Title":null,"Body":"<p>You can set a <a href=\"https:\/\/github.com\/ben-manes\/caffeine\/wiki\/Eviction#time-based\" rel=\"nofollow noreferrer\">per-entry policy<\/a> to evaluate the entry and determine the expiration time. As the token won't be modified within the cache, you can use <code>expireAfterCreate<\/code> and have the other methods return the <code>currentDuration<\/code> to not modify it. From the docs,<\/p>\n\n<pre><code>LoadingCache&lt;Key, Graph&gt; graphs = Caffeine.newBuilder()\n    .expireAfter(new Expiry&lt;Key, Graph&gt;() {\n      public long expireAfterCreate(Key key, Graph graph, long currentTime) {\n        \/\/ Use wall clock time, rather than nanotime, if from an external resource\n        long seconds = graph.creationDate().plusHours(5)\n            .minus(System.currentTimeMillis(), MILLIS)\n            .toEpochSecond();\n        return TimeUnit.SECONDS.toNanos(seconds);\n      }\n      public long expireAfterUpdate(Key key, Graph graph, \n          long currentTime, long currentDuration) {\n        return currentDuration;\n      }\n      public long expireAfterRead(Key key, Graph graph,\n          long currentTime, long currentDuration) {\n        return currentDuration;\n      }\n    })\n    .build(key -&gt; createExpensiveGraph(key));\n<\/code><\/pre>\n"}]}
+{"QuestionId":53089945,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2018-10-31T18:32:47.640","AcceptedAnswerId":58797450.0,"OwnerUserId":10567259.0,"Title":"Tensorflow's Estimator stops training","Body":"<p>I am training a model using Tensorflow's <code>Estimator<\/code> and it suddenly stops training after 2600 steps after performing an evaluation. Isn't it supposed to continue training until the end of the last epoch?<\/p>\n\n<pre><code>def train():\n    train_input_func = lambda: input_fn(mode='train')\n    eval_input_func = lambda: input_fn(mode='eval')\n\n    est_conf = tf.estimator.RunConfig(cfg.model_dir, save_checkpoints_secs=120)\n    estimator = tf.estimator.Estimator(model_fn, cfg.model_dir, est_conf)\n\n\n    Path(estimator.eval_dir()).mkdir(parents=True, exist_ok=True)\n    train_spec = tf.estimator.TrainSpec(input_fn=train_input_func)\n    eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_func, throttle_secs=120)\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n\nif __name__ == '__main__':\n    train()\n<\/code><\/pre>\n\n<p>And this is the <code>input_fn<\/code> function:<\/p>\n\n<pre><code>def input_fn(mode=None):\n        data_generator = lambda: data_loader.data_generator(mode=mode)\n\n        dataset = tf.data.Dataset.from_generator(data_generator,\n                                                 output_types=(tf.int32, tf.int32),\n                                                 output_shapes=([None], [None]))\n\n        if mode is 'train':\n            dataset.shuffle(cfg.shuffle_buffer).repeat(1000)\n\n        dataset = dataset.padded_batch(cfg.batch_size, padded_shapes=([None],[None])).prefetch(1)\n\n        return dataset\n<\/code><\/pre>\n","answers":[{"AnswerId":"55113285","CreationDate":"2019-03-12T02:37:03.133","ParentId":null,"OwnerUserId":"5642366","Title":null,"Body":"<p>When use <code>tf.estimator.train_and_evaluate<\/code>, to make <code>max_steps<\/code> work, you should not use <code>repeat(1000)<\/code>, please use <code>repeat()<\/code>, it will repeat the input indefinitely, and will not throw <code>OutOfRangeError<\/code>.<\/p>\n"},{"AnswerId":"57863811","CreationDate":"2019-09-10T03:51:25.500","ParentId":null,"OwnerUserId":"6724265","Title":null,"Body":"<p><strong>First<\/strong> You need to specify the max_stps in the TrainSpec definition like the following: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_spec = tf.estimator.TrainSpec(input_fn=train_input_func, max_steps=num_steps_you_specify)\n<\/code><\/pre>\n\n<p><strong>Second<\/strong>\nThe training procedure will stop when the input_fn throws \"OutOfRangeError\" on which case the max_step will not work as it was designed to.  So in order to make the training run through the whole epochs, you need to specify the input_fn like the folllowing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = dataset.repeat()# don't specify any number in the repeat()\n<\/code><\/pre>\n\n<p>Hope this will help you.<\/p>\n"},{"AnswerId":"58797450","CreationDate":"2019-11-11T07:54:56.317","ParentId":null,"OwnerUserId":"10567259","Title":null,"Body":"<p>The problem was that I did not assign <code>dataset.shuffle(cfg.shuffle_buffer).repeat(1000)<\/code>. This will fix the problem:<\/p>\n\n<pre><code>dataset = dataset.shuffle(cfg.shuffle_buffer).repeat(1000)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53090165,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><loss-function><non-linear-regression>","CreationDate":"2018-10-31T18:51:04.773","AcceptedAnswerId":null,"OwnerUserId":8889727.0,"Title":"Non linear regression: Outlier sensitive loss function","Body":"<p>I'm trying to come up with a non-linear regression deep learning model to predict the far field of a given nanostructure. <\/p>\n\n<p>The far-field spectrum is highly non-linear and it could have several resonance points (sudden increase in the field). See the graph below.<\/p>\n\n<p>My DNN consists of 3 convolution layers followed by 4 dense layers to address non-linearity. <\/p>\n\n<p>However, I would like to have a loss function which is more sensitive to outliers, so I could find the exact points where resonance occurs. <\/p>\n\n<p>Till now, I have used quite a few loss functions and Huber loss performs best in terms of the settled final loss (0.0101), despite being less sensitive to outliers. The graph below shows the actual and predicted spectra.<\/p>\n\n<p>Actual and predicted spectra:\n<a href=\"https:\/\/i.stack.imgur.com\/zMSRA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zMSRA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I know mean squared error is sensitive to outliers but in this case, the settled loss (0.0198) is higher than the previous case. <\/p>\n\n<p>What loss function is most sensitive to outliers specifically in not linear regression. If you could propose a custom loss function for this problem, that would also be fine.<\/p>\n","answers":[]}
+{"QuestionId":53090359,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><anaconda><conda>","CreationDate":"2018-10-31T19:04:43.507","AcceptedAnswerId":null,"OwnerUserId":10345615.0,"Title":"Tensorflow-GPU Installation ImportError: DLL load failed: The specified module could not be found","Body":"<p>Ok i know this might be already answered question but i have tried almost every trick suggested on stackoverflow to install tensorflow-gpu and suggested on official docs but with no luck i am getting the same error. Firstly i have tried \n<a href=\"https:\/\/towardsdatascience.com\/setup-an-environment-for-machine-learning-and-deep-learning-with-anaconda-in-windows-5d7134a3db10\" rel=\"nofollow noreferrer\">this<\/a> then afterwards offical docs then stackoverflow tricks but still getting error I have installed<\/p>\n\n<ol>\n<li>Anaconda 3.5 which was with python 3.7 but then i downgraded it to 3.6.6<\/li>\n<li>Then created virtual env named tensorflow with python 3.6.6<\/li>\n<li>Then in tensorflow env pip install tensorflow-gpu<\/li>\n<li>It was installed successfully then i installed CUDA 9.0<\/li>\n<li>Then cuDNN 7.3 for CUDA 9.0(I have also tried CUDA 8.0 with cuDNN 6.1)<\/li>\n<li>Copied all DLL cuDNN bin\/lib to CUDA 9.0 bin\/lib<\/li>\n<li><p>Specified CUDA system environment variable path upto bin<\/p>\n\n<p><strong>SAME QUESTION IS ASKED <a href=\"https:\/\/stackoverflow.com\/questions\/52697809\/import-tensorflow-as-tf-gives-errors-even-after-installing-without-errors\">HERE<\/a> BUT HAS NO ANSWER I HAVE THE SAME TRACE.<\/strong><\/p>\n\n<p>Python 3.6.6 |Anaconda, Inc.| (default, Jun 28 2018, 11:27:44) [MSC v.1900 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import tensorflow\n      Traceback (most recent call last):\n      File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *\n      File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 29, in \n      _pywrap_tensorflow_internal = swig_import_helper()\n      File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 25, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n      File \"D:\\conda\\envs\\tensorflow\\lib\\imp.py\", line 243, in load_module\n      return load_dynamic(name, filename, file)\n      File \"D:\\conda\\envs\\tensorflow\\lib\\imp.py\", line 342, in load_dynamic\n      return _load(spec)\n      ImportError: DLL load failed: The specified module could not be found.<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\", line 1, in \n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow__init__.py\", line 22, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python__init__.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 29, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"D:\\conda\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 25, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"D:\\conda\\envs\\tensorflow\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"D:\\conda\\envs\\tensorflow\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.<\/p>\n\n<p>Please help me to get out of this.\nThanks in Advance.<\/p>\n","answers":[{"AnswerId":"53123210","CreationDate":"2018-11-02T17:21:39.383","ParentId":null,"OwnerUserId":"7261743","Title":null,"Body":"<p>I had same problem with you. After try many things, I can solved the problem with these processes.<\/p>\n\n<p>1.Created virtual environment named tensorflow with python 3.5<\/p>\n\n<ul>\n<li>conda create -n tensorflow pip python-3.5<\/li>\n<\/ul>\n\n<p>2.Activate this environment and install tensorflow<\/p>\n\n<ul>\n<li>activate tensorflow<\/li>\n<li>pip install --ignore-installed --upgrade tensorflow-gpu <\/li>\n<\/ul>\n\n<p>For CUDA, after installed CUDA 9.0, I create new folder in drive C and copy cuDNN 7.3 for CUDA 9.0 to its. Then, specified CUDA system environment variable path (bin folder).<\/p>\n\n<p>In my case:  C:\\tools\\cuda\\bin<\/p>\n\n<p>Hope this can help you :)<\/p>\n"}]}
+{"QuestionId":53090438,"AnswerCount":0,"Tags":"<amazon-web-services><tensorflow><amazon-ec2><gpu><tensorflow-serving>","CreationDate":"2018-10-31T19:10:07.430","AcceptedAnswerId":null,"OwnerUserId":6788974.0,"Title":"Tensorflow Serving using GPU on AWS EC2 GPU instance with Deep Learning AMI","Body":"<p>I am running a model using pre-installed tensorflow_model_server (refer <a href=\"https:\/\/docs.aws.amazon.com\/dlami\/latest\/devguide\/tutorial-tfserving.html\" rel=\"nofollow noreferrer\">here<\/a>) on AWS EC2 GPU instance with Deep Learning AMI, but it's only using CPU when I am running a client.<\/p>\n\n<p>Is there any way I can make the pre-installed tensorflow_model_server use GPU for the predictions?<\/p>\n","answers":[]}
+{"QuestionId":53090463,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-10-31T19:11:44.783","AcceptedAnswerId":53100400.0,"OwnerUserId":10429267.0,"Title":"Tensorflow Estimator - High evaluation values on training data","Body":"<p>I'm using Tensorflow 1.10 with a custom Estimator. To test my training\/evaluation loop, I just feed the same image\/label into the network every time, so I expected the network to converge fast, which it does.<\/p>\n\n<p>I'm also using the same image for evaluation, but get a much bigger loss value than when training. After training 2000 steps the loss is:<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Loss for final step: <strong>0.01181452<\/strong><\/p>\n<\/blockquote>\n\n<p>but evaluates to:<\/p>\n\n<blockquote>\n  <p>Eval loss at step 2000: <strong>0.41252694<\/strong><\/p>\n<\/blockquote>\n\n<p>This seems wrong to me. It looks like the same problem as in <a href=\"https:\/\/stackoverflow.com\/questions\/51798423\/tf-estimator-train-and-evaluate-got-wrong-evaluation-accuray-and-loss\">this<\/a> thread. Is there something special to consider, when using the <code>evaluate<\/code> method of <code>Estimator<\/code>?<\/p>\n\n<hr>\n\n<h2>Some more details about my code:<\/h2>\n\n<p>I've defined my model (FeatureNet) like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/autograph\/examples\/notebooks\/rnn_keras_estimator.ipynb\" rel=\"nofollow noreferrer\">here<\/a> as an inheritance of <code>tf.keras.Model<\/code> with <code>init<\/code> and <code>call<\/code> method.<\/p>\n\n<p>My <code>model_fn<\/code> looks like this:<\/p>\n\n<pre><code>def model_fn(features, labels, mode):\n\n    resize_shape = (180, 320)\n    num_dimensions = 16\n\n    model = featurenet.FeatureNet(resize_shape, num_dimensions=num_dimensions)\n\n    training = (mode == tf.estimator.ModeKeys.TRAIN)\n    seg_pred = model(features, training)\n\n    predictions = {\n       # Generate predictions (for PREDICT mode)\n       \"seg_pred\": seg_pred\n    }\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)\n\n    # Calculate Loss (for both TRAIN and EVAL modes)\n    seg_loss = tf.reduce_mean(tf.keras.backend.binary_crossentropy(labels['seg_true'], seg_pred))\n    loss = seg_loss\n\n    # Configure the Training Op (for TRAIN mode)\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        optimizer = tf.train.MomentumOptimizer(learning_rate=1e-4, momentum=0.9)\n\n        train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n\n        return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)\n\n    # Add evaluation metrics (for EVAL mode)\n    return tf.estimator.EstimatorSpec(mode=mode, loss=loss)\n<\/code><\/pre>\n\n<p>Then in the main-part I train and evaluate with an custom Estimator:<\/p>\n\n<pre><code># Create the Estimator\nestimator = tf.estimator.Estimator(\n    model_fn=model_fn,\n    model_dir=\"\/tmp\/discriminative_model\"\n    )\n\ndef input_fn():\n    features, labels = create_synthetic_image()\n\n    training_data = tf.data.Dataset.from_tensors((features, labels))\n    training_data = training_data.repeat(None)\n    training_data = training_data.batch(1)\n    training_data = training_data.prefetch(1)\n    return training_data\n\nestimator.train(input_fn=lambda: input_fn(), steps=2000)\neval_results = estimator.evaluate(input_fn=lambda: input_fn(), steps=50)\nprint('Eval loss at step %d: %s' % (eval_results['global_step'], eval_results['loss']))\n<\/code><\/pre>\n\n<p>Where <code>create_synthetic_image<\/code> creates the same image\/label every time.<\/p>\n","answers":[{"AnswerId":"53100400","CreationDate":"2018-11-01T11:29:59.973","ParentId":null,"OwnerUserId":"10429267","Title":null,"Body":"<p>I've found, that the handling of <code>BatchNormalization<\/code> can cause such errors, like described <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16455\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The usage of the <code>control_dependencies<\/code> in the <code>model-fn<\/code> solved the issue for me (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19643\" rel=\"nofollow noreferrer\">see here<\/a>).<\/p>\n\n<pre><code>if mode == tf.estimator.ModeKeys.TRAIN:\n    optimizer = tf.train.MomentumOptimizer(learning_rate=1e-4, momentum=0.9)\n\n    with tf.control_dependencies(model.get_updates_for(features)):\n        train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())\n\n    return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53090570,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-10-31T19:20:49.903","AcceptedAnswerId":null,"OwnerUserId":887074.0,"Title":"Pytorch: set indices in a particular Tensor dimension (similar to torch.index_select)","Body":"<p>I'm trying to get and set indices in a particular tensor dimension without reshaping if possible. I've been able to find the <code>torch.index_select<\/code> function which does what I want when getting values, but I haven't found an analogous setter function yet. Does one exist?<\/p>\n\n<p>For context I have a tensor and a set of indexes<\/p>\n\n<pre><code>class_energy = torch.rand(3, 10, 32, 32)\nclass_logits = torch.empty_like(class_energy)\nidxs = [2, 3, 5, 7]\n<\/code><\/pre>\n\n<p>I want to access items at those indices <em>in a particular dimension<\/em> so I can perform a log_softmax. <\/p>\n\n<p>If I knew dim a-priori then I could simply use the fancy <code>__getitem__<\/code> \/ <code>__setitem__<\/code> syntax: For example, if <code>dim=1<\/code>, then <code>class_energy[:, idxs]<\/code>. Likewise, if <code>dim=2<\/code> -> <code>class_energy[:, :, idxs]<\/code>, <code>dim=0<\/code> -> <code>class_energy[idxs]<\/code>, etc...<\/p>\n\n<p>In the case where <code>dim=1<\/code>, I essentially want this:<\/p>\n\n<pre><code>class_logits[:, idxs] = F.log_softmax(class_energy[:, idxs], dim=1)\n<\/code><\/pre>\n\n<p>Unfortunately, I don't know the value of <code>dim<\/code> ahead of time. Sure I could build up the fancy index ahead of time via:<\/p>\n\n<pre><code>fancy_index = tuple([slice(None)] * dim + [idxs])\nclass_logits[fancy_index] = F.log_softmax(class_energy[fancy_index], dim=dim)\n<\/code><\/pre>\n\n<p>However, I'm wondering if there is a better way to do this. For the case of <code>__getitem__<\/code>, I know for a fact that there is. The following code using <code>torch.index_select<\/code> is equivalent<\/p>\n\n<pre><code>fancy_index = tuple([slice(None)] * dim + [idxs])\nindex = torch.LongTensor(idxs).to(class_energy.device)\nclass_logits[fancy_index] = F.log_softmax(torch.index_select(class_energy, dim=dim, index=index, dim=dim))\n<\/code><\/pre>\n\n<p>Not only is it functionally the same, index_select is much faster (I've seen 2x improvement) than using the fancy getitem syntax. <\/p>\n\n<p>My issue is that I can't seem to similar functionality for the <code>__setitem__<\/code> portion of the code. It would be really nice if I could get rid of the fancy index all together. I've looked into <code>Tensor.put_<\/code>, <code>torch.index_put<\/code>, and <code>torch.select<\/code>, but none of these seem to have the functionality I want. Is there something I'm missing, or is the fancy index the only way to solve this problem currently?<\/p>\n","answers":[]}
+{"QuestionId":53090781,"AnswerCount":1,"Tags":"<python><tensorflow><pytorch>","CreationDate":"2018-10-31T19:36:36.077","AcceptedAnswerId":53091645.0,"OwnerUserId":10454184.0,"Title":"what is the tensorflow equivalent for pytorch probability function: torch.bernoulli?","Body":"<p>In Pytorch, you can do following:<\/p>\n\n<pre><code>x = torch.bernoulli(my_data)\n<\/code><\/pre>\n\n<p>Any similar functionality in tensorflow? Can the input be 2-D tensor, such as (batch, len)?<\/p>\n\n<h1>I tried tensorflow.contrib.distributions.Bernoulli:<\/h1>\n\n<pre><code>import numpy as np\ntmp_x1 = np.random.rand(20,5) \nnew_data_2 = tf.convert_to_tensor(tmp_x1)\nfrom tensorflow.contrib.distributions import  Bernoulli\ntmp2_x1 = Bernoulli(probs=new_data_2)\n<\/code><\/pre>\n\n<h1>got error:<\/h1>\n\n<pre><code>return math_ops.log(probs) - math_ops.log1p(-1. * probs), probs\nTypeError: unsupported operand type(s) for *: 'float' and 'Tensor'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53091645","CreationDate":"2018-10-31T20:40:31.427","ParentId":null,"OwnerUserId":"1310104","Title":null,"Body":"<p>It seems <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/distributions\/Bernoulli\" rel=\"nofollow noreferrer\"><code>tf.distributions.Bernoulli<\/code><\/a> does what you need.  The input can be an N-D tensor, which includes a 2D tensor.<\/p>\n\n<p><strong>EDIT: example use<\/strong><\/p>\n\n<p>After your comment, I tried the following, which worked for me (using tensorflow 1.11):<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow\nimport tensorflow as tf\nfrom tensorflow.distributions import  Bernoulli\n\ntmp_x1 = np.random.rand(20,5)\nnew_data_2 = tf.convert_to_tensor(tmp_x1)\ntmp2_x1 = Bernoulli(probs=new_data_2)\nsess = tf.Session()\nprint sess.run(tmp2_x1.sample())\n<\/code><\/pre>\n"}]}
+{"QuestionId":53090796,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-10-31T19:38:02.340","AcceptedAnswerId":null,"OwnerUserId":8157810.0,"Title":"Keras ImageDataGenerator saturating 16-bit images","Body":"<p>I have 16-bit images (grayscale if that's relevant) that I'm trying to augment using Keras' <code>ImageDataGenerator<\/code>. For now, I'm using <code>flow()<\/code> to take a single test image and save out samples of the augmented images (ultimately I'll use <code>flow_from_directory()<\/code>).<\/p>\n\n<p>The problem is that the images that get saved are always saturated, i.e. most pixels in the image have a value of (255, 255, 255). This is despite the original image having a max value of ~6000 out of 65535. My expectation would be for the max value of the augmented images to be ~23.<\/p>\n\n<p>I definitely don't want to be feeding my model saturated images, since that means a lot of the dynamic range is lost. Any ideas what's going on here?<\/p>\n\n<p>Here's the code I'm using:<\/p>\n\n<pre><code>from keras.preprocessing.image import ImageDataGenerator, img_to_array, load_img\n\ntest_img = img_to_array(load_img('image10.tif'))\n\ntest_img = test_img.reshape((1,) + test_img.shape) \n\ntrain_datagen = ImageDataGenerator(\n    rotation_range = 90,\n    horizontal_flip = True,\n    vertical_flip = True,\n    fill_mode = 'constant',\n    rescale = 1.\/65535)\n\ni = 0\nfor batch in train_datagen.flow(test_img_array, batch_size=1, save_to_dir=preview_output, save_prefix='test_img', save_format='tif'):\n    i += 1\n    if i &gt; 10:\n        break\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53090979,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-10-31T19:51:38.560","AcceptedAnswerId":null,"OwnerUserId":8901144.0,"Title":"Stateful Bidirectional LSTM batch size problem","Body":"<p>I am trying to implement a stateful Bidirectional LSTM but I continue to get the following error:<\/p>\n\n<p>\" Cannot feed value of shape (72, 24, 8) for Tensor 'bidirectional_18_input:0', which has shape '(96, 24, 8) \" <\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>mdl = Sequential()\n# create and fit the LSTM network\nmdl.add(Bidirectional(LSTM(12 ,return_sequences=True, stateful=True),batch_input_shape=(96, lags, n_features )))\nmdl.add(Bidirectional(LSTM(7,stateful=True)))\nmdl.add(Dense(12))\nmdl.add(Dense(12))\nmdl.add(Dense(1))\n\n\nmdl.compile(loss='mean_squared_error', optimizer='adam')\n\n# early stopping implementation\nfilepath=\"weights.best.hdf5\"\ncheckpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min')\nearly_stop = EarlyStopping(monitor='val_loss', patience=200, mode='min') \ncallbacks_list = [checkpoint, early_stop]\n\n\n# fit network\nprint(mdl.summary())\n\nfor i in range(100):\n    mdl.fit(X_train, y_train, epochs=1, batch_size=96, validation_data=(X_test, y_test), verbose=2, shuffle=False,\n                   callbacks=callbacks_list)\n    mdl.reset_states()\n<\/code><\/pre>\n\n<p>I have the following architecture:\n<a href=\"https:\/\/i.stack.imgur.com\/1TmEc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1TmEc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I can't understand what is the why it says that I am feeding a value of shape (72,24,8) even if I am using a batch size of 96<\/p>\n","answers":[]}
+{"QuestionId":53091085,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-10-31T19:59:33.917","AcceptedAnswerId":53095409.0,"OwnerUserId":1496554.0,"Title":"raise ValueError('Image with id {} already added.'.format(image_id)) in Tensorflow object detection api","Body":"<p>Image training is ok with ssd_mobilenet_v1_coco in tensorflow object detection api.<\/p>\n\n<p>getting the error while testing:<\/p>\n\n<pre><code>File \"\/home\/hipstudents\/anaconda3\/envs\/tensorflow_gpuenv\/lib\/python3.6\/site-packages\/object_detection-0.1-py3.6.egg\/object_detection\/utils\/object_detection_evaluation.py\", line 203, in add_single_ground_truth_image_info\nraise ValueError('Image with id {} already added.'.format(image_id))\n<\/code><\/pre>\n\n<p>Please help.<\/p>\n\n<pre><code>System Info:\n\nWhat is the top-level directory of the model you are using: ~\/\nHave I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes, written scripts to convert .xml files to tf record \nOS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04\nTensorFlow installed from (source or binary): Compiled from source\nTensorFlow version (use command below): 1.11.0\nBazel version (if compiling from source): 0.16.1\nCUDA\/cuDNN version: 9.0.176, cuDNN: 9.0\nGPU model and memory: GeForce GTX1080Ti, 11GB\nExact command to reproduce: python eval.py --logtostderr --pipeline_config_path=training\/ssd_mobilenet_v1_coco.config --checkpoint_dir=training\/ --eval_dir=eval\/\n<\/code><\/pre>\n\n<p>I created dataset manually. Then label it using labelimg. after labeling I created csv file for image annotation and file name. then I create tf record. I follow this tutorial: <a href=\"https:\/\/towardsdatascience.com\/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9\" rel=\"noreferrer\">https:\/\/towardsdatascience.com\/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9<\/a><\/p>\n\n<p>My tfrecord generator for training and testing image:<\/p>\n\n<pre><code>\"\"\"\nUsage:\n  # From tensorflow\/models\/\n  # Create train data:\n  python generate_tfrecord.py --csv_input=data\/train_labels.csv  --output_path=train.record\n  # Create test data:\n  python generate_tfrecord.py --csv_input=data\/test_labels.csv  --output_path=test.record\n\"\"\"\nfrom __future__ import division\nfrom __future__ import print_function\nfrom __future__ import absolute_import\n\nimport os\nimport io\nimport pandas as pd\nimport tensorflow as tf\n\nfrom PIL import Image\nfrom object_detection.utils import dataset_util\nfrom collections import namedtuple, OrderedDict\n\nflags = tf.app.flags\nflags.DEFINE_string('csv_input', '', 'Path to the CSV input')\nflags.DEFINE_string('output_path', '', 'Path to output TFRecord')\nFLAGS = flags.FLAGS\n\n\n# TO-DO replace this with label map\ndef class_text_to_int(row_label):\n    if row_label == 'Field':\n        return 1\n    else:\n        None\n\n\ndef split(df, group):\n    data = namedtuple('data', ['filename', 'object'])\n    gb = df.groupby(group)\n    return [data(filename, gb.get_group(x)) for filename, x in zip(gb.groups.keys(), gb.groups)]\n\n\ndef create_tf_example(group, path):\n    with tf.gfile.GFile(os.path.join(path, '{}'.format(group.filename)), 'rb') as fid:\n        encoded_jpg = fid.read()\n    encoded_jpg_io = io.BytesIO(encoded_jpg)\n    image = Image.open(encoded_jpg_io)\n    width, height = image.size\n\n    filename = group.filename.encode('utf8')\n    image_format = b'jpg'\n    xmins = []\n    xmaxs = []\n    ymins = []\n    ymaxs = []\n    classes_text = []\n    classes = []\n\n    for index, row in group.object.iterrows():\n        xmins.append(row['xmin'] \/ width)\n        xmaxs.append(row['xmax'] \/ width)\n        ymins.append(row['ymin'] \/ height)\n        ymaxs.append(row['ymax'] \/ height)\n        classes_text.append(row['class'].encode('utf8'))\n        classes.append(class_text_to_int(row['class']))\n\n    tf_example = tf.train.Example(features=tf.train.Features(feature={\n        'image\/height': dataset_util.int64_feature(height),\n        'image\/width': dataset_util.int64_feature(width),\n        'image\/filename': dataset_util.bytes_feature(filename),\n        'image\/source_id': dataset_util.bytes_feature(filename),\n        'image\/encoded': dataset_util.bytes_feature(encoded_jpg),\n        'image\/format': dataset_util.bytes_feature(image_format),\n        'image\/object\/bbox\/xmin': dataset_util.float_list_feature(xmins),\n        'image\/object\/bbox\/xmax': dataset_util.float_list_feature(xmaxs),\n        'image\/object\/bbox\/ymin': dataset_util.float_list_feature(ymins),\n        'image\/object\/bbox\/ymax': dataset_util.float_list_feature(ymaxs),\n        'image\/object\/class\/text': dataset_util.bytes_list_feature(classes_text),\n        'image\/object\/class\/label': dataset_util.int64_list_feature(classes),\n    }))\n    return tf_example\n\n\ndef main(_):\n    writer = tf.python_io.TFRecordWriter(FLAGS.output_path)\n    path = os.path.join(os.getcwd(), 'Images')\n    examples = pd.read_csv(FLAGS.csv_input)\n    grouped = split(examples, 'filename')\n    for group in grouped:\n        tf_example = create_tf_example(group, path)\n        writer.write(tf_example.SerializeToString())\n\n    writer.close()\n    output_path = os.path.join(os.getcwd(), FLAGS.output_path)\n    print('Successfully created the TFRecords: {}'.format(output_path))\n\n\nif __name__ == '__main__':\n    tf.app.run()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53095409","CreationDate":"2018-11-01T04:57:38.303","ParentId":null,"OwnerUserId":"1496554","Title":null,"Body":"<p>In the ssd_mobilenet_coco_v1.config file, num_examples was 8000. In my case, test dataset only has 121 samples. I forgot to update that and got new kind of error that I couldn't find on the Internet. As it is a silly mistake, so I think a very few people did that. this answer might help someone who will do this kind of mistake. I changed the following in the config file and the error is resolved:     <\/p>\n\n<pre><code>eval_config: {\n        #num of test images. In my case 121. Previously It was 8000\n        num_examples: 121\n        # Note: The below line limits the evaluation process to 10 evaluations.\n        # Remove the below line to evaluate indefinitely.\n        max_evals: 10\n      }\n<\/code><\/pre>\n"},{"AnswerId":"58090052","CreationDate":"2019-09-25T01:38:16.963","ParentId":null,"OwnerUserId":"1134841","Title":null,"Body":"<p>In my case, the problem was that I'd included images multiple times when constructing  tfrecord files. Though now obvious, I hadn't noticed that many categories of the Open Images Dataset share the same images (which would have the same id in the evaluation, thus the error...). Once I'd corrected the algorithm creating the tfrecords, the error was gone. <\/p>\n"}]}
+{"QuestionId":53091121,"AnswerCount":1,"Tags":"<python><machine-learning><keras><activation-function>","CreationDate":"2018-10-31T20:01:32.647","AcceptedAnswerId":53091486.0,"OwnerUserId":1298461.0,"Title":"How to plot keras activation functions in a notebook","Body":"<p>I wanted to plot all Keras activation functions but some of them are not working. i.e. <code>linear<\/code> throws an error: <\/p>\n\n<blockquote>\n  <p>AttributeError: 'Series' object has no attribute 'eval'<\/p>\n<\/blockquote>\n\n<p>which is weird. How can I plot the rest of my activation functions?<\/p>\n\n<pre><code>points = 100\nzeros = np.zeros((points,1))\n\ndf = pd.DataFrame({\"activation\": np.linspace(-1.2,1.2,points)})\ndf[\"softmax\"] = K.eval(activations.elu(df[\"activation\"]))\n#df[\"linear\"] = K.eval(activations.linear(df[\"activation\"]))\ndf[\"tanh\"] = K.eval(activations.tanh(df[\"activation\"]))\ndf[\"sigmoid\"] = K.eval(activations.sigmoid(df[\"activation\"]))\ndf[\"relu\"] = K.eval(activations.relu(df[\"activation\"]))\n#df[\"hard_sigmoid\"] = K.eval(activations.hard_sigmoid(df[\"activation\"]))\n#df[\"exponential\"] = K.eval(activations.exponential(df[\"activation\"]))\ndf[\"softsign\"] = K.eval(activations.softsign(df[\"activation\"]))\ndf[\"softplus\"] = K.eval(activations.softplus(df[\"activation\"]))\n#df[\"selu\"] = K.eval(activations.selu(df[\"activation\"]))\ndf[\"elu\"] = K.eval(activations.elu(df[\"activation\"]))\n\ndf.plot(x=\"activation\", figsize=(15,15))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53091486","CreationDate":"2018-10-31T20:30:18.633","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>That's because the <code>linear<\/code> activation returns the input without any modifications:<\/p>\n\n<pre><code>def linear(x):\n    \"\"\"Linear (i.e. identity) activation function.\n    \"\"\"\n    return x\n<\/code><\/pre>\n\n<p>Since you are passing a Pandas Series as input, the same Pandas Series will be returned and therefore you don't need to use <code>K.eval()<\/code>:<\/p>\n\n<pre><code>df[\"linear\"] = activations.linear(df[\"activation\"])\n<\/code><\/pre>\n\n<p>As for the <code>selu<\/code> activation, you need to reshape the input to <code>(n_samples, n_output)<\/code>:<\/p>\n\n<pre><code>df[\"selu\"] = K.eval(activations.selu(df[\"activation\"].values.reshape(-1,1)))\n<\/code><\/pre>\n\n<p>And as for the <code>hard_sigmoid<\/code> activation, its input should be explicitly a Tensor which you can create using <code>K.variable()<\/code>:<\/p>\n\n<pre><code>df[\"hard_sigmoid\"] = K.eval(activations.hard_sigmoid(K.variable(df[\"activation\"].values)))\n<\/code><\/pre>\n\n<p>Further, <code>exponential<\/code> activation works as you have written and there is no need for modifications.<\/p>\n"}]}
+{"QuestionId":53091620,"AnswerCount":1,"Tags":"<python-3.x><python-2.7><tensorflow><deep-learning><keras-layer>","CreationDate":"2018-10-31T20:39:32.123","AcceptedAnswerId":null,"OwnerUserId":1410665.0,"Title":"TypeError:__init__() takes exactly 1 argument (3 given)","Body":"<p>I am facing <strong>TypeError:__init__() takes exactly 1 argument (3 given)<\/strong> error when calling below function:<\/p>\n\n<p><strong>Code Snippet:<\/strong><\/p>\n\n<p><code>super(abc, self).__init__(incoming, **kwargs);<\/code><\/p>\n","answers":[{"AnswerId":"53122351","CreationDate":"2018-11-02T16:21:24.170","ParentId":null,"OwnerUserId":"1410665","Title":null,"Body":"<p><strong>Before<\/strong><\/p>\n\n<pre><code>super(abc, self).__init__(incoming, **kwargs);\n<\/code><\/pre>\n\n<p>After modification where \"incoming\" need to be removed.<\/p>\n\n<pre><code>super(abc, self).__init__(**kwargs);\n<\/code><\/pre>\n"}]}
+{"QuestionId":53092712,"AnswerCount":0,"Tags":"<tensorflow><neural-network><keras><lstm><dropout>","CreationDate":"2018-10-31T22:16:22.557","AcceptedAnswerId":null,"OwnerUserId":10501472.0,"Title":"Subsequent dropout layers all connected to the first in keras?","Body":"<p>So I just installed TensorBoard and looked at a deep LSTM I have been working on and saw something on the graph that I wasn't expecting.  I have 5 hidden LSTM layers, each followed by a dropout layer, and it shows that the first dropout layer is connected directly to each one of the 4 following dropout layers:<\/p>\n\n<p>LSTM deep neural network graph:\n<a href=\"https:\/\/i.stack.imgur.com\/5SvWW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5SvWW.png\" alt=\"LSTM deep neural network graph\"><\/a><\/p>\n\n<p>What exactly is going on here?  I was expecting it all to just be in a simple straight line top to bottom.  Here's what my structure looks like in code, I didn't change very many of the default parameters...  <\/p>\n\n<pre><code># dataset.NUM_STEPS = 64\n# dataset.NUM_FEATURES = 1109\n# len(dataset.note_to_one_hot) = 123\n# len(dataset.duration_to_one_hot) = 846\n# len(dataset.offset_to_one_hot) = 140\n\ninputs = keras.layers.Input(shape=(dataset.NUM_STEPS, dataset.NUM_FEATURES))\nx = keras.layers.LSTM(units=512, return_sequences=True)(inputs)\nx = keras.layers.Dropout(.333)(x)\nx = keras.layers.LSTM(units=512, return_sequences=True)(x)\nx = keras.layers.Dropout(.333)(x)\nx = keras.layers.LSTM(units=512, return_sequences=True)(x)\nx = keras.layers.Dropout(.333)(x)\nx = keras.layers.LSTM(units=512, return_sequences=True)(x)\nx = keras.layers.Dropout(.333)(x)\nx = keras.layers.LSTM(units=512)(x)\nx = keras.layers.Dropout(.333)(x)\n\nnote_output = keras.layers.Dense(name=\"n\", units=len(dataset.note_to_one_hot), activation='softmax')(x)\nduration_output = keras.layers.Dense(name=\"d\", units=len(dataset.duration_to_one_hot), activation='softmax')(x)\noffset_output = keras.layers.Dense(name=\"o\", units=len(dataset.offset_to_one_hot), activation='softmax')(x)\n\nmodel = keras.models.Model(name=model_name, inputs=inputs, outputs=[note_output, duration_output, offset_output])\noptimizer = keras.optimizers.RMSprop(lr=6.66e-5, rho=0.9, epsilon=None, decay=0.0)\nlosses = {\"n\": \"categorical_crossentropy\", \"d\": \"categorical_crossentropy\", \"o\": \"categorical_crossentropy\"}\nmetrics = {\"n\": \"categorical_accuracy\", \"d\": \"categorical_accuracy\", \"o\": \"categorical_accuracy\"}\nmodel.compile(optimizer=optimizer, loss=losses, metrics=metrics)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53092971,"AnswerCount":3,"Tags":"<tensorflow><keras><deep-learning><vgg-net><image-preprocessing>","CreationDate":"2018-10-31T22:45:43.660","AcceptedAnswerId":53094188.0,"OwnerUserId":3731622.0,"Title":"Keras VGG16 preprocess_input modes","Body":"<p>I'm using the <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"noreferrer\">Keras VGG16 model<\/a>.<\/p>\n\n<p>I've seen it there is a <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/applications\/vgg16.py\" rel=\"noreferrer\">preprocess_input method to use in conjunction with the VGG16 model<\/a>.  This method appears to call the <a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/imagenet_utils.py\" rel=\"noreferrer\">preprocess_input method in imagenet_utils.py<\/a> which (depending on the case) calls <a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/imagenet_utils.py\" rel=\"noreferrer\">_preprocess_numpy_input method in imagenet_utils.py<\/a>.<\/p>\n\n<p>The <code>preprocess_input<\/code> has a <code>mode<\/code> argument which expects \"caffe\", \"tf\", or \"torch\".  If I'm using the model in Keras with TensorFlow backend, should I absolutely use <code>mode=\"tf\"<\/code>?<\/p>\n\n<p>If yes, is this because the VGG16 model loaded by Keras was trained with images which underwent the same preprocessing (i.e. changed input image's range from [0,255] to input range [-1,1])?<\/p>\n\n<p>Also, should the input images for testing mode also undergo this preprocessing? I'm confident the answer to the last question is yes, but I would like some reassurance.<\/p>\n\n<p>I would expect Francois Chollet to have done it correctly, but looking at <a href=\"https:\/\/github.com\/fchollet\/deep-learning-models\/blob\/master\/vgg16.py\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/deep-learning-models\/blob\/master\/vgg16.py<\/a> either he is or I am wrong about using <code>mode=\"tf\"<\/code>.<\/p>\n\n<p><strong>Updated info<\/strong><\/p>\n\n<p>@FalconUA directed me to the <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/research\/very_deep\/\" rel=\"noreferrer\">VGG at Oxford<\/a> which has a <strong>Models<\/strong> section with links for the 16-layer model.  The information about the <code>preprocessing_input<\/code> <code>mode<\/code> argument <code>tf<\/code> scaling to -1 to 1 and <code>caffe<\/code> subtracting some mean values is found by following the link in the <strong>Models<\/strong> 16-layer model: <a href=\"https:\/\/gist.github.com\/ksimonyan\/211839e770f7b538e2d8#file-readme-md\" rel=\"noreferrer\">information page<\/a>. In the Description section it says:<\/p>\n\n<blockquote>\n  <p>\"In the paper, the model is denoted as the configuration D trained with scale jittering. The input images should be zero-centered by mean pixel (rather than mean image) subtraction. Namely, the following BGR values should be subtracted: [103.939, 116.779, 123.68].\"<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"54483545","CreationDate":"2019-02-01T16:34:00.220","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>Trying to use VGG16 myself again lately, i had troubles getting descent results by just importing <code>preprocess_input<\/code> from vgg16 like this:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16, preprocess_input\n<\/code><\/pre>\n\n<p>Doing so, preprocess_input by default is set to <code>'caffe'<\/code> mode but having a closer look at <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/2.1.6\/keras\/applications\/vgg16.py\" rel=\"nofollow noreferrer\">keras vgg16 code<\/a>, i noticed that weights name <\/p>\n\n<blockquote>\n  <p>'<a href=\"https:\/\/github.com\/fchollet\/deep-learning-models\/releases\/download\/v0.1\/vgg16_weights_tf_dim_ordering_tf_kernels.h5\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/deep-learning-models\/releases\/download\/v0.1\/vgg16_weights_tf_dim_ordering_tf_kernels.h5<\/a>'<\/p>\n<\/blockquote>\n\n<p>is referring to tensorflow twice. I think that preprocess mode should be <code>'tf'<\/code>.<\/p>\n\n<pre><code>processed_img = preprocess_input(img, mode='tf')\n<\/code><\/pre>\n"},{"AnswerId":"53094218","CreationDate":"2018-11-01T01:55:41.970","ParentId":null,"OwnerUserId":"4987390","Title":null,"Body":"<p>In my experience in training VGG16 in Keras, the inputs should be from 0 to 255, subtracting the mean <code>[103.939, 116.779, 123.68]<\/code>. I've tried transfer learning (freezing the bottom and stack a classifier on top) with inputs centering from <code>-1<\/code> to <code>1<\/code>, and the results are much worse than <code>0..255 - [103.939, 116.779, 123.68]<\/code>.<\/p>\n"},{"AnswerId":"53094188","CreationDate":"2018-11-01T01:51:09.460","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<p>The <code>mode<\/code> here is not about the backend, but rather about <em>on what framework the model was trained on and ported from.<\/em> In the <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"noreferrer\">keras link<\/a> to VGG16, it is stated that:<\/p>\n\n<blockquote>\n  <p>These weights are ported from the ones released by <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/research\/very_deep\/\" rel=\"noreferrer\">VGG at Oxford<\/a><\/p>\n<\/blockquote>\n\n<p>So the VGG16 and VGG19 models were trained in Caffe and ported to TensorFlow, hence <code>mode == 'caffe'<\/code> here (range from 0 to 255 and then extract the mean <code>[103.939, 116.779, 123.68]<\/code>).<\/p>\n\n<p>Newer networks, like <em>MobileNet<\/em> and <em>ShuffleNet<\/em> were trained on TensorFlow, so <code>mode<\/code> is <code>'tf'<\/code> for them and the inputs are zero-centered in the range from -1 to 1.<\/p>\n"}]}
+{"QuestionId":53093542,"AnswerCount":1,"Tags":"<python><neural-network><keras><lstm><loss-function>","CreationDate":"2018-11-01T00:07:19.383","AcceptedAnswerId":53094215.0,"OwnerUserId":3303844.0,"Title":"Keras official example of LSTM classifier using real values for training target?","Body":"<p>From the official example in Keras docs, the stacked LSTM classifier is trained using <code>categorical_crossentropy<\/code> as a loss function, as expected.  <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/#examples\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/sequential-model-guide\/#examples<\/a><\/p>\n\n<p>But the <code>y_train<\/code> values are seeded using <code>numpy.random.random()<\/code> which outputs real numbers, versus <code>0,1<\/code> binary classification ( which is typical )   <\/p>\n\n<p>Are the <code>y_train<\/code> values being promoted to <code>0,1<\/code> values under the hood? <\/p>\n\n<p>Can you even train this loss function against real values between <code>0,1<\/code> ? <\/p>\n\n<p>How is <code>accuracy<\/code> then calculated ?<\/p>\n\n<p>Confusing.. no?<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import LSTM, Dense\nimport numpy as np\n\ndata_dim = 16\ntimesteps = 8\nnum_classes = 10\n\n# expected input data shape: (batch_size, timesteps, data_dim)\nmodel = Sequential()\nmodel.add(LSTM(32, return_sequences=True,\n               input_shape=(timesteps, data_dim)))  # returns a sequence of vectors of dimension 32\nmodel.add(LSTM(32, return_sequences=True))  # returns a sequence of vectors of dimension 32\nmodel.add(LSTM(32))  # return a single vector of dimension 32\nmodel.add(Dense(10, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='rmsprop',\n              metrics=['accuracy'])\n\n# Generate dummy training data\nx_train = np.random.random((1000, timesteps, data_dim))\ny_train = np.random.random((1000, num_classes))\n\n# Generate dummy validation data\nx_val = np.random.random((100, timesteps, data_dim))\ny_val = np.random.random((100, num_classes))\n\nmodel.fit(x_train, y_train,\n          batch_size=64, epochs=5,\n          validation_data=(x_val, y_val))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53094215","CreationDate":"2018-11-01T01:55:21.277","ParentId":null,"OwnerUserId":"3963013","Title":null,"Body":"<p>For this example, the y_train and y_test are not the one-hot encoding anymore, but the probabilities of each classes. So it is still applicable for cross-entropy. And we can treat the one-hot encoding as the special case of the probabilities vector.<\/p>\n\n<pre><code>y_train[0]\narray([0.30172708, 0.69581121, 0.23264601, 0.87881279, 0.46294832,\n       0.5876406 , 0.16881395, 0.38856604, 0.00193709, 0.80681196])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53093948,"AnswerCount":0,"Tags":"<python><tensorflow><gradient>","CreationDate":"2018-11-01T01:10:51.840","AcceptedAnswerId":null,"OwnerUserId":1986318.0,"Title":"How to enable gradient flow in scatter_update?","Body":"<p>I am trying to compute the local variance map of an image by taking data from all possible window of fixed-size (eg 5x5), inside a training loop. To vectorize this operation I am thinking about expanding the original image with an operation similar to <a href=\"https:\/\/imgur.com\/Ic3IL5Q\" rel=\"nofollow noreferrer\">this<\/a> using <code>scatter_update\/scatter_nd_update<\/code> inside the training loop. What this operation  essentially does is to map each element in the original tensor to potentially many locations in the new tensor, and the locations are computed inside the training loop.<\/p>\n\n<p>However, <code>scatter_update<\/code> does not allow gradient propagation, and my attempt at creating a simple custom gradient for the <code>scatter_update<\/code> did not work.<\/p>\n\n<pre><code>@tf.RegisterGradient(\"CustomGrad\")\ndef _clip_grad(unused_op, grad):\n  return tf.constant(5., dtype=tf.float32, shape=(1)) # tf.clip_by_value(grad, -0.1, 0.1)\n\nx = tf.Variable([3.0], dtype=tf.float32)\ny = tf.get_variable('y', shape=(1), dtype=tf.float32)\n\ng = tf.get_default_graph()\nwith g.gradient_override_map({\"ScatterNdUpdate1\": \"CustomGrad\"}):\n    output = tf.scatter_nd_update(y, [[0]], x, name=\"ScatterNdUpdate1\")\ngrad_custom = tf.gradients(output, y)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(grad_custom)\n<\/code><\/pre>\n\n<p>Running the code above shows that <code>grad_custom<\/code> contains <code>None<\/code>. Does any one have any idea of how to properly implement a local variance map that can be used in the training loop? Solving the gradient problem would also help me with another problem I am having.<\/p>\n","answers":[]}
+{"QuestionId":53094060,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-01T01:28:53.460","AcceptedAnswerId":null,"OwnerUserId":10523094.0,"Title":"Define a custom accuracy in Keras to ignore samples with a particular label","Body":"<p>I am writing to you because I have a problem, the same as the question in the link below that has not been answered.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47270722\/how-to-define-a-custom-accuracy-in-keras-to-ignore-samples-with-a-particular-gol\/53093965?noredirect=1#comment93084625_53093965\">Previous question<\/a><\/p>\n\n<p>I have defined a loss function that excludes all pixel = 0:<\/p>\n\n<pre><code>def ignore_unknown_xentropy(ytrue, ypred):\n    return (1-ytrue[:, :, :, 0]) * K.categorical_crossentropy(ytrue, ypred)\n<\/code><\/pre>\n\n<p>I noticed that during the training accuracy is distorted, going to consider the pixels equal to zero that I had previously excluded.<\/p>\n\n<p>There is a way to define a custom metric to exclude zero pixels on <code>y_true<\/code> from its calculation?<\/p>\n\n<p>I'm using this approach, that is to calculate the accuracy for each relevant label and add them, but it does not give me reliable results:<\/p>\n\n<pre><code>def single_class_accuracy(y_true, y_pred, class_id):\n    class_id_true = K.argmax(y_true, axis=-1)\n    class_id_preds = K.argmax(y_pred, axis=-1)\n    accuracy_mask = K.cast(K.equal(class_id_preds, class_id), 'int32')\n    class_acc_tensor = K.cast(K.equal(class_id_true, class_id_preds), 'int32') * accuracy_mask\n    class_acc = K.sum(class_acc_tensor) \/ K.maximum(K.sum(accuracy_mask), 1)\n    return class_acc\n\ndef custom_accuracy(y_true, y_pred):\n    return single_class_accuracy(y_true, y_pred, 1) + single_class_accuracy(y_true, y_pred, 2)\n<\/code><\/pre>\n\n<p>And my <code>ytrue<\/code> is a one-hot encoded mask.<\/p>\n","answers":[{"AnswerId":"53099705","CreationDate":"2018-11-01T10:47:08.553","ParentId":null,"OwnerUserId":"9545636","Title":null,"Body":"<p>Is there no way to achieve this just with class weighting?\nSet 0 for the class you are not interested in?\nI may have misunderstood, but this approach lept out to me (ignore class zero):<\/p>\n\n<pre><code>class_weights = {0.0:0.2,0.2:0.2:0.2}\nhistory=model.fit(x=in_train,y=target_train,class_weight=class_weights,\n              validation_data=.... etc\n<\/code><\/pre>\n"}]}
+{"QuestionId":53094100,"AnswerCount":1,"Tags":"<python><tensorflow><matplotlib><deep-learning><conv-neural-network>","CreationDate":"2018-11-01T01:36:26.377","AcceptedAnswerId":null,"OwnerUserId":10588753.0,"Title":"Errno 13 while Transfer-learning Inception v3","Body":"<p>So I just learned python a couple of weeks ago, and I'm new to tensorflow. But I need to fine-tune the inception model. While trying to follow some tutorials, when I finally started to see some light, I got this horrible error that I can\u00b4t get rid of. Here is the part of the code that has the error:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\n\ntrain_dir=r'C:\\tf_files'\n\nimage_paths=train_dir\n\nimages = [plt.imread(train_dir) for path in image_paths]\n\ndef plot_images(images, cls_true, cls_pred=None, smooth=True):\n    assert len(images) == len(cls_true)\n\n    # Create figure with sub-plots.\n    fig, axes = plt.subplots(3, 3)\n\n    # Adjust vertical spacing.\n    if cls_pred is None:\n        hspace = 0.3\n    else:\n        hspace = 0.6\n    fig.subplots_adjust(hspace=hspace, wspace=0.3)\n\n    # Interpolation type.\n    if smooth:\n        interpolation = 'spline16'\n    else:\n        interpolation = 'nearest'\n\n    for i, ax in enumerate(axes.flat):\n        # There may be less than 9 images, ensure it doesn't crash.\n        if i &lt; len(images):\n            # Plot image.\n            ax.imshow(images[i],\n                      interpolation=interpolation)\n\n            # Name of the true class.\n            cls_true_name = class_names[cls_true[i]]\n\n            # Show true and predicted classes.\n            if cls_pred is None:\n                xlabel = \"True: {0}\".format(cls_true_name)\n            else:\n                # Name of the predicted class.\n                cls_pred_name = class_names[cls_pred[i]]\n\n                xlabel = \"True: {0}\\nPred: {1}\".format(cls_true_name, cls_pred_name)\n\n            # Show the classes as the label on the x-axis.\n            ax.set_xlabel(xlabel)\n\n        # Remove ticks from the plot.\n        ax.set_xticks([])\n        ax.set_yticks([])\n\n    # Ensure the plot is shown correctly with multiple plots\n    # in a single Notebook cell.\n    plt.show()\n<\/code><\/pre>\n\n<p>and this is what the terminal spits:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/Shangai\/PycharmProjects\/PSai\/Testing.py\", line 14, in &lt;module&gt;\n    images = [plt.imread(train_dir) for path in image_paths]\n  File \"C:\/Users\/Shangai\/PycharmProjects\/PSai\/Testing.py\", line 14, in &lt;listcomp&gt;\n    images = [plt.imread(train_dir) for path in image_paths]\n  File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\matplotlib\\pyplot.py\", line 2152, in imread\n    return matplotlib.image.imread(fname, format)\n  File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\matplotlib\\image.py\", line 1359, in imread\n    with Image.open(fname) as image:\n  File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\PIL\\Image.py\", line 2609, in open\n    fp = builtins.open(filename, \"rb\")\nPermissionError: [Errno 13] Permission denied: 'C:\\\\tf_files'\n<\/code><\/pre>\n\n<p>I basically need a way to make an image an array, but I can\u00b4t figure it out.\nAny help would be appreciated, thank you so much. Peace!<\/p>\n\n<p>-I'm using pycharm with an anaconda environment-<\/p>\n","answers":[{"AnswerId":"53094124","CreationDate":"2018-11-01T01:39:31.970","ParentId":null,"OwnerUserId":"4102479","Title":null,"Body":"<p>EDIT: You called <code>imread<\/code> for a directory, which is actually not an image. Are your images in <code>C:\\tf_files<\/code>? If so, use <code>os.listdir()<\/code> to get ALL (including non-image) files within a directory. So your code (should be) like this:  <\/p>\n\n<pre><code>train_dir=r'C:\\tf_files' # The directory contains images    \nimage_paths=os.listdir(train_dir)  # Get the image file only, you can check by checking its extension   \nimages = [plt.imread(train_dir) for path in image_paths]  \n<\/code><\/pre>\n\n<p>There is one more thing: The <code>Permission denied<\/code> error is raised, which is confused to me. If my answer help you solve the problem, which mean the error is not due to ownership, so I have no idea why this <code>Permission denied<\/code> happen. <\/p>\n\n<hr>\n\n<p>This is the problem <code>PermissionError: [Errno 13] Permission denied: 'C:\\\\tf_files'<\/code><br>\nDo you have access to <code>C:\\\\tf_files<\/code> directory? Try to gain ownership, or just edit the path to somewhere you have access<\/p>\n"}]}
+{"QuestionId":53094212,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T01:55:06.197","AcceptedAnswerId":53100658.0,"OwnerUserId":4258928.0,"Title":"Tensorflow SxN matrix multiply with SxD matrix to output SxNxD array","Body":"<p>I have a matrix SxN where each K_ij is a scalar <\/p>\n\n<pre><code>[[K_11,K_12,K_13,..,K_1N],\n [K_21,K_22,K_23,..,K_2N],\n ..,\n [K_S1,K_S2,K_S3,..,K_SN]]\n<\/code><\/pre>\n\n<p>Another matrix SxD such that each X_i is a Dx1 vector<\/p>\n\n<pre><code>[X_1,\n X_2,\n ...,\n X_S]\n<\/code><\/pre>\n\n<p>I want a results such that gives me <\/p>\n\n<pre><code>[[K_11 * X_1],[K_12 * X_1] ..., [K_1N * X_1],\n [K_21 * X_2],[K_22 * X_2] ..., [K_2N * X_2],\n [K_31 * X_3],[K_32 * X_2] ..., [K_3N * X_3],\n  ......,\n  [K_S1 * X_S],[K_S2 * X_S] ..., [K_SN * X_S]]\n<\/code><\/pre>\n\n<p>where * stands for multiply<\/p>\n\n<p>Any efficient way of doing this?<\/p>\n","answers":[{"AnswerId":"53100658","CreationDate":"2018-11-01T11:46:12.920","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think this what you want:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nK = tf.placeholder(tf.float32, [None, None])  # (S, N)\nX = tf.placeholder(tf.float32, [None, None])  # (S, D)\n# All-to-all product: (S, N, D)\nresult = K[:, :, tf.newaxis] * X[:, tf.newaxis, :]\n# Test\nwith tf.Session() as sess:\n    # (2, 2)\n    K_val = np.array([[10., 20.],\n                      [30., 40.]])\n    # (2, 3)\n    X_val = np.array([[1., 2., 3.],\n                      [4., 5., 6.]])\n    # (2, 2, 3)\n    result_val = sess.run(result, feed_dict={K: K_val, X: X_val})\n    print(result_val)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[ 10.  20.  30.]\n  [ 20.  40.  60.]]\n\n [[120. 150. 180.]\n  [160. 200. 240.]]]\n<\/code><\/pre>\n"},{"AnswerId":"53121877","CreationDate":"2018-11-02T15:53:13.577","ParentId":null,"OwnerUserId":"10597060","Title":null,"Body":"<p>Alternatively (using TF APIs):<\/p>\n\n<pre><code>tf.matmul(tf.expand_dims(in1, axis=2), tf.expand_dims(in2, axis=1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53094233,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T01:56:56.427","AcceptedAnswerId":53104797.0,"OwnerUserId":1610428.0,"Title":"Tensorflow: Does sequence matter in encoding and decoding a TFRecord","Body":"<p>I have some practice data that I want to encode to a <code>TFRecord<\/code> format and then decode to <code>tf.features<\/code> in Tensorflow. My question is very basic, but I could not find a clear answer to this. <\/p>\n\n<p><strong>Question<\/strong>: Do I need to decode the features in a dataset in the same sequence as they are encoded? In other words, I can't seem to find a way to reference features by field name in a TFRecord. This is really important for 2 reasons. <\/p>\n\n<ol>\n<li>I just wanted to get my assumption validated, so that I know how to avoid breaking my code in the future. Here is some simple code, though this is not a complete example.<\/li>\n<li>Python makes a big deal about dictionaries being <em>un-ordered<\/em>. So how can I guarantee sequence when I am using a data structure that is supposed to be unordered? I was not sure if this was handled in some way that I don't know about.<\/li>\n<\/ol>\n\n<p>To encode data into TFRecord format, you can do something like:<\/p>\n\n<pre><code>#Fields in Dataframe: ['DIVISION','SPORDER','PUMA','REGION']\n\ndf = pd.DataFrame(...)\nwith tf.python_io.TFRecordWriter('myfile.tfrecord') as writer:\n\n    for row in df.itertuples():\n        example = tf.train.Example(features=tf.train.Features(feature={\n          'feat\/division': tf.train.Feature(int64_list=tf.train.Int64List(value=row.DIVISION)),\n          'label\/sporder': tf.train.Feature(int64_list=tf.train.Int64List(value=row.SPORDER)),\n          'feat\/puma': tf.train.Feature(bytes_list=tf.train.BytesList(value=[row.PUMA])),\n          'feat\/region': tf.train.Feature(bytes_list=tf.train.BytesList(value=[row.REGION]))))\n        writer.write(example.SerializeToString())\n<\/code><\/pre>\n\n<p>Then to ingest the dataset you would need something like the code below. Notice that the fields are referenced again in order. NOTE: I used the same dictionary keys in the TFRecords versus the decoded form, but I don't think that is necessary--just a convenience. I was not sure if that is the way things have to be? Meaning,<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset('myfile.tfrecord')\ndataset = dataset.map(_parse_function)\n\ndef _parse_function(example_proto):\n    features = {'feat\/division': tf.FixedLenFeature((), tf.string, default_value=\"\"),\n                'label\/sporder': tf.FixedLenFeature((), tf.int64, default_value=0),\n                'feat\/puma': tf.VarLenFeature(dtype=tf.string),\n                'feat\/region': tf.VarLenFeature(dtype=tf.string)}\n\n    parsed_example = tf.parse_single_example(example_proto, features)\n    parsed_label = parsed_example.pop(\"label\/sporder\", None)\n\n\n    return parsed_example, parsed_label\n<\/code><\/pre>\n","answers":[{"AnswerId":"53104797","CreationDate":"2018-11-01T15:53:11.423","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>The <code>tfrecord<\/code> format uses <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/overview\" rel=\"nofollow noreferrer\">protobuf<\/a> for serialization of the struct. You can think about it as a binary json\/xml format. Json\/xml and protobuf don't care about the order of the fields. So, the order of the feature definitions is not important. It's the same in your snippet because it was just convenient for reading.<\/p>\n"}]}
+{"QuestionId":53094238,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nvidia>","CreationDate":"2018-11-01T01:57:38.767","AcceptedAnswerId":null,"OwnerUserId":8606370.0,"Title":"GPU crashes when running Keras\/tensorflow-gpu, specifically when clock speed goes to idle at 0 MHz","Body":"<p>I'm using Jupyter Notebook to run Keras with a Tensorflow GPU backend. I've done some testing with various dummy models while simultaneously monitoring my GPU usage using MSI Afterburner, GPU-Z, nvidia-smi and Task Manager. My GPU is a GeForce GTX 960M, which has no issues running games. The temperatures are also low when running Keras.<\/p>\n\n<p>What I've noticed is that the Keras runs fine (e.g. loading or training a model) in the beginning but whenever Keras is not running anything, the GPU naturally wants to idle from 1097 MHz to 0 MHz and as soon as it does that the GPU crashes. I can see that the \"GPU is lost\" on NVSMI. I have to then disable and re-enable my GPU in the Device Manager to get it to work.<\/p>\n\n<p>Does anyone have any idea why this might be happening?<\/p>\n\n<p>Edit: I can temporarily prevent this from happening for very small programs by using the \"allow_growth\" feature as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.backend.tensorflow_backend import set_session\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nsess = tf.Session(config=config)\nset_session(sess)\n<\/code><\/pre>\n\n<p>However, this only works if the operation is really small such that it uses only about 0.1 GB of GPU memory such as loading a model or running a really small model. However, if the program is using memory of even 0.3 GB of memory my GPU crashes since the memory does not go to 0 GB <em>before<\/em> the clock speed drops to 0 MHz (lower power state).<\/p>\n","answers":[{"AnswerId":"53542450","CreationDate":"2018-11-29T15:32:25.953","ParentId":null,"OwnerUserId":"8606370","Title":null,"Body":"<p>I was finally able to figure out the issue thanks to someone from another forum. It was a driver issue. The latest drivers provided by Nvidia are causing the issue unlike the old drivers provided by my laptop manufacturer.<\/p>\n\n<p>Since I was not able to run tensorflow with my old drivers and do more troubleshooting, what I did was download eDrawings Viewer and open up some random assembly drawings I found online. First I tried with the latest Nvidia drivers, and I see that when I manipulate the models, my card is at P0 state but if I don't do anything and let the software idle, my card goes to a lower power state and crashes my GPU. But when I did the same exercise with my ASUS manufacturer-certified drivers (since this software was compatible even with the older drivers unlike TF), my GPU did NOT crash.<\/p>\n\n<p>What I also discovered was that eDrawings Viewer does not crash even with the latest Nvidia drivers if I go into the Nvidia Control Panel and select \"Prefer Maximum Performance\" under Power Management Mode. The card stays at P0 state whenever I have the software open even after idling for minutes. Unfortunately, since python.exe does not have a graphical interface, this option does not work for my case. As a workaround, I can still run tensorflow without getting it to crash by running eDrawings Viewer in the background (or really any program that uses a graphical interface), which keeps my card at the P0 State.<\/p>\n"}]}
+{"QuestionId":53094383,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2018-11-01T02:17:12.840","AcceptedAnswerId":null,"OwnerUserId":10588870.0,"Title":"Tensorflow inference very slow on first time run in docker","Body":"<p>I built a docker image for a server that does inference with tensorflow. I installed tensorflow-gpu with pip in the docker image. It works fine for my machine with titan x gpus. But when I ran the docker container on another machine with 1080 ti gpus. The first run becomes incredibly slow, takes about 90 seconds, usually it takes 7 seconds on the first run and 1 second in the following runs. I tried to set tf_cudnn_use_autotune to 0, and also mount a folder to save the cuda cache. But it doesn't really solve the problem. Any one has any suggestion?<\/p>\n","answers":[{"AnswerId":"57246234","CreationDate":"2019-07-29T01:27:11.493","ParentId":null,"OwnerUserId":"7847748","Title":null,"Body":"<p>Here's <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/overview\" rel=\"nofollow noreferrer\">a link<\/a>.\nI find this.<\/p>\n\n<blockquote>\n  <p><em>After running TensorFlow once, the compiled kernels are cached by CUDA. If using a docker container, the data is not cached and the penalty is paid each time TensorFlow starts<\/em>.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53095061,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2018-11-01T04:04:03.333","AcceptedAnswerId":null,"OwnerUserId":9113303.0,"Title":"How to choose the right neural network in a binary classification problem for a unbalanced data?","Body":"<p>I am using <code>keras sequential<\/code> model for binary classification. But My data is unbalanced. I have 2 features column and 1 output column(1\/0). I have 10000 of data. Among that only 20 results in output 1, all others are 0. Then i have extended the data size to 40000. Now also only 20 results in output 1, all others are 0. Since the data is unbalanced(0 dominates 1), which neural network will be better for correct prediction?<\/p>\n","answers":[{"AnswerId":"53096595","CreationDate":"2018-11-01T07:06:43.980","ParentId":null,"OwnerUserId":"9041710","Title":null,"Body":"<p>First of all, two features is a really small amount. Neural Networks are highly non-linear models with a really really high amount of freedom degrees, thus if you try to train a network with more than just a couple of networks it will overfit even with balanced classes. You can find more suitable models for a small dimensionality like Support Vector Machines in scikit-learn library.<\/p>\n\n<p>Now about unbalanced data, the most common techniques are Undersampling and Oversampling. Undersampling is basically training your model several times with a fraction of the dataset, that contains the non dominant class and a random sample of the dominant so that the ratio is acceptable, where as oversampling consist on generating artificial data to balance the classes. In most cases undersampling works better.<\/p>\n\n<p>Also when working with unbalanced data it's quite important to choose the right metric based on what is more important for the problem (is minimizing false positives more important than false negatives, etc).<\/p>\n"}]}
+{"QuestionId":53095125,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-01T04:14:54.947","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"Why does Keras prepend path information to variable names?","Body":"<p>I wanted to make a copy of a Keras variable and assign it a name of my choosing. What I found was that Keras would prepend path information to the name I chose.<\/p>\n\n<pre><code>&gt;&gt;&gt; orig\n&lt;tf.Variable 'conv2d_1\/kernel:0' shape=(3, 3, 3, 32) dtype=float32_ref&gt;\n&gt;&gt;&gt; cpy = K.variable(K.get_value(orig), name=\"copy\")\n\n&gt;&gt;&gt; cpy.name\nu'training\/SGD0\/copy:0'\n&gt;&gt;&gt; orig.name\nu'conv2d_1\/kernel:0'\n<\/code><\/pre>\n\n<p>I don't understand why the path info was prepended to the name of my copied tensor, but was seemingly not prepended to the original tensor - orig. I'd like to know either the reason this happened, or how to avoid it.<\/p>\n","answers":[{"AnswerId":"53106706","CreationDate":"2018-11-01T17:50:59.223","ParentId":null,"OwnerUserId":"1245262","Title":null,"Body":"<p>For reasons I don't entirely understand, Tensorflow seems to look for a '\/' as the final character of a variable's name in order to ensure it's unique - lines 4106 - 4117 of tensorflow\/python\/framework\/ops.py:<\/p>\n\n<pre><code>try:\n  old_stack = self._name_stack\n  if not name:  # Both for name=None and name=\"\" we re-set to empty scope.\n    new_stack = None\n  elif name[-1] == \"\/\":\n    new_stack = _name_from_scope_name(name)\n  else:\n    new_stack = self.unique_name(name)\n  self._name_stack = new_stack\n  yield \"\" if new_stack is None else new_stack + \"\/\"\nfinally:\n  self._name_stack = old_stack\n<\/code><\/pre>\n\n<p>So, adding a '\/' to the end of the variable's name eliminates the path prepending, although the '\/' is replaced by a ':0'<\/p>\n\n<pre><code>&gt;&gt;&gt; orig\n&lt;tf.Variable 'conv2d_1\/kernel:0' shape=(3, 3, 3, 32) dtype=float32_ref&gt;\n&gt;&gt;&gt; cpy = K.variable(K.get_value(orig), name=\"copy\/\")\n&gt;&gt;&gt; cpy.name\nu'copy:0'\n<\/code><\/pre>\n"}]}
+{"QuestionId":53095362,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><nlp>","CreationDate":"2018-11-01T04:50:27.860","AcceptedAnswerId":53157245.0,"OwnerUserId":9633432.0,"Title":"Concatenating hidden units using Keras","Body":"<p>I am trying to concatenate the hidden units. For example, I have <code>3<\/code> units, <code>h1,h2,h3<\/code> then I want the new layer to have <code>[h1;h1],[h1;h2],[h1;h3],[h2;h1]...<\/code>.<\/p>\n\n<p>So, I have tried:<\/p>\n\n<pre><code>class MyLayer(Layer):\n    def __init__(self,W_regularizer=None,W_constraint=None, **kwargs):\n        self.init = initializers.get('glorot_uniform')\n        self.W_regularizer = regularizers.get(W_regularizer)\n        self.W_constraint = constraints.get(W_constraint)\n        super(MyLayer, self).__init__(**kwargs)\n\ndef build(self, input_shape):\n    assert len(input_shape) == 3\n    # Create a trainable weight variable for this layer.\n    self.W = self.add_weight((input_shape[-1],input_shape[-1]),\n                             initializer=self.init,\n                             name='{}_W'.format(self.name),\n                             regularizer=self.W_regularizer,\n                             constraint=self.W_constraint,\n                            trainable=True)\n    super(MyLayer, self).build(input_shape)\n\ndef call(self, x,input_shape):\n    conc=K.concatenate([x[:, :-1, :], x[:, 1:, :]],axis=1)# help needed here\n    uit = K.dot(conc, self.W)# W has input_shape[-1],input_shape[-1]\n    return uit\n\ndef compute_output_shape(self, input_shape):\n    return input_shape[0], input_shape[1],input_shape[-1]\n<\/code><\/pre>\n\n<p>I am not sure what should I return for the second argument of my output shape.<\/p>\n\n<pre><code>from keras.layers import Input, Lambda, LSTM\nfrom keras.models import Model\nimport keras.backend as K\nfrom keras.layers import Lambda\n\nlstm=LSTM(64, return_sequences=True)(input)\nsomething=MyLayer()(lstm)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53157245","CreationDate":"2018-11-05T15:22:32.900","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You could implement the concatenation operation that you described by leveraging <a href=\"https:\/\/docs.python.org\/3\/library\/itertools.html#itertools.product\" rel=\"nofollow noreferrer\">itertools.product<\/a> in order to compute the cartesian product of the temporal dimension's indices. The call method could be coded as follows:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>def call(self, x):\n    prod = product(range(nb_timesteps), repeat=2)\n    conc_prod = []\n    for i, j in prod:\n        c = K.concatenate([x[:, i, :], x[:, j, :]], axis=-1)  # Shape=(batch_size, 2*nb_features)\n        c_expanded = c[:, None, :]  # Shape=(batch_size, 1, 2*nb_features)\n        conc_prod.append(c_expanded)\n    conc = K.concatenate(conc_prod, axis=1)  # Shape=(batch_size, nb_timesteps**2, 2*nb_features)\n    uit = K.dot(conc, self.W)  # W has shape 2*input_shape[-1], input_shape[-1]\n    return uit  # Shape=(batch_size, nb_timesteps**2, nb_features)\n<\/code><\/pre>\n\n<p>In the <a href=\"https:\/\/stackoverflow.com\/questions\/53095362\/concatenating-hidden-units-using-keras?noredirect=1#comment93178283_53095362\">example<\/a> that you provided, <code>nb_timesteps<\/code> would be 3. Note also that the shape of the weights should be <code>(2*input_shape[-1], input_shape[-1])<\/code> for the dot product to be valid. <\/p>\n\n<p><strong>Disclaimer<\/strong>: I am not sure what you want to achieve.<\/p>\n"}]}
+{"QuestionId":53095901,"AnswerCount":1,"Tags":"<windows><tensorflow><build><bazel><intel-mkl>","CreationDate":"2018-11-01T05:59:27.797","AcceptedAnswerId":null,"OwnerUserId":10589283.0,"Title":"build tensorflow r1.12 mkl on Windows using bazel, result in link error","Body":"<p>I'm trying to compile tensorflow with MKL support on Windows 10 64bit via Bazel with command:<\/p>\n\n<pre><code>bazel build --config=mkl --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>on the following environments:<\/p>\n\n<ul>\n<li>Windows 10 64 bit<\/li>\n<li>Tensorflow 1.12<\/li>\n<li>Bazel 0.18<\/li>\n<\/ul>\n\n<p>I have no idea what's wrong with my config, I have tried tensorflow r1.11 and r1.12. both ending up with same problem, the log is:<\/p>\n\n<pre><code>    ERROR: F:\/tools\/tf1.12\/tensorflow\/tensorflow\/python\/BUILD:3766:1: Linking of rule '\/\/tensorflow\/python:_pywrap_tensorflow_internal.so' failed (Exit 1000): link.exe failed: error executing command\n  cd C:\/users\/10267\/_bazel_10267\/udaytyio\/execroot\/org_tensorflow\n  SET LIB=C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\LIB\\amd64;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\ATLMFC\\LIB\\amd64;C:\\Program Files (x86)\\Windows Kits\\10\\lib\\10.0.10240.0\\ucrt\\x64;C:\\Program Files (x86)\\Windows Kits\\NETFXSDK\\4.6.1\\lib\\um\\x64;C:\\Program Files (x86)\\Windows Kits\\8.1\\lib\\winv6.3\\um\\x64;\n    SET PATH=C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\Common7\\IDE\\CommonExtensions\\Microsoft\\TestWindow;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\BIN\\amd64;C:\\WINDOWS\\Microsoft.NET\\Framework64\\v4.0.30319;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\VCPackages;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\Common7\\IDE;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\Common7\\Tools;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\Team Tools\\Performance Tools\\x64;C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\Team Tools\\Performance Tools;C:\\Program Files (x86)\\Windows Kits\\8.1\\bin\\x64;C:\\Program Files (x86)\\Windows Kits\\8.1\\bin\\x86;C:\\Program Files (x86)\\Microsoft SDKs\\Windows\\v10.0A\\bin\\NETFX 4.6.1 Tools\\x64\\;;C:\\WINDOWS\\system32\n    SET PWD=\/proc\/self\/cwd\n    SET PYTHON_BIN_PATH=C:\/Program Files\/Python36\/python.exe\n    SET PYTHON_LIB_PATH=C:\/Program Files\/Python36\/lib\/site-packages\n    SET TEMP=C:\\Users\\10267\\AppData\\Local\\Temp\n    SET TF_DOWNLOAD_CLANG=0\n    SET TF_NEED_CUDA=0\n    SET TF_NEED_OPENCL_SYCL=0\n    SET TMP=C:\\Users\\10267\\AppData\\Local\\Temp\n  C:\/Program Files (x86)\/Microsoft Visual Studio 14.0\/VC\/bin\/amd64\/link.exe \/nologo \/DLL \/SUBSYSTEM:CONSOLE -defaultlib:advapi32.lib -DEFAULTLIB:advapi32.lib \/MACHINE:X64 @bazel-out\/x64_windows-opt\/bin\/tensorflow\/python\/_pywrap_tensorflow_internal.so-2.params \/OPT:ICF \/OPT:REF \/DEF:bazel-out\/x64_windows-opt\/genfiles\/tensorflow\/python\/pywrap_tensorflow_filtered_def_file.def \/ignore:4070\nLINK : warning LNK4044: unrecognized option '\/lpthread'; ignored\n\nbazel-out\/x64_windows-opt\/bin\/external\/nsync\/nsync_cpp.lib : fatal error LNK1000: Internal error during CImplib::EmitThunk\n\n  Version 14.00.24215.1\n\n  ExceptionCode            = C0000005\n  ExceptionFlags           = 00000000\n  ExceptionAddress         = 00007FF7E0686896 (00007FF7E0670000) \"C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\bin\\amd64\\link.exe\"\n  NumberParameters         = 00000002\n  ExceptionInformation[ 0] = 0000000000000000\n  ExceptionInformation[ 1] = 0000000000000008\n\nCONTEXT:\n  Rax    = 0000000000000000  R8     = 00007FF7E076FBE0\n  Rbx    = 0000000000000000  R9     = 00007FF7E076E9F0\n  Rcx    = 0000000000000000  R10    = 0000000000000000\n  Rdx    = 00007FF7E076FBD8  R11    = 0000000000000000\n  Rsp    = 000000082B71DE68  R12    = 00007FF7E073D950\n  Rbp    = 000001A60F8F8360  E13    = 0000000000000000\n  Rsi    = 0000000000008000  R14    = 0000000000000000\n  Rdi    = 000001A6102AA300  R15    = 0000000000000000\n  Rip    = 00007FF7E0686896  EFlags = 0000000000010246\n  SegCs  = 0000000000000033  SegDs  = 000000000000002B\n  SegSs  = 000000000000002B  SegEs  = 000000000000002B\n  SegFs  = 0000000000000053  SegGs  = 000000000000002B\n  Dr0    = 0000000000000000  Dr3    = 0000000000000000\n  Dr1    = 0000000000000000  Dr6    = 0000000000000000\n  Dr2    = 0000000000000000  Dr7    = 0000000000000000\nTarget \/\/tensorflow\/tools\/pip_package:build_pip_package failed to build\nINFO: Elapsed time: 3172.917s, Critical Path: 253.76s\nINFO: 3875 processes: 3875 local.\nFAILED: Build did NOT complete successfully\n<\/code><\/pre>\n\n<p>And I've also tried to build it without --config=mkl term, which is successful, so I wonder if the error has kinda relation with mkl ?<\/p>\n","answers":[{"AnswerId":"54152255","CreationDate":"2019-01-11T18:39:16.807","ParentId":null,"OwnerUserId":"10880709","Title":null,"Body":"<p>please try the workaround described in this github issue response from Intel rep.\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23420\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23420<\/a><\/p>\n\n<p>or use conda and follow the instructions in the \"Windows and MacOS\" section here:\n<a href=\"https:\/\/software.intel.com\/en-us\/articles\/intel-optimization-for-tensorflow-installation-guide\" rel=\"nofollow noreferrer\">https:\/\/software.intel.com\/en-us\/articles\/intel-optimization-for-tensorflow-installation-guide<\/a><\/p>\n"}]}
+{"QuestionId":53096145,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-01T06:22:37.733","AcceptedAnswerId":null,"OwnerUserId":5330102.0,"Title":"How does Tensorflow schedule GPU operations? What is the right GPU scheduling mental model while using TF?","Body":"<p>I'm using Tensorflow's Eager mode to train some neural net models. I'm trying to identify quick wins while optimizing my models for training speed, and my question is about TF's GPU scheduling.<\/p>\n\n<p>How does Tensorflow decide which operations to schedule on a GPU vs which ones to keep on CPU? Is this a function of operand tensor sizes and operator type? What else?<\/p>\n\n<p>Is there a way to view which ops were executed on CPU vs GPU? How does it decide to tradeoff the parallelization advantage of GPU vs the time spent shipping data back and forth between GPU and CPU?<\/p>\n\n<p>Does any of this change for graph mode?<\/p>\n","answers":[]}
+{"QuestionId":53096870,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-11-01T07:28:53.483","AcceptedAnswerId":null,"OwnerUserId":10589692.0,"Title":"How does differ performance of neural network in TensorFlow and Keras?","Body":"<p>I am novice in Machine Learning (ML) and I'm trying to implement algorithm to understand basic syntax of ML frameworks etc. Now I am working on MNIST database of handwritten digits dataset.<\/p>\n\n<p>I implemented just one layer (I mean: Input layer has 784 inputs, Hidden layer has 512 nodes, Output layer has 10 outputs) Neural Network using TensorFlow framework, no data preprocessing, 128 batch size, 10 epochs, ADAM optimizer. And the algorithm achieved about 0.95 accuracy on train set. <\/p>\n\n<p>After that I tried to implement exactly the same architecture in Keras. However, the accuracy (train set) is about 0.3. I tried to find many different implementations founded on the internet but I still cannot find where is the issue. I believe that it is something stupid (as always is) :-\/<\/p>\n\n<p>I presume that the same architecture in Keras should give the same results as the implementation in TensorFlow, am I correct?<\/p>\n\n<p>My Keras implementation is:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\nfrom keras.utils.np_utils import to_categorical\n\ndf_train = pd.read_csv('datasets\/MNIST_train.csv', delimiter=',', header=0)\nY_train, X_train = np.split(df_train.values, [1], axis=1)\n\nm, n_x = X_train.shape\nn_y = len(np.unique(Y_train))\nn_layer1 = 512\nbatch_size = 128\nnum_epochs = 10\n\nY_train = to_categorical(Y_train)\n\nX_input = Input(shape=(n_x,), name='input')\nX = Dense(n_layer1, activation='relu', name='hidden')(X_input)\nX = Dense(n_y, activation='softmax', name='output')(X)\n\nmodel = Model(inputs=X_input, outputs=X, name='Neural Network')\n\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n\nmodel.fit(X_train, Y_train, epochs=num_epochs, batch_size=batch_size)\n<\/code><\/pre>\n\n<p>My TensorFlow implementation is:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\n\ndef one_hot(a, num_classes):\n  return np.eye(num_classes)[a.reshape(-1)]\n\ndef get_minibatches(batch_size, m, X, Y):\n    output_batches = []\n\n    for index in range(0, m, batch_size):\n        index_end = index + batch_size\n        batch = [X[index:index_end], Y[index:index_end]]\n        output_batches.append(batch)\n\n    return output_batches\n\ndef dense_layer(input, channels_in, channels_out, activation=None):\n   initializer = tf.contrib.layers.xavier_initializer()\n   w = tf.Variable(initializer([channels_in, channels_out]), name=\"w\")\n   b = tf.Variable(tf.zeros([1, channels_out]), name=\"b\")\n\n   if (activation == 'relu'):\n       a = tf.nn.relu(tf.matmul(input, w) + b)\n       return a\n   else:\n       z = tf.matmul(input, w) + b\n       return z\n\ndf_train = pd.read_csv('datasets\/MNIST_train.csv', delimiter=',', header=0)\nY_train, X_train = np.split(df_train.values, [1], axis=1)\n\nm, n_x = X_train.shape\nn_y = len(np.unique(Y_train))\nn_layer1 = 512\nbatch_size = 128\nnum_epochs = 10\n\nY_train = one_hot(Y_train, n_y)\n\nX = tf.placeholder(tf.float32, [None, n_x], name=\"X\")\nY = tf.placeholder(tf.float32, [None, n_y], name=\"Y\")\n\nhidden = dense_layer(X, n_x, n_layer1, 'relu')\noutput = dense_layer(hidden, n_layer1, n_y)\n\nloss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output, labels=Y))\n\noptimizer = tf.train.AdamOptimizer().minimize(loss)\n\npredict = tf.argmax(output, 1)\ncorrect_prediction = tf.equal(predict, tf.argmax(Y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\nminibatches = get_minibatches(batch_size, m, X_train, Y_train)\n\nwith tf.Session() as sess:        \n    sess.run(tf.global_variables_initializer())\n\n    current_cost = sess.run(loss, feed_dict={X: X_train, Y: Y_train})\n    train_accuracy = sess.run(accuracy, feed_dict={X: X_train, Y: Y_train})\n    print('Epoch: {:&lt;4} - Loss: {:&lt;8.3} Train Accuracy: {:&lt;5.3} '.format(0, current_cost, train_accuracy))\n\n    for epoch in range(num_epochs):            \n        for minibatch in minibatches:\n            minibatch_X, minibatch_Y = minibatch\n\n            sess.run(optimizer, feed_dict={ X: minibatch_X, Y: minibatch_Y })\n\n        current_cost = sess.run(loss, feed_dict={X: X_train, Y: Y_train})\n        train_accuracy = sess.run(accuracy, feed_dict={X: X_train, Y: Y_train})\n        print('Epoch: {:&lt;4} - Loss: {:&lt;8.3} Train Accuracy: {:&lt;5.3} '.format(epoch + 1, current_cost, train_accuracy))\n<\/code><\/pre>\n\n<p>Could you help me and advice what I am doing wrong?\nThank you\nPetr<\/p>\n","answers":[{"AnswerId":"53152363","CreationDate":"2018-11-05T10:19:16.890","ParentId":null,"OwnerUserId":"10589692","Title":null,"Body":"<p>I figured it out. At least partially. I standardized the input ((x - xmean) \/ xstd) and the Keras implementation has been started to return similar results as TensorFlow implementation\u2026<\/p>\n"}]}
+{"QuestionId":53097319,"AnswerCount":1,"Tags":"<python><keras><time-series><classification><multilabel-classification>","CreationDate":"2018-11-01T08:08:59.267","AcceptedAnswerId":null,"OwnerUserId":7228335.0,"Title":"Keras multi-label time-series classification considering time-series as an input image vector","Body":"<p>I am trying to build a multi-class classifier using Keras. I am not quite sure I have implemented it correctly.\nData is like this<\/p>\n\n<p>label time-series variables [0:25728}<\/p>\n\n<pre><code>index 0  1   2   3   4            25728\n  0   1  2.5 3.2 1.6 1.05 ........ 2.54\n  1   5  3.2 1.6 1.5 1.49 ........ 1.41\n  2   1  2.3 3.2 1.5 1.52 ........ 2.11\n  3   3  0.2 3.1 1.5 1.89 ........ 0.81\n  4   8  1.2 1.1 0.2 1.19 ........ 3.71\n  .   5  .    .   .   .   ........   .\n  .   7  .    .   .   .   ........   .\n1323  5  .    .   .   .   ........   .\n<\/code><\/pre>\n\n<p>Here is the code.\nI split data by 68 % then reshaping 1D array to a 2D array.\nas 384*67 = 25728\nSo forming an image of vector 384 by 67 for one label<\/p>\n\n<pre><code>x_train =x_train.reshape(x_train.shape[0],384,67)\nx_test =x_test.reshape(x_test.shape[0],384,67)\n\ntrain_label_y = df_train_y[2].values\ntest_label_y = df_test_y[2].values\n\ntrain_Y = np.array(train_label_y[:], dtype=int)\ntest_Y  = np.array(test_label_y[:], dtype=int)\n\nY_train = convert_to_one_hot(train_Y, C = 8)\nY_test  = convert_to_one_hot(test_Y,  C = 8)\n\nx_train = x_train.reshape(x_train.shape + (1,))\nx_test = x_test.reshape(x_test.shape + (1,))\n\ninput_shape = x_train.shape[1:]\nmodel = Sequential()\nmodel.add(Conv2D(2, kernel_size=(3, 3), padding='same', input_shape=input_shape))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(64, kernel_size=(3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(128, kernel_size=(3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(8, kernel_size=(3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(GlobalAveragePooling2D())\nmodel.add(Dense(8, activation='sigmoid'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\nhist = model.fit(x_train, Y_train, batch_size=batch_size, epochs=nb_epochs, verbose=1)\nscore = model.evaluate(x_test, Y_test)\nprint(\"Accuracy: %.2f%%\" % (score[1] * 100))\n<\/code><\/pre>\n\n<p>It gives 96.16% accuracy but I don't believe it is true. \nI want to predict the labels. <\/p>\n\n<ol>\n<li>How can I predict labels? <\/li>\n<li>What I am doing wrong?<\/li>\n<\/ol>\n\n<p>Please help! Thank you.<\/p>\n","answers":[{"AnswerId":"55525845","CreationDate":"2019-04-04T22:31:11.607","ParentId":null,"OwnerUserId":"5096832","Title":null,"Body":"<p>The reason you are getting weird results is you should use the 'softmax' activation in your last layer.<\/p>\n"}]}
+{"QuestionId":53097477,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-11-01T08:20:26.520","AcceptedAnswerId":null,"OwnerUserId":10580289.0,"Title":"An error happens while following pytorch tutorial","Body":"<p>I was following pytorch tutorial, training a classifier using CIFAR10 dataset. All I've done was copy and paste codes in tutorial page to Pycharm project, yet I faced unknown errors. And I also used Jupyter notebook to run same code, and met errors again. Below is the code and errors. What should I do?<\/p>\n\n<pre><code>import torch\nimport torchvision\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport torchvision.transforms as transforms\nimport torchvision.models as models\nimport torch.autograd as Variable\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\n\ntransform = transforms.Compose(\n    [transforms.ToTensor(),\n     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n\ntrainset = torchvision.datasets.CIFAR10(root='.\/data', train=True,\n                                    download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=4,\n                                      shuffle=True, num_workers=2)\n\ntestset = torchvision.datasets.CIFAR10(root='.\/data', train=False,\n                                   download=True, transform=transform)\ntestloader = torch.utils.data.DataLoader(testset, batch_size=4,\n                                     shuffle=False, num_workers=2)\n\nclasses = ('plane', 'car', 'bird', 'cat',\n       'deer', 'dog', 'frog', 'horse', 'ship', 'truck')\n\n\nclass Net(nn.Module):\n  def __init__(self):\n    super(Net, self).__init__()\n    self.conv1 = nn.Conv2d(3, 6, 5)\n    self.pool = nn.MaxPool2d(2, 2)\n    self.conv2 = nn.Conv2d(6, 16, 5)\n    self.fc1 = nn.Linear(16 * 5 * 5, 120)\n    self.fc2 = nn.Linear(120, 84)\n    self.fc3 = nn.Linear(84, 10)\n\n  def forward(self, x):\n    x = self.pool(F.relu(self.conv1(x)))\n    x = self.pool(F.relu(self.conv2(x)))\n    x = x.view(-1, 16 * 5 * 5)\n    x = F.relu(self.fc1(x))\n    x = F.relu(self.fc2(x))\n    x = self.fc3(x)\n    return x\n\n\nnet = Net()\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)\n\nfor epoch in range(2):\n\n  running_loss = 0.0\n    for i, data in enumerate(trainloader, 0):\n\n      inputs, labels = data\n\n      optimizer.zero_grad()\n\n      outputs = net(inputs)\n      loss = criterion(outputs, labels)\n      loss.backward()\n      optimizer.step()\n\n      running_loss += loss.item()\n      if i % 2000 == 1999:    # print every 2000 mini-batches\n        print('[%d, %5d] loss: %.3f' %\n              (epoch + 1, i + 1, running_loss \/ 2000))\n        running_loss = 0.0\n\nprint('Finished Training')\n<\/code><\/pre>\n\n<p>Error in Pycharm:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"&lt;string&gt;\", line 1, in &lt;module&gt;\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 105, in spawn_main\nexitcode = _main(fd)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 114, in _main\nprepare(preparation_data)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 225, in prepare\n_fixup_main_from_path(data['init_main_from_path'])\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 277, in _fixup_main_from_path\nrun_name=\"__mp_main__\")\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\runpy.py\", line 263, in run_path\npkg_name=pkg_name, script_name=fname)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\runpy.py\", line 96, in _run_module_code\nmod_name, mod_spec, pkg_name, script_name)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\runpy.py\", line 85, in _run_code\nexec(code, run_globals)\nFile \"D:\\Users\\VML1\\PycharmProjects\\untitled1\\HY1.py\", line 57, in &lt;module&gt;\nfor i, data in enumerate(trainloader, 0):\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 417, in __iter__\nreturn DataLoaderIter(self)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 234, in __init__\nw.start()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\process.py\", line 105, in start\nself._popen = self._Popen(self)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\context.py\", line 223, in _Popen\nreturn _default_context.get_context().Process._Popen(process_obj)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\context.py\", line 322, in _Popen\nreturn Popen(process_obj)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\popen_spawn_win32.py\", line 33, in __init__\nprep_data = spawn.get_preparation_data(process_obj._name)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 143, in get_preparation_data\n_check_not_importing_main()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\spawn.py\", line 136, in _check_not_importing_main\nis not going to be frozen to produce an executable.''')\nRuntimeError: \n    An attempt has been made to start a new process before the\n    current process has finished its bootstrapping phase.\n\n    This probably means that you are not using fork to start your\n    child processes and you have forgotten to use the proper idiom\n    in the main module:\n\n        if __name__ == '__main__':\n            freeze_support()\n            ...\n\n    The \"freeze_support()\" line can be omitted if the program\n    is not going to be frozen to produce an executable.\nTraceback (most recent call last):\nFile \"D:\/Users\/VML1\/PycharmProjects\/untitled1\/HY1.py\", line 57, in &lt;module&gt;\nfor i, data in enumerate(trainloader, 0):\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 417, in __iter__\nreturn DataLoaderIter(self)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 234, in __init__\nw.start()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\process.py\", line 105, in start\nself._popen = self._Popen(self)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\context.py\", line 223, in _Popen\nreturn _default_context.get_context().Process._Popen(process_obj)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\context.py\", line 322, in _Popen\nreturn Popen(process_obj)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\popen_spawn_win32.py\", line 65, in __init__\nreduction.dump(process_obj, to_child)\n  File \"D:\\ProgramData\\Anaconda3\\envs\\pytorch\\lib\\multiprocessing\\reduction.py\", line 60, in dump\n    ForkingPickler(file, protocol).dump(obj)\nBrokenPipeError: [Errno 32] Broken pipe\n<\/code><\/pre>\n\n<p>Error in Jupyter:<\/p>\n\n<pre><code>TypeError                                 \nTraceback (most recent call last)\n&lt;ipython-input-2-1dd1ecde814e&gt; in &lt;module&gt;\n---&gt; 45         outputs = net(inputs)\n     46         loss = criterion(outputs, labels)\n     47         loss.backward()\n\nc:\\users\\vml1\\anaconda3\\envs\\hypytorch\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    355             result = self._slow_forward(*input, **kwargs)\n    356         else:\n--&gt; 357             result = self.forward(*input, **kwargs)\n    358         for hook in self._forward_hooks.values():\n    359             hook_result = hook(self, input, result)\n\nc:\\users\\vml1\\anaconda3\\envs\\hypytorch\\lib\\site-packages\\torchvision\\models\\resnet.py in forward(self, x)\n    137 \n    138     def forward(self, x):\n--&gt; 139         x = self.conv1(x)\n    140         x = self.bn1(x)\n    141         x = self.relu(x)\n\nc:\\users\\vml1\\anaconda3\\envs\\hypytorch\\lib\\site-                \npackages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    355             result = self._slow_forward(*input, **kwargs)\n    356         else:\n--&gt; 357             result = self.forward(*input, **kwargs)\n    358         for hook in self._forward_hooks.values():\n    359             hook_result = hook(self, input, result)\n\n    c:\\users\\vml1\\anaconda3\\envs\\hypytorch\\lib\\site-packages\\torch\\nn\\modules\\conv.py in forward(self, input)\n    280     def forward(self, input):\n    281         return F.conv2d(input, self.weight, self.bias, self.stride,\n--&gt; 282                         self.padding, self.dilation, self.groups)\n    283 \n    284 \n\nc:\\users\\vml1\\anaconda3\\envs\\hypytorch\\lib\\site-    \npackages\\torch\\nn\\functional.py in conv2d(input, weight, bias, stride,     \npadding, dilation, groups)\n     88                 _pair(0), groups, torch.backends.cudnn.benchmark,\n     89                 torch.backends.cudnn.deterministic,     \ntorch.backends.cudnn.enabled)\n---&gt; 90     return f(input, weight, bias)\n     91 \n     92 \n\nTypeError: argument 0 is not a Variable\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53098178,"AnswerCount":2,"Tags":"<python><r><tensorflow><keras><mini-batch>","CreationDate":"2018-11-01T09:12:26.957","AcceptedAnswerId":53103509.0,"OwnerUserId":9770403.0,"Title":"Generator function in LSTM Keras for outputting mini batches of one files","Body":"<p>I have a generator function which works fine. I have a large list of .txt files in which each file is also quite long. The task would now be to write a generator function which Takes:<\/p>\n\n<ol>\n<li>a batch of Files<\/li>\n<li>and then a batch of size 128 out of one file<\/li>\n<\/ol>\n\n<p>my code now:<\/p>\n\n<pre><code>data_files_generator &lt;- function(train_set) {\n\n  files &lt;- train_set\n  next_file &lt;- 0\n\n  function() {\n\n    # move to the next file (note the &lt;&lt;- assignment operator)\n    next_file &lt;&lt;- next_file + 1\n\n    # if we've exhausted all of the files then start again at the\n    # beginning of the list (keras generators need to yield\n    # data infinitely -- termination is controlled by the epochs\n    # and steps_per_epoch arguments to fit_generator())\n    if (next_file &gt; length(files))\n    {next_file &lt;&lt;- 1}\n\n    # determine the file name\n    file &lt;- files[[next_file]]\n\n    text &lt;- read_lines(paste(data_dir, file, sep = \"\" )) %&gt;%\n      str_to_lower() %&gt;%\n      str_c(collapse = \"\\n\") %&gt;%\n      removeNumbers() %&gt;%\n      tokenize_characters(strip_non_alphanum = FALSE, simplify = TRUE)\n\n    text &lt;- text[text %in% chars]\n\n    dataset &lt;- map(\n      seq(1, length(text) - maxlen - 1, by = 3), \n      ~list(sentece = text[.x:(.x + maxlen - 1)], next_char = text[.x + maxlen])\n    )\n\n    dataset &lt;- transpose(dataset)\n\n    # Vectorization\n    x &lt;- array(0, dim = c(length(dataset$sentece), maxlen, length(chars)))\n    y &lt;- array(0, dim = c(length(dataset$sentece), length(chars)))\n\n    for(i in 1:length(dataset$sentece)){\n\n      x[i,,] &lt;- sapply(chars, function(x){\n        as.integer(x == dataset$sentece[[i]])\n      })\n\n      y[i,] &lt;- as.integer(chars == dataset$next_char[[i]])\n\n    }\n    rounded_dim &lt;- floor(dim(x)[1]\/mini_batch_size)\n    match_size_to_batch &lt;- 128 * rounded_dim\n\n    x &lt;- x[1:match_size_to_batch, 1:maxlen, 1:length(chars)]\n    y &lt;- y_val[1:match_size_to_batch, 1:length(chars)]\n\n    return(list(x, y))\n\n  }\n}\n<\/code><\/pre>\n\n<p>So what is coming is coming in is a Text file which is transformed into smaller pieces of text (of length <code>maxlen<\/code>) and is then one hot encoded into 0 and 1 matrices. <\/p>\n\n<p>The problem is that from my code the output is one Data Cube of size <code>maxlen x lenght(chars) x samples<\/code> where the number of samples is very big and that why I would like my generator function to output always a cube of size <code>maxlen x lenght(chars) x samples(128)<\/code> and then output the next batch of size <code>maxlen x lenght(chars) x samples<\/code> until the whole text file is read in and then go to the next text file...<\/p>\n\n<p>The output for now is an error:<\/p>\n\n<pre><code> Error in py_call_impl(callable, dots$args, dots$keywords) : \n  ValueError: Cannot feed value of shape (112512, 40, 43) for Tensor 'lstm_layer_input_1:0', which has shape '(128, 40, 43)' \n<\/code><\/pre>\n\n<p>hope I have explained it good enough to understand. I think I have to input some kind of for loop to iterate over the sample length but I have no Idea how to include this into the gen. function.<\/p>\n","answers":[{"AnswerId":"53102892","CreationDate":"2018-11-01T14:03:18.373","ParentId":null,"OwnerUserId":"10582903","Title":null,"Body":"<p>According to the error, you're trying to feed in an object of shape <code>(112512, 40, 43)<\/code> but your LSTM layer is expecting an object of shape <code>(128, 40, 43)<\/code>. There seems to be some missing code, but when you're defining the input layer, are you fixing the batch size? I've had luck with defining my input layer as:<\/p>\n\n<pre><code>l_input = Input(shape = (None, num_features), name = 'input_layer')\n<\/code><\/pre>\n\n<p>I suspect the error is due to these lines of code:<\/p>\n\n<pre><code>rounded_dim &lt;- floor(dim(x)[1]\/mini_batch_size)\nmatch_size_to_batch &lt;- 128 * rounded_dim\n<\/code><\/pre>\n\n<p>This gives you a batch size much larger than 128. From the <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">Keras documentation<\/a>, the input shape should be <code>(batch_size, timesteps, input_dim)<\/code>. The batch sizes need not be the same throughout an entire epic, but for a batch they all need to have the same number of <code>timesteps<\/code> (which it looks like you handle with <code>maxlen<\/code>).<\/p>\n"},{"AnswerId":"53103509","CreationDate":"2018-11-01T14:39:03.687","ParentId":null,"OwnerUserId":"9770403","Title":null,"Body":"<p>I have implemented an for loop which is returning now batches of size 128:<\/p>\n\n<p>Changed Code:<\/p>\n\n<pre><code>data_files_generator &lt;- function(train_set) {\n\n  files &lt;- train_set\n  next_file &lt;- 0\n\n  function() {\n\n    # move to the next file (note the &lt;&lt;- assignment operator)\n    next_file &lt;&lt;- next_file + 1\n\n    # if we've exhausted all of the files then start again at the\n    # beginning of the list (keras generators need to yield\n    # data infinitely -- termination is controlled by the epochs\n    # and steps_per_epoch arguments to fit_generator())\n    if (next_file &gt; length(files))\n    {next_file &lt;&lt;- 1}\n\n    # determine the file name\n    file &lt;- files[[next_file]]\n\n    text &lt;- read_lines(paste(data_dir, file, sep = \"\" )) %&gt;%\n      str_to_lower() %&gt;%\n      str_c(collapse = \"\\n\") %&gt;%\n      removeNumbers() %&gt;%\n      tokenize_characters(strip_non_alphanum = FALSE, simplify = TRUE)\n\n    text &lt;- text[text %in% chars]\n\n    dataset &lt;- map(\n      seq(1, length(text) - maxlen - 1, by = 3), \n      ~list(sentece = text[.x:(.x + maxlen - 1)], next_char = text[.x + maxlen])\n    )\n\n    dataset &lt;- transpose(dataset)\n\n    # Vectorization\n    x &lt;- array(0, dim = c(length(dataset$sentece), maxlen, length(chars)))\n    y &lt;- array(0, dim = c(length(dataset$sentece), length(chars)))\n\n    for(i in 1:length(dataset$sentece)){\n\n      x[i,,] &lt;- sapply(chars, function(x){\n        as.integer(x == dataset$sentece[[i]])\n      })\n\n      y[i,] &lt;- as.integer(chars == dataset$next_char[[i]])\n\n    }\n    rounded_dim &lt;- floor(dim(x)[1]\/mini_batch_size)\n    match_size_to_batch &lt;- 128 * rounded_dim\n\n    x &lt;- x[1:match_size_to_batch, 1:maxlen, 1:length(chars)]\n    y &lt;- y_val[1:match_size_to_batch, 1:length(chars)]\n\n    #Edit:\n    span_start &lt;-1\n    for (iter in 1:rounded_dim){\n     i &lt;- iter * 128\n     span_end &lt;- iter * 128\n     x &lt;- x[span_start:span_end, 1:maxlen, 1:length(chars)]\n     y &lt;- y[span_start:span_end, 1:length(chars)]\n     span_start &lt;- i\n     return(list(x, y))\n    }\n  }\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":53098220,"AnswerCount":1,"Tags":"<python-3.x><pandas><tensorflow><machine-learning>","CreationDate":"2018-11-01T09:14:41.657","AcceptedAnswerId":null,"OwnerUserId":10117424.0,"Title":"Appending a generator class object to a pandas dataframe","Body":"<p>I have a feature dataset that contains many fruits and the label tells with a logistic probability whether one is an apple or not using tensorflow.\nHowever I can't seem to add the predicted data to the original dataframe. <\/p>\n\n<pre><code>    predictions = DNN_classifier.predict(input_fn=is_apple_ds)\n<\/code><\/pre>\n\n<p>This returns a generator class object and if I create a column and add it to the original dataframe it just says \"generator class object at some address\"<\/p>\n\n<p>I instead want to add a column to the original dataframe that would give the actual predicted values instead of this<\/p>\n\n<p>Edit:\nMy question simply is to how to extract all the values from a generator object and add it to a pandas dataframe<\/p>\n","answers":[{"AnswerId":"53098727","CreationDate":"2018-11-01T09:47:52.300","ParentId":null,"OwnerUserId":"3471881","Title":null,"Body":"<p>According to the <code>tensorflow<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/DNNClassifier#predict\" rel=\"nofollow noreferrer\">documentation<\/a> <code>predict()<\/code> will <code>yield<\/code> each predicition. This is why you get a generator object. <\/p>\n\n<hr>\n\n<pre><code>def generate_sequence():\n    for x in range(0,10):\n        yield x\n\ngenerated_sequence = generate_sequence()\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>&gt;&gt; print(generated_sequence)\n&lt;generator object generate_sequence at 0x7fdc31f23990&gt;\n<\/code><\/pre>\n\n<hr>\n\n<p>You can put a generator object into a <code>DataFrame<\/code> by converting it to a <code>list<\/code> or by simply calling <code>pd.DataFrame()<\/code> with it.<\/p>\n\n<pre><code>&gt;&gt; list(generated_sequence)\n[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]\n\n&gt;&gt; pd.DataFrame(generated_sequence)\n\n0\n0   0\n1   1\n2   2\n3   3\n4   4\n5   5\n6   6\n7   7\n8   8\n9   9\n<\/code><\/pre>\n"}]}
+{"QuestionId":53098225,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-01T09:14:51.540","AcceptedAnswerId":null,"OwnerUserId":5820161.0,"Title":"Training stop after step 1 in Tensorflow model api ssd mobilenet in google colab","Body":"<p>Recently i am trying to train ssd mobilenet object detection model of tensorflow model api on my custom data set in google colab, after step 1 the training session stopped without showing or throwing any exception or message.I can not figure out the issue.Can anyone please give any explanation?<\/p>\n","answers":[{"AnswerId":"56070425","CreationDate":"2019-05-10T03:46:51.477","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>Training stops after 1 step means you have only 1 batch to process.<\/p>\n\n<pre><code>num_batches = num_epochs * (num_samples \/ batch_size)\n<\/code><\/pre>\n\n<p>Even setting <code>steps<\/code> to a very big value, it will stop after step 1 if <code>num_batches<\/code> above is just 1.<\/p>\n"},{"AnswerId":"53293578","CreationDate":"2018-11-14T05:15:07.580","ParentId":null,"OwnerUserId":"4618512","Title":null,"Body":"<p>I think there might be the issue due to limited RAM. You can try:-<\/p>\n\n<ul>\n<li>See the size of Training Data sample(resize to smaller dimension)<\/li>\n<\/ul>\n\n<p>In configuring file:-<\/p>\n\n<ul>\n<li>Reduce your batch size <\/li>\n<li>Disable Data augmentation<\/li>\n<li>Set the following parameters low.<\/li>\n<\/ul>\n\n<p>For example:-<\/p>\n\n<pre><code>  train_config: {\n  batch_size: 1\n  ...\n  batch_queue_capacity: 50\n  num_batch_queue_threads: 8\n  prefetch_queue_capacity: 10\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":53098495,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><conv-neural-network><loadimage>","CreationDate":"2018-11-01T09:33:11.023","AcceptedAnswerId":53099358.0,"OwnerUserId":10148648.0,"Title":"How to load images for classification problem using Keras","Body":"<p>I am working on image classification problem using Keras framework. This is binary classification problem and I have 2 folders <strong>training set<\/strong> and <strong>test set<\/strong> which contains images of both the classes. I don't have separate folder for each class (say cat vs. dog). Keras <strong>ImageDataGenerator<\/strong> works when we have separate folders for each class (cat folder &amp; dog folder). But I have all the images in single folder <strong>training set<\/strong> and I am not understanding how to proceed further. Kindly suggest how to load the images.<\/p>\n\n<p>I also have 2 CSV files - train.csv and test.csv. train.csv contain 2 columns namely image_id and class_name. test.csv contains image_id. Note that image_id is matching with the name of files in the images folders.<\/p>\n","answers":[{"AnswerId":"53099358","CreationDate":"2018-11-01T10:26:42.100","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The latest versions of <code>ImageDataGenerator<\/code> have a method called <code>flow_from_dataframe<\/code>, which does exactly what you want.<\/p>\n\n<p>Basically it is used by first loading your CSV file in a pandas <code>DataFrame<\/code>, instantiate a <code>ImageDataGenerator<\/code> and then call <code>flow_from_dataframe<\/code> with three important parameters:<\/p>\n\n<ul>\n<li><code>directory<\/code>: Folder where your data lives.<\/li>\n<li><code>x_col<\/code>: Column in the <code>DataFrame<\/code> that contains the filenames inside the <code>directory<\/code> that correspond to your training\/testing data.<\/li>\n<li><code>y_col<\/code>: Column in the <code>DataFrame<\/code> corresponding to the labels that will be output by the generator.<\/li>\n<\/ul>\n\n<p>Then you use this generator as any other, by calling <code>fit_generator<\/code>. More information and examples are available <a href=\"https:\/\/medium.com\/@vijayabhaskar96\/tutorial-on-keras-imagedatagenerator-with-flow-from-dataframe-8bd5776e45c1\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53098566,"AnswerCount":1,"Tags":"<python><keras><deep-learning><keras-layer>","CreationDate":"2018-11-01T09:37:10.813","AcceptedAnswerId":53099120.0,"OwnerUserId":6384539.0,"Title":"TypeError: __init__() got an unexpected keyword argument 'trainable'","Body":"<p>I am trying to load a RNN model architecture trained in Keras using keras.models.model_from_json and I am getting the mentioned error<\/p>\n\n<pre><code>with open('model_architecture.json', 'r') as f:\n    model = model_from_json(f.read(), custom_objects={'AttLayer':AttLayer})\n\n# Load weights into the new model\nmodel.load_weights('model_weights.h5')\n<\/code><\/pre>\n\n<p>Here is the custom layer I am using <\/p>\n\n<pre><code>class AttLayer(Layer):\n    def __init__(self, attention_dim):\n        self.init = initializers.get('normal')\n        self.supports_masking = True\n        self.attention_dim = attention_dim\n        super(AttLayer, self).__init__()\n\n    def build(self, input_shape):\n        assert len(input_shape) == 3\n        self.W = K.variable(self.init((input_shape[-1], self.attention_dim)))\n        self.b = K.variable(self.init((self.attention_dim, )))\n        self.u = K.variable(self.init((self.attention_dim, 1)))\n        self.trainable_weights = [self.W, self.b, self.u]\n        super(AttLayer, self).build(input_shape)\n\n    def compute_mask(self, inputs, mask=None):\n        return None\n\n    def call(self, x, mask=None):\n        # size of x :[batch_size, sel_len, attention_dim]\n        # size of u :[batch_size, attention_dim]\n        # uit = tanh(xW+b)\n        uit = K.tanh(K.bias_add(K.dot(x, self.W), self.b))\n        ait = K.dot(uit, self.u)\n        ait = K.squeeze(ait, -1)\n\n        ait = K.exp(ait)\n\n        if mask is not None:\n            # Cast the mask to floatX to avoid float64 upcasting in theano\n            ait *= K.cast(mask, K.floatx())\n        ait \/= K.cast(K.sum(ait, axis=1, keepdims=True) + K.epsilon(), K.floatx())\n        ait = K.expand_dims(ait)\n        weighted_input = x * ait\n        output = K.sum(weighted_input, axis=1)\n\n        return output\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], input_shape[-1])\n\n    def get_config(self):\n        config = {'attention_dim': self.attention_dim}\n        base_config = super(AttLayer, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n\n<p>error:<\/p>\n\n<pre><code>File \"scripts\/Classifier.py\", line 254, in test\n    model = model_from_json(f.read(), custom_objects={'AttLayer':AttLayer})\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/models.py\", line 345, in model_from_json\n    return layer_module.deserialize(config, custom_objects=custom_objects)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/__init__.py\", line 54, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/utils\/generic_utils.py\", line 139, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/topology.py\", line 2489, in from_config\n    process_layer(layer_data)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/topology.py\", line 2475, in process_layer\n    custom_objects=custom_objects)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/__init__.py\", line 54, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/utils\/generic_utils.py\", line 139, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/wrappers.py\", line 100, in from_config\n    custom_objects=custom_objects)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/__init__.py\", line 54, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/utils\/generic_utils.py\", line 139, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/topology.py\", line 2489, in from_config\n    process_layer(layer_data)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/topology.py\", line 2475, in process_layer\n    custom_objects=custom_objects)\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/__init__.py\", line 54, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/utils\/generic_utils.py\", line 141, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"\/home\/biswadip\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/topology.py\", line 1254, in from_config\n    return cls(**config)\nTypeError: __init__() got an unexpected keyword argument 'trainable'\n<\/code><\/pre>\n\n<p>Versions:<\/p>\n\n<pre><code>Keras==2.0.8\ntensorflow==1.4.1\n<\/code><\/pre>\n\n<p>I tried training and loading using different versions, but with no luck. Finally I removed 'trainable' and 'name' (key value pairs)from my custom layer detail in the model architecture file(model_architecture.json) and model seems to be loading without any error. But this looks like a fix and I have to do this every time I train the model.<\/p>\n","answers":[{"AnswerId":"53099120","CreationDate":"2018-11-01T10:11:15.627","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I think you missed a small detail in your layer definition. You layers' <code>__init__<\/code> method should take keyword arguments (<code>**kwargs<\/code>) and you should pass these keyword arguments to the parent class <code>__init__<\/code>, like this:<\/p>\n\n<pre><code>class AttLayer(Layer):\n    def __init__(self, attention_dim, **kwargs):\n        self.init = initializers.get('normal')\n        self.supports_masking = True\n        self.attention_dim = attention_dim\n        super(AttLayer, self).__init__(**kwargs)\n<\/code><\/pre>\n\n<p>This way any generic layer parameter will be correctly passed to the parent class, in your case, the <code>trainable<\/code> flag.<\/p>\n"}]}
+{"QuestionId":53098663,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T09:43:41.857","AcceptedAnswerId":null,"OwnerUserId":10590148.0,"Title":"float16 is much slower than float32 in tensorflow when do embedding_lookup_sparse","Body":"<p>When run the following code on Nvidia GTX1080ti, if FLOAT_DTYPE=tf.float16, the run time is 0.61s, if FLOAT_DTYPE = tf.float32, the time is 0.26s. The timeline show that scatter_sub take up a large part of the time. Why float16 is so much slower than float32? Tensorflow version is 1.8.<\/p>\n\n<pre><code>import time\nimport tensorflow as tf\n\nFLOAT_DTYPE = tf.float16\n\nvar = tf.get_variable(\"embedding\", initializer = tf.ones(shape = [1000000, 5], dtype = FLOAT_DTYPE), dtype = FLOAT_DTYPE)\nembeddings = tf.nn.embedding_lookup(var, [i for i in range(100000)])\n\nw = tf.get_variable(\"w\", shape = [5, 1], dtype = FLOAT_DTYPE)\nz = tf.matmul(embeddings, w)\n\nopt = tf.train.AdamOptimizer(1e-3, epsilon = 1e-4)\n\ng = tf.gradients(z, xs = [var])\ntrain = opt.apply_gradients([(g[0], var)])\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n\n    start = time.time()\n    for _ in range(10):\n        sess.run(train)\n    elapsed = time.time() - start\n    print 'Time: %f' % elapsed\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53098686,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2018-11-01T09:45:04.207","AcceptedAnswerId":null,"OwnerUserId":6140888.0,"Title":"Dump multidimensional tensor to file","Body":"<p>I have a multidimensional tensor and I want to dump it as a matrix to a file. Whats the best way to do this.<\/p>\n\n<p>For now I have the following for a 4-dimensional tensor: <\/p>\n\n<pre><code>const tensorflow::TensorShape&amp; tensor_shape = tensor.shape();\nauto _tensor = tensor.tensor&lt;float, 4&gt;();\nfor(int i=0; i&lt;tensor_shape.dim_size(0); i++)\n{\n    for(int j=0; j&lt;tensor_shape.dim_size(1); j++)\n    {\n\n        for(int k=0; k&lt;tensor_shape.dim_size(2); k++)\n        {\n            for(int l=0; l&lt;tensor_shape.dim_size(3); k++)\n            {\n               std::cout &lt;&lt; _tensor(i,j,k,l) &lt;&lt; std::endl; \/\/right one\n            }\n        }\n    }\n}\n<\/code><\/pre>\n\n<p>However this is extremely slow.. is there a better and more efficient way to do this?<\/p>\n","answers":[]}
+{"QuestionId":53098800,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2018-11-01T09:51:33.087","AcceptedAnswerId":null,"OwnerUserId":6898439.0,"Title":"what does it mean when weights of a layer is not normally distributed","Body":"<p>I plot all my weights of my neural network on tensorboard, I found that some <\/p>\n\n<p>weights of some layer is normally distributed:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TcmJy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TcmJy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>but, some are not.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1gfyP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1gfyP.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>what does this imply? should I increase or decrease the capacity of this layer?\nUpdate:\nMy network is a LSTM-based netowrk. the non-normal distributed weights is the weights multiply with input feature,  the normal distributed weights is the weights multiply with states.<\/p>\n","answers":[{"AnswerId":"53103451","CreationDate":"2018-11-01T14:35:56.650","ParentId":null,"OwnerUserId":"10528078","Title":null,"Body":"<p>one explanation base on convolutional networks might be this(I don't know if this is true for any other kind of artificial neural models or not), hence the first layer tries to find distinct small features weights are distributed very widely and network tries to find any useful feature it can, then in the next layers combination of these distinct features are used, which make sense to put a normal distribution of weights hence every one of the previous features are going to be part of a single bigger or more representative feature  in next layers.<\/p>\n\n<p>but this was only my intuition I am not sure if this is the reason with proof now.<\/p>\n"}]}
+{"QuestionId":53099262,"AnswerCount":0,"Tags":"<multithreading><keras>","CreationDate":"2018-11-01T10:21:01.290","AcceptedAnswerId":null,"OwnerUserId":10293613.0,"Title":"Using Keras model predict function from multiprocessing Pool","Body":"<p>I've seen some <a href=\"https:\/\/stackoverflow.com\/questions\/43136293\/running-keras-model-for-prediction-in-multiple-threads\">few<\/a> <a href=\"https:\/\/stackoverflow.com\/questions\/50099745\/keras-error-when-predicting-on-multithreading\">similar<\/a> <a href=\"https:\/\/stackoverflow.com\/questions\/51657013\/python-multiprocessing-with-keras-prediction\">posts<\/a> on this topic, but none seem to address my issue.<\/p>\n\n<p>I have trained a Keras model (CPU only) and want to call the predict function asynchronously using a <code>multithreading.Pool<\/code>. However, the call to <code>predict<\/code> just hangs. There is no exception thrown or anything. Calling it from the main thread works fine. I tried using <code>model._make_predict_function()<\/code> as suggested before, but this doesn't resolve this for me.<\/p>\n\n<p>I've set up a Jupyter notebook to reproduce this (Keras==2.2.4, tensorflow==1.11.0):<\/p>\n\n<pre><code>In  [1]: from keras.models import Sequential\n         from keras.layers import Dense\n         from multiprocessing.pool import Pool\n\nIn  [2]: # Create sample model from Keras documentation\n         model = Sequential()\n         model.add(Dense(32, activation='relu', input_dim=100))\n         model.add(Dense(1, activation='sigmoid'))\n         model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n\n         # Generate dummy data\n         import numpy as np\n         data = np.random.random((1000, 100))\n         labels = np.random.randint(2, size=(1000, 1))\n\n         # Train the model, iterating on the data in batches of 32 samples\n         model.fit(data, labels, epochs=10, batch_size=32, verbose=0)\n\nIn  [3]: test_data = np.random.random((1,100))\n\n         def predict(model, data):\n             return model.predict(data)\n\n         def do_predict(_=1):\n             print('Prediction:', predict(model, test_data))\n             print('Done')\n\nIn  [4]: do_predict()\nOut [4]: Prediction: [[0.5553096]]\n         Done\n\nIn  [5]: with Pool(1) as pool:\n             pool.apply_async(do_predict, [1]).get()\n             pool.close()\n             pool.join()\n<\/code><\/pre>\n\n<p>At the last step it just hangs. Can anybody help me finding out what's going on here? Is it not possible to use <code>predict<\/code> asynchronously?<\/p>\n","answers":[]}
+{"QuestionId":53099728,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning><pre-trained-model>","CreationDate":"2018-11-01T10:48:31.040","AcceptedAnswerId":null,"OwnerUserId":9200834.0,"Title":"Tensorflow label_image NotFoundError after retraining","Body":"<p>I've run the tensoflow's <code>retrain.py<\/code> script on my custom dataset. After the training process I got the expected output files <code>output_graph.pb<\/code> and <code>output_labels.txt<\/code>. \nHowever, when I tried to predict the label on an unseen image by using the latest tensorflow's <code>label_image.py<\/code> script I got this error:<\/p>\n\n<pre><code>NotFoundError (see above for traceback): Error while reading resource variable module\/InceptionV3\/Mixed_7c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/moving_mean from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost\/module\/InceptionV3\/Mixed_7c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/moving_mean)\n         [[Node: import\/module_apply_default\/InceptionV3\/InceptionV3\/Mixed_7c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/FusedBatchNorm\/ReadVariableOp = ReadVariableOp[dtype=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](import\/module\/InceptionV3\/Mixed_7c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/moving_mean)]]\n<\/code><\/pre>\n\n<p>Does anyone have an idea, why some variables might be incorrectly initialized?<\/p>\n","answers":[]}
+{"QuestionId":53099925,"AnswerCount":2,"Tags":"<python><numpy><machine-learning><keras><numpy-ndarray>","CreationDate":"2018-11-01T11:00:35.897","AcceptedAnswerId":53100197.0,"OwnerUserId":7525181.0,"Title":"How to input a 2D array in Keras-Python?","Body":"<p>I am very new to machine-learning and keras and was stuck on trying to input data; I have data which looks like:<\/p>\n\n<pre><code>[[[0.01363717 0.        ]\n  [0.01577874 0.        ]\n  [0.01463021 0.        ]]\n\n [[0.01577874 0.        ]\n  [0.01463021 0.        ]\n  [0.01006721 0.        ]]\n\n [[0.01463021 0.        ]\n  [0.01006721 0.        ]\n  [0.00762504 0.        ]]...]\n<\/code><\/pre>\n\n<p>The data's shape is: (1607, 3, 2). \nHow can I pass:<\/p>\n\n<pre><code> [[0.01363717 0.        ]\n  [0.01577874 0.        ]\n  [0.01463021 0.        ]]\n<\/code><\/pre>\n\n<p>as an input in a layer of 512 CuDNNLSTM cells? <\/p>\n\n<p>Here is my whole network:<\/p>\n\n<pre><code>def create_model():\n    model = Sequential()\n\n    model.add(CuDNNLSTM(512, input_shape=(3,2), return_sequences=True, name='inputlstm1'))\n    model.add(Dropout(0.2))\n\n    model.add(CuDNNLSTM(512, return_sequences=True,name='lstm2'))\n    model.add(Dropout(0.2))\n\n    model.add(CuDNNLSTM(512, return_sequences=True,name='lstm3'))\n    model.add(Dropout(0.2))\n\n    model.add(Dense(32, activation='relu', name='dense1'))\n    model.add(Dropout(0.2))\n\n    model.add(Dense(1, activation='softmax', name='denseoutput2'))\n\n\n    # Compile model\n    model.compile(\n        loss='mse',\n        optimizer='adam',\n        metrics=['accuracy'],\n    )\n    return model \n<\/code><\/pre>\n\n<p>And its fitting:<\/p>\n\n<pre><code>model=create_model()\nhistory=model.fit(xtrain, ytrain,batch_size=1, epochs=5, validation_data=(xtest, ytest), verbose=1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53100197","CreationDate":"2018-11-01T11:16:49.957","ParentId":null,"OwnerUserId":"8208006","Title":null,"Body":"<p>Building of keras layers includes specifying the shape of passed array, here the array shape to be trained is (3,2) with 1607 samples,<\/p>\n\n<pre><code>input_shape = (3,2)\nX = LSTM(124, activation = 'sigmoid', name='layer1', dropout = 0.4) (temp)\n<\/code><\/pre>\n\n<p>If you want to use stacked LSTM, you can use this<\/p>\n\n<pre><code>input_shape = (3,2)\n    X = LSTM(124, activation = 'sigmoid', name='layer1', dropout = 0.4,return_sequences=True) (temp)\n    X = LSTM(64, activation = 'sigmoid', name='layer2', dropout = 0.4) (X)\n<\/code><\/pre>\n\n<p><strong>edit<\/strong><\/p>\n\n<pre><code>def create_model():\n    model = keras.models.Sequential()\n\n    model.add(keras.layers.CuDNNLSTM(512, input_shape=(3,2), return_sequences=True, name='inputlstm1'))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.CuDNNLSTM(512, return_sequences=True,name='lstm2'))\n    model.add(keras.layers.Dropout(0.2))\n\n    # The last layer of Stacked LSTM need not to return the input sequences\n    model.add(keras.layers.CuDNNLSTM(512,name='lstm3'))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.Dense(32, activation='relu', name='dense1'))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.Dense(1, activation='softmax', name='denseoutput2'))\n\n\n    # Compile model\n    model.compile(\n        loss='mse',\n        optimizer='adam',\n        metrics=['accuracy'],\n    )\n    return model \n<\/code><\/pre>\n"},{"AnswerId":"53099998","CreationDate":"2018-11-01T11:05:19.327","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You can set <code>batch_size = 1<\/code> and Keras will take care of it. Just put the whole array to the model input.<\/p>\n"}]}
+{"QuestionId":53100206,"AnswerCount":1,"Tags":"<git><tensorflow>","CreationDate":"2018-11-01T11:17:26.727","AcceptedAnswerId":53103169.0,"OwnerUserId":4003077.0,"Title":"How to clone specific version of TensorFlow Models + Reasearch Models","Body":"<p>I am trying to use a specific version of <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">TensorFlow Models<\/a> (1.8.0) for object detection purpose. As mentioned in that page, Research Models are not officially supported or available in release branches. So, when I use the following command to clone the git repository:<\/p>\n\n<pre><code>git clone --branch r1.8.0 https:\/\/github.com\/tensorflow\/models.git\n<\/code><\/pre>\n\n<p>After downloading about 600MB, the <code>Research<\/code> directory doesn\u2019t exist in <code>models<\/code> directory.<\/p>\n\n<p>How can I get <code>Research<\/code> directory which is compatible with r1.8.0<\/p>\n","answers":[{"AnswerId":"53103169","CreationDate":"2018-11-01T14:20:32.630","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>You can checkout a few commits before r1.8.0:  <code>git checkout r1.8.0~~~<\/code><\/p>\n\n<p>The <code>models<\/code> directory was intentionally deleted. There is a <a href=\"https:\/\/github.com\/tensorflow\/models\/pull\/4752\" rel=\"nofollow noreferrer\">pull request<\/a> which says \"We remove research models from tagged releases, but they should remain in master indefinitely. Closing this under the assumption that it was made in error.\" That description doesn't tell me much. Would appreciate if anybody would explain why the tagged releases shouldn't contain research models.<\/p>\n"}]}
+{"QuestionId":53101182,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T12:21:03.707","AcceptedAnswerId":null,"OwnerUserId":7149097.0,"Title":"How to inherit a typical Tensorflow model class to reuse most graph building code?","Body":"<p>Typical tensorflow model class looks like this:<\/p>\n\n<pre><code>class Model:\n    def __init__(self):\n        build()\n    def build(self):\n        self.x = tf.placeholder()\n        self.y = f(self.x)\n        self.z = g(self.y)\n<\/code><\/pre>\n\n<p>If we need a slight modification (ie, change <code>self.y=f(self.x)<\/code> to <code>slef.y=h(self.x)<\/code>), we would really like to inherit this <code>Model<\/code> class and add some code to do this.<\/p>\n\n<p>However, once the <code>build<\/code> function was called, a complete graph is built. Overriding an attribute won't change the graph structure. Is there any way to do this work neatly?  <\/p>\n","answers":[{"AnswerId":"53101308","CreationDate":"2018-11-01T12:29:43.323","ParentId":null,"OwnerUserId":"51685","Title":null,"Body":"<p>You could parametrize <code>f<\/code> and <code>g<\/code> (or whatever you have) and pass them in the constructor:<\/p>\n\n<pre><code>class Model:\n    def __init__(self, f=default_f, g=default_g):\n        self.f = f\n        self.g = g\n        self.build()\n\n    def build(self):\n        self.x = tf.placeholder()\n        self.y = self.f(self.x)\n        self.z = self.g(self.y)\n<\/code><\/pre>\n\n<p>Or you can make them overridable class-level variables to avoid the constructor's signature getting bloated, but then you can't implicitly call <code>.build()<\/code> in the constructor:<\/p>\n\n<pre><code>class Model:\n    f = default_f\n    g = default_g\n\n    def build(self):\n        self.x = tf.placeholder()\n        self.y = self.f(self.x)\n        self.z = self.g(self.y)\n\n# ...\n\nm = Model()\nm.f = some_other_f\nm.build()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53101664,"AnswerCount":0,"Tags":"<machine-learning><neural-network><keras><loading>","CreationDate":"2018-11-01T12:51:55.253","AcceptedAnswerId":null,"OwnerUserId":1017290.0,"Title":"Loading weights after a training run in KERAS not recognising the highest level of accuracy achieved in previous run","Body":"<p>I conduct a training run and finish with the following output from a KERAS NN model. I have added a confusion matrix output for model quality<\/p>\n\n<p>...<\/p>\n\n<p>Epoch 09998: val_acc did not improve from 0.83780\nEpoch 9999\/10000\n12232\/12232 [==============================] - 1s 49us\/step - loss: 0.2656 - acc: 0.9133 - val_loss: 0.6134 - val_acc: 0.8051<\/p>\n\n<p>Epoch 09999: val_acc did not improve from 0.83780\nEpoch 10000\/10000\n12232\/12232 [==============================] - 1s 48us\/step - loss: 0.2655 - acc: 0.9124 - val_loss: 0.5918 - val_acc: 0.8283<\/p>\n\n<p>Epoch 10000: val_acc did not improve from 0.83780\n3058\/3058 [==============================] - 0s 46us\/step<\/p>\n\n<p>acc: 82.83%\nModel Quality\nTn: 806  Tp: 1727  Fp: 262  Fn: 263\nPrecision: 0.8683 Recall: 0.8678 Accuracy 0.8283 F score 0.8681<\/p>\n\n<p>I then make a change to a hyper parameter and reload everything and recompile using the following code<\/p>\n\n<pre><code># prep checkpointing\nmodel_file = output_file.rsplit('.',1)[0] + '_model.h5'\ncheckpoint = ModelCheckpoint(model_file, monitor='val_acc', verbose=1, \n                             save_best_only=True, save_weights_only=True, mode='max')\ncallbacks_list = [checkpoint]\nmodel.load_weights(model_file) # - commented out first time thru, reload second time thru\n\nadam = optimizers.Adam(lr=l_rate, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\nmodel.compile(loss='binary_crossentropy', optimizer=adam, metrics=['accuracy'])\n\n\nmodel.fit(X_train, Y_train, validation_data=(model_test_X, model_test_y),\n          batch_size = batch_s, epochs=num_epochs, \n          callbacks=callbacks_list, verbose=1)\n<\/code><\/pre>\n\n<p>Upon re-starting I would expect the best accuracy to be as it was above. In this instance val_acc = 0.83780<\/p>\n\n<p>However, after the first two epochs I get this output:<\/p>\n\n<p>Epoch 1\/10000\n12232\/12232 [==============================] - 1s 99us\/step - loss: 0.2747 - acc: 0.9097 - val_loss: 0.6191 - val_acc: 0.8143<\/p>\n\n<p>Epoch 00001: val_acc improved from -inf to 0.81426, saving model to Data_model.h5<\/p>\n\n<p>Epoch 2\/10000\n12232\/12232 [==============================] - 1s 42us\/step - loss: 0.2591 - acc: 0.9168 - val_loss: 0.6367 - val_acc: 0.8322<\/p>\n\n<p>Epoch 00002: val_acc improved from 0.81426 to 0.83224, saving model to Data_model.h5<\/p>\n\n<p>Epoch 3\/10000\n12232\/12232 [==============================] - 1s 44us\/step - loss: 0.2699 - acc: 0.9140 - val_loss: 0.6157 - val_acc: 0.8313<\/p>\n\n<p>Epoch 00003: val_acc did not improve from 0.83224\n......<\/p>\n\n<p>While I understand the model may start from a different place, my assumption was that the best level of accuracy (val_acc) would have been carried over from the previous run.<\/p>\n\n<p>My question is am i missing something?<\/p>\n","answers":[]}
+{"QuestionId":53102436,"AnswerCount":1,"Tags":"<anaconda><conda><pytorch><pycuda>","CreationDate":"2018-11-01T13:37:54.607","AcceptedAnswerId":null,"OwnerUserId":4346687.0,"Title":"PyTorch Cuda with anaconda not available","Body":"<p>I'm using anaconda to regulate my environment,\nfor a project i have to use my GPU for network training.\nI use pytorch for my project and i'm trying to get CUDA working.<\/p>\n\n<p>I installed cudatoolkit, numba, cudnn<\/p>\n\n<p>still, when i try this command:<\/p>\n\n<pre><code>torch.cuda.is_available()\n<\/code><\/pre>\n\n<p>I get \"False\" as output.\nThis is my environment:<\/p>\n\n<pre><code># Name                    Version                   Build  Channel\nblas                      1.0                         mkl  \nbzip2                     1.0.6                h470a237_2    conda-forge\nca-certificates           2018.03.07                    0  \ncairo                     1.14.12              he6fea26_5    conda-forge\ncertifi                   2018.8.24                py35_1  \ncffi                      1.11.5           py35he75722e_1  \ncloudpickle               0.5.5                    py35_0  \ncudatoolkit               9.2                           0    anaconda\ncudnn                     7.2.1                 cuda9.2_0    anaconda\ncycler                    0.10.0                     py_1    conda-forge\ncython                    0.28.5           py35hf484d3e_0    anaconda\ndask-core                 0.19.2                   py35_0  \ndbus                      1.13.0               h3a4f0e9_0    conda-forge\ndecorator                 4.3.0                    py35_0  \nexpat                     2.2.5                hfc679d8_2    conda-forge\nffmpeg                    4.0.2                ha0c5888_1    conda-forge\nfontconfig                2.13.1               h65d0f4c_0    conda-forge\nfreetype                  2.9.1                h6debe1e_4    conda-forge\ngettext                   0.19.8.1             h5e8e0c9_1    conda-forge\ngiflib                    5.1.4                h470a237_1    conda-forge\nglib                      2.55.0               h464dc38_2    conda-forge\ngmp                       6.1.2                hfc679d8_0    conda-forge\ngnutls                    3.5.19               h2a4e5f8_1    conda-forge\ngraphite2                 1.3.12               hfc679d8_1    conda-forge\ngst-plugins-base          1.12.5               hde13a9d_0    conda-forge\ngstreamer                 1.12.5               h61a6719_0    conda-forge\nharfbuzz                  1.9.0                h08d66d9_0    conda-forge\nhdf5                      1.10.2               hc401514_2    conda-forge\nicu                       58.2                 hfc679d8_0    conda-forge\nimageio                   2.4.1                    py35_0  \nintel-openmp              2019.0                      118  \njasper                    1.900.1              hff1ad4c_5    conda-forge\njpeg                      9c                   h470a237_1    conda-forge\nkiwisolver                1.0.1            py35h2d50403_2    conda-forge\nlibedit                   3.1.20170329         h6b74fdf_2  \nlibffi                    3.2.1                hd88cf55_4  \nlibgcc-ng                 8.2.0                hdf63c60_1  \nlibgfortran               3.0.0                         1    conda-forge\nlibgfortran-ng            7.3.0                hdf63c60_0  \nlibiconv                  1.15                 h470a237_3    conda-forge\nlibopenblas               0.3.3                h5a2b251_3  \nlibpng                    1.6.35               ha92aebf_2    conda-forge\nlibstdcxx-ng              8.2.0                hdf63c60_1  \nlibtiff                   4.0.9                he6b73bb_2    conda-forge\nlibuuid                   2.32.1               h470a237_2    conda-forge\nlibwebp                   0.5.2                         7    conda-forge\nlibxcb                    1.13                 h470a237_2    conda-forge\nlibxml2                   2.9.8                h422b904_5    conda-forge\nllvmlite                  0.24.0           py35hdbcaa40_0  \nmatplotlib                3.0.0            py35h0b34cb6_1    conda-forge\nmkl                       2019.0                      118  \nmkl_fft                   1.0.6                    py35_0    conda-forge\nmkl_random                1.0.1                    py35_0    conda-forge\nncurses                   6.1                  hf484d3e_0  \nnettle                    3.3                           0    conda-forge\nnetworkx                  2.1                      py35_0  \nninja                     1.8.2            py35h6bb024c_1  \nnumba                     0.39.0           py35h04863e7_0  \nnumpy                     1.15.2           py35h1d66e8a_0  \nnumpy-base                1.15.2           py35h81de0dd_0  \nolefile                   0.46                     py35_0  \nopenblas                  0.2.20                        8    conda-forge\nopencv                    3.4.1            py35h6fd60c2_1  \nopencv-python             3.4.3.18                  &lt;pip&gt;\nopenh264                  1.7.0                         0    conda-forge\nopenssl                   1.0.2p               h14c3975_0  \npandas                    0.23.4           py35h04863e7_0  \npcre                      8.41                 hfc679d8_3    conda-forge\npillow                    5.2.0            py35heded4f4_0  \nPillow                    5.3.0                     &lt;pip&gt;\npip                       10.0.1                   py35_0  \npixman                    0.34.0               h470a237_3    conda-forge\npthread-stubs             0.4                  h470a237_1    conda-forge\npycparser                 2.19                     py35_0  \npyparsing                 2.2.2                      py_0    conda-forge\npyqt                      5.6.0            py35h8210e8a_7    conda-forge\npython                    3.5.6                hc3d631a_0  \npython-dateutil           2.7.3                      py_0    conda-forge\npytorch                   0.4.1           py35_py27__9.0.176_7.1.2_2    pytorch\npytz                      2018.5                   py35_0  \npywavelets                1.0.0            py35hdd07704_0  \nqt                        5.6.2                hf70d934_9    conda-forge\nreadline                  7.0                  h7b6447c_5  \nscikit-image              0.14.0           py35hf484d3e_1  \nscipy                     1.1.0            py35hfa4b5c9_1  \nsetuptools                40.2.0                   py35_0  \nsip                       4.18.1           py35hfc679d8_0    conda-forge\nsix                       1.11.0                   py35_1    conda-forge\nsqlite                    3.25.2               h7b6447c_0  \ntk                        8.6.8                hbc83047_0  \ntoolz                     0.9.0                    py35_0  \ntorchvision               0.1.9            py35h72e4c6f_1    soumith\ntornado                   5.1.1            py35h470a237_0    conda-forge\nwheel                     0.31.1                   py35_0  \nx264                      1!152.20180717       h470a237_1    conda-forge\nxorg-kbproto              1.0.7                h470a237_2    conda-forge\nxorg-libice               1.0.9                h470a237_4    conda-forge\nxorg-libsm                1.2.2                h8c8a85c_6    conda-forge\nxorg-libx11               1.6.6                h470a237_0    conda-forge\nxorg-libxau               1.0.8                h470a237_6    conda-forge\nxorg-libxdmcp             1.1.2                h470a237_7    conda-forge\nxorg-libxext              1.3.3                h470a237_4    conda-forge\nxorg-libxrender           0.9.10               h470a237_2    conda-forge\nxorg-renderproto          0.11.1               h470a237_2    conda-forge\nxorg-xextproto            7.3.0                h470a237_2    conda-forge\nxorg-xproto               7.0.31               h470a237_7    conda-forge\nxz                        5.2.4                h14c3975_4  \nzlib                      1.2.11               ha838bed_2 \n<\/code><\/pre>\n\n<p>My desktop has a NVIDIA GeForce GTX 970 (so it is cuda available)\nAlso for some reason, as you can see here:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4yJdm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4yJdm.png\" alt=\"About my system\"><\/a><\/p>\n\n<p>My graphics card doesn't show, however when using the<\/p>\n\n<pre><code>lspci -v\n<\/code><\/pre>\n\n<p>command, i can see my graphics card there. Don't know if that has something to do with it. Anyone knows how i can fix this? <\/p>\n","answers":[{"AnswerId":"53102992","CreationDate":"2018-11-01T14:09:00.707","ParentId":null,"OwnerUserId":"4346687","Title":null,"Body":"<p>FIXED:\nI installed CUDA seperately, not through anaconda and now it works. If anyone knows why it doesn't work when installing cuda in anaconda, feel free to answer<\/p>\n"}]}
+{"QuestionId":53102664,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T13:50:06.663","AcceptedAnswerId":53182833.0,"OwnerUserId":567879.0,"Title":"Calculating F1 score, precision, recall in tfhub retraining script","Body":"<p>I am using tensorflow hub for image retraining classification task. The tensorflow script <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">retrain.py<\/a> by default calculates cross_entropy and accuracy. <\/p>\n\n<pre><code>train_accuracy, cross_entropy_value = sess.run([evaluation_step, cross_entropy],feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth})\n<\/code><\/pre>\n\n<p>I would like to get F1 score, precision, recall and confusion matrix. How could I get these values using this script ? <\/p>\n","answers":[{"AnswerId":"53182833","CreationDate":"2018-11-07T02:27:47.083","ParentId":null,"OwnerUserId":"9535747","Title":null,"Body":"<p>Below I include a method to calculate desired metrics using <em>scikit-learn<\/em> package.<\/p>\n\n<p>You can calculate F1 score, precision and recall using <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.precision_recall_fscore_support.html\" rel=\"noreferrer\">precision_recall_fscore_support<\/a> method and the confusion matrix using <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.confusion_matrix.html\" rel=\"noreferrer\">confusion_matrix<\/a> method:<\/p>\n\n<pre><code>from sklearn.metrics import precision_recall_fscore_support, confusion_matrix\n<\/code><\/pre>\n\n<p>Both methods take two 1D array-like objects which store ground truth and predicted labels respectively.<\/p>\n\n<p>In the code provided, ground-truth labels for training data are stored in <code>train_ground_truth<\/code> variable which is defined in lines <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1054\" rel=\"noreferrer\">1054<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1060\" rel=\"noreferrer\">1060<\/a>, while <code>validation_ground_truth<\/code> stores ground-truth labels for validation data and is defined in line <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1087\" rel=\"noreferrer\">1087<\/a>.<\/p>\n\n<p>The tensor that calculates predicted class labels is defined and returned by <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L804\" rel=\"noreferrer\">add_evaluation_step<\/a> function. You can modify line <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1034\" rel=\"noreferrer\">1034<\/a> in order to capture that tensor object:<\/p>\n\n<pre><code>evaluation_step, prediction = add_evaluation_step(final_tensor, ground_truth_input)\n# now prediction stores the tensor object that \n# calculates predicted class labels\n<\/code><\/pre>\n\n<p>Now you can update line <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1076\" rel=\"noreferrer\">1076<\/a> in order to evaluate <code>prediction<\/code> when calling <code>sess.run()<\/code>:<\/p>\n\n<pre><code>train_accuracy, cross_entropy_value, train_predictions = sess.run(\n    [evaluation_step, cross_entropy, prediction],\n    feed_dict={bottleneck_input: train_bottlenecks,\n               ground_truth_input: train_ground_truth})\n\n# train_predictions now stores class labels predicted by model\n\n# calculate precision, recall and F1 score\n(train_precision,\n train_recall,\n train_f1_score, _) = precision_recall_fscore_support(y_true=train_ground_truth,\n                                                      y_pred=train_predictions,\n                                                      average='micro')\n# calculate confusion matrix\ntrain_confusion_matrix = confusion_matrix(y_true=train_ground_truth,\n                                          y_pred=train_predictions)\n<\/code><\/pre>\n\n<p>Similarly, you can compute metrics for validation subset by modifying line <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L1095\" rel=\"noreferrer\">1095<\/a>:<\/p>\n\n<pre><code>validation_summary, validation_accuracy, validation_predictions = sess.run(\n    [merged, evaluation_step, prediction],\n    feed_dict={bottleneck_input: validation_bottlenecks,\n               ground_truth_input: validation_ground_truth})\n\n# validation_predictions now stores class labels predicted by model\n\n# calculate precision, recall and F1 score\n(validation_precision,\n validation_recall,\n validation_f1_score, _) = precision_recall_fscore_support(y_true=validation_ground_truth,\n                                                           y_pred=validation_predictions,\n                                                           average='micro')\n# calculate confusion matrix\nvalidation_confusion_matrix = confusion_matrix(y_true=validation_ground_truth,\n                                               y_pred=validation_predictions)\n<\/code><\/pre>\n\n<p>Finally, the code calls <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py#L825\" rel=\"noreferrer\">run_final_eval<\/a> to evaluate trained model on test data. In this function, <code>prediction<\/code> and <code>test_ground_truth<\/code> are already defined, so you only need to include code to calculate required metrics:<\/p>\n\n<pre><code>test_accuracy, predictions = eval_session.run(\n    [evaluation_step, prediction],\n    feed_dict={\n        bottleneck_input: test_bottlenecks,\n        ground_truth_input: test_ground_truth\n    })\n\n# calculate precision, recall and F1 score\n(test_precision,\n test_recall,\n test_f1_score, _) = precision_recall_fscore_support(y_true=test_ground_truth,\n                                                     y_pred=predictions,\n                                                     average='micro')\n# calculate confusion matrix\ntest_confusion_matrix = confusion_matrix(y_true=test_ground_truth,\n                                         y_pred=predictions)\n<\/code><\/pre>\n\n<p>Note that the provided code calculates <em>global<\/em> F1-scores by setting <code>average='micro'<\/code>. The different averaging methods that are supported by scikit-learn package are described in <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/model_evaluation.html#multiclass-and-multilabel-classification\" rel=\"noreferrer\">User Guide<\/a>.<\/p>\n"}]}
+{"QuestionId":53102921,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-01T14:04:36.760","AcceptedAnswerId":null,"OwnerUserId":10580626.0,"Title":"Why do I need a dictionary datatype for input data in TensorFlow","Body":"<pre><code>(X_train, y_train), (X_test, y_test) = tf.keras.datasets.mnist.load_data()\nX_train = X_train.astype(np.float32).reshape(-1, 28*28) \/ 255.0\nX_test = X_test.astype(np.float32).reshape(-1, 28*28) \/ 255.0\ny_train = y_train.astype(np.int32)\ny_test = y_test.astype(np.int32)\nX_valid, X_train = X_train[:5000], X_train[5000:]\ny_valid, y_train = y_train[:5000], y_train[5000:]\n\ndnn_clf = tf.estimator.DNNClassifier(hidden_units=[300,100], n_classes=10,\nfeature_columns=feature_cols)\n\ninput_fn = tf.estimator.inputs.numpy_input_fn(\nx={\"X\": X_train}, y=y_train, num_epochs=40, batch_size=50, shuffle=True)\n\ndnn_clf.train(input_fn=input_fn)\n<\/code><\/pre>\n\n<p>My question is:\nWhat's the difference between<\/p>\n\n<p><code>x={\"X\": X_train}, y=y_train, num_epochs=40, batch_size=50, shuffle=True)<\/code><\/p>\n\n<p>and<\/p>\n\n<p><code>x=X_train, y=y_train, num_epochs=40, batch_size=50, shuffle=True)<\/code><\/p>\n\n<p>Why can't I just use a <strong>list<\/strong> <code>X_train<\/code> for the input <code>x<\/code>, but have to convert <code>X_train<\/code> to a <strong>dictionary<\/strong> <code>x={\"X\": X_train}<\/code><\/p>\n\n<p>Why I don't need to do the same thing for <code>y<\/code>?<\/p>\n","answers":[{"AnswerId":"53103699","CreationDate":"2018-11-01T14:49:53.537","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>According to documentation TensorFlow 1.11 for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/inputs\/numpy_input_fn\" rel=\"nofollow noreferrer\">numpy_input_fn<\/a>, you can use both approaches. But I've tested it for TensorFlow 1.9 and it didn't work. So, it seems that this functionality was added later. <\/p>\n\n<p>If I understood, another question was \"why x parameter is a map and y is just an array\". The thing is that model can have multiple input feature columns\/tensors, but the output tensor can be only one. That's true for, at least, standard estimators and for general use-cases. So, you need to somehow specify multiple inputs and the most simple way is just to pass a map. <\/p>\n"}]}
+{"QuestionId":53103135,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><time-series><forecasting>","CreationDate":"2018-11-01T14:18:29.403","AcceptedAnswerId":null,"OwnerUserId":4983920.0,"Title":"time-series prediction for price forecasting (problems with predictions)","Body":"<p>I am working on a project for price movement forecasting and I am stuck with poor quality predictions.<\/p>\n\n<p>At every time-step I am using an LSTM to predict the next 10 time-steps. The input is the sequence of the last 45-60 observations. I tested several different ideas, but they all seems to give similar results. The model is trained to minimize MSE.<\/p>\n\n<p>For each idea I tried a model predicting 1 step at a time where each prediction is fed back as an input for the next prediction, and a model directly predicting the next 10 steps(multiple outputs). For each idea I also tried using as input just the moving average of the previous prices, and extending the input to input the order book at those time-steps.\nEach time-step corresponds to a second.<\/p>\n\n<p>These are the results so far:<\/p>\n\n<p>1- The first attempt was using as input the moving average of the last N steps, and predict the moving average of the next 10. \nAt time t, I use the ground truth value of the price and use the model to predict t+1....t+10<\/p>\n\n<p>This is the result<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YOXxu.png\" rel=\"nofollow noreferrer\">Predicting moving average<\/a><\/p>\n\n<p>On closer inspection we can see what's going wrong:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/m4Xb5.png\" rel=\"nofollow noreferrer\">Prediction seems to be a flat line. Does not care much about the input data.<\/a><\/p>\n\n<p>2) The second attempt was trying to predict differences, instead of simply the price movement. The input this time instead of simply being X[t] (where X is my input matrix) would be X[t]-X[t-1].\nThis did not really help.\nThe plot this time looks like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/toFGz.png\" rel=\"nofollow noreferrer\">Predicting differences<\/a><\/p>\n\n<p>But on close inspection, when plotting the differences, the predictions are always basically 0.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eSqGU.png\" rel=\"nofollow noreferrer\">Plot of differences<\/a><\/p>\n\n<p>At this point, I am stuck here and running our of ideas to try. I was hoping someone with more experience in this type of data could point me in the right direction.<\/p>\n\n<p>Am I using the right objective to train the model? Are there any details when dealing with this type of data that I am missing?\nAre there any \"tricks\" to prevent your model from always predicting similar values to what it last saw? (They do incur in low error, but they become meaningless at that point).<\/p>\n\n<p>At least just a hint on where to dig for further info would be highly appreciated.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53105206","CreationDate":"2018-11-01T16:16:58.687","ParentId":null,"OwnerUserId":"3967806","Title":null,"Body":"<blockquote>\n  <p>Am I using the right objective to train the model? <\/p>\n<\/blockquote>\n\n<p>Yes, but LSTM are always very tricky for forecasting time series. And are very prone to overfitting compared to other time series models. <\/p>\n\n<blockquote>\n  <p>Are there any details when dealing with this type of data that I am missing? <\/p>\n  \n  <p>Are there any \"tricks\" to prevent your model from always predicting similar values to what it last saw?<\/p>\n<\/blockquote>\n\n<p>I haven't seen your code, or the details of the LSTM you are using. Make sure you are using a very small network, and you are avoiding overfitting. Make sure that after you differenced the data - you then reintegrate it before evaluating the final forecast. <\/p>\n\n<p>On trick to try to build a model that forecasts 10 steps ahead directly instead of building a one-step ahead model and then forecasting recursively. <\/p>\n"}]}
+{"QuestionId":53103824,"AnswerCount":1,"Tags":"<python><tensorflow><keras><categorical-data><loss-function>","CreationDate":"2018-11-01T14:57:09.647","AcceptedAnswerId":null,"OwnerUserId":274627.0,"Title":"What Keras loss function to use for ternary classification?","Body":"<p>My LSTM input samples consist of real numbers in [-1, 1] range and the targets (categories) are -1, 0 and 1.<\/p>\n\n<p>I've chosen sparse_categorical_crossentropy for the loss function but it seems it doesn't support negative numbers - it says \"InvalidArgumentError: Received a label value of -1 which is outside the valid range of [0, 1)\" when I start training.<\/p>\n\n<p>What loss function should I use instead for this task?<\/p>\n\n<p>I also can't understand why does it say [0, 1) which means a number equal or higher than zero but smaller than 1 AFAIK but <a href=\"https:\/\/jovianlin.io\/cat-crossentropy-vs-sparse-cat-crossentropy\/\" rel=\"nofollow noreferrer\">what I've read<\/a> about sparse_categorical_crossentropy says it's meant to support any integer number.<\/p>\n","answers":[{"AnswerId":"53103973","CreationDate":"2018-11-01T15:06:34.880","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Your targets have to be integers from <code>0<\/code> to <code>num_classes - 1<\/code>, they can't be negative numbers, as you are doing classification. This is kind of the standard for multi-class classification and you have to follow it.<\/p>\n\n<p>Also make sure that your last layer outputs 3 numbers with a softmax activation, which will represent a distribution over classes.<\/p>\n"}]}
+{"QuestionId":53104299,"AnswerCount":1,"Tags":"<keras><forecasting>","CreationDate":"2018-11-01T15:24:57.530","AcceptedAnswerId":53105163.0,"OwnerUserId":9524145.0,"Title":"ambiguity of understanding SMAPE","Body":"<p>I have implemented several regression forecast approaches and now I want to compare them. I picked the MAE, RMSE and SMAPE ratings. My result looks like follows:<\/p>\n\n<ol>\n<li>Approach 1: MAE= 0,6 , RMSE= 0,9 and SMAPE 531<\/li>\n<li>Approach 2: MAE= 3,0 ,RMSE= 6,1 and SMAPE 510<\/li>\n<li>Approach 3: MAE= 10,1 , RMSE= 17,00 and SMAPE 420<\/li>\n<\/ol>\n\n<p>When I plot my predictions and compare them with my test set, I can see that Approach 1 > Approach 2 > Approach 3. This is also evident from the values of MAE and RMSE. But I thought that the lower the resulting SMAPE, the better the prediction. <\/p>\n\n<p>Did I misunderstand SMAPE?\nSince there is no predefined method in phyton, my SMAPE calculation looks like this:<\/p>\n\n<pre><code>def smape(A, F):\n    return 100\/len(A) * np.sum(2 * np.abs(F - A) \/ (np.abs(A) + np.abs(F)))\n<\/code><\/pre>\n\n<p>Or is the calculation wrong?<\/p>\n\n<p>thanks in advance<\/p>\n","answers":[{"AnswerId":"53105163","CreationDate":"2018-11-01T16:13:37.437","ParentId":null,"OwnerUserId":"9524145","Title":null,"Body":"<p>Okay maybe the method was wrong.. instead i used this one from Kaggle:<\/p>\n\n<pre><code>from numba import jit\nimport math\n\n@jit\ndef smape_fast(y_true, y_pred):\n    out = 0\n    for i in range(y_true.shape[0]):\n        a = y_true[i]\n        b = y_pred[i]\n        c = a+b\n        if c == 0:\n            continue\n        out += math.fabs(a - b) \/ c\n    out *= (200.0 \/ y_true.shape[0])\n    return out\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.kaggle.com\/c\/web-traffic-time-series-forecasting\/discussion\/37232\" rel=\"nofollow noreferrer\">URL<\/a>\nnow my results look more plausible from SMAPE compared to MAE and RMSE<\/p>\n"}]}
+{"QuestionId":53104300,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><google-cloud-ml><object-detection-api>","CreationDate":"2018-11-01T15:25:11.183","AcceptedAnswerId":null,"OwnerUserId":229005.0,"Title":"How to continue training an object detection model using Tensorflow Object Detection API?","Body":"<p>I'm using <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection#tensorflow-object-detection-api\" rel=\"nofollow noreferrer\">Tensorflow Object Detection API<\/a> to train an object detection model using transfer learning. Specifically, I'm using <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md#coco-trained-models\" rel=\"nofollow noreferrer\">ssd_mobilenet_v1_fpn_coco from the model zoo<\/a>, and using the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/samples\/configs\/ssd_mobilenet_v1_fpn_shared_box_predictor_640x640_coco14_sync.config\" rel=\"nofollow noreferrer\">sample pipeline provided<\/a>, having of course replaced the placeholders with actual links to my training and eval tfrecords and labels.<\/p>\n\n<p>I was able able to successfully train a model on my ~5000 images (and corresponding bounding boxes) using the above pipeline (I'm mainly using Google's ML Engine on TPU, if revelant).<\/p>\n\n<p>Now, I prepared an additional ~2000 images, and would like continue training my model with those new images, without restarting from scratch (training the initial model took ~6h of TPU time). How can I do that?<\/p>\n","answers":[{"AnswerId":"53138979","CreationDate":"2018-11-04T08:24:01.940","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>You have two options, in both you need to change the <code>input_path<\/code> of the <code>train_input_reader<\/code> of your new dataset:<\/p>\n\n<ol>\n<li>When specifying a checkpoint to fine-tune in the training configuration, specify the checkpoint of your trained model<\/li>\n<\/ol>\n\n<pre><code>train_config{\n    fine_tune_checkpoint: &lt;path_to_your_checkpoint&gt;\n    fine_tune_checkpoint_type: \"detection\"\n    load_all_detection_checkpoint_vars: true\n}\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Simply keep using the same configuration (except the <code>train_input_reader<\/code>) with the same <code>model_dir<\/code> of your previous model. That way, the API will create a graph and will check whether a checkpoint already exists in <code>model_dir<\/code> and fits the graph. If so - it will restore it and continue training it.<\/li>\n<\/ol>\n\n<p>Edit: fine_tune_checkpoint_type was previously set as true by mistake, while it should be \"detection\" or \"classification\" in general, and \"detection\" in this specific case. Thanks Krish for noticing.<\/p>\n"},{"AnswerId":"53104638","CreationDate":"2018-11-01T15:44:26.330","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>I haven't retrained the object detection model on a new dataset, but it looks like\nincreasing the number of training steps <code>train_config.num_steps<\/code> in the config file and also adding images in the tfrecord files should be enough for that.<\/p>\n"}]}
+{"QuestionId":53104887,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-lite>","CreationDate":"2018-11-01T15:58:18.033","AcceptedAnswerId":53105809.0,"OwnerUserId":9104473.0,"Title":"Access internal tensors and add a new node to a tflite model?","Body":"<p>I am fairly new to TensorFlow and TensorFlow Lite. I have followed the tutorials on how to quantize and convert the model to fixed point calculations using <code>toco<\/code>. Now I have a <code>tflite<\/code> file which is supposed to perform only fixed point operations. I have two questions<\/p>\n\n<ol>\n<li>How do I test this in python? How do i access all the operations and results in the tflite file?<\/li>\n<li>Is there a way to add a new node or operation in this tflite file? If so how?<\/li>\n<\/ol>\n\n<p>I would be really grateful if anyone could guide me.<\/p>\n\n<p>Thanks and Regards,<br>\nAbhinav George<\/p>\n","answers":[{"AnswerId":"53105809","CreationDate":"2018-11-01T16:52:04.160","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<blockquote>\n  <p>Is there a way to add a new node or operation in this tflite file? If so how?<\/p>\n<\/blockquote>\n\n<p>Unfortunately, <strong>no<\/strong>, and it is actually a <em>good thing<\/em>. TF-Lite was designed to be extremely light yet effective, using mapped files, flat buffers, static execution plan and so on to decrease memory footprint. The cost of that is that you loose any flexibility of TensorFlow.<\/p>\n\n<p>TF-Lite is a framework for deployment. However, earlier on <a href=\"https:\/\/www.youtube.com\/watch?v=ByJnpbDd-zc\" rel=\"nofollow noreferrer\">Google IO<\/a>, the TF team mentioned the possibility of on-device training, so maybe some kind of flexibility will be available in the future, but not now.<\/p>\n\n<hr>\n\n<blockquote>\n  <p>How do I test this in python? How do i access all the operations and results in the tflite file?<\/p>\n<\/blockquote>\n\n<p>You <strong>cannot<\/strong> access all internal operations, only inputs and outputs. The reason is simple: <em>the internal tensors wouldn't be saved, since the memory sections for them are also used for other operations<\/em> (which is why the memory footprint of it is so low).<\/p>\n\n<p>If you just want to see the outputs, you can use the Python API as below (the code is self explanatory):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import pprint\nfrom tensorflow.contrib.lite.python import interpreter as interpreter_wrapper\n\n# Load the model and allocate the static memory plan\ninterpreter = interpreter_wrapper.Interpreter(model_path='model.tflite')\ninterpreter.allocate_tensors()\n\n# print out the input details\ninput_details = interpreter.get_input_details()\nprint('input_details:')\npp = pprint.PrettyPrinter(indent=2)\npp.pprint(input_details)\n\n# print out the output details\noutput_details = interpreter.get_output_details()\nprint('output_details:')\npp = pprint.PrettyPrinter(indent=2)\npp.pprint(output_details)\n\n# set input (img is a `numpy array`)\ninterpreter.set_tensor(input_details[0]['index'], img)\n\n# forward pass\ninterpreter.invoke()\n\n# get output of the network\noutput = interpreter.get_tensor(output_details[0]['index'])\n<\/code><\/pre>\n\n<hr>\n\n<blockquote>\n  <p>What if I call <code>interpreter.get_tensor<\/code> for non-input and non-output tensors?<\/p>\n<\/blockquote>\n\n<p>You will not get the actual data that contained in that tensor after execution of the corresponding operation. As mentioned earlier, the memory section for tensors are shared with other tensors for maximum efficiency. <\/p>\n"}]}
+{"QuestionId":53105294,"AnswerCount":2,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2018-11-01T16:21:38.850","AcceptedAnswerId":53211660.0,"OwnerUserId":6046281.0,"Title":"Implementing a batch dependent loss in Keras","Body":"<p>I have an autoencoder set up in Keras. I want to be able to weight the features of the input vector according to a predetermined 'precision' vector. This continuous valued vector has the same length as the input, and each element lies in the range <code>[0, 1]<\/code>, corresponding to the confidence in the corresponding input element, where 1 is completely confident and 0 is no confidence.<\/p>\n\n<p>I have a precision vector for every example.<\/p>\n\n<p>I have defined a loss that takes into account this precision vector. Here, reconstructions of low-confidence features are down-weighted.<\/p>\n\n<pre><code>def MAEpw_wrapper(y_prec):\n    def MAEpw(y_true, y_pred):\n        return K.mean(K.square(y_prec * (y_pred - y_true)))\n    return MAEpw\n<\/code><\/pre>\n\n<p>My issue is that the precision tensor <code>y_prec<\/code> depends on the batch. I want to be able to update <code>y_prec<\/code> according to the current batch so that each precision vector is correctly associated with its observation.<\/p>\n\n<p><strong>I have the done the following:<\/strong><\/p>\n\n<pre><code>global y_prec\ny_prec = K.variable(P[:32])\n<\/code><\/pre>\n\n<p>Here <code>P<\/code> is a numpy array containing all precision vectors with the indices corresponding to the examples. I initialize <code>y_prec<\/code> to have the correct shape for a batch size of 32. I then define the following <code>DataGenerator<\/code>:<\/p>\n\n<pre><code>class DataGenerator(Sequence):\n\n    def __init__(self, batch_size, y, shuffle=True):\n\n        self.batch_size = batch_size\n\n        self.y = y\n\n        self.shuffle = shuffle\n        self.on_epoch_end()\n\n    def on_epoch_end(self):\n        self.indexes = np.arange(len(self.y))\n\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n\n    def __len__(self):\n        return int(np.floor(len(self.y) \/ self.batch_size))\n\n    def __getitem__(self, index):\n        indexes = self.indexes[index * self.batch_size: (index+1) * self.batch_size]\n\n        # Set precision vector.\n        global y_prec\n        new_y_prec = K.variable(P[indexes])\n        y_prec = K.update(y_prec, new_y_prec)\n\n        # Get training examples.\n        y = self.y[indexes]\n\n        return y, y\n<\/code><\/pre>\n\n<p>Here I am aiming to update <code>y_prec<\/code> in the same function that generates the batch. This seems to be updating <code>y_prec<\/code> as expected. I then define my model architecture:<\/p>\n\n<pre><code>dims = [40, 20, 2]\n\nmodel2 = Sequential()\nmodel2.add(Dense(dims[0], input_dim=64, activation='relu'))\nmodel2.add(Dense(dims[1], input_dim=dims[0], activation='relu'))\nmodel2.add(Dense(dims[2], input_dim=dims[1], activation='relu', name='bottleneck'))\nmodel2.add(Dense(dims[1], input_dim=dims[2], activation='relu'))\nmodel2.add(Dense(dims[0], input_dim=dims[1], activation='relu'))\nmodel2.add(Dense(64, input_dim=dims[0], activation='linear'))\n<\/code><\/pre>\n\n<p>And finally, I compile and run:<\/p>\n\n<pre><code>model2.compile(optimizer='adam', loss=MAEpw_wrapper(y_prec))\nmodel2.fit_generator(DataGenerator(32, digits.data), epochs=100)\n<\/code><\/pre>\n\n<p>Where <code>digits.data<\/code> is a numpy array of observations.<\/p>\n\n<p>However, this ends up defining separate graphs:<\/p>\n\n<pre><code>StopIteration: Tensor(\"Variable:0\", shape=(32, 64), dtype=float32_ref) must be from the same graph as Tensor(\"Variable_4:0\", shape=(32, 64), dtype=float32_ref).\n<\/code><\/pre>\n\n<p>I've scoured SO for a solution to my problem but nothing I've found works. Any help on how to do this properly is appreciated.<\/p>\n","answers":[{"AnswerId":"53211660","CreationDate":"2018-11-08T16:07:02.763","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>This autoencoder can be easily implemented using the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Keras functional API<\/a>. This will allow to have an additional input placeholder <code>y_prec_input<\/code>, which will be fed with the \"precision\" vector. The full source code can be found <a href=\"https:\/\/gist.github.com\/rvinas\/9e81ae0f17e61cc2c54b63f45fb07a28\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<hr>\n\n<p><strong>Data generator<\/strong><\/p>\n\n<p>First, let's reimplement your data generator as follows:<\/p>\n\n<pre><code>class DataGenerator(Sequence):\n    def __init__(self, batch_size, y, prec, shuffle=True):\n        self.batch_size = batch_size\n        self.y = y\n        self.shuffle = shuffle\n        self.prec = prec\n        self.on_epoch_end()\n\n    def on_epoch_end(self):\n        self.indexes = np.arange(len(self.y))\n        if self.shuffle:\n            np.random.shuffle(self.indexes)\n\n    def __len__(self):\n        return int(np.floor(len(self.y) \/ self.batch_size))\n\n    def __getitem__(self, index):\n        indexes = self.indexes[index * self.batch_size: (index + 1) * self.batch_size]\n        y = self.y[indexes]\n        y_prec = self.prec[indexes]\n        return [y, y_prec], y\n<\/code><\/pre>\n\n<p>Note that I got rid of the global variable. Now, instead, the precision vector <code>P<\/code> is provided as input argument (<code>prec<\/code>), and the generator yields an additional input that will be fed to the precision placeholder <code>y_prec_input<\/code> (see model definition).<\/p>\n\n<hr>\n\n<p><strong>Model<\/strong><\/p>\n\n<p>Finally, your model can be defined and trained as follows:<\/p>\n\n<pre><code>y_input = Input(shape=(input_dim,))\ny_prec_input = Input(shape=(1,))\nh_enc = Dense(dims[0], activation='relu')(y_input)\nh_enc = Dense(dims[1], activation='relu')(h_enc)\nh_enc = Dense(dims[2], activation='relu', name='bottleneck')(h_enc)\nh_dec = Dense(dims[1], activation='relu')(h_enc)\nh_dec = Dense(input_dim, activation='relu')(h_dec)\nmodel2 = Model(inputs=[y_input, y_prec_input], outputs=h_dec)\nmodel2.compile(optimizer='adam', loss=MAEpw_wrapper(y_prec_input))\n\n# Train model\nmodel2.fit_generator(DataGenerator(32, digits.data, P), epochs=100)\n<\/code><\/pre>\n\n<p>where <code>input_dim = digits.data.shape[1]<\/code>. Note that I also changed the output dimension of the decoder to <code>input_dim<\/code>, since it must match the input dimension.<\/p>\n"},{"AnswerId":"53182225","CreationDate":"2018-11-07T00:48:32.757","ParentId":null,"OwnerUserId":"1163373","Title":null,"Body":"<p>Try to test your code with worker=0 when you call fit_generator, if it works normally then threading is your problem.<\/p>\n\n<p>If threading is the cause, try this:<\/p>\n\n<pre><code># In the code that executes on the main thread\ngraph = tf.get_default_graph()\n\n# In code that executes in other threads(e.g. your generator)\nwith graph.as_default():\n    ...\n    ...\n    new_y_prec = K.variable(P[indexes])\n    y_prec = K.update(y_prec, new_y_prec)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53105325,"AnswerCount":0,"Tags":"<python><neural-network><keras><conv-neural-network>","CreationDate":"2018-11-01T16:23:24.393","AcceptedAnswerId":null,"OwnerUserId":5750238.0,"Title":"Keras conv nn predicting only one class?","Body":"<p>So I've been building a convolutional neural network. I'm trying to predict whether a boardgame state (10x10 matrix) will lead to a win (binary 0 or 1) or not.<\/p>\n\n<p>I have six million examples, which you would think would be enough, but clearly not, as my network is predicting all of one class...<\/p>\n\n<p>Is there something obvious I'm missing? I tried giving it even 10 examples and it still predicts them all as the same class. <\/p>\n\n<p>The input matrices are 10x10 of integers. <\/p>\n\n<p>Input reshaping:<\/p>\n\n<pre><code>x_train = x_train.reshape(len(x_train),10,10,1)\n<\/code><\/pre>\n\n<p>Actual model building:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(3, kernel_size=(1, 1), strides=(1, 1), activation='relu', input_shape=(10,10,1)))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(1, 1)))\nmodel.add(keras.layers.Flatten())\nmodel.add(keras.layers.Dense(500, activation='tanh'))\nmodel.add(Dropout(0.5))\nmodel.add(keras.layers.Dense(75, activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(keras.layers.Dense(10, activation='sigmoid'))\nmodel.add(keras.layers.Dense(1,kernel_initializer='normal',activation='sigmoid'))\noptimizerr = keras.optimizers.SGD(lr=0.001, momentum=0.9, decay=0.01, nesterov=True)\nmodel.compile(optimizer=optimizerr, loss='binary_crossentropy', metrics=[metrics.binary_accuracy])\n\nmodel.fit(x_train, y_train,epochs = 100, batch_size = 128, verbose=1)\n<\/code><\/pre>\n\n<p>I've tried modifying the learning rate, momentum, decay, the kernel_sizes, layer types, sizes... I checked for dying relu and that didn't seem to be the problem. Removing the dropout\/batch normalization layers (or various random layers) didn't do anything either.<\/p>\n\n<p>The data have roughly 53\/47% split across the labels, so it's not that either. <\/p>\n\n<p>I'm more confused because even when I ask it to predict the train set, it STILL insists on only labeling things one class, even if there are only ~20 samples or fewer.<\/p>\n","answers":[]}
+{"QuestionId":53105522,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2018-11-01T16:34:38.817","AcceptedAnswerId":null,"OwnerUserId":7380691.0,"Title":"Keras ImageDataGenerator messing up with grey images, how to correctly augment greyscale image?","Body":"<p>I am analysing 16bit grey scale images.\nAfter uploading greyscale image, everything looks fine.<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\nimg = plt.imread('example_image.tif')\nplt.imshow(img)\nplt.show()\n<\/code><\/pre>\n\n<h1>Or one can also do<\/h1>\n\n<pre><code>import cv2\nimg = cv2.imread('example_image.tif', -cv2.IMREAD_ANYDEPTH)\nimg1 = np.expand_dims(img, axis=2)\nimg2 = np.reshape(img1,[1,696,520,1])\nplt.imshow(img)\nplt.show()\n<\/code><\/pre>\n\n<p>When I checked my augmented images, they looks distorted. What am I doing wrong?<\/p>\n\n<pre><code>img1 = np.expand_dims(img, axis=2)\nimg2 = np.reshape(img1,[1,696,520,1])\nfrom keras.preprocessing.image import ImageDataGenerator\ndatagen = ImageDataGenerator(\n    rescale=1. \/ 4095,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)\n\ni=0\nfor batch in datagen.flow(img2, batch_size=1,\n                          save_to_dir='C:\/image_analysis', save_prefix='grey_check_', save_format='tif'):\n    i += 1\n    if i &gt; 20:\n        break\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YMV1u.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YMV1u.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/M2G68.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/M2G68.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53106244","CreationDate":"2018-11-01T17:21:12.607","ParentId":null,"OwnerUserId":"10418812","Title":null,"Body":"<p>replace this line:<\/p>\n\n<pre><code>img2 = np.reshape(img1,[1,696,520,1])\n<\/code><\/pre>\n\n<p>on this:<\/p>\n\n<pre><code>img2 = np.reshape(img1,[1,520,696,1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53105694,"AnswerCount":0,"Tags":"<tensorflow><virtual-machine><ram><rnn>","CreationDate":"2018-11-01T16:45:32.760","AcceptedAnswerId":null,"OwnerUserId":5647864.0,"Title":"Tensorflow memory consumption is different on VM","Body":"<p>I have an implementation of a RNN for sequence classification, on my local machine (Win 10, 16Gb ram) when I run the training, it reaches sometimes 100% memory usage. \nWhen I try to run it on a Azure VM (Linux Ubuntu, 14gb ram) the process get killed as soon as it reaches high ram usages.<\/p>\n\n<p>I am currently using a batch size of 5000 on the local machine, so I tried to reduce the size for the VM, but even if I put it at 2000, the process got killed anyway. <\/p>\n\n<p>Also on the vm it gives me warnings:\ntensorflow\/core\/framework\/allocator.cc:113] Allocation of 1152000000 exceeds 10% of system memory.<\/p>\n\n<p>What I don't understand is how the local machine can handle the memory usage, and the vm cannot even if they differ by only 2gb of ram.<\/p>\n\n<p>Does someone have any clue?<\/p>\n\n<p>Let me know if you need more information, thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":53105870,"AnswerCount":0,"Tags":"<tensorflow><lambda><enumerate>","CreationDate":"2018-11-01T16:56:13.883","AcceptedAnswerId":null,"OwnerUserId":10454184.0,"Title":"how to iterate over tensorflow tensor_list referencing both index and individual tensor","Body":"<p>I tried following code, used tf.map_fn, and still got error as: <em>\"Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.\"<\/em><\/p>\n\n<pre><code>def get_tensor_A(i):\n    return mask[i]   #another tensor returned\n\ndef get_tensor_B(x):\n    return tf.nn.softmax(x, 1)\n\ndef final_rt(i, inp):\n    return get_tensor_A(idx)*get_tensor_B(inp)\n\ntensor_2_list = [ tf.map_fn(lambda x: final_rt(x(0), x(1)), (idx, inp))  for idx, inp in enumerate(tensor_1_list)]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53106111,"AnswerCount":1,"Tags":"<machine-learning><keras><lstm>","CreationDate":"2018-11-01T17:12:17.057","AcceptedAnswerId":53106682.0,"OwnerUserId":1233523.0,"Title":"In Keras, when should I use input_shape instead of input_dim?","Body":"<p>I have found the use of input_shape instead of input_dim in Keras, especially in LSTM problems? My concern is that input_shape limits the number of rows in the input. It doesn't leave the scope to give complete Dataframe as input. When should we use input_shape instead of input_dim? <\/p>\n\n<p>Here are the examples <a href=\"https:\/\/machinelearningmastery.com\/timedistributed-layer-for-long-short-term-memory-networks-in-python\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/timedistributed-layer-for-long-short-term-memory-networks-in-python\/<\/a><\/p>\n","answers":[{"AnswerId":"53106682","CreationDate":"2018-11-01T17:49:16.410","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>To build on the comment and address the point of confusion. You <em>can<\/em> specify an unknown dimension using <code>None<\/code> to give varying values at runtime. For example, <code>input_shape=(None, 10)<\/code> means varying number of rows each with 10 entries. <code>input_dim<\/code> is just a short cut for specifying the final dimension and is there for convenience. <\/p>\n"}]}
+{"QuestionId":53106369,"AnswerCount":1,"Tags":"<python><pytorch><jupyter-lab>","CreationDate":"2018-11-01T17:29:10.243","AcceptedAnswerId":null,"OwnerUserId":5812695.0,"Title":"Error in training PyTorch classifier from the 60 minute blitz in GPU","Body":"<p>I've started to learn pytorch with their official 60 minute blitz tutorial in a jupyter lab (using their .ipynb file, <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py\" rel=\"nofollow noreferrer\">link to the tutorial<\/a>), and have completed it successfully until the conversion and training of the classifier using the gpu. I think that I have managed to change the device for the net, inputs and labels according to these results:<\/p>\n\n<pre><code>net=net.to(device)\nnet.fc1.weight.type()\n<\/code><\/pre>\n\n<p>With output: <\/p>\n\n<pre><code>'torch.cuda.FloatTensor'\n<\/code><\/pre>\n\n<p>And: <\/p>\n\n<pre><code>inputs, labels = inputs.to(device), labels.to(device)\ninputs.type(),labels.type()\n<\/code><\/pre>\n\n<p>With output:<\/p>\n\n<pre><code>('torch.cuda.FloatTensor', 'torch.cuda.LongTensor')\n<\/code><\/pre>\n\n<p>After running these cells, I ran the cell for training the model, containing this code:<\/p>\n\n<pre><code>for epoch in range(2):  # loop over the dataset multiple times\n\nrunning_loss = 0.0\nfor i, data in enumerate(trainloader, 0):\n    # get the inputs\n    inputs, labels = data\n\n    # zero the parameter gradients\n    optimizer.zero_grad()\n\n    # forward + backward + optimize\n    outputs = net(inputs)\n    loss = criterion(outputs, labels)\n    loss.backward()\n    optimizer.step()\n\n    # print statistics\n    running_loss += loss.item()\n    if i % 2000 == 1999:    # print every 2000 mini-batches\n        print('[%d, %5d] loss: %.3f' %\n              (epoch + 1, i + 1, running_loss \/ 2000))\n        running_loss = 0.0\n\nprint('Finished Training') \n<\/code><\/pre>\n\n<p>And received this error:<\/p>\n\n<pre><code>RuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-55-fe85c778b0e6&gt; in &lt;module&gt;()\n     10 \n     11         # forward + backward + optimize\n---&gt; 12         outputs = net(inputs)\n     13         loss = criterion(outputs, labels)\n     14         loss.backward()\n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, \n*input, **kwargs)\n    475             result = self._slow_forward(*input, **kwargs)\n    476         else:\n--&gt; 477             result = self.forward(*input, **kwargs)\n    478         for hook in self._forward_hooks.values():\n    479             hook_result = hook(self, input, result)\n\n&lt;ipython-input-52-725d44154459&gt; in forward(self, x)\n    14 \n    15     def forward(self, x):\n---&gt;16         x=self.conv1(x)\n    17         x = self.pool(F.relu(x))\n    18         x = self.pool(F.relu(self.conv2(x)))\n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, \n*input, **kwargs)\n    475             result = self._slow_forward(*input, **kwargs)\n    476         else:\n--&gt; 477             result = self.forward(*input, **kwargs)\n    478         for hook in self._forward_hooks.values():\n    479             hook_result = hook(self, input, result)\n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\modules\\conv.py in forward(self, \ninput)\n    299     def forward(self, input):\n    300         return F.conv2d(input, self.weight, self.bias, self.stride,\n--&gt; 301                         self.padding, self.dilation, self.groups)\n    302 \n    303 \n\nRuntimeError: Expected object of type torch.FloatTensor but found type \ntorch.cuda.FloatTensor for argument #2 'weight' \n<\/code><\/pre>\n\n<p>Why did I receive this error and how can I fix it? <\/p>\n","answers":[{"AnswerId":"53574334","CreationDate":"2018-12-01T19:29:59.863","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>You also need to move your <code>inputs<\/code> and <code>labels<\/code> to GPU inside the training loop.<\/p>\n\n<pre><code>for i, data in enumerate(trainloader, 0):\n    # get the inputs\n    inputs, labels = data\n\n    # move to GPU\n    inputs = inputs.to(device)\n    labels = labels.to(device)\n\n    ...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53106470,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><style-transfer>","CreationDate":"2018-11-01T17:35:08.970","AcceptedAnswerId":null,"OwnerUserId":9040678.0,"Title":"How to to feed keras pre-trained model to computational graph?","Body":"<p>lately I've been playing around with CNNs written in Keras, TF. Unfortunately I've come across one problem there:<\/p>\n\n<p>In these magnificent tutorials \n( <a href=\"https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/vgg16.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/vgg16.py<\/a> ; rest of the code is here),\n which I highly recommend, maestro loads pre-trained vgg16.tfmodel in very, very ugly way.<\/p>\n\n<pre><code>def __init__(self):\n    # Now load the model from file. The way TensorFlow\n    # does this is confusing and requires several steps.\n\n    # Create a new TensorFlow computational graph.\n    self.graph = tf.Graph()\n\n    # Set the new graph as the default.\n    with self.graph.as_default():\n\n        # TensorFlow graphs are saved to disk as so-called Protocol Buffers\n        # aka. proto-bufs which is a file-format that works on multiple\n        # platforms. In this case it is saved as a binary file.\n\n        # Open the graph-def file for binary reading.\n        path = os.path.join(data_dir, path_graph_def)\n        with tf.gfile.FastGFile(path, 'rb') as file:\n            # The graph-def is a saved copy of a TensorFlow graph.\n            # First we need to create an empty graph-def.\n            graph_def = tf.GraphDef()\n\n            # Then we load the proto-buf file into the graph-def.\n            graph_def.ParseFromString(file.read())\n\n            # Finally we import the graph-def to the default TensorFlow graph.\n            tf.import_graph_def(graph_def, name='')\n\n            # Now self.graph holds the VGG16 model from the proto-buf file.\n\n        # Get a reference to the tensor for inputting images to the graph.\n        self.input = self.graph.get_tensor_by_name(self.tensor_name_input_image)\n\n        # Get references to the tensors for the commonly used layers.\n        self.layer_tensors = [self.graph.get_tensor_by_name(name + \":0\") for name in self.layer_names]\n<\/code><\/pre>\n\n<p>And the problem is- I want my own pre-trained model load in the same\/similiar way, so I can put the model into the graph of the class I am calling later and if possible, get the last lines of the code here working.(Meaning getting the tensors of wanted layers from the graph.) <\/p>\n\n<p>All my tries based on <strong>load_model<\/strong> imported from keras and the comp. graph failed me. Also I didn't want to load it in a completely different way, because I would have to change A LOT of code afterwards- for a newbie kind of a big problem.<\/p>\n\n<p>Ok, I hope the question will reach the right person and also that it is not too trivial for you :D.<\/p>\n\n<p>BTW: The complex problem I'm solving, for you to make picture is the <strong>style transfer<\/strong> also in the same github repository. (<a href=\"https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/15_Style_Transfer.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/15_Style_Transfer.ipynb<\/a>)<\/p>\n","answers":[{"AnswerId":"53109476","CreationDate":"2018-11-01T21:17:34.207","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>So you want to load in a keras model into tensorflow basically? Can easily be done with the following code:<\/p>\n\n<pre><code>import keras.backend as k\nfrom keras.models import load_model\nimport tensorflow as tf\n\nmodel = load_model(\"your model.h5\")  # now it's in the memory of keras\nwith k.get_session() as sess:\n    # here you have a tensorflow computational graph, view it by:\n    tf.summary.FileWriter(\"folder name\", sess.graph)\n\n    # if you need a certain tensor do:\n    sess.graph.get_tensor_by_name(\"tensor name\") \n<\/code><\/pre>\n\n<p>To have a little read about the get_session function, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/get_session\" rel=\"nofollow noreferrer\">click here<\/a><\/p>\n\n<p>to view the graph, you need to load in the folder from the FileWriter with <a href=\"https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard#launching_tensorboard\" rel=\"nofollow noreferrer\">tensorboard like this<\/a>:<\/p>\n\n<pre><code>tensorboard --logdir path\/to\/folder\n<\/code><\/pre>\n\n<p>Hope this provided some help, good luck!<\/p>\n"}]}
+{"QuestionId":53106983,"AnswerCount":2,"Tags":"<python><tensorflow><scikit-learn>","CreationDate":"2018-11-01T18:10:31.420","AcceptedAnswerId":null,"OwnerUserId":10301016.0,"Title":"KerasClassifier TypeError: __call__() takes exactly 2 arguments (1 given) on cross_val_score","Body":"<p><strong>I try to cross_val_score on my model and I get the following error:<\/strong><\/p>\n\n<p><code>Traceback (most recent call last):\n  File \"\/home\/dinhnha1402\/.local\/lib\/python2.7\/site-packages\/keras\/wrappers\/scikit_learn.py\", line 210, in fit\n    return super(KerasClassifier, self).fit(x, y, **kwargs)\n  File \"\/home\/dinhnha1402\/.local\/lib\/python2.7\/site-packages\/keras\/wrappers\/scikit_learn.py\", line 139, in fit\n    **self.filter_sk_params(self.build_fn.__call__))\nTypeError: __call__() takes exactly 2 arguments (1 given)<\/code><\/p>\n\n<p>This is my model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(int(128), input_shape=(timesteps, int(128)),return_sequences=False))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(20, activation='relu', input_shape=(128,),kernel_initializer=initializers.glorot_uniform(seed=0)))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(20, activation='softmax'))\n\nmodel.add(Dense(2, activation='softmax')) \n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy'])\n\n### fit model\n\nmodel.fit(X_train, Y_train, batch_size=batch_size, epochs= epochs, verbose=1, validation_data=(X_test, Y_test))\n\n####Applying K-fold cross validation\nclassifier = KerasClassifier(build_fn=binary_classify_lstm_fc_model(), batch_size=10, epochs=100, verbose=0)\naccuracies = cross_val_score(estimator= classifier, X = X_train, y = Y_train, cv=10, scoring=\"accuracy\")#n_jobs= -1\nprint(accuracies)\n<\/code><\/pre>\n\n<p>I cannot find this error anywhere (on google). Does anyone have any ideas on how to resolve this issue?<\/p>\n","answers":[{"AnswerId":"58077166","CreationDate":"2019-09-24T09:29:17.547","ParentId":null,"OwnerUserId":"8451546","Title":null,"Body":"<p>Just change loss from categorical_crossentropy to mean_squared_error.<\/p>\n"},{"AnswerId":"53358722","CreationDate":"2018-11-18T07:21:53.343","ParentId":null,"OwnerUserId":"10669577","Title":null,"Body":"<p>just drop the \"()\"<\/p>\n\n<p>like that:<\/p>\n\n<p>classifier = KerasClassifier(build_fn=binary_classify_lstm_fc_model(), batch_size=10, epochs=100, verbose=0)<\/p>\n\n<p>=====><\/p>\n\n<p>classifier = KerasClassifier(build_fn=binary_classify_lstm_fc_model, batch_size=10, epochs=100, verbose=0)<\/p>\n"}]}
+{"QuestionId":53107087,"AnswerCount":2,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2018-11-01T18:17:36.400","AcceptedAnswerId":53111371.0,"OwnerUserId":10480432.0,"Title":"Import installed package pyCharm","Body":"<p>I have pyCharm and Python 3.7. When I go to project <code>interpreter &gt; packages<\/code> and try to download tensorflow, it outputs an error that tells me to do the command <code>pip install tensorflow<\/code>. I did that. Now, how do I import to pyCharm? <a href=\"https:\/\/i.stack.imgur.com\/lWFaH.png\" rel=\"nofollow noreferrer\">This is the error, but I already did pip update<\/a><\/p>\n","answers":[{"AnswerId":"53111371","CreationDate":"2018-11-02T00:58:06.647","ParentId":null,"OwnerUserId":"10474999","Title":null,"Body":"<p>Tensorflow is not availble for 3.7 version as of now.<\/p>\n\n<p>Downgrade the python version to 3.6.4<\/p>\n"},{"AnswerId":"53107182","CreationDate":"2018-11-01T18:24:00.283","ParentId":null,"OwnerUserId":"10562360","Title":null,"Body":"<p>In PyCharm:<\/p>\n\n<p>Option 1: If you have installed TensorFlow without GPU support, you don\u2019t need to do anything. Simply import TensorFlow to your script and you\u2019re done.<\/p>\n\n<p>Option 2: If you have installed TensorFlow with GPU support, then you would have also installed CUDA Toolkit 9.0 and cuDNN v7.0 (24.05.2018). You can add the appropriate directories to your %PATH% environment variable. However, even when doing so, PyCharm was complaining! Therefore, it\u2019s much easier to add these directories directly to your PyCharm project structure.<\/p>\n\n<p>File > Setting > Project Structure > Add Content Root > Add<\/p>\n\n<p>\u00b7 path_to_Cuda\\v9.0\\bin<\/p>\n\n<p>\u00b7 path_to_cuDNN\\v7.0\\bin<\/p>\n\n<p>Import TensorFlow to your script and you\u2019re done :)<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":53107126,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><neural-network><keras><deep-learning>","CreationDate":"2018-11-01T18:20:13.013","AcceptedAnswerId":53183860.0,"OwnerUserId":10193890.0,"Title":"What are the uses of TimeDistributed wrapper for LSTM or any other layers","Body":"<p>I am trying to understand the use of TimeDistributed layer in keras\/tensorflow. I have read some threads and articles but still I didn't get it properly.<\/p>\n\n<p>The threads that gave me some understanding of what the TImeDistributed layer does are -<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47305618\/what-is-the-role-of-timedistributed-layer-in-keras\">What is the role of TimeDistributed layer in Keras?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/44611006\/timedistributeddense-vs-dense-in-keras-same-number-of-parameters\">TimeDistributed(Dense) vs Dense in Keras - Same number of parameters<\/a><\/p>\n\n<p>But I still don't know why the layer is actually used!<\/p>\n\n<p>For example, both the below codes will provide same output (&amp; output_shape):<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(TimeDistributed(LSTM(5, input_shape = (10, 20), return_sequences = True)))\nprint(model.output_shape)\n\nmodel = Sequential()\nmodel.add(LSTM(5, input_shape = (10, 20), return_sequences = True))\nprint(model.output_shape)\n<\/code><\/pre>\n\n<p>And the output shape will be (according to my knowledge) -<\/p>\n\n<pre><code>(None, 10, 5)\n<\/code><\/pre>\n\n<p>So, if both the models provide same output, what is actually the use of TimeDistributed Layer?<\/p>\n\n<p>And I also had one other question. TimeDistributed layer applies time related data to separate layers (sharing same weights). So, how is it different from unrolling the LSTM layer which is provided in keras API as:<\/p>\n\n<blockquote>\n  <p>unroll: Boolean (default False). If True, the network will be unrolled, else a symbolic loop will be used. Unrolling can speed-up a RNN, although it tends to be more memory-intensive. Unrolling is only suitable for short sequences.<\/p>\n<\/blockquote>\n\n<p>What is the difference between these two?<\/p>\n\n<p>Thank you.. I am still a newbie and so have many questions.<\/p>\n","answers":[{"AnswerId":"53183860","CreationDate":"2018-11-07T05:08:50.887","ParentId":null,"OwnerUserId":"10161091","Title":null,"Body":"<p>As Keras documentation suggests TimeDistributed is a wrapper that applies a layer to every temporal slice of an input.<\/p>\n\n<p>Here is an example which might help:<\/p>\n\n<p>Let's say that you have video samples of cats and your task is a simple video classification problem, returning 0 if the cat is not moving or 1 if the cat is moving. Let's assume your input dim is (None, 50, 25, 25, 3) which means you have 50 time steps or frames per sample, and your frames are 25 by 25 and have 3 channels, rgb.<\/p>\n\n<p>Well, one aporoach would be to extract some \"features\" from each frame using CNN, like Conv2D, and then pass them to an LSTM layer. But the feature extraction would be the same for each frame. Now TimeDistributed comes to the rescue. You can wrap your Conv2D with it, then pass the output to a Flatten layer wrapped also by TimeDistributed. So after applying TimeDistributed(Conv2D(...)), the output would be something of dim like (None, 50, 5, 5, 16), and after TimeDistributed(Flatten()), the output would be of dim (None, 50, 400). (The actual dim would depend on Conv2D parameters.)<\/p>\n\n<p>The output at this layer now can be passes through LSTM. <\/p>\n\n<p>So obviously, LSTM itself does not need a TimeDistributed wrapper.<\/p>\n"}]}
+{"QuestionId":53107594,"AnswerCount":0,"Tags":"<python><tensorflow><matrix-multiplication><quantization>","CreationDate":"2018-11-01T18:53:16.130","AcceptedAnswerId":null,"OwnerUserId":9104473.0,"Title":"How do I import and use quantized_matmul and quantized_biasadd operations in tensorflow","Body":"<p>I am trying to use the quantized matmul operation in tensorflow. However I am not sure if I am doing it right. I could not find an example as to how to call the function and use it.<\/p>\n\n<p>I know that quantized conv2D can be called as tf.nn.conv2D(), just wanted to know a similar way of calling quantized_matmul operation as well.<\/p>\n\n<p>The class and the functions are present in tensorflow library and are documented below:<\/p>\n\n<p>Quantized Bias Add:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/quantized-bias-add\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/quantized-bias-add<\/a><\/p>\n\n<p>Quantized Matmul:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/quantized-mat-mul\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/quantized-mat-mul<\/a><\/p>\n\n<p>Thanks and Regards,\nAbhinav George<\/p>\n","answers":[]}
+{"QuestionId":53107605,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-11-01T18:53:51.867","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Tensorflow: ValueError: Cannot create a tensor proto whose content is larger than 2GB","Body":"<p>I'm using the <code>tf.data.Dataset<\/code> API and have a dataset that's ~500K rows and ~1,800 columns.  When I try:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({\"reviews\": data}, labels))\n<\/code><\/pre>\n\n<p>I get back:<\/p>\n\n<pre><code>ValueError: Cannot create a tensor proto whose content is larger than 2GB.\n<\/code><\/pre>\n\n<p>I've googled around and seen a lot of people run into this issue but no satisfactory answers.  Is there a way to get around this limit or a tensorflow approach that will break up my <code>dataset<\/code>?  I already batch it but that happens after calling:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(({\"reviews\": data}, labels))\n<\/code><\/pre>\n\n<p>For what it's worth, my code to read the data from CSV into a <code>tf.data.Dataset<\/code> works when I use 10% of the data.<\/p>\n\n<p>Any suggestions would be awesome!<\/p>\n","answers":[{"AnswerId":"53133352","CreationDate":"2018-11-03T16:35:34.163","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>Depending on your dataset you could try using the tf <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\" rel=\"nofollow noreferrer\">dataset api<\/a>This means you could try converting your dataset into a tfrecord or you could try <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/reading_data\" rel=\"nofollow noreferrer\">csv<\/a> files. The datasetapi takes care of loading your data in the background while you are training other data. This speeds up the training process significantly as well <\/p>\n"}]}
+{"QuestionId":53107763,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-11-01T19:04:14.873","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras: Use categorical_crossentropy without one-hot encoded array of targets","Body":"<p>I have a Keras model that I'm using for a multi-class classification problem.  I'm doing this:<\/p>\n\n<pre><code>model.compile(\n    loss='categorical_crossentropy',\n    optimizer='adam',\n    metrics=['accuracy'],\n)\n<\/code><\/pre>\n\n<p>I currently have ~100 features and there are ~2000 possible classes.  One-hot encoding the class is leading to memory issues.<\/p>\n\n<p>Is it possible to use <code>categorical_crossentropy<\/code> with this Keras model while <em>not<\/em> one-hot encoding the class labels.  E.g. instead of having a target look like:<\/p>\n\n<p><code>[0, 0, 0, 1, 0, 0, ...]<\/code><\/p>\n\n<p>It would just be:<\/p>\n\n<p><code>3<\/code><\/p>\n\n<p>I looked at the source for <code>categorical_crossentropy<\/code> in Keras and it assumes two tensors of the same shape.  Is there a way to get around this and use the approach I described?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53108290","CreationDate":"2018-11-01T19:40:07.393","ParentId":null,"OwnerUserId":"9041710","Title":null,"Body":"<p>Could you post the rest of your code? by my understanding when using categorical crossentropy as loss function, the last layer should use a softmax activation function, yielding for each output neuron the probability of the input corresponding to said neuron's class, and not directly the one-hot vector. Then the categorical crossentropy is calculated as<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LF3Re.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LF3Re.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where the <code>p<\/code>'s are these probabilities. By just outputting the class you wouldn't have access to these probabilities and thus wouldn't be able to compute the categorical crossentropy.<\/p>\n"},{"AnswerId":"53109180","CreationDate":"2018-11-01T20:52:21.283","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>If your targets are one-hot encoded, use <code>categorical_crossentropy<\/code>.\nExamples of one-hot encodings:<\/p>\n\n<pre><code>[1,0,0]\n[0,1,0]\n[0,0,1]\n<\/code><\/pre>\n\n<p>However, if your targets are integers, use <code>sparse_categorical_crossentropy<\/code>.\nExamples of integer encodings:<\/p>\n\n<pre><code>1\n2\n3\n<\/code><\/pre>\n"}]}
+{"QuestionId":53107867,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-11-01T19:11:00.343","AcceptedAnswerId":null,"OwnerUserId":9302655.0,"Title":"LSTM keras MaxMinScale function producing different values for the same number","Body":"<p>I am trying to apply LSTM model to do timeseries forecasting.<\/p>\n\n<p>One step in LSTM is to scale the data using MinMaxScaler. I'm trying to scale the data in the range (-1,1).<\/p>\n\n<p>My train data looks something like below:<\/p>\n\n<pre><code>[[0.         2.54583333]\n [2.54583333 2.6125    ]\n [2.6125     2.875     ]\n ...\n [2.69583333 2.5125    ]\n [2.5125     2.91666667]\n [2.91666667 3.4375    ]]\n\nMinMaxScaler(copy=True, feature_range=(-1, 1))\n[[-1.         -0.95342466]\n [-0.80649248 -0.94794521]\n [-0.80142518 -0.92636986]\n ...\n [-0.79509105 -0.95616438]\n [-0.80902613 -0.92294521]\n [-0.77830562 -0.88013699]]\n<\/code><\/pre>\n\n<p>As, it can be seen above the values are not getting scaled properly. For eg: 2.54583333 is getting scaled to -0.95342466 as well as -0.80649248.<\/p>\n\n<p>Actually the scaled values should look something like below:<\/p>\n\n<pre><code>[[-1.         -0.95342466]\n [-0.95342466 -0.94794521]\n [0.94794521 -0.92636986]\n ...\n [-0.79509105 -0.95616438]\n [-0.95616438 -0.92294521]\n [-0.92294521 -0.88013699]]\n<\/code><\/pre>\n\n<p>Scaled values should also follow the pattern in the training set.<\/p>\n\n<p>Can someone tell me what i am doing wrong, why multiple scaled values are coming for the same number?<\/p>\n\n<p>I'm scaling values using below code:<\/p>\n\n<pre><code>def scale(train, test):\n    # fit scaler\n    scaler = MinMaxScaler(feature_range=(-1, 1))\n    scaler = scaler.fit(train)\n\n    # transform train\n    train = train.reshape(train.shape[0], train.shape[1])\n    train_scaled = scaler.transform(train)\n\n    # transform test\n    test = test.reshape(test.shape[0], test.shape[1])\n    test_scaled = scaler.transform(test)\n    return scaler, train_scaled, test_scaled\n<\/code><\/pre>\n\n<p>But something is really wrong here. Please help.<\/p>\n\n<p>My whole dataset is of shape (720,1)<\/p>\n\n<p>Length of train data : 519\nLength of test data : 200.<\/p>\n\n<p>In the above method the shape of train data is (519,2) and test data is (200,2). Please let me know if more information or any other clarity required. I am new to this and I'm trying LSTM for the first time.<\/p>\n","answers":[{"AnswerId":"53109330","CreationDate":"2018-11-01T21:03:49.620","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>There is nothing wrong with this. Each feature is scaled independently, and each has its own normalization factors, so the same feature value in a different feature dimension will not be normalized in the same way.<\/p>\n"}]}
+{"QuestionId":53108391,"AnswerCount":1,"Tags":"<python><keras><embedding><keras-layer>","CreationDate":"2018-11-01T19:48:56.750","AcceptedAnswerId":null,"OwnerUserId":3272089.0,"Title":"Keras - Embedding Layer and GRU Layer Shape Error","Body":"<pre><code># input_shape = (137861, 21, 1)\n# output_sequence_length = 21\n# english_vocab_size = 199\n# french_vocab_size = 344\n\ndef embed_model(input_shape, output_sequence_length, english_vocab_size, french_vocab_size):\n    '''\n    Build and train a RNN model using word embedding on x and y\n    :param input_shape: Tuple of input shape\n    :param output_sequence_length: Length of output sequence\n    :param english_vocab_size: Number of unique English words in the dataset\n    :param french_vocab_size: Number of unique French words in the dataset\n    :return: Keras model built, but not trained\n    '''\n\n    learning_rate = 1e-3\n    model = Sequential()\n\n    model.add(Embedding(english_vocab_size, 128, input_length=output_sequence_length, input_shape=input_shape[1:]))\n\n    model.add(GRU(units=128, return_sequences=True))\n    model.add(TimeDistributed(Dense(french_vocab_size)))\n    model.add(Activation('softmax'))\n\n    model.summary()\n\n    model.compile(loss=sparse_categorical_crossentropy,\n                  optimizer=Adam(learning_rate),\n                  metrics=['accuracy'])\n\n    return model\n<\/code><\/pre>\n\n<p>When invoking this method to train a model, it gets the error: <\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer gru_1: expected ndim=3, found ndim=4\n<\/code><\/pre>\n\n<p>How to fix the shape error between Embedding Layer and GRU Layer?<\/p>\n","answers":[{"AnswerId":"53118199","CreationDate":"2018-11-02T11:59:27.757","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that the Embedding layer takes a 2D array as the input. However, the shape of the input array is <code>(137861, 21, 1)<\/code> which makes it a 3D array. Simply remove the last axis using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.1\/reference\/generated\/numpy.squeeze.html\" rel=\"nofollow noreferrer\"><code>squeeze()<\/code><\/a> method from numpy:<\/p>\n\n<pre><code>data = np.squeeze(data, axis=-1)\n<\/code><\/pre>\n\n<hr>\n\n<p>As a side, there is no need to use <code>TimeDistributed<\/code> layer here, since the <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">Dense layer is applied on the last axis by defualt<\/a>.<\/p>\n"}]}
+{"QuestionId":53108547,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><flask><keras>","CreationDate":"2018-11-01T20:01:36.997","AcceptedAnswerId":null,"OwnerUserId":10581436.0,"Title":"ValueError while predicting on saved Machine learning model in flask rest api","Body":"<p>I have trained and saved a model using <code>model.save('filename.h5')<\/code>. It works to make predictions, but when I load the same model into the flask_app and try to make a prediction on it I get a ValueError:<\/p>\n\n<p><strong>Trace for that:<\/strong><\/p>\n\n<pre><code>The data type of the acquired data from the post call: &lt;class 'str'&gt;\n\nPreprocessing\nPre-processing complete\nData type of the input data: &lt;class 'numpy.ndarray'&gt;\nPredicting for requirement:\n[2018-11-01 15:34:56,261] ERROR in app: Exception on \/predict [POST]\nTraceback (most recent call last):\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\app.py\", line 2292, in wsgi_app\n    response = self.full_dispatch_request()\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\app.py\", line 1815, in full_dispatch_request\n    rv = self.handle_user_exception(e)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask_cors\\extension.py\", line 161, in wrapped_function\n    return cors_after_request(app.make_response(f(*args, **kwargs)))\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\app.py\", line 1718, in handle_user_exception\n    reraise(exc_type, exc_value, tb)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\_compat.py\", line 35, in reraise\n    raise value\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\app.py\", line 1813, in full_dispatch_request\n    rv = self.dispatch_request()\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\flask\\app.py\", line 1799, in dispatch_request\n    return self.view_functions[rule.endpoint](**req.view_args)\n  File \"C:\\Users\\singroha\\PycharmProjects\\smart_classification_ML_model\\flask_app\\predict_app.py\", line 99, in predict\n    prediction = requirement_model.predict(r_predict_text)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 1164, in predict\n    self._make_predict_function()\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 554, in _make_predict_function\n    **kwargs)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 2744, in function\n    return Function(inputs, outputs, updates=updates, **kwargs)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 2546, in __init__\n    with tf.control_dependencies(self.outputs):\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 5002, in control_dependencies\n    return get_default_graph().control_dependencies(control_inputs)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 4541, in control_dependencies\n    c = self.as_graph_element(c)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3488, in as_graph_element\n    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3567, in _as_graph_element_locked\n    raise ValueError(\"Tensor %s is not an element of this graph.\" % obj)\nValueError: Tensor Tensor(\"dense_7\/Softmax:0\", shape=(?, 2), dtype=float32) is not an element of this graph.\n<\/code><\/pre>\n\n<p><strong>The flow of things:<\/strong><\/p>\n\n<ul>\n<li>I run the flask server that loads the saved models<\/li>\n<li>When it receives a request from the client predict method gets called<\/li>\n<li>I preprocess the data to get it ready to put into the model<\/li>\n<li>Then I try to predict and this happens<\/li>\n<\/ul>\n\n<p><strong>What I thought it was and what I did:<\/strong><\/p>\n\n<ul>\n<li>I thought it just be that I'm not preprocessing the data correctly<\/li>\n<li>So I pickled the preprocessed data and went back to training the model and used this pickled object to make a prediction and it works. <\/li>\n<\/ul>\n\n<p>I tried playing around with _make_predict_function and try some things that I found on the internet, but didn't seem to work for me. Any insight about this is much appreciated. Thank you<\/p>\n\n<p><strong>Here is the predict function for the flask_app:<\/strong><\/p>\n\n<pre><code>@app.route(\"\/predict\", methods=[\"POST\"])\n# @cross_origin(origin='localhost',headers=['Content- Type','Authorization'])\ndef predict():\n    post_message = request.get_json(force=True)\n\n    # Logging\n    print('Data message received from the client', post_message)\n    print(\"\\nThe data type of the acquired data from the post call:\", type(post_message))\n\n    requirement = []\n    requirement.append(post_message)\n\n    # Preprocessing data\n    r_predict_text, (sent_text, s_predict_text) = preprocess_data(requirement)\n\n    # print(\"\\n\\n printing preprocessed data:\")\n    # print(r_predict_text)\n\n    # Saving the preprocessing data for varification\n    # with open('text_req.pickle', 'wb') as pickle_file:\n    #     pickle.dump(r_predict_text, pickle_file, protocol=pickle.HIGHEST_PROTOCOL)\n\n    print('Data type of the input data:', type(r_predict_text))\n    print('Predicting for requirement:')\n\n    # Everything goes down at this point after receiving request\n    prediction = requirement_model.predict(r_predict_text) \n\n    # Ignore this, for now, I was trying to see if the response gets send back if take out the predict function call\n    prediction = 'SMART'\n\n    response = {\n        'prediction': {\n            'SMART': prediction\n        }\n    }\n    print('sending response to the client')\n    return jsonify(response)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53108595,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><lstm>","CreationDate":"2018-11-01T20:05:03.167","AcceptedAnswerId":53109739.0,"OwnerUserId":10231366.0,"Title":"Using LSTM layers in combination with CNN","Body":"<p>I am trying to improve the accuracy of an EEG classifier. Currently I was only classifying with conv layers and fully connected ones. In the literature I found people using LSTM layers in their CNN model. I would like to try this out but I get following error.<\/p>\n\n<p>The data that I'm using is time series EEG data 64 channels x 325 samples (sampled at 500 Hz so 650 ms). X = (1923,63,325,1) and y = (1923,)<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(TimeDistributed(Conv2D(64, (1, 3), input_shape=X.shape[1:])))\nmodel.add(Activation('relu'))\n\nmodel.add(TimeDistributed(Conv2D(64, (1, 3))))\nmodel.add(Activation('relu')) \nmodel.add(MaxPooling2D(pool_size=(1, 2)))\n\nmodel.add(TimeDistributed(Conv2D(64, (1, 3))))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(1, 2)))\n\nmodel.add(TimeDistributed(Flatten()))\n\nmodel.add(LSTM(128,return_sequences=True))\n\nmodel.add(LSTM(128,return_sequences=True))\n\nmodel.add(Dense(16))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.3)) \n\nmodel.add(Dense(2))\nmodel.add(Activation('softmax'))\n\nmodel.compile(loss='sparse_categorical_crossentropy',\n                  optimizer='adam',\n                  metrics=['accuracy'])\nreturn model\n<\/code><\/pre>\n\n<p>The error that I'm getting:<\/p>\n\n<pre><code>IndexError: list index out of range\n<\/code><\/pre>\n\n<p>I have seen some other questions about this online but most of them do not really apply on my application I think.<\/p>\n\n<p>Edit: error gets thrown<\/p>\n\n<pre><code>  File \"&lt;ipython-input-8-0b3d7307ea53&gt;\", line 1, in &lt;module&gt;\n    runfile('D:\/ AA TestPrograms\/LALALAL\/ModelV20.py', wdir='D:\/ AA TestPrograms\/LALALAL')\n\n  File \"d:\\anaconda\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 668, in runfile\n    execfile(filename, namespace)\n\n  File \"d:\\anaconda\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 108, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"D:\/ AA TestPrograms\/LALALAL\/ModelV20.py\", line 115, in &lt;module&gt;\n    history = train_model(model, xtrain, ytrain, xval, yval)\n\n  File \"D:\/ AA TestPrograms\/LALALAL\/ModelV20.py\", line 94, in train_model\n    shuffle=True\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\engine\\training.py\", line 950, in fit\n    batch_size=batch_size)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\engine\\training.py\", line 671, in _standardize_user_data\n    self._set_inputs(x)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\engine\\training.py\", line 577, in _set_inputs\n    self.build(input_shape=(None,) + inputs.shape[1:])\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\engine\\sequential.py\", line 225, in build\n    x = layer(x)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 457, in __call__\n    output = self.call(inputs, **kwargs)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\layers\\wrappers.py\", line 248, in call\n    y = self.layer.call(inputs, **kwargs)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\layers\\convolutional.py\", line 168, in call\n    dilation_rate=self.dilation_rate)\n\n  File \"d:\\anaconda\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 3565, in conv2d\n    data_format=tf_data_format)\n\n  File \"d:\\anaconda\\lib\\site-packages\\tensorflow\\python\\ops\\nn_ops.py\", line 779, in convolution\n    data_format=data_format)\n\n  File \"d:\\anaconda\\lib\\site-packages\\tensorflow\\python\\ops\\nn_ops.py\", line 828, in __init__\n    input_channels_dim = input_shape[num_spatial_dims + 1]\n\n  File \"d:\\anaconda\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 615, in __getitem__\n    return self._dims[key]\n\nIndexError: list index out of range\n<\/code><\/pre>\n\n<p>I am a beginner when it comes to Python and Tensorflow but I think it is the last line where it checks the dimensions? If I look in tensor_shape.py I can't find an IndexError.<\/p>\n\n<p>Thanks for the help!<\/p>\n","answers":[{"AnswerId":"53109739","CreationDate":"2018-11-01T21:40:11.320","ParentId":null,"OwnerUserId":"5245893","Title":null,"Body":"<p>I tried to reproduce it and got this error after I replaced \n<code>model.add(TimeDistributed(Conv2D(64, (1, 3), input_shape=X.shape[1:])))<\/code> \nwith this\n<code>model.add(TimeDistributed(Conv2D(64, (1, 3)), input_shape=X.shape[1:]))<\/code> with X being a random array of shape (1923,63,325,1).\nThe error is raised because the dimension of the array must equal 4 in order to apply Conv2D sequentially for each sample (or without the dimension of a batch). \nIn other words, X must have shape (batch dimension, temporal dimension, row dimension, col dimension, channel dimension).<\/p>\n"}]}
+{"QuestionId":53109028,"AnswerCount":1,"Tags":"<tensorflow><autoencoder>","CreationDate":"2018-11-01T20:39:07.560","AcceptedAnswerId":53116014.0,"OwnerUserId":9203344.0,"Title":"AutoEncoder shape","Body":"<p>I try to create a autoencoder in Tensorflow without using contriib.\nHere is the original code<\/p>\n\n<p><a href=\"https:\/\/github.com\/Machinelearninguru\/Deep_Learning\/blob\/master\/TensorFlow\/neural_networks\/autoencoder\/simple_autoencoder.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Machinelearninguru\/Deep_Learning\/blob\/master\/TensorFlow\/neural_networks\/autoencoder\/simple_autoencoder.py<\/a><\/p>\n\n<p>Here is the program I modify:<\/p>\n\n<pre><code>    import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n\n\n\nae_inputs = tf.placeholder(tf.float32, (None, 32, 32, 1))  # input to the network (MNIST images)\n\n\nxi = tf.nn.conv2d(ae_inputs, \n                 filter=tf.Variable(tf.random_normal([5,5,1,32])), \n                 strides=[1,2,2,1],\n                 padding='SAME')\nprint(\"xi {0}\".format(xi))\n\nxi = tf.nn.conv2d(xi, \n                 filter=tf.Variable(tf.random_normal([5,5,32,16])), \n                 strides=[1,2,2,32],\n                 padding='SAME')\nprint(\"xi {0}\".format(xi))\n\nxi = tf.nn.conv2d(xi, \n                 filter=tf.Variable(tf.random_normal([5,5,16,8])), \n                 strides=[1,4,4,16],\n                 padding='SAME')\nprint(\"xi {0}\".format(xi))\n\n\n\n\n\n\nxo = tf.nn.conv2d_transpose(xi, \n                 filter=tf.Variable(tf.random_normal([5,5,16,8])), \n                 output_shape=[1, 8, 8, 16],\n                 strides=[1,4,4,1],\n                 padding='SAME')\nprint(\"xo {0}\".format(xo))\n\nxo = tf.nn.conv2d_transpose(xo, \n                 filter=tf.Variable(tf.random_normal([5,5,32,16])), \n                 output_shape=[1, 16, 16, 32],\n                 strides=[1,2,2,1],\n                 padding='SAME')\nprint(\"xo {0}\".format(xo))\n\nxo = tf.nn.conv2d_transpose(xo, \n                 filter=tf.Variable(tf.random_normal([5,5,1,32])), \n                 output_shape=[1, 32, 32, 1],\n                 strides=[1,2,2,1],\n                 padding='SAME')\n\nprint(\"xo {0}\".format(xo))\n<\/code><\/pre>\n\n<p>And the result from the print  is that:<\/p>\n\n<blockquote>\n  <p>xi Tensor(\"Conv2D:0\", shape=(?, 16, 16, 32), dtype=float32) \n  xi Tensor(\"Conv2D_1:0\", shape=(?, 8, 8, 16), dtype=float32) \n  xi Tensor(\"Conv2D_2:0\", shape=(?, 2, 2, 8), dtype=float32)\n  xo Tensor(\"conv2d_transpose:0\", shape=(1, 8, 8, 16), dtype=float32) \n  xo Tensor(\"conv2d_transpose_1:0\", shape=(1, 16, 16, 32), dtype=float32)\n  xo Tensor(\"conv2d_transpose_2:0\", shape=(1, 32, 32, 1), dtype=float32)<\/p>\n<\/blockquote>\n\n<p>It seems the output has the good shape but I don't really sure about all the parameters in the conv2 and conv2_transpose.<\/p>\n\n<p>Can someone correct my code if needed<\/p>\n\n<p>edit: \n@Lau I add the relu function as you tel me but I don't known where to add the bias:<\/p>\n\n<pre><code>xi = tf.nn.conv2d(ae_inputs,\n             filter=tf.Variable(tf.random_normal([5,5,1,32])),\n             strides=[1,2,2,1],\n             padding='SAME')\nxi = tf.nn.relu(xi)\n# xi = max_pool(xi,2)\nprint(\"xi {0}\".format(xi))\n\nxi = tf.nn.conv2d(xi,\n                 filter=tf.Variable(tf.random_normal([5,5,32,16])),\n                 strides=[1,2,2,1],\n                 padding='SAME')\nxi = tf.nn.relu(xi)\n# xi = max_pool(xi,2)\nprint(\"xi {0}\".format(xi))\n\nxi = tf.nn.conv2d(xi,\n                 filter=tf.Variable(tf.random_normal([5,5,16,8])),\n                 strides=[1,4,4,1],\n                 padding='SAME')\nxi = tf.nn.relu(xi)\n# xi = max_pool(xi,4)\nprint(\"xi {0}\".format(xi))\n\n\n\n\n\n\nxo = tf.nn.conv2d_transpose(xi,\n                 filter=tf.Variable(tf.random_normal([5,5,16,8])),\n                 output_shape=[tf.shape(xi)[0], 8, 8, 16],\n                 strides=[1,4,4,1],\n                 padding='SAME')\nxo = tf.nn.relu(xo)\n\nprint(\"xo {0}\".format(xo))\n\nxo = tf.nn.conv2d_transpose(xo,\n                 filter=tf.Variable(tf.random_normal([5,5,32,16])),\n                 output_shape=[tf.shape(xo)[0], 16, 16, 32],\n                 strides=[1,2,2,1],\n                 padding='SAME')\nxo = tf.nn.relu(xo)\n\nprint(\"xo {0}\".format(xo))\n\nxo = tf.nn.conv2d_transpose(xo,\n                 filter=tf.Variable(tf.random_normal([5,5,1,32])),\n                 output_shape=[tf.shape(xo)[0], 32, 32, 1],\n                 strides=[1,2,2,1],\n                 padding='SAME')\nxo = tf.nn.tanh(xo)\nprint(\"xo {0}\".format(xo))\nreturn xo\n<\/code><\/pre>\n\n<p>I don't understand what is the difference with the original code:<\/p>\n\n<pre><code># encoder\n# 32 x 32 x 1   -&gt;  16 x 16 x 32\n# 16 x 16 x 32  -&gt;  8 x 8 x 16\n# 8 x 8 x 16    -&gt;  2 x 2 x 8\nprint('inputs {0}'.format(inputs))\n\nnet = lays.conv2d(inputs, 32, [5, 5], stride=2, padding='SAME')\nprint('net {0}'.format(net))\n\nnet = lays.conv2d(net, 16, [5, 5], stride=2, padding='SAME')\nprint('net {0}'.format(net))\n\nnet = lays.conv2d(net, 8, [5, 5], stride=4, padding='SAME')\nprint('net {0}'.format(net))\n\n# decoder\n# 2 x 2 x 8    -&gt;  8 x 8 x 16\n# 8 x 8 x 16   -&gt;  16 x 16 x 32\n# 16 x 16 x 32  -&gt;  32 x 32 x 1\nnet = lays.conv2d_transpose(net, 16, [5, 5], stride=4, padding='SAME')\nprint('net {0}'.format(net))\n\nnet = lays.conv2d_transpose(net, 32, [5, 5], stride=2, padding='SAME')\nprint('net {0}'.format(net))\n\nnet = lays.conv2d_transpose(net, 1, [5, 5], stride=2, padding='SAME', activation_fn=tf.nn.tanh)\n\nprint('net {0}'.format(net))\nreturn net\n<\/code><\/pre>\n\n<p>Edit2 :<\/p>\n\n<p>@Lau I make the new version of the autoencoder with your modifications:<\/p>\n\n<pre><code>mean = 0\n    stdvev = 0.1\n    with tf.name_scope('L0'):\n        xi = tf.nn.conv2d(ae_inputs,\n                     filter=tf.truncated_normal([5,5,1,32], mean = mean, stddev=stdvev),\n                     strides=[1,1,1,1],\n                     padding='SAME')\n        xi =  tf.nn.bias_add(xi, bias_variable([32]))\n        xi = max_pool(xi,2)\n        print(\"xi {0}\".format(xi))\n\n    with tf.name_scope('L1'):\n        xi = tf.nn.conv2d(xi,\n                         filter=tf.truncated_normal([5,5,32,16], mean = mean, stddev=stdvev),\n                         strides=[1,1,1,1],\n                         padding='SAME')\n        xi =  tf.nn.bias_add(xi, bias_variable([16]))\n        xi = max_pool(xi,2)\n        print(\"xi {0}\".format(xi))\n\n    with tf.name_scope('L2'):\n        xi = tf.nn.conv2d(xi,\n                         filter=tf.truncated_normal([5,5,16,8], mean = mean, stddev=stdvev),\n                         strides=[1,1,1,1],\n                         padding='SAME')\n        xi =  tf.nn.bias_add(xi, bias_variable([8]))\n        xi = max_pool(xi,4)\n        print(\"xi {0}\".format(xi))\n\n\n    with tf.name_scope('L3'):\n        xo = tf.nn.conv2d_transpose(xi,\n                         filter=tf.truncated_normal([5,5,16,8], mean = mean, stddev=stdvev),\n                         output_shape=[tf.shape(xi)[0], 8, 8, 16],\n                         strides=[1,4,4,1],\n                         padding='SAME')\n        xo =  tf.nn.bias_add(xo, bias_variable([16]))\n        print(\"xo {0}\".format(xo))\n\n    with tf.name_scope('L4'):\n        xo = tf.nn.conv2d_transpose(xo,\n                         filter=tf.truncated_normal([5,5,32,16], mean = mean, stddev=stdvev),\n                         output_shape=[tf.shape(xo)[0], 16, 16, 32],\n                         strides=[1,2,2,1],\n                         padding='SAME')\n        xo =  tf.nn.bias_add(xo, bias_variable([32]))\n        print(\"xo {0}\".format(xo))\n\n    with tf.name_scope('L5'):\n        xo = tf.nn.conv2d_transpose(xo,\n                         filter=tf.truncated_normal([5,5,1,32], mean = mean, stddev=stdvev),\n                         output_shape=[tf.shape(xo)[0], 32, 32, 1],\n                         strides=[1,2,2,1],\n                         padding='SAME')\n        xo =  tf.nn.bias_add(xo, bias_variable([1]))\n        xo = tf.nn.tanh(xo)\n        print(\"xo {0}\".format(xo))\n<\/code><\/pre>\n\n<p>But the result is the same, the decoded value are not the same. <\/p>\n\n<p>Edit3:<\/p>\n\n<p>I change thefilter definition from<\/p>\n\n<pre><code>filter=tf.truncated_normal([5,5,16,8], mean = mean, stddev=stdvev),\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code> filter= tf.get_variable('filter2',[5,5,16,8]),\n<\/code><\/pre>\n\n<p>The result seems to converge to better result but still converge to a different value. In the original code (0.006) and my version 0.015. I think it comes from the initialize value of the filter and the bias. How can I manage that?<\/p>\n","answers":[{"AnswerId":"53116014","CreationDate":"2018-11-02T09:37:49.370","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>You forgot a bias and an activation. So your network is weaker than a PCA. I suggest you use <code>tf.layers<\/code>instead. If you want to use <code>tf.nn<\/code>, then please use <code>tf.get_variable<\/code>.\nFurthermore you have to add:\n<code>tf.nn.bias_add<\/code> \n<code>tf.nn.relu<\/code> (or any other activation)<\/p>\n\n<p>If you want to know if the code works just test it with:<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.tf.global_variables_initializer())\ntest_output = sess.run(xo, feed_dict={ae_inputs : np.random.random((1, 32, 32, 1))}\nprint(test_output)\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong>\nOk, so the code you posted uses basically the <code>tf.layers<\/code> API, where bias and activation is included. The <code>tf.nn<\/code> API is more basic and only applies convolution, but without activation or bias. <\/p>\n\n<p>Based on your edit I think you want to implement the the CAE in the nn API. A typical encoder layer would be this:<\/p>\n\n<pre><code>conv = tf.nn.conv2d(\n                     nput=input_tensor,\n                     filter=tf.get_variable(\"conv_weight_name\", shape=[height,\n                                                                width,\n                                                                number_input_feature_maps,\n                                                                number_output_feature_maps]),\n                     strides=[1, 1, 1, 1],\n                     padding=\"SAME\")\nbias = tf.nn.bias_add(conv, tf.get_variable(\"name_bias\",\n                                            [number_output_feature_maps]))\nlayer_out = tf.nn.relu(bias)\n<\/code><\/pre>\n\n<p>Here is a typical layer for transpose convolution.<\/p>\n\n<pre><code>conv_transpose = tf.nn.conv2d_transpose(value=input_tensor,\n                       filter=tf.get_variable(\"deconnv_weight_name\", shape=[height,\n                                                                     width,\n                                                                     number_output_feature_maps,\n                                                                     number_input_feature_maps]),\n                       output_shape=[batc_size, height_output, width_ouput, feature_maps_output],\n                       strides=[1, 1, 1, 1])\nbias = tf.nn.bias_add(conv_transpose, tf.get_variable(\"name_bias\", shape=[number_output_feature_maps]))\n\nlayer_out = tf.nn.relu(bias)\n           `\n<\/code><\/pre>\n\n<p>If you have questions about the names, just ask in the commnets.<\/p>\n"}]}
+{"QuestionId":53109580,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2018-11-01T21:25:46.013","AcceptedAnswerId":null,"OwnerUserId":8507381.0,"Title":"Encoding issues with cloud ml","Body":"<p>I'm running the code made in this <a href=\"https:\/\/cloud.google.com\/blog\/products\/gcp\/performing-prediction-with-tensorflow-object-detection-models-on-google-cloud-machine-learning-engine\" rel=\"nofollow noreferrer\">tutorial<\/a> and when I'm trying to make an online prediction with a JSON that contains my picture in b64 encoding I get the message that it Expected uint8 and got a hexadecimal string instead. \nI already checked the JSON and it's format is ok. What can be the problem?<\/p>\n\n<p>I'm using google's CLI for the prediction, and my model's version is tensorflow 1.10 as well as my runtime version. I used the tf API for object detection with faster-rcnn<\/p>\n\n<p>Code to conver image into b64 and into JSON and tfrecord<\/p>\n\n<pre><code>import base64\nimport io\nimport json\nfrom PIL import Image\nimport tensorflow as tf\nwidth = 1024\nheight = 768\npredict_instance_json = \"inputs.json\"\npredict_instance_tfr = \"inputs.tfr\"\nwith tf.python_io.TFRecordWriter(predict_instance_tfr) as tfr_writer:\n  with open(predict_instance_json, \"wb\") as fp:\n    for image in [\"image1.jpg\", \"image2.jpg\"]:\n      img = Image.open(image)\n      img = img.resize((width, height), Image.ANTIALIAS)\n      output_str = io.BytesIO()\n      img.save(output_str, \"JPEG\")\n      fp.write(\n          json.dumps({\"b64\": base64.b64encode(output_str.getvalue())}) + \"\\n\")\n      tfr_writer.write(output_str.getvalue())\n      output_str.close()\n<\/code><\/pre>\n\n<p>Command for prediction:<\/p>\n\n<pre><code>gcloud ml-engine predict --model=${YOUR_MODEL} --version=${YOUR_VERSION} --json-instances=inputs.json\n<\/code><\/pre>\n\n<p>I already tested my model localy and created a docker container with tensorflow serving and it works fine, but no success on cloud ml.<\/p>\n\n<p>The error promts:<\/p>\n\n<pre><code>\"error\": \"Prediction failed: Error processing input: Expected uint8, got '\\\\xff\\\\xd8\\\\...\n\n\\\\xff\\\\xd9' of type 'str' instead.\"\n<\/code><\/pre>\n","answers":[{"AnswerId":"53110594","CreationDate":"2018-11-01T23:04:56.547","ParentId":null,"OwnerUserId":"8507381","Title":null,"Body":"<p>The issue was with the way the graph was exported it is very important to add the flag --input_type when calling the script export_inference_graph.py otherwise the API models have their inputs a UINT8 instead of string.<\/p>\n\n<pre><code>--input_type encoded_image_string_tensor\n<\/code><\/pre>\n"}]}
+{"QuestionId":53109990,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-lite>","CreationDate":"2018-11-01T22:01:33.200","AcceptedAnswerId":53110304.0,"OwnerUserId":4987390.0,"Title":"tflite: get_tensor on non-output tensors gives random values","Body":"<p>I'm trying to debug my <code>tflite<\/code> model, that uses custom ops. I've found the correspondence between op names (in <code>*.pb<\/code>) and op ids (in <code>*.tflite<\/code>), and I'm doing a layer-per-layer comparison (to make sure the outputs difference are always in range <code>1e-4<\/code> (since it blows up at the end, I want to find the exact place where my custom layer fails) as follows:<\/p>\n\n<hr>\n\n<p><strong>Method 1:<\/strong> I use <code>get_tensor<\/code> to get the output as follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from tensorflow.contrib.lite.python import interpreter\n\n# load the model\nmodel = interpreter.Interpreter(model_path='model.tflite')\nmodel.allocate_tensors()\n\n# get tensors\nfor i in tensor_ids:\n    tensor_output[i] = model.get_tensor(i)\n<\/code><\/pre>\n\n<p>It show totally inadequate random values (comparing to the outputs of the TensorFlow model).<\/p>\n\n<hr>\n\n<p><strong>Method 2:<\/strong> convert the <code>*.pb<\/code> only up to a certain layer, then repeat, basically:<\/p>\n\n<ol>\n<li><p>Create a <code>*.pb<\/code> so that it contains the network only from <code>input<\/code> up to <code>layer_1<\/code>.<\/p><\/li>\n<li><p>Convert to <code>tflite<\/code> (so the output is now <code>layer_1<\/code>) and check the outputs of TF-Lite with TensorFlow.<\/p><\/li>\n<li><p>Repeat steps 1-2 for <code>layer_2<\/code>, <code>layer_3<\/code>, ... <code>outputs<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>This method requires much more work and executions, but it correctly shows that for built-in operations the outputs of <code>tflite<\/code> and <code>pb<\/code> models were identical, and only starts to differ in my custom ops (while in <em>Method 1<\/em> the outputs diverges right away from first layers).<\/p>\n\n<hr>\n\n<blockquote>\n  <p><strong>Question:<\/strong> Why the behaviour of <code>get_tensor<\/code> is so strange? Maybe it is because I am using <code>tensorflow 1.9<\/code>(when  TF-Lite was still not released and available only in developer preview)?<\/p>\n<\/blockquote>\n\n<p><em>PS: I am aware about the release of TF-Lite, but I've manually compiled TensorFlow 1.9 for my project and now it is hard to change the versioning.<\/em><\/p>\n","answers":[{"AnswerId":"53110304","CreationDate":"2018-11-01T22:32:39.393","ParentId":null,"OwnerUserId":"4789373","Title":null,"Body":"<p>I had the same problem few month ago. The thing is, TF-Lite is completely different from TensorFlow &ndash; it uses static memory and execution plans, memory mapping files for faster loading, and it is supposed to optimize everything possible in the network's forward propagation pipeline.<\/p>\n\n<p>I'm not a developer of TF-Lite, but I suppose it keeps its memory footprint extremely low by re-using the memory areas that were used for previously computed ops. Let's see the idea on following illustration:<\/p>\n\n<hr>\n\n<p><strong>Step 1:<\/strong> first, we're feeding the inputs to a symbolic tensor <code>I<\/code> (in parentheses). Let's say the value of it is stored in a buffer called <code>buffer_1<\/code>.<\/p>\n\n<pre><code>     op1       op2       op3\n(I) ----&gt;  A  ----&gt;  B  ----&gt;  O\n_________________________________\n^^^        ^^^^^^^^^^^^       ^^^\ninput      intermediate    output\ntensor     tensors         tensor\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Step 2:<\/strong> Now, we need to compute <code>op1<\/code> on symbolic tensor <code>I<\/code> to attain the symbolic tensor <code>A<\/code>. We compute on <code>buffer_1<\/code> and store the value of symbolic tensor <code>A<\/code> in a buffer called <code>buffer_2<\/code>.<\/p>\n\n<pre><code>    [op1]      op2       op3\n(I) ----&gt; (A) ----&gt;  B  ----&gt;  O\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Step 3:<\/strong> Now, we're computing <code>op2<\/code> on symbolic tensor <code>A<\/code> to attain the symbolic tensor <code>B<\/code>. We compute on <code>buffer_2<\/code> and store the value of symbolic tensor <code>B<\/code> in a buffer called <code>buffer_3<\/code>...<\/p>\n\n<pre><code>     op1      [op2]      op3\n I  ----&gt; (A) ----&gt; (B) ----&gt;  O\n<\/code><\/pre>\n\n<p><strong>But wait!<\/strong> Why waste our memory to store in <code>buffer_3<\/code> if we now have <code>buffer_1<\/code> that is unused, and the value of which is now useless for getting the output <code>O<\/code>? So, instead of storing in <code>buffer_3<\/code>, we will actually store results of this operation in <code>buffer_1<\/code>!<\/p>\n\n<p>That's the basic idea of efficient memory re-usage, which I think is implemented in TF-Lite, given its built-in static graph analyzer in <code>toco<\/code> and other stuffs. And that's why you can't simply apply <code>get_tensor<\/code> on non-output tensors.<\/p>\n\n<hr>\n\n<p><strong>An easier way to debug?<\/strong><\/p>\n\n<p>You've mentioned that you're writing a custom op, so I suppose you've built <code>tflite<\/code> with <code>bazel<\/code>, right? Then you can actually inject some logging code to <code>Interpreter::Invoke()<\/code> in the file <code>tensorflow\/lite\/interpreter.cc<\/code>. An ugly hack, but it works.<\/p>\n\n<p><em>PS: I would be glad if any TensorFlow Lite developers come across and give a comment on this :)<\/em><\/p>\n"},{"AnswerId":"53650265","CreationDate":"2018-12-06T11:18:28.023","ParentId":null,"OwnerUserId":"10375062","Title":null,"Body":"<p>Yes intermediate tensors can be overwritten unless specified as outputs.<\/p>\n\n<p>Edit:\nI managed to fix the problem by making all ops be in the output list during conversion. They are then preserved at runtime and the values can be read correctly.<\/p>\n\n<p>See:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/52369111\/obtaining-quantized-activations-in-tensorflow-lite\/52392629#52392629\">Obtaining quantized activations in tensorflow lite<\/a><\/p>\n"}]}
+{"QuestionId":53110159,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><conv-neural-network>","CreationDate":"2018-11-01T22:18:04.257","AcceptedAnswerId":null,"OwnerUserId":10588753.0,"Title":"Could not broadcast input array from shape (8,8,3) into shape (8,8,2048,3)","Body":"<p>So, I am again trying to run this Inception v3 fine tuning with keras, i think there is some problem with the arrays. I've already seen the other posts on these topics, but now i don\u00b4t know what to do. I don't have a lot of experience. So any help would be appreciated.<\/p>\n\n<p>Here is the entire code <a href=\"https:\/\/github.com\/c3s4grod\/PecuScope\" rel=\"nofollow noreferrer\">https:\/\/github.com\/c3s4grod\/PecuScope<\/a><\/p>\n\n<p>And this is the error that appears:<\/p>\n\n<p>Traceback (most recent call last):   <\/p>\n\n<blockquote>\n  <p>File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\utils\\data_utils.py\",\n  line 578, in get\n      inputs = self.queue.get(block=True).get()   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\multiprocessing\\pool.py\",\n  line 644, in get\n      raise self._value   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\multiprocessing\\pool.py\",\n  line 119, in worker\n      result = (True, func(*args, **kwds))   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\utils\\data_utils.py\",\n  line 401, in get_index\n      return _SHARED_SEQUENCES[uid][i]   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\preprocessing\\image.py\",\n  line 1034, in <strong>getitem<\/strong>\n      return self._get_batches_of_transformed_samples(index_array)   File\n  \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\preprocessing\\image.py\",\n  line 1442, in _get_batches_of_transformed_samples\n      batch_x[i] = x ValueError: could not broadcast input array from shape (8,8,3) into shape (8,8,2048,3)<\/p>\n<\/blockquote>\n\n<p>The above exception was the direct cause of the following exception:<\/p>\n\n<p>Traceback (most recent call last):   <\/p>\n\n<blockquote>\n  <p>File \"C:\/Users\/Shangai\/PycharmProjects\/PSai\/INCEPTION.py\", line 234, in\n  \n      validation_steps=steps_test)   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\legacy\\interfaces.py\",\n  line 91, in wrapper\n      return func(*args, **kwargs)   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\models.py\",\n  line 1315, in fit_generator\n      initial_epoch=initial_epoch)   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\legacy\\interfaces.py\",\n  line 91, in wrapper\n      return func(*args, **kwargs)   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\engine\\training.py\",\n  line 2194, in fit_generator\n      generator_output = next(output_generator)   File \"C:\\Users\\Shangai\\AppData\\Local\\conda\\conda\\envs\\PSAI\\lib\\site-packages\\keras\\utils\\data_utils.py\",\n  line 584, in get\n      six.raise_from(StopIteration(e), e)   File \"\", line 2, in raise_from StopIteration: could not broadcast input array from shape\n  (8,8,3) into shape (8,8,2048,3)<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53110536","CreationDate":"2018-11-01T22:58:12.877","ParentId":null,"OwnerUserId":"7570817","Title":null,"Body":"<p>It looks like your input tensor has not the correct size.\nIn general broadcasting only works if both arrays \/ tensors have the same number of dimensions. In your case that is not true: <\/p>\n\n<pre><code>8 8 3       -&gt; dim = 3\n8 8 2048 3  -&gt; dim = 4\n<\/code><\/pre>\n\n<p>You need to expand the first array and add an axis with size 1 at the right position:<\/p>\n\n<pre><code>8 8 1    3\n8 8 2048 3\n<\/code><\/pre>\n\n<p>Now the sizes of each axis match and for the 3. dimension broadcasting is possible.\nTo achieve this you can use the <em>newaxis<\/em>- or <em>None<\/em>-notation. <\/p>\n\n<pre><code>import numpy as np\nA = np.ones((8, 8, 3))\nB = np.ones((8, 8, 2048, 3))\nC = A + B # Fails\nA = A[:, :, np.newaxis, :]\nC = A + B # Works\n<\/code><\/pre>\n\n<p>Another possibility is to use the function <em>numpy.expand_dims<\/em>.\nThe same notation is available for tensors in <em>keras<\/em> \/ <em>tensorflow<\/em> (i.e.  <em>tensorflow.newaxis<\/em>, <em>tensorflow.expand_dims<\/em>).<\/p>\n"}]}
+{"QuestionId":53110171,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T22:19:30.747","AcceptedAnswerId":53126505.0,"OwnerUserId":4468490.0,"Title":"GPU device mapping information in Tensorflow","Body":"<p>This could be a trivial question. The problem is, I only have 1 GPU device. To test if <code>tensorflow<\/code> is using GPU as the accelerator, like discussed in this <a href=\"https:\/\/stackoverflow.com\/a\/38019608\/4468490\">question<\/a>, I run:<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<p>and similar to that answer, I have such device mapping information appeared twice:<\/p>\n\n<pre><code>Device mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: Tesla K80, pci bus id: 0000:00:04.0, compute capability: 3.7\n2018-11-01 22:00:25.678626: I tensorflow\/core\/common_runtime\/direct_session.cc:291] Device mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: Tesla K80, pci bus id: 0000:00:04.0, compute capability: 3.7\n<\/code><\/pre>\n\n<p>So what makes this happen?<\/p>\n","answers":[{"AnswerId":"53126505","CreationDate":"2018-11-02T22:23:54.717","ParentId":null,"OwnerUserId":"5879698","Title":null,"Body":"<p>The reason it is printed twice, is the first time the C code is doing a printf to screen and the 2nd time it is doing a log statement that also gets directed to the screen.<\/p>\n\n<p>If you want to disable all logging from C code run:<\/p>\n\n<pre><code>export TF_CPP_MIN_LOG_LEVEL=3\n<\/code><\/pre>\n\n<p>before running your python program.<\/p>\n\n<p>This is the code from tensorflow\/core\/common_runtime\/direct_session.cc<\/p>\n\n<pre><code>  if (options.config.log_device_placement()) {\n    const string mapping_str = device_mgr_-&gt;DeviceMappingString();\n    if (mapping_str.empty()) {\n      printf(\"Device mapping: no known devices.\\n\");\n    } else {\n      printf(\"Device mapping:\\n%s\", mapping_str.c_str());\n    }\n    LOG(INFO) &lt;&lt; \"Device mapping:\\n\" &lt;&lt; mapping_str;\n<\/code><\/pre>\n\n<p>from here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/direct_session.cc#L301-L309\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/direct_session.cc#L301-L309<\/a><\/p>\n"}]}
+{"QuestionId":53110770,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-01T23:28:39.270","AcceptedAnswerId":53118637.0,"OwnerUserId":7232350.0,"Title":"Tensorflow: create boolean matrix given True indices","Body":"<p>Suppose I have below two vectors, one for start index, the other for end index<\/p>\n\n<pre><code>start_index = [1, 1, 2]\nend_index = [3, 4, 3]\n<\/code><\/pre>\n\n<p>and below I have the shape of the final boolean matrix <\/p>\n\n<pre><code>shape = [3, 6]\n<\/code><\/pre>\n\n<p>I want to generate below boolean matrix<\/p>\n\n<pre><code>bool_mat = [[False, True,  True, True, False, False]\n            [False, True,  True, True, True,  False]\n            [False, False, True, True, False, False]]\n<\/code><\/pre>\n\n<p>for each row True starts from index in start_index and end at index in end_index, and False elsewhere,<\/p>\n\n<pre><code>bool_mat[i, start_index[i]:end_index[i]+1] = True\n<\/code><\/pre>\n\n<p>How to do this in TensorFlow? thanks!<\/p>\n","answers":[{"AnswerId":"53118637","CreationDate":"2018-11-02T12:28:31.510","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do that like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nstart_index = tf.constant([1, 1, 2])\nend_index = tf.constant([3, 4, 3])\nshape = tf.constant([3, 6])\ncol = tf.range(shape[1])\nresult = (col &gt;= start_index[:, tf.newaxis]) &amp; (col &lt;= end_index[:, tf.newaxis])\nwith tf.Session() as sess:\n    print(sess.run(result))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[False  True  True  True False False]\n [False  True  True  True  True False]\n [False False  True  True False False]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53110838,"AnswerCount":1,"Tags":"<python><image-processing><machine-learning><computer-vision><caffe>","CreationDate":"2018-11-01T23:37:10.877","AcceptedAnswerId":null,"OwnerUserId":3303546.0,"Title":"Error indicates flattened dimensions when loading pre-trained network","Body":"<h2>Question<\/h2>\n\n<p>I'm trying to load a pretrained network, and I'm getting the following error<\/p>\n\n<blockquote>\n  <p>F1101 23:03:41.857909    73 net.cpp:757] Cannot copy param 0 weights\n  from layer 'fc4'; shape mismatch.  Source param shape is 512 4096\n  (2097152); target param shape is 512 256 4 4 (2097152). To learn this\n  layer's parameters from scratch rather than copying from a saved net,\n  rename the layer.<\/p>\n<\/blockquote>\n\n<p>I notice that 512 x 256 x 4 x 4 == 512 x 4096, so it seems that in saving and reloading the network weights the layer was somehow flattened.<\/p>\n\n<p><strong>How can I counteract this error?<\/strong><\/p>\n\n<h2>To reproduce<\/h2>\n\n<p>I'm trying to use the D-CNN pretrained network in <a href=\"https:\/\/github.com\/songxinhang\/D-CNN\" rel=\"nofollow noreferrer\">this GitHub repository<\/a>. <\/p>\n\n<p>I load the network with <\/p>\n\n<pre><code>import caffe\nnet = caffe.Net('deploy_D-CNN.prototxt', 'D-CNN.caffemodel', caffe.TEST)\n<\/code><\/pre>\n\n<p>The prototxt file is <\/p>\n\n<pre><code>name: \"D-CNN\"\ninput: \"data\"\ninput_dim: 10\ninput_dim: 3\ninput_dim: 259\ninput_dim: 259\nlayer {\n  name: \"conv1\"\n  type: \"Convolution\"\n  bottom: \"data\"\n  top: \"conv1\"\n  convolution_param {\n    num_output: 64\n    kernel_size: 5\n    stride: 2\n  }\n}\nlayer {\n  name: \"relu1\"\n  type: \"ReLU\"\n  bottom: \"conv1\"\n  top: \"conv1\"\n}\nlayer {\n  name: \"pool1\"\n  type: \"Pooling\"\n  bottom: \"conv1\"\n  top: \"pool1\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 2\n    stride: 2\n  }\n}\nlayer {\n  name: \"norm1\"\n  type: \"LRN\"\n  bottom: \"pool1\"\n  top: \"norm1\"\n  lrn_param {\n    local_size: 5\n    alpha: 0.0001\n    beta: 0.75\n  }\n}\nlayer {\n  name: \"conv2\"\n  type: \"Convolution\"\n  bottom: \"norm1\"\n  top: \"conv2\"\n  convolution_param {\n    num_output: 128\n    pad: 1\n    kernel_size: 3\n  }\n}\nlayer {\n  name: \"relu2\"\n  type: \"ReLU\"\n  bottom: \"conv2\"\n  top: \"conv2\"\n}\nlayer {\n  name: \"pool2\"\n  type: \"Pooling\"\n  bottom: \"conv2\"\n  top: \"pool2\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 2\n    stride: 2\n  }\n}\nlayer {\n  name: \"conv3\"\n  type: \"Convolution\"\n  bottom: \"pool2\"\n  top: \"conv3\"\n  convolution_param {\n    num_output: 256\n    pad: 1\n    kernel_size: 3\n    stride: 1\n  }\n}\nlayer {\n  name: \"relu3\"\n  type: \"ReLU\"\n  bottom: \"conv3\"\n  top: \"conv3\"\n}\nlayer {\n  name: \"fc4\"\n  type: \"Convolution\"\n  bottom: \"conv3\"\n  top: \"fc4\"\n  convolution_param {\n    num_output: 512\n    pad: 0\n    kernel_size: 4\n  }\n}\nlayer {\n  name: \"relu4\"\n  type: \"ReLU\"\n  bottom: \"fc4\"\n  top: \"fc4\"\n}\nlayer {\n  name: \"drop4\"\n  type: \"Dropout\"\n  bottom: \"fc4\"\n  top: \"fc4\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\nlayer { \n  name: \"pool5_spm3\"\n  type: \"Pooling\"\n  bottom: \"fc4\"\n  top: \"pool5_spm3\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 10\n    stride: 10\n  }\n}\nlayer {\n  name: \"pool5_spm3_flatten\"\n  type: \"Flatten\"\n  bottom: \"pool5_spm3\"\n  top: \"pool5_spm3_flatten\"\n}\nlayer { \n  name: \"pool5_spm2\"\n  type: \"Pooling\"\n  bottom: \"fc4\"\n  top: \"pool5_spm2\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 14\n    stride: 14\n  }\n}\nlayer {\n  name: \"pool5_spm2_flatten\"\n  type: \"Flatten\"\n  bottom: \"pool5_spm2\"\n  top: \"pool5_spm2_flatten\"\n}\nlayer { \n  name: \"pool5_spm1\"\n  type: \"Pooling\"\n  bottom: \"fc4\"\n  top: \"pool5_spm1\"\n  pooling_param {\n    pool: MAX\n    kernel_size: 29\n    stride: 29\n  }\n}\nlayer {\n  name: \"pool5_spm1_flatten\"\n  type: \"Flatten\"\n  bottom: \"pool5_spm1\"\n  top: \"pool5_spm1_flatten\"\n}\nlayer {\n  name: \"pool5_spm\"\n  type: \"Concat\"\n  bottom: \"pool5_spm1_flatten\"\n  bottom: \"pool5_spm2_flatten\"\n  bottom: \"pool5_spm3_flatten\"\n  top: \"pool5_spm\"\n  concat_param {\n    concat_dim: 1\n  }\n}\n\n\nlayer {\n  name: \"fc4_2\"\n  type: \"InnerProduct\"\n  bottom: \"pool5_spm\"\n  top: \"fc4_2\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 512\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.005\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0.1\n    }\n  }\n}\nlayer {\n  name: \"relu4\"\n  type: \"ReLU\"\n  bottom: \"fc4_2\"\n  top: \"fc4_2\"\n}\nlayer {\n  name: \"drop4\"\n  type: \"Dropout\"\n  bottom: \"fc4_2\"\n  top: \"fc4_2\"\n  dropout_param {\n    dropout_ratio: 0.5\n  }\n}\n\nlayer {\n  name: \"fc5\"\n  type: \"InnerProduct\"\n  bottom: \"fc4_2\"\n  top: \"fc5\"\n  param {\n    lr_mult: 1\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 19\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayer {\n  name: \"prob\"\n  type: \"Softmax\"\n  bottom: \"fc5\"\n  top: \"prob\"\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"54996684","CreationDate":"2019-03-05T06:30:11.357","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you are taking a pretrained net where layer <code>\"fc4\"<\/code> was a fully connected layer (aka <code>type: \"InnerProduct\"<\/code> layer) and it was \"reshaped\" into a convolutional layer.<br>\nSince both inner product layer and convolutional layer performs roughly the same linear operation on the inputs this change can be made under certain assumptions (see, e.g. <a href=\"https:\/\/stats.stackexchange.com\/q\/263349\/66467\">here<\/a>).<br>\nAs you already identified correctly the weights of the original pre-trained fully connected layer were saved \"flattened\" w.r.t the shape caffe expects for a convolutional layer.<\/p>\n\n<p>I think a solution to this problem can be using <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L307\" rel=\"nofollow noreferrer\"><code>share_mode: PERMISSIVE<\/code><\/a>:<\/p>\n\n<pre><code>layer {\n  name: \"fc4\"\n  type: \"Convolution\"\n  bottom: \"conv3\"\n  top: \"fc4\"\n  convolution_param {\n    num_output: 512\n    pad: 0\n    kernel_size: 4\n  }\n  param {\n    lr_mult: 1\n    decay_mult: 1\n    share_mode: PERMISSIVE  # should help caffe overcome the shape mismatch\n  }\n  param {\n    lr_mult: 2\n    decay_mult: 0\n    share_mode: PERMISSIVE\n  }\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":53110884,"AnswerCount":0,"Tags":"<elasticsearch><tensorflow>","CreationDate":"2018-11-01T23:43:45.550","AcceptedAnswerId":null,"OwnerUserId":2051586.0,"Title":"Deep Learning Model that predicts optimal ElasticSearch Query-time Boost Values","Body":"<p>I would like to use a machine learning model to train the optimal boost values to use when calling <a href=\"https:\/\/www.elastic.co\/guide\/en\/elasticsearch\/guide\/current\/query-time-boosting.html\" rel=\"nofollow noreferrer\">ElasticSearch Query-time Boosting<\/a>.<\/p>\n\n<p>The ElasticSearch documentation recommends a trial-and-error approach:<\/p>\n\n<blockquote>\n  <p>Apply a boost and check the results. Change the boost and check again.<\/p>\n<\/blockquote>\n\n<p>I would like to be more calculated in what these values should be and leverage machine learning with randomized input values to predict the optimal boost values for given search terms. I looked at <a href=\"https:\/\/ai.stackexchange.com\/questions\/3850\/using-machine-deep-learning-for-guessing-pseudo-random-generator\">this post<\/a> as potentially similar to the problem I am trying to solve, but wanted to make sure I am heading in the right direction before getting too far along.<\/p>\n","answers":[]}
+{"QuestionId":53111182,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-02T00:25:53.683","AcceptedAnswerId":53111643.0,"OwnerUserId":4795896.0,"Title":"Why does the keras model compile but the fit_generator command throws a 'model not compiled runtime error'?","Body":"<p>I've set up a custom CNN with K-fold cross-validation in keras with tensorflow back-end. \nThe model.compile() function is called before starting the training, but the model.fit_generator() function call results in an runtime error: \"You must compile your model before using it.\"<\/p>\n\n<p>I do data augmentation with ImageDataGenerator and use the fit_generator function for training.<\/p>\n\n<p>The only related issue I've found up untill now had to do with the tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18287\" rel=\"nofollow noreferrer\">eager execution<\/a> feature which seems to not be enabled in Keras. <\/p>\n\n<p>Here is the code:<\/p>\n\n<p>model definition:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(24, (5, 5), \n          strides=(1, 1), \n          padding=\"valid\", \n          data_format=\"channels_last\",\n          activation='relu', \n          use_bias=True,\n          ))#out=96\nmodel.add(Dropout(.25))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))#out=48\nmodel.add(Conv2D(32, (3, 3), \n          strides=(1, 1), \n          padding=\"valid\", \n          data_format=\"channels_last\",\n          activation='relu', \n          use_bias=True,\n          ))#out=46\nmodel.add(MaxPooling2D(pool_size=(2, 2)))#out=23\nmodel.add(Conv2D(48, (3, 3), \n          strides=(1, 1), \n          padding=\"valid\", \n          data_format=\"channels_last\",\n          activation='relu', \n          use_bias=True,\n          ))#out=21\nmodel.add(MaxPooling2D(pool_size=(2, 2)))#padding???\nmodel.add(Flatten())\nmodel.add(Dense(3, activation='softmax'))\n<\/code><\/pre>\n\n<p>...here would be the data initialization...<\/p>\n\n<p>setting up the ImageGenerator:<\/p>\n\n<pre><code>datagen_training = ImageDataGenerator(\n            rotation_range = 20,\n            width_shift_range = 0.3,\n            height_shift_range=0.3,\n            zoom_range=0.2,\n            fill_mode = \"constant\",\n            cval = 0,\n            vertical_flip = True,\n            validation_split = 0.2\n            )\ndatagen_training.fit(data)\n<\/code><\/pre>\n\n<p>model setup and training:<\/p>\n\n<pre><code>rmsprop = optimizers.RMSprop(lr=0.001)#docu says to only tune the learning rate\nkf = KFold(n_splits=FOLDS, shuffle = True, random_state=78945)\nmodel.compile(rmsprop, loss = losses.categorical_crossentropy, metrics=[metrics.categorical_accuracy])\nacc_hist = []\nwhile True:\n    history = object()\n    for train_idx, val_idx in kf.split(data, labels):\n        x_train, y_train = data[train_idx], labels[train_idx]\n        x_val, y_val = data[val_idx], labels[val_idx]\n        data_iterator = datagen_training.flow(x_train, y_train, batch_size=BATCH_SIZE)\n        history = model.fit_generator(data_iterator, steps_per_epoch=len(x_train) \/\/ BATCH_SIZE, epochs=1)\n    acc_hist.append(history.history['categorical_accuracy'][0])\n    #stop if accuracy doesn't change within 3 epochs\n    if stopping_criterion_met:\n        break\n<\/code><\/pre>\n","answers":[{"AnswerId":"53111643","CreationDate":"2018-11-02T01:41:47.693","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The model isn't compiling because it isn't being built, and it isn't being built because the <code>input_shape<\/code> isn't being specified in the first layer.  I'm not sure the input shape of your data, but something like this for your first layer would get the model to compile:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(24, (5, 5), \n      strides=(1, 1), \n      padding=\"valid\", \n      data_format=\"channels_last\",\n      activation='relu', \n      use_bias=True,\n      input_shape=(100,100,1)\n      ))#out=96\n<\/code><\/pre>\n"}]}
+{"QuestionId":53111339,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning>","CreationDate":"2018-11-02T00:53:07.353","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"Need help on stop training of Keras network","Body":"<p>I have the following question. When we teach the Keras NN, we have something like <\/p>\n\n<pre><code>    earlyStopping = EarlyStopping(\n    monitor='val_loss', patience=30, verbose=0, mode='auto')\n\nresults = model.fit_generator(\n    Generator(train_ids, \n        batch_size = TRAINING_BATCH_SIZE), \n    steps_per_epoch = STEPS_PER_EPOCH, # fill_in? \n    epochs = MAX_TRAIN_EPOCHS,\n    validation_data=Generator(valid_ids, \n        batch_size = VALIDATION_BATCH_SIZE),\n    validation_steps=VALIDATION_STEPS, # fill_in?\n    callbacks=[#checkpoint, # Commented as Kaggle does not save\n        earlyStopping #, plot_losses\n        ])\n<\/code><\/pre>\n\n<p>Now, this code snippet will stop when loss stop improving, but... \nLet's say, it goes like 0.9, 0.8, 0.7, 0.6, 0.5, 0.5, 0.6\nThen, when it stops, will my NN be the one that produced 0.5 (best) or 0.6 (last)?<\/p>\n\n<p>In other words, should I keep testing the loss every epoch, and if a new min. reached, to save the NN, in order to get the best one? Or is there a way of doing it automatically?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53121065","CreationDate":"2018-11-02T15:04:07.283","ParentId":null,"OwnerUserId":"6898439","Title":null,"Body":"<p>I think the question is:<\/p>\n\n<p><strong>1. how to save the best checkpoint?<\/strong><\/p>\n\n<p>you can either save manually or use keras callback, but either way, you need some metric to choose the best weiths.<\/p>\n\n<blockquote>\n  <p>keras.callbacks.ModelCheckpoint(filepath, monitor='val_loss',\n  verbose=0, save_best_only=False, save_weights_only=False, mode='auto',\n  period=1)<\/p>\n<\/blockquote>\n\n<p><strong>So, Do I need keep evaluating my model on validation dataset to find out the best weights?<\/strong><\/p>\n\n<p>It depends on how you choose the best weights. Usually we want the weights perform best on validation dataset, so you need periodically evaluate your model on validation dataset, otherwise you just don't know .<\/p>\n"}]}
+{"QuestionId":53111518,"AnswerCount":1,"Tags":"<keras><conv-neural-network>","CreationDate":"2018-11-02T01:23:42.577","AcceptedAnswerId":53223222.0,"OwnerUserId":10593920.0,"Title":"Wrong number of classes in Keras","Body":"<p>Keras is finding a wrong number of classes in train and test set folders. I have 3 classes, but it keeps saying that there are 4. Can anyone help me, please?<\/p>\n\n<p>Here the code:<\/p>\n\n<pre><code>cnn = Sequential()\n\ncnn.add(Conv2D(32, (3, 3), input_shape = (64, 64, 3), activation = 'relu'))\ncnn.add(Dropout(0.5))\ncnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n\ncnn.add(Conv2D(32, (3, 3), activation = 'relu'))\ncnn.add(Dropout(0.5))\ncnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n\ncnn.add(Conv2D(64, (3, 3), activation = 'relu'))\ncnn.add(Dropout(0.5))\ncnn.add(MaxPooling2D(pool_size = (2, 2)))\n\ncnn.add(Conv2D(128, (3, 3), activation = 'relu'))\ncnn.add(Dropout(0.5))\ncnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n#Full connection\ncnn.add(Dense(units = 64, activation = 'relu'))\ncnn.add(Dense(units = 64, activation = 'relu'))\ncnn.add(Dense(units = 3, activation = 'softmax'))\n\n# Compiling the CNN\ncnn.compile(optimizer = OPTIMIZER, loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\n\n     #Fitting\n    from keras.preprocessing.image import ImageDataGenerator\n\n    train_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                       shear_range = 0.2,\n                                       zoom_range = 0.2,\n                                       horizontal_flip = True)\n\n    test_datagen = ImageDataGenerator(rescale = 1.\/255)\n\n    training_set = train_datagen.flow_from_directory('dataset\/training_set',\n                                                     target_size = tgt_size,\n                                                     batch_size = batch_size,\n                                                     class_mode = 'categorical')\n\n    test_set = test_datagen.flow_from_directory('dataset\/test_set',\n                                                target_size = tgt_size,\n                                                batch_size = batch_size,\n                                                class_mode = 'categorical')\n<\/code><\/pre>\n\n<p>And the error:<\/p>\n\n<pre><code>Found 12000 images belonging to 4 classes.\nFound 3000 images belonging to 4 classes.\n\nEpoch 1\/10\n---------------------------------------------------------------------------\nValueError: Error when checking target: expected dense_15 to have 4 dimensions, but got array with shape (3, 4)\n<\/code><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>It only happens with Jupyter Notebook in Google Cloud. When I use Spyder locally, it finds the correct number of classes.<\/p>\n","answers":[{"AnswerId":"53223222","CreationDate":"2018-11-09T09:41:10.800","ParentId":null,"OwnerUserId":"3415213","Title":null,"Body":"<p>As you might have found out by yourself by now, Jupyter creates hidden checkpoint folders for backup purposes. That's why there is always one class (as in folder) extra when using flow_from_directory.\nThe simplest solution would be to just delete that hidden folder.<\/p>\n"}]}
+{"QuestionId":53111751,"AnswerCount":0,"Tags":"<python><neural-network><keras><multi-layer>","CreationDate":"2018-11-02T02:00:26.820","AcceptedAnswerId":null,"OwnerUserId":3837778.0,"Title":"Keras: can I use a Flatten layer as input to multilayer perceptron model","Body":"<p>I am trying to use MLP to classify images in Keras. My codes are:<\/p>\n\n<pre><code>from keras.preprocessing import sequence\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten\nfrom keras.preprocessing.image import ImageDataGenerator\n\n# Prepare data generators for train and test\nSIZE = 32\ntrain_datagen = ImageDataGenerator(\n    rescale=1.\/255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True,\n    validation_split=0.2)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    'train',\n    target_size=(SIZE, SIZE),\n    class_mode='categorical',\n    batch_size=32,\n    subset='training')\n\nval_generator = train_datagen.flow_from_directory(\n    'train',\n    target_size=(SIZE, SIZE),\n    class_mode='categorical',\n    batch_size=32,\n    subset='validation')\n\ntest_generator = test_datagen.flow_from_directory(\n    'test',\n    target_size=(SIZE, SIZE),\n    class_mode='categorical',\n    batch_size=32)\n\nmodel_mlp = Sequential()\nmodel_mlp.add(Flatten(input_shape=(32,32, 3)))\nmodel_mlp.add(Dense(256, activation='relu'))\nmodel_mlp.add(Dropout(0.5))\nmodel_mlp.add(Dense(128, activation='relu'))\nmodel_mlp.add(Dropout(0.5))\nmodel_mlp.add(Dense(68, activation='softmax'))\n\n# Compile\nmodel_mlp.compile(optimizer='rmsprop',  # other choices: adam or SGD\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\nprint(model_mlp.summary())\n\nhistory = model_mlp.fit_generator(\n    train_generator,\n    steps_per_epoch=100,\n    epochs=100,\n    validation_data=val_generator,\n    validation_steps=100)\n\n# Showing\nshow(history, model_mlp)\n<\/code><\/pre>\n\n<p>Can I use <code>Flatten<\/code> layer as input to multilayer perceptron model to reshape 32*32*3 image into vector as above? Thanks<\/p>\n\n<p><strong>Any problems with my codes? because the <code>accuracy<\/code> is less than 5%, which is too low.<\/strong><\/p>\n","answers":[]}
+{"QuestionId":53112118,"AnswerCount":0,"Tags":"<python><tensorflow><tpu>","CreationDate":"2018-11-02T02:59:43.873","AcceptedAnswerId":null,"OwnerUserId":6642285.0,"Title":"TensorFlow TPU | Initializer for variable is from a control-flow construct","Body":"<p>I'm using TensorFlow TPUs to train a model but when I try to convert to the code to TPU compatible using tf.contrib.tpu.rewrite, I get an error:<\/p>\n\n<blockquote>\n  <p>ValueError: Initializer for variable generator\/decoder\/layer4\/kernel\/ is from inside a control-flow construct, such as a loop or conditional. When creating a variable inside a loop or conditional, use a lambda as the initializer<\/p>\n<\/blockquote>\n\n<p>I looked around, and it seems that there's usually an if-statement or while loop that causes the error. I however don't have any of these.<\/p>\n\n<p>Now, here's the variable definition (kernel):<\/p>\n\n<pre><code>def conv1d_transpose(input, filters, kernel_size, strides, activation=None):\n  kernel = tf.Variable(tf.zeros(dtype=tf.float32, shape=[kernel_size, filters, input.shape[-1].value]), name='kernel')\n  output_shape = [tf.shape(input)[0], strides * tf.shape(input)[1], filters]\n  net = tf.contrib.nn.conv1d_transpose(input, filter=kernel, output_shape=output_shape, stride=strides, padding='SAME')\n  net = activation(net)\n  return net\n<\/code><\/pre>\n\n<p>I call that function multiple times. Here's what the call looks like inside the model function:<\/p>\n\n<pre><code>  with tf.name_scope(\"layer4\"):\n    net = conv1d_transpose(\n        net,\n        filters=15,\n        kernel_size=5,\n        strides=2,\n        activation=tf.nn.leaky_relu)\n    net = tf.concat([net, conv4e], axis=-1)\n    net = tf.layers.batch_normalization(net, training=training)\n<\/code><\/pre>\n\n<p>One more thing, before raising the error, I get a warning\/error from the tpu.replicator:<\/p>\n\n<blockquote>\n  <p>[[{{node input1}} = TPUReplicatedInputN=1, T=DT_BOOL]]\n  2018-11-02 02:55:44.242495: E tensorflow\/core\/common_runtime\/executor.cc:630] Executor failed to create kernel. Not found: No registered 'TPUReplicatedInput' OpKernel for CPU devices compatible with node {{node input1}} = TPUReplicatedInputN=1, T=DT_BOOL\n      .  Registered:  <\/p>\n<\/blockquote>\n\n<p>However it's pretty hard to debug since it doesn't specify the name of the node that is causing the issue.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53112255,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-02T03:22:22.363","AcceptedAnswerId":null,"OwnerUserId":4500293.0,"Title":"Tensorflow Broadcasting across one dimension","Body":"<p>I have a tensor A that is [m,n,k] and I have another tensor B that is [m,n]. I would like to add B in an element wise fashion across the k dimension of A. Is there a way of doing this efficiently in TensorFlow? I was looking into broadcast_to but I don't think that is what I'm looking for. <\/p>\n\n<p>Does the adding operation automatically broadcast? As in, if I did output = A + B explicitly?<\/p>\n\n<p>Thank you for the help!<\/p>\n","answers":[{"AnswerId":"53112264","CreationDate":"2018-11-02T03:24:06.220","ParentId":null,"OwnerUserId":"4500293","Title":null,"Body":"<p>tf.add seems to support broadcasting. Needed to read the documentation more carefully!<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/add\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/add<\/a><\/p>\n"}]}
+{"QuestionId":53112751,"AnswerCount":1,"Tags":"<python><r><tensorflow><keras><anaconda>","CreationDate":"2018-11-02T04:41:53.513","AcceptedAnswerId":null,"OwnerUserId":9873347.0,"Title":"Error when calling keras in R with keras_model_sequential()","Body":"<p>I want to use Keras in R but, already install anaconda, then in R run<\/p>\n\n<pre><code>library(keras)\ninstall_keras()\n<\/code><\/pre>\n\n<p>to start the installation process, but when I want to use the library using  <code>keras_model_sequential()<\/code> it shows the following error.<\/p>\n\n<p>how can i fix it??<\/p>\n\n<pre><code>  &gt; model_keras &lt;- keras_model_sequential()\n    Error: ImportError: Traceback (most recent call last):\n      File \"C:\\Users\\maty\\ANACON~1\\envs\\R-TENS~1\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 14, in swig_import_helper\n        return importlib.import_module(mname)\n      File \"C:\\Users\\maty\\ANACON~1\\envs\\R-TENS~1\\lib\\importlib\\__init__.py\", line 126, in import_module\n        return _bootstrap._gcd_import(name[level:], package, level)\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 994, in _gcd_import\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 971, in _find_and_load\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 955, in _find_and_load_unlocked\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 658, in _load_unlocked\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 571, in module_from_spec\n      File \"&lt;frozen importlib._bootstrap_external&gt;\", line 922, in create_module\n      File \"&lt;frozen importlib._bootstrap&gt;\", line 219, in _call_with_frames_removed\n    ImportError: DLL load failed: Error en una rutina de inicializaci\u00c3\u00b3n de biblioteca\n<\/code><\/pre>\n","answers":[{"AnswerId":"55622349","CreationDate":"2019-04-10T22:51:07.930","ParentId":null,"OwnerUserId":"10578691","Title":null,"Body":"<p>Did you install keras and tensorflow in conda?<\/p>\n\n<p>I know the general tips you get from online resources are to install anaconda for python, keras and tensorflow, but I approached it from a different way. I didn't install anaconda, just python3.7, which automatically has pip installed. Then I use pip3 install tensorflow and keras, R was able to run keras just fine.<\/p>\n"}]}
+{"QuestionId":53113316,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-02T05:55:06.923","AcceptedAnswerId":null,"OwnerUserId":9520630.0,"Title":"Different AUC result from tf.py_func and sklearn during training in Keras","Body":"<p>I want to calculate the AUC of validation set(x_val,y_val) each epoch when I train my model by roc_auc_curve in sklearn.metrics.<\/p>\n\n<p>After I stack my neural network model:<\/p>\n\n<pre><code>model=Sequential()\n      .\n      .\n      .\nmodel.add(Dense(...))\n<\/code><\/pre>\n\n<p>I have two different method to calculate AUC.\nThe first one is by tf.py_func:<\/p>\n\n<pre><code>def auc(y_true, y_pred):\n    return tf.py_func(roc_auc_score, (y_true, y_pred), tf.double)\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy',auc])\nmodel.fit(...,validation_data=(x_val,y_val),...)\n<\/code><\/pre>\n\n<p>And the second is by keras.callbacks:<\/p>\n\n<pre><code>class train_history(Callback):\n    def __init__(self,validation_data):\n        self.x_val = validation_data[0]\n        self.y_val = validation_data[1]\n    def on_train_begin(self, logs={}):\n        self.val_auc= []\n    def on_epoch_end(self, epoch, logs={}):\n        y_pred_val = self.model.predict(self.x_val)\n        val_rocauc = roc_auc_score(self.y_val, y_pred_val)\n        self.val_auc.append(logs['val_rocauc'])\nhis=train_history(validation_data=(x_val, y_val))\ncheckpoint=[his]\nmodel.fit(...,validation_data=(x_val,y_val),callbacks=checkpoint,...)\n<\/code><\/pre>\n\n<p>But the result from these two methods are different. And the second one seems correctly after I check these two methods. <\/p>\n\n<p>Is there any one know why? Thanks a lot.<\/p>\n","answers":[]}
+{"QuestionId":53113327,"AnswerCount":1,"Tags":"<c++11><tensorflow>","CreationDate":"2018-11-02T05:56:14.830","AcceptedAnswerId":null,"OwnerUserId":10594621.0,"Title":"tensorflow c++ error: Unary VariantShapeFn for type_name: int already registered","Body":"<p>i compliled tensorflow with bazel successfully and got libtensorflow_cc.so and libtensorflow_framework.so. Then i built a c++ binary with those share libraries successfully, but errors occurred when i run the binary, info is like:<\/p>\n\n<pre><code>F tensorflow\/core\/framework\/variant_op_registry.cc:51] Check failed: existing == nullptr (0x24fc538 vs. nullptr)Unary VariantShapeFn for type_name: int already registered\n<\/code><\/pre>\n\n<p>c++ code:<\/p>\n\n<pre><code>#include \"tensorflow\/cc\/client\/client_session.h\"\n#include \"tensorflow\/cc\/ops\/standard_ops.h\"\n#include \"tensorflow\/core\/framework\/tensor.h\"\nint main() {\n    using namespace tensorflow;\n    using namespace tensorflow::ops;\n    Scope root = Scope::NewRootScope();auto A = Const(root, { {3.f, 2.f}, {-1.f, 0.f}  });\n    \/\/     \/\/ Vector b = [3 5\n    auto b = Const(root, { {3.f, 5.f}  });\n    auto v = MatMul(root.WithOpName(\"v\"), A, b, MatMul::TransposeB(true));\n    std::vector&lt;Tensor&gt; outputs;\n    ClientSession session(root);\n    \/\/ Run and fetch v\n    TF_CHECK_OK(session.Run({v}, &amp;outputs));\n    \/\/ Expect outputs[0] == [19; -3]\n    LOG(INFO) &lt;&lt; outputs[0].matrix&lt;float&gt;();\n    return 0;\n}\n<\/code><\/pre>\n\n<p>i got my c++ binary using cmd:<\/p>\n\n<pre><code>g++ -std=c++11 test.cpp -o test -I.\/include -L.\/lib -L\/home\/work\/chengjy\/tools\/protobuf\/lib -I\/home\/work\/chengjy\/tools\/tensorflow\/eigen\/eigen\/eigen-eigen-fd6845384b86 -I\/home\/work\/chengjy\/tools\/protobuf\/include -ltensorflow_cc -ltensorflow_framework  -lprotobuf -lpthread -ldl -O3 -Wall\n<\/code><\/pre>\n\n<p>did anyone have this problem before?<\/p>\n","answers":[{"AnswerId":"53113893","CreationDate":"2018-11-02T06:53:39.390","ParentId":null,"OwnerUserId":"10594621","Title":null,"Body":"<p>it seems that i fixed this problem by removing -ltensorflow_framework from the compile cmd.<\/p>\n"}]}
+{"QuestionId":53113976,"AnswerCount":1,"Tags":"<python><tensorflow><keras><kaggle>","CreationDate":"2018-11-02T07:03:21.480","AcceptedAnswerId":null,"OwnerUserId":10114131.0,"Title":"Tensorflow and Keras Problem ValueError: Error when checking input: expected dense_9_input to have shape (24,) but got array with shape (0,)","Body":"<p>Hello I was doing a competition for a kaggle but then while I was practicing with Keras, I encountered a serious problem. The error is \"ValueError: Error when checking input: expected dense_9_input to have shape (24,) but got array with shape (0,)\". I tried to solve this problem by looking through Stack overflow but I couldn't find a right answer.<\/p>\n\n<p>The variable X contains all the variables except the target value which is winPlacePerc and string variables such as ID, match id and group id.<\/p>\n\n<pre><code>#Libraries\n\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\n\nseed = 0\nnp.random.seed(seed)\ntf.set_random_seed(seed)\n\ntrain = pd.read_csv(\"train_V2.csv\")\n\n\nX = train.iloc[:,3:1]\nY = train.iloc[:,-1]\n\n\nmodel = Sequential()\nmodel.add(Dense(8, input_dim=24,activation='relu'))\nmodel.add(Dense(8, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam',metrics=['accuracy'])\n\nmodel.fit(X, Y, epochs=200, batch_size=10)\n\nprint(\"\\n Accuracy: %.4f\" % (model.evaluate(X, Y)[1]))\n<\/code><\/pre>\n\n<p>The Data set can be found here :<a href=\"https:\/\/www.kaggle.com\/overload10\/pubg-predicting-chicken-dinner\/data\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/overload10\/pubg-predicting-chicken-dinner\/data<\/a><\/p>\n\n<p>I am here for comments to solve this problem. I started learning Tensorflow\/ Keras right after doing Automate the Boring Stuff with Python, so I am a super duper noob. I take comments, and advises. Thank you for looking at my first question. Please help! :((((<\/p>\n","answers":[{"AnswerId":"53115035","CreationDate":"2018-11-02T08:28:24.903","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>Your input of Keras does not match your numpy array. Figure out the size of your numpy array and change the input shape accordingly.  In case of more than 2 dimensions use a flatten layer before dense layers. <\/p>\n\n<p>Remember that the first dimension of your np array will be interpreted as the batch size<\/p>\n"}]}
+{"QuestionId":53114494,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-02T07:42:18.770","AcceptedAnswerId":null,"OwnerUserId":7829500.0,"Title":"tensorflow calculating loss with GPU vs. CPU","Body":"<p>I'm a bit confused with the difference I'm getting in loss calculations when calculating loss with a GPU vs. CPU.<\/p>\n\n<p>Model is a 6 layer CNN. <\/p>\n\n<p>Loaded the model from checkpoints and ran it with the same data. Calculated the loss with a CPU and then a GPU.<\/p>\n\n<p>CPU loss: 0.4687191\nGPU loss: 0.46873742<\/p>\n\n<p>Could someone explain to me why these losses were calculated differently?<\/p>\n\n<pre><code>#WITH CPU! - testing cpu vs cpu optimizer calculations\nimport time\ntf.reset_default_graph()\n\n\n\nnew_saver = tf.train.import_meta_graph('.\/graph\/final.meta')\n\nwith tf.Session() as sess:\n  new_saver.restore(sess, tf.train.latest_checkpoint('.\/tmp'))\n\n  optimize = tf.get_default_graph().get_operation_by_name( \"optimizer\" )\n  c_loss = sess.run('loss:0', feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n  print('initial:  c_loss', c_loss)  \n\n  sess.run(optimize, feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n  c_loss = sess.run('loss:0', feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n\n  print('post:  c_loss', c_loss)\n<\/code><\/pre>\n\n<p>Output:<br>\ninitial:  c_loss 0.4687191<br>\npost:  c_loss 0.5455321<br><\/p>\n\n<pre><code>#WITH GPU! - testing cpu vs gpu optimizer calculations\nimport time\ntf.reset_default_graph()\n\n\nnew_saver = tf.train.import_meta_graph('gdrive\/My Drive\/graph\/final.meta')\n\nwith tf.Session() as sess:\n  new_saver.restore(sess, tf.train.latest_checkpoint('gdrive\/My Drive\/tmp'))\n\n  optimize = tf.get_default_graph().get_operation_by_name( \"optimizer\" )\n  c_loss = sess.run('loss:0', feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n  print('initial:  c_loss', c_loss)  \n\n  sess.run(optimize, feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n  c_loss = sess.run('loss:0', feed_dict={'x:0': x_train[0:128], 'y:0': y_train[0:128], 'is_training:0': True})\n\n  print('post:  c_loss', c_loss)\n<\/code><\/pre>\n\n<p>Output: <br>\ninitial: c_loss 0.46873742 <br>\npost:  c_loss 0.5432756 <br><\/p>\n\n<p>EDIT:\nAlso. Want to add I loaded the model with two different CPU sessions and found that the loss calculations above were identical. They only vary when I calculate the loss with GPU.<\/p>\n","answers":[{"AnswerId":"53114685","CreationDate":"2018-11-02T07:58:07.457","ParentId":null,"OwnerUserId":"6067528","Title":null,"Body":"<p>They might be different if you're shuffling the data you train the model on. Try fixing a <code>numpy.random.seed(123)<\/code> above your code (but after your imports) then disable shuffling and the losses should be identical.<\/p>\n\n<p>For example, the <code>tensorlayers<\/code> mini-batch generator allows you to set the shuffle argument to False.<\/p>\n\n<pre><code>import numpy as np\nimport tensorlayers as tl\n\nX = np.asarray([['a','a'], ['b','b'], ['c','c'], ['d','d'], ['e','e'], ['f','f']])\ny = np.asarray([0,1,2,3,4,5])\n\nfor batch in tl.iterate.minibatches(inputs=X, targets=y, batch_size=16, shuffle=False):\n    print(batch)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53114882,"AnswerCount":2,"Tags":"<computer-vision><conv-neural-network><pytorch><vgg-net>","CreationDate":"2018-11-02T08:15:10.490","AcceptedAnswerId":53171862.0,"OwnerUserId":4777141.0,"Title":"Pytorch: Modifying VGG16 Architecture","Body":"<p>I'm currently trying to modify the VGG16 network architecture so that it's able to accept 400x400 px images. <\/p>\n\n<p>Based on literature that I've read, the way to do it would be to covert the fully connected (FC) layers into convolutional (CONV) layers. This would essentially \" allow the network to efficiently \u201cslide\u201d across a larger input image and make multiple evaluations of different parts of the image, incorporating all available contextual information.\" Afterwards, an Average Pooling layer is used to \"average the multiple feature vectors into a single feature vector that summarizes the input image\". <\/p>\n\n<p>I've done this <a href=\"https:\/\/stackoverflow.com\/a\/52981440\/4777141\">using this function<\/a>, and have come up with the following network architecture: <\/p>\n\n<pre><code>----------------------------------------------------------------\n        Layer (type)               Output Shape         Param #\n================================================================\n            Conv2d-1         [-1, 64, 400, 400]           1,792\n              ReLU-2         [-1, 64, 400, 400]               0\n            Conv2d-3         [-1, 64, 400, 400]          36,928\n              ReLU-4         [-1, 64, 400, 400]               0\n         MaxPool2d-5         [-1, 64, 200, 200]               0\n            Conv2d-6        [-1, 128, 200, 200]          73,856\n              ReLU-7        [-1, 128, 200, 200]               0\n            Conv2d-8        [-1, 128, 200, 200]         147,584\n              ReLU-9        [-1, 128, 200, 200]               0\n        MaxPool2d-10        [-1, 128, 100, 100]               0\n           Conv2d-11        [-1, 256, 100, 100]         295,168\n             ReLU-12        [-1, 256, 100, 100]               0\n           Conv2d-13        [-1, 256, 100, 100]         590,080\n             ReLU-14        [-1, 256, 100, 100]               0\n           Conv2d-15        [-1, 256, 100, 100]         590,080\n             ReLU-16        [-1, 256, 100, 100]               0\n        MaxPool2d-17          [-1, 256, 50, 50]               0\n           Conv2d-18          [-1, 512, 50, 50]       1,180,160\n             ReLU-19          [-1, 512, 50, 50]               0\n           Conv2d-20          [-1, 512, 50, 50]       2,359,808\n             ReLU-21          [-1, 512, 50, 50]               0\n           Conv2d-22          [-1, 512, 50, 50]       2,359,808\n             ReLU-23          [-1, 512, 50, 50]               0\n        MaxPool2d-24          [-1, 512, 25, 25]               0\n           Conv2d-25          [-1, 512, 25, 25]       2,359,808\n             ReLU-26          [-1, 512, 25, 25]               0\n           Conv2d-27          [-1, 512, 25, 25]       2,359,808\n             ReLU-28          [-1, 512, 25, 25]               0\n           Conv2d-29          [-1, 512, 25, 25]       2,359,808\n             ReLU-30          [-1, 512, 25, 25]               0\n        MaxPool2d-31          [-1, 512, 12, 12]               0\n           Conv2d-32           [-1, 4096, 1, 1]     301,993,984\n             ReLU-33           [-1, 4096, 1, 1]               0\n          Dropout-34           [-1, 4096, 1, 1]               0\n           Conv2d-35           [-1, 4096, 1, 1]      16,781,312\n             ReLU-36           [-1, 4096, 1, 1]               0\n          Dropout-37           [-1, 4096, 1, 1]               0\n           Conv2d-38              [-1, 3, 1, 1]          12,291\nAdaptiveAvgPool2d-39              [-1, 3, 1, 1]               0\n          Softmax-40              [-1, 3, 1, 1]               0\n================================================================\nTotal params: 333,502,275\nTrainable params: 318,787,587\nNon-trainable params: 14,714,688\n----------------------------------------------------------------\nInput size (MB): 1.83\nForward\/backward pass size (MB): 696.55\nParams size (MB): 1272.21\nEstimated Total Size (MB): 1970.59\n----------------------------------------------------------------\n<\/code><\/pre>\n\n<p>My question is simple: Is the use of the average pooling layer at the end necessary? It seems like by the last convolutional layer, we get a 1x1 image with 3 channels. Doing an average pooling on that would seem to not have any effect. <\/p>\n\n<p>If there is anything amiss in my logic\/ architecture, kindly feel free to point it out. \nThanks!<\/p>\n","answers":[{"AnswerId":"53134070","CreationDate":"2018-11-03T18:00:32.747","ParentId":null,"OwnerUserId":"2353472","Title":null,"Body":"<p>Purpose of <code>AdaptiveAvgPool2d<\/code> is to make the convnet work on input of any arbitrary size (and produce an output of fixed size). In your case, since input size is fixed to 400x400, you probably do not need it.<\/p>\n\n<p>I think this paper might give you a better idea of this method - <a href=\"https:\/\/arxiv.org\/pdf\/1406.4729v3.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1406.4729v3.pdf<\/a><\/p>\n"},{"AnswerId":"53171862","CreationDate":"2018-11-06T12:25:37.103","ParentId":null,"OwnerUserId":"5373738","Title":null,"Body":"<blockquote>\n  <p>How to convert VGG to except input size of 400 x 400 ?<\/p>\n<\/blockquote>\n\n<p><strong>First Approach<\/strong><\/p>\n\n<p>The problem with <code>VGG<\/code> style architecture is we are hardcoding the number of input &amp; output features in our Linear Layers.\ni.e <\/p>\n\n<pre><code>vgg.classifier[0]: Linear(in_features=25088, out_features=4096, bias=True)\n<\/code><\/pre>\n\n<p>It is expecting 25,088 input features.<\/p>\n\n<p>If we pass an image of size <code>(3, 224, 224)<\/code>\nthrough <code>vgg.features<\/code> the output feature map will be of dimensions:<\/p>\n\n<pre><code>(512, 7, 7) =&gt; 512 * 7 * 7 =&gt; 25,088\n<\/code><\/pre>\n\n<p>If we change the input image size to <code>(3, 400, 400)<\/code> and pass\nthrough <code>vgg.features<\/code> the output feature map will be of dimensions:<\/p>\n\n<pre><code>(512, 12, 12) =&gt; 512 * 12 * 12 =&gt;  73,728\n\nthrows `sizemismatch` error.\n<\/code><\/pre>\n\n<p>One way to fix this issue is by using <code>nn.AdaptiveAvgPool<\/code> in place of <code>nn.AvgPool<\/code>. AdaptiveAvgPool helps to define the output size of the layer which remains constant irrespective of the size of the input through the <code>vgg.features<\/code> layer.<\/p>\n\n<p>for eg:<\/p>\n\n<pre><code>vgg.features[30] = nn.AdaptiveAvgPool(output_size=(7,7))\n\nwill make sure the final feature maps have a dimension of `(512, 7, 7)` \nirrespective of the input size.\n<\/code><\/pre>\n\n<p>You can read more about Adaptive Pooling in <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.AdaptiveAvgPool2d\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p><strong>Second Approach<\/strong><\/p>\n\n<p>If you use the technique <a href=\"https:\/\/stackoverflow.com\/a\/52981440\/5373738\">here<\/a> to convert your Linear layers to Convolutional Layers, you don't have to worry about the input dimension, however you have to change the weight initialisation techniques because of the change in number of parameters.  <\/p>\n\n<blockquote>\n  <p>Is the use of the average pooling layer at the end necessary?<\/p>\n<\/blockquote>\n\n<p>No, in this case. It is not changing the size of the input feature map, hence it not doing an average over a set of nodes. <\/p>\n"}]}
+{"QuestionId":53115876,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-02T09:29:17.227","AcceptedAnswerId":null,"OwnerUserId":8764887.0,"Title":"Alternative method to handle large patches generated from image dataset using tensor-flow","Body":"<p>Step 1:\nI have a single tf-record which contains 4644 large images. \nI have to decode each image from the tf-record, \nconvert them to patches of size 50 <\/p>\n\n<p>Step 2:\nOnce I create patches for all 4644 images. I have to write the created patches back to another tf-record which i will use it later for training the network.<\/p>\n\n<p>I don't know how i can do the step 1 and 2 efficiently. i have written code for step 1 and 2 but the system hangs since i have total patches of size [369969,50,50,3] generated from 4644 images.<\/p>\n\n<p>This is the code I have written<\/p>\n\n<pre><code>For decoding each image and convert to patches of size 50x50x3\n\ndataset = (tf.data.TFRecordDataset('Image.tfrecords')\n           .map(read_and_decode, num_parallel_calls=num_parallel_calls )  # step 2\n           .map( get_patches_fn, num_parallel_calls=num_parallel_calls )  # step 3\n           .apply( tf.data.experimental.unbatch())  # unbatch the patches we just produced\n           .prefetch( 1 )  # step 6: make sure you always have one batch ready to serve\n           )\n\niterator = dataset.make_one_shot_iterator()\npatches_image = iterator.get_next()\n\ninit=tf.global_variables_initializer()\n\ntemp_patches_image=[]\nimage_batch_size=369969\nwith tf.Session() as sess:\n    sess.run(init)\n    for i in range(0,image_batch_size):\n      res=sess.run(patches_image)\n      temp_patches_image.append(res)\n    temp_patches_image=tf.stack(temp_patches_image)-------&gt; This step \n                                                            causes my \n                                                            system to \n                                                            hang\n    final_patches_image=sess.run(temp_patches_image)\n\n\nStep 2:\n For writing the patches created to another tf-record\n\n create_data_record_patch(final_patches_image,'patches.tfrecord')\n\n\n    def create_data_record_patch(patches, tfrecords_filename ):\n        # Function to write all the filenames to a tf-record\n        #   Input: file_names - list of filenames to be converted to tf record\n        #          tf-records_filename- name of tf-file-record\n\n        writer = tf.python_io.TFRecordWriter( tfrecords_filename )\n        total_batch_size=0\n        for i in range(patches.shape[0]):\n\n            height = patches.shape[1]\n            width = patches.shape[2]\n\n            img_raw = patches[i,:,:,:].tostring()\n            example = tf.train.Example( features=tf.train.Features( feature={\n                'height': _int64_feature( height ),\n                'width': _int64_feature(width ),\n                'image_raw': _bytes_feature(img_raw)}))\n\n            writer.write( example.SerializeToString() )\n\n        writer.close()\n        return total_batch_size\n<\/code><\/pre>\n\n<p>Additional information: The above code works when i have a tf record of 50 images(generated ~5000 patches) instead of 4644 images. I know tf.stack is causing problem since it has to process all the 369969 patches, but I don't know any alternative method to create a single tf-record for all patches generated.<\/p>\n\n<p>Any help is appreciated.<\/p>\n\n<p>Thanks in Advance.<\/p>\n","answers":[]}
+{"QuestionId":53116036,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-11-02T09:39:25.287","AcceptedAnswerId":null,"OwnerUserId":987397.0,"Title":"Tensorflow prefetching to multiple GPUs","Body":"<p>When having a Tensorflow dataset and multiple GPUs it is beneficial to preload data to the multiple GPUs.<\/p>\n\n<p>The only example I could find is in this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/13610#issuecomment-411893139\" rel=\"nofollow noreferrer\">issue<\/a><\/p>\n\n<pre><code>host_dataset_on_cpu = .....\nmulti_device_iterator = prefetching_ops.MultiDeviceIterator(host_dataset_on_cpu, devices=['\/gpu:0', '\/gpu:1'])\nelem_on_gpu_0, elem_on_gpu_1 = multi_device_iterator.get_next()\n\nwith tf.Session() as sess:\n    sess.run(multi_device_iterator.initializer)\n    elem_0 = sess.run(elem_on_gpu_0)\n    elem_1 = sess.run(elem_on_gpu_1)\n<\/code><\/pre>\n\n<p>What I am wondering is:\n- is it really necessary to have multiple calls to sess.run (one for each GPU?)\n- Can this work with a configurable number of GPUs? (splits = multi_device_iterator.get_next())\n- It seems like this MultiDeviceIterator is implemented, but can not be found in the documentation. Is it still under development?<\/p>\n\n<p>Also when I try to implement this I get a lot of errors<\/p>\n\n<blockquote>\n  <p>2018-11-02 09:22:15.740637: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.741142: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.741670: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.742512: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.748161: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.749587: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.752334: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.752756: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.753581: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.753974: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported object type list 2018-11-02 09:22:15.754434: W\n  tensorflow\/core\/framework\/op_kernel.cc:1261] Unimplemented:\n  Unsupported<\/p>\n<\/blockquote>\n\n<p>which seem not really helpful<\/p>\n","answers":[]}
+{"QuestionId":53116200,"AnswerCount":0,"Tags":"<python><tensorflow><keras><raspberry-pi3>","CreationDate":"2018-11-02T09:49:35.093","AcceptedAnswerId":null,"OwnerUserId":6504342.0,"Title":"Raspberry Pi Tensorflow\/Keras Error: Allocation of 67108864 exceeds 10% of system memory","Body":"<p>Hi I am having difficulties running a ML model based on the above on Pi 3\nThe code is:<\/p>\n\n<pre><code>import pandas as pd\n\nimport tensorflow as tf\n# deep learning library. Tensors are just multi-dimensional arrays\nimport numpy as np\n\n\nimport time\nimport skimage\nfrom PIL import Image\nimport numpy as np\nimport sys\nimport os\nimport csv\n\n# Import the `transform` module from `skimage`\nfrom skimage import transform \n\nfrom skimage import data\n\nimport glob\nfrom os import walk\n\n\nfileLists = []\nimage28 = []\nvalue = []\n\nformat='.jpg'\n\npath = '\/testpic'\nfolder = os.fsencode(path)\n\n\nfor root, directories, filenames in os.walk(path):\n\n\n  for filename in filenames: \n    print (os.path.join(root,filename)) \n    if filename.endswith(('.jpg', '.png', '.gif')):\n\n        fileLists.append(skimage.data.imread(os.path.join(root,filename)))\n\n\n   time1 = time.time()        \n   images28  = [transform.resize(fileList, (227, 227)) for fileList in \n   fileLists]\n\n   train_x = np.array(images28)\n\n   new_model = tf.keras.models.load_model('ml.h5')\n\n   predictions = new_model.predict(train_x)\n\n\n   for i in range(10):\n\ntime1 = time.time()\n\nprint(np.argmax(predictions[i]))\n\ntime2 = time.time()\n\nprint((time2-time1))\n<\/code><\/pre>\n\n<p>When I run it, I get the following error:<\/p>\n\n<blockquote>\n  <p>Using TensorFlow backend.\n  2018-11-02 09:43:34.904211: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:43:34.937663: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:43:42.740079: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:43:43.125458: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:43:43.720775: W tensorflow\/core\/framework\/allocator.cc:122] Allocation of 67108864 exceeds 10% of system memory.\n  2018-11-02 09:43:44.152943: W tensorflow\/core\/framework\/allocator.cc:122] Allocation of 67108864 exceeds 10% of system memory.\n  2018-11-02 09:43:44.298756: W tensorflow\/core\/framework\/allocator.cc:122] Allocation of 67108864 exceeds 10% of system memory.\n  2018-11-02 09:43:44.299426: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:44:08.801283: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:44:09.582225: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph\n  2018-11-02 09:44:09.788274: W tensorflow\/core\/framework\/allocator.cc:122] Allocation of 67108864 exceeds 10% of system memory.\n  2018-11-02 09:44:10.183454: W tensorflow\/core\/framework\/allocator.cc:122] Allocation of 67108864 exceeds 10% of system memory.\n  2018-11-02 09:44:28.074020: I tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:344] Starting optimization for grappler item: tf_graph<\/p>\n<\/blockquote>\n\n<p>After which the program is killed.<\/p>\n\n<p>I have changed the batch size to 1 during my training process, but nothing has been fixed.<\/p>\n\n<p>Please help.<\/p>\n","answers":[]}
+{"QuestionId":53116290,"AnswerCount":1,"Tags":"<python-3.x><keras><pycharm><lstm>","CreationDate":"2018-11-02T09:55:33.457","AcceptedAnswerId":null,"OwnerUserId":8107567.0,"Title":"Training and evaluating accuracies different in keras LSTM model","Body":"<p>I am training a LSTM model using keras. Training the training set gives accuracy:83% in the last epoch, but evaluating or predicting the model with the same training set gives accuracy:47%. I cant figure out the problem. \nAlso, I have commented out the training part in the following code, since i save the model &amp; weights on disk, but I used it before to train i.e. model.fit(...) etc.  <\/p>\n\n<pre><code>import pandas as pd\nimport Preprocessing as pre\nfrom sklearn.feature_extraction.text import CountVectorizer\nfrom sklearn.utils import shuffle\nimport pickle\nimport numpy as np\nimport sys\nfrom keras.models import Sequential, load_model\nfrom keras.layers import Dense, Dropout, Activation\nfrom keras.layers import Embedding\nfrom keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\nfrom keras.layers import LSTM\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras.models import model_from_json\nfrom keras.preprocessing.text import Tokenizer\n\ndf = pd.read_csv('tweets.csv',header=None,encoding = \"ISO-8859-1\")\ndf=shuffle(df)\nlength=df.size\n\ntrain=[]\ntest=[]\nY=[]\nY2=[]\n\ncount=450000\nfor a in range(450000):   #loading data\n    # b=pre.preprocess_tweet(df[1][a])\n    label=df[0][a]\n    # train.append(b)\n    Y.append(label)\n    count-=1\n    print(\"Loading training data...\",  count)\n\n# with open('training_data_no_stopwords.obj', 'wb') as fp:\n#     pickle.dump(train, fp)\nwith open ('training_data.obj', 'rb') as fp:\n    train = pickle.load(fp)\n\ncount=156884\nfor a in range(450000,606884):   #loading testin data\n    # b = pre.preprocess_tweet(df[1][a])\n    label=df[0][a]\n    # test.append(b)\n    Y2.append(label)\n    count-=1\n    print(\"Loading testing data...\",  count)\n\n# with open('testing_data_no_stopwords.obj', 'wb') as fp:\n#     pickle.dump(test, fp)\n\nwith open ('testing_data.obj', 'rb') as fp:\n    test = pickle.load(fp)\n\n# vectorizer = CountVectorizer(analyzer = \"word\",tokenizer = None, preprocessor = None, stop_words = None, max_features = 3500)\n# # # fit_transform() does two functions: First, it fits the model\n# # # and learns the vocabulary; second, it transforms our training data\n# # # into feature vectors. The input to fit_transform should be a list of\n# # # strings.\n#\n# train = vectorizer.fit_transform(train)\n# test = vectorizer.transform(test)\ntokenizer = Tokenizer(split=' ')\ntokenizer.fit_on_texts(train)\ntrain = tokenizer.texts_to_sequences(train)\nmax_words = 134\ntrain = pad_sequences(train, maxlen=max_words)\ntokenizer.fit_on_texts(test)\ntest = tokenizer.texts_to_sequences(test)\ntest = pad_sequences(test, maxlen=max_words)\n\nprint ('Extracting features &amp; training batches')\n\nprint(\"Training...\")\nembedding_size=32\nmodel = Sequential()\nmodel.add(Embedding(606884, 32, input_length=134))\nmodel.add(Dropout(0.4))\nmodel.add(LSTM(128))\nmodel.add(Dense(64))\nmodel.add(Dropout(0.5))\nmodel.add(Activation('relu'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\nprint(model.summary())\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# batch_size = 501\n# num_epochs = 3\n#\n# model.fit(train, np.array(Y),  batch_size=batch_size, epochs=num_epochs ,validation_split=0.2,shuffle=True)\n\n# # Save the weights\n# model.save_weights('LSTM_model_weights.h5')\n#\n# # Save the model architecture\n# with open('LSTM_model.json', 'w') as f:\n#     f.write(model.to_json())\n\n# Model reconstruction from JSON file\nwith open('LSTM_model.json', 'r') as f:\n    model = model_from_json(f.read())\n\n# Load weights into the new model\nmodel.load_weights('LSTM_model_weights.h5')\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nscores = model.evaluate(train, np.array(Y))\nprint('Test accuracy:', scores[1])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53117962","CreationDate":"2018-11-02T11:43:07.603","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Maybe try training with <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.StratifiedKFold.html\" rel=\"nofollow noreferrer\">Stratified K-Fold<\/a>. This probably happens because your test data is very different from training data. Also, you can try shuffling the data before train-test splitting.<\/p>\n"}]}
+{"QuestionId":53116477,"AnswerCount":1,"Tags":"<python><pytorch><tensor>","CreationDate":"2018-11-02T10:07:39.140","AcceptedAnswerId":null,"OwnerUserId":9066532.0,"Title":"Pytorch: is there a function similar to torch.argmax which can really keep the dimension of the original data?","Body":"<p>For example, \nthe code is<\/p>\n\n<pre><code>input = torch.randn(3, 10)\nresult = torch.argmax(input, dim=0, keepdim=True)\n<\/code><\/pre>\n\n<p><code>input<\/code> is<\/p>\n\n<pre><code>tensor([[ 1.5742,  0.8183, -2.3005, -1.1650, -0.2451],\n       [ 1.0553,  0.6021, -0.4938, -1.5379, -1.2054],\n       [-0.1728,  0.8372, -1.9181, -0.9110,  0.2422]])\n<\/code><\/pre>\n\n<p>and <code>result<\/code> is<\/p>\n\n<pre><code>tensor([[ 0,  2,  1,  2,  2]])\n<\/code><\/pre>\n\n<p>However, I want a result like this<\/p>\n\n<pre><code>tensor([[ 1,  0,  0,  0,  0],\n        [ 0,  0,  1,  0,  0],\n        [ 0,  1,  0,  1,  1]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53187157","CreationDate":"2018-11-07T10:00:02.670","ParentId":null,"OwnerUserId":"9066532","Title":null,"Body":"<p>Finally, I solved it. But this solution may not be efficient.\nCode as follow,<\/p>\n\n<pre><code>input = torch.randn(3, 10)\nresult = torch.argmax(input, dim=0, keepdim=True)\nresult_0 = result == 0\nresult_1 = result == 1\nresult_2 = result == 2\nresult = torch.cat((result_0, result_1, result_2), 0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53116808,"AnswerCount":0,"Tags":"<apache-spark><keras><pyspark><autoencoder><systemml>","CreationDate":"2018-11-02T10:28:39.637","AcceptedAnswerId":null,"OwnerUserId":4230234.0,"Title":"How to train a neural network autoencoder (Keras) on Spark dataframe","Body":"<p>I created a very large <strong>Spark Dataframe<\/strong> with PySpark on my cluster, which is too big to fit into memory. I also have an <strong>autoencoder model with Keras<\/strong>, which takes in a Pandas dataframe (in-memory object). <\/p>\n\n<p>What is the best way to bring those two worlds together? <\/p>\n\n<p>I found some libraries that provide Deep Learning on Spark, but is seems only for <a href=\"https:\/\/databricks.com\/blog\/2016\/01\/25\/deep-learning-with-apache-spark-and-tensorflow.html\" rel=\"nofollow noreferrer\">hyper parameter tuning<\/a> or wont support autoencoders like <a href=\"https:\/\/systemml.apache.org\/\" rel=\"nofollow noreferrer\">Apache SystemML<\/a><\/p>\n\n<p>I am surely not the first one to train a NN on Spark Dataframes. I have a conceptual gap here, please help!<\/p>\n","answers":[]}
+{"QuestionId":53116974,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><conv-neural-network>","CreationDate":"2018-11-02T10:38:42.693","AcceptedAnswerId":null,"OwnerUserId":9739661.0,"Title":"Tensorflow variable\/graph key missing while reloading the model in Python 3.6","Body":"<p>ERROR MESSAGE:<\/p>\n\n<p>NotFoundError: Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:<\/p>\n\n<pre><code>\"\"\" PLACEHOLDER \"\"\"\nx=tf.placeholder(tf.float32,shape=[None,784])\ny_true=tf.placeholder(tf.float32,shape=[None,4])\n\n\nWeight=tf.global_variables()\nbias=tf.global_variables()\n\n\"\"\" LAYERS:\"\"\"\n\n\"the image input layer 28*28 input\"\nx_image = tf.reshape(x,[-1,28,28,1])\n\n\"\"\"first convolution, using 6*6 filter, and then max pooling 2by2, final \noutput will have depth 32\nhere we are calculating 32 features\n\"\"\"\nconvo_1 = convolutional_layer(x_image,shape=[6,6,1,32])\nconvo_1_pooling = max_pool_2by2(convo_1)\n\n\"\"\"first convolution, using 6*6 filter, and then max pooling 2by2, final \noutput will have 64\nfeatures hence depth 64\n\"\"\"\nconvo_2 = convolutional_layer(convo_1_pooling,shape=[6,6,32,64])\nconvo_2_pooling = max_pool_2by2(convo_2)\n\n\"flattening the output of the last pooling layer to fuly connect it\"\nconvo_2_flat = tf.reshape(convo_2_pooling,[-1,7*7*64])\nWt,bs=normal_full_layer(convo_2_flat,1024)#1024 nodes in the final layer\n\nfull_layer_one = tf.nn.relu(tf.matmul(convo_2_flat,Wt)+bs)\n\n# NOTE THE PLACEHOLDER HERE!\nhold_prob = tf.placeholder(tf.float32)\nfull_one_dropout = tf.nn.dropout(full_layer_one,keep_prob=hold_prob) \n\nWeight, bias= normal_full_layer(full_one_dropout,2)\n\ny_pred=tf.matmul(full_one_dropout,Weight)+bias\n\n\"\"\"INITIALIZE VARIABLES\"\"\"\ninit = tf.global_variables_initializer()\n\n\"\"\"Add ops to save and restore all the variables\"\"\"\nsaver = tf.train.Saver()\n\n\"\"\"Later, launch the model, use the saver to restore variables from disk, \nand\n# do some work with the model\"\"\"\n\nwith tf.Session() as sess:\n   # Restore variables from disk.\n   saver.restore(sess,SAVE_PATH)\n   print(\"Model restored.\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"55177370","CreationDate":"2019-03-15T07:12:09.533","ParentId":null,"OwnerUserId":"11207171","Title":null,"Body":"<p>If you use the spyder, please Restart kernel! It can work well.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/R7w58.png\" alt=\"https:\/\/i.stack.imgur.com\/R7w58.png\"><\/p>\n"}]}
+{"QuestionId":53117437,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2018-11-02T11:07:16.260","AcceptedAnswerId":null,"OwnerUserId":4841832.0,"Title":"Use `tf.image.resize_image_with_crop_or_pad` to resize numpy array","Body":"<p>I want to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_image_with_crop_or_pad\" rel=\"nofollow noreferrer\"><code>tf.image.resize_image_with_crop_or_pad<\/code><\/a> on a Numpy array of shape <code>(100,100,2)<\/code> to get it cropped or padded to a target shape <code>(h,w,2)<\/code>.<\/p>\n\n<p>However, when I do:<\/p>\n\n<pre><code>img = resize_image_with_crop_or_pad(img, target_height, target_width)\nimg = np.array(img)\n<\/code><\/pre>\n\n<p><code>img.shape<\/code> evaluates to <code>()<\/code>, which is not what I expected. How do I turn the output of this function into a properly shaped numpy array?<\/p>\n","answers":[{"AnswerId":"53278282","CreationDate":"2018-11-13T09:54:59.083","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<pre><code>img = resize_image_with_crop_or_pad(img_tensor, target_height, target_width)\nwith tf.Session as sess:\n    img_output = sess.run(img)\n<\/code><\/pre>\n\n<p>Now <code>img_output<\/code> is a numpy array, but note that img has to be a <code>tf.Tensor<\/code> of shape <code>[1, height, width, channels]<\/code> so you might do this beforehand, suggesting your input image is already a numpy array:<\/p>\n\n<pre><code>img_input = np.expand_dims(img_input, 0)\nimg_tensor = tf.convert_to_tensor(img_input)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53117593,"AnswerCount":0,"Tags":"<tensorflow><go><equivalent>","CreationDate":"2018-11-02T11:17:41.730","AcceptedAnswerId":null,"OwnerUserId":4780378.0,"Title":"tensorflow retrain.py equivalent golang","Body":"<p>I would like to know if there is any equivalent of <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">retrain.py<\/a> in golang.<\/p>\n\n<p>Thanks for help<\/p>\n","answers":[]}
+{"QuestionId":53117786,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-11-02T11:30:45.297","AcceptedAnswerId":null,"OwnerUserId":680074.0,"Title":"Keras - How to convert LSTM code into CNN","Body":"<p>I am working on binary classification sentiment analysis either positive or negative my lstm code is working fine but i am converting my lstm code into cnn  having <strong>Value error of \"input_length\" is 30, but received input has shape (None, 1)<\/strong><\/p>\n\n<p><strong>my input shape is (30,1) my batch size is 24 in lstm<\/strong><\/p>\n\n<pre><code>model.add(Embedding(30,30,input_length=30))\nmodel.add(Conv1D(padding='valid',activation='relu',strides=1))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(30))\nmodel.add(Dropout(0.2))\nmodel.add(Activation('relu'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])\nmodel.fit(inputBatch, ponlabel,batch_size=24,epochs=20,validation_data=(inputBatch, ponlabel))\n<\/code><\/pre>\n\n<p>for the sake of reference I have put my LSTM Code.<\/p>\n\n<pre><code>model.add(LSTM(100, input_shape=(30, 1)))\nmodel.add(Dense(30, activation=\"relu\"))\nmodel.add(Dense(1, activation=\"sigmoid\"))\nmodel.compile(loss='mean_absolute_error', optimizer='adam',metrics=[\"accuracy\"])\nmodel.fit(inputBatch, ponlabel,\n          batch_size=24, epochs=20, verbose=1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53117869","CreationDate":"2018-11-02T11:35:50.463","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You are using <code>Convolution2D<\/code> but your data over timesteps is 1 dimensional. So you need to use <code>Convolutional1D<\/code> to convolve over the tokens in your sentence. There is a CNN text classification in the Keras examples folder <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/imdb_cnn.py\" rel=\"nofollow noreferrer\">imdb_cnn.py<\/a>.<\/p>\n"}]}
+{"QuestionId":53118012,"AnswerCount":1,"Tags":"<python><keras><spyder>","CreationDate":"2018-11-02T11:46:20.350","AcceptedAnswerId":53119212.0,"OwnerUserId":9110576.0,"Title":"How to restore core.py file from keras.layers that I accidentally changed","Body":"<p>I was working with keras in Spyder and I accidentally changed the core.py file and compiled it. Of course, some errors occurred. Then I reversed the changes but for some odd reasons my actual code that uses the file is giving the following error.<\/p>\n\n<pre><code>File \"&lt;ipython-input-18-18c6a7ab4e93&gt;\", line 1, in &lt;module&gt;\nrunfile('F:\/Python Codes\/HFE\/HCM_LSTM_LDA_1.py', wdir='F:\/Python Codes\/HFE')\n\nFile \"e:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\spyder_kernels\\customize\\spydercustomize.py\", line 668, in runfile\nexecfile(filename, namespace)\n\nFile \"e:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\spyder_kernels\\customize\\spydercustomize.py\", line 108, in execfile\nexec(compile(f.read(), filename, 'exec'), namespace)\n\nFile \"F:\/Python Codes\/HFE\/HCM_LSTM_LDA_1.py\", line 195, in &lt;module&gt;\nclassifier.add(Dense(1))\n\nFile \"e:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\nreturn func(*args, **kwargs)\n\nFile \"e:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\layers\\core.py\", line \n844, in __init__\nsuper(Dense, self).__init__(**kwargs)\n\nTypeError: super(type, obj): obj must be an instance or subtype of type\n<\/code><\/pre>\n\n<p>Now I know there is no problem in my code which is:<\/p>\n\n<pre><code>#...\nfrom keras.layers import Dropout\n#...\nclassifier = Sequential()\nclassifier.add(LSTM(units = 100, activation = 'relu'))\nclassifier.add(Dense(1))\nclassifier.add(Dropout(0.0001));                                                                        \nclassifier.add(Dense(1))                                                                                 \nclassifier.compile(optimizer = 'Nadam', loss ='mean_absolute_percentage_error', metrics = ['accuracy']) \nclassifier.fit(Feature_train, Label_train, epochs=100, batch_size=168, \nverbose=2)                         #Fit classifier\nRNN_Predict=classifier.predict(Feature_test) \n<\/code><\/pre>\n\n<p>And I reversed the changes i made to core.py file but i want to tackle this problem without having to uninstall and reinstall everything<\/p>\n","answers":[{"AnswerId":"53119212","CreationDate":"2018-11-02T13:10:00.577","ParentId":null,"OwnerUserId":"9110576","Title":null,"Body":"<p>So After i replaced the code of my core.py file with the code from github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/core.py it didn't work at first but then i restated spyder and it worked.<\/p>\n"}]}
+{"QuestionId":53118589,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-02T12:25:00.780","AcceptedAnswerId":53118754.0,"OwnerUserId":10596177.0,"Title":"Extracting values from a tensor in tensorflow","Body":"<p>If there are two tensor matrices<\/p>\n\n<pre><code>a = [[1 2 3 4][5 6 7 8]]\nb = [[0 1][1 2]],\n<\/code><\/pre>\n\n<p>how can we get this:<\/p>\n\n<pre><code>c = [[1 2][6 7]]\n<\/code><\/pre>\n\n<p>i.e. from first row extracting column 0 and 1, from second row extracting column 1 and 2.<\/p>\n","answers":[{"AnswerId":"53118754","CreationDate":"2018-11-02T12:36:46.527","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a way to do that:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\na = tf.constant([[1, 2, 3, 4],\n                 [5, 6, 7, 8]])\nb = tf.constant([[0, 1],\n                 [1, 2]])\nrow = tf.range(tf.shape(a)[0])\nrow = tf.tile(row[:, tf.newaxis], (1, tf.shape(b)[1]))\nidx = tf.stack([row, b], axis=-1)\nc = tf.gather_nd(a, idx)\nwith tf.Session() as sess:\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[1 2]\n [6 7]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53119402,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2018-11-02T13:24:03.460","AcceptedAnswerId":53119543.0,"OwnerUserId":3806397.0,"Title":"Autoencoder Decoded Output","Body":"<p>I am trying to build an AutoEncoder, where I am trying to de-noise the signal.<\/p>\n\n<p>Now, for example, the amplitude range of my input signal varies in between -47 to +32. But while I am getting the decoded signal (reconstructed one), that signal only ranges in between 0 to +1 amplitude. <\/p>\n\n<p>How can I get my reconstructed signal with the same amplitude range of  -47 to +32?<\/p>\n","answers":[{"AnswerId":"53119543","CreationDate":"2018-11-02T13:32:11.493","ParentId":null,"OwnerUserId":"7678562","Title":null,"Body":"<p>I guess the question is whether your returned signal is a faithful representation of the input signal (but it's just constrained to the range 0 to 1)?<\/p>\n\n<p>If so, you could simply multiply it by 79, and then subtract 47.<\/p>\n\n<p>We'd need to see code if it's more than just a scaling issue.<\/p>\n"}]}
+{"QuestionId":53119432,"AnswerCount":4,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-02T13:25:32.530","AcceptedAnswerId":53120279.0,"OwnerUserId":680074.0,"Title":"Keras - CNN input shape incompatible","Body":"<p>I am working on binary classification my code is working fine on <strong>Keras Lstm<\/strong> when I am working with CNN I am getting input shape incompatibility error.<\/p>\n\n<p>this is the value error i am getting <\/p>\n\n<p><strong>ValueError: Error when checking target: expected dense_61 to have 3 dimensions, but got array with shape (24, 1)<\/strong><\/p>\n\n<p>this is my cnn code using keras<\/p>\n\n<pre><code>model=Sequential()\ninputBatch = inputBatch.reshape(24,30, 1)\nmodel.add(Conv1D(64, 3, activation='relu', input_shape=(30, 1)))\nmodel.add(Conv1D(64, 3, activation='relu'))\nmodel.add(MaxPooling1D(pool_size=4,strides=None, padding='valid'))\nmodel.add(Conv1D(128, 3, activation='relu'))\nmodel.add(Conv1D(128, 3, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])\nmodel.fit(inputBatch,ponlabel,batch_size=24,epochs=20,validation_data=(inputBatch, ponlabel))\n<\/code><\/pre>\n\n<p><strong>I am working on binary classification either it will be positive or either negative<\/strong><\/p>\n\n<p>for reference this is my lstm code<\/p>\n\n<pre><code>inputBatch =inputBatch.reshape(24,30,1)\nmodel=Sequential()\nmodel.add(LSTM(50, input_shape=(30, 1)))\nmodel.add(Dense(1, activation=\"relu\"))\nmodel.compile(loss='mean_absolute_error',optimizer='adam')\nmodel.fit(inputBatch,ponlabel,batch_size=24,epochs=100,verbose=1)\n<\/code><\/pre>\n\n<p><strong>inputBatch is something like this it is working on LSTM code but not working on CNN this is the input which I have used for training on both codes separately<\/strong><\/p>\n\n<pre><code>[[    0.  1288.  1288.  2214. 11266.  6923.   420.     0.     0.  8123.\n      0.  7619.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.     0.     0.     0.     0.     0. 11516.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  9929. 11501.  6573. 11266.  7566.  9963.  4420. 10936.  3657.\n   7050.     0.   408. 11501.  9988.  9963.  8455.  2879.  9322.  2047.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 11956.  5222.     0.     0. 12106.  6481.     0.  7093. 13756.\n  12152.     0.     0.     0.     0. 10173.     0.  5173. 13756.  9371.\n      0.  9956.     0.     0.  9716.     0.     0.     0.     0.     0.]\n [    0.     0.   420.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 11501.  1916.  2073. 10936.  6312.     0. 10193. 10322.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  2879.  7852. 11501.  1934.   286. 11483.     0. 12004. 11118.\n      0. 12007.  9917. 12111.  1520. 10364.     0.  8840.  4195.  2910.\n  10773. 11386. 12117.  9321.     0.     0.     0.     0.     0.     0.]\n [    0.  7885.  7171.  1034. 11501.  3103.  5842.  4395. 11871.  3328.\n   6719.  5407.  1087.  8935.  2937.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  8894.   450. 11516.  7353. 11501. 11502. 11499.     0.  1319.\n  11693. 11501.  5735. 12111.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  1087.  9565.    23.     0.  3045.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  5015. 11501.  3306. 12111.  9307.  5050. 11501.  3306.     0.\n   3306. 12111.  1981. 11516.   615. 11516.     0.  3925. 11956.  9371.\n   9013.  4395. 12111.  5048.     0.  3925.     0.     0.     0.     0.]\n [    0.  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.   128.  2073. 10936.  6312.     0. 10193. 10322.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 10173.  9435.  1320.  9322. 12018.  1055.  8840.  6684. 12051.\n   2879.     0. 12018.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  1570.  5466.  9322.    34. 11480.  1356. 11270.   420.  2153.\n  12006.  5157.  8840.  1055. 11516.  7387.  2356.  2163.  2879.  5541.\n   9443.  7441.  1295.  5473.     0.     0.     0.     0.     0.     0.]\n [    0.  5014.     0.     0.  3651.  1087.    63.  6153.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0. 10608. 10855.  9562.     0.     0.     0.  4202.     0.     0.\n      0. 10818. 10818.  5842.     0.  9963.     0. 11516. 10464.  7491.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  5952.  6133.   450.  7520.  5842.  3412. 10400.  3412.  2149.\n   4891.  2979.  3456.   505.  9929. 11501.  9322.  1836. 11501. 12111.\n   3435. 11105. 11266.   420.  9322.    34.     0.     0.     0.     0.]\n [    0.  1570.  5466.  9322.    34. 11480.  1356. 11270.   420.  2153.\n  12006.  5157.  8840.  1055. 11516.  7387.  2356.  2163.  2879.  5541.\n   9443.  7441.  1295.  5473.     0.     0.     0.     0.     0.     0.]\n [    0.  7544.     0.  1709.   420. 10936.  5222.  5842. 10407.  6937.\n  11329.  2937.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  7785.  8840.     0.   420.  8603. 12003.  2879.  1087.  2356.\n   2390. 12111.     0.     0.     0.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  8695.  8744.   420.  8840.  6697.  9267. 11516. 11203.  2260.\n   8840.  7309.     0. 11100.  6041.     0.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.  9307. 12003.  2879.  6398.  9372.  4614.  5222.     0.     0.\n   2879. 10364.  6923.  4709.  4860. 11871.     0.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]\n [    0.     0.  2844.  1287.   420. 11501.   610. 11501.   596.     0.\n  12111.  3690.  6343.  9963.     0.     0.  8840.     0.     0.     0.\n      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53119621","CreationDate":"2018-11-02T13:37:49.700","ParentId":null,"OwnerUserId":"7358042","Title":null,"Body":"<p>Such an error may arise when input images are not of the same size.<\/p>\n\n<p>Add more information ( I could not comment to say that as I don't have enough rep), possibly including the full stack trace.  <\/p>\n"},{"AnswerId":"53120279","CreationDate":"2018-11-02T14:18:02.140","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The problem is the output shape, since you use an CNN, the output is 3D (samples, width, channels), and the Dense layer will operate on the last dimension, giving you a 3D output. But you want a 2D output, so you need to add a Flatten layer:<\/p>\n\n<pre><code>model=Sequential()\nmodel.add(Conv1D(64, 3, activation='relu', input_shape=(30, 1)))\nmodel.add(Conv1D(64, 3, activation='relu'))\nmodel.add(MaxPooling1D(pool_size=4,strides=None, padding='valid'))\nmodel.add(Conv1D(128, 3, activation='relu'))\nmodel.add(Conv1D(128, 3, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>You can compare the output shape of this model and your original model by doing <code>model.summary()<\/code><\/p>\n"},{"AnswerId":"53119929","CreationDate":"2018-11-02T13:57:22.573","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The incompatible input shape is due to <code>ponlabel<\/code>.  For the LSTM, it has shape (24,1).  But the CNN is using <code>binary_crossentropy<\/code> for the loss, so it would have two target classes.  This means for the CNN the <code>ponlabels<\/code> would need to have shape (24,2,1).<\/p>\n"},{"AnswerId":"53119961","CreationDate":"2018-11-02T13:59:48.283","ParentId":null,"OwnerUserId":"9137052","Title":null,"Body":"<p>You need to use either MSE or Categorical Cross entropy for your CNN loss<\/p>\n"}]}
+{"QuestionId":53119555,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-02T13:32:57.423","AcceptedAnswerId":null,"OwnerUserId":3055351.0,"Title":"How to initialize iterator over dataset when you pass it to keras","Body":"<p>I would like to pass an iterator over Tensorflow dataset to Keras, and I get an error that the iterator was not initialized. How should I do it properly?<\/p>\n\n<p>Here there is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nimport tensorflow.keras as keras\n\nK = keras.backend\n\n# Input parameters:\nbatch_size = 2\n\n# Get some data:\nnum_data_points = 100\nimages = np.random.normal(size=(num_data_points, 5, 5, 3)).astype(np.float32)\nmasks = np.random.normal(size=(num_data_points, 5, 5, 3)).astype(np.float32)\n\n# Number of batches:\nnum_batches = num_data_points \/\/ batch_size\nif num_batches * batch_size &lt; num_data_points:\n    num_batches += 1\n    assert num_batches * batch_size &gt; num_data_points\nelse:\n    assert num_batches * batch_size == num_data_points\n\n# Initialize model:\ngraph = tf.Graph()\nsess = tf.Session(graph=graph)\nK.set_session(session=sess)\n\nwith graph.as_default():\n\n    dataset = tf.data.Dataset.from_tensor_slices(tensors=(images, masks))\n    dataset = dataset.batch(batch_size=batch_size).repeat()\n\n    iterator = tf.data.Iterator.from_structure(\n        output_types=dataset.output_types, output_shapes=dataset.output_shapes\n    )\n    iterator_init_op = iterator.make_initializer(dataset=dataset)\n    iterator_images, iterator_masks = iterator.get_next()\n\n    # Import some model:\n    complex_model = keras.layers.Conv2D(\n        filters=3,\n        kernel_size=(1, 1),\n        activation=\"relu\",\n        padding=\"same\",\n        data_format=\"channels_last\",\n    )\n\n    inputs = keras.layers.Input(tensor=iterator_images)\n    outputs = complex_model(inputs)\n    model = keras.models.Model(inputs=inputs, outputs=outputs)\n\n    model.compile(\n        optimizer=keras.optimizers.RMSprop(),\n        loss=lambda y_true, y_pred: K.mean(\n            K.binary_crossentropy(target=y_true, output=y_pred)\n        ),\n        target_tensors=[iterator_masks]\n    )\n\n    model.fit(epochs=5, steps_per_epoch=num_batches)\n<\/code><\/pre>\n\n<p>It results in the following error:<\/p>\n\n<blockquote>\n  <p>FailedPreconditionError: GetNext() failed because the iterator has not been initialized. Ensure that you have run the initializer operation for this iterator before getting the next element.\n       [[{{node IteratorGetNext}} = IteratorGetNextoutput_shapes=[[?,5,5,3], [?,5,5,3]], output_types=[DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>In the case of using only Tensorflow, I would have to do something like:<\/p>\n\n<pre><code>sess.run(iterator_init_op)\n<\/code><\/pre>\n\n<p>How should I do that using Keras APIs?<\/p>\n","answers":[]}
+{"QuestionId":53119642,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-02T13:39:16.577","AcceptedAnswerId":53154925.0,"OwnerUserId":7945192.0,"Title":"Tensorflow: Determinism of AdamOptimizer when running in CPU","Body":"<p>In my application I optimize a neural network using the apply_gradient() and the AdamOptimizer. I noticed that I cannot reproduce the results exactly every time I run my application and it is required in order to compare results. While searching for similar issues I noticed that the issue of non-determinism can be encountered when Tensorflow runs in GPU, but my problem arises even in CPU.<\/p>\n\n<p>I simplified my code and I think that the problem is the apply_gradients, but I am not sure because I am not experienced with tensorflow. The following code trains a neural network with random data (with a properly seeded numpy of course), and writes the network's parameters on a log file. If you run the program twice you will notice substantial differences in the net's params between different runs of the program, even though the training data are the same.<\/p>\n\n<p>I use Tensorflow 1.11.0<\/p>\n\n<p>Why this is happening? Is it probably due to arithmetic precision errors or am I doing something wrong?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport logging\nimport sys\nimport math\nimport datetime\n\nif __name__ == '__main__':\n\n    # Just configure logging\n    logging.basicConfig(filename='example.log', filemode='w', level=logging.DEBUG)\n    logging.debug('Python version: %s', sys.version)\n    logging.debug('Tensorflow version: %s', tf.__version__)\n    logging.debug('Date time: %s', datetime.datetime.now())\n\n    # Seed everything\n    randomseed = 999\n    np.random.seed(randomseed)\n    tf.set_random_seed(randomseed)\n\n    # Create a Neural Network\n    inputs = tf.placeholder(tf.float64, [None, 10])\n    hidden_dims = [100, 200]\n    net = inputs\n    for dim in hidden_dims:\n        weight_initializer = tf.initializers.random_uniform(minval= - 1 \/ math.sqrt(10), maxval = 1 \/ math.sqrt(10))\n        net = tf.layers.dense(inputs=net, units=dim, kernel_initializer=weight_initializer, bias_initializer=weight_initializer)\n        net = tf.layers.batch_normalization(inputs=net)\n        net = tf.nn.relu(net)\n    weight_initializer = tf.initializers.random_uniform(minval=-3e-3, maxval=3e-3)\n    net = tf.layers.dense(inputs=net, units=20, kernel_initializer= weight_initializer, bias_initializer= weight_initializer)\n    out = tf.nn.tanh(net)\n    net_params = tf.trainable_variables()\n\n    # Define some gradient ops\n    grad = tf.gradients(out, net_params)\n    gradients = list(map(lambda x: tf.div(x, 10), grad))\n    optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).apply_gradients(zip(gradients, net_params))\n\n    with tf.Session() as sess:\n\n        sess.run(tf.global_variables_initializer())\n        for i in range(100):\n             random_input = np.random.rand(10, 10)\n             logging.debug('random_input: %s', random_input.__str__())\n\n             sess.run(optimizer, feed_dict={inputs: random_input})\n             params = sess.run(net_params)\n             logging.debug('params: %d' + params.__str__())\n<\/code><\/pre>\n","answers":[{"AnswerId":"53154925","CreationDate":"2018-11-05T12:58:33.470","ParentId":null,"OwnerUserId":"7945192","Title":null,"Body":"<p>In case some else encounters the same problem, the solution is to set the gate_gradients flag to True in tf.gradients(). Using this you lose parallelism and performance during the calculation and the application of the gradients to the optimizer, but you gain reproducibility.<\/p>\n"}]}
+{"QuestionId":53119748,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-02T13:45:51.290","AcceptedAnswerId":null,"OwnerUserId":10596277.0,"Title":"Layer which works on single sample to layer which works on batches","Body":"<p>I currently have a layer like the following:<\/p>\n\n<pre><code>class MyLayer(tf.keras.layers.Layer):\n    def __init__(self):\n        super(MyLayer, self).__init__()\n\n    def call(self, img, params):\n        tf.foo(img)\n        tf.bar(img, params)\n        return img\n<\/code><\/pre>\n\n<p>The call method gets and input <code>img<\/code> with shape <code>(128, 128, 3)<\/code> and <code>params<\/code> with shape <code>(15)<\/code>.<\/p>\n\n<p>What do I have to change such that the layer can operate on batches? The input <code>img<\/code> would for example have shape <code>(32, 128, 128, 3)<\/code> and <code>params<\/code> would have shape <code>(32, 15)<\/code>.    <\/p>\n\n<p>So the question is basically: how do I have to edit the layer such that it does the same stuff it does now but for each image in a batch?<\/p>\n","answers":[{"AnswerId":"53119864","CreationDate":"2018-11-02T13:53:23.490","ParentId":null,"OwnerUserId":"9137052","Title":null,"Body":"<p>Well by default your layer should recognize the tensor coming from the previous layer and be able to make sure its input shape conforms to that spec:\ni.e.<\/p>\n\n<pre><code>(None, 128,128,3)\n<\/code><\/pre>\n\n<p>So the batch size shouldn't matter. Your problem is most likely that you have no build() function as well as you're not passing **kwargs through your init and its super.<\/p>\n\n<p>Also, the call() should only have self and x as arguments, if you need more than one input, you should use a concatenation layer before it to push the two tensors together, or make a list with them<\/p>\n"},{"AnswerId":"53119843","CreationDate":"2018-11-02T13:51:57.537","ParentId":null,"OwnerUserId":"7391472","Title":null,"Body":"<p>perhaps you simply wanna reshape the size of the image tensor??\nwhat you wanna do would be something like, x=np.reshape(x,(32,128,128,3)). \nLemme know if that helps.<\/p>\n"}]}
+{"QuestionId":53120830,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-02T14:50:44.840","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"How to define loss function on Deep dream like Gradient decent on Model input","Body":"<p>So I have a trained sequential model (Categorizer) in python Kearns \/ Tensorflow and I have some Input. I want to optimize the input to maximize the Category hit.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef dream(input, model, iterations, learning_rate):\n    target = model.predict(input)\n    var = tf.Variable(\"var\", Input.shape)\n    loss = np.linalg.norm(model.predict(var)-goal)\n    for i in range(iterations):\n        input -= learning_rate * tf.gradients(loss,input)\n    return input\n<\/code><\/pre>\n\n<p>However, this doesn't work.<\/p>\n\n<p>How do I define loss correctly?<\/p>\n","answers":[{"AnswerId":"53212337","CreationDate":"2018-11-08T16:43:48.943","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>It is hard to say exactly what the problem is without seeing the rest of your code, so this is going to be based on some assumptions. <\/p>\n\n<ol>\n<li>The first is that this is a problem where you intend to learn a variable input image to produce a model output that better fits into one of several categories.<\/li>\n<li>The second is that the <code>model.predict()<\/code> function is the one specified <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/models\/Model#predict\" rel=\"nofollow noreferrer\">here<\/a>.<\/li>\n<li>The third is that <code>goal<\/code> is an ideal classification of your modified input image into the class into which it was most strongly classified before modification.<\/li>\n<li>The fourth is that the argument <code>image<\/code> has been instantated as a variable datatype. <\/li>\n<\/ol>\n\n<p>If all of this is true then it seems what you are looking for is a means of defining a loss function that compares a set of logits to a set of labels. In which case a function such as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/metrics\/categorical_crossentropy\" rel=\"nofollow noreferrer\">tf.keras.metrics.categorical_crossentropy()<\/a>, should do that for you. Using a loss function based on a <code>numpy<\/code> operation will not work because numpy lacks the gradient handling functionality required for back-propagation to function. <code>tf.<\/code> functions only, please. <\/p>\n\n<p>However if, as you say, \"var has shape ()\" this is unlikely to work. Possibly because you have defined it using <code>Input.shape()<\/code>, when the argument passed to the function is called <code>input<\/code> (cases very definitely matter).<\/p>\n\n<p>See below for a rough adaption of your code to something that seems plausible. <\/p>\n\n<hr>\n\n<pre><code>def dream(input, model, iterations, learning_rate):\n    # decide into which class your input should fit\n    target = model.predict(input)\n    # create ideal output of classification (1 for target class, 0 for others)\n    goal = tf.one_hot(indices=tf.cast(target, tf.int32), depth=1)\n    # define loss according to the cross entropy of the model prediciton vs goal\n    loss =  tf.keras.metrics.categorical_crossentropy(model.predict(input), goal)\n    # iterateively modify the image according to the gradients produced by \n    # differentiating the loss with respect to the variable image pixels\n    for i in range(iterations):\n        input -= learning_rate * tf.gradients(loss,input)\n    return input\n\n# SETUP\noriginal_image = tf.keras.Input(shape=(a,b,3))         \ninput = tf.Variable(original_image, \"variable_image\")  # create trainable image\n\n# add some layers\nx = tf.keras.layers.Dense(n_layers, activation=tf.nn.relu)(input)\noutput = tf.keras.layers.Dense(n_categories, activation=tf.nn.softmax)(x)\n# link model\nmodel = tf.keras.Model(inputs=input, outputs=output)\n\n# call function\ndream(input, model, iterations, learning_rate)\n<\/code><\/pre>\n\n<p>CAVEAT: This code is a hypothetical framework, based on assumptions about the intended purpose of the code. <strong>This code is untested and is not intended to be run verbatim<\/strong>. <\/p>\n"}]}
+{"QuestionId":53120951,"AnswerCount":0,"Tags":"<python><tensorflow><python-3.6>","CreationDate":"2018-11-02T14:57:54.083","AcceptedAnswerId":null,"OwnerUserId":10550693.0,"Title":"TensorFlow not importing properly","Body":"<p>I am using Python 3.6.0. I have installed TensorFlow using this command <\/p>\n\n<pre><code>pip install tensorflow \npip3 install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>I got success messages after isntalling these packages but now i am trying to execute this code in Python idle<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport sklearn\nfrom glob import glob\nimport os\nprint(os.listdir(\"C:\\\\Users\\\\TechProBox\\\\Desktop\\\\A_DeviceMotion_data\"))\n<\/code><\/pre>\n\n<p>But it gives me following error<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"C:\\Users\\TechProBox\\Desktop\\python.py\", line 4, in &lt;module&gt;\nimport tensorflow as tf\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\nFile \"C:\\Program Files\\Python36\\lib\\site- \npackages\\tensorflow\\python\\__init__.py\", line 52, in &lt;module&gt;\nfrom tensorflow.core.framework.graph_pb2 import *\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 6, in &lt;module&gt;\nfrom google.protobuf import descriptor as _descriptor\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in &lt;module&gt;\nfrom google.protobuf.pyext import _message\nImportError: DLL load failed: The specified procedure could not be found.\n<\/code><\/pre>\n\n<p>If i remove the tensorflow import then i get the required output like this which i need<\/p>\n\n<pre><code>['dws_1', 'dws_11', 'dws_2', 'jog_16', 'jog_9', 'sit_13', 'sit_5', 'std_14', 'std_6', 'ups_12', 'ups_3', 'ups_4', 'wlk_15', 'wlk_7', 'wlk_8']\n<\/code><\/pre>\n\n<p>But tensorflow import gives error.What could be the possible cause of this? What is this error and what should i do?<\/p>\n","answers":[]}
+{"QuestionId":53120977,"AnswerCount":0,"Tags":"<neural-network><pytorch><grayscale><batch-normalization>","CreationDate":"2018-11-02T14:59:19.610","AcceptedAnswerId":null,"OwnerUserId":9072735.0,"Title":"BatchNormalization on 2D data in pytorch","Body":"<p>I'm really confused about using batch normalization. For example, i have <code>(256,256)<\/code> image , and i train my network with <code>batch_size = 4<\/code>. I need <code>(4,64,64)<\/code> feature map for each batch, so I have following model:<\/p>\n\n<pre><code>Conv1d(1,1,kernel_size = 1 , stride = 2, padding = 0)\nBatchNorm1d(x)\nConv1d(1,1,kernel_size = 1, stride = 2,padding = 0)\nBatchNorm1(x)\n<\/code><\/pre>\n\n<p>For above model, what should be proper number for <code>x<\/code>? <\/p>\n\n<p><code>BatchNorm<\/code> is performed, generally, on channel axis. However, desired feature map should have <code>(4,64,64)<\/code> size. Setting <code>x<\/code> to <code>1<\/code> is not make sense. Am I wrong?<\/p>\n\n<p>Also I have one question. What is the difference between performing <code>batchNorm<\/code> on channel axis and other axis?<\/p>\n","answers":[]}
+{"QuestionId":53121232,"AnswerCount":1,"Tags":"<tensorflow><protocol-buffers><flatbuffers>","CreationDate":"2018-11-02T15:13:39.557","AcceptedAnswerId":null,"OwnerUserId":403994.0,"Title":"Why tensorflow does not use flatbuffers by default","Body":"<p>I can see that TensorFlow Lite is using flatbuffers by default and documentation page notes that in fact it's more efficient.<\/p>\n\n<p>Why isn't TensorFlow using it by default?<\/p>\n","answers":[{"AnswerId":"53121820","CreationDate":"2018-11-02T15:49:15.087","ParentId":null,"OwnerUserId":"17419","Title":null,"Body":"<p>Probably because the team didn't know of its existence when they started it. FlatBuffers is a relatively new technology, whereas Protocol Buffers has been in use at Google almost since the start, and is used for everything by default.<\/p>\n"}]}
+{"QuestionId":53121846,"AnswerCount":1,"Tags":"<docker><tensorflow><nvidia><google-colaboratory>","CreationDate":"2018-11-02T15:51:02.940","AcceptedAnswerId":53123593.0,"OwnerUserId":10507928.0,"Title":"tensorflow-gpu on a Docker of host running on radeon 560?","Body":"<p>I am trying out tensorflow for deep learning, but for GPU support, tensorflow needs Nvidia GPU, but I have a Radeon 560. Is there any way of running tensorflow GPU in a Docker container? I couldn't wrap my head around Docker. Are there any other possibilities other than using Colab or Docker?<\/p>\n","answers":[{"AnswerId":"53123593","CreationDate":"2018-11-02T17:52:19.160","ParentId":null,"OwnerUserId":"5879698","Title":null,"Body":"<p>The official tensorflow-gpu docker images at this time only support NVIDIA gpus and host running the NVIDIA cuda-driver. In order to use an AMD GPU, you need to use a version of TensorFlow built with ROCm support. Google for \"tensorflow ROCm\" and you'll find some pages on how to do this. The best link is probably:<\/p>\n\n<p><a href=\"https:\/\/github.com\/ROCmSoftwarePlatform\/tensorflow-upstream\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ROCmSoftwarePlatform\/tensorflow-upstream<\/a><\/p>\n"}]}
+{"QuestionId":53122005,"AnswerCount":1,"Tags":"<docker><pytorch><google-colaboratory><fast-ai>","CreationDate":"2018-11-02T16:00:40.600","AcceptedAnswerId":53842203.0,"OwnerUserId":4590385.0,"Title":"Set higher shared memory to avoid RuntimeError with PyTorch on Google Colab","Body":"<p>Using pytorch 1.0 Preview with fastai v1.0 in Colab. <\/p>\n\n<p>I often get <code>RuntimeError: DataLoader worker (pid 13) is killed by signal: Bus error.<\/code> for more memory intensive tasks (nothing huge).<\/p>\n\n<p>Looks like a shared memory issue:\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/5040#issue-294274594\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/5040#issue-294274594<\/a><\/p>\n\n<p>Fix looks like it is to change shared memory in the docker container:<\/p>\n\n<p><a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/2244#issuecomment-318864552\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/2244#issuecomment-318864552<\/a><\/p>\n\n<blockquote>\n  <p>Looks like the shared memory of the docker container wasn't set high enough. Setting a higher amount by adding --shm-size 8G to the docker run command seems to be the trick as mentioned here. <\/p>\n<\/blockquote>\n\n<p>How can I increase the shared memory of the docker container running in Colab or otherwise avoid this error?<\/p>\n","answers":[{"AnswerId":"53842203","CreationDate":"2018-12-18T22:44:20.893","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p>It's not possible to modify this setting in colab, but the default was raised to fix this issue already so you should not need to change the setting further: <a href=\"https:\/\/github.com\/googlecolab\/colabtools\/issues\/329\" rel=\"nofollow noreferrer\">https:\/\/github.com\/googlecolab\/colabtools\/issues\/329<\/a><\/p>\n"}]}
+{"QuestionId":53122041,"AnswerCount":0,"Tags":"<python><image><tensorflow><h5py><cudamat>","CreationDate":"2018-11-02T16:02:33.493","AcceptedAnswerId":null,"OwnerUserId":8323113.0,"Title":"Python H5PY IOError: Unable to open file..... error message = 'no such file or directory'","Body":"<pre><code>data_loc = os.path.join('data', 'datasets', data_code + '.hdf5')\ndata = h5py.File(data_loc, 'r')\n<\/code><\/pre>\n\n<p>By the way I am using Window 10 x64<\/p>\n\n<p>I am getting error while I run\n<code>python train.py fcn_rffc4 brats_fold0 brats_fold0 600 -ch False<\/code> command<\/p>\n\n<p>Error is shortly, IOError Unable to open file.... beacuse 'no such file or directory'\nLong error that I got is :<\/p>\n\n<pre><code>Traceback (most recent call last):\n\nFile \"train.py\", line 305, in \nstart_training(model_code, data_code, checkpoint, max_passes=n_epochs, train_dir=train_dir)\n\nFile \"train.py\", line 273, in start_training\ncoach, model_def = setup_training(model_code, data_code, checkpoint, max_passes)\n\nFile \"train.py\", line 140, in setup_training\ntrain, valid, test = retrieve_data(data_code=data_code)\n\nFile \"train.py\", line 57, in retrieve_data\ndata = h5py.File(data_loc, 'r')\n\nFile \"C:\\Users\\yunus\\Anaconda3\\envs\\medical_image\\lib\\site-packages\\h5py_hl\\files.py\", line 271, in init\nfid = make_fid(name, mode, userblock_size, fapl, swmr=swmr)\n\nFile \"C:\\Users\\yunus\\Anaconda3\\envs\\medical_image\\lib\\site-packages\\h5py_hl\\files.py\", line 101, in make_fid\nfid = h5f.open(name, flags, fapl=fapl)\nFile \"h5py_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\n\nFile \"h5py_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\n\nFile \"h5py\\h5f.pyx\", line 78, in h5py.h5f.open\n\nIOError: Unable to open file (Unable to open file: name = \n'data\\datasets\\brats_fold0.hdf5', errno = 2, error message = 'no such file or directory', flags = 0, o_flags = 0)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53122116,"AnswerCount":1,"Tags":"<keras><lstm><word-embedding><sequence-to-sequence>","CreationDate":"2018-11-02T16:06:37.720","AcceptedAnswerId":53123343.0,"OwnerUserId":3487059.0,"Title":"Merging sequence embedding with Time Series Features","Body":"<p>I am having trouble around certain aspects of the <strong>Keras<\/strong> implementation of <strong>LSTM<\/strong>. This is a description of my problem:<\/p>\n\n<p>I am trying to train a model for word correctness prediction. My model has two types of inputs: <\/p>\n\n<ol>\n<li>A word sequence (sentence) <\/li>\n<li>And a sequence of features vector (for each word I compute a features victor of 6).<\/li>\n<\/ol>\n\n<p><strong>e.g.<\/strong> <\/p>\n\n<pre><code>input_1 = ['we', 'have', 'two', 'review'] \ninput_2 = [\n           [1.25, 0.01, 0.000787, 5.235, 0.0, 0.002091], \n           [ 0.0787, 0.02342, 5.4595, 0.002091, 0.003477, 0.0], \n           [0.371533, 0.529893, 0.371533, 0.6, 0.0194156, 0.003297],\n           [0.471533, 0.635, 0.458, 0.7, 0.0194156, 0.0287]\n          ] \n\n gives output = [1, 1, 2, 1]\n<\/code><\/pre>\n\n<p>As each sentence in my training set has different length, I should zero-pad all of my sentences such that they all have the same length. <\/p>\n\n<p>My question is how about the second input, should I do padding! and how? as they are vectors.<\/p>\n\n<p><strong>Model Architecture :<\/strong><\/p>\n\n<pre><code>input1 = Input(shape=(seq_length,), dtype='int32')\nemb = Embedding(input_dim=num_words, output_dim = num_dimension, \ninput_length=seq_length, weights=[embeddings], mask_zero=True,trainable=False)(input_layer)\n\ninput2 = Input(shape=(seq_length,6 ))\nx = keras.layers.concatenate([emb, input2],axis=2)\n\nlstm = LSTM(64, return_sequences=True)(x)\nackwards = LSTM(128, return_sequences=True, go_backwards=True)(x)\n\ncommon = merge([forwards, backwards], mode='concat', concat_axis=-1)\nout = TimeDistributed(Dense(no_targets, activation='softmax'))(lstm)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53123343","CreationDate":"2018-11-02T17:31:30.240","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You are on the right track and yes you would need to pad your second input with zero rows to match the sentence lengths. Essentially it would look like this:<\/p>\n\n<pre><code># Input 1\nX1 = [[12, 34, 3], [6, 7, 0]] # where numbers are word indices and 0 is padding\n# Input 2\nX2 = [[[1.23,...,2.4], [1.24, ...], [0.6, ...]], [[3.25, ...], [2.4, ...], [0,0,0,0,0]]]\n# So the padded words get zero feature vectors as well and the shapes match\n<\/code><\/pre>\n\n<p>But fear not, because you concatenate <code>emb<\/code> with <code>input2<\/code> the <code>mask_zero=True<\/code> also gets propagated to the concatenated vector so the LSTM actually ignores the padding from second input as well.<\/p>\n"}]}
+{"QuestionId":53122412,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-11-02T16:26:28.680","AcceptedAnswerId":null,"OwnerUserId":886724.0,"Title":"AOT compilation of an Estimator","Body":"<p>I find it very challenging to find in the documentation any help on this important issue.. Indeed, after creating an estimator (canned or custom), one wants to <code>tf.compile<\/code> the resulting predictor, produce the <code>.so<\/code> and link it to one's project..<\/p>\n\n<p>So I have my calib class in which I define a simple linear estimator<\/p>\n\n<pre><code>self.model = tf.estimator.LinearRegressor(\n        feature_columns=self.feature_columns,\n        model_dir = self.model_dir)\n<\/code><\/pre>\n\n<p>After training, I want to \n1- get the trained model with optimal parameters (load it in my variable self.model)<\/p>\n\n<p>2- extract the graph and freeze it<\/p>\n\n<p>3- tf.compile that graph<\/p>\n\n<p>I could not find any way to do parts 1- and 2-. Once I have them, part 3 is solved by using tf.compile<\/p>\n\n<p>Can you please point me to a good way to it?<\/p>\n","answers":[{"AnswerId":"53174296","CreationDate":"2018-11-06T14:50:26.987","ParentId":null,"OwnerUserId":"886724","Title":null,"Body":"<p><a href=\"https:\/\/medium.com\/google-cloud\/optimizing-tensorflow-models-for-serving-959080e9ddbf\" rel=\"nofollow noreferrer\">This<\/a> helped a lot. It may help anyone passing by here<\/p>\n"}]}
+{"QuestionId":53122714,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-02T16:46:50.053","AcceptedAnswerId":53178556.0,"OwnerUserId":3697273.0,"Title":"TensorFlow Probability MCMC with Bernoulli distribution","Body":"<p>I need to use TensorFlow Probability to implement Markov Chain Monte Carlo with sampling from a Bernoulli distribution. \nHowever, my attempts are showing results not consistent with what I would expect from a Bernoulli distribution.<\/p>\n\n<p>I modified the example given in the documentation of tfp.mcmc.sample_chain (sampling from a diagonal-variance Gaussian) <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/mcmc\/sample_chain\" rel=\"nofollow noreferrer\">example here<\/a> to draw from a Bernoulli distribution. Since the Bernoulli distribution is discrete, I used the RandomWalkMetropolis transition kernel instead of the\nHamiltonian Monte Carlo kernel, which I expect would not work since it computes a gradient.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport tensorflow as tf\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\n\ndef make_likelihood(event_prob):\n    return tfd.Bernoulli(probs=event_prob,dtype=tf.float32)\n\n\ndims=1\nevent_prob = 0.3\nnum_results = 30000\nlikelihood = make_likelihood(event_prob)\n\n\nstates, kernel_results = tfp.mcmc.sample_chain(\n    num_results=num_results,\n    current_state=tf.zeros(dims),\n    kernel = tfp.mcmc.RandomWalkMetropolis(\n              target_log_prob_fn=likelihood.log_prob,\n              new_state_fn=tfp.mcmc.random_walk_normal_fn(scale=1.0),\n              seed=124\n             ),\n    num_burnin_steps=5000)\n\nchain_vals = states\n\n# Compute sample stats.\nsample_mean = tf.reduce_mean(states, axis=0)\nsample_var = tf.reduce_mean(\n    tf.squared_difference(states, sample_mean),\n    axis=0)\n\n#initialize the variable\ninit_op = tf.global_variables_initializer()\n\n#run the graph\nwith tf.Session() as sess:\n    sess.run(init_op) \n    [sample_mean_, sample_var_, chain_vals_] = sess.run([sample_mean,sample_var,chain_vals])\n\nchain_samples = (chain_vals_[:] )   \nprint ('Sample mean = {}'.format(sample_mean_))\nprint ('Sample var = {}'.format(sample_var_))\nfig, axes = plt.subplots(2, 1)\nfig.set_size_inches(12, 10)\n\naxes[0].plot(chain_samples[:])\naxes[0].title.set_text(\"values sample chain tfd.Bernoulli\")\nsns.kdeplot(chain_samples[:,0], ax=axes[1], shade=True)\naxes[1].title.set_text(\"chain tfd.Bernoulli distribution\")\nfig.tight_layout()\nplt.show()\n<\/code><\/pre>\n\n<p>I expected to see values for the states of the Markov chain in the interval [0,1].<\/p>\n\n<p>The result for the Markov chain values does not look like what is expected for a Bernoulli distribution, nor does the KDE plot, as shown\nin this figure:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DSnre.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DSnre.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Do I have a conceptual flaw with my example, or mistake in using the TensorFlow Probability API ?<\/p>\n\n<p>Or is there possibly an issue with the TF.Probability implementation of Markov Chain Monte Carlo \nusing a discrete distribution such as the  Bernoulli distribution?<\/p>\n","answers":[{"AnswerId":"53178556","CreationDate":"2018-11-06T19:17:12.047","ParentId":null,"OwnerUserId":"10614872","Title":null,"Body":"<p>I think the source of your confusing experience is that you are still using a continuous proposal distribution in the <code>RandomWalkMetropolis<\/code> transition.  The convention in TensorFlow Probability for integer distributions (including <code>Bernoulli<\/code>) is to implement a continuous relaxation by default.  IIRC, for Bernoulli, that's <code>pdf(x) ~ p ** x * (1 - p) ** (1 - x)<\/code>; once <code>x<\/code> becomes negative, this will steadily drive your random walk Markov chain toward <code>-inf<\/code>, as you observe.<\/p>\n\n<p>Several things you can do about this:<\/p>\n\n<ol>\n<li>Use pass <code>validate_args=True<\/code> to the <code>Bernoulli<\/code> constructor.  This will crash if <code>x<\/code> is not 0 or 1, helping you detect the issue (but don't if you want non-integer results in the interval [0, 1]).<\/li>\n<li>Use a different proposal function, such as an independent uniform between 0 and 1.  Writing your own is not very hard---here's literally the code of the Gaussian drift proposal function you were using: <a href=\"https:\/\/github.com\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/python\/mcmc\/random_walk_metropolis.py#L97-L107\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/python\/mcmc\/random_walk_metropolis.py#L97-L107<\/a>.  Note that the proposal needs to be symmetric to work correctly with <code>RandomWalkMetropolis<\/code>.<\/li>\n<li>Use a different MCMC transition operator altogether.<\/li>\n<\/ol>\n\n<p>I also filed a ticket about making a TransitionKernel for independence proposals (such as what I imagine you might need): <a href=\"https:\/\/github.com\/tensorflow\/probability\/issues\/218\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/issues\/218<\/a><\/p>\n"}]}
+{"QuestionId":53122788,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><quantization>","CreationDate":"2018-11-02T16:50:53.947","AcceptedAnswerId":null,"OwnerUserId":9104473.0,"Title":"How to convert the results of qint32 to quint8","Body":"<p>I am fairly new to quantization in tensorflow. I have the following code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.ops.gen_math_ops import quantized_mat_mul\nfrom tensorflow.python.ops.gen_math_ops import quantized_add\n\na = tf.constant([1.4,-1.2],shape =[1,2],dtype = tf.float32)\na1 = tf.constant([[1.0],[1.0]],shape = [2,1], dtype=tf.float32)\nbias_a = tf.constant([-1.2,1.2],shape=[1,2],dtype = tf.float32)\n\nmin_a= tf.reduce_min(a)\nmax_a= tf.reduce_max(a)\nx = tf.quantize(a,min_a,max_a,tf.quint8)\n\nmin_a1= tf.reduce_min(a1)\nmax_a1= tf.reduce_max(a1)\nx1 = tf.quantize(a1,min_a1,max_a1,tf.quint8)\n\nmin_bias_a= tf.reduce_min(bias_a)\nmax_bias_a= tf.reduce_max(bias_a)\nbias_a_x = tf.quantize(bias_a,min_bias_a,max_bias_a,tf.qint32)\n\ny =tf.matmul(a,a1)\ny1 = quantized_mat_mul(x.output,x1.output,x.output_min,x.output_max,x1.output_min,x1.output_max)\n\no = quantized_add(y1.out,bias_a_x.output,y1.min_out,y1.max_out,bias_a_x.output_min,bias_a_x.output_max)\n<\/code><\/pre>\n\n<p>The results of y1 are as follows (qint32,float_max,float_min) where float_max and float_min represent the min and max float values for the quantized int 32 bit output.<\/p>\n\n<p>However, quantized_add requires inputs in the following format (quint8,float_max,float_min) where float_max and float_min is the max and min float values representing the quantized uint8 result.<\/p>\n\n<p>I wanted to know if there is a way to convert the results of qint32 to quint8 so that i can feed it into quantized_add function?<\/p>\n\n<p>Would be really grateful if anyone could help me.<\/p>\n\n<p>Thanks and Regards,\nAbhinav George<\/p>\n","answers":[{"AnswerId":"53133616","CreationDate":"2018-11-03T17:09:04.890","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You can cast the types from <code>tf.int32<\/code> to <code>tf.quint8<\/code> as follows:\n<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y1_out = tf.cast(y1.out, tf.quint8)\nbias_a_x_out = tf.cast(bias_a_x.output, tf.quint8)\n<\/code><\/pre>\n\n<p>And then perform <code>quantized_add<\/code> as:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>o = quantized_add(y1_out, \n              bias_a_x_out, \n              y1.min_out, \n              y1.max_out, \n              bias_a_x.output_min, \n              bias_a_x.output_max)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53123137,"AnswerCount":2,"Tags":"<python><keras><lstm><keras-layer><autoencoder>","CreationDate":"2018-11-02T17:16:52.300","AcceptedAnswerId":null,"OwnerUserId":6894995.0,"Title":"Adding a muplitiply layer to an autoencoder in Keras","Body":"<p>I want to add a multiply layer on top of an LSTM autoencoder.\nThe multiply layer should multiply the tensor for a constant value.\nI wrote the following code which work without the multiply layer.\nDoes anyone know how to adjust and make this working?<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\nfrom keras.models import Sequential, Model\nfrom keras.layers import Input, LSTM, RepeatVector, TimeDistributed\nfrom keras.layers.core import Flatten, Dense, Dropout, Lambda\nfrom keras.optimizers import SGD, RMSprop, Adam\nfrom keras import objectives\nfrom keras.engine.topology import Layer\nimport numpy as np\n\nclass LayerKMultiply(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        self.k = Null\n        super(LayerKMultiply, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        # Create a trainable weight variable for this layer.\n        self.k = self.add_weight(\n            name='k',\n            shape=(),\n            initializer='ones',\n            dtype='float32',\n            trainable=True,\n        )\n        super(LayerKMultiply, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        #return K.tf.multiply(self.k, x)\n        return self.k * x\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.output_dim)\n\n    timesteps, input_dim, latent_dim = 10, 3, 32\n\n\ninputs = Input(shape=(timesteps, input_dim))\nencoded = LSTM(latent_dim, return_sequences=False, activation='linear')(inputs)\ndecoded = RepeatVector(timesteps)(encoded)\ndecoded = LSTM(input_dim, return_sequences=True, activation='linear')(decoded)\ndecoded = TimeDistributed(Dense(input_dim, activation='linear'))(decoded)\n#decoded = LayerKMultiply(k = 20)(decoded)\n\nsequence_autoencoder = Model(inputs, decoded)\nencoder = Model(inputs, encoded)\n\nautoencoder = Model(inputs, decoded)\nautoencoder.compile(optimizer='adam', loss='mse')\n\n    X = np.array([[[1,2,3,4,5,6,7,8,9,10],[1,2,3,4,5,6,7,8,9,10],[1,2,3,4,5,6,7,8,9,10]]])\nX = X.reshape(1,10,3)\np = autoencoder.predict(x=X, batch_size=1)\nprint(p)\n<\/code><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-3-b2f9497bbf47&gt; in &lt;module&gt;()\n      7 decoded = LSTM(input_dim, return_sequences=True, activation='linear')(decoded)\n      8 decoded = TimeDistributed(Dense(input_dim, activation='linear'))(decoded)\n----&gt; 9 decoded = LayerKMultiply(k = 20)(decoded)\n     10 \n     11 sequence_autoencoder = Model(inputs, decoded)\n\nTypeError: __init__() missing 1 required positional argument: 'output_dim'\n<\/code><\/pre>\n\n<h1>EDIT<\/h1>\n\n<p>What I want to achieve is the architecture described in the following images:<\/p>\n\n<p><a href=\"https:\/\/github.com\/mg64ve\/SMTDAE\/blob\/master\/images\/SMTDAE.png\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mg64ve\/SMTDAE\/blob\/master\/images\/SMTDAE.png<\/a>\n<a href=\"https:\/\/github.com\/mg64ve\/SMTDAE\/blob\/master\/images\/REF.png\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mg64ve\/SMTDAE\/blob\/master\/images\/REF.png<\/a><\/p>\n\n<p>So in this sense, I believe the multiply layer hase to be before the TimeDistributed and Dense layers.\nI modified the code as following:<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\nfrom keras.models import Sequential, Model\nfrom keras.layers import Input, LSTM, RepeatVector, TimeDistributed\nfrom keras.layers.core import Flatten, Dense, Dropout, Lambda\nfrom keras.optimizers import SGD, RMSprop, Adam\nfrom keras import objectives\nfrom keras.engine.topology import Layer\nimport numpy as np\n\nclass LayerKMultiply(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        self.k = None\n        super(LayerKMultiply, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        # Create a trainable weight variable for this layer.\n        self.k = self.add_weight(\n            name='k',\n            shape=(),\n            initializer='ones',\n            dtype='float32',\n            trainable=True,\n        )\n        super(LayerKMultiply, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return self.k * x\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], input_shape[1], self.output_dim)\n\ntimesteps, input_dim, latent_dim = 31, 31, 32\n\n\ninputs = Input(shape=(timesteps, input_dim))\nencoded = LSTM(latent_dim, return_sequences=False, activation='linear')(inputs)\ndecoded = RepeatVector(timesteps)(encoded)\ndecoded = LSTM(input_dim, return_sequences=True, activation='linear')(decoded)\ndecoded = LayerKMultiply(20)(decoded)\ndecoded = TimeDistributed(Dense(input_dim, activation='linear'))(decoded)\n\nautoencoder = Model(inputs, decoded)\n\nbatch_size = 100\nX = np.zeros([5000,31,31])\nautoencoder.fit(X, X, batch_size = batch_size, epochs=3)\n\nautoencoder.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>But I am still getting the following error:<\/p>\n\n<pre><code>InvalidArgumentError: Incompatible shapes: [155,31,31] vs. [100,31,31]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53126625","CreationDate":"2018-11-02T22:36:35.253","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>You are mixing <em>positional arguments<\/em> with <em>keyword arguments<\/em>. When you define a function like <code>def __init__(self, output_dim, **kwargs)<\/code> <code>output_dim<\/code> is a positional argument. You need to:<\/p>\n\n<ul>\n<li>either pass 20 on its own <code>LayerMultiply(20)(decoded)<\/code><\/li>\n<li>or change <code>def __init__(self, k=10, **kwargs)<\/code><\/li>\n<li>or remove <code>output_dim<\/code> from definition and use <code>self.output_dim = kwargs['k']<\/code><\/li>\n<\/ul>\n\n<p>More information <a href=\"https:\/\/stackoverflow.com\/questions\/9450656\/positional-argument-v-s-keyword-argument?lq=1\">here<\/a>.<\/p>\n"},{"AnswerId":"53168027","CreationDate":"2018-11-06T08:11:33.903","ParentId":null,"OwnerUserId":"6894995","Title":null,"Body":"<p>I believe the solution is the following:<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\nfrom keras.models import Sequential, Model\nfrom keras.layers import Input, LSTM, RepeatVector, TimeDistributed\nfrom keras.layers.core import Flatten, Dense, Dropout, Lambda\nfrom keras.optimizers import SGD, RMSprop, Adam\nfrom keras import objectives\nfrom keras.engine.topology import Layer\nimport numpy as np\n\nclass LayerKMultiply(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        self.k = None\n        super(LayerKMultiply, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        # Create a trainable weight variable for this layer.\n        self.k = self.add_weight(\n            name='k',\n            shape=(),\n            initializer='ones',\n            dtype='float32',\n            trainable=True,\n        )\n        super(LayerKMultiply, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return self.k * x\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], input_shape[1], input_shape[2])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53123763,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-02T18:06:48.113","AcceptedAnswerId":null,"OwnerUserId":10597685.0,"Title":"How to perform cartesian product with Tensorflow?","Body":"<p>I'm trying to cross stack two tensors, for example<\/p>\n\n<pre><code>[0,1,2],[2,3]-&gt;[[0,2],[0,3],[1,2],[1,3],[2,2],[2,3]]\n<\/code><\/pre>\n\n<p>Does anybody know which function can do that?<\/p>\n","answers":[{"AnswerId":"53123861","CreationDate":"2018-11-02T18:15:47.590","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think this does what you need:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant([0, 1, 2])\nb = tf.constant([2, 3])\nc = tf.stack(tf.meshgrid(a, b, indexing='ij'), axis=-1)\nc = tf.reshape(c, (-1, 2))\nwith tf.Session() as sess:\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[0 2]\n [0 3]\n [1 2]\n [1 3]\n [2 2]\n [2 3]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53124135,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-02T18:37:47.317","AcceptedAnswerId":null,"OwnerUserId":2872583.0,"Title":"Keras custom layer using argmax kills gradients","Body":"<p>After implementing a custom layer in Keras and using it in my network, I realized that all previous layers are not updated during training anymore. In fact, analyzing the <em>tensorboard distributions<\/em> of the weights and gradients computed for my network, the gradients for all layers preceding the custom layer are zero. <\/p>\n\n<p>The layer I am trying to implement performs the following:<\/p>\n\n<p><strong>Goal:<\/strong>\nGiven a tensor of size <code>B x H x W x D<\/code>, with <code>D=1<\/code>, add two new channels of size <code>H x W<\/code> to each sample of the batch (the values of these channels are given as a constant input, so they are not learned or dynamic). For the resulting tensor of shape <code>B x H x W x 3<\/code>, compute the <em>argmax<\/em> on the first channel sample-wise (i.e. over the 2 dimensional image) and apply the resulting indices on the second and third channel.<\/p>\n\n<p><strong>Shapes:<\/strong>\nGiven <code>B x H x W x D = (B x 8 x 8 x 1)<\/code>, the layer outputs a tensor of shape <code>B x 2 x 1<\/code> with an intermediate representation of size <code>B x H x W x 3<\/code>.<\/p>\n\n<p><strong>Numerical example:<\/strong>\nLet's assume <code>B = 1<\/code>, <code>H = W = 2<\/code>, and let the input tensor have the following values:<\/p>\n\n<pre><code>input[0, :, :, 0] = [[0.1, 0.2],\n                     [0.3, 0.4]]\n<\/code><\/pre>\n\n<p>Adding the new channels:<\/p>\n\n<pre><code>input[0, :, :, 0] = [[0.1, 0.2],\n                     [0.3, 0.4]]\n\ninput[0, :, :, 1] = [[a, b],\n                     [c, d]]\n\ninput[0, :, :, 2] = [[e, f],\n                     [g, h]]\n<\/code><\/pre>\n\n<p>The filter computes the <em>argmax<\/em> on the first channel, and returns the corresponding elements of the second and third channel:<\/p>\n\n<pre><code>indices = argmax(input[0, :, :, 0]) = [1, 1]\ninput[0,indices[0], indices[1], 1] = [d]\ninput[0,indices[0], indices[1], 1] = [h]\n\nLayer(input) = [\n                 [[d], [h]]\n               ]\n<\/code><\/pre>\n\n<p>Here I attach the code I wrote to implement the filter's logic:<\/p>\n\n<pre><code>class CustomGlobalMaxPooling(Layer):\n\n    def __init__(self, **kwargs):\n        \"\"\"Layer computing the argmax in the first channel of each sample in the input\n        tensor and returning the values associated to the second and third channel.\n        \"\"\"\n\n        super(CustomGlobalMaxPooling, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        assert len(input_shape) == 4, \"Only 4 dimensional input supported (B, H, W, D)\"\n        assert input_shape[-1] == 1, \"Only supports one channels. Given shape: {}\".format(input_shape)\n        super(CustomGlobalMaxPooling, self).build(input_shape)\n\n    def call(self, x, **kwargs):\n        batch_size = tf.shape(x)[0]\n        r = tf.cast(tf.range(batch_size), tf.int64)\n\n        img_h = tf.shape(x)[1]\n        img_w = tf.shape(x)[2]\n\n        ## HERE WE ADD THE NEW CHANNELS ON EACH SAMPLE (code omitted) ##\n        x = add_channels(x)\n\n        flattened_values = tf.reshape(x[:, :, :, 0], (batch_size, img_h * img_w))\n        flattened_1 = tf.reshape(x[:, :, :, 1], (batch_size, img_h * img_w))\n        flattened_2 = tf.reshape(x[:, :, :, 2], (batch_size, img_h * img_w))\n\n        argmax = tf.argmax(flattened_values, axis=1)\n        argmax = tf.transpose(tf.stack([r, argmax]), [1, 0])\n\n        max_1 = tf.gather_nd(flattened_1, argmax)\n        max_2 = tf.gather_nd(flattened_2, argmax)\n\n        max_1 = tf.expand_dims(max_1, -1)\n        max_2 = tf.expand_dims(max_2, -1)\n\n        return tf.transpose(tf.stack((max_1, max_2)), [1, 0, 2])\n\n    def compute_output_shape(self, input_shape):\n        batch_size = input_shape[0]\n        return (batch_size, 2, 1)\n<\/code><\/pre>\n\n<p>I guess that by performing the armgax on one channel and applying the indices on new channels, my network is not differentiable anymore, so no gradients can propagate.<\/p>\n\n<p>Is there a way around this problem?<\/p>\n","answers":[]}
+{"QuestionId":53124450,"AnswerCount":2,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2018-11-02T19:03:30.817","AcceptedAnswerId":53142457.0,"OwnerUserId":8500858.0,"Title":"ImportError: No module named absl.testing","Body":"<p>I've been trying to run the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md\" rel=\"nofollow noreferrer\">Object detection API<\/a>, at the point to test if the instalation worked: <\/p>\n\n<pre><code>python object_detection\/builders\/model_builder_test.py\n<\/code><\/pre>\n\n<p>I got the following error:<\/p>\n\n<pre><code>    File \"object_detection\/builders\/model_builder_test.py\", line 18, in &lt;module&gt;\n    from absl.testing import parameterized\nImportError: No module named absl.testing\n<\/code><\/pre>\n\n<p>I couldn't find a solution anywhere. Already tried <code>pip install absl-py<\/code> and <code>pip3 install absl-py<\/code> but without success.\nHope you could help me.<\/p>\n","answers":[{"AnswerId":"53142457","CreationDate":"2018-11-04T15:41:02.487","ParentId":null,"OwnerUserId":"8500858","Title":null,"Body":"<p>Got it working by using <code>python3 object_detection\/builders\/model_builder_test.py<\/code>\n instead of <code>python object_detection\/builders\/model_builder_test.py<\/code>. Forgot that I was using TF for python3.<\/p>\n"},{"AnswerId":"53319814","CreationDate":"2018-11-15T12:46:30.090","ParentId":null,"OwnerUserId":"4418009","Title":null,"Body":"<ul>\n<li><p>First install absl-py by running command if you are using tensorflow\nwith   python3<\/p>\n\n<p><em>sudo pip3 install absl-py<\/em> <\/p>\n\n<p>otherwise<\/p>\n\n<p><em>sudo pip install absl-py<\/em> <\/p><\/li>\n<li><p>After that run<\/p>\n\n<p><em>python3 object_detection\/builders\/model_builder_test.py<\/em><\/p>\n\n<p>Cheers, You done it !!<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53124755,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-02T19:27:25.997","AcceptedAnswerId":null,"OwnerUserId":4528518.0,"Title":"How to prevent Tensorflow from allocating the totality of a GPU memory when using eager execution?","Body":"<p>I have pretty much the same question that has already been answered <a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory\">here<\/a>, with a slight difference though:<\/p>\n\n<p>I'm working on a server with a few GPUs that I share with my colleagues for training our deep learning models. The server should also run a small web application that samples from our models. The sampling script uses the relatively new <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">eager execution<\/a>. In theory it allows me to stop Tensorflow from allocating all the GPU memory by providing a configuration like this:<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\ntf.enable_eager_execution(config=config)\n<\/code><\/pre>\n\n<p>In practice this does not work though. The documentation of the eager execution also states that not all the configuration options that work for sessions will work in eager execution (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/enable_eager_execution#args\" rel=\"nofollow noreferrer\">here<\/a>). But how can I limit the used memory then?<\/p>\n\n<p>I know that I can limit the visible devices like this:<\/p>\n\n<pre><code>os.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"1\"\n<\/code><\/pre>\n\n<p>But I don't want to constantly block an entire GPU for a task that gets called very occasionally and actually needs way less memory.<\/p>\n","answers":[{"AnswerId":"53563317","CreationDate":"2018-11-30T18:56:38.753","ParentId":null,"OwnerUserId":"5010481","Title":null,"Body":"<p>Which tensorflow version are you using?\nI'm using tensorflow v1.12 built from source (r1.12 branch) on Ubuntu 16.04 with Cuda 9.2 and Cudnn 7.2.1 and the allow_growth option seems to work as expected with eager execution.<\/p>\n"}]}
+{"QuestionId":53124809,"AnswerCount":1,"Tags":"<pytorch><embedding><collaborative-filtering>","CreationDate":"2018-11-02T19:32:00.220","AcceptedAnswerId":53169760.0,"OwnerUserId":2118596.0,"Title":"How to learn the embeddings in Pytorch and retrieve it later","Body":"<p>I am building a recommendation system where I predict the best item for each user given their purchase history of items. I have userIDs and itemIDs and how much itemID was purchased by userID. I have Millions of users and thousands of products. Not all products are purchased(there are some products that no one has bought them yet). Since the users and items are big I don't want to use one-hot vectors. I am using pytorch and I want to create and train the embeddings so that I can make the predictions for each user-item pair. I followed this tutorial <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/nlp\/word_embeddings_tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/nlp\/word_embeddings_tutorial.html<\/a>. If it's an accurate assumption that the embedding layer is being trained, then do I retrieve the learned weights through model.parameters() method or should I use the embedding.data.weight option?<\/p>\n","answers":[{"AnswerId":"53169760","CreationDate":"2018-11-06T10:10:37.767","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p><code>model.parameters()<\/code> returns all the <code>parameters<\/code> of your <code>model<\/code>, including the <code>embeddings<\/code>. <\/p>\n\n<p>So all these <code>parameters<\/code> of your <code>model<\/code> are handed over to the <code>optimizer<\/code> <em>(line below)<\/em> and will be trained later when calling <code>optimizer.step()<\/code> - so yes your <code>embeddings<\/code> are trained along with all other <code>parameters<\/code> of the network.<br><em>(you can also freeze certain layers by setting i.e. <code>embedding.weight.requires_grad = False<\/code>, but this is not the case here).<\/em><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># summing it up:\n# this line specifies which parameters are trained with the optimizer\n# model.parameters() just returns all parameters\n# embedding class weights are also parameters and will thus be trained\noptimizer = optim.SGD(model.parameters(), lr=0.001)\n<\/code><\/pre>\n\n<p>You can see that your <em>embedding weights<\/em> are also of type <code>Parameter<\/code> by doing so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nembedding_maxtrix = torch.nn.Embedding(10, 10)\nprint(type(embedding_maxtrix.weight))\n<\/code><\/pre>\n\n<p>This will output the type of the weights, which is <code>Parameter<\/code>:<\/p>\n\n<pre><code>&lt;class 'torch.nn.parameter.Parameter'&gt;\n<\/code><\/pre>\n\n<hr>\n\n<p>I'm not entirely sure what mean by <em>retrieve<\/em>. Do you mean getting a single vector, or do you want just the whole matrix to save it, or do something else?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>embedding_maxtrix = torch.nn.Embedding(5, 5)\n# this will get you a single embedding vector\nprint('Getting a single vector:\\n', embedding_maxtrix(torch.LongTensor([0])))\n# of course you can do the same for a seqeunce\nprint('Getting vectors for a sequence:\\n', embedding_maxtrix(torch.LongTensor([1, 2, 3])))\n# this will give the the whole embedding matrix\nprint('Getting weights:\\n', embedding_maxtrix.weight.data)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Getting a single vector:\n tensor([[-0.0144, -0.6245,  1.3611, -1.0753,  0.5020]], grad_fn=&lt;EmbeddingBackward&gt;)\nGetting vectors for a sequence:\n tensor([[ 0.9277, -0.1879, -1.4999,  0.2895,  0.8367],\n        [-0.1167, -2.2139,  1.6918, -0.3483,  0.3508],\n        [ 2.3763, -1.3408, -0.9531,  2.2081, -1.5502]],\n       grad_fn=&lt;EmbeddingBackward&gt;)\nGetting weights:\n tensor([[-0.0144, -0.6245,  1.3611, -1.0753,  0.5020],\n        [ 0.9277, -0.1879, -1.4999,  0.2895,  0.8367],\n        [-0.1167, -2.2139,  1.6918, -0.3483,  0.3508],\n        [ 2.3763, -1.3408, -0.9531,  2.2081, -1.5502],\n        [-0.5829, -0.1918, -0.8079,  0.6922, -0.2627]])\n<\/code><\/pre>\n\n<p>I hope this answers your question, you can also take a look at the documentation, there you can find some useful examples as well.<\/p>\n\n<p><em><a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Embedding\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Embedding<\/a><\/em><\/p>\n"}]}
+{"QuestionId":53125097,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2018-11-02T19:57:57.873","AcceptedAnswerId":53125798.0,"OwnerUserId":7953712.0,"Title":"Attempting to use uninitialized variable - tensorboard","Body":"<p>I am just starting to play around with Tensorboard, and want to create a simple example where I have a loop that calls a function.  Inside that function I have a tensor variable that gets incremented by one and then I add it to a summary.<\/p>\n\n<p>I am getting a FailedPreconditionError: Attempting to use uninitalized value x_scalar<\/p>\n\n<p>But I thought I was initializing the x_scalar with lines 10 and 14.  What is the proper way to initialize?<\/p>\n\n<pre><code>import tensorflow as tf\ntf.reset_default_graph()   # To clear the defined variables and operations of the previous cell\n\n# create the scalar variable\nx_scalar = tf.get_variable('x_scalar', shape=[], initializer=tf.truncated_normal_initializer(mean=0, stddev=1))\n\n# ____step 1:____ create the scalar summary\nfirst_summary = tf.summary.scalar(name='My_first_scalar_summary', tensor=x_scalar)\nstep = 1\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    writer = tf.summary.FileWriter('.\/graphs', sess.graph)\n    sess.run(x_scalar.assign(1))\n    print(sess.run(x_scalar))\n    print(\"---------------------------\")\ndef main():\n    global init\n    global first_summary\n    global step\n\n    # launch the graph in a session\n    # with tf.Session() as sess:\n        # # ____step 2:____ creating the writer inside the session\n        # writer = tf.summary.FileWriter('.\/graphs', sess.graph)\n    for s in range(100):\n        func()\n\ndef func():\n    global init\n    global first_summary\n    global step\n    global x_scalar\n\n    with tf.Session() as sess:\n        # ____step 2:____ creating the writer inside the session\n\n        # loop over several initializations of the variable\n        sess.run(x_scalar.assign(x_scalar + 1))\n        # ____step 3:____ evaluate the scalar summary\n        summary = sess.run(first_summary)\n        # ____step 4:____ add the summary to the writer (i.e. to the event file)\n        writer.add_summary(summary, step)\n        step = step + 1\n        print('Done with writing the scalar summary') \nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53125798","CreationDate":"2018-11-02T21:03:08.440","ParentId":null,"OwnerUserId":"7829500","Title":null,"Body":"<p>You initialized your variable in a different tf.Session(). When using your tf.Session() as a context manager the session automatically closes after the block of code has completed.<\/p>\n\n<p>You could use a checkpoint and metagraph to save your graph+weights and then loading them into your newly created session. <\/p>\n\n<p>Or you can try passing around a session<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run([CODE])\nsess.run([CODE])\nsess.run([CODE])\nsess.run([CODE])\nsess.close()\n<\/code><\/pre>\n\n<p>edited: made a correction<\/p>\n"}]}
+{"QuestionId":53125515,"AnswerCount":2,"Tags":"<google-app-engine><tensorflow>","CreationDate":"2018-11-02T20:37:24.863","AcceptedAnswerId":53192655.0,"OwnerUserId":3711940.0,"Title":"How to restore Tensorflow model from Google bucket without writing to filesystem?","Body":"<p>I have a 2gb Tensorflow model that I'd like to add to a Flask project I have on App Engine but I can't seem to find any documentation stating what I'm trying to do is possible.<\/p>\n\n<p>Since App Engine doesn't allow writing to the file system, I'm storing my model's files in a Google Bucket and attempting to restore the model from there. These are the files there:<\/p>\n\n<ul>\n<li>model.ckpt.data-00000-of-00001<\/li>\n<li>model.ckpt.index<\/li>\n<li>model.ckpt.meta<\/li>\n<li>checkpoint<\/li>\n<\/ul>\n\n<p>Working locally, I can just use<\/p>\n\n<pre><code>with tf.Session() as sess:\n    logger.info(\"Importing model into TF\")\n    saver = tf.train.import_meta_graph('model.ckpt.meta')\n    saver.restore(sess, model.ckpt)\n<\/code><\/pre>\n\n<p>Where the model is loaded into memory using Flask's <code>@before_first_request<\/code>.<\/p>\n\n<p>Once it's on App Engine, I assumed I could to this:<\/p>\n\n<pre><code>blob = bucket.get_blob('blob_name')\nfilename = os.path.join(model_dir, blob.name)\nblob.download_to_filename(filename)\n<\/code><\/pre>\n\n<p>Then do the same restore. But App Engine won't allow it.<\/p>\n\n<p>Is there a way to stream these files into Tensorflow's restore functions so the files don't have to be written to the file system?<\/p>\n","answers":[{"AnswerId":"53132622","CreationDate":"2018-11-03T15:13:20.147","ParentId":null,"OwnerUserId":"4495081","Title":null,"Body":"<p>I didn't actually use Tensorflow, the answer is based on docs and GAE-related knowledge.<\/p>\n\n<p>In general using GCS objects as files in GAE to avoid the lack of a writable filesystem access relies on one of 2 alternate approaches instead of just passing <strong>a filename<\/strong> to be directly read\/written (which can't be done with GCS objects) by your app code (and\/or any 3rd party utility\/library it may be using):<\/p>\n\n<ul>\n<li><p>using an already open <strong>file-like handler<\/strong> for reading\/writing the data from\/to GCS. Which your app would obtain from using either of:<\/p>\n\n<ul>\n<li>the <code>open<\/code> call from a GCS client library instead of the generic one typically used for a regular filesystem. See, for example <a href=\"https:\/\/stackoverflow.com\/questions\/39524581\/write-a-csv-to-store-in-google-cloud-storage\/39525052#39525052\">Write a CSV to store in Google Cloud Storage<\/a> or <a href=\"https:\/\/stackoverflow.com\/questions\/37778621\/pickling-python-objects-to-google-cloud-storage\/37778921#37778921\">pickling python objects to google cloud storage<\/a><\/li>\n<li>some in-memory faking of a file, using something like <code>StringIO<\/code>, see <a href=\"https:\/\/stackoverflow.com\/questions\/38885764\/how-to-zip-or-tar-a-static-folder-without-writing-anything-to-the-filesystem-in\/38888667#38888667\">How to zip or tar a static folder without writing anything to the filesystem in python?<\/a>. The in-memory fake file also gives easy access to the raw data in case it needs to be persisted in GCS, see below.<\/li>\n<\/ul><\/li>\n<li><p>directly using or producing just the respective <strong>raw data<\/strong> which your app would be entirely responsible for actually reading from\/writing to GCS (again using a GCS client library's <code>open<\/code> calls), see <a href=\"https:\/\/stackoverflow.com\/questions\/35708725\/how-to-open-gzip-file-on-gae-cloud\/35709022#35709022\">How to open gzip file on gae cloud?<\/a><\/p><\/li>\n<\/ul>\n\n<p>In your particular case it seems the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/import_meta_graph\" rel=\"nofollow noreferrer\"><code>tf.train.import_meta_graph()<\/code><\/a> call supports passing a <code>MetaGraphDef<\/code> protocol buffer (i.e. raw data) instead of the filename from which it should be loaded:<\/p>\n\n<blockquote>\n  <p>Args:<\/p>\n  \n  <ul>\n  <li><code>meta_graph_or_file<\/code>: <code>MetaGraphDef<\/code> protocol buffer or filename (including the path) containing a <code>MetaGraphDef<\/code>.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>So restoring models from GCS <em>should<\/em> be possible, something along these lines:<\/p>\n\n<pre><code>import cloudstorage\n\nwith cloudstorage.open('gcs_path_to_meta_graph_file', 'r') as fd:\n    meta_graph = fd.read()\n\n# and later:\n\nsaver = tf.train.import_meta_graph(meta_graph)\n<\/code><\/pre>\n\n<p>However from the quick doc scan saving\/checkpointing the modes back to GCS may be tricky, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\/builder\/SavedModelBuilder#save\" rel=\"nofollow noreferrer\"><code>save()<\/code><\/a> seem to want to want to write the data to disk itself. But I didn't dig too deep.<\/p>\n"},{"AnswerId":"53192655","CreationDate":"2018-11-07T15:33:00.123","ParentId":null,"OwnerUserId":"3711940","Title":null,"Body":"<p>After some tips from Dan Cornilescu and digging into it I found that Tensorflow builds the <code>MetaGraphDef<\/code> with a function called <code>ParseFromString<\/code>, so here's what I ended up doing:<\/p>\n\n<pre><code>from google.cloud import storage\nfrom tensorflow import MetaGraphDef\n\nclient = storage.Client()\nbucket = client.get_bucket(Config.MODEL_BUCKET)\nblob = bucket.get_blob('model.ckpt.meta')\nmodel_graph = blob.download_as_string()\n\nmgd = MetaGraphDef()\nmgd.ParseFromString(model_graph)\n\nwith tf.Session() as sess:\n    saver = tf.train.import_meta_graph(mgd)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53125564,"AnswerCount":1,"Tags":"<python><python-3.x><keras><keras-layer>","CreationDate":"2018-11-02T20:42:39.580","AcceptedAnswerId":null,"OwnerUserId":2943311.0,"Title":"How do I \"Flatten\" a model in a model in Keras","Body":"<p>So lets say I have some Keras model I have built that I like:<\/p>\n\n<pre><code>from keras.layers import Input, Dense, BatchNormalization, Subtract\nfrom keras.models import Model\n\ninput_layer = Input((10, ))\nx = Dense(5)(input_layer)\noutput_layer = Dense(10)(x)\n\nmodel = Model(input_layer, output_layer)\n<\/code><\/pre>\n\n<p>I can learn about my model with the <code>summary()<\/code> function and I get:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 10)                0\n_________________________________________________________________\ndense_1 (Dense)              (None, 5)                 55\n_________________________________________________________________\ndense_2 (Dense)              (None, 10)                60\n=================================================================\nTotal params: 115\nTrainable params: 115\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Now I want to try out adding some pre and post processing steps to the model so for example I might do the following:<\/p>\n\n<pre><code># Add preprocessing layers\nnew_input = Input((10,))\nx = BatchNormalization()(new_input)\nmodel.layers.pop(0) # remove original input\nx = model(x)\n# Change the model to residual modeling with a subtract layer\nnew_output = Subtract()([new_input, x])\nnew_model = Model(new_input, new_output)\n<\/code><\/pre>\n\n<p>However now when I call <code>summary()<\/code> I only learn about the pre and post processing layers, not my original model:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to\n==================================================================================================\ninput_2 (InputLayer)            (None, 10)           0\n__________________________________________________________________________________________________\nbatch_normalization_1 (BatchNor (None, 10)           40          input_2[0][0]\n__________________________________________________________________________________________________\nmodel_1 (Model)                 (None, 10)           115         batch_normalization_1[0][0]\n__________________________________________________________________________________________________\nsubtract_1 (Subtract)           (None, 10)           0           input_2[0][0]\n                                                                 model_1[1][0]\n==================================================================================================\nTotal params: 155\nTrainable params: 135\nNon-trainable params: 20\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>This means that internally, rather than having each layer from the first model get added to the <code>new_model.layers<\/code> of the new model, instead the whole model is getting added as a single element to the <code>new_model.layers<\/code> list.  I want it to actually have that single element get flattened out so the summary will look more like this:<\/p>\n\n<hr>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to\n==================================================================================================\ninput_2 (InputLayer)            (None, 10)           0\n__________________________________________________________________________________________________\nbatch_normalization_1 (BatchNor (None, 10)           40          input_2[0][0]\n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 5)            55          batch_normalization_1[0][0]\n__________________________________________________________________________________________________\ndense_2 (Dense)                 (None, 10)           60          dense_1[0]\n__________________________________________________________________________________________________\nsubtract_1 (Subtract)           (None, 10)           0           input_2[0][0]\n                                                                 dense_2[1][0]\n==================================================================================================\nTotal params: 155\nTrainable params: 135\nNon-trainable params: 20\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>Why do I care?  Basically I am trying out some different modeling approaches where I can try out different combinations of pre and post processing along with different base models and see how that impacts my results.  However it is really hard to do the comparative analysis when the information about the base model and its parameters are all wrapped up in a single \"model\" layer of the wrapper model that just holds information about the pre and post processing.<\/p>\n","answers":[{"AnswerId":"53126428","CreationDate":"2018-11-02T22:14:01.980","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>This is the correct behaviour of Keras since <code>Model<\/code> actually inherits from <code>Layer<\/code>. You can wrap the base model into a function to stop it from wrapping into a <code>Model<\/code>:<\/p>\n\n<pre><code>def base_model(input=None):\n  input_layer = input or Input((10, ))\n  x = Dense(5)(input_layer)\n  output_layer = Dense(10)(x)\n  if input is None:\n    return Model(input_layer, output_layer)\n  return output_layer\n# Then use it\n# Add preprocessing layers\nnew_input = Input((10,))\nx = BatchNormalization()(new_input)\nx = base_model(x)\n# Change the model to residual modeling with a subtract layer\nnew_output = Subtract()([new_input, x])\nnew_model = Model(new_input, new_output)\n<\/code><\/pre>\n\n<p>If you pass an existing layer it will not wrap, if you don't it will return you a <code>Model<\/code> instance.<\/p>\n"}]}
+{"QuestionId":53125730,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><nlp>","CreationDate":"2018-11-02T20:55:58.227","AcceptedAnswerId":null,"OwnerUserId":10199563.0,"Title":"How to save list of lists of varying length in a TFRecord?","Body":"<p>I have a dataset which is made up of nested sequences, so for example one data point looks like this: <\/p>\n\n<pre><code>p1 = [[1,1,1], [2,2], [3,3,3]]\n<\/code><\/pre>\n\n<p>but another data point could look like this:<\/p>\n\n<pre><code>p2 = [[1,1], [2,2,2,2,2]]\n<\/code><\/pre>\n\n<p>I am not sure how I can save lists of lists in TFRecords as all the examples I found only save <strong>non-nested<\/strong> lists of varying length. I plan to then uses padded_batch to create 3D tensors of the shape \n<code>[batch size, max number of lists, max number of tokens in a list]<\/code><\/p>\n","answers":[{"AnswerId":"53128236","CreationDate":"2018-11-03T03:46:04.473","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>One option is to just store each value sequentially. For instance, your first example would be stored as <code>[1,1,1,2,2,3,3,3]<\/code>. To reproduce the original sequence, you can store the original lengths as well - <code>[3,2,3]<\/code> and then reconstruct in preprocessing. I recommend padding the nested sequences from the start so that you can go to\/from flat\/nested with a simple reshape operation.<\/p>\n"}]}
+{"QuestionId":53125854,"AnswerCount":0,"Tags":"<keras><deep-learning>","CreationDate":"2018-11-02T21:10:28.327","AcceptedAnswerId":null,"OwnerUserId":10598190.0,"Title":"adding conv1d layer in my model gives me a dimension error","Body":"<p>I am new to python and deep learning. I am trying to make a Keras sequential model in which I am adding a convolutional 1d layer for a RNA sequence of length 20 bases, which I have hard coded as numbers using the Keras hashing_trick function. My input is in pandas dataframe and looks like below<\/p>\n\n<pre><code> b1  b2  b3  b4  b5  b6  b7  b8  b9  b10  b11  b12  b13  b14  b15  b16  \\\n<\/code><\/pre>\n\n<p>0     2   2   4   2   3   2   4   2   2    3    2    4    2    4    2    2<br>\n1     2   3   3   3   4   4   4   4   3    3    2    3    2    2    2    2<br>\n2     2   2   3   3   3   2   2   4   3    4    3    3    3    3    2    2<br>\n3     2   4   2   2   2   2   3   3   2    2    2    3    2    2    2    2<br>\n4     3   4   4   3   2   2   3   2   4    2    2    2    2    2    2    4<br>\n5     2   3   2   2   2   3   2   2   2    3    4    2    2    2    2    4<br>\n6     2   4   2   4   4   2   2   2   2    2    2    2    2    2    3    2<br>\n7     2   4   3   3   3   4   2   2   4    4    4    4    3    2    3    3<br>\n8     2   3   2   2   2   2   4   2   2    4    2    2    2    2    2    4<br>\n9     3   2   2   2   2   4   2   3   2    2    2    2    4    2    4    2<br>\n10    2   2   4   2   2   2   2   2   2    2    2    2    3    4    2    2<br>\n11    2   2   3   2   2   2   3   4   4    2    2    4    2    2    2    4 <\/p>\n\n<p>Now my data frame shape is (500 rows,20 columns) and my labels in (500 rows,1 column)<\/p>\n\n<p>So when I am adding the conv1D layer and specifying the filter as 32 and kernel size as 5, am I doing it right?? because I am getting the error <\/p>\n\n<p>ValueError: Input 0 is incompatible with layer conv1d_29: expected ndim=3, found ndim=2<\/p>\n\n<p>Since I am not using the conv1d layer as first I am not specifying the shape<\/p>\n\n<p>What am I doing wrong<\/p>\n\n<p>please find the code snippet below<\/p>\n\n<pre><code>data = pd.read_csv(\"Sequence_file.csv\")\ntest = pd.read_csv(\"test_file.csv\")\ntdf = test[['b1','b2','b3','b4','b5','b6','b7','b8','b9','b10','b11','b12','b13','b14','b15','b16','b17','b18','b19','b20']]\ndf1 = data[['b1','b2','b3','b4','b5','b6','b7','b8','b9','b10','b11','b12','b13','b14','b15','b16','b17','b18','b19','b20']]\nlabel = data[['status']]\ntlabel = test[['status']]\nmodel = Sequential()\nmodel.add(Dense(32,activation='relu', input_dim=20))\nmodel.add(Dropout(0.5))\nmodel.add(Conv1D(32,5, activation='relu',padding='same',use_bias=True))\nmodel.add(Dense(64,activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\nsgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='binary_crossentropy', optimizer=sgd, metrics=['accuracy'])\nmodel.fit(df1,label,epochs=10,batch_size=32)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53126147,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda><jupyter-notebook>","CreationDate":"2018-11-02T21:42:36.753","AcceptedAnswerId":null,"OwnerUserId":1610428.0,"Title":"Tensorflow: Trouble importing Tensorflow in a Jupyter Notebook","Body":"<p>I have a <code>conda<\/code> environment with <code>Tensorflow-gpu<\/code> installed for python 3.6.6. I am running Ubuntu linux 16.04 LTS. When I activate the environment and start a python shell, I can import <code>Tensorflow<\/code> just fine. But when I start up a <code>Jupyter<\/code> notebook or <code>Jupyter lab<\/code> session, I can't import <code>Tensorflow<\/code>. <\/p>\n\n<p>Now there is a pre-existing SE <a href=\"https:\/\/stackoverflow.com\/questions\/37061089\/trouble-with-tensorflow-in-jupyter-notebook\">post<\/a> on this issue. However, that post is 2 years old, so there are changes in Jupyter since then. Second, I tried the fixed mentioned in that post, and none of them have worked. So I was hoping there might be some progress or some new ways to fix this. <\/p>\n\n<p>So let me provide a little more detail on the issue. <\/p>\n\n<ol>\n<li><p>First, when I activated a python 3.6.6 environment in anaconda and did <code>conda list<\/code> in a regular Tilix terminal session, I can see that <code>Tensorflow-gpu<\/code> is listed.<\/p><\/li>\n<li><p>When I start <code>ipython<\/code> in a standard Tilix terminal and type <code>import tensorflow as tf<\/code>, that works fine. <\/p><\/li>\n<li><p>Next, I started a jupyter lab session and opened a terminal from the jupyter launcher. In the terminal I started an <code>ipython<\/code> session and tried to import tensorflow. That does not work. <\/p><\/li>\n<li><p>I opened a new jupyter notebook and tried to import <code>Tensorflow<\/code> then I \"ModuleNotFoundError: No module named 'tensorflow'\" error. <\/p><\/li>\n<li><p>If I just run a python file with tensorflow code from the regular tilix terminal, then it will run fine and find Tensorflow. So only the notebook seems to be the big problem. <\/p><\/li>\n<\/ol>\n\n<p>Things that I have tried:<\/p>\n\n<p>I did try to implement a few of the fixes suggested in the older SE post, as well as the closed <a href=\"https:\/\/github.com\/jupyter\/notebook\/issues\/2120\" rel=\"nofollow noreferrer\">github issue<\/a> on this problem. <\/p>\n\n<ol>\n<li>I tried to reinstall <code>tensorflow<\/code><\/li>\n<li>I did a conda install of <code>ipython<\/code>, <code>jupyter<\/code> because some folks suggested installing those into the environment directly. <\/li>\n<li>I looked at the <code>jupyter kernelspec list<\/code>. Some folks suggested making a change to this, but it was not clear what to change. <\/li>\n<li>I installed the <code>conda install -c anaconda-nb-extensions nb_conda<\/code> in the environment too, but still no luck. <\/li>\n<\/ol>\n\n<p>Anyone else have a suggestion about how to fix this.<\/p>\n","answers":[]}
+{"QuestionId":53126939,"AnswerCount":1,"Tags":"<opencv><tensorflow>","CreationDate":"2018-11-02T23:26:39.473","AcceptedAnswerId":53177921.0,"OwnerUserId":10581999.0,"Title":"error loading faster_rcnn models into Opencv","Body":"<p>I followed exactly every step from TensorFlow Object Detection API and trained the faster_rcnn_resnet50 model. Then I referenced link:Wiki to generate the pbtxt file for cv2 read net. <\/p>\n\n<p>When I ran the model using opencv, it gave no error most of the time and this error sometimes:<\/p>\n\n<blockquote>\n  <p>cv2.error: OpenCV(3.4.3)\n  C:\\projects\\opencv-python\\opencv\\modules\\dnn\\src\\tensorflow\\tf_importer.cpp:495:\n  error: (-2:Unspecified error) Input layer not found:\n  CropAndResize\/Reshape\/shape in function\n  'cv::dnn::experimental_dnn_34_v7::anonymous-namespace'::TFImporter::connect'<\/p>\n<\/blockquote>\n\n<p>I tried the optimized tool and re-generated the pbtxt file, but still experienced the same. <\/p>\n\n<p>Any suggestions to make the model work? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"53177921","CreationDate":"2018-11-06T18:32:21.917","ParentId":null,"OwnerUserId":"10581999","Title":null,"Body":"<p>Got answer from @dkurt at <a href=\"https:\/\/github.com\/opencv\/opencv\/issues\/13050\" rel=\"nofollow noreferrer\">https:\/\/github.com\/opencv\/opencv\/issues\/13050<\/a> <\/p>\n\n<p>Wish that can help if you are experiencing the same problem <\/p>\n"}]}
+{"QuestionId":53127232,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T00:12:04.443","AcceptedAnswerId":53128220.0,"OwnerUserId":3259896.0,"Title":"Most space\/memory efficient way to save Tensorflow model for prediction only?","Body":"<p>I have a huge Tensorflow model (the checkpoint file is 4-5 gbs). I was wondering if there's a different way to save Tensorflow models, besides the checkpoint way, that is space\/memory efficient. <\/p>\n\n<p>I know that a checkpoint file also saves all the optimizer gradients, so maybe those can be cut out too. <\/p>\n\n<p>My model is very simple, just two matrices of embeddings, perhaps I can only save those matrices to .npy directly?<\/p>\n","answers":[{"AnswerId":"53128220","CreationDate":"2018-11-03T03:42:51.197","ParentId":null,"OwnerUserId":"2117367","Title":null,"Body":"<p>What you want to do with the checkpoint is to <em>freeze<\/em> it. Check out <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/model_files#freezing\" rel=\"nofollow noreferrer\">this page<\/a> from tensorflow's official documentation.<\/p>\n\n<p>The freezing process strips off all extraneous information from the checkpoint that isn't used for forward inference. Tensorflow provides an easy to use script for it called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow noreferrer\">freeze_graph.py<\/a>.<\/p>\n"}]}
+{"QuestionId":53127718,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-11-03T01:54:16.153","AcceptedAnswerId":null,"OwnerUserId":358614.0,"Title":"Keras - Interference in models","Body":"<p>I am developing something in Django and have two models in two different views. One model has nothing to do with another, but both are called one after another.<\/p>\n\n<p><strong>Model 1:<\/strong><\/p>\n\n<pre><code>    classificador = Sequential()\n    classificador.add(Dense(units=60, activation=\"relu\", kernel_initializer=\"random_uniform\", input_shape=(38,)))\n    classificador.add(Dense(units=60, activation=\"relu\", kernel_initializer=\"random_uniform\"))\n    classificador.add(Dense(units=1, activation=\"sigmoid\"))\n    classificador.compile(optimizer=\"adam\", loss=\"binary_crossentropy\", metrics=[\"binary_accuracy\"])\n    classificador.fit(X_Treino, Y_Treino, batch_size=10, epochs=100)\n<\/code><\/pre>\n\n<p><strong>Model 2<\/strong><\/p>\n\n<pre><code>    classificador_outlier = Sequential()\n    classificador_outlier.add(Dense(units=45, activation=\"relu\", kernel_initializer=\"random_uniform\", input_shape=(6,)))\n    classificador_outlier.add(Dense(units=45, activation=\"relu\", kernel_initializer=\"random_uniform\"))\n    classificador_outlier.add(Dense(units=1, activation=\"sigmoid\"))\n    classificador_outlier.compile(optimizer=\"adam\", loss=\"binary_crossentropy\", metrics=[\"binary_accuracy\"])\n    classificador_outlier.fit(X_Treino_Outlier, Y_Treino_Outlier, batch_size=10, epochs=100)\n<\/code><\/pre>\n\n<p>I think one model is interfering in another. If I run just one model (anyone), it works well. When running both, I receive the following message running <code>Model 1<\/code> (last one to run):<\/p>\n\n<blockquote>\n  <p>Matrix size-incompatible: In[0]: [10,38], In[1]: [6,45]<\/p>\n<\/blockquote>\n\n<p>Observe that the input of <code>Model 1<\/code> has (10, 38) shape, and message error is referecing values that are related to  <code>Model 2<\/code>.<\/p>\n\n<p>What could it be?<\/p>\n","answers":[{"AnswerId":"53128214","CreationDate":"2018-11-03T03:42:11.530","ParentId":null,"OwnerUserId":"358614","Title":null,"Body":"<p>It worked after clearing session, as sugested <a href=\"https:\/\/stackoverflow.com\/questions\/45063602\/attempting-to-reset-tensorflow-graph-when-using-keras-failing\">here<\/a><\/p>\n\n<pre><code>from keras import backend as K\nK.clear_session()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53127841,"AnswerCount":1,"Tags":"<python><keras><keras-layer>","CreationDate":"2018-11-03T02:22:19.663","AcceptedAnswerId":null,"OwnerUserId":7112888.0,"Title":"Graph disconnect - keras","Body":"<h1>Relevant Code<\/h1>\n\n<p>feel free to ask for more if there are missing portions of code.<\/p>\n\n<pre><code>def get_decoder_layers(char_list, encoder_state=None):\n    input = layers.Input(shape=(None, len(char_list)))\n\n    cells = []\n    for hidden in [256, 256]:\n        cells.append(layers.GRUCell(hidden))\n\n    decoder = layers.RNN(cells, return_state=True, return_sequences=True)\n    state = decoder(input, initial_state=encoder_state)\n    output = state[0] # only get the output\n    dense = layers.Dense(len(char_list), activation='linear')\n    output = dense(output)\n\n    return input, output, decoder, dense\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>def get_encoder_layers(char_list):\n    input = layers.Input(shape=(None, len(char_list)))\n\n    cells = []\n    for hidden in [256, 256]:\n        cells.append(layers.GRUCell(hidden))\n\n    encoder = layers.RNN(cells, return_state=True)\n    output = encoder(input)\n    states = output[1:] # get rid of the ouput\n\n    return input, states\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>decoder_input, decoder_output, decoder, dense = m.get_decoder_layers(french_chars)\nencoder_input, encoder_states = m.get_encoder_layers(english_chars)\n<\/code><\/pre>\n\n<hr>\n\n<p>Error is here:<\/p>\n\n<pre><code>encoder_inference_model, decoder_inference_model = m.get_inference_models(\n        encoder_input=encoder_input,\n        decoder_input=decoder_input,\n        states=encoder_states,\n        decoder=decoder,\n        dense_layer=dense)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>def get_inference_models(encoder_input=None, decoder_input=None, states=None, decoder=None, dense_layer=None):\n    encoder_inference = keras.Model(encoder_input, states)\n\n    decoder_state_input_a = layers.Input(shape=(256,)) # these have to be different variables\n    decoder_state_input_b = layers.Input(shape=(256,))\n\n    decoder_output_states = decoder(decoder_input, initial_state=states)\n\n    output = decoder_output_states[0] # the first element is the output\n    decoder_states = decoder_output_states[1:] # the rest are states\n\n    output = dense_layer(output)\n\n    return encoder_inference, keras.Model(inputs=[decoder_input, decoder_state_input_a, decoder_state_input_b], outputs=[output] + decoder_states) # error is at `keras.Model`\n<\/code><\/pre>\n\n<h3>Error<\/h3>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/Applications\/PyCharm.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1664, in &lt;module&gt;\n    main()\n  File \"\/Applications\/PyCharm.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1658, in main\n    globals = debugger.run(setup['file'], None, None, is_module)\n  File \"\/Applications\/PyCharm.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1068, in run\n    pydev_imports.execfile(file, globals, locals)  # execute the script\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/main.py\", line 51, in &lt;module&gt;\n    dense_layer=dense)\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/model.py\", line 86, in get_decoder_inference\n    return encoder_inference, keras.Model(inputs=[decoder_input, decoder_state_input_a, decoder_state_input_b], outputs=[output] + decoder_states)\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/venv\/lib\/python2.7\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/venv\/lib\/python2.7\/site-packages\/keras\/engine\/network.py\", line 93, in __init__\n    self._init_graph_network(*args, **kwargs)\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/venv\/lib\/python2.7\/site-packages\/keras\/engine\/network.py\", line 231, in _init_graph_network\n    self.inputs, self.outputs)\n  File \"\/Users\/zoe\/Developer\/CoreText\/Lis\/venv\/lib\/python2.7\/site-packages\/keras\/engine\/network.py\", line 1443, in _map_graph_network\n    str(layers_with_complete_input))\nValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_1:0\", shape=(?, ?, 31), dtype=float32) at layer \"input_1\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>The error indicates that some of my layers are disconnected but, it looks like all my layers are connected (except for <code>decoder_state_input_a<\/code> and <code>decoder_state_input_b<\/code> which I <em>think<\/em> is okay). <\/p>\n","answers":[{"AnswerId":"53127878","CreationDate":"2018-11-03T02:29:07.903","ParentId":null,"OwnerUserId":"7112888","Title":null,"Body":"<p>I did actually need to connect <code>decoder_state_input_a<\/code> and <code>decoder_state_input_b<\/code>.<\/p>\n\n<p>this:<\/p>\n\n<pre><code>decoder_output_states = decoder(decoder_input, initial_state=states)\n<\/code><\/pre>\n\n<p>needed to be this:<\/p>\n\n<pre><code>decoder_output_states = decoder(decoder_input, initial_state=[decoder_state_input_a, decoder_state_input_b])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53128203,"AnswerCount":0,"Tags":"<machine-learning><neural-network><keras><text-classification><rnn>","CreationDate":"2018-11-03T03:41:11.507","AcceptedAnswerId":null,"OwnerUserId":10588684.0,"Title":"Maybe Overfitting trying to implement rnn with keras","Body":"<p>I am trying to create a RNN for sentiment classification (-1(neg) or 1(pos)).My dataset file looks like this:\ntext,polarity\n\"this is a line\",1\n\"this is  asecond line\",-1\n.....\netc.\nwith 2603 rows like these.\nThis is what i got so far and it reaches maximum accuracy very fast, so i assume it is wrong. What am i doing wrong though? <\/p>\n\n<p>Code:    <\/p>\n\n<pre><code>from __future__ import print_function\nfrom keras.preprocessing import sequence\nimport tensorflow as tf\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Embedding, Activation, Dropout\nfrom keras.preprocessing.text import Tokenizer\nfrom keras.layers import LSTM\nfrom keras.datasets import imdb\nimport pandas as pd\nimport string\nimport nltk\nfrom sklearn.model_selection import train_test_split\nfrom collections import Counter\n\n\nbatch_size = 32\nnum_words=2000\nepochs=15\n\n\nprint('Loading data...')\ndata = pd.read_csv(\"yolo.csv\",header=0,encoding = 'UTF-8')\nX = data['text']\nY = data['polarity']\n\n\nx_train,x_test,y_train,y_test = train_test_split(X,Y,test_size=0.2)  #split train\/test data\n\ntokenizer = Tokenizer(num_words=num_words, lower=False, split=' ')\n\ntokenizer.fit_on_texts(x_train)\nx_train= tokenizer.texts_to_sequences(x_train)\nx_test= tokenizer.texts_to_sequences(x_test)\n\nx_train= tokenizer.sequences_to_matrix(x_train, mode='binary')\nx_test= tokenizer.sequences_to_matrix(x_test, mode='binary')\n\ny_train = tf.keras.utils.to_categorical(y_train,2)\ny_test = tf.keras.utils.to_categorical(y_test,2)\n\n\nprint('x_train shape:', x_train.shape)\nprint('x_test shape:', x_test.shape)\n\n\n\nprint('Build model...')\nmodel = Sequential()\nmodel.add(Dense(512, input_shape=(num_words, )))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(2))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nprint (model.metrics_names)\nhistory= model.fit (x_train,y_train, batch_size=batch_size, epochs=epochs, verbose=1)\nscore= model.evaluate(x_test, y_test, batch_size=batch_size, verbose=1)\n\nprint ('Test loss: ()'.format(score[0]))\nprint ('Test accuracy:{}'.format(score[1]))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53128407,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T04:28:33.653","AcceptedAnswerId":null,"OwnerUserId":4258928.0,"Title":"Tensorflow. NxD matrix multiply NxD matrix to give NxDxD matrix","Body":"<p>for example<\/p>\n\n<pre><code>   A=[[3,2],\n      [1,3],\n      [4,3]]\n\n   B=[[4,1],\n      [2,1],\n      [2,4]]\n<\/code><\/pre>\n\n<p>For each row vector of matrix; I want to perform column-row multiplication<\/p>\n\n<pre><code>result = []\n\nfor i in range(3):\n    x = tf.matmul(tf.reshape(A[i],[2,1]), tf.reshape(B[i],[1,2])) # gives 2x2 matrix\n    result.append(x)\n\ntf.stack(result)\n<\/code><\/pre>\n\n<p>I'm wondering if there is a more efficient way of completely get rid off the loop as in practical I need loop thousands of times<\/p>\n","answers":[{"AnswerId":"53128578","CreationDate":"2018-11-03T05:08:13.763","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>Luckily, Tensorflow supports \"batched\" matrix multiplies, so the following should be sufficient:<\/p>\n\n<pre><code>A = tf.placeholder(tf.float32, [3,2])\nB = tf.placeholder(tf.float32, [3,2])\nAA = tf.reshape(A,[3,2,1])\nBB = tf.reshape(B,[3,1,2])\nprint(tf.matmul(AA,BB))\n&gt;&gt;&gt; &lt;tf.Tensor 'MatMul_1:0' shape=(3, 2, 2) dtype=float32&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":53128629,"AnswerCount":1,"Tags":"<keras><theano><theano-cuda>","CreationDate":"2018-11-03T05:19:42.603","AcceptedAnswerId":null,"OwnerUserId":8016385.0,"Title":"AttributeError: 'module' object has no attribute 'cuda'","Body":"<p>I was trying to run this repository: <a href=\"https:\/\/github.com\/WaqasSultani\/AnomalyDetectionCVPR2018\" rel=\"nofollow noreferrer\">https:\/\/github.com\/WaqasSultani\/AnomalyDetectionCVPR2018<\/a>\n In the Test_Anomaly_Detector_public.py I am stuck with error:theano.sandbox.cuda.use('gpu0')\nAttributeError: 'module' object has no attribute 'cuda'.\nI am using theano as backend<\/p>\n\n<p>This is Test_Anomaly_Detector_public.py:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation\nfrom keras.regularizers import l2\nfrom keras.optimizers import SGD ,Adagrad\nfrom scipy.io import loadmat, savemat\nfrom keras.models import model_from_json\nimport theano.tensor as T\nimport theano\nimport csv\nimport ConfigParser\nimport collections\nimport time\nimport csv\nimport os\nfrom os import listdir\nimport skimage.transform\nfrom skimage import color\nfrom os.path import isfile, join\nimport numpy as np\nimport numpy\nfrom datetime import datetime\nfrom scipy.spatial.distance import cdist,pdist,squareform\nimport theano.sandbox\nimport shutil\ntheano.sandbox.cuda.use('gpu0')\n\n\n\n\nseed = 7\nnumpy.random.seed(seed)\n\n\ndef load_model(json_path):  # Function to load the model\n    model = model_from_json(open(json_path).read())\n    return model\n\ndef load_weights(model, weight_path):  # Function to load the model weights\n    dict2 = loadmat(weight_path)\n    dict = conv_dict(dict2)\n    i = 0\n    for layer in model.layers:\n        weights = dict[str(i)]\n        layer.set_weights(weights)\n        i += 1\n    return model\n\ndef conv_dict(dict2):\n    i = 0\n    dict = {}\n    for i in range(len(dict2)):\n        if str(i) in dict2:\n            if dict2[str(i)].shape == (0, 0):\n                dict[str(i)] = dict2[str(i)]\n            else:\n                weights = dict2[str(i)][0]\n                weights2 = []\n                for weight in weights:\n                    if weight.shape in [(1, x) for x in range(0, 5000)]:\n                        weights2.append(weight[0])\n                    else:\n                        weights2.append(weight)\n                dict[str(i)] = weights2\n    return dict\n\n# Load Video\n\ndef load_dataset_One_Video_Features(Test_Video_Path):\n\n    VideoPath =Test_Video_Path\n    f = open(VideoPath, \"r\")\n    words = f.read().split()\n    num_feat = len(words) \/ 4096\n    # Number of features per video to be loaded. In our case num_feat=32, as we divide the video into 32 segments. Note that\n    # we have already computed C3D features for the whole video and divided the video features into 32 segments.\n\n    count = -1;\n    VideoFeatues = []\n    for feat in xrange(0, num_feat):\n        feat_row1 = np.float32(words[feat * 4096:feat * 4096 + 4096])\n        count = count + 1\n        if count == 0:\n            VideoFeatues = feat_row1\n        if count &gt; 0:\n            VideoFeatues = np.vstack((VideoFeatues, feat_row1))\n    AllFeatures = VideoFeatues\n\n    return  AllFeatures\n\n\n\nprint(\"Starting testing...\")\n\n\nAllTest_Video_Path = '\/newdata\/UCF_Anomaly_Dataset\/Dataset\/CVPR_Data\/C3D_Complete_Video_txt\/Test\/'\n# AllTest_Video_Path contains C3D features (txt file)  of each video. Each file contains 32 features, each of 4096 dimensions.\nResults_Path = '..\/Eval_Res\/'\n# Results_Path is the folder where you can save your results\nModel_dir='..\/Trained_AnomalyModel\/'\n# Model_dir is the folder where we have placed our trained weights\nweights_path = Model_dir + 'weights_L1L2.mat'\n# weights_path is Trained model weights\n\nmodel_path = Model_dir + 'model.json'\n\nif not os.path.exists(Results_Path):\n       os.makedirs(Results_Path)\n\nAll_Test_files= listdir(AllTest_Video_Path)\nAll_Test_files.sort()\n\nmodel=load_model(model_path)\nload_weights(model, weights_path)\nnVideos=len(All_Test_files)\ntime_before = datetime.now()\n\nfor iv in range(nVideos):\n\n    Test_Video_Path = os.path.join(AllTest_Video_Path, All_Test_files[iv])\n    inputs=load_dataset_One_Video_Features(Test_Video_Path) # 32 segments features for one testing video\n    predictions = model.predict_on_batch(inputs)   # Get anomaly prediction for each of 32 video segments.\n    aa=All_Test_files[iv]\n    aa=aa[0:-4]\n    A_predictions_path = Results_Path + aa + '.mat'  # Save array of 1*32, containing anomaly score for each segment. Please see Evaluate Anomaly Detector to compute  ROC.\n    print \"Total Time took: \" + str(datetime.now() - time_before)\n<\/code><\/pre>\n\n<p>My .theanorc file:<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = cuda0\n\n[gpuarray]\npreallocate = 1\n<\/code><\/pre>\n","answers":[{"AnswerId":"53799735","CreationDate":"2018-12-16T05:39:25.890","ParentId":null,"OwnerUserId":"7990199","Title":null,"Body":"<p>You can comment out this line. When you run please follow this<\/p>\n\n<pre><code>THEANO_FLAGS=mode=FAST_RUN,device=cuda0,floatX=float32 python [...]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53128752,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-03T05:43:27.227","AcceptedAnswerId":53131883.0,"OwnerUserId":7467476.0,"Title":"Error while running Convolutional Autoencoder RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn","Body":"<p>I am a noob and am creating a model in PyTorch for the first time. I am trying to create a convolutional autoencoder and am getting the error while running the model. The code I am using is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyDataset(Dataset):\n    def __init__(self, image_paths, target_paths, train=True):\n        self.image_paths = image_paths\n        self.target_paths = target_paths\n\n    def transform(self, image, target):\n        # Transform to tensor\n        resize = transforms.Resize(size=(2350,1650))\n        image = resize(image)\n        target = resize(target)\n        grayscale = transforms.Grayscale(1)\n        image = grayscale(image)\n        target = grayscale(target)\n        image = TF.to_tensor(image)\n        target = TF.to_tensor(target)\n        return image, target\n\n    def __getitem__(self, index):\n        image = Image.open(self.image_paths[index])\n        target = Image.open(self.target_paths[index])\n        x, y = self.transform(image, target)\n        return x, y\n\n    def __len__(self):\n        return len(self.image_paths)\n\ntraindata = MyDataset(image_paths=train_data, target_paths=target_data, train=True)\ntestdata = MyDataset(image_paths=test_data, target_paths=None, train=False)\n\ntrain_loader = DataLoader(traindata, batch_size=1, shuffle=True, num_workers=4)\ntest_loader = DataLoader(testdata, batch_size=1, shuffle=False, num_workers=4)\n\nclass ConvolutionalAutoEncoder(nn.Module):\n    def __init__(self):\n        super(ConvolutionalAutoEncoder, self).__init__()\n        self.encoder_block1 = nn.Sequential(\n            nn.Conv2d(1, 64, 3, stride=1, padding=1),\n            nn.ReLU(True),\n            nn.Conv2d(64, 64, 3, stride=1, padding=1),\n            nn.ReLU(True)\n        )\n        self.decoder_block1 = nn.Sequential(   \n            nn.ConvTranspose2d(64, 64, 3, stride=1, padding=1),\n            nn.ReLU(True),\n            nn.ConvTranspose2d(64, 64, 3, stride=1, padding=1),\n            nn.ReLU(True)\n         )\n        self.decoder_block0 = nn.Sequential(  \n            nn.ConvTranspose2d(64, 1, 3, stride=1, padding=1),\n            nn.Sigmoid()\n        )\n    def forward(self, x):\n        x1 = self.encoder_block1(x)\n        y1 = self.decoder_block1(x1)\n        y0 = self.decoder_block0(y1)\n        return x\n\ndevice = torch.device(\"cuda:2\" if torch.cuda.is_available() else \"cpu\")\nprint(device)\n\nmodel = ConvolutionalAutoEncoder().to(device)\n# Loss and optimizer\nlearning_rate = 0.001\ncriterion = nn.MSELoss()\noptimizer = optim.Adam(model.parameters(), lr=learning_rate)\n\nparams = list(model.parameters())\nprint(len(params))\nprint(params[0].size())  # conv1's .weight\n\nnum_epochs = 30\ntotal_step = len(train_loader)\nfor epoch in range(num_epochs):\n    for batch_idx, data in enumerate(train_loader):\n        inp, targ = data\n        inp = inp.to(device)\n        targ = targ.to(device)\n\n        output = model(inp)\n        loss = criterion(output, targ)\n\n        model.zero_grad()\n        loss.backward()\n        optimizer.step()\n\n        # print statistics\n        running_loss += loss.item()\n\n        if (batch_idx+1) % 10 == 0:\n            print ('Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}' \n                   .format(epoch+1, num_epochs, i+1, total_step, loss.item()))\n<\/code><\/pre>\n\n<p>The full error I am getting is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-17-28fa0c94d845&gt; in &lt;module&gt;\n     13 \n     14         model.zero_grad()\n---&gt; 15         loss.backward()\n     16         optimizer.step()\n     17 \n\n~\/anaconda3\/envs\/gautam_new\/lib\/python3.6\/site-packages\/torch\/tensor.py in backward(self, gradient, retain_graph, create_graph)\n     91                 products. Defaults to ``False``.\n     92         \"\"\"\n---&gt; 93         torch.autograd.backward(self, gradient, retain_graph, create_graph)\n     94 \n     95     def register_hook(self, hook):\n\n~\/anaconda3\/envs\/gautam_new\/lib\/python3.6\/site-packages\/torch\/autograd\/__init__.py in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables)\n     88     Variable._execution_engine.run_backward(\n     89         tensors, grad_tensors, retain_graph, create_graph,\n---&gt; 90         allow_unreachable=True)  # allow_unreachable flag\n     91 \n     92 \n\nRuntimeError: element 0 of tensors does not require grad and does not have a grad_fn\n<\/code><\/pre>\n\n<p>Please help. Also, if possible also advice on how I can make my model deeper. I keep getting CUDA  out of memory error.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53131883","CreationDate":"2018-11-03T13:38:57.147","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>I cannot test your model, but considering the error message it makes sense the cause of your problem lies in the return value of your <code>forward<\/code>. <\/p>\n\n<p>Currently you are returning <code>x<\/code> which is your actual <strong>input<\/strong> not the <strong>output<\/strong>:<\/p>\n\n<pre><code>def forward(self, x):\n    x1 = self.encoder_block1(x)\n    y1 = self.decoder_block1(x1)\n    y0 = self.decoder_block0(y1)\n    return x\n<\/code><\/pre>\n\n<p>So to return the <strong>output<\/strong> you might want to change the return value form <code>x<\/code> to <code>y0<\/code>:<\/p>\n\n<pre><code>def forward(self, x):\n    x1 = self.encoder_block1(x)\n    y1 = self.decoder_block1(x1)\n    y0 = self.decoder_block0(y1)\n    return y0\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p><em>About the memory: <br><br>Please don't put too many issues in one question. Imagine you have three completely different issues in one question, and there are three people out there where each of them is able to solve one of your issues, you might end up with <strong>no answer<\/strong> at all. <br>Because none of them is able to give you a complete answer addressing all of these issues.<br> But if your split your issues into three questions, you might get just three answers, solving all your issues. In many cases it also can improve the question, because one can be more specific to the problem without writing an entire novel in the question.<br>Of course if your issues are very related you can put them into one question, but this doesn't seem to be the case here.<br><br> I guess there is still a slight chance your <code>forward<\/code> function had some side effects leading to the memory problem (wild guess - not sure at all about this). So if you're lucky it might solve your memory problem too, but if not you definitely should open a new question about it.<\/em><\/p>\n"}]}
+{"QuestionId":53128780,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T05:50:43.250","AcceptedAnswerId":null,"OwnerUserId":9097002.0,"Title":"Tensorflow NMT with Attention Tutorial -- need help understanding loss function","Body":"<p>I'm following Tensorflow's Neural Machine Translation with Attention tutorial (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/contrib\/eager\/python\/examples\/nmt_with_attention\/nmt_with_attention.ipynb\" rel=\"nofollow noreferrer\">link<\/a>) but am unclear about some implementation details. It'd be great if someone could help clarify or refer me to a source\/better place to ask:<\/p>\n\n<p>1) <code>def loss_function(real, pred)<\/code>: This function computes loss at a specific time step (say <code>t<\/code>), averaged over the entire the batch. Examples whose labels at <code>t<\/code> is <code>&lt;pad&gt;<\/code> (i.e. no real data, only padded so that all example sequences are of same length) are masked so as not to count towards loss.<\/p>\n\n<p><strong>My question<\/strong>: It seems loss should get smaller the bigger <code>t<\/code> is (since more examples are <code>&lt;pad&gt;<\/code> the further we get to maximum length). So why is loss averaged over the entire batch, and not just over the number of valid (non-<code>&lt;pad&gt;<\/code>) examples? (This is analogous to using <code>tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS<\/code> instead of <code>tf.losses.Reduction.SUM_OVER_BATCH_SIZE<\/code>)<\/p>\n\n<p>2) <code>for epoch in range(EPOCHS)<\/code> \u2014\u2014> Two loss variables are defined in the training loop:<\/p>\n\n<ul>\n<li><code>loss<\/code> = sum of <code>loss_function()<\/code> outputs over all time steps<\/li>\n<li><code>batch_loss<\/code> = <code>loss<\/code> divided by number of time steps<\/li>\n<\/ul>\n\n<p><strong>My question<\/strong>: Why are gradients computed w.r.t. <code>loss<\/code> and not <code>batch_loss<\/code>? Shouldn't <code>batch_loss<\/code> be the average loss over all time steps and the entire batch?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"53397270","CreationDate":"2018-11-20T16:22:08.840","ParentId":null,"OwnerUserId":"10677249","Title":null,"Body":"<blockquote>\n  <p>It seems loss should get smaller the bigger t <\/p>\n<\/blockquote>\n\n<p>The loss does get smaller since the pad token is getting masked while calculating the loss.<\/p>\n\n<p><strong>Batch_loss is used only to print the loss calculated of each batch.<\/strong> Batch loss is calculated for every batch and across all the time steps.<\/p>\n\n<pre><code>for t in range(1, targ.shape[1])\n<\/code><\/pre>\n\n<p>This loop runs over the batch for all timesteps and calculates the loss by masking the padded values.<\/p>\n\n<p>I hope this clears it up :)<\/p>\n"}]}
+{"QuestionId":53128958,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorboard><pre-trained-model>","CreationDate":"2018-11-03T06:22:43.223","AcceptedAnswerId":53129874.0,"OwnerUserId":10085601.0,"Title":"How can I quickly check which tensorflow variables are updated during training and which are frozen?","Body":"<p>I believe in a lot of scenarios, we need to freeze some layers in the tensorflow graph and keep other layers trainable.<\/p>\n\n<p>Is there a method to quickly check if the network is trained as we expected? For example, the variables in the frozen layer are actually not updated during training.<\/p>\n\n<p>I am using the following method to freeze all the variables in scope \"ABC\":<\/p>\n\n<pre><code>    with slim.arg_scope(inception.inceptionb_v2_arg_scope()):\n        with tf.variable_scope('ABC'):\n          _, end_points = getattr(inception, 'inception_v2'(..., is_training = False))\n                         ......\n    trainables = [v for v in tf.trainable_variables() if 'ABC\/' not in v.name]\n    optimizer = tf.train.AdamOptimizer().minimize(loss, var_list= trainables)\n<\/code><\/pre>\n\n<p>What is the suggested way to quickly confirm these variables are really not changed during training?<\/p>\n","answers":[{"AnswerId":"53129874","CreationDate":"2018-11-03T09:03:08.303","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>You can just check them after a couple of iterations:<\/p>\n\n<pre><code>frozen_variables = [v for v in tf.trainable_variables() if 'ABC\/' in v.name]\ntmp_frozen_variables_np = sess.run(frozen_variables)\n# Training Code\nassert np.allclose(tmp_frozen_variables_np, sess.run(frozen_variables))\n<\/code><\/pre>\n\n<p>However, as long as they are not in the var list for the optimizer you should be fine.<\/p>\n"}]}
+{"QuestionId":53129288,"AnswerCount":2,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-11-03T07:20:36.740","AcceptedAnswerId":null,"OwnerUserId":4704607.0,"Title":"LSTM - Predicting the same constant values after a while","Body":"<p>I have a variable that i want to forecast until for the next 30 years. Unfortunately i don't have many samples.<\/p>\n\n<pre><code>df = pd.DataFrame({'FISCAL_YEAR': [1979,1980,1981,1982,1983,  1984,  \n1985,  1986,  1987,  1988,  1989,  1990,  1991,  1992,  1993,  1994,  \n1995,  1996,\n  1997,  1998,  1999,  2000,  2001,  2002,  2003,  2004,  2005,  2006,  \n2007,  2008,  2009,  2010,  2011,  2012,  2013,  2014,  2015,  2016,  \n2017,  2018,  2019],\n 'VALS': [1341.9,  1966.95,  2085.75,  2087.1000000000004,  2760.75,  \n3461.4,  3156.3,  3061.8,  2309.8500000000004,  2320.65,  2535.3,  \n2964.6000000000004,  2949.75,  2339.55,\n  2327.4,  2571.75,  2299.05,  1560.6000000000001,  1370.25,  1301.4,  \n1215.0,  5691.6,  6281.55,  6529.950000000001,  17666.100000000002,  \n14467.95,  15205.050000000001,  14717.7,  14426.1,  12946.5,\n  13000.5,  12761.550000000001,  13076.1,  13444.650000000001,  \n13444.650000000001,  13321.800000000001,  13536.45,  13331.25,  \n12630.6,  12741.300000000001,  12658.95]})\n<\/code><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><code>def build_model(n_neurons,dropout,s):\n    lstm = Sequential()\n    if cudnn:\n        lstm.add(CuDNNLSTM(n_neurons))\n        n_epochs = 200\n    else:\n        lstm.add(Masking(mask_value=-1,input_shape=(s[1],s[2])))\n        lstm.add(LSTM(n_neurons,dropout=dropout))\n        n_epochs = 500\n\n    lstm.add(Dense(1))\n    #lstm.add(Activation('softmax'))\n    lstm.compile(loss='mean_squared_error',optimizer='adam')\n    return lstm\n\ndef create_df(dfin,fwd,lstmws):\n    ''' Input Normalization '''\n    idx = dfin.FISCAL_YEAR.values[fwd:]\n    dfx = dfin[[varn]].copy()\n    dfy = dfin[[varn]].copy()\n\n    # LSTM window - use last lstmws values\n    for i in range(0,lstmws-1):\n        dfx = dfx.join(dfin[[varn]].shift(-i-1),how='left',rsuffix='{:02d}'.format(i+1))\n\n    dfx = (dfx-vmnx).divide(vmxx-vmnx)\n    dfx.fillna(-1,inplace=True) # replace missing values with -1\n\n    dfy = (dfy-vmnx).divide(vmxx-vmnx)\n    dfy.fillna(-1,inplace=True) # replace missing values with -1\n    return dfx,dfy,idx\n\ndef forecast(dfin,dfx,lstm,idx,gapyr=1):\n    ''' Model Forecast '''\n    xhat = dfx.values\n    xhat = xhat.reshape(xhat.shape[0],lstmws,int(xhat.shape[1]\/lstmws))\n    yhat = lstm.predict(xhat)\n\n    yhat = yhat*(vmxx-vmnx)+vmnx\n    dfout = pd.DataFrame(list(zip(idx+gapyr,yhat.reshape(1,-1)[0])),columns=['FISCAL_YEAR',varn])\n    dfout = pd.concat([dfin.head(1),dfout],axis=0).reset_index(drop=True)\n    #append last prediction to X and use for prediction\n    dfin = pd.concat([dfin,dfout.tail(1)],axis=0).reset_index(drop=True)\n    return dfin\n\ndef lstm_training(dfin,lstmws,fwd,num_years,batchsize=4,cudnn=False,n_neurons=47,dropout=0.05,retrain=False):\n    ''' LSTM Parameter '''\n    seed(2018)\n    set_random_seed(2018)\n    gapyr = 1 # Forecast +1 Year\n\n    dfx,dfy,idx = create_df(dfin,fwd,lstmws)\n\n    X,y = dfx.iloc[fwd:-gapyr].values,dfy[fwd+gapyr:].values[:,0]\n    X,y = X.reshape(X.shape[0],lstmws,int(X.shape[1]\/lstmws)),y.reshape(len(y), 1)\n\n    lstm = build_model(n_neurons,dropout,X.shape)\n    ''' LSTM Training Start '''\n    if batchsize == 1:\n        history_i = \nlstm.fit(X,y,epochs=25,batch_size=batchsize,verbose=0,shuffle=False)\n    else:\n        history_i = lstm.fit(X,y,epochs=n_epochs,batch_size=batchsize,verbose=0,shuffle=False)\n\n    dfin = forecast(dfin,dfx,lstm,idx)\n\n\n    lstm.reset_states()\n    if not retrain:\n        for fwd in range(1,num_years):\n\n            dfx,dfy,idx = create_df(dfin,fwd,lstmws)\n\n            dfin = forecast(dfin,dfx,lstm,idx)\n\n            lstm.reset_states()\n\n    del dfy,X,y,lstm\n    gc.collect();\nclear_session();\nreturn dfin,history_i\n\nvarn = \"VALS\"\n#LSTM-window\nlstmws = 10\nvmnx,vmxx = df[varn].astype(float).min(),df[varn].astype(float).max()\ndfin,history_i = lstm_training(dfin,lstmws,0,2051-2018)\n<\/code><\/pre>\n\n<p>In my first version i retrained the model every time after appending the new prediction, and the predictions never converged. But because it's very time consuming to train after every new observation, i had to change.<\/p>\n\n<p>My result:<\/p>\n\n<pre><code>dfin.VALS.values\narray([  1341.9       ,   1966.95      ,   2085.75      ,   2087.1       ,\n     2760.75      ,   3461.4       ,   3156.3       ,   3061.8       ,\n     2309.85      ,   2320.65      ,   2535.3       ,   2964.6       ,\n     2949.75      ,   2339.55      ,   2327.4       ,   2571.75      ,\n     2299.05      ,   1560.6       ,   1370.25      ,   1301.4       ,\n     1215.        ,   5691.6       ,   6281.55      ,   6529.95      ,\n    17666.1       ,  14467.95      ,  15205.05      ,  14717.7       ,\n    14426.1       ,  12946.5       ,  13000.5       ,  12761.55      ,\n    13076.1       ,  13444.65      ,  13444.65      ,  13321.8       ,\n    13536.45      ,  13331.25      ,  12630.6       ,  12741.3       ,\n    12658.95      ,  10345.97167969,  12192.12792969,  13074.4296875 ,\n    13264.40917969,  12956.1796875 ,  12354.1953125 ,  11659.03125   ,\n    11044.06933594,  10643.19921875,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094,  10552.52246094,  10552.52246094,\n    10552.52246094,  10552.52246094])\n<\/code><\/pre>\n\n<p>How can I avoid to get the same prediction for the last 20+ years?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I prepended more random data to see if it is because of the little sample size, but the predictions are again constant after a while.<\/p>\n\n<pre><code>df0 = pd.DataFrame([range(1900,1979),list(np.random.rand(1979-1900)*(vmxx-vmnx)+vmnx)],index=[\"FISCAL_YEAR\",\"VALS\"]).T\ndf = pd.concat([df0,df])\ndf[\"FISCAL_YEAR\"] = df[\"FISCAL_YEAR\"].astype(int)\ndf.index = range(1900,2020)\n<\/code><\/pre>\n\n<p>A strange thing I have observed is that the predictions are the same after 10 years, i.e. the window size, but if I increase lstmws to 20, the predictions converge after 20 years:<\/p>\n\n<pre><code>lstmws = 20\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>{'FISCAL_YEAR': [2020,  2021,  2022,  2023,  2024,  2025,  2026,  027,  028,  2029,  2030,  2031,  2032,  2033,  2034,  2035,  2036,  2037,  2038,  039,  2040,  2041,  2042,  2043,  2044,  2045,  2046,  2047,  2048,  2049,  050,  2051,  2052],\n 'VALS': [11183.32421875,  12388.28125,  13151.013671875,  12543.6796875,  2590.0888671875,  12002.583984375,  11822.8857421875,  11479.6572265625,  1423.1279296875,  11444.5751953125,  11506.60546875,  11563.3173828125,  1595.0029296875,  11599.8955078125,  11586.8037109375,  11571.337890625,  1574.541015625,  11620.7900390625,  11734.2431640625,  11934.216796875,  1934.216796875,  11934.216796875,  11934.216796875,  11934.216796875,  1934.216796875,  11934.216796875,  11934.216796875,  11934.216796875,  1934.216796875,  11934.216796875,  11934.216796875,  11934.216796875,  1934.216796875]}\n<\/code><\/pre>\n","answers":[{"AnswerId":"53129754","CreationDate":"2018-11-03T08:45:49.620","ParentId":null,"OwnerUserId":"9545636","Title":null,"Body":"<p>I can see that your predict step just uses the default batch size? Try setting batch size to that you used for the training step. See if that helps.<\/p>\n\n<pre><code>yhat = lstm.predict(xhat, batch_size=batchsize)\n<\/code><\/pre>\n"},{"AnswerId":"53133607","CreationDate":"2018-11-03T17:07:10.657","ParentId":null,"OwnerUserId":"1727392","Title":null,"Body":"<p>In My experience with LSTM's (I've been generating dance sequences like <a href=\"https:\/\/bl.ocks.org\/duhaime\/b035d551d816a608a58c3ea4f464891e\" rel=\"nofollow noreferrer\">this<\/a>), I've found two things in particular help prevent the model from stagnating and predicting the same output.<\/p>\n\n<p>First, it's helpful to use a Mixture Density Network instead of an L2 loss (as you have). Read Christopher Bishop's <a href=\"https:\/\/publications.aston.ac.uk\/373\/1\/NCRG_94_004.pdf\" rel=\"nofollow noreferrer\">paper on MDN layers<\/a> for the details, but basically the L2 loss tries to predict the conditional average of some input's error terms as y. If for one value x you have multiple possible outputs y0, y1, y2, each with some probability (as many complex systems will), you'll want to consider the MDN layer and a negative log likelihood loss. <a href=\"https:\/\/github.com\/omimo\/Keras-MDN\" rel=\"nofollow noreferrer\">Here<\/a> is a Keras implementation that I'm using.<\/p>\n\n<p>Reading your situation a little more closely now, this may not be helpful for your case, as you seem to be predicting a time series for which by definition each x maps to a single y.<\/p>\n\n<p>Next, I've found it helpful to feed my LSTM <code>n<\/code> sequence values prior to the one I'm trying to predict. The larger n is, the better the results I've found (though slower the training goes). Many papers I've read use 1024 prior sequence values to predict the next sequence value.<\/p>\n\n<p>You don't have many observations, but you could try feeding the prior 8 observations in to predict the next observation.<\/p>\n"}]}
+{"QuestionId":53129439,"AnswerCount":0,"Tags":"<image><keras><tensorboard>","CreationDate":"2018-11-03T07:46:20.407","AcceptedAnswerId":null,"OwnerUserId":7333228.0,"Title":"How do I add multiple images on TensorBoard callback for Keras","Body":"<p>I'm trying to add my validation set images to the TensorBoard callbacks of Keras.  The following code works and even gives me the slider for each batch step.<br> \nHowever, the batch step gets appended to the tag so it ends up showing each image as a separate tag (e.g. <code>predictions_1<\/code>, <code>predictions_2<\/code>, <code>predictions_3<\/code>). <\/p>\n\n<p><strong>Is there some way I can keep the tag from getting appended?<\/strong> <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class TensorBoardWithImages(K.callbacks.TensorBoard):\n    def __init__(self, **kwargs):\n        super(TensorBoardWithImages, self).__init__(**kwargs)\n\n    def on_batch_end(self, batch, logs=None):\n\n        logs = logs or {}\n\n        val_data = self.validation_data\n\n        a=np.random.randint(0, val_data[0].shape[0], 10)\n        tf.summary.image(name=\"predictions\", tensor=val_data[0][a],\n                         max_outputs=3)\n\n        summary_op = tf.summary.merge_all()\n\n        self.writer.add_summary(summary_op.eval(session=self.sess), batch)\n\n        super(TensorBoardWithImages, self).on_batch_end(batch, logs)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53129544,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T08:04:25.383","AcceptedAnswerId":53130159.0,"OwnerUserId":5863503.0,"Title":"Tensorflow: how to save variables and load them to different variables?","Body":"<p>Let's say I have two identical networks, <code>A<\/code> and <code>B<\/code>. I saved (using <code>Saver<\/code>) a previous state of network <code>A<\/code>, and now I would like to load it into network <code>B<\/code> (all happens during the same run). How can I do this?<\/p>\n","answers":[{"AnswerId":"53130159","CreationDate":"2018-11-03T09:48:26.443","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Let me provide an example. First, let's define and save some variables:<\/p>\n\n<pre><code>import tensorflow as tf\n\nv1 = tf.Variable(tf.ones(1), name='v1')\nv2 = tf.Variable(2 * tf.ones(1), name='v2')\n\nsaver = tf.train.Saver(tf.trainable_variables())\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    saver.save(sess, '.\/tmp.ckpt')\n<\/code><\/pre>\n\n<p>Now, let's define some variables with the same names in a new graph, and load their values from the checkpoint:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n    assert len(tf.trainable_variables()) == 0\n    v1 = tf.Variable(tf.zeros(1), name='v1')\n    v2 = tf.Variable(tf.zeros(1), name='v2')\n\n    saver = tf.train.Saver(tf.trainable_variables())\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        saver.restore(sess, '.\/tmp.ckpt')\n        print(sess.run([v1, v2]))\n<\/code><\/pre>\n\n<p>The last line prints:<\/p>\n\n<pre><code>[array([1.], dtype=float32), array([2.], dtype=float32)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53129905,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-03T09:07:52.063","AcceptedAnswerId":null,"OwnerUserId":8005524.0,"Title":"Tensorflow error:The graph couldn't be sorted in topological order","Body":"<p>When I run my loss function and it will be occur this error or warning.\nI really can not figure out what cause it.<br\/>\nI guess that maybe I didn't use the origin input,for example:<\/p>\n\n<pre><code>def loss(predict,label):\n    #because some reason I need to extract some values in predict\n    predictProcessed = process(predict)\n    #predictProcessed is a subset of predict\n    loss = tf.square(predict - label)\n    return loss\n<\/code><\/pre>\n\n<p>My guess is right or not?\nAnd I also use double for-loop in this code,Should the code use less for for-loop?<br\/>thanks<\/p>\n","answers":[]}
+{"QuestionId":53130401,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><object-detection><object-detection-api>","CreationDate":"2018-11-03T10:23:12.633","AcceptedAnswerId":null,"OwnerUserId":4229637.0,"Title":"get bounding box coordinated from object detection api after eval","Body":"<p>I used Object Detection API from Tensorflow to train my model. The eval script also worked fine. <\/p>\n\n<p>However, I want to get the coordinates of the bounding box in eval images. Is that possible to do so in eval script? How can I do it?<\/p>\n\n<p>I used <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/legacy\/eval.py\" rel=\"nofollow noreferrer\">this script<\/a>.  <\/p>\n","answers":[{"AnswerId":"53169487","CreationDate":"2018-11-06T09:55:01.243","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>A few examples (how many is set by <code>num_visualizations<\/code> in the config file) are shown on TensorBoard.<\/p>\n\n<p>If you want more than that, I suggest to export the model and use inference.\nSee the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\" rel=\"nofollow noreferrer\">TF OD tutorial notebook<\/a> for how to use an exported model for inference and visualizing results.<\/p>\n"},{"AnswerId":"57264998","CreationDate":"2019-07-30T05:44:23.420","ParentId":null,"OwnerUserId":"10037689","Title":null,"Body":"<p>You can refer to the following <a href=\"https:\/\/stackoverflow.com\/questions\/52424348\/tensorflow-object-detection-api-multiple-objects-coordinates\/57263771#57263771\">link<\/a>.\nAlthough this does not have any relation with the <code>eval.py<\/code> script, it gets the job done. If you specify a directory consisting of the test images, you can get the bounding box coordinates of all the detected objects using the model you have trained.<\/p>\n"}]}
+{"QuestionId":53131244,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><conv-neural-network><nvidia><tf-slim>","CreationDate":"2018-11-03T12:21:10.897","AcceptedAnswerId":null,"OwnerUserId":9424025.0,"Title":"Implementing mixed precision training in tf-slim","Body":"<p>I am trying to use mixed-precision training with tf-slim in order to speed up the training of networks and make use of the tensorcores available on my GPUs.\nI also want to make use of multiple network architectures with pre-trained checkpoints.<\/p>\n\n<p>An example of what Mixed-Precision training is and how it works can be found at <a href=\"https:\/\/devblogs.nvidia.com\/mixed-precision-resnet-50-tensor-cores\/\" rel=\"nofollow noreferrer\">https:\/\/devblogs.nvidia.com\/mixed-precision-resnet-50-tensor-cores\/<\/a><\/p>\n\n<p>The basic idea is to\n1. Cast the inputs and to fp16 for the forward and backward pass\n2. Cast the values back to fp32 when adjusting loss and weights\n3. When using the Loss for the backward pass, multiply it by a loss scale\n4. When updating the weights, divide it by the same loss scale<\/p>\n\n<p>This reduces the memory bandwidth and makes use of the Tensor Cores on Volta and Turing GPUs through the use of fp16.<\/p>\n\n<p>My problem is that I can't figure out where to put the casts to fp16 and fp32 with tf-slim.<\/p>\n\n<p>To start the training, I use the train_image_classifier.py script from models.research.slim<\/p>\n\n<p>Do I need to do the cast within the definition files for the network architectures? \nOr do I need to apply the changes within the tf.contrib.slim files?<\/p>\n","answers":[{"AnswerId":"53209049","CreationDate":"2018-11-08T13:47:20.770","ParentId":null,"OwnerUserId":"9368966","Title":null,"Body":"<p>NVIDIA's <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/sdk\/mixed-precision-training\/index.html#example_tensorflow\" rel=\"nofollow noreferrer\">documentation on mixed precision training<\/a> gives a clear example on how to do this with tensorflow.<\/p>\n\n<p>Tensorflow has implemented the loss scaling in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/mixed_precision\/LossScaleOptimizer\" rel=\"nofollow noreferrer\">tf.contrib.mixed_precision.LossScaleOptimizer<\/a>. From what I understood, it uses the same strategy as discribed in NVIDIA's documentation on mixed precision training.<\/p>\n\n<pre><code>loss = loss_fn()\nopt = tf.AdamOptimizer(learning_rate=...)\n\n# Choose a loss scale manager which decides how to pick the right loss scale \n# throughout the training process.\n# Use fixed loss scaling factor\nloss_scale_manager = tf.contrib.mixed_precision.FixedLossScaleManager(loss_scale)\n# Use dynamic loss scaling factor\nloss_scale_manager = tf.contrib.mixed_precision.ExponentialUpdateLossScaleManager(init_loss_scale, \n                                                                                  incr_every_n_steps)\n\n# Wrap the original optimizer in a LossScaleOptimizer.\nloss_scale_optimizer = tf.contrib.mixed_precision.LossScaleOptimizer(opt, loss_scale_manager)\n\n# Call minimize() on the loss scale optimizer.\ntrain_op = loss_scale_optimizer.minimize(loss)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53131311,"AnswerCount":1,"Tags":"<docker><tensorflow><jupyter-notebook><jupyter><jupyter-lab>","CreationDate":"2018-11-03T12:28:04.210","AcceptedAnswerId":null,"OwnerUserId":3880039.0,"Title":"Starting jupyter lab on mounted drive with tensorflow docker image","Body":"<p>On the official Tensorflow website documentation <a href=\"https:\/\/www.tensorflow.org\/install\/docker\" rel=\"nofollow noreferrer\">TF<\/a>, it details how to install the latest nvidia and tensorflow Docker images.  I then mounted my host directory and launch jupyter notebook by:<\/p>\n\n<pre><code>$ docker run -it --rm \\\n-p 8888:8888\n--runtime=nvidia -v ~\/my_dir:\/container_dir \\\ntensorflow\/tensorflow:lastest-gpu-py3\n<\/code><\/pre>\n\n<p>This command will start a jupyter notebook from the \\notebook directory in the container by default.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>How do I launch a jupyter notebook from mounted host directory, which contains my own notebooks?  I realize I could just launch a bash shell from the docker image, cd to the mounted host directory, and start a notebook from there, but launching a notebook from root may not be a best practice.<\/li>\n<li>How do I launch jupyter LAB instead of jupyter NOTEBOOK, as I believe both their ports are 8888, and this TF image seems to launch a notebook by default.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53134049","CreationDate":"2018-11-03T17:58:26.283","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>You could write your own dockerfile which installs all dependencies and is based the tensorflow image. You could for example alter the run command that it changes the directory.  Or you create a different image which installs Jupyter hub and starts the Instance on run. You can find a nice explanation on how to create dockerfiles <a href=\"https:\/\/rominirani.com\/docker-tutorial-series-writing-a-dockerfile-ce5746617cd\" rel=\"nofollow noreferrer\">here. <\/a><\/p>\n"}]}
+{"QuestionId":53131397,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><loss-function>","CreationDate":"2018-11-03T12:40:06.293","AcceptedAnswerId":53131649.0,"OwnerUserId":4324126.0,"Title":"Loss function Keras out_dim > 1","Body":"<p>I have a training data:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GYyeW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GYyeW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And, I have a model in Keras with more than one dimension of output. I want to predict A, B and C:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(GRU(32, input_shape=(train_X.shape[1], train_X.shape[2])))\nmodel.add(Dense(3))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\n<\/code><\/pre>\n\n<p>But I want the minimum <code>mean_squared_error<\/code> in A, i.e. only want to consider A for the loss function.<\/p>\n\n<p>What I can do?<\/p>\n","answers":[{"AnswerId":"53131651","CreationDate":"2018-11-03T13:09:40.400","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>What you need to look into is a custom loss function:<\/p>\n\n<pre><code>def only_A_mean_squared(y_true, y_pred):\n     return keras.losses.mean_squared_error(y_true[:,0], y_pred[:,0])\n<\/code><\/pre>\n\n<p>And in order to use it:<\/p>\n\n<pre><code>model.compile(loss=only_A_mean_squared, optimizer='adam')\n<\/code><\/pre>\n\n<p>What i am doing in the above is creating a custom loss function, which only takes the first dimension (the 'A') and feed it to the normal keras mean squared error loss function.<\/p>\n"},{"AnswerId":"53131649","CreationDate":"2018-11-03T13:09:28.253","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can define a custom loss function and only compute the <code>mean_squared_error()<\/code> loss based on the value of <code>A<\/code>:<\/p>\n\n<pre><code>from keras import losses\n\ndef loss_A(y_true, y_pred):\n    return losses.mean_squared_error(y_true[:,0], y_pred[:,0])\n\n#...\nmodel.compile(loss=loss_A, optimizer='adam')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53131429,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T12:44:22.210","AcceptedAnswerId":53131643.0,"OwnerUserId":7835633.0,"Title":"ValueError: Error when checking input: expected flatten_input to have 3 dimensions, but got array with shape (22, 12)","Body":"<p>I am new to Tensorflow, I am trying to do text recognize for image(size 12x22) using tensorflow.\nTraining model seems to be ok, but when do prediction, there is error:<\/p>\n\n<pre><code>  File \"C:\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py\", line 316, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking input: expected flatten_input to have 3 dimensions, but got array with shape (22, 12)\n<\/code><\/pre>\n\n<p>I don't know why it needs 3 dimension while I just input a image in 2D.<\/p>\n\n<p>train_model.py, this is the trainer, export a training model<\/p>\n\n<pre><code>import cv2\nimport pickle\nimport os.path\nimport numpy as np\nfrom imutils import paths\nimport tensorflow as tf\n# from sklearn.preprocessing import LabelBinarizer\n# from sklearn.model_selection import train_test_split\nfrom tensorflow.keras.models import Sequential\n# from keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom tensorflow.keras.layers import Flatten, Dense\n\nimport image_fit\n\nLETTER_IMAGES_FOLDER = \"training_generate\/divided_sample\"\nMODEL_FILENAME = \"captcha_model.hdf5\"\nMODEL_LABELS_FILENAME = \"model_labels.dat\"\n\n\n# initialize the data and labels\ndata = []\nlabels = []\n\n# loop over the input images\nprint('Read images')\nfor image_file in paths.list_images(LETTER_IMAGES_FOLDER):\n    # Load the image and convert it to grayscale\n    image = cv2.imread(image_file)\n    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n\n    # resize the image as 12, 22\n    image = image_fit.resize_to_fit(image, 12, 22)\n\n    # Grab the name of the letter based on the folder it was in\n    label = image_file.split(os.sep)[-2]\n    label = chr(int(label))\n    # Add the letter image and it's label to our training data\n    data.append(image)\n    labels.append(label)\n\nprint('Done reading')\ndata = tf.keras.utils.normalize(np.array(data), axis=1)\n# print(data[0])\nlabels = np.array(labels)\n\nmodel = Sequential()\nmodel.add(Flatten(input_shape=(22, 12)))\nmodel.add(Dense(128, activation=tf.nn.relu))\nmodel.add(Dense(128, activation=tf.nn.relu))\nmodel.add(Dense(62, activation=tf.nn.softmax))\n\nmodel.compile(optimizer='Adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(data, labels, epochs=3)\nprint('Finish training')\n\n# val_loss, val_acc = model.evaluate()\nmodel.save(MODEL_FILENAME)\n<\/code><\/pre>\n\n<p>solve.py, this is the python script that solve the letter.<\/p>\n\n<pre><code>import cv2\nimport pickle\nimport os.path\nimport numpy as np\nfrom imutils import paths\nimport tensorflow as tf\n\nimport image_fit\n\n\nLETTER_IMAGES_FOLDER = \"training_generate\/divided_sample\"\nMODEL_FILENAME = \"captcha_model.hdf5\"\nMODEL_LABELS_FILENAME = \"model_labels.dat\"\n\nimage_file = 'test_0.bmp'\nimage = cv2.imread(image_file)\nimage = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\nimage = image_fit.resize_to_fit(image, 12, 22)\n\nnew_model = tf.keras.models.load_model(MODEL_FILENAME)\npredicitions = new_model.predict(image)\nprint(np.argmax(predicitions[0]))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53131869","CreationDate":"2018-11-03T13:36:26.643","ParentId":null,"OwnerUserId":"7835633","Title":null,"Body":"<p>According to Kruger's answer, the code should be look like this:<\/p>\n\n<pre><code>    image_file = 'test_0.bmp'\n    image = cv2.imread(image_file)\n    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n    image = image_fit.resize_to_fit(image, 12, 22)\n\n    data = [image]\n\n    new_model = tf.keras.models.load_model(MODEL_FILENAME)\n    predicitions = new_model.predict(np.array(data))\n    print(np.argmax(predicitions[0]))\n<\/code><\/pre>\n"},{"AnswerId":"53131643","CreationDate":"2018-11-03T13:08:52.287","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>When you make the prediction with the line<\/p>\n\n<pre><code>predicitions = new_model.predict(image)\n<\/code><\/pre>\n\n<p>it is expecting the first dimension to be the number of images in the batch.  Since you only have one image, the first dimension would be of size one.  You could resize the image to be (1,22,12), and this would pass:<\/p>\n\n<pre><code>predicitions = new_model.predict(image.reshape(1,22,12))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53131471,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><neural-network><deep-learning>","CreationDate":"2018-11-03T12:49:22.117","AcceptedAnswerId":53133293.0,"OwnerUserId":8349549.0,"Title":"How to feed array into a TensorFlow placeholder correctly","Body":"<p>I am trying to feed the array in batches into Tensorflow placeholders. But even after providing correct shape I am getting an <code>InvalidArgumentError<\/code>. <\/p>\n\n<p>Here is the part of my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nxdata = np.linspace(1,50, 10000)\nnoise = np.random.rand(len(xdata))\ny_true = (1.5*xdata) + 5 + noise    #m = 1.5 and c = 5\n\nm = tf.Variable(0.1)  #initial values\nc = tf.Variable(0.2)\n\nbatch_size = 10\n\nx = tf.placeholder(tf.float32, [batch_size])\ny = tf.placeholder(tf.float32, [batch_size])\n\ny_hat = (m*x) + c\nerror = tf.reduce_sum(tf.square(y-y_hat))\noptimizer = tf.train.GradientDescentOptimizer(learning_rate= 0.01)\ntrain = optimizer.minimize(error)\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n\n    sess.run(init)\n\n    n_batches = 1000\n\n    for i in range(n_batches):\n\n        rand_int = np.random.randint(len(xdata), size =batch_size)\n        feed_dict = {x:xdata[rand_int], y: y_true[rand_int]}\n\n        sess.run(train, feed_dict = feed_dict)\n        print('Batch:',i, ' loss: ', sess.run(error))\n\n    m_final, slope_final = sess.run([m , c])\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: You must feed a value for  placeholder tensor\n  'Placeholder' with dtype float  and shape [10]<\/p>\n<\/blockquote>\n\n<p>Why is this happening?<\/p>\n","answers":[{"AnswerId":"53133293","CreationDate":"2018-11-03T16:28:55.703","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The error occurs in this line:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>print('Batch:', i, ' loss: ', sess.run(error))\n<\/code><\/pre>\n\n<p>In order to compute the value for tensor <code>error<\/code>, you must feed a value for placeholders <code>x<\/code> and <code>y<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess.run(error, feed_dict)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53131685,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-03T13:13:24.520","AcceptedAnswerId":null,"OwnerUserId":9349789.0,"Title":"Keras shared weights network is inconsistent","Body":"<p>I try to implement a network in Keras for a symmetric problem - a model that predicts the distance between inputs a and b.<\/p>\n\n<p>I used the following official references:\n<a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#shared-layers\" rel=\"nofollow noreferrer\">1<\/a> and <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_siamese.py\" rel=\"nofollow noreferrer\">2<\/a> to create the following simple implementation:<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import print_function\nfrom __future__ import absolute_import\nfrom __future__ import print_function\nimport keras\nfrom keras.models import Model\nfrom keras.layers import Input, Flatten, Dense, Dropout\nimport numpy as np\n\n\ndef create_base_network(input_shape):\n    '''Base network to be shared (eq. to feature extraction).\n    '''\n    input = Input(shape=input_shape)\n    x = Flatten()(input)\n    x = Dense(128, activation='relu')(x)\n    x = Dropout(0.1)(x)\n    x = Dense(128, activation='relu')(x)\n    x = Dropout(0.1)(x)\n    x = Dense(128, activation='relu')(x)\n    return Model(input, x)\n\nnum_of_features = 256\nnum_of_samples = 280\ndata_a = np.random.random((num_of_samples, 1, num_of_features))\ndata_b = np.random.random((num_of_samples, 1, num_of_features))\n# binary label\nlabels = np.random.randint(2, size=num_of_samples)\n\ninput_shape= (1, num_of_features)\n# network definition\nbase_network = create_base_network(input_shape)\n\ninput_a = Input(shape=input_shape)\ninput_b = Input(shape=input_shape)\n\n# because we re-use the same instance `base_network`, the weights of the network \n# will be shared across the two branches\nencoded_a = base_network(input_a)\nencoded_b = base_network(input_b)\n\n# We can then concatenate the two vectors:\nmerged_vector = keras.layers.concatenate([encoded_a, encoded_b], axis=-1)\n\n# And add a logistic regression on top\npredictions = Dense(1, activation='sigmoid')(merged_vector)\n\n# We define a trainable model linking the inputs to the predictions\nmodel = Model(inputs=[input_a, input_b], outputs=predictions)\n\nmodel.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit([data_a, data_b], labels, epochs=10)   \n<\/code><\/pre>\n\n<p>Yet, when I use model evaluation, I get a different metric value over the test set when I switch between a and b:<\/p>\n\n<pre><code>data_a_test = np.random.random((num_of_samples, 1, num_of_features))\ndata_b_test = np.random.random((num_of_samples, 1, num_of_features))\nlabels_test = np.random.randint(2, size=num_of_samples)\n\nloss_ab, metric_ab = model.evaluate([data_a_test, data_b_test], labels_test, batch_size=32, verbose=2)\nloss_ba, metric_ba = model.evaluate([data_b_test, data_a_test], labels_test, batch_size=32, verbose=2)\n\nloss_ab: 0.9805058070591518 metric_ab: 0.48928571343421934\nloss_ba: 1.0541694641113282 metric_ba: 0.5\n<\/code><\/pre>\n\n<p><strong>What do I miss here?<\/strong>\nplease help me with some inputs...<\/p>\n","answers":[]}
+{"QuestionId":53132037,"AnswerCount":0,"Tags":"<keras><dask>","CreationDate":"2018-11-03T14:00:04.340","AcceptedAnswerId":null,"OwnerUserId":4230234.0,"Title":"Dask DataFrame for fitting Keras model","Body":"<p>I have a Dask DataFrame that I want to use for fitting a Keras autoencoder model:<\/p>\n\n<p><strong>DataFrame:<\/strong> <\/p>\n\n<pre><code>import dask.dataframe as dd\ninput_df = dd.read_csv(file_path)\ninput_df.dtypes\n<\/code><\/pre>\n\n<p>_2          float64\n_3          float64\n_4          float64\n_5          float64 ...<\/p>\n\n<p><strong>Keras model:<\/strong><\/p>\n\n<pre><code>autoencoder = Sequential()\nautoencoder.add(Dense(dense[0], input_shape=(dense[0],), activation = 'relu' ))\nautoencoder.add(Dense(dense[1], activation = 'relu' ))\nautoencoder.add(Dense(dense[2], activation = 'relu' ))\nautoencoder.add(Dense(dense[3], activation = 'relu' ))\nautoencoder.add(Dense(dense[0], activation = 'relu' ))\n\nautoencoder.compile(loss='mse',\n          optimizer='adam',\n          metrics=['mse'])\n<\/code><\/pre>\n\n<p>When I pass the DataFrame for fitting:  <\/p>\n\n<pre><code>autoencoder.fit(input_df, input_df,\n                batch_size=batch_size,\n                epochs=epochs,\n                verbose=1,\n                validation_split = val_split)\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-23-d0480d8a460d&gt; in &lt;module&gt;()\n      3                     epochs=epochs,\n      4                     verbose=1,\n----&gt; 5                     validation_split = val_split)\n\n~\/anaconda3\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n    950             sample_weight=sample_weight,\n    951             class_weight=class_weight,\n--&gt; 952             batch_size=batch_size)\n    953         # Prepare validation data.\n    954         do_validation = False\n\n~\/anaconda3\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    799                 for (ref, sw, cw, mode) in\n    800                 zip(y, sample_weights, class_weights,\n--&gt; 801                     feed_sample_weight_modes)\n    802             ]\n    803             # Check that all arrays have the same length.\n\n~\/anaconda3\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in &lt;listcomp&gt;(.0)\n    797             sample_weights = [\n    798                 standardize_weights(ref, sw, cw, mode)\n--&gt; 799                 for (ref, sw, cw, mode) in\n    800                 zip(y, sample_weights, class_weights,\n    801                     feed_sample_weight_modes)\n\n~\/anaconda3\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py in standardize_weights(y, sample_weight, class_weight, sample_weight_mode)\n    522     else:\n    523         if sample_weight_mode is None:\n--&gt; 524             return np.ones((y.shape[0],), dtype=K.floatx())\n    525         else:\n    526             return np.ones((y.shape[0], y.shape[1]), dtype=K.floatx())\n\n~\/anaconda3\/envs\/py36\/lib\/python3.6\/site-packages\/numpy\/core\/numeric.py in ones(shape, dtype, order)\n    201 \n    202     \"\"\"\n--&gt; 203     a = empty(shape, dtype, order)\n    204     multiarray.copyto(a, 1, casting='unsafe')\n    205     return a\n\nTypeError: 'float' object cannot be interpreted as an integer\n<\/code><\/pre>\n\n<p>Would appreciate some help! Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53132158,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-11-03T14:14:50.823","AcceptedAnswerId":null,"OwnerUserId":10179861.0,"Title":"Issues with importing tensorflow python 3.7 on windows 10","Body":"<p>I have issues with importing tensorflow on my computer.\nAfter I received errors initially, I reinstalled Anaconda with the newest version Anaconda 5.3 for wihndows - python 3.7  . \nFor the installation I used the following commands on cmd:<\/p>\n\n<ul>\n<li><code>conda install tensorflow-gpu<\/code> - got an error because of different versions, and continued with the following:<\/li>\n<li><code>conda create -n Tensorflow anaconda python=3.7<\/code> - installed some packages , as required<\/li>\n<li><code>activate Tensorflow<\/code> <\/li>\n<li><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0-cp35-cp35m-win_amd64.whl<\/code> --> Got the following error : <code>tensorflow-1.1.0-cp35-cp35m-win_amd64.whl is not a supported wheel on this platform.<\/code><\/li>\n<\/ul>\n\n<p>Afterwards I tried to import it on jupyter notebook anyway and got error again:<\/p>\n\n<pre><code>ModuleNotFoundError                       Traceback (most recent call last)\n&lt;ipython-input-1-64156d691fe5&gt; in &lt;module&gt;()\n----&gt; 1 import tensorflow as tf\n\nModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>Can anyone advise how can I solve it?<\/p>\n","answers":[{"AnswerId":"53250621","CreationDate":"2018-11-11T16:12:24.170","ParentId":null,"OwnerUserId":"10497954","Title":null,"Body":"<p>just uninstall your python 3.7.xxxx versions,then install python 3.6.xx or lower version,now you can run pip install tensorflow successfully.\ndon't forget reset your ide's python path and reinstall other packages.<\/p>\n"}]}
+{"QuestionId":53132649,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T15:16:26.537","AcceptedAnswerId":null,"OwnerUserId":10600687.0,"Title":"Tensorflow graph running without providing placeholders","Body":"<p>I have been struggling for hours trying to understand this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/6_lstm.ipynb\" rel=\"nofollow noreferrer\">code<\/a> but I just can't. <\/p>\n\n<p>In cell 9, line 13 it loops:<\/p>\n\n<pre><code>for i in range(num_unrollings + 1):       \n    feed_dict[train_data[i]] = batches[i]\n<\/code><\/pre>\n\n<p>Which means that in the <em>first<\/em> execution only <code>train_data[0]<\/code> has a value. <code>train_data[1]<\/code> is still a placeholder.<\/p>\n\n<p>When you look at the graph definition in cell 8, line 48:<\/p>\n\n<pre><code>train_labels = train_data[1:]\n<\/code><\/pre>\n\n<p>The graph requires <code>train_data[1]<\/code> to exist because if it doesn't you don't have any labels!<\/p>\n\n<p>So my question is: how does this work? It doesn't crash, it works perfectly. Why doesn't the first execution raise an exception? How can the session be run with a placeholder still in place?<\/p>\n\n<p>It should raise a<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float and shape [64,27] <\/p>\n<\/blockquote>\n\n<p>exception shouldn't it?<\/p>\n","answers":[]}
+{"QuestionId":53133190,"AnswerCount":1,"Tags":"<tensorflow><keras><text-processing>","CreationDate":"2018-11-03T16:18:27.940","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Some keys in session_kwargs are not supported at this time: %s', dict_keys(['learning_rate'])","Body":"<p>I am trying to build a movie rating predictor based on the reviews. I am using Keras with tensorflow as backend. I downloaded Tensorflow and keras using Anaconda. X and y in the code are numpy arrays. I am trying to run the following code, <\/p>\n\n<pre><code>import tensorflow as tf\nimport keras\nfrom keras.models import Sequential\nfrom keras.callbacks import History\nfrom keras.layers import Input, Dense, Dropout\nfrom keras.layers import BatchNormalization\nfrom keras.layers.core import Flatten\nfrom keras.layers import Embedding\nmodel = Sequential()\n\nmodel.add(keras.layers.Embedding(vocab_size, 10))\nmodel.add(keras.layers.GlobalAveragePooling1D())\nmodel.add(Dense(17, activation = 'relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10, activation = 'relu'))\nmodel.add(Dropout(0.5))\n\n\nmodel.add(Dense(5, activation = 'softmax'))\noptimizer = keras.optimizers.sgd(lr = 0.1)\n\nmodel.compile(loss = \"categorical_crossentropy\", optimizer = optimizer, metrics = ['accuracy'], learning_rate = 0.2)\nhistory = model.fit(X_train_indices, y_train_oh,batch_size = 128,epochs = 25, verbose = 1)\nprint(model.summary())\n<\/code><\/pre>\n\n<p>And I get the following error message,<\/p>\n\n<pre><code>    ValueError                                Traceback (most recent call last)\n&lt;ipython-input-47-a638769cab16&gt; in &lt;module&gt;()\n----&gt; 1 history = model.fit(X_train_indices, y_train_oh,batch_size = 128,epochs = 25, verbose = 1)\n      2 print(model.summary())\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1006         else:\n   1007             ins = x + y + sample_weights\n-&gt; 1008         self._make_train_function()\n   1009         f = self.train_function\n   1010 \n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in _make_train_function(self)\n    506                     updates=updates,\n    507                     name='train_function',\n--&gt; 508                     **self._function_kwargs)\n    509 \n    510     def _make_test_function(self):\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in function(inputs, outputs, updates, **kwargs)\n   2693                 msg = 'Invalid argument \"%s\" passed to K.function with TensorFlow backend' % key\n   2694                 raise ValueError(msg)\n-&gt; 2695     return Function(inputs, outputs, updates=updates, **kwargs)\n   2696 \n   2697 \n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in __init__(self, inputs, outputs, updates, name, **session_kwargs)\n   2540             raise ValueError('Some keys in session_kwargs are not '\n   2541                              'supported at this '\n-&gt; 2542                              'time: %s', session_kwargs.keys())\n   2543         self._callable_fn = None\n   2544         self._feed_arrays = None\n\nValueError: ('Some keys in session_kwargs are not supported at this time: %s', dict_keys(['learning_rate']))\n<\/code><\/pre>\n\n<p>The shape of input X_train_indices is 6396x28 and the shape of y_train_oh is 6396x5.<\/p>\n\n<p>If I remove the embedding and the GlobalAveragePooling1D layer from the model it works but gives a very poor accuracy of just 27% and the accuracy fluctuates wildly.<\/p>\n\n<p>Please help. Thanks.<\/p>\n","answers":[{"AnswerId":"53133627","CreationDate":"2018-11-03T17:10:35.453","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p><code>learning_rate<\/code> is not a parameter of <code>model.compile<\/code>, it is a parameter for the optimizer (usually called <code>lr<\/code>). Just remove that parameter from your call.<\/p>\n"}]}
+{"QuestionId":53133345,"AnswerCount":0,"Tags":"<keras><resnet><imagenet>","CreationDate":"2018-11-03T16:34:46.907","AcceptedAnswerId":null,"OwnerUserId":5116406.0,"Title":"Use pre-trained ResNet 50 in Keras for images with 4 channels","Body":"<p>Is there any way I can use the ImageNet weights for ResNet50 for my project which has images of shape (224,224,4)? The image has R,G,B,Y channels. <\/p>\n\n<p>At the moment, I am simply using <\/p>\n\n<pre><code> model = ResNet50(include_top=True, weights=None, input_tensor=None, input_shape=input_shape, pooling=None, classes=num_classes)\n<\/code><\/pre>\n\n<p>Now, if I need to use the ImageNet weights, I need to always set the number of classes to 1000. I tried doing that, and then popping the last layer, and adding my own Dense(num_classes) layer. However, now, the number of channels is an issue.<\/p>\n\n<p>Can anyone suggest a way to accommodate 4 channels in the model while using the ImageNet weights?<\/p>\n","answers":[]}
+{"QuestionId":53133642,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-03T17:11:46.613","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Fitting an RNN estimator in tensor flow","Body":"<p>I'm trying to train a TF estimator using the <code>RNNEstimator()<\/code> class, but I'm having trouble with defining the estimator. My goal is the following:<\/p>\n\n<ol>\n<li>Create a tf.data.Dataset.<\/li>\n<li>Feed it into the RNN estimator.<\/li>\n<\/ol>\n\n<p>The first part seems to be working correctly. I define the <\/p>\n\n<pre><code>def _parse_func(record):\n    # takes tf record as input and returns the following tensors\n    # numeric_tensor.shape = (5,170) and y.shape=()\n    return {'numerical': numeric_tensor,}, y\n\ndef input_fn(filenames=['data.tfrecord']):\n    # Returns parsed tf record i.e. the tf.data.Dataset\n    dataset = tf.data.TFRecordDataset(filenames=filenames)\n    dataset = dataset.map(map_func=_parse_func)\n    dataset = dataset.repeat()\n    dataset = dataset.batch(batch_size=BATCH_SIZE)\n    return dataset\n<\/code><\/pre>\n\n<p><em>Now let's move onto the meaty part<\/em>.<\/p>\n\n<p>Estimators take care of creating the session and graph. So I simply create the estimator in the following format:<\/p>\n\n<pre><code># create the column\ncolumn = tf.contrib.feature_column.sequence_numeric_column('numerical')\n\n# create the estimator\nestimator = RNNEstimator(\n    head=tf.contrib.estimator.regression_head(),\n    sequence_feature_columns=[column],\n    num_units=[32, 16], cell_type='lstm')\n\n# train the estimator\nestimator.train(input_fn=input_fn, steps=100)\n<\/code><\/pre>\n\n<p>However, this doesn't work. It gives me a variety of errors! In particularly, at the moment I get:<\/p>\n\n<blockquote>\n  <p>TypeError: Input must be a SparseTensor.<\/p>\n<\/blockquote>\n\n<p>Additionally, I seem to be unable to change the loss to log-loss. I tried setting it by passing it to the head parameter using:<\/p>\n\n<p><code>head = tf.contrib.estimator.regression_head(loss_fn=tf.losses.log_loss)<\/code><\/p>\n","answers":[]}
+{"QuestionId":53134404,"AnswerCount":1,"Tags":"<python><keras><resnet><transfer-learning>","CreationDate":"2018-11-03T18:43:42.217","AcceptedAnswerId":null,"OwnerUserId":10396961.0,"Title":"Keras Transfer Learning Resnet50 using fit_generator got high acc but low val_acc problem","Body":"<p>I'm using Resnet50 model to do transfer learning, using 100,000 images in total of 20 scenes(MIT Place365 dataset). I trained only the last 160 layers(due to the memory restriction). The problem is I got a pretty high accuracy but extremely low validation accuracy, I think this might be an overfitting problem but I don't know how to solve it. I will really appreciate if anyone can give me advice about how to solve my low val_acc problem, thank you very much.\nMy code is as follows:<\/p>\n\n<pre><code>V1 = np.load(\"C:\/Users\/Desktop\/numpydataKeras_20_val\/imgonehot_val_500.npy\")\nV2 = np.load(\"C:\/Users\/Desktop\/numpydataKeras_20_val\/labelonehot_val_500.npy\") \n\n\nnet = keras.applications.resnet50.ResNet50(include_top=False, weights='imagenet', input_tensor=None, input_shape=(224, 224, 3))\n\nx = net.output\nx = Flatten()(x)\nx = Dense(128)(x)\nx = Activation('relu')(x)\nx = Dropout(0.5)(x)\noutput_layer = Dense(20, activation='softmax', name='softmax')(x)\nnet_final = Model(inputs=net.input, outputs=output_layer)\n\nfor layer in net_final.layers[:-160]:\n    layer.trainable = False\nfor layer in net_final.layers[-160:]:\n    layer.trainable = True\n\nnet_final.compile(Adam(lr=.00002122), loss='categorical_crossentropy', metrics=['accuracy'])\n\ndef data_generator():\n    n = 100000\n    Num_batch = 100000\/100\n    arr = np.arange(1000)\n    np.random.shuffle(arr)\n    while (True):\n        for i in arr:\n            seed01 = random.randint(0,1000000)\n\n            X_batch  = np.load( \"C:\/Users\/Desktop\/numpydataKeras\/imgonehot_\"+str((i+1)*100)+\".npy\" )\n            np.random.seed(seed01)\n            np.random.shuffle(X_batch)\n\n            y_batch = np.load( \"C:\/Users\/Desktop\/numpydataKeras\/labelonehot_\"+str((i+1)*100)+\".npy\" )\n            np.random.seed(seed01)\n            np.random.shuffle(y_batch)\n\n            yield X_batch, y_batch\n\nweights_file = 'C:\/Users\/Desktop\/Transfer_learning_resnet50_fit_generator_02s.h5'\nearly_stopping = EarlyStopping(monitor='val_acc', patience=5, mode='auto', verbose=2)\nmodel_checkpoint = ModelCheckpoint(weights_file, monitor='val_acc', save_best_only=True, verbose=2)\ncallbacks = [early_stopping, model_checkpoint]\n\nmodel_fit = net_final.fit_generator(\n    data_generator(),\n    steps_per_epoch=1000,\n    epochs=5,\n    validation_data=(V1, V2),\n    callbacks=callbacks,\n    verbose=1,\n    pickle_safe=False)\n<\/code><\/pre>\n\n<p>The followings are the printouts:<\/p>\n\n<pre><code>Epoch 1\/5\n1000\/1000 [==============================] - 3481s 3s\/step - loss: 1.7917 - acc: 0.4757 - val_loss: 3.5872 - val_acc: 0.0560\n\nEpoch 00001: val_acc improved from -inf to 0.05600, saving model to C:\/Users\/Desktop\/Transfer_learning_resnet50_fit_generator_02s.h5\nEpoch 2\/5\n1000\/1000 [==============================] - 4884s 5s\/step - loss: 1.1287 - acc: 0.6595 - val_loss: 4.2113 - val_acc: 0.0520\n\nEpoch 00002: val_acc did not improve from 0.05600\nEpoch 3\/5\n1000\/1000 [==============================] - 4964s 5s\/step - loss: 0.8033 - acc: 0.7464 - val_loss: 4.9595 - val_acc: 0.0520\n\nEpoch 00003: val_acc did not improve from 0.05600\nEpoch 4\/5\n1000\/1000 [==============================] - 4961s 5s\/step - loss: 0.5677 - acc: 0.8143 - val_loss: 4.5484 - val_acc: 0.0520\n\nEpoch 00004: val_acc did not improve from 0.05600\nEpoch 5\/5\n1000\/1000 [==============================] - 4928s 5s\/step - loss: 0.3999 - acc: 0.8672 - val_loss: 4.6155 - val_acc: 0.0400\n\nEpoch 00005: val_acc did not improve from 0.05600\n<\/code><\/pre>\n","answers":[{"AnswerId":"54325478","CreationDate":"2019-01-23T10:47:36.837","ParentId":null,"OwnerUserId":"1472048","Title":null,"Body":"<p>Following <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9214#issuecomment-397916155\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/9214#issuecomment-397916155<\/a> it seems that batch normalization should be trainable.<\/p>\n\n<p>The following code can replace the loop where you set\/unset trainable layers:<\/p>\n\n<pre><code>for layer in model.layers:\n    if hasattr(layer, 'moving_mean') and hasattr(layer, 'moving_variance'):\n        layer.trainable = True\n        K.eval(K.update(layer.moving_mean, K.zeros_like(layer.moving_mean)))\n        K.eval(K.update(layer.moving_variance, K.zeros_like(layer.moving_variance)))\n    else:\n        layer.trainable = False\n<\/code><\/pre>\n\n<p>On my own data, I needed to reduce batch size to avoid OOM, and I now have:<\/p>\n\n<pre><code>Epoch 1\/10\n470\/470 [==============================] - 90s 192ms\/step - loss: 0.3513 - acc: 0.8660 - val_loss: 0.1299 - val_acc: 0.9590\nEpoch 2\/10\n470\/470 [==============================] - 83s 177ms\/step - loss: 0.2204 - acc: 0.9163 - val_loss: 0.1276 - val_acc: 0.9471\nEpoch 3\/10\n470\/470 [==============================] - 83s 177ms\/step - loss: 0.2219 - acc: 0.9184 - val_loss: 0.1048 - val_acc: 0.9589\nEpoch 4\/10\n470\/470 [==============================] - 83s 177ms\/step - loss: 0.1813 - acc: 0.9327 - val_loss: 0.1857 - val_acc: 0.9303\n<\/code><\/pre>\n\n<p>Warning, it may impact accuracy and you must freeze your model to avoid weird inference. But it seems to be the only way that worked for me.<\/p>\n\n<p>Another comment <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9214#issuecomment-422490253\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/9214#issuecomment-422490253<\/a> only checks the layer names to set it trainable if it's a batch normalization, but it didn't changed anything for me. Maybe it can help for your dataset.<\/p>\n"}]}
+{"QuestionId":53134423,"AnswerCount":1,"Tags":"<python><tensorflow><memory><storage>","CreationDate":"2018-11-03T18:46:46.030","AcceptedAnswerId":53135081.0,"OwnerUserId":5555557.0,"Title":"Reading images from sub hard drive for tensorflow","Body":"<p>I want to run a tensorflow script that will read a large number of images. However, the images will not fit on my local disk drive so I am storing them on a remote usb drive. <\/p>\n\n<p>I'm wondering how I will be able to read the images from the usb drive and use them to train my model if the images will not fit on my local hard drive? Will I need to read the images from the usb drive \"batch_size\" at a time? Or can I just read all of the images at once from the usb drive and divide them into batches as I normally would if they were stored on my local hard drive? <\/p>\n","answers":[{"AnswerId":"53135081","CreationDate":"2018-11-03T20:09:02.417","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>it's not recommended to load all image at once to be honest. Your usb-stick is just another path that can read in, in linux it will be stored under the \/media folder in your root. Then I recommend to get all image paths in memory like this:<\/p>\n\n<pre><code>import os\n\nfolder_to_image_map = \"\"\nimage_paths = [os.path.join(folder_to_image_map, p) for p in os.listdir(folder_to_image_map)]\n<\/code><\/pre>\n\n<p>Image paths will load in all image paths from the folder provided.\nThen you make a generator something like this:<\/p>\n\n<pre><code>from PIL import Image\n\ndef generator(BATCH_SIZE)\n    batch = []\n    for file_name in image_paths:\n        img = Image.open(file_name)\n        \"\"\" implement your preprocessing \"\"\"\n        batch.append(img)\n        if len(batch) &gt;= BATCH_SIZE:\n            g_batch = batch\n            batch = []\n            yield g_batch\n<\/code><\/pre>\n\n<p>The generator will continue in the for loop after every call. Check out this example to see how it runs, play around with it.<\/p>\n\n<pre><code>def generator():\n    ll = [1,2,3,4,5,6,7,8,9,10]\n    for l in ll:\n        yield l\n\nrun_gen = generator()\nprint(next(run_gen)) \nprint(next(run_gen))\n\n# or:\nfor j in run_gen:\n    print(j)\n<\/code><\/pre>\n\n<p>This code is not the most efficient one, but it should work, I can't test it at the moment. <\/p>\n\n<p>I Hopefully helped a bit, ask away if you've questions<\/p>\n"}]}
+{"QuestionId":53134596,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-11-03T19:07:51.710","AcceptedAnswerId":null,"OwnerUserId":10597885.0,"Title":"Tensorflow: Same input data, different output","Body":"<p>After training the model, I save it and load to make some tests. But every time I reload the model I get a different accuracy and results with the exactly same input data. After training the model I print the accuracy and it always gets a nice value (0.8 ~ 0.9), but when I reload it goes down to something like (0.1 ~ 0.5) - I dont know if it is something related to the problem btw thats weird.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport json\n\nn_nodes_hl1 = 1600\nn_nodes_hl2 = 800\nn_nodes_hl3 = 400\nn_nodes_hl4 = 200\n\nn_classes = 4\nbatch_size = 50\n\ninput_lenght = 65\nx = tf.placeholder('float', [None, input_lenght])\ny = tf.placeholder('float')\n\n\ndef train_network(x):\n    prediction = neural_network_model(x)\n    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=prediction, labels=y))\n    optimizer = tf.train.AdamOptimizer(learning_rate=0.00001).minimize(cost)\n\n    hm_epochs = 20000\n    saver = tf.train.Saver()\n    init_op = tf.global_variables_initializer()\n    sess = tf.Session()\n    sess.run(init_op)\n    epoch = 0 \n\n    for epoch in range(hm_epochs):\n        epoch_cost = 0\n        i = 0\n        while i &lt; len(train_x):\n            start = i\n            end = i + batch_size\n\n            batch_x = np.array(train_x[start:end])\n            batch_y = np.array(train_y[start:end])\n\n            _, c = sess.run([optimizer, cost], feed_dict={x: batch_x, y: batch_y})\n\n            epoch_cost += c\n            i += batch_size\n\n    save_path = saver.save(sess, \"drive\/My Drive\/datasets\/tensorflow\/model\")\n\n    correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n    print(\"accuracy:\", accuracy.eval({x: test_x, y: test_y}, session=sess))\n\n\n    sess.close()    \n\ndef group_test_train(features_data, labels_data, test_size):\n    featureset = []\n\n    for i in range(test_size):\n        featureset += [[features_data[i], labels_data[i]]]\n\n    featureset = np.array(featureset)\n\n    np.random.shuffle(featureset)\n\n    train_x = list(featureset[:, 0][:test_size \/\/ 2])\n    train_y = list(featureset[:, 1][:test_size \/\/ 2])\n\n    test_x = list(featureset[:, 0][test_size \/\/ 2:])\n    test_y = list(featureset[:, 1][test_size \/\/ 2:])\n\n    return train_x, train_y, test_x, test_y\n\ndef neural_network_model(data):\n    hidden1 = {'weights': tf.Variable(tf.random_uniform([input_lenght, n_nodes_hl1], -1, 1)),\n               'biases': tf.Variable(tf.random_normal([n_nodes_hl1]))\n               }\n\n    hidden2 = {'weights': tf.Variable(tf.random_uniform([n_nodes_hl1, n_nodes_hl2], -1, 1)),\n               'biases': tf.Variable(tf.random_normal([n_nodes_hl2]))\n               }\n\n    hidden3 = {'weights': tf.Variable(tf.random_uniform([n_nodes_hl2, n_nodes_hl3], -1, 1)),\n               'biases': tf.Variable(tf.random_normal([n_nodes_hl3]))\n               }\n\n    hidden4 = {'weights': tf.Variable(tf.random_uniform([n_nodes_hl3, n_nodes_hl4], -1, 1)),\n               'biases': tf.Variable(tf.random_normal([n_nodes_hl4]))\n               }\n\n    l_output = {'weights': tf.Variable(tf.random_uniform([n_nodes_hl4, n_classes], -1, 1)),\n                'biases': tf.Variable(tf.random_normal([n_classes]))\n                }\n\n    l1 = tf.add(tf.matmul(data, hidden1['weights']), hidden1['biases'])\n    l1 = tf.nn.relu(l1)\n\n    l2 = tf.add(tf.matmul(l1, hidden2['weights']), hidden2['biases'])\n    l2 = tf.nn.relu(l2)\n\n    l3 = tf.add(tf.matmul(l2, hidden3['weights']), hidden3['biases'])\n    l3 = tf.nn.relu(l3)\n\n    l4 = tf.add(tf.matmul(l3, hidden4['weights']), hidden4['biases'])\n    l4 = tf.nn.relu(l4)\n\n    output = tf.add(tf.matmul(l4, l_output['weights']), l_output['biases'])\n\n    return output\n\n\n\nversion = 'end'\nwith open('drive\/My Drive\/datasets\/json\/' + 'data-'+ version +'.json') as json_file:  \n    x_, y_ = json.load(json_file)\n\n\ntrain_x, train_y, test_x, test_y = group_test_train(x_, y_, len(x_) )\ntrain_network(x)\n<\/code><\/pre>\n\n<p>Every time I run this part down bellow the accuracy changes and the output as well. <\/p>\n\n<pre><code>prediction = neural_network_model(x)\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=prediction, labels=y))\noptimizer = tf.train.AdamOptimizer(learning_rate=0.00001).minimize(cost)\n\ninit_op = tf.global_variables_initializer()\n\n\nsess = tf.Session()\nsess.run(init_op)      \nnew_saver = tf.train.import_meta_graph('drive\/My Drive\/datasets\/tensorflow\/model.meta')\nnew_saver.restore(sess, tf.train.latest_checkpoint('drive\/My Drive\/datasets\/tensorflow\/'))\n\n\ncorrect = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n\nprint(\"accuracy:\", accuracy.eval({x: train_x, y: train_y}, session=sess))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53134689,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-03T19:18:31.693","AcceptedAnswerId":null,"OwnerUserId":8292871.0,"Title":"Out of Memory training sequential models in for loop; previous solutions not working","Body":"<p>I'm training a series of models in a for loop - to test a certain architecture. While doing so, I run out of memory and the system shuts down the process. <\/p>\n\n<p>The same problem appears in <a href=\"https:\/\/stackoverflow.com\/questions\/42886049\/keras-tensorflow-cpu-training-sequential-models-in-loop-eats-memory\">this question<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/42765832\/training-multiple-sequential-models-in-a-row-slows-down\/42816860#42816860\">this question<\/a>. To try their solutions, I did a test run with a similar loop to the one that is giving me problems. The code is:<\/p>\n\n<pre><code>def mem_test(n):\n    train_data = np.random.rand(1000,1500)\n    train_labels = np.random.randint(2,size= 1000)\n    mem = []\n    for i in range(n):\n        model = keras.Sequential([keras.layers.Dense(1000, activation= tf.nn.relu), \n                          keras.layers.Dense(2,activation = tf.nn.softmax)])\n        model.compile(optimizer= tf.train.AdamOptimizer(.001), loss = 'sparse_categorical_crossentropy', \n                      metrics = ['accuracy'])\n        model.fit(train_data,train_labels, epochs = 1)\n        mem.append(psutil.virtual_memory())\n    return mem\n\n\ndef mem_test_clear(n):\n    train_data = np.random.rand(1000,1500)\n    train_labels = np.random.randint(2,size= 1000)\n    mem = []\n    for i in range(n):\n        model = keras.Sequential([keras.layers.Dense(1000, activation= tf.nn.relu), \n                          keras.layers.Dense(2,activation = tf.nn.softmax)])\n        model.compile(optimizer= tf.train.AdamOptimizer(.001), loss = 'sparse_categorical_crossentropy', \n                      metrics = ['accuracy'])\n        model.fit(train_data,train_labels, epochs = 1)\n        mem.append(psutil.virtual_memory())\n        keras.backend.clear_session()\n        tf.reset_default_graph()\n    return mem\n<\/code><\/pre>\n\n<p>while the latter seems to do slightly better than the former, they both still end up accumulating memory usage. So, for my actual application of this, I'm left without a solution. What do I need to do in order to actually free up memory in this situation? What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"53134933","CreationDate":"2018-11-03T19:49:16.770","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You have to compile only once the model.\nThen you can build a loop for fitting it:<\/p>\n\n<pre><code>import numpy as np\nimport psutil\nimport keras\nimport tensorflow as tf\n\ndef mem_test(n):\n    train_data = np.random.rand(1000,1500)\n    train_labels = np.random.randint(2,size= 1000)\n    mem = []\n\n    model = keras.Sequential([keras.layers.Dense(1000, activation= tf.nn.relu), \n                  keras.layers.Dense(2,activation = tf.nn.softmax)])\n    model.compile(optimizer= tf.train.AdamOptimizer(.001), loss = 'sparse_categorical_crossentropy', \n                  metrics = ['accuracy'])\n\n    for i in range(n):\n        model.fit(train_data,train_labels, epochs = 1)\n        mem.append(psutil.virtual_memory())\n    return mem\n\nmem_test(50)\n<\/code><\/pre>\n\n<p>This way it will consume just a tiny amount of memory and will not accumulate anything. Furthermore this is the way how your model will work correctly.<\/p>\n"}]}
+{"QuestionId":53134695,"AnswerCount":1,"Tags":"<tensorflow><jupyter-notebook>","CreationDate":"2018-11-03T19:19:15.153","AcceptedAnswerId":null,"OwnerUserId":2350784.0,"Title":"Tensorflow version different in conda env & jupyter notebook","Body":"<p>I created a conda env and installed tensorflow 1.4<\/p>\n\n<pre><code>pip install tensorflow==1.4\n<\/code><\/pre>\n\n<p>This installed, and from within the environment, when i type<\/p>\n\n<pre><code>import tensorflow as tf\ntf.__version__\n<\/code><\/pre>\n\n<p>it prints <code>'1.4.0'<\/code><\/p>\n\n<p>Now, i open jupyter notebook from the same environment, and when i type the same in the notebook, it gives me <code>1.10.0<\/code><\/p>\n\n<p>How is this possible? Jupyter notebook should take the same libraries within which it is running. How does it have a different version of tensorflow<\/p>\n\n<p>And btw I'm concerned with this because i wanna use 1.4.0 in my notebook<\/p>\n","answers":[{"AnswerId":"53136338","CreationDate":"2018-11-03T23:14:29.987","ParentId":null,"OwnerUserId":"10553976","Title":null,"Body":"<p>These are the questions I ask myself in situations like this, and they usually resolve my issue:<\/p>\n\n<p>Was the kernel of my notebook running while I <code>pip<\/code>ped? Shut down and restart.<\/p>\n\n<p>Are <code>pip<\/code> and <code>conda<\/code> sharing PATH, et cetera? I don't take it for granted that my pip is modifying my conda env. You can <a href=\"https:\/\/conda.io\/docs\/user-guide\/tasks\/manage-environments.html#viewing-a-list-of-the-packages-in-an-environment\" rel=\"nofollow noreferrer\">check with these commands.<\/a><\/p>\n\n<p>Your conda env is created, but is it active? You <a href=\"https:\/\/conda.io\/docs\/user-guide\/tasks\/manage-environments.html#viewing-a-list-of-your-environments\" rel=\"nofollow noreferrer\">list created envs with these,<\/a> and you can <a href=\"https:\/\/conda.io\/docs\/user-guide\/tasks\/manage-environments.html#activating-an-environment\" rel=\"nofollow noreferrer\">activate\/deactivate with these.<\/a><\/p>\n\n<p>Besides checking all of these, <em>where am I launching jupyter from?<\/em> I like to install jupyter with a simply <code>pip install jupyter<\/code> and that could add <code>jupyter<\/code> to the command line globally. Anaconda navigator\/command line reliably launches the jupyter that lives in your Anaconda distribution.<\/p>\n"}]}
+{"QuestionId":53135131,"AnswerCount":2,"Tags":"<tensorflow><build>","CreationDate":"2018-11-03T20:15:48.760","AcceptedAnswerId":null,"OwnerUserId":10601485.0,"Title":"TensorFlow build failed on win10 ( error C2338 )","Body":"<p>I try build TensorFlow on win10. Sources download 02 november 2018.\nError<\/p>\n\n<blockquote>\n  <p>C:\\Program Files (x86)\\Windows\n  Kits\\10\\include\\10.0.17134.0\\ucrt\\assert.h\n  .\\tensorflow\/compiler\/tf2xla\/cpu_function_runtime.h(71): error C2338:<\/p>\n<\/blockquote>\n\n<pre><code>std::pair&lt;uint64, uint64&gt; Encode() const {\nstatic_assert(sizeof(*this) == 16, \"\"); &lt;----- assert\nuint64 upper = Pack(kind(), size_);\nuint64 lower = entry_param_number_;\nreturn {upper, lower};\n<\/code><\/pre>\n\n<p>}<\/p>\n\n<p>Bazel config<\/p>\n\n<pre><code>build:xla --define with_xla_support=true\nbuild --config=xla\nbuild --action_env TF_NEED_OPENCL_SYCL=\"0\"\nbuild --action_env TF_NEED_ROCM=\"0\"\nbuild --action_env TF_NEED_CUDA=\"1\"\nbuild --action_env CUDA_TOOLKIT_PATH=\"c:\/Program Files\/NVIDIA GPU Computing Toolkit\/CUDA\/v9.2\"\nbuild --action_env TF_CUDA_VERSION=\"9.2\"\nbuild --action_env CUDNN_INSTALL_PATH=\"c:\/Program Files\/NVIDIA GPU Computing Toolkit\/CUDA\/v9.2\"\nbuild --action_env TF_CUDNN_VERSION=\"7\"\nbuild --action_env TF_CUDA_COMPUTE_CAPABILITIES=\"6.1\"\nbuild --action_env TF_CUDA_CLANG=\"0\"\nbuild --config=cuda\ntest --config=cuda\nbuild:opt --copt=\/arch:AVX\nbuild:opt --define with_default_optimizations=true\nbuild --config monolithic\nbuild --copt=-w --host_copt=-w\nbuild --verbose_failures\nbuild --distinct_host_configuration=false\nbuild --experimental_shortened_obj_file_path=true\nbuild:v2 --define=tf_api_version=2\n<\/code><\/pre>\n","answers":[{"AnswerId":"53159125","CreationDate":"2018-11-05T17:16:25.447","ParentId":null,"OwnerUserId":"1666054","Title":null,"Body":"<p>A couple of things to look at:<\/p>\n\n<p>1.) It looks like you are telling the compiler to find the cuDNN in the same path as CUDA. I downloaded cuDNN separate and unzipped it to C:\\tools\\Cuda\\<\/p>\n\n<p>2.) I had an error similar to this and I set cuDNN version to 7.3.1 (not just 7) and opted to NOT use the C++ eigen bypass. <\/p>\n"},{"AnswerId":"53214777","CreationDate":"2018-11-08T19:24:02.517","ParentId":null,"OwnerUserId":"10601485","Title":null,"Body":"<p>I download another release - r1.12 -another error was occurred (in protobuf - invalid type cast). I just download the binaries for cuda 9.2 from \n<a href=\"https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\/blob\/master\/1.8.0\/py36\/GPU\/cuda92cudnn71sse2\/tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\/blob\/master\/1.8.0\/py36\/GPU\/cuda92cudnn71sse2\/tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl<\/a><\/p>\n"}]}
+{"QuestionId":53135439,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-11-03T20:54:47.747","AcceptedAnswerId":null,"OwnerUserId":8443897.0,"Title":"Issue with add method in tensorflow : AttributeError: module 'tensorflow.python.framework.ops' has no attribute '_TensorLike'","Body":"<pre><code>import keras as K\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom tensorflow import set_random_seed\n\nfor hidden_neuron in hidden_neurons:\n  model = Sequential()\n<\/code><\/pre>\n\n<p><code>model.add(Dense(hidden_neuron, input_dim=61, activation='relu'))<\/code><\/p>\n\n<p>-> i am getting error at this line. I am not really sure what am i missing here.<\/p>\n\n<p>Traceback (most recent call last):<\/p>\n\n<blockquote>\n  <p>File \"PycharmProjects\/HW2\/venv\/bin\/hw3q4.py\", line 46, in \n      model.add(Dense(hidden_neuron, input_dim=61, activation='relu'))   File\n  \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/engine\/sequential.py\",\n  line 165, in add\n      layer(x)   File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py\",\n  line 414, in <strong>call<\/strong>\n      self.assert_input_compatibility(inputs)   File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py\",\n  line 279, in assert_input_compatibility\n      K.is_keras_tensor(x)   File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\",\n  line 472, in is_keras_tensor\n      if not is_tensor(x):   File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\",\n  line 480, in is_tensor\n      return isinstance(x, tf_ops._TensorLike) or tf_ops.is_dense_tensor_like(x)\n  <strong>AttributeError: module 'tensorflow.python.framework.ops' has no attribute '_TensorLike'<\/strong><\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53135494,"AnswerCount":2,"Tags":"<python><keras><multiprocessing>","CreationDate":"2018-11-03T21:01:19.717","AcceptedAnswerId":null,"OwnerUserId":8437974.0,"Title":"BrokenProcessPool on using n_jobs parameter in cross_val_score","Body":"<p>I'm getting error when using <code>n_jobs = -1<\/code> as a parameter in  <code>sklearn.model_selection.cross_val_score<\/code>. I'm a beginner in Deep Learning and ANN and as per the instructor in <a href=\"https:\/\/www.udemy.com\/deeplearning\/\" rel=\"noreferrer\">this<\/a> course in k-fold Cross-Validation, use <code>n_jobs = -1<\/code> to use all the processors of CPU so as to reduce the time but it's throwing an error in my case.<\/p>\n\n<p>Error -<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>Full stack trace can be found <a href=\"https:\/\/pastebin.com\/wscHkN8H\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<pre><code>import keras\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom sklearn.model_selection import cross_val_score\ndef build_classifier():\n    classifier = Sequential()\n    classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu', input_dim = 11))\n    classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu'))\n    classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n    classifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n    return classifier\nclassifier = KerasClassifier(build_fn = build_classifier, batch_size = 10, nb_epoch = 100)\naccuracies = cross_val_score(estimator = classifier, X = X_train, y = Y_train, cv = 10, n_jobs = -1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"54249892","CreationDate":"2019-01-18T08:10:05.043","ParentId":null,"OwnerUserId":"10932232","Title":null,"Body":"<p>n_jobs =-1 isn't working because your gpu is off or I say gpu is not activated,u can activate it with cuda command or activate tensorflow-gpu using these links<\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=Ebo8BklTtmc&amp;t=138s\" rel=\"nofollow noreferrer\">Install Tensorflow (GPU version) for Windows and Anaconda<\/a> <\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=HExRhnO5Mqs\" rel=\"nofollow noreferrer\">How to install Tensorflow-GPU on Windows 10<\/a> <\/p>\n\n<p>Or u can refer to this, for n_jobs understanding<\/p>\n\n<p><a href=\"https:\/\/scikit-learn.org\/stable\/glossary.html#term-n-jobs\" rel=\"nofollow noreferrer\">Glossary of Common Terms and API Elements-term-n-jobs<\/a><\/p>\n"},{"AnswerId":"58420112","CreationDate":"2019-10-16T19:05:02.010","ParentId":null,"OwnerUserId":"12228424","Title":null,"Body":"<p>Try creating your <code>build_classifier<\/code> function in an external file and importing it. E.g:\n<\/p>\n\n<p>in file <code>classifier_builder.py<\/code>: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras\n\ndef build_classifier():\n  classifier = Sequential()\n  classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu', input_dim = 11))\n  classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu'))\n  classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n  classifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\nreturn classifier\n<\/code><\/pre>\n\n<p>and then in your notebook:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import classifier_builder\n\nclassifier = KerasClassifier(build_fn = build_classifier, batch_size = 10, nb_epoch = 100)\naccuracies = cross_val_score(estimator = classifier, X = X_train, y = Y_train, cv = 10, n_jobs = -1)\n<\/code><\/pre>\n\n<p>This solved the issue for me. Apparently, the inline function is not picklable.<\/p>\n"}]}
+{"QuestionId":53136160,"AnswerCount":0,"Tags":"<python><tensorflow><flask><keras>","CreationDate":"2018-11-03T22:43:45.207","AcceptedAnswerId":null,"OwnerUserId":4414359.0,"Title":"OSError: cannot identify image file <_io.BytesIO object at ... > when trying to predict with Keras model","Body":"<p>I was following a \"Deploy Keras neural network to Flask web service\" tutorial by deeplizard on YouTube, but got stuck.\nI found several suggestions to similar problems which have helped others (like <a href=\"https:\/\/stackoverflow.com\/questions\/31077366\/pil-cannot-identify-image-file-for-io-bytesio-object\/46394242\">this one<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/23587426\/pil-open-method-not-working-with-bytesio\">this one<\/a>), but for some reason they aren't working for me. Or maybe I'm applying them wrong.<\/p>\n\n<p>It's upset with the line <code>image = Image.open(io.BytesIO(decoded))<\/code>.<\/p>\n\n<p>Here is the code I have (sorry it's not minimal, I'm not sure how to simplify it without dropping details that might be relevant).<\/p>\n\n<p>If you have any suggestions, please let me know.<\/p>\n\n<p>Much appreciated.<\/p>\n\n<pre><code>app = Flask(__name__)\n\n\ndef get_model():\n    global model, graph\n    model = load_model('model.h5')\n    print(' * Model loaded!')\n    graph = tf.Graph()\n\n\ndef preprocess_image(image, target_size):\n    if image.mode != 'RGB':\n        image = image.convert('RGB')\n    image = image.resize(target_size)\n    image = img_to_array(image)\n    image = np.expand_dims(image, axis=0)\n\n    return image\n\n\nprint(' * Loading model...')\nget_model()\n\n\n@app.route('\/predict', methods=[\"POST\"])\ndef predict():\n    message = request.get_json(force=True)\n    encoded = message['image']\n    decoded = base64.b64decode(encoded)\n    with graph.as_default():\n        image = Image.open(io.BytesIO(decoded))\n        preprocessed_image = preprocess_image(image, target_size(50, 50))\n        prediction = model.predict(preprocessed_image).tolist()\n\n        response = {\n            'prediction': {\n                'food': prediction[0][0],\n                'notfood': prediction[0][1]\n            }\n        }\n    return jsonify(response)\n<\/code><\/pre>\n\n<p>I suspect it might be because my model takes input as:<\/p>\n\n<p><code>model.predict_classes(i.reshape((-1, 50, 50, 3)), batch_size=32, verbose=0)[0]<\/code><\/p>\n\n<p>but the images uploaded through html by user aren't' getting reshaped ... I was trying to finagle that into the code, but no luck so far.<\/p>\n","answers":[]}
+{"QuestionId":53136212,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-11-03T22:53:52.067","AcceptedAnswerId":null,"OwnerUserId":5068933.0,"Title":"keras model.predict randomises the output order?","Body":"<p>I am new to keras and Tensorflow. I have trained a model in Keras using Tensorflow as backend, for reducing a specific noise of images. When I use this model to predict a set of test images(noisy, shape=[batchsize_1, width_1, height_1]), I have got correct predicted images(noise removed, shape=[batchsize_1, width_1, height_1]), but the order of the predicted images are randomised. As the order of the predicted images are important to my next step, how could I keep the order of predicted image identical to test images?<\/p>\n\n<p>Here are my code for loading the model and predict the test data:<\/p>\n\n<pre><code>json_file = open('model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\nloaded_model.load_weights(\"model.h5\")\nadam=Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-8)\nloaded_model.compile(loss=losses.mean_absolute_error, optimizer=adam)\npredict_images= loaded_model.predict(test_images, batch_size=16)\n<\/code><\/pre>\n\n<p>I am sure that the <code>test_images<\/code> are in order in the dimension of batchsize, but the batchsize order of <code>predict_images<\/code> is randomised. Does the order of <code>predict_image<\/code> also depend on how I trained the model? How could I remain the order of <code>predict_image<\/code> unchanged?<\/p>\n\n<p>Any comments and help would be much appreciated. Thanks a lot!<\/p>\n","answers":[]}
+{"QuestionId":53137115,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-04T01:56:10.990","AcceptedAnswerId":53137515.0,"OwnerUserId":10395181.0,"Title":"Why does my tensorflow code runs more and more slower as the training goes on?","Body":"<pre><code>import tensorflow as tf\n\ninput_data = tf.constant([[1.,1.]])\noutput_data = tf.constant([[1.,0.]])\nweight = tf.Variable([[1.,1.],\n                      [1.,1.]])\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\noptimizer = tf.train.GradientDescentOptimizer(0.1)\n\nfor epoch in range(1000):\n    y = tf.matmul(input_data, weight)\n    loss = (output_data[0][0] - y[0][0])**2 + (output_data[0][1] - y[0][1])**2\n    sess.run(optimizer.minimize(loss))\n    print(epoch)\n<\/code><\/pre>\n\n<p>The above program gets more and more slower as the epoch goes on. I think it's because new nodes are keep being added for each epoch. How do I handle this?<\/p>\n","answers":[{"AnswerId":"53137515","CreationDate":"2018-11-04T03:30:48.800","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>Try this...<\/p>\n\n<pre><code>import time\nimport tensorflow as tf\n\ninput_data = tf.constant([[1.,1.]])\noutput_data = tf.constant([[1.,0.]])\nweight = tf.Variable([[1.,1.],\n                      [1.,1.]])\n\noptimizer = tf.train.GradientDescentOptimizer(0.1)\ny = tf.matmul(input_data, weight)\nloss = (output_data[0][0] - y[0][0])**2 + (output_data[0][1] - y[0][1])**2\ntrain = optimizer.minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print('Initial weights: ', sess.run(weight))\n    for epoch in range(1000):\n        st = time.time()\n        sess.run(train)\n        print('Epoch %3d : %.3f ms' %(epoch, 1e3*(time.time()-st)))\n    print('Weights: ', sess.run(weight))\n<\/code><\/pre>\n\n<p>The original code is recreating the graph each epoch.  If you do it this way, the graph is only created once and the only work in the loop is the gradient calculations\/updates.<\/p>\n"}]}
+{"QuestionId":53137546,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-04T03:36:24.900","AcceptedAnswerId":53218807.0,"OwnerUserId":3381072.0,"Title":"Distributed Tensorflow: The worker does not responded","Body":"<p>I am running Tensorflow Distributed mode on AWS instances. PS is on one machine and each worker is on a different machine. I am running in the following problem:<\/p>\n\n<pre><code>tensorflow\/core\/distributed_runtime\/master.cc:267] CreateSession still waiting for response from worker: \/job:ps\/replica:0\/task:0\n<\/code><\/pre>\n\n<p>I found somebody already posts exactly the same problem that I face right now but his answer is not clear for me, <a href=\"https:\/\/stackoverflow.com\/questions\/51219374\/tensorflow-distributed-createsession-still-waiting-for-response-from-worker-j?answertab=oldest#tab-top\">Tensorflow distributed: CreateSession still waiting for response from worker: \/job:ps\/replica:0\/task:0<\/a>. <\/p>\n\n<p>Can anybody suggest what might be the solution? <\/p>\n","answers":[{"AnswerId":"53218807","CreationDate":"2018-11-09T02:01:43.150","ParentId":null,"OwnerUserId":"3381072","Title":null,"Body":"<p>Now, after I solved my error, I will share my solution. It was not a bug in the TF code but one of two things that I tried solved my problem stated above. Because I am working on EC2 instances, the firewall prevents the connection between nodes. So, I made the rule to accept all the traffics to the instances. Second, I was using only IP-address:port-No in the command line. Instead, I wrote it like this, ec2-IP.compute-1.amazonaws.com:2222.<\/p>\n"}]}
+{"QuestionId":53137588,"AnswerCount":1,"Tags":"<python-2.7><pytorch>","CreationDate":"2018-11-04T03:47:06.050","AcceptedAnswerId":null,"OwnerUserId":10552423.0,"Title":"pytorch backports.functools_lru_cache conflict","Body":"<p>I'm using <em>Windows 10<\/em>, and my the installation dir is:\n<code>anacoda2\/python2.7\/python3.6\/opencv\/cdua10\/cudnn ....<\/code> <\/p>\n\n<p>Now I want to install pytorch, with this command: <code>conda install pytorch -c pytorch<\/code><\/p>\n\n<p>But as result I'm getting this error:<\/p>\n\n<pre><code>C:\\Users\\MM&gt;conda install pytorch -c pytorch\nSolving environment: failed\n\nUnsatisfiableError: The following specifications were found to be in conflict:\n  - backports.functools_lru_cache\n  - pytorch\nUse \"conda info &lt;package&gt;\" to see the dependencies for each package.\n<\/code><\/pre>\n\n<p>The version of <code>backports.functools_lru_cache<\/code> is <code>1.4<\/code>.<\/p>\n\n<p><strong>Does anyone know how to solve this?<\/strong><\/p>\n","answers":[{"AnswerId":"53820612","CreationDate":"2018-12-17T17:51:37.830","ParentId":null,"OwnerUserId":"9035405","Title":null,"Body":"<p>I faced a similar problem while installing pytorch. I was able to resolve it by\ncreating another environment on anaconda which had python version 3.6.   <\/p>\n"}]}
+{"QuestionId":53138822,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-04T07:58:05.583","AcceptedAnswerId":53199383.0,"OwnerUserId":10598704.0,"Title":"Tensorflow - can't initialize saved variables unless I recreate the \"saver\" object. Why?","Body":"<p>I'm pretty sure I'm missing something about how tensorflow works because my solution doesn't make any sense.<\/p>\n\n<p>I'm trying to train a neural network (from scratch, without using Estimators or other abstractions), save it, and load a simplified version of it for inference.<\/p>\n\n<p>The following code trains but gives me the error: <code>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value hidden0\/biases\/Variable\n     [[Node: hidden0\/biases\/Variable\/read = Identity[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](hidden0\/biases\/Variable)]]<\/code>. If I add the commented line - if I recreate the saver obect <strong>that I'm not going to use nor return<\/strong> - the code works just fine.<\/p>\n\n<p>Why do I need to create a (useless) saver object in order to restore the saved weights?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef add_fc_layer(input_tensor, input_dimensions, output_dimensions, layer_name, activation=None):\n    with tf.variable_scope(layer_name):\n        with tf.variable_scope('weights'):\n            weights = tf.Variable(tf.truncated_normal([input_dimensions, output_dimensions]))\n        with tf.variable_scope('biases'):\n            biases = tf.Variable(tf.zeros([output_dimensions]))\n        with tf.variable_scope('Wx_plus_b'):\n            preactivate = tf.matmul(input_tensor, weights) + biases\n        if activation is None:\n            return preactivate\n        with tf.variable_scope('activation'):\n            activations = activation(preactivate)\n        return activations\n\n\ndef make_network(model_phase):\n    if model_phase not in {\"train\", \"test\"}:\n        raise ValueError(\"invalid type\")\n\n    hidden0_units = 25\n    hidden1_units = 15\n    hidden2_units = 10\n    input_size = 10\n    output_size = 4\n\n    with tf.variable_scope('InputVector'):\n        inputs = tf.placeholder(shape=[1, input_size], dtype=tf.float32)\n\n    hidden0_out = add_fc_layer(inputs, input_size, hidden0_units, \"hidden0\", activation=tf.nn.sigmoid)\n    hidden1_out = add_fc_layer(hidden0_out, hidden0_units, hidden1_units, \"hidden1\", activation=tf.nn.sigmoid)\n    hidden2_out = add_fc_layer(hidden1_out, hidden1_units, hidden2_units, \"hidden2\", activation=tf.nn.sigmoid)\n\n    out =  add_fc_layer(hidden2_out, hidden2_units, output_size, \"regression\")\n\n    if model_phase == \"test\":\n        # UNCOMMENTIN THIS LINE MAKES THE SCRIPT WORK\n        # saver = tf.train.Saver(var_list=tf.trainable_variables())\n        return inputs, out\n\n    saver = tf.train.Saver(var_list=tf.trainable_variables())\n\n    with tf.variable_scope('training'):\n        with tf.variable_scope('groundTruth'):\n            ground_truth = tf.placeholder(shape=[1, output_size], dtype=tf.float32)\n        with tf.variable_scope('loss'):\n            loss = tf.reduce_sum(tf.square(ground_truth - out))\n            tf.summary.scalar('loss', loss)\n        with tf.variable_scope('optimizer'):\n            trainer = tf.train.AdamOptimizer(learning_rate=0.001)\n        with tf.variable_scope('gradient'):\n            updateModel = trainer.minimize(loss)\n\n\n    with tf.variable_scope('predict'):\n        predict = tf.random_shuffle(tf.boolean_mask(out, tf.equal(out, tf.reduce_max(out, axis=None))))[0]\n\n\n    writer = tf.summary.FileWriter('\/tmp\/test', tf.get_default_graph())\n\n\n    return inputs, out, ground_truth, updateModel, writer, saver\n\n\ntrain_graph = tf.Graph()\nwith tf.Session(graph=train_graph) as sess:\n    tf.set_random_seed(42)\n    inputs, out, ground_truth, updateModel, writer, saver = make_network(model_phase='train')\n    init = tf.initialize_all_variables()\n    sess.run(init)\n    print('\\nLearning...')\n    for _ in range(10):\n        sess.run([updateModel], feed_dict={inputs:np.arange(10)+np.random.random((1,10)), ground_truth:np.arange(4).reshape(1, 4)})\n    saver.save(sess,'.\/tensorflowModel.ckpt')\n\nnew_graph = tf.Graph()\nwith tf.Session(graph=new_graph) as sess:\n    inputs, out = make_network(model_phase='test')\n    saver = tf.train.import_meta_graph('.\/tensorflowModel.ckpt.meta')\n    saver.restore(sess, tf.train.latest_checkpoint('.\/'))\n\n    # evaluation\n    print('\\nEvaluation...')\n    for _ in range(10):\n        _ = sess.run(out, feed_dict={inputs:np.arange(10).reshape(1,10)})\n<\/code><\/pre>\n","answers":[{"AnswerId":"53199383","CreationDate":"2018-11-07T23:20:42.127","ParentId":null,"OwnerUserId":"6780025","Title":null,"Body":"<p>I don't know why creating an unused Saver makes the problem go away, but the code betrays a misunderstanding.<\/p>\n\n<p>When you are restoring, you are creating the model graph twice. First, you call make_network() which creates the computation graph and variables. You then also call <code>import_meta_graph<\/code> which also creates a graph and variables. You should create a saver with simple <code>saver = tf.train.Saver()<\/code> instead of <code>saver = tf.train.import_meta_graph('.\/tensorflowModel.ckpt.meta')<\/code><\/p>\n"}]}
+{"QuestionId":53138899,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-04T08:10:43.853","AcceptedAnswerId":53139651.0,"OwnerUserId":10602820.0,"Title":"Tensorflow returns same weights in hidden layer","Body":"<p>I'm trying to train a model that will emulate a sine function.\nUsing 3 fully connected layers with 1-15-1 neurons. So I have 2 weight variables, 1x15 and 15x1. After training they have the same values:<\/p>\n\n<pre><code>[[0.02432664] [0.02432664] [0.02432664] [0.02432664] [0.02432664] \n[0.02432664] [0.02432664] [0.02432664] [0.02432664] [0.02432664] \n[0.02432664] [0.02432664] [0.02432665] [0.02432665] [0.02432665]]\n\n[[0.02432664 0.02432664 0.02432664 0.02432664 0.02432664 0.02432664\n  0.02432664 0.02432664 0.02432664 0.02432664 0.02432664 0.02432664\n  0.02432665 0.02432665 0.02432665]]\n<\/code><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><code>W1 = tf.Variable(tf.ones([15, 1]) * 0.3)\nW2 = tf.Variable(tf.ones([1, 15]) * 0.3)\nlabel = tf.placeholder(dtype=tf.float32, shape=[1, None])\nresult = tf.placeholder(dtype=tf.float32)\n\nl1 = tf.matmul(W1, label)\nl2 = tf.matmul(W2, l1)\nloss = tf.square(l2 - result)\nloss = tf.reduce_sum(loss)\ntrain_op = tf.train.AdamOptimizer().minimize(loss)\n\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nfeed_dict = {label: [labels], result: values}\n\nfor i in range(0, 500):\n    sess.run(train_op, feed_dict)\n<\/code><\/pre>\n\n<p>Values and labels from feed_dict look like this:<\/p>\n\n<pre><code>[0.041410759078128294, 0.5139904997395468, 0.97, 0.9710565162951535, \n 0.6471067811865476, 0.20643446504023055]\n\n[1, 15, 50, 60, 75, 95]\n<\/code><\/pre>\n\n<p>(Range 1-100 equivalent to 0 - PI)<\/p>\n\n<p>Have no idea why tensorflow is doing so. Can you suggest what has gone wrong and how to fix it?<\/p>\n","answers":[{"AnswerId":"53139651","CreationDate":"2018-11-04T10:08:15.067","ParentId":null,"OwnerUserId":"6067528","Title":null,"Body":"<p>It's because you're not randomly initializing your weights matrices <code>W1<\/code> and <code>W2<\/code>. You're using:<\/p>\n\n<pre><code>W1 = tf.Variable(tf.ones([15, 1]) * 0.3)\nW2 = tf.Variable(tf.ones([1, 15]) * 0.3)\n<\/code><\/pre>\n\n<p>But you should be initializing them using something like a random normal distribution:<\/p>\n\n<pre><code>W1 = tf.Variable(tf.random_normal((15,1), stddev=0.01)\/tf.sqrt(n))\n<\/code><\/pre>\n\n<p>This is because no what what the input is, if all weights are initialized to the same values, as in your case, all units in the hidden layer will be the same, and your network won't be able to learn.<\/p>\n"}]}
+{"QuestionId":53139104,"AnswerCount":1,"Tags":"<c#><python><unity3d><tensorflow><tensorflowsharp>","CreationDate":"2018-11-04T08:44:31.687","AcceptedAnswerId":53139216.0,"OwnerUserId":10602888.0,"Title":"tensorflow batch normalization freeze","Body":"<p>I'm trying to implement frozen graph on unity. But I have some error. I think the error related on batch normalization but I have no idea to solve this problem. <br>\nWhen I train my model <em>BN<\/em> is necessary so I can't just remove the <em>BN<\/em>. <\/p>\n\n<p>So I have to find some way to how do proper freeze the graph with <em>BN<\/em> without causing error when it implement on unity. Any Idea?<\/p>\n\n<p>I had already some researches for this problem I know not only the <em>BN<\/em> cause this error also the dropout has same problem.<\/p>\n\n<p><em>note: I'm using tensorflowsharp. network is based on ResNet34 with little modified.<\/em> <\/p>\n\n<pre><code>error: TFException: Input 0 of node bn2c_branch2a\/Assign was passed float from bn2c_branch2a\/mean:0 incompatible with expected float_ref.\nTensorFlow.TFStatus.CheckMaybeRaise (TensorFlow.TFStatus incomingStatus, System.Boolean last) (at &lt;013bee3700ac45f2a3b8843fa1d4d914&gt;:0)\nTensorFlow.TFGraph.Import (TensorFlow.TFBuffer graphDef, TensorFlow.TFImportGraphDefOptions options, TensorFlow.TFStatus status) (at &lt;013bee3700ac45f2a3b8843fa1d4d914&gt;:0)\nTensorFlow.TFGraph.Import (System.Byte[] buffer, TensorFlow.TFImportGraphDefOptions options, TensorFlow.TFStatus status) (at &lt;013bee3700ac45f2a3b8843fa1d4d914&gt;:0)\nTensorFlow.TFGraph.Import (System.Byte[] buffer, System.String prefix, TensorFlow.TFStatus status) (at &lt;013bee3700ac45f2a3b8843fa1d4d914&gt;:0)\nDeepGes.GClassifier.InitClassifier () (at Assets\/Scripts\/Gesture_Main\/GClassifier.cs:42)\nG_Game.Awake () (at Assets\/Scripts\/Gesture_Main\/G_Game.cs:67)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53139216","CreationDate":"2018-11-04T09:05:38.230","ParentId":null,"OwnerUserId":"6519679","Title":null,"Body":"<p>I am not exactly sure on what you are asking or working on, however I did some research for you to hopefully point you in the right direction.<\/p>\n\n<p><a href=\"https:\/\/blog.goodaudience.com\/tensorflow-unity-how-to-set-up-a-custom-tensorflow-graph-in-unity-d65cc1bd1ab1\" rel=\"nofollow noreferrer\">Good Audience - TensorFlow + Unity: How to set up a custom TensorFlow graph in Unity<\/a><\/p>\n\n<p>There is a guy here who is an ML engineer, so maybe try asking him.<br>\n<a href=\"https:\/\/answers.unity.com\/questions\/1517623\/using-tensorflows-inference-graph-in-unity.html\" rel=\"nofollow noreferrer\">Unity Forums - Using Tensorflow's inference graph in Unity<\/a><\/p>\n\n<p>Best of luck.<\/p>\n"}]}
+{"QuestionId":53139678,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-04T10:11:26.347","AcceptedAnswerId":53142329.0,"OwnerUserId":3885491.0,"Title":"Assigning zero to tensor at indices specified in a list","Body":"<p>I have to tensors, for example <\/p>\n\n<pre><code>A = tf.Tensor(\n       [[1.0986123 0.6931472 0.        0.6931472 0.       ]\n        [0.        0.        0.        0.        0.       ]\n        [3.7376697 3.7612002 3.7841897 3.8066626 3.8286414]], shape=(3, 5), dtype=float32)\n\nB = tf.Tensor(\n   [[2 1]\n    [2 2]], shape=(2, 2), dtype=int64)\n<\/code><\/pre>\n\n<p>Tensor <code>B<\/code> holds indices in tensor <code>A<\/code>. I want to update every value in tensor <code>A<\/code> to zero that is listed in the index list <code>B<\/code>.<\/p>\n\n<p>So, the expected result would be <\/p>\n\n<pre><code>tf.Tensor(\n       [[1.0986123 0.6931472 0.        0.6931472 0.       ]\n        [0.        0.        0.        0.        0.       ]\n        [3.7376697 0 0 3.8066626 3.8286414]], shape=(3, 5), dtype=float32)\n<\/code><\/pre>\n\n<p>So the entries at index [2,1] and [2, 2] are set to 0.<\/p>\n\n<p>I looked at <code>tf.assign<\/code> but they can only be used for <code>tf.Variable<\/code>'s. <code>tf.boolean_mask<\/code> would be a nice way to do it, but i do not know and could not find out how i can create a boolean mask with a list of indices. <\/p>\n\n<p>I looked at the tensor flow functions i could find and related S\/O answers but couldn't find a satisfying solution.<\/p>\n","answers":[{"AnswerId":"53142329","CreationDate":"2018-11-04T15:24:29.547","ParentId":null,"OwnerUserId":"8315956","Title":null,"Body":"<p>You can use <code>tf.scatter_nd_update<\/code> for this. For example:<\/p>\n\n<pre><code>A = tf.Variable(\n    [[1.0986123, 0.6931472, 0.       , 0.6931472, 0.       ],\n     [0.       , 0.       , 0.       , 0.       , 0.       ],\n     [3.7376697, 3.7612002, 3.7841897, 3.8066626, 3.8286414]], dtype=tf.float32)\n\nB = tf.Variable(\n    [[2, 1],\n     [2, 2]], dtype=tf.int64)\n\nC = tf.scatter_nd_update(A, B, tf.zeros(shape=tf.shape(B)[0]))\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(C))\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>A = tf.constant(\n    [[1.0986123, 0.6931472, 0.       , 0.6931472, 0.       ],\n     [0.       , 0.       , 0.       , 0.       , 0.       ],\n     [3.7376697, 3.7612002, 3.7841897, 3.8066626, 3.8286414]], dtype=tf.float32)\n\nB = tf.constant(\n    [[2, 1],\n     [2, 2]], dtype=tf.int64)\n\nAV = tf.Variable(A)\n\nC = tf.scatter_nd_update(AV, B, tf.zeros(shape=tf.shape(B)[0]))\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(C))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53140089,"AnswerCount":1,"Tags":"<makefile><caffe>","CreationDate":"2018-11-04T11:00:02.907","AcceptedAnswerId":null,"OwnerUserId":10603111.0,"Title":"Make all In Caffe: could not find -l lmdbboost_system","Body":"<p>The path is: \/usr\/local\/caffe<\/p>\n\n<p>When I run <code>sudo make all -j8<\/code>\nI get an error:<\/p>\n\n<pre><code>AR -o .build_release\/lib\/libcaffe.a\nLD -o .build_release\/lib\/libcaffe.so.1.0.0\n\/usr\/bin\/ld: cannot find -llmdbboost_system\ncollect2: error: ld returned 1 exit status\nMakefile:585: recipe for target '.build_release\/lib\/libcaffe.so.1.0.0' failed\nmake: *** [.build_release\/lib\/libcaffe.so.1.0.0] Error 1\n<\/code><\/pre>\n\n<p>I use caffe-cuda on ubuntu 18.04, cuda V10.0.130, nvcc V9.1.85, cuDNN V7.3.1<\/p>\n","answers":[{"AnswerId":"53140595","CreationDate":"2018-11-04T11:59:47.720","ParentId":null,"OwnerUserId":"10603111","Title":null,"Body":"<p>In the Makefile<\/p>\n\n<p>I forgot to add a space between lmdb and boost_system<\/p>\n\n<pre><code>LIBRARIES += lamdb boost_system\n<\/code><\/pre>\n"}]}
+{"QuestionId":53140853,"AnswerCount":2,"Tags":"<python><arrays><numpy><tensorflow><scikit-learn>","CreationDate":"2018-11-04T12:31:08.420","AcceptedAnswerId":53140877.0,"OwnerUserId":9771955.0,"Title":"How do I add an element to numpy nd array?","Body":"<p>I want to add a matrix of shape (60,4) to an empty numpy array of shape (1000,60,4)\nHowever when i use the following code it fails<\/p>\n\n<pre><code>np.append(arr=array,values=i)\n<\/code><\/pre>\n\n<p>In the above code , the shape of object <strong><em>i<\/em><\/strong> is (60,4),\nBy <strong>it fails<\/strong> i mean it adds a matrix full of zeroes to the array which is not expected . Am I doing anything wrong\nHow can I add the matrix to the array keeping my final output as (1000,60,4)?<\/p>\n\n<p><strong>NOTE<\/strong>\nDon't duplicate this question . It is different from questions like <a href=\"https:\/\/stackoverflow.com\/questions\/7332841\/add-single-element-to-array-in-numpy\">these<\/a>\nI have not a 1d array but a more complex nd array.\nThanks for help<\/p>\n","answers":[{"AnswerId":"53140912","CreationDate":"2018-11-04T12:40:32.200","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>If you want to <strong>assign<\/strong> a value to one of the element of array, you have to use <strong>assign<\/strong> operation, not <strong>concatenation<\/strong>, what you made:<\/p>\n\n<pre><code>import numpy as np\nnp.append(arr=array,values=i)\n<\/code><\/pre>\n\n<p>will results in a list with the shape (<strong>240240,<\/strong>)<\/p>\n\n<p><strong>Assigning<\/strong> to one of the element of array:<\/p>\n\n<pre><code>   arr[0] = i\n<\/code><\/pre>\n\n<p>will results in a list with the shape (<strong>1000,60,4<\/strong>)<\/p>\n"},{"AnswerId":"53140877","CreationDate":"2018-11-04T12:35:11.503","ParentId":null,"OwnerUserId":"8111755","Title":null,"Body":"<p>You can assign your 60,4 to any of your 1000 positions.<\/p>\n\n<pre><code>big_array[pos] = little_array\n<\/code><\/pre>\n\n<p>example:<\/p>\n\n<pre><code>import numpy as np\n\na = np.zeros((3,6,4))\nb = np.ones((6,4))\n\na[1] = b\n<\/code><\/pre>\n\n<p>yields:<\/p>\n\n<pre><code>           a[1]=  \n           [[1., 1., 1., 1.],\n           [1., 1., 1., 1.],\n           [1., 1., 1., 1.],\n           [1., 1., 1., 1.],\n           [1., 1., 1., 1.],\n           [1., 1., 1., 1.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53141144,"AnswerCount":0,"Tags":"<python><tensorflow><hashmap><computer-vision>","CreationDate":"2018-11-04T13:06:55.383","AcceptedAnswerId":null,"OwnerUserId":9088536.0,"Title":"Did my model failes to learn due to \"Bad loss function design\" or \"Mistake in training loop\"?","Body":"<p>I'm trying to build a siamese neural network for human pose estimation based on Hash matching.<\/p>\n\n<p>The <strong>basic concept and references<\/strong> are described in my <a href=\"https:\/\/stackoverflow.com\/questions\/52961507\/facing-no-gradients-for-any-variable-error-while-training-a-siamese-network\">previous post<\/a><\/p>\n\n<p>I fixed the \"no gradients provided for any variable\" error but found that my model's loss value was not improving no matter how long the training was.<\/p>\n\n<p>I saved checkpoints on 100th,10000th and 500000th iteration and the restored models have the same magnitude of loss value.<\/p>\n\n<p>Currently I consider there might be two causes of this problem:<\/p>\n\n<p><strong>1.Bad loss function design:<\/strong><\/p>\n\n<p>The design was borrowed from Openpose, which is the difference between \"distance map\" and the \"label map\".<\/p>\n\n<p>Unlike the openpose project, the \"distance map\" was obtained by computing Hamming distances to the \"Hint Hash\".\nThis process contain lots of unconventional operations and binarization, which may cause the model to be not trainable.<\/p>\n\n<p><strong>2.Wrong traing loop design:<\/strong><\/p>\n\n<p>During the training, I found every saved checkpoint files were named \"XXXX.ckpt.data-<strong>00000-of-00001<\/strong>\", the XXXX part will change, but the suffix \"00000-of-00001\"kept reappearing.<\/p>\n\n<p>I suspect there was something wrong in my model or training loop and caused it to repeat doing the <strong>first step of training<\/strong>?<\/p>\n\n<p>I'm still trying to find resources to fix this problem, any of your comment may help me a lot!<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport time\nfrom imageLoader import getPaddedROI,training_data_feeder\nimport math\nimport cv2\n\ntf.reset_default_graph()\n\ndef truncated_normal_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype,     initializer=tf.truncated_normal_initializer(stddev=0.01)))\ndef zero_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.constant_initializer(0.0)))\n\nroi_size = 23\nimage_input_size = 301\n\n#input placeholders\n#batch1 hints\ninputs_b1h1 = tf.placeholder(tf.float32, ( 16, roi_size, roi_size, 3), name='inputs_b1h1')\n#inputs_b1h2 = tf.placeholder(tf.float32, ( 16, roi_size, roi_size, 3), name='inputs_b1h2')\n\ninputs_s = tf.placeholder(tf.float32, (None, image_input_size, image_input_size, 3), name='inputs_s')\nlabels = tf.placeholder(tf.float32,(16,76,76), name='labels')\n\n#define the model\n\ndef paraNet(inputs, inputs_s , ground_truth_labels ):\n    with tf.variable_scope('conv'):\n        out_l1 = tf.layers.conv2d(inputs, 16, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_1')\n        out_l1r = tf.nn.relu(out_l1)\n        out_l2 = tf.layers.conv2d(out_l1r, 48, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_2')\n        out_l2r = tf.nn.relu(out_l2)\n        out_l3 = tf.layers.conv2d(out_l2r, 96, [5, 5],strides=(1, 1), padding ='valid' ,name='para_conv_3')\n        out_l3r = tf.nn.relu(out_l3)\n        out_l4 = tf.layers.conv2d(out_l3r, 32, [1, 1],strides=(1, 1), padding ='valid' ,name='para_conv_4')\n        hint = tf.squeeze(  tf.sign( tf.sigmoid(out_l4) ) )\n\n    with tf.variable_scope('conv', reuse=tf.AUTO_REUSE ):\n        out_2_l1 = tf.layers.conv2d(inputs_s,  16, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_1')\n        out_2_l1r = tf.nn.relu(out_2_l1)\n        out_2_l2 = tf.layers.conv2d(out_2_l1r, 48, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_2')\n        out_2_l2r = tf.nn.relu(out_2_l2)\n        out_2_l3 = tf.layers.conv2d(out_2_l2r, 96, [5, 5],strides=(1, 1), padding ='same' ,name='para_conv_3')\n        out_2_l3r = tf.nn.relu(out_2_l3)\n        out_2_l4 = tf.layers.conv2d(out_2_l3r, 32, [1, 1],strides=(1, 1), padding ='same' ,name='para_conv_4')\n        sample =tf.sign( tf.sigmoid(out_2_l4))\n\n    map0 = tf.reduce_sum ( tf.abs (tf.subtract( hint[0] , sample ) ) , axis=3 )  \n    map1 = tf.reduce_sum ( tf.abs (tf.subtract( hint[1] , sample ) ) , axis=3 )  \n    map2 = tf.reduce_sum ( tf.abs (tf.subtract( hint[2] , sample ) ) , axis=3 )  \n    map3 = tf.reduce_sum ( tf.abs (tf.subtract( hint[3] , sample ) ) , axis=3 )  \n    map4 = tf.reduce_sum ( tf.abs (tf.subtract( hint[4] , sample ) ) , axis=3 )  \n    map5 = tf.reduce_sum ( tf.abs (tf.subtract( hint[5] , sample ) ) , axis=3 )  \n    map6 = tf.reduce_sum ( tf.abs (tf.subtract( hint[6] , sample ) ) , axis=3 )  \n    map7 = tf.reduce_sum ( tf.abs (tf.subtract( hint[7] , sample ) ) , axis=3 )  \n    map8 = tf.reduce_sum ( tf.abs (tf.subtract( hint[8] , sample ) ) , axis=3 )  \n    map9 = tf.reduce_sum ( tf.abs (tf.subtract( hint[9] , sample ) ) , axis=3 )  \n    map10 = tf.reduce_sum ( tf.abs (tf.subtract( hint[10] , sample ) ) , axis=3 )  \n    map11 = tf.reduce_sum ( tf.abs (tf.subtract( hint[11] , sample ) ) , axis=3 )  \n    map12 = tf.reduce_sum ( tf.abs (tf.subtract( hint[12] , sample ) ) , axis=3 )  \n    map13 = tf.reduce_sum ( tf.abs (tf.subtract( hint[13] , sample ) ) , axis=3 )  \n    map14 = tf.reduce_sum ( tf.abs (tf.subtract( hint[14] , sample ) ) , axis=3 )  \n    map15 = tf.reduce_sum ( tf.abs (tf.subtract( hint[15] , sample ) ) , axis=3 )  \n\n    totoal_map =tf.div( tf.concat([map0, map1, map2, map3, map4, map5, map6, map7,\n                               map8, map9, map10,map11,map12, map13, map14, map15], 0) , 64)\n    loss = tf.nn.l2_loss( totoal_map -  ground_truth_labels , name = 'loss'  )\n\n    return loss, totoal_map\n\nloss, totoal_map = paraNet(inputs_b1h1, inputs_s, labels)\ntrain_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n\ninit =  tf.global_variables_initializer()\n\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess:    \n    #writer = tf.summary.FileWriter(\".\/variable_graph\",graph = sess.graph)\n    sess.run(init)\n\n    #load image from dataset(train set)\n    joint_data_path = \".\/custom_data.json\"\n    train_val_path = \".\/train_val_indices.json\"\n    imgpath = \".\/000\/\"\n    input_size = 301\n    hint_roi_size = 23\n\n    #saver.restore(sess, \".\/temp_model\/model5.ckpt\")\n\n\n    for i in range(5000):\n\n        #load data\n        hintSet01,hintSet02,t_img,t_label_norm = training_data_feeder(joint_data_path, train_val_path, imgpath, input_size, hint_roi_size )\n        #Normalize the image pixel values to 0~1\n        hintSet01_norm = []\n        hintSet02_norm = []\n\n        t_img =[ np.float32(t_img \/255.0) ]\n\n        for rois in hintSet01:\n            tmp = np.float32(rois \/ 255.0)\n            hintSet01_norm.append(tmp.tolist())\n        for rois in hintSet02:\n            tmp = np.float32(rois \/ 255.0)\n            hintSet02_norm.append(tmp.tolist())\n        loss_val, _ = sess.run([loss, train_step] , \n                      feed_dict = {inputs_s:  t_img, \n                                   inputs_b1h1: hintSet01_norm, \n                                   labels: t_label_norm })\n        if i % 50 == 0:\n            print(loss_val)\n\n    #save_path = saver.save(sess, \".\/temp_model\/model\" + '5' + \".ckpt\")\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/gitpharm01\/Parapose\/\" rel=\"nofollow noreferrer\">Here<\/a> is the github repo and the dataset links on github <a href=\"https:\/\/github.com\/gitpharm01\/Parapose\/blob\/master\/000\/readme.md\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<blockquote>\n  <p><strong>2018.11.9 Update:<\/strong><\/p>\n<\/blockquote>\n\n<p>I borrowed a binarization method from \"Binarized Neural Network\".<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><code>def ste_binarize( value ):\n\"\"\"\nClip and binarize tensor using the straight through estimator (STE) for the gradient.\nThe gradient of tf.sign(x) will always be zero thus it use:\ng.gradient_override_map({\"Sign\" : \"Identity\"}) \nThis will replace the gradient of tf.sign(x) with the original gradient of x(which is the Identity). \n\"\"\"\ng = tf.get_default_graph()\n\nwith ops.name_scope(\"Binarized\") as name:\n    with g.gradient_override_map({\"Sign\" : \"Identity\"}):\n        return tf.sign(value)\n<\/code><\/pre>\n\n<p>This method works well in the original project on <a href=\"https:\/\/github.com\/itayhubara\/BinaryNet.tf\" rel=\"nofollow noreferrer\">Github<\/a>\nI replaced the original tf.sign() with this method and tried training again.\nSadly, it didn't work on my case.\nI'll change the structure of my model and try to record the gradient value.\nIf I make any progress, I'll update this post. <\/p>\n\n<blockquote>\n  <p><strong>2018.11.22 Update:<\/strong>\n  I changed the model structure and trained it again.\n  This time I discovered a strange behavior of the loss value:\n  While training, the loss value was actually going up and freeze at a certain point.<\/p>\n<\/blockquote>\n\n<p>Here's the new model:<\/p>\n\n<pre><code>import tensorflow as tf\n\nfrom tensorflow.python.framework import ops\n\nfrom tensorflow.python.platform import gfile\n\n\nfrom progress.bar import Bar\n\nimport numpy as np\nimport time\nfrom datetime import datetime\nimport math\nimport matplotlib.pyplot as plt\nimport cv2\n\nfrom imageLoader import getPaddedROI,training_batch_generator\n\ntf.reset_default_graph()\n\n\n\n#load image from dataset(train set)\njoint_data_path = \".\/custom_data.json\"\ntrain_val_path = \".\/train_val_indices.json\"\nimgpath = \".\/000\/\"\ninput_size = 301\nhint_roi_size = 23\n\nroi_size = 23\nimage_input_size = 301\nbatch_number =10\n\ndef truncated_normal_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.truncated_normal_initializer(stddev=0.01)))\ndef zero_var(name,shape,dtype):\n    return(tf.get_variable(name=name, shape=shape, dtype=dtype, initializer=tf.constant_initializer(0.0)))\n\ndef ste_binarize( value ):\n\"\"\"\nClip and binarize tensor using the straight through estimator (STE) for the gradient.\nThe gradient of tf.sign(x) will always be zero so it use:\ng.gradient_override_map({\"Sign\" : \"Identity\"}) \nThis will replace the gradient of tf.sign(x) with the original gradient of x(which is the Identity). \n\"\"\"\n    g = tf.get_default_graph()\n\n    with ops.name_scope(\"Binarized\") as name:\n        with g.gradient_override_map({\"Sign\" : \"Identity\"}):\n            return tf.sign(value)\n\n#define the model\ndef paraNet(hint_inputs, sample_inputs):\n    def paraConv(inputs):\n        out_l1 = tf.layers.conv2d(inputs , 16, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_1')\n        out_l1r = tf.nn.relu(out_l1)\n        out_l2 = tf.layers.conv2d(out_l1r, 48, [3, 3],strides=(2, 2), padding ='valid' ,name='para_conv_2')\n        out_l2r = tf.nn.relu(out_l2)\n        out_l3 = tf.layers.conv2d(out_l2r, 96, [5, 5],strides=(1, 1), padding ='valid' ,name='para_conv_3')\n        out_l3r = tf.nn.relu(out_l3)\n        out_l4 = tf.layers.conv2d(out_l3r, 32, [1, 1],strides=(1, 1), padding ='valid' ,name='para_conv_4')\n        hint =tf.reshape( tf.squeeze(  ste_binarize( out_l4 ) ), [batch_number, 1 , 1 , 32] )\n        return hint\n\nwith tf.variable_scope('conv'):\n    hint00 = paraConv(hint_inputs[:,0,:,:,:])\nwith tf.variable_scope('conv', reuse= True ):\n    hint01 = paraConv(hint_inputs[:,1,:,:,:])\n    hint02 = paraConv(hint_inputs[:,2,:,:,:])\n    hint03 = paraConv(hint_inputs[:,3,:,:,:])\n    hint04 = paraConv(hint_inputs[:,4,:,:,:])\n    hint05 = paraConv(hint_inputs[:,5,:,:,:])\n    hint06 = paraConv(hint_inputs[:,6,:,:,:])\n    hint07 = paraConv(hint_inputs[:,7,:,:,:])\n    hint08 = paraConv(hint_inputs[:,8,:,:,:])\n    hint09 = paraConv(hint_inputs[:,9,:,:,:])\n    hint10 = paraConv(hint_inputs[:,10,:,:,:])\n    hint11 = paraConv(hint_inputs[:,11,:,:,:])\n    hint12 = paraConv(hint_inputs[:,12,:,:,:])\n    hint13 = paraConv(hint_inputs[:,13,:,:,:])\n    hint14 = paraConv(hint_inputs[:,14,:,:,:])\n    hint15 = paraConv(hint_inputs[:,15,:,:,:]) \n\n    out_2_l1 = tf.layers.conv2d(sample_inputs,  16, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_1')\n    out_2_l1r = tf.nn.relu(out_2_l1)\n    out_2_l2 = tf.layers.conv2d(out_2_l1r, 48, [3, 3],strides=(2, 2), padding ='same' ,name='para_conv_2')\n    out_2_l2r = tf.nn.relu(out_2_l2)\n    out_2_l3 = tf.layers.conv2d(out_2_l2r, 96, [5, 5],strides=(1, 1), padding ='same' ,name='para_conv_3')\n    out_2_l3r = tf.nn.relu(out_2_l3)\n    out_2_l4 = tf.layers.conv2d(out_2_l3r, 32, [1, 1],strides=(1, 1), padding ='same' ,name='para_conv_4')\n    #sample =tf.sign( tf.sigmoid(out_2_l4))\n    sample =ste_binarize( out_2_l4)\n\n    #originalMap = tf.reduce_sum ( tf.abs (tf.subtract( hint00 , sample ) ) , axis=3 )\n\n    map0 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint00 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )\n    map1 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint01 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map2 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint02 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map3 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint03 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map4 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint04 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map5 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint05 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map6 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint06 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map7 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint07 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map8 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint08 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map9 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint09 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map10 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint10 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map11 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint11 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map12 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint12 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map13 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint13 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map14 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint14 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n    map15 = tf.reshape( tf.reduce_sum ( tf.abs (tf.subtract( hint15 , sample ) ) , axis=3 ), [batch_number, 1, 76, 76]  )  \n\n    totoal_map =tf.div( tf.concat([map0, map1, map2, map3, map4, map5, map6, map7,\n                               map8, map9, map10,map11,map12, map13, map14, map15], 1) , 32)\nreturn totoal_map\n\n\n\n\ninputs_b1h1 = tf.placeholder(tf.float32, ( None, 16, roi_size, roi_size, 3), name='inputs_b1h1')\ninputs_s = tf.placeholder(tf.float32, (None, image_input_size, image_input_size, 3), name='inputs_s')\nground_truth_labels = tf.placeholder(tf.float32,(None, 16,76,76), name='labels')\n\nmtotoal_map = paraNet(inputs_b1h1 , inputs_s )\nmloss = tf.nn.l2_loss( ground_truth_labels -  mtotoal_map, name = 'loss'  )\ntrain_step = tf.train.GradientDescentOptimizer(0.0001).minimize(mloss)\n\ninit =  tf.global_variables_initializer()\n\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess:\n    #writer = tf.summary.FileWriter(\".\/variable_graph\",graph = sess.graph)\n    sess.run(init)\n    #saver.restore(sess, \".\/temp_model\/model7.ckpt\")\n\n    loss_per_iteration = []\n    for i in range(100):\n\n        #load data\n        hintSet01,hintSet02,t_img,t_label_norm = training_batch_generator(joint_data_path, train_val_path, imgpath, input_size, hint_roi_size,batch_number)\n\n        loss_val, _ = sess.run([mloss, train_step] , \n                      feed_dict = {inputs_s:  t_img, \n                                   inputs_b1h1: hintSet01, \n                                   ground_truth_labels: t_label_norm })\n        loss_per_iteration.append(loss_val)\n        if i % 5 == 0:\n            print(loss_val)\n\n    plt.plot(loss_per_iteration)\n    plt.show()\n    save_path = saver.save(sess, \".\/temp_model\/model\" + '7' + \".ckpt\")\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53141762,"AnswerCount":1,"Tags":"<python><tensorflow><assign><tensor>","CreationDate":"2018-11-04T14:15:42.327","AcceptedAnswerId":53143985.0,"OwnerUserId":1973451.0,"Title":"tensorflow: how to assign an updated numpy","Body":"<p>I'm new in tensorflow and I'm trying to understand its behaviors; I'm trying to define all the operations outside the session scope so to optimize the computation time.\nIn the following code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np  \nZ_tensor = tf.Variable(np.float32( np.zeros((1, 10)) ), name=\"Z_tensor\")\nZ_np = np.zeros((1,10))\nupdate_Z = tf.assign(Z_tensor, Z_np)\n\nZ_np[0][2:4] = 4\n\nwith tf.Session() as sess:\n    sess.run(Z_tensor.initializer)\n    print(Z_tensor.eval())\n    print(update_Z.eval(session=sess))\n<\/code><\/pre>\n\n<p>I obtain as ouput:<\/p>\n\n<pre><code>[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]\n[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]\n<\/code><\/pre>\n\n<p>Instead I expected as output:<\/p>\n\n<pre><code>[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]\n[[0. 0. 4. 4. 0. 0. 0. 0. 0. 0.]]\n<\/code><\/pre>\n\n<p>It seems that the <code>Z_np<\/code> array is not updated in the assign operation and I don't understand why.\nDoesn't the operation<\/p>\n\n<pre><code>  update_Z = tf.assign(Z_tensor, Z_np)\n<\/code><\/pre>\n\n<p>make a link with <code>Z_np<\/code>?<\/p>\n","answers":[{"AnswerId":"53143985","CreationDate":"2018-11-04T18:24:12.397","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>When you use tf.assign, it expects a tensor as the second argument.  Because you provided a Numpy array, it automatically promotes it to a CONSTANT tensor and places it in the graph at that moment.  Because of this, no changes you make to the Numpy array will have any effect on the TensorFlow graph.  In order to get the desired functionality, you should use a placeholder:<\/p>\n\n<pre><code>Z_placeholder = tf.placeholder(tf.float32, Z_np.shape)\n\nwith tf.Session() as sess:\n    sess.run(Z_tensor.initializer)\n    print(Z_tensor.eval(feed_dict={Z_placeholder: Z_np}, session=sess))\n    Z_np[0][2:4] = 4\n    print(Z_tensor.eval(feed_dict={Z_placeholder: Z_np}, session=sess))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53141930,"AnswerCount":1,"Tags":"<tensorflow><segment>","CreationDate":"2018-11-04T14:37:29.900","AcceptedAnswerId":null,"OwnerUserId":5991532.0,"Title":"tensorflow how do one get the output the same size as input tensor after segment sum","Body":"<p>I'm using the tf.unsorted_segment_sum method of TensorFlow and it works.\nFor example:<\/p>\n\n<pre><code>tf.unsorted_segment_sum(tf.constant([0.2, 0.1, 0.5, 0.7, 0.8]),\n                        tf.constant([0, 0, 1, 2, 2]), 3)\n<\/code><\/pre>\n\n<p>Gives the right result:<\/p>\n\n<pre><code>array([ 0.3,  0.5 , 1.5 ], dtype=float32)\n<\/code><\/pre>\n\n<p>I want to get:<\/p>\n\n<pre><code>array([0.3, 0.3, 0.5, 1.5, 1.5], dtype=float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53142316","CreationDate":"2018-11-04T15:22:51.157","ParentId":null,"OwnerUserId":"5991532","Title":null,"Body":"<p>I've solved it.<\/p>\n\n<pre><code>data = tf.constant([0.2, 0.1, 0.5, 0.7, 0.8])\ngr_idx = tf.constant([0, 0, 1, 2, 2])\ny, idx, count = tf.unique_with_count(gr_idx)\ngroup_sum = tf.segment_sum(data, gr_idx)\ngroup_sup = tf.gather(group_sum, idx)\n<\/code><\/pre>\n\n<p>answer:<\/p>\n\n<pre><code>array([0.3, 0.3, 0.5, 1.5, 1.5], dtype=float32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53142006,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-04T14:46:08.773","AcceptedAnswerId":null,"OwnerUserId":10550693.0,"Title":"Unable to import tensor flow and keras properly","Body":"<p>I have wrote this code in python here. I have a dataset in json and i am trying to train that dataset using keras. I was able to load data perfectly but after writing the complete code to train my dataset and running it, it started giving me errors in imports of my program. I tried installing tensorflowjs and keras using pip again but everytime i try tensorflowjs it start giving me compatibilty error on kerasApplications version. I tried installing every version but it is still giving me the same error. How can i get out of this trouble?<\/p>\n\n<pre><code>import json\nimport numpy as np\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nimport tensorflowjs as tfjs\nwith open(\"C:\\\\Users\\\\TechProBox\\\\Desktop\\\\Model.json\") as f:\ndata = json.load(f)\nx1 = np.array(data['attiude.roll'])\ny1 = np.array(data['attitude.pitch'])\nz1 = np.array(data['attitude.yaw'])\nx2 = np.array(data['gravity.x'])\ny2 = np.array(data['gravity.y'])\nz2 = np.array(data['gravity.z'])\nx3 = np.array(data['rotationRate.x'])\ny3 = np.array(data['rotationRate.y'])\nz3 = np.array(data['rotationRate.z'])\nx4 = np.array(data['userAcceleration.x'])\ny4 = np.array(data['userAcceleration.y'])\nz4 = np.array(data['userAcceleration.z'])\n\nx1_train = x1[:-10000]\ny1_train = y1[:-10000]\nz1_train = z1[:-10000]\nx2_train = x2[:-10000]\ny2_train = y2[:-10000]\nz2_train = z2[:-10000]\nx3_train = x3[:-10000]\ny3_train = y3[:-10000]\nz3_train = z3[:-10000]\nx4_train = x4[:-10000]\ny4_train = y4[:-10000]\nz4_train = z4[:-10000]\nx1_test = x1[:-10000]\ny1_test = y1[:-10000]\nz1_test = z1[:-10000]\nx2_test = x2[:-10000]\ny2_test = y2[:-10000]\nz2_test = z2[:-10000]\nx3_test = x3[:-10000]\ny3_test = y3[:-10000]\nz3_test = z3[:-10000]\nx4_test = x4[:-10000]\ny4_test = y4[:-10000]\nz4_test = z4[:-10000]\n\n\nmodel = Sequential()\nmodel.add(Dense(64, activation='relu', input_dim=6))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(3, activation='softmax'))\n\nadam = keras.optimizers.Adam(lr=0.0001)\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=adam,\n              metrics=['accuracy'])\n\nmodel.fit(x1_train, y1_train, z1_train, x2_train, y2_train, z2_train,x3_train, y3_train, z3_train,\n            x4_train, y4_train, z4_train,\n            epochs=14,\n            batch_size=128)\n\nscore = model.evaluate(x1_test, y2_test, z3_test, x2_test, y2_test, z2_test, x3_test, y3_test, z3_test,\n                        x4_test, y4_test, z4_test, batch_size=128)\n\nprint(score)\nmodel.save(\"Keras-64*2-10epoch\")\ntfjs.converters.save_keras_model(model,\"tfjsv3\")\n<\/code><\/pre>\n\n<p>Here are the errors<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\\Users\\TechProBox\\Desktop\\Python1.py\", line 3, in &lt;module&gt;\n    import keras\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\backend\\__init__.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 52, in &lt;module&gt;\n    from tensorflow.core.framework.graph_pb2 import *\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in &lt;module&gt;\n    from google.protobuf.pyext import _message\n\nImportError: DLL load failed: The specified procedure could not be found.\n<\/code><\/pre>\n","answers":[{"AnswerId":"53142513","CreationDate":"2018-11-04T15:46:45.843","ParentId":null,"OwnerUserId":"10550693","Title":null,"Body":"<p>Just needed to isntall protobuff and it worked<\/p>\n\n<pre><code>pip install protobuf==3.6.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":53142182,"AnswerCount":0,"Tags":"<input><keras><lstm><rnn>","CreationDate":"2018-11-04T15:06:01.793","AcceptedAnswerId":null,"OwnerUserId":10603961.0,"Title":"Keras RNN input timesteps order","Body":"<p>I have a question regarding the LSTM input shap in keras. \nThe input shape should be a 3D tensor where the dimensions are:<br> <\/p>\n\n<blockquote>\n  <p>batch size, time step, feutures <\/p>\n<\/blockquote>\n\n<p>I want an univariate time series model. The only dimension where the order is important is time step. Lets say my time series looks the following:<\/p>\n\n<pre><code>Data = [[1, 10],[2, 9],[3, 8],[4, 7],[5, 6],[6, 5]]\n<\/code><\/pre>\n\n<p>with 2 input features and lets say <code>timestep = 3<\/code>.  Reshaped is my reshaped data:<\/p>\n\n<ul>\n<li><p><code>Reshaped(1, :, :) = [[1, 10],[2,9],[3,8]]<\/code><\/p><\/li>\n<li><p><code>Reshaped(1, :, :) = [[3, 8],[2,9],[1,10]]<\/code><\/p><\/li>\n<\/ul>\n\n<p><strong>Which of them is correct? The first or the second approach?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":53142833,"AnswerCount":0,"Tags":"<python><arrays><json><tensorflow><keras>","CreationDate":"2018-11-04T16:22:32.240","AcceptedAnswerId":null,"OwnerUserId":10550693.0,"Title":"Get data from list containg objects","Body":"<p>I have a json format list.There are more than 3000 records in json. I am trying to train my model using keras and tensorflowjs in python but i need to iterate through this list data and take objects data one by one. I need to know that how i can iterate through this data using python so that my model will start reading the values. I just need to iterate through the numpy array in specific variables. so that the data within each variable could be unique with every iteration. I am not well experiences with python so please if you can edit my code then that would be of great help<\/p>\n\n<pre><code>[\n  {\n    \"\": 0,\n    \"attitude.roll\": 1.528132,\n    \"attitude.pitch\": -0.733896,\n    \"attitude.yaw\": 0.696372,\n    \"gravity.x\": 0.741895,\n    \"gravity.y\": 0.669768,\n    \"gravity.z\": -0.031672000000000006,\n    \"rotationRate.x\": 0.316738,\n    \"rotationRate.y\": 0.77818,\n    \"rotationRate.z\": 1.0827639999999998,\n    \"userAcceleration.x\": 0.294894,\n    \"userAcceleration.y\": -0.18449300000000002,\n    \"userAcceleration.z\": 0.377542\n  },\n  {\n    \"\": 1,\n    \"attitude.roll\": 1.527992,\n    \"attitude.pitch\": -0.716987,\n    \"attitude.yaw\": 0.677762,\n    \"gravity.x\": 0.753099,\n    \"gravity.y\": 0.657116,\n    \"gravity.z\": -0.032255,\n    \"rotationRate.x\": 0.842032,\n    \"rotationRate.y\": 0.424446,\n    \"rotationRate.z\": 0.643574,\n    \"userAcceleration.x\": 0.21940500000000002,\n    \"userAcceleration.y\": 0.035845999999999996,\n    \"userAcceleration.z\": 0.11486600000000001\n  }\n]\n<\/code><\/pre>\n\n<p>Here is my Python code<\/p>\n\n<pre><code>import json\nimport numpy as np\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nimport tensorflowjs as tfjs\nwith open(\"C:\\\\Users\\\\TechProBox\\\\Desktop\\\\Model.json\") as f:\ndata = json.load(f)\nx1 = np.array(data['attiude.roll'])\ny1 = np.array(data['attitude.pitch'])\nz1 = np.array(data['attitude.yaw'])\nx2 = np.array(data['gravity.x'])\ny2 = np.array(data['gravity.y'])\nz2 = np.array(data['gravity.z'])\nx3 = np.array(data['rotationRate.x'])\ny3 = np.array(data['rotationRate.y'])\nz3 = np.array(data['rotationRate.z'])\nx4 = np.array(data['userAcceleration.x'])\ny4 = np.array(data['userAcceleration.y'])\nz4 = np.array(data['userAcceleration.z'])\n\nx1_train = x1[:-10000]\ny1_train = y1[:-10000]\nz1_train = z1[:-10000]\nx2_train = x2[:-10000]\ny2_train = y2[:-10000]\nz2_train = z2[:-10000]\nx3_train = x3[:-10000]\ny3_train = y3[:-10000]\nz3_train = z3[:-10000]\nx4_train = x4[:-10000]\ny4_train = y4[:-10000]\nz4_train = z4[:-10000]\nx1_test = x1[:-10000]\ny1_test = y1[:-10000]\nz1_test = z1[:-10000]\nx2_test = x2[:-10000]\ny2_test = y2[:-10000]\nz2_test = z2[:-10000]\nx3_test = x3[:-10000]\ny3_test = y3[:-10000]\nz3_test = z3[:-10000]\nx4_test = x4[:-10000]\ny4_test = y4[:-10000]\nz4_test = z4[:-10000]\n\n\nmodel = Sequential()\nmodel.add(Dense(64, activation='relu', input_dim=6))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(64, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(3, activation='softmax'))\n\nadam = keras.optimizers.Adam(lr=0.0001)\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=adam,\n              metrics=['accuracy'])\n\nmodel.fit(x1_train, y1_train, z1_train, x2_train, y2_train, z2_train,x3_train, y3_train, z3_train,\n            x4_train, y4_train, z4_train,\n            epochs=14,\n            batch_size=128)\n\nscore = model.evaluate(x1_test, y2_test, z3_test, x2_test, y2_test, z2_test, x3_test, y3_test, z3_test,\n                        x4_test, y4_test, z4_test, batch_size=128)\n\nprint(score)\nmodel.save(\"Keras-64*2-10epoch\")\ntfjs.converters.save_keras_model(model,\"tfjsv3\")\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53142940,"AnswerCount":1,"Tags":"<r><keras>","CreationDate":"2018-11-04T16:35:19.257","AcceptedAnswerId":53143052.0,"OwnerUserId":10604232.0,"Title":"Installation of Python not found, Python bindings not loaded","Body":"<p>I am new to R.\nI am trying to do one hot coding  and am using <code>to_categorical<\/code> function and I am getting the error: <\/p>\n\n<pre><code>\"Installation of Python not found, Python bindings not loaded\"\n<\/code><\/pre>\n\n<p>I have already installed keras package and have loaded it as well. Even tried removing it and installing again.<\/p>\n\n<p><em>This is my code:<\/em><\/p>\n\n<pre><code>trainy &lt;- c(0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2)\n\ntrainLabels &lt;- to_categorical(trainy)\n<\/code><\/pre>\n\n<p>I also tried the following code and got the same error: <\/p>\n\n<pre><code>model &lt;- keras_model_sequential()\n<\/code><\/pre>\n\n<p>Please let me know if you require any other details!<br>\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"53143052","CreationDate":"2018-11-04T16:46:35.893","ParentId":null,"OwnerUserId":"5224236","Title":null,"Body":"<p>You probably should either install Anaconda or add it to the path.\nIf the following:\ngo to start menu \/ type <code>cmd<\/code> \/ open the command line \/ type <code>python --version<\/code>\nDoesn't give you something like:<\/p>\n\n<pre><code>Python 3.6.4 :: Anaconda, Inc.\n<\/code><\/pre>\n\n<p>Then you need to install Conda: go <a href=\"https:\/\/www.anaconda.com\/download\/\" rel=\"nofollow noreferrer\">here<\/a> and add it to your path after install by following <a href=\"https:\/\/stackoverflow.com\/a\/4855685\/5224236\">this answer<\/a> for detailed instructions<\/p>\n"}]}
+{"QuestionId":53143118,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><vgg-net>","CreationDate":"2018-11-04T16:55:11.497","AcceptedAnswerId":53558743.0,"OwnerUserId":8809044.0,"Title":"Feature extraction with VGGnet from my own data with TensorFlow?","Body":"<p>I pre-trained VGG19 with SIRI-WHU dataset, and now I want to extract the features, and I don't know how to do it.\nCan someone help me please?\nthanks<\/p>\n","answers":[{"AnswerId":"53146381","CreationDate":"2018-11-04T23:07:13.490","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>Well there is not much information to help you. Can you load the model, if so you can do something like this:<\/p>\n\n<pre><code>with tf.Session() as sess:    \n    # the tensor you want to feed your image to\n    input_tensor = sess.graph.get_tensor_by_name(\"name of your input tensor\") \n\n    # the tensor you're interested in, most likely last_dense_layer_name\/BiasAdd:0\n    output_tensor = sess.graph.get_tensor_by_name(\"name of your output tensor\")\n\n    feature_vector = sess.run(output_tensor, feed_dict={input_tensor: **insert numpy array of your image **})\n<\/code><\/pre>\n\n<p>This code assumes that your graph is in memory, if you've problems doing that, just ask<\/p>\n"},{"AnswerId":"53558743","CreationDate":"2018-11-30T13:42:56.770","ParentId":null,"OwnerUserId":"8809044","Title":null,"Body":"<p>I solved this issue by fixing the FC6 layer as an output layer<\/p>\n\n<pre><code>prob = sess.run(vgg.fc6, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>then I stored the features in h5 file<\/p>\n\n<pre><code>f = h5py.File('sample.h5','a')\nf.create_dataset('data',data=prob,dtype=np.float32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53144038,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><reinforcement-learning>","CreationDate":"2018-11-04T18:29:32.263","AcceptedAnswerId":null,"OwnerUserId":10210261.0,"Title":"What does DeepMind's Sonnet afford that Keras doesn't?","Body":"<p>I'm really confused about the purpose of DeepMind's Sonnet library for TensorFlow. As far as I can tell from the documentation, it seems to do essentially what Keras does (flexible functional abstractions). Can someone tell me what the advantage of Sonnet is?<\/p>\n","answers":[{"AnswerId":"53239472","CreationDate":"2018-11-10T13:32:37.983","ParentId":null,"OwnerUserId":"9184597","Title":null,"Body":"<p>There isn't much difference between them. They are both:<\/p>\n\n<ul>\n<li>High-level object oriented libraries that bring about abstraction when developing neural networks (NN) or other machine learning (ML) algorithms.<\/li>\n<li>Built on top of TensorFlow (with the addition of Theano for Keras).<\/li>\n<\/ul>\n\n<p>So why did they make Sonnet? It appears that Keras doesn't seem to suit the needs of DeepMind. So DeepMind came up with Sonnet, a high-level object oriented programming library built on top of TensorFlow to address its research needs.<\/p>\n\n<p>Keras and Sonnet are both trying to simplify deep reinforcement learning, with the major difference being Sonnet is specifically adapted to the problems that DeepMind explores.<\/p>\n\n<p>The main advantage of Sonnet, from my perspective, is you can use it to reproduce the research demonstrated in DeepMind's papers with greater ease than keras, since DeepMind will be using Sonnet themselves. Aside from that advantage, it's just yet another framework with which to explore deep RL problems.<\/p>\n"}]}
+{"QuestionId":53144166,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-04T18:45:35.083","AcceptedAnswerId":53152966.0,"OwnerUserId":7326400.0,"Title":"How do I assign a 2D subset of my tensor with another 2D tensor in Keras?","Body":"<p>If I have two 3-D tensors <code>img<\/code> and <code>gen<\/code>. How do I assign a 2D subset of <code>img<\/code> with a 2D subset of <code>gen<\/code>? The following doesn't work as tensorflow doesn't allow direct assignment of tensors.<\/p>\n\n<pre><code>img[96:160 , 144:240 , :] = gen[96:160 , 144:240 , :]\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>This is the surrounding code. So I'm using a custom keras layer. This layer must receive the input image <code>img<\/code> and the generated image <code>x<\/code>. It must replace a portion of <code>img<\/code> with <code>x<\/code> and must return the modified <code>img<\/code>.<\/p>\n\n<pre><code>def patcher(tensors):\n    img = tensor[1]\n    gen = tensor[0]\n    #This is where the slicing must happen\n    img[96:160 , 144:240 , :] = gen[96:160 , 144:240 , :]\n    return [img]\n\nimg = Input( .. )\nx = Conv( .. )(img)\nout = Lambda(patcher,lambda a : [a[1]] )([x , img])\nmodel = Model(img, out)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53156335","CreationDate":"2018-11-05T14:27:12.523","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>I modified my initial solution, which only worked when the batch size was set, based on @jdehesa's solution. This should work in Keras across all the backends (TensorFlow, Theano and CNTK):<\/p>\n\n<pre><code>from keras import backend as K\nimport numpy as np\n\ndef replace_slices(ts, row_start, row_end, col_start, col_end):\n    shape = K.int_shape(ts[0])[1:-1]\n    np_mask = np.zeros(shape + (1,))\n    np_mask[row_start:row_end, col_start:col_end] = 1.\n    mask = K.variable(np_mask, dtype=K.dtype(ts[0]))\n    # ts[0] is the img and ts[1] is the x tensor\n    return ts[0] * (1 - mask) + ts[1] * mask\n\nargs = {'row_start': 96, 'row_end': 160, 'col_start': 144, 'col_end': 240}\n\nimg = Input(shape=(256,384,3))\nx = Conv2D(3, (3,3), padding='same')(img) # this must have 3 filters since img has 3 channels\nout = Lambda(replace_slices, arguments=args)([img, x])\nmodel = Model(img, out)\n<\/code><\/pre>\n"},{"AnswerId":"53152966","CreationDate":"2018-11-05T10:54:33.697","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Currently, you cannot replace slices of a tensor in a straightforward manner. I actually <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18383\" rel=\"nofollow noreferrer\">opened an issue about it<\/a> because it is something that people keep asking for. With the current API, you have to sort of figure out the best way to build the tensor you want. In this case, assuming <code>img<\/code> and <code>gen<\/code> have both the same shape, this is one way you could do that:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Input\nimg = tf.placeholder(tf.float32, [None, None, None])\ngen = tf.placeholder(tf.float32, [None, None, None])\nrow_start = tf.placeholder(tf.int32, [])\nrow_end = tf.placeholder(tf.int32, [])\ncol_start = tf.placeholder(tf.int32, [])\ncol_end = tf.placeholder(tf.int32, [])\n# Masks rows and columns to be replaced\nshape = tf.shape(img)\nrows = shape[0]\ncols = shape[1]\nchannels = shape[2]\ni = tf.range(rows)\nrow_mask = (row_start &lt;= i) &amp; (i &lt; row_end)\nj = tf.range(cols)\ncol_mask = (col_start &lt;= j) &amp; (j &lt; col_end)\n# Full mask of replaced elements\nmask = row_mask[:, tf.newaxis] &amp; col_mask\n# Select elements from flattened arrays\nimg_flat = tf.reshape(img, [-1, channels])\ngen_flat = tf.reshape(gen, [-1, channels])\nmask_flat = tf.reshape(mask, [-1])\nresult_flat = tf.where(mask_flat, gen_flat, img_flat)\n# Reshape back\nresult = tf.reshape(result_flat, shape)\n<\/code><\/pre>\n\n<p>Here is a small test:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    # img is positive and gen is negative\n    img_val = np.arange(60).reshape((4, 5, 3))\n    gen_val = -img_val\n    # Do img[2:4, 0:3, :] = gen[2:4, 0:3, :]\n    result_val = sess.run(result, feed_dict={\n        img: img_val,\n        gen: gen_val,\n        row_start: 2,\n        row_end: 4,\n        col_start: 0,\n        col_end: 3,\n    })\n    # Print one channel only for clarity\n    print(result_val[:, :, 0])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[  0.   3.   6.   9.  12.]\n [ 15.  18.  21.  24.  27.]\n [-30. -33. -36.  39.  42.]\n [-45. -48. -51.  54.  57.]]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>Here is a possible implementation for the code that you have posted. I am using a slightly different method here based on multiplication, which I think is better when you have many images.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef replace_slices(img, gen, row_start, row_end, col_start, col_end):\n    # Masks rows and columns to be replaced\n    shape = tf.shape(img)\n    rows = shape[1]\n    cols = shape[2]\n    i = tf.range(rows)\n    row_mask = (row_start &lt;= i) &amp; (i &lt; row_end)\n    j = tf.range(cols)\n    col_mask = (col_start &lt;= j) &amp; (j &lt; col_end)\n    # Full mask of replaced elements\n    mask = row_mask[:, tf.newaxis] &amp; col_mask\n    # Add channel dimension to mask and cast\n    mask = tf.cast(mask[:, :, tf.newaxis], img.dtype)\n    # Compute result\n    result = img * (1 - mask) + gen * mask\n    return result\n\ndef patcher(tensors):\n    img = tensor[1]\n    gen = tensor[0]\n    img = replace_slices(img, gen, 96, 160, 144, 240)\n    return [img]\n\nimg = Input( .. )\nx = Conv( .. )(img)\nout = Lambda(patcher, ambda a: [a[1]])([x , img])\nmodel = Model(img, out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53144385,"AnswerCount":0,"Tags":"<python><linux><tensorflow><tensorboard><libpng>","CreationDate":"2018-11-04T19:08:59.403","AcceptedAnswerId":null,"OwnerUserId":9548638.0,"Title":"tf.summary.image provoks PNG encoding failed","Body":"<p>I am working on MNIST dataset and having trouble with tensorboard. I want to record an image with the following function : <\/p>\n\n<pre><code>tf.summary.image('input', input_layer, 1)\n<\/code><\/pre>\n\n<p>But I get the following warning and error : <\/p>\n\n<blockquote>\n  <p>2018-11-04 19:32:44.798869: W tensorflow\/core\/lib\/png\/png_io.cc:87] PNG warning: Application built with libpng-1.2.50 but running with 1.6.34\n  2018-11-04 19:32:44.838003: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at summary_image_op.cc:131 : Internal: PNG encoding failed\n  2018-11-04 19:32:44.798869: W tensorflow\/core\/lib\/png\/png_io.cc:87] PNG warning: Application built with libpng-1.2.50 but running with 1.6.34\n  2018-11-04 19:32:44.838003: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at summary_image_op.cc:131 : Internal: PNG encoding failed<\/p>\n<\/blockquote>\n\n<p>I read about the warning, it is possible that it is not a problem from what I read. \nWhat I cannot find out or understand is which is the \"application\" the warning is talking about, tensorflow ? tensorboard ? \nOn of the advice I get from internet is to have the same version in my system than the one of the application. I want to update the one of the application.<\/p>\n\n<p>Another point, to be sure that the input to tf.summary.image is fine, the following command : <\/p>\n\n<pre><code>input_layer = tf.reshape(features, [-1, 28, 28, 1])\n<\/code><\/pre>\n\n<p><code>print(input_layer)<\/code><\/p>\n\n<p>Gives the following output :<\/p>\n\n<blockquote>\n  <p>Tensor(\"Reshape:0\", shape=(?, 28, 28, 1), dtype=float32)<\/p>\n<\/blockquote>\n\n<p>Which seems to correspond to the input requirement for the image function.<\/p>\n\n<p>Could you please tell me if : \n- the issue is known\n- or how to investigate the problem to be more independant next time (I am a bit lost about version investigation, what is installed or not, if I should check the version in python or in Linux).<\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[]}
+{"QuestionId":53144606,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><conv-neural-network><transfer-learning>","CreationDate":"2018-11-04T19:32:37.600","AcceptedAnswerId":53144652.0,"OwnerUserId":9745931.0,"Title":"ValueError, Error when checking target: expected dense_1 to have 4 dimensions","Body":"<p>I'm trying to fine-tune a MobileNet but I'm receiving the following error:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>ValueError, Error when checking target: expected dense_1 to have 4 dimensions, but got array with shape (10, 14)\n<\/code><\/pre>\n\n<p>Is there any conflict with how I set my directory iterators?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_batches = ImageDataGenerator(preprocessing_function=keras.applications.mobilenet_v2.preprocess_input).flow_from_directory(\ntrain_path, target_size=(224, 224), batch_size=10)\nvalid_batches = ImageDataGenerator(preprocessing_function=keras.applications.mobilenet_v2.preprocess_input).flow_from_directory(\nvalid_path, target_size=(224, 224), batch_size=10)\n\ntest_batches = ImageDataGenerator(preprocessing_function=keras.applications.mobilenet_v2.preprocess_input).flow_from_directory(\ntest_path, target_size=(224, 224), batch_size=10, shuffle=False)\n<\/code><\/pre>\n\n<p>and my new bottleneck layer is as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x=mobile.layers[-6].output\npredictions = Dense(14, activation='softmax')(x)\n\nmodel = Model(inputs=mobile.input, outputs=predictions)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53144652","CreationDate":"2018-11-04T19:37:01.277","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since the <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">Dense layer is applied on the last axis of its input<\/a> and considering the fact that <code>x<\/code> is a 4D tensor, the <code>predictions<\/code> tensor would also be a 4D tensor. That's why the model expects a 4D output (i.e. <code>expected dense_1 to have 4 dimensions<\/code>) but your labels are 2D (i.e. <code>but got array with shape (10, 14)<\/code>). To resolve this, you need to make <code>x<\/code> as a 2D tensor. One way of doing it is to use a <code>Flatten<\/code> layer:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>x = mobile.layers[-6].output\nx = Flatten()(x)\npredictions = Dense(14, activation='softmax')(x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53144762,"AnswerCount":1,"Tags":"<python><keras><lstm><lasagne>","CreationDate":"2018-11-04T19:48:57.293","AcceptedAnswerId":53196692.0,"OwnerUserId":6145828.0,"Title":"convert Lasagne to Keras code (CNN -> LSTM)","Body":"<p>I would like to convert this Lasagne code:<\/p>\n\n<pre><code>et = {}\nnet['input'] = lasagne.layers.InputLayer((100, 1, 24, 113))\nnet['conv1\/5x1'] = lasagne.layers.Conv2DLayer(net['input'], 64, (5, 1))\nnet['shuff'] = lasagne.layers.DimshuffleLayer(net['conv1\/5x1'], (0, 2, 1, 3))\nnet['lstm1'] = lasagne.layers.LSTMLayer(net['shuff'], 128)\n<\/code><\/pre>\n\n<p>in Keras code. Currently I came up with this:<\/p>\n\n<pre><code>multi_input = Input(shape=(1, 24, 113), name='multi_input')\ny = Conv2D(64, (5, 1), activation='relu', data_format='channels_first')(multi_input)\ny = LSTM(128)(y)\n<\/code><\/pre>\n\n<p>But I get the error: <code>Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=4<\/code><\/p>\n","answers":[{"AnswerId":"53196692","CreationDate":"2018-11-07T19:44:38.613","ParentId":null,"OwnerUserId":"2995038","Title":null,"Body":"<h1>Solution<\/h1>\n\n<pre><code>from keras.layers import Input, Conv2D, LSTM, Permute, Reshape\n\nmulti_input = Input(shape=(1, 24, 113), name='multi_input')\nprint(multi_input.shape)  # (?, 1, 24, 113)\n\ny = Conv2D(64, (5, 1), activation='relu', data_format='channels_first')(multi_input)\nprint(y.shape)  # (?, 64, 20, 113)\n\ny = Permute((2, 1, 3))(y)\nprint(y.shape)  # (?, 20, 64, 113)\n\n# This line is what you missed\n# ==================================================================\ny = Reshape((int(y.shape[1]), int(y.shape[2]) * int(y.shape[3])))(y)\n# ==================================================================\nprint(y.shape)  # (?, 20, 7232)\n\ny = LSTM(128)(y)\nprint(y.shape)  # (?, 128)\n<\/code><\/pre>\n\n<h1>Explanations<\/h1>\n\n<p>I put the documents of Lasagne and Keras here so you can do cross-referencing:<\/p>\n\n<p><a href=\"https:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/recurrent.html\" rel=\"nofollow noreferrer\"><strong>Lasagne<\/strong><\/a><\/p>\n\n<blockquote>\n  <p>Recurrent layers can be used similarly to feed-forward layers except\n  that the input shape is expected to be <code>(batch_size, sequence_length, num_inputs)<\/code><\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\"><strong>Keras<\/strong><\/a><\/p>\n\n<blockquote>\n  <p><strong>Input shape<\/strong><\/p>\n  \n  <p>3D tensor with shape <code>(batch_size, timesteps, input_dim)<\/code>.<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Basically the API is the same, but Lasagne probably does reshape for you (I need to check the source code later). That's why you got this error:<\/p>\n\n<pre><code>Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=4\n<\/code><\/pre>\n\n<p>, since the tensor shape after <code>Conv2D<\/code> is <code>(?, 64, 20, 113)<\/code> of <code>ndim=4<\/code><\/p>\n\n<p>Therefore, the solution is to reshape it to <code>(?, 20, 7232)<\/code>.<\/p>\n\n<h2>Edit<\/h2>\n\n<p>Confirmed with the Lasagne <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/blob\/master\/lasagne\/layers\/recurrent.py#L863\" rel=\"nofollow noreferrer\">source code<\/a>, it does the trick for you:<\/p>\n\n<pre><code>num_inputs = np.prod(input_shape[2:])\n<\/code><\/pre>\n\n<p>So the correct tensor shape as input for LSTM is <code>(?, 20, 64 * 113)<\/code> = <code>(?, 20, 7232)<\/code><\/p>\n\n<hr>\n\n<h2>Note<\/h2>\n\n<p><code>Permute<\/code> is redundant here in Keras since you have to reshape anyway. The reason why I put it here is to have a \"full translation\" from Lasagne to Keras, and it does what <code>DimshuffleLaye<\/code> does in Lasagne. <\/p>\n\n<p><code>DimshuffleLaye<\/code> is however needed in Lasagne because of the reason I mentioned in <strong>Edit<\/strong>, the new dimension created by Lasagne LSTM is from the multiplication of \"the last two\" dimensions.<\/p>\n"}]}
+{"QuestionId":53144832,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-04T19:55:25.197","AcceptedAnswerId":null,"OwnerUserId":10520111.0,"Title":"Tensorflow CNN 'tuple' object has no attribute 'initializer'","Body":"<p>Seems like I am messing up a step in preparing the dataset. Couldn't find a proper answer or look up the correct solution in documentation. I have pointed out the problem line with ###, bottom part.<\/p>\n\n<pre><code>def parse_file(data_path):\n\n\n    imagepaths = list()\n    labels = list()\n    # a working parser for os is here\n\n    imagepaths = tf.constant(imagepaths, dtype=tf.string)\n    labels = tf.constant(labels, dtype=tf.float32)\n\n    return imagepaths, labels\n\n\ndef parse_image(imagepath, label):\n\n    image_string = tf.read_file(imagepath)\n    image_decoded = tf.image.decode_png(image_string, channels=3)\n    # The image size is 425x425.\n    image_resized = tf.image.resize_images(image_decoded, [img_size, img_size])\n    image_normalized = image_resized * 1.0\/255\n    print(image_normalized)\n    print(label)\n    return image_normalized, label\n\ndataset = tf.data.Dataset.from_tensor_slices((parsed_files))\ndataset = dataset.map(parse_image)\ndataset = dataset.batch(batch_size)\n\niterator = dataset.make_initializable_iterator()\niterator = iterator.get_next()\n\nx = tf.placeholder(tf.float32, [None, img_size, img_size, channels])\ny = tf.placeholder(tf.float32, [None, 1])\n<\/code><\/pre>\n\n<p>(Model goes here, irrelevant.)<\/p>\n\n<pre><code>with tf.Session() as sess:\n\n    ### AttributeError: 'tuple' object has no attribute 'initializer'\n    sess.run(iterator.initializer)\n    batch_x, batch_y = iterator.get_next()\n    test1, test2 = sess.run([batch_x, batch_y])\n    total_batch = int(total_input[0] \/ batch_size)\n    # define the iterator for the network\n    for epoch in range(epochs):\n        avg_cost = 0\n        for i in range(total_batch):\n            batch_x, batch_y = sess.run(iterator)\n            _, c = sess.run([optimiser, cross_entropy], feed_dict={x: batch_x, y: batch_y})\n            avg_cost += c \/ total_batch\n\n        test_acc = sess.run(accuracy,feed_dict={x: test_x, y: np.expand_dims(test_y, axis=-1)})\n        print(\"Epoch:\", (epoch + 1), \"cost =\", \"{:.3f}\".format(avg_cost), \" test accuracy: {:.3f}\".format(test_acc))\n        summary = sess.run(merged, feed_dict={x: test_x, y: np.expand_dims(test_y, axis=-1)})\n\n    print(\"\\nTraining complete!\")\n    print(sess.run(accuracy, feed_dict={x: test_x, y: np.expand_dims(test_y, axis=-1)}))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53145826","CreationDate":"2018-11-04T21:48:43.970","ParentId":null,"OwnerUserId":"7996768","Title":null,"Body":"<p>I've no experience with tf.Datasets but this is what might go wrong:<\/p>\n\n<pre><code>iterator = dataset.make_initializable_iterator()\niterator = iterator.get_next()\n<\/code><\/pre>\n\n<p>First you create an iterator, and you overwrite that by asking data from it with the .get_next method. This apparently gives you a tuple. Then you do:<\/p>\n\n<pre><code>sess.run(iterator.initializer)\n<\/code><\/pre>\n\n<p>And you get your error because your iterator is no longer the iterator from make_initializable_iterator(). Have you tried this:<\/p>\n\n<pre><code>iterator = dataset.make_initializable_iterator()\nwith tf.Session() as sess:\n    sess.run(iterator.initializer)\n<\/code><\/pre>\n\n<p>You might get more errors after that though, but maybe I'm wrong since I'm not used to using tf.Datasets. <\/p>\n\n<p>Check out this example I found <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#creating_an_iterator\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><code>max_value = tf.placeholder(tf.int64, shape=[])\ndataset = tf.data.Dataset.range(max_value)\niterator = dataset.make_initializable_iterator()\nnext_element = iterator.get_next()\n\n# Initialize an iterator over a dataset with 10 elements.\nsess.run(iterator.initializer, feed_dict={max_value: 10})\nfor i in range(10):\n    value = sess.run(next_element)\n    assert i == value\n\n# Initialize the same iterator over a dataset with 100 elements.\nsess.run(iterator.initializer, feed_dict={max_value: 100})\nfor i in range(100):\n    value = sess.run(next_element)\n    assert i == value\n<\/code><\/pre>\n"}]}
+{"QuestionId":53144917,"AnswerCount":0,"Tags":"<python><tensorflow><keras><precision><numerical-stability>","CreationDate":"2018-11-04T20:05:01.880","AcceptedAnswerId":null,"OwnerUserId":1727392.0,"Title":"Keras: loss fluctuates between inf and number","Body":"<p>I am working on a Keras model with a custom loss function provided by a <a href=\"https:\/\/github.com\/omimo\/Keras-MDN\" rel=\"nofollow noreferrer\">Mixture Density Network final layer<\/a> (the loss tries to minimize the negative log likelihood of some Gaussian models).<\/p>\n\n<p>What confuses me is the loss will sometimes hit an epoch in which it returns -inf as the resulting loss. Then the next iteration the loss will be a number again (e.g. -2.1). The loss sometimes bounces between negative infinity and a number every other epoch.<\/p>\n\n<p>The negative loss is evidently to be expected with a NLL loss, but this fluctuation is confusing to me. What explains this behavior within Keras? My understanding is the -inf loss is caused by numeric underflow somewhere, but I'm not sure how the model can recover from this and re-establish numeric stability thereafter.<\/p>\n\n<p>Does anyone know how this works? I'd be very grateful for any suggestions others can offer on this question.<\/p>\n","answers":[]}
+{"QuestionId":53145330,"AnswerCount":0,"Tags":"<python><opencv><tensorflow>","CreationDate":"2018-11-04T20:49:11.250","AcceptedAnswerId":null,"OwnerUserId":1907337.0,"Title":"Specify GPU device with OpenCV Python API","Body":"<p>When working on a project using Tensorflow (w\/ GPU support) to process some features extracted from OpenCV (3.4.3, using Python API) captures in real time, I am getting the following error from cuDNN whenever I try to <code>read<\/code> from the capture after I start the tf session:<\/p>\n\n<blockquote>\n  <p>F1028 02:37:31.456640 xxxxx cudnn_conv_layer.cu:28] Check failed: status == CUDNN_STATUS_SUCCESS (8 vs. 0)  CUDNN_STATUS_EXECUTION_FAILED<\/p>\n<\/blockquote>\n\n<p>I suppose that the issue is with OpenCV and Tensorflow both using GPU via CUDA at the same time and that the GPU ran out of memory. The workaround I am currently using is to start capturing with OpenCV first, and then start the tf session whenever it is actually needed. This way, TensorFlow knows that the GPU is busy and opted to use CPU only. However, the frame rate drops significantly as a result.<\/p>\n\n<p>Considering that I only use OpenCV for capturing and basic preprocessing, I don't think GPU support is necessary and it would be preferable to let TensorFlow use the GPU.<\/p>\n\n<p><strong>Is there a way to specify which GPU device to use with OpenCV Python API (or if it should use GPU at all)?<\/strong> I see with the C++ API there is a <code>setDevice()<\/code> method under the <code>gpu<\/code> namespace. Is there an equivalent for the Python API?<\/p>\n","answers":[]}
+{"QuestionId":53145845,"AnswerCount":1,"Tags":"<python><dataframe><machine-learning><keras><regression>","CreationDate":"2018-11-04T21:50:50.557","AcceptedAnswerId":53155286.0,"OwnerUserId":10322193.0,"Title":"Strange output for keras regression model","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/4qxo8.png\" rel=\"nofollow noreferrer\">data<\/a><\/p>\n\n<p>I just started to make regression model in Keras. I tried on my own photos, but I got loss as <code>Nan<\/code> and accuracy zero as output. Can anyone explain why?<\/p>\n\n<pre><code>data=pd.read_excel('Data.xlsx',sheet_name=2)\ndata.head(2)\nfeatures=data.drop('PRHP',axis=1)\nlabels=data['PRHP']\n# Normalizing data:\nm_features=features.mean(axis=0)\nm_labels=labels.mean(axis=0)\nstd_features=features.std(axis=0)\nstd_labels=labels.std(axis=0)\nnor_features=(features-m_features)\/std_features\nnor_labels=(labels-m_labels)\/std_labels\n# Build Model:\nmodel=Sequential([Dense(12,input_shape=(7,),activation='linear'),Dense(1,activation='linear')])\nmodel.summary()\nmodel.compile(adam(lr=0.01),loss='mean_squared_error',metrics=['accuracy'])\nmodel.fit(nor_features,nor_labels, batch_size=10, epochs=10, shuffle=True,verbose=2)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Epoch 1\/10\n - 0s - loss: nan - acc: 0.0000e+00\nEpoch 2\/10\n - 0s - loss: nan - acc: 0.0000e+00\nEpoch 3\/10\n - 0s - loss: nan - acc: 0.0000e+00\nEpoch 4\/10\n - 0s - loss: nan - acc: 0.0000e+00\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"53155286","CreationDate":"2018-11-05T13:23:51.103","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One of the reason for a <code>nan<\/code> loss is the existence of <code>nan<\/code> values in the training data itself. If there is any <code>nan<\/code> values in the original data try to fill them with some value or entirely remove their corresponding row sample. Otherwise, if there is no missing data in the original data, since you are normalizing by dividing by std, it might be that the value of std is zero and therefore causes division by zero. To prevent this, add a small constant when performing division:<\/p>\n\n<pre><code>nor_features = (features-m_features) \/ (std_features + 1e-7)\nnor_labels = (labels-m_labels) \/ (std_labels + 1e-7)\n<\/code><\/pre>\n\n<p>As for the accuracy being zero, you must keep in mind that the <code>'accuracy'<\/code> metric in Keras is only used for classification problems. Here, since you are doing regression it does not make sense at all to use <code>'accuracy'<\/code> as the metric. Instead, you can use other metrics like <code>mae<\/code> (i.e. mean absolute error) or don't use any metric at all.<\/p>\n"}]}
+{"QuestionId":53146142,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2018-11-04T22:27:48.353","AcceptedAnswerId":null,"OwnerUserId":10605219.0,"Title":"Installing tensorflow-gpu broke tensorflow","Body":"<p>I'm using tensorflow cpu via anaconda and after using the command<\/p>\n\n<p>conda install tensorflow-gpu<\/p>\n\n<p>I got the following error message<\/p>\n\n<blockquote>\n  <p>File \"\", line 1, in \n  runfile('C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC\n  example\/Neural Network Worked Example.py',\n  wdir='C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC example')<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\",\n  line 705, in runfile\n      execfile(filename, namespace)<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\",\n  line 102, in execfile\n      exec(compile(f.read(), filename, 'exec'), namespace)<\/p>\n  \n  <p>File \"C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC example\/Neural\n  Network Worked Example.py\", line 7, in \n      from keras.models import Sequential as sequential<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras__init__.py\",\n  line 3, in \n      from . import utils<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\utils__init__.py\",\n  line 6, in \n      from . import conv_utils<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\utils\\conv_utils.py\",\n  line 9, in \n      from .. import backend as K<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend__init__.py\",\n  line 89, in \n      from .tensorflow_backend import *<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\",\n  line 5, in \n      import tensorflow as tf<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow__init__.py\",\n  line 22, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py\",\n  line 49, in \n      from tensorflow.python import pywrap_tensorflow<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 74, in \n      raise ImportError(msg)<\/p>\n  \n  <p>ImportError: Traceback (most recent call last):   File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\",\n  line 243, in load_module\n      return load_dynamic(name, filename, file)   File \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n      return _load(spec) ImportError: DLL load failed: The specified module could not be found.<\/p>\n  \n  <p>Failed to load the native TensorFlow runtime.<\/p>\n  \n  <p>See\n  <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n  \n  <p>for some common reasons and solutions.  Include the entire stack trace\n  above this error message when asking for help.<\/p>\n<\/blockquote>\n\n<p>I'm not even trying to use the GPU at the minute since it hasn't arrived yet and this is CPU only mode, I'm very confused as to what could be wrong.<\/p>\n","answers":[{"AnswerId":"53152742","CreationDate":"2018-11-05T10:40:15.507","ParentId":null,"OwnerUserId":"977098","Title":null,"Body":"<p><code>tensorflow-gpu<\/code> is not just an additional module for tensorflow but rather a separate version of <code>tensorflow<\/code>. When you installed <code>tensorflow-gpu<\/code> the <code>tensorflow<\/code> package got uninstalled. Now loading the tensorflow-gpu version fails, since you probably are missing all of the CUDA libraries. Just remove the <code>tensorflow-gpu<\/code> package and reinstall the old <code>tensorflow<\/code> package again until you actually have a CUDA enabled GPU.<\/p>\n"}]}
+{"QuestionId":53146273,"AnswerCount":1,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2018-11-04T22:48:02.253","AcceptedAnswerId":53146563.0,"OwnerUserId":1114272.0,"Title":"Tensorflow computing an aggregate function over lots of data in tfrecord format","Body":"<p>I have a very large dataset stored in chunks of 5000 across several tfrecord files. Together all of these records are way larger than my RAM. What I would like to do is sample from N = 0.05 * TOTAL_SIZE random indexes in the dataset and compute the mean and std deviation to normalize my data.<\/p>\n\n<p>If it wasn't for the dataset size this would be easy, but I run out of memory even when I try and compute a sum of all the tensors I'm interested in.<\/p>\n\n<pre><code># NOTE: count is computed ahead of time by looping over all the tfrecord entries\n\nwith tf.device('\/cpu:0'):\n    sample_size = int(count * 0.05)\n    random_indexes = set(np.random.randint(low=0, high=count, size=sample_size))\n    stat_graph = tf.Graph()\n    with tf.Session(graph=stat_graph) as sess:\n        val_sum = np.zeros(shape=(180, 2050))\n        for file in files:\n            print(\"Reading from file: %s\" % file)\n            for record in tf.python_io.tf_record_iterator(file):\n                features = tf.parse_single_example(\n                    record,\n                    features={\n                        \"val\": tf.FixedLenFeature((180, 2050), tf.float32),\n                    })\n                if index in random_indexes:\n                    val_sum += features[\"val\"].eval(session=sess)\n                index += 1\n        val_mean = val_sum \/ sample_size\n<\/code><\/pre>\n\n<p>What is the right way to compute some aggregate function, i.e. mean and\/or standard deviation, over a tfrecord dataset?<\/p>\n","answers":[{"AnswerId":"53146563","CreationDate":"2018-11-04T23:37:26.853","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>I think that tf.parse_single_example adds a new tensor to the graph every time it's called.  Instead of the above, you should feed the string with a placeholder:<\/p>\n\n<pre><code>...\nrecord_placeholder = tf.placeholder(tf.string)\nfeatures = tf.parse_single_example(\n    record_placeholder,\n    features={\n        \"val\": tf.FixedLenFeature((180, 2050), tf.float32),\n    })\nfor record in tf.python_io.tf_record_iterator(file):\n...\nval_sum += features[\"val\"].eval(feed_dict={record_placeholder: record}, session=sess)\n<\/code><\/pre>\n\n<p>Let me know if this works since I have no way of testing it.<\/p>\n"}]}
+{"QuestionId":53146525,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow-estimator><tfrecord>","CreationDate":"2018-11-04T23:30:08.523","AcceptedAnswerId":null,"OwnerUserId":6695432.0,"Title":"convert strings extracted from TFRecord to one_hot with tensorflow","Body":"<p>I have encoded some images in TFrecord format in order to train a CNN.\nHere's there is the function I use to read the TFrecord files and extract images.\nI have problems reading labels (strings composed of 5 digits in range 10000-50000 and sparse) from tfrecord and convert those strings into \"one hot\" encoded tensors in order train my classifier.\nThe training is made with a custom Estimator using tensorflow.\nhere's the snippet of the function I used to read the TFRecords files<\/p>\n\n<pre><code>def imgs_input_fn(filenames, classes, perform_shuffle=False, repeat_count=1, batch_size=1):\n    def _parse_function(serialized):\n        features = \\\n        {\n            'image\/encoded': tf.FixedLenFeature([], tf.string),\n            'image\/width': tf.FixedLenFeature([], tf.int64),\n            'image\/height': tf.FixedLenFeature([], tf.int64),\n            'image\/channels': tf.FixedLenFeature([], tf.int64),\n            'image\/colorspace': tf.FixedLenFeature([], tf.string),\n            'image\/class\/label': tf.FixedLenFeature([], tf.string),\n            'image\/class\/text_label': tf.FixedLenFeature([], tf.string),\n            'image\/filename': tf.FixedLenFeature([], tf.string)\n        }\n\n        # Parse the serialized data so we get a dict with our data.\n        parsed_example = tf.parse_single_example(serialized=serialized,\n                                                 features=features)\n        # Get the image as raw bytes.\n        # in image_shape I can't use parsed_example['image\/channels'] \n        # read from file but need to pass 1 to the shape...\n        # how to get this?\n        channels = parsed_example['image\/channels']\n        image_shape = tf.stack([parsed_example['image\/width'], \n                      parsed_example['image\/height'], 1])\n        image_raw = parsed_example['image\/encoded']\n        # Labels are string representing numbers but are sparse\n        label = tf.string_to_number(parsed_example['image\/class\/label'], out_type=tf.int32)\n        # Check how to pass the value read from the tfrecord file\n        image = tf.image.decode_image(image_raw)\n        image = tf.divide(tf.cast(image, tf.float32), tf.constant(255., dtype=tf.float32))\n        image = tf.reshape(image, image_shape)\n\n        num_classes = classes\n        #The following operation does not give me expecte result \n        # as as labels are strings like 12345, 34234, 53453, \n        # and I have only ie 100 classes so tf.one_hot(10000, 100)\n        # will give me a tensor with only 0s in it\n        d = dict(zip([input_name], [image])), tf.one_hot(label, num_classes)\n        return d\n\n\n    dataset = tf.data.TFRecordDataset(filenames=filenames)\n    # Parse the serialized data in the TFRecords files.\n    # This returns TensorFlow tensors for the image and labels.\n    dataset = dataset.map(_parse_function)\n    if perform_shuffle:\n        # Randomizes input using a window of 1024 elements (read into memory)\n        dataset = dataset.shuffle(buffer_size=1024)\n    dataset = dataset.repeat(repeat_count)  # Repeats dataset this # times\n    dataset = dataset.batch(batch_size)  # Batch size to use\n    iterator = dataset.make_one_shot_iterator()\n    batch_features, batch_labels = iterator.get_next()\n    return batch_features, batch_labels\n<\/code><\/pre>\n\n<p>So how can I fill ie a structure like a lookup table, using for example tf.contrib.lookup.index_table_from_tensor, reading the information directly from the TFRecord files as the images are read for training  not providing a file in advance or reading all the TFRecords extracting beforehand the labels?\nI would like to leverage the fact that if a lablel is \"unknown\" for the lookup table the \"index_table_from_tensor\" will use the hash value of the label to give a coherent result.\nThe function I wrote is called from a train loop tf.estimator.train_and_evaluate after I defined the tf.estimator.TrainSpec and tf.estimator.EvalSpec and I use a keras model<\/p>\n\n<p>Is there a way to achieve this?<\/p>\n\n<p>Thanks a lot.<\/p>\n\n<p>Seba<\/p>\n","answers":[]}
+{"QuestionId":53146638,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><conv-neural-network><convolution><tensor>","CreationDate":"2018-11-04T23:49:47.537","AcceptedAnswerId":null,"OwnerUserId":1782677.0,"Title":"Tensorflow: How can I convolve image with custom constant filter","Body":"<p>I have 3 5x5 filters I want to convolve on a grayscale image (shape [nx,ny,1]) input. I have preset hardcoded values each of these 5x5 filters need to be and I don't want them to be \"learned\" by my model but just be a constant operation. <\/p>\n\n<p>How do I achieve this? <\/p>\n\n<p>I was looking into using tf.nn.conv2d() and it says its filters need to be of shape [height, width, input, output] so I tried to use tf.constant() to create a tensor for my filters of shape [5,5,1,3] (so 3 filters of shape 5x5 applied to an input with 1 channel) but the results from tf.constant() did not look right. It came out like this: <\/p>\n\n<pre><code> [[[[ -5   7  -12]]\n\n   [[  21  0   2]]\n\n   [[ -6   9  -6]]\n\n   [[  2  -2   8]]\n\n   [[-6   4  -1]]]\n\n\n  [[[  2  -6   8]]\n\n   [[ -6   2  -1]]\n\n   [[  2  -2   2]]\n\n   [[ -1   1   5]]\n\n   [[  4   3   2]]]\n\n ...etc\n<\/code><\/pre>\n\n<p>which does not look like the shape of 3 5x5 filters. <\/p>\n\n<p>If I use tf.constant() with the shape as [1,3,5,5] I get this: <\/p>\n\n<pre><code>[[[[ -5   7  -12   21  0]\n   [  2  -6   9  -6   2]\n   [ -2   8 -6   4  -1]\n   [  2  -6   8  -6   2]\n   [ -1   2  -2   2  -1]]\n\n  [[  1   5   4   3   2]\n   [  4   0  -2   0   4]\n   [  2  -1   7  -3   5]\n   [  -1   0  -1   0  -1]\n   [  5   0   9   0   5]]\n\n   ...etc\n<\/code><\/pre>\n\n<p>which <em>does<\/em> look like 5x5 filters but it's not the correct shape taken in by tf.nn.conv2d() <\/p>\n\n<p>so I'm confused by this mismatch and don't know what the right thing to do is.<\/p>\n","answers":[{"AnswerId":"53146948","CreationDate":"2018-11-05T00:45:45.617","ParentId":null,"OwnerUserId":"2060205","Title":null,"Body":"<p>It's better not to worry about how the filters look like. Just keep track of the shape to make sure they make sense.<\/p>\n\n<p>Here is an example to apply 2 Sobel filters to an image:<\/p>\n\n<pre><code>from skimage import data\nimg = np.expand_dims(data.camera(), -1)\nimg = np.expand_dims(img, 0)  # shape: (1, 512, 512, 1)\n\nsobel_x = np.array([[-0.25, -0.2 ,  0.  ,  0.2 ,  0.25],\n                   [-0.4 , -0.5 ,  0.  ,  0.5 ,  0.4 ],\n                   [-0.5 , -1.  ,  0.  ,  1.  ,  0.5 ],\n                   [-0.4 , -0.5 ,  0.  ,  0.5 ,  0.4 ],\n                   [-0.25, -0.2 ,  0.  ,  0.2 ,  0.25]])\n\nsobel_y = np.array([[-0.25, -0.4 , -0.5 , -0.4 , -0.25],\n                   [-0.2 , -0.5 , -1.  , -0.5 , -0.2 ],\n                   [ 0.  ,  0.  ,  0.  ,  0.  ,  0.  ],\n                   [ 0.2 ,  0.5 ,  1.  ,  0.5 ,  0.2 ],\n                   [ 0.25,  0.4 ,  0.5 ,  0.4 ,  0.25]])\n\nfilters = np.concatenate([[sobel_x], [sobel_y]])  # shape: (2, 5, 5)\nfilters = np.expand_dims(filters, -1)  # shape: (2, 5, 5, 1)\nfilters = filters.transpose(1, 2, 3, 0)  # shape: (5, 5, 1, 2)\n\n# Convolve image\nans = tf.nn.conv2d((img \/ 255.0).astype('float32'),\n                   filters,\n                   strides=[1, 1, 1, 1],\n                   padding='SAME')\n\nwith tf.Session() as sess:\n    ans_np = sess.run(ans)  # shape: (1, 512, 512, 2)\n\nfiltered1 = ans_np[0, ..., 0]\nfiltered2 = ans_np[0, ..., 1]\n<\/code><\/pre>\n\n<p>The image is correctly convolved with the 2 filters and the resulting images look like:<\/p>\n\n<pre><code>plt.matshow(filtered1)\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fLYwE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fLYwE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>plt.matshow(filtered2)\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rRRiW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rRRiW.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"53146733","CreationDate":"2018-11-05T00:05:45.037","ParentId":null,"OwnerUserId":"8163791","Title":null,"Body":"<p>The first case seems to be correctly showing filters of shape <code>[5, 5, 1, 3]<\/code>. Take a look at the number of square brackets - tensorflow is showing the 4th dimension across the horizontal of each of those little boxes, and the 2nd dimension down the vertical. (The 3rd has dimensionality of 1, so there's no point in showing that)<\/p>\n"}]}
+{"QuestionId":53146876,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2018-11-05T00:32:39.087","AcceptedAnswerId":null,"OwnerUserId":10588132.0,"Title":"weights values per iteration for ANN in Keras","Body":"<p>I am trying to output weights values for the last layer of my ANN per iteration. But from the following code I wrote, I could only output it for the last iteration, which means the final result. So how can I add it to per iteration? <\/p>\n\n<pre><code>model=Sequential()\nmodel.add(Dense(3,activation='sigmoid',input_dim=8))\nmodel.add(Dense(3,activation='sigmoid'))\nmodel.add(Dense(10,activation='softmax'))\n\nsgd=SGD(lr=0.1)\nmodel.compile(optimizer=sgd,loss='categorical_crossentropy',metrics=['accuracy'])\n\nprint(\"Training----------\")\nmodel.fit(features_train,class_train,validation_data=(features_test,class_test),nb_epoch=100,batch_size=1)\n\nweights = np.array(model.layers[2].get_weights())\nprint(weights)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53146937","CreationDate":"2018-11-05T00:43:16.797","ParentId":null,"OwnerUserId":"10574412","Title":null,"Body":"<p>I would strongly reccomend you to read this:<\/p>\n\n<p><a href=\"https:\/\/machinelearningmastery.com\/check-point-deep-learning-models-keras\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/check-point-deep-learning-models-keras\/<\/a><\/p>\n"}]}
+{"QuestionId":53147068,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-05T01:08:19.253","AcceptedAnswerId":null,"OwnerUserId":10605522.0,"Title":"How to address Error when I import tensorflow as tf","Body":"<p>I install the tensorflow python3.6 cpu version but when I import tensorflow as tf, it shows import: not authorized `tf' @ error\/constitute.c\/WriteImage\/1028. Is there anyone can help me? My system is ubuntu 16.04<\/p>\n","answers":[{"AnswerId":"53153601","CreationDate":"2018-11-05T11:34:26.547","ParentId":null,"OwnerUserId":"9427281","Title":null,"Body":"<p>Try to create a conda environment with python 3.6(you can change the python version to suit your needs) using following commands and try installing tensorflow <\/p>\n\n<pre><code>conda create -n &lt;env-name&gt; python=3.6\nsource activate &lt;env-name&gt;\npip install tensorflow\n<\/code><\/pre>\n\n<p>It will resolve your issue<\/p>\n"},{"AnswerId":"58596890","CreationDate":"2019-10-28T19:00:39.747","ParentId":null,"OwnerUserId":"6089653","Title":null,"Body":"<p>I got a similar issue.<\/p>\n\n<p>Did you mention which interpreter to use?<\/p>\n\n<p>You can either add the shebang <code>#!\/usr\/bin\/env python3<\/code> at the beginning of your script or run the script as <code>python3 script.py<\/code>. This fixed the issue for me.<\/p>\n"}]}
+{"QuestionId":53147241,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras><keras-layer>","CreationDate":"2018-11-05T01:37:22.367","AcceptedAnswerId":53149891.0,"OwnerUserId":10395181.0,"Title":"In using keras Lambda, how do I handle \"TypeError: Object arrays are not currently supported\"?","Body":"<p>I'm using Keras, and I want to make a layer that takes <code>[a0, a1]<\/code>, <code>[b0, b1, b2]<\/code> as inputs and gives <code>[a0*b0, a0*b1, a0*b2, a1*b0, a1*b1, a1*b2]<\/code> as output. I tried to use <code>Lambda<\/code>, but I couldn't succeed. Here's my code:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Input\nfrom keras.layers import Lambda\n\ndef mix(A):\n    reshaped = [np.reshape(A[m], (1,np.size(A[m]))) for m in range(len(A))]\n    mixed = reshaped[-1]\n\n    for i in range(len(A)-1):\n        mixed = np.matmul(np.transpose(reshaped[-i-2]), mixed)\n        mixed = np.reshape(mixed, (1,np.size(mixed)))\n\n    return np.reshape(mixed, np.size(mixed))\n\na = Input(shape=(2,))\nb = Input(shape=(3,))\nc = Lambda(mix)([a, b])\n<\/code><\/pre>\n\n<p>Here's the error I got:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-32-07bbf930b48b&gt; in &lt;module&gt;()\n      1 a = Input(shape=(2,))\n      2 b = Input(shape=(3,))\n----&gt; 3 c = Lambda(mix)([a, b])\n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\engine\\base_layer.py in __call__(self, inputs, **kwargs)\n    455             # Actually call the layer,\n    456             # collecting output(s), mask(s), and shape(s).\n--&gt; 457             output = self.call(inputs, **kwargs)\n    458             output_mask = self.compute_mask(inputs, previous_mask)\n    459 \n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\layers\\core.py in call(self, inputs, mask)\n    685         if has_arg(self.function, 'mask'):\n    686             arguments['mask'] = mask\n--&gt; 687         return self.function(inputs, **arguments)\n    688 \n    689     def compute_mask(self, inputs, mask=None):\n\n&lt;ipython-input-31-bbc21320d8af&gt; in mix(A)\n      4 \n      5     for i in range(len(A)-1):\n----&gt; 6         mixed = np.matmul(np.transpose(reshaped[-i-2]), mixed)\n      7         mixed = np.reshape(mixed, (1,np.size(mixed)))\n      8 \n\nTypeError: Object arrays are not currently supported\n<\/code><\/pre>\n\n<p>But if I put:<\/p>\n\n<pre><code>a = np.array([1,2])\nb = np.array([3,4,5])\nprint(mix([a,b]))\n<\/code><\/pre>\n\n<p>then I get:<\/p>\n\n<pre><code>[ 3  4  5  6  8 10]\n<\/code><\/pre>\n\n<p>which is exactly what I intended. But I don't know how to put this in <code>Lambda<\/code> properly.<\/p>\n\n<p>Can anyone tell me how to handle this? I'm new to Keras, so I don't know the internal structure of <code>Lambda<\/code>, <code>Input<\/code> or other stuffs.<\/p>\n\n<hr>\n\n<p>Following Abhijit's comment, I changed the code like this:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom keras.models import Input\nfrom keras.layers import Lambda\n\ndef mix(A):\n    reshaped = [tf.reshape(A[m], (1,tf.size(A[m]))) for m in range(len(A))]\n    mixed = reshaped[-1]\n\n    for i in range(len(A)-1):\n        mixed = tf.matmul(tf.transpose(reshaped[-i-2]), mixed)\n        mixed = tf.reshape(mixed, (1,tf.size(mixed)))\n\n    return tf.reshape(mixed, [tf.size(mixed)])\n\na = Input(shape=(2,))\nb = Input(shape=(3,))\nc = Lambda(mix)([a, b])\n<\/code><\/pre>\n\n<p>Now I don't get any errors, but I don't think I got the right neural network. Because executing:<\/p>\n\n<pre><code>model = Model(inputs=[a,b], outputs=c)\nprint(model.summary())\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_22 (InputLayer)           (None, 2)            0                                            \n__________________________________________________________________________________________________\ninput_23 (InputLayer)           (None, 3)            0                                            \n__________________________________________________________________________________________________\nlambda_3 (Lambda)               (None,)              0           input_22[0][0]                   \n                                                                 input_23[0][0]                   \n==================================================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>But see the layer <code>lambda_3<\/code>. Shouldn't the output shape be <code>(None, 6)<\/code>?<\/p>\n","answers":[{"AnswerId":"53149891","CreationDate":"2018-11-05T07:19:08.660","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Apart from the fact that you need to use Keras backend functions (i.e. <code>keras.backend.*<\/code>) or use backend functions directly (i.e. <code>tf.*<\/code> or <code>th.*<\/code>), I think you are making the definition of <code>mix<\/code> unnecessarily complicated. It can be done much simpler like this:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef mix(ts):\n    t0 = K.expand_dims(ts[0], axis=-1)\n    t1 = K.expand_dims(ts[1], axis=1)\n    return K.batch_flatten(t0 * t1)\n\na = Input(shape=(2,))\nb = Input(shape=(3,))\nc = Lambda(mix)([a, b])\n\nmodel = Model(inputs=[a,b], outputs=c)\n<\/code><\/pre>\n\n<p>Here is the test:<\/p>\n\n<pre><code># the reshapes are necessary to make them a batch\na = np.array([1,2]).reshape(1,2)\nb = np.array([3,4,5]).reshape(1,3)\nprint(model.predict([a, b]))\n\n# output\n[[ 3.  4.  5.  6.  8. 10.]]\n<\/code><\/pre>\n\n<p>Further, sometimes the <code>Lambda<\/code> layer could automatically infer the output shape. However, if you would like you can explicitly set its output shape:<\/p>\n\n<pre><code>c = Lambda(mix, output_shape=(6,))([a, b])\n<\/code><\/pre>\n\n<p>Model summary:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_9 (InputLayer)            (None, 2)            0                                            \n__________________________________________________________________________________________________\ninput_10 (InputLayer)           (None, 3)            0                                            \n__________________________________________________________________________________________________\nlambda_5 (Lambda)               (None, 6)            0           input_9[0][0]                    \n                                                                 input_10[0][0]                   \n==================================================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n__________________________________________________________________________________________________\n<\/code><\/pre>\n"}]}
+{"QuestionId":53147365,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><createml>","CreationDate":"2018-11-05T01:58:06.107","AcceptedAnswerId":null,"OwnerUserId":7966026.0,"Title":"How to develop a neural network that can identify a song by composer","Body":"<p>I have recently undertaken a project to test the efficiency of machine learning algorithms. However, I have had extensive trouble finding a platform to bring my idea into realization. Here is what I am attempting to do.<\/p>\n\n<p>I am attempting to create a neural network that can identify classical music pieces by composer. I will train it using a set of classical clips, each 5 seconds long and by 5 different composers. There will be 25 such songs in all. The algorithm, without any integration into an application, has to be able to identify the songs. There is a specific set of requirements I have for this algorithm.<\/p>\n\n<p>1: It must be available for MacOS Mojave or High Sierra as an application or library in the Python programming language.<\/p>\n\n<p>2: It must have some sort of documentation online as to how to create a neural network in it.<\/p>\n\n<p>3: It must be able to handle auditory stimuli. (For example, Xcode 10.1 cannot handle auditory stimuli using an MLClassifier.)<\/p>\n\n<p>4: The user must be able to alter the number of hidden layers in the algorithm easily.<\/p>\n\n<p>If all of these requirements are met and if I am able to resolve additional problems in correspondence, my problem will be solved.  I look forward to working with members of the Stack Overflow community in order to find a solution. Thank you for any assistance you can provide.<\/p>\n\n<p>Sincerely, \nSuren Grigorian<\/p>\n","answers":[{"AnswerId":"53147672","CreationDate":"2018-11-05T02:46:35.627","ParentId":null,"OwnerUserId":"10553976","Title":null,"Body":"<p>Problem: <\/p>\n\n<pre><code>The algorithm, without any integration into an application, has to be able to identify the songs.\n<\/code><\/pre>\n\n<p>Requirements: <\/p>\n\n<pre><code>1: It must be available for MacOS Mojave or High Sierra as an application or library in the Python programming language.\n\n2: It must have some sort of documentation online as to how to create a neural network in it.\n\n3: It must be able to handle auditory stimuli. (For example, Xcode 10.1 cannot handle auditory stimuli using an MLClassifier.)\n\n4: The user must be able to alter the number of hidden layers in the algorithm easily.\n<\/code><\/pre>\n\n<p>This is troublesome. 1) contradicts the problem statement - it states a MacOS, or Python integration requirement. 2) states an integration requirement (online docs.) 4) compromises the efficacy requirement (the user could reshape the model until it is useless or too heavy to operate on available resources.) Why are you obligated to prepare the data in 5 second snippets? Data science projects with confusing requirements can do more harm than good.<\/p>\n\n<p>Setting all this aside, I guess your project sounds like a good candidate for tensorflow. You can train in any environment that lets you persist to tf graphs and weights (e.g. keras) and then use the converter to persist your model for tensorflow.js if you want seamless integration with d3 or a modern webapp.<\/p>\n\n<p>If you haven't implemented an audio classifier before, tf has a good starter project in their docs: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition<\/a><\/p>\n"}]}
+{"QuestionId":53147724,"AnswerCount":0,"Tags":"<python><tensorflow><image-processing><deep-learning><tfrecord>","CreationDate":"2018-11-05T02:55:03.173","AcceptedAnswerId":null,"OwnerUserId":7353970.0,"Title":"How to decode a list of bytes stored in tfrecords using TensorFlow?","Body":"<p>I have stored a varying number of images as a list of bytes:<\/p>\n\n<pre><code>img.append(trajectory_step['img'].tostring())\nfeature['img'] = _bytes_feature(img)\n...\nexample = tf.train.Example(features=tf.train.Features(feature=feature))\nwriter.write(example.SerializeToString())\n<\/code><\/pre>\n\n<p>I also made sure to save the number of images for later decoding (<code>experiment_length<\/code>).<\/p>\n\n<p>Now I fail to decode the images in the following way:<\/p>\n\n<pre><code>features = {\n'img': tf.VarLenFeature(tf.string),\n...\n}\nparsed_features = tf.parse_single_example(example_proto, features)\nimg = tf.decode_raw(parsed_features['img'], out_type=tf.uint8)\nimg = tf.reshape(img, tf.stack([experiment_length, 120, 160, 3]))\n<\/code><\/pre>\n\n<p>Which yields following error:<\/p>\n\n<blockquote>\n  <p>TypeError: Expected string passed to parameter 'bytes' of op\n  'DecodeRaw', got\n   of type 'SparseTensor' instead.<\/p>\n<\/blockquote>\n\n<p>If I choose to use <code>tf.FixedLenFeature<\/code> I get the following error:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Name:\n  , Key: img, Index: 0.  Number of bytes values != expected. \n  Values size: 5 but output shape: []<\/p>\n<\/blockquote>\n\n<p>How do I properly decode a list of bytes? And: Is <code>tf.VarLenFeature<\/code> in this case correct or should I use <code>tf.FixedLenFeature<\/code>?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":53148176,"AnswerCount":0,"Tags":"<python><macos><tensorflow><neural-network><pip>","CreationDate":"2018-11-05T04:03:09.833","AcceptedAnswerId":null,"OwnerUserId":7966026.0,"Title":"Tensorflow installation Mac error: unable to find matching distribution","Body":"<p>I am attempting to install TensorFlow on my Macintosh computer. I was following the instructions as provided on their website when I reached a problem. I had established a virtual environment in the MacOS terminal and attempted to use pip to install TensorFlow with the command<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>when I received the following message:<\/p>\n\n<pre><code>Collecting tensorflow\n  Could not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow\n<\/code><\/pre>\n\n<p>How can I resolve this? Thank you for any assistance you can provide.<\/p>\n\n<p>Sincerely, \nSuren Grigorian<\/p>\n","answers":[]}
+{"QuestionId":53148889,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T05:37:40.813","AcceptedAnswerId":null,"OwnerUserId":10606215.0,"Title":"How to reuse a placeholder in tensorflow?","Body":"<p>I tried to reuse a placeholder for different shapes of input but failed. And I search for the answer on the Internet while I only found that how to reuse variables by variable_scope(\"\",reuse=tf.AUTO_REUSE). I try the same method while it didn't work. Python throw an exception \"ValueError: Duplicate node name in graph: 'input_image'\". I'm confused and thanks for the help in advance.<\/p>\n\n<p>For example, I want to train a net with pictures in 128*128 and I want to test the net with pictures in 1080*960:<\/p>\n\n<pre><code>        with tf.variable_scope('test\/', reuse=tf.AUTO_REUSE):\n            input_image = tf.placeholder(shape=[batch_size, 128, 128, 3], dtype=tf.float32, name='input_image\/')\n        aug_image= train_net(input_image)\n        sess = tf.Session()\n        init = tf.global_variables_initializer()\n        sess.run(init)\n        for i in range(100000):\n           augg = sess.run([aug_image],feed_dict={input_image: tmp})\n           if i % 1000 == 0:\n               with tf.variable_scope('test\/', reuse=True):\n                   input_image = tf.placeholder(shape=[1, 1080, 960, 3],dtype=tf.float32,name='input_image\/')\n                   aug = sess.run([aug_image],{input_image: pic1})\n                   show(aug)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53148901,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-05T05:38:54.950","AcceptedAnswerId":53149250.0,"OwnerUserId":10606197.0,"Title":"error with checking target: expected dense_49 to have 4 dimensions, but got array with shape (2250, 3)","Body":"<pre><code>model=models.Sequential()\n\nmodel.add(layers.Dense(7, activation='relu', input_shape=(50,50,3)))\n\nmodel.add(layers.Dense(7, activation='relu'))\n\nmodel.add(layers.Dense(3, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nmodel.fit(X_train,y_train,batch_size=50,epochs=10, verbose=1)\n\nresults = model.evaluate(X_test, y_test)\n<\/code><\/pre>\n\n<p>The above is my python code. Can anyone tell my what am I missing. FYI, I have 2250 images in my training set, each of the shape <code>(50,50,3)<\/code>.<\/p>\n","answers":[{"AnswerId":"53149250","CreationDate":"2018-11-05T06:17:26.223","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>Let's look to the following dense layers Image from <a href=\"https:\/\/blog.datawow.io\/interns-explain-basic-neural-network-ebc555708c9\" rel=\"nofollow noreferrer\">here<\/a>. \n<a href=\"https:\/\/i.stack.imgur.com\/tdblE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tdblE.png\" alt=\"enter image description here\"><\/a>\nDense layer connects each node in <em>layer-L<\/em> to every node in  <em>layer-L+1<\/em>. Even-though it is not necessarily true, the input to dense layer often have shape (batch_size, width, channels). Keras documentation says: \"<a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\">If the input to the layer has a rank greater than 2, then it is flattened prior to the initial dot product with  kernel<\/a>\".<br>\nBack to your question. In your case you specified the input shape to be((50,50,3). Now the first dense layer expects inputs with shape(batch_size, 50, 50,3) and produces output with shape [batch_size, 50,50, 7],i.e inputs and outputs have rank 4. The same goes for the last layer with softmax activation. So the output of the last layer will have rank 4, i.e shape (batch_size, 50, 50, 3). For this reason the model is expecting the labels to have this(batch_size, 50, 50, 3) shape.<\/p>\n\n<p><strong>Solution<\/strong><br>\nUse vectors as input to dense layer. Steps to follow.<\/p>\n\n<ol>\n<li><p>Change the input shape argument<\/p>\n\n<p><code>model.add(layers.Dense(7, activation='relu', input_shape=(50*50*3,)))<\/code><\/p><\/li>\n<li><p>Reshape X_train and X_test to deal with new input_shape   <\/p><\/li>\n<\/ol>\n\n<p><code>#assuming X_train and X_test are numpy array<\/code><br>\n<code>X_train  = X_train.reshape(-1,50*50*3)<\/code><br>\n<code>X_test  = X_test.reshape(-1,50*50*3)<\/code><\/p>\n"}]}
+{"QuestionId":53149059,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T05:56:09.103","AcceptedAnswerId":null,"OwnerUserId":7212365.0,"Title":"Why do these two fc api act differently?","Body":"<p>I was implementing the resnet and training on cifar-10 dataset. After \"global average pooling\" layer, I add 10-way fully connected layer. I used two tensorflow API to implement fc layer. However, the method 2 works to me, but the method 1 doesn't work. The method 1 is according to tensorflow official api: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/fully_connected\" rel=\"nofollow noreferrer\">tf.contrib.layers.fully_connected<\/a>. The second method is according to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/dense\" rel=\"nofollow noreferrer\">tf.layers.dense<\/a>. I'm so confused why the first method doesn't work here. FYI, the output of global average pooling layer is [batch_size, 1, 1, 64]. Looking for some help! Thanks in advance!<\/p>\n\n<p>Edit: Doesn't work means if I use the first method, all output logit of fc become zero, which doesn't make any sense to me. <\/p>\n\n<pre><code># Method 1: \ndef fc(x, n_outputs, scope):\n  with tf.variable_scope(scope):\n    # x = tf.layers.flatten(x)\n    x = tf.contrib.layers.fully_connected(x, num_outputs=n_outputs, weights_initializer=weight_init, weights_regularizer=weight_reg, scope=scope)\n    return x\n\n# Method 2:\ndef fc(x, n_outputs, scope):\n  with tf.variable_scope(scope):\n     x = tf.layers.flatten(x)\n     x = tf.layers.dense(x, units=n_outputs, kernel_initializer=weight_init, kernel_regularizer=weight_reg)\n     return x\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53149126,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-05T06:04:05.357","AcceptedAnswerId":53176825.0,"OwnerUserId":260826.0,"Title":"Tensorflow Estimator shuffle","Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets_for_estimators\" rel=\"nofollow noreferrer\">this<\/a> link <\/p>\n\n<p>The <code>tf.data.Dataset.shuffle<\/code> method uses a fixed-size buffer to shuffle the items as they pass through. In this case the <code>buffer_size<\/code> is greater than the number of examples in the Dataset, ensuring that the data is completely shuffled (The Iris data set only contains 150 examples).<\/p>\n\n<p>Is shuffle method created in addition to feature engineering when I would shuffle and split my dataset to create training data? What is the advantage of: <code>\"to shuffle the items as they pass through\".<\/code><\/p>\n","answers":[{"AnswerId":"53176825","CreationDate":"2018-11-06T17:16:31.893","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>shuffle is a generally useful method if you want to shuffle your data and you didnt have an option to be able to do it before, like when you are reading from a generators for example. The shuffle method is also really useful if you want to shuffle your data each time you iterate through it (typically in combination with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#repeat\" rel=\"nofollow noreferrer\">repeat<\/a> method). Here is a quick example that shows why:<\/p>\n\n<pre><code>dataset1 = tf.data.Dataset.range(5).shuffle(5).repeat(3).make_one_shot_iterator().get_next()\ndataset2 = tf.data.Dataset.range(5).shuffle(5, reshuffle_each_iteration=False).repeat(3).make_one_shot_iterator().get_next()\n\nsess = tf.Session()\n\nfor i in range(15):\n    d1, d2 = sess.run([dataset1, dataset2])\n    print(\"Dataset1: {}, Dataset2:, {}\".format(d1, d2))\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>Dataset1: 0, Dataset2:, 3\nDataset1: 3, Dataset2:, 0\nDataset1: 4, Dataset2:, 1\nDataset1: 1, Dataset2:, 2\nDataset1: 2, Dataset2:, 4\nDataset1: 1, Dataset2:, 3\nDataset1: 2, Dataset2:, 0\nDataset1: 3, Dataset2:, 1\nDataset1: 4, Dataset2:, 2\nDataset1: 0, Dataset2:, 4\nDataset1: 3, Dataset2:, 3\nDataset1: 0, Dataset2:, 0\nDataset1: 4, Dataset2:, 1\nDataset1: 1, Dataset2:, 2\nDataset1: 2, Dataset2:, 4\n<\/code><\/pre>\n\n<p>As you can see, the second dataset is repeating the elements it shuffled once, while the first dataset shuffles the data each time it iterates through the dataset.<\/p>\n"}]}
+{"QuestionId":53149189,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T06:10:46.143","AcceptedAnswerId":null,"OwnerUserId":10606279.0,"Title":"How get trainable variable count when using Tf Estimator?","Body":"<p>I created a CNN classifier model using the tf estimator framework. However, I could not access variables defined in the model. tf.trainable_variables() always return 0.\n  How can i access variables using tf estimator? In particular, how can i get a count of the total number of parameters (adding up the dimensions of all variables.<\/p>\n\n<p>Thanks,\nHarold<\/p>\n","answers":[{"AnswerId":"56390017","CreationDate":"2019-05-31T06:47:03.340","ParentId":null,"OwnerUserId":"6055926","Title":null,"Body":"<p>You can use <code>get_variable_names()<\/code> to get all variable names,\nand use <code>get_variable_value(name)<\/code> to get the variable value by name.<\/p>\n\n<p>Please use your code like:<\/p>\n\n<pre><code>estimator = tf.estimator.Estimator(...)\nparams = estimator.get_variable_names()\nfor p in params:\n    print(p, estimator.get_variable_value(p).shape)\n<\/code><\/pre>\n\n<p>The more information is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_names\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_names<\/a>\nand \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_value\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_value<\/a>.<\/p>\n\n<p>Note: you must create the graph first and then will get the variables.<\/p>\n"},{"AnswerId":"57989549","CreationDate":"2019-09-18T09:32:47.123","ParentId":null,"OwnerUserId":"2247040","Title":null,"Body":"<p>As mentioned above, you should use:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_names\" rel=\"nofollow noreferrer\"><code>tf.estimator.Estimator.get_variable_names()<\/code><\/a> in order to get all estimator variables<\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#get_variable_value\" rel=\"nofollow noreferrer\"><code>tf.estimator.Estimator.get_variable_value(name)<\/code><\/a> in order to get variables values<\/li>\n<\/ul>\n\n<p>Once you have the variables, you can use one of the following ways in order to get the total number of estimator parameters.<\/p>\n\n<ul>\n<li><p>Either multiply the shape dims of each one of the variables with <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.prod.html\" rel=\"nofollow noreferrer\"><code>numpy.prod<\/code><\/a> and then sum it:<\/p>\n\n<p><code>sum([np.prod(est.get_variable_value(var).shape) for var in est.get_variable_names()])<\/code><\/p><\/li>\n<li><p>Or sum the sizes of the variables with <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ndarray.size.html\" rel=\"nofollow noreferrer\"><code>numpy.ndarray.size<\/code><\/a> and then sum it:<\/p>\n\n<p><code>sum([est.get_variable_value(var).size for var in est.get_variable_names()])<\/code><\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53149386,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras-layer><index-error>","CreationDate":"2018-11-05T06:30:11.810","AcceptedAnswerId":null,"OwnerUserId":10606272.0,"Title":"IndexError: list index out of range in model.fit()","Body":"<p>I am new in using tensorflow. I am trying to train my network with images of shape (16*16). I have divided 3 grayscale images of 512*512 into 16*16 and appended all. so i have 3072*16*16. while training I am getting error. I am using jupyter notebook.Can anyone please help me? <\/p>\n\n<p>Here is the code<\/p>\n\n<pre><code>import tensorflow as tf \nimport numpy as np\nfrom numpy import newaxis\nimport glob\nimport os\nfrom PIL import Image,ImageOps\nimport random\nfrom os.path import join\nimport matplotlib.pyplot as plt\nfrom tensorflow import keras\nTRAIN_PATH = 'dataset\/2\/*.jpg'\nLOGS_Path = \"dataset\/logs\/\"\nCHECKPOINTS_PATH = 'dataset\/checkpoints\/'\nBETA = .75\nEXP_NAME = f\"beta_{BETA}\"\n\nfiles_list = glob.glob(join(TRAIN_PATH))\nleng=len(files_list)\nnew_cover = []\nfor i in range(leng):\n    img_cover_path = files_list[i]   \n    for j in range (0,512,16):\n        for k in range (0,512,16):\n        img_cover = Image.open(img_cover_path)\n        area=(k,j,k+16,j+16)\n        img_cover1=img_cover.crop(area)\n        img_cover1 = np.array(ImageOps.fit(img_cover1(16,16)),dtype=np.float32)\n        img_cover1 \/= 255.\n        n1.append(img_cover1)\n\n\n    new_cover.append(n1)\n\n\nnew_cover = np.array(new_cover) \nnew_cover1=np.swapaxes(new_cover, 1,3) \n\ntf.reset_default_graph()\nmodel=keras.Sequential()\n\n#1st\nmodel.add(keras.layers.Conv2D(64, (3, 3), strides=1,padding='SAME', input_shape = (16, 16, 3072))) #number of filters,shape of filter,input image size,activation function\nmodel.add(keras.layers.BatchNormalization())\nmodel.add(keras.layers.Activation(activation='relu'))\n#2\nmodel.add(keras.layers.Conv2D(64, (3, 3),strides=1,padding='SAME')) #number of filters,shape of filter,input image size,activation function\nmodel.add(keras.layers.BatchNormalization())\nmodel.add(keras.layers.Activation(activation='relu'))\n#3\nmodel.add(keras.layers.Conv2D(64, (3, 3),strides=1,padding='SAME')) #number of filters,shape of filter,input image size,activation function\nmodel.add(keras.layers.BatchNormalization())\nmodel.add(keras.layers.Activation(activation='relu'))\n#4\nmodel.add(keras.layers.Conv2D(64, (3, 3),strides=1,padding='SAME')) #number of filters,shape of filter,input image size,activation function\nmodel.add(keras.layers.BatchNormalization())\nmodel.add(keras.layers.Activation(activation='relu'))\n#message\n\n#compiling\nmodel.compile(optimizer = tf.train.AdamOptimizer(0.001),loss='mse', metrics = ['accuracy'])\nmodel.summary()\n# Store training stats\nmodel.fit(x=new_cover1,y=None, batch_size=32, epochs=1, verbose=1, callbacks=None, validation_split=0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)\n<\/code><\/pre>\n\n<p>and it is giving error:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d (Conv2D)              (None, 16, 16, 64)        1769536   \n_________________________________________________________________\nbatch_normalization (BatchNo (None, 16, 16, 64)        256       \n_________________________________________________________________\nactivation (Activation)      (None, 16, 16, 64)        0         \n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 16, 16, 64)        36928     \n_________________________________________________________________\nbatch_normalization_1 (Batch (None, 16, 16, 64)        256       \n_________________________________________________________________\nactivation_1 (Activation)    (None, 16, 16, 64)        0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 16, 16, 64)        36928     \n_________________________________________________________________\nbatch_normalization_2 (Batch (None, 16, 16, 64)        256       \n_________________________________________________________________\nactivation_2 (Activation)    (None, 16, 16, 64)        0         \n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 16, 16, 64)        36928     \n_________________________________________________________________\nbatch_normalization_3 (Batch (None, 16, 16, 64)        256       \n_________________________________________________________________\nactivation_3 (Activation)    (None, 16, 16, 64)        0         \n=================================================================\nTotal params: 1,881,344\nTrainable params: 1,880,832\nNon-trainable params: 512\n_________________________________________________________________\n\n---------------------------------------------------------------------------\nIndexError                                Traceback (most recent call last)\n&lt;ipython-input-20-49da746cee1b&gt; in &lt;module&gt;()\n     24 model.summary()\n     25 # Store training stats\n---&gt; 26 model.fit(x=new_cover1,y=None, batch_size=32, epochs=1, verbose=1, callbacks=None, validation_split=0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)\n     27 \n     28 #return model\n\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, max_queue_size, workers, use_multiprocessing, **kwargs)\n   1654           initial_epoch=initial_epoch,\n   1655           steps_per_epoch=steps_per_epoch,\n-&gt; 1656           validation_steps=validation_steps)\n   1657 \n   1658   def evaluate(self,\n\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py in fit_loop(model, inputs, targets, sample_weights, batch_size, epochs, verbose, callbacks, val_inputs, val_targets, val_sample_weights, shuffle, initial_epoch, steps_per_epoch, validation_steps)\n    135   indices_for_conversion_to_dense = []\n    136   for i in range(len(feed)):\n--&gt; 137     if issparse is not None and issparse(ins[i]) and not K.is_sparse(feed[i]):\n    138       indices_for_conversion_to_dense.append(i)\n    139 \n\nIndexError: list index out of range\n<\/code><\/pre>\n","answers":[{"AnswerId":"53153346","CreationDate":"2018-11-05T11:17:54.657","ParentId":null,"OwnerUserId":"7470786","Title":null,"Body":"<p>After my research, it became clearer that this is an already know issue and the official solution has not been released yet. Although, still there are some suggestions claimed to work.<\/p>\n\n<p>It is suggested to update to the update nightly build version (<code>pip install tf-nightly<\/code> or <code>pip install tf-nightly-gpu<\/code>) <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/21894#issuecomment-418552609\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/21894#issuecomment-418552609<\/a><\/p>\n"}]}
+{"QuestionId":53149435,"AnswerCount":0,"Tags":"<docker><tensorflow><azure-devops><continuous-deployment>","CreationDate":"2018-11-05T06:34:58.953","AcceptedAnswerId":null,"OwnerUserId":10287469.0,"Title":"Docker TensorFlow project","Body":"<p>I am struggling to find a solution to deploy a recommendation system using containers. <\/p>\n\n<p><strong>Scenario<\/strong> -- I have built a TensorFlow model and a Django API server on top of that. \n- The model is saved on disk. \n- APi requests are related to new users, new items and new user\/item interactions.\n - All these API requests load the model from disk, re-train the model with new information and return the updated predictions.<\/p>\n\n<p><strong>Problem<\/strong> -- I want to containerize this thing to run on Azure App Services under Web App for Containers. The problem i am facing is that how can i have persistent volumes along container to keep the saved model there?<\/p>\n\n<p>TO my knowledge, i have two options - <\/p>\n\n<ul>\n<li>Use Azure Blob Storage to store all the model files - But not sure if this will work with TensorFlow.<\/li>\n<li>Use the container volumes as is. - But i am afraid if do a release pipeline to destroy the container and push the new container to production, it will delete all the model files and the volumes will be re-created.<\/li>\n<\/ul>\n\n<p>I am pretty lost in this space right now with no clear idea how to approach a good option.<\/p>\n","answers":[]}
+{"QuestionId":53149907,"AnswerCount":0,"Tags":"<python><flask><keras><uwsgi>","CreationDate":"2018-11-05T07:20:17.603","AcceptedAnswerId":null,"OwnerUserId":10606483.0,"Title":"Flask + uwsgi + keras error \" DAMN ! worker 1 \uff08pid: 5240\uff09died, killed by signal 11:\uff08trying respawn\u2026\u2026\uff09\"","Body":"<p>I encountered a problem when I run <code>flask uwsgi<\/code> to call keras (using tensorflow backend) object detection API, an error \u201cDAMN ! worker 1 \uff08pid: 5240\uff09died, killed by signal 11:\uff08trying respawn\u2026\u2026\uff09\u201d is reported. Then I try to use only one thread, but it doesn't work. Instead, another error occurs which is \" !!!uWSGI process 347 got Segmentation Fault!!!\". Could anyone give me some help? Thanks very much<\/p>\n\n<p>The following is my configuration file:<\/p>\n\n<pre><code>[uwsgi]\nmodule = img_recognize_demo_3:app\nprotocol = http\nmaster = true\n\nprocessess = 2\nthreads = 2\nchdir = \/media\/xiu\/\u65b0\u52a0\u5377\/FlasklrProj\n\nthunder-lock=true\n\nhttp-socket = 0.0.0.0:5000\nlogto = \/media\/xiu\/\u65b0\u52a0\u5377\/FlasklrProj\/img_demo.log\nchmod-socket = 666\nvacuum = true\n<\/code><\/pre>\n\n<p>And this is the error in my log file:<\/p>\n\n<pre><code>*** Operational MODE: threaded ***\nUsing TensorFlow backend.\n\u8bbe\u7f6etensorflow\u540e\u53f0\uff1a \n2018-11-05 15:39:05.848219: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2018-11-05 15:39:05.848850: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1524] Found device 0 with properties: \nname: Ellesmere [Radeon Pro WX 7100]\nAMDGPU ISA: gfx803\nmemoryClockRate (GHz) 1.243\npciBusID 0000:02:00.0\nTotal memory: 8.00GiB\nFree memory: 7.75GiB\n2018-11-05 15:39:05.848918: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1635] Adding visible gpu devices: 0\n2018-11-05 15:39:05.848975: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1044] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-05 15:39:05.849021: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1050]      0 \n2018-11-05 15:39:05.849068: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1063] 0:   N \n2018-11-05 15:39:05.849139: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1183] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 7539 MB memory) -&gt; physical GPU (device: 0, name: Ellesmere [Radeon Pro WX 7100], pci bus id: 0000:02:00.0)\nWSGI app 0 (mountpoint='') ready in 9 seconds on interpreter 0x559665933a40 pid: 10593 (default app)\n*** uWSGI is running in multiple interpreter mode ***\nspawned uWSGI master process (pid: 10593)\nspawned uWSGI worker 1 (pid: 10662, cores: 2)\n[pid: 10662|app: 0|req: 1\/1] 127.0.0.1 () {32 vars in 817 bytes} [Mon Nov  5 15:39:14 2018] GET \/demo\/config?data_dir=\/media\/xiu\/%E6%96%B0%E5%8A%A0%E5%8D%B7\/upload\/testImg&amp;save_dir=\/media\/xiu\/%E6%96%B0%E5%8A%A0%E5%8D%B7\/upload\/testout2&amp;if_object=1 =&gt; generated 237 bytes in 2 msecs (HTTP\/1.1 200) 2 headers in 72 bytes (1 switches on core 0)\n\u5f00\u59cb\u8c03\u7528\u7269\u4f53\u68c0\u6d4b\u51fd\u6570\n\u8bfb\u53d6\u5f85\u68c0\u6d4b\u56fe\u7247\uff0c \u6267\u884c\u6a21\u578b\u9884\u6d4b \/media\/xiu\/\u65b0\u52a0\u5377\/upload\/testImg\/69c732e4gy1fuw8eljyssj20u00u0gn3.jpg\nDAMN ! worker 1 (pid: 10662) died, killed by signal 11 :( trying respawn ...\nRespawned uWSGI worker 1 (new pid: 10739)\n[deadlock-detector] a process holding a robust mutex died. recovering...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53150047,"AnswerCount":1,"Tags":"<tensorflow><compilation>","CreationDate":"2018-11-05T07:30:18.970","AcceptedAnswerId":53165437.0,"OwnerUserId":10207367.0,"Title":"tensorflow lite add custom op","Body":"<p>I install tensorflow from source code. As I need to add some OPs in tensorflow lite model, I followed the steps of the tensorflow offical instructions. But in the end, I grep my op name from the libtensorflow-lite.a, and nothing was found. When use the model in my c++ demo, the error is \"Didn't find custom op for name 'MatmulCalc' with version 1\", however the name can be greped out  in libtensorflow_cc.so after compiling. And the model transfering process is ok.<\/p>\n\n<p>So my question is \nhow to add new OP in lite, maybe the official steps is not complete.<a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/extend\/op<\/a><\/p>\n","answers":[{"AnswerId":"53165437","CreationDate":"2018-11-06T03:40:25.460","ParentId":null,"OwnerUserId":"10146304","Title":null,"Body":"<p>The link you shared is for creating custom Op in TensorFlow. If you are trying to add a custom Op in TensorFlow Lite, the guide is available here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/g3doc\/custom_operators.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/g3doc\/custom_operators.md<\/a><\/p>\n"}]}
+{"QuestionId":53150169,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-05T07:41:47.307","AcceptedAnswerId":null,"OwnerUserId":9697833.0,"Title":"tensorflow nested training, how to implement it","Body":"<p>I meet a problem in my experiments.\n Briefly speaking, I have a series of tensor computation flow, such as <code>a =&gt; b =&gt; c =&gt; loss =&gt; train_op<\/code>, this is a typical scenario of tensorflow usage. However, my own problem is that, when doing <code>b =&gt; c<\/code>, i need to training a extra network for several steps, that's \n<code>a=&gt;b=&gt; training another network =&gt;c=&gt;loss=&gt;train_op<\/code>, hence, I dnt know how to implement this?\nSomeone hit me that I can put the extra training network part in <code>tf.py_func<\/code> , but I still feel confused.<\/p>\n","answers":[]}
+{"QuestionId":53150235,"AnswerCount":0,"Tags":"<python><neural-network><keras><data-science>","CreationDate":"2018-11-05T07:47:24.090","AcceptedAnswerId":null,"OwnerUserId":9960182.0,"Title":"how to create specificity custom metric for Keras Neural Nets","Body":"<p>I found this function here \n<a href=\"https:\/\/stackoverflow.com\/questions\/43547402\/how-to-calculate-f1-macro-in-keras\">How to calculate F1 Macro in Keras?<\/a> but i am not sure how i can write the same way for specificity? I am using tensorflow backend for keras.<\/p>\n\n<pre><code>def recall(y_true, y_pred):\n    \"\"\"Recall metric.\n\n    Only computes a batch-wise average of recall.\n\n    Computes the recall, a metric for multi-label classification of\n    how many relevant items are selected.\n    \"\"\"\n    true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))\n    possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))\n    recall = true_positives \/ (possible_positives + K.epsilon())\n    return recall\n<\/code><\/pre>\n\n<p>I tried this solution but it gives error,        <\/p>\n\n<pre><code>def compute_binary_specificity(y_pred, y_true):\n    \"\"\"Compute the confusion matrix for a set of predictions.\n    Returns\n    -------\n    out : the specificity\n    \"\"\"\n    TN = np.logical_and(K.eval(y_true) == 0, K.eval(y_pred) == 0)\n    FP = np.logical_and(K.eval(y_true) == 0, K.eval(y_pred) == 1)\n    # as Keras Tensors\n    TN = K.sum(K.variable(TN))\n    FP = K.sum(K.variable(FP))\n    specificity = TN \/ (TN + FP + K.epsilon())\n    return specificity\n<\/code><\/pre>\n\n<p>Error: InvalidArgumentError: You must feed a value for placeholder tensor 'dense_95_input' with dtype float and shape [?,140]\n     [[Node: dense_95_input = Placeholderdtype=DT_FLOAT, shape=[?,140], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n\n<p>and points here<br>\n    ---> TN = np.logical_and(K.eval(y_true) == 0, K.eval(y_pred) == 0)<\/p>\n","answers":[]}
+{"QuestionId":53150430,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-05T08:03:50.333","AcceptedAnswerId":null,"OwnerUserId":7467476.0,"Title":"Deep learning model to clean documents","Body":"<p>I am trying to build a Convolutional Autoencoder which can remove pen marks such as circles, underlines, etc from official documents.<\/p>\n\n<p>I have the original, clean soft copy of the document and the xeroxed copy with the pen marks. <\/p>\n\n<p>The xeroxed copy would obviously not completely match up with the original doc and it would be slightly skewed or shifted, etc while scanning.<\/p>\n\n<p>Additionally, since the doc size is huge (2360, 1650), I had to split the image into 4 halves of size (587, 412) to input into my model.<\/p>\n\n<p>What I would like to know is would the above issues cause any problems while training and is there any way to rectify it?<\/p>\n\n<p>Any help would be highly appreciated.<\/p>\n\n<p>Thanks <\/p>\n\n<p>EDIT:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GHrQn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GHrQn.png\" alt=\"Original Image\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rOrdr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rOrdr.png\" alt=\"Noisy Image\"><\/a><\/p>\n\n<p>As you can see (hopefully!!), the noisy image is slightly different because of skewness or translation while xeroxing.<\/p>\n","answers":[{"AnswerId":"53150518","CreationDate":"2018-11-05T08:11:09.600","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I don't think it will cause any problems. But, if it does, you can always load the image, resize it to a desired shape and then input it into the model.  <\/p>\n"}]}
+{"QuestionId":53150642,"AnswerCount":0,"Tags":"<python><tensorflow><cluster-computing><nvidia><slurm>","CreationDate":"2018-11-05T08:20:33.850","AcceptedAnswerId":null,"OwnerUserId":4564882.0,"Title":"Tensorflow run on cluster (40k cores, 500TFLOP\/s and 40k of RAM ) vs NVIDIA GTX 1080","Body":"<p>I have the opportunity to run my Tensorflow training on a cluster computer with slurm workload manager (the cluster contains nearly 400000 cores, 40000 GB of RAM, Perfomance is Rmax=500 TFlop\/s and Rpeak=1000 TFlop\/s, AMD GPUs). The problem is that I will use the CPU version of Tensorflow, I don't know how to use AMD for training in Tensorflow. <\/p>\n\n<p>I would like to how much faster is to run may training on this cluster (Tensorflow-CPU) versus running it on my laptop equipped with NVIDIA GTX 1080 (using Tensorlfow-GPU)<\/p>\n","answers":[]}
+{"QuestionId":53151209,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm><attention-model>","CreationDate":"2018-11-05T09:03:40.167","AcceptedAnswerId":null,"OwnerUserId":1115237.0,"Title":"Keras - Add attention mechanism to an LSTM model","Body":"<p>With the following code:<\/p>\n\n<pre><code>model = Sequential()\n\nnum_features = data.shape[2]\nnum_samples = data.shape[1]\n\nmodel.add(\n    LSTM(16, batch_input_shape=(None, num_samples, num_features), return_sequences=True, activation='tanh'))\nmodel.add(PReLU())\nmodel.add(Dropout(0.5))\nmodel.add(LSTM(8, return_sequences=True, activation='tanh'))\nmodel.add(Dropout(0.1))\nmodel.add(PReLU())\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>I'm trying to understand how can I add an attention mechanism before the first LSTM layer. I've found the following GitHub: <a href=\"https:\/\/github.com\/philipperemy\/keras-attention-mechanism\" rel=\"nofollow noreferrer\">keras-attention-mechanism by Philippe R\u00e9my<\/a> but couldn't figure out how exactly to use it with my code.<\/p>\n\n<p>I would like to visualize the attention mechanism and see what are the features that the model focus on. <\/p>\n\n<p>Any help would be appreciated, especially a code modification. Thanks :)<\/p>\n","answers":[{"AnswerId":"53151599","CreationDate":"2018-11-05T09:30:15.673","ParentId":null,"OwnerUserId":"3192021","Title":null,"Body":"<p>You may find an example of how to use a LSTM with an activation mechanism in Keras in this gist<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/mbollmann\/ccc735366221e4dba9f89d2aab86da1e\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/mbollmann\/ccc735366221e4dba9f89d2aab86da1e<\/a> <\/p>\n\n<p>And in the following answer on SO:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42918446\/how-to-add-an-attention-mechanism-in-keras?answertab=votes#tab-top\">How to add an attention mechanism in keras?<\/a> <\/p>\n\n<p>And to visualize your activations you can use the following repository <a href=\"https:\/\/github.com\/philipperemy\/keras-activations\" rel=\"nofollow noreferrer\">https:\/\/github.com\/philipperemy\/keras-activations<\/a> <\/p>\n\n<p>Hope that helped :) <\/p>\n"}]}
+{"QuestionId":53151522,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T09:25:23.447","AcceptedAnswerId":null,"OwnerUserId":8106257.0,"Title":"could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873","Body":"<p>I was try to copy a code which implement adaptive gradient<\/p>\n\n<p>error comes from this code<\/p>\n\n<pre><code>    for i, n_hidden in enumerate(n_hiddens):\n    if i == 0:\n        input = x\n        input_dim = n_in\n    else:\n        input = output\n        input_dim = n_hiddens[i - 1]\n\n    W = weight_variable([input_dim, n_hidden])\n    b = bias_variable([n_hidden])\n\n    h = tf.nn.relu(tf.matmul(input, W) + b)\n    output = tf.nn.dropout(h, keep_prob)\n<\/code><\/pre>\n\n<p>at <\/p>\n\n<pre><code>h = tf.nn.relu(tf.matmul(input, w) + b)\n<\/code><\/pre>\n\n<p>and the interpreter said<\/p>\n\n<pre><code>AbortedError (see above for traceback): Operation received an exception:Status: 3, message: could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873\n [[{{node Relu}} = _MklRelu[T=DT_FLOAT, _kernel=\"MklOp\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](add, DMT\/_0)]]\n<\/code><\/pre>\n\n<p>there's other errors too, but error occur same like that.<\/p>\n\n<p>I was really good at copying this code, so I think there's nothing special mistake at code but I can't confirm it.\nI think that something wrong at 'relu' function which comes with tensorflow, but I don't know what to do.<\/p>\n\n<hr>\n\n<p>here is my full code, if you need<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom sklearn import datasets\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.utils import shuffle\nimport matplotlib.pyplot as plt\n\nnp.random.seed(0)\ntf.set_random_seed(1234)\n\ndef inference(x, keep_prob, n_in, n_hiddens, n_out):\n    def weight_variable(shape):\n        initial = np.sqrt(2.0 \/ shape[0]) * tf.truncated_normal(shape)\n        return tf.Variable(initial)\n\n    def bias_variable(shape):\n        initial = tf.zeros(shape)\n        return tf.Variable(initial)\n\n    for i, n_hidden in enumerate(n_hiddens):\n        if i == 0:\n            input = x\n            input_dim = n_in\n        else:\n            input = output\n            input_dim = n_hiddens[i - 1]\n\n        W = weight_variable([input_dim, n_hidden])\n        b = bias_variable([n_hidden])\n\n        h = tf.nn.relu(tf.matmul(input, W) + b)\n        output = tf.nn.dropout(h, keep_prob)\n\n    W_out = weight_variable([n_hiddens[-1], n_out])\n    b_out = bias_variable([n_out])\n    y = tf.nn.softmax(tf.matmul(output, W_out) + b_out)\n    return y\n\ndef loss(y, t):\n    cross_entropy = tf.reduce_mean(-tf.reduce_sum(t * tf.log(tf.clip_by_value(y, 1e-10, 1.0)), reduction_indices=[1]))\n    return cross_entropy\n\ndef training(loss):\n    optimizer = tf.train.AdagradOptimizer(0.01)\n    train_step = optimizer.minimize(loss)\n    return train_step\n\ndef accuracy(y, t):\n    correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(t, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n    return accuracy\n\nif __name__ == '__main__':\n    mnist = datasets.fetch_mldata('MNIST original')\n\n    n = len(mnist.data)\n    N = 30000\n    N_train = 20000\n    N_validation = 4000\n    indices = np.random.permutation(range(n))[:N]\n\n    X = mnist.data[indices]\n    X = X \/ 255.0\n    X = X - X.mean(axis=1).reshape(len(X), 1)\n    y = mnist.target[indices]\n    Y = np.eye(10)[y.astype(int)]\n\n    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, train_size=N_train)\n\n    X_train, X_validation, Y_train, Y_validation = train_test_split(X_train, Y_train, test_size=N_validation)\n\n    n_in = len(X[0])\n    n_hiddens = [200, 200, 200]\n    n_out = len(Y[0])\n    p_keep = 0.5\n\n    x = tf.placeholder(tf.float32, shape=[None, n_in])\n    t = tf.placeholder(tf.float32, shape=[None, n_out])\n    keep_prob = tf.placeholder(tf.float32)\n\n    y = inference(x, keep_prob, n_in=n_in, n_hiddens=n_hiddens, n_out=n_out)\n    loss = loss(y, t)\n    train_step = training(loss)\n\n    accuracy = accuracy(y, t)\n\n    history = {'val_loss' : [], 'val_acc' : []}\n\n    epochs = 50\n    batch_size = 200\n\n    init = tf.global_variables_initializer()\n    sess = tf.Session()\n    sess.run(init)\n\n    n_batches = N_train \/\/ batch_size\n\n    for epoch in range(epochs):\n        X_, Y_ = shuffle(X_train, Y_train)\n\n        for i in range(n_batches):\n            start = i * batch_size\n            end = start + batch_size\n\n            sess.run(train_step, feed_dict = {x : X_[start:end], t : Y_[start:end], keep_prob : p_keep})\n\n        val_loss = loss.eval(session=sess, feed_dict = {x : X_validation, t : Y_validation, keep_prob : 1.0})\n        val_acc = accuracy.eval(session = sess, feed_dict = {x : X_validation, t : Y_validation, keep_prob : 1.0})\n\n        history['val_loss'].append(val_loss)\n        history['val_acc'].append(val_acc)\n\n        print('epoch : ', epoch, ' validation loss : ', val_loss, ' validation accuracy : ', val_acc)\n\n    plt.rc('font', family = 'serif')\n    fig = plt.figure()\n    plt.plot(range(epochs), history['val_loss'], label = 'loss', color = 'black')\n    plt.xlabel('epochs')\n    plt.show()\n\n    accuracy_rate = accuracy.eval(session = sess, feed_dict = {x : X_test, t : Y_test, keep_prob : 1.0})\n    print('accuracy : ', accuracy_rate)\n<\/code><\/pre>\n\n<p>and here is my whole error log<\/p>\n\n<pre><code> \/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/model_selection\/_split.py:2069: FutureWarning: From version 0.21, test_size will always complement train_size unless both are specified.\n  FutureWarning)\n2018-11-05 18:11:21.403663: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2018-11-05 18:11:21.406748: I tensorflow\/core\/common_runtime\/process_util.cc:69] Creating new thread pool with default inter op setting: 2. Tune using inter_op_parallelism_threads for best performance.\n2018-11-05 18:11:22.406667: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at mkl_relu_op.cc:876 : Aborted: Operation received an exception:Status: 3, message: could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873\nTraceback (most recent call last):\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1292, in _do_call\n    return fn(*args)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1277, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1367, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.AbortedError: Operation received an exception:Status: 3, message: could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873\n     [[{{node Relu}} = _MklRelu[T=DT_FLOAT, _kernel=\"MklOp\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](add, DMT\/_0)]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/shin\/pythonClass\/AdaGrad.py\", line 105, in &lt;module&gt;\n    sess.run(train_step, feed_dict = {x : X_[start:end], t : Y_[start:end], keep_prob : p_keep})\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 887, in run\n    run_metadata_ptr)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1110, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1286, in _do_run\n    run_metadata)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1308, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.AbortedError: Operation received an exception:Status: 3, message: could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873\n     [[{{node Relu}} = _MklRelu[T=DT_FLOAT, _kernel=\"MklOp\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](add, DMT\/_0)]]\n\nCaused by op 'Relu', defined at:\n  File \"\/home\/shin\/pythonClass\/AdaGrad.py\", line 81, in &lt;module&gt;\n    y = inference(x, keep_prob, n_in=n_in, n_hiddens=n_hiddens, n_out=n_out)\n  File \"\/home\/shin\/pythonClass\/AdaGrad.py\", line 31, in inference\n    h = tf.nn.relu(tf.matmul(input, W) + b)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_nn_ops.py\", line 6780, in relu\n    \"Relu\", features=features, name=name)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3272, in create_op\n    op_def=op_def)\n  File \"\/home\/shin\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1768, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nAbortedError (see above for traceback): Operation received an exception:Status: 3, message: could not initialize a memory primitive descriptor, in file tensorflow\/core\/kernels\/mkl_relu_op.cc:873\n     [[{{node Relu}} = _MklRelu[T=DT_FLOAT, _kernel=\"MklOp\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](add, DMT\/_0)]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53152149,"AnswerCount":1,"Tags":"<python><machine-learning><keras><nlp><embedding>","CreationDate":"2018-11-05T10:06:04.637","AcceptedAnswerId":53152566.0,"OwnerUserId":7908461.0,"Title":"Can a matrix be given as input to Keras's embedding layer?","Body":"<p>I am using Keras to capture semantic information for a dataset. And I already tokenize the data to integer vectors. It has a form like this:<\/p>\n\n<pre><code>texts=[[1,2,3,2,1],\n       [2,3,4,2,2],\n       [3,33,2,1,3]]\n\nlabels=[1,0,1]\n<\/code><\/pre>\n\n<p>And the labels only contains 0 or 1, each list contain one label.\nI want to use Keras's embedding layer to embed this. But the examples on the Internet only contain a list:<\/p>\n\n<pre><code>texts=[1,2,3,4,2,1]\n<\/code><\/pre>\n\n<p>I am wondering can I input a matrix to the embedding layer?<\/p>\n","answers":[{"AnswerId":"53152566","CreationDate":"2018-11-05T10:30:20.800","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Each list in the <code>texts<\/code> list is a training sample and there is a corresponding label for each of them in the <code>labels<\/code> list. Therefore, each training sample is just a vector of integers (i.e. word indices) which you can easily feed to an Embedding layer:<\/p>\n\n<pre><code>inp = Input(shape=(num_words_per_sample,))\nx = Embedding(vocab_size, emb_dim)(inp)\n<\/code><\/pre>\n\n<p>Note that you may need to convert the training data and labels to a numpy array (if it is not already):<\/p>\n\n<pre><code>import numpy as np\ntexts = np.array(texts)\nprint(texts.shape)  # (3,5)  &lt;--- three samples each containing 5 words\n\nlabels = np.array(labels)\nprint(labels.shape) # (3,)   &lt;--- three labels, one for each sample\n<\/code><\/pre>\n"}]}
+{"QuestionId":53152165,"AnswerCount":1,"Tags":"<python><tensorflow><keras><computer-vision>","CreationDate":"2018-11-05T10:06:53.107","AcceptedAnswerId":null,"OwnerUserId":8113093.0,"Title":"Keras model evaluate() and predict_classes() give contradicting results","Body":"<p>I have been working for a while with VGG16 from Keras and after finetuning on my own 4 classes dataset, the training seems to be going well, the accuracy on training and validation sets increased and at the end, even the evaluation using <code>model.evaluate()<\/code> gave a good accuracy on testing data. I tried <code>confusion_matrix<\/code> and it also showed good classification results. In order to confirm, I tried to predict on images from my dataset using the following:<\/p>\n\n<pre><code>im = cv2.resize(frame,(image_size, image_size), interpolation = cv2.INTER_AREA)\n\n#convert the image pixels to a numpy array\nframer = img_to_array(im)             \nimage = framer.reshape((1, framer.shape[0], framer.shape[1], framer.shape[2]))\n\n# prepare the image for the VGG model\nimage = preprocess_input(image)  \n\nlabel = FLOW1_model.predict_classes(image, verbose=0)\n<\/code><\/pre>\n\n<p>I loaded flow frames for only 1 class (class 0) but about 30% only of the pictures were classified correctly, in another class (class 1) it was 50% correct classification. Given that the accuracy went above 85% when I trained, and also the confusion matrix showing very good results, I find this very abnormal. I looked at all possible causes online but couldn't find what went wrong... Is there a known issue that explains this difference in the results ?<\/p>\n\n<p>Edit: <\/p>\n\n<p>I basically load the Keras vgg without top, add my own classifier and compile the model like this:<\/p>\n\n<pre><code># Compile the model method 2\nsgd = SGD(lr=0.00001, decay = 1e-6, momentum=0.9, nesterov=True)\nmodel.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53164435","CreationDate":"2018-11-06T01:02:21.793","ParentId":null,"OwnerUserId":"9742432","Title":null,"Body":"<p>As far as I know, you cannot use the predict_classes() method in case of multi-class output.\nHave you already tried with:<\/p>\n\n<pre><code>FLOW1_model.predict(image, verbose=0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53152243,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-05T10:10:59.757","AcceptedAnswerId":null,"OwnerUserId":6329284.0,"Title":"Keras metric based on output of an intermediate layer","Body":"<p>Problem:\nI want to monitor my model better during training. Because in some cases the loss suddenly turn to nan during training, and I want to know what the model is doing when this happens.\nBesides that, I want to see if a certain layer obeys a specific condition (rows and columns should sum to one).<\/p>\n\n<p>Approach:\nDefining a custom metric wont help, since this one will only work on <code>y_pred<\/code> and <code>y_true<\/code>. Maybe there is some complex solution with building a model within a model, and trying to somehow calculate a metric on the output of the intermediate-model layer. But that feels a bit too complex.<\/p>\n\n<p>Solution:\nThe only thing I could think of otherwise is to switch to Tensorflow itself, so that I have more control over the training process. Any other ideas?<\/p>\n","answers":[{"AnswerId":"55025604","CreationDate":"2019-03-06T14:36:19.347","ParentId":null,"OwnerUserId":"7405530","Title":null,"Body":"<p>There are several ways to do this without the need to construct a callback, depending on how you add your losses.<\/p>\n\n<p>If you add the loss with model.add_loss, you need to display it through a workaround by adding the metric after the compile step (as discussed <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9459#issuecomment-469282443\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<p>This results in something like this (specifically for a VAE, one is interested in kl_loss which depends on the intermediate layer): <\/p>\n\n<pre><code>reconstruction_loss = mse(K.flatten(inputs), K.flatten(outputs))\nkl_loss = beta*K.mean(- 0.5 * 1\/latent_dim * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1))\n\nmodel.add_loss(reconstruction_loss)\nmodel.add_loss(kl_loss)\nmodel.compile(optimizer='adam')\n\nmodel.metrics_tensors.append(kl_loss)\nmodel.metrics_names.append(\"kl_loss\")\n\nmodel.metrics_tensors.append(reconstruction_loss)\nmodel.metrics_names.append(\"mse_loss\")\n<\/code><\/pre>\n\n<p>For me this gives an output like this: <\/p>\n\n<pre><code>Epoch 1\/1\n252\/252 [==============================] - 23s 92ms\/step - loss: 0.4336 - kl_loss: 0.0823 - mse_loss: 0.3513 - val_loss: 0.2624 - val_kl_loss: 0.0436 - val_mse_loss: 0.2188\n<\/code><\/pre>\n\n<p>If you don't use model.add_loss but pass your losses directly in the compiler, than you need to define a custom metric (similar to <a href=\"https:\/\/stackoverflow.com\/questions\/46858016\/keras-custom-loss-function-to-pass-arguments-other-than-y-true-and-y-pred\">custom loss metric<\/a>) and pass the metric to the compile step. In the case above: <\/p>\n\n<pre><code>def customMetric(kl_loss):\n\n    def klLoss(y_true,y_pred):    \n        return kl_loss\n\n    return klLoss\n\nmodel.compile(..., metrics=[customMetric(kl_loss)])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53153790,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><keras-layer>","CreationDate":"2018-11-05T11:46:59.063","AcceptedAnswerId":53154110.0,"OwnerUserId":10607259.0,"Title":"Tensor' object has no attribute 'lower'","Body":"<p>I am fine tuning a MobileNet with 14 new classes. When I add new layers by:<\/p>\n\n<pre><code>x=mobile.layers[-6].output\nx=Flatten(x)\npredictions = Dense(14, activation='softmax')(x)\nmodel = Model(inputs=mobile.input, outputs=predictions)\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>'Tensor' object has no attribute 'lower'\n<\/code><\/pre>\n\n<p>Also using:<\/p>\n\n<pre><code>model.compile(Adam(lr=.0001), loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit_generator(train_batches, steps_per_epoch=18,\n                validation_data=valid_batches, validation_steps=3, epochs=60, verbose=2)\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>Error when checking target: expected dense_1 to have 4 dimensions, but got array with shape (10, 14)\n<\/code><\/pre>\n\n<p>What does <code>lower<\/code> mean? I saw other fine-tuning scripts and there were no other arguments other than the name of the model which is X in this case.<\/p>\n","answers":[{"AnswerId":"53154110","CreationDate":"2018-11-05T12:07:09.370","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The tensor must be passed to the layer when you are calling it, and not as an argument. Therefore it must be like this:<\/p>\n\n<pre><code>x = Flatten()(x)  # first the layer is constructed and then it is called on x\n<\/code><\/pre>\n\n<p>To make it more clear, it is equivalent to this:<\/p>\n\n<pre><code>flatten_layer = Flatten()  # instantiate the layer\nx = flatten_layer(x)       # call it on the given tensor\n<\/code><\/pre>\n"}]}
+{"QuestionId":53154917,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-05T12:58:03.217","AcceptedAnswerId":54402330.0,"OwnerUserId":9881203.0,"Title":"Load LSUN dataset with tensorflow","Body":"<p>recently I try to find the right way to read LSUN dataset which is in the form of lmdb. However, I do not find any useful information. I want to know how to read image data from lmdb and what the advantage is in that way. Thank you!<\/p>\n","answers":[{"AnswerId":"54402330","CreationDate":"2019-01-28T12:45:08.770","ParentId":null,"OwnerUserId":"9881203","Title":null,"Body":"<p>Finally, I use the following code to extract LUSN images from lmbd file.<\/p>\n\n<pre><code>import os\nimport lmdb\nfrom PIL import Image\nimport tempfile\n\ndef _export_mdb_images(db_path, out_dir=None, flat=True, limit=-1, size=256):\n    out_dir = out_dir\n    env = lmdb.open(\n        db_path, map_size=1099511627776,\n        max_readers=1000, readonly=True\n    )\n    count = 0\n    with env.begin(write=False) as txn:\n        cursor = txn.cursor()\n        for key, val in cursor:\n            key = str(key, 'utf-8')\n            # decide image out directory\n            if not flat:\n                image_out_dir = os.path.join(out_dir, '\/'.join(key[:6]))\n            else:\n                image_out_dir = out_dir\n\n            # create the directory if an image out directory doesn't exist\n            if not os.path.exists(image_out_dir):\n                os.makedirs(image_out_dir)\n\n            with tempfile.NamedTemporaryFile('wb') as temp:\n                temp.write(val)\n                temp.flush()\n                temp.seek(0)\n                image_out_path = os.path.join(image_out_dir, key + '.jpg')\n                Image.open(temp.name).resize((size, size)).save(image_out_path)\n            count += 1\n            if count == limit:\n                break\n            if count % 1000 == 0:\n                print('Finished', count, 'images')\n\nprint(\"start\")\ndb_path = \"path to lmbd\"\nout_dir = os.path.join(db_path, \"data\")\n_export_mdb_images(db_path, out_dir)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53156244,"AnswerCount":1,"Tags":"<python><machine-learning><keras><loss-function><mse>","CreationDate":"2018-11-05T14:22:08.570","AcceptedAnswerId":53156685.0,"OwnerUserId":10341061.0,"Title":"different mse result for training set","Body":"<p>I get different results for <code>mse<\/code>. During trainig I get 0.296 after my last training epoch and when I evaluate my model I get 0.112. Does any one know why that is so?<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(100, return_sequences=True,batch_input_shape=(batch_size,look_back,dim_x)))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(150,return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(100,return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(50,return_sequences=False))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1, activation='linear'))\n\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nhistory=model.fit(x_train_r, y_train_r, validation_data=(x_test_r, y_test_r),\\\n                  epochs=epochs, batch_size=batch_size, callbacks=[es])\n\nscore_test = model.evaluate(x_test_r, y_test_r,batch_size=batch_size)\nscore_train = model.evaluate(x_train_r, y_train_r,batch_size=batch_size)\n\nprint(\"Score Training Data:\")\nprint(score_train)\n<\/code><\/pre>\n\n<p>Batch size and everything stays the same. Does anyone knows why I get so different results for <code>mse<\/code>?<\/p>\n","answers":[{"AnswerId":"53156685","CreationDate":"2018-11-05T14:46:53.383","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The reason for the discrepancy between the training loss and the loss obtained on the training data after the training is finished, is the existence of <code>Dropout<\/code> layer in the model. That's because this layer has different behavior during training and inference time. As I have mentioned in <a href=\"https:\/\/stackoverflow.com\/a\/52760112\/2099607\">another answer<\/a>, you can make this behavior the same either by passing <code>training=True<\/code> to dropout call, or using <code>K.learning_phase()<\/code> flag and a backend function.<\/p>\n"}]}
+{"QuestionId":53156298,"AnswerCount":2,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-05T14:24:57.073","AcceptedAnswerId":null,"OwnerUserId":8804960.0,"Title":"Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #2 'weight'","Body":"<p>Firstly I have used like 'model.cuda()' to convert model and data to cuda. But it still has such a problem. I debug every layer of the model, and weights for every module has iscuda=True. So does anyone know why there is such a problem?<\/p>\n\n<p>I have two models, one is resnet50 and another one which contains the first one as backbone.<\/p>\n\n<pre><code>class FC_Resnet(nn.Module):\n    def __init__(self, model, num_classes):\n        super(FC_Resnet, self).__init__()\n\n        # feature encoding\n        self.features = nn.Sequential(\n            model.conv1,\n            model.bn1,\n            model.relu,\n            model.maxpool,\n            model.layer1,\n            model.layer2,\n            model.layer3,\n            model.layer4)\n\n        # classifier\n        num_features = model.layer4[1].conv1.in_channels\n        self.classifier = nn.Sequential(\n            nn.Conv2d(num_features, num_classes, kernel_size=1, bias=True))\n\n    def forward(self, x):\n        # children=self.features.children()\n        # for child in children:\n        #     if child.weight is not None:\n        #         print(child.weight.device)\n        x = self.features(x)\n        x = self.classifier(x)\n        return x\n\ndef fc_resnet50(num_classes=20, pre_trained=True):\n    model = FC_Resnet(models.resnet50(pre_trained), num_classes)\n\n    return model\n<\/code><\/pre>\n\n<p>And another one:<\/p>\n\n<pre><code>class PeakResponseMapping(nn.Sequential):\n    def __init__(self, *args, **kargs):\n        super(PeakResponseMapping, self).__init__(*args)\n        ...\n\n    def forward(self, input, class_threshold=0, peak_threshold=30, retrieval_cfg=None):\n        assert input.dim() == 4\n        if self.inferencing:\n            input.requires_grad_()\n\n        class_response_maps = super(PeakResponseMapping, self).forward(input)\n\n        return class_response_maps\n<\/code><\/pre>\n\n<p>And the main is very simple:<\/p>\n\n<pre><code>def main():\n    dataset = VOC(img_transform=image_transform())\n    dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)\n\n    model = peak_response_mapping(fc_resnet50(), win_size=3, sub_pixel_locating_factor=8, enable_peak_stimulation=True)\n    model=model.cuda()\n\n    for step, (b_x, b_y) in enumerate(dataloader):\n        b_x.cuda()\n        b_y.cuda()\n\n        result = model.forward(b_x)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53570255","CreationDate":"2018-12-01T11:18:51.997","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>Somewhere down in the stack trace, Torch is expecting a CPU tensor (<code>torch.FloatTensor<\/code>) but is getting a GPU \/ CUDA tensor (<code>torch.cuda.FloatTensor<\/code>).<\/p>\n\n<p>Given a tensor <code>tensor<\/code>:<\/p>\n\n<ul>\n<li><code>tensor.to('cpu')<\/code> returns the CPU version of the tensor<\/li>\n<li><code>tensor.to('cuda')<\/code> returns the CUDA version of the tensor<\/li>\n<\/ul>\n\n<p>To write hardware-agnostic code:<\/p>\n\n<pre><code>device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n<\/code><\/pre>\n\n<p>Then you can do:<\/p>\n\n<pre><code>tensor.to(device)\n<\/code><\/pre>\n\n<p>For the OP, this becomes:<\/p>\n\n<pre><code>result = model.forward(b_x.to(device))\n<\/code><\/pre>\n"},{"AnswerId":"53156544","CreationDate":"2018-11-05T14:39:13.053","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to assign <code>b_x.cuda()<\/code> back to <code>b_x<\/code>:<\/p>\n\n<pre><code>b_x = b_x.cuda()\nb_y = b_y.cuda()\n<\/code><\/pre>\n\n<p>Looking at the documentation of <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.cuda\" rel=\"nofollow noreferrer\"><code>.cuda()<\/code><\/a>:<\/p>\n\n<blockquote>\n  <p>Returns a copy of this object in CUDA memory.<\/p>\n<\/blockquote>\n\n<p>so, <code>b_x.cuda()<\/code> returns a <em>copy<\/em> of <code>b_x<\/code> and does not affect <code>b_x<\/code> in an  <em>in-place<\/em> manner.<\/p>\n"}]}
+{"QuestionId":53156985,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-05T15:05:52.497","AcceptedAnswerId":null,"OwnerUserId":8519078.0,"Title":"Cannot get above 55% TRAINING accuracy on wine quality dataset","Body":"<p>I have tried a logistic regression model as well as just making the NN a little bigger, tried to tune learning rate, excluding some useless features but I still can't get above 0.55 training accuracy. Here's is some of my code:<\/p>\n\n<p><strong>Importing files<\/strong><\/p>\n\n<pre><code>df_red = pd.read_csv('winequality-red.csv', sep=';')\ndf_white = pd.read_csv('winequality-white.csv', sep=';')\n\ndf = df_red.append(df_white, ignore_index=True)\n<\/code><\/pre>\n\n<p><strong>Excluding outliers<\/strong><\/p>\n\n<pre><code>from scipy import stats\n\ndf_new = df[(np.abs(stats.zscore(df)) &lt; 3).all(axis=1)]\n<\/code><\/pre>\n\n<p><strong>Scaling the data<\/strong><\/p>\n\n<pre><code>Y = df_new['quality']\n\ndf_new = df_new.drop(['quality'], axis=1)\n\nfrom sklearn.preprocessing import StandardScaler\n\nscaler = StandardScaler()\n\ndf_new = scaler.fit_transform(df_new)\n\nX = df_new[:, 0:11]\nY = Y.reshape(-1, 1)\n<\/code><\/pre>\n\n<p><strong>Splitting the data<\/strong><\/p>\n\n<pre><code>X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.20)\n<\/code><\/pre>\n\n<p><strong>One-hot encoding<\/strong><\/p>\n\n<pre><code>Y_train = to_categorical(Y_train)\nprint(Y_train[0:7])\nY_test = to_categorical(Y_test)\nprint(Y_test[0:7])\n<\/code><\/pre>\n\n<p><strong>Model<\/strong><\/p>\n\n<pre><code># Define our model\ndef model(X_train, Y_train, X_test, Y_test):\nglobal network_history\nmodel = Sequential()\nmodel.add(Dense(10, input_shape=(11,), activation='relu'))\nmodel.add(Dense(9, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nnetwork_history = model.fit(X_train, Y_train, batch_size=64, epochs=200, shuffle=True, validation_data=(X_test, Y_test))\n\nscore = model.evaluate(X_test, Y_test)\n\nprint(\"Test score \", score[0])\nprint(\"Test accuracy \", score[1])\n\nreturn model\n<\/code><\/pre>\n\n<p>What can I do to improve this?<\/p>\n","answers":[{"AnswerId":"53157063","CreationDate":"2018-11-05T15:10:43.253","ParentId":null,"OwnerUserId":"3523464","Title":null,"Body":"<p>The problem is, you are using real test data for validation, which is not OK! Validation data set is part of the training data.<\/p>\n\n<p>I would suggest you use:<\/p>\n\n<pre><code>validation_split = 0.1\n<\/code><\/pre>\n\n<p>to use 10% of training data for validation purposes.<\/p>\n\n<p>One of the options is that your network is overfitting. Try adding Dropout layers, this might help. Further, have you tried non-scaled data with outliers?<\/p>\n\n<p>Further, the number of epochs seems too high. Try 5 or 10 epochs for such simple data set for starters. Similarly, the batch size could be further experimented with.<\/p>\n"}]}
+{"QuestionId":53157197,"AnswerCount":1,"Tags":"<python><tensorflow><gpgpu>","CreationDate":"2018-11-05T15:19:47.380","AcceptedAnswerId":53158031.0,"OwnerUserId":5535270.0,"Title":"Why can't I run tensorflow session on CPU while one GPU device's memory is all allocated?","Body":"<p>From the tensorflow website (<a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/using_gpu<\/a>) I found the following code to manually specify the use of a CPU instead of a GPU:<\/p>\n\n<pre><code># Creates a graph.\nwith tf.device('\/cpu:0'):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n    c = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>I tried running this on my machine (with 4 GPUs) and got the following error:<\/p>\n\n<blockquote>\n  <p>2018-11-05 10:02:30.636733: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1392] Found device 0 with properties: \n  name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\n  pciBusID: 0000:18:00.0\n  totalMemory: 10.92GiB freeMemory: 10.76GiB\n  2018-11-05 10:02:30.863280: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1392] Found device 1 with properties: \n  name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582\n  pciBusID: 0000:3b:00.0\n  totalMemory: 10.92GiB freeMemory: 10.76GiB\n  2018-11-05 10:02:31.117729: E tensorflow\/core\/common_runtime\/direct_session.cc:158] Internal: failed initializing StreamExecutor for CUDA device ordinal 2: Internal: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_OUT_OF_MEMORY; total memory reported: 11721506816\n  Traceback (most recent call last):\n    File \".\/tf_test.py\", line 10, in \n      sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n    File \"...\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1566, in <strong>init<\/strong>\n      super(Session, self).<strong>init<\/strong>(target, graph, config=config)\n    File \"...\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 636, in <strong>init<\/strong>\n      self._session = tf_session.TF_NewSession(self._graph._c_graph, opts)\n  tensorflow.python.framework.errors_impl.InternalError: Failed to create session.<\/p>\n<\/blockquote>\n\n<p>It seems that when I create the session, tensorflow tries to initialize a stream executor on all devices.  Unfortunately, one of the GPUs is being used by my colleague right now.  I would hope that his full use of one GPU would not prevent me from using another device (whether GPU or CPU), but this does not seem to be the case.<\/p>\n\n<p>Does anyone know a workaround to this?  Perhaps something to add to the config?  Is this something that could be fixed in tensorflow?<\/p>\n\n<p>FYI... here is the output of \"gpustat -upc\":<\/p>\n\n<pre><code>&lt;my_hostname&gt;  Mon Nov  5 10:19:47 2018\n[0] GeForce GTX 1080 Ti | 36'C,   0 % |    10 \/ 11178 MB |\n[1] GeForce GTX 1080 Ti | 41'C,   0 % |    10 \/ 11178 MB |\n[2] GeForce GTX 1080 Ti | 38'C,   0 % | 11097 \/ 11178 MB | &lt;my_colleague&gt;:python2\/148901(11087M)\n[3] GeForce GTX 1080 Ti | 37'C,   0 % |    10 \/ 11178 MB |\n<\/code><\/pre>\n","answers":[{"AnswerId":"53158031","CreationDate":"2018-11-05T16:07:50.173","ParentId":null,"OwnerUserId":"5535270","Title":null,"Body":"<p>OK... so with the help of my colleague, I have a workable solution.  The key is, in fact, a modification to the config.  Specifically, something like this:<\/p>\n\n<p>config.gpu_options.visible_device_list = '0'<\/p>\n\n<p>This will ensure that tensorflow only sees GPU 0.<\/p>\n\n<p>In fact, I was able to run the following:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python                                                                                                                                                                                                                        \n\nimport tensorflow as tf\n\nwith tf.device('\/gpu:2'):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n    c = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.                                                                                                                                                                                   \nconfig=tf.ConfigProto(log_device_placement=True)\nconfig.gpu_options.visible_device_list = '0,1,3'\nsess = tf.Session(config=config)\n# Runs the op.                                                                                                                                                                                                                               \nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>Notice that this code actually specifies to run on GPU 2 (which you might remember is the one that is full).  This is an important point... the GPUs are renumbered according to the visible_device_list, so in the above code, when we say \"with gpu:2\", this is referring to the 3rd GPU in the list ('0,1,3'), which is actually GPU 3.  This can bite you if you try this:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python                                                                                                                                                                                                                        \n\nimport tensorflow as tf\n\nwith tf.device('\/gpu:1'):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n    c = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.                                                                                                                                                                                   \nconfig=tf.ConfigProto(log_device_placement=True)\nconfig.gpu_options.visible_device_list = '1'\nsess = tf.Session(config=config)\n# Runs the op.                                                                                                                                                                                                                               \nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>The problem is that its looking for the 2nd GPU in the list, but there is only one GPU in the visible list.  The error you will get is as follows:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'a': Operation was explicitly assigned to \/device:GPU:1 but available devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0, \/job:localhost\/replica:0\/task:0\/device:GPU:0 ]. Make sure the device specification refers to a valid device.\n       [[Node: a = Constdtype=DT_FLOAT, value=Tensor, _device=\"\/device:GPU:1\"]]<\/p>\n<\/blockquote>\n\n<p>It still seems odd to me that I must specify a GPU list when I want to run on the CPU.  I tried using an empty list and it failed, so if all 4 GPUs were in use, I would not have a workaround.  Anyone else have a better idea?<\/p>\n"}]}
+{"QuestionId":53157214,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-05T15:20:46.220","AcceptedAnswerId":null,"OwnerUserId":7844769.0,"Title":"Tensorflow: How to improve the image preprocessing speed?","Body":"<p>I am now using the tensorlfow dataset to implement the SSD image preporcessing method. The code of dataset is:<\/p>\n\n<pre><code>dataset = dataset.prefetch(buffer_size=batch_size)\n\nis_training = True if mode == 'Train' else False   \n\nif is_training:    \n  dataset = dataset.shuffle(buffer_size=shuffle_buffer)\n\n# If we are training over multiple epochs before \n#evaluating, repeat the\n# dataset for the appropriate number of epochs.\nif num_epochs is not None:\n    dataset = dataset.repeat(num_epochs)\n\nif is_training and num_gpus and examples_per_epoch:\n    total_examples = num_epochs * examples_per_epoch\n\ntotal_batches = total_examples \/\/ batch_size \/\/ \n num_gpus * num_gpus\ndataset.take(total_batches * batch_size)\n\n# Parse the raw records into images and labels. Testing \nhas shown that setting\n# num_parallel_batches &gt; 1 produces no improvement in \nthroughput, since\n# batch_size is almost always much greater than the \nnumber of CPU cores.\n\ndataset = dataset.apply(\n  tf.contrib.data.map_and_batch(\n      parse_record_fn,\n      batch_size=batch_size,\n      num_parallel_batches=1,\n      drop_remainder=True))\ndataset = \ndataset.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)\n<\/code><\/pre>\n\n<p>Where \"parse_record_fn\" includes random crop and color saturation. The \"dataset\" is conducted under CPU. I set the batch size 32 and image size 640*640*3. Experimental results shows that the GPU utils is about 34% which means it takes too much time on image preprocessing. Is there any method to improve the speed?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53157978,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><pytorch>","CreationDate":"2018-11-05T16:05:00.160","AcceptedAnswerId":53160140.0,"OwnerUserId":1072825.0,"Title":"coverting roi pooling in pytorch to nn layer","Body":"<p>I have a an mlmodel  using ROI pooling for which I am using this \n(adapted from <a href=\"https:\/\/github.com\/longcw\/faster_rcnn_pytorch\/blob\/master\/faster_rcnn\/roi_pooling\/modules\/roi_pool_py.py\" rel=\"nofollow noreferrer\">here<\/a>) (non NN layer version)<\/p>\n\n<pre><code>def forward(self, features, rois):\n        batch_size, num_channels, data_height, data_width = features.size()\n        num_rois = rois.size()[0]\n        outputs = Variable(torch.zeros(num_rois, num_channels, self.pooled_height, self.pooled_width)).cuda()\n\n        for roi_ind, roi in enumerate(rois):\n            batch_ind = int(roi[0].data[0])\n            roi_start_w, roi_start_h, roi_end_w, roi_end_h = np.round(\n                roi[1:].data.cpu().numpy() * self.spatial_scale).astype(int)\n            roi_width = max(roi_end_w - roi_start_w + 1, 1)\n            roi_height = max(roi_end_h - roi_start_h + 1, 1)\n            bin_size_w = float(roi_width) \/ float(self.pooled_width)\n            bin_size_h = float(roi_height) \/ float(self.pooled_height)\n\n            for ph in range(self.pooled_height):\n                hstart = int(np.floor(ph * bin_size_h))\n                hend = int(np.ceil((ph + 1) * bin_size_h))\n                hstart = min(data_height, max(0, hstart + roi_start_h))\n                hend = min(data_height, max(0, hend + roi_start_h))\n                for pw in range(self.pooled_width):\n                    wstart = int(np.floor(pw * bin_size_w))\n                    wend = int(np.ceil((pw + 1) * bin_size_w))\n                    wstart = min(data_width, max(0, wstart + roi_start_w))\n                    wend = min(data_width, max(0, wend + roi_start_w))\n\n                    is_empty = (hend &lt;= hstart) or(wend &lt;= wstart)\n                    if is_empty:\n                        outputs[roi_ind, :, ph, pw] = 0\n                    else:\n                        data = features[batch_ind]\n                        outputs[roi_ind, :, ph, pw] = torch.max(\n                            torch.max(data[:, hstart:hend, wstart:wend], 1)[0], 2)[0].view(-1)\n\n        return outputs\n<\/code><\/pre>\n\n<p>I want to convert the pytorch model to caffe and hence I need to convert the above to an NN layer for which I am using the below (adapted from <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/d8d378c31d2766009db400ac03f41dd837a56c2a\/fast_rcnn\/roi_pooling.py\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<pre><code>def forward(self, input, rois):\n    output = []\n    rois = rois.data.float()\n    num_rois = rois.size(0)\n\n    rois[:,1:].mul_(self.spatial_scale)\n    rois = rois.long()\n    for i in range(num_rois):\n        roi = rois[i]\n        im_idx = roi[0]\n        im = input.narrow(0, im_idx, 1)[..., roi[2]:(roi[4]+1), roi[1]:(roi[3]+1)]\n        op = nn.functional.adaptive_max_pool2d(input = im, output_size = self.size) \n        output.append(op)\n\n    return torch.cat(tuple(output), dim=0)\n<\/code><\/pre>\n\n<p>The output returned doesn't seem to match for the above methods even though they are performing the same function. I seem to hit a deadlock. Could anyone point if I am making any obvious mistake in the above ?<\/p>\n","answers":[{"AnswerId":"53160140","CreationDate":"2018-11-05T18:29:55.037","ParentId":null,"OwnerUserId":"1072825","Title":null,"Body":"<p>Found the issue - \nThe rois after multiplication with spatial scale were being rounded down and had to call round function before calling long like so <\/p>\n\n<pre><code>rois = rois.data.float()\nnum_rois = rois.size(0)\n\nrois[:,1:].mul_(self.spatial_scale)\nrois = rois.round().long() ## Check this here !!\n<\/code><\/pre>\n\n<p>Hope this helps someone!<\/p>\n"}]}
+{"QuestionId":53158340,"AnswerCount":1,"Tags":"<java><python><image><tensorflow><classification>","CreationDate":"2018-11-05T16:25:13.887","AcceptedAnswerId":53208128.0,"OwnerUserId":9850717.0,"Title":"Tensorflow dataset produced in Python has different readings in Tensorflow Java API (Label Image)","Body":"<p>Background: I am new to Tensorflow and AI and wish to try out Tensorflow in a Java based environment. Found Tensorflow has a Java API and tried it out but results don't seem to be similar to my Python implementations. I used Maven for my Java implementation.<\/p>\n\n<p>I had retrained a dataset using Tensorflow in Python by following the <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/index.html?index=..%2F..%2Findex#0\" rel=\"nofollow noreferrer\">Tensorflow for Poets<\/a> tutorial.<\/p>\n\n<p>In a Python based Tensorflow environment the dataset outputs values which are more or less correct. For example, if I have an image of a car and my classification options are car, bug,or building; it would output car.<\/p>\n\n<p>But in my Java equivalent which I tried using the Tensorflow Java API and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.4\/tensorflow\/java\/src\/main\/java\/org\/tensorflow\/examples\/LabelImage.java\" rel=\"nofollow noreferrer\">Image label demo<\/a>; the datasets I produced earlier are readable but the outputs are completely different from the Python implementation. Unlike the Python example, if I throw a car it would output it as a bug or building.<\/p>\n\n<p>I had tried the output data on several different machines with processors and gpus from different generations to see if it would affect the output; but it appears to be exclusive with the language\/API.<\/p>\n\n<ol>\n<li><p>Any ideas as to why the two implementations using the same dataset\nproduce completely different outputs?<\/p><\/li>\n<li><p>And is there any way I could get my Java version to output similar\nresults to my Python example?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53208128","CreationDate":"2018-11-08T12:50:49.020","ParentId":null,"OwnerUserId":"9850717","Title":null,"Body":"<p>Problem solved. Apparently the graph construction in the \"tensorflow for poets\" and java \"label image\" feature different dimensions when plotting\/constructing the graph hence the skewed results I received. The difference in dimensions led to completely different constructed results.<\/p>\n\n<ol>\n<li><p>Output is different due to graph mapping issue pertaining to image itself. The tutorial dimensions and java sample dimensions were different.<\/p><\/li>\n<li><p>Readjust dimensions to fit. Both  require a common or similar dimension reading for graph reconstruction.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53159232,"AnswerCount":1,"Tags":"<python><python-3.x><opencv><keras>","CreationDate":"2018-11-05T17:23:35.813","AcceptedAnswerId":null,"OwnerUserId":5617580.0,"Title":"Cannot reshape array of size 921600 into shape (1,128,128,1)","Body":"<p>I am taking image as a form input from the user and then trying to resizing it to (1,128,128,1) but getting \n<code>ValueError: cannot reshape array of size 921600 into shape (1,128,128,1)<\/code><\/p>\n\n<p>Following is the snippet for taking image and then resizing it to (1,128,128,1)-<\/p>\n\n<pre><code>def predit():\n   im = Image.open(request.files['image'])\n   image_data = np.asarray(im)\n   input_img_flatten = cv2.resize(image_data,(128,128)).flatten()\n   im2arr = np.reshape(input_img_flatten,(1,128,128,1))\n   return str(np.argmax(model.predict(im2arr)))\n<\/code><\/pre>\n\n<p>When I am not taking input image as a form input but just opening it from my desktop using the following approach , my code is running correctly.<\/p>\n\n<pre><code>model = load_model('.\/latest.hdf5')\nimg = cv2.imread('\/Users\/swastik\/thapar\/test\/test2.jpg',0)\ninput_img_flatten = cv2.resize(img,(128,128).flatten()\nim2arr = np.array(input_img_flatten).reshape((1, 128, 128, 1)) \nprint(np.argmax(model.predict(im2arr)))\n<\/code><\/pre>\n\n<p>How to do it?<\/p>\n\n<p>I have seen this answer <a href=\"https:\/\/stackoverflow.com\/questions\/45408496\/getting-error-cannot-reshape-array-of-size-122304-into-shape-52-28-28\">Getting error: Cannot reshape array of size 122304 into shape (52,28,28)<\/a> but the op hasn't accepted any of the answers and even I am not able to understand the given solutions correctly.<\/p>\n\n<p>Complete Code-<\/p>\n\n<pre><code>from keras.models import load_model\nfrom PIL import Image\nimport numpy as np\nimport cv2\n\nmodel = load_model('.\/latest.hdf5')\nim = Image.open('Anyimageinyourfolder.jpg')\nimage_data = np.asarray(im)\ninput_img_flatten = cv2.resize(image_data,(128,128)).flatten()\nim2arr = np.array(input_img_flatten).reshape((1, 128, 128, 1))\nprint(np.argmax(model.predict(im2arr)))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53160126","CreationDate":"2018-11-05T18:28:36.517","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>Its hard to tell what is causing your issue. Assuming the images are grey-scale (single color channel) like they are in your example, maybe this code helps:<\/p>\n\n<pre><code>img = cv2.imread('messi5.jpg', cv2.IMREAD_GRAYSCALE)\nimg = cv2.resize(img, (128,128))\nimg = img[np.newaxis,:,:,np.newaxis]\nprint(img.shape)\n&gt;&gt;&gt; (1, 128, 128, 1)\n<\/code><\/pre>\n\n<p>Also, if you are only predicting one image, you still need to index the returned predictions like so: <code>print(np.argmax(model.predict(img)[0]))<\/code><\/p>\n\n<p>Hope this helps<\/p>\n"}]}
+{"QuestionId":53159251,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T17:24:59.800","AcceptedAnswerId":53159696.0,"OwnerUserId":6068550.0,"Title":"Matmul Eror shape","Body":"<h2>Defining weights as variables in a linear model<\/h2>\n\n<p>I'm new in tensorflow , and i ran this code using <code>tf.matmul<\/code>, \nAt first- I didn't understand why the shape in <code>matmul<\/code> is not good.- i fixed it with another [] int the definition of the variable.<\/p>\n\n<p>now- i don't understand why it's still doesn't work.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>import tensorflow as tf\r\n\r\nW = tf.Variable([[.3]], tf.float32)\r\nb = tf.Variable([[-.3]], tf.float32)\r\nx = tf.placeholder(tf.float32)\r\n\r\nlinear_model = tf.matmul(x, W) + b\r\n\r\nsess = tf.Session()\r\n\r\ninit = tf.global_variables_initializer()\r\nsess.run(init)\r\n\r\nprint(sess.run(linear_model, {x: [[1, 2, 3, 4]]}))<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-css lang-css prettyprint-override\"><code>C:\\Users\\hagayj\\AppData\\Local\\Programs\\Python\\Python35\\python.exe \"C:\/Users\/hagayj\/OneDrive\/\u05dc\u05d9\u05de\u05d5\u05d3\u05d9\u05dd\/untitled1\/Defining weights as variables in a linear model.py\"\r\n2018-11-05 20:21:31.580447: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\r\nTraceback (most recent call last):\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1292, in _do_call\r\n    return fn(*args)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1277, in _run_fn\r\n    options, feed_dict, fetch_list, target_list, run_metadata)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1367, in _call_tf_sessionrun\r\n    run_metadata)\r\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Matrix size-incompatible: In[0]: [1,4], In[1]: [1,1]\r\n  [[{{node MatMul}} = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_Placeholder_0_0, Variable\/read)]]\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n  File \"&lt;encoding error&gt;\", line 14, in &lt;module&gt;\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 887, in run\r\n    run_metadata_ptr)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1110, in _run\r\n    feed_dict_tensor, options, run_metadata)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1286, in _do_run\r\n    run_metadata)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\client\\session.py\", line 1308, in _do_call\r\n    raise type(e)(node_def, op, message)\r\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Matrix size-incompatible: In[0]: [1,4], In[1]: [1,1]\r\n  [[{{node MatMul}} = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_Placeholder_0_0, Variable\/read)]]\r\n\r\nCaused by op 'MatMul', defined at:\r\n  File \"&lt;encoding error&gt;\", line 7, in &lt;module&gt;\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\ops\\math_ops.py\", line 2053, in matmul\r\n    a, b, transpose_a=transpose_a, transpose_b=transpose_b, name=name)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\ops\\gen_math_ops.py\", line 4856, in mat_mul\r\n    name=name)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\r\n    op_def=op_def)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 488, in new_func\r\n    return func(*args, **kwargs)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3272, in create_op\r\n    op_def=op_def)\r\n  File \"C:\\Users\\hagayj\\AppData\\Roaming\\Python\\Python35\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1768, in __init__\r\n    self._traceback = tf_stack.extract_stack()\r\n\r\nInvalidArgumentError (see above for traceback): Matrix size-incompatible: In[0]: [1,4], In[1]: [1,1]\r\n  [[{{node MatMul}} = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_Placeholder_0_0, Variable\/read)]]\r\n\r\n\r\nProcess finished with exit code 1<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"53159696","CreationDate":"2018-11-05T17:55:15.810","ParentId":null,"OwnerUserId":"10349960","Title":null,"Body":"<p>Ok, it should work if you add additional brackets:<\/p>\n\n<pre><code>W = tf.Variable([[.3]], tf.float32)\nb = tf.Variable([[-.3]], tf.float32)\nx = tf.placeholder(tf.float32)\n...\nprint(sess.run(linear_model, {x: [[1],[2],[3],[4]]}))\n<\/code><\/pre>\n"},{"AnswerId":"53159694","CreationDate":"2018-11-05T17:55:12.127","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The issue is you're using <code>matmul<\/code> to multiply one-dimensional vectors.  If you check the shape of <code>W<\/code> by <code>W.get_shape()<\/code>, it will return (1,) while it should be a 2D matrix of shape (1,1).  You can do this by simply adding brackets: <code>W = tf.Variable([[.3]], tf.float32)<\/code>.<\/p>\n\n<p>Doing this for <code>x<\/code> as well (<code>sess.run(linear_model, {x: [[1, 2, 3, 4]]})<\/code>) creates a (1,4) matrix.  But if you multiply <code>x<\/code> and <code>W<\/code>, you will get an error because you are trying to multiply a (1,4) matrix with a (1,1) matrix so they are incompatible (the second value in the first matrix shape needs to be the same as the first value in the second matrix shape).  Instead, you would need to take a transpose of <code>x<\/code> so that you are multiplying a (4,1) matrix with a (1,1) matrix.  This can be done by using the <code>transpose_a=True<\/code> flag in <code>matmul<\/code>.  Here's the final code that would run without errors:<\/p>\n\n<pre><code>import tensorflow as tf\n\nW = tf.Variable([[.3]], tf.float32)\nb = tf.Variable([[-.3]], tf.float32)\nx = tf.placeholder(tf.float32)\n\nlinear_model = tf.matmul(x, W, transpose_a=True) + b\n\nsess = tf.Session()\n\ninit = tf.global_variables_initializer()\nsess.run(init)\n\nprint(sess.run(linear_model, {x: [[1, 2, 3, 4]]}))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53159427,"AnswerCount":1,"Tags":"<python><machine-learning><computer-vision><pytorch>","CreationDate":"2018-11-05T17:36:49.710","AcceptedAnswerId":53170821.0,"OwnerUserId":1415817.0,"Title":"pytorch freeze weights and update param_groups","Body":"<p>Freezing weights in pytorch for <code>param_groups<\/code> setting.<\/p>\n\n<p>So if one wants to freeze weights during training:<\/p>\n\n<pre><code>for param in child.parameters():\n    param.requires_grad = False\n<\/code><\/pre>\n\n<p>the optimizer also has to be updated to not include the non gradient weights:<\/p>\n\n<pre><code>optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=opt.lr, amsgrad=True)\n<\/code><\/pre>\n\n<p>If one wants to use different <code>weight_decay<\/code> \/ learning rates for bias and weights\/this also allows for differing learning rates: <\/p>\n\n<pre><code>param_groups = [{'params': model.module.bias_parameters(), 'weight_decay': args.bias_decay},\n                {'params': model.module.weight_parameters(), 'weight_decay': args.weight_decay}]\n<\/code><\/pre>\n\n<p><code>param_groups<\/code> a <em>list<\/em> of <em>dics<\/em> is defined and passed into <code>SGD<\/code> as follows:<\/p>\n\n<pre><code>optimizer = torch.optim.Adam(param_groups, args.lr,\n                                 betas=(args.momentum, args.beta))\n<\/code><\/pre>\n\n<p><strong>How can this be achieved with freezing individual weights<\/strong>? Running filter over a list of dics or is there a way of adding tensors to the optimizer separately?<\/p>\n","answers":[{"AnswerId":"53170821","CreationDate":"2018-11-06T11:18:06.317","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Actually I think you don't have to update the <code>optimizer<\/code>. The <code>Parameters<\/code> handed over to the <code>optimizer<\/code> are just references. <\/p>\n\n<p>So when you change the <code>requires_grad<\/code> flag it will immediately be updated.<\/p>\n\n<p>But even if that would for some reason not be the case - as soon as you you set the <code>requires_grad<\/code> flag to be <code>False<\/code> you cannot anymore calculate gradients any <strong><em>new gradients<\/strong> (see at the bottom with <code>None<\/code> and zero gradients)<\/em> for this weight, so the gradient won't change anymore and if you use <code>optimizer.zero_grad()<\/code> it will just stay <code>zero<\/code>.<\/p>\n\n<p>So if there is no gradient, then there is also no need to exclude these from the <code>optimizer<\/code>. Because without gradient the <code>optimizer<\/code> will just do nothing, no matter what learning rate you use.<\/p>\n\n<p>Here is an small example to show this behaviour:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\nimport torch.optim as optim\n\nn_dim = 5\n\np1 = nn.Linear(n_dim, 1)\np2 = nn.Linear(n_dim, 1)\n\noptimizer = optim.Adam(list(p1.parameters())+list(p2.parameters()))\np2.weight.requires_grad = False\nfor i in range(4):\n    dummy_loss = (p1(torch.rand(n_dim)) + p2(torch.rand(n_dim))).squeeze()\n    optimizer.zero_grad()\n    dummy_loss.backward()\n    optimizer.step()\n    print('p1: requires_grad =', p1.weight.requires_grad, ', gradient:', p1.weight.grad)\n    print('p2: requires_grad =', p2.weight.requires_grad, ', gradient:', p2.weight.grad)\n    print()\n\n    if i == 1:\n        p1.weight.requires_grad = False\n        p2.weight.requires_grad = True\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>p1: requires_grad = True , gradient: tensor([[0.8522, 0.0020, 0.1092, 0.8167, 0.2144]])\np2: requires_grad = False , gradient: None\n\np1: requires_grad = True , gradient: tensor([[0.7635, 0.0652, 0.0902, 0.8549, 0.6273]])\np2: requires_grad = False , gradient: None\n\np1: requires_grad = False , gradient: tensor([[0., 0., 0., 0., 0.]])\np2: requires_grad = True , gradient: tensor([[0.1343, 0.1323, 0.9590, 0.9937, 0.2270]])\n\np1: requires_grad = False , gradient: tensor([[0., 0., 0., 0., 0.]])\np2: requires_grad = True , gradient: tensor([[0.0100, 0.0123, 0.8054, 0.9976, 0.6397]])\n<\/code><\/pre>\n\n<p>Here you can see that no gradients are calculated. You may have notice the gradient for <code>p2<\/code> is <code>None<\/code> at the beginning and later on it is <code>tensor([[0., 0., 0., 0., 0.]])<\/code> for <code>p1<\/code> instead of <code>None<\/code> after deactivating gradients.<\/p>\n\n<p>This is the case because <code>p1.weight.grad<\/code> is just a variable which is modified by <code>backward()<\/code> and <code>optimizer.zero_grad()<\/code>.<\/p>\n\n<p>So at the beginning <code>p1.weight.grad<\/code> is just initialized with <code>None<\/code>, after the gradients are written or accumulated to this variable they won't be cleared automatically. But because <code>optimizer.zero_grad()<\/code> is called they are set to zero and stay like this since <code>backward()<\/code> cannot anymore calculate new gradients with <code>requires_grad=False<\/code>.<\/p>\n\n<p>You can also change the code in the <code>if<\/code>-statement to:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if i == 1:\n    p1.weight.requires_grad = False\n    p1.weight.grad = None\n    p2.weight.requires_grad = True\n<\/code><\/pre>\n\n<p>So once reset to <code>None<\/code> they are left untouched and stay <code>None<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>p1: requires_grad = True , gradient: tensor([[0.2375, 0.7528, 0.1501, 0.3516, 0.3470]])\np2: requires_grad = False , gradient: None\n\np1: requires_grad = True , gradient: tensor([[0.5181, 0.5178, 0.6590, 0.6950, 0.2743]])\np2: requires_grad = False , gradient: None\n\np1: requires_grad = False , gradient: None\np2: requires_grad = True , gradient: tensor([[0.4797, 0.7203, 0.2284, 0.9045, 0.6671]])\n\np1: requires_grad = False , gradient: None\np2: requires_grad = True , gradient: tensor([[0.8344, 0.1245, 0.0295, 0.2968, 0.8816]])\n<\/code><\/pre>\n\n<p>I hope this makes sense to you!<\/p>\n"}]}
+{"QuestionId":53159930,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-11-05T18:14:16.803","AcceptedAnswerId":53159976.0,"OwnerUserId":8085890.0,"Title":"What should be the 5th dimension for the input to 3D-CNN while working with hyper-spectral images?","Body":"<p>I have a hyperspectral image having dimension <code>S * S * L<\/code> where <code>S*S<\/code> is the spatial size and <code>L<\/code> denotes the number of spectral bands.<\/p>\n\n<p>Now the shape of my X (image array) is: <code>(1, 145, 145, 200)<\/code> where 1 is the number of examples, 145 is the length and width of the image and 200 is no. of channels of the image.<\/p>\n\n<p>I want to input this small windows of this image (having dimension like <code>W * W * L<\/code>; W &lt; S) into a 3D CNN, but for that, I need to have 5 dimensions in the following format: <code>(batch, length, width, depth, channels)<\/code>.<\/p>\n\n<p>It seems to me I am missing one of the spatial dimensions, how do I convert my image array into a 5-dimensional array without losing any information?<\/p>\n\n<p>I am using python and Keras for the above.<\/p>\n","answers":[{"AnswerId":"53159976","CreationDate":"2018-11-05T18:17:10.757","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>If you want to convolve along the dimension of your channels, you should add a singleton dimension in the position of channel.  If you don't want to convolve along the dimension of your channels, you should use a 2D CNN.<\/p>\n"},{"AnswerId":"53163769","CreationDate":"2018-11-05T23:25:06.390","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>What you want is a 2D CNN, not a 3D one. A 2D CNN already supports multiple channels, so you should have no problem using it with a hyperspectral image.<\/p>\n"}]}
+{"QuestionId":53160222,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-hub>","CreationDate":"2018-11-05T18:37:41.777","AcceptedAnswerId":53224168.0,"OwnerUserId":6595663.0,"Title":"create_module_spec - tfhub","Body":"<p>I am trying to create a <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">simple audio recognition<\/a> to spot key words. Since my data set is small i am performing transfer learning. This is how the <a href=\"https:\/\/github.com\/BhanuKiranChaluvadi\/tensorflow_exercise\/blob\/master\/audio_recognition\/google_simple_audio_recognition.png\" rel=\"nofollow noreferrer\">graph looks<\/a>. Following this <a href=\"https:\/\/github.com\/tensorflow\/hub\/issues\/27\" rel=\"nofollow noreferrer\">link<\/a> i created a module. And here is the code <\/p>\n\n<pre><code>import tensorflow_hub as hub\nimport tensorflow as tf\n\n# pylint: disable=unused-import\nfrom tensorflow.contrib.framework.python.ops import audio_ops as contrib_audio\n# pylint: enable=unused-import\n\ndef module_fn():\n    input_name = \"Reshape:0\"\n    output_name = \"Reshape_2:0\"\n    graph_def = tf.GraphDef()\n    with open('my_frozen_graph.pb', \"rb\") as f:\n        graph_def.ParseFromString(f.read())\n\n    input_ten=tf.placeholder(tf.float32, shape = (1, 98, 40))\n    output_ten,=tf.import_graph_def(graph_def, input_map = {input_name: input_ten}, return_elements = [output_name])\n    hub.add_signature(inputs = input_ten, outputs = output_ten)\n\n\nspec = hub.create_module_spec(module_fn)\nmodule = hub.Module(spec)\nwith tf.Session() as session:\n    module.export('test_module',session)\n<\/code><\/pre>\n\n<p>Although it does executed and created a 'test_module' folder. <\/p>\n\n<pre><code>test_module\n    |--&gt; assets\n    |--&gt; variables\n    |--&gt; saved_model.pb\n    |--&gt; tfhub_module.pb\n<\/code><\/pre>\n\n<p>How ever I have few questions<\/p>\n\n<ol>\n<li><p>The variables folder is empty. Not sure if this is how it supposed to be ?<\/p><\/li>\n<li><p><code>input_ten=tf.placeholder(tf.float32, shape = (1, 98, 40))<\/code> \nIs this correct ? 98X48 are the image size and first tuple usually represent batch size. Should it be kept as '1' or for unknown batch size 'None' ?<\/p><\/li>\n<li><p>After loading the module into the script <\/p>\n\n<p>height, width = hub.get_expected_image_size('test_module')<\/p>\n\n<p>is giving me an error.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53224168","CreationDate":"2018-11-09T10:42:29.180","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>Let me try to answer your questions in turn.<\/p>\n\n<ol>\n<li><p>If the graph def you build your model from is indeed frozen (i.e., all variables have been replaced by constants), there are no variables that need writing to the checkpoint that is commonly located at variables\/variables*. So this looks explicable to me. --\nThat said, Hub modules would give you a way to avoid freezing graph defs: call the original graph building code in the module_fn, and restore pre-trained variables in the session before calling Module.export().<\/p><\/li>\n<li><p>For your type of module, you get to make the rules. ;-) Hub Modules can accommodate all sorts of input and output shapes, including partially or fully unknown shapes. An input placeholder like above will have to have a shape that is compatible with the graph you are plugging it into. That graph, in turn, will use shapes that work with the convolutions it is doing.\nGenerally speaking, it is often useful to use the leading dimension for batch size and leave that unspecified (<code>None<\/code>).<\/p><\/li>\n<li><p>hub.get_expected_image_size() is meant for use with image inputs. I would avoid it here.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53160230,"AnswerCount":1,"Tags":"<python><keras><deep-learning><rnn><grid-search>","CreationDate":"2018-11-05T18:38:11.157","AcceptedAnswerId":null,"OwnerUserId":10449878.0,"Title":"GridSearch fit method error : TypeError [Tuning the parameters for RNN model]","Body":"<p>CONTEXT : \" Building a Next-word Predictor using RNN \"<\/p>\n\n<p>I have made a LSTM model around which I want to tune number of epochs and batch_size. However I can't get past fit method of GridSearch.<\/p>\n\n<p>Please click on the links to watch the code:<\/p>\n\n<p>X_train is a 50 words list from a book, while y_train is 1 word list that comes after 50 words.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LcWxf.png\" rel=\"nofollow noreferrer\">CODE FOR THE MODEL<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xbnuQ.png\" rel=\"nofollow noreferrer\">CODE FOR GRID SEARCH<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/C6vH3.png\" rel=\"nofollow noreferrer\">ERROR WHILE EXECUTING<\/a><\/p>\n\n<p>THANKS FOR THE HELP<\/p>\n","answers":[{"AnswerId":"55183508","CreationDate":"2019-03-15T13:16:57.567","ParentId":null,"OwnerUserId":"7959545","Title":null,"Body":"<p>I think you are forgetting to put your keras model into a sk-learn wrapper without wrapper it is not possible to run grid or random search on keras models. Like pass your model in keras Classifier function build_fn.<\/p>\n\n<pre><code>from keras.wrappers.scikit_learn import KerasClassifier\nfrom sklearn.model_selection import RandomizedSearchCV\n\n    model = KerasClassifier(build_fn=Model, epochs=10, batch_size\n=20, verbose=1)\nRandom = RandomizedSearchCV(estimator=model, cv=KFold(3), param_distributions=params, \n                          verbose=20,  n_iter=10, n_jobs=1)\n\nRandom_result = Random.fit(X_train, y_train)\n<\/code><\/pre>\n\n<p>If you want to see more Please go to the link. Here Working random search model is performed.<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/sohaibanwaar1203\/hyperparameter-tunning-and-cnn-visualization?scriptVersionId=11619218\" rel=\"nofollow noreferrer\">Random Search Link Kaggle Kernal<\/a><\/p>\n"}]}
+{"QuestionId":53160825,"AnswerCount":0,"Tags":"<python><tensorflow><keras><keras-layer>","CreationDate":"2018-11-05T19:21:24.320","AcceptedAnswerId":null,"OwnerUserId":10608708.0,"Title":"keras: record layer-wise learning rate","Body":"<p>I'm implementing a simple CNN with Keras, and trying to set up layer-wise learning rate in Adam. I take a reference of <a href=\"https:\/\/erikbrorson.github.io\/2018\/04\/30\/Adam-with-learning-rate-multipliers\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>. The revised Adam as shown below:<\/p>\n\n<pre><code>class Adam_lr_mult(Optimizer): \n    def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,\n                 epsilon=None, decay=0., amsgrad=False,\n                 multipliers=None, debug_verbose=True,**kwargs):\n        ...'''Omitted'''\n\n        self.multipliers = multipliers\n        self.layerwise_lr={}    # record layer-wise lr\n        self.debug_verbose = debug_verbose\n\n    @interfaces.legacy_get_updates_support\n    def get_updates(self, loss, params):\n        grads = self.get_gradients(loss, params)\n        self.updates = [K.update_add(self.iterations, 1)]\n\n        lr = self.lr\n        if self.initial_decay &gt; 0:\n        lr *= (1. \/ (1. + self.decay * K.cast(self.iterations,\n                                              K.dtype(self.decay))))\n\n        t = K.cast(self.iterations, K.floatx()) + 1\n        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) \/\n                     (1. - K.pow(self.beta_1, t)))\n\n        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]\n        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]\n        if self.amsgrad:\n            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]\n        else:\n            vhats = [K.zeros(1) for _ in params]\n\n        self.weights = [self.iterations] + ms + vs + vhats\n\n        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):\n\n            # Learning rate multipliers\n            if self.multipliers:\n                multiplier = [mult for mult in self.multipliers if mult in p.name]\n                if self.debug_verbose:\n                    print('parameter: ',p.name)\n            else:\n                multiplier = None\n\n            if multiplier:\n                new_lr_t = lr_t * self.multipliers[multiplier[0]]\n                self.layerwise_lr[multiplier[0]] = K.get_value(new_lr_t)\n\n                if self.debug_verbose:\n                    print('Setting {} to learning rate : {}'.format(multiplier[0], new_lr_t))\n                    print('learning rate:',K.get_value(new_lr_t))\n                    print('Dict:',self.layerwise_lr)\n                    print('\\n')\n            else:\n                new_lr_t = lr_t\n                self.layerwise_lr[p.name.split('\/')[0]] = K.get_value(new_lr_t)\n\n                if self.debug_verbose:\n                    print('No change in learning rate : {}'.format(p.name))\n                    print('learning rate:',K.get_value(new_lr_t))\n                    print('Dict:',self.layerwise_lr)\n                    print('\\n')\n\n            ...'''Omitted'''\n\n        print(***__Hello__***)\n        return self.updates\n<\/code><\/pre>\n\n<p>Besides, I use ReduceLROnPlateau and CSVLogger callback function to record learning rate. In order to record more information about layer-wise learning rate, I also revise ReduceLROnPlateau:<\/p>\n\n<pre><code>class ReduceLROnPlateau_lr_mult(Callback):\n    def __init__(self, monitor='val_loss', factor=0.1, patience=10,\n             verbose=0, mode='auto', min_delta=1e-4, cooldown=0, min_lr=0,\n             **kwargs):\n        ...'''Omitted'''\n\n    def on_epoch_end(self, epoch, logs=None):\n        logs = logs or {}\n        logs['lr'] = K.get_value(self.model.optimizer.lr)\n        logs.update(self.model.optimizer.layerwise_lr) # only add this line\n        ...'''Omitted'''\n<\/code><\/pre>\n\n<p>In order to test revised Adam and ReduceLROnPlateau function, I use MNIST dataset and build a simple CNN with only 4 convolution layers, 4 batch_normalization layers and 1 dense layer, the code and result as shown below:<\/p>\n\n<pre><code># Learning multiplier\nlr_multipliers = {}\nlr_multipliers['conv2d_1'] = 0.8\nlr_multipliers['batch_normalization_1'] = 0.8\nlr_multipliers['conv2d_2'] = 0.6\nlr_multipliers['conv2d_3'] = 0.4\nlr_multipliers['conv2d_4'] = 0.2\n\n# Adam with layer-wise lr\nadam_lr_mult = Adam_lr_mult(multipliers=lr_multipliers)\n\n# ReduceLROnPlateau with layer-wise lr dict\nlr_reducer_mult = ReduceLROnPlateau_lr_mult(monitor='val_loss',factor=np.sqrt(0.1), \n                      cooldown=0, patience=5, min_lr=0.00001, mode='auto', verbose=1)\n\nmodel.compile(loss='categorical_crossentropy', optimizer=adam_lr_mult, metrics=['accuracy'])\n\nmodel.fit(...,verbose=2,\n             callbacks=[lr_reducer_mult, early_stopper, csv_logger, ...])\n<\/code><\/pre>\n\n<p>the result shows:<\/p>\n\n<pre><code>...'''Omitted'''\nparameter:  batch_normalization_4\/beta:0\nNo change in learning rate : batch_normalization_4\/beta:0\nlearning rate: 0.00031623512\nDict: {'conv2d_1': 0.00025298807, 'batch_normalization_1': 0.00025298807,...}\n\nparameter:  dense_1\/kernel:0\nNo change in learning rate : dense_1\/kernel:0\nlearning rate: 0.00031623512\nDict: {'conv2d_1': 0.00025298807, 'batch_normalization_1': 0.00025298807,...}\n\nparameter:  dense_1\/bias:0\nNo change in learning rate : dense_1\/bias:0\nlearning rate: 0.00031623512\nDict: {'conv2d_1': 0.00025298807, 'batch_normalization_1': 0.00025298807,...}\n\n***__Hello__***\nTrain on 48000 samples, validate on 12000 samples\nEpoch 1\/100\n - 4s - loss: 0.4517 - acc: 0.9098 - val_loss: 0.2027 - val_acc: 0.9572\nEpoch 2\/100\n - 2s - loss: 0.1029 - acc: 0.9827 - val_loss: 0.1374 - val_acc: 0.9718\nEpoch 3\/100\n - 2s - loss: 0.0739 - acc: 0.9905 - val_loss: 0.0929 - val_acc: 0.9833\nEpoch 4\/100\n - 2s - loss: 0.0604 - acc: 0.9939 - val_loss: 0.0815 - val_acc: 0.9865\nEpoch 5\/100\n - 2s - loss: 0.0513 - acc: 0.9959 - val_loss: 0.0785 - val_acc: 0.9864\nEpoch 6\/100\n - 2s - loss: 0.0448 - acc: 0.9979 - val_loss: 0.1081 - val_acc: 0.9759\nEpoch 7\/100\n - 2s - loss: 0.0405 - acc: 0.9984 - val_loss: 0.0752 - val_acc: 0.9864\nEpoch 8\/100\n - 2s - loss: 0.0368 - acc: 0.9990 - val_loss: 0.1382 - val_acc: 0.9666\nEpoch 9\/100\n - 2s - loss: 0.0337 - acc: 0.9996 - val_loss: 0.0659 - val_acc: 0.9890\nEpoch 10\/100\n - 2s - loss: 0.0314 - acc: 0.9998 - val_loss: 0.0746 - val_acc: 0.9860\n...\n\nEpoch 25\/100\n - 2s - loss: 0.0177 - acc: 0.9991 - val_loss: 0.1212 - val_acc: 0.9731\n\nEpoch 00025: ReduceLROnPlateau reducing learning rate to 0.00031622778103685084.\nEpoch 26\/100\n - 2s - loss: 0.0155 - acc: 0.9998 - val_loss: 0.0446 - val_acc: 0.9915\nEpoch 27\/100\n - 2s - loss: 0.0146 - acc: 1.0000 - val_loss: 0.0422 - val_acc: 0.9926\n<\/code><\/pre>\n\n<p>My question:<\/p>\n\n<p>I'm not sure where I'm going wrong, <code>logs['lr']<\/code> changes in <a href=\"https:\/\/imgur.com\/YREOTQx\" rel=\"nofollow noreferrer\">CSV file<\/a> but the <code>dictionary \"layerwise_lr\"<\/code> doesn't. In order to find out the problem, I add a line\n<code>print(***__Hello__***)<\/code> in Adam and it only appear one time. Which makes me confused, the information about setting learning rate only appeared before first epoch and never appear again. Could anybody give me some suggestion? Thanks a lot!<\/p>\n","answers":[]}
+{"QuestionId":53161217,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-05T19:49:40.973","AcceptedAnswerId":53161394.0,"OwnerUserId":7031763.0,"Title":"How to set working directory for tensorflow","Body":"<p>I installed tensorflow in my machine using both <strong>pip<\/strong> and <strong>conda<\/strong> installation. I can import it in my <strong>python idle<\/strong> as well as in <code>C:\\Users\\Anaconda3\\envs\\tensorflow\\Scripts<\/code> this directory using jupyter notebook. when I want to import in another directory in jupyter notebook, it shows me <\/p>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'tensorflow'.<\/p>\n<\/blockquote>\n\n<p>How can I import it in my working directory?<\/p>\n","answers":[{"AnswerId":"53161394","CreationDate":"2018-11-05T20:03:14.893","ParentId":null,"OwnerUserId":"7505383","Title":null,"Body":"<p>From your comments it seems you are running your jupyter-notebook from the wrong conda environment (your base environment rather than the tensorflow one you have created). This is why the packages don't appear to be installed.<\/p>\n\n<p>Make sure your have <code>jupyter<\/code> installed in your tensorflow environment and also install <code>nb_conda<\/code>. You can do that with:<\/p>\n\n<pre><code>conda install jupyter nb_conda -n tensorflow\n<\/code><\/pre>\n\n<p>Then launch jupyter from within your tensorflow environment and you should see it in the list of kernels<\/p>\n"}]}
+{"QuestionId":53161268,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-05T19:52:46.660","AcceptedAnswerId":null,"OwnerUserId":7322746.0,"Title":"flag provided but not defined: --mount","Body":"<p>I'm pretty new with Tensorflow Serving, and I followed the basic instruction from: <a href=\"https:\/\/tensorflow.google.cn\/serving\/serving_basic\" rel=\"nofollow noreferrer\">https:\/\/tensorflow.google.cn\/serving\/serving_basic<\/a><\/p>\n\n<p>All previous steps are ok, but when I use:<\/p>\n\n<p>docker run -p 8500:8500 --mount type=bind,source=\/tmp\/mnist,target=\/models\/mnist -e MODEL_NAME=mnist -t tensorflow\/serving &amp;<\/p>\n\n<p>It gave me:\nflag provided but not defined: --mount<\/p>\n\n<p>Can someone help me with it?<\/p>\n","answers":[]}
+{"QuestionId":53161638,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-11-05T20:21:39.363","AcceptedAnswerId":53417427.0,"OwnerUserId":5998829.0,"Title":"How large can a TensorFlow .record file be before it causes performance issues?","Body":"<p>In the TensorFlow Object Detection API, they advocate sharding if the dataset contains \"more than a few thousand examples\", <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/using_your_own_dataset.md\" rel=\"nofollow noreferrer\">noting that<\/a>:<\/p>\n\n<ul>\n<li>tf.data.Dataset API can read input examples in parallel improving throughput.<\/li>\n<li>tf.data.Dataset API can shuffle the examples better with sharded files which improves performance of the model slightly.<\/li>\n<\/ul>\n\n<p>A few thousand is a bit vague, and it would be nice to have a more precise answer, such as a file size.  In other words, how big can a .record file before it starts causing performance issues?  What file size should we aim for when sharding our data?<\/p>\n","answers":[{"AnswerId":"53417427","CreationDate":"2018-11-21T17:17:18.077","ParentId":null,"OwnerUserId":"2623964","Title":null,"Body":"<p>It seems like the TensorFlow team recommends ~100MB shards. <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/overview\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/performance\/overview<\/a> You might also consider the performance implications related to batch size while training. <a href=\"https:\/\/www.pugetsystems.com\/labs\/hpc\/GPU-Memory-Size-and-Deep-Learning-Performance-batch-size-12GB-vs-32GB----1080Ti-vs-Titan-V-vs-GV100-1146\/\" rel=\"nofollow noreferrer\">https:\/\/www.pugetsystems.com\/labs\/hpc\/GPU-Memory-Size-and-Deep-Learning-Performance-batch-size-12GB-vs-32GB----1080Ti-vs-Titan-V-vs-GV100-1146\/<\/a><\/p>\n"}]}
+{"QuestionId":53161875,"AnswerCount":1,"Tags":"<python><tensorflow><concat>","CreationDate":"2018-11-05T20:39:49.077","AcceptedAnswerId":null,"OwnerUserId":6091064.0,"Title":"string concatenation in tensorflow","Body":"<p>I have a <code>tf.string<\/code> tensor, <code>chars<\/code>, with shape <code>chars[Batch][None]<\/code> where <code>None<\/code> denotes a dynamic shaped tensor (output from a variable length sequence).  <\/p>\n\n<p>If this tensor's shape were known (e.g. <code>chars[Batch][Time]<\/code>), then I could achieve concatenation of strings along the last dimension as:<\/p>\n\n<pre><code>chars = tf.split(chars,chars.shape[-1],axis=-1)\nwords = tf.squeeze(tf.strings.join(chars))\n<\/code><\/pre>\n\n<p>However, since the shape is unknown until runtime, I cannot use <code>split<\/code>.<\/p>\n\n<p>Is there another way to accomplish this for a dynamic shaped <code>string<\/code> tensor?<\/p>\n\n<p>In other words, I would like the string analogy of<\/p>\n\n<pre><code>words = tf.reduce_sum(chars,axis=-1)\n<\/code><\/pre>\n\n<p>along a dynamic shaped dimension.<\/p>\n","answers":[{"AnswerId":"53162389","CreationDate":"2018-11-05T21:19:53.133","ParentId":null,"OwnerUserId":"6091064","Title":null,"Body":"<p>This can be accomplished via:<\/p>\n\n<pre><code>words = tf.reduce_join(chars,axis=-1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53162151,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-11-05T21:00:53.627","AcceptedAnswerId":null,"OwnerUserId":457901.0,"Title":"Keras model overfiting","Body":"<p>im working on a multi class image classification problem in keras. Using the dog-breeds dataset on kaggle. My accuracy for 12 breeds is 95% yet, my validation accuracy is only 50%. It looks like the model is overfitting, but im not sure what i would need to do to prevent overfitting<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gGDPf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gGDPf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's my basic training setup<\/p>\n\n<pre><code>from keras.utils.np_utils import to_categorical\nfrom keras.layers import Conv2D, Activation, MaxPooling2D\nfrom keras import optimizers\nfrom keras.layers.normalization import BatchNormalization\n\nimg_width, img_height = 224, 224\n\ndatagen_top = ImageDataGenerator(\n        rotation_range=180,\n        width_shift_range=0.2,\n        height_shift_range=0.2,\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True,\n        fill_mode='nearest')\n\ngenerator_top = datagen_top.flow_from_directory(\n    train_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode='categorical',\n    shuffle=False)\n\nnb_train_samples = len(generator_top.filenames)\nnum_classes = len(generator_top.class_indices)\n\n\ntrain_data = bottleneck_features_train\n\n# get the class lebels for the training data, in the original order\ntrain_labels = generator_top.classes\n\n# https:\/\/github.com\/fchollet\/keras\/issues\/3467\n# convert the training labels to categorical vectors\ntrain_labels = to_categorical(train_labels, num_classes=num_classes)\n\ngenerator_top = datagen_top.flow_from_directory(\n    validation_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode=None,\n    shuffle=False)\n\nnb_validation_samples = len(generator_top.filenames)\n\nvalidation_data = bottleneck_features_validation\n\nvalidation_labels = generator_top.classes\nvalidation_labels = to_categorical(\n    validation_labels, num_classes=num_classes)\n\ninput_shape = train_data.shape[1:]\n\nmodel = Sequential()\nmodel.add(Flatten(input_shape=input_shape))\nmodel.add(Dense(num_classes, activation='softmax'))\n\nmodel.compile(optimizer=optimizers.RMSprop(lr=2e-4),\n              loss='categorical_crossentropy', metrics=['accuracy'])\n\nhistory = model.fit(train_data, train_labels,\n                    epochs=epochs,\n                    batch_size=batch_size,\n                    callbacks=[],\n                    validation_data=(validation_data, validation_labels))\n\nmodel.save_weights(top_model_weights_path)\n\n(eval_loss, eval_accuracy) = model.evaluate(\n    validation_data, validation_labels, batch_size=batch_size, verbose=1)\n<\/code><\/pre>\n\n<p>notebook is on colab.\n<a href=\"https:\/\/colab.research.google.com\/drive\/13RzXpxE-yMEuMFPHnmBpzD1gFXWxVyXK\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/13RzXpxE-yMEuMFPHnmBpzD1gFXWxVyXK<\/a><\/p>\n","answers":[{"AnswerId":"53162861","CreationDate":"2018-11-05T22:01:14.460","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>A single layer network isn't gonna fly with an image classification problem. The network will never be able to generalize because there is no opportunity to. Try expanding the network with a few more layers and maybe try a CNN.<\/p>\n\n<p>Example: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3, 3), padding='same',\n                 activation='relu',\n                 input_shape=input_shape))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=optimizers.RMSprop(),\n              metrics=['accuracy'])\n<\/code><\/pre>\n"},{"AnswerId":"53164318","CreationDate":"2018-11-06T00:45:59.993","ParentId":null,"OwnerUserId":"9742432","Title":null,"Body":"<p>This usually happens when you have too many layers and the resulting dimensionality (after striding and pooling) is lower than the minimum input size (convolutional kernel) of a subsequent layer.<\/p>\n\n<p>Which is the image size of the dog-breeds data?<\/p>\n\n<p>Have you made sure that the reshaping works correctly?<\/p>\n"}]}
+{"QuestionId":53162693,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-05T21:46:30.843","AcceptedAnswerId":null,"OwnerUserId":7715968.0,"Title":"How do I specify the feature columns in tf.estimator.DNNClassifier?","Body":"<p>I have a dataset that looks basically like this<\/p>\n\n<pre><code>FeatureA FeatureB FeatureC Target\n\n3        2        5        1\n\n4        1        8        0\n\n3        0        2        0\n<\/code><\/pre>\n\n<p>The target is categorical (1,0), features are numerical. The goal is to learn an algorithm to classify based on all 3 features whether target is 0 or 1.<\/p>\n\n<p>When calling the classifier e.g. on the Google colab's example on building a text classifier, they have written the code as such<\/p>\n\n<pre><code>estimator = tf.estimator.DNNClassifier(\nhidden_units=[500, 100],\nfeature_columns=[embedded_text_feature_column],\nn_classes=2,\noptimizer=tf.train.AdagradOptimizer(learning_rate=0.003))\n<\/code><\/pre>\n\n<p>However, in my example, where I just want to use the 3 feature columns as <code>feature_columns<\/code>, how do I tell <code>tf.estimator.DNNClassifier<\/code> that? I.e. what do I pass into the argument <code>feature_columns<\/code>?<\/p>\n","answers":[{"AnswerId":"53165193","CreationDate":"2018-11-06T03:02:06.123","ParentId":null,"OwnerUserId":"10240511","Title":null,"Body":"<p>Its actually quite simple...though it took me a while to understand it.<\/p>\n\n<p>Take a close look this code below:<\/p>\n\n<pre><code># Represent a tf.float64 scalar.\nnumeric_feature_column = tf.feature_column.numeric_column(key=\"SepalLength\",\n                                                          dtype=tf.float64)\n<\/code><\/pre>\n\n<p>One can say that you are \"building\" a column for this particular model. In this scenario, the column name is  \"SepalLength\" and its data type is float64.<\/p>\n\n<p>Lets apply it to your scenario:<\/p>\n\n<pre><code>numeric_feature_column_1 = tf.feature_column.numeric_column(key=\"_1\",\n                                                          dtype=tf.float64)\n\nnumeric_feature_column_2 = tf.feature_column.numeric_column(key=\"_2\",\n                                                          dtype=tf.float64)\n\nnumeric_feature_column_3 = tf.feature_column.numeric_column(key=\"_3\",\n                                                          dtype=tf.float64)\n\nestimator   = tf.estimator.DNNClassifier(\n              hidden_units=[500, 100],\n              feature_columns=[numeric_feature_column_1,\n                               numeric_feature_column_2,\n                               numeric_feature_column_3],\n              n_classes=2,\n              optimizer = tf.train.AdagradOptimizer(learning_rate=0.003)\n              )\n<\/code><\/pre>\n\n<p>I strongly recommend that you read these links below to get a better understanding of Tensorflow's data pipeline and how it connects to the Estimator class.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/datasets<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/guide\/feature_columns\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/feature_columns<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/numeric_column\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/feature_column\/numeric_column<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/census\/estimator\/trainer\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/census\/estimator\/trainer\/model.py<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53164055,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-11-06T00:05:33.490","AcceptedAnswerId":null,"OwnerUserId":214488.0,"Title":"training tensorflow data API with procedural on-the-fly data generation","Body":"<p>My question is similar to <a href=\"https:\/\/stackoverflow.com\/questions\/47318734\/on-the-fly-generation-with-dataset-api-tensorflow\">this<\/a> one, in that I would like to generate batches of training data on the fly. I have a function <code>get_random_batch(batch_size, input_path, target_path, **other_kwargs)<\/code> which returns <code>inputs<\/code> and <code>targets<\/code>, but it is a pure vanilla python \/ numpy function, not tensorflow. So it returns numpy arrays, not tensors. (The function is quite a large complex one with some 3rd party libs. To port it to tensorflow is not very feasible. To preprocess everything is also not feasible as the data is huge, and I don't have the space to store it all preprocessed! In fact I don't even train a single epoch, I just take random selections for hundreds of thousands of iterations).<\/p>\n\n<p>For a few years I've been using tensorflow's low level training API: generate a batch of input-target pairs, run forward pass with feed (into placeholders) and fetch, calculate loss, apply gradients, repeat etc.<\/p>\n\n<p>Now I finally would like to try out the newer data API (so that I can use the Keras API for building the model and 'fitting' etc), but I can't figure out how to migrate. All of the documentation I've seen assumes that the data loading and preprocessing is part of the graph, and the output of the dataset are already tensors. <\/p>\n\n<p>--<\/p>\n\n<p><strong>Update<\/strong>: Ok this seems to work with <strong>tf.data.Dataset.from_generator<\/strong><\/p>\n\n<pre><code>compile_kwargs = dict(\n    optimizer = tf.train.AdamOptimizer(3e-4),\n    loss = 'mse',\n    metrics = ['accuracy', 'mse']\n)\n\ndef prepare_data():\n    x,t = get_random_training_pair() # this returns one training pair (each np.float32 ndarrays)\n    yield (x, t)\n\ndataset = tf.data.Dataset.from_generator(prepare_data, (tf.float32, tf.float32))\ndataset = dataset.batch(128).repeat()\nmodel = build_keras_model()\nmodel.compile(**compile_kwargs)\nmodel.summary()\nmodel.fit(dataset, epochs=10, steps_per_epoch=100, shuffle=False)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53164105,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-06T00:12:16.570","AcceptedAnswerId":null,"OwnerUserId":7509978.0,"Title":"How can I make my model to overfit to my data?","Body":"<p>I use Acrobot-v1 of openai gym to make my model\nThe observation state size is 6 and action is -1 or 1 or 0. \nI want to check whether my model is capable of learning.<\/p>\n\n<p>So, I used cross-validation method and my set is divided to training set of 64000 state-action pair and validation set of 16000 state-action pair.<\/p>\n\n<p>But whenever I increase the number of neurons in my hidden layer, validation loss keep decreasing in training process. \nI've learned many parameter makes overfitting and the overfitting can be detected by watching increase of validation loss.\nI do not know what should I do.<\/p>\n\n<pre><code>self.DEncoder = tf.keras.models.Sequential([\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nfeature),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU()\n])\n\nself.DDecoder = tf.keras.models.Sequential([\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nInHidden),\n    tf.keras.layers.BatchNormalization(),\n    tf.keras.layers.LeakyReLU(),\n    tf.keras.layers.Dense(self.nState*self.npred),\n    tf.keras.layers.Reshape((self.npred,self.nState,1))\n])\n<\/code><\/pre>\n\n<p>This is my model. I have used l2 loss (mse, mean squared error).<\/p>\n","answers":[{"AnswerId":"53164269","CreationDate":"2018-11-06T00:36:57.230","ParentId":null,"OwnerUserId":"9041710","Title":null,"Body":"<p>It's not quite clear what you're asking, mainly because you <strong>never<\/strong> want your model to overfit.<\/p>\n\n<p>Overfitting occurs when the <strong>validation loss<\/strong> increases while <strong>train loss<\/strong> keeps decreasing, which means your model will only work with the training set data and not with the rest of the data as intended. If your validation loss keeps decreasing your model may just need more training epochs to achieve an optimal performance.<\/p>\n\n<p>Also separating part of the Train set as Validation set is called <strong>Hold-out Validation<\/strong>, not Cross-Validation.<\/p>\n"}]}
+{"QuestionId":53164569,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-06T01:25:54.410","AcceptedAnswerId":null,"OwnerUserId":4015074.0,"Title":"How to get the structure of a neural network in tensorflow","Body":"<p>How to extract the structure of a neural network from an existing graph? That is, assuming a network exists in the current graph, I would like to have a dictionary that maps each variable to the variables it is connected to. Tensorboard somehow does this under the hood to visualize the graph but I am not able to deduce it from their code.<\/p>\n","answers":[{"AnswerId":"53165365","CreationDate":"2018-11-06T03:27:46.657","ParentId":null,"OwnerUserId":"6300743","Title":null,"Body":"<ol>\n<li><p>you can save the model.<a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#save_variables\" rel=\"nofollow noreferrer\">links<\/a><\/p><\/li>\n<li><p>you can expore the model to *.pb.<a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#manually_build_a_savedmodel\" rel=\"nofollow noreferrer\">links<\/a><\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53164605,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-06T01:31:33.227","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Google Cloud ML Engine: Using Tensorflow 1.10?","Body":"<p>I've trained a Keras model using Tensorflow 1.10.  However, I only see up to Tensorflow 1.9 on ML Engine when I create a model version in the GUI.  If I use 1.9, I can get predictions back but it gives me pause for concern since the versions are mismatched.<\/p>\n\n<p>Having said that, the docs here seem to indicate that 1.10 is supported: <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/runtime-version-list\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/runtime-version-list<\/a>.<\/p>\n\n<p>How can I create a model version using Tensorflow 1.10?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57106191","CreationDate":"2019-07-19T05:44:48.363","ParentId":null,"OwnerUserId":"11414441","Title":null,"Body":"<p>It seems the document in the link you provided has updated the following:<\/p>\n\n<blockquote>\n  <p>Note: The default AI Platform runtime version used by the AI Platform\n  Training and\n      Prediction API is version 1.0. When you submit training job requests and deploy models,\n      you must specify your desired runtime version. Otherwise, AI Platform uses version 1.0. \n      See how to set runtime versions.<\/p>\n<\/blockquote>\n\n<p>Example:<\/p>\n\n<pre><code>gcloud ai-platform jobs submit training my_job \\\n    --module-name trainer.task \\\n    --job-dir gs:\/\/my\/training\/job\/directory \\\n    --package-path \/path\/to\/my\/project\/trainer \\\n    --python-version 3.5\n    --region us-central1 \\\n    --runtime-version 1.13\n<\/code><\/pre>\n"}]}
+{"QuestionId":53164733,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><convolution>","CreationDate":"2018-11-06T01:51:37.140","AcceptedAnswerId":null,"OwnerUserId":10332319.0,"Title":"Difference between Conv3d vs Conv2d","Body":"<p>I am a little confused with the difference between conv2d and conv3d functions. \nFor example, if I have a stack of N images with H height and W width, and 3 RGB channels. The input to the network can be two forms\nform1: (batch_size, N, H, W, 3) this is a rank 5 tensor\nform2: (batch_size, H, W, 3N ) this is a rank 4 tensor<\/p>\n\n<p>The question is\uff0c if I apply conv3d with M filters with size (N,3,3) to form1 and apply conv2d with M filters with size (3,3)<\/p>\n\n<p>Do they have basicly the same feature operations? I think both of these forms convolve in temporal and spatial dimension.<\/p>\n\n<p>I really appreciate if anyone can help me figure this out. <\/p>\n","answers":[{"AnswerId":"53169648","CreationDate":"2018-11-06T10:04:25.510","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>If you have a stack of images, you have a video. You can not have two input forms. You have either images or videos. For the video case you can use 3D convolution and 2D convolution is not defined for it. If you stack the channels as you mentioned it (3N) the 2D convolution will interpret the stack as one image with a lot of channels, but not as stack.<\/p>\n\n<p>Note here that a 2D convolution with (batch, H, W, Channels) is the same as an 3D convolution with (batch, H, W, Channels, 1). <\/p>\n"}]}
+{"QuestionId":53164779,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><generative-adversarial-network>","CreationDate":"2018-11-06T01:58:49.733","AcceptedAnswerId":null,"OwnerUserId":2414957.0,"Title":"DCGAN: ValueError: Input 0 of layer deconv is incompatible with the layer: expected ndim=4, found ndim=2. Full shape received: [None, 100]","Body":"<p>My code (DCGAN for CelebA dataset) is neither complete nor correct as I can't get passed this error (so please just help me figure how to fix the first error):<\/p>\n\n<pre><code>tf.reset_default_graph()\nLOGDIR = \"logs\"\n\ndef train(args):\n    data_loader = Dataset(args.data_path, args.num_images, args.image_size) \n    #batch_z = sample_z(args.dim_z, args.batch_size)\n    print(args.dim_z)\n    print(args.image_size)\n    X = tf.placeholder(tf.float32, shape=[args.dim_z, 3,args.image_size , args.image_size])\n    Z = tf.placeholder(tf.float32, shape=[None, args.dim_z])\n    G_sample, _ = generator(Z, args)\n    #G_sample, _ = generator(batch_z, args)\n    print(type(G))\n    #D_real, D_real_logits = discriminator(real_batch, args, reuse=False)\n    D_real, D_real_logits = discriminator(X, args, reuse=False)\n    D_fake, D_fake_logits = discriminator(G_sample, args, reuse=True)\n    tf.summary.image('generated image', tf.reshape(G, [-1, 28, 28, 1], 3))\n    d_loss, g_loss = get_losses(d_real_logits, d_fake_logits)\n    z_sum = tf.histogram.summary('z', batch_z)\n    d_sum = tf.histogram.summary('d', D_real)\n    G_sum = tf.histogram.summary('g', G)\n    d_loss_sum = tf.scalar.summary('d_loss', d_loss)\n    g_loss_sum = tf.scalar.summary('g_loss', g_loss)\n    d_sum = tf.merged_summary([z_sum, d_sum, d_loss_sum])\n    g_sum = tf.merged_summary([z_sum, G_sum, g_loss_sum])\n    saver = tf.train.Saver()\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        writer = tf.summary.FileWriter('log', sess.graph)\n\n        for epoch in range(args.n_epoch):\n            for itr, real_batch in enumerate(data_loader.get_nextbatch(args.batch_size)):\n                #fill with your codes\n                sample = sess.run(G_sample, feed_dict={Z:sample_z(args.dim_z, args.batch_size)})\n                d_optimizer, g_optimizer = get_optimizers(agrs.learning_rate, args.beta1, args.beta2)\n                d_step, g_step = optimize(d_optimizer, g_optimizer, d_loss, g_loss)\n                writer = tf.train.SummaryWriter(train_dir, sess.gLraph)  \n                #background = np.ones((10 + (112 + 10) * 5, 10 + (112 + 10) * 8, 3)).astype(np.uint8) * 255\n                #scipy.misc.imsave(os.path.join('log', 'generated-ep-' + str(ep) + '.jpg'), background)\n                tf.summary.image('raw image', tf.reshape())\n                tf.summary.image('generated', image_output)    \n                summary_op = tf.summary.merge_all()\n                writer = tf.summary.FileWriter(logdir, graph)\n                summary = sess.run(summary_op)\n                writer.add_summary(image, step)\n                writer.add_summary(merged_summary, itr)\n                d_loss_summary = tf.summary.scalar(\"Discriminator_Total_Loss\", d_loss)\n                g_loss_summary = tf.summary.scalar(\"Generator_Total_Loss\", g_loss)\n                merged_summary = tf.summary.merge_all()\n                #latest_checkpoint = tf.train.latest_checkpoint(LOGDIR+\"\/checkpoints\/\")\n                writer.add_graph(sess.graph)\n                saver.save(sess, save_path='.\/gan.ckpt')\n\n\ntrain(args)\n<\/code><\/pre>\n\n<p>Basically how to fix the following error?<\/p>\n\n<pre><code>100\n64\n\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-51-6622063733a2&gt; in &lt;module&gt;()\n     54 \n     55 \n---&gt; 56 train(args)\n\n&lt;ipython-input-51-6622063733a2&gt; in train(args)\n      9     X = tf.placeholder(tf.float32, shape=[args.dim_z, 3,args.image_size , args.image_size])\n     10     Z = tf.placeholder(tf.float32, shape=[None, args.dim_z])\n---&gt; 11     G_sample, _ = generator(Z, args)\n     12     #G_sample, _ = generator(batch_z, args)\n     13     print(type(G))\n\n&lt;ipython-input-31-15807fb5df23&gt; in generator(x, args, reuse)\n     10                                              padding='valid',\n     11                                              use_bias=False,\n---&gt; 12                                              name='deconv')\n     13                 batch_norm1=tf.layers.batch_normalization(deconv1,\n     14                                              name = 'batch_norm')\n\n\/share\/pkg\/tensorflow\/r1.10\/install\/py3-gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/convolutional.py in conv2d_transpose(inputs, filters, kernel_size, strides, padding, data_format, activation, use_bias, kernel_initializer, bias_initializer, kernel_regularizer, bias_regularizer, activity_regularizer, kernel_constraint, bias_constraint, trainable, name, reuse)\n   1266       _reuse=reuse,\n   1267       _scope=name)\n-&gt; 1268   return layer.apply(inputs)\n   1269 \n   1270 \n\n\/share\/pkg\/tensorflow\/r1.10\/install\/py3-gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in apply(self, inputs, *args, **kwargs)\n    803       Output tensor(s).\n    804     \"\"\"\n--&gt; 805     return self.__call__(inputs, *args, **kwargs)\n    806 \n    807   def _set_learning_phase_metadata(self, inputs, outputs):\n\n\/share\/pkg\/tensorflow\/r1.10\/install\/py3-gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/base.py in __call__(self, inputs, *args, **kwargs)\n    360 \n    361       # Actually call layer\n--&gt; 362       outputs = super(Layer, self).__call__(inputs, *args, **kwargs)\n    363 \n    364     if not context.executing_eagerly():\n\n\/share\/pkg\/tensorflow\/r1.10\/install\/py3-gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    718 \n    719         # Check input assumptions set before layer building, e.g. input rank.\n--&gt; 720         self._assert_input_compatibility(inputs)\n    721         if input_list and self._dtype is None:\n    722           try:\n\n\/share\/pkg\/tensorflow\/r1.10\/install\/py3-gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _assert_input_compatibility(self, inputs)\n   1422                            'expected ndim=' + str(spec.ndim) + ', found ndim=' +\n   1423                            str(ndim) + '. Full shape received: ' +\n-&gt; 1424                            str(x.shape.as_list()))\n   1425       if spec.max_ndim is not None:\n   1426         ndim = x.shape.ndims\n\nValueError: Input 0 of layer deconv is incompatible with the layer: expected ndim=4, found ndim=2. Full shape received: [None, 100]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53345236","CreationDate":"2018-11-16T20:54:14.650","ParentId":null,"OwnerUserId":"2414957","Title":null,"Body":"<p>I missed the step to preprocess the CelebA dataset images. I did it in the code below and also for once and the above error faded:<\/p>\n\n<p>Basically the line:<\/p>\n\n<pre><code>data_loader.preprocess_and_save_images('preprocessed', 'results_celebA') #preprocess the images once\n<\/code><\/pre>\n\n<p>In the code below:<\/p>\n\n<pre><code>def train(args):\n    tf.reset_default_graph()\n    data_loader = Dataset(args.data_path, args.num_images, args.image_size)\n    data_loader.preprocess_and_save_images('preprocessed', 'results_celebA') #preprocess the images once\n    X = tf.placeholder(tf.float32, shape=[args.batch_size, args.image_size , args.image_size, 3])\n    Z = tf.placeholder(tf.float32, shape=[args.batch_size, 1, 1, args.dim_z])\n\n    G_sample, _ = generator(Z, args)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53164864,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2018-11-06T02:10:48.693","AcceptedAnswerId":null,"OwnerUserId":10560602.0,"Title":"tensorflow object api faster_rcnn_resnet101 training image resizing","Body":"<p>I am currently using the Tensorflow Object API to train my own classes. I am retraining using the faster_rcnn_resnet101_coco model.<\/p>\n\n<p>To create the training data, I used RectLabel to put bounding boxes around objects in approx 100 images. Each image has approx 30 classes in them, for a total of 40 classes present in all the images.<\/p>\n\n<p>My images are 1920\u200a\u00d7\u200a1080 in size. The images are produced by pulling random frames from videos of the objects I would like to detect.<\/p>\n\n<p>My issue is that I am not getting any detections (Tensorboard is not showing any) and I think it is because the training images are being resized and the objects in the images are getting too small. I am using the default faster_rcnn_resnet101_coco.config file with no changes (except for locations to the data).<\/p>\n\n<p>Would it be a good idea to perform a random crop of the images (instead of resizing as below) so as to keep the object size the same for training?<\/p>\n\n<pre><code>image_resizer {\n  keep_aspect_ratio_resizer {\n    min_dimension: 600\n    max_dimension: 1024\n  }\n<\/code><\/pre>\n\n<p>Could there be another issue I am overlooking?<\/p>\n","answers":[{"AnswerId":"53165179","CreationDate":"2018-11-06T02:58:27.440","ParentId":null,"OwnerUserId":"6300743","Title":null,"Body":"<p>I used to deal with a object detection problem,I got nothing at first.I trained the model for two more days ,I got the right results.<\/p>\n\n<p>more training and more data may be helpful.<\/p>\n"},{"AnswerId":"53169317","CreationDate":"2018-11-06T09:45:06.180","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>If you're worried that the resizing is making the objects to small to detect, you can use a larger input resolution. Theoretically you can do it only on your training data, but I'm not sure it would give good result with such tiny training set.\nInstead, you can first fine-tune the pre-trained model with the same dataset (COCO?) on the larger input resolution, and only then fine-tune it on your training data with the larger resolution.\nThis way, the model will theoretically first learn to adapt to the larger resolution, and then will learn your classes.<\/p>\n\n<p>I would also like to side with Friday2013 and suggest to get more training data, possibly more augmentation, and then more training time. Only training longer might not help if you still train on the same small amount of images, since you would get overfitting.<\/p>\n"}]}
+{"QuestionId":53164892,"AnswerCount":0,"Tags":"<keras><google-cloud-ml>","CreationDate":"2018-11-06T02:13:22.153","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Google Cloud ML Engine: Keras model prediction includes name of last layer","Body":"<p>I have a keras model with tensorflow backend hosted on ML Engine.  I ping it for a prediction with:<\/p>\n\n<pre><code>gcloud ml-engine predict --model my_model --version my_version --json-instances my_instances.json\n<\/code><\/pre>\n\n<p>and I get back:<\/p>\n\n<p><code>MY_SECOND_DENSE\n[0.0003, 0.2819, ...]<\/code><\/p>\n\n<p><code>my_second_dense<\/code> is the name of the output layer in my keras model.  Do we know why the name of my model's output layer is included in the prediction result?  Is there a way to avoid this?<\/p>\n\n<p>More to the point, what is the data type of the response here?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Edit:<\/p>\n\n<pre><code>def get_model(model):\n    estimator = tf.keras.estimator.model_to_estimator(model)\n    def serving():\n        inputs = {\n            'reviews': tf.placeholder(shape=[None, 100], dtype='float32')\n        }\n        preproc = inputs.copy()\n        return tf.estimator.export.ServingInputReceiver(preproc, inputs)\n\n    estimator.export_savedmodel('.\/test_export', serving)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53165418,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T03:38:05.107","AcceptedAnswerId":null,"OwnerUserId":9697833.0,"Title":"Order of sess.run([op1, op2...]) in Tensorflow","Body":"<p>I wonder what's the running order of the op list in sess.run(ops_list, ...). for example:\n  for a typical classification scenario: <code>_, loss = sess.run([train_op, loss_op])<\/code>, if <code>train_op<\/code> run first,then the loss is the loss after current backprop. But if <code>loss<\/code> run first, then the loss is the loss before current backprop.\n  anyone help me? Thanks. <\/p>\n","answers":[{"AnswerId":"53170306","CreationDate":"2018-11-06T10:45:08.513","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p><code>loss<\/code> will be the value of the loss <em>before<\/em> the update caused by <code>train_op<\/code>. Note that <code>loss_op<\/code> is an input to the optimizer, so it necessarily goes \"before\" <code>train_op<\/code> in the graph. <code>loss_op<\/code> is computed with the variable values <em>at the beginning<\/em> of the <code>run<\/code> operation. If you wanted to compute the loss <em>after<\/em> <code>train_op<\/code>, you could do that for example using a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/control_dependencies\" rel=\"nofollow noreferrer\"><code>tf.control_dependencies<\/code><\/a> block with the optimizer and computing the loss again, but in that case you would be doing two forward passes of your model on each step, with the associated cost. Usually, if you just want to plot the loss for monitorization or something like that it is okay to use the value from the previous step.<\/p>\n\n<p>For further explanation, in general the order in which TensorFlow operations are executed is only guaranteed to the extent that these operations depend on each other, and it is not related to the order in which the are passed to <code>run<\/code>. In your case, <code>train_op<\/code> depends on <code>loss_op<\/code>, so <code>loss_op<\/code> has to go first. However, in other cases operations do not depend directly on each other, and in that case the order is not guaranteed. Many times that does not really matter, but in some cases it does. Consider the next example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nv = tf.Variable(0)\nv2 = 2 * v\nv_update = v.assign(v + 1)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    for i in range(5):\n        print(sess.run([v_update, v2]))\n<\/code><\/pre>\n\n<p>A run in my computer produced this output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[1, 0]\n[2, 2]\n[3, 4]\n[4, 8]\n[5, 10]\n<\/code><\/pre>\n\n<p>As you see, <code>v2<\/code> is sometimes twice the updated value and sometimes twice the non-updated value. If for example we wanted to make sure that <code>v2<\/code> always runs <em>before<\/em> <code>v_update<\/code> we could do:<\/p>\n\n<pre><code>import tensorflow as tf\n\nv = tf.Variable(0)\nv2 = 2 * v\nwith tf.control_dependencies([v2]):\n    v_update = v.assign(v + 1)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    for i in range(5):\n        print(sess.run([v_update, v2]))\n<\/code><\/pre>\n\n<p>Which consistently produces:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[1, 0]\n[2, 2]\n[3, 4]\n[4, 6]\n[5, 8]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53165443,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T03:41:12.193","AcceptedAnswerId":null,"OwnerUserId":10610947.0,"Title":"tensorflow performance bottleneck on IteratorGetNext and less efficient than tf.FIFOQueue","Body":"<p>I have training a speech recognition network with tensorflow and use tf.FIFIQueue to feed data before. the training speed didn't meet the requirements.and the speed is slow down when i use tf.data.TFRecordDataset and transform binary file to tfrecords, i have refer to the <a href=\"https:\/\/i.stack.imgur.com\/5I5tw.png\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/performance\/datasets<\/a>, the pre-process code as follow\uff1a<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>def read_and_decode(loader, handle, num_epochs=1):\n  \"\"\" read tfrecord format data\"\"\"\n  batch_size = loader.batch_size()\n  feature_size = model_settings['fingerprint_size']\n  def parse_exmp(serialized_example):\n    features = tf.parse_single_example(serialized_example, features={\n                                                'feature':  tf.VarLenFeature(tf.float32),\n                                                'label':    tf.VarLenFeature(tf.int64),\n                                                'mask':     tf.VarLenFeature(tf.int64),\n                                                'length':   tf.FixedLenFeature((),tf.int64)\n                                                })\n    length = tf.cast(features['length'], tf.int32)\n    feature = tf.sparse_tensor_to_dense(features['feature'])\n    feature = tf.reshape(feature, [length, feature_size])\n    label = tf.sparse_tensor_to_dense(features['label'])\n    mask = tf.sparse_tensor_to_dense(features['mask'])\n    return feature, label, mask, length\n  '''\n  filenames = tf.data.Dataset.list_files(\".\/train_input\/tfrecords_file\/train_dataset_*.tfrecords\")\n  dataset = filenames.apply(tf.contrib.data.parallel_interleave(\\\n            lambda filename: tf.data.TFRecordDataset(filename),cycle_length=4))\n  '''\n  filenames = ['.\/train_input\/tfrecords_file\/train_dataset_%d.tfrecords'%i for i in range(cpu_count()-1)]\n  dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads=10)\n\n  dataset = dataset.map(parse_exmp, num_parallel_calls=48)\n  dataset = dataset.prefetch(buffer_size=batch_size)\n  dataset = dataset.apply(tf.contrib.data.shuffle_and_repeat(10000, num_epochs))\n  dataset = dataset.padded_batch(batch_size, padded_shapes=([None, feature_size],[None],[None],[]))\n  train_iterator = dataset.make_initializable_iterator()\n\n  iterator = tf.data.Iterator.from_string_handle(handle, \\\n    dataset.output_types, dataset.output_shapes)\n\n  batch_data, batch_label, batch_mask, batch_length = iterator.get_next()\n  batch_data =[tf.transpose(data, (1,0,2)) for data in tf.split(batch_data, FLAGS.gpu_num, axis=0)]\n  batch_label = tf.split(batch_label, FLAGS.gpu_num, axis=0)\n  batch_mask = tf.split(batch_mask, FLAGS.gpu_num, axis=0)\n\n  return train_iterator,batch_data,[tf.transpose(label, (1,0)) for label in batch_label], [tf.transpose(mask, (1,0)) for mask in batch_mask], batch_length\n<\/code><\/pre>\n\n<p>here is my Speed chart:\n <a href=\"https:\/\/i.stack.imgur.com\/5I5tw.png\" rel=\"nofollow noreferrer\">enter image description here<\/a>\nwe can see that tf.Dataset is slow than tf.FIFOQueue. and the timeline for tf.FIFOQueue:<a href=\"https:\/\/i.stack.imgur.com\/loevD.png\" rel=\"nofollow noreferrer\">enter image description here<\/a> the timeline for tf.Dataset:\n<a href=\"https:\/\/i.stack.imgur.com\/AKHXH.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>we can see IteratorGetNext is too long about 200ms, tf.Dataset is also slow even no IteratorGetNext.\nmy question is why tf.Dataset is so long than tf.FIFOQueue even i have Optimize the code. \nthanks.<\/p>\n","answers":[]}
+{"QuestionId":53165492,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2018-11-06T03:48:25.157","AcceptedAnswerId":53169528.0,"OwnerUserId":3959965.0,"Title":"How to use filter tags in tensorboard?","Body":"<p>I'm using tensorboard ver:1.10.0 scalars to visualize the loss. How to use these filter tags field? Can I use regular expression to filter my x and y axis, eg: y>100? Or do I need to add tags in <code>writer.add_summary<\/code>? Your help is much appreciated.<\/p>\n\n<p>Please find the screenshot below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9N3LT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9N3LT.png\" alt=\"Tensorboard scalar\"><\/a><\/p>\n","answers":[{"AnswerId":"53169528","CreationDate":"2018-11-06T09:57:08.693","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>The tag is specified by the names you give, when calling <code>tf.summary.scalar(\"loss\", tensor)<\/code>. For this example and yours, too, \"loss\" is the tag. It makes sense, when you have multiple summaries of several different runs in your tfboard. <\/p>\n"}]}
+{"QuestionId":53165531,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T03:54:53.107","AcceptedAnswerId":null,"OwnerUserId":7966026.0,"Title":"ModuleNotFoundError: No module named 'tensorflow' after pip install -- following TF official install docs","Body":"<p>I have recently installed TensorFlow using the pip install instructions on the TensorFlow official website. Due to technical issues, I downgraded to Python version 3.6.2. However, when I entered IDLE, opened an additional file and entered:<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>Python responded with the message:<\/p>\n\n<pre><code>   Traceback (most recent call last):\n  File \"\/Users\/gg-mac\/Documents\/firsttensorflow.py\", line 1, in  &lt;module&gt;\n import tensorflow as tf\nModuleNotFoundError: No module named 'tensorflow'\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>Is there any way to resolve this issue? I am simply trying to know where to run Tensorflow, as it is not specified anywhere on the website. I am developing an audio recognition program, if that is relevant. Thank you for any assistance you can provide.<\/p>\n\n<p>Sincerely,\nSuren Grigorian<\/p>\n","answers":[{"AnswerId":"53165743","CreationDate":"2018-11-06T04:27:41.973","ParentId":null,"OwnerUserId":"9108532","Title":null,"Body":"<p>This is hard to diagnose without more information, since we cannot test the absence of it. Perhaps try this:<\/p>\n\n<pre><code>pip3 install tensorflow\n<\/code><\/pre>\n\n<p>and try again, or perhaps run<\/p>\n\n<pre><code>python3 \/Users\/gg-mac\/Documents\/firsttensorflow.py\n<\/code><\/pre>\n\n<p>from the Terminal (looks like you're running Mac OS).<\/p>\n"}]}
+{"QuestionId":53165544,"AnswerCount":0,"Tags":"<python><keras><classification>","CreationDate":"2018-11-06T03:57:18.740","AcceptedAnswerId":null,"OwnerUserId":9673278.0,"Title":"VGG16 Keras Implementation Model.Fit Error","Body":"<p>I have a data set with 32X32X3 input sizes. I would like to classify this data set with the VGG16 Keras Implementation. But at the Train stage I get the following error. I'd appreciate it if you could help. Thanks...<\/p>\n\n<p>Error : <\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"VGG16_3.py\", line 211, in &lt;module&gt;\n   model = DroneVgg()\nFile \"VGG16_3.py\", line 31, in __init__\n   self.model = self.train(self.model)\nFile \"VGG16_3.py\", line 204, in train\n   validation_data=(x_test, y_test),callbacks=[reduce_lr],verbose=2)\nFile \"\/truba\/home\/atkorez\/anaconda3\/envs\/CPU_Env\/lib\/python3.6\/site- packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n   return func(*args, **kwargs)\nFile \"\/truba\/home\/atkorez\/anaconda3\/envs\/CPU_Env\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1418, in fit_generator\n   initial_epoch=initial_epoch)\nFile \"\/truba\/home\/atkorez\/anaconda3\/envs\/CPU_Env\/lib\/python3.6\/site-packages\/keras\/engine\/training_generator.py\", line 144, in fit_generator\n  val_x, val_y, val_sample_weight)\nFile \"\/truba\/home\/atkorez\/anaconda3\/envs\/CPU_Env\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 789, in _standardize_user_data\n  exception_prefix='target')\nFile \"\/truba\/home\/atkorez\/anaconda3\/envs\/CPU_Env\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\", line 128, in standardize_input_data\n  'with shape ' + str(data_shape))\nValueError: Error when checking target: expected activation_15 to have 2 dimensions, but got array with shape (329038, 6, 6)\n<\/code><\/pre>\n\n<p><strong>Edit<\/strong>\nI sent 2 messages because I received too many code errors in a single message.<\/p>\n\n<p>Python code here :<\/p>\n\n<pre><code>def veri_yukle():\n\n   x_train = np.load('\/truba\/home\/atkorez\/CapsNet\/x_train.npy')\n   x_test = np.load('\/truba\/home\/atkorez\/CapsNet\/x_test.npy')\n   y_train = np.load('\/truba\/home\/atkorez\/CapsNet\/y_train.npy')\n   y_test = np.load('\/truba\/home\/atkorez\/CapsNet\/y_test.npy')\n\n   return (x_train, y_train), (x_test, y_test)\n\nclass DroneVgg:\ndef __init__(self,train=True):\n    self.num_classes = 6\n    self.weight_decay = 0.0005\n    self.x_shape = [32,32,1]\n\n    self.model = self.build_model()\n    if train:\n        self.model = self.train(self.model)\n    else:\n        self.model.load_weights('DroneVgg.h5')\n\n\ndef build_model(self):\n\n    model = Sequential()\n    weight_decay = self.weight_decay\n\n    model.add(Conv2D(64, (3, 3), padding='same',\n                     input_shape=self.x_shape,kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.3))\n\n    model.add(Conv2D(64, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(128, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(256, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.4))\n\n    model.add(Conv2D(512, (3, 3), padding='same',kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.5))\n\n    model.add(Flatten())\n    model.add(Dense(512,kernel_regularizer=regularizers.l2(weight_decay)))\n    model.add(Activation('relu'))\n    model.add(BatchNormalization())\n\n    model.add(Dropout(0.5))\n    model.add(Dense(self.num_classes))\n    model.add(Activation('softmax'))\n    return model\n\n\ndef normalize(self,X_train,X_test):\n    mean = np.mean(X_train,axis=(0,1,2,3))\n    std = np.std(X_train, axis=(0, 1, 2, 3))\n    X_train = (X_train-mean)\/(std+1e-7)\n    X_test = (X_test-mean)\/(std+1e-7)\n    return X_train, X_test\n\n\ndef train(self,model):\n\n    #training parameters\n    batch_size = 128\n    maxepoches = 250\n    learning_rate = 0.1\n    lr_decay = 1e-6\n    lr_drop = 20\n    # The data, shuffled and split between train and test sets:\n    (x_train, y_train), (x_test, y_test) = veri_yukle()\n    print('Veriler Yuklendi')\n\n    x_train = x_train.astype('float32')\n    x_test = x_test.astype('float32')\n    x_train, x_test = self.normalize(x_train, x_test)\n\n    y_train = keras.utils.to_categorical(y_train, self.num_classes)\n    y_test = keras.utils.to_categorical(y_test, self.num_classes)\n\n    def lr_scheduler(epoch):\n        return learning_rate * (0.5 ** (epoch \/\/ lr_drop))\n    reduce_lr = keras.callbacks.LearningRateScheduler(lr_scheduler)\n\n    #data augmentation\n    datagen = ImageDataGenerator(\n        featurewise_center=False,  # set input mean to 0 over the dataset\n        samplewise_center=False,  # set each sample mean to 0\n        featurewise_std_normalization=False,  # divide inputs by std of the dataset\n        samplewise_std_normalization=False,  # divide each input by its std\n        zca_whitening=False,  # apply ZCA whitening\n        rotation_range=15,  # randomly rotate images in the range (degrees, 0 to 180)\n        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)\n        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)\n        horizontal_flip=True,  # randomly flip images\n        vertical_flip=False)  # randomly flip images\n    # (std, mean, and principal components if ZCA whitening is applied).\n    datagen.fit(x_train)\n\n\n\n    #optimization details\n    sgd = optimizers.SGD(lr=learning_rate, decay=lr_decay, momentum=0.9, nesterov=True)\n    model.compile(loss='categorical_crossentropy', optimizer=sgd,metrics=['accuracy'])\n\n\n    # training process in a for loop with learning rate drop every 25 epoches.\n\n    historytemp = model.fit_generator(datagen.flow(x_train, y_train,\n                                     batch_size=batch_size),\n                        steps_per_epoch=x_train.shape[0] \/\/ batch_size,\n                        epochs=maxepoches,\n                        validation_data=(x_test, y_test),callbacks=[reduce_lr],verbose=2)\n    model.save_weights('DroneVgg.h5')\n    return model\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53165818,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2018-11-06T04:38:58.950","AcceptedAnswerId":null,"OwnerUserId":7057994.0,"Title":"Keras layers explaination","Body":"<p>I want to get a deep idea about how this keras layers works in a model. What does each layer doing in the model etc. I followed kers documentation and information isn't enough. If any of you know place to get more knowledge let me know.Thanks in advance<\/p>\n","answers":[{"AnswerId":"53166221","CreationDate":"2018-11-06T05:29:20.060","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Keras layers are widely used CNN, DNN and RNN layers. There is atleast one research paper for each of them and there is a lot of educational material out there. If you are really curious you could look at keras' code. Some links for you:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/tree\/master\/keras\/layers\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/tree\/master\/keras\/layers<\/a><\/p>\n\n<p><a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow noreferrer\">http:\/\/cs231n.github.io\/convolutional-networks\/<\/a><\/p>\n\n<p><a href=\"https:\/\/leonardoaraujosantos.gitbooks.io\/artificial-inteligence\" rel=\"nofollow noreferrer\">https:\/\/leonardoaraujosantos.gitbooks.io\/artificial-inteligence<\/a><\/p>\n\n<p><a href=\"http:\/\/www.jmlr.org\/papers\/volume15\/srivastava14a\/srivastava14a.pdf\" rel=\"nofollow noreferrer\">http:\/\/www.jmlr.org\/papers\/volume15\/srivastava14a\/srivastava14a.pdf<\/a><\/p>\n"}]}
+{"QuestionId":53165823,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-06T04:39:51.007","AcceptedAnswerId":null,"OwnerUserId":8222067.0,"Title":"Trained Keras Model fails to load with load_model","Body":"<p>I have trained a Keras model with Tensorflow backend. It was saved with model.save. I now want to reload the model using model_load, however, I get the following error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"&lt;ipython-input-235-387752c910a4&gt;\", line 1, in &lt;module&gt;\n    load_model('MyModel.h5')\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\models.py\", line 243, in load_model\n    model = model_from_config(model_config, custom_objects=custom_objects)\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\models.py\", line 317, in model_from_config\n    return layer_module.deserialize(config, custom_objects=custom_objects)\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 144, in deserialize_keras_object\n    list(custom_objects.items())))\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\engine\\topology.py\", line 2514, in from_config\n    process_layer(layer_data)\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\engine\\topology.py\", line 2500, in process_layer\n    custom_objects=custom_objects)\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 144, in deserialize_keras_object\n    list(custom_objects.items())))\n\n  File \"C:\\Anaconda\\envs\\tensorflow\\lib\\site-packages\\keras\\models.py\", line 1367, in from_config\n    if 'class_name' not in config[0] or config[0]['class_name'] == 'Merge':\n\nKeyError: 0\n<\/code><\/pre>\n\n<p>From what I read, there seems to be a bug in Keras when a model that was trained with an older version of Keras is loaded with a recent version. So there might be a version mismatch. However, I couldn't find a report that corresponds to my situation. Downgrading Keras or retraining is not an option.<\/p>\n\n<p>Did anyone have this issue and maybe even found a solution? I would appreciate it a lot!<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53179802","CreationDate":"2018-11-06T20:49:34.350","ParentId":null,"OwnerUserId":"8222067","Title":null,"Body":"<p>For future reference: It is an issue in the config files. Keras 2.2.4 has a fix for this: <\/p>\n\n<p>Keras 2.2.4<\/p>\n\n<p>@fchollet fchollet released this on Oct 3 \u00b7 79 commits to master since this release\nAssets 2<\/p>\n\n<p>This is a bugfix release, addressing two issues:<\/p>\n\n<p>Ability to save a model when a file with the same name already exists.\nIssue with loading legacy config files for the Sequential model.<\/p>\n\n<p>So I ended up creating a new virtual environment with the most recent TF and Keras versions.<\/p>\n"}]}
+{"QuestionId":53165953,"AnswerCount":1,"Tags":"<amazon-web-services><keras><amazon-sagemaker>","CreationDate":"2018-11-06T04:57:19.330","AcceptedAnswerId":53165954.0,"OwnerUserId":6164611.0,"Title":"ValueError: Tensor is not an element of this graph, when hosting a model in Sagemaker with Gunicorn and Flask and Keras","Body":"<p>I created a time series predictor with Keras and  Dockerized the model with with Flash and Gunicorn as per AWS docs. I am loading the serialized model with this code.<\/p>\n\n<pre><code>@classmethod\ndef get_model(cls):\n    if cls.model == None:\n        cls.model = load_model('\/opt\/ml\/bitcoin_model.h5')\n    return cls.model\n<\/code><\/pre>\n\n<p>Then I used the predict method to produce the results , the dockerized container is working perfectly in the local environment , but when I try to host the model in sagemaker it produces this error.<\/p>\n\n<pre><code>ValueError: Tensor Tensor(\"dense_1\/BiasAdd:0\", shape=(?, 1), dtype=float32) is not an element of this graph.\n<\/code><\/pre>\n\n<p>So how can I resolve this issue ?<\/p>\n","answers":[{"AnswerId":"53165954","CreationDate":"2018-11-06T04:57:19.330","ParentId":null,"OwnerUserId":"6164611","Title":null,"Body":"<p>The issue was resolved by calling _make_predict_function() method in the model load phase.<\/p>\n\n<pre><code>@classmethod\ndef get_model(cls):\n    if cls.model == None:\n        cls.model = load_model('\/opt\/ml\/bitcoin_model.h5')\n        cls.model._make_predict_function()\n    return cls.model\n<\/code><\/pre>\n\n<p>Bug Reference : <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6462\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/6462<\/a><\/p>\n"}]}
+{"QuestionId":53165990,"AnswerCount":1,"Tags":"<python-3.x><ubuntu-16.04><pytorch>","CreationDate":"2018-11-06T05:02:08.160","AcceptedAnswerId":53935803.0,"OwnerUserId":9853650.0,"Title":"PyTorch is installed but not imported","Body":"<p>I am trying to build PyTorch.\nReference site\uff1a<a href=\"https:\/\/github.com\/hughperkins\/pytorch\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hughperkins\/pytorch<\/a><\/p>\n\n<p>but, When we performed unit test, The following error occur.<\/p>\n\n<pre><code>ImportError while importing test module '\/home\/usr2\/pytorch\/test\/testByteTensor.py'.\nHint: make sure your test modules\/packages have valid Python names.\nTraceback:\ntest\/testByteTensor.py:2: in &lt;module&gt;\n    import PyTorch\nE   ImportError: No module named 'PyTorch'\n__________________ ERROR collecting test\/testDoubleTensor.py ___________________\n<\/code><\/pre>\n\n<p>I understand that PyTorch is not imported.\nbut It is confirmed that pytorch is installed.\nIs there a way to solve this problem?<\/p>\n\n<p>environment<\/p>\n\n<pre><code>ubuntu 16.04\npython3.5\ncuda9.2\n<\/code><\/pre>\n","answers":[{"AnswerId":"53935803","CreationDate":"2018-12-26T18:36:22.560","ParentId":null,"OwnerUserId":"9569969","Title":null,"Body":"<p>In fact, you should do <code>import torch<\/code> instead of <code>import PyTorch<\/code><br>\nHere is what's work for me: (I installed it using conda)<\/p>\n\n<pre><code>&gt;&gt;&gt; import torch\n\n&gt;&gt;&gt; torch.version\n&gt;&gt;&gt; &lt;module 'torch.version' from '\/home\/koke_cacao\/miniconda3\/envs\/ml\/lib\/python3.6\/site-packages\/torch\/version.py'&gt;\n\n&gt;&gt;&gt; print(torch.__version__)\n&gt;&gt;&gt; 0.4.1.post2\n\n&gt;&gt;&gt; a = torch.FloatTensor(2,3)\n&gt;&gt;&gt; tensor([[-7.4368e-13,  3.0911e-41, -9.6122e-13],\n    [ 3.0911e-41, -7.3734e-13,  3.0911e-41]])\n<\/code><\/pre>\n\n<p>Edit: The version works with no problem at all for me. But if you insist to perform the unit test, maybe other people can solve your problem.<\/p>\n"}]}
+{"QuestionId":53166615,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><tensorboard>","CreationDate":"2018-11-06T06:13:39.780","AcceptedAnswerId":null,"OwnerUserId":1935883.0,"Title":"Tensorboard projector visualisation - PCA keeps loading","Body":"<p>Tensorboard projector visualisation - PCA keeps hanging. <\/p>\n\n<p>I wrote a simple NN to predict the class type of iris dataset. \nNN model works fine. <\/p>\n\n<pre><code>import pandas as pd \nimport numpy as np\n\nfrom sklearn.datasets import load_iris\nfrom sklearn.model_selection import train_test_split\nfrom sklearn import preprocessing\n\nimport tensorflow as tf\nimport tensorflow as tf\nfrom tensorflow import keras\n\niris_data = load_iris()\nx = pd.DataFrame(iris_data.data, columns=iris_data.feature_names)\ny = pd.DataFrame(iris_data.target, columns=['class'])\n\nencoder = preprocessing.OneHotEncoder(categories='auto')\nencoder.fit(y)\n#Transform\ny_enc = encoder.transform(y).toarray()\n\nx_train, x_test, y_train, y_test = train_test_split(x, y_enc)\n\n\nmodel = keras.Sequential()\nmodel.add(keras.layers.Dense(8, name='input_layer', activation=tf.nn.relu, input_shape=(x_train.shape[1],)))\nmodel.add(keras.layers.Dense(4, name='hidden_layer', activation=tf.nn.relu))\nmodel.add(keras.layers.Dense(3, name='out_layer', activation=tf.nn.softmax))\n\nmodel.compile(optimizer=tf.keras.optimizers.Adam(0.005),\n              loss=keras.losses.binary_crossentropy,\n              metrics=[keras.metrics.categorical_accuracy])\nmodel.fit(x_train, y_train, epochs=50, verbose=0)\nresult = model.predict(x_test)\n<\/code><\/pre>\n\n<p>Now I am trying to visualise the output of the test set. \nBelow is the code for Tensorboard projector. \nI don't know what I am missing but PCA keeps loading even after starting the Tensorboard several minutes ago. <\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.contrib.tensorboard.plugins import projector\n\nimport numpy as np\nimport os\n\nLOG_DIR = 'logs'  # FULL PATH HERE!!!\n\nmetadata_file = os.path.join(LOG_DIR, 'metadata.tsv')\nwith open(metadata_file, 'w') as f:\nf.write('{}\\t{}\\n'.format('class_name','class_id'))\nwith open(metadata_file, 'a') as f:\nfor i in range(len(y_test)):\n    c = np.nonzero(y_test[i])[0][0]\n    f.write('{}\\t{}\\n'.format(iris_data.target_names[c],c))\n\n\nembedding_var = tf.Variable(result,  name='final_layer_embedding')\nsess = tf.Session()\nsess.run(embedding_var.initializer)\nsummary_writer = tf.summary.FileWriter(LOG_DIR)\nconfig = projector.ProjectorConfig()\nembedding = config.embeddings.add()\nembedding.tensor_name = embedding_var.name\n\nembedding.metadata_path = 'metadata.tsv'\n\nprojector.visualize_embeddings(summary_writer, config)\nsaver = tf.train.Saver([embedding_var])\nsaver.save(sess, os.path.join(LOG_DIR, 'model.ckpt'), 1)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3tPJo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3tPJo.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I googled to understand what I am doing wrong but I could not fix it. Despite being my model is small very I could not visualise. Any help to resolve this problem would be highly appreciable. <\/p>\n","answers":[{"AnswerId":"53373700","CreationDate":"2018-11-19T11:29:59.227","ParentId":null,"OwnerUserId":"1935883","Title":null,"Body":"<p>I am answering for my own question. <\/p>\n\n<p>As suggested by @Tay2510 in the comment.<\/p>\n\n<p>Same code works after upgrading <strong>tensorboard<\/strong> version to <strong>1.12.0<\/strong> from 1.11.0.<\/p>\n\n<p>However my <strong>tensorflow<\/strong> version remained the the same to <strong>1.11.0<\/strong>. <\/p>\n"}]}
+{"QuestionId":53167108,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><tensor>","CreationDate":"2018-11-06T07:00:20.663","AcceptedAnswerId":53167709.0,"OwnerUserId":6407393.0,"Title":"How to conditionally scale values in Keras Lambda layer?","Body":"<p>The input tensor <code>rnn_pv<\/code> is of shape <code>(?, 48, 1)<\/code>. I want to scale every element in this tensor, so I try to use <code>Lambda<\/code> layer as below:<\/p>\n\n<pre><code>rnn_pv_scale = Lambda(lambda x: 1 if x &gt;=1000 else x\/1000.0 )(rnn_pv)\n<\/code><\/pre>\n\n<p>But it comes the error:<\/p>\n\n<pre><code>TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.\n<\/code><\/pre>\n\n<p>So what is the proper way to realize this function ?<\/p>\n","answers":[{"AnswerId":"53167709","CreationDate":"2018-11-06T07:47:40.423","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can't use Python control flow statements such as if-else statements to perform conditional operations in the definition of a model. Instead you need to use methods defined in Keras backends. Since you are using TensorFlow as the backend you can use <code>tf.where()<\/code> to achieve that:<\/p>\n\n<pre><code>import tensorflow as tf\n\nscaled = Lambda(lambda x: tf.where(x &gt;= 1000, tf.ones_like(x), x\/1000.))(input_tensor)\n<\/code><\/pre>\n\n<p>Alternatively, to support all the backends, you can create a mask to do this:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef rescale(x):\n    mask = K.cast(x &gt;= 1000., dtype=K.floatx())\n    return mask + (x\/1000.0) * (1-mask)\n\n#...\nscaled = Lambda(rescale)(input_tensor)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Update:<\/strong> An alternative way to support all the backends is to use <code>K.switch<\/code> method:<\/p>\n\n<pre><code>from keras import backend as K\n\nscaled = Lambda(lambda x: K.switch(x &gt;= 1000., K.ones_like(x), x \/ 1000.))(input_tensor)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53167161,"AnswerCount":1,"Tags":"<python><tensorflow><gmm>","CreationDate":"2018-11-06T07:03:33.130","AcceptedAnswerId":null,"OwnerUserId":7403477.0,"Title":"Tensorflow MixtureSameFamily and gaussian mixture model","Body":"<p>I am really new to Tensorflow as well as gaussian mixture model.<\/p>\n\n<p>I have recently used tensorflow.contrib.distribution.MixtureSameFamily class for predicting probability density function which is derived from gaussian mixture of 4 components.<\/p>\n\n<p>When I plotted the predicted density function using \"prob()\" function as Tensorflow tutorial explains, I found the plotted pdf with only one mode. I expected to see 4 modes as the mixture components are 4.<\/p>\n\n<p>I would like to ask whether Tensorflow uses any global mode predicting algorithm in their MixtureSameFamily class. If not, I would also like to know how MixtureSameFamily class forms the pdf with statistical values.  <\/p>\n\n<p>Thank you very much.   <\/p>\n","answers":[{"AnswerId":"53184428","CreationDate":"2018-11-07T06:17:27.807","ParentId":null,"OwnerUserId":"7403477","Title":null,"Body":"<p>I found an answer for above question thanks to my collegue. <\/p>\n\n<p>The 4 components of gaussian mixture have had very similar means that the mixture seems like it has only one mode.<\/p>\n\n<p>If I put four explicitly different values as means to the MixtureSameFamily class, I could get a plot of gaussian mixture with 4 different modes.<\/p>\n\n<p>Thank you very much for reading this.     <\/p>\n"}]}
+{"QuestionId":53167302,"AnswerCount":1,"Tags":"<tensorflow><keras-layer>","CreationDate":"2018-11-06T07:14:54.260","AcceptedAnswerId":null,"OwnerUserId":8384504.0,"Title":"Does tf.keras.layers.Conv2D as first layer in model truly need input_shape?","Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Conv2D\" rel=\"nofollow noreferrer\">the official document<\/a> on <code>tf.keras.layers.Conv2D<\/code>,<\/p>\n\n<blockquote>\n  <p>When using this layer as the first layer in a model, provide the keyword argument input_shape (tuple of integers, does not include the sample axis), e.g. input_shape=(128, 128, 3) for 128x128 RGB pictures in data_format=\"channels_last\".<\/p>\n<\/blockquote>\n\n<p>but actually without input_shape it does work in both graph execution and eager execution environment.<\/p>\n\n<p>In graph execution, <\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Conv2D, Flatten, Dense\n\nclass CNN(tf.keras.Model):\n\n    def __init__(self):\n        super(CNN, self).__init__()\n        self.conv = Conv2D(1, 3, padding='same', data_format='channels_first')\n        self.flatten = Flatten()\n        self.dense = Dense(1)\n\n    def call(self, inputs):\n        x = self.conv(inputs)\n        x = self.flatten(x)\n        return self.dense(x)\n\n\ncnn = CNN()\ninputs = tf.random_uniform([2, 3, 16, 16])\noutputs = cnn(inputs)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    outputs = sess.run(outputs)\n    print(outputs)\n<\/code><\/pre>\n\n<p>works without any error and in eager execution,<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Conv2D, Flatten, Dense\ntf.enable_eager_execution()\n\nclass CNN(tf.keras.Model):\n    def __init__(self):\n        super(CNN, self).__init__()\n        self.conv = Conv2D(1, 3, padding='same', data_format='channels_first')\n        self.flatten = Flatten()\n        self.dense = Dense(1)\n\n    def call(self, inputs):\n        x = self.conv(inputs)\n        x = self.flatten(x)\n        return self.dense(x)\n\n\ncnn = CNN()\ninputs = tf.random_uniform([2, 3, 16, 16])\noutputs = cnn(inputs)\nprint(outputs)\n<\/code><\/pre>\n\n<p>also does.<\/p>\n\n<p>Q1: Does <code>tf.keras.layers.Conv2D<\/code> as the first layer in a model truly need to specifying <code>input_shape<\/code>?<\/p>\n\n<p>Q2: If not, when is it needed and why is it mentioned so in the official document?<\/p>\n\n<p>UPDATE1:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_layers#layers_common_sets_of_useful_operations\" rel=\"nofollow noreferrer\">Tutorial on tf.keras<\/a> says <\/p>\n\n<blockquote>\n  <p>The number of input dimensions is often unnecessary, as it can be inferred\n  the first time the layer is used, but it can be provided if you want to \n  specify it manually, which is useful in some complex models.<\/p>\n<\/blockquote>\n\n<p>UPDATE2:\n<code>git blame<\/code> of docstring in TensorFlow source revealed that this document is copied from Keras API (which is not TensorFlow keras API).<\/p>\n","answers":[{"AnswerId":"53169453","CreationDate":"2018-11-06T09:52:35.433","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>Convolution does generally does not need an input shape. Actually you can feed the same network different input shapes, but it is much faster, when you give tensorflow an input shape. I think that the reason, why its stated in the docs. <\/p>\n\n<p>The original method of tensorflow does not even have an argument for input shapes.<\/p>\n"}]}
+{"QuestionId":53168198,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><neural-network>","CreationDate":"2018-11-06T08:25:06.690","AcceptedAnswerId":null,"OwnerUserId":10587031.0,"Title":"how to processing image like mnist using python?","Body":"<p>I'm making digit recoginition model using mnist.\nlearning is quite well, model's accuarcy is about 98.6%.\nBut, when i test other images(not mnist image), the accuarcy is very low<\/p>\n\n<p>This is my code processing the image.\nthe image what i want to use is IAM dataset. here is a link.\n<a href=\"http:\/\/www.fki.inf.unibe.ch\/databases\/iam-handwriting-database\" rel=\"nofollow noreferrer\">http:\/\/www.fki.inf.unibe.ch\/databases\/iam-handwriting-database<\/a><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom PIL import Image\nfrom glob import glob\nfrom skimage import color\nfrom skimage import io\nimport matplotlib.cm as cm\nfrom scipy.misc import imresize\n\n\ndef read_image(path):\n    image = io.imread(path, as_grey= 1)\n    image_resized = imresize(image, [28, 28])\n    image_reshaped = image_resized.reshape([784])\n    image_reshaped = np.array(image_reshaped, dtype= np.float32)\n\n    for i in range(len(image_reshaped)):\n        if image_reshaped[i] &gt;= 255:\n            image_reshaped[i] = 0\n        elif image_reshaped[i] != 255:\n            image_reshaped[i] \/= 255.0000\n    return image_reshaped\n\ntesting_image = read_image(path)\nplt.imshow(a.reshape((28, 28)))\n<\/code><\/pre>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/yW1d8.png\" alt=\"enter image description here\"><\/p>\n\n<p>this is my processing image(sorry for link, i'm newbee, so i can't upload directly)<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/RKXnD.png\" alt=\"enter image description here\"><\/p>\n\n<p>and this is mnist image.<\/p>\n\n<p>so, my question is, what's the different of my images and mnist images,\nand how can i make my images like mnist images?<\/p>\n","answers":[{"AnswerId":"53175509","CreationDate":"2018-11-06T15:59:00.360","ParentId":null,"OwnerUserId":"10574412","Title":null,"Body":"<p>The model related to MNIST isn't able to generalize to different images! As you can see the '7' in your data is completely different from the one in the MNIST set. I would suggest you to create a new ANN and train it with the AIM dataset.<\/p>\n"}]}
+{"QuestionId":53168648,"AnswerCount":1,"Tags":"<keras><constraints><loss-function><unsupervised-learning>","CreationDate":"2018-11-06T09:01:20.413","AcceptedAnswerId":null,"OwnerUserId":6215584.0,"Title":"keras: unsupervised learning with external constraint","Body":"<p>I have to train a network on unlabelled data of binary type (True\/False), which sounds like unsupervised learning. This is what the normalised data look like:<\/p>\n\n<pre><code>array([[-0.05744527, -1.03575495, -0.1940105 , -1.15348956, -0.62664491,\n    -0.98484037],\n   [-0.05497629, -0.50935675, -0.19396862, -0.68990988, -0.10551919,\n    -0.72375012],\n   [-0.03275552,  0.31480204, -0.1834951 ,  0.23724946,  0.15504367,\n     0.29810553],\n   ...,\n   [-0.05744527, -0.68482282, -0.1940105 , -0.87534175, -0.23580062,\n    -0.98484037],\n   [-0.05744527, -1.50366446, -0.1940105 , -1.52435329, -1.14777063,\n    -0.98484037],\n   [-0.05744527, -1.26970971, -0.1940105 , -1.33892142, -0.88720777,\n    -0.98484037]])\n<\/code><\/pre>\n\n<p>However, I do have a <em>constraint on the total number of True<\/em> labels in my data. This doesn't mean I can build a classical custom loss function in Keras taking <code>(y_true, y_pred)<\/code> arguments as required: my external constraint is just on the predicted total of <code>True<\/code> and <code>False<\/code>, not on the individual labels.<\/p>\n\n<p>My question is whether there is a somewhat \"standard\" approach to this kind of problems, and how that is implementable in Keras.<\/p>\n\n<p><strong>POSSIBLE SOLUTION<\/strong><\/p>\n\n<p>Should I assign <code>y_true<\/code> randomly as 0\/1, have a network return <code>y_pred<\/code> as 1\/0 with a sigmoid activation function, and then define my loss function as<\/p>\n\n<pre><code>sum_y_true = 500 # arbitrary constant known a priori\n\ndef loss_function(y_true, y_pred):\n    loss = np.abs(y_pred.sum() - sum_y_true)\n    return loss\n<\/code><\/pre>\n","answers":[{"AnswerId":"53377223","CreationDate":"2018-11-19T14:55:24.573","ParentId":null,"OwnerUserId":"6215584","Title":null,"Body":"<p>In the end, I went with the following solution, which worked.<\/p>\n\n<p>1) Define batches in your dataframe <code>df<\/code> with a <code>batch_id<\/code> column, so that in each batch <code>Y_train<\/code> is your <em>identical<\/em> \"batch ground truth\" (in my case, the total number of True labels in the batch). You can then pass these instances together to the network. This can be done with a generator:<\/p>\n\n<pre><code>def grouper(g,x,y):\n    while True:\n        for gr in g.unique():\n            # this assigns indices to the entire set of values in g,\n            # then subsects to all the rows in which g == gr\n            indices = g == gr\n            yield (x[indices],y[indices])\n\n# train set\ntrain_generator = grouper(df.loc[df['set'] == 'train','batch_id'], X_train, Y_train)\n# validation set\nval_generator = grouper(df.loc[df['set'] == 'val','batch_id'], X_val, Y_val)\n<\/code><\/pre>\n\n<p>2) define a custom loss function, to track how close the <em>total number<\/em> of instances predicted as true matches the ground truth:<\/p>\n\n<pre><code>def custom_delta(y_true, y_pred):\n\nloss = K.abs(K.mean(y_true) - K.sum(y_pred))\n\nreturn loss\n\n\ndef custom_wrapper():\n\n    def custom_loss_function(y_true, y_pred):\n        return custom_delta(y_true, y_pred)\n\n    return custom_loss_function\n<\/code><\/pre>\n\n<p>Note that here <\/p>\n\n<p>a) Each <code>y_true<\/code> label is <em>already<\/em> the sum of the ground truth in our batch (cause we don't have individual values). That's why <code>y_true<\/code> is not summed over;<\/p>\n\n<p>b) <code>K.mean<\/code> is actually a bit of an overkill to extract a single scalar from this uniform tensor, in which all y_true values in each batch are identical - <code>K.min<\/code> or <code>K.max<\/code> would also work, but I haven't tested whether their performance is faster.<\/p>\n\n<p>3) Use <code>fit_generator<\/code> instead of <code>fit<\/code>:<\/p>\n\n<pre><code>fmodel = Sequential()\n# ...your layers...\n# Create the loss function object using the wrapper function above\nloss_ = custom_wrapper()\nfmodel.compile(loss=loss_, optimizer='adam')\n\nhistory1 = fmodel.fit_generator(train_generator, steps_per_epoch=total_batches, \n    validation_data=val_generator,\n    validation_steps=df.loc[encs.df['set'] == 'val','batch_id'].nunique(),\n    epochs=20, verbose = 2)\n<\/code><\/pre>\n\n<p>This way the problem is basically addressed as one of supervised learning, although without individual labels, which means that notions like true\/false positive are meaningless here. <\/p>\n\n<p>This approach not only managed to give me a <code>y_pred<\/code> that closely matches the totals I know per batch. It actually finds two groups (True\/False) that occupy the expected different portions of parameter space.<\/p>\n"}]}
+{"QuestionId":53168652,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp><word-embedding>","CreationDate":"2018-11-06T09:01:27.627","AcceptedAnswerId":53269758.0,"OwnerUserId":1181744.0,"Title":"Test data giving prediction error in Keras in the model with Embedding layer","Body":"<p>I have trained a Bi-LSTM model to find NER on a set of sentences. For this I took the different words present and I did a mapping between a word and a number and then created the Bi-LSTM model using those numbers. I then create and pickle that model object.<\/p>\n\n<p>Now I get a set of new sentences containing certain words that the training model has not seen. Thus these words do not have a numeric value till now. Thus when I test it on my previously existing model, it would give an error. It is not able to find the words or features as the numeric values for those do not exist.<\/p>\n\n<p>To circumvent this error I gave a new integer value to all the new words that I see.<\/p>\n\n<p>However, when I load the model and test it, it gives the error that: <\/p>\n\n<pre><code>InvalidArgumentError: indices[0,24] = 5444 is not in [0, 5442)   [[Node: embedding_14_16\/Gather = Gather[Tindices=DT_INT32, Tparams=DT_FLOAT, validate_indices=true,\n_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](embedding_14_16\/embeddings\/read, embedding_14_16\/Cast)]]\n<\/code><\/pre>\n\n<p>The training data contains 5445 words including the padding word. Thus = [0, 5444]<\/p>\n\n<p>5444 is the index value I have given to the paddings in the test sentences. Not clear why it is assuming the index values to range between [0, 5442).<\/p>\n\n<p>I have used the base code available on the following link: <a href=\"https:\/\/www.kaggle.com\/gagandeep16\/ner-using-bidirectional-lstm\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/gagandeep16\/ner-using-bidirectional-lstm<\/a><\/p>\n\n<p>The code:<\/p>\n\n<pre><code>input = Input(shape=(max_len,))\nmodel = Embedding(input_dim=n_words, output_dim=50\n                  , input_length=max_len)(input)\n\nmodel = Dropout(0.1)(model)\nmodel = Bidirectional(LSTM(units=100, return_sequences=True, recurrent_dropout=0.1))(model)\nout = TimeDistributed(Dense(n_tags, activation=\"softmax\"))(model)  # softmax output layer\n\nmodel = Model(input, out)\nmodel.compile(optimizer=\"rmsprop\", loss=\"categorical_crossentropy\", metrics=[\"accuracy\"])\n\n#number of  epochs - Also for output file naming\nepoch_num=20\ndomain=\"..\/data\/Laptop_Prediction_Corrected\"\noutput_file_name=domain+\"_E\"+str(epoch_num)+\".xlsx\"\n\nmodel_name=\"..\/models\/Laptop_Prediction_Corrected\"\noutput_model_filename=model_name+\"_E\"+str(epoch_num)+\".sav\"\n\n\nhistory = model.fit(X_tr, np.array(y_tr), batch_size=32, epochs=epoch_num, validation_split=0.1, verbose=1)\n<\/code><\/pre>\n\n<p><code>max_len<\/code> is the total number of words in a sentence and <code>n_words<\/code> is the vocab size. In the model the padding has been done using the following code where <code>n_words=5441<\/code>:<\/p>\n\n<pre><code>X = pad_sequences(maxlen=max_len, sequences=X, padding=\"post\", value=n_words)\n<\/code><\/pre>\n\n<p>The padding in the new dataset:<\/p>\n\n<pre><code>max_len = 50\n# this is to pad sentences to the maximum length possible\n#-&gt; so all records of X will be of the same length\n\n#X = pad_sequences(maxlen=max_len, sequences=X, padding=\"post\", value=res_new_word2idx[\"pad_blank\"])\n\n#X = pad_sequences(maxlen=max_len, sequences=X, padding=\"post\", value=5441)\n<\/code><\/pre>\n\n<p>Not sure which of these paddings is correct?<\/p>\n\n<p>However, the vocab only includes the words in the training data. When I say:<\/p>\n\n<pre><code>p = loaded_model.predict(X)\n<\/code><\/pre>\n\n<p>How to use <code>predict<\/code> for text sentences which contain words that are not present in the initial vocab?<\/p>\n","answers":[{"AnswerId":"53269758","CreationDate":"2018-11-12T20:39:39.693","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can use Keras <a href=\"https:\/\/keras.io\/preprocessing\/text\/#tokenizer\" rel=\"nofollow noreferrer\"><code>Tokenizer<\/code><\/a> class and its methods to easily tokenize and preprocess the input data. Specify the vocab size when instantiating it and then use its <code>fit_on_texts()<\/code> method on the training data to construct a vocabulary based on the given texts. After that you can use its <code>text_to_sequences()<\/code> method to convert each text string to a list of word indices. The good thing is that only the words in the vocabulary is considered and all the other words are ignored (you can set those words to one by passing <code>oov_token=1<\/code> to <code>Tokenizer<\/code> class):<\/p>\n\n<pre><code>from keras.preprocessing.text import Tokenizer\n\n# set num_words to limit the vocabulary to the most frequent words\ntok = Tokenizer(num_words=n_words)\n\n# you can also pass an arbitrary token as `oov_token` argument \n# which will represent out-of-vocabulary words and its index would be 1\n# tok = Tokenizer(num_words=n_words, oov_token='[unk]')\n\ntok.fit_on_texts(X_train)\n\nX_train = tok.text_to_sequences(X_train)\nX_test = tok.text_to_sequences(X_test)  # use the same vocab to convert test data to sequences\n<\/code><\/pre>\n\n<p>You can optionally use <code>pad_sequences<\/code> function to pad them with zeros or truncate them to make them all have the same length:<\/p>\n\n<pre><code>from keras.preprocessing.sequence import pad_sequences\n\nX_train = pad_sequences(X_train, maxlen=max_len)\nX_test = pad_sequences(X_test, maxlen=max_len)\n<\/code><\/pre>\n\n<p>Now, the vocab size would be equal to <code>n_words+1<\/code> if you have not used oov token or <code>n_words+2<\/code> if you have used it. And then you can pass the correct number to embedding layer as its <code>input_dim<\/code> argument (first positional argument):<\/p>\n\n<pre><code>Embedding(correct_num_words, embd_size, ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53169369,"AnswerCount":0,"Tags":"<android><opencv><caffe><forward>","CreationDate":"2018-11-06T09:47:40.437","AcceptedAnswerId":null,"OwnerUserId":1125721.0,"Title":"Is there implicit limit for DNN input size of OpenCV Android?","Body":"<p>I trained a Caffe model which does NOT specify input sizes. Then I ran it on Android using OpenCV. The code is like below.<\/p>\n\n<pre><code>private void runEntireImageTest(Mat inY) {\n    float scaleFactor = 1.0f \/ 255.0f;\n    Scalar mean = new Scalar(0);\n    Mat resized = new Mat();\n\n    \/\/ (550, 441) runs OK. (smaller sized runs also fine)\n    \/\/ (551, 441), (550, 442) crash.\n    \/\/ (441, 550) crashes\n\n    Imgproc.resize(inY, resized, new Size(550, 441));\n    Mat segBlob = Dnn.blobFromImage(resized, scaleFactor, resized.size(), mean, false, false);\n    mNet.setInput(segBlob);\n\n    \/\/ if input size is above some value, crash will happen here\n    Mat lastLayer = mNet.forward();\n    Mat outY = lastLayer.reshape(1, 1);\n}\n<\/code><\/pre>\n\n<p>As it is written in the comment, there seems to be some <em>intrinsic<\/em>, or <em>implicit<\/em> limitation for the input size. <code>(550, 441)<\/code> ran okay, but <code>(551, 441)<\/code> will cause <code>SIGSEGV<\/code>:<\/p>\n\n<p><code>A\/libc: Fatal signal 11 (SIGSEGV), code 1, fault addr 0x6d494bd744 in tid 19693 (myappname), pid 19661 (myappname)<\/code><\/p>\n\n<p>I think it's not a memory problem because <code>(550, 441)<\/code> runs fine but <code>(441, 550)<\/code> will crash. What is the cause of this problem?<\/p>\n","answers":[]}
+{"QuestionId":53169455,"AnswerCount":1,"Tags":"<python><machine-learning><keras><keras-layer>","CreationDate":"2018-11-06T09:52:50.343","AcceptedAnswerId":53169711.0,"OwnerUserId":10395181.0,"Title":"How to properly connect Dense layer to Lambda layer in Keras?","Body":"<p>I'm using Keras. In the following code, <code>model<\/code> takes <code>[a0, a1]<\/code>, <code>[b0, b1, b2]<\/code> as inputs and gives <code>[a0*b0, a0*b1, a0*b2, a1*b0, a1*b1, a1*b2]<\/code> as an output:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.models import Model\nfrom keras.models import Input\nfrom keras.layers import Dense\n\ndef mix(ts):\n    t0 = K.expand_dims(ts[0], axis=-1)\n    t1 = K.expand_dims(ts[1], axis=1)\n    return K.batch_flatten(t0 * t1)\n\na = Input(shape=(2,))\nb = Input(shape=(3,))\nc = Lambda(mix)([a, b])\n\nmodel = Model(inputs=[a,b], outputs=c)\n<\/code><\/pre>\n\n<p>Here's the test:<\/p>\n\n<pre><code>u = np.array([1,2]).reshape(1,2)\nv = np.array([3,4,5]).reshape(1,3)\nprint(model.predict([u,v]))\n<\/code><\/pre>\n\n<p><code>[[ 3.  4.  5.  6.  8. 10.]]<\/code><\/p>\n\n<p>But if I try to connect <code>Dense<\/code> layer to <code>Lambda<\/code> layer, I get an error:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.models import Model\nfrom keras.models import Input\nfrom keras.layers import Dense\n\ndef mix(ts):\n    t0 = K.expand_dims(ts[0], axis=-1)\n    t1 = K.expand_dims(ts[1], axis=1)\n    return K.batch_flatten(t0 * t1)\n\na = Input(shape=(2,))\nb = Input(shape=(3,))\nc = Lambda(mix)([a, b])\nd = Dense(2)(c)\n\nmodel = Model(inputs=[a,b], outputs=d)\n<\/code><\/pre>\n\n<p>Here's the error I get:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-6-0f7f977a1e79&gt; in &lt;module&gt;()\n      7 b = Input(shape=(3,))\n      8 c = Lambda(mix)([a, b])\n----&gt; 9 d = Dense(2)(c)\n     10 \n     11 model = Model(inputs=[a,b], outputs=d)\n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\engine\\base_layer.py in __call__(self, inputs, **kwargs)\n    429                                          'You can build it manually via: '\n    430                                          '`layer.build(batch_input_shape)`')\n--&gt; 431                 self.build(unpack_singleton(input_shapes))\n    432                 self.built = True\n    433 \n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\layers\\core.py in build(self, input_shape)\n    864                                       name='kernel',\n    865                                       regularizer=self.kernel_regularizer,\n--&gt; 866                                       constraint=self.kernel_constraint)\n    867         if self.use_bias:\n    868             self.bias = self.add_weight(shape=(self.units,),\n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\legacy\\interfaces.py in wrapper(*args, **kwargs)\n     89                 warnings.warn('Update your `' + object_name + '` call to the ' +\n     90                               'Keras 2 API: ' + signature, stacklevel=2)\n---&gt; 91             return func(*args, **kwargs)\n     92         wrapper._original_function = func\n     93         return wrapper\n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\engine\\base_layer.py in add_weight(self, name, shape, dtype, initializer, regularizer, trainable, constraint)\n    247         if dtype is None:\n    248             dtype = K.floatx()\n--&gt; 249         weight = K.variable(initializer(shape),\n    250                             dtype=dtype,\n    251                             name=name,\n\n~\\Anaconda3\\envs\\mind\\lib\\site-packages\\keras\\initializers.py in __call__(self, shape, dtype)\n    207             scale \/= max(1., fan_out)\n    208         else:\n--&gt; 209             scale \/= max(1., float(fan_in + fan_out) \/ 2)\n    210         if self.distribution == 'normal':\n    211             # 0.879... = scipy.stats.truncnorm.std(a=-2, b=2, loc=0., scale=1.)\n\nTypeError: unsupported operand type(s) for +: 'NoneType' and 'int'\n<\/code><\/pre>\n\n<p>How can I properly connect <code>Dense<\/code> layer to <code>Lambda<\/code> layer?<\/p>\n","answers":[{"AnswerId":"53169711","CreationDate":"2018-11-06T10:08:17.533","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to set the output shape of the <code>Lambda<\/code> layer in this case, since it cannot be inferred automatically. Either pass the <code>output_shape<\/code> manually:<\/p>\n\n<pre><code>c = Lambda(mix, output_shape=(6,))([a, b])\n<\/code><\/pre>\n\n<p>Or better than that, pass a function to compute the output shape based on the shapes of input tensors of the layer:<\/p>\n\n<pre><code>def mix_output_shape(input_shape):\n    # input_shape[0] is the shape of first input tensor\n    # input_shape[1] is the shape of second input tensor\n    return (input_shape[0][0], input_shape[0][1] * input_shape[1][1])\n\n# ...\nc = Lambda(mix, mix_output_shape)([a, b])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53169997,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T10:26:32.893","AcceptedAnswerId":53171978.0,"OwnerUserId":9540764.0,"Title":"Apply a function (tf.square()) to some values of a Tensor - TensorFlow","Body":"<p>Let's start with an example. I have a tensor with shape of <code>[3, 3, 3]<\/code> like this<\/p>\n\n<pre><code>input = tf.constant([[[1.2, -2.3, 4.5], [1.03, 2.89, -1.2], [2.1, 1.9, -1.5]],\n                     [[1.3, -2.4, 4.6], [1.04, 2.88, -1.3], [2.2, 1.8, -1.6]],\n                     [[1.4, -2.5, 4.7], [1.05, 2.87, -1.4], [2.3, 1.7, -1.7]]],\n                    dtype=tf.float32)\n<\/code><\/pre>\n\n<p>and I want to apply <code>tf.square<\/code> only to the elements lower than <code>2<\/code> for example.\nWhat I'm doing is this<\/p>\n\n<pre><code>indices = tf.where(input &lt;= 2)\nbase = tf.zeros_like(input)\nones = tf.constant(1, shape=[18])\nmask = tf.scatter_nd(indices, ones, tf.cast(tf.shape(base), indices.dtype))\nmask = tf.cast(mask, tf.float32)\nmasked = tf.square(tf.multiply(input, mask))\nneg_mask = 1 - mask\nneg_masked = tf.multiply(input, neg_mask)\noutput = tf.add(masked, neg_masked)\n<\/code><\/pre>\n\n<p>and it works. The final output is this<\/p>\n\n<pre><code>output = tf.constant([[[1.44, 5.29, 4.5], [1.0609, 2.89, 1.44], [2.1, 3.61, 2.25]],\n                      [[1.69, 5.76, 4.6], [1.0816, 2.88, 1.69], [2.2, 3.24, 2.56]],\n                      [[1.96, 6.25, 4.7], [1.1025, 2.87, 1.96], [2.3, 2.89, 2.89]]],\n                     dtype=tf.float32)\n<\/code><\/pre>\n\n<p>The problem is that it's quite tricky because this is a toy example but in my case, the tensor has a shape of thousands of elements. Moreover as you can see in this line <code>ones = tf.constant(1, shape=[18])<\/code> I have coded <code>18<\/code> because I know that they are 18 and if I try to use for example <code>ones = tf.constant(1, shape=[indices.get_shape()[0]])<\/code> I get this error <code>TypeError: <strong>long<\/strong> returned non-long (type NoneType)<\/code>.<\/p>\n\n<p>So my two questions are:<\/p>\n\n<ol>\n<li>How can I solve this problem of the shape?<\/li>\n<li>Is there a faster and more efficient way to do this?<\/li>\n<\/ol>\n\n<p>Thanks in advance<\/p>\n\n<hr>\n\n<p><strong>edit<\/strong><\/p>\n\n<p>Question 1 solved doing this <code>ones = tf.ones(shape=[tf.shape(indices)[0]])<\/code><\/p>\n","answers":[{"AnswerId":"53171978","CreationDate":"2018-11-06T12:31:34.010","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The easiest thing is to square all values and then select the right ones:<\/p>\n\n<pre><code>import tensorflow as tf\n\ninput = tf.constant([[[1.2, -2.3, 4.5], [1.03, 2.89, -1.2], [2.1, 1.9, -1.5]],\n                     [[1.3, -2.4, 4.6], [1.04, 2.88, -1.3], [2.2, 1.8, -1.6]],\n                     [[1.4, -2.5, 4.7], [1.05, 2.87, -1.4], [2.3, 1.7, -1.7]]],\n                    dtype=tf.float32)\noutput = tf.where(input &lt; 2, tf.square(input), input)\nwith tf.Session() as sess:\n    print(sess.run(output))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[1.44      5.29      4.5      ]\n  [1.0609    2.89      1.44     ]\n  [2.1       3.61      2.25     ]]\n\n [[1.6899998 5.76      4.6      ]\n  [1.0816    2.88      1.6899998]\n  [2.2       3.2399998 2.5600002]]\n\n [[1.9599999 6.25      4.7      ]\n  [1.1024998 2.87      1.9599999]\n  [2.3       2.89      2.89     ]]]\n<\/code><\/pre>\n\n<p>If you had a really big tensor where only few values will be squared then you could consider squaring only the necessary values. I'm not sure this could ever be actually faster, though, since it requires some additional work and intermediate values, but you could benchmark it. If, instead of just squaring, you were doing a particularly expensive operation, I guess it could make a difference. This is more or less the same you did but a bit simpler:<\/p>\n\n<pre><code>import tensorflow as tf\n\ninput = tf.constant([[[1.2, -2.3, 4.5], [1.03, 2.89, -1.2], [2.1, 1.9, -1.5]],\n                     [[1.3, -2.4, 4.6], [1.04, 2.88, -1.3], [2.2, 1.8, -1.6]],\n                     [[1.4, -2.5, 4.7], [1.05, 2.87, -1.4], [2.3, 1.7, -1.7]]],\n                    dtype=tf.float32)\nm = input &lt; 2\nv = tf.boolean_mask(input, m)\nv2 = tf.square(v)\nv2_scatter = tf.scatter_nd(tf.where(m), v2, tf.cast(tf.shape(input), tf.int64))\noutput = input * tf.cast(~m, input.dtype) + v2_scatter\nwith tf.Session() as sess:\n    print(sess.run(output))\n    # Output is the same as before\n<\/code><\/pre>\n"},{"AnswerId":"53173345","CreationDate":"2018-11-06T13:55:29.557","ParentId":null,"OwnerUserId":"8386202","Title":null,"Body":"<p>This is an easier but not necessarily faster implementation:<\/p>\n\n<pre><code>input = tf.constant([[[1.2, -2.3, 4.5], [1.03, 2.89, -1.2], [2.1, 1.9, -1.5]],\n                     [[1.3, -2.4, 4.6], [1.04, 2.88, -1.3], [2.2, 1.8, -1.6]],\n                     [[1.4, -2.5, 4.7], [1.05, 2.87, -1.4], [2.3, 1.7, -1.7]]],\n                    dtype=tf.float32)\n\noriginal_shape = input.get_shape()\ninput = tf.reshape(input, shape=[-1])\noutput = tf.map_fn(lambda e:tf.cond(e &lt; 2, lambda:tf.square(e), lambda:e), input)\noutput = tf.reshape(output, shape=original_shape)\n\nwith tf.Session() as sess:\n    print(sess.run(output))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[[1.44      5.29      4.5      ]\n  [1.0609    2.89      1.44     ]\n  [2.1       3.61      2.25     ]]\n\n [[1.6899998 5.76      4.6      ]\n  [1.0816    2.88      1.6899998]\n  [2.2       3.2399998 2.5600002]]\n\n [[1.9599999 6.25      4.7      ]\n  [1.1024998 2.87      1.9599999]\n  [2.3       2.89      2.89     ]]]\n<\/code><\/pre>\n\n<p>The <code>tf.map_fn(fn, elems, ...)<\/code> function unpacks the N-dimensional input <code>elems<\/code> along the first dimension into multiple N-1-dimensional subtensors and applies <code>fn<\/code> to each subtensor. So I reshape input into one-dimensional tensor, apply function on each element, and then reshape output to original shape.<\/p>\n"}]}
+{"QuestionId":53170149,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-11-06T10:35:37.847","AcceptedAnswerId":53177266.0,"OwnerUserId":6831406.0,"Title":"Memory usage of neural network, Keras","Body":"<p>I am trying to develop a model for denoising images. I've been reading up on how to calculate memory usage of a neural network and the standard approach seems to be:<\/p>\n\n<pre><code>    params = depth_n x (kernel_width x kernel_height) x depth_n-1 + depth\n<\/code><\/pre>\n\n<p>By summing all parameters together in my network, I end up getting <strong>1,038,097<\/strong> which approximates to <strong>4.2MB<\/strong>. It seems I have done a slight miscalculation in the last layer since Keras ends up getting <strong>1,038,497<\/strong> params. Nevertheless, this is a small difference. 4.2MB is just the parameters, and I've seen somewhere that one should <strong>multiply by 3<\/strong> to include backprop and other needed calculations. This would then approximate to 13MB.<\/p>\n\n<p>I have approximately 11 GB of GPU memory to work with, yet this model gets exhausted. Where does all the extra needed memory come from? What am I missing? I know this post might be labeled as duplicate, but none of the others seems to catch the topic which I am asking about.<\/p>\n\n<p>My model:<\/p>\n\n<pre><code>    def network(self):\n        weights = RandomUniform(minval=-0.05, maxval=0.05, seed=None)\n        input_img = Input(shape=(self.img_rows, self.img_cols, self.channels))\n        conv1 = Conv2D(1024, (3,3), activation='tanh', kernel_initializer=weights,\n                padding='same', use_bias=True)(input_img)\n        conv2 = Conv2D(64, (3,3), activation='tanh', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv1)\n        conv3 = Conv2D(64, (3,3), activation='tanh', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv2)\n        conv4 = Conv2D(64, (3,3), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv3)\n        conv5 = Conv2D(64, (7,7), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv4)\n        conv6 = Conv2D(64, (5,5), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv5)\n        conv7 = Conv2D(32, (5,5), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv6)\n        conv8 = Conv2D(32, (3,3), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv7)\n        conv9 = Conv2D(16, (3,3), activation='relu', kernel_initializer=weights,\n                padding='same', use_bias=True)(conv8)\n        decoded = Conv2D(1, (5,5), kernel_initializer=weights,\n                padding='same', activation='sigmoid', use_bias=True)(conv8)\n        return input_img, decoded\n\n\n    def compiler(self):\n        self.model.compile(optimizer='RMSprop', loss='mse')\n        self.model.summary()\n<\/code><\/pre>\n\n<p>I assume my model is silly in a lot of ways and that there are multiple things to improve (dropout, other filter sizes and numbers, optimizers etc.) and all suggestions are received gladly, but the actual question still remain. Why does this model consume so much memory? Is it due to the extremely high depth of <code>conv1<\/code>?<\/p>\n\n<p>Model summary:<\/p>\n\n<pre><code>    Using TensorFlow backend.\n    _________________________________________________________________\n    Layer (type)                 Output Shape              Param #   \n    =================================================================\n    input_1 (InputLayer)         (None, 1751, 480, 1)      0         \n    _________________________________________________________________\n    conv2d_1 (Conv2D)            (None, 1751, 480, 1024)   10240     \n    _________________________________________________________________\n    conv2d_2 (Conv2D)            (None, 1751, 480, 64)     589888    \n    _________________________________________________________________\n    conv2d_3 (Conv2D)            (None, 1751, 480, 64)     36928     \n    _________________________________________________________________\n    conv2d_4 (Conv2D)            (None, 1751, 480, 64)     36928     \n    _________________________________________________________________\n    conv2d_5 (Conv2D)            (None, 1751, 480, 64)     200768    \n    _________________________________________________________________\n    conv2d_6 (Conv2D)            (None, 1751, 480, 64)     102464    \n    _________________________________________________________________\n    conv2d_7 (Conv2D)            (None, 1751, 480, 32)     51232     \n    _________________________________________________________________\n    conv2d_8 (Conv2D)            (None, 1751, 480, 32)     9248      \n    _________________________________________________________________\n    conv2d_10 (Conv2D)           (None, 1751, 480, 1)      801       \n    =================================================================\n    Total params: 1,038,497\n    Trainable params: 1,038,497\n    Non-trainable params: 0\n    _________________________________________________________________\n<\/code><\/pre>\n","answers":[{"AnswerId":"53177266","CreationDate":"2018-11-06T17:46:19.753","ParentId":null,"OwnerUserId":"1479778","Title":null,"Body":"<p>You are correct, this is due to the number of filters in <code>conv1<\/code>. What you must compute is the memory required to store the activations:<\/p>\n\n<p>As shown by your <code>model.summary()<\/code>, the output size of this layer is <code>(None, 1751, 480, 1024)<\/code>. For a single image, this is a total of <code>1751*480*1024<\/code> pixels. As your image is likely in <code>float32<\/code>, each pixel takes 4 bytes to store. So the output of this layer requires <code>1751*480*1024*4<\/code> bytes, which is around 3.2 GB per image just for this layer. <\/p>\n\n<p>If you were to change the number of filters to, say, 64, you would only need around 200 MB per image. <\/p>\n\n<p>Either change the number of filters or change the batch size to 1.<\/p>\n"}]}
+{"QuestionId":53170600,"AnswerCount":1,"Tags":"<python><machine-learning><time><deployment><keras>","CreationDate":"2018-11-06T11:03:46.050","AcceptedAnswerId":null,"OwnerUserId":9094995.0,"Title":"Keras models training taking up more time after some time","Body":"<p>I have a link to a database of 500 users. From these users, data is extracted. Total training dataset size for each user is <code>[6000*4]<\/code> i.e. 4 features and 6000 training examples. A neural network is trained for each of these users. <\/p>\n\n<p>Here is a brief code of what is happening in the code:<\/p>\n\n<pre><code>user_list = get_users_list()\n\nfor each_user in user_list:\n   is_succesful = run_model(each_user)\n\ndef run_model():\n   training_data = extract_data() # 6000 * 4 samples\n   X_nn,Y_nn = training_data[:,0],training_data[:,1]\n   start_time = time.time()\n   model = fit_nn(X_nn,Y,nb_epoch = 200,batch_size = 20) # code of this is appended below\n   end_time = time.time()\n   duration = end_time - start_time\n   print(str(duration))\n   model.save_model('some_model')\n   return True\n<\/code><\/pre>\n\n<p>The problem that is arising is: for the first 15-20 users, training time is ~ 90 seconds.\nBut after training of 15-20 users, training time keeps on increasing and reaches to <strong>1600 seconds.<\/strong> This doesn't make sense because <\/p>\n\n<ol>\n<li>The models are independent (one model gets destroyed before next training is started).<\/li>\n<li>Model architecture is the same<\/li>\n<li>Data size is almost the same.<\/li>\n<\/ol>\n\n<p>So, my question is: shouldn't models training take almost the same time? <\/p>\n\n<p>Here is the code of <code>fit_nn<\/code><\/p>\n\n<pre><code>def fit_nn(x_train,y_train,nb_epoch,batch_size):\n\n  model = Sequential()\n  model.add(Dense(300, activation=\"relu\",input_shape=(x_train.shape[1],)))\n  model.add(Dropout(0.3))\n  model.add(Dense(150, activation=\"relu\"))\n  model.add(Dense(50, activation=\"relu\"))\n  model.add(Dense(30, activation=\"relu\"))\n  model.add(Dense(1))\n\n  optimizer = optimizers.RMSprop(0.001)      \n  model.compile(loss='mse',optimizer=optimizer,metrics=['mae'])              \n\n  model.fit(x_train,y_train,epochs=nb_epoch,batch_size=batch_size,verbose=2,shuffle=True)\n\n  return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"53171342","CreationDate":"2018-11-06T11:50:50.497","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>If you are using tensorflow as your keras backend I suggest calling <\/p>\n\n<pre><code>K.clear_session()\n<\/code><\/pre>\n\n<p>at the end of <code>run_model()<\/code><\/p>\n\n<p>make sure you add the needed import <code>from keras import backend as K<\/code> <\/p>\n\n<p>If you don't reset your graph which <code>K.clear_session()<\/code> does, tensorflow will keep adding nodes to it, as a result each iteration will take longer time until the process itself is killed.<\/p>\n"}]}
+{"QuestionId":53171188,"AnswerCount":1,"Tags":"<linux><docker><ubuntu><tensorflow><nvidia-docker>","CreationDate":"2018-11-06T11:41:08.623","AcceptedAnswerId":null,"OwnerUserId":10612886.0,"Title":"why is my NVIDIA docker not working? Error response from daemon?","Body":"<p>I'm trying to install NVIDIA docker.\nI used these lines:<\/p>\n\n<pre><code>curl -s -L https:\/\/nvidia.github.io\/nvidia-docker\/gpgkey | \\\n  sudo apt-key add -\ndistribution=$(. \/etc\/os-release;echo $ID$VERSION_ID)\n\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-docker\/$distribution\/nvidia-docker.list | \\\n  sudo tee \/etc\/apt\/sources.list.d\/nvidia-docker.list\n\nsudo apt-get update\n<\/code><\/pre>\n\n<p>and then:<\/p>\n\n<p><code>$ sudo apt-get install nvidia-docker<\/code><\/p>\n\n<p>Now trying to check if it installed correctly by typing:<\/p>\n\n<p><code>nvidia-docker run --rm nvidia\/cuda nvidia-smi<\/code><\/p>\n\n<p>This error appears: <\/p>\n\n<blockquote>\n  <p>nvidia-docker | 2018\/11\/06 13:09:24 Error: Got permission denied while\n  trying to connect to the Docker daemon socket at\n  unix:\/\/\/var\/run\/docker.sock: Get\n  <a href=\"http:\/\/%2Fvar%2Frun%2Fdocker.sock\/v1.38\/version\" rel=\"nofollow noreferrer\">http:\/\/%2Fvar%2Frun%2Fdocker.sock\/v1.38\/version<\/a>: dial unix\n  \/var\/run\/docker.sock: connect: permission denied<\/p>\n<\/blockquote>\n\n<p>Later I tried:<\/p>\n\n<p><code>sudo nvidia-docker run --rm nvidia\/cuda nvidia-smi<\/code><\/p>\n\n<p>This error appears:<\/p>\n\n<blockquote>\n  <p>Using default tag: latest latest: Pulling from nvidia\/cuda\n  473ede7ed136: Pull complete  c46b5fa4d940: Pull complete \n  93ae3df89c92: Pull complete  6b1eed27cade: Pull complete \n  d31e9163d0a5: Pull complete  8668af631f88: Pull complete \n  0d99f8ab6ae2: Pull complete  74440c29d798: Pull complete  Digest:\n  sha256:a6b5fd418d1cd0bc6d8a60c1c4ba33670508487039b828904f8494ec29e6b450\n  Status: Downloaded newer image for nvidia\/cuda:latest docker: Error\n  response from daemon: OCI runtime create failed:\n  container_linux.go:348: starting container process caused \"exec:\n  \\\"nvidia-smi\\\": executable file not found in $PATH\": unknown.<\/p>\n<\/blockquote>\n\n<p>I install Linux with Ubuntu.\nCan someone help me, please?<\/p>\n","answers":[{"AnswerId":"53173338","CreationDate":"2018-11-06T13:54:44.500","ParentId":null,"OwnerUserId":"4650277","Title":null,"Body":"<p>You have two errors which are quite self explicites.<\/p>\n\n<p>First of all, it seems your login user is not allowed to connect to docker daemon. \nThis is quite a standard issue, you just need to add your user login to docker group, It should solve this issue. You will need to logout\/login again for this change to become active. <\/p>\n\n<p>Second, this is also quite a standard linux issue, your shell has a environment variable called PATH , containing all folders where it will be looking for a command binary, when this command doesn't contain full path to the binary. <\/p>\n\n<p>For exemple, when you typed curl to download docker-nvidia, your shell find it in <code>\/usr\/bin\/<\/code> folder, because this folder is declared into the PATH variable. <\/p>\n\n<p>Same applies for containers you download and different users on your local system. You can investigate this specific error message and find this issue on github: <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\/issues\/388\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/nvidia-docker\/issues\/388<\/a><\/p>\n"}]}
+{"QuestionId":53171885,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-06T12:27:10.473","AcceptedAnswerId":null,"OwnerUserId":6949639.0,"Title":"How to use tf.data.Dataset and tf.keras do multi-inputs and multi-outpus?","Body":"<p>I meet one question about multi-output using tf.keras to build model and also using tf.data.Dataset as the input pipeline. Here is my code below:<\/p>\n\n<pre><code>  a = tf.keras.layers.Input(shape=(368, 368, 3))\n  conv1 = tf.keras.layers.Conv2D(64, 3, 1)(a)\n  conv2 = tf.keras.layers.Conv2D(64, 3, 1)(conv1)\n  maxpool = tf.keras.layers.MaxPooling2D(pool_size=8, strides=8, \n   padding='same')(conv2)\n  conv3 = tf.keras.layers.Conv2D(5, 1, 1)(maxpool)\n  conv4 = tf.keras.layers.Conv2D(6, 1, 1)(maxpool)\n\n  inputs = a\n  outputs = [conv3, conv4]\n\n  model = tf.keras.models.Model(inputs=inputs, outputs=outputs)\n\n\n  model.compile(optimizer=tf.keras.optimizers.SGD(),\n          loss=tf.keras.losses.mean_squared_error)\n\n\n  import numpy as np\n  data = np.random.rand(10, 368, 368, 3)\n  cpm  = np.random.rand(10, 46, 46, 5)\n  paf  = np.random.rand(10, 46, 46, 6)\n\n  dataset1 = tf.data.Dataset.from_tensor_slices((data))\n  dataset2 = tf.data.Dataset.from_tensor_slices((cpm, paf))\n  dataset1 = dataset1.batch(10).repeat()\n  dataset2 = dataset2.batch(10).repeat()\n\n  dataset  = tf.data.Dataset.zip((dataset1, dataset2))\n\n  model.fit(dataset, epochs=200, steps_per_epoch=30)\n<\/code><\/pre>\n\n<p>I'm using tensorflow==1.10.1 and I got error like this:<\/p>\n\n<pre><code> File \"\/home\/ulsee\/work\/tensorflow-HalfBodyPose\/learnkeras.py\", line 123, in &lt;module&gt;\nmodel.fit(dataset, epochs=200, steps_per_epoch=30)\n File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1278, in fit\nvalidation_split=validation_split)\n File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 917, in _standardize_user_data\nexception_prefix='target')\n File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 143, in standardize_input_data\ndata = [standardize_single_array(x) for x in data]\n File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 143, in &lt;listcomp&gt;\ndata = [standardize_single_array(x) for x in data]\n File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 81, in standardize_single_array\nelif x.ndim == 1:\nAttributeError: 'tuple' object has no attribute 'ndim'\n<\/code><\/pre>\n\n<p><strong>Update<\/strong> :\nI have made this code work after i upgrading tf==1.11.0. So maybe i think it's a version bug. <\/p>\n","answers":[{"AnswerId":"53192526","CreationDate":"2018-11-07T15:25:56.490","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>One thing you can try is to concatenate your outputs and then do the same for your target numpy arrays. Whether it makes sense for your application logic wise is something I am not sure.<\/p>\n\n<pre><code>def conc_op(tensors):\n    return K.concatenate(tensors) # K refers to Keras backend\n\ndef conc_op_shape(input_shapes):\n    shape1 = list(input_shapes[0])\n    shape2 = list(input_shapes[1])\n    return tuple(shape1[:-1], shape1[-1]+shape2[-1])\n\na = tf.keras.layers.Input(shape=(368, 368, 3))\nconv1 = tf.keras.layers.Conv2D(64, 3, 1)(a)\nconv2 = tf.keras.layers.Conv2D(64, 3, 1)(conv1)\nmaxpool = tf.keras.layers.MaxPooling2D(pool_size=8, strides=8, padding='same')(conv2)\nconv3 = tf.keras.layers.Conv2D(5, 1, 1)(maxpool)\nconv4 = tf.keras.layers.Conv2D(6, 1, 1)(maxpool)\n\ninputs = a\noutputs = [conv3, conv4]\nconc_outputs = Lambda(conc_op, output_shape=conc_op_shape)(outputs) # This is a keras layer\nmodel = tf.keras.models.Model(inputs=inputs, outputs=conc_outputs)\n\nmodel.compile(optimizer=tf.keras.optimizers.SGD(), loss=keras.losses.mean_squared_error)\nmodel.summary()\ndata = np.random.rand(10, 368, 368, 3)\ncpm  = np.random.rand(10, 46, 46, 5)\npaf  = np.random.rand(10, 46, 46, 6)\nlabel = np.concatenate((cpm, paf), axis=-1)\n\ndataset = tf.data.Dataset.from_tensor_slices((data, label))\ndataset = dataset.batch(2).repeat()\nmodel.fit(dataset.make_one_shot_iterator(), epochs=2, steps_per_epoch=5) # Just to check if it runs\n<\/code><\/pre>\n\n<p>Returns the Result:<\/p>\n\n<pre><code>Epoch 1\/2\n5\/5 [==============================] - 15s 3s\/step - loss: 0.4057\nEpoch 2\/2\n5\/5 [==============================] - 0s 32ms\/step - loss: 0.2282\n<\/code><\/pre>\n"}]}
+{"QuestionId":53172215,"AnswerCount":1,"Tags":"<python><tensorflow><pre-trained-model>","CreationDate":"2018-11-06T12:46:12.493","AcceptedAnswerId":null,"OwnerUserId":8005524.0,"Title":"Duplicate node name in graph: 'conv2d_0\/kernel\/Adam'","Body":"<p>I just saved a model, by that code:<\/p>\n\n<pre><code>def train():    \nwith tf.Session() as sess:\n    saver = tf.train.Saver(max_to_keep = 2)\n    Loss = myYoloLoss([Scale1,Scale2,Scale3],[Y1, Y2 ,Y3])\n    opt = tf.train.AdamOptimizer(2e-4).minimize(Loss)\n    init = tf.global_variables_initializer()\n    sess.run(init)\n    imageNum = 0\n    Num = 0\n    while(1):\n        #get batchInput\n        batchImg,batchScale1,batchScale2,batchScale3 = getBatchImage(batchSize = BATCHSIZE)\n        for epoch in range(75):\n            _ , epochloss = sess.run([opt,Loss],feed_dict={X:batchImg,Y1:batchScale1,Y2:batchScale2,Y3:batchScale3})\n            if(epoch%15 == 0):\n                print(epochloss)\n        imageNum = imageNum + BATCHSIZE\n        Num = Num + 1\n        if(Num%4 == 0):\n            saver.save(sess,MODELPATH + 'MyModle__' + str(imageNum))            \n        if(os.path.exists(STOPFLAGPATH)):\n            saver.save(sess,MODELPATH + 'MyModle__Stop_' + str(imageNum))   \n            print('checked stopfile,stop')\n            break\nreturn 0\n<\/code><\/pre>\n\n<p>And then I get some files:<br\/><\/p>\n\n<p>MyModle__Stop_288.index<br\/>\nMyModle__Stop_288.meta<br\/>\nMyModle__Stop_288.data-00000-of-00001<br\/>\ncheckpoint<br\/><\/p>\n\n<p>Then I continue to train this model:<\/p>\n\n<pre><code>def reTrain():\nwith tf.Session() as sess:\n    loder = tf.train.import_meta_graph('E:\/MyYoloModel\/MyModle__Stop_288.meta')\n    loder.restore(sess, tf.train.latest_checkpoint('E:\/MyYoloModel\/'))\n    graph = tf.get_default_graph()\n    X = graph.get_tensor_by_name(\"X:0\")\n    Y1 = graph.get_tensor_by_name(\"Y1:0\")\n    Y2 = graph.get_tensor_by_name(\"Y2:0\")\n    Y3 = graph.get_tensor_by_name(\"Y3:0\")\n    Scale1 = graph.get_tensor_by_name(\"Scale1:0\")\n    Scale2 = graph.get_tensor_by_name(\"Scale2:0\")\n    Scale3 = graph.get_tensor_by_name(\"Scale3:0\")  \n    Loss = myYoloLoss([Scale1,Scale2,Scale3],[Y1, Y2 ,Y3])\n    #error code \n    opt = tf.train.AdamOptimizer(2e-4).minimize(Loss)\n    init = tf.global_variables_initializer()\n    sess.run(init)\n    batchImg,batchScale1,batchScale2,batchScale3 = getBatchImage(batchSize = BATCHSIZE)\n    for epoch in range(10):\n        _ ,epochloss = sess.run([opt,Loss],feed_dict={X:batchImg,Y1:batchScale1,Y2:batchScale2,Y3:batchScale3})\n        print(epochloss)\n<\/code><\/pre>\n\n<p>And it will be occur this error:\nValueError: Duplicate node name in graph: 'conv2d_0\/kernel\/Adam'<br\/>\nHow can fix it?<\/p>\n","answers":[{"AnswerId":"57535830","CreationDate":"2019-08-17T11:23:05.970","ParentId":null,"OwnerUserId":"9221127","Title":null,"Body":"<p>The reason is that AdamOptimizer  creates additional variables and operation in your graph. When you store your model, those operations are stored and loaded with the graph when you restore the model. \nIf you run <\/p>\n\n<blockquote>\n  <p>tf.Graph.get_operations(graph)<\/p>\n<\/blockquote>\n\n<p>you can see the list of operations that are loaded with you model. You will see operations that have <em>\/Adam or train\/Adam<\/em> init. When you try to find-tune or reuse you model, the new AdamOptimizer tries to create those operations again, hence it raises the \"Duplicate node name\" error.\nOne way to fix the issue is to give a name to your new AdampOptimzer. <\/p>\n\n<blockquote>\n  <p>opt = tf.train.AdamOptimizer(2e-4m name='MyNewAdam').minimize(Loss)<\/p>\n<\/blockquote>\n\n<p>However, We are not done yet. \nAs you want to reuse the weight, you cannot initialize variable. However, if you will get error of uninitialized parameters when you run your training which is raised due to new AdamOptimizer variables which have not been initialized yet. To get around it, you need to initialize those new variables by :<\/p>\n\n<pre><code>uninitialized_vars = []\nfor var in tf.all_variables():\n    try:\n        sess.run(var)\n    except tf.errors.FailedPreconditionError:\n        uninitialized_vars.append(var)\n\ntf.initialize_variables(uninitialized_vars)\n<\/code><\/pre>\n\n<p>Note: Unused nodes will not be executed and hence they won't affect training time.<\/p>\n"}]}
+{"QuestionId":53172852,"AnswerCount":0,"Tags":"<keras><lstm><embedding>","CreationDate":"2018-11-06T13:25:14.367","AcceptedAnswerId":null,"OwnerUserId":5617507.0,"Title":"Masking zero inputs in LSTM in keras without using embedding","Body":"<p>I am training an LSTM in Keras:<\/p>\n\n<pre><code>iclf = Sequential()\niclf.add(Bidirectional(LSTM(units=10, return_sequences=True, recurrent_dropout=0.3), input_shape=(None,2048)))\niclf.add(TimeDistributed(Dense(1, activation='sigmoid')))\n<\/code><\/pre>\n\n<p>The input to each cell is a 2048 vector which is known and need not to be learned (if you will, they are the ELMo embeddings of the words in the input sentences). Therefore, here I have not the Embedding layer. <\/p>\n\n<p>Since the input sequences have variable lengths, they are padded using <code>pad_sequences<\/code>:<\/p>\n\n<pre><code>X = pad_sequences(sequences=X, padding='post', truncating='post', value=0.0, dtype='float32')\n<\/code><\/pre>\n\n<p>Now, I want to tell the LSTM to ignore these padded elements. The official way is to use the Embedding layer with <code>mask_zero=True<\/code>. But, here there is no Embedding layer. How can I inform the LSTM to mask zero elements?<\/p>\n","answers":[]}
+{"QuestionId":53173449,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T14:01:34.390","AcceptedAnswerId":null,"OwnerUserId":8165066.0,"Title":"Is it a bug of tensorflow for feed type?","Body":"<p>See the code snippet:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n\nz = tf.placeholder(tf.float32, [None, 100], name='z')\nuc = tf.placeholder(tf.float32, [None, 2], name='uc')\nsc = tf.placeholder(tf.int32, [None], name='sc')\n\n\nsc_one_hot = tf.one_hot(sc, 10)\nzc = tf.concat([z, uc, sc_one_hot], axis=1)\n\n\nsample_z = np.random.normal(size=[32, 100])\nsample_uc = np.random.uniform(0., 1., size=[32, 2])\nsample_sc1 = np.random.randint(10, size=32)\nsample_sc2 = np.mod(np.arange(32), 10)\n\nwith tf.Session() as sess:\n    sess.run(zc, feed_dict={z: sample_z, uc: sample_uc, sc: sample_sc1})     # works well\n    sess.run(zc, feed_dict={z: sample_z, uc: sample_uc, sc: sample_sc2})   # raises an error\n<\/code><\/pre>\n\n<p><code>sample_sc1<\/code> and <code>sample_sc2<\/code> are both <code>ndarray<\/code> with the <code>size<\/code> of <code>32<\/code> and <code>dtype<\/code> of <code>int32<\/code>, and all numbers in them range from <code>0<\/code> to <code>9<\/code>. <\/p>\n\n<p><code>sc<\/code> is <code>tf.placeholder<\/code> with <code>dtype<\/code> of <code>tf.int32<\/code> and <code>size<\/code> of <code>[None]<\/code>. \nWhen I feed <code>sample_sc1<\/code> into <code>sc<\/code>, it works well, but when I feed <code>sample_sc2<\/code> into <code>sc<\/code>, it raises <code>tensorflow.python.framework.errors_impl.InternalError: Unsupported feed type<\/code>. <\/p>\n\n<p>It is an amazing error, is it a bug of tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":53173884,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><keras>","CreationDate":"2018-11-06T14:25:31.257","AcceptedAnswerId":53174026.0,"OwnerUserId":7660421.0,"Title":"Design a neural network to predict array of numbers in Keras","Body":"<p>I'm new to machine learning and have little experience with ML libraries. I'm developing a neural network to predict array of numbers given an input array. So my input is like follows:<\/p>\n\n<blockquote>\n  <p>Input to the neural network is a array of length 100.<\/p>\n<\/blockquote>\n\n<p>And the output:<\/p>\n\n<blockquote>\n  <p>Output of the model is array of length 60.<\/p>\n<\/blockquote>\n\n<p>Here I post my code so far. It is currently not working. Can someone help me to do this in a correct way? <\/p>\n\n<pre><code>model = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Flatten(input_shape=(100,)))\nmodel.add(tf.keras.layers.Dense(768, activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(768, activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(60, activation=tf.nn.softmax))\n\nmodel.compile(optimizer = 'adam',\n         loss = 'sparse_categorical_crossentropy',\n         metrices = ['accuracy'])\nmodel.fit(x_train_new, y_train_new, epochs = 20)\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> <code>x_train_new<\/code> is a array of arrays of length 100 and <code>y_train_new<\/code> is a array of arrays of length 60.<\/p>\n\n<p>When I run this and call <code>model.fit()<\/code> following error will be appeared.<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-62-33e8f1dfe4c8&gt; in &lt;module&gt;()\n      5 # print(y_train_new.shape)\n      6 \n----&gt; 7 model.fit(x_train_new, y_train_new, epochs = 20)\n\n\/home\/vajira\/tf_cpu_py3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1276         steps_name='steps_per_epoch',\n   1277         steps=steps_per_epoch,\n-&gt; 1278         validation_split=validation_split)\n   1279 \n   1280     # Prepare validation data.\n\n\/home\/vajira\/tf_cpu_py3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split)\n    915           feed_output_shapes,\n    916           check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 917           exception_prefix='target')\n    918 \n    919       # Generate sample-wise weight values given the `sample_weight` and\n\n\/home\/vajira\/tf_cpu_py3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    189                 'Error when checking ' + exception_prefix + ': expected ' +\n    190                 names[i] + ' to have shape ' + str(shape) +\n--&gt; 191                 ' but got array with shape ' + str(data_shape))\n    192   return data\n    193 \n\nValueError: Error when checking target: expected dense_14 to have shape (1,) but got array with shape (60,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53174026","CreationDate":"2018-11-06T14:34:18.720","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You must specify the input shape for the first layer:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Flatten(input_shape=(100,)))\n<\/code><\/pre>\n\n<p>However, if the input is a vector of length 100, then there is no need for a <code>Flatten<\/code> layer as it is already flattened. Instead, set the <code>input_shape<\/code> for the first <code>Dense<\/code> layer and remove the <code>Flatten<\/code> layer.<\/p>\n\n<p><strong>Update:<\/strong> Regarding the shape mismatch error, either use <code>categorical_crossentropy<\/code> as the loss function and pass one-hot encoded labels of size 60, or use <code>sparse_categorical_crossentropy<\/code> and pass integer labels (i.e. 0 or 1 or 2, etc.). Don't mix these two things: currently you are using <code>sparse_categorical_crossentropy<\/code> and pass one-hot encoded labels to <code>fit()<\/code> method. Therefore you get the shape mismatch error.<\/p>\n\n<hr>\n\n<p><strong>Side note:<\/strong> as @Alexis mentioned in the comments, please add the error you get (if applicable) when asking questions.<\/p>\n"}]}
+{"QuestionId":53174066,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-11-06T14:36:19.527","AcceptedAnswerId":null,"OwnerUserId":7867517.0,"Title":"Why my sparse tensor cannot convert to Tensor ?","Body":"<p>I want to feed my sparse data to Tensorflow by batch, so I use these code:<\/p>\n\n<pre><code>with graph.as_default():\n    global_step = tf.Variable(0, trainable=False)\n    # Here we use the indices and values to reproduce the input SparseTensor\n    sp_indice = tf.placeholder(tf.int64)\n    sp_value = tf.placeholder(tf.float32)\n    x =  tf.SparseTensor(sp_indice, sp_value, [batch_size, feature_num])\n    y_ = tf.placeholder(tf.float32, [None, 1])\n    keep_prob = tf.placeholder(\"float32\")\n    W = tf.Variable(tf.zeros([feature_num, 1]))\n    b = tf.Variable(tf.zeros([1]))\n\n    # Construct model\n    pred = tf.nn.softmax(tf.matmul(x, W) + b) # Softmax\n    # Minimize error using cross entropy\n    cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred), reduction_indices=1))\n<\/code><\/pre>\n\n<p>But I got an error:<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class 'tensorflow.python.framework.sparse_tensor.SparseTensor'&gt; to Tensor. Contents: SparseTensor(indices=Tensor(\"Placeholder:0\", dtype=int64), values=Tensor(\"Placeholder_1:0\", dtype=float32), dense_shape=Tensor(\"SparseTensor\/dense_shape:0\", shape=(2,), dtype=int64)). Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>I am confused about this error, it seems like my <code>sp_indice = tf.placeholder(tf.int64)<\/code> was wrong, maybe because I did not give the <code>shape<\/code>.  But I cannot make sure the shape since the shape is different in every batch. How can I deliver index array and value array to the tensor?<\/p>\n","answers":[{"AnswerId":"53174553","CreationDate":"2018-11-06T15:04:33.033","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>When you are doing matrix multiplication of a sparse tensor, you need to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_tensor_dense_matmul\" rel=\"nofollow noreferrer\"><code>tf.sparse_tensor_dense_matmul<\/code><\/a>.<\/p>\n\n<pre><code>pred = tf.nn.softmax(tf.sparse_tensor_dense_matmul(x, W) + b)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53174250,"AnswerCount":0,"Tags":"<python><matlab><tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-06T14:47:12.770","AcceptedAnswerId":null,"OwnerUserId":5931298.0,"Title":"\"Same\" network on MATLAB and Keras has very different results","Body":"<p>I have been trying to replicate the same simple network structure in MATLAB and Keras. The problem is the accuracy I get is very different. MATLAB code gets accuracy near 0.84 and loss near 17 and Keras code gets accuracy near 0.63 and loss near 130, with Keras using double epochs to train and the same data. I think the difference is too big to be a matter of implementation, so I think I'm missing something.<\/p>\n\n<p>The original code is from a MATLAB example in which I have made a little change to avoid normalization in the first layer.<\/p>\n\n<p>Here is the MATLAB code:<\/p>\n\n<pre class=\"lang-matlab prettyprint-override\"><code>% Load the digit training set as 4-D array data using\n% |digitTrain4DArrayData|.\n[trainImages,~,trainAngles] = digitTrain4DArrayData;\n\ndisp(\"Train Images:\")\ndisp(trainImages(:,:,:,1))\n% Display 20 random sample training digits using |imshow|.\nnumTrainImages = size(trainImages,4);\n\nfigure\nidx = randperm(numTrainImages,20);\nfor i = 1:numel(idx)\n    subplot(4,5,i)\n\n    imshow(trainImages(:,:,:,idx(i)))\n    drawnow\nend\n\n%%\n% Combine all the layers together in a |Layer| array.\nlayers = [ ...\n    imageInputLayer([28 28 1], 'Normalization', 'none')\n    convolution2dLayer(12,25)\n    reluLayer\n    fullyConnectedLayer(1)\n    regressionLayer];\n\n%% Train Network'\n\noptions = trainingOptions('sgdm','InitialLearnRate',0.001, ...\n    'MaxEpochs',15)\n\nnet = trainNetwork(trainImages,trainAngles,layers,options)\n\nnet.Layers\n\n%% Test Network\n[testImages,~,testAngles] = digitTest4DArrayData;\n\npredictedTestAngles = predict(net,testImages);\n\n% *Evaluate Performance*\npredictionError = testAngles - predictedTestAngles;\n\nthr = 10;\nnumCorrect = sum(abs(predictionError) &lt; thr);\nnumTestImages = size(testImages,4);\n\naccuracy = numCorrect\/numTestImages\n\n%%\n% Use the root-mean-square error (RMSE) to measure the differences between\n% the predicted and actual angles of rotation.\nsquares = predictionError.^2;\nrmse = sqrt(mean(squares))\n\n%%\n% *Display Box Plot of Residuals for Each Digit Class*\nresiduals = testAngles - predictedTestAngles;\n\nresidualMatrix = reshape(residuals,500,10);\n\nfigure\nboxplot(residualMatrix, ...\n    'Labels',{'0','1','2','3','4','5','6','7','8','9'})\n\nxlabel('Digit Class')\nylabel('Degrees Error')\ntitle('Residuals')\n\nidx = randperm(numTestImages,49);\nfor i = 1:numel(idx)\n    image = testImages(:,:,:,idx(i));\n    predictedAngle = predictedTestAngles(idx(i));\n\n    imagesRotated(:,:,:,i) = imrotate(image,predictedAngle,'bicubic','crop');\nend\n\nfigure\nsubplot(1,2,1)\nmontage(testImages(:,:,:,idx))\ntitle('Original')\n\nsubplot(1,2,2)\nmontage(imagesRotated)\ntitle('Corrected')\n<\/code><\/pre>\n\n<p>Here is the Keras code:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.io\nimport matplotlib.pyplot as plt\nfrom keras.datasets import mnist\nfrom keras.models import Sequential\nfrom keras.layers import Dropout, Flatten, Dense, Activation, Conv2D, BatchNormalization, AveragePooling2D\nfrom keras.optimizers import SGD\nfrom keras.utils import np_utils\nfrom keras import regularizers\n\nnp.random.seed(1671) # for reproducibility\n# network and training\nNB_EPOCH = 30\nBATCH_SIZE = 128\nVERBOSE = 1\nNB_CLASSES = 10 # number of outputs = number of digits\nOPTIMIZER = SGD() # SGD optimizer, explained later in this chapter\nN_HIDDEN = 128\nVALIDATION_SPLIT=0.2 # how much TRAIN is reserved for VALIDATION\n\ndata = scipy.io.loadmat('RegressionImageData.mat')\n\nXTrain = np.rollaxis(data['XTrain'],3,0)\nXTest = np.rollaxis(data['XTest'],3,0)\nYTest = np.squeeze(data['YTest'])\nYTrain = np.squeeze(data['YTrain'])\n\nprint(\"Train Images:\")\nprint(XTrain.shape)\nprint(type(XTrain))\nprint(XTrain)\n\nXTrain_test = np.reshape(XTrain, (5000,28,28))\n\nwith open(\".\/test.txt\", \"a+\") as file:\n    np.set_printoptions(threshold=np.nan)\n    file.write(np.array2string(XTrain_test[0], max_line_width=np.inf))\n\nmodel = Sequential()\nmodel.add(Conv2D(25,(12,12), input_shape=(28,28,1), strides=(1,1), activation = \"relu\"))\nmodel.add(Flatten())\nmodel.add(Dense(1))\nmodel.summary()\n\nsgd = SGD(lr=0.001, decay=0.1, momentum=0.9, nesterov=False)\nmodel.compile(loss='mean_squared_error', optimizer=sgd)\n\nhistory = model.fit(XTrain, YTrain,\nbatch_size=BATCH_SIZE, epochs=NB_EPOCH,\nverbose=VERBOSE, validation_split=VALIDATION_SPLIT,\nshuffle=False)\n\npredictions= model.predict(XTrain)\n[np.transpose(predictions[1:50]), np.transpose(YTrain[1:50]), np.abs(np.transpose(predictions[1:50])- np.transpose(YTrain[1:50]))]\n\nplt.plot(history.history['loss'])\nplt.plot(history.history['val_loss'])\nplt.title('rmse')\nplt.ylabel('loss')\nplt.xlabel('epoch')\nplt.legend(['train', 'test'], loc='upper left')\nplt.show()\n\nYpred_test= model.predict(XTest)\nYpred_test=np.reshape(Ypred_test, (5000,))\npredictionError = Ypred_test - YTest\nthr = 10;\nnumCorrect = np.sum((np.abs(predictionError) &lt; thr)*1)\nnumValidationImages = len(YTest)\n\naccuracy = numCorrect\/numValidationImages\nprint(accuracy)\nsquares = np.power(predictionError,2)\nrmse = np.sqrt(np.mean(squares))\nprint(rmse)\n<\/code><\/pre>\n\n<p>Anyone know where could be the gap?<\/p>\n","answers":[]}
+{"QuestionId":53174590,"AnswerCount":2,"Tags":"<java><android><tensorflow><tensorflow-lite>","CreationDate":"2018-11-06T15:06:54.010","AcceptedAnswerId":null,"OwnerUserId":8531947.0,"Title":"Cannot copy between TFLite tensor - illegal shape","Body":"<p>Im currently making my own model and all works fine with the tensorflow-for-poets-2 demo. I trained multiple pictures in different folders, and the app recognized it.<\/p>\n\n<p>Now I want to display a bounding box around the object. I found an example <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/lite\/examples\/android\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>My problem is that my app is returning following error when I am adding my own tflite model:<\/p>\n\n<pre><code>E\/AndroidRuntime: FATAL EXCEPTION: inference\nProcess: org.tensorflow.lite.demo, PID: 3495\njava.lang.IllegalArgumentException: Cannot copy between a TensorFlowLite tensor with shape [1, 6] and a Java object with shape [1, 10, 4].\n    at org.tensorflow.lite.Tensor.throwExceptionIfTypeIsIncompatible(Tensor.java:240)\n    at org.tensorflow.lite.Tensor.copyTo(Tensor.java:116)\n    at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:157)\n    at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:229)\n    at org.tensorflow.demo.TFLiteObjectDetectionAPIModel.recognizeImage(TFLiteObjectDetectionAPIModel.java:194)\n    at org.tensorflow.demo.DetectorActivity$3.run(DetectorActivity.java:247)\n    at android.os.Handler.handleCallback(Handler.java:789)\n    at android.os.Handler.dispatchMessage(Handler.java:98)\n    at android.os.Looper.loop(Looper.java:164)\n    at android.os.HandlerThread.run(HandlerThread.java:65)\n<\/code><\/pre>\n\n<p>How I trained them:<\/p>\n\n<pre><code>python3 scripts\/retrain.py \\\n  --bottleneck_dir=bottlenecks \\\n  --how_many_training_steps=500 \\\n  --model_dir=inception \\\n  --output_graph=tf_files\/retrained_graph.pb \\\n  --output_labels=tf_files\/retrained_labels.txt \\\n  --image_dir=tf_files\/ \\\n  --architecture mobilenet_1.0_224\n<\/code><\/pre>\n\n<p>Generating:<\/p>\n\n<pre><code>toco \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --input_file=tf_files\/retrained_graph.pb \\\n  --output_format=TFLITE \\\n  --output_file=tf_files\/optimized_graph.lite \\\n  --inference_type=FLOAT \\\n  --inference_input_type=FLOAT \\\n  --input_arrays=input \\\n  --output_arrays=final_result \\\n  --input_shapes=1,224,224,3\\\n  --mean_values=128 \\\n  --std_values=128 \\\n  --default_ranges_min=0 \\\n  --default_ranges_max=6\n<\/code><\/pre>\n\n<p>DetectorActivity.java<\/p>\n\n<pre><code>\/\/ Configuration values for the prepackaged SSD model.\nprivate static final int TF_OD_API_INPUT_SIZE = 224; \/\/ 300\nprivate static final boolean TF_OD_API_IS_QUANTIZED = false; \/\/ true\nprivate static final String TF_OD_API_MODEL_FILE = \"optimized_graph.lite\"; \/\/detect.tflite\nprivate static final String TF_OD_API_LABELS_FILE = \"file:\/\/\/android_asset\/retrained_labels.txt\";\n<\/code><\/pre>\n","answers":[{"AnswerId":"53322657","CreationDate":"2018-11-15T15:24:48.957","ParentId":null,"OwnerUserId":"8694824","Title":null,"Body":"<p>What is your current use case? Is it the same as of Tensorflow Poets. The error is evident that your tensorflow model output shape is not in par with the application. The use case of the application may be different from your use case.<\/p>\n"},{"AnswerId":"57522217","CreationDate":"2019-08-16T09:33:08.303","ParentId":null,"OwnerUserId":"5542648","Title":null,"Body":"<p>I came to the solution.\nit's just because of the model you prepare to use<\/p>\n\n<p>tfLite.run(imgData,outputScores);<\/p>\n\n<p>private static final int NUM_DETECTIONS = 6;<\/p>\n\n<p>It works at my side <\/p>\n"}]}
+{"QuestionId":53174610,"AnswerCount":1,"Tags":"<tensorflow><random><initializer>","CreationDate":"2018-11-06T15:08:13.067","AcceptedAnswerId":53183434.0,"OwnerUserId":6299165.0,"Title":"Tensorflow: tf.random_normal get different results with the same initial seed","Body":"<p>I want to make a reusable <code>RANDOM TENSOR x<\/code> and assign the SAME tensor to  <code>VARIABLE y<\/code>. That means they should have the same value during <code>Session.run()<\/code>.<\/p>\n\n<p>But it turns out not the case. So why does <code>y<\/code> NOT equal <code>x<\/code>?<\/p>\n\n<p>Update:\nAfter applying <code>sess.run(x)<\/code> and <code>sess.run(y)<\/code> multiple times in line, confirmed that <code>x<\/code> changes every time while <code>y<\/code> stays steady. Why?<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.random_normal([3], seed = 1)\ny = tf.Variable(initial_value = x) # expect y get the same random tensor as x\n\ndiff = tf.subtract(x, y)\navg = tf.reduce_mean(diff)\n\nsess = tf.InteractiveSession()\nsess.run(y.initializer)\n\nprint('x0:', sess.run(x))\nprint('y0:', sess.run(y))\nprint('x1:', sess.run(x))\nprint('y1:', sess.run(y))\nprint('x2:', sess.run(x))\nprint('y2:', sess.run(y))\nprint('diff:', sess.run(diff))\nprint('avg:', sess.run(avg)) # expected as 0.0\n\nsess.close()\n<\/code><\/pre>\n\n<p>Ouputs: TENSOR x changes every sess.run(x)<\/p>\n\n<pre><code>x0: [ 0.55171245 -0.13107552 -0.04481386]\ny0: [-0.8113182   1.4845988   0.06532937]\nx1: [-0.67590594  0.28665832  0.3215887 ]\ny1: [-0.8113182   1.4845988   0.06532937]\nx2: [1.2409041  0.44875884 0.33140722]\ny2: [-0.8113182   1.4845988   0.06532937]\ndiff: [ 1.2404865  -1.4525002   0.05412297]\navg: -0.04116\n<\/code><\/pre>\n","answers":[{"AnswerId":"53183434","CreationDate":"2018-11-07T04:06:12.113","ParentId":null,"OwnerUserId":"6299165","Title":null,"Body":"<p>The true cause is that:\n<code>x = tf.random_normal(seed = initial_seed)<\/code>is evolving every time when applying <code>sess.run()<\/code> but produces the same tensor series <code>x0-x1-x2<\/code> if <strong>restart<\/strong> running the script. <a href=\"https:\/\/stackoverflow.com\/questions\/22639587\/random-seed-what-does-it-do\">Here<\/a> provides some explanation on <strong>random seed<\/strong>.<\/p>\n\n<p>To guarantee the same <code>x<\/code> after every first run, we need <strong>reinitialize<\/strong> it. Not sure there is a decent way for my case. But we can set <code>x<\/code> as a variable and initialize with a fixed seed. Either <code>tf.get_variable<\/code> or <code>tf.Variable<\/code> is OK. I find <a href=\"https:\/\/stackoverflow.com\/questions\/34888235\/how-do-i-generate-a-random-vector-in-tensorflow-and-maintain-it-for-further-use\/34888586#34888586\">this answer<\/a> fit my question.<\/p>\n\n<p>Here is my final code. It works.<\/p>\n\n<pre><code>import tensorflow as tf\n\ninitializer = tf.random_normal_initializer(seed = 1)\nx = tf.get_variable(name = 'x', shape = [3], dtype = tf.float32, initializer = initializer)\ny = tf.Variable(initial_value = x)\n\ndiff = tf.subtract(x, y)\navg = tf.reduce_mean(diff)\n\nsess = tf.InteractiveSession()\nsess.run(tf.global_variables_initializer())\n\nprint('x0:', sess.run(x))\nprint('y0:', sess.run(y))\n\nprint('x1:', sess.run(x))\nprint('y1:', sess.run(y))\n\nprint('x2:', sess.run(x))\nprint('y2:', sess.run(y))\n\nprint('diff:', sess.run(diff))\nprint('avg:', sess.run(avg))\nsess.close()\n\nx0: [-0.8113182   1.4845988   0.06532937]\ny0: [-0.8113182   1.4845988   0.06532937]\nx1: [-0.8113182   1.4845988   0.06532937]\ny1: [-0.8113182   1.4845988   0.06532937]\nx2: [-0.8113182   1.4845988   0.06532937]\ny2: [-0.8113182   1.4845988   0.06532937]\ndiff: [0. 0. 0.]\navg: 0.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":53175678,"AnswerCount":3,"Tags":"<python><tensorflow><mnist>","CreationDate":"2018-11-06T16:10:04.547","AcceptedAnswerId":53217492.0,"OwnerUserId":10226157.0,"Title":"Loss for classification using MNIST data stays around the same through every epoch","Body":"<pre><code>import numpy as np\nimport tensorflow as tf\nimport pandas as pd\ndata = pd.read_csv('mnist_train.csv')\nX = data.drop('label', axis=1).values\ny = data['label'].values\nwith tf.Session() as sess:\n    Y = tf.one_hot(y, 10).eval()\nhidden  = [5, 4, 3]\n\ndef costa(y, yhat):\n    loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=yhat, labels=y)\n    loss = tf.reduce_sum(loss)\n    return loss\n\ndef train(cost):\n    train_op = tf.train.GradientDescentOptimizer(0.0001).minimize(cost)\n    return train_op\n\nwith tf.Graph().as_default():\n    X1 = tf.placeholder(tf.float32, [None, 784])\n    y1 = tf.placeholder(tf.float32, [None, 10])\n    w1 = tf.Variable(tf.random_normal((784, hidden[0])))\n    w2 = tf.Variable(tf.random_normal((hidden[0], hidden[1])))\n    w3 = tf.Variable(tf.random_normal((hidden[1], hidden[2])))\n    wo = tf.Variable(tf.random_normal((hidden[2], 10)))\n    b1 = tf.Variable(tf.random_normal((1, hidden[0])))\n    b2 = tf.Variable(tf.random_normal((1, hidden[1])))\n    b3 = tf.Variable(tf.random_normal((1, hidden[2])))\n    bo = tf.Variable(tf.random_normal((1, 10)))\n    layer1 = tf.nn.relu(tf.matmul(X1, w1) + b1)\n    layer2 = tf.nn.relu(tf.matmul(layer1, w2) + b2)\n    layer3 = tf.nn.relu(tf.matmul(layer2, w3) + b3)\n    layerout = (tf.matmul(layer3, wo) + bo)\n    yhat = layerout\n    cost = costa(y1, yhat)\n    train_op = train(cost)\n    init_op = tf.global_variables_initializer()\n    for epoch in range(1000):\n\n        with tf.Session() as sess:\n            sess.run(init_op)\n            sess.run(train_op, feed_dict={X1:X, y1:Y})\n            loss = sess.run(cost, feed_dict={X1:X, y1:Y})\n            print(\"Loss for epoch {}: {}\".format(epoch, loss))\n<\/code><\/pre>\n\n<p>The loss stays around the same, jumps up and down a lot, but does not decrease accordingly.<\/p>\n\n<p>I can't seem to find what is going wrong here, any help would be appeciated.<\/p>\n\n<p>Is it the activations to the layers or am I getting the cost function wrong?<\/p>\n","answers":[{"AnswerId":"53178941","CreationDate":"2018-11-06T19:45:51.860","ParentId":null,"OwnerUserId":"4830338","Title":null,"Body":"<p>There are a couple of issues here:<\/p>\n\n<ol>\n<li>You are running <code>sess.run(init_op)<\/code> every epoch. This means that the model parameters are being reset to random numbers every epoch, and therefore will be unable to learn. Try putting this op before <code>for epoch in range(1000)<\/code><\/li>\n<li><p>You are creating a new session every epoch. Change your code so it looks like this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(init_op)\n\n    for epoch in range(1000):\n        sess.run(train_op, feed_dict={X1:X, y1:Y})\n        loss = sess.run(cost, feed_dict={X1:X, y1:Y})\n        print(\"Loss for epoch {}: {}\".format(epoch, loss))\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"53209372","CreationDate":"2018-11-08T14:05:28.747","ParentId":null,"OwnerUserId":"10226157","Title":null,"Body":"<p>Okay one little tweak did the trick, I used RMSPropOptimizer instead and the loss started decreasing as expected. <\/p>\n\n<p>I still have to figure out as to why this works, I\u2019m still learning, but for now this is the solution I have. <\/p>\n\n<p>Although the loss decreases very slowly. <\/p>\n"},{"AnswerId":"53217492","CreationDate":"2018-11-08T23:03:34.123","ParentId":null,"OwnerUserId":"10226157","Title":null,"Body":"<p>Initialising weights with a standard deviation of <code>(2.0\/neurons_in_prev_layer)**0.5<\/code> worked like a charm for me! \nAlso changed the hidden layers to, 2 hidden layers of 256, 256 neurons. <\/p>\n"}]}
+{"QuestionId":53175747,"AnswerCount":0,"Tags":"<indexing><pytorch>","CreationDate":"2018-11-06T16:13:31.977","AcceptedAnswerId":null,"OwnerUserId":1099913.0,"Title":"tensorflow matrix_band_part function equivalent in pytorch","Body":"<p>I need to create upper triangular masking tensor in pytorch. in tensorflow it's easy using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/linalg\/band_part\" rel=\"nofollow noreferrer\">matrix_band_part<\/a>.  is it any pytorch equivalent for this function?\nit's like numpy triu_indices but I need for tensor, not just matrix.  <\/p>\n","answers":[]}
+{"QuestionId":53175789,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><tensorflow-serving>","CreationDate":"2018-11-06T16:16:09.123","AcceptedAnswerId":53208983.0,"OwnerUserId":10307998.0,"Title":"How to do predictions without charging the model everytime - tensorflow?","Body":"<p>Hi I'm making a conv net with tf.estimators , and I want to predict with my trained model but when I upload an image always loads and closes the model, how to make it stay loaded and continue to upload images to make predictions.I have read something about tf.serving but I just want to run it on my computer<\/p>\n\n<p>this is my code:<\/p>\n\n<pre><code>#Convolutional Funcion .......................................\n\ndef cnn_model_fn(features, labels, mode):\n\n    \"\"\"Model\"\"\"\n\n    input_layer = tf.reshape(features[\"image\"], [-1, 224, 224, 3])\n\n    # Convolutional Layer #1\n    conv1 = tf.layers.conv2d(\n      inputs=input_layer,\n      filters=64,\n      kernel_size=[7, 7],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n    # Pooling Layer #1\n    pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)\n\n    # Convolutional Layer #2\n    conv2 = tf.layers.conv2d(\n      inputs=pool1,\n      filters=128,\n      kernel_size=[5, 5],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n    # Pooling Layer #2\n    pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)\n\n    # Convolutional Layer #3\n    conv3 = tf.layers.conv2d(\n      inputs=pool2,\n      filters=192,\n      kernel_size=[5, 5],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n    # Pooling Layer #3\n    pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2], strides=2)\n\n    # Convolutional Layer #4\n    conv4 = tf.layers.conv2d(\n      inputs=pool3,\n      filters=192,\n      kernel_size=[3, 3],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n    # Pooling Layer #4\n    pool4 = tf.layers.max_pooling2d(inputs=conv4, pool_size=[2, 2], strides=2)\n\n    # Convolutional Layer #5\n\n    conv5 = tf.layers.conv2d(\n      inputs=pool4,\n      filters=128,\n      kernel_size=[3, 3],\n      padding=\"same\",\n      activation=tf.nn.relu)\n\n    # Pooling Layer #5\n\n    pool5 = tf.layers.max_pooling2d(inputs=conv5, pool_size=[2, 2], strides=2)\n\n    # Flatten tensor into a batch of vectors\n    pool5_flat = tf.reshape(pool5, [-1, 7 * 7 * 128])\n\n    # Dense Layer\n    dense = tf.layers.dense(inputs=pool5_flat, units=2048, activation=tf.nn.relu)\n\n    # Add dropout operation; 0.6 probability that element will be kept\n    dropout = tf.layers.dropout(inputs=dense, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN)\n\n    dense_1 = tf.layers.dense(inputs=dropout, units=2048, activation=tf.nn.relu)\n\n    dropout_1 = tf.layers.dropout(inputs=dense_1, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN)\n\n    #Final Layer with 2 outputs\n    logits = tf.layers.dense(inputs=dropout_1, units=2)\n\n    #Predictions\n    output=tf.nn.softmax(logits)\n    predictions=tf.argmax(input=output, axis=1)\n    predictions={'Probabilities':output,'Prediction':predictions}\n\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        return tf.estimator.EstimatorSpec(mode=mode,predictions=predictions)\n\ndef read_img():\n\n    filenames = tf.constant([data_path])\n\n    dataset = tf.data.Dataset.from_tensor_slices(filenames)\n\n    def _parse_function(filename):\n        image_string = tf.read_file(filename)\n        image_decoded = tf.image.decode_jpeg(image_string)\n        image_resized = tf.image.resize_images(image_decoded, [224, 224])\n        return {'image':image_resized}\n\n    dataset = dataset.map(_parse_function)\n\n    return dataset\n\ndef main(params):\n\n    det = tf.estimator.Estimator(model_fn=cnn_model_fn, model_dir='\/Users\/David\/Desktop\/David\/Tesis')\n\n    pred_results=det.predict(input_fn=read_img)\n\n    print(next(pred_results))\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong>\nI obtain this every time:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/h3SA7.png\" rel=\"nofollow noreferrer\">results<\/a><\/p>\n\n<p>every time restore the parameters<\/p>\n\n<p><strong>EDIT 2<\/strong><\/p>\n\n<p>I have my code like this:<\/p>\n\n<pre><code>def first():\n  data_path=tf.placeholder(dtype=tf.string)\n  filenames = tf.constant([data_path])\n  dataset = tf.data.Dataset.from_tensor_slices(filenames)\n  dataset = dataset.map(_parse_function)\n\n  iterator = dataset.make_one_shot_iterator()\n  features= iterator.get_next()\n\n  global preds\n  preds=cnn_model_fn(features,None,tf.estimator.ModeKeys.PREDICT).predictions\n  tf.train.Saver().restore(sess,tf.train.latest_checkpoint('\/Users\/David\/Desktop'))\n\n\ndef second(path):\n\n  try:\n\n    while True:\n      print(sess.run(preds,feed_dict={data_path:path}))\n\n  except tf.errors.OutOfRangeError:\n    print('Done')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53208983","CreationDate":"2018-11-08T13:43:01.340","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>One option would be to swap entirely to TF Eager and simply rewrite your main, like so:<\/p>\n\n<pre><code>tf.enable_eager_execution()\n\ndef main(params):\n  for data in read_img():\n    print(det.model_fn(data, None, tf.estimator.ModeKeys.PREDICT).predictions)\n<\/code><\/pre>\n\n<p>Another would be to use sessions:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  data = read_img().one_shot_iterator()\n  preds = det.model_fn(data, None, tf.estimator.ModeKeys.PREDICT).predictions\n  tf.train.Saver().restore(sess, det.model_dir)\n  while True:\n    sess.run(preds)\n<\/code><\/pre>\n\n<p>both of these are just pseudocode, so the API names etc. may be off.<\/p>\n"}]}
+{"QuestionId":53175991,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-06T16:27:10.067","AcceptedAnswerId":null,"OwnerUserId":10563517.0,"Title":"How can I make predictions from a trained model inside a Tensorflow input pipeline?","Body":"<p>I am trying to train a model for emotion recognition, which uses one of VGG's layer's output as an input.<\/p>\n\n<p>I could manage what I want by running the prediction in a first step, saving the extracted features and then using them as input to my network, but I am looking for a way to do the whole process at once.<\/p>\n\n<p>The second model uses a concatenated array of feature maps as input (I am working with video data), so I am not able to simply wire it to the output of VGG.<\/p>\n\n<p>I tried to use a map operation as depicted in the <code>tf.data.dataset<\/code> API documentations this way :<\/p>\n\n<pre><code>def trimmed_vgg16():\n  vgg16 = tf.keras.applications.vgg16.VGG16(input_shape=(224,224,3))\n  trimmed = tf.keras.models.Model(inputs=vgg16.get_input_at(0),\n                                  outputs=vgg16.layers[-3].get_output_at(0))\n  return trimmed\n\nvgg16 = trimmed_vgg16()\n\ndef _extract_vgg_features(images, labels):\n    pred = vgg16_model.predict(images, batch_size=batch_size, steps=1)\n    return pred, labels\n\ndataset = #load the dataset (image, label) as usual\ndataset = dataset.map(_extract_vgg_features)\n<\/code><\/pre>\n\n<p>But I'm getting this error : <code>Tensor Tensor(\"fc1\/Relu:0\", shape=(?, 4096), dtype=float32) is not an element of this graph<\/code> which is pretty explicit. I'm stuck here, as I don't see a good way of inserting the trained model in the same graph and getting predictions \"on the fly\".<\/p>\n\n<p>Is there a clean way of doing this or something similar ?<\/p>\n\n<p>Edit: missed a line.<br>\nEdit2: added details<\/p>\n","answers":[{"AnswerId":"53176283","CreationDate":"2018-11-06T16:44:26.993","ParentId":null,"OwnerUserId":"2542043","Title":null,"Body":"<p>You should be able to connect the layers by first creating the <code>vgg16<\/code> and then retrieving the output of the model as such and afterward you can use that tensor as an input to your own network.<\/p>\n\n<pre><code>vgg16 = tf.keras.applications.vgg16.VGG16(input_shape=(224,224,3))\n\nnetwork_input = vgg16.get_input_at(0)\nvgg16_out = vgg16.layers[-3].get_output_at(0) # use this tensor as input to your own network\n<\/code><\/pre>\n"}]}
+{"QuestionId":53176018,"AnswerCount":0,"Tags":"<python><ubuntu><tensorflow><environment-variables>","CreationDate":"2018-11-06T16:28:53.337","AcceptedAnswerId":null,"OwnerUserId":3445394.0,"Title":"setting python path environment variable for tensorflow models in ubuntu","Body":"<p>I am following <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/v1.11\/official\" rel=\"nofollow noreferrer\">TensorFlow Models version 1.11<\/a> to add the path to models to my <code>PYTHONPATH<\/code>.<\/p>\n\n<p>I wrote in the end of <code>~\/.bashrc<\/code> the following command:<\/p>\n\n<pre><code># setting PYTHONPATH\nexport PYTHONPATH=\"$PYTHONPATH:\/home\/yij8\/Documents\/notchange\/models\"\n<\/code><\/pre>\n\n<p>where <code>\/home\/yij8\/Documents\/notchange\/models<\/code> is the path to my models folder.<\/p>\n\n<p>However after I restarted the terminal, and typed in <\/p>\n\n<pre><code>pip3 install --user -r official\/requirements.txt\n<\/code><\/pre>\n\n<p>it showed<\/p>\n\n<pre><code>Could not open requirements file: [Errno 2] No such file or directory: 'official\/requirements.txt'\n<\/code><\/pre>\n\n<p>which made me really confused.<\/p>\n\n<p>I checked the python path by <code>echo $PYTHONPATH<\/code>, it showed<\/p>\n\n<pre><code>:\/home\/yij8\/Documents\/notchange\/models\n<\/code><\/pre>\n\n<p>Could anyone tell me where I did wrong?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":53176173,"AnswerCount":0,"Tags":"<python><tensorflow><raspberry-pi>","CreationDate":"2018-11-06T16:37:56.287","AcceptedAnswerId":null,"OwnerUserId":938512.0,"Title":"Convert Tensorflow model to TFLITE format","Body":"<p>I've followed <a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">these instructions<\/a> many times to train a model on my own images.  It works perfectly and I can then label my images fine after training.   I want to use this trained model on a Raspberry Pi Zero.  <\/p>\n\n<p>I moved the model to the pi, installed tensorflow with pip and tried to use the standard label_image.py script.   This understandably runs out of resources and dies because the model is 83MB.  I want to use tfite to slim the model down.   When I issue:<\/p>\n\n<pre><code>   toco \\\n    --graph_def_file=output_graph.pb \\\n    --output_file=tf_files\/optimized_graph.lite \\\n    --input_format=TENSORFLOW_GRAPHDEF \\\n    --output_format=TFLITE \\\n    --input_array=Placeholder \\\n    --output_array=final_result \\\n    --inference_type=FLOAT \\\n    --input_data_type=FLOAT\n<\/code><\/pre>\n\n<p>I get a toco usage message and then just the error \"toco: error: \"<\/p>\n\n<p>It's unclear to me which parameters are mandatory and what they should be.    What's the easiest way to convert a model to TFLITE for the pi zero?<\/p>\n","answers":[]}
+{"QuestionId":53176735,"AnswerCount":1,"Tags":"<python><tensorflow><keras><object-detection>","CreationDate":"2018-11-06T17:10:15.423","AcceptedAnswerId":53190939.0,"OwnerUserId":8478967.0,"Title":"Implementing a custom loss function for object detection","Body":"<p>I am new to keras and tensorflow . How do I go about implementing a custom loss function while doing object detection , right now I have 5 parameters -  4 for bounding box coordinates and 1 for whether the object is present or not . <strong>Loss function should return square of difference between coordinates if object is present else if object is absent it should return a huge value as loss<\/strong> .  This is the code I am tring right now:<\/p>\n\n<pre><code>def loss_func(y_true,y_pred):\n  mask = np.array([False, False, False,False,True])   # check column of the class of object\n  mask1 = np.array([True, True, True,True,False])     # get the columns of the coordinates of B box\n  check_class = K.mean(K.square(tf.subtract(tf.boolean_mask(y_true,mask),tf.boolean_mask(y_pred,mask))))\n  mean_square = K.mean(K.square(tf.subtract(tf.boolean_mask(y_true,mask1),tf.boolean_mask(y_pred,mask1))))\n  value=K.mean(tf.boolean_mask(y_pred,mask))\n\n  return value*mean_square + check_class\n<\/code><\/pre>\n\n<p>Here I am masking other values to obtain the last value which is 1000--> object present 0 --> object absent.\nIs there any other better way to do this?<\/p>\n\n<p>The value of loss when I am running this in Kaggle decreases rapidly , by 2nd epoch the loss becomes 0.<\/p>\n","answers":[{"AnswerId":"53190939","CreationDate":"2018-11-07T13:58:44.517","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>First of all i would recommend using 1 and not 1000 for the \"image exist\" parameter.<br>\nYou can manipulate y_true and y_pred.<\/p>\n\n<pre><code>penalty = 100\n\ndef lf(y_true,y_pred):\n    mean_square = tf.keras.losses.mean_squared_error(y_true[:,0:4], y_pred[:,0:4])\n    check_class = tf.subtract(y_true[:,4], y_pred[:,4])\n    check_class = check_class * -penalty\n    check_class = tf.keras.backend.mean(check_class)\n    return mean_square + check_class\n<\/code><\/pre>\n\n<p>The above function first check the mean squared error for the first 4 parameters.\n<br>\nthe second part then check for the \"present\" parameter.\n<br>\nIf they are different it will output -1, if they are the same it will output 0.\n<br>\nThen it use the penalty for punishing it for a wrong parameter.\n<br><br>\nUsing \"punish\" by some constant can be difficult to train. I would recommend changing the optimizer to <strong>SGD<\/strong>, adam will not work well in the situation, and playing with the penalty until you reach satisfying results.<\/p>\n"}]}
+{"QuestionId":53177032,"AnswerCount":1,"Tags":"<python><tensorflow><recommendation-engine><recommender-systems>","CreationDate":"2018-11-06T17:30:21.440","AcceptedAnswerId":null,"OwnerUserId":3790814.0,"Title":"WALS Model Tensorflow - Get recommendations for new user","Body":"<p>I wonder if there is any way I can get recommendations for a new user, using an already trained WALS model, and given the list of items the user liked. <\/p>\n\n<p>Currently, to get a recommendation you must provide the id of the user, which must be among the users the model was trained with. I would like to get a recommendation by providing the list of items that were liked by a new user.<\/p>\n\n<p>There is a similar feature in the <a href=\"https:\/\/github.com\/benfred\/implicit\/pull\/35\" rel=\"nofollow noreferrer\">implicit<\/a> python library <\/p>\n","answers":[{"AnswerId":"54593574","CreationDate":"2019-02-08T13:35:07.733","ParentId":null,"OwnerUserId":"5161762","Title":null,"Body":"<p>What you are describing is called \"cold start problem\".<\/p>\n\n<p>You want a recommendation from no information on user's past behaviour.<\/p>\n\n<p>As Matrix factorization needs historical data from the user, the best recommendation you can make if a most-popular recommendation.<\/p>\n\n<p>To tackle this problem, you can add another type of model that handle this. You can, for example, add user's data in the algorithm in order not to be absolutely dependant on past data or by doing the online recommendation by analysing data during session.<\/p>\n"}]}
+{"QuestionId":53177270,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-06T17:46:49.677","AcceptedAnswerId":null,"OwnerUserId":8067584.0,"Title":"Google colab TPU bugged or I do something wrong?","Body":"<p>Trying to port my dcgan to tpu.\nBut this error:<\/p>\n\n<pre><code>Error recorded from training_loop: File system scheme '[local]' not implemented\n<\/code><\/pre>\n\n<p>Here is notebook\n<a href=\"https:\/\/colab.research.google.com\/drive\/101FjBAIMVuXyNyeUvq_Vfx-Z6CR3g4df\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/101FjBAIMVuXyNyeUvq_Vfx-Z6CR3g4df<\/a><\/p>\n","answers":[{"AnswerId":"53215745","CreationDate":"2018-11-08T20:35:28.703","ParentId":null,"OwnerUserId":"4716286","Title":null,"Body":"<p>The local filesystem is not <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting#cannot_use_local_filesystem\" rel=\"nofollow noreferrer\">available<\/a> on Cloud TPU's. You can use Google Cloud Storage (GCS), like in this <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/shakespeare_with_tpuestimator.ipynb\" rel=\"nofollow noreferrer\">example<\/a>. GCS is <a href=\"https:\/\/cloud.google.com\/free\/\" rel=\"nofollow noreferrer\">free<\/a> up to 5GB a month.<\/p>\n\n<p>Alternatively Keras models running on TPU's copy variables back to the CPU so they can be used without GCS for the model directory.<\/p>\n"}]}
+{"QuestionId":53177494,"AnswerCount":1,"Tags":"<python><tensorflow><batch-normalization>","CreationDate":"2018-11-06T18:03:23.220","AcceptedAnswerId":53290273.0,"OwnerUserId":6254131.0,"Title":"How to deal with bool valued needed operation like tf.nn.fused_batch_norm in Tensorflow","Body":"<p>For functions like <code>tf.layers.batch_normalization<\/code>, it is easy to use a placeholder <code>train_flag<\/code> as input for the <code>training<\/code> argument, and after we define the whole network, we can feed in <code>True<\/code> or <code>False<\/code> for <code>train_flag<\/code> under training or inference phases.<\/p>\n\n<p>However, for operations like <code>tf.nn.fused_batch_norm<\/code> (defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/ops\/nn_impl.py\" rel=\"nofollow noreferrer\">nn_impl.py<\/a>), which only accepts python bool as input for argument <code>is_training<\/code>, does it mean that we will need to construct the network twice under training and inference phases with difference <code>is_training<\/code> arguments? <\/p>\n","answers":[{"AnswerId":"53290273","CreationDate":"2018-11-13T22:11:30.250","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<p>In the definition file of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/keras\/layers\/normalization.py\" rel=\"nofollow noreferrer\"><code>tf.keras.layers.BatchNormalization<\/code><\/a>, <\/p>\n\n<pre><code>def _fused_batch_norm(self, inputs, training):\n    \"\"\"Returns the output of fused batch norm.\"\"\"\n\n    def _fused_batch_norm_training():\n        return nn.fused_batch_norm(\n            inputs,\n            self.gamma,\n            self.beta,\n            epsilon=self.epsilon)\n\n    def _fused_batch_norm_inference():\n        return nn.fused_batch_norm(\n            inputs,\n            self.gamma,\n            self.beta,\n            mean=self.moving_mean,\n            variance=self.moving_variance,\n            epsilon=self.epsilon,\n            is_training=False)\n\n    output, mean, variance = tf_utils.smart_cond(\n        training, _fused_batch_norm_training, _fused_batch_norm_inference)\n    if not self._bessels_correction_test_only:\n        # Remove Bessel's correction to be consistent with non-fused batch norm.\n        # Note that the variance computed by fused batch norm is\n        # with Bessel's correction.\n        sample_size = math_ops.cast(\n                array_ops.size(inputs) \/ array_ops.size(variance), variance.dtype)\n        factor = (sample_size - math_ops.cast(1.0, variance.dtype)) \/ sample_size\n        variance *= factor\n\n    training_value = tf_utils.constant_value(training)\n    if training_value is None:\n        momentum = tf_utils.smart_cond(training,\n                                       lambda: self.momentum,\n                                       lambda: 1.0)\n    else:\n        momentum = ops.convert_to_tensor(self.momentum)\n    if training_value or training_value is None:\n        mean_update = self._assign_moving_average(self.moving_mean, mean,\n                                                  momentum)\n        variance_update = self._assign_moving_average(self.moving_variance,\n                                                      variance, momentum)\n        self.add_update(mean_update, inputs=True)\n        self.add_update(variance_update, inputs=True)\n\n    return output\n<\/code><\/pre>\n\n<p>It utilizes <code>training_value = tf_utils.constant_value(training)<\/code> and then uses <code>if training_value is None<\/code>, where <code>tf_utils.constant_value<\/code> should be <code>tf.contrib.util.constant_value<\/code><\/p>\n"}]}
+{"QuestionId":53177605,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-06T18:11:01.570","AcceptedAnswerId":53178130.0,"OwnerUserId":7112888.0,"Title":"Model not learning - keras","Body":"<h3>Background<\/h3>\n\n<p>I have a pretty simple script that creates a keras model designed to act like an XOR gate.<\/p>\n\n<p>I generate 40000 datapoints in the <code>get_data<\/code> function. It creates two arrays; an input array containing 1s and 0s in some order, and an output which is either a 1 or a 0. <\/p>\n\n<h3>Issue<\/h3>\n\n<p>When I run the code it does not appear to learn and the results I get vary dramatically every time I train it. <\/p>\n\n<h3>Code<\/h3>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>from keras import models\nfrom keras import layers\n\nimport numpy as np\n\nfrom random import randint\n\n\ndef get_output(a, b): return 0 if a == b else 1\n\n\ndef get_data ():\n    data = []\n    targets = []\n\n    for _ in range(40010):\n        a, b = randint(0, 1), randint(0, 1)\n\n        targets.append(get_output(a, b))\n        data.append([a, b])\n\n    return data, targets\n\n\ndata, targets = get_data()\n\ndata = np.array(data).astype(\"float32\")\ntargets = np.array(targets).astype(\"float32\")\n\ntest_x = data[40000:]\ntest_y = targets[40000:]\n\ntrain_x = data[:40000]\ntrain_y = targets[:40000]\n\nmodel = models.Sequential()\n\n# input\nmodel.add(layers.Dense(2, activation='relu', input_shape=(2,)))\n\n# hidden\n# model.add(layers.Dropout(0.3, noise_shape=None, seed=None))\nmodel.add(layers.Dense(2, activation='relu'))\n# model.add(layers.Dropout(0.2, noise_shape=None, seed=None))\nmodel.add(layers.Dense(2, activation='relu'))\n\n# output\nmodel.add(layers.Dense(1, activation='sigmoid')) # sigmoid puts between 0 and 1\n\nmodel.summary() # print out summary of model\n\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n\nres = model.fit(train_x, train_y, epochs=2000, batch_size=200, validation_data=(test_x, test_y)) # train\n\nprint 'predict: \\n', test_x\nprint model.predict(test_x)\n<\/code><\/pre>\n\n<h3>Output<\/h3>\n\n<pre class=\"lang-python prettyprint-override\"><code>[[0. 1.]\n [1. 1.]\n [1. 1.]\n [0. 0.]\n [1. 0.]\n [0. 0.]\n [0. 0.]\n [0. 1.]\n [1. 1.]\n [1. 0.]]\n[[0.6629775 ]\n [0.00603844]\n [0.00603844]\n [0.6629775 ]\n [0.6629775 ]\n [0.6629775 ]\n [0.6629775 ]\n [0.6629775 ]\n [0.00603844]\n [0.6629775 ]]\n<\/code><\/pre>\n\n<h3>Edit<\/h3>\n\n<p>Even without the dropout layers, I got very similar results. <\/p>\n","answers":[{"AnswerId":"53178130","CreationDate":"2018-11-06T18:47:33.340","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>There are several issues with your question.<\/p>\n\n<p>To start with, your imports are rather unorthodox (irrelevant to your issue, true, but it helps sticking to some conventions):<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>from keras.models import Sequential\nfrom keras.layers import Dense\nimport numpy as np\n<\/code><\/pre>\n\n<p>Second, you don't need some thousands of examples for the XOR problem; there are only four combinations:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>X = np.array([[0,0],[0,1],[1,0],[1,1]])\ny = np.array([[0],[1],[1],[0]])\n<\/code><\/pre>\n\n<p>and that's all.<\/p>\n\n<p>Third, for the very same reason, you can't actually have \"validation\" or \"test\" data with XOR; in the simplest approach (i.e. what you are arguably trying to do here), you can only test how well the model has learnt the function, using these 4 combinations (since there are no more!).<\/p>\n\n<p>Fourth, you should start with a simple one-hidden layer model (with somewhat more than 2 units and no dropout), and then proceed gradually <em>if needed<\/em>:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>model = Sequential()\nmodel.add(Dense(8, activation=\"relu\", input_dim=2))\nmodel.add(Dense(1, activation=\"sigmoid\"))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam')\nmodel.fit(X, y, batch_size=1, epochs=1000)\n<\/code><\/pre>\n\n<p>This should take your loss down to ~ 0.12; how well has it learnt the function?<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>model.predict(X)\n# result:\narray([[0.31054294],\n       [0.9702552 ],\n       [0.93392825],\n       [0.04611744]], dtype=float32)\n\ny\n# result:\narray([[0],\n       [1],\n       [1],\n       [0]])\n<\/code><\/pre>\n\n<p>Is this good enough? Well, I don't know - the correct answer is always \"it depends\"! But you now have a starting point (i.e. a network that arguably <em>learns<\/em> something), from which you can proceed to further experiments...<\/p>\n"}]}
+{"QuestionId":53177958,"AnswerCount":1,"Tags":"<python><keras><deep-learning><computer-vision><conv-neural-network>","CreationDate":"2018-11-06T18:35:37.503","AcceptedAnswerId":null,"OwnerUserId":10032996.0,"Title":"How to reduce overfiting while using VGG16 for regression?","Body":"<p>i'm using transfer learning from VGG16 for a regression task but i get over-fit very quickly. I want to reduce the number of parameters for the regression (last layer), how can i do it?<\/p>\n","answers":[{"AnswerId":"53180887","CreationDate":"2018-11-06T22:14:15.633","ParentId":null,"OwnerUserId":"9041710","Title":null,"Body":"<p>Assuming you're not re-training the inception layers and just adding an output layer there's no way to reduce the number of trainable parameters, you could however try to fight overfitting by adding a <code>Dropout<\/code> layer in between or something like that.<\/p>\n\n<p>Although you should bear in mind that VGG16's weights were trained with a loss function suitable for classification tasks (like categorical crossentropy), which gives me the impression that your model will only guess the length based on what the object is (which may even be the idea but a large car that looks a bit like an insect might be given a smaller size than a small car that looks like a building).<\/p>\n"}]}
+{"QuestionId":53178052,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><keras><deep-learning>","CreationDate":"2018-11-06T18:42:49.913","AcceptedAnswerId":null,"OwnerUserId":10614772.0,"Title":"Real time Keras Object Recognition with OpenCV","Body":"<p>I tried loading a fine tuned model saved as a .h5 file to this script but it just kept loading in Jupyter Notebook. What is the correct argument to use so I can deploy the model on a real time camera feed?<\/p>\n\n<pre><code> from keras.preprocessing import image\nfrom keras import load_model\nimport argparse\nimport cv2\nimport numpy as np\nimport os\nimport random\nimport sys\n\nimport threading\n\nlabel = ''\n\nframe = None\n\nclass MyThread(threading.Thread):\ndef __init__(self):\n    threading.Thread.__init__(self)\n\ndef run(self):\n    global label\n    # Load the VGG16 network\n    print(\"[INFO] loading network...\")\n    self.model = load_model(PiVision.h5)\n\n    while (~(frame is None)):\n        (inID, label) = self.predict(frame)\n\ndef predict(self, frame):\n    image = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB).astype(np.float32)\n    image = image.transpose((2, 0, 1))\n    image = image.reshape((1,) + image.shape)\n\n    image = preprocess_input(image)\n    preds = self.model.predict(image)\n    return decode_predictions(preds)[0]\n\ncap = cv2.VideoCapture(0)\n\nif (cap.isOpened()):\nprint(\"Camera OK\")\nelse:\ncap.open()\n\nkeras_thread = MyThread()\n\nkeras_thread.start()\n\nwhile (True):\nret, original = cap.read()\n\nframe = cv2.resize(original, (224, 224))\n\n# Display the predictions\n# print(\"ImageNet ID: {}, Label: {}\".format(inID, label))\ncv2.putText(original, \"Label: {}\".format(label), (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.9, (0, 255, 0), 2)\ncv2.imshow(\"Classification\", original)\n\nif (cv2.waitKey(1) &amp; 0xFF == ord('q')):\n    break;\n\ncap.release()\nframe = None\ncv2.destroyAllWindows()\nsys.exit()\n<\/code><\/pre>\n\n<p>also keras does not have the option to save labels.pbtxt so how is it gonna print the labels?<\/p>\n\n<p>is there a way I can create an instance to pass in the labels I have for my classifier<\/p>\n","answers":[]}
+{"QuestionId":53178458,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-11-06T19:10:07.347","AcceptedAnswerId":null,"OwnerUserId":7733184.0,"Title":"Giving a time series input to Pytorch-LSTM using a Batch size of 128","Body":"<p>I have a dataset of n training samples (Where X matrix = n * 30 * 20) where 30 is the number of sequences, 20 is the number of features. I am using categorical cross-entropy loss since this is a classification task. My input matrix is of shape 10000 *30 * 20 and target matrix is 1000 * 1. I was wondering what is the correct way to give the input to the LSTM. My entire code is attached below. <\/p>\n\n<pre><code>import torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils.data as data\nimport torch.backends.cudnn as cudnn\nimport torch.optim as optim\nimport torchvision.transforms as transforms\nimport torch.utils.data as utils\nfrom torch.utils.data import DataLoader\nimport torch\nfrom torch.autograd import Variable\n\nclass MyData(data.Dataset):\n    def __init__(self,X,y,transform=None, target_transform=None):\n        self.data = X\n        self.targets = y\n        self.transform = transform\n        self.target_transform= target_transform\n    def __getitem__(self, index):\n        item, target = self.data[index], self.targets[index]\n        item = torch.FloatTensor(item)\n        target = torch.LongTensor([target])\n        return item, target\n\n    def __len__(self):\n        return len(self.data)\n\nclass MyPredictor(nn.Module):\n    def __init__(self, num_blocks=1, input_dim=22, batch_size=128, hidden_dim1=64, hidden_dim2=128, GPU=True):\n        super(MyPredictor, self).__init__()\n        self.batch_size= batch_size\n        self.hidden_dim1=hidden_dim1\n        self.hidden_dim2=hidden_dim2\n        self.GPU=GPU\n\n        self.lstm1 = nn.LSTM(input_dim,self.hidden_dim1,num_layers=2)\n        self.lstm2 = nn.LSTM(self.hidden_dim1, self.hidden_dim2)\n        self.fc1 = nn.Linear(self.hidden_dim2, 256)\n        self.fc2=nn.Linear(256,1)\n\n\n    if self.GPU:\n        self.hx1 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim1).cuda())\n        self.cx1 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim1).cuda())\n        self.hx2 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim2).cuda())\n        self.cx2 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim2).cuda())\n    else:\n        self.hx1 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim1))\n        self.cx1 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim1))\n        self.hx2 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim2))\n        self.cx2 = Variable(torch.zeros(30, self.batch_size, self.hidden_dim2))\n\n        self.hidden1=(self.hx1,self.cx1)\n        self.hidden2=(self.hx2,self.cx2)\n\n    def forward(self, x):\n\n        out1, self.hidden1 = self.lstm1(x, self.hidden1)\n        out2, self.hidden2 = self.lstm1(x, self.hidden2)\n        out3=self.fc1(out2)\n        out3= F.relu(out3)\n        out4=self.fc2(out3)\n        out=F.relu(out4)\n\n        return out\n<\/code><\/pre>\n\n<p>Moreover, I am initializing the hidden and cell states as 1 * Batch_size * hidden size. However since I have 30 as sequence length, should it be 30 * batch_size * hidden size?<\/p>\n\n<pre><code>def train(epoch):\n    print('\\nEpoch: %d' % epoch)\n    net.train()\n    train_loss = 0\n    correct = 0\n    total = 0\n    store=[]\n    for batch_idx, (inputs, targets) in enumerate(trainloader):\n        inputs, targets = inputs.to(device), targets.to(device)\n        optimizer.zero_grad()\n        outputs = net(inputs)\n        loss = criterion(outputs, targets)\n        loss.backward()\n        optimizer.step()\n\n        train_loss += loss.item()\n        _, predicted = outputs.max(1)\n        total += targets.size(0)\n        correct += predicted.eq(targets).sum().item()\n\n#         progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d\/%d)'\n#             % (train_loss\/(batch_idx+1), 100.*correct\/total, correct, total))\n    print(\"Train-&gt;Epoch: \",epoch,\" Accuracy: \",100.*correct\/total,\" Loss \",train_loss\/len(trainloader))\n    return[100.*correct\/total,train_loss\/len(trainloader)]\n\n\ndef test(epoch):\n    global best_acc\n    criterion = nn.CrossEntropyLoss()\n    net.eval()\n    test_loss = 0\n    correct = 0\n    total = 0\n    with torch.no_grad():\n        for batch_idx, (inputs, targets) in enumerate(testloader):\n            inputs, targets = inputs.to(device), targets.to(device)\n            outputs = net(inputs)\n            loss = criterion(outputs, targets)\n\n            test_loss += loss.item()\n            _, predicted = outputs.max(1)\n            total += targets.size(0)\n            correct += predicted.eq(targets).sum().item()\n\n#             progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d\/%d)'\n#                 % (test_loss\/(batch_idx+1), 100.*correct\/total, correct, total))\n\n    print(\"Test-&gt;Epoch: \",epoch,\" Accuracy: \",100.*correct\/total,\" Loss \",test_loss\/len(testloader))\n<\/code><\/pre>\n\n<p>I am loading my inputs as below:<\/p>\n\n<pre><code>from sklearn.externals import joblib\nimport numpy as np\n\n#See if Gpu is available\ndevice = 'cuda' if torch.cuda.is_available() else 'cpu'\nnet=MyPredictor()      \nnet = net.to(device)\nGPU=False\nif device == 'cuda':\n    net = torch.nn.DataParallel(net)\n    cudnn.benchmark = True\n    GPU= True\n\n#optimizer = optim.SGD(model.parameters(), lr=0.1)\noptimizer = optim.Adam(net.parameters(), lr=0.001)\ncriterion = nn.CrossEntropyLoss()\n\n#Load the datset\n\n\nX, y = joblib.load(\"\/app\/dhanush\/lstm_training_data.pkl\")\nX_train=np.array(X[0:30000])\nX_test=np.array(X[30000:])\ny_train= np.array(y[0:30000])\ny_test= np.array(y[30000:])\n\nprint (X_train.shape)\nprint (y_train.shape)\n\ntrainset = MyData(X_train,y_train)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=False, num_workers=2)\n\ntestset = MyData(X_test, y_test)\ntestloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False, num_workers=2)\n\n\nfor epoch in range(0, 21):\n    train(epoch)\n    test(epoch)\n<\/code><\/pre>\n\n<p>I get the below shape mismatch error. I definitely think it has something to do with my input tensors or the way I am initializing my hidden states. Also, I am not sure if my input to the first fully connected layer is correct. Should I give only the last hidden state output from the LSTM 2 <code>(self.hx2[-1]?)<\/code><\/p>\n\n<p>On reading other posts, I see that for 2D inputs, you are sometimes asked to reshape to <code>(sequence * batch size * feature)<\/code> for direct input to LSTM. How does this work ?<\/p>\n\n<p>And most important I get the below error :<\/p>\n\n<pre><code>RuntimeError: Expected hidden[0] size (1, 30, 64), got (30, 128, 64)\n<\/code><\/pre>\n\n<p>I understand this is all a lot but my goal here is to provide as much information as possible. I have read other posts, sample GitHub codes but most of them deal with using an embedding later which I don't want to use.<\/p>\n","answers":[]}
+{"QuestionId":53178462,"AnswerCount":1,"Tags":"<python><pandas><tensorflow>","CreationDate":"2018-11-06T19:10:14.317","AcceptedAnswerId":53191103.0,"OwnerUserId":10550693.0,"Title":"Unable to read folder files correctly using pandas","Body":"<p>I am trying to implement simple features DNN using tensorflow. I need to iterate through every folder in the directory and the code will read every csv files contain in a particular folder. I have got multiple folders in A_Device_Motion directory with name as 'dws_1', 'dws_2', 'jog_9' etc. Each folder contains data of 24 subjects activity in csv files. I am trying to iterate throught the folder data and giving Activity variable a specific name with each iteration. But the problem i am facing is that my code is not able to read the folder name completely and giving me a keyError. And my folder strcuture looks like this <a href=\"https:\/\/i.stack.imgur.com\/FGvHw.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FGvHw.jpg\" alt=\"Here\"><\/a>\nHere is my Traceback <\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"C:\\Users\\TechProBox\\Desktop\\python.py\", line 21, in &lt;module&gt;\ndf['Activity'] = Activety_Types[j[49:52]]\nKeyError: 'ws_'\n<\/code><\/pre>\n\n<p>My Python Code<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport sklearn\nfrom glob import glob\nimport os\nimport itertools\nimport tensorflow as tf\nprint(os.listdir(\"C:\/Users\/TechProBox\/Desktop\/A_DeviceMotion_data\"))\nActivety_Types = {'dws':1,'jog':2,'sit':3,'std':4,'ups':5,'wlk':6}        \nlistDict = list(Activety_Types.keys())\nFolders = glob('C:\/Users\/TechProBox\/Desktop\/A_DeviceMotion_data\/*_*')\nFolders = [s for s in Folders if \"csv\" not in s]\nDf_all_list = []\nExp = 0\nSegment_Size = 400\nfor j  in Folders:\n  Csv = glob(j + '\/*' )\n  for i in Csv:\n    df = pd.read_csv(i)\n    df['Activity'] = Activety_Types[j[49:52]]\n    df['Sub_Num'] = i[len(j)+5:-4]\n    df['Exp_num'] = 1\n    ExpNum = np.zeros((df.shape[0])) \n    for i in range(0,df.shape[0]-Segment_Size,Segment_Size):\n        ExpNum[range(i,i+Segment_Size)] = i\/Segment_Size +Exp*100 \n    df['Exp_num'] = ExpNum\n    Df_all_list.append(df)\n    Exp += 1        \n\nDf_all = pd.concat(Df_all_list,axis=0)  \nDf_all.head()\nplt.plot([1,2,3])\nfor i in range(6):\nD = Df_all[Df_all['Activity']==i+1]\nplt.subplot(3,2,i+1)\nplt.plot(D['userAcceleration.z'][:200])\nplt.title(listDict[i])\nplt.ylim([-1, 1])\n\nplt.tight_layout()\n<\/code><\/pre>\n\n<p>I did not pasted the whole code here but i think this code is enough for understanding. I am facing error on this line <code>df['Activity'] = Activety_Types[j[49:52]]<\/code>\nPS: Code is not mine i took it from kaggle.<\/p>\n","answers":[{"AnswerId":"53191103","CreationDate":"2018-11-07T14:08:42.193","ParentId":null,"OwnerUserId":"10550693","Title":null,"Body":"<p>Apparently i was giving wrong names of the folders that needed to be added and there were some other minor mistakes in the code too. I had to give exact names of each and every folder to run the program and it worked here is the correct code<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport sklearn\nfrom glob import glob\nimport os\nimport itertools\nimport tensorflow as tf\nprint(os.listdir(\"A_DeviceMotion_data\"))\nActivety_Types = {'dws_1':1,'jog_9':2,'sit_5':3,'std_6':4,'ups_3':5,'wlk_7':6, 'dws_2':7, 'dws_11':8, 'jog_16':9, 'sit_13':10, 'std_14':11, 'ups_4':12, 'ups_12':13,'wlk_8':14,'wlk_15':15}        \nlistDict = list(Activety_Types.keys())\nFolders = glob('C:\/Users\/TechProBox\/Desktop\/A_DeviceMotion_data\/*_*')\nFolders = [s for s in Folders if \"csv\" not in s]\nDf_all_list = []\nExp = 0\nSegment_Size = 400\nfor j  in Folders:\n    print(\"Folder {}\\n\".format(j))\n    Csv = glob(j + '\/*' )\n    for i in Csv:\n        print (i)\n        print (j)\n        print(\"HERE\")\n        df = pd.read_csv(i)\n        print(Activety_Types[j[48:56]])\n        df['Activity'] = Activety_Types[j[48:56]]\n        print(\"HERE1\")\n\n        df['Sub_Num'] = i[len(j)+5:-4]\n        df['Exp_num'] = 1\n        ExpNum = np.zeros((df.shape[0])) \n        for i in range(0,df.shape[0]-Segment_Size,Segment_Size):\n            ExpNum[range(i,i+Segment_Size)] = i\/Segment_Size +Exp*100 \n        df['Exp_num'] = ExpNum\n        Df_all_list.append(df)\n        Exp += 1        \n\nDf_all = pd.concat(Df_all_list,axis=0)  \nDf_all.head()\nplt.plot([1,2,3])\nfor i in range(6):\n    D = Df_all[Df_all['Activity']==i+1]\n    plt.subplot(3,2,i+1)\n    plt.plot(D['userAcceleration.z'][:200])\n    plt.title(listDict[i])\n    plt.ylim([-1, 1])\nplt.tight_layout()\n\ndf_sum = Df_all.groupby('Exp_num', axis=0).mean().reset_index()\ndf_sum.columns = df_sum.columns.str.replace('.','_sum_')\n\ndf_sum_SS = np.power(Df_all.astype(float),2).groupby('Exp_num', axis=0).median().reset_index() \ndf_sum_SS.columns = df_sum_SS.columns.str.replace('.','_sumSS_')\n\ndf_max = Df_all.groupby('Exp_num', axis=0).max().reset_index()\ndf_max.columns = df_max.columns.str.replace('.','_max_')\n\ndf_min = Df_all.groupby('Exp_num', axis=0).min().reset_index()\ndf_min.columns = df_min.columns.str.replace('.','_min_')\n\ndf_skew = Df_all.groupby('Exp_num', axis=0).skew().reset_index()\ndf_skew.columns = df_skew.columns.str.replace('.','_skew_')\n\ndf_std = Df_all.groupby('Exp_num', axis=0).std().reset_index()\ndf_std.columns = df_std.columns.str.replace('.','_std_')\n\nDf_Features = pd.concat([ df_max , df_sum[df_sum.columns[2:-2]], \n                         df_min[df_min.columns[2:-2]], df_sum_SS[df_sum_SS.columns[2:-2]], \n                         df_std[df_std.columns[2:-2]], df_skew[df_skew.columns[2:-2]]], axis=1)\n\nDf_Features_1 = Df_Features.drop(['Exp_num','Unnamed: 0','Activity','Sub_Num'],axis=1)\nLabels = Df_Features['Activity']\n\nfrom sklearn.model_selection import train_test_split\nX_train, X_test, y_train, y_test = train_test_split(Df_Features_1, Labels, test_size=0.25, random_state=42)\n\nX_shape = Df_Features_1.shape[1]\nCluss_Num = len(Activety_Types)\nX = tf.placeholder(dtype=tf.float32,shape=[None,X_shape])\ny = tf.placeholder(dtype=tf.float32,shape = [None,Cluss_Num])\nHold_prob = tf.placeholder(tf.float32)\n\n\ndef Init_Wightes(shape1):\n    W = tf.truncated_normal(shape1,stddev=0.1)\n    return tf.Variable(W)\n\ndef Init_bias(shape1):\n    b =  tf.constant(0.1,shape=shape1)\n    return tf.Variable(b)\n\ndef FC_layer(input1,shape1):\n\n    W = Init_Wightes(shape1)\n    B = Init_bias([shape1[1]])\n    Wx = tf.matmul(input1,W)\n    Wx_b = tf.add(Wx,B)\n    return tf.nn.relu(Wx_b)\n\n\nH1_size = 54\nH2_size = 24\n\nH1 = FC_layer(X, [X_shape,H1_size])\nH1_drop = tf.nn.dropout(H1,keep_prob=Hold_prob)\ny_pred = FC_layer(H1_drop,[H1_size,len(Activety_Types)])\n\nErr = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y,logits=y_pred))\nOptimaizer = tf.train.AdamOptimizer()\nTrain = Optimaizer.minimize(Err)\nInit = tf.global_variables_initializer()\n\nBatchSize = 64\nlabel_binarizer = sklearn.preprocessing.LabelBinarizer()\nlabel_binarizer.fit(range(Cluss_Num))\n\n\nX_train_Np = np.array(X_train,dtype=np.float32)\nY_train_Np = np.array(y_train,dtype=np.float32)\nX_test_Np = np.array(X_test,dtype=np.float32)\nY_test_Np = np.array(y_test,dtype=np.float32)\nY_test_OH  = label_binarizer.transform(Y_test_Np)\nY_test_OH = np.array(Y_test_OH,dtype=np.float32)\nBatches = np.array(range(0,X_train_Np.shape[0]-BatchSize,BatchSize))\naccuracy = 10\nmisclass = 1\nsteps = 5000\n\nwith tf.Session() as sess:\n    sess.run(Init)\n    for i in range(steps): \n        BastchNum = np.mod(i,len(Batches)-1)\n        batch_x = X_train_Np[Batches[BastchNum] : Batches[BastchNum+1] ,:]\n        batch_y = Y_train_Np[Batches[BastchNum] : Batches[BastchNum+1] ]\n        batch_y_OH  = label_binarizer.transform(batch_y)\n        batch_y_OH = np.array(batch_y_OH,dtype=np.float32)\n        sess.run(Train,feed_dict={X:batch_x,y:batch_y_OH,Hold_prob:0.5}) \n        if i%100 == 0:   \n            print('Currently on step {}'.format(i))\n            print('Accuracy is:')\n            # Test the Train Model\n            matches = tf.equal(tf.argmax(y_pred,1),tf.argmax(y,1))\n            acc = tf.reduce_mean(tf.cast(matches,tf.float32))\n            print(sess.run(acc,feed_dict={X:X_test_Np,y:Y_test_OH,Hold_prob:1.0}))\n            print('\\n')\n            Conf = tf.confusion_matrix(tf.arg_max(y_pred,1),tf.arg_max(y,1))\n            C1 = sess.run(Conf,feed_dict={X:X_test_Np,y:Y_test_OH,Hold_prob:1.0})\ndef plot_confusion_matrix(cm,\n    target_names,\n    title='Confusion matrix',\n    cmap=None,\n    normalize=True):\n    accuracy = np.trace(cm) \/ float(np.sum(cm))\n    misclass = 1 - accuracy\n    if cmap is None:\n        cmap = plt.get_cmap('Blues')\n        plt.figure(figsize=(8, 6))\n        plt.imshow(cm, interpolation='nearest', cmap=cmap)\n        plt.title(title)\n        plt.colorbar()\n        if target_names is not None:\n            tick_marks = np.arange(len(target_names))\n            plt.xticks(tick_marks, target_names, rotation=45)\n            plt.yticks(tick_marks, target_names)\n        if normalize:\n            cm = cm.astype('float') \/ cm.sum(axis=1)[:, np.newaxis]\n        thresh = cm.max() \/ 1.5 if normalize else cm.max() \/ 2\n        for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n            if normalize:\n                plt.text(j, i, \"{:0.4f}\".format(cm[i, j]),\n                    horizontalalignment=\"center\",\n                    color=\"white\" if cm[i, j] &gt; thresh else \"black\")\n            else:\n                plt.text(j, i, \"{:,}\".format(cm[i, j]),\n                    horizontalalignment=\"center\",\n                    color=\"white\" if cm[i, j] &gt; thresh else \"black\")\nplt.tight_layout()\nplt.ylabel('True label')\nplt.xlabel('Predicted label\\naccuracy={:0.4f}; misclass={:0.4f}'.format(accuracy, misclass))\nplt.show()\nConf = tf.confusion_matrix(tf.arg_max(y_pred,1),tf.arg_max(y,1))\nplot_confusion_matrix(C1,target_names=[*Activety_Types])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53178596,"AnswerCount":0,"Tags":"<r><tensorflow><keras><rstudio><keras-layer>","CreationDate":"2018-11-06T19:19:55.200","AcceptedAnswerId":null,"OwnerUserId":8715382.0,"Title":"R Keras Error: All inputs to the layer should be tensors","Body":"<p>I'm running the following code and getting an error when I run the last part adding the layers. I'm not sure how to fix the error. I'm using version 2.1.5 of Keras and 3.5.1 of R. <\/p>\n\n<pre><code>library(keras)\n\n    mnist &lt;- dataset_mnist()\n    x_train &lt;- mnist$train$x\n    y_train &lt;- mnist$train$y\n    x_test &lt;- mnist$test$x\n    y_test &lt;- mnist$test$y\n\n    #     The x data is a 3-d array (images,width,height) of grayscale values . \n    #     To prepare the data for training we convert the 3-d arrays into matrices by reshaping \n    #     width and height into a single dimension (28x28 images are flattened into length 784 vectors). \n    #     Then, we convert the grayscale values from integers ranging between 0 to 255 into \n    #     floating point values ranging between 0 and 1:\n\n    # reshape\n\n    x_train &lt;- array_reshape(x_train, c(nrow(x_train), 784))\n    x_test &lt;- array_reshape(x_test, c(nrow(x_test), 784))\n\n    # rescale\n\n    x_train &lt;- x_train \/ 255\n    x_test &lt;- x_test \/ 255\n\n    #   Note that we use the array_reshape() function rather than the dim&lt;-() function to reshape the array. \n    #   This is so that the data is re-interpreted using row-major semantics (as opposed to R\u2019s default \n    #   column-major semantics), which is in turn compatible with the way that the numerical libraries \n    #   called by Keras interpret array dimensions.\n\n    #   The y data is an integer vector with values ranging from 0 to 9. To prepare this data for training \n    #   we one-hot encode the vectors into binary class matrices using the Keras to_categorical() function:\n\n    y_train &lt;- to_categorical(y_train, 10)\n    y_test &lt;- to_categorical(y_test, 10)\n\n    model &lt;- keras_model_sequential() \n    model %&gt;% \n      layer_dense(units = 256, activation = 'relu', input_shape = c(784)) %&gt;% \n      layer_dropout(rate = 0.4) %&gt;% \n      layer_dense(units = 128, activation = 'relu') %&gt;%\n      layer_dropout(rate = 0.3) %&gt;%\n      layer_dense(units = 10, activation = 'softmax')\n<\/code><\/pre>\n\n<p>Error in py_call_impl(callable, dots$args, dots$keywords) : \n  ValueError: Layer dense_1 was called with an input that isn't a symbolic tensor. Received type: class 'keras.engine.sequential.Sequential'. Full input: keras.engine.sequential.Sequential object at 0x000000002A2B90F0. All inputs to the layer should be tensors. <\/p>\n","answers":[]}
+{"QuestionId":53179255,"AnswerCount":1,"Tags":"<python><lstm><pytorch>","CreationDate":"2018-11-06T20:09:10.513","AcceptedAnswerId":53179615.0,"OwnerUserId":3262154.0,"Title":"Question about terminology used in LSTM inputs - seq_length vs context_size for sliding window approach","Body":"<p>I have time series of a sequence of <em>n<\/em> vectors that I need to feed to a LSTM with a sliding window approach. <\/p>\n\n<p>In different resources that I read online, <code>seq_length<\/code> is often referred to as the window length [or the number of LSTM cells] and <code>context_size<\/code> is defined as the input size to LSTM at a timestep. (for eg. one input vector at a time step).<\/p>\n\n<p>At any given point of time <em>t<\/em>, I want to pass points <em>x_{t-m}, \u2026, x_{t-1},\u2026,x_{t}<\/em> to LSTM followed by a dense layer and predict a categorical target attribute at every tilmestep. <\/p>\n\n<p>If I want to follow the sliding window approach, do I need to explicitly split input data into windows of size <em>m<\/em> ? <\/p>\n\n<p>For instance: <\/p>\n\n<p><em>[x_{t-m}, \u2026, x_{t-1},\u2026,x_{t}], [x_{t-m+1}, \u2026, x_{t},\u2026,x_{t+1}], [x_{t-m+2}, \u2026, x_{t+1}, x_{t+2}]<\/em> etc.<\/p>\n\n<p>Or, can I split the input data into non-overlapping chunks <em>[x_{t-m}, \u2026, x_{t-1},\u2026,x_{t}], [x_{t+1}, \u2026, x_{t+m-1},x_{t+m}]<\/em> etc. and instead resize the context_size ?<\/p>\n\n<pre><code>     embeds = embeds.unfold(1, context_size, 1)  # Keeping the step size to be 1\n     embeds = embeds.view(embeds.size(0), embeds.size(1), -1)\n<\/code><\/pre>\n\n<p>Is there a better way to implement sliding window approach for timeseries data ?<\/p>\n","answers":[{"AnswerId":"53179615","CreationDate":"2018-11-06T20:35:45.403","ParentId":null,"OwnerUserId":"4400634","Title":null,"Body":"<p>LSTM by default recurses on data, in that the prediction at time t will depend on all of the past(depending on memory). In this case it seems like you want input at time t to depend on m+1 instances into the past. You do not need a recurrent net if so, you can simple use linear and feed in the sliding window at each instant. However, if you're using a recurrent net, you don't need to pass the same inputs again and again. <\/p>\n"}]}
+{"QuestionId":53179795,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-11-06T20:48:25.850","AcceptedAnswerId":null,"OwnerUserId":10266265.0,"Title":"AttributeError: module 'tensorflow.app' has no attribute 'flags'","Body":"<pre><code>import tensorflow as tf\nfrom models.nets import cpm_hand_slim\nimport numpy as np\nfrom utils import cpm_utils\nimport cv2\nimport time\nimport math\nimport sys\ndef del_all_flags(FLAGS):\n    flags_dict = FLAGS._flags()    \n    keys_list = [keys for keys in flags_dict]    \n    for keys in keys_list:\n        FLAGS.__delattr__(keys)\n\ndel_all_flags(tf.flags.FLAGS)\nFLAGS = tf.app.flags.FLAGS\ntf.app.flags.DEFINE_string('DEMO_TYPE',\n                           #default_value='test_imgs\/roger.png',\n                           default_value='test_imgs\/longhand.jpg',\n                           #default_value='SINGLE',\n                           docstring='MULTI: show multiple stage,'\n                                    'SINGLE: only last stage,'\n                                     'HM: show last stage heatmap,'\n                                     'paths to .jpg or .png image')\ntf.app.flags.DEFINE_string('model_path',\n                           default_value='models\/weights\/cpm_hand.pkl',\n                           docstring='Your model')\ntf.app.flags.DEFINE_integer('input_size',\n                            default_value=368,\n                            docstring='Input image size')\ntf.app.flags.DEFINE_integer('hmap_size',\n                            default_value=46,\n                            docstring='Output heatmap size')\ntf.app.flags.DEFINE_integer('cmap_radius',\n                            default_value=21,\n                            docstring='Center map gaussian variance')\ntf.app.flags.DEFINE_integer('joints',\n                            default_value=21,\n                            docstring='Number of joints')\ntf.app.flags.DEFINE_integer('stages',\n                            default_value=6,\n                            docstring='How many CPM stages')\ntf.app.flags.DEFINE_integer('cam_num',\n                            default_value=0,\n                            docstring='Webcam device number')\ntf.app.flags.DEFINE_bool('KALMAN_ON',\n                         default_value=True,\n                         docstring='enalbe kalman filter')\ntf.app.flags.DEFINE_float('kalman_noise',\n                            default_value=3e-2,\n                            docstring='Kalman filter noise value')\ntf.app.flags.DEFINE_string('color_channel',\n                           default_value='RGB',\n                           docstring='')\n\n# Set color for each finger\njoint_color_code = [[139, 53, 255],\n                    [0, 56, 255],\n                    [43, 140, 237],\n                    [37, 168, 36],\n                    [147, 147, 0],\n                    [70, 17, 145]]\n\n\nlimbs = [[0, 1],\n         [1, 2],\n         [2, 3],\n         [3, 4],\n         [0, 5],\n         [5, 6],\n         [6, 7],\n         [7, 8],\n         [0, 9],\n         [9, 10],\n         [10, 11],\n         [11, 12],\n         [0, 13],\n         [13, 14],\n         [14, 15],\n         [15, 16],\n         [0, 17],\n         [17, 18],\n         [18, 19],\n         [19, 20]\n         ]\n\nif sys.version_info.major == 3:\n    PYTHON_VERSION = 3\nelse:\n    PYTHON_VERSION = 2\n\n\ndef main(argv):\n    tf_device = '\/gpu:0'\n    with tf.device(tf_device):\n        \"\"\"Build graph\n        \"\"\"\n        if FLAGS.color_channel == 'RGB':\n            input_data = tf.placeholder(dtype=tf.float32, shape=[None, FLAGS.input_size, FLAGS.input_size, 3],\n                                        name='input_image')\n        else:\n            input_data = tf.placeholder(dtype=tf.float32, shape=[None, FLAGS.input_size, FLAGS.input_size, 1],\n                                        name='input_image')\n\n        center_map = tf.placeholder(dtype=tf.float32, shape=[None, FLAGS.input_size, FLAGS.input_size, 1],\n                                    name='center_map')\n\n        model = cpm_hand_slim.CPM_Model(FLAGS.stages, FLAGS.joints + 1)\n        model.build_model(input_data, center_map, 1)\n\n    saver = tf.train.Saver()\n\n    \"\"\"Create session and restore weights\n    \"\"\"\n    sess = tf.Session()\n\n    sess.run(tf.global_variables_initializer())\n    if FLAGS.model_path.endswith('pkl'):\n        model.load_weights_from_file(FLAGS.model_path, sess, False)\n    else:\n        saver.restore(sess, FLAGS.model_path)\n\n    test_center_map = cpm_utils.gaussian_img(FLAGS.input_size, FLAGS.input_size, FLAGS.input_size \/ 2,\n                                             FLAGS.input_size \/ 2,\n                                             FLAGS.cmap_radius)\n    test_center_map = np.reshape(test_center_map, [1, FLAGS.input_size, FLAGS.input_size, 1])\n\n    # Check weights\n    for variable in tf.trainable_variables():\n        with tf.variable_scope('', reuse=True):\n            var = tf.get_variable(variable.name.split(':0')[0])\n            print(variable.name, np.mean(sess.run(var)))\n\n    if not FLAGS.DEMO_TYPE.endswith(('png', 'jpg')):\n        cam = cv2.VideoCapture(FLAGS.cam_num)\n\n    # Create kalman filters\n    if FLAGS.KALMAN_ON:\n        kalman_filter_array = [cv2.KalmanFilter(4, 2) for _ in range(FLAGS.joints)]\n        for _, joint_kalman_filter in enumerate(kalman_filter_array):\n            joint_kalman_filter.transitionMatrix = np.array([[1, 0, 1, 0], [0, 1, 0, 1], [0, 0, 1, 0], [0, 0, 0, 1]],\n                                                            np.float32)\n            joint_kalman_filter.measurementMatrix = np.array([[1, 0, 0, 0], [0, 1, 0, 0]], np.float32)\n            joint_kalman_filter.processNoiseCov = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]],\n                                                           np.float32) * FLAGS.kalman_noise\n    else:\n        kalman_filter_array = None\n\n    with tf.device(tf_device):\n\n        while True:\n            t1 = time.time()\n            if FLAGS.DEMO_TYPE.endswith(('png', 'jpg')):\n                test_img = cpm_utils.read_image(FLAGS.DEMO_TYPE, [], FLAGS.input_size, 'IMAGE')\n            else:\n                test_img = cpm_utils.read_image([], cam, FLAGS.input_size, 'WEBCAM')\n\n            test_img_resize = cv2.resize(test_img, (FLAGS.input_size, FLAGS.input_size))\n            print('img read time %f' % (time.time() - t1))\n\n            if FLAGS.color_channel == 'GRAY':\n                test_img_resize = np.dot(test_img_resize[..., :3], [0.299, 0.587, 0.114]).reshape(\n                    (FLAGS.input_size, FLAGS.input_size, 1))\n                cv2.imshow('color', test_img.astype(np.uint8))\n                cv2.imshow('gray', test_img_resize.astype(np.uint8))\n                cv2.waitKey(1)\n\n            test_img_input = test_img_resize \/ 256.0 - 0.5\n            test_img_input = np.expand_dims(test_img_input, axis=0)\n\n\n            if FLAGS.DEMO_TYPE.endswith(('png', 'jpg')):\n                # Inference\n                t1 = time.time()\n                predict_heatmap, stage_heatmap_np = sess.run([model.current_heatmap,\n                                                              model.stage_heatmap,\n                                                              ],\n                                                             feed_dict={'input_image:0': test_img_input,\n                                                                        'center_map:0': test_center_map})\n\n                # Show visualized image\n                demo_img = visualize_result(test_img, FLAGS, stage_heatmap_np, kalman_filter_array)\n                cv2.imshow('demo_img', demo_img.astype(np.uint8))\n                if cv2.waitKey(0) == ord('q'): break\n                print('fps: %.2f' % (1 \/ (time.time() - t1)))\n            elif FLAGS.DEMO_TYPE == 'MULTI':\n\n                # Inference\n                t1 = time.time()\n                predict_heatmap, stage_heatmap_np = sess.run([model.current_heatmap,\n                                                              model.stage_heatmap,\n                                                              ],\n                                                             feed_dict={'input_image:0': test_img_input,\n                                                                        'center_map:0': test_center_map})\n\n                # Show visualized image\n                demo_img = visualize_result(test_img, FLAGS, stage_heatmap_np, kalman_filter_array)\n                cv2.imshow('demo_img', demo_img.astype(np.uint8))\n                if cv2.waitKey(1) == ord('q'): break\n                print('fps: %.2f' % (1 \/ (time.time() - t1)))\n\n\n            elif FLAGS.DEMO_TYPE == 'SINGLE':\n\n                # Inference\n                t1 = time.time()\n                stage_heatmap_np = sess.run([model.stage_heatmap[5]],\n                                            feed_dict={'input_image:0': test_img_input,\n                                                       'center_map:0': test_center_map})\n\n                # Show visualized image\n                demo_img = visualize_result(test_img, FLAGS, stage_heatmap_np, kalman_filter_array)\n                cv2.imshow('current heatmap', (demo_img).astype(np.uint8))\n                if cv2.waitKey(1) == ord('q'): break\n                print('fps: %.2f' % (1 \/ (time.time() - t1)))\n\n\n            elif FLAGS.DEMO_TYPE == 'HM':\n\n                # Inference\n                t1 = time.time()\n                stage_heatmap_np = sess.run([model.stage_heatmap[FLAGS.stages - 1]],\n                                            feed_dict={'input_image:0': test_img_input,\n                                                       'center_map:0': test_center_map})\n                print('fps: %.2f' % (1 \/ (time.time() - t1)))\n\n                demo_stage_heatmap = stage_heatmap_np[len(stage_heatmap_np) - 1][0, :, :, 0:FLAGS.joints].reshape(\n                    (FLAGS.hmap_size, FLAGS.hmap_size, FLAGS.joints))\n                demo_stage_heatmap = cv2.resize(demo_stage_heatmap, (FLAGS.input_size, FLAGS.input_size))\n\n                vertical_imgs = []\n                tmp_img = None\n                joint_coord_set = np.zeros((FLAGS.joints, 2))\n\n                for joint_num in range(FLAGS.joints):\n                    # Concat until 4 img\n                    if (joint_num % 4) == 0 and joint_num != 0:\n                        vertical_imgs.append(tmp_img)\n                        tmp_img = None\n\n                    demo_stage_heatmap[:, :, joint_num] *= (255 \/ np.max(demo_stage_heatmap[:, :, joint_num]))\n\n                    # Plot color joints\n                    if np.min(demo_stage_heatmap[:, :, joint_num]) &gt; -50:\n                        joint_coord = np.unravel_index(np.argmax(demo_stage_heatmap[:, :, joint_num]),\n                                                       (FLAGS.input_size, FLAGS.input_size))\n                        joint_coord_set[joint_num, :] = joint_coord\n                        color_code_num = (joint_num \/\/ 4)\n\n                        if joint_num in [0, 4, 8, 12, 16]:\n                            if PYTHON_VERSION == 3:\n                                joint_color = list(\n                                    map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                            else:\n                                joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                            cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color,\n                                       thickness=-1)\n                        else:\n                            if PYTHON_VERSION == 3:\n                                joint_color = list(\n                                    map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                            else:\n                                joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                            cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color,\n                                       thickness=-1)\n\n                    # Put text\n                    tmp = demo_stage_heatmap[:, :, joint_num].astype(np.uint8)\n                    tmp = cv2.putText(tmp, 'Min:' + str(np.min(demo_stage_heatmap[:, :, joint_num])),\n                                      org=(5, 20), fontFace=cv2.FONT_HERSHEY_COMPLEX, fontScale=0.3, color=150)\n                    tmp = cv2.putText(tmp, 'Mean:' + str(np.mean(demo_stage_heatmap[:, :, joint_num])),\n                                      org=(5, 30), fontFace=cv2.FONT_HERSHEY_COMPLEX, fontScale=0.3, color=150)\n                    tmp_img = np.concatenate((tmp_img, tmp), axis=0) \\\n                        if tmp_img is not None else tmp\n\n                # Plot limbs\n                for limb_num in range(len(limbs)):\n                    if np.min(demo_stage_heatmap[:, :, limbs[limb_num][0]]) &gt; -2000 and np.min(\n                            demo_stage_heatmap[:, :, limbs[limb_num][1]]) &gt; -2000:\n                        x1 = joint_coord_set[limbs[limb_num][0], 0]\n                        y1 = joint_coord_set[limbs[limb_num][0], 1]\n                        x2 = joint_coord_set[limbs[limb_num][1], 0]\n                        y2 = joint_coord_set[limbs[limb_num][1], 1]\n                        length = ((x1 - x2) ** 2 + (y1 - y2) ** 2) ** 0.5\n                        if length &lt; 10000 and length &gt; 5:\n                            deg = math.degrees(math.atan2(x1 - x2, y1 - y2))\n                            polygon = cv2.ellipse2Poly((int((y1 + y2) \/ 2), int((x1 + x2) \/ 2)),\n                                                       (int(length \/ 2), 3),\n                                                       int(deg),\n                                                       0, 360, 1)\n                            color_code_num = limb_num \/\/ 4\n                            if PYTHON_VERSION == 3:\n                                limb_color = list(\n                                    map(lambda x: x + 35 * (limb_num % 4), joint_color_code[color_code_num]))\n                            else:\n                                limb_color = map(lambda x: x + 35 * (limb_num % 4), joint_color_code[color_code_num])\n\n                            cv2.fillConvexPoly(test_img, polygon, color=limb_color)\n\n                if tmp_img is not None:\n                    tmp_img = np.lib.pad(tmp_img, ((0, vertical_imgs[0].shape[0] - tmp_img.shape[0]), (0, 0)),\n                                         'constant', constant_values=(0, 0))\n                    vertical_imgs.append(tmp_img)\n\n                # Concat horizontally\n                output_img = None\n                for col in range(len(vertical_imgs)):\n                    output_img = np.concatenate((output_img, vertical_imgs[col]), axis=1) if output_img is not None else \\\n                        vertical_imgs[col]\n\n                output_img = output_img.astype(np.uint8)\n                output_img = cv2.applyColorMap(output_img, cv2.COLORMAP_JET)\n                test_img = cv2.resize(test_img, (300, 300), cv2.INTER_LANCZOS4)\n                cv2.imshow('hm', output_img)\n                cv2.moveWindow('hm', 2000, 200)\n                cv2.imshow('rgb', test_img)\n                cv2.moveWindow('rgb', 2000, 750) \n                if cv2.waitKey(1) == ord('q'): break\n\n\ndef visualize_result(test_img, FLAGS, stage_heatmap_np, kalman_filter_array):\n    t1 = time.time()\n    demo_stage_heatmaps = []\n    if FLAGS.DEMO_TYPE == 'MULTI':\n        for stage in range(len(stage_heatmap_np)):\n            demo_stage_heatmap = stage_heatmap_np[stage][0, :, :, 0:FLAGS.joints].reshape(\n                (FLAGS.hmap_size, FLAGS.hmap_size, FLAGS.joints))\n            demo_stage_heatmap = cv2.resize(demo_stage_heatmap, (test_img.shape[1], test_img.shape[0]))\n            demo_stage_heatmap = np.amax(demo_stage_heatmap, axis=2)\n            demo_stage_heatmap = np.reshape(demo_stage_heatmap, (test_img.shape[1], test_img.shape[0], 1))\n            demo_stage_heatmap = np.repeat(demo_stage_heatmap, 3, axis=2)\n            demo_stage_heatmap *= 255\n            demo_stage_heatmaps.append(demo_stage_heatmap)\n\n        last_heatmap = stage_heatmap_np[len(stage_heatmap_np) - 1][0, :, :, 0:FLAGS.joints].reshape(\n            (FLAGS.hmap_size, FLAGS.hmap_size, FLAGS.joints))\n        last_heatmap = cv2.resize(last_heatmap, (test_img.shape[1], test_img.shape[0]))\n    else:\n        last_heatmap = stage_heatmap_np[len(stage_heatmap_np) - 1][0, :, :, 0:FLAGS.joints].reshape(\n            (FLAGS.hmap_size, FLAGS.hmap_size, FLAGS.joints))\n        last_heatmap = cv2.resize(last_heatmap, (test_img.shape[1], test_img.shape[0]))\n    print('hm resize time %f' % (time.time() - t1))\n\n    t1 = time.time()\n    joint_coord_set = np.zeros((FLAGS.joints, 2))\n\n    # Plot joint colors\n    if kalman_filter_array is not None:\n        for joint_num in range(FLAGS.joints):\n            joint_coord = np.unravel_index(np.argmax(last_heatmap[:, :, joint_num]),\n                                           (test_img.shape[0], test_img.shape[1]))\n            joint_coord = np.array(joint_coord).reshape((2, 1)).astype(np.float32)\n            kalman_filter_array[joint_num].correct(joint_coord)\n            kalman_pred = kalman_filter_array[joint_num].predict()\n            joint_coord_set[joint_num, :] = np.array([kalman_pred[0], kalman_pred[1]]).reshape((2))\n\n            color_code_num = (joint_num \/\/ 4)\n            if joint_num in [0, 4, 8, 12, 16]:\n                if PYTHON_VERSION == 3:\n                    joint_color = list(map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                else:\n                    joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color, thickness=-1)\n            else:\n                if PYTHON_VERSION == 3:\n                    joint_color = list(map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                else:\n                    joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color, thickness=-1)\n    else:\n        for joint_num in range(FLAGS.joints):\n            joint_coord = np.unravel_index(np.argmax(last_heatmap[:, :, joint_num]),\n                                           (test_img.shape[0], test_img.shape[1]))\n            joint_coord_set[joint_num, :] = [joint_coord[0], joint_coord[1]]\n\n            color_code_num = (joint_num \/\/ 4)\n            if joint_num in [0, 4, 8, 12, 16]:\n                if PYTHON_VERSION == 3:\n                    joint_color = list(map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                else:\n                    joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color, thickness=-1)\n            else:\n                if PYTHON_VERSION == 3:\n                    joint_color = list(map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num]))\n                else:\n                    joint_color = map(lambda x: x + 35 * (joint_num % 4), joint_color_code[color_code_num])\n\n                cv2.circle(test_img, center=(joint_coord[1], joint_coord[0]), radius=3, color=joint_color, thickness=-1)\n    print('plot joint time %f' % (time.time() - t1))\n\n    t1 = time.time()\n    # Plot limb colors\n    for limb_num in range(len(limbs)):\n\n        x1 = joint_coord_set[limbs[limb_num][0], 0]\n        y1 = joint_coord_set[limbs[limb_num][0], 1]\n        x2 = joint_coord_set[limbs[limb_num][1], 0]\n        y2 = joint_coord_set[limbs[limb_num][1], 1]\n        length = ((x1 - x2) ** 2 + (y1 - y2) ** 2) ** 0.5\n        if length &lt; 150 and length &gt; 5:\n            deg = math.degrees(math.atan2(x1 - x2, y1 - y2))\n            polygon = cv2.ellipse2Poly((int((y1 + y2) \/ 2), int((x1 + x2) \/ 2)),\n                                       (int(length \/ 2), 3),\n                                       int(deg),\n                                       0, 360, 1)\n            color_code_num = limb_num \/\/ 4\n            if PYTHON_VERSION == 3:\n                limb_color = list(map(lambda x: x + 35 * (limb_num % 4), joint_color_code[color_code_num]))\n            else:\n                limb_color = map(lambda x: x + 35 * (limb_num % 4), joint_color_code[color_code_num])\n\n            cv2.fillConvexPoly(test_img, polygon, color=limb_color)\n    print('plot limb time %f' % (time.time() - t1))\n\n    if FLAGS.DEMO_TYPE == 'MULTI':\n        upper_img = np.concatenate((demo_stage_heatmaps[0], demo_stage_heatmaps[1], demo_stage_heatmaps[2]), axis=1)\n        lower_img = np.concatenate((demo_stage_heatmaps[3], demo_stage_heatmaps[len(stage_heatmap_np) - 1], test_img),\n                                   axis=1)\n        demo_img = np.concatenate((upper_img, lower_img), axis=0)\n        return demo_img\n    else:\n        return test_img\n\n\nif __name__ == '__main__':\n    tf.app.run()\n<\/code><\/pre>\n\n<p>**I tried with reinstalling tensorflow but still getting the same error  currently use tensorflow , and when i run the module im getting the error. but when I run the module, it shows: AttributeError: module AttributeError: module 'tensorflow.app' has no attribute 'flags'. I am really confused why im i getting the error could you please help it in rectifying it\n** <\/p>\n","answers":[{"AnswerId":"58853530","CreationDate":"2019-11-14T09:36:21.963","ParentId":null,"OwnerUserId":"12371762","Title":null,"Body":"<p>Change:<\/p>\n\n<pre><code>tf.app.run\n<\/code><\/pre>\n\n<p>to: <\/p>\n\n<pre><code>tf.compat.v1.app.run\n<\/code><\/pre>\n"}]}
+{"QuestionId":53180173,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><word-embedding>","CreationDate":"2018-11-06T21:20:07.243","AcceptedAnswerId":53183456.0,"OwnerUserId":4872343.0,"Title":"Keras autoencoder with pretrained embeddings returning incorrect number of dimensions","Body":"<p>I have been attempting to replicate a sentence autoencoder loosely based off of <a href=\"https:\/\/github.com\/PacktPublishing\/Deep-Learning-with-Keras\/blob\/master\/Chapter07\/sent-thoughts-rnn.py\" rel=\"nofollow noreferrer\">an example from the Deep Learning with Keras book<\/a>.<\/p>\n\n<p>I recoded the example to use an embedding layer instead of the sentence generator and to use <code>fit<\/code> vs. <code>fit_generator<\/code>. <\/p>\n\n<p>My code is as follows: <\/p>\n\n<pre><code>df_train_text = df['string']\n\nmax_length = 80\nembedding_dim = 300\nlatent_dim = 512\nbatch_size = 64\nnum_epochs = 10\n\n# prepare tokenizer\nt = Tokenizer(filters='')\nt.fit_on_texts(df_train_text)\nword_index = t.word_index\nvocab_size = len(t.word_index) + 1\n\n# integer encode the documents\nencoded_train_text = t.texts_to_matrix(df_train_text)\n\npadded_train_text = pad_sequences(encoded_train_text, maxlen=max_length, padding='post')\n\npadding_train_text = np.asarray(padded_train_text, dtype='int32')\n\nembeddings_index = {}\nf = open('\/Users\/embedding_file.txt')\nfor line in f:\n    values = line.split()\n    word = values[0]\n    coefs = np.asarray(values[1:], dtype='float32')\n    embeddings_index[word] = coefs\nf.close()\n\nprint('Found %s word vectors.' % len(embeddings_index))\n#Found 51328 word vectors.\n\nembedding_matrix = np.zeros((vocab_size, embedding_dim))\nfor word, i in word_index.items():\n    embedding_vector = embeddings_index.get(word)\n    if embedding_vector is not None:\n        # words not found in embedding index will be all-zeros.\n        embedding_matrix[i] = embedding_vector\n\n\nembedding_layer = Embedding(vocab_size,\n                            embedding_dim,\n                            weights=[embedding_matrix],\n                            input_length=max_length,\n                            trainable=False)\ninputs = Input(shape=(max_length,), name=\"input\")\nembedding_layer = embedding_layer(inputs)\nencoder = Bidirectional(LSTM(latent_dim), name=\"encoder_lstm\", merge_mode=\"sum\")(embedding_layer)\ndecoder = RepeatVector(max_length)(encoder)\ndecoder = Bidirectional(LSTM(embedding_dim, name='decoder_lstm', return_sequences=True), merge_mode=\"sum\")(decoder)\nautoencoder = Model(inputs, decoder)\nautoencoder.compile(optimizer=\"adam\", loss=\"mse\")\n\n\nautoencoder.fit(padded_train_text, padded_train_text,\n                epochs=num_epochs, \n                batch_size=batch_size,\n                callbacks=[checkpoint])\n<\/code><\/pre>\n\n<p>I verified that my layer shapes are the same as those in the example, however when I try to fit my autoencoder, I get the following error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected bidirectional_1 to have 3 dimensions, but got array with shape (36320, 80)\n<\/code><\/pre>\n\n<p>A few other things I tried included switching <code>texts_to_matrix<\/code> to <code>texts_to_sequence<\/code> and wrapping\/not wrapping my padded strings <\/p>\n\n<p>I also came across <a href=\"https:\/\/stackoverflow.com\/questions\/44731059\/keras-lstm-autoencoder-with-embedding-layer\">this post<\/a> which seems to indicate that I am going about this the wrong way. Is it possible to fit an autoencoder with the embedding layer as I have coded it? If not, can someone help explain the fundamental difference between what is going on with the provided example and my version? <\/p>\n\n<p>EDIT: I removed the <code>return_sequences=True<\/code> argument in the last layer and got the following error: <code>ValueError: Error when checking target: expected bidirectional_1 to have shape (300,) but got array with shape (80,)<\/code><\/p>\n\n<p>After updating my layer shapes are: <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput (InputLayer)           (None, 80)                0         \n_________________________________________________________________\nembedding_8 (Embedding)      (None, 80, 300)           2440200   \n_________________________________________________________________\nencoder_lstm (Bidirectional) (None, 512)               3330048   \n_________________________________________________________________\nrepeat_vector_8 (RepeatVecto (None, 80, 512)           0         \n_________________________________________________________________\nbidirectional_8 (Bidirection (None, 300)               1951200   \n=================================================================\nTotal params: 7,721,448\nTrainable params: 5,281,248\nNon-trainable params: 2,440,200\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Am I missing a step between the <code>RepeatVector<\/code> layer and the last layer of the model so that I can return a shape of (None, 80, 300) rather than the (None, 300) shape it is currently generating? <\/p>\n","answers":[{"AnswerId":"53183456","CreationDate":"2018-11-07T04:08:31.787","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>Embedding<\/code> layer takes as input a sequence of integers (i.e. word indices) with a shape of <code>(num_words,)<\/code> and gives the corresponding embeddings as output with a shape of <code>(num_words, embd_dim)<\/code>. So after fitting the <code>Tokenizer<\/code> instance on the given texts, you need to use its <code>texts_to_sequences()<\/code> method to transform each text to a sequence of integers:<\/p>\n\n<pre><code>encoded_train_text = t.texts_to_sequences(df_train_text)\n<\/code><\/pre>\n\n<p>Further, since after padding <code>encoded_train_text<\/code> it would have a shape of <code>(num_samples, max_length)<\/code>, the output shape of the network must also have the same shape (i.e. since we are creating an autoencoder) and therefore you need to remove the <code>return_sequences=True<\/code> argument of last layer. Otherwise, it would give us a 3D tensor as output which does not make sense.<\/p>\n\n<p>As a side note, the following line is redundant as <code>padded_train_text<\/code> is already a numpy array (and by the way you have not used <code>padding_train_text<\/code> at all):<\/p>\n\n<pre><code>padding_train_text = np.asarray(padded_train_text, dtype='int32')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53180481,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><recurrent-neural-network>","CreationDate":"2018-11-06T21:41:38.417","AcceptedAnswerId":null,"OwnerUserId":9047195.0,"Title":"Generate a Prediction List","Body":"<p>With Tensorflow, is there a go to way to output the networks predictions?<\/p>\n\n<p>My Outputs have been using a One Hot Representation for 12 classes <\/p>\n\n<pre><code>[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]\n[0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0]\n[0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0]\netc...\n<\/code><\/pre>\n\n<p>When trying to get predictions out of my model for a given input, I ran the following code<\/p>\n\n<pre><code>    prediction=tf.argmax(y,1)\n    best = sess.run([prediction],feed_dict={x: batch_x, y: batch_y,\n                                  seqlen: batch_seqlen})\n    print(\"Prediction: \")\n    print(best)\n<\/code><\/pre>\n\n<p>My output when I ran this code and printed the prediction was: <\/p>\n\n<pre><code>[array([1, 5, 7, 7, 7, 4, 7, 9, 4, 4, 9, 6, 7, 8, 3, 2], dtype=int64)]\n<\/code><\/pre>\n\n<p>my batch size that I input was 16, so it did make sense to have 16 outputs. However, none of these are One Hot representation (Not sure if tensorflow's output is intended to be interperated as index's, so the 1 would actually be some form of onehot<\/p>\n\n<p>Is there a way for each particular X to create a ranked list of predictions, what the model finds the most likely given that X?<\/p>\n\n<p>Does this make sense? <\/p>\n","answers":[{"AnswerId":"53191798","CreationDate":"2018-11-07T14:47:04.393","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>You're taking the <code>tf.argmax<\/code> of 1-hot vectors, so that's why you're seeing indices instead of the 1-hot vectors you expected.<\/p>\n\n<p>To get a sorted list of class predictions you could take your prediction vector and apply <code>values, indices = tf.nn.top_k(prediction)<\/code> the <code>values<\/code> will be your predictions sorted in DESCENDING order and <code>indices<\/code> will be those sorted <code>values<\/code>' indices.<\/p>\n"}]}
+{"QuestionId":53180589,"AnswerCount":2,"Tags":"<python><tensorflow><keras><keras-layer><nvidia-jetson>","CreationDate":"2018-11-06T21:51:14.913","AcceptedAnswerId":null,"OwnerUserId":10615423.0,"Title":"Keras ValueError: Unknown layer:name, when trying to load model to another platform","Body":"<p>I have trained a convolutional neural network using Keras 2.2.4 on Nvidia Quadro board. I have saved the trained model in tow separate files: one file (model.json) that describes the architecture and another file (model.h5) that has all the weights. <\/p>\n\n<p>I want to load the saved model on the Nvidia Jetson TX2 board that runs Keras 2.2.2 and I'm trying to do it as follows:<\/p>\n\n<pre><code># load json and create model\n    json_file = open(prefix+'final_model.json', 'r')\n    loaded_model_json = json_file.read()\n    json_file.close()\n    loaded_model = model_from_json(loaded_model_json)\n# load weights into new model\n    loaded_model.load_weights(prefix+\"model.h5\")\n    model = loaded_model\n<\/code><\/pre>\n\n<p>However, when I tried to load I got the following error:<\/p>\n\n<p>loaded_model = model_from_json(loaded_model_json)<\/p>\n\n<p>File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/engine\/saving.py\", line 368, in model_from_json\n    return deserialize(config, custom_objects=custom_objects)\n  File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/layers\/<strong>init<\/strong>.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/utils\/generic_utils.py\", line 145, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/engine\/sequential.py\", line 292, in from_config\n    custom_objects=custom_objects)\n  File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/layers\/<strong>init<\/strong>.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/nvidia\/.local\/lib\/python3.5\/site-packages\/keras\/utils\/generic_utils.py\", line 165, in deserialize_keras_object\n    ':' + function_name)\n<strong>ValueError: Unknown layer:name<\/strong><\/p>\n\n<p>I've also tried to save the whole model in one file, but got the same error.<\/p>\n\n<p>I have tried the solution from <a href=\"https:\/\/stackoverflow.com\/questions\/50837728\/valueerror-unknown-layer-capsulelayer\">here<\/a> but wasn't able to solve it. <\/p>\n\n<p>Has anyone seen this error before? Any suggestions?<\/p>\n","answers":[{"AnswerId":"53235982","CreationDate":"2018-11-10T04:25:13.283","ParentId":null,"OwnerUserId":"3646527","Title":null,"Body":"<p>I had the same problem yesterday, I just updated keras through conda and everything worked perfectly.<\/p>\n"},{"AnswerId":"53598446","CreationDate":"2018-12-03T17:03:26.840","ParentId":null,"OwnerUserId":"6835092","Title":null,"Body":"<p>just update the keras package to recent one using following command<\/p>\n\n<p><code>conda update keras<\/code><\/p>\n"}]}
+{"QuestionId":53181434,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><anaconda>","CreationDate":"2018-11-06T23:06:13.373","AcceptedAnswerId":null,"OwnerUserId":10605219.0,"Title":"Tensorflow-gpu is refusing to run with a strange error","Body":"<p>I have just purchased an RX 2070, installed microsoft visual studio, then CUDA, then cuDNN. After this I installed anaconda and issues commands<\/p>\n\n<p>pip install tensorflow-gpu<\/p>\n\n<p>pip install keras<\/p>\n\n<p>With in theory a functioning version of tensorflow-gpu I attempted to run my neural network which was working under the cpu version and the following error appeared.<\/p>\n\n<p>The same error appears under any attempt to import tensorflow<\/p>\n\n<blockquote>\n  <p>File \"\", line 1, in\n  runfile('C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC\n  example\/Neural Network Worked Example.py',\n  wdir='C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC example')<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\",\n  line 705, in runfile execfile(filename, namespace)<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\",\n  line 102, in execfile exec(compile(f.read(), filename, 'exec'),\n  namespace)<\/p>\n  \n  <p>File \"C:\/Users\/g\/Desktop\/Masters-Project\/Gaussian ROC example\/Neural\n  Network Worked Example.py\", line 7, in from keras.models import\n  Sequential as sequential<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras__init__.py\",\n  line 3, in from . import utils<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\utils__init__.py\",\n  line 6, in from . import conv_utils<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\utils\\conv_utils.py\",\n  line 9, in from .. import backend as K<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend__init__.py\",\n  line 89, in from .tensorflow_backend import *<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\",\n  line 5, in import tensorflow as tf<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow__init__.py\",\n  line 22, in from tensorflow.python import pywrap_tensorflow # pylint:\n  disable=unused-import<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py\",\n  line 49, in from tensorflow.python import pywrap_tensorflow<\/p>\n  \n  <p>File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 74, in raise ImportError(msg)<\/p>\n  \n  <p>ImportError: Traceback (most recent call last): File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 58, in from tensorflow.python.pywrap_tensorflow_internal import *\n  File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 28, in _pywrap_tensorflow_internal = swig_import_helper() File\n  \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper _mod =\n  imp.load_module('_pywrap_tensorflow_internal', fp, pathname,\n  description) File \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\", line 243, in\n  load_module return load_dynamic(name, filename, file) File\n  \"C:\\ProgramData\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n  return _load(spec) ImportError: DLL load failed: The specified module\n  could not be found.<\/p>\n  \n  <p>Failed to load the native TensorFlow runtime.<\/p>\n  \n  <p>See\n  <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n  \n  <p>for some common reasons and solutions. Include the entire stack trace\n  above this error message when asking for help.<\/p>\n<\/blockquote>\n\n<p>I've been unable to resolve this, I can confirm to the best of my ability the other packages are correct and I'm absolutely stumped as to what's wrong.<\/p>\n","answers":[{"AnswerId":"53184714","CreationDate":"2018-11-07T06:49:18.407","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>make sure you install the correct Cuda and Cudnn versions. You can check the required versions in <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">here<\/a>, if the error still persists try  reinstalling tensorflow. <\/p>\n"}]}
+{"QuestionId":53181495,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-06T23:14:36.853","AcceptedAnswerId":53198785.0,"OwnerUserId":6254131.0,"Title":"How does tf.layers create variables?","Body":"<p>Everytime we call <code>tf.layers.conv2d<\/code>, Tensorflow will automatically creates new kernel, bias and new name for this tensor by adding <code>_1<\/code>, <code>_2<\/code>, like:<\/p>\n\n<pre><code>&lt;tf.Variable 'conv2d\/kernel:0' shape=(3, 3, 10, 10) dtype=float32_ref&gt;, \n&lt;tf.Variable 'conv2d\/bias:0' shape=(10,) dtype=float32_ref&gt;, \n&lt;tf.Variable 'conv2d_1\/kernel:0' shape=(3, 3, 10, 10) dtype=float32_ref&gt;,  \n&lt;tf.Variable 'conv2d_1\/bias:0' shape=(10,) dtype=float32_ref&gt;\nconv2d\/BiasAdd:0\nconv2d_1\/BiasAdd:0\n<\/code><\/pre>\n\n<p>If I would like to define a similar layer with variables inside, for example,<\/p>\n\n<pre><code>def some_layer(input):\n    gamma = tf.get_variable(name='gamma', shape=[10], dtype=tf.float32,\n                        initializer=tf.constant_initializer(1.0))\n\nx = some_layer(input)\ny = some_layer(input)\n<\/code><\/pre>\n\n<p>It will cause <code>ValueError: Variable gamma already exists<\/code>. I know there are methods to give each variable a name scope or variable scope, but I am wondering is there any method to automatically create new variables <code>gamma_1:0<\/code>, <code>gamma_2:0<\/code> like <code>conv2d\/kernel:0<\/code>, <code>conv2d_1\/kernel:0<\/code> after calling <code>tf.layers.conv2d<\/code>.<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.11\/tensorflow\/python\/layers\/convolutional.py\" rel=\"nofollow noreferrer\">definition<\/a> of <code>tf.layers.conv2d<\/code> gives me no hint.<\/p>\n","answers":[{"AnswerId":"53198785","CreationDate":"2018-11-07T22:22:20.447","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<p>The Tensorflow tutorial <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_layers\" rel=\"nofollow noreferrer\">custom layers<\/a> gives the answer. If we need to build a custom layer, we will need to extend the class <code>tf.keras.layers.Layer<\/code>.<\/p>\n"}]}
+{"QuestionId":53181837,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><multilabel-classification>","CreationDate":"2018-11-06T23:55:03.860","AcceptedAnswerId":53183302.0,"OwnerUserId":7933418.0,"Title":"Convolutional Neural Network : Weights and Bias initialization","Body":"<p>I'm building convolutional neural network for classification of the data into different categories\nThe input data is of shape : 30000, 6, 15, 1\nthe data has 30000 samples, 15 predictors and 6 possible categories.<\/p>\n\n<p>My model is defined as follows.<\/p>\n\n<pre><code>x = tf.placeholder(\"float\", [None, 6,15,1])\ny = tf.placeholder(\"float\", [None, n_classes])\n\n#Define Weights\nweights = {\n    'wc1': tf.get_variable('W0', shape=(3,3,1,8), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc2': tf.get_variable('W1', shape=(3,3,32,12), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc3': tf.get_variable('W2', shape=(3,3,64,16), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc4': tf.get_variable('W3', shape=(3,3,64,20), initializer=tf.contrib.layers.xavier_initializer()),\n    'wd1': tf.get_variable('W4', shape=(4*4*15,15), initializer=tf.contrib.layers.xavier_initializer()), \n    'out': tf.get_variable('W6', shape=(15,n_classes), initializer=tf.contrib.layers.xavier_initializer()), \n}\n\nbiases = {\n    'bc1': tf.get_variable('B0', shape=(8), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc2': tf.get_variable('B1', shape=(12), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc3': tf.get_variable('B2', shape=(16), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc4': tf.get_variable('B3', shape=(20), initializer=tf.contrib.layers.xavier_initializer()),\n    'bd1': tf.get_variable('B4', shape=(15), initializer=tf.contrib.layers.xavier_initializer()),\n    'out': tf.get_variable('B5', shape=(6), initializer=tf.contrib.layers.xavier_initializer()),\n}\n\n#Define convolutional layer\ndef conv2d(x, W, b, strides=1, reuse=True):\n    # Conv2D wrapper, with bias and relu activation\n    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')\n    x = tf.nn.bias_add(x, b)\n    return tf.nn.relu(x)\n\n#Define Maxpool layer\ndef maxpool2d(x, k=2):\n    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],padding='SAME')\n\n#Define a convolutional neural network function\ndef conv_net(x, weights, biases):  \n    conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n    conv1 = maxpool2d(conv1, k=2)\n\n    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n    conv2 = maxpool2d(conv2, k=2)\n\n    conv3 = conv2d(conv2, weights['wc3'], biases['bc3'])\n    conv3 = maxpool2d(conv3, k=2)\n\n    conv4 = conv2d(conv3, weights['wc4'], biases['bc4'])\n    conv4 = maxpool2d(conv4, k=2)\n\n\n    # Fully connected layer\n    # Reshape conv2 output to fit fully connected layer input\n    fc1 = tf.reshape(conv4, [-1, weights['wd1'].get_shape().as_list()[0]])\n    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n    fc1 = tf.nn.relu(fc1)\n    # Output, class prediction \n    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n    return out\n<\/code><\/pre>\n\n<p>I'm getting the error : <code>ValueError: Dimensions must be equal, but are 8 and 32 for 'Conv2D_1' (op: 'Conv2D') with input shapes: [?,8,3,8], [3,3,32,12].<\/code> <\/p>\n\n<p>When I execute : <\/p>\n\n<pre><code>pred = conv_net(x, weights, biases)\n<\/code><\/pre>\n\n<p>I went through multiple conv2D models but most of them are for image classification and I might be missing something here that I'm not able to identify. Please help.<\/p>\n","answers":[{"AnswerId":"53183302","CreationDate":"2018-11-07T03:45:15.173","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The number of input channels for the weights <code>wc2<\/code>, <code>wc3<\/code>, and <code>wc4<\/code> need to be the same as the number of output channels for the previous layers.  Keeping the number of output channels the way you have it, they would be changed to:<\/p>\n\n<pre><code>    'wc1': tf.get_variable('W0', shape=(3,3,1,8), initializer=tf.contrib.layers.xavier_initializer()),\n    'wc2': tf.get_variable('W1', shape=(3,3,8,12), initializer=tf.contrib.layers.xavier_initializer()),\n    'wc3': tf.get_variable('W2', shape=(3,3,12,16), initializer=tf.contrib.layers.xavier_initializer()),\n    'wc4': tf.get_variable('W3', shape=(3,3,16,20), initializer=tf.contrib.layers.xavier_initializer()),\n<\/code><\/pre>\n"}]}
+{"QuestionId":53182177,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><pytorch><onnx>","CreationDate":"2018-11-07T00:42:39.840","AcceptedAnswerId":null,"OwnerUserId":10615993.0,"Title":"How do you convert a .onnx to tflite?","Body":"<p>I've exported my model to ONNX via:<\/p>\n\n<pre><code># Export the model\ntorch_out = torch.onnx._export(learn.model,             # model being run\n                           x,                       # model input (or a tuple for multiple inputs)\n                          EXPORT_PATH + \"mnist.onnx\", # where to save the model (can be a file or file-like object)\n                           export_params=True)      # store the trained parameter weights inside the model file\n<\/code><\/pre>\n\n<p>And now I am trying to convert the model to a Tensorflow Lite file so that I can do inference on Android. Unfortunately, PyTorch\/Caffe2 support is fairly lacking or too complex for Android but Tensorflow appears much simpler.<\/p>\n\n<p>The documentation for ONNX to Tflite is pretty light on this.<\/p>\n\n<p>I've tried exporting to a Tensorflow GraphDef proto via:<\/p>\n\n<p><code>tf_rep.export_graph(EXPORT_PATH + 'mnist-test\/mnist-tf-export.pb')<\/code><\/p>\n\n<p>And then running <code>toco<\/code>:<\/p>\n\n<pre><code>toco \\\n--graph_def_file=mnist-tf-export.pb \\\n--input_format=TENSORFLOW_GRAPHDEF \\\n--output_format=TFLITE \\\n--inference_type=FLOAT \\\n--input_type=FLOAT \\\n--input_arrays=0 \\\n--output_arrays=add_10 \\\n--input_shapes=1,3,28,28 \\\n--output_file=mnist.tflite`\n<\/code><\/pre>\n\n<p>When I do though I get the following error:<\/p>\n\n<pre><code>File \"anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/lite\/python\/convert.py\", line 172, in toco_convert_protos\n    \"TOCO failed. See console for info.\\n%s\\n%s\\n\" % (stdout, stderr))\ntensorflow.lite.python.convert.ConverterError: TOCO failed. See console for info.\n2018-11-06 16:28:33.864889: I tensorflow\/lite\/toco\/import_tensorflow.cc:1268] Converting unsupported operation: PyFunc\n2018-11-06 16:28:33.874130: F tensorflow\/lite\/toco\/import_tensorflow.cc:114] Check failed: attr.value_case() == AttrValue::kType (1 vs. 6)\n<\/code><\/pre>\n\n<p>Further, even when I run the command I don't know what to specify for the input_arrays or output_arrays since the model was originally built in PyTorch.<\/p>\n\n<p>Has anyone successfully converted their ONNX model to TFlite?<\/p>\n\n<p>Here's the ONNX file I'm trying to convert: <a href=\"https:\/\/drive.google.com\/file\/d\/1sM4RpeBVqPNw1WeCROpKLdzbSJPWSK79\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/1sM4RpeBVqPNw1WeCROpKLdzbSJPWSK79\/view?usp=sharing<\/a><\/p>\n\n<p>Extra info<\/p>\n\n<ul>\n<li>Python 3.6.6 :: Anaconda custom (64-bit)<\/li>\n<li>onnx.<strong>version<\/strong> = '1.3.0'<\/li>\n<li>tf.<strong>version<\/strong> = '1.13.0-dev20181106'<\/li>\n<li>torch.<strong>version<\/strong> = '1.0.0.dev20181029'<\/li>\n<\/ul>\n","answers":[{"AnswerId":"58576060","CreationDate":"2019-10-27T01:50:02.123","ParentId":null,"OwnerUserId":"7001213","Title":null,"Body":"<p>I think your onnx file that you have given is corrupted i don't know what is the issue but it is not doing any inference on onnx runtime<br>\nThe best practise to convert model from <strong>Pytorch<\/strong> to <strong>Onnx<\/strong> is that you should add following paremeters to specify names of input and ouput layer of your model in <strong>torch.onnx.export()<\/strong> function<br>\nSo in your case:<\/p>\n\n<pre><code>    # Export the model\ntorch_out = torch.onnx._export(learn.model,             # model being run\n                               x,                       # model input (or a tuple for multiple inputs)\n                              EXPORT_PATH + \"mnist.onnx\", # where to save the model (can be a file or file-like object)\n                              export_params=True,       # store the trained parameter weights inside the model file\n                               input_names=['main_input'],  # specify the name of input layer in onnx model\n                               output_names=['main_output'])     # specify the name of input layer in onnx model\n<\/code><\/pre>\n\n<p>Now export this model to tensorFlow protobuf FreezeGraph using onnx-tf<br>\nPlease note that this method is only working when tensorflow_version &lt; 2  <\/p>\n\n<pre><code>from onnx_tf.backend import prepare\n\n# load the model saved in onnx format\nmodel_onnx = onnx.load('\/content\/mnist.onnx')\n\n# prepare model for exporting to tensorFlow using tensorFlow backend\ntf_rep = prepare(model_onnx)\nprint(tf_rep.run(dummy_input)) # run sample inference of your model\nprint(tf_rep.inputs) # Input nodes to the model\nprint('-----')\nprint(tf_rep.outputs) # Output nodes from the model\nprint('-----')\nprint(tf_rep.tensor_dict) # All nodes in the model\n\n# export tensorFlow backend to tensorflow tf file\ntf_rep.export_graph('mnist.pb')\n<\/code><\/pre>\n\n<p>The next thing you need to do is to get detail about ouput layer of your pb model  <\/p>\n\n<pre><code>print(tf_rep.tensor_dict) # All nodes in the model\n<\/code><\/pre>\n\n<p>What output it prints find the node by the name of your output layer specified in first function for example in my model is  <\/p>\n\n<pre><code>'main_output': &lt;tf.Tensor 'add_10:0' shape=(1, 2) dtype=float32  \n<\/code><\/pre>\n\n<p>Now remember the operator of the tensor which is add_10  <\/p>\n\n<p>The best way to convert model from protobuf freezeGraph is to use official tensorflow lite convert <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/get_started#2_convert_the_model_format\" rel=\"nofollow noreferrer\">documentations<\/a> Now to convert this model from .pb file to tflite model use this code <\/p>\n\n<pre><code>import tensorflow as tf\n# make a converter object from the saved tensorflow file\nconverter = tf.lite.TFLiteConverter.from_frozen_graph('\/content\/mnist.pb',\n                                                      input_arrays=['main_input'], # input arrays \n                                                      output_arrays=['add_10']  # output arrays as told in upper in my model case it si add_10\n)\n\n# tell converter which type of optimization techniques to use\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\n# to view the best option for optimization read documentation of tflite about optimization go to this link https:\/\/www.tensorflow.org\/lite\/guide\/get_started#4_optimize_your_model_optional\n\n# convert the model \ntf_lite_model = converter.convert()\n# save the converted model \nopen('eye_state_model_tensorFlowopt.tflite', 'wb').write(tf_lite_model)\n<\/code><\/pre>\n\n<p>To choose which option is best for optimization of your model use see this official guide <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/get_started#4_optimize_your_model_optional\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/guide\/get_started#4_optimize_your_model_optional<\/a><\/p>\n\n<p>Please review my code<\/p>\n"}]}
+{"QuestionId":53182285,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-11-07T00:56:50.383","AcceptedAnswerId":null,"OwnerUserId":634042.0,"Title":"Use `tf.data` to efficiently count the number of Examples in TFRecords?","Body":"<p>I'm writing some simple code to count the number of Examples in TFRecord files. Previously I was using something like this:<\/p>\n\n<pre><code>def _count_example(path):\n    return sum(1 for _ in tf.python_io.tf_record_iterator(path))\n\ndef _count_total_examples(tfrecord_glob_pattern):\n    example_count = 0\n    paths = glob.glob(tfrecord_glob_pattern)\n\n    with futures.ProcessPoolExecutor(max_workers=os.cpu_count()) as executor:\n        results = executor.map(_count_example, paths, chunksize=max(len(paths) \/\/ os.cpu_count(), 1))\n        for result in results:\n            example_count += result\n\n    return example_count\n<\/code><\/pre>\n\n<p>I've copied my test TFRecord input to a RAM disk to make sure we are not bottlenecked by I\/O. My test TFRecord files have 635,424 image files have 128 shards. The above code finishes in 13 seconds.<\/p>\n\n<p>I'm trying to change the same code to using <code>tf.data<\/code> and see if I can get similar results. So I'm doing something like:<\/p>\n\n<pre><code>def _count_total_examples(tfrecord_glob_pattern):\n    dataset = tf.data.Dataset.list_files(tfrecord_glob_pattern, shuffle=False)\n    dataset = tf.data.TFRecordDataset(dataset, num_parallel_reads=os.cpu_count())\n    iterator = dataset.make_one_shot_iterator()\n    next_element = iterator.get_next()\n    with tf.Session() as sess:\n        example_count = 0\n        for _ in trange(635424):\n            sess.run(next_element)\n            example_count += 1\n    return example_count\n<\/code><\/pre>\n\n<p>This version takes more than 2 minutes to finish. I guess this is because I'm calling <code>sess.run()<\/code> too many times and this has a non-trivial overhead. To verify this claim, I've changed the above code to:<\/p>\n\n<pre><code>def _count_total_examples(tfrecord_glob_pattern):\n    dataset = tf.data.Dataset.list_files(tfrecord_glob_pattern, shuffle=False)\n    dataset = tf.data.TFRecordDataset(dataset, num_parallel_reads=os.cpu_count())\n    dataset = dataset.map(lambda _: 1, num_parallel_calls=os.cpu_count())\n    dataset = dataset.batch(4096)\n    iterator = dataset.make_one_shot_iterator()\n    next_element = iterator.get_next()\n    with tf.Session() as sess:\n        example_count = 0\n        for _ in trange(635424 \/\/ 4096):\n            sess.run(next_element)\n            example_count += 1\n    return example_count\n<\/code><\/pre>\n\n<p>This time the code finishes within 30 seconds - which is still not ideal but I guess better. Basically, based on my above observations, I have two questions:<\/p>\n\n<p>1) What is the efficient approach to address this issue? Or should I just stay using <code>tf.python_io.tf_record_iterator<\/code>?<\/p>\n\n<p>2) Is there a way to \"loop the whole dataset and then return the final result\" in the TensorFlow graph? I've tried something like <code>tf.contrib.data.Reducer<\/code> but it seems to be still quite slow (takes around one minute to finish):<\/p>\n\n<pre><code>def _count_total_examples(tfrecord_glob_pattern):\n    dataset = tf.data.Dataset.list_files(tfrecord_glob_pattern, shuffle=False)\n    dataset = tf.data.TFRecordDataset(dataset, num_parallel_reads=os.cpu_count() * 2)\n    reducer = tf.contrib.data.Reducer(\n        init_func=lambda _: 0,\n        reduce_func=lambda curr, _: curr + 1,\n        finalize_func=lambda curr: curr\n    )\n    dataset = tf.contrib.data.reduce_dataset(dataset, reducer)\n    with tf.Session() as sess:\n        result = sess.run(dataset)\n        example_count = result\n    return example_count\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53182436,"AnswerCount":1,"Tags":"<tensorflow><data-visualization><tensorboard><amazon-sagemaker>","CreationDate":"2018-11-07T01:18:04.340","AcceptedAnswerId":53453394.0,"OwnerUserId":2089899.0,"Title":"SageMaker: visualizing training statistics","Body":"<p>If I send a TensorFlow training job to a SageMaker instance, what is the typical way to view training progress? Can I access TensorBoard for this launched EC2 instance? Is there some other alternative? What I'm looking for specifically are things like graphs of current training epoch and mAP.<\/p>\n","answers":[{"AnswerId":"53453394","CreationDate":"2018-11-23T22:07:36.183","ParentId":null,"OwnerUserId":"10697340","Title":null,"Body":"<p>you can now specify metrics(metricName, Regex) that you want to track by using AWS management console or Amazon SageMaker Python SDK APIs. After the model training starts, Amazon SageMaker will automatically monitor and stream the specified metrics in real time to the Amazon CloudWatch console for visualizing time-series curves. <\/p>\n\n<p>Ref: \n<a href=\"https:\/\/docs.aws.amazon.com\/sagemaker\/latest\/dg\/API_MetricDefinition.html\" rel=\"nofollow noreferrer\">https:\/\/docs.aws.amazon.com\/sagemaker\/latest\/dg\/API_MetricDefinition.html<\/a><\/p>\n"}]}
+{"QuestionId":53182716,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-07T02:06:17.150","AcceptedAnswerId":null,"OwnerUserId":7239131.0,"Title":"reuse kernel weight in keras with tensorflow backend","Body":"<p>I want to reuse kernel weight in keras with tensorflow backend.<\/p>\n\n<p>tf.version = 1.11.0<\/p>\n\n<p>My code is as following :<\/p>\n\n<pre><code>from tensorflow.python.keras.layers import  Input\nfrom tensorflow.python.keras import regularizers\ndef infer3(data_input, Reuse):\n    with tf.variable_scope('Network', reuse=Reuse):\n        inputs = Input(tensor = data_input)\n        network = tf.keras.layers.Conv2D(kernel_size=3, strides=2, filters=64, padding='same',kernel_regularizer=regularizers.l2(1),\n                    activation='linear', kernel_initializer=\"glorot_normal\", name='conv1', bias_initializer='zeros')(inputs)\n\n    return network\n\ntf.reset_default_graph() \ninput_tensor = tf.placeholder(tf.float32,shape=[BATCH_SIZE,img_H,img_W,1])\noutput_tensor = tf.placeholder(tf.float32,shape=[BATCH_SIZE,img_H,img_W,1]) \nin_training_mode = tf.placeholder(tf.bool)     \n\n\nnetwork = infer3(input_tensor,False)\nnetwork_test = infer3(input_tensor,True)\n<\/code><\/pre>\n\n<p>When I type \" tf.trainable_variables() \"<\/p>\n\n<p>It show message:<\/p>\n\n<pre><code>[&lt;tf.Variable 'Network\/conv1\/kernel:0' shape=(3, 3, 1, 64) dtype=float32&gt;,\n &lt;tf.Variable 'Network\/conv1\/bias:0' shape=(64,) dtype=float32&gt;,\n &lt;tf.Variable 'Network_1\/conv1\/kernel:0' shape=(3, 3, 1, 64) dtype=float32&gt;,\n &lt;tf.Variable 'Network_1\/conv1\/bias:0' shape=(64,) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>How can I reuse kernel weight?<\/p>\n","answers":[{"AnswerId":"53185368","CreationDate":"2018-11-07T07:51:32.257","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>As stated <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Layer#get_weights\" rel=\"nofollow noreferrer\">here<\/a>, you can use <code>get_weights()<\/code> method on an instance of layer to get its weights. Then you can assign those weights to a variable and initialize another layer of the same type and shape with those weights.<\/p>\n"}]}
+{"QuestionId":53182773,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras><lstm><rnn>","CreationDate":"2018-11-07T02:17:07.470","AcceptedAnswerId":53183497.0,"OwnerUserId":8879559.0,"Title":"Pit in LSTM programming by python","Body":"<p>As we all Know, if we want to train a LSTM network, we must reshape the train dataset by the function <code>numpy.reshape()<\/code>, and reshaping result is like <code>[samples,time_steps,features]<\/code>. However, the new shape is influenced by the original one. I have seen some blogs teaching LSTM programming taking 1 as time_steps, and if time_steps is another number, samples will change relevently. My question is that <em>does the <code>samples<\/code>equal to <code>batch_size<\/code>?<\/em><\/p>\n\n<p><code>X = X.reshape(X.shape[0], 1, X.shape[1])<\/code><\/p>\n","answers":[{"AnswerId":"53183497","CreationDate":"2018-11-07T04:17:50.880","ParentId":null,"OwnerUserId":"10161091","Title":null,"Body":"<p>No, samples is different from batch_size. samples is the total number of samples you would have. batch_size would be the size of each batch or the number of samples per each batch used in training, like by .fit. <\/p>\n\n<p>For example, if samples=128 and batch_size=16, then your data would be divided into 8 batches with each having 16 samples inside during .fit call. <\/p>\n\n<p>As another note, time_steps is the total time steps or observations within each sample. It does not make much sense to have it as 1 with LSTM as the main advantage of RNN's in general is to learn the temporal patterns. With time_step=1, there won't be any history to leverage. Here as an example that might help:<\/p>\n\n<p>Assume that your job is to determine if someone is active or not every hour by looking at they breathing rate and heart rate provided every minute, i.e. 2 features measured at 60 samples per hour. (This is just an example, use accelerometers if you really wanted to do this :)) Let's say you have 128 hours of labeled data. Then your input data would be of shape (128, 60, 2) and your output would be of shape (128, 1).<\/p>\n\n<p>Here, you have 128 samples, 60 time steps or observations per sample, and two features.<\/p>\n\n<p>Next you split the data into train, validation, and testing according to the samples. For example, your train, validation, and test data would be of shapes (96, 60, 2), (16, 60, 2), and (16, 60, 2), respectively. <\/p>\n\n<p>If you use batch_size=16, your training, validation, and test data would have 6, 1, and 1 batches, respectively.<\/p>\n"},{"AnswerId":"53186754","CreationDate":"2018-11-07T09:35:33.717","ParentId":null,"OwnerUserId":"6844910","Title":null,"Body":"<p>No. Samples are not equal to batch size. Samples means the number of rows in your data-set. Your training data-set is divided into number of batches and pass it to the network to train. <\/p>\n\n<p>In simple words,<\/p>\n\n<p>Imagine your data-set has 30 samples, and you define your batch_size as 3. \nThat means the 30 samples divided into 10 batches(30 divided by you defined batch_size = 10). When you train you model, at a time only 3 rows of data will be be pushed to the neural network and then next 3 rows will be push to the neural network. Like wise whole data-set will push to the neural network.<\/p>\n\n<p><em>Samples\/Batch_size = Number of batches<\/em><\/p>\n\n<p>Remember that batch_size and number of batches are two different things.<\/p>\n"}]}
+{"QuestionId":53183092,"AnswerCount":0,"Tags":"<python><tensorflow><cudnn>","CreationDate":"2018-11-07T03:09:35.837","AcceptedAnswerId":null,"OwnerUserId":10003211.0,"Title":"How to solve libcudnn.so.6: cannot open shared object file without sudo","Body":"<p>When I use import tensorflow some error appeared.\nImportError: libcudnn.so.6: cannot open shared object file: No such file or directory.\n<a href=\"https:\/\/i.stack.imgur.com\/OzdJm.png\" rel=\"nofollow noreferrer\">enter image description here<\/a>\nBecause I used it on the server, so there is no sudo permission.\nHow can I solve it?<\/p>\n","answers":[]}
+{"QuestionId":53183424,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T04:05:14.833","AcceptedAnswerId":53186321.0,"OwnerUserId":6319495.0,"Title":"Training ssd inception_v3 using pretrained model from slim","Body":"<p>I want to train ssd inception_v3 model using object detection API with pretrained model from SLIM (<a href=\"https:\/\/:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\" rel=\"nofollow noreferrer\">link<\/a>)\nI try to train object detection ssd inception v3 model using config:<\/p>\n\n<pre><code>model {\n  ssd {\n    num_classes: 1\n    image_resizer {\n      fixed_shape_resizer {\n        height: 240\n        width: 320\n      }\n    }\n    feature_extractor {\n      type: \"ssd_inception_v3\"\n      depth_multiplier: 1.0\n      min_depth: 16\n      conv_hyperparams {\n        regularizer {\n          l2_regularizer {\n            weight: 3.99999989895e-05\n          }\n        }\n        initializer {\n          truncated_normal_initializer {\n            mean: 0.0\n            stddev: 0.0299999993294\n          }\n        }\n        activation: RELU_6\n        batch_norm {\n          decay: 0.999700009823\n          center: true\n          scale: true\n          epsilon: 0.0010000000475\n          train: true\n        }\n      }\n      override_base_feature_extractor_hyperparams: true\n    }\n...\n<\/code><\/pre>\n\n<p>I stopped procces after creating files model.ckpt-0.*, loaded and printed all tensors' names.<\/p>\n\n<p>After that I loaded pretrained model from <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim<\/a> using <\/p>\n\n<pre><code>reader = pywrap_tensorflow.NewCheckpointReader(os.path.join(folder, 'model.ckpt'))\nvar_to_shape_map = reader.get_variable_to_shape_map()\n<\/code><\/pre>\n\n<p>When I compared output I didn't see a lot layers for ssd incpetion v3 model. For example: <\/p>\n\n<blockquote>\n  <p>InceptionV3\/AuxLogits\/Conv2d_2a_5x5\/weights\n  InceptionV3\/Mixed_7c\/Branch_3\/Conv2d_0b_1x1\/weight<\/p>\n<\/blockquote>\n\n<p>In model from ssd_inception_v3 I saw mixed layers before 5c.<\/p>\n\n<p>What are the differences of Feature Extractor in SSD_inception and SLIM models? In general, is it possible to load weights from SLIM for the classifier in the Object detection API for detection.<\/p>\n","answers":[{"AnswerId":"53186321","CreationDate":"2018-11-07T09:06:24.053","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>You can see what happens in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/models\/ssd_inception_v3_feature_extractor.py\" rel=\"nofollow noreferrer\">ssd_inception_v3_feature_extractor<\/a>.\nIt uses the outputs of 'Mixed_5d', 'Mixed_6e', 'Mixed_7c' of InceptionV3 from <code>inception_v3.inception_v3_base<\/code> (Note the <code>_base<\/code>) and creating 3 additional feature maps with 512, 256, 128 number of channels (this happens in <code>feature_map_generators.multi_resolution_feature_maps<\/code> by <code>feature_map_layout<\/code>).\nLoading the weights of the classifier for the detection model can be done by configuration:<\/p>\n\n<pre><code>train_config{\n    ...\n    fine_tune_checkpoint: &lt;path_to_inception_checkpoint&gt;\n    fine_tune_checkpoint_type: \"classification\"\n}\n<\/code><\/pre>\n\n<p>Of course that the checkpoint has to match the model you're using, e.g. <code>ssd_inception_v3<\/code>.<\/p>\n"}]}
+{"QuestionId":53183865,"AnswerCount":9,"Tags":"<python><tensorflow><machine-learning><keras><google-colaboratory>","CreationDate":"2018-11-07T05:09:17.157","AcceptedAnswerId":53656253.0,"OwnerUserId":8614829.0,"Title":"Unknown initializer: GlorotUniform when loading Keras model","Body":"<p>I trained my CNN (VGG) through google colab and generated .h5 file. Now problem is, I can predict my output successfully through google colab but when i download that .h5 trained model file and try to predict output on my laptop, I am getting error when loading the model.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport h5py\n\n# Initialization\n\nloaded_model = keras.models.load_model('.\/train_personCount_model.h5')\n<\/code><\/pre>\n\n<p>And the error:<\/p>\n\n<pre><code>ValueError: Unknown initializer: GlorotUniform\n<\/code><\/pre>\n","answers":[{"AnswerId":"54770296","CreationDate":"2019-02-19T15:56:50.730","ParentId":null,"OwnerUserId":"6650382","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/issues\/798#issuecomment-452442569\" rel=\"nofollow noreferrer\">Something<\/a> that helped me which wasn't in any of the answers:<\/p>\n\n<p><code>custom_objects={'GlorotUniform': glorot_uniform()}<\/code><\/p>\n"},{"AnswerId":"56691652","CreationDate":"2019-06-20T18:08:21.020","ParentId":null,"OwnerUserId":"9389666","Title":null,"Body":"<p><strong>Changing<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import load_model\n<\/code><\/pre>\n\n<p><strong>to<\/strong> <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.models import load_model\n<\/code><\/pre>\n\n<p><strong>solved my problem!<\/strong><\/p>\n\n<p>To eliminate errors, import <strong>all<\/strong> things directly from Keras or TensorFlow. Mixing both of them in same project may result in problems.<\/p>\n"},{"AnswerId":"55406788","CreationDate":"2019-03-28T21:00:46.300","ParentId":null,"OwnerUserId":"7257643","Title":null,"Body":"<p>I had the same problem with a model built with tensorflow 1.11.0 (using tensorflow.python.keras.models.save_model) and loaded with tensoflow 1.11.0 (using tensorflow.python.keras.models.load_model).<\/p>\n\n<p>I solved it by upgrading everything to tensorflow 1.13.1, after building the model again with the new version, I could load it without this error.<\/p>\n"},{"AnswerId":"53369637","CreationDate":"2018-11-19T06:51:53.117","ParentId":null,"OwnerUserId":"1918190","Title":null,"Body":"<p>I had a same problem and was fixed this way. just don't save the optimizer with the model!\njust change the save line like this:<\/p>\n\n<pre><code>the_model.save(file_path,True\/False,False)\n<\/code><\/pre>\n\n<p>Second parameter tells Keras to overwrite the model if the file existed or not and the 3rd one tells it not to save the optimizer with the model.<\/p>\n\n<hr>\n\n<p><strong>Edit<\/strong>:\nI ran over the problem again on another system today and this did not helped me this time. so i saved the model conf as json and weights as h5 and used them to rebuild the model in another machine. you can do it like this.\nsave like this:<\/p>\n\n<pre><code>json = model.to_json()\n# Save the json on a file\nmodel.save_weights(weights_filepath,save_format=\"h5\")\n<\/code><\/pre>\n\n<p>rebuild the model like this:<\/p>\n\n<pre><code># load the json file\n# here i use json as loaded content of json file\nmodel = keras.models.model_from_json(json)\nmodel.load_weights(weights_file_path)\n<\/code><\/pre>\n"},{"AnswerId":"53656253","CreationDate":"2018-12-06T16:56:17.903","ParentId":null,"OwnerUserId":"10755953","Title":null,"Body":"<p>I ran into the same issue. After changing:<\/p>\n\n<p><code>from tensorflow import keras<\/code><\/p>\n\n<p>to:<\/p>\n\n<p><code>import keras<\/code><\/p>\n\n<p>life is once again worth living.<\/p>\n"},{"AnswerId":"54879211","CreationDate":"2019-02-26T05:54:34.103","ParentId":null,"OwnerUserId":"11117505","Title":null,"Body":"<p>I fixed the problem:<\/p>\n\n<p>Before:<\/p>\n\n<p><code>from keras.models import load_model\nclassifierLoad = load_model('model\/modeltest.h5')<\/code><\/p>\n\n<p>Works for me<\/p>\n\n<p><code>import tensorflow as tf \nclassifierLoad = tf.keras.models.load_model('model\/modeltest.h5')\n<\/code><\/p>\n"},{"AnswerId":"57044939","CreationDate":"2019-07-15T17:57:40.223","ParentId":null,"OwnerUserId":"8188662","Title":null,"Body":"<pre><code>from tensorflow.keras.initializers import glorot_uniform\n\nloaded_model = tf.keras.models.load_model(\"pruned.h5\",custom_objects={'GlorotUniform': glorot_uniform()})\n\n<\/code><\/pre>\n\n<p>this worked for me when importing tensorflow keras<\/p>\n"},{"AnswerId":"56993437","CreationDate":"2019-07-11T16:17:01.363","ParentId":null,"OwnerUserId":"11450077","Title":null,"Body":"<p>if you are loading the architecture and weights separtly, while loading archtiecture of the model change :<\/p>\n\n<pre><code>models.model_from_json(json)\n<\/code><\/pre>\n\n<p>to : <\/p>\n\n<pre><code>tf.keras.models.model_from_json(json)\n<\/code><\/pre>\n\n<p>and the problem is solved<\/p>\n"},{"AnswerId":"53689541","CreationDate":"2018-12-09T05:01:21.463","ParentId":null,"OwnerUserId":"6029751","Title":null,"Body":"<p>Wow I, just spent 6 Hours of my life trying to figure this out.. Dmitri posted a solution to this here: <a href=\"https:\/\/stackoverflow.com\/questions\/53051274\/i-trained-a-keras-model-on-google-colab-now-not-able-to-load-it-locally-on-my-s\">I trained a keras model on google colab. Now not able to load it locally on my system.<\/a><\/p>\n\n<p>I'm just basically reposting it here because it worked for me.<\/p>\n\n<p>This looks like some kind of a serialization bug in keras. \nIf you wrap your load_model with the below CustomObjectScope thingy... all should work..<\/p>\n\n<pre><code>import keras\nfrom keras.models import load_model\nfrom keras.utils import CustomObjectScope\nfrom keras.initializers import glorot_uniform\n\nwith CustomObjectScope({'GlorotUniform': glorot_uniform()}):\n        model = load_model('imdb_mlp_model.h5')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53184109,"AnswerCount":3,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-11-07T05:40:44.247","AcceptedAnswerId":53184879.0,"OwnerUserId":5686692.0,"Title":"Is it possible to restore a tensorflow estimator from saved model?","Body":"<p>I use <code>tf.estimator.train_and_evaluate()<\/code> to train my custom estimator. My dataset is partitioned 8:1:1 for training, evaluation and test. At the end of the training, I would like to restore the best model, and evaluate the model using <code>tf.estimator.Estimator.evaluate()<\/code> with the test data. The best model is currently exported using <code>tf.estimator.BestExporter<\/code>.<\/p>\n\n<p>While <code>tf.estimator.Estimator.evaluate()<\/code> accepts <code>checkpoint_path<\/code> and restores variables, I cannot find any easy way to use the exported model generated by <code>tf.estimator.BestExporter<\/code>. I could of course keep all checkpoints during training, and look for the best model by myself, but that seems quite suboptimal.<\/p>\n\n<p>Could anyone tell me an easy workaround? Maybe it is possible to convert a saved model to a checkpoint?<\/p>\n","answers":[{"AnswerId":"53184879","CreationDate":"2018-11-07T07:06:22.587","ParentId":null,"OwnerUserId":"5686692","Title":null,"Body":"<p>Hope someone else will find a cleaner way..<\/p>\n\n<p><code>tf.estimator.BestExporter<\/code> exports the best model like this:<\/p>\n\n<pre><code>&lt;your_estimator.model_dir&gt;\n+--export\n   +--best_exporter\n      +--xxxxxxxxxx(timestamp)\n         +--saved_model.pb\n         +--variables\n            +--variables.data-00000-of-00001\n            +--variables.index\n<\/code><\/pre>\n\n<p>On the other hand, in <code>your_estimator.model_dir<\/code>, checkpoints are stored in three files. <\/p>\n\n<pre><code>model.ckpt-xxxx.data-00000-of-00001\nmodel.ckpt-xxxx.index\nmodel.ckpt-xxxx.meta\n<\/code><\/pre>\n\n<p>First, I used <code>tf.estimator.Estimator.evaluate(..., checkpoint_path='&lt;your_estimator.model_dir&gt;\/export\/best_exporter\/&lt;xxxxxxxxxx&gt;\/variables\/variables')<\/code>, but this did not work.<\/p>\n\n<p>After copying one of the metafiles in <code>your_estimator.model_dir<\/code>, and renaming it \"variables.meta\", the evaluation seemed to work all right.<\/p>\n"},{"AnswerId":"53415153","CreationDate":"2018-11-21T15:17:15.823","ParentId":null,"OwnerUserId":"5623899","Title":null,"Body":"<p>I am also new to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> API but I think I know what you are looking for, although it is equally annoying.<\/p>\n\n<p>From this <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a>, which is a toy custom <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> with some bells and whistles added on:<\/p>\n\n<pre><code>from tensorflow.contrib import predictor\npredict_fn = predictor.from_saved_model(&lt;model_dir&gt;)\npredict_fn(pred_features) # pred_features corresponds to your input features\n<\/code><\/pre>\n\n<p>and this estimator both uses a <code>BestExporter<\/code><\/p>\n\n<pre><code>exporter = tf.estimator.BestExporter(\n    name=\"best_exporter\",\n    serving_input_receiver_fn=serving_input_receiver_fn,\n    exports_to_keep=5\n) # this will keep the 5 best checkpoints\n<\/code><\/pre>\n\n<p>as well as just exports the model after training:<\/p>\n\n<p><code>est.export_savedmodel('.\/here', serving_input_receiver_fn)<\/code> <\/p>\n\n<p>If it irks you that the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> API has no \"proper\" way to load a <code>SavedModel<\/code>, I already created an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23824\" rel=\"nofollow noreferrer\">issue<\/a> on GitHub.<\/p>\n\n<p>However, if you are trying to load it onto a different device, see my other questions:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/53356558\/tensorflow-v1-10-load-savedmodel-with-different-device-placement-or-manually-se\">TensorFlow v1.10+ load SavedModel with different device placement or manually set dynamic device placement?<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/53409652\/tensorflow-estimator-clear-deivces-in-exporters\">TensorFlow Estimator clear_deivces in exporters?<\/a><\/p><\/li>\n<\/ul>\n\n<p>which address device placement, for which there are other GitHub issues<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23900\" rel=\"nofollow noreferrer\">No clear_devices in BestExporter #23900<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23900\" rel=\"nofollow noreferrer\">Relative Device Placement #23834<\/a><\/p><\/li>\n<\/ul>\n\n<p>In short, at the moment, the device you train on is the device you <strong>MUST<\/strong> have available when you load your <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> if you export using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> exporters. If you manually export your <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> in the <code>model_fn<\/code> if you set <code>clear_devices<\/code>, then you should be good to go. At the moment there does not seem to be a way to change this after you export your model.<\/p>\n"},{"AnswerId":"58345886","CreationDate":"2019-10-11T17:20:07.280","ParentId":null,"OwnerUserId":"3662820","Title":null,"Body":"<p>Based on the resolution to @SumNeuron's Github issue <code>tf.contrib.estimator.SavedModelEstimator<\/code> is the way to load from a saved model to an <code>Estimator<\/code>. <\/p>\n\n<p>The following works for me:<\/p>\n\n<pre><code>estimator = tf.contrib.estimator.SavedModelEstimator(saved_model_dir)\nprediction_results = estimator.predict(input_fn)\n<\/code><\/pre>\n\n<p>Baffling that this is essentially completely undocumented.<\/p>\n"}]}
+{"QuestionId":53184153,"AnswerCount":1,"Tags":"<python><tensorflow><keras><cluster-computing><slurm>","CreationDate":"2018-11-07T05:46:17.233","AcceptedAnswerId":null,"OwnerUserId":4564882.0,"Title":"keras (tensorflow backend) run on a cluster using slurm","Body":"<p>I have the opportunity to run my Tensorflow training on a cluster computer with slurm workload manager (the cluster contains nearly 400000 cores, 40000 GB of RAM, Performance is Rmax=500 TFlop\/s and Rpeak=1000 TFlop\/s, AMD GPUs). <\/p>\n\n<p>I work on image processing projects using deep learning algorithms.<\/p>\n\n<p>My question is how to scale my keras deep learning to run on this cluster using slurm as workload manager ?<\/p>\n","answers":[{"AnswerId":"53186344","CreationDate":"2018-11-07T09:07:57.393","ParentId":null,"OwnerUserId":"10617409","Title":null,"Body":"<p>Use Horovod to scale out the Keras training - <a href=\"https:\/\/github.com\/uber\/horovod\" rel=\"nofollow noreferrer\">https:\/\/github.com\/uber\/horovod<\/a><\/p>\n"}]}
+{"QuestionId":53184978,"AnswerCount":0,"Tags":"<python><lstm><pytorch>","CreationDate":"2018-11-07T07:16:12.810","AcceptedAnswerId":null,"OwnerUserId":10608399.0,"Title":"Pytorch DataLoader - working with individual batches","Body":"<p>I am building an LSTM for part-of-speech tagging, and am currently using <a href=\"https:\/\/github.com\/jiangqy\/LSTM-Classification-Pytorch\/blob\/master\/main.py\" rel=\"nofollow noreferrer\">this LSTM<\/a> as reference to understand how to use the Pytorch DataLoader. However, I'm confused as to how to work with individual batches (official Pytorch documentation doesn't shed any light for me). More specifically I need to pad my sentences so that all the sentences <em>in the same batch<\/em> are of the same length (different batches can have different lengths).<\/p>\n\n<p>My code currently looks like:<\/p>\n\n<pre><code>train_loader = DataLoader(train_set, batch_size=BATCH_SIZE)\n\n# other stuff...\n\nfor epoch in range(0, EPOCH_COUNT):\n    for iter, sentence in enumerate(train_loader):\n        model.zero_grad()\n        model.hidden = model.init_hidden()\n        sentence_in = prep_sentence(sentence, word_to_ix)\n        targets = prep_tags(tags, tag_to_ix)\n        # do forward pass and then backprop\n<\/code><\/pre>\n\n<p>Since I'm enumerating train_loader and iterating through each sentence, it's not clear to me how\/at what point I can get the max sentence length of a single batch and then run my add_padding() function.<\/p>\n","answers":[]}
+{"QuestionId":53185089,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><word-embedding>","CreationDate":"2018-11-07T07:27:13.080","AcceptedAnswerId":null,"OwnerUserId":1259561.0,"Title":"TensorFlow: How to constrain\/update only some rows of a variable?","Body":"<p>I have an embedding matrix variable <\/p>\n\n<pre><code>embs = tf.get_variable(name=\"embs\", shape=[total, hidden_size])\n<\/code><\/pre>\n\n<p>I want to constrain the norm to be unit<\/p>\n\n<pre><code>embs_unit_norm_op = tf.assign(embs, tf.keras.constraints.unit_norm(axis=1)(embs))\n<\/code><\/pre>\n\n<p>But the <code>embs<\/code> matrix is too large, and only some rows changed by the previous gradient applying step. I want to reduce computation by only constraining\/updating these 'active' embeddings. <\/p>\n\n<p>I have a list of row numbers, e.g., <code>e = [1, 3, 5]<\/code>, same as one-hot ID used in embedding lookup. How could I constrain\/update only these \"active\" embeddings?<\/p>\n","answers":[{"AnswerId":"53186487","CreationDate":"2018-11-07T09:17:55.410","ParentId":null,"OwnerUserId":"8128643","Title":null,"Body":"<p>You can try <code>tf.scatter_update<\/code> or <code>tf.scatter_nd_update<\/code><\/p>\n"}]}
+{"QuestionId":53185283,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T07:44:06.337","AcceptedAnswerId":null,"OwnerUserId":8254102.0,"Title":"How to correctly implement validation when training using Tensorflow?","Body":"<p>there. I want to validate my neural network during the process of training. \nA placeholder is created to decide whether train_batch or test_batch should be feed into the network to realize validation.\nHere is my sample code. In the <strong>train.tfrecords<\/strong>, the number is from 0 to 100, and in the <strong>test.tfrecords<\/strong>, it is from 200 to 240.<\/p>\n\n<pre><code>create_tfrecord(filename=r'test', begin_num=200, end_num=240)\n\nfilename = r'.\/train.tfrecords'\nfilename_que = tf.train.string_input_producer([filename],\n                                              num_epochs=1)\nnum = decode_tfrecord(filename_que)\n\ntestfilename = r'.\/test.tfrecords'\ntest_que = tf.train.string_input_producer([testfilename],\n                                          num_epochs=None)\ntest_num = decode_tfrecord(test_que)\n\nis_training = tf.placeholder(dtype=tf.bool, shape=())\n\nnum = tf.cond(is_training, lambda:num, lambda:test_num)    \n\ncoord = tf.train.Coordinator()\n\ninit_op = tf.local_variables_initializer()\nwith tf.Session() as sess:\n\n    sess.run(init_op)\n    threads = tf.train.start_queue_runners(coord=coord, sess=sess)\n\n    try:\n        i = 0\n        while not coord.should_stop():\n\n            feed_dict = {is_training: True}\n            number = sess.run(num, feed_dict=feed_dict)\n            print(\"global_step: %i, read_num: %i\" %(i, number))\n\n            if i % 5 == 0 and i != 0:\n                feed_dict = {is_training: False}\n                number = sess.run(num, feed_dict=feed_dict)\n                print(\"global_step: %i, read_num: %i\" %(i, number))\n\n            i += 1\n\n\n    except tf.errors.OutOfRangeError:\n        print(\"done.\")\n\n    finally:\n        coord.request_stop()\n        coord.join(threads)\n<\/code><\/pre>\n\n<p>However, the output is not fully the same as what i want. The test_batch is not begin from 200 that is the first number in <strong>test.tfrecords<\/strong> rather than begin from 206. After that, the train_batch is begin from 7 rather than 6. Here is some outputs.<\/p>\n\n<pre><code>global_step: 0, read_num: 0\nglobal_step: 1, read_num: 1\nglobal_step: 2, read_num: 2\nglobal_step: 3, read_num: 3\nglobal_step: 4, read_num: 4\nglobal_step: 5, read_num: 5\nglobal_step: 5, read_num: 206\nglobal_step: 6, read_num: 7\nglobal_step: 7, read_num: 8\nglobal_step: 8, read_num: 9\nglobal_step: 9, read_num: 10\nglobal_step: 10, read_num: 11\nglobal_step: 10, read_num: 212\n<\/code><\/pre>\n\n<p>Those outputs acts like <strong>there exists a counter which is used both in training and validation.<\/strong> Once the epochs is set to 1, not all my training set could be used in the training stage, and the order my test_batch be fetched is not what i want. I think this way is a wrong way to realize validation when training.<\/p>\n\n<p>So, is there anyone could tell me how can i correctly implement validating during training my network?<\/p>\n\n<p>Thanks in advance. <\/p>\n","answers":[]}
+{"QuestionId":53186129,"AnswerCount":3,"Tags":"<python><machine-learning><keras><image-preprocessing>","CreationDate":"2018-11-07T08:53:52.907","AcceptedAnswerId":null,"OwnerUserId":8796118.0,"Title":"How to apply normalization to images in testing phase when using keras ImageDataGenerator?","Body":"<p>I'm trying to predict a new image using trained model. My accuracy is 95%. But the predict_classes always return the first label [0] whatever I input. \nI guess one of the reason is I use <code>featurewise_center=True<\/code> and <code>samplewise_center=True<\/code> in <code>ImageDataGenerator<\/code>. I think I should do the same thing on my input image. But I can't find what did these function do to the image.<\/p>\n\n<p>Any suggestion will be appreciated.<\/p>\n\n<p><code>ImageDataGenerator<\/code> code:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\nsamplewise_center=True,\nrescale=1. \/ 255,\nshear_range=30,\nzoom_range=30,\nrotation_range=20,\nwidth_shift_range=0.2,\nheight_shift_range=0.2)\n\ntest_datagen = ImageDataGenerator(\nsamplewise_center=True,\nrescale=1. \/ 255,\nshear_range=30,\nzoom_range=30,\nrotation_range=20,\nwidth_shift_range=0.2,\nheight_shift_range=0.2,\nhorizontal_flip=True)\n<\/code><\/pre>\n\n<p>Prediction code (I use 100*100*3 image to train the model):<\/p>\n\n<pre><code>model = load_model('CNN_model.h5')\nimg = cv2.imread('train\/defect\/6.png')\nimg = cv2.resize(img,(100,100))\nimg = np.reshape(img,[1,100,100,3])\nimg = img\/255.\n\nclasses = model.predict_classes(img)\n\nprint (classes)\n<\/code><\/pre>\n\n<p>updated 11\/14:<\/p>\n\n<p>I change my code to predict image like below. But the model still predict the same class even if I feed the image which I have used to train my model(and got 95% accuarcy). Is there anything I missed?<\/p>\n\n<pre><code>model = load_model('CNN_model.h5')\nimg = cv2.imread('train\/defect\/6.png')\nimg = cv2.resize(img,(100,100))\nimg = np.reshape(img,[1,100,100,3])\nimg = np.array(img, dtype=np.float64) \nimg = train_datagen.standardize(img)\n\nclasses = model.predict_classes(img)\nprint(classes)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53188575","CreationDate":"2018-11-07T11:27:29.930","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to use the <code>standardize()<\/code> method of <code>ImageDataGenerator<\/code> instance. From <a href=\"https:\/\/keras.io\/preprocessing\/image\/#standardize\" rel=\"nofollow noreferrer\">Keras documentation<\/a>:<\/p>\n\n<blockquote>\n  <h2>standardize<\/h2>\n\n<pre><code>standardize(x)\n<\/code><\/pre>\n  \n  <p>Applies the normalization configuration to a batch of inputs.<\/p>\n  \n  <p><strong>Arguments<\/strong><\/p>\n  \n  <ul>\n  <li>x: Batch of inputs to be normalized.<\/li>\n  <\/ul>\n  \n  <p><strong>Returns<\/strong><\/p>\n  \n  <p>The inputs, normalized.<\/p>\n<\/blockquote>\n\n<p>So it would be like this:<\/p>\n\n<pre><code>img = cv2.imread('train\/defect\/6.png')\nimg = cv2.resize(img,(100,100))\nimg = np.reshape(img,[1,100,100,3])\nimg = train_datagen.standardize(img)\n\nclasses = model.predict_classes(img)\n<\/code><\/pre>\n\n<p>Note that it would apply the rescaling as well so there is no need to do it yourself (i.e. remove <code>img = img\/255.<\/code>).<\/p>\n\n<p>Further, keep in mind that since you have set <code>featurewise_ceneter=True<\/code> you need to use <a href=\"https:\/\/keras.io\/preprocessing\/image\/#fit\" rel=\"nofollow noreferrer\"><code>fit()<\/code><\/a> method of generator before using it for training:<\/p>\n\n<pre><code>train_datagen.fit(training_data)\n\n# then use fit_generator method\nmodel.fit_generator(train_datagen, ...)\n<\/code><\/pre>\n"},{"AnswerId":"54512628","CreationDate":"2019-02-04T08:45:01.187","ParentId":null,"OwnerUserId":"4983201","Title":null,"Body":"<p>Not a complete answer but some information:<\/p>\n\n<p>From <a href=\"https:\/\/gist.github.com\/fchollet\/0830affa1f7f19fd47b06d4cf89ed44d\" rel=\"nofollow noreferrer\">this link<\/a> that is referenced in keras docs:<\/p>\n\n<pre><code># this is the augmentation configuration we will use for training\ntrain_datagen = ImageDataGenerator(\n    rescale=1. \/ 255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)\n\n# this is the augmentation configuration we will use for testing:\n# only rescaling\ntest_datagen = ImageDataGenerator(rescale=1. \/ 255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode='binary')\n\nvalidation_generator = test_datagen.flow_from_directory(\n    validation_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=batch_size,\n    class_mode='binary')\n<\/code><\/pre>\n\n<p>I think you should do this way for training. Then for test I think using <code>train_datagen.standardize<\/code> is true.<\/p>\n"},{"AnswerId":"56275747","CreationDate":"2019-05-23T12:56:00.760","ParentId":null,"OwnerUserId":"11544937","Title":null,"Body":"<p>I think it's the problem that you used <code>cv2<\/code> to import your images, because when you use <code>cv2.imread<\/code>, the channels are not \"r,g,b\" but \"b,g,r\".<\/p>\n\n<p>for example,<\/p>\n\n<pre><code>import cv2\nfrom tensorflow.keras.preprocessing import image\n\nbgr = cv2.imread('r.jpg')\nrgb = np.array(image.load_img('r.jpg'))\nprint(bgr[1,1,:],rgb[1,1,:])\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>[ 83 113   0] [  0 114  83]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53186310,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><indexing><tensor>","CreationDate":"2018-11-07T09:05:21.220","AcceptedAnswerId":null,"OwnerUserId":10550766.0,"Title":"Tensorflow: Filtering 3D Index duplicates by their maximum Values","Body":"<p>I am trying to create a filter mask that removes duplicate Indices from an vector by comparing which of their respective values is greater. <\/p>\n\n<p>My current approach is:<br><\/p>\n\n<ol>\n<li>Transform 3-D Index to 1-D<\/li>\n<li>Check the 1-D Index for uniqueness<\/li>\n<li>Calculate the maximum values of each unique index<\/li>\n<li>Compare the maximum values with the original values. If the same value exists, keep that 3-D Index.<\/li>\n<\/ol>\n\n<p>I want to get an filter array so I can apply a <code>boolean_mask<\/code> to other tensors as well. For this example the mask should look the following way:\n<code>[False True  True  True  True]<\/code>.<\/p>\n\n<p>My current code kind of works unless the values themselves are also duplicated. However this seems to be the case when I am using it therefore I need to find a better solution to it.<\/p>\n\n<p>Here is an examplary of how my Code looks<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Dummy Input values with same Structure as the real\nx_cells   = tf.constant([1,2,3,4,1], dtype=tf.int32)   # Index_1\ny_cells   = tf.constant([4,4,4,4,4], dtype=tf.int32)   # Index_2\niou_index = tf.constant([1,2,3,4,1], dtype=tf.int32) # Index_3\niou_max   = tf.constant([1.,2.,3.,4.,5.], dtype=tf.float32) # Values\n\n# my Output should be a mask that is [False True True True True]\n# So if i filter this i get e.g. x_cells = [2,3,4,1] or iou_max = [2.,3.,4.,5.]\n\nmax_dim_y = tf.constant(10)\nmax_dim_x = tf.constant(20)\nnum_anchors = 5\nstride = 32\n<\/code><\/pre>\n\n<hr>\n\n<pre><code># 1. Transforming the 3D-Index to 1D\ntmp = tf.stack([x_cells, y_cells, iou_index], axis=1)\nindices = tf.matmul(tmp, [[max_dim_y * num_anchors],     [num_anchors],[1]])\n\n# 2. Looking for unique \/ duplicate indices\ny, idx = tf.unique(tf.squeeze(indices))\n\n# 3. Calculating the maximum values of each unique index.\n# An function like unsorted_segment_argmax() would be awesome here\nnum_segments = tf.shape(y)[0]\nious = tf.unsorted_segment_max(iou_max, idx, num_segments)\n\niou_max_length = tf.shape(iou_max)[0]\nious_length = tf.shape(ious)[0]\n\n# 4. Compare all max values to original values.\niou_max_tiled = tf.tile(iou_max, [ious_length])\niou_reshaped = tf.reshape(iou_max_tiled, [ious_length, iou_max_length])\niou_max_reshaped = tf.transpose(iou_reshaped)\nfilter_mask = tf.reduce_any(tf.equal(iou_max_reshaped, ious), -1)\nfilter_mask = tf.reshape(filter_mask, shape=[-1])\n<\/code><\/pre>\n\n<hr>\n\n<p>This code above will fail if we simply change the value of the <code>iou_max<\/code> Variable in the beginning to:<\/p>\n\n<pre><code>x_cells = tf.constant([1,2,3,4,1], dtype=tf.int32)\ny_cells = tf.constant([4,4,4,4,4], dtype=tf.int32)\niou_index = tf.constant([1,2,3,4,1], dtype=tf.int32)\niou_max = tf.constant([2.,2.,3.,4.,5.], dtype=tf.float32)\n<\/code><\/pre>\n\n<hr>\n","answers":[{"AnswerId":"53188096","CreationDate":"2018-11-07T10:56:00.030","ParentId":null,"OwnerUserId":"10550766","Title":null,"Body":"<h2>My current workaround changed point 4 of my question:<\/h2>\n\n<p>Basically I changed that I compare the tuples instead of the single values. This leads to me beeing able to logically check if both, index AND value are in the remaining values from <strong>3.<\/strong><\/p>\n\n<pre><code># 4. Compare a Max Value and Indices with original values\nrem_index_val_pair = tf.stack([ious, tf.cast(y, dtype=tf.float32)], axis=1)\norig_val_index_pair = tf.stack([iou_max, tf.cast(indices, dtype=tf.float32)], axis=1)\n\norig_val_index_pair_t = tf.tile(orig_val_index_pair, [1, ious_length])\norig_val_index_pair_s = tf.reshape(orig_val_index_pair_t, [iou_max_length, ious_length, 2])\nfilter_mask_1 = tf.equal(orig_val_index_pair_s, rem_index_val_pair)\nfilter_mask_2 = tf.reduce_all(filter_mask_1, -1)\nfilter_mask_3 = tf.reduce_any(filter_mask_2, -1)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code># The orig_val_index_pair_s looks like the following\na =  [[[  2.  71.][  2.  71.][  2.  71.][  2.  71.]\n     [[  2. 122.][  2. 122.][  2. 122.][  2. 122.]]\n     [[  3. 173.][  3. 173.][  3. 173.][  3. 173.]]\n     [[  4. 224.][  4. 224.][  4. 224.][  4. 224.]]\n     [[  5.  71.][  5.  71.][  5.  71.][  5.  71.]]]\n# I then compare it to the rem_max_val_pair which looks like this.\nb =  [[  5.  71.][  2. 122.][  3. 173.][  4. 224.]]\n\n# Using equal(a,b) will now compare each of the values resulting in:\nc = [[[False  True][ True False][False False][False False]]\n     [[False False][ True  True][False False][False False]]\n     [[False False][False False][ True  True][False False]]\n     [[False False][False False][False False][ True  True]]\n     [[ True  True][False False][False False][False False]]]\n\n# Using tf.reduce_all(c, -1) I can filter the bool pairs with a logical And. \n# (This kicks out my false positives from before).\n# Afterwards I can check if the line has any true value by tf.reduce_any().\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>IMO this solution still is a dirty workaround. So if you have any better solution proposals please share them. :)<\/strong><\/p>\n"}]}
+{"QuestionId":53186736,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-07T09:34:47.003","AcceptedAnswerId":53189865.0,"OwnerUserId":5516760.0,"Title":"RuntimeError: Attempting to deserialize object on CUDA device 2 but torch.cuda.device_count() is 1","Body":"<p>I've got a snippet of python code for training a model. The problem is that after running:<\/p>\n\n<pre><code>loaded_state = torch.load(model_path+seq_to_seq_test_model_fname)\n<\/code><\/pre>\n\n<p>to load a pretrained model,I'm getting: <\/p>\n\n<pre><code>  Traceback (most recent call last):\n  File \"img_to_text.py\", line 480, in &lt;module&gt;\n    main()\n  File \"img_to_text.py\", line 475, in main\n    r = setup_test()\n  File \"img_to_text.py\", line 259, in setup_test\n    s2s_data = s2s.setup_test()\n  File \"\/media\/ahrzb\/datasets\/notebooks\/mzh\/SemStyle\/semstyle\/code\/seq2seq_pytorch.py\", line 220, in setup_test\n    loaded_state= torch.load(model_path+seq_to_seq_test_model_fname)\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 358, in load\n    return _load(f, map_location, pickle_module)\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 542, in _load\n    result = unpickler.load()\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 508, in persistent_load\n    data_type(size), location)\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 372, in restore_location\n    return default_restore_location(storage, location)\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 104, in default_restore_location\n    result = fn(storage, location)\n  File \"\/home\/ahrzb\/.pyenv\/versions\/2.7.15\/envs\/mzh2.7\/lib\/python2.7\/site-packages\/torch\/serialization.py\", line 85, in _cuda_deserialize\n    device, torch.cuda.device_count()))\n<\/code><\/pre>\n\n<p>I think this is because they have trained the model on two GPUs and I need to load it in one GPU. I changed this line: <\/p>\n\n<pre><code>loaded_state = torch.load(model_path+seq_to_seq_test_model_fname) \n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>loaded_state = torch.load(model_path+seq_to_seq_test_model_fname, map_location={'cuda:1': 'cuda:0'} ) \n<\/code><\/pre>\n\n<p>in order to map data of cuda 1 to cuda 0 but it did not work.<\/p>\n","answers":[{"AnswerId":"53189865","CreationDate":"2018-11-07T12:54:54.850","ParentId":null,"OwnerUserId":"5516760","Title":null,"Body":"<p>I just figured it out:<\/p>\n\n<pre><code> loaded_state = torch.load(model_path+seq_to_seq_test_model_fname,map_location='cuda:0')\n<\/code><\/pre>\n\n<p>is the solution<\/p>\n"}]}
+{"QuestionId":53186793,"AnswerCount":1,"Tags":"<tensorflow><matplotlib>","CreationDate":"2018-11-07T09:37:38.757","AcceptedAnswerId":null,"OwnerUserId":7909469.0,"Title":"How to plot weights of fully connected layer?","Body":"<p>I am using a a fully connected network with 4 input and 2 output nodes. I store the Weights of my network after completely training it. Suppose here is my weight matrix<\/p>\n\n<pre><code>`W = np.array([[0.8,0.02],[0.5,0.4],[0.3,0.2],[0.1,0.7]])`\n<\/code><\/pre>\n\n<p>I want to visualize that what weights each class has adopted. How I can do that? I searched the codes related to this they are are using plt.imshow. Should I simply mention plt.imshow(W) to visualize weights? <\/p>\n","answers":[{"AnswerId":"53196489","CreationDate":"2018-11-07T19:30:12.620","ParentId":null,"OwnerUserId":"10458815","Title":null,"Body":"<p>You should use TensorBoard for this. Also, you should not need to store the weights manually, as they are stored by TensorFlow. You can access them in a couple of different ways, such as with <code>tf.trainable_variables()<\/code>, or <code>tape.watched_variables()<\/code> in eager mode. Then its just a matter of looping through the variables for the weights you want.<\/p>\n\n<p>To plot your weights in TensorBoard, check out this: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/summary\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/summary<\/a><\/p>\n"}]}
+{"QuestionId":53187009,"AnswerCount":1,"Tags":"<tensorflow><keras><classification>","CreationDate":"2018-11-07T09:50:28.727","AcceptedAnswerId":53187100.0,"OwnerUserId":9880187.0,"Title":"Classfying image with multi standard","Body":"<p>I'm not sure if it is right to describe this as \"multi standard\" but my problem is this.<\/p>\n\n<p>I'm classfying pictures of fashion products,\nand I want it to be classified some how like:<\/p>\n\n<pre><code>id, brand, product_kind, sex\n1, gucci, wallet, woman\n2, H&amp;M, backpack, unisex\n3, zara, coat, man\n<\/code><\/pre>\n\n<p>after searching I think it is similar to multi label classfication, but as for my understanding, multi label is more like:<\/p>\n\n<pre><code>id, gucci, H&amp;M, zara, wallet, backpack, coat, woman, unisex, man\n1, 1, 0, 0, 1, 0, 0, 1, 0, 0\n2, 0, 1, 0, 0, 1, 0, 0, 1, 0\n3, 0, 0, 1, 0, 0, 1, 0, 0, 1\n<\/code><\/pre>\n\n<p>My purpose is to predict one from brand, one from kind, one from sex.\nCould anyone recommend me any article or tutorial for this problem?<\/p>\n","answers":[{"AnswerId":"53187100","CreationDate":"2018-11-07T09:56:29.963","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>Your problem is basically 3 different classifiers, so the \"easiest\" would be to train three independent ones.<\/p>\n\n<p>You can also have 3 independent output layers (with perhaps additional hidden layers for each of them), each of them with the traditional softmax outputs, and your cost function is the sum of the loss for each of these 3 layers. This way, the three classifiers share feature extractions and part of the logic.<\/p>\n\n<p>Something like:<\/p>\n\n<pre><code>Features\n   |\nHidden layer 1\n   |\nHidden layer 2\n   |-----------------------|--------------------|\nHidden layer 3-1      Hidden layer 3-2      Hidden layer 3-3\n   |                       |                    |\nHidden layer 4-1      Hidden layer 4-2      Hidden layer 4-3\n   |                       |                    |\nsoftmax 1              softmax 2              softmax 3\n<\/code><\/pre>\n"}]}
+{"QuestionId":53187210,"AnswerCount":0,"Tags":"<python><tensorflow><image-processing><conv-neural-network><image-scaling>","CreationDate":"2018-11-07T10:02:46.900","AcceptedAnswerId":null,"OwnerUserId":9723527.0,"Title":"reduce image size without losing details","Body":"<p>I am new to image processing. I have hand scan images and they are huge in size (~14000x12000 pixels). I need to reduce image size without losing much details (like patterns and ridges present on the fingers and palm). My aim is to use those for classification using neural network. Currently I am getting out of memory error, when ever i tried to run the codes. I used thumb prints and it ran well because those are smaller in size. Please suggest algorithms or methods in which this problem can be solved. \nThank you in advance. <\/p>\n","answers":[]}
+{"QuestionId":53187570,"AnswerCount":1,"Tags":"<apache-spark><keras><pyspark><deep-learning>","CreationDate":"2018-11-07T10:25:10.977","AcceptedAnswerId":null,"OwnerUserId":3156383.0,"Title":"How to pass deep learning model data to map function in Spark","Body":"<p>I have a very simple use-case where I am reading large number of images as rdd from s3 using sc.binaryFiles method. Once this RDD is created I am passing the content inside the rdd to the vgg16 feature extractor function. So, in this I will need the model data using which the feature extraction will be done, so I am putting the model data into broadcast variable and then accesing the value in each map function. Below is the code:-<\/p>\n\n<pre><code>s3_files_rdd = sc.binaryFiles(RESOLVED_IMAGE_PATH)\n\ns3_files_rdd.persist()\n\nmodel_data = initVGG16()\nbroadcast_model = sc.broadcast(model_data)\n\nfeatures_rdd = s3_files_rdd.mapPartitions(extract_features_)\n\nresponse_rdd = features_rdd.map(lambda x: (x[0], write_to_s3(x, OUTPUT, FORMAT_NAME)))\n<\/code><\/pre>\n\n<p>extract_features_ method:-<\/p>\n\n<pre><code>def extract_features_(xs):\n    model_data = initVGG16()\n    for k, v in xs:\n        yield k, extract_features2(model_data,v)\n<\/code><\/pre>\n\n<p>extract_features method:-<\/p>\n\n<pre><code>from keras.preprocessing import image\nfrom keras.applications.vgg16 import VGG16\nfrom keras.models import Model\nfrom io import BytesIO\nfrom keras.applications.vgg16 import preprocess_input\ndef extract_features(model,obj):\n    try:\n        print('executing vgg16 feature extractor...')\n        img = image.load_img(BytesIO(obj), target_size=(224, 224,3))\n        img_data = image.img_to_array(img)\n        img_data = np.expand_dims(img_data, axis=0)\n        img_data = preprocess_input(img_data)\n        vgg16_feature = model.predict(img_data)[0]\n        print('++++++++++++++++++++++++++++',vgg16_feature.shape)\n        return vgg16_feature\n    except Exception as e:\n        print('Error......{}'.format(e.args))\n        return []\n<\/code><\/pre>\n\n<p>write to s3 method:-<\/p>\n\n<pre><code>def write_to_s3(rdd, output_path, format_name):\n    file_path = rdd[0]\n    file_name_without_ext = get_file_name_without_ext(file_name)\n    bucket_name = output_path.split('\/', 1)[0]\n\n    final_path = 'deepak' + '\/' + file_name_without_ext + '.' + format_name\n\n    LOGGER.info(\"Saving to S3....\")\n    cci = cc.get_interface(bucket_name, ACCESS_KEY=os.environ.get(\"AWS_ACCESS_KEY_ID\"),\n                           SECRET_KEY=os.environ.get(\"AWS_SECRET_ACCESS_KEY\"), endpoint_url='https:\/\/s3.amazonaws.com')\n    response = cci.upload_npy_array(final_path, rdd[1])\n    return response\n<\/code><\/pre>\n\n<p>Inside the write_to_s3 method I am getting the RDD, extracting the key name to be saved and bucket. then using a library called cottoncandy to drectly save the RDD content which is numpy array in my case instead of saving any intermediate file.<\/p>\n\n<p>I am getting below error :-<\/p>\n\n<pre><code>127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/cloudpickle.py\", line 600, in save_reduce\n    save(state)\n  File \"\/usr\/lib64\/python2.7\/pickle.py\", line 286, in save\n    f(self, obj) # Call unbound method with explicit self\n  File \"\/usr\/lib64\/python2.7\/pickle.py\", line 655, in save_dict\n    self._batch_setitems(obj.iteritems())\n  File \"\/usr\/lib64\/python2.7\/pickle.py\", line 687, in _batch_setitems\n    save(v)\n  File \"\/usr\/lib64\/python2.7\/pickle.py\", line 306, in save\n    rv = reduce(self.proto)\nTypeError: can't pickle thread.lock objects\nTraceback (most recent call last):\n  File \"one_file5.py\", line 98, in &lt;module&gt;\n    run()\n  File \"one_file5.py\", line 89, in run\n    LOGGER.info('features_rdd rdd created,...... %s',features_rdd.count())    \n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 1041, in count\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 1032, in sum\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 906, in fold\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 809, in collect\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 2455, in _jrdd\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 2388, in _wrap_function\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/rdd.py\", line 2374, in _prepare_for_python_RDD\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/serializers.py\", line 464, in dumps\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/cloudpickle.py\", line 704, in dumps\n  File \"\/mnt\/yarn\/usercache\/hadoop\/appcache\/application_1541576150127_0010\/container_1541576150127_0010_01_000001\/pyspark.zip\/pyspark\/cloudpickle.py\", line 162, in dump\npickle.PicklingError: Could not serialize object: TypeError: can't pickle thread.lock objects.\n<\/code><\/pre>\n\n<p>When I am commenting out the the code part of features_rdd, then the program runs fine which means something is not proper in the features_rdd part. Not sure what I am doing wrong here.<\/p>\n\n<p>I am running the program in AWS EMR, with 4 executors. \nexecutor core 7 \nexecutor RAM 8GB\nSpark version 2.2.1<\/p>\n","answers":[{"AnswerId":"53188965","CreationDate":"2018-11-07T11:54:46.523","ParentId":null,"OwnerUserId":"10618151","Title":null,"Body":"<p>Replace your current code with <code>mapPartitions<\/code>:<\/p>\n\n<pre><code>def extract_features_(xs):\n    model_data = initVGG16()\n    for k, v in xs:\n        yield k, extract_features(model_data, v)\n\nfeatures_rdd = s3_files_rdd.mapPartitions(extract_features_)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53187734,"AnswerCount":1,"Tags":"<tensorflow><mnist>","CreationDate":"2018-11-07T10:36:02.880","AcceptedAnswerId":null,"OwnerUserId":7061580.0,"Title":"Tensorflow cannot load mnist","Body":"<p>I downloaded a copy of mnist data from a random website because the official website was down.\nWhich directory can I put the mnist file for tensorflow's input_data method to pick it up<\/p>\n","answers":[{"AnswerId":"54466062","CreationDate":"2019-01-31T17:23:26.887","ParentId":null,"OwnerUserId":"4398242","Title":null,"Body":"<p>Maybe this <a href=\"https:\/\/stackoverflow.com\/questions\/42927841\/trouble-reading-mnist-with-tensorflow\">Link<\/a> help you.<\/p>\n\n<p>Another way to use MNIST dataset: some Frameworks like Tensorflow and Keras has this dataset. So you can use this:<\/p>\n\n<pre><code># Import MNIST\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n\n# Load data\nX_train = mnist.train.images\nY_train = mnist.train.labels\nX_test = mnist.test.images\nY_test = mnist.test.labels\n<\/code><\/pre>\n\n<p>The sample example in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn\" rel=\"nofollow noreferrer\">here<\/a>:<br>\nIn Keras also its same and you can import and use. Both Frameworks download it first and then just need to import.<\/p>\n"}]}
+{"QuestionId":53189050,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T11:59:27.613","AcceptedAnswerId":null,"OwnerUserId":7745839.0,"Title":"what does var_list in tf.train.Saver mean?","Body":"<p>I check the tensorflow API, If None, the var_list is defaults to the list of all saveable objects. I want to know what does the all saveable objects mean? tf.global_variables? \nW = tf.get_variable(\"W\", shape=[784, 256], initializer=tf.contrib.layers.xavier_initializer())\nb = tf.get_variable(\"b\", shape=[784, 256], initializer=tf.contrib.layers.xavier_initializer())\nm = tf.add(W, b)\nDoes m belongs to tf.global_variables? \nI am really confused. \n<a href=\"https:\/\/i.stack.imgur.com\/gyoJf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gyoJf.png\" alt=\"enter image description here\"><\/a>\nHere is my graph, I found train and loss node exists, so do I save the train and loss node? I just want to save model weights.....<\/p>\n","answers":[{"AnswerId":"53189184","CreationDate":"2018-11-07T12:08:21.543","ParentId":null,"OwnerUserId":"10288309","Title":null,"Body":"<p>The var_list variable says what it does, you may sometimes not want to save the entire model and only a part of it, for these purposes this list can be used, let me elaborate<\/p>\n\n<p>I am working on face recognition and i trained a CNN that extracts information from face images and gives a 512 size array of encodings that i can then pass to a SVM that can map these embeddings to a name, now at training i need the svm (or a simple ANN) i dont need it during inference, and saving its weights will lead to a larger model size and will consume more GPU memory, so i can while saving decide to store only the CNN variables, not the SVM\/ANN ones, hence i will pass names of the layers of the CNN in var_list and not names of the layers of the SVM<\/p>\n\n<p>Another aspect of saving for inference is that layers like dropout are useless during inference, so its best not to store them at all as they take a chunk of memory<\/p>\n\n<p>In your case i would suggest dont use it and save the model as it is<\/p>\n"}]}
+{"QuestionId":53189299,"AnswerCount":1,"Tags":"<image><tensorflow><machine-learning>","CreationDate":"2018-11-07T12:15:51.677","AcceptedAnswerId":null,"OwnerUserId":1546548.0,"Title":"How to retrain an Image Classifier with new classes and keep old classes","Body":"<p>I'm trying to make an image classifier that can identify how likely it is that an image is an image of a watermelon. To do this I followed the flower classifier example here: <a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retrainin\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retrainin<\/a> and trained the model using this command<\/p>\n\n<pre><code>python retrain.py --image_dir ~\/flower_photos\n<\/code><\/pre>\n\n<p>The problem I found when trying this classifier out is that it only classifies among the new classes, that is the flower classes in this case. So when I tried to classify an image of a dog (which I know is present in the Inception module) it classified it as a rose<\/p>\n\n<pre><code>python label_image.py \\\n--graph=\/tmp\/output_graph.pb --labels=\/tmp\/output_labels.txt \\\n--input_layer=Placeholder \\\n--output_layer=final_result \\\n--image=\/images\/dog.jpg\n<\/code><\/pre>\n\n<p>Result<\/p>\n\n<pre><code>roses 0.7626607\ntulips 0.12247563\ndandelion 0.071335025\nsunflowers 0.028395686\ndaisy 0.0151329385\n<\/code><\/pre>\n\n<p>How could I use TensorFlow to extend the model with an additional class instead of creating a new model with only the new classes?<\/p>\n","answers":[{"AnswerId":"53191986","CreationDate":"2018-11-07T14:56:10.783","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>What you can do is join the two datasets and train them together or just leave the classes of model you are retraining in the possible classes and add a few images of those classes to the dataset just for the model to not forget what it already learned. <\/p>\n"}]}
+{"QuestionId":53189379,"AnswerCount":2,"Tags":"<python><tensorflow><hashtable>","CreationDate":"2018-11-07T12:21:03.040","AcceptedAnswerId":null,"OwnerUserId":9555777.0,"Title":"No OpKernel was registered to support Op 'HashTableV2' with these attrs. Registered devices: [CPU,GPU], Registered kernels:","Body":"<p>I am coding with tensorflow 1.5.0, python 3.5. I want to create a hashtable. Since I intend to assign values to it later, I create it in the init function like this.(the values and shape are randomly given)\n<a href=\"https:\/\/i.stack.imgur.com\/Cis2c.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>but then I encounter a problem like this\n<a href=\"https:\/\/i.stack.imgur.com\/NVaAg.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>Can anyone help me?<\/p>\n","answers":[{"AnswerId":"53190030","CreationDate":"2018-11-07T13:04:24.750","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>It seems that the implementation of HashTable in your version of TensorFlow does not provide kernels for every possible combination of key and value types. There are two things you can do:<\/p>\n\n<ul>\n<li><p>According to your error message, there is a kernel implementation for 64-bit integer keys and 32-bit float values. So one possible fix is to simply change the data type of <code>keys<\/code> to <code>tf.int64<\/code>:<\/p>\n\n<pre><code>keys = tf.constant([1, 2, 3]), dtype=tf.int64)\n<\/code><\/pre><\/li>\n<li><p>Another possibility is to update TensorFlow to a version where this combination of key and value is implemented. It seems this was added in version <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/v1.11.0-rc0\" rel=\"nofollow noreferrer\">v1.11.0-rc0<\/a> (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/aaa33652fa8a74570708f9ef8ac355e91b399bfc\" rel=\"nofollow noreferrer\">see commmit<\/a>), so upgrading to that or a later version (in general it is more recommendable to upgrade to a stable version instead of a release candidate) should also fix the problem.<\/p><\/li>\n<\/ul>\n"},{"AnswerId":"53239060","CreationDate":"2018-11-10T12:41:26.510","ParentId":null,"OwnerUserId":"9762372","Title":null,"Body":"<p>The answers by @jdehesa is really great. It works for me!!! my tf version is 1.4, python=3.6<\/p>\n\n<p>Here is my code which works:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.contrib.lookup import *\n\nk = tf.range(1, 3, dtype=tf.int64)\nv = tf.range(5, 7, dtype=tf.int64)\ntable = tf.contrib.lookup.HashTable(\n    tf.contrib.lookup.KeyValueTensorInitializer(k, v, key_dtype=tf.int64, value_dtype=tf.int64), -1)\n\nout = table.lookup(tf.constant([2,1], dtype=tf.int64))\n\nwith tf.Session() as sess:\n    print(sess.run([k, v]))\n    table.init.run()\n    print(out.eval())\n<\/code><\/pre>\n"}]}
+{"QuestionId":53189435,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T12:24:32.803","AcceptedAnswerId":null,"OwnerUserId":10484396.0,"Title":"Incompatible shapes of test and train set in tensorflow","Body":"<p>I am trying to build a CNN network with tensorflow package. My train data sets are 237 images with 100X100 in 6 classes and my test sets are 30 images with same size and classes.\nEverything is going well except last part that I want to measure accuracy with<\/p>\n\n<pre><code>acc=tf.reduce_mean(tf.cast(matches,tf.float32))\n<\/code><\/pre>\n\n<p>I cannot understand why it make error while I am testing with test sets (both data and label) that it is compatible \u201c(30,100,100,1) vs (30,6)\u201d. \nCan anyone explain why and what is possible solution.<\/p>\n\n<p>Note: when I apply batch size 30 (equal to number my test set row) on train data it works.i assume the model expect that the train and test set should be equal number of row that it doesn't make set (usually 70% train vs 30% test)  i checked some post with same error title but i couldn\u2019t get it, in face, i think although the error title was same but their problem was different. mostly they had issue with flatten layer.<\/p>\n\n<p>screenshot from data set info\n<a href=\"https:\/\/i.stack.imgur.com\/g4uY6.png\" rel=\"nofollow noreferrer\">enter image description here<\/a> <\/p>\n\n<p>here is the simple code<\/p>\n\n<pre><code>\u2026\nconvo_1=\u2026(x_img,shape=[5,5,1,32])\nconvo_1_pooling=\u2026(convo_1)\n\nconvo_2=\u2026(x_img,shape=[5,5,32,64])\nconvo_2_pooling=\u2026(convo_2)\n\nconvo_3=\u2026(x_img,shape=[5,5,64,128])\nconvo_3_pooling=\u2026(convo_3)\n\nconvo_3_flat=tf.reshape(convo_3_pooling,[-1,13*13*128])\nfull_layer_one=tf.nn.relu(sp.normal_full_layer(convo_3_flat,21632))\n\n# DROPOUT\nhold_prob=tf.placeholder(tf.float32)\nfull_one_dropout=tf.nn.dropout(full_layer_one,keep_prob=hold_prob)\ny_pred=sp.normal_full_layer(full_one_dropout,6)\n\n# LOOS FUNCTION\ncross_entropy=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=y_pred))\n\n# OPTIMIZER\noptimizer=tf.train.AdamOptimizer(learning_rate=learning_rate)\ntrain=optimizer.minimize(cross_entropy)\n\nwith tf.Session() as sess:\n    sess.run(init)\n    b=0\n    for i in range(training_iters):\n        sess.run(train,feed_dict={x_img:train_x, y:train_y,hold_prob:.5})\n        if i%2 ==0:\n            print(\"NO STEP:{}\",format(i))\n            print(\"ACCURACY: \")\n            matches=tf.equal(tf.argmax(y_pred,1),tf.argmax(train_y,1))\n            acc=tf.reduce_mean(tf.cast(matches,tf.float32))\n            print(sess.run(acc,feed_dict={x_img:test_x, y:test_y, hold_prob:1.0}))\n            print('\\n')\n<\/code><\/pre>\n\n<p>and this is the error <\/p>\n\n<pre><code>InvalidArgumentError: Incompatible shapes: [30] vs. [237]\n     [[{{node Equal_1}} = Equal[T=DT_INT64, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](ArgMax_2, ArgMax_3)]]\n<\/code><\/pre>\n\n<p>thank you so much <\/p>\n","answers":[{"AnswerId":"57192258","CreationDate":"2019-07-24T22:57:15.943","ParentId":null,"OwnerUserId":"11812454","Title":null,"Body":"<p>It must be happening because while training your model you might have used <strong>train_x<\/strong> or <strong>train_y<\/strong> data instead of passing them through <strong>feed_dict,<\/strong> as a value for placeholder.<\/p>\n\n<p>A similar thing was happening with my code which got resolved after hours of debugging.<\/p>\n"}]}
+{"QuestionId":53189543,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2018-11-07T12:32:02.010","AcceptedAnswerId":null,"OwnerUserId":4481339.0,"Title":"AdamOptimizer returns invalid datatype error","Body":"<p>I have been struggling to fix this for the past couple of days, any help would be appreciated.<\/p>\n\n<p>I have a network that outputs numeric locations <code>x,y coordinates<\/code> in an image, hence the <code>int32<\/code> datatype. The output is of size <code>(2,2)<\/code>. Ground truth also has the locations as <code>int32<\/code>.<\/p>\n\n<p>Then I run, following two lines of code:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.square(y_true - y_pred_cls))\noptimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cost)\n<\/code><\/pre>\n\n<p>and receive following error:<\/p>\n\n<pre><code>ValueError: Invalid type tf.int32 for Mean_7:0, expected: [tf.float32, tf.float64, tf.float16, tf.bfloat16].\n<\/code><\/pre>\n\n<p>I understand that average should be a float value, but if I convert predicted value to float, I start getting 'No gradients provided' error. <\/p>\n\n<p>I have also tried to use <code>softmax_cross_entropy_with_logits_v2<\/code> of Tensorflow to turn my problem into a classification problem, but same error. I'd really appreciate any help with the solution. <\/p>\n\n<p><strong>Code<\/strong><\/p>\n\n<p>Code for Input and Labels<\/p>\n\n<pre><code>x_p = tf.placeholder(tf.float32, shape=[None, img_size_flat*num_channels], name='x_pred')\nx_pred_image = tf.reshape(x_p, [-1, 10, 10, num_channels])\nk = tf.constant(npatches)\ny_true = tf.placeholder(tf.int32, shape=[None, 2, 2], name='y_true')\ny_true_cls = tf.placeholder(tf.int32, shape=[None, 2, 2], name='y_true_cls')\n<\/code><\/pre>\n\n<p>Convolution Layer<\/p>\n\n<pre><code>def new_weights(shape, layer_name):\n    initializer = tf.contrib.layers.xavier_initializer()\n    return tf.Variable(initializer(shape), name=layer_name+'_W')\n\ndef new_bias(length, layer_name):\n    return tf.Variable(tf.constant(0.05, shape=[length]), name=layer_name+'_b')\n\ndef new_conv_layer(input,\n                   num_input_channels,\n                   filter_size,\n                   num_filters,\n                   name_scope,\n                   layer_name='',\n                   use_pooling=True):\n\n    with tf.name_scope(name_scope):\n        shape = [filter_size, filter_size, num_input_channels, num_filters]\n        weights = new_weights(shape, layer_name)\n        biases = new_bias(num_filters, layer_name)\n\n        layer = tf.add(tf.nn.conv2d(input=input, filter=weights, strides=[1,1,1,1], padding='SAME'), biases, name=layer_name)\n\n        if use_pooling:\n            layer = tf.nn.max_pool(value=layer,\n                                   ksize=[1, 3, 3, 1],\n                                   strides=[1, 2, 2, 1],\n                                   padding='SAME', name=layer_name+'_max')\n\n        layer = tf.nn.relu(layer, name=layer_name+'_activation')\n\n    return layer, weights\n<\/code><\/pre>\n\n<p>Fully Connected Layer<\/p>\n\n<pre><code>def new_fc_layer(input,\n                num_inputs,\n                num_outputs,\n                name_scope,\n                layer_name='',\n                use_relu=True):\n\n    with tf.name_scope(name_scope):\n        weights = new_weights([num_inputs, num_outputs], layer_name)\n        biases = new_bias(num_outputs, layer_name)\n\n        layer = tf.add(tf.matmul(input, weights),biases,name=layer_name)\n    #     layer = tf.matmul(input, weights) + biases\n\n        if use_relu:\n            layer = tf.nn.relu(layer, layer_name+'_activation')\n\n    return layer\n<\/code><\/pre>\n\n<p>Custom Layer Code <\/p>\n\n<pre><code>with tf.name_scope(name_scope):\n    resh_inp = tf.reshape(input, [-1, 10, 10])\n    input_shape = tf.shape(resh_inp)\n    rows, cols = input_shape[1], input_shape[2]\n    d_rows, d_cols = 2, 2\n    subm_rows, subm_cols = rows - d_rows + 1, cols - d_cols + 1\n    ii, jj = tf.meshgrid(tf.range(subm_rows), tf.range(subm_cols), indexing='ij')\n    d_ii, d_jj = tf.meshgrid(tf.range(d_rows), tf.range(d_cols), indexing='ij')\n\n    subm_ii = ii[:, :, tf.newaxis, tf.newaxis] + d_ii\n    subm_jj = jj[:, :, tf.newaxis, tf.newaxis] + d_jj\n    subm = tf.gather_nd(resh_inp[0, :, :], tf.stack([subm_ii, subm_jj], axis=-1), name=layer_name + \"_gather\")\n    subm_sum = tf.reduce_sum(subm, axis=(2, 3), name=layer_name + \"_subm_sum\")\n    _, top_idx = tf.nn.top_k(tf.reshape(subm_sum, [-1]), tf.minimum(k, tf.size(subm_sum)), name=layer_name + \"_top_idx\")\n    top_row = top_idx \/\/ subm_cols\n    top_col = top_idx % subm_cols\n    result = tf.stack([top_row, top_col], axis=-1, name=layer_name + \"_result\")\n    result_shape = tf.shape(result)\n    result = tf.reshape(result, [-1, result_shape[0], result_shape[1]], name=layer_name + \"_shaped_result\")\n<\/code><\/pre>\n\n<p>Result from above code is passed as <code>y_pred_cls<\/code> in below code<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.square(y_true - y_pred_cls))\noptimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cost)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53189712,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><image-processing><computer-vision>","CreationDate":"2018-11-07T12:44:10.343","AcceptedAnswerId":null,"OwnerUserId":7994456.0,"Title":"Generating binary segmentation file for lane detection in Python\/Tensorflow (Tusimple Lanenet Dataset)","Body":"<p>I was using the code from <a href=\"https:\/\/github.com\/MaybeShewill-CV\/lanenet-lane-detection\" rel=\"nofollow noreferrer\">https:\/\/github.com\/MaybeShewill-CV\/lanenet-lane-detection<\/a> which using Deep Learning to detect road lane lines.<\/p>\n\n<p>I successfully tested the model. Now I want to retrain the model on my own data.<\/p>\n\n<p>The training data consisted of three parts: the original image, the binary segmentation file and the instance segmentation file. Please refer to both <code>gt_image_binary<\/code> and <code>gt_image_instance<\/code> folders inside the <code>\/data\/training_data_example<\/code> folder from the repo.<\/p>\n\n<p>The binary segmentation use 255 to represent the lane field and 0 for the rest. The instance use different pixel value to represent different lane field and 0 for the rest.<\/p>\n\n<p>My question is how do I generate these two labels (binary and instance segmentation files)?<\/p>\n\n<p>The author said that you just need to follow the guidelines in the readme file for the Tusimple Lanenet Dataset found here: <a href=\"https:\/\/github.com\/TuSimple\/tusimple-benchmark\/blob\/master\/doc\/lane_detection\/readme.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/TuSimple\/tusimple-benchmark\/blob\/master\/doc\/lane_detection\/readme.md<\/a><\/p>\n\n<p>And from this, it states that you can generate these files using this format in json file:<\/p>\n\n<pre><code>{\n      'raw_file': str. 20th frame file path in a clip.\n      'lanes': list. A list of lanes. For each list of one lane, the elements are width values on image.\n      'h_samples': list. A list of height values corresponding to the 'lanes', which means len(h_samples) == len(lanes[i])\n    }\n<\/code><\/pre>\n\n<p>Where each json line in 'label_data_(date).json' is the label data for the frame.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>{\n  \"lanes\": [\n        [-2, -2, -2, -2, 632, 625, 617, 609, 601, 594, 586, 578, 570, 563, 555, 547, 539, 532, 524, 516, 508, 501, 493, 485, 477, 469, 462, 454, 446, 438, 431, 423, 415, 407, 400, 392, 384, 376, 369, 361, 353, 345, 338, 330, 322, 314, 307, 299],\n        [-2, -2, -2, -2, 719, 734, 748, 762, 777, 791, 805, 820, 834, 848, 863, 877, 891, 906, 920, 934, 949, 963, 978, 992, 1006, 1021, 1035, 1049, 1064, 1078, 1092, 1107, 1121, 1135, 1150, 1164, 1178, 1193, 1207, 1221, 1236, 1250, 1265, -2, -2, -2, -2, -2],\n        [-2, -2, -2, -2, -2, 532, 503, 474, 445, 416, 387, 358, 329, 300, 271, 241, 212, 183, 154, 125, 96, 67, 38, 9, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2],\n        [-2, -2, -2, 781, 822, 862, 903, 944, 984, 1025, 1066, 1107, 1147, 1188, 1229, 1269, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2]\n       ],\n  \"h_samples\": [240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, 400, 410, 420, 430, 440, 450, 460, 470, 480, 490, 500, 510, 520, 530, 540, 550, 560, 570, 580, 590, 600, 610, 620, 630, 640, 650, 660, 670, 680, 690, 700, 710],\n  \"raw_file\": \"path_to_clip\"\n}\n<\/code><\/pre>\n\n<p>Then you will just overlay the points on the original image (using cv2.polylines?).<\/p>\n\n<p>How do I do this?<\/p>\n\n<p>I tried to create both binary and instance segmentation files by just drawing a line (in paint, yes) on a black background. Then I got a shape error, so I converted them to grayscale. I also checked the original files which are also in grayscale so I followed this image format.<\/p>\n\n<p>But again, I got error doing this.<\/p>\n\n<p>What is the best way to generate the binary and instance segmentation files?<\/p>\n","answers":[{"AnswerId":"56125579","CreationDate":"2019-05-14T07:58:58.520","ParentId":null,"OwnerUserId":"8427549","Title":null,"Body":"<p>I will just drop an answer here (Don't know if it's good to answer an old question).<\/p>\n\n<pre><code>label = '0601'\nlabe_filename = 'label_data_%s.json' % label\nclips = [json.loads(line) for line in open(labe_filename).readlines()]\n\nfor clip in clips:\n    lanes = clip['lanes']\n    filepath = clip['raw_file']\n    ysamples = clip['h_samples']\n\n    lanes = [[(x, y) for (x, y) in zip(lane, ysamples) if x &gt;= 0] for lane in lanes]\n\n    raw_image = cv2.imread(filepath)\n    label_image = np.zeros(raw_image.shape[:2], dtype=np.uint8)\n\n    for lane in lanes:\n        cv2.polylines(label_image, np.int32([lane]), isClosed=False, color=(255,255,255), thickness=5)\n\n    cv2.imshow(\"win1\", raw_image)\n    cv2.imshow(\"win2\", label_image)\n    cv2.waitKey(0)\n    break\n<\/code><\/pre>\n\n<p>Tusimple dataset uses something like scanline to form markings for lanes (<a href=\"https:\/\/raw.githubusercontent.com\/TuSimple\/tusimple-benchmark\/master\/doc\/lane_detection\/assets\/examples\/lane_example.jpg\" rel=\"nofollow noreferrer\">this<\/a>). <code>h_samples<\/code> are the y coordinates of scanlines. So you need to combine <code>h_samples<\/code> and <code>lanes<\/code> (which are the x coordinates) to form a lane marking. And if there is no existing lane marking, <code>x<\/code> coordinate will be 0. That's what this line does:<\/p>\n\n<pre><code>lanes = [[(x, y) for (x, y) in zip(lane, ysamples) if x &gt;= 0] for lane in lanes]\n<\/code><\/pre>\n\n<p>For instance segmentation, use different color for each lane. You may refer to their data provider codes to address the shape errors.<\/p>\n"}]}
+{"QuestionId":53189730,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2018-11-07T12:46:11.577","AcceptedAnswerId":null,"OwnerUserId":10618247.0,"Title":"Using Keras masking layer with 2D convolutions (Conv2D)","Body":"<p>I'm trying to design a neural network including time dependent input with different lengths and I'm currently using a Masking layer.\nThis network worked well with TensorFlow version 1.9.0 but after updating to version 1.11.0, I get the following error:<\/p>\n\n<blockquote>\n  <p>Layer conv2d_1 does not support masking, but was passed an input_mask: Tensor(\"cnn1\/Reshape_2:0\", shape=(?, 81, 81), dtype=bool)<\/p>\n<\/blockquote>\n\n<p>Any idea on how to solve this problem?<\/p>\n\n<p>I'm using the following code:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D, \\\n    TimeDistributed, Dense, Masking, Activation, BatchNormalization\n\nmodel= Sequential()\n# first layer\nmodel.add(TimeDistributed(Masking(0., input_shape=(81,81,3)),\n                                  input_shape=(None,81,81,3), name='mask'))\n# CNN layers\nmodel.add(TimeDistributed(Conv2D(filters=10,\n                                 kernel_size=5,\n                                 strides=1,\n                                 padding='same'),\n                         name='cnn1'))\n\nmodel.add(Activation('relu', name='relu1'))\nmodel.add(BatchNormalization())\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2, 2))))\n\n# output layer\nmodel.add(TimeDistributed(Dense(3, name='output')))\nmodel.add(Activation('softmax'))\n\n# compilation\nmodel.compile(loss='categorical_crossentropy')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53189796,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2018-11-07T12:50:31.137","AcceptedAnswerId":null,"OwnerUserId":5771979.0,"Title":"In Keras, CNN layer result is different from the result of model.predict","Body":"<pre><code>history = model.fit(x_spectro_train, y_train_onehot, batch_size=batch_size, epochs=training_epochs, validation_data =(x_spectro_test, y_test_onehot), shuffle=True, callbacks=callbacks_list,class_weight=class_weights, verbose=1)\n\n\nmodel=load_model(model_name)\npredict_prob_train = model.predict(x_spectro_train,batch_size=batch_size) \n\n\ninp = model.input                                           # input placeholder\noutputs = [layer.output for layer in model.layers]          # all layer outputs\nfunctors = [K.function([inp, K.learning_phase()], [out]) for out in outputs]    # evaluation functions\nlayer_outs = [func([x_spectro_train, 0.]) for func in functors] #test mode (0.0), train mode(1.0)\n<\/code><\/pre>\n\n<p>I want to save CNN layer outputs. \nI want to train the svm model with CNN layer outputs (not probability)<\/p>\n\n<p>So I used code from <a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\/41712013#41712013\">Keras, How to get the output of each layer?<\/a> and I saw the result. <\/p>\n\n<p>But my result of CNN layer is different from the result of model.predict.\nI monitored the val accuracy, save the best model, and load it.\nThis is structure of my model. (below image)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zw71S.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>I expected that result of layer_outs[13] (last layer) is same with predict_prob_train. However, the results were different. (like below image)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/atpRo.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>Why the results are different? <\/p>\n","answers":[{"AnswerId":"53191804","CreationDate":"2018-11-07T14:47:12.270","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>You have 7 layers after <code>Conv<\/code> layer (2 of which are <code>Dense<\/code>). They also learn stuff and they are 'making the decision' of the model output. <\/p>\n\n<p>Think about it like this: <code>Conv<\/code> outputs something, that is the input to <code>Dense1<\/code> -> <code>Dense2<\/code>. All those layers are learning simultaneously. So the goal of <code>Dense1<\/code> layer is to learn what <code>Conv<\/code> layer is 'trying to tell it', how to interpret the results of <code>Conv<\/code> layer. If you input the image to this <code>Dense1<\/code> layer and then to <code>Dense2<\/code> layer, you won't get the same result (nor correct one). All of those layers are working together to get the correct prediction.<\/p>\n\n<p>You cannot isolate 1 layer and expect the correct result.<\/p>\n"}]}
+{"QuestionId":53189837,"AnswerCount":0,"Tags":"<python><tensorflow><keras><classification>","CreationDate":"2018-11-07T12:52:45.173","AcceptedAnswerId":null,"OwnerUserId":9511699.0,"Title":"Keras Classifier giving absurd accuracy for hand gesture classifcation","Body":"<p>I am trying to classify hand gestures with a Keras classifier model. But it is giving absurd accuracy while training, like on very first epoc it gives around 80 % a ccuracy and then jumps to 90 %  and final accuracy which comes after 25 epocs is 98 % . I created my own dataset for training which can be found <a href=\"https:\/\/drive.google.com\/drive\/folders\/1onT10LCM5cEkdYv_H7wv6Or0ErxA4nEd?usp=sharing\" rel=\"nofollow noreferrer\">here<\/a>. The 0 folder in dataset is for blank screen.<\/p>\n\n<pre><code># Importing the Keras libraries and packages\nfrom keras.models import Sequential\nfrom keras.layers import Convolution2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense , Dropout\nimport os\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"1\"\n\n# Step 1 - Building the CNN\n\n# Initializing the CNN\nclassifier = Sequential()\n\n# First convolution layer and pooling\nclassifier.add(Convolution2D(32, (3, 3), input_shape=(310, 310, 1), activation='relu'))\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))\n# Second convolution layer and pooling\nclassifier.add(Convolution2D(32, (3, 3), activation='relu'))\n# input_shape is going to be the pooled feature maps from the previous convolution layer\nclassifier.add(MaxPooling2D(pool_size=(2, 2)))\n\n# Flattening the layers\nclassifier.add(Flatten())\n\n# Adding a fully connected layer\nclassifier.add(Dense(units=128, activation='relu'))\nclassifier.add(Dropout(0.40))\nclassifier.add(Dense(units=96, activation='relu'))\nclassifier.add(Dropout(0.40))\nclassifier.add(Dense(units=64, activation='relu'))\nclassifier.add(Dense(units=27, activation='softmax')) # softmax for more than 2\n\n# Compiling the CNN\nclassifier.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy']) # categorical_crossentropy for more than 2\n\n\n# Step 2 - Preparing the train\/test data and training the model\n\n# Code copied from - https:\/\/keras.io\/preprocessing\/image\/\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        shear_range=0.2,\n        zoom_range=0.2,\n        horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntraining_set = train_datagen.flow_from_directory('data\/train',\n                                                 target_size=(310, 310),\n                                                 batch_size=10,\n                                                 color_mode='grayscale',\n                                                 class_mode='categorical')\n\ntest_set = test_datagen.flow_from_directory('data\/test',\n                                            target_size=(310 , 310),\n                                            batch_size=10,\n                                            color_mode='grayscale',\n                                            class_mode='categorical') \n\n\nclassifier.fit_generator(\n        training_set,\n        steps_per_epoch=12841,#9603, # No of images in training set\n        epochs=25,\n        validation_data=test_set,\n        validation_steps=4268)#2580)# No of images in test set\n\n\n# Saving the model\nmodel_json = classifier.to_json()\nwith open(\"model-bw.json\", \"w\") as json_file:\n    json_file.write(model_json)\nclassifier.save_weights('model-bw.h5')\n<\/code><\/pre>\n\n<p>When I actually test it using the following code it doesn't shows correct output, like for blan symbol it doesn't shows blank and for all other symbol its hows blank. <\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import model_from_json\nfrom image_processing import func\nimport operator\nimport time \nimport cv2\nimport sys, os\nimport matplotlib.pyplot as plt\n\n# Loading the model\nminValue = 70\njson_file = open(\"model-bw.json\", \"r\")\nmodel_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(model_json)\n# load weights into new model\n#loaded model = load_model('model-bw.h5')\nloaded_model.load_weights(\"model-bw.h5\")\nprint(\"Loaded model from disk\")\n\ncap = cv2.VideoCapture(0)\n\n# Category dictionary\ncategories = {0: 'ZERO', 1: 'ONE', 2: 'TWO', 3: 'THREE', 4: 'FOUR', 5: 'FIVE'}\n\nwhile True:\n    _, frame = cap.read()\n    # Simulating mirror image\n    frame = cv2.flip(frame, 1)\n    #frame = (cv2.imread(\"\/home\/rc\/Downloads\/soe\/train\/A\/001.jpg\"))\n    # Got this from collect-data.py\n    # Coordinates of the ROI\n    x1 = int(0.5*frame.shape[1])\n    y1 = 10\n    x2 = frame.shape[1]-10\n    y2 = int(0.5*frame.shape[1])\n    # x1 = 220\n    # x2 = 620\n    # y1 = 10\n    # y2 = 310\n    print(x1 , y1 , x2 , y2)\n    # Drawing the ROI\n    # The increment\/decrement by 1 is to compensate for the bounding box\n    cv2.rectangle(frame, (x1-1, y1-1), (x2+1, y2+1), (255,0,0) ,1)\n    # Extracting the ROI\n    roi = frame[y1:y2, x1:x2]\n\n    # Resizing the ROI so it can be fed to the model for prediction\n    #roi = cv2.resize(roi, (64, 64)) \n    #roi = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)\n    #_, test_image = cv2.threshold(roi, 120, 255, cv2.THRESH_BINARY)\n    gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)\n\n    blur = cv2.GaussianBlur(gray,(5,5),2)\n    # #blur = cv2.bilateralFilter(roi,9,75,75)\n\n    th3 = cv2.adaptiveThreshold(blur,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,cv2.THRESH_BINARY_INV,11,2)\n    ret, test_image = cv2.threshold(th3, minValue, 255, cv2.THRESH_BINARY_INV+cv2.THRESH_OTSU)\n    #time.sleep(5)\n    #cv2.imwrite(\"\/home\/rc\/Downloads\/soe\/im1.jpg\", roi)\n    #test_image = func(\"\/home\/rc\/Downloads\/soe\/im1.jpg\")\n\n\n\n    test_image = cv2.resize(test_image, (310,310))\n    cv2.imshow(\"test\", test_image)\n    # Batch of 1\n    # print(test_image.reshape(1, 64, 64, 1))\n    # break\n    result = loaded_model.predict(test_image.reshape(1, 310, 310, 1))\n    prediction = {\n                  'blank': result[0][0],\n                  'A': result[0][1], \n                  'B': result[0][2], \n                  'C': result[0][3],\n                  'D': result[0][4],\n                  'E': result[0][5],\n                  'F': result[0][6],\n                  'G': result[0][7],\n                  'H': result[0][8],\n                  'I': result[0][9],\n                  'J': result[0][10],\n                  'K': result[0][11],\n                  'L': result[0][12],\n                  'M': result[0][13],\n                  'N': result[0][14],\n                  'O': result[0][15],\n                  'P': result[0][16],\n                  'Q': result[0][17],\n                  'R': result[0][18],\n                  'S': result[0][19],\n                  'T': result[0][20],\n                  'U': result[0][21],\n                  'V': result[0][22],\n                  'W': result[0][23],\n                  'X': result[0][24],\n                  'Y': result[0][25],\n                  'Z': result[0][26],\n                }\n    # Sorting based on top prediction\n    prediction = sorted(prediction.items(), key=operator.itemgetter(1), reverse=True)\n\n    # Displaying the predictions\n    cv2.putText(frame, prediction[0][0], (10, 120), cv2.FONT_HERSHEY_PLAIN, 1, (0,255,255), 1)  \n    # print(prediction)\n    cv2.imshow(\"Frame\", frame)\n\n    interrupt = cv2.waitKey(10)\n    if interrupt &amp; 0xFF == 27: # esc key\n        break\n\n\ncap.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p>I trained with the same model with another dataset i have which doesn't have blank symbols and only has 24 letters. It used to get trained properly with accuracy in 1st epoch being started from 50% and it used to gradually increase. Input Size we kept in it was 300*300. But it could detect only 13-15 symbols correctly. Not more than that. So i made my own larger dataset(That dataset hadonly 4000 total images and mine has 17000) and kept size as 310*310 and its behaving in an absurd way.If anyone could point out the the mistake, it would be really helpfull<\/p>\n","answers":[]}
+{"QuestionId":53189994,"AnswerCount":0,"Tags":"<python><machine-learning><keras><conv-neural-network><lstm>","CreationDate":"2018-11-07T13:01:53.457","AcceptedAnswerId":null,"OwnerUserId":6938248.0,"Title":"Keras CNN-LSTM RuntimeError","Body":"<p>I'm trying to use the following model Inceptionv3 base CNN and LSTM layer for a regression problem. My input data is pictures with continuous target values. I'd like to feed the sequence of images to a CNN and after to an LSTM layer. However i get a <\/p>\n\n<pre><code>RuntimeError: You must compile your model before using it \n<\/code><\/pre>\n\n<p>message. Any idea what can be the cause? I tried to find out in github and on several pages but i didnt succeed.<\/p>\n\n<pre><code>from keras.applications.inception_v3 import InceptionV3\nfrom keras.models import Sequential, Model\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.layers import Dense, Activation, Flatten, Dropout, BatchNormalization, Conv2D, MaxPooling2D, GlobalAveragePooling2D, LSTM, TimeDistributed, Input\nfrom keras.optimizers import SGD, RMSprop\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint, CSVLogger, ReduceLROnPlateau\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport numpy as np\n\n\n# custom R2-score metrics for keras backend\n#https:\/\/www.kaggle.com\/c\/mercedes-benz-greener-manufacturing\/discussion\/34019\nfrom keras import backend as K\n\ndef r2_keras(y_true, y_pred):\n    SS_res =  K.sum(K.square(y_true - y_pred))\n    SS_tot = K.sum(K.square(y_true - K.mean(y_true)))\n    return ( 1 - SS_res\/(SS_tot + K.epsilon()) )  \n\ntrain_data_dir = '...'\ntest_data_dir = '...'\n\ntrain_df = pd.read_csv('...')\nvalid_df = pd.read_csv('...')\n\nfilepath_loss = '...'\nfilepath_csv = '...'\n\ndatagen=ImageDataGenerator(rescale=1.\/255.,)\n\n\nimg_width, img_height = 380, 380\nframes = 5\nchannels = 3\n\npictures = Input(shape=(frames, img_width, img_height, channels))\n\ntrain_generator=datagen.flow_from_dataframe(\ndataframe=train_df,\ndirectory=train_data_dir,\nx_col=\"block_heights\",\ny_col=\"weighted_prices\",\nhas_ext=False,      #x_col column doesnt has the file extensions\n#subset=\"training\",     if validation split is set in ImageDataGenerator\nbatch_size=16,\nseed=42,\nshuffle=False,\nclass_mode=\"other\",  #for regression other should be used\ntarget_size=(img_width, img_height))\n\nvalid_generator=datagen.flow_from_dataframe(\ndataframe=valid_df,\ndirectory=train_data_dir,\nx_col=\"block_heights\",\ny_col=\"weighted_prices\",\nhas_ext=False,              #x_col column doesnt has the file extensions\n#subset=\"validation\",      if validation split is set in ImageDataGenerator\nbatch_size=16,\nseed=42,\nshuffle=False,\nclass_mode=\"other\",\ntarget_size=(img_width, img_height))  \n\nconv_base = InceptionV3(weights=None, include_top=False, input_shape=(img_width,img_height,3))\nconv_base.trainable = True\n\nmodel = Sequential()\nmodel.add(TimeDistributed(conv_base))\nmodel.add(TimeDistributed(Flatten()))\nmodel.add(LSTM(10, return_sequences=True))\nmodel.add(Dense(512, activation='relu'))\nmodel.add(Dense(1, activation='linear'))\n\n#error at callbacks if the learning rate is explicitly set somewhere\nrms = RMSprop(lr=0.1, rho=0.9, epsilon=None, decay=0.0)\n\nmodel.compile(loss='mse', optimizer=rms, metrics=['mae', r2_keras])\n\nSTEP_SIZE_TRAIN=train_generator.n\/\/train_generator.batch_size\nSTEP_SIZE_VALID=valid_generator.n\/\/valid_generator.batch_size\n\ncallbacks = [EarlyStopping(monitor='val_loss', patience=10, verbose=1, mode='auto'),\n            ReduceLROnPlateau(monitor='val_loss', factor=0.02, patience=3, min_lr=0.001),\n            ModelCheckpoint(filepath_loss, monitor='val_loss', verbose=1, save_best_only=True, mode='min'),\n            CSVLogger(filepath_csv, separator = \",\", append = False)]\n\n\nhistory = model.fit_generator(generator=train_generator,steps_per_epoch=STEP_SIZE_TRAIN, validation_data=valid_generator, validation_steps=STEP_SIZE_VALID, epochs=50, callbacks=callbacks)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53190016,"AnswerCount":1,"Tags":"<machine-learning><keras><dataset><cross-validation><train-test-split>","CreationDate":"2018-11-07T13:03:28.960","AcceptedAnswerId":53190630.0,"OwnerUserId":1511274.0,"Title":"Difference between doing cross-validation and validation_data\/validation_split in Keras","Body":"<p>First, I split the dataset into train and test, for example:<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.4, random_state=999)\n<\/code><\/pre>\n\n<p>I then use <code>GridSearchCV<\/code> with cross-validation to find the best performing model:<\/p>\n\n<pre><code>validator  = GridSearchCV(estimator=clf, param_grid=param_grid, scoring=\"accuracy\", cv=cv)\n<\/code><\/pre>\n\n<p>And by doing this, I have:<\/p>\n\n<blockquote>\n  <p>A model is trained using k-1 of the folds as training data; the resulting\n  model is validated on the remaining part of the data (<a href=\"https:\/\/scikit-learn.org\/stable\/modules\/cross_validation.html\" rel=\"nofollow noreferrer\">scikit-learn.org<\/a>)<\/p>\n<\/blockquote>\n\n<p>But then, when reading about Keras <code>fit<\/code> fuction, the document introduces 2 more terms:<\/p>\n\n<blockquote>\n  <p>validation_split: Float between 0 and 1. Fraction of the training data\n  to be used as validation data. The model will set apart this fraction\n  of the training data, will not train on it, and will evaluate the loss\n  and any model metrics on this data at the end of each epoch. The\n  validation data is selected from the last samples in the x and y data\n  provided, before shuffling.<\/p>\n  \n  <p>validation_data: tuple (x_val, y_val) or tuple  (x_val, y_val,\n  val_sample_weights) on which to evaluate the loss and any model\n  metrics at the end of each epoch. The model will not be trained on\n  this data.  validation_data will override  validation_split.<\/p>\n<\/blockquote>\n\n<p>From what I understand, <code>validation_split<\/code> (to be overridden by <code>validation_data<\/code>) will be used as an <strong><em>unchanged validation dataset<\/em><\/strong>, meanwhile hold-out set in cross-validation changes during each cross-validation step.<\/p>\n\n<ul>\n<li>First question:  <em>is it necessary to use <code>validation_split<\/code> or <code>validation_data<\/code> since I already do cross validation?<\/em><\/li>\n<li><p>Second question: <em>if it is not necessary, then should I set <code>validation_split<\/code> and <code>validation_data<\/code> to 0 and None, respectively?<\/em><\/p>\n\n<pre><code>grid_result = validator.fit(train_images, train_labels, validation_data=None, validation_split=0)\n<\/code><\/pre><\/li>\n<li><p>Question 3: <em>If I do so, what will happen during the training, would Keras just simply ignore the validation step?<\/em><\/p><\/li>\n<li><p>Question 4: <em>Does the <code>validation_split<\/code> belong to <code>k-1 folds<\/code> or the <code>hold-out fold<\/code>, or will it be considered as \"test set\" (like in the case of <code>cross validation<\/code>) which will never be used to train the model.<\/em><\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"53190630","CreationDate":"2018-11-07T13:41:45.097","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Validation is performed to ensure that the model is not overfitting on the dataset and it would generalize to new data. Since in the parameters grid search you are also doing validation then there is no need to perform the validation step by the Keras model itself during training. Therefore to answer your questions:<\/p>\n\n<blockquote>\n  <p>is it necessary to use validation_split or validation_data since I already do cross validation?<\/p>\n<\/blockquote>\n\n<p>No, as I mentioned above.<\/p>\n\n<blockquote>\n  <p>if it is not necessary, then should I set validation_split and validation_data to 0 and None, respectively?<\/p>\n<\/blockquote>\n\n<p>No, since by default no validation is done in Keras (i.e. by default we have <code>validation_split=0.0, validation_data=None<\/code> in <a href=\"https:\/\/keras.io\/models\/sequential\/#fit\" rel=\"nofollow noreferrer\"><code>fit()<\/code><\/a> method).<\/p>\n\n<blockquote>\n  <p>If I do so, what will happen during the training, would Keras just simply ignore the validation step?<\/p>\n<\/blockquote>\n\n<p>Yes, Keras won't perform the validation when training the model. However note that, as I mentioned above, the grid search procedure would perform validation to better estimate the performance of the model with a specific set of parameters.<\/p>\n"}]}
+{"QuestionId":53190253,"AnswerCount":4,"Tags":"<python><tensorflow><keras><lstm><stateful>","CreationDate":"2018-11-07T13:17:00.730","AcceptedAnswerId":53303620.0,"OwnerUserId":1115237.0,"Title":"Stateful LSTM and stream predictions","Body":"<p>I've trained an LSTM model (built with Keras and TF) on multiple batches of 7 samples with 3 features each, with a shape the like below sample (numbers below are just placeholders for the purpose of explanation), each batch is labeled 0 or 1:<\/p>\n\n<p>Data:<\/p>\n\n<pre><code>[\n   [[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3]]\n   [[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3]]\n   [[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3],[1,2,3]]\n   ...\n]\n<\/code><\/pre>\n\n<p>i.e: batches of m sequences, each of length 7, whose elements are 3-dimensional vectors (so batch has shape (m*7*3))<\/p>\n\n<p>Target:<\/p>\n\n<pre><code>[\n   [1]\n   [0]\n   [1]\n   ...\n]\n<\/code><\/pre>\n\n<p>On my production environment data is a stream of samples with 3 features (<code>[1,2,3],[1,2,3]...<\/code>). I would like to stream each sample as it arrives to my model and get the intermediate probability without waiting for the entire batch (7) - see the animation below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4D0yt.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4D0yt.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>One of my thoughts was padding the batch with 0 for the missing samples, \n<code>[[0,0,0],[0,0,0],[0,0,0],[0,0,0],[0,0,0],[0,0,0],[1,2,3]]<\/code> but that seems to be inefficient.<\/p>\n\n<p>Will appreciate any help that will point me in the right direction of both saving the LSTM intermediate state in a persistent way, while waiting for the next sample and predicting on a model trained on a specific batch size with partial data.<\/p>\n\n<hr>\n\n<p><strong>Update,<\/strong> including model code:<\/p>\n\n<pre><code>opt = optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=10e-8, decay=0.001)\nmodel = Sequential()\n\nnum_features = data.shape[2]\nnum_samples = data.shape[1]\n\nfirst_lstm = LSTM(32, batch_input_shape=(None, num_samples, num_features), return_sequences=True, activation='tanh')\nmodel.add(\n    first_lstm)\nmodel.add(LeakyReLU())\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(16, return_sequences=True, activation='tanh'))\nmodel.add(Dropout(0.2))\nmodel.add(LeakyReLU())\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer=opt,\n              metrics=['accuracy', keras_metrics.precision(), keras_metrics.recall(), f1])\n<\/code><\/pre>\n\n<p><strong>Model Summary:<\/strong><\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlstm_1 (LSTM)                (None, 100, 32)           6272      \n_________________________________________________________________\nleaky_re_lu_1 (LeakyReLU)    (None, 100, 32)           0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 100, 32)           0         \n_________________________________________________________________\nlstm_2 (LSTM)                (None, 100, 16)           3136      \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 100, 16)           0         \n_________________________________________________________________\nleaky_re_lu_2 (LeakyReLU)    (None, 100, 16)           0         \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 1600)              0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 1)                 1601      \n=================================================================\nTotal params: 11,009\nTrainable params: 11,009\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n","answers":[{"AnswerId":"53284954","CreationDate":"2018-11-13T16:04:05.087","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><strong>Note: This answer assumes that your model in training phase is not stateful. You must understand what an stateful RNN layer is and make sure that the training data has the corresponding properties of statefulness. In short it means there is a dependency between the sequences, i.e. one sequence is the follow-up to another sequence, which you want to consider in your model. If your model and training data is stateful then I think other answers which involve setting <code>stateful=True<\/code> for the RNN layers from the beginning are simpler.<\/strong><\/p>\n\n<p><strong>Update: No matter the training model is stateful or not, you can always copy its weights to the inference model and enable statefulness. So I think solutions based on setting <code>stateful=True<\/code> are shorter and better than mine. Their only drawback is that the batch size in these solutions must be fixed.<\/strong><\/p>\n\n<hr>\n\n<p>Note that the output of a LSTM layer over a single sequence is determined by its weight matrices, which are fixed, and its internal states which depends on the <strong>previous processed timestep<\/strong>. Now to get the output of LSTM layer for a single sequence of length <code>m<\/code>, one obvious way is to feed the entire sequence to the LSTM layer in one go. However, as I stated earlier, since its internal states depends on the previous timestep, we can exploit this fact and feed that single sequence chunk by chunk by getting the state of LSTM layer at the end of processing a chunk and pass it to the LSTM layer for processing the next chunk. To make it more clear, suppose the sequence length is 7 (i.e. it has 7 timesteps of fixed-length feature vectors). As an example, it is possible to process this sequence like this:<\/p>\n\n<ol>\n<li>Feed the timesteps 1 and 2 to the LSTM layer; get the final state (call it <code>C1<\/code>).<\/li>\n<li>Feed the timesteps 3, 4 and 5 and state <code>C1<\/code> as the initial state to the LSTM layer; get the final state (call it <code>C2<\/code>).<\/li>\n<li>Feed the timesteps 6 and 7 and state <code>C2<\/code> as the initial state to the LSTM layer; get the final output.<\/li>\n<\/ol>\n\n<p>That final output is equivalent to the output produced by the LSTM layer if we had feed it the entire 7 timesteps at once.<\/p>\n\n<p>So to realize this in Keras, you can set the <code>return_state<\/code> argument of LSTM layer to <code>True<\/code> so that you can get the intermediate state. Further, don't specify a fixed timestep length when defining the input layer. Instead use <code>None<\/code> to be able to feed the model with sequences of arbitrary length which enables us to process each sequence progressively (it's fine if your input data in training time are sequences of fixed-length). <\/p>\n\n<p>Since you need this chuck processing capability in inference time, we need to define a new model which shares the LSTM layer used in training model and can get the initial states as input and also gives the resulting states as output. The following is a general sketch of it could be done (note that the returned state of LSTM layer is not used when training the model, we only need it in test time):<\/p>\n\n<pre><code># define training model\ntrain_input = Input(shape=(None, n_feats))   # note that the number of timesteps is None\nlstm_layer = LSTM(n_units, return_state=True)\nlstm_output, _, _ =  lstm_layer(train_input) # note that we ignore the returned states\nclassifier = Dense(1, activation='sigmoid')\ntrain_output = classifier(lstm_output)\n\ntrain_model = Model(train_input, train_output)\n\n# compile and fit the model on training data ...\n\n# ==================================================\n\n# define inference model\ninf_input = Input(shape=(None, n_feats))\nstate_h_input = Input(shape=(n_units,))\nstate_c_input = Input(shape=(n_units,))\n\n# we use the layers of previous model\nlstm_output, state_h, state_c = lstm_layer(inf_input,\n                                           initial_state=[state_h_input, state_c_input])\noutput = classifier(lstm_output)\n\ninf_model = Model([inf_input, state_h_input, state_c_input],\n                  [output, state_h, state_c])  # note that we return the states as output\n<\/code><\/pre>\n\n<p>Now you can feed the <code>inf_model<\/code> as much as the timesteps of a sequence are available right now. However, note that initially you must feed the states with vectors of all zeros (which is the default initial value of states). For example, if the sequence length is 7, a sketch of what happens when new data stream is available is as follows:<\/p>\n\n<pre><code>state_h = np.zeros((1, n_units,))\nstate_c = np.zeros((1, n_units))\n\n# three new timesteps are available\noutputs = inf_model.predict([timesteps, state_h, state_c])\n\nout = output[0,0]  # you may ignore this output since the entire sequence has not been processed yet\nstate_h = outputs[0,1]\nstate_c = outputs[0,2]\n\n# after some time another four new timesteps are available\noutputs = inf_model.predict([timesteps, state_h, state_c])\n\n# we have processed 7 timesteps, so the output is valid\nout = output[0,0]  # store it, pass it to another thread or do whatever you want to do with it\n\n# reinitialize the state to make them ready for the next sequence chunk\nstate_h = np.zeros((1, n_units))\nstate_c = np.zeros((1, n_units))\n\n# to be continued...\n<\/code><\/pre>\n\n<p>Of course you need to do this in some kind of loop or implement a control flow structure to process the data stream, but I think you get what the general idea looks like.<\/p>\n\n<p>Finally, although your specific example is not a sequence-to-sequence model, but I highly recommend to read the <a href=\"https:\/\/blog.keras.io\/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html\" rel=\"nofollow noreferrer\">official Keras seq2seq tutorial<\/a> which I think one can learn a lot of ideas from it.<\/p>\n"},{"AnswerId":"53193408","CreationDate":"2018-11-07T16:11:43.370","ParentId":null,"OwnerUserId":"4511978","Title":null,"Body":"<p>If I understood correctly, you have batches of <code>m<\/code> sequences, each of length 7, whose elements are 3-dimensional vectors (so batch has shape <code>(m*7*3)<\/code>).\nIn any <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">Keras RNN<\/a> you can set the \n<code>return_sequences<\/code> flag to <code>True<\/code> to become the intermediate states, i.e., for every batch, instead of the definitive prediction, you will get the corresponding 7 outputs, where output <code>i<\/code> represents the prediction at stage <code>i<\/code> given all inputs from 0 to <code>i<\/code>.<\/p>\n\n<p>But you would be getting all at once at the end. As far as I know, <strong>Keras doesn't provide a direct interface for retrieving the throughput whilst the batch is being processed<\/strong>. This may be even more constrained if you are using any of the <code>CUDNN<\/code>-optimized variants. What you can do is basically to <strong>regard your batch as 7 succesive batches of shape <code>(m*1*3)<\/code>, and feed them progressively to your LSTM<\/strong>, recording the hidden state and prediction at each step. For that, you can either set <code>return_state<\/code> to <code>True<\/code> and do it manually, or you can simply set <code>stateful<\/code>to <code>True<\/code> and let the object keep track of it.<\/p>\n\n<hr>\n\n<p>The following Python2+Keras example should exactly represent what you want. Specifically:<\/p>\n\n<ul>\n<li>allowing to save the whole LSTM intermediate state in a persistent way<\/li>\n<li>while waiting for the next sample<\/li>\n<li>and predicting on a model trained on a specific batch size that may be arbitrary and unknown.<\/li>\n<\/ul>\n\n<p>For that, it includes an example of <code>stateful=True<\/code> for easiest training, and <code>return_state=True<\/code> for most precise inference, so you get a flavor of both approaches. It also assumes that you get a model that has been serialized and from which you don't know much about. The structure is closely related to the one in Andrew Ng's course, who is definitely more authoritative than me in the topic. Since you don't specify how the model has been trained, I assumed a many-to-one training setup, but this could be easily adapted.<\/p>\n\n<pre><code>from __future__ import print_function\nfrom keras.layers import Input, LSTM, Dense\nfrom keras.models import Model, load_model\nfrom keras.optimizers import Adam\nimport numpy as np\n\n# globals\nSEQ_LEN = 7\nHID_DIMS = 32\nOUTPUT_DIMS = 3 # outputs are assumed to be scalars\n\n\n##############################################################################\n# define the model to be trained on a fixed batch size:\n# assume many-to-one training setup (otherwise set return_sequences=True)\nTRAIN_BATCH_SIZE = 20\n\nx_in = Input(batch_shape=[TRAIN_BATCH_SIZE, SEQ_LEN, 3])\nlstm = LSTM(HID_DIMS, activation=\"tanh\", return_sequences=False, stateful=True)\ndense = Dense(OUTPUT_DIMS, activation='linear')\nm_train = Model(inputs=x_in, outputs=dense(lstm(x_in)))\nm_train.summary()\n\n# a dummy batch of training data of shape (TRAIN_BATCH_SIZE, SEQ_LEN, 3), with targets of shape (TRAIN_BATCH_SIZE, 3):\nbatch123 = np.repeat([[1, 2, 3]], SEQ_LEN, axis=0).reshape(1, SEQ_LEN, 3).repeat(TRAIN_BATCH_SIZE, axis=0)\ntargets = np.repeat([[123,234,345]], TRAIN_BATCH_SIZE, axis=0) # dummy [[1,2,3],,,]-&gt; [123,234,345] mapping to be learned\n\n\n# train the model on a fixed batch size and save it\nprint(\"&gt;&gt; INFERECE BEFORE TRAINING MODEL:\", m_train.predict(batch123, batch_size=TRAIN_BATCH_SIZE, verbose=0))\nm_train.compile(optimizer=Adam(lr=0.5), loss='mean_squared_error', metrics=['mae'])\nm_train.fit(batch123, targets, epochs=100, batch_size=TRAIN_BATCH_SIZE)\nm_train.save(\"trained_lstm.h5\")\nprint(\"&gt;&gt; INFERECE AFTER TRAINING MODEL:\", m_train.predict(batch123, batch_size=TRAIN_BATCH_SIZE, verbose=0))\n\n\n##############################################################################\n# Now, although we aren't training anymore, we want to do step-wise predictions\n# that do alter the inner state of the model, and keep track of that.\n\n\nm_trained = load_model(\"trained_lstm.h5\")\nprint(\"&gt;&gt; INFERECE AFTER RELOADING TRAINED MODEL:\", m_trained.predict(batch123, batch_size=TRAIN_BATCH_SIZE, verbose=0))\n\n# now define an analogous model that allows a flexible batch size for inference:\nx_in = Input(shape=[SEQ_LEN, 3])\nh_in = Input(shape=[HID_DIMS])\nc_in = Input(shape=[HID_DIMS])\npred_lstm = LSTM(HID_DIMS, activation=\"tanh\", return_sequences=False, return_state=True, name=\"lstm_infer\")\nh, cc, c = pred_lstm(x_in, initial_state=[h_in, c_in])\nprediction = Dense(OUTPUT_DIMS, activation='linear', name=\"dense_infer\")(h)\nm_inference = Model(inputs=[x_in, h_in, c_in], outputs=[prediction, h,cc,c])\n\n#  Let's confirm that this model is able to load the trained parameters:\n# first, check that the performance from scratch is not good:\nprint(\"&gt;&gt; INFERENCE BEFORE SWAPPING MODEL:\")\npredictions, hs, zs, cs = m_inference.predict([batch123,\n                                               np.zeros((TRAIN_BATCH_SIZE, HID_DIMS)),\n                                               np.zeros((TRAIN_BATCH_SIZE, HID_DIMS))],\n                                              batch_size=1)\nprint(predictions)\n\n\n# import state from the trained model state and check that it works:\nprint(\"&gt;&gt; INFERENCE AFTER SWAPPING MODEL:\")\nfor layer in m_trained.layers:\n    if \"lstm\" in layer.name:\n        m_inference.get_layer(\"lstm_infer\").set_weights(layer.get_weights())\n    elif \"dense\" in layer.name:\n        m_inference.get_layer(\"dense_infer\").set_weights(layer.get_weights())\n\npredictions, _, _, _ = m_inference.predict([batch123,\n                                            np.zeros((TRAIN_BATCH_SIZE, HID_DIMS)),\n                                            np.zeros((TRAIN_BATCH_SIZE, HID_DIMS))],\n                                           batch_size=1)\nprint(predictions)\n\n\n# finally perform granular predictions while keeping the recurrent activations. Starting the sequence with zeros is a common practice, but depending on how you trained, you might have an &lt;END_OF_SEQUENCE&gt; character that you might want to propagate instead:\nh, c = np.zeros((TRAIN_BATCH_SIZE, HID_DIMS)), np.zeros((TRAIN_BATCH_SIZE, HID_DIMS))\nfor i in range(len(batch123)):\n    # about output shape: https:\/\/keras.io\/layers\/recurrent\/#rnn\n    # h,z,c hold the network's throughput: h is the proper LSTM output, c is the accumulator and cc is (probably) the candidate\n    current_input = batch123[i:i+1] # the length of this feed is arbitrary, doesn't have to be 1\n    pred, h, cc, c = m_inference.predict([current_input, h, c])\n    print(\"input:\", current_input)\n    print(\"output:\", pred)\n    print(h.shape, cc.shape, c.shape)\n    raw_input(\"do something with your prediction and hidden state and press any key to continue\")\n<\/code><\/pre>\n\n<hr>\n\n<h2>Additional information:<\/h2>\n\n<p>Since we have two forms of state persistency:<br>\n1. The saved\/trained parameters of the model that are the same for each sequence<br>\n2. The <code>a<\/code>, <code>c<\/code> states that evolve throughout the sequences and may be \"restarted\"<\/p>\n\n<p>It is interesting to take a look at the guts of the LSTM object. In the Python example that I provide, the <code>a<\/code> and <code>c<\/code> weights are explicitly handled, but the trained parameters aren't, and it may not be obvious how they are internally implemented or what do they mean. They can be inspected as follows:<\/p>\n\n<pre><code>for w in lstm.weights:\n    print(w.name, w.shape)\n<\/code><\/pre>\n\n<p>In our case (32 hidden states) returns the following:<\/p>\n\n<pre><code>lstm_1\/kernel:0 (3, 128)\nlstm_1\/recurrent_kernel:0 (32, 128)\nlstm_1\/bias:0 (128,)\n<\/code><\/pre>\n\n<p>We observe a dimensionality of 128. Why is that? <a href=\"https:\/\/adamtiger.github.io\/NNSharp\/recurrents\/#lstm\" rel=\"nofollow noreferrer\">this link<\/a> describes the Keras LSTM implementation as follows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lBLdV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lBLdV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<blockquote>\n  <p>The g is the recurrent activation, p is the activation, Ws are the kernels, Us are the recurrent kernels, h is the hidden variable which is the output too and the notation * is an element-wise multiplication.<\/p>\n<\/blockquote>\n\n<p>Which explains the <code>128=32*4<\/code> being the parameters for the affine transformation happening inside each one of the 4 gates, concatenated:<\/p>\n\n<ul>\n<li>The matrix of shape <code>(3, 128)<\/code> (named <code>kernel<\/code>) handles the input for a given sequence element  <\/li>\n<li>The matrix of shape <code>(32, 128)<\/code> (named <code>recurrent_kernel<\/code>) handles the input for the last recurrent state <code>h<\/code>.<\/li>\n<li>The vector of shape <code>(128,)<\/code> (named <code>bias<\/code>), as usual in any other NN setup.<\/li>\n<\/ul>\n"},{"AnswerId":"53303620","CreationDate":"2018-11-14T15:30:12.810","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>I think there might be an easier solution. <\/p>\n\n<p>If your model does not have convolutional layers or any other layers that act upon the length\/steps dimension, you can simply mark it as <code>stateful=True<\/code><\/p>\n\n<h2>Warning: your model has layers that act on the length dimension !!<\/h2>\n\n<p>The <code>Flatten<\/code> layer transforms the length dimension into a feature dimension. This will completely prevent you from achieving your goal. If the <code>Flatten<\/code> layer is expecting 7 steps, you will always need 7 steps. <\/p>\n\n<p>So, before applying my answer below, fix your model to not use the <code>Flatten<\/code> layer. Instead, it can just remove the <code>return_sequences=True<\/code> for the <strong>last<\/strong> LSTM layer. <\/p>\n\n<p>The following code fixed that and also prepares a few things to be used with the answer below: <\/p>\n\n<pre><code>def createModel(forTraining):\n\n    #model for training, stateful=False, any batch size   \n    if forTraining == True:\n        batchSize = None\n        stateful = False\n\n    #model for predicting, stateful=True, fixed batch size\n    else:\n        batchSize = 1\n        stateful = True\n\n    model = Sequential()\n\n    first_lstm = LSTM(32, \n        batch_input_shape=(batchSize, num_samples, num_features), \n        return_sequences=True, activation='tanh', \n        stateful=stateful)   \n\n    model.add(first_lstm)\n    model.add(LeakyReLU())\n    model.add(Dropout(0.2))\n\n    #this is the last LSTM layer, use return_sequences=False\n    model.add(LSTM(16, return_sequences=False, stateful=stateful,  activation='tanh'))\n\n    model.add(Dropout(0.2))\n    model.add(LeakyReLU())\n\n    #don't add a Flatten!!!\n    #model.add(Flatten())\n\n    model.add(Dense(1, activation='sigmoid'))\n\n    if forTraining == True:\n        compileThisModel(model)\n<\/code><\/pre>\n\n<p>With this, you will be able to train with 7 steps and predict with one step. Otherwise it will not be possible. <\/p>\n\n<h2>The usage of a stateful model as a solution for your question<\/h2>\n\n<p>First, train this new model again, because it has no Flatten layer:<\/p>\n\n<pre><code>trainingModel = createModel(forTraining=True)\ntrainThisModel(trainingModel)\n<\/code><\/pre>\n\n<p>Now, with this trained model, you can simply create a <strong>new model<\/strong> exactly the same way you created the trained model, but marking <code>stateful=True<\/code> in all its LSTM layers. And we should copy the weights from the trained model.<\/p>\n\n<p>Since these new layers will need a fixed batch size (Keras' rules), I assumed it would be 1 (one single stream is coming, not m streams) and added it to the model creation above.<\/p>\n\n<pre><code>predictingModel = createModel(forTraining=False)\npredictingModel.set_weights(trainingModel.get_weights())\n<\/code><\/pre>\n\n<p>And voil\u00e0. Just predict the outputs of the model with a single step:<\/p>\n\n<pre><code>pseudo for loop as samples arrive to your model:\n    prob = predictingModel.predict_on_batch(sample)\n\n    #where sample.shape == (1, 1, 3)\n<\/code><\/pre>\n\n<p>When you decide that you reached the end of what you consider a continuous sequence, call <code>predictingModel.reset_states()<\/code> so you can safely start a new sequence without the model thinking it should be mended at the end of the previous one.<\/p>\n\n<hr>\n\n<h2>Saving and loading states<\/h2>\n\n<p>Just get and set them, saving with h5py:<\/p>\n\n<pre><code>def saveStates(model, saveName):\n\n    f = h5py.File(saveName,'w')\n\n    for l, lay in enumerate(model.layers):\n        #if you have nested models, \n            #consider making this recurrent testing for layers in layers\n        if isinstance(lay,RNN):\n            for s, stat in enumerate(lay.states):\n                f.create_dataset('states_' + str(l) + '_' + str(s),\n                                 data=K.eval(stat), \n                                 dtype=K.dtype(stat))\n\n    f.close()\n\n\ndef loadStates(model, saveName):\n\n    f = h5py.File(saveName, 'r')\n    allStates = list(f.keys())\n\n    for stateKey in allStates:\n        name, layer, state = stateKey.split('_')\n        layer = int(layer)\n        state = int(state)\n\n        K.set_value(model.layers[layer].states[state], f.get(stateKey))\n\n    f.close()\n<\/code><\/pre>\n\n<h2>Working test for saving\/loading states<\/h2>\n\n<pre><code>import h5py, numpy as np\nfrom keras.layers import RNN, LSTM, Dense, Input\nfrom keras.models import Model\nimport keras.backend as K\n\n\n\n\ndef createModel():\n    inp = Input(batch_shape=(1,None,3))\n    out = LSTM(5,return_sequences=True, stateful=True)(inp)\n    out = LSTM(2, stateful=True)(out)\n    out = Dense(1)(out)\n    model = Model(inp,out)\n    return model\n\n\ndef saveStates(model, saveName):\n\n    f = h5py.File(saveName,'w')\n\n    for l, lay in enumerate(model.layers):\n        #if you have nested models, consider making this recurrent testing for layers in layers\n        if isinstance(lay,RNN):\n            for s, stat in enumerate(lay.states):\n                f.create_dataset('states_' + str(l) + '_' + str(s), data=K.eval(stat), dtype=K.dtype(stat))\n\n    f.close()\n\n\ndef loadStates(model, saveName):\n\n    f = h5py.File(saveName, 'r')\n    allStates = list(f.keys())\n\n    for stateKey in allStates:\n        name, layer, state = stateKey.split('_')\n        layer = int(layer)\n        state = int(state)\n\n        K.set_value(model.layers[layer].states[state], f.get(stateKey))\n\n    f.close()\n\ndef printStates(model):\n\n    for l in model.layers:\n        #if you have nested models, consider making this recurrent testing for layers in layers\n        if isinstance(l,RNN):\n            for s in l.states:\n                print(K.eval(s))   \n\nmodel1 = createModel()\nmodel2 = createModel()\nmodel1.predict_on_batch(np.ones((1,5,3))) #changes model 1 states\n\nprint('model1')\nprintStates(model1)\nprint('model2')\nprintStates(model2)\n\nsaveStates(model1,'testStates5')\nloadStates(model2,'testStates5')\n\nprint('model1')\nprintStates(model1)\nprint('model2')\nprintStates(model2)\n<\/code><\/pre>\n\n<h2>Considerations on the aspects of the data<\/h2>\n\n<p>In your first model (if it is <code>stateful=False<\/code>), it considers that each sequence in <code>m<\/code> is individual and not connected to the others. It also considers that each batch contains unique sequences. <\/p>\n\n<p>If this is not the case, you might want to train the stateful model instead (considering that each sequence is actually connected to the previous sequence). And then you would need <code>m<\/code> batches of 1 sequence. -> <code>m x (1, 7 or None, 3)<\/code>.<\/p>\n"},{"AnswerId":"53344603","CreationDate":"2018-11-16T20:00:41.613","ParentId":null,"OwnerUserId":"7476764","Title":null,"Body":"<p>As far as I know, because of the static graph in Tensorflow, there is no efficient way to feed inputs with different length from the training input length. <\/p>\n\n<p>Padding is the official way to work around with that, but it is less efficient and memory consuming. I suggest you look into Pytorch, which will be trivial to fix your problem. <\/p>\n\n<p>There are a lot of <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/time_sequence_prediction\/train.py\" rel=\"nofollow noreferrer\">great posts<\/a> to build lstm with Pytorch, and you will understand the benefit of dynamic graph once you see them. <\/p>\n"}]}
+{"QuestionId":53190752,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras><deep-learning>","CreationDate":"2018-11-07T13:47:28.597","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"Renaming Layers in Deep Learning Decreases the Accuracy Rate","Body":"<p>I have used Keras to implemente two deep learning models. The two models resulted high accuracy rate. Also, I created a multi input model which concatenate the two models together, but I got this error<\/p>\n\n<pre><code>RuntimeError: ('The name \"block1_conv1\" is used 2 times in the model. All layer names should be unique. \n<\/code><\/pre>\n\n<p>So, I have renamed the layers in the second model:<\/p>\n\n<pre><code>for layer in model2.layers:\n    layer.name = layer.name + str(\"2\")\n<\/code><\/pre>\n\n<p>After renaming the layers in the second model, the accuracy of the second model is decrease. When I remove the code of renaming, the second model produce a \nhigh accuracy. There is no difference in the structure of the second model before and after renaming the layers of the model. But I do not <strong>why the performance of the second model is changed after renaming the layers<\/strong>!.<\/p>\n\n<p>Also, I have tried to rename the first model only, and I faced the same problem. <\/p>\n\n<p><strong>Do you have any solution to solve this problem?<\/strong><\/p>\n\n<p><strong>Is there another way to change the layers names?<\/strong><\/p>\n\n<pre><code>model = vgg16.VGG16(include_top=False, input_shape=(64,64,3))\nin = model.input\nout = model.output\n\nmodel2 = vgg16.VGG16(include_top=False, input_shape=(64,64,3))\nin2 = model2.input\nout2 = model2.output\n\nfor layer in model2.layers:\n    layer.name = layer.name + str(\"2\")\n\nmerged = concatenate([out, out2])\nmerged = Dense(1024, activation='relu')(merged)\nmerged = Dense(num_classes, activation='softmax')(merged)\n\nmodel_fusion = Model([in, in], merged)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53190765,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T13:48:06.463","AcceptedAnswerId":null,"OwnerUserId":3861775.0,"Title":"Tensorflow: Undo global average pooling","Body":"<p>In Tensorflow I do at the end of my network the following global average pooling:<\/p>\n\n<pre><code>x_ = tf.reduce_mean(x, axis=[1,2])\n<\/code><\/pre>\n\n<p>My tensor <code>x<\/code> has the shape <code>(n, h, w, c)<\/code> where <code>n<\/code> is the number of inputs, <code>w<\/code> and <code>h<\/code> correspond to the width and height dimensions, and <code>c<\/code> is the number of channels\/filters.<\/p>\n\n<p>Starting with a tensor <code>x<\/code> of size <code>(n, h, w, c)<\/code> after calling <code>tf.reduce_mean()<\/code> the resulting tensor is of size <code>(n, c)<\/code>.<\/p>\n\n<p>How can I reverse that process? How can I do the unpooling operation?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Here is an example which did not work as expected:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nn, c = 1, 2 \nh, w = 2, 2\n\nx = tf.ones([n, h, w, c])\ny = tf.reduce_mean(x, axis=[1,2], keepdims=True)\nz = tf.reshape(y, [n, 1, 1, c])\nu = tf.tile(z, [n, h, w, c])\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(x)\n    print(\"x\", sess.run(x))\n    print(\"\\n\")\n    print(y)\n    print(\"y\", sess.run(y))\n    print(\"\\n\")\n    print(z)\n    print(\"z\", sess.run(z))\n    print(\"\\n\")\n    print(u)\n    print(\"u\", sess.run(u))\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>Tensor(\"ones:0\", shape=(1, 2, 2, 2), dtype=float32)\nx [[[[1. 1.]\n   [1. 1.]]\n\n  [[1. 1.]\n   [1. 1.]]]]\n\n\nTensor(\"Mean:0\", shape=(1, 1, 1, 2), dtype=float32)\ny [[[[1. 1.]]]]\n\n\nTensor(\"Reshape:0\", shape=(1, 1, 1, 2), dtype=float32)\nz [[[[1. 1.]]]]\n\n\nTensor(\"Tile:0\", shape=(1, 2, 2, 4), dtype=float32)\nu [[[[1. 1. 1. 1.]\n   [1. 1. 1. 1.]]\n\n  [[1. 1. 1. 1.]\n   [1. 1. 1. 1.]]]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53198212","CreationDate":"2018-11-07T21:37:55.440","ParentId":null,"OwnerUserId":"8384504","Title":null,"Body":"<p>You can do the unpooling operation with <code>tf.reshape<\/code> and <code>tf.tile<\/code>.<\/p>\n\n<pre><code>x = tf.random_uniform([n, c])\ny = tf.reshape(x, [n, 1, 1, c])\nz = tf.tile(y, [1, h, w, 1])\n<\/code><\/pre>\n\n<p>Starting with a tensor <code>x<\/code> of size <code>(n, c)<\/code> after calling <code>tf.reshape<\/code> and <code>tf.tile<\/code> the resulting tensor <code>z<\/code> is of size <code>(n, h, w, c)<\/code>.<\/p>\n"}]}
+{"QuestionId":53191114,"AnswerCount":1,"Tags":"<python><numpy><keras><imdb>","CreationDate":"2018-11-07T14:09:14.303","AcceptedAnswerId":null,"OwnerUserId":3612693.0,"Title":"Keras with tensorflow backend---MemoryError","Body":"<p>I am trying to follow <a href=\"https:\/\/www.kaggle.com\/drscarlat\/imdb-sentiment-analysis-keras-and-tensorflow\" rel=\"nofollow noreferrer\">this<\/a> tutorial to learn a bit about deep learning with keras, however I keep getting MemoryError. Can you please point out what is causing it and how to take care of it?<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import numpy as np\nfrom keras import models, regularizers, layers\nfrom keras.datasets import imdb\n\n(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)\n\ndef vectorize_sequences(sequences, dimension=10000):\n    results = np.zeros((len(sequences), dimension))\n    for i, sequence in enumerate(sequences):\n        results[i, sequence] = 1.\n    return results\n\n\nx_train = vectorize_sequences(train_data)\n<\/code><\/pre>\n\n<p>Here is the traceback (line number doesn't match the line number from the code mentioned above)<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/uttam\/pycharm-2018.2.4\/helpers\/pydev\/_pydev_bundle\/pydev_umd.py\", line 197, in runfile\n    pydev_imports.execfile(filename, global_vars, local_vars)  # execute the script\n  File \"\/home\/uttam\/pycharm-2018.2.4\/helpers\/pydev\/_pydev_imps\/_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"\/home\/uttam\/PycharmProjects\/IMDB\/imdb.py\", line 33, in &lt;module&gt;\n    x_train = vectorize_sequences(train_data)\n  File \"\/home\/uttam\/PycharmProjects\/IMDB\/imdb.py\", line 27, in vectorize_sequences\n    results = np.zeros((len(sequences), dimension))\nMemoryError\n<\/code><\/pre>\n","answers":[{"AnswerId":"53191465","CreationDate":"2018-11-07T14:29:12.167","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Yes, you are correct. The problem does arise from <code>vectorize_sequences<\/code>.<\/p>\n\n<p>You should do that logic in batches (with slicing data like for <code>partial_x_train<\/code>) or use generators (<a href=\"https:\/\/www.programiz.com\/python-programming\/generator\" rel=\"nofollow noreferrer\">here<\/a> is a good explanation and example).<\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
+{"QuestionId":53191168,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><mouseevent><object-detection-api>","CreationDate":"2018-11-07T14:12:23.717","AcceptedAnswerId":53204672.0,"OwnerUserId":9709132.0,"Title":"List index out of range error when using mouse events functions with tensorflow for object detection","Body":"<p>I am using Tensorflow for object detection. I successfully trained the neural network and it can detect the object I want to detect in the livestream. It does this by making a bounding box around the object. <\/p>\n\n<p>Now I want to mark areas in the video frame(initially) such that if the object comes in the marked area and is detected,(i.e. if a bounding box is made in the marked area) then I want to print a message in the terminal. <\/p>\n\n<p>For this purpose, I am using OpenCV. I found a nice tutorial on how to use Mouse callback functions to do this. The link is given below.<\/p>\n\n<p><a href=\"https:\/\/www.pyimagesearch.com\/2015\/03\/09\/capturing-mouse-click-events-with-python-and-opencv\/\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2015\/03\/09\/capturing-mouse-click-events-with-python-and-opencv\/<\/a><\/p>\n\n<p>But there is an error when I execute my code. The error is shown below.<\/p>\n\n<pre><code> ---------------------------------------------------------------------------\n IndexError                                Traceback (most recent call last)\n&lt;ipython-input-1-10159c26292b&gt; in click_and_crop(event, x, y, flags, params)\n    201           refPt.append((x,y))\n    202           cropping = False\n--&gt; 203           cv2.rectangle(image_np,refPt[0],refPt[1],(0,255,0),2)\n    204       ret, image_np = cap.read()\n    205       # Expand dimensions since the model expects images to have \nshape: [1, None, None, 3]\n\nIndexError: list index out of range\n<\/code><\/pre>\n\n<p>My main program is as follows:<\/p>\n\n<pre><code>import numpy as np\nimport os\nimport six.moves.urllib as urllib\nimport sys\nimport tarfile\nimport tensorflow as tf\nimport zipfile\n\nfrom collections import defaultdict\nfrom io import StringIO\nfrom PIL import Image\n\nimport cv2\ncap = cv2.VideoCapture(0)\n\n\n# This is needed since the notebook is stored in the object_detection \nfolder.\nsys.path.append(\"..\")\n\nfrom object_detection.utils import ops as utils_ops\n\nif tf.__version__ &lt; '1.4.0' and tf.__version__ != '1.10.0':\nraise ImportError('Please upgrade your tensorflow installation to v1.4.* or \nlater!')\n\n\n# ## Env setup\n\n# In[3]:\n\n\n# This is needed to display the images.\n#get_ipython().run_line_magic('matplotlib', 'inline')\n\n\n# ## Object detection imports\n# Here are the imports from the object detection module.\n\n# In[5]:\n\n\nfrom utils import label_map_util\n\nfrom utils import visualization_utils as vis_util\n\n\n# # Model preparation \n\n# ## Variables\n# \n# Any model exported using the `export_inference_graph.py` tool can be \nloaded here simply by changing `PATH_TO_FROZEN_GRAPH` to point to a new .pb \nfile.  \n# \n\n\n# In[6]:\n\n\n# What model to download.\nMODEL_NAME = 'car_inference_graph'\n\n# Path to frozen detection graph. This is the actual model that is used for \nthe object detection.\nPATH_TO_CKPT = MODEL_NAME + '\/frozen_inference_graph.pb'\n\n# List of the strings that is used to add correct label for each box.\nPATH_TO_LABELS = os.path.join('training', 'object-detection.pbtxt')\n\nNUM_CLASSES = 1\n\n\n# ## Download Model\n\n# ## Load a (frozen) Tensorflow model into memory.\n\n# In[7]:\n\n\ndetection_graph = tf.Graph()\nwith detection_graph.as_default():\n  od_graph_def = tf.GraphDef()\n  with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:\n  serialized_graph = fid.read()\n  od_graph_def.ParseFromString(serialized_graph)\n  tf.import_graph_def(od_graph_def, name='')\n\n\n# ## Loading label map\n# Label maps map indices to category names, so that when our convolution \nnetwork predicts `5`, we know that this corresponds to `airplane`.  Here we \nuse internal utility functions, but anything that returns a dictionary \nmapping integers to appropriate string labels would be fine\n\n# In[8]:\n\n\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\ncategories = label_map_util.convert_label_map_to_categories(label_map, \nmax_num_classes=NUM_CLASSES, use_display_name=True)\ncategory_index = label_map_util.create_category_index(categories)\n\n\n# ## Helper code\n\n# In[9]:\n\n\ndef load_image_into_numpy_array(image):\n  (im_width, im_height) = image.size\n  return np.array(image.getdata()).reshape(\n  (im_height, im_width, 3)).astype(np.uint8)\n\n\n# # Detection\n\n# In[10]:\n\n\n# For the sake of simplicity we will use only 2 images:\n# image1.jpg\n# image2.jpg\n# If you want to test the code with your images, just add path to the images \nto the TEST_IMAGE_PATHS.\nPATH_TO_TEST_IMAGES_DIR = 'test_images'\nTEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, \n'image{}.jpg'.format(i)) for i in range(1,44) ]\n\n# Size, in inches, of the output images.\nIMAGE_SIZE = (12, 8)\n\n\n# In[11]:\n\n\ndef run_inference_for_single_image(image, graph):\n  with graph.as_default():\n  with tf.Session() as sess:\n    # Get handles to input and output tensors\n    ops = tf.get_default_graph().get_operations()\n    all_tensor_names = {output.name for op in ops for output in op.outputs}\n    tensor_dict = {}\n    for key in [\n      'num_detections', 'detection_boxes', 'detection_scores',\n      'detection_classes', 'detection_masks'\n  ]:\n    tensor_name = key + ':0'\n    if tensor_name in all_tensor_names:\n      tensor_dict[key] = tf.get_default_graph().get_tensor_by_name(\n          tensor_name)\n    if 'detection_masks' in tensor_dict:\n      # The following processing is only for single image\n      detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])\n      detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])\n      # Reframe is required to translate mask from box coordinates to image \n  coordinates and fit the image size.\n      real_num_detection = tf.cast(tensor_dict['num_detections'][0], \n  tf.int32)\n      detection_boxes = tf.slice(detection_boxes, [0, 0], \n  [real_num_detection, -1])\n      detection_masks = tf.slice(detection_masks, [0, 0, 0], \n  [real_num_detection, -1, -1])\n      detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(\n          detection_masks, detection_boxes, image.shape[0], image.shape[1])\n      detection_masks_reframed = tf.cast(\n          tf.greater(detection_masks_reframed, 0.5), tf.uint8)\n      # Follow the convention by adding back the batch dimension\n      tensor_dict['detection_masks'] = tf.expand_dims(\n          detection_masks_reframed, 0)\n    image_tensor = \n  tf.get_default_graph().get_tensor_by_name('image_tensor:0')\n\n    # Run inference\n    output_dict = sess.run(tensor_dict,\n                         feed_dict={image_tensor: np.expand_dims(image, 0)})\n\n    # all outputs are float32 numpy arrays, so convert types as appropriate\n    output_dict['num_detections'] = int(output_dict['num_detections'][0])\n    output_dict['detection_classes'] = output_dict[\n        'detection_classes'][0].astype(np.uint8)\n    output_dict['detection_boxes'] = output_dict['detection_boxes'][0]\n    output_dict['detection_scores'] = output_dict['detection_scores'][0]\n    if 'detection_masks' in output_dict:\n      output_dict['detection_masks'] = output_dict['detection_masks'][0]\nreturn output_dict\n\n\n# In[12]:\n\n\n# In[10]:\n\nwith detection_graph.as_default():\n  with tf.Session(graph=detection_graph) as sess:\n    while True:\n      refPt = [] #ROI code starts from here\n      cropping = False\n\n    def click_and_crop(event,x,y,flags,params):\n      global refPt,cropping\n\n      if event == cv2.EVENT_LBUTTONDOWN:\n        refPt = [(x,y)]\n        cropping = True\n\n      elif event == cv2.EVENT_LBUTTONUP:\n        refPt.append((x,y))\n        cropping = False\n        cv2.rectangle(image_np,refPt[0],refPt[1],(0,255,0),2) # ROI code end\n    ret, image_np = cap.read()\n    # Expand dimensions since the model expects images to have shape: [1, \n  None, None, 3]\n    image_np_expanded = np.expand_dims(image_np, axis=0)\n    image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n    # Each box represents a part of the image where a particular object was \n  detected.\n    boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n    # Each score represent how level of confidence for each of the objects.\n    # Score is shown on the result image, together with the class label.\n    scores = detection_graph.get_tensor_by_name('detection_scores:0')\n    classes = detection_graph.get_tensor_by_name('detection_classes:0')\n    num_detections = detection_graph.get_tensor_by_name('num_detections:0')\n    # Actual detection.\n    (boxes, scores, classes, num_detections) = sess.run(\n        [boxes, scores, classes, num_detections],\n        feed_dict={image_tensor: image_np_expanded})\n    # Visualization of the results of a detection.\n    vis_util.visualize_boxes_and_labels_on_image_array(\n        image_np,\n        np.squeeze(boxes),\n        np.squeeze(classes).astype(np.int32),\n        np.squeeze(scores),\n        category_index,\n        use_normalized_coordinates=True,\n        line_thickness=8)\n\n    cv2.imshow(\"object detection\", image_np)\n    cv2.setMouseCallback(\"object detection\", click_and_crop)\n\n    if cv2.waitKey(25) &amp; 0xFF == ord('q'):\n      cv2.destroyAllWindows()\n      cap.release()\n      break\n<\/code><\/pre>\n\n<p>Using this code:<\/p>\n\n<ul>\n<li>I can see the livestream properly.<\/li>\n<li>The desired object is successfully detected.<\/li>\n<li>But as soon as I drag the left mouse button to draw ROI in the frame, I get \nthe error mentioned above.<\/li>\n<\/ul>\n\n<p>I understand this has something to do with <code>refPt[0],refPt[1]<\/code> but I don't understand where to make the necessary changes!<\/p>\n\n<p>Technical information:<\/p>\n\n<ol>\n<li>Tensorflow 1.10<\/li>\n<li>OS - Ubuntu 18.04<\/li>\n<li>Python 3.6<\/li>\n<li>OpenCV 3.4.2<\/li>\n<\/ol>\n\n<p>Please help.<\/p>\n\n<p>Thanks :)<\/p>\n","answers":[{"AnswerId":"53204672","CreationDate":"2018-11-08T09:18:36.747","ParentId":null,"OwnerUserId":"888688","Title":null,"Body":"<p>You have a couple of problems in your code, but basically the problem can be summarized in the scope of the variables you are using.<\/p>\n\n<ol>\n<li><p>avoid creating a function in a loop... this will redefine a function every time... not a problem here, but it is better not to do it. <\/p><\/li>\n<li><p>You have a <code>refPt = []<\/code> inside the while, this will empty the array at every iteration... as in case 1, it should be outside. Anyways you have in the function <code>refPt = [(x,y)]<\/code> that will remove old values and \"clean\" the variable.<\/p><\/li>\n<li><p>Inside the function you have <code>cv2.rectangle(image_np,refPt[0],refPt[1],(0,255,0),2)<\/code> which changes <code>image_np<\/code>, but this image gets changed locally and not globally. <\/p><\/li>\n<li><p>In the loop you have <code>ret, image_np = cap.read()<\/code> which will remove any rectangle almost immediately without being displayed.... you need to draw the rectangle on the new images. Something like:<\/p><\/li>\n<\/ol>\n\n<hr>\n\n<pre><code> ret, image_np = cap.read()\n # if no image was obtained quit the loop\n if !ret:\n   break\n tmpPt = refPt.copy() # to avoid it being changed in the callback\n if len(tmpPt ) ==2:\n   cv2.rectangle(image_np,tmpPt [0],tmpPt [1],(0,255,0),2)\n<\/code><\/pre>\n\n<hr>\n\n<ol start=\"5\">\n<li>It is recommended to use <code>cv2.setMouseCallback(\"object detection\", click_and_crop)<\/code> outside the loop... You can use <code>cv2.namedWindow(\"object detection\")<\/code> to create the window without having the image.<\/li>\n<\/ol>\n\n<p>These are the problems I see, maybe you encounter more once these are corrected...One more thing, you are just drawing a rectangle, but I do not see that you are actually using it to select a roi (cropping the image to the rectangle size), I do not know if this is intended...<\/p>\n\n<p>I hope this helps you, and if you have a question, just ask in a comment.<\/p>\n\n<hr>\n\n<h3>UPDATE<\/h3>\n\n<p>To make myself a little bit clear and to add the selection first and then the detection part, the code should look like this:<\/p>\n\n<pre><code>refPt = [] \ncropping = False\n\ndef click_and_crop(event,x,y,flags,params):\n  global refPt,cropping\n\n  if event == cv2.EVENT_LBUTTONDOWN:\n    refPt = [(x,y)]\n    cropping = True\n\n  elif event == cv2.EVENT_LBUTTONUP:\n    refPt.append((x,y))\n    cropping = False\n\ncv2.namedWindow(\"object detection\")\ncv2.setMouseCallback(\"object detection\", click_and_crop)\n\ndetect = False\nwith detection_graph.as_default():\n  with tf.Session(graph=detection_graph) as sess:\n    while True:\n     ret, image_np = cap.read()\n     # if no image was obtained quit the loop\n     if !ret:\n       break\n     tmpPt = refPt.copy() # to avoid it being changed in the callback\n     if len(tmpPt ) ==2:\n       cv2.rectangle(image_np,tmpPt [0],tmpPt [1],(0,255,0),2)\n\n    if detect:\n      # Expand dimensions since the model expects images to have shape: [1, None, None, 3]\n      image_np_expanded = np.expand_dims(image_np, axis=0)\n      image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n      # Each box represents a part of the image where a particular object was detected.\n      boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n      # Each score represent how level of confidence for each of the objects.\n      # Score is shown on the result image, together with the class label.\n      scores = detection_graph.get_tensor_by_name('detection_scores:0')\n      classes = detection_graph.get_tensor_by_name('detection_classes:0')\n      num_detections = detection_graph.get_tensor_by_name('num_detections:0')\n      # Actual detection.\n      (boxes, scores, classes, num_detections) = sess.run(\n          [boxes, scores, classes, num_detections],\n          feed_dict={image_tensor: image_np_expanded})\n      # Visualization of the results of a detection.\n      vis_util.visualize_boxes_and_labels_on_image_array(\n          image_np,\n          np.squeeze(boxes),\n          np.squeeze(classes).astype(np.int32),\n          np.squeeze(scores),\n          category_index,\n          use_normalized_coordinates=True,\n          line_thickness=8)\n\n    cv2.imshow(\"object detection\", image_np)\n\n    key = cv2.waitKey(25) &amp; 0xFF \n    if key == ord('q'):\n      cv2.destroyAllWindows()\n      cap.release()\n      break\n    elif key == ord('s'):\n      detect = True # start detecting\n<\/code><\/pre>\n\n<p>Once again this will only draw the rectangle... it does not crop<\/p>\n"}]}
+{"QuestionId":53191408,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><autoencoder>","CreationDate":"2018-11-07T14:25:59.730","AcceptedAnswerId":null,"OwnerUserId":6104742.0,"Title":"Always same output for tensorflow autoencoder","Body":"<p>At the moment I try to build an Autoencoder for timeseries data in tensorflow. I have nearly 500 days of data where each day have 24 datapoints. Since this is my first try my architecture is very simple. After my input of size <code>24<\/code> the hidden layers are of size: <code>10; 3; 10<\/code> with an output of again <code>24<\/code>. I normalized the data (datapoints are in range <code>[-0.5; 0.5]<\/code>), use the sigmoid activation function and the RMSPropOptimizer.<\/p>\n\n<p>After training (loss function in picture) the output is the same for every timedata i give into the network. Does someone know what is the reason for that? Is it possible that my Dataset is the issue (code below)?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0KcbD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0KcbD.png\" alt=\"loss function of training\"><\/a><\/p>\n\n<pre><code>class TimeDataset:\ndef __init__(self,data):\n    self._index_in_epoch = 0\n    self._epochs_completed = 0\n    self._data = data\n    self._num_examples = data.shape[0]\n    pass\n\n\n@property\ndef data(self):\n    return self._data\n\ndef next_batch(self, batch_size, shuffle=True):\n    start = self._index_in_epoch\n\n    # first call\n    if start == 0 and self._epochs_completed == 0:\n        idx = np.arange(0, self._num_examples)  # get all possible indexes\n        np.random.shuffle(idx)  # shuffle indexe\n        self._data = self.data[idx]  # get list of `num` random samples\n\n    if start + batch_size &gt; self._num_examples:\n        # not enough samples left -&gt; go to the next batch\n        self._epochs_completed += 1\n        rest_num_examples = self._num_examples - start\n        data_rest_part = self.data[start:self._num_examples]\n        idx0 = np.arange(0, self._num_examples)  # get all possible indexes\n        np.random.shuffle(idx0)  # shuffle indexes\n        self._data = self.data[idx0]  # get list of `num` random samples\n\n        start = 0\n        self._index_in_epoch = batch_size - rest_num_examples #avoid the case where the #sample != integar times of batch_size\n        end =  self._index_in_epoch  \n        data_new_part =  self._data[start:end]  \n        return np.concatenate((data_rest_part, data_new_part), axis=0)\n    else:\n        # get next batch\n        self._index_in_epoch += batch_size\n        end = self._index_in_epoch\n        return self._data[start:end]\n<\/code><\/pre>\n\n<p>*edit: here are some examples of the output (red original, blue reconstructed):\n<a href=\"https:\/\/i.stack.imgur.com\/XtUyB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XtUyB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>**edit: I just saw an autoencoder example with a more complicant luss function than mine. Someone know if the loss function <code>self.loss = tf.reduce_mean(tf.pow(self.X - self.decoded, 2))<\/code> is sufficient?<\/p>\n\n<p>***edit: some more code to describe my training\nThis is my Autoencoder Class:<\/p>\n\n<pre><code>class AutoEncoder():\ndef __init__(self):\n    # Training Parameters\n    self.learning_rate = 0.005\n    self.alpha = 0.5\n\n    # Network Parameters\n    self.num_input = 24 # one day as input\n    self.num_hidden_1 = 10 # 2nd layer num features\n    self.num_hidden_2 = 3 # 2nd layer num features (the latent dim)\n\n    self.X = tf.placeholder(\"float\", [None, self.num_input])\n\n    self.weights = {\n        'encoder_h1': tf.Variable(tf.random_normal([self.num_input, self.num_hidden_1])),\n        'encoder_h2': tf.Variable(tf.random_normal([self.num_hidden_1, self.num_hidden_2])),\n        'decoder_h1': tf.Variable(tf.random_normal([self.num_hidden_2, self.num_hidden_1])),\n        'decoder_h2': tf.Variable(tf.random_normal([self.num_hidden_1, self.num_input])),\n    }\n    self.biases = {\n        'encoder_b1': tf.Variable(tf.random_normal([self.num_hidden_1])),\n        'encoder_b2': tf.Variable(tf.random_normal([self.num_hidden_2])),\n        'decoder_b1': tf.Variable(tf.random_normal([self.num_hidden_1])),\n        'decoder_b2': tf.Variable(tf.random_normal([self.num_input])),\n    }    \n\n    self.encoded = self.encoder(self.X)\n    self.decoded = self.decoder(self.encoded)\n\n    # Define loss and optimizer, minimize the squared error\n    self.loss = tf.reduce_mean(tf.pow(self.X - self.decoded, 2))\n\n    self.optimizer = tf.train.RMSPropOptimizer(self.learning_rate).minimize(self.loss)\n\ndef encoder(self, x):\n    # sigmoid, tanh, relu\n    en_layer_1 = tf.nn.sigmoid (tf.add(tf.matmul(x, self.weights['encoder_h1']),\n                                   self.biases['encoder_b1']))\n\n    en_layer_2 = tf.nn.sigmoid (tf.add(tf.matmul(en_layer_1, self.weights['encoder_h2']),\n                                   self.biases['encoder_b2']))\n\n    return en_layer_2\n\ndef decoder(self, x):\n    de_layer_1 = tf.nn.sigmoid (tf.add(tf.matmul(x, self.weights['decoder_h1']),\n                                   self.biases['decoder_b1']))\n\n    de_layer_2 = tf.nn.sigmoid (tf.add(tf.matmul(de_layer_1, self.weights['decoder_h2']),\n                                   self.biases['decoder_b2']))\n\n    return de_layer_2\n<\/code><\/pre>\n\n<p>and this is how I train my network (input data have shape (number_days, 24)):<\/p>\n\n<pre><code>model = autoencoder.AutoEncoder()\n\nnum_epochs = 3\nbatch_size = 50\nnum_batches = 300\n\ndisplay_batch = 50\nexamples_to_show = 16\n\nloss_values = []\n\nwith tf.Session() as sess:\n\nsess.run(tf.global_variables_initializer())\n\n#training\nfor e in range(1, num_epochs+1):\n    print('starting epoch {}'.format(e))\n    for b in range(num_batches):\n        # get next batch of data\n        batch_x = dataset.next_batch(batch_size)\n\n        # Run optimization op (backprop) and cost op (to get loss value)\n        l = sess.run([model.loss], feed_dict={model.X: batch_x})\n        sess.run(model.optimizer, feed_dict={model.X: batch_x})            \n\n        # Display logs\n        if b % display_batch == 0:\n            print('Epoch {}: Batch ({}) Loss: {}'.format(e, b, l))\n            loss_values.append(l)\n\n\n# testing\ntest_data = dataset.next_batch(batch_size)\ndecoded_test_data = sess.run(model.decoded, feed_dict={model.X: test_data})\n<\/code><\/pre>\n","answers":[{"AnswerId":"53192319","CreationDate":"2018-11-07T15:14:30.763","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>The sigmoid output is floored at 0, so it cannot reproduce your data that is below 0.<\/p>\n\n<p>If you want to use a sigmoid output, then rescale your data between ]0;1[ (0 and 1 excluded).<\/p>\n"},{"AnswerId":"53191644","CreationDate":"2018-11-07T14:38:38.500","ParentId":null,"OwnerUserId":"10618796","Title":null,"Body":"<p>Just a suggestion, I have had some issues with autoencoders using the <code>sigmoid<\/code> function.<\/p>\n\n<p>I switched to <code>tanh<\/code> or <code>relu<\/code> and those improved the results. \nWith the autoencoder it is basically learning to recreate the output from the input, by encoding and decoding. If you mean it's the same as the input, then you are getting what you want. It has learned the data set.<\/p>\n\n<p>Ultimately you can compare by reviewing the Mean Squared Error between the input and output and see if it is exactly the same. If you mean that the output is exactly the same regardless of the input, that isn't something I've run into.  I guess if your input doesn't vary much from day to day, then I could imagine that would have some impact. Are you looking for anomalies?<\/p>\n\n<p>Also, if you have a time series for training, I wouldn't shuffle the data in this particular case. If the temporal order is significant, you introduce data leakage (basically introducing future data into the training set) depending on what you are trying to achieve. <\/p>\n\n<p>Ah, I didn't initially see your post with the graph results.. thanks for adding.<\/p>\n"}]}
+{"QuestionId":53191799,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><neural-network><backpropagation>","CreationDate":"2018-11-07T14:47:07.690","AcceptedAnswerId":null,"OwnerUserId":7285732.0,"Title":"Backpropagation shapes do not match","Body":"<p>I have been trying to create a 3 layer (1 input - 2 hidden - 1 output) neural network from scratch using numpy. The output layer has only 1 neuron. I am trying to use stochastic gradient descent with mini batches. What I have done so far is as follows:<\/p>\n\n<pre><code>N1 = 50 # number of neurons in the first hidden layer\nN2 = 20 # number of neurons in the second hidden layer\nfor n in range(10000):\n\n    weight_increment2 = 0\n    weight_increment1 = 0\n    weight_increment0 = 0\n\n    for j in range( batch_size):\n\n        # choose a sample randomly from training set\n        index = np.random.randint( trainims.shape[2] - 1)\n        layer0 = trainims[:,:,index]\n\n        # convert it to a column vector and add 1 for bias\n        layer0 = layer0.flatten()\n        layer0 = np.append( layer0, 1)\n        layer0 = layer0.reshape((1025,1))\n\n        # feed forward\n        layer1 = np.tanh( np.dot( w0.T, layer0))\n        layer1 = np.append( layer1, 1)\n        layer1 = layer1.reshape((N1+1,1))\n        layer2 = np.tanh( np.dot( w1.T, layer1))\n        layer2 = np.append( layer2, 1)\n        layer2 = layer2.reshape((N2+1,1))\n        layer3 = math.tanh( np.dot( w2.T, layer2))\n\n        # backpropagation\n        layer3_error = trainlbls[0,index] - layer3\n        layer3_gradient = 1 - layer3 * layer3\n        layer3_delta = layer3_error * layer3_gradient\n\n        layer2_error = layer3_delta * w2\n        layer2_gradient = 1 - np.multiply(layer2, layer2)\n        layer2_delta = np.multiply(layer2_error, layer2_gradient)\n        strippedlayer2delta = layer2_delta[0:N2,:]\n\n        layer1_error = strippedlayer2delta.T * w1\n        layer1_gradient = 1 - np.multiply( layer1, layer1)\n        layer1_delta = np.multiply( layer1_error, layer1_gradient)\n        strippedlayer1delta = layer1_delta[0:N1,:]\n\n\n        weight_increment2 = weight_increment2 + learning_rate * layer2 * layer3_delta\n        weight_increment1 = weight_increment1 + learning_rate * np.dot( layer1, strippedlayer2delta.T)\n        weight_increment0 = weight_increment0 + learning_rate * np.dot( layer0, strippedlayer1delta.T)\n\n    # update the weights\n    w2 = w2 + weight_increment2\n    w1 = w1 + weight_increment1\n    w0 = w0 + weight_increment0\n<\/code><\/pre>\n\n<ul>\n<li><code>w0<\/code> is the input weights: 1025x50 (1024=input, +1 is the bias term)<\/li>\n<li><code>w1<\/code> is the first hidden layer weights: 51x20 (50=number of first hidden layer neurons, +1 is bias term)<\/li>\n<li><code>w2<\/code> is the second hidden layer weights: 21x1 (20=number of second hidden lyaer neurons, +1 is bias term)<\/li>\n<li><code>layer0<\/code> is the input layer vector: an input sample (1024x1) and <code>1<\/code> appended to the end, making <code>layer0<\/code> a 1025x1 vector.<\/li>\n<li><code>layer1<\/code> is the first hidden layer vector: After calculating <code>tanh( np.dot( w0.T, layer0))<\/code>, I add <code>1<\/code> to the end, which results with a 51x1 vector<\/li>\n<li><code>layer2<\/code> is the second hidden layer vector: After calculating <code>tanh( np.dot( w1.T, layer1))<\/code>, I add <code>1<\/code> to the end, which results with a 21x1 vector<\/li>\n<li><code>layer3<\/code> is the output vector: <code>tanh( np.dot( w2.T, layer2))<\/code>, which is just a nunmber<\/li>\n<li><code>trainims<\/code> is the training set with size 32x32x1900, meaning 1900 32x32 images<\/li>\n<\/ul>\n\n<p>While doing backpropagation, I omit the last rows of <code>layer2_delta<\/code> and <code>layer1_delta<\/code>, but I do not exactly know why I do it. I suppose it is because of omitting the <code>1<\/code> terms I append to <code>layer1<\/code> and <code>layer2<\/code>.<\/p>\n\n<p>When I try to run the code, it gives the following error:<\/p>\n\n<pre><code>ValueError Traceback (most recent call last)\n&lt;ipython-input-11-98c3beb3b346&gt; in &lt;module&gt;()\n     62         print( \"weight_increment1:\", weight_increment1.shape)\n     63 \n---&gt; 64         weight_increment0 = weight_increment0 + learning_rate * np.dot( layer0, strippedlayer1delta.T)\n     65         print( \"weight_increment0:\", weight_increment0.shape)\n     66 \n\nValueError: shapes (1025,1) and (20,50) not aligned: 1 (dim 1) != 20 (dim 0)\n<\/code><\/pre>\n\n<p>I know that the dimensions should match to obtain a 1025x50 matrix and update <code>w0<\/code>, but I could not come up with a solution. Maybe I am doing backpropagation wrong, I do not know. Can anybody suggest a solution for this problem?<\/p>\n\n<p>Edit:\nOK I have noticed my mistake. Previously I was doing<\/p>\n\n<pre><code>layer1_error = strippedlayer2delta.T * w1\n<\/code><\/pre>\n\n<p>which was giving a matrix, but an error must be a vector for a layer, not a matrix. Changing this line to<\/p>\n\n<pre><code>layer1_error = np.dot( w1, strippedlayer2delta)\n<\/code><\/pre>\n\n<p>fixed the shape error.<\/p>\n\n<p>But still, can anybody check whether my implementations of stochastic gradient descent with mini batches, and backpropagation is correct or not?<\/p>\n","answers":[]}
+{"QuestionId":53192096,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2018-11-07T15:02:51.040","AcceptedAnswerId":53223108.0,"OwnerUserId":5581481.0,"Title":"Keras: Dimensions must be equal","Body":"<p>I was doing some classification with keras, when met this error:<\/p>\n\n<pre><code>InvalidArgumentError: Dimensions must be equal, but are 256 and 8 for 'dense_185\/MatMul' (op: 'MatMul') with input shapes: [?,256], [8,300].\n<\/code><\/pre>\n\n<p>It surprised me because the dimension of the input to the dense is 1.<\/p>\n\n<p>This is a sequential model with a few custom layers. I have no idea why 8 appears in the error of dense layer.<\/p>\n\n<pre><code>class Residual(Layer):\n    def __init__(self,input_shape,**kwargs):\n        super(Residual, self).__init__(**kwargs)\n        self.input_shapes = input_shape\n\n    def call(self, x):\n        print(np.shape(x)) #    (?, 128, 8) \n        first_layer =   Conv1D(256, 4, activation='relu', input_shape = self.input_shapes)(x)\n        print(np.shape(first_layer))     (?, 125, 256)\n        x =             Conv1D(256, 4, activation='relu')(first_layer)\n        print(np.shape(x)) (?, 122, 256) \n        x =             Conv1D(256, 4, activation='relu')(x)\n        print(np.shape(x)) (?, 119, 256)\n        x =             ZeroPadding1D(padding=3)(x)\n        residual =      Add()([x, first_layer])\n        x = Activation(\"relu\")(residual)\n        return x\n\nclass Pooling(Layer):\n    def __init__(self,**kwargs):\n        super(Pooling, self).__init__(**kwargs)\n\n    def call(self, x):\n        first_layer =   GlobalMaxPooling1D(data_format='channels_last')(x)\n        second_layer =  GlobalAveragePooling1D(data_format='channels_last')(x)\n        pooling =      Add()([first_layer, second_layer])\n        print(np.shape(pooling)) (?, 256)\n        return pooling\n\nmodel = Sequential()\nmodel.add(Residual(input_shape=(128,8)))\nmodel.add(Pooling())\nmodel.add(Dense(300, activation='relu'))\nmodel.add(Dense(150, activation='relu'))\nmodel.add(Dense(10, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer=keras.optimizers.Adadelta(), metrics=['accuracy'])\nmodel.fit(np.array(dataset_data), dataset_target, epochs=1000, validation_split=0.1, verbose=1, batch_size=8)\n<\/code><\/pre>\n\n<p>Dimensions:<\/p>\n\n<p>(1000, 128, 8) - input (1000 audio, 8 features, 128 seq_length)<\/p>\n\n<p>(1000, 10) - target (1000 audio, 10 classes)<\/p>\n","answers":[{"AnswerId":"53223108","CreationDate":"2018-11-09T09:34:53.290","ParentId":null,"OwnerUserId":"7141506","Title":null,"Body":"<p>I think there are two edits required:<\/p>\n\n<ol>\n<li>Add <code>InputLayer<\/code> as entrance for the data<\/li>\n<li>Define <code>compute_output_shape<\/code> method at least for <code>Pooling<\/code> layer (<a href=\"https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/\" rel=\"nofollow noreferrer\">link<\/a>). If this method is not defined, <code>Dense<\/code> layer can't figure out what's input shape for it, I guess, and then fails.<\/li>\n<\/ol>\n\n<p>Also there's minor editing - since model have <code>InputLayer<\/code>, you need no more <code>input_shape<\/code> kwarg in <code>Residual<\/code> layer.<\/p>\n\n<pre><code>class Residual(Layer):\n    def __init__(self, **kwargs):  # remove input shape\n        super(Residual, self).__init__(**kwargs)\n\n    def call(self, x):\n        print(np.shape(x))\n        first_layer = Conv1D(256, 4, activation='relu')(x)\n        print(np.shape(first_layer))    \n        x = Conv1D(256, 4, activation='relu')(first_layer)\n        print(np.shape(x)) \n        x = Conv1D(256, 4, activation='relu')(x)\n        print(np.shape(x)) \n        x = ZeroPadding1D(padding=3)(x)\n        residual = Add()([x, first_layer])\n        x = Activation(\"relu\")(residual)\n        return x\n\n\nclass Pooling(Layer):\n    def __init__(self, **kwargs):\n        super(Pooling, self).__init__(**kwargs)\n\n    def call(self, x):\n        # !!! I build model without data_format argument - my version of keras\n        # doesn't support it !!!\n        first_layer = GlobalMaxPooling1D(data_format='channels_last')(x)  \n        second_layer = GlobalAveragePooling1D(data_format='channels_last')(x)\n        pooling = Add()([first_layer, second_layer])\n        print(np.shape(pooling)) \n        self.output_dim = int(np.shape(pooling)[-1])  # save output shape\n        return pooling\n\n    def compute_output_shape(self, input_shape):\n        # compute output shape here\n        return (input_shape[0], self.output_dim)  \n<\/code><\/pre>\n\n<p>Initialize model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(InputLayer((128,8)))\nmodel.add(Residual())\nmodel.add(Pooling())\nmodel.add(Dense(300, activation='relu'))\nmodel.add(Dense(150, activation='relu'))\nmodel.add(Dense(10, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', \n              optimizer=keras.optimizers.Adadelta(), \n              metrics=['accuracy'])\n\nOut:\n(?, 128, 8)\n(?, 125, 256)\n(?, 122, 256)\n(?, 119, 256)\n(?, 256)\n<\/code><\/pre>\n\n<p>Summary of the model (don't know why Residual and Pooling don't show params the have. I guess some additional method required for this classes to count internal params):<\/p>\n\n<pre><code>model.summary()\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nresidual_10 (Residual)       (None, 128, 8)            0         \n_________________________________________________________________\npooling_8 (Pooling)          (None, 256)               0         \n_________________________________________________________________\ndense_15 (Dense)             (None, 300)               77100     \n_________________________________________________________________\ndense_16 (Dense)             (None, 150)               45150     \n_________________________________________________________________\ndense_17 (Dense)             (None, 10)                1510      \n=================================================================\nTotal params: 123,760\nTrainable params: 123,760\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Create fake data and check training process: <\/p>\n\n<pre><code>dataset_data = np.random.randn(1000, 128, 8)\ndataset_target = np.zeros((1000, 10))\ndataset_target[:, 0] = 1\nmodel.fit(np.array(dataset_data), dataset_target, epochs=1000,\n          validation_split=0.1, verbose=1, batch_size=8)\nTrain on 900 samples, validate on 100 samples\nEpoch 1\/1000\n900\/900 [==============================] - 2s 2ms\/step - loss: 0.0235 - acc: 0.9911 - val_loss: 9.4426e-05 - val_acc: 1.0000\nEpoch 2\/1000\n900\/900 [==============================] - 1s 1ms\/step - loss: 4.2552e-05 - acc: 1.0000 - val_loss: 1.7458e-05 - val_acc: 1.0000\nEpoch 3\/1000\n900\/900 [==============================] - 1s 1ms\/step - loss: 1.1342e-05 - acc: 1.0000 - val_loss: 7.3141e-06 - val_acc: 1.0000\n... and so on\n<\/code><\/pre>\n\n<p>Looks like it works.<\/p>\n"}]}
+{"QuestionId":53192160,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-07T15:05:29.567","AcceptedAnswerId":null,"OwnerUserId":10533662.0,"Title":"FileNotFoundError: [WinError 3] The system cannot find the path specified: '\/Users\/User\/Downloads\/TrafficSigns\\\\TrafficSigns\/Training'","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/bPvGc.png\" rel=\"nofollow noreferrer\">path<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/ROjlk.png\" rel=\"nofollow noreferrer\">ROOT_PATH = \"C:\/Users\/User\/Downloads\/TrafficSigns\"<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/fERf1.png\" rel=\"nofollow noreferrer\">I am having problem in uploading data but can't figure out the problem. the code is given in the attached image<\/a><\/p>\n","answers":[{"AnswerId":"53192243","CreationDate":"2018-11-07T15:10:25.263","ParentId":null,"OwnerUserId":"8154557","Title":null,"Body":"<p>You are running on Windows. There is no such thing as the '\/' root path on Windows.<\/p>\n\n<p>Provide the correct path to the TrafficSigns directory. It is porbably \"C:\/Users\/User\/...\". If forward slashes are not allowed, just replace them with double backslashes.<\/p>\n"}]}
+{"QuestionId":53192619,"AnswerCount":1,"Tags":"<python><tensorflow><vectorization>","CreationDate":"2018-11-07T15:31:33.913","AcceptedAnswerId":53193612.0,"OwnerUserId":6091064.0,"Title":"Using gather_nd to build matrix","Body":"<p>I am trying to construct the rank-2 <code>tf.float<\/code> tensor, <code>x<\/code>, from a rank-3 <code>tf.float<\/code> tensor, <code>y<\/code>, and rank-2 <code>tf.int32<\/code> tensor, <code>z<\/code> as:<\/p>\n\n<p><code>x[i][j] = y[i,z[i][j],j]<\/code><\/p>\n\n<p>I know that I need to use <code>tf.gather_nd<\/code> as:<\/p>\n\n<pre><code>x = tf.gather_nd(y,indices)\n<\/code><\/pre>\n\n<p>where <\/p>\n\n<p><code>indices[i][j][:] = [i,z[i][j],j]<\/code><\/p>\n\n<p>However, I am having trouble using tensorflow functions to essentially expand <code>z<\/code> into a higher-rank to construct <code>indices<\/code>.<\/p>\n\n<p>I am trying to maintain these operations in a vectorized form.<\/p>\n\n<p>Is it more practical to simply use <code>tf.stack<\/code> as,<\/p>\n\n<p><code>indices = tf.stack([ii,z,jj],axis=-1)<\/code><\/p>\n\n<p>where <\/p>\n\n<p><code>ii[i,:] = i<\/code> <\/p>\n\n<p>and <\/p>\n\n<p><code>jj[:,j] = j<\/code>?<\/p>\n","answers":[{"AnswerId":"53193612","CreationDate":"2018-11-07T16:22:34.403","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think this does what you need:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Inputs\ny = tf.placeholder(tf.float32, [None, None, None])\nz = tf.placeholder(tf.int32, [None, None])\n# Make first and last indices\ny_shape = tf.shape(y)\nii, jj = tf.meshgrid(tf.range(y_shape[0]), tf.range(y_shape[2]), indexing='ij')\n# Make full ND index\nidx = tf.stack([ii, z, jj], axis=-1)\n# Gather result\nx = tf.gather_nd(y, idx)\n# Test\nwith tf.Session() as sess:\n    # Numbers from 0 to 11 in a (3, 4) matrix\n    a = np.arange(12).reshape((3, 4))\n    # Make Y with replicas of the matrix multiplied by 1, 10 and 100\n    y_val = np.stack([a, a * 10, a * 100], axis=1).astype(np.float32)\n    # Z will be a (3, 4) matrix of values 0, 1, 2, 0, 1, 2, ...\n    z_val = (a % 3).astype(np.int32)\n    # X  should have numbers from 0 to 11 multiplied by 1, 10, 100, 1, 10, 100, ...\n    x_val = sess.run(x,  feed_dict={y: y_val, z: z_val}) #, feed_dict={y: y_val, z: z_val})\n    print(x_val)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[   0.   10.  200.    3.]\n [  40.  500.    6.   70.]\n [ 800.    9.  100. 1100.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53192824,"AnswerCount":3,"Tags":"<python><web-services><tensorflow><initialization><multiple-models>","CreationDate":"2018-11-07T15:42:30.753","AcceptedAnswerId":53347180.0,"OwnerUserId":9864814.0,"Title":"Initialize a tensorflow model in main(), pass it to __init__ and execute inside another method","Body":"<p>I want to build a web service with flask where multiple deep learning models will be applied to certain types of data to give back a result. Currently, I want to load them locally on main() once at start, pass them to <strong>init<\/strong> to just initialize them once when the execution of the script starts and then call it every time it is needed to perform a forward pass to return something. So far that's what I ve done with the rest but I don't know how to handle a pure tensorflow model initialization. The below code works fine. Any Suggestions, alterations are appreciated:<\/p>\n\n<pre><code>def evaluate_sample(numpy_array, no_of_frames):\n    _IMAGE_SIZE = 224\n    _SAMPLE_VIDEO_FRAMES = no_of_frames\n    _CHECKPOINT_PATHS = {'flow': 'data\/checkpoints\/flow_scratch\/model.ckpt'}\n    NUM_CLASSES = 400\n\n    flow_input = tf.placeholder(\n        tf.float32,\n        shape=(1, _SAMPLE_VIDEO_FRAMES, _IMAGE_SIZE, _IMAGE_SIZE, 2))\n    with tf.variable_scope('Flow', reuse=tf.AUTO_REUSE):\n        flow_model = i3d.InceptionI3d(NUM_CLASSES, spatial_squeeze=True, final_endpoint='Logits')\n        flow_logits, _ = flow_model(flow_input, is_training=False, dropout_keep_prob=1.0)\n    flow_variable_map = {}\n    for variable in tf.global_variables():\n        if variable.name.split('\/')[0] == 'Flow':\n            flow_variable_map[variable.name.replace(':0', '')] = variable\n    flow_saver = tf.train.Saver(var_list=flow_variable_map, reshape=True)    \n    model_logits = flow_logits\n    model_predictions = tf.nn.softmax(model_logits)\n\n    with tf.Session() as sess:\n        feed_dict = {}\n        flow_saver.restore(sess, _CHECKPOINT_PATHS['flow'])\n        flow_sample = numpy_array\n        feed_dict[flow_input] = flow_sample\n        out_logits, out_predictions = sess.run(\n            [model_logits, model_predictions],\n            feed_dict=feed_dict)\n        logits2=np.asarray(out_logits)\n\n    return logits2\n\n\n\n\ndef get_flow_features(video_path):\n         .....\n         aggregated_flow_vector = evaluate_sample(final_np_cropped_flow, len(all_frames_flow))\n         .....\n\n\n    class GetOutOfContext:\n        def __init__(self, keras_model, pytorch_model, word2vec_model, max_pooling):\n            self.keras_model = keras_model\n            self.pytorch_model = pytorch_model\n            self.word2vec_model = word2vec_model\n            self.max_pooling = max_pooling\n            #self.kineticsi3d = kineticsi3d\n            print(\"Similarity Between Video and Text Service Initialized...\")\n\n        def get(self):\n            dirpath = tempfile.mkdtemp()+\"\/\"\n            video_path = download_video(url,dirpath)\n            aggregated_audio = get_audio_features(video_path)\n            aggregated_flow = get_flow_features(video_path)\n            aggregated_video = get_visual_features(video_path, dirpath)\n            aggregated_text = get_word_features(text)\n            .......\n\n    if __name__ == \"__main__\":\n        \"\"\"Loading Prediction Model\"\"\"\n        video_modality_dim = {'face': (128,128), 'audio': (128*16,128),'visual': (2048,2048), 'motion': (1024,1024)}\n        the_model = Net(video_modality_dim, 300, audio_cluster=16)\n        the_model.load_state_dict(torch.load('\/home\/estathop\/Desktop\/journalmodel\/msrvttjournal.pt', map_location=lambda storage, loc: storage))\n        the_model.eval()\n        \"\"\"Loading Image Feature Extraction Model\"\"\"\n        model = ResNet152(include_top=False, weights='imagenet', pooling = 'avg') #cons\n        \"\"\"Loading Word2Vec Model\"\"\"\n        model2 = api.load(\"word2vec-google-news-300\")  \n        maxpoolingmodel = keras.layers.pooling.GlobalMaxPooling1D()\n        word_vectors = model2.wv\n        nltk.download('stopwords')\n\n        x = GetOutOfContext(model,the_model,model2, maxpoolingmodel)\n        y = x.get()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53347034","CreationDate":"2018-11-17T00:22:36.227","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>I'd hold on to the <code>Session<\/code> and just <code>run<\/code> multiple times. <code>saver.restore<\/code> should happen just once. For error checking you can <code>tf.get_default_graph().finalize()<\/code> after you specify the model to make sure the graph isn't changing each request, which would slow things down.<\/p>\n"},{"AnswerId":"53347180","CreationDate":"2018-11-17T00:49:21.190","ParentId":null,"OwnerUserId":"7978903","Title":null,"Body":"<p>In your current code, the model is defined and loaded in <code>evaluate_sample<\/code>, you can simply move the majority of the code from <code>evaluate_sample<\/code> to main or <strong>init<\/strong> and pass the tf.Session object and the graph to <code>evaluate_sample<\/code>.\nHere is a clumsy example:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef main():\n    sess=tf.Session()\n    a=tf.placeholder(tf.float32)\n    b=tf.constant(5.0)\n    c=a+b\n    evaluate_sample(sess,c,a)\n\ndef evaluate_sample(session,graph,input):\n    print(session.run(graph,feed_dict={input:3}))\n\nif __name__ == \"__main__\":\n    main()\n<\/code><\/pre>\n"},{"AnswerId":"53193866","CreationDate":"2018-11-07T16:36:07.780","ParentId":null,"OwnerUserId":"7829500","Title":null,"Body":"<p>Are you trying to load a pretrained model and run an inference? By initializing are you referring to loading a model or initializing new weights for each instance this is executed?<\/p>\n"}]}
+{"QuestionId":53193208,"AnswerCount":1,"Tags":"<keras><metrics><autoencoder>","CreationDate":"2018-11-07T16:01:11.680","AcceptedAnswerId":null,"OwnerUserId":2491687.0,"Title":"Accuracy on middle layer of autoencoder implemente using Keras","Body":"<p>I have implemented an autoencoder using Keras. I understand that I can add accuracy performance metric as follows:<\/p>\n\n<blockquote>\n  <p>autoencoder.compile(optimizer='adam', \n                      loss='mean_squared_error', \n                      metrics=['accuracy'])<\/p>\n<\/blockquote>\n\n<p>My question is:\nIs the accuracy metric applied on the last layer of the decoder by default? If so, how can I set it so that it would get the representations from middle (hidden) layer to compute accuracy performance? Do I need to define a custom metric? How would that work?<\/p>\n","answers":[{"AnswerId":"53193543","CreationDate":"2018-11-07T16:19:03.770","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>It seems that what you really want is a multiple output network.<\/p>\n\n<p>So on top of your middle layer that defines your embedding, add a layer (or more) to do your classification.<\/p>\n\n<p>Then have a look at <a href=\"https:\/\/stackoverflow.com\/questions\/44036971\/multiple-outputs-in-keras\">Multiple outputs in Keras<\/a> to create your global cost.<\/p>\n\n<p>You may also want to start by training the autoendoder only, then the classifier additional layers only to see the performance, you can also balance the accuracy of the encoder vs the accuracy of the classifier as a loss, training \"both\" networks at the same time.<\/p>\n"}]}
+{"QuestionId":53193410,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-07T16:11:54.320","AcceptedAnswerId":null,"OwnerUserId":3192738.0,"Title":"How to winsorize a tensorflow tensor?","Body":"<p>During training, I generate a tensor which contains some outlier values.How to set these outlier values to a proper smaller value? \nI have tried batch normalization but does not help.<\/p>\n","answers":[{"AnswerId":"53200271","CreationDate":"2018-11-08T01:21:07.807","ParentId":null,"OwnerUserId":"8384504","Title":null,"Body":"<p>You can use tf.clip_by_value. Put this operation at desired place.<\/p>\n"}]}
+{"QuestionId":53193522,"AnswerCount":1,"Tags":"<python-3.x><oop><tensorflow><low-level-api>","CreationDate":"2018-11-07T16:18:22.177","AcceptedAnswerId":53197954.0,"OwnerUserId":7779411.0,"Title":"Create class object with a tuple having tensorflow objects","Body":"<p>I have a parametersTheta class which creates neural network as follows:<\/p>\n\n<pre><code>class parametersTheta:\n    def __init__(self, weight1, weight2,....):\n        self.weightName1 = weight1\n        self.weightName2 = weight2\n        ...\n        self.sess = tf.Session()\n    def makeWorkerTheta(self, param):\n        return parametersTheta(self.sess.run(functionCalculatingTensorWeights, feed_dict={...}))\n<\/code><\/pre>\n\n<p>self.sess.run creates a tuple of all the weight tensors. However, error pops up saying you need to input weight2 and onwards, i.e. the tuple goes into weight1<\/p>\n\n<p>How can I solve this? Basically, how can I create an instance of class parametersTheta with a tuple?<\/p>\n","answers":[{"AnswerId":"53197954","CreationDate":"2018-11-07T21:18:02.360","ParentId":null,"OwnerUserId":"8384504","Title":null,"Body":"<p>You can instantiate class with tuple expanded to arguments like this.<\/p>\n\n<pre><code>parametersTheta(*(weight1, weight2, ...))\n<\/code><\/pre>\n\n<p>An asterisk before a tuple expand it to a corresponding arguments list.<\/p>\n"}]}
+{"QuestionId":53193863,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><flask><keras><artificial-intelligence>","CreationDate":"2018-11-07T16:36:05.163","AcceptedAnswerId":null,"OwnerUserId":5763193.0,"Title":"Loading multiple models in Python to do comparison between results of model","Body":"<p>I'm trying to predict classes using two different models. \nThese models are saved and loaded using a .json and .h5 file. <\/p>\n\n<p>The problem I'm running into is that 50% of the time everything works fine, both models are correctly created and are able to do their predictions.<\/p>\n\n<p>Half of the time there seems to go something wrong with the creation of the model(s) because then, when I try to predict something it gives me an error like this <\/p>\n\n<pre><code>InvalidArgumentError: You must feed a value for placeholder tensor 'input_1_1' with dtype float and shape [?,10693]\n [[{{node input_1_1}} = Placeholder[dtype=DT_FLOAT, shape=[?,10693], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>Running both models separate\/not at the same time works perfectly. No problems whatsoever. <\/p>\n\n<p>After some googling I found that Keras keeps a global graph of the model(s) and this might cause the error. I feel like I have tried most of the suggestions I found online but I still can't seem to fix the problem of having two different models at the same time in a program.<\/p>\n\n<p>So based on a given argument I want to either predict using model1, or model2:<\/p>\n\n<pre><code>def predict_using_cnn(text, model, tokenizer, to_predict):\nif to_predict == 'functiegroep':\n    with graph.as_default():\n        sequences_test = tokenizer.texts_to_sequences([text])\n        X_test = pad_sequences(sequences_test, maxlen=6616)  # TODO fix this hardcoded thingy\n        y_pred = model.predict(X_test)\n    return build_response(y_pred, to_predict)\nelif to_predict == 'kernfunctie':\n    with graph.as_default():\n        sequences_test = tokenizer.texts_to_sequences([text])\n        X_test = pad_sequences(sequences_test, maxlen=10693)  # TODO fix this hardcoded thingy\n        y_pred = model.predict(X_test)\n    return build_response(y_pred, to_predict)\n<\/code><\/pre>\n\n<p>This is where sometimes it goes intoe the first if statement thinking that it needs an input of 10693, which is the <em>other<\/em> model. So the program seems mix these things up at random.<\/p>\n\n<p>Both models are loaded using the same function (ofcourse they are saved to two different variables) <\/p>\n\n<pre><code>def load_cnn_model(model_name):\nprint(\"loading\" + str(model_name))\njson_file = open(str(model_name) + '.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nmodel = model_from_json(loaded_model_json)\nmodel.load_weights(str(model_name) + '.h5')\nreturn model\n<\/code><\/pre>\n\n<p>Anything to help me fix this problem is very much appreciated!<\/p>\n","answers":[{"AnswerId":"53275720","CreationDate":"2018-11-13T07:15:52.497","ParentId":null,"OwnerUserId":"5763193","Title":null,"Body":"<p>I eventually got it working like this:<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    session = tf.Session()\n    with session.as_default():\n        graph_fg = graph\n        session_fg = session\n        cnn_fg_model = predict.load_cnn_model('models\/cnn\/new_model_fg\/new_model_fg')\n\nwith graph.as_default():\n    session = tf.Session()\n    with session.as_default():\n        graph_kf = graph\n        session_kf = session\n        cnn_kf_model = predict.load_cnn_model('models\/cnn\/new_model_kf\/new_model_kf')\n\nwith graph.as_default():\n    with session.as_default():\n        if to_predict == 'functiegroep':\n            sequences_test = tokenizer.texts_to_sequences([text])\n            X_test = pad_sequences(sequences_test, maxlen=6616)  # TODO fix this hardcoded thingy\n            y_pred = model.predict(X_test)\n            return build_response_cnn(y_pred, to_predict)\n        elif to_predict == 'kernfunctie':\n            sequences_test = tokenizer.texts_to_sequences([text])\n            X_test = pad_sequences(sequences_test, maxlen=10693)  # TODO fix this hardcoded thingy\n            y_pred = model.predict(X_test)\n            return build_response_cnn(y_pred, to_predict)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53194185,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-07T16:54:27.807","AcceptedAnswerId":null,"OwnerUserId":4500293.0,"Title":"Adam Optimizer and Regularization in Tensorflow","Body":"<p>I've been creating a CNN using the Adam Optimizer. My model has been really overfitting so I decided to add some regularization. I initially thought that adding the \"kernel_regularization\" parameter to each convolutional layer and collecting all the regulaizations and adding it to the loss would work. All this does to my network is make it learn what it did without the regularization much slower but there is still overfitting after many more epochs.<\/p>\n\n<p>However, reading around a bit it looks like that regularization is not the correct way to go when using Adam in Tensorflow, but I could not find a definitive answer on this. <\/p>\n\n<p>If someone could please clarify and direct me on how to add regularization when using the Adam optimizer, I would be very grateful.<\/p>\n\n<p>EDIT: A little more description about the task<\/p>\n\n<p>I am doing a binary segmentation task using a UNet architecture. <\/p>\n","answers":[]}
+{"QuestionId":53194564,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-07T17:18:30.587","AcceptedAnswerId":null,"OwnerUserId":10619689.0,"Title":"Tensorflow Keras CuDNN LSTM layers do not support Masking","Body":"<p>I am trying to build a NER using TF keras, when I do zero padding to get the batches to same length and then add a tf.keras.layers.Masking() layer to mask the padding, the error I get is that CuDNNLSTM does not support Masking.<\/p>\n\n<p>My Tensorflow-gpu version is 1.11,  Cuda 9.0, CudNN 7.3.1.<\/p>\n\n<p>I currently do have the version working where I just group same length samples into each batch to avoid padding, I would like to know if we should just not pad from now on or if CuDNN LSTM will support masking paddings in the future?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"54073864","CreationDate":"2019-01-07T11:48:00.540","ParentId":null,"OwnerUserId":"9957560","Title":null,"Body":"<p>Just found this<\/p>\n\n<blockquote>\n  <p>CuDNN RNNs cannot be used with masking for the time being, and likely never will be. It would need to be implemented at the CuDNN level.<\/p>\n<\/blockquote>\n\n<p>see <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8620\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":53194704,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nan><loss>","CreationDate":"2018-11-07T17:27:46.377","AcceptedAnswerId":53212731.0,"OwnerUserId":4444978.0,"Title":"Keras loss function understanding","Body":"<p>In order to understand some callbacks of Keras better, I want to artificially create a <code>nan<\/code> loss.<\/p>\n\n<p>This is the function<\/p>\n\n<pre><code>def soft_dice_loss(y_true, y_pred):\n\n  from keras import backend as K\n  if K.eval(K.random_normal((1, 1), mean=2, stddev=2))[0][0] \/\/ 1 == 2.0:\n    # return nan\n    return K.exp(1.0) \/ K.exp(-10000000000.0) - K.exp(1.0) \/ K.exp(-10000000000.0)\n\n  epsilon = 1e-6\n\n  axes = tuple(range(1, len(y_pred.shape) - 1))\n  numerator = 2. * K.sum(y_pred * y_true, axes)\n  denominator = K.sum(K.square(y_pred) + K.square(y_true), axes)\n\n return 1 - K.mean(numerator \/ (denominator + epsilon))\n<\/code><\/pre>\n\n<p>So normally, it calculates the dice loss, but from time to time it should randomly return a <code>nan<\/code>. However, this does not seem to happen:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/psKL9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/psKL9.png\" alt=\"keras outputs\"><\/a><\/p>\n\n<p>From time to time though, when I try to run the code, it stops right at the start (before the first epoch) with an error, saying that <code>An operation has None for gradient. Please make sure that all of your ops have a gradient defined<\/code><\/p>\n\n<p>Does that mean, that the the random function of Keras is just evaluated once and then always returns the same value?\nIf so, why is that and how can I create a loss function that returns <code>nan<\/code> from time to time?<\/p>\n","answers":[{"AnswerId":"53212731","CreationDate":"2018-11-08T17:03:48.560","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Your first conditional statement is only evaluated once the loss function is defined (i.e. called; that is why Keras stops right at the start). Instead, you could use <a href=\"https:\/\/keras.io\/backend\/#switch\" rel=\"nofollow noreferrer\">keras.backend.switch<\/a> to integrate your conditional into the graph's logic. Your loss function could be something along the lines of:<\/p>\n\n<pre><code>import keras.backend as K\nimport numpy as np\n\n\ndef soft_dice_loss(y_true, y_pred):\n    epsilon = 1e-6\n    axes = tuple(range(1, len(y_pred.shape) - 1))\n    numerator = 2. * K.sum(y_pred * y_true, axes)\n    denominator = K.sum(K.square(y_pred) + K.square(y_true), axes)\n    loss = 1 - K.mean(numerator \/ (denominator + epsilon))\n\n    return K.switch(condition=K.random_normal((), mean=0, stddev=1) &gt; 3,\n                    then_expression=K.variable(np.nan),\n                    else_expression=loss)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53194827,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-07T17:37:05.647","AcceptedAnswerId":53204364.0,"OwnerUserId":639616.0,"Title":"cannot train Keras convolution network on GPU","Body":"<p>I can train a Keras network with <code>Dense<\/code> layer using <code>keras.datasets.fashion_mnist<\/code> dataset.  However, when I tried to train a convolutional network, I got an error.<\/p>\n\n<p>Here is some part of the code:<\/p>\n\n<pre><code>from tensorflow.keras.layers import *\n\nmodel = keras.Sequential([\n        Convolution2D(16, (3,3), activation='relu', input_shape=(28,28,1)),\n        MaxPooling2D(pool_size=(2,2)),\n        Flatten(),\n        Dense(16, activation='relu'),\n        Dense(10, activation='softmax')    \n])\nmodel.compile(optimizer=tf.train.AdamOptimizer(), \n          loss='sparse_categorical_crossentropy',\n          metrics=['accuracy'])\n\nmodel.fit(train_images, train_labels, epochs=5)\n<\/code><\/pre>\n\n<p>and its error when I tried to fit.<\/p>\n\n<blockquote>\n  <p>UnknownError: Failed to get convolution algorithm. This is probably\n  because cuDNN failed to initialize, so try looking to see if a warning\n  log message was printed above.     [[{{node conv2d\/Conv2D}} =\n  Conv2D[T=DT_FLOAT, data_format=\"NCHW\", dilations=[1, 1, 1, 1],\n  padding=\"VALID\", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](training\/TFOptimizer\/gradients\/conv2d\/Conv2D_grad\/Conv2DBackpropFilter-0-TransposeNHWCToNCHW-LayoutOptimizer,\n  conv2d\/Conv2D\/ReadVariableOp)]]    [[{{node\n  loss\/dense_1_loss\/broadcast_weights\/assert_broadcastable\/AssertGuard\/Assert\/Switch_2\/_69}}\n  = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\",\n  send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\",\n  send_device_incarnation=1, tensor_name=\"edge_112_l...t\/Switch_2\",\n  tensor_type=DT_INT32,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>I have <code>cudnn64_7.dll<\/code> in <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v9.0\\bin<\/code> and the <code>PATH<\/code> is already contain that folder.<\/p>\n","answers":[{"AnswerId":"53204364","CreationDate":"2018-11-08T08:59:40.683","ParentId":null,"OwnerUserId":"6081433","Title":null,"Body":"<p>I think this <a href=\"https:\/\/stackoverflow.com\/a\/49960236\/6081433\">link<\/a> would solve your problem, its because the cnDNN version you installed is not compatible with the cuDNN version that compiled in tensorflow.<\/p>\n"}]}
+{"QuestionId":53194848,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-07T17:38:40.513","AcceptedAnswerId":null,"OwnerUserId":5084566.0,"Title":"Fail to feed forward 1D signal to 1D convolution in Keras","Body":"<p>I am newbie of Keras. I want to perform 1D convolution, given a input array. How should I do it? I have write the code but it cannot run<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv1D\nimport numpy as np\nimport tensorflow as tf\nfrom keras.models import Model\n\ninput = np.array(tf.constant([1,2,3,4,5,6,7,8]))   \nmodel = Sequential()\nmodel.add(Conv1D(1,3,strides=1,padding='same', name='conv'))\n\nlayer_model = Model(inputs=input,outputs=model.get_layer('conv').output)\nconv_output = layer_model.predict(f)    \nprint (conv_output)\n<\/code><\/pre>\n\n<p>This is error<\/p>\n\n<pre><code>    AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-13-31db6e7dcb90&gt; in &lt;module&gt;()\n     10 \n     11 layer_model = Model(inputs=input,\n---&gt; 12                                 outputs=model.get_layer('conv').output)\n     13 conv_output = layer_model.predict(f)\n     14 print (conv_output)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/base_layer.py in output(self)\n    808         if not self._inbound_nodes:\n    809             raise AttributeError('Layer ' + self.name +\n--&gt; 810                                  ' has no inbound nodes.')\n    811         if len(self._inbound_nodes) &gt; 1:\n    812             raise AttributeError('Layer ' + self.name +\n\nAttributeError: Layer conv has no inbound nodes.\n<\/code><\/pre>\n","answers":[{"AnswerId":"53195818","CreationDate":"2018-11-07T18:43:42.727","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You shouldn't complicate yourself by mixing the Functional and Sequential APIs, a simpler way is:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv1D\nimport numpy as np\n\ninput = np.array([1,2,3,4,5,6,7,8])\ninput = input.reshape(1, 8, 1) # (samples, width, channels)   \n\nmodel = Sequential()\nmodel.add(Conv1D(1,3,padding='same', input_shape=(8, 1)))\n\nconv_output = model.predict(input)    \nprint(conv_output)\n<\/code><\/pre>\n\n<p>Note that we haven't set the convolutional weights, so they are automatically initialized to random values.<\/p>\n"}]}
+{"QuestionId":53194918,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-11-07T17:43:27.987","AcceptedAnswerId":53195867.0,"OwnerUserId":3642162.0,"Title":"Tensorflow tf.layers Dense Neural Net function vs. Class Interface","Body":"<p>I am trying to implement a helper class to create a standard Feedforward Neural network in python.\n<\/p>\n\n<p>Since I want the class to be general, there is a method called addHiddenLayer() which should append layers to the Flow Graph.\n<\/p>\n\n<p>To add layers to the flow graph I went through the tf.layers module which provides two options <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/dense\" rel=\"nofollow noreferrer\">tf.layers.dense<\/a> : A function which returns an object which can act as the input to the next layer.\n<\/p>\n\n<p>There is also <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/Dense\" rel=\"nofollow noreferrer\">tf.layers.Dense<\/a> : A class which has almost identical attributes as the parameters of tf.layers.dense(), and implements essentially the same operation on the inputs.\n<\/p>\n\n<p>After going through the documentation for both, I fail to see any extra functionality added by using the class version. I think the function implementation should suffice for my use case the skeleton for which is given below.<\/p>\n\n<pre><code>class myNeuralNet:\ndef __init__(self, dim_input_data, dim_output_data): \n    #Member variable for dimension of Input Data Vectors (Number of features...)\n    self.dim_input_data = dim_input_data\n    #Variable for dimension of output labels\n    self.dim_output_data = dim_output_data\n    #TF Placeholder for input data   \n    self.x =  tf.placeholder(tf.float32, [None, 784])\n    #TF Placeholder for labels \n    self.y_ = tf.placeholder(tf.float32, [None, 10])\n    #Container to store all the layers of the network \n    #Containter to hold layers of NN\n    self.layer_list = []\n\ndef addHiddenLayer(self, layer_dim, activation_fn=None, regularizer_fn=None):\n    # Add a layer to the network of layer_dim\n    # append the new layer to the container of layers \n    pass\n\ndef addFinalLayer(self, activation_fn=None, regularizer_fn=None):\n    pass\n\ndef setup_training(self, learn_rate):\n    # Define loss, you might want to store it as self.loss\n    # Define the train step as self.train_step = ..., use an optimizer from tf.train and call minimize(self.loss)\n    pass\n\ndef setup_metrics(self):\n    # Use the predicted labels and compare them with the input labels(placeholder defined in __init__)\n    # to calculate accuracy, and store it as self.accuracy\n    pass\n\n# add other arguments to this function as given below\ndef train(self, sess, max_epochs, batch_size, train_size, print_step = 100):                \n    pass\n<\/code><\/pre>\n\n<p>Can someone give an example of a situation where the class version would be required? \nReferences:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50029121\/how-to-use-tf-layers-classes-instead-of-functions\">Related question<\/a> on SO<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/koaning\/c26b2dd5c2bdeaf6d7479a68bd7023bb\" rel=\"nofollow noreferrer\">Example<\/a> of function usage <\/p>\n","answers":[{"AnswerId":"53195036","CreationDate":"2018-11-07T17:52:04.937","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>I've always used the <code>dense<\/code> because you get the output tensor that you can reuse for next layers.<\/p>\n\n<p>Then it's probably a matter of taste than anything else though.<\/p>\n"},{"AnswerId":"53195867","CreationDate":"2018-11-07T18:46:57.673","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>Using <code>Dense<\/code> has the advantage that you get the \"layer object\" that you can refer back to later. <code>dense<\/code> actually just calls <code>Dense<\/code> and then uses its <code>apply()<\/code> method immediately, discarding the layer object afterwards. Here are two example scenarios where <code>Dense<\/code> would be useful:<\/p>\n\n<ol>\n<li>Accessing variables. Let's say you want to do something with the dense layer weights, e.g. visualize them, use them in some kind of regularization etc. If you used <code>dense<\/code>, you have a problem: The layer object storing the variables was discarded. You can only get them back from the computation graph, which is really annoying and ugly -- see <a href=\"https:\/\/stackoverflow.com\/questions\/45372291\/how-to-get-weights-in-tf-layers-dense\">this question<\/a> for an example. If you created a <code>Dense<\/code> layer object on the other hand you can simply ask for the <code>trainable_variables<\/code> attribute of the layer.<br>\nFurthermore, if you use eager execution I believe you <em>have<\/em> to have explicit variable storage because there is no computational graph -- if you use <code>dense<\/code> the variables would be discarded along with the layer object and your training wouldn't work (but don't quote me on this, I don't know much about eager execution).<\/li>\n<li>Re-using a layer. Let's say you want to apply the layer to some input, and then later apply it to a different input as well. With <code>dense<\/code> you have to use variable scopes and the <code>reuse<\/code> feature which I personally find quite unintuitive, as well as making your code harder to understand. If you used <code>Dense<\/code>, you can simply call the <code>apply<\/code> method of the layer object again.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53195439,"AnswerCount":1,"Tags":"<python><pandas><numpy><keras><deep-learning>","CreationDate":"2018-11-07T18:16:55.780","AcceptedAnswerId":53195670.0,"OwnerUserId":10614772.0,"Title":"Converting array to percentage in Numpy","Body":"<p>I fine tuned a model and return the predictions into the variable result  but I got a array of floats. How can I convert this to percentage and print it with the labels. Also how to return the top 5 predictions.<\/p>\n\n<pre><code>result = model.predict(test_image) \n\n\n\nprint (result)\n\n[[6.4043561e-05 4.0387777e-06 5.7413094e-05 4.3322724e-01 4.8211630e-05\n\n\n 3.2328039e-03 5.4211664e-04 5.6212872e-01 6.7689107e-04 1.8414050e-05]]\n<\/code><\/pre>\n\n<p>My labels are stored in class indices<\/p>\n\n<pre><code>test_labels = test_batches.classes\ntest_labels\n\ntest_batches.class_indices\n{'Name1': 0,\n'Name2': 1,\n'Name3': 2,\n'Name4': 3,\n'Name5': 4,\n'Name6': 5,\n'Name7': 6,\n'Name8': 7,\n'Name9': 8,\n'Name10': 9}\n<\/code><\/pre>\n","answers":[{"AnswerId":"53195670","CreationDate":"2018-11-07T18:34:02.207","ParentId":null,"OwnerUserId":"2722504","Title":null,"Body":"<p>I'm guessing what you want is to get the accuracy of your model as the percentage of correctly classified examples. The values in the array are probably logits (confidence of the model that this particular input example corresponds to each possible label. The largest value in the array correspond to the most likely class (prediction) of your model for the input. If you want to calculate top-5 accuracy, you check (for every input) if one of the 5 largest values in the array correspond to the correct label. If that's the case, you treat this example as correctly classified, then count the total number of correctly classified examples, divide by the total number of examples, and multiply by 100 to get your percentage.<\/p>\n"}]}
+{"QuestionId":53195482,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-07T18:20:10.257","AcceptedAnswerId":53204495.0,"OwnerUserId":7826511.0,"Title":"Tensorflow training, how to prevent training node deletion","Body":"<p>I am using Tensorflow with python for object detection.<\/p>\n\n<p>I want to start training and leave it for a while and keep all training nodes (model-cpk). Standard Tensorflow training seems to delete nodes and only keep the last few nodes. How do I prevent that?<\/p>\n\n<p>Please excuse me if this is the wrong place to ask such questions. I would be oblidged if been told a proper place. Thank you.<\/p>\n","answers":[{"AnswerId":"53205258","CreationDate":"2018-11-08T09:54:59.227","ParentId":null,"OwnerUserId":"8990787","Title":null,"Body":"<p>You can save modelcheckpoints as .hdf5 files are load them again when wanting to predict on test data.\nHope that helps.<\/p>\n"},{"AnswerId":"53204495","CreationDate":"2018-11-08T09:07:13.117","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>You can use the <code>keep_checkpoint_max<\/code> flag to <code>tf.estimator.RunConfig<\/code> in <code>model_main.py<\/code>.\nYou can set it to a very large number to practically save all checkpoints.\nYou should be warned though that depending on the model size and saving frequency, it might fill up your disk (and therefore crash during training).\nYou can change saving frequency by the flags <code>save_checkpoints_steps<\/code> or <code>save_checkpoints_secs<\/code> of <code>RunConfig<\/code>. The default is to use <code>save_checkpoints_secs<\/code>, with a default value of 600 (10 minutes).<\/p>\n"}]}
+{"QuestionId":53195966,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><conv-neural-network><multiclass-classification>","CreationDate":"2018-11-07T18:53:44.030","AcceptedAnswerId":null,"OwnerUserId":7933418.0,"Title":"Filter shape in fully connected layer and output layer in Convolutional Neural Network","Body":"<p>I'm building convolutional neural network for classification of the data into different categories The input data is of shape : 30000, 6, 15, 1 the data has 30000 samples, 15 predictors and 6 possible categories. <\/p>\n\n<p>The weight and bias dictionary I'm using is as follows:<\/p>\n\n<pre><code>weights = {\n    'wc1': tf.get_variable('W0', shape=(3,3,1,8), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc2': tf.get_variable('W1', shape=(3,3,8,12), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc3': tf.get_variable('W2', shape=(3,3,12,16), initializer=tf.contrib.layers.xavier_initializer()), \n    'wc4': tf.get_variable('W3', shape=(3,3,16,20), initializer=tf.contrib.layers.xavier_initializer()),\n    'wd1': tf.get_variable('W4', shape=(4*4*20,20), initializer=tf.contrib.layers.xavier_initializer()), \n    'out': tf.get_variable('W6', shape=(20,n_classes), initializer=tf.contrib.layers.xavier_initializer()), \n}\n\n\nbiases = {\n    'bc1': tf.get_variable('B0', shape=(8), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc2': tf.get_variable('B1', shape=(12), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc3': tf.get_variable('B2', shape=(16), initializer=tf.contrib.layers.xavier_initializer()),\n    'bc4': tf.get_variable('B3', shape=(20), initializer=tf.contrib.layers.xavier_initializer()),\n    'bd1': tf.get_variable('B4', shape=(20), initializer=tf.contrib.layers.xavier_initializer()),\n    'out': tf.get_variable('B5', shape=(6), initializer=tf.contrib.layers.xavier_initializer()),\n}\n<\/code><\/pre>\n\n<p>As expected the output tensor of :<\/p>\n\n<pre><code>def conv_net(x, weights, biases):  \n    conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n    conv1 = maxpool2d(conv1, k=2)\n\n    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n    conv2 = maxpool2d(conv2, k=2)\n\n    conv3 = conv2d(conv2, weights['wc3'], biases['bc3'])\n    conv3 = maxpool2d(conv3, k=2)\n\n    conv4 = conv2d(conv3, weights['wc4'], biases['bc4'])\n    conv4 = maxpool2d(conv4, k=2)\n\n\n    # Fully connected layer\n    # Reshape conv2 output to fit fully connected layer input\n    fc1 = tf.reshape(conv4, [-1, weights['wd1'].get_shape().as_list()[0]])\n    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n    fc1 = tf.nn.relu(fc1)\n    # Output, class prediction \n    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n    return out\n<\/code><\/pre>\n\n<p>When x = x(batch size = 64) is of shape (4,6). <\/p>\n\n<p>But since the labels for the batch os 64 is of shape [64,6] where 6 is the number of categories, the cost function defined as <\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))\n\nwhere pred = conv_net(x, weights, biases)\n<\/code><\/pre>\n\n<p>gives the error : <\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): logits and labels must be broadcastable: logits_size=[4,6] labels_size=[64,6]\n     [[Node: softmax_cross_entropy_with_logits_sg = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Add_1, softmax_cross_entropy_with_logits_sg\/Reshape_1)]]\n<\/code><\/pre>\n\n<p>If my understanding is correct it has to do with the definition of fully connected layer and output layer filter size in weights library. Am I understanding right and if yes, what should be the filter shape in FC and output layers? and what's the underlying logic?<\/p>\n","answers":[]}
+{"QuestionId":53196156,"AnswerCount":1,"Tags":"<python><tensorflow><speech-recognition>","CreationDate":"2018-11-07T19:06:48.567","AcceptedAnswerId":null,"OwnerUserId":10620037.0,"Title":"Different spectrogram between audio_ops and tf.contrib.signal","Body":"<p>I am trying to update the feature extraction pipeline of an speech command recognition model replacing the function <code>audio_ops.audio_spectrogram()<\/code> by <code>tf.contrib.signal.stft()<\/code>. I assumed that they were equivalent, but I am obtaining different spectrogram values with the same input audio. Could someone explain the relation between the two methods, or whether it is possible to obtain the same results using <code>tf.contrib.signal.stft()<\/code>?<\/p>\n\n<p>My code:<\/p>\n\n<p>1) <code>audio_ops<\/code> method:<\/p>\n\n<pre><code>from tensorflow.contrib.framework.python.ops import audio_ops\nimport tensorflow as tf\nimport numpy as np\nfrom tensorflow.python.ops import io_ops\n\n#WAV audio loader\nwav_filename_placeholder_ = tf.placeholder(tf.string, [], name='wav_filename')\nwav_loader = io_ops.read_file(wav_filename_placeholder_)\nsample_rate = 16000\ndesired_samples = 16000 #1 sec audio\nwav_decoder = audio_ops.decode_wav(wav_loader, desired_channels=1, desired_samples=desired_samples)\n\n#Computing the spectrograms\nspectrogram = audio_ops.audio_spectrogram(wav_decoder.audio,\n                                              window_size=320,\n                                              stride=160,\n                                              magnitude_squared=False)\nwith tf.Session() as sess:\n    feed_dict={wav_filename_placeholder_:\"\/&lt;folder_path&gt;\/audio_sample.wav\"}\n    #Get the input audio and the spectrogram\n    audio_ops_wav_decoder_audio, audio_ops_spectrogram = sess.run([wav_decoder.audio, spectrogram], feed_dict)\n<\/code><\/pre>\n\n<p>2) <code>tf.contrib.signal<\/code> method:<\/p>\n\n<pre><code>#Input WAV audio (will be initialized with the same audio signal: wav_decoder.audio )\nsignals = tf.placeholder(tf.float32, [None, None])\n\n#Compute the spectrograms and get the absolute values\nstfts = tf.contrib.signal.stft(signals, \n                               frame_length=320, \n                               frame_step=160, \n                               fft_length=512, \n                               window_fn=None)\nmagnitude_spectrograms = tf.abs(stfts)\nwith tf.Session() as sess:\n    feed_dict = {signals : audio_ops_wav_decoder_audio.reshape(1,16000)}\n    tf_original, tf_stfts, tf_spectrogram, = sess.run([signals, stfts, magnitude_spectrograms], feed_dict)\n<\/code><\/pre>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"56468024","CreationDate":"2019-06-05T21:06:49.073","ParentId":null,"OwnerUserId":"3968485","Title":null,"Body":"<p>Found these helpful comments in github that discuss the differences:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11339#issuecomment-345741527\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11339#issuecomment-345741527<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11339#issuecomment-443553788\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/11339#issuecomment-443553788<\/a><\/p>\n\n<blockquote>\n  <p>You can think of audio_ops.audio_spectrogram and audio_ops.mfcc as\n  \"fused\" ops (like fused batch-norm or fused LSTM cells that TensorFlow\n  has) for the ops in tf.contrib.signal. I think the original motivation\n  of them was that a fused op makes it easier to provide mobile support.\n  I think long term it would be nice if we removed them and provided\n  automatic fusing via XLA, or unified the API to match\n  tf.contrib.signal API, and provided fused keyword arguments to\n  tf.contrib.signal functions, like we do for\n  tf.layers.batch_normalization.<\/p>\n  \n  <p>audio_spectrogram is a C++ implementation of an STFT, while\n  tf.signal.stft uses TensorFlow ops to compute the STFT (and thus has\n  CPU, GPU and TPU support).<\/p>\n  \n  <p>The main cause of difference between them is that audio_spectrogram\n  uses fft2d to compute FFTs while tf.contrib.signal.stft uses Eigen\n  (CPU), cuFFT (GPU), and XLA (TPU). There is another very minor\n  difference, which is that the default periodic Hann window used by\n  each is slightly different. tf.contrib.signal.stft follows\n  numpy\/scipy's definition.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53196255,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T19:13:22.130","AcceptedAnswerId":53197080.0,"OwnerUserId":1245262.0,"Title":"How to find zeros of tensorflow function","Body":"<p>I'm starting to learn Tensorflow and set as an exercise for myself to find the zeros of a function. The approach I took was as follows:<\/p>\n\n<pre><code>x = tf.Variable(0.0, trainable=True)  # Independent variable\ny = 2*tf.pow(x,2) - 6*x + 4           # Function for which to find 0's\nloss = tf.pow(y,2)                    # Function with minima at 0 of y(x)\nopt = tf.train.GradientDescentOptimizer(0.1).minimize(loss) # Optimizer\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(1000): # Minimizing loop\n        print sess.run([x,y, loss])\n        sess.run(opt)\n<\/code><\/pre>\n\n<p>I expected that finding the minima of loss (i.e. y^2) would give me the the zeros of y.  However, when I tried the above code, I got these results:<\/p>\n\n<pre><code>[0.0, 4.0, 16.0]\n[4.8, 21.280005, 452.83856]\n[-51.37921, 5591.9224, 31269596.0]\n[236505.78, 111868550000.0, 1.2514572e+22]\n[-2.1165915e+16, 8.959919e+32, inf]\n[inf, nan, nan]\n[nan, nan, nan]\n<\/code><\/pre>\n\n<p>What did I do wrong? I was expecting to \"discover\" that y had a zero at x=1.<\/p>\n","answers":[{"AnswerId":"53197080","CreationDate":"2018-11-07T20:09:47.297","ParentId":null,"OwnerUserId":"8384504","Title":null,"Body":"<p>Learning rate is too large. Smaller learning rate like 0.01 produces the result below:<\/p>\n\n<pre><code>[0.0, 4.0, 16.0]\n[0.48, 1.5808, 2.4989288]\n[0.6089933, 1.087786, 1.1832783]\n[0.68653125, 0.82346296, 0.6780912]\n[0.7401202, 0.65483475, 0.42880854]\n[0.77992785, 0.5370076, 0.28837714]\n[0.8108626, 0.44982052, 0.2023385]\n[0.83566165, 0.3826909, 0.14645232]\n[0.85600054, 0.3294704, 0.10855074]\n[0.8729749, 0.28632116, 0.08197981]\n[0.8873373, 0.25071096, 0.06285599]\n[0.8996254, 0.2208991, 0.048796415]\n[0.9102352, 0.19564486, 0.03827691]\n[0.91946596, 0.17403984, 0.030289866]\n[0.9275488, 0.15540075, 0.024149394]\n[0.93466556, 0.13920641, 0.019378424]     \n[0.9409614, 0.12504816, 0.015637042]\n[0.94655395, 0.112605095, 0.0126799075]\n[0.95153964, 0.101617336, 0.010326083]\n[0.9559983, 0.09187603, 0.008441205]\n<\/code><\/pre>\n\n<p>This will converge to 1.<\/p>\n\n<p>UPDATE: Description of divergent phenomenon of the original code.<\/p>\n\n<p>As you can see, x coordinates of your results are interpreted as the oscillation around the optimal solution x = 1. More distant x ordinate is, a larger gradient is produced. Finally, loss exceeds the largest value <code>tf.float32<\/code> can express, resulting in <code>inf<\/code>.<\/p>\n"}]}
+{"QuestionId":53196467,"AnswerCount":2,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2018-11-07T19:29:00.847","AcceptedAnswerId":null,"OwnerUserId":1185242.0,"Title":"Can you post process results from Cloud ML's prediction output?","Body":"<p>I have a model for object detection (Faster RCNN from Tensorflow's Object Detection API) running on Google Cloud ML. I also have some code to filter the resulting bounding boxes based on size, aspect ratio etc.<\/p>\n\n<ol>\n<li><p>Is it possible to run this code as part of the prediction process so I don't need to run a separate process to do it afterwards.<\/p><\/li>\n<li><p>Is it possible to limit the number of bounding boxes predicted by the model based on some confidence threshold, as currently outputting a lot of extraneous data.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53203082","CreationDate":"2018-11-08T07:25:47.253","ParentId":null,"OwnerUserId":"4392373","Title":null,"Body":"<ol>\n<li>You can simply add your filter logic to prediction process code and deploy it back.<\/li>\n<li>Yes you can use <strong>min_score_thresh<\/strong> argument in <strong>visualize_boxes_and_labels_on_image_array<\/strong>. use below code<\/li>\n<\/ol>\n\n<blockquote>\n<pre><code>vis_util.visualize_boxes_and_labels_on_image_array(\n        image_np,\n        np.squeeze(boxes),\n        np.squeeze(classes).astype(np.int32),\n        np.squeeze(scores),\n        category_index,\n        use_normalized_coordinates=True,\n        line_thickness=4,\n        min_score_thresh=.9)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>This will result in bounding boxes with more than or equal to 90% confidence score. change it to any value between 0 and 1.<\/p>\n"},{"AnswerId":"53202686","CreationDate":"2018-11-08T06:46:04.660","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>I'll answer (1): we have an Alpha API that will permit this. Please contact cloudml-feedback@google.com for more information.<\/p>\n"}]}
+{"QuestionId":53197231,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network><text-classification>","CreationDate":"2018-11-07T20:22:00.660","AcceptedAnswerId":null,"OwnerUserId":6078821.0,"Title":"Tensorflow - Sparse embedding lookup that remains sparse","Body":"<p>I'm implementing a text classifier with a CNN similar to <a href=\"https:\/\/arxiv.org\/pdf\/1408.5882.pdf\" rel=\"nofollow noreferrer\">Kim 2014<\/a> with Tensorflow. Tensorflow provides <code>tf.nn.embedding_lookup_sparse<\/code>, which allows you to provide the word IDs as a sparse tensor. This is nice, especially for enabling variable length sequences. However, this function requires a \"combination\" step after the lookup, such as \"mean\" or \"sum\". This coerces it back to the dense tensor space. I don't want to do any combination. I want to keep my vectors in the sparse representation, so I can do other convolutions afterwards. Is this possible in TF?<\/p>\n\n<p>EDIT: I want to avoid padding the input prior to the embedding lookup. This is because Tensorflow's embedding lookup generates vectors for the pad value, and its a kludge trying to mask it with zeros <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2373\" rel=\"nofollow noreferrer\">(see here)<\/a>.<\/p>\n","answers":[{"AnswerId":"53197686","CreationDate":"2018-11-07T20:56:15.223","ParentId":null,"OwnerUserId":"567620","Title":null,"Body":"<p>I think there are two points of confusion in the question. Firstly, the combiner operation happens across the set of embedding IDs <em>for each row<\/em> of the sparse indices input <code>sp_ids<\/code>. So if <code>sp_ids<\/code> has a shape of N x 1, then you are \"combining\" just one embedding vector per each row of <code>sp_ids<\/code>, which will just retrieve that embedding vector (which is I think what you are saying you want).<\/p>\n\n<p>Secondly though, the return value <em>is the embedding vector<\/em> for each row of input. The embedding vector itself is a dense vector, by very definition of what the embedding is and what the TensorFlow embedding operations calculate. So this return result will always be dense, and that's what you want. A sparse matrix representation would be horribly inefficient, since the matrix truly will be dense (full of dense embeddings), regardless of whether any 'combiner' operation happens or not.<\/p>\n\n<p>The research paper you linked does not seem to be doing any type of special methodology that would result in a special case of a sparse embedding vector, so I don't see a reason here for expecting or desiring sparse outputs.<\/p>\n\n<p>Maybe I am incorrect, can you provide more details about why you expect the embedding vectors themselves to be sparse vectors? That would be a highly unusual situation if so.<\/p>\n"}]}
+{"QuestionId":53197383,"AnswerCount":1,"Tags":"<python><tensorflow><dataset><conv-neural-network><data-augmentation>","CreationDate":"2018-11-07T20:33:05.477","AcceptedAnswerId":53197652.0,"OwnerUserId":9981271.0,"Title":"Automatically make a composite image for cnn training","Body":"<p>i would like to train a CNN for detection and classification of any kind of signs (mainly laboratory and safety markers) using tensorflow. \nWhile I can gather enough training data for the classification training set, using e.g. The Bing API, I\u2018m struggeling to think about a solution to get enough images for the object detection training set. Since these markers are mostly not public available, I thought I could make a composite of a natrual scene image with the image of the marker itself, to get a training set. Is there any way to do that automatically? \nI looked at tensorflow data augmentation class, but it seems it only provides functionality for simpler data augmentation tasks.<\/p>\n","answers":[{"AnswerId":"53197652","CreationDate":"2018-11-07T20:54:01.037","ParentId":null,"OwnerUserId":"8384504","Title":null,"Body":"<p>You can do it with OpenCV as preprocessing.<\/p>\n\n<p>The algorithm follows:<\/p>\n\n<ol>\n<li>Choose a combination of a natural scene image and a sign image randomly.<\/li>\n<li>Sample random position in the natural scene image where the sign image is pasted.<\/li>\n<li>Paste the sign image at the position.<\/li>\n<li>Obtain the pasted image and the position as a part of training data.<\/li>\n<\/ol>\n\n<p>Step1 and 2 is done with python standard <code>random<\/code> module or <code>numpy<\/code>.<\/p>\n\n<p>Step3 is done with opencv-python. See <a href=\"https:\/\/stackoverflow.com\/questions\/14063070\/overlay-a-smaller-image-on-a-larger-image-python-opencv\">overlay a smaller image on a larger image python OpenCv\n<\/a>.<\/p>\n"}]}
+{"QuestionId":53197659,"AnswerCount":1,"Tags":"<lstm><pytorch>","CreationDate":"2018-11-07T20:54:56.363","AcceptedAnswerId":53197787.0,"OwnerUserId":4819795.0,"Title":"Pytorch passing PackSequence argument to LSTM","Body":"<p>As per my understanding, <code>pack_sequence<\/code> and <code>pack_padded_sequence<\/code> returns a <code>PackedSequence<\/code>, for which its <code>data<\/code> attribute should always be 1 dimension.<\/p>\n\n<p>However, the following code gives the error:\n<code>RuntimeError: input must have 2 dimensions, got 1<\/code>.<\/p>\n\n<pre><code>import torch\nimport torch.nn.utils.rnn as rnn_utils\na = torch.Tensor([1, 2, 3])\nb = torch.Tensor([4, 5])\nc = torch.Tensor([6])\nseq = rnn_utils.pack_sequence([a, b, c])\nlstm = torch.nn.LSTM(\n    input_size=3,\n    hidden_size=3\n)\nlstm(seq)\n<\/code><\/pre>\n\n<p>I've tried <code>batch_first=True<\/code>, that did not work.<\/p>\n\n<p>The source file of error code: <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/nn\/modules\/rnn.py#L127\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/nn\/modules\/rnn.py#L127<\/a><\/p>\n","answers":[{"AnswerId":"53197787","CreationDate":"2018-11-07T21:03:44.347","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>There are a few things wrong with your code. <\/p>\n\n<ol>\n<li>Input size should be 1<\/li>\n<li>LSTM takes a vector at each time step, you are passing scalars and hence the dimension error. <\/li>\n<\/ol>\n\n<p>Below code should fix the problem. I made each input a (1,) size array and changed input size to 1. <\/p>\n\n<pre><code>import torch \nimport torch.nn.utils.rnn as rnn_utils \na = torch.Tensor([[1], [2], [3]]) \nb = torch.Tensor([[4], [5]]) \nc = torch.Tensor([[6]]) \nseq = rnn_utils.pack_sequence([a, b, c]) \nlstm = torch.nn.LSTM(input_size=1, hidden_size=3) \nlstm(seq) \n<\/code><\/pre>\n\n<p>The packed sequence can be multi-dimensional and it depends on the data at each time step. For example in this case <code>seq<\/code> will be <\/p>\n\n<pre><code>PackedSequence(data=tensor([[1.],\n        [4.],\n        [6.],\n        [2.],\n        [5.],\n        [3.]]), batch_sizes=tensor([3, 2, 1]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53197828,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-07T21:07:16.157","AcceptedAnswerId":null,"OwnerUserId":3443786.0,"Title":"Importing data from CSV to tensor flow, Invalid Argument Error","Body":"<p>I am trying to use data i have in CSV files for training.\nTo use them I'm following these instructions: <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/reading_data#_QueueRunner_\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/reading_data#<em>QueueRunner<\/em><\/a><\/p>\n\n<p>since my data are no integers but decimals i have changed <\/p>\n\n<pre><code>record_defaults = [[1], [1], [1], [1], [1]]\n<\/code><\/pre>\n\n<p>to <code>record_defaults = [[1.], [1.], [1.], [1.], [1.]]<\/code>\nso it accepts floats.<\/p>\n\n<p>I still get following error though:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Field 0 in record 0 is not a valid float: 15.5<\/p>\n  \n  <p>[[Node: DecodeCSV = DecodeCSV[OUT_TYPE=[DT_FLOAT, ... , DT_FLOAT], field_delim=\";\", na_value=\"\", select_cols=[], use_quote_delim=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](ReaderReadV2:1, DecodeCSV\/record_defaults_0, ... , DecodeCSV\/record_defaults_0)]]<\/p>\n<\/blockquote>\n\n<p>Which doesn't make sense to me, 15.5 is in fact the first cell from my CSV file and if it isn't a float, what else?<\/p>\n\n<p>any help on how to eat with this would be greatly appreciated<\/p>\n","answers":[{"AnswerId":"53345179","CreationDate":"2018-11-16T20:49:21.747","ParentId":null,"OwnerUserId":"861974","Title":null,"Body":"<p>It appears to be a known issue with Tensorflow when reading CSV using the <code>UTF-8 + BOM<\/code> encoding.  You should use your favorite (non notepad) text editor to convert the file to a normal <code>UTF-8<\/code> encoding then it should read the csv properly.<\/p>\n\n<p>Issue for reference: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8415\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8415<\/a><\/p>\n"}]}
+{"QuestionId":53199099,"AnswerCount":2,"Tags":"<neural-network><keras><lstm><rnn><training-data>","CreationDate":"2018-11-07T22:49:31.627","AcceptedAnswerId":53223568.0,"OwnerUserId":1408483.0,"Title":"Interesting results from LSTM RNN : lagged results for train and validation data","Body":"<p>As an introduction to RNN\/LSTM (stateless) I'm training a model with sequences of 200 days of previous data (X), including things like daily price change, daily volume change, etc and for the labels\/Y I have the % price change from current price to that in 4 months. Basically I want to estimate the market direction, not to be 100% accurate. But I'm getting some odd results...<\/p>\n\n<p>When I then test my model with the training data, I notice the output from the model is a perfect fit when compared to the actual data, it just lags by <em>exactly<\/em> 4 months:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1rvmF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1rvmF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When I shift the data by 4 months, you can see it's a perfect fit. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/11S0C.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/11S0C.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I can obviously understand why the training data would be a very close fit as it has seen it all during training - but why the 4 months lag?<\/p>\n\n<p>It does the same thing with the validation data (note the area I highlighted with the red box for future reference):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9FIO4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9FIO4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Time-shifted:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MoZvY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MoZvY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It's not as close-fitting as the training data, as you'd expect, but still <em>too<\/em> close for my liking - I just don't think it can be this accurate (see the little blip in the red rectangle as an example). I think the model is acting as a naive predictor, I just can't work out how\/why it's possibly doing it.<\/p>\n\n<p>To generate this output from the validation data, I input a sequence of 200 timesteps, but there's nothing in the data sequence that says what the %price change will be in 4 months - it's entirely disconnected, so how is it <em>so<\/em> accurate?  The 4-month lag is obviously another indicator that something's not right here, I don't know how to explain that, but I suspect the two are linked.<\/p>\n","answers":[{"AnswerId":"53223568","CreationDate":"2018-11-09T10:01:54.203","ParentId":null,"OwnerUserId":"1408483","Title":null,"Body":"<p>Okay, I realised my error; the way I was using the model to generate the forecast line was naive. For every date in the graph above, I was getting an output from the model, and then apply the forecasted % change to the actual price for that date - that would give predicted price in 4 months' time.  <\/p>\n\n<p>Given the markets usually only move within a margin of 0-3% (plus or minus) over a 4 month period, that would mean my forecasts was always going to closely mirror the current price, just with a 4 month lag.<\/p>\n\n<p>So at every date the predicted output was being re-based, so the model line would never deviate far from the actual; it'd be the same, but within a margin of 0-3% (plus or minus).<\/p>\n\n<p>Really, the graph isn't important, and it doesn't reflect the way I'll use the output anyway, so I'm going to ditch trying to get a visual representation, and concentrate on trying to find different metrics that lower the validation loss.<\/p>\n"},{"AnswerId":"53521525","CreationDate":"2018-11-28T14:18:34.497","ParentId":null,"OwnerUserId":"9292516","Title":null,"Body":"<p>I tried to explain the observation based on some general underlying concept:<\/p>\n\n<ol>\n<li><p>If you don't provide a time-lagged X input dataset (lagged t-k where k is the time steps), then basically you will be feeding the LSTM with like today's closing price to predict the same today's closing price..in the training stage. The model will (over fit) and behave Exactly as the answer is known already (data leakage)<\/p><\/li>\n<li><p>If the Y is the predicted percentage change (ie. X * (1 + Y%) = 4 months future price), the present value Yvalue predicted really is just the future discounted by the Y%\nso the predicted value will have 4 months shift<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53199675,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-07T23:57:23.607","AcceptedAnswerId":53286280.0,"OwnerUserId":8793975.0,"Title":"Importing tensorflow makes python 3.6.5 error","Body":"<p>Tensorflow used to work on my computer. But now when I try to <code>import tensorflow<\/code> python itself errors out. I am not given a traceback call to tell me what the error  is. I get a window's prompt that says \"Python has stopped working\". When I click \"debug\" all I get is \"An unhandled win32 exception occurred in python.exe\". I've never had a python package actually error out python itself for me, I've always just had a traceback error thrown by python if I didn't install something right. <\/p>\n\n<p>I've tried uninstalling and reinstalling tensorflow (effectively updating from 1.7.0 to 1.12.0) but that has not helped. I'm not sure how to search for a solution to this problem either since I'm not given a traceback or an error code or an error message aside from the very generic one above.<\/p>\n\n<p>I'm currently using python 3.6.5 with tensorflow 1.12.0 (CPU only) installed. My OS is Windows 7 Enterprise 64 bit.  <\/p>\n\n<p>Any ideas? <\/p>\n\n<p>EDIT: The python distro I am using is through Anaconda and I'm trying to run python directly through the anaconda prompt (command line interface). <\/p>\n\n<p>EDIT2: I used the <code>faulthandler<\/code> module to see if I can get a stack trace out of it, and I got a <code>Windows fatal exception: code 0xc0000139<\/code> and a  <code>Windows fatal exception: access violation<\/code>, along with a bunch of lines linking to various <code>frozen importlib._bootstrap<\/code> lines of code in various <code>__init__.py<\/code> modules.<\/p>\n\n<p>EDIT3: For a bit more context, this is on a workplace machine with a lot of security software installed on it.  <\/p>\n","answers":[{"AnswerId":"55010016","CreationDate":"2019-03-05T19:18:25.813","ParentId":null,"OwnerUserId":"7484853","Title":null,"Body":"<p>Please try following Steps<\/p>\n\n<pre><code>conda install -c anaconda qt\nconda update qt pyqt\n<\/code><\/pre>\n\n<p>start again<\/p>\n"},{"AnswerId":"53286280","CreationDate":"2018-11-13T17:11:40.237","ParentId":null,"OwnerUserId":"8793975","Title":null,"Body":"<p>I have solved the issue. The following procedure was used to find and fix the problem:<\/p>\n\n<p>I used the <code>faulthandler<\/code> module to force python to print out a stack trace and recieved a <code>Windows fatal exception: access violation<\/code> error which seems to suggest the problem was indeed a segfault caused by some module used by tensorflow.<\/p>\n\n<p>I tried to fix dependencies by doing a <code>conda update --all<\/code> and then a <code>conda clean --all<\/code> which didn't fix the problem. <\/p>\n\n<p>I noticed though that the problems seems to arise from the <code>h5py<\/code> and <code>keras<\/code> modules so I did <code>pip install --upgrade h5py<\/code> and <code>pip install --upgrade keras<\/code> and <code>pip install --upgrade tensorflow<\/code> and the problem was fixed. I am now using <code>tensorflow version 1.12.0<\/code>, <code>keras version 2.2.4<\/code>, and <code>h5py version 2.8.0<\/code>.<\/p>\n\n<p>The key to solving this problem seems to be the <code>faulthander<\/code> module which showed me which modules (h5py and keras) were leading to the segfault.     <\/p>\n"}]}
+{"QuestionId":53199702,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2018-11-08T00:01:29.673","AcceptedAnswerId":null,"OwnerUserId":10621026.0,"Title":"Why am I getting a dimensional error when I use a 2-dim convolutional input layer?","Body":"<p>I have a Keras deep learning question:<\/p>\n\n<p>I'm building a mnist classifier, but the input layer of my neural network is a 2nd Dimensional Convolutional layer. The training set has 60000 samples, and the testing set has 10000 samples. The dimensions of each image is 28 by 28 pixels. I take the pixel tensor slices and squash it down to a two parameter tensor, divide by 255, and make it a 'float32'.<\/p>\n\n<p>imports: <\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, MaxPooling2D, Conv2D, Dropout, Flatten\nfrom keras.datasets import mnist\n<\/code><\/pre>\n\n<p>Loading the Data:<\/p>\n\n<pre><code>(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\n<\/code><\/pre>\n\n<p>Pre-processing:<\/p>\n\n<pre><code>train_images = train_images.reshape((60000, 28 * 28))\ntrain_images = train_images.astype('float32') \/255\ntest_images = test_images.reshape((10000, 28 * 28))\ntest_images = test_images.astype('float32') \/255\n<\/code><\/pre>\n\n<p>Then I run it through a 2-Dim Convolutional Input Layer:<\/p>\n\n<pre><code>network = Sequential()\nnetwork.add(Conv2D(64, kernel_size = (3, 3), padding='valid', data_format=\"channels_last\", input_shape = (60000, 784)))\n<\/code><\/pre>\n\n<p>Full Model: <\/p>\n\n<pre><code>network = Sequential()\nnetwork.add(Conv2D(64, kernel_size = (3, 3), padding='valid', data_format=\"channels_last\", input_shape = (60000, 784)))\nnetwork.add(MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=\"channels_last\"))\nnetwork.add(Conv2D(64, (2,2), strides=(1, 1), padding='valid',  data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, \nkernel_initializer='glorot_uniform', bias_initializer='zeros', \nkernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, \nkernel_constraint=None, bias_constraint=None))\nnetwork.add(MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None))\nnetwork.add(Flatten())\nnetwork.add(Dense(128, activation= 'relu'))\nnetwork.add(Dropout(0.2, noise_shape=None, seed=7))\nnetwork.add(Dense(10, activation='softmax'))\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer conv2d_7: expected ndim=4, found ndim=3\n<\/code><\/pre>\n\n<p>So I tried to add a channel dimension (to represent black white):<\/p>\n\n<pre><code>network = Sequential()\nnetwork.add(Conv2D(64, kernel_size = (3, 3), padding='valid', \ndata_format=\"channels_last\", input_shape = (60000, 784, 1)))\n<\/code><\/pre>\n\n<p>I still get an error when I try to fit the data, but this time my model compiles<\/p>\n\n<pre><code>ValueError: Error when checking input: expected conv2d_3_input to have 4 dimensions, but got array with shape (60000, 784)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53200400","CreationDate":"2018-11-08T01:43:21.103","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>You first need to make sure that Conv2d takes the input shape as <code>( height , width , number of channels )<\/code> Also a batch size dimension has to be added at the beginning. The whole dimension becomes <code>( batch size , height , width , number of channels )<\/code>. <\/p>\n\n<ul>\n<li><p>Your input shape is <code>( 60000 , 784 , 1 )<\/code>\nHere 60000 is your <strong>batch size<\/strong> and the number of channels is 1.<\/p><\/li>\n<li><p>For the Conv2d , you input shape should be <code>( 28 , 28 ,1 )<\/code>. You <strong>do not need to mention the batch size for Conv2d<\/strong>.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53199729,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras>","CreationDate":"2018-11-08T00:04:31.313","AcceptedAnswerId":null,"OwnerUserId":715386.0,"Title":"Tensorflow Keras composite Model","Body":"<p>I want to make a Model composed of other models which supports various training strategies. The main goals are:<\/p>\n\n<ul>\n<li>One model for input that has a specified <code>input_shape<\/code>, a few layers, and a single output layer<\/li>\n<li>Then <em>N<\/em> additional models all connected to the input model<\/li>\n<li>An additional Model which composes the above smaller models into one large multi-output Model<\/li>\n<\/ul>\n\n<p>I could of course just build this as one large model but I have a specific bunch of training experiments I would like my composite model to support:<\/p>\n\n<ul>\n<li>Train the entire network together<\/li>\n<li>Train <em>1<\/em> to <em>N<\/em> outputs\n\n<ul>\n<li>Only effecting certain paths but propagating changes all the way through the input model<\/li>\n<li>Support training the sub models on their own (which is why I have been thinking to build this as a composite of other models vs. one large one)<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>In pseudocode:<\/p>\n\n<pre><code>def build_model_a(input_layer):\n    ... build model a using input_layer arg as input...\n    model = Model(inputs=input_layer,\n                  outputs=output_layer)\n    model.compile(...)\n    return model\n\n\n\n\ninput_model = build_input_model(args)\nmodel_a = build_model_a(input_model.output)\nmodel_b = build_model_b(input_model.output)\n\ncomposite_model = Model(inputs=input_model.input,\n                        outputs=[model_a.output,\n                                 model_b.output])\n\n# train from input_model through model_a but don't change model_b weights\nmodel_b.trainable = False\ncomposite_model.fit(...)\n\n# train from input_model through model_b but don't change model_a weights\nmodel_a.trainable = False\nmodel_b.trainable = True\ncomposite_model.fit(...)\n\n# train from input_model through model_a but don't change model_b weights\nmodel_b.trainable = True\nmodel_a.trainable = True\ncomposite_model.fit(...)\n\n# Train just input_model\ninput_model.fit(...)\n\n# Eval through entire model\ncomposite_model.predict(xs)\n<\/code><\/pre>\n\n<p>That is my current approach which leads to an error with:<\/p>\n\n<pre><code>AttributeError: 'Model' object has no attribute '_name'\n<\/code><\/pre>\n\n<p>This makes me think I am not going about this correctly so I'm hoping for other advice or references. I'm fairly confident that this can be done with Keras functional API but am struggling to get it right.<\/p>\n\n<p>I'm also not clear, if using this approach with <code>trainable = False<\/code>, is allowed. Based on the warnings from the fit run I need to re-compile these models after changing the trainable flag...<\/p>\n","answers":[{"AnswerId":"53206116","CreationDate":"2018-11-08T10:46:07.187","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p>At least for the first part of the question, you can use models as layers using the functional api, not using their inputs,outputs. Something like this:<\/p>\n\n<pre><code># creation of a model\ndef build_model_a():\n    ... build model create your own input...\n    input_layer = Input(shape=i_shape)\n    model = Model(inputs=input_layer,\n                  outputs=output_layer)\n    model.compile(...)\n    return model\n\n# model creations\ninput_model = build_input_model(args)\nmodel_a = build_model_a()\nmodel_b = build_model_b()\n\n#connect models\ninput_layer = Input(shape=i_shape)\nmiddle = input_model(input_layer)\nout_a = model_a(middle )\nout_b = model_b(middle )\n\ncomposite_model = Model(inputs=input_layer ,\n                        outputs=[out_a ,\n                                 out_b ])\n<\/code><\/pre>\n\n<p>Maybe someone can complete the answer or you will be able to test yourself<\/p>\n"}]}
+{"QuestionId":53199786,"AnswerCount":1,"Tags":"<python><json><keras><data-science>","CreationDate":"2018-11-08T00:11:45.100","AcceptedAnswerId":null,"OwnerUserId":7933418.0,"Title":"Load model stored as json file Keras","Body":"<p>I'm loading the model stored as json file using Keras library model_from_json. Following are the steps I'm executing <\/p>\n\n<pre><code>json_file = open('file.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n<\/code><\/pre>\n\n<p>Its throwing me an error when I execute : <code>loaded_model = model_from_json(loaded_model_json)<\/code><\/p>\n\n<p>Following is the error : TypeError: <strong>init<\/strong>() takes at least 4 arguments (3 given)<\/p>\n\n<p>As per the documentation, I'm on the right track. Not sure where I'm going wrong. Appreciate any recommendations<\/p>\n","answers":[{"AnswerId":"53204607","CreationDate":"2018-11-08T09:14:37.497","ParentId":null,"OwnerUserId":"8990787","Title":null,"Body":"<p>Update:<\/p>\n\n<p>Your code is correct and it should work, you can alternatively try:<\/p>\n\n<pre><code>with open('model_architecture.json', 'r') as f:\n      model = model_from_json(f.read())\n<\/code><\/pre>\n\n<p>If this does not work, check your json file, it might me corrupt?<\/p>\n"}]}
+{"QuestionId":53199932,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving><tensorflow-estimator><tensorflow-hub>","CreationDate":"2018-11-08T00:30:38.850","AcceptedAnswerId":53223967.0,"OwnerUserId":6150310.0,"Title":"Serving retrained Tensorflow Hub module with new features","Body":"<p>I have trained a Tensorflow model using estimators and TF Hub Elmo word embedding + new features, and have been trying to get it working with Tensorflow Serving for predictions. When I try to run a prediction, I get the following error:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>Traceback (most recent call last):\r\n  File \"client.py\", line 49, in &lt;module&gt;\r\n    run(args.host, args.port, args.text, args.model, args.signature_name)\r\n  File \"client.py\", line 29, in run\r\n    response = stub.Predict(request, 10.0)\r\n  File \"\/home\/user\/anaconda3\/envs\/twitter\/lib\/python3.6\/site-packages\/grpc\/_channel.py\", line 500, in __call__\r\n    return _end_unary_response_blocking(state, call, False, None)\r\n  File \"\/home\/user\/anaconda3\/envs\/twitter\/lib\/python3.6\/site-packages\/grpc\/_channel.py\", line 434, in _end_unary_response_blocking\r\n    raise _Rendezvous(state, None, None, deadline)\r\ngrpc._channel._Rendezvous: &lt;_Rendezvous of RPC that terminated with (StatusCode.FAILED_PRECONDITION, Serving signature key \"serving_default\" not found.)&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Here are all the relevant files needed to reproduce the steps:<\/p>\n\n<pre><code>train.py\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>from __future__ import absolute_import\r\nfrom __future__ import division\r\nfrom __future__ import print_function\r\n\r\nimport os\r\nimport numpy as np\r\nimport pandas as pd\r\nimport tensorflow as tf\r\nimport tensorflow_hub as hub\r\nimport json\r\nimport urllib\r\n\r\nfrom sklearn.preprocessing import MultiLabelBinarizer\r\n\r\ntf.logging.set_verbosity(tf.logging.INFO)\r\n\r\ntf.app.flags.DEFINE_integer(\r\n    'steps', 10, 'The number of steps to train a model')\r\ntf.app.flags.DEFINE_string(\r\n    'model_dir', '.\/models\/ckpt\/', 'Dir to save a model and checkpoints')\r\ntf.app.flags.DEFINE_string(\r\n    'saved_dir', '.\/models\/pb\/', 'Dir to save a model for TF serving')\r\nFLAGS = tf.app.flags.FLAGS\r\n\r\n'''\r\nData loading and preprocessing method\r\nDropped rows that do not contain sentiment\r\nCasted text column to all lowercase to normalize tweets\r\nCreated new feature column for tweets containing a website or link\r\nCreated new feature column to check for profanity from a text file\r\n'''\r\n\r\n\r\ndef load_and_preprocess():\r\n    data = pd.read_csv('https:\/\/www.figure-eight.com\/wp-content\/uploads\/2016\/03\/Apple-Twitter-Sentiment-DFE.csv',\r\n                       encoding=\"ISO-8859-1\")\r\n    data['text'] = data['text'].str.lower()\r\n    data = data[data.sentiment.str.contains(\"not_relevant\") == False]\r\n    data['contains_url'] = data['text'].str.contains('http').astype(int)\r\n    profanity = pd.read_fwf('data\/profanity.txt', header=None)\r\n    words = [any(i in words for i in profanity[0].values)\r\n             for words in data['text'].str.split().values]\r\n    words = np.array(words, dtype=np.float32)\r\n    data['contains_profanity'] = words.astype(int)\r\n    sentiment = data['sentiment']\r\n    text = data['text']\r\n    url = data['contains_url']\r\n    profanity = data['contains_profanity']\r\n\r\n    return data, sentiment, text, url, profanity\r\n\r\n\r\ndata, sentiment, text, url, profanity = load_and_preprocess()\r\n\r\n\r\ndef serving_input_receiver_fn():\r\n    feature_spec = {\r\n        \"text\": tf.placeholder(dtype=tf.string, shape=[None]),\r\n        \"url\": tf.placeholder(dtype=tf.bool, shape=[None]),\r\n        \"profanity\": tf.placeholder(dtype=tf.bool, shape=[None]),\r\n    }\r\n    return tf.estimator.export.ServingInputReceiver(feature_spec, feature_spec)\r\n\r\n\r\ndef main(unused_argv):\r\n    train_size = int(len(text) * .8)\r\n    train_text = text[:train_size]\r\n    train_sentiment = sentiment[:train_size]\r\n    train_url = url[:train_size]\r\n    train_profanity = profanity[:train_size]\r\n\r\n    test_text = text[train_size:]\r\n    test_sentiment = sentiment[train_size:]\r\n    test_url = url[train_size:]\r\n    test_profanity = profanity[train_size:]\r\n\r\n    text_embeddings = hub.text_embedding_column(\r\n        \"text\",\r\n        module_spec=\"https:\/\/tfhub.dev\/google\/elmo\/2\", trainable=True\r\n    )\r\n\r\n    encoder = MultiLabelBinarizer()\r\n    encoder.fit_transform(train_sentiment)\r\n    train_encoded = encoder.transform(train_sentiment)\r\n    test_encoded = encoder.transform(test_sentiment)\r\n    num_classes = len(encoder.classes_)\r\n\r\n    multi_label_head = tf.contrib.estimator.multi_label_head(\r\n        num_classes,\r\n        loss_reduction=tf.losses.Reduction.SUM_OVER_BATCH_SIZE\r\n    )\r\n\r\n    estimator = tf.contrib.estimator.DNNEstimator(\r\n        head=multi_label_head,\r\n        hidden_units=[64, 10],\r\n        feature_columns=[text_embeddings],\r\n        model_dir=FLAGS.model_dir\r\n    )\r\n\r\n    # Format our data for the numpy_input_fn\r\n    features = {\r\n        \"text\": np.array(train_text),\r\n        \"url\": np.array(train_url),\r\n        \"profanity\": np.array(train_profanity)\r\n    }\r\n    labels = np.array(train_encoded)\r\n\r\n    train_input_fn = tf.estimator.inputs.numpy_input_fn(\r\n        features,\r\n        labels,\r\n        shuffle=True,\r\n        batch_size=FLAGS.steps,\r\n        num_epochs=10\r\n    )\r\n\r\n    estimator.train(input_fn=train_input_fn)\r\n    estimator.export_savedmodel(\r\n        FLAGS.saved_dir, serving_input_receiver_fn=serving_input_receiver_fn)\r\n\r\n    eval_input_fn = tf.estimator.inputs.numpy_input_fn({\"text\": np.array(\r\n        test_text).astype(np.str)}, test_encoded.astype(np.int32), shuffle=False)\r\n    estimator.evaluate(input_fn=eval_input_fn)\r\n\r\n\r\nif __name__ == \"__main__\":\r\n    tf.app.run()<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<pre><code>client.py\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>import numpy as np\r\nimport tensorflow as tf\r\n\r\nfrom grpc.beta import implementations\r\nimport argparse\r\n\r\nfrom tensorflow_serving.apis import predict_pb2\r\nfrom tensorflow_serving.apis import prediction_service_pb2\r\nfrom tensorflow_serving.apis import classification_pb2\r\n\r\n\r\ndef run(host, port, text, model, signature_name):\r\n    text = \"My cat only chews @apple cords. Such an #AppleSnob.\".encode()\r\n    url = np.array([0]).astype(np.bool)\r\n    profanity = np.array([0]).astype(np.bool)\r\n\r\n    # establish a connection\r\n    channel = implementations.insecure_channel(host, port)\r\n    stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)\r\n\r\n    request = predict_pb2.PredictRequest()\r\n    request.model_spec.name = 'elmo'\r\n    request.model_spec.signature_name = signature_name\r\n\r\n    request.inputs['text'].CopyFrom(tf.make_tensor_proto(text))\r\n    request.inputs['url'].CopyFrom(tf.make_tensor_proto(url))\r\n    request.inputs['profanity'].CopyFrom(tf.make_tensor_proto(profanity))\r\n\r\n    response = stub.Predict(request, 10.0)\r\n    print(response)\r\n\r\n\r\ndef parse_args(parser):\r\n\r\n\r\n    options = parser.parse_args()\r\n    return options\r\n\r\n\r\nif __name__ == \"__main__\":\r\n    parser = argparse.ArgumentParser()\r\n    parser.add_argument(\"--host\", type=str, default=\"localhost\")\r\n    parser.add_argument(\"--port\", type=int, default=8500)\r\n    parser.add_argument(\"--text\", type=str)\r\n    parser.add_argument(\"--model\", type=str)\r\n    parser.add_argument('--signature_name', default='serving_default', type=str)\r\n\r\n    args = parser.parse_args()\r\n    run(args.host, args.port, args.text, args.model, args.signature_name)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<pre><code>Dockerfile\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>FROM ubuntu:18.04\r\n\r\nRUN apt-get update &amp;&amp; apt-get install -y \\\r\n        curl \\\r\n        libcurl3-dev \\\r\n        unzip \\\r\n        wget \\\r\n        &amp;&amp; \\\r\n    apt-get clean &amp;&amp; \\\r\n    rm -rf \/var\/lib\/apt\/lists\/*\r\n\r\n# Installing tensorflow-model-server\r\nRUN TEMP_DEB=\"$(mktemp)\" \\\r\n    &amp;&amp; wget -O \"$TEMP_DEB\" 'http:\/\/storage.googleapis.com\/tensorflow-serving-apt\/pool\/tensorflow-model-server-1.8.0\/t\/tensorflow-model-server\/tensorflow-model-server_1.8.0_all.deb' \\\r\n    &amp;&amp; dpkg -i \"$TEMP_DEB\" \\\r\n    &amp;&amp; rm -f \"$TEMP_DEB\"\r\n\r\n# gRPC port\r\nEXPOSE 8500\r\n\r\n# REST API port\r\nEXPOSE 8501\r\n\r\n# Serving the model\r\nCMD tensorflow_model_server \\\r\n  --port=8500 \\\r\n  --rest_api_port=8501 \\\r\n  --model_name=\"$MODEL_NAME\" \\\r\n  --model_base_path=\"$MODEL_PATH\"<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<pre><code>env.yml\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>name: elmo\r\nchannels:\r\n    - https:\/\/conda.anaconda.org\/menpo\r\n    - conda-forge\r\ndependencies:\r\n    - python=3\r\n    - scikit-learn\r\n    - protobuf\r\n    - openblas\r\n    - scipy\r\n    - numpy\r\n    - pandas\r\n    - pillow\r\n    - h5py\r\n    - pip:\r\n        - grpcio\r\n        - grpcio-tools\r\n        - tensorflow\r\n        - tensorflow-serving-api\r\n        - tensorflow-hub\r\n        - flake8<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<pre><code>Steps to reproduce\n<\/code><\/pre>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>1. conda env create -f env.yml\r\n2. python train.py --steps 100 --saved_dir .\/models\/ --model_dir .\/elmo_ckpt\r\n  - This takes a while to train, so I have provided my repo with the model files already in it\r\n3. docker build --rm -f Dockerfile -t tensorflow-elmo:latest .\r\n4. docker run --rm -v ${PWD}\/models:\/models -e MODEL_NAME='elmo' -e MODEL_PATH='\/models' -p 8500:8500 -p 8501:8501 --name tensorflow-server tensorflow-elmo:latest\r\n5. python client.py --model elmo<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I've tried with different options in the signature_name and the model was named after the Docker container, which I found out fixed some other peoples problems. Here's the repo with the trained model and checkpoint files already available, in the interest of time: <a href=\"https:\/\/gitlab.com\/sadiq.quasem\/twitter_sentiment_analysis\" rel=\"nofollow noreferrer\">Repository with model files<\/a><\/p>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[{"AnswerId":"53223967","CreationDate":"2018-11-09T10:28:14.420","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>The error you're seeing is about a mismatch between the signatures in the SavedModel: the estimator does not write the one you are trying to serve. The <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#cli_to_inspect_and_execute_savedmodel\" rel=\"nofollow noreferrer\">saved_model_cli<\/a> should help you inspect it.<\/p>\n\n<p>Also, your code uses the long-obsolete tf.contrib.estimator and mixes it with tf.estimator. I highly recommend switching to tf.estimator throughout.<\/p>\n\n<p>This seems unrelated to the use of TensorFlow Hub. By the time your estimator exports a model, all module contents have been inlined into it.<\/p>\n"}]}
+{"QuestionId":53200335,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-08T01:31:52.993","AcceptedAnswerId":53202984.0,"OwnerUserId":1316501.0,"Title":"Tensorflow: ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`","Body":"<p>I've created a keras model with a tensorflow backend but am having difficulty exporting my model for use on ML Engine (as a <code>saved_model.pb<\/code>).  Here's what I'm doing:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((data_train, labels_train))\ndataset = dataset.map(lambda x, y: ({'reviews': x}, y))\nval_dataset = tf.data.Dataset.from_tensor_slices((data_test, labels_test))\nval_dataset = val_dataset.map(lambda x, y: ({'reviews': x}, y))\ndataset = dataset.batch(self.batch_size).repeat()  # repeat infinitely\nval_dataset = val_dataset.batch(self.batch_size).repeat()\n<\/code><\/pre>\n\n<p>Then I perform some preprocessing on my <code>Dataset<\/code> objects:<\/p>\n\n<pre><code>dataset = dataset.map(lambda x, y: preprocess_text_and_y(x,y))\nval_dataset = val_dataset.map(lambda x, y: preprocess_text_and_y(x,y))\n<\/code><\/pre>\n\n<p>I build my keras model and call <code>.fit(...)<\/code>.  It all works.<\/p>\n\n<p>Then I try to export my model, with something like this:<\/p>\n\n<pre><code>def export(data_vocab):\n\n    estimator = tf.keras.estimator.model_to_estimator(model)\n\n    def serving():\n        data_table = tf.contrib.lookup.index_table_from_tensor(tf.constant(self.data_vocab),\n                                                                    default_value=0)\n        inputs = {\n            'reviews': tf.placeholder(shape=[1], dtype=tf.string)\n        }\n        preproc = inputs.copy()\n        preproc = preprocess_text(preproc, data_table)\n        return tf.estimator.export.ServingInputReceiver(preproc, inputs)\n\n    estimator.export_savedmodel('.\/test_export', serving)\n<\/code><\/pre>\n\n<p>And unfortunately, I get back:<\/p>\n\n<pre><code>ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`.\n<\/code><\/pre>\n\n<p>I googled around and found this:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/46221420\/how-to-use-tensorflow-dataset-api-in-combination-with-dense-layers\">How to use TensorFlow Dataset API in combination with dense layers<\/a><\/p>\n\n<p>which says I need to call <code>tf.set_shape(...)<\/code>.  I'm preprocessing strings into an array of integers with length 100.  I've tried adding <code>x['reviews'].set_shape([100])<\/code> in my <code>preprocess_text<\/code> function<\/p>\n\n<p>But then that breaks training with:<\/p>\n\n<pre><code>ValueError: Shapes must be equal rank, but are 2 and 1\n<\/code><\/pre>\n\n<p>Any thoughts on how to fix?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53202984","CreationDate":"2018-11-08T07:15:51.600","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>If you set the shape after batching, you will need to set it to <code>[None, 100]<\/code> to include the batch axis:<\/p>\n\n<pre><code>x['reviews'].set_shape([None, 100])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53200395,"AnswerCount":0,"Tags":"<docker><tensorflow><keras><tensorflow-serving>","CreationDate":"2018-11-08T01:42:33.573","AcceptedAnswerId":null,"OwnerUserId":10621363.0,"Title":"got NodeDef mentions attr 'T' not in Op<name=NonMaxSuppressionV3 when request grpc api via tensorserving on deocker","Body":"<p>I have installed Tensorflow Serving on Docker, and built my own keras version yolov3 model via tf.saved_model module. But when I requested my model via Tensorflow Serving gprc api, got this error\uff1a<\/p>\n\n<pre><code>&lt;_Rendezvous of RPC that terminated with:\nstatus = StatusCode.INVALID_ARGUMENT\ndetails = \"NodeDef mentions attr 'T' not in Op&lt;name=NonMaxSuppressionV3; signature=boxes:float, scores:float, max_output_size:int32, iou_threshold:float, score_threshold:float -&gt; selected_indices:int32&gt;; NodeDef: {{node non_max_suppression\/NonMaxSuppressionV3}} = NonMaxSuppressionV3[T=DT_FLOAT, _output_shapes=[[?]], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](boolean_mask\/GatherV2\/_861, boolean_mask_1\/GatherV2\/_863, Const_6, non_max_suppression\/iou_threshold, non_max_suppression\/score_threshold). (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.).\n[[{{node non_max_suppression\/NonMaxSuppressionV3}} = NonMaxSuppressionV3[T=DT_FLOAT, _output_shapes=[[?]], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](boolean_mask\/GatherV2\/_861, boolean_mask_1\/GatherV2\/_863, Const_6, non_max_suppression\/iou_threshold, non_max_suppression\/score_threshold)]]\"\ndebug_error_string = \"{\"created\":\"@1541581883.448168389\",\"description\":\"Error received from peer\",\"file\":\"src\/core\/lib\/surface\/call.cc\",\"file_line\":1017,\"grpc_message\":\"NodeDef mentions attr 'T' not in Op&lt;name=NonMaxSuppressionV3; signature=boxes:float, scores:float, max_output_size:int32, iou_threshold:float, score_threshold:float -&gt; selected_indices:int32&gt;; NodeDef: {{node non_max_suppression\/NonMaxSuppressionV3}} = NonMaxSuppressionV3[T=DT_FLOAT, _output_shapes=[[?]], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](boolean_mask\/GatherV2\/_861, boolean_mask_1\/GatherV2\/_863, Const_6, non_max_suppression\/iou_threshold, non_max_suppression\/score_threshold). (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.).\\n\\t [[{{node non_max_suppression\/NonMaxSuppressionV3}} = NonMaxSuppressionV3[T=DT_FLOAT, _output_shapes=[[?]], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](boolean_mask\/GatherV2\/_861, boolean_mask_1\/GatherV2\/_863, Const_6, non_max_suppression\/iou_threshold, non_max_suppression\/score_threshold)]]\",\"grpc_status\":3}\"\n<\/code><\/pre>\n\n<p>PS: It worked fine when I changed yolov3 model to a simple classification model<\/p>\n\n<p>my env\uff1a<\/p>\n\n<pre><code>**OS Platform and Distribution **:Linux Ubuntu 16.04\n**TensorFlow Serving installed from **: docker image tensorflow\/serving:latest-gpu(Created\"2018-10-23T21:38:54.625545811Z\")\nTensorFlow Serving version:1.11.1\nTensorFlow version:1.12.0rc2\nTensorFlow-Serving-api version:1.11.0\n<\/code><\/pre>\n\n<p>here is where I used NonMaxSuppression:<\/p>\n\n<pre><code>nms_index = tf.image.non_max_suppression(\n            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)\nclass_boxes = K.gather(class_boxes, nms_index)\n<\/code><\/pre>\n\n<p>here are my codes\uff1a<\/p>\n\n<pre><code>#######save keras model########\nperson_detect_model = YOLO()\noutputs = person_detect_model.outputs()\ninputs = person_detect_model.inputs()\n\ndef save_model_to_serving(inputs, outputs, export_version, export_path):\n    for i in range(len(inputs)):\n        inputs[i] = tf.saved_model.utils.build_tensor_info(inputs[i])\n    for i in range(len(outputs)):\n        outputs[i] = tf.saved_model.utils.build_tensor_info(outputs[i])\n\n    signature = tf.saved_model.signature_def_utils.build_signature_def(                                                                        \n        inputs={'x': inputs[0],'shape':inputs[1]}, \n        outputs={'boxes':outputs[0], 'scores':outputs[1], 'classes': outputs[2]},\n        method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)\n\n    export_path = export_path+'\/'+export_version\n    builder = tf.saved_model.builder.SavedModelBuilder(export_path)\n    main_op = tf.group(tf.tables_initializer(), name='main_op')\n    builder.add_meta_graph_and_variables(\n        sess=person_detect_model.sess,                                                                                                                    \n        tags=[tf.saved_model.tag_constants.SERVING],                                                                                             \n        signature_def_map={                                                                                                                      \n            'detect_image': signature,                                                                                                                     \n        },\n        main_op=main_op)\n    builder.save()\n\nwith person_detect_model.graph.as_default():\n    save_model_to_serving(inputs, outputs, '1111', '.\/yolov3')\n\n#######request grpc api########\nchannel = grpc.insecure_channel('127.0.0.1:8500')\nstub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n\nrequest = predict_pb2.PredictRequest()\nrequest.model_spec.name = 'yolov3'\nrequest.model_spec.signature_name = 'detect_image'\n\nfile_path = '\/home\/sw\/Work\/tf\/test2\/0000.jpg'\n\nimg = Image.open(file_path)\nimg = np.array(img)\nimage = pretreat(img)\n\nrequest.inputs['x'].CopyFrom(tf.contrib.util.make_tensor_proto(image))\nrequest.inputs['shape'].CopyFrom(tf.contrib.util.make_tensor_proto([img.shape[0], img.shape[1]]))\n\nresult_future = stub.Predict.future(request, 10)\nexception = result_future.exception()\nif exception:\n    print(exception)\nelse:\n    response = numpy.array([\n    result_future.result().outputs['boxes'].float_val,\n    result_future.result().outputs['scores'].float_val,\n    result_future.result().outputs['classes'].int_val])\n    print response\n<\/code><\/pre>\n\n<p>thx for help<\/p>\n","answers":[]}
+{"QuestionId":53200550,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T02:06:48.413","AcceptedAnswerId":null,"OwnerUserId":5940710.0,"Title":"How to extract t[i,i,:] using tensorflow?","Body":"<p>For example:<\/p>\n\n<pre><code>t = tf.constant(np.array([[[1,2],[3,4]],[[5,6],[7,8]]]))\n# I want to extract h = [[1,2],[7,8]]\n<\/code><\/pre>\n\n<p>How to use tensorflow to do this? Thank you!<\/p>\n","answers":[{"AnswerId":"53204955","CreationDate":"2018-11-08T09:35:57.810","ParentId":null,"OwnerUserId":"10550766","Title":null,"Body":"<p>You can use <code>tf.gather_nd()<\/code>. It gathers slices from the last dimension given some indices. You can read more on that <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<pre><code>import tensorflow as tf\n\n\nt = tf.constant(np.array([[[1,2],[3,4]],[[5,6],[7,8]]]))\nt_res = tf.gather_nd(t, [(0,0),(1,1)])\nsess = tf.Session()\nh = sess.run(t_res) # &gt;&gt;&gt;[[1,2],[7,8]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53200839,"AnswerCount":2,"Tags":"<python><image-processing><keras><neural-network><deep-learning>","CreationDate":"2018-11-08T02:46:06.443","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"How to feed in images into the model Keras","Body":"<p>I have a dataset for the self-driving car. My <code>X<\/code> values are the names of the images. Example would be <\/p>\n\n<pre><code>array([['img_2.png'],\n       ['img_3.png'],\n       ['img_4.png'],\n       ...,\n       ['img_6405.png'],\n       ['img_6406.png'],\n       ['img_6407.png']], dtype=object)\n<\/code><\/pre>\n\n<p>I found out that model performs good if we have some kind of <code>batch_generator<\/code>. I found that template code. <\/p>\n\n<pre><code>def batch_generator(image_paths, steering_ang, batch_size, istraining):\n\n  while True:\n    batch_img = []\n    batch_steering = []\n\n    for i in range(batch_size):\n      random_index = random.randint(0, len(image_paths) - 1)\n\n      if istraining:\n        im = random_augment(image_paths[random_index])\n        steering = steering_ang[random_index]\n      else:\n        im = mpimg.imread(image_paths[random_index])\n        steering = steering_ang[random_index]\n\n      im = img_preprocess(im)\n      batch_img.append(im)\n      batch_steering.append(steering)\n    yield (np.asarray(batch_img), np.asarray(batch_steering))  \n<\/code><\/pre>\n\n<p>I changed this function to for my use but when i apply it.<\/p>\n\n<pre><code>x_train_gen, y_train_gen = next(batch_generator(X_train, y_train, 1, 1))\nx_valid_gen, y_valid_gen = next(batch_generator(X_valid, y_valid, 1, 0))\n<\/code><\/pre>\n\n<p>I get the following error <code>TypeError: Object does not appear to be a 8-bit string path or a Python file-like object<\/code>. I understand the error, image is not an array but a string. How can i convert string of the image path to the array <\/p>\n","answers":[{"AnswerId":"53203520","CreationDate":"2018-11-08T08:01:39.193","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>I don't know what you're doing in the <code>img_preprocess()<\/code> function but from what I see there are 2 possible problems:<\/p>\n\n<ol>\n<li><p>You have to append the path to the image to the image name: <code>path_to_image = path_to_image_dir + '\/' + image<\/code><\/p><\/li>\n<li><p>You have to actually open the image to get it's array. You can use Pillow or OpenCV:\n<code>PIL.Image.open(path_to_image)<\/code> or <code>cv2.imread(path_to_image)<\/code><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"57351002","CreationDate":"2019-08-05T00:02:53.850","ParentId":null,"OwnerUserId":"11623095","Title":null,"Body":"<p>It's because at some point you converted <code>X_train<\/code> and <code>y_train<\/code> into numpy arrays instead of image paths.<\/p>\n\n<p>That's why python is complaining. You were probably doing something else with the code which needed you to convert the entire training dataset, but now you don't need that because you have <code>imread()<\/code> in the <code>batch_generator<\/code> function. I'd go back to earlier in the code and recreate <code>X_train<\/code> and <code>y_train<\/code> as the file paths to the images, and then re-run this portion of the code.<\/p>\n"}]}
+{"QuestionId":53201297,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T03:50:42.830","AcceptedAnswerId":53201901.0,"OwnerUserId":7177402.0,"Title":"how to perform element wise multiply for two 4d tensor of shape [1,16,16,3] and [1,4,4,3]","Body":"<p>I want to multiply element-wise two 4d tensor and store the result in a 4d tensor.I have a tensor say 'A' of shape [batch_size,16,16,3] and tensor 'B' of shape [batch_size,4,4,3]. I want to perform a sort of tiling operation such that each 4x4x3 block of tensor 'A' does element wise multiplication with tensor 'B'.\nThe result is stored in a tensor 'C' of shape same as tensor 'A' i.e. [batch_size,16,16,3]. AS tensor 'B' has height and width are multiples of height and width of tensor 'A'. I should concatenate or accumulate or assign the results of 4 elementwise multiplication in tensor 'C'.<\/p>\n","answers":[{"AnswerId":"53201901","CreationDate":"2018-11-08T05:13:41.640","ParentId":null,"OwnerUserId":"5737630","Title":null,"Body":"<p>This can be done by reshaping and using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">broadcasting<\/a>.<\/p>\n\n<pre><code>import tensorflow as tf\n\nbatch_size = 50\n\n# These are the tensors we want to multiply\na = tf.random_normal(shape=(batch_size,16,16,3))\nb = tf.random_normal(shape=(batch_size,4,4,3))\n\n# We reshape the tensors so that they their shapes are\n# compatible for broadcasting\na_reshape = tf.reshape(a,(-1,4,4,4,4,3))\nb_reshape = tf.reshape(b,(-1,4,1,4,1,3))\n\n# we perform the multiplication. Each dimenshion of size 1 in `b_reshape`\n# will be tiled to have the corresponding size 4 of `a_reshape`\nc_reshape = a_reshape*b_reshape\n\n# convert result to required shape\nc = tf.reshape(c_reshape,(-1,16,16,3))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53201534,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-11-08T04:23:20.517","AcceptedAnswerId":53212781.0,"OwnerUserId":10621791.0,"Title":"Calling cuda() with async results in SyntaxError","Body":"<p>I'm trying to run this PyTorch code:<\/p>\n\n<pre><code>for i, (input, target) in enumerate(train_loader):\n\n    input = input.float().cuda(async=True)\n    target = target.cuda(async=True)\n    input_var = torch.autograd.Variable(input)\n    target_var = torch.autograd.Variable(target)\n\n    output = model(input_var)\n<\/code><\/pre>\n\n<p>But when I try I am getting this error message:<\/p>\n\n<pre><code>input = input.float().cuda(async=True)\n                               ^\nSyntaxError: invalid syntax\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>What am I doing wrong? I already installed cuda.<\/p>\n","answers":[{"AnswerId":"53212781","CreationDate":"2018-11-08T17:07:07.420","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p><strong>Your code does not work because:<\/strong><\/p>\n\n<ul>\n<li><p><code>async<\/code> is a reserved keyword in python which cannot be used in that way, that is why you get the <code>SyntaxError<\/code><\/p><\/li>\n<li><p><code>cuda()<\/code> itself also has no argument <code>async<\/code>. The constructor looks like this:<\/p>\n\n<blockquote>\n  <p><strong>cuda<\/strong>(<em>device=None, non_blocking=False<\/em>) \u2192 Tensor<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p><strong>What you can do:<\/strong><\/p>\n\n<ul>\n<li><p>Just calling <code>cuda()<\/code> without any arguments should work fine.<\/p><\/li>\n<li><p>There are two arguments (<code>device<\/code> and <code>non_blocking<\/code>) you can use to call <code>cuda()<\/code>.<br>\nYou did not write what you want to do, but <code>non_blocking<\/code> might be what you were looking for:<\/p>\n\n<blockquote>\n  <ul>\n  <li><code>non_blocking<\/code> <em>(bool):<\/em> <br>If <code>True<\/code> and the source is in pinned memory, the\n  copy will be asynchronous with respect to the host. Otherwise, the\n  argument has no effect. Default: <code>False<\/code>.<\/li>\n  <\/ul>\n<\/blockquote><\/li>\n<li><p>Always good to take a look at the Documentation:<br> <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.cuda\" rel=\"noreferrer\">https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.cuda<\/a><\/p><\/li>\n<\/ul>\n\n<p><br><\/p>\n\n<p><hr>\n<em>As an add-on: If you are interested in what <code>async<\/code> is actually used for you can take a look here:\n<a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0492\/#new-syntax\" rel=\"noreferrer\">https:\/\/www.python.org\/dev\/peps\/pep-0492\/#new-syntax<\/a><\/em><\/p>\n"},{"AnswerId":"58280213","CreationDate":"2019-10-08T04:31:19.883","ParentId":null,"OwnerUserId":"11284323","Title":null,"Body":"<p>There was an <code>async<\/code> parameter but it was deprecated now, because <code>async<\/code> became a reserved word in Python 3.7. Details are included in this issue <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/3431\" rel=\"nofollow noreferrer\">rename .cuda(async=..) parameters<\/a>. You can use <code>non_blocking<\/code> as an alternative.<\/p>\n"}]}
+{"QuestionId":53201921,"AnswerCount":1,"Tags":"<python><pytorch><tensor>","CreationDate":"2018-11-08T05:15:53.920","AcceptedAnswerId":53202025.0,"OwnerUserId":3131604.0,"Title":"torch.Tensor() with requires_grad parameter","Body":"<p>I can't use torch.Tensor() with requires_grad parameter (torch version : 0.4.1)<\/p>\n\n<p>without requires_grad :<\/p>\n\n<pre><code>x = torch.Tensor([[.5, .3, 2.1]])\nprint(x)\n&gt; tensor([[0.5000, 0.3000, 2.1000]])\n<\/code><\/pre>\n\n<p>with requires_grad=True or requires_grad=False :<\/p>\n\n<pre><code>x = torch.Tensor([[.5, .3, 2.1]], requires_grad=False)\nprint(x)\nTraceback (most recent call last):\n  File \"D:\/_P\/dev\/ai\/pytorch\/notes\/tensor01.py\", line 4, in &lt;module&gt;\n    x = torch.Tensor([[.5, .3, 2.1]], requires_grad=False)\nTypeError: new() received an invalid combination of arguments - got (list, requires_grad=bool), but expected one of:\n * (torch.device device)\n * (torch.Storage storage)\n * (Tensor other)\n * (tuple of ints size, torch.device device)\n      didn't match because some of the keywords were incorrect: requires_grad\n * (object data, torch.device device)\n      didn't match because some of the keywords were incorrect: requires_grad\n<\/code><\/pre>\n","answers":[{"AnswerId":"53202025","CreationDate":"2018-11-08T05:30:54.017","ParentId":null,"OwnerUserId":"4313195","Title":null,"Body":"<p>You are creating the tensor <code>x<\/code> by using the <code>torch.Tensor<\/code> class constructor which doesn't take the <code>requires_grad<\/code> flag. Instead, you want to use <code>torch.tensor()<\/code> (lowercase 't') method<\/p>\n\n<pre><code>x = torch.tensor([[.5, .3, 2.1]], requires_grad=False)\n<\/code><\/pre>\n\n<p>Edit: adding a link to docs: <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html\" rel=\"nofollow noreferrer\">torch.Tensor<\/a><\/p>\n"}]}
+{"QuestionId":53202283,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard><eager>","CreationDate":"2018-11-08T06:04:14.667","AcceptedAnswerId":null,"OwnerUserId":4450159.0,"Title":"Tensorflow eager mode support for text summary","Body":"<p>I am trying to make a summary of a text using tensorflow with eager mode enabled, for this I am using this code:<\/p>\n\n<pre><code>writer = summary_ops_v2.create_file_writer('some_path', flush_millis=10000)\nwriter.set_as_default()\n\ntensor = tf.convert_to_tensor(\"Some text\")\nmeta = tf.SummaryMetadata()\nmeta.plugin_data.plugin_name = \"text\"\n\nwith summary_ops_v2.always_record_summaries():\n   summary_ops_v2.generic(name=\"text\", tensor=tensor)\n<\/code><\/pre>\n\n<p>but when reviewing tensorboard the event is created, but the text is not being saved (<a href=\"https:\/\/i.stack.imgur.com\/WK8De.png\" rel=\"nofollow noreferrer\">Tensorboard image<\/a>)<\/p>\n\n<p>Try using <code>tf.make_tensor_proto<\/code> but it did not work either.<\/p>\n\n<pre><code>writer = summary_ops_v2.create_file_writer('some_path', flush_millis=10000)\nwriter.set_as_default()\n\ntensor = tf.make_tensor_proto(\"Some text\", dtype=tf.string)\nmeta = tf.SummaryMetadata()\nmeta.plugin_data.plugin_name = \"text\"\n\nwith summary_ops_v2.always_record_summaries():\n   summary_ops_v2.generic(name=\"text\", tensor=tensor)\n<\/code><\/pre>\n","answers":[{"AnswerId":"55421067","CreationDate":"2019-03-29T15:49:19.647","ParentId":null,"OwnerUserId":"9762143","Title":null,"Body":"<p>I think this might have been a bug. I updated to tensorflow 1.13.1 and the summaries started been displayed correctly.<\/p>\n"}]}
+{"QuestionId":53202362,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><slurm>","CreationDate":"2018-11-08T06:13:46.957","AcceptedAnswerId":54925839.0,"OwnerUserId":4564882.0,"Title":"Scale a keras training using horovod and slurm","Body":"<p>I have this code on keras library used to train an alexnet model on MNIST dataset.<\/p>\n\n<p>I want to scale the training on a cluster running Slurm as workload manager and horovod (<a href=\"https:\/\/github.com\/uber\/horovod\" rel=\"nofollow noreferrer\">https:\/\/github.com\/uber\/horovod<\/a>) for distributed training.<\/p>\n\n<p>The code contains a function that define the alexnet layers and a main function that load the MNIST data and prepare it to be trained on the ALexNet model. \nThe code contains only keras code. <\/p>\n\n<pre><code>import ....\n\ndef alexnet_model(img_shape=(28, 28, 1), n_classes=10, l2_reg=0.):\n\n    alexnet = Sequential()\n    alexnet.add(Conv2D(96, (11, 11), input_shape=img_shape,\n        padding='same', kernel_regularizer=l2(l2_reg)))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Conv2D(256, (5, 5), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(512, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(1024, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(1024, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Flatten())\n    alexnet.add(Dense(3072))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(4096))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(n_classes))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('softmax'))\n    return alexnet\n\nif __name__ == \"__main__\":\n\n\n    batch_size = 32\n    num_classes = 10\n    epochs = 3\n\n    # input image dimensions\n    img_rows, img_cols = 28, 28\n\n    # the data, split between train and test sets\n    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n\n    if K.image_data_format() == 'channels_first':\n        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n        input_shape = (1, img_rows, img_cols)\n    else:\n        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)\n        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)\n        input_shape = (img_rows, img_cols, 1)\n\n    x_train = x_train.astype('float32')\n    x_test = x_test.astype('float32')\n    x_train \/= 255\n    x_test \/= 255\n    print('x_train shape:', x_train.shape)\n    print(x_train.shape[0], 'train samples')\n    print(x_test.shape[0], 'test samples')\n\n    # convert class vectors to binary class matrices\n    y_train = keras.utils.to_categorical(y_train, num_classes)\n    y_test = keras.utils.to_categorical(y_test, num_classes)\n\n    model = alexnet_model()\n\n    model.compile(loss=keras.losses.categorical_crossentropy,\n                optimizer=keras.optimizers.Adadelta(),\n                metrics=['accuracy'])\n\n    checkpoint = ModelCheckpoint(filepath='alexnet_mnist_checkpoint.hdf5')\n\n    history= model.fit(x_train, y_train,\n            batch_size=batch_size,\n            epochs=epochs,\n            verbose=1,\n            validation_data=(x_test, y_test),\n            callbacks=[tbCallback, checkpoint])\n    score = model.evaluate(x_test, y_test, verbose=0)\n}\n<\/code><\/pre>\n\n<p>I want to know which lines should I add to scale this code on Slurm and Horovod?<\/p>\n","answers":[{"AnswerId":"54925839","CreationDate":"2019-02-28T12:33:07.643","ParentId":null,"OwnerUserId":"9320434","Title":null,"Body":"<p>The changes that needs to be made is already mentioned in the link <a href=\"https:\/\/github.com\/horovod\/horovod\" rel=\"nofollow noreferrer\">https:\/\/github.com\/horovod\/horovod<\/a>. It basically boils down to changes in epochs, learning rate, callbacks,optimizer, checkpoint saving and some initialization &amp; configuration settings. Please find the horovod enabled version of your code below:<\/p>\n\n<pre><code>import keras\nfrom keras.datasets import mnist\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras import backend as K\nfrom keras.regularizers import l2\nfrom keras.layers import BatchNormalization, Activation, ZeroPadding2D\nimport math\nimport tensorflow as tf\nimport horovod.keras as hvd\n\n# Horovod: initialize Horovod.\nhvd.init()\n\n# Horovod: pin GPU to be used to process local rank (one GPU per process)\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nconfig.gpu_options.visible_device_list = str(hvd.local_rank())\nK.set_session(tf.Session(config=config))\n\ndef alexnet_model(img_shape=(28, 28, 1), n_classes=10, l2_reg=0.):\n\n    alexnet = Sequential()\n    alexnet.add(Conv2D(96, (11, 11), input_shape=img_shape,\n        padding='same', kernel_regularizer=l2(l2_reg)))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Conv2D(256, (5, 5), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(512, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(1024, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n\n    alexnet.add(ZeroPadding2D((1, 1)))\n    alexnet.add(Conv2D(1024, (3, 3), padding='same'))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(MaxPooling2D(pool_size=(2, 2)))\n    alexnet.add(Flatten())\n    alexnet.add(Dense(3072))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(4096))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('relu'))\n    alexnet.add(Dropout(0.5))\n    alexnet.add(Dense(n_classes))\n    alexnet.add(BatchNormalization())\n    alexnet.add(Activation('softmax'))\n    return alexnet\n\nif __name__ == \"__main__\":\n\n\n    batch_size = 32\n    num_classes = 10\n    epochs = int(math.ceil(3.0 \/ hvd.size()))\n\n    # input image dimensions\n    img_rows, img_cols = 28, 28\n\n    # the data, split between train and test sets\n    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n\n    if K.image_data_format() == 'channels_first':\n        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n        input_shape = (1, img_rows, img_cols)\n    else:\n        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)\n        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)\n        input_shape = (img_rows, img_cols, 1)\n\n    x_train = x_train.astype('float32')\n    x_test = x_test.astype('float32')\n    x_train \/= 255\n    x_test \/= 255\n    print('x_train shape:', x_train.shape)\n    print(x_train.shape[0], 'train samples')\n    print(x_test.shape[0], 'test samples')\n\n    # convert class vectors to binary class matrices\n    y_train = keras.utils.to_categorical(y_train, num_classes)\n    y_test = keras.utils.to_categorical(y_test, num_classes)\n\n    model = alexnet_model()\n\n    model.compile(loss=keras.losses.categorical_crossentropy,\n                optimizer=hvd.DistributedOptimizer(keras.optimizers.Adadelta(1.0 * hvd.size())),\n                metrics=['accuracy'])\n\n    callbacks=[hvd.callbacks.BroadcastGlobalVariablesCallback(0)]\n\n# Horovod: save checkpoints only on worker 0 to prevent other workers from corrupting them.\n    if hvd.rank() == 0:\n        callbacks.append(keras.callbacks.ModelCheckpoint(filepath='alexnet_mnist_checkpoint.hdf5'))\n\n    history= model.fit(x_train, y_train,\n            batch_size=batch_size,\n            epochs=epochs,\n            verbose=1,\n            validation_data=(x_test, y_test),\n            callbacks=callbacks)\n    score = model.evaluate(x_test, y_test, verbose=0)\n<\/code><\/pre>\n\n<p>After changing the code, please run it with mpirun\/mpiexec so that horovod can effectively communicate with its processes. For eg: <code>mpirun -hosts &lt;hostname0,hostname1&gt; -np 2 -genv OMP_NUM_THREADS 4 python alexnet_mnist_horovod.py<\/code>. Please note that you need to have OpenMPI\/IntelMPI already installed on your cluster. Example given is for IntelMPI. Please change accordingly if you are using OpenMPI<\/p>\n"}]}
+{"QuestionId":53202440,"AnswerCount":1,"Tags":"<python><numpy><pytorch><tensor>","CreationDate":"2018-11-08T06:21:19.140","AcceptedAnswerId":53202742.0,"OwnerUserId":8898218.0,"Title":"is there any difference between matmul and usual multiplication of tensors","Body":"<p>I am confused between the multiplication between two tensors using * and matmul.\nBelow is my code<\/p>\n\n<pre><code>import torch\ntorch.manual_seed(7)\nfeatures = torch.randn((2, 5))\nweights = torch.randn_like(features)\n<\/code><\/pre>\n\n<p>here, i want to multiply weights and features. so, one way to do it is as follows<\/p>\n\n<pre><code>print(torch.sum(features * weights))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor(-2.6123)\n<\/code><\/pre>\n\n<p>Another way to do is using matmul<\/p>\n\n<pre><code>print(torch.mm(features,weights.view((5,2))))\n<\/code><\/pre>\n\n<p>but, here output is <\/p>\n\n<pre><code>tensor([[ 2.8089,  4.6439],\n        [-2.3988, -1.9238]])\n<\/code><\/pre>\n\n<p>What i don't understand here is that why <code>matmul<\/code> and usual multiplication are giving different outputs, when both are same. Am i doing anything wrong here?<\/p>\n\n<p>Edit: When, i am using feature of shape <code>(1,5)<\/code> both * and <code>matmul<\/code> outputs are same. \nbut, its not the same when the shape is <code>(2,5)<\/code>.<\/p>\n","answers":[{"AnswerId":"53202742","CreationDate":"2018-11-08T06:51:41.043","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>When you use <code>*<\/code>, the multiplication is elementwise, when you use <code>torch.mm<\/code> it is matrix multiplication.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>a = torch.rand(2,5)\nb = torch.rand(2,5)\nresult = a*b \n<\/code><\/pre>\n\n<p><code>result<\/code> will be shaped the same as <code>a<\/code> or <code>b<\/code> i.e <code>(2,5)<\/code>\nwhereas considering operation <\/p>\n\n<pre><code>result = torch.mm(a,b)\n<\/code><\/pre>\n\n<p>It will give a size mismatch error, as this is proper matrix multiplication (as we study in linear algebra) and <code>a.shape[1] != b.shape[0]<\/code>. When you apply the view operation in <code>torch.mm<\/code> you are trying to match the dimensions. <\/p>\n\n<p>In the special case of the shape in some particular dimension being 1, it becomes a dot product and hence <code>sum (a*b)<\/code> is same as <code>mm(a, b.view(5,1))<\/code><\/p>\n"}]}
+{"QuestionId":53202563,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><artificial-intelligence>","CreationDate":"2018-11-08T06:33:34.490","AcceptedAnswerId":null,"OwnerUserId":3492226.0,"Title":"Tensorflow - Nan loss and constant accuracy when training","Body":"<p>as the title says, I am trying to train a neural network to predict outcomes, and I can't figure out what is wrong with my model. I keep getting the exact same accuracy level, and the loss is Nan. I'm so confused... I have looked at other similar questions and still can't seem to get it working. My code for the model and training is below:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport tensorflow as tf\nimport urllib.request as request\nimport matplotlib.pyplot as plt\nfrom FlowersCustom import get_MY_data\n\ndef get_data():\n    IRIS_TRAIN_URL = \"http:\/\/download.tensorflow.org\/data\/iris_training.csv\"\n    IRIS_TEST_URL = \"http:\/\/download.tensorflow.org\/data\/iris_test.csv\"\n\n    names = ['sepal-length', 'sepal-width', 'petal-length', 'petal-width', 'species']\n    train = pd.read_csv(IRIS_TRAIN_URL, names=names, skiprows=1)\n    test = pd.read_csv(IRIS_TEST_URL, names=names, skiprows=1)\n\n    # Train and test input data\n    Xtrain = train.drop(\"species\", axis=1)\n    Xtest = test.drop(\"species\", axis=1)\n\n    # Encode target values into binary ('one-hot' style) representation\n    ytrain = pd.get_dummies(train.species)\n    ytest = pd.get_dummies(test.species)\n\n    return Xtrain, Xtest, ytrain, ytest\n\n\ndef create_graph(hidden_nodes):\n    # Reset the graph\n    tf.reset_default_graph()\n\n    # Placeholders for input and output data\n    X = tf.placeholder(shape=Xtrain.shape, dtype=tf.float64, name='X')\n    y = tf.placeholder(shape=ytrain.shape, dtype=tf.float64, name='y')\n\n    # Variables for two group of weights between the three layers of the network\n    print(Xtrain.shape, ytrain.shape)\n    W1 = tf.Variable(np.random.rand(Xtrain.shape[1], hidden_nodes), dtype=tf.float64)\n    W2 = tf.Variable(np.random.rand(hidden_nodes, ytrain.shape[1]), dtype=tf.float64)\n\n    # Create the neural net graph\n    A1 = tf.sigmoid(tf.matmul(X, W1))\n    y_est = tf.sigmoid(tf.matmul(A1, W2))\n\n    # Define a loss function\n    deltas = tf.square(y_est - y)\n    loss = tf.reduce_sum(deltas)\n\n    # Define a train operation to minimize the loss\n    # optimizer = tf.train.GradientDescentOptimizer(0.005)\n    optimizer = tf.train.AdamOptimizer(0.001)\n    opt = optimizer.minimize(loss)\n\n    return opt, X, y, loss, W1, W2, y_est\n\n\ndef train_model(hidden_nodes, num_iters, opt, X, y, loss, W1, W2, y_est):\n    # Initialize variables and run session\n    init = tf.global_variables_initializer()\n    sess = tf.Session()\n    sess.run(init)\n    losses = []\n\n    # Go through num_iters iterations\n    for i in range(num_iters):\n        sess.run(opt, feed_dict={X: Xtrain, y: ytrain})\n        local_loss = sess.run(loss, feed_dict={X: Xtrain.values, y: ytrain.values})\n        losses.append(local_loss)\n        weights1 = sess.run(W1)\n        weights2 = sess.run(W2)\n\n        y_est_np = sess.run(y_est, feed_dict={X: Xtrain.values, y: ytrain.values})\n        correct = [estimate.argmax(axis=0) == target.argmax(axis=0)\n                   for estimate, target in zip(y_est_np, ytrain.values)]\n        acc = 100 * sum(correct) \/ len(correct)\n\n        if i % 10 == 0:\n            print('Epoch: %d, Accuracy: %.2f, Loss: %.2f' % (i, acc, local_loss))\n\n    print(\"loss (hidden nodes: %d, iterations: %d): %.2f\" % (hidden_nodes, num_iters, losses[-1]))\n    sess.close()\n    return weights1, weights2\n\n\ndef test_accuracy(weights1, weights2):\n    X = tf.placeholder(shape=Xtest.shape, dtype=tf.float64, name='X')\n    y = tf.placeholder(shape=ytest.shape, dtype=tf.float64, name='y')\n    W1 = tf.Variable(weights1)\n    W2 = tf.Variable(weights2)\n    A1 = tf.sigmoid(tf.matmul(X, W1))\n    y_est = tf.sigmoid(tf.matmul(A1, W2))\n\n    # Calculate the predicted outputs\n    init = tf.global_variables_initializer()\n    with tf.Session() as sess:\n        sess.run(init)\n        y_est_np = sess.run(y_est, feed_dict={X: Xtest, y: ytest})\n\n    # Calculate the prediction accuracy\n    correct = [estimate.argmax(axis=0) == target.argmax(axis=0)\n               for estimate, target in zip(y_est_np, ytest.values)]\n    accuracy = 100 * sum(correct) \/ len(correct)\n    print('final accuracy: %.2f%%' % accuracy)\n\n\ndef get_inputs_and_outputs(train, test, output_column_name):\n    Xtrain = train.drop(output_column_name, axis=1)\n    Xtest = test.drop(output_column_name, axis=1)\n    ytrain = pd.get_dummies(getattr(train, output_column_name))\n    ytest = pd.get_dummies(getattr(test, output_column_name))\n\n    return Xtrain, Xtest, ytrain, ytest\n\n\n\n\nif __name__ == '__main__':\n\n    train, test = get_MY_data('output')\n\n    Xtrain, Xtest, ytrain, ytest = get_inputs_and_outputs(train, test, 'output')#get_data()\n    # Xtrain, Xtest, ytrain, ytest = get_data()\n\n    hidden_layers = 10\n    num_epochs = 500\n\n\n    opt, X, y, loss, W1, W2, y_est = create_graph(hidden_layers)\n    w1, w2 = train_model(hidden_layers, num_epochs, opt, X, y, loss, W1, W2, y_est)\n    # test_accuracy(w1, w2)\n<\/code><\/pre>\n\n<p>Here is a screenshot of what the training is printing out:\n<a href=\"https:\/\/i.stack.imgur.com\/Kg7gL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Kg7gL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And this is a screenshot of the Pandas Dataframe that I am using for the input data (5 columns of floats):\n<a href=\"https:\/\/i.stack.imgur.com\/DBakr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DBakr.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And finally, here is the Pandas Dataframe that I am using for the expected outputs (1 column of either -1 or 1):\n<a href=\"https:\/\/i.stack.imgur.com\/YqN1I.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YqN1I.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53238120","CreationDate":"2018-11-10T10:40:51.270","ParentId":null,"OwnerUserId":"1709080","Title":null,"Body":"<p>This is almost always a problem with the input data.<\/p>\n\n<p>I would suggest inspecting in detail the values you are feeding into the model to make sure the model is receiving what you think it is.<\/p>\n"}]}
+{"QuestionId":53203218,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-11-08T07:39:12.547","AcceptedAnswerId":53203580.0,"OwnerUserId":8916723.0,"Title":"Reshaping tensor in custom loss function in Keras","Body":"<p>I'm getting None type when reshaping the tensor. This happens when compiling the model with loss function and the optimizer (before starting the training). What do I do?<\/p>\n\n<p>Error:<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class 'tuple'&gt; to Tensor. Contents: (None, -1). Consider casting elements to a supported type.  \n<\/code><\/pre>\n\n<p>Custom loss function:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n\n        y_pred = K.reshape(y_pred, (K.get_variable_shape(y_pred)[0], -1))\n        y_true = K.reshape(y_true, (K.get_variable_shape(y_true)[0], -1))\n        y_pred = K.std(y_pred, axis=0)\n        y_true = K.std(y_true, axis=0)\n        loss = (1\/2) * (y_pred - y_true) ** 2\n        loss = K.mean(loss)\n\n        return loss\n<\/code><\/pre>\n","answers":[{"AnswerId":"53203580","CreationDate":"2018-11-08T08:05:58.707","ParentId":null,"OwnerUserId":"5108062","Title":null,"Body":"<p>This happens because your <code>y_true<\/code> or <code>y_pred<\/code> tensors shape is not defined properly. <code>None<\/code> means here that the tensors shape is not strictly set, but it can vary based on the previous operations what we can't see. Or you just initialized your tensor like this.<\/p>\n\n<p>How to fix it:<\/p>\n\n<ul>\n<li>First you should investigate how the <code>y_true<\/code> or <code>y_pred<\/code> gets its shape and avoid getting a None shape, so the tensors will have a deterministic number of rows and cols<\/li>\n<\/ul>\n\n<p>Example:<\/p>\n\n<p>Your loss function works for proper inputs:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n\n\ndef custom_loss(y_true, y_pred):\n    y_pred = K.reshape(y_pred, (K.get_variable_shape(y_pred)[0], -1))\n    y_true = K.reshape(y_true, (K.get_variable_shape(y_true)[0], -1))\n    y_pred = K.std(y_pred, axis=0)\n    y_true = K.std(y_true, axis=0)\n    loss = (1 \/ 2) * (y_pred - y_true) ** 2\n\n    return loss\n\n\na = tf.constant([[1.0, 2., 3.]])\nb = tf.constant([[1., 2., 3.]])\nloss = custom_loss(a, b)\nloss = tf.Print(loss, [loss], \"loss\")\n\nwith tf.Session() as sess:\n    _ = sess.run([loss])\n<\/code><\/pre>\n\n<p>But when using a placeholder where I have not defined the shape, will throw the same exception<\/p>\n\n<pre><code>a = tf.placeholder(tf.float32, (None, 32))\n<\/code><\/pre>\n"},{"AnswerId":"53203643","CreationDate":"2018-11-08T08:09:44.203","ParentId":null,"OwnerUserId":"8916723","Title":null,"Body":"<p>I fixed it:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n\n        y_pred = K.reshape(y_pred, (K.shape(y_pred)[0], -1))\n        y_true = K.reshape(y_true, (K.shape(y_true)[0], -1))\n        y_pred = K.std(y_pred, axis=0)\n        y_true = K.std(y_true, axis=0)\n        loss = (1\/2) * (y_pred - y_true) ** 2\n        loss = K.mean(loss)\n\n        return loss\n<\/code><\/pre>\n\n<p>The problem was with finding the first dimension of the y_true and y_predict. During compilation, it will not get a true shape and hence it returns a None value. So, instead of getting the integer value of the shape, I get the tf.Tensor shape. I changed  K.get_variable_shape(y_true) to just K.shape(y_true) and it solves. <\/p>\n"}]}
+{"QuestionId":53203644,"AnswerCount":1,"Tags":"<opencv><caffe>","CreationDate":"2018-11-08T08:09:47.120","AcceptedAnswerId":null,"OwnerUserId":6430408.0,"Title":"caffe is conflicted with python cv2","Body":"<p>I want to capture camera info with python ,and predict with caffe. When import caffe and import cv2 at the same time ,as the code:<\/p>\n\n<pre><code>import caffe\nimport cv2\n\n# caffe.set_mode_cpu()\ncap = cv2.VideoCapture(0)\nwhile(1):\n    # get a frame\n    ret, frame = cap.read()\n    # show a frame\n    cv2.imshow(\"capture\", frame)\n    if cv2.waitKey(1) &amp; 0xFF == ord('q'):\n        break\ncap.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p><strong>the following error will happen:<\/strong><\/p>\n\n<p>objc[35793]: Class CaptureDelegate is implemented in both \/usr\/local\/opt\/opencv\/lib\/libopencv_videoio.3.4.dylib (0x104dc0938) and \/Users\/python-envs\/py27\/lib\/python2.7\/site-packages\/cv2\/cv2.so (0x1111d9380). One of the two will be used. Which one is undefined.\nsrc\/tcmalloc.cc:332] Attempt to free invalid pointer 0x7f9f37403fc0 <\/p>\n\n<p>I know the reason is that ,caffe and cv2 using different opencv lib , but how to solve this problem?What's more ,I don't want to rebuilt the caffe<\/p>\n","answers":[{"AnswerId":"53386302","CreationDate":"2018-11-20T04:34:50.700","ParentId":null,"OwnerUserId":"10678081","Title":null,"Body":"<p>This is because caffe and opencv use different memory allocator. Caffe uses tcmalloc but opencv uses the default one. \nleveldb is dependent on gperftools, which contains tcmalloc.\nChick whether you've set USE_LEVELDB := 0 in the MakeFile.config of caffe. If you haven't, it's better to set USE_LEVELDB := 0 and rebuilt caffe. Or you can change the memory allocator used by opencv to be the same with caffe.<\/p>\n"}]}
+{"QuestionId":53203775,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-11-08T08:20:09.323","AcceptedAnswerId":null,"OwnerUserId":3604106.0,"Title":"Are there different between origin model and Model(origin model's input, origin model's output)?","Body":"<p>I would like to use middle layers output in the loss function, so I use<\/p>\n\n<pre><code>dis_loss = Model(input=dis_model.get_input_at(0), output=dis_model.get_output_at(-1))\nfunc_x = dis_loss(y_true)\nfunc_x = dis_loss(y_pred) \n<\/code><\/pre>\n\n<p>Then, I got an error message as following.<\/p>\n\n<pre><code>ValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_layer_1:0\", shape=(?, 224, 224, 3), dtype=float32) at layer \"input_layer\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>I had checked the name of <code>dis_model.get_input_at(0)<\/code>, and <code>dis_model.layer[0].input<\/code> and they are same.\nfirst layer's name of <code>dis_model.summary()<\/code> is also same.<\/p>\n\n<p>However I try to directly use <code>dis_loss = dis_model<\/code>, and then the model compiled successfully. Are the two models different? Input and output are from the same model, shouldn't they connect?<\/p>\n","answers":[{"AnswerId":"53205651","CreationDate":"2018-11-08T10:17:57.530","ParentId":null,"OwnerUserId":"3604106","Title":null,"Body":"<p>2 input layers would cause 2 output layers, so <\/p>\n\n<pre><code>dis_loss = Model(input=dis_model.get_input_at(0), output=dis_model.get_output_at(-1))\n<\/code><\/pre>\n\n<p>has to be revised as <\/p>\n\n<pre><code>dis_loss = Model(input=dis_model.get_input_at(0), output=dis_model.get_output_at(0))\n<\/code><\/pre>\n\n<p>When combing 2 model,<\/p>\n\n<pre><code>gen_x = gen_model.output\ndis_z = dis_model(gen_x)\nd_on_g = model(input=gen_x.input, output=[gen_x, dis_z])\n<\/code><\/pre>\n\n<p>it creates second output, so\n<code>dis_model.get_output_at(0)<\/code> is <code>dense_1\/Tanh:0<\/code> which is dis_model's output, so it connect to the dis_model's input.<\/p>\n\n<p><code>dis_model.get_output_at(1)<\/code> is <code>model_2\/dense_1\/Tanh:0<\/code> which is d_on_g's output, so it coneect to d_on_g's input which is equal to gen_model's input.<\/p>\n"}]}
+{"QuestionId":53203795,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-11-08T08:21:25.923","AcceptedAnswerId":null,"OwnerUserId":10439611.0,"Title":"tensorflow estimator accuracy and loss is zero","Body":"<p>My model's accuracy and loss are evaluating to 0.<br>\nThe global steps should be 1625 but it's 1.<br>\nThe acc and loss shouldn't be equal to 0 as both of them are contradicting each other.<\/p>\n\n<p>My input function,keras estimator,train_and_evaluate are as follows:<\/p>\n\n<pre><code>def make_input_fn(addrs,labels,batch_size,mode):\n\n filename_dataset = tf.data.Dataset.from_tensor_slices((addrs,labels))     \n\n dataset = filename_dataset.apply(tf.contrib.data.map_and_batch(lambda \n addrs, labels: tuple(tf.py_func(\n    process, [addrs, labels], [tf.uint8, labels.dtype])),batch_size,\n\n num_parallel_batches=2,\n\n drop_remainder=False))\n if mode == tf.estimator.ModeKeys.TRAIN:\n  num_epochs = None # indefinitely\n  dataset = dataset.apply(tf.contrib.data.shuffle_and_repeat(buffer_size = 10000))\n else:\n  num_epochs = 1\n  dataset = dataset.repeat(num_epochs)\n\n dataset = dataset.prefetch(buffer_size=batch_size)\n images,labels = dataset.make_one_shot_iterator().get_next()\n images.set_shape([None,512,512,3])\n labels.set_shape([None,1])\n return images,labels\n\ndef keras_estimator(model_dir,config):\n base_model = Xception(weights='imagenet', include_top=False,input_shape = \n  (512,512,3),classes = 5)\n x = base_model.output\n x = GlobalAveragePooling2D()(x)\n\n x = Dense(1024, activation='relu')(x)\n x = Dropout(0.2)(x)\n x = Dense(256, activation='relu')(x)\n x = Dropout(0.2)(x)\n\n predictions = Dense(5, activation='softmax')(x)\n\n\n model = Model(inputs=base_model.input, outputs=predictions)\n\n\n for layer in base_model.layers:\n   layer.trainable = False\n model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', \n       metrics=['acc'])\n\n\n estimator=tf.keras.estimator.model_to_estimator(keras_model=model,\n      model_dir=model_dir,\n      config=config)\n return estimator\n\ndef train_and_evaluate(model_dir):\n t_batch_size = 512\n e_batch_size = 64\n num_epochs = 25\n import pandas as pd\n df = pd.read_csv('\/content\/trainLabels.csv')\n from random import shuffle\n addrs = ['\/content\/train\/train\/' + str(df.iloc[i]['image']) + '.jpeg' for i \n in range(len(df))]\n labels = df['level'].values.tolist()\n c = list(zip(addrs, labels))\n shuffle(c)\n addrs1, labels1 = zip(*c)\n train_addrs = addrs1[0 : int(0.9 * len(addrs))]\n train_labels = labels1[0 : int(0.9 * len(labels))]\n val_addrs = addrs1[ int(0.9 * len(addrs)) : ]\n val_labels = labels1[ int(0.9 * len(addrs)) : ]\n train_addrs = list(train_addrs)\n train_labels = list(train_labels)\n val_addrs = list(val_addrs)\n val_labels = list(val_labels)\n\n run_config = tf.estimator.RunConfig(save_checkpoints_secs=300)\n\n estimator = keras_estimator(model_dir,run_config)\n\n t_max_steps = (len(train_addrs) \/\/ t_batch_size) * num_epochs\n\n train_spec = tf.estimator.TrainSpec(input_fn = lambda : \n make_input_fn(train_addrs,train_labels,\n t_batch_size,mode=tf.estimator.ModeKeys.TRAIN),max_steps = t_max_steps)\n\n eval_spec = tf.estimator.EvalSpec(input_fn = lambda : \n make_input_fn(val_addrs,val_labels,\n e_batch_size,mode=tf.estimator.ModeKeys.EVAL),steps = \n None,start_delay_secs=10,\n    throttle_secs=300)\n\n\n tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>Here are the log files:<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Running training and evaluation locally\n  (non-distributed). INFO:tensorflow:Start train and evaluate loop. The\n  evaluate will happen after every checkpoint. Checkpoint frequency is\n  determined based on RunConfig arguments: save_checkpoints_steps None\n  or save_checkpoints_secs 300. WARNING:tensorflow:From\n  :9: map_and_batch (from\n  tensorflow.contrib.data.python.ops.batching) is deprecated and will be\n  removed in a future version. Instructions for updating: Use\n  tf.data.experimental.map_and_batch(...). WARNING:tensorflow:From\n  :12: shuffle_and_repeat (from\n  tensorflow.contrib.data.python.ops.shuffle_ops) is deprecated and will\n  be removed in a future version. Instructions for updating: Use\n  tf.data.experimental.shuffle_and_repeat(...). INFO:tensorflow:Calling\n  model_fn. INFO:tensorflow:Done calling model_fn.\n  INFO:tensorflow:Warm-starting with WarmStartSettings:\n  WarmStartSettings(ckpt_to_initialize_from='\/content\/training\/keras\/keras_model.ckpt',\n  vars_to_warm_start='.*', var_name_to_vocab_info={},\n  var_name_to_prev_var_name={}) INFO:tensorflow:Warm-starting from:\n  ('\/content\/training\/keras\/keras_model.ckpt',)\n  INFO:tensorflow:Warm-starting variable: dense\/kernel; prev_var_name:\n  Unchanged INFO:tensorflow:Warm-starting variable: dense\/bias;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  dense_1\/kernel; prev_var_name: Unchanged INFO:tensorflow:Warm-starting\n  variable: dense_1\/bias; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: dense_2\/kernel; prev_var_name:\n  Unchanged INFO:tensorflow:Warm-starting variable: dense_2\/bias;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  Adam\/iterations; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: Adam\/lr; prev_var_name:\n  Unchanged INFO:tensorflow:Warm-starting variable: Adam\/beta_1;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  Adam\/beta_2; prev_var_name: Unchanged INFO:tensorflow:Warm-starting\n  variable: Adam\/decay; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_1; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_2;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_3; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_4;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_5; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_6;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_7; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_8;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_9; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_10;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_11; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_12;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_13; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_14;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_15; prev_var_name: Unchanged\n  INFO:tensorflow:Warm-starting variable: training\/Adam\/Variable_16;\n  prev_var_name: Unchanged INFO:tensorflow:Warm-starting variable:\n  training\/Adam\/Variable_17; prev_var_name: Unchanged\n  INFO:tensorflow:Create CheckpointSaverHook. INFO:tensorflow:Graph was\n  finalized. INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done\n  running local_init_op. INFO:tensorflow:Saving checkpoints for 0 into\n  \/content\/training\/model.ckpt. INFO:tensorflow:Saving checkpoints for 1\n  into \/content\/training\/model.ckpt. INFO:tensorflow:Calling model_fn.\n  INFO:tensorflow:Done calling model_fn. INFO:tensorflow:Starting\n  evaluation at 2018-11-05-13:21:17 INFO:tensorflow:Graph was finalized.\n  INFO:tensorflow:Restoring parameters from\n  \/content\/training\/model.ckpt-1 INFO:tensorflow:Running local_init_op.\n  INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Finished\n  evaluation at 2018-11-05-13:22:08 INFO:tensorflow:Saving dict for\n  global step 1: acc = 0.0, global_step = 1, loss = 0.0\n  INFO:tensorflow:Saving 'checkpoint_path' summary for global step 1:\n  \/content\/training\/model.ckpt-1 INFO:tensorflow:Loss for final step:\n  None.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"56963311","CreationDate":"2019-07-10T04:12:45.550","ParentId":null,"OwnerUserId":"9826546","Title":null,"Body":"<p>I had this issue before. It was because that I specified the wrong directory for the data sets. Ultimately tensorflow had no input data. I hope this helps. <\/p>\n"}]}
+{"QuestionId":53203986,"AnswerCount":0,"Tags":"<python><lua><pytorch><torch>","CreationDate":"2018-11-08T08:32:59.540","AcceptedAnswerId":null,"OwnerUserId":4393898.0,"Title":"Creating \".t7\" torch file in python","Body":"<p>By using pytorch, I can manage to save file like the following:<\/p>\n\n<pre><code>torch.save(object, \"object.t7\")\n<\/code><\/pre>\n\n<p>However, I cannot open it through Lua.<\/p>\n\n<p>When I use the following command in Lua,<\/p>\n\n<pre><code>local info1 = torch.load(\"object.t7\")\n<\/code><\/pre>\n\n<p>I receive the following error:<\/p>\n\n<blockquote>\n  <p>\/home\/hasan\/build\/torch\/install\/bin\/luajit:\n  ...an\/build\/torch\/install\/share\/lua\/5.1\/threads\/threads.lua:315:\n  ...an\/build\/torch\/install\/share\/lua\/5.1\/threads\/threads.lua:183:\n  [thread 4 callback]\n  ...e\/hasan\/build\/torch\/install\/share\/lua\/5.1\/torch\/File.lua:375:\n  unknown object stack traceback:   [C]: in function 'error'\n    ...e\/hasan\/build\/torch\/install\/share\/lua\/5.1\/torch\/File.lua:375: in\n  function 'readObject'\n    ...e\/hasan\/build\/torch\/install\/share\/lua\/5.1\/torch\/File.lua:409: in\n  function 'load'\n    \/home\/hasan\/GAN\/icml2016\/data\/donkey_folder_txt_coco.lua:112: in\n  function \n    [C]: in function 'xpcall'\n    ...an\/build\/torch\/install\/share\/lua\/5.1\/threads\/threads.lua:234: in\n  function 'callback'\n    ...asan\/build\/torch\/install\/share\/lua\/5.1\/threads\/queue.lua:65: in\n  function\n  &lt;...asan\/build\/torch\/install\/share\/lua\/5.1\/threads\/queue.lua:41>  [C]:\n  in function 'pcall'\n    ...asan\/build\/torch\/install\/share\/lua\/5.1\/threads\/queue.lua:40: in\n  function 'dojob'  [string \"  local Queue = require\n  'threads.queue'...\"]:15: in main chunk<\/p>\n<\/blockquote>\n\n<p>I need to run a torch training in Lua, with preprocessed data through Pytorch.<\/p>\n\n<p>Is there any way to save \".t7\" file via pytorch, and open it through Lua?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[]}
+{"QuestionId":53204604,"AnswerCount":0,"Tags":"<python><keras><deep-learning><backpropagation>","CreationDate":"2018-11-08T09:14:31.103","AcceptedAnswerId":null,"OwnerUserId":4657751.0,"Title":"Truncated Backpropagation in keras with one sequence per batch","Body":"<p>If I understood correctly, to perform TBPTT in keras we have to split our sequences into smaller parts of k timesteps. To re-use the state of our LSTM accross all the parts of the sequence we have to use the stateful parameter, according to the documentation of keras :<\/p>\n\n<blockquote>\n  <p>You can set RNN layers to be 'stateful', which means that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch. This assumes a one-to-one mapping between samples in different successive batches.<\/p>\n<\/blockquote>\n\n<p>So if I understand correctly the 1st sample of the 1st batch is the 1s part of the 1st sequence, the 1st sample of the 2nd batch is the 2nd part of the 1 sequence, etc. I have 125973 sequences of length 1000 that I split into 40 sequences of k=25 timesteps. So my model should train on 40 batches containing 125973 sequences of 25 timesteps. My issue is the memory of my GPU (quadro K2200, I'm poor), a batch size of 125973 seems to be too much. I'd like to know if it is possible to keep the state of the LSTM inside the same batch and reset it between batches, so I should have a batch size of 40 and 125973 batches instead.<\/p>\n\n<p>Here is my model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(len(char_to_num), 200, mask_zero=True, batch_input_shape=(batch_size, k)))\nmodel.add(Dropout(0.5))\nmodel.add(LSTM(512, activation='relu', return_sequences=True, stateful=True))\nmodel.add(Dropout(0.5))\nmodel.add(TimeDistributed(Dense(len(char_to_num), activation='softmax')))\n\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer='adam')\nmodel.summary()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53204747,"AnswerCount":2,"Tags":"<python><tensorflow><memory-leaks><keras>","CreationDate":"2018-11-08T09:23:59.343","AcceptedAnswerId":null,"OwnerUserId":10319899.0,"Title":"Keras variable() memory leak","Body":"<p>I am new to Keras, and tensorflow in general, and have a problem. I am using some of the loss functions (binary_crossentropy and mean_squared_error mainly) to calculate the loss after prediction. Since Keras only accepts it's own variable type, I am creating one and supply it as an argument. This scenario is executed in a loop (with sleep) as such:<\/p>\n\n<p>Get appropriate data -> predict -> calculate the lost -> return it.<\/p>\n\n<p>Since I have multiple models that follow this pattern I created tensorflow graphs and sessions to prevent collision (also when exporting the models' weights I had problem with single graph and session so I had to create distinct ones for every single model).<\/p>\n\n<p>However, now the memory is rising uncontrollably, from couple of MiB to 700MiB in couple of iterations. I am aware of Keras's clear_session() and gc.collect(), and I use them at the end of every iteration, but the problem is still present. Here I provide a code snippet, which is not the actual code, from the project. I've created separate script in order to isolate the problem:<\/p>\n\n<pre><code>import tensorflow as tf\n\nfrom keras import backend as K\nfrom keras.losses import binary_crossentropy, mean_squared_error\n\nfrom time import time, sleep\nimport gc\nfrom numpy.random import rand\n\nfrom os import getpid\nfrom psutil import Process\n\nfrom csv import DictWriter\nfrom keras import backend as K\n\nthis_process = Process(getpid())\n\ngraph = tf.Graph()\nsess = tf.Session(graph=graph)\n\ncnt = 0\nmax_c = 500\n\nwith open('\/home\/quark\/Desktop\/python-test\/leak-7.csv', 'a') as file:\n    writer = DictWriter(file, fieldnames=['time', 'mem'])\n    writer.writeheader()\n\n    while cnt &lt; max_c:  \n        with graph.as_default(), sess.as_default():         \n            y_true = K.variable(rand(36, 6))\n            y_pred = K.variable(rand(36, 6))\n\n            rec_loss = K.eval(binary_crossentropy(y_true, y_pred))\n            val_loss = K.eval(mean_squared_error(y_true, y_pred))\n\n            writer.writerow({\n                'time': int(time()),\n                'mem': this_process.memory_info().rss\n            })\n\n        K.clear_session()\n        gc.collect()\n\n        cnt += 1\n        print(max_c - cnt)\n        sleep(0.1)\n<\/code><\/pre>\n\n<p>Additionally, I've added the plot of the memory usage:\n<a href=\"https:\/\/imgur.com\/u3m7nGp\" rel=\"nofollow noreferrer\">Keras memory leak<\/a><\/p>\n\n<p>Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"53205795","CreationDate":"2018-11-08T10:25:52.913","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p>I just removed the <code>with<\/code> statement (probably some tf code), and I don't see any leak. I believe there is a difference between the keras session and the tf default session. So you were not clearing the correct session with <code>K.clear_session()<\/code>. Probably using <code>tf.reset_default_graph()<\/code> could work too.<\/p>\n\n<pre><code>while True: \n    y_true = K.variable(rand(36, 6))\n    y_pred = K.variable(rand(36, 6))\n\n    val_loss = K.eval(binary_crossentropy(y_true, y_pred))\n    rec_loss = K.eval(mean_squared_error(y_true, y_pred))\n\n    K.clear_session()\n    gc.collect()\n\n    sleep(0.1)\n<\/code><\/pre>\n"},{"AnswerId":"53224737","CreationDate":"2018-11-09T11:18:47.183","ParentId":null,"OwnerUserId":"10319899","Title":null,"Body":"<p>At the end, what I did was remove the <code>K.variable()<\/code> code out of the <code>where<\/code> statement. In that way, the variables are part of the default graph which is later cleared by <code>K.clear_session()<\/code>.<\/p>\n"}]}
+{"QuestionId":53204794,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T09:27:05.970","AcceptedAnswerId":53256465.0,"OwnerUserId":9589731.0,"Title":"A strange bug in tf.scatter_add when I implement unpool in tensorflow","Body":"<p>I'm trying to implement unpool in tensorflow with tf.scatter_add, but I meet a strange bug, here is my code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport random\n\ntf.reset_default_graph()\n\nmat = list(range(64))\nrandom.shuffle(mat)\nmat = np.array(mat)\nmat = np.reshape(mat, [1,8,8,1])\nM = tf.constant(mat, dtype=tf.float32)\npool1, argmax1 = tf.nn.max_pool_with_argmax(M, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\npool2, argmax2 = tf.nn.max_pool_with_argmax(pool1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\npool3, argmax3 = tf.nn.max_pool_with_argmax(pool2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\n\n\ndef unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):\n    x_shape = x.get_shape().as_list()\n    argmax_shape = argmax.get_shape().as_list()\n    assert not(x_shape[0] is None and batch_size is None), \"must input batch_size if number of batch is alterable\"\n    if x_shape[0] is None:\n        x_shape[0] = batch_size\n    if argmax_shape[0] is None:\n        argmax_shape[0] = x_shape[0]\n    if unpool_shape is None:\n        unpool_shape = [x_shape[i] * strides[i] for i in range(4)]\n    x_unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)\n    argmax = tf.cast(argmax, tf.int32)\n    argmax = tf.reshape(argmax, [np.prod(argmax_shape)])\n    x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])\n    x_unpool = tf.scatter_add(x_unpool , argmax, x)\n    x_unpool = tf.reshape(x_unpool , unpool_shape)\n    return x_unpool \n\n\nunpool2 = unpool(pool3, argmax3, strides=[1,2,2,1], name='unpool3')\nunpool1 = unpool(unpool2, argmax2, strides=[1,2,2,1], name='unpool2')\nunpool0 = unpool(unpool1, argmax1, strides=[1,2,2,1], name='unpool1')\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    mat_out = mat[:,:,:,0]\n    pool1_out = sess.run(pool1)[0,:,:,0]\n    pool2_out = sess.run(pool2)[0,:,:,0]\n    pool3_out = sess.run(pool3)[0,:,:,0]\n    argmax1_out = sess.run(argmax1)[0,:,:,0]\n    argmax2_out = sess.run(argmax2)[0,:,:,0]\n    argmax3_out = sess.run(argmax3)[0,:,:,0]\n    unpool2_out = sess.run(unpool2)[0,:,:,0]\n    unpool1_out = sess.run(unpool1)[0,:,:,0]\n    unpool0_out = sess.run(unpool0)[0,:,:,0]\n    print(unpool2_out)\n    print(unpool1_out)\n    print(unpool0_out)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[[ 0.  0.]\n [ 0. 63.]]\n[[  0.   0.   0.   0.]\n [  0.   0.   0.   0.]\n [  0.   0. 126.   0.]\n [  0.   0.   0.   0.]]\n[[  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0. 315.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]\n [  0.   0.   0.   0.   0.   0.   0.   0.]]\n<\/code><\/pre>\n\n<p>The location is right, but the value is wrong. unpool2 is right, unpool1 is double of expected value, and unpool2 is quintuple of expected value. I don't know what's wrong, can anyone tell me how to fix this bug?<\/p>\n\n<p>Great thanks in advance.<\/p>\n","answers":[{"AnswerId":"53204888","CreationDate":"2018-11-08T09:31:44.447","ParentId":null,"OwnerUserId":"9589731","Title":null,"Body":"<p>Use tf.scatter_update can avoid this.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport random\n\ntf.reset_default_graph()\n\nmat = list(range(64))\nrandom.shuffle(mat)\nmat = np.array(mat)\nmat = np.reshape(mat, [1,8,8,1])\nM = tf.constant(mat, dtype=tf.float32)\npool1, argmax1 = tf.nn.max_pool_with_argmax(M, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\npool2, argmax2 = tf.nn.max_pool_with_argmax(pool1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\npool3, argmax3 = tf.nn.max_pool_with_argmax(pool2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')\n\n\ndef unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):\n    x_shape = x.get_shape().as_list()\n    argmax_shape = argmax.get_shape().as_list()\n    assert not(x_shape[0] is None and batch_size is None), \"must input batch_size if number of batch is alterable\"\n    if x_shape[0] is None:\n        x_shape[0] = batch_size\n    if argmax_shape[0] is None:\n        argmax_shape[0] = x_shape[0]\n    if unpool_shape is None:\n        unpool_shape = [x_shape[i] * strides[i] for i in range(4)]\n    unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)\n    argmax = tf.cast(argmax, tf.int32)\n    argmax = tf.reshape(argmax, [np.prod(argmax_shape)])\n    x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])\n    unpool = tf.scatter_update(unpool, argmax, x)\n    unpool = tf.reshape(unpool, unpool_shape)\n    return unpool\n\n\nunpool2 = unpool(pool3, argmax3, strides=[1,2,2,1], name='unpool3')\nunpool1 = unpool(unpool2, argmax2, strides=[1,2,2,1], name='unpool2')\nunpool0 = unpool(unpool1, argmax1, strides=[1,2,2,1], name='unpool1')\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    mat_out = mat[:,:,:,0]\n    pool1_out = sess.run(pool1)[0,:,:,0]\n    pool2_out = sess.run(pool2)[0,:,:,0]\n    pool3_out = sess.run(pool3)[0,:,:,0]\n    argmax1_out = sess.run(argmax1)[0,:,:,0]\n    argmax2_out = sess.run(argmax2)[0,:,:,0]\n    argmax3_out = sess.run(argmax3)[0,:,:,0]\n    unpool2_out = sess.run(unpool2)[0,:,:,0]\n    unpool1_out = sess.run(unpool1)[0,:,:,0]\n    unpool0_out = sess.run(unpool0)[0,:,:,0]\n    print(unpool2_out)\n    print(unpool1_out)\n    print(unpool0_out)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[[ 0.  0.]\n [ 0. 63.]]\n[[ 0.  0.  0.  0.]\n [ 0.  0.  0.  0.]\n [ 0.  0.  0. 63.]\n [ 0.  0.  0.  0.]]\n[[ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0. 63.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  0.  0.  0.]]\n<\/code><\/pre>\n"},{"AnswerId":"53256465","CreationDate":"2018-11-12T05:41:51.657","ParentId":null,"OwnerUserId":"9589731","Title":null,"Body":"<p>In fact, the answer is simple. For convenience, I rename some variables, look this code: <\/p>\n\n<pre><code>def unpool(x, argmax, strides, unpool_shape=None, batch_size=None, name='unpool'):\n    x_shape = x.get_shape().as_list()\n    argmax_shape = argmax.get_shape().as_list()\n    assert not(x_shape[0] is None and batch_size is None), \"must input batch_size if number of batch is alterable\"\n    if x_shape[0] is None:\n        x_shape[0] = batch_size\n    if argmax_shape[0] is None:\n        argmax_shape[0] = x_shape[0]\n    if unpool_shape is None:\n        unpool_shape = [x_shape[i] * strides[i] for i in range(4)]\n    x_unpool = tf.get_variable(name=name, shape=[np.prod(unpool_shape)], initializer=tf.zeros_initializer(), trainable=False)\n    argmax = tf.cast(argmax, tf.int32)\n    argmax = tf.reshape(argmax, [np.prod(argmax_shape)])\n    x = tf.reshape(x, [np.prod(argmax.get_shape().as_list())])\n    x_unpool_add = tf.scatter_add(x_unpool , argmax, x)\n    x_unpool_reshape = tf.reshape(x_unpool_add , unpool_shape)\n    return x_unpool_reshape \n<\/code><\/pre>\n\n<p>x_unpool_add is a op of tf.scatter_add, everytime we compute x_unpool_reshape, x_unpool_add will be called. So x_unpool will add x twice if we compute unpool2 twice. In my origin code, I compute unpool0, unpool1, unpool2 in order, x_unpool_add of unpool1 is called firstly, then when we compute unpool2, because of we need compute unpool1, x_unpool_add will be called again, so it's equal to call x_unpool_add twice, the value is wrong. If we compute unpool2 directly, we will get right result. So replacling tf.scatter_add with tf.scatter_update can avoid this bug.<\/p>\n\n<p>This code can reproducible this intuitively:<\/p>\n\n<pre><code>import tensorflow as tf\n\nt1 = tf.get_variable(name='t1', shape=[1], dtype=tf.float32, initializer=tf.zeros_initializer())\nt2 = tf.get_variable(name='t2', shape=[1], dtype=tf.float32, initializer=tf.zeros_initializer())\nd = tf.scatter_add(t1, [0], [1])\ne = tf.scatter_add(t2, [0], d)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    d_out1 = sess.run(d)\n    d_out2 = sess.run(d)\n    e_out = sess.run(e)\n    print(d_out1)\n    print(d_out2)\n    print(e_out)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[1.]\n[2.]\n[3.]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53206037,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-08T10:41:43.540","AcceptedAnswerId":53206820.0,"OwnerUserId":10027028.0,"Title":"What means mean and stddev in keras\/tensorflow","Body":"<p>in below part of code what means mean and stddev ?\nI know the seed is put to 1 so if you generate random values those are always the same. But don't know about mean and stddev?<\/p>\n\n<p>I know the seed is put on 1 so you should have always the same values generated<\/p>\n\n<pre><code>    with tf.Session() as test:\n    tf.set_random_seed(1)\n    y_true = (None, None, None)\n    y_pred = (tf.random_normal([3, 128], mean=6, stddev=0.1, seed = 1),\n              tf.random_normal([3, 128], mean=1, stddev=1, seed = 1),\n              tf.random_normal([3, 128], mean=3, stddev=4, seed = 1))\n    loss = triplet_loss(y_true, y_pred)\n\n    print(\"loss = \" + str(loss.eval()))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53206820","CreationDate":"2018-11-08T11:27:37.480","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p><a href=\"https:\/\/www.mathsisfun.com\/definitions\/mean.html\" rel=\"nofollow noreferrer\">Mean<\/a> is the average of the numbers which will be generated by tf.random_normal(). Similarly, stddev stands for <a href=\"https:\/\/www.mathsisfun.com\/data\/standard-deviation.html\" rel=\"nofollow noreferrer\">Standard deviation<\/a> of the data.<\/p>\n\n<ul>\n<li>if mean=1.2 and stddev=3.4 then, the number matrix generated by tf.random_normal() will have mean and standard deviation as 1.2 and 3.4 respectively.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53206147,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-08T10:48:29.433","AcceptedAnswerId":53206448.0,"OwnerUserId":1563247.0,"Title":"Merging keras models before compile or fit?","Body":"<p>working in Keras 2.2.4<\/p>\n\n<p>I have three data features (feature1, feature2, feature3) that correspond to a set of users and how they rank certain items.  At present, I have modelled them all separately using a model architecture as follows:<\/p>\n\n<pre><code>input_vecs = add([feature1_vec, user_vec])\nnn = Dropout(0.5)(Dense(128, activation='relu')(input_vecs))\nnn = BatchNormalization()(nn)\nnn = Dropout(0.5)(Dense(128, activation='relu')(nn))\nnn = BatchNormalization()(nn)\nnn = Dense(128, activation='relu')(nn)\nfeature1result = Dense(9, activation='softmax')(nn)\n\nfeature1model = Model([feature1_input, user_input], feature1result)\nfeaturemodel.compile('adam', 'categorical_crossentropy')\n<\/code><\/pre>\n\n<p>Each of the models have similar architectures (each separately tuned), and the same style\/shape of outputs.  I would like to take the preliminary result of the three models and put those into a new layer and then create a final result.<\/p>\n\n<p>I guess I could run the three models separately, take their output, and then put that output into an entirely new (likely Sequential) model using a structure like this:<\/p>\n\n<pre><code>model1 = load_model(\"feature1.h5\")\nmodel2 = load_model(\"feature2.h5\")\nmodel3 = load_model(\"feature3.h5\")\nmerged_model = Sequential()\nmerged_model.add(merge([model1.layers[-1].output,model2.layers[-1].output,model3.layers[-1].output]))\nmerged_model.add(Dense(units = 9, activation='relu')) #or whatever units\nmerged_model.add(Dense(units = 12, activation='relu'))#or whatever units\nmerged_model.add(Dense(9, activation='softmax'))\n<\/code><\/pre>\n\n<p>However, I'd like to merge them beforehand as it will make managing the models easier later on (i.e., only updating one large model instead of updating 4 smaller models).<\/p>\n\n<p>How can I do this before doing the <code>compile()<\/code> and <code>fit()<\/code> steps?<\/p>\n","answers":[{"AnswerId":"53206448","CreationDate":"2018-11-08T11:04:32.137","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p>You need to use the functional API to use the merge layer. I hope this gives you the idea:<\/p>\n\n<pre><code>input = Input( ... ) \nmodel1 = load_model(\"feature1.h5\")\nmodel2 = load_model(\"feature2.h5\")\nmodel3 = load_model(\"feature3.h5\")\n\nm1 = model1(input)\nm2 = model2(input)\nm3 = model3(input)\nmerged_model = merge([m1,m2,m3]) \n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53206554,"AnswerCount":1,"Tags":"<python><torch>","CreationDate":"2018-11-08T11:10:39.880","AcceptedAnswerId":null,"OwnerUserId":4393898.0,"Title":"Saving data in \".t7\" format with python","Body":"<p>Is there any possible way to save data in \".t7\" format with python?<\/p>\n\n<p>\".t7\" is the serialization format, which can be opened by lua torch. However, when I save my data in pickle with python with \".t7\" at the end of the naming, it does not work.<\/p>\n\n<p>I have  been searching through internet, but I could not find out any working answer.<\/p>\n","answers":[{"AnswerId":"53207190","CreationDate":"2018-11-08T11:52:02.610","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>there is currently no such converter but there is a workaround. What you can do is convert pyTorch model to Caffe model and from Caffe to Torch (lua) model. <a href=\"https:\/\/github.com\/ysh329\/deep-learning-model-convertor\" rel=\"nofollow noreferrer\">Here<\/a> is the table of converters by framework.<\/p>\n"}]}
+{"QuestionId":53206781,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-11-08T11:24:45.047","AcceptedAnswerId":null,"OwnerUserId":7205164.0,"Title":"CUDA runtime implicit initialization on GPU:0 failed. Status: unknown error","Body":"<p>For ImageDataGenerator using keras library i use the following code.Number of training set is 8000 and number of test set is 2000.<\/p>\n\n<pre><code>classifier.fit_generator(\n        generator = training_set,\n        steps_per_epoch=8000,\n        validation_data = test_set,\n        validation_steps = 2000,\n        epochs=25)\n<\/code><\/pre>\n\n<p>But when i run the code i get the following errors.Is there any problem in my tensorflow version or keras version? Currently i am using keras version = 2.2.4, python=3.6, tensorflow version = 1.11.  <\/p>\n\n<pre><code>InternalError                             Traceback (most recent call last)\n&lt;ipython-input-88-16fbb44d18e3&gt; in &lt;module&gt;()\n      4         validation_data = test_set,\n      5         validation_steps = 2000,\n----&gt; 6         epochs=25)\n\n~\\Anaconda3\\lib\\site-packages\\keras\\legacy\\interfaces.py in wrapper(*args, **kwargs)\n     89                 warnings.warn('Update your `' + object_name + '` call to the ' +\n     90                               'Keras 2 API: ' + signature, stacklevel=2)\n---&gt; 91             return func(*args, **kwargs)\n     92         wrapper._original_function = func\n     93         return wrapper\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n   1416             use_multiprocessing=use_multiprocessing,\n   1417             shuffle=shuffle,\n-&gt; 1418             initial_epoch=initial_epoch)\n   1419 \n   1420     @interfaces.legacy_generator_methods_support\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n    215                 outs = model.train_on_batch(x, y,\n    216                                             sample_weight=sample_weight,\n--&gt; 217                                             class_weight=class_weight)\n    218 \n    219                 outs = to_list(outs)\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py in train_on_batch(self, x, y, sample_weight, class_weight)\n   1215             ins = x + y + sample_weights\n   1216         self._make_train_function()\n-&gt; 1217         outputs = self.train_function(ins)\n   1218         return unpack_singleton(outputs)\n   1219 \n\n~\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in __call__(self, inputs)\n   2695 \n   2696     def __call__(self, inputs):\n-&gt; 2697         if hasattr(get_session(), '_make_callable_from_options'):\n   2698             if py_any(is_sparse(x) for x in self.inputs):\n   2699                 if py_any(is_tensor(x) for x in inputs):\n\n~\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in get_session()\n    184                 config = tf.ConfigProto(intra_op_parallelism_threads=num_thread,\n    185                                         allow_soft_placement=True)\n--&gt; 186             _SESSION = tf.Session(config=config)\n    187         session = _SESSION\n    188     if not _MANUAL_VAR_INIT:\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in __init__(self, target, graph, config)\n   1509 \n   1510     \"\"\"\n-&gt; 1511     super(Session, self).__init__(target, graph, config=config)\n   1512     # NOTE(mrry): Create these on first `__enter__` to avoid a reference cycle.\n   1513     self._default_graph_context_manager = None\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py in __init__(self, target, graph, config)\n    632     try:\n    633       # pylint: disable=protected-access\n--&gt; 634       self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)\n    635       # pylint: enable=protected-access\n    636     finally:\n\nInternalError: CUDA runtime implicit initialization on GPU:0 failed. Status: unknown error\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53206900,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-08T11:31:55.827","AcceptedAnswerId":53207330.0,"OwnerUserId":7089239.0,"Title":"Sound way of managing multiple sessions and graphs","Body":"<p>I'd like to manage multiple Keras models in multiple sessions. My application is constructed such that models can be live at the same time, in addition to creating, saving and loading them.<\/p>\n\n<p><strong>What is the proper way of managing this situation?<\/strong><\/p>\n\n<p>Currently one model is represented by an instance of a wrapper class. This is used in the training, saving, loading and prediction. One <code>tf.Graph<\/code> and <code>tf.Session<\/code> is created per instance, and they are used in every function requiring the actual model.<\/p>\n\n<pre><code>class NN:\n    def __init__(self):\n        self.graph = tf.Graph()\n        self.session = tf.Session(graph=self.graph)\n\n    def predict(self, x):\n        with self.graph.as_default():\n            with self.session.as_default():\n                return self.model.predict(x)\n<\/code><\/pre>\n\n<p>Similar functions using the <code>with<\/code> statements are created for compiling the network, fitting, saving (weights to .h5 and model to JSON) and loading. So whenever the model is needed, the graph and session are brought to context.<\/p>\n\n<p>This resulted in a strange error (<a href=\"https:\/\/stackoverflow.com\/questions\/53002518\/poor-exit-code-when-managing-multiple-sessions\">Q<\/a> for further context), and I was left wondering, what is the standard way of dealing with this. I tried to release all possible resources before creating or loading a model, but it hasn't helped. This function is just a compilation of all possible routines scraped off the internet, and is purely guesswork.<\/p>\n\n<pre><code>def _new_session(self):\n    if self.session is not None:\n        self.session.close()\n    k.clear_session()\n    gc.collect()\n    self.graph = tf.Graph()\n    self.session = tf.Session(graph=self.graph)\n<\/code><\/pre>\n\n<p>I've not found good documentation of a similar situation. So I'd very much appreciate any real insight into this.<\/p>\n\n<hr>\n\n<p>I might need to delete the old question, as it's quite all over the place. At the time of asking I had no idea what was going on. But it's there for now.<\/p>\n\n<hr>\n\n<p>Some specific questions have arisen.<\/p>\n\n<ul>\n<li>Loading and making predictions on a model works, compiling and fitting doesn't, although just compiling does. Do the two contexts differ in any way? Is the loaded model exactly the same?<\/li>\n<li>At which points should a new context be created when manipulating the models? (e.g. at load, compilation, fitting, probably not with every prediction)<\/li>\n<li>Which actions are necessary to take when releasing the resources of previous contexts? Either when a network is disposed of or when creating a new context.<\/li>\n<li>Why exactly is the context switch needed for multiple models?<\/li>\n<li>What are the roles of graph vs. session, given that different things are executed on the graph and session?<\/li>\n<\/ul>\n\n<h3>Updates<\/h3>\n\n<ul>\n<li>Compiling, fitting and saving one network works without any context trickery. Doing the same for another model in the same context works too (or at least does not produce an error).<\/li>\n<li>In addition to above, <strong>loading the saved model and predicting works too<\/strong>, right after the training and for both models! Now I'm not sure whether the prediction is made correctly, but again, no error. This only begs the question I posed above even more: <em>why are the different contexts needed?<\/em><\/li>\n<\/ul>\n\n<hr>\n\n<p>The underlying issue with the error has been finally (and somewhat embarassingly) <a href=\"https:\/\/stackoverflow.com\/questions\/53002518\/poor-exit-code-when-managing-multiple-sessions\">resolved<\/a> by updating all packages.<\/p>\n","answers":[{"AnswerId":"53207330","CreationDate":"2018-11-08T12:00:39.117","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>EDIT: Actually, seeing again how <code>K.get_session()<\/code> works, it should return the current default session, so I'm not sure <code>set_session<\/code> is doing anything meaningful there. I'll leave the answer just in case you want to try but probably this won't help.<\/p>\n\n<hr>\n\n<p>Maybe you can get it to work with something like this:<\/p>\n\n<pre><code>from contextlib import contextmanager\n\nclass NN:\n    def __init__(self):\n        self.graph = tf.Graph()\n        self.session = tf.Session(graph=self.graph)\n\n    def predict(self, x):\n        with self._context():\n            return self.model.predict(x)\n\n    @contextmanager\n    def _context(self):\n        prev_sess = K.get_session()\n        K.set_session(self.session)\n        with self.graph.as_default(), self.session.as_default():\n            yield\n        K.set_session(prev_sess)\n<\/code><\/pre>\n\n<p>Note that the Keras session object is a global variable, so I suppose this should work as long as you don't try to use these contexts from multiple threads.<\/p>\n"}]}
+{"QuestionId":53207057,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-lite><toco>","CreationDate":"2018-11-08T11:43:22.697","AcceptedAnswerId":null,"OwnerUserId":1223817.0,"Title":"Tensorflow [Toco] convert Model to optimized format cause ValueError","Body":"<p>I just started trying Tensorflow, I'd like to do some experiments on Android, but I read that Tensorflow for Poets (the version for mobile) support .tflite while the models I downloaded and found googling are .pb.<\/p>\n\n<p>So I decided to convert them using toco, but when I execute the command to convert I always get this error<\/p>\n\n<blockquote>\n  <p>ValueError: Invalid tensors 'input' were found.<\/p>\n<\/blockquote>\n\n<p>The CLI command is:<\/p>\n\n<pre><code>toco   --graph_def_file=tensorflow_inception_graph.pb   --output_file=optimized_graph.tflite   --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE   --input_shape=1,300,300,3   --input_array=input   --o\nutput_array=final_result   --inference_type=FLOAT   --inference_input_type=FLOAT\n<\/code><\/pre>\n\n<p>What's wrong with that?<\/p>\n\n<p><strong>Stacktrace<\/strong><\/p>\n\n<pre><code>File \"\/home\/Mitro\/venv\/bin\/toco\", line 11, in &lt;module&gt;\n    sys.exit(main())\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 401, in main\n    app.run(main=run_main, argv=sys.argv[:1])\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 397, in run_main\n    _convert_model(tflite_flags)\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 100, in _convert_model\n    converter = _get_toco_converter(flags)\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 87, in _get_toco_converter\n    return converter_fn(**converter_kwargs)\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", line 271, in from_frozen_graph\n    sess.graph, input_arrays)\n  File \"\/home\/Mitro\/venv\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/convert_saved_model.py\", line 189, in get_tensors_from_tensor_names\n    \",\".join(invalid_tensors)))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets-2-tflite\/#2\" rel=\"nofollow noreferrer\">My reference from Google<\/a><\/p>\n","answers":[{"AnswerId":"53404445","CreationDate":"2018-11-21T02:20:14.883","ParentId":null,"OwnerUserId":"4584138","Title":null,"Body":"<p>Run this <a href=\"https:\/\/github.com\/rdeepc\/Tensorflow_lite_tutorial\/blob\/master\/modelinfo.py\" rel=\"nofollow noreferrer\">script<\/a> changing the model_file parameter to the pb file, and as an output you will have something like this:<\/p>\n\n<p>Tensor(\"import\/<strong>Placeholder<\/strong>:0\", shape=(?, 299, 299, 3), dtype=float32)<br>\nTensor(\"import\/<strong>final_result<\/strong>:0\", shape=(?, 5), dtype=float32)<\/p>\n\n<p>in this case they are the parameters that you must pass through<br>\n--input_array = <strong>Placeholder<\/strong><br>\n--output_array = <strong>final_result<\/strong><\/p>\n\n<p><a href=\"https:\/\/medium.com\/@rdeep\/tensorflow-lite-tutorial-easy-implementation-in-android-145443ec3775\" rel=\"nofollow noreferrer\">Here<\/a> is the complete information<\/p>\n"}]}
+{"QuestionId":53207842,"AnswerCount":0,"Tags":"<apache-spark><keras><deep-learning><google-cloud-ml>","CreationDate":"2018-11-08T12:34:22.043","AcceptedAnswerId":null,"OwnerUserId":10139404.0,"Title":"How to make realtime predictions in Spark Streaming with a pretrained h5py model from Keras?","Body":"<p>I want to produce realtime predictions with a pretrained model. The data cannot fit in RAM so, I use Keras in Google Cloud ML to training and I want to use Apache Spark Streaming for predictions but I couldn't find any resources to be able to use pretrained models (h5py) in Spark streaming.<\/p>\n\n<p>Another question is effiency of my strategy. Is there any better way for making predictions based on sensor data in realtime with a pretrained model? Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53208035,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-08T12:45:38.140","AcceptedAnswerId":null,"OwnerUserId":8909870.0,"Title":"Append channels of zero to tensor in tensorflow","Body":"<p>I want to append channels with al zero as values to an existing tensor with the limitation of a flexible batch size. So I can not use <code>tf.concat<\/code> and concatenate my tensor with <code>tf.zeros<\/code>. I thought about something like <code>tf.pad<\/code>, but I need to do it in the last\/fourth dimension.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Here is an example:<\/p>\n\n<p>I have an array of [8, 300, 300, 3] with specific values:\ne.g. <\/p>\n\n<p><code>tensor_1 = tf.convert_to_tensor(np.random.random([8,300, 300, 3]))<\/code> <\/p>\n\n<p>Now I want to expand the number of the last dimensions to e.g. 6 with  a complete zero tensor.<\/p>\n\n<p>If I have a static batch size I just can do the following:<\/p>\n\n<p><code>zeros = tf.zeros(tensor_1.shape) \nresult = tf.concat([tensor_1, zeros])<\/code><\/p>\n\n<p>But if the first dimension is <code>None<\/code> <code>tf.zeros<\/code> returns an error. So I need method that does not depend on knowing the batch size.<\/p>\n","answers":[]}
+{"QuestionId":53208293,"AnswerCount":1,"Tags":"<python><machine-learning><keras><classification><mnist>","CreationDate":"2018-11-08T13:01:07.600","AcceptedAnswerId":53209168.0,"OwnerUserId":4921713.0,"Title":"Getting different accuracy on test data in MNIST digit recognition in Keras","Body":"<p>I am doing <em>handwritten digit recognition<\/em> using Keras and I have two files: <em>predict.py<\/em> and <em>train.py<\/em>.<\/p>\n\n<p><em>train.py<\/em> trains the model (if it is not already trained) and saves it to a directory, otherwise it would just load the trained model from the directory it was saved to and prints the <code>Test Loss<\/code> and <code>Test Accuracy<\/code>.<\/p>\n\n<pre><code>def getData():\n    (X_train, y_train), (X_test, y_test) = mnist.load_data()\n    y_train = to_categorical(y_train, num_classes=10)\n    y_test = to_categorical(y_test, num_classes=10)\n    X_train = X_train.reshape(X_train.shape[0], 784)\n    X_test = X_test.reshape(X_test.shape[0], 784)\n\n    # normalizing the data to help with the training\n    X_train \/= 255\n    X_test \/= 255\n\n\n    return X_train, y_train, X_test, y_test\n\ndef trainModel(X_train, y_train, X_test, y_test):\n    # training parameters\n    batch_size = 1\n    epochs = 10\n    # create model and add layers\n    model = Sequential()    \n    model.add(Dense(64, activation='relu', input_shape=(784,)))\n    model.add(Dense(10, activation = 'softmax'))\n\n\n    # compiling the sequential model\n    model.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer='adam')\n    # training the model and saving metrics in history\n    history = model.fit(X_train, y_train,\n          batch_size=batch_size, epochs=epochs,\n          verbose=2,\n          validation_data=(X_test, y_test))\n\n    loss_and_metrics = model.evaluate(X_test, y_test, verbose=2)\n    print(\"Test Loss\", loss_and_metrics[0])\n    print(\"Test Accuracy\", loss_and_metrics[1])\n\n    # Save model structure and weights\n    model_json = model.to_json()\n    with open('model.json', 'w') as json_file:\n        json_file.write(model_json)\n    model.save_weights('mnist_model.h5')\n    return model\n\ndef loadModel():\n    json_file = open('model.json', 'r')\n    model_json = json_file.read()\n    json_file.close()\n    model = model_from_json(model_json)\n    model.load_weights(\"mnist_model.h5\")\n    return model\n\nX_train, y_train, X_test, y_test = getData()\n\nif(not os.path.exists('mnist_model.h5')):\n    model = trainModel(X_train, y_train, X_test, y_test)\n    print('trained model')\n    print(model.summary())\nelse:\n    model = loadModel()\n    print('loaded model')\n    print(model.summary())\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n    loss_and_metrics = model.evaluate(X_test, y_test, verbose=2)\n    print(\"Test Loss\", loss_and_metrics[0])\n    print(\"Test Accuracy\", loss_and_metrics[1])\n<\/code><\/pre>\n\n<p>Here is the output (assuming model was trained earlier and this time model will just be loaded):<\/p>\n\n<blockquote>\n  <p>('Test Loss', 1.741784990310669)<\/p>\n  \n  <p>('Test Accuracy', 0.414)<\/p>\n<\/blockquote>\n\n<p><em>predict.py<\/em>, on the other hand, predicts a handwritten number:<\/p>\n\n<pre><code>def loadModel():\n    json_file = open('model.json', 'r')\n    model_json = json_file.read()\n    json_file.close()\n    model = model_from_json(model_json)\n    model.load_weights(\"mnist_model.h5\")\n    return model\n\nmodel = loadModel()\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nprint(model.summary())\n\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\ny_test = to_categorical(y_test, num_classes=10)\nX_test = X_test.reshape(X_test.shape[0], 28*28)\n\n\nloss_and_metrics = model.evaluate(X_test, y_test, verbose=2)\n\nprint(\"Test Loss\", loss_and_metrics[0])\nprint(\"Test Accuracy\", loss_and_metrics[1])\n<\/code><\/pre>\n\n<p>In this case, to my surprise, getting the following result:<\/p>\n\n<blockquote>\n  <p>('Test Loss', 1.8380377866744995)<\/p>\n  \n  <p>('Test Accuracy', 0.8856)<\/p>\n<\/blockquote>\n\n<p>In the second file, I am getting a <code>Test Accuracy<\/code> of 0.88 (more than double that I was getting before).<\/p>\n\n<p>Also, <code>model.summery()<\/code> is the same in both of the files: <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense_1 (Dense)              (None, 64)                50240     \n_________________________________________________________________\ndense_2 (Dense)              (None, 10)                650       \n=================================================================\nTotal params: 50,890\nTrainable params: 50,890\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>I can't figure out the reason behind this behavior. Is it normal? Or am I missing something?<\/p>\n","answers":[{"AnswerId":"53209168","CreationDate":"2018-11-08T13:54:39.510","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The discrepancy results from the fact that one time you are calling <code>evaluate()<\/code> method with normalized data (i.e. divided by 255) and the other time (i.e. in \"predict.py\" file) you are calling it with un-normalized data. In inference time (i.e. test time) you should always use the same pre-processing step you have used for the training data.<\/p>\n\n<p>Further, first convert the data to floating point and then divide it by 255 (otherwise, with <code>\/<\/code>, a true division is done in Python 2.x and in Python 3.x you would get errors when running <code>X_train \/= 255<\/code> and <code>X_test \/= 255<\/code>):<\/p>\n\n<pre><code>X_train = X_train.astype('float32')\nX_test = X_test.astype('float32')\n\nX_train \/= 255.\nX_test \/= 255.\n<\/code><\/pre>\n"}]}
+{"QuestionId":53208334,"AnswerCount":1,"Tags":"<python><tensorflow><backpropagation><automatic-differentiation>","CreationDate":"2018-11-08T13:03:47.597","AcceptedAnswerId":53212479.0,"OwnerUserId":4296244.0,"Title":"How does tensorflow handle non differentiable nodes during gradient calculation?","Body":"<p>I understood the concept of automatic differentiation, but couldn't find any explanation how tensorflow calculates the error gradient for non differentiable functions as for example <code>tf.where<\/code> in my loss function or <code>tf.cond<\/code> in my graph. It works just fine, but I would like to understand how tensorflow backpropagates the error through such nodes, since there is no formula to calculate the gradient from them.<\/p>\n","answers":[{"AnswerId":"53212479","CreationDate":"2018-11-08T16:51:33.193","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>In the case of <code>tf.where<\/code>, you have a function with three inputs, condition <code>C<\/code>, value on true <code>T<\/code> and value on false <code>F<\/code>, and one output <code>Out<\/code>. The gradient receives one value and has to return three values. Currently, no gradient is computed for the condition (that would hardly make sense), so you just need to do the gradients for <code>T<\/code> and <code>F<\/code>. Assuming the input and the outputs are vectors, imagine <code>C[0]<\/code> is <code>True<\/code>. Then <code>Out[0]<\/code> comes from <code>T[0]<\/code>, and its gradient should propagate back. On the other hand, <code>F[0]<\/code> would have been discarded, so its gradient should be made zero. If <code>Out[1]<\/code> were <code>False<\/code>, then the gradient for <code>F[1]<\/code> should propagate but not for <code>T[1]<\/code>. So, in short, for <code>T<\/code> you should propagate the given gradient where <code>C<\/code> is <code>True<\/code> and make it zero where it is <code>False<\/code>, and the opposite for <code>F<\/code>. If you look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.12.0\/tensorflow\/python\/ops\/math_grad.py#L1112-L1118\" rel=\"nofollow noreferrer\">the implementation of the gradient of <code>tf.where<\/code> (<code>Select<\/code> operation)<\/a>, it does exactly that:<\/p>\n\n<pre><code>@ops.RegisterGradient(\"Select\")\ndef _SelectGrad(op, grad):\n  c = op.inputs[0]\n  x = op.inputs[1]\n  zeros = array_ops.zeros_like(x)\n  return (None, array_ops.where(c, grad, zeros), array_ops.where(\n      c, zeros, grad))\n<\/code><\/pre>\n\n<p>Note the input values themselves are not used in the computation, that will be done by the gradients of the operation producing those inputs. For <code>tf.cond<\/code>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.12.0\/tensorflow\/python\/ops\/control_flow_grad.py#L95-L138\" rel=\"nofollow noreferrer\">the code is a bit more complicated<\/a>, because the same operation (<code>Merge<\/code>) is used in different contexts, and also <code>tf.cond<\/code> also uses <code>Switch<\/code> operations inside. However the idea is the same. Essentially, <code>Switch<\/code> operations are used for each input, so the input that was activated (the first if the condition was <code>True<\/code> and the second otherwise) gets the received gradient and the other input gets a \"switched off\" gradient (like <code>None<\/code>), and does not propagate back further.<\/p>\n"}]}
+{"QuestionId":53208646,"AnswerCount":0,"Tags":"<python><callback><keras><neural-network><metrics>","CreationDate":"2018-11-08T13:22:49.427","AcceptedAnswerId":null,"OwnerUserId":9629391.0,"Title":"Why do I have different results calculating True Positive rate in Keras Neural Network?","Body":"<p>I'm training a neural network using Python Keras package. I care about the True Positive rate, so I added it to Callbacks and Metrics. Surprisingly, I'm getting different results using the same formula (Callbacks shows 81%, which is correct: I can see the same manually after I join Labels and Predictions; Metrics shows higher, around 86%). What is the matter? Any comment on the code is also appreciated<\/p>\n\n<pre><code>def sensitivity(y_true, y_pred):\n    true_positives = K.sum(K.round(K.clip(y_true, 0, 1)) * K.round(K.clip(y_pred, 0, 1)))\n    possible_positives = K.sum(y_true)\n    return true_positives \/ (possible_positives + K.epsilon())\n....\ndef calculate_rates(model, data, label):\n    num_positive_prediction = np.sum(label)\n    prediction = np.round(np.clip(model.predict(data, batch_size = 1024)[:,1], 0, 1))\n    true_positive = np.sum(np.multiply(prediction, label)) \/ num_positive_prediction\n    return(true_positive)\n....\nclass TestCallback(keras.callbacks.Callback):\n    def __init__(self, is_train, data, labels):\n        self.data = data\n        self.labels = labels[:,1]\n        self.is_train = is_train\n\n    def on_epoch_end(self, epoch, logs={}):\n        true_positive = calculate_rates(model, self.data, self.labels)\n        if (epoch + 1) % 10 == 0 or epoch == 0:\n            if self.is_train:\n                print(\"Epoch: %d\" % epoch + 1)\n                print(\"Training Set:\")\n            else:\n                print(\"Testing Set:\")\n            print(\"True Positive Rate: %4g\" % true_positive)\n....\nmodel = keras.Sequential()\nmy_init = keras.initializers.RandomNormal(stddev=0.1)\nmodel.add(keras.layers.Dense(units=200, activation='relu', input_dim=num_variables))\nmodel.add(keras.layers.Dropout(dropout_rate))\nmodel.add(keras.layers.Dense(units = 200, activation = 'relu', kernel_initializer=my_init, bias_initializer=my_init))\nmodel.add(keras.layers.Dropout(dropout_rate))\nmodel.add(keras.layers.Dense(units=num_outputs, activation='softmax', kernel_initializer=my_init, bias_initializer=my_init))\nmodel.compile(loss='binary_crossentropy',\n              optimizer=keras.optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-8),\n              metrics=[sensitivity]\n              )\nhistory = model.fit(train_data, train_labels, epochs=num_epochs, batch_size=1024, class_weight={0: 1, 1: weight},\n                    callbacks=[TestCallback(1, train_data, train_labels), TestCallback(0, test_data, test_labels)], verbose=1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53208970,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorflow-datasets>","CreationDate":"2018-11-08T13:42:25.117","AcceptedAnswerId":null,"OwnerUserId":6547250.0,"Title":"tf.data pipeline design for optimized performance","Body":"<p>I am new to TensorFlow and I would like to know of there is any specific order to set the dataset using tfdata. For example:<\/p>\n\n<pre><code>    data_files = tf.gfile.Glob(\"%s%s%s\" % (\".\/data\/cifar-100-binary\/\", self.data_key, \".bin\"))\n    data = tf.data.FixedLengthRecordDataset(data_files, record_bytes=3074)\n    data = data.map(self.load_transform)\n    if self.shuffle_key:\n        data = data.shuffle(5000)\n\n    data = data.batch(self.batch_size).repeat(100)\n    iterator = data.make_one_shot_iterator()\n    img, label = iterator.get_next()\n    # label = tf.one_hot(label, depth=100)\n    print('img_shape:', img.shape)\n<\/code><\/pre>\n\n<p>In this case I read the data then shuffle the data followed by batch and repeat specifications. With this method my computer's RAM increases by 2% <\/p>\n\n<p>and then I tried one more method:<\/p>\n\n<pre><code>    data_files = tf.gfile.Glob(\"%s%s%s\" % (\".\/data\/cifar-100-binary\/\", self.data_key, \".bin\"))\n    data = tf.data.FixedLengthRecordDataset(data_files, record_bytes=3074)\n    data = data.map(self.load_transform)\n    data = data.batch(self.batch_size).repeat(100)\n    if self.shuffle_key:\n        data = data.shuffle(5000)\n    iterator = data.make_one_shot_iterator()\n    img, label = iterator.get_next()\n    # label = tf.one_hot(label, depth=100)\n    print('img_shape:', img.shape)\n<\/code><\/pre>\n\n<p>so with this case when I first specify the batch size, repeat and then shuffle the RAM utilization increases by 40% (I do not know why) it would be great if someone helps me figure that out. \nSo is there a sequence which I should always follow to define the dataset in tensorflow using tf.data ? <\/p>\n","answers":[{"AnswerId":"58087081","CreationDate":"2019-09-24T19:33:31.523","ParentId":null,"OwnerUserId":"401884","Title":null,"Body":"<p>The memory usage increases because you are shuffling batches instead of single records.<\/p>\n\n<p><code>data.shuffle(5000)<\/code> will fill a buffer of <code>5000<\/code> elements, then sample randomly from the buffer to produce the next element.<\/p>\n\n<p><code>data.batch(self.batch_size)<\/code> changes the element type from single records to batches of records. So if you call <code>batch<\/code> before <code>shuffle<\/code>, the shuffle buffer will contain <code>5000 * self.batch_size<\/code> records instead of just <code>5000<\/code>.<\/p>\n\n<p>The order of calling <code>shuffle<\/code> and <code>batch<\/code> will also affect the data itself. Batching before shuffling will result in all elements of a batch being sequential.<\/p>\n\n<p><code>batch<\/code> before <code>shuffle<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; dataset = tf.data.Dataset.range(12)\n&gt;&gt;&gt; dataset = dataset.batch(3)\n&gt;&gt;&gt; dataset = dataset.shuffle(4)\n&gt;&gt;&gt; print([element.numpy() for element in dataset])\n[array([ 9, 10, 11]), array([0, 1, 2]), array([3, 4, 5]), array([6, 7, 8])]\n<\/code><\/pre>\n\n<p><code>shuffle<\/code> before <code>batch<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; dataset = tf.data.Dataset.range(12)\n&gt;&gt;&gt; dataset = dataset.shuffle(4)\n&gt;&gt;&gt; dataset = dataset.batch(3)\n&gt;&gt;&gt; print([element.numpy() for element in dataset])\n[array([1, 2, 5]), array([4, 7, 8]), array([0, 3, 9]), array([ 6, 10, 11])]\n<\/code><\/pre>\n\n<p>Usually shuffling is done before batching to avoid all elements in the batch being sequential.<\/p>\n"}]}
+{"QuestionId":53209495,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T14:11:17.787","AcceptedAnswerId":53210523.0,"OwnerUserId":851699.0,"Title":"Tensorflow: copy existing graph into new graph multiple times","Body":"<p>I want to paste an existing tensorflow graph into a new graph.<\/p>\n\n<p>Suppose I create a graph computing <code>y = tanh(x @ w)<\/code><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef some_function(x):\n    w = tf.Variable(initial_value=np.random.randn(4, 5), dtype=tf.float32)\n    return tf.tanh(x @ w)\n\nx = tf.placeholder(shape=(None, 4), dtype = tf.float32)\ny = some_function(x)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nval_x = np.random.randn(3, 4)\nval_y, = sess.run([y], feed_dict={x: val_x})\n<\/code><\/pre>\n\n<p>Great.  Now suppose I've lost the code that generated that graph, but I still have access to variables (<code>x<\/code>, <code>y<\/code>).  Now I want to take this graph (using the current value of w), and copy it twice into a new graph (the two paths should share the same <code>w<\/code>), so that I now compute <code>d = tf.reduce_sum((tanh(x1 @ w)-tanh(x2 @ w))**2)<\/code> by adding the line:<\/p>\n\n<pre><code># Starting with access to tensors: x, y\n&lt;SOMETHING HERE&gt;\nd = tf.reduce_sum((y1-y2)**2)\nval_x1 = np.random.randn(3, 4)\nval_x2 = np.random.randn(3, 4)\nval_d = sess.run([d], feed_dict = {x1: val_x1, x2: val_x2})\n<\/code><\/pre>\n\n<p>What do I fill in for <code>&lt;SOMETHING HERE&gt;<\/code> to make this work?  (Obviously, without recreating the first graph)<\/p>\n","answers":[{"AnswerId":"53210523","CreationDate":"2018-11-08T15:07:14.627","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>There is the <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.graph_editor\" rel=\"nofollow noreferrer\">Graph Editor<\/a> module to help with this sort of operations. Its main disadvantage is that you cannot have a running session while you modify the graph. However, you can checkpoint the session, modify the graph and the restore it back if you need so.<\/p>\n\n<p>The problem with what you want is that you basically need to replicate a subgraph except you do no want to replicate variables. So you can simply exclude variable types (mainly <code>Variable<\/code>, <code>VariableV2<\/code> and maybe <code>VarHandleOp<\/code>, although I threw in a few more I found in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.12.0\/tensorflow\/python\/training\/device_setter.py#L30-L34\" rel=\"nofollow noreferrer\">TensorFlow code<\/a>). You can do that with a function like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Receives the outputs to recalculate and the input replacements\ndef replicate_subgraph(outputs, mappings):\n    # Types of operation that should not be replicated\n    # Taken from tensorflow\/python\/training\/device_setter.py\n    NON_REPLICABLE = {'Variable', 'VariableV2', 'AutoReloadVariable',\n                      'MutableHashTable', 'MutableHashTableV2',\n                      'MutableHashTableOfTensors', 'MutableHashTableOfTensorsV2',\n                      'MutableDenseHashTable', 'MutableDenseHashTableV2',\n                      'VarHandleOp', 'BoostedTreesEnsembleResourceHandleOp'}\n    # Find subgraph ops\n    ops = tf.contrib.graph_editor.get_backward_walk_ops(outputs, stop_at_ts=mappings.keys())\n    # Exclude non-replicable operations\n    ops_replicate = [op for op in ops if op.type not in NON_REPLICABLE]\n    # Make subgraph viewitems\n    sgv = tf.contrib.graph_editor.make_view(*ops_replicate)\n    # Make the copy\n    _, info = tf.contrib.graph_editor.copy_with_input_replacements(sgv, mappings)\n    # Return new outputs\n    return info.transformed(outputs)\n<\/code><\/pre>\n\n<p>For an example similar to yours (I edited it a bit so it is easy to see that the output is correct because the second value is ten times the first one).<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef some_function(x):\n    w = tf.Variable(initial_value=tf.random_normal((5,)), dtype=tf.float32)\n    return 2 * (x * w)\n\nx1 = tf.placeholder(shape=(), dtype=tf.float32, name='X1')\nx2 = tf.placeholder(shape=(), dtype=tf.float32, name='X2')\ny1 = some_function(x1)\ny2, = replicate_subgraph([y1], {x1: x2})\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    print(*sess.run([y1, y2], feed_dict={x1: 1, x2: 10}), sep='\\n')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[ 2.3356955   2.277849    0.58513653  2.0919807  -0.15102367]\n[23.356955  22.77849    5.851365  20.919807  -1.5102367]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>Here is another solution using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/make_template\" rel=\"nofollow noreferrer\"><code>tf.make_template<\/code><\/a>. This requires you to actually have the code for the function, but it is a cleaner and \"more official\" way of supporting subgraph reuse.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef some_function(x):\n    w = tf.get_variable('W', (5,), initializer=tf.random_normal_initializer())\n    # Or if the variable is only local and not trainable\n    # w = tf.Variable(initial_value=tf.random_normal(5,), dtype=tf.float32, trainable=False)\n    return 2 * (x * w)\n\nx1 = tf.placeholder(shape=(), dtype=tf.float32, name='X1')\nx2 = tf.placeholder(shape=(), dtype=tf.float32, name='X2')\nsome_function_tpl = tf.make_template('some_function', some_function)\ny1 = some_function_tpl(x1)\ny2 = some_function_tpl(x2)\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    print(*sess.run([y1, y2], feed_dict={x1: 1, x2: 10}), sep='\\n')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53209513,"AnswerCount":0,"Tags":"<python><memory><neural-network><pytorch>","CreationDate":"2018-11-08T14:12:33.600","AcceptedAnswerId":null,"OwnerUserId":6224975.0,"Title":"Memory keeps rising when training FFNN pyTorch","Body":"<p>I have a Feed forward Neural Network which classifies the MNIST data set.\nFor some reason the memory keeps approaching 99% no matter how big the batch size is.\nI don't have anything which is increasing in size - every dynamic variable is getting overridden after the first epoch but even after epoch number 70 the memory keeps rising. <\/p>\n\n<p>Im running it on 8GB memory, 2.8 ghZ intel i-5 (7th gen) quad core (Ubuntu 18.04)<\/p>\n\n<pre><code>batch_size =50 #Number of x's we pass through the net at each iteration\nnum_epochs =100 #Number of times the entire training data is through the    net\n\nn_train = len(xtrain)\nn_batch_train = n_train \/\/ batch_size\nn_val =len(xval)\nn_batch_val = n_val \/\/ batch_size\n\n# loss\/acc\ntrain_acc,train_loss =[],[]\nval_acc,val_loss =[],[]\n\ntest_acc,test_loss =[],[]\n\n\n#Get parameters from the net\n\npar=[]\n\nfor i in range(len(layers)-1):\n    par=par+list(net.L[i].parameters())\n\n\n\n#Optimizer\n\noptimizer = optim.Adam(par,lr=0.001)\n\n#interval of x\n\nget_slice = lambda i,size: range(i*size,(i+1)*size)\n\n\nfor e in range(num_epochs):\n  curr_loss =0\n  net.train()\n  for i in range(n_batch_train):\n\n    x_interval = get_slice(i,batch_size)\n    slze = get_slice(i,batch_size)\n    #Batchnorm\n    bn = nn.BatchNorm1d(num_features = num_features)\n    x_batch = bn(Variable(torch.from_numpy(xtrain[slze])))\n\n    out = (net(x_batch)).double()\n    target_batch = Variable(torch.from_numpy(ytrain[slze]).double())\n    L=criterion(out,target_batch)\n\n    #Update gradients\n    optimizer.zero_grad()\n    L.backward()\n    optimizer.step()\n\n    #Store training accuracy and loss\n    train_acc.append(accuracy(target_batch, out).data)\n    train_loss.append(L.data.numpy())\n\n#### Validate ####\n  net.eval()\n  for j in range(n_batch_val):\n    slze = get_slice(j,batch_size)\n    val_batch = Variable(torch.from_numpy(xval[slze]))\n    val_out = (net(bn(val_batch))).double()\n    target_batch = Variable(torch.from_numpy(yval[slze]).double())\n\n    #Store val acc and loss\n    val_acc.append(accuracy(target_batch,val_out).data)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53209734,"AnswerCount":0,"Tags":"<android><tensorflow><cmake>","CreationDate":"2018-11-08T14:25:46.367","AcceptedAnswerId":null,"OwnerUserId":4829166.0,"Title":"Cannot compile tensorflowlite C++ API on android","Body":"<p>I think the problem is I'm not including the shared library in the right away, but I'm not sure. <\/p>\n\n<p>The error I get is<\/p>\n\n<pre><code>error: undefined reference to 'tflite::InterpreterBuilder::operator()(std::__ndk1::unique_ptr&lt;tflite::Interpreter, std::__ndk1::default_delete&lt;tflite::Interpreter&gt; &gt;*)\n<\/code><\/pre>\n\n<p>Which points at the last line of:<\/p>\n\n<pre><code>std::unique_ptr&lt;tflite::FlatBufferModel&gt; model = tflite::FlatBufferModel::BuildFromFile(casemodel_path.c_str());\ntflite::ops::builtin::BuiltinOpResolver resolver;\ntflite::InterpreterBuilder builder(*model.get(), resolver);\nstd::unique_ptr&lt;tflite::Interpreter&gt; interpreter;\nbuilder(&amp;interpreter);\n<\/code><\/pre>\n\n<p>I took this from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/minimal\/minimal.cc\" rel=\"nofollow noreferrer\">here<\/a>, as the <a href=\"https:\/\/www.tensorflow.org\/lite\/apis\" rel=\"nofollow noreferrer\">documentation<\/a> seems to be out of date.<\/p>\n\n<p>I compiled tensorflow from source using NDK16b. I followed <a href=\"https:\/\/stackoverflow.com\/questions\/49834875\/problems-with-using-tensorflow-lite-c-api-in-android-studio-project\">this<\/a> to compile it.<\/p>\n\n<p>The relevant portion of my cmake file looks like:<\/p>\n\n<pre><code>[...]\n# Flatbuffer\nset(FLATB_DIR &lt;path-to&gt;\/git\/flatbuffers)\ninclude_directories(${FLATB_DIR}\/include)\ninclude_directories(${FLATB_DIR}\/grpc)\n\nfile(GLOB flatb_src \"${FLATB_DIR}\/src\/*.cpp\")\nadd_library(flatbuffer ${flatb_src})\n\nadd_library(libtensorflowlite SHARED IMPORTED)\nset_target_properties(libtensorflowlite PROPERTIES IMPORTED_LOCATION\n        &lt;path-to&gt;\/app\/src\/main\/jniLibs\/armeabi-v7a\/libtensorflowlite.so)\n\n\ninclude_directories(&lt;path-to&gt;\/git\/tensorflow-android)\n[...]\ntarget_link_libraries(flatbuffer libtensorflowlite &lt;tons-of-other-libraries&gt;)\n<\/code><\/pre>\n\n<p>As I mentioned already I think the problem is I'm not including the shared library in the right way. All I've done is created the folder(s) <code>jniLibs\/armeabi-v7a\/<\/code> under <code>src\/main<\/code>, where I put the <code>libtensorflowlite.so<\/code>. Googling around that seems to be one way to do it? I've tried some other ways (with <code>sourceSets<\/code> and <code>implementation fileTree<\/code> but nothing worked).<\/p>\n\n<p>I've got some other precompiled libraries that I'm using just fine, but they are static (in <code>target_link_libraries<\/code> I point directly to their path). Is it a problem mixing static\/shared libraries like that?<\/p>\n\n<p>edit: Following <a href=\"https:\/\/heartbeat.fritz.ai\/compiling-a-tensorflow-lite-build-with-custom-operations-cf6330ee30e2\" rel=\"nofollow noreferrer\">this<\/a>, I also tried using ndk15c and editing <code>ANDROID_NDK_API_LEVEL<\/code>, but that did not help.<\/p>\n","answers":[]}
+{"QuestionId":53209822,"AnswerCount":1,"Tags":"<scikit-learn><keras>","CreationDate":"2018-11-08T14:30:40.607","AcceptedAnswerId":53210026.0,"OwnerUserId":1511274.0,"Title":"Load and use saved Keras model.h5","Body":"<p>I try to a KerasClassifier (wrapper)  into <code>final_model.h5<\/code><\/p>\n\n<pre><code>validator  = GridSearchCV(estimator=clf, param_grid=param_grid)\ngrid_result = validator.fit(train_images, train_labels)\nbest_estimator = grid_result.best_estimator_\nbest_estimator.model.save(\"final_model.h5\")\n<\/code><\/pre>\n\n<p>And then I want to reuse the model<\/p>\n\n<pre><code>from keras.models import load_model\nloaded_model = load_model(\"final_model.h5\")\n<\/code><\/pre>\n\n<p>But it seems like <code>loaded_model<\/code> is now a <code>Sequential<\/code> object instead. In other words it is different from <code>KerasClassifier<\/code> object like <code>best_estimator<\/code><\/p>\n\n<p>I want to reuse some method like <code>score<\/code> which is available in <code>KerasClassifier<\/code>, which is not available in <code>Sequential<\/code> model. What should I do?<\/p>\n\n<p>Also, I would like to know more about how to continue the training process left off on <code>final_model.h5<\/code>. What can I do next?<\/p>\n","answers":[{"AnswerId":"53210026","CreationDate":"2018-11-08T14:41:31.577","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, in the end you saved the Keras model as HDF5, not the KerasClassifier that is just an adapter to use with scikit-learn.<\/p>\n\n<p>But you don't really need the KerasClassifier instance, you want the <code>score<\/code> function and this in keras is called <code>evaluate<\/code>, so just call <code>model.evaluate(X, Y)<\/code> and this will return a list containing first the loss and then any metrics that your model used (most likely accuracy).<\/p>\n\n<p>To continue training the model, just load it and call <code>model.fit<\/code> with the new training set and that's it.<\/p>\n"}]}
+{"QuestionId":53210130,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm>","CreationDate":"2018-11-08T14:47:06.840","AcceptedAnswerId":53211105.0,"OwnerUserId":9261416.0,"Title":"Return states Keras- multivariate output","Body":"<p>With reference to this blog in the <code>return state<\/code> section: \n<a href=\"https:\/\/machinelearningmastery.com\/return-sequences-and-return-states-for-lstms-in-keras\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/return-sequences-and-return-states-for-lstms-in-keras\/<\/a><\/p>\n\n<p>I am trying to implement a multivariate (predicting 2 outputs- <code>y1<\/code> &amp; <code>y2<\/code>) <code>stateful<\/code> <code>LSTM<\/code> model. Here is the snippet:<\/p>\n\n<pre><code>## defining the model\ndef my_model():\n    input_x = Input(batch_shape=(batch_size, look_back, x_train.shape[2]), name='input')\n    drop = Dropout(0.5)\n\n    lstm_1, state_h, state_c = LSTM(50, return_sequences=False,batch_input_shape=(batch_size, look_back, x_train.shape[2]),return_state=True)(input_x)\n    lstm_1_drop = drop(lstm_1)\n\n    y1 = Dense(1, activation='linear', name='op1')(lstm_1_drop)\n    y2 = Dense(1, activation='linear', name='op2')(lstm_1_drop)\n\n    model = Model(inputs=input_x, outputs=[y1,y2])\n\n    optimizer = Adam(lr=0.0005, decay=0.00001)\n    model.compile(loss='mse', optimizer=optimizer,metrics=['mse'])\n    model.summary()\n    return model\n\nmodel = my_model()\n\nhistory = model.fit(x_train, [y_11_train,y_22_train], epochs=1, batch_size=batch_size, verbose=0, shuffle=False)\n<\/code><\/pre>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>I have some issues here: I am not sure if it has been properly implemented here yet as I didn't feed the <code>lstm_1, state_h, state_c<\/code> in the <code>outputs<\/code> of <code>Model()<\/code> as explained in the blog, because here I have two different predictions instead of 1 as stated in the blog. <\/p>\n\n<p>If I have to feed <code>lstm_1, state_h, state_c<\/code> in the <code>outputs<\/code> of <code>Model()<\/code> how can it be implemented and how it will affect the <code>model.fit<\/code> section. <\/p>\n\n<p>Any help will be highly appreciated. <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53211105","CreationDate":"2018-11-08T15:37:11.780","ParentId":null,"OwnerUserId":"9305450","Title":null,"Body":"<p>Reading the Blog post it seems like the author did implement state_h in the output to investigate the insides of the lstm layers. Such  an output can be usefull for an already trained network but not for training.<\/p>\n\n<p>For training you can safely leave the information out of your output. \nIf yu want to have the information during prediction, simply define a second model:<\/p>\n\n<pre><code>model = Model(inputs=input_x, outputs=[y1,y2,state_h,state_c])\n<\/code><\/pre>\n\n<p>Keras will then reuse your already trained layers and you have the information in your output without worrying about your training.<\/p>\n"}]}
+{"QuestionId":53210423,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><conv-neural-network><data-science>","CreationDate":"2018-11-08T15:02:22.707","AcceptedAnswerId":null,"OwnerUserId":7933418.0,"Title":"Input shape to Convolutional Neural Network","Body":"<p>I'm predicting the preference of a customer from a set of objects using convolutional neural network. The input is of the following format.<\/p>\n\n<pre><code>Customer    Objects  x1   x2   x3    x4 . .......x15\na1           fruits  0.5  .9   0.9   0.7\na1           veggies 0.2  .6   0.3   0.2\na1        condiments 0.0  .8   0.9   0.0\na1           dairy   0.4  .2   0.3   0.3\na1        pastries   0.6  .7   0.8   0.0\na1        other      0.9  .0   0.6   0.4\na2           fruits  0.5  .9   0.9   0.7\na2           veggies 0.2  .6   0.3   0.2\na2        condiments 0.0  .8   0.9   0.0\na2           dairy   0.4  .2   0.3   0.3\na2        pastries   0.6  .7   0.8   0.0\na2        other      0.9  .0   0.6   0.4\n<\/code><\/pre>\n\n<p>and so on. x1...x15 are continuous variables that represent buying habbit of that particular customer at different perios. I have normalized these variables and hot encoded the labels which is objects column. I'm retaining the objects column in predictors as the variables x combined with objects will have a meaning else they are just random numbers. Following is my code that creates X, Y and reshapes X in the format that could be feeded into conv net.,<\/p>\n\n<pre><code>X = df\nY = df.loc[:, df.columns == 'objects']\nY = Y.values.ravel()\n#Encode the class values to integers\nencoder = LabelEncoder()\nencoder.fit(Y)\nY_int = encoder.transform(Y)\n#Hot encode the integer values\nY_cat = np_utils.to_categorical(Y_int)\n\n# Generate Training and Validation Sets\nX_train, X_test, y_train, y_test = train_test_split(X,Y_cat, test_size=.3)\n\nprint(X_train.shape, y_train.shape)\nprint(X_test.shape, y_test.shape)\n\n((35000, 17), (35000, 6))\n((15000, 17), (15000, 6))\n\ndef data_formatting_score(df):\n  pred_col = [ 'x1', 'x2', 'x3',\n               'x4', 'x5', 'x6', 'x7',....'x15']\n  value_col = pred_col\n  df_cust=df.pivot_table(index=['customer','objects'],dropna=False,aggfunc=np.sum)[value_col]\n  df_c=df_c.fillna(value=0)\n  df_np=np.array(list(df_c.groupby(by=[df_c.index.get_level_values(0),df_c.index.get_level_values(0)]).apply(pd.DataFrame.as_matrix)))\n  return (df_np, df_c.index.get_level_values(0).unique())\n\nX_train, indx = data_formatting_score(X_train)\na,b,c=X_train.shape\n\nX_train = X_train.reshape(a,b,c,1)\n\nprint X_train.shape\n(34952, 6, 15, 1)\n<\/code><\/pre>\n\n<ol>\n<li><p>Why am I only getting 34952 columns when I reshape? Also I noted that everytime I run the script the sample size changes.<\/p><\/li>\n<li><p>Can we feed data with different height and width (which is 6 X15 in this case) as input to convnet?<\/p><\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":53210551,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><rnn>","CreationDate":"2018-11-08T15:09:05.987","AcceptedAnswerId":null,"OwnerUserId":2539738.0,"Title":"how does tensorflow train RNNs?","Body":"<p>I just read through and ran the code here: <a href=\"https:\/\/tomaxent.com\/2017\/04\/26\/LSTM-by-Example-using-Tensorflow-Text-Generate\/\" rel=\"nofollow noreferrer\">https:\/\/tomaxent.com\/2017\/04\/26\/LSTM-by-Example-using-Tensorflow-Text-Generate\/<\/a><\/p>\n\n<p>(this guy rips off the following medium.com article, but I can't access medium.com from my work computer); <a href=\"https:\/\/medium.com\/towards-data-science\/lstm-by-example-using-tensorflow-feb0c1968537\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/towards-data-science\/lstm-by-example-using-tensorflow-feb0c1968537<\/a><\/p>\n\n<p>From my previous reading, it is my understanding that to train RNNs, we have to 'unwrap' them into feed forward networks (FFN) for a certain number of steps (along with an extra input for the \"x at time t\"), and set it so that all the weights in the FFN that correspond to a single weight in the RNN are equal.<\/p>\n\n<p>I'm looking at the code and I don't see any 'unwrapping' step, or even a variable indicating the number of steps for which we want to unwrap.<\/p>\n\n<p>Is there another way to train an RNN? Am I just missing the line in the code where that variable is defined?<\/p>\n","answers":[{"AnswerId":"53230869","CreationDate":"2018-11-09T17:45:46.577","ParentId":null,"OwnerUserId":"10382807","Title":null,"Body":"<p>If i am not mistaken there is no such 'unwrapping step'.We generally \"unroll\" a RNN in-order to understand its working properly (through each time step).\nNow,coming to Tensorflow Implementation I found this repo:<a href=\"https:\/\/github.com\/MuhammedBuyukkinaci\/TensorFlow-Text-Generator\/blob\/master\/tensorflow_text_generator.py\" rel=\"nofollow noreferrer\">MuhammedBuyukkinaci\/TensorFlow-Text-Generator<\/a> to be very useful and this might clear up your most of the doubts.<\/p>\n\n<p>Other Useful Links:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/recurrent\" rel=\"nofollow noreferrer\">Tensorflow-RNN<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/rnn_cell\/BasicRNNCell\" rel=\"nofollow noreferrer\">Basic_Rnn_Cell<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/static_rnn\" rel=\"nofollow noreferrer\">Static_RNN Cell<\/a><\/p>\n"}]}
+{"QuestionId":53212098,"AnswerCount":0,"Tags":"<python><apache-spark><keras><pyspark>","CreationDate":"2018-11-08T16:32:31.470","AcceptedAnswerId":null,"OwnerUserId":5640161.0,"Title":"Mapping RDDs using an object's member function","Body":"<p>I would like to map a function <code>foo(x, my_object)<\/code> to an RDD <code>rdd<\/code> <\/p>\n\n<pre><code>rdd2 = rdd.map(lambda x: foo(x, my_object))\n<\/code><\/pre>\n\n<p>where <code>foo<\/code> is defined as<\/p>\n\n<pre><code>def foo(x, some_object):\n\n    return [x, some_object.member_function(x)]\n<\/code><\/pre>\n\n<p>In other words, the mapping applies a member function of the passed object. (In my particular case, the <code>member_function()<\/code> is essentially a call to a Keras <code>predict()<\/code> function.)<\/p>\n\n<p>The problem is that I get the error <code>PicklingError: Could not serialize object: TypeError: can't pickle _thread.lock objects<\/code>. Apparently, Spark has a problem handling certain objects.<\/p>\n\n<p>What is the best way to do this? <\/p>\n\n<p>PS: I'm not dead-set on using RDDs, actually I would have preferred working with DataFrames.<\/p>\n\n<p><strong>EDIT<\/strong>: Unlike the <a href=\"https:\/\/stackoverflow.com\/questions\/53187570\/how-to-pass-deep-learning-model-data-to-map-function-in-spark\">example<\/a> proposed as a duplicate, this problem isn't specific to deep learning. Besides, I am looking for a minimum working example, which the proposed duplicate isn't (references to AWS' S3, etc.). Finally, there is no explanation as to why <code>mapPartitions()<\/code> is the solution. I just want a way to perform a mapping and pass an object. Why can't that work? This has supposedly something to do with Spark's serialization mechanisms, but I'd like to know what the (elegant) workarounds could be.<\/p>\n","answers":[]}
+{"QuestionId":53212485,"AnswerCount":1,"Tags":"<python><tensorflow><scikit-learn><deep-learning><coreml>","CreationDate":"2018-11-08T16:52:07.190","AcceptedAnswerId":53213797.0,"OwnerUserId":5535448.0,"Title":"Is there a way to train an ML model to classify words into categories?","Body":"<p>I'm looking to train an ML model to classify words into several broad categories, in my case: color. So I'll have some pre-defined color buckets like these for example:<\/p>\n\n<pre><code>let blue = [\"blue\", \"royal\", \"dark blue\", \"light blue\"]\nlet red = [\"red\", \"cardinal\", \"dusty red\", \"red polka dot\"]\n<\/code><\/pre>\n\n<p>And I want<\/p>\n\n<p>a) For the model to classify colors already existing in the buckets, i.e. if given \"blue\" it will know that \"blue\" is in the <code>blue<\/code> bucket.<\/p>\n\n<p>b) For the model to take words not seen before, such as \"faded blue\", and to classify them in the correct bucket, in this case <code>blue<\/code> based on a confidence score of some sort.<\/p>\n\n<p>I'm not sure if this is possible and the current method I have is a series of if statements to go about the classifying, but I'm wondering if there is a more intuitive way to do this with an ML model.<\/p>\n","answers":[{"AnswerId":"53213797","CreationDate":"2018-11-08T18:14:04.617","ParentId":null,"OwnerUserId":"6858244","Title":null,"Body":"<p>You can try scikit-learn:<\/p>\n\n<pre><code>import pandas as pd\nfrom sklearn.pipeline import Pipeline\nfrom sklearn.feature_extraction.text import CountVectorizer\nfrom sklearn.linear_model import  LogisticRegression\n\n\ndata = {'blue': [\"blue\", \"royal\", \"dark blue\", \"light blue\"],\n        'red': [\"red\", \"cardinal\", \"dusty red\", \"red polka dot\"]}\n\ntrain_data = pd.DataFrame(data).T.reset_index()\ntrain_data.rename(columns={'index':'target'}, inplace=True)\n\n# predictors\nX = train_data.drop('target', axis=1)\nX = X.apply(lambda x: ','.join(x), axis=1)\n\n# target\ny = train_data.target\n\n# simple toy model\nclf  = Pipeline(steps=[\n        ('vec',  CountVectorizer(ngram_range=(1, 2),)),\n        ('clf', LogisticRegression())\n])\n\n# train a model\nclf.fit(X,y)\n\n# predict a new value\nprint(clf.predict(['faded blue']))\n<\/code><\/pre>\n\n<p>Hope this will set you to the right path :)<\/p>\n\n<p>Results from above:\n<a href=\"https:\/\/i.stack.imgur.com\/cyFDk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cyFDk.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53212507,"AnswerCount":1,"Tags":"<max><pytorch><indices>","CreationDate":"2018-11-08T16:53:00.940","AcceptedAnswerId":53215649.0,"OwnerUserId":1332314.0,"Title":"How to efficiently retrieve the indices of maximum values in a Torch tensor?","Body":"<p>Assume to have a torch tensor, for example of the following shape:<\/p>\n\n<pre><code>x = torch.rand(20, 1, 120, 120)\n<\/code><\/pre>\n\n<p>What I would like now, is to get the indices of the maximum values of each 120x120 matrix. To simplify the problem I would first <code>x.squeeze()<\/code> to work with shape <code>[20, 120, 120]<\/code>. I would then like to get torch tensor which is a list of indices with shape <code>[20, 2]<\/code>.<\/p>\n\n<p>How can I do this fast?<\/p>\n","answers":[{"AnswerId":"53215649","CreationDate":"2018-11-08T20:28:17.023","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>If I get you correctly you don't want the values, but the indices. Unfortunately there is no out of the box solution. There exists an <code>argmax()<\/code> function, but I cannot see how to get it to do exactly what you want.<\/p>\n\n<p>So here is a small workaround, the efficiency should also be okay since we're just dividing tensors:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>n = torch.tensor(4)\nd = torch.tensor(4)\nx = torch.rand(n, 1, d, d)\nm = x.view(n, -1).argmax(1)\n# since argmax() does only return the index of the flattened\n# matrix block we have to calculate the indices by ourself \n# by using \/ and % (\/\/ would also work, but as we are dealing with\n# type torch.long \/ works as well\nindices = torch.cat(((m \/ d).view(-1, 1), (m % d).view(-1, 1)), dim=1)\nprint(x)\nprint(indices)\n<\/code><\/pre>\n\n<p><code>n<\/code> represents your first dimension, and <code>d<\/code> the last two dimensions. I take smaller numbers here to show the result. But of course this will also work for <code>n=20<\/code> and <code>d=120<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>n = torch.tensor(20)\nd = torch.tensor(120)\nx = torch.rand(n, 1, d, d)\nm = x.view(n, -1).argmax(1)\nindices = torch.cat(((m \/ d).view(-1, 1), (m % d).view(-1, 1)), dim=1)\n#print(x)\nprint(indices)\n<\/code><\/pre>\n\n<p>Here is the output for <code>n=4<\/code> and <code>d=4<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensor([[[[0.3699, 0.3584, 0.4940, 0.8618],\n          [0.6767, 0.7439, 0.5984, 0.5499],\n          [0.8465, 0.7276, 0.3078, 0.3882],\n          [0.1001, 0.0705, 0.2007, 0.4051]]],\n\n\n        [[[0.7520, 0.4528, 0.0525, 0.9253],\n          [0.6946, 0.0318, 0.5650, 0.7385],\n          [0.0671, 0.6493, 0.3243, 0.2383],\n          [0.6119, 0.7762, 0.9687, 0.0896]]],\n\n\n        [[[0.3504, 0.7431, 0.8336, 0.0336],\n          [0.8208, 0.9051, 0.1681, 0.8722],\n          [0.5751, 0.7903, 0.0046, 0.1471],\n          [0.4875, 0.1592, 0.2783, 0.6338]]],\n\n\n        [[[0.9398, 0.7589, 0.6645, 0.8017],\n          [0.9469, 0.2822, 0.9042, 0.2516],\n          [0.2576, 0.3852, 0.7349, 0.2806],\n          [0.7062, 0.1214, 0.0922, 0.1385]]]])\ntensor([[0, 3],\n        [3, 2],\n        [1, 1],\n        [1, 0]])\n<\/code><\/pre>\n\n<p>I hope this is what you wanted to get! :)<\/p>\n\n<p><strong><em>Edit:<\/em><\/strong><\/p>\n\n<p>Here is a slightly modified which might be minimally faster (not much I guess :), but it is a bit simpler and prettier:<\/p>\n\n<p>Instead of this like before:<\/p>\n\n<pre><code>m = x.view(n, -1).argmax(1)\nindices = torch.cat(((m \/\/ d).view(-1, 1), (m % d).view(-1, 1)), dim=1)\n<\/code><\/pre>\n\n<p>The necessary reshaping already done on the <code>argmax<\/code> values:<\/p>\n\n<pre><code>m = x.view(n, -1).argmax(1).view(-1, 1)\nindices = torch.cat((m \/\/ d, m % d), dim=1)\n<\/code><\/pre>\n\n<p>But as mentioned in the comments. I don't think it is possible to get much more out of it. <\/p>\n\n<p>One thing you could do, if it is <em>really<\/em> important for you to get the last possible bit of performance improvement out of it, is implementing this above function as a low-level extension (like in C++) for pytorch. <\/p>\n\n<p>This would give you just one function you can call for it and would avoid slow python code.<\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/tutorials\/advanced\/cpp_extension.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/advanced\/cpp_extension.html<\/a><\/p>\n"}]}
+{"QuestionId":53212559,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow><keras>","CreationDate":"2018-11-08T16:55:37.737","AcceptedAnswerId":53226416.0,"OwnerUserId":1391609.0,"Title":"Standard Deviation Pooling with Keras","Body":"<p>I am trying to implement a standard deviation pooling layer using keras. The idea is similar to implement a layer with a functionality similar to <code>AveragePooling1D<\/code>, but calculating standard deviation instead.<\/p>\n\n<p>My first course of action was to try and implement this as a Lambda layer. It should take a 3d tensor such as (batch_size,time,features) and a stride integer (indicating the size of the window). It should return a tensor with shape (batch_size,time,features).<\/p>\n\n<p>My implementation follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow\nimport keras\nfrom keras.layers import Dense, TimeDistributed, Lambda, Input\nimport numpy as np\nimport keras.backend as K\n\ndef stdev_pooling(inputs):\n    data, stride = inputs\n\n    stride = K.cast(stride, dtype='int32')\n\n    print K.dtype(stride), K.dtype(data), '---'\n\n    num_windows = K.shape(data)[1] \/ stride\n\n    idxs = K.arange(num_windows) * stride\n\n    windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())\n\n    windows = K.permute_dimensions(windows, (1,0,2,3))\n\n    stds = K.map_fn(lambda w: K.std(w, axis=1), windows)\n\n    return stds\n\nipt = Input(shape=(None,10))\nd = TimeDistributed(Dense(10))(ipt)\nout = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])\n\nm = keras.Model(inputs=ipt, outputs=out)\nx = np.arange(1000).reshape(1,-1,10)\nm.predict(x).shape\n<\/code><\/pre>\n\n<p>However, my output (which shows the data types for both the stride and the data tensors, in this order) is this:<\/p>\n\n<pre><code>int32 float32 ---\nfloat32 float32 ---\n<\/code><\/pre>\n\n<p>And the stack trace is this:<\/p>\n\n<pre><code>\n\n    ---------------------------------------------------------------------------\n    ValueError                                Traceback (most recent call last)\n     in ()\n          1 ipt = Input(shape=(None,10))\n          2 d = TimeDistributed(Dense(10))(ipt)\n    ----> 3 out = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/base_layer.pyc in __call__(self, inputs, **kwargs)\n        472             if all([s is not None\n        473                     for s in to_list(input_shape)]):\n    --> 474                 output_shape = self.compute_output_shape(input_shape)\n        475             else:\n        476                 if isinstance(input_shape, list):\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/core.pyc in compute_output_shape(self, input_shape)\n        643                 if isinstance(input_shape, list):\n        644                     xs = [K.placeholder(shape=shape) for shape in input_shape]\n    --> 645                     x = self.call(xs)\n        646                 else:\n        647                     x = K.placeholder(shape=input_shape)\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/layers\/core.pyc in call(self, inputs, mask)\n        680         if has_arg(self.function, 'mask'):\n        681             arguments['mask'] = mask\n    --> 682         return self.function(inputs, **arguments)\n        683 \n        684     def compute_mask(self, inputs, mask=None):\n\n     in stdev_pooling(inputs)\n          5     print K.dtype(stride), K.dtype(data), '---'\n          6 \n    ----> 7     num_windows = K.shape(data)[1] \/ stride\n          8 \n          9     idxs = K.arange(num_windows-1) * stride\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/math_ops.pyc in binary_op_wrapper(x, y)\n        848     with ops.name_scope(None, op_name, [x, y]) as name:\n        849       if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):\n    --> 850         return func(x, y, name=name)\n        851       elif not isinstance(y, sparse_tensor.SparseTensor):\n        852         try:\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/math_ops.pyc in _div_python2(x, y, name)\n        972   with ops.name_scope(name, \"div\", [x, y]) as name:\n        973     x = ops.convert_to_tensor(x, name=\"x\")\n    --> 974     y = ops.convert_to_tensor(y, name=\"y\", dtype=x.dtype.base_dtype)\n        975     x_dtype = x.dtype.base_dtype\n        976     y_dtype = y.dtype.base_dtype\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.pyc in convert_to_tensor(value, dtype, name, preferred_dtype)\n        996       name=name,\n        997       preferred_dtype=preferred_dtype,\n    --> 998       as_ref=False)\n        999 \n       1000 \n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.pyc in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx)\n       1092 \n       1093     if ret is None:\n    -> 1094       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n       1095 \n       1096     if ret is NotImplemented:\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.pyc in _TensorTensorConversionFunction(t, dtype, name, as_ref)\n        929     raise ValueError(\n        930         \"Tensor conversion requested dtype %s for Tensor with dtype %s: %r\" %\n    --> 931         (dtype.name, t.dtype.name, str(t)))\n        932   return t\n        933 \n\n    ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor(\"lambda_9\/Placeholder_1:0\", shape=(), dtype=float32)'\n\n<\/code><\/pre>\n\n<p>Interestingly, as far as I understand, it says that the <code>stride<\/code> variable is a <code>float32<\/code>, which shoud be converted to <code>int32<\/code>, although it was declared an <code>int32<\/code> variable <code>K.variable(20, dtype='int32', name='stride_var')<\/code>.<\/p>\n\n<p>What is wrong here? Any help would be much appreciated! Thanks!<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>As @BlackBear suggested, I added an explicit cast to stride and it seems to have solved part of the problem:<\/p>\n\n<pre><code>def stdev_pooling(inputs):\n    data, stride = inputs\n\n    stride = K.cast(stride, dtype='int32')\n\n    print K.dtype(stride), K.dtype(data), '---'\n\n    num_windows = K.shape(data)[1] \/ stride\n\n    idxs = K.arange(num_windows) * stride\n\n    windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())\n\n    windows = K.permute_dimensions(windows, (1,0,2,3))\n\n    stds = K.map_fn(lambda w: K.std(w, axis=1), windows)\n\n    return stds\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>int32 float32 ---\nint32 float32 ---\n<\/code><\/pre>\n\n<p>However, now I have a new error that I have no idea where it comes from!<\/p>\n\n<p>Here is the stack trace:<\/p>\n\n<pre><code>\n\n    ---------------------------------------------------------------------------\n    AttributeError                            Traceback (most recent call last)\n     in ()\n          3 x = np.arange(2000).reshape(2,-1,10)\n          4 \n    ----> 5 m = keras.Model(inputs=ipt, outputs=out)\n          6 \n          7 m.predict(x).shape\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/legacy\/interfaces.pyc in wrapper(*args, **kwargs)\n         89                 warnings.warn('Update your `' + object_name +\n         90                               '` call to the Keras 2 API: ' + signature, stacklevel=2)\n    ---> 91             return func(*args, **kwargs)\n         92         wrapper._original_function = func\n         93         return wrapper\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in __init__(self, *args, **kwargs)\n         91                 'inputs' in kwargs and 'outputs' in kwargs):\n         92             # Graph network\n    ---> 93             self._init_graph_network(*args, **kwargs)\n         94         else:\n         95             # Subclassed network\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in _init_graph_network(self, inputs, outputs, name)\n        235         # Keep track of the network's nodes and layers.\n        236         nodes, nodes_by_depth, layers, layers_by_depth = _map_graph_network(\n    --> 237             self.inputs, self.outputs)\n        238         self._network_nodes = nodes\n        239         self._nodes_by_depth = nodes_by_depth\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in _map_graph_network(inputs, outputs)\n       1351                   layer=layer,\n       1352                   node_index=node_index,\n    -> 1353                   tensor_index=tensor_index)\n       1354 \n       1355     for node in reversed(nodes_in_decreasing_depth):\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index)\n       1338             tensor_index = node.tensor_indices[i]\n       1339             build_map(x, finished_nodes, nodes_in_progress, layer,\n    -> 1340                       node_index, tensor_index)\n       1341 \n       1342         finished_nodes.add(node)\n\n    \/home\/juliano\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index)\n       1310             ValueError: if a cycle is detected.\n       1311         \"\"\"\n    -> 1312         node = layer._inbound_nodes[node_index]\n       1313 \n       1314         # Prevent cycles.\n\n    AttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong>: I've updated my stdev_pooling function and it now returns the correct output. However, I'm still having the <code>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'<\/code> error...<\/p>\n","answers":[{"AnswerId":"53226416","CreationDate":"2018-11-09T13:15:43.480","ParentId":null,"OwnerUserId":"1391609","Title":null,"Body":"<p>After fiddling a bit more with the code and reading about how keras interacts with tensorflow (in many different places, including the source code for tensorflow and keras) I figured out what was wrong.<\/p>\n\n<p>First of all, here's a minimal working example of what I wanted to do:<\/p>\n\n\n\n<pre><code>import tensorflow\nimport keras\nfrom keras.layers import Dense, TimeDistributed, Lambda, Input\nimport numpy as np\nimport keras.backend as K\n\n\ndef stdev_pooling(inputs, stride):\n\n    data = inputs\n\n    padding = K.shape(data)[1] % stride\n\n    data = K.switch(padding &gt; 0, K.temporal_padding(data, padding=(0,stride-padding)), data )\n\n    num_windows = K.shape(data)[1] \/ stride\n\n    idxs = K.arange(num_windows) * stride\n\n    windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())\n\n    windows = K.permute_dimensions(windows, (1,0,2,3))\n\n    stds = K.map_fn(lambda w: K.std(w, axis=1), windows)\n\n    return stds\n\nipt = Input(shape=(None,10))\nd = TimeDistributed(Dense(10))(ipt)\n#stride is an argument to stdev_pooling, not a signal coming from\n#a previous layer. Thus it must be passed in the `arguments`\n#dictionary of the `Lambda` layer.\nout = Lambda(stdev_pooling, arguments={'stride':15})(d)\n\nx = np.arange(2000).reshape(2,-1,10)    \nm = keras.Model(inputs=ipt, outputs=out)    \ny = m.predict(x)\nprint y\nprint y.shape\n<\/code><\/pre>\n\n<p>The problem stemmed from the line <code>out = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])<\/code> in the previous code.<\/p>\n\n<p>When a signal (such as K.variable(...)) is provided as input to the Lambda layer, keras expects it to be connected to an Input layer. Thus, the error <code>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'<\/code>.<\/p>\n\n<p>The solution was simply to provide the stride argument through the <code>arguments<\/code> dictionary of the Lambda layer constructor:<\/p>\n\n<p><code>out = Lambda(stdev_pooling, arguments={'stride':15})(d)<\/code><\/p>\n\n<p>I hope this code helps everyone trying to build some sort of pooling layer in keras. When I have some time I shall write it as a proper Pooling Layer. For now, this Lambda version should do.<\/p>\n"}]}
+{"QuestionId":53212580,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2018-11-08T16:56:45.967","AcceptedAnswerId":53212776.0,"OwnerUserId":10624301.0,"Title":"CPU and GPU Tensorflow Installation","Body":"<p>I'm very new with TensorFlow.<\/p>\n\n<p>I want to run my code on my CUDA gpu. So I've installed TensorFlow -gpu, after I've installed normal TensorFlow.<\/p>\n\n<p>How can I tell Python that it takes gpu based TensorFlow?<\/p>\n","answers":[{"AnswerId":"53212776","CreationDate":"2018-11-08T17:06:26.113","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>If you have tensorflow-gpu installed there really isn't any reason to also have tensorflow. Without the presence of a gpu it will just run on cpu anyway. <\/p>\n\n<p>To be specific in which GPU you use:<\/p>\n\n<pre><code>import os\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\"\n<\/code><\/pre>\n\n<p>in place of <code>\"0\"<\/code> you can either list GPUs (if you have multiple), or <code>\"\"<\/code> if you want it to run on cpu. <\/p>\n\n<p>alternatively, specify in the session:<\/p>\n\n<pre><code>sess = tf.Session(config=tf.ConfigProto(device_count={'GPU': 0}))\n<\/code><\/pre>\n\n<p>Furthermore you can check which version your computer prioritizes by opening python console and typing:<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; tensorflow\n&lt;module 'tensorflow' from \n'\/home\/...\/python3.6\/site- packages\/tensorflow\/__init__.py'&gt;\n                                       ^\n                                       |\n                                      here\n<\/code><\/pre>\n"}]}
+{"QuestionId":53212672,"AnswerCount":4,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-08T17:00:41.520","AcceptedAnswerId":53214077.0,"OwnerUserId":7576115.0,"Title":"Read only mode in keras","Body":"<p>I have cloned human pose estimation keras model from this link <a href=\"https:\/\/github.com\/michalfaber\/keras_Realtime_Multi-Person_Pose_Estimation\" rel=\"noreferrer\">human pose estimation keras<\/a> <\/p>\n\n<p>When I try to load the model on google colab, I get the following error<\/p>\n\n<p>code<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('model.h5')\n<\/code><\/pre>\n\n<p>error<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call \n\nlast)\n&lt;ipython-input-29-bdcc7d8d338b&gt; in &lt;module&gt;()\n      1 from keras.models import load_model\n----&gt; 2 model = load_model('model.h5')\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/saving.py in load_model(filepath, custom_objects, compile)\n    417     f = h5dict(filepath, 'r')\n    418     try:\n--&gt; 419         model = _deserialize_model(f, custom_objects, compile)\n    420     finally:\n    421         if opened_new_file:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/saving.py in _deserialize_model(f, custom_objects, compile)\n    219         return obj\n    220 \n--&gt; 221     model_config = f['model_config']\n    222     if model_config is None:\n    223         raise ValueError('No model found in config.')\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/utils\/io_utils.py in __getitem__(self, attr)\n    300             else:\n    301                 if self.read_only:\n--&gt; 302                     raise ValueError('Cannot create group in read only mode.')\n    303                 val = H5Dict(self.data.create_group(attr))\n    304         return val\n\nValueError: Cannot create group in read only mode.\n<\/code><\/pre>\n\n<p>Can someone please help me understand this read-only mode? How do I load this model?<\/p>\n","answers":[{"AnswerId":"55624161","CreationDate":"2019-04-11T03:15:44.473","ParentId":null,"OwnerUserId":"6652751","Title":null,"Body":"<p>I had a similar issue and solved this way<\/p>\n\n<p><strong>store<\/strong> the <code>graph\\architecture<\/code> in <code>JSON<\/code> format and <code>weights<\/code> in <code>h5<\/code> format <\/p>\n\n<pre><code>import json\n\n# lets assume `model` is main model \nmodel_json = model.to_json()\nwith open(\"model_in_json.json\", \"w\") as json_file:\n    json.dump(model_json, json_file)\n\nmodel.save_weights(\"model_weights.h5\")\n<\/code><\/pre>\n\n<p>then need to <code>load model<\/code> first to <code>create<\/code> <code>graph\\architecture<\/code> and <code>load_weights<\/code> in model<\/p>\n\n<pre><code>from keras.models import load_model\nfrom keras.models import model_from_json\nimport json\n\nwith open('model_in_json.json','r') as f:\n    model_json = json.load(f)\n\nmodel = model_from_json(model_json)\nmodel.load_weights('model_weights.h5')\n<\/code><\/pre>\n"},{"AnswerId":"55126505","CreationDate":"2019-03-12T16:35:28.750","ParentId":null,"OwnerUserId":"1731823","Title":null,"Body":"<p>Easy solution is you have to load the save model in the same way it is saved <\/p>\n\n<p>for example you had created a model like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Build the model\n\nmodel = Sequential()\nmodel.add(Dense(100, input_shape=(10,)))\nmodel.add(Activation('relu'))\n\nmodel.add(Dense(10))\nmodel.add(Activation('softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n\n#fit the model\nmodel.fit(x_train, y_train,\n                    batch_size=10,\n                    epochs=5,\n                    verbose=1,\n                    validation_split=0.1)\n\n#save model\nmodel.save(\"model.h5\")\n<\/code><\/pre>\n\n<p>so if you want to load this model for prediction you have to follow the same steps<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import load_model\n# Build the model\n\nmodel = Sequential()\nmodel.add(Dense(100, input_shape=(10,)))\nmodel.add(Activation('relu'))\n\nmodel.add(Dense(10))\nmodel.add(Activation('softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n\n# load model\nmodel.load_weights('model.h5')\n\nprediction = model.predict( ''' data to predict ''')\n\n<\/code><\/pre>\n"},{"AnswerId":"56000582","CreationDate":"2019-05-06T07:27:47.483","ParentId":null,"OwnerUserId":"4376839","Title":null,"Body":"<p>I used <code>callbacks.ModelCheckpoint<\/code> to save the weights and I had a similar error. I found out that there is a parameter called <code>save_weights_only<\/code><\/p>\n\n<p>If I set <code>save_weights_only=True<\/code>, then when I use load_model() to load the model in another process, it will raise the 'Cannot create group in read only mode.' error.<\/p>\n\n<p>If I set <code>save_weights_only=False<\/code>(which is the default), then I can use load_model() to load the model and use it to do prediction, without compiling the model first.<\/p>\n"},{"AnswerId":"53214077","CreationDate":"2018-11-08T18:33:02.800","ParentId":null,"OwnerUserId":"2103431","Title":null,"Body":"<p>Here is an example Git gist created on Google Collab for you: <a href=\"https:\/\/gist.github.com\/kolygri\/835ccea6b87089fbfd64395c3895c01f\" rel=\"noreferrer\">https:\/\/gist.github.com\/kolygri\/835ccea6b87089fbfd64395c3895c01f<\/a><\/p>\n\n<p>As far as I understand:<\/p>\n\n<blockquote>\n  <p>You have to set and define the architecture of your model and then use model.load_weights('alexnet_weights.h5').<\/p>\n<\/blockquote>\n\n<p>Here is a useful Github conversation link, which hopefully will help you understand the issue better:\n<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6937\" rel=\"noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/6937<\/a><\/p>\n"}]}
+{"QuestionId":53212713,"AnswerCount":1,"Tags":"<python><django><keras>","CreationDate":"2018-11-08T17:03:07.383","AcceptedAnswerId":53214235.0,"OwnerUserId":10619689.0,"Title":"Keras Model Serving- GPU vs CPU","Body":"<p>I would like to put my Keras NER model up to my website which is Django based.<\/p>\n\n<p>My question is that when served, should the model be run on CPU or GPU? How would a GPU handle let's say hundreds of users sending requests at the same time since it cant do multi threading like CPUs can.  <\/p>\n\n<p>Thanks for your time.<\/p>\n","answers":[{"AnswerId":"53214235","CreationDate":"2018-11-08T18:45:24.397","ParentId":null,"OwnerUserId":"1269942","Title":null,"Body":"<p>\"hundreds of users\" will necessitate using a GPU.<\/p>\n\n<p>knowing what your users are doing might give more insight, but I would consider having a dedicated instance for your model.  Design a small program that is persistent and waits for input data using a queue strategy.<\/p>\n\n<p>Let's say you have \"100s\" of users uploading \"text documents\" for your ner model. Your web application would gather\/validate the text upload\/post and make a socket connection to your model server, pass the text, receive the response, parse the response and respond to the user accordingly.<\/p>\n\n<p>Your model server has a socket listener that upon connection, passes the socket+data request to a queue. Your model sits and waits...looking at that queue--takes any item in the queue, processes and passes the results back via the included socket(serialized handle). It's a bit messy in python but works fine.\nThe reasons to have a persistent model running:\n1) for \"100s\" of users, spinning up a model app each time takes several seconds. having it already initialized and ready to go will help user experience.\n2) using one gpu server you don't want competition for gpu resources.<\/p>\n"}]}
+{"QuestionId":53213372,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T17:43:24.963","AcceptedAnswerId":53224701.0,"OwnerUserId":2213201.0,"Title":"Tensorflow, change Tensor values given a condition","Body":"<p>I am translating a numpy code to Tensorflow.<\/p>\n\n<p>It has the following line:<\/p>\n\n<pre><code>netout[..., 5:] *= netout[..., 5:] &gt; obj_threshold\n<\/code><\/pre>\n\n<p>This is not the same Tensorflow syntax, I'm having trouble finding the functions with the same behavior.<\/p>\n\n<p>Firstly I tried:<\/p>\n\n<pre><code>netout[..., 5:] * netout[..., 5:] &gt; obj_threshold\n<\/code><\/pre>\n\n<p>But the return is a boolean only Tensor.\nIn this case I want all values below the <code>obj_threshold<\/code> to be 0.<\/p>\n","answers":[{"AnswerId":"53224701","CreationDate":"2018-11-09T11:16:29.143","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>If you just wanted to make 0 all values below <code>obj_threshold<\/code> you could just do:<\/p>\n\n<pre><code>netout = tf.where(netout &gt; obj_threshold, netout, tf.zeros_like(netout))\n<\/code><\/pre>\n\n<p>Or:<\/p>\n\n<pre><code>netout = netout * tf.cast(netout &gt; obj_threshold, netout.dtype)\n<\/code><\/pre>\n\n<p>However, your case is a bit more complicated, because you only want the change to affect part of the tensor. So one thing you can do is make a boolean mask that is <code>True<\/code> for values over <code>obj_threshold<\/code> or values where the last index is below 5.<\/p>\n\n<pre><code>mask = (netout &gt; obj_threshold) | (tf.range(tf.shape(netout)[-1]) &lt; 5)\n<\/code><\/pre>\n\n<p>Then you can use that with any of the previous methods:<\/p>\n\n<pre><code>netout = tf.where(mask, netout, tf.zeros_like(netout))\n\nnetout = netout * tf.cast(mask, netout.dtype)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53213653,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-08T18:04:21.647","AcceptedAnswerId":null,"OwnerUserId":10458815.0,"Title":"Is an eager-graph compatible same code solution possible?","Body":"<p>I am trying to write code that is eager and graph compatible. However, there is very little information online for how to do this, being a literal footnote on TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/guide\/eager#write_compatible_code\" rel=\"nofollow noreferrer\">website<\/a>. Furthermore, what they have wrote is confusing, saying:<\/p>\n\n<blockquote>\n  <p>The same code written for eager execution will also build a graph during graph execution. Do this by simply running the same code in a new Python session where eager execution is not enabled.<\/p>\n<\/blockquote>\n\n<p>This implies that a same code solution is possible, where the only change required is the addition or removal of <code>tf.enable_eager_execution()<\/code>. <\/p>\n\n<p>Currently I use <code>tf.keras<\/code> to define my model and <code>tf.data<\/code> for my input pipeline. However, many eager operations don't work in graph, with the opposite also being true.<\/p>\n\n<p>For example, I keep track of my number of <code>epochs<\/code> using <code>tf.train.Checkpoint()<\/code>. In eager mode, after restoring I can access it using <code>epochs.numpy()<\/code> to assign its value to a local variable. However, this does not work with graphs, which instead would require <code>sess.run(epochs)<\/code> due to the values not being defined during execution. <\/p>\n\n<p>Again, to compute my gradients in eager I need to use some form of autograd, in my case <code>tf.GradientTape()<\/code>. This is not compatible with graphs, as \"tf.GradientTape.gradients() does not support graph control flow.\"<\/p>\n\n<p>I see that <code>tfe.py_func<\/code> exists, but once again, this only works when eager is not enabled, thus not helping for this problem.<\/p>\n\n<p>So how do I make a same code solution, when it seems that many aspects of eager and graph directly conflict with each other?<\/p>\n","answers":[]}
+{"QuestionId":53213703,"AnswerCount":1,"Tags":"<python><tensorflow><regression><linear-regression>","CreationDate":"2018-11-08T18:08:07.347","AcceptedAnswerId":53347000.0,"OwnerUserId":10625147.0,"Title":"Tensorflow Linear Regression NaN output","Body":"<p>I am trying to write code for a Machine Learning algorithm to learn both Machine Learning concepts and Tensorflow. The algorithm I am trying to write up is:<\/p>\n\n<p>(Not enough reputation to embed an image) <a href=\"https:\/\/i.imgur.com\/lxgC7YV.png\" rel=\"nofollow noreferrer\">https:\/\/i.imgur.com\/lxgC7YV.png<\/a><\/p>\n\n<p>\"Which is equivalent to a piece wise linear regression model.\"<\/p>\n\n<p>From (Equation 7):<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1411.3315.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1411.3315.pdf<\/a><\/p>\n\n<p>I've loaded in the vectors I want to do this on. And initialised my placeholders and variables:<\/p>\n\n<pre><code>size = len(originalVecs)\n_x1 = tf.placeholder(tf.float64, shape=[size, 300], name=\"x1-input\")\n_x2 = tf.placeholder(tf.float64, shape=[size, 300], name=\"x2-input\")\n\n_w = tf.Variable(tf.random_uniform([300,300], -1, 1, dtype = tf.float64), name=\"weight1\")\n<\/code><\/pre>\n\n<p>My prediction, cost, and training step I've set as I have set as:<\/p>\n\n<pre><code>prediction = tf.matmul(_x1,_w)\ncost = tf.reduce_sum(tf.square(tf.norm(prediction - _x2)))\ntrain_step = tf.train.GradientDescentOptimizer(0.01).minimize(cost)\n<\/code><\/pre>\n\n<p>After I initialize I train with the following:<\/p>\n\n<pre><code>for i in range(10000):\n    sess.run(train_step, feed_dict={_x1: timedVecs, _x2 : originalVecs})\n    if i % 1001 == 0:\n        print('Epoch ', i)\n        print('Prediction ', sess.run(prediction, feed_dict={_x1: timedVecs, _x2 : originalVecs}).shape)\n<\/code><\/pre>\n\n<p>When I run my code it is wildly unstable and just grows within about 20 iterations to just print NaNs. I think I'm doing a couple of things wrong but I do not know how to correct.<\/p>\n\n<p>The shape of the prediction is [20,300] when I would expect it to be [1,300]. I want it to predict based off a single x1 and x2, rather than all at once then try to learn from the sum of the error for all data points (what I assume piecewise is). I'm not sure how to go about this however as I think currently I'm minimising based on the 20,300 matrix rather than the sum of 20 1,300 matrices. <\/p>\n\n<p>I assume matmul is correct as multiply is element wise?<\/p>\n\n<p>I am entering my input data as a list of np arrays. Each np array being a data point with 300 dimensions.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53347000","CreationDate":"2018-11-17T00:16:11.553","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Generally I'd avoid square roots in losses. The issue is that the derivative of <code>x**0.5<\/code> is <code>0.5 * x**-0.5<\/code>, which means dividing by <code>x<\/code>. If <code>x<\/code> is ever zero, this will produce NaNs. In this case the square root comes from <code>tf.norm<\/code> and is immediately followed by <code>tf.square<\/code>, but the operations aren't fused together and don't cancel.<\/p>\n\n<p>Simplifying your loss expression to <code>tf.reduce_sum(tf.square(prediction - _x2))<\/code> should make things more stable.<\/p>\n"}]}
+{"QuestionId":53214189,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><keras-layer>","CreationDate":"2018-11-08T18:41:29.687","AcceptedAnswerId":53214433.0,"OwnerUserId":1958843.0,"Title":"Custom Keras layer failing with functional APIs","Body":"<p>I am using Keras Functional API to introduce a custom layer, <code>GaussianLayer<\/code>, that returns a list of two elements (see <code>call<\/code> method below):<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\nfrom tensorflow.keras.layers import Layer\n\ndef gaussian_loss(y_true, y_pred):\n    \"\"\"\n    y_true is a scalar (float)\n    y_pred is a tensor [mu, sigma]\n    \"\"\"\n    print(y_pred.shape)\n    return tf.reduce_mean(0.5*tf.log(y_pred[1]) + 0.5*tf.div(tf.square(y_true - y_pred[0]), y_pred[0])) + 1\n\nclass GaussianLayer(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(GaussianLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.kernel_1 = self.add_weight(name='kernel', \n                                      shape=(30, self.output_dim),\n                                      initializer='uniform',\n                                      trainable=True)\n        self.kernel_2 = self.add_weight(name='kernel', \n                                      shape=(30, self.output_dim),\n                                      initializer='uniform',\n                                      trainable=True)\n        self.bias_1 = self.add_weight(name='bias',\n                                    shape=(self.output_dim),\n                                    initializer='zeros',\n                                    trainable=True)\n        self.bias_2 = self.add_weight(name='bias',\n                                    shape=(self.output_dim),\n                                    initializer='zeros',\n                                    trainable=True)\n        super(GaussianLayer, self).build(input_shape) \n\n    def call(self, x):\n        output_mu  = K.dot(x, self.kernel_1) + self.bias_1\n        output_sig = K.dot(x, self.kernel_2) + self.bias_2\n        output_sig_pos = K.log(1 + K.exp(output_sig)) + 1e-06  \n        return [output_mu, output_sig_pos]\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.output_dim)\n\n# This returns a tensor\ninputs = Input(shape=(1,))\nx = Dense(24, activation='relu')(inputs)\nx = Dense(30, activation='relu')(x)\npredictions = GaussianLayer(1)(x)\n\nmodel = Model(inputs, predictions)\nmodel.compile(loss=gaussian_loss, optimizer='adam')\nmodel.fit(train_x, train_y, epochs=600)\n<\/code><\/pre>\n\n<p>Unfortunately this code fails with an <code>AttributeError<\/code>:<\/p>\n\n<pre><code>----------------------------------------------------------------\nAttributeError                 Traceback (most recent call last)\n&lt;ipython-input-200-e93218491799&gt; in &lt;module&gt;()\n     49 x = Dense(24, activation='relu')(inputs)\n     50 x = Dense(30, activation='relu')(x)\n---&gt; 51 predictions = GaussianLayer(1)(x)\n     52 \n     53 model = Model(inputs, predictions)\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    783       if in_deferred_mode or build_graph and have_all_keras_metadata(inputs):\n    784         inputs, outputs = self._set_connectivity_metadata_(\n--&gt; 785             inputs, outputs, args, kwargs)\n    786       if context.executing_eagerly():\n    787         return outputs\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _set_connectivity_metadata_(self, inputs, outputs, args, kwargs)\n    895     kwargs.pop('mask', None)  # `mask` should not be serialized.\n    896     self._add_inbound_node(\n--&gt; 897         input_tensors=inputs, output_tensors=outputs, arguments=kwargs)\n    898     return inputs, outputs\n    899 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _add_inbound_node(self, input_tensors, output_tensors, arguments)\n   1076         input_tensors=input_tensors,\n   1077         output_tensors=output_tensors,\n-&gt; 1078         arguments=arguments)\n   1079 \n   1080     # Update tensor history metadata.\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __init__(self, outbound_layer, inbound_layers, node_indices, tensor_indices, input_tensors, output_tensors, arguments)\n   1747         # For compatibility with external Keras, we use the deprecated\n   1748         # accessor here.\n-&gt; 1749         layer.outbound_nodes.append(self)\n   1750     # For compatibility with external Keras, we use the deprecated\n   1751     # accessor here.\n\nAttributeError: 'Dense' object has no attribute 'outbound_nodes'\n<\/code><\/pre>\n\n<p>I know I need to use the Functional API in order to return a list to be passed to <code>gaussian_loss<\/code>, but unfortunately this does not seem to work either.<\/p>\n","answers":[{"AnswerId":"53214433","CreationDate":"2018-11-08T18:59:03.650","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First, don't mix <code>keras<\/code> and <code>tensorflow.keras<\/code> modules. Only use one of them in your code.<\/p>\n\n<p>Second, fix the shape of both <code>bias<\/code> variables by adding a <code>,<\/code> at the end to make them a tuple, i.e. <code>shape=(self.output_dim,)<\/code>.<\/p>\n\n<p>Third, if your custom layer returns a list of two tensors as output then the output shape must be consistent with this, i.e.:<\/p>\n\n<pre><code>def compute_output_shape(self, input_shape):\n    return [(input_shape[0], self.output_dim), (input_shape[0], self.output_dim)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53214444,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-08T18:59:29.737","AcceptedAnswerId":53228976.0,"OwnerUserId":7322746.0,"Title":"In tensorflow serving, how to store a list in feature dictionary?","Body":"<p>I'm pretty new with tensorflow serving, now I'm working with client-end coding.<\/p>\n\n<p>With the basic tutorial, I know I need to build a feature dictionary like:<\/p>\n\n<pre><code>feature_dict={\n    'input_content':tf.train.Feature(...)\n    'input_label':tf.train.Feature(...)\n}\n<\/code><\/pre>\n\n<p>Then,<\/p>\n\n<pre><code>model_input=tf.train.Example(feature=tf.train.Features(feature=feature_dict))\n<\/code><\/pre>\n\n<p>Now, my question is, how can I put a list into the feature_dict?\nLike, I have a 10 dimension list, I want to set it as the 'input_content', how can I get that?<\/p>\n","answers":[{"AnswerId":"53228976","CreationDate":"2018-11-09T15:47:30.603","ParentId":null,"OwnerUserId":"3549221","Title":null,"Body":"<p>A <code>tf.train.Feature<\/code> contains lists which may hold <em>zero or more<\/em> values. The lists could be of type <code>BytesList, FloatList,<\/code> or <code>Int64List<\/code>. <\/p>\n\n<p>The following code adds a single float element (<code>float_element<\/code>) to the <code>tf.train.Feature<\/code> <\/p>\n\n<p><code>tf.train.Feature(float_list=tf.train.FloatList(value=[float_element]))<\/code><\/p>\n\n<p>Notice that the <code>float_element<\/code> is surrounded by square brackets (<code>[]<\/code>), i.e., a list is being created with a single element.\nWhile trying to add a list (<code>float_list<\/code>), one should not use square brackets like the following code snippet.<\/p>\n\n<p><code>tf.train.Feature(float_list=tf.train.FloatList(value=float_list))<\/code><\/p>\n"}]}
+{"QuestionId":53214462,"AnswerCount":1,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2018-11-08T19:00:45.177","AcceptedAnswerId":53392598.0,"OwnerUserId":4829166.0,"Title":"Compiling the tensorflow lite minimal example","Body":"<p>I'm trying to use tensorflow lite on android. I'm getting an error during compilation with android studio: <\/p>\n\n<pre><code>undefined reference to 'tflite::InterpreterBuilder::operator()(std::__ndk1::unique_ptr&lt;tflite::Interpreter, std::__ndk1::default_delete&lt;tflite::Interpreter&gt; &gt;*)'\n<\/code><\/pre>\n\n<p>I've managed to get a reproducible example while trying to compile <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/lite\/examples\/minimal\" rel=\"nofollow noreferrer\">this minimal example from the tf repo<\/a>.<\/p>\n\n<p>Steps to get the error<\/p>\n\n<ol>\n<li><p>Download the android ndk (I've tried both 15c and 16b and they both failed) . Call <code>build\/tools\/make_standalone_toolchain.py --arch arm --api 21 --stl=libc++ --install-dir android-toolchain<\/code><\/p><\/li>\n<li><p>Build libtensorflow, this is in my build file.<\/p><\/li>\n<\/ol>\n\n<p><code>tensorflow\/lite\/BUILD<\/code><\/p>\n\n<pre><code>cc_binary(\n    name = \"libtensorflowlite.so\",\n    linkopts=[\n        \"-shared\",\n        \"-Wl,-soname=libtensorflowlite.so\",\n    ],  \n    linkshared = 1,\n    copts = tflite_copts(),\n    deps = [ \n        \":framework\",\n        \"\/\/tensorflow\/lite\/kernels:builtin_ops\",\n    ],  \n)\n<\/code><\/pre>\n\n<p>I call it like:<\/p>\n\n<pre><code>bazel build \/\/tensorflow\/lite:libtensorflowlite.so --crosstool_top=\/\/external:android\/crosstool --cpu=armeabi-v7a --host_crosstool_top=@bazel_tools\/\/tools\/cpp:toolchain --cxxopt=\"-std=c++11\"\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li><p>I clone the flatbuffer repo (a dependency) and compile it.<\/p><\/li>\n<li><p>Then I try and compile <code>minimal.cc<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>I call it like:<\/p>\n\n<pre><code>android-toolchain\/bin\/clang++  -std=c++11 -Igit\/tensorflow-android -Igit\/flatbuffers\/include -Lgit\/tensorflow-android -Lgit\/flatbuffers\/build minimal.cc -ltensorflowlite -lflatbuffers\n<\/code><\/pre>\n\n<p>(<code>libtensorflow.so<\/code> is in <code>git\/tensorflow-android<\/code>, <code>libflatbuffers.a<\/code> in <code>git\/flatbuffers\/build<\/code>)<\/p>\n\n<p>What's going on? I have no idea why this method is being called, I'm not calling the object with two arguments.<\/p>\n","answers":[{"AnswerId":"53392598","CreationDate":"2018-11-20T12:02:28.997","ParentId":null,"OwnerUserId":"4829166","Title":null,"Body":"<p>The error was me not being on ndk 18. I hadn't switched to it because when compiling bazel was specifically telling me to use &lt;=16!<\/p>\n"}]}
+{"QuestionId":53214616,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network><mnist>","CreationDate":"2018-11-08T19:12:04.677","AcceptedAnswerId":null,"OwnerUserId":10614496.0,"Title":"recognise misaligned pictures of handwritten digits with CNN","Body":"<p>Hello I'm trying to make a programm that recognises handwritten digits (from MNIST). I've actually done it and it works quite good. (I made a CNN that reaches 98% of accuracy on the test dataset)\nThen I created a programm where you could wrtie the digits with your mouse and the computer said wich digit you had written. When you write the digits in the middle of the screen, it works good, but when you write almost the same thing but a little bit misaligned, it dosen't work anymore. Is there a special cnn architecture for that? I tryed several architecture but it's alway the same. Maybe the problem is in the training data? --> because all the the digits are written in the middle of the picture.<\/p>\n","answers":[]}
+{"QuestionId":53214858,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-08T19:29:46.490","AcceptedAnswerId":53215339.0,"OwnerUserId":2723396.0,"Title":"Problems restoring a saved model in tensorflow, how to debug?","Body":"<p>After training a model in tensorflow, it is saved as following:<\/p>\n\n<pre><code>saver = tf.train.Saver()\nsaver.save(sess,'myModel\/Path\/Model_1')\n<\/code><\/pre>\n\n<p>Generating files called:<\/p>\n\n<ul>\n<li>Model_1.meta<\/li>\n<li>Model_1.index<\/li>\n<li>Model_1.data-000000-of-000001<\/li>\n<li>checkpoint<\/li>\n<\/ul>\n\n<p>Now to restore the model after creating a new session, and initializing the tensorflow graph in exactly the same way as originally created, I restore it as follows:<\/p>\n\n<pre><code>sess = tf.Session()    \n# Initialize the variables (i.e. assign their default value)\ninit = tf.global_variables_initializer()\nsess.run(init)\n\nimported_meta = tf.train.Saver()\nimported_meta.restore(sess,'myModel\/Path\/Model_1.meta')\n<\/code><\/pre>\n\n<p>Which throws the following error:<\/p>\n\n<pre><code>InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [6152,32] rhs shape= [6164,80]\n [[Node: save_2\/Assign_3 = Assign[T=DT_FLOAT, _class=[\"loc:@DGNS\/bidirectional_rnn\/bw\/basic_lstm_cell\/kernel\"], use_locking=true, validate_shape=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](DGNS\/bidirectional_rnn\/bw\/basic_lstm_cell\/kernel, save_2\/RestoreV2\/_111)]]\n\nCaused by op u'save_2\/Assign_3', defined at:\n  File \"\/usr\/lib\/python2.7\/dist-packages\/spyderlib\/widgets    \/externalshell\/start_ipython_kernel.py\", line 205, in &lt;module&gt;\n __ipythonkernel__.start()\n \"\/usr\/lib\/python2.7\/dist-packages\/IPython\/kernel\/zmq\/kernelapp.py\", line 459, in start\nioloop.IOLoop.instance().start()\n  File \"\/usr\/lib\/python2.7\/dist-packages\/zmq\/eventloop\/ioloop.py\", line 162, in start\nsuper(ZMQIOLoop, self).start()\n  File \"\/usr\/lib\/python2.7\/dist-packages\/zmq\/eventloop\/minitornado\/ioloop.py\", line 830, in start\nself._run_callback(callback)\n  File \"\/usr\/lib\/python2.7\/dist-packages\/zmq\/eventloop\/minitornado\/ioloop.py\", line 603, in _run_callback\nret = callback()\n<\/code><\/pre>\n\n<p>...\n... etc<\/p>\n\n<p>I need help understanding what is happening here. The error hints to some shape mismatch issue. But I do not understand how this can be as I have used exactly the same code for generating the model and initializing a new graph. The only difference in the code is the model loading part. <\/p>\n\n<p>How can I start debugging this error in order to get a hint on how to load my model correclty?<\/p>\n","answers":[{"AnswerId":"53215339","CreationDate":"2018-11-08T20:04:42.300","ParentId":null,"OwnerUserId":"6150310","Title":null,"Body":"<p>I'm pretty sure you are not supposed to load the .meta file. It's tricky to understand since it outputs 3 different files for the checkpoints. Try this:<\/p>\n\n<pre><code>with tf.Session() as sess:\nnew_saver = tf.train.import_meta_graph(\n    'myModel\/Path\/Model_1.meta', clear_devices=True)\nnew_saver.restore(sess, 'myModel\/Path\/Model_1')\n<\/code><\/pre>\n\n<p>Also, just for clarification, are you storing your full model in a .pb file as well, or just generating these checkpoints?<\/p>\n"}]}
+{"QuestionId":53215433,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-08T20:12:44.113","AcceptedAnswerId":null,"OwnerUserId":10616367.0,"Title":"How to do a \"Tree search\" style fit in keras","Body":"<p>I'm trying to apply the following algorithm to how I'm fitting my model<\/p>\n\n<ol>\n<li>Train the model for X epochs<\/li>\n<li>Select the best model from those X epochs<\/li>\n<li>Train N models starting from the best model to come from the first fit<\/li>\n<li>Train another N models on M of the most successful networks<\/li>\n<li>Continue expanding this tree for G generations and take the best model out of all the fitting attempts<\/li>\n<\/ol>\n\n<p>I'm a beginner and have no idea of where to start so some help would be appreciated, thanks.<\/p>\n","answers":[]}
+{"QuestionId":53215477,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-08T20:15:52.137","AcceptedAnswerId":54835266.0,"OwnerUserId":4058766.0,"Title":"Where can I identify which Keras methods are TPU compatible methods?","Body":"<p>I recently learned that UpSampling2D is not a TPU supported operation.<\/p>\n\n<p>How do I find out which Keras functions are compatible with a TPU in Colab? I was unable to find any documentation on the subject. <\/p>\n","answers":[{"AnswerId":"54835266","CreationDate":"2019-02-22T21:12:38.057","ParentId":null,"OwnerUserId":"10355421","Title":null,"Body":"<p>Generally all Keras layers should be supported, but there is not a nice summary. There is a more general summary that is related to TPU Ops <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tensorflow-ops\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>In particular, UpSampling2D is a work in progress at this time, but you should be encouraged to file issues <a href=\"https:\/\/github.com\/tensorflow\/tpu\" rel=\"nofollow noreferrer\">here<\/a> if you are having more problems with Keras and TPUs. <\/p>\n"}]}
+{"QuestionId":53215722,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2018-11-08T20:33:27.423","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"How to Create a Multi-inputs Convolutional Neural Network Model for Images Classification?","Body":"<p>I am beginner in deep learning and I hope if you help me to solve my issue.<\/p>\n\n<p>I want to create a CNN model that takes two inputs of images and produce one output which is the class of the two images. The model takes one image from dataset type1 and one image from dataset type2.\nI have two datasets: type1 and type2, and each dataset contains the same number of classes, but the number of images in each class in the dataset type1 is higher than the number of images in each class in the dataset type2. The following is the structure of the datasets.<\/p>\n\n<p>The model should take one image from Type1 dataset and one image from Type2 dataset and then classify these images to one class (ClassA or ClassB or------).<\/p>\n\n<pre><code>Type1 dataset\n|Train\n              |ClassA\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassB\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassC\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassD\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n       ----------------\n|Validate\n            -----------\n|Test\n           --------------\n\nType2 dataset\n|Train\n              |ClassA\n                             |image1\n                             |image2\n                            -----\n              |ClassB\n                             |image1\n                             |image2\n                            -----\n              |ClassC\n                             |image1\n                             |image2\n                            -----\n              |ClassD\n                             |image1\n                             |image2\n                            -----\n       ----------------\n|Validate\n            -----------\n|Test\n           --------------\n<\/code><\/pre>\n\n<p>So, I want to create a model that inputs two images (from type 1 and 2), as long as they are from the same class. Also, I want each image from type1 be paired with every images from type2 from the same class. How can I do this???<\/p>\n\n<p>The code:<\/p>\n\n<pre><code>in1 = Input(...)  \n x = Conv2D(...)(in1)\n--------\n--------\n out1 = Dense(...)(x)  \n\n in2 = Input(...)  \nx = Conv2D(...)(in2)\n--------\n--------\nout2 = Dense(...)(x)  \n\nconcatenated_layer = concatenate([out1, out2])  # merge the outputs of the two models\noutput_layer = Dense(no_classes, activation='softmax', name='prediction')(concatenated_layer)\nmodal= Model(inputs=[in1, in2], outputs=[output_layer])\n\n\n\ninput_imgen = ImageDataGenerator(rescale = 1.\/255, \n                                   shear_range = 0.2, \n                                   zoom_range = 0.2,\n                                   rotation_range=5.,\n                                   horizontal_flip = True)\n\ntest_imgen = ImageDataGenerator()\n\n\n\ndef generate_generator_multiple(generator,dir1, dir2, batch_size, img_height,img_width):\n    genX1 = generator.flow_from_directory(dir1,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n\n    genX2 = generator.flow_from_directory(dir2,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n    while True:\n            X1i = genX1.next()\n            X2i = genX2.next()\n            yield [X1i[0], X2i[0]], X2i[1]  #Yield both images and their mutual label\n\n\ninputgenerator=generate_generator_multiple(generator=input_imgen,\n                                           dir1=train_iris_data,\n                                           dir2=train_face_data,\n                                           batch_size=32,\n                                           img_height=224,\n                                           img_width=224)       \n\ntestgenerator=generate_generator_multiple(generator=test_imgen,\n                                          dir1=valid_iris_data,\n                                          dir2=valid_face_data,\n                                          batch_size=1,\n                                          img_height=224,\n                                          img_width=224) \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53215816,"AnswerCount":2,"Tags":"<python><machine-learning><deep-learning><pytorch><gradient-descent>","CreationDate":"2018-11-08T20:41:31.850","AcceptedAnswerId":null,"OwnerUserId":10625599.0,"Title":"Making custom non-trivial loss function in pytorch","Body":"<p>I'm just started with pytorch and trying to understand how to deal with custom loss functions, especially with some non trivial ones. <\/p>\n\n<p><strong>Problem 1<\/strong>. I'd like to stimulate my nn to maximize true positive rate and at the same time minimize\nfalse discovery rate. For example increase total score on +2 for true positive, and decrease on -5 for false positive.<\/p>\n\n<pre><code>def tp_fp_loss(yhat, y):\n    total_score = 0\n    for i in range(y.size()):\n        if is_tp(yhat[i],y[i]):\n            total_score += 2\n        if is_fp(yhat[i],y[i]):\n            total_score -= 5\n    return -total_score\n<\/code><\/pre>\n\n<p><strong>Problem 2<\/strong>. In case when y is a list of positive and negative rewards (y = [10,-5, -40, 23, 11, -7]), \nstimulate nn to maximize sum of rewards. <\/p>\n\n<pre><code>def max_reward_loss(yhat,y):\n    r = torch.autograd.Variable(torch.Tensor(y[yhat &gt;= .5]), requires_grad=True).sum()                    \n    return -r\n<\/code><\/pre>\n\n<p>Maybe I'm not completely understand some autograd mechanics, functions which I implemented correctly calculate loss but\nlearning with them doesnt work :( What I'm doing wrong? Can anybody help me with some working solution of any of that problems?<\/p>\n","answers":[{"AnswerId":"53217308","CreationDate":"2018-11-08T22:47:39.553","ParentId":null,"OwnerUserId":"361609","Title":null,"Body":"<p>@Shai already summed it up: Your loss function is not differentiable.<\/p>\n\n<p>One way to think about it is that your loss function should be plottable, and the \"downhill\" slope should \"roll\" toward the desired model output.  In order to plot your loss function, fix <code>y_true=1<\/code> then plot <code>[loss(y_pred) for y_pred in np.linspace(0, 1, 101)]<\/code> where <code>loss<\/code> is your loss function, and make sure your plotted loss function has the slope as desired.  In your case, it sounds like you want to weight the the loss more strongly when it is on the wrong side of the threshold.  As long as you can plot it, and the slope is always downhill toward your target value (no flat spots or uphill slopes on the way from a valid prediction to the target value), your model should learn from it.<\/p>\n\n<p>Also note that if you're just trying to take into account some business objective which prioritizes precision over recall, you could accomplish this by training to convergence with cross entropy or some well-known loss function, and then by tuning your model threshold based on your use case.  A higher threshold would normally prioritize precision, and a lower threshold would normally prioritize recall.  After you've trained, you can then evaluate your model at a variety of thresholds and choose the most appropriate.<\/p>\n"},{"AnswerId":"53216489","CreationDate":"2018-11-08T21:34:22.270","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your loss function is not differentiable - you cannot compute its gradient (go ahead and try).<br>\nYou should look as something like <a href=\"https:\/\/stackoverflow.com\/questions\/27632440\/infogainloss-layer\">infogain loss<\/a><\/p>\n"}]}
+{"QuestionId":53216162,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><object-detection><object-detection-api>","CreationDate":"2018-11-08T21:08:30.437","AcceptedAnswerId":null,"OwnerUserId":416206.0,"Title":"How to train Tensorflow Object Detection images that do not contain objects?","Body":"<p>I am training an object detection network using Tensorflow's object detection,<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection<\/a><\/p>\n\n<p>I can successfully train a network based on my own images and labels.\nHowever, I have a large dataset of images that do not contain any of my labeled objects, and I want to be able to train the network to not detect anything in these images.<\/p>\n\n<p>From what I understand with Tensorflow object detection, I need to give it a set of images and corresponding XML files that box and label the objects in the image. The scripts convert the XML to CSV and then to another format for the training, and do not allow XML files that have no objects.<\/p>\n\n<p>How to give an image and XML files that have no objects?<\/p>\n\n<p>Or, how does the network learn what is not an object?<\/p>\n\n<p>For example if you want to detect \"hot dogs\" you can train it with a set of images with hot dogs. But how to train it what is not a hot dog?<\/p>\n","answers":[{"AnswerId":"54575263","CreationDate":"2019-02-07T14:11:04.860","ParentId":null,"OwnerUserId":"11027619","Title":null,"Body":"<p>You might want to take a look at <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/3578#issuecomment-375267920\" rel=\"nofollow noreferrer\">this solution<\/a>. <\/p>\n\n<p>For for the tensorflow object detection API to include your negative examples, you need to add the negative examples to the csv file you have created from the xml, either by modifying the script that generates the csv file or by adding the examples afterwards.<\/p>\n\n<p>For generating xml-files without class labels using <a href=\"https:\/\/github.com\/tzutalin\/labelImg\" rel=\"nofollow noreferrer\">LabelImg<\/a>, you can do this by pressing \"Verify Image\".<\/p>\n"},{"AnswerId":"53247641","CreationDate":"2018-11-11T10:06:15.767","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>An Object Detection CNN can learn what is not an object, simply by letting it see examples of images without any labels.<\/p>\n\n<p>There are two main architecture types: <\/p>\n\n<ol>\n<li>two-stages, with first stage object\/region proposal (RPN), and second - classification and bounding box fine-tuning; <\/li>\n<li>one-stage, which directly classifies and regresses BB based on the feature vector corresponding to a certain cell in the feature map.<\/li>\n<\/ol>\n\n<p>In any case, there's a part which is responsible to decide what is an object and what's not. In RPN you have \"objectness\" score, and in one-stages there's the confidence of classification, where you usually a background class (i.e. everything which is not the supported classes).<\/p>\n\n<p>So in both cases, in case a specific example in an image doesn't have any supported class, you teach the CNN to decrease the objectness score or increase the background confidence correspondingly.<\/p>\n"}]}
+{"QuestionId":53216184,"AnswerCount":2,"Tags":"<android><tensorflow><object-detection><object-detection-api>","CreationDate":"2018-11-08T21:10:57.187","AcceptedAnswerId":53315381.0,"OwnerUserId":416206.0,"Title":"How to configure Tensorflow object detection Android demo to work with Inception v2","Body":"<p>We have built an Android app based on the Tensorflow object detection Android demo app. It works when using a Mobilenet network, but crashes if we try to use an Inception v2 based network.<\/p>\n\n<p>It is possible for Tensorflow Inception v2 object detection to work on Android? <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection<\/a><\/p>\n\n<p>We are using the exact same code as the Tensorflow detector demo here.\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/TensorFlowObjectDetectionAPIModel.java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/TensorFlowObjectDetectionAPIModel.java<\/a><\/p>\n\n<p>Models from zoo here,\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md<\/a><\/p>\n\n<p>Using TF Object Detection API.\nIf we use a network pb file trained using ssd_mobilenet_v2_coco the demo app works.\nIf we use a network pb file trained using faster_rcnn_inception_v2_coco it crashes, (see below)<\/p>\n\n<p>Is it possible for the Android app to work with the Inception v2 model?\n(the mobile-net accuracy is very bad, but Inception is much better)<\/p>\n\n<pre><code>11-14 12:11:47.817 7122-7199\/org.tensorflow.demo E\/AndroidRuntime: FATAL EXCEPTION: inference\nProcess: org.tensorflow.demo, PID: 7122\njava.nio.BufferOverflowException\nat java.nio.FloatBuffer.put(FloatBuffer.java:444)\nat org.tensorflow.Tensor.writeTo(Tensor.java:488)\nat org.tensorflow.contrib.android.TensorFlowInferenceInterface.fetch(TensorFlowInferenceInterface.java:488)\nat org.tensorflow.contrib.android.TensorFlowInferenceInterface.fetch(TensorFlowInferenceInterface.java:442)\nat org.tensorflow.demo.TensorFlowObjectDetectionAPIModel.recognizeImage(TensorFlowObjectDetectionAPIModel.java:170)\nat org.tensorflow.demo.DetectorActivity$3.run(DetectorActivity.java:288)\nat android.os.Handler.handleCallback(Handler.java:739)\nat android.os.Handler.dispatchMessage(Handler.java:95)\nat android.os.Looper.loop(Looper.java:148)\nat android.os.HandlerThread.run(HandlerThread.java:61)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53267418","CreationDate":"2018-11-12T17:42:26.153","ParentId":null,"OwnerUserId":"8185230","Title":null,"Body":"<p>It should be possible to use SSD Inception, although not advisable. Inception is quite large for mobile, and I don't believe we don't have quantization support for it right now.<\/p>\n"},{"AnswerId":"53315381","CreationDate":"2018-11-15T08:42:35.077","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>I read about this issue once.<\/p>\n\n<p>I think the problem is in this line of your code:<\/p>\n\n<pre><code>  private static final int MAX_RESULTS = 100;\n<\/code><\/pre>\n\n<p>This creates an array for output with the specified length. I think SSD mobilenet gives this number of predictions at maximum, but default faster RCNN (without any configurations from your side) gives you more. Try to increase this value for example to 500.<\/p>\n"}]}
+{"QuestionId":53216245,"AnswerCount":2,"Tags":"<tensorflow><model><keras><save><load>","CreationDate":"2018-11-08T21:16:04.580","AcceptedAnswerId":null,"OwnerUserId":10625975.0,"Title":"tensorflow keras save and load model","Body":"<p>I have run this example and I got the following error when I try to save the model.<\/p>\n\n<pre><code>import tensorflow as tf\nimport h5py\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(512, activation=tf.nn.relu),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer='adam',\n          loss='sparse_categorical_crossentropy',\n          metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=2)\nval_loss, val_acc = model.evaluate(x_test, y_test)\nprint(val_loss, val_acc)\n\nmodel.save('model.h5')\n\nnew_model = tf.keras.models.load_model('model.h5')\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"\/home\/zneic\/PycharmProjects\/test\/venv\/test.py\", line 23, in &lt;module&gt;\nmodel.save('model.h5')\nFile \"\/home\/zneic\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/network.py\", line 1359, in save\n'Currently `save` requires model to be a graph network. Consider '\nNotImplementedError: Currently `save` requires model to be a graph network. Consider using `save_weights`, in order to save the weights of the model.\n<\/code><\/pre>\n","answers":[{"AnswerId":"54255886","CreationDate":"2019-01-18T14:19:47.117","ParentId":null,"OwnerUserId":"6345161","Title":null,"Body":"<p>I have the same problem, and I solved it. I don't konw why, but it works. You can modify as this:<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n  layers.Flatten(input_shape=(28, 28)),\n  layers.Dense(512, activation=tf.nn.relu, input_shape=(784,)),\n  layers.Dropout(0.2),\n  layers.Dense(10, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n"},{"AnswerId":"53216815","CreationDate":"2018-11-08T22:01:57.990","ParentId":null,"OwnerUserId":"6150310","Title":null,"Body":"<p>Your weights don't seem to be saved or loaded back into the session. Can you try saving the graph and the weights separately and loading them separately?<\/p>\n\n<pre><code>model_json = model.to_json()\nwith open(\"model.json\", \"w\") as json_file:\n    json_file.write(model_json)\n\nmodel.save_weights(\"model.h5\")\n<\/code><\/pre>\n\n<p>Then you can load them:<\/p>\n\n<pre><code>def loadModel(jsonStr, weightStr):\n    json_file = open(jsonStr, 'r')\n    loaded_nnet = json_file.read()\n    json_file.close()\n\n    serve_model = tf.keras.models.model_from_json(loaded_nnet)\n    serve_model.load_weights(weightStr)\n\n    serve_model.compile(optimizer=tf.train.AdamOptimizer(),\n                        loss='categorical_crossentropy',\n                        metrics=['accuracy'])\n    return serve_model\n\nmodel = loadModel('model.json', 'model.h5')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53216564,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network><loss-function>","CreationDate":"2018-11-08T21:41:30.997","AcceptedAnswerId":null,"OwnerUserId":9548759.0,"Title":"How to ignore part of input and output in Keras?","Body":"<p>I'm trying to train a model that takes n values as input and output n values. The problem is that n can be from 1 to 700. So I build a network with 700 as input and 700 as output. The extra inputs and outputs are set to zero.\nWhen training the model, I don't care about if the extra outputs are accurate or not. So I tried to define my own loss function as follows:<\/p>\n\n<pre><code>def mse_truncate(y_true, y_pred):\n    def fn(x):\n        return tf.cond(x &lt; 0.01,lambda: 0.0,lambda: 1.0)\n    #Ignore the square error if y_true[i] is near zero\n    sgn = tf.map_fn(fn,y_true)\n    return K.mean(sgn * K.square(y_true-y_pred),axis=-1)\n<\/code><\/pre>\n\n<p>This function works on console.\nBut when I compile the model, I get an error:<\/p>\n\n<pre><code>model.compile(optimizer='sgd',loss=mse_truncate, metrics=['accuracy'])\nValueError: Shape must be rank 0 but is rank 1 for 'loss_5\/dense_2_loss\/map\/while\/cond\/Switch' (op: 'Switch') with input shapes: [?], [?].\n<\/code><\/pre>\n\n<p>Can someone tell me what's wrong here?\nOr are there better ways to handle the variable length input and output?<\/p>\n\n<p>Note: \nMore on the problem, the input is a sequence(length &lt;= 700) and the output is the distance between the first element and each element in the sequence.<\/p>\n","answers":[{"AnswerId":"55930625","CreationDate":"2019-05-01T01:30:54.427","ParentId":null,"OwnerUserId":"5339857","Title":null,"Body":"<p>You could use <code>tf.where<\/code> and <code>tf.gather<\/code> to only take those values that you care about into consideration, e.g.:<\/p>\n\n<pre><code>indices = tf.where(tf.greater(y_true, 0.01)) #  or `tf.less`, `tf.equal` etc.\nloss = K.mean(K.square(tf.gather(y_true, indices) - tf.gather(y_pred, indices))))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53217564,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><tensorflow-estimator>","CreationDate":"2018-11-08T23:12:43.300","AcceptedAnswerId":53289161.0,"OwnerUserId":1605795.0,"Title":"SessionRunHook returning empty SessionRunValues after run","Body":"<p>I'm trying to write a hook that will allow me to compute some global metrics (rather than batch-wise metrics). To prototype, I thought I'd get a simple hook up and running that would capture and remember true positives. It looks like this:<\/p>\n\n<pre><code>class TPHook(tf.train.SessionRunHook):\n\n    def after_create_session(self, session, coord):\n        print(\"Starting Hook\")\n\n        tp_name = 'metrics\/f1_macro\/TP'\n        self.tp = []\n        self.args = session.graph.get_operation_by_name(tp_name)\n        print(f\"Got Args: {self.args}\")\n\n    def before_run(self, run_context):\n        print(\"Starting Before Run\")\n        return tf.train.SessionRunArgs(self.args)\n\n    def after_run(self, run_context, run_values):\n        print(\"After Run\")\n        print(f\"Got Values: {run_values.results}\")\n<\/code><\/pre>\n\n<p>However, the values returned in the \"after_run\" part of the hook are always None. I tested this in both the train and evaluation phase. Am I misunderstanding something about how the SessionRunHooks are supposed to work?<\/p>\n\n<hr>\n\n<p>Maybe relevant information:\nThe model was build in keras and converted to an estimator with the <code>keras.estimator.model_to_estimator()<\/code> function. The model has been tested and works fine, and the op that I'm trying to retrieve in the hook is defined in this code block:<\/p>\n\n<pre><code>def _f1_macro_vector(y_true, y_pred):\n    \"\"\"Computes the F1-score with Macro averaging.\n\n    Arguments:\n        y_true {tf.Tensor} -- Ground-truth labels\n        y_pred {tf.Tensor} -- Predicted labels\n\n    Returns:\n        tf.Tensor -- The computed F1-Score\n    \"\"\"\n    y_true = K.cast(y_true, tf.float64)\n    y_pred = K.cast(y_pred, tf.float64)\n\n    TP = tf.reduce_sum(y_true * K.round(y_pred), axis=0, name='TP')\n    FN = tf.reduce_sum(y_true * (1 - K.round(y_pred)), axis=0, name='FN')\n    FP = tf.reduce_sum((1 - y_true) * K.round(y_pred), axis=0, name='FP')\n\n    prec = TP \/ (TP + FP)\n    rec = TP \/ (TP + FN)\n\n    # Convert NaNs to Zero\n    prec = tf.where(tf.is_nan(prec), tf.zeros_like(prec), prec)\n    rec = tf.where(tf.is_nan(rec), tf.zeros_like(rec), rec)\n\n    f1 = 2 * (prec * rec) \/ (prec + rec)\n\n    # Convert NaN to Zero\n    f1 = tf.where(tf.is_nan(f1), tf.zeros_like(f1), f1)\n\n    return f1\n<\/code><\/pre>\n","answers":[{"AnswerId":"53289161","CreationDate":"2018-11-13T20:40:01.453","ParentId":null,"OwnerUserId":"1605795","Title":null,"Body":"<p>In case anyone runs into the same problem, I found out how to restructure the program so that it worked. Although the documentation makes it sound like I can pass raw ops into the <code>SessionRunArgs<\/code>, it seems like it requires actual tensors (maybe this is a misreading on my part).\nThis is pretty easy to accomplish - I just changed the <code>after_create_session<\/code> code to what's shown below.<\/p>\n\n<pre><code>def after_create_session(self, session, coord):\n\n    tp_name = 'metrics\/f1_macro\/TP'\n    self.tp = []\n    tp_tensor = session.graph.get_tensor_by_name(tp_name+':0')\n\n    self.args = [tp_tensor]\n<\/code><\/pre>\n\n<p>And this successfully runs. <\/p>\n"}]}
+{"QuestionId":53217596,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-08T23:15:47.410","AcceptedAnswerId":53217826.0,"OwnerUserId":4567788.0,"Title":"Tensorflow AttributeError: 'numpy.float32' object has no attribute 'value'","Body":"<p>I'm using this code as a base and wish to view the learning progress by the loss within Tensorboard. After adding the writer and attempting to add_summary, I receive the above error presented within the title. <\/p>\n\n<pre><code>\"\"\" Convolutional Neural Network.\n\nBuild and train a convolutional neural network with TensorFlow.\nThis example is using the MNIST database of handwritten digits\n(http:\/\/yann.lecun.com\/exdb\/mnist\/)\n\nAuthor: Aymeric Damien\nProject: https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/\n\"\"\"\n\nfrom __future__ import division, print_function, absolute_import\n\nimport tensorflow as tf\n\n# Import MNIST data\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n\n# Training Parameters\nlearning_rate = 0.001\nnum_steps = 200\nbatch_size = 128\ndisplay_step = 10\n\n# Network Parameters\nnum_input = 784 # MNIST data input (img shape: 28*28)\nnum_classes = 10 # MNIST total classes (0-9 digits)\ndropout = 0.75 # Dropout, probability to keep units\n\n# tf Graph input\nX = tf.placeholder(tf.float32, [None, num_input])\nY = tf.placeholder(tf.float32, [None, num_classes])\nkeep_prob = tf.placeholder(tf.float32) # dropout (keep probability)\n\n\n# Create some wrappers for simplicity\ndef conv2d(x, W, b, strides=1):\n    # Conv2D wrapper, with bias and relu activation\n    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')\n    x = tf.nn.bias_add(x, b)\n    return tf.nn.relu(x)\n\n\ndef maxpool2d(x, k=2):\n    # MaxPool2D wrapper\n    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],\n                          padding='SAME')\n\n\n# Create model\ndef conv_net(x, weights, biases, dropout):\n    # MNIST data input is a 1-D vector of 784 features (28*28 pixels)\n    # Reshape to match picture format [Height x Width x Channel]\n    # Tensor input become 4-D: [Batch Size, Height, Width, Channel]\n    x = tf.reshape(x, shape=[-1, 28, 28, 1])\n\n    # Convolution Layer\n    conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n    # Max Pooling (down-sampling)\n    conv1 = maxpool2d(conv1, k=2)\n\n    # Convolution Layer\n    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n    # Max Pooling (down-sampling)\n    conv2 = maxpool2d(conv2, k=2)\n\n    # Fully connected layer\n    # Reshape conv2 output to fit fully connected layer input\n    fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])\n    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n    fc1 = tf.nn.relu(fc1)\n    # Apply Dropout\n    fc1 = tf.nn.dropout(fc1, dropout)\n\n    # Output, class prediction\n    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n    return out\n\n# Store layers weight &amp; bias\nweights = {\n    # 5x5 conv, 1 input, 32 outputs\n    'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),\n    # 5x5 conv, 32 inputs, 64 outputs\n    'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),\n    # fully connected, 7*7*64 inputs, 1024 outputs\n    'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),\n    # 1024 inputs, 10 outputs (class prediction)\n    'out': tf.Variable(tf.random_normal([1024, num_classes]))\n}\n\nbiases = {\n    'bc1': tf.Variable(tf.random_normal([32])),\n    'bc2': tf.Variable(tf.random_normal([64])),\n    'bd1': tf.Variable(tf.random_normal([1024])),\n    'out': tf.Variable(tf.random_normal([num_classes]))\n}\n\n# Construct model\nlogits = conv_net(X, weights, biases, keep_prob)\nprediction = tf.nn.softmax(logits)\n\n# Define loss and optimizer\nloss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(\n    logits=logits, labels=Y))\noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\ntrain_op = optimizer.minimize(loss_op)\n\n\n# Evaluate model\ncorrect_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n# Initialize the variables (i.e. assign their default value)\ninit = tf.global_variables_initializer()\n\n# Start training\nwith tf.Session() as sess:\n\n    # Run the initializer\n    sess.run(init)\n\n    for step in range(1, num_steps+1):\n        batch_x, batch_y = mnist.train.next_batch(batch_size)\n        # Run optimization op (backprop)\n        sess.run(train_op, feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.8})\n        if step % display_step == 0 or step == 1:\n            # Calculate batch loss and accuracy\n            loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,\n                                                                 Y: batch_y,\n                                                                 keep_prob: 1.0})\n            print(\"Step \" + str(step) + \", Minibatch Loss= \" + \\\n                  \"{:.4f}\".format(loss) + \", Training Accuracy= \" + \\\n                  \"{:.3f}\".format(acc))\n\n    print(\"Optimization Finished!\")\n\n    # Calculate accuracy for 256 MNIST test images\n    print(\"Testing Accuracy:\", \\\n        sess.run(accuracy, feed_dict={X: mnist.test.images[:256],\n                                      Y: mnist.test.labels[:256],\n                                      keep_prob: 1.0}))\n<\/code><\/pre>\n\n<p>Running it works completely fine, I then added the writer variable to write to the root tmp directory of my OS.<\/p>\n\n<pre><code># Start training\nwith tf.Session() as sess:\n\n    writer = tf.summary.FileWriter(\"\/tmp\/log\/\", graph=sess.graph)\n\n    # Run the initializer\n    sess.run(init)\n\n    for step in range(1, num_steps+1):\n        batch_x, batch_y = mnist.train.next_batch(batch_size)\n        # Run optimization op (backprop)\n        sess.run(train_op, feed_dict={X: batch_x, Y: batch_y, keep_prob: 0.8})\n        if step % display_step == 0 or step == 1:\n            # Calculate batch loss and accuracy\n            summary, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,\n                                                                 Y: batch_y,\n                                                                 keep_prob: 1.0})\n            print(\"Step \" + str(step) + \", Minibatch Loss= \" + \\\n                  \"{:.4f}\".format(summary) + \", Training Accuracy= \" + \\\n                  \"{:.3f}\".format(acc))\n\n            writer.add_summary(summary=summary, global_step=step)\n            writer.flush()\n        writer.close()\n\n    print(\"Optimization Finished!\")\n\n    # Calculate accuracy for 256 MNIST test images\n    print(\"Testing Accuracy:\", \\\n        sess.run(accuracy, feed_dict={X: mnist.test.images[:256],\n                                      Y: mnist.test.labels[:256],\n                                      keep_prob: 1.0}))\n<\/code><\/pre>\n\n<p>But that's when the error occurs. What exactly am I doing wrong and what is the correction that needs to be applied to fix this. I'm still quite new to Tensorflow but I've used summary writer before without this error.<\/p>\n\n<p>I also used this tutorial as well in hopes of fixing this issue:\n<a href=\"https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard<\/a><\/p>\n\n<p>Error just so everyone knows the culprit<\/p>\n\n<pre><code>Step 1, Minibatch Loss= 98733.8750, Training Accuracy= 0.102\nTraceback (most recent call last):\n  File \"convolutional_network_raw.py\", line 137, in &lt;module&gt;\n    writer.add_summary(summary=summary, global_step=step)\n  File \"\/home\/kyle\/.conda\/envs\/csc\/lib\/python3.6\/site-packages\/tensorflow\/python\/summary\/writer\/writer.py\", line 126, in add_summary\n    for value in summary.value:\nAttributeError: 'numpy.float32' object has no attribute 'value'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53217826","CreationDate":"2018-11-08T23:45:22.740","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The summary that goes into the file writer is a <code>tf.Summary<\/code> object.  You need to first convert the loss value into a summary object, for example below (also change the variable in the print so you don't get an error):<\/p>\n\n<pre><code>        loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,\n                                                             Y: batch_y,\n                                                             keep_prob: 1.0})\n        summary = tf.Summary(value=[tf.Summary.Value(tag=\"loss\", simple_value=loss)])\n        print(\"Step \" + str(step) + \", Minibatch Loss= \" + \\\n              \"{:.4f}\".format(loss) + \", Training Accuracy= \" + \\\n              \"{:.3f}\".format(acc))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53217965,"AnswerCount":1,"Tags":"<python><tensorflow><pickle>","CreationDate":"2018-11-09T00:03:56.250","AcceptedAnswerId":53218175.0,"OwnerUserId":10499813.0,"Title":"'ascii' codec can't decode byte 0x8b in position 6: ordinal not in range(128)","Body":"<p>I am trying to code a cnn with tensorflow but I keep coming up with this error:<\/p>\n\n<pre><code>UnicodeDecodeError                        Traceback (most recent call last)\n&lt;ipython-input-20-a02172d91c0c&gt; in &lt;module&gt;()\n     39 # Load all the data batches.\n     40 for i in range(5):\n---&gt; 41     data_batch = unpickle( 'data_batch_' + str(i + 1))\n     42 \n     43     train_data = np.append(train_data, data_batch[b'data'])\n\n&lt;ipython-input-20-a02172d91c0c&gt; in unpickle(file)\n     27     import pickle\n     28     with open(file, 'rb') as fo:\n---&gt; 29         dict = pickle.load(fo)\n     30         dict = dict.encode('ascii', 'ignore')\n     31     return dict\n\nUnicodeDecodeError: 'ascii' codec can't decode byte 0x8b in position 6: ordinal not in range(128)\n<\/code><\/pre>\n\n<p>I am not sure what to do I have tried everything but I am still getting the same error. Here is my code: <\/p>\n\n<pre><code># IMAGE RECOGNITION\n\n# Tensorflow and numpy to create the neural network\nimport tensorflow as tf\nimport numpy as np\n\n# Matplotlib to plot info to show our results\nimport matplotlib.pyplot as plt\n\n# OS to load files and save checkpoints\nimport os\n\n# LOADING THE DATA:\n\n# LOADING CIFAR data from file:\n# Load cifar data from file\n# Load MNIST data from tf examples\n\n# Load cifar data from file\n\nimage_height = 32\nimage_width = 32\n\ncolor_channels = 3\n\nmodel_name = \"cifar\"\n\n\ndef unpickle(file):\n    import pickle\n    with open(file, 'rb') as fo:\n        dict = pickle.load(fo)\n\n    return dict\n\n\n\n\ntrain_data = np.array([])\ntrain_labels = np.array([])\n\n# Load all the data batches.\nfor i in range(5):\n    data_batch = unpickle( 'data_batch_' + str(i + 1))\n\n    train_data = np.append(train_data, data_batch[b'data'])\n    train_labels = np.append(train_labels, data_batch[b'labels'])\n\n# Load the eval batch.\neval_batch = unpickle( 'test_batch')\n\neval_data = eval_batch[b'data']\neval_labels = eval_batch[b'labels']\n\n# Load the english category names.\ncategory_names_bytes = unpickle('batches.meta')[b'label_names']\ncategory_names = list(map(lambda x: x.decode(\"utf-8\"), category_names_bytes))\n\n\n# TODO: Process Cifar data\n\ndef process_data(data):\n    float_data = np.array(data, dtype=float) \/ 255.0\n\n    reshaped_data = np.reshape(float_data, (-1, color_channels, image_height, image_width))\n\n    # The incorrect image\n\n    transposed_data = np.transpose(reshaped_data, [0, 2, 3, 1])\n\n    return transposed_data\n\n\ntrain_data = process_data(train_data)\n\neval_data = process_data(eval_data)\n\n# TODO: The Neural Network\n# CONVOLUTIONAL NEURAL NETWORK CLASS:\n<\/code><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"53218175","CreationDate":"2018-11-09T00:31:24.227","ParentId":null,"OwnerUserId":"10611983","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>pickle.load(fo, encoding='latin1')\n<\/code><\/pre>\n\n<p>It's probably a Python 2\/3 compatibility issue. BTW, you should try not to use things like <code>dict<\/code> as variable names, as it overwrites Python built-ins.<\/p>\n"}]}
+{"QuestionId":53218325,"AnswerCount":1,"Tags":"<c++><c><api><tensorflow>","CreationDate":"2018-11-09T00:54:01.717","AcceptedAnswerId":null,"OwnerUserId":10325666.0,"Title":"tensorflow c_api.h Why GPU is slower than CPU","Body":"<p>I have trained a <code>lstm mode<\/code> and I load the model in my C project to predict my data.I find the GPU is slowly than CPU. <\/p>\n\n<p>Here is my configuration:<\/p>\n\n<ul>\n<li>gpu:nvidia TITAN X<\/li>\n<li>CUDA\uff1a9.0<\/li>\n<li>cudnn:7.0<\/li>\n<li>cpu: intel E5<\/li>\n<li>tensorflow: 1.11.0<\/li>\n<\/ul>\n\n<p>I predicted about 200 data items and every data items calls the function \uff1a<\/p>\n\n<pre><code>TF_SessionRun(sess,\nnullptr, \/\/ Run options.\n&amp;input_op, &amp;input_tensor, 1, \/\/ Input tensors, input tensor values, number of inputs.\n&amp;out_op, &amp;output_tensor, 1, \/\/ Output tensors, output tensor values, number of outputs.\nnullptr, 0, \/\/ Target operations, number of targets.\nnullptr, \/\/ Run metadata.\nstatus \/\/ Output status.\n);\n<\/code><\/pre>\n\n<p>Every time GPU is slower than the CPU.<\/p>\n\n<p>Is my method wrong ?<\/p>\n\n<p>Is there a way to increase the speed? <\/p>\n\n<p>Can I enter data in batches for prediction and if so how?<\/p>\n\n<p>This is the c_api.h <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/c_api.h\" rel=\"nofollow noreferrer\">enter link description here<\/a><\/p>\n","answers":[{"AnswerId":"53220422","CreationDate":"2018-11-09T05:38:30.320","ParentId":null,"OwnerUserId":"8858832","Title":null,"Body":"<p>There is an overhead to transfer data to the GPUs memory , you see this trend sometimes even in Python for extremely small datasets , where the CPU sometimes outperforms the GPU as most of time the gpu isn't utilized but rather is waiting for the next batch of data to come in , with a dataset size of 200 this seems like the most plausible reason . <\/p>\n\n<p>As for making it faster it faster the Python api had an option in batch mode to preload the data into the gpu which somewhat offsets this issue , check for a similar option in the c api<\/p>\n"}]}
+{"QuestionId":53218337,"AnswerCount":1,"Tags":"<tensorflow><amazon-s3>","CreationDate":"2018-11-09T00:55:32.593","AcceptedAnswerId":null,"OwnerUserId":10626567.0,"Title":"Issue accessing S3 from Tensorflow","Body":"<p>With the following config.<\/p>\n\n<pre><code>os.environ['AWS_ACCESS_KEY_ID'] = 'xxxxxx'\nos.environ['AWS_SECRET_ACCESS_KEY'] = 'xxxxxxxx'\nos.environ['AWS_REGION'] = 'us-west-2'\nos.environ['S3_ENDPOINT'] = 's3-us-west-2.amazonaws.com'\nos.environ['S3_USE_HTTPS'] = '1'                          \nos.environ['S3_VERIFY_SSL'] = '1' \n\nprint(file_io.stat('s3:\/\/abcd\/def.txt')) \n<\/code><\/pre>\n\n<p>I get the error  <\/p>\n\n<blockquote>\n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/lib\/io\/file_io.py in stat(filename)\n      556   with errors.raise_exception_on_not_ok_status() as status:\n      557     pywrap_tensorflow.Stat(compat.as_bytes(filename), file_statistics, status)\n  --> 558     return file_statistics\n      559 \n      560 <\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)\n      526             None, None,\n      527             compat.as_text(c_api.TF_Message(self.status.status)),\n  --> 528             c_api.TF_GetCode(self.status.status))<br>\n      529     # Delete the underlying status object from memory otherwise it stays alive<br>\n      530     # as there is a reference to status from this from the traceback due to<\/p>\n  \n  <p>NotFoundError: Object s3:\/\/abcd\/def.txt does not exist<\/p>\n<\/blockquote>\n\n<p>Note this file does exist.<\/p>\n\n<p>I also get the following error on a write and close.  <\/p>\n\n<blockquote>\n  <p>UnknownError: PermanentRedirect: Unable to parse ExceptionName: PermanentRedirect Message: The bucket you are attempting to access must be addressed using the specified endpoint. Please send all future requests to this endpoint.<\/p>\n<\/blockquote>\n\n<p>What more is needed to fix this?<\/p>\n","answers":[{"AnswerId":"53219031","CreationDate":"2018-11-09T02:33:19.910","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>you can use endpoint like \n mybucket.s3-us-west-2.amazonaws.com<\/p>\n\n<p>then use s3:\/\/pathtofile to access <\/p>\n"}]}
+{"QuestionId":53218360,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T01:01:25.163","AcceptedAnswerId":null,"OwnerUserId":7909469.0,"Title":"How to print first channel of the filter?","Body":"<p>In the following code I want to print 1st channel of the filter. How can I do that? I want to print 2 by 2 filter but if I print filter[:,:,0] it gives me a 3rd order tensor shouldn't it be simply 2 by 2 filter?<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    filter= tf.Variable(tf.random_normal([2,2,3,32]))\n\n    init = tf.initialize_all_variables()\n    sess = tf.Session()\n    sess.run(init)\n    print(sess.run(filter))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53219600,"AnswerCount":2,"Tags":"<artificial-intelligence><pytorch><gradient-descent>","CreationDate":"2018-11-09T03:49:50.130","AcceptedAnswerId":53225404.0,"OwnerUserId":3131604.0,"Title":"Pytorch - Getting gradient for intermediate variables \/ tensors","Body":"<p>As an exercice in pytorch framework (0.4.1) , I am trying to display the gradient of X (gX or dSdX) in a simple Linear layer (Z = X.W + B). To simplify my toy example, I backward() from a sum of Z (not a loss).<\/p>\n\n<p>To sum up, I want gX(dSdX) of S=sum(XW+B).<\/p>\n\n<p><strong>The problem is that the gradient of Z (dSdZ) is None<\/strong>. As a result, gX is wrong too of course.<\/p>\n\n<pre><code>import torch\nX = torch.tensor([[0.5, 0.3, 2.1], [0.2, 0.1, 1.1]], requires_grad=True)\nW = torch.tensor([[2.1, 1.5], [-1.4, 0.5], [0.2, 1.1]])\nB = torch.tensor([1.1, -0.3])\nZ = torch.nn.functional.linear(X, weight=W.t(), bias=B)\nS = torch.sum(Z)\nS.backward()\nprint(\"Z:\\n\", Z)\nprint(\"gZ:\\n\", Z.grad)\nprint(\"gX:\\n\", X.grad)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>Z:\n tensor([[2.1500, 2.9100],\n        [1.6000, 1.2600]], grad_fn=&lt;ThAddmmBackward&gt;)\ngZ:\n None\ngX:\n tensor([[ 3.6000, -0.9000,  1.3000],\n        [ 3.6000, -0.9000,  1.3000]])\n<\/code><\/pre>\n\n<p>I have exactly the same result if I use nn.Module as below:<\/p>\n\n<pre><code>class Net1Linear(torch.nn.Module):\n    def __init__(self, wi, wo,W,B):\n        super(Net1Linear, self).__init__()\n        self.linear1 = torch.nn.Linear(wi, wo)\n        self.linear1.weight = torch.nn.Parameter(W.t())\n        self.linear1.bias = torch.nn.Parameter(B)\n    def forward(self, x):\n        return self.linear1(x)\nnet = Net1Linear(3,2,W,B)\nZ = net(X)\nS = torch.sum(Z)\nS.backward()\nprint(\"Z:\\n\", Z)\nprint(\"gZ:\\n\", Z.grad)\nprint(\"gX:\\n\", X.grad)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53225404","CreationDate":"2018-11-09T12:04:32.210","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>First of all you only calculate gradients for tensors where you enable the gradient by setting the <code>requires_grad<\/code> to <code>True<\/code>.<\/p>\n\n<p>So your output is just as one would expect. You get the gradient for <code>X<\/code>.<\/p>\n\n<p>PyTorch does not save gradients of intermediate results for performance reasons. So you will just get the gradient for those tensors you set <code>requires_grad<\/code> to <code>True<\/code>.<\/p>\n\n<p>However you can use <code>register_hook<\/code> to extract the intermediate grad during calculation or to save it manually. Here I just save it to the <code>grad<\/code> variable of tensor <code>Z<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\n# function to extract grad\ndef set_grad(var):\n    def hook(grad):\n        var.grad = grad\n    return hook\n\nX = torch.tensor([[0.5, 0.3, 2.1], [0.2, 0.1, 1.1]], requires_grad=True)\nW = torch.tensor([[2.1, 1.5], [-1.4, 0.5], [0.2, 1.1]])\nB = torch.tensor([1.1, -0.3])\nZ = torch.nn.functional.linear(X, weight=W.t(), bias=B)\n\n# register_hook for Z\nZ.register_hook(set_grad(Z))\n\nS = torch.sum(Z)\nS.backward()\nprint(\"Z:\\n\", Z)\nprint(\"gZ:\\n\", Z.grad)\nprint(\"gX:\\n\", X.grad)\n<\/code><\/pre>\n\n<p>This will output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Z:\n tensor([[2.1500, 2.9100],\n        [1.6000, 1.2600]], grad_fn=&lt;ThAddmmBackward&gt;)\ngZ:\n tensor([[1., 1.],\n        [1., 1.]])\ngX:\n tensor([[ 3.6000, -0.9000,  1.3000],\n        [ 3.6000, -0.9000,  1.3000]])\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n\n<p><em>Btw.: Normally you would want the gradient to be activated for your parameters - so your weights and biases. Because what you would do right now when using the optimizer, is altering your inputs <code>X<\/code> and not your weights <code>W<\/code> and bias <code>B<\/code>. So usually gradient is activated for <code>W<\/code> and <code>B<\/code> in such a case.<\/em><\/p>\n"},{"AnswerId":"53232291","CreationDate":"2018-11-09T19:40:56.347","ParentId":null,"OwnerUserId":"3131604","Title":null,"Body":"<p>blue-phoenox, thanks for your answer. I am pretty happy to have heard about register_hook().<\/p>\n\n<p>What led me to think that I had a wrong gX is that it was independant of the values of X. I will have to do the math to understand it. But using CCE Loss instead of SUM makes things much more clean. So I updated the example for those who might be interested. Using SUM was a bad idea in this case.<\/p>\n\n<pre><code>T_dec = torch.tensor([0, 1])\nX = torch.tensor([[0.5, 0.8, 2.1], [0.7, 0.1, 1.1]], requires_grad=True)\nW = torch.tensor([[2.7, 0.5], [-1.4, 0.5], [0.2, 1.1]])\nB = torch.tensor([1.1, -0.3])\nZ = torch.nn.functional.linear(X, weight=W.t(), bias=B)\nprint(\"Z:\\n\", Z)\nL = torch.nn.CrossEntropyLoss()(Z,T_dec)\nZ.register_hook(lambda gZ: print(\"gZ:\\n\",gZ))\nL.backward()\nprint(\"gX:\\n\", X.grad)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>Z:\n tensor([[1.7500, 2.6600],\n        [3.0700, 1.3100]], grad_fn=&lt;ThAddmmBackward&gt;)\ngZ:\n tensor([[-0.3565,  0.3565],\n        [ 0.4266, -0.4266]])\ngX:\n tensor([[-0.7843,  0.6774,  0.3209],\n        [ 0.9385, -0.8105, -0.3839]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53219666,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T03:57:08.337","AcceptedAnswerId":null,"OwnerUserId":1558715.0,"Title":"How can I parallelize Tensorflow's map_fn in GPU?","Body":"<p>can't post the code in question, but I'll do my best to describe my scenario.<\/p>\n\n<p>I have a custom tensorflow operation that I defined using Tensorflow's Autograph. It contains a lot of tensorflow control flow operations automatically converted by autograph. I run this operation on a lot of batches and realized it runs really slow on the GPU because it's not being parallelized through map_fn, yet on the CPU it runs really fast and is being parallelized. I verified CPU is being parallelized and GPU is not by sticking a tf.Print and the beginning and end of the operation, if there is only one operation running at a time the \"Start\" and \"End\" would show up in order, otherwise they wouldn't. <\/p>\n\n<p>In summary, any ideas why my operation would be automatically parallelized on CPU, but not on GPU considering the fact that it has a lot of control flow operations? I've also already tried manually setting the \"parallel_iterations\" parameter of map_fn.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53220245,"AnswerCount":0,"Tags":"<keras><deep-learning><artificial-intelligence>","CreationDate":"2018-11-09T05:17:38.653","AcceptedAnswerId":null,"OwnerUserId":10074721.0,"Title":"freeze layers while passing a part of features","Body":"<p>I want to design a network using keras like this pic. And there is a problem I need some help. F* is a feature tensor like a sentence or a picture, the shared layers is a block comprise several layers. F* will merge together after passing the shared layer. Then merged feature will pass an output layer. The structure is described below.<\/p>\n\n<p><strong>The problem is I want to train this network only use F1. Namely, when F2 pass the shared layers, shared layers are frozen.<\/strong><\/p>\n\n<p>I would very appreciate if you could answer me with pseudocode.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/M0iO6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/M0iO6.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53220397,"AnswerCount":0,"Tags":"<python><tensorflow><asynchronous><flask><keras>","CreationDate":"2018-11-09T05:34:56.087","AcceptedAnswerId":null,"OwnerUserId":10165516.0,"Title":"Update the training accuracy to the front end","Body":"<p>I am doing a web app , when the button is clicked newly updated data will be trained. When the training is on progress, I should show the  progress on the the front end. Progress means for every epoch I should show the accuracy and error level. I could easily get it from keras but How to update it in the front end. <\/p>\n\n<p>I am using KERAS for the model training and FLASK framework for the web. Is there any one face the same problem as me. <\/p>\n","answers":[]}
+{"QuestionId":53220915,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-11-09T06:35:01.883","AcceptedAnswerId":53223739.0,"OwnerUserId":7482828.0,"Title":"External loss function in Tensorflow","Body":"<p>I have a simple 2 layer dense NN which I want to use a regression model to compute 4 number of given <code>~ 700<\/code> features of an image. Unfortunately, I do not have ground truth elements, so I use custom loss function. Here's the source of the function:<\/p>\n\n<pre><code>def loss_function(logits, img, g, compare_img):\n\n    final_img = img_pipeline(vga_8b=img, g=g%external color gamma function%)\n\n    with tf.name_scope('Loss'):\n        loss = score(gt_image=compare_img, curr_img=final_img)\n        return loss\n<\/code><\/pre>\n\n<p>Where <code>logits<\/code> are the current evaluated 4 numbers, g is just a interpolated function used as color gamma for the image, img is external grayscale image used to generate the final result image used for the <code>score<\/code> function. <code>compare_img<\/code> is not a ground truth image, but some statistical values <code>(kept in python dict)<\/code> used in the score function to evaluate the current produced image.\nUnfortunately, I can't feed <code>g<\/code> and <code>compare_img<\/code> as they are python function and python dictionary which cannot be converted to tensors.<\/p>\n\n<p>Is there a way to hack it somehow and achieve the desired result?<\/p>\n\n<p>Thanks in advance!  <\/p>\n","answers":[{"AnswerId":"53223739","CreationDate":"2018-11-09T10:13:04.713","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>You can use external functions with tensorflow with <code>tf.map<\/code> but I'm afraid to to say, that these are not able to calculate gradients through it. but you loss functions needs to be derivable in every case. So you have to write the function in tensorflow.<\/p>\n\n<p>For your dict values you can create a  lookuptable with <\/p>\n\n<p><code>table = tf.contrib.lookup.HashTable(\n  tf.contrib.lookup.KeyValueTensorInitializer(keys, values), -1)<\/code><\/p>\n"}]}
+{"QuestionId":53220979,"AnswerCount":1,"Tags":"<python-3.x><keras><time-series><lstm><recurrent-neural-network>","CreationDate":"2018-11-09T06:41:41.743","AcceptedAnswerId":null,"OwnerUserId":7655447.0,"Title":"How to get data in proper shape to feed to LSTM layer in keras for sequence to sequence prediction","Body":"<p>I have dataframe as following for time series where <code>SETTLEMENTDATE<\/code> is index. I want to take first row, i.e <code>2018-11-01 14:30:00<\/code> and values of <code>T_1<\/code>, <code>T_2<\/code>, <code>T_3<\/code>, <code>T_4<\/code>, <code>T_5<\/code>, <code>T_6<\/code> as a sequence and predict sequence of <code>DE_1<\/code>, <code>DE_2<\/code>, <code>DE_3<\/code>, <code>DE_4<\/code>.<\/p>\n\n<p>I am using keras for Sequence to sequence time series using LSTM. I tried to take <code>T_1<\/code> to <code>T_6<\/code> as input dataframe <code>'X'<\/code> and <code>DE_1<\/code> to <code>DE_4<\/code> as output dataframe <code>'y'<\/code>. I reshaped it using <code>X = np.array(X)<\/code> <code>y = np.array(y)<\/code> and then <code>X = X.reshape(4,6,1)<\/code> and <code>y = y.reshape(4,4,1)<\/code> to feed to <code>batch_input_shape()<\/code> but it does not work. <\/p>\n\n<p>How to get data in proper shape to feed to LSTM layer?<\/p>\n\n<pre><code>                      T_1     T_2     T_3     T_4     T_5     T_6    DE_1    DE_2    DE_3    DE_4\nSETTLEMENTDATE                                          \n2018-11-01 14:30:00 1645.82 1623.23 1619.09 1581.94 1538.20 1543.48 1624.23 1722.85 1773.77 1807.04\n2018-11-01 15:00:00 1628.60 1645.82 1623.23 1619.09 1581.94 1538.20 1722.85 1773.77 1807.04 1873.53\n2018-11-01 15:30:00 1624.23 1628.60 1645.82 1623.23 1619.09 1581.94 1773.77 1807.04 1873.53 1889.06\n2018-11-01 16:00:00 1722.85 1624.23 1628.60 1645.82 1623.23 1619.09 1807.04 1873.53 1889.06 1924.57\n<\/code><\/pre>\n","answers":[{"AnswerId":"53229165","CreationDate":"2018-11-09T15:57:52.240","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>LSTM accepts two arguments: <code>input_shape<\/code> and <code>batch_input_shape<\/code>. The difference is in convention that <code>input_shape<\/code> does not contain the batch size, while <code>batch_input_shape<\/code> is the full input shape including the batch size.<\/p>\n\n<p>LSTM layer is a recurrent layer, hence it expects a 3-dimensional input <code>(batch_size, timesteps, input_dim)<\/code>. That's why the correct specification is <code>input_shape=(6, 1)<\/code> or <code>batch_input_shape=(BATCH_SIZE, 6, 1)<\/code>, where <code>BATCH_SIZE<\/code> is the size of your batch.<\/p>\n\n<p>I hope it helps :)<\/p>\n"}]}
+{"QuestionId":53221372,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T07:22:39.757","AcceptedAnswerId":53223212.0,"OwnerUserId":3861775.0,"Title":"Tensorflow Apply softmax function only to filters","Body":"<p>Let <code>x<\/code> be an tensor of filters of size <code>(n, w, h, c)<\/code>.<\/p>\n\n<p>I want to apply the <code>tf.nn.softmax()<\/code> function to every filter in that tensor. How can I do that?<\/p>\n\n<p>I tried the following but got an error:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nn, c = 2, 2 \nh, w = 2, 2\n\nx = tf.ones([n, h, w, c])\ny = tf.nn.softmax(x, axis=[1,2])\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(x)\n    print(\"x\", sess.run(x))\n    print(\"\\n\")\n    print(y)\n    print(\"y\", sess.run(y))\n<\/code><\/pre>\n\n<p>After the operation, I would expect every filter to be <\/p>\n\n<pre><code>0.25 0.25\n0.25 0.25\n<\/code><\/pre>\n","answers":[{"AnswerId":"53223212","CreationDate":"2018-11-09T09:40:32.523","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>\nHere is my solution:<\/p>\n\n<ol>\n<li><p>Reshape <code>x<\/code> as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_r = tf.reshape(x, [n, -1, c])\n<\/code><\/pre><\/li>\n<li><p>Apply softmax to filter dimension:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y_r = tf.nn.softmax(x_r, axis=1)\n<\/code><\/pre><\/li>\n<li><p>Recover the original shape:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y = tf.reshape(y_r, [n, h, w, c])\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53221523,"AnswerCount":2,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-11-09T07:36:13.180","AcceptedAnswerId":null,"OwnerUserId":4246716.0,"Title":"How to check if tensorflow is using all available GPU's","Body":"<p>I am learning to use Tensorflow for object detection.  To speed up the training process, I have taken a AWS g3.16xlarge instance which has 4 GPUs.  I am using the following code to run training process:<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0,1,2,3\n python object_detection\/train.py --logtostderr --pipeline_config_path=\/home\/ubuntu\/builder\/rcnn.config --train_dir=\/home\/ubuntu\/builder\/experiments\/training\/\n<\/code><\/pre>\n\n<p>Inside the rcnn.config - i have set the <code>batch-size = 1<\/code>.  During runtime I get the following output:<\/p>\n\n<p><strong>console output<\/strong><\/p>\n\n<pre><code>2018-11-09 07:25:50.104310: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1045] Device peer to peer matrix\n2018-11-09 07:25:50.104385: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1051] DMA: 0 1 2 3 \n2018-11-09 07:25:50.104395: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1061] 0:   Y N N N \n2018-11-09 07:25:50.104402: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1061] 1:   N Y N N \n2018-11-09 07:25:50.104409: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1061] 2:   N N Y N \n2018-11-09 07:25:50.104416: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1061] 3:   N N N Y \n2018-11-09 07:25:50.104429: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1120] Creating TensorFlow device (\/device:GPU:0) -&gt; (device: 0, name: Tesla M60, pci bus id: 0000:00:1b.0, compute capability: 5.2)\n2018-11-09 07:25:50.104439: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1120] Creating TensorFlow device (\/device:GPU:1) -&gt; (device: 1, name: Tesla M60, pci bus id: 0000:00:1c.0, compute capability: 5.2)\n2018-11-09 07:25:50.104446: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1120] Creating TensorFlow device (\/device:GPU:2) -&gt; (device: 2, name: Tesla M60, pci bus id: 0000:00:1d.0, compute capability: 5.2)\n2018-11-09 07:25:50.104455: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1120] Creating TensorFlow device (\/device:GPU:3) -&gt; (device: 3, name: Tesla M60, pci bus id: 0000:00:1e.0, compute capability: 5.2)\n<\/code><\/pre>\n\n<p>When I run <code>nvidia-smi<\/code>, I get the following output:\n<strong>nvidia-smi output<\/strong><\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 375.26                 Driver Version: 375.26                    |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  Tesla M60           Off  | 0000:00:1B.0     Off |                    0 |\n| N\/A   52C    P0   129W \/ 150W |   7382MiB \/  7612MiB |     92%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  Tesla M60           Off  | 0000:00:1C.0     Off |                    0 |\n| N\/A   33C    P0    38W \/ 150W |   7237MiB \/  7612MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   2  Tesla M60           Off  | 0000:00:1D.0     Off |                    0 |\n| N\/A   40C    P0    38W \/ 150W |   7237MiB \/  7612MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   3  Tesla M60           Off  | 0000:00:1E.0     Off |                    0 |\n| N\/A   34C    P0    39W \/ 150W |   7237MiB \/  7612MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID  Type  Process name                               Usage      |\n|=============================================================================|\n|    0     97860    C   python                                        7378MiB |\n|    1     97860    C   python                                        7233MiB |\n|    2     97860    C   python                                        7233MiB |\n|    3     97860    C   python                                        7233MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>and <code>**nvidia-smi dmon**<\/code> provides the following output:<\/p>\n\n<pre><code># gpu   pwr  temp    sm   mem   enc   dec  mclk  pclk\n# Idx     W     C     %     %     %     %   MHz   MHz\n    0   158    69    90    69     0     0  2505  1177\n    1    38    36     0     0     0     0  2505   556\n    2    38    45     0     0     0     0  2505   556\n    3    39    37     0     0     0     0  2505   556\n<\/code><\/pre>\n\n<p>I am confused with each of the output.  While I read the console output as the program is recognizing the availability of 4 different gpus, in the nvidia-smi output the volatile GPU-Util percentage is shown only for the first GPU and for the rest it is zero.  However the same table prints memory usage for all the 4 gpu's at the bottom.  And the nvidia-smi dmon prints the sm values only for first gpu and for the others it is zero.  From this <a href=\"https:\/\/www.microway.com\/hpc-tech-tips\/nvidia-smi_control-your-gpus\/\" rel=\"nofollow noreferrer\">blog<\/a> I understand the zero in <code>dmon<\/code> indicates that GPU is free.  <\/p>\n\n<p>What I want to understand is, does the train.py utilizes all the 4 GPU's that I have in my instance. If it is not utilizing all the GPU's how do I ensure the <code>object_detection\/train.py<\/code> of tensorflow is optimized for all the GPU's.  <\/p>\n","answers":[{"AnswerId":"53221637","CreationDate":"2018-11-09T07:45:10.950","ParentId":null,"OwnerUserId":"5435253","Title":null,"Body":"<p>Check if it's returning list of all GPUs.<\/p>\n\n<pre><code>tf.test.gpu_device_name()\n<\/code><\/pre>\n\n<p>Returns the name of a GPU device if available or the empty string.<\/p>\n\n<p>then you can do something like this  to use all the available GPUs.<\/p>\n\n<pre><code># Creates a graph.\nc = []\nfor d in ['\/device:GPU:2', '\/device:GPU:3']:\n  with tf.device(d):\n    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3])\n    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2])\n    c.append(tf.matmul(a, b))\nwith tf.device('\/cpu:0'):\n  sum = tf.add_n(c)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint(sess.run(sum))\n<\/code><\/pre>\n\n<p>You see below output:<\/p>\n\n<pre><code>Device mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: Tesla K20m, pci bus\nid: 0000:02:00.0\n\/job:localhost\/replica:0\/task:0\/device:GPU:1 -&gt; device: 1, name: Tesla K20m, pci bus\nid: 0000:03:00.0\n\/job:localhost\/replica:0\/task:0\/device:GPU:2 -&gt; device: 2, name: Tesla K20m, pci bus\nid: 0000:83:00.0\n\/job:localhost\/replica:0\/task:0\/device:GPU:3 -&gt; device: 3, name: Tesla K20m, pci bus\nid: 0000:84:00.0\nConst_3: \/job:localhost\/replica:0\/task:0\/device:GPU:3\nConst_2: \/job:localhost\/replica:0\/task:0\/device:GPU:3\nMatMul_1: \/job:localhost\/replica:0\/task:0\/device:GPU:3\nConst_1: \/job:localhost\/replica:0\/task:0\/device:GPU:2\nConst: \/job:localhost\/replica:0\/task:0\/device:GPU:2\nMatMul: \/job:localhost\/replica:0\/task:0\/device:GPU:2\nAddN: \/job:localhost\/replica:0\/task:0\/cpu:0\n[[  44.   56.]\n [  98.  128.]]\n<\/code><\/pre>\n"},{"AnswerId":"57018143","CreationDate":"2019-07-13T10:07:42.457","ParentId":null,"OwnerUserId":"4791116","Title":null,"Body":"<p>Python code to check if GPU is found and available for using with <code>tensorflow<\/code>:<\/p>\n\n<pre><code>## Libraries import\nimport tensorflow as tf\n\n## Test GPU\ndevice_name = tf.test.gpu_device_name()\nif device_name != '\/device:GPU:0':\n  raise SystemError('GPU device not found')\nprint('Found GPU at: {}'.format(device_name))\nprint('')\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\n<\/code><\/pre>\n"}]}
+{"QuestionId":53221959,"AnswerCount":1,"Tags":"<python><keras><keras-layer>","CreationDate":"2018-11-09T08:11:28.467","AcceptedAnswerId":53222206.0,"OwnerUserId":1563247.0,"Title":"Keras model order of inputs","Body":"<p>I had 3 separate models that were basically in the structure of user-item.  I wanted to merge these and then run a a few layers post merege.  However, when it came time for the inputs, I first got an error.  I had assumed that I would need inputs in the structure of <code>[item1, user, item2, users, item3 users]<\/code>, which matched the inputs from my 3 initial separate models.  However, on doing so, it basically said, 'you can't repeat inputs'.  However, something feels wrong with the <code>[item1, item2, item3, user]<\/code> structure to me, even though it runs with decent results.  Should I simply duplicate the users to make a user1, user2, user3 that are identical?<\/p>\n\n<p>Code below:<\/p>\n\n<pre><code>#Making the vctors for the primary categories\nitem1_input = Input(shape=[1])\nitem2_input = Input(shape=[1])\nitem3_input = Input(shape=[1])\nuser_input = Input(shape=[1])\n\nitem1_vec = Flatten()(Embedding(nb_item1s + 1, 32)(item1_input))\nitem1_vec = Dropout(0.5)(item1_vec)\n\n\nitem2_vec = Flatten()(Embedding(nb_breweries + 1, 32)(item2_input))\nitem2_vec = Dropout(0.5)(item2_vec)\n\n\nitem3_vec = Flatten()(Embedding(nb_item3s + 1, 32)(item3_input))\nitem3_vec = Dropout(0.5)(item3_vec)\n\nuser_vec = Flatten()(Embedding(nb_users + 1, 32)(user_input))\nuser_vec = Dropout(0.5)(user_vec)\n\n# Next, we join them all together and put them\n# through a pretty standard deep learning architecture\nitem1_input_vecs = add([item1_vec, user_vec])\nitem1_nn = Dropout(0.5)(Dense(128, activation='relu')(item1_input_vecs))\nitem1_nn = BatchNormalization()(item1_nn)\nitem1_nn = Dropout(0.5)(Dense(128, activation='relu')(item1_nn))\n#item1_nn = BatchNormalization()(item1_nn)\n#item1_nn = Dense(128, activation='relu')(item1_nn)\nitem1_result = Dense(9, activation='softmax')(item1_nn)\n\nitem2_input_vecs = add([item2_vec, user_vec])\nitem2_nn = Dropout(0.5)(Dense(128, activation='relu')(item2_input_vecs))\nitem2_nn = BatchNormalization()(item2_nn)\nitem2_nn = Dropout(0.5)(Dense(128, activation='relu')(item2_nn))\n#item2_nn = BatchNormalization()(item2_nn)\n#item2_nn = Dense(128, activation='relu')(item2_nn)\nitem2_result = Dense(9, activation='softmax')(item2_nn)\n\nitem3_input_vecs = add([item3_vec, user_vec])\nitem3_nn = Dropout(0.5)(Dense(128, activation='relu')(item3_input_vecs))\nitem3_nn = BatchNormalization()(item3_nn)\nitem3_nn = Dropout(0.5)(Dense(128, activation='relu')(item3_nn))\n#item3_nn = BatchNormalization()(item3_nn)\n#item3_nn = Dense(128, activation='relu')(item3_nn)\nitem3_result = Dense(9, activation='softmax')(item3_nn)\n\nresult_vecs = Concatenate()([item1_result, item2_result, item3_result])\nresult_vecs = Dropout(0.5)(result_vecs)\n\nfinal_nn = Dense(128,activation='relu')(result_vecs)\nfinal_nn = Dropout(0.5)(final_nn)\nfinal_result = Dense(9, activation='softmax')(final_nn)\n\nfinalmodel = Model(inputs=[item1_input, item2_input, item3_input, user_input], outputs=final_result)\nfinalmodel.compile(optimizer='adam', loss = 'categorical_crossentropy')\n\n\nfinalhistory = finalmodel.fit([a_item1id, a_item2id, a_item3id, a_userid], a_y, \n                     epochs=20, \n                     validation_data=([b_item1id, b_item2id, b_item3id, b_userid], b_y), verbose=1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53222206","CreationDate":"2018-11-09T08:30:54.803","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p>Your code is correct as it is. Your model has one input (<code>user_input<\/code>), then inside the model you are using this <code>Input()<\/code> in three different layers. <\/p>\n"}]}
+{"QuestionId":53222326,"AnswerCount":0,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-11-09T08:41:08.847","AcceptedAnswerId":null,"OwnerUserId":264052.0,"Title":"Where is the `torch.empty` defined?","Body":"<p>(I am not a Python veteran so pardon me if this is a naive question.)<\/p>\n\n<p>Following the <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/tensor_tutorial.html\" rel=\"nofollow noreferrer\">PyTorch tutorial<\/a>, below code works fine. <\/p>\n\n<pre><code>import torch\nx= torch.empty(5,3) # &lt;========= HERE\nprint(x)\n<\/code><\/pre>\n\n<p>But the Visual Studio Code complains that:<\/p>\n\n<blockquote>\n  <p>E1101:Module 'torch' has no 'empty' member<\/p>\n<\/blockquote>\n\n<p>Eclipse PyDev has a similar warning.<\/p>\n\n<p>So I use below code to check what <code>torch<\/code> offers:<\/p>\n\n<pre><code>for (k, v) in torch.__dict__.items():\n    print(k, v)\n<\/code><\/pre>\n\n<p>And I find that it is a built-in:<\/p>\n\n<pre><code>empty &lt;built-in method empty of type object at 0x00007FFA380E0A80&gt;\n<\/code><\/pre>\n\n<p>Since it is a method, I searched all the <code>*.py<\/code> file in the torch package folder for something like <code>def empty<\/code> and <code>empty<\/code>, no luck.<\/p>\n\n<p>So where is this <code>empty<\/code> method defined? Some pyd\/DLL?<\/p>\n\n<p>How is it imported into the namespace?<\/p>\n\n<p>Or put it another way, how to create a built-in function for my own package?<\/p>\n\n<h2>ADD 1 -- 10:38 AM 11\/12\/2018<\/h2>\n\n<p>Extending Python Built-ins with C\/C++<\/p>\n\n<p><a href=\"https:\/\/docs.python.org\/3\/extending\/extending.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/extending\/extending.html<\/a><\/p>\n\n<p>I guess it's the <code>torch::empty<\/code> method. I have found <a href=\"https:\/\/github.com\/pytorch\/pytorch\/search?l=C%2B%2B&amp;p=1&amp;q=torch%3A%3Aempty\" rel=\"nofollow noreferrer\">a lot of matches<\/a> in the github PyTorch codebase. But I haven't nailed the definition yet.<\/p>\n","answers":[]}
+{"QuestionId":53222648,"AnswerCount":1,"Tags":"<deep-learning><pytorch><cudnn>","CreationDate":"2018-11-09T09:03:57.063","AcceptedAnswerId":null,"OwnerUserId":4642437.0,"Title":"CUDNN_STATUS_MAPPING_ERROR when training with pose2body","Body":"<p>I'm trying to train <a href=\"https:\/\/github.com\/NVIDIA\/vid2vid\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/vid2vid<\/a>. I'm...<\/p>\n\n<ul>\n<li><p>...executing with pretty much the vanilla parametrization shown in the readme, I had to change the number of GPUs though and increased the number of threads for reading the dataset. Command:<\/p>\n\n<p><code>python train.py \\\n--name pose2body_256p  \\\n--dataroot datasets\/pose  \\\n--dataset_mode pose  \\\n--input_nc 6  \\\n--num_D 2  \\ \n--resize_or_crop ScaleHeight_and_scaledCrop  \\\n--loadSize 384  \\\n--fineSize 256  \\\n--gpu_ids 0,1  \\\n--batchSize 1  \\\n--max_frames_per_gpu 3  \\\n--no_first_img  \\\n--n_frames_total 12  \\\n--max_t_step 4  \\\n--nThreads 6<\/code> <\/p><\/li>\n<li><p>...training on the supplied example datasets.<\/p><\/li>\n<li>...running a docker container built with the scripts in <code>vid2vid\/docker<\/code>, e. g. with CUDA 9.0 and CUDNN 7.<\/li>\n<li>...using two NVIDIA V100 GPUs.<\/li>\n<\/ul>\n\n<p>Whenever I start training the script crashes after a couple of minutes with the message <strong><code>RuntimeError: CUDNN_STATUS_MAPPING_ERROR<\/code><\/strong>. Full error message:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 329, in &lt;module&gt;\n    train()\n  File \"train.py\", line 104, in train\n    fake_B, fake_B_raw, flow, weight, real_A, real_Bp, fake_B_last = modelG(input_A, input_B, inst_A, fake_B_last)            \n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 491, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/parallel\/data_parallel.py\", line 114, in forward\n    outputs = self.parallel_apply(replicas, inputs, kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/parallel\/data_parallel.py\", line 124, in parallel_apply\n    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 65, in parallel_apply\n    raise output\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 41, in _worker\n    output = module(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 491, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/vid2vid\/models\/vid2vid_model_G.py\", line 130, in forward\n    fake_B, fake_B_raw, flow, weight = self.generate_frame_train(netG, real_A_all, fake_B_prev, start_gpu, is_first_frame)        \n  File \"\/vid2vid\/models\/vid2vid_model_G.py\", line 175, in generate_frame_train\n    fake_B_feat, flow_feat, fake_B_fg_feat, use_raw_only)\n  File \"\/vid2vid\/models\/networks.py\", line 171, in forward\n    downsample = self.model_down_seg(input) + self.model_down_img(img_prev)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 491, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/container.py\", line 91, in forward\n    input = module(input)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 491, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/conv.py\", line 301, in forward\n    self.padding, self.dilation, self.groups)\nRuntimeError: CUDNN_STATUS_MAPPING_ERROR\n<\/code><\/pre>\n\n<p>From reading the issues in the vid2vid using two V100 should work with this setup. The error also occurs if CUDA 8\/CUDNN 6 are used. I checked the flags but haven't found any indication of further necessary changes to the arguments supplied to <code>train.py<\/code>.<\/p>\n\n<p>Any ideas on how to solve (or work around) this?<\/p>\n","answers":[{"AnswerId":"53234907","CreationDate":"2018-11-10T00:20:29.473","ParentId":null,"OwnerUserId":"4642437","Title":null,"Body":"<p>In case anybody deals the same issue: Training on P100 cards worked. Seems like the V100 architecture clashes with version of pytorch used in the supplied Dockerfile at some point. Not quite a solution, but a workaround.<\/p>\n"}]}
+{"QuestionId":53222664,"AnswerCount":1,"Tags":"<python><theano><bayesian><normal-distribution><pymc3>","CreationDate":"2018-11-09T09:05:01.233","AcceptedAnswerId":null,"OwnerUserId":2124204.0,"Title":"PyMC3 passing stochastic covariance matrix to pm.MvNormal()","Body":"<p>I've tried to fit a simple 2D gaussian model to observed data by using PyMC3.<\/p>\n\n<pre><code>import numpy as np\nimport pymc3 as pm\n\nn = 10000;\nnp.random.seed(0)\nX = np.random.multivariate_normal([0,0], [[1,0],[0,1]], n);\n\nwith pm.Model() as model:\n    # PRIORS\n    mu = [pm.Uniform('mux', lower=-1, upper=1), \n          pm.Uniform('muy', lower=-1, upper=1)]\n    cov = np.array([[pm.Uniform('a11', lower=0.1, upper=2), 0],\n                    [0, pm.Uniform('a22', lower=0.1, upper=2)]])\n\n    # LIKELIHOOD\n    likelihood = pm.MvNormal('likelihood', mu=mu, cov=cov, observed=X)\n\nwith model:\n    trace = pm.sample(draws=1000, chains=2, tune=1000)\n<\/code><\/pre>\n\n<p>while I can do this in 1D by passing the <code>sd<\/code> to <code>pm.Normal<\/code>, I have some trouble in passing the covariance matrix to <code>pm.MvNormal<\/code>.<\/p>\n\n<p>Where am I going wrong?<\/p>\n","answers":[{"AnswerId":"53308406","CreationDate":"2018-11-14T20:40:27.360","ParentId":null,"OwnerUserId":"570918","Title":null,"Body":"<p><strong>PyMC3 distribution objects are not simple numeric objects or numpy arrays<\/strong>. Instead, they are nodes in a theano computation graph and often require operations from either <a href=\"https:\/\/docs.pymc.io\/api\/math.html\" rel=\"nofollow noreferrer\"><code>pymc3.math<\/code><\/a> or <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html\" rel=\"nofollow noreferrer\"><code>theano.tensor<\/code><\/a> to manipulate them. Moreover, placing PyMC3 objects in numpy arrays is unnecessary since they already are multidimensional.<\/p>\n\n<h2>Original Model<\/h2>\n\n<p>Keeping with the intent of your code, a working version would go something like<\/p>\n\n<pre><code>import numpy as np\nimport pymc3 as pm\nimport theano.tensor as tt\n\nN = 10000\nnp.random.seed(0)\nX = np.random.multivariate_normal(np.zeros(2), np.eye(2), size=N)\n\nwith pm.Model() as model:\n    # use `shape` argument to define tensor dimensions\n    mu = pm.Uniform('mu', lower=-1, upper=1, shape=2)\n\n    # diagonal values on covariance matrix\n    a = pm.Uniform('a', lower=0.1, upper=2, shape=2)\n\n    # convert vector to a 2x2 matrix with `a` on the diagonal\n    cov = tt.diag(a)\n\n    likelihood = pm.MvNormal('likelihood', mu=mu, cov=cov, observed=X)\n<\/code><\/pre>\n\n<h2>Alternative Model<\/h2>\n\n<p>I assume the example you provided is just a toy to communicate the problem. But just in case, I'll mention that the main advantage of using a multivariate normal (modeling covariance between parameters) is lost when restricting the covariance matrix to be diagonal. Furthermore, the theory of priors for covariance matrices is well-developed, so it's worth one's time considering existing solutions. In particular, there is <a href=\"https:\/\/docs.pymc.io\/notebooks\/LKJ.html\" rel=\"nofollow noreferrer\">a PyMC3 example using the LKJ prior for covariance matrices<\/a>.<\/p>\n\n<p>Here's a simple application of that example in this context:<\/p>\n\n<pre><code>with pm.Model() as model_lkj:\n    # use `shape` argument to define tensor dimensions\n    mu = pm.Uniform('mu', lower=-1, upper=1, shape=2)\n\n    # LKJ prior for covariance matrix (see example)\n    packed_L = pm.LKJCholeskyCov('packed_L', n=2,\n                                 eta=2., sd_dist=pm.HalfCauchy.dist(2.5))\n    # convert to (2,2)\n    L = pm.expand_packed_triangular(2, packed_L)\n\n    likelihood = pm.MvNormal('likelihood', mu=mu, chol=L, observed=X)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53223019,"AnswerCount":1,"Tags":"<numpy><tensorflow>","CreationDate":"2018-11-09T09:29:15.177","AcceptedAnswerId":53223532.0,"OwnerUserId":10304086.0,"Title":"How do I turn a numpy array into a tensor in \"Tensorflow\"?","Body":"<p>I tried to test some learning networks after I completed training with a tensorflow.<\/p>\n\n<p>But my test image is [512 512 1] data of channel 1 in 512 horizontal and 512 vertical pixels.<\/p>\n\n<p>I changed the image data to a numpy array.<\/p>\n\n<p>The tensor network should be [? 512 512 1] It looks like this.<\/p>\n\n<p>How do I convert a numpy array to a tensor? ([512 512 1] -> [? 512 512 1])<\/p>\n","answers":[{"AnswerId":"53223532","CreationDate":"2018-11-09T09:59:22.663","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>You just have to append one dimension<\/p>\n\n<pre><code>arr = your_image # [512, 512, 1]\nnew_arr = np.expand_dims(arr, 0)\n\ntensor = tf.convert_to_tensor(new_arr)\n<\/code><\/pre>\n\n<p>Now you can use feed dict or something else.<\/p>\n"}]}
+{"QuestionId":53223280,"AnswerCount":0,"Tags":"<tensorflow><predict>","CreationDate":"2018-11-09T09:44:22.343","AcceptedAnswerId":null,"OwnerUserId":1808715.0,"Title":"Predicting using adanet for binary classification problem","Body":"<p>I followed the tutorial of adanet:\n<a href=\"https:\/\/github.com\/tensorflow\/adanet\/tree\/master\/adanet\/examples\/tutorials\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/adanet\/tree\/master\/adanet\/examples\/tutorials<\/a>\nand was able to apply adanet to my own binary classification problem.<\/p>\n\n<p>But how can I predict using the train model? I have a very little knowledge of TensorFlow. Any helps would be really appreciated<\/p>\n","answers":[]}
+{"QuestionId":53223398,"AnswerCount":1,"Tags":"<python><opencv><image-processing><keras><conv-neural-network>","CreationDate":"2018-11-09T09:51:19.623","AcceptedAnswerId":53227955.0,"OwnerUserId":10551333.0,"Title":"How to use convolutional neural network on binary image using Keras?","Body":"<p>I am trying to train a cnn model for ocr using keras. I preprocessed the images by converting to grayscale, removing noise and then converting it to binary, as binary images work better in ocr. But the problem I am getting is that binary image has 2 dimensions and no channel dimension and conv2d in keras(well any conv layer in general) require 3 dimensions. So what should I do to add a dimension but keep image binary? I am using cv2 for image processing so please tell solutions using that preferably. Also tell me whether I am right that using binary image dataset is better for ocr.<\/p>\n","answers":[{"AnswerId":"53227955","CreationDate":"2018-11-09T14:45:49.303","ParentId":null,"OwnerUserId":"10551333","Title":null,"Body":"<p>I got my solution. I used numpy function numpy.expand_dims() to add empty dimension. so it became (width,height,1). Here is what I did:-<\/p>\n\n<pre><code>img = np.expand_dims(img,axis=2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53223457,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T09:54:31.323","AcceptedAnswerId":null,"OwnerUserId":10627756.0,"Title":"getting attentionwrapper issue in spellchecker program","Body":"<p><a href=\"https:\/\/github.com\/Currie32\/Spell-Checker\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Currie32\/Spell-Checker<\/a>.<\/p>\n\n<p>In the above link code I'm getting error that DynamicAttentionWrapper is not defined. I'm using TensorFlow version as 1.2. I couldn't able to over come with this error. Please help me with this.<\/p>\n","answers":[{"AnswerId":"53223616","CreationDate":"2018-11-09T10:05:27.323","ParentId":null,"OwnerUserId":"8382118","Title":null,"Body":"<p>There is issue with <code>DynamicAttentionWrapper<\/code> in your version of <code>tensorflow<\/code>.<\/p>\n\n<p>Try changing <code>DynamicAttentionWrapper<\/code> to <code>AttentionWrapper<\/code> or downgrade to <code>tensorflow 1.1<\/code>.<\/p>\n\n<p>For your version of <code>tensorflow<\/code>, try these changes for <code>initial_state, inference_logits and training_logits<\/code>:<\/p>\n\n<pre><code>initial_state = dec_cell.zero_state(batch_size=batch_size,dtype=tf.float32).clone(cell_state=enc_state)\n\ninference_logits, _ ,_ = tf.contrib.seq2seq.dynamic_decode(inference_decoder,\noutput_time_major=False,\nimpute_finished=True,\nmaximum_iterations=max_target_length)\n\ntraining_logits, _ ,_ = tf.contrib.seq2seq.dynamic_decode(training_decoder,\noutput_time_major=False,\nimpute_finished=True,\nmaximum_iterations=max_target_length)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53223499,"AnswerCount":2,"Tags":"<function><tensorflow><keras><loss>","CreationDate":"2018-11-09T09:57:17.527","AcceptedAnswerId":null,"OwnerUserId":1958843.0,"Title":"Keras Functional API and loss function with multiple inputs","Body":"<p>I am trying to use a custom Keras loss function that apart from the usual  signature <code>(y_true, y_pred)<\/code> takes another parameter <code>sigma<\/code> (which is also produced by the last layer of the network).\nThe training works fine, but then I am not sure how to perform forward propagation and return <code>sigma<\/code> (while <code>mu<\/code>is the output of the <code>model.predict<\/code> method).\nThis is the code I am using, which features a custom layer <code>GaussianLayer<\/code> that returns the list <code>[mu, sigma]<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\nfrom keras.layers import Input, Dense, Layer, Dropout\nfrom keras.models import Model\nfrom keras.initializers import glorot_normal\nimport numpy as np\n\ndef custom_loss(sigma):\n    def gaussian_loss(y_true, y_pred):\n        return tf.reduce_mean(0.5*tf.log(sigma) + 0.5*tf.div(tf.square(y_true - y_pred), sigma)) + 10\n    return gaussian_loss\n\nclass GaussianLayer(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(GaussianLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.kernel_1 = self.add_weight(name='kernel_1', \n                                      shape=(30, self.output_dim),\n                                      initializer=glorot_normal(),\n                                      trainable=True)\n        self.kernel_2 = self.add_weight(name='kernel_2', \n                                      shape=(30, self.output_dim),\n                                      initializer=glorot_normal(),\n                                      trainable=True)\n        self.bias_1 = self.add_weight(name='bias_1',\n                                    shape=(self.output_dim, ),\n                                    initializer=glorot_normal(),\n                                    trainable=True)\n        self.bias_2 = self.add_weight(name='bias_2',\n                                    shape=(self.output_dim, ),\n                                    initializer=glorot_normal(),\n                                    trainable=True)\n        super(GaussianLayer, self).build(input_shape) \n\n    def call(self, x):\n        output_mu  = K.dot(x, self.kernel_1) + self.bias_1\n        output_sig = K.dot(x, self.kernel_2) + self.bias_2\n        output_sig_pos = K.log(1 + K.exp(output_sig)) + 1e-06  \n        return [output_mu, output_sig_pos]\n\n    def compute_output_shape(self, input_shape):\n        return [(input_shape[0], self.output_dim), (input_shape[0], self.output_dim)]\n\n# This returns a tensor\ninputs = Input(shape=(1,))\nx = Dense(30, activation='relu')(inputs)\nx = Dropout(0.3)(x)\nx = Dense(30, activation='relu')(x)\nx = Dense(40, activation='relu')(x)\nx = Dropout(0.3)(x)\nx = Dense(30, activation='relu')(x)\nmu, sigma = GaussianLayer(1)(x)\n\nmodel = Model(inputs, mu)\nmodel.compile(loss=custom_loss(sigma), optimizer='adam')\nmodel.fit(train_x, train_y, epochs=150)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53223615","CreationDate":"2018-11-09T10:05:25.793","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since your model returns two tensors as output, you also need to pass <strong>a list of two arrays<\/strong> as the output when calling <code>fit()<\/code> method. That's essentially what the error is trying to convey:<\/p>\n\n<blockquote>\n  <p>Error when checking <strong>model target<\/strong>:<\/p>\n<\/blockquote>\n\n<p>So the error is in targets (i.e. labels). What is wrong?<\/p>\n\n<blockquote>\n  <p>the list of Numpy arrays that you are passing to your model <strong>is not the size the model expected<\/strong>. <strong>Expected<\/strong> to see <strong>2 array(s)<\/strong>, but instead <strong>got<\/strong> the following list of <strong>1 arrays<\/strong>:<\/p>\n<\/blockquote>\n"},{"AnswerId":"53247542","CreationDate":"2018-11-11T09:51:56.273","ParentId":null,"OwnerUserId":"1958843","Title":null,"Body":"<p>I may have found the answer among <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">Keras FAQs<\/a>.\nI found out that it is possible to retrieve intermediate steps' output using the code snippet below:  <\/p>\n\n<pre><code>layer_name = 'main_output'\nintermediate_layer_model = Model(inputs=model.input,\n                                 outputs=model.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(train_x[0])\nintermediate_output\n<\/code><\/pre>\n\n<p>In this case <code>intermediate_output<\/code> is a list of two values <code>[mu, sigma]<\/code> (just needed to name the output layer <code>main_output<\/code> and retrieve it later)<\/p>\n"}]}
+{"QuestionId":53224172,"AnswerCount":1,"Tags":"<python><amazon-web-services><keras><amazon-sagemaker>","CreationDate":"2018-11-09T10:42:38.833","AcceptedAnswerId":53347423.0,"OwnerUserId":2032998.0,"Title":"How can I update the keras.json file on Sagemaker","Body":"<p>I need to make some amendments to the keras.json file for a sagemaker notebook instance but I am unsure where it is located. Any help would be greatly appreciated.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53347423","CreationDate":"2018-11-17T01:36:42.423","ParentId":null,"OwnerUserId":"4570570","Title":null,"Body":"<p>Thank you for using Amazon SageMaker! <\/p>\n\n<p>You can find keras.json file located in \/home\/ec2-user\/.keras . <\/p>\n\n<p>Thanks,<br>\nNeelam <\/p>\n"}]}
+{"QuestionId":53224685,"AnswerCount":1,"Tags":"<python><opencv><pytorch><onnx>","CreationDate":"2018-11-09T11:15:35.967","AcceptedAnswerId":null,"OwnerUserId":8288264.0,"Title":"Trouble using openCV to load a net from ONNX (python\/pytorch)","Body":"<p>I'm trying to load a trained .onnx model (from a neural-style-transfer algorithm) into cv2. I've seen that there is a <\/p>\n\n<pre><code>cv.dnn.readNetFromONNX() \n<\/code><\/pre>\n\n<p>function, but there is no such function in cv2. I can't seem to import or load opencv as cv, and as such can't seem to load my model in cv2. Does anyone know a solution?<\/p>\n\n<p>I've basically trained a model with \n<a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/fast_neural_style\/neural_style\/neural_style.py#L122-L150\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/examples\/blob\/master\/fast_neural_style\/neural_style\/neural_style.py#L122-L150<\/a>\nthis script, and made an export of an onnx model by adding<\/p>\n\n<pre><code>torch.onnx.export(style_model, dummy_input, \"chipsoft_mod.onnx\", verbose=True)\n<\/code><\/pre>\n\n<p>Now I want to run the trained model trough the cv2 reader, but I fail spectacularly.<\/p>\n","answers":[{"AnswerId":"56583296","CreationDate":"2019-06-13T14:56:24.407","ParentId":null,"OwnerUserId":"10974829","Title":null,"Body":"<p>Update your opencv to a newer version. It should help.<\/p>\n\n<pre><code>pip install opencv-python==4.1.0.25\n<\/code><\/pre>\n"}]}
+{"QuestionId":53225442,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-09T12:07:28.943","AcceptedAnswerId":null,"OwnerUserId":9290285.0,"Title":"tensorflow estimator: filter initialization values of the model","Body":"<p>i am using the tensorflow estimator api and have successfully trained my model. However, i want to see how my filters are initialized at the beginning of the training. <\/p>\n\n<p>At the moment i can force the model to give me the kernel values from a certain training stage with: <\/p>\n\n<pre><code>print(tf.get_default_graph().get_tensor_by_name('resnet_model\/conv2d\/kernel:0'))\n<\/code><\/pre>\n\n<p>when i predict with the estimator<\/p>\n\n<pre><code>estimator.predict(input_fn=input_fn,checkpoint_path = checkpointfile)\n<\/code><\/pre>\n\n<p>But i want also to get the initialization values before the network has anything learned. So my idea was not to give any checkpoint to the model. This results in an error:<\/p>\n\n<pre><code>ValueError: Could not find trained model in model_dir\n<\/code><\/pre>\n\n<p>So how can i get the initialization values of my filter when using the estimator api?<\/p>\n","answers":[]}
+{"QuestionId":53225804,"AnswerCount":0,"Tags":"<python><tensorflow><record>","CreationDate":"2018-11-09T12:33:54.037","AcceptedAnswerId":null,"OwnerUserId":10596177.0,"Title":"Passing tensorflow records one by one","Body":"<p>The tensorflow record size is very huge 100 GB storing 1000 mini-batches.\nBut I will need 10000 mini-batches.\nIf a tensorflow record is created while the program is running replacing the old tensorflow record, how can I place the newly created record in the queue?<\/p>\n\n<p>Do I need to have all the tensorflow records before running the tensorflow code or can I pass one by one, creating each when the last one is over.<\/p>\n\n<pre><code># Writing to a tf.record.\nresult_tf_file = file_path_prefix + '.tfrecords'\nwriter = tf.python_io.TFRecordWriter(result_tf_file)\n# Create feature\nfeatures = tf.train.Features(feature=d_feature)\nexample = tf.train.Example(features=features)\nserialized = example.SerializeToString()\nwriter.write(serialized)\n\n\n# For Initializing the tf.record.\ndataset = tf.data.TFRecordDataset([file_name_tf])\ndataset = dataset.map(m)\ndataset = dataset.batch(1)\ndataset = dataset.shuffle(buffer_size=50)\ndataset = dataset.repeat(1)\niterator = dataset.make_initializable_iterator()\n\n# Initializing iterator\nsess.run(iterator.initializer, feed_dict={file_name_tf: result_tf_file})\n<\/code><\/pre>\n\n<p>Can I initialize like\nsess.run(iterator.initializer, serialized)?<\/p>\n","answers":[]}
+{"QuestionId":53226038,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-09T12:50:40.993","AcceptedAnswerId":null,"OwnerUserId":7286275.0,"Title":"Adding LSTM layers in Keras produces input error","Body":"<p>First of all apologies if I word this wrong, I'm relatively new to TensorFlow<\/p>\n\n<p>I am designing a model for simple classification of a dataset, each column is an attribute and the final column is the class. I split these and generate a dataframe in the usual way. <\/p>\n\n<p>If I generated a model with dense layers, it works great:<\/p>\n\n<pre><code>def baseline_model():\n# create model\nmodel = Sequential()\nmodel.add(Dense(30, input_dim=len(dataframe.columns)-2, activation='sigmoid'))\nmodel.add(Dense(20,activation='sigmoid'))   \nmodel.add(Dense(unique, activation='softmax'))\n\n# Compile model\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.summary()\nreturn model\n<\/code><\/pre>\n\n<p>If I were to add, say, an LSTM layer to the model:<\/p>\n\n<pre><code>def baseline_model():\n# create model\nmodel = Sequential()\nmodel.add(Dense(30, input_dim=len(dataframe.columns)-2, activation='sigmoid'))\n\n\n#this bit here &gt;\nmodel.add(LSTM(20, return_sequences=True))\nmodel.add(Dense(20,activation='sigmoid'))   \nmodel.add(Dense(unique, activation='softmax'))\n\n# Compile model\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.summary()\nreturn model\n<\/code><\/pre>\n\n<p>I get the following error when I execute the code:<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2\n<\/code><\/pre>\n\n<p>I'm not sure where these variables are coming from although, maybe it's the classes? I have three classes of data ('POSITIVE', 'NEGATIVE', 'NEUTRAL') mapped to sets of well over 2,000 attribute values - they're statistical extractions of timewindowed EEG brainwave data from multiple electrodes classed by emotional state.<\/p>\n\n<p>Note:\nThe <em>'input_dim=len(dataframe.columns)-2'<\/em> produces the number of attributes (inputs), I do this as I'd like the script to work with CSV datasets of different sizes on the fly<\/p>\n\n<p>Also, there are no tabs in my code pasted but it is indented and will compile<\/p>\n\n<p>The full code is pasted here: <a href=\"https:\/\/pastebin.com\/1aXp9uDA\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/1aXp9uDA<\/a> for presentation purposes. Apologies in advance for the terrible practices! This is just an initial project, I do plan on cleaning it all up later on!<\/p>\n","answers":[{"AnswerId":"53226952","CreationDate":"2018-11-09T13:48:26.783","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>In your original code you have an input dimension of 2 which is shaped (batch, feature).  When you add an LSTM, you're telling Keras you want to do the classification given the last N timesteps, hence you need an input that is shaped (batch, timestep, feature).  It's easy to think that an LSTM will look back across all inputs in the batch but unfortunately it doesn't.  You have to manually organize your data to present all the timestep elements together.<\/p>\n\n<p>To split up your data you generally do a sliding window of length N (where N is how many values you wish the LSTM to look back across).  You can slide the window N steps each time, meaning there's no overlap of the data or you can simply slide it one sample, meaning you get multiple copies of your data.  There are numerous blogs on how to do this.  Take a look at this one <a href=\"https:\/\/machinelearningmastery.com\/reshape-input-data-long-short-term-memory-networks-keras\/\" rel=\"nofollow noreferrer\">How to Reshape Input Data for LSTM<\/a>.<\/p>\n\n<p>You also have one other issue.  For your LSTM, you probably want \"return_sequences=False\".  With this equal to True, you would need to have an output \"Y\" value for each element of your timestep.  You probably want your \"Y\" value to represent the next value in your time-series.  Keep this in mind when organizing your data.<\/p>\n\n<p>The above link provides some nice examples or you can search for more in-depth ones.  If you follow those it should be clear how to reorganize things for an LSTM.<\/p>\n"}]}
+{"QuestionId":53226553,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-09T13:23:53.827","AcceptedAnswerId":null,"OwnerUserId":10257622.0,"Title":"Train accuracy is great, but test accuracy is bad","Body":"<p>I used <code>ImageDataGenerator.flow_from_directory()<\/code> to train and test my model.\nWhen it is training, the accuracy was up to 90%.<br>\nBut when I test my model, the accuracy was just the average(50%).<\/p>\n\n<p>This is how I use <code>ImageDataGenerator<\/code>:<\/p>\n\n<pre><code>datagen = ImageDataGenerator(rescale=1. \/ 255)\n\ntrain_data = datagen.flow_from_directory(img_path + '\\\\train', target_size=(224, 224),\n                                     classes=['NORMAL', 'PNEUMONIA'],\n                                     batch_size=10)\n\ntest_data = datagen.flow_from_directory(img_path + '\\\\test', target_size=(224, 224),\n                                    classes=['NORMAL', 'PNEUMONIA'],\n                                    batch_size=10)\n<\/code><\/pre>\n\n<p>This is how I train and test my model:<\/p>\n\n<pre><code>model.fit_generator(train_data, steps_per_epoch=32, epochs=100)\n\ntest_info = model.evaluate_generator(test_data, steps=10)\n\nprint(test_info)\n<\/code><\/pre>\n\n<p>I tried to replace <code>test_data<\/code> with <code>train_data<\/code> when test the model. \nBut the result was identical.<\/p>\n","answers":[{"AnswerId":"53226857","CreationDate":"2018-11-09T13:42:10.390","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>It looks like your model is over fitting on the training data. You can try some methods to prevent it like:<\/p>\n\n<ul>\n<li>early stopping<\/li>\n<li><p>Regularization( dropout, L1\/L2 regularization)<\/p><\/li>\n<li><p>cross-validation <\/p><\/li>\n<li><p>Increasing dataset(if possible)<\/p><\/li>\n<li><p>Reduce architecture complexity<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53226702,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-09T13:34:29.950","AcceptedAnswerId":null,"OwnerUserId":10041301.0,"Title":"Issue with converting tensorflow model to Intel Movidius graph","Body":"<p>Hello I faced with the problem when trying to use Intel Movidius Neural Stick with tensorflow. I have keras model and I convert it to tensorflow model. When I convert it to Movidius graph I got error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"\/usr\/local\/bin\/mvNCCompile\", line 118, in\ncreate_graph(args.network, args.inputnode, args.outputnode, args.outfile, args.nshaves, args.inputsize, args.weights)\nFile \"\/usr\/local\/bin\/mvNCCompile\", line 104, in create_graph\nnet = parse_tensor(args, myriad_config)\nFile \"\/usr\/local\/bin\/ncsdk\/Controllers\/TensorFlowParser.py\", line 290, in parse_tensor\nif have_first_input(strip_tensor_id(node.outputs[0].name)):\nIndexError: list index out of range\n<\/code><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><code>from keras.models import model_from_json\nfrom keras.models import load_model\nfrom keras import backend as K\nimport tensorflow as tf\nimport nn\nimport os\n\nweights_file = \"weights.h5\"\n\nsess = K.get_session()\nK.set_learning_phase(0)\nmodel = nn.alexnet_model() # get keras model\nmodel.load_weights(weights_file)\n\nsaver = tf.train.Saver()\nsaver.save(sess, \".\/TF_Model\/tf_model\") # convert keras to tensorflow model\n\ntf_model_path = \".\/TF_Model\/tf_model\"\n\nfw = tf.summary.FileWriter('logs', sess.graph)\nfw.close()\n\nos.system('mvNCCompile TF_Model\/tf_model.meta -in=conv2d_1_input -on=activation_7\/Softmax') # get Movidius graph\n<\/code><\/pre>\n\n<p>Python version: 2.7\nOS: Ubuntu 16.04\nTensorflow version: 1.12<\/p>\n","answers":[{"AnswerId":"53274009","CreationDate":"2018-11-13T04:49:54.413","ParentId":null,"OwnerUserId":"8743494","Title":null,"Body":"<p>As I know, the ncsdk compiler does not resolve every part of a normal tensorflow network, so you have to modify the network and re-save it in an NCS-friendly way in order to successfully make a Movidius graph. <\/p>\n\n<p>For more information about how to modify tensorflow network, have a look at <a href=\"https:\/\/movidius.github.io\/ncsdk\/tf_compile_guidance.html\" rel=\"nofollow noreferrer\">the official guidance<\/a>.<\/p>\n\n<p>Hope it will help you.<\/p>\n"}]}
+{"QuestionId":53226898,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T13:44:17.797","AcceptedAnswerId":53227406.0,"OwnerUserId":5623899.0,"Title":"TensorFlow custom estimator stuck when calling evaluate after training","Body":"<p>I made a custom estimator (see this <a href=\"https:\/\/colab.research.google.com\/drive\/10VKcE1sZh8F2-HYASxx7OgV8Mf4DQKFy\" rel=\"nofollow noreferrer\">colab<\/a>) in TensorFlow (v1.10) based on their <a href=\"https:\/\/www.tensorflow.org\/guide\/custom_estimators\" rel=\"nofollow noreferrer\">guide<\/a>.<\/p>\n\n<p>I trained the toy model with:<\/p>\n\n<pre><code>tf.estimator.train_and_evaluate(est, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>and then, with some test set data, try to evaluate the model with:<\/p>\n\n<pre><code>test_fn = lambda: input_fn(DATASET['test'], run_params)\ntest_res = est.evaluate(input_fn=test_fn)\n<\/code><\/pre>\n\n<p>(where the <code>train_fn<\/code> and <code>valid_fn<\/code> are functionally identical to <code>test_fn<\/code>, e.g. sufficient for <code>tf.estimator.train_and_evaluate<\/code> to work).<\/p>\n\n<p>I would expect something to happen, however this is what I get:<\/p>\n\n<pre><code>INFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Starting evaluation at 2018-11-09-13:38:44\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from .\/test\/model.ckpt-100\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\n<\/code><\/pre>\n\n<p>and then it just runs forever.<\/p>\n\n<p>How come?<\/p>\n","answers":[{"AnswerId":"53227406","CreationDate":"2018-11-09T14:16:36.220","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>This is because you repeat the dataset indefinitely:<\/p>\n\n<pre><code># In input_fn\ndataset = dataset.repeat().batch(batch_size)\n<\/code><\/pre>\n\n<p>By default, estimator.evaluate() runs until the input_fn raises an end-of-input exception. Because you repeat the test dataset indefinitely, it never raises the exception and keeps running.<\/p>\n\n<p>You can either remove the repeat when testing, or run the evaluation for a given number of steps using the 'steps' argument as it is used in your original 'eval_spec'.<\/p>\n"}]}
+{"QuestionId":53227410,"AnswerCount":0,"Tags":"<keras><gensim>","CreationDate":"2018-11-09T14:16:51.443","AcceptedAnswerId":null,"OwnerUserId":7227411.0,"Title":"How to print the weights of Keras embedding?","Body":"<p>import gensim<\/p>\n\n<p>Load Google's pre-trained Word2Vec model.<\/p>\n\n<pre><code>model = gensim.models.KeyedVectors.load_word2vec_format('GoogleNews-vectors-negative300.bin', binary=True)\n\nembed = model.get_keras_embedding()\n\nprint(embed.weights)\n<\/code><\/pre>\n\n<p>output: []<\/p>\n\n<p>How to print the weights of embedding layer in Keras, currently it is printing empty list?<\/p>\n","answers":[]}
+{"QuestionId":53227954,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T14:45:46.337","AcceptedAnswerId":53377498.0,"OwnerUserId":8119509.0,"Title":"Unable to import tensorflow, error for importing pywrap_tensorflow","Body":"<p>I am trying to use Keras Sequential, however, my jupyter notebook is flooded with error as it's not able to import tensorflow in the backend (i think). Later I found that, its not with Keras, but I am not able to do 'import tensorflow as tf' as well.<\/p>\n\n<p>Any suggestions, please?\nI am using python 3.5.6\ntensorflow 1.12<\/p>\n\n<p>I did, pip install tensorflow for installation.<\/p>\n\n<p>ImportError                               Traceback (most recent call last)\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in swig_import_helper()\n     17         try:\n---> 18             fp, pathname, description = imp.find_module('_pywrap_tensorflow', [dirname(<strong>file<\/strong>)])\n     19         except ImportError:<\/p>\n\n<p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\imp.py in find_module(name, path)\n    295     else:\n--> 296         raise ImportError(_ERR_MSG.format(name), name=name)\n    297 <\/p>\n\n<p>ImportError: No module named '_pywrap_tensorflow'<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>ModuleNotFoundError                       Traceback (most recent call last)\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py in \n     53     # use <code>dlopen()<\/code> for dynamic loading.\n---> 54     from tensorflow.python import pywrap_tensorflow\n     55 except ImportError:<\/p>\n\n<p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in \n     27             return _mod\n---> 28     _pywrap_tensorflow = swig_import_helper()\n     29     del swig_import_helper<\/p>\n\n<p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py in swig_import_helper()\n     19         except ImportError:\n---> 20             import _pywrap_tensorflow\n     21             return _pywrap_tensorflow<\/p>\n\n<p>ModuleNotFoundError: No module named '_pywrap_tensorflow'<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>ImportError                               Traceback (most recent call last)\n in \n----> 1 import tensorflow as tf<\/p>\n\n<p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow__init__.py in \n     22 \n     23 # pylint: disable=wildcard-import\n---> 24 from tensorflow.python import *\n     25 # pylint: enable=wildcard-import\n     26 <\/p>\n\n<p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py in \n     58 please exit the tensorflow source tree, and relaunch your python interpreter\n     59 from there.\"\"\" % traceback.format_exc()\n---> 60   raise ImportError(msg)\n     61 \n     62 # Protocol buffers<\/p>\n\n<p>ImportError: Traceback (most recent call last):\n  File \"C:\\Users\\ritesh.kankonkar\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 18, in swig_import_helper\n    fp, pathname, description = imp.find_module('_pywrap_tensorflow', [dirname(<strong>file<\/strong>)])\n  File \"C:\\Users\\ritesh.kankonkar\\AppData\\Local\\Continuum\\anaconda3\\lib\\imp.py\", line 296, in find_module\n    raise ImportError(_ERR_MSG.format(name), name=name)\nImportError: No module named '_pywrap_tensorflow'<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"C:\\Users\\ritesh.kankonkar\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py\", line 54, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\ritesh.kankonkar\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 28, in \n    _pywrap_tensorflow = swig_import_helper()\n  File \"C:\\Users\\ritesh.kankonkar\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 20, in swig_import_helper\n    import _pywrap_tensorflow\nModuleNotFoundError: No module named '_pywrap_tensorflow'<\/p>\n\n<p>Error importing tensorflow.  Unless you are using bazel,\nyou should not try to import tensorflow from its source directory;\nplease exit the tensorflow source tree, and relaunch your python interpreter\nfrom there.<\/p>\n","answers":[{"AnswerId":"53377498","CreationDate":"2018-11-19T15:08:51.277","ParentId":null,"OwnerUserId":"8119509","Title":null,"Body":"<p>Well, I am answering my own question since the error seems to have multiple causes.\nI am not sure what was the cause, however, after downgrading python to 3.5 and installing tensorflow with pip (pip install tensorflow), resolved the issue.<\/p>\n\n<p>Note: I uninstalled everything before installing Anaconda again.<\/p>\n"}]}
+{"QuestionId":53228244,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T15:04:58.837","AcceptedAnswerId":null,"OwnerUserId":9290285.0,"Title":"tensorflow: two initializers ops found for one convlayer","Body":"<p>when i read my model operations with <\/p>\n\n<pre><code>tf.get_default_graph().as_graph_def().node\n<\/code><\/pre>\n\n<p>if find multiple initializer ops for the same weight layer:<\/p>\n\n<pre><code>'model\/conv2d\/kernel\/Initializer\/truncated_normal\/TruncatedNormal'\n'model\/conv2d\/kernel\/Initializer\/truncated_normal'\n<\/code><\/pre>\n\n<p>when i read back the tensors with <\/p>\n\n<pre><code>tf.get_default_graph().get_tensor_by_name('model\/conv2d\/kernel\/Initializer\/truncated_normal\/TruncatedNormal:0')\n<\/code><\/pre>\n\n<p>, they have the same shape\/dimension but other values.<\/p>\n\n<p>Any idea what's happening here?<\/p>\n","answers":[]}
+{"QuestionId":53228768,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-11-09T15:35:28.667","AcceptedAnswerId":null,"OwnerUserId":10072879.0,"Title":"how to make a input tensor trainable","Body":"<p>The following is the code trying to use the input image as a training variable during the optimization. It starts with a keras model and convert that to a tensorflow model. This tensorflow model take a tensor as input and try to optimize the cost function using the input tensor as a trainable variable. <\/p>\n\n<p>The error is: <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>NotImplementedError: ('Trying to update a Tensor ', )<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>The reason is that the input tensor is not a variable. The question is how to make the input image trainable or cast a tensor to a tf.variable. Help is appreciated: <\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.models import Sequential, load_model, Model\nfrom keras import backend as K\nfrom keras.layers.core import Dense, Dropout, Activation\nimport os\nfrom tensorflow.python.framework import graph_util\nfrom tensorflow.python.framework import graph_io\n\nn_classes = 10\n\nmodel = Sequential()\nmodel.add(Dense(10, input_shape=(784,)))\nmodel.add(Activation('relu'))                            \nmodel.add(Dense(n_classes, name='logits'))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='categorical_crossentropy', metrics=['accuracy'],  optimizer='adam')\n\n# Write the graph in binary .pb file\noutdir = \"model4_tf\"\ntry:\n    os.mkdir(outdir )\nexcept:\n    pass\n\n\nprefix = \"simple_nn\" \nname = 'output_graph.pb'\n# Alias the outputs in the model - this sometimes makes them easier to access in TF\npred = []\npred_node_names = []\nfor i, o in enumerate(model.outputs):\n    pred_node_names.append(prefix+'_'+str(i))\n    pred.append(tf.identity(o, \n                            name=pred_node_names[i]))\nprint('Output nodes names are: ', pred_node_names)\n\n\nsess = K.get_session()\n\n\nconstant_graph = graph_util.convert_variables_to_constants(sess,                                         \nsess.graph.as_graph_def(), pred_node_names)\ngraph_io.write_graph(constant_graph, outdir, name, as_text=False)\n\n\ntf.reset_default_graph()\n\ndef load_graph(model_name):\n    #graph = tf.Graph()\n    graph = tf.get_default_graph()\n    graph_def = tf.GraphDef()\n    with open(model_name, \"rb\") as f:\n        graph_def.ParseFromString(f.read())\n    with graph.as_default():\n        tf.import_graph_def(graph_def)\n    return graph\n\nmy_graph = load_graph(model_name=os.path.join(outdir, name))\n\n\n# In[15]:\n\ninput_op = my_graph.get_operation_by_name(\"import\/dense_1_input\")\noutput_op = my_graph.get_operation_by_name(\"import\/simple_nn_0\")\nlogit_op = my_graph.get_operation_by_name(\"import\/logits\/BiasAdd\")\n\n\nx_hat = input_op.outputs[0] # input tensor\nlabels = output_op.outputs[0] # label tensor\nlogits = logit_op.outputs[0] # logits tensor\n\nlearning_rate = tf.placeholder(tf.float32, ())\n\nloss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=[labels])\noptim_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss,     var_list=[x_hat])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53228969,"AnswerCount":2,"Tags":"<android><tensorflow><tensorflow-lite><deeplab>","CreationDate":"2018-11-09T15:47:11.087","AcceptedAnswerId":null,"OwnerUserId":3177661.0,"Title":"Unable to test and deploy a deeplabv3-mobilenetv2 tensorflow-lite segmentation model for inference","Body":"<p>We are trying to run a semantic segmentation model on android using deeplabv3 and mobilenetv2.We followed the official tensorflow lite conversion procedure using TOCO and tflite_convert with the help of bazel.The source frozen graph was obtained from the official TensorFlow DeepLab Model Zoo.<\/p>\n\n<p>We were able to successfully convert the model with the following command:-<\/p>\n\n<blockquote>\n  <p>CUDA_VISIBLE_DEVICES=\"0\" toco --output_file=toco256.tflite\n  --graph_def_file=path\/to\/deeplab\/deeplabv3_mnv2_pascal_trainval\/frozen_inference_graph.pb\n  --input_arrays=ImageTensor --output_arrays=SemanticPredictions --input_shapes=1,256,256,3 --inference_input_type=QUANTIZED_UINT8 --inference_type=FLOAT --mean_values=128 --std_dev_values=127 --allow_custom_ops --post_training_quantize<\/p>\n<\/blockquote>\n\n<p>The size of the tflite file was around 2.25 Mb.But when we tried to test the model using the official benchmark tool, it failed with the following  error report :-<\/p>\n\n<pre><code>bazel run -c opt tensorflow\/contrib\/lite\/tools\/benchmark:benchmark_model -- --graph=`realpath toco256.tflite`\nINFO: Analysed target \/\/tensorflow\/contrib\/lite\/tools\/benchmark:benchmark_model (0 packages loaded).\nINFO: Found 1 target...\nTarget \/\/tensorflow\/contrib\/lite\/tools\/benchmark:benchmark_model up-to-date:\n  bazel-bin\/tensorflow\/contrib\/lite\/tools\/benchmark\/benchmark_model\nINFO: Elapsed time: 0.154s, Critical Path: 0.00s\nINFO: 0 processes.\nINFO: Build completed successfully, 1 total action\nINFO: Running command line: bazel-bin\/tensorflow\/contrib\/lite\/tools\/benchmark\/benchmark_model '--graph=path\/to\/deeplab\/venINFO: Build completed successfully, 1 total action\nSTARTING!\nNum runs: [50]\nInter-run delay (seconds): [-1]\nNum threads: [1]\nBenchmark name: []\nOutput prefix: []\nWarmup runs: [1]\nGraph: path\/to\/venv\/tensorflow\/toco256.tflite]\nInput layers: []\nInput shapes: []\nUse nnapi : [0]\nLoaded model path\/to\/venv\/tensorflow\/toco256.tflite\nresolved reporter\nInitialized session in 45.556ms\nRunning benchmark for 1 iterations \ntensorflow\/contrib\/lite\/kernels\/pad.cc:96 op_context.dims != 4 (3 != 4)\nNode number 24 (PAD) failed to prepare.\n\nFailed to invoke!\nAborted (core dumped)\n<\/code><\/pre>\n\n<p>We also tried the same command without including the 'allow_custom_ops' and 'post_training_quantize' options and even used the same input size as 1,513,513,3; but the result was the same.<\/p>\n\n<p>This issue seems to be similar to the following github issue:\n (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/21266\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/21266<\/a>). However in the latest version of TensorFlow the issue  is supposed to be fixed.<\/p>\n\n<p>Model: <a href=\"http:\/\/download.tensorflow.org\/models\/deeplabv3_mnv2_pascal_trainval_2018_01_29.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/deeplabv3_mnv2_pascal_trainval_2018_01_29.tar.gz<\/a>\nTensorflow version: 1.11\nBazel version: 0.17.2\nOS: Ubuntu 18.04<\/p>\n\n<p>Also the android application was not able to load the model properly (tflite interpretr)<\/p>\n\n<p>So, how can we convert a segmentation model properly to a tflite format which can be used for inference on an android device?<\/p>\n\n<p>UPDATE:-<\/p>\n\n<p>Using tensorflow 1.12, we got a new error :<\/p>\n\n<pre><code>$ bazel run -c opt tensorflow\/lite\/tools\/benchmark:benchmark_model -- --graph=`realpath \/path\/to\/research\/deeplab\/venv\/tensorflow\/toco256.tflite`\n\n    tensorflow\/lite\/kernels\/depthwise_conv.cc:99 params-&gt;depth_multiplier * SizeOfDimension(input, 3) != SizeOfDimension(filter, 3) (0 != 32)\n    Node number 30 (DEPTHWISE_CONV_2D) failed to prepare.\n<\/code><\/pre>\n\n<p>Also,while using a newer version of the same model(3 Mb .pb file) with depth_multiplier=0.5 from the tensorflow deeplab <a href=\"http:\/\/download.tensorflow.org\/models\/deeplabv3_mnv2_dm05_pascal_trainval_2018_10_01.tar.gz\" rel=\"nofollow noreferrer\">model zoo<\/a>, we got a different error:-<\/p>\n\n<pre><code>F tensorflow\/lite\/toco\/graph_transformations\/propagate_fixed_sizes.cc:116] Check failed: dim_x == dim_y (3 vs. 32)Dimensions must match\n<\/code><\/pre>\n\n<p>In this case we used the same aforementioned command for tflite conversion ;but we were not even  able to produce a 'tflite' file as output.It seems to be an issue with depth multiplier values.(Even we tried giving the depth_multiplier parameter as argument at the time of conversion).<\/p>\n","answers":[{"AnswerId":"53439660","CreationDate":"2018-11-23T01:25:08.960","ParentId":null,"OwnerUserId":"9797365","Title":null,"Body":"<p>I have same issue. from <a href=\"https:\/\/github.com\/tantara\/JejuNet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tantara\/JejuNet<\/a> I see that he was successfully convert model to tflite. I PM him for help, but unfortunately no response right now.<\/p>\n"},{"AnswerId":"59108370","CreationDate":"2019-11-29T16:35:05.417","ParentId":null,"OwnerUserId":"9282580","Title":null,"Body":"<p>I also ran into this problem. There seems to be 2 issues in the conversion:<\/p>\n\n<ul>\n<li>The input tensor has a dynamic shape, that is [?,?,?,3]<\/li>\n<li>The pad_to_bounding_box node part is not auto converted to a static shape<\/li>\n<\/ul>\n\n<p>For the solution below, this is tested on:<\/p>\n\n<ul>\n<li>Tensorflow 1.15<\/li>\n<li>Ubuntu 16.0.4<\/li>\n<\/ul>\n\n<h1>Solution<\/h1>\n\n<p>I assume you've already created a .pb file using the export_model.py file in the deeplab folder and named this file deeplab_mobilenet_v2.pb. From here on:<\/p>\n\n<h2>STEP 1: OPTIMIZE FOR INFERENCE<\/h2>\n\n<ul>\n<li>download optimize_for_inference.py <a href=\"https:\/\/raw.githubusercontent.com\/benoitsteiner\/tensorflow-opencl\/master\/tensorflow\/python\/tools\/optimize_for_inference.py\" rel=\"nofollow noreferrer\">https:\/\/raw.githubusercontent.com\/benoitsteiner\/tensorflow-opencl\/master\/tensorflow\/python\/tools\/optimize_for_inference.py<\/a><\/li>\n<li>run optimization (change parameters according to your config)<\/li>\n<\/ul>\n\n<pre><code>python3 optimize_for_inference.py \\\n        --input \"path\/to\/your\/deeplab_mobilenet_v2.pb\" \\\n        --output \"path\/to\/deeplab_mobilenet_v2_opt.pb\" \\\n        --frozen_graph True \\\n        --input_names ImageTensor \\\n        --output_names SemanticPredictions \\\n        --placeholder_type_enum=4\n<\/code><\/pre>\n\n<p>placeholder_type_enum=4 is the uint8 data type (dtypes.uint8.as_datatype_enum)<\/p>\n\n<h1>STEP 2: APPLY GRAPH TRANSFORM TOOL<\/h1>\n\n<p>Make sure you have installed bazel and have downloaded the tensorflow r1.15 branch from github. Then make the transform_graph tool from the tensorflow repo:<\/p>\n\n<p><code>bazel build tensorflow\/tools\/graph_transforms:transform_graph<\/code><\/p>\n\n<p>Then run the transform_graph tool (make sure to set the shape to whatever shape you are using as input):<\/p>\n\n<pre><code>bazel-bin\/tensorflow\/tools\/graph_transforms\/transform_graph \\\n--in_graph=\"\/path\/to\/deeplab_mobilenet_v2_opt.pb\" \\\n--out_graph=\"\/path\/to\/deeplab_mobilenet_v2_opt_flatten.pb\" \\\n--inputs='ImageTensor' \\\n--outputs='SemanticPredictions' \\\n--transforms='\n    strip_unused_nodes(type=quint8, shape=\"1,400,225,3\")\n    flatten_atrous_conv\n    fold_constants(ignore_errors=true, clear_output_shapes=false)\n    fold_batch_norms\n    fold_old_batch_norms\n    remove_device\n    sort_by_execution_order'\n<\/code><\/pre>\n\n<h1>STEP 3: BYPASS pad_to_bounding_box NODE AND MAKE INPUT STATIC<\/h1>\n\n<p>Run the python file below, making sure to change the model_filepath, save_folder and save_name to whatever suits your needs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.contrib import graph_editor as ge\n\ndef freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):\n    \"\"\"\n    Freezes the state of a session into a pruned computation graph.\n\n    Creates a new computation graph where variable nodes are replaced by\n    constants taking their current value in the session. The new graph will be\n    pruned so subgraphs that are not necessary to compute the requested\n    outputs are removed.\n    @param session The TensorFlow session to be frozen.\n    @param keep_var_names A list of variable names that should not be frozen,\n                          or None to freeze all the variables in the graph.\n    @param output_names Names of the relevant graph outputs.\n    @param clear_devices Remove the device directives from the graph for better portability.\n    @return The frozen graph definition.\n    \"\"\"\n    graph = session.graph\n    with graph.as_default():\n        freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))\n        output_names = output_names or []\n        output_names += [v.op.name for v in tf.global_variables()]\n        input_graph_def = graph.as_graph_def()\n        if clear_devices:\n            for node in input_graph_def.node:\n                node.device = \"\"\n        frozen_graph = tf.graph_util.convert_variables_to_constants(\n            session, input_graph_def, output_names, freeze_var_names)\n        return frozen_graph\n\ndef load_convert_save_graph(model_filepath, save_folder, save_name):\n    '''\n    Lode trained model.\n    '''\n    print('Loading model...')\n    graph = tf.Graph()\n    sess = tf.InteractiveSession(graph = graph)\n\n    with tf.gfile.GFile(model_filepath, 'rb') as f:\n        graph_def = tf.GraphDef()\n        graph_def.ParseFromString(f.read())\n\n    print('Check out the input placeholders:')\n    nodes = [n.name + ' =&gt; ' +  n.op for n in graph_def.node if n.op in ('Placeholder')]\n    for node in nodes:\n        print(node)\n\n    # Define input tensor\n    input = tf.placeholder(np.uint8, shape = [1,400,225,3], name='ImageTensor')\n\n    tf.import_graph_def(graph_def, {'ImageTensor': input}, name='')\n\n    print('Model loading complete!')\n\n    # remove the pad to bounding box node\n    name = \"pad_to_bounding_box\"\n    print(name)\n    sgv = ge.make_view_from_scope(name, tf.get_default_graph())\n    print(\"\\t\" + sgv.inputs[0].name)\n    for node in sgv.inputs:\n        print(\"name in = \" + node.name)\n    for node in sgv.outputs:\n        print(\"name out = \" + node.name)\n    print(\"\\t\" + sgv.outputs[len(sgv.outputs)-1].name)\n    sgv = sgv.remap_inputs([0])\n    sgv = sgv.remap_outputs([len(sgv.outputs)-1])\n    (sgv2, det_inputs) = ge.bypass(sgv)\n\n\n    frozen_graph = freeze_session(sess,\n                              output_names=['SemanticPredictions'])\n    tf.train.write_graph(frozen_graph, save_folder, save_name, as_text=False)\n\n\nload_convert_save_graph(\"path\/to\/deeplab_mobilenet_v2_opt_flatten.pb\", \"\/path\/to\", \"deeplab_mobilenet_v2_opt_flatten_static.pb\")\n\n<\/code><\/pre>\n\n<h1>STEP 4: CONVERT TO TFLITE<\/h1>\n\n<pre><code>tflite_convert \\\n  --graph_def_file=\"\/path\/to\/deeplab_mobilenet_v2_opt_flatten_static.pb\" \\\n  --output_file=\"\/path\/to\/deeplab_mobilenet_v2_opt_flatten_static.tflite\" \\\n  --output_format=TFLITE \\\n  --input_shape=1,400,225,3 \\\n  --input_arrays=\"ImageTensor\" \\\n  --inference_type=FLOAT \\\n  --inference_input_type=QUANTIZED_UINT8 \\\n  --std_dev_values=128 \\\n  --mean_values=128 \\\n  --change_concat_input_ranges=true \\\n  --output_arrays=\"SemanticPredictions\" \\\n  --allow_custom_ops\n<\/code><\/pre>\n\n<h1>DONE<\/h1>\n\n<p>You can now run your tflite model<\/p>\n"}]}
+{"QuestionId":53229283,"AnswerCount":1,"Tags":"<keras><auto-keras>","CreationDate":"2018-11-09T16:04:19.490","AcceptedAnswerId":null,"OwnerUserId":10629741.0,"Title":"Autokeras exported model performs not as expected","Body":"<p>I just installed autokeras on python3.6. After some bug fixing it works well and I can train models with my dataset. \nAfter training i get a model with an accuracy of 0.991 and loss of 0.06. With the predict-function it classifies my test dataset almost without any fault.\nBut when I export and save it as a keras model the model performs really bad. It classifies everything wrongly and the predict-function returns just random decimals and not the number of the expected class.\nI think I am just missing some commands since the image classifier works well.\nAny help or tips would be very nice.<\/p>\n\n<p>Edit:\nI basically followed the example on <a href=\"https:\/\/autokeras.com\/start\/\" rel=\"nofollow noreferrer\">https:\/\/autokeras.com\/start\/<\/a>. I preprocessed my data to be in shape (n, 150, 150, 3) and range between 0 and 1. The code for each image is the following:<\/p>\n\n<pre><code>img='\/home\/example_image.png'\nx=image.img_to_array(image.load_img(img, target_size=(150, 150)))\nx=x.reshape(1, 150, 150, 3)\nx = x.astype('float32') \/ 255\n<\/code><\/pre>\n\n<p>I also tried with 50x50 shape for the images because autokeras seems to work better with smaller shapes. My labels are a list of length n. Then I use the code from the website:<\/p>\n\n<pre><code>clf = ImageClassifier(verbose=True)\nclf.fit(x_train, y_train, time_limit=12 * 60 * 60)\nclf.final_fit(x_train, y_train, x_test, y_test, retrain=True)\ny = clf.evaluate(x_test, y_test)\nprint(y)\n<\/code><\/pre>\n\n<p>It returns a very good model with high accuracy and small loss. I can't show you the screenshots of the output because I am currently at home and my laptop is too weak to reproduce this problem. When I use the following command:<\/p>\n\n<pre><code>clf.predict(example_img)\n<\/code><\/pre>\n\n<p>I also get a correct result with the number of object in the image. \nMy problem starts when I try to export the model with the commands given on their website:<\/p>\n\n<pre><code>from autokeras import ImageClassifier\nclf = ImageClassifier(verbose=True, augment=False)\nclf.load_searcher().load_best_model().produce_keras_model().save('my_model.h5')\n<\/code><\/pre>\n\n<p>After that I load my model.<\/p>\n\n<pre><code>import keras\nfrom keras.models import load_model\nmodel = load_model('my_model.h5')\n<\/code><\/pre>\n\n<p>It gives me some error:<\/p>\n\n<blockquote>\n  <p>\"UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually.\n  warnings.warn('No training configuration found in save file: '\"<\/p>\n<\/blockquote>\n\n<p>I searched for it but people said to ignore it.\nBut when I try to use predict with this loaded model I always get an array of some random decimal values which are often negative. The classification is always wrong since the highest value in this array is for a wrong class.<\/p>\n\n<p>I will edit this text with more screenshots and details when I am at the office again with my ML computer.<\/p>\n\n<p>Edit 2:\nWhen I train the model this is the best model :<\/p>\n\n<pre><code>+--------------------------------------------------------------------------+\n|    Father Model ID     |                 Added Operation                 |\n+--------------------------------------------------------------------------+\n|                        |           ('to_add_skip_model', 1, 5)           |\n|                        |            ('to_wider_model', 1, 64)            |\n|                        |            ('to_wider_model', 5, 64)            |\n|           5            |          ('to_conv_deeper_model', 9, 3)         |\n|                        |         ('to_conv_deeper_model', 23, 3)         |\n|                        |          ('to_concat_skip_model', 5, 9)         |\n+--------------------------------------------------------------------------+\n\nSaving model.\n+--------------------------------------------------------------------------+\n|        Model ID        |          Loss          |      Metric Value      |\n+--------------------------------------------------------------------------+\n|           6            |  0.014135746611282229  |          1.0           |\n+--------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>After that I used final_fit and tried the predict_function:<\/p>\n\n<pre><code>&gt;&gt;&gt; clf.predict(test_images)\narray([ 0.,  0.,  0., ..., 12., 12., 12.])\n<\/code><\/pre>\n\n<p>This is the expected output. Evaluation gives also good results:<\/p>\n\n<pre><code>&gt;&gt;&gt; y = clf.evaluate(test_images, test_labels)\n&gt;&gt;&gt; print(y)\n0.9969230769230769\n<\/code><\/pre>\n\n<p>Then I export the model as a keras model and load it again:<\/p>\n\n<pre><code>&gt;&gt;&gt; from keras import models\n&gt;&gt;&gt; clf.load_searcher().load_best_model().produce_keras_model().save('keras_best_model.h5')\n&gt;&gt;&gt; model = models.load_model('keras_best_model.h5')\n<\/code><\/pre>\n\n<p>Everything does fine and I get no error. But when I use the predict_function now it returns wrong results:<\/p>\n\n<pre><code>&gt;&gt;&gt; model.predict(test_images[0].reshape(1, 50, 50, 3))\narray([[ 2.5287893, -2.2281592, -2.8172228,  1.1171696, -5.8477755,\n        -3.1250796,  1.4904132,  1.1068834, -4.982565 , -1.6350467,\n        -4.3806715, -2.7464929, -6.1051216]], dtype=float32)\n<\/code><\/pre>\n\n<p>So clf.predict gives me one number and model.predict give me an array. Shouldn't it be the same? Almost all predictions are wrong. When I evaluate the model this is the output:<\/p>\n\n<pre><code>&gt;&gt;&gt; model.evaluate(test_images, test_labels1)\n2600\/2600 [==============================] - 1s 569us\/step\n[5.251570468682509, 0.10115384615384615]\n<\/code><\/pre>\n\n<p>It is totally bad in comparison to the clf evaluation and I don't know why.<\/p>\n","answers":[{"AnswerId":"54571718","CreationDate":"2019-02-07T10:55:47.570","ParentId":null,"OwnerUserId":"9442808","Title":null,"Body":"<p>In the new version, (0.3.6), it seems<\/p>\n\n<pre><code>clf.load_searcher().load_best_model().produce_keras_model().save('keras_best_model.h5') \n<\/code><\/pre>\n\n<p>is removed. For the moment you can use \"autokeras.utils.pickle_to_file\" for saving and \"autokeras.utils.pickle_from_file\" to load the model. This is not a usual keras model, however we can get same results.  <\/p>\n"}]}
+{"QuestionId":53230290,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-09T17:04:20.213","AcceptedAnswerId":55721363.0,"OwnerUserId":10458815.0,"Title":"How do I export an eager execution model?","Body":"<p>Having completed my model, I now wish to export and deploy it, following <a href=\"https:\/\/www.tensorflow.org\/serving\/serving_basic#train_and_export_tensorflow_model\" rel=\"nofollow noreferrer\">this<\/a> tutorial on TensorFlow's website. However, there is no indication of how to do this in eager execution, where I am unable to provide a session or graph to <code>builder.add_meta_graph_and_variables()<\/code>.<\/p>\n\n<p>Is this a case where my code needs to be eager and graph compatible, or where I need to save my model, import it to a session, and export it from there?<\/p>\n","answers":[{"AnswerId":"55721363","CreationDate":"2019-04-17T06:39:15.997","ParentId":null,"OwnerUserId":"11253331","Title":null,"Body":"<p>TF 2.0 Alpha supports exporting\/saving an Eager Execution Model (Default is Eager Execution in 2.0). A SavedModel contains a complete TensorFlow program, including weights and computation. <\/p>\n\n<p><strong>Sample code for the same is shown below:<\/strong><\/p>\n\n<pre><code>!pip install -q tensorflow==2.0.0-alpha0\nimport tensorflow as tf\n\npretrained_model = tf.keras.applications.MobileNet()\ntf.saved_model.save(pretrained_model, \"\/tmp\/mobilenet\/1\/\")\n\n#Loading the saved model\nloaded = tf.saved_model.load(\"\/tmp\/mobilenet\/1\/\")\ninfer = loaded.signatures[\"serving_default\"]\n<\/code><\/pre>\n\n<p><strong>Serving the Model:<\/strong><\/p>\n\n<pre><code>nohup tensorflow_model_server \\\n  --rest_api_port=8501 \\\n  --model_name=mobilenet \\\n  --model_base_path=\"\/tmp\/mobilenet\" &gt;server.log 2&gt;&amp;1\n\n#Sending the Request for Inference\n\n!pip install -q requests\nimport json\nimport numpy\nimport requests\ndata = json.dumps({\"signature_name\": \"serving_default\",\n                   \"instances\": x.tolist()})\nheaders = {\"content-type\": \"application\/json\"}\njson_response = requests.post('http:\/\/localhost:8501\/v1\/models\/mobilenet:predict',\n                              data=data, headers=headers)\npredictions = numpy.array(json.loads(json_response.text)[\"predictions\"])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53230357,"AnswerCount":0,"Tags":"<ubuntu><anaconda><nvidia><pytorch>","CreationDate":"2018-11-09T17:08:34.893","AcceptedAnswerId":null,"OwnerUserId":4794147.0,"Title":"Pytorch\/CUDA - Cannot find GPU","Body":"<p>I'm trying to use my GPU to train a network. I installed PyTorch and CUDA with anaconda. I have an NVIDA GTX 860M on my computer I would like to use. <\/p>\n\n<p>When using the command :<\/p>\n\n<pre><code>$ lspci | grep NVIDIA\n<\/code><\/pre>\n\n<p>I get the output :<\/p>\n\n<pre><code>01:00.0 3D controller: NVIDIA Corporation GM107M [GeForce GTX 860M] (rev a2)\n<\/code><\/pre>\n\n<p>So the graphic card is there. But in a Jupyter notebook, the following lines :<\/p>\n\n<pre><code>print(torch.cuda.device_count())\nprint(torch.cuda.is_available())\nprint(torch.version.cuda)\n<\/code><\/pre>\n\n<p>Return :<\/p>\n\n<pre><code>0\nFalse\n9.0.176\n<\/code><\/pre>\n\n<p>I have tried updating the latest Cuda driver using conda it did not solve the issue. My PyTorch version is 0.4.1.post2. I couldn't find an answer browsing through the different online forums. \nI am on Ubuntu 18.04.1 LTS.\nThanks' in advance !<\/p>\n","answers":[]}
+{"QuestionId":53230422,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-09T17:13:29.080","AcceptedAnswerId":null,"OwnerUserId":7867665.0,"Title":"How to add scikit-kearn based loss to neural network?","Body":"<p>I want to train a neural network that has a loss component coming from an AdaBoost classifier. <\/p>\n\n<pre><code>myLoss(y_true, y_pred, adaClf, Z_):\n    loss = BinaryCrossEntropy(y_true, y_pred) + MeanSquareError(adaClf.predict(Y_pred), Z_)\n    return loss\n<\/code><\/pre>\n\n<p>where <code>adaClf<\/code> is the scikit-learn classifier. How can I implement this in Keras?<\/p>\n","answers":[{"AnswerId":"53231053","CreationDate":"2018-11-09T17:59:18.683","ParentId":null,"OwnerUserId":"8909870","Title":null,"Body":"<p>Sorry, but you can not use external functions as loss in tensorflow. With <code>tf.map<\/code> you can use external functions, but no gradients can flow through it and you definitely need gradients in your loss. So you probably have to implement the classifier in tensorflow. Maybe <a href=\"https:\/\/stackoverflow.com\/questions\/35691636\/adaboost-with-neural-networks\">this<\/a> helps you.<\/p>\n"}]}
+{"QuestionId":53230682,"AnswerCount":1,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-11-09T17:32:09.750","AcceptedAnswerId":null,"OwnerUserId":10630058.0,"Title":"calculate equation with non linear data","Body":"<p>I'm making my first steps with Tensorflow.js and I need to calculate the equation<\/p>\n\n<pre><code>y = [(x * 100) \/ a]\/100\n<\/code><\/pre>\n\n<p>I.E. given 4 tensors like:<\/p>\n\n<pre><code>[1,1,1,0], [2,2,2,1], [0,0,1,0], [0,2,2,1]\n<\/code><\/pre>\n\n<p>the sums of all values for each tensor would be:<\/p>\n\n<pre><code>3, 7, 1, 5\n<\/code><\/pre>\n\n<p>the sums of those values would be:<\/p>\n\n<pre><code>15\n<\/code><\/pre>\n\n<p>and the equation above would be:<\/p>\n\n<pre><code>y = [(3 * 100) \/ 15]\/100\ny = [(7 * 100) \/ 15]\/100\ny = [(1 * 100) \/ 15]\/100\ny = [(5 * 100) \/ 15]\/100\n<\/code><\/pre>\n\n<p>so the output tensor should be: <\/p>\n\n<pre><code>[0.19], [0.44], [0.06], [0.31]\n<\/code><\/pre>\n\n<p>I made the code below where I tried to train a model to solve the equation, but the results are far from being acceptable.\nI even tried generating 60 couples of input and output examples, augmenting epochs of training to 50k and augmenting the number of units of the input layer, but the results seem to be even worse.\nCan you give me some help? Where am I mistaking?\nThank you!<\/p>\n\n<pre><code>&lt;script&gt;\n    async function predictOutput() {\n\n        const model = tf.sequential();\n        \/\/config for the hidden layer\n        const config_hidden = {\n          inputShape:[4],\n          activation:'sigmoid',\n          units:4\n        }\n        \/\/config for the output layer\n        const config_output={\n          units:1,\n          activation:'sigmoid'\n        }\n        \/\/defining the hidden and output layer\n        const hidden = tf.layers.dense(config_hidden);\n        const output = tf.layers.dense(config_output);\n        \/\/adding layers to model\n        model.add(hidden);\n        model.add(output);\n        \/\/define an optimizer\n        const optimize=tf.train.sgd(0.1);\n        \/\/config for model\n        const config={\n        optimizer:optimize,\n        loss:'meanSquaredError'\n        }\n        \/\/compiling the model\n        model.compile(config);\n\n        \/\/Dummy training data\n        const x_train = tf.tensor([\n        [1,0,0,3], [0,0,3,0], [1,0,0,0], [0,1,0,4],\n        [0,0,0,1], [2,0,2,1], [2,4,1,0], [0,2,0,1],\n        [1,1,1,0], [2,2,2,1], [0,0,1,0], [0,2,2,1],\n        [1,0,0,0], [0,1,0,0], [1,1,1,0], [2,2,2,2],\n        [2,5,7,9], [2,1,0,10], [22,5,7,9], [2,0,3,1],\n        [1,1,1,1], [2,2,2,2], [0,5,8,1], [5,5,8,1],\n        [3,4,1,5], [1,0,3,1], [5,5,1,0], [4,2,6,0],\n        [1,0,0,0], [1,1,2,1], [1,3,2,1], [1,2,0,0],\n        [1,0,0,2], [0,0,0,7], [0,1,0,0], [5,0,0,0],\n        [0,4,0,0], [1,0,7,0], [3,2,8,1], [0,10,9,0]\n        ]);\n\n        \/\/Dummy training labels\n        const y_train = tf.tensor([\n        [0.31], [0.23], [0.08], [0.38],\n        [0.07], [0.31], [0.44], [0.18],\n        [0.19], [0.44], [0.06], [0.31],\n        [0.08], [0.08], [0.23], [0.61],\n        [0.27], [0.15], [0.51], [0.07],\n        [0.09], [0.18], [0.31], [0.42],\n        [0.32], [0.12], [0.27], [0.29],\n        [0.07], [0.31], [0.44], [0.18],\n        [0.19], [0.44], [0.06], [0.31],\n        [0.09], [0.18], [0.31], [0.42]\n        ]);\n\n        \/\/Dummy testing data\n        const x_test = tf.tensor([\n            [1,0,0,1], [0,1,1,0], [2,0,1,2], [0,0,0,1]\n        ]);\n\n        \/\/ expected result: [0.20], [0.20], [0.50], [0.10]\n\n\n        await model.fit(x_train, y_train, {\n            batchSize: 1,\n            epochs: 5000\n        });\n\n        \/\/ Test the model and display output \n        document.getElementById(\"output\").innerText = model.predict(x_test);\n    }\n    predictOutput();\n&lt;\/script&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"53235987","CreationDate":"2018-11-10T04:27:24.187","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>If the linear equation is known, i.e <code>y = 1\/15x<\/code>, then the output can be calculated directly using the following function<\/p>\n\n<pre><code>const calc = (t) =&gt; t.add(tf.scalar(1\/15))\n<\/code><\/pre>\n\n<p>If the slope <code>a<\/code> and the bias <code>b<\/code> are now known in advance and need to be computed, a perceptron neural network can be used. \nThe inputShape of the model is wrong since the prediction is to be done for a single value. It should be [1] instead of 4. As for the x_test, it should look like an array of array of length 1 which means the following:<code>[[1], [2], [3], [56], [49]]<\/code> <\/p>\n\n<p>Also using a smaller learning rate might lead to better accuracy<\/p>\n\n<p>EDIT:<\/p>\n\n<p>If the model has to predict on the sum of the inputs, then, <code>inputShape<\/code> should be [4]. Your input values are small and that is affecting the error loss using a perceptron model. Before giving the data to the model, one can process the input values multiplying by 10 for instance. As a result, the model prediction will be 10 times the correct values.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential()\r\nmodel.add(tf.layers.dense({inputShape: [4], units: 1 }))\r\nmodel.add(tf.layers.dense({units: 1}))\r\nconst x_train = tf.tensor([\r\n        [1,0,0,3], [0,0,3,0], [1,0,0,0], [0,1,0,4], [0,0,0,1]]);\r\nconst y = tf.tensor([\r\n        [0.31], [0.23], [0.08], [0.38],[0.07]\r\n        ]);\r\n\r\nconst y_train = y.mul(tf.scalar(10))\r\nconst optimizer = tf.train.sgd( 0.01 )\r\nmodel.compile({optimizer: optimizer, loss: 'meanSquaredError' })\r\nmodel.fit(x_train, y_train, {epochs: 1000, \r\n      callbacks: {\r\n      onEpochEnd: (epoch, log) =&gt; {\r\n        console.log(epoch, log.loss);\r\n        if (epoch === 100 ) {\r\n           model.optimizer.setLearningRate(.001);\r\n        }\r\n      }\r\n    }}).then(() =&gt; {\r\n  const y_predict = model.predict(x_train)\r\n  const y_correct = y_predict.div(tf.scalar(10)).print()\r\n})<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.3\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53230925,"AnswerCount":1,"Tags":"<python><keras><predict>","CreationDate":"2018-11-09T17:50:34.950","AcceptedAnswerId":null,"OwnerUserId":7313100.0,"Title":"Wrong predictions using model.predict()","Body":"<p>I have a problem with making predictions in transfer-learned VGG16 net. I have a model trained with Adam on 7 classes. It was trained with fit_generator(), using ImageDataGenerator. I am loading model using:<\/p>\n\n<pre><code># load the model we saved\nmodel = load_model('models\/vgg16_9.h5')\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=optimizers.Adam(lr=1e-4),\n              metrics=['acc'])\n<\/code><\/pre>\n\n<p>And then trying to made predictions. First, I used predict_generator() to make a .CSV file with results:<\/p>\n\n<pre><code>test_datagen = ImageDataGenerator(rescale = 1.\/255)\ntest_generator = test_datagen.flow_from_directory(\"dataset\/test_set\",\n                                                  target_size=(227, 454),\n                                                  batch_size=1,\n                                                  class_mode=None,\n                                                  shuffle=False,\n                                                  seed=42)\n\ntest_generator.reset()\npred = model.predict_generator(test_generator, verbose = 1)\npredicted_class_indices = np.argmax(pred, axis = 1)\n\nlabels = (valid_generator.class_indices)\nlabels = dict((v,k) for k,v in labels.items())\npredictions = [labels[k] for k in predicted_class_indices]\n\nfilenames=test_generator.filenames\nresults=pd.DataFrame({\"Filename\":filenames,\n                      \"Predictions\":predictions})\nresults.to_csv(\"results.csv\",index=False)\n<\/code><\/pre>\n\n<p>It works ok, and I get a result like:<\/p>\n\n<pre><code>...\nFilename,Predictions\ntest\\green.1191.png,Green\ntest\\green.1195.png,Green\ntest\\green.1196.png,Green\ntest\\green.1197.png,Green\ntest\\green.1198.png,OK\ntest\\green.1199.png,Green\ntest\\green.1200.png,Green\ntest\\green.1201.png,Green\ntest\\green.1202.png,OK\ntest\\green.1203.png,Green\ntest\\green.1204.png,OK\ntest\\green.1205.png,Green\ntest\\green.1206.png,Green\ntest\\green.1207.png,Green\n...\n<\/code><\/pre>\n\n<p>But when I am trying to make a single image prediction with:<\/p>\n\n<pre><code># predicting images\ntest_image = image.load_img('dataset\/test_set\/test\/green.1230.png', target_size = (227, 454))\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis = 0)\nresult = model.predict(test_image, verbose=1)\ny_pred = np.argmax(result, axis = 1)\n<\/code><\/pre>\n\n<p>My y_pred is [6] and if I am reading class_indices correctly [6] is totaly other class (predict_generator() is doing well). <\/p>\n\n<p>Class indices:<\/p>\n\n<pre><code>class_names = (valid_generator.class_indices)\n\nclass_names = dict((v,k) for k,v in class_names.items())\nclass_names_list = []\ntemp = []\n\nfor key, value in class_names.items():\n    temp = value\n    class_names_list.append(temp)\n<\/code><\/pre>\n\n<p>Gives me:<\/p>\n\n<pre><code>{0: 'Green', 1: 'Half', 2: 'Moldy', 3: 'NoEmbryo', 4: 'OK', 5: 'Organic', 6: 'Stones'}\n<\/code><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"53231110","CreationDate":"2018-11-09T18:02:25.873","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>Your problem are rooted probably from python dict to list conversion.\nWhen your prediction goes from predict_generator(), then it chooses from a <strong>dict by key<\/strong>.<\/p>\n\n<p>In the second example you convert it to a list and uses the <strong>list's indices<\/strong>, which will be quite other result.<\/p>\n\n<p>I don't really get what is the purpose of:<\/p>\n\n<pre><code>for key, value in class_names.items():\n    temp = value\n    class_names_list.append(temp)\n<\/code><\/pre>\n\n<p>But if you get the class result from the class_names_list, you will get wrong results. So:<\/p>\n\n<pre><code>y_pred = np.argmax(result, axis = 1)\nclass_names[y_pred]\n<\/code><\/pre>\n\n<p>should give you the correct value.<\/p>\n"}]}
+{"QuestionId":53231168,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-11-09T18:06:56.767","AcceptedAnswerId":null,"OwnerUserId":3687776.0,"Title":"How to use keras.utils.multi_gpu_model with custom subclassed Keras models","Body":"<p>I am using the Keras Subclassing API to create a custom model. I tried to parallelize my model using:<\/p>\n\n<pre><code>parallel_model = keras.utils.multi_gpu_model(subclassed_model, gpus=num_gpus)\n<\/code><\/pre>\n\n<p>However I encountered the following error:<\/p>\n\n<pre><code>line 203, in multi_gpu_model\n    for i in range(len(model.outputs)):\nTypeError: object of type 'NoneType' has no len()\n<\/code><\/pre>\n\n<p>After, reading the docs in more detail, I noticed the following:<\/p>\n\n<blockquote>\n  <p>In subclassed models, the model's topology is defined as Python code(rather than as a static graph of layers). That means the model's topology cannot be inspected or serialized. As a result, the following methods and attributes are not available for subclassed models: <code>model.inputs<\/code> and <code>model.outputs<\/code>.<\/p>\n<\/blockquote>\n\n<p>So how can I train subclassed models on multiple GPUs? Have in mind that I am only interested in data parallelization.<\/p>\n","answers":[]}
+{"QuestionId":53231571,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-09T18:38:37.890","AcceptedAnswerId":null,"OwnerUserId":4931135.0,"Title":"What does flatten_parameters() do?","Body":"<p>I saw many Pytorch examples using flatten_parameters in the forward function of the RNN<\/p>\n\n<p><code>self.rnn.flatten_parameters()<\/code>  <\/p>\n\n<p>I saw this <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/rnn.html\" rel=\"noreferrer\">RNNBase<\/a> and it is written that it<\/p>\n\n<blockquote>\n  <p>Resets parameter data pointer so that they can use faster code paths<\/p>\n<\/blockquote>\n\n<p>What does that mean?<\/p>\n","answers":[{"AnswerId":"56042155","CreationDate":"2019-05-08T13:39:13.700","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>It may not be a full answer to your question. But, if you give a look at the <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/eb5d28ecefb9d78d4fff5fac099e70e5eb3fbe2e\/torch\/csrc\/api\/src\/nn\/modules\/rnn.cpp#L111-L129\" rel=\"nofollow noreferrer\"><code>flatten_parameters<\/code><\/a>'s  source code , you will notice that it calls <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/fbd690c1fec0651ee9e6cc07ddcb12217ffb31bc\/aten\/src\/ATen\/native\/cudnn\/RNN.cpp#L622-L674\" rel=\"nofollow noreferrer\"><code>_cudnn_rnn_flatten_weight<\/code><\/a> in<\/p>\n\n<pre><code>...\nNoGradGuard no_grad;\ntorch::_cudnn_rnn_flatten_weight(...)\n...\n<\/code><\/pre>\n\n<p>is the function that does the job. You will find that what it actually does is copying the model's weights into a <code>vector&lt;Tensor&gt;<\/code> (check the <code>params_arr<\/code> declaration) in: <\/p>\n\n<pre><code>  \/\/ Slice off views into weight_buf\n  std::vector&lt;Tensor&gt; params_arr;\n  size_t params_stride0;\n  std::tie(params_arr, params_stride0) = get_parameters(handle, rnn, rnn_desc, x_desc, w_desc, weight_buf);\n\n  MatrixRef&lt;Tensor&gt; weight{weight_arr, static_cast&lt;size_t&gt;(weight_stride0)},\n                    params{params_arr, params_stride0};\n<\/code><\/pre>\n\n<p>And the weights copying in <\/p>\n\n<pre><code>  \/\/ Copy weights\n  _copyParams(weight, params);\n<\/code><\/pre>\n\n<p>Also note that they update (or <code>Reset<\/code> as they explicitly say in docs) the original pointers  of <code>weights<\/code> with the new pointers of <code>params<\/code> by doing an in-place operation <code>.set_<\/code> (<code>_<\/code> is their notation for the in-place operations) in <code>orig_param.set_(new_param.view_as(orig_param));<\/code><\/p>\n\n<pre><code>  \/\/ Update the storage\n  for (size_t i = 0; i &lt; weight.size(0); i++) {\n    for (auto orig_param_it = weight[i].begin(), new_param_it = params[i].begin();\n         orig_param_it != weight[i].end() &amp;&amp; new_param_it != params[i].end();\n         orig_param_it++, new_param_it++) {\n      auto orig_param = *orig_param_it, new_param = *new_param_it;\n      orig_param.set_(new_param.view_as(orig_param));\n    }\n  }\n<\/code><\/pre>\n\n<p>And according to <a href=\"http:\/\/www.open-std.org\/jtc1\/sc22\/wg21\/docs\/papers\/2010\/n3092.pdf\" rel=\"nofollow noreferrer\">n2798 (draft of C++0x)<\/a><\/p>\n\n<p>\u00a9ISO\/IECN3092 <\/p>\n\n<blockquote>\n  <p>23.3.6  Class template vector<\/p>\n  \n  <p>A vector is a sequence container that supports random access iterators. In addition, it supports (amortized)constant time insert and erase operations at the end; insert and erase in the middle take linear time. Storage management is handled automatically, though hints can be given to improve efficiency. <strong>The elements of a vector are stored contiguously<\/strong>, meaning that if <code>v<\/code> is a vector <code>&lt;T, Allocator&gt;<\/code> where <code>T<\/code> is some type other than bool, then it obeys the <code>identity&amp;v[n] == &amp;v[0] + n<\/code> for all <code>0 &lt;= n &lt; v.size()<\/code>.<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>In some <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/2460\" rel=\"nofollow noreferrer\">situations<\/a><\/p>\n\n<blockquote>\n  <p>UserWarning: RNN module weights are not part of single contiguous chunk of memory. This means they need to be compacted at every call, possibly greately increasing memory usage. To compact weights again call <code>flatten_parameters()<\/code>.<\/p>\n<\/blockquote>\n\n<p>They explicitly advise people in code warnings to have a contiguous chunk of memory.<\/p>\n"}]}
+{"QuestionId":53231758,"AnswerCount":1,"Tags":"<python><tensorflow><keras><google-cloud-ml>","CreationDate":"2018-11-09T18:54:29.043","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Speeding up TFRecords feed into Keras model on CloudML for GPU","Body":"<p>I would like to feed TFRecords into my model at a super fast rate. However, currently, my GPU(Single K80 on GCP) is at 0% load which is super slow on <em>CloudML<\/em>. <\/p>\n\n<p>I have TFRecords in GCS: <code>train_directory = gs:\/\/bucket\/train\/*.tfrecord<\/code>,  (around 100 files of 30mb-800mb in size), but for some reason it struggles to feed the data into my model fast enough for GPU.<\/p>\n\n<p>Interestingly, loading data into memory and using numpy arrays using <code>fit_generator()<\/code> is 7x faster. There I can specify multi-processing and multi workers. <\/p>\n\n<p>My current set up parses tf records and loads an infinite <code>tf.Dataset<\/code>. Ideally, the solution would save\/prefecth some batches in memory, for the gpu to use on demand.<\/p>\n\n<pre><code>def _parse_func(record):\n    \"\"\" Parses TF Record\"\"\"\n    keys_to_features = {}\n    for _ in feature_list: # 300 features ['height', 'weights', 'salary']\n         keys_to_features[_] = tf.FixedLenFeature([TIME_STEPS], tf.float32)\n    parsed = tf.parse_single_example(record, keys_to_features)\n    t = [tf.manip.reshape(parsed[_], [-1, 1]) for _ in feature_list]\n    numeric_tensor = tf.concat(values=t, axis=1)\n\n    x = dict()\n    x['numeric'] = numeric_tensor\n    y = ...\n    w = ...\n\n    return x, y, w\n\ndef input_fn(file_pattern, b=BATCH_SIZE):\n    \"\"\"\n    :param file_pattern: GCS bucket to read from\n    :param b: Batch size, defaults to BATCH_SIZE in hparams.py\n    :return: And infinitely iterable data set using tf records of tf.data.Dataset class\n    \"\"\"\n    files = tf.data.Dataset.list_files(file_pattern=file_pattern)\n    d = files.apply(\n        tf.data.experimental.parallel_interleave(\n            lambda filename: tf.data.TFRecordDataset(filename),\n            cycle_length=4,\n            block_length=16,\n            buffer_output_elements=16,\n            prefetch_input_elements=16,\n            sloppy=True))\n    d = d.apply(tf.contrib.data.map_and_batch(\n        map_func=_parse_func, batch_size=b,\n        num_parallel_batches=4))\n    d = d.cache()\n    d = d.repeat()\n    d = d.prefetch(1)\n    return d\n<\/code><\/pre>\n\n<p>Get train data<\/p>\n\n<pre><code># get files from GCS bucket and load them into dataset\ntrain_data = input_fn(train_directory, b=BATCH_SIZE)\n<\/code><\/pre>\n\n<p>Fit the model<\/p>\n\n<pre><code>model.fit(x=train_data.make_one_shot_iterator())\n<\/code><\/pre>\n\n<p>I am running it on CloudML so GCS and CloudML should be pretty fast.<\/p>\n\n<hr>\n\n<h2>CloudML CPU Usage:<\/h2>\n\n<p>As we can see below, the CPU is at 70% and the memory doesn't increase past 10%. So what does the <code>dataset.cache()<\/code> do?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/e1L6z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/e1L6z.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h2>GPU metrics in CloudML logs<\/h2>\n\n<p>As seen below, it seems that the GPU is off! Also the memory is at 0mb. Where is the cache stored?<\/p>\n\n<p><strong>No processes running on GPU!<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/biMiH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/biMiH.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h3>Edit:<\/h3>\n\n<p>It seems that indeed, there are not processes running on GPU. I tried to explicitly state:<\/p>\n\n<pre><code>tf.keras.backend.set_session(tf.Session(config=tf.ConfigProto(\n    allow_soft_placement=True,\n    log_device_placement=True)))\n\ntrain_data = input_fn(file_pattern=train_directory, b=BATCH_SIZE)\n\nmodel = create_model()\n\nwith tf.device('\/gpu:0'):\n    model.fit(x=train_data.make_one_shot_iterator(),\n              epochs=EPOCHS,\n              steps_per_epoch=STEPS_PER_EPOCH,\n              validation_data=test_data.make_one_shot_iterator(),\n              validation_steps=VALIDATION_STEPS)\n<\/code><\/pre>\n\n<p>but everything still utilises the CPU!<\/p>\n","answers":[{"AnswerId":"53268293","CreationDate":"2018-11-12T18:46:25.127","ParentId":null,"OwnerUserId":"4729764","Title":null,"Body":"<p>In my case, I was using a custom <code>setup.py<\/code> file which used a CPU-only Tensorflow version. <\/p>\n\n<p>I am kicking myself, please install <code>tensorflow-gpu<\/code> instead.<\/p>\n"}]}
+{"QuestionId":53232159,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-09T19:28:36.273","AcceptedAnswerId":null,"OwnerUserId":4400904.0,"Title":"How to handle 0 batch size in tf.layers.max_pooling2d Tensorflow?","Body":"<p>I am using Tensorflow fold to write a model, and Tensorflow fold often times has 0 batch sizes (<a href=\"https:\/\/github.com\/tensorflow\/fold\/issues\/23\" rel=\"nofollow noreferrer\">Error on github<\/a>). This causes issues for certain Tensorflow operations, saying an error like so: <code>F tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:466] could not set cudnn tensor descriptor: CUDNN_STATUS_BAD_PARAM<\/code><\/p>\n\n<p>However, it is possible to solve this by writing a custom gradient operation, like described <a href=\"https:\/\/github.com\/ronghanghu\/n2nmn\/blob\/master\/util\/empty_safe_conv.py\" rel=\"nofollow noreferrer\">here<\/a> and below. <\/p>\n\n<pre><code># TensorFlow Fold can generate zero-size batch for conv layer\n# which will crash cuDNN on backward pass. So use this\n# for arbitrary convolution in modules to avoid the crash.\ndef _conv_safe(inputs, filters, kernel_size, strides, activation):\n    g = tf.get_default_graph()\n    with g.gradient_override_map({'Conv2D': 'Conv2D_handle_empty_batch'}):\n        return tf.layers.conv2d(inputs=inputs, filters=filters, kernel_size=kernel_size,strides=strides, activation=activation)\n\n@tf.RegisterGradient('Conv2D_handle_empty_batch')\ndef _Conv2DGrad(op, grad):\n    with tf.device('\/cpu:0'):\n        return [tf.nn.conv2d_backprop_input(\n                tf.shape(op.inputs[0]), op.inputs[1], grad, op.get_attr('strides'),\n                op.get_attr('padding'), op.get_attr('use_cudnn_on_gpu'),\n                op.get_attr('data_format')),\n                tf.nn.conv2d_backprop_filter(op.inputs[0],\n                                             tf.shape(op.inputs[1]), grad,\n                                             op.get_attr('strides'),\n                                             op.get_attr('padding'),\n                                             op.get_attr('use_cudnn_on_gpu'),\n                                             op.get_attr('data_format'))]\n<\/code><\/pre>\n\n<p>I am now wondering how I can do a similar thing to avoid this crash when using the <code>tf.layers.max_pooling2d<\/code> operation, or any other form of max pooling. You can see in the example for <code>tf.layers.conv2d<\/code>, we are able to get around it by custom implementing the gradient to handle the 0 batch size. How can I do this for <code>tf.layers.max_pooling2d<\/code>?<\/p>\n\n<p>Note: I am using Tensorflow 1.0 since that is what is supported by Tensorflow Fold.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53234550","CreationDate":"2018-11-09T23:28:27.533","ParentId":null,"OwnerUserId":"4400904","Title":null,"Body":"<p>I think we can do it like this:<\/p>\n\n<pre><code>from tensorflow.python.ops import gen_nn_ops\n\ndef max_pooling_zero_batch(inputs, pool_size, strides, name):\n\n    g = tf.get_default_graph()\n    with g.gradient_override_map({'MaxPool': 'MaxPool_handle_empty_batch'}):  \n        return tf.layers.max_pooling2d(inputs=inputs, pool_size=pool_size, strides=strides, name=name)\n\n@tf.RegisterGradient(\"MaxPool_handle_empty_batch\")\ndef _MaxPoolGrad(op, grad):\n    with tf.device('\/cpu:0'):\n        return gen_nn_ops._max_pool_grad(op.inputs[0], op.outputs[0], grad, op.get_attr(\"ksize\"), op.get_attr(\"strides\"), padding=op.get_attr(\"padding\"), data_format=op.get_attr(\"data_format\"))\n<\/code><\/pre>\n\n<p>It seems to work with 0 batch size.<\/p>\n"}]}
+{"QuestionId":53232171,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2018-11-09T19:29:26.180","AcceptedAnswerId":null,"OwnerUserId":4931135.0,"Title":"How to use DistributedDataParallel() and init_process_group() in a HPC?","Body":"<p>I will use HPC for my research and I don't know a lot about parallel or distributed computing. \nI really don't understand the <code>DistributedDataParallel()<\/code> in pytorch. Especially <code>init_process_group()<\/code>.\nWhat is the meaning of initializing processes group? and what is<\/p>\n\n<blockquote>\n  <p>init_method : URL specifying how to initialize the package.    <\/p>\n<\/blockquote>\n\n<p>for example (I found those in the documentation):<br>\n<code>'tcp:\/\/10.1.1.20:23456'<\/code> or <code>'file:\/\/\/mnt\/nfs\/sharedfile'<\/code><br>\nWhat are those URLs?<\/p>\n\n<p>What is the Rank of the current process?<br>\nIs <code>world_size<\/code> the number of GPUs?   <\/p>\n\n<p>It would be really appreciated if someone explained to me what is and How to use <code>DistributedDataParallel()<\/code> and <code>init_process_group()<\/code> because I don't know parallel or distributed computing.<\/p>\n\n<p>I will use things like Slurm(sbatch) in the HPC.<\/p>\n","answers":[]}
+{"QuestionId":53232617,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-11-09T20:05:47.673","AcceptedAnswerId":null,"OwnerUserId":125571.0,"Title":"conversion error while converting tf saved model to tflite","Body":"<p>I am trying to convert a saved model tf to a tflite mode. I am getting following error in tensorflow c++ while conversion:-<\/p>\n\n<pre><code>tensorflow\/contrib\/lite\/toco\/tooling_util.cc:981] Check failed: name.substr(colon_pos + 1).find_first_not_of(\"0123456789\") == string::npos (1 vs. 18446744073709551615)Array name must only have digits after colon\\n'\n<\/code><\/pre>\n\n<p>Here is the source code :-<\/p>\n\n<pre><code>def convert_to_tflite_util(model_dir, output_dir):\n    tag_constants = set([\"train\"])\n    converter = tf.contrib.lite.TocoConverter.from_saved_model(model_dir, tag_set=tag_constants)\n    #converter = tf.contrib.lite.TocoConverter.from_saved_model(model_dir)\n    tflite_model = converter.convert()\n<\/code><\/pre>\n\n<p>Logs :-<\/p>\n\n<pre><code>2018-11-09 11:44:11.410085: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nINFO:tensorflow:Restoring parameters from \/Users\/deosingh\/code\/mldev\/sensei-on-device\/converters\/deeplasso_converter\/model\/variables\/variables\nINFO:tensorflow:The given SavedModel MetaGraphDef contains SignatureDefs with the following keys: {'serving_default'}\nINFO:tensorflow:input tensors info: \nINFO:tensorflow:Tensor's key in saved_model's tensor_map: input\nINFO:tensorflow: tensor name: data:0, shape: (-1, 320, 320, 4), type: DT_FLOAT\nINFO:tensorflow:output tensors info: \nINFO:tensorflow:Tensor's key in saved_model's tensor_map: output\nINFO:tensorflow: tensor name: prob:0, shape: (1, 320, 320, 2), type: DT_FLOAT\nINFO:tensorflow:Restoring parameters from \/Users\/deosingh\/code\/mldev\/model\/variables\/variables\nINFO:tensorflow:Froze 387 variables.\nINFO:tensorflow:Converted 387 variables to const ops.\nTraceback (most recent call last):\n  File \"\/Applications\/PyCharm CE.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1664, in &lt;module&gt;\n    main()\n  File \"\/Applications\/PyCharm CE.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1658, in main\n    globals = debugger.run(setup['file'], None, None, is_module)\n  File \"\/Applications\/PyCharm CE.app\/Contents\/helpers\/pydev\/pydevd.py\", line 1068, in run\n    pydev_imports.execfile(file, globals, locals)  # execute the script\n  File \"\/Applications\/PyCharm CE.app\/Contents\/helpers\/pydev\/_pydev_imps\/_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"\/Users\/deosingh\/code\/mldev\/src\/tf_to_tflite.py\", line 112, in &lt;module&gt;\n    main()\n  File \"\/Users\/deosingh\/code\/mldev\/src\/tf_to_tflite.py\", line 108, in main\n    convert_to_tflite_util(args.src_dir, args.dest_dir)\n  File \"\/Users\/deosingh\/code\/mldev\/src\/tf_to_tflite.py\", line 91, in convert_to_tflite_util\n    tflite_model = converter.convert()\n  File \"\/Users\/deosingh\/code\/mldev\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", line 439, in convert\n    **converter_kwargs)\n  File \"\/Users\/deosingh\/code\/mldev\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 309, in toco_convert_impl\n    input_data.SerializeToString())\n  File \"\/Users\/deosingh\/code\/mldev\/venv\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/convert.py\", line 109, in toco_convert_protos\n    (stdout, stderr))\nRuntimeError: TOCO failed see console for info.\nb'2018-11-09 11:44:18.328373: F tensorflow\/contrib\/lite\/toco\/tooling_util.cc:981] Check failed: name.substr(colon_pos + 1).find_first_not_of(\"0123456789\") == string::npos (1 vs. 18446744073709551615)Array name must only have digits after colon\\n'\nNone\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c19e29306ce1777456b2dbb3a14f511edf7883a8\/tensorflow\/contrib\/lite\/toco\/tooling_util.cc#L980\" rel=\"nofollow noreferrer\">Link to tensorflow source<\/a> which is throwing this error.<\/p>\n\n<p>The original model was in caffe , then was converted to tensorflow using <a href=\"https:\/\/github.com\/Microsoft\/MMdnn\" rel=\"nofollow noreferrer\">mmdnn<\/a> . Converter is complaining about name not being in the correct format i.e doesn't expect _ after : as per the tf source. Please let me know how to resolve\/debug this error. Since the tf source which isthrowing this error is in c++  and I am invoking it through python I am not able to debug it through pycharm on macosx.<\/p>\n","answers":[]}
+{"QuestionId":53232969,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-09T20:39:43.400","AcceptedAnswerId":53241303.0,"OwnerUserId":7571462.0,"Title":"How to monitor validation loss in the training of estimators in TensorFlow?","Body":"<p>I want to ask a question about how to monitor validation loss in the training process of estimators in TensorFlow. I have checked a similar question (<a href=\"https:\/\/stackoverflow.com\/questions\/45417502\/validation-during-training-of-estimator\">validation during training of Estimator<\/a>) asked before, but it did not help much.<\/p>\n\n<p>If I use estimators to build a model, I will give an input function to the Estimator.train() function. But there is no way to add another validation_x, and validation_y data in the training process. Therefore, when the training started, I can only see the training loss. The training loss is expected to decrease when the training process running longer. However, this information is not helpful to prevent overfitting. The more valuable information is validation loss. Usually, the validation loss is the U-shape with the number of epochs. To prevent overfitting, we want to find the number of epochs that the validation loss is minimum.<\/p>\n\n<p>So this is my problem. How can I get validation loss for each epoch in the training process of using estimators?<\/p>\n","answers":[{"AnswerId":"53241303","CreationDate":"2018-11-10T17:03:21.000","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>You need to create a validation input_fn and either use estimator.train() and estimator.evaluate() alternatively or simpy use tf.estimator.train_and_evaluate()<\/p>\n\n<pre><code>x = ...\ny = ...\n\n...\n\n# For example, if x and y are numpy arrays &lt; 2 GB\ntrain_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))\nval_dataset = tf.data.Dataset.from_tensor_slices((x_val_, y_val))\n\n...\n\nestimator = ...\n\nfor epoch in n_epochs:\n    estimator.train(input_fn = train_dataset)\n    estimator.evaluate(input_fn = val_dataset)\n<\/code><\/pre>\n\n<p>estimator.evaluate() will compute the loss and any other metrics that are defined in your model_fn and will save the events in a new \"eval\" directory inside your job_dir.<\/p>\n"}]}
+{"QuestionId":53233123,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-09T20:52:32.527","AcceptedAnswerId":null,"OwnerUserId":10630478.0,"Title":"tf.data: Combining multiple from_generator() datasets to create batches padded across time windows","Body":"<p>I am working on a timeseries problem where each timeseries is fairly long (10^3-10^4 timesteps, and each timeseries is of different length).<\/p>\n\n<p>For each sequence, I can define a Python generator that yields values one timestep at a time. I am using the <code>tf.data.Dataset.from_generator()<\/code> constructor to wrap these generators into the tf.data API. The documentation suggests using <code>from_generator()<\/code> along with the <code>tf.contrib.data.parallel_interleave()<\/code> transformation to parallelize the extraction from my Python generators.<\/p>\n\n<p>My downstream use for these data is a stateful RNN (e.g. LSTM or GRU). I want to chunk up the timeseries into smaller (~10^2) windows and use each chunk as a training example (i.e., truncated BPTT). Since my data are streaming, I think that means saving up <code>window_size<\/code> timesteps of each generator before passing it on through the pipeline, to be batched with the other generators' data. I also want to save the RNN state across these chunks so I can still learn long-term dependencies.<\/p>\n\n<p>My issue comes with wanting to create padded batches of these generators' batched outputs. Ideally, I would want to present to my neural network windows of the generator outputs, with padding as necessary when some subset of the generators exhaust themselves before others. I know that if I consume the entire generator output for each generator, then use <code>Dataset.padded_batch()<\/code> I can do this (and can then slice the padded batch across the time dimension into windowed chunks as necessary). However, I want to pass each window to the neural network it becomes available. If one of the generators exhausts itself before the others, I want to pad it with the padding value until all others have, so I can reset the RNN state and begin the next batch of generators with an empty initial RNN state. I am stuck here because the dataset resulting from <code>tf.contrib.data.parallel_interleave()<\/code> transformation discards each generator when it becomes exhausted, and the timeseries do not maintain a consistent ordering across samples from it.<\/p>\n\n<p>Here is a small example:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef stepwise_generator(length):\n    for i in range(length):\n        yield i\n\nlengths = list(range(1,10,2)) # [1, 3, 5, 7, 9]\n\nwindow_length = 4\nbatch_size = 3\n\ndataset = tf.data.Dataset.from_tensor_slices(lengths)\n\ngen = lambda length: tf.data.Dataset.from_generator(\n    stepwise_generator, tf.float32, output_shapes=[], args=(length,)\n).batch(window_length) # this batching saves window_length timesteps per generator\n\ndataset = dataset.apply(\n    tf.contrib.data.parallel_interleave(gen, cycle_length=batch_size)\n)\n\ndataset = dataset.padded_batch(batch_size, (-1,), np.inf)\n# batching 3 generators at once, and padding exhausted ones with inf.\n# using a batch_size value no more than cycle_length above means we\n# shouldn't start a new generator mid-batch (i think)\n\niterator = dataset.make_one_shot_iterator()\ntensor = iterator.get_next()\n\nouts = []\nwith tf.Session() as sess:\n    while True:\n        try:\n            out = sess.run(tensor)\n            outs.append(out)\n        except tf.errors.OutOfRangeError:\n            break\n\nprint(np.asarray(outs))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[[ 0. inf inf inf]   # batch 1\n  [ 0.  1.  2. inf]\n  [ 0.  1.  2.  3.]]\n\n [[ 4. inf inf inf]   # batch 2 - the generator in index -1 in the\n  [ 0.  1.  2.  3.]   # previous batch gets cycled to index 0 and two\n  [ 0.  1.  2.  3.]]  # new generators are initiated\n\n [[ 4.  5.  6. inf]   # batch 3 - more generator cycling, and the one in\n  [ 4.  5.  6.  7.]   # index 1 also gets cycled to index 2 in the same\n  [ 8. inf inf inf]]] # batch (because we have run out of generators in\n                      # parallel_interleave)\n<\/code><\/pre>\n\n<p>My desired output would be something like<\/p>\n\n<pre><code>[[[ 0. inf inf inf]   # batch 1\n  [ 0.  1.  2. inf]\n  [ 0.  1.  2.  3.]]\n\n [[inf]               # batch 2 - the leftover timestep from a padded \n  [inf]               # batch of the first 3 generators\n  [4. ]]\n\n [[ 0.  1.  2.  3.]   # batch 3 - only two generators are left so this is \n  [ 0.  1.  2.  3.]]  # an end-of-epoch smaller batch\n\n [[ 4.  5.  6. inf]   # batch 4\n  [ 4.  5.  6.  7.]]\n\n [[inf]               # batch 5\n  [ 8.]]]\n<\/code><\/pre>\n\n<p>Here, the internal states of the RNNs would be reset after batch 2 and 5.<\/p>\n\n<p>Again, the desired output can be simple to create if I consume the entirety of each generator's output, then pad, batch, and slice, but I want to produce batches as the generators, which may be each receiving data in real-time from e.g. a separate simulation, make them available.<\/p>\n","answers":[]}
+{"QuestionId":53233457,"AnswerCount":0,"Tags":"<python><linux><tensorflow>","CreationDate":"2018-11-09T21:25:13.610","AcceptedAnswerId":null,"OwnerUserId":1531248.0,"Title":"Failed call to cuInit CUDA_ERROR_NOT_INITIALIZED (Device mapping: no known devices) Tensorflow-gpu 1.2.1","Body":"<p>I have 2 GPU's but not able to utilize via tensorflow-GPU, could anyone help to figure it out the issue.<\/p>\n\n<p><strong>Versions:<\/strong><\/p>\n\n<ul>\n<li>Tensorflow GPU 1.2.1 <\/li>\n<li>Cudnn 6.0.21 <\/li>\n<li><p>Cuda 10.0 (toolkit)<\/p>\n\n<p><code>import tensorflow as tf\n a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\n b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\n c = tf.matmul(a, b)\n with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:\n   print (sess.run(c))<\/code><\/p><\/li>\n<\/ul>\n\n<h2>Warning\/Error message:<\/h2>\n\n<pre><code>2018-11-09 12:06:42.827220: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:406] **failed call to cuInit: CUDA_ERROR_NOT_INITIALIZED**\n2018-11-09 12:06:42.827292: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:158] retrieving CUDA diagnostic information for host: ca47dppaia001\n2018-11-09 12:06:42.827309: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:165] hostname: ca47dppaia001\n2018-11-09 12:06:42.827370: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:189] libcuda reported version is: 410.72.0\n2018-11-09 12:06:42.827419: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:369] driver version file contents: \"\"\"NVRM version: NVIDIA UNIX ppc64le Kernel Module  410.72  Wed Oct 17 20:19:50 CDT 2018\nGCC version:  gcc version 4.8.5 20150623 (Red Hat 4.8.5-36) (GCC)\n\"\"\"\n2018-11-09 12:06:42.827457: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:193] kernel reported version is: 410.72.0\n2018-11-09 12:06:42.827471: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:300] kernel version seems to match DSO: 410.72.0\n**Device mapping: no known devices.**\n2018-11-09 12:06:42.871454: I tensorflow\/core\/common_runtime\/direct_session.cc:265] Device mapping:\n<\/code><\/pre>\n\n<hr>\n","answers":[]}
+{"QuestionId":53233633,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2018-11-09T21:46:03.210","AcceptedAnswerId":53234381.0,"OwnerUserId":7784797.0,"Title":"2 Layer Neural Network Does not Converge","Body":"<h1>Background<\/h1>\n\n<p>I am a newbie to TensorFlow and I am trying to understand the basics of deep learning. I started from writing a two-layer neural network from scratch and it achieved 89% accuracy on MNIST dataset and now I am trying to implement the same network in TensorFlow and compare their performance. <\/p>\n\n<h1>Problem<\/h1>\n\n<p>I am not sure if I miss something basic in the code, but the following implementation seems to be unable to update weights and therefore could not output anything meaningful.<\/p>\n\n<pre><code>num_hidden = 100\n# x -&gt; (batch_size, 784)\nx = tf.placeholder(tf.float32, [None, 784])\n\nW1 = tf.Variable(tf.zeros((784, num_hidden)))\nb1 = tf.Variable(tf.zeros((1, num_hidden)))\nW2 = tf.Variable(tf.zeros((num_hidden, 10)))\nb2 = tf.Variable(tf.zeros((1, 10)))\n# z -&gt; (batch_size, num_hidden)\nz = tf.nn.relu(tf.matmul(x, W1) + b1)\n# y -&gt; (batch_size, 10)\ny = tf.nn.softmax(tf.matmul(z, W2) + b2)\n\n# y_ -&gt; (batch_size, 10)\ny_ =  tf.placeholder(tf.float32, [None, 10])\n# y_ * tf.log(y) -&gt; (batch_size, 10)\ncross_entropy =  -tf.reduce_sum(y_ * tf.log(y+1e-10))\ntrain_step = tf.train.GradientDescentOptimizer(0.05).minimize(cross_entropy)\nsess = tf.InteractiveSession()\ntf.global_variables_initializer().run()\n# tf.argmax(y, axis=1) returns the maximum index in each row\ncorrect_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\nfor epoch in range(1000):\n    # batch_xs -&gt; (100, 784)\n    # batch_ys -&gt; (100, 10), one-hot encoded\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n    train_data = {x: batch_xs, y_: batch_ys}\n    sess.run(train_step, feed_dict=train_data)\nprint(sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels}))\nW1_e, b1_e, W2_e, b2_e = W1.eval(), b1.eval(), W2.eval(), b2.eval()\nsess.close()\n<\/code><\/pre>\n\n<h1>What I Have Done<\/h1>\n\n<p>I checked many the official docs and many other implementations, but I feel totally confused since they may use different versions and API varies greatly.<\/p>\n\n<p>So could someone help me, thank you in advance.<\/p>\n","answers":[{"AnswerId":"53234381","CreationDate":"2018-11-09T23:05:24.287","ParentId":null,"OwnerUserId":"1803124","Title":null,"Body":"<p>There are two problems with what you have done so far. First, you have initialised all of the weights to zero, which will prevent the network from learning. And secondly, the learning rate was too high. The below code got me 0.9665 accuracy. For why not to set all the weights to zero you can see <a href=\"https:\/\/stackoverflow.com\/questions\/20027598\/why-should-weights-of-neural-networks-be-initialized-to-random-numbers\">here<\/a> .<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n\n\nnum_hidden = 100\n\n# x -&gt; (batch_size, 784)\nx = tf.placeholder(tf.float32, [None, 784])\nlabel_place = tf.placeholder(tf.float32, [None, 10])\n\n\n# WONT WORK as EVERYTHING IS ZERO!\n# # Get accuracy at chance \\approx 0.1\n# W1 = tf.Variable(tf.zeros((784, num_hidden)))\n# b1 = tf.Variable(tf.zeros((1, num_hidden)))\n# W2 = tf.Variable(tf.zeros((num_hidden, 10)))\n# b2 = tf.Variable(tf.zeros((1, 10)))\n\n# Will work, you will need to train a bit more than 1000 steps\n# though\nW1 = tf.Variable(tf.random_normal((784, num_hidden), 0., 0.1))\nb1 = tf.Variable(tf.zeros((1, num_hidden)))\nW2 = tf.Variable(tf.random_normal((num_hidden, 10), 0, 0.1))\nb2 = tf.Variable(tf.zeros((1, 10)))\n\n# network, we only go as far as the linear output after the hidden layer\n# so we can feed it into the tf.nn.softmax_cross_entropy_with_logits below\n# this is more numerically stable\nz = tf.nn.relu(tf.matmul(x, W1) + b1)\nlogits = tf.matmul(z, W2) + b2\n\n# define our loss etc as before. however note that the learning rate is lower as\n# with a higher learning rate it wasnt really working\ncross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=label_place, logits=logits)\ntrain_step = tf.train.GradientDescentOptimizer(.001).minimize(cross_entropy)\n\n# continue as before\nsess = tf.InteractiveSession()\ntf.global_variables_initializer().run()\ncorrect_prediction = tf.equal(tf.argmax(tf.nn.softmax(logits), 1), tf.argmax(label_place, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\nfor epoch in range(5000):\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n    train_data = {x: batch_xs, label_place: batch_ys}\n    sess.run(train_step, feed_dict=train_data)\nprint(sess.run(accuracy, feed_dict={x: mnist.test.images, label_place: mnist.test.labels}))\nW1_e, b1_e, W2_e, b2_e = W1.eval(), b1.eval(), W2.eval(), b2.eval()\nsess.close()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53233782,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-hub>","CreationDate":"2018-11-09T21:59:09.123","AcceptedAnswerId":null,"OwnerUserId":6334174.0,"Title":"Problems with tensorflow hub: Table not initialized","Body":"<p>I am trying to use tf_hub for universal-sentence-encoder-large when I have the following problem:<\/p>\n\n<pre><code>FailedPreconditionError (see above for traceback): Table not initialized.\n<\/code><\/pre>\n\n<p>It seems that TensorFlow thinks I did not run init op, but actually, I have run the init op:<\/p>\n\n<pre><code>embed = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-large\/3\")\nembeddings = embed([\n\"The quick brown fox jumps over the lazy dog.\"])    \ninit = tf.global_variables_initializer()\n\n\nwith tf.Session() as sess:\n    sess.run(init)\n    embeddings = sess.run(embeddings)\n\n    print(embeddings)\n<\/code><\/pre>\n\n<p>The same code structure has been fine with other tf_hub models like <code>elmo<\/code>.<\/p>\n","answers":[{"AnswerId":"53240680","CreationDate":"2018-11-10T15:55:54.713","ParentId":null,"OwnerUserId":"6334174","Title":null,"Body":"<p>Looks like to use this tensorflow hub, I need to run an addtional initializer:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\ntable_init = tf.tables_initializer()\n\nwith tf.Session() as sess:\n    sess.run([init, table_init])\n    embeddings_ = sess.run(embeddings)\n\n    print(embeddings)\n<\/code><\/pre>\n"},{"AnswerId":"56495855","CreationDate":"2019-06-07T14:04:25.537","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>You could try<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.train.SingularMonitoredSession() as sess:\n  ...\n<\/code><\/pre>\n\n<p>which does all standard initializations by itself (including \"shared resources\", for which there was no public API last time I checked).<\/p>\n"}]}
+{"QuestionId":53234062,"AnswerCount":1,"Tags":"<tensorflow><batch-normalization>","CreationDate":"2018-11-09T22:26:03.427","AcceptedAnswerId":null,"OwnerUserId":10453487.0,"Title":"Problem with batch normalization in tensorflow","Body":"<p>I am having trouble understanding the implementation of batch normalization in Tensorflow. To illustrate, I have created  a simple network with one input node, one hidden node, and one output node and run with 1 batch, with a batch size of 2.  My input x consists of a scalar with 2 values (ie a batch size of 2), one set to 0 and other set to 1.<\/p>\n\n<p>I run for one epoch, and write out the output from the hidden layer (before and after batch normalization) as well as the batch norm moving mean, variance, gamma, and beta. <\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimport numpy as np\n\nN_HIDDEN_1 = 1\nN_INPUT= 1\nN_OUTPUT = 1\n\n###########################################################\n\n# DEFINE THE Network\n\n# Define placeholders for data that will be fed in during execution\nx = tf.placeholder(tf.float32, (None, N_INPUT))\ny = tf.placeholder(tf.float32, (None, N_OUTPUT))\nlx = tf.placeholder(tf.float32, [])\ntraining = tf.placeholder_with_default(False, shape=(), name='training')\n\n# Hidden layers with relu activation\nwith tf.variable_scope('hidden1'):\n      hidden_1 = tf.layers.dense(x, N_HIDDEN_1, activation=None, use_bias=False)\n      bn_1 = tf.layers.batch_normalization(hidden_1, training=training, momentum=0.5)\n      bn_1x = tf.nn.relu(bn_1)\n\n# Output layer\nwith tf.variable_scope('output'):\n      predx = tf.layers.dense(bn_1x, N_OUTPUT, activation=None, use_bias=False)\n      pred = tf.layers.batch_normalization(predx, training=training, momentum=0.5)\n\n###########################################################\n\n# Define the cost function that is optimized when\n# training the network and the optimizer\n\ncost = tf.reduce_mean(tf.square(pred-y))\n\noptimizer = tf.train.AdamOptimizer(learning_rate=lx).minimize(cost)\n\nbout1 = tf.global_variables('hidden1\/batch_normalization\/moving_mean:0')\nbout2 = tf.global_variables('hidden1\/batch_normalization\/moving_variance:0')\nbout3 = tf.global_variables('hidden1\/batch_normalization\/gamma:0')\nbout4 = tf.global_variables('hidden1\/batch_normalization\/beta:0')\n\n###########################################################\n\n# Train network\n\ninit = tf.global_variables_initializer()\nextra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\nwith tf.Session() as sess:\n\n    sess.run(init)\n\n    # Create dummy data\n    batchx = np.zeros((2,1))\n    batchy = np.zeros((2,1))\n    batchx[0,0]=0.0\n    batchx[1,0]=1.0\n    batchy[0,0]=3.0\n    batchy[1,0]=4.0\n\n    _,_ = sess.run([optimizer, extra_update_ops], feed_dict={training: True, x:batchx, y:batchy, lx: 0.001})\n\n    print('weight of hidden layer')\n    W1 = np.array(sess.run(tf.global_variables('hidden1\/dense\/kernel:0')))\n    W1x = np.sum(W1, axis=1)\n    print(W1x)\n\n    print()\n    print('output from hidden layer, batch norm layer, and relu layer')\n    hid1,b1,b1x = sess.run([hidden_1, bn_1, bn_1x], feed_dict={training: False, x:batchx})\n    print('hidden_1', hid1)\n    print('bn_1', b1)\n    print('bn_1x', b1x)\n\n    print()\n    print('batchnorm parameters')\n    print('moving mean', sess.run(bout1))\n    print('moving variance', sess.run(bout2))\n    print('gamma', sess.run(bout3))\n    print('beta', sess.run(bout4))\n<\/code><\/pre>\n\n<p>Here is the output I get when I run the code:<\/p>\n\n<pre><code>weight of hidden layer [[1.404974]]\n\noutput from hidden layer, batch norm layer, and relu layer\nhidden_1 [[0.      ]\n          [1.404974]]\n\nbn_1 [[-0.40697935]\n      [ 1.215785  ]]\n\nbn_1x [[0.      ]\n      [1.215785]]\n\nbatchnorm parameters\nmoving mean [array([0.3514931], dtype=float32)]\nmoving variance [array([0.74709475], dtype=float32)]\ngamma [array([0.999], dtype=float32)]\nbeta [array([-0.001], dtype=float32)]\n<\/code><\/pre>\n\n<p>I am puzzled by the resulting batchnorm parameters. In this particular case, the output from the hidden layer prior to applying the batch norm\nare the scalars 0 and 1.404974. But the batch norm parameter moving mean is 0.3514931. This is for the case where I use momentum = 0.5. It is not clear to me why the moving mean after 1 iteration is not exactly the average of 0 and 1.404974 in this case. I was under the impression that the momentum parameter would only kick in from the second batch on.<\/p>\n\n<p>Any help would be much appreciated.<\/p>\n","answers":[{"AnswerId":"53241106","CreationDate":"2018-11-10T16:42:08.757","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>Because you ran the optimizer, it's hard to know what is really happening inside: the hidden_1 values you are printing are not the ones that were used to update the batch norm statistics; they are the post update values.<\/p>\n\n<p>Anyways, I don't really see the issue:<\/p>\n\n<pre><code>Moving mean original value = 0.0\nbatch mean value = (1.404974 - 0.0) \/ 2.0 = ~0.7\nMoving mean value = momentum * Moving mean original value + (1 - momentum) * batch mean value\n                  = 0.0 * 0.5 + (1 - 0.5) * 0.7\n                  = 0.35\n\nMoving variance original value = 1.0\nbatch variance value = ~0.5\nMoving variance value = momentum * Moving variance original value + (1 - momentum) * batch variance value\n                  = 1.0 * 0.5 + (1.0 - 0.5) * 0.5\n                  = 0.75\n<\/code><\/pre>\n"}]}
+{"QuestionId":53234585,"AnswerCount":1,"Tags":"<keras><tensorflow-datasets>","CreationDate":"2018-11-09T23:32:24.060","AcceptedAnswerId":null,"OwnerUserId":4095551.0,"Title":"adding and accessing auxiliary tf.Dataset attributes with Keras","Body":"<p>I use a <code>tf.py_func<\/code> call to parse data (features, labels and sample_weights) from file to a <code>tf.Dataset<\/code>:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((records, labels, sample_weights))    \ndataset = dataset.map(\n   lambda filename, label, sample_weight: tuple(tf.py_func(\n     self._my_parse_function, [filename, label, sample_weights], [tf.float32, label.dtype, tf.float32])))\n<\/code><\/pre>\n\n<p>The data is variable-length 1-D sequences, so I also pad the sequences to a fixed length in <code>my_parse_function<\/code>.  <\/p>\n\n<p>I use <code>tensorflow.python.keras.models.Sequential.fit(...)<\/code> to train the data (which now accepts datasets as input, including datasets with sample_weights) and <code>tensorflow.python.keras.models.Sequential.predict<\/code> to predict outputs.<\/p>\n\n<p>Once I have predictions I would like to do some post-processing to make sense of the outputs. For example, I'd like to truncate the padded data to the actual sequence length. Also, I'd like to know for sure which file the data came from, since I am not sure that ordering is guaranteed with dataset iterators, especially if batching is used (I do batch the dataset as well) or multi-GPU or multi-workers are involved (I hope to try the multi- scenarios). Even if order was 'guaranteed' this is a decent sanity check.<\/p>\n\n<p>This information, filename (i.e, a string) and sequence length (i.e, an integer), is not currently conveniently accessible, so I'd like to add these two attributes to the dataset elements and be able to retrieve them during\/after the call to predict.<\/p>\n\n<p>What is the best approach to do this?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53325786","CreationDate":"2018-11-15T18:28:11.323","ParentId":null,"OwnerUserId":"4095551","Title":null,"Body":"<p>As a workaround, I store this auxiliary information in a 'global' dictionary in <code>my_parse_fn<\/code>, so it stores (and re-stores) on every iteration through the <code>tf.Dataset<\/code>.  This is ok for now since there are only about 1000 examples in the training set, so storing 1000 strings and integers is not a problem.  But if this auxiliary information were larger or the training set were larger, this approach would not be very scalable. In my case, the input data for each training example is significantly large, about 50MB in size, which is why reading a <code>tf.Dataset<\/code> from file (i.e., on every epoch) is important. <\/p>\n\n<p>I still think that it would be helpful to be able to more conveniently extend a <code>tf.Dataset<\/code> with this information.  Also I noticed that when I adding a field to a <code>tf.Dataset<\/code> like dataset.tag to identify, say, dataset.tag = 'training', dataset.tag ='validation' or dataset.tag = 'test' sets, the field did not survive the iterations of training.  <\/p>\n\n<p>So again in this case I'm wondering how a <code>tf.Dataset<\/code> can be extended.<\/p>\n\n<p>On the other question, it looks like the order of <code>tf.Dataset<\/code> elements <strong>is<\/strong> respected through iterations, so predictions, say, from <code>tensorflow.python.keras.models.Sequential.predict(...)<\/code> are ordered as the file ids were presented to <code>my_parse_fn<\/code> (at least batching respects this ordering, but I still don't know about whether a multi-GPU scenario would as well).<\/p>\n\n<p>Thanks for any insights.<\/p>\n"}]}
+{"QuestionId":53235029,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-10T00:47:11.027","AcceptedAnswerId":53238098.0,"OwnerUserId":10479259.0,"Title":"Issue of batch sizes when using custom loss functions in Keras","Body":"<p>I am doing a slight modification of a standard neural network by defining a custom loss function. The custom loss function depends not only on y_true and y_pred, but also on the training data. I implemented it using the wrapping solution described <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2121\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Specifically, I wanted to define a custom loss function that is the standard mse plus the mse between the input and the square of y_pred:<\/p>\n\n<pre><code>def custom_loss(x_true)\n    def loss(y_true, y_pred):\n        return K.mean(K.square(y_pred - y_true) + K.square(y_true - x_true))\n    return loss\n<\/code><\/pre>\n\n<p>Then I compile the model using<\/p>\n\n<pre><code>model_custom.compile(loss = custom_loss( x_true=training_data ), optimizer='adam')\n<\/code><\/pre>\n\n<p>fit the model using<\/p>\n\n<pre><code>model_custom.fit(training_data, training_label, epochs=100, batch_size = training_data.shape[0])\n<\/code><\/pre>\n\n<p>All of the above works fine, because the batch size is actually the number of all the training samples.<\/p>\n\n<p>But if I set a different batch_size (e.g., 10) when I have 1000 training samples, there will be an error<\/p>\n\n<blockquote>\n  <p>Incompatible shapes: [1000] vs. [10].<\/p>\n<\/blockquote>\n\n<p>It seems that Keras is able to automatically adjust the size of the inputs to its own loss function base on the batch size, but cannot do so for the custom loss function.<\/p>\n\n<p>Do you know how to solve this issue?<\/p>\n\n<p>Thank you!<\/p>\n\n<p>==========================================================================<\/p>\n\n<h1>* Update: the batch size issue is solved, but another issue occurred<\/h1>\n\n<p>Thank you, Ori, for the suggestion of concatenating the input and output layers! It \"worked\", in the sense that the codes can run under any batch size. However, it seems that the result from training the new model is wrong... Below is a simplified version of the codes to demonstrate the problem:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.io\nimport keras\nfrom keras import backend as K\nfrom keras.models import Model\nfrom keras.layers import Input, Dense, Activation\nfrom numpy.random import seed\nfrom tensorflow import set_random_seed\n\ndef custom_loss(y_true, y_pred): # this is essentially the mean_square_error\n    mse = K.mean( K.square( y_pred[:,2] - y_true ) )\n    return mse\n\n# set the seeds so that we get the same initialization across different trials\nseed_numpy = 0\nseed_tensorflow = 0\n\n# generate data of x = [ y^3 y^2 ]\ny = np.random.rand(5000+1000,1) * 2 # generate 5000 training and 1000 testing samples\nx = np.concatenate( ( np.power(y, 3) , np.power(y, 2) ) , axis=1 )\n\ntraining_data  = x[0:5000:1,:]\ntraining_label = y[0:5000:1]\ntesting_data   = x[5000:6000:1,:]\ntesting_label  = y[5000:6000:1]\n\n# build the standard neural network with one hidden layer\nseed(seed_numpy)\nset_random_seed(seed_tensorflow)\n\ninput_standard = Input(shape=(2,))                                               # input\nhidden_standard = Dense(10, activation='relu', input_shape=(2,))(input_standard) # hidden layer\noutput_standard = Dense(1, activation='linear')(hidden_standard)                 # output layer\n\nmodel_standard = Model(inputs=[input_standard], outputs=[output_standard])     # build the model\nmodel_standard.compile(loss='mean_squared_error', optimizer='adam')            # compile the model\nmodel_standard.fit(training_data, training_label, epochs=50, batch_size = 500) # train the model\ntesting_label_pred_standard = model_standard.predict(testing_data)             # make prediction\n\n# get the mean squared error\nmse_standard = np.sum( np.power( testing_label_pred_standard - testing_label , 2 ) ) \/ 1000\n\n# build the neural network with the custom loss\nseed(seed_numpy)\nset_random_seed(seed_tensorflow)\n\ninput_custom = Input(shape=(2,))                                             # input\nhidden_custom = Dense(10, activation='relu', input_shape=(2,))(input_custom) # hidden layer\noutput_custom_temp = Dense(1, activation='linear')(hidden_custom)            # output layer\noutput_custom = keras.layers.concatenate([input_custom, output_custom_temp])\n\nmodel_custom = Model(inputs=[input_custom], outputs=[output_custom])         # build the model\nmodel_custom.compile(loss = custom_loss, optimizer='adam')                   # compile the model\nmodel_custom.fit(training_data, training_label, epochs=50, batch_size = 500) # train the model\ntesting_label_pred_custom = model_custom.predict(testing_data)               # make prediction\n\n# get the mean squared error\nmse_custom = np.sum( np.power( testing_label_pred_custom[:,2:3:1] - testing_label , 2 ) ) \/ 1000\n\n# compare the result\nprint( [ mse_standard , mse_custom ] )\n<\/code><\/pre>\n\n<p>Basically, I have a standard one-hidden-layer neural network, and a custom one-hidden-layer neural network whose output layer is concatenated with the input layer. For testing purpose, I did not use the concatenated input layer in the custom loss function, because I wanted to see if the custom network can reproduce the standard neural network. Since the custom loss function is equivalent to the standard 'mean_squared_error' loss, both networks should have the same training results (I also reset the random seeds to make sure that they have the same initialization).<\/p>\n\n<p>However, the training results are very different. It seems that the concatenation makes the training process different? Any ideas?<\/p>\n\n<p>Thank you again for all your help!<\/p>\n\n<p><strong>Final update: Ori's approach of concatenating input and output layers works, and is verified by using the generator. Thanks!!<\/strong><\/p>\n","answers":[{"AnswerId":"53238098","CreationDate":"2018-11-10T10:37:20.560","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>The problem is that when compiling the model, you set x_true to be a static tensor, in the size of all the samples. While the input for keras loss functions are the y_true and y_pred, where each of them is of size <code>[batch_size, :]<\/code>. <br><\/p>\n\n<p>As I see it there are 2 options you can solve this, the first one is using a generator for creating the batches, in such a way that you will have control over which indices are evaluated each time, and at the loss function you could slice the <code>x_true<\/code> tensor to fit the samples being evaluated: <br><\/p>\n\n<pre><code>def custom_loss(x_true)\n    def loss(y_true, y_pred):\n        x_true_samples = relevant_samples(x_true)\n        return K.mean(K.square(y_pred - y_true) + K.square(y_true - x_true_samples))\n    return loss\n<\/code><\/pre>\n\n<p>This solution can be complicated, what I would suggest is a simpler workaround - <br>\nConcatenate the input layer with the output layer, such that your new output will be of the form <code>original_output , input<\/code>. <br><br>\nNow you can use a new modified loss function:<\/p>\n\n<pre><code>def loss(y_true, y_pred):\n    return K.mean(K.square(y_pred[:,:output_shape] - y_true[:,:output_shape]) +\n                  K.square(y_true[:,:output_shape] - y_pred[:,outputshape:))\n<\/code><\/pre>\n\n<p>Now your new loss function will take in account both the input data, and the prediction. <br><br><\/p>\n\n<p><strong>Edit:<\/strong><br>\nNote that while you set the seed, your models are not exactly the same, and as you did not use a generator, you let keras choose the batches, and for different models he might pick different samples. <br>\nAs your model does not converge, different samples can lead to different results. <br><br><\/p>\n\n<p>I added a generator to your code, to verify the samples we pick for training, now you can see both results are the same: <br><\/p>\n\n<pre><code>def custom_loss(y_true, y_pred): # this is essentially the mean_square_error\n    mse = keras.losses.mean_squared_error(y_true, y_pred[:,2])\n    return mse\n\n\ndef generator(x, y, batch_size):\n    curIndex = 0\n    batch_x = np.zeros((batch_size,2))\n    batch_y = np.zeros((batch_size,1))\n    while True:\n        for i in range(batch_size):            \n            batch_x[i] = x[curIndex,:]\n            batch_y[i] = y[curIndex,:]\n            i += 1;\n            if i == 5000:\n                i = 0\n        yield batch_x, batch_y\n\n# set the seeds so that we get the same initialization across different trials\nseed_numpy = 0\nseed_tensorflow = 0\n\n# generate data of x = [ y^3 y^2 ]\ny = np.random.rand(5000+1000,1) * 2 # generate 5000 training and 1000 testing samples\nx = np.concatenate( ( np.power(y, 3) , np.power(y, 2) ) , axis=1 )\n\ntraining_data  = x[0:5000:1,:]\ntraining_label = y[0:5000:1]\ntesting_data   = x[5000:6000:1,:]\ntesting_label  = y[5000:6000:1]\n\nbatch_size = 32\n\n\n\n# build the standard neural network with one hidden layer\nseed(seed_numpy)\nset_random_seed(seed_tensorflow)\n\ninput_standard = Input(shape=(2,))                                               # input\nhidden_standard = Dense(10, activation='relu', input_shape=(2,))(input_standard) # hidden layer\noutput_standard = Dense(1, activation='linear')(hidden_standard)                 # output layer\n\nmodel_standard = Model(inputs=[input_standard], outputs=[output_standard])     # build the model\nmodel_standard.compile(loss='mse', optimizer='adam')            # compile the model\n#model_standard.fit(training_data, training_label, epochs=50, batch_size = 10) # train the model\nmodel_standard.fit_generator(generator(training_data,training_label,batch_size),  steps_per_epoch= 32, epochs= 100)\ntesting_label_pred_standard = model_standard.predict(testing_data)             # make prediction\n\n# get the mean squared error\nmse_standard = np.sum( np.power( testing_label_pred_standard - testing_label , 2 ) ) \/ 1000\n\n# build the neural network with the custom loss\nseed(seed_numpy)\nset_random_seed(seed_tensorflow)\n\n\ninput_custom = Input(shape=(2,))                                               # input\nhidden_custom = Dense(10, activation='relu', input_shape=(2,))(input_custom) # hidden layer\noutput_custom_temp = Dense(1, activation='linear')(hidden_custom)            # output layer\noutput_custom = keras.layers.concatenate([input_custom, output_custom_temp])\n\nmodel_custom = Model(inputs=input_custom, outputs=output_custom)         # build the model\nmodel_custom.compile(loss = custom_loss, optimizer='adam')                   # compile the model\n#model_custom.fit(training_data, training_label, epochs=50, batch_size = 10) # train the model\nmodel_custom.fit_generator(generator(training_data,training_label,batch_size),  steps_per_epoch= 32, epochs= 100)\ntesting_label_pred_custom = model_custom.predict(testing_data)\n\n# get the mean squared error\nmse_custom = np.sum( np.power( testing_label_pred_custom[:,2:3:1] - testing_label , 2 ) ) \/ 1000\n\n# compare the result\nprint( [ mse_standard , mse_custom ] )\n<\/code><\/pre>\n"}]}
+{"QuestionId":53235045,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-10T00:50:10.827","AcceptedAnswerId":53235092.0,"OwnerUserId":7353970.0,"Title":"Decoding a rank 5 tensor from tfrecords yields \"CopyElementToLargerSlice Unhandled rank: 5\"","Body":"<p>I am currently trying to decode a sequence feature of type <code>tf.string()<\/code> with the length <code>FixedLenSequenceFeature<\/code> of shape (None, None, 120, 160, 7) in TensorFlow version 1.4.0. I am using padded batches and <code>tf.parse_single_sequence_example()<\/code> but when sampling batches from <code>tf.data.TFRecordDataset<\/code>, I get the following error:<\/p>\n\n<pre><code>Unimplemented: CopyElementToLargerSlice Unhandled rank: 5\n<\/code><\/pre>\n\n<p>and further:<\/p>\n\n<pre><code>[[Node: IteratorGetNext = IteratorGetNext[output_shapes=[[?,?,120,160,3], [?], [?], [?,?,3], [?,?,120,160,3], [?], [?], [?,?,?,120,160,7], [?,?,?,3], [?,?,?,3], [?,?,120,160]], output_types=[DT_INT16, DT_INT64, DT_INT64, DT_DOUBLE, DT_INT16, DT_INT64, DT_INT64, DT_INT16, DT_DOUBLE, DT_DOUBLE, DT_INT16], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Iterator)]]\n<\/code><\/pre>\n\n<p>Handling rank 5 tensors in this way doesn't seem implemented. When can this be expected to be implemented?<\/p>\n\n<p>Thank you <\/p>\n","answers":[{"AnswerId":"53235092","CreationDate":"2018-11-10T00:56:57.953","ParentId":null,"OwnerUserId":"7353970","Title":null,"Body":"<p>It turns out, updating from TensorFlow <code>version 1.4.0<\/code> to <code>version 1.12.0<\/code> solved the issue.<\/p>\n"}]}
+{"QuestionId":53235440,"AnswerCount":1,"Tags":"<deep-learning><pytorch>","CreationDate":"2018-11-10T02:10:28.003","AcceptedAnswerId":53238726.0,"OwnerUserId":7784797.0,"Title":"Two-Layer Neural Network in PyTorch does not Converge","Body":"<h1>Problem<\/h1>\n\n<p>I am trying to implement 2-layer neural network using different methods (TensorFlow, PyTorch and from scratch) and then compare their performance based on MNIST dataset.<\/p>\n\n<p>I am not sure what mistakes I have made, but the accuracy in PyTorch is only about 10%, which is basically random guess. I think probably the weights does not get updated at all. <\/p>\n\n<p>Note that I intentionally use the dataset provided by TensorFlow to keep the data I use through 3 different methods consistent for accurate comparison. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.examples.tutorials.mnist import input_data\nimport torch\n\nclass Net(torch.nn.Module):\n    def __init__(self):\n      super(Net, self).__init__()\n      self.fc1 =  torch.nn.Linear(784, 100)\n      self.fc2 =  torch.nn.Linear(100, 10)\n\n    def forward(self, x):\n      # x -&gt; (batch_size, 784)\n      x = torch.relu(x)\n      # x -&gt; (batch_size, 10)\n      x = torch.softmax(x, dim=1)\n      return x\n\nnet = Net()\nnet.zero_grad()\nLoss = torch.nn.CrossEntropyLoss()\noptimizer =  torch.optim.SGD(net.parameters(), lr=0.01)\n\nfor epoch in range(1000):  # loop over the dataset multiple times\n\n    batch_xs, batch_ys = mnist_m.train.next_batch(100)\n    # convert to appropriate settins\n    # note the input to the linear layer should be (n_sample, n_features)\n    batch_xs = torch.tensor(batch_xs, requires_grad=True)\n    # batch_ys -&gt; (batch_size,)\n    batch_ys = torch.tensor(batch_ys, dtype=torch.int64)\n\n    # forward\n    # output -&gt; (batch_size, 10)\n    output = net(batch_xs)\n    # result -&gt; (batch_size,)\n    result = torch.argmax(output, dim=1)\n    loss = Loss(output, batch_ys)\n\n    # backward\n    optimizer.zero_grad()\n    loss.backward()\n    optimizer.step()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53238726","CreationDate":"2018-11-10T12:00:42.723","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The problem here is that you don't apply your fully connected layers <code>fc1<\/code> and <code>fc2<\/code>.<\/p>\n\n<p>Your <code>forward()<\/code> currently looks like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def forward(self, x):\n    # x -&gt; (batch_size, 784)\n    x = torch.relu(x)\n    # x -&gt; (batch_size, 10)\n    x = torch.softmax(x, dim=1)\n    return x\n<\/code><\/pre>\n\n<p>So if you change it to:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def forward(self, x):\n    # x -&gt; (batch_size, 784)\n    x = self.fc1(x)             # added layer fc1\n    x = torch.relu(x)  \n\n    # x -&gt; (batch_size, 10)\n    x = self.fc2(x)             # added layer fc2\n    x = torch.softmax(x, dim=1)\n    return x\n<\/code><\/pre>\n\n<p>It should work.<\/p>\n\n<p><em>Regarding Umang Guptas answer: As I see it, calling <code>zero_grad()<\/code> before calling <code>backward()<\/code> as Mr.Robot did, is just fine. This shouldn't be a problem.<\/em>\n<br><\/p>\n\n<hr>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>So I did a short test - I set iterations from <code>1000<\/code> to <code>10000<\/code> to see the bigger picture if it is really decreasing. <em>( Of course I also loaded the data to <code>mnist_m<\/code> as this wasn't included in the code you've posted )<\/em><\/p>\n\n<p>I added a print condition to the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if epoch % 1000 == 0:\n    print('Epoch', epoch, '- Loss:', round(loss.item(), 3))\n<\/code><\/pre>\n\n<p>Which prints out the loss every <code>1000<\/code> iterations:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 0 - Loss: 2.305\nEpoch 1000 - Loss: 2.263\nEpoch 2000 - Loss: 2.187\nEpoch 3000 - Loss: 2.024\nEpoch 4000 - Loss: 1.819\nEpoch 5000 - Loss: 1.699\nEpoch 6000 - Loss: 1.699\nEpoch 7000 - Loss: 1.656\nEpoch 8000 - Loss: 1.675\nEpoch 9000 - Loss: 1.659\n<\/code><\/pre>\n\n<p><em>Tested with PyTorch version 0.4.1<\/em><\/p>\n\n<p>So you can see that with the changed <code>forward()<\/code> the network is learning now, the rest of the code I left untouched. <\/p>\n\n<p>Good luck further!<\/p>\n"}]}
+{"QuestionId":53235485,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-10T02:19:47.357","AcceptedAnswerId":null,"OwnerUserId":3161192.0,"Title":"I want to use Tensorflow on CPU for everything except back propagation","Body":"<p>I recently built my first TensorFlow model (converted from hand-coded python). I'm using tensorflow-gpu, but I only want to use GPU for backprop during training. For everything else I want to use CPU. I've seen <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">this article<\/a> showing how to force CPU use on a system that will use GPU by default. However, you have to specify every single operation where you want to force CPU use. Instead I'd like to do the opposite. I'd like to default to CPU use, but then specify GPU just for the backprop that I do during training. Is there a way to do that?<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>Looks like things are just going to run slower over tensorflow because of how my model and scenario are built at present. I tried using a different environment that just uses regular (non-gpu) tensorflow, and it still runs significantly slower than hand-coded python. The reason for this, I suspect, is it's a reinforcement learning model that plays checkers (see below) and makes one single forward prop \"prediction\" at a time as it plays against a computer opponent. At the time I designed the architecture, that made sense. But it's not very efficient to do predictions one at a time, and less so with whatever overhead there is for tensorflow.<\/p>\n\n<p>So, now I'm thinking that I'm going to need to change the game playing architecture to play, say, a thousand games simultaneously and run a thousand forward prop moves in a batch. But, man, changing the architecture now is going to be tricky at best.<\/p>\n","answers":[{"AnswerId":"53325362","CreationDate":"2018-11-15T17:57:19.620","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>TensorFlow lets you control device placement with the <code>tf.device<\/code> context manager.<\/p>\n\n<p>So for example to run some code on the CPU do<\/p>\n\n<pre><code>with tf.device('cpu:0'):\n  &lt;your code goes here&gt;\n<\/code><\/pre>\n\n<p>Similarly to force GPU usage.<\/p>\n\n<p>Instead of always running your forward pass on the CPU though you're better off making two graphs: a forward-only cpu-only graph to be used when rolling out the policy and a gpu-only forward-and-backward graph to be used when training.<\/p>\n"}]}
+{"QuestionId":53235490,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-10T02:20:43.287","AcceptedAnswerId":null,"OwnerUserId":10631590.0,"Title":"Bugs of getting gradient of loss function with respect to input using Keras","Body":"<pre><code>model = keras.Sequential([         #my model, trained on Fashion_MNIST\n    keras.layers.Conv2D(32, (3,3), activation=tf.nn.relu, input_shape=(28,28,1)),\n    keras.layers.MaxPooling2D(),\n    keras.layers.Conv2D(64, (3,3), activation=tf.nn.relu),\n    keras.layers.MaxPooling2D(),\n    keras.layers.Flatten(),\n    keras.layers.Dense(1000, activation='relu'),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.load_weights('.\/weights\/235model')            #load weights\nmodel.compile(optimizer=tf.train.AdamOptimizer(),   \n          loss='sparse_categorical_crossentropy',\n          metrics=['accuracy'])\n\ntmp = np.full((10,1),0.0)                           #get correct label\ntmp[test_labels[0]] = 1.0\ny_tensor=tf.convert_to_tensor(tmp)\n\nloss = keras.losses.sparse_categorical_crossentropy(model.output,    y_tensor)\ngradients = keras.backend.gradients(loss, model.input)\nprint(gradients)\n<\/code><\/pre>\n\n<p>I am doing an iterative FGSM attack, so I need to get the gradient of the loss with respect to the input, but when I print gradients, it shows\n[None] instead of a valid tensor.<\/p>\n\n<p>Later I run it,<\/p>\n\n<pre><code>with keras.backend.get_session() as sess:\n\n    evaluated_gradients = sess.run(gradients, feed_dict={model.input:np.reshape(test_images[0],(1,28,28,1))})\n<\/code><\/pre>\n\n<p>It gives me error \n\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 258, in for_fetch\n    type(fetch)))\nTypeError: Fetch argument None has invalid type <\/p>\n\n<p>If I change loss to model.output in setting the gradients, it works ok, but if I use the crossentropy loss there, it doesn't work, could someone helps me? I am a beginner to keras and tensorflow.<\/p>\n","answers":[{"AnswerId":"53290643","CreationDate":"2018-11-13T22:45:09.467","ParentId":null,"OwnerUserId":"6780025","Title":null,"Body":"<p>You flipped the order of arguments to <code>keras.losses.sparse_categorical_crossentropy<\/code>. It should be <\/p>\n\n<pre><code>loss = keras.losses.sparse_categorical_crossentropy(y_tensor, model.output)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/keras.io\/losses\/#sparse_categorical_crossentropy\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/losses\/#sparse_categorical_crossentropy<\/a><\/p>\n"}]}
+{"QuestionId":53236430,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-10T06:04:09.783","AcceptedAnswerId":null,"OwnerUserId":1867083.0,"Title":"How to add custom summaries to tensorboard when training with tf.keras.Model.fit","Body":"<p>I'm training a model as:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n        with tf.Session(config=tf.ConfigProto(allow_soft_placement = True)) as sess:\n                K.set_session(sess)\n                tf.train.create_global_step()\n                #with tf.device('\/gpu:0:'):\n                m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init,\n                        nhidden, embed_dim, dropout, train_emb,\n                        char_dim, use_feat, gating_fn, words).build_network()\n                m.compile(optimizer=tf.train.AdamOptimizer(0.01),\n                          loss=tf.keras.losses.categorical_crossentropy,\n                          metrics=[tf.keras.metrics.categorical_accuracy])\n                tensorboard = TensorBoardCustom(log_dir=\"logs\", sess=sess)\n                m.fit_generator(generator=batch_loader_train, steps_per_epoch=len(batch_loader_train.batch_pool), epochs=100, callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>and I defined a custom callback extending the keras.callbacks.Tensorboard as:<\/p>\n\n<pre><code>class TensorBoardCustom(TensorBoard):\n\n    def __init__(self, log_dir, sess, **kwargs):\n        super(TensorBoardCustom, self).__init__(log_dir, **kwargs)\n        self.sess = sess\n\n    def on_batch_end(self, batch, logs={}):\n        summary = tf.summary.merge_all()\n        writer = tf.summary.FileWriter(self.log_dir)\n        s = self.sess.run(summary)\n        writer.add_summary(s, batch)\n        writer.close()\n        super(TensorBoardCustom, self).on_batch_end(batch, logs)\n<\/code><\/pre>\n\n<p>and I'm adding a new summary as:<\/p>\n\n<pre><code>l_docin = tf.keras.layers.Input(shape=(None,))\nwith tf.name_scope('summaries'):\n            table = tf.contrib.lookup.index_to_string_table_from_tensor(\n                    self.mapping_string, default_value=\"UNKNOWN\")\n            words = table.lookup(tf.cast(l_qin, tf.int64))\n            text = tf.reduce_join(words, 1, separator=' ')\n            tf.summary.text('text', text)\n<\/code><\/pre>\n\n<p>However, this is not working and I'm getting the following error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_2' with dtype float and shape [?,?]\n     [[{{node input_2}} = Placeholder[dtype=DT_FLOAT, shape=[?,?], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>Can someone explain why this is happening and how I can correct it? Is there a simpler\/better way of adding custom summaries?<\/p>\n","answers":[]}
+{"QuestionId":53236537,"AnswerCount":2,"Tags":"<keras><nlp><embedding>","CreationDate":"2018-11-10T06:28:43.523","AcceptedAnswerId":null,"OwnerUserId":7383245.0,"Title":"keras , Incompatible shapes: [64] vs. [64,280]","Body":"<p>I am doing a simple binary classifier , but something wrong happened, and why ?\nnote: I have override the related layers for supporting mask_zero\nthe details are following...\n my model is following:<\/p>\n\n<pre><code>def mymodel_v1(conf_dt):\n    input_ = Input(shape=(conf_dt['sent_len'],), dtype='int32')\n    embedding = Embedding(input_dim=conf_dt['vocab_size']+1, output_dim=conf_dt['embed_size'], input_length=conf_dt['sent_len'], mask_zero=True)(input_)\n    flat = FlattenWithMasking()(embedding)\n    dropout = Dropout(rate=conf_dt['dropout'])(flat)\n    dense = Dense(300)(dropout)\n    out = Dense(1, activation='sigmoid')(dense)\n\n    model = Model(inputs=input_, outputs=out)\n    print model.summary()\n    return model\n<\/code><\/pre>\n\n<p>and the train step is :<\/p>\n\n<pre><code>def train(train_fp, test_fp, conf_dt):\n    train_X, train_y, test_X, test_y, _ = load_data(train_fp, test_fp)\n    train_X = pad_sequences(train_X, maxlen=conf_dt['sent_len'], padding='post', truncating='post')\n    test_X = pad_sequences(test_X, maxlen=conf_dt['sent_len'], padding='post', truncating='post')\n    train_y = np.array(train_y, ndmin=2)\n    test_y = np.array(test_y, ndmin=2)\n    print 'data load and preprocess done'\n    print 'train_X.shape: ', train_X.shape\n    print 'train_y.shape: ', train_y.shape\n    sys.stdout.flush()\n\n    model = mymodel_v1(conf_dt)\n    model.compile(optimizer='rmsprop', loss='binary_crossentropy')\n\n    model.fit(train_X, train_y, batch_size=64, nb_epoch=2, verbose=2)\n    print model.summary()\n\n    model.evaluate(test_X, test_y)\n<\/code><\/pre>\n\n<p>the config.dict is:<\/p>\n\n<pre><code>conf_dt = {'vocab_size': 200000, 'dropout': 0.3, 'sent_len': 280, 'embed_size': 50}\n<\/code><\/pre>\n\n<p>the output of shape and model.summary is:<\/p>\n\n<pre><code>train_X.shape:  (116389, 280)\ntrain_y.shape:  (116389, 1)\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, 280)               0         \n_________________________________________________________________\nembedding_1 (Embedding)      (None, 280, 50)           10000050  \n_________________________________________________________________\nflatten_with_masking_1 (Flat (None, 14000)             0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 14000)             0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 300)               4200300   \n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 301       \n=================================================================\nTotal params: 14,200,651\nTrainable params: 14,200,651\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>the traceback is following:<\/p>\n\n<pre><code>      File \"train.py\", line 31, in train\n    model.fit(train_X, train_y, batch_size=64, nb_epoch=2, verbose=2)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1598, in fit\n    validation_steps=validation_steps)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1183, in _fit_loop\n    outs = f(ins_batch)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2273, in __call__\n    **self.session_kwargs)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 895, in run\n    run_metadata_ptr)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1124, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1321, in _do_run\n    options, run_metadata)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1340, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [64] vs. [64,280]\n         [[Node: loss\/dense_2_loss\/mul = Mul[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](loss\/dense_2_loss\/Mean, loss\/dense_2_loss\/Cast)]]\n\nCaused by op u'loss\/dense_2_loss\/mul', defined at:\n  File \"train.py\", line 49, in &lt;module&gt;\n    train(train_pt, test_pt, conf_dt)\n  File \"train.py\", line 29, in train\n    model.compile(optimizer='rmsprop', loss='binary_crossentropy')\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 850, in compile\n    sample_weight, mask)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 455, in weighted\n    score_array *= mask\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 865, in binary_op_wrapper\n    return func(x, y, name=name)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 1088, in _mul_dispatch\n    return gen_math_ops._mul(x, y, name=name)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\", line 1449, in _mul\n    result = _op_def_lib.apply_op(\"Mul\", x=x, y=y, name=name)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 767, in apply_op\n    op_def=op_def)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2630, in create_op\n    original_op=self._default_original_op, op_def=op_def)\n  File \"\/home\/homework\/.jumbo\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1204, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): Incompatible shapes: [64] vs. [64,280]\n         [[Node: loss\/dense_2_loss\/mul = Mul[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](loss\/dense_2_loss\/Mean, loss\/dense_2_loss\/Cast)]]\n<\/code><\/pre>\n\n<p>why is this happened? can somebody help me? ty.<\/p>\n","answers":[{"AnswerId":"53258241","CreationDate":"2018-11-12T08:23:50.700","ParentId":null,"OwnerUserId":"7383245","Title":null,"Body":"<p>and I read some keras source code, the layers of <code>Dense<\/code> and <code>Dropout<\/code> are <code>support the masking<\/code> by default, and the <code>FlattenWithMasking<\/code> which I override is following:<\/p>\n\n<pre><code>class FlattenWithMasking(Flatten):\n        def __init__(self, **kwargs):\n                self.supports_masking = True\n                super(FlattenWithMasking, self).__init__(**kwargs)\n\n        def compute_mask(self, x, mask):\n                return mask    \n<\/code><\/pre>\n"},{"AnswerId":"55573295","CreationDate":"2019-04-08T12:16:48.973","ParentId":null,"OwnerUserId":"6434566","Title":null,"Body":"<p>First, why your dense2 layer output only a shape 1? This sounds weird to me generally the output is at least 2 as the model rank outputs with a level of confidence. <\/p>\n\n<p>The shape 64 which does not fit with the [280,64]\u00a0you should at least have an something like [1,64] to do so you can use :<\/p>\n\n<pre><code> tf.convert_to_tensor()\n<\/code><\/pre>\n\n<p>(you should multiply matrices and not matrix with a variable):<\/p>\n"}]}
+{"QuestionId":53237161,"AnswerCount":3,"Tags":"<python><dataset><python-imaging-library><pytorch>","CreationDate":"2018-11-10T08:10:35.540","AcceptedAnswerId":null,"OwnerUserId":9312750.0,"Title":"Error in Google Colaboratory - AttributeError: module 'PIL.Image' has no attribute 'register_decoder'","Body":"<p>I am running this code on Google Colaboratory and I am getting error of register decoder<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>image_data = dset.ImageFolder(root=\"drive\/SemanticDataset\/train\/\", transform = transforms.Compose([\n                                        transforms.Scale(size=img_size),\n                                        transforms.CenterCrop(size=(img_size,img_size*2)),\n                                        transforms.ToTensor(),\n                                        ]))\nenter code herelabel_data = dset.ImageFolder(root=\"drive\/SemanticDataset\/label\/\", transform = transforms.Compose([\n                                        transforms.Scale(size=img_size),\n                                        transforms.CenterCrop(size=(img_size,img_size*2)),\n                                        transforms.ToTensor(),\n                                        ]))\n\nimage_batch = data.DataLoader(image_data, batch_size=batch_size, shuffle=False, num_workers=2)\nlabel_batch = data.DataLoader(label_data, batch_size=batch_size, shuffle=False, num_workers=2)\nfor i in range(epoch):\n   for _, (image, label) in enumerate(zip(image_batch, label_batch)):\n       optimizer.zero_grad()\n\n       x = Variable(image, requires_grad=True).cuda()\n       y = Variable(label).cuda()\n\n       out = model.forward(x)\n       loss = loss_func(out, y)\n\n       loss.backward()\n       optimizer.step()\n\n       if _ % 100 == 0:\n           print(\"Epoch: \"+i+\"| Loss: \" , loss)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Zb53v.png\" rel=\"noreferrer\">here is the screenshot of error<\/a><\/p>\n","answers":[{"AnswerId":"53245643","CreationDate":"2018-11-11T03:41:11.687","ParentId":null,"OwnerUserId":"4786728","Title":null,"Body":"<p>First, check the version of pillow you have by using:<\/p>\n\n<pre><code>import PIL\nprint(PIL.PILLOW_VERSION)\n<\/code><\/pre>\n\n<p>and make sure you have the newest version, the one I am using right now is <code>5.3.0<\/code>\nIf you have like <code>4.0.0<\/code>, install a new version by using:\n<code>!pip install Pillow==5.3.0<\/code> in the Colab environment.<\/p>\n\n<p>Second, <strong>restart<\/strong> your Google colab environment, and check the version again, it should be updated.<\/p>\n\n<p>I had the same problem, and I spent some time trying to solve it. <\/p>\n\n<p><strong>Note<\/strong>: I was using PyTorch 0.4. <\/p>\n\n<p>I hope this will solve your problem.<\/p>\n"},{"AnswerId":"53578342","CreationDate":"2018-12-02T07:39:04.847","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>I'd recommend using:<\/p>\n\n<pre><code>!pip install -U pillow\n<\/code><\/pre>\n\n<p>The runtime needs to be restarted after the upgrade.<\/p>\n\n<p>The <code>-U<\/code> will ensure that <code>pillow<\/code> is only installed if there is a newer version available, which will save time the 2nd time the cell is run after the kernel restart.<\/p>\n"},{"AnswerId":"53893326","CreationDate":"2018-12-22T05:37:33.727","ParentId":null,"OwnerUserId":"6042903","Title":null,"Body":"<p>I tried this:<\/p>\n\n<pre><code>!pip uninstall -y Pillow\n# install the new one\n!pip install Pillow==5.3.0\n# import the new one\nimport PIL\nprint(PIL.PILLOW_VERSION)\n<\/code><\/pre>\n\n<p>and it printed <code>5.3.0<\/code> and I resarted runtime enviorment via <code>Menu &gt; Runtime &gt; Restart Runtime<\/code>. It worked like a charm for me.<\/p>\n"}]}
+{"QuestionId":53237598,"AnswerCount":1,"Tags":"<python><tensorflow><matrix-multiplication><tensor>","CreationDate":"2018-11-10T09:25:56.213","AcceptedAnswerId":null,"OwnerUserId":10336700.0,"Title":"Tensorflow matmul how to deal with channel=None","Body":"<p>I am trying to do matmul of two tensors, one is [None, 4, 256], and the other is its transpose (I use tf.transpose and got a tensor [256, 4, None]). The expected result is a [4, 4] tensor. When I use matmul, it returns an error. I am wondering, how can I get the expected result? Thank you!<\/p>\n\n<pre><code>#inputs: [None, 4, 256]\ninputs_transpose = tf.transpose(inputs, perm = [0, 2, 1]) #[None, 256, 4]\ntemp_weights = tf.matmul(inputs, inputs_transpose) #[4, 4]expected\n<\/code><\/pre>\n","answers":[{"AnswerId":"53238131","CreationDate":"2018-11-10T10:42:30.107","ParentId":null,"OwnerUserId":"10618540","Title":null,"Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/convert_to_tensor\" rel=\"nofollow noreferrer\">Tensowflow docs<\/a><\/p>\n\n<p>\"Note: This function diverges from default Numpy behavior for float and string types when None is present in a Python list or scalar. Rather than silently converting None values, an error will be thrown.\"<\/p>\n\n<p>Tensowflow does not accept None values while creating tensors.\nI tried to recreate the setup you had on my system and got this error message:<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class 'list'&gt; to Tensor. Contents: [None, 4, 256]. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>You would need to explicitly handle the None by replacing it with a 0 before attempting to use the python list.<\/p>\n"}]}
+{"QuestionId":53238019,"AnswerCount":1,"Tags":"<c++><tensorflow><build><bazel><debug-symbols>","CreationDate":"2018-11-10T10:26:38.333","AcceptedAnswerId":null,"OwnerUserId":3609835.0,"Title":"How do I create a debug build of a recent Tensorflow version with CUDA Support?","Body":"<p>I tried and tried to create a debug build for a recent version of Tensorflow , using the official docker images (latest-cuda-devel-py3 -> r1.12.0) but nothing seems to work. Has someone recently created a successful debug build for Tensorflow (>= r1.11.0) and can share his approach ?<\/p>\n\n<p>This is what I tried so far.<\/p>\n\n<p>I basically tried to follow the instructions at <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/source<\/a>, but tried to modify them to generate a debug build. Nothing I tried resulted in a successful build.<\/p>\n\n<p>The Host System is a Linux x86-64 machine with lots of RAM (e.g. 512 GB of RAM -> DGX-1). The CUDA Version within the Docker-Image is CUDA-9.0. The recent \"latest\" Tensorflow Version which is inside the docker image is r1.12.0<\/p>\n\n<p>In order to get any cuda-build working, I needed to use \"nvidia-docker\", otherwise I get a linker error with \"libcuda.so.1\".<\/p>\n\n<p>I started like this:<\/p>\n\n<pre><code>nvidia-docker pull tensorflow\/tensorflow:latest-devel-gpu-py3\nnvidia-docker run --runtime=nvidia -it -w \/tensorflow -v $PWD:\/mnt -e HOST_PERMS=\"$(id -u):$(id -g)\" \\\n    tensorflow\/tensorflow:latest-devel-gpu-py3 bash\n<\/code><\/pre>\n\n<p>Then I tried to configure the project using<\/p>\n\n<pre><code>cd \/tensorflow\n.\/configure\n<\/code><\/pre>\n\n<p>I tried various configs. I tried keeping all values at their defaults. I tried enabling only the parts which I need. I tried not running .\/configure at all. I pointed it to my own cuda-9.0 and tensorrt installtion. But not running .\/configure at all (in the docker image) seems to produce best results (e.g. I can do optimized builds successfully with least effort).<\/p>\n\n<p>If I build it using the exact official build instructions, i.e. creating an <strong>optimized\/non-debug<\/strong> build, everything works as expected. So running the following seems to succeed.<\/p>\n\n<pre><code>bazel build --config=opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>Same thing, if I run the following, which includes debug info, but does not turn off optimization (e.g. I cannot really use this for debug purposes).<\/p>\n\n<pre><code>bazel build --config cuda --strip=never -c opt --copt=\"-ggdb\"  \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>But everything which disables optimizations does not seem to work. If I run the following (with or without the --strip=never flag )<\/p>\n\n<pre><code>bazel build --config cuda --strip=never -c dbg\n\/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>I arrive at the following error:<\/p>\n\n<blockquote>\n  <p>INFO: From Compiling\n  tensorflow\/contrib\/framework\/kernels\/zero_initializer_op_gpu.cu.cc:\n  external\/com_google_absl\/absl\/strings\/string_view.h(496): error:\n  constexpr function return is non-constant<\/p>\n<\/blockquote>\n\n<p>Which can be resolved by defining -DNDEBUG (see <a href=\"https:\/\/stackoverflow.com\/questions\/52665441\/nvcc-error-string-view-h-constexpr-function-return-is-non-constant\">nvcc error: string_view.h: constexpr function return is non-constant<\/a> ). <\/p>\n\n<p>But If I run the following:<\/p>\n\n<pre><code>bazel build --config cuda --strip=never -c dbg --copt=\"-DNDEBUG\"  \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>I get these linking errors at the final step of the build:<\/p>\n\n<blockquote>\n  <p>ERROR:\n  \/tensorflow\/python\/BUILD:3865:1:\n  Linking of rule '\/\/tensorflow\/python:_pywrap_tensorflow_internal.so'\n  failed (Exit 1)\n  \/usr\/lib\/gcc\/x86_64-linux-gnu\/5\/..\/..\/..\/x86_64-linux-gnu\/crti.o: In\n  function <code>_init': (.init+0x7): relocation truncated to fit:\n  R_X86_64_REX_GOTPCRELX against undefined symbol<\/code><strong>gmon_start<\/strong>'\n  \/usr\/lib\/gcc\/x86_64-linux-gnu\/5\/crtbeginS.o: In function\n  <code>deregister_tm_clones': crtstuff.c:(.text+0x3): relocation truncated\n  to fit: R_X86_64_PC32 against<\/code>.tm_clone_table'\n  crtstuff.c:(.text+0xa): relocation truncated to fit: R_X86_64_PC32\n  against symbol <code>__TMC_END__' defined in .nvFatBinSegment section in\n  bazel-out\/k8-dbg\/bin\/tensorflow\/python\/_pywrap_tensorflow_internal.so\n  crtstuff.c:(.text+0x1e): relocation truncated to fit:\n  R_X86_64_REX_GOTPCRELX against undefined symbol\n  <\/code>_ITM_deregisterTMCloneTable'\n  \/usr\/lib\/gcc\/x86_64-linux-gnu\/5\/crtbeginS.o: In function\n  <code>register_tm_clones': crtstuff.c:(.text+0x43): relocation truncated to\n  fit: R_X86_64_PC32 against<\/code>.tm_clone_table' crtstuff.c:(.text+0x4a):\n  relocation truncated to fit: R_X86_64_PC32 against symbol\n  <code>__TMC_END__' defined in .nvFatBinSegment section in\n  bazel-out\/k8-dbg\/bin\/tensorflow\/python\/_pywrap_tensorflow_internal.so\n  crtstuff.c:(.text+0x6b): relocation truncated to fit:\n  R_X86_64_REX_GOTPCRELX against undefined symbol\n  <\/code>_ITM_registerTMCloneTable'\n  \/usr\/lib\/gcc\/x86_64-linux-gnu\/5\/crtbeginS.o: In function\n  <code>__do_global_dtors_aux': crtstuff.c:(.text+0x92): relocation truncated\n  to fit: R_X86_64_PC32 against<\/code>.bss' crtstuff.c:(.text+0x9c):\n  relocation truncated to fit: R_X86_64_GOTPCREL against symbol\n  <code>__cxa_finalize@@GLIBC_2.2.5' defined in .text section in\n  \/lib\/x86_64-linux-gnu\/libc.so.6 crtstuff.c:(.text+0xaa): relocation\n  truncated to fit: R_X86_64_PC32 against symbol<\/code>__dso_handle' defined\n  in .data.rel.local section in\n  \/usr\/lib\/gcc\/x86_64-linux-gnu\/5\/crtbeginS.o crtstuff.c:(.text+0xbb):\n  additional relocation overflows omitted from the output\n  bazel-out\/k8-dbg\/bin\/tensorflow\/python\/_pywrap_tensorflow_internal.so:\n  PC-relative offset overflow in GOT PLT entry for\n  `_ZNK5Eigen10TensorBaseINS_9TensorMapINS_6TensorIKjLi1ELi1EiEELi16ENS_11MakePointerEEELi0EE9unaryExprINS_8internal11scalar_leftIjjN10tensorflow7functor14right_shift_opIjEEEEEEKNS_18TensorCwiseUnaryOpIT_KS6_EERKSH_'\n  collect2: error: ld returned 1 exit status Target\n  \/\/tensorflow\/tools\/pip_package:build_pip_package failed to build<\/p>\n<\/blockquote>\n\n<p>I hoped to be able to solve that by doing a monolithic build. So I tried that, and got essentially the same error.<\/p>\n\n<pre><code>bazel build --config cuda -c dbg --config=monolithic --copt=\"-DNDEBUG\"  \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>I also tried the approaches from <a href=\"https:\/\/stackoverflow.com\/questions\/40520146\/tensorflow-doesnt-build-with-debug-mode\">TensorFlow doesnt build with debug mode<\/a> and several other variants I found by extensive googling. I'm running out of options.<\/p>\n\n<p>I'd take any Tensorflow version from 1.11 onwards, including (working) nightly builds. It just needs to work with CUDA 9 on x86 linux, include debug symbols and disabled optimizations.<\/p>\n\n<p>thank you very much in Advance..<\/p>\n","answers":[{"AnswerId":"53239413","CreationDate":"2018-11-10T13:26:54.893","ParentId":null,"OwnerUserId":"3609835","Title":null,"Body":"<p>Just in case someone else stumbles over this problem. I finally got it to compile, using the following command:<\/p>\n\n<pre><code>bazel build --config cuda --strip=never --copt=\"-DNDEBUG\" --copt=\"-march=native\" --copt=\"-Og\" --copt=\"-g3\" --copt=\"-mcmodel=medium\" --copt=\"-fPIC\"  \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>After that, installation is a bit of a hazzle, since the wheel cannot be built anymore. But the tensorflow build can be installed anyway:<\/p>\n\n<p>When building the wheel, via <\/p>\n\n<pre><code>.\/bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n<\/code><\/pre>\n\n<p>The process fails with an error which seems to be a problem with python's builtin zip compression library (i.e. it cannot compress the resulting archive, since it's too large). <\/p>\n\n<p>It's important to run it anyway, since it only fails at the final step (archiving). When running build_pip_package, it prints to the console right at the start of the process, that it's building the package in a temporary directory (say, \/tmp\/Shjwejweu ) - the contents of that temp directory can be used to install tf debug version. Simply copy it to the target machine, then make sure you have any old tensorflow package removed (e.g. pip uninstall tensorflow),  and run within:<\/p>\n\n<pre><code>python setup.py install\n<\/code><\/pre>\n\n<p>But be careful to actively uninstall the \"tensorflow\" package first, otherwise you can get two simultaneously installed tensorflow versions..<\/p>\n"}]}
+{"QuestionId":53238597,"AnswerCount":0,"Tags":"<r><tensorflow><conv-neural-network>","CreationDate":"2018-11-10T11:44:23.653","AcceptedAnswerId":null,"OwnerUserId":10604232.0,"Title":"how to improve accuracy in image classification using CNN in R","Body":"<p>I am new to image classification.\nI am using CNN to classify 3 images of 3 products.\nBelow is my code -<\/p>\n\n<pre><code>model &lt;- keras_model_sequential()\n\nmodel %&gt;%\n\n     layer_conv_2d(filters = 64, \n                   kernel_size = c(3,3),\n                   activation = 'relu',\n                   input_shape = c(100,100,3)) %&gt;% \n\n   layer_conv_2d(filters = 32,\n                   kernel_size = c(3,3),\n                   activation = 'relu') %&gt;%\nlayer_max_pooling_2d(pool_size = c(2,2)) %&gt;%\n\n     layer_dropout(rate = 0.25) %&gt;%\n\n     layer_conv_2d(filters = 64,\n                   kernel_size = c(3,3),\n                   activation = 'relu') %&gt;%\n\n     layer_conv_2d(filters = 64,\n                   kernel_size = c(3,3),\n                   activation = 'relu') %&gt;%\n     layer_max_pooling_2d(pool_size = c(2,2)) %&gt;%\n     layer_dropout(rate = 0.25) %&gt;%\n      layer_flatten() %&gt;%\n     layer_dense(units = 256, activation = 'relu') %&gt;%\n     layer_dropout(rate=0.25) %&gt;%\n     layer_dense(units = 3, activation = 'softmax') %&gt;%\n\n     compile(loss = 'categorical_crossentropy',\n             optimizer = optimizer_sgd(lr = 0.01,\n                                       decay = 1e-6,\n                                       momentum = 0.9,\n                                       nesterov = T),\n             metrics = c('accuracy'))\n# Fit model\n<\/code><\/pre>\n\n<p>history &lt;- model %>%<\/p>\n\n<pre><code>     fit(train,\n         trainLabels,\n         epochs = 60,\n         batch_size =  32,\n         validation_split = 0.2,\n         validation_data = list(test, testLabels))\n<\/code><\/pre>\n\n<p>The accuracy that i am getting from this is 0.333\ni would like to improve this accuracy. i am not able to find any tutorials which explains on how to decide the appropriate  kernel size or the appropriate no of filters.\nCan someone explain same to me?\nPlease let me know if you require any other details.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[]}
+{"QuestionId":53239332,"AnswerCount":1,"Tags":"<tensorflow><slice>","CreationDate":"2018-11-10T13:17:14.113","AcceptedAnswerId":null,"OwnerUserId":10632990.0,"Title":"How to slice a tensor using given indices in tensorflow?","Body":"<p>I have a tensor with probabilities. This is a dynamic tensor with shape (?, 30) and I am selecting index with the best probability of these 30 values as :<\/p>\n\n<pre><code>    best_probability = tf.argmax(probability, axis = 1)\n<\/code><\/pre>\n\n<p>Now the dimensions of tensor best_probability is (?,). Now I want to select the values with these indices from another tensor called data with dimensions (?, 30, 1024, 3). Essentially from each of the 30 values select one with best probability using best_probability tensor.<\/p>\n\n<p>The final output should have dimensions of (?, 1024, 3). <\/p>\n\n<p>PS:- I tried gather_nd but it need indexing of best_probability tensor something like [[0, 9], [1, 10], [2, 15], [3, 25]]. To do so I wrote following snippet.<\/p>\n\n<pre><code> selected_data = tf.stack(tf.range(probability.shape[0]),\n                                tf.argmax(probability, axis = 1))\n<\/code><\/pre>\n\n<p>This doesn't work as I am dealing with a dynamic tensor. Is there any alternative to solve this problem. <\/p>\n","answers":[{"AnswerId":"53239897","CreationDate":"2018-11-10T14:24:35.637","ParentId":null,"OwnerUserId":"10632990","Title":null,"Body":"<p>I was able to solve this issue using tf.batch_gather and tf.reshape<\/p>\n\n<pre><code>selected_data = tf.reshape(tf.batch_gather(data, best_probability),\n                           (-1, data.shape[2],data.shape[3]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53240175,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><gpu><inference>","CreationDate":"2018-11-10T14:58:51.380","AcceptedAnswerId":null,"OwnerUserId":10632801.0,"Title":"Neural network inference with one image: Why is GPU utilization not 100%?","Body":"<p>First of all: this question is connected to neural network inference and not training.<\/p>\n\n<p>I have discovered, that when doing inference of a trained neural network with only one image over and over on a GPU (e.g. P100) the utilization of the computing power with Tensorflow is not reaching 100%, but instead around 70%. This is also the case if the image does not have to be transferred to the GPU. Therefore, the issue has to be connected to constraints in the parallelization of the calculations. My best guesses for the reasons are:<\/p>\n\n<ul>\n<li><p><em>Tensorflow<\/em> can only <em>utilize<\/em> the parallelization capabilities of a GPU up to a certain level. (Also the higher utilization of the same model as a TensorRT models suggest that). In this case, the question is: What is the reason for that?<\/p><\/li>\n<li><p>The inherent <em>neural network structure<\/em> with several subsequent layers avoids that a higher usage is possible. Therefore the problem is not overhead of a framework but lies in the general design of neural networks. In this case, the question is: What are the restrictions to that?<\/p><\/li>\n<li><p>Both of the above combined.<\/p><\/li>\n<\/ul>\n\n<p>Thanks for your ideas on the issue!<\/p>\n","answers":[{"AnswerId":"53240354","CreationDate":"2018-11-10T15:20:10.150","ParentId":null,"OwnerUserId":"3009636","Title":null,"Body":"<p>Why do you expect the GPU utilization to go to 100% when you run the neuronal network prediction for one image?<\/p>\n\n<p>The GPU utilization is per time unit (e.g. 1 second). This means, when the neuronal network algorithm finished before this time unit elapsed (e.g within 0.5s) Then the rest of the time the GPU may get used by other programs or not get used at all. If the GPU is not used by any other programs neither then well you will not reach 100%.<\/p>\n"}]}
+{"QuestionId":53240288,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-10T15:12:41.337","AcceptedAnswerId":null,"OwnerUserId":938512.0,"Title":"Cannot Reload saved Keras model using tensorflow","Body":"<p>I am working in a single jupyter notebook.   I create and train a very simple CNN with keras.  It compiles, fits, and predicts fine.   I save it with:<\/p>\n\n<pre><code>model.save(\"mymodel.hd5\")\n<\/code><\/pre>\n\n<p>Model is a keras.models.Sequential.  <\/p>\n\n<p>I then read that back in with:<\/p>\n\n<pre><code>reload_keras_model = keras.models.load_model(\"mymodel.hd5\")\n<\/code><\/pre>\n\n<p>That also works fine.  However if I try to read the model in using tensorflow via:<\/p>\n\n<pre><code>from tensorflow.keras.models import load_model\nreload_tf_mmodel = load_model(\"mymodel.hd5\")\n<\/code><\/pre>\n\n<p>That fails with:<\/p>\n\n<pre><code>ValueError: Unknown layer:layers\n<\/code><\/pre>\n\n<p>Most of the threads I've read on github say \"update your model\" or comments about custom objects (I'm not using any).   My target platform is the rpi zero and I've been able to install tf but unable to install keras, and that's why I want to load via tf.   Why would keras and tf.keras handle this model differently and what do I need to update\/change to read it in with tf.keras?<\/p>\n","answers":[{"AnswerId":"53240859","CreationDate":"2018-11-10T16:14:39.550","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>While keras (can) use TF as Backend, it does not guarantee that the saved model is readable in TF as well. <br><\/p>\n\n<p>Note that you can use keras with both theano and tf, thus <code>reload_keras_model = keras.models.load_model(\"mymodel.hd5\")<\/code> will work good with both backends, as the saving\/loading is done in the \"keras\" part, and not using the backend.<\/p>\n\n<p>You can use this tool: <a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"nofollow noreferrer\">keras_to_tensorflow<\/a><\/p>\n\n<p>Or something similar.<\/p>\n"}]}
+{"QuestionId":53240394,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-10T15:24:25.460","AcceptedAnswerId":null,"OwnerUserId":8384504.0,"Title":"How to draw a line in an image by connecting points in TensorFlow","Body":"<p>I want to know if I could draw a line in an image by connecting two points in TensorFlow.\nI should do this in a differentiable way.<\/p>\n","answers":[]}
+{"QuestionId":53240581,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-10T15:45:51.100","AcceptedAnswerId":53241567.0,"OwnerUserId":4785824.0,"Title":"Understanding keras.backend.max usage with tf.random_normal","Body":"<pre><code>import numpy as np\nimport tensorflow as tf\nfrom keras import backend as K\n\nsess = tf.InteractiveSession()\n\nbox_scores1 = tf.constant([[[ 9.188682,  11.484599 ],\n                      [10.06533,    7.557296 ]],\n\n                    [[10.099248,  10.591225 ],\n                     [10.592823 ,  7.8770704]]])\n\nbox_scores2 = tf.random_normal([2,2,2], mean=10, stddev=1, dtype=tf.float32, seed = 1)\n\nbox_class_scores1 = K.max(box_scores1, axis=-1)\nbox_class_scores2 = K.max(box_scores2, axis=-1)\n\nprint(box_scores1.eval())\nprint(box_scores2.eval())\n\nprint(box_class_scores1.eval())\nprint(box_class_scores2.eval())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[[ 9.188682  11.484599 ]\n  [10.06533    7.557296 ]]\n\n [[10.099248  10.591225 ]\n  [10.592823   7.8770704]]]\n[[[ 9.188682  11.484599 ]\n  [10.06533    7.557296 ]]\n\n [[10.099248  10.591225 ]\n  [10.592823   7.8770704]]]\n[[11.484599 10.06533 ]\n  [10.591225 10.592823]]\n[[10.242094 10.515779]\n [12.083789 11.397354]]\n<\/code><\/pre>\n\n<p>As, we can see values in box_scores1 and box_scores2 are same but the result obtained after applying max operation differs. How can the values of box_class_scores1 and box_class_scores2 be different?<\/p>\n","answers":[{"AnswerId":"53241567","CreationDate":"2018-11-10T17:28:12.310","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Your problem has nothing to do with the <code>max<\/code> function, but a misunderstanding with tensorflow, as most of its operations are symbolic, so when you use <code>tf.random_mormal<\/code>, this does not produce random numbers, but a symbolic normal distribution with the given mean and standard distribution.<\/p>\n\n<p>Then, each time you evaluate this distribution, <strong>it generates different outputs<\/strong>, so your first eval looks ok, but the second produces a different output that is given to <code>max<\/code>, so it produces a different output than just giving a constant vector.<\/p>\n"}]}
+{"QuestionId":53241021,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-10T16:32:26.073","AcceptedAnswerId":53272742.0,"OwnerUserId":2518806.0,"Title":"Tensorflow installation issue: ImportError: DLL load failed with error code -1073741795","Body":"<p>I am unable to install and import Tensorflow on my PC. I have tried the below approach as mentioned on <a href=\"https:\/\/www.tensorflow.org\/install\/pip#package-location\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip#package-location<\/a>. I have also tried installation using Conda but the same problem appears. MS Visual C++ is also installed as mentioned on <a href=\"https:\/\/www.tensorflow.org\/install\/pip#package-location\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip#package-location<\/a><\/p>\n\n<p>System information<\/p>\n\n<blockquote>\n  <p>OS Platform: Windows 7 Service Pack1 TensorFlow installed from (source\n  or binary):\n  <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.12.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.12.0-cp35-cp35m-win_amd64.whl<\/a>\n  TensorFlow version: 1.12 Python version: 3.5.4 64-bit Installed using\n  virtualenv<\/p>\n<\/blockquote>\n\n<p>Traceback (most recent call last):<\/p>\n\n<blockquote>\n  <p>File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in  from\n  tensorflow.python.pywrap_tensorflow_internal import * File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28,\n  in \n  _pywrap_tensorflow_internal = swig_import_helper() File \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24,\n  in swig_i mport_helper\n  _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, descript ion) File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\im\n  p.py\", line 243, in load_module return load_dynamic(name, filename,\n  file) File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\im\n  p.py\", line 343, in load_dynamic return _load(spec) ImportError: DLL\n  load failed with error code -1073741795<\/p>\n<\/blockquote>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):<\/p>\n\n<blockquote>\n  <p>File \"\", line 1, in  File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow_init_.py\", line 24, in  from tensorflow.python\n  import pywrap_tensorflow # pylint: disable=unused-im port File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python_init_.py\", line 49, in  from\n  tensorflow.python import pywrap_tensorflow File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in \n  raise ImportError(msg) ImportError: Traceback (most recent call last):\n  File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in  from\n  tensorflow.python.pywrap_tensorflow_internal import * File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28,\n  in \n  _pywrap_tensorflow_internal = swig_import_helper() File \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\si\n  te-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24,\n  in swig_i mport_helper\n  _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, descript ion) File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\im\n  p.py\", line 243, in load_module return load_dynamic(name, filename,\n  file) File\n  \"C:\\Users\\Desk\\AppData\\Local\\Programs\\Python\\Python35\\Scripts\\venv\\lib\\im\n  p.py\", line 343, in load_dynamic return _load(spec) ImportError: DLL\n  load failed with error code -1073741795<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53272742","CreationDate":"2018-11-13T02:07:13.577","ParentId":null,"OwnerUserId":"2518806","Title":null,"Body":"<p>Got it fixed post looking for similar answers in internet...I had to install the tensorflow version 1.5 and the tensorflow backend loaded seamlessly<\/p>\n"},{"AnswerId":"53261948","CreationDate":"2018-11-12T12:11:57.623","ParentId":null,"OwnerUserId":"10389355","Title":null,"Body":"<p>If your CPU didn't support AVX instructions, you will get ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed. (Win 10) or ImportError: DLL load failed with error code -1073741795 (Win 7) when using tensorflow official release 1.6.0 and up (pip install tensorflow)\nYou can use pip install [filename].whl which file download from sse2 folder instead of using official AVX binary.<\/p>\n\n<p>Please verify whether the below command is working.\n<strong>pip3 install --upgrade tensorflow<\/strong><\/p>\n\n<p>You can use the below link for reference of tensorflow-windows-wheel.\n<strong><a href=\"https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fo40225\/tensorflow-windows-wheel<\/a><\/strong><\/p>\n\n<p>I have built tensorflow wheel without AVX.<\/p>\n\n<p><a href=\"https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\/tree\/master\/1.6.0\/py36\/CPU\/sse2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\/tree\/master\/1.6.0\/py36\/CPU\/sse2<\/a><\/p>\n\n<p>I wish this .whl may help you.<\/p>\n\n<p>You can install the wheel file with pip. First change the current directory to install location. Then,<\/p>\n\n<p><strong>pip install tensorflow-1.6.0-cp36-cp36m-win_amd64.whl<\/strong><\/p>\n\n<p>If its not working please try to resolve it by downgrading protobuf from 3.6.1 to 3.6.0:<\/p>\n\n<p><strong>pip install protobuf==3.6.0<\/strong><\/p>\n"}]}
+{"QuestionId":53241028,"AnswerCount":0,"Tags":"<keras><large-data-volumes><imagedata>","CreationDate":"2018-11-10T16:33:21.887","AcceptedAnswerId":null,"OwnerUserId":8156744.0,"Title":"Keras : Dealing with large image datasets","Body":"<p>I am trying to fit a model using a large image datasets. I have a memory RAM of 14 GB, and the dataset have the size of 40 GB. I tried to use <code>fit_generator<\/code>, but I end up with a method that does not delete the loaded batchs after using theme.<\/p>\n\n<p>If there is anyway to sole the problem or resources, thanks to point me to it.<\/p>\n\n<p>Thanks.<\/p>\n\n<p>The generator code is :<\/p>\n\n<pre><code>class Data_Generator(Sequence):\n\n    def __init__(self, image_filenames, labels, batch_size):\n        self.image_filenames, self.labels = image_filenames, labels\n        self.batch_size = batch_size\n\n    def __len__(self):\n        return int(np.ceil(len(self.image_filenames) \/ float(self.batch_size)))\n    def __format_labels__(self, gd_truth):\n        cols=gd_truth.columns\n        y=[]\n        for col in cols:\n            y.append(gd_truth[col].values)\n        return y\n\n    def __getitem__(self, idx):\n        batch_x = self.image_filenames[idx * self.batch_size:(idx + 1) * self.batch_size]\n        batch_y = self.labels[idx * self.batch_size:(idx + 1) * self.batch_size]\n        gd_truth=pd.DataFrame(data=batch_y,columns=self.labels.columns)\n        #gd_truth=batch_y\n        return np.array([read_image(file_name) for file_name in batch_x]),self.__format_labels__(gd_truth) #np.array(batch_y)\n<\/code><\/pre>\n\n<p>Then I have created two generators for train and validation images: <\/p>\n\n<pre><code>training_batch_generator = Data_Generator(training_filenames, trainTargets, batch_size)\nmvalidation_batch_generator = Data_Generator(validation_filenames, valTargets, batch_size)\n<\/code><\/pre>\n\n<p>The fit_generator call is as follow : <\/p>\n\n<pre><code>num_epochs=10\nmodel.fit_generator(generator=my_training_batch_generator,\n                                          steps_per_epoch=(num_training_samples \/\/ batch_size),\n                                          epochs=num_epochs,\n                                          verbose=1,\n                                          validation_data=my_validation_batch_generator,\n                                          validation_steps=(num_validation_samples \/\/ batch_size),\n                                          max_queue_size=16)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53241056,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><deep-learning>","CreationDate":"2018-11-10T16:36:31.340","AcceptedAnswerId":null,"OwnerUserId":681911.0,"Title":"Tensorflow gradients for every item of tensor","Body":"<p>I have a network that takes as input an <code>Nx3<\/code> matrix and produces an N-dimensional vector. Let's say batch size is 1 and N=1024, so the output would have the shape <code>(1,1024)<\/code>. I want to compute the gradients for every dimension of the output, with respect to the input. That is, <code>dy\/dx<\/code> for every <code>y<\/code>. However tensorflow's <code>tf.gradients<\/code> computes <code>d sum(y)\/dx<\/code>, aggregate. I know there's no straightforward way to compute the gradients for every output dimension, so I finally decided to run <code>tf.gradients<\/code> 1024 times, because I only have to do this once in the project, and never again. <\/p>\n\n<p>So I do this:<\/p>\n\n<pre><code>start = datetime.datetime.now()\noutput_code_split = tf.split(output_code,1024)\n#output shape = (1024,)\ngrad_ops = []\nfor i in range(1024):\n    gr = tf.gradients(output_code_split[i],input)\n    #output shape = (1024,1,16,1024,3) , where 16= batch size\n\n    gr = tf.reduce_mean(gr,[0,1,2,3])\n    #output shape = (1024,)\n\n    grad_ops.append(gr)\n    present = datetime.datetime.now()\n    print(i,(present-start).seconds,flush=True)\n    #prints time taken to finish previous computation.\n    start = datetime.datetime.now()\n<\/code><\/pre>\n\n<p>When the code started running, the time <strong>between<\/strong> two iterations was 4 seconds, so I figured it'll run for roughly 4096 seconds. However, as the number of iterations increase, the time taken for subsequent runs keeps increasing. The gap, which was 4 seconds when the code started, eventually got to 30 seconds after about 500 iterations, which is too much. <\/p>\n\n<p>Is the list holding the gradient ops <code>grad_ops<\/code> growing bigger and occupying more memory. I'm unfortunately not in the position to do a detailed memory profiling of this code..Any ideas about what causes the iteration time to blow up as time goes on?<\/p>\n\n<p>(Note that in the code, I'm only creating the gradient ops and not actually evaluating them. That part comes later, but my code doesn't reach there on account of the extreme slowdown mentioned above)<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53243209","CreationDate":"2018-11-10T20:40:52.700","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>What blows up your execution time is that you define a new operation on the graph in every iteration of your for loop. Every call to <code>tf.gradient<\/code> and <code>tf.reduce_mean<\/code> pushes a new node onto the graph. Then it needs to recompile to be run. What should actually work for you is to use <code>tf.gather<\/code> with an int32 placeholder, which supplies the dimension to your gradient operation. So something like this:<\/p>\n\n<pre><code>idx_placeholder = tf.placeholder(tf.int32, shape=(None,))\ngrad_operation = tf.gradients(tf.gather(output_code_split, idx_placeholder))\nfor i in range(1024):\n      sess.run(grad_operation, {idx_placeholder: np.array([i])})\n<\/code><\/pre>\n"}]}
+{"QuestionId":53241220,"AnswerCount":0,"Tags":"<java><arrays><tensorflow>","CreationDate":"2018-11-10T16:53:58.627","AcceptedAnswerId":null,"OwnerUserId":10633625.0,"Title":"Tensorflow Mobile- Creating a Conditional Statement based on the output","Body":"<p>I am a novice at programming. I have setup Tensorflow mobile (I know tensor flow lite exists, but I am on windows, so I can't convert the .pb file to .lite)successfully on my pc, including a custom image recognition file. However, I would like to create a conditional statement based on the outputs of the Tensorflow prediction. E.g.- if the name of the highest predicted object is \"Android\" then print out an additional statement. I believe this is possible through the id class, which is the name of the output of the prediction, but am not quite sure what to do. Would you point me in the right direction?<\/p>\n","answers":[]}
+{"QuestionId":53241578,"AnswerCount":1,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-11-10T17:29:31.840","AcceptedAnswerId":53243789.0,"OwnerUserId":9148806.0,"Title":"How to extract the value of a tensor in tensorflow.js?","Body":"<p>My model outputs ten predictions from which I take the largest one with argmax. Unfortunately, this also returns a tensor (with single value like this: Tensor: [8]) and I can't figure out how to make this tensor either into integer, or extract the only value. <\/p>\n\n<p>So far I've tried gather, asType, dataSync, and get.<\/p>\n","answers":[{"AnswerId":"53243789","CreationDate":"2018-11-10T21:54:48.777","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>dataSync returns an array. To get the single element, one needs to use the array indexing as follows<\/p>\n\n<pre><code>const t = model.predict(tensor)\n\/\/ Use of argmax and get the value\nconst v = t.argMax().dataSync()[0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53241974,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><neural-network>","CreationDate":"2018-11-10T18:12:59.293","AcceptedAnswerId":null,"OwnerUserId":10002393.0,"Title":"Problem after install tensorflow in python 3","Body":"<p>ive installed tensorflow<\/p>\n\n<p>But when i try to start it i give this error:<\/p>\n\n<pre><code>2018-11-10 19:07:16.129068: F tensorflow\/core\/platform\/cpu_feature_guard.cc:37] The TensorFlow library was compiled to use SSE4.1 instructions, but these aren't available on your machine.\n<\/code><\/pre>\n\n<p>Why?<\/p>\n","answers":[{"AnswerId":"53242755","CreationDate":"2018-11-10T19:43:44.427","ParentId":null,"OwnerUserId":"10634103","Title":null,"Body":"<p>Your CPU is too old for precompiled libraries. Or you compiled TF from source with SSE instructions enabled and your CPU does not support SSE instructions. <\/p>\n\n<p>If you used precompiled libraries, you will need to compile from source without SSE instructions. Check this <a href=\"https:\/\/stackoverflow.com\/questions\/41474136\/disable-sse4-1-when-compiling-tensorflow\/41477681#41477681\">post<\/a> for more details.<\/p>\n"}]}
+{"QuestionId":53242025,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2018-11-10T18:18:21.623","AcceptedAnswerId":null,"OwnerUserId":8499072.0,"Title":"Make error in caffe (I guess relate on Openmp,gcc)","Body":"<p>I build to <a href=\"https:\/\/github.com\/liulei01\/DRBox\" rel=\"nofollow noreferrer\">https:\/\/github.com\/liulei01\/DRBox<\/a><\/p>\n\n<p><code>make -j8<\/code> 's Output<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/O0nsy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O0nsy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>OS : ubuntu 14.04 <\/p>\n\n<p>cuda : 7.5<\/p>\n\n<p>cudnn : 5.1v<\/p>\n\n<p>gcc : 4.6.4<\/p>\n\n<p>what can i do?<\/p>\n","answers":[]}
+{"QuestionId":53242146,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2018-11-10T18:31:13.680","AcceptedAnswerId":53243153.0,"OwnerUserId":10594993.0,"Title":"How does the multi-input deep learning work in Keras?","Body":"<p>I have a multi-input convolutional neural network model that inputs 2 images from 2 datasets to give one output which is the class of the two inputs. The two datasets have the same classes. I used 2 vgg16 models and concatenate them to classify the two images.<\/p>\n\n<pre><code>vgg16_model = keras.applications.vgg16.VGG16()\ninput_layer1= vgg16_model .input\nlast_layer1 = vgg16_model.get_layer('fc2').output\n\n\nvgg16_model2 = keras.applications.vgg16.VGG16()\ninput_layer2= vgg16_model .input\nlast_layer2 = vgg16_model.get_layer('fc2').output\n\n\ncon = concatenate([last_layer1, last_layer2])  # merge the outputs of the two models\noutput_layer = Dense(no_classes, activation='softmax', name='prediction')(con)\nmultimodal_model1 = Model(inputs=[input_layer1, input_layer2], outputs=[output_layer])\n<\/code><\/pre>\n\n<p>My questions are:<\/p>\n\n<p>1- Which case from the following represents how the images enter to the model?<\/p>\n\n<p>One to One<\/p>\n\n<p>database1-img1 + database2-img1<\/p>\n\n<p>database1-img2 + database2-img2<\/p>\n\n<p>database1-img3 + database2-img3<\/p>\n\n<p>database1-img4 + database2-img4<\/p>\n\n<p>.........<\/p>\n\n<p>Many to many<\/p>\n\n<p>database1-img1 + database2-img1<\/p>\n\n<p>database1-img1 + database2-img2<\/p>\n\n<p>database1-img1 + database2-img3<\/p>\n\n<p>database1-img1 + database2-img4<\/p>\n\n<p>database1-img2 + database2-img1<\/p>\n\n<p>database1-img2 + database2-img2<\/p>\n\n<p>database1-img2 + database2-img3<\/p>\n\n<p>database1-img2 + database2-img4<\/p>\n\n<p>.........<\/p>\n\n<p>2- In general in deep learning, Does the images enter from the two datasets to the model at the same time have the same class (labels) or not?<\/p>\n","answers":[{"AnswerId":"53243153","CreationDate":"2018-11-10T20:32:28.727","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<ol>\n<li>It is a 1:1 mapping, the same should be with multiple outputs as well. <br><\/li>\n<\/ol>\n\n<p>When you have a model such as <code>Model(inputs=[input_layer1, input_layer2], outputs=[output_layer])<\/code> or even <code>Model(inputs=[input_layer1, input_layer2], outputs=[output_layer1, output_layer2])<\/code> , You must feed it with inputs \/ output of the same shape.<br>\nAssume the other case - You will need to have <code>ds1.shape[0] * ds2.shape[0]<\/code> different labels, for each possible mix of the 2 datasets, and will need to have them ordered at a certain way. That is not really feasible, at least not simply.\n<br><br>\n2. Its not as if the same images have the same label, but the <strong>Pair<\/strong> of both images have a single label.<\/p>\n"}]}
+{"QuestionId":53242203,"AnswerCount":1,"Tags":"<python><tensorflow><convolution>","CreationDate":"2018-11-10T18:37:58.980","AcceptedAnswerId":53349095.0,"OwnerUserId":10602820.0,"Title":"How to train convolution filters in tensorflow","Body":"<p>I'm trying to train a neural network that will be activated by horisontal lines (and in future, by corners).<\/p>\n\n<p>I draw an image that contains both vertical and horisontal lines (see link below). Then in photoshop I draw marks on top of horisontal lines, the places where I want the net to activate.<\/p>\n\n<p>Original and label images look like this:\n<a href=\"https:\/\/i.stack.imgur.com\/SHUAB.png\" rel=\"nofollow noreferrer\">original<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/6uZXo.png\" rel=\"nofollow noreferrer\">lbl_hor<\/a><\/p>\n\n<p>Now the task seems easy, but I can't get tensorflow to train such a net. Here is my code:<\/p>\n\n<pre><code>input = tf.convert_to_tensor(original, tf.float32)\nlabel = tf.convert_to_tensor(lbl_hor, tf.float32)\n\nW1 = tf.Variable(tf.random_normal([5,5,1,1]))\nb1 = tf.Variable(0.1)\nW2 = tf.Variable(tf.random_normal([5,5,1,1]))\nb2 = tf.Variable(0.1)\n\nl1 = tf.nn.conv2d(input, W1, [1,1,1,1], 'SAME')\nl1 = l1 + b1\nl1 = tf.nn.relu(l1)\n\nl2 = tf.nn.conv2d(l1, W2, [1,1,1,1], 'SAME')\nl2 = l2 + b2\nl2 = tf.nn.relu(l2)\n\nloss = tf.square(l2 - label)\nloss_sum = tf.reduce_sum(loss)\ntrain_op = tf.train.AdamOptimizer(0.01).minimize(loss)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nfor i in range(0, 100):\n    sess.run(train_op)\n<\/code><\/pre>\n\n<p>The network gives every time unpredictable output, sometimes just black. I tried to multiply the label by some constant so that the activations have more weight - does not have much effect...<\/p>\n\n<p>Can you help me to figure out how to train such a network or find what's wrong with my code?<\/p>\n","answers":[{"AnswerId":"53349095","CreationDate":"2018-11-17T07:16:40.397","ParentId":null,"OwnerUserId":"10602820","Title":null,"Body":"<p>Finally figured out the answer.\nIt's embarrassing, but in case someone needs it, posting it here.<\/p>\n\n<p>First, I should make sure that my labels and the network output are normalized:<\/p>\n\n<pre><code>output_normalized = tf.nn.l2_normalize(l2)\nlabel_normalized = tf.nn.l2_normalize(label)\n<\/code><\/pre>\n\n<p>Second, tensor <code>tf.square(l2 - label)<\/code> is of shape <code>[1, 200, 200, 200]<\/code> - that's clearly not right. So I modified it further:<\/p>\n\n<pre><code>output_normalized = tf.reshape(tf.nn.l2_normalize(l2), [200, 200])\nlabel_normalized = tf.nn.l2_normalize(label)\n\nloss = tf.square(output_normalized - label_normalized)\n<\/code><\/pre>\n\n<p>Here is new code, that works completely fine:<\/p>\n\n<pre><code>input = tf.convert_to_tensor(original, tf.float32)\nlabel = tf.convert_to_tensor(lbl_hor, tf.float32)\n\nW1 = tf.Variable(tf.ones([5,5,1,1]))\nb1 = tf.Variable(0.1)\nW2 = tf.Variable(tf.ones([5,5,1,1]))\nb2 = tf.Variable(0.1)\n\nl1 = tf.nn.conv2d(input, W1, [1,1,1,1], 'SAME')\nl1 = l1 + b1\nl1 = tf.nn.relu(l1)\n\nl2 = tf.nn.conv2d(l1, W2, [1,1,1,1], 'SAME')\nl2 = l2 + b2\nl2 = tf.nn.relu(l2)\n\noutput_normalized = tf.reshape(tf.nn.l2_normalize(l2), [200, 200])\nlabel_normalized = tf.nn.l2_normalize(label)\n\nloss = tf.square(output_normalized - label_normalized)\nloss_sum = tf.reduce_sum(loss)\ntrain_op = tf.train.AdamOptimizer(0.1).minimize(loss_sum)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nfor i in range(0, 100):\n    sess.run(train_op)\n<\/code><\/pre>\n\n<p>For some reason, <code>tf.random_normal<\/code> as initial values does not produce stable learning every time, so I have used <code>tf.ones<\/code> instead. Not sure why it happens, but that works...<\/p>\n"}]}
+{"QuestionId":53242299,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network><tflearn>","CreationDate":"2018-11-10T18:49:43.523","AcceptedAnswerId":null,"OwnerUserId":7673365.0,"Title":"Image Classification Using TensorFlow, Value error","Body":"<p>I am new to deep learning and tensorflow and i'm trying to create an image classifier using tensorflow that will classify 5 classes of images.\nMy training dataset is 25000 images and testing dataset is 5000 images.\nbelow is my code:<\/p>\n\n<pre><code>import os\nfrom random import shuffle\n\nimport cv2\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport tensorflow as tf\nimport tflearn\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\nfrom tqdm import tqdm\nimport keras\nfrom keras.models import save_model\n\"\"\"from keras.models import Sequential\nfrom keras.layers import Dense\"\"\"\n\nTRAIN_DIR = 'train'\nTEST_DIR = 'test'\nIMG_SIZE = 16\nLR = 1e-3\nMODEL_NAME = 'cifar 10 -convnet'\n\ndef create_label(image_name):\n    \"\"\" Create an one-hot encoded vector from image name \"\"\"\n    word_label = image_name.split('.')[0:2]\n    if word_label == 'cat':\n        return np.array([1,0,0,0,0])\n    elif word_label == 'Dog':\n        return np.array([0,1,0,0,0])\n    elif word_label == 'Automobile':\n        return np.array([0,0,1,0,0])\n    elif word_label == 'Airplane':\n        return np.array([0,0,0,1,0])\n    elif word_label == 'Ship':\n        return np.array([0,0,0,0,1])\n\n\ndef create_train_data():\n    training_data = []\n    for img in tqdm(os.listdir(TRAIN_DIR)):\n        path = os.path.join(TRAIN_DIR, img)\n        img_data = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n        img_data = cv2.resize(img_data, (IMG_SIZE, IMG_SIZE))\n        training_data.append([np.array(img_data), create_label(img)])\n    shuffle(training_data)\n    np.save('train_data.npy', training_data)\n    return training_data\n\n\ndef create_test_data():\n    testing_data = []\n    for img in tqdm(os.listdir(TEST_DIR)):\n        path = os.path.join(TEST_DIR, img)\n        img_num = img.split('.')[0:2]\n        img_data = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n        img_data = cv2.resize(img_data, (IMG_SIZE, IMG_SIZE))\n        testing_data.append([np.array(img_data), img_num])\n\n    shuffle(testing_data)\n    np.save('test_data.npy', testing_data)\n    return testing_data\n\n\n# If dataset is not created:\ntrain_data = create_train_data()\ntest_data = create_test_data()\n# If you have already created the dataset:\n# train_data = np.load('train_data.npy')\n# test_data = np.load('test_data.npy')\ntrain = train_data[:25000]\ntest = train_data[:5000]\nX_train = np.array([i[0] for i in train]).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\ny_train = [i[1] for i in train]\nX_test = np.array([i[0] for i in test]).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\ny_test = [i[1] for i in test]\n\n# Building The Model\n\ntf.reset_default_graph()\nconvnet = input_data(shape=[None, IMG_SIZE, IMG_SIZE, 1], name='input')\nconvnet = conv_2d(convnet, 32, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\nconvnet = conv_2d(convnet, 64, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\nconvnet = conv_2d(convnet, 128, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\nconvnet = conv_2d(convnet, 64, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\nconvnet = conv_2d(convnet, 32, 5, activation='relu')\nconvnet = max_pool_2d(convnet, 5)\nconvnet = fully_connected(convnet, 1024, activation='relu')\nconvnet = dropout(convnet, 0.8)\nconvnet = fully_connected(convnet, 2, activation='softmax')\nconvnet = regression(convnet, optimizer='adam', learning_rate=LR, loss='categorical_crossentropy', name='targets')\nmodel = tflearn.DNN(convnet, tensorboard_dir='log', tensorboard_verbose=0)\nhistory = model.fit({'input': X_train}, {'targets': y_train}, n_epoch=25,\n          validation_set=({'input': X_test}, {'targets': y_test}),\n          snapshot_step=500, show_metric=True, run_id=MODEL_NAME)\n\n#print(history.history.keys())\n#plt.figure(1)\n\n# summarize history for accuracy\n\n\"\"\"plt.subplot(211)\nplt.plot(history.history['acc'])\nplt.plot(history.history['val_acc'])\nplt.title('model accuracy')\nplt.ylabel('accuracy')\nplt.xlabel('epoch')\nplt.legend(['train', 'test1'], loc='upper left')\nplt.show()\"\"\"\n\nfig = plt.figure(figsize=(25, 12))\n\nfor num, data in enumerate(test_data[:25]):\n\n    img_num = data[1]\n    img_data = data[0]\n\n    y = fig.add_subplot(5, 5, num + 1)\n    orig = img_data\n    data = img_data.reshape(IMG_SIZE, IMG_SIZE, 1)\n    model_out = model.predict([data])[0]\n\n    if np.argmax(model_out) == 1:\n        str_label = 'Dog'\n    if np.argmax(model_out) == 2:\n        str_label = 'Automobile'\n    if np.argmax(model_out) == 3:\n        str_label = 'Airplane'\n    if np.argmax(model_out) == 4:\n        str_label = 'Ship'\n    else:\n        str_label = 'Cat'\n\n    y.imshow(orig, cmap='gray')\n    plt.title(str_label)\n    y.axes.get_xaxis().set_visible(False)\n    y.axes.get_yaxis().set_visible(False)\nplt.show()\n<\/code><\/pre>\n\n<p>i'm getting following error:\n ValueError: Cannot feed value of shape (64,) for Tensor 'targets\/Y:0', which has shape '(?, 2)'<\/p>\n\n<p>Can someone help me with this?\nthanks in advance.<\/p>\n","answers":[{"AnswerId":"53279911","CreationDate":"2018-11-13T11:20:26.140","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>The problem is with the shape of your labels, so check the shape for <code>y_train<\/code> and <code>y_test<\/code> . Both of them should have to have a shape similar to the out put of your model, which is shown at the error <code>...which has shape '(?, 2)'<\/code> and defined at the line where you create the last fully connected layer <code>convnet = fully_connected(convnet, 2, activation='softmax')\n<\/code> . <\/p>\n"}]}
+{"QuestionId":53242520,"AnswerCount":0,"Tags":"<python-2.7><docker><tensorflow>","CreationDate":"2018-11-10T19:14:24.903","AcceptedAnswerId":null,"OwnerUserId":3174459.0,"Title":"tensorflow 1.11 and cuda 8","Body":"<p>TensorFlow 1.11 fails to build with CUDA 8. I tried opening an issue on github (Issue opened on Github #23256 [<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23256]\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23256]<\/a>) but the tensorflow team's response is to just upgrade CUDA to 9 or downgrade Tensorflow to 1.10, which isn't an option for me. Trying to find a way to get TF1.11 to work with CUDA 8.<\/p>\n\n<p>Attempting to build a docker container with TF 1.11 and CUDA 8 on an GeForce 1060 3GB GPU. An error keeps occurring in the build. Github Issue 22729 (#22729) was looked at but the work around didn't work for TF 1.11 and that's what is needed. The docker file is also below. Any help you can provide would be greatly appreciated.<\/p>\n\n<p>System information<\/p>\n\n<p>OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04\nTensorFlow installed from (source or binary): Source<\/p>\n\n<p>TensorFlow version: TF 1.11<\/p>\n\n<p>Python version: 2.7<\/p>\n\n<p>Installed using virtualenv? pip? conda?: Docker<\/p>\n\n<p>Bazel version (if compiling from source): 0.15.0<\/p>\n\n<p>GCC\/Compiler version (if compiling from source): 7.3.0<\/p>\n\n<p>CUDA\/cuDNN version: 8.0\/7<\/p>\n\n<p>GPU model and memory: GeForce GTX 1060 3GB<\/p>\n\n<p>Provide the exact sequence of commands \/ steps that you executed before running into the problem\nsudo docker build --no-cache . -f Dockerfile.tf-1.11-py27-gpu.txt -t tf-1.11-py27-gpu<\/p>\n\n<p>Thank you,\nKyle<\/p>\n\n<p>Dockerfile.tf-1.11-py27-gpu<\/p>\n\n<pre><code>FROM nvidia\/cuda:8.0-cudnn7-devel-ubuntu16.04\n\nLABEL maintainer=\"Craig Citro &lt;craigcitro@google.com&gt;; Modified for Cuda 8 by Jack Harris\"\n\nRUN apt-get update &amp;&amp; apt-get install -y --allow-downgrades --allow-change-held-packages --no-install-recommends \\\n    build-essential \\\n    cuda-command-line-tools-8-0 \\\n    cuda-cublas-dev-8-0 \\\n    cuda-cudart-dev-8-0 \\\n    cuda-cufft-dev-8-0 \\\n    cuda-curand-dev-8-0 \\\n    cuda-cusolver-dev-8-0 \\\n    cuda-cusparse-dev-8-0 \\\n    curl \\\n    git \\\n    libcudnn7=7.2.1.38-1+cuda8.0 \\\n    libcudnn7-dev=7.2.1.38-1+cuda8.0 \\\n    libnccl2=2.2.13-1+cuda8.0 \\\n    libnccl-dev=2.2.13-1+cuda8.0 \\\n    libcurl3-dev \\\n    libfreetype6-dev \\\n    libhdf5-serial-dev \\\n    libpng12-dev \\\n    libzmq3-dev \\\n    pkg-config \\\n    python-dev \\\n    rsync \\\n    software-properties-common \\\n    unzip \\\n    zip \\\n    zlib1g-dev \\\n    wget \\\n    &amp;&amp; \\\nrm -rf \/var\/lib\/apt\/lists\/* &amp;&amp; \\\nfind \/usr\/local\/cuda-8.0\/lib64\/ -type f -name 'lib*_static.a' -not -name 'libcudart_static.a' -delete &amp;&amp; \\\nrm -f \/usr\/lib\/x86_64-linux-gnu\/libcudnn_static_v7.a\n\nRUN apt-get update &amp;&amp; \\\n    apt-get install nvinfer-runtime-trt-repo-ubuntu1604-4.0.1-ga-cuda8.0 &amp;&amp; \\\n    apt-get update &amp;&amp; \\\n    apt-get install libnvinfer4=4.1.2-1+cuda8.0 &amp;&amp; \\\n    apt-get install libnvinfer-dev=4.1.2-1+cuda8.0\n\n# Link NCCL libray and header where the build script expects them.\nRUN mkdir \/usr\/local\/cuda-8.0\/lib &amp;&amp;  \\\nln -s \/usr\/lib\/x86_64-linux-gnu\/libnccl.so.2 \/usr\/local\/cuda\/lib\/libnccl.so.2 &amp;&amp; \\\nln -s \/usr\/include\/nccl.h \/usr\/local\/cuda\/include\/nccl.h\n\n# TODO(tobyboyd): Remove after license is excluded from BUILD file.\n#RUN gunzip \/usr\/share\/doc\/libnccl2\/NCCL-SLA.txt.gz &amp;&amp; \\\n#    cp \/usr\/share\/doc\/libnccl2\/NCCL-SLA.txt \/usr\/local\/cuda\/\n\n# Add External Mount Points\nRUN mkdir -p \/external_lib\nRUN mkdir -p \/external_bin\n\nRUN curl -fSsL -O https:\/\/bootstrap.pypa.io\/get-pip.py &amp;&amp; \\\n    python get-pip.py &amp;&amp; \\\n    rm get-pip.py\n\nRUN pip --no-cache-dir install \\\n    ipykernel \\\n    jupyter \\\n    keras_applications==1.0.5 \\\n    keras_preprocessing==1.0.3 \\\n    matplotlib \\\n    numpy \\\n    pandas \\\n    scipy \\\n    sklearn \\\n    mock \\ \n    &amp;&amp; \\\npython -m ipykernel.kernelspec\n\n# Set up our notebook config.\n#COPY jupyter_notebook_config.py \/root\/.jupyter\/\n\n# Jupyter has issues with being run directly:\n#   https:\/\/github.com\/ipython\/ipython\/issues\/7062\n# We just add a little wrapper script.\n# COPY run_jupyter.sh \/\n\n# Set up Bazel.\n\n# Running bazel inside a `docker build` command causes trouble, cf:\n#   https:\/\/github.com\/bazelbuild\/bazel\/issues\/134\n# The easiest solution is to set up a bazelrc file forcing --batch.\nRUN echo \"startup --batch\" &gt;&gt;\/etc\/bazel.bazelrc\n# Similarly, we need to workaround sandboxing issues:\n#   https:\/\/github.com\/bazelbuild\/bazel\/issues\/418\nRUN echo \"build --spawn_strategy=standalone --genrule_strategy=standalone\" \\\n    &gt;&gt;\/etc\/bazel.bazelrc\n# Install the most recent bazel release.\nENV BAZEL_VERSION 0.15.0\nWORKDIR \/\nRUN mkdir \/bazel &amp;&amp; \\\n    cd \/bazel &amp;&amp; \\\n    curl -H \"User-Agent: Mozilla\/5.0 (X11; Linux x86_64) AppleWebKit\/537.36     (KHTML, like Gecko) Chrome\/57.0.2987.133 Safari\/537.36\" -fSsL -O https:\/\/github.com\/bazelbuild\/bazel\/releases\/download\/$BAZEL_VERSION\/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh &amp;&amp; \\\ncurl -H \"User-Agent: Mozilla\/5.0 (X11; Linux x86_64) AppleWebKit\/537.36 (KHTML, like Gecko) Chrome\/57.0.2987.133 Safari\/537.36\" -fSsL -o \/bazel\/LICENSE.txt https:\/\/raw.githubusercontent.com\/bazelbuild\/bazel\/master\/LICENSE &amp;&amp; \\\nchmod +x bazel-*.sh &amp;&amp; \\\n.\/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh &amp;&amp; \\\ncd \/ &amp;&amp; \\\nrm -f \/bazel\/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh\n\n# Download and build TensorFlow.\nRUN git clone http:\/\/github.com\/tensorflow\/tensorflow --branch r1.11 --depth=1\n\nWORKDIR \/tensorflow\n\nRUN sed -i 's\/^#if TF_HAS_.*$\/#if !defined(__NVCC__)\/g' tensorflow\/core\/platform\/macros.h\n\nENV TF_NCCL_VERSION=2\n\n#RUN \/bin\/echo -e \"\/usr\/bin\/python\\n\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\ny\\n8.0\\n\/usr\/local\/cuda\\n7.0\\n\/usr\/local\/cuda\\n\\n\\n\\nn\\n\\nn\\n-march=native\\nn\\n\" | .\/configure\nRUN \/bin\/echo -e \"\/usr\/bin\/python\\n\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\nn\\ny\\n8.0\\n\/usr\/local\/cuda\\n7.0\\n\/usr\/local\/cuda\\nn\\n\\n\\n\\n\\n\\nn\\n\\nn\\n-march=native\\nn\\n\" | .\/configure\n#RUN \/bin\/echo -e \"\\n\\nn\\nn\\nn\\nn\\nn\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n-march=native\\nn\\n\" | .\/configure\n\n# Configure the build for our CUDA configuration.\nENV CI_BUILD_PYTHON python\nENV PATH \/external_bin:$PATH\nENV LD_LIBRARY_PATH \/external_lib:\/usr\/local\/cuda\/extras\/CUPTI\/lib64:$LD_LIBRARY_PATH\nENV TF_NEED_CUDA 1\nENV TF_NEED_TENSORRT 1\nENV TF_CUDA_COMPUTE_CAPABILITIES=3.0,3.5,5.2,6.0,6.1\nENV TF_CUDA_VERSION=8.0\nENV TF_CUDNN_VERSION=7\n\n# https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17801\nRUN ln -s \/usr\/local\/cuda\/lib64\/stubs\/libcuda.so \/usr\/local\/cuda\/lib64\/stubs\/libcuda.so.1 &amp;&amp; \\\nln -s \/usr\/local\/cuda\/nvvm\/libdevice\/libdevice.compute_50.10.bc \/usr\/local\/cuda\/nvvm\/libdevice\/libdevice.10.bc &amp;&amp; \\\nLD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\/stubs:${LD_LIBRARY_PATH} \\\ntensorflow\/tools\/ci_build\/builds\/configured GPU \\\nbazel build -c opt --copt=-mavx --config=cuda \\\n--cxxopt=\"-D_GLIBCXX_USE_CXX11_ABI=0\" \\\n    tensorflow\/tools\/pip_package\/build_pip_package &amp;&amp; \\\n    rm \/usr\/local\/cuda\/lib64\/stubs\/libcuda.so.1 \n\nRUN bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/pip \n\nRUN pip --no-cache-dir install --upgrade \/tmp\/pip\/tensorflow-*.whl &amp;&amp; \\\nrm -rf \/tmp\/pip &amp;&amp; \\\nrm -rf \/root\/.cache\n# Clean up pip wheel and Bazel cache when done.\n\nWORKDIR \/root\n\n# TensorBoard\nEXPOSE 6006\n# IPython\nEXPOSE 8888\n\nCMD [ \"\/bin\/bash\" ]\n<\/code><\/pre>\n\n<p>tf11cuda8.log - Log attached to github issue (too long to post here)<\/p>\n","answers":[]}
+{"QuestionId":53242817,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-10T19:50:59.017","AcceptedAnswerId":53267309.0,"OwnerUserId":4590385.0,"Title":"Which sorting algorithm does TensorFlow use?","Body":"<p>I'm trying to figure out how TensorFlow's <code>sort<\/code> function works.<\/p>\n\n<p>I've looked in the docs and tried to find in the source code without luck.<\/p>\n\n<p>The function is <code>tf.contrib.framework.sort(values)<\/code>?\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/framework\/sort\" rel=\"nofollow noreferrer\">1<\/a><\/p>\n","answers":[{"AnswerId":"53267309","CreationDate":"2018-11-12T17:34:39.347","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>From the sort source code, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/contrib\/framework\/python\/ops\/sort_ops.py#L175\" rel=\"nofollow noreferrer\">this<\/a> is the line where the actual sort operation is called. The topk op is actually part of the core kernels and is written in C\/C++. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/topk_op.cc\" rel=\"nofollow noreferrer\">This<\/a> is the corresponding source code for it. A quick high level glance seems to employ some form of heap sort.  <\/p>\n"}]}
+{"QuestionId":53242875,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><classification>","CreationDate":"2018-11-10T19:58:13.467","AcceptedAnswerId":53246611.0,"OwnerUserId":5796133.0,"Title":"Accuracy Decreasing with higher epochs","Body":"<p>I am a newbie to Keras and machine learning in general. I'm trying to build a binary classification model using the Sequential model. After some experimenting, I saw that on multiple runs(not always) I was getting a an accuracy of even 97% on my validation data in the second or third epoch itself but this dramatically decreased to as much as 12%. What is the reason behind this ? How do I fine tune my model ? \nHere's my code - <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Flatten(input_shape=(6,size)))\nmodel.add(Dense(6,activation='relu'))\nmodel.add(Dropout(0.35))\nmodel.add(Dense(3,activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(1,activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['binary_accuracy'])\nmodel.fit(x, y,epochs=60,batch_size=40,validation_split=0.2)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53246611","CreationDate":"2018-11-11T07:14:25.650","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>According to me, you can take into consideration the following factors.<\/p>\n\n<ol>\n<li><strong>Reduce your learning rate<\/strong> to a very small number like 0.001 or even 0.0001.<\/li>\n<li>Provide <strong>more data<\/strong>.<\/li>\n<li>Set <strong>Dropout rates<\/strong> to a number like 0.2. <strong>Keep them uniform across the network<\/strong>.<\/li>\n<li>Try <strong>decreasing the batch size<\/strong>.<\/li>\n<li>Using appropriate optimizer: You may need to experiment a bit on this. Use different optimizers on the same network, and <strong>select an optimizer which gives you the least loss<\/strong>.<\/li>\n<\/ol>\n\n<p>If any of the above factors work for you, please let me know about it in the comments section.<\/p>\n"}]}
+{"QuestionId":53243067,"AnswerCount":1,"Tags":"<python><keras><hyperparameters><hyperopt>","CreationDate":"2018-11-10T20:20:20.060","AcceptedAnswerId":null,"OwnerUserId":3261292.0,"Title":"reinitializing keras model weights after each training pass","Body":"<p>I noticed few similar questions similar to this one in Stack-overflow, but none has an answer ..<\/p>\n\n<p>I have a simple Keras model:<\/p>\n\n<pre><code>def create_model(x_train, y_train, x_val, y_val):\n    # building the model\n    # compile\n    # fit\n    # return the score using model.predict\n<\/code><\/pre>\n\n<p>I'm applying cross validation (Kfold stratified) as following:<\/p>\n\n<pre><code>    skf = StratifiedKFold(y, n_folds=5, shuffle=True, random_state=0)\n    scores = []\n    for train_index, val_index in skf:\n        X_train, X_val = df[train_index], df[val_index]\n        y_train, y_val = y[train_index], y[val_index]\n\n        scores.append(create_model(X_train, y_train, X_val, y_val))\n        # point A\n<\/code><\/pre>\n\n<p>Do I have to reinitialize the model weights after each training pass (point A), or Keras library manage this process?<\/p>\n\n<p>If not, any suggestion that can improve the processing time (maybe flushing the memory? .. if possible).<\/p>\n\n<p>I'm asking this question because I'm applying this process with Hyperopt library for Hyperparameter optimization, and I noticed that after many trials the model starts to take more time than in the beginning .. <\/p>\n\n<p>Edit: as an Example, you can notice the processing time below for Hyperopt evals, where in each pass the 5-folds method is applied:<\/p>\n\n<pre><code>Hyperopt evals:   3%|\u258e         | 5\/150 [16:09&lt;7:54:20, 196.28s\/it]\n\nHyperopt evals:   4%|\u258d         | 6\/150 [22:33&lt;10:06:20, 252.64s\/it]\n\nHyperopt evals:   5%|\u258d         | 7\/150 [26:20&lt;9:43:55, 245.01s\/it] \n\nHyperopt evals:   5%|\u258c         | 8\/150 [33:33&lt;11:53:16, 301.38s\/it]\n\nHyperopt evals:   6%|\u258c         | 9\/150 [41:56&lt;14:10:16, 361.82s\/it]\n\nHyperopt evals:   7%|\u258b         | 10\/150 [45:56&lt;12:38:50, 325.22s\/it]\n\nHyperopt evals:   7%|\u258b         | 11\/150 [48:19&lt;10:26:55, 270.61s\/it]\n\nHyperopt evals:   8%|\u258a         | 12\/150 [54:11&lt;11:18:28, 294.99s\/it]\n\nHyperopt evals:   9%|\u258a         | 13\/150 [58:45&lt;10:58:57, 288.59s\/it]\n\nHyperopt evals:   9%|\u2589         | 14\/150 [1:05:57&lt;12:31:47, 331.68s\/it]\n\nHyperopt evals:  10%|\u2588         | 15\/150 [1:13:38&lt;13:53:30, 370.45s\/it]\n\nHyperopt evals:  11%|\u2588         | 16\/150 [1:17:36&lt;12:18:28, 330.66s\/it]\n\nHyperopt evals:  11%|\u2588\u258f        | 17\/150 [1:25:56&lt;14:06:13, 381.75s\/it]\n\nHyperopt evals:  12%|\u2588\u258f        | 18\/150 [1:31:54&lt;13:43:38, 374.39s\/it]\n\nHyperopt evals:  13%|\u2588\u258e        | 19\/150 [1:36:11&lt;12:20:55, 339.35s\/it]\n\nHyperopt evals:  13%|\u2588\u258e        | 20\/150 [1:45:06&lt;14:22:20, 398.01s\/it]\n\nHyperopt evals:  14%|\u2588\u258d        | 21\/150 [1:49:14&lt;12:38:51, 352.95s\/it]\n\nHyperopt evals:  15%|\u2588\u258d        | 22\/150 [1:54:45&lt;12:18:47, 346.31s\/it]\n\nHyperopt evals:  15%|\u2588\u258c        | 23\/150 [1:59:04&lt;11:17:24, 320.04s\/it]\n\nHyperopt evals:  16%|\u2588\u258c        | 24\/150 [2:04:05&lt;11:00:29, 314.52s\/it]\n\nHyperopt evals:  17%|\u2588\u258b        | 25\/150 [2:07:47&lt;9:57:11, 286.65s\/it] \n\nHyperopt evals:  17%|\u2588\u258b        | 26\/150 [2:12:47&lt;10:00:37, 290.62s\/it]\n\nHyperopt evals:  18%|\u2588\u258a        | 27\/150 [2:17:08&lt;9:37:55, 281.91s\/it] \n\nHyperopt evals:  19%|\u2588\u258a        | 28\/150 [2:22:46&lt;10:07:15, 298.65s\/it]\n\nHyperopt evals:  19%|\u2588\u2589        | 29\/150 [2:28:56&lt;10:45:29, 320.08s\/it]\n\nHyperopt evals:  20%|\u2588\u2588        | 30\/150 [2:34:55&lt;11:03:44, 331.87s\/it]\n\nHyperopt evals:  21%|\u2588\u2588        | 31\/150 [2:40:20&lt;10:53:43, 329.61s\/it]\n\nHyperopt evals:  21%|\u2588\u2588\u258f       | 32\/150 [2:46:19&lt;11:05:42, 338.50s\/it]\n\nHyperopt evals:  22%|\u2588\u2588\u258f       | 33\/150 [2:51:47&lt;10:53:54, 335.34s\/it]\n\nHyperopt evals:  23%|\u2588\u2588\u258e       | 34\/150 [2:58:14&lt;11:18:06, 350.75s\/it]\n\nHyperopt evals:  23%|\u2588\u2588\u258e       | 35\/150 [3:04:10&lt;11:15:41, 352.53s\/it]\n\nHyperopt evals:  24%|\u2588\u2588\u258d       | 36\/150 [3:13:59&lt;13:24:26, 423.39s\/it]\n\nHyperopt evals:  25%|\u2588\u2588\u258d       | 37\/150 [3:20:13&lt;12:49:38, 408.66s\/it]\n\nHyperopt evals:  25%|\u2588\u2588\u258c       | 38\/150 [3:25:55&lt;12:05:23, 388.61s\/it]\n\nHyperopt evals:  26%|\u2588\u2588\u258c       | 39\/150 [3:35:53&lt;13:54:59, 451.35s\/it]\n\nHyperopt evals:  27%|\u2588\u2588\u258b       | 40\/150 [3:44:26&lt;14:21:12, 469.75s\/it]\n\nHyperopt evals:  27%|\u2588\u2588\u258b       | 41\/150 [3:50:42&lt;13:22:33, 441.77s\/it]\n\nHyperopt evals:  28%|\u2588\u2588\u258a       | 42\/150 [3:58:03&lt;13:14:29, 441.39s\/it]\n\nHyperopt evals:  29%|\u2588\u2588\u258a       | 43\/150 [4:11:11&lt;16:12:35, 545.38s\/it]\n\nHyperopt evals:  29%|\u2588\u2588\u2589       | 44\/150 [4:19:18&lt;15:32:40, 527.93s\/it]\n\nHyperopt evals:  30%|\u2588\u2588\u2588       | 45\/150 [4:26:03&lt;14:19:21, 491.06s\/it]\n\nHyperopt evals:  31%|\u2588\u2588\u2588       | 46\/150 [4:34:32&lt;14:20:31, 496.46s\/it]\n\nHyperopt evals:  31%|\u2588\u2588\u2588\u258f      | 47\/150 [4:45:01&lt;15:20:25, 536.17s\/it]\n\nHyperopt evals:  32%|\u2588\u2588\u2588\u258f      | 48\/150 [4:54:11&lt;15:18:45, 540.45s\/it]\n\nHyperopt evals:  33%|\u2588\u2588\u2588\u258e      | 49\/150 [4:58:42&lt;12:53:19, 459.40s\/it]\n\nHyperopt evals:  33%|\u2588\u2588\u2588\u258e      | 50\/150 [5:04:07&lt;11:38:30, 419.11s\/it]\n\nHyperopt evals:  34%|\u2588\u2588\u2588\u258d      | 51\/150 [5:12:48&lt;12:22:14, 449.85s\/it]\n\nHyperopt evals:  35%|\u2588\u2588\u2588\u258d      | 52\/150 [5:20:37&lt;12:23:57, 455.49s\/it]\n\nHyperopt evals:  35%|\u2588\u2588\u2588\u258c      | 53\/150 [5:28:18&lt;12:19:19, 457.31s\/it]\n\nHyperopt evals:  36%|\u2588\u2588\u2588\u258c      | 54\/150 [5:37:02&lt;12:43:26, 477.15s\/it]\n\nHyperopt evals:  37%|\u2588\u2588\u2588\u258b      | 55\/150 [5:45:21&lt;12:46:00, 483.80s\/it]\n\nHyperopt evals:  37%|\u2588\u2588\u2588\u258b      | 56\/150 [5:51:07&lt;11:33:16, 442.51s\/it]\n\nHyperopt evals:  38%|\u2588\u2588\u2588\u258a      | 57\/150 [5:59:38&lt;11:57:39, 463.00s\/it]\n\nHyperopt evals:  39%|\u2588\u2588\u2588\u258a      | 58\/150 [6:11:19&lt;13:39:13, 534.27s\/it]\n\nHyperopt evals:  39%|\u2588\u2588\u2588\u2589      | 59\/150 [6:28:06&lt;17:05:39, 676.26s\/it]\n\nHyperopt evals:  40%|\u2588\u2588\u2588\u2588      | 60\/150 [6:37:29&lt;16:03:23, 642.27s\/it]\n\nHyperopt evals:  41%|\u2588\u2588\u2588\u2588      | 61\/150 [6:43:38&lt;13:51:06, 560.30s\/it]\n\nHyperopt evals:  41%|\u2588\u2588\u2588\u2588\u258f     | 62\/150 [6:52:41&lt;13:33:52, 554.92s\/it]\n\nHyperopt evals:  42%|\u2588\u2588\u2588\u2588\u258f     | 63\/150 [7:00:05&lt;12:36:40, 521.84s\/it]\n\nHyperopt evals:  43%|\u2588\u2588\u2588\u2588\u258e     | 64\/150 [7:12:13&lt;13:56:21, 583.50s\/it]\n\nHyperopt evals:  43%|\u2588\u2588\u2588\u2588\u258e     | 65\/150 [7:20:03&lt;12:58:38, 549.62s\/it]\n\nHyperopt evals:  44%|\u2588\u2588\u2588\u2588\u258d     | 66\/150 [7:31:56&lt;13:58:08, 598.68s\/it]\n\nHyperopt evals:  45%|\u2588\u2588\u2588\u2588\u258d     | 67\/150 [7:44:48&lt;15:00:05, 650.67s\/it]\n\nHyperopt evals:  45%|\u2588\u2588\u2588\u2588\u258c     | 68\/150 [7:57:32&lt;15:35:45, 684.70s\/it]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53437103","CreationDate":"2018-11-22T19:30:24.780","ParentId":null,"OwnerUserId":"3261292","Title":null,"Body":"<blockquote>\n  <p>Do I have to reinitialize the model weights after each training pass\n  (point A), or Keras library manage this process?<\/p>\n<\/blockquote>\n\n<p>After checking the documentation + manual experiments: It seemed to me that Keras manage re-initializing the weights, and it's not needed.<\/p>\n\n<blockquote>\n  <p>If not, any suggestion that can improve the processing time (maybe\n  flushing the memory? .. if possible).<\/p>\n<\/blockquote>\n\n<p>The processing time in my case was increasing because: <\/p>\n\n<p>1- Hyperopt uses bayesian optimization technique, so it tries each time when it pick the next parameter set to choose something better based on the prior probabilities<br>\n2- I'm using early stopping.<\/p>\n\n<p>Thus, in each next eval, the hyperopt library starts to choose better parameter set where the model also starts to converge better than previously .. which means, less using of early stopping and more processing time (to complete the whole epoches).<\/p>\n"}]}
+{"QuestionId":53243386,"AnswerCount":1,"Tags":"<python-3.x><keras><rasa-nlu><rasa-core>","CreationDate":"2018-11-10T21:03:52.507","AcceptedAnswerId":53296557.0,"OwnerUserId":10180167.0,"Title":"Features and Story graph || input to keras policy","Body":"<p>Could you please let me know <strong>how are you identifying the features<\/strong> you're passing to keras policy.\nI could see story graphs are being created during agent.load_data.<\/p>\n\n<p>Could you please share live example so that I can tune the parameters and hyperparameters to get the best out of keras lstm model.<\/p>\n\n<p>Rasa Core version: 0.11.12<\/p>\n\n<p>Python version: 3.5<\/p>\n\n<p>Operating system (windows, osx, ...):windows 10<\/p>\n","answers":[{"AnswerId":"53296557","CreationDate":"2018-11-14T09:12:35.327","ParentId":null,"OwnerUserId":"3429596","Title":null,"Body":"<p>The selected feature depends on the used policy and its configuration.\nYou can specify your configuration in a <a href=\"https:\/\/rasa.com\/docs\/core\/policies\/\" rel=\"nofollow noreferrer\">policy configuration file<\/a>. If you use the \"embedding policy\", you can also define the layers etc. of the used LSTM in this configuration file.<\/p>\n\n<p>The features are identified from<\/p>\n\n<ul>\n<li>intents so far<\/li>\n<li>last actions<\/li>\n<li>current slot values \/ entities values<\/li>\n<\/ul>\n\n<p>Have a look at the <a href=\"https:\/\/rasa.com\/docs\/core\/api\/featurizer\/#\" rel=\"nofollow noreferrer\">documentation on featurization<\/a> for more details since this highly depends on the used policy configuration (you can select different featurizers).<\/p>\n\n<p>With rasa_core version 0.12 you can only <a href=\"https:\/\/rasa.com\/docs\/core\/evaluation\/#comparing-policies\" rel=\"nofollow noreferrer\">compare the accuracy of different policies<\/a> with the command <code>python -m rasa_core.train compare<\/code>. This is probably helpful if you want to finetune that. <\/p>\n"}]}
+{"QuestionId":53244152,"AnswerCount":0,"Tags":"<python><tensorflow><generative-adversarial-network>","CreationDate":"2018-11-10T22:43:25.550","AcceptedAnswerId":null,"OwnerUserId":5278088.0,"Title":"Problem GAN conversion when applying variable reuse on tensorflow","Body":"<p>I am building an GAN and when i started calling my discriminator twice, using reuse, my GAN started to diverge. I first created my discriminator as following:<\/p>\n\n<pre><code>def discriminator(self, x_past, x_future, gen_future):\n        os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' \n        with tf.variable_scope(\"disc\") as disc:\n            gen_future = tf.concat([gen_future, x_past], 2)\n            x_future = tf.concat([x_future, x_past], 2)\n            x_in = tf.concat([gen_future, x_future], 0)\n            conv1 = tf.layers.conv1d(inputs=x_in, filters=20, kernel_size=3, strides=1,\n                padding='same', activation=tf.nn.relu)\n            max_pool_1 = tf.layers.max_pooling1d(inputs=conv1, pool_size=2, strides=2, padding='same')\n            conv2 = tf.layers.conv1d(inputs=max_pool_1, filters=3, kernel_size=2, strides=1,\n            padding='same', activation=tf.nn.relu)\n            max_pool_2 = tf.layers.max_pooling1d(inputs=conv2, pool_size=2, strides=2, padding='same')\n\n            # Flatten and add dropout\n            flat = tf.reshape(max_pool_2, (-1, 9))\n            flat = tf.nn.dropout(flat, keep_prob=self.keep_prob)\n\n            # Predictions\n            logits = tf.layers.dense(flat, 2)\n\n            y_true = logits[:self.batch_size]\n            y_gen = logits[self.batch_size:]\n\n            return y_true, y_gen\n<\/code><\/pre>\n\n<p>And I was calling it like this:<\/p>\n\n<pre><code>y_true, y_gen = self.discriminator(x_past, x_future, gen_future)\n<\/code><\/pre>\n\n<p>I was able to train the GAN properly. Now I need to use reuse to be able to call it without having to send real and fake data at once. I changed it to:<\/p>\n\n<pre><code>def discriminator(self, x_past, x_future, reuse=False):\n        os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' \n        with tf.variable_scope(\"disc\", reuse=reuse) as disc:\n            x_in = tf.concat([x_future, x_past], 2)\n            conv1 = tf.layers.conv1d(inputs=x_in, filters=20, kernel_size=3, strides=1,\n                padding='same', activation=tf.nn.relu)\n            max_pool_1 = tf.layers.max_pooling1d(inputs=conv1, pool_size=2, strides=2, padding='same')\n            conv2 = tf.layers.conv1d(inputs=max_pool_1, filters=3, kernel_size=2, strides=1,\n            padding='same', activation=tf.nn.relu)\n            max_pool_2 = tf.layers.max_pooling1d(inputs=conv2, pool_size=2, strides=2, padding='same')\n\n            # Flatten and add dropout\n            flat = tf.reshape(max_pool_2, (-1, 9))\n            flat = tf.nn.dropout(flat, keep_prob=self.keep_prob)\n\n            # Predictions\n            logits = tf.layers.dense(flat, 2)\n            return logits\n<\/code><\/pre>\n\n<p>And calling it like this:<\/p>\n\n<pre><code>y_true = self.discriminator(x_past, x_future)\ny_gen = self.discriminator(x_past, gen_future, reuse=True)\n<\/code><\/pre>\n\n<p>Now it started to diverge. Any idea why is that?<\/p>\n","answers":[]}
+{"QuestionId":53244407,"AnswerCount":1,"Tags":"<keras><output><lstm>","CreationDate":"2018-11-10T23:25:07.530","AcceptedAnswerId":null,"OwnerUserId":1376574.0,"Title":"Output sequences for multivariable output for LSTM in Keras","Body":"<p>Please help to understand how to form output sequence for LSTM network in Keras.\nFor example I have data<\/p>\n\n<pre><code>[[1]\n [2]\n [3]\n [4]\n [5]]\n<\/code><\/pre>\n\n<p>I can for input sequences as per [samples, timesteps, features] tensor \n- assume 3 input sequence and 2 output. So<\/p>\n\n<pre><code>Input seq      Output seq\n[[[1,2,3]]] -&gt; [[4, 5]] \n<\/code><\/pre>\n\n<p>but what about multivariable input and output?!<br>\nNeural network may be several neurons - one neuron for each feature\nExample (two features data, two features input and two features output):<\/p>\n\n<p>Data:<\/p>\n\n<pre><code>[[1, 10]\n [2, 20]\n [3, 30]\n [4, 40]\n [5, 50]]\n<\/code><\/pre>\n\n<p>Input tensor (samples, timesteps - 3, features - 2)<\/p>\n\n<pre><code>[[[1, 2, 3], [10, 20, 30]]]\n<\/code><\/pre>\n\n<p>But output matrix for Y<\/p>\n\n<pre><code>[[4, 5, 30, 40]] \n<\/code><\/pre>\n\n<ul>\n<li>but it doesn't sense we have a mess with completely \ndifferent features!!! <\/li>\n<\/ul>\n\n<p>Can I conclude that in case of Keras we can form output sequences for only\none feature?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"53247402","CreationDate":"2018-11-11T09:29:55.370","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>For multivariate also it would be similar, if your data has 2 features as you have mentioned in your question:<\/p>\n\n<pre><code>[[1, 10]\n [2, 20]\n [3, 30]\n [4, 40]\n [5, 50]]\n<\/code><\/pre>\n\n<p>And you want to prepare it with<code>Input tensor (samples, timesteps - 3, features - 2)<\/code> ,ie 3 timestep and each record has 2 variables then it would look like below:<\/p>\n\n<pre><code> x=[\n      [[1,10],[2,20],[3,30]],\n      [[2,20],[3,30],[4,40]],\n      [[3,30],[4,40],[5,50]]       \n  ]\ny=[[4, 5],[5,6],[6,7]] \n<\/code><\/pre>\n\n<p>I just assumed you had value 6 and 7 otherwise according to your given data you can only have just one(1) record in x ie (1 sample,3 timestep,2featues)<\/p>\n\n<pre><code> x=[\n          [[1,10],[2,20],[3,30]]\n   ]\n<\/code><\/pre>\n\n<p>and just one record of y<\/p>\n\n<pre><code>y=[[4, 5]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53244539,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-10T23:46:12.970","AcceptedAnswerId":null,"OwnerUserId":4006471.0,"Title":"TensorFlow Estimator makes different prediction on each call of predict","Body":"<p>I trained a classifier for the Iris dataset using TF Estimators, but each prediction call I'm getting different results. I wonder if I'm making something wrong in the training or if it is an issue in the prediction.<\/p>\n\n<p>I'm loading an already trained model and just making the .predict call. This is my input function for prediction. <\/p>\n\n<pre><code>def get_predict_fn(features,batch_size):\n    def predict_input_fn():\n        dataset = tf.data.Dataset.from_tensor_slices(dict(features))\n        dataset = dataset.batch(batch_size)\n        return dataset.make_one_shot_iterator().get_next()\n\n    return predict_input_fn\n<\/code><\/pre>\n\n<p>This is the result of one call<\/p>\n\n<blockquote>\n  <p>[{'logits': array([-3.5082035, -1.074667 , -3.8533034],\n  dtype=float32), 'probabilities': array([0.07629351, 0.8696793 ,\n  0.05402722], dtype=float32), 'class_ids': array([1]), 'classes': array([b'Iris-versicolor'], dtype=object)}]<\/p>\n<\/blockquote>\n\n<p>This is another call<\/p>\n\n<blockquote>\n  <p>[{'logits': array([ 3.0530725, -1.0889677,  2.3922846],\n  dtype=float32), 'probabilities': array([0.6525989 , 0.01037006,\n  0.337031  ], dtype=float32), 'class_ids': array([0]), 'classes': array([b'Iris-setosa'], dtype=object)}]<\/p>\n<\/blockquote>\n\n<p>Both are making a call to the same model, sending the same example DataFrame.<\/p>\n\n<blockquote>\n  <p>sepal_length  sepal_width  petal_length  petal_width<br>\n  5.7             2.5           5.0          2.0<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53245079,"AnswerCount":1,"Tags":"<python><keras><autoencoder>","CreationDate":"2018-11-11T01:31:36.567","AcceptedAnswerId":null,"OwnerUserId":3837778.0,"Title":"VAE in Keras: how to define the end-to-end model?","Body":"<p>I am learning the <a href=\"http:\/\/nbviewer.jupyter.org\/github\/fchollet\/deep-learning-with-python-notebooks\/blob\/master\/8.4-generating-images-with-vaes.ipynb\" rel=\"nofollow noreferrer\">tutorial<\/a> here. My Model part is:<\/p>\n\n<pre><code>input_img = keras.Input(shape=img_shape)\n\nx = layers.Conv2D(32, (3, 3),\n                  padding='same', activation='relu')(input_img)\n\n...\nx = layers.Conv2D(64, (3, 3),\n                  padding='same', activation='relu')(x)\nshape_before_flattening = K.int_shape(x)\n\nx = layers.Flatten()(x)\nx = layers.Dense(32, activation='relu')(x)\n\nz_mean = layers.Dense(latent_dim)(x)\nz_log_var = layers.Dense(latent_dim)(x)\n\ndef sampling(args):\n    ...\n\nz = layers.Lambda(sampling)([z_mean, z_log_var])\n\ndecoder_input = layers.Input(K.int_shape(z)[1:])\n\nx = layers.Dense(np.prod(shape_before_flattening[1:]),\n                 activation='relu')(decoder_input)\n\nx = layers.Reshape(shape_before_flattening[1:])(x)\n\nx = layers.Conv2DTranspose(32, 3,\n                           padding='same', activation='relu',\n                           strides=(2, 2))(x)\nx = layers.Conv2D(1, 3,\n                  padding='same', activation='sigmoid')(x)\n\n# This is our decoder model from letent space to reconstructed images\ndecoder = Model(decoder_input, x)\n\n# We then apply it to `z` to recover the decoded `z`.\nz_decoded = decoder(z)\n\ndef vae_loss(self, x, z_decoded):\n    ...\n\n\n# Fit the end-to-end model\nvae = Model(input_img, z_decoded) # vae = Model(input_img, x)\nvae.compile(optimizer='rmsprop', loss=vae_loss)\nvae.summary()\n<\/code><\/pre>\n\n<p>My question is: the end-to-end is <code>vae = Model(input_img, z_decoded)<\/code> or <code>vae = Model(input_img, x)<\/code>. Should we compute loss on <code>input_img<\/code> and <code>z_decoded<\/code> OR between <code>input_img<\/code> and <code>x<\/code>? Thanks<\/p>\n","answers":[{"AnswerId":"53246597","CreationDate":"2018-11-11T07:11:44.913","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>x is changing throughout the model, where <code>x = layers.Conv2D(1, 3,padding='same', activation='sigmoid')(x)<\/code> you set <code>x<\/code> to be the last layer of your decoder model. <br><br>\nWhen doing <code>z_decoded = decoder(z)<\/code> you chain your decoder straight after the encoder, <code>z_decoded<\/code> is actually the output layer of your decoder, thus, the same <code>x<\/code> as earlier. Also, you create the link between the actual input and the output. <br><br><\/p>\n\n<p>Computing the loss would yield the same results on both (as they both represent the same layer). <br>\nIn short - Both <code>vae = Model(input_img, z_decoded)<\/code> and <code>vae = Model(input_img, x)<\/code> are the end to end model, i would suggest using the z_decoded version, for readability.<\/p>\n"}]}
+{"QuestionId":53245413,"AnswerCount":0,"Tags":"<tensorflow><tensor>","CreationDate":"2018-11-11T02:46:52.967","AcceptedAnswerId":null,"OwnerUserId":10634878.0,"Title":"How to lookup tables initialized with estimators?","Body":"<p>I train a linear model with tf.estimator.train_and_evaluate and save model by <code>tf.estimator.RunConfig(model_dir=FLAGS.checkpoints_dir,save_checkpoints_steps=FLAGS.save_checkpoints_steps)<\/code>. Now I want to inference with the saved model, meet the error \"Table not initialized\", how should solve this problem? <\/p>\n","answers":[]}
+{"QuestionId":53246056,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><conv-neural-network>","CreationDate":"2018-11-11T05:19:32.753","AcceptedAnswerId":null,"OwnerUserId":9115361.0,"Title":"Problem with retrain.py. Getting error tensorflow.python.framework.errors_impl.NotFoundError","Body":"<p>I'm using <code>retrain.py<\/code> to retrain an object detector on photos of my hand (to detect how many fingers I'm holding). On the Tensorflow site, I followed the tutorial where I retrained it on their images of flowers. So I wrote python <code>retrain.py --image_dir ~\/flower_photo<\/code>s and it worked. I had my <code>hand_photos<\/code> directory which contained sub-directories of me holding various fingers up. However I received this error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.NotFoundError: \/Users\/spencerkraisler\/hand_photos\/untitled folder; No such file or directory\n<\/code><\/pre>\n\n<p>I don't understand why it's looking for some untitled folder in my <code>hand_photos<\/code>directory, nor why it works find on <code>flower_photos<\/code>.<\/p>\n\n<p>I am using the current tensor flow and python 3.6. I am in an <code>anaconda<\/code> environment however this works just fine with flower_photos.<\/p>\n\n<p>Furthermore, flower_photos just contains sub-directories of photos of various flowers, all <code>.jpg<\/code>. My photos are also in <code>.jpg<\/code> format.<\/p>\n","answers":[{"AnswerId":"53246102","CreationDate":"2018-11-11T05:29:37.707","ParentId":null,"OwnerUserId":"9115361","Title":null,"Body":"<p>Well problem solved. I actually had an empty directory named \"untitled folder\" in my training data. I must've created it by accident. Case closed. <\/p>\n"}]}
+{"QuestionId":53246288,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><tfrecord><protoc>","CreationDate":"2018-11-11T06:06:08.690","AcceptedAnswerId":null,"OwnerUserId":5029100.0,"Title":"I am doing image processing in Python and TensorFlow I try to create TF_Record file it give error","Body":"<p>I already install protoc3 and my Python version is 3.6. I already created the CSV file using that I have to create TF_record. <\/p>\n\n<pre><code>anoj@anoj-Lenovo-G50-80:~\/**PycharmProjects\/untitled1\/venv$ python3 tf_Record.py --csv_input=data\/train_labels.csv  --output_path=train.record**\n<\/code><\/pre>\n\n<p>this is my error<\/p>\n\n<pre><code>*Traceback (most recent call last):\n  File \"tf_Record.py\", line 99, in &lt;module&gt;\n    tf.app.run()\n  File \"\/home\/anoj\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"tf_Record.py\", line 90, in main\n    tf_example = create_tf_example(group, path)\n  File \"tf_Record.py\", line 45, in create_tf_example\n    encoded_jpg = fid.read()\n  File \"\/home\/anoj\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 125, in read\n    self._preread_check()\n  File \"\/home\/anoj\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 85, in _preread_check\n    compat.as_bytes(self.__name), 1024 * 512, status)\n  File \"\/home\/anoj\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py\", line 528, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.NotFoundError: \/home\/anoj\/PycharmProjects\/untitled1\/venv\/images\/20081023231835!Hk-p9s-1; No such file or directory*\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53246331,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-probability>","CreationDate":"2018-11-11T06:14:55.630","AcceptedAnswerId":53453868.0,"OwnerUserId":798689.0,"Title":"Saving and restoring functions in TensorFlow","Body":"<p>I am working on a VAE project in TensorFlow where the encoder\/decoder networks are build in functions. The idea is to be able to save, then load the trained model and do sampling, using the encoder function.<\/p>\n\n<p>After restoring the model, I am having trouble getting the decoder function to run and give me back the restored, trained variables, getting an \"Uninitialized value\" error. I assume it is because the function is either creating a new new one, overwriting the existing, or otherwise. But I cannot figure out how to solve this. Here is some code:<\/p>\n\n<pre><code>class VAE(object):    \n    def __init__(self, restore=True):\n        self.session = tf.Session()\n        if restore:\n            self.restore_model()\n            self.build_decoder = tf.make_template('decoder', self._build_decoder)\n\n@staticmethod\ndef _build_decoder(z, output_size=768, hidden_size=200,\n                  hidden_activation=tf.nn.elu, output_activation=tf.nn.sigmoid):\n    x = tf.layers.dense(z, hidden_size, activation=hidden_activation)\n    x = tf.layers.dense(x, hidden_size, activation=hidden_activation)\n    logits = tf.layers.dense(x, output_size, activation=output_activation)\n    return distributions.Independent(distributions.Bernoulli(logits), 2)\n\ndef sample_decoder(self, n_samples):\n    prior = self.build_prior(self.latent_dim)\n    samples = self.build_decoder(prior.sample(n_samples), self.input_size).mean()\n    return self.session.run([samples])\n\ndef restore_model(self):\n    print(\"Restoring\")\n    self.saver = tf.train.import_meta_graph(os.path.join(self.save_dir, \"turbolearn.meta\"))\n    self.saver.restore(self.sess, tf.train.latest_checkpoint(self.save_dir))\n    self._restored = True\n<\/code><\/pre>\n\n<p>want to run <code>samples = vae.sample_decoder(5)<\/code><\/p>\n\n<p>In my training routine, I run:<\/p>\n\n<pre><code>        if self.checkpoint:\n            self.saver.save(self.session, os.path.join(self.save_dir, \"myvae\"), write_meta_graph=True)\n<\/code><\/pre>\n\n<h3>UPDATE<\/h3>\n\n<p>Based on the suggested answer below, I changed the restore method<\/p>\n\n<pre><code>self.saver = tf.train.Saver()\nself.saver.restore(self.session, tf.train.latest_checkpoint(self.save_dir))\n<\/code><\/pre>\n\n<p>But now get a value error when it creates the Saver() object:<\/p>\n\n<pre><code>ValueError: No variables to save\n<\/code><\/pre>\n","answers":[{"AnswerId":"53453868","CreationDate":"2018-11-23T23:30:57.823","ParentId":null,"OwnerUserId":"4958717","Title":null,"Body":"<p>The <code>tf.train.import_meta_graph<\/code> restores the graph, meaning rebuilds the network architecture that was stored to the file. The call to <code>tf.train.Saver.restore<\/code> on the other hand only restores the variable values from the file to the current graph in the session (this naturally fails if the some values of in the file belong to variables that do not exist in the currently active graph).<\/p>\n\n<p>So if you already build the network layers in the code, you don't need to call <code>tf.train.import_meta_graph<\/code>. Otherwise this might be causing you problems.<\/p>\n\n<p>Not sure how the rest of your code looks like but here are some suggestions. First build the graph, then create the session, and finally restore if applicable. Your init might look like this then<\/p>\n\n<pre><code>def __init__(self, restore=True):\n    self.build_decoder = tf.make_template('decoder', self._build_decoder)\n    self.session = tf.Session()\n    if restore:\n        self.restore_model()\n<\/code><\/pre>\n\n<p>However if you are only restoring the encoder, and building the decoder anew, you might build the decoder last. But then don't forget to initialize its variables before usage.<\/p>\n"}]}
+{"QuestionId":53246624,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-11T07:16:08.267","AcceptedAnswerId":null,"OwnerUserId":1082410.0,"Title":"fashion_mnist.load_data() fails do to force connection closed by remote","Body":"<p>I'm trying to go through the tensorflow tutorial which requires me to execture the line <\/p>\n\n<pre><code>fashion_mnist.load_data()\n<\/code><\/pre>\n\n<p>in my code to download the data set. In the middle of the download I get the following error. <\/p>\n\n<pre><code>ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host\n<\/code><\/pre>\n\n<p>How can I prevent the host from closing the connection. <\/p>\n\n<p>File Being Executed<\/p>\n\n<pre><code>import _thread\nimport time\nimport chess\n\n# TensorFlow and tf.keras\nimport tensorflow as tf\nfrom tensorflow import keras\n\n# Helper libraries\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nprint(tf.__version__)\nfashion_mnist = keras.datasets.fashion_mnist\n\n\nclass_names = ['T-shirt\/top', 'Trouser', 'Pullover', 'Dress', 'Coat', \n               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']\n\n(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()\ntrain_images.shape\n<\/code><\/pre>\n\n<p>Output I Get<\/p>\n\n<pre><code>1.12.0\nDownloading data from https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/t10k-labels-idx1-ubyte.gz\n\n8192\/5148 [===============================================] - 0s 0us\/step\nDownloading data from https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/t10k-images-idx3-ubyte.gz\n\n   8192\/4422102 [..............................] - ETA: 4s\n  57344\/4422102 [..............................] - ETA: 5s\n 147456\/4422102 [&gt;.............................] - ETA: 3s\n 278528\/4422102 [&gt;.............................] - ETA: 2s\n 376832\/4422102 [=&gt;............................] - ETA: 2s\n 385024\/4422102 [=&gt;............................] - ETA: 3s\n 393216\/4422102 [=&gt;..........\n\n\n... [more lines] ... \n\n   ====&gt;........................] - ETA: 42:25\n 5758976\/26421880 [=====&gt;........................] - ETA: 42:32\n 5767168\/26421880 [=====&gt;........................] - ETA: 42:39\n 5775360\/26421880 [=====&gt;........................] - ETA: 42:46\n 5783552\/26421880 [=====&gt;........................] - ETA: 42:54\n 5791744\/26421880 [=====&gt;........................] - ETA: 43:09\n 5799936\/26421880 [=====&gt;........................] - ETA: 43:18\n 5808128\/26421880 [=====&gt;........................] - ETA: 43:26\n 5816320\/26421880 [=====&gt;........................] - ETA: 43:34Traceback (most recent call last):\n  File \"D:\\PiChess\\Core\\hello - Copy.py\", line 20, in &lt;module&gt;\n    (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\datasets\\fashion_mnist.py\", line 52, in load_data\n    paths.append(get_file(fname, origin=base + fname, cache_subdir=dirname))\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\data_utils.py\", line 249, in get_file\n    urlretrieve(origin, fpath, dl_progress)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\urllib\\request.py\", line 277, in urlretrieve\n    block = fp.read(bs)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\http\\client.py\", line 449, in read\n    n = self.readinto(b)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\http\\client.py\", line 493, in readinto\n    n = self.fp.readinto(b)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\socket.py\", line 586, in readinto\n    return self._sock.recv_into(b)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\ssl.py\", line 1009, in recv_into\n    return self.read(nbytes, buffer)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\ssl.py\", line 871, in read\n    return self._sslobj.read(len, buffer)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\ssl.py\", line 631, in read\n    v = self._sslobj.read(len, buffer)\nConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53247146,"AnswerCount":0,"Tags":"<tensorflow><keras><computer-vision><medical>","CreationDate":"2018-11-11T08:50:03.003","AcceptedAnswerId":null,"OwnerUserId":766541.0,"Title":"How to increase validation accuracy on medical images","Body":"<p>I'm working on a medical Xray images dataset trying to do a binary classification.<\/p>\n\n<p>After many tries, I have a found a model that can overfit my training set with > 99% accuracy but from the validation curve look it seems like my model has only learned irrelevant details.\nWhat do you think ?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eRbGf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eRbGf.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/qRsAY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qRsAY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When I try to introduce dropout, training become incredibly slow with bad acc.\n<a href=\"https:\/\/i.stack.imgur.com\/MluX8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MluX8.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/6h0Iu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6h0Iu.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If I try image augmentation, results are more promising but of course much slower.\n<a href=\"https:\/\/i.stack.imgur.com\/eulYb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eulYb.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/G5uOF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G5uOF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I wonder what to look next:<\/p>\n\n<ul>\n<li>try running more epochs on the image augmented model<\/li>\n<li>try some medical pretrained model (do you know where to look)<\/li>\n<\/ul>\n\n<p>What would you use as parameters for image augmentation (preferably in Keras) with Xray images ?<\/p>\n","answers":[]}
+{"QuestionId":53247475,"AnswerCount":4,"Tags":"<python><nlp><pytorch><chatbot>","CreationDate":"2018-11-11T09:43:51.150","AcceptedAnswerId":null,"OwnerUserId":10635543.0,"Title":"Pytorch Chatbot Tutorial problem: How can I solve List Index Out of Range","Body":"<p>I\u2019m new to pytorch and have been following the many tutorials available.\nBut, When I did The <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/chatbot_tutorial.html?highlight=chatbot%20tutorial\" rel=\"nofollow noreferrer\">CHATBOT TUTORIAL<\/a> is not work.<\/p>\n\n<p>Like the figure below<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CzLli.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CzLli.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What should I do and what is causing this?<\/p>\n","answers":[{"AnswerId":"53263215","CreationDate":"2018-11-12T13:28:34.100","ParentId":null,"OwnerUserId":"9475362","Title":null,"Body":"<p>Are you running Windows? Currently most Pytorch tutorial requires Pytorch 1.0 which is currently unavailable on Windows, I'm trying it in a different OS.<\/p>\n\n<p>Will edit the answer as soon as I get it running on Ubuntu 18.<\/p>\n"},{"AnswerId":"53326606","CreationDate":"2018-11-15T19:24:31.313","ParentId":null,"OwnerUserId":"10659493","Title":null,"Body":"<pre><code>def filter_pair(self, p):\n    try:\n        return len(p[0].split(' ')) &lt; self.MAX_LENGTH and len(p[1].split(' '))&lt;\\ self.MAX_LENGTH\n    except:\n        return False\n<\/code><\/pre>\n\n<p>This is because some of pairs are empty list <\/p>\n"},{"AnswerId":"53359535","CreationDate":"2018-11-18T09:42:46.487","ParentId":null,"OwnerUserId":"10667737","Title":null,"Body":"<p>I think the problem is the way the lines are read in resulting in blank lines. You can get over the problem rather crudely by filtering out blank lines. Here is such a crude solution in the readVocs routine.<\/p>\n\n<pre><code># Read query\/response pairs and return a voc object\ndef readVocs(datafile, corpus_name):\n    print(\"Reading lines...\")\n    # Read the file and split into lines\n    lines = open(datafile, encoding='utf-8').\\\n        read().strip().split('\\n')\n    #Now on windows you seem to get alternate blank lines so filter them out.\n    lines2=[]\n    for l in lines:\n        if len(l)&gt;0:\n            lines2.append(l)\n    #And as a check just print the first 10\n    for index, line in enumerate(lines2[:10]):\n        print(index,' - ',line)\n\n    # Split every line into pairs and normalize\n    pairs = [[normalizeString(s) for s in l.split('\\t')] for l in lines2]\n    voc = Voc(corpus_name)\n    return voc, pair\n<\/code><\/pre>\n"},{"AnswerId":"59010767","CreationDate":"2019-11-23T18:19:05.970","ParentId":null,"OwnerUserId":"6454266","Title":null,"Body":"<p>A little late, but I hit this same error at this line before on this same tutorial.  I am also running on 3.6 on Windows, had no issues loading PyTorch or running anything with CUDA.  The issue for me was in the data.<\/p>\n\n<p>It was because there was a blank line somewhere in the source data, so when it went to split the words in the list for that function, it threw this error - there were no words to split.<\/p>\n"}]}
+{"QuestionId":53247534,"AnswerCount":0,"Tags":"<tensorflow><hdf5><tensorflow-datasets><tfrecord>","CreationDate":"2018-11-11T09:51:08.433","AcceptedAnswerId":null,"OwnerUserId":6088463.0,"Title":"TFRecord IO slower than python hdf5 reader, how do I improve its speed?","Body":"<p>I was following the official TF giude to use the <code>tf.data.Dataset<\/code> API for building data pipeline, but I found it ~2 times slower than my python data pipeline using hdf5.<\/p>\n\n<h2>Here's my experiment result:<\/h2>\n\n<pre>\nTFRecordDataset: 660 thousand samples\/second \nhdf5 reader + feed_dict to placeholder: 1.1 million samples\/second\n<\/pre>\n\n<p>My experiment setting is:<br>\n    - batch size=1000,<br>\n    - print log every 10 million samples to show the time,<br>\n    - run a <strong>fake model<\/strong> that only takes data as input and do not compute anything. <\/p>\n\n<p>what's worse is that when I set batch size=10000, it becomes:  <\/p>\n\n<pre>\nTFRecordDataset: 760 thousand samples\/second \nhdf5 reader + feed_dict to placeholder: 2.1 million samples\/second\n<\/pre>\n\n<h2>Here's my code for reading tfrecord<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>def parse_tfrecord_batch(record_batch):\n    FEATURE_LEN = 29\n    dics = {\n        'label': tf.FixedLenFeature(shape=(), dtype=tf.int64),\n        'feature_id': tf.FixedLenFeature(shape=(FEATURE_LEN,), dtype=tf.int64),\n        'feature_val': tf.FixedLenFeature(shape=(FEATURE_LEN,), dtype=tf.float32)\n    }\n    parsed_example = tf.parse_example(record_batch, dics)\n    return parsed_example['feature_id'], parsed_example['feature_val'], parsed_example['label']\n\nclass fakeModel:\n    def __init__(self, train_filenames):\n        self.graph = tf.Graph()\n        with self.graph.as_default():\n            dataset = tf.data.TFRecordDataset(train_filenames)\n            dataset = dataset.repeat()\n            dataset = dataset.batch(1000)\n            dataset = dataset.map(parse_tfrecord_batch, num_parallel_calls=1)\n            dataset = dataset.prefetch(1000)\n            self.iterator = dataset.make_initializable_iterator()\n            self.id, self.wt, self.label = self.iterator.get_next()\n\n            self.train_preds = tf.identity(self.lbl_hldr)\n<\/code><\/pre>\n\n<p>I've tune the <code>num_parallel_calls<\/code> to 2, 10. Not working.  <\/p>\n\n<p>I've also tuned the <code>prefetch(n)<\/code> from 1 to 1000, which has little improvement.  <\/p>\n\n<p>My question is:<br>\nIs there any way to improve my tfrecord data pipeline? Am I missing something in my code?   <\/p>\n\n<p>Appreciate it for any help.<\/p>\n","answers":[]}
+{"QuestionId":53247877,"AnswerCount":1,"Tags":"<keras><neural-network><deep-learning><computer-vision><conv-neural-network>","CreationDate":"2018-11-11T10:34:48.520","AcceptedAnswerId":null,"OwnerUserId":9838554.0,"Title":"Gender Classification -VGG model","Body":"<p>I am using the below code for classifying human gender(M vs F). However its overfitting and val accuracy is even not going 90% .  Need your suggestion in this.<\/p>\n\n<pre><code>img_width, img_height =128,128\n\ntop_model_weights_path = 'bottleneck_fc_model.h5'\ntrain_data_dir = 'Train'\nvalidation_data_dir = 'Test'\nnb_train_samples = 30000\nnb_validation_samples = 7000\nepochs = 150\nbatch_size = 128\n\n\ndef save_bottlebeck_features():\n    datagen = ImageDataGenerator(rescale=1. \/ 255)\n\n    # build the VGG16 network\n    model = applications.VGG16(include_top=False, weights='imagenet')\n\n    generator = datagen.flow_from_directory(\n        train_data_dir,\n        target_size=(img_width, img_height),\n        batch_size=batch_size,\n        class_mode=None,\n        shuffle=False)\n    predict_size_train = int(math.ceil(nb_train_samples \/ batch_size))\n    bottleneck_features_train = model.predict_generator(generator, predict_size_train)\n    np.save('bottleneck_features_train.npy',\n            bottleneck_features_train)\n\n    generator = datagen.flow_from_directory(\n        validation_data_dir,\n        target_size=(img_width, img_height),\n        batch_size=batch_size,\n        class_mode=None,\n        shuffle=False)\n    predict_size_validation = int(math.ceil(nb_validation_samples \/ batch_size))\n    bottleneck_features_validation = model.predict_generator(generator, predict_size_validation)\n    np.save('bottleneck_features_validation.npy',\n            bottleneck_features_validation)\n\n\ndef train_top_model():\n    train_data = np.load('bottleneck_features_train.npy')\n    train_labels = np.array(\n        [0] * (nb_train_samples \/\/ 2) + [1] * (nb_train_samples \/\/ 2))\n\n    validation_data = np.load('bottleneck_features_validation.npy')\n    validation_labels = np.array(\n        [0] * (nb_validation_samples \/\/ 2) + [1] * (nb_validation_samples \/\/ 2))\n\n    model = Sequential()\n    model.add(Flatten(input_shape=train_data.shape[1:]))\n    model.add(Dense(512, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(1, activation='sigmoid'))\n\n    model.compile(optimizer='rmsprop',\n                  loss='binary_crossentropy', metrics=['accuracy'])\n\n    model.fit(train_data, train_labels,\n              epochs=epochs,\n              batch_size=batch_size,\n              validation_data=(validation_data, validation_labels))\n    model.save_weights(top_model_weights_path)\n\n\nsave_bottlebeck_features()\ntrain_top_model()\n<\/code><\/pre>\n\n<p>Here is  some of the last epochs <\/p>\n\n<pre><code>Epoch 130\/150\nloss: 0.0337 - acc: 0.9902 - val_loss: 1.1683 - val_acc: 0.8356\nEpoch 131\/150\nloss: 0.0307 - acc: 0.9919 - val_loss: 1.0721 - val_acc: 0.8345\nEpoch 132\/150\nloss: 0.0313 - acc: 0.9914 - val_loss: 1.1606 - val_acc: 0.8342\nEpoch 133\/150\nloss: 0.0316 - acc: 0.9914 - val_loss: 1.1487 - val_acc: 0.8347\nEpoch 134\/150\nloss: 0.0311 - acc: 0.9909 - val_loss: 1.1363 - val_acc: 0.8356\nEpoch 135\/150\nloss: 0.0295 - acc: 0.9914 - val_loss: 1.2289 - val_acc: 0.8355\nEpoch 136\/150\nloss: 0.0325 - acc: 0.9912 - val_loss: 1.1787 - val_acc: 0.8345\nEpoch 137\/150\nloss: 0.0276 - acc: 0.9922 - val_loss: 1.2281 - val_acc: 0.8337\nEpoch 138\/150\nloss: 0.0314 - acc: 0.9918 - val_loss: 1.1973 - val_acc: 0.8352\nEpoch 139\/150\nloss: 0.0298 - acc: 0.9913 - val_loss: 1.1551 - val_acc: 0.8311\nEpoch 140\/150\nloss: 0.0301 - acc: 0.9919 - val_loss: 1.2301 - val_acc: 0.8339\nEpoch 141\/150\nloss: 0.0315 - acc: 0.9917 - val_loss: 1.1344 - val_acc: 0.8328\nEpoch 142\/150\nloss: 0.0290 - acc: 0.9918 - val_loss: 1.2094 - val_acc: 0.8286\nEpoch 143\/150\nloss: 0.0292 - acc: 0.9919 - val_loss: 1.1449 - val_acc: 0.8358\nEpoch 144\/150\nloss: 0.0284 - acc: 0.9925 - val_loss: 1.2666 - val_acc: 0.8267\nEpoch 145\/150\nloss: 0.0328 - acc: 0.9913 - val_loss: 1.1720 - val_acc: 0.8331\nEpoch 146\/150\nloss: 0.0270 - acc: 0.9928 - val_loss: 1.2077 - val_acc: 0.8355\nEpoch 147\/150\nloss: 0.0338 - acc: 0.9907 - val_loss: 1.2715 - val_acc: 0.8313\nEpoch 148\/150\nloss: 0.0276 - acc: 0.9923 - val_loss: 1.3014 - val_acc: 0.8223\nEpoch 149\/150\nloss: 0.0290 - acc: 0.9923 - val_loss: 1.2123 - val_acc: 0.8291\nEpoch 150\/150\nloss: 0.0317 - acc: 0.9920 - val_loss: 1.2682 - val_acc: 0.8277\n<\/code><\/pre>\n\n<p>Its clearly over fitting and needs more data . However for Cats vs Dogs having 10K data this code works and val accuracy goes above 90% in 4-5 epoch . Need help on this.<\/p>\n","answers":[{"AnswerId":"53256956","CreationDate":"2018-11-12T06:32:24.273","ParentId":null,"OwnerUserId":"6341397","Title":null,"Body":"<p>Try these suggestions (tune until you reach to the wanted result):<\/p>\n\n<ul>\n<li>Change optimizer (try adam) .<\/li>\n<li>Change learning rate (try small learning rate) <\/li>\n<li>Add regularize L2<\/li>\n<li>Add batch normalization layer.<\/li>\n<li>The layers after flatten >> make it as vgg16.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53248099,"AnswerCount":2,"Tags":"<python><machine-learning><keras><generator><image-segmentation>","CreationDate":"2018-11-11T11:07:15.077","AcceptedAnswerId":null,"OwnerUserId":10635788.0,"Title":"Keras: Image segmentation using grayscale masks and ImageDataGenerator class","Body":"<p>I am currently trying to implement a convolutional network using Keras 2.1.6 (with TensorFlow as backend) and its <code>ImageDataGenerator<\/code> to segment an image using a grayscale mask. I try to use an image as input, and a mask as label. Due to a low amount of training images, and memory constraints I utilize the <code>ImageDataGenerator<\/code> class provided in Keras.<\/p>\n\n<p>However I get this error, after changing the values provided in the Keras example to the ones described later:<\/p>\n\n<pre><code>File \"C:\\Users\\XXX\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 2223, in fit_generator\nbatch_size = x.shape[0]\nAttributeError: 'tuple' object has no attribute 'shape'\n<\/code><\/pre>\n\n<p>Which, as far as I know, happens because the generator does generate a tuple, and not an array. This first happened after I changed following parameters from the standard values provided in the Keras example to the following: <code>color_mode='grayscale'<\/code> for all mask generators, and <code>class_mode='input'<\/code> due to this being recommended for autoencoders.<\/p>\n\n<p>The Keras example can be found in <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The dataset I am using consists of 100 images (<code>jpg<\/code>) and 100 corresponding grayscale masks (<code>png<\/code>) and can be downloaded at <a href=\"http:\/\/www.ntu.edu.sg\/home\/asjfcai\/Benchmark_Website\/Semantic%20dataset100.zip\" rel=\"nofollow noreferrer\">this link<\/a><\/p>\n\n<p>The architecture I wanted to implement is an autoencoder\/U-Net based network and it is shown in the provided code:<\/p>\n\n<pre><code>from keras.preprocessing import image\nfrom keras.models import Model\nfrom keras import optimizers \nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D\nfrom keras import initializers\n\nimage_path = \nmask_path = \nvalid_image_path = \nvalid_mask_path = \n\nimg_size=160\nbatchsize=10\nsamplesize = 60\nsteps = samplesize \/ batchsize\n\ntrain_datagen = image.ImageDataGenerator(shear_range=0.2,\n                                        zoom_range=0.2,\n                                        horizontal_flip=True)\n\ndata_gen_args = dict(rotation_range=90,\n                     width_shift_range=0.1,\n                     height_shift_range=0.1,\n                     zoom_range=0.2)\n\nimage_datagen = ImageDataGenerator(**data_gen_args)\nmask_datagen = ImageDataGenerator(**data_gen_args)\n\nseed = 1\n\nimage_generator = image_datagen.flow_from_directory(\n    image_path,\n    target_size=(img_size, img_size),\n    class_mode='input',\n    batch_size = batchsize,\n    seed=seed)\n\nmask_generator = mask_datagen.flow_from_directory(\n    mask_path,\n    target_size=(img_size, img_size),\n    class_mode='input',\n    color_mode = 'grayscale',\n    batch_size = batchsize,\n    seed=seed)\n\nvimage_generator = image_datagen.flow_from_directory(\n    valid_image_path,\n    target_size=(img_size, img_size),\n    class_mode='input',\n    batch_size = batchsize,\n    seed=seed)\n\nvmask_generator = mask_datagen.flow_from_directory(\n    valid_mask_path,\n    target_size=(img_size, img_size),\n    class_mode='input',\n    color_mode = 'grayscale',\n    batch_size = batchsize,\n    seed=seed)\n\n#Model \ninput_img = Input(shape=(img_size,img_size,3))\nc11 = Conv2D(16, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(input_img)\nmp1 = MaxPooling2D((2, 2), padding='same')(c11)\nc21 = Conv2D(16, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(mp1)\nmp2 = MaxPooling2D((2, 2), padding='same')(c21)\nc31 = Conv2D(32, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(mp2)\nencoded = MaxPooling2D((5, 5), padding='same')(c31)\n\nc12 = Conv2D(32, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(encoded)\nus12 = UpSampling2D((5,5))(c12)\nc22 = Conv2D(16, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(us12)\nus22 = UpSampling2D((2, 2))(c22)\nc32 = Conv2D(16, (3, 3), activation='relu', padding='same', kernel_initializer=initializers.random_normal(stddev=0.01))(us22)\nus32 = UpSampling2D((2, 2))(c32)\ndecoded = Conv2D(1, (3, 3), activation='softmax', padding='same')(us32)\n\nmodel = Model(input_img, decoded)\n\nmodel.compile(loss=\"mean_squared_error\", optimizer=optimizers.Adam(),metrics=[\"accuracy\"])\n#model.summary()\n\n#Generators, tr: training, v: validation\ntrgen = zip(image_generator,mask_generator)\nvgen = zip(vimage_generator,vmask_generator)\n\nmodel.fit_generator(\n    trgen,\n    steps_per_epoch= steps,\n    epochs=5,\n    validation_data = vgen,\n    validation_steps=10)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53252258","CreationDate":"2018-11-11T19:14:57.567","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>What you are trying to build is an image segmentation model and not an autoencoder. Therefore, since you have separate generators for the images and the labels (i.e. masks), you need to set the <code>class_mode<\/code> argument to <code>None<\/code> to prevent generator from producing any labels arrays.<\/p>\n\n<p>Further, you need to change the activation function of last layer from <code>softmax<\/code> to <code>sigmoid<\/code>, otherwise since the softmax normalizes the sum of its input elements to 1, the output would be all ones. You can also use <code>binary_crossentropy<\/code> for the loss function as well.<\/p>\n"},{"AnswerId":"53504311","CreationDate":"2018-11-27T16:43:00.533","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p><strong>Here is a better version of Unet, you can use this code<\/strong><\/p>\n\n<pre><code>def conv_block(tensor, nfilters, size=3, padding='same', initializer=\"he_normal\"):\n    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(tensor)\n    x = BatchNormalization()(x)\n    x = Activation(\"relu\")(x)\n    x = Conv2D(filters=nfilters, kernel_size=(size, size), padding=padding, kernel_initializer=initializer)(x)\n    x = BatchNormalization()(x)\n    x = Activation(\"relu\")(x)\n    return x\n\n\ndef deconv_block(tensor, residual, nfilters, size=3, padding='same', strides=(2, 2)):\n    y = Conv2DTranspose(nfilters, kernel_size=(size, size), strides=strides, padding=padding)(tensor)\n    y = concatenate([y, residual], axis=3)\n    y = conv_block(y, nfilters)\n    return y\n\n\ndef Unet(img_height, img_width, nclasses=3, filters=64):\n# down\n    input_layer = Input(shape=(img_height, img_width, 3), name='image_input')\n    conv1 = conv_block(input_layer, nfilters=filters)\n    conv1_out = MaxPooling2D(pool_size=(2, 2))(conv1)\n    conv2 = conv_block(conv1_out, nfilters=filters*2)\n    conv2_out = MaxPooling2D(pool_size=(2, 2))(conv2)\n    conv3 = conv_block(conv2_out, nfilters=filters*4)\n    conv3_out = MaxPooling2D(pool_size=(2, 2))(conv3)\n    conv4 = conv_block(conv3_out, nfilters=filters*8)\n    conv4_out = MaxPooling2D(pool_size=(2, 2))(conv4)\n    conv4_out = Dropout(0.5)(conv4_out)\n    conv5 = conv_block(conv4_out, nfilters=filters*16)\n    conv5 = Dropout(0.5)(conv5)\n# up\n    deconv6 = deconv_block(conv5, residual=conv4, nfilters=filters*8)\n    deconv6 = Dropout(0.5)(deconv6)\n    deconv7 = deconv_block(deconv6, residual=conv3, nfilters=filters*4)\n    deconv7 = Dropout(0.5)(deconv7) \n    deconv8 = deconv_block(deconv7, residual=conv2, nfilters=filters*2)\n    deconv9 = deconv_block(deconv8, residual=conv1, nfilters=filters)\n# output\n    output_layer = Conv2D(filters=nclasses, kernel_size=(1, 1))(deconv9)\n    output_layer = BatchNormalization()(output_layer)\n    output_layer = Activation('softmax')(output_layer)\n\n    model = Model(inputs=input_layer, outputs=output_layer, name='Unet')\n    return model\n<\/code><\/pre>\n\n<p>Note if you have only two classes ie nclasses=2, you need to change<\/p>\n\n<pre><code>output_layer = Conv2D(filters=nclasses, kernel_size=(1, 1))(deconv9)\noutput_layer = BatchNormalization()(output_layer)\noutput_layer = Activation('softmax')(output_layer)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>output_layer = Conv2D(filters=2, kernel_size=(1, 1))(deconv9)\noutput_layer = BatchNormalization()(output_layer)\noutput_layer = Activation('sigmoid')(output_layer)\n<\/code><\/pre>\n\n<p><strong>Now for the data generators, you can use the builtin ImageDataGenerator class\nhere is the code from Keras docs<\/strong><\/p>\n\n<pre><code># we create two instances with the same arguments\ndata_gen_args = dict(featurewise_center=True,\n                     featurewise_std_normalization=True,\n                     rotation_range=90,\n                     width_shift_range=0.1,\n                     height_shift_range=0.1,\n                     zoom_range=0.2)\nimage_datagen = ImageDataGenerator(**data_gen_args)\nmask_datagen = ImageDataGenerator(**data_gen_args)\n\n# Provide the same seed and keyword arguments to the fit and flow methods\nseed = 1\nimage_datagen.fit(images, augment=True, seed=seed)\nmask_datagen.fit(masks, augment=True, seed=seed)\n\nimage_generator = image_datagen.flow_from_directory(\n    'data\/images',\n    class_mode=None,\n    seed=seed)\n\nmask_generator = mask_datagen.flow_from_directory(\n    'data\/masks',\n    class_mode=None,\n    seed=seed)\n\n# combine generators into one which yields image and masks\ntrain_generator = zip(image_generator, mask_generator)\n\nmodel.fit_generator(\n    train_generator,\n    steps_per_epoch=2000,\n    epochs=50)\n<\/code><\/pre>\n\n<p><strong>Another way to go is implement your own generator by extending the Sequence class from Keras<\/strong><\/p>\n\n<pre><code>class seg_gen(Sequence):\n    def __init__(self, x_set, y_set, batch_size, image_dir, mask_dir):\n        self.x, self.y = x_set, y_set\n        self.batch_size = batch_size\n        self.samples = len(self.x)\n        self.image_dir = image_dir\n        self.mask_dir = mask_dir\n\n    def __len__(self):\n        return int(np.ceil(len(self.x) \/ float(self.batch_size)))\n\n    def __getitem__(self, idx):\n        idx = np.random.randint(0, self.samples, batch_size)\n        batch_x, batch_y = [], []\n        drawn = 0\n        for i in idx:\n            _image = image.img_to_array(image.load_img(f'{self.image_dir}\/{self.x[i]}', target_size=(img_height, img_width)))\/255.   \n            mask = image.img_to_array(image.load_img(f'{self.mask_dir}\/{self.y[i]}', grayscale=True, target_size=(img_height, img_width)))\n#             mask = np.resize(mask,(img_height*img_width, classes))\n            batch_y.append(mask)\n            batch_x.append(_image)\n        return np.array(batch_x), np.array(batch_y)\n<\/code><\/pre>\n\n<p><strong>Here is a sample code to train the model<\/strong><\/p>\n\n<pre><code>unet = Unet(256, 256, nclasses=66, filters=64)\nprint(unet.output_shape)\np_unet = multi_gpu_model(unet, 4)\np_unet.load_weights('models-dr\/top_weights.h5')\np_unet.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\ntb = TensorBoard(log_dir='logs', write_graph=True)\nmc = ModelCheckpoint(mode='max', filepath='models-dr\/top_weights.h5', monitor='acc', save_best_only='True', save_weights_only='True', verbose=1)\nes = EarlyStopping(mode='max', monitor='acc', patience=6, verbose=1)\ncallbacks = [tb, mc, es]\ntrain_gen = seg_gen(image_list, mask_list, batch_size)\n\n\np_unet.fit_generator(train_gen, steps_per_epoch=steps, epochs=13, callbacks=callbacks, workers=8)\n<\/code><\/pre>\n\n<p><strong>I got good results when i had only 2 classes by using dice loss, here is the code for it<\/strong><\/p>\n\n<pre><code>def dice_coeff(y_true, y_pred):\n    smooth = 1.\n    y_true_f = K.flatten(y_true)\n    y_pred_f = K.flatten(y_pred)\n    intersection = K.sum(y_true_f * y_pred_f)\n    score = (2. * intersection + smooth) \/ (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)\n    return score\n\ndef dice_loss(y_true, y_pred):\n    loss = 1 - dice_coeff(y_true, y_pred)\n    return loss\n<\/code><\/pre>\n"}]}
+{"QuestionId":53248435,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><time-series>","CreationDate":"2018-11-11T11:48:33.717","AcceptedAnswerId":null,"OwnerUserId":4296481.0,"Title":"LSTM model for time series predictions predicts irregular values like a sawtooth","Body":"<p>I am training a Keras model to predict availability of bike-sharing stations. I am giving in the training set a whole row with day of the year, time, weekday, station and free bikes. Each sample contains the availability for the previous day (144 samples) and I am trying to predict the availability for the next day (144 samples). The shapes for the sets used are<\/p>\n\n<pre><code>Train X (2362, 144, 5)\nTrain Y (2362, 144)\nTest X (39, 144, 5)\nTest Y (39, 144)\nValidation X (1535, 144, 5)\nValidation Y (1535, 144)\n<\/code><\/pre>\n\n<p>The model I am using is this one<\/p>\n\n<pre><code>    model.add(LSTM(20, input_shape=(self.train_x.shape[1], self.train_x.shape[2]), return_sequences = True))\n    model.add(Dropout(0.2))\n    model.add(LSTM(20))\n    model.add(Dense(144))\n    model.compile(loss='mse', optimizer='adam', metrics = ['acc', 'mape', 'mse'])\n\n    history = self.model.fit(self.train_x, self.train_y, batch_size=50, epochs=20, validation_data=(self.validation_x, self.validation_y), verbose=1, shuffle = True)\n<\/code><\/pre>\n\n<p>The predictions made after training have nothing to do with the expected output, they have like a sawtooth shape with values that exceed the original size.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/G0ivi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G0ivi.png\" alt=\"Prediction Sample\"><\/a><\/p>\n\n<p>The accuracy rarely goes up but loss has a normal shape.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XLnt1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XLnt1.png\" alt=\"Accuracy\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jH5Ph.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jH5Ph.png\" alt=\"Loss\"><\/a><\/p>\n\n<p>As an example the history after each epoch looks like this<\/p>\n\n<pre><code>Epoch 17\/20\n2362\/2362 [==============================] - 12s 5ms\/step - loss: 9.1214 - acc: 0.0000e+00 - mean_absolute_percentage_error: 21925846.0813 - mean_squared_error: 9.1214 - val_loss: 9.0642 - val_acc: 0.0000e+00 - val_mean_absolute_percentage_error: 24162847.3779 - val_mean_squared_error: 9.0642\nEpoch 18\/20\n2362\/2362 [==============================] - 12s 5ms\/step - loss: 8.2241 - acc: 0.0013 - mean_absolute_percentage_error: 21906919.9136 - mean_squared_error: 8.2241 - val_loss: 8.1923 - val_acc: 0.0000e+00 - val_mean_absolute_percentage_error: 22754663.8013 - val_mean_squared_error: 8.1923\nEpoch 19\/20\n2362\/2362 [==============================] - 12s 5ms\/step - loss: 7.4190 - acc: 0.0000e+00 - mean_absolute_percentage_error: 21910003.1744 - mean_squared_error: 7.4190 - val_loss: 7.3926 - val_acc: 0.0000e+00 - val_mean_absolute_percentage_error: 24673277.8420 - val_mean_squared_error: 7.3926\nEpoch 20\/20\n2362\/2362 [==============================] - 12s 5ms\/step - loss: 6.7067 - acc: 0.0013 - mean_absolute_percentage_error: 22076339.2168 - mean_squared_error: 6.7067 - val_loss: 6.6758 - val_acc: 6.5147e-04 - val_mean_absolute_percentage_error: 22987089.8436 - val_mean_squared_error: 6.6758\n<\/code><\/pre>\n\n<p>I really don't know where the problem might be, more layers?, less layers?, different approach?<\/p>\n\n<p><strong>UPDATE: Plots of training\/test data<\/strong>. Left part of the plot shows the previous day of availability that is fed to the model and the right part shows what the result should be and the prediction made by the model.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FuEIb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FuEIb.png\" alt=\"Predicted\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53249341,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-11T13:42:50.150","AcceptedAnswerId":null,"OwnerUserId":10635304.0,"Title":"ValueError: Cannot feed value of shape (100,) for Tensor 'Placeholder_1:0', which has shape '(?, 10)'","Body":"<p>I'm newer to tensorflow, I implement the logistic regression using tensorflow, my code is same as the book \"Getting started with TensorFlow\", but when I run it, occured error: ValueError: Cannot feed value of shape (100,) for Tensor 'Placeholder_1:0', which has shape '(?, 10)'. Could anyone help me?<\/p>\n\n<pre><code>mnist = input_data.read_data_sets(\"tmp\/data\/\", one_hot=False)\n\n# set the total number of epochs of the training phase:\ntraining_epochs = 25\nlearning_rate = 0.01\nbatch_size = 100\ndisplay_step = 1\n\n#                           building the model\n# define 'x' as the input tensor, it represents the MNIST data\n# images of size 28 x 28 = 784 pixels:\nx = tf.placeholder(\"float\", [None, 784])\ny = tf.placeholder(\"float\", [None, 10])\nW = tf.Variable(tf.zeros([784, 10]))\nb = tf.Variable(tf.zeros([10]))\nevidence = tf.matmul(x, W) + b\n\nactivation = tf.nn.softmax(evidence)\n# error function\ncross_entropy = y * tf.lgamma(activation)\n# cost function\ncost = tf.reduce_mean(-tf.reduce_sum(cross_entropy, reduction_indices = 1))\n# minimize the cost function\noptimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)\n\navg_set = []\nepoch_set = []\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    for epoch in range(training_epochs):\n        avg_cost = 0\n        total_batch = int(mnist.train.num_examples \/ batch_size)\n        for i in range(total_batch):\n            batch_xs, batch_ys = mnist.train.next_batch(batch_size)\n            # Fit the training using the batch data:\n            sess.run(optimizer, feed_dict = {x: batch_xs, y: batch_ys})\n            # compute the average loss:\n            avg_cost += sess.run(cost, feed_dict = {x: batch_xs,\n                                                    y: batch_ys}) \/ total_batch\n        # During the computation, we display a log per epoch step:\n        if epoch % display_step == 0:\n            print(\"Epoch: \", '%04d' % (epoch + 1),\n                  \"cost=\", \"{:.9f}\".format(avg_cost))\n    print(\"Training phase finished.\")\n\n    correct_prediction = tf.equal(tf.argmax(activation, 1), tf.argmax(y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n    print(\"MODEL accuracy: \", accuracy.eval({x: mnist.test.images,\n                                             y: mnist.test.labels}))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53249386,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network><valueerror>","CreationDate":"2018-11-11T13:48:08.887","AcceptedAnswerId":null,"OwnerUserId":10636279.0,"Title":"ValueError: Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3","Body":"<p>After inquiring into the questions already asked about this problem, I keep presenting it. Im trying to classify letters from A to D. All input images are 64x64 and graycolor. <\/p>\n\n<p>The first layer of my CNN is:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3,  3), input_shape = input_shape, activation = 'relu'))\n<\/code><\/pre>\n\n<p>And <code>input_shape<\/code> it's coming from:<\/p>\n\n<pre><code># Define the number of classes\nnum_classes = 4\nlabels_name={'A':0,'B':1,'C':2,'D':3}\n\nimg_data_list=[]\nlabels_list=[]\n\nfor dataset in data_dir_list:\n    img_list=os.listdir(data_path+'\/'+ dataset)\n    print ('Loading the images of dataset-'+'{}\\n'.format(dataset))\n    label = labels_name[dataset]\n    for img in img_list:\n    input_img=cv2.imread(data_path + '\/'+ dataset + '\/'+ img )\n          input_img=cv2.cvtColor(input_img, cv2.COLOR_BGR2GRAY)\n          input_img_resize=cv2.resize(input_img,(128,128))\n          img_data_list.append(input_img_resize)\n          labels_list.append(label)\n\nimg_data = np.array(img_data_list)\nimg_data = img_data.astype('float32')\nimg_data \/= 255\nprint (img_data.shape)\n\nlabels = np.array(labels_list)\nprint(np.unique(labels,return_counts=True))\n\n#convert class labels to on-hot encoding\nY = np_utils.to_categorical(labels, num_classes)\n\n#Shuffle the dataset\nx,y = shuffle(img_data,Y, random_state=2)\n\n# Split the dataset\nX_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=2)\n\n#Defining the model\ninput_shape=img_data[0].shape\nprint(input_shape)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53252796","CreationDate":"2018-11-11T20:14:42.250","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>Conv2d expects input of  shape (batchsize, w, h, filters). <br><br><\/p>\n\n<p>You need to add a reshape to fit the data before the conv layer:<br><\/p>\n\n<pre><code> model.add(Reshape((64, 64, 1)))\n<\/code><\/pre>\n\n<p>This will set your model dimensions to [None, 64,64,1] and should be fine for Conv2d.<\/p>\n"},{"AnswerId":"54037025","CreationDate":"2019-01-04T10:23:19.643","ParentId":null,"OwnerUserId":"10738302","Title":null,"Body":"<p>The CNN model needs a data set with dimensions greater when used multiple layer(Convolutional layers and pooling). To avoid the negative dimensional problems increase the image dimensions or reduce the CNN layers. It works..<\/p>\n"}]}
+{"QuestionId":53249556,"AnswerCount":0,"Tags":"<tensorflow><gpu>","CreationDate":"2018-11-11T14:09:45.767","AcceptedAnswerId":null,"OwnerUserId":10612886.0,"Title":"tensorflow-gpu error: Could not install packages due to an EnvironmentError: [Errno 2] No such file or directory:","Body":"<p>Trying to install tensorflow-gpu version 1.8.<\/p>\n\n<p>Already installed NVIDIA,CUDA and cudnn.\nI'm using windows 10 and python 3.\ncreated an environment.<\/p>\n\n<p>When trying:<\/p>\n\n<pre><code>pip install tensorflow-gpu==1.8\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<blockquote>\n  <p>Collecting tensorflow-gpu==1.8   Using cached\n  <a href=\"https:\/\/files.pythonhosted.org\/packages\/42\/a8\/4c96a2b4f88f5d6dfd70313ebf38de1fe4d49ba9bf2ef34dc12dd198ab9a\/tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/files.pythonhosted.org\/packages\/42\/a8\/4c96a2b4f88f5d6dfd70313ebf38de1fe4d49ba9bf2ef34dc12dd198ab9a\/tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl<\/a>\n  Requirement already satisfied: gast>=0.2.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (0.2.0) Requirement already satisfied: protobuf>=3.4.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (3.5.2) Requirement already satisfied: wheel>=0.26 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (0.32.1) Requirement already satisfied: termcolor>=1.1.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (1.1.0) Requirement already satisfied: absl-py>=0.1.6 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (0.5.0) Requirement already satisfied: astor>=0.6.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (0.7.1) Collecting tensorboard&lt;1.9.0,>=1.8.0 (from\n  tensorflow-gpu==1.8)   Using cached\n  <a href=\"https:\/\/files.pythonhosted.org\/packages\/59\/a6\/0ae6092b7542cfedba6b2a1c9b8dceaf278238c39484f3ba03b03f07803c\/tensorboard-1.8.0-py3-none-any.whl\" rel=\"nofollow noreferrer\">https:\/\/files.pythonhosted.org\/packages\/59\/a6\/0ae6092b7542cfedba6b2a1c9b8dceaf278238c39484f3ba03b03f07803c\/tensorboard-1.8.0-py3-none-any.whl<\/a>\n  Requirement already satisfied: grpcio>=1.8.6 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (1.12.1) Requirement already satisfied: six>=1.10.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (1.11.0) Requirement already satisfied: numpy>=1.13.3 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from tensorflow-gpu==1.8)\n  (1.15.2) Requirement already satisfied: setuptools in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from\n  protobuf>=3.4.0->tensorflow-gpu==1.8) (39.1.0) Requirement already\n  satisfied: html5lib==0.9999999 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from\n  tensorboard&lt;1.9.0,>=1.8.0->tensorflow-gpu==1.8) (0.9999999)\n  Requirement already satisfied: werkzeug>=0.11.10 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from\n  tensorboard&lt;1.9.0,>=1.8.0->tensorflow-gpu==1.8) (0.14.1) Requirement\n  already satisfied: markdown>=2.6.8 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from\n  tensorboard&lt;1.9.0,>=1.8.0->tensorflow-gpu==1.8) (2.6.8) Requirement\n  already satisfied: bleach==1.5.0 in\n  c:\\users\\sarit\\anaconda3\\lib\\site-packages (from\n  tensorboard&lt;1.9.0,>=1.8.0->tensorflow-gpu==1.8) (1.5.0)\n  <strong><em>Could not install packages due to an EnvironmentError: [Errno 2] No such file or directory:\n  'c:\\users\\sarit\\anaconda3\\lib\\site-packages\\tensorflow-1.11.0.dist-info\\METADATA'<\/em><\/strong><\/p>\n<\/blockquote>\n\n<p>Couldn't find a solution for that. Be happy to get some help<\/p>\n","answers":[]}
+{"QuestionId":53249636,"AnswerCount":1,"Tags":"<python><keras><sentiment-analysis>","CreationDate":"2018-11-11T14:20:34.453","AcceptedAnswerId":53256020.0,"OwnerUserId":9417884.0,"Title":"Workaround for python MemoryError","Body":"<p>How can I change this function to make it more efficient? I keep getting MemoryError<\/p>\n\n<pre><code>def vectorize_sequences(sequences, dimension=10000):\n    results = np.zeros((len(sequences), dimension))\n    for i, sequence in enumerate(sequences):\n        results[i, sequence] = 1.\n    return results\n<\/code><\/pre>\n\n<p>I call the function here:<\/p>\n\n<pre><code>x_train = vectorize_sequences(train_data)\nx_test = vectorize_sequences(test_data)\n<\/code><\/pre>\n\n<p>Train and Test data are IMDB dataset for sentiment analysis, i.e.<\/p>\n\n<pre><code>(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)\n<\/code><\/pre>\n\n<p>EDIT: I am running this on 64 bit Ubuntu system with 4 GB RAM.<\/p>\n\n<p>Here is the Traceback:<\/p>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"\/home\/uttam\/PycharmProjects\/IMDB\/imdb.py\", line 29, in &lt;module&gt;\n    x_test = vectorize_sequences(test_data)\n  File \"\/home\/uttam\/PycharmProjects\/IMDB\/imdb.py\", line 20, in vectorize_sequences\n    results = np.zeros((len(sequences), dimension))\nMemoryError\n<\/code><\/pre>\n","answers":[{"AnswerId":"53256020","CreationDate":"2018-11-12T04:34:46.117","ParentId":null,"OwnerUserId":"4323","Title":null,"Body":"<p>Your array appears to be 10k x 10k which is 100 million elements of 64 bits each (because the default dtype is float64).  So that's 800 million bytes, aka 763 megabytes.<\/p>\n\n<p>If you use float32 it will cut the memory usage in half:<\/p>\n\n<pre><code>np.zeros((len(sequences), dimension), dtype=np.float32)\n<\/code><\/pre>\n\n<p>Or if you only care about 0 and 1, this will cut it by 88%:<\/p>\n\n<pre><code>np.zeros((len(sequences), dimension), dtype=np.int8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53249807,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2018-11-11T14:42:42.297","AcceptedAnswerId":53249987.0,"OwnerUserId":10492592.0,"Title":"ValueError: Cannot feed value of shape (2, 4) for Tensor u'InputData\/X:0', which has shape '(?, 2, 4","Body":"<p>I got an error,ValueError: Cannot feed value of shape (2, 4) for Tensor u'InputData\/X:0', which has shape '(?, 2, 4, 104)'.\nI wrote codes,<\/p>\n\n<pre><code># coding: utf-8\nimport tensorflow as tf\nimport tflearn\n\nfrom tflearn.layers.core import input_data,dropout,fully_connected\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.normalization import local_response_normalization\nfrom tflearn.layers.estimator import regression\n\nimport pandas as pd\nimport numpy as np\nfrom sklearn import metrics\n\ntf.reset_default_graph()\nnet = input_data(shape=[2, 4, 104])\nnet = conv_2d(net, 4, 16, activation='relu')\nnet = max_pool_2d(net, 1)\nnet = tflearn.activations.relu(net)\nnet = dropout(net, 0.5)\nnet = tflearn.fully_connected(net, 10, activation='softmax')\nnet = tflearn.regression(net, optimizer='adam', learning_rate=0.5, loss='categorical_crossentropy')\n\nmodel = tflearn.DNN(net)\n\ntrainDataSet = [[0.25,0.25,1,1],[0,0,1,1],[0.25,0.25,1,1]]\ntrainLabel = [[0,1],[0,1],[1,0]]\nmodel.fit(trainDataSet, trainLabel, n_epoch=100, batch_size=32, validation_set=0.1, show_metric=True)\n<\/code><\/pre>\n\n<p>Traceback says<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"cnn.py\", line 16, in &lt;module&gt;\n    model.fit(trainDataSet, trainLabel, n_epoch=100, batch_size=32, validation_set=0.1, show_metric=True)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tflearn\/models\/dnn.py\", line 216, in fit\n    callbacks=callbacks)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tflearn\/helpers\/trainer.py\", line 339, in fit\n    show_metric)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tflearn\/helpers\/trainer.py\", line 818, in _train\n    feed_batch)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1128, in _run\n    str(subfeed_t.get_shape())))\nValueError: Cannot feed value of shape (2, 4) for Tensor u'InputData\/X:0', which has shape '(?, 2, 4, 104)'\n<\/code><\/pre>\n\n<p>I rewrote into<\/p>\n\n<pre><code>trainDataSet = np.array([[0.25,0.25,1,1],[0,0,1,1],[0.25,0.25,1,1]])\ntrainLabel = np.array([[0,1],[0,1],[1,0]])\n<\/code><\/pre>\n\n<p>but same error happens.What is wrong in my codes?How should I fix this?<\/p>\n","answers":[{"AnswerId":"53249987","CreationDate":"2018-11-11T15:01:34.210","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>Citing from the <a href=\"http:\/\/tflearn.org\/layers\/conv\/\" rel=\"nofollow noreferrer\">Tensorflow<\/a> documentation: <\/p>\n\n<pre><code>tflearn.layers.conv.conv_2d \n<\/code><\/pre>\n\n<blockquote>\n  <p><strong>Input:<\/strong><\/p>\n  \n  <p><strong>4-D Tensor<\/strong> [batch, height, width, in_channels].<\/p>\n<\/blockquote>\n\n<p>From other Tensorflow documentation:<\/p>\n\n<pre><code>tf.nn.conv2d\n<\/code><\/pre>\n\n<blockquote>\n  <p>Computes a 2-D convolution given <strong>4-D input<\/strong> and filter tensors.<\/p>\n  \n  <p>Given an input tensor of shape [batch, in_height, in_width,\n  in_channels] and a filter \/ kernel tensor of shape [filter_height,\n  filter_width, in_channels, out_channels], this op performs the\n  following:<\/p>\n<\/blockquote>\n\n<p>Your dataset, label and input shape aren't aligning i.e. doesn't fit to each other.<\/p>\n\n<p>Currently your <code>trainDataSet<\/code> has the shape of (3,4):<\/p>\n\n<pre><code>import numpy as np\ntrainDataSet = np.array([[0.25,0.25,1,1],[0,0,1,1],[0.25,0.25,1,1]])\nprint(trainDataSet.shape)\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>(3, 4)\n<\/code><\/pre>\n\n<p>But you defined the input shape as:<\/p>\n\n<pre><code>net = input_data(shape=[2, 4, 104])\n<\/code><\/pre>\n\n<p>Ambiguous what you really want to achieve, but if you wanted to see a simple working example, that your code should have been seen as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tflearn\n\nfrom tflearn.layers.core import input_data,dropout,fully_connected\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.normalization import local_response_normalization\nfrom tflearn.layers.estimator import regression\n\nimport pandas as pd\nimport numpy as np\nfrom sklearn import metrics\n\ntf.reset_default_graph()\nnet = input_data(shape=[3, 4, 1])\nnet = conv_2d(net, 4, 16, activation='relu')\nnet = max_pool_2d(net, 1)\nnet = tflearn.activations.relu(net)\nnet = dropout(net, 0.5)\nnet = tflearn.fully_connected(net, 2, activation='softmax')\nnet = tflearn.regression(net, optimizer='adam', learning_rate=0.5, loss='categorical_crossentropy')\n\nmodel = tflearn.DNN(net)\n\ntrainDataSet = [\n    [\n        [[0.25], [0.25], [1], [1]],\n        [[0], [0], [1], [1]],\n        [[0.25], [0.25], [1], [1]]\n    ],\n    [\n        [[0.25], [0.25], [1], [1]],\n        [[0], [0], [1], [1]],\n        [[0.25], [0.25], [1], [1]]\n    ],\n    [\n        [[0.25], [0.25], [1], [1]],\n        [[0], [0], [1], [1]],\n        [[0.25], [0.25], [1], [1]]\n    ]\n]\n\ntrainLabel = [[0,1],[0,1],[1,0]]\nmodel.fit(trainDataSet, trainLabel, n_epoch=100, batch_size=32, validation_set=0.1, show_metric=True)\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>---------------------------------\nRun id: NHHJV7\nLog directory: \/tmp\/tflearn_logs\/\nINFO:tensorflow:Summary name Accuracy\/ (raw) is illegal; using Accuracy\/__raw_ instead.\n---------------------------------\nTraining samples: 2\nValidation samples: 1\n--\nTraining Step: 1  | time: 1.160s\n| Adam | epoch: 001 | loss: 0.00000 - acc: 0.0000 | val_loss: 23.02585 - val_acc: 0.0000 -- iter: 2\/2\n--\nTraining Step: 2  | total loss: 0.62966 | time: 1.008s\n| Adam | epoch: 002 | loss: 0.62966 - acc: 0.0000 | val_loss: 10.76885 - val_acc: 0.0000 -- iter: 2\/2\n.\n.\n.\nTraining Step: 99  | total loss: 0.00000 | time: 1.013s\n| Adam | epoch: 099 | loss: 0.00000 - acc: 1.0000 | val_loss: 23.02585 - val_acc: 0.0000 -- iter: 2\/2\n--\nTraining Step: 100  | total loss: 0.00000 | time: 1.011s\n| Adam | epoch: 100 | loss: 0.00000 - acc: 1.0000 | val_loss: 23.02585 - val_acc: 0.0000 -- iter: 2\/2\n--\n<\/code><\/pre>\n"}]}
+{"QuestionId":53249988,"AnswerCount":0,"Tags":"<tensorflow><image-preprocessing><data-augmentation>","CreationDate":"2018-11-11T15:01:53.493","AcceptedAnswerId":null,"OwnerUserId":7886651.0,"Title":"What does \"color ordering\" mean when used with distorted inputs with inception model","Body":"<p>I have been looking for pre-processing techniques of input images. After looking into the implementation of the inception model at: <a href=\"https:\/\/github.com\/awslabs\/deeplearning-benchmark\/blob\/master\/tensorflow\/inception\/inception\/image_processing.py#L164\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/deeplearning-benchmark\/blob\/master\/tensorflow\/inception\/inception\/image_processing.py#L164<\/a>, I found that they have used <code>color_ordering<\/code> parameter which could be 0 or 1. What does color ordering mean in this context?<\/p>\n","answers":[]}
+{"QuestionId":53250206,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-11T15:26:44.513","AcceptedAnswerId":null,"OwnerUserId":9491737.0,"Title":"initializing tensor of zeors and ones","Body":"<p>I have a tensor that gives the row indices of rows that I would like to be initialized to ones. It looks something like this:<\/p>\n\n<pre><code>[0 1 5 9]\n<\/code><\/pre>\n\n<p>For all other rows, I would like these to be initialized to zero. That is, I would like to initialize a rank 2 tensor where the rows 0, 1, 5, 9 are ones, whilst the rest are zeros. <\/p>\n\n<p>I have looked at the initialize functions such as ones, zeros, and fill, but I am not having much luck creating a tensor of this specification. <\/p>\n\n<p>I realise that I could evaluate this tensor to have it returned as a numpy array, and the use numpy to create a matrix of this form, but this does not fit within the structure of my code. I have tried to make it so I can do this, but I would have to remove placeholders and instead use numpy arrays within my tensorflow graph, which is not good practice. <\/p>\n\n<p>Is there a function, or particular way where I would be able to creatr a tensor as above?<\/p>\n","answers":[{"AnswerId":"53250347","CreationDate":"2018-11-11T15:41:47.990","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>If you are trying to assign values to a tensor (<code>t<\/code>) based on the values within another tensor (<code>t_0159<\/code>) i believe the correct method would be to evaluate <code>t_0159<\/code> within the session, and use the values returned to call a <code>tf.assign<\/code> operation to overwrite the content of <code>t<\/code> with values passed to a placeholder. Im not sure if this is what you described above, but it should not require any <code>numpy<\/code> arrays <strong>within<\/strong> the graph, just during the session. <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# create placeholder for raplacement variables\nvalues = tf.placeholder(dtype = tf.float32, shape = [10,10], name=\"values\")\n\n# sample tensors, one containing row indeces, other is arbitrary Variable\nt_0159 = tf.constant([0,1,5,9])\nt = tf.Variable(tf.zeros([10,10]), name = \"my_tensor\")\n\n# define assignment operation to overwrite tensor t\nassignment_op = tf.assign(t, values, name = \"assignment_op\")\n\n\n# run session and call assignment_op\nwith tf.Session() as sess:\n    # initialize variables and print tensor state\n    sess.run(tf.global_variables_initializer())\n    print(tf.get_default_graph().get_tensor_by_name(\"my_tensor:0\").eval())\n\n    # get indeces\n    r_indeces = sess.run(t_0159)\n\n    # create replacement tensor in numpy\n    assignment_value = np.zeros([10,10])\n    for i in r_indeces:\n        assignment_value[i,:]= 1\n\n    # run assignment op and printe new tensor state\n    sess.run(assignment_op, feed_dict = {\"values:0\": assignment_value})\n    print(tf.get_default_graph().get_tensor_by_name(\"my_tensor:0\").eval())\n<\/code><\/pre>\n\n<p>I believe this is the \"proper\" way to do it, and produces an output:<\/p>\n\n<pre><code>[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]\n[[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]\n [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]]\n<\/code><\/pre>\n\n<p>I hope this helps, if not let me know. <\/p>\n"},{"AnswerId":"53251124","CreationDate":"2018-11-11T17:06:41.117","ParentId":null,"OwnerUserId":"8733347","Title":null,"Body":"<p>Here is one way (maybe not the best) to do it.\nSay that you want to create a tensor, with shape <code>(10, n_cols)<\/code>, with the constraint that the rows with indices in <code>rows=[0, 1, 5, 9]<\/code> are initialized to ones and the other ones are initialized to zeros.<\/p>\n\n<pre><code>import tensorflow as tf\n\nrows = [0, 1, 5, 9]\nn_cols = 10  # for the sake of the example...\nmy_tensor = [tf.get_variable(name='my_tensor_row_' + str(idx), \n                             shape=(1, n_cols),\n                             initializer=tf.initializers.ones if idx in rows else tf.initalizers.zeros) \n             for idx in range(10)]\nmy_tensor = tf.concat(my_tensor, axis=0)\n<\/code><\/pre>\n\n<p>Then, you can access the value of <code>my_tensor<\/code> in a tensorflow Session:<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.global_variables_initializer()) # initialize the rows of `my_tensor`\nmy_tensor_val = sess.run(my_tensor)\n<\/code><\/pre>\n\n<p>will give you a Numpy array <code>my_tensor_val<\/code> with shape <code>(10, n_cols)<\/code> which is initialized as you wanted.<\/p>\n"}]}
+{"QuestionId":53250360,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-11T15:43:06.447","AcceptedAnswerId":null,"OwnerUserId":9137167.0,"Title":"TypeError: non_max_suppression() got an unexpected keyword argument 'score_threshold'","Body":"<p>Hi I am using win 7 64 bit and tensorflow version 1.5 I've tried 1.9 and higher but isnt work and I've tried tensorflow-gpu version but again isnt work all the error this <\/p>\n","answers":[{"AnswerId":"54779304","CreationDate":"2019-02-20T05:18:50.130","ParentId":null,"OwnerUserId":"8029140","Title":null,"Body":"<p>I encountered same issue using tf 1.8. Tensorflow versions &lt; 1.9 did not support the <code>score_threshold<\/code> param.\nNeed to be sure you're using version 1.9 or newer. <\/p>\n"}]}
+{"QuestionId":53250533,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-11T16:01:51.623","AcceptedAnswerId":53252006.0,"OwnerUserId":7275118.0,"Title":"How to use tile function in Keras?","Body":"<p>I want to build a neural network with Keras,but I got a error:\n<code>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'<\/code>,and this is my example code:<\/p>\n\n<pre><code>from keras.layers.merge import concatenate\n\nimg = Input(shape=(64,64,3))\ntext_input = Input(shape=(192,))\ntext_emb = Reshape(target_shape=(1, 1, 256))(Dense(256, activation='relu')(text_input))\ntiled_emb = keras.backend.tile(text_emb, (-1, 64, 64, 1))\nimg_feat = Conv2D(400,4,padding='same')(img)\ncon = concatenate([tiled_emb,img_feat])\nconv4 = Conv2D(512, 1)(con)\n\nflat = Flatten()(conv4)\nvalidity = Dense(1, activation='sigmoid')(flat)\n\nModel([img, text_input], validity)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53252006","CreationDate":"2018-11-11T18:47:39.293","ParentId":null,"OwnerUserId":"5536887","Title":null,"Body":"<p>This error occurs because keras.backend.tile is a function and not a layer, making tiled_emb a tensor. The error is then generated when trying to construct the network and encountering just a tensor where it expects a layer (so the attr _inbound_nodes is not defined).<\/p>\n\n<p>You can turn any function into a layer by using the <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\">keras.layers.lambda<\/a> layer, eg:<\/p>\n\n<pre><code>tiled_emb = Lambda(keras.backend.tile, arguments={'n':(-1, 64, 64, 1)})(text_emb)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53251060,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2018-11-11T16:59:18.390","AcceptedAnswerId":null,"OwnerUserId":5346264.0,"Title":"Tensorflow on Ubuntu","Body":"<p>I tried to run keras in ubuntu 14.04 LTS.<\/p>\n\n<pre><code>import tensorflow as tf\nmnist = tf.keras.datasets.mnist\n<\/code><\/pre>\n\n<p>Installed:<\/p>\n\n<pre><code>tensorflow   - 1.12.0\nKeras        - 2.2.4\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>F tensorflow\/core\/platform\/cpu_feature_guard.cc:37] The TensorFlow\nlibrary was compiled to use SSE4.1 instructions, but these aren't\navailable on your machine. Aborted (core dumped)\n<\/code><\/pre>\n\n<p>How can I resolve the above issue?<\/p>\n","answers":[]}
+{"QuestionId":53251390,"AnswerCount":0,"Tags":"<python><tensorflow><keras><rnn>","CreationDate":"2018-11-11T17:35:11.177","AcceptedAnswerId":null,"OwnerUserId":7886651.0,"Title":"Using Native tensorflow RNNLayer with dropout within keras model","Body":"<p>I have a model implemented in Keras, but I need to implement the same model in tensorflow. So, I am looking to implement only the RNN layer of the model and keep the rest the same, that is, the prediction method, fitting the model... are all implemented in keras. Therefore, here is the code:<\/p>\n\n<p>Keras model:<\/p>\n\n<pre><code>def emotion_model(max_seq_len, num_features, learning_rate, num_units_1, num_units_2, bidirectional, dropout, num_targets):\n    # Input layer\n    inputs = Input(shape=(max_seq_len, num_features))\n\n    # 1st layer\n    net = LSTM(num_units_1, return_sequences=True, dropout=dropout, recurrent_dropout=dropout)(net)\n\n    # 2nd layer\n    net = LSTM(num_units_2, return_sequences=True, dropout=dropout, recurrent_dropout=dropout)(net)\n\n    # Output layer\n    outputs = []\n    out1 = TimeDistributed(Dense(1))(net)  # linear activation\n    outputs.append(out1)\n    if num_targets &gt;= 2:\n        out2 = TimeDistributed(Dense(1))(net)  # linear activation\n        outputs.append(out2)\n    if num_targets == 3:\n        out3 = TimeDistributed(Dense(1))(net)  # linear activation\n        outputs.append(out3)\n\n    # Create and compile model\n    rmsprop = RMSprop(lr=learning_rate)\n    model   = Model(inputs=inputs, outputs=outputs)\n    model.compile(optimizer=rmsprop, loss=ccc_loss)  # CCC-based loss function\n    return model\n<\/code><\/pre>\n\n<p>Now, I would like to replace the LSTM layers above with the equivalent code in tensorflow. Therefore, in a different Module I have implemented the following:<\/p>\n\n<pre><code>def baseline_model(inputs, cell_Size1, cell_Size2, dropout):\n    with tf.variable_scope('model', reuse=tf.AUTO_REUSE):\n        cell1 = tf.nn.rnn_cell.LSTMCell(cell_Size1)\n        cell1 = tf.nn.rnn_cell.DropoutWrapper(cell1, input_keep_prob=1.0 - dropout, state_keep_prob=1.0 - dropout)\n\n        cell2 = tf.nn.rnn_cell.LSTMCell(cell_Size2)\n        cell2 = tf.nn.rnn_cell.DropoutWrapper(cell2, input_keep_prob=1.0 - dropout, state_keep_prob=1.0 - dropout)\n\n        cell = tf.nn.rnn_cell.MultiRNNCell([cell1, cell2], state_is_tuple=True)\n\n        # output1: shape=[1, time_steps, 32]\n        output, new_state = tf.nn.dynamic_rnn(cell, inputs, dtype=tf.float32)\n\n        return output\n<\/code><\/pre>\n\n<p>I have tried <code>net = Lambda(partial(baseline_model, dropout))(net)<\/code> where I removed the <code>cell_size1<\/code> and <code>cell_size2<\/code> from the method \"baseline_model\" arguments, yet didn't work<\/p>\n\n<p>Second, I have tried dumping directly the LSTM layers implemented in tensorflow instead of the <code>LSTM<\/code> layers in keras above, and this doesn't solve my problem.<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[]}
+{"QuestionId":53251422,"AnswerCount":0,"Tags":"<lstm><keras>","CreationDate":"2018-11-07T16:31:37.527","AcceptedAnswerId":null,"OwnerUserId":8456120.0,"Title":"LSTM Autoencoder Keras variable batch size","Body":"<p>I have a collection of sequences.  I would like to encode them as vectors of length <code>encoded_dim<\/code>.  To that end, I made an autoencoder using stacked LSTMs in Keras:<\/p>\n\n<pre><code>inputs = Input(shape=(None, input_dim))\n\nencoded = LSTM(encoded_dim)(LSTM(latent_dim, return_sequences=True)(inputs))\n\ndecoded = RepeatVector(timesteps)(encoded)\ndecoded = Dense(8)(Dense(100)(LSTM(input_dim, return_sequences=True)(decoded)))\n\nsequence_autoencoder = Model(inputs, decoded)\n\nsequence_autoencoder.compile(optimizers.Adam(lr=0.01), loss='mean_squared_error')\nsequence_autoencoder.fit(input, input, epochs=3000, batch_size=16, shuffle=False, callbacks=[...], validation_split=.2)\n<\/code><\/pre>\n\n<p>This works well when I restrict my attention to sequences of length <code>timesteps<\/code>.  In general, however, my sequences are not the same length, so <code>RepeatVector(timesteps)<\/code> won't work as is. (This will also mess up batching, but I have had decent training with <code>batch_size=1<\/code> so I'm not too worried about that for my particular use case.)<\/p>\n\n<p>There is a fruitful conversation <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6776\" rel=\"nofollow noreferrer\">here<\/a> involving variable length input sequences and batching, in the context of general seq2seq modeling.  Unfortunately, there is no concern there for retrieving the fixed size encoding.<\/p>\n\n<p>Any suggestions\/code snippets are greatly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53251827,"AnswerCount":1,"Tags":"<tensorflow><split><keras><reusability><pre-trained-model>","CreationDate":"2018-11-11T18:25:16.553","AcceptedAnswerId":53267270.0,"OwnerUserId":1919956.0,"Title":"Pretrained Tensorflow model RGB -> RGBY channel extension","Body":"<p>I am working on the <a href=\"https:\/\/www.kaggle.com\/c\/human-protein-atlas-image-classification\/data\" rel=\"nofollow noreferrer\">protein analysis project<\/a>. We receive the <a href=\"https:\/\/www.kaggleusercontent.com\/kf\/7215060\/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..LPZ9_xvlaqTrcG2SKlup1g.cLeDCrM3KtoP5P8ST0YZ92L0LYO9Jkg4HH0-kcrOivqGrs3nn35sgxLIMGVabuY6d-DNJ8qhHukxM8kEqFz3qP_kcC2mE-sxsywBuKsXTVg9EmHSddJT_4U_zAg4-ZUPS23LLUI4Gid5baxgWNGjmmsExp6wjs_uBbZJowwKfps.kxXJmjlVMWA1ROgOiCcKLg\/__results___files\/__results___54_0.png\" rel=\"nofollow noreferrer\">images<\/a>* of proteins with 4 filters (Red, Green, Blue and Yellow). Every of those RGBY channels contains unique data as different cellular structures are visible with different filters. <\/p>\n\n<p>The idea is to use a pre-trained network e.g. VGG19 and extend the number of channels from default 3 to 4. Something like this:<\/p>\n\n<p><em>(My appologies, I am not allowed to add images directly before 10 reputation, please press the \"Run code snippet\" button to visualize):<\/em>\n<div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;img src=\"https:\/\/i.stack.imgur.com\/TZKka.png\" alt=\"Italian Trulli\"&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p><em>Picture: VGG model with RGB extended to RGBY<\/em><\/p>\n\n<p>The Y channel should be the copy of the existing pretrained channel. Then it is possible to make use of the pretrained weights. <\/p>\n\n<p>Does anyone have an idea of how such extension of a pretrained network can be achieved?<\/p>\n\n<p>*\n<em>Author of the collage - Allunia from Kaggle, \"Protein Atlas - Exploration and Baseline\" kernel.<\/em><\/p>\n","answers":[{"AnswerId":"53267270","CreationDate":"2018-11-12T17:31:50.153","ParentId":null,"OwnerUserId":"1919956","Title":null,"Body":"<p>Use the <code>layer.get_weights()<\/code> and <code>layer.set_weights()<\/code> functions of <a href=\"https:\/\/keras.io\/layers\/about-keras-layers\/\" rel=\"nofollow noreferrer\">Keras api<\/a>. <\/p>\n\n<p>Create a template structure for 4-layers VGG (set input <code>shape=(width, height, 4)<\/code>). Then load the weights from 3-channel RGB model into 4-channel as RGBB. <\/p>\n\n<p>Below is the code that does the procedure. In case of sequential VGG, the only layer that needs to be modified is the first Convolution layer. The structure of the subsequent layers is independent on the number of channels. <\/p>\n\n<pre><code>#!\/usr\/bin\/env python3\n# -*- coding: utf-8 -*-\n\nfrom keras.applications.vgg19 import VGG19\nfrom keras.models import Model\n\nvgg19 = VGG19(weights='imagenet')\nvgg19.summary() # To check which layers will be omitted in 'pretrained' model\n\n# Load part of the VGG without the top layers into 'pretrained' model\npretrained = Model(inputs=vgg19.input, outputs=vgg19.get_layer('block5_pool').output)\npretrained.summary()\n\n#%% Prepare model template with 4 input channels\nconfig = pretrained.get_config() # run config['layers'][i] for reference\n                                 # to restore layer-by layer structure\n\nfrom keras.layers import Input, Conv2D, MaxPooling2D\nfrom keras import optimizers\n\n# For training from scratch change kernel_initializer to e.g.'VarianceScaling'\ninputs = Input(shape=(224, 224, 4), name='input_17')\n# block 1\nx = Conv2D(64, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block1_conv1')(inputs)\nx = Conv2D(64, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block1_conv2')(x)\nx = MaxPooling2D(pool_size=(2, 2), name='block1_pool')(x)\n\n# block 2\nx = Conv2D(128, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block2_conv1')(x)\nx = Conv2D(128, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block2_conv2')(x)\nx = MaxPooling2D(pool_size=(2, 2), strides=(2,2), name='block2_pool')(x)\n\n# block 3\nx = Conv2D(256, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block3_conv1')(x)\nx = Conv2D(256, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block3_conv2')(x)\nx = Conv2D(256, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block3_conv3')(x)\nx = Conv2D(256, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block3_conv4')(x)\nx = MaxPooling2D(pool_size=(2, 2), strides=(2,2), name='block3_pool')(x)\n\n# block 4\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block4_conv1')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block4_conv2')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block4_conv3')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block4_conv4')(x)\nx = MaxPooling2D(pool_size=(2, 2), strides=(2,2), name='block4_pool')(x)\n\n# block 5\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block5_conv1')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block5_conv2')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block5_conv3')(x)\nx = Conv2D(512, (3,3), padding='same', activation='relu', kernel_initializer='zeros', name='block5_conv4')(x)\nx = MaxPooling2D(pool_size=(2, 2), strides=(2,2), name='block5_pool')(x)\n\nvgg_template = Model(inputs=inputs, outputs=x)\n\nvgg_template.compile(optimizer=optimizers.RMSprop(lr=2e-4),\n                     loss='categorical_crossentropy',\n                     metrics=['acc'])\n\n\n#%% Rewrite the weight loading\/modification function\nimport numpy as np\n\nlayers_to_modify = ['block1_conv1'] # Turns out the only layer that changes\n                                    # shape due to 4th channel is the first\n                                    # convolution layer.\n\nfor layer in pretrained.layers: # pretrained Model and template have the same\n                                # layers, so it doesn't matter which to \n                                # iterate over.\n\n    if layer.get_weights() != []: # Skip input, pooling and no weights layers\n\n        target_layer = vgg_template.get_layer(name=layer.name)\n\n        if layer.name in layers_to_modify:\n\n            kernels = layer.get_weights()[0]\n            biases  = layer.get_weights()[1]\n\n            kernels_extra_channel = np.concatenate((kernels,\n                                                    kernels[:,:,-1:,:]),\n                                                    axis=-2) # For channels_last\n\n            target_layer.set_weights([kernels_extra_channel, biases])\n\n        else:\n            target_layer.set_weights(layer.get_weights())\n\n\n#%% Save 4 channel model populated with weights for futher use    \n\nvgg_template.save('vgg19_modified_clear.hdf5')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53252152,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm>","CreationDate":"2018-11-11T19:02:54.023","AcceptedAnswerId":null,"OwnerUserId":10313194.0,"Title":"Keras time series can I predict next 6 month in one time","Body":"<p>I use keras for time series prediction. My code can predict next 6 months by predict next one month and then get it to be input for predict next month again untill complete 6 months. That means predict one month 6 times. Can I predict next 6 month in one time.<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom keras.layers import LSTM\nfrom pandas.tseries.offsets import MonthEnd\nfrom sklearn.preprocessing import MinMaxScaler\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nimport keras.backend as K\nfrom keras.layers import Bidirectional\nfrom keras.layers import Embedding\nfrom keras.layers import GRU\n\ndf = pd.read_csv('D:\/\/data.csv',\n             engine='python')\n\ndf['DATE_'] = pd.to_datetime(df['DATE_']) + MonthEnd(1)\ndf = df.set_index('DATE_')\ndf.head()\n\nsplit_date = pd.Timestamp('03-01-2015')\n\ntrain = df.loc[:split_date, ['data']]\ntest = df.loc[split_date:, ['data']]\nsc = MinMaxScaler()\n\ntrain_sc = sc.fit_transform(train)\ntest_sc = sc.transform(test)\n\nX_train = train_sc[:-1]\ny_train = train_sc[1:]\n\nX_test = test_sc[:-1]\ny_test = test_sc[1:]\n\nK.clear_session()\nmodel = Sequential()\nmodel.add(Dense(12, input_dim=1, activation='relu'))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.summary()\n\nmodel.fit(X_train, y_train, epochs=200, batch_size=2)\n\ny_pred = model.predict(X_test)\n\nreal_pred = sc.inverse_transform(y_pred)\nreal_test = sc.inverse_transform(y_test)\n\nprint(\"Predict Value\")\nprint(real_pred)\n\nprint(\"Test Value\")\nprint(real_test)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53252969","CreationDate":"2018-11-11T20:33:37.977","ParentId":null,"OwnerUserId":"8111755","Title":null,"Body":"<p>Yes, by changing your output layer (the last layer) from Dense(1) to Dense(6).  Of course you also have to change your y_train and y_test to have shape (1,6) instead of (1,1).<\/p>\n\n<p>Best of luck.<\/p>\n"}]}
+{"QuestionId":53252177,"AnswerCount":0,"Tags":"<tensorflow><autoencoder>","CreationDate":"2018-11-11T19:06:24.050","AcceptedAnswerId":null,"OwnerUserId":7409794.0,"Title":"Considering Gaussian decoder for Variational autoencoders","Body":"<p>I am trying to implement variation auto-encoder for real data where both encoder and decoder are modeled via multivariate Gaussian. I have found several implementations online for the case where the encoder is Gaussian and decoder is Bernoulli, but nothing for Gaussian decoder case. For the case of Bernoulli decoder the reconstruction loss can be defined as follows<\/p>\n\n<pre><code>reconstr_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=x,logits=x_out_logit)\n<\/code><\/pre>\n\n<p>where x_out_logit is modeled by a DNN. I am not sure how to write reconstruction loss for the Gaussian case. I assumed the decoder should output mean (gz_mean) and variance (gz_log_sigma_sq) as well (similar to the Gaussian encoder) and since the reconstruction loss is the Gaussian probability, I defined it to be<\/p>\n\n<pre><code>mvn = tf.contrib.distributions.MultivariateNormalDiag(loc=self.gz_mean,scale_diag=tf.sqrt(tf.exp(self.gz_log_sigma_sq)))\nreconstr_loss = tf.log(1e-20+mvn.prob(self.x))\n<\/code><\/pre>\n\n<p>However this loss does not seem to work, mvn.prob(self.x) is always zero no matter what training step. Please let me know of any ideas or any git-hub source which considers this case.<\/p>\n","answers":[]}
+{"QuestionId":53252284,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><deep-learning><computer-vision>","CreationDate":"2018-11-11T19:18:03.577","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"Multi-input Convolutional Neural Network for Images Classification","Body":"<p>I am beginner in deep learning and I want to create a multi-input Convolutional Neural Network (CNN) model in Keras for <strong>Images Classification<\/strong>. <\/p>\n\n<p>I am working to create a CNN model that takes <strong>two images<\/strong> and gives <strong>one<\/strong> output which is the <strong>class of the two images<\/strong>. <\/p>\n\n<p>I have two datasets: type1 and type2, and each dataset contains <strong>the same  classes<\/strong>, but <strong>the number of images in each class in the dataset type1 is higher than the number of images in each class in the dataset type2.<\/strong> \nThe model should take one image from Type1 dataset and one image from Type2 dataset and then classify these images to <strong>one class<\/strong> (ClassA or ClassB or------).<\/p>\n\n<p>The following is the structure of the datasets.<\/p>\n\n<pre><code>Type1 dataset\n|Train\n              |ClassA\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassB\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassC\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n              |ClassD\n                             |image1\n                             |image2\n                             |image3\n                             |image4\n                            -----\n       ----------------\n|Validate\n            -----------\n|Test\n           --------------\n\nType2 dataset\n|Train\n              |ClassA\n                             |image1\n                             |image2\n                            -----\n              |ClassB\n                             |image1\n                             |image2\n                            -----\n              |ClassC\n                             |image1\n                             |image2\n                            -----\n              |ClassD\n                             |image1\n                             |image2\n                            -----\n       ----------------\n|Validate\n            -----------\n|Test\n           --------------\n<\/code><\/pre>\n\n<p>The model is very similar to the model in this images, but it has more layers before flatten layer.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Dg2QX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Dg2QX.png\" alt=\"Multi-input\"><\/a><\/p>\n\n<p>I created a custom generator that inputs two images (from type 1 and 2), and each image from type1 be paired with every images from type2 as long as these images belong to the <strong>same class (label)<\/strong>. <\/p>\n\n<p>The problem is when executing <code>fit_generator<\/code> I get infinite loop as the following:<\/p>\n\n<pre><code>  Found *** images belonging to 100 classes.\n\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes.\n    Found *** images belonging to 100 classes. ......\n.................................................................\n<\/code><\/pre>\n\n<p>Here is my custom generator code:<\/p>\n\n<pre><code>input_imgen = ImageDataGenerator( \n                                  rotation_range=10,\n                                  shear_range=0.2,\n                                  zoom_range=0.1,\n                                  width_shift_range=0.1,\n                                  height_shift_range=0.1\n                                  )\n\n\n\ntest_imgen = ImageDataGenerator()\n\n\n\ndef generate_generator_multiple(generator,dir1, dir2, batch_size, img_height,img_width):\n\n\n    genX1 = generator.flow_from_directory(dir1,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n\n    genX2 = generator.flow_from_directory(dir2,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n    while True:\n      X2i = genX2.next() \n      Type1 = []\n      Type2 = []\n      image1 = []\n      image2 = []\n\n      while True:\n        X1i = genX1.next() \n        for i in range(len(X2i[1])): #Type2\n          for j in range(len(X1i[1])): #Type1\n            if all(X2i[1][i]) == all(X1i[1][j]): # have same label\n              image1.append(X1i[0][j]) # add image\n              image1.append(X1i[1][j]) # add label\n              image2.append(X2i[0][i]) # add image\n              image2.append(X2i[1][i]) # add label\n      Type1.append(image1)\n      Type2.append(image2)\n      yield [Type1 [0], Type2 [0]], Type2 [1]  #Yield both images and their mutual label\n\n\ninputgenerator=generate_generator_multiple(generator=input_imgen,\n                                           dir1=train_iris_data,\n                                           dir2=train_face_data,\n                                           batch_size=32,\n                                           img_height=224,\n                                           img_width=224)       \n\nvalidgenerator=generate_generator_multiple(generator=test_imgen,\n                                          dir1=valid_iris_data,\n                                          dir2=valid_face_data,\n                                          batch_size=32,\n                                          img_height=224,\n                                          img_width=224) \n\ntestgenerator=generate_generator_multiple(generator=test_imgen,\n                                          dir1=test_face_data,\n                                          dir2=test_face_data,\n                                          batch_size=32,\n                                          img_height=224,\n                                          img_width=224)\n\n\n    # compile the model\n    multi_model.compile(\n            loss='categorical_crossentropy',\n            optimizer=Adam(lr=0.0001),\n            metrics=['accuracy']\n        )\n\n\n# train the model and save the history\nhistory = multi_model.fit_generator(\ninputgenerator,\nsteps_per_epoch=len(train_data) \/\/ batch_size,\nepochs=10,\nverbose=1,\nvalidation_data=validgenerator,\nvalidation_steps=len(valid_data) \/\/ batch_size,\nuse_multiprocessing=True,\nshuffle=False\n)\n<\/code><\/pre>\n\n<p>Could you please help me to solve this problem and create the custom generator??<\/p>\n","answers":[]}
+{"QuestionId":53252286,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-11-11T19:18:10.527","AcceptedAnswerId":null,"OwnerUserId":3143269.0,"Title":"Building Tensorflow r1.12 with Cuda 10 on Ubuntu 18.04","Body":"<p>I really have tried getting built. I have followed the instructions <a href=\"https:\/\/www.python36.com\/how-to-install-tensorflow-gpu-with-cuda-10-0-for-python-on-ubuntu\/\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/medium.com\/@vitali.usau\/install-cuda-10-0-cudnn-7-3-and-build-tensorflow-gpu-from-source-on-ubuntu-18-04-3daf720b83fe\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">here<\/a>. Clean system:<\/p>\n\n<ul>\n<li>Ubuntu 18.04 server<\/li>\n<li>Ubuntu Gnome<\/li>\n<li>Cuda 10.0<\/li>\n<li>cuDNN 7.4.1<\/li>\n<li>NCCL 2.3.7<\/li>\n<li>TensorRT 5.0.2.1<\/li>\n<li>Tensorflow r1.12<\/li>\n<li>Python 3.6<\/li>\n<li><p>GCC 7.3.0\nusing the command line:<\/p>\n\n<p>bazel build --config=cuda --config=mkl \/\/tensorflow\/tools\/pip_package:build_pip_package<\/p><\/li>\n<\/ul>\n\n<p>This is the error:<\/p>\n\n<pre><code>WARNING: The following rc files are no longer being read, please transfer their contents or import their path into one of the standard rc files:\n\/devel\/software\/tensorflow\/tools\/bazel.rc\nINFO: Options provided by the client:\n  Inherited 'common' options: --isatty=1 --terminal_columns=171\nERROR: Config value cuda is not defined in any .rc file\n<\/code><\/pre>\n\n<p>Anybody know which parts of bazel.rc the need to be imported?<\/p>\n\n<p>Why is cuda not defined?<\/p>\n\n<p>If I try:<\/p>\n\n<p>bazel test -c opt -- \/\/tensorflow\/... -\/\/tensorflow\/compiler\/... -\/\/tensorflow\/contrib\/lite\/...<\/p>\n\n<p>It fails with this error:<\/p>\n\n<p>ERROR: \/devel\/software\/tensorflow\/tensorflow\/python\/BUILD:3768:1: Linking of rule '\/\/tensorflow\/python:_pywrap_tensorflow_internal.so' failed (Exit 1) gcc failed: error executing command \/usr\/bin\/gcc -shared -o bazel-out\/k8-opt\/bin\/tensorflow\/python\/_pywrap_tensorflow_internal.so -Wl,--version-script bazel-out\/k8-opt\/bin\/tensorflow\/python\/pywrap_tensorflow_internal_versionscript.lds ... (remaining 65 argument(s) skipped)<\/p>\n\n<p>Anyone had success? (Obviously at least 2 people have) Maybe someone could tell me what I am doing wrong? Or how to get this built?<\/p>\n","answers":[{"AnswerId":"53338061","CreationDate":"2018-11-16T12:37:25.557","ParentId":null,"OwnerUserId":"4404341","Title":null,"Body":"<p>Do as what is said in the warning and it will work. <\/p>\n\n<pre><code>\"WARNING: The following rc files are no longer being read, please transfer their contents or import their path into one of the standard rc files:\n\/devel\/software\/tensorflow\/tools\/bazel.rc\"\n<\/code><\/pre>\n\n<p>You should have a file .bazelrc at  \/devel\/software\/tensorflow\/.bazelrc. At the top of this file add the following code snippet.<\/p>\n\n<pre><code>import \/devel\/software\/tensorflow\/tools\/bazel.rc\n<\/code><\/pre>\n\n<p>Enjoy!<\/p>\n"}]}
+{"QuestionId":53252529,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2018-11-11T19:45:16.570","AcceptedAnswerId":53252593.0,"OwnerUserId":8325122.0,"Title":"Import tensorflow error: DLL load failed: The specified procedure could not be found","Body":"<pre><code>import tensorflow\n<\/code><\/pre>\n\n<p>When i try to import tensorflow and run the code I get this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/Ivan\/PycharmProjects\/AI_Numers\/script.py\", line 1, in &lt;module&gt;\nimport tensorflow\n  File \"C:\\Users\\Ivan\\PycharmProjects\\AI_Numers\\venv\\lib\\site-packages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Ivan\\PycharmProjects\\AI_Numers\\venv\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 59, in &lt;module&gt;\nfrom tensorflow.core.framework.graph_pb2 import *\n  File \"C:\\Users\\Ivan\\PycharmProjects\\AI_Numers\\venv\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"C:\\Users\\Ivan\\PycharmProjects\\AI_Numers\\venv\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in &lt;module&gt;\n    from google.protobuf.pyext import _message\nImportError: DLL load failed: The specified procedure could not be found.`\n<\/code><\/pre>\n\n<p>I am currently using: \nPython: 3.6.0\nTensorflow: 1.12<\/p>\n\n<p>I've already looked at <a href=\"https:\/\/stackoverflow.com\/questions\/52092810\/tensorflow-error-dll-load-failed-the-specified-procedure-could-not-be-found\">this<\/a> thread, but I could not fix my problem.<\/p>\n","answers":[{"AnswerId":"53252593","CreationDate":"2018-11-11T19:52:26.540","ParentId":null,"OwnerUserId":"10637425","Title":null,"Body":"<p>It seems some backward compatibility issues on the library called protobuf. You can use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/21719\" rel=\"nofollow noreferrer\">this<\/a>. But I would recommend a fresh install of python 3.6.6 with anaconda or miniconda and use conda to install tensorflow.<\/p>\n"},{"AnswerId":"53252611","CreationDate":"2018-11-11T19:53:42.617","ParentId":null,"OwnerUserId":"8929127","Title":null,"Body":"<p>Maybe <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22794\" rel=\"nofollow noreferrer\">this<\/a> or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22512\" rel=\"nofollow noreferrer\">this<\/a> could help you. Many people had problems coused by <code>DLL load failed<\/code>. Also <a href=\"https:\/\/www.tensorflow.org\/install\/errors\" rel=\"nofollow noreferrer\">tensorflow<\/a> is providing a list with possible errors<\/p>\n"},{"AnswerId":"54339018","CreationDate":"2019-01-24T03:35:53.220","ParentId":null,"OwnerUserId":"6881202","Title":null,"Body":"<p>Facing the same issue on my machine after some trials found a solution which worked for me.\nSteps to solve the issue:<\/p>\n\n<ol>\n<li>Re-Install Cuda(<strong>Till Date Tensorflow only supports Cuda V9.0<\/strong> so install Cuda9.0 only)<\/li>\n<li>Add Cuda library, bin paths to system variables<\/li>\n<li>Add <strong>Cudnn(Any Version is Ohkey) library<\/strong> to Cuda9.0<\/li>\n<li>Restart PC<\/li>\n<li>Install \"pip install <strong>tensorflow-gpu==1.8<\/strong>\"(I tried with higher version also but not supported in my case but 1.8 easily got installed)<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53253465,"AnswerCount":1,"Tags":"<python><keras><deep-learning><artificial-intelligence><mnist>","CreationDate":"2018-11-11T21:35:20.717","AcceptedAnswerId":null,"OwnerUserId":6131173.0,"Title":"How to combine Keras MNIST dataset with my own MNIST images?","Body":"<p>I created some MNIST digits by using a generative adversarial neural network and saved them in png format. I know that Keras has the MNIST dataset, but I want to combine the digit images that I created with the original MNIST dataset in the Keras. Is this possible, if so, how can I do this ?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53253480","CreationDate":"2018-11-11T21:37:10.593","ParentId":null,"OwnerUserId":"7904315","Title":null,"Body":"<p>i'm pretty sure keras uses the lecun set found <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n\n<p>just decode that set and add your own to the pile<\/p>\n"}]}
+{"QuestionId":53253905,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2018-11-11T22:31:28.443","AcceptedAnswerId":53254078.0,"OwnerUserId":6637004.0,"Title":"Mac OS High Sierra: Tensorflow verions returned by `pip3 upgrade ` and `python3 -c 'import tensorflow as tf; print(tf.__version__)'` differ","Body":"<p>I am currently getting the error, <code>TypeError: softmax() got an unexpected keyword argument 'axis'<\/code>.<\/p>\n\n<p>It seems this error is common if you don't have up-to-date keras\/tensorflow. <\/p>\n\n<p>I checked what version of tensorflow I have with <code>python3 -c 'import tensorflow as tf; print(tf.__version__)'<\/code> It returned <code>0.12.0<\/code> which indeed looks out of date. <\/p>\n\n<p>However, when I try to upgrade tensorflow, with <code>pip3 install tensorflow --upgrade<\/code> I get <\/p>\n\n<blockquote>\n  <p>Requirement already up-to-date: tensorflow in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\n  (0.12.0) Requirement already satisfied, skipping upgrade:\n  numpy>=1.11.0 in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages(from\n  tensorflow) (1.15.3) Requirement already satisfied, skipping upgrade:\n  protobuf==3.1.0 in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\n  (from tensorflow) (3.1.0) Requirement already satisfied, skipping\n  upgrade: six>=1.10.0 in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\n  (from tensorflow) (1.11.0) Requirement already satisfied, skipping\n  upgrade: wheel>=0.26 in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\n  (from tensorflow) (0.32.2) Requirement already satisfied, skipping\n  upgrade: setuptools in\n  \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\n  (from protobuf==3.1.0->tensorflow) (39.0.1)<\/p>\n<\/blockquote>\n\n<p>I tried to force the update with <code>pip3 install --ignore-installed --upgrade tensorflow<\/code> but got:<\/p>\n\n<blockquote>\n  <p>Collecting tensorflow   Could not find a version that satisfies the\n  requirement tensorflow (from versions: ) No matching distribution\n  found for tensorflow<\/p>\n<\/blockquote>\n\n<p>When I tried to upgrade with <code>python3 -m pip install tensorflow --upgrade<\/code>:<\/p>\n\n<pre><code>Requirement already up-to-date: tensorflow in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages (0.12.0)\nRequirement already satisfied, skipping upgrade: wheel&gt;=0.26 in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages (from tensorflow) (0.32.2)\nRequirement already satisfied, skipping upgrade: protobuf==3.1.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages (from tensorflow) (3.1.0)\nRequirement already satisfied, skipping upgrade: numpy&gt;=1.11.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages(from tensorflow) (1.15.3)\nRequirement already satisfied, skipping upgrade: six&gt;=1.10.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages (from tensorflow) (1.11.0)\nRequirement already satisfied, skipping upgrade: setuptools in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages (from protobuf==3.1.0-&gt;tensorflow) (39.0.1)\n<\/code><\/pre>\n\n<p>When I print sys.path in my program, I get:<\/p>\n\n<pre><code>['\/Users\/myname\/topdirect', '\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python37.zip', '\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7', '\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/lib-dynload', '\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages']\n<\/code><\/pre>\n\n<p><code>pip3 show tensorflow<\/code> I get:<\/p>\n\n<pre><code>Name: tensorflow\nVersion: 0.12.0\nSummary: TensorFlow helps the tensors flow\nHome-page: http:\/\/tensorflow.org\/\nAuthor: Google Inc.\nAuthor-email: opensource@google.com\nLicense: Apache 2.0\nLocation: \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\nRequires: six, protobuf, numpy, wheel\nRequired-by:\n<\/code><\/pre>\n","answers":[{"AnswerId":"53254078","CreationDate":"2018-11-11T22:56:48.393","ParentId":null,"OwnerUserId":"9674061","Title":null,"Body":"<p>First of all, I would recommend using anaconda  environment to manage your tensorflow version. On the other side, you need to add your operating system to make the question more clear.<\/p>\n"}]}
+{"QuestionId":53254056,"AnswerCount":2,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-11T22:54:55.097","AcceptedAnswerId":53254057.0,"OwnerUserId":246028.0,"Title":"Running cloud TPU profiler in Google Colab environment","Body":"<p>I am running a Google Colab notebook and am trying to capture TPU profiling data for use in TensorBoard, however I can't get <code>capture_tpu_profile<\/code> to run in the background while running my TensorFlow code.<\/p>\n\n<p>So far I tried to run the capture process in the background with:<\/p>\n\n<pre><code>!capture_tpu_profile --logdir=gs:\/\/&lt;my_logdir&gt; --tpu=$COLAB_TPU_ADDR &amp;\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>!bg capture_tpu_profile --logdir=gs:\/\/&lt;my_logdir&gt; --tpu=$COLAB_TPU_ADDR\n<\/code><\/pre>\n","answers":[{"AnswerId":"54522174","CreationDate":"2019-02-04T18:26:26.463","ParentId":null,"OwnerUserId":"4716286","Title":null,"Body":"<p>One way to do this is to use the TPUProfilerHook<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tpu\/blob\/master\/models\/common\/tpu_profiler_hook.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tpu\/blob\/master\/models\/common\/tpu_profiler_hook.py<\/a><\/p>\n\n<p>Which runs the profiler as a session hook.<\/p>\n\n<p>Example here <a href=\"https:\/\/github.com\/tensorflow\/tpu\/blob\/5d838047af0163bdf7b97b9404648dc2961c4b63\/models\/official\/resnet\/resnet_main.py#L699\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tpu\/blob\/5d838047af0163bdf7b97b9404648dc2961c4b63\/models\/official\/resnet\/resnet_main.py#L699<\/a><\/p>\n"},{"AnswerId":"53254057","CreationDate":"2018-11-11T22:54:55.097","ParentId":null,"OwnerUserId":"246028","Title":null,"Body":"<p>Turns out a way to do this is to start the process from python directly like this (I also had to modify the parameter from <code>--tpu<\/code> to <code>--service_addr<\/code>):<\/p>\n\n<pre><code>import subprocess\nsubprocess.Popen([\"capture_tpu_profile\",\"--logdir=gs:\/\/&lt;my_logdir&gt;\", \"--service_addr={}\".format(os.environ['COLAB_TPU_ADDR'])])\n<\/code><\/pre>\n\n<p>the <code>check=True<\/code> makes the command raise an Exception if it fails.<\/p>\n"}]}
+{"QuestionId":53254061,"AnswerCount":1,"Tags":"<c++><tensorflow><bazel>","CreationDate":"2018-11-11T22:55:11.963","AcceptedAnswerId":53272388.0,"OwnerUserId":10630432.0,"Title":"Failing to Bazel build C++ project with tensorflow as a dependency","Body":"<p>I see a lot of tutorials that explain how to build projects inside Tensorflow's Bazel WORKSPACE (like <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/cc\" rel=\"nofollow noreferrer\">this one<\/a>). But I can't seem to be able to find a way to build my own project and include tensorflow as a dependency. I looked at <a href=\"https:\/\/docs.bazel.build\/versions\/master\/external.html#bazel-projects\" rel=\"nofollow noreferrer\">this<\/a> Bazel documentation, and there is clearly a way to build with external dependencies, which I tried to follow myself. (since tf is also built with bazel).<\/p>\n\n<p>Here is what my directory structure looks like:<\/p>\n\n<pre><code>.\n\u251c\u2500\u2500 perception\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 BUILD\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 graph_loader.cc\n\u251c\u2500\u2500 third-party\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 tensorflow # I cloned tf repo into this folder\n\u2514\u2500\u2500 WORKSPACE\n<\/code><\/pre>\n\n<p>Here is what's inside my <code>perception\/BUILD<\/code> file:<\/p>\n\n<pre><code>cc_binary(\n    name = \"graph-loader\",\n    srcs = [\n        \"graph_loader.cc\",\n    ],\n    deps = [\n        \"@tensorflow\/\/tensorflow:libtensorflow.so\",\n    ]\n)\n<\/code><\/pre>\n\n<p>Here is what's inside my <code>WORKSPACE<\/code> file:<\/p>\n\n<pre><code>local_repository(\n    name = \"tensorflow\",\n    path = \"path\/to\/my\/project\/third-party\/tensorflow\",\n)\n<\/code><\/pre>\n\n<p>Here is what's inside my <code>perception\/graph_loader.cc<\/code> file:<\/p>\n\n<pre><code>#include \"tensorflow\/cc\/client\/client_session.h\"\n#include \"tensorflow\/cc\/ops\/standard_ops.h\"\n#include \"tensorflow\/core\/framework\/tensor.h\"\n\nint main() {\n  using namespace tensorflow;\n  using namespace tensorflow::ops;\n  Scope root = Scope::NewRootScope();\n  \/\/ Matrix A = [3 2; -1 0]\n  auto A = Const(root, { {3.f, 2.f}, {-1.f, 0.f} });\n  \/\/ Vector b = [3 5]\n  auto b = Const(root, { {3.f, 5.f} });\n  \/\/ v = Ab^T\n  auto v = MatMul(root.WithOpName(\"v\"), A, b, MatMul::TransposeB(true));\n  std::vector&lt;Tensor&gt; outputs;\n  ClientSession session(root);\n  \/\/ Run and fetch v\n  TF_CHECK_OK(session.Run({v}, &amp;outputs));\n  \/\/ Expect outputs[0] == [19; -3]\n  LOG(INFO) &lt;&lt; outputs[0].matrix&lt;float&gt;();\n  return 0;\n}\n<\/code><\/pre>\n\n<p>I run my build with the following command:<\/p>\n\n<pre><code>build \/\/perception:graph-loader\n<\/code><\/pre>\n\n<p>But it fails with this message:<\/p>\n\n<pre><code>ERROR: path\/to\/my\/project\/perception\/BUILD:1:1: error loading package '@tensorflow\/\/tensorflow': Extension file not found. Unable to load package for '@local_config_cuda\/\/cuda:build_defs.bzl': The repository could not be resolved and referenced by '\/\/perception:graph-loader'\nERROR: Analysis of target '\/\/perception:graph-loader' failed; build aborted: error loading package '@tensorflow\/\/tensorflow': Extension file not found. Unable to load package for '@local_config_cuda\/\/cuda:build_defs.bzl': The repository could not be resolved\nINFO: Elapsed time: 0.037s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (1 packages loaded, 0 targets configured)\n    currently loading: @tensorflow\/\/tensorflow\n<\/code><\/pre>\n\n<p>Here are the quesitions:<\/p>\n\n<ol>\n<li>What am I doing wrong so that my build keeps failing? <\/li>\n<li>Is it even possible to do what I'm trying here, i.e. build tensorflow inside of my project?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53272388","CreationDate":"2018-11-13T01:17:39.687","ParentId":null,"OwnerUserId":"6521181","Title":null,"Body":"<p>You are getting this error because you have not added the required repository rules in your <code>WORKSPACE<\/code>. Bazel currently does not have recursive workspace, so you need to manually copy those repositories of all your dependencies into the main <code>WORKSPACE<\/code>.<\/p>\n\n<p>In your <code>WORKSPACE<\/code> file, copy this:<\/p>\n\n<pre><code>local_repository(\n    name = \"org_tensorflow\",\n    path = \"third-party\/tensorflow\",\n)\n<\/code><\/pre>\n\n<p>Append all content of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/WORKSPACE\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/WORKSPACE<\/a> into the <code>WORKSPACE<\/code> file. Remove the <code>workspace(name = \"org_tensorflow\")<\/code> line.<\/p>\n\n<p>Finally, change all <code>\/\/*<\/code> in <code>WOKRSPACE<\/code> to <code>@org_tensorflow\/\/*<\/code>.<\/p>\n\n<p>Note that building Tensorflow in sub-folder is not officially supported.<\/p>\n"}]}
+{"QuestionId":53254100,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-11T22:59:50.887","AcceptedAnswerId":null,"OwnerUserId":10510693.0,"Title":"ValueError: You are trying to load a weight file containing 5 layers into a model with 0 layers","Body":"<p>I have been searching to find an answer to this problem in the past few days, no luck.<\/p>\n\n<p>Just training a simple <code>CNN<\/code>, an saving the <code>weights<\/code> and also the <code>model<\/code>.<\/p>\n\n<p>While loading the <code>model<\/code> and <code>weights<\/code>, i receive the error saying :   <\/p>\n\n<blockquote>\n  <p>ValueError: You are trying to load a weight file containing 5 layers into a model with 0 layers.<\/p>\n<\/blockquote>\n\n<p>I have used this before without any issue, long time a go. But all new attempts to load the model fails. <\/p>\n\n<p>Can you help?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/5fKTS.jpg\" alt=\"Screen Shot\"><\/p>\n","answers":[]}
+{"QuestionId":53254420,"AnswerCount":0,"Tags":"<python><keras><rnn>","CreationDate":"2018-11-11T23:51:28.220","AcceptedAnswerId":null,"OwnerUserId":5405378.0,"Title":"Creating a multiple output RNN with keras","Body":"<p>I'm trying to use a csv file as my input. The csv file has 100000 rows and 10 columns. I would like the first 7 columns to be the input of the network and the last 3 columns to be the output of the network. First 6 columns are demographic information and the last 4 are the number of products they purchased. I would like to predict the number of purchased items for the next 3 visits, (based off demographics and their first purchase).<\/p>\n\n<p>Can this be done with keras and is RNN the right choice?<\/p>\n\n<p>Data in csv format looks like...<\/p>\n\n<p>age,zip,income,weight,waistsize,necksize,p1#,p2#,p3#,p4#<\/p>\n\n<p>19,92201,30000,220,39,17,9,1,3,1<\/p>\n\n<p>33,60173,45000,245,44,15,2,5,3,2<\/p>\n","answers":[]}
+{"QuestionId":53254586,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-11-12T00:23:50.650","AcceptedAnswerId":53256297.0,"OwnerUserId":5495304.0,"Title":"Keras train and validation metric values are different even when using same data (Logistic regression)","Body":"<p>I have been trying to better understand the train\/validation sequence in the <code>keras<\/code> model <code>fit()<\/code> loop. So I tried out a simple training loop where I attempted to fit a simple logistic regression model with input data consisting of a single feature. <\/p>\n\n<p>I feed the <strong>same data<\/strong> for both <strong>training<\/strong> and <strong>validation<\/strong>. Under those conditions, and by specifying batch size to be the same and total data size, one would expect to obtain exactly the same loss and accuracy. But this is not the case. <\/p>\n\n<p>Here is my code:<\/p>\n\n<p>Generate some two random data with two classes:<\/p>\n\n<pre><code>N = 100\nx = np.concatenate([np.random.randn(N\/\/2, 1), np.random.randn(N\/\/2, 1)+2])\ny = np.concatenate([np.zeros(N\/\/2), np.ones(N\/\/2)])\n<\/code><\/pre>\n\n<p>And plotting the two class data distribution (one feature x):<\/p>\n\n<pre><code>data = pd.DataFrame({'x': x.ravel(), 'y': y})\nsns.violinplot(x='x', y='y', inner='point', data=data, orient='h')\npyplot.tight_layout(0)\npyplot.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DNy36.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DNy36.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Build and fit the keras model:<\/p>\n\n<pre><code>model = tf.keras.Sequential([tf.keras.layers.Dense(1, activation='sigmoid', input_dim=1)])\nmodel.compile(optimizer=tf.keras.optimizers.SGD(2), loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(x, y, epochs=10, validation_data=(x, y), batch_size=N)\n<\/code><\/pre>\n\n<p>Notice that I have specified the data <code>x<\/code> and targets <code>y<\/code> for both training and for <code>validation_data<\/code>. Also, the batch_size is same as total size <code>batch_size=N<\/code>. <\/p>\n\n<p>The training results are:<\/p>\n\n<pre><code>100\/100 [==============================] - 1s 5ms\/step - loss: 1.4500 - acc: 0.2300 - val_loss: 0.5439 - val_acc: 0.7200\nEpoch 2\/10\n100\/100 [==============================] - 0s 18us\/step - loss: 0.5439 - acc: 0.7200 - val_loss: 0.4408 - val_acc: 0.8000\nEpoch 3\/10\n100\/100 [==============================] - 0s 16us\/step - loss: 0.4408 - acc: 0.8000 - val_loss: 0.3922 - val_acc: 0.8300\nEpoch 4\/10\n100\/100 [==============================] - 0s 16us\/step - loss: 0.3922 - acc: 0.8300 - val_loss: 0.3659 - val_acc: 0.8400\nEpoch 5\/10\n100\/100 [==============================] - 0s 17us\/step - loss: 0.3659 - acc: 0.8400 - val_loss: 0.3483 - val_acc: 0.8500\nEpoch 6\/10\n100\/100 [==============================] - 0s 16us\/step - loss: 0.3483 - acc: 0.8500 - val_loss: 0.3356 - val_acc: 0.8600\nEpoch 7\/10\n100\/100 [==============================] - 0s 17us\/step - loss: 0.3356 - acc: 0.8600 - val_loss: 0.3260 - val_acc: 0.8600\nEpoch 8\/10\n100\/100 [==============================] - 0s 18us\/step - loss: 0.3260 - acc: 0.8600 - val_loss: 0.3186 - val_acc: 0.8600\nEpoch 9\/10\n100\/100 [==============================] - 0s 18us\/step - loss: 0.3186 - acc: 0.8600 - val_loss: 0.3127 - val_acc: 0.8700\nEpoch 10\/10\n100\/100 [==============================] - 0s 23us\/step - loss: 0.3127 - acc: 0.8700 - val_loss: 0.3079 - val_acc: 0.8800\n<\/code><\/pre>\n\n<p>The results show that <strong>val_loss<\/strong> and <strong>loss<\/strong> are not the same at the end of each epoch, and also <strong>acc<\/strong> and <strong>val_acc<\/strong> are not exactly the same. However, based on this setup, one would expect them to be the same. <\/p>\n\n<p>I have been going through the code in keras, particularly this part: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training.py#L1364\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training.py#L1364<\/a><\/p>\n\n<p>and so far, all I can say that the difference is due to some different computation through the computation graph. <\/p>\n\n<p><strong>Does anyone has any idea why there would be such difference?<\/strong><\/p>\n","answers":[{"AnswerId":"53256297","CreationDate":"2018-11-12T05:17:11.610","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>So after looking more closely at the results, the <code>loss<\/code> and <code>acc<\/code> values from the training step are computed BEFORE the current batch is used to update the model.<\/p>\n\n<p>Thus, in the case of a single batch per epoch, the train <code>acc<\/code> and <code>loss<\/code> are evaluated when the batch is fed in, then the model parameters are updated based on the provided optimizer. After the train step is finished, we compute loss and accuracy by feeding in the validation data, which is now evaluated using a new updated model. <\/p>\n\n<p>This is evident from the training results output, where validation accuracy and loss are in epoch 1 are equal to train accuracy and loss in epoch 2, etc...<\/p>\n\n<p>A quick check using tensorflow confirmed that values are fetched before variables are updated: <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nnp.random.seed(1)\n\nx = tf.placeholder(dtype=tf.float32, shape=(None, 1), name=\"x\")\ny = tf.placeholder(dtype=tf.float32, shape=(None), name=\"y\")\n\nW = tf.get_variable(name=\"W\", shape=(1, 1), dtype=tf.float32, initializer=tf.constant_initializer(0))\nb = tf.get_variable(name=\"b\", shape=1, dtype=tf.float32, initializer=tf.constant_initializer(0))\nz = tf.matmul(x, W) + b\n\nerror = tf.square(z - y)\nobj = tf.reduce_mean(error, name=\"obj\")\n\nopt = tf.train.MomentumOptimizer(learning_rate=0.025, momentum=0.9)\ngrads = opt.compute_gradients(obj)\ntrain_step = opt.apply_gradients(grads)\n\nN = 100\nx_np = np.random.randn(N).reshape(-1, 1)\ny_np = 2*x_np + 3 + np.random.randn(N)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(2):\n        res = sess.run([obj, W, b, train_step], feed_dict={x: x_np, y: y_np})\n        print('MSE: {}, W: {}, b: {}'.format(res[0], res[1][0, 0], res[2][0]))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>MSE: 14.721437454223633, W: 0.0, b: 0.0\nMSE: 13.372591018676758, W: 0.08826743811368942, b: 0.1636980175971985\n<\/code><\/pre>\n\n<p>Since the parameters <code>W<\/code> and <code>b<\/code> were initialized to 0, then it is clear that the fetched values is still 0 even though session was run with gradient update request...<\/p>\n"}]}
+{"QuestionId":53254870,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><mnist>","CreationDate":"2018-11-12T01:17:08.633","AcceptedAnswerId":53255191.0,"OwnerUserId":2361174.0,"Title":"How to use a different CNN without losing accuracy","Body":"<p>I have been given a task to implement a Convolutional neural network that can evaluate hand-written digits found in the <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">MNIST dataset<\/a> with the architecture of the network looking like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Ct8a8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ct8a8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have implemented a CNN that matches the architecture, unfortunately it only has about a 10% accuracy to it. I've looked online and tried other example CNNs to make sure if anything else causing the issue, however they seem to work fine and give me a ~99% accuracy. I've placed both CNNs in my code and made a boolean switch to show the difference between the two:<\/p>\n\n<pre><code>import tensorflow\nfrom tensorflow.keras.datasets import mnist\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, Flatten\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D\n\nbatch_size = 128\nnum_classes = 10\nepochs = 1\nimg_rows, img_cols = 28, 28\n\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n\nx_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)\nx_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)\ninput_shape = (img_rows, img_cols, 1)\n\nx_train = x_train.astype('float32')\nx_test = x_test.astype('float32')\nx_train \/= 255\nx_test \/= 255\nprint('x_train shape:', x_train.shape)\nprint(x_train.shape[0], 'train samples')\nprint(x_test.shape[0], 'test samples')\n\ny_train = tensorflow.keras.utils.to_categorical(y_train, num_classes)\ny_test = tensorflow.keras.utils.to_categorical(y_test, num_classes)\n\nexampleModel = False  # Use to toggle which CNN goes into the model\n\nif exampleModel:  # An example CNN that I found for MNIST\n    model = Sequential()\n    model.add(Conv2D(32, kernel_size=(3, 3),\n                     activation='relu',\n                     input_shape=input_shape))\n    model.add(Conv2D(64, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.25))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(num_classes, activation='softmax'))\nelse:  # The CNN I created\n    input_layer = tensorflow.keras.layers.Input(shape=input_shape)\n    conv1 = Conv2D(32, (1, 1), activation='relu')(input_layer)\n    pool1 = MaxPooling2D(2, 2)(conv1)\n    conv2_1 = Conv2D(64, (1, 1), activation='relu', padding='same')(pool1)\n    pool2_1 = MaxPooling2D(2, 2)(conv2_1)\n    drop2_1 = Dropout(0.5)(pool2_1)\n    conv2_2 = Conv2D(64, (1, 1), activation='relu', padding='same')(pool1)\n    pool2_2 = MaxPooling2D(2, 2)(conv2_2)\n    drop2_2 = Dropout(0.5)(pool2_2)\n    conv3_1 = Conv2D(256, (1, 1), activation='relu', padding='same')(drop2_1)\n    conv3_2 = Conv2D(256, (1, 1), activation='relu', padding='same')(drop2_2)\n    merged = tensorflow.keras.layers.concatenate([conv3_1, conv3_2], axis=-1)\n    merged = Dropout(0.5)(merged)\n    merged = Flatten()(merged)\n    fc1 = Dense(1000, activation='relu')(merged)\n    fc2 = Dense(500, activation='relu')(fc1)\n    out = Dense(10)(fc2)\n    model = tensorflow.keras.models.Model(input_layer, out)\n\nmodel.compile(loss=tensorflow.keras.losses.categorical_crossentropy,\n              optimizer=tensorflow.keras.optimizers.Adadelta(),\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train,\n          batch_size=batch_size,\n          epochs=epochs,\n          verbose=1,\n          validation_data=(x_test, y_test))\nscore = model.evaluate(x_test, y_test, verbose=0)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])\n<\/code><\/pre>\n\n<p>In order to complete my task, I believe I have to convert my example CNN piece-by-piece into the required architecture. Although I have no idea how to do this,  they look completely different from each other (one is purely sequential, the other uses parallel layers and merging). I'm a beginner in machine learning so there might be something that I'm missing although I can't find and resource online that goes into this conversion process. Any help with this is appreciated. <\/p>\n","answers":[{"AnswerId":"53255191","CreationDate":"2018-11-12T02:11:31.900","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You simply have to add an <strong>softmax activation<\/strong> to the last, <code>out<\/code> layer:<\/p>\n\n<pre><code>out = Dense(10, activation=\"softmax\")(fc2)\n<\/code><\/pre>\n\n<p>Thus your model in completed form:<\/p>\n\n<pre><code>input_layer = tensorflow.keras.layers.Input(shape=input_shape)\nconv1 = Conv2D(32, (1, 1), activation='relu')(input_layer)\npool1 = MaxPooling2D(2, 2)(conv1)\nconv2_1 = Conv2D(64, (1, 1), activation='relu', padding='same')(pool1)\npool2_1 = MaxPooling2D(2, 2)(conv2_1)\ndrop2_1 = Dropout(0.5)(pool2_1)\nconv2_2 = Conv2D(64, (1, 1), activation='relu', padding='same')(pool1)\npool2_2 = MaxPooling2D(2, 2)(conv2_2)\ndrop2_2 = Dropout(0.5)(pool2_2)\nconv3_1 = Conv2D(256, (1, 1), activation='relu', padding='same')(drop2_1)\nconv3_2 = Conv2D(256, (1, 1), activation='relu', padding='same')(drop2_2)\nmerged = tensorflow.keras.layers.concatenate([conv3_1, conv3_2], axis=-1)\nmerged = Dropout(0.5)(merged)\nmerged = Flatten()(merged)\nfc1 = Dense(1000, activation='relu')(merged)\nfc2 = Dense(500, activation='relu')(fc1)\nout = Dense(10, activation=\"softmax\")(fc2)\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>x_train shape: (60000, 28, 28, 1)\n60000 train samples\n10000 test samples\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/1\n60000\/60000 [==============================] - 25s 416us\/step - loss: 0.6394 - acc: 0.7858 - val_loss: 0.2956 - val_acc: 0.9047\nTest loss: 0.29562548571825026\nTest accuracy: 0.9047\n<\/code><\/pre>\n"}]}
+{"QuestionId":53254879,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><confusion-matrix>","CreationDate":"2018-11-12T01:18:54.730","AcceptedAnswerId":null,"OwnerUserId":10636279.0,"Title":"confusion matrix in keras cnn model without xtrain xtest ytrain ytest","Body":"<p>I am currently trying to implement a confusion matrix into my cnn model code. All the examples that I've been watching includes using <code>x_train, x_test, y_train, y_test<\/code>, but I don't know how to do that on my code, or if I will be able to do it from my_model.h5 file. Hoping you could help me. Thanks.<\/p>\n\n<p>I leave my model code below here:<\/p>\n\n<pre><code>classifier = Sequential()\nclassifier.add(Conv2D(32, (3,  3), input_shape = (64, 64, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size =(2,2)))\nclassifier.add(Conv2D(32, (3,  3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size =(2,2)))\nclassifier.add(Conv2D(64, (3,  3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size =(2,2)))\nclassifier.add(Flatten())\nclassifier.add(Dense(256, activation = 'relu'))\nclassifier.add(Dropout(0.5))\nclassifier.add(Dense(26, activation = 'softmax'))\n\nclassifier.compile(\n          optimizer = optimizers.SGD(lr = 0.01),\n          loss = 'categorical_crossentropy',\n          metrics = ['accuracy'])\n\nclassifier.summary()\n\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(\n    rescale=1.\/255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True)\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntraining_set = train_datagen.flow_from_directory(\n    'mydata\/training_set',\n    target_size=(64, 64),\n    batch_size=32,\n    class_mode='categorical')\n\ntest_set = test_datagen.flow_from_directory(\n    'mydata\/test_set',\n    target_size=(64, 64),\n    batch_size=32,\n    class_mode='categorical')\n\nmodel = classifier.fit_generator(\n    training_set,\n    steps_per_epoch=int(steps_per_epoch_user),\n    epochs=int(epochs_user),\n    validation_data = test_set,\n    validation_steps = int(validation_steps_user)\n  )\n\nimport h5py\nclassifier.save('my_model.h5')\n\nprint(model.history.keys()) \n\nimport matplotlib.pyplot as plt\n\nplt.plot(model.history['acc'])\nplt.plot(model.history['val_acc'])\nplt.title('model accuracy')\nplt.ylabel('acc')\nplt.xlabel('epoch')\nplt.legend(['train', 'test'], loc='upper left')\nplt.show()\n\nplt.plot(model.history['loss'])\nplt.plot(model.history['val_loss'])\nplt.title('model loss')\nplt.ylabel('loss')\nplt.xlabel('epoch')\nplt.legend(['train', 'test'], loc='upper left')\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53490276","CreationDate":"2018-11-26T22:47:26.877","ParentId":null,"OwnerUserId":"8634111","Title":null,"Body":"<p>I would add the following code:<\/p>\n\n<p>from sklearn.metrics import confusion_matrix<\/p>\n\n<p>cm = confusion_matrix(Y_test,Y_test_predicted)<\/p>\n\n<p>print('\\n','cm=','\\n',cm)<\/p>\n\n<p>for more descriptive confusion matrix look at scikit-learn documentation in the following link:<\/p>\n\n<p><a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.confusion_matrix.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.confusion_matrix.html<\/a><\/p>\n"}]}
+{"QuestionId":53255054,"AnswerCount":1,"Tags":"<python><keras><deep-learning><computer-vision>","CreationDate":"2018-11-12T01:47:31.717","AcceptedAnswerId":null,"OwnerUserId":5740246.0,"Title":"Weight only the negative or the positive samples in the class","Body":"<p>Algorithm i am trying to implement deals with class imbalances in the dataset. I have 40 classes of CelebA dataset which is a multi label dataset and wants to balance the dataset dynamically. I am using keras but having trouble coming up with the right logic to solve the problem. <\/p>\n\n<p><strong>Algorithm<\/strong><\/p>\n\n<ul>\n<li>Positive and Negative samples in the class should be equal<\/li>\n<li>If positive sample instance are greater in a class then simply select the targeted amount of samples from them and weight the negative instances i.e if batch size is 100 and there are 70 positive instances and 30 negative instances in a class then randomly choose 50 from the positive instances and assign weight 50\/30=5\/3 to negative samples and vice versa in case of negative over represented instances<\/li>\n<\/ul>\n\n<p><strong>Problem<\/strong><\/p>\n\n<ul>\n<li>I am using Keras to implement a deep learning problem with the mentioned algorithm. How can i assign only negative or positive instances a weight instead of all the instances like it does when you use <strong><em>sample_weights<\/em><\/strong>.<\/li>\n<li>If i try to balance the samples by simply duplicating then i run into some sort of circular paradox i.e if i balance one class of over represented instances by removing samples then another class may get under represented and when i duplicate this new under represented class, the previous class may get over represented. <\/li>\n<\/ul>\n\n<p>Kindly let me know what is the most efficient way to solve the mentioned algorithm in Keras. For further help you can check out the research paper here: <a href=\"https:\/\/www.cs.umd.edu\/~emhand\/Papers\/AAAI2018_SelectiveLearning.pdf?fbclid=IwAR13MMtV4TpVSerMZqgmt_t1-8VhD-XUQYJ4qORjXz-mbMDv1W7JqgwrSN8\" rel=\"nofollow noreferrer\">https:\/\/www.cs.umd.edu\/~emhand\/Papers\/AAAI2018_SelectiveLearning.pdf?fbclid=IwAR13MMtV4TpVSerMZqgmt_t1-8VhD-XUQYJ4qORjXz-mbMDv1W7JqgwrSN8<\/a><\/p>\n\n<p>Algorithm is stated on page 3 under <strong>batch balancing<\/strong> heading. <\/p>\n","answers":[{"AnswerId":"53255088","CreationDate":"2018-11-12T01:54:09.127","ParentId":null,"OwnerUserId":"10498246","Title":null,"Body":"<p>You can use class weights like in:<\/p>\n\n<pre><code>    model.fit(X_train, y_train, class_weight=class_weights)\n<\/code><\/pre>\n\n<p>Where class_weights is a dictionary. For example, class_weights = {0: 4, 1: 10, 2:1}, which means that for every 4 samples in class 0, we have 10 samples in class 1 and 1 sample in class 2. <\/p>\n\n<p>In your particular example, you might just set the appropriate parameters for only the positive and negative class of interest.<\/p>\n"}]}
+{"QuestionId":53255677,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-11-12T03:41:33.153","AcceptedAnswerId":null,"OwnerUserId":10631580.0,"Title":"Why model performs poor after normalization?","Body":"<p>I'm using fully connected neural network and I am using normalized data such that every single sample values range from 0 to 1. I have used 100 neurons in first layer and 10 in second layer and used almost 50 lack samples during training. I want to classify my data into two classes. But my networks performance is too low, almost 49 percent on training and test data. I tried to increase the performance by changing the values of hyper parameters. But it didn't work. Can some one please tell me what should I do to get higher performance?<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, nPixels])\nW1 = tf.Variable(tf.random_normal([nPixels, nNodes1], stddev=0.01))\nb1 = tf.Variable(tf.zeros([nNodes1]))\ny1 = tf.nn.relu(tf.matmul(x, W1) + b1)\n\nW2 = tf.Variable(tf.random_normal([nNodes1, nNodes2], stddev=0.01))\nb2 = tf.Variable(tf.zeros([nNodes2]))\ny2 = tf.nn.relu(tf.matmul(y1, W2) + b2)\n\nW3 = tf.Variable(tf.random_normal([nNodes2, nLabels], stddev=0.01))\nb3 = tf.Variable(tf.zeros([nLabels]))\ny = tf.nn.softmax(tf.matmul(y2, W3) + b3)\n\ny_ = tf.placeholder(dtype=tf.float32, shape=[None, 2])\n\ncross_entropy = -1*tf.reduce_sum(y_* tf.log(y), axis=1)    \nloss = tf.reduce_mean(cross_entropy)\n\noptimizer = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\ncorrect_prediction = tf.equal(tf.argmax(y_,axis=1), tf.argmax(y, axis=1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53258511","CreationDate":"2018-11-12T08:45:07.767","ParentId":null,"OwnerUserId":"9703830","Title":null,"Body":"<p>Your computational model knows nothing about \"images\", it only sees numbers. So if you trained it with pixels of values from 0-255, it has learned what \"light\" means, what \"dark\" means and how do these combine to give you whatever target value you try model.<\/p>\n\n<p>And what you did by the normalization is that you forced all pixel to be 0-1. So as far as the model cares, they are all black as night. No surprise that it cannot extract anything meaningful.<\/p>\n\n<p><strong>You need to apply the same input normalization during both training and testing.<\/strong><\/p>\n\n<p>And speaking about normalization for NN models, it is better to normalize to zero mean.<\/p>\n"}]}
+{"QuestionId":53255873,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><lstm>","CreationDate":"2018-11-12T04:11:46.767","AcceptedAnswerId":53257244.0,"OwnerUserId":1757224.0,"Title":"tf.metrics.accuracy and hand-written accuracy function give different results","Body":"<p>I am trying to see how <code>tf.metrics.accuracy<\/code> works. I want to compare batch accuracy results of the function given below<\/p>\n\n<pre><code>with tf.name_scope('Accuracy1'):\n        correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(y, 1))\n        accuracy1 = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name=\"accuracy\")\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>with tf.name_scope('Accuracy2'):\n        accuracy2, accuracy_op = tf.metrics.accuracy(labels=tf.argmax(y, 1), predictions=tf.argmax(predictions, 1))\n<\/code><\/pre>\n\n<p>Minimal working example is provided below:<\/p>\n\n<pre><code>import numpy as np \nimport pandas as pd \nimport tensorflow as tf\nimport math\n\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n\nnum_steps=28\nnum_inputs = 28\nnum_classes = 10\nnum_neurons = 128\nnum_layers = 3\nbatch_size = 500\n\ngraph = tf.Graph()\nwith graph.as_default():\n    with tf.name_scope(\"graph_inputs\"):\n        X = tf.placeholder(tf.float32, [None, num_steps, num_inputs], name='input_placeholder')\n        y = tf.placeholder(tf.float32, [None, num_classes], name='labels_placeholder')\n       output_keep_prob = tf.placeholder_with_default(1.0, shape=(), name =\"output_dropout\")\n\ndef build_lstm_cell(num_neurons, output_keep_prob):\n    \"\"\"Returns a dropout-wrapped LSTM-cell.\n    See https:\/\/stackoverflow.com\/a\/44882273\/2628369 for why this local function is necessary.\n    Returns:\n    tf.contrib.rnn.DropoutWrapper: The dropout-wrapped LSTM cell.\n    \"\"\"\n    initializer = tf.contrib.layers.xavier_initializer()\n    lstm_cell = tf.contrib.rnn.LSTMCell(num_units=num_neurons, initializer=initializer, forget_bias=1.0, state_is_tuple=True, name='LSTM_cell')\n    lstm_cell_drop = tf.contrib.rnn.DropoutWrapper(lstm_cell, output_keep_prob=output_keep_prob)\n    return lstm_cell_drop\n\nwith tf.name_scope(\"LSTM\"):\n    with tf.name_scope(\"Cell\"):\n        multi_layer_cell = tf.contrib.rnn.MultiRNNCell([build_lstm_cell(num_neurons, output_keep_prob) for _ in range(num_layers)], state_is_tuple=True)\n    with tf.name_scope(\"Model\"):\n        outputs, states = tf.nn.dynamic_rnn(cell=multi_layer_cell, inputs=X, swap_memory=False, time_major = False, dtype=tf.float32)#[Batch_size, time_steps, num_neurons]\n    with tf.name_scope(\"Graph_Outputs\"):\n        outputs = tf.transpose(outputs, [1, 0, 2]) # [num_timesteps, batch_size, num_neurons]\n        outputs = tf.gather(outputs, int(outputs.get_shape()[0]) - 1) # [batch_size, num_neurons]\n    with tf.variable_scope('Softmax'):\n        logits =  tf.layers.dense(inputs = outputs, units = num_classes, name=\"logits\") #[Batch_size, num_classes]\n    with tf.name_scope('Predictions'):\n        predictions = tf.nn.softmax(logits, name=\"predictions\")  #[Batch_size, num_classes]\n    with tf.name_scope('Accuracy1'):\n        correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(y, 1))\n        accuracy1 = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name=\"accuracy\")\n    with tf.name_scope('Accuracy2'):\n        accuracy2, accuracy_op = tf.metrics.accuracy(labels=tf.argmax(y, 1), predictions=tf.argmax(predictions, 1))\n    with tf.name_scope('Loss'):\n        xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y)\n        loss = tf.reduce_mean(xentropy, name=\"loss\")\n    with tf.name_scope('Train'):\n        optimizer= tf.train.AdamOptimizer(learning_rate=0.0001)\n        trainer=optimizer.minimize(loss, name=\"training_op\")\n\nwith tf.Session(graph = graph) as sess:\n    tf.global_variables_initializer().run()\n    total_batch = mnist.train.num_examples \/\/ batch_size\n    for batch in range(total_batch):\n        tf.local_variables_initializer().run()\n        xBatch, yBatch = mnist.train.next_batch(batch_size)\n        xBatch = xBatch.reshape((batch_size, num_steps, num_inputs))\n        sess.run(trainer, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})\n        miniBatchAccuracy1 = sess.run(accuracy1, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})\n        print('[hand-written] Batch {} accuracy: {}'.format(batch, miniBatchAccuracy1))\n        accuracy_op_val = sess.run(accuracy_op, feed_dict={X: xBatch, y: yBatch, output_keep_prob: 0.5})\n        miniBatchAccuracy2 = sess.run(accuracy2)\n        print(\"[tf.metrics.accuracy] Batch {} accuracy: {}\".format(batch, miniBatchAccuracy2))\n    sess.close()\n<\/code><\/pre>\n\n<p>I print the accuracy values of each batches using these two approaches and they are different. Should not the results be the same? <\/p>\n\n<pre><code>[hand-written] Batch 0 accuracy: 0.09600000083446503\n[tf.metrics.accuracy] Batch 0 accuracy: 0.09399999678134918\n\n[hand-written] Batch 1 accuracy: 0.1120000034570694\n[tf.metrics.accuracy] Batch 1 accuracy: 0.07800000160932541\n\n[hand-written] Batch 2 accuracy: 0.10199999809265137\n[tf.metrics.accuracy] Batch 2 accuracy: 0.09600000083446503\n\n[hand-written] Batch 3 accuracy: 0.12999999523162842\n[tf.metrics.accuracy] Batch 3 accuracy: 0.12800000607967377\n\n[hand-written] Batch 4 accuracy: 0.1379999965429306\n[tf.metrics.accuracy] Batch 4 accuracy: 0.10199999809265137\n\n[hand-written] Batch 5 accuracy: 0.16200000047683716\n[tf.metrics.accuracy] Batch 5 accuracy: 0.1340000033378601\n\n[hand-written] Batch 6 accuracy: 0.1340000033378601\n[tf.metrics.accuracy] Batch 6 accuracy: 0.12600000202655792\n\n[hand-written] Batch 7 accuracy: 0.12999999523162842\n[tf.metrics.accuracy] Batch 7 accuracy: 0.16200000047683716\n...\n...\n...\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"53257244","CreationDate":"2018-11-12T07:00:34.180","ParentId":null,"OwnerUserId":"8143158","Title":null,"Body":"<p>When measuring the accuracy for both cases, you are passing the <code>dropout rate<\/code> as 0.5. This is the reason its giving two different values. Set the <code>dropout<\/code> value at 1.0 and you should see similar values for both cases.<\/p>\n"}]}
+{"QuestionId":53256136,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><keras-layer>","CreationDate":"2018-11-12T04:52:01.660","AcceptedAnswerId":null,"OwnerUserId":10336700.0,"Title":"Keras Lambda layer error: did not return a tensor","Body":"<p>I am using Lambda to create a self attention layer, but it raise an error that the output of lambda layer is not a tensor.<\/p>\n\n<p>My code:<\/p>\n\n<pre><code>def selfAttention(x):\n    # input shape [None, n_window_sizes, n_hidden]\n    temp_transpose = K.transpose(x) \n    inputs_transpose = K.permute_dimensions(temp_transpose, [2, 0, 1]) # [None, n_hidden, n_window_sizes]\n    temp_weights = tf.matmul(x, inputs_transpose)\n    weights = tf.nn.softmax(temp_weights)\n    output = tf.matmul(weights, x)\n    return output\n<\/code><\/pre>\n\n<p>I call Lambda function as below:<\/p>\n\n<pre><code>attention_input = K.stack([lstm[0], lstm[1], lstm[2]], axis = 1)\nl_attention = Lambda(selfAttention)(attention_input)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53274508","CreationDate":"2018-11-13T05:40:03.330","ParentId":null,"OwnerUserId":"10336700","Title":null,"Body":"<p>Use lambda function to wrap K.stack as follow will solve the problem. <\/p>\n\n<pre><code> attention_input = Lambda(lambda x: K.stack([x[0], x[1], x[2]], axis = 1))([lstm[0], lstm[1], lstm[2]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53256551,"AnswerCount":1,"Tags":"<tensorflow><metadata><conv-neural-network>","CreationDate":"2018-11-12T05:54:31.693","AcceptedAnswerId":null,"OwnerUserId":10638765.0,"Title":"adding metadata to tensorflow tflearn CNN","Body":"<p>I built a simple CNN network for (medical) image classification successfully, using tflearn.  When I tried to add metadata to the CNN, I ran into this problem:ValueError: Cannot feed value of shape (96, 2) for Tensor 'TargetsData\/Y:0', which has shape '(1390, 2)'. Any help is appreciated:<\/p>\n\n<pre><code>#extract pictures (0 thru 4095), next two bytes for the selection, and the rest for metadata\nX, Y, Z  = train_data[:,0:4096],train_data[:,4096:4098], train_data[:,4098:]\nX = X.reshape([-1,64,64,1])\n\n\nnetwork = input_data(shape=[None, 64, 64, 1])\nmdnetwork = input_data(shape=[None, 100])\nnetwork = conv_2d(network, 30, 3, activation='relu')\nnetwork = max_pool_2d(network, 2)\nnetwork = conv_2d(network, 30, 3, activation='relu')\nnetwork = conv_2d(network, 40, 3, activation='relu')\nnetwork = max_pool_2d(network, 2)\nnetwork = conv_2d(network, 40, 3, activation='relu')\nnetwork = conv_2d(network, 40, 3, activation='relu')\nnetwork = conv_2d(network, 30, 3, activation='relu')\nnetwork = max_pool_2d(network, 2)\nnetwork = fully_connected(network, 100, activation='relu')\n\nZt= fully_connected(Z, 100, activation='relu')\nnetwork = merge([network,Zt], 'concat')\n\nnetwork = dropout(network, 0.5)\nnetwork = fully_connected(network, 50, activation='relu')\nnetwork = fully_connected(network, 2, activation='softmax')\n\n# Train using classifier\nnetwork = regression(network, optimizer='adam',\n                 loss='categorical_crossentropy',\n                 learning_rate=0.001)\n\nmodel = tflearn.DNN(network, tensorboard_verbose=3)\n\nmodel.fit([np.array(X).reshape(-1, 64, 64, 1), np.array(Z).reshape(-1, 100)], Y, n_epoch=5, shuffle=True, validation_set=0,\n    show_metric=True, batch_size=96, run_id='my_cnn')\n\nmodel.save('my_cnn.tflearn')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53348963","CreationDate":"2018-11-17T06:52:49.070","ParentId":null,"OwnerUserId":"10638765","Title":null,"Body":"<p>It is good I can answer my question here! Anyway, I found the issue in my code above. It was a simple error.  The error message led me astray.  Here is the solution: Replace this code snippet<\/p>\n\n<pre><code>Zt= fully_connected(Z, 100, activation='relu')\nnetwork = merge([network,Zt], 'concat')\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>network = merge([network, mdnetwork], 'concat')\n<\/code><\/pre>\n\n<p>Thanks for those who read my request.  Let me know if there are other options.<\/p>\n"}]}
+{"QuestionId":53256877,"AnswerCount":4,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-12T06:25:31.220","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"How to convert keras(h5) file to a tflite file?","Body":"<p>I got an keras(h5) file. I need to convert it to tflite??\nI researched, First i need to go via h5 -> pb -> tflite\n(because h5 - tflite sometimes results in some issue)<\/p>\n","answers":[{"AnswerId":"57224367","CreationDate":"2019-07-26T17:04:40.863","ParentId":null,"OwnerUserId":"4311535","Title":null,"Body":"<p>There is one factor, which you <strong>must<\/strong> to consider. You need to <strong>change the learning phase<\/strong>, before converting. It's super important, when you have Dropout or Batch Normalization. You can take a look at <a href=\"https:\/\/ai-pool.com\/d\/keras_model_to_tflite\" rel=\"nofollow noreferrer\">'Keras model to tflite'<\/a> or <a href=\"https:\/\/ai-pool.com\/d\/problem_after_converting_keras_model_into_tensorflow_pb\" rel=\"nofollow noreferrer\">'Problem after converting keras model into Tensorflow pb'<\/a> discussions<\/p>\n"},{"AnswerId":"57995945","CreationDate":"2019-09-18T15:10:46.477","ParentId":null,"OwnerUserId":"5944114","Title":null,"Body":"<p>Only some specific version of Tensorflow and Keras works properly in all the os. I even tried toco command line but it has issues too. Use tensorflow==1.13.0-rc1 and keras==2.1.3<\/p>\n\n<p>and then after this will work<\/p>\n\n<pre><code>from tensorflow.contrib import lite\nconverter = lite.TFLiteConverter.from_keras_model_file( 'model.h5' ) # Your model's name\nmodel = converter.convert()\nfile = open( 'model.tflite' , 'wb' ) \nfile.write( model )\n<\/code><\/pre>\n"},{"AnswerId":"53263124","CreationDate":"2018-11-12T13:23:12.640","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<pre><code>from tensorflow.contrib import lite\nconverter = lite.TFLiteConverter.from_keras_model_file( 'model.h5')\ntfmodel = converter.convert()\nopen (\"model.tflite\" , \"wb\") .write(tfmodel)\n<\/code><\/pre>\n\n<p>You can use the TFLiteConverter to directly convert .h5 files to .tflite file.\n<strong>This does not work on Windows.<\/strong><\/p>\n\n<p>For Windows, use this <a href=\"https:\/\/drive.google.com\/file\/d\/1IUIn9ffk5ICKujqPyuGaHL2irQ9Wmtpm\/view?usp=drivesdk\" rel=\"noreferrer\">Google Colab notebook<\/a> to convert. Upload the .h5 file and it will convert it .tflite file.<\/p>\n\n<p>Follow, if you want to try it yourself :<\/p>\n\n<ol>\n<li>Create a Google Colab Notebook. In the left top corner, click the \"UPLOAD\" button and upload your .h5 file.<\/li>\n<li><p>Create a code cell and insert this code.<\/p>\n\n<pre><code>from tensorflow.contrib import lite\nconverter = lite.TFLiteConverter.from_keras_model_file( 'model.h5' ) # Your model's name\nmodel = converter.convert()\nfile = open( 'model.tflite' , 'wb' ) \nfile.write( model )\n<\/code><\/pre><\/li>\n<li><p>Run the cell. You will get a model.tflite file. Right click on the file and select \"DOWNLOAD\" option.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"53267871","CreationDate":"2018-11-12T18:14:51.150","ParentId":null,"OwnerUserId":"6150310","Title":null,"Body":"<p>Use this tool to convert to Tensorflow:\n<a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"nofollow noreferrer\">Keras to Tensorflow Converter<\/a><\/p>\n\n<p>Then use this one of the 3 code snippets(depending on how you saved your TF file) to convert that file to TF-Lite:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code># Converting a GraphDef from session.\r\nconverter = lite.TFLiteConverter.from_session(sess, in_tensors, out_tensors)\r\ntflite_model = converter.convert()\r\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\r\n\r\n# Converting a GraphDef from file.\r\nconverter = lite.TFLiteConverter.from_frozen_graph(\r\n  graph_def_file, input_arrays, output_arrays)\r\ntflite_model = converter.convert()\r\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\r\n\r\n# Converting a SavedModel.\r\nconverter = lite.TFLiteConverter.from_saved_model(saved_model_dir)\r\ntflite_model = converter.convert()<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53257165,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2018-11-12T06:53:49.253","AcceptedAnswerId":53266448.0,"OwnerUserId":10506048.0,"Title":"Preprocessing the image on the server","Body":"<p>How can I implement the resize function in this construction?\nFor example, <code>in_image<\/code> has <code>shape = (845, 594, 3)<\/code>, but I want to resize this image to  <code>shape = (299, 299, 3)<\/code><\/p>\n\n<pre><code>def main(_):\n    with tf.Graph().as_default() as graph:\n        input_graph = FLAGS.input_graph\n        saved_model_dir = FLAGS.saved_model_dir\n        # Read in the export graph\n        with tf.gfile.FastGFile(input_graph, 'rb') as f:\n            graph_def = tf.GraphDef()\n            graph_def.ParseFromString(f.read())\n            tf.import_graph_def(graph_def, name='')\n\n        # Define SavedModel Signature (inputs and outputs)\n        in_image = graph.get_tensor_by_name('input:0')\n        inputs = {'image_bytes': tf.saved_model.utils.build_tensor_info(in_image)}\n\n        out_classes = graph.get_tensor_by_name('InceptionV3\/Predictions\/Reshape_1:0')\n        outputs = {'prediction': \n                    tf.saved_model.utils.build_tensor_info(out_classes)}\n\n        signature = tf.saved_model.signature_def_utils.build_signature_def(\n            inputs=inputs,\n            outputs=outputs,\n            method_name='tensorflow\/serving\/predict'\n        ) \n\n        with tf.Session(graph=graph) as sess:\n            # Save out the SavedModel.\n            b = saved_model_builder.SavedModelBuilder(saved_model_dir)\n            b.add_meta_graph_and_variables(sess,\n                                    [tf.saved_model.tag_constants.SERVING],\n                                    signature_def_map={'serving_default': \n                                                        signature})\n           b.save()\n\nif __name__ == '__main__':\n  tf.app.run()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53266448","CreationDate":"2018-11-12T16:36:09.107","ParentId":null,"OwnerUserId":"10506048","Title":null,"Body":"<p>I figured it out:<\/p>\n\n<pre><code>def main(_):\n    with tf.Graph().as_default() as graph:\n        input_graph = FLAGS.input_graph\n        saved_model_dir = FLAGS.saved_model_dir\n        # Read in the export graph\n        with tf.gfile.FastGFile(input_graph, 'rb') as f:\n            graph_def = tf.GraphDef()\n            graph_def.ParseFromString(f.read())\n            tf.import_graph_def(graph_def, name='')\n\n        serialized_tf_example = tf.placeholder(tf.string, name='b64')\n        jpeg = preprocess_image(serialized_tf_example)\n\n        out, = tf.import_graph_def(graph.as_graph_def(),\n                         input_map={'input:0': jpeg},\n                         return_elements = ['InceptionV3\/Predictions\/Reshape_1:0'])\n\n        inputs = {'inputs': tf.saved_model.utils.build_tensor_info(serialized_tf_example)}\n        outputs = {'prediction': tf.saved_model.utils.build_tensor_info(out)}\n\n        signature = tf.saved_model.signature_def_utils.build_signature_def(\n            inputs=inputs,\n            outputs=outputs,\n            method_name='tensorflow\/serving\/predict'\n        )\n        with tf.Session(graph=graph) as sess:\n            # Save out the SavedModel.\n            b = saved_model_builder.SavedModelBuilder(saved_model_dir)\n            b.add_meta_graph_and_variables(sess,\n                                    [tf.saved_model.tag_constants.SERVING],\n                                    signature_def_map={'serving_default': signature})\n            b.save()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53257506,"AnswerCount":0,"Tags":"<scikit-learn><keras>","CreationDate":"2018-11-12T07:23:38.257","AcceptedAnswerId":null,"OwnerUserId":10639147.0,"Title":"Can I use sklearn cross_val_score with two different input?","Body":"<p>My problem is:\n'can I use sklearn cross_val_score with two different input?'<\/p>\n\n<p>I concatenated two LSTM model with each different X and with same target Y.\nAnd I tried to check cross_val_score. but it isn't work.<\/p>\n\n<pre><code>def concat():\ninput1 = Input(shape=(6,1))\npay = LSTM(10)(input1)\npay = Dense(10, activation='relu')(pay)\n\n# DEMO_data net\ninput2 = Input(shape=(6,1))\nbill = LSTM(10)(input2)\nbill = Dense(10, activation='relu')(bill)\n\nmerge = concatenate([pay, bill])\n\nhidden1 = Dense(10, activation='relu')(merge)\n\noutput = Dense(1, activation='sigmoid')(hidden1)\nmodel = Model(inputs=[input1, input2], outputs=output)\n\nmodel.summary()\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics='accuracy'])\n\nreturn model\n<\/code><\/pre>\n\n<p>and I tried this<\/p>\n\n<pre><code>dnn = KerasClassifier(build_fn=concat, batch_size=BATCH_SIZE, epochs=NB_EPOCH, verbose=VERBOSE)\n\n# Evaluate using 5-fold cross validation\ncv_results = cross_val_score(dnn, PAY_data, BILL_data, y, cv=5)\nprint(cv_results)\nprint(cv_results.mean(), cv_results.std())\n<\/code><\/pre>\n\n<p>but it didn't work with this below error message<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/itaWt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/itaWt.png\" alt=\"image of error message\"><\/a><\/p>\n\n<p>is it impossible to use cross_val_score with two different input?<\/p>\n","answers":[]}
+{"QuestionId":53257563,"AnswerCount":1,"Tags":"<tensorflow><rnn><loss>","CreationDate":"2018-11-12T07:28:46.557","AcceptedAnswerId":null,"OwnerUserId":10201498.0,"Title":"RNN use mean square error does not converge","Body":"<p>I am learning RNN through <a href=\"https:\/\/medium.com\/@erikhallstrm\/hello-world-rnn-83cd7105b767\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@erikhallstrm\/hello-world-rnn-83cd7105b767<\/a>. I change the loss function to mean square error and found it does not converge. The output is stuck at 0.5. Somehow, I feel the mistake is inside<\/p>\n\n<pre><code>midlosses = [tf.squeeze(logits)-tf.squeeze(labels)  for logits, labels in zip(logits_series,labels_series)]\n<\/code><\/pre>\n\n<p>But I don't how. I am not familiar with datatype. This may be a silly question. In case I don't make myself clear, the full code is below:<\/p>\n\n<pre><code>from __future__ import print_function, division\nimport numpy as np\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\n\nnum_epochs = 100\ntotal_series_length = 50000\ntruncated_backprop_length = 15\nstate_size = 4\nnum_classes = 1\necho_step = 3\nbatch_size = 5\nnum_batches = total_series_length\/\/batch_size\/\/truncated_backprop_length\n\ndef generateData():\n    x = np.array(np.random.choice(2, total_series_length, p=[0.5, 0.5]))\n    y = np.roll(x, echo_step)\n    y[0:echo_step] = 0\n\n    x = x.reshape((batch_size, -1))  # The first index changing slowest, subseries as rows\n    y = y.reshape((batch_size, -1))\n\n    return (x, y)\n\ntf.reset_default_graph()\nbatchX_placeholder = tf.placeholder(tf.float32, [batch_size, truncated_backprop_length])\nbatchY_placeholder = tf.placeholder(tf.float32, [batch_size, truncated_backprop_length])\n\ninit_state = tf.placeholder(tf.float32, [batch_size, state_size])\n\nW = tf.Variable(np.random.rand(state_size+1, state_size), dtype=tf.float32)\nb = tf.Variable(np.zeros((1,state_size)), dtype=tf.float32)\n\nW2 = tf.Variable(np.random.rand(state_size, num_classes),dtype=tf.float32)\nb2 = tf.Variable(np.zeros((1,num_classes)), dtype=tf.float32)\n\n# Unpack columns\ninputs_series = tf.unstack(batchX_placeholder, axis=1)\nlabels_series = tf.unstack(batchY_placeholder, axis=1)\n\n# Forward pass\ncurrent_state = init_state\nstates_series = []\nfor current_input in inputs_series:\n    current_input = tf.reshape(current_input, [batch_size, 1])\n    input_and_state_concatenated = tf.concat([current_input, current_state],axis=1)  # Increasing number of columns\n\n    next_state = tf.tanh(tf.matmul(input_and_state_concatenated, W) + b)  # Broadcasted addition\n    states_series.append(next_state)\n    current_state = next_state\n\nlogits_series = [tf.matmul(state, W2) + b2 for state in states_series] \n#Loss function HERE\nmidlosses = [tf.squeeze(logits)-tf.squeeze(labels)  for logits, labels in zip(logits_series,labels_series)]\nlosses = tf.square(midlosses)\ntotal_loss = tf.reduce_mean(losses)\ntrain_step = tf.train.AdagradOptimizer(0.3).minimize(total_loss)\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    loss_list = []\n\n    for epoch_idx in range(num_epochs):\n        x,y = generateData()\n        _current_state = np.zeros((batch_size, state_size))\n\n        print(\"New data, epoch\", epoch_idx)\n\n        for batch_idx in range(num_batches):\n            start_idx = batch_idx * truncated_backprop_length\n            end_idx = start_idx + truncated_backprop_length\n\n            batchX = x[:,start_idx:end_idx]\n            batchY = y[:,start_idx:end_idx]\n\n            _total_loss, _train_step, _current_state,_logits_series,_midlosses = sess.run(\n                [total_loss, train_step, current_state,logits_series,midlosses],\n                feed_dict={\n                    batchX_placeholder:batchX,\n                    batchY_placeholder:batchY,\n                    init_state:_current_state\n                }) \n            loss_list.append(_total_loss)\n            if batch_idx%100 == 0:\n                print(\"Step\",batch_idx, \"Loss\", _total_loss) \n<\/code><\/pre>\n","answers":[{"AnswerId":"53289028","CreationDate":"2018-11-13T20:30:15.740","ParentId":null,"OwnerUserId":"10201498","Title":null,"Body":"<p>Just need to replace <\/p>\n\n<pre><code>logits_series = [tf.matmul(state, W2) + b2 for state in states_series] \n<\/code><\/pre>\n\n<p>by <\/p>\n\n<pre><code>logits_series = [tf.squeeze(tf.matmul(state, W2) + b2) for state in states_series] #Broadcasted addition\n<\/code><\/pre>\n\n<p>Problem can solved.<\/p>\n"}]}
+{"QuestionId":53257986,"AnswerCount":0,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2018-11-12T08:04:23.123","AcceptedAnswerId":null,"OwnerUserId":10052998.0,"Title":"Prediction is slower when model is loaded than if it is fited during the process","Body":"<p>I have a strange issue, the DNN.predict method is quite slower when I load my model's weight than when I train with the fit method. I've also noted that when I run a prediction over a batch of images, it's getting faster and faster to predict.<\/p>\n\n<p>Here is my code<\/p>\n\n<pre><code>class Reseau(object):\n\ndef init(self, img_size, lr=-1, activation=\" \"):\ntf.logging.set_verbosity(tf.logging.ERROR)\nself.lr = lr\nself.activation = activation\nself.img_size = img_size\nself.alreadySaved = 0\n\ndef setting(self, X, Y, test_x, test_y, nbEpoch):\ntflearn.init_graph(num_cores=32, gpu_memory_fraction=1)\nwith tf.device(\"\/device:GPU:0\"):\nconvnet = input_data(shape=[None, self.img_size, self.img_size, 3], name='input')\n\n        convnet = conv_2d(convnet, 32, 5, activation=self.activation)\n        convnet = max_pool_2d(convnet, 5)\n\n        convnet = conv_2d(convnet, 64, 5, activation=self.activation)\n        convnet = max_pool_2d(convnet, 5)\n\n        convnet = conv_2d(convnet, 128, 5, activation=self.activation)\n        convnet = max_pool_2d(convnet, 5)\n\n        convnet = conv_2d(convnet, 64, 5, activation=self.activation)\n        convnet = max_pool_2d(convnet, 5)\n\n        convnet = conv_2d(convnet, 32, 5, activation=self.activation)\n        convnet = max_pool_2d(convnet, 5)\n\n        convnet = flatten(convnet)\n\n        convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')\n        convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')\n        convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')\n        convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')\n        convnet = dropout(convnet, 0.8)\n\n        convnet = fully_connected(convnet, 2, activation='softmax')\n        convnet = regression(convnet, optimizer='adam', learning_rate=self.lr, loss='categorical_crossentropy', name='targets')\n\n        self.model = tflearn.DNN(convnet, tensorboard_dir='log')\n\n        if self.alreadySaved == 0:\n            self.model.fit({'input': X}, {'targets': Y}, n_epoch=nbEpoch, validation_set=({'input': test_x}, {'targets': test_y}), snapshot_step=500, show_metric=True, run_id=\"model\")\n            self.model.save(\".\/model\")\n        else:\n            self.model.load(\".\/model\", weights_only=True)\n    return self.model\n\ndef predire(self, img, label):\nimage = array(img).reshape(1, self.img_size,self.img_size,3)\nmodel_out = self.model.predict(image)\nrep = 0\n\n    if np.argmax(model_out) == np.argmax(label): rep = 1\n    else: rep = 0\n\n    return rep\n<\/code><\/pre>\n\n<p>Here is a part of my main<\/p>\n\n<pre><code>reseau.setting(X, Y, test_x, test_y, NB_EPOCH)\n\nX = np.array([i[0] for i in test]).reshape(-1,IMG_SIZE,IMG_SIZE,3)\nY = [i[1] for i in test]\ncpt = 0\n\nvrai = 0\nstart_time = time.time()\n\nfor i in range(20):\n    cpt = 0\n\n    vrai = 0\n    start_time = time.time()\n    for img in tqdm(X):\n        prediction = reseau.predire(img, Y[cpt])\n        cpt += 1\n        if prediction == 1:\n            vrai += 1\n<\/code><\/pre>\n\n<p>As you can see, I predict the same batch of images 20 times. The first time is always slower than the other ones (without fitting, I predict 82 images the first and then 340 a second, with fitting, it's 255 images the first time and 340 a seconde then).<\/p>\n\n<p>I'm really out of idea to fix this.<\/p>\n","answers":[]}
+{"QuestionId":53258000,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2018-11-12T08:05:25.423","AcceptedAnswerId":null,"OwnerUserId":4869481.0,"Title":"How to convert a HED model to Tensorflow Lite model","Body":"<p>I'm creating a model from using <a href=\"https:\/\/github.com\/moabitcoin\/holy-edge\" rel=\"nofollow noreferrer\">holy-edge<\/a>. Then, I would like to convert this model to Tensorflow Lite for Mobile usage. I followed <a href=\"https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc\" rel=\"nofollow noreferrer\">steps here<\/a> and was able to convert the model to .pg file with output_node_names=predictions . For next step, I need to convert .pg file to Tensorflow Lite model by using <a href=\"https:\/\/www.tensorflow.org\/lite\/devguide\" rel=\"nofollow noreferrer\">toco<\/a> but I don't know where to find some required parameters such as input_array, output_array, input_shape and output_node_names. <\/p>\n\n<p>Does anyone know how to convert a HED model to Tensorflow Lite mode? Please help.<\/p>\n\n<p>Thanks,\nDuc<\/p>\n","answers":[{"AnswerId":"57273322","CreationDate":"2019-07-30T13:49:48.360","ParentId":null,"OwnerUserId":"4663377","Title":null,"Body":"<p>You can use this code:<\/p>\n\n<pre><code>import tensorflow as tf\ngf = tf.GraphDef()   \nm_file = open('frozen_inference_graph.pb','rb')\ngf.ParseFromString(m_file.read())\n\nwith open('somefile.txt', 'a') as the_file:\n    for n in gf.node:\n        the_file.write(n.name+'\\n')\n\nfile = open('somefile.txt','r')\ndata = file.readlines()\nprint (\"\\noutput name = \")\nprint (data[len(data)-1])\n\nprint (\"Input name = \")\nfile.seek ( 0 )\nprint (file.readline())\n<\/code><\/pre>\n\n<p>In my case I had <\/p>\n\n<pre><code>output name: SemanticPredictions\ninput name: ImageTensor\n<\/code><\/pre>\n"}]}
+{"QuestionId":53258219,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch><fast-ai>","CreationDate":"2018-11-12T08:22:25.853","AcceptedAnswerId":null,"OwnerUserId":3171335.0,"Title":"Python - importing fastai results in SyntaxError","Body":"<p>I am trying to setup an ML model using fastai and have to do the following imports: <\/p>\n\n<pre><code>import fastai.models\nimport fastai.nlp\nimport fastai.dataset\n<\/code><\/pre>\n\n<p>However, it gives me the following error by the <code>fastai<\/code> imports. <\/p>\n\n<pre><code>Python 2.7.15rc1 (default, Apr 15 2018, 21:51:34) \n[GCC 7.3.0] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import fastai.nlp\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/fastai\/nlp.py\", line 172\n    if os.path.isdir(path): paths=glob(f'{path}\/*.*')\n                                                   ^\nSyntaxError: invalid syntax\n<\/code><\/pre>\n\n<p>Apparently, the character <code>f<\/code> in <code>glob(f'{path}\/*.*')<\/code> is causing the error. I fixed the error by removing <code>f<\/code>, but it seems that there are lots of these errors in the <code>fastai<\/code> library. <\/p>\n\n<p>My current thought is that I am using an incorrect python version. Could anyone give me some pointer? <\/p>\n","answers":[{"AnswerId":"53258674","CreationDate":"2018-11-12T08:55:18.960","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Strings in the shape of:<\/p>\n\n<pre><code>f'{path}\/*.*'\n<\/code><\/pre>\n\n<p>are called <em>f-strings<\/em> and were introduced in Python3.6.\nThat's why you get the <code>SyntaxError<\/code> - for versions lower than Python3.6 a <code>SyntaxError<\/code> will be raised, as this syntax just doesn't exist in lower versions. <\/p>\n\n<p>So obviously <em>fast-ai<\/em> is programmed for Python3.6 or higher.<\/p>\n\n<p>When you take a look at the <a href=\"https:\/\/pypi.org\/project\/fastai\/\" rel=\"nofollow noreferrer\">installation issues (you have to scroll down a bit)<\/a>,<br>\nyou can see under <strong>Is My System Supported?<\/strong> the first point:<\/p>\n\n<blockquote>\n  <p><strong>Python: You need to have python 3.6 or higher<\/strong><\/p>\n<\/blockquote>\n\n<p>So I'm afraid updating your python is the easiest way to solve the problem!<\/p>\n\n<hr>\n\n<p><em>If you like to learn more about f-strings you can take a look here: <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0498\/\" rel=\"nofollow noreferrer\">https:\/\/www.python.org\/dev\/peps\/pep-0498\/<\/a><\/em><\/p>\n"}]}
+{"QuestionId":53258236,"AnswerCount":4,"Tags":"<python><tensorflow><keras><deep-learning><ibm-watson>","CreationDate":"2018-11-12T08:23:21.013","AcceptedAnswerId":null,"OwnerUserId":3966036.0,"Title":"How to save keras model in WML repository in Watson Studio?","Body":"<p>I am trying to deploy Keras model by training on MNIST dataset on Watson studio but unable to save and successfully deploy it. <\/p>\n\n<p>When I am trying to save the model object, it says it can't save Sequential Object. \nWhen I am trying to convert hd5 to tgz and save it, it gets saved but on deployment I get error<\/p>\n\n<pre><code>\"{\"code\":\"load_model_failure\",\"message\":\"SavedModel file does not exist at: \/opt\/ibm\/s...\"\n<\/code><\/pre>\n\n<p>When I am trying to deploy hd5 file, it says its not in compressed format.<\/p>\n\n<p>Can any help me how exactly to save keras model and deploy it on watson studio?<\/p>\n\n<pre><code># \n\nconvert class vectors to binary class matrices\ny_train = keras.utils.to_categorical(y_train, num_classes)\ny_test = keras.utils.to_categorical(y_test, num_classes)\n\nmodel = Sequential()\nmodel.add(Conv2D(32, kernel_size=(3, 3),\n                 activation='relu',\n                 input_shape=input_shape))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\nmodel.add(Flatten())\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes, activation='softmax'))\n\nmodel.compile(loss=keras.losses.categorical_crossentropy,\n              optimizer=keras.optimizers.Adadelta(),\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train,\n          batch_size=batch_size,\n          epochs=epochs,\n          verbose=1,\n          validation_data=(x_test, y_test))\nscore = model.evaluate(x_test, y_test, verbose=0)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])\n\nmodel_result_path = \"keras_model.h5\"\nmodel.save(model_result_path)\n\npublished_model = client.repository.store_model(model='keras_model.h5', meta_props=model_props,training_data=x_train, training_target=y_train)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53267546","CreationDate":"2018-11-12T17:51:46.390","ParentId":null,"OwnerUserId":"7639069","Title":null,"Body":"<p>Have you tried just passing the Python model object itself to the <code>store_model<\/code> function?<\/p>\n\n<p>For example, see section 4.2 of this sample notebook:\n<a href=\"https:\/\/dataplatform.cloud.ibm.com\/exchange\/public\/entry\/view\/1438a61212a64ac435c837ba040e6934\" rel=\"nofollow noreferrer\">https:\/\/dataplatform.cloud.ibm.com\/exchange\/public\/entry\/view\/1438a61212a64ac435c837ba040e6934<\/a><\/p>\n"},{"AnswerId":"53293055","CreationDate":"2018-11-14T04:06:42.640","ParentId":null,"OwnerUserId":"7445374","Title":null,"Body":"<p>can you try compressing the h5 file(i.e forming a tar.gz) and then try giving it to the client.repository.store_model instead of directly giving .h5 file.<\/p>\n"},{"AnswerId":"53316247","CreationDate":"2018-11-15T09:30:23.460","ParentId":null,"OwnerUserId":"10656702","Title":null,"Body":"<p>You will have to provide the path to the compressed keras file.\nE.g.:<\/p>\n\n<pre><code>keras_file_path = \"\/Users\/jsmith\/keras\/ker_seq_mnist_model.tar.gz\"\npublished_model = client.repository.store_model(model=keras_file_path, meta_props=model_props,training_data=x_train, training_target=y_train)\n<\/code><\/pre>\n"},{"AnswerId":"56029138","CreationDate":"2019-05-07T19:14:45.267","ParentId":null,"OwnerUserId":"6741838","Title":null,"Body":"<p><strong>Step 1<\/strong>: Save your model to .h5 file.<\/p>\n\n<pre><code>model_result_path = \"keras_model.h5\"\nmodel.save(model_result_path)\n<\/code><\/pre>\n\n<p><strong>Step 2<\/strong>: compress .h5 file to tgz.<\/p>\n\n<pre><code>!tar -zcvf keras_model.tgz keras_model.h5\n<\/code><\/pre>\n\n<p><strong>Step 3<\/strong>: <strong>Important ><\/strong>  For deploying a keras model, it is mandatory to pass the FRAMEWORK_LIBRARIES along with other metaprops. <a href=\"https:\/\/wml-api-pyclient.mybluemix.net\/index.html?highlight=store_model#client.Repository.store_model\" rel=\"nofollow noreferrer\">store_model WML documentation<\/a>.<\/p>\n\n<pre><code>metadata = {\nclient.repository.ModelMetaNames.NAME: 'Image-classifier',\nclient.repository.ModelMetaNames.FRAMEWORK_NAME: 'tensorflow',\nclient.repository.ModelMetaNames.FRAMEWORK_VERSION: '1.5',\nclient.repository.ModelMetaNames.FRAMEWORK_LIBRARIES:[{'name':'keras', 'version': '2.1.3'}]\n}\n<\/code><\/pre>\n\n<p><strong>Step 4<\/strong>: Store the model.<\/p>\n\n<pre><code>published_model = client.repository.store_model(model= 'keras_model.tgz', meta_props=metadata, training_data= X_train,training_target= y_train)\n<\/code><\/pre>\n\n<p><strong>step 5<\/strong>: Deploy the model.<\/p>\n\n<pre><code>model_id = published_model[\"metadata\"][\"guid\"]\n\nmodel_deployment_details = client.deployments.create(artifact_uid=model_id, name=\"deployment_name\" )\n<\/code><\/pre>\n"}]}
+{"QuestionId":53258661,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2018-11-12T08:54:37.143","AcceptedAnswerId":null,"OwnerUserId":6555133.0,"Title":"How to serve retrain model with tensorflow-serving and RESTful API","Body":"<p>I am new to Tensorflow. Just managed to build a simple image classifier with this tutorial <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/\" rel=\"nofollow noreferrer\">https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/<\/a><\/p>\n\n<p>I am trying to serve the model with Docker. It seems that I need to export trained model to SavedModel <\/p>\n\n<p>I managed to create SavedModel using answer in this question: <a href=\"https:\/\/stackoverflow.com\/questions\/37237940\/how-to-serve-retrained-inception-model-using-tensorflow-serving\">How to serve retrained Inception model using Tensorflow Serving?<\/a><\/p>\n\n<p>The problem is I have no I idea how to call image classification with RESTful API.<\/p>\n","answers":[]}
+{"QuestionId":53259021,"AnswerCount":0,"Tags":"<python><python-3.x><set><pytorch><tensor>","CreationDate":"2018-11-12T09:16:34.847","AcceptedAnswerId":null,"OwnerUserId":1269281.0,"Title":"Get PyTorch tensor values into python set","Body":"<p>I've a set of pytorch tensors of integers representing image masks (shape is <code>(1, x, x)<\/code>), I don't have the labels and I need to grab them manually from this pytorch tensor.<\/p>\n\n<p>Is there an effective way to get these values into a set without having to iterate over every cell.<\/p>\n","answers":[]}
+{"QuestionId":53259434,"AnswerCount":1,"Tags":"<python><machine-learning><keras><progress-bar><conv-neural-network>","CreationDate":"2018-11-12T09:43:03.947","AcceptedAnswerId":53262068.0,"OwnerUserId":10639632.0,"Title":"AttributeError when training CNN 1D with Python Keras","Body":"<p>I have tried to build a CNN 1D but the interpreter says me:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'ProgbarLogger' object has no attribute 'log_values'<\/p>\n<\/blockquote>\n\n<p>Here is the code snippet:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv1D(200, 20, activation='relu', padding='same',input_shape=(1154,1024))\nprint(model.summary())\nmodel.compile(optimizer='adam',\n          loss='sparse_categorical_crossentropy',\n          metrics=['accuracy'])\nhistory=model.fit(X, y,batch_size=10, epochs=25,validation_split=0.7)\n<\/code><\/pre>\n\n<p>and this is the error:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nconv1d_1 (Conv1D)            (None, 1154, 200)         4096200   \n=================================================================\nTotal params: 4,096,200\nTrainable params: 4,096,200\nNon-trainable params: 0\n_________________________________________________________________\nNone\n\nTrain on 0 samples, validate on 1 samples\nEpoch 1\/25\nTraceback (most recent call last):\n  File \"binary_classification.py\", line 59, in &lt;module&gt;\n    history=model.fit(X, y,batch_size=10, epochs=25,validation_split=0.7)\n  File \"\/home\/isabella\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\",\nline 1039, in fit\n    validation_steps=validation_steps)\n  File \"\/home\/isabella\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py\",\nline 217, in fit_loop\n    callbacks.on_epoch_end(epoch, epoch_logs)\n  File \"\/home\/isabella\/.local\/lib\/python3.6\/site-packages\/keras\/callbacks.py\",\nline 79, in on_epoch_end\n    callback.on_epoch_end(epoch, logs)\n  File \"\/home\/isabella\/.local\/lib\/python3.6\/site-packages\/keras\/callbacks.py\",\nline 338, in on_epoch_end\n    self.progbar.update(self.seen, self.log_values)\nAttributeError: 'ProgbarLogger' object has no attribute 'log_values'\n\nX shape is :(1,1154,1024),\ny shape is :(1,1154, 1 )\n<\/code><\/pre>\n","answers":[{"AnswerId":"53262068","CreationDate":"2018-11-12T12:19:44.527","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you look carefully you will see this line right before stack trace output:<\/p>\n\n<pre><code>Train on 0 samples, validate on 1 samples\n<\/code><\/pre>\n\n<p>There is no training data! Why? That's because you have set the <code>validation_split<\/code> to <code>0.7<\/code> so at first 70% of the data points in <code>X<\/code> (and <code>y<\/code>) are put aside for validation and the remaining 30% is used for training. Probably the number of data points in <code>X<\/code> is less than 4 and therefore its 30% would amount to less than 1 which means zero data points remains for training. Either use more than 4 data points or remove the <code>validation_split<\/code> argument (or lower it such that at least one sample remains for training).<\/p>\n"}]}
+{"QuestionId":53259645,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><opencv3.0><object-detection><bounding-box>","CreationDate":"2018-11-12T09:55:21.823","AcceptedAnswerId":null,"OwnerUserId":9709132.0,"Title":"Check for bounding box availability in only certain areas of the image\/video frame","Body":"<p>I am trying to detect cars in a certain area of live video stream. For that I used Tensorflow's object detection API. Now, the detection is fair enough and almost all the cars in the live video stream are detected as \"car\" with bounding boxes around them and with some amount of detection confidence score in percentage.<\/p>\n\n<p>My question is how do I check for availability of only the desired on required bounding boxes?<\/p>\n\n<p>For example, since the desired area and the camera used for detection are both fixed in position, I used OpenCV's <code>cv2.rectangle()<\/code> function and passed the <code>(x1,y1)<\/code> and <code>(x2,y2)<\/code> co-ordinates of the desired area. So now, I have <strong>constant<\/strong> rectangular box around that area. My task is to somehow know that the car has arrived in this marked rectangle area by printing a logging message \"detected\" to the Ubuntu terminal.<\/p>\n\n<p>I am having difficulties comparing bounding box co-ordinates with the rectangle co-ordinates. So the question arises how to <\/p>\n\n<ol>\n<li>Grab only required bounding boxes(thereby required detected cars)?<\/li>\n<li>detect whenever these bounding boxes are inside the rectangle\/marked area?<\/li>\n<\/ol>\n\n<p>Here is the code I used.<\/p>\n\n<pre><code>import numpy as np\nimport os\nimport six.moves.urllib as urllib\nimport sys\nimport tarfile\nimport tensorflow as tf\nimport zipfile\n\nfrom collections import defaultdict\nfrom io import StringIO\nfrom PIL import Image\n\nimport cv2\ncap = cv2.VideoCapture(0)\n# This is needed since the notebook is stored in the object_detection \nfolder.\nsys.path.append(\"..\")\nfrom object_detection.utils import ops as utils_ops\n\nif tf.__version__ != '1.10.1':\n  raise ImportError('Please upgrade your tensorflow installation to \nv1.10.1* or later!')\n\n\n# ## Env setup\n\n# In[3]:\n\n# ## Object detection imports\n# Here are the imports from the object detection module.\n\n# In[5]:\n\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util\n\n# # Model preparation \n\n# ## Variables\n\n# Any model exported using the `export_inference_graph.py` tool can be \nloaded here simply by changing `PATH_TO_FROZEN_GRAPH` to point to a \nnew .pb file.  \n# \n# By default we use an \"SSD with Mobilenet\" model here. See the \n[detection model zoo] \n\n# In[6]:\n# What model to download.\nMODEL_NAME = 'car_inference_graph'\n\n# Path to frozen detection graph. This is the actual model that is \nused for the object detection.\nPATH_TO_CKPT = MODEL_NAME + '\/frozen_inference_graph.pb'\n\n# List of the strings that is used to add correct label for each box.\nPATH_TO_LABELS = os.path.join('training', 'object-detection.pbtxt')\n\nNUM_CLASSES = 1\n\n# ## Load a (frozen) Tensorflow model into memory.\n\n# In[7]:\ndetection_graph = tf.Graph()\nwith detection_graph.as_default():\nod_graph_def = tf.GraphDef()\n  with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:\n    serialized_graph = fid.read()\n    od_graph_def.ParseFromString(serialized_graph)\n    tf.import_graph_def(od_graph_def, name='')\n\n# ## Loading label map\n# Label maps map indices to category names, so that when our \nconvolution network predicts `5`, we know that this corresponds to \n`airplane`.  Here we use internal utility functions, but anything that \nreturns a dictionary mapping integers to appropriate string labels \nwould be fine\n\n# In[8]:\n\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\ncategories = label_map_util.convert_label_map_to_categories(label_map, \nmax_num_classes=NUM_CLASSES, use_display_name=True)\ncategory_index = label_map_util.create_category_index(categories)\n\n# ## Helper code\n\n# In[9]:\n\ndef load_image_into_numpy_array(image):\n  (im_width, im_height) = image.size\n  return np.array(image.getdata()).reshape(\n      (im_height, im_width, 3)).astype(np.uint8)\n\n# # Detection\n\ndef run_inference_for_single_image(image, graph):\n  with graph.as_default():\n    with tf.Session() as sess:\n      # Get handles to input and output tensors\n      ops = tf.get_default_graph().get_operations()\n      all_tensor_names = {output.name for op in ops for output in \n op.outputs}\n      tensor_dict = {}\n      for key in [\n          'num_detections', 'detection_boxes', 'detection_scores',\n          'detection_classes', 'detection_masks'\n      ]:\n        tensor_name = key + ':0'\n        if tensor_name in all_tensor_names:\n          tensor_dict[key] = \ntf.get_default_graph().get_tensor_by_name(\n              tensor_name)\n      if 'detection_masks' in tensor_dict:\n        # The following processing is only for single image\n        detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], \n[0])\n        detection_masks = tf.squeeze(tensor_dict['detection_masks'], \n[0])\n        # Reframe is required to translate mask from box coordinates \nto image coordinates and fit the image size.\n        real_num_detection = tf.cast(tensor_dict['num_detections'][0], \ntf.int32)\n        detection_boxes = tf.slice(detection_boxes, [0, 0], \n[real_num_detection, -1])\n        detection_masks = tf.slice(detection_masks, [0, 0, 0], \n[real_num_detection, -1, -1])\n        detection_masks_reframed = \nutils_ops.reframe_box_masks_to_image_masks(\n            detection_masks, detection_boxes, image.shape[0], \nimage.shape[1])\n        detection_masks_reframed = tf.cast(\n            tf.greater(detection_masks_reframed, 0.5), tf.uint8)\n        # Follow the convention by adding back the batch dimension\n        tensor_dict['detection_masks'] = tf.expand_dims(\n            detection_masks_reframed, 0)\n      image_tensor = \ntf.get_default_graph().get_tensor_by_name('image_tensor:0')\n\n      # Run inference\n      output_dict = sess.run(tensor_dict,\n                             feed_dict={image_tensor: \nnp.expand_dims(image, 0)})\n\n      # all outputs are float32 numpy arrays, so convert types as \nappropriate\n      output_dict['num_detections'] = \nint(output_dict['num_detections'][0])\n      output_dict['detection_classes'] = output_dict[\n          'detection_classes'][0].astype(np.uint8)\n      output_dict['detection_boxes'] = output_dict['detection_boxes'] \n[0]\n      output_dict['detection_scores'] = \noutput_dict['detection_scores'][0]\n      if 'detection_masks' in output_dict:\n        output_dict['detection_masks'] = \noutput_dict['detection_masks'][0]\n  return output_dict\n\nwith detection_graph.as_default():\n  with tf.Session(graph=detection_graph) as sess:\n    while True:\n      ret, image_np = cap.read()\n\n     # Expand dimensions since the model expects images to have shape: \n[1, None, None, 3]\n      image_np_expanded = np.expand_dims(image_np, axis=0)\n      image_tensor = \ndetection_graph.get_tensor_by_name('image_tensor:0')\n      # Each box represents a part of the image where a particular \nobject was detected.\n      boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n      # Each score represent how level of confidence for each of the \nobjects.\n      # Score is shown on the result image, together with the class \nlabel.\n      scores = \ndetection_graph.get_tensor_by_name('detection_scores:0')\n      classes = \ndetection_graph.get_tensor_by_name('detection_classes:0')\n      num_detections = \ndetection_graph.get_tensor_by_name('num_detections:0')\n      # Actual detection.\n      (boxes, scores, classes, num_detections) = sess.run(\n          [boxes, scores, classes, num_detections],\n          feed_dict={image_tensor: image_np_expanded})\n      # Visualization of the results of a detection.\n      vis_util.visualize_boxes_and_labels_on_image_array(\n          image_np,\n          np.squeeze(boxes),\n          np.squeeze(classes).astype(np.int32),\n          np.squeeze(scores),\n          category_index,\n          use_normalized_coordinates=True,\n          line_thickness=8)\n\n      area1 = cv2.rectangle(image_np,(201,267),(355,476), \n   (0,255,0),2)\n      area2 = cv2.rectangle(image_np,(354,271),(562,454), \n   (255,0,0),2)\n      cv2.imshow(\"object detection\", image_np)\n\n      if 'detection_boxes:0' == 1 in area1[(201,267),(353,468)]:\n        print(\"area1 occupied!\")\n      else:\n        print(\"area1 free!\")\n\n      if 'detection_boxes:1' == 1 in area2[(354,271),(562,454)]:\n        print(\"area2 occupied!\")\n      else:\n        print(\"area2 free!\")\n\n      if cv2.waitKey(1) &amp; 0xFF == ord('q'):\n        cv2.destroyAllWindows()\n        cap.release()\n        break\n<\/code><\/pre>\n\n<p>I find it difficult to find a solution. Please help.<\/p>\n\n<p>Technical information:<\/p>\n\n<p>Tensorflow 1.10<\/p>\n\n<p>OS - Ubuntu 18.04<\/p>\n\n<p>Python 3.6<\/p>\n\n<p>OpenCV 3.4.2<\/p>\n\n<p>Thanks :)<\/p>\n","answers":[{"AnswerId":"53674062","CreationDate":"2018-12-07T17:11:15.777","ParentId":null,"OwnerUserId":"6663675","Title":null,"Body":"<p>You can use Intersection over Union for this. If the car is in your desired marked rectangle. IOU will have some value otherwise it will be zero.<\/p>\n\n<p>When car rectangle is exactly in your marked rectangle, it will be close to 1 and that will be your solution<\/p>\n"}]}
+{"QuestionId":53259647,"AnswerCount":1,"Tags":"<python><tensorflow><keras><amazon-sagemaker><amazon-machine-learning>","CreationDate":"2018-11-12T09:55:25.347","AcceptedAnswerId":53285949.0,"OwnerUserId":312605.0,"Title":"Issues in upgrading ML libraries using AWS Sagemaker Notebook's Lifecycle configurations","Body":"<p>I use the following script to automate upgrading of my libraries.<\/p>\n\n<p><strong>My script (Start Notebook):<\/strong><\/p>\n\n<pre><code>#!\/bin\/bash\n\nset -e\n\necho 'Before:'\necho $PATH\n\nexport PATH=\/home\/ec2-user\/anaconda3\/envs\/JupyterSystemEnv\/bin:\/home\/ec2-user\/anaconda3\/bin\/:\/usr\/libexec\/gcc\/x86_64-amazon-linux\/4.8.5:\/usr\/local\/cuda\/bin:\/usr\/local\/bin:\/opt\/aws\/bin:\/usr\/local\/mpi\/bin:\/usr\/local\/bin:\/bin:\/usr\/bin:\/usr\/local\/sbin:\/usr\/sbin:\/sbin:\/opt\/aws\/bin:$PATH\n\necho 'After:'\necho $PATH\n\necho `pwd`\n\nsource \/home\/ec2-user\/anaconda3\/bin\/activate tensorflow_p36\n\npip install pandas --upgrade\n\npip install tensorflow-gpu --upgrade\n<\/code><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<p>I get the following error, how can I point to the correct location(\/home\/ec2-user) of keras instead of <strong><em>\/root<\/em><\/strong><\/p>\n\n<pre><code>cp: cannot stat \u2018\/root\/.keras\/keras_tensorflow.json\u2019: No such file or directory \n<\/code><\/pre>\n\n<p><strong>Full Logs:<\/strong><\/p>\n\n<pre><code>Before:\n\/sbin:\/bin:\/usr\/sbin:\/usr\/bin\nAfter:\n\/home\/ec2-user\/anaconda3\/envs\/JupyterSystemEnv\/bin:\/home\/ec2-user\/anaconda3\/bin\/:\/usr\/libexec\/gcc\/x86_64-amazon-linux\/4.8.5:\/usr\/local\/cuda\/bin:\/usr\/local\/bin:\/opt\/aws\/bin:\/usr\/local\/mpi\/bin:\/usr\/local\/bin:\/bin:\/usr\/bin:\/usr\/local\/sbin:\/usr\/sbin:\/sbin:\/opt\/aws\/bin:\/sbin:\/bin:\/usr\/sbin:\/usr\/bin\n\/home\/ec2-user\ncp: cannot stat \u2018\/root\/.keras\/keras_tensorflow.json\u2019: No such file or directory\n<\/code><\/pre>\n\n<p><strong>Without lifecycle configuration:<\/strong><\/p>\n\n<p>All the commands in the above script works.<\/p>\n\n<p>Actual keras.json file is exising, under \/home\/ec2-user when I remove the lifecycle configuration with the following value.<\/p>\n\n<pre><code>sh-4.2$ cat .keras\/keras.json\n{\n    \"backend\": \"tensorflow\"\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"53285949","CreationDate":"2018-11-13T16:55:19.243","ParentId":null,"OwnerUserId":"4249707","Title":null,"Body":"<p>Looks like <code>pip install<\/code> in your case executed \"outside\" of virtualenv  <\/p>\n\n<p>try to change from:  <\/p>\n\n<p><code>source \/home\/ec2-user\/anaconda3\/bin\/activate tensorflow_p36<\/code>  <\/p>\n\n<p>to: <\/p>\n\n<p><code>source \/home\/ec2-user\/anaconda3\/bin\/activate tensorflow_p36 &amp;&amp; pip install pandas tensorflow-gpu --upgrade<\/code>  <\/p>\n\n<p>and delete redundant lines<\/p>\n"}]}
+{"QuestionId":53259746,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-12T10:00:05.707","AcceptedAnswerId":53261479.0,"OwnerUserId":10027028.0,"Title":"module 'tensorflow' has no attribute 'run'","Body":"<p>Code:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\n\na3dim = np.array([[[1,2],[3,4]],\n                  [[5,6],[7,8]]\n                 ])\n\nprint(\"a3dim Shape: \", a3dim.shape)\n\ntf_t=tf.convert_to_tensor(a3dim,dtype=tf.float64)\n\nprint('tf_t : ',tf_t)\nprint('tf_t[0][0][0] : ',tf_t[0][0][0])\nprint('tf_t[1][1][1] : ',tf_t[1][1][1])\n\nprint('run(tf_t) : \\n', tf.run(tf_t))\n<\/code><\/pre>\n\n<p>When I run this program, I have the following error:<\/p>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-9-3506c45f6784&gt; in &lt;module&gt;()\n     15 print('tf_t[1][1][1] : ',tf_t[1][1][1])\n     16 \n---&gt; 17 print('run(tf_t) : \\n', tf.run(tf_t))\n\nAttributeError: module 'tensorflow' has no attribute 'run'\n<\/code><\/pre>\n\n<p>How do I solve this tensorflow issue?\nIs it a version problem?<\/p>\n","answers":[{"AnswerId":"53260425","CreationDate":"2018-11-12T10:40:27.337","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>you have to first create a session to run  <code>tf_t<\/code> then something like <code>session.run(tf_t)<\/code> would work. <\/p>\n"},{"AnswerId":"53261479","CreationDate":"2018-11-12T11:45:22.493","ParentId":null,"OwnerUserId":"10535783","Title":null,"Body":"<pre><code>import numpy as np\nimport tensorflow as tf\n\n\na3dim = np.array([[[1,2],[3,4]],\n                  [[5,6],[7,8]]\n                 ])\n\nprint(\"a3dim Shape: \", a3dim.shape)\n\ntf_t=tf.convert_to_tensor(a3dim,dtype=tf.float64)\n\nprint('tf_t : ',tf_t)\nprint('tf_t[0][0][0] : ',tf_t[0][0][0])\nprint('tf_t[1][1][1] : ',tf_t[1][1][1])\nsess=tf.Session()#create session\nprint('run(tf_t) : \\n', sess.run(tf_t))\nsess.close()#close session\n<\/code><\/pre>\n\n<p>Tensorflow needs graph and Session to compute. The first step of the startup graph is to create a Session object. If there are no creation parameters, the Session builder will start the default graph. Session manages all resources of the TensorFlow program runtime. The Session needs to be closed after all calculations are completed to help the system recycle resources, otherwise the problem of resource leakage may occur.<\/p>\n"}]}
+{"QuestionId":53260221,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-11-12T10:27:51.100","AcceptedAnswerId":null,"OwnerUserId":5177587.0,"Title":"tensorflow data api with keras (passing tensors to keras model)","Body":"<p>I am trying to train a pretrained keras model on new data. I came across tensorflow's dataset api and I am trying to use it with my old keras model. I understand that tf data api returns tensors, so the data api as well as model should be part of the same graph and the output of the data api should be connected as input to the model. Here is the code<\/p>\n\n<pre><code>import tensorflow as tf   \nfrom data_pipeline import ImageDataGenerator\nimport os\nimport keras\nfrom keras.engine import InputLayer\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n###################### to check visible devices ###############\nfrom tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n###############################################################\n\n_EPOCHS      = 10\n_NUM_CLASSES = 2\n_BATCH_SIZE  = 32\n\n\ndef training_pipeline():\n  # #############\n  # Load Dataset\n  # #############\n  training_set = ImageDataGenerator(directory=\"\\\\\\\\in-pdc-sem2\\\\training\",\n                                  horizontal_flip=True, vertical_flip=True, rescale=True, normalize=True,\n                                  color_jitter=True, batch_size=_BATCH_SIZE,\n                                  num_cpus=8, epochs=60, output_patch_size=389, validation=False).dataset_pipeline()\n  testing_set = ImageDataGenerator(directory=\"\\\\\\\\in-pdc-sem2\\\\training\",\n                                  horizontal_flip=False, vertical_flip=False, rescale=False, normalize=True,\n                                  color_jitter=False, batch_size=_BATCH_SIZE,\n                                  num_cpus=8, epochs=60, output_patch_size=389, validation=True).dataset_pipeline()\n\n  print(training_set.output_types, training_set.output_shapes)\n\n  iterator = tf.data.Iterator.from_structure(training_set.output_types, training_set.output_shapes)#((None, 389, 389, 3), (None)))\n\n  train_initializer = iterator.make_initializer(training_set)\n  validation_initializer = iterator.make_initializer(testing_set)\n\n  img, labels = iterator.get_next()\n  img = img.set_shape((None, 389, 389, 3))\n\n  model = baseline_model(img, labels)  # keras model defined here\n  model.summary()\n\n  keras.backend.get_session().run(tf.global_variables_initializer())\n  for epoch in range(_EPOCHS):\n\n      # #############\n      # Train Model\n      # #############\n      keras.backend.get_session().run(train_initializer)\n      model.fit(\n          steps_per_epoch=1000000 \/\/ _BATCH_SIZE,\n          epochs=1,\n          # validation_steps=11970 \/\/ _BATCH_SIZE,\n          callbacks=callbacks(),\n          verbose = 1)\n\n      keras.backend.get_session().run(validation_initializer)\n\n      loss, acc, cross_entropy = model.evaluate(verbose=1, steps=11970 \/\/ 32)\n      filepath = \".\/weights\/ResNet_16_Best\/weights-improvement-Run1-\" + str(epoch) + \"-\" + str(loss) + \".hdf5\"\n      model.save_weights(filepath, overwrite=True)\n\n\ndef baseline_model(input_tensor, labels):\n    jsonFile = '\\\\\\\\in-pdc-sem2\\\\resnetV4_2Best.json'\n    weightsFile = '\\\\\\\\in-pdc-sem1\\\\resnetV4_2BestWeightsOnly.hdf5'\n    with open(jsonFile, \"r\") as file:\n        jsonDef = file.read()\n    from keras.models import model_from_json\n    model_single = model_from_json(jsonDef)\n\n    model_single.load_weights(weightsFile)\n    model_single.layers[0] = InputLayer(input_tensor=input_tensor, input_shape=(389, 389, 3))\n    model_single.compile(target_tensors=[labels], loss='categorical_crossentropy', optimizer='Adam', metrics=[keras.metrics.categorical_accuracy])\n    return model_single\n\ndef callbacks():\n    tensorboard = keras.callbacks.TensorBoard(log_dir='.\/tensorboard', write_grads=False, write_images=False, histogram_freq=0)\n    callbacks_list = [tensorboard]\n    return callbacks_list\n\nif __name__ == '__main__':\n    training_pipeline()\n<\/code><\/pre>\n\n<p>The \"training set\" returns image and label tuple, image is a tensor of shape (32, 389, 389, 3), its a batch of 32 images. I verified the shape in a separate script, it is correct. I am defining the input layer of the model using the tensor, and target tensors in the model.compile part.<\/p>\n\n<p>This is what the model.summary output looks like:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, 389, 389, 3)  0                                            \n__________________________________________________________________________________________________\nconv1 (Conv2D)                  (None, 383, 383, 13) 1924        input_1[0][0]                    \n__________________________________________________________________________________________________\nbn_conv1 (BatchNormalization)   (None, 383, 383, 13) 52          conv1[0][0]                      \n__________________________________________________________________________________________________\nactivation_1 (Activation)       (None, 383, 383, 13) 0           bn_conv1[0][0]                   \n__________________________________________________________________________________________________\nmax_pooling2d_1 (MaxPooling2D)  (None, 191, 191, 13) 0           activation_1[0][0]               \n__________________________________________________________________________________________________\nres2a_branch2a (Conv2D)         (None, 191, 191, 4)  56          max_pooling2d_1[0][0]            \n__________________________________________________________________________________________________\nbn2a_branch2a (BatchNormalizati (None, 191, 191, 4)  16          res2a_branch2a[0][0]             \n__________________________________________________________________________________________________\nactivation_2 (Activation)       (None, 191, 191, 4)  0           bn2a_branch2a[0][0]              \n__________________________________________________________________________________________________\nres2a_branch2b (Conv2D)         (None, 191, 191, 4)  148         activation_2[0][0]               \n__________________________________________________________________________________________________\nbn2a_branch2b (BatchNormalizati (None, 191, 191, 4)  16          res2a_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_3 (Activation)       (None, 191, 191, 4)  0           bn2a_branch2b[0][0]              \n__________________________________________________________________________________________________\nres2a_branch2c (Conv2D)         (None, 191, 191, 8)  40          activation_3[0][0]               \n__________________________________________________________________________________________________\nres2a_branch1 (Conv2D)          (None, 191, 191, 8)  112         max_pooling2d_1[0][0]            \n__________________________________________________________________________________________________\nbn2a_branch2c (BatchNormalizati (None, 191, 191, 8)  32          res2a_branch2c[0][0]             \n__________________________________________________________________________________________________\nbn2a_branch1 (BatchNormalizatio (None, 191, 191, 8)  32          res2a_branch1[0][0]              \n__________________________________________________________________________________________________\nadd_1 (Add)                     (None, 191, 191, 8)  0           bn2a_branch2c[0][0]              \n                                                                 bn2a_branch1[0][0]               \n__________________________________________________________________________________________________\nactivation_4 (Activation)       (None, 191, 191, 8)  0           add_1[0][0]                      \n__________________________________________________________________________________________________\nbn2b_branch2a (BatchNormalizati (None, 191, 191, 8)  32          activation_4[0][0]               \n__________________________________________________________________________________________________\nactivation_5 (Activation)       (None, 191, 191, 8)  0           bn2b_branch2a[0][0]              \n__________________________________________________________________________________________________\nres2b_branch2b (Conv2D)         (None, 191, 191, 4)  292         activation_5[0][0]               \n__________________________________________________________________________________________________\nbn2b_branch2b (BatchNormalizati (None, 191, 191, 4)  16          res2b_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_6 (Activation)       (None, 191, 191, 4)  0           bn2b_branch2b[0][0]              \n__________________________________________________________________________________________________\nres2b_branch2c (Conv2D)         (None, 191, 191, 8)  40          activation_6[0][0]               \n__________________________________________________________________________________________________\nadd_2 (Add)                     (None, 191, 191, 8)  0           res2b_branch2c[0][0]             \n                                                                 activation_4[0][0]               \n__________________________________________________________________________________________________\nbn2c_branch2a (BatchNormalizati (None, 191, 191, 8)  32          add_2[0][0]                      \n__________________________________________________________________________________________________\nactivation_7 (Activation)       (None, 191, 191, 8)  0           bn2c_branch2a[0][0]              \n__________________________________________________________________________________________________\nres2c_branch2b (Conv2D)         (None, 191, 191, 4)  292         activation_7[0][0]               \n__________________________________________________________________________________________________\nbn2c_branch2b (BatchNormalizati (None, 191, 191, 4)  16          res2c_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_8 (Activation)       (None, 191, 191, 4)  0           bn2c_branch2b[0][0]              \n__________________________________________________________________________________________________\nres2c_branch2c (Conv2D)         (None, 191, 191, 8)  40          activation_8[0][0]               \n__________________________________________________________________________________________________\nadd_3 (Add)                     (None, 191, 191, 8)  0           res2c_branch2c[0][0]             \n                                                                 add_2[0][0]                      \n__________________________________________________________________________________________________\nres3a_branch2a (Conv2D)         (None, 96, 96, 8)    72          add_3[0][0]                      \n__________________________________________________________________________________________________\nbn3a_branch2a (BatchNormalizati (None, 96, 96, 8)    32          res3a_branch2a[0][0]             \n__________________________________________________________________________________________________\nactivation_9 (Activation)       (None, 96, 96, 8)    0           bn3a_branch2a[0][0]              \n__________________________________________________________________________________________________\nres3a_branch2b (Conv2D)         (None, 96, 96, 8)    584         activation_9[0][0]               \n__________________________________________________________________________________________________\nbn3a_branch2b (BatchNormalizati (None, 96, 96, 8)    32          res3a_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_10 (Activation)      (None, 96, 96, 8)    0           bn3a_branch2b[0][0]              \n__________________________________________________________________________________________________\nres3a_branch2c (Conv2D)         (None, 96, 96, 16)   144         activation_10[0][0]              \n__________________________________________________________________________________________________\nres3a_branch1 (Conv2D)          (None, 96, 96, 16)   144         add_3[0][0]                      \n__________________________________________________________________________________________________\nbn3a_branch2c (BatchNormalizati (None, 96, 96, 16)   64          res3a_branch2c[0][0]             \n__________________________________________________________________________________________________\nbn3a_branch1 (BatchNormalizatio (None, 96, 96, 16)   64          res3a_branch1[0][0]              \n__________________________________________________________________________________________________\nadd_4 (Add)                     (None, 96, 96, 16)   0           bn3a_branch2c[0][0]              \n                                                                 bn3a_branch1[0][0]               \n__________________________________________________________________________________________________\nactivation_11 (Activation)      (None, 96, 96, 16)   0           add_4[0][0]                      \n__________________________________________________________________________________________________\nbn3b_branch2a (BatchNormalizati (None, 96, 96, 16)   64          activation_11[0][0]              \n__________________________________________________________________________________________________\nactivation_12 (Activation)      (None, 96, 96, 16)   0           bn3b_branch2a[0][0]              \n__________________________________________________________________________________________________\nres3b_branch2b (Conv2D)         (None, 96, 96, 8)    1160        activation_12[0][0]              \n__________________________________________________________________________________________________\nbn3b_branch2b (BatchNormalizati (None, 96, 96, 8)    32          res3b_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_13 (Activation)      (None, 96, 96, 8)    0           bn3b_branch2b[0][0]              \n__________________________________________________________________________________________________\nres3b_branch2c (Conv2D)         (None, 96, 96, 16)   144         activation_13[0][0]              \n__________________________________________________________________________________________________\nadd_5 (Add)                     (None, 96, 96, 16)   0           res3b_branch2c[0][0]             \n                                                                 activation_11[0][0]              \n__________________________________________________________________________________________________\nres4a_branch2a (Conv2D)         (None, 48, 48, 16)   272         add_5[0][0]                      \n__________________________________________________________________________________________________\nbn4a_branch2a (BatchNormalizati (None, 48, 48, 16)   64          res4a_branch2a[0][0]             \n__________________________________________________________________________________________________\nactivation_14 (Activation)      (None, 48, 48, 16)   0           bn4a_branch2a[0][0]              \n__________________________________________________________________________________________________\nres4a_branch2b (Conv2D)         (None, 48, 48, 16)   2320        activation_14[0][0]              \n__________________________________________________________________________________________________\nbn4a_branch2b (BatchNormalizati (None, 48, 48, 16)   64          res4a_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_15 (Activation)      (None, 48, 48, 16)   0           bn4a_branch2b[0][0]              \n__________________________________________________________________________________________________\nres4a_branch2c (Conv2D)         (None, 48, 48, 64)   1088        activation_15[0][0]              \n__________________________________________________________________________________________________\nres4a_branch1 (Conv2D)          (None, 48, 48, 64)   1088        add_5[0][0]                      \n__________________________________________________________________________________________________\nbn4a_branch2c (BatchNormalizati (None, 48, 48, 64)   256         res4a_branch2c[0][0]             \n__________________________________________________________________________________________________\nbn4a_branch1 (BatchNormalizatio (None, 48, 48, 64)   256         res4a_branch1[0][0]              \n__________________________________________________________________________________________________\nadd_6 (Add)                     (None, 48, 48, 64)   0           bn4a_branch2c[0][0]              \n                                                                 bn4a_branch1[0][0]               \n__________________________________________________________________________________________________\nactivation_16 (Activation)      (None, 48, 48, 64)   0           add_6[0][0]                      \n__________________________________________________________________________________________________\nbn4b_branch2a (BatchNormalizati (None, 48, 48, 64)   256         activation_16[0][0]              \n__________________________________________________________________________________________________\nactivation_17 (Activation)      (None, 48, 48, 64)   0           bn4b_branch2a[0][0]              \n__________________________________________________________________________________________________\nres4b_branch2b (Conv2D)         (None, 48, 48, 16)   9232        activation_17[0][0]              \n__________________________________________________________________________________________________\nbn4b_branch2b (BatchNormalizati (None, 48, 48, 16)   64          res4b_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_18 (Activation)      (None, 48, 48, 16)   0           bn4b_branch2b[0][0]              \n__________________________________________________________________________________________________\nres4b_branch2c (Conv2D)         (None, 48, 48, 64)   1088        activation_18[0][0]              \n__________________________________________________________________________________________________\nadd_7 (Add)                     (None, 48, 48, 64)   0           res4b_branch2c[0][0]             \n                                                                 activation_16[0][0]              \n__________________________________________________________________________________________________\nres5a_branch2a (Conv2D)         (None, 24, 24, 32)   2080        add_7[0][0]                      \n__________________________________________________________________________________________________\nbn5a_branch2a (BatchNormalizati (None, 24, 24, 32)   128         res5a_branch2a[0][0]             \n__________________________________________________________________________________________________\nactivation_19 (Activation)      (None, 24, 24, 32)   0           bn5a_branch2a[0][0]              \n__________________________________________________________________________________________________\nres5a_branch2b (Conv2D)         (None, 24, 24, 32)   9248        activation_19[0][0]              \n__________________________________________________________________________________________________\nbn5a_branch2b (BatchNormalizati (None, 24, 24, 32)   128         res5a_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_20 (Activation)      (None, 24, 24, 32)   0           bn5a_branch2b[0][0]              \n__________________________________________________________________________________________________\nres5a_branch2c (Conv2D)         (None, 24, 24, 128)  4224        activation_20[0][0]              \n__________________________________________________________________________________________________\nres5a_branch1 (Conv2D)          (None, 24, 24, 128)  8320        add_7[0][0]                      \n__________________________________________________________________________________________________\nbn5a_branch2c (BatchNormalizati (None, 24, 24, 128)  512         res5a_branch2c[0][0]             \n__________________________________________________________________________________________________\nbn5a_branch1 (BatchNormalizatio (None, 24, 24, 128)  512         res5a_branch1[0][0]              \n__________________________________________________________________________________________________\nadd_8 (Add)                     (None, 24, 24, 128)  0           bn5a_branch2c[0][0]              \n                                                                 bn5a_branch1[0][0]               \n__________________________________________________________________________________________________\nactivation_21 (Activation)      (None, 24, 24, 128)  0           add_8[0][0]                      \n__________________________________________________________________________________________________\nres6a_branch2a (Conv2D)         (None, 12, 12, 64)   8256        activation_21[0][0]              \n__________________________________________________________________________________________________\nbn6a_branch2a (BatchNormalizati (None, 12, 12, 64)   256         res6a_branch2a[0][0]             \n__________________________________________________________________________________________________\nactivation_22 (Activation)      (None, 12, 12, 64)   0           bn6a_branch2a[0][0]              \n__________________________________________________________________________________________________\nres6a_branch2b (Conv2D)         (None, 12, 12, 64)   36928       activation_22[0][0]              \n__________________________________________________________________________________________________\nbn6a_branch2b (BatchNormalizati (None, 12, 12, 64)   256         res6a_branch2b[0][0]             \n__________________________________________________________________________________________________\nactivation_23 (Activation)      (None, 12, 12, 64)   0           bn6a_branch2b[0][0]              \n__________________________________________________________________________________________________\nres6a_branch2c (Conv2D)         (None, 12, 12, 512)  33280       activation_23[0][0]              \n__________________________________________________________________________________________________\nres6a_branch1 (Conv2D)          (None, 12, 12, 512)  66048       activation_21[0][0]              \n__________________________________________________________________________________________________\nbn6a_branch2c (BatchNormalizati (None, 12, 12, 512)  2048        res6a_branch2c[0][0]             \n__________________________________________________________________________________________________\nbn6a_branch1 (BatchNormalizatio (None, 12, 12, 512)  2048        res6a_branch1[0][0]              \n__________________________________________________________________________________________________\nadd_9 (Add)                     (None, 12, 12, 512)  0           bn6a_branch2c[0][0]              \n                                                                 bn6a_branch1[0][0]               \n__________________________________________________________________________________________________\nactivation_24 (Activation)      (None, 12, 12, 512)  0           add_9[0][0]                      \n__________________________________________________________________________________________________\navg_pool (GlobalAveragePooling2 (None, 512)          0           activation_24[0][0]              \n__________________________________________________________________________________________________\ndropout_1 (Dropout)             (None, 512)          0           avg_pool[0][0]                   \n__________________________________________________________________________________________________\nFC1 (Dense)                     (None, 1)            513         dropout_1[0][0]                  \n__________________________________________________________________________________________________\nactivation_25 (Activation)      (None, 1)            0           FC1[0][0]                        \n==================================================================================================\nTotal params: 196,557\nTrainable params: 192,867\nNon-trainable params: 3,690\n<\/code><\/pre>\n\n<p>Everything looks correct. However When I run the code, I get the following error:<\/p>\n\n<pre><code>Epoch 1\/1\nTraceback (most recent call last):\n  File \"C:\/Users\/ASista162282\/Desktop\/code\/camleyon_17\/train.py\", line 114, in &lt;module&gt;\n    training_pipeline()\n  File \"C:\/Users\/ASista162282\/Desktop\/code\/camleyon_17\/train.py\", line 71, in training_pipeline\n    verbose = 1)\n  File \"C:\\ProgramData\\Miniconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 1705, in fit\n    validation_steps=validation_steps)\n  File \"C:\\ProgramData\\Miniconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 1188, in _fit_loop\n    outs = f(ins)\n  File \"C:\\ProgramData\\Miniconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 2478, in __call__\n    **self.session_kwargs)\n  File \"C:\\ProgramData\\Miniconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 900, in run\n    run_metadata_ptr)\n  File \"C:\\ProgramData\\Miniconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1111, in _run\n    str(subfeed_t.get_shape())))\nValueError: Cannot feed value of shape () for Tensor 'input_1:0', which has shape '(?, 389, 389, 3)'\n<\/code><\/pre>\n\n<p>It doesn't make any sense. I even added the set_shape function before defining the model, and it still shows empty shape. Any help will be really appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"53263038","CreationDate":"2018-11-12T13:18:32.703","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>The way you are replacing the input layer doesn't seem to connect the new layer correctly. Try replacing this:<\/p>\n\n<p><code>model_single.layers[0] = InputLayer(input_tensor=input_tensor, input_shape=(389, 389, 3))<\/code><\/p>\n\n<p>with this:<\/p>\n\n<p><code>from keras.models import Model\nmodel_single.layers.pop(0)\nnew_input = InputLayer(input_tensor=input_tensor, input_shape=(389, 389, 3))\nnew_output = model_single(new_input)\nmodel_single = Model(new_input, new_output)<\/code><\/p>\n"}]}
+{"QuestionId":53260459,"AnswerCount":2,"Tags":"<python><tensorflow><ffmpeg>","CreationDate":"2018-11-12T10:42:40.500","AcceptedAnswerId":null,"OwnerUserId":8853991.0,"Title":"tf.contrib.ffmpeg.decode_audio replacement?","Body":"<p>In the tensorflow docs, it stated that tf.contrib.ffmpeg.decode_audio is depreciated, however it does not state what it's replacement is...<\/p>\n\n<blockquote>\n  <p>THIS FUNCTION IS DEPRECATED. It will be removed after 2018-09-04. Instructions for updating: This will be deleted and should not be used.<\/p>\n<\/blockquote>\n\n<p>Here is the link: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/ffmpeg\/decode_audio\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/ffmpeg\/decode_audio<\/a><\/p>\n","answers":[{"AnswerId":"56210900","CreationDate":"2019-05-19T18:40:30.797","ParentId":null,"OwnerUserId":"8321660","Title":null,"Body":"<p>Try \n<code>tensorflow.contrib.framework.python.ops.audio_ops.decode_wav<\/code>.<\/p>\n\n<p>This might be gone in tensorflow 2<\/p>\n"},{"AnswerId":"54899747","CreationDate":"2019-02-27T06:58:38.030","ParentId":null,"OwnerUserId":"479153","Title":null,"Body":"<p>two replacements for <code>tf.contrib.ffmpeg.decode_audio<\/code> are available for  tensorflow 2.0:<\/p>\n\n<ul>\n<li>Tensorflow 2 will have a <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/audio\/decode_wav\" rel=\"nofollow noreferrer\">builtin wav decoder<\/a>  <\/li>\n<li>The ffmpeg ops from <code>contrib<\/code> move to a new external <a href=\"https:\/\/github.com\/tensorflow\/io\" rel=\"nofollow noreferrer\">I\/O project<\/a>. However the audio ops are <a href=\"https:\/\/github.com\/tensorflow\/io\/issues\/49\" rel=\"nofollow noreferrer\">not implemented yet<\/a>.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53260857,"AnswerCount":1,"Tags":"<performance><numpy><tensorflow><keras><generator>","CreationDate":"2018-11-12T11:07:09.960","AcceptedAnswerId":null,"OwnerUserId":10639500.0,"Title":"Slow numpy array indexing for keras time series generator","Body":"<p>I use the keras time series generator for training a neural network with LSTM cells, which unfortunately proved to be a bottleneck in training. <\/p>\n\n<p>Below is a simplified example to run, which shows the high runtime of the batch generator. It is important to note that the rows from the dataset are chosen randomly and thus a sliding window is not possible. During the training the CPUs are running continuously at about 80%, whereas the GPU is running at a single-digit percentage rate.<\/p>\n\n<pre><code>def get_time_series(data, index, look_back, batch_size):\n    samples1 = np.empty((batch_size, look_back, np.size(data, axis=1)))\n    rows = np.random.randint(look_back, np.size(data, axis=1), size=batch_size)\n    for j, row in enumerate(rows):\n        indices = range(rows[j] - look_back, rows[j], 1)\n        samples1[j] = data[indices]\n    return samples1\n\n\ndata = np.random.rand(100000, 20)\nstart = time.time()\nbatch = get_time_series(data, index=50, look_back=1000, batch_size=2**12)\nprint(\"Batch generator needs\",  time.time()-start,  \"seconds\")\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>Batch generator needs 0.6224319934844971 seconds\n<\/code><\/pre>\n\n<p>I already tried to build the 3-d array first, so I only have to index the array-rows in the *get_time_series-*Function. This was about 60 times faster during the training, but leads to an \"out of memory error\" with large datasets.<\/p>\n\n<p>Does anyone have ideas on how to improve the performance of this bottleneck? Work with pointer, faster indexing methods, ...<\/p>\n\n<p>Thanks,\nMax<\/p>\n","answers":[{"AnswerId":"53261800","CreationDate":"2018-11-12T12:03:43.093","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>EDIT 2:<\/p>\n\n<p>Not sure if this is going to be any faster, but you can also just do something like this. It still relies on advanced indexing, although over contiguous data, so maybe it's a bit better?:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\n\ndef get_time_series(data, indices, look_back):\n    # Make sure indices are big enough\n    indices = indices[indices &gt;= look_back]\n    # Make indexing matrix\n    idx = indices[:, np.newaxis] + np.arange(-look_back, 0)\n    # Make batch\n    return data[idx]\n<\/code><\/pre>\n\n<p>You would use it for example like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\n\ndef get_time_series(data, indices, look_back):\n    indices = indices[indices &gt;= look_back]\n    idx = indices[:, np.newaxis] + np.arange(-look_back, 0)\n    return data[idx]\n\ndef make_batches(data, look_back, batch_size):\n    indices = np.random.permutation(np.arange(look_back, len(data) + 1))\n    for i in range(0, len(indices), batch_size):\n        yield get_time_series(data, indices[i:i + batch_size], look_back)\n\ndata = ...\nlook_back = ...\nbatch_size = ...\nfor batch in make_batches(data, look_back, batch_size):\n    # Use batch\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>If you want to shuffle the examples, you could first make the sliding window for the whole dataset (which should not take any memory or time) and then take batches from a shuffled index:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Make sliding window with the previous function\ndata_sw = get_time_series(data, 0, look_back, len(data))\n# Random index\nbatch_idx = np.random.permutation(len(data_sw))\n# To get the first batch\nbatch = data_sw[batch_idx[:batch_size]]\n<\/code><\/pre>\n\n<hr>\n\n<p>I think this does what you want, and should be quite faster than using loops:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\n\ndef get_time_series(data, index, look_back, batch_size):\n    from numpy.lib.stride_tricks import as_strided\n    # Index should be at least as big as look_back to have enough elements before it\n    index = max(index, look_back)\n    # Batch size should not go beyond the array\n    batch_size = min(batch_size, len(data) - index + 1)\n    # Relevant slice for the batch\n    data_slice = data[index - look_back:index + batch_size]\n    # Reshape with stride tricks as a \"sliding window\"\n    data_strides = data_slice.strides\n    batch_shape = (batch_size, look_back, data_slice.shape[-1])\n    batch_strides = (data_strides[0], data_strides[0], data_strides[1])\n    return as_strided(data_slice, batch_shape, batch_strides, writeable=False)\n\n# Test\ndata = np.arange(300).reshape((100, 3))\nbatch = get_time_series(data, 20, 5, 4)\nprint(batch)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[45 46 47]\n  [48 49 50]\n  [51 52 53]\n  [54 55 56]\n  [57 58 59]]\n\n [[48 49 50]\n  [51 52 53]\n  [54 55 56]\n  [57 58 59]\n  [60 61 62]]\n\n [[51 52 53]\n  [54 55 56]\n  [57 58 59]\n  [60 61 62]\n  [63 64 65]]\n\n [[54 55 56]\n  [57 58 59]\n  [60 61 62]\n  [63 64 65]\n  [66 67 68]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53261074,"AnswerCount":1,"Tags":"<python><linux><tensorflow><anaconda><inspect>","CreationDate":"2018-11-12T11:19:40.083","AcceptedAnswerId":null,"OwnerUserId":9689358.0,"Title":"python3.5 import tensorflow.contrib error: Function has keyword-only arguments or annotations, use getfullargspec() API which can support them","Body":"<p>I am using python 3.5 and installed tensorflow-gpu 1.12.0 on Linux. When I run this in python3.5:<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>It works fine with no errors.<\/p>\n\n<p>However, when I run:<\/p>\n\n<pre><code>import tensorflow.contrib\n<\/code><\/pre>\n\n<p>There is an error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nFile \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.contrib import distributions\nFile \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/distributions\/__init__.py\", line 38, in &lt;module&gt;\nfrom tensorflow.contrib.distributions.python.ops.estimator import *\n     File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/distributions\/python\/ops\/estimator.py\", line 21, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.estimators.head import _compute_weighted_loss\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/__init__.py\", line 96, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn import *\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn import *\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/__init__.py\", line 30, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn import estimators\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/__init__.py\", line 302, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.estimators.dnn import DNNClassifier\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn.py\", line 35, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.estimators import dnn_linear_combined\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn_linear_combined.py\", line 36, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.estimators import estimator\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 52, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.learn_io import data_feeder\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/learn_io\/__init__.py\", line 26, in &lt;module&gt;\n    from tensorflow.contrib.learn.python.learn.learn_io.dask_io import extract_dask_data\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/learn_io\/dask_io.py\", line 33, in &lt;module&gt;\n    import dask.dataframe as dd\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/dask\/dataframe\/__init__.py\", line 1, in &lt;module&gt;\n    from .core import (DataFrame, Series, Index, _Frame, map_partitions,\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/dask\/dataframe\/core.py\", line 23, in &lt;module&gt;\n    from .. import array as da\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/dask\/array\/__init__.py\", line 4, in &lt;module&gt;\n    from .core import (Array, stack, concatenate, take, tensordot, transpose,\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/dask\/array\/core.py\", line 13, in &lt;module&gt;\n    from toolz.curried import (pipe, partition, concat, unique, pluck, join, first,\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/toolz\/curried\/__init__.py\", line 53, in &lt;module&gt;\n    _curry_namespace(vars(toolz)),\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/toolz\/curried\/__init__.py\", line 48, in _curry_namespace\n    for name, f in ns.items() if '__' not in name\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/toolz\/curried\/__init__.py\", line 48, in &lt;genexpr&gt;\n    for name, f in ns.items() if '__' not in name\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/toolz\/curried\/__init__.py\", line 42, in _should_curry\n    return (callable(f) and _nargs(f) &gt; 1 or f in do_curry)\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/site-packages\/toolz\/curried\/__init__.py\", line 35, in _nargs\n    return len(inspect.getargspec(f).args)\n  File \"\/home\/myname\/anaconda3\/lib\/python3.5\/inspect.py\", line 1050, in getargspec\n    raise ValueError(\"Function has keyword-only arguments or annotations\"\nValueError: Function has keyword-only arguments or annotations, use getfullargspec() API which can support them\n<\/code><\/pre>\n\n<p>I searched for getfullargspec() API but got even more confused. As far as I know, the getfullargspec() API from inspect can only do with something like <\/p>\n\n<pre><code>getfullargsspec(function_name)\n<\/code><\/pre>\n\n<p>How can I do with it? I could appreciate it a lot if you could explain the cause and also possibly a simple solution.<\/p>\n\n<hr>\n\n<p>Plus, I am running the code on a remote server without root permission. Command lines such as\n<strong>sudo<\/strong> xxxxx\ndo not work for me.<\/p>\n\n<p>Thank you for your help!<\/p>\n","answers":[{"AnswerId":"53630738","CreationDate":"2018-12-05T10:59:54.083","ParentId":null,"OwnerUserId":"10748979","Title":null,"Body":"<p>I also face this problem. I found this way to solve this problem.<\/p>\n\n<pre><code>conda install toolz\n<\/code><\/pre>\n\n<p>Good Luck<\/p>\n"}]}
+{"QuestionId":53261322,"AnswerCount":0,"Tags":"<python><tensorflow><pytorch>","CreationDate":"2018-11-12T11:34:54.877","AcceptedAnswerId":null,"OwnerUserId":9724030.0,"Title":"pytorch runs much slower than tensorflow even for a shallow CNN","Body":"<p>The input data is an EEG data which was converted into PSD. And its shape is <strong>[103600, 59, 51]<\/strong>, where 103600 is the number of samples, i.e. the total samples for an epoch. The data was loaded into memory.<\/p>\n\n<p>And the model I used is very simple, however, the time required <strong>for training an epoch was 50s with pytorch 0.4 and a GTX 1080<\/strong>. However, the time for val set interface was 0.3s.<\/p>\n\n<p>I also tested this model using <strong>tensorflow 1.2.0 the time was 8s<\/strong>. I have no idea if I got it wrong.<\/p>\n\n<p>So I made some test. This img shows the time for runing.\n<a href=\"https:\/\/i.stack.imgur.com\/WCRM8.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>where\n<strong><em>time_to_tensor<\/em><\/strong> is the time for <code>inputs, targets = torch.tensor(inputs, dtype=torch.float), torch.tensor(targets, dtype=torch.int64)<\/code><\/p>\n\n<p><strong><em>time_to_cuda<\/em><\/strong> for <code>inputs, targets = inputs.to(self.device), targets.to(self.device)<\/code><\/p>\n\n<p><strong><em>time_to_model<\/em><\/strong> for <code>out = self.model(inputs)<\/code><\/p>\n\n<p><strong><em>time_criterion for<\/em><\/strong> <code>loss = self.criterion(out, targets)<\/code><\/p>\n\n<p><strong><em>time_backward<\/em><\/strong> for <code>self.optimizer.step()<\/code><\/p>\n\n<p>It seems the operation for tensor.to(self.device) (or tensor.cuda()) wasted much time. So <em>I tried to move the data to GPU at once<\/em>, using <\/p>\n\n<pre><code>self.X_train = torch.tensor(self.X_train, dtype=torch.float).to(self.device)\nself.y_train = torch.tensor(self.y_train, dtype=torch.int64).to(self.device)\n<\/code><\/pre>\n\n<p>and commented<\/p>\n\n<pre><code>inputs, targets = torch.tensor(inputs, dtype=torch.float), torch.tensor(targets, dtype=torch.int64)\ninputs, targets = inputs.to(self.device), targets.to(self.device)\n<\/code><\/pre>\n\n<p>However I got the running time in the following img.<a href=\"https:\/\/i.stack.imgur.com\/EDcRN.png\" rel=\"nofollow noreferrer\">enter image description here<\/a>. The time required for an epoch still was about 50s, and time for each operation changed. I am confused. <\/p>\n\n<p>Could someone help me to find the problem? Thanks.<\/p>\n\n<p>Code was here.<\/p>\n\n<h1>It's the model I used<\/h1>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass TxtCNN4(nn.Module):\n\n    def __init__(self, n_classes=6, time_steps=59, psd_lenght=51, filter_numbers=128):\n        super(TxtCNN4, self).__init__()\n        self.n_classes = n_classes\n        self.time_steps = time_steps\n        self.psd_lenght = psd_lenght\n        self.filters = filter_numbers\n\n        self.conv1 = nn.Conv2d(in_channels=1, out_channels=self.filters\/\/2, \n                                kernel_size=(3, psd_lenght), stride=1, padding=0, bias=False)\n        self.bn1 = nn.BatchNorm2d(self.filters\/\/2, momentum=0.05)\n        self.pool1 = nn.MaxPool2d(kernel_size=(self.time_steps - 3 + 1, 1))\n\n        self.conv2 = nn.Conv2d(in_channels=self.filters\/\/2, out_channels=self.filters,\n                                    kernel_size=(3, 1), stride=1, padding=0, bias=False)\n        self.bn2 = nn.BatchNorm2d(self.filters, momentum=0.05)\n        self.pool2 = nn.MaxPool2d(kernel_size=(self.time_steps - 5 + 1, 1))\n\n        self.conv3 = nn.Conv2d(in_channels=self.filters, out_channels=self.filters,\n                                    kernel_size=(3, 1), stride=1, padding=0, bias=False)\n        self.bn3 = nn.BatchNorm2d(self.filters, momentum=0.05)\n        self.pool3 = nn.MaxPool2d(kernel_size=(self.time_steps - 7 + 1, 1))\n\n        self.fc = nn.Linear(in_features=2*self.filters+self.filters\/\/2, out_features=128)\n        self.cls = nn.Linear(in_features=128, out_features=self.n_classes)\n        self.criterion = nn.CrossEntropyLoss(weight=None)\n\n    def forward(self, inputs):\n        inputs = inputs.view(-1, self.time_steps, self.psd_lenght, 1)\n        inputs = inputs.permute(0, 3, 1, 2)\n        conv1 = F.relu(self.bn1(self.conv1(inputs)), inplace=True)\n        conv2 = F.relu(self.bn2(self.conv2(conv1)), inplace=True)\n        conv3 = F.relu(self.bn3(self.conv3(conv2)), inplace=True)\n\n        x = torch.cat((self.pool1(conv1), self.pool2(conv2), self.pool3(conv3)), dim=1)\n        x = x.view(-1, 2*self.filters+self.filters\/\/2)\n        x = F.dropout(x, p=0.5, training=self.training)\n        x = F.relu(self.fc(x), inplace=True)\n        x = self.cls(x)\n\n        return x\n<\/code><\/pre>\n\n<h1>It's the main python file for traning the model<\/h1>\n\n<pre><code>file_path = '\/mnt\/disk2\/wy\/SLEEP_EDF\/'\n\nbatch_size = 32*4*2\ndropout_rate = 0.5\nnb_classes = 5\nmax_epochs = 100\nearly_stop_epoch = 10\n\nlearning_rate = 1e-3\n\nmodel_type='TxtCNN4'\nfilter_numbers = 128\n\nfs = 100    # Sample frequence\nn_Channels = 2\nn_Samples = 30*100\n\nload_flag = False\nmodel_path = ''\n\ndef iterate_minibatches(inputs, targets, batchsize, shuffle=False):\n    input_len = inputs.shape[0]\n    assert input_len == len(targets)\n\n    if shuffle:\n        indices = np.arange(input_len)  \n        np.random.shuffle(indices)\n    for start_idx in range(0, input_len, batchsize):\n        if shuffle:\n            excerpt = indices[start_idx:start_idx + batchsize]\n        else:\n            excerpt = slice(start_idx, start_idx + batchsize)\n\nclass Trainer():\n\n    def __init__(self, X_inputs, labels, fold, subj_id, log_path, model_type=model_type, args=None):\n        self.subj_id = subj_id\n        self.model_type = model_type\n        self.writer_root_path = os.path.abspath(\n                                os.path.join(\n                                    os.path.curdir,\n                                    log_path+'_'+model_type,\n                                    model_type + '_' + str(subj_id)))\n        self.train_writer = SummaryWriter(os.path.join(self.writer_root_path, 'train'))\n        self.val_writer = SummaryWriter(os.path.join(self.writer_root_path, 'val'))\n        self.test_writer = SummaryWriter(os.path.join(self.writer_root_path, 'test'))\n\n        self.curr_epcoh = 0\n        self.curr_iter = 0\n\n        self.train_end_flag = False\n\n        (self.X_train, self.y_train), (self.X_val, self.y_val), (self.X_test, self.y_test) = reformatInput(X_inputs, labels, fold)\n        print('Test set label and BiLi:\\t', np.unique(self.y_test, return_counts=True))\n        # normalization between all data\n        X_mean = self.X_train.mean()\n        X_std = self.X_train.std()\n        self.X_train = (self.X_train - X_mean)\/X_std\n        self.X_val = (self.X_val - X_mean)\/X_std\n        self.X_test = (self.X_test - X_mean)\/X_std\n\n\n        if model_type == 'TxtCNN4':\n            model = TxtCNN4()\n        else:\n            model = DeepSleepNet()\n        self.device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n        self.model = model.to(self.device)\n        self.criterion = nn.CrossEntropyLoss(weight=None)\n\n        # self.X_train = torch.tensor(self.X_train, dtype=torch.float).to(self.device)\n        # self.y_train = torch.tensor(self.y_train, dtype=torch.int64).to(self.device)\n\n        self.optimizer = optim.Adam(self.model.parameters(), lr=learning_rate)\n\n\n    def train(self, epoch):\n        train_loss = AvgMeter()\n        train_acc = AvgMeter()\n        start_time = time.time()\n\n        time_to_tensor = 0\n        time_to_cuda = 0\n        time_model = 0\n        time_criterion = 0\n        time_backward = 0\n        time_step = 0\n        time_writer = 0\n        self.curr_epcoh = epoch\n        self.model.train()\n        for (inputs, targets) in iterate_minibatches(self.X_train, self.y_train, batch_size, shuffle=False):\n            self.curr_iter += 1\n            time_1 = time.time()\n            inputs, targets = torch.tensor(inputs, dtype=torch.float), torch.tensor(targets, dtype=torch.int64)  # dtype=torch.int64\n            time_to_tensor += (time.time() - time_1)\n            time_1 = time.time()\n            inputs, targets = inputs.to(self.device), targets.to(self.device)\n            time_to_cuda += (time.time() - time_1)\n\n            # zero the parameter gradients\n            self.optimizer.zero_grad()\n\n            # forward + backward + optimize\n            time_1 = time.time()\n            out = self.model(inputs)\n            time_model += (time.time() - time_1)\n\n            time_1 = time.time()\n            loss = self.criterion(out, targets)\n            time_criterion += (time.time() - time_1)\n\n            time_1 = time.time()\n            loss.backward()\n            time_backward += (time.time() - time_1)\n\n            time_1 = time.time()\n            self.optimizer.step()\n            time_step += (time.time() - time_1)\n\n            if self.curr_iter % 20 == 0:\n                time_1 = time.time()\n                pred = out.max(dim=1)[1].cpu().numpy()\n                gt = targets.cpu().numpy()\n                acc = np.mean(pred==gt)\n                _loss = loss.detach().cpu().numpy()\n                train_acc.update(acc)\n                train_loss.update(_loss)\n\n                self.train_writer.add_scalar('learning_rate', curr_lr, self.curr_iter)\n                self.train_writer.add_scalar('acc', acc, self.curr_iter)\n                self.train_writer.add_scalar('loss', _loss, self.curr_iter)\n                time_writer += (time.time() - time_1)\n\n\n        fmt_str = \"Train\\tEpoch [{:d}\/{:d}]  train_Loss: {:.4f}\\ttrain_Acc: {:.2f}\\tTime per Epoch: {:.4f}\"\n        print_str = fmt_str.format(self.curr_epcoh,\n                                    max_epochs,\n                                    train_loss.avg, \n                                    train_acc.avg*100, \n                                    time.time()-start_time)\n        print(print_str)\n        print('time_to_tensor: ',time_to_tensor)\n        print('time_to_cuda: ',time_to_cuda)\n        print('time_model: ',time_model)\n        print('time_criterion: ',time_criterion)\n        print('time_backward: ',time_backward)\n        print('time_step: ',time_step)\n        print('time_writer: ',time_writer)\n        print(len(self.y_train))\n        train_acc.reset()\n        train_loss.reset()\n\n    def validation(self):\n        self.model.eval()\n        with torch.no_grad():\n            # val set\n            loss_gather = AvgMeter()\n            pred_gather = PreGather()\n            tran_time = 0\n            for (inputs, targets) in iterate_minibatches(self.X_val, self.y_val, batch_size, shuffle=False):\n                time_1 = time.time()\n                inputs, targets = torch.tensor(inputs, dtype=torch.float), torch.tensor(targets, dtype=torch.int64)  # dtype=torch.int64\n                inputs, targets = inputs.to(self.device), targets.to(self.device)\n                tran_time += (time.time() - time_1)\n\n                out = self.model(inputs)\n                _loss = self.criterion(out, targets).detach().cpu().numpy()\n                pred = out.max(dim=1)[1].cpu().numpy()\n                loss_gather.update(_loss)\n                pred_gather.update(pred)\n\n\n            val_loss = loss_gather.avg\n            pred = pred_gather.pred\n            val_acc = np.mean(pred==self.y_val)\n            val_kappa = cohen_kappa_score(self.y_val, pred)\n            val_BCA = recall_score(self.y_val, pred, average='macro')\n            loss_gather.reset()\n            pred_gather.reset()\n\n            # Then we print the results for this epoch:\n            fmt_str = \"VAL \\tEpoch [{:d}\/{:d}]  val_Loss: {:.4f}\\tval_Acc: {:.2f}\\tval_kappa: {:.2f}\\tval_BCA: {:.2f}\"\n            print_str = fmt_str.format(self.curr_epcoh,\n                                        max_epochs,\n                                        val_loss,\n                                        val_acc*100,\n                                        val_kappa*100,\n                                        val_BCA*100)\n            print(print_str)\n            print(tran_time)\n            print(len(self.y_val))\n            self.val_writer.add_scalar('acc', val_acc, self.curr_iter)\n            self.val_writer.add_scalar('loss', val_loss, self.curr_iter)\n            self.val_writer.add_scalar('kappa', val_kappa, self.curr_iter)\n            self.val_writer.add_scalar('bca', val_BCA, self.curr_iter)\n\n\ndef train_all_subject(num_epochs=max_epochs, log_path=None):\n    # Leave-Subject-Out cross validation\n    subj_nums, fold_pairs, EEGs, labels = load_data(file_path, subj_nums=38, channel=1)  \n    for subj_id in range(subj_nums):\n        print('The subj_id', subj_id, '\\t\\t Training the ' + model_type + ' Model...')\n        trainer = Trainer(EEGs, labels, fold_pairs[subj_id], subj_id, log_path)\n        for epoch in range(trainer.curr_epcoh, max_epochs):\n            if trainer.train_end_flag is False:\n                trainer.train(epoch)\n                trainer.validation()\n                print('-'*50)\n            else:\n                break\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53261966,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><tfrecord>","CreationDate":"2018-11-12T12:13:12.667","AcceptedAnswerId":null,"OwnerUserId":10640511.0,"Title":"Multivariate Variable Length Timeseries in TFRecord dataset","Body":"<p>I am trying to compose a TFRecord dataset from the UCI repository: <a href=\"https:\/\/archive.ics.uci.edu\/ml\/datasets\/Character+Trajectories\" rel=\"nofollow noreferrer\">https:\/\/archive.ics.uci.edu\/ml\/datasets\/Character+Trajectories<\/a><\/p>\n\n<p>The data consists of 2858 samples with each a class and a multivariate (=3) sequence of measurements. Maximum length is 205 steps.<\/p>\n\n<p>When I create the TFRecord dataset, it seems that a tf.train.SequenceExample is the object to go for. But all the examples cover only univariate timeseries.<\/p>\n\n<pre><code>with tf.python_io.TFRecordWriter('\/tmp\/test.tfrecord') as writer:\nfor seq, cls in zip(sequences, classes)):\n    LEN = seq.shape[1]\n    example = tf.train.SequenceExample()\n    example.context.feature[\"length\"].int64_list.value.append(LEN)\n    example.context.feature[\"class\"].int64_list.value.append(cls)\n    example_tokens = record_seq.feature_lists.feature_list['sequence']\n    for i in range(LEN):\n        #THIS does not work for 3xLEN sequences\n        example_tokens.feature.add().float_list.value.append(seq[:,i])\n<\/code><\/pre>\n\n<p>How do I write the 3xLEN sequence to TFrecord using SequenceExample?<\/p>\n","answers":[]}
+{"QuestionId":53262744,"AnswerCount":1,"Tags":"<python><tensorflow><vector>","CreationDate":"2018-11-12T13:01:39.957","AcceptedAnswerId":null,"OwnerUserId":5180220.0,"Title":"how to create sparseVector with index and value in Tensorflow, python?","Body":"<p>i have two Tensors, like this:<\/p>\n\n<pre><code>&gt;&gt;&gt; xx_idx\n&lt;tf.Tensor 'Placeholder:0' shape=(100, ?) dtype=int64&gt;\n&gt;&gt;&gt; xx_val\n&lt;tf.Tensor 'Placeholder_1:0' shape=(100, ?) dtype=float64&gt;\n<\/code><\/pre>\n\n<p>How to create a SparseTensor from them? xx_idx are the indexes, xx_val are the values. \nThere are 100 samples.\nThe dimension of the vector is unknown, maybe 22000.<\/p>\n\n<p>I tried this:<\/p>\n\n<pre><code>xx_vec = tf.SparseTensor(xx_idx, xx_val, 25000)\n<\/code><\/pre>\n\n<p>but here comes the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nNameError: name 'user_idx' is not defined\n&gt;&gt;&gt; xx_vec = tf.SparseTensor(xx_idx, xx_val, 25000)\nTraceback (most recent call last):\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 671, in merge_with\n    self.assert_same_rank(other)\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 716, in assert_same_rank\n    other))\nValueError: Shapes (100, ?) and (?,) must have the same rank\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 746, in with_rank\n    return self.merge_with(unknown_shape(ndims=rank))\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 677, in merge_with\n    raise ValueError(\"Shapes %s and %s are not compatible\" % (self, other))\nValueError: Shapes (100, ?) and (?,) are not compatible\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/sparse_tensor.py\", line 133, in __init__\n    values_shape = values.get_shape().with_rank(1)\n  File \"\/home\/work\/tf\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 748, in with_rank\n    raise ValueError(\"Shape %s must have rank %d\" % (self, rank))\nValueError: Shape (100, ?) must have rank 1\n<\/code><\/pre>\n","answers":[{"AnswerId":"53276063","CreationDate":"2018-11-13T07:40:14.380","ParentId":null,"OwnerUserId":"5180220","Title":null,"Body":"<p>problem solved<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import TensorShape, Dimension\n\n\nclass GetVector:\n\n    @classmethod\n    def get_sparse_vector(cls, idx_all_0, val_all, dim_num):\n        batch_size = idx_all_0.shape[0].value\n        '''\n        cur_idx = tf.placeholder(tf.int64, [batch_size, None, 1])\n        cur_val = tf.placeholder(tf.float64, [batch_size, None])\n        cur_vec = tf.placeholder(tf.float64, [batch_size, dim_num])\n        '''\n        idx_all = cls.idx_reform(idx_all_0)\n\n        ans = []\n        for i in range(batch_size):\n            cur_idx = idx_all[i]\n            cur_val = val_all[i]\n            cur_vec = tf.SparseTensor(cur_idx, cur_val, TensorShape(Dimension(dim_num)))\n            cur_vec_tensor = tf.sparse_tensor_to_dense(cur_vec)\n            ans = tf.concat([ans, cur_vec_tensor], 0)\n        ans = tf.reshape(ans, [batch_size, dim_num])\n        return ans\n\n    @classmethod\n    def idx_reform(cls, idx_all):\n        batch_size = idx_all.shape[0].value\n        return tf.reshape(idx_all, [batch_size, -1, 1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53263216,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><generative-adversarial-network>","CreationDate":"2018-11-12T13:28:37.233","AcceptedAnswerId":53271942.0,"OwnerUserId":8354233.0,"Title":"Tensorflow - keras - Shape errors with 'strided_slice' (used resized MNIST dataset)","Body":"<p>I am going to make some GAN-Model Tester using tf.keras with MNIST hand-writed digits dataset. Because my model is going to be used in 128x128 images, I resized MNIST dataset to 128x128x1. but, the program makes some errors, that I never seen.<\/p>\n\n<pre><code>(x_train, _), (_, _) = mnist.load_data()\nx_train = (x_train.astype(np.float32) - 127.5) \/ 127.5\nx_train = tf.image.resize_images(x_train, [128, 128])\n\nidx = np.random.randint(0, x_train.shape[0], batch_size)  # picks some data, count is batch_size=32.\nimgs = x_train[idx]  # This line made errors\n<\/code><\/pre>\n\n<p>The last line made two errors:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape must be rank 1 but is rank 2 for 'strided_slice_1' (op: 'StridedSlice') with input shapes: [60000,128,128,1], [1,32], [1,32], [1].<\/p>\n<\/blockquote>\n\n<p>and,<\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice_1' (op: 'StridedSlice') with input shapes: [60000,128,128,1], [1,32], [1,32], [1].<\/p>\n<\/blockquote>\n\n<p>I think the number '32' means batch_size(=32).<\/p>\n\n<p>I tried to find about this error, but I couldn't find like this error.<\/p>\n\n<p>I don't have any ideas to solve this problems (because I started to use keras a week ago, before I used pytorch).<\/p>\n","answers":[{"AnswerId":"53271942","CreationDate":"2018-11-13T00:11:27.410","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You have more issue with your code above, but the primer cause of error is that tensorflow doesn't  support the numpy type advanced slicing.\nActually the error message is because tensorflow try to align your input array in his <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/strided_slice\" rel=\"nofollow noreferrer\">strided-slices<\/a>:<\/p>\n\n<p>An example for the strided-slices:<\/p>\n\n<blockquote>\n  <p>foo[5:,:,:3] on a 7x8x9 tensor is equivalent to foo[5:7,0:8,0:3].\n  foo[::-1] reverses a tensor with shape 8.<\/p>\n<\/blockquote>\n\n<p>Unfortunately only basic type indexing is available in Tensorflow currently. Advanced type indexing is under development.<\/p>\n\n<p>The secondary problem, that your resizing was not proper.\nTensorflow assumes a 3D or 4D input. You tried to pass a 2D image to `tf.image.resize_images(), which doesn't return the required new image dimensions. So we have to reshape the raw images like this:<\/p>\n\n<pre><code>x_train = x_train.reshape((-1, x_train.shape[1], x_train.shape[1], 1))\n<\/code><\/pre>\n\n<p>only then we can pass them to:<\/p>\n\n<pre><code>`x_train = tf.image.resize_images(x_train, [128, 128])\n<\/code><\/pre>\n\n<p>It will return then the proper dimensions:<\/p>\n\n<pre><code>print(x_train.shape)\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>(60000, 128, 128, 1)\n<\/code><\/pre>\n\n<p>So summarizing the whole solution, currently you can do it as follows:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nbatch_size = 32\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, _), (_, _) = mnist.load_data()\n\nx_train = x_train.reshape((-1, x_train.shape[1], x_train.shape[1], 1))\nx_train = (x_train.astype(np.float32) - 127.5) \/ 127.5\nx_train = tf.image.resize_images(x_train, [128, 128])\n\nidx = np.random.randint(0, x_train.shape[0], batch_size)\n\nimgs = [x_train[i,:,:,:] for i in idx]\n<\/code><\/pre>\n\n<p>which is a pretty messy \"solution\".<\/p>\n\n<p><strong>Other, actually a real solution with rearrange the original code we can achieve what we aimed to as a workaround to the tensorflow indexing issue:<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nbatch_size = 32\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, _), (_, _) = mnist.load_data()\n\nx_train = (x_train.astype(np.float32) - 127.5) \/ 127.5\nidx = np.random.randint(0, x_train.shape[0], batch_size)\nx_train = x_train[idx]\n\nx_train = x_train.reshape((-1, x_train.shape[1], x_train.shape[1], 1))\nx_train = tf.image.resize_images(x_train, [128, 128])\n\nprint(x_train.shape)\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>(32, 128, 128, 1)\n<\/code><\/pre>\n\n<p>That's it!<\/p>\n\n<p>Alternatively instead of tf.image.resize_images() you can use an additional image tools like <a href=\"http:\/\/scikit-image.org\/docs\/dev\/api\/skimage.transform.html\" rel=\"nofollow noreferrer\">skimage.transform.resize()<\/a> from <a href=\"https:\/\/scikit-image.org\/\" rel=\"nofollow noreferrer\">scikit-image<\/a> which returns numpy array type data.<\/p>\n"}]}
+{"QuestionId":53263516,"AnswerCount":1,"Tags":"<python><tensorflow><rasa-nlu>","CreationDate":"2018-11-12T13:45:42.257","AcceptedAnswerId":53426894.0,"OwnerUserId":null,"Title":"is there a way to change RASA_NLU config from visual c++ 14.0 to MinGW","Body":"<p>is there a way to change RASA_NLU config from visual c++ 14.0 to MinGW<\/p>\n\n<p>error: Microsoft Visual C++ 14.0 is required. Get it with \"Microsoft Visual C++ Build Tools\": <a href=\"http:\/\/landinghub.visualstudio.com\/visual-cpp-build-tools\" rel=\"nofollow noreferrer\">http:\/\/landinghub.visualstudio.com\/visual-cpp-build-tools<\/a><\/p>\n","answers":[{"AnswerId":"53426894","CreationDate":"2018-11-22T08:42:20.317","ParentId":null,"OwnerUserId":"3429596","Title":null,"Body":"<p>Please make sure that <code>C:\\MinGW_w64\\bin<\/code> is in your <code>PATH<\/code> environment variable. Then create a file <code>distutils.cfg<\/code> in the directory <code>&lt;Python Installation Directory&gt;\\Lib\\distutils<\/code> and add the following content:<\/p>\n\n<pre><code>[build]\ncompiler=mingw32\n\n[build_ext]\ncompiler=mingw32\n<\/code><\/pre>\n\n<p>Note, that mingw apparently only works with Python versions up to 3.4. See this <a href=\"https:\/\/wiki.python.org\/moin\/WindowsCompilers#GCC_-_MinGW-w64_.28x86.2C_x64.29\" rel=\"nofollow noreferrer\">wikipage<\/a> for more information.<\/p>\n"}]}
+{"QuestionId":53263546,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorflow-serving>","CreationDate":"2018-11-12T13:47:45.240","AcceptedAnswerId":53265554.0,"OwnerUserId":8334261.0,"Title":"Requests to TensorFlow serving's predict API returns error \"Missing inputs\"","Body":"<p>I have trained a simple regression model to fit a linear function with the following equation: <code>y = 3x + 1<\/code>. For testing purposes, I saved the model as checkpoints, so that I could resume training and wouldn't have to start from scratch every time. <\/p>\n\n<p>Now I want to make this model available via TF serving. For this reason, I had to convert it into the <code>SavedModel<\/code> format of tensorflow via this script:<\/p>\n\n<pre><code>import tensorflow as tf\nimport restoretest as rt  ## just the module that contains the linear model\n\ntf.reset_default_graph()        \n\nlatest_checkpoint = tf.train.latest_checkpoint('path\/to\/checkpoints')\nmodel = rt.LinearModel()\nsaver = tf.train.Saver()\n\nexport_path = 'path\/to\/export\/folder'\n\nwith tf.Session() as sess:\n\n    if latest_checkpoint:\n        saver.restore(sess, latest_checkpoint)\n    else:\n        raise ValueError('No checkpoint file found') \n\n    print('Exporting trained model to', export_path)\n\n    builder = tf.saved_model.builder.SavedModelBuilder(export_path)\n\n    ## define inputs and outputs\n\n    tensor_info_x = tf.saved_model.utils.build_tensor_info(model.x)\n    tensor_info_y = tf.saved_model.utils.build_tensor_info(model.y_pred)\n\n    prediction_signature = (\n            tf.saved_model.signature_def_utils.build_signature_def(\n                    inputs={'xvals': tensor_info_x},\n                    outputs={'yvals': tensor_info_y},\n                    method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))\n\n\n    builder.add_meta_graph_and_variables(sess, \n                                         [tf.saved_model.tag_constants.SERVING],\n                                         signature_def_map={tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature},\n                                         main_op=tf.tables_initializer(),\n                                         strip_default_attrs=True)\n\n    builder.save()\n\n    print('Done exporting')\n<\/code><\/pre>\n\n<p>This creates a folder (as expected) with the contents:<\/p>\n\n<pre><code>export_folder\n    |-saved_model.pb\n    |-variables\n        |-variables.index\n        |-variables.data-00000-of-00001\n<\/code><\/pre>\n\n<p>To serve this with tf serving and docker, I pulled the tensorflow\/serving image from docker and ran the container via the command:<\/p>\n\n<pre><code>sudo docker run -p 8501:8501 --mount type=bind,source=path\/to\/export\/folder,target=models\/linear -e MODEL_NAME=linear -t tensorflow\/serving\n<\/code><\/pre>\n\n<p>This seems to execute without problems, as I get a lot of infos. In the last line of the output it says<\/p>\n\n<blockquote>\n  <p>[evhttp_server.cc : 237] RAW: Entering the event loop ...<\/p>\n<\/blockquote>\n\n<p>I guess the server is waiting for requests. Now, when I try to send a request to it via curl, I get an error:<\/p>\n\n<pre><code>curl -d '{\"xvals\": [1.0 2.0 5.0]}' -X POST http:\/\/localhost:8501\/v1\/models\/linear:predict\n<\/code><\/pre>\n\n<blockquote>\n  <p>{ \"error\": \"Missing \\'inputs\\' or \\'instances\\' key\" }<\/p>\n<\/blockquote>\n\n<p>What am I doing wrong? The model works when I send dummy values via the <code>saved_model_cli<\/code>.<\/p>\n","answers":[{"AnswerId":"53265554","CreationDate":"2018-11-12T15:45:31.197","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>Looks like the body of the POST request should be modified. According to <a href=\"https:\/\/www.tensorflow.org\/serving\/api_rest#request_format_2\" rel=\"nofollow noreferrer\">documentation<\/a> the format should be<\/p>\n\n<p><code>{ \"inputs\": {\"xvals\": [1.0 2.0 5.0]} }<\/code><\/p>\n"}]}
+{"QuestionId":53264074,"AnswerCount":0,"Tags":"<android><tensorflow><android-camera2><android-things>","CreationDate":"2018-11-12T14:17:53.313","AcceptedAnswerId":null,"OwnerUserId":1374416.0,"Title":"Android Things - Creating a Camera Preview Session Fails, and no Preview is shown","Body":"<p>I am trying to deploy the Android TensorFlow-Lite Example, specifically, the Detector Activity.<\/p>\n\n<p>I have had success in deploying it in a Tablet. The app works great, it is able to detect objects, put a bounding rectangle around it, with a label as well as a confidence level.<\/p>\n\n<p>I then set up my Raspberry Pi 3 Model B Board, installed Android Things in it, connected via ADB, and then deployed the same program from Android Studio. However, the screen I was using for my R\u03c0 board was blank. <\/p>\n\n<p>Upon checking a <a href=\"https:\/\/github.com\/weichen2046\/CameraDemoForAndroidThings\" rel=\"nofollow noreferrer\">Camera Demo For Android Things tutorial<\/a>, I had this idea to enable hardware acceleration in order to support the Camera Preview. I added in:<\/p>\n\n<pre><code>android:hardwareAccelerated=\"true\"\n<\/code><\/pre>\n\n<p>in the <code>application<\/code> tag of the Manifest.  <\/p>\n\n<p>I also added in the following within the application tag:<\/p>\n\n<pre><code>&lt;uses-library android:name=\"com.google.android.things\" \/&gt;\n<\/code><\/pre>\n\n<p>And an intent filter in my activity tag:<\/p>\n\n<pre><code>&lt;intent-filter&gt;\n    &lt;action android:name=\"android.intent.action.MAIN\" \/&gt;\n    &lt;category android:name=\"android.intent.category.IOT_LAUNCHER\" \/&gt;\n    &lt;category android:name=\"android.intent.category.DEFAULT\" \/&gt;\n&lt;\/intent-filter&gt;\n<\/code><\/pre>\n\n<p>So that the TensorFlow App runs after boot.<\/p>\n\n<p>I deployed the application again, but the same error persists -- I am unable to configure the preview screen session.<\/p>\n\n<p>Here is the following code that was included in the TensorFlow Example:<\/p>\n\n<pre><code>private void createCameraPreviewSession() {\n    try {\n        final SurfaceTexture texture = textureView.getSurfaceTexture();\n        assert texture != null;\n\n        \/\/ We configure the size of default buffer to be the size of camera preview we want.\n        texture.setDefaultBufferSize(previewSize.getWidth(), previewSize.getHeight());\n\n        \/\/ This is the output Surface we need to start preview.\n        final Surface surface = new Surface(texture);\n\n        \/\/ We set up a CaptureRequest.Builder with the output Surface.\n        previewRequestBuilder = cameraDevice.createCaptureRequest(CameraDevice.TEMPLATE_PREVIEW);\n        previewRequestBuilder.addTarget(surface);\n\n        LOGGER.e(\"Opening camera preview: \" + previewSize.getWidth() + \"x\" + previewSize.getHeight());\n\n        \/\/ Create the reader for the preview frames.\n        previewReader =\n                ImageReader.newInstance(\n                        previewSize.getWidth(), previewSize.getHeight(), ImageFormat.YUV_420_888, 2);\n\n        previewReader.setOnImageAvailableListener(imageListener, backgroundHandler);\n        previewRequestBuilder.addTarget(previewReader.getSurface());\n\n        \/\/ Here, we create a CameraCaptureSession for camera preview.\n        cameraDevice.createCaptureSession(\n                Arrays.asList(surface, previewReader.getSurface()),\n                new CameraCaptureSession.StateCallback() {\n\n                    @Override\n                    public void onConfigured(final CameraCaptureSession cameraCaptureSession) {\n                        \/\/ The camera is already closed\n                        if (null == cameraDevice) {\n                            return;\n                        }\n\n                        \/\/ When the session is ready, we start displaying the preview.\n                        captureSession = cameraCaptureSession;\n                        try {\n                            \/\/ Auto focus should be continuous for camera preview.\n                            previewRequestBuilder.set(\n                                    CaptureRequest.CONTROL_AF_MODE,\n                                    CaptureRequest.CONTROL_AF_MODE_CONTINUOUS_PICTURE);\n                            \/\/ Flash is automatically enabled when necessary.\n                            previewRequestBuilder.set(\n                                    CaptureRequest.CONTROL_AE_MODE, CaptureRequest.CONTROL_AE_MODE_ON_AUTO_FLASH);\n\n                            \/\/ Finally, we start displaying the camera preview.\n                            previewRequest = previewRequestBuilder.build();\n                            captureSession.setRepeatingRequest(\n                                    previewRequest, captureCallback, backgroundHandler);\n                        } catch (final CameraAccessException e) {\n                            LOGGER.e(e, \"Exception!\");\n                            LOGGER.e(\"camera access exception!\");\n                        }\n                    }\n\n                    @Override\n                    public void onConfigureFailed(final CameraCaptureSession cameraCaptureSession) {\n                        showToast(\"Failed\");\n                        LOGGER.e(\"configure failed!!\");\n                    }\n                },\n                null);\n    } catch (final CameraAccessException e) {\n        LOGGER.e(\"camera access exception!\");\n        LOGGER.e(e, \"Exception!\");\n    }\n}\n<\/code><\/pre>\n\n<p>The error log is the one that is in the <code>onConfigureFailed<\/code> override method, and the relevant error logs leading towards that statement are:<\/p>\n\n<pre><code>11-12 14:02:40.677 1991-2035\/org.tensorflow.demo E\/CameraCaptureSession: Session 0: Failed to create capture session; configuration failed\n11-12 14:02:40.679 1991-2035\/org.tensorflow.demo E\/tensorflow: CameraConnectionFragment: configure failed!!\n<\/code><\/pre>\n\n<p>However, I couldn't trace the <code>Session 0:<\/code> stack trace. <\/p>\n\n<p>Aside from turning on hardware acceleration and adding several other tags to the Manifest, I have not tried anything.<\/p>\n\n<p>I have done my research and I have seen other examples, but they only take a photo at the click of a button. I need a working cameraPreview. <\/p>\n\n<p>I also have that CameraDemoForAndroidThings example, but I'm afraid that I don't know a lick of Kotlin to be able to guess how it works.<\/p>\n\n<p>If there's any one who managed to make a Java version of the TensorFlow Detection Activity run on Raspberry Pi Android Things, kindly contribute and let us know how you did it. <\/p>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>Apparently, the camera can only support one stream configuration at a time. I was also able to infer that I have to modify the <code>createCaptureSession()<\/code> function to only use one surface, my function now looks like this:<\/p>\n\n<pre><code>cameraDevice.createCaptureSession(\n\/\/                    Arrays.asList(surface, previewReader.getSurface()),\n        Arrays.asList(surface),\n        new CameraCaptureSession.StateCallback() {\n\n            @Override\n            public void onConfigured(final CameraCaptureSession cameraCaptureSession) {\n                \/\/ The camera is already closed\n                if (null == cameraDevice) {\n                    return;\n                }\n\n                \/\/ When the session is ready, we start displaying the preview.\n                captureSession = cameraCaptureSession;\n                try {\n                    \/\/ Auto focus should be continuous for camera preview.\n\/\/                                previewRequestBuilder.set(\n\/\/                                        CaptureRequest.CONTROL_AF_MODE,\n\/\/                                        CaptureRequest.CONTROL_AF_MODE_CONTINUOUS_PICTURE);\n                    \/\/ Flash is automatically enabled when necessary.\n\/\/                                previewRequestBuilder.set(\n\/\/                                        CaptureRequest.CONTROL_AE_MODE, CaptureRequest.CONTROL_AE_MODE_ON_AUTO_FLASH);\n\n                    \/\/ Finally, we start displaying the camera preview.\n                    previewRequest = previewRequestBuilder.build();\n                    captureSession.setRepeatingRequest(\n                            previewRequest, captureCallback, backgroundHandler);\n\n                    previewRequestBuilder.addTarget(previewReader.getSurface());\n\n                } catch (final CameraAccessException e) {\n                    LOGGER.e(\"exception hit while configuring camera!\");\n                    LOGGER.e(e, \"Exception!\");\n                }\n            }\n\n            @Override\n            public void onConfigureFailed(final CameraCaptureSession cameraCaptureSession) {\n                LOGGER.e(\"Configure failed!\");\n                showToast(\"Failed\");\n            }\n        },\n        null);\n<\/code><\/pre>\n\n<p>This enables me to get a live preview. However, the code does not proceed with sending the image from the preview to the <code>processImage()<\/code> block. <\/p>\n\n<p>Has anyone successfully implemented the TensorFlow-Lite examples that involve live Camera Previews to Android Things?<\/p>\n","answers":[]}
+{"QuestionId":53264181,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-12T14:23:34.403","AcceptedAnswerId":53269437.0,"OwnerUserId":9880592.0,"Title":"How can I get best val_acc and val_loss simultaneously in CNN using keras?","Body":"<p>I've used <code>ModelCheckpoint(file_path, monitor='val_acc', verbose=1, save_best_only=True, mode='max')<\/code> to get the max val_acc. So, when i run my program I get a many of sample such as: <\/p>\n\n<ol>\n<li>val_acc = 0.92857 &amp; val_loss = 0.2495, <\/li>\n<li>val_acc = 0.98750 &amp; val_loss = 0.6820.<\/li>\n<\/ol>\n\n<p>So, it takes 2 no., although it's val_loss in large.\nHow can I choose the no. 1 with lower val_loss?\nAlternatively, how can i choose the best fitted model keeping val_acc and val_loss value in mind?<\/p>\n","answers":[{"AnswerId":"53269437","CreationDate":"2018-11-12T20:14:48.583","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>You should really just factor one of them for your checkpoint. However, you could just create two callbacks and save the network based on both metrics.<\/p>\n\n<pre><code>call_acc = ModelCheckpoint(weights.{epoch:02d}-{val_acc:.2f}.hdf5, monitor='val_acc', save_best_only=True, mode='max')\ncall_loss = ModelCheckpoint(weights.{epoch:02d}-{val_loss:.2f}.hdf5, monitor='val_loss', save_best_only=True, mode='min')\nmodel.fit(X, Y, epochs, batch_siz, callbacks=[call_loss, call_acc])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53264206,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-12T14:25:34.840","AcceptedAnswerId":null,"OwnerUserId":10550766.0,"Title":"How to feed a Tensorflow Dataset with more than two Values to Keras","Body":"<p>I have a <code>tf.Dataset<\/code> that consists of three seperate entries. This means when I call from my dataset.iterator <code>get_next()<\/code> I get the following values:<\/p>\n\n<pre><code>dataset = inputs(...) # Extracts Data from a TFRecord and preprocesses it\niterator = dataset.make_initializable_iterator()\nsess.run(iterator.initializer)\nsess.run(tf.tables_initializer())\ninput_1, label_1, label_2 = iterator.get_next()\n<\/code><\/pre>\n\n<p>Furthermore i built a custom loss function so my loss can takes the two labels to compute a single loss even though it has only a single output. This Loss is inspired by <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2121#issuecomment-214551349\" rel=\"nofollow noreferrer\">this<\/a>: <\/p>\n\n<pre><code>def loss_func(label_1, label_2):\n    def loss(y_true, y_pred):  #  y_true is a dummy that is not used.\n        \"\"\"...\"\"\"\n    return scalar\nreturn loss\n<\/code><\/pre>\n\n<p>I would like to feed my Dataset into the Keras <code>model.fit(Dataset, ...)<\/code> function. However I don't know how Keras handles the three different Values of the Dataset instead of the standard (value, label) pair.<br>\nMy current guess is that Keras creates an iterator when calling fit going over the entire Dataset. For the iterator it assigns the first value as input x and the second value as label y.<\/p>\n\n<p><strong>Which way can I feed my Dataset to the model.fit() function or what alternatives to I have?<\/strong><\/p>\n\n<hr>\n\n<p>The Code looks in the end like this:<\/p>\n\n<pre><code>def loss_func(label_1, label_2):\n    def loss(y_true, y_pred):  #  y_true is a dummy that is not used.\n        \"\"\"...\"\"\"\n    return scalar\nreturn loss\n\nmodel = my_keras_model.get_model(...) # Returns the Keras Model\n\ndataset = inputs(...)\niterator = dataset.make_initializable_iterator()\nsess.run(iterator.initializer)\nsess.run(tf.tables_initializer())\ninput_1, label_1, label_2 = iterator.get_next()\n\nloss = loss_func(label_1, label_2)\n\nmodel.compile(optimizer='adam',\n              loss=loss)\n\nmodel.fit(???)\n# or model.train_on_batch()?\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53264316,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-12T14:31:13.667","AcceptedAnswerId":null,"OwnerUserId":3107479.0,"Title":"The tensorflow tutorial cannot load keras library","Body":"<p>I'm completely new to Tensorflow, and I've tried installing it on a Windows 10 desktop.<\/p>\n\n<p>I've followed the instructions on the Tensorflow web site.<\/p>\n\n<p>Windows 10 Enterprise 10.0.16299 Build 16299\npython x64 3.6.7\npip 18.1<\/p>\n\n<p>I've set up a virtual environment and it has these packages.<\/p>\n\n<pre><code>Package             Version\n------------------- -------\nabsl-py             0.6.1\nastor               0.7.1\ngast                0.2.0\ngrpcio              1.16.0\nh5py                2.8.0\nKeras-Applications  1.0.6\nKeras-Preprocessing 1.0.5\nMarkdown            3.0.1\nnumpy               1.15.4\npip                 18.1\nprotobuf            3.6.1\nsetuptools          40.5.0\nsix                 1.11.0\ntensorboard         1.12.0\ntensorflow          1.12.0\ntermcolor           1.1.0\nvirtualenv          16.1.0\nWerkzeug            0.14.1\nwheel               0.32.2\n<\/code><\/pre>\n\n<p>If I run the verify install :<\/p>\n\n<pre><code>python -c \"import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))\"\n<\/code><\/pre>\n\n<p>then it responds with :<\/p>\n\n<pre><code>2018-11-12 14:22:54.667533: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\ntf.Tensor(-344.4322, shape=(), dtype=float32)\n<\/code><\/pre>\n\n<p>However if I try to run the example script in the tutorials (the one you see on the page that uses the mnist dataset) then I get the error :<\/p>\n\n<pre><code>AttributeError: module 'tensorflow' has no attribute 'keras'\n<\/code><\/pre>\n\n<p>However if I run the following the tutorial script line by line via the python prompt then it executes without exception.<\/p>\n\n<p>Ideas anyone?<\/p>\n","answers":[{"AnswerId":"53264733","CreationDate":"2018-11-12T14:56:04.647","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>Keras was added to tensorflow in version <code>1.4<\/code>.\nContrary to the package list you pasted above, the error message suggests to me you're probably running an older version of Tensorflow when you run the example script.<\/p>\n\n<p>You can verify that by adding<\/p>\n\n<p><code>print(tf.VERSION)<\/code><\/p>\n\n<p>to your script and running it and if the Tensorflow version is older than <code>1.4<\/code> you will just need to update it.<\/p>\n"},{"AnswerId":"53264764","CreationDate":"2018-11-12T14:57:14.020","ParentId":null,"OwnerUserId":"3107479","Title":null,"Body":"<p>omg.  school boy error.<\/p>\n\n<p>when trying the tutorial code, I created a python script called tensorflow.py<\/p>\n\n<p>hence there was a namespace\/class-name clash.<\/p>\n\n<p>i renamed the script to bob.py and hey presto all works.<\/p>\n"}]}
+{"QuestionId":53265082,"AnswerCount":2,"Tags":"<tensorflow><pre-trained-model>","CreationDate":"2018-11-12T15:16:06.087","AcceptedAnswerId":53266300.0,"OwnerUserId":3655743.0,"Title":"Where to find tensorflow pretrained models (list or download link)","Body":"<p>I am starting to work with Intel movidius neural compute stick.\nTo start working, in my case, it is necessary to download pretrained models.\nIn the tutorials they refer to <a href=\"http:\/\/download.tensorflow.org\/models\/\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/<\/a>.<\/p>\n\n<p>However, there is not a list that shows all the models available for download.<\/p>\n\n<p>If the latest version of a net, lets say inception_v4, is not known, I cannot download the corresponding .tar.gz file.<\/p>\n\n<p>Does anyone know a method to have an updated list of the .tar.gz files of the pretrained models available for download?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53266300","CreationDate":"2018-11-12T16:27:33.437","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>The following two links may help <\/p>\n\n<ol>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">detection_model_zoo<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim#Pretrained\" rel=\"nofollow noreferrer\">\nTensorFlow-Slim image classification model library<\/a><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"53270473","CreationDate":"2018-11-12T21:38:15.623","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p><a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">keras.applications<\/a> makes it easy to load models and their pretrained weights if you can use Keras<\/p>\n"}]}
+{"QuestionId":53265124,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-12T15:18:46.853","AcceptedAnswerId":null,"OwnerUserId":7702772.0,"Title":"Keras LSTM Input - ValueError: Error when checking input: expected input_1 to have 3 dimensions, but got array with shape (1745, 1)","Body":"<p>My Keras RNN code is as follows:<\/p>\n\n<pre><code>def RNN(): \n   inputs = Input(shape = (None, word_vector_size))\n   layer = LSTM(64)(inputs)\n   layer = Dense(256,name='FC1')(layer)\n   layer = Dropout(0.5)(layer)\n   layer = Dense(num_classes,name='out_layer')(layer)\n   layer = Activation('softmax')(layer)\n   model = Model(inputs=inputs,outputs=layer)\n   return model\n<\/code><\/pre>\n\n<p>I'm getting the error when I call model.fit()<\/p>\n\n<pre><code>model.fit(np.array(word_vector_matrix), np.array(Y_binary), batch_size=128, epochs=10, validation_split=0.2, callbacks=[EarlyStopping(monitor='val_loss',min_delta=0.0001)])\n<\/code><\/pre>\n\n<p>Word_vector_matrix is a 3-dim numpy array. \nI have printed the following :<\/p>\n\n<pre><code>print(type(word_vector_matrix), type(word_vector_matrix[0]), type(word_vector_matrix[0][0]), type(word_vector_matrix[0][0][0]))\n<\/code><\/pre>\n\n<p>and the answer is : <\/p>\n\n<pre><code>&lt;class 'numpy.ndarray'&gt; &lt;class 'numpy.ndarray'&gt; &lt;class 'numpy.ndarray'&gt; &lt;class 'numpy.float32'&gt;\n<\/code><\/pre>\n\n<p>It's shape is 1745 x sentence length x word vector size. \nThe sentence length is variable and I'm trying to pass this entire word vector matrix to the RNN, but I get the error above. <\/p>\n\n<p>The shape is printed like:<\/p>\n\n<pre><code>print(word_vector_matrix.shape)\n<\/code><\/pre>\n\n<p>The answer is <code>(1745,)<\/code><\/p>\n\n<p>The shape of the nested arrays are printed like:<\/p>\n\n<pre><code>print(word_vector_matrix[10].shape)\n<\/code><\/pre>\n\n<p>The answer is <code>(7, 300)<\/code>\nThe first number 7 denotes the sentence length, which is variable and changes for each sentence, and the second number is 300, which is fixed for all words and is the word vector size. <\/p>\n\n<p>I have converted everything to np.array() as suggested by the other posts, but still the same error. Can someone please help me. I'm using python3 btw. The similar thing is working in python2 for me, but not in python3. Thanks!<\/p>\n","answers":[{"AnswerId":"53622327","CreationDate":"2018-12-04T22:21:10.517","ParentId":null,"OwnerUserId":"6812182","Title":null,"Body":"<p><code>word_vector_matrix<\/code> is <strong>not<\/strong> a 3-D ndarray. It's a 1-D ndarray of 2-D arrays. This is due to variable sentence length. <\/p>\n\n<p>Numpy allows <code>ndarray<\/code> to be list-like structures that may contain a complex element (another <code>ndarray<\/code>). In Keras however, the <code>ndarray<\/code> must be converted into a <code>Tensor<\/code> (which has to be a \"mathematical\" matrix of some dimension - this is required for the sake of efficient computation).<\/p>\n\n<p>Therefore, each batch must have fixed size sentences (and not the entire data).<\/p>\n\n<p>Here are a few alternatives:<\/p>\n\n<ol>\n<li>Use batch size of 1 - simplest approach, but impedes your network's convergence. I would suggest to only use it as a temporary sanity check.<\/li>\n<li>If sequence length <strong>variability is low<\/strong>, pad all your batches to be of the same length.<\/li>\n<li>If sequence length <strong>variability is high<\/strong>, pad each batch with the max length within that batch. This would require you to use a custom data generator.<\/li>\n<\/ol>\n\n<p>Note: After you padded your data, you need to use <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">Masking<\/a>, so that the padded part will be ignored during training.<\/p>\n"}]}
+{"QuestionId":53265281,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-12T15:28:06.293","AcceptedAnswerId":53265466.0,"OwnerUserId":4961771.0,"Title":"How to improve prediction with keras and tensorflow","Body":"<p>I am using tensorflow with keras to perform regression on some historical data. Data type as follows:<\/p>\n\n<blockquote>\n  <p>id,timestamp,ratio\n  \"santalucia\",\"2018-07-04T16:55:59.020000\",21.8\n  \"santalucia\",\"2018-07-04T16:50:58.043000\",22.2\n  \"santalucia\",\"2018-07-04T16:45:56.912000\",21.9\n  \"santalucia\",\"2018-07-04T16:40:56.572000\",22.5\n  \"santalucia\",\"2018-07-04T16:35:56.133000\",22.5\n  \"santalucia\",\"2018-07-04T16:30:55.767000\",22.5<\/p>\n<\/blockquote>\n\n<p>And I am reformulating it as a time series problem (25 time steps) so that I can predict (make a regression) for the next values of the series (variance should not be high). I am using also <code>sklearn.preprocessing MinMaxScaler<\/code> to scale the data to range (-1,1) or (0,1) depending if I use LSTM or Dense (respectively). \nI am training with two different architectures:<\/p>\n\n<p>Dense is as follows: <\/p>\n\n<pre><code>def get_model(self, layers, activation='relu'):\n    model = Sequential()\n    # Input arrays of shape (*, layers[1])\n    # Output = arrays of shape (*, layers[1] * 16)\n    model.add(Dense(units=int(64), input_shape=(layers[1],), activation=activation))\n    model.add(Dense(units=int(64), activation=activation))\n    # model.add(Dropout(0.2))\n\n    model.add(Dense(units=layers[3], activation='linear'))\n    # activation=activation))\n\n    # opt = optimizers.Adagrad(lr=self.learning_rate, epsilon=None, decay=self.decay_lr)\n    opt = optimizers.rmsprop(lr=0.001)\n    model.compile(optimizer=opt, loss=self.loss_fn, metrics=['mae'])\n    model.summary()\n    return model\n<\/code><\/pre>\n\n<p>Which more or less provides with good results (same architecture as in tensorflows' tutorial for predicting house prices).<\/p>\n\n<p>However, LSTM is not giving good results, it usually ends up stuck around a value (for example, 40 (40.0123123, 40.123123,41.09090...) and I do not see why or how to improve it. Architecture is:<\/p>\n\n<pre><code>def get_model(self, layers, activation='tanh'):\n    model = Sequential()\n    # Shape = (Samples, Timesteps, Features)\n    model.add(LSTM(units=128, input_shape=(layers[1], layers[2]),\n                   return_sequences=True, activation=activation))\n\n    model.add(LSTM(64, return_sequences=True, activation=activation))\n\n    model.add(LSTM(layers[2], return_sequences=False, activation=activation))\n    model.add(Dense(units=layers[3], activation='linear'))\n    # activation=activation))\n\n    opt = optimizers.Adagrad(lr=0.001, decay=self.decay_lr)\n    model.compile(optimizer=opt, loss='mean_squared_error', metrics=['accuracy'])\n    model.summary()\n    return model\n<\/code><\/pre>\n\n<p>I currently train with a batch size of 200 that increases by a rate of 1.5 every fit. Each fit is made of 50 epochs, and I use a keras earlystopping callback with at least 20 epoch. <\/p>\n\n<p>I have tried adding more layers, more units, reducing layers, units, increasing and decreasing learning rate, etc, but every time it gets stuck around a value. Any reason for this?<\/p>\n\n<p>Also, do you know any good practices that can be applied to this problem?<\/p>\n\n<p>Cheers<\/p>\n","answers":[{"AnswerId":"53265466","CreationDate":"2018-11-12T15:39:54.097","ParentId":null,"OwnerUserId":"10553976","Title":null,"Body":"<p>Have you tried holding back a validation set seeing how well the model performance on the training set tracks with the validation set? This is often how I catch myself overfitting.<\/p>\n\n<p>A simple function for doing this (<a href=\"https:\/\/www.kaggle.com\/amneves\/quick-draw-keras-cnn-model\" rel=\"nofollow noreferrer\">adapted from here<\/a>) can help you do that:<\/p>\n\n<pre><code>hist = model.fit_generator(...)\ndef gen_graph(history, title):\n    plt.plot(history.history['categorical_accuracy'])\n    plt.plot(history.history['val_categorical_accuracy'])\n    plt.title(title)\ngen_graph(hist, \"Accuracy, training vs. validation scores\")\n<\/code><\/pre>\n\n<p>Also, do you have enough samples? If you're really, really sure that you have done as much as you can in terms of preprocessing, and in terms of hyperparameter tuning... generating some synthetic data or doing some data augmentation has occasionally helped me.<\/p>\n"}]}
+{"QuestionId":53265473,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><hyperparameters>","CreationDate":"2018-11-12T15:40:16.993","AcceptedAnswerId":53282783.0,"OwnerUserId":712452.0,"Title":"Different models do incremental fit for RNN model while hyperparameter tuning","Body":"<p>I am quite new to deep learning and I was studying <a href=\"https:\/\/www.youtube.com\/watch?v=BSpXCRTOLJA\" rel=\"nofollow noreferrer\">this RNN example<\/a>. <\/p>\n\n<p>After completing the tutorial, I decided to see the effect of various hyperparameters such as the number of nodes in each layer and dropout factor etc. <\/p>\n\n<p>What I do is, for each value in my lists, create a new model using a set of parameters and test the performance in my dataset. Below is the basic code:<\/p>\n\n<pre><code>def build_model(MODELNAME, l1,l2,l3, l4, d):\n    tf.global_variables_initializer() \n    tf.reset_default_graph()\n    model = Sequential(name = MODELNAME)\n    model.reset_states\n\n    model.add(CuDNNLSTM(l1, input_shape=(x_train.shape[1:]), return_sequences=True) )\n    model.add(Dropout(d))\n    model.add(BatchNormalization())\n\n    model.add(CuDNNLSTM(l2, input_shape=(x_train.shape[1:]), return_sequences=True) )\n\n    # Definition of other layers of the model ...\n\n    model.compile(loss=\"sparse_categorical_crossentropy\",\n                 optimizer=opt,\n                 metrics=['accuracy'])\n\n    history = model.fit(x_train, y_train,\n                        epochs=EPOCHS,\n                        batch_size=BATCH_SIZE,\n                        validation_data=(x_validation, y_validation))\n    return model\n\nlayer1 = [64, 128, 256]\nlayer2,3,4 = [...]\ndrop = [0.2, 0.3, 0.4]\n\nconfig = tf.ConfigProto()\nconfig.gpu_options.per_process_gpu_memory_fraction = 0.4\n\n\nfor l1 in layer1:\n     #for l2, l3, l4 for layer2, layer3, layer4  \n        for d in drop:\n            sess = tf.Session(config=config)\n            set_session(sess)\n\n            MODELNAME = 'RNN-l1={}-l2={}-l3={}-l4={}-drop={} '.format(l1, l2, l3, l4, d)\n            print(MODELNAME)\n\n            model = build_model(MODELNAME, l1,l2,l3, l4, d)\n            sess.close()\n            print('-----&gt; training &amp; validation loss &amp; accuracies)\n<\/code><\/pre>\n\n<p>The problem is when the new model is built using the new parameters, it works as if the next epoch of the previous model, rather than epoch 1 of the new one. Below is some of the results.<\/p>\n\n<pre><code>RNN-l1=64-l2=64-l3=64-l4=32-drop=0.2 \nTrain on 90116 samples, validate on 4458 samples\nEpoch 1\/6\n90116\/90116 [==============================] - 139s 2ms\/step - loss: 0.5558 - acc: 0.7116 - val_loss: 0.8857 - val_acc: 0.5213\n... # results for other epochs\nEpoch 6\/6\nRNN-l1=64-l2=64-l3=64-l4=32-drop=0.3 \n90116\/90116 [==============================] - 140s 2ms\/step - loss: 0.5233 - acc: 0.7369 - val_loss: 0.9760 - val_acc: 0.5336\nEpoch 1\/6\n90116\/90116 [==============================] - 142s 2ms\/step - loss: 0.5170 - acc: 0.7403 - val_loss: 0.9671 - val_acc: 0.5310\n... # results for other epochs\n90116\/90116 [==============================] - 142s 2ms\/step - loss: 0.4953 - acc: 0.7577 - val_loss: 0.9587 - val_acc: 0.5354\nEpoch 6\/6\n90116\/90116 [==============================] - 143s 2ms\/step - loss: 0.4908 - acc: 0.7614 - val_loss: 1.0319 - val_acc: 0.5397\n# -------------------AFTER 31TH SET OF PARAMETERS\nRNN-l1=64-l2=256-l3=128-l4=32-drop=0.2\nEpoch 1\/6\n90116\/90116 [==============================] - 144s 2ms\/step - loss: 0.1080 - acc: 0.9596 - val_loss: 1.8910 - val_acc: 0.5372\n<\/code><\/pre>\n\n<p>As seen, the first epoch of 31th set of parameters behaves as if it is 181th epoch. Similarly, if I stop the run at one point and re-run again, the accuracy and loss look as if it is the next epoch as below. <\/p>\n\n<pre><code>Epoch 1\/6\n90116\/90116 [==============================] - 144s 2ms\/step - loss: 0.1053 - acc: 0.9621 - val_loss: 1.9120 - val_acc: 0.5375\n<\/code><\/pre>\n\n<p>I tried a bunch of things (as you can see in the code), such as <code>model=None<\/code>, <code>reinitializing the variables<\/code>, <code>resetting_status of the model<\/code>, <code>closing session in each iteration<\/code> etc but none helped. I searched for similar question with no luck.<\/p>\n\n<p>I am trying to understand what I am doing wrong. \nAny help is appreciated, <\/p>\n\n<p>Note: Title is not very explanatory, I am open to suggestions for a better title.<\/p>\n","answers":[{"AnswerId":"53282783","CreationDate":"2018-11-13T14:04:58.403","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>Looks like you are using a Keras setting, which means you need to import keras backend and then clear that session before you run your new model. It would be something like this:<\/p>\n\n<pre><code>from keras import backend as K \nK.clear_session()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53265549,"AnswerCount":1,"Tags":"<python><scikit-learn><keras><lstm>","CreationDate":"2018-11-12T15:45:08.983","AcceptedAnswerId":53266873.0,"OwnerUserId":887611.0,"Title":"MinMaxScaler inverse_transform diferente size array","Body":"<p>So I have a Scaler:<\/p>\n\n<pre><code>scaler = MinMaxScaler(feature_range=(0, 1))\n<\/code><\/pre>\n\n<p>that has 9 columns, where column 0 is my Y, and i use my scaler to work all columns.<\/p>\n\n<p>When I make the predict:<\/p>\n\n<pre><code>yhat = model.predict(test_X)\n<\/code><\/pre>\n\n<p>I want to use the same scaler so I can transform back my values to normal, but now my output only has 1 column, and my scaler has 9 and this is a problem.<\/p>\n\n<p>So what I am hopping to find is a way where I can do something like, grab the scaler and tell him \"inverse_transform using the [0] column to work my prediction out.\"<\/p>\n\n<p>Is there a way to do this?<\/p>\n\n<p>Or the only way is to do other Scaler for my Y column and use it?<\/p>\n","answers":[{"AnswerId":"53266873","CreationDate":"2018-11-12T17:03:16.670","ParentId":null,"OwnerUserId":"10305819","Title":null,"Body":"<p>You can combine the new predicted  y value with the x value,get a 9 column matrix and  scale it back.But it would be just easier to use two different instances of MinmaxScaler for x and y , so that you can just scale the predicted output back by in-versing the scale for y.<\/p>\n"}]}
+{"QuestionId":53266213,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><tensor>","CreationDate":"2018-11-12T16:22:10.987","AcceptedAnswerId":53266992.0,"OwnerUserId":6346529.0,"Title":"Getting \"You must feed a value for placeholder tensor\" error when using Keras model","Body":"<p>I am trying to implement a model in Keras and got the following error:<\/p>\n\n<blockquote>\n  <p>You must feed a value for placeholder tensor<\/p>\n<\/blockquote>\n\n<p>Here is my model:<\/p>\n\n<pre><code>def create_base_network(input_shape, out_dims):\n    model = Sequential()\n    model.add(Conv2D(32, kernel_size=(3, 3),\n                     activation='relu',\n                     input_shape=input_shape))\n    model.add(Conv2D(64, (3, 3), activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Dropout(0.25))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(out_dims, activation='linear'))\n\n    return model\n\ninput_shape=(28,28,3)\nanchor_in = Input(shape=input_shape)\npos_in = Input(shape=input_shape)\nneg_in = Input(shape=input_shape)\n\nbase_network = create_base_network(input_shape, 128)\nanchor_out = base_network(anchor_in)\npos_out = base_network(pos_in)\nneg_out = base_network(neg_in)\n\nmerged = concatenate([anchor_out, pos_out, neg_out], axis=-1)\n\nmodel = Model(inputs=[anchor_in, pos_in, neg_in], outputs=merged)    \n<\/code><\/pre>\n\n<p>I then try to get the output from the sequential model using:<\/p>\n\n<pre><code>seq_fun = K.function([model.layers[0].input, model.layers[1].input, model.layers[2].input], [model.layers[3].get_output_at(0)])\nseq_output = seq_fun([a, p, n])[0]\n<\/code><\/pre>\n\n<p>The input to this is coming from a generator in form of a numpy array each with the required shape. The error message then is:<\/p>\n\n<pre><code>InvalidArgumentError: You must feed a value for placeholder tensor 'conv2d_1_input' with dtype float and shape [?,28,28,3]\n [[{{node conv2d_1_input}} = Placeholder[dtype=DT_FLOAT, shape=[?,28,28,3], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n [[{{node dense_2\/BiasAdd\/_175}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_102_dense_2\/BiasAdd\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>I don't really know what's going on.<\/p>\n","answers":[{"AnswerId":"53266992","CreationDate":"2018-11-12T17:12:25.983","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The sequential model you have created have 4 output nodes. The one with index zero, i.e. <code>get_output_at(0)<\/code>, is the result of feeding it directly and the other three are its outputs when it is fed using one of the input layers you have defined. Obviously the first one is not connected to the input layers you have defined, therefore you get the error:<\/p>\n\n<blockquote>\n  <p>You must feed a value for placeholder tensor ...<\/p>\n<\/blockquote>\n\n<p>So you need to specify other three output nodes (index 1, 2 or 3) as the output of your custom function:<\/p>\n\n<pre><code>seq_fun = K.function([model.layers[0].input, model.layers[1].input, model.layers[2].input], \n                      [model.layers[3].get_output_at(i)])  # i must be 1, 2 or 3\n<\/code><\/pre>\n\n<p>As a side note, you can define the custom function more succinctly using <code>inputs<\/code> attribute of the model:<\/p>\n\n<pre><code>seq_fun = K.function(model.inputs, [model.layers[3].get_output_at(i)])\n<\/code><\/pre>\n"},{"AnswerId":"57333067","CreationDate":"2019-08-02T20:28:29.570","ParentId":null,"OwnerUserId":"11875974","Title":null,"Body":"<p>reinstall tensorflow==1.11.0 and keras==2.1.2 works for me. <\/p>\n"}]}
+{"QuestionId":53266303,"AnswerCount":0,"Tags":"<python-3.x><keras><sentiment-analysis>","CreationDate":"2018-11-12T16:27:37.540","AcceptedAnswerId":null,"OwnerUserId":6413657.0,"Title":"sentiment classification using keras","Body":"<p>I very new to deep learning classification. I have reviews data with the label(pos, neg) and I 'm trying to classify the data using keras. here is my code:<\/p>\n\n<pre><code>from sklearn.model_selection import train_test_split\nfrom sklearn.feature_extraction.text import CountVectorizer\n\nsentences_train, sentences_test, y_train, y_test = \n     train_test_split(review_list2, label_list, test_size=0.25, random_state=1000)\n\nvectorizer = CountVectorizer()\nvectorizer.fit(sentences_train)\n\nX_train = vectorizer.transform(sentences_train)\nX_test  = vectorizer.transform(sentences_test)\n\n#build the model \n\nfrom keras.models import Sequential\nfrom keras import layers\n\ninput_dim = X_train.shape[1]  \n\nmodel = Sequential()\nmodel.add(layers.Dense(8, input_dim=input_dim, activation='relu'))\nmodel.add(layers.Dense(2, activation='softmax'))\nmodel.compile(loss='binary_crossentropy', \n              optimizer='adam',  metrics=['accuracy'])\n\n\n model.summary()\n history = model.fit(X_train, y_train, epochs=100, verbose=False, validation_data=(X_test, y_test), batch_size=10)\n<\/code><\/pre>\n\n<p>I got an error:<\/p>\n\n<pre><code>   AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-52-34c39f53e335&gt; in &lt;module&gt;\n----&gt; 1 history = model.fit(X_train, y_train, epochs=100, verbose=False, validation_data=(X_test, y_test), batch_size=10)\n\nd:\\py-ver35\\lib\\site-packages\\keras\\engine\\training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n    950             sample_weight=sample_weight,\n    951             class_weight=class_weight,\n--&gt; 952             batch_size=batch_size)\n    953         # Prepare validation data.\n    954         do_validation = False\n\nd:\\py-ver35\\lib\\site-packages\\keras\\engine\\training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    787                 feed_output_shapes,\n    788                 check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 789                 exception_prefix='target')\n    790 \n    791             # Generate sample-wise weight values given the `sample_weight` and\n\nd:\\py-ver35\\lib\\site-packages\\keras\\engine\\training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n     90         data = data.values if data.__class__.__name__ == 'DataFrame' else data\n     91         data = [data]\n---&gt; 92     data = [standardize_single_array(x) for x in data]\n     93 \n     94     if len(data) != len(names):\n\nd:\\py-ver35\\lib\\site-packages\\keras\\engine\\training_utils.py in &lt;listcomp&gt;(.0)\n     90         data = data.values if data.__class__.__name__ == 'DataFrame' else data\n     91         data = [data]\n---&gt; 92     data = [standardize_single_array(x) for x in data]\n     93 \n     94     if len(data) != len(names):\n\nd:\\py-ver35\\lib\\site-packages\\keras\\engine\\training_utils.py in standardize_single_array(x)\n     25                 'Got tensor with shape: %s' % str(shape))\n     26         return x\n---&gt; 27     elif x.ndim == 1:\n     28         x = np.expand_dims(x, 1)\n     29     return x\n\nAttributeError: 'str' object has no attribute 'ndim' \n<\/code><\/pre>\n\n<p>I have tried every solution mention related to this error but still cannot fix.Any help? thanks in Advance<\/p>\n","answers":[]}
+{"QuestionId":53266350,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-11-12T16:29:59.890","AcceptedAnswerId":53266932.0,"OwnerUserId":46190.0,"Title":"How to tell PyTorch to not use the GPU?","Body":"<p>I want to do some timing comparisons between CPU &amp; GPU as well as some profiling and would like to know if there's a way to tell PyTorch to not use the GPU and instead use the CPU only? I realize I could install another CPU-only PyTorch, but hoping there's an easier way.<\/p>\n","answers":[{"AnswerId":"53266932","CreationDate":"2018-11-12T17:07:29.927","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>you can just set <code>CUDA_VISIBLE_DEVICES<\/code> variable to empty via shell before running your torch code. <\/p>\n\n<p><code>export CUDA_VISIBLE_DEVICES=\"\"<\/code> <\/p>\n\n<p>should tell torch that there are no GPUs.<\/p>\n\n<p><code>export CUDA_VISIBLE_DEVICES=\"0\"<\/code> will tell it to use only one GPU (the one with id 0) and so on.<\/p>\n"},{"AnswerId":"53267643","CreationDate":"2018-11-12T17:58:42.283","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>I just wanted to add that it is also possible to do so within the PyTorch Code:<\/p>\n\n<p>Here is a small example taken from the <a href=\"https:\/\/pytorch.org\/blog\/pytorch-0_4_0-migration-guide\/#writing-device-agnostic-code\" rel=\"nofollow noreferrer\">PyTorch Migration Guide for 0.4.0<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># at beginning of the script\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n\n...\n\n# then whenever you get a new Tensor or Module\n# this won't copy if they are already on the desired device\ninput = data.to(device)\nmodel = MyModule(...).to(device)\n<\/code><\/pre>\n\n<p><em>I think the example is pretty self-explaining. But if there are any questions just ask! <br>One big advantage is when using this syntax like in the example above is, that you can create code which runs on CPU if no GPU is available but also on GPU without changing a single line.<\/em><\/p>\n\n<p>Instead of using the <em>if-statement<\/em> with <code>torch.cuda.is_available()<\/code> you can also just set the device to <em>CPU<\/em> like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>device = torch.device(\"cpu\")\n<\/code><\/pre>\n\n<p>Further you can create tensors on the desired <em>device<\/em> using the <code>device<\/code> flag:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mytensor = torch.rand(5, 5, device=device)\n<\/code><\/pre>\n\n<p>This will create a tensor directly on the <code>device<\/code> you specified previously.<\/p>\n\n<p>I want to point out, that you can switch between <em>CPU<\/em> and <em>GPU<\/em>  using this syntax, but also between different <em>GPUs<\/em>.<\/p>\n\n<p>I hope this is helpful!<\/p>\n"}]}
+{"QuestionId":53266491,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><mnist>","CreationDate":"2018-11-12T16:39:21.617","AcceptedAnswerId":53294036.0,"OwnerUserId":9727212.0,"Title":"Input numerical arrays instead of images into Keras\/TF CNN","Body":"<p>I have been building some variations of CNN's off of Keras\/Tensorflow examples that use the MNIST data images (ubyte files) for feature extraction. My eventual goal is to do a similar thing but with a collection (~10000) 2D FFT arrays of signal data that I have made (n x m ~ 1000 x 50)(32 bite float data)<\/p>\n\n<p>I have been looking for an example that uses something other than image files but can not seem to find any.<\/p>\n\n<p>My questions are: Is this possible to do without converting them to images. Can the dataset be exported to a pickle or some other file that I could input? Whats the best way to achieve this?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53294036","CreationDate":"2018-11-14T06:03:21.293","ParentId":null,"OwnerUserId":"10619045","Title":null,"Body":"<p>Yes, you can use CNN for data other than images like sequential\/time-series data(1D convolution but you can use 2D convolution as well).<\/p>\n\n<p>CNN does its job pretty good for these types of data.<\/p>\n\n<p>You should provide your input as an image matrix i.e a window on which CNN can perform convolution on.<\/p>\n\n<p>And you can store those input matrices\/window in a numpy array and then load those file and train your CNN on it.<\/p>\n"}]}
+{"QuestionId":53266608,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-12T16:46:54.580","AcceptedAnswerId":null,"OwnerUserId":1376574.0,"Title":"Multivariate data for LSTM in Keras","Body":"<p>Please help to understand how to prepare multivariate data for LSTM in Keras.\nExample - there is three features data  <\/p>\n\n<pre><code>x1 y1 z1\nx2 y2 z2\nx3 y3 z3\nx4 y4 z4\nx5 y5 z5\n<\/code><\/pre>\n\n<p>(numbers is chronology - 1 is earliest and 5 is latests and so on)\nFor LSTM in Keras I need to transform this data to 3 dimensional tensor form\n(samples, timesteps, features). If I want 3 timesteps which of transformation is correct<\/p>\n\n<p>1<\/p>\n\n<pre><code>[\n[[x1,y1,z1],[x2,y2,z2],[x3,y3,z3]],\n[[x2,y2,z2],[x3,y3,z3],[x4,y4,z4]],\n[[x3,y3,z3],[x4,y4,z4],[x5,y5,z5]]\n]\n<\/code><\/pre>\n\n<p>Or<\/p>\n\n<p>2<\/p>\n\n<pre><code>[\n[[x1,x2,x3],[y1,y2,y3],[z1,z2,z3]],\n[[x2,x3,x4],[y2,y3,y4],[z2,z3,z4]],\n[[x3,x4,x5],[y3,y4,y5],[z3,z4,z5]]\n]\n<\/code><\/pre>\n\n<p>?<\/p>\n\n<p>Thanks for any help<\/p>\n","answers":[]}
+{"QuestionId":53266771,"AnswerCount":1,"Tags":"<python><keras><lstm><word-embedding>","CreationDate":"2018-11-12T16:55:53.830","AcceptedAnswerId":null,"OwnerUserId":10640491.0,"Title":"How to load pre-trained word embeddings in keras and output different classes","Body":"<p>I want to compare two LSTMs, one trained on Wikipedia data and one on my own. I have a problem when designing the first model. I defined two word embeddings vocabularies, one that is only built on wikipedia data and one that is also trained on my corpus. I would like to define a first LSTM that has an embedding layer (which loads the pre-trained weights from Wikipedia), but the output should be different than its vocabulary size(10000), in fact I would like to output as many classes as the ones contained into the other vocabulary (50000). This is what I have now:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(vocab_size_wikipedia, embedding_size, input_length=55, weights=[pretrained_weights_wikipedia])) \nmodel.add(Bidirectional(LSTM(units=embedding_size)))\nmodel.add(Dense(vocab_size, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy',\n          optimizer = RMSprop(lr=0.0005),\n          metrics=['accuracy'])\n\nmodel.fit(np.array(X_train), np.array(y_train), epochs=10, validation_data=(np.array(X_val), np.array(y_val)))\n<\/code><\/pre>\n\n<p>Here my variables and shapes:<\/p>\n\n<pre><code>shape of pretrained_weights_wikipedia = (10000, 100)\nvocab_size = 50000\nembedding_size = 100\nvocab_size_wikipedia = 10000\n\nX_train.shape() = (1600,55)\ny_train.shape() = (1600,50000)\nX_train.shape() = (400,55)\nX_train.shape() = (400,50000)\nX_train.shape() = (200,55)\nX_train.shape() = (200,50000) #the labels are padded \n<\/code><\/pre>\n\n<p>Thank you for your help!<\/p>\n","answers":[{"AnswerId":"58990280","CreationDate":"2019-11-22T08:29:57.803","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>You want to build a model in which you want the pretrained weights with different vocabulary size (here 10000) and finetune it for your own corpus with different vocabulary size (here 50000).<\/p>\n\n<p>Because the problem is symmetric, the mechanism that encodes the first sentence should be reused (weights and all) to encode the second sentence. Here we use a shared layer to encode the inputs.<\/p>\n\n<p>Click <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">here<\/a> for more information about <code>shared layers<\/code>.<\/p>\n\n<pre><code>input = Input(shape=55)\n\nemb1 = Embedding(vocab_size_wikipedia, embedding_size, weights=pretrained_weights_wikipedia, trainable=False)(input)\nemb2 = Embedding(50000, embedding_size)(input)\n\nbi = Bidirectional(LSTM(embedding_size))\n\nx1 = bi(emb1)\nx2 = bi(emb2)\n\ndense = Dense(vocab_size, activation='softmax')\n\nop1 = dense(x1)\nop2 = dense(x2)\n\nmodel = Model(inputs=[input1, input2], outputs=[op1, op2])\nmodel.compile(loss='categorical_crossentropy',\n          optimizer = RMSprop(lr=0.0005),\n          metrics=['accuracy'])\nmodel.summary()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53266795,"AnswerCount":0,"Tags":"<python><tensorflow><scikit-learn><svm>","CreationDate":"2018-11-12T16:57:59.200","AcceptedAnswerId":null,"OwnerUserId":10641797.0,"Title":"ValueError: X.shape[1] = 128 should be equal to 512, the number of features at training time","Body":"<p>Currently i am working on facial recognition and i am having the following issue with the given code.<\/p>\n\n<pre><code>for i in range(nrof_faces):\n    emb_array = np.zeros((1, embedding_size))\n\n    bb[i][0] = det[i][0]\n    bb[i][1] = det[i][1]    \n    bb[i][2] = det[i][2]\n    bb[i][3] = det[i][3]\n\n    # inner exception\n    if bb[i][0] &lt;= 0 or bb[i][1] &lt;= 0 or bb[i][2] &gt;= len(frame[0]) or bb[i] \n        [3] &gt;= len(frame):\n        print('Face is very close!')\n        continue\n\n    cropped.append(frame[bb[i][1]:bb[i][3], bb[i][0]:bb[i][2], :])\n    cropped[i] = facenet.flip(cropped[i], False)\n    scaled.append(misc.imresize(cropped[i], (image_size, image_size), \n             interp='bilinear'))\n    scaled[i] = cv2.resize(scaled[i], (input_image_size,input_image_size),\n                       interpolation=cv2.INTER_CUBIC)\n    scaled[i] = facenet.prewhiten(scaled[i]) \n<\/code><\/pre>\n\n<p>scaled_reshape.append(scaled[i].reshape(-1,input_image_size,input_image_size,3))<\/p>\n\n<pre><code>    feed_dict = {images_placeholder: scaled_reshape[i], phase_train_placeholder: False}\n    emb_array[0, :] = sess.run(embeddings, feed_dict=feed_dict)\n    predictions = model.predict_proba(emb_array)\n    print(predictions)\n<\/code><\/pre>\n\n<p>And it is giving me following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"F:\\std\\programs\\python\\Camera\\Facenet-Real-time-face-recognition-using-deep-learning-Tensorflow\\test_video.py\", line 107, in &lt;module&gt;\n  predictions = model.predict_proba(emb_array)\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\sklearn\\svm\\base.py\", line 613, in _predict_proba\n  X = self._validate_for_predict(X)\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\sklearn\\svm\\base.py\", line 478, in _validate_for_predict\n  (n_features, self.shape_fit_[1]))\nValueError: X.shape[1] = 128 should be equal to 512, the number of features at training time\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53267183,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras><deep-learning>","CreationDate":"2018-11-12T17:26:12.817","AcceptedAnswerId":null,"OwnerUserId":10237680.0,"Title":"Keras Model: Same array that is used for model.fit is not being processed in model.predict","Body":"<p>I have a model:<\/p>\n\n<pre><code>model.add(Dense(16, input_dim = X.shape[1], activation = 'tanh'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(8, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(4, activation = 'tanh'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(2, activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['mae'])\n<\/code><\/pre>\n\n<p>And during Model.evaluvate it works just fine with 'X' s input:<\/p>\n\n<pre><code>history = model.fit(X, Y, validation_split=0.2, epochs=10, callbacks=   [PrintDot()], batch_size=10, verbose=0)\n<\/code><\/pre>\n\n<p>But during prediction as I use X[1] it throws an error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected dense_8_input to have shape (500,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>But X[1].Shape is (500,):<\/p>\n\n<pre><code>X[1].shape\n--&gt; (500,)\n<\/code><\/pre>\n\n<p>How can I mend this error, any help appreciated<\/p>\n","answers":[{"AnswerId":"53305038","CreationDate":"2018-11-14T16:45:09.890","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>Keras <code>model.predict<\/code> expects to receive input of <code>(amount_of_items, features)<\/code>. <br><\/p>\n\n<p>So even when attempting to predict a single sample, you must reshape it to <code>(1, features)<\/code> , and in your case, <code>(1, 500)<\/code>.<\/p>\n"}]}
+{"QuestionId":53268085,"AnswerCount":0,"Tags":"<python><tensorflow><keras><vgg-net>","CreationDate":"2018-11-12T18:31:47.837","AcceptedAnswerId":null,"OwnerUserId":8737789.0,"Title":"VGG16 Keras Getting very low accuracy(~2%), while simple LeNet achieves up to 98%","Body":"<p>I have been trying to train a VGG16 net in keras with a data set of about 2000 images. The problem occurs when consistently after each epoch I achieve a very low accuracy of around 2 percent, and it remains and fluctuates around that number. But when I train it with a simple leNet, it achieves a high accuracy of around 99 percent. I do not know what the problem is. Is there an inherent flaw in my code? Does more depth sometimes negatively affect your accuracy? Both pieces of code are attached below. Sorry, I do not have a screen shot of the training process.<\/p>\n\n<p>VGG16 Code:<\/p>\n\n<pre><code>from keras.layers import Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense\nfrom keras.models import Sequential\nclassifier = Sequential()\n\nclassifier.add(Conv2D(64, 3, 3, input_shape = (224,224, 3), activation = 'relu'))\nclassifier.add(Conv2D(64, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(128, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(128, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(256, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(256, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(256, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(Conv2D(512, 3, 3, activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Flatten())\nclassifier.add(Dense(4096, activation = 'relu'))\nclassifier.add(Dense(4096, activation = 'relu'))\nclassifier.add(Dense(1000, activation = 'relu'))\nclassifier.add(Dense(output_dim = 30, activation = 'softmax'))\n\nclassifier.compile(optimizer = 'adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True,\n    validation_split=0.25)\n\ntrain_generator = train_datagen.flow_from_directory(\n    '\/home\/aslt\/signData\/top30Classes',\n    target_size=(224, 224),\n    batch_size= 15,\n    class_mode='categorical',\n    subset='training')\n\nvalidation_generator = train_datagen.flow_from_directory(\n    '\/home\/aslt\/signData\/top30Classes',\n    target_size=(224, 224),\n    batch_size= 15,\n    class_mode='categorical',\n    subset='validation')\n\nfilePath=\"\/home\/aslt\/modelFiles\/callbacksVGGNetWeights.hdf5\"\ncheckpoint = ModelCheckpoint(filePath, monitor='val_acc', verbose=1,     save_best_only=True, mode='max')\nclassifier.fit_generator(\n    train_generator,\n    steps_per_epoch = train_generator.samples,\n    validation_data = validation_generator, \n    validation_steps = validation_generator.samples,\n    epochs = 10,\n    callbacks = [checkpoint])\n\n\n\nclassifier.save('aslModelVGG.h5')\n<\/code><\/pre>\n\n<p>LeNet Code:<\/p>\n\n<pre><code>import keras\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.layers import Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense\nfrom keras.models import Sequential\nclassifier = Sequential()\n\nclassifier.add(Conv2D(32, 3, 3, input_shape = (64,64, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2,2)))\nclassifier.add(Flatten())\nclassifier.add(Dense(1000, activation = 'relu'))\nclassifier.add(Dense(output_dim = 30, activation = 'softmax'))\n\nclassifier.compile(optimizer = 'adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,\n    shear_range=0.2,\n    zoom_range=0.2,\n    horizontal_flip=True,\n    validation_split=0.25)\n\ntrain_generator = train_datagen.flow_from_directory(\n    '\/home\/aslt\/signData\/top30Classes',\n    target_size=(64, 64),\n    batch_size= 15,\n    class_mode='categorical',\n    subset='training')\n\nvalidation_generator = train_datagen.flow_from_directory(\n    '\/home\/aslt\/signData\/top30Classes',\n    target_size=(64, 64),\n    batch_size= 15,\n    class_mode='categorical',\n    subset='validation')\nfilepath=\"\/home\/aslt\/modelFiles\"\ncheckpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1,     save_best_only=True, mode='max')\nclassifier.fit_generator(\n    train_generator,\n    steps_per_epoch = train_generator.samples,\n    validation_data = validation_generator, \n    validation_steps = validation_generator.samples,\n    epochs = 10,\n    callbacks = [checkpoint])\n\n\n\nclassifier.save('aslModelLeNet.h5')\n<\/code><\/pre>\n\n<p>Also, to be clear I am using a dataset from the hand speak dictionary. I split the videos into images. The goal is to be able to recognize American Sign Language words through image classification.<\/p>\n","answers":[]}
+{"QuestionId":53268374,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-12T18:53:29.310","AcceptedAnswerId":null,"OwnerUserId":3141840.0,"Title":"Concatenate layer in Keras makes fitting fail","Body":"<p>Whenever I concatenate the outputs of two layers (for example, because I want to use softmax on some outputs and another activation function on the rest), the network always fails to learn.<\/p>\n\n<p>This is some example code to demonstrate the problem:<\/p>\n\n<pre><code>from tensorflow.keras.layers import Lambda, Input, Dense, Concatenate, Dropout, Reshape, \\\n                                    Conv2D, Flatten, MaxPooling2D\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.datasets import mnist\nfrom tensorflow.keras.losses import mse, categorical_crossentropy, binary_crossentropy\nfrom tensorflow.keras.utils import plot_model, to_categorical\nfrom tensorflow.keras import backend as K\nfrom tensorflow.keras import optimizers\n\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport argparse\nimport os\nimport pygameVisualise as pyvis\n\n# MNIST dataset\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n\nno_cls = max(y_train)+1\nwidth = 20\n\nextra_dims = True\n\nimage_size = x_train.shape[1]\noriginal_dim = image_size * image_size\nx_train = np.reshape(x_train, [-1, original_dim])\nx_test = np.reshape(x_test, [-1, original_dim])\nx_train = x_train.astype('float32') \/ 255\nx_test = x_test.astype('float32') \/ 255\ny_train = to_categorical(y_train, num_classes=width if extra_dims else no_cls)\ny_test = to_categorical(y_test, num_classes=width if extra_dims else no_cls)\n\nhidden_dim = 512\nbatch_sz = 256\neps = 10\n\nins = Input(shape=(original_dim,))\nx = Dense(hidden_dim)(ins)\ncls_pred = Dense(no_cls, activation=\"softmax\")(x)\nother    = Dense(width-no_cls)(x)\nouts = Concatenate()([cls_pred, other])\n\nencoder = Model(ins, outs if extra_dims else cls_pred, name=\"encoder\")\nencoder.summary()\n\ndef cust_loss_fn(y_true, y_pred):\n    return categorical_crossentropy(y_true[:no_cls], y_pred[:no_cls])\n\noptimiser = optimizers.SGD(lr=0.003, clipvalue=0.1)\nencoder.compile(optimizer=optimiser, loss=cust_loss_fn,\n                metrics=[\"accuracy\"])\n\nencoder.fit(x_train, y_train,\n            batch_size=batch_sz,\n            epochs=eps,\n            validation_data=(x_test, y_test))\n\nscore = encoder.evaluate(x_test, y_test)\nprint(score)\n\nprint(encoder.predict(x_train[0:10]))\n<\/code><\/pre>\n\n<p>With <code>extra_dims = False<\/code>, i.e. no concatenate layer, the network will consistently reach 88% accuracy in the 10 epochs. When it is <code>True<\/code>, the network will stay at around 8% accuracy and the loss will not drop at all during training.<\/p>\n\n<p>Am I doing something wrong?<\/p>\n","answers":[]}
+{"QuestionId":53268442,"AnswerCount":3,"Tags":"<python><python-3.x><computer-vision><pytorch><google-colaboratory>","CreationDate":"2018-11-12T18:58:23.223","AcceptedAnswerId":null,"OwnerUserId":1269281.0,"Title":"pytorch RuntimeError: CUDA error: device-side assert triggered","Body":"<p>I've a notebook on google colab that fails with following error <\/p>\n\n<pre><code>---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in fit(epochs, model, loss_func, opt, data, callbacks, metrics)\n     93         exception = e\n---&gt; 94         raise e\n     95     finally: cb_handler.on_train_end(exception)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in fit(epochs, model, loss_func, opt, data, callbacks, metrics)\n     83                 xb, yb = cb_handler.on_batch_begin(xb, yb)\n---&gt; 84                 loss = loss_batch(model, xb, yb, loss_func, opt, cb_handler)\n     85                 if cb_handler.on_batch_end(loss): break\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in loss_batch(model, xb, yb, loss_func, opt, cb_handler)\n     24     if opt is not None:\n---&gt; 25         loss = cb_handler.on_backward_begin(loss)\n     26         loss.backward()\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/callback.py in on_backward_begin(self, loss)\n    223         for cb in self.callbacks:\n--&gt; 224             a = cb.on_backward_begin(**self.state_dict)\n    225             if a is not None: self.state_dict['last_loss'] = a\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in on_backward_begin(self, smooth_loss, **kwargs)\n    266         if self.pbar is not None and hasattr(self.pbar,'child'):\n--&gt; 267             self.pbar.child.comment = f'{smooth_loss:.4f}'\n    268 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/tensor.py in __format__(self, format_spec)\n    377         if self.dim() == 0:\n--&gt; 378             return self.item().__format__(format_spec)\n    379         return object.__format__(self, format_spec)\n\nRuntimeError: CUDA error: device-side assert triggered\n\nDuring handling of the above exception, another exception occurred:\n\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-33-dd390b1c8108&gt; in &lt;module&gt;()\n----&gt; 1 lr_find(learn)\n      2 learn.recorder.plot()\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/train.py in lr_find(learn, start_lr, end_lr, num_it, stop_div, **kwargs)\n     26     cb = LRFinder(learn, start_lr, end_lr, num_it, stop_div)\n     27     a = int(np.ceil(num_it\/len(learn.data.train_dl)))\n---&gt; 28     learn.fit(a, start_lr, callbacks=[cb], **kwargs)\n     29 \n     30 def to_fp16(learn:Learner, loss_scale:float=512., flat_master:bool=False)-&gt;Learner:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in fit(self, epochs, lr, wd, callbacks)\n    160         callbacks = [cb(self) for cb in self.callback_fns] + listify(callbacks)\n    161         fit(epochs, self.model, self.loss_func, opt=self.opt, data=self.data, metrics=self.metrics,\n--&gt; 162             callbacks=self.callbacks+callbacks)\n    163 \n    164     def create_opt(self, lr:Floats, wd:Floats=0.)-&gt;None:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in fit(epochs, model, loss_func, opt, data, callbacks, metrics)\n     93         exception = e\n     94         raise e\n---&gt; 95     finally: cb_handler.on_train_end(exception)\n     96 \n     97 loss_func_name2activ = {'cross_entropy_loss': partial(F.softmax, dim=1), 'nll_loss': torch.exp, 'poisson_nll_loss': torch.exp,\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/callback.py in on_train_end(self, exception)\n    254     def on_train_end(self, exception:Union[bool,Exception])-&gt;None:\n    255         \"Handle end of training, `exception` is an `Exception` or False if no exceptions during training.\"\n--&gt; 256         self('train_end', exception=exception)\n    257 \n    258 class AverageMetric(Callback):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/callback.py in __call__(self, cb_name, call_mets, **kwargs)\n    185         \"Call through to all of the `CallbakHandler` functions.\"\n    186         if call_mets: [getattr(met, f'on_{cb_name}')(**self.state_dict, **kwargs) for met in self.metrics]\n--&gt; 187         return [getattr(cb, f'on_{cb_name}')(**self.state_dict, **kwargs) for cb in self.callbacks]\n    188 \n    189     def on_train_begin(self, epochs:int, pbar:PBar, metrics:MetricFuncList)-&gt;None:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/callback.py in &lt;listcomp&gt;(.0)\n    185         \"Call through to all of the `CallbakHandler` functions.\"\n    186         if call_mets: [getattr(met, f'on_{cb_name}')(**self.state_dict, **kwargs) for met in self.metrics]\n--&gt; 187         return [getattr(cb, f'on_{cb_name}')(**self.state_dict, **kwargs) for cb in self.callbacks]\n    188 \n    189     def on_train_begin(self, epochs:int, pbar:PBar, metrics:MetricFuncList)-&gt;None:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/callbacks\/lr_finder.py in on_train_end(self, **kwargs)\n     45         # restore the valid_dl we turned of on `__init__`\n     46         self.data.valid_dl = self.valid_dl\n---&gt; 47         self.learn.load('tmp')\n     48         if hasattr(self.learn.model, 'reset'): self.learn.model.reset()\n     49         print('LR Finder complete, type {learner_name}.recorder.plot() to see the graph.')\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/fastai\/basic_train.py in load(self, name, device)\n    202         \"Load model `name` from `self.model_dir` using `device`, defaulting to `self.data.device`.\"\n    203         if device is None: device = self.data.device\n--&gt; 204         self.model.load_state_dict(torch.load(self.path\/self.model_dir\/f'{name}.pth', map_location=device))\n    205         return self\n    206 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in load(f, map_location, pickle_module)\n    356         f = open(f, 'rb')\n    357     try:\n--&gt; 358         return _load(f, map_location, pickle_module)\n    359     finally:\n    360         if new_fd:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in _load(f, map_location, pickle_module)\n    527     unpickler = pickle_module.Unpickler(f)\n    528     unpickler.persistent_load = persistent_load\n--&gt; 529     result = unpickler.load()\n    530 \n    531     deserialized_storage_keys = pickle_module.load(f)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in persistent_load(saved_id)\n    493             if root_key not in deserialized_objects:\n    494                 deserialized_objects[root_key] = restore_location(\n--&gt; 495                     data_type(size), location)\n    496             storage = deserialized_objects[root_key]\n    497             if view_metadata is not None:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in restore_location(storage, location)\n    376     elif isinstance(map_location, torch.device):\n    377         def restore_location(storage, location):\n--&gt; 378             return default_restore_location(storage, str(map_location))\n    379     else:\n    380         def restore_location(storage, location):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in default_restore_location(storage, location)\n    102 def default_restore_location(storage, location):\n    103     for _, _, fn in _package_registry:\n--&gt; 104         result = fn(storage, location)\n    105         if result is not None:\n    106             return result\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py in _cuda_deserialize(obj, location)\n     84                                'to an existing device.'.format(\n     85                                    device, torch.cuda.device_count()))\n---&gt; 86         return obj.cuda(device)\n     87 \n     88 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/_utils.py in _cuda(self, device, non_blocking, **kwargs)\n     74         else:\n     75             new_type = getattr(torch.cuda, self.__class__.__name__)\n---&gt; 76             return new_type(self.size()).copy_(self, non_blocking)\n     77 \n     78 \n\nRuntimeError: cuda runtime error (59) : device-side assert triggered at \/pytorch\/aten\/src\/THC\/generic\/THCTensorCopy.cpp:20\n<\/code><\/pre>\n\n<p>There is no information about the real cause, I tried to get the stack trace by forcing cuda to run on one gpu (as suggested <a href=\"https:\/\/stackoverflow.com\/questions\/51691563\/cuda-runtime-error-59-device-side-assert-triggered\">here<\/a>) using a cell like this<\/p>\n\n<pre><code>!export CUDA_LAUNCH_BLOCKING=1\n<\/code><\/pre>\n\n<p>But this does not seem to work, still having the same error with. <\/p>\n\n<p>Is there another way that works with Google Colab?  <\/p>\n","answers":[{"AnswerId":"58383294","CreationDate":"2019-10-14T19:56:48.147","ParentId":null,"OwnerUserId":"3614578","Title":null,"Body":"<p>Switch Hardware Accelerator Type to \"None\" under <code>Runtime-&gt;Change Runtime Type<\/code> . This should give you a more meaningful error message.<\/p>\n"},{"AnswerId":"55532605","CreationDate":"2019-04-05T09:47:01.233","ParentId":null,"OwnerUserId":"5473777","Title":null,"Body":"<p>Be sure that your targets values starts from zero to number of classes - 1. Ex: you have 100 classification class so your target should be from 0 to 99 <\/p>\n"},{"AnswerId":"53325110","CreationDate":"2018-11-15T17:39:59.497","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p><code>!export FOO=blah<\/code> is usually not useful to run in a notebook because <code>!<\/code> means run the following command in a <em>sub<\/em>-shell, so the effect of the statement is gone by the time the <code>!<\/code> returns.\nYou might have more success by storing your python code in a file and then executing that file in a subshell:<\/p>\n\n<p>In one cell:<\/p>\n\n<pre><code>%%writefile foo.py\n[...your code...]\n<\/code><\/pre>\n\n<p>In the next cell:<\/p>\n\n<pre><code>!export CUDA_LAUNCH_BLOCKING=1; python3 foo.py\n<\/code><\/pre>\n\n<p>(or s\/python3\/python2\/ if you're writing py2)<\/p>\n"}]}
+{"QuestionId":53268818,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-11-12T19:26:30.517","AcceptedAnswerId":53268998.0,"OwnerUserId":9933449.0,"Title":"Installing tensorflow gpu in anaconda 5.3","Body":"<p>I have recently installed <code>anaconda 5.3<\/code> and it came with <code>python 3.7<\/code> preinstalled. Now when I checked tensorflow website it's says tensorflow does not support 3.7 only 3.6.<\/p>\n\n<blockquote>\n  <p>Tensorflow Requires Python 3.4, 3.5, or 3.6<\/p>\n<\/blockquote>\n\n<p>What can I do now to have <code>tensorflow gpu<\/code> for my pc. Any help is appreciated. Is there any workaround?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53268998","CreationDate":"2018-11-12T19:40:08.910","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You should create a <strong>new conda virtual environment<\/strong> with python 3.6, then install <code>tensorflow<\/code> into that.\nAt the creation of the new <code>env<\/code> you can freely choose the python version you want to use.\nThis is one of the essence of <code>conda<\/code> and virtual <code>envs<\/code> in general.<\/p>\n\n<p><strong>A bit more details:<\/strong><\/p>\n\n<p>Inside of the <a href=\"https:\/\/anaconda.org\/anaconda\/anaconda-navigator\" rel=\"nofollow noreferrer\">anaconda-navigator<\/a> you can choose <strong>Environments<\/strong>, then choose <strong>create<\/strong>. Here you can give the new environment name and the  <strong>packages basis<\/strong> i.e. <code>Python<\/code> or <code>R<\/code> and the versions you want to use it for the environment. Then you have to install your custom packages beside of the default ones. Here you can install tensorflow-gpu too.\nFor installing packages you can choose <code>Anaconda navigator GUI<\/code> or inside of the Anaconda command shell with <a href=\"https:\/\/conda.io\/docs\/user-guide\/tasks\/manage-pkgs.html\" rel=\"nofollow noreferrer\">conda<\/a> <code>conda install &lt;package&gt;<\/code> or in the Anaconda command shell <code>pip install &lt;package&gt;<\/code>. \nIn general you should prefer Anaconda package management i.e. GUI or conda to the package versions be consistent, what conda manages well. In some cases -choose always the package developer's suggestion- however, you may choose pip install.<\/p>\n"}]}
+{"QuestionId":53269358,"AnswerCount":1,"Tags":"<python><tensorflow><bazel>","CreationDate":"2018-11-12T20:08:11.203","AcceptedAnswerId":53269359.0,"OwnerUserId":1953800.0,"Title":"Multiple definition of '***' while building Tensorflow","Body":"<p>I recently upgrade to Bazel <code>0.19.0<\/code> and get the following error while trying to build Tensorflow on Ubuntu 16.04:<\/p>\n\n<pre><code>tensorflow multiple definition of 'pb_ostream_from_buffer'\n...\ncollect2: error: ld returned 1 exit status\n<\/code><\/pre>\n\n<p>where the <code>...<\/code> represents more errors of the form <code>multiple definition of '***'<\/code><\/p>\n","answers":[{"AnswerId":"53269359","CreationDate":"2018-11-12T20:08:11.203","ParentId":null,"OwnerUserId":"1953800","Title":null,"Body":"<p>Bazel <code>0.19.0<\/code> no longer reads from <code>tools\/bazel.rc<\/code>. One option in that file (<code>--define=grpc_no_ares=true<\/code>) is needed.<\/p>\n\n<p>You can instead include it on the CLI:<\/p>\n\n<pre><code>bazel build --define=grpc_no_ares=true --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>This problem is detailed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23402\" rel=\"nofollow noreferrer\">this GitHub issue<\/a>.<\/p>\n"}]}
+{"QuestionId":53269744,"AnswerCount":1,"Tags":"<keras><autoencoder><maximize><loss>","CreationDate":"2018-11-12T20:38:43.933","AcceptedAnswerId":53695812.0,"OwnerUserId":6431410.0,"Title":"Minimizing and maximizing the loss","Body":"<p>I would like to train an autoencoder in such a way that the reconstruction error will be low on some observations, and high on the others.<\/p>\n\n<pre><code>from keras.model import Sequential\nfrom keras.layers import Dense\nimport keras.backend as K\n\ndef l1Loss(y_true, y_pred):\n    return K.mean(K.abs(y_true - y_pred))\n\nmodel = Sequential()\nmodel.add(Dense(5, input_dim=10, activation='relu'))\nmodel.add(Dense(10, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss=l1Loss)\n\nfor i in range(1000):\n    model.train_on_batch(x_good, x_good) # minimize on low\n    model.train_on_batch(x_bad, x_bad, ???) # need to maximize this part, so that mse(x_bad, x_bad_reconstructed is high)\n<\/code><\/pre>\n\n<p>I saw something about replacing <code>???<\/code> with <code>sample_weight=-np.ones(batch_size)<\/code>, but I have no idea if this is fitting for my goal.<\/p>\n","answers":[{"AnswerId":"53695812","CreationDate":"2018-12-09T19:19:55.953","ParentId":null,"OwnerUserId":"9296295","Title":null,"Body":"<p>If you set sample weight to negative numbers, then minimizing it would in fact lead to maximization of its absolute value.<\/p>\n"}]}
+{"QuestionId":53269883,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-11-12T20:49:41.720","AcceptedAnswerId":null,"OwnerUserId":10364969.0,"Title":"Anaconda Tensorflow python","Body":"<p>I'm trying to run my code in Anaconda but I'm getting this error:<\/p>\n\n<pre><code>ImportError: Traceback (most recent call last):\n  File \"C:\\Users\\MHD\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\MHD\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\MHD\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\MHD\\Anaconda3\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\MHD\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n<\/code><\/pre>\n\n<p>and I don\u2019t understand why.<\/p>\n","answers":[]}
+{"QuestionId":53270187,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-12T21:12:22.620","AcceptedAnswerId":null,"OwnerUserId":10047819.0,"Title":"Is Tenserflow the right tool for my problem?","Body":"<p>I have the following data:<\/p>\n\n<ul>\n<li>Teacher with the subjects he teaches in the school<\/li>\n<li>School subjects with the respective areas<\/li>\n<li>Available rooms in which classes *are held<\/li>\n<\/ul>\n\n<p>The next step is to transfer the data to TensorFlow to automatically create a timetable. <\/p>\n\n<p>My questions:<\/p>\n\n<ol>\n<li>Is TensorFlow too big for such a task?<\/li>\n<li>Should this be worked out differently (functions) (Switch Case etc.)?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53270417","CreationDate":"2018-11-12T21:32:53.830","ParentId":null,"OwnerUserId":"1153180","Title":null,"Body":"<p>TensorFlow is the correct tool if you want to write code that trains a deep neural network on a large amount of historical data to answer a straightforward question about some input (e.g., is this a photo of a cat?). It has no ability to produce class-teacher-room timetables, and attempting to write code to do this yourself will be extremely difficult.<\/p>\n\n<p>If you want to construct such timetables automatically, I would suggest just <a href=\"https:\/\/www.google.com\/search?q=school+timetable+software\" rel=\"nofollow noreferrer\">searching for dedicated software<\/a> that does this.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":53270378,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-11-12T21:29:28.610","AcceptedAnswerId":56516518.0,"OwnerUserId":5100295.0,"Title":"What are collective ops in Tensorflow?","Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/distribute\/CollectiveAllReduceStrategy\" rel=\"nofollow noreferrer\">CollectiveAllReduce documentation<\/a> mentions 'collective ops':<\/p>\n\n<blockquote>\n  <p>It is similar to the MirroredStrategy but it uses collective ops for reduction.<\/p>\n<\/blockquote>\n\n<p>The question is simple, what are these?<\/p>\n","answers":[{"AnswerId":"56516518","CreationDate":"2019-06-09T16:47:32.887","ParentId":null,"OwnerUserId":"3210627","Title":null,"Body":"<p>Even though this is a bit of an old question, i thought i'd might as well answer. <\/p>\n\n<p>When it comes to mirroring strategies, Tensorflow (2.0) has 2 types, the MirroredStrategy and the MultiWorkerMirroredStrategy. The MirrorStrategy mirrors the variables on each replica - where a single replica is created per GPU on the machine. On the the hand, the MultiWorkerMirroredStrategy copies the variables on all the workers in a cluster. This is why the multi-worker one would need the TF_CONFIG environment variable setup. <\/p>\n\n<p>As per the documentation, CollectiveOps help keeping the variables in sync between the devices. These ops perform gather, broadcast, reduce and other functionalities, collectively across the different workers.<\/p>\n"}]}
+{"QuestionId":53271529,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-12T23:21:19.717","AcceptedAnswerId":null,"OwnerUserId":155137.0,"Title":"Multiplex between Iterators in TensorFlow","Body":"<p><strong>What I need, functionally:<\/strong> My dataset is partitioned in blocks, and each block sits in a binary file. \nI have an algorithm that also operates on blocks to reduce computational complexity, and then merges the results together, after visiting all blocks. It's important to have a single minibatch of data to originate from a single block, and to know which block exactly to be able to pass some parameters into the graph specific to that particular block. The next iteration, when starting again at block 0, the next minibatch from all blocks should be used. Blocks can have <strong>non-equal lengths<\/strong> and should be forever repeating.<\/p>\n\n<p><strong>My current solution:<\/strong> Currently, I create a <code>tf.Iterator<\/code> per block (i.e.: per file), created with a <code>tf.data.FixedLengthRecordDataset<\/code>:<\/p>\n\n<pre><code>\/\/ for every file:\nds = tf.dataFixedLengthRecordDataset(...)\nds = ds.repeat()\nds = ds.batch(...)\nds = ds.map(...)\nds = ds.prefetch(buffer_size=1)\n\nit = ds.make_one_shot_iterator()\n<\/code><\/pre>\n\n<p>Then I have an \"master\" iterator that multiplexes between the file-level iterators. This is done through:<\/p>\n\n<pre><code>itr_handle = tf.placeholder(tf.string, shape=())\nmaster_itr = tf.data.Iterator.from_string_handle(itr_handle, output_types)\nmaster_next = master_itr.get_next()\n<\/code><\/pre>\n\n<p>So each time the graph is executed, I pass into the placeholder the string handle of the respective iterator I want to use for this execution. This way every file-level iterator still <strong>has its own state<\/strong>; so when the same block-file is asked for the next minibatch, it returns effectively the next minibatch, instead of reopening the file, and simply returning the first minibatch again.<\/p>\n\n<p><strong>The problem:<\/strong> Creating the file-level iterators is <em>slow<\/em>. It takes at least 200ms to create an Iterator per file. The dataset I use can easily contain up to 100 block-files, which causes TensorFlow\/Python to be sitting there making these Iterator objects and graph nodes for 20 seconds, not actually processing any data.<\/p>\n\n<p><strong>Question:<\/strong><\/p>\n\n<ol>\n<li>Another approach to tackle this problem with only 1 iterator for example?<\/li>\n<li>If not, how to speed up Iterator creation?<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":53271754,"AnswerCount":1,"Tags":"<keras><dropout>","CreationDate":"2018-11-12T23:49:00.597","AcceptedAnswerId":null,"OwnerUserId":10476359.0,"Title":"How can I use dropout in keras","Body":"<p>An error occurred while cnn modeling.\nWhen using dropout, the following error message occurs.<\/p>\n\n<p>this is error message<\/p>\n\n<hr>\n\n<pre><code>UnboundLocalError: local variable 'a' referenced before assignment\n<\/code><\/pre>\n\n<hr>\n\n<p>model<\/p>\n\n<pre><code>def getModel(input_shape,filter_size=32,pool_size=(2,2),dropout=0.2): \n\nmodel = Sequential()\nmodel.add(Conv2D(16, (3, 3), input_shape=input_shape, activation='elu', kernel_initializer=\"he_normal\", padding='same', kernel_regularizer=regularizers.l2(0.01)))\n<\/code><\/pre>\n\n<p>I want to use dropout after maxpooling<\/p>\n\n<pre><code>model.add(MaxPooling2D(pool_size=pool_size))\nmodel.add(Dropout(dropout))\n\nmodel.add(Conv2D(32, (2, 2), kernel_initializer=\"he_normal\", padding='same',strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\n\nmodel.add(Conv2D(32, (2, 3), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\n\nmodel.add(Conv2D(16, (2, 2), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\n\nmodel.add(Conv2D(32, (2, 3), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\n\nmodel.add(Conv2D(32, (2, 2), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\n<\/code><\/pre>\n\n<p>this is flatten area<\/p>\n\n<pre><code>model.add(Flatten())\nmodel.add(Dense(126, kernel_initializer=\"glorot_normal\" ,kernel_regularizer=regularizers.l2(0.01)))\nmodel.add(Activation('tanh'))\nmodel.add(Dense(classes))\nmodel.add(Activation('sigmoid'))\n<\/code><\/pre>\n\n<p>complile<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy',\n              optimizer='adadelta',  #SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\n              metrics=['accuracy'])\nreturn model\n<\/code><\/pre>\n\n<p>model fit<\/p>\n\n<pre><code>np.random.seed(42)\nhist = model.fit(X_train, Y_train, batch_size = batch_size, epochs = epochs, verbose = 1, validation_split = .2)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53274573","CreationDate":"2018-11-13T05:46:51.347","ParentId":null,"OwnerUserId":"7458186","Title":null,"Body":"<p>I couldn't figure out what is 'a' here and hence the error,but I think following code should help:<\/p>\n\n<pre><code>model.add(Conv2D(32, (2, 2), kernel_initializer=\"he_normal\",padding='same',strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\nmodel.add(Conv2D(32, (2, 3), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(32, (2, 3), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\nmodel.add(Conv2D(32, (2, 2), kernel_initializer=\"he_normal\", padding='same', strides=1, kernel_regularizer=regularizers.l2(0.02)))\nmodel.add(Activation('elu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53271865,"AnswerCount":6,"Tags":"<tensorflow><keras><anaconda>","CreationDate":"2018-11-13T00:01:34.537","AcceptedAnswerId":53274728.0,"OwnerUserId":6773806.0,"Title":"Anaconda prompt crashes as soon as I activate tensorflow env","Body":"<p>I have just installed Anaconda 3.7 in Windows 10. Then I have created a new env for\ntensorflow and installed it there. It got installed without any problem. Then I\nused the command \nconda install -c conda-forge keras \nto install Keras. While Keras installation was running, Anaconda Prompt crashed suddenly. I restarted it and I tried to activate my tensorflow env; but as soon as I try to activate it, Anaconda Prompt crashes!! Please take a look at my screenshot. How can I fix this? Thank you very much for your support.\nFerari<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NMhLD.png\" rel=\"noreferrer\">Anaconda Prompt Crashes<\/a><\/p>\n","answers":[{"AnswerId":"55612436","CreationDate":"2019-04-10T12:21:16.863","ParentId":null,"OwnerUserId":"8741314","Title":null,"Body":"<p>when you open Anaconda command prompt just press ctrl+c and press Y to avoid the crash for now and then start again with the task of setting the environment for tensorflow...\nthis worked for me..<\/p>\n"},{"AnswerId":"53818083","CreationDate":"2018-12-17T15:17:43.100","ParentId":null,"OwnerUserId":"6835092","Title":null,"Body":"<p>I was also facing the same error and i have resolved it by re creating the conda environment. It is happening due to version mismatch between couple of packages. Just delete the conda environment and re-create the environment but this time do not add tensorflow GPU package instead just add keras-gpu, it will take care everything.<\/p>\n\n<p>I did the following things:<\/p>\n\n<ol>\n<li><p>deleted the same env<\/p><\/li>\n<li><p>created the same environment using  <\/p><\/li>\n<\/ol>\n\n<p><code>conda create -n myEnv python==3.6<\/code><\/p>\n\n<ol start=\"3\">\n<li>install keras gpu <\/li>\n<\/ol>\n\n<p><code>conda install -c conda-forge keras-gpu<\/code><\/p>\n\n<p>now it will install necessary packages and then you can activate the environment and use it<\/p>\n"},{"AnswerId":"53384999","CreationDate":"2018-11-20T01:33:04.473","ParentId":null,"OwnerUserId":"9842544","Title":null,"Body":"<p>Same problem here..<\/p>\n\n<p>I just have installed tensorflow-gpu in tf-gpu environment using <\/p>\n\n<p>conda install tensorflow-gpu<\/p>\n\n<p>It has successfully installed and works perfectly. But after installation whenever I am going to use environment by this command<\/p>\n\n<p>activate tf-gpu <\/p>\n\n<p>it shows the following<a href=\"https:\/\/i.stack.imgur.com\/NkV0u.jpg\" rel=\"nofollow noreferrer\">enter image description here<\/a> output and close the command prompt soon. I am using anaconda for python 3.6.<\/p>\n"},{"AnswerId":"53274728","CreationDate":"2018-11-13T06:02:01.570","ParentId":null,"OwnerUserId":"9427337","Title":null,"Body":"<p>The problem could be due to the version of tensorflow - tensorboard mismatch. When you give the command <em>conda install -c conda-forge keras<\/em> for installing keras, the tensorflow and tensorboard versions gets changed. <\/p>\n\n<p>I tried the following steps and it worked fine for me.<\/p>\n\n<ul>\n<li>conda create -n tf python=3.6<\/li>\n<li>activate tf<\/li>\n<li>conda install keras<\/li>\n<\/ul>\n\n<p>Installing keras will automatically install tensorflow.<\/p>\n"},{"AnswerId":"53282004","CreationDate":"2018-11-13T13:23:36.737","ParentId":null,"OwnerUserId":"1967571","Title":null,"Body":"<p>I just had an issue with the same symptom. Instead of printing the error message, the activation of the conda environment exited the console application.<\/p>\n\n<p>The scripts that run on activation can be found in <code>.\/etc\/conda\/activate.d<\/code> inside the environment. The path to the environments can be found using <code>conda env list<\/code>. Check these scripts for any 'exit' commands, commenting them out if necessary. This should let you at least see the error message.<\/p>\n\n<p>In my case the issue was caused by packages requiring MSVC compiler, and Visual Studio studio was not installed.<\/p>\n"},{"AnswerId":"54074660","CreationDate":"2019-01-07T12:41:24.737","ParentId":null,"OwnerUserId":"9493599","Title":null,"Body":"<p>I encountered the same problem. Solved by creating a new env and using <\/p>\n\n<p><code>conda install -c hesi_m keras<\/code><\/p>\n\n<p>This will install the latest versions of <code>Keras<\/code> and <code>Tensorflow<\/code>.<\/p>\n\n<p>Turns out you will most likely get an outdated version of keras if you installed it with <code>conda install -c conda-forge keras<\/code>. <\/p>\n"}]}
+{"QuestionId":53272171,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-11-13T00:44:34.527","AcceptedAnswerId":null,"OwnerUserId":3040335.0,"Title":"How to normalize LSTM input data in Keras with BatchNormalization","Body":"<p>I have my neural network constructed like the following:<\/p>\n\n<pre><code>tempIn = Input(shape = (None, 4))\ntempModel = LSTM(data.xRnnLosFeatures)(tempIn)\ntempModel = BatchNormalization()(tempModel)\ntempModel = Activation('tanh')(tempModel)\ntempModel = Dropout(0.5)(tempModel)\ntempModel = Dense(1)(tempModel)\nmodel = Model(inputs=tempIn, outputs=tempModel)\n<\/code><\/pre>\n\n<p>However, I keep getting a very large error if I do not manually normalize my input data before I feed this network. What is a way to normalize my input data correctly. I tried to add another one before LSTM layer but that doesn't work. Thanks!<\/p>\n","answers":[{"AnswerId":"54876341","CreationDate":"2019-02-25T23:33:44.463","ParentId":null,"OwnerUserId":"776515","Title":null,"Body":"<pre><code>from sklearn.preprocessing import MinMaxScaler\n\nscalerx = MinMaxScaler( feature_range=(0, 1) )  # To normalize the inputs\nscalery = MinMaxScaler( feature_range=(0, 1) )  # To normalize the outputs\n\ndatax = scalerx.fit_transform( data['inputs'] ) # Assuming a dictionary with inputs\ndatay = scalerx.fit_transform( data['outputs'] ) # and outputs\n\nmodel.fit( datax, datay )\n<\/code><\/pre>\n\n<p>After you run the model, you will get values in the range [0,1], and you need to revert the normalization to make sense of them:<\/p>\n\n<pre><code>y_hat = model.predict( some_data_normalized_with_scalerx )\n\ny_hat_denorm = scalery.inverse_transform( y_hat )\n<\/code><\/pre>\n\n<p>and <code>y_hat_denorm<\/code> will have the same units from the beginning, i.e. those from <code>data['outputs']<\/code>, used to create <code>scalery<\/code>.<\/p>\n"},{"AnswerId":"53981867","CreationDate":"2018-12-30T22:15:03.727","ParentId":null,"OwnerUserId":"8355515","Title":null,"Body":"<p>You can use the <strong>keras<\/strong> <code>normalise<\/code> function or alternatively you could use <strong>scikit-learn<\/strong> <code>preprocessing<\/code> function.<\/p>\n"}]}
+{"QuestionId":53272508,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T01:35:08.450","AcceptedAnswerId":null,"OwnerUserId":8383470.0,"Title":"inception v3 using tf.data?","Body":"<p>I'm using a bit of code that is derived from inception v3 as distributed by the Google folks, but it's now complaining that the queue runners used to read the data are deprecated (tf.train.string_input_producer in image_processing.py, and similar).  Apparently I'm supposed to switch to tf.data for this kind of stuff.<\/p>\n\n<p>Unfortunately, the documentation on tf.data isn't doing much to relieve my concern that I've got too much data to fit in memory, especially given that I want to batch it in a reusable way, etc. I'm confident that the tf.data stuff <em>can<\/em> do this; I just don't know <em>how<\/em> to do it. Can anyone point me to a full example of code that uses tf.data to deal with batches of data that won't all fit in memory?  Ideally, it would simply be an updated version of the inception-v3 code, but I'd be happy to try and work with anything.  Thanks!<\/p>\n","answers":[{"AnswerId":"53452436","CreationDate":"2018-11-23T20:09:09.057","ParentId":null,"OwnerUserId":"8383470","Title":null,"Body":"<p>Well, I eventually got this working.  The various documents referenced in the comment on my question had what I needed, and I gradually figured out which parameters passed to queuerunners corresponded to which parameters in the tf.data stuff.<\/p>\n\n<p>There was one gotcha that took a while for me to sort out.  In the inception implementation, the number of examples used for validation is rounded up to be a multiple of the batch size; presumably the validation set is reshuffled and some examples are used more than once.  (This does not strike me as great practice, but generally the number of validation instances is way larger than the batch size, so only a relative few are double counted.)<\/p>\n\n<p>In the tf.data stuff, enabling shuffling and reuse is a separate thing and I didn't do it on the validation data.  Then things broke because there weren't enough unique validation instances, and I had to track that down.<\/p>\n\n<p>I hope this helps the next person with this issue.  Unfortunately, my code has drifted quite far from Inception v3 and I doubt that it would be helpful for me to post my modification.  Thanks!<\/p>\n"}]}
+{"QuestionId":53272519,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-13T01:35:49.217","AcceptedAnswerId":null,"OwnerUserId":2359772.0,"Title":"Custom C++ extension : torch\/torch.h not found","Body":"<p>I am following <a href=\"https:\/\/pytorch.org\/tutorials\/advanced\/cpp_extension.html\" rel=\"nofollow noreferrer\">this tutorial<\/a> to create a C++ extension for Pytorch. My C++ code is giving following error :<\/p>\n\n<pre><code>test.cpp:3:10: fatal error: torch\/torch.h: No such file or directory\n #include &lt;torch\/torch.h&gt;\n<\/code><\/pre>\n\n<p>How to get <code>torch.h<\/code> header file ? Is there some <code>pytorch-dev<\/code> version?<\/p>\n","answers":[{"AnswerId":"53462489","CreationDate":"2018-11-24T21:24:58.643","ParentId":null,"OwnerUserId":"84278","Title":null,"Body":"<p>You most likely need to install the c++ version of torch. <\/p>\n\n<p>If you head over to www.pytorch.org they will guide you to which download is appropriate for you. Just make sure to select the c++ version. <\/p>\n\n<p>Hopefully that works! <\/p>\n"}]}
+{"QuestionId":53272609,"AnswerCount":0,"Tags":"<c++><linux><caffe>","CreationDate":"2018-11-13T01:47:56.793","AcceptedAnswerId":null,"OwnerUserId":8018636.0,"Title":"protoc shared libraries error in ubuntu 16.04","Body":"<p>I tried to install caffe in Ubuntu 16.04 and I encountered the error when I typed <code>make all<\/code> in the root caffe directory.<\/p>\n\n<pre><code>protoc: error while loading shared libraries: libprotoc.so.9: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>I tried <code>protoc --version<\/code>, the error is still:<\/p>\n\n<pre><code>protoc: error while loading shared libraries: libprotoc.so.9: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>So I used <code>locate libprotoc.so.9<\/code>, and found that the libprotoc exists in <\/p>\n\n<pre><code>\/usr\/lib\/x86_64-linux-gnu\/libprotoc.so.9\n\/usr\/lib\/x86_64-linux-gnu\/libprotoc.so.9.0.1\n<\/code><\/pre>\n\n<p>So I used the following methods try to let system knows where is libprotoc,<\/p>\n\n<p>1)<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/usr\/lib\n<\/code><\/pre>\n\n<p>2) go to <code>\/etc<\/code> and add the path by <\/p>\n\n<pre><code>vim ld.so.conf \n<\/code><\/pre>\n\n<p>Although I used <code>sudo ldconfig<\/code> everytime I made any change to the files, still the error exists. I also found that when I tried <code>gedit filename<\/code>, the similar error will show up as :<\/p>\n\n<pre><code>gedit: error while loading shared libraries: libmirprotobuf.so.3: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>I really don't know how to deal with this shared libraries problem. <\/p>\n\n<p>And idea would be helpful ! <\/p>\n\n<p><strong>UPDATE:<\/strong>\nso I remove the original protobuf by <code>sudo uninstall protobuf<\/code>,however a new error shows up when I use <code>make all<\/code> in caffe file, <\/p>\n\n<pre><code>.build_release\/src\/caffe\/proto\/caffe.pb.h:10:40: fatal error: google\/protobuf\/port_def.inc: No such file or directory\n<\/code><\/pre>\n\n<p>And <code>port_def.inc<\/code> actually exists at <code>~\/google\/protobuf\/port_def.inc<\/code>.Seems I cannot let caffe know the path, that's the problem. <\/p>\n\n<p>But how to solve it ? <\/p>\n","answers":[]}
+{"QuestionId":53272743,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-13T02:07:24.417","AcceptedAnswerId":null,"OwnerUserId":10244760.0,"Title":"How to handle single output multiple losses in keras?","Body":"<p>My model has a single output. But I want to separate the loss into 3 different components so that I can track the progress of each. Is there a way to do this with keras?<\/p>\n\n<p>Maybe I could define the same loss components as metrics but is there a more elegant way?<\/p>\n","answers":[{"AnswerId":"53276127","CreationDate":"2018-11-13T07:45:40.713","ParentId":null,"OwnerUserId":"5108062","Title":null,"Body":"<p>You can define the losses as Keras layers and then you can add all of your losses and metrics (if you want) manually.<\/p>\n\n<p><a href=\"https:\/\/gaborvecsei.wordpress.com\/2018\/10\/15\/implement-loss-functions-inside-keras-models\/\" rel=\"nofollow noreferrer\">You can see a complete tutorial about this topic here<\/a><\/p>\n\n<p>TL;DR:<\/p>\n\n<ul>\n<li>Define layers where you calculate the loss<\/li>\n<li>Write your own compile() function where you add the optimizer, and the losses and the metrics<\/li>\n<li>At model.compile(optimizer=\"adam\", loss=...) add <code>None<\/code> as loss<\/li>\n<\/ul>\n\n<p>This is how manually adding the losses looks like in code:<\/p>\n\n<pre><code>loss_layer_names = {\"my_loss\", ...}\n\n# Adding losses\nfor name in loss_layer_names:\n    layer = model.get_layer(name)\n    loss = (tf.reduce_mean(layer.output, keepdims=True))\n    model.add_loss(loss)\n\n# Adding metrics\nfor name in loss_layer_names:\n    layer = model.get_layer(name)\n    loss = (tf.reduce_mean(layer.output, keepdims=True))\n    model.metrics_names.append(name)\n    model.metrics_tensors.append(loss)\n\nmodel.compile(optimizer=\"adam\", loss=[None] * len(model.outputs))\n<\/code><\/pre>\n\n<p>Where <code>model<\/code> is a Keras model<\/p>\n"}]}
+{"QuestionId":53272771,"AnswerCount":1,"Tags":"<tensorflow><while-loop><tensorflow-probability>","CreationDate":"2018-11-13T02:10:43.240","AcceptedAnswerId":null,"OwnerUserId":1189112.0,"Title":"tensorflow.while_loop with sample from tensorflow_distribution in condition","Body":"<p>I'm struggling to make the following code work. It keeps performing Bernoulli trials until one succeeds.<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\n\ndef geometric(p):\n    def cond(_):\n        return tf.equal(1, tfd.Bernoulli(p).sample())\n    def body(t):\n        return tf.add(t, 1)\n    return tf.while_loop(\n        cond, # name is automatically generated\n        body,\n        [tf.constant(0)]\n    )\n\nwith tf.Session() as sess:\n    acc = sess.run(geometric(0.001))\n    print(acc)\n<\/code><\/pre>\n\n<p>The above code prints out values from <code>0<\/code> to <code>3<\/code>, which does not makes sense. I'm expecting it to print something in the hundreds. Furthermore, when I change the <code>geometric<\/code> call to <code>geometric(0.000001)<\/code>, I still get the same result.<\/p>\n\n<p>Can anyone please point out what is wrong with my code above?<\/p>\n","answers":[{"AnswerId":"53305521","CreationDate":"2018-11-14T17:13:59.513","ParentId":null,"OwnerUserId":"187421","Title":null,"Body":"<p>Do you have the while condition reversed? I think you want while <em>not<\/em> 1 (while 0)<\/p>\n"}]}
+{"QuestionId":53272808,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2018-11-13T02:16:31.577","AcceptedAnswerId":null,"OwnerUserId":9450018.0,"Title":"TensorFlow, Keras - TypeError: Value passed to parameter 'labels' has DataType float32 not in list of allowed values: int32, int64","Body":"<p>The code below is giving <code>TypeError: Value passed to parameter 'labels' has DataType float32 not in list of allowed values: int32, int64<\/code>. <\/p>\n\n<pre><code>import tensorflow as tf\n#tf.enable_eager_execution()\nvocab_size=100\nembedding_dim=256\nseq_len=10\nmodel=tf.keras.Sequential()\nmodel.add(tf.keras.layers.Embedding(vocab_size, embedding_dim, input_length=seq_len))\nmodel.add(tf.keras.layers.Dense(vocab_size))\nmodel.compile(optimizer = tf.train.GradientDescentOptimizer(1.0),loss = tf.losses.sparse_softmax_cross_entropy)#tf.keras.metrics.sparse_categorical_crossentropy\n<\/code><\/pre>\n\n<p>The error is avoided by any one of the following three actions:<\/p>\n\n<ol>\n<li>execute <code>tf.enable_eager_execution()<\/code> after importing tensorflow<\/li>\n<li>change <code>tf.losses.sparse_softmax_cross_entropy<\/code> to <code>tf.keras.metrics.sparse_categorical_crossentropy<\/code><\/li>\n<li>remove the <code>tf.keras.layers.Embedding<\/code> layer<\/li>\n<\/ol>\n\n<p>Any insights into this interaction between Keras Embedding layer and tf.losses will be greatly appreciated! I am using TensorFlow version 1.12.0.<\/p>\n","answers":[]}
+{"QuestionId":53273169,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-13T03:02:34.620","AcceptedAnswerId":53273714.0,"OwnerUserId":6590393.0,"Title":"How to train a CNN on an unlabeled dataset?","Body":"<p>I want to train a CNN on my unlabeled data, and from what I read on Keras\/Kaggle\/TF documentation or Reddit threads, it looks like I will have to label my dataset beforehand. Is there a way to train the CNN in an unsupervised way?<br>\nI cannot understand how to initialize <i><b>y_train<\/b><\/i> and <i><b>y_test<\/b><\/i> (where y_train and y_test represent usual meanings)<br>\nThe information about my dataset is as follows:<br><\/p>\n\n<ol>\n<li>I have 50,000 matrices of dimension 30 x 30.<\/li>\n<li>Each matrix is divided into 9 subareas (for understanding, as separated by the vertical and horizontal bars). <\/li>\n<li>A subarea is said to be <i><b>active<\/b><\/i> if it has at least one element equal to 1. If all elements for that subarea are equal to 0, the subarea is <i><b>inactive<\/b><\/i>.  <\/li>\n<li>For the first example shown below, I should get as output the names of subareas that are active, so here, (1, 4, 5, 6, 7, 9).<\/li>\n<li>If no subarea is active, as in the second example, the output should be 0.<\/li>\n<\/ol>\n\n<p>First example: Output - (1, 4, 5, 6, 7, 9)\n<a href=\"https:\/\/i.stack.imgur.com\/ll8dv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ll8dv.png\" alt=\"First example image\"><\/a><\/p>\n\n<p>Second example: Output - 0\n<a href=\"https:\/\/i.stack.imgur.com\/BONcd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BONcd.png\" alt=\"Second example image\"><\/a>\nAfter creating these matrices, I did the following:<\/p>\n\n<ol>\n<li>I put these matrices in a CSV file after reshaping them into vectors of dimension 900 x 1.<\/li>\n<li>So basically, each row in the CSV contains 900 columns with values either 0 or 1.<\/li>\n<li>The classes for my classification problem are numbers from 0-9 where 0 represents the class where no label has an active (value=1) value. <\/li>\n<\/ol>\n\n<p>For my model, I want the following:<\/p>\n\n<ul>\n<li><b>Input: <\/b> a 900 x 1 vector as described above.<\/li>\n<li><b>Output: <\/b> one of the values from 0-9,<br> where 1-9 represent the active subareas, and 0 represents no active subarea.<\/li>\n<\/ul>\n\n<p><b>What I have done:<\/b><br>\nI am able to retrieve the data from the CSV file into a data frame and split the data frame into <i><b>x_train<\/b><\/i> and <i><b>x_test<\/b><\/i>. But I am unable to understand how to set my <i><b>y_train<\/b><\/i> and <i><b>y_test<\/b><\/i> values.<br>\nMy problem seems very similar to the MNIST dataset, except I don't have the labels. Would it be possible for me to train the model without the labels?\n<br><br> My code currently looks like this:<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nfrom sklearn.model_selection import train_test_split\n\n# Read the dataset from the CSV file into a dataframe\ndf = pd.read_csv(\"bci_dataset.csv\")\n\n# Split the dataframe into training and test dataset\ntrain, test = train_test_split(df, test_size=0.2)\n\nx_train = train.iloc[:, :]\nx_test = test.iloc[:, :]\n\nprint(x_train.shape)\nprint(x_test.shape)\n<\/code><\/pre>\n\n<p>Thank you, in advance, for reading this whole thing and helping me out!<\/p>\n","answers":[{"AnswerId":"53273714","CreationDate":"2018-11-13T04:14:58.347","ParentId":null,"OwnerUserId":"10626636","Title":null,"Body":"<p>Can you tell us why you want to use a CNN specifically?  Generally neural networks are used when there's some complication involved in going from feature to output - the artificial neurons are able to learn different behavior as a result of being exposed to the ground truth (i.e., the labels).  Most of the time, the researcher using the neural network doesn't even know what features of the input data are being used by the network to come to its output conclusions.<\/p>\n\n<p>In the case you have given us, it looks a little bit more like you know what features are important (that is, the sum of a subarea has to be greater than 0 in order to be active).  The neural network wouldn't need to really learn anything in particular to do its job.  Although it doesn't seem necessary to use a neural network for this process, it does make sense for you to automate it, given the size of your input data! :)<\/p>\n\n<p>Let me know if I'm misunderstanding your situation, though?<\/p>\n\n<p>Edit: To contrast this with the MNIST dataset - so for identifying handwritten digits, there's some ambiguity that the network has to learn to deal with.  Not every kind of handwriting is going to render a 7 the same way.  A neural network is able to figure out a couple of the features of a 7 (i.e., there is a high probability that a 7 will have a diagonal line going from top-right-to-bottom-left, which, depending on how you write, could be slightly curved or offset or whatever), as well as a couple of different versions of a 7 (some people do a horizontal slash through the middle of it, other versions of a 7 don't have that slash).  The utility of a neural network here is in figuring out all that ambiguity and probabilistically classifying an input as a 7 (because it has seen previous images that it \"knows\" are 7s).  However, in your case, there's only one way for your answer to be rendered - if there's any element greater than 0 in a subarea, it's active!  So you don't need to train a network to do anything - you will just need to write some code that automates the summing of the subareas.<\/p>\n"}]}
+{"QuestionId":53273206,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><google-cloud-tpu><tpu>","CreationDate":"2018-11-13T03:08:41.657","AcceptedAnswerId":null,"OwnerUserId":229005.0,"Title":"How to manually use Google TPU with Tensorflow's Object Detection API?","Body":"<p>I've successfully trained models using Tensorflow's Object Detection API running both locally on GPU (using <code>model_main.py<\/code>) and using Google's ML Engine (both GPU and TPU). However, I can't seem to be able to use <code>model_tpu_main.py<\/code> to train a model, when running on on Google's Cloud (using a manually provisionned VM and TPU). <\/p>\n\n<p>When I launch <code>model_tpu_main.py<\/code> using something like <code>python -m object_detection.model_tpu_main --model_dir=gs:\/\/bucket\/training --tpu_zone us-central1-b --pipeline_config_path=gs:\/\/bucket\/training\/pipeline.config --job-dir gs:\/\/bucket\/training --tpu_name mytpu_name<\/code>, it gets stuck on:<\/p>\n\n<pre><code>...\nW1113 03:05:16.628712 139998232708864 variables_helper.py:144] Variable [resnet_v1_50\/fpn\/smoothing_2\/BatchNorm\/moving_mean] is not available in checkpoint\nW1113 03:05:16.629062 139998232708864 variables_helper.py:144] Variable [resnet_v1_50\/fpn\/smoothing_2\/BatchNorm\/moving_variance] is not available in checkpoint\nW1113 03:05:16.629330 139998232708864 variables_helper.py:144] Variable [resnet_v1_50\/fpn\/smoothing_2\/weights] is not available in checkpoint\n2018-11-13 03:06:08.618834: W tensorflow\/core\/distributed_runtime\/rpc\/grpc_session.cc:349] GrpcSession::ListDevices will initialize the session with an empty graph and other defaults because the session has not yet been created.\n...\n<\/code><\/pre>\n\n<p>Looking at the TPU logs, pretty much all I get is:<\/p>\n\n<pre><code>...\nStart master session b9186abfa4e15b1d with config: isolate_session_state: true A \nStart master session 48b812f9ca0d3ebf with config: isolate_session_state: true A \nStart master session 33048226cb131f4c with config: isolate_session_state: true A \nStart master session cab95e277a429f9d with config: isolate_session_state: true A \nStart master session 56b5d3296c9bfe15 with config: isolate_session_state: true A \nStart master session 3fdac64b285c365d with config: isolate_session_state: true A \nStart master session ec1fa14806ad9351 with config: isolate_session_state: true A \n...\n<\/code><\/pre>\n\n<p>Any idea what I'm doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":53273226,"AnswerCount":1,"Tags":"<python><tensorflow><image-rotation><data-augmentation>","CreationDate":"2018-11-13T03:11:27.873","AcceptedAnswerId":null,"OwnerUserId":7886651.0,"Title":"How to rotate images at different angles randomly in tensorflow","Body":"<p>I know that I can rotate images in tensorflow using <code>tf.contrib.image.rotate<\/code>. But suppose I want to apply the rotation randomly at an angle between -0.3 and 0.3 in radians as follows:<\/p>\n\n<pre><code>images = tf.contrib.image.rotate(images, tf.random_uniform(shape=[batch_size], minval=-0.3, maxval=0.3, seed=mseed), interpolation='BILINEAR')\n<\/code><\/pre>\n\n<p>So far this will work fine. But the problem arises when the batch size changes on the last iteration and I got an error. So how to fix this code and make it work in all case scenarios? Please note that the inputs images are fed using <code>tf.data.Dataset<\/code> api.<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[{"AnswerId":"53855704","CreationDate":"2018-12-19T16:47:29.297","ParentId":null,"OwnerUserId":"1477064","Title":null,"Body":"<p>You can't feed <code>tf.contrib.image.rotate<\/code> with an angles tensor.<\/p>\n\n<p>But if you inspect the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.12.0\/tensorflow\/contrib\/image\/python\/ops\/image_ops.py\" rel=\"nofollow noreferrer\">source code<\/a> you can see it just makes a bunch of argument validations, and then:<\/p>\n\n<pre><code>image_height = math_ops.cast(array_ops.shape(images)[1],\n                             dtypes.float32)[None]\nimage_width = math_ops.cast(array_ops.shape(images)[2],\n                            dtypes.float32)[None]\noutput = transform(\n    images,\n    angles_to_projective_transforms(angles, image_height, image_width),\n                                    interpolation=interpolation)\n<\/code><\/pre>\n\n<p><code>tf.contrib.image.transform()<\/code> receives a projective transform matrix.\n<code>tf.contrib.image.angles_to_projective_transforms()<\/code> generates projective transforms from the rotation angles.<\/p>\n\n<p>Both accept tensors as arguments, so you can just call the underlying functions.<\/p>\n\n<hr>\n\n<p>Here is an example using MNIST<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# load mnist\nfrom tensorflow.examples.tutorials.mnist\nimport input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot = True)\n\n# Tensorflow random angle rotation\ninput_size = mnist.train.images.shape[1]\nside_size = int(np.sqrt(input_size))\n\ndataset = tf.placeholder(tf.float32, [None, input_size])\nimages = tf.reshape(dataset, (-1, side_size, side_size, 1))\nrandom_angles = tf.random.uniform(shape = (tf.shape(images)[0], ), minval = -np\n    .pi \/ 4, maxval = np.pi \/ 4)\n\nrotated_images = tf.contrib.image.transform(\n    images,\n    tf.contrib.image.angles_to_projective_transforms(\n        random_angles, tf.cast(tf.shape(images)[1], tf.float32), tf.cast(tf\n            .shape(images)[2], tf.float32)\n    ))\n\n# Run and Print\nsess = tf.Session()\nresult = sess.run(rotated_images, feed_dict = {\n    dataset: mnist.train.images,\n})\n\noriginal = np.reshape(mnist.train.images * 255, (-1, side_size, side_size)).astype(\n    np.uint8)\nrotated = np.reshape(result * 255, (-1, side_size, side_size)).astype(np.uint8)\n\n\n# Print 10 random samples\nfig, axes = plt.subplots(2, 10, figsize = (15, 4.5))\nchoice = np.random.choice(range(len(mnist.test.labels)), 10)\nfor k in range(10):\n    axes[0][k].set_axis_off()\naxes[0][k].imshow(original[choice[k, ]], interpolation = 'nearest', \\\n    cmap = 'gray')\naxes[1][k].set_axis_off()\naxes[1][k].imshow(rotated[choice[k, ]], interpolation = 'nearest', \\\n    cmap = 'gray')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2fR5l.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2fR5l.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53273456,"AnswerCount":1,"Tags":"<python><machine-learning><keras><keras-layer><softmax>","CreationDate":"2018-11-13T03:41:44.730","AcceptedAnswerId":53279514.0,"OwnerUserId":10395181.0,"Title":"In Keras, how to apply softmax function on each row of the weight matrix?","Body":"<pre><code>from keras.models import Model\nfrom keras.models import Input\nfrom keras.layers import Dense\n\na = Input(shape=(3,))\nb = Dense(2, use_bias=False)(a)\nmodel = Model(inputs=a, outputs=b)\n<\/code><\/pre>\n\n<p>Suppose that the weights of the <code>Dense<\/code> layer in the above code is <code>[[2, 3], [3, 1], [-1, 1]]<\/code>. If we give <code>[[2, 1, 3]]<\/code> as an input to the <code>model<\/code>, then the output will be:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mVuF4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mVuF4.png\" alt=\"no softmax\"><\/a><\/p>\n\n<p>But I want to apply the softmax function to each row of the <code>Dense<\/code> layer, so that the output will be:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/N4byN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/N4byN.png\" alt=\"with softmax\"><\/a><\/p>\n\n<p>How can I do this?<\/p>\n","answers":[{"AnswerId":"53279514","CreationDate":"2018-11-13T10:58:47.750","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One way to achieve what you are looking for is to define a custom layer by subclassing the <code>Dense<\/code> layer and overriding its <code>call<\/code> method:<\/p>\n\n<pre><code>from keras import backend as K\n\nclass CustomDense(Dense):\n    def __init__(self, units, **kwargs):\n        super(CustomDense, self).__init__(units, **kwargs)\n\n    def call(self, inputs):\n        output = K.dot(inputs, K.softmax(self.kernel, axis=-1))\n        if self.use_bias:\n            output = K.bias_add(output, self.bias, data_format='channels_last')\n        if self.activation is not None:\n            output = self.activation(output)\n        return output\n<\/code><\/pre>\n\n<p>Test to make sure it works:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(CustomDense(2, use_bias=False, input_shape=(3,)))\n\nmodel.compile(loss='mse', optimizer='adam')\n\nimport numpy as np\n\nw = np.array([[2,3], [3,1], [1,-1]])\ninp = np.array([[2,1,3]])\n\nmodel.layers[0].set_weights([w])\nprint(model.predict(inp))\n\n# output\n[[4.0610714 1.9389288]]\n<\/code><\/pre>\n\n<p>Verify it using numpy:<\/p>\n\n<pre><code>soft_w = np.exp(w) \/ np.sum(np.exp(w), axis=-1, keepdims=True)\nprint(np.dot(inp, soft_w))\n\n[[4.06107115 1.93892885]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53273568,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2018-11-13T03:58:47.127","AcceptedAnswerId":null,"OwnerUserId":10550386.0,"Title":"How to Update parameters every certain steps?","Body":"<p>I'm using a tensorflow based keras to train a object detection network.<\/p>\n\n<p>Due to small GPU memory size, have to update params once per some steps to \"expand\" batch size.\nBut my implementation can't successfully converge in Mask Rcnn training or perform badly comparing normal optimizers in simple cifar10 classification:<\/p>\n\n<pre><code>class custom_SGD(Optimizer):\ndef get_updates(self, loss, params):\n    \"\"\"Main params updates operation\n    \"\"\"\n    shapes = [K.int_shape(p) for p in params]\n    if self.nesterov:\n        sum_grads = [K.zeros(shape) for shape in shapes]\n\n    moments = [K.zeros(shape) for shape in shapes]\n    sum_moments = [K.zeros(shape) for shape in shapes]\n\n    # Current gradients\n    grads = self.get_gradients(loss, params)\n    self.updates = [K.update_add(self.iterations, 1)] \n    self.weights = [self.iterations] + moments\n\n    cond1 = K.equal(self.iterations % self.steps_per_update, 0)\n\n    # Learning rate decay\n    lr = self.lr\n    if self.initial_decay &gt; 0:\n        lr = lr * (1. \/ (1. + self.decay * K.cast(self.iterations,\n                                                  K.dtype(self.decay))))\n    if not self.nesterov:\n        for p, g, m,sm in zip(params, grads,moments, sum_moments):\n            v = self.momentum*m - lr*g\n            # updates sum_moments, moments\n            self.updates.append(K.update(sm, sm+v))\n            self.updates.append(K.switch(cond1,K.update(m,sm\/float(self.steps_per_update)),m))\n\n            new_p = p + m\n            # Apply constraint\n            if getattr(p, 'constraint', None) is not None:\n                new_p = p.constraint(new_p)\n            self.updates.append(K.switch(cond1,K.update(p,new_p),p))\n            # Clear up container\n            self.updates.append(K.switch(cond1,K.update(sm,K.zeros_like(sm)),sm))\n\n    return self.updates\n<\/code><\/pre>\n\n<p>Can anybody tell me why. \nThanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53273662,"AnswerCount":1,"Tags":"<python><pytorch><autograd>","CreationDate":"2018-11-13T04:08:30.107","AcceptedAnswerId":53274732.0,"OwnerUserId":1582098.0,"Title":"Pytorch. Can autograd be used when the final tensor has more than a single value in it?","Body":"<p>Can autograd be used when the final tensor has more than a single value in it?<\/p>\n\n<p>I tried the following. <\/p>\n\n<pre><code>x = torch.tensor([4.0, 5.0], requires_grad=True)\ny = x ** 2\n\nprint(y)\n\ny.backward()\n<\/code><\/pre>\n\n<p>Throws an error<\/p>\n\n<pre><code>RuntimeError: grad can be implicitly created only for scalar outputs\n<\/code><\/pre>\n\n<p>The following however works.<\/p>\n\n<pre><code>x = torch.tensor([4.0, 5.0], requires_grad=True)\ny = x ** 2\ny = torch.sum(y)\nprint(y)\n\ny.backward()\nprint(x.grad)\n<\/code><\/pre>\n\n<p>The output is as <\/p>\n\n<pre><code>tensor(41., grad_fn=&lt;SumBackward0&gt;)\ntensor([ 8., 10.])\n<\/code><\/pre>\n\n<p>Am I missing something here or can I proceed with the assumption that autograd only works when the final tensor has a single value in it?<\/p>\n","answers":[{"AnswerId":"53274732","CreationDate":"2018-11-13T06:02:15.560","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>See <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/autograd_tutorial.html#gradients\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/autograd_tutorial.html#gradients<\/a><\/p>\n\n<p><code>y.backward()<\/code> is same as <code>y.backward(torch.tensor(1.0))<\/code><\/p>\n\n<p>Usually, the output is scalar and hence the scalar is passed to backward as a default choice. However, since your output is two dimensional you should call\n<code>y.backward(torch.tensor([1.0,1.0]))<\/code><\/p>\n\n<p>This will give expected results with <code>x.grad<\/code> being <code>tensor([ 8., 10.])<\/code><\/p>\n"}]}
+{"QuestionId":53273692,"AnswerCount":0,"Tags":"<python><tensorflow><bitwise-operators><bitwise-xor>","CreationDate":"2018-11-13T04:12:01.583","AcceptedAnswerId":null,"OwnerUserId":10643919.0,"Title":"Tensorflow cost function based on binary xor operation","Body":"<p>I am trying to create a cost function that is based on the xor operation of x &amp; y. <\/p>\n\n<p>x and y are tf.float32 tensors and both have the same shape. <\/p>\n\n<p>x --> ... --> y auto-encoder network<\/p>\n\n<p>I have tried doing the following:<\/p>\n\n<pre><code>x_value = tf.cast(x, tf.int32)\ny_value = tf.cast(y, tf.int32)\nxor_bitwise_result = tf.bitwise.bitwise_xor(x_value, y_value)   \ncost_value = tf.cast(xor_bitwise_result, tf.float32)\ncost = tf.reduce_mean(cost_value)\n<\/code><\/pre>\n\n<p>but I am getting an error:<\/p>\n\n<p><code>ValueError: No gradients provided for any variable, check your graph for ops that do not support gradient.<\/code><\/p>\n\n<p>if I try to do the following I will get the same error as well<\/p>\n\n<pre><code>cost = tf.sqrt(tf.reduce_mean(tf.square(x_value-y_value)))\n<\/code><\/pre>\n\n<p>It might be the way I am casting is not good, it is possible my understanding of the process is lacking so any pointers would be greatly appreciated. <\/p>\n","answers":[]}
+{"QuestionId":53273723,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-13T04:16:17.397","AcceptedAnswerId":null,"OwnerUserId":3253409.0,"Title":"Error: tensorflow\/contrib\/nccl\/kernels\/nccl_manager.cc:273 check failed: result==ncclSuccess (2 vs 0)system error","Body":"<p>I am trying to run distributed tensorflow code using MirrorStrategy option alongwith tensorflow estimator API and getting the error as mentioned in the title. I am using tensorflow-gpu 1.9.0.  I am following <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/distribute\" rel=\"nofollow noreferrer\">link<\/a>  for distributed tensorflow training.<\/p>\n\n<p>and getting below mentioned warning along with error: \nYou should always run with libnvidia-ml.so that is installed with your NVIDIA Display Driver. By default it's I installed in \/usr\/lib and \/usr\/lib64 . libnvidia-ml.so in GDK package is a stub library that is attached only for build purposes (e.g. machine that you build your application doesn't have to have display driver installed).<\/p>\n","answers":[{"AnswerId":"57280852","CreationDate":"2019-07-30T22:30:06.710","ParentId":null,"OwnerUserId":"6684106","Title":null,"Body":"<p>Chances are you've got the stub libraries for compilation, and your LD_LIBRARY_PATH doesn't include the path for the runtime libraries.<\/p>\n\n<p>Check your library path for \"\/usr\/local\/cuda\/lib64\/stubs\" or something similar. If it exists, you just need to put the correct location before it in your library path.<\/p>\n\n<p>Depending on the driver version you have installed, you may be able to find the libnvidia-ml.so file under \"\/usr\/lib\/nvidia-384\" or some number other than 384 that matches your nvidia driver version.<\/p>\n\n<p>You could add a line to your .bashrc file that looks something like the following:<\/p>\n\n<p><code>export LD_LIBRARY_PATH=\/usr\/lib\/nvidia-(Your driver number here):$LD_LIBRARY_PATH<\/code><\/p>\n"}]}
+{"QuestionId":53273786,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-11-13T04:24:54.273","AcceptedAnswerId":null,"OwnerUserId":10584895.0,"Title":"Having trouble with CNN prediction","Body":"<p>I am using Convolutional Neural Networking for vehicle identification, my first time. Currently, I am working with just 2 classes(bike and car). Training set: 420 car images and 825 bike images. Test set: 44 car images and 110 bike images Car and Bike images are in different format(bmp,jpg). In single prediction, I am always getting 'bike'. I have tried using the Sigmoid function in the output layer. Then I get only 'car'. My code is like following: ``<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense,Dropout\n\n\n\nclassifier = Sequential()\n\n\nclassifier.add(Conv2D(32, (3, 3), input_shape = (128, 128, 3), activation = 'relu'))\n\n\nclassifier.add(MaxPooling2D(pool_size = (3, 3)))\n\n# Adding a second convolutional layer\nclassifier.add(Conv2D(32, (3, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (3, 3)))\n\n# Step 3 - Flattening\nclassifier.add(Flatten())\n\n# Step 4 - Full connection\nclassifier.add(Dropout(0.3))\nclassifier.add(Dense(units = 128, activation = 'relu'))\nclassifier.add(Dense(units = 1, activation = 'sigmoid'))\n\n# Compiling the CNN\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\n# Part 2 - Fitting the CNN to the images\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   rotation_range= 3,\n                                   fill_mode = 'nearest',\n                                   horizontal_flip = True)\n\ntest_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   rotation_range= 3,\n                                   fill_mode = 'nearest',\n                                   horizontal_flip = True)\n\ntraining_set = train_datagen.flow_from_directory('dataset\/training_set',\n                                                 target_size = (128, 128),\n                                                 batch_size = 10,\n                                                 class_mode = 'binary')\n\ntest_set = test_datagen.flow_from_directory('dataset\/test_set',\n                                            target_size = (128, 128),\n                                            batch_size = 10,\n                                            class_mode = 'binary')\n\nclassifier.fit_generator(training_set,\n                         steps_per_epoch = 1092\/\/10,\n                         epochs = 3,\n                         validation_data = test_set,\n                         validation_steps = 20)\n\nclassifier.save(\"car_bike.h5\")\n<\/code><\/pre>\n\n<p>And I wanted to test a single image like the following:<\/p>\n\n<pre><code>test_image = image.load_img('dataset\/single_prediction\/download (3).jpg', target_size = (128, 128))\ntest_image = image.img_to_array(test_image)\ntest_image *= (1\/255.0)\ntest_image = np.expand_dims(test_image, axis = 0)\nresult = classifier.predict(test_image)\nif result[0][0] == 1:\n    prediction = 'bike'\nelse:\n    prediction = 'car'\n\nprint(\" {}\".format(prediction))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53274436","CreationDate":"2018-11-13T05:33:28.517","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>If you print your <code>result<\/code> matrix you'll see that it doesn't have only 1s and 0s but floats between these numbers. You may pick a threshold and set values that exceed it to 1 and everything else to 0.<\/p>\n"}]}
+{"QuestionId":53273828,"AnswerCount":1,"Tags":"<amazon-web-services><tensorflow><keras><gpu>","CreationDate":"2018-11-13T04:29:04.657","AcceptedAnswerId":null,"OwnerUserId":1849429.0,"Title":"Keras, GPU running out of memory","Body":"<p>I just wanted to post the details here as well. But basically I have an implementation which is running out of memory.<\/p>\n\n<p>Referenced Github Issues here:\n<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/11624\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/11624<\/a><\/p>\n\n<p>This is the error message:<\/p>\n\n<pre><code>UserWarning: Viewer requires Qt\n  warn('Viewer requires Qt')\n2018-11-12 09:30:54.179843: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2018-11-12 09:31:11.234972: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.236072: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 0 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:17.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.322354: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.323475: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 1 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:18.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.413172: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.414297: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 2 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:19.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.510326: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.511434: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 3 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:1a.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.617084: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.618204: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 4 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:1b.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.719956: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.721063: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 5 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:1c.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.825226: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.826376: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 6 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:1d.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.935858: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-12 09:31:11.936963: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 7 with properties:\nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:1e.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-12 09:31:11.945353: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0, 1, 2, 3, 4, 5, 6, 7\n2018-11-12 09:31:14.423061: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-12 09:31:14.423126: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 1 2 3 4 5 6 7\n2018-11-12 09:31:14.423139: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N Y Y Y Y Y Y Y\n2018-11-12 09:31:14.423147: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 1:   Y N Y Y Y Y Y Y\n2018-11-12 09:31:14.423155: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 2:   Y Y N Y Y Y Y Y\n2018-11-12 09:31:14.423162: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 3:   Y Y Y N Y Y Y Y\n2018-11-12 09:31:14.423169: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 4:   Y Y Y Y N Y Y Y\n2018-11-12 09:31:14.423177: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 5:   Y Y Y Y Y N Y Y\n2018-11-12 09:31:14.423186: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 6:   Y Y Y Y Y Y N Y\n2018-11-12 09:31:14.423196: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 7:   Y Y Y Y Y Y Y N\n2018-11-12 09:31:14.425010: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 10757 MB memory) -&gt; physical GPU (device: 0, name: Tesla K80, pci bus id: 0000:00:17.0, compute capability: 3.7)\n2018-11-12 09:31:14.425736: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:1 with 10757 MB memory) -&gt; physical GPU (device: 1, name: Tesla K80, pci bus id: 0000:00:18.0, compute capability: 3.7)\n2018-11-12 09:31:14.426309: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:2 with 10757 MB memory) -&gt; physical GPU (device: 2, name: Tesla K80, pci bus id: 0000:00:19.0, compute capability: 3.7)\n2018-11-12 09:31:14.426869: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:3 with 10757 MB memory) -&gt; physical GPU (device: 3, name: Tesla K80, pci bus id: 0000:00:1a.0, compute capability: 3.7)\n2018-11-12 09:31:14.427875: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:4 with 10757 MB memory) -&gt; physical GPU (device: 4, name: Tesla K80, pci bus id: 0000:00:1b.0, compute capability: 3.7)\n2018-11-12 09:31:14.428440: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:5 with 10757 MB memory) -&gt; physical GPU (device: 5, name: Tesla K80, pci bus id: 0000:00:1c.0, compute capability: 3.7)\n2018-11-12 09:31:14.428998: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:6 with 10757 MB memory) -&gt; physical GPU (device: 6, name: Tesla K80, pci bus id: 0000:00:1d.0, compute capability: 3.7)\n2018-11-12 09:31:14.429564: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:7 with 10757 MB memory) -&gt; physical GPU (device: 7, name: Tesla K80, pci bus id: 0000:00:1e.0, compute capability: 3.7)\n\/app\/networks.py:240: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(units=3, activation=\"linear\")`\n  model.add(Dense(output_dim=action_size, activation='linear'))\n2018-11-12 09:31:29.037056: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:267] Allocator (GPU_0_bfc) ran out of memory trying to allocate 3.09GiB.  Current allocation summary follows.\n2018-11-12 09:31:29.037156: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (256):   Total Chunks: 36, Chunks in use: 36. 9.0KiB allocated for chunks. 9.0KiB in use in bin. 1.8KiB client-requested in use in bin.\n2018-11-12 09:31:29.037186: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (512):   Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037207: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (1024):  Total Chunks: 1, Chunks in use: 1. 1.2KiB allocated for chunks. 1.2KiB in use in bin. 1.0KiB client-requested in use in bin.\n2018-11-12 09:31:29.037224: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (2048):  Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037248: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (4096):  Total Chunks: 2, Chunks in use: 2. 12.0KiB allocated for chunks. 12.0KiB in use in bin. 12.0KiB client-requested in use in bin.\n2018-11-12 09:31:29.037272: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (8192):  Total Chunks: 5, Chunks in use: 5. 40.0KiB allocated for chunks. 40.0KiB in use in bin. 40.0KiB client-requested in use in bin.\n2018-11-12 09:31:29.037289: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (16384):         Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037317: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (32768):         Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037336: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (65536):         Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037358: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (131072):        Total Chunks: 4, Chunks in use: 4. 544.0KiB allocated for chunks. 544.0KiB in use in bin. 544.0KiB client-requested in use in bin.\n2018-11-12 09:31:29.037439: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (262144):        Total Chunks: 1, Chunks in use: 0. 417.8KiB allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037449: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (524288):        Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037457: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (1048576):       Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037465: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (2097152):       Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037474: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (4194304):       Total Chunks: 4, Chunks in use: 4. 16.00MiB allocated for chunks. 16.00MiB in use in bin. 16.00MiB client-requested in use in bin.\n2018-11-12 09:31:29.037484: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (8388608):       Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037492: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (16777216):      Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037501: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (33554432):      Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037509: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (67108864):      Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037520: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (134217728):     Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2018-11-12 09:31:29.037530: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:597] Bin (268435456):     Total Chunks: 4, Chunks in use: 2. 8.00GiB allocated for chunks. 6.18GiB in use in bin. 6.18GiB client-requested in use in bin.\n2018-11-12 09:31:29.037540: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:613] Bin for 3.09GiB was 256.00MiB, Chunk State:\n2018-11-12 09:31:29.037553: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:619]   Size: 934.00MiB | Requested Size: 12B | in_use: 0, prev:   Size: 3.09GiB | Requested Size: 3.09GiB | in_use: 1\n2018-11-12 09:31:29.037564: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:619]   Size: 934.00MiB | Requested Size: 0B | in_use: 0, prev:   Size: 3.09GiB | Requested Size: 3.09GiB | in_use: 1\n2018-11-12 09:31:29.037574: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00000 of size 1280\n2018-11-12 09:31:29.037584: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00500 of size 256\n2018-11-12 09:31:29.037591: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00600 of size 256\n2018-11-12 09:31:29.037599: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00700 of size 256\n2018-11-12 09:31:29.037606: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00800 of size 256\n2018-11-12 09:31:29.037619: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00900 of size 256\n2018-11-12 09:31:29.037627: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00a00 of size 256\n2018-11-12 09:31:29.037633: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00b00 of size 256\n2018-11-12 09:31:29.037640: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00c00 of size 256\n2018-11-12 09:31:29.037647: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe00d00 of size 8192\n2018-11-12 09:31:29.037655: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe02d00 of size 256\n2018-11-12 09:31:29.037662: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe02e00 of size 256\n2018-11-12 09:31:29.037668: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe02f00 of size 256\n2018-11-12 09:31:29.037677: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03000 of size 256\n2018-11-12 09:31:29.037683: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03100 of size 256\n2018-11-12 09:31:29.037690: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03200 of size 256\n2018-11-12 09:31:29.037702: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03300 of size 256\n2018-11-12 09:31:29.037709: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03400 of size 256\n2018-11-12 09:31:29.037716: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03500 of size 256\n2018-11-12 09:31:29.037722: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03600 of size 256\n2018-11-12 09:31:29.037731: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03700 of size 256\n2018-11-12 09:31:29.037738: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03800 of size 256\n2018-11-12 09:31:29.037745: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03900 of size 256\n2018-11-12 09:31:29.037754: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03a00 of size 256\n2018-11-12 09:31:29.037761: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03b00 of size 256\n2018-11-12 09:31:29.037767: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03c00 of size 256\n2018-11-12 09:31:29.037774: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03d00 of size 256\n2018-11-12 09:31:29.037784: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03e00 of size 256\n2018-11-12 09:31:29.037791: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe03f00 of size 256\n2018-11-12 09:31:29.037797: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe04000 of size 256\n2018-11-12 09:31:29.037804: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe04100 of size 256\n2018-11-12 09:31:29.037812: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe04200 of size 256\n2018-11-12 09:31:29.037819: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe04300 of size 6144\n2018-11-12 09:31:29.037828: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe05b00 of size 6144\n2018-11-12 09:31:29.037835: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe07300 of size 8192\n2018-11-12 09:31:29.037841: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe09300 of size 8192\n2018-11-12 09:31:29.037848: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe0b300 of size 256\n2018-11-12 09:31:29.037857: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe0b400 of size 256\n2018-11-12 09:31:29.037864: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe0b500 of size 8192\n2018-11-12 09:31:29.037870: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe0d500 of size 8192\n2018-11-12 09:31:29.037880: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe0f500 of size 131072\n2018-11-12 09:31:29.037886: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe2f500 of size 131072\n2018-11-12 09:31:29.037893: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe4f500 of size 256\n2018-11-12 09:31:29.037903: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe4f600 of size 256\n2018-11-12 09:31:29.037910: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe4f700 of size 256\n2018-11-12 09:31:29.037916: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe4f800 of size 256\n2018-11-12 09:31:29.037929: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe4f900 of size 147456\n2018-11-12 09:31:29.037936: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1fe73900 of size 147456\n2018-11-12 09:31:29.037943: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Free  at 0x7c1fe97900 of size 427776\n2018-11-12 09:31:29.037952: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c1ff00000 of size 4194304\n2018-11-12 09:31:29.037958: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c20300000 of size 4194304\n2018-11-12 09:31:29.037967: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c20700000 of size 4194304\n2018-11-12 09:31:29.037974: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c20b00000 of size 4194304\n2018-11-12 09:31:29.037980: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7c20f00000 of size 3315597312\n2018-11-12 09:31:29.037989: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Free  at 0x7ce6900000 of size 979369984\n2018-11-12 09:31:29.037996: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Chunk at 0x7d20f00000 of size 3315597312\n2018-11-12 09:31:29.038003: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:632] Free  at 0x7de6900000 of size 979369984\n2018-11-12 09:31:29.038011: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:638]      Summary of in-use Chunks by size:\n2018-11-12 09:31:29.038020: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 36 Chunks of size 256 totalling 9.0KiB\n2018-11-12 09:31:29.038030: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 1 Chunks of size 1280 totalling 1.2KiB\n2018-11-12 09:31:29.038038: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 2 Chunks of size 6144 totalling 12.0KiB\n2018-11-12 09:31:29.038046: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 5 Chunks of size 8192 totalling 40.0KiB\n2018-11-12 09:31:29.038053: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 2 Chunks of size 131072 totalling 256.0KiB\n2018-11-12 09:31:29.038063: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 2 Chunks of size 147456 totalling 288.0KiB\n2018-11-12 09:31:29.038071: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 4 Chunks of size 4194304 totalling 16.00MiB\n2018-11-12 09:31:29.038079: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:641] 2 Chunks of size 3315597312 totalling 6.18GiB\n2018-11-12 09:31:29.038088: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:645] Sum Total of in-use chunks: 6.19GiB\n2018-11-12 09:31:29.038098: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:647] Stats:\nLimit:                 11279748301\nInUse:                  6648592640\nMaxInUse:               6653229824\nNumAllocs:                      64\nMaxAllocSize:           3315597312\n\n2018-11-12 09:31:29.038112: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:271] ***************************************___________***************************************___________\nTraceback (most recent call last):\n  File \".\/ddrqn_per.py\", line 355, in &lt;module&gt;\nState Size (4, 216, 43, 1)\n    agent.update_target_model()\n  File \".\/ddrqn_per.py\", line 153, in update_target_model\n    self.target_model.set_weights(self.model.get_weights())\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/network.py\", line 508, in set_weights\n    K.batch_set_value(tuples)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 2470, in batch_set_value\n    get_session().run(assign_ops, feed_dict=feed_dict)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1152, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1328, in _do_run\n    run_metadata)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1348, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InternalError: Dst tensor is not initialized.\n         [[{{node _arg_Placeholder_6_0_7\/_101}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_50__arg_Placeholder_6_0_7\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n         [[{{node Assign_6\/_123}} = _Recv[_start_time=0, client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_71_Assign_6\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>This one has got me stumped. Is it that GPU memory cannot actual stack, which means that the more GPU doesnt mean the more memory ?<\/p>\n\n<p>The script is running inside a docker container: FROM tensorflow\/tensorflow:latest-gpu<\/p>\n\n<p>The Docker.gpu looks like this:<\/p>\n\n<pre><code>FROM tensorflow\/tensorflow:latest-gpu\n\nADD .\/gloob-bot-model \/app\nRUN mkdir \/app\/images &amp;&amp; mkdir \/app\/models\nADD .\/recordings.zip \/app\/recordings.zip\n\nRUN apt-get update -qq &amp;&amp;\\\n    apt-get install --no-install-recommends -y \\\n    python-tk unzip\n\nRUN pip install -r \/app\/requirements.gpu.txt\n\nENV PYTHONPATH \"${PYTHONPATH}:\/app\"\n\nCMD cd \/app &amp;&amp; unzip recordings.zip &amp;&amp; python .\/ddrqn_per.py\n<\/code><\/pre>\n\n<p>I have changed my batch size to 1!. Im using an AWS p2.8xlarge. So 8 NVIDIA K80 GPUs with each 12gig<\/p>\n\n<p>My input size is like:<\/p>\n\n<pre><code>batch_size x 4 x 216 x 40 x 1\n<\/code><\/pre>\n\n<p>Im using pillow to make the images,\nThen converting them to grey scale<\/p>\n\n<pre><code>image = image.convert('L')\n<\/code><\/pre>\n\n<p>Which is then being put into an array of 4 and inserted into my model<\/p>\n\n<p>This is my model:<\/p>\n\n<pre><code>model = Sequential()\n        model.add(TimeDistributed(Conv2D(32, 8, activation='relu'), input_shape=input_shape))\n        model.add(TimeDistributed(Conv2D(64, 4, activation='relu')))\n        model.add(TimeDistributed(Conv2D(64, 3, activation='relu')))\n        model.add(TimeDistributed(Flatten()))\n\n        # Use all traces for training\n        #model.add(LSTM(512, return_sequences=True,  activation='tanh'))\n        #model.add(TimeDistributed(Dense(output_dim=action_size, activation='linear')))\n\n        # Use last trace for training\n        model.add(LSTM(512,  activation='tanh'))\n        model.add(Dense(output_dim=action_size, activation='linear'))\n\n        adam = Adam(lr=learning_rate)\n        model.compile(loss=huber_loss, optimizer=adam)\n<\/code><\/pre>\n\n<p>This is my User-Data running on initialization. I am installing NVIDIA stuff on the aws Ubuntu machine. (Not sure if i should be doing this, but it didnt work without it).<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/kevupton\/c963cd237ed8ad24b1140694fe867db2\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/kevupton\/c963cd237ed8ad24b1140694fe867db2<\/a><\/p>\n\n<p>My Config Setup:<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nsess = tf.Session(config=config)\nK.set_session(sess)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53277681","CreationDate":"2018-11-13T09:24:13.963","ParentId":null,"OwnerUserId":"1849429","Title":null,"Body":"<p>Ok the problem was this:<\/p>\n\n<pre><code>with tf.device('\/cpu:0'):\n    model = generate_model()\n<\/code><\/pre>\n\n<p>This code that got rid of the OOM error.<\/p>\n\n<p>using their documentation:\nReference: <a href=\"https:\/\/keras.io\/utils\/#multi_gpu_model\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/utils\/#multi_gpu_model<\/a><\/p>\n\n<pre><code>from keras.utils import multi_gpu_model\n\n# Instantiate the base model (or \"template\" model).\n# We recommend doing this with under a CPU device scope,\n# so that the model's weights are hosted on CPU memory.\n# Otherwise they may end up hosted on a GPU, which would\n# complicate weight sharing.\nwith tf.device('\/cpu:0'):\n    model = Xception(weights=None,\n                     input_shape=(height, width, 3),\n                     classes=num_classes)\n\n# Replicates the model on 8 GPUs.\n# This assumes that your machine has 8 available GPUs.\nparallel_model = multi_gpu_model(model, gpus=8)\nparallel_model.compile(loss='categorical_crossentropy',\n                       optimizer='rmsprop')\n<\/code><\/pre>\n\n<p>I implemented:<\/p>\n\n<pre><code>    def generate_model():\n        model = Sequential()\n        model.add(TimeDistributed(Conv2D(32, 8, activation='relu'), input_shape=input_shape))\n        model.add(TimeDistributed(Conv2D(64, 4, activation='relu')))\n        model.add(TimeDistributed(Conv2D(64, 3, activation='relu')))\n        model.add(TimeDistributed(Flatten()))\n\n        # Use all traces for training\n        #model.add(LSTM(512, return_sequences=True,  activation='tanh'))\n        #model.add(TimeDistributed(Dense(output_dim=action_size, activation='linear')))\n\n        # Use last trace for training\n        model.add(LSTM(512,  activation='tanh'))\n        model.add(Dense(output_dim=action_size, activation='linear'))\n\n        return model\n\n    if total_gpus is None:\n        model = generate_model()\n    else:\n        with tf.device('\/cpu:0'):\n            model = generate_model()\n\n        model = multi_gpu_model(model, gpus=total_gpus)\n\n    adam = Adam(lr=learning_rate)\n    model.compile(loss=huber_loss, optimizer=adam)\n<\/code><\/pre>\n\n<p><strong>EDIT!<\/strong>\nThe only problem with this method is that it is is slower than my CPU's<\/p>\n"}]}
+{"QuestionId":53274141,"AnswerCount":1,"Tags":"<python><machine-learning><keras><concatenation><lstm>","CreationDate":"2018-11-13T05:05:14.530","AcceptedAnswerId":null,"OwnerUserId":10639147.0,"Title":"Keras concatenate LSTM model with non-LSTM model","Body":"<p>I have two data. One is time series and the other contains features such as Sex, education, etc. and I want to concatenate output of LSTM model and a dense model. However, I got an error message (look at the end).<\/p>\n\n<p>This is what the data looks like:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/izC01.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/izC01.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0TXCU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0TXCU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And this is the code:<\/p>\n\n<pre><code># PAY_data net\ninput1 = Input(shape=(6,1))\npay = LSTM(10)(input1)\npay = Dense(10, activation='relu')(pay)\n\n# DEMO_data net\ninput2 = Input(shape=(5,1))\ndemo = Dense(10, activation='relu')(input2)\ndemo = Dense(10, activation='relu')(demo)\n\nmerge = concatenate([pay, demo])\n\nhidden1 = Dense(10, activation='relu')(merge)\n\noutput = Dense(1, activation='sigmoid')(merge)\nmodel = Model(inputs=[input1, input2], outputs=output)\n\nprint(model.summary())\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=    ['accuracy'])\n\nmodel.fit([PAY_data, DEMO_data], y,nb_epoch=20, batch_size=50, verbose=2, validation_split=0.2)\n<\/code><\/pre>\n\n<p>and this is the error I get:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lIqae.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lIqae.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53275765","CreationDate":"2018-11-13T07:19:09.480","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since the <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">Dense layer is applied on the last axis of its input data<\/a>, and considering that you have specified an input shape of <code>(5,1)<\/code> for your \"Demo_data net\", the output shape of this model would be <code>(None, 5, 10)<\/code> and therefore it cannot be concatenated with the output of the \"Pay_data net\" which has an output shape of <code>(None, 10)<\/code>. To resolve this, you can remove the redundant last axis from <code>PAY_data<\/code> using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.1\/reference\/generated\/numpy.squeeze.html\" rel=\"nofollow noreferrer\"><code>np.squeeze()<\/code><\/a>:<\/p>\n\n<pre><code>PAY_data = np.squeeze(PAY_data)\n<\/code><\/pre>\n\n<p>and also set the input shape accordingly:<\/p>\n\n<pre><code>input2 = Input(shape=(5,))  # now the input shape is (5,) and not (5,1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53274587,"AnswerCount":1,"Tags":"<python><gradient><pytorch><derivative>","CreationDate":"2018-11-13T05:48:37.553","AcceptedAnswerId":53278914.0,"OwnerUserId":8433424.0,"Title":"PyTorch gradient differs from manually calculated gradient","Body":"<p>I'm trying to compute the gradient of 1\/x without using Pytorch's autograd. I use the formula grad(1\/x, x) = -1\/x**2. When I compare my result with this formula to the gradient given by Pytorch's autograd, they're different.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>a = torch.tensor(np.random.randn(), dtype=dtype, requires_grad=True)\nloss = 1\/a\nloss.backward()\nprint(a.grad - (-1\/(a**2)))\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>tensor(5.9605e-08, grad_fn=&lt;ThAddBackward&gt;)\n<\/code><\/pre>\n\n<p>Can anyone explain to me what the problem is?<\/p>\n","answers":[{"AnswerId":"53278914","CreationDate":"2018-11-13T10:27:34.937","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>So I guess you expect zero as result. When you take a closer look you see that it is quite close. When deviding numbers on a binary system (computer) then you often get round-off errors.<\/p>\n\n<p>Lets take a look at your example with an additional <em>print-statement<\/em> added:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = torch.tensor(np.random.randn(), requires_grad=True)\nloss = 1\/a\nloss.backward()\nprint(a.grad, (-1\/(a**2)))\nprint(a.grad - (-1\/(a**2)))\n<\/code><\/pre>\n\n<p>The because of the random input the output is of course random too <em>(so you won't get this very same numbers, but just repeat and you will have similar examples)<\/em>, sometimes you will also get zero as result, here it is note the case:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensor(-0.9074) tensor(-0.9074, grad_fn=&lt;MulBackward&gt;)\ntensor(5.9605e-08, grad_fn=&lt;ThSubBackward&gt;)\n<\/code><\/pre>\n\n<p>You see even though both are displayed as the same number but they differ in one of the last decimal places. That is why you get this very small difference when subtracting both.<\/p>\n\n<p>This problem as a general problem of computers, some fractions just have an large or infinite number of decimal places, but your memory has not. So so they are cut off at some point.<\/p>\n\n<p>So what you experience here is actually a lack of precision. And the precision is depending on the numerical data type you are using (i.e. <code>torch.float32<\/code> or <code>torch.float64<\/code>). <\/p>\n\n<p><em>You can also take a look here more information:<br>\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Double-precision_floating-point_format\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Double-precision_floating-point_format<\/a><\/em><\/p>\n\n<hr>\n\n<p>But this is not specific to PyTorch or so, here is a Python example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print(29\/100*100)\n<\/code><\/pre>\n\n<p>Results in:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>28.999999999999996\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>As @HOANG GIANG pointed out, changing the equation to <em>-(1\/a)*(1\/a)<\/em> works well and the result is zero.\nProbably this the case because the calculation in done to calculate the gradient is very similar (or the same) to <em>-(1\/a)*(1\/a)<\/em> in this case. Therefore it shares the same round-off errors therefore the difference is zero.<\/p>\n\n<p>So then here is another more fitting example than the one above. Even though <em>-(1\/x)*(1\/x)<\/em> is mathematically equivalent to <em>-1\/x^2<\/em> it is not always the same when calculating it on the computer, depending on the value of <em>x<\/em>:<\/p>\n\n<pre><code>import numpy as np\nprint('e1 == e2','x value', '\\t'*2, 'round-off error', sep='\\t')\nprint('='*70)\nfor i in range(10):\n    x = np.random.randn()\n    e1 = -(1\/x)*(1\/x)\n    e2 = (-1\/(x**2))\n    print(e1 == e2, x, e1-e2, sep='\\t\\t')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>e1 == e2    x value                 round-off error\n======================================================================\nTrue        0.2934154339948173      0.0\nTrue        -1.2881863891014191     0.0\nTrue        1.0463038021843876      0.0\nTrue        -0.3388766143622498     0.0\nTrue        -0.6915415747192347     0.0\nFalse       1.3299049850551317      1.1102230246251565e-16\nTrue        -1.2392046539563553     0.0\nFalse       -0.42534236747121645    8.881784197001252e-16\nTrue        1.407198823994324       0.0\nFalse       -0.21798652132356966    3.552713678800501e-15\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p>Even though the <em>round-off error<\/em> seems to be a bit less (I tried different random values, and rarely more than two out of ten had a <em>round-off error<\/em>), but still there are already small differences when just calculating <strong><em>1\/x<\/em><\/strong>:<\/p>\n\n<pre><code>import numpy as np\nprint('e1 == e2','x value', '\\t'*2, 'round-off error', sep='\\t')\nprint('='*70)\nfor i in range(10):\n    x = np.random.randn()\n    # calculate 1\/x\n    result = 1\/x\n    # apply inverse function\n    reconstructed_x = 1\/result\n    # mathematically this should be the same as x\n    print(x == reconstructed_x, x, x-reconstructed_x, sep='\\t\\t')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>e1 == e2    x value             round-off error\n======================================================================\nFalse       0.9382823115235075      1.1102230246251565e-16\nTrue        -0.5081217386356917     0.0\nTrue        -0.04229436058156134    0.0\nTrue        1.1121100294357302      0.0\nFalse       0.4974618312372863      -5.551115123125783e-17\nTrue        -0.20409933212316553    0.0\nTrue        -0.6501652554924282     0.0\nTrue        -3.048057937738731      0.0\nTrue        1.6236075700470816      0.0\nTrue        0.4936926651641918      0.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":53275045,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2018-11-13T06:29:03.487","AcceptedAnswerId":53275323.0,"OwnerUserId":1773216.0,"Title":"Why does model.fit() raise ValueError with tf.train.AdamOptimizer using categorical_crossentropy loss function?","Body":"<p>I'm following the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">TensorFlow basic classification example<\/a> with the Keras API provided in the \"Getting Started\" docs. I get through the tutorial as-is just fine, but if I change the loss function from <code>sparse_categorical_crossentropy<\/code> to <code>categorical_crossentropy<\/code>, the code below:<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Flatten(input_shape=(28, 28)),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(), \n          loss='categorical_crossentropy',\n          metrics=['accuracy'])\n\nmodel.fit(train_images, train_labels, epochs=5)\n<\/code><\/pre>\n\n<p>fails during the training\/fitting step with the following error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_1 to have shape (10,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>The documentation on the loss functions doesn't delve much into expected input and output. Obviously there is a dimensionality issue here, but if any experts can give a detailed explanation, what is it about this loss function or any other loss function that raises this <code>ValueError<\/code>?<\/p>\n","answers":[{"AnswerId":"53275323","CreationDate":"2018-11-13T06:49:22.227","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>sparse_categorical_crossentropy<\/code> loss expects the provided labels to be <strong>integers<\/strong> like 0, 1, 2 and so on, where each integer indicates a particular class. For example class 0 might be dogs, class 1 might be cats and class 2 might be lions. On the other hand, <code>categorical_crossentropy<\/code> loss takes <strong>one-hot encoded<\/strong> labels such as <code>[1,0,0]<\/code>, <code>[0,1,0]<\/code>, <code>[0,0,1]<\/code> and they are interpreted such that the index of 1 indicates the class of the sample. For example <code>[0,0,1]<\/code> means this sample belongs to class 2 (i.e. lions). Further, in the context of classification models, since the output is usually a probability distribution produced by the output of softmax layer, this form of labels also corresponds to a probability distribution and match with the output of the model. Again, <code>[0,0,1]<\/code> means that with probability of one we know that this sample belongs to class two.<\/p>\n\n<p><code>sparse_categorical_crossentropy<\/code> is almost a convenient way to use <code>categorical_crossentropy<\/code> as the loss function where Keras (or its backend) would handle the integer labels internally and you don't need to manually convert labels to one-hot encoded form. However, if the labels you provide are one-hot encoded then you must use <code>categorical_crossentropy<\/code> as the loss function.<\/p>\n\n<p>Also you might be interested to look at <a href=\"https:\/\/stackoverflow.com\/a\/51892084\/2099607\">this answer<\/a> as well, where I have explained briefly about the activation and loss functions and the format of labels used in the context of different kinds of classification tasks.<\/p>\n"}]}
+{"QuestionId":53275380,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><artificial-intelligence>","CreationDate":"2018-11-13T06:52:59.603","AcceptedAnswerId":null,"OwnerUserId":10241857.0,"Title":"How to use output of one network in a graph to feed to another network in another graph in tensorflow?","Body":"<p>I have a problem where I have two tensorflow graph networks\nI'am feeding data into the first network and, manipulating the output of the network to create a new batch, I need to feed this new batch the second network in another graph in tensorflow, How to do it?<\/p>\n\n<p>NOTE : The batch size of the two input data, i.e. the batch size of the first input data, and the batch size of the input data fed to the second network, are different.<\/p>\n","answers":[]}
+{"QuestionId":53275530,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><python-3.6>","CreationDate":"2018-11-13T07:02:51.810","AcceptedAnswerId":null,"OwnerUserId":2637953.0,"Title":"Categorize a not trained object as unknown object in machine learning using python","Body":"<p>I'm new in machine learning. Currently I'm working on a fruit\/vegetable classification project using deep learning with keras. I was able to train the model. So far, it can detect already trained objects correctly but when I'm providing an other fruit\/vegetable which was not trained, its predicts from the trained model whereas, It should recognize it as unknown fruit.<\/p>\n\n<p>Suppose, I trained a model with Apple, Potato and Banana images. As long as I provide the images of Apple\/Potato\/Banana its predicts correctly. But the moment I provide a image of an Orange, it predicts as an Potato or lemon predicts as an Apple. Here is the code snippet that I'm using to train &amp; predicts fruits:<\/p>\n\n<pre><code>import os.path\nimport numpy as np\nnp.random.seed(123)\nfrom keras import applications\nfrom keras.models import Sequential\nfrom keras.layers import Dropout, Flatten, Dense\nfrom keras.utils.np_utils import to_categorical\nfrom keras.preprocessing.image import ImageDataGenerator, img_to_array, load_img\nfrom keras.layers.normalization import BatchNormalization\nimport matplotlib.pyplot as plt\nimport math\nimport cv2   \n\n#dimensions of images\nimg_width, img_height = 224, 224  \n\n#file paths &amp; directories\ntop_model_weights_path = 'bottleneck_fc_model.h5'\nbottleneck_train_path = 'bottleneck_features_train.npy'\nbottleneck_validation_path = 'bottleneck_features_validation.npy'\ntrain_data_dir = 'data\/train'\nvalidation_data_dir = 'data\/validation\/'\n\n#hyperparameters\nepochs = 10\nbatch_size = 16\n\n\ndef save_bottleneck_features():\n    model = applications.VGG16(include_top=False, weights='imagenet')\n\n    datagen = ImageDataGenerator(rescale=1.\/255,\n                                 shear_range=0.2,\n                                 zoom_range=0.2,\n                                 horizontal_flip=True)\n\n    generator = datagen.flow_from_directory(train_data_dir,\n                                            target_size=(img_width, img_height),\n                                            batch_size=batch_size,\n                                            shuffle=False)\n\n    no_train_samples = len(generator.filenames)\n\n    predict_size_train = int(math.ceil(no_train_samples \/ batch_size))\n    bottleneck_features_train = model.predict_generator(generator, predict_size_train)\n    np.save(bottleneck_train_path, bottleneck_features_train)\n\n    datagen = ImageDataGenerator(rescale=1.\/255)\n\n    generator = datagen.flow_from_directory(validation_data_dir,\n                                            target_size=(img_width, img_height),\n                                            batch_size=batch_size,\n                                            class_mode=None,\n                                            shuffle=False)\n\n    no_validation_samples = len(generator.filenames)\n\n    predict_size_validation = int(math.ceil(no_validation_samples \/ batch_size))\n    bottleneck_features_validation = model.predict_generator(generator, predict_size_validation)\n    np.save(bottleneck_validation_path, bottleneck_features_validation)\n\n\n\ndef train_top_model():\n    datagen_top = ImageDataGenerator(rescale=1.\/255)\n\n    generator_top = datagen_top.flow_from_directory(train_data_dir, \n                                                    target_size=(img_width, img_height),\n                                                    batch_size=batch_size,\n                                                    class_mode='categorical',\n                                                    shuffle=False)\n\n    num_classes = len(generator_top.class_indices)\n\n    # save the class indices to use later in predictions\n    np.save('class_indices.npy', generator_top.class_indices)\n\n    # get the class labels for the training data, in the original order\n    train_labels = generator_top.classes\n    # convert the training labels to categorical vectors\n    train_labels = to_categorical(train_labels, num_classes=num_classes)\n\n    generator_top = datagen_top.flow_from_directory(validation_data_dir,\n                                                    target_size=(img_width, img_height),\n                                                    batch_size=batch_size,\n                                                    class_mode=None,\n                                                    shuffle=False)\n\n    validation_labels = generator_top.classes\n    validation_labels = to_categorical(validation_labels, num_classes=num_classes)\n\n    # load the bottleneck features saved earlier\n    train_data = np.load('bottleneck_features_train.npy')\n    validation_data = np.load('bottleneck_features_validation.npy')\n\n\n    # build the model\n    model = Sequential()\n    model.add(Flatten(input_shape=train_data.shape[1:]))\n    model.add(BatchNormalization())\n    model.add(Dense(256, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(num_classes, activation='softmax'))\n\n    model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\n\n    history = model.fit(train_data, train_labels, epochs=epochs, batch_size=batch_size)\n\n    model.save_weights(top_model_weights_path)\n\n\ndef predict(image_path):\n    class_dictionary = np.load('class_indices.npy').item()\n    #print(\"Values: \",class_dictionary)\n\n    num_classes = len(class_dictionary)\n\n    orig = cv2.imread(image_path)\n\n    print('[INFO] loading and preprocessing image...')\n    image = load_img(image_path, target_size=(224, 224))\n    image = img_to_array(image)\n    image = image \/ 255\n    image = np.expand_dims(image, axis=0)\n\n    model = applications.VGG16(include_top=False, weights='imagenet')\n\n    bottleneck_prediction = model.predict(image)\n\n    # build top model\n    model = Sequential()\n    model.add(Flatten(input_shape=bottleneck_prediction.shape[1:]))\n    model.add(BatchNormalization())\n    model.add(Dense(256, activation='relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(num_classes, activation='sigmoid'))\n\n    model.load_weights(top_model_weights_path)\n\n    # use the bottleneck prediction on the top model to get the final classification\n    class_predicted = model.predict_classes(bottleneck_prediction)\n\n    probabilities = model.predict_proba(bottleneck_prediction)\n\n    inID = class_predicted[0]\n\n    inv_map = {v: k for k, v in class_dictionary.items()}\n\n    label = inv_map[inID]\n\n    print(\"Image ID: {}, Label: {}\".format(inID, label))\n\n    cv2.putText(orig, \"Predicted: {}\".format(label), (10, 30),\n                cv2.FONT_HERSHEY_PLAIN, 1.5, (0, 0, 0), 2)\n\n    cv2.imshow('Classification', orig)\n    cv2.imwrite('predicted.jpg', orig)\n    cv2.waitKey(0)\n\n\nif __name__ == '__main__': \n    if not os.path.exists(bottleneck_train_path):\n        save_bottleneck_features()\n    if not os.path.exists(top_model_weights_path):\n        train_top_model()\n\n    image_path = 'fruits\/orange.jpg'  \n    predict(image_path)\n<\/code><\/pre>\n\n<p>How to overcome this situation? Any help will be appreciated.<\/p>\n","answers":[{"AnswerId":"53300550","CreationDate":"2018-11-14T12:44:46.083","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>When training with <code>loss='categorical_crossentropy'<\/code> , and <code>activation='softmax'<\/code> at the last layer, each class is represented by 1 node in the graph at the last layer. softmax ensures the value of all the nodes is summed to 1, by normalizing all the values.<br>\ncategorial crossentropy assumes the node with the highest value is the predicted class, and compares it with the labels. <br><br>\nFor example, after predicting a sample you might can have the following values for nodes: <code>0.33 banana, 0.33 apple, 0.34 potato<\/code> , thus potato will be the class selected for the prediction. <br><br>\nIf you wish to have an 'unknown class' prediction, you should add it as another possible class, and have some samples labeled as unknown class during training.\n<br><br><\/p>\n\n<p><strong>Edit:<\/strong> <br>\nNote that while what i suggested above can work for you, training it can be very difficulty, as it might require a larger number of samples, both known and unkown, as the 'unkown' features are harder to generalize, and might damage the other classes generalization. <br><br>\nA more practical way would be to define a threshold, as i mentioned earlier, after predicting you recieve some value for each one of the possible labels. When picking the label, this value is the 'confidence' score of the label. <br>\nExamining the example above, we will pick potato with 0.34 confidence.<br>\nYou could define some threshold, say 0.6, and when assuming the label for the image, check the different confidence score, if the highest (the 'chosen') label is lower then the threshold, you can mark it as 'unknown fruit'.<\/p>\n"}]}
+{"QuestionId":53276197,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T07:50:53.627","AcceptedAnswerId":null,"OwnerUserId":4232454.0,"Title":"KeyError in freeze model","Body":"<p>I've used the TensorFlow's <code>graph_util.convert_variables_to_constants<\/code> function to freeze a <code>ckpt<\/code> model. I have added all nodes in the <code>ckpt<\/code> model to the whitelist of the <code>convert_variables_to_constants<\/code>.<\/p>\n\n<p>However, when I want to load the freezed model and use it, apparently a few nodes have not been added. I mean in the bellow line:<\/p>\n\n<p><code>logits = tf.get_default_graph().get_tensor_by_name(\"Logits\/BiasAdd:0\")<\/code><\/p>\n\n<p>it gives me the following error:<\/p>\n\n<pre><code>KeyError: \"The name 'Logits\/BiasAdd:0' refers to a Tensor which does not exist. The operation, 'Logits\/BiasAdd', does not exist in the graph.\"\n<\/code><\/pre>\n\n<p>Any ideas how can I add this node to the freezed model?<\/p>\n","answers":[]}
+{"QuestionId":53276981,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2018-11-13T08:43:41.673","AcceptedAnswerId":53277363.0,"OwnerUserId":5368297.0,"Title":"How to adopt multiple different loss functions in each steps of LSTM in Keras","Body":"<p>I have a set of sentences and their scores, I would like to train a marking system that could predict the score for a given sentence, such one example is like this:<\/p>\n\n<pre><code>(X =Tomorrow is a good day, Y = 0.9)\n<\/code><\/pre>\n\n<p>I would like to use LSTM to build such a marking system, and also consider the sequential relationship between each word in the sentence, so the training example shown above is transformed as following:<\/p>\n\n<pre><code>(x1=Tomorrow, y1=is) (x2=is, y2=a) (x3=a, y3=good) (x4=day, y4=0.9)\n<\/code><\/pre>\n\n<p>When training this LSTM, I would like the first three time steps using a softmax classifier, and the final step using a MSE. It is obvious that the loss function used in this LSTM is composed of two different loss functions. In this case, it seems the Keras does not provide the way to address my problem directly. In addition, I am not sure whether my method to build the marking system is correct or not.<\/p>\n","answers":[{"AnswerId":"53277363","CreationDate":"2018-11-13T09:07:20.670","ParentId":null,"OwnerUserId":"1462770","Title":null,"Body":"<p>Keras support multiple loss functions as well:<\/p>\n\n<pre><code>   model = Model(inputs=inputs,\n                 outputs=[lang_model, sent_model])\n\n    model.compile(optimizer='sgd', \n                  loss=['categorical_crossentropy', 'mse'],\n                  metrics=['accuracy'], loss_weights=[1., 1.])\n<\/code><\/pre>\n\n<p>Based on your explanation, I think you need a model that first, predict a token based on previous tokens, in NLP domain it usually called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Language_model\" rel=\"nofollow noreferrer\">Language model<\/a>, and then compute a score which I assume it is a sentiment (it is applicable to other domain).<\/p>\n\n<p>To do so, you can train your language model with LSTM and pick the last output of LSTM for your ranking task. To this end, you need to define two loss function: <code>categorical_crossentropy<\/code> for the language model and <code>MSE<\/code> for the ranking task.<\/p>\n\n<p>This tutorial would be helpful: <a href=\"https:\/\/www.pyimagesearch.com\/2018\/06\/04\/keras-multiple-outputs-and-multiple-losses\/\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2018\/06\/04\/keras-multiple-outputs-and-multiple-losses\/<\/a><\/p>\n"}]}
+{"QuestionId":53277246,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-13T09:01:02.380","AcceptedAnswerId":53277417.0,"OwnerUserId":10449174.0,"Title":"The results after each training keras model are different","Body":"<p>I'm a newbie in Machine Learning. I want to build a keras model which will be used for facial recognition. I am currently using the model at: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3, 3), padding='same',\n             input_shape=x_train.shape[1:]))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes))\nmodel.add(Activation('softmax'))\n\n# initiate RMSprop optimizer\nopt = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6)\n\n# Let's train the model using RMSprop\nmodel.compile(loss='categorical_crossentropy',\n          optimizer=opt,\n           metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I trained with the same data and parameters the same, but the training results are very different.There are 100% results or 28% results.\nWhat made that difference? <\/p>\n","answers":[{"AnswerId":"53277417","CreationDate":"2018-11-13T09:10:21.830","ParentId":null,"OwnerUserId":"6347629","Title":null,"Body":"<p>Setting the seed, when training the model will solve the problem. This will give you the repeatability. <\/p>\n\n<pre><code>np.random.seed(10)\ntf.set_random_seed(10)\n<\/code><\/pre>\n\n<p>Also make sure train and test split also does not change ever instance. Hence, you can set the seed for data splitting also. <\/p>\n"}]}
+{"QuestionId":53277465,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><generator>","CreationDate":"2018-11-13T09:12:18.170","AcceptedAnswerId":53279655.0,"OwnerUserId":7326400.0,"Title":"Keras custom generator TypeError: 'NoneType' object is not callable","Body":"<p>I've written a custom generator to load my dataset and feed it into the <code>fit_generator<\/code> method. But I'm getting an error. <code>question_train<\/code>,<code>img_lis_train<\/code> and <code>answers_train<\/code> are lists of strings. I want <code>mygen<\/code> to return batches of 32 samples of the format:<\/p>\n\n<pre><code>[images,questions] , answers\n<\/code><\/pre>\n\n<p>Here is the code:<\/p>\n\n<pre><code>def mygen(questions_train,img_lis_train,answers_train):\n    start = 0  \n    data_size = len(questions_train)\n    batch_size = 32        \n    while True:          \n        if( start+batch_size &lt;= data_size ):\n            batch_ques = questions_train[ start : start+batch_size ] \n            batch_ans = answers_train[ start : start+batch_size ] \n            batch_img_names = img_lis_train[ start : start+batch_size ] \n        elif(start &lt; data_size):\n            batch_ques = questions_train[ start : ] \n            batch_ans = answers_train[ start : ] \n            batch_img_names = img_lis_train[ start : start+batch_size ] \n        else:\n            start = 0\n            continue       \n\n        batch_img = []\n        for img_name in batch_img_names:\n            img = load_img('.\/dataset\/images\/' + str(img_name) + '.png' , target_size = (224,224))\n            img = img_to_array(img)   \n            batch_img.append( preprocess_input(img) )    \n\n        start += 32\n        print('start = ' + str(start))\n        yield [batch_img, batch_ques] ,batch_ans\n\nfc_model.fit_generator(mygen, steps_per_epoch=3376, epochs = 10)\n<\/code><\/pre>\n\n<p>Here is the error I get:<\/p>\n\n<pre><code>  File \"mycode.py\", line 210, in &lt;module&gt;\n    fc_model.fit_generator(mygen, steps_per_epoch=3376, epochs = 10)\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/keras\/legacy\/interfaces.py\", line 87, in wrapper\n    return func(*args, **kwargs)\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/keras\/models.py\", line 1223, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/keras\/legacy\/interfaces.py\", line 87, in wrapper\n    return func(*args, **kwargs)\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/keras\/engine\/training.py\", line 2083, in fit_generator\n    generator_output = next(output_generator)\nStopIteration\nException ignored in: &lt;bound method BaseSession.__del__ of &lt;tensorflow.python.client.session.Session object at 0x7f936527fcc0&gt;&gt;\nTraceback (most recent call last):\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 712, in __del__\n  File \"\/opt\/apps\/Python-3.5.1\/lib\/python3.5\/site-packages\/tensorflow\/python\/framework\/c_api_util.py\", line 31, in __init__\nTypeError: 'NoneType' object is not callable\n<\/code><\/pre>\n","answers":[{"AnswerId":"53279655","CreationDate":"2018-11-13T11:06:40.150","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to construct the generator by calling it and passing your data to it, otherwise how is it supposed to generate batches? Here is the correct way of doing it:<\/p>\n\n<pre><code>fc_model.fit_generator(mygen(q_train, i_train, a_train), ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53277988,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T09:39:35.887","AcceptedAnswerId":null,"OwnerUserId":1286165.0,"Title":"How to read tensorflow dataset caches without building the dataset again","Body":"<p>I have created a dataset (<code>tf.data.Dataset<\/code>) with fairly complicated transformations and I have cached it to a file. My question is how can I read the content of that dataset again without reconstructing the dataset object again. For example:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\ndb = tf.data.Dataset.range(10)\ndb = db.cache('\/tmp\/range')\nfor v in db:\n    print(v)\n# \/tmp\/range.data-00000-of-00001  \/tmp\/range.index files are created\n\n# later, you could restore the dataset from a method like this:\nnew_db = tf.data.Dataset.from_cache('\/tmp\/range')\n<\/code><\/pre>\n\n<p>The idea is to build the dataset in another context and use it without building the complicated input pipeline that I had before.<\/p>\n","answers":[{"AnswerId":"56385853","CreationDate":"2019-05-30T21:27:44.393","ParentId":null,"OwnerUserId":"11245741","Title":null,"Body":"<p>I am pretty sure if you complete one epoch of reading the dataset, the next time you start a new training, it will still read from this cache file even if you make changes to your pipeline.<\/p>\n"}]}
+{"QuestionId":53278007,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-13T09:41:00.623","AcceptedAnswerId":null,"OwnerUserId":10645296.0,"Title":"I'd like to call 'on_epoch_end()' at each end of epoch. [keras custom generator]","Body":"<p>I use keras and tried to define custom generator.<\/p>\n\n<p>In generator, I expect that the function \"on_epoch_end()\" is called at each end of epoch, But \"on_epoch_end()\" never be called anytime. <\/p>\n\n<p>Could you tell me why? Please.<\/p>\n\n<pre><code>from pathlib import Path\nimport math\n\nfrom tensorflow.keras.utils import Sequence\nfrom keras.utils import np_utils\n\nclass ImageSequence(Sequence):\n    def __init__(self, x, batch_size=512):\n        self.x_positive = x[0]\n        self.x_negative = x[1]\n        self.batch_size = batch_size\n\n    def __getitem__(self, idx):\n        hbs = self.batch_size\/\/2\n        idx_p = np.random.randint(0, self.x_positive.shape[0], hbs)\n        batch_x_positive = self.x_positive[idx_p]\n        #\n        idx_n = np.random.randint(0, self.x_negative.shape[0], hbs)\n        batch_x_negative = self.x_negative[idx_n]\n        #batch_x_negative = self.x_negative[idx*hbs : (idx+1)*hbs]\n        #\n        batch_x = np.r_[batch_x_positive, batch_x_negative]\n        #\n        batch_y = np.r_[np.ones(len(batch_x_positive)), np.zeros(len(batch_x_negative))]\n        return batch_x, batch_y\n\n    def __len__(self):\n        return math.ceil(2 * len(self.x_negative) \/ self.batch_size)\n\n    def _shuffle(self):        \n        self.x_negative = shuffle(self.x_negative)\n\n    def on_epoch_end(self):\n        self._shuffle()\n\n\ndata_gen = ImageSequence([train_positive, train_negative], batch_size=BATCH_SIZE)\n\nhistory = model.fit_generator(\n    generator=data_gen,\n    use_multiprocessing=True,\n    validation_data=(x_valid, y_valid),\n    steps_per_epoch=2 * len(train_positive) \/ BATCH_SIZE, \n    epochs=30,\n    verbose=2,\n    callbacks=[])\n<\/code><\/pre>\n\n<p>I wrote my environment(version info )<\/p>\n\n<pre><code>import tensorflow.keras\nprint(tensorflow.keras.__version__)\n2.1.6-tf\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53278261,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T09:54:01.470","AcceptedAnswerId":53279982.0,"OwnerUserId":192373.0,"Title":"Tensorflow batch: keep result as strings","Body":"<p>This simple program<\/p>\n\n<pre><code>import tensorflow as tf\n\ninput = 'string'\nbatch = tf.train.batch([tf.constant(input)], batch_size=1)\nwith tf.Session() as sess:\n    tf.train.start_queue_runners()\n    output, = sess.run(batch)\n\n    print(1, input, output)\n    print(2, str(output, 'utf-8'))\n    print(3, input.split('i'))\n    print(4, str(output, 'utf-8').split('i'))\n    print(5, output.split('i'))\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<blockquote>\n  <p>1 string b'string'<br\/>\n  2 string<br\/>\n  3 ['str', 'ng']<br\/>\n  4 ['str', 'ng']<br\/>\n  ERROR:tensorflow:Exception in QueueRunner: Session has been closed.<br\/>\n  &nbsp; &nbsp; &nbsp; print(5, output.split('i'))<br\/>\n  TypeError: a bytes-like object is required, not 'str'<\/p>\n<\/blockquote>\n\n<p>Why isn't the result a list of strings, if the input is?<\/p>\n\n<p><em>OK, @jdehesa <a href=\"https:\/\/stackoverflow.com\/questions\/53278261\/tensorflow-batch-keep-result-as-strings\/53279982?noredirect=1#comment93444348_53279982\">explained<\/a> <strong>WHY<\/strong>, but not how to 'fix' it.<\/em> I can apply <strong>bytes.decode()<\/strong> to the results of session:<\/p>\n\n<pre><code>output, = map(bytes.decode, sess.run(batch))\n<\/code><\/pre>\n\n<p>And there exists <strong>tf.map_fn()<\/strong> that should do the same on tensors. The only question is how I can use this in my scenario?<\/p>\n\n<hr>\n\n<p><em>PS<\/em>: actually, the error message is puzzling, too. The problem is that we provide a bytes object, not a string. But the <strong>TypeError<\/strong> suggests exactly the opposite.<\/p>\n\n<p><em>PPS<\/em>: the error message explained, thanks to @jdehesa: it was about the <em>split()<\/em>'s parameter, not the object. <code>output.split(b'i')<\/code> works well!<\/p>\n","answers":[{"AnswerId":"53279982","CreationDate":"2018-11-13T11:24:17.847","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The problem is that <code>output<\/code> is a <code>bytes<\/code> object, because TensorFlow <code>tf.string<\/code> tensors are indeed made of <code>bytes<\/code>. But then you are trying to use <code>split<\/code> with a <code>str<\/code> separator, and that is why it complains. Try:<\/p>\n\n<pre><code>output.split(b'i')\n<\/code><\/pre>\n\n<p>or:<\/p>\n\n<pre><code>output.decode().split('i')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53278677,"AnswerCount":1,"Tags":"<tensorflow><memory><keras><neural-network><backpropagation>","CreationDate":"2018-11-13T10:16:04.003","AcceptedAnswerId":53582740.0,"OwnerUserId":9673730.0,"Title":"Memory requirements for back propagation - why not use the mean activation?","Body":"<p>I need help understanding the memory requirements of a neural network and their differences between training and evaluation processes. More specifically, the memory requirements of the training process (I'm using a Keras API running on top of TensorFlow).<\/p>\n\n<p>For a CNN that contains N weights, when using a batch of size x, there is a constant amount of memory required for the weights themselves and the input data. During the <strong>forward pass<\/strong> the GPU needs additional x*N units of memory (the specific required amount is not crucial to the question) for passing all the samples simultaneously and calculating the activation of each neuron.<\/p>\n\n<p>My question is regarding the <strong>back propagation<\/strong> process, it seems that the process requires additional x*N units of memory(*) for the specific gradient of every weight for every sample. According to my understanding, it means that the algorithm calculates the specific gradients of each sample and then sums them up for the back-propagation to the previous layer.<\/p>\n\n<p><strong>Q.<\/strong> Since there is only a single update step per batch, why isn't the gradient calculation performed on the mean activation of each neuron? That way the additional required memory for training will only be (x+1)*N and not 2*x*N.<\/p>\n\n<p>(*) This is according to my own little experiment of the maximal allowed batch size during evaluation (~4200) and training (~1200). Obviously it is a very simplified way of looking at the memory requirments <\/p>\n","answers":[{"AnswerId":"53582740","CreationDate":"2018-12-02T17:23:44.503","ParentId":null,"OwnerUserId":"9673730","Title":null,"Body":"<p><strong>The short answer is<\/strong>: that is just the way the mini-batch SGD back-propagation algorithm works.\nLooking back at its origins and difference between using the standard SGD and mini-batch SGD it is clearer why.<\/p>\n\n<p>The <strong>standard stochastic gradient decent<\/strong> algorithm passes a single sample thru the model, then back-propagates its gradients and updates model weights before repeating the process with the next sample. The main downside is that it is a serial process (can't run samples simultaneously because the each sample needs to run on a model that was already updated by the previous sample), so it is very computationally expensive. In addition using just a single sample for each update results in a very noisy gradient.<\/p>\n\n<p>The <strong>mini-batch SGD<\/strong> utilizes the same principle, with one difference - the gradients are accumulated from multiple samples and an update is only performed once every x samples. This helps to get a smooth gradient during training and enables passing multiple samples thru the model in parallel. This is the algorithm which is used when training with keras\/tensorflow in mini-batches (commonly called batches but that term actually means using the batch gradient decent which is slightly different algorithm).<\/p>\n\n<p>I haven't found any work regarding using the mean of the gradients in each layer for the update. It is interesting to check the results of such an algorithm. It would be more memory efficient however it is likely that it will also be less capable of reaching good minimum points.<\/p>\n"}]}
+{"QuestionId":53278845,"AnswerCount":1,"Tags":"<keras><pytorch><torch>","CreationDate":"2018-11-13T10:24:06.480","AcceptedAnswerId":53286613.0,"OwnerUserId":4724057.0,"Title":"Initialize weights and bias in torch","Body":"<p>What is the equivalent command for the below keras code in Pytorch<\/p>\n\n<pre><code>Dense(64, kernel_initializer='he_normal', bias_initializer='zeros', name='uhat_digitcaps')(d5)\n<\/code><\/pre>\n\n<p>How to I initialize weights and bias?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53286613","CreationDate":"2018-11-13T17:34:34.957","ParentId":null,"OwnerUserId":"4400634","Title":null,"Body":"<pre><code>class Net(nn.Module):\n    def __init__(self, in_channels, out_channels):\n        self.linear = nn.Linear(in_channels, 64)\n        nn.init.kaiming_normal_(self.linear.weight, mode='fan_out')\n        nn.init.constant_(self.linear.bias, 0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53278915,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-13T10:27:36.790","AcceptedAnswerId":53280737.0,"OwnerUserId":10143111.0,"Title":"keras: how to add weights to loss evaluation","Body":"<p><strong>Todo :<\/strong><\/p>\n\n<ul>\n<li>I would like to add a weight for each pattern loss in a given Keras loss function.\n\n<blockquote>\n  <p><strong>For example:<\/strong> <br>if the error on pattern i is l_i, <br>I would like to consider, instead, an error l_i * c_i, where c_i is an input scalar.<\/p>\n<\/blockquote><\/li>\n<\/ul>\n","answers":[{"AnswerId":"53280737","CreationDate":"2018-11-13T12:09:31.007","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<pre><code>def customloss(y_true, y_pred):\n    c_i = ...\n    loss = ...(only use tensor operations on y_true and y_pred or use built in keras losses)\n    return c_i*loss\n<\/code><\/pre>\n\n<p>Now compile your model passing the loss function.<\/p>\n\n<pre><code>model.compile(loss = customloss)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53279390,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T10:52:13.020","AcceptedAnswerId":null,"OwnerUserId":5963279.0,"Title":"Tensorflow DLL import failed","Body":"<p>Using python 3.6.1, I've installed tensorflow-gpu via pip.<\/p>\n\n<p>Whenever I execute my script it bonks out on the first tensorflow import statement with the following stack trace:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\info\\Desktop\\Tools\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Users\\info\\Desktop\\Tools\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Users\\info\\Desktop\\Tools\\Python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Users\\info\\Desktop\\Tools\\Python36\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\info\\Desktop\\Tools\\Python36\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n<\/code><\/pre>\n\n<p>I've followed tensorflow-gpu setup by the book (i.e. following the steps in the official tensorflow documentaion to the letter). Has anyone had problems with this before? <\/p>\n\n<p>EDIT:<\/p>\n\n<p>I've also tried installing tensorflow cpu via pip, but then I get a different error message:<\/p>\n\n<pre><code>Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n<\/code><\/pre>\n\n<p>My computers specs:<\/p>\n\n<p>i7 4500K @ 5GHz\nNVidea 1070\nSamsung SSD\n32GB DDR4 RAM<\/p>\n","answers":[]}
+{"QuestionId":53279787,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2018-11-13T11:14:28.550","AcceptedAnswerId":null,"OwnerUserId":4229637.0,"Title":"error while running eval.py tensorflow object detection api","Body":"<p>When I run the eval.py script, the images are evaluated and in the logs I can see that visualizations are also created on the images. <\/p>\n\n<p>I have set num_examples to 50 in the pipeline.config. However, 50 images are not evaluated. After evaluation of some 9 images (which varies a lot - sometimes 5, sometimes 4), I get an error that <code>ValueError: Image with id 1531471339_visible.png already added.<\/code> <\/p>\n\n<p>I am not sure where I am going wrong. <\/p>\n\n<p>Note: This is my own dataset that I've trained. \nI also tried different machines, and still the same error. <\/p>\n\n<p>eval config:<\/p>\n\n<pre><code>eval_config {\n  num_examples: 50\n  use_moving_averages: false\n}\neval_input_reader {\n  label_map_path: \"\/home\/path\/to\/labelmap.pbtxt\"\n  tf_record_input_reader {\n    input_path: \"\/home\/path\/to\/file.tfrecord\"\n  }\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"55966174","CreationDate":"2019-05-03T08:40:33.423","ParentId":null,"OwnerUserId":"4151886","Title":null,"Body":"<p>I got the same error message. I solved it by adding<\/p>\n\n<p><strong>metrics_set: \"coco_detection_metrics\"<\/strong><\/p>\n\n<p>into my config file.<\/p>\n\n<pre><code>eval_config {\n  num_examples: 50\n  use_moving_averages: false\n  metrics_set: \"coco_detection_metrics\"\n}\n<\/code><\/pre>\n\n<p>Then the error message disappear... not sure why. Maybe you can give it a go. Let me know if it helps.<\/p>\n"}]}
+{"QuestionId":53280894,"AnswerCount":1,"Tags":"<python><tensorflow><indexing><tensor><softmax>","CreationDate":"2018-11-13T12:20:19.797","AcceptedAnswerId":53281169.0,"OwnerUserId":5319143.0,"Title":"return the top_k masked softmax of each row for a 2D tensor","Body":"<p>For any 2D tensor like <\/p>\n\n<p>[[2,5,4,7],\n[7,5,6,8]], <\/p>\n\n<p>I want to do softmax for the top <em>k<\/em> element in each row and then construct a new tensor by replacing all the other elements to 0.<\/p>\n\n<p>The result should be to get the softmax of top <em>k<\/em> (here k=2) elements for each row [[7,5],[8,7]], \nwhich is thus\n[[0.880797,0.11920291],\n[0.7310586,0.26894143]] \nand then reconstruct a new tensor according to the index of the top <em>k<\/em> elements in the original tensor, the final result should be <\/p>\n\n<p>[[0,0.11920291,0,0.880797],\n[0.26894143,0,0,0.7310586]].<\/p>\n\n<p>Is it possible to implement this kind of <em>masked softmax<\/em> in tensorflow? Many thanks in advance!<\/p>\n","answers":[{"AnswerId":"53281169","CreationDate":"2018-11-13T12:37:35.867","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is how you can do that:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input data\na = tf.placeholder(tf.float32, [None, None])\nnum_top = tf.placeholder(tf.int32, [])\n# Find top elements\na_top, a_top_idx = tf.nn.top_k(a, num_top, sorted=False)\n# Apply softmax\na_top_sm = tf.nn.softmax(a_top)\n# Reconstruct into original shape\na_shape = tf.shape(a)\na_row_idx = tf.tile(tf.range(a_shape[0])[:, tf.newaxis], (1, num_top))\nscatter_idx = tf.stack([a_row_idx, a_top_idx], axis=-1)\nresult = tf.scatter_nd(scatter_idx, a_top_sm, a_shape)\n# Test\nwith tf.Session() as sess:\n    result_val = sess.run(result, feed_dict={a: [[2, 5, 4, 7], [7, 5, 6, 8]], num_top: 2})\n    print(result_val)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[0.         0.11920291 0.         0.880797  ]\n [0.26894143 0.         0.         0.7310586 ]]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>Actually, there is a function that more closely does what you intend, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse\/softmax\" rel=\"nofollow noreferrer\"><code>tf.sparse.softmax<\/code><\/a>. However, it requires a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse\/SparseTensor\" rel=\"nofollow noreferrer\"><code>SparseTensor<\/code><\/a> as input, and I'm not sure it should be faster since it has to figure out which sparse values go together in the softmax. The good thing about this function is that you could have different number of elements to softmax in each row, but in your case that does not seem to be important. Anyway, here is an implementation with that, in case you find it useful.<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.placeholder(tf.float32, [None, None])\nnum_top = tf.placeholder(tf.int32, [])\n# Find top elements\na_top, a_top_idx = tf.nn.top_k(a, num_top, sorted=False)\n# Flatten values\nsparse_values = tf.reshape(a_top, [-1])\n# Make sparse indices\nshape = tf.cast(tf.shape(a), tf.int64)\na_row_idx = tf.tile(tf.range(shape[0])[:, tf.newaxis], (1, num_top))\nsparse_idx = tf.stack([a_row_idx, tf.cast(a_top_idx, tf.int64)], axis=-1)\nsparse_idx = tf.reshape(sparse_idx, [-1, 2])\n# Make sparse tensor\na_top_sparse = tf.SparseTensor(sparse_idx, sparse_values, shape)\n# Reorder sparse tensor\na_top_sparse = tf.sparse.reorder(a_top_sparse)\n# Softmax\nresult_sparse = tf.sparse.softmax(a_top_sparse)\n# Convert back to dense (or you can keep working with the sparse tensor)\nresult = tf.sparse.to_dense(result_sparse)\n# Test\nwith tf.Session() as sess:\n    result_val = sess.run(result, feed_dict={a: [[2, 5, 4, 7], [7, 5, 6, 8]], num_top: 2})\n    print(result_val)\n    # Same as before\n<\/code><\/pre>\n"}]}
+{"QuestionId":53280916,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><natural-language-processing>","CreationDate":"2018-11-13T12:22:02.280","AcceptedAnswerId":53285676.0,"OwnerUserId":4532083.0,"Title":"Getting vector obtained in the last layer of CNN before softmax layer","Body":"<p>I am trying to implement a system by encoding inputs using CNN. After CNN, I need to get a vector and use it in another deep learning method.<\/p>\n\n<pre><code>  def get_input_representation(self):\n    # get word vectors from embedding\n    inputs = tf.nn.embedding_lookup(self.embeddings, self.input_placeholder)\n\n\n    sequence_length = inputs.shape[1] # 56\n    vocabulary_size = 160 # 18765\n    embedding_dim = 256\n    filter_sizes = [3,4,5]\n    num_filters = 3\n    drop = 0.5\n\n    epochs = 10\n    batch_size = 30\n\n    # this returns a tensor\n    print(\"Creating Model...\")\n    inputs = Input(shape=(sequence_length,), dtype='int32')\n    embedding = Embedding(input_dim=vocabulary_size, output_dim=embedding_dim, input_length=sequence_length)(inputs)\n    reshape = Reshape((sequence_length,embedding_dim,1))(embedding)\n\n    conv_0 = Conv2D(num_filters, kernel_size=(filter_sizes[0], embedding_dim), padding='valid', kernel_initializer='normal', activation='relu')(reshape)\n    conv_1 = Conv2D(num_filters, kernel_size=(filter_sizes[1], embedding_dim), padding='valid', kernel_initializer='normal', activation='relu')(reshape)\n    conv_2 = Conv2D(num_filters, kernel_size=(filter_sizes[2], embedding_dim), padding='valid', kernel_initializer='normal', activation='relu')(reshape)\n\n    maxpool_0 = MaxPool2D(pool_size=(sequence_length - filter_sizes[0] + 1, 1), strides=(1,1), padding='valid')(conv_0)\n    maxpool_1 = MaxPool2D(pool_size=(sequence_length - filter_sizes[1] + 1, 1), strides=(1,1), padding='valid')(conv_1)\n    maxpool_2 = MaxPool2D(pool_size=(sequence_length - filter_sizes[2] + 1, 1), strides=(1,1), padding='valid')(conv_2)\n\n    concatenated_tensor = Concatenate(axis=1)([maxpool_0, maxpool_1, maxpool_2])\n    flatten = Flatten()(concatenated_tensor)\n    dropout = Dropout(drop)(flatten)\n    output = Dense(units=2, activation='softmax')(dropout)\n    model = Model(inputs=inputs, outputs=output)\n    adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)\n\n    model.compile(optimizer=adam, loss='binary_crossentropy', metrics=['accuracy'])\n    adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)\n\n    model.compile(optimizer=adam, loss='binary_crossentropy', metrics=['accuracy'])\n    print(\"Traning Model...\")\n    model.fit(X_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1, callbacks=[checkpoint], validation_data=(X_test, y_test))  # starts training\n\n\n    return ??\n<\/code><\/pre>\n\n<p>The above code, trains the model using <code>X_train<\/code> and <code>Y_train<\/code> and then tests it. However in my system I do not have <code>Y_train<\/code> or <code>Y_test<\/code>, I only need the vector in the last hidden layer before softmax layer. How can I obtain it?<\/p>\n","answers":[{"AnswerId":"53285676","CreationDate":"2018-11-13T16:41:17.837","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>For that you can define a backend function to get the output of arbitrary layer(s):<\/p>\n\n<pre><code>from keras import backend as K\n\nfunc = K.function([model.input], [model.layers[index_of_layer].output])\n<\/code><\/pre>\n\n<p>You can find the index of your desired layer using <code>model.summary()<\/code> where the layers are listed starting from index zero. If you need the layer before the last layer you can use <code>-2<\/code> as the index (i.e. <code>.layers<\/code> attribute is actually a list so you can index it like a list in python). Then you can use the function you have defined by passing a <strong>list<\/strong> of input array(s):<\/p>\n\n<pre><code>outputs = func(inputs)\n<\/code><\/pre>\n\n<p>Alternatively, you can also define a model for this purpose. This has been covered in <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">Keras documentation<\/a> more thoroughly so I advise you to read that.<\/p>\n"}]}
+{"QuestionId":53280967,"AnswerCount":1,"Tags":"<python><performance><machine-learning><iterator><pytorch>","CreationDate":"2018-11-13T12:24:49.217","AcceptedAnswerId":53293991.0,"OwnerUserId":1014809.0,"Title":"PyTorch next(iter(training_loader)) extremely slow, simple data, can't num_workers?","Body":"<p>Here <code>x_dat<\/code> and <code>y_dat<\/code> are just really long 1-dimensional tensors.<\/p>\n\n<pre><code>class FunctionDataset(Dataset):\n    def __init__(self):\n        x_dat, y_dat = data_product()\n\n        self.length = len(x_dat)\n        self.y_dat = y_dat\n        self.x_dat = x_dat\n\n    def __getitem__(self, index):\n        sample = self.x_dat[index]\n        label = self.y_dat[index]\n        return sample, label\n\n    def __len__(self):\n        return self.length\n\n...\n\ndata_set = FunctionDataset()\n\n...\n\ntraining_sampler = SubsetRandomSampler(train_indices)\nvalidation_sampler = SubsetRandomSampler(validation_indices)\n\ntraining_loader = DataLoader(data_set, sampler=training_sampler, batch_size=params['batch_size'], shuffle=False)\nvalidation_loader = DataLoader(data_set, sampler=validation_sampler, batch_size=valid_size, shuffle=False)\n<\/code><\/pre>\n\n<p>I have also  tried pinning the memory for the two loaders. Setting <code>num_workers<\/code> to > 0 gives me run-time errors between the processes (like EOF error and interruption errors). I get my batch with:<\/p>\n\n<pre><code>x_val, target = next(iter(training_loader))\n<\/code><\/pre>\n\n<p>The entire data-set would fit into memory\/gpu but I would like to emulate batches for this experiment. Profiling my process gives me the following:<\/p>\n\n<pre><code>16276989 function calls (16254744 primitive calls) in 38.779 seconds\n\n   Ordered by: cumulative time\n\n   ncalls  tottime  percall  cumtime  percall filename:lineno(function)\n   1745\/1    0.028    0.000   38.780   38.780 {built-in method builtins.exec}\n        1    0.052    0.052   38.780   38.780 simple aprox.py:3(&lt;module&gt;)\n        1    0.000    0.000   36.900   36.900 simple aprox.py:519(exploreHeatmap)\n        1    0.000    0.000   36.900   36.900 simple aprox.py:497(optFromSample)\n        1    0.033    0.033   36.900   36.900 simple aprox.py:274(train)\n  705\/483    0.001    0.000   34.495    0.071 {built-in method builtins.next}\n      222    1.525    0.007   34.493    0.155 dataloader.py:311(__next__)\n      222    0.851    0.004   12.752    0.057 dataloader.py:314(&lt;listcomp&gt;)\n  3016001   11.901    0.000   11.901    0.000 simple aprox.py:176(__getitem__)\n       21    0.010    0.000   10.891    0.519 simple aprox.py:413(validationError)\n      443    1.380    0.003    9.664    0.022 sampler.py:136(__iter__)\n  663\/221    2.209    0.003    8.652    0.039 dataloader.py:151(default_collate)\n      221    0.070    0.000    6.441    0.029 dataloader.py:187(&lt;listcomp&gt;)\n      442    6.369    0.014    6.369    0.014 {built-in method stack}\n  3060221    2.799    0.000    5.890    0.000 sampler.py:68(&lt;genexpr&gt;)\n  3060000    3.091    0.000    3.091    0.000 tensor.py:382(&lt;lambda&gt;)\n      222    0.001    0.000    1.985    0.009 sampler.py:67(__iter__)\n      222    1.982    0.009    1.982    0.009 {built-in method randperm}\n  663\/221    0.002    0.000    1.901    0.009 dataloader.py:192(pin_memory_batch)\n      221    0.000    0.000    1.899    0.009 dataloader.py:200(&lt;listcomp&gt;)\n....\n<\/code><\/pre>\n\n<p>Suggesting the data loader is immensely slow compared to the remaining activity of my experiment (training the model and lots of other computations etc.). What's going wrong and what would be the best way to speed this up?<\/p>\n","answers":[{"AnswerId":"53293991","CreationDate":"2018-11-14T05:58:48.817","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When retrieving a batch with<\/p>\n\n<pre><code>x, y = next(iter(training_loader))\n<\/code><\/pre>\n\n<p>you actually create a new instance of dataloader iterator at each call (!) See <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/1917#issuecomment-433698337\" rel=\"noreferrer\">this thread<\/a> for more infotrmation.<br>\nWhat you should do instead is create the iterator <em>once<\/em> (per epoch):<\/p>\n\n<pre><code>training_loader_iter = iter(training_loader)\n<\/code><\/pre>\n\n<p>and then call <code>next<\/code> for each batch on the iterator<\/p>\n\n<pre><code>for i in range(num_batches_in_epoch):\n  x, y = next(training_loader_iter)\n<\/code><\/pre>\n\n<p>I had similar issue before, and this also made the EOF errors you experience when using multiple workers go away.<\/p>\n"}]}
+{"QuestionId":53281860,"AnswerCount":0,"Tags":"<tensorflow><lstm><attention-model>","CreationDate":"2018-11-13T13:16:56.930","AcceptedAnswerId":null,"OwnerUserId":7675470.0,"Title":"How to create an LSTM with no input in Tensorflow?","Body":"<p>I am implementing an attention mechanism which uses an LSTM cell as encoder where the LSTM does not take\/have any input (only a hidden state). How would one specify an LSTM with no input in TensorFlow?<\/p>\n\n<p>Essentially, the hidden state would get concatenated with an empty input at the \"entrypoint\" of the LSTM cell.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Edit: related question I have asked at <a href=\"https:\/\/stackoverflow.com\/questions\/53272066\/attention-mechanism-with-an-lstm-which-takes-no-input\">Attention mechanism with an LSTM which takes no input<\/a><\/p>\n","answers":[]}
+{"QuestionId":53281990,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-11-13T13:23:02.937","AcceptedAnswerId":53282754.0,"OwnerUserId":6243129.0,"Title":"How to only detect humans in object detection API Tensorflow","Body":"<p>I am using <code>tensorflow<\/code> object detection API to detect objects. It is working fine in my windows system. How can I make changes in it to only detect mentioned objects, for example, I only want to detect humans and not all the objects.<\/p>\n\n<p>As per the 1st comment in this <a href=\"https:\/\/stackoverflow.com\/questions\/46515682\/how-to-detect-only-objects-of-a-specific-category-in-tensorflow-object-detection\">answer<\/a>, I have checked the visualization file but didn't find anything related to categories of objects. Then I looked into <code>category_util.py<\/code> and found out that there is <code>csv<\/code> file from which all the categories are being loaded but didnt found this <code>csv<\/code> file in the project. Can anyone please point me into the right direction. Thanks<\/p>\n","answers":[{"AnswerId":"53282754","CreationDate":"2018-11-13T14:03:34.827","ParentId":null,"OwnerUserId":"10240101","Title":null,"Body":"<p>I assume from your question, that you did not finetune your model yourself, but just used a pretrained one from the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">model zoo<\/a>!?<\/p>\n\n<p>In this case, I think the model already detect humans AND other objects and you want these other objects to disappear!? For doing so, you just have to change your <code>label_map.pbtxt<\/code> by deleting all classes which you don't need. If you are not sure where to find this file have a look into your <code>.config<\/code> file and search for <code>label_map_path=\"PATH\"<\/code>.<\/p>\n"}]}
+{"QuestionId":53282169,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T13:32:16.057","AcceptedAnswerId":53282991.0,"OwnerUserId":10646426.0,"Title":"ValueError: too many values to unpack 3","Body":"<pre><code>  1 i#:coding:utf-8\n  2 #0\u5bfc\u5165\u6a21\u5757\uff0c\u751f\u6210\u6a21\u62df\u6570\u636e\u96c6\n  3 import tensorflow as tf\n  4 import numpy as np\n  5 BATCH_SIZE = 8\n  6 seed = 23455\n  7 \n  8 #\u7ed9\u4e88seed\u4ea7\u751f\u968f\u673a\u6570\n  9 rng = np.random.RandomState(seed)\n 10 #\u968f\u673a\u6570\u8fd4\u56de32\u884c2\u5217\u77e9\u9635 \u8868\u793a32\u7ec4 \u4f53\u79ef\u548c\u91cd\u91cf \u4f5c\u4e3a\u8f93\u5165\u6570\u636e\u96c6\n 11 X = rng.rand(32,3)\n 12 \n 13 Y = [[int(x0+x1&lt;1)] for (x0,x1) in X]\n 14 print \"X:\\n\",X\n 15 print \"Y:\\n\",Y\n 16 \n 17 #1\u5b9a\u4e49\u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u5165\uff0c\u53c2\u6570\u548c\u8f93\u51fa\uff0c\u5b9a\u4e49\u5411\u524d\u4f20\u64ad\u8fc7\u7a0b\n 18 x = tf.placeholder(tf.float32, shape=(None, 2))\n 19 y_= tf.placeholder(tf.float32, shape=(None, 1))\n 20 \n 21 w1= tf.variable(tf.random([2,3], stddev=1, seed=1))\n 22 w2= tf.variable(tf.random([3,1], stddev=1, seed=1))\n 23 \n 24 a =tf.matmul(x,w1)\n 25 y =tf.matmul(a,w2)\n 26 \n 27 #\u5b9a\u4e49\u635f\u5931\u51fd\u6570\u96c6\u53cd\u5411\u4f20\u64ad\u65b9\u6cd5\n 28 loss = tf.reduce_mean(tf.square(y-y_))\n 29 #train_step = tf.train.MomentumOptimizer(0.001,0.9).minimize(loss)\n 30 #train_step = tf.train.AdamOptimizer(0.001).minimize(loss)\n 31 \n 32 #3\u751f\u6210\u4f1a\u8bdd\uff0c\u8bad\u7ec3steps\u8f6e\n 33 with tf.Session() as sess:\n 34     init_op = tf.global_variables_initializer()\n 35     sess.run(init_op)\n 36     # \u8f93\u51fa\u76ee\u524d\u672a\u7ecf\u8bad\u7ec3\u7684\u53c2\u6570\u53d6\u503c\u3002\n 37     print \"w1:\\n\", sess.run(w1)\n 38     print \"w2:\\n\", sess.run(w2)\n 39     print \"\\n\"\n 40 \n 41     #train the model\n 42     STEPS=3000\n 43     for i in range(STEPS):\n 44         start =(i*BATCH_SIZE) % 32\n 45         end = start + BATCH_SIZE\n 46         sess.run(train_step, feed_dict={x: X[start:end], y_: Y[start:end]})\n 47         if i % 500 == 0:\n 48             total_loss = sess.run(loss, feed_dict={x: X, y_: Y})\n 49             print(\"After %d training steps(s), loss on all data in %g\" % (i,     total_loss))\n 50 \n 51     #output the trained value of variables\n 52     print \"\\n\"\n 53     print \"w1:\\n\", sess.run(w1)\n 54     print \"w2:\\n\", sess.run(w2)\n<\/code><\/pre>\n\n<p>File \"tf3_6.py\", line 13, in \n    Y = [[int(x0+x1&lt;1)] for (x0,x1) in X]\nValueError: too many values to unpack.<\/p>\n\n<p>The code i don't think it is wrong but i still noticed the value error so i hope you guys to help me cope this question thanks a lot<\/p>\n","answers":[{"AnswerId":"53282991","CreationDate":"2018-11-13T14:15:26.813","ParentId":null,"OwnerUserId":"6560549","Title":null,"Body":"<p>The shape of <code>X<\/code> is <code>(32, 3)<\/code>, but in your list comprehension, you are only trying to unpack 2 values: <\/p>\n\n<pre><code>Y = [[int(x0+x1&lt;1)] for (x0,x1) in X]\n<\/code><\/pre>\n\n<p>Either change the shape of your array of rands:<\/p>\n\n<pre><code>X = rng.rand(32,2)\n<\/code><\/pre>\n\n<p>Or throw away the third rand in your list comp:<\/p>\n\n<pre><code>Y = [[int(x0+x1&lt;1)] for (x0,x1, _) in X]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53282439,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><neural-network><python-3.6>","CreationDate":"2018-11-13T13:46:41.523","AcceptedAnswerId":53283709.0,"OwnerUserId":8755792.0,"Title":"In TensorFlow, how can I look at the batch normalization parameters?","Body":"<p>I'm using a <code>tf.layers.batch_normalization<\/code> layer in my network. As you may know, batch  normalization employs trainable parameters gamma and beta to each unit u_i in this layer, to choose its own standard deviation and mean across u_i(x) for various inputs x. Typically gamma is initialized to 1 and beta to 0.<\/p>\n\n<p>I'm interested in peeking at the values of gamma and beta that are being learned at various units, to collect statistics about where they tend to end up after the network trains. How can I peek at their current values during each training instance?<\/p>\n","answers":[{"AnswerId":"53283709","CreationDate":"2018-11-13T14:53:57.660","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>You could get all the variables inside the scope of the batch normalization layer and print them. Example:<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.reset_default_graph()\nx = tf.constant(3.0, shape=(3,))\nx = tf.layers.batch_normalization(x)\n\nprint(x.name) # batch_normalization\/batchnorm\/add_1:0\n\nvariables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,\n                              scope='batch_normalization')\nprint(variables)\n\n#[&lt;tf.Variable 'batch_normalization\/gamma:0' shape=(3,) dtype=float32_ref&gt;,\n# &lt;tf.Variable 'batch_normalization\/beta:0' shape=(3,) dtype=float32_ref&gt;,\n# &lt;tf.Variable 'batch_normalization\/moving_mean:0' shape=(3,) dtype=float32_ref&gt;,\n#  &lt;tf.Variable 'batch_normalization\/moving_variance:0' shape=(3,) dtype=float32_ref&gt;]\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    gamma = sess.run(variables[0])\n    print(gamma) # [1. 1. 1.]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53282870,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-13T14:09:16.017","AcceptedAnswerId":null,"OwnerUserId":6823483.0,"Title":"How to predict a None (not in the categories)?","Body":"<p>Im using tensorflow.keras<\/p>\n\n<p>I made a model that can predict cat and a dog but when I try to predict image without dog or a cat I still get an answer of a cat\/dog.<\/p>\n\n<p>I tried find the % value of the prediction to know how low\/high the prediction was. But I couldn't find much updated answers, how can I get the probability values?<\/p>\n\n<p>Like:\nDog - 0.91\nCat - 0.09\nSomething like that...<\/p>\n\n<p>So how can I basically detect that thats not a cat or a dog?<\/p>\n","answers":[]}
+{"QuestionId":53283056,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-13T14:18:49.163","AcceptedAnswerId":53283958.0,"OwnerUserId":1984680.0,"Title":"How do the loss weights work in Tensorflow?","Body":"<p>I am training a recurrent binary classifier on a significantly underrepresented target class. Let's say our target class 1 represents &lt;1% of all the training data we have and class 0 >99%. In order to punish the model more for mispredicting the minority class I'd like to use weights in the loss function. For each minibatch, I have create a corresponding minibatch of weights where our target class gets a weight scalar >1.0 and our majority class &lt;1.0 accordingly. For example, in the code below we used 2.0 for class 1 and 0.6 for class 2.<\/p>\n\n<pre><code>loss_sum = 0.0\nfor t, o, tw in zip(self._targets_uns, self._logits_uns, self._targets_weight_uns):\n    # t -- targets tensor [batchsize x 1], tw -- weights tensor [batchsize x 1]\n    # e.g. [0, 0, 0, 0, 1, 1, 0] -- [0.5, 0.5, 0.5, 0.5, 2.0, 2.0, 0.5]\n    _loss = tf.losses.sigmoid_cross_entropy(t, o, weights=tw, label_smoothing=0,\n                                scope=\"sigmoid_cross_entropy\",\n                                loss_collection=tf.GraphKeys.LOSSES)\n    loss_sum += _loss\n<\/code><\/pre>\n\n<p>Once the model is trained, I check the prediction accuracy and find that it is slightly lower than the accuracy without weights. I continue experimenting trying out weight pairs of <code>[1.4, 0.8], [1.6, 0.4], [4.0, 0.1], [3.0, 1.0], ...<\/code> and so on. However, I am not getting any improvement over the unweighted training except marginal differences in 2-3% lower. Ok, maybe I misunderstood the docs for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/losses\/sigmoid_cross_entropy\" rel=\"nofollow noreferrer\">tf.losses.sigmoid_cross_entropy<\/a> function.<\/p>\n\n<blockquote>\n  <p>weights acts as a coefficient for the loss. If a scalar is provided, then the loss is simply scaled by the given value. If weights is a tensor of shape [batch_size], then the loss weights apply to each corresponding sample.<\/p>\n<\/blockquote>\n\n<p>I just reverse the pairs and use higher weight for class 0 and lower for class 1: <code>[0.5, 2.0], [0.8, 1.3], [0.2, 1.0], ...<\/code>. This also does not provide any improvement except being slightly worse than unweighted version. <\/p>\n\n<p>Can somebody please explain to me the behaviour of a weighted loss? Am I doing it correctly and what should I do to upweight the minority class?<\/p>\n","answers":[{"AnswerId":"53283958","CreationDate":"2018-11-13T15:09:07.023","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>Weighting is a general mathematical technique used for solving an over-specified system of equations of the form <code>Wx=y<\/code>, where <code>x<\/code> in the input vector, <code>y<\/code> is the output vector and <code>W<\/code> is the transformation matrix you wish to find.  Often times, these problems are solved using techniques such as SVD.  SVD will find the solution for <code>W<\/code> by minimizing the <code>least-squared error<\/code> for the over-specified system.  Tensorflow is basically solving a similar problem through its minimization process.<\/p>\n\n<p>In your case, what is happening is that you have 1 sample of class A and 99 samples of class B.  Because the solving process works to minimize the overall error, class B contributes to the solution by a factor of 99 to class A's 1.  In order to solve this, you should adjust your weights to so that class A and B have an even contribution to the solution, ie.. weight down class B by 0.01.<\/p>\n\n<p>More generally you can do...<\/p>\n\n<pre><code>ratio = num_B \/ (num_A + num_B)\nweights = [ratio, 1.0 - ratio]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53283485,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-11-13T14:42:46.113","AcceptedAnswerId":53292772.0,"OwnerUserId":6243129.0,"Title":"Tensorflow: Does object detection API returns detected object id","Body":"<p>I am using the <code>tensorflow<\/code> <code>object detection api<\/code> on windows system and it has been working fine. I am using the pre-trained model as of now which is <code>ssd_mobilenet_v1_coco_11_06_2017<\/code>. It is easily detecting all the objects in a given input video. I wanted to record time of each object so lets say, if in a video a car enters at <code>00:05<\/code> and leaves at <code>00:15<\/code> which means that it was in for <code>10secs<\/code>. <\/p>\n\n<p>To achieve this, I was looking if there is anything like id of each detected object which the API returns so that I can start a timer from the code to calculate the time of an object. Is there any already built functionality for this in the API.?<\/p>\n","answers":[{"AnswerId":"53290746","CreationDate":"2018-11-13T22:56:17.193","ParentId":null,"OwnerUserId":"8185230","Title":null,"Body":"<p>The Tensorflow Object Detection API does not currently track objects between frames.<\/p>\n"},{"AnswerId":"53292772","CreationDate":"2018-11-14T03:28:49.367","ParentId":null,"OwnerUserId":"6037560","Title":null,"Body":"<p>Tensorflow Object detection does not provide such functionality, but you can user KFC algorithm(easily available using Open CV) to track the object.<\/p>\n\n<p><a href=\"https:\/\/www.docs.opencv.org\/3.4.1\/d2\/dff\/classcv_1_1TrackerKCF.html\" rel=\"nofollow noreferrer\">https:\/\/www.docs.opencv.org\/3.4.1\/d2\/dff\/classcv_1_1TrackerKCF.html<\/a><\/p>\n\n<p>or You can implement SORT above object detection API which uses Kalman Filter but easy to integrate.<\/p>\n\n<p><a href=\"https:\/\/github.com\/abewley\/sort\/blob\/master\/sort.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/abewley\/sort\/blob\/master\/sort.py<\/a><\/p>\n"}]}
+{"QuestionId":53283711,"AnswerCount":0,"Tags":"<r><tensorflow><keras><deep-learning>","CreationDate":"2018-11-13T14:54:02.323","AcceptedAnswerId":null,"OwnerUserId":10646738.0,"Title":"Set class weights in Keras of R when there are multiple outputs","Body":"<p>I'm using the <code>keras<\/code> package in <code>R<\/code> to fit a neural network model. The model I'm working on has two outputs: <code>output1<\/code> is continuous(for regression), <code>output2<\/code> is binary(for classification). <\/p>\n\n<p>Since we have a very imbalanced dataset for the classification problem(<code>output2<\/code>), I want to assign different class weights to deal with the imbalance, but apparently we don't need to do that for <code>output1<\/code>(the regression).<\/p>\n\n<p>Here is the sample code for the NN model that I'm working on:<\/p>\n\n<pre><code>input &lt;- layer_input(shape = c(32,24))\noutput &lt;- input %&gt;%\n  layer_lstm(units = 64, dropout = 0.2, recurrent_dropout = 0.2) \npred1 &lt;- output %&gt;%\n  layer_dense(units = 1, name = \"output1\")\npred2 &lt;- output %&gt;%\n  layer_dense(units = 1, activation = \"sigmoid\", name = \"output2\")\nmodel &lt;- keras_model(\n  input,\n  list(pred1, pred2)\n)  \nsummary(model)\n\nmodel %&gt;% compile(\n  optimizer = \"rmsprop\",\n  loss = list(\n    output1 = \"mse\",\n    output2 = \"binary_crossentropy\"\n  ),\n  loss_weights = list(\n    output1 = 0.25,\n    output2 = 10\n  )\n)\n\nhistory &lt;- model %&gt;% fit(\n  train_x, list(output1 = train_y1,output2 = train_y2),\n  epochs = 10, \n  batch_size = 5000,\n  class_weight = ???,\n  validation_data = list(valid_x, list(output1 = valid_y1,output2 = valid_y2))\n)\n<\/code><\/pre>\n\n<p>If we just have one binary output, I know that the class weights can be assigned by:<\/p>\n\n<pre><code>class_weight = list(\"0\"=1,\"1\"=100),\n<\/code><\/pre>\n\n<p>but it doesn't work anymore when we have two outputs and just want to assign the weights to one of them. I guess I may need to somehow specify the name of the binary output in \"class_weight\" so that it knows the weights only apply to output2, but I don't know how to do it in R.<\/p>\n\n<p>Does anyone know how to assign class weights to the binary output only when we have two outputs(one is regression, one is classification)? Thank you very much for the help!<\/p>\n","answers":[]}
+{"QuestionId":53284138,"AnswerCount":2,"Tags":"<python><tensorflow><one-hot-encoding>","CreationDate":"2018-11-13T15:19:59.007","AcceptedAnswerId":53284512.0,"OwnerUserId":3741207.0,"Title":"One hot encoding huge 3D array","Body":"<p>As the title my data looks like this:\n<code>[\"test\", \"bob\", \"romeo\"]<\/code> - etc just random words\nI have converted them into numbers based on position in alphabet for each letter in the word so now it would be:<\/p>\n\n<pre><code>[[19, 4, 18, 19], [1, 14, 1], [17, 14, 12, 4, 14]]\n<\/code><\/pre>\n\n<p>and now I'd want to hot-encode it<\/p>\n\n<pre><code>tf.one_hot(featuresVectors, longestWordLen)\n<\/code><\/pre>\n\n<p>results in <\/p>\n\n<pre><code>ResourceExhaustedError: OOM when allocating tensor with shape[512996,62,62]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53284499","CreationDate":"2018-11-13T15:39:19.473","ParentId":null,"OwnerUserId":"5632058","Title":null,"Body":"<p>You basically run out of memory. Two approaches which could help are using less features (e.g count the words and just keep the top 10000 or so and a \"unknown toekn\" for therest) to make the onehot size smaller. Or you could use an embedding layer in your network and feed the integers directly. <\/p>\n"},{"AnswerId":"53284512","CreationDate":"2018-11-13T15:40:03.193","ParentId":null,"OwnerUserId":"2034218","Title":null,"Body":"<p>You are running out of memory, meaning that there isn't enough memory left on your device to create such a tensor. Given that your batch size is <code>512996<\/code> and your depth dimension is <code>62<\/code>, you are trying to create a tensor of <code>512996x62x62xsizeof(float): ~7.34Go !<\/code><\/p>\n\n<p>Since the indices are never going to be greater than 26. You can try to use a smaller data type for this tensor, like <code>int8<\/code>:  <code>tf.one_hot(featuresVectors, longestWordLen, dtype=tf.int8)<\/code>.\nThat should take <code>512996x62x62x1: ~1.83 Go<\/code> on your device. <\/p>\n\n<p>If your device still cannot allocate the tensor, then you'll have to reduce your batch size. (Aka the number of words) <\/p>\n"}]}
+{"QuestionId":53284223,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorflow-estimator>","CreationDate":"2018-11-13T15:24:15.760","AcceptedAnswerId":null,"OwnerUserId":6547250.0,"Title":"Custom input function for estimator instead of tf.data.dataset","Body":"<p>I want to know if anyone has created their own custom input function for tensorflow's estimator ? like in (<a href=\"https:\/\/www.tensorflow.org\/guide\/premade_estimators#create_input_functions\" rel=\"nofollow noreferrer\">link<\/a>) this image:<a href=\"https:\/\/i.stack.imgur.com\/RKynA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RKynA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where the say it is recommended to use tf.data.dataset. But I do not want to use that one, as I want to write my own iterator which yields data in batches and shuffles it as well. <\/p>\n\n<pre><code>    def data_in(train_data):\n      data = next(train_data)\n      ff = list(data)\n      tf.enable_eager_execution()\n      imgs = tf.stack([tf.convert_to_tensor(np.reshape(f[0], [img_size[0], img_size[1], img_size[2]])) for f\n                                        in ff])\n      lbls = tf.stack([f[1] for f in ff])\n      print('TRAIN data: %s %s ' % (imgs.get_shape(), lbls.get_shape()))\n      return imgs, lbls\n<\/code><\/pre>\n\n<p>output: <code>TRAIN data: (10, 32, 32, 3) (10,)<\/code><\/p>\n\n<p>where train_data is a generator object which iterates through my dataset using <strong>iter<\/strong> and np.reshape(f[0], [img_size[0], img_size<a href=\"https:\/\/i.stack.imgur.com\/RKynA.png\" rel=\"nofollow noreferrer\">2<\/a>, img_size<a href=\"https:\/\/i.stack.imgur.com\/RKynA.png\" rel=\"nofollow noreferrer\">2<\/a>] basically reshapes the data I extract to the required dimensions and it is a batch of an entire dataset. I use stack to convert the list of tensors to convert to stacked tensors. But when I use this with estimators I get an error for the features provided to the model saying the features do not have get_shape(). When I test it without an estimator it works well and it get_shape() also works well. <\/p>\n","answers":[{"AnswerId":"53302664","CreationDate":"2018-11-14T14:36:28.160","ParentId":null,"OwnerUserId":"6547250","Title":null,"Body":"<p>Hey kvish I figured it out how to do it. I just had to add these lines<\/p>\n\n<pre><code>experiment = tf.contrib.learn.Experiment(\ncifar_classifier,\ntrain_input_fn=lambda: data_in(),\neval_input_fn=lambda: data_in_eval(),\ntrain_steps=train_steps)\n<\/code><\/pre>\n\n<p>I know experiment is deprecated, I will also do it with estimator now :) <\/p>\n"}]}
+{"QuestionId":53284674,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2018-11-13T15:50:01.500","AcceptedAnswerId":53618000.0,"OwnerUserId":9702962.0,"Title":"Tensorflow serving with contrib operations","Body":"<p>How can I serve model with <a href=\"https:\/\/www.tensorflow.org\/serving\/\" rel=\"nofollow noreferrer\">tensorflow-serving<\/a>, if there are tf.contrib operations. I use Tensorflow Serving via Docker (latest) (version of tf 1.11) and when I serve model there is the next message:<\/p>\n\n<pre><code>\u201cFailed to start server. Error: Unknown: 1 servable(s) did not become available: {{{name: slider_universal version: 1} due to error: Not found: Op type not registered \u2018ImageProjectiveTransformV2\u2019 in binary running on 254345a5d9f1. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib, accessing (e.g.) tf.contrib.resampler should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.}, }\u201d\n<\/code><\/pre>\n\n<p>I also built with bazel but there was the same error<\/p>\n\n<p>I use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/transform\" rel=\"nofollow noreferrer\">tf.contrib.image.transform<\/a> <\/p>\n\n<p>If I delete this operation during exporting model it can be served by tensorflow serving<\/p>\n","answers":[{"AnswerId":"53618000","CreationDate":"2018-12-04T16:59:24.993","ParentId":null,"OwnerUserId":"2372357","Title":null,"Body":"<p>I had problems with the same op and it seems like the only way is to build it.<\/p>\n\n<p><code>tensorflow_serving\/model_servers\/BUILD<\/code> defines which Tensorflow ops will be included in variable <code>SUPPORTED_TENSORFLOW_OPS<\/code>, and I was a bit confused about this since it specifies that contrib ops should be included. However, since the tensorflow contrib build rule doesn't seem to include the ops under <code>contrib.image<\/code> so instead I explicitly added these by updating this variable to the following<\/p>\n\n<pre><code>SUPPORTED_TENSORFLOW_OPS = [\n    \"@org_tensorflow\/\/tensorflow\/contrib:contrib_kernels\",\n    \"@org_tensorflow\/\/tensorflow\/contrib:contrib_ops_op_lib\",\n    \"@org_tensorflow\/\/tensorflow\/contrib\/image:image_ops_kernels\",\n    \"@org_tensorflow\/\/tensorflow\/contrib\/image:image_ops_op_lib\",\n]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53284746,"AnswerCount":1,"Tags":"<keras><anaconda>","CreationDate":"2018-11-13T15:53:07.243","AcceptedAnswerId":null,"OwnerUserId":8412906.0,"Title":"ImportError: cannot import name '_obtain_input_shape'","Body":"<p>When I run <\/p>\n\n<pre><code>from keras_vggface.vggface import VGGFace\n\n# Based on VGG16 architecture -&gt; old paper(2015)\nvggface = VGGFace(model='vgg16') # or VGGFace() as default\n\n# Based on RESNET50 architecture -&gt; new paper(2017)\nvggface = VGGFace(model='resnet50')\n\n# Based on SENET50 architecture -&gt; new paper(2017)\nvggface = VGGFace(model='senet50')\n<\/code><\/pre>\n\n<p>in <a href=\"https:\/\/github.com\/rcmalli\/keras-vggface\" rel=\"nofollow noreferrer\">instructions<\/a>.<\/p>\n\n<p>I got this error<\/p>\n\n<p>ImportError: cannot import name '_obtain_input_shape'<\/p>\n\n<p>I follow the <a href=\"https:\/\/github.com\/rcmalli\/keras-vggface\" rel=\"nofollow noreferrer\">instructions<\/a> in install the keras_vggface. And here is a list info of my packages.<\/p>\n\n<pre><code># packages in environment at \/Users\/lzq\/anaconda3\/envs\/py36_faceRec:\n#\n# Name                    Version                   Build  Channel\nbleach                    1.5.0                     &lt;pip&gt;\nca-certificates           2018.03.07                    0  \ncertifi                   2018.10.15               py36_0  \nenum34                    1.1.6                     &lt;pip&gt;\nh5py                      2.8.0                     &lt;pip&gt;\nhtml5lib                  0.9999999                 &lt;pip&gt;\nKeras                     2.1.1                     &lt;pip&gt;\nkeras-vggface             0.5                       &lt;pip&gt;\nlibcxx                    4.0.1                hcfea43d_1  \nlibcxxabi                 4.0.1                hcfea43d_1  \nlibedit                   3.1.20170329         hb402a30_2  \nlibffi                    3.2.1                h475c297_4  \nMarkdown                  3.0.1                     &lt;pip&gt;\nncurses                   6.1                  h0a44026_0  \nnumpy                     1.15.4                    &lt;pip&gt;\nopenssl                   1.1.1                h1de35cc_0  \nPillow                    5.3.0                     &lt;pip&gt;\npip                       18.1                     py36_0  \nprotobuf                  3.6.1                     &lt;pip&gt;\npython                    3.6.7                haf84260_0  \nPyYAML                    3.13                      &lt;pip&gt;\nreadline                  7.0                  h1de35cc_5  \nscipy                     1.1.0                     &lt;pip&gt;\nsetuptools                40.5.0                   py36_0  \nsix                       1.11.0                    &lt;pip&gt;\nsqlite                    3.25.2               ha441bb4_0  \ntensorflow                1.4.0                     &lt;pip&gt;\ntensorflow-tensorboard    0.4.0                     &lt;pip&gt;\ntk                        8.6.8                ha441bb4_0  \nWerkzeug                  0.14.1                    &lt;pip&gt;\nwheel                     0.32.2                   py36_0  \nxz                        5.2.4                h1de35cc_4  \nzlib                      1.2.11               hf3cbc9b_2 \n<\/code><\/pre>\n","answers":[{"AnswerId":"53285146","CreationDate":"2018-11-13T16:14:42.340","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>The <code>VGGFace<\/code> package you're using is configured for an older version of Keras. The easiest solution for you is to edit the <code>keras-vggface\/keras_vggface\/models.py<\/code> file inside the <code>vvgface<\/code> directory and change line 15 from:<\/p>\n\n<p><code>from keras.applications.imagenet_utils import _obtain_input_shape<\/code><\/p>\n\n<p>to<\/p>\n\n<p><code>from keras_applications.imagenet_utils import _obtain_input_shape<\/code><\/p>\n\n<p>Alternatively you can also just downgrade Keras to 2.2.0 if you find that easier<\/p>\n"}]}
+{"QuestionId":53285128,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving><google-cloud-ml>","CreationDate":"2018-11-13T16:13:35.463","AcceptedAnswerId":53285850.0,"OwnerUserId":6695801.0,"Title":"Export tensorflow graph with export_saved_model","Body":"<p>I'm trying to train and deploy simplified Quick, Draw! classifier from <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/recurrent_quickdraw\" rel=\"nofollow noreferrer\">here<\/a> on Google Cloud. I've managed to train model in GC, now stuck at deploying it, more precisely, at <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/deploying-models#creating_serving_input_functions\" rel=\"nofollow noreferrer\">creating serving input functions<\/a>.<\/p>\n\n<p>I'm following instructions from <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#using_savedmodel_with_estimators\" rel=\"nofollow noreferrer\">here<\/a> and having tough times trying to understand what type of input tensor should be.<\/p>\n\n<p>Error: <\/p>\n\n<blockquote>\n  <p>TypeError: Failed to convert object of type  to Tensor. Contents: SparseTensor(indices=Tensor(\"ParseExample\/ParseExample:0\", shape=(?, 2), dtype=int64), values=Tensor(\"ParseExample\/ParseExample:1\", shape=(?,), dtype=float32), dense_shape=Tensor(\"ParseExample\/ParseExample:2\", shape=(2,), dtype=int64)). Consider casting elements to a supported type. <\/p>\n<\/blockquote>\n\n<p>Serving function:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  serialized_tf_example = tf.placeholder(dtype=tf.string, shape=[None], name='input_tensors')\n  receiver_tensors = {'infer_inputs': serialized_tf_example}\n  features = tf.parse_example(serialized_tf_example, feature_spec)\n  return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n<\/code><\/pre>\n\n<p>Feature specification:<\/p>\n\n<pre><code>feature_spec = {\n    \"ink\": tf.VarLenFeature(dtype=tf.float32),\n    \"shape\": tf.FixedLenFeature([2], dtype=tf.int64)\n}\n<\/code><\/pre>\n\n<p>Input layer:<\/p>\n\n<pre><code>def _get_input_tensors(features, labels):\n  shapes = features[\"shape\"]\n  lengths = tf.squeeze(\n    tf.slice(shapes, begin=[0, 0], size=[params.batch_size, 1]))\n  inks = tf.reshape(features[\"ink\"], [params.batch_size, -1, 3])\n\n  if labels is not None:\n    labels = tf.squeeze(labels)\n  return inks, lengths, labels\n<\/code><\/pre>\n\n<p>Code of model and training data were taken <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/recurrent_quickdraw#run_the_tutorial_code\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"53285850","CreationDate":"2018-11-13T16:49:52.337","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  ink = tf.placeholder(dtype=tf.float32, shape=[None, None, 3], name='ink')\n  length = tf.placeholder(dtype=tf.int64, shape=[None, 1])\n  features = {\"ink\": inks, \"length\": lengths}\n  return tf.estimator.export.ServingInputReceiver(features, features)\n<\/code><\/pre>\n\n<p>An example payload would be:<\/p>\n\n<pre><code>{\"instances\": [{\"ink\": [[0.1, 1.0, 2.0]], \"length\":[[1]]}]}\n<\/code><\/pre>\n\n<p>or as input to <code>gcloud predict --json-instances<\/code>:<\/p>\n\n<pre><code>{\"ink\": [[0.1, 1.0, 2.0]], \"length\":[[1]]}]\n<\/code><\/pre>\n\n<p>I didn't look into the actual code; if ink is generally going to hold a lot of floats, you may want to consider an alternative encoding system.<\/p>\n"}]}
+{"QuestionId":53285424,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-13T16:28:38.433","AcceptedAnswerId":53341233.0,"OwnerUserId":2300867.0,"Title":"AttributeError: 'Tensor' object has no attribute '_keras_shape'","Body":"<p>I'm trying to run code below to generate a JSON file and use it to built a t-SNE with a set of images. However my experience with Keras and machine learning is limited and I'm unable to run code below and getting error: AttributeError: 'Tensor' object has no attribute '_keras_shape'<\/p>\n\n<pre><code>import argparse\nimport sys\nimport numpy as np\nimport json\nimport os\nfrom os.path import isfile, join\nimport keras\nfrom keras.preprocessing import image\nfrom keras.applications.imagenet_utils import decode_predictions, preprocess_input\nfrom keras.models import Model\nfrom sklearn.decomposition import PCA\nfrom sklearn.manifold import TSNE\nfrom scipy.spatial import distance\n\ndef process_arguments(args):\n    parser = argparse.ArgumentParser(description='tSNE on audio')\n    parser.add_argument('--images_path', action='store', help='path to directory of images')\n    parser.add_argument('--output_path', action='store', help='path to where to put output json file')\n    parser.add_argument('--num_dimensions', action='store', default=2, help='dimensionality of t-SNE points (default 2)')\n    parser.add_argument('--perplexity', action='store', default=30, help='perplexity of t-SNE (default 30)')\n    parser.add_argument('--learning_rate', action='store', default=150, help='learning rate of t-SNE (default 150)')\n    params = vars(parser.parse_args(args))\n    return params\n\ndef get_image(path, input_shape):\n    img = image.load_img(path, target_size=input_shape)\n    x = image.img_to_array(img)\n    x = np.expand_dims(x, axis=0)\n    x = preprocess_input(x)\n    return x\n\ndef find_candidate_images(images_path):\n    \"\"\"\n    Finds all candidate images in the given folder and its sub-folders.\n    Returns:\n        images: a list of absolute paths to the discovered images.\n    \"\"\"\n    images = []\n    for root, dirs, files in os.walk(images_path):\n        for name in files:\n            file_path = os.path.abspath(os.path.join(root, name))\n            if ((os.path.splitext(name)[1]).lower() in ['.jpg','.png','.jpeg']):\n                images.append(file_path)\n    return images\n\ndef analyze_images(images_path):\n    # make feature_extractor\n    model = keras.applications.VGG16(weights='imagenet', include_top=True)\n    feat_extractor = Model(inputs=model.input, outputs=model.get_layer(\"fc2\").output)\n    input_shape = model.input_shape[1:3]\n    # get images\n    candidate_images = find_candidate_images(images_path)\n    # analyze images and grab activations\n    activations = []\n    images = []\n    for idx,image_path in enumerate(candidate_images):\n        file_path = join(images_path,image_path)\n        img = get_image(file_path, input_shape);\n        if img is not None:\n            print(\"getting activations for %s %d\/%d\" % (image_path,idx,len(candidate_images)))\n            acts = feat_extractor.predict(img)[0]\n            activations.append(acts)\n            images.append(image_path)\n    # run PCA firt\n    print(\"Running PCA on %d images...\" % len(activations))\n    features = np.array(activations)\n    pca = PCA(n_components=300)\n    pca.fit(features)\n    pca_features = pca.transform(features)\n    return images, pca_features\n\ndef run_tsne(images_path, output_path, tsne_dimensions, tsne_perplexity, tsne_learning_rate):\n    images, pca_features = analyze_images(images_path)\n    print(\"Running t-SNE on %d images...\" % len(images))\n    X = np.array(pca_features)\n    tsne = TSNE(n_components=tsne_dimensions, learning_rate=tsne_learning_rate, perplexity=tsne_perplexity, verbose=2).fit_transform(X)\n    # save data to json\n    data = []\n    for i,f in enumerate(images):\n        point = [float((tsne[i,k] - np.min(tsne[:,k]))\/(np.max(tsne[:,k]) - np.min(tsne[:,k]))) for k in range(tsne_dimensions) ]\n        data.append({\"path\":os.path.abspath(join(images_path,images[i])), \"point\":point})\n    with open(output_path, 'w') as outfile:\n        json.dump(data, outfile)\n\n\nif __name__ == '__main__':\n    params = process_arguments(sys.argv[1:])\n    images_path = params['images_path']\n    output_path = params['output_path']\n    tsne_dimensions = int(params['num_dimensions'])\n    tsne_perplexity = int(params['perplexity'])\n    tsne_learning_rate = int(params['learning_rate'])\n    run_tsne(images_path, output_path, tsne_dimensions, tsne_perplexity, tsne_learning_rate)\n    print(\"finished saving %s\" % output_path)\n<\/code><\/pre>\n\n<p>from: <code>https:\/\/github.com\/ml4a\/ml4a-ofx\/blob\/master\/scripts\/tSNE-images.py<\/code><\/p>\n\n<p>Here is what I'm getting:<\/p>\n\n<pre><code>    Traceback (most recent call last):\n  File \"tSNE-images.py\", line 95, in &lt;module&gt;\n    run_tsne(images_path, output_path, tsne_dimensions, tsne_perplexity, tsne_learning_rate)\n  File \"tSNE-images.py\", line 75, in run_tsne\n    images, pca_features = analyze_images(images_path)\n  File \"tSNE-images.py\", line 50, in analyze_images\n    feat_extractor = Model(inputs=model.input, outputs=model.get_layer(\"fc2\").output)\n  File \"\/Users\/...\/anaconda3\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/Users\/...\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/network.py\", line 91, in __init__\n    self._init_graph_network(*args, **kwargs)\n  File \"\/Users\/...\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/network.py\", line 251, in _init_graph_network\n    input_shapes=[x._keras_shape for x in self.inputs],\n  File \"\/Users\/...\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/network.py\", line 251, in &lt;listcomp&gt;\n    input_shapes=[x._keras_shape for x in self.inputs],\nAttributeError: 'Tensor' object has no attribute '_keras_shape'\n<\/code><\/pre>\n\n<p>I found similar error in here: <\/p>\n\n<pre><code>`https:\/\/stackoverflow.com\/questions\/47616588\/keras-throws-tensor-object-has-no-attribute-keras-shape-when-splitting-a`\n<\/code><\/pre>\n\n<p>However I can't seem to figure out how to go about updating code using Lambda. How can I solve this error?<\/p>\n","answers":[{"AnswerId":"53341233","CreationDate":"2018-11-16T15:49:47.940","ParentId":null,"OwnerUserId":"2300867","Title":null,"Body":"<p>I followed @user2300867 suggestion and updated tensorflow with:<\/p>\n\n<pre><code>pip3 install --upgrade tensorflow-gpu\n<\/code><\/pre>\n\n<p>and updated keras to 2.2.4<\/p>\n\n<pre><code>pip install Keras==2.2.4\n<\/code><\/pre>\n\n<p>I still got error: <\/p>\n\n<pre><code>TypeError: expected str, bytes or os.PathLike object, not NoneType\n<\/code><\/pre>\n\n<p>but this was easy to fix by simply editing the code for local paths<\/p>\n"}]}
+{"QuestionId":53286657,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-13T17:37:56.030","AcceptedAnswerId":null,"OwnerUserId":10647572.0,"Title":"what's meaning of diagram from tensorboard","Body":"<p>I follow some tutorial to try to visualize my network (word2vec) and get a training log. luckily, I make it work and get some diagram. I think the result is not bad because of loss diagram is look like this.<a href=\"https:\/\/i.stack.imgur.com\/FbKdj.png\" rel=\"nofollow noreferrer\">loss diagram<\/a>\nBut I don't understand what meaning of other diagrams.\nhow to understand about following diagrams? <\/p>\n\n<ul>\n<li>Bias    <a href=\"https:\/\/i.stack.imgur.com\/oACJ3.png\" rel=\"nofollow noreferrer\">bias diagram<\/a><\/li>\n<li>Weight  <a href=\"https:\/\/i.stack.imgur.com\/MHUZg.png\" rel=\"nofollow noreferrer\">weight diagram<\/a><\/li>\n<li>Embedding    <a href=\"https:\/\/i.stack.imgur.com\/NjV3L.png\" rel=\"nofollow noreferrer\">Embedding diagram<\/a><\/li>\n<\/ul>\n\n<p>ps. the code is here <\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    # Input data.\n    train_inputs = tf.placeholder(tf.int32, shape=[batch_size],name='train_inputs')\n    train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1],name='train_labels')\n    valid_dataset = tf.constant(valid_examples, dtype=tf.int32)\n\n    # Ops and variables pinned to the CPU because of missing GPU implementation\n    with tf.device('\/cpu:0'):\n        # Look up embeddings for inputs.\n        with tf.name_scope('Embeddings'):\n            embeddings = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0),name='Embeddings')\n            embed = tf.nn.embedding_lookup(embeddings, train_inputs)\n            tf.summary.histogram(name ='Embeddings', values = embeddings)\n\n        # Construct the variables for the NCE loss\n        with tf.name_scope('Weights'):\n            nce_weights = tf.Variable(tf.truncated_normal([vocabulary_size, embedding_size],stddev=1.0 \/ math.sqrt(embedding_size)),name='Weights')\n            tf.summary.histogram(name ='Weights', values = nce_weights)\n        with tf.name_scope('Biases'):\n            nce_biases = tf.Variable(tf.zeros([vocabulary_size]),dtype=tf.float32,name='Biases')\n            tf.summary.histogram(name ='Biases', values = nce_biases)\n\n    # Compute the average NCE loss for the batch.\n    # tf.nce_loss automatically draws a new sample of the negative labels each\n    # time we evaluate the loss.\n    with tf.name_scope('Loss'):\n        output_layer = tf.nn.nce_loss(weights=nce_weights,biases=nce_biases,inputs=embed,labels=train_labels,num_sampled=num_sampled,num_classes=vocabulary_size)\n        loss = tf.reduce_mean(output_layer)\n        tf.summary.scalar('loss', loss)\n    # Construct the SGD optimizer using a learning rate of 1.0.\n    with tf.name_scope('Train'):\n        optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)\n        merged = tf.summary.merge_all()\n        #train_set = [optimizer, loss]\n        train_set = [optimizer, merged]\n    # Compute the cosine similarity between minibatch examples and all embeddings.\n    norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))\n    normalized_embeddings = embeddings \/ norm\n    valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings, valid_dataset)\n    similarity = tf.nn.softmax(tf.matmul(valid_embeddings, normalized_embeddings, transpose_b=True))\n\n    # Add variable initializer.\n\n    init = tf.global_variables_initializer()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53286833,"AnswerCount":1,"Tags":"<tensorflow><proxy><pip><anaconda>","CreationDate":"2018-11-13T17:49:35.180","AcceptedAnswerId":53295761.0,"OwnerUserId":9026380.0,"Title":"Proxy Connection Error when trying to install packages with pip in anaconda environment","Body":"<p>I'm trying to install a package (tensorflow) utilizing pip in an anaconda environment <code>pip install --upgrade tensorflow<\/code> on Windows Server 2016, resulting in the following error:<\/p>\n\n<blockquote>\n  <p>Retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ProxyError('Cannot connect to proxy.', NewConnectionError(': Failed to establish a new connection: [WinError 10061] No connection could be made because the target machine actively refused it',))': \/simple\/tensorflow\/<\/p>\n<\/blockquote>\n\n<p>I'm in a corporate network and the proxy for anaconda is set.<\/p>\n\n<p>I've already tried the following things which don't help, the same error is still popping up:<\/p>\n\n<ul>\n<li>disabling firewall <\/li>\n<li>installing another package with pip (matplotlib)<\/li>\n<li>adding the flag <code>--proxy=http:\/\/myproxy.bla:8080<\/code> to the pip command <\/li>\n<li>adding the flag <code>trusted-host=pypi.python.org<\/code> to the pip command<\/li>\n<li>re-installing anaconda<\/li>\n<li>downloading the whl-file of tensorflow and <code>pip install<\/code> that. However, during that process, pip throws the error again when downloading any other package (e.g. astor)<\/li>\n<li>installation of other pip and python versions in other anaconda environments<\/li>\n<\/ul>\n\n<p>Any help or hint is appreciated!<\/p>\n","answers":[{"AnswerId":"53295761","CreationDate":"2018-11-14T08:23:29.617","ParentId":null,"OwnerUserId":"9026380","Title":null,"Body":"<p>I solved the problem through the command <code>pip install --upgrade tensorflow--proxy=\"myproxy.blabla:8080<\/code> after searching in Windows for \"Services\", scrolling to \"Cntlm Authentication Proxy\" and hitting \"Start the Service\".<\/p>\n\n<p>Note that executung \"Start cntlm authentication proxy\"-script didn't work. <\/p>\n"}]}
+{"QuestionId":53287034,"AnswerCount":0,"Tags":"<opencv><tensorflow>","CreationDate":"2018-11-13T18:05:03.497","AcceptedAnswerId":null,"OwnerUserId":10647715.0,"Title":"Using retrained Tensorflow model with Open CV","Body":"<p>I am trying to use a retrained TensorFlow model to work for live image recognition with Open CV.<\/p>\n\n<p>I followed the guide <a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">here<\/a> to retrain the model.<\/p>\n\n<p>The outputted files are shown below:\n<a href=\"https:\/\/i.stack.imgur.com\/Hvty9.png\" rel=\"nofollow noreferrer\">files<\/a><\/p>\n\n<p>I want to create the frozen_inference_graph.pb and graph.pbtxt file to use with Open CV.<\/p>\n\n<p>I am trying to follow <a href=\"https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc\" rel=\"nofollow noreferrer\">https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc<\/a> guide but I just cant seam to replicate it.<\/p>\n\n<p>Have I the correct files generated in order to make a frozen_inference_graph.pb and graph.pbtxt?<\/p>\n\n<p>Also could anyone be able to supply me with a sample commands if possible.<\/p>\n","answers":[]}
+{"QuestionId":53287556,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-tpu>","CreationDate":"2018-11-13T18:40:48.247","AcceptedAnswerId":53384174.0,"OwnerUserId":8463061.0,"Title":"How to find out more about the Cloud TPU device you are running your programs against?","Body":"<p>Whether we are using Google Colab or accessing Cloud TPUs directly, the below program gives only limited information about the underlying TPUs:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport tensorflow as tf\n\ntpu_address = 'grpc:\/\/' + os.environ['COLAB_TPU_ADDR']\nprint ('TPU address is', tpu_address)\n\ndef printTPUDevices():\n   with tf.Session(tpu_address) as session:\n      devices = session.list_devices()\n\n      print ('TPU devices:')\n      return devices\n\nprintTPUDevices()\n<\/code><\/pre>\n\n<p>Is there any documentation of programmatically or via bash commands to display more information, see this gist for e.g. <a href=\"https:\/\/gist.github.com\/neomatrix369\/256913dcf77cdbb5855dd2d7f5d81b84\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/neomatrix369\/256913dcf77cdbb5855dd2d7f5d81b84<\/a>.<\/p>\n","answers":[{"AnswerId":"53384174","CreationDate":"2018-11-19T23:34:26.637","ParentId":null,"OwnerUserId":"4716286","Title":null,"Body":"<p>The Cloud TPU system architecture is a bit different from GPU's so this level of information is not available.<\/p>\n\n<p>Because the client talks to a remote TensorFlow <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/system-architecture#software_architecture\" rel=\"nofollow noreferrer\">server<\/a> and uses <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/compiler\/xla\/g3doc\/overview.md\" rel=\"nofollow noreferrer\">XLA<\/a>, client code doesn't need to change based on the available features on the TPU, the remote server will compile machine instructions based on the TPU's capabilities.<\/p>\n\n<p>However the <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/cloud-tpu-tools\" rel=\"nofollow noreferrer\">Cloud TPU Profiler<\/a> does give a lower level view of the TPU for performance optimization. You can see a trace level view of what operations are using up memory and compute time.<\/p>\n"}]}
+{"QuestionId":53287587,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><computer-vision>","CreationDate":"2018-11-13T18:42:31.300","AcceptedAnswerId":null,"OwnerUserId":10644279.0,"Title":"TF Session graph is empty - Creating new tf graph from two imported graphs","Body":"<p>Right now, I am working with two models, one of which I load from a checkpoint, and the other I load a frozen model. Inside a class, I have a loadmodel function, and another processing function where I want to apply both of these models onto an image, so that when a future program can easily process everything on the image just by calling that method of the class.<\/p>\n\n<p>Here is a simplified example of the code, for just the methods I am struggling with in the class:<\/p>\n\n<pre><code>def segmented_stylization(self, img, target_size = 3840):       \n    self.g1 = self.seg_graph.as_graph_def()\n    self.g2 = self.eval_graph.as_graph_def()\n    start = time.time()  \n    # Here we need to create a combined graph\n    with tf.Graph().as_default() as g_combined:\n\n        # Get both our stylized frame and segmentation mappings by passing in the original frame input \n        style_input = self.eval_graph.get_tensor_by_name(\"img_placeholder:0\")\n        style_output = self.eval_graph.get_tensor_by_name(\"style_outputs:0\")            \n\n        seg_input = self.seg_graph.get_tensor_by_name(\"ImageTensor:0\")\n        seg_output = self.seg_graph.get_tensor_by_name(\"SemanticPredictions:0\")\n        seg_map = seg_output[0]\n\n        # Take our segmented mapping and identify the pixels corresponding to any human(s) in the picture\n        hPixels = tf.where(tf.equal(seg_map, self.humanIdx))\n\n        # Replace the pixels corresponding to a human \n        #mframe = tf.scatter_nd_update(style_output, hPixels, tf.gather_nd(resized_img, hPixels))\n\n        # mframe = tf.image.resize_nearest_neighbor(style_output (self.output_width, self.output_height))\n        #mframe = tf.clip_by_value(mframe, 0, 255)\n\n    sess_comb = tf.Session(graph = g_combined)\n    a = sess_comb.run([hPixels], feed_dict = {\"img_placeholder:0\" : resized_img, \"ImageTensor:0\" : resized_img})\n\n    output_img = sess_comb.run([mframe], feed_dict = {\"img_placeholder:0\" : resized_img, \"ImageTensor:0\" : resized_img})\n    return output_img\n<\/code><\/pre>\n\n<p>The error I get above is <\/p>\n\n<pre><code>raise RuntimeError('The Session graph is empty.  Add operations to the '\nRuntimeError: The Session graph is empty.  Add operations to the graph before calling run().\n<\/code><\/pre>\n\n<p>I've also tried wrapping in the combined session in a with statement, trying both nested and outside but haven't gotten it to work yet. <\/p>\n\n<p>Another thing I've tried instead was instead of get_tensor_by_name:<\/p>\n\n<pre><code>self.g1 = self.seg_graph.as_graph_def()\nself.g2 = self.eval_graph.as_graph_def()\nseg_output, = tf.import_graph_def(self.g1, input_map = {\"ImageTensor:0\" : resized_img, \"img_placeholder:0\" : resized_img}, return_elements = [\"SemanticPredictions:0\"])\nstyle_output, = tf.import_graph_def(self.g2, input_map={\"img_placeholder:0\" : resized_img}, return_elements=[\"style_outputs:0\"])\n<\/code><\/pre>\n\n<p>But this returns the error:<\/p>\n\n<pre><code>ValueError: Input 0 of node import\/Shape was passed double from import\/_inputs\/Const:0 incompatible with expected uint8.\n<\/code><\/pre>\n\n<p>Any insight would appreciate it, thanks! Still new to tf :). <\/p>\n","answers":[]}
+{"QuestionId":53288270,"AnswerCount":0,"Tags":"<python-3.x><performance><tensorflow><lstm><multilabel-classification>","CreationDate":"2018-11-13T19:31:56.037","AcceptedAnswerId":null,"OwnerUserId":1531248.0,"Title":"Second GPU not utilized - tensorflow gpu with python Bi-LSTM - Multi class Multi label classification","Body":"<p>My second GPU is not used and below is my model and its associated training code. (Bi-LSTM tensorflow and python).<\/p>\n\n<p>Could anyone give me where to do changes (tf.devices &amp; tf.variable_scope and so on). <\/p>\n\n<p>I am getting lot of confusion reading tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow noreferrer\">cifar10<\/a> examples<\/p>\n\n<p>Model - Using BI-LSTM architecture to predict multi-label-multi-class classification:<\/p>\n\n<pre><code>import tensorflow as tf   \ndef _variable_on_cpu(name, shape, initializer):\n    with tf.device('\/cpu:0'):\n        var = tf.get_variable(name, shape, initializer=initializer)\n    return var\n\ndef _variable_with_weight_decay(name, shape, initializer, wd):\n    var = _variable_on_cpu(name, shape, initializer)\n    if wd is not None and wd != 0.:\n        weight_decay = tf.multiply(tf.nn.l2_loss(var), wd, name='weight_loss')\n    else:\n        weight_decay = tf.constant(0.0, dtype=tf.float32)\n    return var, weight_decay\n\n\ndef _auc_pr(true, prob, threshold):\n    pred = tf.where(prob &gt; threshold, tf.ones_like(prob), tf.zeros_like(prob))\n    tp = tf.logical_and(tf.cast(pred, tf.bool), tf.cast(true, tf.bool))\n    fp = tf.logical_and(tf.cast(pred, tf.bool), tf.logical_not(tf.cast(true, tf.bool)))\n    fn = tf.logical_and(tf.logical_not(tf.cast(pred, tf.bool)), tf.cast(true, tf.bool))\n    pre = tf.truediv(tf.reduce_sum(tf.cast(tp, tf.int32)), tf.reduce_sum(tf.cast(tf.logical_or(tp, fp), tf.int32)))\n    rec = tf.truediv(tf.reduce_sum(tf.cast(tp, tf.int32)), tf.reduce_sum(tf.cast(tf.logical_or(tp, fn), tf.int32)))\n    return pre, rec\n\n\nclass Model(object):\n\n    def __init__(self, config, is_train=True):\n        self.num_uni = 100\n        self.is_train = is_train\n        self.emb_size = config['emb_size']\n        self.thresholdVal = config['thresholdVal']\n        self.batch_size = config['batch_size']\n        self.num_kernel = config['num_kernel']\n        self.min_window = config['min_window']\n        self.max_window = config['max_window']\n        self.vocab_size = config['vocab_size']\n        self.num_classes = config['num_classes']\n        self.sent_len = config['sent_len']\n        self.l2_reg = config['l2_reg']\n        self.multi_instance = config['attention']\n        self.multi_label = config['multi_label']\n        if is_train:\n            self.optimizer = config['optimizer']\n            self.dropout = config['dropout']\n        self.build_graph()\n\n    def build_graph(self):\n        \"\"\" Build the computation graph. \"\"\"\n        self._inputs = tf.placeholder(dtype=tf.int64, shape=[None, self.sent_len], name='input_x')\n        self._labels = tf.placeholder(dtype=tf.float32, shape=[None, self.num_classes], name='input_y')\n        losses = []\n\n        # lookup layer\n        with tf.variable_scope('embedding') as scope:\n            self._W_emb = _variable_on_cpu(name='embedding', shape=[self.vocab_size, self.emb_size],\n                                           initializer=tf.random_uniform_initializer(minval=-1.0, maxval=1.0))\n            # sent_batch is of shape: (batch_size, sent_len, emb_size, 1), in order to use conv2d\n            sent_batch = tf.nn.embedding_lookup(params=self._W_emb, ids=self._inputs)\n            #sent_batch = tf.expand_dims(sent_batch, -1)\n\n        with tf.variable_scope('forward-lstm') as scope:\n            frw_cell = tf.contrib.rnn.LSTMCell(num_units = self.num_uni)\n            dropout_frw = tf.contrib.rnn.DropoutWrapper(frw_cell)\n\n        with tf.variable_scope('backward-lstm') as scope:\n            bcw_cell = tf.contrib.rnn.LSTMCell(num_units = self.num_uni)\n            dropout_bcw = tf.contrib.rnn.DropoutWrapper(bcw_cell)\n\n        with tf.variable_scope('encoder') as scope:\n            rnn_output,last_state = tf.nn.bidirectional_dynamic_rnn(cell_fw = dropout_frw,\n                                                                    cell_bw = dropout_bcw,\n                                                                    inputs = sent_batch,\n                                                                    #sequence_length = self.sent_len,\n                                                                    dtype=tf.float32)\n\n        # Make use of output of last_state\n        pool_flat = tf.concat([last_state[0].c,last_state[1].c],axis=-1)\n\n        # drop out layer\n        if self.is_train and self.dropout &gt; 0:\n            pool_dropout = tf.nn.dropout(pool_flat, 1 - self.dropout)\n        else:\n            pool_dropout = pool_flat\n\n        # fully-connected layer\n        with tf.variable_scope('output') as scope:\n            W, wd = _variable_with_weight_decay('W', shape=[2*self.num_uni, self.num_classes],\n                                                initializer=tf.truncated_normal_initializer(stddev=0.05),\n                                                wd=self.l2_reg)\n            losses.append(wd)\n            biases = _variable_on_cpu('bias', shape=[self.num_classes],\n                                      initializer=tf.constant_initializer(0.01))\n            self.logits = tf.nn.bias_add(tf.matmul(pool_dropout, W), biases, name='logits')\n\n        # loss\n        with tf.variable_scope('loss') as scope:\n            if self.multi_label:\n                cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self._labels,\n                                                                        name='cross_entropy_per_example')\n            else:\n                cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self._labels,\n                                                                        name='cross_entropy_per_example')\n\n            cross_entropy_loss = tf.reduce_mean(cross_entropy, name='cross_entropy_loss')\n\n            losses.append(cross_entropy_loss)\n            self._total_loss = tf.add_n(losses, name='total_loss')\n\n        # eval with precision-recall\n        with tf.variable_scope('evaluation') as scope:\n            precision = []\n            recall = []\n            pre, rec = _auc_pr(self._labels, tf.sigmoid(self.logits), self.thresholdVal)\n            precision.append(pre)\n            recall.append(rec)\n            self._pre_op = precision\n            self._rec_op = recall\n\n        # train on a batch\n        self._lr = tf.Variable(0.0, trainable=False)\n        if self.is_train:\n            if self.optimizer == 'adadelta':\n                opt = tf.train.AdadeltaOptimizer(self._lr)\n            elif self.optimizer == 'adagrad':\n                opt = tf.train.AdagradOptimizer(self._lr)\n            elif self.optimizer == 'adam':\n                opt = tf.train.AdamOptimizer(self._lr)\n            elif self.optimizer == 'sgd':\n                opt = tf.train.GradientDescentOptimizer(self._lr)\n            else:\n                raise ValueError(\"Optimizer not supported.\")\n            grads = opt.compute_gradients(self._total_loss)\n            self._train_op = opt.apply_gradients(grads)\n\n            for var in tf.trainable_variables():\n                tf.summary.histogram(var.op.name, var)\n        else:\n            self._train_op = tf.no_op()\n\n        return\n\n    @property\n    def inputs(self):\n        return self._inputs\n\n    @property\n    def labels(self):\n        return self._labels\n\n    @property\n    def lr(self):\n        return self._lr\n\n    @property\n    def train_op(self):\n        return self._train_op\n\n    @property\n    def total_loss(self):\n        return self._total_loss\n\n    @property\n    def eval_op(self):\n        return self._eval_op\n\n    @property\n    def pre_op(self):\n        return self._pre_op\n\n    @property\n    def rec_op(self):\n        return self._rec_op\n\n    @property\n    def scores(self):\n        return tf.sigmoid(self.logits)\n\n    @property\n    def W_emb(self):\n        return self._W_emb\n\n    def assign_lr(self, session, lr_value):\n        session.run(tf.assign(self.lr, lr_value))\n\n    def assign_embedding(self, session, pretrained):\n        session.run(tf.assign(self.W_emb, pretrained))\n<\/code><\/pre>\n\n<p>Training:<\/p>\n\n<pre><code>import numpy as np\nfrom datetime import datetime\nimport tensorflow as tf\nimport pickle\nimport lstmModel as lstm\nimport time\nimport os\n\ndef frameDataBasedOnKey(config,keys,inputIdWithKey, labelIdWithKey):\n    num_classes = config['num_classes']\n    sent_len = config['sent_len']\n    stg_len = len(keys)\n\n    stg_inputIds = [inputIdWithKey.get(key) for key in keys]\n    stg_labelIds = [labelIdWithKey.get(key) for key in keys]\n\n    stg_input = np.zeros([stg_len,sent_len],dtype=int)\n    for rowId,rowVal in enumerate(stg_inputIds):\n        for colId,colVal in enumerate(rowVal):\n            stg_input[rowId][colId] = colVal\n\n    stg_Label = np.zeros([stg_len,num_classes],dtype=int)\n    scalingTo100 = 0-config['minlblIndex']\n    for rowId,rowVal in enumerate(stg_labelIds):\n        for colVal in [rVal for rVal in rowVal if rVal &gt;= config['minlblIndex'] and rVal &lt; config['maxlblIndex']]:\n            stg_Label[rowId][colVal+scalingTo100] = 1\n    return (stg_input,stg_Label)\n\ndef batch_iter(inputIdWithKey,labelIdWithKey,keys,config, isTest = False):\n    data_size = len(keys)\n    num_batches_per_epoch = int(np.ceil(float(data_size)\/config['batch_size']))\n    num_epochs= 1 if isTest else config['num_epochs']\n    for epoch in range(num_epochs):\n        for batch_num in range(num_batches_per_epoch):\n            start_index = batch_num * config['batch_size']\n            end_index = min((batch_num + 1) * config['batch_size'], data_size)\n            yield frameDataBasedOnKey(config,keys[start_index:end_index], inputIdWithKey, labelIdWithKey)\n\ndef offset(array, pre, post):\n    ret = np.array(array)\n    ret = np.insert(ret, 0, pre)\n    ret = np.append(ret, post)\n    return ret\n\ndef calc_auc_pr(precision, recall):\n    assert len(precision) == len(recall)\n    return np.trapz(offset(precision, 1, 0), x=offset(recall, 0, 1), dx=5)\n\ndef _summary_for_scalar(name, value):\n    return tf.Summary(value=[tf.Summary.Value(tag=name, simple_value=float(value))])\n\ndef train(inputIdWithKey,labelIdWithKey,train_key, test_key,config,embeddingWordMatrix):\n    num_batches_per_epoch = int(np.ceil(float(len(train_key))\/config['batch_size']))\n    max_steps = num_batches_per_epoch * config['num_epochs']\n\n    with tf.Graph().as_default():\n        with tf.variable_scope('lstm', reuse=None):\n            m = lstm.Model(config, is_train=True)\n        with tf.variable_scope('lstm', reuse=True):\n            mtest = lstm.Model(config, is_train=False)\n\n        saver = tf.train.Saver(tf.global_variables())\n        save_path = os.path.join(config['ModelOutput'].format(config['minlblIndex'],config['maxlblIndex']), 'model.ckpt')\n        summary_op = tf.summary.merge_all()\n\n        # session\n        #sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n        #with tf.Session().as_default() as sess:\n        with tf.Session(config=tf.ConfigProto(log_device_placement=True)).as_default() as sess:\n\n            proj_config = tf.contrib.tensorboard.plugins.projector.ProjectorConfig()\n            '''embedding = proj_config.embeddings.add()\n            embedding.tensor_name = m.W_emb.name\n            embedding.metadata_path = os.path.join(FLAGS.data_dir, 'vocab.txt')'''\n\n            train_summary_writer = tf.summary.FileWriter(os.path.join(config['ModelOutput'].format(config['minlblIndex'],config['maxlblIndex']), \"train\"), graph=sess.graph)\n            dev_summary_writer = tf.summary.FileWriter(os.path.join(config['ModelOutput'].format(config['minlblIndex'],config['maxlblIndex']), \"dev\"), graph=sess.graph)\n            tf.contrib.tensorboard.plugins.projector.visualize_embeddings(train_summary_writer, proj_config)\n            tf.contrib.tensorboard.plugins.projector.visualize_embeddings(dev_summary_writer, proj_config)\n\n            sess.run(tf.global_variables_initializer())\n\n            # assign pretrained embeddings\n            if config['use_pretrain']:\n                pretrained_embedding = np.array(embeddingWordMatrix.get('Embedding_matrix'))\n                m.assign_embedding(sess, pretrained_embedding)\n\n            # initialize parameters\n            current_lr = config['init_lr']\n            lowest_loss_value = float(\"inf\")\n            decay_step_counter = 0\n            global_step = 0\n\n            # evaluate on dev set\n            def dev_step(mtest, sess):\n                dev_loss = []\n                dev_auc = []\n                dev_f1_score = []\n\n                # create batch\n                test_batches = batch_iter(inputIdWithKey,labelIdWithKey,test_key, config, isTest = True)\n                for x_batch,y_batch in test_batches:\n                    #loss_value, eval_value = sess.run([mtest.total_loss, mtest.eval_op],\n                    loss_value, pre, rec = sess.run([mtest.total_loss, mtest.pre_op, mtest.rec_op],\n                        feed_dict={mtest.inputs: np.array(x_batch), mtest.labels: np.array(y_batch)})\n                    dev_loss.append(loss_value)\n                    dev_auc.append(calc_auc_pr(pre, rec))\n                    fOne = 0.0\n                    if pre[0] != 0.0 and rec[0] != 0.0:\n                        fOne = (2.0 * pre[0] * rec[0]) \/ (pre[0] + rec[0])\n                    dev_f1_score.append(fOne) # threshold = 0.5\n\n                return np.mean(dev_loss), np.mean(dev_auc), np.mean(dev_f1_score)\n\n            # train loop\n            train_loss = []\n            train_auc = []\n            train_f1_score = []\n            train_batches = batch_iter(inputIdWithKey,labelIdWithKey,train_key, config)\n            for x_batch,y_batch in train_batches:\n                batch_size = len(x_batch)\n\n                m.assign_lr(sess, current_lr)\n                global_step += 1\n\n                feed = {m.inputs: np.array(x_batch), m.labels: np.array(y_batch)}\n                start_time = time.time()\n                #_, loss_value, eval_value = sess.run([m.train_op, m.total_loss, m.eval_op], feed_dict=feed)\n                _, loss_value, pre, rec = sess.run([m.train_op, m.total_loss, m.pre_op, m.rec_op], feed_dict=feed)\n                proc_duration = time.time() - start_time\n                train_loss.append(loss_value)\n                #pre, rec = zip(*eval_value)\n                #print(\"Pre: \", pre,\" Rec: \", rec)\n                auc = calc_auc_pr(pre, rec)\n                #f1 = (2.0 * pre[5] * rec[5]) \/ (pre[5] + rec[5]) # threshold = 0.5\n                f1 =  0.0\n                if pre[0] != 0.0 and rec[0] != 0:\n                    f1 = (2.0 * pre[0] * rec[0]) \/ (pre[0] + rec[0])\n                train_auc.append(auc)\n                train_f1_score.append(f1)\n\n                assert not np.isnan(loss_value), \"Model loss is NaN.\"\n\n                # print log\n                if global_step % config['log_step'] == 0:\n                    examples_per_sec = batch_size \/ proc_duration\n                    format_str = '%s: step %d\/%d, f1 = %.4f, auc = %.4f, loss = %.4f, pre = %.4f, rec = %.4f ' + \\\n                                 '(%.1f examples\/sec; %.3f sec\/batch), lr: %.6f'\n                    print(format_str % (datetime.now(), global_step, max_steps, f1, auc, loss_value,pre[0], rec[0],\n                                        examples_per_sec, proc_duration, current_lr))\n\n                # write summary\n                if global_step % config['summary_step'] == 0:\n\n                    summary_str = sess.run(summary_op)\n                    train_summary_writer.add_summary(summary_str, global_step)\n                    dev_summary_writer.add_summary(summary_str, global_step)\n\n                    # summary loss, f1\n                    train_summary_writer.add_summary(\n                        _summary_for_scalar('loss', np.mean(train_loss)), global_step=global_step)\n                    train_summary_writer.add_summary(\n                        _summary_for_scalar('auc', np.mean(train_auc)), global_step=global_step)\n                    train_summary_writer.add_summary(\n                        _summary_for_scalar('f1', np.mean(train_f1_score)), global_step=global_step)\n\n                    dev_loss, dev_auc, dev_f1 = dev_step(mtest, sess)\n                    dev_summary_writer.add_summary(\n                        _summary_for_scalar('loss', dev_loss), global_step=global_step)\n                    dev_summary_writer.add_summary(\n                        _summary_for_scalar('auc', dev_auc), global_step=global_step)\n                    dev_summary_writer.add_summary(\n                        _summary_for_scalar('f1', dev_f1), global_step=global_step)\n\n\n                    dev_loss, dev_auc, dev_f1 = dev_step(mtest, sess)\n                    print(\"\\n===== write summary =====\")\n                    print(\"%s: step %d\/%d: train_loss = %.6f, train_auc = %.4f, train_f1 = %.4f\" \\\n                          % (datetime.now(), global_step, max_steps,\n                             np.mean(train_loss), np.mean(train_auc), np.mean(train_f1_score)))\n                    print(\"%s: step %d\/%d:   dev_loss = %.6f,   dev_auc = %.4f,   dev_f1 = %.4f\\n\" \\\n                          % (datetime.now(), global_step, max_steps, dev_loss, dev_auc, dev_f1))\n\n                    # reset container\n                    train_loss = []\n                    train_auc = []\n                    train_f1_score = []\n\n                # decay learning rate if necessary\n                if loss_value &lt; lowest_loss_value:\n                    lowest_loss_value = loss_value\n                    decay_step_counter = 0\n                else:\n                    decay_step_counter += 1\n                if decay_step_counter &gt;= config['tolerance_step']:\n                    current_lr *= config['lr_decay']\n                    print('%s: step %d\/%d, Learning rate decays to %.5f' % \\\n                          (datetime.now(), global_step, max_steps, current_lr))\n                    decay_step_counter = 0\n\n                # stop learning if learning rate is too low\n                if current_lr &lt; 1e-5:\n                    break\n\n                if global_step % config['checkpoint_step'] == 0:\n                    saver.save(sess, save_path, global_step=global_step)\n            saver.save(sess, save_path, global_step=global_step)\n\nconfig = {'batch_size':128,'num_epochs' : 2,'emb_size':300,'num_kernel':100,'min_window': 3,\n      'max_window':5,'vocab_size':0,'num_classes':150,\n      'sent_len':600,'l2_reg': 1e-4,'attention': False,'multi_label': True, 'optimizer':'adam','dropout':0.5,\n      'use_pretrain': True, 'init_lr' : 1e-3,'summary_step': 150,'lr_decay' : 0.95,'checkpoint_step':8000,\n      'tolerance_step':500,'log_step':25, 'thresholdVal': 0.5, 'ModelOutput':'ModelOutput_{0}-{1}',\n      'minlblIndex':0,'maxlblIndex':150}\n\nwith tf.device('\/cpu:0'):\n    with open('shortListedEmbeddingMatrix_Full.pk', 'rb') as fin:\n        embeddingWordMatrix = pickle.load(fin)\n\n    with open('InputEmbIdWithKey.pk', 'rb') as fin:\n        inputIdWithKey = pickle.load(fin)\n\n    with open('ICDLabelFull_Id_WithKey.pk', 'rb') as fin:\n        labelIdWithKey = pickle.load(fin)\n\n    inputKey = set(inputIdWithKey.keys())\n    labelKey = set(labelIdWithKey.keys())\n\n    inputData = list(inputKey.intersection(labelKey))\n\n    np.random.shuffle(inputData) # Shuffle the training data\n\n    train_size = int(len(inputData) * 0.99)\n    train_key = inputData[:train_size]\n    test_key = inputData[train_size:]\n\n    config['vocab_size'] = len(embeddingWordMatrix.get('word_vocab'))\n\nprint('Training in progress ...')\ntrain(inputIdWithKey,labelIdWithKey,train_key, test_key, config, embeddingWordMatrix)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CIOQM.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CIOQM.jpg\" alt=\"nvidia-smi and epoch progress: \"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53289156,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><dimensions>","CreationDate":"2018-11-13T20:39:35.967","AcceptedAnswerId":null,"OwnerUserId":2552306.0,"Title":"Adding an extra dimension with size None to a tensor","Body":"<p>How do you add a <code>None<\/code> outer dimension to a tensor ? <\/p>\n\n<p>Specifically, I have a  tensor of shape [299,299] that I want to return as output from a tensorflow <code>SavedModel<\/code> object. But the <code>SavedModel<\/code> object wants an unknown outer dimension to be able to deal with batches.<\/p>\n\n<p><code>tf.expand_dims()<\/code> adds an extra dimension but it is set to 1. <\/p>\n","answers":[{"AnswerId":"53289602","CreationDate":"2018-11-13T21:16:25.140","ParentId":null,"OwnerUserId":"5763590","Title":null,"Body":"<pre><code>In [1]: import tensorflow as tf    \n        x = tf.constant([3., 2.])\n<\/code><\/pre>\n\n<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\">tf.reshape()<\/a> for this,<\/p>\n\n<pre><code>In [1]: tf.reshape(x, [2, 1])\nOut[1]: TensorShape([Dimension(2), Dimension(1)])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53289208,"AnswerCount":0,"Tags":"<python><keras><neural-network><one-hot-encoding><language-model>","CreationDate":"2018-11-13T20:44:00.843","AcceptedAnswerId":null,"OwnerUserId":4466132.0,"Title":"One hot encoding of 1 million category","Body":"<p>For a language model, I have to predict a word for a given sequence of words. My vocabulary contains 1 million words. I'm trying to predict the words from it. I tried to use one hot encoding using keras (to_categorical) for predicted words. But for such a large vocabulary, I'm getting memory error in python. Is there any way to overcome this or my approach is wrong?<\/p>\n","answers":[]}
+{"QuestionId":53289212,"AnswerCount":3,"Tags":"<python><keras><deep-learning><lstm>","CreationDate":"2018-11-13T20:44:28.243","AcceptedAnswerId":53289565.0,"OwnerUserId":10605037.0,"Title":"create a masking layer in deep learning generative model","Body":"<p>I have created a generative model, which trains on padded sequences with a length of 129. I have tried to add a masking layer, but I get the following error message:<\/p>\n\n<pre><code>input_length\" is 129, but received input has shape (None, 129, 7905).\n<\/code><\/pre>\n\n<p><code>vocab_size = 7905<\/code><\/p>\n\n<p><code>seq_length_sent = 129<\/code><\/p>\n\n<p>Here is the model: <\/p>\n\n<pre><code>model = Sequential() \nmodel.add(Masking(mask_value=0, input_shape=(seq_length_sent, vocab_size)))\nmodel.add(Embedding(vocab_size, 50, input_length=seq_length_sent)) \nmodel.add(LSTM(100, return_sequences=True)) \nmodel.add(LSTM(100))\nmodel.add(Dense(100, activation='relu')) \nmodel.add(Dense(vocab_size, activation='softmax')) \nprint(model.summary())\n<\/code><\/pre>\n\n<p>The input are padded sequences, that have been tokenized so each word is represented by an integer. The output word is one-hot encoded by the following code:<\/p>\n\n<pre><code>X, y = sequences_sent[:,:-1], sequences_sent[:,-1] \ny = to_categorical(y, num_classes=vocab_size)\n<\/code><\/pre>\n\n<p>This is the first time, I'm making such a model, so I'm not sure whether the approach is right.<\/p>\n","answers":[{"AnswerId":"53358781","CreationDate":"2018-11-18T07:32:28.280","ParentId":null,"OwnerUserId":"10605037","Title":null,"Body":"<p>The input are padded sequences, that have been tokenized so each word is represented by an integer. The output word is one-hot encoded by the following code: <\/p>\n\n<pre><code>X, y = sequences_sent[:,:-1], sequences_sent[:,-1] \ny = to_categorical(y, num_classes=vocab_size)\n<\/code><\/pre>\n\n<p>This is the first time, I'm making such a model, so I'm not sure whether the approach is right. <\/p>\n"},{"AnswerId":"53295562","CreationDate":"2018-11-14T08:08:28.830","ParentId":null,"OwnerUserId":"6670629","Title":null,"Body":"<p>The error comes from how you define your masking layer. If your input is padded sequences of words and each word is one hot encoded, then you have correctly defined your masking layer. However, if your input is padded sequences of words without one hot encoding then you have to define your masking layer as <\/p>\n\n<p><code>model.add(Masking(mask_value=0, input_shape=(seq_length_sent,)))<\/code><\/p>\n\n<p>instead of<\/p>\n\n<p><code>model.add(Masking(mask_value=0, input_shape=(seq_length_sent,vocab_size))<\/code><\/p>\n"},{"AnswerId":"53289565","CreationDate":"2018-11-13T21:12:55.890","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>The problem is in where you define your input shape. I assume the input to your model are tokenized padded sequences. Therefore, the input should not be of shape of the <code>seq_length_sent*vocab_size<\/code>, but only of shape <code>seq_length_sent<\/code><\/p>\n\n<p>So to fix your issue just replace:<\/p>\n\n<p><code>model.add(Masking(mask_value=0, input_shape=(seq_length_sent, vocab_size)))<\/code><\/p>\n\n<p>with<\/p>\n\n<p><code>model.add(Masking(mask_value=0, input_shape=(seq_length_sent,)))<\/code><\/p>\n"}]}
+{"QuestionId":53289863,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2018-11-13T21:36:31.610","AcceptedAnswerId":null,"OwnerUserId":10648491.0,"Title":"TensorFlow Keras Error: expected flatten_input to have shape","Body":"<p>I am using TensorFlow Keras to create a machine learning model for a set of data.  I have been able to get the model to compile, however when I try to run a prediction, I get the following error:<\/p>\n\n<pre><code>    prediction = model.predict(test)\n  File \"C:\\Users\\mbaker\\PycharmProjects\\UnileverPatentWatch\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1864, in predict\n    x, check_steps=True, steps_name='steps', steps=steps)\n  File \"C:\\Users\\mbaker\\PycharmProjects\\UnileverPatentWatch\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 992, in _standardize_user_data\n    class_weight, batch_size)\n  File \"C:\\Users\\mbaker\\PycharmProjects\\UnileverPatentWatch\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1117, in _standardize_weights\n    exception_prefix='input')\n  File \"C:\\Users\\mbaker\\PycharmProjects\\UnileverPatentWatch\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py\", line 332, in standardize_input_data\n    ' but got array with shape ' + str(data_shape))\nValueError: Error when checking input: expected flatten_input to have shape (1872,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>My ML code is:<\/p>\n\n<pre><code>    train_features = np.array(train_features)\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(len(train_features[0]),)),\n    keras.layers.Dense(512, activation=tf.nn.relu),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(7, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(),\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(train_features, train_labels, epochs=10)\n\n    x = 0\nfor test in test_features:\n    test = np.array(test)\n    print(test)\n    print(len(test))\n    prediction = model.predict(test)\n    print(test_patents[x], prediction, labelIndex[prediction])\n    x += 1\n<\/code><\/pre>\n\n<p>test_features is a list of feature sets for which I want a prediction.  I have tried alternately passing the feature set as a numpy array as seen above or as a list and both methods have resulted in the above error.  The data is being passed to the predict function in the same format as was provided to train the model.  Any help would be greatly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53290220,"AnswerCount":1,"Tags":"<python><tensorflow><keras><training-data>","CreationDate":"2018-11-13T22:05:57.157","AcceptedAnswerId":null,"OwnerUserId":10648778.0,"Title":"Training speed on a shallow neural network with a small dataset","Body":"<p>My data contains of 1 feature and a label per feature \nie. <code>[\"smallBigTest\", \"toastBob\"]<\/code> &lt;- feature\n4 labels <code>[\"mix\", \"small\", \"big\", \"medium\"]<\/code><\/p>\n\n<p>I have converted my features to numbers based on alphabet\nie. <\/p>\n\n<pre><code>smallBigTest -&gt; 18, 12,  0, 53, 53, 27,  8,  6, 45,  4, 18, 19\ntoastBob -&gt;     19, 14,  0, 18, 19, 27, 14,  1, -1, -1, -1, -1\n<\/code><\/pre>\n\n<p>which later on I hot-encoded and reshaped so the final array of features would look like <code>[[hotencoded(18,12,0,53,53,27,8,6,45,4,18,19)], [hotencoded(19,14,0,18,19,27,14,1,-1,-1,-1,-1)]<\/code><\/p>\n\n<p>simply made it into a 2d array from 3d array to match my labels shape, \ni have also hot encoded labels<\/p>\n\n<p>the training data is about 60k lines of text 1.2mb csv file<\/p>\n\n<p>and here is my model:<\/p>\n\n<pre><code>model = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Dense(16, activation=tf.nn.sigmoid))\nmodel.add(tf.keras.layers.Dense(labelsDictSize, activation=tf.nn.softmax))\n\noptimizer = tf.train.GradientDescentOptimizer(0.05)\nmodel.compile(optimizer, loss=tf.losses.softmax_cross_entropy)\nmodel.fit(featuresOneHot,labelsOneHot, steps_per_epoch=dataCount, epochs=5, verbose=1)\n<\/code><\/pre>\n\n<p>I'm new to ML, so I might be doing something completely wrong or completely stupid, I thought though that this amount of data would be fine.\nTraining on my machine with gtx870m takes an hour per epoch and on google collaboratory around 20-30 minutes per epoch<\/p>\n","answers":[{"AnswerId":"53290760","CreationDate":"2018-11-13T22:57:58.567","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>It's not unusual for NLP models to take so much time to train. The only thing I would change about your model to speed up the learning process is changing the optimizer to something that doesn't have a fixed learning rate. That should speed up the process. I would suggest using adam as it's one of the fastest optimizers with good performance.<\/p>\n\n<p>Just replace <\/p>\n\n<pre><code>model.compile(optimizer, loss=tf.losses.softmax_cross_entropy)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>model.compile(optimizer='adam', loss=tf.losses.softmax_cross_entropy)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53290306,"AnswerCount":3,"Tags":"<python><scikit-learn><pytorch><confusion-matrix><transfer-learning>","CreationDate":"2018-11-13T22:13:35.747","AcceptedAnswerId":53291323.0,"OwnerUserId":2414957.0,"Title":"Confusion matrix and test accuracy for PyTorch Transfer Learning tutorial","Body":"<p>Following the Pytorch Transfer learning tutorial, I am interested in reporting only train and test accuracy as well as confusion matrix (say using sklearn confusionmatrix). How can I do that? The current tutorial only reports train\/val accuracy and I am having hard time figuring how to incorporate the sklearn confusionmatrix code there. Link to original tutorial here: <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/transfer_learning_tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/transfer_learning_tutorial.html<\/a><\/p>\n\n<pre><code>%matplotlib inline\nfrom graphviz import Digraph\nimport torch\nfrom torch.autograd import Variable\n# Author: Sasank Chilamkurthy\n\nfrom __future__ import print_function, division\n\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.optim import lr_scheduler\nimport numpy as np\nimport torchvision\nfrom torchvision import datasets, models, transforms\nimport matplotlib.pyplot as plt\nimport time\nimport os\nimport copy\n\nplt.ion()\ndata_transforms = {\n    'train': transforms.Compose([\n        transforms.RandomResizedCrop(224),\n        transforms.RandomHorizontalFlip(),\n        transforms.ToTensor(),\n        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n    ]),\n    'val': transforms.Compose([\n        transforms.Resize(256),\n        transforms.CenterCrop(224),\n        transforms.ToTensor(),\n        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])\n    ]),\n}\n\n\ndata_dir = \"images\"\nimage_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),\n                                          data_transforms[x])\n                  for x in ['train', 'val']}\ndataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,\n                                             shuffle=True, num_workers=4)\n              for x in ['train', 'val']}\ndataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}\nclass_names = image_datasets['train'].classes\n\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n\ndef imshow(inp, title=None):\n    \"\"\"Imshow for Tensor.\"\"\"\n    inp = inp.numpy().transpose((1, 2, 0))\n    mean = np.array([0.485, 0.456, 0.406])\n    std = np.array([0.229, 0.224, 0.225])\n    inp = std * inp + mean\n    inp = np.clip(inp, 0, 1)\n    plt.imshow(inp)\n    if title is not None:\n        plt.title(title)\n    plt.pause(0.001)  # pause a bit so that plots are updated\n\n\n# Get a batch of training data\ninputs, classes = next(iter(dataloaders['train']))\n\n# Make a grid from batch\nout = torchvision.utils.make_grid(inputs)\n\nimshow(out, title=[class_names[x] for x in classes])\n\ndef train_model(model, criterion, optimizer, scheduler, num_epochs=25):\n    since = time.time()\n\n    best_model_wts = copy.deepcopy(model.state_dict())\n    best_acc = 0.0\n\n    for epoch in range(num_epochs):\n        print('Epoch {}\/{}'.format(epoch, num_epochs - 1))\n        print('-' * 10)\n\n        # Each epoch has a training and validation phase\n        for phase in ['train', 'val']:\n            if phase == 'train':\n                scheduler.step()\n                model.train()  # Set model to training mode\n            else:\n                model.eval()   # Set model to evaluate mode\n\n            running_loss = 0.0\n            running_corrects = 0\n\n            # Iterate over data.\n            for inputs, labels in dataloaders[phase]:\n                inputs = inputs.to(device)\n                labels = labels.to(device)\n\n                # zero the parameter gradients\n                optimizer.zero_grad()\n\n                # forward\n                # track history if only in train\n                with torch.set_grad_enabled(phase == 'train'):\n                    outputs = model(inputs)\n                    _, preds = torch.max(outputs, 1)\n                    loss = criterion(outputs, labels)\n\n                    # backward + optimize only if in training phase\n                    if phase == 'train':\n                        loss.backward()\n                        optimizer.step()\n\n                # statistics\n                running_loss += loss.item() * inputs.size(0)\n                running_corrects += torch.sum(preds == labels.data)\n\n            epoch_loss = running_loss \/ dataset_sizes[phase]\n            epoch_acc = running_corrects.double() \/ dataset_sizes[phase]\n\n            print('{} Loss: {:.4f} Acc: {:.4f}'.format(\n                phase, epoch_loss, epoch_acc))\n\n            # deep copy the model\n            if phase == 'val' and epoch_acc &gt; best_acc:\n                best_acc = epoch_acc\n                best_model_wts = copy.deepcopy(model.state_dict())\n\n        print()\n\n    time_elapsed = time.time() - since\n    print('Training complete in {:.0f}m {:.0f}s'.format(\n        time_elapsed \/\/ 60, time_elapsed % 60))\n    print('Best val Acc: {:4f}'.format(best_acc))\n\n    # load best model weights\n    model.load_state_dict(best_model_wts)\n    return model\n\ndef visualize_model(model, num_images=6):\n    was_training = model.training\n    model.eval()\n    images_so_far = 0\n    fig = plt.figure()\n\n    with torch.no_grad():\n        for i, (inputs, labels) in enumerate(dataloaders['val']):\n            inputs = inputs.to(device)\n            labels = labels.to(device)\n\n            outputs = model(inputs)\n            _, preds = torch.max(outputs, 1)\n\n            for j in range(inputs.size()[0]):\n                images_so_far += 1\n                ax = plt.subplot(num_images\/\/2, 2, images_so_far)\n                ax.axis('off')\n                ax.set_title('predicted: {}'.format(class_names[preds[j]]))\n                imshow(inputs.cpu().data[j])\n\n                if images_so_far == num_images:\n                    model.train(mode=was_training)\n                    return\n        model.train(mode=was_training)\n\nmodel_ft = models.resnet18(pretrained=True)\nnum_ftrs = model_ft.fc.in_features\nmodel_ft.fc = nn.Linear(num_ftrs, 9)\n\nmodel_ft = model_ft.to(device)\n\ncriterion = nn.CrossEntropyLoss()\n\n# Observe that all parameters are being optimized\noptimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)\n\n# Decay LR by a factor of 0.1 every 7 epochs\nexp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)\n\nmodel_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,\n                       num_epochs=25)\n\nvisualize_model(model_ft)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58690208","CreationDate":"2019-11-04T09:13:27.090","ParentId":null,"OwnerUserId":"3177661","Title":null,"Body":"<p>Here is a slightly modified(direct) approach using sklearn's confusion_matrix:-<\/p>\n\n<pre><code>from sklearn.metrics import confusion_matrix\n\nnb_classes = 9\n\n# Initialize the prediction and label lists(tensors)\npredlist=torch.zeros(0,dtype=torch.long, device='cpu')\nlbllist=torch.zeros(0,dtype=torch.long, device='cpu')\n\nwith torch.no_grad():\n    for i, (inputs, classes) in enumerate(dataloaders['val']):\n        inputs = inputs.to(device)\n        classes = classes.to(device)\n        outputs = model_ft(inputs)\n        _, preds = torch.max(outputs, 1)\n\n        # Append batch prediction results\n        predlist=torch.cat([predlist,preds.view(-1).cpu()])\n        lbllist=torch.cat([lbllist,classes.view(-1).cpu()])\n\n# Confusion matrix\nconf_mat=confusion_matrix(lbllist.numpy(), predlist.numpy())\nprint(conf_mat)\n\n# Per-class accuracy\nclass_accuracy=100*conf_mat.diagonal()\/conf_mat.sum(1)\nprint(class_accuracy)\n<\/code><\/pre>\n"},{"AnswerId":"54496741","CreationDate":"2019-02-02T19:31:41.757","ParentId":null,"OwnerUserId":"7482900","Title":null,"Body":"<p>Another simple way to get accuracy is to use sklearns \"accuracy_score\". \nHeres an example:<\/p>\n\n<pre><code>from sklearn.metrics import accuracy_score\ny_pred = y_pred.data.numpy()\naccuracy = accuracy_score(labels, np.argmax(y_pred, axis=1))\n<\/code><\/pre>\n\n<p>First you need to get the data from the variable.\n\"y_pred\" is the predictions from your model, and labels are of course your labels. <\/p>\n\n<p>np.argmax returns the index of the largest value inside the array. We want the largest value as it corresponds to the highest probability class when using softmax for multi-class classification. Accuracy score will return a percentage of matches between the labels and y_pred.<\/p>\n"},{"AnswerId":"53291323","CreationDate":"2018-11-14T00:12:04.767","ParentId":null,"OwnerUserId":"2414957","Title":null,"Body":"<p>Answer given by <code>ptrblck<\/code> of PyTorch community. Thanks a lot!  <\/p>\n\n<pre><code>nb_classes = 9\n\nconfusion_matrix = torch.zeros(nb_classes, nb_classes)\nwith torch.no_grad():\n    for i, (inputs, classes) in enumerate(dataloaders['val']):\n        inputs = inputs.to(device)\n        classes = classes.to(device)\n        outputs = model_ft(inputs)\n        _, preds = torch.max(outputs, 1)\n        for t, p in zip(classes.view(-1), preds.view(-1)):\n                confusion_matrix[t.long(), p.long()] += 1\n\nprint(confusion_matrix)\n<\/code><\/pre>\n\n<p>To get the per-class accuracy:<\/p>\n\n<pre><code>print(confusion_matrix.diag()\/confusion_matrix.sum(1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53291571,"AnswerCount":1,"Tags":"<python><tensorflow><cuda><deep-learning>","CreationDate":"2018-11-14T00:43:58.837","AcceptedAnswerId":null,"OwnerUserId":4104334.0,"Title":"Do I need to install CUDA driver for tensorflow-gpu manually if I install tf through conda","Body":"<p>I followed this tutorial and installed tf-gpu using conda (<a href=\"https:\/\/www.pugetsystems.com\/labs\/hpc\/The-Best-Way-to-Install-TensorFlow-with-GPU-Support-on-Windows-10-Without-Installing-CUDA-1187\/\" rel=\"nofollow noreferrer\">https:\/\/www.pugetsystems.com\/labs\/hpc\/The-Best-Way-to-Install-TensorFlow-with-GPU-Support-on-Windows-10-Without-Installing-CUDA-1187\/<\/a>) and it worked because I am seeing \"...gpu:0\" in my printed out log. Before I did the installation, I already have CUDA driver installed, so I am not sure.<\/p>\n\n<p>Seems to me that conda install tensorflow-gpu comes with cuda toolkit and cuDNN,etc. I was wondering if installing CUDA driver is a require step. Another post I found did't mention driver either (<a href=\"https:\/\/towardsdatascience.com\/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc<\/a>). But the official GPU guide says it's required, so I am confused. I am doing it on Windows 10.<\/p>\n","answers":[{"AnswerId":"53295441","CreationDate":"2018-11-14T08:00:22.490","ParentId":null,"OwnerUserId":"10298883","Title":null,"Body":"<p>In my experience you do not need to install cuda or cudnn. Just your graphics driver is enough.  <\/p>\n\n<p>But depending on your system it might not be optimized. For that you would need to compile tensorflow from scratch and optimize it for your system. <\/p>\n"}]}
+{"QuestionId":53291738,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T01:05:35.887","AcceptedAnswerId":null,"OwnerUserId":8201451.0,"Title":"tf.image.decode_jpeg returns wrong image","Body":"<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport cv2 \nimport numpy as np\nsess = tf.Session()\n\n\njpg1 = cv2.imread(\"test2.jpg\") \n\njpg2 = tf.read_file(\"test2.jpg\")\njpg2 = tf.image.decode_jpeg(jpg2) \njpg2 = sess.run(jpg2) \n\n\ncv2.imshow('jpg1', jpg1)\ncv2.waitKey(0)\ncv2.destroyAllWindows()\n\ncv2.imshow('jpg2', jpg2)\ncv2.waitKey(0)\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fwSPG.jpg\" rel=\"nofollow noreferrer\">original<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/UhTSQ.jpg\" rel=\"nofollow noreferrer\">tensorflow's reconstruction<\/a><\/p>\n\n<p>Total difference between pixels on original vs. reconstructed: 12563624 <\/p>\n\n<p>I am trying to load images into tensorflow, but the pixel differences are obviously too large and the color on the reconstructed image isn't true to the original image.<\/p>\n","answers":[]}
+{"QuestionId":53291939,"AnswerCount":1,"Tags":"<tensorflow><tensor2tensor>","CreationDate":"2018-11-14T01:34:31.173","AcceptedAnswerId":null,"OwnerUserId":1018733.0,"Title":"How do I use tensor2tensor's distillation.py to distill the knowledge from a teacher network to student network?","Body":"<p><strong>Top Level Problem<\/strong><\/p>\n\n<p>I want to use a teacher network and distill its performance\/knowledge on a small subset of its power to another simpler model<\/p>\n\n<p><strong>Attempted solution<\/strong><\/p>\n\n<p>I am trying to get started with the T2T distillation code.\n<a href=\"https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/models\/distillation.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/models\/distillation.py<\/a><\/p>\n\n<p><strong>Problems with Attempted solution<\/strong><\/p>\n\n<p>I am struggling to understand how to use it for my teacher and my student. \nAre there any examples that show how it works I could get running first? \nHow do I get it work on an existing model in T2T? \nHow do I get it to work on a model defined in Keras?<\/p>\n\n<p>I understand how this block is what I'm supposed to use to register my model in T2T. But how do I start the training? I searched github for <code>distill_resnet_32_to_15_cifar20x5<\/code>, and the only hits were duplicate forks of the T2T repo with no examples of how to use this.<\/p>\n\n<pre><code>@registry.register_hparams\ndef distill_resnet_32_to_15_cifar20x5():\n  \"\"\"Set of hyperparameters.\"\"\"\n  hparams = distill_base()\n  hparams.teacher_model = \"resnet\"\n  hparams.teacher_hparams = \"resnet_cifar_32\"\n  hparams.student_model = \"resnet\"\n  hparams.student_hparams = \"resnet_cifar_15\"\n\n  hparams.optimizer_momentum_nesterov = True\n  # (base_lr=0.1) * (batch_size=128*8 (on TPU, or 8 GPUs)=1024) \/ (256.)\n  hparams.teacher_learning_rate = 0.25 * 128. * 8. \/ 256.\n  hparams.student_learning_rate = 0.2 * 128. * 8. \/ 256.\n  hparams.learning_rate_decay_scheme = \"piecewise\"\n  hparams.add_hparam(\"learning_rate_boundaries\", [40000, 60000, 80000])\n  hparams.add_hparam(\"learning_rate_multiples\", [0.1, 0.01, 0.001])\n\n  hparams.task_balance = 0.28\n  hparams.distill_temperature = 2.0\n\n  hparams.num_classes = 20\n\n  return hparams\n<\/code><\/pre>\n\n<p><strong>Full Code: distillation.py<\/strong><\/p>\n\n<pre><code># coding=utf-8\n# Copyright 2018 The Tensor2Tensor Authors.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#     http:\/\/www.apache.org\/licenses\/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n\"\"\"Traditional Student-Teacher Distillation.\"\"\"\n\nfrom __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\nfrom tensor2tensor.layers import common_hparams\nfrom tensor2tensor.utils import registry\nfrom tensor2tensor.utils import t2t_model\n\nimport tensorflow as tf\n\n\n@registry.register_model\nclass Distillation(t2t_model.T2TModel):\n  \"\"\"Distillation from a teacher to student network.\n  First, a teacher is trained on a task; Second, a student is trained to perform\n  the task while matching the teacher's softened outputs. For more details, see\n  the paper below.\n  In the hparams passed to this model include the desired\n  {teacher\/student}_model and {teacher\/student}_hparams to be used. Also,\n  specify the distillation temperature and task-distillation balance.\n  Distilling the Knowledge in a Neural Network\n  Hinton, Vinyals and Dean\n  https:\/\/arxiv.org\/abs\/1503.02531\n  \"\"\"\n\n  def __init__(self,\n               hparams,\n               mode=tf.estimator.ModeKeys.TRAIN,\n               problem_hparams=None,\n               data_parallelism=None,\n               decode_hparams=None):\n    assert hparams.distill_phase in [\"train\", \"distill\"]\n\n    if hparams.distill_phase == \"train\" and hparams.teacher_learning_rate:\n      hparams.learning_rate = hparams.teacher_learning_rate\n    elif hparams.distill_phase == \"distill\" and hparams.student_learning_rate:\n      hparams.learning_rate = hparams.student_learning_rate\n\n    self.teacher_hparams = registry.hparams(hparams.teacher_hparams)\n    self.teacher_model = registry.model(\n        hparams.teacher_model)(self.teacher_hparams, mode, problem_hparams,\n                               data_parallelism, decode_hparams)\n    self.student_hparams = registry.hparams(hparams.student_hparams)\n    self.student_model = registry.model(\n        hparams.student_model)(self.student_hparams, mode, problem_hparams,\n                               data_parallelism, decode_hparams)\n    super(Distillation, self).__init__(hparams, mode, problem_hparams,\n                                       data_parallelism, decode_hparams)\n\n  def body(self, features):\n    hp = self.hparams\n    is_distill = hp.distill_phase == \"distill\"\n\n    targets = features[\"targets_raw\"]\n    targets = tf.squeeze(targets, [1, 2, 3])\n    one_hot_targets = tf.one_hot(targets, hp.num_classes, dtype=tf.float32)\n\n    # Teacher Network\n    with tf.variable_scope(\"teacher\"):\n      teacher_outputs = self.teacher_model.body(features)\n      tf.logging.info(\"teacher output shape: %s\" % teacher_outputs.get_shape())\n      teacher_outputs = tf.reduce_mean(teacher_outputs, axis=[1, 2])\n      teacher_logits = tf.layers.dense(teacher_outputs, hp.num_classes)\n\n      teacher_task_xent = tf.nn.softmax_cross_entropy_with_logits_v2(\n          labels=one_hot_targets, logits=teacher_logits)\n      outputs = teacher_logits\n\n    if is_distill:\n      # Load teacher weights\n      tf.train.init_from_checkpoint(hp.teacher_dir, {\"teacher\/\": \"teacher\/\"})\n      # Do not train the teacher\n      trainable_vars = tf.get_collection_ref(tf.GraphKeys.TRAINABLE_VARIABLES)\n      del trainable_vars[:]\n\n    # Student Network\n    if is_distill:\n      with tf.variable_scope(\"student\"):\n        student_outputs = self.student_model.body(features)\n        tf.logging.info(\n            \"student output shape: %s\" % student_outputs.get_shape())\n        student_outputs = tf.reduce_mean(student_outputs, axis=[1, 2])\n        student_logits = tf.layers.dense(student_outputs, hp.num_classes)\n\n        student_task_xent = tf.nn.softmax_cross_entropy_with_logits_v2(\n            labels=one_hot_targets, logits=student_logits)\n        teacher_targets = tf.nn.softmax(teacher_logits \/ hp.distill_temperature)\n        student_distill_xent = tf.nn.softmax_cross_entropy_with_logits_v2(\n            labels=tf.stop_gradient(teacher_targets), logits=student_logits)\n\n        outputs = student_logits\n\n        # Summaries\n        tf.summary.scalar(\"distill_xent\", student_distill_xent)\n\n    if not is_distill:\n      phase_loss = teacher_task_xent\n    else:\n      phase_loss = hp.task_balance * student_task_xent\n      phase_loss += (1 - hp.task_balance) * student_distill_xent\n\n    losses = {\"training\": phase_loss}\n    outputs = tf.reshape(outputs, [-1, 1, 1, 1, outputs.shape[1]])\n\n    return outputs, losses\n\n  def top(self, body_output, features):\n    return body_output\n\n\ndef distill_base():\n  \"\"\"Set of hyperparameters.\"\"\"\n  # Base\n  hparams = common_hparams.basic_params1()\n\n  # teacher\/student parameters\n  hparams.add_hparam(\"teacher_model\", \"\")\n  hparams.add_hparam(\"teacher_hparams\", \"\")\n  hparams.add_hparam(\"student_model\", \"\")\n  hparams.add_hparam(\"student_hparams\", \"\")\n\n  # Distillation parameters\n  # WARNING: distill_phase hparam will be overwritten in \/bin\/t2t_distill.py\n  hparams.add_hparam(\"distill_phase\", None)\n  hparams.add_hparam(\"task_balance\", 1.0)\n  hparams.add_hparam(\"distill_temperature\", 1.0)\n  hparams.add_hparam(\"num_classes\", 10)\n\n  # Optional Phase-specific hyperparameters\n  hparams.add_hparam(\"teacher_learning_rate\", None)\n  hparams.add_hparam(\"student_learning_rate\", None)\n\n  # Training parameters (stolen from ResNet)\n  hparams.batch_size = 128\n  hparams.optimizer = \"Momentum\"\n  hparams.optimizer_momentum_momentum = 0.9\n  hparams.optimizer_momentum_nesterov = True\n  hparams.weight_decay = 1e-4\n  hparams.clip_grad_norm = 0.0\n  # (base_lr=0.1) * (batch_size=128*8 (on TPU, or 8 GPUs)=1024) \/ (256.)\n  hparams.learning_rate = 0.4\n  hparams.learning_rate_decay_scheme = \"cosine\"\n  # For image_imagenet224, 120k training steps, which effectively makes this a\n  # cosine decay (i.e. no cycles).\n  hparams.learning_rate_cosine_cycle_steps = 120000\n  hparams.initializer = \"normal_unit_scaling\"\n  hparams.initializer_gain = 2.\n\n  return hparams\n\n\n@registry.register_hparams\ndef distill_resnet_32_to_15_cifar20x5():\n  \"\"\"Set of hyperparameters.\"\"\"\n  hparams = distill_base()\n  hparams.teacher_model = \"resnet\"\n  hparams.teacher_hparams = \"resnet_cifar_32\"\n  hparams.student_model = \"resnet\"\n  hparams.student_hparams = \"resnet_cifar_15\"\n\n  hparams.optimizer_momentum_nesterov = True\n  # (base_lr=0.1) * (batch_size=128*8 (on TPU, or 8 GPUs)=1024) \/ (256.)\n  hparams.teacher_learning_rate = 0.25 * 128. * 8. \/ 256.\n  hparams.student_learning_rate = 0.2 * 128. * 8. \/ 256.\n  hparams.learning_rate_decay_scheme = \"piecewise\"\n  hparams.add_hparam(\"learning_rate_boundaries\", [40000, 60000, 80000])\n  hparams.add_hparam(\"learning_rate_multiples\", [0.1, 0.01, 0.001])\n\n  hparams.task_balance = 0.28\n  hparams.distill_temperature = 2.0\n\n  hparams.num_classes = 20\n\n  return hparams\n<\/code><\/pre>\n","answers":[{"AnswerId":"54501956","CreationDate":"2019-02-03T10:32:28.073","ParentId":null,"OwnerUserId":"4975126","Title":null,"Body":"<p>Have a look at the training script here: <a href=\"https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/bin\/t2t_distill.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/bin\/t2t_distill.py<\/a><\/p>\n\n<p>To run the example you should just need to checkout the T2T repo and run:\n<code>python bin\/t2t_distill.py --model=distillation --hparams= distill_resnet_32_to_15_cifar20x5 --problem=image_cifar100 etc...<\/code><\/p>\n\n<p>My apologies for the lack of documentation, I would welcome your contribution on cleaning things up and making it easier for people to figure out how to use this.<\/p>\n"}]}
+{"QuestionId":53292093,"AnswerCount":3,"Tags":"<tensorflow><bazel>","CreationDate":"2018-11-14T01:54:36.523","AcceptedAnswerId":null,"OwnerUserId":6586992.0,"Title":"How do I specify a \"toolchain_identifier\" when building tensorflow from source","Body":"<p>I am building tensorflow from source in order to use the GPU version with an older card with a compute capability of 3.0.<\/p>\n\n<p>When building, I get an error:<\/p>\n\n<pre><code>ERROR: \/home\/[user]\/.cache\/bazel\/_bazel_[user]\/35191c369325bea6db75133a187a58d6\/external\/local_config_cc\/BUILD:57:1: in cc_toolchain rule @local_config_cc\/\/:cc-compiler-k8: Error while selecting cc_toolchain: Toolchain identifier 'local' was not found, valid identifiers are [local_linux, local_darwin, local_windows]\n<\/code><\/pre>\n\n<p>I worked around this by hand editing <code>~\/.cache\/bazel\/_bazel_[user]\/35191c369325bea6db75133a187a58d6\/external\/local_config_cc\/BUILD<\/code><\/p>\n\n<p>to change the toolchain_identifier from \"local\" to \"local_linux\" under cc_toolchain.<\/p>\n\n<p>With that change, everything compiles. But, that seems unconventional to me.<\/p>\n\n<p>Is there something should be specifying elsewhere so that bazel gets the identifier correct on its own?<\/p>\n","answers":[{"AnswerId":"53329021","CreationDate":"2018-11-15T22:54:30.483","ParentId":null,"OwnerUserId":"8800732","Title":null,"Body":"<p>open \/home\/[user]\/.cache\/bazel\/_bazel_jeff\/35191c369325bea6db75133a187a58d6\/external\/local_config_cc\/BUILD with any text editor and change the line 57 as local_linux<\/p>\n"},{"AnswerId":"53334117","CreationDate":"2018-11-16T08:35:52.263","ParentId":null,"OwnerUserId":"10661771","Title":null,"Body":"<p>I got the same error in building tensorflow r1.9 for one older Nvidia GPU card. I downgraded the bazel from 0.19.1 to 0.18.1. The error was fixed in compiling.<\/p>\n"},{"AnswerId":"53293037","CreationDate":"2018-11-14T04:04:15.960","ParentId":null,"OwnerUserId":"3723596","Title":null,"Body":"<p>Not sure this is related but ... I was having the same problem, tried a bunch of things that did not work, including alternating between clang and gcc, and then told configure I was using cudnn 7.2 instead of just 7 and it worked after that.<\/p>\n"}]}
+{"QuestionId":53292107,"AnswerCount":1,"Tags":"<cuda><anaconda><pytorch>","CreationDate":"2018-11-14T01:57:24.897","AcceptedAnswerId":null,"OwnerUserId":9066532.0,"Title":"When using Pytorch-GPU in Anaconda, is it not necessary to install CUDA?","Body":"<p>I found that after installing Pytorch 0.4 GPU version in Anaconda, you don't need to install CUDA locally to call gpu acceleration. When running code, the GPU core can be used at more than 90%.<\/p>\n\n<p>Edit:I used it in Windows 10. Don't know if it works in Linux.<\/p>\n","answers":[{"AnswerId":"53335414","CreationDate":"2018-11-16T09:59:27.943","ParentId":null,"OwnerUserId":"9066532","Title":null,"Body":"<p>@talonmies<\/p>\n\n<p>Thanks for your url. It seems that pytorch don't need cuda in Windows, since its dependencies are cffi, mkl, numpy, and python.<\/p>\n\n<p>I entered this command <code>conda search -c pytorch pytorch=0.4.0 --info<\/code> in Anaconda Prompt and it says<\/p>\n\n<pre><code>Loading channels: done\npytorch 0.4.0 py35_cuda80_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py35_cuda80_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py35_cuda80_cudnn7he774522_1\nbuild number: 1\nsize        : 528.5 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py35_cuda80_cudnn7he774522_1.tar.bz2\nmd5         : 7db3971bb054079d7c7ff84b6286c58e\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.5,&lt;3.6.0a0\n\n\npytorch 0.4.0 py35_cuda90_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py35_cuda90_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py35_cuda90_cudnn7he774522_1\nbuild number: 1\nsize        : 578.5 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py35_cuda90_cudnn7he774522_1.tar.bz2\nmd5         : 8200c9841f9cad6f2e605015812aa3f2\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.5,&lt;3.6.0a0\n\n\npytorch 0.4.0 py35_cuda91_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py35_cuda91_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py35_cuda91_cudnn7he774522_1\nbuild number: 1\nsize        : 546.1 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py35_cuda91_cudnn7he774522_1.tar.bz2\nmd5         : 79d99a825f66b55b1aa6f04d22d68aac\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.5,&lt;3.6.0a0\n\n\npytorch 0.4.0 py36_cuda80_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py36_cuda80_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py36_cuda80_cudnn7he774522_1\nbuild number: 1\nsize        : 529.2 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py36_cuda80_cudnn7he774522_1.tar.bz2\nmd5         : 27d20c9869fb57ffe0d6d014cf348855\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.6,&lt;3.7.0a0\n\n\npytorch 0.4.0 py36_cuda90_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py36_cuda90_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py36_cuda90_cudnn7he774522_1\nbuild number: 1\nsize        : 577.6 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py36_cuda90_cudnn7he774522_1.tar.bz2\nmd5         : 138dcca8eeff1d58a8fd9b1febf702f6\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.6,&lt;3.7.0a0\n\n\npytorch 0.4.0 py36_cuda91_cudnn7he774522_1\n------------------------------------------\nfile name   : pytorch-0.4.0-py36_cuda91_cudnn7he774522_1.tar.bz2\nname        : pytorch\nversion     : 0.4.0\nbuild string: py36_cuda91_cudnn7he774522_1\nbuild number: 1\nsize        : 546.4 MB\narch        : x86_64\nconstrains  : ()\nplatform    : Platform.win\nlicense     : BSD 3-Clause\nsubdir      : win-64\nurl         : https:\/\/conda.anaconda.org\/pytorch\/win-64\/pytorch-0.4.0-py36_cuda91_cudnn7he774522_1.tar.bz2\nmd5         : 326265665000de6f7501160b10b089c8\ndependencies:\n  - cffi\n  - mkl &gt;=2018\n  - numpy &gt;=1.11\n  - python &gt;=3.6,&lt;3.7.0a0\n<\/code><\/pre>\n"}]}
+{"QuestionId":53292289,"AnswerCount":0,"Tags":"<python-3.x><keras><neural-network>","CreationDate":"2018-11-14T02:22:00.173","AcceptedAnswerId":null,"OwnerUserId":9808963.0,"Title":"Using a Custom Dual Input Keras Layer","Body":"<p>I am attempting to create a simple multi-layer perceptron in Keras. The general structure I would like to create is one where a matrix A of dimension [n_a1, n_a2] is sent through a number of layers of a multilayer perceptron, and at a certain point, the dot product of the morphed A matrix is taken with a randomly selected y vector [n_y, 1] from a set of y vectors, and the result then continues through a number more layers before reaching the end where it is compared with the input labels as normal.<\/p>\n\n<p>Unfortunately, I am having issues with figuring how to implement this properly. I created a custom multi-input layer per the simple example offered at <a href=\"https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/<\/a>, but it seems that I still can't have multiple inputs in the network. I am getting an error for setting an array element as a sequence: <code>ValueError: setting an array element with a sequence.<\/code><\/p>\n\n<p>Also, its unclear to me how the network would know what to do with the list formatted inputs for the other layers. Do I need to specify the list shape for each layer, and to somehow only use the A in the [A, y] list?<\/p>\n\n<p><strong>Custom Layer<\/strong><\/p>\n\n<pre><code>class DualLayer(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(DualLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        #Trainable weight variable for layer\n        self.kernel = self.add_weight(name='kernel',\n                                     shape=(input_shape[1], self.output_dim),\n                                     initializer='uniform',\n                                     trainable=True)\n        super(DualLayer, self).build(input_shape)\n\n    def call(self, x):\n        aOpt, y = x\n        return [K.dot(aOpt, y)]\n\n    def compute_outpute_shape(self, input_shape):\n        assert isinstance(input_shape, list)\n        shape_y, shape_aOpt = input_shape\n        return [shape_y[0]]\n<\/code><\/pre>\n\n<p><strong>Model<\/strong><\/p>\n\n<pre><code>def modFunc(y1, A1, y2, A2, xSim):\n    model = Sequential()\n    model.add(Flatten(input_shape = np.shape(A1)))\n    model.add(Dense(np.shape(A1)[0]*np.shape(A1)[1], activation = 'relu',\n                    kernel_regularizer = 'l1', activity_regularizer = 'l1'))\n    model.add(DualLayer([y1, A1]))\n    model.add(Dense(5, activation = 'relu'))\n\n    clf = model.fit([y1, A1], xSim, epochs=5, batch_size=1, verbose=2)\n    return clf.coef_\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53293243,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T04:33:52.300","AcceptedAnswerId":53300634.0,"OwnerUserId":9357193.0,"Title":"Reduce the rank of a one one tensor of Tensorflow?","Body":"<p>I have a Tensorflow tensor [[1]], how can I change this tensor to [1] which is just a number? <\/p>\n\n<p>Thank you so much!<\/p>\n","answers":[{"AnswerId":"53300634","CreationDate":"2018-11-14T12:48:58.530","ParentId":null,"OwnerUserId":"8889727","Title":null,"Body":"\n\n<p>Here I create a dummy 2D tensor and convert it to the scalar you are looking for. Note that your could run this inside a session, <code>tf.Session().run(tensor)<\/code> instead of the following <code>InteractiveSession<\/code>.<\/p>\n\n<p>One thing to note here is that any returned tensor of <code>Session.run<\/code> or <code>eval<\/code> is a numpy array. If the returned array is multi dimensional, you could call <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.0\/reference\/generated\/numpy.ravel.html\" rel=\"nofollow noreferrer\">np.ravel()<\/a> to make it 1D and take the 1st element to make it a scalar.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; sess = tf.InteractiveSession()\n&gt;&gt;&gt; t=[[10]]\n&gt;&gt;&gt; tensor=tf.convert_to_tensor(t)\n&gt;&gt;&gt; tensor.eval().ravel()[0]\n10\n<\/code><\/pre>\n"}]}
+{"QuestionId":53293798,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T05:37:38.713","AcceptedAnswerId":null,"OwnerUserId":10628840.0,"Title":"iterating tensor objects in tf.data api without eager execution","Body":"<p>data.Dataset api for converting my data into tensorflow objects and the code is <\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((input_tensor_train, target_tensor_train)).shuffle(BUFFER_SIZE)\n\ndataset = dataset.batch(BATCH_SIZE, drop_remainder=True)\n<\/code><\/pre>\n\n<p>in eager execution it is working fine with enumerate. But i like to iterator over the data without eager execution and i found some resources regarding this then i used <\/p>\n\n<pre><code>dataset.make_initializable_iterator()\niterator.get_next()\n<\/code><\/pre>\n\n<p>to iterate over the dataset.<\/p>\n\n<p>But its giving the error <\/p>\n\n<blockquote>\n  <p>Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.<\/p>\n<\/blockquote>\n\n<p>could anybody is there to help me to solve the error<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/eager\/python\/examples\/nmt_with_attention\/nmt_with_attention.ipynb\" rel=\"nofollow noreferrer\">and this is the original source code unchanged<\/a><\/p>\n","answers":[]}
+{"QuestionId":53294415,"AnswerCount":1,"Tags":"<python><docker><tensorflow><dockerfile><tensorflow-serving>","CreationDate":"2018-11-14T06:38:56.537","AcceptedAnswerId":null,"OwnerUserId":7462428.0,"Title":"What is the difference between tensorflow serving Dockerfile and Dockerfile.devel?","Body":"<p>Why are there two different docker files for tensorflow serving - Dockerfile &amp; Dockerfile.devel - for both CPU and GPUs?<\/p>\n\n<p>Which one is necessary for deploying and testing?<\/p>\n","answers":[{"AnswerId":"53308093","CreationDate":"2018-11-14T20:14:35.583","ParentId":null,"OwnerUserId":"6713869","Title":null,"Body":"<p>a Dockerfile is a file where your write the configurations to create a docker image. <br><\/p>\n\n<p>The tensorflow\/serving cpu and gpu are docker images which means they are already configured to work with tensorflow, tensorflow_model_server and, in the case of gpu, with CUDA. <br><\/p>\n\n<p>If you have a GPU, then you can use a tensorflow\/serving gpu version which would reduce the latency of your predictions. If you don't have a GPU, then you can use a tensorflow\/serving cpu version which would do exactly the same but will be slower.<\/p>\n"}]}
+{"QuestionId":53294681,"AnswerCount":2,"Tags":"<python-3.x><keras><typeerror><chatbot><rasa-core>","CreationDate":"2018-11-14T06:59:23.447","AcceptedAnswerId":53349798.0,"OwnerUserId":9214835.0,"Title":"RASA Chatbot Framework gives error while training :fit() got multiple values for keyword argument 'batch_size'","Body":"<p>I was trying to train the Chat Bot built with rasa using the snippet<\/p>\n\n<pre><code>agent.train(data,augmentation_factor=50,\n        epochs=500,\n        batch_size=10)\n<\/code><\/pre>\n\n<p>and got the following error. I Know its not rasa_core error but something related to keras probably. <\/p>\n\n<pre><code>rasa_core\/policies\/keras_policy.py\", line 177, in train\n    **params)\nTypeError: fit() got multiple values for keyword argument 'batch_size'\n<\/code><\/pre>\n\n<p>I am new to bot building and never done in deep learning Project before that.Using following dependencies\nrasa_core==0.12.x\nkeras==2.1.6<\/p>\n","answers":[{"AnswerId":"53349798","CreationDate":"2018-11-17T09:13:47.017","ParentId":null,"OwnerUserId":"10665033","Title":null,"Body":"<p>I had the same error and after several search I did find a solution. It is not the best solution but it can help!<\/p>\n\n<p>Go to rasa_core\\rasa_core\\policies\\keras_policy.py in lines 172-177 and delete epochs and batch_size arguemnts from model.fit.\nChange this:<\/p>\n\n<pre><code>    params = self._get_valid_params(self.model.fit, **kwargs)\n    self.model.fit(shuffled_X, shuffled_y,\n                   epochs=self.epochs, batch_size=self.batch_size,\n                   **params)\n<\/code><\/pre>\n\n<p>to this :<\/p>\n\n<pre><code>    params = self._get_valid_params(self.model.fit, **kwargs)\n    self.model.fit(shuffled_X, shuffled_y,**params)\n<\/code><\/pre>\n\n<p>Then you can pass epochos and batch_size arguments in agent.train()<\/p>\n\n<p>I hope it will help!<\/p>\n"},{"AnswerId":"53441452","CreationDate":"2018-11-23T06:13:10.200","ParentId":null,"OwnerUserId":"3708401","Title":null,"Body":"<p>My friend had the same issue. He has latest version of rasa_core installed. However the keras_policy.py got similar issue. I gave him my version of keras_policy.py and he replaced it with my version and it worked. I know it is not right solution, but you can use that as work around<\/p>\n"}]}
+{"QuestionId":53294987,"AnswerCount":0,"Tags":"<tensorflow><object-detection><object-detection-api>","CreationDate":"2018-11-14T07:23:01.803","AcceptedAnswerId":null,"OwnerUserId":2825492.0,"Title":"Exception in quantization aware training for faster rcnn","Body":"<p>I am getting  <\/p>\n\n<pre><code>File \"\/home\/ubuntu\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/quantize\/python\/fold_batch_norms.py\", line 338, in _ComputeBatchNormCorrections\n    match.moving_variance_tensor + match.batch_epsilon)\nTypeError: unsupported operand type(s) for +: 'NoneType' and 'float'\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZaZOv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZaZOv.png\" alt=\"enter image description here\"><\/a>\nException while doing quantization aware training using tensorflow object detection api on faster rcnn architecture.\nI was able to succesfully get it working for ssd mobilenet architecture. I am using TF 1.11.0 on GPU and master branch of object detection api.\nI added <\/p>\n\n<pre><code>graph_rewriter {\n  quantization {\n    delay: 48\n    weight_bits: 8\n    activation_bits: 8\n  }\n}\n<\/code><\/pre>\n\n<p>to my pipeline.config file for quantization. I tried to set --fine_tune_batch_norm=true but it didn't help.<\/p>\n","answers":[]}
+{"QuestionId":53295253,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-14T07:44:56.010","AcceptedAnswerId":null,"OwnerUserId":7587433.0,"Title":"When use Dataset API, got device placement error with tensorflow >= 1.11","Body":"<p>My script used dataset API to realize input pipeline. When I use tensorflow 1.10, everything is ok, but when I upgrade to tensorflow 1.11 or 1.12. I got the following error message at the beginning of training:<\/p>\n\n<p>Attempted create an iterator on device \"\/job:localhost\/replica:0\/task:0\/device:GPU:0\" from handle defined on device \"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"<\/p>\n\n<p>Here are piece of my code:<\/p>\n\n<pre><code>def build_all_dataset(self):\n    self.build_dataset(ROUTE_TRAIN)\n    self.build_dataset(ROUTE_VALIDATION)\n    self.build_dataset(ROUTE_TEST)\n\ndef build_dataset(self, p_route):\n    # Create dataset instance.\n    # ......\n    # ......\n\n    self.dataset[p_route] = dataset\n    self.iterators[p_route] = dataset.make_initializable_iterator()\n    self.handles[p_route] = (self.iterators[p_route].string_handle()).eval()\n\ndef build_model(self):\n    with tf.device(\"\/GPU:0\"):\n        # Build dataset for training\/validation\/test.\n        self.build_all_dataset()\n        self.ph_dataset_handle = tf.placeholder(tf.string, shape=[])\n        iterator = tf.data.Iterator.from_string_handle(self.ph_dataset_handle, self.dataset[ROUTE_TRAIN].output_types,\n                                                       self.dataset[ROUTE_TRAIN].output_shapes)\n\n        xxx, xxx, xxx = iterator.get_next()\n        # Build the following graph.\n        # ......\n        # ......\n\nwhile True:\n    try:\n        session.run(xxx, {self.ph_dataset_handle: self.handles[ROUTE_TRAIN]})\n    except tf.errors.OutOfRangeError:\n        break\n<\/code><\/pre>\n\n<p>I checked that the placeholder \"self.ph_dataset_handle\" is placed on \/GPU:0, why tensorflow said \"handle defined on device\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"?<\/p>\n\n<p>Could you please give any insight? Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53295272,"AnswerCount":1,"Tags":"<python><keras><neural-network><deep-learning><lstm>","CreationDate":"2018-11-14T07:46:04.343","AcceptedAnswerId":null,"OwnerUserId":3625281.0,"Title":"Use both losses on a subnetwork of combined networks","Body":"<p>I am trying to stack two networks together. I want to calculate loss of each network separately. For example in the image below; loss of LSTM1 should be <em>(Loss1 + Loss2)<\/em> and loss of system should be just <em>(Loss2)<\/em><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SscXv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SscXv.png\" alt=\"stacked networks\"><\/a><\/p>\n\n<p>I implemented a network like below with the idea above but have no idea how to compile and run it.<\/p>\n\n<pre><code>def build_lstm1():\n    x = Input(shape=(self.timesteps, self.input_dim,), name = 'input')\n    h = LSTM(1024, return_sequences=True))(x)\n    scores = TimeDistributed(Dense(self.input_dim, activation='sigmoid', name='dense'))(h)\n    LSTM1 = Model(x, scores)\n    return LSTM1\n\n\ndef build_lstm2():\n    x = Input(shape=(self.timesteps, self.input_dim,), name = 'input')\n    h = LSTM(1024, return_sequences=True))(x)\n    labels = TimeDistributed(Dense(self.input_dim, activation='sigmoid', name='dense'))(h)\n    LSTM2 = Model(x, labels)\n    return LSTM2\n\n\n\n\nlstm1 = build_lstm1()\nlstm2 = build_lstm2()\n\n\ncombined = Model(inputs = lstm1.input ,\n                 outputs = [lstm1.output,\n                            lstm2(lstm1.output).output)])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53297883","CreationDate":"2018-11-14T10:22:18.390","ParentId":null,"OwnerUserId":"6670629","Title":null,"Body":"<p>This is a wrong way of using the Model functional API of Keras. Also it's not possible to have loss of LSTM1 as Loss1+ Loss2. It will be only Loss1. Similarly for LSTM2 it will be only Loss2. However, for the combined network you can have any linear combination of Loss1 and Loss2 as the overall Loss i.e. <\/p>\n\n<p>Loss_overall = a.Loss1 + b.Loss2. where a,b are non-negative real numbers<\/p>\n\n<p>The real essence of Model Functional API is that it allows you create Deep learning architecture with multiple outputs and multiple inputs in single model.     <\/p>\n\n<pre><code>def build_lstm_combined():\n    x = Input(shape=(self.timesteps, self.input_dim,), name = 'input')\n    h_1 = LSTM(1024, return_sequences=True))(x)\n    scores = TimeDistributed(Dense(self.input_dim, activation='sigmoid', name='dense'))(h_1)\n    h_2 = LSTM(1024, return_sequences=True))(h_1)\n    labels = TimeDistributed(Dense(self.input_dim, activation='sigmoid', name='dense'))(h_2)\n    LSTM_combined = Model(x,[scores,labels])\n    return LSTM_combined\n<\/code><\/pre>\n\n<p>This combined model has loss which is combination of both Loss1 and Loss2. While compiling the model you can specify the weights of each loss to get overall loss. If your desired loss is 0.5Loss1 + Loss2 you can do this by:<\/p>\n\n<pre><code>model_1 = build_lstm_combined()\nmodel_1.compile(optimizer=Adam(0.001), loss = ['categorical_crossentropy','categorical_crossentropy'],loss_weights= [0.5,1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53295570,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-14T08:09:14.457","AcceptedAnswerId":null,"OwnerUserId":8846515.0,"Title":"UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually","Body":"<p>After a training procedure, I wanted to check the accuracy by loading the created <code>model.h5<\/code> and executing an evaluation procedure. However, I am getting a following warning:<\/p>\n\n<blockquote>\n  <p>\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/saving.py:269:\n  UserWarning: No training configuration found in save file: the model\n  was <em>not<\/em> compiled. Compile it manually.   warnings.warn('No training\n  configuration found in save file:<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qSbNC.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qSbNC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This <code>dist-packages\/keras\/engine\/saving.py<\/code> file<\/p>\n\n<p>so the problem in loading created model -> this line of code<\/p>\n\n<pre><code>train_model = load_model('model.h5')\n<\/code><\/pre>\n\n<p>Problem indicates that the model was not compiled, however, I did it.<\/p>\n\n<pre><code>optimizer = Adam(lr=lr, clipnorm=0.001)\ntrain_model.compile(loss=dummy_loss, optimizer=optimizer)\n<\/code><\/pre>\n\n<p>I can't understand what I am doing wrong . . .\nPlease help me! SOS :-(<\/p>\n","answers":[{"AnswerId":"54180527","CreationDate":"2019-01-14T11:21:18.623","ParentId":null,"OwnerUserId":"1517969","Title":null,"Body":"<p>Do you get this warning when saving the model?<\/p>\n\n<pre><code>WARNING:tensorflow:TensorFlow optimizers do not make it possible to access \noptimizer attributes or optimizer state after instantiation. As a result, we \ncannot save the optimizer as part of the model save file.You will have to \ncompile your model again after loading it. Prefer using a Keras optimizer \ninstead (see keras.io\/optimizers).\n<\/code><\/pre>\n\n<p>Seems tensorflow optimizers can't be preserved by keras :\/<\/p>\n"},{"AnswerId":"56805866","CreationDate":"2019-06-28T11:11:01.847","ParentId":null,"OwnerUserId":"11680024","Title":null,"Body":"<p>As mentioned keras can't save Tensorflow optimizers. Use the keras one:<\/p>\n\n<pre><code>optimizer = keras.optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\n\nmodel.compile(optimizer=optimizer,\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(...)\nmodel.save('...')\n<\/code><\/pre>\n\n<p>This way works for me without manual compiling after calling load.<\/p>\n"},{"AnswerId":"57054106","CreationDate":"2019-07-16T09:26:54.333","ParentId":null,"OwnerUserId":"1338797","Title":null,"Body":"<h2>Intro<\/h2>\n\n<p>I'd like to add to olejorgenb's answer - for a specific scenario, where you don't want to train the model, just use it (e.g. in production).<\/p>\n\n<p>\"Compile\" means \"prepare for training\", which includes mainly setting up the optimizer. It could also have been saved before, and then you can continue the \"same\" training after loading the saved model.<\/p>\n\n<h2>The fix<\/h2>\n\n<p>But, what about the scenario - I want to just run the model? Well, use the <code>compile=False<\/code> argument to <code>load_model<\/code> like that:<\/p>\n\n<pre><code>trained_model = load_model('model.h5', compile=False)\n<\/code><\/pre>\n\n<p>You won't be able to <code>.fit()<\/code> this model without using <code>trained_model.compile(...)<\/code> first, but most importantly - the warning will go away.<\/p>\n\n<h2>Misc Notes<\/h2>\n\n<p>Btw, in my Keras version, the argument <code>include_optimizer<\/code> has a default of <code>True<\/code>. This should work also for trainig callbacks like <code>Checkpoint<\/code>. This means, when loading a model saved by Keras, you can usually count on the optimizer being included (except for the situation: see Hull Gasper's answer).<\/p>\n\n<p>But, when you have a model which was not trained by Keras (e.g. when converting a model trained by Darknet), the model is saved un-compiled. This produces the warning, and you can get rid of it in the way described above.<\/p>\n"}]}
+{"QuestionId":53296165,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-14T08:49:39.680","AcceptedAnswerId":53296293.0,"OwnerUserId":9663229.0,"Title":"Keras: update model with a bigger training set","Body":"<p>I trained a model with Keras for text classification (supervised learning) using a training set. Let's say that there are 50.000 sentences in this training set.<\/p>\n\n<p>During a week I collect 5.000 new sentences and I add them to the old training set.<\/p>\n\n<p>If next week I want to train a new model with the new and bigger training set (50.000 old sentences + 5.000 new sentences), should I restart the training phase from the beginning, or can I take the old model and \"update\" it in some way to save some time? <\/p>\n","answers":[{"AnswerId":"53296293","CreationDate":"2018-11-14T08:57:50.490","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>You can save\/load model\/weights. Check out this <a href=\"https:\/\/machinelearningmastery.com\/save-load-keras-deep-learning-models\/\" rel=\"nofollow noreferrer\">tutorial<\/a> by Jason Brownlee. <br><\/p>\n\n<p>After you loaded the weights, you can start training with the new dataset (the 55000 samples). As the 'training' is basically just updating weights, and you loaded your trained weights, you are now 'updating' the already trained model.<\/p>\n"}]}
+{"QuestionId":53296374,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-14T09:02:34.147","AcceptedAnswerId":null,"OwnerUserId":7620627.0,"Title":"Something wrong when computing the receptive field using non-zero gradient in Keras","Body":"<p>I'm trying to compute the receptive field of some specific neurons based on the non-zero gradient but found one strange thing. <\/p>\n\n<p>The following is a simple NN model built in keras. The remaining parts are to calculate the gradient of the output (here the targeted neuron of which pos is (0,2) on the first channel) of conv2d_4 w.r.t its input. Through finding those non-zero values on the gradient map, we can easily locate the receptive field of one neuron. The ideal receptive field of one neuron in the output of conv2d_4 w.r.t its input should be 3x3 since the kernel size of conv2d_4 is 3x3, but the non-zero gradient map is a 4x5 patch (given by those TRUE values in f_sum).<\/p>\n\n<pre><code>import numpy as np\nimport keras.backend as K\nimport matplotlib.pyplot as plt\nfrom keras.models import load_model, Model\nfrom keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D,Input, AveragePooling2D, Lambda\n\ndef model_build_func(input_shape=(25,25,1)):\n    inp = Input(shape=input_shape, name='input')\n\n    x = Conv2D(32, (3,3), activation='linear', name='conv2d_1')(inp)\n    x = Conv2D(32, (3,3), activation='linear', name='conv2d_2')(x)\n    x = AveragePooling2D(pool_size=(2,2))(x)\n\n    x = Conv2D(64, (3,3), activation='linear', name='conv2d_3')(x)\n    x = Conv2D(64, (3,3), activation='linear', name='conv2d_4')(x)\n    x = AveragePooling2D(pool_size=(2,2))(x)\n\n    x = Flatten()(x)\n    x = Dense(units=64, name='dense_1')(x)\n    x = Dense(units=2, name='dense_2')(x)\n\n    model = Model(inputs=inp, outputs=x)\n    return model\n\n# used for building the Lambda layer\ndef get_mask_tensor(input_tensors, x_pos, y_pos, channel_idx):\n    mask_tensor = K.tf.gradients(input_tensors[0][:,x_pos,y_pos,channel_idx], input_tensors[1])[0]\n    return mask_tensor\n\n#specify the position of the neuron that we want to compute the RF\nx_pos = 0\ny_pos = 2\nchannel_idx = 0\n\nlayer_idx = 5 # the layer: conv2d_4\n\nmodel = model_build_func()\ncurrent_layer = model.layers[layer_idx]\n\n#get the gradient tensor\nmask_tensor = Lambda(get_mask_tensor, output_shape=K.int_shape(model.input),\n                     arguments={'x_pos':x_pos, 'y_pos':y_pos, 'channel_idx':channel_idx})([current_layer.output, current_layer.input])\n\n#create a keras model\nnew_model = Model(inputs=[model.input], outputs=[mask_tensor])\n\n#get the value of the gradient map\ngradient_map = new_model.predict(0.1*(np.random.random(size=(32,25,25,1))-0.05))\nf_sum = np.sum(np.abs(gradient_map), axis=-1)\nf_sum = np.sum(np.abs(f_sum), axis=0)\n\n#f_sum is a binary array. \n#It should has a 3x3 patch with TRUE values, but here it's 4x5\nplt.imshow(f_sum!=0)\nplt.grid()\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53296961,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-14T09:35:01.157","AcceptedAnswerId":53297197.0,"OwnerUserId":10483420.0,"Title":"site to download python full tutorials video in zip file","Body":"<h1>python and tensor flow best tutorial video<\/h1>\n\nI want to download the best python tutorial video and also tensor flow tutorial for machine learning projects\n","answers":[{"AnswerId":"53297197","CreationDate":"2018-11-14T09:47:24.380","ParentId":null,"OwnerUserId":"2335138","Title":null,"Body":"<p>Stackoverflow is really a site for answering programming or development questions, this isn't really a question.<\/p>\n\n<p>In terms of a place where you can get python and tensor flow videos, look at youtube, if you want to download them have a look at <a href=\"https:\/\/rg3.github.io\/youtube-dl\/\" rel=\"nofollow noreferrer\">youtube-dl<\/a>.<\/p>\n\n<p>If you struggle to use youtube-dl, then feel free to ask another question, but make sure you try to use it first and follow the <a href=\"https:\/\/stackoverflow.com\/help\/how-to-ask\">guidelines<\/a>.<\/p>\n"}]}
+{"QuestionId":53298021,"AnswerCount":0,"Tags":"<python><keras><categorical-data>","CreationDate":"2018-11-14T10:29:26.207","AcceptedAnswerId":null,"OwnerUserId":8767101.0,"Title":"Using Keras's Categorical Crossentropy loss the right way","Body":"<p>My model's output shape is<\/p>\n\n<pre><code>(batch_size, n_timesteps, n_outputs)\n<\/code><\/pre>\n\n<p>the last axis contains the outputs of a softmax layer, ranging over <code>n_outputs<\/code> classes.<\/p>\n\n<p>I just want to be sure that Keras' <code>categorical_crossentropy<\/code> loss and <code>categorical_accuracy<\/code> metrics are computed timestep-wise and not in some weird way (e.g. over the flattened <code>n_timesteps*n_outputs<\/code> outputs, which would be very wrong).<\/p>\n\n<p>How can I be sure of this?<\/p>\n","answers":[]}
+{"QuestionId":53298106,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><keras>","CreationDate":"2018-11-14T10:32:55.977","AcceptedAnswerId":null,"OwnerUserId":10048730.0,"Title":"How do we compare the performance of different ConvNets?","Body":"<p>I am currently training a net to play a game with a CNN having the following architecture:<br><\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(100, kernel_size=(2, 2), strides=(2, 2), activation='relu', input_shape=input_shape))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(250, activation='relu'))\nmodel.add(Dense(classifications, activation='softmax'))\nmodel.compile(loss=\"categorical_crossentropy\", optimizer=\"adam\", metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Now I wish to introduce some complexity in the architecture and make the net deep. How can I tabulate the performance of the CNNs of different complexities and ultimately conclude by giving the best choice for the particular task?<\/p>\n\n<p>Am i going in the wrong direction? How to decide the depth of a CNN and how does it affect the performance on the same dataset?<\/p>\n\n<p>Thanks in advance (I am new to this site, kindly excuse the immaturity of this post)<\/p>\n\n<p>Edit: Information about the dataset I am using: dataset consists of images and each image has 3 possible lables (0, 1, 2) stored in a CSV file with each row corresponding to that particular image.<\/p>\n","answers":[{"AnswerId":"53298871","CreationDate":"2018-11-14T11:12:03.907","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>The simplest thing you can do is generate a few different model architectures, train them on a train set and evaluate them on the test set. Then compare their accuracies and the one with the highest accuracy should in theory be the best performing model.<\/p>\n\n<p>To make the model deeper you can add extra dense or convolutional layers. For example:<\/p>\n\n<p>changing this:<\/p>\n\n<pre><code>model.add(Dense(250, activation='relu'))\n<\/code><\/pre>\n\n<p>to this:<\/p>\n\n<pre><code>model.add(Dense(250, activation='relu'))\nmodel.add(Dense(250, activation='relu'))\nmodel.add(Dense(250, activation='relu'))\n<\/code><\/pre>\n\n<p>will add three extra dense layers. Hence making the network deeper.<\/p>\n\n<p>You can do the same with duplicating the convolutional layers by duplicating the Conv2D and MaxPooling2D lines.<\/p>\n\n<p>The alternative to 'trial and error' approach to finding the best architecture and hyperparameters is to use a search approach <a href=\"https:\/\/machinelearningmastery.com\/grid-search-hyperparameters-deep-learning-models-python-keras\/\" rel=\"nofollow noreferrer\">like explained in this tutorial<\/a> that user grid search. It will, however, take significantly longer than just trying out a few versions you can up with yourself.<\/p>\n"}]}
+{"QuestionId":53299082,"AnswerCount":1,"Tags":"<machine-learning><caffe><pycaffe>","CreationDate":"2018-11-14T11:23:33.487","AcceptedAnswerId":null,"OwnerUserId":7677894.0,"Title":"Any method to limit the CPU usage of pycaffe compiled with openblas?","Body":"<p>The pycaffe is compiled with openblas.<\/p>\n\n<p>My pycaffe model occupies more than 20 CPU cores while training, any method to limit that?<\/p>\n","answers":[{"AnswerId":"53313691","CreationDate":"2018-11-15T06:33:48.263","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Try setting environment variable<\/p>\n\n<pre><code>OPENBLAS_NUM_THREADS=10\n<\/code><\/pre>\n\n<p><strong>before<\/strong> running pycaffe to restrict openblas to only 10 threads.<\/p>\n"}]}
+{"QuestionId":53299410,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-14T11:40:46.310","AcceptedAnswerId":null,"OwnerUserId":4251377.0,"Title":"DLL load failed during tensorflow import","Body":"<p>Im trying to import tensorflow in Python3.6.\nBut i keep getting this log:\nDuring handling of the above exception, another exception occurred:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"C:\\Python36_64\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"C:\\Python36_64\\lib\\imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Python36_64\\lib\\imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: The specified module could not be found.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/errors\n<\/code><\/pre>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.<\/p>\n\n<p>Does anyone have an idea whats happening?<\/p>\n","answers":[]}
+{"QuestionId":53299543,"AnswerCount":1,"Tags":"<tensorflow><theano><lasagne>","CreationDate":"2018-11-14T11:48:07.473","AcceptedAnswerId":null,"OwnerUserId":6494531.0,"Title":"Is there any method in tensorflow like get_output in lasagne","Body":"<p>I found that it is easy to use lasagne to make a graph like this.<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>import lasagne.layers as L\nclass A:\n  def __init__(self):\n    self.x = L.InputLayer(shape=(None, 3), name='x')\n    self.y = x + 1\n\n  def get_y_sym(self, x_var, **kwargs):\n    y = L.get_output(self.y, {self.x: x_var}, **kwargs)\n    return y\n<\/code><\/pre>\n\n<p>through the method <code>get_y_sym<\/code>, we could get a tensor not a value, then I could use this tensor as the input of another graph.<\/p>\n\n<p>But if I use tensorflow, how could I implement this?<\/p>\n","answers":[{"AnswerId":"53309926","CreationDate":"2018-11-14T22:52:24.030","ParentId":null,"OwnerUserId":"2426955","Title":null,"Body":"<p>I'm not familiar with lasagne but you should know that ALL of TensorFlow uses graph based computation (unless you use tf.Eager, but that's another story).  So by default something like:<\/p>\n\n<p><code>net = tf.nn.conv2d(...)<\/code> <\/p>\n\n<p>returns a reference to a Tensor object. In other words, <code>net<\/code> is NOT a value, it is a reference to the output of the convolution node created by <code>tf.nn.conv2d(...)<\/code>. <\/p>\n\n<p>These can then be chained:<\/p>\n\n<p><code>net2 = tf.nn.conv2d(net, ...)<\/code> and so on.<\/p>\n\n<p>To get \"values\" one has to open a <code>tf.Session<\/code>:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  net2_eval = sess.run(net2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53299599,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-11-14T11:50:41.787","AcceptedAnswerId":53302712.0,"OwnerUserId":10618716.0,"Title":"TensorFlow 1D model, wrong input shape for MaxPooling","Body":"<p>I'm building a 1D model with TensorFlow for audio but I have a problem with the input shape during the second MaxPool1D in the model.<\/p>\n\n<p>The problem is here, after this Pooling:<\/p>\n\n<pre><code>x = Convolution1D(32, 3, activation=relu, padding='valid')(x)\nx = MaxPool1D(4)(x)\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<p><em>ValueError: Negative dimension size caused by subtracting 4 from 1 for 'max_pooling1d_5\/MaxPool' (op: 'MaxPool') with input shapes: [?,1,1,32].<\/em><\/p>\n\n<p>I tried to reshape <em>x<\/em> (which is a tensor) but I don't think I'm going in the right way.<\/p>\n\n<p>In this same model, before that, I have a couple convolutional layers and a maxpooling that are working proporly.<\/p>\n\n<p>Anyone have suggestions?\nThanks<\/p>\n","answers":[{"AnswerId":"53302712","CreationDate":"2018-11-14T14:38:58.517","ParentId":null,"OwnerUserId":"10513982","Title":null,"Body":"<p>The number of steps in the input to the <code>MaxPool1D<\/code> layer is smaller than the pool size.  <\/p>\n\n<p>In the error, it says <code>...input shapes: [?,1,1,32]<\/code>, which means the output from the <code>Convolution1D<\/code> layer has shape <code>[1,32]<\/code>.  It needs to be at least 4 steps to be used as input to the <code>MaxPool1D(4)<\/code> layer, so have a minimum size of <code>[4,32]<\/code>.<\/p>\n\n<p>You can continue walking this back.  For example, the <code>Convolution1D<\/code> layer will decrease the step size by kernel_size-1=2.  This means the input to the <code>Convolution1D<\/code> layer needs to have at least 4+2=6 steps, meaning a shape of at least <code>[6,?]<\/code>.  Continuing up to the input layer, you'll find the input size is too small.<\/p>\n\n<p>You'll need to change the architecture to allow the input size, or, if applicable, change the input size.<\/p>\n"}]}
+{"QuestionId":53300155,"AnswerCount":0,"Tags":"<pytorch><torchtext>","CreationDate":"2018-11-14T12:21:57.883","AcceptedAnswerId":null,"OwnerUserId":4575486.0,"Title":"AttributeError: 'Field' object has no attribute 'vocab'","Body":"<p>Returned iterator of BucketIterator is giving an error.<\/p>\n\n<pre><code>train_iterator = data.BucketIterator.splits(\n    train_data, \n    batch_size=BATCH_SIZE, \n    device=device\n)\n<\/code><\/pre>\n\n<p>Value of train_data also looks good.<\/p>\n\n<pre><code>print(vars(train_data.examples[10]))\n{'text': ['plus',\n  \"you've\",\n  'added',\n  'commercials',\n  'to',\n  'the',\n  'experience...',\n  'tacky.'],\n 'label': 'positive'}\n<\/code><\/pre>\n\n<p>Stack trace:<\/p>\n\n<pre><code>print([x for x in train_iterator])\n\n---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-50-f14edc400c88&gt; in &lt;module&gt;()\n----&gt; 1 [x for x in train_iter]\n\n&lt;ipython-input-50-f14edc400c88&gt; in &lt;listcomp&gt;(.0)\n----&gt; 1 [x for x in train_iter]\n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/iterator.py in __iter__(self)\n    154                     else:\n    155                         minibatch.sort(key=self.sort_key, reverse=True)\n--&gt; 156                 yield Batch(minibatch, self.dataset, self.device)\n    157             if not self.repeat:\n    158                 return\n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/batch.py in __init__(self, data, dataset, device)\n     32                 if field is not None:\n     33                     batch = [getattr(x, name) for x in data]\n---&gt; 34                     setattr(self, name, field.process(batch, device=device))\n     35 \n     36     @classmethod\n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/field.py in process(self, batch, device)\n    235         \"\"\"\n    236         padded = self.pad(batch)\n--&gt; 237         tensor = self.numericalize(padded, device=device)\n    238         return tensor\n    239 \n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/field.py in numericalize(self, arr, device)\n    334         if self.use_vocab:\n    335             if self.sequential:\n--&gt; 336                 arr = [[self.vocab.stoi[x] for x in ex] for ex in arr]\n    337             else:\n    338                 arr = [self.vocab.stoi[x] for x in arr]\n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/field.py in &lt;listcomp&gt;(.0)\n    334         if self.use_vocab:\n    335             if self.sequential:\n--&gt; 336                 arr = [[self.vocab.stoi[x] for x in ex] for ex in arr]\n    337             else:\n    338                 arr = [self.vocab.stoi[x] for x in arr]\n\n~\/anaconda3\/envs\/env\/lib\/python3.6\/site-packages\/torchtext\/data\/field.py in &lt;listcomp&gt;(.0)\n    334         if self.use_vocab:\n    335             if self.sequential:\n--&gt; 336                 arr = [[self.vocab.stoi[x] for x in ex] for ex in arr]\n    337             else:\n    338                 arr = [self.vocab.stoi[x] for x in arr]\n\nAttributeError: 'Field' object has no attribute 'vocab'\n<\/code><\/pre>\n\n<p>May I know what is it that we are missing here?<\/p>\n","answers":[]}
+{"QuestionId":53300337,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T12:32:22.863","AcceptedAnswerId":53390234.0,"OwnerUserId":9973879.0,"Title":"variable_scope does not get reused when using default scope name","Body":"<p>I have a question regarding sub-scopes when reusing variables. This<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef make_bar():\n  with tf.variable_scope('bar'):\n    tf.get_variable('baz', ()) \n\nwith tf.variable_scope('foo') as scope:\n  make_bar()\n  scope.reuse_variables()\n  make_bar()\n<\/code><\/pre>\n\n<p>works perfectly fine, only a single variable <code>foo\/bar\/baz<\/code> is created.<\/p>\n\n<p>However, if I change <code>make_bar<\/code> to<\/p>\n\n<pre><code>def make_bar(scope=None):\n  with tf.variable_scope(scope, 'bar'):\n    tf.get_variable('baz', ()) \n<\/code><\/pre>\n\n<p>the code now fails with a <\/p>\n\n<pre><code>ValueError: Variable foo\/bar_1\/baz does not exist\n<\/code><\/pre>\n\n<p>Question: why does variable scope reuse fail when using <code>default name<\/code>s? If it is on purpose, what is the rationale behind this choice?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>Some precisions on the <code>default_name<\/code> argument of <code>tf.variable_scope<\/code>. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope#__init__\" rel=\"nofollow noreferrer\">From the documentation,<\/a><\/p>\n\n<blockquote>\n  <ul>\n  <li><code>default_name<\/code>: The default name to use if the <code>name_or_scope<\/code> argument is <code>None<\/code>, this name will be uniquified. If <code>name_or_scope<\/code> is provided it won't be used and therefore it is not required and can be <code>None<\/code>.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>So as it name underlies, it is a way to provide a default scope name.<\/p>\n\n<p>In the first version of <code>make_bar<\/code>, the scope name is forced to be <code>bar<\/code> -- the function has no parameter to change it.<\/p>\n\n<p>In the second version of <code>make_bar<\/code>, I enhance this function to make it parameterizable. So <code>bar<\/code> is still the default scope name (provided this time as the <code>default_name<\/code> argument of <code>tf.variable_scope<\/code>), but this time the caller has the possibility to change it by setting the default argument <code>scope<\/code> of <code>make_bar<\/code> to anything else than <code>None<\/code>.`<\/p>\n\n<p>When this second version of <code>make_bar<\/code> is used without an argument, it should, I think, fall back to the behavior of the first version -- which it does not.<\/p>\n\n<p>Note that in my example, <code>bar<\/code> is intended to be a subscope of <code>foo<\/code>. The variable to be reused is <em>meant<\/em> to be <code>foo\/bar\/baz<\/code>.<\/p>\n","answers":[{"AnswerId":"53378104","CreationDate":"2018-11-19T15:41:41.577","ParentId":null,"OwnerUserId":"6565406","Title":null,"Body":"<p>You're not actually using the scope 'foo' in your example. You need to pass the parameter to <code>tf.variable_scope('foo', 'bar')<\/code> or <code>tf.variable_scope(scope, 'bar')<\/code>.\nyou're calling the method <code>make_bar<\/code> without the parameter in either case, which means in your first example <code>name_or_scope='bar'<\/code>, in the second example <code>name_or_scope=scope<\/code> (with value <code>None<\/code>) and <code>default_name='bar'<\/code>.<\/p>\n\n<p>this is probably what you want:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef make_bar(scope=None):\n    with tf.variable_scope(scope, 'bar'):\n        tf.get_variable('baz', ())\n\nwith tf.variable_scope('foo') as scope:\n    make_bar(scope)\n    scope.reuse_variables()\n    make_bar(scope)\n<\/code><\/pre>\n\n<p>I would actually advise against using default parameters because they reduce readability like in your example. When is a <code>None<\/code> scope ever the answer you want? it would make more sense if you test it, maybe like this?<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef make_bar(scope=None):\n    if scope is None:\n        scope = 'default_scope'\n    with tf.variable_scope(scope, 'bar'):\n        tf.get_variable('baz', ())\n\n\nwith tf.variable_scope('foo') as scope:\n    make_bar(scope)   # use foo scope\n    scope.reuse_variables()\n    make_bar()        # use 'default_scope'\n<\/code><\/pre>\n\n<p>but this makes code less readable and more likely to lead to bugs<\/p>\n"},{"AnswerId":"53390234","CreationDate":"2018-11-20T09:48:09.477","ParentId":null,"OwnerUserId":"6565406","Title":null,"Body":"<p>From the documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope<\/a><\/p>\n\n<blockquote>\n  <p>If <code>name_or_scope<\/code> is not None, it is used as is. If <code>name_or_scope<\/code>\n  is None, then <code>default_name<\/code> is used. In that case, if the same name\n  has been previously used in the same scope, it will be made <strong>unique<\/strong>\n  by appending <code>_N<\/code> to it.<\/p>\n<\/blockquote>\n\n<p>...<\/p>\n\n<blockquote>\n  <p>To prevent accidental sharing of variables, we raise an exception when\n  getting an existing variable in a non-reusing scope.<\/p>\n<\/blockquote>\n\n<p>...<\/p>\n\n<blockquote>\n  <p>Similarly, we raise an exception when trying to get a variable that\n  does not exist in reuse mode.<\/p>\n<\/blockquote>\n\n<p>By using <code>None<\/code> scope you are forcing each call to <code>get_variable<\/code> to generate unique scope\/variable, but by calling <code>reuse_variables<\/code> you are forcing subsequent calls to only return existing variables. There is no variable to return so <code>ValueError<\/code> is raised. This is expected behaviour.<\/p>\n"}]}
+{"QuestionId":53300568,"AnswerCount":4,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-14T12:45:42.497","AcceptedAnswerId":null,"OwnerUserId":10651990.0,"Title":"How to import Your own data to google colab","Body":"<p>I'm kinda new in python and tensorflow. I want to run my Deep network on a collab for human activity recognition.\nThe problem is I don't know how to import my dataset to collab!\nI already did upload dataset to Drive. (The dataset link is in below)\n<a href=\"http:\/\/www.cis.fordham.edu\/wisdm\/dataset.php\" rel=\"nofollow noreferrer\">http:\/\/www.cis.fordham.edu\/wisdm\/dataset.php<\/a>\nand I want to just use the WISDM_ar_v1.1_raw.txt file<\/p>\n","answers":[{"AnswerId":"53304212","CreationDate":"2018-11-14T15:59:08.690","ParentId":null,"OwnerUserId":"7376885","Title":null,"Body":"<p>According to <a href=\"https:\/\/stackoverflow.com\/a\/53302366\/7376885\">your post<\/a> you have one redundant bracket that leads to syntax error. Try this<\/p>\n\n<pre><code>data = pd.read_csv(filepath_or_buffer = 'My Drive\/app\/WISDM_ar_latest.tar.gz\/WISDM_ar_v1.1\/WISDM_ar_v1.1_raw.txt',header = None, names = column_names)\n<\/code><\/pre>\n"},{"AnswerId":"53304367","CreationDate":"2018-11-14T16:07:14.057","ParentId":null,"OwnerUserId":"10651990","Title":null,"Body":"<p>Okey It's solved now!\nAll think you need to do is just upload or data to your drive.\n(I uploaded in folder name: \"app\")\nThen<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/drive\/')\n<\/code><\/pre>\n\n<p>Now by this code you can import the data (Edit: Moved drive.mount to next line):<\/p>\n\n<pre><code>data = pd.read_csv(\"\/content\/drive\/My Drive\/app\/WISDM_ar_v1.1_raw.txt\")\n<\/code><\/pre>\n"},{"AnswerId":"53302366","CreationDate":"2018-11-14T14:20:19.880","ParentId":null,"OwnerUserId":"10651990","Title":null,"Body":"<blockquote>\n  <p>column_names = ['user-id','activity','timestamp', 'x-axis', 'y-axis',\n  'z-axis'] data = pd.read_csv(('My\n  Drive\/app\/WISDM_ar_latest.tar.gz\/WISDM_ar_v1.1\/WISDM_ar_v1.1_raw.txt',header\n  = None, names = column_names)<\/p>\n<\/blockquote>\n\n<pre><code>  File \"&lt;ipython-input-26-033bf3a3f419&gt;\", line 2\n    data = pd.read_csv(('My Drive\/app\/WISDM_ar_latest.tar.gz\/WISDM_ar_v1.1\/WISDM_ar_v1.1_raw.txt',header = None, names = column_names)\n                                                                                                         ^\nSyntaxError: invalid syntax\n<\/code><\/pre>\n"},{"AnswerId":"53300731","CreationDate":"2018-11-14T12:53:57.393","ParentId":null,"OwnerUserId":"7376885","Title":null,"Body":"<p>You got to mount your google drive in your notebook<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/drive')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1srw_HFWQ2SMgmWIawucXfusGzrj1_U0q\" rel=\"nofollow noreferrer\">Reference<\/a><\/p>\n"}]}
+{"QuestionId":53300965,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-11-14T13:07:16.130","AcceptedAnswerId":null,"OwnerUserId":10652100.0,"Title":"Pytorch Exception in Thread: ValueError: signal number 32 out of range","Body":"<p>I'm getting this error:<\/p>\n\n<pre><code>Exception in Thread: ValueError: signal number 32 out of range\n<\/code><\/pre>\n\n<p>The specific tutorial that raises an issue for me is the training a classifier (<a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html<\/a>), the specific line is: <\/p>\n\n<pre><code>dataiter = iter(trainloader) \n<\/code><\/pre>\n\n<p>and the full error traceback is:<\/p>\n\n<pre><code>Exception in thread Thread-5:\nTraceback (most recent call last):\nFile \"\/home\/chenchen\/anaconda3\/lib\/python3.6\/threading.py\", line 916, in _bootstrap_inner\nself.run()\nFile \"\/home\/chenchen\/anaconda3\/lib\/python3.6\/threading.py\", line 864, in run\nself._target(*self._args, **self._kwargs)\nFile \"\/home\/chenchen\/anaconda3\/lib\/python3.6\/multiprocessing\/resource_sharer.py\", line 139, in _serve\nsignal.pthread_sigmask(signal.SIG_BLOCK, range(1, signal.NSIG))\nFile \"\/home\/chenchen\/anaconda3\/lib\/python3.6\/signal.py\", line 60, in pthread_sigmask\nsigs_set = _signal.pthread_sigmask(how, mask)\nValueError: signal number 32 out of range\n<\/code><\/pre>\n\n<p>My operation system is Ubuntu 18.10 and my python env is Anaconda3 for python 3.6. I installed pytorch from the latest source. My cuda version is 10.0.<\/p>\n","answers":[{"AnswerId":"54670454","CreationDate":"2019-02-13T12:40:50.333","ParentId":null,"OwnerUserId":"3509475","Title":null,"Body":"<p>I have faced a similar issue and it got resolved when I set:<\/p>\n\n<pre><code>num_workers=0\n<\/code><\/pre>\n"},{"AnswerId":"58861719","CreationDate":"2019-11-14T16:40:16.437","ParentId":null,"OwnerUserId":"391376","Title":null,"Body":"<p>Looks like this was a bug in Python (<a href=\"https:\/\/bugs.python.org\/issue33329\" rel=\"nofollow noreferrer\">https:\/\/bugs.python.org\/issue33329<\/a>).<\/p>\n\n<p>For Python <strong>3.6<\/strong>, it got fixed in 3.6.6, released around July 2019 (see the <a href=\"https:\/\/docs.python.org\/3.6\/whatsnew\/changelog.html#python-3-6-6-release-candidate-1\" rel=\"nofollow noreferrer\">release notes<\/a> mentioning it). (You can check which version you have by running <code>python3 --version<\/code>.)<\/p>\n\n<p>For Python <strong>3.5<\/strong>, you'd want at least 3.5.7 (... <a href=\"https:\/\/docs.python.org\/3.5\/whatsnew\/changelog.html#python-3-5-7-final\" rel=\"nofollow noreferrer\">release notes<\/a>).<\/p>\n\n<p>Meanwhile... I haven't tried this personally, but Ubuntu 18.10 <a href=\"https:\/\/packages.ubuntu.com\/cosmic\/python\/python3.6\" rel=\"nofollow noreferrer\">seems to have Python 3.6.7 available<\/a>, so a simple <code>sudo apt install python3.6<\/code> might solve this?<\/p>\n"}]}
+{"QuestionId":53301015,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T13:09:49.820","AcceptedAnswerId":53305286.0,"OwnerUserId":2455494.0,"Title":"Tensorflow - How to Convert int32 to string (using Python API for Tensorflow)","Body":"<p>Simple question really but cannot seem to find a function in the TensorFlow docs or by googling.<\/p>\n\n<p>How can I convert a tensor of type <code>tf.int32<\/code> to one of type <code>tf.string<\/code>?<\/p>\n\n<p>I tried simply casting it with something like this:<\/p>\n\n<pre><code>x = tf.constant([1,2,3], dtype=tf.int32)\nx_as_string = tf.cast(x, dtype=tf.string) # hoping for this output: [ '1', '2', '3' ]\n\nwith tf.Session() as sess:\n  res = sess.run(x_as_string)\n<\/code><\/pre>\n\n<p>but hit the error message:<\/p>\n\n<blockquote>\n  <p>Cast int32 to string is not supported<\/p>\n<\/blockquote>\n\n<p>Is there a simple function somewhere in the documentation that I am missing?<\/p>\n\n<hr>\n\n<p>UPDATE:<\/p>\n\n<p>To clarify: I realise I could 'work around' this issue using a python function with <code>tf.py_func<\/code> but asking if there is a solution in TensorFlow itself<\/p>\n","answers":[{"AnswerId":"55584182","CreationDate":"2019-04-09T02:18:18.887","ParentId":null,"OwnerUserId":"6599380","Title":null,"Body":"<p>tf.as_string()\nConverts each entry in the given tensor to strings.  Supports many numeric<\/p>\n"},{"AnswerId":"53305286","CreationDate":"2018-11-14T16:59:42.390","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do that with the newly added (v1.12.0) <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/strings\/format\" rel=\"nofollow noreferrer\"><code>tf.strings.format<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.constant([1, 2, 3], dtype=tf.int32)\nx_as_string = tf.map_fn(lambda xi: tf.strings.format('{}', xi), x, dtype=tf.string)\n\nwith tf.Session() as sess:\n  res = sess.run(x_as_string)\n  print(res)\n  # [b'1' b'2' b'3']\n<\/code><\/pre>\n"}]}
+{"QuestionId":53301664,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2018-11-14T13:43:52.603","AcceptedAnswerId":null,"OwnerUserId":3655743.0,"Title":"how to compile summarize_graph with bazel in Tensorflow","Body":"<p>I am testing a Movidius Neural Compute stick.\nMy first test is to replicate this guide with Inception_v4 model:\n<a href=\"https:\/\/movidius.github.io\/ncsdk\/tf_modelzoo.html\" rel=\"nofollow noreferrer\">https:\/\/movidius.github.io\/ncsdk\/tf_modelzoo.html<\/a><\/p>\n\n<p>I do need to know the \"output_node_name\" for the freeze_graph.py.\nI have found instructions to do so in tensorflow docs: <\/p>\n\n<pre><code>bazel build tensorflow\/tools\/graph_transforms:summarize_graph\nbazel-bin\/tensorflow\/tools\/graph_transforms\/summarize_graph --in_graph=MY_PB_FILE.pb\n<\/code><\/pre>\n\n<p>However, when trying to follow steps, I get the following error:<\/p>\n\n<pre><code>lberto@alberto-OptiPlex-9010:~\/movidius_ws$ bazel build  tensorflow\/tools\/graph_transforms:summarize_graph\nStarting local Bazel server and connecting to it...\nERROR: Skipping 'tensorflow\/tools\/graph_transforms:summarize_graph': no such package 'tensorflow\/tools\/graph_transforms': BUILD file not found on package path\nWARNING: Target pattern parsing failed.\nERROR: no such package 'tensorflow\/tools\/graph_transforms': BUILD file not found on package path\nINFO: Elapsed time: 0.947s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)\n<\/code><\/pre>\n\n<p>According to the guide provided by Movidius, it is not necessary to install tensorflow in the host, only download the source code. I have cloned the Tensorflow source code in a directory to follow the guide, and the movidius tools in a docker container sharing that directory.<\/p>\n\n<p>What am I doing wrong building the summarize_graph application?\nI need to have this application running to know the output_node_names that I want to use with the movidius neural compute stick<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"53318421","CreationDate":"2018-11-15T11:26:01.303","ParentId":null,"OwnerUserId":"7778502","Title":null,"Body":"<p>Make sure the following files exist:<\/p>\n\n<ul>\n<li>\"~\/movidius_ws\/WORKSPACE\"<\/li>\n<li>\"~\/movidius_ws\/tensorflow\/tools\/graph_transforms\/BUILD\" or \"~\/movidius_ws\/tensorflow\/tools\/graph_transforms\/BUILD.bazel\"<\/li>\n<\/ul>\n\n<p>If they do, check the list of targets in the \"\/\/tensorflow\/tools\/graph_transforms\" package:<\/p>\n\n<pre><code>bazel query \/\/tensorflow\/tools\/graph_transforms:*\n<\/code><\/pre>\n\n<p>and see if \"summarize_graph\" is there.<\/p>\n"}]}
+{"QuestionId":53301877,"AnswerCount":2,"Tags":"<python><keras><lstm>","CreationDate":"2018-11-14T13:54:49.257","AcceptedAnswerId":53304924.0,"OwnerUserId":4912188.0,"Title":"Sequence-to-sequence classification with variable sequence lengths in Keras","Body":"<p>I would like to train an LSTM or GRU network in TensorFlow\/Keras to continuously recognize whether a user is walking or not based on input from motion sensors (accelerometer and gyroscope). I have 50 input sequences with lengths varying from 581 to 5629 time steps and 6 features and 50 corresponding output sequences of boolean values. My problem is that I don't know how to feed the training data to the fit() method.<\/p>\n\n<p>I know approximately what I need to do: I'd like to train with 5 batches of 10 sequences each, and for each batch I have to pad all but the longest sequence so all 10 sequences have the same lengths and apply masking. I just don't know how to build the data structures. I know that I can make one big 3D tensor of size (50,5629,6) and that works, but it's painfully slow, so I'd really like to make the sequence length of each batch as small as possible.<\/p>\n\n<p>Here's the problem in code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Load data from file\nx_list, y_list = loadSequences(\"train.csv\")\n\n# x_list is now a list of arrays (n,6) of float64, where n is the timesteps\n# and 6 is the number of features, sorted by increasing sequence lengths.\n# y_list is a list of arrays (n,1) of Boolean.\n\nx_train = # WHAT DO I WRITE HERE?\ny_train = # AND HERE?\n\nmodel = tf.keras.models.Sequential([\n            tf.keras.layers.Masking(),\n            tf.keras.layers.LSTM(32, return_sequences=True),\n            tf.keras.layers.Dense(2, activation=tf.nn.softmax)\n        ])\nmodel.compile(optimizer='adam',\n            loss='sparse_categorical_crossentropy',\n            metrics=['accuracy'])\nmodel.fit(x_train, y_train, batch_size=10, epochs=100)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53450534","CreationDate":"2018-11-23T17:00:36.503","ParentId":null,"OwnerUserId":"4912188","Title":null,"Body":"<p>In case it helps someone, here's how I ended up implementing a solution:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Load data from file\nx_list, y_list = loadSequences(\"train.csv\")\n\n# x_list is now a list of arrays (m,n) of float64, where m is the timesteps\n# and n is the number of features.\n# y_list is a list of arrays (m,1) of Boolean.\nassert len(x_list) == len(y_list)\nnum_sequences = len(x_list)\nnum_features = len(x_list[0][0])\nbatch_size = 10\nbatches_per_epoch = 5\nassert batch_size * batches_per_epoch == num_sequences\n\ndef train_generator():\n    # Sort by length so the number of timesteps in each batch is minimized\n    x_list.sort(key=len)\n    y_list.sort(key=len)\n    # Generate batches\n    while True:\n        for b in range(batches_per_epoch):\n            longest_index = (b + 1) * batch_size - 1\n            timesteps = len(x_list[longest_index])\n            x_train = np.zeros((batch_size, timesteps, num_features))\n            y_train = np.zeros((batch_size, timesteps, 1))\n            for i in range(batch_size):\n                li = b * batch_size + i\n                x_train[i, 0:len(x_list[li]), :] = x_list[li]\n                y_train[i, 0:len(y_list[li]), 0] = y_list[li]\n            yield x_train, y_train\n\nmodel = tf.keras.models.Sequential([\n            tf.keras.layers.Masking(mask_value=0., input_shape=(None,num_features)),\n            tf.keras.layers.LSTM(32, return_sequences=True),\n            tf.keras.layers.Dense(2, activation=tf.nn.softmax)\n        ])\nmodel.compile(optimizer='adam',\n            loss='sparse_categorical_crossentropy',\n            metrics=['accuracy'])\nmodel.fit_generator(train_generator(), steps_per_epoch=batches_per_epoch, epochs=100)\n<\/code><\/pre>\n"},{"AnswerId":"53304924","CreationDate":"2018-11-14T16:38:50.117","ParentId":null,"OwnerUserId":"9194674","Title":null,"Body":"<p>You can do some thing like this <\/p>\n\n<p>use generator function take a look at this link <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">fit_generator<\/a> look for fit_generator method.<\/p>\n\n<pre><code>def data_generater(batch_size):\nprint(\"reading data\")\ntraining_file = 'data_location', 'r')\n\n# assuming data is in json format so feels free to change accordingly\n\ntraining_set = json.loads(training_file.read())\ntraining_file.close()\n\nbatch_i = 0  # Counter inside the current batch vector\nbatch_x = []  # The current batch's x data\nbatch_y = []  # The current batch's y data\n\nwhile True:\n\n    for obj in training_set:\n        batch_x.append(your input sequences one by one)\n        if obj['val'] == True:\n            batch_y.append([1])\n        elif obj['val'] == False:\n            batch_y.append([0])\n        batch_i += 1\n\n        if batch_i == batch_size:\n            # Ready to yield the batch\n            # pad input to max length in the batch\n            batch_x = pad_txt_data(batch_x)\n            yield batch_x, np.array(batch_y)\n            batch_x = []\n            batch_y = []\n            batch_i = 0\n\ndef pad_txt_data(arr):\n# expecting arr to be in the shape of (10, m, 6)\npaded_arr = []\nprefered_len = len(max(arr, key=len))\n\n# Now pad all your sequences to preferred length in the batch(arr)\n\nreturn np.array(paded_arr)\n<\/code><\/pre>\n\n<p>and in the model <\/p>\n\n<pre><code>model = keras.Sequential()\nmodel.add(keras.layers.Masking(mask_value=0., input_shape=(None,6)))\nmodel.add(keras.layers.LSTM(32))\nmodel.add(keras.layers.Dense(1, activation=\"softmax\"))\nmodel.compile(optimizer=\"Adam\", loss='categorical_crossentropy', metrics=['categorical_accuracy'])\nmodel.fit_generator(data_generater(10), steps_per_epoch=5, epochs=10)\n<\/code><\/pre>\n\n<p>Batch_size, steps_per_epoch, epoch can be different.\nGenerally <\/p>\n\n<p><code>steps_per_epoch = (number of sequences\/batch_size)<\/code><\/p>\n\n<p>Note: Form reading your description your task appears to be Binary classification problem not like an Sequence to sequence problem. A good example for sequence to sequence is a language translation. Just google around you will find what i mean.<\/p>\n\n<p>And if you really want to see the difference in training times I suggest using a GPU if available and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/CuDNNLSTM\" rel=\"nofollow noreferrer\">CuDNNLSTM<\/a>.  <\/p>\n"}]}
+{"QuestionId":53302713,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-14T14:39:06.090","AcceptedAnswerId":53303083.0,"OwnerUserId":5223033.0,"Title":"Does it differ to use optimizer.step or model.step in pytorch?","Body":"<p>In pytorch, to update the model, should I use <code>optimizer.step()<\/code> or <code>model.step()<\/code> ?<\/p>\n\n<p>Here is a example snippet:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch nn\nclass SomeNeuralNet(nn.Module):\n    def __init__(self,hs,es,dropout):\n          SomeNeuralNet(ClaimRecognizer, self).__init__()\n          # Some initialization here\n    def forward(x):\n          # forward propagation here\n\nmodel = SomeNeuralNet(es,hs,dp)\noptimizer = optim.Adam(model.parameters())\nloss_function = nn.NLLLoss()\nfor epoch in N:\n   for x in data:\n      logp = model(x)\n      loss = loss_function(logp,gold_outs)\n      loss.backward()\n       # Which one I should call ? Optimizer.step() or model.step() or both ?\n       optimizer.step()\n       model.step()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53303083","CreationDate":"2018-11-14T14:58:43.587","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>To make the gradient descent step, you normally use just <code>optimizer.step()<\/code>. <br>Here is also an example taken from the <em><a href=\"https:\/\/pytorch.org\/docs\/stable\/optim.html#taking-an-optimization-step\" rel=\"nofollow noreferrer\">documentation (same link at bottom)<\/a><\/em>, what it looks like in general:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for input, target in dataset:\n    optimizer.zero_grad()\n    output = model(input)\n    loss = loss_fn(output, target)\n    loss.backward()\n    optimizer.step()\n<\/code><\/pre>\n\n<p>I don't know where you got this <code>model.step()<\/code>? Does did you try it? <\/p>\n\n<p>If your model really possesses some kind of <em><code>step()<\/code>-functionality<\/em>, it probably does something different.<\/p>\n\n<p>But unless you define something extra, your <code>model<\/code> gets its functions from <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module\" rel=\"nofollow noreferrer\"><code>nn.Module<\/code><\/a> and this does not have <code>step<\/code> function!<\/p>\n\n<p>See this example from the the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module\" rel=\"nofollow noreferrer\">Pytorch Documentation<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch.nn as nn\nimport torch.nn.functional as F\n\nclass Model(nn.Module):\n    def __init__(self):\n        super(Model, self).__init__()\n        self.conv1 = nn.Conv2d(1, 20, 5)\n        self.conv2 = nn.Conv2d(20, 20, 5)\n\n    def forward(self, x):\n        x = F.relu(self.conv1(x))\n        return F.relu(self.conv2(x))\n\nmodel = Model()\nmodel.step()\n<\/code><\/pre>\n\n<p>Trying to call <code>step()<\/code> result in an <code>AttributeError<\/code>:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-41-b032813f7eda&gt; in &lt;module&gt;\n     13 \n     14 model = Model()\n---&gt; 15 model.step()\n\n~\/miniconda3\/envs\/py37\/lib\/python3.7\/site-packages\/torch\/nn\/modules\/module.py in __getattr__(self, name)\n    530                 return modules[name]\n    531         raise AttributeError(\"'{}' object has no attribute '{}'\".format(\n--&gt; 532             type(self).__name__, name))\n    533 \n    534     def __setattr__(self, name, value):\n\nAttributeError: 'Model' object has no attribute 'step'\n<\/code><\/pre>\n\n<p>To sum it up, normally your <code>model<\/code> should <em>not<\/em> have a <code>step()<\/code> function, <code>optimizer.step()<\/code> is the way to go if you want to do the <em>optimization step<\/em>.<\/p>\n\n<p>See also here:\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/optim.html#taking-an-optimization-step\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/optim.html#taking-an-optimization-step<\/a><\/p>\n"}]}
+{"QuestionId":53302958,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-14T14:51:57.523","AcceptedAnswerId":54084297.0,"OwnerUserId":1138455.0,"Title":"How to debug Tensorflow segmentation fault in model.fit()?","Body":"<p>I am trying to run the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"noreferrer\">Keras MINST example<\/a> using tensorflow-gpu with a Geforce 2080. My environment is Anaconda on a Linux system.<\/p>\n\n<p>I am running the unmodified example from a command line python session. I get the following output:<\/p>\n\n<pre><code>Using TensorFlow backend.\nDevice mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: GeForce RTX 2080, pci bus id: 0000:01:00.0, compute capability: 7.5\nx_train shape: (60000, 28, 28, 1)\n60000 train samples\n10000 test samples\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/12\nconv2d_1\/random_uniform\/RandomUniform: (RandomUniform): \n\/job:localhost\/replica:0\/task:0\/device:GPU:0\nconv2d_1\/random_uniform\/sub: (Sub): \n\/job:localhost\/replica:0\/task:0\/device:GPU:0\nconv2d_1\/random_uniform\/mul: (Mul): \n\/job:localhost\/replica:0\/task:0\/device:GPU:0\nconv2d_1\/random_uniform: (Add): \n\/job:localhost\/replica:0\/task:0\/device:GPU:0\n[...]\n<\/code><\/pre>\n\n<p>The last lines I receive are:<\/p>\n\n<pre><code>training\/Adadelta\/Const_31: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\ntraining\/Adadelta\/mul_46\/x: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\ntraining\/Adadelta\/mul_47\/x: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nSegmentation fault (core dumped)\n<\/code><\/pre>\n\n<p>From reading around I assumed this might be a memory problem and added these lines to prevent the GPU from running out of memory:<\/p>\n\n<pre><code>config = tf.ConfigProto(log_device_placement=True)\nconfig.gpu_options.per_process_gpu_memory_fraction=0.3\nK.tensorflow_backend.set_session(tf.Session(config=config))\n<\/code><\/pre>\n\n<p>Checking with the <code>nvidia-smi<\/code> tool that the GPU is actually used (<code>watch -n1 nvidia-smi<\/code>)I can confirm from the following output (in this run no <code>per_process_gpu_memory_fraction<\/code> was set to 1):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CpmLD.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CpmLD.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I suspect a version incompatibility somewhere between CUDA, Keras and Tensorflow to be the issue, but I don't know, how to debug this.<\/p>\n\n<p>What debugging measures are available to get to the bottom of this? What other issues might be the reason for this segfault?<\/p>\n\n<p>EDIT: I experimented further and replacing the model with this code works fine:<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Flatten(input_shape=input_shape),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n\n<p>However once I introduce a convolution layer like so<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape),\n#    keras.layers.Flatten(input_shape=input_shape),\n    keras.layers.Flatten(),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n<\/code><\/pre>\n\n<p>])<\/p>\n\n<p>then I again get the aforementioned segfault.<\/p>\n\n<p>All packets have been installed through Anaconda. I have installed<\/p>\n\n<ul>\n<li>conda 4.5.11<\/li>\n<li>python 3.6.6<\/li>\n<li>keras-gpu 2.2.4<\/li>\n<li>tensorflow 1.12.0<\/li>\n<li>tensorflow-gpu 1.12.0<\/li>\n<li>cudnn 7.2.1<\/li>\n<li>cudatoolkit 9.2<\/li>\n<\/ul>\n\n<p>EDIT: I tried the same code in a non anaconda environment and it works flawlessly. I would prefer to use anaconda though to avoid system updates breaking things.<\/p>\n","answers":[{"AnswerId":"53965613","CreationDate":"2018-12-29T00:10:40.937","ParentId":null,"OwnerUserId":"439431","Title":null,"Body":"<p>I had the same Conv2D problem with:<\/p>\n\n<ul>\n<li>Ubuntu 18.04<\/li>\n<li>Graphic card: GeForce RTX 2080<\/li>\n<li>CUDA: cuda_10.0.130_410<\/li>\n<li>CUDNN: cudnn-10.0-linux-x64-v7.4.2<\/li>\n<li>conda with Python 3.6<\/li>\n<\/ul>\n\n<p>Best advice was from this link: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/24383\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/24383<\/a><\/p>\n\n<p>So a fix should come with <strong>Tensorflow 1.13<\/strong>.\nIn the meantime, using Tensorflow 1.13 nightly build (Dec 26, 2018) + using tensorflow.keras instead of keras <strong>solved the issue<\/strong>.<\/p>\n"},{"AnswerId":"54712904","CreationDate":"2019-02-15T15:59:02.927","ParentId":null,"OwnerUserId":"9228901","Title":null,"Body":"<p>I had the exact same problem on a very similar system as Francois but using a RTX2070 on which I could reliably reproduce the segmentation fault error when using the conv2d function executed on the GPU. My setting:<\/p>\n\n<ul>\n<li>Ubuntu: 18.04<\/li>\n<li>GPU: RTX 2070<\/li>\n<li>CUDA: 10<\/li>\n<li>cudnn: 7<\/li>\n<li>conda with python 3.6<\/li>\n<\/ul>\n\n<p>I finally solved it by <strong><em>building tensorflow from source<\/em><\/strong> into a new conda environment. For a fantastic guide see e.g. the following link:\n<a href=\"https:\/\/gist.github.com\/Brainiarc7\/6d6c3f23ea057775b72c52817759b25c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/Brainiarc7\/6d6c3f23ea057775b72c52817759b25c<\/a><\/p>\n\n<p>This is basically like any other build-tensorflow-from-source guide and consisted in my case of the following steps:<\/p>\n\n<ol>\n<li>insalling bazel<\/li>\n<li>cloning tensorflow from git and running <code>.\/configure<\/code> <\/li>\n<li>running the appropriate <code>bazel build<\/code> command (see link for details)<\/li>\n<\/ol>\n\n<p>Some minor issues came up during the build, one of which was solved by installing 3 packages manually, using:<\/p>\n\n<pre><code>pip install keras_applications==1.0.4 --no-deps\npip install keras_preprocessing==1.0.2 --no-deps\npip install h5py==2.8.0\n<\/code><\/pre>\n\n<p>which I found out using this answer here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/51771039\/error-compiling-tensorflow-from-source-no-module-named-keras-applications\/51774943#51774943\">Error Compiling Tensorflow From Source - No module named &#39;keras_applications&#39;<\/a><\/p>\n\n<p>conv2d now works like a charm when using the gpu!<\/p>\n\n<p>However, since all this took a fairly long time (building from source takes over an hour, not counting the search for the solution on the internet) I recommend to make a backup of the system after you get it working, e.g. using timeshift or any other program that you like.<\/p>\n"},{"AnswerId":"54084297","CreationDate":"2019-01-08T01:55:45.093","ParentId":null,"OwnerUserId":"10881494","Title":null,"Body":"<p>Build the tensorflow from source(r1.13) .Conv2D segmentation fault fixed.<\/p>\n\n<p>follow <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"noreferrer\">Build from Source<\/a><\/p>\n\n<p>my GPU : RTX 2070\nUbuntu 16.04\nPython 3.5.2\nNvidia Driver 410.78\nCUDA - 10.0.130\ncuDNN-10.0 - 7.4.2.24\nTensorRT-5.0.0\nCompute Capability: 7.5<\/p>\n\n<p>Build : tensorflow-1.13.0rc0-cp35-cp35m-linux_x86_64<\/p>\n\n<p>Download prebuilt  from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22706\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22706<\/a><\/p>\n"}]}
+{"QuestionId":53303418,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2018-11-14T15:17:53.803","AcceptedAnswerId":null,"OwnerUserId":1614232.0,"Title":"How to build tenssorflow op with bazel with additional include directories","Body":"<p>I got tensorflow binaries (already compiled)\nI have added to tensorflow source:\n<code>tensorflow\\core\\user_ops\\icp_op_kernel.cc<\/code> - contains:\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/vid2depth\/ops\/icp_op_kernel.cc\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/vid2depth\/ops\/icp_op_kernel.cc<\/a>\n<code>tensorflow\\core\\user_ops\\BUILD<\/code> - contains:<\/p>\n\n<pre><code>load(\"\/\/tensorflow:tensorflow.bzl\", \"tf_custom_op_library\")\n\ntf_custom_op_library(\n    name = \"icp_op_kernel.so\",\n    srcs = [\"icp_op_kernel.cc\"],\n)\n<\/code><\/pre>\n\n<p>I am trying to build with:<\/p>\n\n<pre><code>bazel build --config opt \/\/tensorflow\/core\/user_ops:icp_op_kernel.so\n<\/code><\/pre>\n\n<p>And I get:<\/p>\n\n<pre><code>tensorflow\/core\/user_ops\/icp_op_kernel.cc(16): fatal error C1083: Cannot open include file: 'pcl\/point_types.h': No such file or directory\n<\/code><\/pre>\n\n<p>Because bazel don't know where the pcl include files are.\nI have installed pcl and the include directory is in:<\/p>\n\n<pre><code>C:\\Program Files\\PCL 1.6.0\\include\\pcl-1.6\n<\/code><\/pre>\n\n<p>How do I tell bazel to also include this directory?<\/p>\n\n<p>Also I will probably need to add <code>C:\\Program Files\\PCL 1.6.0\\lib<\/code> to the link, How do I do that?<\/p>\n","answers":[{"AnswerId":"53306145","CreationDate":"2018-11-14T17:53:41.723","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<p>You don't need <code>bazel<\/code> for building ops if it fails.<\/p>\n\n<p>I have implemented customized ops both in CPU and GPU, and basically follow the  two Tensorflow tutorials.<\/p>\n\n<p>For CPU ops, follow Tensorflow tutorial on <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#build_the_op_library\" rel=\"nofollow noreferrer\">Build the op library<\/a>:<\/p>\n\n<pre><code>TF_CFLAGS=( $(python -c 'import tensorflow as tf; print(\" \".join(tf.sysconfig.get_compile_flags()))') )\nTF_LFLAGS=( $(python -c 'import tensorflow as tf; print(\" \".join(tf.sysconfig.get_link_flags()))') )\ng++ -std=c++11 -shared zero_out.cc -o zero_out.so -fPIC ${TF_CFLAGS[@]} ${TF_LFLAGS[@]} -O2\n<\/code><\/pre>\n\n<p>Note on <code>gcc<\/code> version <code>&gt;=5<\/code>: gcc uses the new <code>C++ ABI<\/code> since version 5. The binary pip packages available on the TensorFlow website are built with gcc4 that uses the older ABI. If you compile your op library with <code>gcc&gt;=5<\/code>, add <code>-D_GLIBCXX_USE_CXX11_ABI=0<\/code> to the command line to make the library compatible with the older abi.<\/p>\n\n<p>For GPU ops, check the current official GPU ops building instructions on <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#gpu_support\" rel=\"nofollow noreferrer\">Tensorflow adding GPU op support<\/a><\/p>\n\n<pre><code>nvcc -std=c++11 -c -o cuda_op_kernel.cu.o cuda_op_kernel.cu.cc \\\n  ${TF_CFLAGS[@]} -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC\n\ng++ -std=c++11 -shared -o cuda_op_kernel.so cuda_op_kernel.cc \\\n  cuda_op_kernel.cu.o ${TF_CFLAGS[@]} -fPIC -lcudart ${TF_LFLAGS[@]}\n<\/code><\/pre>\n\n<p>As it says, Note that if your CUDA libraries are not installed in <code>\/usr\/local\/lib64<\/code>, you'll need to specify the path explicitly in the second (g++) command above. For example, add <code>-L \/usr\/local\/cuda-8.0\/lib64\/<\/code> if your CUDA is installed in <code>\/usr\/local\/cuda-8.0<\/code>.<\/p>\n\n<p>Also, Note in some linux settings, additional options to nvcc compiling step are needed. Add <code>-D_MWAITXINTRIN_H_INCLUDED<\/code> to the nvcc command line to avoid errors from <code>mwaitxintrin.h<\/code>.<\/p>\n"}]}
+{"QuestionId":53303513,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-14T15:23:37.230","AcceptedAnswerId":55117680.0,"OwnerUserId":8030268.0,"Title":"Tensorflow seq2seq tutorial 404","Body":"<p>i'm currently trying to use the seq2seq library but i can't access to the tutorial it is pointing \n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/index.html<\/a><\/p>\n\n<p>Has it been replaced by something else or function are deprecated ? <\/p>\n\n<p>wants to try different seq2seq and understand them. <\/p>\n","answers":[{"AnswerId":"55117680","CreationDate":"2019-03-12T09:08:37.507","ParentId":null,"OwnerUserId":"11171670","Title":null,"Body":"<p>Sorry, for the late answer, but you can always try to find broken or unmaintained older websites or links at <a href=\"https:\/\/archive.org\/\" rel=\"nofollow noreferrer\"><code>archive.org<\/code><\/a> with their wayback machine.<\/p>\n\n<p>I've found <a href=\"https:\/\/web.archive.org\/web\/20170129002833\/https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">your link<\/a> there, too.<\/p>\n"}]}
+{"QuestionId":53303724,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-14T15:35:00.257","AcceptedAnswerId":null,"OwnerUserId":9758922.0,"Title":"How to average only non-zero entries in Tensor?","Body":"<p>I've come across a case in which the averaging includes padded values. Given a tensor <code>X<\/code> of some shape <code>(batch_size, ..., features)<\/code>, there could be zero padded features to get the same shape.<\/p>\n\n<p>How can I average the final dimension of <code>X<\/code> (the features) but only the non-zero entries? So, we divide by the sum by the number of non-zero entries.<\/p>\n\n<p>Example input:<\/p>\n\n<pre><code>x = [[[[1,2,3], [2,3,4], [0,0,0]],\n       [[1,2,3], [2,0,4], [3,4,5]],\n       [[1,2,3], [0,0,0], [0,0,0]],\n       [[1,2,3], [1,2,3], [0,0,0]]],\n      [[[1,2,3], [0,1,0], [0,0,0]],\n       [[1,2,3], [2,3,4], [0,0,0]],                                                         \n       [[1,2,3], [0,0,0], [0,0,0]],                                                         \n       [[1,2,3], [1,2,3], [1,2,3]]]]\n# Desired output\ny = [[[1.5 2.5 3.5]\n      [2.  2.  4. ]\n      [1.  2.  3. ]\n      [1.  2.  3. ]]\n     [[0.5 1.5 1.5]\n      [1.5 2.5 3.5]\n      [1.  2.  3. ]\n      [1.  2.  3. ]]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53303725","CreationDate":"2018-11-14T15:35:00.257","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>A pure Keras solution counts the number of non-zero entries and then divides the sum accordingly. Here is a custom layer:<\/p>\n\n<pre><code>import keras.layers as L\nimport keras.backend as K\n\nclass NonZeroMean(L.Layer):\n  \"\"\"Compute mean of non-zero entries.\"\"\"\n  def call(self, x): \n    \"\"\"Calculate non-zero mean.\"\"\"\n    # count the number of nonzero features, last axis\n    nonzero = K.any(K.not_equal(x, 0.0), axis=-1)\n    n = K.sum(K.cast(nonzero, 'float32'), axis=-1, keepdims=True)\n    x_mean = K.sum(x, axis=-2) \/ n\n    return x_mean\n\n  def compute_output_shape(self, input_shape):\n    \"\"\"Collapse summation axis.\"\"\"\n    return input_shape[:-2] + (input_shape[-1],)\n<\/code><\/pre>\n\n<p>I suppose a condition needs to be added to check if all the features are zero and return zero, otherwise we get a division by zero error. Current example tested with:<\/p>\n\n<pre><code># Dummy data\nx = [[[[1,2,3], [2,3,4], [0,0,0]],\n      [[1,2,3], [2,0,4], [3,4,5]],\n      [[1,2,3], [0,0,0], [0,0,0]],\n      [[1,2,3], [1,2,3], [0,0,0]]],\n     [[[1,2,3], [0,1,0], [0,0,0]],\n      [[1,2,3], [2,3,4], [0,0,0]],\n      [[1,2,3], [0,0,0], [0,0,0]],\n      [[1,2,3], [1,2,3], [1,2,3]]]]\nx = np.array(x, dtype='float32')\n\n# Example run\nx_input = K.placeholder(shape=x.shape, name='x_input')\nout = NonZeroMean()(x_input)\ns = K.get_session()\nprint(\"INPUT:\", x)\nprint(\"OUTPUT:\", s.run(out, feed_dict={x_input: x}))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53303934,"AnswerCount":0,"Tags":"<tensorflow><kubernetes><cluster-computing><dask>","CreationDate":"2018-11-14T15:45:45.540","AcceptedAnswerId":null,"OwnerUserId":4717048.0,"Title":"Dask, Tensorflow serving (and Kubernetes and Streamz)","Body":"<p>What is the current 'state of technology' when having a pipeline composed of python code and Tensorflow\/Keras models?<\/p>\n\n<p>We are trying to have scalability and reactive design using <a href=\"https:\/\/github.com\/dask\/dask\" rel=\"nofollow noreferrer\">dask<\/a> and <a href=\"https:\/\/streamz.readthedocs.io\/en\/latest\/\" rel=\"nofollow noreferrer\">Streamz<\/a> (for servers registered using Kubernetes).\nBut currently, we do not know what is the right way to design such infrastructure concerning the fact, that we do want our Tensorflow models to persist and not to be repeatedly created and deleted.<\/p>\n\n<p>Is Tensorflow serving the technology to be used for this task?<\/p>\n\n<p>(I was able to find only the basic examples like <a href=\"https:\/\/stackoverflow.com\/questions\/42988110\/persistent-dataflows-with-dask\">Persistent dataflows with dask<\/a> and <a href=\"http:\/\/matthewrocklin.com\/blog\/work\/2017\/02\/11\/dask-tensorflow\" rel=\"nofollow noreferrer\">http:\/\/matthewrocklin.com\/blog\/work\/2017\/02\/11\/dask-tensorflow<\/a>)<\/p>\n","answers":[]}
+{"QuestionId":53304490,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-14T16:13:51.143","AcceptedAnswerId":53304909.0,"OwnerUserId":2213201.0,"Title":"Tensorflow - Iterate over two Tensors","Body":"<p>I have two similar tensors; one has all the found valid boxes, and the other has all of the indexes where they belonged.<\/p>\n\n<p><code>Tensor(\"valid_boxes:0\", shape=(?, 9), dtype=float32)<\/code><\/p>\n\n<p><code>Tensor(\"valid_boxes_indexes:0\", shape=(?, 4), dtype=int64)<\/code><\/p>\n\n<p>I need a <code>map_fun<\/code> which access both variables. I tried this:<\/p>\n\n<pre><code>operation = tf.map_fn(lambda x: generate_bounding_box(x[0], x[1][1], x[1][0], x[1][2], grid_h, grid_w, anchors), (valid_boxes, valid_boxes_indexes))\n<\/code><\/pre>\n\n<p>Tensorflow gave me the following:<\/p>\n\n<blockquote>\n  <p>ValueError: The two structures don't have the same nested structure.<\/p>\n  \n  <p>First structure: type=tuple str=(tf.float32, tf.int64)<\/p>\n  \n  <p>Second structure: type=Tensor str=Tensor(\"map_14\/while\/stack:0\",\n  shape=(5,), dtype=float32)<\/p>\n  \n  <p>More specifically: Substructure \"type=tuple str=(tf.float32,\n  tf.int64)\" is a sequence, while substructure \"type=Tensor\n  str=Tensor(\"map_14\/while\/stack:0\", shape=(5,), dtype=float32)\" is not<\/p>\n<\/blockquote>\n\n<p>Is there any way to do this properly?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53304909","CreationDate":"2018-11-14T16:37:48.780","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You need to specify a <code>dtype<\/code> when the input and output values do not have the same structure. From the documentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\"><code>tf.map_fn<\/code><\/a>:<\/p>\n\n<blockquote>\n  <p>Furthermore, <code>fn<\/code> may emit a different structure than its input. For example, <code>fn<\/code> may look like: <code>fn = lambda t1: return (t1 + 1, t1 - 1)<\/code>. In this case, the <code>dtype<\/code> parameter is not optional: <code>dtype<\/code> must be a type or (possibly nested) tuple of types matching the output of <code>fn<\/code>.<\/p>\n<\/blockquote>\n\n<p>Try with this:<\/p>\n\n<pre><code>operation = tf.map_fn(\n    lambda x: generate_bounding_box(x[0], x[1][1], x[1][0], x[1][2],\n                                    grid_h, grid_w, anchors),\n    (valid_boxes, valid_boxes_indexes)\n    dtype=tf.float32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53305259,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><magenta>","CreationDate":"2018-11-14T16:58:24.797","AcceptedAnswerId":53506101.0,"OwnerUserId":2594778.0,"Title":"TensorFlow: extract data with a given feature, from NSynth Dataset","Body":"<p>I have a data set of TFRecord files of serialized TensorFlow Example protocol buffers with one Example proto per note, downloaded from <a href=\"https:\/\/magenta.tensorflow.org\/datasets\/nsynth\" rel=\"nofollow noreferrer\">https:\/\/magenta.tensorflow.org\/datasets\/nsynth<\/a>. I am using the test set, which is approximately 1 Gb, in case someone wants to download it, to check the code below. Each Example contains many features: pitch, instrument ...<\/p>\n\n<p>The code that reads in this data is:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.InteractiveSession()\n\n# Reading input data\ndataset = tf.data.TFRecordDataset('..\/data\/nsynth-test.tfrecord')\n\n# Convert features into tensors\nfeatures = {\n\"pitch\": tf.FixedLenFeature([1], dtype=tf.int64),\n\"audio\": tf.FixedLenFeature([64000], dtype=tf.float32),\n\"instrument_family\": tf.FixedLenFeature([1], dtype=tf.int64)}\n\nparse_function = lambda example_proto: tf.parse_single_example(example_proto,features)\ndataset = dataset.map(parse_function)\n\n# Consuming TFRecord data.\ndataset = dataset.shuffle(buffer_size=10000)\ndataset = dataset.batch(batch_size=3)\ndataset = dataset.repeat()\niterator = dataset.make_one_shot_iterator()\nbatch = iterator.get_next()\nsess.run(batch)\n<\/code><\/pre>\n\n<p>Now, the pitch ranges from 21 to 108. But I want to consider data of a given pitch only, e.g. pitch = 51. How do I extract this \"pitch=51\" subset from the whole dataset? Or alternatively, what do I do to make my iterator go through this subset only?<\/p>\n","answers":[{"AnswerId":"53506101","CreationDate":"2018-11-27T18:36:29.413","ParentId":null,"OwnerUserId":"2969950","Title":null,"Body":"<p>What you have looks pretty good, all you're missing is a filter function.<\/p>\n\n<p>For example if you only wanted to extract pitch=51, you should add after your map function<\/p>\n\n<pre><code>dataset = dataset.filter(lambda example: tf.equal(example[\"pitch\"][0], 51))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53305946,"AnswerCount":0,"Tags":"<docker><tensorflow><nvidia><gpu-programming><nvidia-docker>","CreationDate":"2018-11-14T17:41:04.337","AcceptedAnswerId":null,"OwnerUserId":8710430.0,"Title":"Deep Learning on GPU","Body":"<p>A DGX-1 has quite a lot of power. However, when using it I only utilize 34% of one of eight cards. Are there some points and tweaks I might missed? I know I can't parallelize everything, there is an upper limit. But to be honest, I expected a better performance.<\/p>\n\n<p>I'm running a nvidia-docker with tensorflow preinstalled. The <a href=\"https:\/\/github.com\/dennybritz\/reinforcement-learning\/blob\/master\/DQN\/dqn.py\" rel=\"nofollow noreferrer\">script<\/a> running is from dennybritz which is quite well used. I ran the docker one one card. Since it didn't utilized the whole card I didn't gave him second one. Would this have any benefit? Of course I could ran multiple instances and pick the best one. But I'd rather have results sooner when having 170TFLOPS accessible. <\/p>\n","answers":[]}
+{"QuestionId":53306252,"AnswerCount":0,"Tags":"<python><image><tensorflow>","CreationDate":"2018-11-14T18:00:36.660","AcceptedAnswerId":null,"OwnerUserId":10304086.0,"Title":"I want to use learned weights with a tensor flow","Body":"<p>I trained my neural network with a tensorflow and I learned 1,000,000 times.<\/p>\n\n<p>Three files have been created in \"C \/ folder\". (meta, index and data files).<\/p>\n\n<p>I would like to load only the my weight and bias. <\/p>\n\n<p>Please look following code.<\/p>\n\n<pre><code>c_dim = 1\nscale = 3\nim = Image.open('test.bmp') \n#shape of im is(256, 256, 3)\nimage_size_width, image_size_height = im.width, im.height\nimg = im.convert('YCbCr')\n# I need only Y channel\n# shape of img is (256, 256, 3)\n\narr_img = np.asarray(img)\narr_img = arr_img[:, :, 0]\n#shape of arr_img is (256, 256)\narrimg = np.expand_dims(arr_img, 0)\narrimg = np.expand_dims(arrimg, 3)\n# Tensorflow needs... [?, 256, 256, 1] so, i expand dimention of 'arrimg'\n\nimages = tf.placeholder(tf.float32, [None, image_size_width, image_size_height, c_dim], name='images')\n# I define plachholder\nw1 = tf.Variable(tf.random_normal([9, 9, 1, 64], stddev=1e-3), name='w1')\nw2 = tf.Variable(tf.random_normal([1, 1, 64, 32], stddev=1e-3), name='w2')\nw3 = tf.Variable(tf.random_normal([5, 5, 32, 1], stddev=1e-3), name='w3')\n# I define weight\nb1 = tf.Variable(tf.zeros([64]), name='b1')\nb2 = tf.Variable(tf.zeros([32]), name='b2')\nb3 = tf.Variable(tf.zeros([1]), name='b3')\n# I define bias\n\nconv1 = tf.nn.relu(tf.nn.conv2d(images, w1, strides=[1,1,1,1], padding='VALID') + b1)\nconv2 = tf.nn.relu(tf.nn.conv2d(conv1, w2, strides=[1,1,1,1], padding='VALID') + b2)\nresult = tf.nn.conv2d(conv2, w3, strides=[1,1,1,1], padding='VALID') + b3\n# After restoring the saved my weights, I want to put it into the calculation graph I want.\n\nsess = tf.Session()\nsaver = tf.train.Saver()\nsaver.restore(sess, 'C:\/folder\/my.model-1000000')\nsaver.restore(sess, tf.train.latest_checkpoint('C:\/folder\/'))\n#I restore my weight and bias \n\nsess.run(tf.global_variables_initializer())\naa = sess.run(result, {images : arrimg})\n\n#aa = aa[0,:,:,0]\nprint(type(aa))\n# this is numpy array\nprint(np.shape(aa))\n# (1, 244, 244, 1)\n# I can not change this(shape of (1, 244, 244, 1)) to image! \n\naa = np.reshape(aa, (244, 244))\n# so i change shape\n\nresultimage = Image.fromarray(aa, 'L')\nresultimage.save('C:\/SRCNN\/result.bmp')\n<\/code><\/pre>\n\n<p>However, there is only meaningless black and white image.<\/p>\n\n<p>The tensorflow must have rank 4 for graph computation.<\/p>\n\n<p>So I changed the dimensions of the original RGB image (256, 256, 3) at will.<\/p>\n\n<p>Is it because I made a mistake in image processing?<\/p>\n\n<p>Or did i make a mistake in how to restore weights and bias?<\/p>\n","answers":[]}
+{"QuestionId":53306630,"AnswerCount":1,"Tags":"<python><machine-learning><keras><text-classification><tweets>","CreationDate":"2018-11-14T18:29:08.983","AcceptedAnswerId":53307294.0,"OwnerUserId":7613534.0,"Title":"Keras: tweets classification","Body":"<p>Hello dear forum members,<\/p>\n\n<p>I have a data set of 20 Million randomly collected individual tweets (no two tweets come from the same account). Let me refer to this data set as \"general\" data set. Also, I have another \"specific\" data set that includes 100,000 tweets collected from drug (opioid) abusers. Each tweet has at least one tag associated with it, e.g., opioids, addiction, overdose, hydrocodone, etc. (max 25 tags).<\/p>\n\n<p>My goal is to use the \"specific\" data set to train the model using Keras and then use it to tag tweets in the \"general\" data set to identify tweets that might have been written by drug abusers.<\/p>\n\n<p>Following examples in <a href=\"https:\/\/cloud.google.com\/blog\/products\/gcp\/intro-to-text-classification-with-keras-automatically-tagging-stack-overflow-posts\" rel=\"nofollow noreferrer\">source1<\/a> and <a href=\"http:\/\/%20%20%20%20https:\/\/datascience.stackexchange.com\/questions\/38280\/keras-multiple-text-features-input-and-single-text-label-output-classification\" rel=\"nofollow noreferrer\">source2<\/a>, I managed to build a simple working version of such model:<\/p>\n\n<pre><code>from tensorflow.python import keras\nimport pandas as pd\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\nfrom sklearn.preprocessing import LabelBinarizer, LabelEncoder\nfrom sklearn.metrics import confusion_matrix\nfrom tensorflow import keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Activation, Dropout\nfrom keras.preprocessing import text, sequence\nfrom keras import utils\n\n# load opioid-specific data set, where post is a tweet and tags is a single tag associated with a tweet\n# how would I include multiple tags to be used in training?\ndata = pd.read_csv(\"filename.csv\")\ntrain_size = int(len(data) * .8)\ntrain_posts = data['post'][:train_size]\ntrain_tags = data['tags'][:train_size]\ntest_posts = data['post'][train_size:]\ntest_tags = data['tags'][train_size:]\n\n# tokenize tweets\nvocab_size = 100000 # what does vocabulary size really mean?\ntokenize = text.Tokenizer(num_words=vocab_size)\ntokenize.fit_on_texts(train_posts)\nx_train = tokenize.texts_to_matrix(train_posts)\nx_test = tokenize.texts_to_matrix(test_posts)\n\n# make sure columns are strings\ndata['post'] = data['post'].astype(str)\ndata['tags'] = data['tags'].astype(str)\n\n# labeling\n# is this where I add more columns with tags for training?\nencoder = LabelBinarizer()\nencoder.fit(train_tags)\ny_train = encoder.transform(train_tags)\ny_test = encoder.transform(test_tags)\n\n# model building\nbatch_size = 32\nmodel = Sequential()\nmodel.add(Dense(512, input_shape=(vocab_size,)))\nmodel.add(Activation('relu'))\nnum_labels = np.max(y_train) + 1 #what does this +1 really mean?\nmodel.add(Dense(1865))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nhistory = model.fit(x_train, y_train, batch_size = batch_size, epochs = 5, verbose = 1, validation_split = 0.1)\n\n# test prediction accuracy\nscore = model.evaluate(x_test, y_test, \n                       batch_size=batch_size, verbose=1)\nprint('Test score:', score[0])\nprint('Test accuracy:', score[1])\n\n# make predictions using a test set\nfor i in range(1000):    \n    prediction = model.predict(np.array([x_test[i]]))\ntext_labels = encoder.classes_ \npredicted_label = text_labels[np.argmax(prediction[0])]\nprint(test_posts.iloc[i][:50], \"...\")\nprint('Actual label:' + test_tags.iloc[i])\nprint(\"Predicted label: \" + predicted_label)\n<\/code><\/pre>\n\n<p>In order to move forward, I would like to clarify a few things:<\/p>\n\n<ol>\n<li>Let's say all my training tweets have a single tag -- opioids. Then if I pass the non-tagged tweets through it, isn't it likely that the model simply tags all of them as opioids as it doesn't know anything else? Should I be using a variety of different tweets\/tags then for the learning purpose? Perhaps, there are any general guidelines for the selection of the tweets\/tags for the training purposes?<\/li>\n<li>How can I add more columns with tags for training (not a single one like is used in the code)?<\/li>\n<li>Once I train the model and achieve appropriate accuracy, how do I pass non-tagged tweets through it to make predictions?<\/li>\n<li>How do I add a confusion matrix?<\/li>\n<\/ol>\n\n<p>Any other relevant feedback is also greatly appreciated.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Examples of \"general\" tweets:<\/p>\n\n<pre><code>everybody messages me when im in class but never communicates on the weekends like this when im free. feels like that anyway lol.\ni woke up late, and now i look like shit. im the type of person who will still be early to whatever, ill just look like i just woke up.\n<\/code><\/pre>\n\n<p>Examples of \"specific\" tweets:<\/p>\n\n<pre><code>$2 million grant to educate clinicians who prescribe opioids\nearly and regular marijuana use is associated with use of other illicit drugs, including opioids\n<\/code><\/pre>\n","answers":[{"AnswerId":"53307294","CreationDate":"2018-11-14T19:15:37.823","ParentId":null,"OwnerUserId":"6109920","Title":null,"Body":"<p>My shot to this is:<\/p>\n\n<ol>\n<li><p>Create a new dataset with tweets from general + specific data. Let's say 200k-250K where 100K is you specific data set, rest is general<\/p><\/li>\n<li><p>Take your 25 keywords\/tags and write a rule if any one or more than one exists in a tweet it is DA (Drug Abuser) or NDA(Non Drug Abuser). This will be your dependent variable.<\/p><\/li>\n<li><p>Your new dataset will be one column with all the tweets and another column with the dependent variable telling it is DA or NDA<\/p><\/li>\n<li><p>Now divide into train\/test and use keras or any other algo. so that it can learn. <\/p><\/li>\n<li><p>Then test the model by plotting Confusion Matrix<\/p><\/li>\n<li><p>Passs you other remaining data set from General to this model and check,<\/p><\/li>\n<\/ol>\n\n<p>If their are new words other than 25 which are not in the specific dataset, from the model you built it will still try to intelligently guess the right category by the group of words that come together, tone etc.<\/p>\n"}]}
+{"QuestionId":53306696,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-14T18:33:50.207","AcceptedAnswerId":null,"OwnerUserId":7749267.0,"Title":"Keras cnn: training accuracy and validation accuracy are incredibly high","Body":"<pre><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Conv2D\nfrom tensorflow.keras.layers import MaxPooling2D\nfrom tensorflow.keras.layers import Flatten\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.layers import Dropout\nfrom tensorflow.keras.utils import to_categorical\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import LabelEncoder\nimport numpy as np\nimport cv2\nimport os\nfrom imutils import paths\nimport imutils\n\n\ndef image_to_feature_vector(image, size=(32, 32)):\n    # resize the image to a fixed size, then flatten the image into\n    # a list of raw pixel intensities\n    return np.reshape(cv2.resize(image, size).flatten(), (32, 32, 3))\n\n\ndef extract_color_histogram(image, bins=(8, 8, 8)):\n    # extract a 3D color histogram from the HSV color space using\n    # the supplied number of `bins` per channel\n    hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n    hist = cv2.calcHist([hsv], [0, 1, 2], None, bins,\n                        [0, 180, 0, 256, 0, 256])\n\n    # handle normalizing the histogram if we are using OpenCV 2.4.X\n    if imutils.is_cv2():\n        hist = cv2.normalize(hist)\n\n    # otherwise, perform \"in place\" normalization in OpenCV 3\n    else:\n        cv2.normalize(hist, hist)\n\n    # return the flattened histogram as the feature vector\n    return hist.flatten()\n\n\nDATA_SET_PATH = '..\/training_images'\n# grab the list of images that we'll be describing\nprint(\"[INFO] describing images...\")\nimagePaths = list(paths.list_images(DATA_SET_PATH))\n\n# initialize the raw pixel intensities matrix, the features matrix,\n# and labels list\nrawImages = []\nfeatures = []\nlabels = []\n\n# loop over the input images\nfor (i, imagePath) in enumerate(imagePaths):\n    # load the image and extract the class label (assuming that our\n    # path as the format: \/path\/to\/dataset\/{class}.{image_num}.jpg\n    image = cv2.imread(imagePath)\n    label = imagePath.split(os.path.sep)[-2].split(\".\")[1]\n\n    # extract raw pixel intensity \"features\", followed by a color\n    # histogram to characterize the color distribution of the pixels\n    # in the image\n    pixels = image_to_feature_vector(image)\n    hist = extract_color_histogram(image)\n\n    # update the raw images, features, and labels matricies,\n    # respectively\n    rawImages.append(pixels)\n    labels.append(label)\n    features.append(hist)\n\n    # show an update every 1,000 images\n    if i &gt; 0 and i % 1000 == 0:\n        print(\"[INFO] processed {}\/{}\".format(i, len(imagePaths)))\n\n# show some information on the memory consumed by the raw images\n# matrix and features matrix\n\nrawImages = np.array(rawImages)\nfeatures = np.array(features)\nlabels = np.array(labels)\nprint(\"[INFO] pixels matrix: {:.2f}MB\".format(\n    rawImages.nbytes \/ (1024 * 1000.0)))\n\nencoder = LabelEncoder()\nencoder.fit(labels)\nlabels = to_categorical(encoder.transform(labels))\n\n# partition the data into training and testing splits, using 75%\n# of the data for training and the remaining 25% for testing\n(trainRI, testRI, trainRL, testRL) = train_test_split(\n    rawImages, labels, test_size=0.15, random_state=42)\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3), padding='same', activation='relu', input_shape=(32, 32, 3)))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(257, activation='softmax'))\n\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\naug = ImageDataGenerator(rotation_range=30, width_shift_range=0.1,\n                         height_shift_range=0.1, shear_range=0.2, zoom_range=0.2,\n                         horizontal_flip=True, fill_mode=\"nearest\")\n\nmodel.fit_generator(aug.flow(trainRI, trainRL, batch_size=256), epochs=10, validation_data=(testRI, testRL))\n\nmodel.evaluate(testRI, testRL)\n<\/code><\/pre>\n\n<p>I'm a beginner in cnn and keras and I'm trying to create a cnn to classify 257 different objects using keras. Above is my code. When I run the code, both training accuracy and validation are incredibly high. <\/p>\n\n<p>100\/102 [============================>.] - ETA: 1s - loss: 0.0265 - acc: 0.9960\n101\/102 [============================>.] - ETA: 0s - loss: 0.0264 - acc: 0.9960\n102\/102 [==============================] - 56s 553ms\/step - loss: 0.0264 - acc: 0.9960 - val_loss: 0.0250 - val_acc: 0.9961<\/p>\n\n<p>Could anyone tell me what's wrong with my code? Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53306907,"AnswerCount":1,"Tags":"<python><tensorflow><gradient-descent>","CreationDate":"2018-11-14T18:48:32.787","AcceptedAnswerId":53307124.0,"OwnerUserId":7957937.0,"Title":"tf.gradients: ValueError: Shapes must be equal rank, but are 2 and 1","Body":"<p>I need to get the gradient of weights and biases with tf.gradients():<\/p>\n\n<pre><code>        x = tf.placeholder(tf.float32, [batch_size, x_train.shape[1]])\n        y = tf.placeholder(tf.float32, [batch_size, y_train.shape[1]])\n        y_ = tf.placeholder(tf.float32, [batch_size, y_train.shape[1]])\n\n        Wx=tf.Variable(tf.random_normal(stddev=0.1,shape=[x_train.shape[1],n_hidden]))\n        Wy=tf.Variable(tf.random_normal(stddev=0.1,shape=[y_train.shape[1],n_hidden]))\n        b=tf.Variable(tf.constant(0.1,shape=[n_hidden]))\n\n        hidden_joint=tf.nn.relu((tf.matmul(x,Wx)+tf.matmul(y,Wy))+b)\n        hidden_marg=tf.nn.relu(tf.matmul(x,Wx)+tf.matmul(y_,Wy)+b)\n\n        Wout=tf.Variable(tf.random_normal(stddev=0.1,shape=[n_hidden, 1]))\n        bout=tf.Variable(tf.constant(0.1,shape=[1]))\n\n        out_joint=tf.matmul(hidden_joint,Wout)+bout\n        out_marg=tf.matmul(hidden_marg,Wout)+bout\n\n        optimizer = tf.train.AdamOptimizer(0.005)\n\n\n        t = out_joint\n        et = tf.exp(out_marg)\n\n        ex_delta_t = tf.reduce_mean(tf.gradients(t, tf.trainable_variables()))\n        ex_delta_et = tf.reduce_mean(tf.gradients(et, tf.trainable_variables()))\n<\/code><\/pre>\n\n<p>But I always get the following error: <\/p>\n\n<pre><code>  File \"\/home\/ferdi\/Documents\/mine\/mine.py\", line 77, in get_mi_batched\n    ex_delta_t = tf.reduce_mean(tf.gradients(t, tf.trainable_variables()))\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 1490, in reduce_mean\n    reduction_indices),\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 1272, in _ReductionDims\n    return range(0, array_ops.rank(x))\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 368, in rank\n    return rank_internal(input, name, optimize=True)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 388, in rank_internal\n    input_tensor = ops.convert_to_tensor(input)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1048, in convert_to_tensor\n    as_ref=False)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1144, in internal_convert_to_tensor\n    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 971, in _autopacking_conversion_function\n    return _autopacking_helper(v, dtype, name or \"packed\")\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 923, in _autopacking_helper\n    return gen_array_ops.pack(elems_as_tensors, name=scope)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 4689, in pack\n    \"Pack\", values=values, axis=axis, name=name)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3272, in create_op\n    op_def=op_def)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1790, in __init__\n    control_input_ops)\n  File \"\/home\/ferdi\/anaconda3\/envs\/ml_all\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1629, in _create_c_op\n    raise ValueError(str(e))\nValueError: Shapes must be equal rank, but are 2 and 1\n    From merging shape 3 with other shapes. for 'Rank\/packed' (op: 'Pack') with input shapes: [512,20], [10,20], [20], [20,1], [1].\n<\/code><\/pre>\n\n<p>If I reshape or do similar things, there are other error occurring. I know that there are many similar question but I still couldn't figure it out. What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"53307124","CreationDate":"2018-11-14T19:04:43.373","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p><strong>The solution:<\/strong><\/p>\n\n<pre><code>ex_delta_t = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in tf.gradients(t, tf.trainable_variables())], axis=0))\nex_delta_et = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in tf.gradients(et, tf.trainable_variables())], axis=0))\n<\/code><\/pre>\n\n<p><strong>Or the same code unfolded:<\/strong><\/p>\n\n<pre><code>grads_t_0 = tf.gradients(t, tf.trainable_variables())\ngrads_et_0 = tf.gradients(t, tf.trainable_variables())\n\ngrads_t = []\ngrads_et = []\nfor gt,get in zip(grads_t_0, grads_et_0):\n    grads_t.append(tf.reshape(gt, [-1]))\n    grads_et.append(tf.reshape(get, [-1]))\n\ngrads_t_flatten = tf.concat(grads_t, axis=0)\ngrads_et_flatten = tf.concat(grads_et, axis=0)\n\nex_delta_t = tf.reduce_mean(grads_t_flatten)\nex_delta_et = tf.reduce_mean(grads_et_flatten)\n<\/code><\/pre>\n\n<p><strong>Explanation:<\/strong><\/p>\n\n<p>You may get this error message, because your gradient functions<\/p>\n\n<pre><code>tf.gradients(t, tf.trainable_variables())\ntf.gradients(et, tf.trainable_variables()\n<\/code><\/pre>\n\n<p>returns multiply shaped tensors.\nAs a result your <code>tf.reduce_mean()<\/code> operation complains, that it can't work with this multiply shaped tensors. <\/p>\n\n<p>As a possibility to workaround this, you should first flatten than concatenate the gradient list, then pass it to the reduce_mean function.<\/p>\n\n<blockquote>\n  <p><strong>Let's see a simple example code to simulate the error and its solution!<\/strong><\/p>\n<\/blockquote>\n\n<pre><code>#You dummy gradients as the output of tf.gradients()\ngrad_wx = tf.constant(0.1, shape=[512, 20])\ngrad_wy = tf.constant(0.2, shape=[10, 20])\ngrad_b = tf.constant(0.3, shape=[20])\ngrad_wout = tf.constant(0.4, shape=[20, 1])\ngrad_bout = tf.constant(0.5, shape=[1])\n\ngrads_0 = [grad_wx, grad_wy, grad_b, grad_wout, grad_bout]\n\nsess = tf.Session()\n\nresult = tf.reduce_mean(grads_0)\nprint(sess.run(result)\n<\/code><\/pre>\n\n<p><strong>Out(error):<\/strong><\/p>\n\n<pre><code>ValueError: Shapes must be equal rank, but are 2 and 1\n    From merging shape 3 with other shapes. for 'Rank\/packed' (op: 'Pack') with input shapes: [512,20], [10,20], [20], [20,1], [1].\n<\/code><\/pre>\n\n<p><strong>Solution:<\/strong><\/p>\n\n<pre><code>result = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in grads_0], axis=0))\nprint(sess.run(result))\n<\/code><\/pre>\n\n<p><strong>Out(fixed):<\/strong><\/p>\n\n<pre><code>0.102899365\n<\/code><\/pre>\n"}]}
+{"QuestionId":53307555,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-14T19:32:52.000","AcceptedAnswerId":null,"OwnerUserId":9155143.0,"Title":"Is it worth learning TensorFlow with TensorFlow 2.0 coming out so soon?","Body":"<p>I have been programming for a few years and I am looking to get into using ML python libraries. I have familiarity with a few but I am looking to take a deeper dive. <\/p>\n\n<p><strong>I know these are more opinion based but my questions are:<\/strong><\/p>\n\n<ul>\n<li>Is it worth learning TF right now with 2.0 around the corner? <\/li>\n<li>If not is there something that would be more practical to learn from a professional perspective in the mean time?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"53307630","CreationDate":"2018-11-14T19:39:05.967","ParentId":null,"OwnerUserId":"1762211","Title":null,"Body":"<p>Based on the <a href=\"https:\/\/www.tensorflow.org\/guide\/version_compat\" rel=\"nofollow noreferrer\">changelog<\/a>: <\/p>\n\n<blockquote>\n  <p>Only the public APIs of TensorFlow are backwards compatible across\n  minor and patch versions. The public APIs consist of<\/p>\n  \n  <ul>\n  <li>All the documented Python functions and classes in the tensorflow module and its submodules, except for ..<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>In nutshell, most of the python API won't change, so what you will with the current version is transferrable to 2.0. <\/p>\n"}]}
+{"QuestionId":53307748,"AnswerCount":1,"Tags":"<tensorflow><tensor><coreml><object-detection-api>","CreationDate":"2018-11-14T19:48:11.163","AcceptedAnswerId":null,"OwnerUserId":7110470.0,"Title":"Extract Grind-Anchors from Object-Detection API Model","Body":"<p>I'm currently trying to get my SSDLite Network, which I trained with the Tensroflow Object-detection API working, with iOS.<\/p>\n\n<p>So I'm using the Open Source Code of <a href=\"https:\/\/github.com\/vonholst\/SSDMobileNet_CoreML\" rel=\"nofollow noreferrer\">SSDMobileNet_CoreML<\/a>.<\/p>\n\n<p>The Graph allready works with some limitations. For running on iOS I had to extract the FeatureExtractor from my Graph and where unable to keep Preprocessor, Posprocessor and MutlipleGrindAnchorBox, same as they did in <a href=\"https:\/\/github.com\/vonholst\/SSDMobileNet_CoreML\" rel=\"nofollow noreferrer\">SSDMobileNet_CoreML<\/a>.<\/p>\n\n<p>Here you can see the <a href=\"https:\/\/github.com\/vonholst\/SSDMobileNet_CoreML\/blob\/master\/SSDMobileNet\/SSDMobileNet\/Anchors.swift\" rel=\"nofollow noreferrer\">Anchors<\/a> they have used.<\/p>\n\n<p>So cause my Anchors seem to be a little different I tried to undestand how they got this array. <\/p>\n\n<p>So I found in an GitHub <a href=\"https:\/\/github.com\/tf-coreml\/tf-coreml\/issues\/107#issuecomment-359675509\" rel=\"nofollow noreferrer\">Issue<\/a> an explenation, where the User who created the <a href=\"https:\/\/github.com\/vonholst\/SSDMobileNet_CoreML\/blob\/master\/SSDMobileNet\/SSDMobileNet\/Anchors.swift\" rel=\"nofollow noreferrer\">Anchors<\/a> explains how he got them. <\/p>\n\n<p>He says:<\/p>\n\n<blockquote>\n  <p>I just exported them out of the Tensorflow Graph from the import\/MultipleGridAnchorGenerator\/Identity tensor<\/p>\n<\/blockquote>\n\n<p>I allready found the matching tensor in my Graph but I don't know how to export the Graph and retrive the correct Anchor encoding.\nCan sombody explain this to me?<\/p>\n","answers":[{"AnswerId":"53308471","CreationDate":"2018-11-14T20:46:45.760","ParentId":null,"OwnerUserId":"7110470","Title":null,"Body":"<p>I allready figured it out. A little below quote was a link to a <a href=\"https:\/\/gist.github.com\/vincentchu\/5f2f669aeb6df9864ec6c43252261269\" rel=\"nofollow noreferrer\">Python Notebook<\/a> which explains everything in detail.<\/p>\n"}]}
+{"QuestionId":53307943,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network><semantic-segmentation>","CreationDate":"2018-11-14T20:01:55.203","AcceptedAnswerId":null,"OwnerUserId":3241790.0,"Title":"Keras combine value of two loss funcation","Body":"<p>I have a model content one encoder and two decoder with two loss function:<\/p>\n\n<pre><code>input_shape = (384, 512, 3)\nmodel = Model(inputs=input, outputs=[1_features, 2_features])\nmodel = build_model(input_shape, 3)\nlosses = {\n\"loss1_output\": \"categorical_crossentropy\",\n\"loss2_output\": \"categorical_crossentropy\"}\n\nlossWeights = {\"loss1_output\": 1.0, \"loss2_output\": 1.0}\nEPOCHS = 50\nINIT_LR = 1e-3\n\nopt = Adam(lr=INIT_LR, decay=INIT_LR \/ EPOCHS)\nmodel.compile(optimizer=opt, loss=losses, loss_weights=lossWeights,\n          metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>I would combine the value for both those losses in one loss value and backward the result of the combination.\nMy question is close to this <a href=\"https:\/\/stackoverflow.com\/questions\/46163909\/keras-combining-two-losses-with-adjustable-weights\">one<\/a> which I read and tried and I found the model called the loss function one time for each branch (output).<\/p>\n","answers":[]}
+{"QuestionId":53307954,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-14T20:02:54.153","AcceptedAnswerId":53320326.0,"OwnerUserId":5623899.0,"Title":"TensorFlow Custom Estimator predict throwing value error","Body":"<p>Note: this question has an accompanying, documented <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">Colab<\/a> notebook.<\/p>\n\n<p>TensorFlow's documentation can, at times, leave a lot to be desired. Some of the older docs for lower level apis seem to have been expunged, and most newer documents point towards using higher level apis such as TensorFlow's subset of <code>keras<\/code> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>estimators<\/code><\/a>. This would not be so problematic if the higher level apis did not so often rely closely on their lower levels. Case in point, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>estimators<\/code><\/a> (especially the <code>input_fn<\/code> when using TensorFlow Records). <\/p>\n\n<p>Over the following Stack Overflow posts:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/52035692\/tensorflow-v1-10-store-images-as-byte-strings-or-per-channel\">Tensorflow v1.10: store images as byte strings or per channel?<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/52064866\/tensorflow-1-10-tfrecorddataset-recovering-tfrecords\">Tensorflow 1.10 TFRecordDataset - recovering TFRecords<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/52874647\/tensorflow-v1-10-why-is-an-input-serving-receiver-function-needed-when-checkpoi\">Tensorflow v1.10+ why is an input serving receiver function needed when checkpoints are made without it?<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/52641737\/tensorflow-1-10-custom-estimator-early-stopping-with-train-and-evaluate\">TensorFlow 1.10+ custom estimator early stopping with train_and_evaluate<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/53226898\/tensorflow-custom-estimator-stuck-when-calling-evaluate-after-training\">TensorFlow custom estimator stuck when calling evaluate after training<\/a><\/li>\n<\/ul>\n\n<p>and with the gracious assistance of the TensorFlow \/ StackOverflow community, we have moved closer to doing what the TensorFlow <a href=\"https:\/\/www.tensorflow.org\/guide\/custom_estimators\" rel=\"nofollow noreferrer\">\"Creating Custom Estimators\" guide<\/a> has not, demonstrating how to make an estimator one might actually use in practice (rather than toy example) e.g. one which:<\/p>\n\n<ul>\n<li>has a validation set for early stopping if performance worsen, <\/li>\n<li>reads from TF Records because many datasets are larger than the TensorFlow recommend 1Gb for in memory, and <\/li>\n<li>that saves its best version whilst training<\/li>\n<\/ul>\n\n<p>While I still have many questions regarding this (from the best way to encode data into a TF Record, to what exactly the <code>serving_input_fn<\/code> expects), there is one question that stands out more prominently than the rest:<\/p>\n\n<p>How to predict with the custom estimator we just made?<\/p>\n\n<p>Under the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#predict\" rel=\"nofollow noreferrer\">predict<\/a>, it states:<\/p>\n\n<blockquote>\n  <p><code>input_fn<\/code>: A function that constructs the features. Prediction continues until <code>input_fn<\/code> raises an end-of-input exception (<code>tf.errors.OutOfRangeError<\/code> or <code>StopIteration<\/code>). See Premade Estimators for more information. The function should construct and return one of the following:<\/p>\n  \n  <ul>\n  <li>A tf.data.Dataset object: Outputs of Dataset object must have same constraints as below.<\/li>\n  <li>features: A tf.Tensor or a dictionary of string feature name to Tensor. features are consumed by model_fn. They should satisfy the expectation of model_fn from inputs.<\/li>\n  <li>A tuple, in which case the first item is extracted as features.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>(perhaps) Most likely, if one is using <code>estimator.predict<\/code>, they are using data in memory such as a dense tensor (because a held out test set would likely go through <code>evaluate<\/code>).<\/p>\n\n<p>So I, in the accompanying <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">Colab<\/a>, create a single dense example, wrap it up in a <code>tf.data.Dataset<\/code>, and call <code>predict<\/code> to get a <code>ValueError<\/code>.<\/p>\n\n<p>I would greatly appreciate it if someone could explain to me how I can:<\/p>\n\n<ol>\n<li>load my saved estimator<\/li>\n<li>given a dense, in memory example, predict the output with the estimator<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53320326","CreationDate":"2018-11-15T13:14:34.443","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<pre><code>to_predict = random_onehot((1, SEQUENCE_LENGTH, SEQUENCE_CHANNELS))\\\n        .astype(tf_type_string(I_DTYPE))\npred_features = {'input_tensors': to_predict}\n\npred_ds = tf.data.Dataset.from_tensor_slices(pred_features)\npredicted = est.predict(lambda: pred_ds, yield_single_examples=True)\n\nnext(predicted)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Tensor(\"IteratorV2:0\", shape=(), dtype=resource) must be from the same graph as Tensor(\"TensorSliceDataset:0\", shape=(), dtype=variant).<\/p>\n<\/blockquote>\n\n<p>When you use the <code>tf.data.Dataset<\/code> module, it actually defines an input graph which is independant from the model graph. What happens here is that you first created a small graph by calling <code>tf.data.Dataset.from_tensor_slices()<\/code>, then the estimator API created a second graph by calling <code>dataset.make_one_shot_iterator()<\/code> automatically. These 2 graphs can't communicate so it throws an error.<\/p>\n\n<p>To circumvent this, you should never create a dataset outside of estimator.train\/evaluate\/predict. This is why everything data related is wrapped inside input functions.<\/p>\n\n<pre><code>def predict_input_fn(data, batch_size=1):\n  dataset = tf.data.Dataset.from_tensor_slices(data)\n  return dataset.batch(batch_size).prefetch(None)\n\npredicted = est.predict(lambda: predict_input_fn(pred_features), yield_single_examples=True)\nnext(predicted)\n<\/code><\/pre>\n\n<p>Now, the graph is not created outside of the predict call.<\/p>\n\n<p>I also added <code>dataset.batch()<\/code> because the rest of your code expect batched data and it was throwing a shape error. Prefetch just speed things up.<\/p>\n"}]}
+{"QuestionId":53308711,"AnswerCount":1,"Tags":"<python-2.7><theano><pymc3>","CreationDate":"2018-11-14T21:05:37.063","AcceptedAnswerId":53324788.0,"OwnerUserId":10654030.0,"Title":"TensorVariable to Array","Body":"<p>I'm trying to evaluate a theano TensorValue expression:<\/p>\n\n<pre><code>import pymc3\nimport numpy as np\nwith pymc3.Model():\n    growth = pymc3.Normal('growth_%s' % 'some_name', 0, 10)\n\nx = np.arange(4)\n(x * growth).eval()\n<\/code><\/pre>\n\n<p>but get the error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/gof\/graph.py\", line 522, in eval\n    self._fn_cache[inputs] = theano.function(inputs, self)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/compile\/function.py\", line 317, in function\n    output_keys=output_keys)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/compile\/pfunc.py\", line 486, in pfunc\n    output_keys=output_keys)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/compile\/function_module.py\", line 1839, in orig_function\n    name=name)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/compile\/function_module.py\", line 1487, in __init__\n    accept_inplace)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/compile\/function_module.py\", line 181, in std_fgraph\n    update_mapping=update_mapping)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/gof\/fg.py\", line 175, in __init__\n    self.__import_r__(output, reason=\"init\")\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/gof\/fg.py\", line 346, in __import_r__\n    self.__import__(variable.owner, reason=reason)\n  File \"\/home\/danna\/.virtualenvs\/lib\/python2.7\/site-packages\/theano\/gof\/fg.py\", line 391, in __import__\n    raise MissingInputError(error_msg, variable=r)\ntheano.gof.fg.MissingInputError: Input 0 of the graph (indices start from 0), used to compute InplaceDimShuffle{x}(growth_some_name), was not provided and not given a value. Use the Theano flag exception_verbosity='high', for more information on this error.\n<\/code><\/pre>\n\n<p>I tried \nCan someone please help me see what the theano variables actually output?\nThank you!\nI'm using Python 2.7 and theano 1.0.3<\/p>\n","answers":[{"AnswerId":"53324788","CreationDate":"2018-11-15T17:19:09.520","ParentId":null,"OwnerUserId":"570918","Title":null,"Body":"<p>While PyMC3 distributions are <code>TensorVariable<\/code> objects, they don't technical have any values to be evaluated outside of sampling. If you want values, you have to at least run sampling on the model:<\/p>\n\n<pre><code>with pymc3.Model():\n    growth = pymc3.Normal('growth', 0, 10)\n\n    trace = pymc3.sample(10)\n\nx = np.arange(4)\nx[:, np.newaxis]*trace['growth'] \n<\/code><\/pre>\n\n<p>If you want to view node values during sampling, you'd need to use <code>theano.tensor.printing.Print<\/code> objects. For more info, see <a href=\"https:\/\/docs.pymc.io\/notebooks\/howto_debugging.html\" rel=\"nofollow noreferrer\">the PyMC3 debugging tips<\/a>.<\/p>\n"}]}
+{"QuestionId":53309074,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-14T21:35:16.023","AcceptedAnswerId":null,"OwnerUserId":10654317.0,"Title":"Tensorflow, movie reviews predict","Body":"<p>I am trying to make the following tutorial: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_text_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_text_classification<\/a><\/p>\n\n<pre><code>train_data = keras.preprocessing.sequence.pad_sequences(train_data,\n                                                        value=word_index[\"&lt;PAD&gt;\"],\n                                                        padding='post',\n                                                        maxlen=256)\n\ntest_data = keras.preprocessing.sequence.pad_sequences(test_data,\n                                                       value=word_index[\"&lt;PAD&gt;\"],\n                                                       padding='post',\n                                                       maxlen=256)\n\nvocab_size = 10000\n\nmodel = keras.Sequential()\nmodel.add(keras.layers.Embedding(vocab_size, 16))\nmodel.add(keras.layers.GlobalAveragePooling1D())\nmodel.add(keras.layers.Dense(16, activation=tf.nn.relu))\nmodel.add(keras.layers.Dense(1, activation=tf.nn.sigmoid))\nmodel.summary()\n\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(),\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\n\nprediction = model.predict(test_data[5])\nprint(prediction)\n<\/code><\/pre>\n\n<p>Why does the prediction return an array of 256? And not 0 or 1? How to do ?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"53309738","CreationDate":"2018-11-14T22:32:50.957","ParentId":null,"OwnerUserId":"847000","Title":null,"Body":"<p>you should try to pass list of numpy arrays, as far as I remember<\/p>\n\n<p>Try something like:<\/p>\n\n<pre><code>model.predict(np.array([test_data[5]]))\n<\/code><\/pre>\n\n<p>Does it what you expected?\nHope this helps!<\/p>\n"}]}
+{"QuestionId":53309192,"AnswerCount":1,"Tags":"<tensorflow><nlp><similarity><spacy><sentence-similarity>","CreationDate":"2018-11-14T21:45:36.767","AcceptedAnswerId":null,"OwnerUserId":2073937.0,"Title":"Similarity between two lists of documents","Body":"<p>I need to find the similarity between two lists of the short texts in Python.\nTexts can be 1-4 word long. The length of the lists can be 10K each.\nI didn't find how to do this effectively in spaCy. Maybe other packages can do this?\nI assume the words are represented by a vector (300d), but any other options are also Ok.\nThis task can be done in a cycle, but there should be a more effective way for sure. This task fits the TensorFlow, pyTorch, and similar packages, but I'm not familiar with details of these packages.<\/p>\n","answers":[{"AnswerId":"53319264","CreationDate":"2018-11-15T12:13:29.347","ParentId":null,"OwnerUserId":"3134774","Title":null,"Body":"<p>I think your question is ambiguous - You might mean to produce a single similarity score for the similarity of the average of list 1 vs the average of list 2. I'm assuming that you want a similarity score for each combination of items from the two lists. For 10K items per list, that will produce 10K pow 2 = 100M similarity scores. <\/p>\n\n<pre><code>import spacy\nspacyModel = spacy.load('en')\n\nlist1 = [\"hello, example 1\", \"right, second example\"]\nlist2 = [\"hello, example 1 in the second list\", \"And now for something completely different\"]\n\nlist1SpacyDocs = [spacyModel(x) for x in list1]\nlist2SpacyDocs = [spacyModel(x) for x in list2]\n\nsimilarityMatrix = [[x.similarity(y) for x in list1SpacyDocs] for y in list2SpacyDocs]\n\nprint(similarityMatrix)\n[[0.8537950408055295, 0.8852732956832498], [0.5802435148988874, 0.7643245611465626]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53309394,"AnswerCount":1,"Tags":"<javascript><performance><tensorflow><tensorflow.js>","CreationDate":"2018-11-14T22:02:27.260","AcceptedAnswerId":null,"OwnerUserId":4610514.0,"Title":"TensorFlow.js speed in the browser","Body":"<p>I've trained a simple bidirectional LSTM network in Keras (20 units) and exported the model via<\/p>\n\n<pre><code>tfjs.converters.save_keras_model(model, 'myModel')\n<\/code><\/pre>\n\n<p>The model is 53kb large. In my JavaScript app, I load the model like this<\/p>\n\n<pre><code>var model;\nasync function loadModel() {;\n    model = await tf.loadModel('https:\/\/example.com\/myModel.json');\n}\n<\/code><\/pre>\n\n<p>and afterwards I run my predictions with<\/p>\n\n<pre><code>async function predict(input) {\n    var pred = model.predict(input);\n    ...\n}\n<\/code><\/pre>\n\n<p>It takes 5-6 seconds till the model is loaded, this is fine. But what bothers me is that every call of predict() also takes 5-6 seconds. Every time. For my use case, I'd need the prediction to be extremely fast, 1 second or less.<\/p>\n\n<p>My question is: Is this normal? Or is something wrong with my code?<\/p>\n\n<p><em>Edit: Here is a codepen: <a href=\"https:\/\/codepen.io\/anon\/pen\/XygXRP\" rel=\"nofollow noreferrer\">https:\/\/codepen.io\/anon\/pen\/XygXRP<\/a><\/em><\/p>\n\n<p>BTW, model.predict is blocking the UI - how can I prevent that?<\/p>\n","answers":[{"AnswerId":"53467837","CreationDate":"2018-11-25T13:14:59.140","ParentId":null,"OwnerUserId":"3092734","Title":null,"Body":"<p>Your UI is being blocked because you have not asked the thread to await the results of the prediction, this means that it is running synchronously instead of asynchronously. You can fix this by using the await keyword e.g. var pred = await model.predict(input).<\/p>\n\n<p>The rest of your code appears to be fine and so it looks like the delay is coming from your actual model as I saw my CPU was barely taxed to run your model.<\/p>\n\n<p>It is worth reading the tensorflowjs blog <a href=\"https:\/\/medium.com\/tensorflow\/introducing-the-model-optimization-toolkit-for-tensorflow-254aca1ba0a3\" rel=\"nofollow noreferrer\">post<\/a> as they give you examples of how you can improve the efficiency of models to make faster inferences in the browser.<\/p>\n"}]}
+{"QuestionId":53309967,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm><rnn>","CreationDate":"2018-11-14T22:57:15.300","AcceptedAnswerId":53314287.0,"OwnerUserId":6937452.0,"Title":"ValueError: Error when checking target: expected dense_3 to have 2 dimensions, but got array with shape (500, 10, 14)","Body":"<p>Keras: 2.1.6, python 3.6, tensorflow 1.8.0<\/p>\n\n<p>I am trying to train a Sequence Model that has two LSTM layers and 3 dense layers. I had done some data preparation beforehand and setup my data in the format the LSTM layer requires it, i.e. <code>(n_samples, n_timesteps, n_features)<\/code>. My data has 14 features and is actually a long sequence of 5000 steps thus I have broken it down to 500 samples to 10 time steps each. Once I completed that, I started with below model, but soon ran into the error of Input shape for the last layer. I tried using Sequential and Functional API both yield the same error.<\/p>\n\n<pre><code>import keras \nfrom keras import callbacks\nimport tensorflow as tf\nfrom keras.models import Model\nfrom keras.layers import Input, Dense\nfrom keras.layers import LSTM\n\n\nX_input = Input(X_train.shape[1:]);\n\n## First LSTM Layer\nX = LSTM(10, return_sequences=True, input_shape = (10,14), name = 'LSTM_1')(X_input);\n\n## Second LSTM Layer\nX = LSTM(10)(X);\n\n## First Dense Layer\nX = Dense(10, activation='relu', name = 'dense_1')(X)\n\n## Second Dense Layer\nX = Dense(5, activation='relu', name = 'dense_2')(X)\n\n## Final Dense Layer\nX = Dense(1, activation='relu', name = 'dense_3')(X)\n\n    ##The model object\nmodel = Model(inputs = X_input, outputs = X, name='LSTMModel')\n\nmodel.compile(optimizer = \"Adam\" , loss = \"mean_squared_error\", metrics = ['mean_squared_error','cosine', 'mae']);\n\nModel.fit(x = X_train, y = Y_train, epochs = 300, callbacks=[tensorboard], validation_data=(X_eval,Y_eval));\n<\/code><\/pre>\n\n<p>My data is of shape <code>(500,10,14)<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; X_train.shape\n(500,10,14)\n<\/code><\/pre>\n\n<p>And my model summary looks like such:<\/p>\n\n<pre><code>model.summary()\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 10, 14)            0\n_________________________________________________________________\nLSTM_1 (LSTM)                (None, 10, 10)            1000\n_________________________________________________________________\nLSTM_2 (LSTM)                (None, 10)                840\n_________________________________________________________________\ndense_1 (Dense)              (None, 10)                110\n_________________________________________________________________\ndense_2 (Dense)              (None, 5)                 55\n_________________________________________________________________\ndense_3 (Dense)              (None, 1)                 6\n=================================================================\nTotal params: 2,011\nTrainable params: 2,011\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Although, I still get the error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_3 to have 2 dimensions, but got array with shape (500, 10, 14)\n<\/code><\/pre>\n\n<p>The shape of my labels are as follows:<\/p>\n\n<pre><code>X_train = np.reshape(Train_data_scaled.values,(500,10,14));\nY_train = np.reshape(Train_labels_scaled.values,(500,10,1));\nX_eval = np.reshape(Validation_data_scaled.values,(10,10,14));\nY_eval = np.reshape(Validation_labels_scaled.values,(10,10,1));\n<\/code><\/pre>\n\n<p>After adding the layer of RepeatVector I find another issue here is the stack trace of the same.<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 10, 14)            0\n_________________________________________________________________\nLSTM_1 (LSTM)                (None, 10)                1000\n_________________________________________________________________\nrepeat_vector_1 (RepeatVecto (None, 10, 10)            0\n_________________________________________________________________\nLSTM_2 (LSTM)                (None, 10, 10)            840\n_________________________________________________________________\ndense_1 (Dense)              (None, 10, 10)            110\n_________________________________________________________________\ndense_2 (Dense)              (None, 10, 5)             55\n_________________________________________________________________\ndense_3 (Dense)              (None, 10, 1)             6\n=================================================================\nTotal params: 2,011\nTrainable params: 2,011\nNon-trainable params: 0\n_________________________________________________________________\nTraceback (most recent call last):\n  File \".\\lstm.py\", line 76, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 126, in run\n    _sys.exit(main(argv))\n  File \".\\lstm.py\", line 67, in main\n    Hist =  Model.fit(x = X_train, y = Y_train, epochs = 300,batch_size=10, callbacks=[tensorboard], validation_data=(X_eval,Y_eval));\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\engine\\training.py\", line 1630, in fit\n    batch_size=batch_size)\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\engine\\training.py\", line 1480, in _standardize_user_data\n    exception_prefix='target')\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\keras\\engine\\training.py\", line 123, in _standardize_input_data\n    str(data_shape))\nValueError: Error when checking target: expected dense_3 to have shape (10, 1) but got array with shape (10, 14)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53314287","CreationDate":"2018-11-15T07:22:59.013","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since you want to predict value of a story 10 days in the future you need to set the <code>return_sequences<\/code> argument of second LSTM layer to <code>True<\/code> to make the output shape of the whole model <code>(None, 10, 1)<\/code>:<\/p>\n\n<pre><code>## Second LSTM Layer\nX = LSTM(10, return_sequences=True)(X)\n<\/code><\/pre>\n\n<p>Further, a more general solution to predict the values <code>d<\/code> days in the future is to use a <code>RepeatVector<\/code> layer right after the first LSTM layer. This time you need to set <code>return_sequences<\/code> argument of the first LSTM layer to <code>False<\/code>:<\/p>\n\n<pre><code>d = 5  # how many days in the future you want to predict?\n\n## First LSTM Layer\nX = LSTM(10, input_shape = (10,14), name = 'LSTM_1')(X_input);\n\nX = RepeatVector(d)(X)\n\n## Second LSTM Layer\nX = LSTM(10, return_sequences=True)(X)\n<\/code><\/pre>\n\n<p>It's as if the first LSTM layer encodes the input data and the second LSTM layer predicts the future value(s) based on that encoded representation. Also, it is needless to say that the shape of label arrays (i.e. <code>y_train<\/code>) must also be <code>(n_samples, d, n_feats)<\/code>.<\/p>\n"}]}
+{"QuestionId":53310539,"AnswerCount":1,"Tags":"<javascript><python><tensorflow><tensorflow.js>","CreationDate":"2018-11-14T23:56:33.870","AcceptedAnswerId":53316830.0,"OwnerUserId":1288230.0,"Title":"Difference between Tensorflow and Tensorflow.js?","Body":"<p>I was expecting these two programs to return the same result:<\/p>\n\n<p><em>main.js<\/em>:<\/p>\n\n<pre><code>const tf = require('@tensorflow\/tfjs');\nrequire('@tensorflow\/tfjs-node');\n\nvar x = tf.tensor([0.101,0.102,0.103,0.104,0.105,0.106,0.107,0.108,0.109,0.110,0.111,0.112,0.113,0.114,0.115,0.116,0.117,0.118,0.119,0.120,0.121,0.122,0.123,0.124,0.125,0.126,0.127,0.128,0.129,0.130,0.131,0.132,0.133,0.134,0.135,0.136,0.137,0.138,0.139,0.140,0.141,0.142,0.143,0.144,0.145,0.146,0.147,0.148,0.149,0.150,0.151,0.152,0.153,0.154,0.155,0.156,0.157,0.158,0.159,0.160,0.161,0.162,0.163,0.164,0.165,0.166,0.167,0.168,0.169,0.170,0.171,0.172,0.173,0.174,0.175,0.176,0.177,0.178,0.179,0.180,0.181,0.182,0.183,0.184,0.185,0.186,0.187,0.188,0.189,0.190,0.191,0.192,0.193,0.194,0.195,0.196,0.197,0.198,0.199,0.200,0.201,0.202,0.203,0.204,0.205,0.206,0.207,0.208,0.209,0.210,0.211,0.212,0.213,0.214,0.215,0.216,0.217,0.218,0.219,0.220,0.221,0.222,0.223,0.224,0.225,0.226,0.227,0.228,0.229,0.230,0.231,0.232,0.233,0.234,0.235,0.236,0.237,0.238,0.239,0.240,0.241,0.242,0.243,0.244,0.245,0.246,0.247,0.248,0.249,0.250,0.251,0.252,0.253,0.254,0.255,0.256,0.257,0.258,0.259,0.260,0.261,0.262,0.263,0.264,0.265,0.266,0.267,0.268,0.269,0.270,0.271,0.272,0.273,0.274,0.275,0.276,0.277,0.278,0.279,0.280,0.281,0.282,0.283,0.284,0.285,0.286,0.287,0.288,0.289,0.290,0.291,0.292,0.293,0.294,0.295,0.296,0.297,0.298,0.299,0.300,0.301,0.302,0.303,0.304,0.305,0.306,0.307,0.308,0.309,0.310,0.311,0.312,0.313,0.314,0.315,0.316,0.317,0.318,0.319,0.320,0.321,0.322,0.323,0.324,0.325,0.326,0.327,0.328,0.329,0.330,0.331,0.332,0.333,0.334,0.335,0.336,0.337,0.338,0.339,0.340,0.341,0.342,0.343,0.344,0.345,0.346,0.347,0.348,0.349,0.350,0.351,0.352,0.353,0.354,0.355,0.356,0.357,0.358,0.359,0.360,0.361,0.362,0.363,0.364,0.365,0.366,0.367,0.368,0.369,0.370,0.371,0.372,0.373,0.374,0.375,0.376,0.377,0.378,0.379,0.380,0.381,0.382,0.383,0.384,0.385,0.386,0.387,0.388,0.389,0.390,0.391,0.392,0.393,0.394,0.395,0.396,0.397,0.398,0.399,0.400]).reshape([1,10,10,3])\nvar filters = tf.tensor([0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,0.10,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.20,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.30,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.40,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.50,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.60,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.70,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.80,0.81]).reshape([3,3,3,3])\nvar bias = tf.tensor([0.1,0.2,0.3])\n\nvar out = tf.add(\n  tf.conv2d(x, filters, [2, 2], 'same'),\n  bias)\n\nconsole.log(x.dataSync())\n<\/code><\/pre>\n\n<p><em>main.py<\/em>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport json\n\nx = tf.constant([0.101,0.102,0.103,0.104,0.105,0.106,0.107,0.108,0.109,0.110,0.111,0.112,0.113,0.114,0.115,0.116,0.117,0.118,0.119,0.120,0.121,0.122,0.123,0.124,0.125,0.126,0.127,0.128,0.129,0.130,0.131,0.132,0.133,0.134,0.135,0.136,0.137,0.138,0.139,0.140,0.141,0.142,0.143,0.144,0.145,0.146,0.147,0.148,0.149,0.150,0.151,0.152,0.153,0.154,0.155,0.156,0.157,0.158,0.159,0.160,0.161,0.162,0.163,0.164,0.165,0.166,0.167,0.168,0.169,0.170,0.171,0.172,0.173,0.174,0.175,0.176,0.177,0.178,0.179,0.180,0.181,0.182,0.183,0.184,0.185,0.186,0.187,0.188,0.189,0.190,0.191,0.192,0.193,0.194,0.195,0.196,0.197,0.198,0.199,0.200,0.201,0.202,0.203,0.204,0.205,0.206,0.207,0.208,0.209,0.210,0.211,0.212,0.213,0.214,0.215,0.216,0.217,0.218,0.219,0.220,0.221,0.222,0.223,0.224,0.225,0.226,0.227,0.228,0.229,0.230,0.231,0.232,0.233,0.234,0.235,0.236,0.237,0.238,0.239,0.240,0.241,0.242,0.243,0.244,0.245,0.246,0.247,0.248,0.249,0.250,0.251,0.252,0.253,0.254,0.255,0.256,0.257,0.258,0.259,0.260,0.261,0.262,0.263,0.264,0.265,0.266,0.267,0.268,0.269,0.270,0.271,0.272,0.273,0.274,0.275,0.276,0.277,0.278,0.279,0.280,0.281,0.282,0.283,0.284,0.285,0.286,0.287,0.288,0.289,0.290,0.291,0.292,0.293,0.294,0.295,0.296,0.297,0.298,0.299,0.300,0.301,0.302,0.303,0.304,0.305,0.306,0.307,0.308,0.309,0.310,0.311,0.312,0.313,0.314,0.315,0.316,0.317,0.318,0.319,0.320,0.321,0.322,0.323,0.324,0.325,0.326,0.327,0.328,0.329,0.330,0.331,0.332,0.333,0.334,0.335,0.336,0.337,0.338,0.339,0.340,0.341,0.342,0.343,0.344,0.345,0.346,0.347,0.348,0.349,0.350,0.351,0.352,0.353,0.354,0.355,0.356,0.357,0.358,0.359,0.360,0.361,0.362,0.363,0.364,0.365,0.366,0.367,0.368,0.369,0.370,0.371,0.372,0.373,0.374,0.375,0.376,0.377,0.378,0.379,0.380,0.381,0.382,0.383,0.384,0.385,0.386,0.387,0.388,0.389,0.390,0.391,0.392,0.393,0.394,0.395,0.396,0.397,0.398,0.399,0.400])\nx = tf.reshape(x, [1,10,10,3])\n\nfilters = tf.constant([0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,0.10,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.20,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.30,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.40,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.50,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.60,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.70,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.80,0.81])\nfilters = tf.reshape(filters, [3,3,3,3])\n\nbias = tf.constant([0.1,0.2,0.3])\n\nout = tf.math.add(\n  tf.nn.conv2d(x, filters, [1, 2, 2, 1], 'SAME'),\n  bias)\n\nwith tf.Session() as sess:\n  print(json.dumps(sess.run(out).tolist()))\n<\/code><\/pre>\n\n<p>Instead, they return very different values:<\/p>\n\n<ul>\n<li>The first 5 values of the tensor returned by the Javascript version are:   0.10100000351667404, 0.10199999809265137, 0.10300000011920929, 0.10400000214576721, 0.10499999672174454<\/li>\n<li>The first 5 values of the tensor returned by the Python version are: 1.8199000358581543, 1.9838900566101074, 2.1478798389434814, 1.8911800384521484, 2.058410167694092<\/li>\n<\/ul>\n\n<p>Did I miss something? Shouldn't I expect these two programs to produce the same outcome?<\/p>\n","answers":[{"AnswerId":"53316830","CreationDate":"2018-11-15T10:01:34.980","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>Both programs outputs the same result. If the js output is different, make sure you are using the latest version 0.13.3. <\/p>\n\n<p>Output of python code <a href=\"https:\/\/repl.it\/@kedevked\/convolution-tensorflow\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>The following is the js code<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>var x = tf.tensor([0.101,0.102,0.103,0.104,0.105,0.106,0.107,0.108,0.109,0.110,0.111,0.112,0.113,0.114,0.115,0.116,0.117,0.118,0.119,0.120,0.121,0.122,0.123,0.124,0.125,0.126,0.127,0.128,0.129,0.130,0.131,0.132,0.133,0.134,0.135,0.136,0.137,0.138,0.139,0.140,0.141,0.142,0.143,0.144,0.145,0.146,0.147,0.148,0.149,0.150,0.151,0.152,0.153,0.154,0.155,0.156,0.157,0.158,0.159,0.160,0.161,0.162,0.163,0.164,0.165,0.166,0.167,0.168,0.169,0.170,0.171,0.172,0.173,0.174,0.175,0.176,0.177,0.178,0.179,0.180,0.181,0.182,0.183,0.184,0.185,0.186,0.187,0.188,0.189,0.190,0.191,0.192,0.193,0.194,0.195,0.196,0.197,0.198,0.199,0.200,0.201,0.202,0.203,0.204,0.205,0.206,0.207,0.208,0.209,0.210,0.211,0.212,0.213,0.214,0.215,0.216,0.217,0.218,0.219,0.220,0.221,0.222,0.223,0.224,0.225,0.226,0.227,0.228,0.229,0.230,0.231,0.232,0.233,0.234,0.235,0.236,0.237,0.238,0.239,0.240,0.241,0.242,0.243,0.244,0.245,0.246,0.247,0.248,0.249,0.250,0.251,0.252,0.253,0.254,0.255,0.256,0.257,0.258,0.259,0.260,0.261,0.262,0.263,0.264,0.265,0.266,0.267,0.268,0.269,0.270,0.271,0.272,0.273,0.274,0.275,0.276,0.277,0.278,0.279,0.280,0.281,0.282,0.283,0.284,0.285,0.286,0.287,0.288,0.289,0.290,0.291,0.292,0.293,0.294,0.295,0.296,0.297,0.298,0.299,0.300,0.301,0.302,0.303,0.304,0.305,0.306,0.307,0.308,0.309,0.310,0.311,0.312,0.313,0.314,0.315,0.316,0.317,0.318,0.319,0.320,0.321,0.322,0.323,0.324,0.325,0.326,0.327,0.328,0.329,0.330,0.331,0.332,0.333,0.334,0.335,0.336,0.337,0.338,0.339,0.340,0.341,0.342,0.343,0.344,0.345,0.346,0.347,0.348,0.349,0.350,0.351,0.352,0.353,0.354,0.355,0.356,0.357,0.358,0.359,0.360,0.361,0.362,0.363,0.364,0.365,0.366,0.367,0.368,0.369,0.370,0.371,0.372,0.373,0.374,0.375,0.376,0.377,0.378,0.379,0.380,0.381,0.382,0.383,0.384,0.385,0.386,0.387,0.388,0.389,0.390,0.391,0.392,0.393,0.394,0.395,0.396,0.397,0.398,0.399,0.400]).reshape([1,10,10,3])\r\nvar filters = tf.tensor([0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,0.10,0.11,0.12,0.13,0.14,0.15,0.16,0.17,0.18,0.19,0.20,0.21,0.22,0.23,0.24,0.25,0.26,0.27,0.28,0.29,0.30,0.31,0.32,0.33,0.34,0.35,0.36,0.37,0.38,0.39,0.40,0.41,0.42,0.43,0.44,0.45,0.46,0.47,0.48,0.49,0.50,0.51,0.52,0.53,0.54,0.55,0.56,0.57,0.58,0.59,0.60,0.61,0.62,0.63,0.64,0.65,0.66,0.67,0.68,0.69,0.70,0.71,0.72,0.73,0.74,0.75,0.76,0.77,0.78,0.79,0.80,0.81]).reshape([3,3,3,3])\r\nvar bias = tf.tensor([0.1,0.2,0.3])\r\n\r\nvar out = tf.add(\r\n  tf.conv2d(x, filters, [2, 2], 'same'),\r\n  bias)\r\n\r\nout.print()<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.3\/dist\/tf.min.js\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53310724,"AnswerCount":1,"Tags":"<python><tensorflow><ubuntu-18.04><tesla>","CreationDate":"2018-11-15T00:19:47.060","AcceptedAnswerId":null,"OwnerUserId":10654822.0,"Title":"Why does Tensorflow 1.11.0 return CUDA_ERROR_NOT_SUPPORTED?","Body":"<p>My machine is <strong>Ubuntu 18.04.1 LTS<\/strong>, with CUDA has been successfully installed. The output of <code>$nvcc --version<\/code> is<\/p>\n\n<pre><code>nvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2017 NVIDIA Corporation\nBuilt on Fri_Sep__1_21:08:03_CDT_2017\nCuda compilation tools, release 9.0, V9.0.176\n<\/code><\/pre>\n\n<p>I have two GPUs of Tesla K80, and the command <strong>nvidia-smi<\/strong> shows: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ifbHG.png\" rel=\"nofollow noreferrer\">output of nvidia-smi<\/a><\/p>\n\n<p>I also tried to test with <code>.\/deviceQuery<\/code> from <code>NVIDIA_CUDA-9.0_Samples<\/code> and its output is as the followings:<\/p>\n\n<pre><code>CUDA Device Query (Runtime API) version (CUDART static linking)\n\nDetected 2 CUDA Capable device(s)`\n\n...\n\ndeviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 10.0, CUDA Runtime Version = 9.0, NumDevs = 2\nResult = PASS\n<\/code><\/pre>\n\n<p>However, once I install Tensorflow GPU <code>version 1.11.0<\/code> from pip, I couldn't open a Tensorflow session.<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; sess = tf.Session()\n<\/code><\/pre>\n\n<p>and it outputs:<\/p>\n\n<pre><code>2018-11-15 00:13:46.593039: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/quoctin.phan\/tools\/anaconda\/envs\/tensorflow_1.11\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 1511, in __init__\n    super(Session, self).__init__(target, graph, config=config)\n  File \"\/home\/quoctin.phan\/tools\/anaconda\/envs\/tensorflow_1.11\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 634, in __init__\n    self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)\ntensorflow.python.framework.errors_impl.InternalError: failed initializing StreamExecutor for CUDA device ordinal 0: Internal: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_NOT_SUPPORTED: operation not supported\n<\/code><\/pre>\n\n<p>I have tried to reinstall Tensorflow 1.12.0, but nothing changes. Your help is appreciated.<\/p>\n","answers":[{"AnswerId":"53797620","CreationDate":"2018-12-15T21:58:28.717","ParentId":null,"OwnerUserId":"10738833","Title":null,"Body":"<p>Do you think your problem might be somehow connected to Compute Capability? The problem is described in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22594\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>You can check them when you run deviceQuery.exe. <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/1024645\/compute-capability-check\/\" rel=\"nofollow noreferrer\">Here<\/a> is a thread about where to find it on windows distribution of CUDA package.<\/p>\n"}]}
+{"QuestionId":53310915,"AnswerCount":0,"Tags":"<python><machine-learning><scikit-learn><keras><text-classification>","CreationDate":"2018-11-15T00:48:11.937","AcceptedAnswerId":null,"OwnerUserId":7376349.0,"Title":"How to prevent overfitting in Keras sequential model?","Body":"<p>I am already adding dropout regularization. I am trying to build a multiclass text classification multilayer perceptron model.\nMy model:<\/p>\n\n<pre><code>model = Sequential([\n                Dropout(rate=0.2, input_shape=features),\n                Dense(units=64, activation='relu'),\n                Dropout(rate=0.2),\n                Dense(units=64, activation='relu'),\n                Dropout(rate=0.2),\n                Dense(units=16, activation='softmax')])\n<\/code><\/pre>\n\n<p>My model.summary():<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ndropout_1 (Dropout)          (None, 20000)             0\n_________________________________________________________________\ndense_1 (Dense)              (None, 64)                1280064\n_________________________________________________________________\ndropout_2 (Dropout)          (None, 64)                0\n_________________________________________________________________\ndense_2 (Dense)              (None, 64)                4160\n_________________________________________________________________\ndropout_3 (Dropout)          (None, 64)                0\n_________________________________________________________________\ndense_3 (Dense)              (None, 16)                1040\n=================================================================\nTotal params: 1,285,264\nTrainable params: 1,285,264\nNon-trainable params: 0\n_________________________________________________________________\nNone\nTrain on 6940 samples, validate on 1735 samples\n<\/code><\/pre>\n\n<p>I am getting: <\/p>\n\n<pre><code>Epoch 16\/1000\n - 4s - loss: 0.4926 - acc: 0.8719 - val_loss: 1.2640 - val_acc: 0.6640\nValidation accuracy: 0.6639769498140736, loss: 1.2639631692545559\n<\/code><\/pre>\n\n<p>The validation accuracy is ~20% less than the accuracy, and the validation loss is way higher than the training loss. <\/p>\n\n<p>I am already using dropout regularization, and using epochs = 1000, batch size = 512 and early stopping on val_loss.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[]}
+{"QuestionId":53311001,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><vgg-net>","CreationDate":"2018-11-15T01:02:10.173","AcceptedAnswerId":53311049.0,"OwnerUserId":2218351.0,"Title":"Cannot convert VGG16 layers to sequential","Body":"<p>I am trying to do image classification using VGG16 pre-trained model. For the same, I did the following:<\/p>\n\n<pre><code>vgg16_model = keras.applications.vgg16.VGG16()\n<\/code><\/pre>\n\n<p>The type of the model is as follows:<\/p>\n\n<pre><code>type(vgg16_model)\n<\/code><\/pre>\n\n<p>And the result is:<\/p>\n\n<pre><code>tensorflow.python.keras.engine.training.Model\n<\/code><\/pre>\n\n<p>Then, I defined a Sequential model as:<\/p>\n\n<pre><code>model = Sequential()\n<\/code><\/pre>\n\n<p>Then, I tried to convert the <code>vgg16_model<\/code> into sequential by:<\/p>\n\n<pre><code>for layer in vgg16_model.layers:\n    model.add(layer)\n<\/code><\/pre>\n\n<p>It shows me an error as follows:<\/p>\n\n<pre><code>TypeError: The added layer must be an instance of class Layer. Found: &lt; tensorflow.python.keras.engine.input_layer.InputLayer object at 0x1ddbce5e80&gt;**\n<\/code><\/pre>\n\n<p>It would be great if anyone could help me on this one.<\/p>\n","answers":[{"AnswerId":"53313141","CreationDate":"2018-11-15T05:45:45.667","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One simpler way to do this is to pass the layers directly to the Sequential model instance, instead of using a for loop:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\n\nvgg = VGG16(weights='imagenet', ...)\nmodel = Sequential(vgg.layers)\n<\/code><\/pre>\n"},{"AnswerId":"53311049","CreationDate":"2018-11-15T01:09:33.913","ParentId":null,"OwnerUserId":"2218351","Title":null,"Body":"<p>Solution:<\/p>\n\n<p>My mistake was that my import statement was:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\n<\/code><\/pre>\n\n<p>Then, again, when I initialised the model, I called it again as:<\/p>\n\n<pre><code>vgg16_model = keras.applications.vgg16.VGG16()\n<\/code><\/pre>\n\n<p>So, a stupid mistake on my part. The fix is as follows:<\/p>\n\n<pre><code>vgg16_model = VGG16()\n<\/code><\/pre>\n\n<p>I realise that the problem is very specific and might not be so much useful to the community. Still, I am posting the solution just in case someone else faces it again. <\/p>\n"}]}
+{"QuestionId":53311701,"AnswerCount":1,"Tags":"<tensorflow><multilabel-classification>","CreationDate":"2018-11-15T02:47:35.413","AcceptedAnswerId":null,"OwnerUserId":1494821.0,"Title":"Tensorflow multi-label with NCE or sampled softmax","Body":"<p>Are there any code examples for using Tensorflow's sampled_softmax_loss or nce_loss functions with multi-label problems? That is, where num_true is more than one?<\/p>\n\n<p>What follows is my attempt to create a wrapper for nce_loss() and sampled_softmax_loss() based Jeff Chao's work (<a href=\"https:\/\/github.com\/joelthchao\/keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/joelthchao\/keras<\/a>). In the following code, if you change num_true to 1, both samplers work. But with num_true > 1, both samplers throw slightly different exceptions involving tensor shape. <\/p>\n\n<p>The main program is a simple autoencoder that replicates the class of problem I'm trying to solve: multi-label testing with a huge number of output classes, with a Zipfian distribution. Comments and stack trace at the end.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport keras.layers as layers\nfrom keras.models import Model\nfrom keras import backend as K\nfrom keras import initializers,regularizers,constraints\nfrom keras.models import Model\nfrom keras.layers import Dense\nfrom keras.engine.base_layer import InputSpec\nfrom keras.engine.topology import Layer\nfrom keras.engine.input_layer import Input\n\nfrom tensorflow.keras.optimizers import Nadam, Adam\nnp.random.seed(10)\nimport random\n\ndef nce_loss_function(weights, biases, labels, inputs, num_sampled, num_classes, num_true):\n    if K.learning_phase() == 1:\n        loss = tf.nn.nce_loss(weights, biases, labels, inputs, num_sampled, num_classes, num_true,\n            partition_strategy=\"div\")\n    else:\n        logits = tf.matmul(inputs, tf.transpose(weights))\n        logits = tf.nn.bias_add(logits, biases)\n        labels_one_hot = tf.one_hot(labels, num_classes)\n        loss = tf.nn.sigmoid_cross_entropy_with_logits(\n            labels=labels_one_hot[:][0][:],\n            logits=logits)\n        loss = tf.reduce_sum(loss, axis=1)\n    return loss\n\ndef sampled_softmax_loss_function(weights, biases, labels, inputs, num_sampled, num_classes, num_true):\n    if K.learning_phase() == 1:\n        return tf.nn.sampled_softmax_loss(weights, biases, labels, inputs, num_sampled, num_classes, num_true, \n            partition_strategy=\"div\")\n    else:\n        logits = tf.matmul(inputs, tf.transpose(weights))\n        logits = tf.nn.bias_add(logits, biases)\n        labels_one_hot = tf.one_hot(labels, num_classes)\n        loss = tf.nn.softmax_cross_entropy_with_logits_v2(\n            labels=labels_one_hot,\n            logits=logits)\n        return loss\n\nclass Sampling(Layer):\n    \"\"\"Regular densely-connected NN layer with various sampling Loss.\n    `Sampling` implements the operation:\n    `output = dot(input, kernel) + bias`\n    `kernel` is a weights matrix created by the layer, and `bias` is a bias vector\n    created by the layer. Also, it adds a sampling Loss to the model.\n    See [reference](http:\/\/proceedings.mlr.press\/v9\/gutmann10a\/gutmann10a.pdf).\n    # Example\n    ```python\n        inputs = Input(shape=(4,))\n        target = Input(shape=(1,))  # sparse format, e.g. [1, 3, 2, 6, ...]\n        net = Dense(8)(inputs)\n        net = Sampling(units=128, num_sampled=32)([net, target])\n        model = Model(inputs=[inputs, target], outputs=net)\n        model.compile(optimizer='adam', loss=None)\n        x = np.random.rand(1000, 4)\n        y = np.random.randint(128, size=1000)\n        model.fit([x, y], None)\n    ```\n    # Arguments\n        units: Positive integer, dimensionality of the output space (num classes).\n        num_sampled: Positive integer, number of classes to sample in Sampling Loss.\n        type: 'sampled_softmax', 'nce'\n        num_true: Max # of positive classes, pad to this for variable inputs\n        kernel_initializer: Initializer for the `kernel` weights matrix\n            (see [initializers](..\/initializers.md)).\n        bias_initializer: Initializer for the bias vector\n            (see [initializers](..\/initializers.md)).\n        kernel_regularizer: Regularizer function applied to\n            the `kernel` weights matrix\n            (see [regularizer](..\/regularizers.md)).\n        bias_regularizer: Regularizer function applied to the bias vector\n            (see [regularizer](..\/regularizers.md)).\n        activity_regularizer: Regularizer function applied to\n            the output of the layer (its \"activation\").\n            (see [regularizer](..\/regularizers.md)).\n        kernel_constraint: Constraint function applied to\n            the `kernel` weights matrix\n            (see [constraints](..\/constraints.md)).\n        bias_constraint: Constraint function applied to the bias vector\n            (see [constraints](..\/constraints.md)).\n    # Input shape\n        Two tensors. First one is 2D tensor with shape: `(batch_size, input_dim)`.\n        Second one is 1D tensor with length `batch_size`\n    # Output shape\n        2D tensor with shape: `(batch_size, units)`.\n        For instance, for a 2D input with shape `(batch_size, input_dim)`,\n        the output would have shape `(batch_size, units)`.\n    \"\"\"\n\n    def __init__(self,\n                 units,\n                 num_sampled,\n                 type='sampled_softmax',\n                 num_true=1,\n                 kernel_initializer='glorot_uniform',\n                 bias_initializer='zeros',\n                 kernel_regularizer=None,\n                 bias_regularizer=None,\n                 activity_regularizer=None,\n                 kernel_constraint=None,\n                 bias_constraint=None,\n                 **kwargs):\n        if 'input_shape' not in kwargs and 'input_dim' in kwargs:\n            kwargs['input_shape'] = (kwargs.pop('input_dim'),)\n        super(Sampling, self).__init__(**kwargs)\n        self.units = units\n        self.num_sampled = num_sampled\n        if self.num_sampled &gt; self.units:\n            raise Exception('num_sample: {} cannot be greater than units: {}'.format(\n                num_sampled, units))\n        self.type = type\n        if not (self.type == 'nce' or self.type == 'sampled_softmax'):\n            raise Exception('type {} is not a valid sampling loss type'.format(type))\n        self.num_true = num_true\n        self.kernel_initializer = initializers.get(kernel_initializer)\n        self.bias_initializer = initializers.get(bias_initializer)\n        self.kernel_regularizer = regularizers.get(kernel_regularizer)\n        self.bias_regularizer = regularizers.get(bias_regularizer)\n        self.activity_regularizer = regularizers.get(activity_regularizer)\n        self.kernel_constraint = constraints.get(kernel_constraint)\n        self.bias_constraint = constraints.get(bias_constraint)\n        self.input_spec = [InputSpec(min_ndim=2), InputSpec(min_ndim=1)]\n        self.supports_masking = True\n\n    def build(self, input_shape):\n        assert len(input_shape) == 2\n        input_dim = input_shape[0][-1]\n\n        self.kernel = self.add_weight(shape=(input_dim, self.units),\n                                      initializer=self.kernel_initializer,\n                                      name='kernel',\n                                      regularizer=self.kernel_regularizer,\n                                      constraint=self.kernel_constraint)\n        self.bias = self.add_weight(shape=(self.units,),\n                                    initializer=self.bias_initializer,\n                                    name='bias',\n                                    regularizer=self.bias_regularizer,\n                                    constraint=self.bias_constraint)\n        self.input_spec[0] = InputSpec(min_ndim=2, axes={-1: input_dim})\n        self.built = True\n\n    def call(self, inputs):\n        pred, target = inputs\n        output = K.dot(pred, self.kernel)\n        output = K.bias_add(output, self.bias, data_format='channels_last')\n\n        # TODO : check train or test mode\n        if self.type == 'nce':\n            nce_loss = nce_loss_function(\n                K.transpose(self.kernel), self.bias, target, pred, self.num_sampled, self.units, self.num_true)\n            self.add_loss(K.mean(nce_loss))\n        else:\n            sampled_softmax_loss = sampled_softmax_loss_function(\n                K.transpose(self.kernel), self.bias, target, pred, self.num_sampled, self.units, self.num_true)\n            self.add_loss(K.mean(sampled_softmax_loss))\n        return output\n\n    def compute_output_shape(self, input_shape):\n        assert input_shape and len(input_shape) == 2\n        assert input_shape[0][-1]\n        output_shape = list(input_shape[0])\n        output_shape[-1] = self.units\n        return tuple(output_shape)\n\n    def get_config(self):\n        config = {\n            'units': self.units,\n            'num_sampled': self.num_sampled,\n            'kernel_initializer': initializers.serialize(self.kernel_initializer),\n            'bias_initializer': initializers.serialize(self.bias_initializer),\n            'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),\n            'bias_regularizer': regularizers.serialize(self.bias_regularizer),\n            'activity_regularizer': regularizers.serialize(self.activity_regularizer),\n            'kernel_constraint': constraints.serialize(self.kernel_constraint),\n            'bias_constraint': constraints.serialize(self.bias_constraint)\n        }\n        base_config = super(Sampling, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n\ndef fill_zipf(length, num_classes, num_true=1):\n  data_onehot = np.zeros((length, num_classes), dtype='float32')\n  data_labels = np.zeros((length, num_true), dtype='int32')\n  # all indexes outside of num_classes scattered in existing space\n  rand = np.random.zipf(1.3, length * num_true) % num_classes\n  for i in range(length):\n    for j in range(num_true):\n      k = rand[i]\n      data_onehot[i][k] = 1.0\n      data_labels[i][j] = k\n  return data_onehot, data_labels\n\n# number of test samples\nnum_train = 32*500\nnum_test = 32*500\nnum_valid = 100\nnum_epochs = 5\nnum_hidden = 10\n# number of classes\nnum_classes = 2000\n# number of samples for NCE\nnum_sampled = 24\n# number of labels\nnum_true = 1\n# type of negative sampler\nsampler_type='sampled_softmax'   \n\ninputs = Input(shape=(num_classes,))\ntarget = Input(shape=(num_true,), dtype=tf.int32)  # sparse format, e.g. [1, 3, 2, 6, ...]\nnet = Dense(num_classes)(inputs)\nnet = Dense(num_hidden, activation='relu')(net)\nnet = Sampling(units=num_classes, num_sampled=num_sampled, type=sampler_type)([net, target])\nmodel = Model(inputs=[inputs, target], outputs=net)\nmodel.compile(optimizer='adam', loss=None, metrics=['binary_crossentropy'])\nmodel.summary()\ntrain_input, train_output = fill_zipf(num_train, num_classes, num_true)\nvalid_input, valid_output = fill_zipf(num_valid, num_classes, num_true)\nhistory = model.fit([train_input, train_output], None, \n    validation_data=([valid_input, valid_output], None),\n    epochs=num_epochs, verbose=2)\n\ntest_input, test_output = fill_zipf(num_test, num_classes, num_true)\npredicts = model.predict([test_input, test_output], batch_size=32)\ncount = 0\nfor test in range(num_test):\n  pred = predicts[test]\n  imax = np.argmax(pred)\n  if imax == test_output[test]:\n    count += 1\nprint(\"Found {0} out of {1}\".format(count\/num_true, num_test))\n<\/code><\/pre>\n\n<p>This test works for the single-label case, both 'nce' and 'sampled_softmax'. But, when I set num_true to greater than one, both NCE and Sampled Softmax throw a tensor mismatch exception. <\/p>\n\n<pre><code>num_true=3\nwidth=2000\nsampler_type='sampled_softmax'\n<\/code><\/pre>\n\n<p>With these parameters, for Sampled Softmax, the code throws this exception trace:<\/p>\n\n<pre><code>  File \"postable_sampling_tests.py\", line 220, in &lt;module&gt;\n    epochs=num_epochs, verbose=2)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1039, in fit\n    validation_steps=validation_steps)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py\", line 199, in fit_loop\n    outs = f(ins_batch)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2715, in __call__\n    return self._call(inputs)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2675, in _call\n    fetched = self._callable_fn(*array_vals)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1399, in __call__\n    run_metadata_ptr)\n  File \"\/opt\/ds\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py\", line 526, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: logits and labels must be broadcastable: logits_size=[32,2000] labels_size=[96,2000]\n     [[{{node sampling_1\/softmax_cross_entropy_with_logits}} = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _class=[\"loc:@train...s_grad\/mul\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](sampling_1\/BiasAdd_1, sampling_1\/softmax_cross_entropy_with_logits\/Reshape_1)]]\n<\/code><\/pre>\n\n<p>32 is the batch_size. Clearly, something is num_true * batch_size but I don't know how to fix this. <\/p>\n\n<p>If we change the sampler to NCE:<\/p>\n\n<pre><code>num_true=3\nwidth=2000\nsampler_type='nce'\n<\/code><\/pre>\n\n<p>The final two lines of the exception stack:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [32,2000] vs. [3,2000]\n     [[{{node sampling_1\/logistic_loss\/mul}} = Mul[T=DT_FLOAT, _class=[\"loc:@training\/Adam\/gradients\/sampling_1\/logistic_loss\/mul_grad\/Reshape\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](sampling_1\/BiasAdd_1, sampling_1\/strided_slice_2)]]\n<\/code><\/pre>\n\n<p>In this case, the labels have not been multiplied by batch_size.<\/p>\n\n<p>What am I doing wrong? How can I get this wrapper system working for multi-label cases?<\/p>\n","answers":[{"AnswerId":"55388401","CreationDate":"2019-03-28T00:21:13.503","ParentId":null,"OwnerUserId":"3259896","Title":null,"Body":"<p>You can also use samples softmax with multiple labels, you just have to take the mean of each samples softmax<\/p>\n\n<pre><code>  embeddings = tf.get_variable( 'embeddings', \n    initializer= tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))\n\n  softmax_weights = tf.get_variable( 'softmax_weights',\n    initializer= tf.truncated_normal([vocabulary_size, embedding_size],\n                         stddev=1.0 \/ math.sqrt(embedding_size)))\n\n  softmax_biases = tf.get_variable('softmax_biases', \n    initializer= tf.zeros([vocabulary_size]),  trainable=False )\n\n  embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n  embed_reshaped = tf.reshape( embed, [batch_size*num_inputs, embedding_size] )\n\n  segments= np.arange(batch_size).repeat(num_inputs)\n\n  averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n  loss = tf.reduce_mean(\n    tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds,\n                               labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size))\n\n  optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) #Original learning rate was 1.0\n<\/code><\/pre>\n\n<p>from<\/p>\n\n<p><a href=\"https:\/\/github.com\/Santosh-Gupta\/Research2Vec\/blob\/master\/Research2VecTraining2.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Santosh-Gupta\/Research2Vec\/blob\/master\/Research2VecTraining2.ipynb<\/a><\/p>\n"}]}
+{"QuestionId":53311759,"AnswerCount":0,"Tags":"<windows><ubuntu><tensorflow><gpu><nvidia>","CreationDate":"2018-11-15T02:55:31.620","AcceptedAnswerId":null,"OwnerUserId":975425.0,"Title":"GPU compute cluster faster on Windows 10 than Ubuntu 18.04","Body":"<p>Info: My windows and Ubuntu system is the same. I just swap out my boot drive using a removable drive bay.<\/p>\n\n<p>*Not using Docker<\/p>\n\n<p>CPU: i9-7920X (128 GB DDR4)<\/p>\n\n<p>GPUs:\n2 Titan Vs and\nRTX 2070 (Display)<\/p>\n\n<p>Windows 10:\nNvidia Driver 416.81 in TCC cluster and\nCuda 10<\/p>\n\n<p>Ubuntu Desktop 18.04:\nNvidia Driver 410.72 (not sure if need to cluster on Linux) and\nCuda 10<\/p>\n\n<p>Test Software:\nRunning TensorFlow cifar10 esitimator - same training data - (link at bottom)<\/p>\n\n<p>python cifar10_main.py --data-dir=\/tmp --job-dir=\/tmp\/cifar10 --num-gpus=2 --train-steps=10000<\/p>\n\n<p>Windows completes in 420 seconds (7.0 minutes)\nUbuntu completes in 614 seconds (10.2 minutes)<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10_estimator\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10_estimator<\/a><\/p>\n\n<p>Why is Ubuntu so much slower? Do the GPUs need to be clustered on Linux or are the different drivers (Ubuntu vs. Windows) the issue?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53311812,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-11-15T03:03:52.127","AcceptedAnswerId":53388837.0,"OwnerUserId":6116330.0,"Title":"TensorFlow Estimator tracking timeline?","Body":"<p>A normal practice of timeline to track TensorFlow session is below:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.client import timeline\n\nx = tf.random_normal([1000, 1000])\ny = tf.random_normal([1000, 1000])\nres = tf.matmul(x, y)\n\n# Run the graph with full trace option\nwith tf.Session() as sess:\n    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\n    run_metadata = tf.RunMetadata()\n    sess.run(res, options=run_options, run_metadata=run_metadata)\n\n    # Create the Timeline object, and write it to a json\n    tl = timeline.Timeline(run_metadata.step_stats)\n    ctf = tl.generate_chrome_trace_format()\n    with open('timeline.json', 'w') as f:\n        f.write(ctf)\n<\/code><\/pre>\n\n<p>But now I am using <code>tf.estimator<\/code>, without defining a session explicitly.\nSo how and when should I define and use <code>tensorflow.python.client.timeline<\/code>\uff1f<\/p>\n","answers":[{"AnswerId":"53388837","CreationDate":"2018-11-20T08:22:01.497","ParentId":null,"OwnerUserId":"10309583","Title":null,"Body":"<p>try this:<\/p>\n\n<pre><code>hook = tf.train.ProfilerHook(save_steps=100, output_dir='\/tmp\/')\nclassifier.train(\n    input_fn=lambda: ltr_dataset.csv_input_fn(train_file_list, args.batch_size)\n    ,hooks=[hook]\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53311885,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><classification>","CreationDate":"2018-11-15T03:15:51.793","AcceptedAnswerId":null,"OwnerUserId":10245126.0,"Title":"Keras' fit_generator() for binary classification predictions always 50%","Body":"<p>I have set up a model to train on classifying whether an image is a certain video game or not. I <code>pre-scaled<\/code> my images into <code>250x250<\/code> pixels and have them separated into two folders (the two binary classes) labelled <code>0<\/code> and <code>1<\/code>. The amount of both classes are within <code>~100<\/code> of each other and I have around <code>3500<\/code> images in total.<\/p>\n\n<p>Here are photos of the training process, the model set up and some predictions: <a href=\"https:\/\/imgur.com\/a\/CN1b6LV\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/CN1b6LV<\/a><\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\n    rescale=1. \/ 255,\n    shear_range=0,\n    zoom_range=0,\n    horizontal_flip=True,\n    width_shift_range=0.1,\n    height_shift_range=0.1,\n    validation_split=0.2)\ntrain_generator = train_datagen.flow_from_directory(\n    'data\\\\',\n    batch_size=batchsize,\n    shuffle=True,\n    target_size=(250, 250),\n    subset=\"training\",\n    class_mode=\"binary\")\nval_generator = train_datagen.flow_from_directory(\n    'data\\\\',\n    batch_size=batchsize,\n    shuffle=True,\n    target_size=(250, 250),\n    subset=\"validation\",\n    class_mode=\"binary\")\npred_datagen = ImageDataGenerator(\n    rescale=1. \/ 255,\n    shear_range=0,\n    zoom_range=0,\n    horizontal_flip=False,\n    width_shift_range=0.1,\n    height_shift_range=0.1)\npred_generator = pred_datagen.flow_from_directory(\n    'batch_pred\\\\',\n    batch_size=30,\n    shuffle=False,\n    target_size=(250, 250))\n\n\nmodel = Sequential()\nmodel.add(Conv2D(input_shape=(250, 250, 3), filters=25, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=32, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=32, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(MaxPooling2D(pool_size=2,  padding=\"same\", strides=(2, 2)))\nmodel.add(BatchNormalization())\nmodel.add(Conv2D(filters=64, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=64, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=64, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(MaxPooling2D(pool_size=2, padding=\"same\", strides=(2, 2)))\nmodel.add(BatchNormalization())\nmodel.add(Conv2D(filters=128, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=128, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=128, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(MaxPooling2D(pool_size=2, padding=\"same\", strides=(2, 2)))\nmodel.add(Conv2D(filters=256, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=256, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(Conv2D(filters=256, kernel_size=3, activation=\"relu\", padding=\"same\"))\nmodel.add(MaxPooling2D(pool_size=2, padding=\"same\", strides=(2, 2)))\nmodel.add(BatchNormalization())\ndense = False\nif dense:\n    model.add(Flatten())\n    model.add(Dense(250, activation=\"relu\"))\n    model.add(BatchNormalization())\n    model.add(Dense(50, activation=\"relu\"))\nelse:\n    model.add(GlobalAveragePooling2D())\nmodel.add(Dense(1, activation=\"softmax\"))\nmodel.compile(loss='binary_crossentropy',\n              optimizer=Adam(0.0005), metrics=[\"acc\"])\ncallbacks = [EarlyStopping(monitor='val_acc', patience=200, verbose=1),\n             ModelCheckpoint(filepath=\"model_checkpoint.h5py\",\n                             monitor='val_acc', save_best_only=True, verbose=1)]\nmodel.fit_generator(\n      train_generator,\n      steps_per_epoch=train_generator.samples \/\/ batchsize,\n      validation_data=val_generator,\n      validation_steps=val_generator.samples \/\/ batchsize,\n      epochs=500,\n      callbacks=callbacks)\n<\/code><\/pre>\n\n<p>Everything appears to run correctly in terms of the <code>model<\/code> iterating the data by epoch, it finding the correct number of images etc. However, my predictions are always <code>50%<\/code> despite a good validation accuracy, low loss, high accuracy etc.<\/p>\n\n<p>I'm not sure what I'm doing wrong and any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"53311897","CreationDate":"2018-11-15T03:17:44.197","ParentId":null,"OwnerUserId":"10655346","Title":null,"Body":"<p>I think your problem is that you're using sigmoid for binary classification, your final layer activation function should be linear.<\/p>\n"},{"AnswerId":"53313039","CreationDate":"2018-11-15T05:34:38.387","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are using <code>softmax<\/code> on a Dense layer with one unit. Softmax function normalizes its input such that the sum of its elements becomes equal to one. So if it has one unit, then the output would be always 1. Instead, for binary classification you need to use <code>sigmoid<\/code> function as the activation function of last layer.<\/p>\n"}]}
+{"QuestionId":53311962,"AnswerCount":1,"Tags":"<python><tensorflow><keras><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-15T03:24:56.007","AcceptedAnswerId":null,"OwnerUserId":8100655.0,"Title":"TPU runs as slow as CPU when using keras_to_tpu_model in colab","Body":"<p>I use <code>tf.contrib.tpu.keras_to_tpu_model<\/code> to make my code be able to run on TPU,but it took 170 hours to finish an epoch while CPU took the same time and GPU took only  40 hours per epoch.I tried to adjust batch size but nothing changed.And I've tested the input function may take up 20% of the run time when running on GPU, so I think it's maybe not the main reason.     <\/p>\n\n<p>Here is my code:<a href=\"https:\/\/github.com\/WangHexie\/DHNE\/blob\/master\/src\/hypergraph_embedding.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/WangHexie\/DHNE\/blob\/master\/src\/hypergraph_embedding.py<\/a>    <\/p>\n\n<p>Run on colab:     <\/p>\n\n<ol>\n<li>TPU:<a href=\"https:\/\/colab.research.google.com\/gist\/WangHexie\/30c385509f9cd93be747f04c39f039a4\/tpu-error.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/gist\/WangHexie\/30c385509f9cd93be747f04c39f039a4\/tpu-error.ipynb<\/a>   <\/li>\n<li>GPU\uff1a<a href=\"https:\/\/colab.research.google.com\/gist\/WangHexie\/5bfac53bf92ef0ad527f15ddbf8705e1\/-gpu-ipynb.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/gist\/WangHexie\/5bfac53bf92ef0ad527f15ddbf8705e1\/-gpu-ipynb.ipynb<\/a><\/li>\n<\/ol>\n\n<p>The model:<\/p>\n\n<pre><code>def build_model(self):\n    self.inputs = [Input(shape=(self.options.dim_feature[i], ), name='input_{}'.format(i), dtype='float') for i in range(3)]\n\n    self.encodeds = [Dense(self.options.embedding_size[i], activation='tanh', name='encode_{}'.format(i))(self.inputs[i]) for i in range(3)]\n    self.decodeds = [Dense(self.options.dim_feature[i], activation='sigmoid', name='decode_{}'.format(i),\n                    activity_regularizer = regularizers.l2(0.0))(self.encodeds[i]) for i in range(3)]\n\n    self.merged = concatenate(self.encodeds, axis=1)\n    self.hidden_layer = Dense(self.options.hidden_size, activation='tanh', name='full_connected_layer')(self.merged)\n    self.ouput_layer = Dense(1, activation='sigmoid', name='classify_layer')(self.hidden_layer)\n\n    self.model = Model(inputs=self.inputs, outputs=self.decodeds+[self.ouput_layer])\n\n    self.model.compile(optimizer=tf.train.AdamOptimizer(learning_rate=self.options.learning_rate),\n            loss=[self.sparse_autoencoder_error]*3+['binary_crossentropy'],\n                          loss_weights=[self.options.alpha]*3+[1.0],\n                          metrics=dict([('decode_{}'.format(i), 'mse') for i in range(3)]+[('classify_layer', 'accuracy')]))\n    self.model = tf.contrib.tpu.keras_to_tpu_model(\n        self.model,\n        strategy=tf.contrib.tpu.TPUDistributionStrategy(\n            tf.contrib.cluster_resolver.TPUClusterResolver(\n                tpu='grpc:\/\/' + os.environ['COLAB_TPU_ADDR'])\n        )\n    )\n    self.model.summary()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57463879","CreationDate":"2019-08-12T15:14:54.433","ParentId":null,"OwnerUserId":"11860456","Title":null,"Body":"<p>As of 2019-02-20, the function <code>tf.contrib.tpu.keras_to_tpu_model<\/code> has been deprecated. You should therefore re-attempt converting your model using the new <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/distribute\/Strategy\" rel=\"nofollow noreferrer\">Distribution Strategy<\/a> function. An in depth guide on distributed training can be found <a href=\"https:\/\/www.tensorflow.org\/guide\/distribute_strategy\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I also noticed that you are using data type float as your input values. In CPython, the default bit value is 64bit. Currently, TPU\u2019s function most optimally with 16-bit floats therefore you should reduce your inputs to either 8-bit or 16-bit. The lower the bit value, the faster the processing will be for your model. <\/p>\n\n<p>Therefore it is also recommended to take advantage of Quantization, converting float weights to 8-bit integers. There are two types of quantized training: <a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization\" rel=\"nofollow noreferrer\">post-training quantization<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/quantize\/README.md\" rel=\"nofollow noreferrer\">quantization-aware training<\/a>. <\/p>\n\n<p>For more information concerning TPU\u2019s on Google Cloud Platform you may refer to the <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/\" rel=\"nofollow noreferrer\">Cloud TPU documentation<\/a>, and for more information on TPU system architecture you may refer to <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/system-architecture\" rel=\"nofollow noreferrer\">this<\/a> documentation by Google as it properly explains how TPU\u2019s are designed.<\/p>\n"}]}
+{"QuestionId":53312025,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><finetunning>","CreationDate":"2018-11-15T03:30:51.977","AcceptedAnswerId":53312991.0,"OwnerUserId":1276374.0,"Title":"Keras finetunning InceptionV3 tensor dimension error","Body":"<p>I am trying to finetune a model in Keras:<\/p>\n\n<pre><code>    inception_model = InceptionV3(weights=None, include_top=False, input_shape=(150, \n150, 1))\n\n    x = inception_model.output\n    x = GlobalAveragePooling2D()(x)\n    x = Dense(256, activation='relu', name='fc1')(x)\n    x = Dropout(0.5)(x)\n    predictions = Dense(10, activation='softmax', name='predictions')(x)\n    classifier = Model(inception_model.input, predictions)\n\n\n    ####training training training ... save weights\n\n\n    classifier.load_weights(\"saved_weights.h5\")\n\n    classifier.layers.pop()\n    classifier.layers.pop()\n    classifier.layers.pop()\n    classifier.layers.pop()\n    ###enough poping to reach standard InceptionV3 \n\n    x = classifier.output\n    x = GlobalAveragePooling2D()(x)\n    x = Dense(256, activation='relu', name='fc1')(x)\n    x = Dropout(0.5)(x)\n    predictions = Dense(10, activation='softmax', name='predictions')(x)\n    classifier = Model(classifier.input, predictions)\n<\/code><\/pre>\n\n<p>But I get the error:<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer global_average_pooling2d_3: expected ndim=4, found ndim=2\n<\/code><\/pre>\n","answers":[{"AnswerId":"53312991","CreationDate":"2018-11-15T05:29:48.157","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2371#issuecomment-211120172\" rel=\"nofollow noreferrer\">shouldn't use<\/a> <code>pop()<\/code> method on models created using functional API (i.e. <code>keras.models.Model<\/code>). Only Sequential models (i.e. <code>keras.models.Sequential<\/code>) have a built-in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/7205d903fbc079bb99fbae0e3c02e6d2b4d227f0\/keras\/engine\/sequential.py#L192\" rel=\"nofollow noreferrer\"><code>pop()<\/code><\/a> method (usage: <code>model.pop()<\/code>). Instead, use index or the names of the layers to access a specific layer:<\/p>\n\n<pre><code>classifier.load_weights(\"saved_weights.h5\")\nx = classifier.layers[-5].output   # use index of the layer directly\nx = GlobalAveragePooling2D()(x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53312308,"AnswerCount":0,"Tags":"<python><tensorflow><shutil>","CreationDate":"2018-11-15T04:09:36.130","AcceptedAnswerId":null,"OwnerUserId":9226290.0,"Title":"Delete tensorflow checkpoint and event files after training","Body":"<p>After training and evaluating a custom tensorflow estimator, I would like to delete all the files that it saves - the checkpoint files, the event files, the eval folders, etc. - from within the same script. I am trying to do this by using shutil.rmtree on the directory that holds all the files; however, it keeps throwing this error:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\weakref.py\", line 548, in <strong>call<\/strong>\n    return info.func(*info.args, **(info.kwargs or {}))\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\tempfile.py\", line 797, in _cleanup\n    _shutil.rmtree(name)\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 494, in rmtree\n    return _rmtree_unsafe(path, onerror)\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 384, in _rmtree_unsafe\n    _rmtree_unsafe(fullname, onerror)\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 384, in _rmtree_unsafe\n    _rmtree_unsafe(fullname, onerror)\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 384, in _rmtree_unsafe\n    _rmtree_unsafe(fullname, onerror)\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 393, in _rmtree_unsafe\n    onerror(os.rmdir, path, sys.exc_info())\n  File \"C:\\Users____\\AppData\\Local\\Continuum\\anaconda3\\lib\\shutil.py\", line 391, in _rmtree_unsafe\n    os.rmdir(path)<\/p>\n\n<p>OSError: [WinError 145] The directory is not empty: 'C:\\Users\\____\\AppData\\Local\\Temp\\4\\tmpfuioqj79\\search_1\\fold_1\\eval'<\/p>\n\n<p>My understanding is that shutil.rmtree will remove the directory and all of its contents, and, when I use shutil.rmtree from a separate script after the script that creates and trains the models is finished, it works fine. Therefore, my I'm guessing that tensorflow is still using the files\/directory. I cannot find anything in the estimator or saver documentation that would suggest I need to manually close something after training with the estimator implementation.<\/p>\n\n<p>Anyone have any idea how to delete these files from the same script that performs the training and evaluation?<\/p>\n\n<p>Edit: After looking through all the files that are being used by python processes, it looks like it's only the event files, not the checkpoint files, that are still in use.<\/p>\n","answers":[]}
+{"QuestionId":53312325,"AnswerCount":0,"Tags":"<android><tensorflow><keras>","CreationDate":"2018-11-15T04:11:58.337","AcceptedAnswerId":null,"OwnerUserId":3138464.0,"Title":"Keras model tensorflow lite conversion input shape","Body":"<p>I'm trying to convert my simple Keras model frozen graph to tensorflowlite but I'm not sure what the input shape is. <\/p>\n\n<pre><code>    toco \n--input_file='my_model.pb'\n--input_format=TENSORFLOW_GRAPHDEF \\\n--output_format=TFLITE \\\n--output_file=\/tmp\/my_model.tflite \\\n--inference_type=FLOAT \\\n--input_type=FLOAT \\\n--input_arrays=input_tensor \\\n--output_arrays=output_tensor \\\n--input_shapes=0,4,2851\n<\/code><\/pre>\n\n<p>My model is:<\/p>\n\n<pre><code># create model\nmodel = Sequential()\nmodel.add(Dense(50, activation=\"tanh\", input_dim=4, kernel_initializer=\"random_uniform\", name=\"input_tensor\"))\nmodel.add(Dense(50, activation=\"tanh\", kernel_initializer=\"random_uniform\"))\nmodel.add(Dense(1, activation=\"linear\", kernel_initializer='random_uniform', name=\"output_tensor\"))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53312339,"AnswerCount":2,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2018-11-15T04:14:06.213","AcceptedAnswerId":53312427.0,"OwnerUserId":8715275.0,"Title":"Installing COCOAPI when 2 versions of python are present","Body":"<p>I need to install COCOAPI for Python 3.5 on my linux machine but when I do \"make\" it automatically installs it for 2.7. Is there an option to choose for a python version while using \"make\" ? <\/p>\n\n<p><strong>EDIT 1 :<\/strong> \nGoing to PythonAPI folder and installing it via python3 setup.py install gives the following error. <\/p>\n\n<pre><code>sudo python3 setup.py install\n\nrunning install\nrunning bdist_egg\nrunning egg_info\ncreating pycocotools.egg-info\nwriting requirements to pycocotools.egg-info\/requires.txt\nwriting dependency_links to pycocotools.egg-info\/dependency_links.txt\nwriting top-level names to pycocotools.egg-info\/top_level.txt\nwriting pycocotools.egg-info\/PKG-INFO\nwriting manifest file 'pycocotools.egg-info\/SOURCES.txt'\nreading manifest file 'pycocotools.egg-info\/SOURCES.txt'\nwriting manifest file 'pycocotools.egg-info\/SOURCES.txt'\ninstalling library code to build\/bdist.linux-x86_64\/egg\nrunning install_lib\nrunning build_py\ncreating build\ncreating build\/lib.linux-x86_64-3.5\ncreating build\/lib.linux-x86_64-3.5\/pycocotools\ncopying pycocotools\/coco.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\ncopying pycocotools\/__init__.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\ncopying pycocotools\/mask.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\ncopying pycocotools\/cocoeval.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\nrunning build_ext\ncythoning pycocotools\/_mask.pyx to pycocotools\/_mask.c\n\/home\/pradyumn\/.local\/lib\/python3.5\/site-packages\/Cython\/Compiler\/Main.py:367: FutureWarning: Cython directive 'language_level' not set, using 2 for now (Py2). This will change in a later release! File: \/home\/pradyumn\/cocoapi-master\/PythonAPI\/pycocotools\/_mask.pyx\n  tree = Parsing.p_module(s, pxd, full_module_name)\nbuilding 'pycocotools._mask' extension\ncreating build\/common\ncreating build\/temp.linux-x86_64-3.5\ncreating build\/temp.linux-x86_64-3.5\/pycocotools\nx86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I\/home\/pradyumn\/.local\/lib\/python3.5\/site-packages\/numpy\/core\/include -I..\/common -I\/usr\/include\/python3.5m -c ..\/common\/maskApi.c -o build\/temp.linux-x86_64-3.5\/..\/common\/maskApi.o -Wno-cpp -Wno-unused-function -std=c99\n..\/common\/maskApi.c: In function \u2018rleToBbox\u2019:\n..\/common\/maskApi.c:141:31: warning: \u2018xp\u2019 may be used uninitialized in this function [-Wmaybe-uninitialized]\n       if(j%2==0) xp=x; else if(xp&lt;x) { ys=0; ye=h-1; }\n                               ^\nx86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I\/home\/pradyumn\/.local\/lib\/python3.5\/site-packages\/numpy\/core\/include -I..\/common -I\/usr\/include\/python3.5m -c pycocotools\/_mask.c -o build\/temp.linux-x86_64-3.5\/pycocotools\/_mask.o -Wno-cpp -Wno-unused-function -std=c99\npycocotools\/_mask.c:4:20: fatal error: Python.h: No such file or directory\ncompilation terminated.\nerror: command 'x86_64-linux-gnu-gcc' failed with exit status 1\nenter code here\n<\/code><\/pre>\n\n<p><strong>Edit 2<\/strong> <\/p>\n\n<p>I tried installing cython and then pycocotools but that gives me the following error. <\/p>\n\n<pre><code>pip3 install pycocotools \nCollecting pycocotools\n  Using cached https:\/\/files.pythonhosted.org\/packages\/96\/84\/9a07b1095fd8555ba3f3d519517c8743c2554a245f9476e5e39869f948d2\/pycocotools-2.0.0.tar.gz\nBuilding wheels for collected packages: pycocotools\n  Running setup.py bdist_wheel for pycocotools ... error\n  Complete output from command \/usr\/bin\/python3 -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-install-p1jp5iir\/pycocotools\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" bdist_wheel -d \/tmp\/pip-wheel-xirt_d3c --python-tag cp35:\n  running bdist_wheel\n  running build\n  running build_py\n  creating build\n  creating build\/lib.linux-x86_64-3.5\n  creating build\/lib.linux-x86_64-3.5\/pycocotools\n  copying pycocotools\/__init__.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n  copying pycocotools\/coco.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n  copying pycocotools\/mask.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n  copying pycocotools\/cocoeval.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n  running build_ext\n  building 'pycocotools._mask' extension\n  creating build\/temp.linux-x86_64-3.5\n  creating build\/temp.linux-x86_64-3.5\/pycocotools\n  creating build\/temp.linux-x86_64-3.5\/common\n  x86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I\/home\/pradyumn\/.local\/lib\/python3.5\/site-packages\/numpy\/core\/include -Icommon -I\/usr\/include\/python3.5m -c pycocotools\/_mask.c -o build\/temp.linux-x86_64-3.5\/pycocotools\/_mask.o -Wno-cpp -Wno-unused-function -std=c99\n  pycocotools\/_mask.c:32:20: fatal error: Python.h: No such file or directory\n  compilation terminated.\n  error: command 'x86_64-linux-gnu-gcc' failed with exit status 1\n\n  ----------------------------------------\n  Failed building wheel for pycocotools\n  Running setup.py clean for pycocotools\nFailed to build pycocotools\nInstalling collected packages: pycocotools\n  Running setup.py install for pycocotools ... error\n    Complete output from command \/usr\/bin\/python3 -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-install-p1jp5iir\/pycocotools\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-record-pipcgtu8\/install-record.txt --single-version-externally-managed --compile:\n    running install\n    running build\n    running build_py\n    creating build\n    creating build\/lib.linux-x86_64-3.5\n    creating build\/lib.linux-x86_64-3.5\/pycocotools\n    copying pycocotools\/__init__.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n    copying pycocotools\/coco.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n    copying pycocotools\/mask.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n    copying pycocotools\/cocoeval.py -&gt; build\/lib.linux-x86_64-3.5\/pycocotools\n    running build_ext\n    building 'pycocotools._mask' extension\n    creating build\/temp.linux-x86_64-3.5\n    creating build\/temp.linux-x86_64-3.5\/pycocotools\n    creating build\/temp.linux-x86_64-3.5\/common\n    x86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I\/home\/pradyumn\/.local\/lib\/python3.5\/site-packages\/numpy\/core\/include -Icommon -I\/usr\/include\/python3.5m -c pycocotools\/_mask.c -o build\/temp.linux-x86_64-3.5\/pycocotools\/_mask.o -Wno-cpp -Wno-unused-function -std=c99\n    pycocotools\/_mask.c:32:20: fatal error: Python.h: No such file or directory\n    compilation terminated.\n    error: command 'x86_64-linux-gnu-gcc' failed with exit status 1\n\n    ----------------------------------------\nCommand \"\/usr\/bin\/python3 -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-install-p1jp5iir\/pycocotools\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-record-pipcgtu8\/install-record.txt --single-version-externally-managed --compile\" failed with error code 1 in \/tmp\/pip-install-p1jp5iir\/pycocotools\/\n<\/code><\/pre>\n\n<p><strong>EDIT 3<\/strong><\/p>\n\n<p>To solve the error above run the below for appropriate version of python.<\/p>\n\n<pre><code>sudo apt-get install python3.5-dev\n<\/code><\/pre>\n\n<p>Then do the following: <\/p>\n\n<pre><code>pip3 install pycocotools --user\n<\/code><\/pre>\n\n<p>And then rerun the following: <\/p>\n\n<pre><code>sudo python3 setup.py install\n<\/code><\/pre>\n","answers":[{"AnswerId":"53312427","CreationDate":"2018-11-15T04:25:44.000","ParentId":null,"OwnerUserId":"8468264","Title":null,"Body":"<p>Clone the repo using <code>git clone https:\/\/github.com\/cocodataset\/cocoapi.git<\/code> then enter the dir where it is located and type <code>python3 setup.py install<\/code> which should install using python3.<\/p>\n\n<p>If the above doesn't work, try <code>pip3 install cython<\/code> followed by <code>pip3 install pycocotools<\/code> (you can add the <code>--user<\/code> flag to all of these if necessary)<\/p>\n"},{"AnswerId":"53312399","CreationDate":"2018-11-15T04:22:12.367","ParentId":null,"OwnerUserId":"399573","Title":null,"Body":"<p>You can make sure the default <code>python<\/code> resolves in the one you are interested in, for example by symbolically linking <code>python<\/code> to <code>python2<\/code> wherever it is defined (possibly <code>\/usr\/bin<\/code>)<\/p>\n"}]}
+{"QuestionId":53312642,"AnswerCount":2,"Tags":"<keras><jupyter-notebook>","CreationDate":"2018-11-15T04:54:52.760","AcceptedAnswerId":53329788.0,"OwnerUserId":2827296.0,"Title":"Does interrupting keras training in a Jupyter notebook save the training?","Body":"<p>So, I was working on a machine learning project using a Jupyter Notebook and Keras, and I started training. I came back a few hours later, only to realize that I had accidentally set the epochs to a really high number.<\/p>\n\n<p>I'm wondering, if I stop running the cell (send a KeyboardInterrupt), will the whole training be canceled? Or will the weights from the epoch I was currently on still be saved?<\/p>\n\n<p>I can still access the model from the next cell.<\/p>\n","answers":[{"AnswerId":"53318439","CreationDate":"2018-11-15T11:27:00.793","ParentId":null,"OwnerUserId":"5108062","Title":null,"Body":"<p>If you have not defined a <code>ModelCheckpoint<\/code> callback or some custom model saver callback then the answer is no.<\/p>\n\n<p>Next time you should include the <a href=\"https:\/\/keras.io\/callbacks\/#modelcheckpoint\" rel=\"nofollow noreferrer\"><code>ModelCheckpoint<\/code><\/a> callback, so at every epoch your model will be saved, and you can restore it<\/p>\n"},{"AnswerId":"53329788","CreationDate":"2018-11-16T00:34:25.663","ParentId":null,"OwnerUserId":"10298844","Title":null,"Body":"<p>The trained model will still be in memory, in the state it was in when the  KeyboardInterrupt happened. As long as the Python kernel isn't stopped or the model isn't reinstantiated, you can continue to use the trained model. To test this, evaluate the model's prediction accuracy.<\/p>\n\n<p>Note that, if you continue training the model, a KeyboardInterrupt restarts the epoch counter. That will effect any callbacks that rely on the epoch number.<\/p>\n"}]}
+{"QuestionId":53313091,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2018-11-15T05:40:07.327","AcceptedAnswerId":null,"OwnerUserId":4003077.0,"Title":"c++ - How to get visible devices for Tensorflow?","Body":"<p>I am trying to fetch visible devices in Tensorflow C API. \nThe following commands shows me <code>0<\/code> devices!<\/p>\n\n<pre><code>#include &lt;tensorflow\/core\/public\/session.h&gt;\nusing namespace tensorflow;\n...\ntensorflow::SessionOptions tf_options;\nstd::cout &lt;&lt; tf_options.config.device_count().size();\n<\/code><\/pre>\n\n<p>Does anyone knows how can I do it?<\/p>\n","answers":[]}
+{"QuestionId":53313301,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2018-11-15T05:59:04.460","AcceptedAnswerId":null,"OwnerUserId":4968861.0,"Title":"Pytorch: why the orders of weights in ConvTranspose2d and Conv2d are different?","Body":"<p>Pytorch code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>up = nn.ConvTranspose2d(3, 128, 2, stride=2)\nconv = nn.Conv2d(3, 128, 2)\n\ninputs = Variable(torch.rand(1, 3, 64, 64))\nprint('up conv output size:', up(inputs).size())\n\ninputs = Variable(torch.rand(1, 3, 64, 64))\nprint('conv output size:', conv(inputs).size())\n\nprint('up conv weight size:', up.weight.data.shape)\nprint('conv weight size:', conv.weight.data.shape)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>up conv output size: torch.Size([1, 128, 128, 128])\nconv output size: torch.Size([1, 128, 63, 63])\nup conv weight size: torch.Size([3, 128, 2, 2])\nconv weight size: torch.Size([128, 3, 2, 2])\n<\/code><\/pre>\n\n<p>Why the orders are different between <code>ConvTranspose2d (3,128)<\/code> and <code>Conv2d (128, 3)<\/code>? <\/p>\n\n<p>Is it supposed to behave like this?<\/p>\n","answers":[]}
+{"QuestionId":53313712,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-15T06:35:26.473","AcceptedAnswerId":null,"OwnerUserId":7058803.0,"Title":"Keras model predicting experience from hours","Body":"<p>I am very new to Keras, neural networks and machine learning having just started to learn yesterday. I decided to try predicting the experience over an hour (0 to 23) (for a game and my own generated data-set) that a user would earn. Currently running what I have the predictions seem to be very low and very poor. I have tried a relu activation, which produced predictions all to be zero and from a bit of research, LeakyReLU. <\/p>\n\n<p>This is the code I have for the prediction model so far:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LeakyReLU\nimport numpy\n\nnumpy.random.seed(7)\ndataset = numpy.loadtxt(\"experience.csv\", delimiter=\",\")\n\nX = dataset[: ,0]\nY = dataset[: ,1]\n\nmodel = Sequential()\nmodel.add(Dense(12, input_dim = 1, activation=LeakyReLU(0.3)))\nmodel.add(Dense(8,  activation=LeakyReLU(0.3)))\nmodel.add(Dense(1, activation=LeakyReLU(0.3)))\n\nmodel.compile(loss = 'mean_absolute_error', optimizer='adam', metrics = ['accuracy'])\n\nmodel.fit(X, Y, epochs=120, batch_size=10, verbose = 0)\n\npredictions = model.predict(X)\n\nrounded = [round(x[0]) for x in predictions]\n\nprint(rounded)\n<\/code><\/pre>\n\n<p>I have also tried playing around with the hidden levels of the network, but honestly have no idea how many there should be or a good way to justify an amount.<\/p>\n\n<p>If it helps here is the data-set I have been using:<\/p>\n\n<p><a href=\"https:\/\/raw.githubusercontent.com\/NightShadeII\/xpPredictor\/master\/experience.csv\" rel=\"nofollow noreferrer\">https:\/\/raw.githubusercontent.com\/NightShadeII\/xpPredictor\/master\/experience.csv<\/a><\/p>\n\n<p>Thankyou for any help<\/p>\n","answers":[{"AnswerId":"53321784","CreationDate":"2018-11-15T14:35:44.147","ParentId":null,"OwnerUserId":"8990787","Title":null,"Body":"<p>Looking at your data it does not seem like a classification problem. \nYou have two options:<\/p>\n\n<p>-> Look at the second column and bucket them depending on the ranges and make classes that can be predicted, for instance: 0, 1, 2 etc. Now it tries to train but does not have enough examples for millions of classes that it thinks you are trying to predict.<\/p>\n\n<p>-> If you want real valued output and not classes, try using linear regression.<\/p>\n"}]}
+{"QuestionId":53313713,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-15T06:35:29.277","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"How can I get a gradient at two points with Keras\/Tensorflow","Body":"<p>I'm trying to get the gradient of a function at two different values of the independent variable. I am doing this in the context of modifying the <code>get_updates<\/code> method of Keras' optimizer class.<\/p>\n\n<p>The relevant section of the code I've written is <\/p>\n\n<pre><code>def get_updates(self, loss, params):\n    grads = self.get_gradients(loss, params)\n    self.updates = [K.update_add(self.iterations, 1)]\n\n    # Copy parameters\n    params2 = []\n    for p in params:\n        params2.append(K.variable(K.get_value(p), name=p.name[:-2] + \"_cpy1\/\"))\n\n    self.weights = [self.iterations]\n    for p, p2, g in zip(params, params2, grads):\n        v = - self.lr * g  \n        new_p = p + 0.5*v # Intermediate\/Partial step\n        new_p2 = p2 + v   # Reference point for 2nd Gradient\n\n    grads2 = self.get_gradients(loss, params2)\n    ....\n<\/code><\/pre>\n\n<p>The error I get when I execute this code is <\/p>\n\n<pre><code>  File \"\/home\/me\/Projects\/RungeKutta\/rk_optimizers2.py\", line 86, in get_updates\n    grads2 = self.get_gradients(loss, params2)\n  File \"\/home\/me\/anaconda2\/lib\/python2.7\/site-packages\/keras\/optimizers.py\", line 91, in get_gradients\n    raise ValueError('An operation has `None` for gradient. '\nValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.\n<\/code><\/pre>\n\n<p>The problematic section of code from Keras's optimizers.py is:<\/p>\n\n<pre><code>def get_gradients(self, loss, params):\n    grads = K.gradients(loss, params)\n    if None in grads:\n        raise ValueError('An operation has `None` for gradient. '\n                         'Please make sure that all of your ops have a '\n                         'gradient defined (i.e. are differentiable). '\n                         'Common ops without gradient: '\n                         'K.argmax, K.round, K.eval.')\n<\/code><\/pre>\n\n<p>In my case, <code>K<\/code> corresponds to Tensorflow. So, the corresponding code for <code>gradients<\/code> is:<\/p>\n\n<pre><code>def gradients(loss, variables):\n    \"\"\"Returns the gradients of `loss` w.r.t. `variables`.\n\n    # Arguments\n        loss: Scalar tensor to minimize.\n        variables: List of variables.\n\n    # Returns\n        A gradients tensor.\n    \"\"\"\n    return tf.gradients(loss, variables, colocate_gradients_with_ops=True)\n<\/code><\/pre>\n\n<p>Why do I get the error I see? How can I get a gradient at the first projected updated parameter so that I can get a final update using the average of the gradient at the starting point and the gradient at an (initially) projected end point?<\/p>\n","answers":[]}
+{"QuestionId":53313913,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-15T06:53:57.707","AcceptedAnswerId":53315228.0,"OwnerUserId":6402099.0,"Title":"How to convert 2D matrix to 3D tensor without blending corresponding entries?","Body":"<p>I have data with the shape of <strong>(3000, 4)<\/strong>, the features are <strong>(product, store, week, quantity)<\/strong>. Quantity is the target. <\/p>\n\n<p>So I want to reconstruct this matrix to a tensor, without blending the corresponding quantities.<\/p>\n\n<p>For example, if there are 30 product, 20 stores and 5 weeks, the shape of the tensor should be (5, 20, 30), with the corresponding quantity. Because there won't be an entry like (store A, product X, week 3) twice in entire data, so every store x product x week pair should have one corresponding quantity.<\/p>\n\n<p>Any suggestions about how to achieve this, or there is any logical error? Thanks.<\/p>\n","answers":[{"AnswerId":"53315228","CreationDate":"2018-11-15T08:32:30.877","ParentId":null,"OwnerUserId":"4016285","Title":null,"Body":"<p>If there is no dummies, you just have to carefully sort our data. np.lexsort can do it.<\/p>\n\n<p>Suppose your data looks like <code>data<\/code>:<\/p>\n\n<pre><code>import numpy as np\ndims=a,b,c=30,20,5\ndata=np.array(list(product(*[np.arange(i) for i in dims+(1,)])))\ndata[:,-1]=np.random.randint(0,100,a*b*c)\nnp.random.shuffle(data)\n#array([[ 4, 15,  0, 56],\n#       [27, 16,  2,  3],\n#       [ 4,  8,  4, 26],\n#       ..., \n#       [20, 14,  3, 28],\n#       [14, 10,  4,  6],\n#       [19, 14,  3, 39]])\n<\/code><\/pre>\n\n<p>You can then sort if necessary and reshape like this:<\/p>\n\n<pre><code>sorteddata=data[np.lexsort(data[:,::-1].T)]\ntensor=sorteddata[:,-1].reshape(dims)\n<\/code><\/pre>\n\n<p>now <code>tensor[4,15,0]<\/code> is <code>56<\/code>.  ok !<\/p>\n"},{"AnswerId":"53314237","CreationDate":"2018-11-15T07:19:21.077","ParentId":null,"OwnerUserId":"10640534","Title":null,"Body":"<p>You can first go through each of your first three columns and count the number of different products, stores and weeks that you have. This will give you the shape of your new array, which you can create using numpy. Importantly now, you need to create a conversion matrix for each category. For example, if product is 'XXX', then you want to know to which row of the first dimension (as product is the first dimension of your array) 'XXX' corresponds; same idea for store and week. Once you have all of this, you can simply iterate through all lines of your existing array and assign the value of quantity to the correct location inside your new array based on the indices stored in your conversion matrices for each value of product, store and week. As you said, it makes sense because there is a one-to-one correspondence.<\/p>\n"}]}
+{"QuestionId":53314152,"AnswerCount":1,"Tags":"<python><keras><neural-network><deep-learning><keras-layer>","CreationDate":"2018-11-15T07:12:13.607","AcceptedAnswerId":53314495.0,"OwnerUserId":1391660.0,"Title":"Copying weights of a specific layer - keras","Body":"<p>According to <a href=\"https:\/\/stackoverflow.com\/questions\/48547688\/tensorflow-keras-copy-weights-from-one-model-to-another\">this<\/a> the following copies weights from one model to another: <\/p>\n\n<pre><code>target_model.set_weights(model.get_weights())\n<\/code><\/pre>\n\n<p>What about copying the weights of a specific layer, would this work?<\/p>\n\n<pre><code>model_1.layers[0].set_weights(source_model.layers[0].get_weights())\nmodel_2.layers[0].set_weights(source_model.layers[0].get_weights())\n<\/code><\/pre>\n\n<p>If I train <code>model_1<\/code> and <code>model_2<\/code> will they have separate weights? The <a href=\"https:\/\/keras.io\/layers\/about-keras-layers\/\" rel=\"nofollow noreferrer\">documentation<\/a> doesn't state whether if this <code>get_weights<\/code> makes a deep copy or not. If this doesn't work, how can this be achieved?<\/p>\n","answers":[{"AnswerId":"53314495","CreationDate":"2018-11-15T07:38:31.647","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Of course, it would be a copy of the weights. It does not make sense the weights object to be shared between two separate models. You can check it for yourself with a simple example like this:<\/p>\n\n<pre><code>model1 = Sequential()\nmodel1.add(Dense(10, input_dim=2))\n\nmodel2 = Sequential()\nmodel2.add(Dense(10, input_dim=2))\n\nmodel1.compile(loss='mse', optimizer='adam')\nmodel2.compile(loss='mse', optimizer='adam')\n<\/code><\/pre>\n\n<p>Test:<\/p>\n\n<pre><code>&gt;&gt;&gt; model1.layers[0].get_weights()\n[array([[-0.42853734,  0.18648076, -0.47137827,  0.1792168 ,  0.0373047 ,\n          0.2765705 ,  0.38383502,  0.09664273, -0.4971757 ,  0.41548246],\n        [ 0.0403192 , -0.01309097,  0.6656211 , -0.0536288 ,  0.58677703,\n          0.21625364,  0.26447064, -0.42619988,  0.17218047, -0.39748642]],\n       dtype=float32),\n array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)]\n\n&gt;&gt;&gt; model2.layers[0].get_weights()\n[array([[-0.30062824, -0.3740575 , -0.3502644 ,  0.28050178, -0.68631136,\n          0.1596322 ,  0.08288956, -0.20988202,  0.34323698,  0.2893324 ],\n        [-0.29182747, -0.2754455 , -0.64082885,  0.29160154,  0.04342002,\n         -0.4996035 ,  0.6608283 ,  0.10293472,  0.11375248, -0.43438092]],\n       dtype=float32),\n array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)]\n\n&gt;&gt;&gt; model2.layers[0].set_weights(model1.layers[0].get_weights())\n&gt;&gt;&gt; model2.layers[0].get_weights()\n[array([[-0.42853734,  0.18648076, -0.47137827,  0.1792168 ,  0.0373047 ,\n          0.2765705 ,  0.38383502,  0.09664273, -0.4971757 ,  0.41548246],\n        [ 0.0403192 , -0.01309097,  0.6656211 , -0.0536288 ,  0.58677703,\n          0.21625364,  0.26447064, -0.42619988,  0.17218047, -0.39748642]],\n       dtype=float32),\n array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)]\n\n&gt;&gt;&gt; id(model1.layers[0].get_weights()[0])\n140494823634144\n\n&gt;&gt;&gt; id(model2.layers[0].get_weights()[0])\n140494823635664\n<\/code><\/pre>\n\n<p>The ids of kernel weights arrays are different so they are different objects, but with the same value.<\/p>\n"}]}
+{"QuestionId":53314895,"AnswerCount":1,"Tags":"<python><amazon-ec2><anaconda><caffe2><ubuntu-17.10>","CreationDate":"2018-11-15T08:07:50.560","AcceptedAnswerId":null,"OwnerUserId":10656139.0,"Title":"caffe2 no gpu support error: libprotobuf.so.14 [In Progress]","Body":"<p>Currently running AWS Deep Learning AMI (Ubuntu) Version 17.0 and have run into a bump with Caffe2. Checking to see if caffe2 was correctly installed by using <\/p>\n\n<pre><code>python -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'\n<\/code><\/pre>\n\n<p>I receive the following error. <\/p>\n\n<p><strong>Current Error:<\/strong><\/p>\n\n<pre><code>ubuntu@ip:~$ python -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'\nWARNING:root:This caffe2 python run does not have GPU support. Will run in CPU only mode.\nWARNING:root:Debug message: libprotobuf.so.14: cannot open shared object file: No such file or directory\nCRITICAL:root:Cannot load caffe2.python. Error: libprotobuf.so.14: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>What am i missing? Not sure if this is helpful but it seems to exist. <\/p>\n\n<p><strong>File Found?<\/strong><\/p>\n\n<pre><code>    ubuntu@ip:~$ find . -name \"libprotobuf.so.14*\"\n.\/anaconda3\/pkgs\/protobuf-3.4.0-py27_1\/lib\/libprotobuf.so.14.0.0\n.\/anaconda3\/pkgs\/protobuf-3.4.0-py27_1\/lib\/libprotobuf.so.14\n.\/anaconda3\/pkgs\/protobuf-3.4.0-py35_1\/lib\/libprotobuf.so.14.0.0\n.\/anaconda3\/pkgs\/protobuf-3.4.0-py35_1\/lib\/libprotobuf.so.14\n.\/anaconda3\/envs\/caffe_p35\/lib\/libprotobuf.so.14.0.0\n.\/anaconda3\/envs\/caffe_p35\/lib\/libprotobuf.so.14\n.\/anaconda3\/envs\/caffe2_p27\/lib\/libprotobuf.so.14.0.0\n.\/anaconda3\/envs\/caffe2_p27\/lib\/libprotobuf.so.14\n.\/anaconda3\/envs\/caffe_p27\/lib\/libprotobuf.so.14.0.0\n.\/anaconda3\/envs\/caffe_p27\/lib\/libprotobuf.so.14\n<\/code><\/pre>\n\n<p><strong>AWS Instance:<\/strong>\nWith latest deep learning frameworks pre-installed: MXNet, TensorFlow, PyTorch, Keras, Chainer, Caffe\/2, Theano &amp; CNTK, configured with NVIDIA CUDA, cuDNN, NCCL &amp; Intel MKL-DNN.<\/p>\n","answers":[{"AnswerId":"55588977","CreationDate":"2019-04-09T09:04:55.473","ParentId":null,"OwnerUserId":"9427281","Title":null,"Body":"<p>Based on your search<\/p>\n\n<p>It is found that <strong><em>\"libprotobuf.so.14\"<\/em><\/strong><\/p>\n\n<p>is available at .\/anaconda3\/envs\/caffe_p35\/lib\/libprotobuf.so.14 [one of the path]<\/p>\n\n<p>Try adding this lib path to LD_LIBRARY_PATH environment variable and Run the code again.<\/p>\n\n<p>Below is the command to set the path from where you have executed the 'find' command<\/p>\n\n<hr>\n\n<blockquote>\n  <p>export\n  LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:.\/anaconda3\/envs\/caffe_p35\/lib\"<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Hope this will solve your issue.<\/p>\n"}]}
+{"QuestionId":53315282,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><deep-learning><google-colaboratory>","CreationDate":"2018-11-15T08:35:53.663","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"Cannot read .7z file in Google Colab","Body":"<p>I am using Google Colab to create a deep learning model, and I face an issue when I run this code at the <strong>first time<\/strong>.<\/p>\n\n<pre><code>!p7zip -d filename.7z\n<\/code><\/pre>\n\n<p>I get the following message:<\/p>\n\n<pre><code>\/usr\/bin\/p7zip: cannot read filename.7z\n<\/code><\/pre>\n\n<p><strong>But<\/strong> when I re-run the same cell again, the code works.<\/p>\n\n<p>Do you know what is the reason of this issue? <\/p>\n","answers":[{"AnswerId":"55283302","CreationDate":"2019-03-21T14:55:42.853","ParentId":null,"OwnerUserId":"10234531","Title":null,"Body":"<p>First, you have to specify path before the file name \nin my case<\/p>\n\n<pre><code>!mkdir ~\/data\n!cd ~\/data\n!mkdir planet\n!cd planet\n\n# -c: competition name\n# -f: which file you want to download\n# -p: path to where the file should be saved\n!kaggle competitions download -c planet-understanding-the-amazon-from-space -f train-jpg.tar.7z -p ~\/data\/planet\/\n\n# Unzip the 7zip files\n# -d: which file to un7zip\n!p7zip -d ~\/data\/planet\/test-jpg.tar.7z\n<\/code><\/pre>\n"}]}
+{"QuestionId":53315723,"AnswerCount":0,"Tags":"<python><tensorflow><keras><python-import>","CreationDate":"2018-11-15T09:01:45.363","AcceptedAnswerId":null,"OwnerUserId":4841832.0,"Title":"Can import elements from tensorflow.python.keras, but not the module as a whole","Body":"<p>How does the following python session make sense?<\/p>\n\n<pre><code>(ctlearn) jsevillamol@jsevillamol-N551JK:~\/Documentos\/ctlearn_keras$ python\nPython 3.6.5 | packaged by conda-forge | (default, Apr  6 2018, 13:39:56) \n[GCC 4.8.2 20140120 (Red Hat 4.8.2-15)] on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; from tensorflow.python.keras.models import Model\n&gt;&gt;&gt; import tensorflow.python.keras as K\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nAttributeError: module 'tensorflow' has no attribute 'python'\n<\/code><\/pre>\n\n<p>I can import <code>tensorflow.python.keras.models<\/code>, but not import <code>tensorflow.python.keras<\/code>. <\/p>\n\n<p>What. I don't even.<\/p>\n","answers":[]}
+{"QuestionId":53316174,"AnswerCount":1,"Tags":"<neural-network><deep-learning><nlp><pytorch><word-embedding>","CreationDate":"2018-11-15T09:26:29.863","AcceptedAnswerId":53329859.0,"OwnerUserId":7483494.0,"Title":"Using pre-trained word embeddings - how to create vector for unknown \/ OOV Token?","Body":"<p>I wan't to add <a href=\"https:\/\/nlp.stanford.edu\/projects\/glove\/\" rel=\"nofollow noreferrer\">pre-trained embeddings<\/a> to a model. But as it seems there is no <em>out-of-vocabulary (OOV)<\/em> token resp. no vector for unseen words existent. <\/p>\n\n<p>So what can I do to handle <em>OOV-tokens<\/em> I come across? I have some ideas, but none of them seem to be very good:<\/p>\n\n<ul>\n<li><p>I could just create a random vector for this token, but ideally I'd like the vector to within the <em>logic<\/em> of the existing model. If I just create it randomly I'm afraid the vector accidentally could be very similar to a very frequent word like <em>'the', 'for', 'that'<\/em> etc. which is not my intention.<\/p><\/li>\n<li><p>Or should I just initialize the vector with plain zeros instead?<\/p><\/li>\n<li><p>Another idea would be averaging the token over other existing vectors. But averaging on what vectors then? On all? This doesn't seem to be very conclusive either.<\/p><\/li>\n<li><p>I also thought about trying to train this vector. However this doesn't come very handy if I want to freeze the rest of the embedding during training. <\/p><\/li>\n<\/ul>\n\n<p><em>(A general solution is appreciated, but I wanted to add that I'm using PyTorch - just in case PyTorch already comes with a handy solution to this problem.)<\/em><\/p>\n\n<p><strong>So what would be a good and <em>easy<\/em> strategy to create such a vector?<\/strong><\/p>\n","answers":[{"AnswerId":"53329859","CreationDate":"2018-11-16T00:44:48.080","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>There are multiple ways you can deal with. I don't think I can cite reference about which works better.<\/p>\n\n<p><strong>Non-trainable option<\/strong>:<\/p>\n\n<ol>\n<li>Random vector as embedding<\/li>\n<li>You can use all zero vector for OOV.<\/li>\n<li>You can mean of all the embedding vector, that way you avoid the risk of being away from actual distribution.<\/li>\n<li>Also embeddings generally come with \"unk\" vectors learned during the training you can use that. <\/li>\n<\/ol>\n\n<p><strong>Trainable Option<\/strong>:<\/p>\n\n<p>You can declare a separate embedding vector for OOV and make it trainable keeping other embedding fixed. You might have to over-write the forward method of embedding lookup for this. You can declare a new trainable <code>Variable<\/code> and in the forward pass use this vector as embedding for OOV instead of doing a look-up.<\/p>\n\n<hr>\n\n<p>Addressing the comments of OP:<\/p>\n\n<p>I am not sure which of the three non-trainable methods may work better and I am not sure if there is some work about this. But method 4) should be working better. <\/p>\n\n<p>For trainable option, you can create a new embedding layer as below. <\/p>\n\n<pre><code>class Embeddings_new(torch.nn.Module): \n    def __init__(self, dim, vocab): \n        super().__init__() \n        self.embedding = torch.nn.Embedding(vocab, dim) \n        self.embedding.weight.requires_grad = False\n        # vector for oov \n        self.oov = torch.nn.Parameter(data=torch.rand(1,dim)) \n        self.oov_index = -1 \n        self.dim = dim \n\n    def forward(self, arr): \n        N = arr.shape[0] \n        mask =  (arr==self.oov_index).long() \n        mask_ = mask.unsqueeze(dim=1).float() \n        embed =(1-mask_)*self.embedding((1-mask)*arr) + mask_*(self.oov.expand((N,self.dim))) \n        return embed \n<\/code><\/pre>\n\n<p>Usage: <\/p>\n\n<pre><code>model = Embeddings_new(10,20000)    \nout = model.forward(torch.tensor([-1,-1, 100, 1, 0]))\n# dummy loss\nloss = torch.sum(a**2)\nloss.backward()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53316561,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><keras-layer>","CreationDate":"2018-11-15T09:47:13.807","AcceptedAnswerId":null,"OwnerUserId":6763456.0,"Title":"Calculating the pearson coefficients in a keras model met with AttributeError: 'NoneType' object has no attribute '_inbound_nodes'","Body":"<p>I'm using pearson correlation coefficients as the input features of the MLP model, and meeting with problem when establishing the model.<\/p>\n\n<p>I've updated and run my code according to <strong>today<\/strong>'s favorable comment that wrapping the function inside a Lambda layer, whereas the same error still occurs. I thought I've fixed the errors <strong>today<\/strong> mentioned. Below is my updated code:<\/p>\n\n<pre><code>from keras.layers import Conv2D, Input, Lambda, MaxPooling2D, BatchNormalization, Add, Flatten, Dense, Concatenate\nfrom keras.models import Model, Sequential\nimport keras.backend as K\n\ndef cal_cor(x):\n    a, b = x\n    assert a.shape[1] == b.shape[1], 'lengths not compatible'\n\n    a_mean = K.mean(a)\n    b_mean = K.mean(b)\n    a_norm = a - a_mean\n    b_norm = b - b_mean\n    numerator = a_norm * b_norm\n\n    a_var = K.square(a_norm)\n    b_var = K.square(b_norm)\n    denominator = (a_var * b_var) ** 0.5\n\n    return numerator \/ denominator\n\ndef cnn_with_pearson():\n\n    # Pearson part\n    cor_layer = Lambda(function=cal_cor, name='correlation_layer')\n\n    cors = []\n    for i in range(28):\n        cor_added = []\n        for j in range(28):\n            if i == j:\n                continue\n            cor_added.append( cor_layer( [input[:, i, :, 0], input[:, j, :, 0]] ) )\n        cors.append(Add()(cor_added))\n\n    con = Concatenate()(cors)\n    f = Flatten()(K.expand_dims(con, -1))\n    d = Dense(units=10, activation='softmax')(f)\n    model = Model(inputs=input, outputs=d)\n\n    return model\n\nmodel = cnn_with_pearson()\n<\/code><\/pre>\n\n<p>Error occurs in the line <code>model = Model(inputs=input, outputs=d)<\/code>:<\/p>\n\n<pre><code>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53317569","CreationDate":"2018-11-15T10:41:02.877","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You must wrap your function inside a <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\"><code>Lambda<\/code><\/a> layer to let Keras be able to track it and augment its returned tensor(s) with necessary information. To do this, you need to modify the definition of your function as well, since it must take a list of tensors as input:<\/p>\n\n<pre><code>def cal_cor(x):\n    a, b = x\n\n    # the rest is the same ...\n<\/code><\/pre>\n\n<p>Then define the Lambda layer for that:<\/p>\n\n<pre><code>cor_layer = Lambda(cal_cor)\n<\/code><\/pre>\n\n<p>and then use it like this:<\/p>\n\n<pre><code>cor_added = []\nfor j in range(28):\n    if i == j:\n        continue\n    cor_added.append(cor_layer(input[:, i, :, 0], input[:, j, :, 0]))\n\ncors.append(Add()(cor_added))\n<\/code><\/pre>\n\n<p>Note that for addition, we also need to use the <a href=\"https:\/\/keras.io\/layers\/merge\/#add\" rel=\"nofollow noreferrer\"><code>Add<\/code><\/a> layer (<code>from keras.layers import Add<\/code>).<\/p>\n"}]}
+{"QuestionId":53317187,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><keras-layer>","CreationDate":"2018-11-15T10:19:23.050","AcceptedAnswerId":53317295.0,"OwnerUserId":3645038.0,"Title":"Batch-size dimension not honored when composing models","Body":"<p>I have a Keras model that I defined during training as:<\/p>\n\n<pre><code>   img = keras.Input(shape=[65, 65, 2])\n   bnorm = keras.layers.BatchNormalization()(img)\n   ...\n   model = keras.Model(img, outputprob)\n<\/code><\/pre>\n\n<p>During serving, though, my inputs come differently. Therefore, I defined an input layer (verifying that the <code>to_img<\/code> shape comes out to be also <code>(65, 65, 2)<\/code>) and tried to do model composition using:<\/p>\n\n<pre><code>  to_img = keras.layers.Lambda(...)(json_input)\n  model_output = model(to_img)\n  serving_model = keras.Model(json_input, model_output)\n<\/code><\/pre>\n\n<p>However, I get this error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError:\n Shape must be rank 4 but is rank 3 for \n'model\/batch_normalization\/cond\/FusedBatchNorm' (op:\n 'FusedBatchNorm') with input shapes: [65,65,2],\n [2], [2], [0], [0].\n<\/code><\/pre>\n\n<p>This seems to indicate the batch-dimension didn't make it through. Why?<\/p>\n\n<p><strong>EDIT:<\/strong>\nThings I have tried:<\/p>\n\n<p>(1) Explicitly set <code>trainable=False<\/code> in all the layers but that didn't seem to make any difference:<\/p>\n\n<pre><code>  model_core = model\n  for layer in model_core.layers:\n    layer.trainable = False\n  model_output = model_core(to_img)\n<\/code><\/pre>\n\n<p>(2) Tried expanding the result of the preproc:<\/p>\n\n<pre><code>   to_img = keras.layers.Lambda(\n      lambda x : preproc(x))(json_input)\n   to_img = keras.layers.Lambda( \n      lambda x : tf.expand_dims(x, axis=0) )(to_img)\n<\/code><\/pre>\n\n<p>This results in an error: <code>AttributeError: 'Model' object has no attribute '_name'<\/code> on the line <code>serving_model = keras.Model(json_input, model_output)<\/code><\/p>\n\n<p>(3) Changed the lambda layer to do a map_fn to process the data individually:<\/p>\n\n<pre><code>to_img = keras.layers.Lambda(\n    lambda items: K.map_fn(lambda x: preproc, items))(json_input)\n<\/code><\/pre>\n\n<p>This caused a shape error that indicates that the preproc function is getting [65,2] items instead of [65,65,2].  This suggests that the Lambda layer applies the function to examples one at a time.<\/p>\n\n<p>(4) Here's the full code for the model:<\/p>\n\n<pre><code>  img = keras.Input(shape=[height, width, 2])\n\n  # convolutional part of model\n  cnn = keras.layers.BatchNormalization()(img)\n  for layer in range(nlayers):\n    nfilters = nfil * (layer + 1)\n    cnn = keras.layers.Conv2D(nfilters, (ksize, ksize), padding='same')(cnn)\n    cnn = keras.layers.Activation('elu')(cnn)\n    cnn = keras.layers.BatchNormalization()(cnn)\n    cnn = keras.layers.MaxPooling2D(pool_size=(2, 2))(cnn)\n  cnn = keras.layers.Flatten()(cnn)\n  cnn = keras.layers.Dropout(dprob)(cnn)\n  cnn = keras.layers.Dense(10, activation='relu')(cnn)\n\n  # feature engineering part of model\n  engfeat = keras.layers.Lambda(\n    lambda x: engineered_features(x, height\/\/2))(img)\n\n  # concatenate the two parts\n  both = keras.layers.concatenate([cnn, engfeat])\n  ltgprob = keras.layers.Dense(1, activation='sigmoid')(both)\n\n  # create a model\n  model = keras.Model(img, ltgprob)\n  def rmse(y_true, y_pred):\n    import tensorflow.keras.backend as K\n    return K.sqrt(K.mean(K.square(y_pred - y_true), axis=-1))\n  optimizer = tf.keras.optimizers.Adam(lr=params['learning_rate'],\n                                       clipnorm=1.)\n  model.compile(optimizer=optimizer,\n                loss='binary_crossentropy',\n                metrics=['accuracy', 'mse', rmse])\n<\/code><\/pre>\n\n<p>and the code for the preprocessing function:<\/p>\n\n<pre><code>def reshape_into_image(features, params):\n  # stack the inputs to form a 2-channel input\n  # features['ref'] is [-1, height*width]\n  # stacked image is [-1, height*width, n_channels]\n  n_channels = 2\n  stacked = tf.concat([features['ref'], features['ltg']], axis=1)\n  height = width = PATCH_SIZE(params)\n  return tf.reshape(stacked, [height, width, n_channels])\n<\/code><\/pre>\n\n<p>and the serving layer:<\/p>\n\n<pre><code>  # 1. layer that extracts multiple inputs from JSON\n  height = width = PATCH_SIZE(hparams)\n  json_input = keras.layers.concatenate([\n    keras.layers.Input(name='ref', dtype=tf.float32, shape=(height * width,)),\n    keras.layers.Input(name='ltg', dtype=tf.float32, shape=(height * width,)),\n  ], axis=0)\n\n  # 2. convert json_input to image (what model wants)\n  to_img = keras.layers.Lambda(\n    lambda x: reshape_into_image(features={\n      'ref': tf.reshape(x[0], [height * width, 1]),\n      'ltg': tf.reshape(x[1], [height * width, 1])\n    }, params=hparams),\n    name='serving_reshape')(json_input)\n\n\n  # 3. now, use trained model to predict\n  model_output = model(to_img)\n\n  # 4. create serving model\n  serving_model = keras.Model(json_input, model_output)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53317295","CreationDate":"2018-11-15T10:25:57.600","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The input shape of your model, considering the samples axis, is <code>(?, 65, 65, 2)<\/code> where <code>?<\/code> can be one or more than one. So you need to modify the Lambda layer (actually the function wrapped inside it) such that its output to be <code>(?, 65, 65, 2)<\/code> as well. One way of doing it is to use <code>K.expand_dims(out, axis=0)<\/code> <strong>in the wrapped function<\/strong> so that the output would have a shape of <code>(1, 65, 65, 2)<\/code>.<\/p>\n\n<p>By the way, <code>K<\/code> refers to backend: <code>from keras import backend as K<\/code>.<\/p>\n\n<p>Further, note that you must define the function wrapped by the Lambda such that it preserves the batch axis; otherwise, it is very likely that you are doing something wrong in the definition of that function.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>The error <code>AttributeError: 'Model' object has no attribute '_name'<\/code> is raised because you are passing <code>json_input<\/code> as the input of the model. However, it is not an input layer. Rather, it is the output of <code>concatenation<\/code> layer. To resolve this, first define the input layers and then pass them to <code>concatenation<\/code> layer and <code>Model<\/code> class, like this:<\/p>\n\n<pre><code>inputs = [keras.layers.Input(name='ref', dtype=tf.float32, shape=(height * width,)), \n          keras.layers.Input(name='ltg', dtype=tf.float32, shape=(height * width,))]\n\njson_input = keras.layers.concatenate(inputs, axis=0)\n\n# ...\nserving_model = keras.Model(inputs, model_output)\n<\/code><\/pre>\n\n<p><strong>Update 2:<\/strong>\nI think you can write this much more simpler and without going into so much unnecessary trouble. You want to go from two tensors of shape <code>(?, h*w)<\/code> to a tensor of shape <code>(?, h, w, 2)<\/code>. You can use <a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\"><code>Reshape<\/code><\/a> layer, so that would be:<\/p>\n\n<pre><code>from keras.layers import Reshape\n\ninputs = [keras.layers.Input(name='ref', dtype=tf.float32, shape=(height * width,)), \n          keras.layers.Input(name='ltg', dtype=tf.float32, shape=(height * width,))]\n\nreshape_layer = Reshape((height, width, 1))\nr_in1 = reshape_layer(inputs[0])\nr_in2 = reshape_layer(inputs[1])\nimg = concatenate([r_in1, r_in2])\n\noutput = model(img)\n\nserving_model = keras.Model(inputs, output)\n<\/code><\/pre>\n\n<p>No need for any custom function or Lambda layer.<\/p>\n\n<p>And by the way, if you are interested to know, the trouble with batch axis removal is caused by this line:<\/p>\n\n<pre><code>return tf.reshape(stacked, [height, width, n_channels])\n<\/code><\/pre>\n\n<p>You are not considering the batch axis when reshaping.<\/p>\n"}]}
+{"QuestionId":53317523,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-15T10:37:52.407","AcceptedAnswerId":53321485.0,"OwnerUserId":7089239.0,"Title":"Part 3: Switching between multiple contexts - no error and a bad exit code","Body":"<p>I've been struggling with managing multiple Keras models with <code>tf.Graphs<\/code> and <code>tf.Sessions<\/code> for several weeks now. In short, I'd like to have multiple models open and switch between them as needed. This includes training new models, opening from file and making predictions.<\/p>\n\n<p>The bottom line is: (almost) everything works fine until the program crashes with exit code <code>0xC0000005<\/code>. No error messages are given. Let me explain.<\/p>\n\n<ul>\n<li>I can load a model and make predictions on it. Results are recieved, and after printing them, the program crashes.<\/li>\n<li>I can load multiple models and make predictions on them. Then the program crashes.<\/li>\n<li>I can create a new model, and make predictions on it. Finally, the program crashes.<\/li>\n<li>I cannot create two models, even the same model twice with different instances of the class below. The program crashes.<\/li>\n<\/ul>\n\n<p>You get the point. This is how I currently manage the graphs and sessions. I use a context manager to set the created graph and session as defaults and later switch to the previous state.<\/p>\n\n<pre><code>class NeuralNetwork:\n    def __init__(self):\n        self.graph = tf.Graph()\n        self.session = tf.Session(graph=self.graph)\n        self.model = None\n\n    def close(self):\n        self.session.close()\n        del self.graph\n        self.graph = None\n        gc.collect()\n\n    @contextmanager\n    def _context(self):\n        prev = k.get_session()\n        k.set_session(self.session)\n        with self.graph.as_default(), self.session.as_default():\n            yield\n        k.set_session(prev)\n\n    def predict(self, x):\n        with self._context():\n            return self.model.predict(x)\n\n    def fit(self, x_train, y_train, n=20, batch=256):\n        with self._context():\n            self.model.fit(x_train, y_train, epochs=n, batch_size=batch, verbose=0)\n\n    def create(self, shape):\n        with self._context():\n            self.model = Sequential()\n            self.model.add(Dense(shape[1], input_dim=shape[0], activation='relu'))\n            self.model.add(Dropout(drop))\n            self.model.add(Dense(shape[2], activation='sigmoid'))\n            self.model.compile(loss='binary_crossentropy', optimizer='rmsprop')\n\n    def load(self, path, sfx=''):\n        with open(path \/ ('architecture' + sfx + '.json'), 'r') as f:\n            js = f.read()\n\n        with self._context():\n            self.model = model_from_json(js)\n            self.model.load_weights(path \/ ('weights' + sfx + '.h5'))\n            self.model.compile(loss='binary_crossentropy', optimizer='rmsprop')\n\n    def save(self, path, sfx=''):\n        path.mkdir(exist_ok=True)\n        with self._context():\n            js = self.model.to_json()\n            with open(path \/ ('architecture' + sfx + '.json'), 'w') as f:\n                f.write(js)\n            self.model.save_weights(path \/ ('weights' + sfx + '.h5'))\n<\/code><\/pre>\n\n<p>And with the above class, here's how a network is used elsewhere:<\/p>\n\n<pre><code>def create(self):\n    x, y = [], []\n    shape = (15, 30, 1)\n\n    self.predictor = NeuralNetwork()\n    self.predictor.create(shape)\n    self.predictor.fit(x, y)\n    self.predictor.save(path=self.path)\n    self.predictor.close()\n\ndef load(self):\n    self.predictor.load(path=self.path)\n\ndef predict(x):\n    # Executed only on loaded networks, never on created networks\n    # due to program structure\n    return self.predictor.predict(x)\n<\/code><\/pre>\n\n<p>Here are my previous efforts at articulating the problem.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/53002518\/poor-exit-code-when-managing-multiple-sessions\">Part 1<\/a>, the one where I had no clue<\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/53206900\/sound-way-of-managing-multiple-sessions-and-graphs\">Part 2<\/a>, the one where I started to figure things out<\/li>\n<\/ul>\n\n<p>To the best of my abilities and with the help of some people, I've tried to come up with a way to manage these resources (context manager and \"closing\" the network after training). But I have not come across documentation or a tutorial describing the process of Tensorflow or Keras resource management in detail.<\/p>\n\n<hr>\n\n<p>My goals are two-fold.<\/p>\n\n<ul>\n<li>First and foremost, get rid of this error<\/li>\n<li>Hopefully learn the absolutely correct way of dealing with this scenario<\/li>\n<\/ul>\n\n<p>If you can help me achieve or even step a tiny amount towards the direction of either one, I'd greatly appreciate it! I have the experience, that my struggles are neither unique nor something that others haven't already thought of. So I must just be lacking the proper approach.<\/p>\n","answers":[{"AnswerId":"53321485","CreationDate":"2018-11-15T14:20:24.833","ParentId":null,"OwnerUserId":"7089239","Title":null,"Body":"<p>The issue was resolved by updating all packages to their latest versions. Sadly, I made the upgrading in one go, which means I'm not sure what actually was the cause. But I'm willing to bet on Tensorflow.<\/p>\n\n<p>Here are the package versions most likely involved in producing the error and their updated versions:<\/p>\n\n<ul>\n<li><code>tensorflow==1.8.0 -&gt; 1.12.0<\/code><\/li>\n<li><code>numpy==1.14.5 -&gt; 1.15.4<\/code><\/li>\n<li><code>scikit-learn==0.19.1 -&gt; 0.20.0<\/code><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53317618,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-15T10:42:57.567","AcceptedAnswerId":53318547.0,"OwnerUserId":9066532.0,"Title":"Pytorch 0.4: Should the requires_grad flag of the network input be set to True during training?","Body":"<p>The <code>requires_grad<\/code> flag of input is default <code>False<\/code> during my last training. I was wondering if I should set it to <code>True<\/code>.<\/p>\n","answers":[{"AnswerId":"53318547","CreationDate":"2018-11-15T11:33:06.200","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>So generally:<\/p>\n\n<ul>\n<li><p>For all <strong>tensors \/ weights that you want to be trained<\/strong> the <code>requires_grad<\/code> flag has to be <strong><code>True<\/code><\/strong>. <br>\nThis would be the case for your <strong>parameters<\/strong> resp. <strong>weights<\/strong> and <strong>biases<\/strong>. So there you want the flag to be <strong><code>True<\/code><\/strong>. But this is already default value for predefined modules like <code>nn.Linear<\/code>, <code>nn.Embedding<\/code>. <code>nn.Conv2d<\/code> etc. So you don't have to change it.<\/p><\/li>\n<li><p>For things like <strong>inputs<\/strong> the <code>requires_grad<\/code> flag needs to be <strong><code>False<\/code><\/strong> since you don't want to train your input data. <\/p><\/li>\n<\/ul>\n\n<p>I hope this answers your question, if you want to know something more specific just ask!<\/p>\n"}]}
+{"QuestionId":53318270,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><loss-function>","CreationDate":"2018-11-15T11:16:52.580","AcceptedAnswerId":null,"OwnerUserId":4729764.0,"Title":"Tensorflow calculates incorrect loss for `tf.keras` models when using Weights","Body":"<p>The loss calculation is not correct when working with <code>tf.keras<\/code>. After building the model, <code>tf.keras.fit_generator<\/code> should accept <code>(inputs, targets, sample_weights)<\/code> as inputs. However, if I multiply the <code>sample_weights<\/code> by 10000, the loss doesn't change.<\/p>\n\n<p>The bug seems to appear from 1.10 version of Tensorflow onwards e.g. (1.11, 1.12)<\/p>\n\n<h2>Code to reproduce<\/h2>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nWEIGHT_VARIABLE = 1\n\nno_of_features = 10\ntimesteps = 3\nbatch_size = 32\n\ndef data_gen():\n\n    while True:\n        numerical = np.random.randint(5, size=(batch_size, timesteps, no_of_features))\n        y = np.random.randint(2, size=batch_size)\n        w = np.ones(batch_size) * WEIGHT_VARIABLE\n\n        yield {'numeric_input': numerical}, y, w\n\n\ndef build_model():\n    numerical_input = tf.keras.layers.Input(shape=(timesteps, no_of_features), name='numeric_input')\n    rnn_out = tf.keras.layers.GRU(32, return_sequences=False)(numerical_input)\n    dense = tf.keras.layers.Dense(1, activation='sigmoid', name='main_output')(rnn_out)\n\n    model = tf.keras.models.Model(numerical_input, dense)\n\n    params = {\n        'loss': 'binary_crossentropy',\n        'optimizer': tf.keras.optimizers.Adam(),\n        'metrics': [tf.keras.metrics.binary_crossentropy, tf.keras.metrics.binary_accuracy]\n    }\n    model.compile(**params)\n\n    return model\n\n\ndef train_model():\n    gen1 = data_gen()\n    model = build_model()\n\n    model.fit_generator(gen1, epochs=30, steps_per_epoch=10)\n\n\nif __name__ == '__main__':\n    train_model()\n<\/code><\/pre>\n\n<p>In the above code, you simply need to change the <code>WEIGHT_VARIABLE = 1<\/code> From 1 to 100000 and rerun the file.<\/p>\n\n<hr>\n\n<h3>Logs<\/h3>\n\n<pre><code>v1.10\n\nWEIGHT_VARIABLE = 1 \n\nEpoch 1\/5 10\/10 [==============================] - \n1s 128ms\/step - loss: 0.7407 - binary_crossentropy: 0.7407 - binary_accuracy: 0.5031 \nEpoch 2\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7043 - binary_crossentropy: 0.7043 - binary_accuracy: 0.5125\nEpoch 3\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7055 - binary_crossentropy: 0.7055 - binary_accuracy: 0.5219\nEpoch 4\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7002 - binary_crossentropy: 0.7002 - binary_accuracy: 0.5250\nEpoch 5\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.6944 - binary_crossentropy: 0.6944 - binary_accuracy: 0.5375\n\nWEIGHT_VARIABLE = 10000\n\nEpoch 1\/5 10\/10 [==============================] - \n1s 131ms\/step - loss: 7235.5976 - binary_crossentropy: 0.7236 - binary_accuracy: 0.4562 \nEpoch 2\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 7271.9184 - binary_crossentropy: 0.7272 - binary_accuracy: 0.4844 \nEpoch 3\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 7276.9147 - binary_crossentropy: 0.7277 - binary_accuracy: 0.4500 \nEpoch 4\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 7052.0121 - binary_crossentropy: 0.7052 - binary_accuracy: 0.4625 \nEpoch 5\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 7187.0285 - binary_crossentropy: 0.7187 - binary_accuracy: 0.4969\n\nv1.12\n\nWEIGHT_VARIABLE = 1 \n\nEpoch 1\/5 10\/10 [==============================] - \n1s 68ms\/step - loss: 0.7188 - binary_crossentropy: 0.7188 - binary_accuracy: 0.5312 \nEpoch 2\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7044 - binary_crossentropy: 0.7044 - binary_accuracy: 0.4969 \nEpoch 3\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7086 - binary_crossentropy: 0.7086 - binary_accuracy: 0.4844 \nEpoch 4\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7075 - binary_crossentropy: 0.7075 - binary_accuracy: 0.4500 \nEpoch 5\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.6950 - binary_crossentropy: 0.6950 - binary_accuracy: 0.5187\n\nWEIGHT_VARIABLE = 10000\n\nEpoch 1\/5 10\/10 [==============================] - \n1s 74ms\/step - loss: 0.9084 - binary_crossentropy: 0.9084 - binary_accuracy: 0.4719\nEpoch 2\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7120 - binary_crossentropy: 0.7120 - binary_accuracy: 0.5062 \nEpoch 3\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7024 - binary_crossentropy: 0.7024 - binary_accuracy: 0.5344\nEpoch 4\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7257 - binary_crossentropy: 0.7257 - binary_accuracy: 0.4500 \nEpoch 5\/5 10\/10 [==============================] - \n0s 4ms\/step - loss: 0.7013 - binary_crossentropy: 0.7013 - binary_accuracy: 0.4844\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23767\" rel=\"nofollow noreferrer\">Link to Github Issue<\/a><\/p>\n","answers":[]}
+{"QuestionId":53320451,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-15T13:22:21.273","AcceptedAnswerId":null,"OwnerUserId":3603324.0,"Title":"Regarding loss weighting in Keras regression problem for multiple outputs","Body":"<p>I am running a Hyperas optimization for regression problem, with 3 predictors (X) and 2 targets (Y).<\/p>\n\n<p>I did this, after ingesting the raw data:<\/p>\n\n<pre><code>X_train, X_val, Y_train, Y_val = train_test_split(X, Y, test_size=0.2, random_state=111)\n\n# Input layers and Hidden Layers\nmodel = Sequential()\nmodel.add(Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}, input_dim = X_train.shape[1]))\nmodel.add(Activation({{choice(['tanh','relu', 'sigmoid'])}}))\nmodel.add(Dropout({{uniform(0, 1)}}))\nmodel.add(Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}))\nmodel.add(Activation({{choice(['tanh','relu', 'sigmoid'])}}))\nmodel.add(Dropout({{uniform(0, 1)}}))\n\n# Output layer\nmodel.add(Dense(Y_train.shape[1]))\nmodel.add(Activation('linear'))\n\nmodel.compile(loss='mae', metrics=['mae'],optimizer=optimizer, loss_weights=[0.6,0.4])\n\nhistory = model.fit(X_train, Y_train,\n          batch_size={{choice([16,32,64,128])}},\n          epochs={{choice([20000])}},\n          verbose=2,\n          validation_data=(X_val, Y_val),\n          callbacks=callbacks_list)\n<\/code><\/pre>\n\n<p>However, when running this, it says:<\/p>\n\n<pre><code>ValueError: When passing a list as loss_weights, it should have one entry per model output. The model has 1 outputs, but you passed loss_weights=[1, 1]\n<\/code><\/pre>\n\n<p>I'm guessing its due to the format of my inputs and outputs. However, I can't figure out the proper format for which I am supposed to feed it into the model.<\/p>\n\n<p>Appreciate your advice please, thank you.<\/p>\n","answers":[]}
+{"QuestionId":53320680,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-15T13:34:43.747","AcceptedAnswerId":null,"OwnerUserId":10520111.0,"Title":"Tensorflow seems to have problems with scientific notation","Body":"<p>Since my network kept giving high loss and zero accuracy I have decided to look into the tensors that are going in it.<\/p>\n\n<p>I have noticed that my normalizing function goes ok with decimal numbers but leaves scientific notation numbers as they are thus creating chaos with the samples.<\/p>\n\n<p>After my simple normalization and batching, loaded from lists of filepaths and labels:<\/p>\n\n<pre><code>def _parse_image(imagepath, label):\n    image_string = tf.read_file(imagepath)\n    image_decoded = tf.image.decode_jpeg(image_string, channels=3)\n    image_resized = tf.image.resize_images(image_decoded, [img_size, img_size])\n    image = image_resized * (1 \/ 255)\n    return image, label\n\ndataset = tf.data.Dataset.from_tensor_slices((dataset_train))\ndataset = dataset.map(_parse_image)\ndataset = dataset.shuffle(buffer_size=1000)\ndataset = dataset.batch(batch_size)\niterator = dataset.make_initializable_iterator()\ntrain_next = iterator.get_next()\n<\/code><\/pre>\n\n<p>And then<\/p>\n\n<pre><code>batch_x, batch_y = sess.run(train_next)\nop, c = sess.run([optimiser, cross_entropy], feed_dict={x: batch_x, y: np.expand_dims(batch_y, axis=-1)})\n<\/code><\/pre>\n\n<p>I get this is fed in feed_dict for decimals<\/p>\n\n<pre><code>[0.3425858  0.42196694 0.35618874]\n[0.28829658 0.311826   0.311826  ]\n[0.23520222 0.2508885  0.26069242]\n<\/code><\/pre>\n\n<p>And this for scientific notation<\/p>\n\n<pre><code>[7.11764753e-01 5.84313750e-01 7.27451026e-01]\n[7.07843184e-01 5.68627477e-01 7.03921616e-01]\n[7.08823562e-01 5.60784340e-01 7.23529458e-01]\n<\/code><\/pre>\n\n<p>Tried to do some transforms into float32 in a couple of points but it didn't work.<\/p>\n","answers":[]}
+{"QuestionId":53320728,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-15T13:37:59.333","AcceptedAnswerId":53323295.0,"OwnerUserId":4296244.0,"Title":"Extract one hot encoding from a file into a dataset","Body":"<p>I have a dataset images and corresponding labels, where to each image file there is a .txt file which contains the one hot encoding:<\/p>\n\n<pre><code>0\n0\n0\n0\n1\n0\n<\/code><\/pre>\n\n<p>My code looks something like this:<\/p>\n\n<pre><code>imageString = tf.read_file('image.jpg')\nimageDecoded = tf.image.decode_jpeg(imageString)\n\nlabelString = tf.read_file(labelPath)\n# decode csv string\n<\/code><\/pre>\n\n<p>but labelString looks like this:<\/p>\n\n<pre><code>tf.Tensor(b'0\\n0\\n0\\n0\\n1\\n', shape=(), dtype=string)\n<\/code><\/pre>\n\n<p>Is there a way to transform this into an array of numbers inside tensorflow?<\/p>\n","answers":[{"AnswerId":"53323295","CreationDate":"2018-11-15T15:59:04.793","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a function to do that.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef read_label_file(labelPath):\n    # Read file\n    labelStr = tf.io.read_file(labelPath)\n    # Split string (returns sparse tensor)\n    labelStrSplit = tf.strings.split([labelStr])\n    # Convert sparse tensor to dense\n    labelStrSplitDense = tf.sparse.to_dense(labelStrSplit, default_value='')[0]\n    # Convert to numbers\n    labelNum = tf.strings.to_number(labelStrSplitDense)\n    return labelNum\n<\/code><\/pre>\n\n<p>A test case:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Write file for test\nlabelPath = 'labelData.txt'\nlabelTxt = '0\\n0\\n0\\n0\\n1\\n0'\nwith open(labelPath, 'w') as f:\n    f.write(labelTxt)\n# Test the function\nwith tf.Session() as sess:\n    label_data = read_label_file(labelPath)\n    print(sess.run(label_data))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[0. 0. 0. 0. 1. 0.]\n<\/code><\/pre>\n\n<p>Note the function, as I wrote it, uses some of the new-ish API endpoints, you can also write it as below for more backwards compatibility, with almost the same meaning (there are slight differences between <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/strings\/split\" rel=\"nofollow noreferrer\"><code>tf.strings.split<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/string_split\" rel=\"nofollow noreferrer\"><code>tf.string_split<\/code><\/a>):<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef read_label_file(labelPath):\n    labelStr = tf.read_file(labelPath)\n    labelStrSplit = tf.string_split([labelStr], delimiter='\\n')\n    labelStrSplitDense = tf.sparse_to_dense(labelStrSplit.indices,\n                                            labelStrSplit.dense_shape,\n                                            labelStrSplit.values, default_value='')[0]\n    labelNum = tf.string_to_number(labelStrSplitDense)\n    return labelNum\n<\/code><\/pre>\n"},{"AnswerId":"53321228","CreationDate":"2018-11-15T14:04:56.300","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>You can use basic python commands and convert it to a tensor.  Try...<\/p>\n\n<pre><code>with open(labelPath) as f:\n    lines = f.readlines()\n    lines = [int(l.strip()) for l in lines if l.strip()]\nlabelString = tf.convert_to_tensor(lines, dtype='int32')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53320823,"AnswerCount":1,"Tags":"<tensorflow><tensorrt><deeplab>","CreationDate":"2018-11-15T13:43:36.023","AcceptedAnswerId":null,"OwnerUserId":5191402.0,"Title":"running deeplab v3+ with tensorRT","Body":"<p>i'm trying to optimize a deeplab v3+ model using tensorRT and i get the following errors:<\/p>\n\n<pre><code>    UFF Version 0.5.5\n=== Automatically deduced input nodes ===\n[name: \"ImageTensor\"\nop: \"Placeholder\"\nattr {\n  key: \"_output_shapes\"\n  value {\n    list {\n      shape {\n        dim {\n          size: 1\n        }\n        dim {\n          size: -1\n        }\n        dim {\n          size: -1\n        }\n        dim {\n          size: 3\n        }\n      }\n    }\n  }\n}\nattr {\n  key: \"dtype\"\n  value {\n    type: DT_UINT8\n  }\n}\nattr {\n  key: \"shape\"\n  value {\n    shape {\n      dim {\n        size: 1\n      }\n      dim {\n        size: -1\n      }\n      dim {\n        size: -1\n      }\n      dim {\n        size: 3\n      }\n    }\n  }\n}\n]\n=========================================\n\n=== Automatically deduced output nodes ===\n[name: \"Squeeze_1\"\nop: \"Squeeze\"\ninput: \"resize_images\/ResizeNearestNeighbor\"\nattr {\n  key: \"T\"\n  value {\n    type: DT_INT64\n  }\n}\nattr {\n  key: \"_output_shapes\"\n  value {\n    list {\n      shape {\n        dim {\n          size: 1\n        }\n        dim {\n          size: -1\n        }\n        dim {\n          size: -1\n        }\n      }\n    }\n  }\n}\nattr {\n  key: \"squeeze_dims\"\n  value {\n    list {\n      i: 3\n    }\n  }\n}\n]\n==========================================\n\nUsing output node Squeeze_1\nConverting to UFF graph\nWarning: No conversion function registered for layer: ResizeNearestNeighbor yet.\nConverting resize_images\/ResizeNearestNeighbor as custom op: ResizeNearestNeighbor\nWarning: No conversion function registered for layer: ExpandDims yet.\nConverting ExpandDims_1 as custom op: ExpandDims\nWarning: No conversion function registered for layer: Slice yet.\nConverting Slice as custom op: Slice\nWarning: No conversion function registered for layer: ArgMax yet.\nConverting ArgMax as custom op: ArgMax\nWarning: No conversion function registered for layer: ResizeBilinear yet.\nConverting ResizeBilinear_2 as custom op: ResizeBilinear\nWarning: No conversion function registered for layer: ResizeBilinear yet.\nConverting ResizeBilinear_1 as custom op: ResizeBilinear\nTraceback (most recent call last):\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"C:\\Users\\iariav\\Anaconda3\\envs\\tensorflow\\Scripts\\convert-to-uff.exe\\__main__.py\", line 9, in &lt;module&gt;\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\site-packages\\uff\\bin\\convert_to_uff.py\", line 89, in main\n    debug_mode=args.debug\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\site-packages\\uff\\converters\\tensorflow\\conversion_helpers.py\", line 187, in from_tensorflow_frozen_model\n    return from_tensorflow(graphdef, output_nodes, preprocessor, **kwargs)\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\site-packages\\uff\\converters\\tensorflow\\conversion_helpers.py\", line 157, in from_tensorflow\n    debug_mode=debug_mode)\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\site-packages\\uff\\converters\\tensorflow\\converter.py\", line 94, in convert_tf2uff_graph\n    uff_graph, input_replacements, debug_mode=debug_mode)\n  File \"c:\\users\\iariav\\anaconda3\\envs\\tensorflow\\lib\\site-packages\\uff\\converters\\tensorflow\\converter.py\", line 72, in convert_tf2uff_node\n    inp_node = tf_nodes[inp_name]\nKeyError: 'logits\/semantic\/biases\/read'\n<\/code><\/pre>\n\n<p>from what i understand, is this caused by some layers which are not supported by the uff converter? has anyone succeeded in converting a deeplab model to uff?\ni'm using the original deeplabv3+ model in tensorflow.<\/p>\n\n<p>thanks<\/p>\n","answers":[{"AnswerId":"53391719","CreationDate":"2018-11-20T11:10:53.740","ParentId":null,"OwnerUserId":"3411080","Title":null,"Body":"<p>yeah sometimes getting a specific model to work in TensorRT is a bit tricky due to layer support.\nWith the new TensorRT 5GA these are the supported layers (taken from the Developer Guide):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mGD8L.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mGD8L.png\" alt=\"Tensorflow Supported Layers\"><\/a><\/p>\n\n<p>Ask you can see you have some layers like <code>ResizeNearestNeighbor<\/code>, <code>ResizeBilinear<\/code> and <code>ArgMax<\/code>, your best approach and what I ended up doing is porting the network to a certain point and use the cpp API to create the layers I needed. Check the <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/sdk\/tensorrt-api\/c_api\/classnvinfer1_1_1_i_plugin_v2.html\" rel=\"nofollow noreferrer\">IPluginV2<\/a> and the IPluginCreator and see if you can implement the layers yourself.<\/p>\n\n<p>I think with time more layer support will roll out but I guess that if you can't wait just give it a try yourself.<\/p>\n"}]}
+{"QuestionId":53321316,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><tensorboard>","CreationDate":"2018-11-15T14:10:38.730","AcceptedAnswerId":null,"OwnerUserId":4408280.0,"Title":"PR-curves plugin (pr_curve_raw_data_op) not plotting completely when using metrics in Estimator API","Body":"<p>TensorBoard version (from pip package, also printed out when running <code>tensorboard<\/code>): 1.12.0\nTensorFlow version if different from TensorBoard: 1.12.0\nOS Platform and version (e.g., Linux Ubuntu 16.04): Arch Linux\nPython version (e.g. 2.7, 3.5): 3.6.5<\/p>\n\n<p>Hi there,<\/p>\n\n<p>I am trying to plot PR curves using <code>pr_curve_raw_data_op<\/code> function with precisions, recalls and other metrics which I have calculated before hand, but I'm getting wrong results in chart. I am using Estimator API, meaning I create my own dictionary which contains tuple of <code>tf.Tensor<\/code> and <code>tf.Operation<\/code> instances which are then forwarded to estimator for dumping summaries. <\/p>\n\n<p>I evaluate multiple steps during each checkpoint save, which means that I accumulate precision and recall through multiple steps (but I don't use pr_curves plugin to do so, I use custom written metrics), and all I want from the plugin is to dump the PR curve into the tensorboard. Here is the part of the code which serves for storing the <code>pr-curves<\/code> summary tensor into the metrics dictionary:<\/p>\n\n<pre><code>metrics_dict['pr_curve\/CLASS_{}'.format(cls_index)] = (\n    summary.raw_data_op(\n        'pr_curve\/CLASS_{}'.format(cls_index),\n        true_positive_counts=tps,\n        false_positive_counts=fps,\n        true_negative_counts=tns,\n        false_negative_counts=fns,\n        precision=precisions,\n        recall=recalls,\n        num_thresholds=11\n    ),\n    tf.no_op()\n)\n<\/code><\/pre>\n\n<p>I also tried to dump precisions as text summary just to be sure that output printed in terminal is the same that is being stored in summaries, and it was I dumped those summaries using this code:<\/p>\n\n<pre><code>metrics_dict['precisions_text\/CLASS_{}'.format(cls_index)] = (\n    tf.summary.text(name=\"aa\",\n                    tensor=tf.as_string(precs)),\n    tf.no_op()\n)\n<\/code><\/pre>\n\n<p>In the end I will attach couple of screenshots. First one shows precisions dumped to text summary and their values ( I only showed first 6 values, as for the remaining 4 values are 0.0):<\/p>\n\n<p><img src=\"https:\/\/user-images.githubusercontent.com\/14930845\/48550395-e5254000-e8d2-11e8-97f7-ce0c06c16df3.png\" alt=\"image\"><\/p>\n\n<p>Second screenshot contains those same precisions for same class and step plotted on PR curve:<\/p>\n\n<p><img src=\"https:\/\/user-images.githubusercontent.com\/14930845\/48552012-7bf3fb80-e8d7-11e8-9db6-19ff31780fcd.png\" alt=\"image\"><\/p>\n\n<p>You can see that there is only first point on the plot, with precision being 1.0 for 0.0 recall. That occurs in several more cases, along with chart cutting off at last point (for example, I have precision values from recalls 0.0 up to 0.6, but chart plots only up to 0.5).<\/p>\n","answers":[]}
+{"QuestionId":53321573,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2018-11-15T14:24:49.537","AcceptedAnswerId":53321920.0,"OwnerUserId":3214872.0,"Title":"How do I hide the GPU from one tf::Session() in a multi-session scenario?","Body":"<p>I have a C++ program that makes use of two different classifiers running in separate <code>tensorflow::Session<\/code>s. One of these models needs to run on the GPU, while the second is very small and I'd like for it to run on the CPU.<\/p>\n\n<p>I tried creating the sessions using:<\/p>\n\n<pre><code>auto options = tensorflow::SessionOptions();\noptions.config.mutable_gpu_options()-&gt;set_per_process_gpu_memory_fraction(0);\nm_session.reset(tensorflow::NewSession(options));\n<\/code><\/pre>\n\n<p>but this seems only to trigger the default \"hog the GPU\" behaviour.<\/p>\n\n<p>I also tried playing around with <code>options.config.mutable_gpu_options()-&gt;set_visible_device_list(\"-1\")<\/code> but there is a specific check in the TF code for invalid devices in the list that raises an error.<\/p>\n\n<p>Setting <code>CUDA_VISIBLE_DEVICES=-1<\/code> in the environment before running the program is of course not an option because the GPU should be visible for the session that needs it.<\/p>\n\n<p>Does anyone know how to forbid <em>just one<\/em> one session to use the GPU?<\/p>\n","answers":[{"AnswerId":"53321920","CreationDate":"2018-11-15T14:42:57.753","ParentId":null,"OwnerUserId":"6426627","Title":null,"Body":"<p>An easy workaround would be to temporarily set <code>CUDA_VISIBLE_DEVICES=-1<\/code> using <code>putenv(\"CUDA_VISIBLE_DEVICES=-1\");<\/code> and reset it after creating the session. <\/p>\n\n<pre><code>std::string str_tmp = \"CUDA_VISIBLE_DEVICES=\";\nstr_tmp += getenv(\"CUDA_VISIBLE_DEVICES\");\nputenv(\"CUDA_VISIBLE_DEVICES=-1\");\n\n#Create session\n\n#Reset environment variable\nputenv(str_tmp);\n<\/code><\/pre>\n\n<p>However, there might be a cleaner way to do it without changing your environment variables. <\/p>\n"}]}
+{"QuestionId":53321844,"AnswerCount":0,"Tags":"<python><numpy><pytorch><autograd>","CreationDate":"2018-11-15T14:39:08.290","AcceptedAnswerId":null,"OwnerUserId":10658041.0,"Title":"Update specific vector elements in PyTorch","Body":"<p>I have a large vector that I would like to update. I'll update it by adding an offset to specific elements in the vector. I specify a vector of indices that I want to update (call the index vector <code>ix<\/code>), and for each index, I specify a value that I want to add to that element (call the value vector <code>vals<\/code>). If all entries of the index vector are unique, then the following code suffices:<\/p>\n\n<pre><code>vec = torch.zeros(4, dtype=torch.float)\nix = torch.tensor([0,2], dtype=torch.long)\nvals = torch.tensor([0.2, 0.5], dtype=torch.float)\nvec[ix] += vals\n<\/code><\/pre>\n\n<p>However, this does not work if there are repeated indices in <code>ix<\/code>. A naive approach for the case of repeated indices is as follows:<\/p>\n\n<pre><code>for i in range(len(ix)):\n    vec[ix[i]] += vals[i]\n<\/code><\/pre>\n\n<p>But this doesn't scale well - it is very slow when <code>ix<\/code> is large. Are there any faster ways to do this? If there were a fast way to sum all entries of <code>vals<\/code> that have the same index in <code>ix<\/code>, then the solution should be easy.<\/p>\n\n<p><strong>Update:<\/strong><br>\nI found one solution that works pretty well, described below. I'd still love feedback for better solutions.<\/p>\n\n<pre><code># get unique indices\nix_unique = torch.unique(ix)\n\n# for each unique index, get sum of all vals with that index\nvals_unique = torch.stack([\n    torch.sum(torch.where(ix==i, vals, torch.zeros_like(vals))) \n    for i in ix_unique\n])\n\n# update vec\nvec[ix_unique] += vals_unique\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53321913,"AnswerCount":0,"Tags":"<python><keras><nlp>","CreationDate":"2018-11-15T14:42:31.327","AcceptedAnswerId":null,"OwnerUserId":10658055.0,"Title":"High accuracy measures when using pretrained embedding layer in Python","Body":"<p>I am trying to implement a pretrained embedding layer into my generative model using GloVe. <\/p>\n\n<p>Into the model I feed sequences of 50 (X) items pulled from a text, and it is to predict the 51. word (y) in the text. <\/p>\n\n<p>I reach an accuracy of 0.99 already when the model only has trained for 1\/100 iterations. What can be the issue?<\/p>\n\n<pre><code># create a weight matrix for words in training docs\nembedding_matrix = zeros((vocab_size, 100))\nfor word, i in tokenizer.word_index.items():\nembedding_vector = embeddings_index.get(word)\nif embedding_vector is not None:\n    embedding_matrix[i] = embedding_vector\n\n # define model\n model = Sequential() #assigning the sequential function to a model\nmodel.add(Embedding(vocab_size, 100, weights=[embedding_matrix], input_length=seq_length, trainable = False)) #defining embedding layer size\nmodel.add(LSTM(100, return_sequences=True)) #adding layer of nodes\nmodel.add(LSTM(100))  #adding layer of nodes\nmodel.add(Dense(100, activation='relu')) #specifying the structure of the hidden layer, recu is an argument of a rectified linear unit. \nmodel.add(Dense(vocab_size, activation='softmax')) #using the softmax function to creating probabilities\nprint(model.summary())\n# compile the model\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n\n# fit the model\nmodel.fit(X, y, batch_size=128, epochs=100, verbose=1)\n<\/code><\/pre>\n\n<p>Link to github: <a href=\"https:\/\/github.com\/KiriKoppelgaard\/Generative_model\" rel=\"nofollow noreferrer\">https:\/\/github.com\/KiriKoppelgaard\/Generative_model<\/a> \ncommit from  Nov 14, 2018 <\/p>\n","answers":[]}
+{"QuestionId":53322045,"AnswerCount":1,"Tags":"<python><machine-learning><keras><hyperas>","CreationDate":"2018-11-15T14:49:32.997","AcceptedAnswerId":53322131.0,"OwnerUserId":3603324.0,"Title":"Regarding error using Keras functional API","Body":"<p>I have a regression dataset:<\/p>\n\n<pre><code>X_train (float64) Size = (1616, 3) -&gt; i.e. 3 predictors\nY_train (float64) Size = (1616, 2) -&gt; i.e. 2 targets\n<\/code><\/pre>\n\n<p>I tried doing Hyperas using Functional API (my main purpose is to use the <code>loss_weights<\/code> option during compiling):<\/p>\n\n<pre><code>inputs1 = Input(shape=(X_train.shape[0], X_train.shape[1]))\n\nx  = Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}, activation={{choice(['tanh','relu', 'sigmoid'])}})(inputs1)\nx  = Dropout({{uniform(0, 1)}})(x)\n\nx  = Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}, activation={{choice(['tanh','relu', 'sigmoid'])}})(x)\nx  = Dropout({{uniform(0, 1)}})(x)\n\nx  = Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}, activation={{choice(['tanh','relu', 'sigmoid'])}})(x)\nx  = Dropout({{uniform(0, 1)}})(x)\n\nif conditional({{choice(['three', 'four'])}}) == 'four':\n    x  = Dense({{choice([np.power(2,1),np.power(2,2),np.power(2,3),np.power(2,4),np.power(2,5)])}}, activation={{choice(['tanh','relu', 'sigmoid'])}})(x)\n    x  = Dropout({{uniform(0, 1)}})(x)\n\noutput1 = Dense(1,  activation='linear')(x)\noutput2 = Dense(1,  activation='linear')(x)\n\nmodel = Model(inputs = inputs1, outputs = [output1,output2])\n\nadam    = keras.optimizers.Adam(lr={{choice([10**-3,10**-2, 10**-1])}})\nrmsprop = keras.optimizers.RMSprop(lr={{choice([10**-3,10**-2, 10**-1])}})\nsgd     = keras.optimizers.SGD(lr={{choice([10**-3,10**-2, 10**-1])}})\n\nchoiceval = {{choice(['adam', 'rmsprop','sgd'])}}\nif choiceval == 'adam':\n    optimizer = adam\nelif choiceval == 'rmsprop':\n    optimizer = rmsprop\nelse:\n    optimizer = sgd\n\nmodel.compile(loss='mae', metrics=['mae'],optimizer=optimizer, loss_weights=[0.5,0.5])\n\nearlyStopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=50, verbose=0, mode='auto')\ncheckpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=2, save_best_only=True, mode='max')\nlr_reducer = ReduceLROnPlateau(monitor='val_loss', factor=0.5, cooldown=1, patience=10, min_lr=1e-4,verbose=2)\ncallbacks_list = [earlyStopping, checkpoint, lr_reducer]\n\nhistory = model.fit(X_train, Y_train,\n          batch_size={{choice([16,32,64,128])}},\n          epochs={{choice([20000])}},\n          verbose=2,\n          validation_data=(X_val, Y_val),\n          callbacks=callbacks_list)\n<\/code><\/pre>\n\n<p>However, upon running it, I get the following error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected input_1 to have 3 dimensions, but got array with shape (1616, 3)\n<\/code><\/pre>\n\n<p>I would greatly appreciate if someone could point me to the direction of what is going wrong here. I suspect the input (i.e. <code>X_train<\/code>, <code>Y_train<\/code>) and also the Input shape might be at fault. Would appreciate any help here.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Ok so, indeed the fault was at the Input line:<\/p>\n\n<p>I changed it to: <code>inputs1 = Input(shape=(X_train.shape[1],))<\/code>.<\/p>\n\n<p>However, now I received another error:<\/p>\n\n<pre><code>ValueError: Error when checking model target: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[0.19204772, 0.04878049],\n   [0.20226056, 0.        ],\n   [0.12029842, 0.04878049],\n   ...,\n   [0.45188627, 0.14634146],\n   [0.26942276, 0.02439024],\n   [0.12942418, 0....\n<\/code><\/pre>\n","answers":[{"AnswerId":"53322131","CreationDate":"2018-11-15T14:53:59.877","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since your model has two output layers, you need to pass <strong>a list<\/strong> of <strong>two arrays<\/strong> as the true target (i.e. <code>y<\/code>) when calling <code>fit()<\/code> method. For example like this:<\/p>\n\n<pre><code>model.fit(X_train, [Y_train[:,0:1], Y_train[:,1:]], ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53322139,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-11-15T14:54:24.217","AcceptedAnswerId":null,"OwnerUserId":1116068.0,"Title":"LSTM model does not learn a simple pattern","Body":"<p>I'm a noob in ML, and tried to write an LSTM model that will process batches of sequences and will detect the following simple pattern: If a sequence starts with an odd number then the target is 0 otherwise it is 1:<\/p>\n\n<p>data:<\/p>\n\n<pre><code>[[[ 1  2  3]\n  [ 2  3  4]\n  [ 3  4  5]\n  [ 4  5  6]\n  [ 5  6  7]] #starts with 1 -&gt; 0\n\n [[ 6  7  8]\n  [ 7  8  9]\n  [ 8  9 10]\n  [ 9 10 11]\n  [10 11 12]] #starts with 6 -&gt; 1\n\n [[11 12 13]\n  [12 13 14]\n  [13 14 15]\n  [14 15 16]\n  [15 16 17]]] #starts with 11 -&gt; 0\n<\/code><\/pre>\n\n<p>target:<\/p>\n\n<pre><code>[0 1 0]\n<\/code><\/pre>\n\n<p>code:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom keras import callbacks\nfrom keras import optimizers\nfrom keras.layers import LSTM, Dense, Flatten, Dropout\nfrom keras.layers.advanced_activations import LeakyReLU\nfrom keras.models import Sequential\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.utils import shuffle\n\n\ndef demo():\n    scaler = StandardScaler()\n    dummy_data = pd.DataFrame(data=[[x, x+1, x+2, int((x - 1) \/ 5 % 2)] for x in range(1, 1001)],\n                              columns=['a', 'b', 'c', 'target'])\n\n    dummy_data[['a', 'b', 'c']] = scaler.fit_transform(dummy_data[['a', 'b', 'c']])\n    data = dummy_data.loc[:, dummy_data.columns != 'target']\n    target = dummy_data['target']\n    data = np.array(np.split(data.values, 200))\n    target = np.array(np.split(target.values, 200))\n    data, target = shuffle(data, target)\n    target = np.array(list(map(lambda x: x[0],target)))\n    print(data[:3,:],target[:3])\n    x_train, x_test, y_train, y_test = train_test_split(data, target, test_size=0.25, random_state=4)\n\n    opt = optimizers.Adam(lr=0.0005, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)\n    # build the model\n    model = Sequential()\n\n\n    num_features = data.shape[2]\n    num_samples = data.shape[1]\n\n    first_lstm = LSTM(32, batch_input_shape=(None, num_samples, num_features), return_sequences=True, activation='tanh')\n    model.add(\n        first_lstm)\n    model.add(LeakyReLU())\n    model.add(Dropout(0.2))\n    model.add(LSTM(16, return_sequences=True, activation='tanh'))\n    model.add(Dropout(0.2))\n    model.add(LeakyReLU())\n    model.add(LSTM(8, return_sequences=True, activation='tanh'))\n    model.add(LeakyReLU())\n    model.add(Flatten())\n    model.add(Dense(1, activation='sigmoid'))\n\n    model.compile(loss='binary_crossentropy', optimizer=opt,\n                  metrics=['accuracy'])\n\n    model.summary()\n\n    tb = callbacks.TensorBoard(log_dir='.\/logs\/', histogram_freq=10,\n                               batch_size=128,\n                               write_graph=True, write_grads=True, write_images=False,\n                               embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None)\n\n    loss_checkpoint = callbacks.ModelCheckpoint('.\/best_loss.hdf5', monitor='val_loss', verbose=1, save_best_only=True,\n                                                mode='min')\n\n    model.fit(x_train, y_train, batch_size=128, epochs=5000, validation_data=(x_test, y_test),\n                        callbacks=[tb, loss_checkpoint])\n\n\ndemo()\n<\/code><\/pre>\n\n<p>I'm expecting the net to learn this simple pattern however it fails, see the loss below:\n<a href=\"https:\/\/i.stack.imgur.com\/CUaWR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CUaWR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What could be improved, in order for the network to perform better?<\/p>\n","answers":[{"AnswerId":"53323271","CreationDate":"2018-11-15T15:57:53.547","ParentId":null,"OwnerUserId":"8609330","Title":null,"Body":"<p>According to the your comment, I would suggest to change dataset. Try something like:<\/p>\n\n<p>data:<\/p>\n\n<pre><code>[\n    [1, 3, 5],\n    [2, 4, 6],\n    [3, 5, 7]\n]\n<\/code><\/pre>\n\n<p>target:<code>[1, 0, 1]<\/code><\/p>\n\n<p>You should try dataset with a more pronounced pattern in the sequence. In theory, LSTM should perform better for such samples.<\/p>\n"}]}
+{"QuestionId":53322483,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-15T15:13:41.550","AcceptedAnswerId":null,"OwnerUserId":10658345.0,"Title":"Error with keras model: expected dense_53 to have 2 dimensions, but got array with shape (1, 1, 28, 28)","Body":"<p>I have a model and although I have a flatten layer it keeps giving me the error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_53 to have 2 dimensions, but got array with shape (1, 1, 28, 28)\n<\/code><\/pre>\n\n<p>Additionaly, if I remove the dense and flatten layers it gives me the same error for the maxPooling and then for the convolutional. Here is my entire code:<\/p>\n\n<pre><code>import os \nimport numpy as np \nfrom keras.preprocessing import image\nfrom keras import backend as K\n\nfrom keras.callbacks import EarlyStopping\nfrom keras.datasets import cifar10\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Flatten\nfrom keras.layers.convolutional import Conv2D\nfrom keras.optimizers import Adam\nfrom keras.layers.pooling import MaxPooling2D\nfrom keras.utils import to_categorical\n\nPATH = os.getcwd()\n\ntrain_path = PATH+'\/Downloads\/KerasStuff\/train_data\/rice\/'\ntrain_batch = os.listdir(train_path)\nx_train = np.empty((0,32,32,3))\n\n# if data are in form of images\nfor sample in train_batch:\n img_path = train_path+sample\n if img_path == '\/Users\/23athreyad\/Downloads\/KerasStuff\/train_data\/rice\/.DS_Store':\n  print(\"INVALID\")\n else:\n  x = image.load_img(img_path)\n  np.append(x_train,x)\n\ntest_path = PATH+'\/Downloads\/KerasStuff\/test_data\/rice\/'\ntest_batch = os.listdir(test_path)\nx_test = []\n\nfor sample in test_batch:\n img_path = test_path+sample\n x = image.load_img(img_path)\n # preprocessing if required\n x_test.append(x)\n\n# Create the model\n\nK.set_image_dim_ordering('tf')\n\nmodel = Sequential()\nmodel.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=(32, 32, 3)))\nmodel.add(MaxPooling2D(pool_size = (2,2),data_format=\"channels_last\"))\nprint(model.output_shape)\nmodel.add(Dropout(0.25))\nmodel.add(Flatten())\nprint(model.output_shape)\nmodel.add(Dense(7200,activation = 'relu'))\n\n# Compile the model\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=Adam(lr=0.0001, decay=1e-6),\n              metrics=['accuracy'])\nmodel.fit(x_train,y_train,batch_size = 10, epochs = 20,callbacks=[EarlyStopping(min_delta=0.001, patience=3)])\n<\/code><\/pre>\n\n<p>Can someone help me with this? I am new to Keras and have looked this up but not found any solutions.<\/p>\n","answers":[]}
+{"QuestionId":53324147,"AnswerCount":0,"Tags":"<python><tensorflow><caffe>","CreationDate":"2018-11-15T16:44:17.250","AcceptedAnswerId":null,"OwnerUserId":1972060.0,"Title":"What could be the cause of the little difference between Caffe convolution and Tensorflow output?","Body":"<p>I have Tensorflow network and I ported it in the Caffe format. All weights and algorithm is correct in Caffe and TF, I checked it several times.<\/p>\n\n<p>But when I run both frameworks and compare their output layer-by-layer, small differences start to appear.<\/p>\n\n<p>For example, if the result of the <code>BatchNormalization<\/code> at the one point in Caffe is  <code>0.940091<\/code> then for TF at the same point it is <code>0.939559<\/code>, so the difference is <code>0.0005319999999999769<\/code> . Very small but it gets amplified deeper in the network we go, and at the end values are very different.\nFor <code>Convolution<\/code> difference is present also(<code>-2.7540846 Caffe<\/code> <code>-2.7540843 TF<\/code>) but it gets more noticeable after first <code>BatchNormalization<\/code> <\/p>\n\n<p>I suppose it might be some internal difference and casting between some sort of <code>FP16, FP32, FP64<\/code>. Have no idea how TF handles conversion from python graph to C++ graph. <\/p>\n\n<p>May be someone encountered same problem?<\/p>\n","answers":[]}
+{"QuestionId":53324467,"AnswerCount":2,"Tags":"<tensorflow><gpu>","CreationDate":"2018-11-15T16:59:29.047","AcceptedAnswerId":53324876.0,"OwnerUserId":5623476.0,"Title":"Why gpu power consumption is at 40% when gpu utilization is 100%?","Body":"<p>nvidia-smi shows less than 100W power consumption out of 250W while GPU-Utilization is at 100%. Shouldn't power consumption be closer to GPU Utilization? How to interpret that?<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 410.73       Driver Version: 410.73       CUDA Version: 10.0     |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  GeForce GTX 108...  Off  | 00000000:01:00.0  On |                  N\/A |\n| 54%   71C    P2    90W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  GeForce GTX 108...  Off  | 00000000:02:00.0 Off |                  N\/A |\n| 45%   64C    P2    99W \/ 250W |  10377MiB \/ 11178MiB |    100%      Default |\n+-------------------------------+----------------------+----------------------+\n<\/code><\/pre>\n","answers":[{"AnswerId":"53324876","CreationDate":"2018-11-15T17:25:01.933","ParentId":null,"OwnerUserId":"295704","Title":null,"Body":"<p>GPU-Util is not intended to represent the GPU's total processing abilities.<\/p>\n\n<p>EXAMPLE:\nThere are two people exercising for the next 10 minutes.  One exerciser decides to walk and the other exerciser decides to run.  They are both busy so both exercisers show \"Human-Util\" = 100%. However, which excerciser is consuming more energy?  The runner is consuming more energy, therefore the runner might be at 90% W.  The walker may be only consuming 30% W.<\/p>\n\n<p>If your GPU-Util shows 65%, it means 650,000 out of every 1 million clock cycles are being used, but it does not mean it is running at max capacity within each of the 650,000 clock cycles that ran.<\/p>\n\n<p>It's the same at 100% or any other percentage.  Even though your GPU-Util shows 100% usage, it only means 1 million out of 1 million clock cycles were busy, but it may have only been using simple computations or a small amount of computations within each of the those clock cycles.<\/p>\n"},{"AnswerId":"53325727","CreationDate":"2018-11-15T18:24:39.790","ParentId":null,"OwnerUserId":"5623476","Title":null,"Body":"<p>Another explanation: power consumption value is ~0.5 seconds delayed from GPU utilization. Consecutive nvidia-smi calls show that average GPU utilization is in accordance with average power consumption.<\/p>\n\n<pre><code>$ while true ; do nvidia-smi | grep 10922MiB ; sleep .2 ; done\n| 54%   71C    P2    92W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   70C    P2    90W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 55%   70C    P2   254W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   72C    P2   155W \/ 250W |  10922MiB \/ 11175MiB |     99%      Default |\n| 54%   70C    P2    79W \/ 250W |  10922MiB \/ 11175MiB |      0%      Default |\n| 54%   72C    P2   193W \/ 250W |  10922MiB \/ 11175MiB |     87%      Default |\n| 54%   71C    P2    97W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   70C    P2    95W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   71C    P2   233W \/ 250W |  10922MiB \/ 11175MiB |     99%      Default |\n| 54%   71C    P2    74W \/ 250W |  10922MiB \/ 11175MiB |     92%      Default |\n| 54%   70C    P2    78W \/ 250W |  10922MiB \/ 11175MiB |      0%      Default |\n| 54%   71C    P2   156W \/ 250W |  10922MiB \/ 11175MiB |     77%      Default |\n| 54%   71C    P2    82W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   71C    P2   264W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   72C    P2    82W \/ 250W |  10922MiB \/ 11175MiB |     99%      Default |\n| 54%   70C    P2    74W \/ 250W |  10922MiB \/ 11175MiB |      6%      Default |\n| 54%   70C    P2    73W \/ 250W |  10922MiB \/ 11175MiB |      4%      Default |\n| 54%   71C    P2   207W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   71C    P2   109W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   70C    P2    91W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   70C    P2   134W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n| 54%   72C    P2   260W \/ 250W |  10922MiB \/ 11175MiB |     98%      Default |\n| 54%   70C    P2    79W \/ 250W |  10922MiB \/ 11175MiB |     64%      Default |\n| 54%   71C    P2   205W \/ 250W |  10922MiB \/ 11175MiB |     92%      Default |\n| 54%   71C    P2    96W \/ 250W |  10922MiB \/ 11175MiB |    100%      Default |\n<\/code><\/pre>\n"}]}
+{"QuestionId":53324596,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><neural-network>","CreationDate":"2018-11-15T17:06:59.237","AcceptedAnswerId":53325468.0,"OwnerUserId":5533928.0,"Title":"Problems when implementing Keras model in Tensorflow","Body":"<p>I'm just starting off with <code>Tensorflow<\/code>.<\/p>\n\n<p>I tried implementing a model to classify digits in the MNSIT dataset.<\/p>\n\n<p>I am familiar with <code>Keras<\/code>, so I first used it to create the model.<\/p>\n\n<p>Keras code:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.datasets import mnist\nfrom os import path\n\nimport numpy as np\n\nnetwork = Sequential()\nnetwork.add(Dense(700, input_dim=784, activation='tanh'))\nnetwork.add(Dense(500, activation='tanh'))\nnetwork.add(Dense(500, activation='tanh'))\nnetwork.add(Dense(500, activation='tanh'))\nnetwork.add(Dense(10, activation='softmax'))\n\nnetwork.compile(loss='categorical_crossentropy', optimizer='adam')\n\n(x_train, y_temp), (x_test, y_test) = mnist.load_data()\ny_train = vectorize(y_temp)  # I defined this function to create vectors of the labels. It works without issues.\n\nx_train = x_train.reshape(x_train.shape[0], x_train.shape[1]*x_train.shape[2])\n\nnetwork.fit(x_train, y_train, batch_size=100, epochs=3)\n\nx_test = x_test.reshape(x_test.shape[0], x_test.shape[1]*x_test.shape[2])\n\n\nscores = network.predict(x_test)\n\ncorrect_pred = 0\nfor i in range(len(scores)):\n    if np.argmax(scores[i]) == y_test[i]:\n        correct_pred += 1\n\nprint((correct_pred\/len(scores))*100)\n<\/code><\/pre>\n\n<p>The above code gives me an accuracy of around 92%.<\/p>\n\n<p>I tried implementing the same model in Tensorflow:<\/p>\n\n<pre><code>import sys\n\nimport tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\n\ndata = input_data.read_data_sets('.', one_hot=True)\n\nsess = tf.InteractiveSession()\n\nx = tf.placeholder(tf.float32, [None, 784])\ny = tf.placeholder(tf.float32, [None, 10])\n\nw = tf.Variable(tf.zeros([784, 700]))\nw2 = tf.Variable(tf.zeros([700, 500]))\nw3 = tf.Variable(tf.zeros([500, 500]))\nw4 = tf.Variable(tf.zeros([500, 500]))\nw5 = tf.Variable(tf.zeros([500, 10]))\n\nh1 = tf.nn.tanh(tf.matmul(x, w))\nh2 = tf.nn.tanh(tf.matmul(h1, w2))\nh3 = tf.nn.tanh(tf.matmul(h2, w3))\nh4 = tf.nn.tanh(tf.matmul(h3, w4))\nh = tf.matmul(h4, w5)\n\nloss = tf.math.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=h, labels=y))\ngradient_descent = tf.train.AdamOptimizer().minimize(loss)\n\ncorrect_mask = tf.equal(tf.argmax(h, 1), tf.argmax(y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_mask, tf.float32))\n\nsess.run(tf.global_variables_initializer())\n\nfor i in range(3):\n    batch_x, batch_y = data.train.next_batch(100)\n    loss_print = tf.print(loss, output_stream=sys.stdout)\n    sess.run([gradient_descent, loss_print], feed_dict={x: batch_x, y: batch_y})\n\nans = sess.run(accuracy, feed_dict={x: data.test.images, y: data.test.labels})\n\nprint(ans)\n<\/code><\/pre>\n\n<p>However, this code only gave me an accuracy of around 11%.\nI tried increasing the number of epochs to 1000, but the result didn't change. Furthermore, the loss in every epoch was the same (2.30).<\/p>\n\n<p>Am I missing something in the Tensorflow code?<\/p>\n","answers":[{"AnswerId":"53325468","CreationDate":"2018-11-15T18:06:17.700","ParentId":null,"OwnerUserId":"5533928","Title":null,"Body":"<p>Turns out, the problem was that I initialized the weights as zeros!<\/p>\n\n<p>Simply changing<\/p>\n\n<pre><code>w = tf.Variable(tf.zeros([784, 700]))\nw2 = tf.Variable(tf.zeros([700, 500]))\nw3 = tf.Variable(tf.zeros([500, 500]))\nw4 = tf.Variable(tf.zeros([500, 500]))\nw5 = tf.Variable(tf.zeros([500, 10]))\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>w = tf.Variable(tf.random_normal([784, 700], seed=42))\nw2 = tf.Variable(tf.random_normal([700, 500], seed=42))\nw3 = tf.Variable(tf.random_normal([500, 500], seed=42))\nw4 = tf.Variable(tf.random_normal([500, 500], seed=42))\nw5 = tf.Variable(tf.random_normal([500, 10], seed=42))\n<\/code><\/pre>\n\n<p>gave significant improvements.<\/p>\n"}]}
+{"QuestionId":53324637,"AnswerCount":0,"Tags":"<tensorflow><machine-learning>","CreationDate":"2018-11-15T17:09:46.740","AcceptedAnswerId":null,"OwnerUserId":509326.0,"Title":"Detect CANBUS Patterns using Machine Learning","Body":"<p>I would like to use AI\/TensorFlow\/Machine Learning of some description in order to recognise patterns in a set of data.<\/p>\n\n<p>Most of the samples of machine learning seem to be based on decision making, whether a value falls above or a below a line, then the decision is good or bad. But what I seem to have is a set of data, that may or may not have any relationship, may or may not have a pattern, and a single entity by itself is neither good nor bad, but the whole set of data together needs to be used to work out the type of data.<\/p>\n\n<p>The data in question is a set of readings over a period of time from an automotive CANBUS reader - so hexadecimal values containing a Command ID, and 1 or more values, usually in the format: FFF#FF:FF:FF:FF:FF:FF:FF:FF<\/p>\n\n<p>eg:<\/p>\n\n<p>35C#F4:C6:4E:7C:31:2B:60:28<\/p>\n\n<p>One canbus command, may contain one or more sensor readers - so for example F4 above might represent the steering position, C6 might indicate the pitch of the vehicle, 4E might indicate the roll.<\/p>\n\n<p>Each sensor may take up one or more octets, 7C:31 might indicate the speed of the vehicle, and \"B\" of \"2B\" might indicate whether the engine is running or not.<\/p>\n\n<p>I can detect the data with a human eye, and can see that the relevent item might be linear, random, static (ie a limited set of values) or it might be a bell curve.<\/p>\n\n<p>Im new to statistical analysis and machine learning so do not know the terminology. In the first instance Im looking for the terminology and references to appropriate material that will help me achieve my goal.<\/p>\n\n<p>My goal is given a sample of data from a CANBUS reader, to scan all the values, using each and every possible combination of numbers (eg octets 1, 1+2, 1+2+3... 2, 2+3, 2+3+4... 3, 3+4 etc) to detect patterns within the data and work out whether those patterns are linear, curve, static, or random.<\/p>\n\n<p>I want to basically read as many CAN-BUS readings from as many cars as I can, throw it at a program to do some analysis and learning and hopefully provide me with possibilities to investigate further so I can monitor various systems on different cars.<\/p>\n\n<p>It seems like a relatively simple premise, but extremely hard for me to define.<\/p>\n","answers":[]}
+{"QuestionId":53324641,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><callback>","CreationDate":"2018-11-15T17:09:58.617","AcceptedAnswerId":null,"OwnerUserId":7750161.0,"Title":"Callback in Tensorflow","Body":"<p>In Keras we can add callback simply as below:<\/p>\n\n<pre><code>self.model.fit(X_train,y_train,callbacks=[Custom_callback])\n<\/code><\/pre>\n\n<p>The callback is defined in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/Callback\" rel=\"nofollow noreferrer\">doc<\/a>, but I cannot find any example to use them. Could anyone tell me how to add custom callbacks into TensorFlow?<\/p>\n","answers":[]}
+{"QuestionId":53325418,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-15T18:01:52.087","AcceptedAnswerId":null,"OwnerUserId":9242076.0,"Title":"Pytorch speed comparison - GPU slower than CPU","Body":"<p>I was trying to find out if GPU tensor operations are actually faster than CPU ones. So, I wrote this particular code below to implement a simple 2D addition of CPU tensors and GPU cuda tensors successively to see the speed difference:<\/p>\n\n<pre><code>import torch\nimport time\n\n###CPU\nstart_time = time.time()\na = torch.ones(4,4)\nfor _ in range(1000000):\n    a += a\nelapsed_time = time.time() - start_time\n\nprint('CPU time = ',elapsed_time)\n\n###GPU\nstart_time = time.time()\nb = torch.ones(4,4).cuda()\nfor _ in range(1000000):\n    b += b\nelapsed_time = time.time() - start_time\n\nprint('GPU time = ',elapsed_time)\n<\/code><\/pre>\n\n<p>To my surprise, the CPU time was 0.93 sec and the GPU time was as high as 63 seconds. Am I doing the cuda tensor operation properly or is the concept of cuda tensors works faster only in very highly complex operations, like in neural networks?<\/p>\n\n<p>Note: My GPU is NVIDIA 940MX and <code>torch.cuda.is_available()<\/code> call returns True.<\/p>\n","answers":[{"AnswerId":"53327162","CreationDate":"2018-11-15T20:06:24.023","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>GPU acceleration works by heavy parallelization of computation. On a GPU you have a huge amount of cores, each of them is not very powerful, but the huge amount of cores here matters.<\/p>\n\n<p>Frameworks like PyTorch do their to make it possible to compute as much as possible in parallel. In general matrix operations are very well suited for parallelization, but still it isn't always possible to parallelize computation!<\/p>\n\n<p>In your example you have a loop:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>b = torch.ones(4,4).cuda()\nfor _ in range(1000000):\n    b += b\n<\/code><\/pre>\n\n<p>You have <em>1000000<\/em> operations, but due to the structure of the code it impossible to parallelize much of these computations. If you think about it, to compute the <em>next<\/em> <code>b<\/code> you need to know the value of the <em>previous<\/em> (or <em>current<\/em>) <code>b<\/code>.<\/p>\n\n<p>So you have <em>1000000<\/em> operations, but each of these has to be computed one after another. Possible parallelization is limited to the size of your tensor. This size though is not very large in your example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>torch.ones(4,4)\n<\/code><\/pre>\n\n<p>So you only can parallelize <em>16<\/em> operations (additions) per iteration.\nAs the CPU has <em>few<\/em>, but much more <em>powerful<\/em> cores, it is just much faster for the given example!<\/p>\n\n<p>But things change if you change the size of the tensor, then PyTorch is able to parallelize much more of the overall computation. I changed the iterations to <em>1000<\/em> (because I did not want to wait so long :), but you can put in any value you like, the relation between CPU and GPU should stay the same.<\/p>\n\n<p>Here are the results for different tensor sizes:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#torch.ones(4,4)       - the size you used\nCPU time =  0.00926661491394043\nGPU time =  0.0431208610534668\n\n#torch.ones(40,40)     - CPU gets slower, but still faster than GPU\nCPU time =  0.014729976654052734\nGPU time =  0.04474186897277832\n\n#torch.ones(400,400)   - CPU now much slower than GPU\nCPU time =  0.9702610969543457\nGPU time =  0.04415607452392578\n\n#torch.ones(4000,4000) - GPU much faster then CPU \nCPU time =  38.088677167892456\nGPU time =  0.044649362564086914\n<\/code><\/pre>\n\n<p>So as you see, where it is possible to parallelize stuff (here the addition of the tensor elements), GPU becomes very powerful. <br>GPU time is not changing at all for the given calculations, the GPU can handle much more! <br><em>(as long as it doesn't run out of memory :)<\/em><\/p>\n"}]}
+{"QuestionId":53325436,"AnswerCount":0,"Tags":"<keras><neural-network><regression><lstm><prediction>","CreationDate":"2018-11-15T18:03:29.140","AcceptedAnswerId":null,"OwnerUserId":10079650.0,"Title":"Predicting the price of the natural gas using LSTM neural network","Body":"<p>I want to build a model using Keras to predict the price of the natural gas.\nThe dataset contains the price for the gas daily and monthly since 1997 and it is available <a href=\"https:\/\/datahub.io\/core\/natural-gas\" rel=\"nofollow noreferrer\">Here<\/a>. <\/p>\n\n<p>The following graph shows the prices during a sequence of days. X is days and Y is the price.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/n7UMB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n7UMB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have tried LSTM with 4,50,100 cell in hidden layer but the accuracy still not was bad and the model failed to predict future price.\nI have added another two hidden layers (full connected) with 100 and 128 cell but it did not work too.<\/p>\n\n<p>This is the model and the result form training process:<\/p>\n\n<pre><code>num_units = 100\nactivation_function = 'sigmoid'\noptimizer = 'adam'\nloss_function = 'mean_squared_error'\nbatch_size = 5\nnum_epochs = 10\nlog_file_name = f\"{SEQ_LEN}-SEQ-{1}-PRED-{int(time.time())}\"\n\n# Initialize the model (of a Sequential type)\nmodel = Sequential()\n\n# Adding the input layer and the LSTM layer\nmodel.add(LSTM(units = num_units, activation = activation_function,input_shape=(None, 1)))\n\n# Adding the output layer\nmodel.add(Dense(units = 1))\n\n# Compiling the RNN\nmodel.compile(optimizer = optimizer, loss = loss_function, metrics=['accuracy'])\n\n# Using the training set to train the model\nhistory = model.fit(train_x, train_y, batch_size = batch_size, epochs =num_epochs,validation_data=(test_x, test_y))\n<\/code><\/pre>\n\n<p>and the output is :<\/p>\n\n<blockquote>\n  <p>Train on 4362 samples, validate on 1082 samples<\/p>\n  \n  <p>Epoch 1\/10\n  4362\/4362 [==============================] - 11s 3ms\/step - loss: 0.0057 - acc: 2.2925e-04 - val_loss: 0.0016 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 2\/10\n  4362\/4362 [==============================] - 9s 2ms\/step - loss: 6.2463e-04 - acc: 4.5851e-04 - val_loss: 0.0013 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 3\/10\n  4362\/4362 [==============================] - 9s 2ms\/step - loss: 6.1073e-04 - acc: 2.2925e-04 - val_loss: 0.0014 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 4\/10\n  4362\/4362 [==============================] - 8s 2ms\/step - loss: 5.4403e-04 - acc: 4.5851e-04 - val_loss: 0.0014 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 5\/10\n  4362\/4362 [==============================] - 7s 2ms\/step - loss: 5.4765e-04 - acc: 4.5851e-04 - val_loss: 0.0012 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 6\/10\n  4362\/4362 [==============================] - 8s 2ms\/step - loss: 5.1991e-04 - acc: 4.5851e-04 - val_loss: 0.0013 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 7\/10\n  4362\/4362 [==============================] - 7s 2ms\/step - loss: 5.7324e-04 - acc: 2.2925e-04 - val_loss: 0.0011 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 8\/10\n  4362\/4362 [==============================] - 7s 2ms\/step - loss: 4.4248e-04 - acc: 4.5851e-04 - val_loss: 0.0011 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 9\/10\n  4362\/4362 [==============================] - 7s 2ms\/step - loss: 4.3868e-04 - acc: 4.5851e-04 - val_loss: 0.0011 - val_acc: 0.0018<\/p>\n  \n  <p>Epoch 10\/10\n  4362\/4362 [==============================] - 7s 2ms\/step - loss: 4.6654e-04 - acc: 4.5851e-04 - val_loss: 0.0011 - val_acc: 0.0018<\/p>\n<\/blockquote>\n\n<p>How to know the number of layers and cells for problem like this? Anyone can suggest a netwrok structure that can solve this problem?<\/p>\n","answers":[]}
+{"QuestionId":53325753,"AnswerCount":1,"Tags":"<tensorflow><mnist>","CreationDate":"2018-11-15T18:26:27.997","AcceptedAnswerId":null,"OwnerUserId":10659241.0,"Title":"Downloading Fashion MNIST file in TensorFlow tutorial is taking forever","Body":"<p>I am trying to do this tutorial for a machine learning class I am taking in college.<\/p>\n\n<blockquote>\n  <p>www.tensorflow.org\/tutorials\/keras\/basic_classification<\/p>\n<\/blockquote>\n\n<p>When it executes the lines<\/p>\n\n<pre><code> fashion_mnist = keras.datasets.fashion_mnist\n\n (train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()\n<\/code><\/pre>\n\n<p>it is taking forever to download the data.  At the rate it is downloading, it is going to take a few days or weeks to download all of it.  I am using a MacBook.  My classmate is also using a MacBook and when he downloads the data it only takes a few seconds.  Please help.<\/p>\n","answers":[{"AnswerId":"58574753","CreationDate":"2019-10-26T21:09:21.243","ParentId":null,"OwnerUserId":"2342537","Title":null,"Body":"<p>In my case the download was giving me an error. By digging into the error I was able to find the file in which the base URL was declared, which in my case was: <\/p>\n\n<pre><code>\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/datasets\/fashion_mnist.py\n<\/code><\/pre>\n\n<p>At line 44 I have commented out the line:<\/p>\n\n<pre><code># base = 'https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/'\n<\/code><\/pre>\n\n<p>And declared a different base URL, which I had found looking into the <a href=\"https:\/\/github.com\/zalandoresearch\/fashion-mnist#get-the-data\" rel=\"nofollow noreferrer\">documentation<\/a> of the original dataset:<\/p>\n\n<pre><code>base = 'http:\/\/fashion-mnist.s3-website.eu-central-1.amazonaws.com\/'\n<\/code><\/pre>\n\n<p>The download started immediately and gave no errors. Hope this helps.<\/p>\n"}]}
+{"QuestionId":53326133,"AnswerCount":0,"Tags":"<tensorflow><timer><object-detection>","CreationDate":"2018-11-15T18:51:19.243","AcceptedAnswerId":null,"OwnerUserId":10466965.0,"Title":"Tensor Flow Object detection API notifications","Body":"<p>Tensorflow 1.12.0\nPython 3.5.0 \nWindows 10<\/p>\n\n<p>Hello all, I've created my own object detection model based on tensorflows object detection tutorial. I want to notify by SMS (via a service like twilio) when a object is detected, but I don't want to be notified by every frame of the same object class, instead I'd like to have a delay between text messages of objects i.e. every 5 seconds (at least) between every call. I've looked at threading and timer, but I fear that I'll restart the threading and timer every for call and was wondering if their is a more efficient way via the object detection API to accomplish this task. I know that I can print the actual class detected via<\/p>\n\n<pre><code>print [category_index.get(value) for index,value in enumerate(classes[0]) if scores[0,index] &gt; 0.5]\n<\/code><\/pre>\n\n<p>in the code<\/p>\n\n<pre><code>with detection_graph.as_default():\n  with tf.Session(graph=detection_graph) as sess:\n    while True:\n      ret, image_np = cap.read()\n      # Expand dimensions since the model expects images to have shape: [1, None, None, 3]\n      image_np_expanded = np.expand_dims(image_np, axis=0)\n      image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n      # Each box represents a part of the image where a particular object was detected.\n      boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n      # Each score represent how level of confidence for each of the objects.\n      # Score is shown on the result image, together with the class label.\n      scores = detection_graph.get_tensor_by_name('detection_scores:0')\n      classes = detection_graph.get_tensor_by_name('detection_classes:0')\n      num_detections = detection_graph.get_tensor_by_name('num_detections:0')\n      # Actual detection.\n      (boxes, scores, classes, num_detections) = sess.run(\n          [boxes, scores, classes, num_detections],\n          feed_dict={image_tensor: image_np_expanded})\n      # Visualization of the results of a detection.\n      vis_util.visualize_boxes_and_labels_on_image_array(\n          image_np,\n          np.squeeze(boxes),\n          np.squeeze(classes).astype(np.int32),\n          np.squeeze(scores),\n          category_index,\n          use_normalized_coordinates=True,\n          line_thickness=8)\n\n      cv2.imshow('object detection', cv2.resize(image_np, (800,600)))\n      if cv2.waitKey(25) &amp; 0xFF == ord('q'):\n        cv2.destroyAllWindows()\n        break\n<\/code><\/pre>\n\n<p>but again, I don't want an overflow of text messages to my phone, and I don't want my script to pause neither.. if any suggestions can be given, I'd appreciate it. Thank you all. <\/p>\n","answers":[]}
+{"QuestionId":53327026,"AnswerCount":1,"Tags":"<python><tensorflow><access-denied>","CreationDate":"2018-11-15T19:56:06.373","AcceptedAnswerId":53365376.0,"OwnerUserId":10330721.0,"Title":"Tensorflow access denied in config_util.py while training","Body":"<p>I have a problem with training a model in <code>tensorflow<\/code>. I am working on Windows 10. When I run the command:<\/p>\n\n<pre><code>python .\/object_detection\/model_main.py --pipeline_config_path=C:\/Tensorflow\/object-detection\/ssd_mobilenet_v1_coco_2018_01_28 --model_dir=C:\/Tensorflow\/object-detection\/output-model --num_train_steps=50000 --sample_1_of_n_eval_examples=1 --alsologtostderr\n<\/code><\/pre>\n\n<p>from <code>C:\/Tensorflow\/models\/research<\/code> to start the training process I get an error in line 95 <code>(proto_str = f.read())<\/code> from the <code>config_util.py<\/code> script. Below \nyou can see my whole console output:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \".\/object_detection\/model_main.py\", line 109, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\lucci\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \".\/object_detection\/model_main.py\", line 71, in main\n    FLAGS.sample_1_of_n_eval_on_train_examples))\n  File \"C:\\Tensorflow\\models\\research\\object_detection\\model_lib.py\", line 536, in create_estimator_and_inputs\n    config_override=config_override)\n  File \"C:\\Tensorflow\\models\\research\\object_detection\\utils\\config_util.py\", line 95, in get_configs_from_pipeline_file\n    proto_str = f.read()\n  File \"C:\\Users\\lucci\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 125, in read\n    self._preread_check()\n  File \"C:\\Users\\lucci\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 85, in _preread_check\n    compat.as_bytes(self.__name), 1024 * 512, status)\n  File \"C:\\Users\\lucci\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 528, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.UnknownError: NewRandomAccessFile failed to Create\/Open: C:\/Tensorflow\/object-detection\/ssd_mobilenet_v1_coco_2018_01_28 : Zugriff verweigert\n; Input\/output error\n<\/code><\/pre>\n\n<p>The error is in the last two lines: <code>Zugriff verweigert<\/code> is german and means <code>access denied<\/code>.<br>\nI am admin on this pc (it's my own pc) and i have FullControl to the folders (I double-checked it with PowerShell). When I try to move the folder to another\nplace, e.g. <code>C:\\Users\\lucci\\Documents\\<\/code> I get the same error. The problem still remains when I run the console as admin, when I try the command <code>runas \/user:lucci ...<\/code> and\nso on.<\/p>\n\n<p>Can anyone help me with this?\nI am using Python3.6.<\/p>\n\n<p><em>EDIT:<\/em> This also not helps: <a href=\"https:\/\/stackoverflow.com\/questions\/41116270\/tensorflow-windows-accessing-folders-deniednewrandomaccessfile-failed-to-creat\">Tensorflow Windows Accessing Folders Denied:&quot;NewRandomAccessFile failed to Create\/Open: Access is denied. ; Input\/output error&quot;<\/a><\/p>\n","answers":[{"AnswerId":"53365376","CreationDate":"2018-11-18T20:57:01.353","ParentId":null,"OwnerUserId":"10330721","Title":null,"Body":"<p>I finally found the solution on my own. It does not matter if you are operating on <strong>Linux<\/strong> or <strong>Windows<\/strong>.\nWhen you run the command you always have to specify the <strong>full path to the pipeline.config<\/strong> file.\nSo, when your pipeline.config-file is located in <code>C:\/ObjectDetection\/Model\/pipeline.config<\/code> it is not\nsufficient to specify the location like <code>C:\/ObjectDetection\/<\/code>. You have to specify the location\nas <code>C:\/ObjectDetection\/pipeline.config<\/code>.<\/p>\n\n<p>Running the command again with the <strong>full path for<\/strong> the Parameter <code>--pipeline_config_file<\/code> works great!!!<\/p>\n"}]}
+{"QuestionId":53328808,"AnswerCount":0,"Tags":"<python><tensorflow><keras><anaconda>","CreationDate":"2018-11-15T22:32:48.573","AcceptedAnswerId":null,"OwnerUserId":10358075.0,"Title":"ImportError: cannot import name 'transpose_shape'","Body":"<p>I am following the Coursera, Deep Learning course with Andrew Ng for object detection with YOLO algorithm, I am trying to run the algorithm on my PC using Windows and Anaconda Navigator, I installed Keras to run on TensorFlow backend. But when I typed the following:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n<\/code><\/pre>\n\n<p>It showed me the following: <\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-8-0bcaeab92875&gt;\", line 2, in &lt;module&gt;\n    from keras import backend as K\n\n  File \"C:\\Users\\AbanoubMamdouh\\Anaconda3\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n\n  File \"C:\\Users\\AbanoubMamdouh\\Anaconda3\\lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n\n  File \"C:\\Users\\AbanoubMamdouh\\Anaconda3\\lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n\n  File \"C:\\Users\\AbanoubMamdouh\\Anaconda3\\lib\\site-packages\\keras\\backend\\__init__.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n\n  File \"C:\\Users\\AbanoubMamdouh\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 24, in &lt;module&gt;\n    from ..utils.generic_utils import transpose_shape\n\nImportError: cannot import name 'transpose_shape'\n<\/code><\/pre>\n\n<p>Seems like it has a problem with this line in the tensorflow_backend.py file:<\/p>\n\n<pre><code>from ..utils.generic_utils import transpose_shape\n<\/code><\/pre>\n\n<p>Help would be much appreciated!\nCheers<\/p>\n","answers":[]}
+{"QuestionId":53328826,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-15T22:35:08.940","AcceptedAnswerId":null,"OwnerUserId":7164853.0,"Title":"Keras: InvalidArgumentError caused by placeholder when setting initial_states of LSTM","Body":"<p>I want to build a \"recurrent\" layer in Keras which contains multiple predefined layers (amongst others two lstm layers).<\/p>\n\n<p>The general scheme is to use a keras model as layer which contains the simple layers I need, instantiate it and apply it multiple times to an input and its own outputs (thus creating multiple \"nodes\" of the model\/layer with shared weights)<\/p>\n\n<p>For the LSTMs to work properly I need to pass the cell and hidden state between the nodes.<\/p>\n\n<p>This is a simplified version, which should emulate a normal LSTM:<\/p>\n\n<pre><code>def silly_lstm():  # the model-layer\n\n  input_1 = Input(shape=(1,4), name='input_1')\n  input_c = Input(shape=(2,), name='input_c')\n  input_h = Input(shape=(2,), name='input_here_happens_the_error')\n\n  out, cell, hidden = LSTM(2, return_state=True, name='lstm')\n                              (input_1, initial_state=[input_c, input_h])\n\n  return Model(inputs=[input_1, input_c, input_h],\n                  outputs=[out, cell, hidden], name ='silly_lstm')\n\ndef silly_lstm_model():  # the actual model \n\n  i1 = Input(shape=(1,4), name='i1')  # the external input\n  i2 = Input(shape=(2,), name='i2')   # initializer of the cell state\n  i3 = Input(shape=(2,), name='i3')   # initializer of the hidden state\n\n  lstm = silly_lstm()\n\n  # the layer is \"recurrently\" applied multiple times\n  o1, o2, o3 = lstm([i1, i2, i3])\n  o1, o2, o3 = lstm([i1, o2, o3])\n\n  model = Model(inputs=[i1, i2, i3], outputs=[o1, o2, o3])\n\n  return model\n\n\nmodel = silly_lstm_model()\n\ni = [np.zeros((1,1,4)), np.zeros((1,2)), np.zeros((1,2))]\nmodel.predict(i)\n<\/code><\/pre>\n\n<p>I get the following error (which goes away when I do not set the initial states in silly_lstm):<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_here_happens_an_error' with dtype float and shape [?,2]\n<\/code><\/pre>\n\n<p>The Traceback is this:<\/p>\n\n<pre><code>    Caused by op 'input_here_happens_an_error', defined at:\n  File \"c:\\program files\\python36\\lib\\site-packages\\spyder\\utils\\ipython\\start_kernel.py\", line 269, in &lt;module&gt;\n    main()\n  File \"c:\\program files\\python36\\lib\\site-packages\\spyder\\utils\\ipython\\start_kernel.py\", line 265, in main\n    kernel.start()\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\kernelapp.py\", line 486, in start\n    self.io_loop.start()\n  File \"c:\\program files\\python36\\lib\\site-packages\\tornado\\platform\\asyncio.py\", line 112, in start\n    self.asyncio_loop.run_forever()\n  File \"c:\\program files\\python36\\lib\\asyncio\\base_events.py\", line 421, in run_forever\n    self._run_once()\n  File \"c:\\program files\\python36\\lib\\asyncio\\base_events.py\", line 1431, in _run_once\n    handle._run()\n  File \"c:\\program files\\python36\\lib\\asyncio\\events.py\", line 145, in _run\n    self._callback(*self._args)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tornado\\platform\\asyncio.py\", line 102, in _handle_events\n    handler_func(fileobj, events)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tornado\\stack_context.py\", line 276, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"c:\\program files\\python36\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 450, in _handle_events\n    self._handle_recv()\n  File \"c:\\program files\\python36\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 480, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"c:\\program files\\python36\\lib\\site-packages\\zmq\\eventloop\\zmqstream.py\", line 432, in _run_callback\n    callback(*args, **kwargs)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tornado\\stack_context.py\", line 276, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 283, in dispatcher\n    return self.dispatch_shell(stream, msg)\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 233, in dispatch_shell\n    handler(stream, idents, msg)\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\kernelbase.py\", line 399, in execute_request\n    user_expressions, allow_stdin)\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\ipkernel.py\", line 208, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"c:\\program files\\python36\\lib\\site-packages\\ipykernel\\zmqshell.py\", line 537, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"c:\\program files\\python36\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2728, in run_cell\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"c:\\program files\\python36\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2850, in run_ast_nodes\n    if self.run_code(code, result):\n  File \"c:\\program files\\python36\\lib\\site-packages\\IPython\\core\\interactiveshell.py\", line 2910, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-82-dd9a4619cb90&gt;\", line 1, in &lt;module&gt;\n    model = silly_lstm_model_3()\n  File \"&lt;ipython-input-81-b0023bfb8995&gt;\", line 89, in silly_lstm_model_3\n    lstm = silly_lstm()\n  File \"&lt;ipython-input-81-b0023bfb8995&gt;\", line 5, in silly_lstm\n    input_h = Input(shape=(2,), dtype='float32', name='input_here_happens_an_error')\n  File \"c:\\program files\\python36\\lib\\site-packages\\keras\\engine\\topology.py\", line 1455, in Input\n    input_tensor=tensor)\n  File \"c:\\program files\\python36\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"c:\\program files\\python36\\lib\\site-packages\\keras\\engine\\topology.py\", line 1364, in __init__\n    name=self.name)\n  File \"c:\\program files\\python36\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 504, in placeholder\n    x = tf.placeholder(dtype, shape=shape, name=name)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1746, in placeholder\n    return gen_array_ops._placeholder(dtype=dtype, shape=shape, name=name)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 4026, in _placeholder\n    \"Placeholder\", dtype=dtype, shape=shape, name=name)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3271, in create_op\n    op_def=op_def)\n  File \"c:\\program files\\python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1650, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n<\/code><\/pre>\n","answers":[{"AnswerId":"53338955","CreationDate":"2018-11-16T13:36:43.300","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Apparently, upgrading Keras to the latest version (2.2.4) resolved the issue.<\/p>\n"}]}
+{"QuestionId":53328905,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><lstm><recurrent-neural-network>","CreationDate":"2018-11-15T22:42:39.707","AcceptedAnswerId":null,"OwnerUserId":5526180.0,"Title":"initial state in a dynamic_rnn as a placeholder","Body":"<p>I'd like to enter the tensor to the intial_state of an LSTM as a placeholder so that later on I assign it's value myself during the learning process. <\/p>\n\n<p>I wrote the following:<\/p>\n\n<pre><code>import tensorflow as tf\nclass PGNetwork:\ndef __init__(self, inpsize, name='PGNetwork'):\n    tf.reset_default_graph()\n    self.inpsize = inpsize\n\n    with tf.variable_scope(name):\n        # We create the placeholders\n        self.inputs_vec = tf.placeholder(tf.float32, [None, \n                            self.inpsize], name=\"inputs_vec\")\n        self.in_state = tf.placeholder(tf.float32, [1, 16], name=\"in_state\")\n\n        self.lstm_layer = tf.contrib.rnn.BasicLSTMCell(16,forget_bias=1)\n#             self.in_state = self.lstm_layer.zero_state(1, dtype=tf.float32) # By uncommenting this line the error is no longer there.\n        self.out_rnn, self.rnn_state = tf.nn.dynamic_rnn(self.lstm_layer, \\\n                                                      tf.expand_dims(self.inputs_vec, 1), initial_state=self.in_state)\n\n        self.output = tf.layers.dense(inputs = self.out_rnn, \n                                       kernel_initializer=tf.contrib.layers.xavier_initializer(),\n                                      units = 5, \n                                    activation=None, name=\"output\")\n\n        self.action_distribution = tf.nn.softmax(self.output, name=\"softmax\")\n\nPGNetwork = PGNetwork(8)\n<\/code><\/pre>\n\n<p>I get this error :\n<code>Tensor objects are not iterable when eager execution is not enabled. To iterate over this tensor use tf.map_fn.<\/code><\/p>\n\n<p>Which is actually there because of the <code>initial_state<\/code> in <code>tf.nn.dynamic_rnn()<\/code> since it is not accepted to be a placeholder.<\/p>\n\n<p>Is it possible to convert the placeholder to something accepted by the <code>dynamic_rnn<\/code>????<\/p>\n","answers":[]}
+{"QuestionId":53329357,"AnswerCount":1,"Tags":"<tensorflow><autodiff>","CreationDate":"2018-11-15T23:35:33.417","AcceptedAnswerId":null,"OwnerUserId":5037235.0,"Title":"Can Tensorflow take gradient on matrix 2-norm?","Body":"<p>Normally the matrix norm we took in Tensorflow is <em>Frobenius<\/em> norm which is easy to compute and easy to understand, e.g., a Bayesian view. But in many cases, it is the <em>largest singular value<\/em> matters. It is possible to optimize that in Tensorflow? It depends on whether tensorflow can take gradient with respect to matrix 2-norm. <\/p>\n","answers":[{"AnswerId":"53336474","CreationDate":"2018-11-16T10:56:56.383","ParentId":null,"OwnerUserId":"1058309","Title":null,"Body":"<p>Actually, the <a href=\"http:\/\/mathworld.wolfram.com\/SpectralNorm.html\" rel=\"nofollow noreferrer\">spectral norm<\/a> is <a href=\"https:\/\/math.stackexchange.com\/questions\/586663\/why-does-the-spectral-norm-equal-the-largest-singular-value\">equal<\/a> the largest singular value. To get to this value you can use TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/linalg\/svd\" rel=\"nofollow noreferrer\"><code>linalg.svd<\/code><\/a>.<\/p>\n"}]}
+{"QuestionId":53329558,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2018-11-16T00:02:25.207","AcceptedAnswerId":53330055.0,"OwnerUserId":2455494.0,"Title":"tensorflow cast tf.int32 to tf.string in version suitable for Google cloud ml-engine","Body":"<p>I would like to convert a tensor of type <code>tf.int32<\/code> to one of type <code>tf.string<\/code>.<\/p>\n\n<p>Thanks to <a href=\"https:\/\/stackoverflow.com\/a\/53305286\/2455494\">this answer<\/a> I know that in version 1.12 I can use <code>tf.strings.format<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.constant([1, 2, 3], dtype=tf.int32)\nx_as_string = tf.map_fn(lambda xi: tf.strings.format('{}', xi), x, dtype=tf.string)\n\nwith tf.Session() as sess:\n  res = sess.run(x_as_string)\n  print(res)\n  # [b'1' b'2' b'3']\n<\/code><\/pre>\n\n<p>But i would like to do this operation on Google cloud ml-engine which (today) only supports up to version 1.10. <\/p>\n\n<p>Is there an alternative operation I can use in an earlier version of tensorflow? <\/p>\n\n<p>Or, alternatively, is there anyway to use the new version of TensorFlow in Google cloud ml-engine?<\/p>\n","answers":[{"AnswerId":"53330055","CreationDate":"2018-11-16T01:09:59.383","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r1.10\/api_docs\/python\/tf\/as_string\" rel=\"nofollow noreferrer\"><code>tf.as_string<\/code><\/a> should work to convert ints to strings.<\/p>\n"}]}
+{"QuestionId":53329659,"AnswerCount":1,"Tags":"<python><keras><neural-network><reshape><flatten>","CreationDate":"2018-11-16T00:15:49.290","AcceptedAnswerId":null,"OwnerUserId":9808963.0,"Title":"Unflattening Layer in Keras","Body":"<p>I would like to create a simple Keras neural network that accepts an input matrix of dimension <code>(rows, columns) = (n, m)<\/code>, flattens the matrix to a dimension <code>(n*m, 1)<\/code>, sends the flattened matrix through a number of arbitrary layers, and in the final layer, once more unflattens the matrix to a dimension of <code>(n, m)<\/code> before releasing this final matrix as an output. <\/p>\n\n<p>The issue I'm having is that I haven't found any documentation for an Unflatten layer at the <a href=\"https:\/\/keras.io\" rel=\"nofollow noreferrer\">keras.io<\/a> page, and I'm wondering whether there is a reason that such a seemingly standard common use layer doesn't exist. Is there a much more natural and easy way to do what I'm proposing?<\/p>\n","answers":[{"AnswerId":"53333623","CreationDate":"2018-11-16T07:57:29.797","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\"><code>Reshape<\/code><\/a> layer for this purpose. It accepts the desired output shape as its argument and would reshape the input tensor to that shape. For example:<\/p>\n\n<pre><code>from keras.layers import Reshape\n\nrsh_inp = Reshape((n*m, 1))(inp)  # if don't want the last axis with dimension 1, you can also use Flatten layer\n\n# rsh_inp goes through a number of arbitrary layers ...\n\n# reshape back the output\nout = Reshape((n,m))(out_rsh_inp)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53329708,"AnswerCount":1,"Tags":"<python><dictionary><tensorflow><neural-network>","CreationDate":"2018-11-16T00:22:59.430","AcceptedAnswerId":53329769.0,"OwnerUserId":10660362.0,"Title":"How to display values of a dictionary?","Body":"<p>I'm trying to create a dictionary of weigths per layer for my ANN implementation.<\/p>\n\n<p>The issue is that although I create the dictionary with strings as keys and tensors as values I don't know how to display them when I call the <code>init_weight<\/code> method<\/p>\n\n<pre><code>def init_weights(topology):\n#topology: dimensions of the network\n\nfor i in range(1,len(topology)):\nparameters['W' + str(i)] = tf.Variable(tf.random_normal([topology[i-1],topology[i]]))\n<\/code><\/pre>\n\n<p>The output of the method shows the following:<\/p>\n\n<pre><code>{'W1': &lt;tf.Variable 'Variable_1:0' shape=(2, 5) dtype=float32_ref&gt;,\n 'W2': &lt;tf.Variable 'Variable_3:0' shape=(5, 5) dtype=float32_ref&gt;,\n 'W3': &lt;tf.Variable 'Variable_5:0' shape=(5, 5) dtype=float32_ref&gt;,\n 'W4': &lt;tf.Variable 'Variable_7:0' shape=(5, 10) dtype=float32_ref&gt;}\n<\/code><\/pre>\n\n<p>How can print the weights matrices?<\/p>\n","answers":[{"AnswerId":"53329769","CreationDate":"2018-11-16T00:31:33.840","ParentId":null,"OwnerUserId":"8199433","Title":null,"Body":"<p>Tensorflow is a statically typed framework (guess this is changing in 2.0). Meaning, you first build a static graph and the graph has values only when run using Tf.Session(). Now to answer your question. There are two ways to get what you want.<\/p>\n\n<ol>\n<li><p>Add <code>tf.enable_eager_execution()<\/code> at the start of your script. This creates a dynamic graph (similar to Pytorch). Your same code without any extra addition will give you what you want.<\/p><\/li>\n<li><p>Wrap everything into a <code>tf.Session()<\/code> and run it. You will get the weight matrices<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53330006,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-11-16T01:03:21.340","AcceptedAnswerId":53352707.0,"OwnerUserId":10593920.0,"Title":"CNN - Wrong prediction with multiclass classification","Body":"<p>I'm making a convolutional network to predict 3 class of images, Cats, Dogs &amp; People. I trained and trained it, but then when I pass a cat image to predict it always gives the wrong output. I tried other pictures of cats, but the result doesn't change. With people and dogs it has no problem, just with cats.<\/p>\n\n<pre><code>     cnn = Sequential()\n\n    #------------------- Convolu\u00e7\u00e3o e Pooling\n    cnn.add(Conv2D(32, (3, 3), input_shape = (64, 64, 3), activation = 'relu'))\n    cnn.add(Dropout(0.5))\n    cnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n    cnn.add(Conv2D(32, (3, 3), activation = 'relu'))\n    cnn.add(Dropout(0.5))\n    cnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n    cnn.add(Conv2D(64, (3, 3), activation = 'relu'))\n    cnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n    cnn.add(Conv2D(64, (3, 3), activation = 'relu'))\n    cnn.add(Dropout(0.5))\n    cnn.add(MaxPooling2D(pool_size = (2, 2)))\n\n    #Full connection\n    cnn.add(Flatten())\n    cnn.add(Dense(units = 128, activation = 'relu'))\n    cnn.add(Dense(units = 4, activation = 'softmax'))\n\n    # Compiling the CNN\n    cnn.compile(optimizer = OPTIMIZER, loss = 'categorical_crossentropy', metrics = ['accuracy'])\n\n    filepath=\"LPT-{epoch:02d}-{loss:.4f}.h5\"\n    checkpoint = ModelCheckpoint(filepath, monitor='loss', verbose=1, save_best_only=True, mode='min')\n    callbacks_list = [checkpoint]\n<\/code><\/pre>\n\n<p>12000 train images - 3000 test images<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   horizontal_flip = True)\n\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ntraining_set = train_datagen.flow_from_directory('data\/train',\n                                                 target_size = tgt_size,\n                                                 batch_size = batch_size,\n                                                 class_mode = 'categorical')\n\ntest_set = test_datagen.flow_from_directory('data\/test',\n                                            target_size = tgt_size,\n                                            batch_size = batch_size,\n                                            class_mode = 'categorical')\n\ncnn.fit_generator(training_set,\n                  #steps_per_epoch = 12000,  \n                  steps_per_epoch = nb_train_samples \/\/ batch_size,                                                   \n                  epochs = EPOCHS,\n                  verbose = VERBOSE,\n                  validation_data = test_set,\n                  validation_steps = nb_validation_samples \/\/ batch_size,\n                  callbacks = callbacks_list)\n<\/code><\/pre>\n\n<p>My best training result:<\/p>\n\n<p><code>loss: 0.6410 - acc: 0.7289 - val_loss: 0.6308 - val_acc: 0.7293<\/code><\/p>\n\n<p>Class indices:<\/p>\n\n<pre><code>{'.ipynb_checkpoints': 0, 'cats': 1, 'dogs':2, 'person':3}\n<\/code><\/pre>\n\n<p>(I can't remove that ipynb folder)<\/p>\n\n<p>Prediction:<\/p>\n\n<pre><code>pred1 = 'single_prediction\/ct.jpg'\npred2 = 'single_prediction\/ps.jpg'\npred3 = 'data\/single_prediction\/dg.jpg'\n\ntest_img = image.load_img(pred1, target_size = tgt_size)\ntest_img = image.img_to_array(test_img)\ntest_img = np.expand_dims(test_img, axis = 0)\npred = new_model.predict(test_img)\nprint(pred)\n\nif pred[0][1] == 1:\n        print('It is a cat!')\nelif pred[0][2] == 1:\n    print('It is a dog!')\nelif pred[0][3] == 1:\n    print('It is a Person!')\n<\/code><\/pre>\n\n<p>And the output for a cat image:<\/p>\n\n<pre><code>[[0.000000e+00 0.000000e+00 8.265931e-34 1.000000e+00]]\n<\/code><\/pre>\n\n<p>I already tried:\n Change the number of layers (added and removed), increase the epochs, decrease the batch... I also tried using np.argmax(). Can someone please give me a light here?<\/p>\n\n<p><strong>UPDATE<\/strong>: I removed the jupyter notebook's hidden folder with the command shutil.rmtree() and trained for about 40 epochs until it stopped improving. At last, I rescaled the prediction image and got it right.<\/p>\n\n<pre><code>test_img = image.img_to_array(test_img)\/255\n<\/code><\/pre>\n\n<p>Thanks for all the help!<\/p>\n","answers":[{"AnswerId":"53352707","CreationDate":"2018-11-17T15:35:30.257","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>The problem is with the ipynb checkpoints folder. This is a hidden folder. You need to delete it first. Then change your output dense layer to have 3 units (classes). \nChange this<\/p>\n\n<pre><code>cnn.add(Dense(units = 4, activation = 'softmax')) \n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>cnn.add(Dense(units = 3, activation = 'softmax')) \n<\/code><\/pre>\n"}]}
+{"QuestionId":53330884,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-16T03:11:04.527","AcceptedAnswerId":null,"OwnerUserId":9808963.0,"Title":"Keras Input Layer Misinterpreting Input Shape","Body":"<p>I am trying to make a very simple functional neural network in Keras. I input a vector of shape <code>(270000,)<\/code> to the network, and have entered this as the shape to accept in the input layer, but I receive the error shown below. Given that the shape printed for the input specified to be at fault, is in fact <code>(270000,)<\/code>, I don't know why I am receiving this error.<\/p>\n\n<p><strong>Model Function<\/strong><\/p>\n\n<pre><code>def spectrify(A1, y1, simData, aOrigShape):\n    print(\"A1: \", np.shape(A1))\n    print(\"y1: \", np.shape(y1))\n    print(\"simData\", np.shape(simData))\n    print(\"aOrigShape:\", aOrigShape)\n    dataIn = Input(shape=np.shape(A1))\n    dataOut = Dense(np.shape(A1)[0])(dataIn)\n    outShaper = Reshape((aOrigShape))(dataOut)\n    model = Model(inputs = dataIn, outputs = outShaper)\n    model.compile(optimizer = 'rmsprop',\n                 loss = 'categorical_crossentropy',\n                 metrics = ['accuracy'])\n    model.fit(A1, simData)\n    return model\n<\/code><\/pre>\n\n<p><strong>Execution<\/strong><\/p>\n\n<p>Running the function above prints the shapes and raises the following error:<\/p>\n\n<pre><code>A1:  (270000,)\ny1:  (200, 540)\nsimData (200, 400)\naOrigShape: (500, 540)\n\n...\n&lt;ipython-input-130-88e6c1dfc1c9&gt; in spectrify(A1, y1, simData, aOrigShape)\n     12                  loss = 'categorical_crossentropy',\n     13                  metrics = ['accuracy'])\n---&gt; 14     model.fit(A1, simData)\n     15     return model\n...\n\nValueError: Error when checking input: expected input_50 to have shape (270000,) but got array with shape (1,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53335982","CreationDate":"2018-11-16T10:29:54.660","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>shape<\/code> argument refers to the shape of <strong>one single sample<\/strong> in the training data. So if you have 270000 training samples of shape <code>(1,)<\/code>, then the <code>shape<\/code> argument must be set to <code>(1,)<\/code>. Otherwise, which is unlikely but possible, if you have one sample of shape <code>(270000,)<\/code> then you need the shape argument must be set to <code>(270000,)<\/code> and <code>A<\/code> must have a shape of <code>(1, 270000)<\/code>, which means one sample of shape <code>(270000,)<\/code>, and not <code>(270000,)<\/code> which means 270000 samples of shape <code>(1,)<\/code>.<\/p>\n\n<p>Generally, if <code>X_train<\/code> is the array which contains your training data, then it's a good practice to use <code>X_train.shape[1:]<\/code> (i.e. the shape of each sample) as the input shape, like this:<\/p>\n\n<pre><code>Input(shape=X_train.shape[1:])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53331211,"AnswerCount":1,"Tags":"<tensorflow><bazel><quantization>","CreationDate":"2018-11-16T03:56:05.107","AcceptedAnswerId":53346729.0,"OwnerUserId":5218450.0,"Title":"Tensorflow bazel quantization build error","Body":"<p>I am trying to build tensorflow tools package with bazel 0.18.0<\/p>\n\n<p>following steps are ok<\/p>\n\n<pre><code>git clone https:\/\/github.com\/tensorflow\/tensoflow\nbazel build --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\nbazel build --config=cuda \/\/tensorflow\/examples\/label_image:label_image\n<\/code><\/pre>\n\n<p>until trying to run this command line <\/p>\n\n<pre><code>bazel build --config=cuda \/\/tensorflow\/contrib\/quantization:quantize_graph\n<\/code><\/pre>\n\n<p>it show error, so should i give something else then quantize_graph? and what i can use or find?<\/p>\n\n<p>root@24663fb1018d:\/srv\/wu\/tensorflow-src\/tensorflow# bazel build --config=cuda \/\/tensorflow\/contrib\/quantization:quantize_graph\nWARNING: Duplicate rc file: \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc is read multiple times, most recently imported from \/srv\/wu\/tensorflow-src\/tensorflow\/.bazelrc\nWARNING: Processed legacy workspace file \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc. This file will not be processedin the next release of Bazel. Please read <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319<\/a> for further information, including how to upgrade.\nStarting local Bazel server and connecting to it...\nWARNING: The following configs were expanded more than once: [cuda]. For repeatable flags, repeats are counted twice and may lead to unexpected behavior.\nERROR: Skipping '\/\/tensorflow\/contrib\/quantization:quantize_graph': no such target '\/\/tensorflow\/contrib\/quantization:quantize_graph': target 'quantize_graph' not declared in package 'tensorflow\/contrib\/quantization' defined by \/srv\/wu\/tensorflow-src\/tensorflow\/tensorflow\/contrib\/quantization\/BUILD\nWARNING: Target pattern parsing failed.\nERROR: <strong>no such target '\/\/tensorflow\/contrib\/quantization:quantize_graph': target 'quantize_graph' not declared in package 'tensorflow\/contrib\/quantization'<\/strong> defined by \/srv\/wu\/tensorflow-src\/tensorflow\/tensorflow\/contrib\/quantization\/BUILD\nINFO: Elapsed time: 1.195s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (1 packages loaded)<\/p>\n\n<p>and then i tried with tools path, no luck<\/p>\n\n<pre><code>bazel build --config=cuda \/\/tensorflow\/tools\/quantization:quantize_graph\n<\/code><\/pre>\n\n<p>WARNING: Duplicate rc file: \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc is read multiple times, most recently imported from \/srv\/wu\/tensorflow-src\/tensorflow\/.bazelrcWARNING: Processed legacy workspace file \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc. This file will not be processed in the next release of Bazel. Please read <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319<\/a> for further information, including how to upgrade.WARNING: The following configs were expanded more than once: [cuda]. For repeatable flags, repeats are counted twice and may lead to unexpected behavior.ERROR: Skipping '\/\/tensorflow\/tools\/quantization:quantize_graph': no such package'tensorflow\/tools\/quantization': BUILD file not found on package pathWARNING: Target pattern parsing failed.\nERROR: no such package 'tensorflow\/tools\/quantization': BUILD file not found on package path\nINFO: Elapsed time: 0.506s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)<\/p>\n\n<p>also the toco is not working<\/p>\n\n<pre><code>bazel build --config=cuda tensorflow\/contrib\/lite\/toco:toco\n<\/code><\/pre>\n\n<p>WARNING: Duplicate rc file: \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc is read multiple times, most recently imported from \/srv\/wu\/tensorflow-src\/tensorflow\/.bazelrc\nWARNING: Processed legacy workspace file \/srv\/wu\/tensorflow-src\/tensorflow\/tools\/bazel.rc. This file will not be processed in the next release of Bazel. Please read <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bazelbuild\/bazel\/issues\/6319<\/a> for further information, including how to upgrade.WARNING: The following configs were expanded more than once: [cuda]. For repeatable flags, repeats are counted twice and may lead to unexpected behavior.ERROR: Skipping 'tensorflow\/contrib\/lite\/toco:toco': no such package 'tensorflow\/contrib\/lite\/toco': BUILD file not found on package path\nWARNING: Target pattern parsing failed.ERROR: no such package 'tensorflow\/contrib\/lite\/toco': BUILD file not found on package pathINFO: Elapsed time: 0.500s\nINFO: 0 processes.FAILED: Build did NOT complete successfully (0 packages loaded)<\/p>\n","answers":[{"AnswerId":"53346729","CreationDate":"2018-11-16T23:32:47.927","ParentId":null,"OwnerUserId":"521209","Title":null,"Body":"<p>To verify where targets are, look into the <code>BUILD<\/code> file in the package directory. <\/p>\n\n<p>The <code>quantize_graph<\/code> target has been <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/6a7bd8b8b34d541eba788b670c2cc7f489a1ae66\/tensorflow\/contrib\/quantize\/BUILD#L228\" rel=\"nofollow noreferrer\">moved to the <code>\/\/tensorflow\/contrib\/quantize<\/code> package<\/a>. This should work:<\/p>\n\n<pre><code>$ bazel build --config=cuda \/\/tensorflow\/contrib\/quantize:quantize_graph\n<\/code><\/pre>\n\n<p>The <code>toco<\/code> target <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1a90c933a806812228a7b290cf670996cf1f674d\/tensorflow\/lite\/toco\/BUILD#L399\" rel=\"nofollow noreferrer\">has been moved<\/a> from <code>\/\/tensorflow\/contrib\/lite\/toco<\/code> to <code>\/\/tensorflow\/lite\/toco<\/code>. Like <code>quantize_graph<\/code>, this should work:<\/p>\n\n<pre><code>$ bazel build --config=cuda \/\/tensorflow\/lite\/toco:toco\n<\/code><\/pre>\n"}]}
+{"QuestionId":53331247,"AnswerCount":1,"Tags":"<pytorch><tensor>","CreationDate":"2018-11-16T04:01:02.313","AcceptedAnswerId":53332659.0,"OwnerUserId":9066532.0,"Title":"Pytorch 0.4.0: There are three ways to create tensors on CUDA device. Is there some difference between them?","Body":"<p>I failed in the third way. <code>t3<\/code> is still on CPU. No idea why.<\/p>\n\n<pre><code>a = np.random.randn(1, 1, 2, 3)\n\nt1 = torch.tensor(a)\nt1 = t3.to(torch.device('cuda'))\n\nt2 = torch.tensor(a)\nt2 = t2.cuda() \n\nt3 = torch.tensor(a, device=torch.device('cuda'))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53332659","CreationDate":"2018-11-16T06:37:23.817","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>All three methods worked for me. <\/p>\n\n<p>In 1 and 2, you create a tensor on CPU and then move it to GPU when you use <code>.to(device)<\/code> or <code>.cuda()<\/code>. They are the same here.<\/p>\n\n<p>However, when you use <code>.to(device)<\/code> method you can explicitly tell torch to move to specific GPU by setting <code>device=torch.device(\"cuda:&lt;id&gt;\")<\/code>. with <code>.cuda()<\/code> you have to do <code>.cuda(&lt;id&gt;)<\/code> to move to some particular GPU. <\/p>\n\n<hr>\n\n<p><em>Why do these two methods exist then?<\/em><\/p>\n\n<p><code>.to(device)<\/code> was introduced in 0.4 because it is easier to declare <code>device<\/code> variable at top of the code as <\/p>\n\n<p><code>device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")<\/code><\/p>\n\n<p>and use <code>.to(device)<\/code> everywhere. This makes it quite easy to switch from CPU to GPU and vice-versa<\/p>\n\n<p>Before this, we had to use <code>.cuda()<\/code> and your code will have <code>if<\/code> check for <code>cuda.is_available()<\/code> everywhere which made it cumbersome to switch between GPU\/CPU.<\/p>\n\n<hr>\n\n<p>The third method doesn't create a tensor on the CPU and directly copies data to GPU, which is more efficient. <\/p>\n"}]}
+{"QuestionId":53331502,"AnswerCount":0,"Tags":"<python><keras><deep-learning><artificial-intelligence><reinforcement-learning>","CreationDate":"2018-11-16T04:35:59.230","AcceptedAnswerId":null,"OwnerUserId":5932428.0,"Title":"Deep Q Pong Learning Failing","Body":"<p>I recently started researching Deep Q learning and have run into issues. Following along with articles on the web (including this one: <a href=\"https:\/\/keon.io\/deep-q-learning\/\" rel=\"nofollow noreferrer\">https:\/\/keon.io\/deep-q-learning\/<\/a>), I attempted to modify the cart pole example to work with the Pong game. However, even though my solution seems to follow other's (minimize the state, keep differences, etc), the ai literally never gets better (upwards of 800 games one time).<\/p>\n\n<p>Base Code<\/p>\n\n<pre><code>import numpy as np\nfrom ddqn import DQNAgent\nimport gym\nfrom PIL import Image\nenv = gym.make('Pong-v0')\nstate_size = 88*80\naction_size = env.action_space.n\nagent = DQNAgent(state_size, action_size)\ndone = False\nbatch_size = 32\nfor e in range(EPISODES):\n    state = env.reset()\n    prev_state = state\n    total_reward = 0\n    for time in range(500000):\n        env.render()\n        current_state_diff = minimize_state(state) - minimize_state(prev_state)\n        action = agent.act(current_state_diff)\n        next_state, reward, done, _ = env.step(action)\n        total_reward += reward\n        next_state_diff = minimize_state(next_state) -         minimize_state(state)\n        agent.remember(current_state_diff, action, reward, next_state_diff, done)\n        state = next_state\n        if done:\n            agent.update_target_model()\n            print(\"episode: {}\/{}, score: {}, e: {:.2}\"\n                  .format(e, EPISODES, total_reward, agent.epsilon))\n            break\n        if len(agent.memory) &gt; batch_size:\n            agent.replay(batch_size)\n    if e % 10 == 0:\n        agent.save(\".\/save\/pong-v0-ddqn.h5\")\n<\/code><\/pre>\n\n<p>Where the minimization function is<\/p>\n\n<pre><code>def minimize_state(state):\n    img = Image.fromarray(state)\n    state = np.array(img.convert(\"L\"))\n    state = state[34:]\n    state = state[::2,::2]\n    return np.reshape(state, [1, state_size])\n<\/code><\/pre>\n\n<p>And the DQNAgent is from: <a href=\"https:\/\/github.com\/keon\/deep-q-learning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keon\/deep-q-learning<\/a><\/p>\n\n<p>What's going wrong?<\/p>\n","answers":[]}
+{"QuestionId":53331540,"AnswerCount":2,"Tags":"<python><pytorch><gradient-descent>","CreationDate":"2018-11-16T04:40:35.513","AcceptedAnswerId":53549374.0,"OwnerUserId":2530674.0,"Title":"Accumulating Gradients","Body":"<p>I want to accumulate the gradients before I do a backward pass. So wondering what the right way of doing it is. According to <a href=\"https:\/\/medium.com\/huggingface\/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255\" rel=\"nofollow noreferrer\">this article<\/a>\nit's:<\/p>\n\n<pre><code>model.zero_grad()                                   # Reset gradients tensors\nfor i, (inputs, labels) in enumerate(training_set):\n    predictions = model(inputs)                     # Forward pass\n    loss = loss_function(predictions, labels)       # Compute loss function\n    loss = loss \/ accumulation_steps                # Normalize our loss (if averaged)\n    loss.backward()                                 # Backward pass\n    if (i+1) % accumulation_steps == 0:             # Wait for several backward steps\n        optimizer.step()                            # Now we can do an optimizer step\n        model.zero_grad()\n<\/code><\/pre>\n\n<p>whereas I expected it to be:<\/p>\n\n<pre><code>model.zero_grad()                                   # Reset gradients tensors\nloss = 0\nfor i, (inputs, labels) in enumerate(training_set):\n    predictions = model(inputs)                     # Forward pass\n    loss += loss_function(predictions, labels)       # Compute loss function                              \n    if (i+1) % accumulation_steps == 0:             # Wait for several backward steps\n        loss = loss \/ accumulation_steps            # Normalize our loss (if averaged)\n        loss.backward()                             # Backward pass\n        optimizer.step()                            # Now we can do an optimizer step\n        model.zero_grad()     \n        loss = 0  \n<\/code><\/pre>\n\n<p>where I accumulate the loss and then divide by the accumulation steps to average it.<\/p>\n\n<p>Secondary question, if I am right, would you expect my method to be quicker considering I only do the backward pass every accumulation steps?       <\/p>\n","answers":[{"AnswerId":"53549374","CreationDate":"2018-11-30T00:03:55.453","ParentId":null,"OwnerUserId":"2530674","Title":null,"Body":"<p>So according to the answer <a href=\"https:\/\/discuss.pytorch.org\/t\/why-do-we-need-to-set-the-gradients-manually-to-zero-in-pytorch\/4903\/20?u=alband\" rel=\"nofollow noreferrer\">here<\/a>, the first method is memory efficient. The amount of work required is more or less the same in both methods.<\/p>\n\n<p>The second method keeps accumulating the graph, so would require <code>accumulation_steps<\/code> times more memory. The first method calculates the gradients straight away (and simply adds gradients) so requires less memory.<\/p>\n"},{"AnswerId":"53335060","CreationDate":"2018-11-16T09:38:11.253","ParentId":null,"OwnerUserId":"7191510","Title":null,"Body":"<p>The backward pass <code>loss.backward()<\/code> is the operation that actually computes the <em>gradients<\/em>.<\/p>\n\n<p>If you only do the forward pass (<code>predictions = model(inputs)<\/code>) no gradients will be computed and no accumulation is thus possible.<\/p>\n"}]}
+{"QuestionId":53331928,"AnswerCount":0,"Tags":"<apache-spark><tensorflow><pyspark>","CreationDate":"2018-11-16T05:25:40.103","AcceptedAnswerId":null,"OwnerUserId":8759138.0,"Title":"Run Tensorflow and Pyspark Together in same program","Body":"<p>When I run a program with tensorflow an pyspark together with <code>spark-submit<\/code> <code>sample.py<\/code> it says no module named <code>tensorflow<\/code> but when I run activate tensorflow and run with python <code>sample.py<\/code> it says no module named pyspark.<\/p>\n\n<pre><code>import tensorflow as tf\n\nimport pyspark # only run after findspark.init()\n\nfrom pyspark.sql import SparkSession\n\nspark = SparkSession.builder.getOrCreate()\n\ndf = spark.sql('''select 'spark' as hello ''')\n\ndf.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53332070,"AnswerCount":0,"Tags":"<ubuntu-16.04><theano><nvcc>","CreationDate":"2018-11-16T05:41:17.890","AcceptedAnswerId":null,"OwnerUserId":4399461.0,"Title":"nvcc compiler not found on $PATH","Body":"<p>I am getting the following error with theano:<\/p>\n\n<blockquote>\n  <p>ERROR (theano.sandbox.cuda): nvcc compiler not found on $PATH. Check\n  your nvcc installation and try again.<\/p>\n<\/blockquote>\n\n<p>I've tried every solution given by previous questions. Running <code>nvcc -V<\/code> returns the following:<\/p>\n\n<pre><code>nvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2017 NVIDIA Corporation\nBuilt on Fri_Nov__3_21:07:56_CDT_2017\nCuda compilation tools, release 9.1, V9.1.85\n<\/code><\/pre>\n\n<p>And <code>which nvcc<\/code> returns the following:<\/p>\n\n<pre><code>\/usr\/local\/cuda-9.1\/bin\/\/nvcc\n<\/code><\/pre>\n\n<p>I've also confirmed I have an nvcc file in <code>\/usr\/local\/cuda-9.1\/bin<\/code>.<\/p>\n\n<p>I've ran the following commands to ensure nvcc can be found:<\/p>\n\n<pre><code>export PATH=\/usr\/local\/cuda-9.1\/bin\/:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-9.1\/lib64:$LD_LIBRARY_PATH\nexec bash\n<\/code><\/pre>\n\n<p>I've confirmed the nvcc location is definitely in $PATH as running <code>echo $PATH<\/code> returns the following:<\/p>\n\n<pre><code>\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/cuda-9.1\/bin\/:\/usr\/local\/cuda-9.1\/bin\/:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/cuda\/bin\/:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/sbin:\/usr\/local\/bin:\/usr\/sbin:\/usr\/bin:\/sbin:\/bin:\/usr\/games:\/usr\/local\/games:\/snap\/bin:\/usr\/lib\/jvm\/java-9-oracle\/bin:\/usr\/lib\/jvm\/java-9-oracle\/db\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin\n<\/code><\/pre>\n\n<p>Also if I open a python console and run <code>import os; print(os.environ[\"PATH\"])<\/code> I get the following (same thing):<\/p>\n\n<pre><code>\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/cuda-9.1\/bin\/:\/usr\/local\/cuda-9.1\/bin\/:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/cuda\/bin\/:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/bin:\/home\/paperspace\/anaconda3\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/home\/paperspace\/src\/torch\/install\/bin:\/usr\/local\/sbin:\/usr\/local\/bin:\/usr\/sbin:\/usr\/bin:\/sbin:\/bin:\/usr\/games:\/usr\/local\/games:\/snap\/bin:\/usr\/lib\/jvm\/java-9-oracle\/bin:\/usr\/lib\/jvm\/java-9-oracle\/db\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/cuda\/bin\n<\/code><\/pre>\n\n<p>My .theanorc file in \/home\/paperspace is as follows:<\/p>\n\n<pre><code>[global]\ndevice = cuda\nfloatX = float32\n[cuda] \nroot = \/usr\/local\/cuda-9.1\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53332122,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-16T05:47:55.533","AcceptedAnswerId":null,"OwnerUserId":9634525.0,"Title":"Pretrained model for deep_speech2 tensorflow","Body":"<p>I found an implementation of the DeepSpeech2 model here: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/deep_speech#deepspeech2-model\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/deep_speech#deepspeech2-model<\/a><\/p>\n\n<p>This implementation has only to train and evaluate and no pre-trained models available. So I trained the model with a minimal dataset myself and tried to run inference but I didn't get meaningful results. So here my question is <\/p>\n\n<ol>\n<li>Does this implementation work well if I train with a large amount of dataset?<\/li>\n<li>Is there a pre-trained model available for this?<\/li>\n<\/ol>\n\n<p>Thanks in advance<\/p>\n","answers":[]}
+{"QuestionId":53332316,"AnswerCount":0,"Tags":"<python><keras><keras-layer>","CreationDate":"2018-11-16T06:06:34.407","AcceptedAnswerId":null,"OwnerUserId":6152018.0,"Title":"I write my own keras layer, but i when build it ,the input shape is None","Body":"<p>The input code: valuelist contains six tensors which shape is (None,1,128), so the allvalue should be (None,6,128), and query shape is (None,128). <\/p>\n\n<pre><code>allvalue = Concatenate(axis=-2,name=\"concatenatet\")(valuelist)\nprint(\"allvalue\",type(allvalue),allvalue.shape)\nval2 = MyAttention.Attention3(name=\"attalpall\")([query,allvalue,allvalue])\n<\/code><\/pre>\n\n<p>The allvalue print result is:<\/p>\n\n<pre><code>allvalue &lt;class 'tensorflow.python.framework.ops.Tensor'&gt; (?, 6, 128)\n<\/code><\/pre>\n\n<p>the Attention3 layer build code:<\/p>\n\n<pre><code>def build(self, input_shape):\n    print(input_shape)\n    self.WQ = self.add_weight(name='WQ', \n                              shape=(input_shape[0][-1], input_shape[0][-1]),\n                              initializer='glorot_uniform',\n                              trainable=True)\n    self.WK = self.add_weight(name='WK', \n                              shape=(input_shape[1][-1], input_shape[1][-1]),\n                              initializer='glorot_uniform',\n                              trainable=True)\n    self.WV = self.add_weight(name='WV', \n                              shape=(input_shape[2][-1], input_shape[2][-1]),\n                              initializer='glorot_uniform',\n                              trainable=True)\n    super(Attention3, self).build(input_shape)\n<\/code><\/pre>\n\n<p>The input_shape print result is\uff1a<\/p>\n\n<pre><code>[(None, 128), None, None]\n<\/code><\/pre>\n\n<p>So I get this error when I build my model.<\/p>\n\n<pre><code>shape=(input_shape[1][-1], input_shape[1][-1]),\n\nTypeError: 'NoneType' object is not subscriptable\n<\/code><\/pre>\n\n<p>It seems query's shape is right, but I don't know why the allvalue becomes None<\/p>\n","answers":[]}
+{"QuestionId":53332547,"AnswerCount":0,"Tags":"<pytorch><cudnn>","CreationDate":"2018-11-16T06:27:10.317","AcceptedAnswerId":null,"OwnerUserId":10661310.0,"Title":"pytorch project, CUDNN_STATUS_INTERNAL_ERROR:an illegal memory access wa encountered","Body":"<p>I get batch data from <code>DataLoader<\/code>. And the length of dataset is not divisible by batch size. All batches work well expect the last one. If I set \u201cdrop_last = True\u201d when instantiate <code>DataLoader<\/code> instance, the error won\u2019t appear.<\/p>\n\n<p>My dataset length is 11522, batch size is 32. So the length of last batch is 2. And I use two gpus.<\/p>\n\n<p>I google for an hour, someone said <code>rm -r ~\/.nv<\/code> and restart computer. But my code runs on server , I can\u2019t restart it. Also someone said to set <code>cudnn.benchmark = False<\/code>. But the same statement on other project works well.<\/p>\n\n<p>Why encounter the error? It upset me and I dont\u2019 wanna drop the last two data actually.<\/p>\n\n<p>My pytorch version is 0.3.1. And I did use <code>DataParallel<\/code>.<\/p>\n\n<p>I print shape of my data as shown below<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/6YEBX.jpg\" alt=\"enter image description here\"><\/p>\n","answers":[]}
+{"QuestionId":53332635,"AnswerCount":0,"Tags":"<python><pip><pytorch>","CreationDate":"2018-11-16T06:34:54.743","AcceptedAnswerId":null,"OwnerUserId":6492094.0,"Title":"Installing pytorch on python 3.7 installed ubuntu gives an error","Body":"<p>Hi all am new to pytorch and trying to run a model based on pytorch.My workstation has python 3.7 installed.So as per pytorch official site when I try to install python with below command<\/p>\n\n<p><code>pip3 install http:\/\/download.pytorch.org\/whl\/cpu\/torch-0.4.1.post2-cp37-cp37m-linux_x86_64.whl<\/code><\/p>\n\n<p>It gives error like below<\/p>\n\n<pre><code>torch-0.4.1.post2-cp37-cp37m-linux_x86_64.whl is not a supported wheel on this platform.\n<\/code><\/pre>\n\n<p>Can any one help me with this.<\/p>\n","answers":[]}
+{"QuestionId":53332663,"AnswerCount":2,"Tags":"<python><deep-learning><pytorch><normalize><torchvision>","CreationDate":"2018-11-16T06:37:49.840","AcceptedAnswerId":53334458.0,"OwnerUserId":5845276.0,"Title":"Torchvision 0.2.1 transforms.Normalize does not work as expected","Body":"<p>I am trying a new code using Pytorch. In this code, to load the dataset (CIFAR10), I am using torchvision's datasets. I define two transform functions ToTensor() and Normalize(). After normalize I expect the data in the dataset should be between 0 and 1. But the max value is still 255. I also inserted a print statement inside the '__call__' function of Normalize class in transforms.py (Lib\\site-packages\\torchvision\\transforms\\transforms.py). This print is not printed while running the code too. Not sure what is happening. Every page I visited in the internet, mentions the usage almost the same way as I do. For example some sites I visited\n<a href=\"https:\/\/github.com\/adventuresinML\/adventures-in-ml-code\/blob\/master\/pytorch_nn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/adventuresinML\/adventures-in-ml-code\/blob\/master\/pytorch_nn.py<\/a>\n<a href=\"https:\/\/github.com\/pytorch\/tutorials\/blob\/master\/beginner_source\/blitz\/cifar10_tutorial.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/tutorials\/blob\/master\/beginner_source\/blitz\/cifar10_tutorial.py<\/a><\/p>\n\n<p>My code is given below. This reads the dataset with and without Normalize, then prints some stats. The min and max printed is an indicator of whether the data is normalized or not.<\/p>\n\n<pre><code>import torchvision as tv\nimport numpy as np\n\ndataDir = 'D:\\\\general\\\\ML_DL\\\\datasets\\\\CIFAR'\n\ntrainTransform  = tv.transforms.Compose([tv.transforms.ToTensor()])\ntrainSet        = tv.datasets.CIFAR10(dataDir, train=True, download=False, transform=trainTransform)\nprint (trainSet.train_data.mean(axis=(0,1,2))\/255)\nprint (trainSet.train_data.min())\nprint (trainSet.train_data.max())\nprint (trainSet.train_data.shape)\n\ntrainTransform  = tv.transforms.Compose([tv.transforms.ToTensor(), tv.transforms.Normalize((0.4914, 0.4822, 0.4466), (0.247, 0.243, 0.261))])\ntrainSet        = tv.datasets.CIFAR10(dataDir, train=True, download=False, transform=trainTransform)\nprint (trainSet.train_data.mean(axis=(0,1,2))\/255)\nprint (trainSet.train_data.min())\nprint (trainSet.train_data.max())\nprint (trainSet.train_data.shape)\n<\/code><\/pre>\n\n<p>The output looks like,<\/p>\n\n<pre><code>[ 0.49139968  0.48215841  0.44653091]\n0\n255\n(50000, 32, 32, 3)\n[ 0.49139968  0.48215841  0.44653091]\n0\n255\n(50000, 32, 32, 3)\n<\/code><\/pre>\n\n<p>Please help me understand this better. As most of the functions I tried, ends up with similar results - for example Grayscale, CenterCrop too.<\/p>\n","answers":[{"AnswerId":"53334458","CreationDate":"2018-11-16T08:59:12.107","ParentId":null,"OwnerUserId":"7179564","Title":null,"Body":"<p>So, in the code you have laid out a plan that how you want to process your data. You have created a data pipeline through which your data will flow and multiple transforms will be applied. <\/p>\n\n<p>However, You forgot to call <code>torch.utils.data.DataLoader<\/code>. Until this is called, transformations on your data won't be applied. You can read more about it <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Now when we add the above to your code like the following -<\/p>\n\n<pre><code>trainTransform  = tv.transforms.Compose([tv.transforms.ToTensor(), \n                  tv.transforms.Normalize((0.4914, 0.4822, 0.4466), (0.247, 0.243, 0.261))])\n\ntrainSet = tv.datasets.CIFAR10(root=dataDir, train=True,\n                                    download=False, transform=trainTransform)\n\ndataloader = torch.utils.data.DataLoader(trainset, batch_size=32, shuffle=False, num_workers=4)\n<\/code><\/pre>\n\n<p>and printed images like the following - <\/p>\n\n<pre><code>images, labels = iter(dataloader).next()\nprint images\nprint images.max()\nprint images.min()\n<\/code><\/pre>\n\n<p>We get <code>Tensors<\/code> having transformations we have applied.<\/p>\n\n<p>A small snippet of the output<\/p>\n\n<pre><code>[[ 1.8649,  1.8198,  1.8348,  ...,  0.3924,  0.3774,  0.2572],\n      [ 1.9701,  1.9550,  1.9851,  ...,  0.7230,  0.6929,  0.6629],\n      [ 2.0001,  1.9550,  2.0001,  ...,  0.7831,  0.7530,  0.7079],\n      ...,\n      [-0.8096, -1.0049, -1.0350,  ..., -1.3355, -1.3655, -1.4256],\n      [-0.7796, -0.8697, -0.9749,  ..., -1.2754, -1.4557, -1.5609],\n      [-0.7645, -0.7946, -0.9298,  ..., -1.4106, -1.5308, -1.5909]]]])\ntensor(2.1309)\ntensor(-1.9895) \n<\/code><\/pre>\n\n<p>Secondly, <code>transforms.Normalize(mean,std)<\/code> applies <code>input[channel] = (input[channel] - mean[channel]) \/ std[channel]<\/code> so according to the mean and standard deviation we are providing we can't get values after transformation in the range of <code>(0,1)<\/code>. In case you want values between <code>(-1,1)<\/code> you can use the following -<\/p>\n\n<pre><code>trainTransform  = tv.transforms.Compose([tv.transforms.ToTensor(), \n                  tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n<\/code><\/pre>\n\n<p>I hope it helps! :)<\/p>\n"},{"AnswerId":"53337082","CreationDate":"2018-11-16T11:36:24.957","ParentId":null,"OwnerUserId":"5845276","Title":null,"Body":"<p>Looks like while reading without Normalization and converting into tensors itself, they are automatically normalized in 0 to 1 range. When we apply Normalization, it applies the formula you mentioned on this data ranging 0 to 1. Below is the modified working code with some print statements showing when the '__call__' function inside Normalize class is called, as well as showing how the values are normalized.The first value is 0.2314. Normalizing with 0.5 makes it (0.2314-0.5)\/0.5 = -0.5372. The first print and the second print of tensor value shows this.<\/p>\n\n<p>Code<\/p>\n\n<pre><code>import torchvision as tv\nimport numpy as np\nimport torch.utils.data as data\n\ndataDir         = 'D:\\\\general\\\\ML_DL\\\\datasets\\\\CIFAR'\n\ntrainTransform  = tv.transforms.Compose([tv.transforms.ToTensor()])\n\ntrainSet        = tv.datasets.CIFAR10(dataDir, train=True, download=False, transform=trainTransform)\nprint ('Approach1 Step1 done')\ndataloader      = data.DataLoader(trainSet, batch_size=1, shuffle=False, num_workers=0)\nprint ('Approach1 Step2 done')\nimages, labels  = iter(dataloader).next()\nprint ('Approach1 Step3 done')\nprint (images[0,0])\nprint (images.max())\nprint (images.min())\nprint (images.mean())\n\n#trainTransform = tv.transforms.Compose([tv.transforms.ToTensor(), tv.transforms.Normalize((0.4914, 0.4822, 0.4466), (0.247, 0.243, 0.261))])\ntrainTransform  = tv.transforms.Compose([tv.transforms.ToTensor(), tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\ntrainSet        = tv.datasets.CIFAR10(dataDir, train=True, download=False, transform=trainTransform)\nprint ('Approach2 Step1 done')\ndataloader      = data.DataLoader(trainSet, batch_size=1, shuffle=False, num_workers=0)\nprint ('Approach2 Step2 done')\nimages, labels  = iter(dataloader).next()\nprint ('Approach2 Step3 done')\nprint (images[0,0])\nprint (images.max())\nprint (images.min())\nprint (images.mean())\n<\/code><\/pre>\n\n<p>And the output for the above code is<\/p>\n\n<pre><code>Approach1 Step1 done\nApproach1 Step2 done\nApproach1 Step3 done\ntensor([[0.2314, 0.1686, 0.1961,  ..., 0.6196, 0.5961, 0.5804],\n        [0.0627, 0.0000, 0.0706,  ..., 0.4824, 0.4667, 0.4784],\n        [0.0980, 0.0627, 0.1922,  ..., 0.4627, 0.4706, 0.4275],\n        ...,\n        [0.8157, 0.7882, 0.7765,  ..., 0.6275, 0.2196, 0.2078],\n        [0.7059, 0.6784, 0.7294,  ..., 0.7216, 0.3804, 0.3255],\n        [0.6941, 0.6588, 0.7020,  ..., 0.8471, 0.5922, 0.4824]])\ntensor(1.)\ntensor(0.)\ntensor(0.4057)\nApproach2 Step1 done\nApproach2 Step2 done\n__call__ inside Normalization is called\nApproach2 Step3 done\ntensor([[-0.5373, -0.6627, -0.6078,  ...,  0.2392,  0.1922,  0.1608],\n        [-0.8745, -1.0000, -0.8588,  ..., -0.0353, -0.0667, -0.0431],\n        [-0.8039, -0.8745, -0.6157,  ..., -0.0745, -0.0588, -0.1451],\n        ...,\n        [ 0.6314,  0.5765,  0.5529,  ...,  0.2549, -0.5608, -0.5843],\n        [ 0.4118,  0.3569,  0.4588,  ...,  0.4431, -0.2392, -0.3490],\n        [ 0.3882,  0.3176,  0.4039,  ...,  0.6941,  0.1843, -0.0353]])\ntensor(1.)\ntensor(-1.)\ntensor(-0.1886)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53333247,"AnswerCount":0,"Tags":"<java><python><django><tensorflow>","CreationDate":"2018-11-16T07:27:59.587","AcceptedAnswerId":null,"OwnerUserId":1799284.0,"Title":"Java REST api vs Django REST api for ML","Body":"<p>It's been trivial thing for me!, and bugging me for over a 2 weeks.<\/p>\n\n<p>My Project Contains MVC in Java - Spring+Hibernate.<\/p>\n\n<p>I've been using Java for more than 4 years, so far it's good, but recent Machine learning bothering me so much because i just want to restart my 4 year old project in Machine Learning(It's best thing for my project, And i can't reinvent algorithm to match for my project).<\/p>\n\n<p>I read lot about Machine learning, I was so keen at Google TensorFlow, Sadly i can't use it in Java as i know, The API by google is not ready for Java.On the other hand i read about DL4J, it's seems okay but i have to explore it.<\/p>\n\n<p>On the Python side TensorFlow works flawless and SciKit too, but i'm new to Python Django REST framework. I'll quickly may catch up DRF. And it's seems that DRF is so light weight and pretty great without like spring annotation and hibernate annotation and too many stupid things.  <\/p>\n\n<p>But one thing bothering so much! is that should i stick to Java with DL4J or switch it to Python DRF?<\/p>\n\n<p>Please suggest me which track should i choose?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":53333756,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-16T08:07:05.430","AcceptedAnswerId":53340552.0,"OwnerUserId":10170730.0,"Title":"How can I add a small image to a bigger one in Tensorflow?","Body":"<p>I want to overlay a smaller image onto a larger one.<\/p>\n\n<p>I have tried adding to a slice but couldn't get it to work.<\/p>\n\n<p>So, as a simple example, how do I perform this NumPy operation in Tensorflow:<\/p>\n\n<pre><code>a = np.array([1, 1, 1, 1])\nb = np.array([5, 5])\nc = a\nc[1:3] = c[1:3] + b\nprint(c)\n# =&gt; [1 6 6 1]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53340552","CreationDate":"2018-11-16T15:13:15.137","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>This is one possible implementation:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# i and j are first row and colum\n# alpha (0..1) is the intensity of the overlay\ndef overlay_patch(img, patch, i, j, alpha=0.5):\n    img_shape = tf.shape(img)\n    img_rows, img_cols = img_shape[0], img_shape[1]\n    patch_shape = tf.shape(patch)\n    patch_rows, patch_cols = patch_shape[0], patch_shape[1]\n    i_end = i + patch_rows\n    j_end = j + patch_cols\n    # Mix patch: alpha from patch, minus alpha from image\n    overlay = alpha * (patch - img[i:i_end, j:j_end])\n    # Pad patch\n    overlay_pad = tf.pad(overlay, [[i, img_rows - i_end], [j, img_cols - j_end], [0, 0]])\n    # Make final image\n    img_overlay = img + overlay_pad\n    return img_overlay\n<\/code><\/pre>\n\n<p>Test:<\/p>\n\n<pre><code>img = tf.placeholder(tf.float32, [None, None, None])\npatch = tf.placeholder(tf.float32, [None, None, None])\ni = tf.placeholder(tf.int32, [])\nj = tf.placeholder(tf.int32, [])\nalpha = tf.placeholder(tf.float32, [])\nimg_overlay = overlay_patch(img, patch, i, j, alpha)\nwith tf.Session() as sess:\n    result = sess.run(img_overlay, feed_dict={\n        img: [[[ 1], [ 2], [ 3], [ 4]],\n              [[ 5], [ 6], [ 7], [ 8]],\n              [[ 9], [10], [11], [12]],\n              [[13], [14], [15], [16]]],\n        patch: [[[10], [20], [30]],\n                [[40], [50], [60]]],\n        i: 2, j: 1, alpha: 0.5\n    })\n    print(result[..., 0])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[ 1.   2.   3.   4. ]\n [ 5.   6.   7.   8. ]\n [ 9.  10.  15.5 21. ]\n [13.  27.  32.5 38. ]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53333769,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-16T08:07:54.067","AcceptedAnswerId":53400786.0,"OwnerUserId":10661534.0,"Title":"InfeedEnqueueTuple issue when trying to restore updated BERT model checkpoint using Cloud TPU","Body":"<p>I'd appreciate any help on the below, thank you in advance. I made a copy of <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/bert_finetuning_with_cloud_tpus.ipynb\" rel=\"nofollow noreferrer\">Google Bert's notebook on fine-tuning<\/a> and trained the SQUAD dataset on it using Cloud TPU and Bucket. The predictions on the dev set are ok, so I downloaded the checkpoint, model.ckpt.meta, model.ckpt.index and model.ckpt.data files locally and tried to restore using code:<\/p>\n\n<pre><code>sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))\nsaver = tf.train.import_meta_graph(META_FILE) # META_FILE being path to .meta\nsaver.restore(sess, 'model.ckpt')\n<\/code><\/pre>\n\n<p>However, I got the error:<\/p>\n\n<pre><code>    op_def = op_dict[node.op]\nKeyError: 'InfeedEnqueueTuple'\n<\/code><\/pre>\n\n<p>I assume it is part of <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/cloud-tpu-tools\" rel=\"nofollow noreferrer\">Cloud TPU Tools<\/a> and I should continue on Cloud TPU, so I tried the below (<a href=\"https:\/\/stackoverflow.com\/questions\/53125515\/how-to-restore-tensorflow-model-from-google-bucket-without-writing-to-filesystem\">reference<\/a>):<\/p>\n\n<pre><code># code from cells before includes\n...\ntf.contrib.cloud.configure_gcs(session, credentials=auth_info)\n...\ntpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(TPU_ADDRESS)\nrun_config = tf.contrib.tpu.RunConfig(\n    cluster=tpu_cluster_resolver,\n    model_dir=OUTPUT_DIR,\n    save_checkpoints_steps=SAVE_CHECKPOINTS_STEPS,\n    tpu_config=tf.contrib.tpu.TPUConfig(\n        iterations_per_loop=ITERATIONS_PER_LOOP,\n        num_shards=NUM_TPU_CORES,\n        per_host_input_for_training=tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2))\n...\n<\/code><\/pre>\n\n<p>Problem cell:<\/p>\n\n<pre><code>\"\"\"\n# not valid checkpoint error. &lt;bucket&gt; placeholder for cloud bucket name\nsess = tf.Session()\nMETA_FILE = \"gs:\/\/&lt;bucket&gt;\/bert\/models\/bertsquad\/model.ckpt-10949.meta\"\nCKPT_FILE = \"gs:\/\/&lt;bucket&gt;\/bert\/models\/bertsquad\/model.ckpt\"\nsaver = tf.train.import_meta_graph(META_FILE)\nsaver.restore(sess, CKPT_FILE)\n\"\"\"\n\nfrom google.cloud import storage\nfrom tensorflow import MetaGraphDef\n\nclient = storage.Client(project=\"agent-helper-4a014\")\nbucket = client.get_bucket(&lt;bucket&gt;)\nmetafile = \"bert\/models\/bertsquad\/model.ckpt-10949.meta\"\n# using full path gs:\/\/&lt;bucket&gt;\/bert\/models\/bertsquad doesn't work\n\nblob = bucket.get_blob(metafile)\n#blob = bucket.blob(metafile)\n#model_graph = blob.download_to_filename(\"model.ckpt\")\nmodel_graph = blob.download_as_string()\n\nmgd = MetaGraphDef()\nmgd.ParseFromString(model_graph)\n\nwith tf.Session() as sess:\n    saver = tf.train.import_meta_graph(mgd, clear_devices=True)\n    init_checkpoint = saver.restore(sess, 'model.ckpt')\n<\/code><\/pre>\n\n<p>That in turn gave the following error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:\n\nNo OpKernel was registered to support Op 'InfeedEnqueueTuple' with these attrs.  Registered devices: [CPU,XLA_CPU], Registered kernels:\n  &lt;no registered kernels&gt;\n\n     [[node input_pipeline_task0\/while\/InfeedQueue\/enqueue\/0 (defined at &lt;ipython-input-67-e4b52b7b5944&gt;:21)  = InfeedEnqueueTuple[_class=[\"loc:@input_pipeline_task0\/while\/IteratorGetNext\"], device_ordinal=0, dtypes=[DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32], shapes=[[2], [2,384], [2,384], [2,384], [2], [2]], _device=\"\/job:worker\/task:0\/device:CPU:0\"](input_pipeline_task0\/while\/IteratorGetNext, input_pipeline_task0\/while\/IteratorGetNext:1, input_pipeline_task0\/while\/IteratorGetNext:2, input_pipeline_task0\/while\/IteratorGetNext:3, input_pipeline_task0\/while\/IteratorGetNext:4, input_pipeline_task0\/while\/IteratorGetNext:5)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53400786","CreationDate":"2018-11-20T20:10:08.743","ParentId":null,"OwnerUserId":"3870291","Title":null,"Body":"<p>If your motive is prediction, then just give the model_dir location(must be a GCS bucket) where the checkpoints and meta file are saved. The code will not go for training again (as the checkpoint is saved for the number of training steps and there is no change in the model graph). It will directly jump to the prediction.<\/p>\n\n<p>But if your use case really want to save the checkpoints, and restore it only for inference then follow the steps:<\/p>\n\n<ul>\n<li>Create the model network for each and every layer manually as that of original model or use saved .meta file to recreate the network using <code>tf.train.import()<\/code> function like this:<\/li>\n<\/ul>\n\n<p><code>saver = tf.train.import_meta_graph('&lt;filename&gt;.meta')<\/code><\/p>\n\n<ul>\n<li>Now, restore the checkpoints using: <code>saver.restore(sess, 'model.ckpt')<\/code><\/li>\n<\/ul>\n\n<p><strong>NOTE:<\/strong> Model graph into which checkpoints are being restored, should be exactly the same of that original graph for which those checkpoints are saved.<\/p>\n\n<p>Hope this solves your issue.<\/p>\n"}]}
+{"QuestionId":53333794,"AnswerCount":2,"Tags":"<python><tensorflow><gradient-descent>","CreationDate":"2018-11-16T08:10:03.563","AcceptedAnswerId":null,"OwnerUserId":10306015.0,"Title":"Multidimensional gradient descent in Tensorflow","Body":"<p>What does Tensorflow really do when the Gradient descent optimizer is applied to a \"loss\" placeholder that is not a number (a tensor of size 1) but rather a vector (a 1-dimensional tensor of size 2, 3, 4, or more)?<\/p>\n\n<p>Is it like doing the descent on the sum of the components?<\/p>\n","answers":[{"AnswerId":"53416420","CreationDate":"2018-11-21T16:23:49.233","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Tensorflow first reduces your loss to a scalar and then optimizes that.<\/p>\n"},{"AnswerId":"53336001","CreationDate":"2018-11-16T10:31:08.690","ParentId":null,"OwnerUserId":"1058309","Title":null,"Body":"<p>The answer to your second question is \"no\". <\/p>\n\n<p>As for the second: just like in the one-dimensional case (e.g. y = f(x), x in R), where the direction the algorithm takes is defined by the derivative of the function with respect to its single variable, in the multidimensional case the 'overall' direction is defined by the derivative of the function with respect to <em>each<\/em> variable. <\/p>\n\n<p>This means the size of the step you'll take in <em>each<\/em> direction will be determined by the value of the derivative of the variable corresponding to that direction.<\/p>\n\n<p>Since there's no way to properly type math in StackOverflow, instead of messing around with it I'll suggest you take a look at this <a href=\"https:\/\/towardsdatascience.com\/machine-learning-bit-by-bit-multivariate-gradient-descent-e198fdd0df85\" rel=\"nofollow noreferrer\">article<\/a>.<\/p>\n"}]}
+{"QuestionId":53334559,"AnswerCount":1,"Tags":"<tensorflow><keras><lstm><keras-layer><vgg-net>","CreationDate":"2018-11-16T09:06:20.120","AcceptedAnswerId":53335595.0,"OwnerUserId":8049537.0,"Title":"Error in Merging one LSTM model output and VGG model output","Body":"\n\n<pre class=\"lang-py prettyprint-override\"><code>#make custom embedding layer\nembedding_layer = Embedding(len(word_index) + 1,\n                        EMBEDDING_DIM,\n                        weights=[embedding_matrix],\n                        input_length=max_length_of_text,\n                        trainable=False)\n\n#make lstm model\ninputs = Input((max_length_of_text, ))\nx = embedding_layer(inputs)\nx = LSTM(units=512, return_sequences=True)(x)\nx = Dropout(0.1)(x)\nx = LSTM(units=512, return_sequences=False)(x)\nx = Dropout(0.1)(x)\nx = Dense(1024, activation='tanh')(x)\nlstm_model = Model(inputs,x)\n\n\n#make vgg model\nvgg = VGG16(weights=None, include_top=True)\nvgg.load_weights('.\/vgg_weights.h5')\nvgg.layers.pop()\nfor l in vgg.layers[:]:\n    l.trainable = False\ninp = vgg.input\nout = Dense(1024, activation='tanh')(vgg.layers[-1].output)\nvgg_model = Model(inp,out)\n\n\n#make final model\nfc_model = Sequential()\nfc_model = Multiply()([vgg_model.output,lstm_model.output]))\nfc_model.add(Merge([vgg_model, lstm_model], mode='mul'))\nfc_model.add(Dropout(0.2))\nfc_model.add(Dense(512, activation='tanh'))\nfc_model.add(Dropout(0.2))\nfc_model.add(Dense(26, activation='softmax'))\nfc_model.compile(optimizer='rmsprop', loss='categorical_crossentropy',\n    metrics=['accuracy'])\n\n\nfc_model.summary()\n\n'''\nif os.path.exists(model_weights_filename):\n   print \"Loading Weights...\"\n   fc_model.load_weights(model_weights_filename)\n\n'''\n\n\n#Train\n\n#--\nimg_lis_train = img_lis[ : 108032]\nimg_lis_test  = img_lis[ 108032 : ]\nquestions_train = questions[ : 108032]\nquestions_test  = questions[ 108032 : ]\nanswers_train = answers[ : 108032]\nanswers_test  = answers[ 108032 : ]\n\n\ndef mygen(questions_train,img_lis_train,answers_train):\n    start = 0  \n    data_size = len(questions_train)\n    batch_size = 64\n    while True:          \n        if( start+batch_size &lt;= data_size ):\n            batch_ques = questions_train[ start : start+batch_size ] \n            batch_ans = answers_train[ start : start+batch_size ] \n            batch_img_names = img_lis_train[ start : start+batch_size ] \n        elif(start &lt; data_size):\n            batch_ques = questions_train[ start : ] \n            batch_ans = answers_train[ start : ] \n            batch_img_names = img_lis_train[ start : start+batch_size ] \n        else:\n            break       \n\n        batch_img = []\n        for img_name in batch_img_names:\n            img = load_img('.\/dataset\/images\/' + str(img_name) + '.png' , target_size = (224,224))\n            img = img_to_array(img)   \n            batch_img.append( preprocess_input(img) )    \n\n        start += batch_size\n        print('start = ' + str(start))\n        yield [np.array(batch_img), np.array(batch_ques)] ,np.array(batch_ans)\n\n\nclass WeightsSaver(Callback):\n    def __init__(self, N):\n        self.N = N\n        self.batch = 0\n\n    def on_batch_end(self, batch, logs={}):\n        if self.batch % self.N == 0:\n            name = '.\/weights\/weights.h5'\n            self.model.save_weights(name)\n        self.batch += 1\n\nfc_model.load_weights('.\/weights\/weights.h5')\nfc_model.fit_generator(mygen(questions_train, img_lis_train , answers_train ), steps_per_epoch = 1688, epochs = 100 ,callbacks=[WeightsSaver(10)])\n<\/code><\/pre>\n\n<p>The errors I am getting is <code>fc_model<\/code> is converting to tensor using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>fc_model = Multiply()([vgg_model.output,lstm_model.output]))\n<\/code><\/pre>\n\n<p>and in earlier version of Keras <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>fc_model.add(Merge([vgg_model, lstm_model], mode='mul'))\n<\/code><\/pre>\n\n<p>was working but <code>Merge<\/code> is removed from new versions of Keras. I want to multiply the output LSTM model and VGG model in <code>fc_model<\/code>. Can anyone suggest me anything?<\/p>\n","answers":[{"AnswerId":"53335595","CreationDate":"2018-11-16T10:09:59.420","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You need to use <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\" rel=\"nofollow noreferrer\">Keras functional API<\/a> to be able construct a non-sequential model. So your code would look like this:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>mult = Multiply()([vgg_model.output,lstm_model.output])\nx = Dropout(0.2)(mult)\nx = Dense(512, activation='tanh')(x)\nx = Dropout(0.2)(x)\nout = Dense(26, activation='softmax')(x)\n\nfc_model = Model([vgg_model.input, lstm_model.input], out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53335038,"AnswerCount":1,"Tags":"<android><tensorflow><face-detection><tensorflow-lite><firebase-mlkit>","CreationDate":"2018-11-16T09:36:44.940","AcceptedAnswerId":null,"OwnerUserId":6180094.0,"Title":"Firebase ML Kit: Pre-Trained Models","Body":"<p>I can see there are 3 pre-trained models included in the <a href=\"https:\/\/github.com\/firebase\/quickstart-android\/tree\/master\/mlkit\/app\/src\/main\/assets\" rel=\"nofollow noreferrer\">assets<\/a> folder of the ML-Kit sample app. <\/p>\n\n<p>Please help me understand<\/p>\n\n<ul>\n<li>Which of these models correspond to face detection?<\/li>\n<li>What is the data set used to train these models?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"53614523","CreationDate":"2018-12-04T13:54:33.547","ParentId":null,"OwnerUserId":"456208","Title":null,"Body":"<p>Those are custom models provided to highlight the ML Kit functionality that allows you to bring your own TFLite models.<\/p>\n\n<p>The pre-trained APIs, also known as Base APIs (text recognition, image labeling, etc) get added by including the appropriate libraries in your app's gradle file. No need to include anything in your assets folder. See an example here:\n<a href=\"https:\/\/firebase.google.com\/docs\/ml-kit\/android\/recognize-text\" rel=\"nofollow noreferrer\">https:\/\/firebase.google.com\/docs\/ml-kit\/android\/recognize-text<\/a><\/p>\n"}]}
+{"QuestionId":53335278,"AnswerCount":0,"Tags":"<keras><lstm>","CreationDate":"2018-11-16T09:51:23.743","AcceptedAnswerId":null,"OwnerUserId":5368297.0,"Title":"How to calculate the softmax in LSTM auto-encoder for input with zero padding","Body":"<p>I am implementing a LSTM auto-encoder that with the input x and hopefully get the same output as x. However, I have several questions in the implementing, and please see my codes below:<\/p>\n\n<pre><code>dict_size = 10\nmax_sentence_length = 5\nembed_dim = 20\n\nx = Input(shape=(max_sentence_length,), dtype='int32')\nencoder_input = Embedding(dict_size, embed_dim)(x)\nencoder_output=LSTM(32, return_sequences=True)(encoder_input)\ndeocder_output = LSTM(32, return_sequences=True)(encoder_output)\ny = Dense(dict_size, activation='softmax')(deocder_output)\n\nmodel = Model(inputs=x, outputs=y)\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\nprint(model.summary())\n\ntrain_x = np.array([[3, 1, 2]])\ntrain_x = pad_sequences(train_x, max_sentence_length, padding='post')\ntrain_y = np_utils.to_categorical(train_x, dict_size)\n\nmodel.fit(train_x, train_y, batch_size=1, epochs=1)\npredict_y = model.predict(train_x)\nprint(predict_y)\n<\/code><\/pre>\n\n<p>In this codes, the first problem is that : \"ValueError: Error when checking target: expected dense_1 to have 3 dimensions, but got array with shape (5, 10)\", I really do not know how to solve this problem first.<\/p>\n\n<p>The second question is that, the length of x is 3, while with extra padding its length becomes to 5; At the final stage, the output would be a  5x10 matrix, i.e., the last two elements of the final output should not participate into the calculation of the softmax. Is there any way to fix this problem quickly? <\/p>\n\n<p>Many many thanks!<\/p>\n","answers":[]}
+{"QuestionId":53335293,"AnswerCount":1,"Tags":"<neural-network><caffe><object-detection><tensorrt><nvidia-jetson>","CreationDate":"2018-11-16T09:52:32.887","AcceptedAnswerId":null,"OwnerUserId":6084488.0,"Title":"Error parsing Mobilenet-SSD using TensorRT Caffe Parser, \"ditcaffe.LayerParameter\" has no field named \"prior_box_param\"","Body":"<p>I'm parsing MobileNet-SSD caffe Model from <a href=\"https:\/\/github.com\/chuanqi305\/MobileNet-SSD\" rel=\"nofollow noreferrer\">https:\/\/github.com\/chuanqi305\/MobileNet-SSD<\/a> using TensorRT caffe parser. I use Jetpack 3.3 in Jetson TX2 as the platform. How to fix this error from nvinfer library? There is even a TensorRT class <code>nvinfer1::plugin::PriorBoxParameters<\/code> in the API. But this layer handle seems to be not defined. Is there any error in the layer param names and format? Or do we have to create a custom layer plugin for this one. Does someone have a implementation for this?<\/p>\n\n<p>The error is as follows,<\/p>\n\n<pre><code>[libprotobuf ERROR google\/protobuf\/text_format.cc:298] Error parsing text-format ditcaffe.NetParameter: 1245:18: Message type \"ditcaffe.LayerParameter\" has no field named \"prior_box_param\".\nERROR: CaffeParser: Could not parse deploy file\n<\/code><\/pre>\n\n<p>The source code line relevant for this error,<\/p>\n\n<pre><code>m_network = m_builder-&gt;createNetwork();\nm_parser = createCaffeParser();\nconst IBlobNameToTensor* blobNameToTensor = m_parser-&gt;parse(deployFpath, modelFpath, *network, DataType::kFLOAT);\n<\/code><\/pre>\n\n<p>Error given layer in prototxt file,<\/p>\n\n<pre><code>layer {\n  name: \"conv11_mbox_priorbox\"\n  type: \"PriorBox\"\n  bottom: \"conv11\"\n  bottom: \"data\"\n  top: \"conv11_mbox_priorbox\"\n  prior_box_param {\n    min_size: 60.0\n    aspect_ratio: 2.0\n    flip: true\n    clip: false\n    variance: 0.1\n    variance: 0.1\n    variance: 0.2\n    variance: 0.2\n    offset: 0.5\n  }\n} \n<\/code><\/pre>\n","answers":[{"AnswerId":"55178850","CreationDate":"2019-03-15T08:58:10.237","ParentId":null,"OwnerUserId":"11207496","Title":null,"Body":"<p>You have to remove all the blocks that have _param from the ptototxt file, then you should removed layers using the plugin API.\nVisit the discussion on this link for clear understanding: <a href=\"https:\/\/github.com\/chenzhi1992\/TensorRT-SSD\/issues\/5\" rel=\"nofollow noreferrer\">https:\/\/github.com\/chenzhi1992\/TensorRT-SSD\/issues\/5<\/a><\/p>\n"}]}
+{"QuestionId":53335338,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><predict>","CreationDate":"2018-11-16T09:55:03.500","AcceptedAnswerId":53335383.0,"OwnerUserId":4782060.0,"Title":"How can I predict the next elements in a dataset with LSTM in Keras, python?","Body":"<p>This is my first time with Keras and LSTMs and I am working in a project in which I have many time series data to train with.<\/p>\n\n<p>I have around 13000 rows of data (1 column) with numerical values regarding to a degradation level of a component ending in a failure; and on the other side I have multiple datasets of 100 rows (1 column) with data regarding to a degradation level of a component, but ending some points before a failure.<\/p>\n\n<p>The challenge is to predict when those datasets will record a failure.<\/p>\n\n<p>So what I have done, is the next:<\/p>\n\n<pre><code>from pandas import DataFrame\nfrom pandas import Series\nfrom pandas import concat\nfrom pandas import read_csv\nfrom pandas import datetime\nfrom sklearn.metrics import mean_squared_error\nfrom sklearn.preprocessing import MinMaxScaler\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom math import sqrt\nfrom matplotlib import pyplot\nfrom numpy import array\n\n\n# convert an array of values into a dataset matrix\ndef create_dataset(dataset, look_back):\n    dataX, dataY = [], []\n    for i in range(len(dataset)-look_back-1):\n        a = dataset[i:(i+look_back), 0]\n        dataX.append(a)\n        dataY.append(dataset[i + look_back, 0])\n    return numpy.array(dataX), numpy.array(dataY)\n\n\nlook_back = 600\nepochs = 500\nbatch_size = 50\n\n\n\n\ndata = np.array(data).reshape(-1,1)\ndata = data.astype('float32')\n\n# Scale the data\nscaler = MinMaxScaler(feature_range=(0, 1))\ndataset = scaler.fit_transform(data)\n\n# split into train and test sets\ntrain_size = int(len(dataset) * 0.67)\ntest_size = len(dataset) - train_size\ntrain, test = dataset[0:train_size,:], dataset[train_size:len(dataset),:]\n\ntrainX, trainY = create_dataset(train, look_back)\ntestX, testY = create_dataset(test, look_back)\n\n\n# reshape input to be [samples, time steps, features]\ntrainX = numpy.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))\ntestX = numpy.reshape(testX, (testX.shape[0], 1, testX.shape[1]))\n\n\n# create and fit the LSTM network\nmodel = Sequential()\nmodel.add(LSTM(100, activation = 'tanh', inner_activation = 'hard_sigmoid', return_sequences=True))\nmodel.add(LSTM(50, activation = 'tanh', inner_activation = 'hard_sigmoid'))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.fit(trainX, trainY, epochs=epochs, batch_size=batch_size, verbose=2)\n\n\n\n# make predictions\ntrainPredict = model.predict(trainX)\ntestPredict = model.predict(testX)\n\n\n# invert predictions\ntrainPredict = scaler.inverse_transform(trainPredict)\ntrainY = scaler.inverse_transform([trainY])\ntestPredict = scaler.inverse_transform(testPredict)\ntestY = scaler.inverse_transform([testY])\n\n\n\n\n\n# calculate root mean squared error\ntrainScore = math.sqrt(mean_squared_error(trainY[0], trainPredict[:,0]))\nprint('Train Score: %.2f RMSE' % (trainScore))\ntestScore = math.sqrt(mean_squared_error(testY[0], testPredict[:,0]))\nprint('Test Score: %.2f RMSE' % (testScore))\n<\/code><\/pre>\n\n<p>These line codes represents the training and the evaluation of the dataset, but how can I predict the next (for example) 50 elements of one dataset of 100 rows using that model?<\/p>\n","answers":[{"AnswerId":"53335383","CreationDate":"2018-11-16T09:57:59.850","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>Try this:\n    model.predict(newX)<\/p>\n"}]}
+{"QuestionId":53337586,"AnswerCount":3,"Tags":"<python><machine-learning><keras><theano><convolution>","CreationDate":"2018-11-16T12:06:11.540","AcceptedAnswerId":53338183.0,"OwnerUserId":1744914.0,"Title":"Convolution2D gives wrong output shape in Keras","Body":"<p>I'm following tutorial from <a href=\"https:\/\/elitedatascience.com\/keras-tutorial-deep-learning-in-python\" rel=\"nofollow noreferrer\">here<\/a> and have stuck with simple model creation<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2D(32, kernel_size=(3, 3), activation='relu', input_shape=(1, 28, 28)))\nmodel.output_shape\n<\/code><\/pre>\n\n<p>This should give <code>(None, 32, 26, 26)<\/code> output but instead I get <code>(None, -1, 26, 32)<\/code>. I guess something wrong with parameters provided or I need to make additional reshape. Maybe something has been changed for <code>Convolution2D<\/code> and <code>Conv2D<\/code> in Keras? I'm running <code>2.2.4<\/code> with <code>Theano<\/code> backend<\/p>\n","answers":[{"AnswerId":"53337765","CreationDate":"2018-11-16T12:17:43.633","ParentId":null,"OwnerUserId":"5108062","Title":null,"Body":"<p>The problem is that it uses Theano as a backend and probably you are using tensorflow (you should check your setting once again). You can change it with: <\/p>\n\n<pre><code>import os\nos.environ['KERAS_BACKEND'] = 'theano'\n<\/code><\/pre>\n\n<p>This should be done before importing <code>keras<\/code><\/p>\n"},{"AnswerId":"53337844","CreationDate":"2018-11-16T12:22:26.897","ParentId":null,"OwnerUserId":"3602069","Title":null,"Body":"<p>Well like I wrote in my comment. I got a error with your code. So see that it looks like you have bad shape on input, if you change to <code>input_shape=(28,28,1)<\/code> then output_shape returns <code>(None, 26, 26, 32)<\/code><\/p>\n\n<p>I don't know if this is what you are exactly looking for.<\/p>\n"},{"AnswerId":"53338183","CreationDate":"2018-11-16T12:44:16.920","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>It seems that the channels are in the first axis of input data. Therefore, you need to set <code>data_format<\/code> argument to <code>\"channels_first\"<\/code>:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2D(32, kernel_size=(3, 3),\n                        data_format='channels_first',\n                        activation='relu', input_shape=(1, 28, 28)))\nmodel.output_shape\n<\/code><\/pre>\n\n<p>Which gives:<\/p>\n\n<pre><code>(None, 32, 26, 26)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53337880,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensor>","CreationDate":"2018-11-16T12:24:31.137","AcceptedAnswerId":53338100.0,"OwnerUserId":7326400.0,"Title":"How to stack (None,1024) tensor into (None,7,7,1024) tensor in keras?","Body":"<p>I have a tensor <code>A<\/code> of shape <code>(None,1024)<\/code> and I want to repeat it 49 times to make a matrix tensor of shape <code>(None,7,7,1024)<\/code>. Every <code>(,1024)<\/code> vector is a copy of the tensor <code>A<\/code>.<\/p>\n\n<p>How do I do this?<\/p>\n\n<p>I tried<\/p>\n\n<pre><code>A = K.stack([A,A,A,A,A,A,A])\nA = K.stack([A,A,A,A,A,A,A])\n<\/code><\/pre>\n\n<p>But this gives me a tensor of shape <code>(7, 7, None, 1024)<\/code><\/p>\n","answers":[{"AnswerId":"53338100","CreationDate":"2018-11-16T12:39:24.640","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First use <a href=\"https:\/\/keras.io\/layers\/core\/#repeatvector\" rel=\"nofollow noreferrer\"><code>RepeatVector<\/code><\/a> layer to repeat it for 49 times, then use <a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\"><code>Reshape<\/code><\/a> layer to reshape it as desired:<\/p>\n\n<pre><code>r_A = RepeatVector(49)(A)\nrsh_A = Reshape((7, 7, -1))(r_A)\n<\/code><\/pre>\n\n<p>Alternatively, you can define a function based on backend functions and wrap it inside a <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef rsh(x):\n    y = K.repeat(x, 49)\n    return K.reshape(y, (-1, 7, 7, K.int_shape(y)[-1]))\n\nrsh_A = Lambda(rsh)(A)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53337927,"AnswerCount":1,"Tags":"<machine-learning><caffe><build-error>","CreationDate":"2018-11-16T12:27:53.060","AcceptedAnswerId":null,"OwnerUserId":10041301.0,"Title":"Caffe build fail: Makefile:591: recipe for target '.build_release\/src\/caffe\/util\/io.o' failed","Body":"<p>Hello I' ve got issue on caffe installation. I did everything in accordance to the instruction <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-16.04-Installation-Guide\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-16.04-Installation-Guide<\/a> (for CPU version)\nOn command make all I've got the following output:<\/p>\n\n<pre><code>CXX src\/caffe\/util\/io.cpp\nIn file included from \/usr\/local\/include\/opencv2\/core.hpp:52:0,\n                 from \/usr\/local\/include\/opencv2\/core\/core.hpp:48,\n                 from src\/caffe\/util\/io.cpp:6:\n\/usr\/local\/include\/opencv2\/core\/cvdef.h:421:4: error: #error \"OpenCV 4.x+ requires enabled C++11 support\"\n #  error \"OpenCV 4.x+ requires enabled C++11 support\"\n    ^\nIn file included from \/usr\/include\/c++\/5\/array:35:0,\n                 from \/usr\/local\/include\/opencv2\/core\/cvdef.h:426,\n                 from \/usr\/local\/include\/opencv2\/core.hpp:52,\n                 from \/usr\/local\/include\/opencv2\/core\/core.hpp:48,\n                 from src\/caffe\/util\/io.cpp:6:\n\/usr\/include\/c++\/5\/bits\/c++0x_warning.h:32:2: error: #error This file requires compiler and library support for the ISO C++ 2011 standard. This support must be enabled with the -std=c++11 or -std=gnu++11 compiler options.\n #error This file requires compiler and library support \\\n  ^\nIn file included from \/usr\/local\/include\/opencv2\/core\/base.hpp:58:0,\n                 from \/usr\/local\/include\/opencv2\/core.hpp:54,\n                 from \/usr\/local\/include\/opencv2\/core\/core.hpp:48,\n                 from src\/caffe\/util\/io.cpp:6:\n\/usr\/local\/include\/opencv2\/core\/cvstd.hpp:389:12: error: expected \u2018,\u2019 or \u2018...\u2019 before \u2018&amp;&amp;\u2019 token\n     Ptr(Ptr&amp;&amp; o);\n            ^\n\/usr\/local\/include\/opencv2\/core\/cvstd.hpp:389:16: error: invalid constructor; you probably meant \u2018cv::Ptr&lt;T&gt; (const cv::Ptr&lt;T&gt;&amp;)\u2019\n     Ptr(Ptr&amp;&amp; o);\n                ^\n\/usr\/local\/include\/opencv2\/core\/cvstd.hpp:390:25: error: expected \u2018,\u2019 or \u2018...\u2019 before \u2018&amp;&amp;\u2019 token\n     Ptr&amp; operator = (Ptr&amp;&amp; o);\n                         ^\nIn file included from \/usr\/local\/include\/opencv2\/core\/cvstd.hpp:1033:0,\n                 from \/usr\/local\/include\/opencv2\/core\/base.hpp:58,\n                 from \/usr\/local\/include\/opencv2\/core.hpp:54,\n                 from \/usr\/local\/include\/opencv2\/core\/core.hpp:48,\n                 from src\/caffe\/util\/io.cpp:6:\n\/usr\/local\/include\/opencv2\/core\/ptr.inl.hpp:256:16: error: expected \u2018,\u2019 or \u2018...\u2019 before \u2018&amp;&amp;\u2019 token\n Ptr&lt;T&gt;::Ptr(Ptr&amp;&amp; o) : owner(o.owner), stored(o.stored)\n                ^\n\/usr\/local\/include\/opencv2\/core\/ptr.inl.hpp:256:1: error: prototype for \u2018cv::Ptr&lt;T&gt;::Ptr(cv::Ptr&lt;T&gt;)\u2019 does not match any in class \u2018cv::Ptr&lt;T&gt;\u2019\n Ptr&lt;T&gt;::Ptr(Ptr&amp;&amp; o) : owner(o.owner), stored(o.stored)\n ^\nIn file included from \/usr\/local\/include\/opencv2\/core\/base.hpp:58:0,\n                 from \/usr\/local\/include\/opencv2\/core.hpp:54,\n                 from \/usr\/local\/include\/opencv2\/core\/core.hpp:48,\n                 from src\/caffe\/util\/io.cpp:6:\n\/usr\/local\/include\/opencv2\/core\/cvstd.hpp:319:5: error: candidates are: template&lt;class T&gt; template&lt;class Y&gt; cv::Ptr&lt;T&gt;::Ptr(const cv::Ptr&lt;Y&gt;&amp;, T*)\n     Ptr(const Ptr&lt;Y&gt;&amp; o, T* p);\n<\/code><\/pre>\n\n<p>Here is also a lot of text in output but I cannot upload it because of limit on number of characters<\/p>\n\n<pre><code>Makefile:591: recipe for target '.build_release\/src\/caffe\/util\/io.o' failed\nmake: *** [.build_release\/src\/caffe\/util\/io.o] Error 1\n<\/code><\/pre>\n\n<p>Can somebody help me with this problem?<\/p>\n\n<p>OS: Ubuntu 16.04 \ncmake version 3.5.1\nCuda version: 10.0.130<\/p>\n","answers":[{"AnswerId":"53760635","CreationDate":"2018-12-13T11:16:14.800","ParentId":null,"OwnerUserId":"10041301","Title":null,"Body":"<p>The problem was in 4th version of OpenCV. Caffe dosen't work with it. When I installed 3rd version everything was ok.<\/p>\n"}]}
+{"QuestionId":53338011,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2018-11-16T12:33:58.360","AcceptedAnswerId":null,"OwnerUserId":10662816.0,"Title":"got Error on start \"unsupported address family\"","Body":"<p>Than you for reading this<\/p>\n\n<p>when I run from source, I got \"unsupported address family\"  Error, because method \"_get_wildcard_address\"  returned a value of  \"::\"<\/p>\n\n<p>in tensorboard source code, I modified program.py, print some vars<\/p>\n\n<p>program.py<\/p>\n\n<pre><code>  def _get_wildcard_address(self, port):\n    \"\"\"Returns a wildcard address for the port in question.\n\n    This will attempt to follow the best practice of calling getaddrinfo() with\n    a null host and AI_PASSIVE to request a server-side socket wildcard address.\n    If that succeeds, this returns the first IPv6 address found, or if none,\n    then returns the first IPv4 address. If that fails, then this returns the\n    hardcoded address \"::\" if socket.has_ipv6 is True, else \"0.0.0.0\".\n    \"\"\"\n    print(\"port:\", port)\n    fallback_address = '::' if socket.has_ipv6 else '0.0.0.0'\n    if hasattr(socket, 'AI_PASSIVE'):\n      try:\n        addrinfos = socket.getaddrinfo(None, port, socket.AF_UNSPEC,\n                                       socket.SOCK_STREAM, socket.IPPROTO_TCP,\n                                       socket.AI_PASSIVE)\n        print(\"addrinfos:\", addrinfos)\n      except socket.gaierror as e:\n        logger.warn('Failed to auto-detect wildcard address, assuming %s: %s',\n                    fallback_address, str(e))\n        return fallback_address\n      addrs_by_family = defaultdict(list)\n      for family, _, _, _, sockaddr in addrinfos:\n        # Format of the \"sockaddr\" socket address varies by address family,\n        # but [0] is always the IP address portion.\n        addrs_by_family[family].append(sockaddr[0])\n      print(\"addrs_by_family:\", addrs_by_family)\n      print(\"socket.AF_INET6:\", socket.AF_INET6)\n      print(\"socket.AF_INET:\", socket.AF_INET)\n      if hasattr(socket, 'AF_INET6') and addrs_by_family[socket.AF_INET6]:\n        print(\"return INET6 addr:\", addrs_by_family[socket.AF_INET6][0])\n        return addrs_by_family[socket.AF_INET6][0]\n      if hasattr(socket, 'AF_INET') and addrs_by_family[socket.AF_INET]:\n        print(\"return INET addr:\", addrs_by_family[socket.AF_INET][0])\n        return addrs_by_family[socket.AF_INET][0]\n    logger.warn('Failed to auto-detect wildcard address, assuming %s',\n            fallback_address)\nreturn fallback_address\n<\/code><\/pre>\n\n<p>run command got this:<\/p>\n\n<pre><code>$ bazel run \/\/tensorboard --  --logdir  \/home\/admin\/sean.xd\/tmp\/   \nINFO: Analysed target \/\/tensorboard:tensorboard (0 packages loaded).\nINFO: Found 1 target...\nTarget \/\/tensorboard:tensorboard up-to-date:\n  bazel-bin\/tensorboard\/tensorboard\nINFO: Elapsed time: 0.280s, Critical Path: 0.00s\nINFO: 0 processes.\nINFO: Build completed successfully, 1 total action\nINFO: Build completed successfully, 1 total action\nport: 6006\naddrinfos: [(2, 1, 6, '', ('0.0.0.0', 6006)), (10, 1, 6, '', ('::', 6006, 0, 0))]\naddrs_by_family: defaultdict(&lt;type 'list'&gt;, {2: ['0.0.0.0'], 10: ['::']})\nsocket.AF_INET6: 10\nsocket.AF_INET: 2\nreturn INET6 addr: ::\nhost: ::\nE1116 20:02:21.122143 MainThread program.py:201] Tensorboard could not bind to unsupported address family ::\nWARNING: Logging before flag parsing goes to stderr.\nE1116 20:02:21.122143 139781956085568 program.py:201] Tensorboard could not bind to unsupported address family ::\nERROR: Tensorboard could not bind to unsupported address family ::\n<\/code><\/pre>\n\n<p>envivonment:<\/p>\n\n<p>$ uname -a<\/p>\n\n<p>Linux arks7.inc.alipay.net 3.10.0-327.ali2000.alios7.x86_64 #1 SMP Tue Dec 29 19:54:05 CST 2015 x86_64 x86_64 x86_64 GNU\/Linux\n(xudong_test) <\/p>\n","answers":[{"AnswerId":"53744651","CreationDate":"2018-12-12T13:57:17.230","ParentId":null,"OwnerUserId":"5979631","Title":null,"Body":"<p>Recently, I faced a similar problem. \nIn case, if you haven't found a solution, you may want to give a look at <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/CSKLSgCqjaM\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/CSKLSgCqjaM<\/a><\/p>\n"},{"AnswerId":"56519993","CreationDate":"2019-06-10T02:58:29.657","ParentId":null,"OwnerUserId":"6055868","Title":null,"Body":"<p>I solve this problem by adding the \"--host\" parameter, that is:<\/p>\n\n<pre><code>tensorboard --logdir .\/  --port xxxx  --host 0.0.0.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":53338621,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><google-drive-api><google-colaboratory>","CreationDate":"2018-11-16T13:12:08.640","AcceptedAnswerId":53343032.0,"OwnerUserId":458105.0,"Title":"Issue with gdrive on Colab","Body":"<p>I am using colab to train resnet on cifar10, after mounting google drive \nI cloned the repository and I was able to run the script.\nHowever, Tensorflow is loaded and the data files are passed to the network  but I am ending with: <\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.NotFoundError:\n  \/content\/drive\/My; No such file or directory<\/p>\n<\/blockquote>\n\n<p>It seems that there is an issue with my path because it contains a space \"\/content\/gdrive\/My Drive\/apps\/PocketFlow\". How I could change the way that gdrive is mounted, in other words can I change \"My drive\" to something else to run the test again?<\/p>\n\n<p>Below you can find the code and the log file:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/gdrive')\nimport os\nos.chdir(\"\/content\/gdrive\/My Drive\/apps\/PocketFlow\")\n!chmod 755 .\/scripts\/run_local.sh\n!.\/scripts\/run_local.sh nets\/resnet_at_cifar10_run.py\n<\/code><\/pre>\n\n<p><strong>the log:<\/strong><\/p>\n\n<pre><code>Python script: nets\/resnet_at_cifar10_run.py\n# of GPUs: 1\nextra arguments:  --model_http_url https:\/\/api.ai.tencent.com\/pocketflow --data_dir_local \/content\/drive\/My Drive\/apps\/datasets\/cifar10\n'nets\/resnet_at_cifar10_run.py' -&gt; 'main.py'\nmulti-GPU training disabled\n[WARNING] TF-Plus &amp; Horovod cannot be imported; multi-GPU training is unsupported\nINFO:tensorflow:FLAGS:\nINFO:tensorflow:data_disk: local\nINFO:tensorflow:data_hdfs_host: None\nINFO:tensorflow:data_dir_local: \/content\/drive\/My\nINFO:tensorflow:data_dir_hdfs: None\nINFO:tensorflow:cycle_length: 4\nINFO:tensorflow:nb_threads: 8\nINFO:tensorflow:buffer_size: 1024\nINFO:tensorflow:prefetch_size: 8\nINFO:tensorflow:nb_classes: 10\nINFO:tensorflow:nb_smpls_train: 50000\nINFO:tensorflow:nb_smpls_val: 5000\nINFO:tensorflow:nb_smpls_eval: 10000\nINFO:tensorflow:batch_size: 128\nINFO:tensorflow:batch_size_eval: 100\nINFO:tensorflow:resnet_size: 20\nINFO:tensorflow:lrn_rate_init: 0.1\nINFO:tensorflow:batch_size_norm: 128.0\nINFO:tensorflow:momentum: 0.9\nINFO:tensorflow:loss_w_dcy: 0.0002\nINFO:tensorflow:model_http_url: https:\/\/api.ai.tencent.com\/pocketflow\nINFO:tensorflow:summ_step: 100\nINFO:tensorflow:save_step: 10000\nINFO:tensorflow:save_path: .\/models\/model.ckpt\nINFO:tensorflow:save_path_eval: .\/models_eval\/model.ckpt\nINFO:tensorflow:enbl_dst: False\nINFO:tensorflow:enbl_warm_start: False\nINFO:tensorflow:loss_w_dst: 4.0\nINFO:tensorflow:tempr_dst: 4.0\nINFO:tensorflow:save_path_dst: .\/models_dst\/model.ckpt\nINFO:tensorflow:nb_epochs_rat: 1.0\nINFO:tensorflow:ddpg_actor_depth: 2\nINFO:tensorflow:ddpg_actor_width: 64\nINFO:tensorflow:ddpg_critic_depth: 2\nINFO:tensorflow:ddpg_critic_width: 64\nINFO:tensorflow:ddpg_noise_type: param\nINFO:tensorflow:ddpg_noise_prtl: tdecy\nINFO:tensorflow:ddpg_noise_std_init: 1.0\nINFO:tensorflow:ddpg_noise_dst_finl: 0.01\nINFO:tensorflow:ddpg_noise_adpt_rat: 1.03\nINFO:tensorflow:ddpg_noise_std_finl: 1e-05\nINFO:tensorflow:ddpg_rms_eps: 0.0001\nINFO:tensorflow:ddpg_tau: 0.01\nINFO:tensorflow:ddpg_gamma: 0.9\nINFO:tensorflow:ddpg_lrn_rate: 0.001\nINFO:tensorflow:ddpg_loss_w_dcy: 0.0\nINFO:tensorflow:ddpg_record_step: 1\nINFO:tensorflow:ddpg_batch_size: 64\nINFO:tensorflow:ddpg_enbl_bsln_func: True\nINFO:tensorflow:ddpg_bsln_decy_rate: 0.95\nINFO:tensorflow:ws_save_path: .\/models_ws\/model.ckpt\nINFO:tensorflow:ws_prune_ratio: 0.75\nINFO:tensorflow:ws_prune_ratio_prtl: optimal\nINFO:tensorflow:ws_nb_rlouts: 200\nINFO:tensorflow:ws_nb_rlouts_min: 50\nINFO:tensorflow:ws_reward_type: single-obj\nINFO:tensorflow:ws_lrn_rate_rg: 0.03\nINFO:tensorflow:ws_nb_iters_rg: 20\nINFO:tensorflow:ws_lrn_rate_ft: 0.0003\nINFO:tensorflow:ws_nb_iters_ft: 400\nINFO:tensorflow:ws_nb_iters_feval: 25\nINFO:tensorflow:ws_prune_ratio_exp: 3.0\nINFO:tensorflow:ws_iter_ratio_beg: 0.1\nINFO:tensorflow:ws_iter_ratio_end: 0.5\nINFO:tensorflow:ws_mask_update_step: 500.0\nINFO:tensorflow:cp_lasso: True\nINFO:tensorflow:cp_quadruple: False\nINFO:tensorflow:cp_reward_policy: accuracy\nINFO:tensorflow:cp_nb_points_per_layer: 10\nINFO:tensorflow:cp_nb_batches: 30\nINFO:tensorflow:cp_prune_option: auto\nINFO:tensorflow:cp_prune_list_file: ratio.list\nINFO:tensorflow:cp_channel_pruned_path: .\/models\/pruned_model.ckpt\nINFO:tensorflow:cp_best_path: .\/models\/best_model.ckpt\nINFO:tensorflow:cp_original_path: .\/models\/original_model.ckpt\nINFO:tensorflow:cp_preserve_ratio: 0.5\nINFO:tensorflow:cp_uniform_preserve_ratio: 0.6\nINFO:tensorflow:cp_noise_tolerance: 0.15\nINFO:tensorflow:cp_lrn_rate_ft: 0.0001\nINFO:tensorflow:cp_nb_iters_ft_ratio: 0.2\nINFO:tensorflow:cp_finetune: False\nINFO:tensorflow:cp_retrain: False\nINFO:tensorflow:cp_list_group: 1000\nINFO:tensorflow:cp_nb_rlouts: 200\nINFO:tensorflow:cp_nb_rlouts_min: 50\nINFO:tensorflow:dcp_save_path: .\/models_dcp\/model.ckpt\nINFO:tensorflow:dcp_save_path_eval: .\/models_dcp_eval\/model.ckpt\nINFO:tensorflow:dcp_prune_ratio: 0.5\nINFO:tensorflow:dcp_nb_stages: 3\nINFO:tensorflow:dcp_lrn_rate_adam: 0.001\nINFO:tensorflow:dcp_nb_iters_block: 10000\nINFO:tensorflow:dcp_nb_iters_layer: 500\nINFO:tensorflow:uql_equivalent_bits: 4\nINFO:tensorflow:uql_nb_rlouts: 200\nINFO:tensorflow:uql_w_bit_min: 2\nINFO:tensorflow:uql_w_bit_max: 8\nINFO:tensorflow:uql_tune_layerwise_steps: 100\nINFO:tensorflow:uql_tune_global_steps: 2000\nINFO:tensorflow:uql_tune_save_path: .\/rl_tune_models\/model.ckpt\nINFO:tensorflow:uql_tune_disp_steps: 300\nINFO:tensorflow:uql_enbl_random_layers: True\nINFO:tensorflow:uql_enbl_rl_agent: False\nINFO:tensorflow:uql_enbl_rl_global_tune: True\nINFO:tensorflow:uql_enbl_rl_layerwise_tune: False\nINFO:tensorflow:uql_weight_bits: 4\nINFO:tensorflow:uql_activation_bits: 32\nINFO:tensorflow:uql_use_buckets: False\nINFO:tensorflow:uql_bucket_size: 256\nINFO:tensorflow:uql_quant_epochs: 60\nINFO:tensorflow:uql_save_quant_model_path: .\/uql_quant_models\/uql_quant_model.ckpt\nINFO:tensorflow:uql_quantize_all_layers: False\nINFO:tensorflow:uql_bucket_type: channel\nINFO:tensorflow:uqtf_save_path: .\/models_uqtf\/model.ckpt\nINFO:tensorflow:uqtf_save_path_eval: .\/models_uqtf_eval\/model.ckpt\nINFO:tensorflow:uqtf_weight_bits: 8\nINFO:tensorflow:uqtf_activation_bits: 8\nINFO:tensorflow:uqtf_quant_delay: 0\nINFO:tensorflow:uqtf_freeze_bn_delay: None\nINFO:tensorflow:uqtf_lrn_rate_dcy: 0.01\nINFO:tensorflow:nuql_equivalent_bits: 4\nINFO:tensorflow:nuql_nb_rlouts: 200\nINFO:tensorflow:nuql_w_bit_min: 2\nINFO:tensorflow:nuql_w_bit_max: 8\nINFO:tensorflow:nuql_tune_layerwise_steps: 100\nINFO:tensorflow:nuql_tune_global_steps: 2101\nINFO:tensorflow:nuql_tune_save_path: .\/rl_tune_models\/model.ckpt\nINFO:tensorflow:nuql_tune_disp_steps: 300\nINFO:tensorflow:nuql_enbl_random_layers: True\nINFO:tensorflow:nuql_enbl_rl_agent: False\nINFO:tensorflow:nuql_enbl_rl_global_tune: True\nINFO:tensorflow:nuql_enbl_rl_layerwise_tune: False\nINFO:tensorflow:nuql_init_style: quantile\nINFO:tensorflow:nuql_opt_mode: weights\nINFO:tensorflow:nuql_weight_bits: 4\nINFO:tensorflow:nuql_activation_bits: 32\nINFO:tensorflow:nuql_use_buckets: False\nINFO:tensorflow:nuql_bucket_size: 256\nINFO:tensorflow:nuql_quant_epochs: 60\nINFO:tensorflow:nuql_save_quant_model_path: .\/nuql_quant_models\/model.ckpt\nINFO:tensorflow:nuql_quantize_all_layers: False\nINFO:tensorflow:nuql_bucket_type: split\nINFO:tensorflow:log_dir: .\/logs\nINFO:tensorflow:enbl_multi_gpu: False\nINFO:tensorflow:learner: full-prec\nINFO:tensorflow:exec_mode: train\nINFO:tensorflow:debug: False\nINFO:tensorflow:h: False\nINFO:tensorflow:help: False\nINFO:tensorflow:helpfull: False\nINFO:tensorflow:helpshort: False\n2018-11-16 12:53:20.147847: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-16 12:53:20.148287: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 0 with properties: \nname: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235\npciBusID: 0000:00:04.0\ntotalMemory: 11.17GiB freeMemory: 11.10GiB\n2018-11-16 12:53:20.148358: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0\n2018-11-16 12:53:20.565167: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-16 12:53:20.565235: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 \n2018-11-16 12:53:20.565262: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N \n2018-11-16 12:53:20.565561: W tensorflow\/core\/common_runtime\/gpu\/gpu_bfc_allocator.cc:42] Overriding allow_growth setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.\n2018-11-16 12:53:20.565637: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 10758 MB memory) -&gt; physical GPU (device: 0, name: Tesla K80, pci bus id: 0000:00:04.0, compute capability: 3.7)\nWARNING:tensorflow:From \/content\/gdrive\/My Drive\/apps\/PocketFlow\/datasets\/abstract_dataset.py:85: parallel_interleave (from tensorflow.contrib.data.python.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.experimental.parallel_interleave(...)`.\nWARNING:tensorflow:From \/content\/gdrive\/My Drive\/apps\/PocketFlow\/datasets\/abstract_dataset.py:106: shuffle_and_repeat (from tensorflow.contrib.data.python.ops.shuffle_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.experimental.shuffle_and_repeat(...)`.\n2018-11-16 12:53:23.066723: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0\n2018-11-16 12:53:23.066814: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-16 12:53:23.066857: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 \n2018-11-16 12:53:23.066882: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N \n2018-11-16 12:53:23.067168: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 10758 MB memory) -&gt; physical GPU (device: 0, name: Tesla K80, pci bus id: 0000:00:04.0, compute capability: 3.7)\n2018-11-16 12:53:24.963790: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at matching_files_op.cc:49 : Not found: \/content\/drive\/My; No such file or directory\n2018-11-16 12:53:24.964542: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at matching_files_op.cc:49 : Not found: \/content\/drive\/My; No such file or directory\n2018-11-16 12:53:24.964744: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at iterator_ops.cc:947 : Not found: \/content\/drive\/My; No such file or directory\n     [[{{node ShuffleDataset\/data\/list_files\/MatchingFiles}} = MatchingFiles[](ShuffleDataset\/data\/list_files\/file_pattern)]]\nTraceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1334, in _do_call\n    return fn(*args)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1319, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1407, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.NotFoundError: \/content\/drive\/My; No such file or directory\n     [[{{node ShuffleDataset\/data\/list_files\/MatchingFiles}} = MatchingFiles[](ShuffleDataset\/data\/list_files\/file_pattern)]]\n     [[{{node data\/OneShotIterator}} = OneShotIterator[container=\"\", dataset_factory=_make_dataset_E02JEaYNEAE[], output_shapes=[[?,32,32,3], [?,10]], output_types=[DT_FLOAT, DT_FLOAT], shared_name=\"\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n     [[{{node data\/IteratorGetNext\/_3}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_107_data\/IteratorGetNext\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"main.py\", line 69, in &lt;module&gt;\n    tf.app.run()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"main.py\", line 55, in main\n    learner.train()\n  File \"\/content\/gdrive\/My Drive\/apps\/PocketFlow\/learners\/full_precision\/learner.py\", line 71, in train\n    self.sess_train.run(self.train_op)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1152, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1328, in _do_run\n    run_metadata)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1348, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.NotFoundError: \/content\/drive\/My; No such file or directory\n     [[{{node ShuffleDataset\/data\/list_files\/MatchingFiles}} = MatchingFiles[](ShuffleDataset\/data\/list_files\/file_pattern)]]\n     [[node data\/OneShotIterator (defined at \/content\/gdrive\/My Drive\/apps\/PocketFlow\/datasets\/abstract_dataset.py:109)  = OneShotIterator[container=\"\", dataset_factory=_make_dataset_E02JEaYNEAE[], output_shapes=[[?,32,32,3], [?,10]], output_types=[DT_FLOAT, DT_FLOAT], shared_name=\"\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n     [[{{node data\/IteratorGetNext\/_3}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_107_data\/IteratorGetNext\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53343032","CreationDate":"2018-11-16T17:53:07.323","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p>It looks like either <code>.\/scripts\/run_local.sh<\/code> or <code>nets\/resnet_at_cifar10_run.py<\/code> is passing the equivalent of <code>$PWD<\/code> to a subprocess with insufficient quoting. You could either fix that or work around it e.g. with:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/gdrive')\n!ln -s \"\/content\/gdrive\/My Drive\" \"\/content\/mydrive\"\nimport os\nos.chdir(\"\/content\/mydrive\/apps\/PocketFlow\")\n!chmod 755 .\/scripts\/run_local.sh\n!.\/scripts\/run_local.sh nets\/resnet_at_cifar10_run.py\n<\/code><\/pre>\n"}]}
+{"QuestionId":53338790,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><regression>","CreationDate":"2018-11-16T13:24:45.573","AcceptedAnswerId":53338970.0,"OwnerUserId":2651073.0,"Title":"KerasRegressor: ValueError: continuous is not supported","Body":"<p>I am trying to apply a <a href=\"https:\/\/machinelearningmastery.com\/regression-tutorial-keras-deep-learning-library-python\/\" rel=\"nofollow noreferrer\">regression learning method<\/a> to my data which has 28 dimensions. <\/p>\n\n<p>Code:<\/p>\n\n<pre><code>import numpy\nimport pandas as pd\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.wrappers.scikit_learn import KerasRegressor\nfrom sklearn.model_selection import cross_val_score\nfrom sklearn.model_selection import KFold\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.pipeline import Pipeline\nfrom sklearn.metrics import accuracy_score\nimport os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n\n# load dataset\ndataframe = pd.read_csv(\"gold_train_small.csv\", header=None)\ndataset = dataframe.values\n# split into input (X) and output (Y) variables\nX = dataset[:,1:29]\nY = dataset[:,0]\n\n# load test set\n\n# load dataset\ndataframe = pd.read_csv(\"gold_test.csv\", header=None)\ndataset = dataframe.values\n# split into input (X) and output (Y) variables\nX_test = dataset[:,1:29]\nY_test = dataset[:,0]\n\n# define base model\ndef baseline_model():\n    # create model\n    model = Sequential()\n    model.add(Dense(28, input_dim=28, kernel_initializer='normal', activation='relu'))\n    model.add(Dense(1, kernel_initializer='normal'))\n    # Compile model\n    model.compile(loss='mean_squared_error', optimizer='adam')\n    return model\n\n# fix random seed for reproducibility\nseed = 7\nnumpy.random.seed(seed)\n# evaluate model with standardized dataset\nestimator = KerasRegressor(build_fn=baseline_model, epochs=100, batch_size=5, verbose=0)\n\nkfold = KFold(n_splits=10, random_state=seed)\nresults = cross_val_score(estimator, X, Y, cv=kfold)\nprint(\"Results: %.2f (%.2f) MSE\" % (results.mean(), results.std()))\n\n#Baseline: 31.64 (26.82) MSE\n\n# evaluate model with standardized dataset\nnumpy.random.seed(seed)\nestimators = []\nestimators.append(('standardize', StandardScaler()))\nestimators.append(('mlp', KerasRegressor(build_fn=baseline_model, epochs=50, batch_size=5, verbose=0)))\npipeline = Pipeline(estimators)\nkfold = KFold(n_splits=10, random_state=seed)\nresults = cross_val_score(pipeline, X, Y, cv=kfold)\nprint(\"Standardized: %.2f (%.2f) MSE\" % (results.mean(), results.std()))\n\n\nestimator.fit(X, Y)\nprediction = estimator.predict(X_test)\naccuracy_score(Y_test, prediction)\n<\/code><\/pre>\n\n<p>However, I receive the following error for the last line:<\/p>\n\n<blockquote>\n  <p>ValueError: continuous is not supported<\/p>\n<\/blockquote>\n\n<p>Should I use other measures?<\/p>\n","answers":[{"AnswerId":"53338970","CreationDate":"2018-11-16T13:37:49.583","ParentId":null,"OwnerUserId":"3339273","Title":null,"Body":"<p>This is because accuracy_score is for a classification model.<\/p>\n\n<p>As this is a regression model, you should try using:<\/p>\n\n<pre><code>estimator.score(X_test,Y_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53338975,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-probability>","CreationDate":"2018-11-16T13:38:02.060","AcceptedAnswerId":53487545.0,"OwnerUserId":9478404.0,"Title":"Use and modify variables in tensorflow bijectors","Body":"<p>In the reference <a href=\"https:\/\/arxiv.org\/abs\/1711.10604\" rel=\"nofollow noreferrer\">paper<\/a> for TensorFlow Distributions (now <a href=\"https:\/\/www.tensorflow.org\/probability\/\" rel=\"nofollow noreferrer\">Probability<\/a>), it is mentioned that TensorFlow <code>Variable<\/code>s can be used to construct <code>Bijector<\/code> and <code>TransformedDistribution<\/code> objects, i.e.:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\n\ntf.enable_eager_execution()\n\nshift = tf.Variable(1., dtype=tf.float32)\nmyBij = tfp.bijectors.Affine(shift=shift)\n\n# Normal distribution centered in zero, then shifted to 1 using the bijection\nmyDistr = tfd.TransformedDistribution(\n            distribution=tfd.Normal(loc=0., scale=1.),\n            bijector=myBij,\n            name=\"test\")\n\n# 2 samples of a normal centered at 1:\ny = myDistr.sample(2)\n# 2 samples of a normal centered at 0, obtained using inverse transform of myBij:\nx = myBij.inverse(y)\n<\/code><\/pre>\n\n<p>I would now like to modify the shift variable (say, I might compute gradients of some likelihood function as a function of the shift and update its value) so I do<\/p>\n\n<pre><code>shift.assign(2.)\ngx = myBij.forward(x)\n<\/code><\/pre>\n\n<p>I would expect that <code>gx=y+1<\/code>, but I see that <code>gx=y<\/code>... And indeed, <code>myBij.shift<\/code> still evalues to <code>1<\/code>.<\/p>\n\n<p>If I try to modify the bijector directly, i.e.:<\/p>\n\n<pre><code>myBij.shift.assign(2.)\n<\/code><\/pre>\n\n<p>I get<\/p>\n\n<pre><code>AttributeError: 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'assign'\n<\/code><\/pre>\n\n<p>Computing gradients also does not work as expected:<\/p>\n\n<pre><code>with tf.GradientTape() as tape:\n    gx = myBij.forward(x)\ngrad = tape.gradient(gx, shift)\n<\/code><\/pre>\n\n<p>Yields <code>None<\/code>, as well as this exception when the script ends:<\/p>\n\n<pre><code>Exception ignored in: &lt;bound method GradientTape.__del__ of &lt;tensorflow.python.eager.backprop.GradientTape object at 0x7f529c4702e8&gt;&gt;\nTraceback (most recent call last):\nFile \"~\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/backprop.py\", line 765, in __del__\nAttributeError: 'NoneType' object has no attribute 'context'\n<\/code><\/pre>\n\n<p>What am I missing here?<\/p>\n\n<p>Edit: I got it working with a graph\/session, so it seems there is an issue with eager execution...<\/p>\n\n<p>Note: I have tensorflow version 1.12.0 and tensorflow_probability version 0.5.0<\/p>\n","answers":[{"AnswerId":"53487545","CreationDate":"2018-11-26T19:10:21.300","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>If you are using eager mode, you will need to recompute everything from the variable forward. Best to capture this logic in a function;<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\n\ntf.enable_eager_execution()\n\nshift = tf.Variable(1., dtype=tf.float32)\ndef f():\n  myBij = tfp.bijectors.Affine(shift=shift)\n\n  # Normal distribution centered in zero, then shifted to 1 using the bijection\n  myDistr = tfd.TransformedDistribution(\n            distribution=tfd.Normal(loc=0., scale=1.),\n            bijector=myBij,\n            name=\"test\")\n\n  # 2 samples of a normal centered at 1:\n  y = myDistr.sample(2)\n  # 2 samples of a normal centered at 0, obtained using inverse\n  # transform of myBij:\n  x = myBij.inverse(y)\n  return x, y\nx, y = f()\nshift.assign(2.)\ngx, _ = f()\n<\/code><\/pre>\n\n<p>Regarding gradients, you will need to wrap calls to <code>f()<\/code> in a <code>GradientTape<\/code><\/p>\n"}]}
+{"QuestionId":53339224,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-16T13:53:31.927","AcceptedAnswerId":53339363.0,"OwnerUserId":8236090.0,"Title":"pytorch save a minibatch of 4D tensor as images","Body":"<p>I have set of 8 tensors in 4D tensor of shape (B, C, H, W). More specifically, the shape of my tensor is (8,3,64,64) i.e 8 images with [3x64x64] format.<\/p>\n\n<p>I would like to save these in 1.png, 2.png ..., 8.png etc. When I try <code>torchvision.utils.save_images(my_tensor)<\/code>, this is working fine. However, that is saving the images in a single grid instead of individual images.<\/p>\n\n<p>How can I fix this issue?<\/p>\n\n<p>I tried below code but that did not work.<\/p>\n\n<pre><code>for i in range(tensor.size(0)):\n\n        np_data = tensor.cpu().numpy()\n        imgplot = plt.imshow(np_data)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53339363","CreationDate":"2018-11-16T14:02:54.017","ParentId":null,"OwnerUserId":"1415817","Title":null,"Body":"<p>If you want to save individual images you could use:<\/p>\n\n<pre><code>for i in range(tensor.size(0)):\n           torchvision.utils.save_image(tensor[i, :, :, :], '{}.png'.format(i))\n<\/code><\/pre>\n\n<p>which will save as : 1.png, 2.png ..., 8.png<\/p>\n"}]}
+{"QuestionId":53340036,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><embedding><tfrecord><transfer-learning>","CreationDate":"2018-11-16T14:43:31.803","AcceptedAnswerId":53364661.0,"OwnerUserId":7744356.0,"Title":"How to decode vggish audioset embeddings from tfrecord?","Body":"<p>I am trying to use the 128 byte embeddings produced by the pre-trained base of the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/audioset\" rel=\"nofollow noreferrer\">VGGish<\/a> model for transfer learning on audio data. Using <code>python vggish_inference_demo.py --wav_file ...<\/code> to encode my training data to a tfrecord worked fine, but now I want to use this as an input to another model (e.g. a neural network I create with keras or something else). Using some <a href=\"https:\/\/stackoverflow.com\/questions\/46204992\/how-can-i-extract-the-audio-embeddings-features-from-google-s-audioset\">similar questions<\/a> and the documentation, I go this far with the first embedding record of one file:<\/p>\n\n<pre><code>tfrecords_filename = 'example1.tfrecord'\nrecord_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)\nstring_record = next(record_iterator)\nexample = tf.train.SequenceExample()\nexample.ParseFromString(string_record)\nprint(example.feature_lists.feature_list['audio_embedding'].feature[0].bytes_list.value)\n<\/code><\/pre>\n\n<p>This produces<\/p>\n\n<pre><code>[b'\\x99\\x07\\xaa&gt;\\xd2_R_\\x9f\\xbbqN\\x99\\xa18V\\xad\\x7f\\x93\\xf0)\\xdd4\\x80~\\xb0\\xa4d\\x8e\\x85\\xb6\\x88\\xa3?U\\xa6Q[\\x9b\\x038\\xff\\x00EE&gt;OJ\\xa5\\xb8\\x828)\\x97^\\x8a\\xaa\\x12h\\xff\\xff\\xc39\\xce\\x9b\\x13\\x80\\x00j\\xcaZ\\xac\\xff\\xff\\x0f\\xac\\x1c\\x90&amp;\\xd2.b\\xe2{\\xc1\\x15\\xe9\\xba\\xed\\xd4\\xa9\\xff\\xdc\\xb5\\x99]!\\x04\\xca\\xff\\xa6;b\\xe0\\x19\\xbfW\\xebP!\\xff\\xc5\\xff\\x82\\xff\\x1a\\xbe\\xec-h\\xff\\x8d\\xff\\r\\x96\\x00\\x00\\xff']\n<\/code><\/pre>\n\n<p>I am not even sure what this <code>b'...'<\/code> is (there's more than 64 and fewer than 128 xs - so not sure how this lines up with anything).<\/p>\n\n<p>Maybe I am missing some basic Python knowledge here, but how do I turn this into a numeric array of numbers that I can use as an input to some other model? <\/p>\n","answers":[{"AnswerId":"53364661","CreationDate":"2018-11-18T19:29:05.880","ParentId":null,"OwnerUserId":"7744356","Title":null,"Body":"<p>It turns out that these are bytes that can be converted to hex, which can be converted to an array of integers between 0 to 255.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntfrecords_filename = 'example1.tfrecord'\nrecord_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)\nstring_record = next(record_iterator)\nexample = tf.train.SequenceExample()\nexample.ParseFromString(string_record)\nhexembed = example.feature_lists.feature_list['audio_embedding'].feature[0].bytes_list.value[0].hex()\narrayembed = [int(hexembed[i:i+2],16) for i in range(0,len(hexembed),2)]\nprint(arrayembed)\n<\/code><\/pre>\n\n<p>This produces output in the format that I desired:<\/p>\n\n<blockquote>\n  <p>[153, 7, 170, 62, 210, 95, 82, 95, 159, 187, 113, 78, 153, 161, 56,\n  86, 173, 127, 147, 240, 41, 221, 52, 128, 126, 176, 164, 100, 142,\n  133, 182, 136, 163, 63, 85, 166, 81, 91, 155, 3, 56, 255, 0, 69, 69,\n  62, 79, 74, 165, 184, 130, 56, 41, 151, 94, 138, 170, 18, 104, 255,\n  255, 195, 57, 206, 155, 19, 128, 0, 106, 202, 90, 172, 255, 255, 15,\n  172, 28, 144, 38, 210, 46, 98, 226, 123, 193, 21, 233, 186, 237, 212,\n  169, 255, 220, 181, 153, 93, 33, 4, 202, 255, 166, 59, 98, 224, 25,\n  191, 87, 235, 80, 33, 255, 197, 255, 130, 255, 26, 190, 236, 45, 104,\n  255, 141, 255, 13, 150, 0, 0, 255]<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53340597,"AnswerCount":1,"Tags":"<python><tensorflow><keras><keras-layer>","CreationDate":"2018-11-16T15:15:34.043","AcceptedAnswerId":53392349.0,"OwnerUserId":10663471.0,"Title":"OR-Lambda-Layer operation with Keras","Body":"<p>I'm creating a different project with Keras, it's a development of a Neural network based on predefined knowledge (described as IF-THEN rules), called Neurules. I've created a Python Module to train each one of my Neurons\/Neurules from a given IF-THEN logical expression and after all I need to use Keras to create it as a network and reuse this Model.<\/p>\n\n<p>I've already tested with one little example and it worked, everything manually added, the weights and bias. Now I've updated my script and it is giving me a JSON with all the weights to add to Keras (working until now).<\/p>\n\n<p>Here comes my problem, I have one first Layer with 20 Neurules (neurons created from IF-THENs) but I only have 2 possible outputs, some of the Neurules\/Neurons give me <code>output[0]<\/code> some of them the <code>output[1]<\/code>, I want to add a layer in-between representing the OR connections.<\/p>\n\n<p>E.g.:<\/p>\n\n<p>Layer 1:\nNEURON1, NEURON2, NEURON3<\/p>\n\n<p><code>Output[0]<\/code> is formed by: <code>NEURON1 or NEURON2<\/code><\/p>\n\n<p><code>Output[1]<\/code> is formed by: <code>NEURON2 or NEURON3<\/code><\/p>\n\n<p>What I've done in my first little example was: I've created and trained an OR neuron with my pre-developed python module and then added a second layer with it. I've then connected the entries manually to the ORs Neurules (putting the weights in the correct connections and putting 0 when they shouldn't influence the OR). Now I have something bigger and I'm automating the whole process.<\/p>\n\n<p>Visualization of the simple net:\nBuffer is just forwarding the value, OR is doing an OR operation with the inputs.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yTFfM.png\" rel=\"nofollow noreferrer\">Visualization of the layers<\/a><\/p>\n\n<p>How can I create a Lambda Layer in Keras which takes some of the outputs, processes a logical OR and connects to one of the outputs?<\/p>\n\n<p>I've found the Backend function: \n<code>tf.keras.backend.any<\/code>, but I'm not able to use it until now, how should I use it? Probably in a Lambda layer, but how?<\/p>\n\n<p>I need to connect, for example<\/p>\n\n<p><code>(NEURON1 or NEURON4 or NEURON5)<\/code> -> <code>output[0]<\/code><\/p>\n\n<p><code>(NEURON3 or NEURON6 or NEURON7)<\/code> -> <code>output[1]<\/code><\/p>\n\n<p>In my system -1 represents False and 1 represents True.\nUntil now I've saved which Neurons are using each one of the 2 outputs in an array in a JSON like:<\/p>\n\n<pre><code>\"secondLayerDescription\": [\n    [0, 1, 4, 5, 6, 8, 12, 13, 14, 16, 18], \n    [2, 3, 7, 9, 10, 11, 15, 17, 19]\n]\n<\/code><\/pre>\n\n<p>I hope someone can help me :)<\/p>\n\n<p><strong>EDIT:<\/strong> Giving an update, I found a solution after some days, I'm splitting my layer into 2 layers and operating them with lambda layers as following:<\/p>\n\n<pre><code>def logical_or_layer(x):\n    \"\"\"Processing an OR operation\"\"\"\n    import keras.backend\n    #normalized to 0,1 \n    aux_array = keras.backend.sign(x)\n    aux_array = keras.backend.relu(aux_array)\n    # OR operation\n    aux_array = keras.backend.any(aux_array)\n    # casting back the True\/False to 1,0\n    aux_array = keras.backend.cast(aux_array, dtype='float32')\n\n    return aux_array\n\n\n#this is the input tensor\ninputs = Input(shape=(inputSize,))\n\n#this is the Neurule layer\nx = Dense(neurulesQt, activation='softsign')(inputs)\n\n#after each neuron layer, the outputs need to be put into SIGNUM (-1 or 1)\nx = Lambda(signumTransform, output_shape=lambda x:x, name='signumAfterNeurules')(x)\n\n#separating into 2 (2 possible outputs)\nlayer_split0 = Lambda( lambda x: x[:, :end_output0], output_shape=(11, ), name='layer_split0')(x)\nlayer_split1 = Lambda( lambda x: x[:, start_output1:end_output1], output_shape=(9,), name='layer_split1')(x)\n#this is the OR layer\ny_0 = Lambda(logical_or_layer, output_shape=(1,), name='or0')(layer_split0)\ny_1 = Lambda(logical_or_layer, output_shape=(1,), name='or1')(layer_split1)\n<\/code><\/pre>\n\n<p>But I'm still having problems, I can't merge them together, I've raised a new question based on this new topic.<\/p>\n","answers":[{"AnswerId":"53392349","CreationDate":"2018-11-20T11:47:29.850","ParentId":null,"OwnerUserId":"10663471","Title":null,"Body":"<p>I found a way to do it, I need to sort my Neurules Layer, split them and then, with one Lambda Layer for each split, do some processing as shown in the EDITed part of the question: normalize the inputs, use the <code>backend.any<\/code> and then cast the <code>True<\/code> or <code>False<\/code> back to float.<\/p>\n"}]}
+{"QuestionId":53340700,"AnswerCount":0,"Tags":"<python><tensorflow><flask><gunicorn><tensorflow-serving>","CreationDate":"2018-11-16T15:21:16.740","AcceptedAnswerId":null,"OwnerUserId":926272.0,"Title":"Performance penalty for Tensorflow Serving in docker hosted app","Body":"<p>I am experiencing a large performance penalty for calls to Tensorflow Serving, when the calling app is hosted in a docker container. I was hoping someone would have some suggestions. Details below on the setup and what I have tried.<\/p>\n\n<p>Scenario 1:<\/p>\n\n<ul>\n<li>Docker (version 18.09.0, build 4d60db4) hosted Tensorflow model, following the instructions <a href=\"https:\/\/www.tensorflow.org\/serving\/docker\" rel=\"nofollow noreferrer\">here<\/a>.<\/li>\n<li>Flask app running on the host machine (not in container).<\/li>\n<li>Using gRPC for sending the request to the model.<\/li>\n<li>Performance: 0.0061 seconds \/ per prediction<\/li>\n<\/ul>\n\n<p>Scenario 2:<\/p>\n\n<ul>\n<li>Same docker container hosted Tensorflow model.<\/li>\n<li>Container hosted Flask app running on the host machine (inside same container as the model).<\/li>\n<li>Using gRPC for sending the request to the model.<\/li>\n<li>Performance: 0.0107 \/ per prediction<\/li>\n<\/ul>\n\n<p>In other words, when the app is hosted in the same container as the model, performance is ~40% lower.<\/p>\n\n<p>I have logged timing on nearly every step in the app and have tracked the difference down to this line:<\/p>\n\n<pre><code>result = self.stub.Predict(self.request, 60.0)\n<\/code><\/pre>\n\n<p>In the container hosted app, the average round-trip for this is 0.006 seconds. For the same app hosted outside the container, the round-trip for this line is 0.002 seconds.<\/p>\n\n<p>This is the function I am using to establish the connection to the model.<\/p>\n\n<pre><code>def TFServerConnection():\n    channel = implementations.insecure_channel('127.0.0.1', 8500)\n    stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)\n    request = predict_pb2.PredictRequest()\n    return (channel, stub, request)\n<\/code><\/pre>\n\n<p>I have tried hosting the app and model in separate containers, building a custom Tensorflow Serving containers (optimized for my VM), and using the REST api (which decreased performance slightly for both scenarios).<\/p>\n\n<p><strong>Edit 1<\/strong><\/p>\n\n<p>To add a bit more information, I am running the docker container with the following command:<\/p>\n\n<pre><code>docker run \\\n    --detach \\\n    --publish 8000:8000 \\\n    --publish 8500:8500 \\\n    --publish 8501:8501 \\\n    --name tfserver \\\n    --mount type=bind,source=\/home\/jason\/models,target=\/models \\\n    --mount type=bind,source=\/home\/jason\/myapp\/serve\/tfserve.conf,target=\/config\/tfserve.conf \\\n    --network host \\\n    jason\/myapp:latest\n<\/code><\/pre>\n\n<p><strong>Edit 2<\/strong><\/p>\n\n<p>I have now tracked this down to being an issue with  <code>stub.Predict(request, 60.0)<\/code> in Flask apps only. It seems Docker is not the issue. Here are the versions of Flask and Tensorflow I am currently running.<\/p>\n\n<pre><code>$ sudo pip3 freeze | grep Flask\nFlask==1.0.2\n\n$ sudo pip3 freeze | grep tensor\ntensorboard==1.12.0\ntensorflow==1.12.0\ntensorflow-serving-api==1.12.0\n<\/code><\/pre>\n\n<p>I am using gunicorn as my WSGI server:<\/p>\n\n<pre><code>gunicorn --preload --config config\/gunicorn.conf app:app\n<\/code><\/pre>\n\n<p>And the contents of config\/gunicorn.conf:<\/p>\n\n<pre><code>bind = \"0.0.0.0:8000\"\nworkers = 3\ntimeout = 60\nworker_class = 'gevent'\nworker_connections = 1000\n<\/code><\/pre>\n\n<p><strong>Edit 3<\/strong><\/p>\n\n<p>I have now narrowed the issue down to Flask. I ran the Flask app directly with <code>app.run()<\/code> and got the same performance as when using gunicorn. What could Flask be doing that would slow the call to Tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":53340723,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network><pytorch><onnx>","CreationDate":"2018-11-16T15:22:18.000","AcceptedAnswerId":null,"OwnerUserId":3297229.0,"Title":"Conversion from pytorch to onnx to tensorflow graph definition to tflite - TOCO failed - type check fail","Body":"<p>I have a simple network of two convolutional layers and one fully connected one defined in pytorch essentially as follows:<\/p>\n\n<pre><code>def __init__([...])\n    [...]\n    self.conv1 = nn.Conv1d(1, channels_conv1, width_conv1)\n    self.conv2 = nn.Conv1d(channels_conv1, channels_conv2, width_conv2)\n    self.fc1 = nn.Linear(hidden_layer_size, 2)\n\ndef forward(self, x):\n    x = functional.max_pool1d(functional.relu(self.conv1(x)), 2, stride=2)\n    x = functional.max_pool1d(functional.relu(self.conv2(x)), 2, stride=2)\n    x = x.view(-1, self.num_flat_features(x))\n    x = functional.softmax(self.fc1(x))\n    return x\n<\/code><\/pre>\n\n<p>I want to convert it to tflite. So first it is converted to onnx <\/p>\n\n<pre><code>torch.onnx.export(model, input, \"net.onnx\",\n                  export_params=True,\n                  input_names=['input'],\n                  output_names=['output'],\n                  verbose=true)\n<\/code><\/pre>\n\n<p>Then I convert the result to a tensorflow graph definition with <code>onnx-tf<\/code>. The resulting <code>net.pb<\/code> is ok as it produces the same output as the original with <code>prepare(onnx.load('net.onnx')).run(...)<\/code>.<\/p>\n\n<p>However, I have two problems: a minor one is that the <code>net.pb<\/code> graph doesn't contain the output node anymore and I have to fish for the output node. \nThe second is that when I try to perform the final conversion with<\/p>\n\n<pre><code>tflite_convert --output_file=net.tflite --graph_def_file=net.pb --input_arrays=input --output_arrays=Softmax\n<\/code><\/pre>\n\n<p>I get a TOCO failure at a type check:<\/p>\n\n<pre><code>tensorflow.lite.python.convert.ConverterError: TOCO failed. See console for info.\n2018-11-16 16:11:37.592030: I tensorflow\/lite\/toco\/import_tensorflow.cc:1280] Converting unsupported operation: Where\n2018-11-16 16:11:37.601384: I tensorflow\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before Removing unused ops: 61 operators, 113 arrays (0 quantized)\n2018-11-16 16:11:37.602005: I tensorflow\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before general graph transformations: 61 operators, 113 arrays (0 quantized)\n2018-11-16 16:11:37.602311: F tensorflow\/lite\/toco\/graph_transformations\/resolve_constant_gather.cc:105] Check failed: coords_array.data_type == ArrayDataType::kInt32 Only int32 indices are supported\nAborted (core dumped)\n<\/code><\/pre>\n\n<p>I tried digging in the network but I can't seem to find the vexing object and I didn't find issues obviously related to this problem. Any pointer to a point where this process might have derailed would be a great!<\/p>\n\n<pre><code>tf-nightly==1.13.0.dev20181116\nonnx==1.3.0\ntorch-nightly==1.0.0.dev201811\n<\/code><\/pre>\n\n<p>and <code>master<\/code> (commit <code>b5fef1b<\/code>) of <a href=\"https:\/\/github.com\/onnx\/onnx-tensorflow\/commit\/b5fef1b6471ad6c5bac214e9799c38d41658ee73\" rel=\"nofollow noreferrer\">onnx-tensorflow<\/a><\/p>\n","answers":[]}
+{"QuestionId":53341653,"AnswerCount":0,"Tags":"<tensorflow><keras><keras-2><vgg-net>","CreationDate":"2018-11-16T16:14:00.320","AcceptedAnswerId":null,"OwnerUserId":3837778.0,"Title":"why my below modified vgg-16 Keras model takes 1 hour to run one epoch","Body":"<p>My model is<\/p>\n\n<pre><code>Using TensorFlow backend.\nFound 8704 images belonging to 68 classes.\nFound 2176 images belonging to 68 classes.\nFound 1360 images belonging to 68 classes.\n\n_________________________________________________________________\nNone\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nvgg16 (Model)                (None, 4, 4, 512)         14714688  \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 8192)              0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 256)               2097408   \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 256)               0         \n_________________________________________________________________\ndense_2 (Dense)              (None, 68)                17476     \n=================================================================\nTotal params: 16,829,572\nTrainable params: 3,850,372\nNon-trainable params: 12,979,200\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Each image is 128*128*3. There are 8 CPUs on my laptop and cpu usage is about 700%. why it takes about 1 hour to run one epoch? How to improve the performance? Thanks<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Below is details of my model:<\/p>\n\n<pre><code>vgg16 = VGG16(include_top=False,\n              weights='imagenet',\n              input_tensor=None,\n              input_shape=(IMG_SIZE, IMG_SIZE, CHANNELS),\n              pooling=None,\n              classes=68)\n\n\nfor layer in vgg16.layers[-8:-1]:\n    layer.trainable = False\n\nmodel = Sequential()\n\nmodel.add(vgg16)\nmodel.add(Flatten())\nmodel.add(Dense(256, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(68, activation='softmax'))\n\n\nprint(model.summary())\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53342619,"AnswerCount":1,"Tags":"<ios><objective-c><tensorflow><xcode10>","CreationDate":"2018-11-16T17:22:39.597","AcceptedAnswerId":null,"OwnerUserId":9643695.0,"Title":"TensorFlow register.h file not found in iOS project","Body":"<p>I am building the TensorFlow iOS project and having just this one issue. <\/p>\n\n<pre><code>#include \"tensorflow\/lite\/kernels\/register.h\"\n<\/code><\/pre>\n\n<p>Xcode error states the file cannot be found. I am able to find the actual file, but am not sure how to \"assign\" it to the project? I followed the steps online, and everything else went okay. <\/p>\n\n<p>Is this a common \/ known bug? <\/p>\n","answers":[{"AnswerId":"54450563","CreationDate":"2019-01-30T22:32:07.377","ParentId":null,"OwnerUserId":"10905823","Title":null,"Body":"<p>In Build Phases > Link Binary with Libraries include the libtensorflow-lite.a file from tensorflow\/lite\/tools\/make\/gen\/lib\/libtensorflow-lite.a (you probably need to go to \"Add Other...\" to do it). Include tensorflow\/lite\/tools\/make\/gen\/lib as a Library search path. Then, make sure your header search paths contain:<\/p>\n\n<ul>\n<li>the root folder of tensorflow,<\/li>\n<li>tensorflow\/lite\/downloads<\/li>\n<li>tensorflow\/lite\/tools\/make\/downloads\/flatbuffers<\/li>\n<li>tensorflow\/lite\/tools\/make\/downloads\/flatbuffers\/include\/flatbuffers<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53342795,"AnswerCount":2,"Tags":"<tensorflow><gpu><cpu>","CreationDate":"2018-11-16T17:35:29.013","AcceptedAnswerId":null,"OwnerUserId":10624301.0,"Title":"How can I check whether I use CPU or GPU in TensorFlow?","Body":"<p>I've read that<\/p>\n\n<pre><code>os.environ[\"CUDA_VISIBLE_DEVICES\"] = ''\n<\/code><\/pre>\n\n<p>takes care that tensorflow will run on CPU and that<\/p>\n\n<pre><code>os.environ[\"CUDA_VISIBLE_DEVICES\"] = '0'\n<\/code><\/pre>\n\n<p>takes care that tensorflow will run on GPU 0.<\/p>\n\n<p>How can I check, which device is used?<\/p>\n\n<p>The code<\/p>\n\n<pre><code># Creates a graph.\na = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\nb = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\nc = tf.matmul(a, b)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>shows only the result<\/p>\n\n<pre><code>[[ 22.  28.]\n [ 49.  64.]]\n<\/code><\/pre>\n\n<p>and no used device etc.<\/p>\n","answers":[{"AnswerId":"53347717","CreationDate":"2018-11-17T02:42:33.463","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>You should be able to do this by turning on the tensorflow logging statements.  There's a few ways to do this.  You can do it with a bash environment variable with..<\/p>\n\n<pre><code>export TF_CPP_MIN_LOG_LEVEL=1\n<\/code><\/pre>\n\n<p>or from within your code with..<\/p>\n\n<pre><code>tf.logging.set_verbosity(tf.logging.INFO)\n<\/code><\/pre>\n\n<p>On my system I get something like...<\/p>\n\n<blockquote>\n  <p>Device mapping: \/job:localhost\/replica:0\/task:0\/device:XLA_CPU:0 ->\n  device: XLA_CPU device\n  \/job:localhost\/replica:0\/task:0\/device:XLA_GPU:0 -> device: XLA_GPU\n  device \/job:localhost\/replica:0\/task:0\/device:GPU:0 -> device: 0,\n  name: GeForce GTX TITAN X, pci bus id: 0000:65:00.0, compute\n  capability: 5.2 MatMul: (MatMul):\n  \/job:localhost\/replica:0\/task:0\/device:GPU:0 a: (Const):\n  \/job:localhost\/replica:0\/task:0\/device:GPU:0 b: (Const):\n  \/job:localhost\/replica:0\/task:0\/device:GPU:0 [[22. 28.]  [49. 64.]]<\/p>\n<\/blockquote>\n"},{"AnswerId":"53361984","CreationDate":"2018-11-18T14:31:44.457","ParentId":null,"OwnerUserId":"10624301","Title":null,"Body":"<p>I found that there are differences between Python and IPython. IPython is the used Kernel in Spyder. So I'm guessing that's the reason for those differing outputs.<\/p>\n"}]}
+{"QuestionId":53343346,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-11-16T18:17:37.260","AcceptedAnswerId":null,"OwnerUserId":10664338.0,"Title":"ValueError: A `Concatenate` layer requires inputs with matching shapes","Body":"<pre><code>    from fr_utils import *\n    from inception_blocks_v2 import *\n\n    def triplet_loss(y_true, y_pred, alpha=0.3):\n        \"\"\"\n        Implementation of the triplet loss as defined by formula (3)\n\n        Arguments:\n        y_pred -- python list containing three objects:\n                anchor -- the encodings for the anchor images, of shape (None, 128)\n                positive -- the encodings for the positive images, of shape (None, 128)\n                negative -- the encodings for the negative images, of shape (None, 128)\n\n        Returns:\n        loss -- real number, value of the loss\n        \"\"\"\n\n        anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2]\n\n        # Step 1: Compute the (encoding) distance between the anchor and the positive, you will need to sum over axis=-1\n        pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), axis=-1)\n        # Step 2: Compute the (encoding) distance between the anchor and the negative, you will need to sum over axis=-1\n        neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), axis=-1)\n        # Step 3: subtract the two previous distances and add alpha.\n        basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), alpha)\n        # Step 4: Take the maximum of basic_loss and 0.0. Sum over the training examples.\n        loss = tf.reduce_sum(tf.maximum(basic_loss, 0.0))\n\n        return loss\n\n    def main():\n        FRmodel = faceRecoModel(input_shape=(3, 96, 96))\n        FRmodel.compile(optimizer='adam', loss=triplet_loss, metrics=['accuracy'])\n        FRmodel.save('face-rec_Google.h5')\n        print_summary(model)\n\n    main()\n<\/code><\/pre>\n\n<p>The error shown in this code is as follows  <\/p>\n\n<p>Got inputs shapes: %s' % (input_shape))<\/p>\n\n<p>ValueError: A <code>Concatenate<\/code> layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 128, 12, 192), (None, 32, 12, 192), (None, 32, 12, 102), (None, 64, 12, 192)]<\/p>\n\n<p>I try looking on the internet for he error but did not find a solution<\/p>\n","answers":[{"AnswerId":"58929654","CreationDate":"2019-11-19T08:32:24.663","ParentId":null,"OwnerUserId":"12396282","Title":null,"Body":"<p>You just need to change representation for your images. I guess, that your images are represented as 3-dimensional array <em>[rows][cols][channels]<\/em> where color channel is in last dimension. This code will move color channel into first dimension <em>[channels][rows][cols]<\/em>:<\/p>\n\n<pre><code>from keras import backend as K\nK.set_image_data_format('channels_first')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53344172,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-16T19:24:32.747","AcceptedAnswerId":54736689.0,"OwnerUserId":1316501.0,"Title":"Tensorflow: tf.contrib deprecation","Body":"<p>I've been doing some NLP using tensorflow and have been using <code>tf.contrib.lookup.index_table_from_tensor<\/code> to create a look-up table for tokens.<\/p>\n\n<p>I know tensorflow is coming out with 2.0 in Spring '19 and they've said:<\/p>\n\n<p>\"tf.contrib<\/p>\n\n<p>TensorFlow\u2019s contrib module has grown beyond what can be maintained and supported in a single repository. Larger projects are better maintained separately, while we will incubate smaller extensions along with the main TensorFlow code. Consequently, as part of releasing TensorFlow 2.0, we will stop distributing tf.contrib. We will work with the respective owners on detailed migration plans in the coming months, including how to publicise your TensorFlow extension in our community pages and documentation. For each of the contrib modules we will either a) integrate the project into TensorFlow; b) move it to a separate repository or c) remove it entirely. This does mean that all of tf.contrib will be deprecated, and we will stop adding new tf.contrib projects today. We are looking for owners\/maintainers for a number of projects currently in tf.contrib, please contact us (reply to this email) if you are interested.\"<\/p>\n\n<p>Two questions then:<\/p>\n\n<p>1) Is there another way to build a lookup-table easily in Tensorflow?<\/p>\n\n<p>2) I couldn't determine how likely it is that this functionality will be included in 2.0.  Is there any indication?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"54736689","CreationDate":"2019-02-17T19:12:04.573","ParentId":null,"OwnerUserId":"1123955","Title":null,"Body":"<ol>\n<li>You should be able to use the <code>lookup<\/code> functionality under tensorflow 2.0.<\/li>\n<li>The <code>tensorflow.contrib.lookup<\/code> will move to the core after tensorflow 2.0 <\/li>\n<\/ol>\n\n<p>According to <a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md#list-of-projects\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md#list-of-projects<\/a>.<\/p>\n\n<blockquote>\n  <p>List of Projects\n  This section lists the possible fate of all projects currently in tf.contrib. The decisions here are preliminary, not all owners have been contacted (those marked with a * have not). Named in parentheses are people with a potential interest in this, but who are not owners.<\/p>\n  \n  <p>If you have an interest in one of these projects and are not named as an owner (or as an interested party), please speak up. I've added some (tentative) move targets to as of yet nonexistent repositories: tensorflow\/io, tensorflow\/contrib (I would prefer a better name for this), and tensorflow\/scientific will all require new SIGs to maintain them.<\/p>\n  \n  <p>Note that any project that has no other fate specified (including those moving to a repo tentatively maintained by a SIG, but for which no SIG is actually formed) will be deleted by default at 2.0.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53345043,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><generative-adversarial-network>","CreationDate":"2018-11-16T20:37:24.630","AcceptedAnswerId":53367183.0,"OwnerUserId":2414957.0,"Title":"DCGAN Tensorflow code doesn't produce faces on CelebA dataset","Body":"<p>I have written the following code but does not produce faces on the celebA dataset. I think it should create some sort of face (even if very blurry) at the last iteration of each epoch. However, it just creates noisy squares with no visible face. I am quite new to GANs and I am not sure how to debug this Deep Convolutional GAN (DCGAN) in order to figure what's going wrong.<\/p>\n\n<p>My code might be easier to be seen here:\n<a href=\"https:\/\/pastebin.com\/c4QUqxJy\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/c4QUqxJy<\/a> \nHere is the code:<\/p>\n\n<pre><code>from __future__ import print_function\nimport random\nimport os\nimport glob\nimport scipy\n\nimport tensorflow as tf\nimport numpy as np\nfrom PIL import Image\nimport skimage.io as io\nimport matplotlib.pyplot as plt\n\n\nclass Arguments(object):\n\n    data_path = 'results_celebA\/preprocessed\/'\n    save_path = 'results_celebA'                           #path to save preprocessed image folder\n    preproc_foldername = 'preprocessed'      #folder name for preprocessed images\n    image_size = 64                          #images are resized to image_size value \n    num_images = 202590                      #the number of training images\n    batch_size = 64                          #batch size\n    dim_z = 100                              #the dimension of z variable (the generator input dimension)        \n    n_g_filters = 64                         #the number of the generator filters (gets multiplied between layers)\n    n_f_filters = 64                         #the number of the discriminator filters (gets multiplied between layers)           \n    n_epoch = 25                             #the number of epochs\n    lr = 0.0002                              #learning rate\n    beta1 = 0.5                              #beta_1 parameter of Adam optimizer\n    beta2 = 0.99                             #beta_2 parameter of Adam optimizer\n\nargs = Arguments()\n\n\n#contains functions that load, preprocess and visualize images. \n\n\nclass Dataset(object):     \n    def __init__(self, data_path, num_imgs, target_imgsize):\n        self.data_path = data_path\n        self.num_imgs = num_imgs \n        self.target_imgsize = target_imgsize \n\n    def normalize_np_image(self, image):\n        return (image \/ 255.0 - 0.5) \/ 0.5\n\n    def denormalize_np_image(self, image):\n        return (image * 0.5 + 0.5) * 255\n\n    def get_input(self, image_path):\n        image = np.array(Image.open(image_path)).astype(np.float32)\n        return self.normalize_np_image(image)\n\n    def get_imagelist(self, data_path, celebA=False): \n        if celebA == True:\n            imgs_path = os.path.join(data_path, 'img_align_celeba\/*.jpg')\n        else:\n            imgs_path = os.path.join(data_path, '*.jpg') \n        all_namelist = glob.glob(imgs_path, recursive=True)\n        return all_namelist[:self.num_imgs]\n\n    def load_and_preprocess_image(self, image_path): \n        image = Image.open(image_path)\n        j = (image.size[0] - 100) \/\/ 2\n        i = (image.size[1] - 100) \/\/ 2\n        image = image.crop([j, i, j + 100, i + 100])    \n        image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)\n        image = np.array(image.convert('RGB')).astype(np.float32)\n        image = self.normalize_np_image(image)\n        return image    \n\n    #reads data, preprocesses and saves to another folder with the given path. \n    def preprocess_and_save_images(self, dir_name, save_path=''): \n        preproc_folder_path = os.path.join(save_path, dir_name)\n        if not os.path.exists(preproc_folder_path):\n            os.makedirs(preproc_folder_path)   \n            imgs_path = os.path.join(self.data_path, 'img_align_celeba\/*.jpg')\n            print('Saving and preprocessing images ...')\n            for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):\n                cur_image = self.load_and_preprocess_image(imgname)\n                cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))\n                cur_image.save(preproc_folder_path + '\/preprocessed_image_%d.jpg' %(num)) \n        self.data_path= preproc_folder_path\n\n    def get_nextbatch(self, batch_size):\n        print(\"nextbatch batchsize is: \", batch_size) \n        assert (batch_size &gt; 0),\"Give a valid batch size\"\n        cur_idx = 0\n        image_namelist = self.get_imagelist(self.data_path)\n        while cur_idx + batch_size &lt;= self.num_imgs:\n            cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]\n            cur_batch = [self.get_input(image_path) for image_path in cur_namelist]\n            cur_batch = np.array(cur_batch).astype(np.float32)\n            cur_idx += batch_size\n            yield cur_batch\n\n    def show_image(self, image, normalized=True):\n        if not type(image).__module__ == np.__name__:\n            image = image.numpy()\n        if normalized:\n            npimg = (image * 0.5) + 0.5 \n        npimg.astype(np.uint8)\n        plt.imshow(npimg, interpolation='nearest')\n\n\n#contains functions that load, preprocess and visualize images. \n\nclass Dataset(object):     \n    def __init__(self, data_path, num_imgs, target_imgsize):\n        self.data_path = data_path\n        self.num_imgs = num_imgs \n        self.target_imgsize = target_imgsize \n\n    def normalize_np_image(self, image):\n        return (image \/ 255.0 - 0.5) \/ 0.5\n\n    def denormalize_np_image(self, image):\n        return (image * 0.5 + 0.5) * 255\n\n    def get_input(self, image_path):\n        image = np.array(Image.open(image_path)).astype(np.float32)\n        return self.normalize_np_image(image)\n\n    def get_imagelist(self, data_path, celebA=False): \n        if celebA == True:\n            imgs_path = os.path.join(data_path, 'img_align_celeba\/*.jpg')\n        else:\n            imgs_path = os.path.join(data_path, '*.jpg') \n\n        all_namelist = glob.glob(imgs_path, recursive=True)\n        return all_namelist[:self.num_imgs]\n\n    def load_and_preprocess_image(self, image_path): \n        image = Image.open(image_path)\n        j = (image.size[0] - 100) \/\/ 2\n        i = (image.size[1] - 100) \/\/ 2\n        image = image.crop([j, i, j + 100, i + 100])    \n        image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)\n        image = np.array(image.convert('RGB')).astype(np.float32)\n        image = self.normalize_np_image(image)\n        return image    \n\n    #reads data, preprocesses and saves to another folder with the given path. \n    def preprocess_and_save_images(self, dir_name, save_path=''): \n        preproc_folder_path = os.path.join(save_path, dir_name)\n        if not os.path.exists(preproc_folder_path):\n            os.makedirs(preproc_folder_path)   \n            imgs_path = os.path.join(self.data_path, 'img_align_celeba\/*.jpg')\n            print('Saving and preprocessing images ...')\n            for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):\n                cur_image = self.load_and_preprocess_image(imgname)\n                cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))\n                cur_image.save(preproc_folder_path + '\/preprocessed_image_%d.jpg' %(num)) \n        self.data_path= preproc_folder_path\n\n    def get_nextbatch(self, batch_size): \n        assert (batch_size &gt; 0),\"Give a valid batch size\"\n        cur_idx = 0\n        image_namelist = self.get_imagelist(self.data_path)\n        while cur_idx + batch_size &lt;= self.num_imgs:\n            cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]\n            cur_batch = [self.get_input(image_path) for image_path in cur_namelist]\n            cur_batch = np.array(cur_batch).astype(np.float32)\n            cur_idx += batch_size\n            yield cur_batch\n\n    def show_image(self, image, normalized=True):\n        if not type(image).__module__ == np.__name__:\n            image = image.numpy()\n        if normalized:\n            npimg = (image * 0.5) + 0.5 \n        npimg.astype(np.uint8)\n        plt.imshow(npimg, interpolation='nearest')\n\n\n\ndef generator(x, args, reuse=False):\n    with tf.device('\/gpu:0'):\n        with tf.variable_scope(\"generator\", reuse=reuse): \n            #Layer Block 1\n            with tf.variable_scope(\"layer1\"):\n                deconv1 = tf.layers.conv2d_transpose(inputs=x, \n                                             filters= args.n_g_filters*8, \n                                             kernel_size=4, \n                                             strides=1,\n                                             padding='valid',\n                                             use_bias=False,\n                                             name='deconv')\n                batch_norm1=tf.layers.batch_normalization(deconv1,\n                                             name = 'batch_norm')\n                relu1 = tf.nn.relu(batch_norm1, name='relu')\n            #Layer Block 2\n            with tf.variable_scope(\"layer2\"):\n                deconv2 = tf.layers.conv2d_transpose(inputs=relu1, \n                                             filters=args.n_g_filters*4, \n                                             kernel_size=4,\n                                             strides=2,\n                                             padding='same', \n                                             use_bias=False,\n                                             name='deconv')\n                batch_norm2 = tf.layers.batch_normalization(deconv2,\n                                             name = 'batch_norm')\n                relu2 = tf.nn.relu(batch_norm2, name='relu')\n            #Layer Block 3\n            with tf.variable_scope(\"layer3\"):\n                deconv3 = tf.layers.conv2d_transpose(inputs=relu2, \n                                             filters=args.n_g_filters*2, \n                                             kernel_size=4, \n                                             strides=2, \n                                             padding='same',\n                                             use_bias = False,\n                                             name='deconv')\n                batch_norm3 = tf.layers.batch_normalization(deconv3, \n                                             name = 'batch_norm')\n                relu3 = tf.nn.relu(batch_norm3, name='relu')\n            #Layer Block 4\n            with tf.variable_scope(\"layer4\"):\n                deconv4 = tf.layers.conv2d_transpose(inputs=relu3, \n                                             filters=args.n_g_filters, \n                                             kernel_size=4, \n                                             strides=2,\n                                             padding='same',\n                                             use_bias=False,\n                                             name='deconv')\n                batch_norm4 = tf.layers.batch_normalization(deconv4,\n                                             name = 'batch_norm')\n                relu4 = tf.nn.relu(batch_norm4, name='relu')\n            #Output Layer\n            with tf.variable_scope(\"last_layer\"):\n                logit = tf.layers.conv2d_transpose(inputs=relu4, \n                                             filters=3, \n                                             kernel_size=4, \n                                             strides=2, \n                                             padding='same',\n                                             use_bias=False,\n                                             name='logit')\n                output = tf.nn.tanh(logit) \n    return output, logit\n\n\n\ndef discriminator(x, args, reuse=False):\n    with tf.device('\/gpu:0'):\n        with tf.variable_scope(\"discriminator\", reuse=reuse): \n            with tf.variable_scope(\"layer1\"):\n                conv1 = tf.layers.conv2d(inputs=x,\n                                         filters=args.n_f_filters,\n                                         kernel_size=4,\n                                         strides=2,\n                                         padding='same',\n                                         use_bias=False,\n                                         name='conv')\n                relu1 = tf.nn.leaky_relu(conv1, alpha=0.2, name='relu')\n            with tf.variable_scope(\"layer2\"):\n                conv2 = tf.layers.conv2d(inputs=relu1,\n                                         filters=args.n_f_filters*2,\n                                         kernel_size=4,\n                                         strides=2,\n                                         padding='same',\n                                         use_bias=False,\n                                         name='conv')\n                batch_norm2 = tf.layers.batch_normalization(conv2,name='batch_norm')\n                relu2 = tf.nn.leaky_relu(batch_norm2, alpha=0.2, name='relu')\n            with tf.variable_scope(\"layer3\"):\n                conv3 = tf.layers.conv2d(inputs=relu2,\n                                         filters=args.n_f_filters*4,\n                                         kernel_size=4,\n                                         strides=2,\n                                         padding='same',\n                                         use_bias=False,\n                                         name='conv')\n                batch_norm3 = tf.layers.batch_normalization(conv3, name='batch_norm')\n                relu3 = tf.nn.leaky_relu(batch_norm3, name='relu')\n            with tf.variable_scope(\"layer4\"):\n                conv4 = tf.layers.conv2d(inputs=relu3,\n                                         filters=args.n_f_filters*8,\n                                         kernel_size=4,\n                                         strides=2,\n                                         padding='same',\n                                         use_bias=False,\n                                         name='conv')\n                batch_norm4 = tf.layers.batch_normalization(conv4, name='batch_norm')\n                relu4 = tf.nn.leaky_relu(batch_norm4, alpha=0.2, name='relu')\n            with tf.variable_scope(\"last_layer\"):\n                logit = tf.layers.conv2d(inputs=relu4,\n                                         filters=1,\n                                         kernel_size=4,\n                                         strides=1,\n                                         padding='valid',\n                                         use_bias=False,\n                                         name='conv')\n                output = tf.nn.sigmoid(logit) \n    return output, logit\n\n\n\ndef sample_z(dim_z, num_batch):\n    mu = 0\n    sigma = 1\n    s = np.random.normal(mu, sigma, num_batch*dim_z)\n    samples = s.reshape(num_batch, 1, 1, dim_z)\n    ##dist = tf.distributions.Normal(0.0, 1.0)\n    ##samples = dist.sample([num_batch, 1, 1, dim_z])\n    return samples\n#64,1,1,100  6400\nsample_z(100, 64)\n\n\n\ndef get_losses(d_real_logits, d_fake_logits):\n    #add new loss function here  \n    ###d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits, labels=tf.ones_like(d_real_logits)))    \n    ###d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.zeros_like(d_fake_logits))) \n    ###d_loss = d_loss_real + d_loss_fake\n    ###g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.ones_like(d_fake_logits)))\n    ###return d_loss, g_loss\n    d_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits,labels=tf.ones_like(d_real_logits)) + tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.zeros_like(d_fake_logits)))\n\n    g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.ones_like(d_fake_logits)))\n    return d_loss, g_loss\n\n\n\ndef get_optimizers(learning_rate, beta1, beta2):\n    d_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)\n    g_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)\n    return d_optimizer, g_optimizer\n\n\ndef optimize(d_optimizer, g_optimizer, d_loss, g_loss):\n    d_step = d_optimizer.minimize(d_loss)\n    g_step = g_optimizer.minimize(g_loss)\n    return d_step, g_step\n\n\nLOGDIR = \"logs_basic_dcgan\"\n\ndef merge_images(image_batch, size):\n    h,w = image_batch.shape[1], image_batch.shape[2]\n    c = image_batch.shape[3]\n    img = np.zeros((int(h*size[0]), w*size[1], c))\n    for idx, im in enumerate(image_batch):\n        i = idx % size[1]\n        j = idx \/\/ size[1]\n        img[j*h:j*h+h, i*w:i*w+w,:] = im\n    return img\nitr_fh = open('basic_gan_itr.txt', 'a+')\n\ndef train(args):\n    tf.reset_default_graph()\n    data_loader = Dataset(args.data_path, args.num_images, args.image_size) \n    #data_loader.preprocess_and_save_images('preprocessed', 'results_celebA') #preprocess the images once\n    X = tf.placeholder(tf.float32, shape=[args.batch_size, args.image_size , args.image_size, 3])\n    Z = tf.placeholder(tf.float32, shape=[args.batch_size, 1, 1, args.dim_z])\n\n    G_sample, _ = generator(Z, args)\n    D_real, D_real_logits = discriminator(X, args)\n    D_fake, D_fake_logits = discriminator(G_sample, args, reuse=True)\n    d_loss, g_loss = get_losses(D_real_logits, D_fake_logits)\n    d_optimizer, g_optimizer = get_optimizers(args.lr, args.beta1, args.beta2)\n    d_step, g_step = optimize(d_optimizer, g_optimizer, d_loss, g_loss)\n    ###z_sum = tf.summary.histogram('z', Z)\n    ###d_sum = tf.summary.histogram('d', D_real)\n    ###G_sum = tf.summary.histogram('g', G_sample)\n    ###d_loss_sum = tf.summary.scalar('d_loss', d_loss)\n    ###g_loss_sum = tf.summary.scalar('g_loss', g_loss)\n    ###d_sum = tf.summary.merge([z_sum, d_sum, d_loss_sum])\n    ###g_sum = tf.summary.merge([z_sum, G_sum, g_loss_sum])\n    ###saver = tf.train.Saver()\n    ###merged_summary = tf.summary.merge_all()\n\n    ###d_loss_summary = tf.summary.scalar(\"Discriminator_Total_Loss\", d_loss)\n    ###g_loss_summary = tf.summary.scalar(\"Generator_Total_Loss\", g_loss)\n    ###merged_summary = tf.summary.merge_all()\n\n    with tf.Session() as sess:\n\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(args.n_epoch):\n            for itr, real_batch in enumerate(data_loader.get_nextbatch(args.batch_size)):\n                print('itr is %d, and epoch is %d' %(itr, epoch))\n                itr_fh.write(\"epoch: \" +  str(epoch) + \" itr: \" + str(itr) + \"\\n\")\n\n                Z_sample = sample_z(args.dim_z, args.batch_size)\n\n                _, _ = sess.run([d_step, g_step], feed_dict={X:real_batch , Z:Z_sample})\n                sample = sess.run(G_sample, feed_dict={Z:Z_sample})\n                print(\"sample size is: \", sample.shape)\n                if itr==3164: #num_images\/batch_size\n                    im_merged = merge_images(sample[:16], [4,4])\n                    plt.imsave('sample_gan_images\/im_merged_epoch_%d.png' %(epoch), im_merged )\n                    scipy.misc.imsave('sample_gan_images\/im_epoch_%d_itr_%d.png' %(epoch,itr), sample[1])\n                    ##merged_summary = sess.run(merged_summary, feed_dict={X:real_batch , Z:Z_sample})\n                    ###writer = tf.summary.FileWriter(LOGDIR)   \n                    ###writer.add_summary(merged_summary, itr)\n                    ###d_loss_summary = tf.summary.scalar(\"Discriminator_Total_Loss\", d_loss)\n                    ###g_loss_summary = tf.summary.scalar(\"Generator_Total_Loss\", g_loss)\n                    ###merged_summary = tf.summary.merge_all()\n                    ###writer.add_graph(sess.graph)\n                    ###saver.save(sess, save_path='logs_basic_dcgan\/gan.ckpt')\n\n\ntrain(args)\n<\/code><\/pre>\n\n<p>Here is the images created at the end of first 5 epochs. I also have commented stuff related to tensorboard because it makes it very slow unfortunately.<\/p>\n\n<p>end of Epoch 0:\n<a href=\"https:\/\/i.stack.imgur.com\/giOUe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/giOUe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>end of Epoch 1:\n<a href=\"https:\/\/i.stack.imgur.com\/p9MBU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p9MBU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>end of Epoch 2:\n<a href=\"https:\/\/i.stack.imgur.com\/HXep2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HXep2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>end of Epoch 3:\n<a href=\"https:\/\/i.stack.imgur.com\/3EEJ0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3EEJ0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>end of Epoch 4:\n<a href=\"https:\/\/i.stack.imgur.com\/rYqqJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rYqqJ.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53367183","CreationDate":"2018-11-19T01:27:19.090","ParentId":null,"OwnerUserId":"10530728","Title":null,"Body":"<p>I think the problem is related with the definition of the optimizers:<\/p>\n\n<pre><code>def optimize(d_optimizer, g_optimizer, d_loss, g_loss):\n    d_step = d_optimizer.minimize(d_loss)\n    g_step = g_optimizer.minimize(g_loss)\n    return d_step, g_step\n<\/code><\/pre>\n\n<p>Although you define each optimizer with the corresponding loss, you are not passing the list of variables that will be trained by each optimizer. Therefore, by <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">default<\/a> the function <code>minimize<\/code> will consider all variables under the graph collection <code>GraphKeys.TRAINABLE_VARIABLES<\/code>. Since all your variables are defined under this graph collection, your current code actually updates all variables from the generator and from the discriminator when you call <code>d_step<\/code> and when you call <code>g_step<\/code>.<\/p>\n\n<p>You have to define the list of variables for each model. Since you are using variable scopes, one way to do that is:<\/p>\n\n<pre><code>def optimize(d_optimizer, g_optimizer, d_loss, g_loss):\n    t_vars = tf.trainable_variables()\n    d_vars = [var for var in t_vars if var.name.startswith('discriminator')]\n    g_vars = [var for var in t_vars if var.name.startswith('generator')]\n\n    d_step = d_optimizer.minimize(d_loss, var_list=d_vars)\n    g_step = g_optimizer.minimize(g_loss, var_list=g_vars)\n    return d_step, g_step\n<\/code><\/pre>\n"}]}
+{"QuestionId":53345624,"AnswerCount":1,"Tags":"<python><tensorflow><statistics><gaussian><gpflow>","CreationDate":"2018-11-16T21:26:54.203","AcceptedAnswerId":null,"OwnerUserId":442920.0,"Title":"GPFLow: get the full covariance matrix and find its entropy","Body":"<p>I would like to compute the determinant of the covariance matrix of a GP regression in GPFlow.  I am guessing I can get the covariance matrix with this function:<\/p>\n\n<pre><code>GPModel.predict_f_full_cov\n<\/code><\/pre>\n\n<p>This function was suggested here: <\/p>\n\n<p><a href=\"https:\/\/gpflow.readthedocs.io\/en\/develop\/notebooks\/regression.html\" rel=\"nofollow noreferrer\">https:\/\/gpflow.readthedocs.io\/en\/develop\/notebooks\/regression.html<\/a><\/p>\n\n<p>However, I have no idea how to use this function or what it returns.  I need to know a function that returns the covariance matrix for my entire model and then I need to know how to compute the determinant of it.<\/p>\n\n<p>After some effort, I figured out how to give predict_f_full_cov some points I am interested in, as we see here:<\/p>\n\n<pre><code>c = m.predict_f_full_cov(np.array([[.2],[.4],[.6],[.8]])))\n<\/code><\/pre>\n\n<p>This returned two arrays, the first of which is the mean of the predicted function for points I asked for along the x-axis.  The second array is a bit of a mystery.  I am guessing this is the covariance matrix.  I pulled it out using this:<\/p>\n\n<pre><code>covMatrix = m.predict_f_full_cov(np.array([[.2],[.4],[.6],[.8]]))[1][0]\n<\/code><\/pre>\n\n<p>Then I looked up how to compute the determinant, like so:<\/p>\n\n<pre><code>x = np.linalg.det(covMatrix)\n<\/code><\/pre>\n\n<p>Then I computed the log of this to get an entropy for the covariance matrix:<\/p>\n\n<pre><code>print(-10*math.log(np.linalg.det(covMatrix)))\n<\/code><\/pre>\n\n<p>I ran this twice using two different sets of data. The first had high noise, the second had low noise.  Strangely, the entropy went up for the <em>lower<\/em> noise data set.  I am at a loss.<\/p>\n\n<p>I found that if I just compute the covariance matrix on a small region, which should be linear, turning the noise up and down does not do what I expect.  Also, if I regress the GP to a large number of points, the determinant goes to 0.0.<\/p>\n\n<p>Here is the code I am using:<\/p>\n\n<pre><code>import gpflow\nimport numpy as np\nN = 300\nnoiseSize = 0.01\nX = np.random.rand(N,1)\nY = np.sin(12*X) + 0.66*np.cos(25*X)  + np.random.randn(N,1)*noiseSize + 3\nk = gpflow.kernels.Matern52(1, lengthscales=0.3)\nm = gpflow.models.GPR(X, Y, kern=k)\nm.likelihood.variance = 0.01\naRange = np.linspace(0.1,0.9,200)\nnewRange = []\nfor point in aRange:\n    newRange.append([point])\ncovMatrix = m.predict_f_full_cov(newRange)[1][0]\nimport math\nprint(\"Determinant: \" + str(np.linalg.det(covMatrix)))\nprint(-10*math.log(np.linalg.det(covMatrix)))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53486742","CreationDate":"2018-11-26T18:07:06.990","ParentId":null,"OwnerUserId":"442920","Title":null,"Body":"<p>So, first things first, the entropy of a multivariate normal (and a GP, given a fixed set of points on which it's evaluated) <a href=\"https:\/\/stats.stackexchange.com\/questions\/372407\/what-is-the-entropy-of-a-gaussian-process\">only depends<\/a> on its covariance matrix.<\/p>\n\n<p>Answers to your questions:<\/p>\n\n<ol>\n<li>Yes - when you make the set $X$ more and more dense, you're making the covariance matrix larger and larger, and for many simple covariance kernels, this makes the determinant smaller and smaller. My guess is that this is due to the fact that determinants of large matrices have a <em><strong>lot<\/strong><\/em> of product terms (see the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Leibniz_formula_for_determinants\" rel=\"nofollow noreferrer\">Leibniz formula<\/a>) and products of terms less than one tend to zero faster than their sums. You can verify this easily:<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/70hcl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/70hcl.png\" alt=\"Dimension and Covar\"><\/a><\/p>\n\n<p>Code for this:<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport sklearn.gaussian_process.kernels as k\n\nplt.style.use(\"ggplot\"); plt.ion()\n\nn = np.linspace(2, 25, 23, dtype = int)\nd = np.zeros(len(n))\n\nfor i in range(len(n)):\n    X = np.linspace(-1, 1, n[i]).reshape(-1, 1)\n    S = k.RBF()(X)\n    d[i] = np.log(np.linalg.det(S))\n\nplt.scatter(n, d)\nplt.ylabel(\"Log Determinant of Covariance Matrix\")\nplt.xlabel(\"Dimension of Covariance Matrix\")\n<\/code><\/pre>\n\n<p>Before moving onto the next point, do note that the entropy of a multivariate normal also has a contribution from <em>size<\/em> of the matrix, so even though the determinant shoots off to zero, there's a small contribution from the dimension.<\/p>\n\n<ol start=\"2\">\n<li>With decreasing noise, as one would expect, <em><strong>the entropy &amp; determinant do decrease<\/strong><\/em> but not tend to zero exactly; they'll decrease to the determinant due to the other kernels present in the covariance. For the demonstration below, the dimension of the covariance is kept constant ($10*10$) and the noise level is increased from 0:<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/t6q9c.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/t6q9c.png\" alt=\"Error and Determinant\"><\/a><\/p>\n\n<p>Code:<\/p>\n\n<pre><code>e = np.logspace(1, -10, 30)\nd = np.zeros(len(e))\nX = np.linspace(-1, 1, 10).reshape(-1, 1)\n\nfor i in range(len(e)):\n    S = (k.RBF() + k.WhiteKernel(e[i])) (X)\n    d[i] = np.log(np.linalg.det(S))\n\ne = np.log(e)\n\nplt.scatter(e, d)\nplt.ylabel(\"Log Determinant\")\nplt.xlabel(\"Log Error\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":53346195,"AnswerCount":1,"Tags":"<python><tensorflow><intersection><set-intersection>","CreationDate":"2018-11-16T22:22:06.260","AcceptedAnswerId":53346390.0,"OwnerUserId":869606.0,"Title":"Set intersection in Tensorflow","Body":"<p>I want to check if any of a set of given values are contained in a sparse tensor. The sparse tensor is called <code>labels<\/code> and has just one dimension containing a list of ids.<\/p>\n\n<p>In the end this seems like a simple set intersection problem, so I tried this.<\/p>\n\n<pre><code>sparse_ids = load_ids_as_sparse_tensor()\nwanted_ids = tf.constant([34, 56, 12])\nintersection = tf.sets.set_intersection(\n    wanted_ids,\n    tf.cast(sparse_ids.values, tf.int32)\n)\ncontains_any_wanted_ids = tf.not_equal(tf.size(intersection), 0)\n<\/code><\/pre>\n\n<p>However, I am getting this error:<\/p>\n\n<p><code>ValueError: Shape must be at least rank 2 but is rank 1 for 'DenseToDenseSetOperation' (op: 'DenseToDenseSetOperation') with input shapes: [3], [?].<\/code><\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"53346390","CreationDate":"2018-11-16T22:46:46.127","ParentId":null,"OwnerUserId":"7978903","Title":null,"Body":"<p>The following code works. However, I am not sure whether the result is what you want.<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.constant([34, 56, 12])\nb = tf.constant([56])\nintersection = tf.sets.set_intersection(a[None,:],b[None,:])\nsess=tf.Session()\nsess.run(intersection)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<blockquote>\n  <p>SparseTensorValue(indices=array([[0, 0]], dtype=int64),\n  values=array([56]), dense_shape=array([1, 1], dtype=int64))<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53346869,"AnswerCount":2,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2018-11-16T23:55:29.653","AcceptedAnswerId":53347561.0,"OwnerUserId":1422096.0,"Title":"How to get the activations after layer k with Keras?","Body":"<p>Note: I already read <a href=\"https:\/\/stackoverflow.com\/questions\/46677926\/keras-forward-pass-with-tensorflow-variable-as-input\">keras forward pass with tensorflow variable as input<\/a> but it did not help.<\/p>\n\n<p>I'm training an <em>auto-encoder<\/em> unsupervised neural-network with Keras with the MNIST database:<\/p>\n\n<pre><code>import keras, cv2\nfrom keras.datasets import mnist\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\nx_train = x_train.reshape(60000, 784).astype('float32') \/ 255.0\nx_test = x_test.reshape(10000, 784).astype('float32') \/ 255.0\n\nmodel = Sequential()\nmodel.add(Dense(100, activation='sigmoid', input_shape=(784,)))\nmodel.add(Dense(10, activation='sigmoid'))\nmodel.add(Dense(100, activation='sigmoid'))\nmodel.add(Dense(784, activation='sigmoid'))\nmodel.compile(loss='mean_squared_error', optimizer='sgd')\nhistory = model.fit(x_train, x_train, batch_size=1, epochs=1, verbose=0)\n<\/code><\/pre>\n\n<p>Then I would like to <strong>get the output vector<\/strong> when the input vector is <code>x_test[i]<\/code>:<\/p>\n\n<pre><code>for i in range(100):\n    x = x_test[i]\n    a = model(x)\n    cv2.imshow('img', a.reshape(28,28))\n    cv2.waitKey(0)\n<\/code><\/pre>\n\n<p>but I get this error:<\/p>\n\n<blockquote>\n  <p>All inputs to the layer should be tensors.<\/p>\n<\/blockquote>\n\n<p><strong>How should I modify this code to do a forward pass of an input vector in the neural network, and get a vector in return?<\/strong><\/p>\n\n<p><strong>Also how to get the activation <em>after, say, the 2nd layer<\/em>?<\/strong> i.e. don't propagate until the last layer, but get the output after the 2nd layer.<br>\nExample: input: vector of size 784, output: vector of size 10<\/p>\n","answers":[{"AnswerId":"53347561","CreationDate":"2018-11-17T02:07:52.140","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>To run a model after you've finished training it you need to use <a href=\"https:\/\/keras.io\/models\/model\/#predict\" rel=\"nofollow noreferrer\">keras predict()<\/a>.  This will evaluate the graph, given your input data.  Note that the input data must be the same dimensions as the specified model inputs, which in your case looks to be <code>[None, 784]<\/code>.  Keras does not require you to specify the batch dimension but you still need a 2D array going in.  Do something like..<\/p>\n\n<pre><code>x = x_test[5]\nx = x[numpy.newaxis,:]\nout_val = model.predict(x)[0]\n<\/code><\/pre>\n\n<p>if you just want to process a single value.  <\/p>\n\n<p>The <code>numpy.newaxis<\/code> is required to make a 2D array and thus match your input size.  You can skip this if you pass in an array of values to evaluate all at once. <\/p>\n\n<p>With Keras\/Tensorflow, your model is a graph\/function, not standard python procedural code.  You can't call it with data directly.  You need to create functions and then call the functions.  To get the output from an intermediate layer you can do something like..<\/p>\n\n<pre><code>OutFunc = K.function([model.input], [model.layers[2].output])\nout_val = OutFunc([x])[0]\n<\/code><\/pre>\n\n<p>again, keep in mind there is a batch dimension on the input which will be produced in the output.  There's a number of posts on getting data from intermediate layers if you need some additional examples.  For instance see <a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\">Keras, How to get the output of each layer?<\/a><\/p>\n"},{"AnswerId":"53356827","CreationDate":"2018-11-18T00:25:12.963","ParentId":null,"OwnerUserId":"1422096","Title":null,"Body":"<p>An other way to do this than the accepted answer: when <code>x<\/code> is just a <code>(784,)<\/code> or <code>(784,1)<\/code> numpy array, we can use this: <\/p>\n\n<pre><code>model.predict([[x]]) \n<\/code><\/pre>\n\n<p>with a double <code>[[...]]<\/code>.<\/p>\n"}]}
+{"QuestionId":53346900,"AnswerCount":0,"Tags":"<cuda><pytorch><nvrtc>","CreationDate":"2018-11-16T23:58:37.967","AcceptedAnswerId":null,"OwnerUserId":10665345.0,"Title":"CUDA Undefined Symbols Error for nvrtc related samples","Body":"<p>The following error occurs when I try to test on the CUDA installation samples which come with the installation (e.g. <code>vectorAdd_nvrtc<\/code>):<\/p>\n\n<pre><code>clang++ -I..\/..\/common\/inc -I\/Developer\/NVIDIA\/CUDA-9.2\/include  -arch x86_64  -o vectorAdd.o -c vectorAdd.cpp\nclang++  -rpath \/Developer\/NVIDIA\/CUDA-9.2\/lib -o vectorAdd_nvrtc \nvectorAdd.o  -L\/Developer\/NVIDIA\/CUDA-9.2\/lib -lnvrtc\nUndefined symbols for architecture x86_64:\n  \"_cuCtxCreate_v2\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\n  \"_cuCtxSynchronize\", referenced from:\n      _main in vectorAdd.o\n  \"_cuDeviceGet\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\n      findCudaDeviceDRV(int, char const**) in vectorAdd.o\n      gpuDeviceInitDRV(int, char const**) in vectorAdd.o\n  \"_cuDeviceGetAttribute\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\n      gpuGetMaxGflopsDeviceIdDRV() in vectorAdd.o\n      void getCudaAttribute&lt;int&gt;(int*, CUdevice_attribute_enum, int) in vectorAdd.o\n  \"_cuDeviceGetCount\", referenced from:\n      gpuDeviceInitDRV(int, char const**) in vectorAdd.o\n      gpuGetMaxGflopsDeviceIdDRV() in vectorAdd.o\n  \"_cuDeviceGetName\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\n      findCudaDeviceDRV(int, char const**) in vectorAdd.o\n      gpuDeviceInitDRV(int, char const**) in vectorAdd.o\n  \"_cuInit\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\n      gpuDeviceInitDRV(int, char const**) in vectorAdd.o\n      gpuGetMaxGflopsDeviceIdDRV() in vectorAdd.o\n  \"_cuLaunchKernel\", referenced from:\n      _main in vectorAdd.o\n  \"_cuMemAlloc_v2\", referenced from:\n      _main in vectorAdd.o\n  \"_cuMemFree_v2\", referenced from:\n      _main in vectorAdd.o\n  \"_cuMemcpyDtoH_v2\", referenced from:\n      _main in vectorAdd.o\n  \"_cuMemcpyHtoD_v2\", referenced from:\n      _main in vectorAdd.o\n  \"_cuModuleGetFunction\", referenced from:\n      _main in vectorAdd.o\n  \"_cuModuleLoadDataEx\", referenced from:\n      loadPTX(char*, int, char**) in vectorAdd.o\nld: symbol(s) not found for architecture x86_64\nclang: error: linker command failed with exit code 1 (use -v to see invocation)\nmake: *** [vectorAdd_nvrtc] Error 1\n<\/code><\/pre>\n\n<p>However, this only happens when the runtime lib <code>nvrtc<\/code> is utilized, i.e. <code>vectorAdd<\/code> will build just fine. I tried with several different samples (e.g. <code>clock<\/code> vs <code>clock_nvrtc<\/code>, etc.) and obversed the same outcome.<\/p>\n\n<p>I researched a little online and it seems like to be related to some header file include problem, but I am not very familiar with c++ and have no idea how to fix this. I do see that in the <code>Makefile<\/code>, it states that <code>nvrtc<\/code> is not supported on <code>ARMv7<\/code> systems, but I do not think that is the issue since my architecture is x86_64? Any help is greatly appreciated!<\/p>\n\n<p>(CUDA is of version 9.2. I am running on MAC OS 10.13.6, with Xcode 9.2, Apple LLVM version 9.0.0. I need CUDA for pytorch, not sure if this is an relevant issue though.)<\/p>\n","answers":[]}
+{"QuestionId":53347293,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory><google-cloud-tpu>","CreationDate":"2018-11-17T01:11:49.517","AcceptedAnswerId":null,"OwnerUserId":8067584.0,"Title":"Google colab TPU and reading from disc while traning","Body":"<p>I have 100k pics, and it doesn't fit into ram, so I need read it from disc while training.<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(in_pics)\ndataset = dataset.map(extract_fn)\n\ndef extract_fn(x):\n    x = tf.read_file(x)\n    x = tf.image.decode_jpeg(x, channels=3)\n    x = tf.image.resize_images(x, [64, 64])\nreturn x\n<\/code><\/pre>\n\n<p>But then I try to train, I get this error <\/p>\n\n<pre><code>File system scheme '[local]' not implemented (file: '\/content\/anime-faces\/black_hair\/danbooru_2629248_487b383a8a6e7cc0e004383300477d66.jpg')\n<\/code><\/pre>\n\n<p>Can I work around it somehow?\nAlso tried with TFRecords API, get the same error.<\/p>\n","answers":[{"AnswerId":"53754063","CreationDate":"2018-12-13T02:05:16.537","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p>The Cloud TPU you use in this scenario is not colocated on the same VM where your python runs. Easiest is to stage your data on GCS and use a gs:\/\/ URI to point the TPU at it.<\/p>\n\n<p>To optimize performance when using GCS add <code>prefetch(AUTOTUNE)<\/code> to your tf.data pipeline, and for small (&lt;50GB) datasets use <code>cache()<\/code>.<\/p>\n"}]}
+{"QuestionId":53347534,"AnswerCount":1,"Tags":"<tensorflow><neural-network><ibm-watson><watson-conversation>","CreationDate":"2018-11-17T02:01:45.203","AcceptedAnswerId":53347590.0,"OwnerUserId":10548846.0,"Title":"How can IBM Watson Assistant access saved model on Tensorflow?","Body":"<p>I have a project that I am working on that involves training a deep neural network to generate a relevance score for users and project (building a recommender system to recommend projects for users). <\/p>\n\n<p>I was wondering if there is any way for me to save an already trained recommender system to Tensorflow and then have the Watson Assistant query that model? <\/p>\n\n<p>I am really just interested in setting up an interactive experience for a user to find projects that they might be interested in. <\/p>\n\n<p>I am not experienced in either Tensorflow\/Watson Assistant so I apologize in advance if this is a foolish question. <\/p>\n","answers":[{"AnswerId":"53347590","CreationDate":"2018-11-17T02:16:17.567","ParentId":null,"OwnerUserId":"1432067","Title":null,"Body":"<p>As I am aware of, there\u2019s no direct way to achieve this. You may have to use IBM Watson Studio to expose your deep learning model as an API and create an application in the middle that interacts with both Watson Assistant API and Watson Studio API serving the responses to the UI for your user.<\/p>\n\n<p>References:<\/p>\n\n<p>Check here on how to deploy a tensorflow batch API -  <a href=\"https:\/\/dat\" rel=\"nofollow noreferrer\">https:\/\/dat<\/a>\n<a href=\"https:\/\/dataplatform.cloud.ibm.com\/docs\/content\/analyze-data\/ml-custom_libs_tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/dataplatform.cloud.ibm.com\/docs\/content\/analyze-data\/ml-custom_libs_tensorflow.html<\/a><\/p>\n\n<p><a href=\"https:\/\/dataplatform.cloud.ibm.com\/docs\/content\/analyze-data\/ml_dlaas_tensorflow_deploy_score.html\" rel=\"nofollow noreferrer\">https:\/\/dataplatform.cloud.ibm.com\/docs\/content\/analyze-data\/ml_dlaas_tensorflow_deploy_score.html<\/a><\/p>\n\n<p>Watson Assistant API reference - <a href=\"https:\/\/console.bluemix.net\/apidocs\/assistant\" rel=\"nofollow noreferrer\">https:\/\/console.bluemix.net\/apidocs\/assistant<\/a><\/p>\n"}]}
+{"QuestionId":53347661,"AnswerCount":1,"Tags":"<python><tensorflow><reshape><image-segmentation><deeplab>","CreationDate":"2018-11-17T02:30:58.763","AcceptedAnswerId":null,"OwnerUserId":8477210.0,"Title":"Deeplab v3+ Shape mismatch in tuple component","Body":"<p>I have trained deeplab v3+ on <code>ADE20K<\/code>dataset,and got the trained <code>ckpt<\/code>jlogs and <code>events<\/code>logs.But when I run <code>eval.py<\/code>and <code>vis.py<\/code>on <code>ADE20K<\/code>,I got the following errors about shape:<\/p>\n\n<pre><code>Shape mismatch in tuple component 1. Expected [513,513,3], got [513,683,3]\n<\/code><\/pre>\n\n<p>These are my <code>eval<\/code>scripts and <code>vis<\/code> scripts:<\/p>\n\n<p><code>eval<\/code>scripts:<\/p>\n\n<pre><code>#!\/bin\/bash\n\ncd ..\/\n\npython deeplab\/eval.py  \\\n--logtostderr \\\n--eval_split=\"val\" \\\n--model_variant=\"xception_65\" \\\n--atrous_rates=6 \\\n--atrous_rates=12 \\\n--atrous_rates=18 \\\n--output_stride=16 \\\n--decoder_output_stride=4 \\\n--eval_crop_size=513 \\\n--eval_crop_size=513 \\\n--checkpoint_dir=deeplab\/datasets\/ADE20K\/exp\/train_on_train_set\/train\/  \\\n--eval_logdir=deeplab\/datasets\/ADE20K\/exp\/train_on_train_set\/eval\/  \\\n--dataset_dir=deeplab\/datasets\/ADE20K\/tfrecord\/ \\\n--max_number_of_iterations=1\n<\/code><\/pre>\n\n<p><code>vis<\/code>scripts:<\/p>\n\n<pre><code>#!\/bin\/bash\n\ncd ..\/\n\npython deeplab\/vis.py  \\\n--logtostderr \\\n--vis_split=\"val\" \\\n--model_variant=\"xception_65\" \\\n--atrous_rates=6 \\\n--atrous_rates=12 \\\n--atrous_rates=18 \\\n--output_stride=16 \\\n--decoder_output_stride=4 \\\n--vis_crop_size=513 \\\n--vis_crop_size=513 \\\n--checkpoint_dir=deeplab\/datasets\/ADE20K\/exp\/train_on_train_set\/train\/  \\\n--vis_logdir=deeplab\/datasets\/ADE20K\/exp\/train_on_train_set\/vis\/  \\\n--dataset_dir=deeplab\/datasets\/ADE20K\/tfrecord\/ \\\n--max_number_of_iterations=1\n<\/code><\/pre>\n\n<p>And my <code>train<\/code>scripts:<\/p>\n\n<pre><code>#!\/bin\/bash\n\ncd ..\/\n\npython deeplab\/train.py  \\\n--logtostderr  \\\n--training_number_of_steps=150000  \\\n--train_split=\"train\"  \\\n--model_variant=\"xception_65\"  \\\n--atrous_rates=6  \\\n--atrous_rates=12  \\\n--atrous_rates=18  \\\n--output_stride=16  \\\n--decoder_output_stride=4  \\\n--train_crop_size=513  \\\n--train_crop_size=513  \\\n--train_batch_size=2  \\\n--min_resize_value=513  \\\n--max_resize_value=513  \\\n--resize_factor=16  \\\n--dataset=\"ade20k\"  \\\n --tf_initial_checkpoint=deeplab\/datasets\/ADE20K\/init_models        \/deeplabv3_xception_ade20k_train\/model.ckpt.index  \\\n--train_logdir=deeplab\/datasets\/ADE20K\/exp\/train_on_train_set\/train  \\\n--dataset_dir=deeplab\/datasets\/ADE20K\/tfrecord\/\n<\/code><\/pre>\n\n<p>Is there any thing I set wrong?\nThanks for any help.<\/p>\n","answers":[{"AnswerId":"58373006","CreationDate":"2019-10-14T08:42:21.097","ParentId":null,"OwnerUserId":"10951310","Title":null,"Body":"<p>Make sure the arguments used in your sh-script match the arguments required by your current code version.<\/p>\n\n<p>Not long ago you had to pass two separated values for the crop size buy the current implementation uses <\/p>\n\n<pre><code>--eval_crop_size=\"513,513\" \\\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code> --vis_crop_size=\"513,513\" \\\n<\/code><\/pre>\n\n<p>(taken from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/local_test.sh\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n\n<p>Hope this helps ;). If not try to print the crop values in the vis.py\/eval.py script and look if the are passed correctly.<\/p>\n"}]}
+{"QuestionId":53347789,"AnswerCount":3,"Tags":"<docker><tensorflow><tensorflow-serving><custom-operator>","CreationDate":"2018-11-17T02:55:56.107","AcceptedAnswerId":null,"OwnerUserId":1679217.0,"Title":"How to use custom operations with tensorflow server","Body":"<p>What is the ideal way to have tensorflow model server recognize my custom operation?<\/p>\n\n<p>I have a custom operation written following this guide:\n<a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/extend\/op<\/a><\/p>\n\n<p>I'm able to use the opp by calling <code>tf.load_op_library<\/code>, but when I try and and run <code>tensorflow_model_server<\/code><\/p>\n\n<pre><code>tensorflow_model_server --port=9000 \\\n                        --model_name=mymodel \\\n                        --model_base_path=\/serving\/mymodel\n<\/code><\/pre>\n\n<p>I get the following error about being unable to find my opp.<\/p>\n\n<blockquote>\n  <p>tensorflow_serving\/util\/retrier.cc:37] Loading servable: {name:\n  mymodel version: 1} failed: Not found: Op type not registered 'MyOpp'\n  in binary running on c37a4ef2d4b4.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53370626","CreationDate":"2018-11-19T08:13:16.733","ParentId":null,"OwnerUserId":"10340061","Title":null,"Body":"<p>Did you add your op lib in BUILD file where you want to call it?<\/p>\n"},{"AnswerId":"53772753","CreationDate":"2018-12-14T02:29:26.107","ParentId":null,"OwnerUserId":"1679217","Title":null,"Body":"<p>Here are the things that I wanted to do with my op:\n    - generate python wrappers\n    - add op too the pip package\n    - have my operation linked to tensorflow so tensorflow-serving could execute the operation<\/p>\n\n<p>I placed my op in tensorflow\/contrib\/foo. Here is what the source tree looked like<\/p>\n\n<pre><code>.\n\u251c\u2500\u2500 BUILD\n\u251c\u2500\u2500 LICENSE\n\u251c\u2500\u2500 __init__.py\n\u251c\u2500\u2500 foo_op.cc\n\u251c\u2500\u2500 foo_op_gpu.cu.cc\n\u2514\u2500\u2500 foo_op.h\n<\/code><\/pre>\n\n<p>My <code>__init__.py<\/code> file had the import for the generated wrappers<\/p>\n\n<pre><code>from tensorflow.contrib.sampling.ops.gen_foo import *\n<\/code><\/pre>\n\n<p>I added an import in the <code>tensorflow\/contrib\/__init__.py<\/code><\/p>\n\n<pre><code>from tensorflow.contrib import foo\n<\/code><\/pre>\n\n<p>Here is my <code>tensorflow\/contrib\/foo\/BUILD<\/code> file:<\/p>\n\n<pre><code>licenses([\"notice\"])  # Apache 2.0\nexports_files([\"LICENSE\"])\n\npackage(default_visibility = [\"\/\/visibility:public\"])\n\nload(\"\/\/tensorflow:tensorflow.bzl\", \"tf_custom_op_py_library\")\nload(\"\/\/tensorflow:tensorflow.bzl\", \"tf_gen_op_libs\")\nload(\"\/\/tensorflow:tensorflow.bzl\", \"tf_gen_op_wrapper_py\")\nload(\"\/\/tensorflow:tensorflow.bzl\", \"tf_kernel_library\")\n\ntf_kernel_library(\n    name = \"foo_op_kernels\",\n    prefix = \"foo\",\n    alwayslink = 1,\n)\ntf_gen_op_libs(\n    op_lib_names = [\"foo\"],\n)\ntf_gen_op_wrapper_py(\n    name = \"foo\",\n    visibility = [\"\/\/visibility:public\"],\n    deps = [\n        \":foo_op_kernels\",\n    ],\n)\ntf_custom_op_py_library(\n    name = \"foo_py\",\n    srcs = [\n        \"__init__.py\",\n    ],\n    kernels = [\n        \":foo_op_kernels\",\n    ],\n    srcs_version = \"PY2AND3\",\n    deps = [\n        \":foo\",\n        \"\/\/tensorflow\/contrib\/util:util_py\",\n        \"\/\/tensorflow\/python:common_shapes\",\n        \"\/\/tensorflow\/python:framework_for_generated_wrappers\",\n        \"\/\/tensorflow\/python:platform\",\n        \"\/\/tensorflow\/python:util\",\n    ],\n)\n<\/code><\/pre>\n\n<p>Here are the tensorflow bazel files I had to touch to get it working.<\/p>\n\n<ul>\n<li><code>tensorflow\/contrib\/BUILD<\/code>\n\n<ul>\n<li>Add <code>foo_op_kernels<\/code> to <code>contrib_kernels<\/code> deps<\/li>\n<li>Add <code>foo_op_lib<\/code> to <code>contrib_ops_op_lib<\/code> deps<\/li>\n<li>Add <code>foo<\/code> to <code>contrib_py<\/code> deps<\/li>\n<\/ul><\/li>\n<li><code>tensorflow\/tools\/pip_package\/BUILD<\/code>\n\n<ul>\n<li>Added my python target to <code>COMMON_PIP_DEPS<\/code><\/li>\n<\/ul><\/li>\n<li><code>tensorflow\/core\/BUILD<\/code>\n\n<ul>\n<li>Added my kernels to <code>all_kernels_statically_linked<\/code>. I might have gone overboard with this one, but It worked.<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Here are the tensorflow serving bazel files:<\/p>\n\n<ul>\n<li><code>WORKSPACE<\/code>\n\n<ul>\n<li>Change <code>org_tensorflow<\/code> to be a <code>local_repository<\/code> pointing to my tensorflow rather than a google's <code>tensorflow_http_archive<\/code><\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Then I modified: <code>tensorflow_serving\/tools\/docker\/Dockerfile.devel-gpu<\/code> to clone my versions of tensorflow and tensorflow-serving.<\/p>\n"},{"AnswerId":"54410399","CreationDate":"2019-01-28T21:15:18.293","ParentId":null,"OwnerUserId":"8185121","Title":null,"Body":"<p>You can also use tensorflow as a submodule or local_repository to use the custom macros in the repo for your ops.<\/p>\n"}]}
+{"QuestionId":53348251,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory>","CreationDate":"2018-11-17T04:35:52.747","AcceptedAnswerId":null,"OwnerUserId":2446515.0,"Title":"Cannot see Tensorflow logs in google colab notebook","Body":"<p>I was playing around with tensorflow apis on google colab virtual notebook. I want to see device mapping of my colab virtual machine.<\/p>\n\n<p>As mentioned on tensorflow developer guide, I can set the flag (log_device_placement=True) to enable logging. <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/using_gpu<\/a><\/p>\n\n<p>Below is my code running on colab notebook -<\/p>\n\n<pre><code>import tensorflow as tf\n# Creates a graph.\na = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\nb = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\nc = tf.matmul(a, b)\n\ntf.logging.set_verbosity(tf.logging.INFO)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n# Runs the op.\nprint(sess.run(c))\n<\/code><\/pre>\n\n<p>But it doesn't seems to work on colab notebook. However it is working with local jupyter notebook terminal console.<\/p>\n\n<p>Any idea how to enable logging on google colab platform ?<\/p>\n","answers":[{"AnswerId":"53353891","CreationDate":"2018-11-17T17:46:49.473","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>Looks like a TensorFlow issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3047\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3047<\/a><\/p>\n\n<p>Or, a jupyter issue: <a href=\"https:\/\/github.com\/ipython\/ipython\/issues\/1230\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ipython\/ipython\/issues\/1230<\/a><\/p>\n\n<p>Here's a work-around using a third-party library:<\/p>\n\n<pre><code>!pip install wurlitzer\n\nimport tensorflow as tf\n# Creates a graph.\na = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')\nb = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')\nc = tf.matmul(a, b)\n\ntf.logging.set_verbosity(tf.logging.INFO)\n# Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n\n# Runs the op.\nfrom wurlitzer import pipes\n\nwith pipes() as (out, err):\n  print(sess.run(c))\n\nprint (out.read())\n<\/code><\/pre>\n\n<p>Full notebook:\n<a href=\"https:\/\/colab.research.google.com\/drive\/1Z5FVCD_z8EMmyd31PsjQffQV_K7dDLfj\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1Z5FVCD_z8EMmyd31PsjQffQV_K7dDLfj<\/a><\/p>\n"}]}
+{"QuestionId":53349245,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><tensorflow-datasets>","CreationDate":"2018-11-17T07:39:39.523","AcceptedAnswerId":null,"OwnerUserId":10232639.0,"Title":"Preparing images to feed into tensorflow as datasets","Body":"<h1>AskTensorFlow<\/h1>\n\n<p>I have used the tensorflow packaged datasets like MNIST, IMDB to study the working of tensorflow. However, in practical applications we have to preprocess and prepare the dataset on our own. Suppose I am working with image dataset, so I want to preprocess them to a format that can be fed into a tensorflow model. How can I preprocess image dataset to a tensorflow format?<\/p>\n","answers":[{"AnswerId":"53349948","CreationDate":"2018-11-17T09:37:16.970","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>When working with images, you will usually use a generator. <br>\nGenerator is a function which output <code>(u,v)<\/code>, where <code>u<\/code> are the samples, and <code>v<\/code> are the labels. <br><\/p>\n\n<p>Example on how to do this can be found here <a href=\"https:\/\/stackoverflow.com\/questions\/39325275\/how-to-train-tensorflow-network-using-a-generator-to-produce-inputs\">How to train TensorFlow network using a generator to produce inputs?<\/a>. <br><\/p>\n\n<p>When building the generator function to deal with image, remember that every image is just an array, of either <code>(x,y)<\/code> for greyscale, or <code>(x,y,channels)<\/code> for color image. <br>\nThus your generator function will need to read a batch of images from the disk, and turn them into arrays. there are plenty of tools to handle this: <a href=\"https:\/\/docs.opencv.org\/3.0-beta\/doc\/py_tutorials\/py_gui\/py_image_display\/py_image_display.html\" rel=\"nofollow noreferrer\">opencv<\/a> , <a href=\"http:\/\/www.scipy-lectures.org\/advanced\/image_processing\/\" rel=\"nofollow noreferrer\">scipy<\/a>, <a href=\"https:\/\/pillow.readthedocs.io\/en\/3.0.x\/handbook\/tutorial.html\" rel=\"nofollow noreferrer\">PIL<\/a>.<br>\nAfter loading the images you can do any manipulations you like on them (using these tools, or others), usualy you will need to <strong>reshape<\/strong> the image to fit your model. <br>\nIn the end you will need to output a pair <code>([batch_size,x,y,channels], [batch_size,labels])<\/code>.<\/p>\n"}]}
+{"QuestionId":53350677,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><pytorch>","CreationDate":"2018-11-17T11:15:20.683","AcceptedAnswerId":53352843.0,"OwnerUserId":1672565.0,"Title":"How to do a \"element by element in-place inverse\" with pytorch?","Body":"<p>Given is an array <code>a<\/code>:<\/p>\n\n<pre><code>a = np.arange(1, 11, dtype = 'float32')\n<\/code><\/pre>\n\n<p>With numpy, I can do the following:<\/p>\n\n<pre><code>np.divide(1.0, a, out = a)\n<\/code><\/pre>\n\n<p>Resulting in:<\/p>\n\n<pre><code>array([1.        , 0.5       , 0.33333334, 0.25      , 0.2       ,\n       0.16666667, 0.14285715, 0.125     , 0.11111111, 0.1       ],\n      dtype=float32)\n<\/code><\/pre>\n\n<p>Assuming that <code>a<\/code> is instead a pytorch tensor, the following operation <strong>fails<\/strong>:<\/p>\n\n<pre><code>torch.div(1.0, a, out = a)\n<\/code><\/pre>\n\n<p>The first parameter of <code>div<\/code> is expected to be a tensor of matching length\/shape.<\/p>\n\n<p>If I substitute <code>1.0<\/code> with an array <code>b<\/code> filled with ones, its length equal to the length of <code>a<\/code>, it works. The downside is that I have to allocate memory for <code>b<\/code>. I can also do something like <code>a = 1.0 \/ a<\/code> which will yet again allocate extra (temporary) memory.<\/p>\n\n<p>How can I do this operation efficiently \"in-place\" (without the allocation of extra memory), ideally with broadcasting?<\/p>\n","answers":[{"AnswerId":"53352843","CreationDate":"2018-11-17T15:52:46.177","ParentId":null,"OwnerUserId":"5373738","Title":null,"Body":"<p>Pytorch follows the convention of using <code>_<\/code> for in-place operations.\nfor eg<\/p>\n\n<pre><code>add -&gt; add_  # in-place equivalent\ndiv -&gt; div_  # in-place equivalent\netc\n<\/code><\/pre>\n\n<p>Element-by-element inplace inverse.<\/p>\n\n<pre><code>&gt;&gt;&gt; a = torch.arange(1, 11, dtype=torch.float32) \n&gt;&gt;&gt; a.pow_(-1) \n&gt;&gt;&gt; a\n&gt;&gt;&gt; tensor([1.0000, 0.5000, 0.3333, 0.2500, 0.2000, 0.1667, 0.1429, 0.1250, 0.1111, 0.1000])\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>&gt;&gt;&gt; a = torch.arange(1, 11, dtype=torch.float32) \n&gt;&gt;&gt; a.div_(a ** a) \n&gt;&gt;&gt; a\n&gt;&gt;&gt; tensor([1.0000, 0.5000, 0.3333, 0.2500, 0.2000, 0.1667, 0.1429, 0.1250, 0.1111, 0.1000])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53350810,"AnswerCount":2,"Tags":"<python-3.x><lambda><closures><pytorch><local-variables>","CreationDate":"2018-11-17T11:32:32.623","AcceptedAnswerId":53350913.0,"OwnerUserId":2397253.0,"Title":"Why is this function parameter identical in every invocation, despite passing different value? (Creating closures in a loop)","Body":"<p>I'm using PyTorch and trying to register hooks on model parameters. The following code creates lambda functions to add to each model parameter, so I can see in the hook which tensor the gradient belongs to<\/p>\n\n<pre><code>import torch\nimport torchvision\n\n# define model and random train batch\nmodel = torchvision.models.alexnet()\ninput = torch.rand(10, 3, 224, 224)   # batch of 10 images\ntargets = torch.zeros(10).long()\n\ndef grad_hook_template(param, name, grad):\n    print(f'Receive grad for {name} w whape {grad.shape}')\n\n# add one lambda hook to each parameter\nfor name, param in model.named_parameters():\n    print(f'Register hook for {name}')\n\n    # use a lambda so we can pass additional information to the hook, which should only take one parameter\n    param.register_hook(lambda grad: grad_hook_template(param, name, grad))\n\nloss_fn = torch.nn.CrossEntropyLoss()\noptimizer = torch.optim.SGD(model.parameters(), lr=0.1)\noptimizer.zero_grad()\n\nprediction = model(input)\nloss = loss_fn(prediction, targets)\nloss.backward()\noptimizer.step()\n<\/code><\/pre>\n\n<p>The result is that the <code>name<\/code> and <code>param<\/code> arguments to <code>grad_hook_template<\/code> always as the same value (and <code>id<\/code>), but the <code>grad<\/code> argument is always different (as expected). Why is it that when I register the hook, the lambdas seem to refer to the same local variables each time?<\/p>\n\n<p>I read e.g. <a href=\"https:\/\/stackoverflow.com\/a\/2295368\/2397253\">here<\/a> that loops do not create new scopes and closures are lexical in Python, i.e. the <code>name<\/code> and <code>param<\/code> which I'm passing to the lambda are just pointers and whatever value they have at the end of the loop is seen by everyone with this pointer. But what can I do about it? <code>copy.copy()<\/code> the variables?<\/p>\n","answers":[{"AnswerId":"53350855","CreationDate":"2018-11-17T11:38:27.673","ParentId":null,"OwnerUserId":"2397253","Title":null,"Body":"<p>This is kind of answered by the <a href=\"https:\/\/docs.python.org\/3\/faq\/programming.html#why-do-lambdas-defined-in-a-loop-with-different-values-all-return-the-same-result\" rel=\"nofollow noreferrer\">FAQ<\/a>.<\/p>\n\n<p>Solutions include<\/p>\n\n<ul>\n<li>using <code>functools.partial<\/code> instead of <code>lambda<\/code><\/li>\n<li>using default params to the lambdas to capture the value of the variables<\/li>\n<\/ul>\n"},{"AnswerId":"53350913","CreationDate":"2018-11-17T11:46:24.160","ParentId":null,"OwnerUserId":"892383","Title":null,"Body":"<p>You've run into <em>late binding closures<\/em>. The variables <code>param<\/code> and <code>name<\/code> are looked up at call time, not when the function they're used in is defined. By the time any of these functions are called, <code>name<\/code> and <code>param<\/code> are at the last values in the loop. To get around this, you could do this:<\/p>\n\n<pre><code>for name, param in model.named_parameters():\n    print(f'Register hook for {name}')\n    param.register_hook(lambda grad, name=name, param=param: grad_hook_template(param, name, grad))\n<\/code><\/pre>\n\n<p>However, I think using <code>functools.partial<\/code> is the right solution here:<\/p>\n\n<pre><code>from functools import partial\n\nfor name, param in model.named_parameters():\n    print(f'Register hook for {name}')\n    param.register_hook(partial(grad_hook_template, name=name, param=param))\n<\/code><\/pre>\n\n<p>You can find more information about <a href=\"https:\/\/docs.python-guide.org\/writing\/gotchas\/#late-binding-closures\" rel=\"nofollow noreferrer\">late binding closures at the Common Gotchas page of the Hitchhiker's Guide to Python<\/a> as well as <a href=\"https:\/\/docs.python.org\/3\/faq\/programming.html#why-do-lambdas-defined-in-a-loop-with-different-values-all-return-the-same-result\" rel=\"nofollow noreferrer\">in the Python docs<\/a>.<\/p>\n\n<p><strong>Note that this applies equally to functions defined with the <code>def<\/code> keyword.<\/strong><\/p>\n"}]}
+{"QuestionId":53350894,"AnswerCount":0,"Tags":"<python><tensorflow><keras><nvidia><cudnn>","CreationDate":"2018-11-17T11:43:10.020","AcceptedAnswerId":null,"OwnerUserId":8049537.0,"Title":"UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory","Body":"<p>This is my model.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code># Import Keras \r\nimport tensorflow as tf\r\nfrom tensorflow.python.keras.layers import Conv2D, MaxPooling2D, Flatten,GlobalMaxPool2D\r\nfrom tensorflow.python.keras.layers import Input, LSTM, Embedding, Dense\r\nfrom tensorflow.python.keras.models import Model, Sequential\r\nfrom tensorflow.python import keras\r\n# Define CNN for Image Input\r\nvision_model = Sequential()\r\nvision_model.add(Conv2D(64, (3, 3), activation='relu', padding='same', input_shape=(120, 160, 3)))\r\nvision_model.add(Conv2D(64, (3, 3), activation='relu'))\r\nvision_model.add(MaxPooling2D((2, 2)))\r\nvision_model.add(Conv2D(128, (3, 3), activation='relu', padding='same'))\r\nvision_model.add(Conv2D(128, (3, 3), activation='relu'))\r\nvision_model.add(MaxPooling2D((2, 2)))\r\nvision_model.add(Conv2D(196, (3, 3), activation='relu', padding='same'))\r\nvision_model.add(Conv2D(196, (3, 3), activation='relu'))\r\nvision_model.add(Conv2D(196, (3, 3), activation='relu'))\r\nvision_model.add(MaxPooling2D((2, 2)))\r\nvision_model.add(Conv2D(256, (3, 3), activation='relu', padding='same'))\r\nvision_model.add(Conv2D(256, (3, 3), activation='relu'))\r\nvision_model.add(Conv2D(256, (3, 3), activation='relu'))\r\nvision_model.add(MaxPooling2D((2, 2)))\r\nvision_model.add(Conv2D(384, (3, 3), activation='relu', padding='same'))\r\nvision_model.add(Conv2D(384, (3, 3), activation='relu'))\r\nvision_model.add(GlobalMaxPool2D())\r\nvision_model.summary()\r\nimage_input = Input(shape=(120, 160, 3))\r\nencoded_image = vision_model(image_input)\r\n\r\n# Define RNN for language input\r\nquestion_input = Input(shape=(42,), dtype='int32')\r\nembedded_question = Embedding(input_dim=500, output_dim=256, input_length=100)(question_input)\r\nencoded_question = LSTM(256,return_sequences=True)(embedded_question)\r\nencoded_question = LSTM(256)(embedded_question)\r\n\r\n# Combine CNN and RNN to create the final model\r\nmerged = keras.layers.concatenate([encoded_question, encoded_image])\r\noutput = Dense(26, activation='softmax')(merged)\r\nvqa_model = Model(inputs=[image_input, question_input], outputs=output)\r\nvqa_model.summary()<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>.\n.\n.\n.<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>vqa_model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['acc'])\r\nvqa_model.fit_generator(train_gen,len(train_data[0])\/\/16,validation_data=val_gen,validation_steps=split\/\/16,verbose=1,epochs=100)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>And after running this I am getting this error<\/p>\n\n<blockquote>\n  <p>UnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.\n       [[{{node sequential\/conv2d\/Conv2D}} = Conv2D[T=DT_FLOAT, _class=[\"loc:@training\/Adam\/gradients\/sequential\/conv2d\/Conv2D_grad\/Conv2DBackpropFilter\"], data_format=\"NCHW\", dilations=[1, 1, 1, 1], padding=\"SAME\", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](training\/Adam\/gradients\/sequential\/conv2d\/Conv2D_grad\/Conv2DBackpropFilter-0-TransposeNHWCToNCHW-LayoutOptimizer, sequential\/conv2d\/Conv2D\/ReadVariableOp)]]\n       [[{{node ConstantFoldingCtrl\/loss\/dense_loss\/broadcast_weights\/assert_broadcastable\/AssertGuard\/Switch_0\/_316}} = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_1886_...d\/Switch_0\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]\n   \"<\/p>\n<\/blockquote>\n\n<p>I am using tensorflow gpu on my local system with NVIDIA 1050Ti.<\/p>\n\n<p>I tried to search the solution but that wasn't working.<\/p>\n","answers":[]}
+{"QuestionId":53350905,"AnswerCount":1,"Tags":"<gpu><pytorch><allennlp>","CreationDate":"2018-11-17T11:44:28.177","AcceptedAnswerId":null,"OwnerUserId":10666819.0,"Title":"pytorch delete model from gpu","Body":"<p>I want to make a cross validation in my project based on Pytorch.\nAnd I didn't find any method that pytorch provided to delete the current model and empty the memory of GPU. Could you tell that how can I do it?<\/p>\n","answers":[{"AnswerId":"53353203","CreationDate":"2018-11-17T16:32:31.443","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Freeing memory in PyTorch works as it does with the normal Python garbage collector. This means once all references to an <em>Python-Object<\/em> are gone it will be deleted.<\/p>\n\n<p>You can delete references by using the <a href=\"https:\/\/stackoverflow.com\/questions\/20847149\/how-does-del-operator-work-in-list-in-python\"><code>del<\/code><\/a> operator:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>del model\n<\/code><\/pre>\n\n<p>You have to make sure though that there is no reference to the respective object left, otherwise the memory won't be freed.<\/p>\n\n<p>So once you've deleted all references of your <code>model<\/code>, it should be deleted and the memory freed.<\/p>\n\n<p>If you want to learn more about memory management you can take a look here:\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/cuda.html#cuda-memory-management\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/notes\/cuda.html#cuda-memory-management<\/a><\/p>\n"}]}
+{"QuestionId":53351130,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-17T12:12:25.153","AcceptedAnswerId":null,"OwnerUserId":10666722.0,"Title":"Tensorflow Dense label shape","Body":"<p>I'm new to using Tensorflow and Python, I've seen all tutorial in the website and now I'm working with my first real dataset.<\/p>\n\n<p>What I want to do with the NN is to predict some power plant energy consumes knowing the daily trends. I have an .xlsx file with all those (real) values. Using Pandas I'splitted and normalized the data in train set and validation set (i.e. train_x and train_y, where train_x is the time and train_y is the label). The x and y array are both numpy.ndarray and formatted as below (just the head):<\/p>\n\n<pre><code>print(train_x)\n[ 644]\n[ 645]\n[ 646]\n<\/code><\/pre>\n\n<p><code>print(train_y)\n [-0.09154356  1.10702972  1.13661838]\n [ 0.05104414  1.39112378  1.5319337 ]\n [-0.05719421  1.40702419  1.48187637]<\/code><\/p>\n\n<p>Then I created the model:<\/p>\n\n<pre><code>model = keras.Sequential([\nkeras.layers.Dense(64, activation=tf.nn.relu, input_shape= (train_x.shape([0]))),     \nkeras.layers.Dense(3, activation=tf.nn.softmax)]) \n\nmodel.compile(loss='categorical_cross_entropy',\n                optimizer='Adam',\n                metrics=['accuracy'])\n\nhistory = model.fit(train_x, train_y, epochs=5, verbose=1)\n<\/code><\/pre>\n\n<p>But when I run the script I got this error:<\/p>\n\n<pre><code>TypeError: 'tuple' object is not callable\n<\/code><\/pre>\n\n<p>I guess the problem is about the input shape of the layer or maybe of the loss function as is suggested <a href=\"https:\/\/stackoverflow.com\/questions\/51617857\/keras-dense-layer-shape-error\">here<\/a>, so I tried to modify the loss function in:<\/p>\n\n<pre><code>LOSS = tf.nn.categorical_cross_entropy_with_logits(logits=3, labels=3)\n<\/code><\/pre>\n\n<p>and, of course, the model.compile:<\/p>\n\n<pre><code>model.compile(loss=LOSS,\n                optimizer='Adam',\n                metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>but I got the same error again:<\/p>\n\n<pre><code>TypeError: 'tuple' object is not callable\n<\/code><\/pre>\n\n<p>Where I go wrong?<\/p>\n","answers":[{"AnswerId":"53351469","CreationDate":"2018-11-17T12:58:10.330","ParentId":null,"OwnerUserId":"8491005","Title":null,"Body":"<p>it should be <code>array.shape[0]<\/code>,not <code>array.shape([0])<\/code>. <code>shape<\/code> is an attribute of a numpy array, not a method. The correct syntax should be:<\/p>\n\n<pre><code>keras.layers.Dense(64, activation=tf.nn.relu, input_shape= (train_x.shape[-1],)),\n<\/code><\/pre>\n\n<p>Also, change <code>train_x<\/code> and <code>train_y<\/code> to 2d arrays, with the shape of [length_of_array,1].<\/p>\n"}]}
+{"QuestionId":53351362,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><lstm>","CreationDate":"2018-11-17T12:41:48.430","AcceptedAnswerId":null,"OwnerUserId":10651990.0,"Title":"ConvLSTM2D data preparation","Body":"<p>I'm trying to use ConvLSTM2D for 1700 of 90x3 data in keras.<\/p>\n\n<p>I already did CONV2D which data is (1700x90x30x1). Data format is (batch, rows, cols, channels)\nNow I want to use CONVLSTM2D but I found out I should change the data format to (samples, time, rows, cols, channels).<\/p>\n\n<blockquote>\n  <p>samples=1700  , row=90 , cols=30,  channels=1<\/p>\n<\/blockquote>\n\n<p>How to determine the \"time\"?<\/p>\n","answers":[{"AnswerId":"53351542","CreationDate":"2018-11-17T13:07:15.710","ParentId":null,"OwnerUserId":"3053918","Title":null,"Body":"<p>ConvLSTM2D, or LSTM as a special type of recurrent neural network in general, are used when the input data is a time series. This enables to take advantage of temporal properties within the data.\nIn case of ConvLSTM2D, the input is usually a video, consisting of multiple frames. Consequently, you have to reshape the data the following way:<\/p>\n\n<pre><code>samples=1700 , time=t, row=90 , cols=30, channels=1\n<\/code><\/pre>\n\n<p>where <em>t<\/em> is the number of frames in the video.<\/p>\n\n<p>As an example, let's say we want to do video classification (or frame prediction) based on a short video clip of 10 frames, then <em>t=10<\/em>.\nThis of course only makes sense in case the image frames you are having are in a temporal order. Simply use <em>tf.reshape(...)<\/em>.<\/p>\n"}]}
+{"QuestionId":53351963,"AnswerCount":2,"Tags":"<python><tensorflow><keras><confusion-matrix>","CreationDate":"2018-11-17T14:04:47.133","AcceptedAnswerId":53352044.0,"OwnerUserId":2651073.0,"Title":"Mnist: get confusion matrix","Body":"<p>I try to get the confusion matrix of mnist dataset.<\/p>\n\n<p>It's my code:<\/p>\n\n<pre><code>mnist = tf.keras.datasets.mnist\n\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(512, activation=tf.nn.tanh),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\n\nmodel.fit(x_train, y_train, epochs=1, callbacks=[history])\n\ntest_predictions = model.predict(x_test)\n\n\n# Compute confusion matrix\nconfusion = tf.confusion_matrix(y_test, test_predictions)\n<\/code><\/pre>\n\n<p>The problem is that <code>test_prediction<\/code> is a 10000 x 10 matrix, while y_test is 10000 x 1 matrix. In fact, the neural net doesn't provide an output for every test sample. How can I calculate confusion matrix for such cases? <\/p>\n\n<p>And how can I then present the confusion matrix? Can I use the sci-kit library for this purpose?<\/p>\n","answers":[{"AnswerId":"53352044","CreationDate":"2018-11-17T14:15:20.660","ParentId":null,"OwnerUserId":"6384539","Title":null,"Body":"<p>This is probably because your prediction consist of probabilities of all the possible classes. You need to pick the class with highest probability, which will result in the same dimension as y_test. You can use the argmax() method from numpy. It works something like this:<\/p>\n\n<pre><code>import numpy as np\na = np.array([[0.9,0.1,0],[0.2,0.3,0.5],[0.4,0.6,0]])\nnp.argmax(a, axis=0)\narray([0, 2, 1])\n<\/code><\/pre>\n\n<p>You can use sklearn for generating confusion matrix. Your code will become something like this<\/p>\n\n<pre><code>from sklearn.metrics import confusion_matrix\nimport numpy as np\n\nconfusion = confusion_matrix(y_test, np.argmax(test_predictions,axis=1))\n<\/code><\/pre>\n"},{"AnswerId":"53352459","CreationDate":"2018-11-17T15:06:25.983","ParentId":null,"OwnerUserId":"10349960","Title":null,"Body":"<p>If you use .predict_classes method instead of just predict, you get a vector of classes with the highest probability.<\/p>\n\n<p>Then, you can use confusion_matrix from sklearn.<\/p>\n\n<pre><code>test_predictions = model.predict_classes(x_test)\n\nfrom sklearn.metrics import confusion_matrix\n\ncm = confusion_matrix(y_true = y_test, y_pred = test_predictions)\nprint(cm)\n<\/code><\/pre>\n\n<p>The shape of test_predictions here is (10000,).<\/p>\n\n<p>The printed result will be something like:<\/p>\n\n<pre><code>array([[ 967,    1,    1,    2,    0,    1,    5,    0,    2,    1],\n   [   0, 1126,    3,    1,    0,    1,    1,    0,    3,    0],\n   [   3,    2, 1001,    8,    1,    0,    3,    6,    8,    0],\n   [   0,    0,    1, 1002,    0,    1,    0,    1,    5,    0],\n   [   3,    1,    2,    2,  955,    2,    6,    1,    3,    7],\n   [   3,    1,    0,   37,    1,  833,    9,    0,    6,    2],\n   [   4,    3,    1,    1,    1,    3,  941,    0,    4,    0],\n   [   2,    9,    8,    5,    0,    0,    0,  988,    8,    8],\n   [   3,    1,    3,   10,    3,    2,    2,    3,  946,    1],\n   [   3,    8,    0,   10,    8,    8,    1,    4,    5,  962]],\n  dtype=int64)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53352119,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-17T14:25:36.517","AcceptedAnswerId":53352567.0,"OwnerUserId":10667324.0,"Title":"error during the import of tensorflow-gpu version","Body":"<blockquote>\n  <p>import tensorflow Traceback (most recent call last):   File\n  \"\", line 1, in \n      import tensorflow   File \"C:\\Users\\nayer meleka\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow__init__.py\",\n  line 39, in \n      from tensorflow._api.v1 import app ModuleNotFoundError: No module named 'tensorflow._api'<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QaugB.png\" rel=\"nofollow noreferrer\">this error happend when I tried to import it from terminal<\/a><\/p>\n\n<p>When I tried to use tensorflow this error pop up.<\/p>\n\n<p>I installed it using pip didn't use any virtualmachine.<\/p>\n","answers":[{"AnswerId":"53352567","CreationDate":"2018-11-17T15:19:07.483","ParentId":null,"OwnerUserId":"10665545","Title":null,"Body":"<p>Try installing tensorflow-gpu again by running the command prompt in <em>administrator mode<\/em> still through pip. And keep in mind that it won't just start using the GPU by itself, you will have to install other dependencies too such as <strong>Visual Studio, CUDA, CUDNN<\/strong>, etc.<\/p>\n\n<p>Read this: <a href=\"https:\/\/www.quantinsti.com\/blog\/install-tensorflow-gpu\" rel=\"nofollow noreferrer\">https:\/\/www.quantinsti.com\/blog\/install-tensorflow-gpu<\/a><\/p>\n"}]}
+{"QuestionId":53352215,"AnswerCount":0,"Tags":"<python><keras><tensorboard>","CreationDate":"2018-11-17T14:36:10.790","AcceptedAnswerId":null,"OwnerUserId":6525034.0,"Title":"Keras TensorBoard visulize Conv Kernels","Body":"<p>I am using Keras with TensorFlow as backend.\nNow i want to use the TensorBoard callback to visualize my conv layer kernels.\nBut i can only see the first conv layer kernel in TensorBoard and my Dense layers at the end.\nFor the other conv layers i can just the the bias values and not the kernels.<\/p>\n\n<p>Here is my sample code for the Keras model.<\/p>\n\n<pre><code>tb = TensorBoard(\n    log_dir=log_dir, \n    histogram_freq=epochs, \n    write_images=True)\n\n# Define the DNN\nmodel = Sequential()\nmodel.add(Conv2D(filters=16, kernel_size=3, input_shape=(width, height, depth), name=\"conv1\"))\nmodel.add(Activation(\"relu\"))\nmodel.add(Conv2D(filters=16, kernel_size=3, name=\"conv2\"))\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPool2D())\n\nmodel.add(Conv2D(filters=32, kernel_size=3, name=\"conv3\"))\nmodel.add(Activation(\"relu\"))\nmodel.add(Conv2D(filters=32, kernel_size=3, name=\"conv4\"))\nmodel.add(Activation(\"relu\"))\nmodel.add(MaxPool2D())\n\nmodel.add(Flatten())\nmodel.add(Dense(128))\nmodel.add(Activation(\"relu\"))\nmodel.add(Dense(num_classes, name=\"features\"))\nmodel.add(Activation(\"softmax\"))\n\n# Print the DNN layers\nmodel.summary()\n\n# Train the DNN\nlr = 1e-3\noptimizer = Adam(lr=lr)\nmodel.compile(loss=\"categorical_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\nmodel.fit(x_train, y_train, verbose=1, \nbatch_size=batch_size, epochs=epochs, \nvalidation_data=(x_test, y_test), \ncallbacks=[tb])\n<\/code><\/pre>\n\n<p>And this is what i see in TensorBoard.\n(I minimized the Kernels of my first conv layer)\n<a href=\"https:\/\/i.stack.imgur.com\/EuiC5.png\" rel=\"nofollow noreferrer\">TB Screenshot<\/a><\/p>\n\n<p>What am i missing to visulize all my kernels?<\/p>\n","answers":[]}
+{"QuestionId":53352604,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-17T15:23:56.957","AcceptedAnswerId":null,"OwnerUserId":5955479.0,"Title":"What is the right way to import data to tensorflow?","Body":"<p>I am new to Tensorflow and trying to make my own little project. I would like to import my CSV file as a dataset and then I would like to split it into training and testing sets and also to be able to make batches from my dataset.<br>\nMy CSV file contains 3 columns of numbers so I managed to find these lines of code<\/p>\n\n<pre><code>filenames = ['mydata.csv']\nrecord_defaults = [tf.float32] * 3\ndataset = tf.contrib.data.CsvDataset(filenames, record_defaults, header=True, select_cols=[1,2,3])\n<\/code><\/pre>\n\n<p>How do I convert this object to tensor or dataset, so I can either split the data or create batches of data?<\/p>\n","answers":[{"AnswerId":"53353302","CreationDate":"2018-11-17T16:43:26.520","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>Use a tool to split your data like <code>sklearn.model_selection.train_test_split<\/code>:<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(\n    dataset[:2], dataset[2], test_size=0.33, random_state=42)\n<\/code><\/pre>\n\n<p>For instance if your dataset consists of two features columns and one output label.<\/p>\n"},{"AnswerId":"53377171","CreationDate":"2018-11-19T14:52:51.880","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>As explained on the tensorflow guide <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets#consuming_csv_data\" rel=\"nofollow noreferrer\">here<\/a> , you have the dataset and after this you can preprocess your data using the <code>Dataset.map()<\/code> transformation for a certain defined function. Batching and shuffling could also be done after wards using <code>dataset.batch(Batch_size)<\/code> and <code>dataset.shuffle(buffer_size=Buffer_Size)<\/code>. you can read the guide for further details. <\/p>\n"}]}
+{"QuestionId":53352954,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2018-11-17T16:04:47.167","AcceptedAnswerId":53353112.0,"OwnerUserId":10640491.0,"Title":"Multitask learning in Keras","Body":"<p>I have two different datasets and I would like to try multi-task learning. My problem is that all the examples I could find have two different training inputs, but the labels are the same. My question is: can I have different labels? This is my code right now:<\/p>\n\n<pre><code>input1 = Sequential()\ninput1.add(Embedding(vocabulary_size, embedding_size, \ninput_length=longest_sen_input1))\ninput1.add(Bidirectional(LSTM(units=embedding_size)))\ninput1.add(Dense(len(document), activation='softmax'))\n\ninput2 = Sequential()\ninput2.add(Embedding(vocabulary_size, embedding_size, \ninput_length=longest_sen_input2))\ninput2.add(Bidirectional(LSTM(units=embedding_size)))\ninput2.add(Dense(len(document), activation='softmax'))\n\nmodel = Sequential()\nmodel.add(Merge([input1, input2], mode='sum'))\nmodel.add(Dense(len(document), activation='softmax'))\nmodel.compile(loss='categorical_crossentropy',optimizer='adam')\n\nmodel.fit([X_train_input1, X_train_input2], [Y_train_input1, Y_train_input2], epochs=100)\n<\/code><\/pre>\n\n<p>I try to insert [Y_train_input1, Y_train_input2], but I have this error:<\/p>\n\n<pre><code>Error when checking model target: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 2 arrays: [array([[0., 0., 0., ..., 0., 0., 0.],\n   [0., 0., 0., ..., 0., 0., 0.],\n   [0., 0., 0., ..., 0., 0., 0.],\n   ...,\n   [0., 0., 0., ..., 1., 0., 0.],\n   [0., 0., 0., ..., 0., 0., 0....\n<\/code><\/pre>\n\n<p>Does anybody know how to perform multitask learning with two inputs (X_train_input1\/Y_train_input1 and X_train_input2\/Y_train_input2) returning a common prediction?<\/p>\n\n<p><em>EDIT<\/em>\nMy model seems to work now, I simply changed<\/p>\n\n<pre><code>model.fit([X_train_input1, X_train_input2], [Y_train_input1, Y_train_input2], epochs=100)\n<\/code><\/pre>\n\n<p>in<\/p>\n\n<pre><code>model.fit([X_train_input1, X_train_input2], Y_train, epochs=100)\n<\/code><\/pre>\n\n<p>but then I try to test the model like this<\/p>\n\n<pre><code>multitask_model.predict_classes(X_test)\n<\/code><\/pre>\n\n<p>and I have this error:<\/p>\n\n<pre><code>ValueError: Error when checking model : the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[  0,   0,   0, ...,  13,   8, 134],\n   [  0,   0,   0, ...,  33,  87,  19],\n   [  0,   0,   0, ...,  27,   1,   4],\n   ...,\n   [  0,   0,   0, ...,   1,  10,   8],\n   [  0...\n<\/code><\/pre>\n\n<p>What am I missing?<\/p>\n","answers":[{"AnswerId":"53353112","CreationDate":"2018-11-17T16:23:00.810","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Your model has only one output, and you're passing two: <code>Y_train_input1<\/code>,  <code>Y_train_input2<\/code>.<\/p>\n\n<p>If your goal is not to merge both models then you should keep them separate. When you merge\/sum the outputs, you end up with only one output. <\/p>\n\n<p>Is your intention to really have two different separate models without any interaction between them? <\/p>\n\n<ul>\n<li>Either you have a common output and a common <code>Y_train<\/code>, or   <\/li>\n<li>You have two separate outputs and two different targets. <\/li>\n<\/ul>\n"}]}
+{"QuestionId":53353067,"AnswerCount":0,"Tags":"<machine-learning><pytorch>","CreationDate":"2018-11-17T16:17:15.313","AcceptedAnswerId":null,"OwnerUserId":4907108.0,"Title":"Pytorch autograd: compute batch gradient w.r.t to input","Body":"<p>I need to compute the gradient of the network output (not the loss) with respect to the input but I'm getting the error <code>grad can be implicitly created only for scalar outputs<\/code>. I'm trying to reproduce what's described in <a href=\"https:\/\/arxiv.org\/pdf\/1712.01026v4.pdf\" rel=\"nofollow noreferrer\">this paper<\/a> and I need the gradient I mentioned to perform step 7 <a href=\"https:\/\/i.stack.imgur.com\/JyE42.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JyE42.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is what I'd like to do:<\/p>\n\n<pre><code>out = netD(input).view(-1)\nout.backward(retain_graph=True)\ngrad = input.grad\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53353493,"AnswerCount":1,"Tags":"<arrays><r><keras>","CreationDate":"2018-11-17T17:03:28.993","AcceptedAnswerId":null,"OwnerUserId":10327890.0,"Title":"Array Dimension","Body":"<p>I am currently learning to use Keras in R. When I run the command <\/p>\n\n<pre><code>dim(mnist$train$x) \n<\/code><\/pre>\n\n<p>I get the output as (output 1)<\/p>\n\n<pre><code>[1] 60000 28 28\n<\/code><\/pre>\n\n<p>which means that there are 60000 matrices each with a 28*28 dimension. <\/p>\n\n<p>Now when I create an array through R code for the same dimension , I use <\/p>\n\n<pre><code>test &lt;- array(28*28*6000,dim=c(28,28,6000)) \n<\/code><\/pre>\n\n<p>where the inner layers are specified first and upon using the statement <code>dim(test)<\/code> I get this output (output 2)<\/p>\n\n<pre><code>[1]   28   28 6000\n<\/code><\/pre>\n\n<p>Both these notations are showing up the same array in different format. Is it possible to get the output in the second case in the output 1 format ?<\/p>\n","answers":[{"AnswerId":"53355285","CreationDate":"2018-11-17T20:29:34.640","ParentId":null,"OwnerUserId":"3587960","Title":null,"Body":"<p>Do you just want output 2 in output 1 format? You mean something like this:<\/p>\n\n<pre><code>test &lt;- array(28*28*6000,dim=c(28,28,6000))\n\nd &lt;- dim(test)\n\nnewdim &lt;- c(d[length(d)], d[-length(d)] )\n<\/code><\/pre>\n\n<p>newdim will be <code>60000 28 28<\/code><\/p>\n"}]}
+{"QuestionId":53353566,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-17T17:13:10.300","AcceptedAnswerId":null,"OwnerUserId":1270741.0,"Title":"How to set input_tensor on a loaded model","Body":"<p>I've created a model based on VGG16:<\/p>\n\n<pre><code>base_model = VGG16(weights='imagenet', include_top=False,\n                   input_tensor=next_batch[\"image\"],\n                   input_shape=INPUT_SHAPE)\nx = base_model.output\n# model customization follows, not relevant\n<\/code><\/pre>\n\n<p>Note that in the snippet I specify <code>input_tensor<\/code> which comes from <code>tf.data.Dataset<\/code>.<\/p>\n\n<p>I've trained the model for some epochs and then saved it using <code>model.save(\"model.h5\")<\/code><\/p>\n\n<p>Now, when I try to load the model using <code>load_model(\"model.h5\")<\/code> and continue training it, I get:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n  tensorflow.python.framework.errors_impl.InvalidArgumentError: You must\n  feed a value for placeholder tensor 'input_1' with dtype float and\n  shape [?,512,512,3]\n           [[{{node input_1}} = Placeholderdtype=DT_FLOAT, shape=[?,512,512,3],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>So the question is, how to specify <code>input_tensor<\/code> on a loaded model?<\/p>\n","answers":[{"AnswerId":"53362876","CreationDate":"2018-11-18T16:07:21.773","ParentId":null,"OwnerUserId":"1270741","Title":null,"Body":"<p>After reading through the code responsible for model loading, I came up with the following solution:<\/p>\n\n<pre><code>class InputLayerFix:\n    @staticmethod\n    def from_config(config):\n        return InputLayer(input_tensor=next_batch[\"image\"],\n                          dtype=config[\"dtype\"],\n                          name=config[\"name\"],\n                          batch_input_shape=config[\"batch_input_shape\"])\n\ncustom_objects = {\"InputLayer\": InputLayerFix}\nmodel.load_model(\"model.h5\", custom_objects=custom_objects)\n<\/code><\/pre>\n\n<p>This code relies on the fact that the deserialization routine first checks <code>custom_objects<\/code> for classes to instantiate (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/keras\/utils\/generic_utils.py#L155\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/keras\/utils\/generic_utils.py#L155<\/a>)<\/p>\n"}]}
+{"QuestionId":53353925,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-11-17T17:50:51.867","AcceptedAnswerId":null,"OwnerUserId":10609392.0,"Title":"Tensorflow import failed - SystemError: initialization of _message raised unreported exception","Body":"<p>I am currently running Anaconda 1.9.2 on Ubuntu 18.04. I updated to the Anaconda 1.9.2 version and since then I am facing problems to import tensorflow, pandas, sklearn etc. I get the following error message when I'm trying to import.<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Python 3.6.5 |Anaconda, Inc.| (default, Apr 29 2018, 16:14:56) \n[GCC 7.2.0] on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n**&gt;&gt;&gt; import tensorflow**\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/__init__.py\", line 52, in &lt;module&gt;\n    from tensorflow.core.framework.graph_pb2 import *\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/core\/framework\/graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/google\/protobuf\/descriptor.py\", line 40, in &lt;module&gt;\n    from google.protobuf.internal import api_implementation\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/google\/protobuf\/internal\/api_implementation.py\", line 154, in &lt;module&gt;\n    from google.protobuf.pyext import _message\nSystemError: initialization of _message raised unreported exception\n<\/code><\/pre>\n\n<pre class=\"lang-none prettyprint-override\"><code>&gt;&gt;&gt; import pandas\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/__init__.py\", line 42, in &lt;module&gt;\n    from pandas.core.api import *\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/api.py\", line 10, in &lt;module&gt;\n    from pandas.core.groupby.groupby import Grouper\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/groupby\/__init__.py\", line 2, in &lt;module&gt;\n    from pandas.core.groupby.groupby import (\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/groupby\/groupby.py\", line 49, in &lt;module&gt;\n    from pandas.core.frame import DataFrame\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/frame.py\", line 66, in &lt;module&gt;\n    from pandas.core.generic import NDFrame, _shared_docs\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/generic.py\", line 102, in &lt;module&gt;\n    class NDFrame(PandasObject, SelectionMixin):\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/pandas\/core\/generic.py\", line 2133, in NDFrame\n    protocol=pkl.HIGHEST_PROTOCOL):\nAttributeError: module 'pickle' has no attribute 'HIGHEST_PROTOCOL'\n\n\n&gt;&gt;&gt; import numpy\n&gt;&gt;&gt; import sklearn\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/__init__.py\", line 64, in &lt;module&gt;\n    from .base import clone\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/base.py\", line 13, in &lt;module&gt;\n    from .utils.fixes import signature\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/utils\/__init__.py\", line 13, in &lt;module&gt;\n    from .validation import (as_float_array,\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/utils\/validation.py\", line 27, in &lt;module&gt;\n    from ..utils._joblib import Memory\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/utils\/_joblib.py\", line 18, in &lt;module&gt;\n    from ..externals.joblib import __all__   # noqa\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/externals\/joblib\/__init__.py\", line 112, in &lt;module&gt;\n    from .memory import Memory, MemorizedResult, register_store_backend\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/externals\/joblib\/memory.py\", line 24, in &lt;module&gt;\n    from . import hashing\n  File \"\/home\/deepak\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/sklearn\/externals\/joblib\/hashing.py\", line 22, in &lt;module&gt;\n    Pickler = pickle._Pickler\nAttributeError: module 'pickle' has no attribute '_Pickler'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53353978,"AnswerCount":1,"Tags":"<tensorflow><gensim><word2vec><tensorboard>","CreationDate":"2018-11-17T17:57:54.303","AcceptedAnswerId":null,"OwnerUserId":9690788.0,"Title":"How to project my word2vec model in Tensorflow","Body":"<p>I am new to using word embedding and want to know how i can project my model in Tensorflow. I was looking at the tensorflow website and it only accepts tsv file (vector\/metadata), but don't know how to generate the required tsv files. I have tried looking it up and can't find any solutions regrading this. Will I try saving my model in a tsv file format, will i need to do some transformations? Any help will be appreciated.<\/p>\n\n<p>I have saved my model as the following files, and just load it up when I need to use it:<\/p>\n\n<blockquote>\n  <p>word2vec.model<\/p>\n  \n  <p>word2vec.model.wv.vectors.npy<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53354320","CreationDate":"2018-11-17T18:39:04.680","ParentId":null,"OwnerUserId":"4620420","Title":null,"Body":"<p>Assuming you're trying to load some pre-trained Gensim word embeddings into a model, you can do this directly with the following code..<\/p>\n\n<pre><code>import numpy\nimport tensorflow as tf\nfrom   gensim.models import KeyedVectors\n\n# Load the word-vector model\nwvec_fn = 'wvecs.kv'\nwvecs = KeyedVectors.load(wvec_fn, mmap='r')\nvec_size = wvecs.vector_size\nvocab_size = len(wvecs.vocab)\n\n# Create the embedding matrix where words are indexed alphabetically\nembedding_mat = numpy.zeros(shape=(vocab_size, vec_size), dtype='int32')\nfor idx, word in enumerate(sorted(wvecs.vocab)):\n    embedding_mat[idx] = wvecs.get_vector(word)\n\n# Setup the embedding matrix for tensorflow\nwith tf.variable_scope(\"input_layer\"):\n    embedding_tf = tf.get_variable(\n       \"embedding\", [vocab_size, vec_size],\n        initializer=tf.constant_initializer(embedding_mat),\n        trainable=False)\n\n# Integrate this into your model\nbatch_size = 32     # just for example\nseq_length = 20\ninput_data = tf.placeholder(tf.int32, [batch_size, seq_length])\ninputs = tf.nn.embedding_lookup(embedding_tf, input_data)\n<\/code><\/pre>\n\n<p>If you've save a model instead of just the KeyedVectors, you may need to modify the code to load the model and then access the KeyedVectors with <code>model.wv<\/code>.<\/p>\n"}]}
+{"QuestionId":53354176,"AnswerCount":2,"Tags":"<python><tensorflow><loss-function><precision-recall>","CreationDate":"2018-11-17T18:20:40.807","AcceptedAnswerId":53355719.0,"OwnerUserId":7382402.0,"Title":"How to use F-score as error function to train neural networks?","Body":"<p>I am pretty new to neural networks. I am training a network in tensorflow, but the number of positive examples is much much less than negative examples in my dataset (it is a medical dataset). \nSo, I know that F-score calculated from precision and recall is a good measure of how well the model is trained. \nI have used error functions like cross-entropy loss or MSE before, but they are all based on accuracy calculation (if I am not wrong). But how do I use this F-score as an error function? Is there a tensorflow function for that? Or I have to create a new one?<\/p>\n\n<p>Thanks in advance. <\/p>\n","answers":[{"AnswerId":"53355719","CreationDate":"2018-11-17T21:24:29.323","ParentId":null,"OwnerUserId":"4267414","Title":null,"Body":"<p>I think you are confusing model evaluation metrics for classification with training losses.<\/p>\n\n<p>Accuracy, precision, F-scores etc. are evaluation metrics computed from binary outcomes and binary predictions. <\/p>\n\n<p>For model training, you need a function that compares a continuous score (your model output) with a binary outcome - like cross-entropy. Ideally, this is calibrated such that it is minimised if the predicted mean matches the population mean (given covariates). These rules are called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Scoring_rule\" rel=\"nofollow noreferrer\">proper scoring rules<\/a>, and the cross-entropy is one of them.<\/p>\n\n<p>Also check the thread <a href=\"https:\/\/stats.stackexchange.com\/questions\/359909\/is-accuracy-an-improper-scoring-rule-in-a-binary-classification-setting\">is-accuracy-an-improper-scoring-rule-in-a-binary-classification-setting<\/a><\/p>\n\n<p>If you want to weigh positive and negative cases differently, two methods are<\/p>\n\n<ul>\n<li>oversample the minority class and correct predicted probabilities when predicting on new examples. For fancier methods, check the under sampling module of <a href=\"https:\/\/github.com\/scikit-learn-contrib\/imbalanced-learn\" rel=\"nofollow noreferrer\"><code>imbalanced-learn<\/code><\/a> to get an overview.<\/li>\n<li>use a different proper scoring rule for training loss. This allows to e.g. build in asymmetry in how you treat positive and negative cases while preserving calibration. <a href=\"https:\/\/www-bcf.usc.edu\/~shaddin\/cs699fa17\/docs\/GR07.pdf\" rel=\"nofollow noreferrer\">Here<\/a> is review of the subject.<\/li>\n<\/ul>\n\n<p>I recommend just using simple oversampling in practice.<\/p>\n"},{"AnswerId":"57740822","CreationDate":"2019-08-31T18:50:34.640","ParentId":null,"OwnerUserId":"8010295","Title":null,"Body":"<p>It appears approaches for optimising directly for these types of metrics have been devised and used successfully, improving scoring and or training times:<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/c\/human-protein-atlas-image-classification\/discussion\/77289\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/human-protein-atlas-image-classification\/discussion\/77289<\/a><\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/c\/human-protein-atlas-image-classification\/discussion\/70328\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/human-protein-atlas-image-classification\/discussion\/70328<\/a><\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/rejpalcz\/best-loss-function-for-f1-score-metric\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/rejpalcz\/best-loss-function-for-f1-score-metric<\/a><\/p>\n\n<p>One such method involves using the sums of probabilities, in place of counts, for the sets of true positives, false positives, and false negative metrics. For example F-beta loss (the generalisation of F1) can be calculated in with Torch in Python as follows:<\/p>\n\n<pre><code>def forward(self, y_logits, y_true):\n    y_pred = self.sigmoid(y_logits)\n    TP = (y_pred * y_true).sum(dim=1)\n    FP = ((1 - y_pred) * y_true).sum(dim=1)\n    FN = (y_pred * (1 - y_true)).sum(dim=1)\n    fbeta = (1 + self.beta**2) * TP \/ ((1 + self.beta**2) * TP + (self.beta**2) * FN + FP + self.epsilon)\n    fbeta = fbeta.clamp(min=self.epsilon, max=1 - self.epsilon)\n    return 1 - fbeta.mean()\n<\/code><\/pre>\n\n<p>An alternative method is described in this paper:<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1608.04802\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1608.04802<\/a><\/p>\n\n<p>The approach taken optimises for a lower bound on the statistic. Other metrics such as AUROC and AUCPR are also discussed. An implementation in TF of such an approach can be found here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/global_objectives\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/global_objectives<\/a><\/p>\n"}]}
+{"QuestionId":53354376,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network><knn>","CreationDate":"2018-11-17T18:44:10.573","AcceptedAnswerId":null,"OwnerUserId":10597829.0,"Title":"Feed CNN features to KNN algorithm in python","Body":"<p>I want to build an end-to-end trainable model that do:<\/p>\n\n<ul>\n<li>extract features from face image by using CNN<\/li>\n<li>apply KNN algorithm on the previously extracted features<\/li>\n<\/ul>\n\n<p>I build my cnn network using python3.5<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=4)\n\nX_train = X_train.astype('float32')\nX_test = X_test.astype('float32')\n\nX_train \/= 255\nX_test \/= 255\n\nY_train = np_utils.to_categorical(y_train, nb_classes)\nY_test = np_utils.to_categorical(y_test, nb_classes)    \nmodel = Sequential()\n    X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols)\n    X_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)\n\n    model.add(Convolution2D(nb_filters, (nb_conv, nb_conv),\n                            padding='valid',\n                            input_shape=(1, img_cols, img_rows),\n                            data_format='channels_first'))\n    model.add(Activation('relu'))\n    model.add(Convolution2D(nb_filters, (nb_conv, nb_conv)))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(nb_pool, nb_pool)))\n    model.add(Dropout(0.5))\n    model.add(Flatten())\n    model.add(Dense(128))\n    model.add(Activation('relu'))\n    model.add(Dropout(0.5))\n    model.add(Dense(nb_classes))\n    model.add(Activation('softmax'))\n\n    model.compile(loss='categorical_crossentropy', optimizer='adadelta', metrics=['accuracy'])\n\n\n    hist = model.fit(X_train, Y_train, batch_size=batch_size, epochs=nb_epoch,\n                     verbose=1, validation_data=(X_test, Y_test))\n\n    hist = model.fit(X_train, Y_train, batch_size=batch_size, epochs=nb_epoch,\n                     verbose=1, validation_split=0.2)\n<\/code><\/pre>\n\n<p>and also i have the knn algorithm from <a href=\"https:\/\/machinelearningmastery.com\/tutorial-to-implement-k-nearest-neighbors-in-python-from-scratch\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/tutorial-to-implement-k-nearest-neighbors-in-python-from-scratch\/<\/a><\/p>\n\n<pre><code>def loadDataset(filename, split, trainingSet=[], testSet=[]):\n    with open(filename, 'r') as csvfile:\n        lines = csv.reader(csvfile)\n        dataset = list(lines)\n        for x in range(len(dataset) - 1):\n            for y in range(4):\n                dataset[x][y] = float(dataset[x][y])\n            if random.random() &lt; split:\n                trainingSet.append(dataset[x])\n            else:\n                testSet.append(dataset[x])\n\ndef euclideanDistance(instance1, instance2, length):\n    distance = 0\n    for x in range(length):\n        distance += pow((instance1[x] - instance2[x]), 2)\n    return math.sqrt(distance)\n\ndef getNeighbors(trainingSet, testInstance, k):\n    distances = []\n    length = len(testInstance) - 1\n    for x in range(len(trainingSet)):\n        dist = euclideanDistance(testInstance, trainingSet[x], length)\n        distances.append((trainingSet[x], dist))\n    distances.sort(key=operator.itemgetter(1))\n    neighbors = []\n    for x in range(k):\n        neighbors.append(distances[x][0])\n    return neighbors\n\ndef getResponse(neighbors):\n    classVotes = {}\n    for x in range(len(neighbors)):\n        response = neighbors[x][-1]\n        if response in classVotes:\n            classVotes[response] += 1\n        else:\n            classVotes[response] = 1\n    sortedVotes = sorted(classVotes.items(), key=operator.itemgetter(1), reverse=True)\n    return sortedVotes[0][0]\n<\/code><\/pre>\n\n<p>I want to use the extracted features from cnn and feed them into the knn algorithm. How can i get, visualize the extracted features and interact with them by taking each k nearest features and classify them in specific group in knn? I cannot able to link the two codes and get the required result. <\/p>\n","answers":[]}
+{"QuestionId":53355034,"AnswerCount":1,"Tags":"<machine-learning><keras><time-series><generator><autoregressive-models>","CreationDate":"2018-11-17T19:59:29.057","AcceptedAnswerId":53539923.0,"OwnerUserId":10235001.0,"Title":"Using sample_weights with fit_generator()","Body":"<p>Inside an autoregressive continuous problem, when the zeros take too much place, it is possible to treat the situation as a zero-inflated problem (i.e. ZIB). In other words, instead of working to fit <code>f(x)<\/code>, we want to fit <code>g(x)*f(x)<\/code> where <code>f(x)<\/code> is the function we want to approximate, i.e. <code>y<\/code>, and <code>g(x)<\/code> is a function which output a value between 0 and 1 depending if a value is zero or non-zero. <\/p>\n\n<p>Currently, I have two models. One model which gives me <code>g(x)<\/code> and another model which fits <code>g(x)*f(x)<\/code>. <\/p>\n\n<p>The first model gives me a set of weights. This is where I need your help. I can use the <code>sample_weights<\/code> arguments with <code>model.fit()<\/code>. As I work with tremendous amount of data, then I need to work with <code>model.fit_generator()<\/code>. However, <code>fit_generator()<\/code> does not have the argument <code>sample_weights<\/code>. <\/p>\n\n<p>Is there a work around to work with <code>sample_weights<\/code> inside <code>fit_generator()<\/code>? Otherwise, how can I fit <code>g(x)*f(x)<\/code> knowing that I have already a trained model for <code>g(x)<\/code>?<\/p>\n","answers":[{"AnswerId":"53539923","CreationDate":"2018-11-29T13:16:26.150","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can provide sample weights as the third element of the tuple returned by the generator. From Keras documentation on <a href=\"https:\/\/keras.io\/models\/sequential\/#fit_generator\" rel=\"noreferrer\"><code>fit_generator<\/code><\/a>:<\/p>\n\n<blockquote>\n  <p><strong>generator:<\/strong> A generator or an instance of <code>Sequence<\/code> (<code>keras.utils.Sequence<\/code>) object in order to avoid duplicate data when using multiprocessing. The output of the generator must be either<\/p>\n  \n  <ul>\n  <li>a tuple <code>(inputs, targets)<\/code><\/li>\n  <li>a tuple <code>(inputs, targets, sample_weights)<\/code>.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p><strong>Update:<\/strong> Here is a rough sketch of a generator that returns the input samples and targets as well as the sample weights obtained from model <code>g(x)<\/code>:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>def gen(args):\n    while True:\n        for i in range(num_batches):\n            # get the i-th batch data\n            inputs = ...\n            targets = ...\n\n            # get the sample weights\n            weights = g.predict(inputs)\n\n            yield inputs, targets, weights\n\n\nmodel.fit_generator(gen(args), steps_per_epoch=num_batches, ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53355055,"AnswerCount":0,"Tags":"<javascript><python><tensorflow><tensorflow-estimator><tensorflowjs-converter>","CreationDate":"2018-11-17T20:02:01.387","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow Estimator to TensorFlow JS","Body":"<p>With the help of the Stack Overflow community, I have started to understand the complexity of TensorFlow's \"high level\" API, <code>Estimator<\/code>. I, perhaps foolishly, did so thinking that if TensorFlow and TensorFlow JS were to get along, most likely it would be through the <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> API... <\/p>\n\n<p>So in this <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">Colab<\/a>, I have a simple custom <a href=\"https:\/\/www.tensorflow.org\/guide\/custom_estimators\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> which is simply wired up so that methods like <code>.train_and_evaluate<\/code> work and after \"training\" the <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> is exported via <code>export_savedmodel<\/code>. <\/p>\n\n<p>Now, suppose I want to go ahead and use this trained model in the browser via TensorFlow js. <\/p>\n\n<p>To my luck, there is a <a href=\"https:\/\/js.tensorflow.org\/tutorials\/import-saved-model.html\" rel=\"nofollow noreferrer\">guide<\/a> on how to convert a <code>saved_model<\/code> for TensorFlow JS:<\/p>\n\n<pre><code>tensorflowjs_converter \\\n    --input_format=tf_saved_model \\\n    --output_node_names='MobilenetV1\/Predictions\/Reshape_1' \\\n    --saved_model_tags=serve \\\n    \/mobilenet\/saved_model \\\n    \/mobilenet\/web_model\n<\/code><\/pre>\n\n<p>While the arguments have a description...<\/p>\n\n<ul>\n<li><code>--output_node_names<\/code> The names of the output nodes, separated by commas.<\/li>\n<li><code>--saved_model_tags<\/code>  Only applicable to SavedModel conversion, Tags of the MetaGraphDef to load, in comma separated format. Defaults to serve.<\/li>\n<li><code>--signature_name<\/code>    Only applicable to TensorFlow Hub module conversion, signature to load. Defaults to default. See <a href=\"https:\/\/www.tensorflow.org\/hub\/common_signatures\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/hub\/common_signatures\/<\/a>.<\/li>\n<\/ul>\n\n<p>I am not sure what I am supposed to replace these with for the demo estimator found in the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">Colab<\/a>.<\/p>\n\n<p>Why? Well for starters, the <code>best_model<\/code> exporter, which uses the <code>serving_input_receiver_fn<\/code>, has different output when loaded via <\/p>\n\n<pre><code>from tensorflow.contrib import predictor\npredict_fn = predictor.from_saved_model('&lt;exported_location&gt;')\n<\/code><\/pre>\n\n<p>than from <\/p>\n\n<pre><code>estimator.predict(lambda: predict_input_fn(pred_features), yield_single_examples=False)\n<\/code><\/pre>\n\n<p>namely, the key of the predicted features features outputted in the former is \"outputs\" and \"labels\" in the latter.<\/p>\n\n<p>which, forgive my mini-rant, why doesn't <code>Estimator<\/code> have a way to load an exported model?<\/p>\n\n<p>so:<\/p>\n\n<ol>\n<li><p>what are my \"<code>output_node_names<\/code>\"?<\/p><\/li>\n<li><p>since this is a SavedModel, what <code>tags<\/code> do I need?<\/p><\/li>\n<\/ol>\n\n<p>I would greatly appreciate any guidance. <\/p>\n","answers":[]}
+{"QuestionId":53355155,"AnswerCount":1,"Tags":"<python><keras><deep-learning><conv-neural-network><transfer-learning>","CreationDate":"2018-11-17T20:14:03.853","AcceptedAnswerId":53355732.0,"OwnerUserId":7631183.0,"Title":"ValueError: Error when checking input: expected input_2 to have shape (224, 224, 3) but got array with shape (244, 244, 3)","Body":"<p>I am trying to use a pretained CNN (VGG16) but I keep getting the following error: <\/p>\n\n<pre><code>ValueError: Error when checking input: expected input_2 to have shape (224, 224, 3) but got array with shape (244, 244, 3)\n<\/code><\/pre>\n\n<p>Here is my full code: <\/p>\n\n<pre><code>import numpy as np \nimport keras \nfrom keras import backend as K \nfrom keras.models import Sequential \nfrom keras.layers import Activation \nfrom keras.layers.core import Dense, Flatten \nfrom keras.optimizers import Adam \nfrom keras.metrics import categorical_crossentropy \nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.layers.normalization import BatchNormalization \nfrom keras.layers.convolutional import *\n\ntrain_path = \"\/DATA\/train\"\nvalid_path = \"\/DATA\/valid\"\ntest_path = \"\/DATA\/test\"\n#creating the training, testing, and validation sets \ntrainBatches = ImageDataGenerator().flow_from_directory(train_path, target_size=(244,244), classes=['classU', 'classH'], batch_size=20)\nvalBatches = ImageDataGenerator().flow_from_directory(valid_path, target_size=(244,244), classes=['classU', 'classH'], batch_size=2)\ntestBatches = ImageDataGenerator().flow_from_directory(test_path, target_size=(244,244), classes=['classU', 'classH'], batch_size=2)\n#loading the model &amp; removing the top layer \nmodel = Sequential() \nfor layer in vgg16_model.layers[:-1]:\n    model.add(layer)\n\n#Fixing the weights \nfor layer in model.layers:\n    layer.trainable = False\n\n#adding the new classier \nmodel.add(Dense(2, activation = 'softmax'))\n\n\nmodel.compile(Adam(lr=0.0001), loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit_generator(trainBatches, steps_per_epoch=89, validation_data=valBatches, validation_steps=11, epochs=5, verbose=2)\n<\/code><\/pre>\n\n<p>But I don't know what I am getting the error. I thought ImageDataGenerator() will take care of the data\/batches generation with the correct dimensions. What I am missing? <\/p>\n","answers":[{"AnswerId":"53355732","CreationDate":"2018-11-17T21:26:24.030","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>The VGG model in this case expects images to be of <code>(224, 224)<\/code> whereas your image generator targets are <code>(244, 244)<\/code> hence your input shapes get a mismatch. You should adjust the target size to the expected shape. The <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"nofollow noreferrer\">documentation<\/a> details the expected inputs and it also has an option <code>include_top<\/code> that will remove the last layer for you so you don't have to do it manually.<\/p>\n"}]}
+{"QuestionId":53355326,"AnswerCount":1,"Tags":"<pytorch><cross-validation>","CreationDate":"2018-11-17T20:35:34.230","AcceptedAnswerId":null,"OwnerUserId":9247213.0,"Title":"Stratified cross validation with Pytorch","Body":"<p>My goal is to make binary classification, using neural network.\nThe problem is that dataset is unbalanced, I have 90% of class 1 and 10 of class 0.\nTo deal with it I want to use Stratified cross-validation. <\/p>\n\n<p>The problem that is I am working with Pytorch, I can't find any example and documentation doesn't provide it, and I'm student, quite new for neural networks.<\/p>\n\n<p>Can anybody help?\nThank you!<\/p>\n","answers":[{"AnswerId":"54488513","CreationDate":"2019-02-01T23:36:41.013","ParentId":null,"OwnerUserId":"1838233","Title":null,"Body":"<p>Have a look at <a href=\"https:\/\/github.com\/dnouri\/skorch\" rel=\"nofollow noreferrer\">skorch<\/a>. It's a scikit-learn compatible neural network library that wraps PyTorch. It has a function <a href=\"https:\/\/skorch.readthedocs.io\/en\/latest\/user\/dataset.html?highlight=fold#cvsplit\" rel=\"nofollow noreferrer\">CVSplit<\/a> for cross validation or you can use sklearn.\nFrom the docs:<\/p>\n\n<pre><code>net = NeuralNetClassifier(\n   module=MyModule,\n   train_split=None,\n)\nfrom sklearn.model_selection import cross_val_predict\ny_pred = cross_val_predict(net, X, y, cv=5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53356029,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2018-11-17T22:14:42.130","AcceptedAnswerId":53456660.0,"OwnerUserId":5623899.0,"Title":"TensorFlow v1.10+ Serving Custom Estimator?","Body":"<p>There are several questions regarding serving with TensorFlow e.g.<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/47856879\/saving-custom-estimators-in-tensorflow\">Saving custom estimators in TensorFlow<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/49040715\/tensorflow-serving\">Tensorflow serving<\/a><\/p><\/li>\n<\/ul>\n\n<p>However many, that I have found, are outdated, related to the <code>Estimator<\/code>'s <code>export_outputs<\/code>, or use a different API (e.g. C#).<\/p>\n\n<p>Further, the \"Basic\" <a href=\"https:\/\/www.tensorflow.org\/serving\/serving_basic#load_exported_model_with_standard_tensorflow_modelserver\" rel=\"nofollow noreferrer\">Serving Guide<\/a> is anything but. It assumes familiarity with <a href=\"https:\/\/www.docker.com\" rel=\"nofollow noreferrer\">docker<\/a>, requires use of a separate TensorFlow <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow noreferrer\">repo<\/a>, and the guide for loading a model is limited to what follows:<\/p>\n\n<blockquote>\n  <p><strong>Load exported model with standard TensorFlow ModelServer<\/strong><\/p>\n  \n  <p>Use a Docker serving image to easily load the model for serving:<\/p>\n\n<pre><code>docker run -p 8500:8500 \\\n--mount type=bind,source=\/tmp\/mnist,target=\/models\/mnist \\\n-e MODEL_NAME=mnist -t tensorflow\/serving &amp;\n<\/code><\/pre>\n<\/blockquote>\n\n<p>without bothering to explain what any of the arguments mean and how one would begin to adapt it to a custom estimator.<\/p>\n\n<p>So here is a <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">simple custom estimator<\/a><\/p>\n\n<p>Can someone please explain to me in layman's terms, for someone who does not know what gRPC service is, how to take my exported model from the above <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> (say I download the files directory as is from the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> to <code>\/tmp\/colab\/&lt;contents-from-linked-colab&gt;<\/code>) and serve it (preferably without docker)<\/p>\n","answers":[{"AnswerId":"53405813","CreationDate":"2018-11-21T05:35:51.757","ParentId":null,"OwnerUserId":"7975433","Title":null,"Body":"<p>In your linked <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> code, after you <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub#scrollTo=bRaPjeKJaT_S\" rel=\"nofollow noreferrer\">run your estimator<\/a>, you should have a <code>saved_model.pb<\/code> and <code>\/variables<\/code> folder in the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a>'s default filesystem. I will refer to where these files are as <code>OUTPUT_PATH<\/code>.<\/p>\n\n<p>To figure out what the <code>OUTPUT_PATH<\/code> is, let us take a quick look at the relevant code from the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> here:<\/p>\n\n<blockquote>\n  <h2><a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub#scrollTo=sB-39Bp-7A_O\" rel=\"nofollow noreferrer\">Estimator > define exporter<\/a><\/h2>\n\n<pre><code>exporter = tf.estimator.BestExporter(\n    name=\"best_exporter\",\n    serving_input_receiver_fn=serving_input_receiver_fn,\n    exports_to_keep=5\n) # this will keep the 5 best checkpoints\n<\/code><\/pre>\n<\/blockquote>\n\n<p>and <\/p>\n\n<blockquote>\n  <h2><a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub#scrollTo=IOXTadwdWsxL&amp;line=2&amp;uniqifier=1\" rel=\"nofollow noreferrer\">Estimator > init estimator<\/a><\/h2>\n\n<pre><code>est = tf.estimator.Estimator(\n    model_fn = model_fn,\n    config = run_config, # &lt;--- model_dir is set in here\n    params = run_params,\n)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Since under, <code>Setup &gt; Constants<\/code> you define <code>MODEL_DIR = '.\/test'<\/code> your <code>BestExporter<\/code> is saved under <code>test\/export\/best_exporter\/&lt;model_num&gt;\/<\/code><\/p>\n\n<p>So your <code>OUTPUT_PATH<\/code> is equal to that. <\/p>\n\n<p>Download this folder to where you want to store you results. For better readability, rename <code>&lt;model_num&gt;<\/code> with something meaningful, e.g. <code>test\/export\/best_exporter\/demo_model<\/code><\/p>\n\n<p>Serving with docker is for clarity, and with a modified docker command:<\/p>\n\n<pre><code>docker run -p 8500:8500 \\\n--mount type=bind,\\\n        source=$OUTPUT_PATH,\\\n        target=\/models\/$MODEL_NAME \\\n-e MODEL_NAME=$MODEL_NAME -t tensorflow\/serving &amp;\n<\/code><\/pre>\n\n<p>For those uninitiated with docker, <code>source=$OUTPUT_PATH,target=\/models\/$MODEL_NAME<\/code> maps the directory <code>OUTPUT_PATH<\/code> to the docker container's directory <code>\/models\/$MODEL_NAME<\/code>. <\/p>\n\n<p>So in this instance you would have:<\/p>\n\n<pre><code>source=&lt;path-to-downloaded-dir&gt;\/test\/export\/best_exporter\/demo_model,\\\ntarget=\/models\/demo_model,\\\n-e MODEL_NAME=demo_model\n<\/code><\/pre>\n\n<p>as we are assuming that the source is the <code>model_dir<\/code> and where <code>&lt;path-to-downloaded-dir&gt;<\/code> is wherever you downloaded <code>\/test\/export\/best_exporter\/demo_model<\/code>. <\/p>\n\n<p>Then following the grpc example to write the client.\nIf you prefer RESTful API, maybe you need to change the docker port to <code>8501:8501<\/code> or using both together <code>8500-8501:8500-8501<\/code>. Here is my another <a href=\"https:\/\/stackoverflow.com\/questions\/53366768\/serving-a-keras-model-with-tensorflow-serving\/53405554#53405554\">answer<\/a> to explain this docker command.<\/p>\n\n<p>If you don't want docker, try to install the tf-serving <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/g3doc\/setup.md#installing-using-apt\" rel=\"nofollow noreferrer\">locally<\/a>, almost the same command to run a server.<\/p>\n"},{"AnswerId":"53456660","CreationDate":"2018-11-24T08:58:05.010","ParentId":null,"OwnerUserId":"10076057","Title":null,"Body":"<p>Suppose you have a custom estimator, trained and ready to serve, like the one linked in your question. The procedure for saving and serving a trained estimator model is:<\/p>\n\n<ol>\n<li>Export the Estimator to the SavedModel format.<\/li>\n<li>Serve the SavedModel using the TensorFlow ModelServer.<\/li>\n<li>Feed inputs to the served model and observe the prediction results.<\/li>\n<\/ol>\n\n<p>For some use cases your trained estimator model might be better deployed and reused without serving it. Sometimes it's better to freeze a model and deploy it directly inside a program. Or sometimes you want to convert a model to the javascript or lite versions of TensorFlow. There are many ways to reuse a trained estimator without serving it. But since your question specifically asks about serving, this answer is about specifically serving with the standard ModelServer.<\/p>\n\n<h2>1. Export to SavedModel format<\/h2>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#using_savedmodel_with_estimators\" rel=\"nofollow noreferrer\">From the docs<\/a>:<\/p>\n\n<blockquote>\n  <p>To prepare a trained Estimator for serving, you must export it in the standard SavedModel format.<\/p>\n<\/blockquote>\n\n<p>For this we can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator#export_saved_model\" rel=\"nofollow noreferrer\"><code>export_saved_model<\/code><\/a> function, and doing that requires that we first define a serving input receiver function. The serving input receiver function specifies and names all the tensors that become inputs to the model at serve time.<\/p>\n\n<p>There are two kinds of serving input receiver functions and each type tells TensorFlow how inputs should be expected in step 3:<\/p>\n\n<ul>\n<li><p><strong>Parsing serving input receiver functions<\/strong>: The inputs are provided as serialized <code>Example<\/code> protobufs. This type tells TensorFlow to expect the model input to come from a string tensor that will be parsed into features. This receiver can be built using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/build_parsing_serving_input_receiver_fn\" rel=\"nofollow noreferrer\"><code>tf.estimator.export.build_parsing_serving_input_receiver_fn<\/code><\/a>.<\/p><\/li>\n<li><p><strong>Raw serving input receiver functions<\/strong>: Inputs are provided directly as <code>Tensor<\/code> protobufs. This receiver can be built using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/build_raw_serving_input_receiver_fn\" rel=\"nofollow noreferrer\"><code>tf.estimator.export.build_raw_serving_input_receiver_fn<\/code><\/a>.<\/p><\/li>\n<\/ul>\n\n<p>Your colab code is building two receiver functions that do the same thing: <\/p>\n\n<pre><code>serving_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(\n    {'input_tensors': tf.placeholder(tf.float32, I_SHAPE(None), name=\"input_tensors\")})\n<\/code><\/pre>\n\n<p>and:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  input_tensors = tf.placeholder(tf.float32, I_SHAPE(None), name=\"input_tensors\")\n\u200b\n\u200b\n  features = {'input_tensors' : input_tensors}\n  receiver_tensors = {'input_tensors': input_tensors}\n  return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n<\/code><\/pre>\n\n<p>But exporting with only one:<\/p>\n\n<pre><code>est.export_savedmodel('.\/here', serving_input_receiver_fn)\n<\/code><\/pre>\n\n<p>You could remove your <code>serving_input_receiver_fn<\/code> method and use the first definition:<\/p>\n\n<pre><code>est.export_savedmodel('.\/here', serving_fn)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>exporter = tf.estimator.BestExporter(\n    name=\"best_exporter\",\n    serving_input_receiver_fn=serving_fn,\n    exports_to_keep=5\n)\n<\/code><\/pre>\n\n<h2>2. Serve the SavedModel<\/h2>\n\n<p>Your question said you prefer to serve your model without using docker. According to <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/tools\/docker\/Dockerfile\" rel=\"nofollow noreferrer\">its Dockerfile<\/a>, the docker image is just running the <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow noreferrer\">TensorFlow ModelServer<\/a> binary, which you can install or build from source outside a container as described in its README, or you can copy it out from the <code>tensorflow\/serving<\/code> container.<\/p>\n\n<p>Once you have the binary installed, run it to launch a gRPC server listening on the port you want, for example 8500:<\/p>\n\n<pre><code>tensorflow_model_server --port=8500 --model_name=my_model --model_base_path=\/path\/to\/export\/dir\n<\/code><\/pre>\n\n<p>Now you are \"serving\" the model. If you only want to run the model without needing anything from the tensorflow_serving repo, you could instead use the <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#cli_to_inspect_and_execute_savedmodel\" rel=\"nofollow noreferrer\">saved model command line interface<\/a> to run the SavedModel without a model server. It should already be installed with TensorFlow if you installed from a pre-built binary.<\/p>\n\n<h2>3. Query the running model server<\/h2>\n\n<p>The standard way of querying the model is with the gRPC service provided by the ModelServer. gRPC is an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Remote_procedure_call\" rel=\"nofollow noreferrer\">RPC<\/a> framework that uses Google's protocol buffer format to define services and to communicate between hosts. It is designed to be fast, cross-platform, and scalable. It is especially convenient when all your data is already handled in protobuf format, like when dealing with TFRecord files.<\/p>\n\n<p>There are gRPC libraries for many different languages, and you can even talk to your server with e.g. cURL, but since your question is tagged for Python I will use the <a href=\"https:\/\/pypi.org\/project\/grpcio\/\" rel=\"nofollow noreferrer\">grpcio<\/a> and <a href=\"https:\/\/pypi.org\/project\/tensorflow-serving-api\/\" rel=\"nofollow noreferrer\">tensorflow-serving-api<\/a> Python packages to execute the gRPC calls needed to predict with the served model.<\/p>\n\n<p>Once the server is running and the Python packages are installed, you can verify the connection by querying the model's signature def metadata:<\/p>\n\n<pre><code>from __future__ import print_function\nimport grpc\nfrom tensorflow_serving.apis import get_model_metadata_pb2\nfrom tensorflow_serving.apis import model_pb2\nfrom tensorflow_serving.apis import prediction_service_pb2_grpc\n\n\nwith grpc.insecure_channel(\"localhost:8500\") as channel:\n  stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n\n  request = get_model_metadata_pb2.GetModelMetadataRequest(\n      model_spec=model_pb2.ModelSpec(name=\"my_model\"),\n      metadata_field=[\"signature_def\"])\n\n  response = stub.GetModelMetadata(request)\n  sigdef_str = response.metadata[\"signature_def\"].value\n\n  print (\"Name:\", response.model_spec.name)\n  print (\"Version:\", response.model_spec.version.value)\n  print (get_model_metadata_pb2.SignatureDefMap.FromString(sigdef_str))\n<\/code><\/pre>\n\n<p>With the model from your colab you would see<\/p>\n\n<pre><code>Name: my_model\nVersion: ...\nsignature_def {\n  key: \"labels\"\n  value {\n    inputs {\n      key: \"input_tensors\"\n      value {\n        name: \"input_tensors:0\"\n        dtype: DT_FLOAT\n        tensor_shape {\n          dim {\n            size: -1\n          }\n          dim {\n            size: 20\n          }\n          dim {\n            size: 7\n          }\n        }\n      }\n    }\n    outputs {\n      key: \"output\"\n      value {\n        name: \"Sigmoid:0\"\n        dtype: DT_FLOAT\n        tensor_shape {\n          dim {\n            size: -1\n          }\n          dim {\n            size: 20\n          }\n          dim {\n            size: 4\n          }\n        }\n      }\n    }\n    method_name: \"tensorflow\/serving\/predict\"\n  }\n}\nsignature_def {\n  key: \"serving_default\"\n  value {\n    inputs {\n      key: \"input_tensors\"\n      value {\n        name: \"input_tensors:0\"\n        dtype: DT_FLOAT\n        tensor_shape {\n          dim {\n            size: -1\n          }\n          dim {\n            size: 20\n          }\n          dim {\n            size: 7\n          }\n        }\n      }\n    }\n    outputs {\n      key: \"output\"\n      value {\n        name: \"Sigmoid:0\"\n        dtype: DT_FLOAT\n        tensor_shape {\n          dim {\n            size: -1\n          }\n          dim {\n            size: 20\n          }\n          dim {\n            size: 4\n          }\n        }\n      }\n    }\n    method_name: \"tensorflow\/serving\/predict\"\n  }\n}\n<\/code><\/pre>\n\n<p>So according to its signature definition the model expects a dictionary mapping an <code>input_tensors<\/code> key to a Tensor proto of floating point type and shape <code>[-1, 20, 7]<\/code> and will output a dictionary mapping an <code>output<\/code> key to a Tensor proto of floating point type and shape <code>[-1, 20, 4]<\/code>. We can create a Tensor proto in Python from a numpy array using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/make_tensor_proto\" rel=\"nofollow noreferrer\"><code>tf.make_tensor_proto<\/code><\/a> and convert back using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/make_ndarray\" rel=\"nofollow noreferrer\"><code>tf.make_ndarray<\/code><\/a>:<\/p>\n\n<pre><code>from __future__ import print_function\nimport grpc\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow_serving.apis import model_pb2\nfrom tensorflow_serving.apis import predict_pb2\nfrom tensorflow_serving.apis import prediction_service_pb2_grpc\n\n\n# Dummy input data for batch size 3.\nbatch_input = np.ones((3, 20, 7), dtype=\"float32\")\n\n\nwith grpc.insecure_channel(\"localhost:8500\") as channel:\n  stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n\n  request = predict_pb2.PredictRequest(\n      model_spec=model_pb2.ModelSpec(name=\"my_model\"),\n      inputs={\"input_tensors\": tf.make_tensor_proto(batch_input)})\n\n  response = stub.Predict(request)\n\n  batch_output = tf.make_ndarray(response.outputs[\"output\"])\n  print (batch_output.shape)\n<\/code><\/pre>\n\n<p>Indeed you should have a floating point array of shape <code>(3, 20, 4)<\/code> returned by your served estimator model.<\/p>\n\n<p>For more information about how gRPC services are defined and used in Python, see the <a href=\"https:\/\/grpc.io\/docs\/tutorials\/basic\/python.html\" rel=\"nofollow noreferrer\">tutorial on the gRPC website<\/a>. For <code>tensorflow_serving<\/code> API details see the <a href=\"https:\/\/github.com\/tensorflow\/serving\/tree\/master\/tensorflow_serving\/apis\" rel=\"nofollow noreferrer\"><code>.proto<\/code> protobuf definitions<\/a>.<\/p>\n"}]}
+{"QuestionId":53356558,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving><tensorflow-estimator>","CreationDate":"2018-11-17T23:31:30.710","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow v1.10+ load SavedModel with different device placement or manually set dynamic device placement?","Body":"<p>So in TensorFlow's guide for using <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu#manual_device_placement\" rel=\"nofollow noreferrer\">GPUs<\/a> there is a part about using multiple GPUs in a \"multi-tower fashion\":<\/p>\n\n<pre><code>...\nfor d in ['\/device:GPU:2', '\/device:GPU:3']:\n  with tf.device(d): # &lt;---- manual device placement\n...\n<\/code><\/pre>\n\n<p>Seeing this, one might be tempted to leverage this style for multiple GPU training in a custom <code>Estimator<\/code> to indicate to the model that it can be distributed across multiple GPUs efficiently.<\/p>\n\n<p>To my knowledge, if manual device placement is absent TensorFlow does not have some form of optimal device mapping (expect perhaps if you have the GPU version installed and a GPU is available, using it over the CPU). So what other choice do you have?<\/p>\n\n<p>Anyway, you carry on with training your estimator and export it to a <code>SavedModel<\/code> via <code>estimator.export_savedmodel(...)<\/code> and wish to use this <code>SavedModel<\/code> later... perhaps on a different machine, one which may not have as many GPUs as the device on which the model was trained (or maybe no GPUs) <\/p>\n\n<p>so when you run <\/p>\n\n<pre><code>from tensorflow.contrib import predictor\npredict_fn = predictor.from_saved_model(model_dir)\n<\/code><\/pre>\n\n<p>you get<\/p>\n\n<pre><code>Cannot assign a device for operation &lt;OP-NAME&gt;. Operation was \nexplicitly assigned to &lt;DEVICE-NAME&gt; but available devices are \n[&lt;AVAILABLE-DEVICE-0&gt;,...]\n<\/code><\/pre>\n\n<p>An older <a href=\"https:\/\/stackoverflow.com\/questions\/42300475\/change-device-placement-of-an-existing-tensorflow-variable\">S.O. Post<\/a> suggests that changing device placement was not possible... but hopefully over time things have changed.<\/p>\n\n<p>Thus my question is:<\/p>\n\n<ol>\n<li><p>when loading a <code>SavedModel<\/code> can I change the device placement to be appropriate for the device it is loaded on. E.g. if I train a model with 6 GPUs and a friend wants to run it at home with their e-GPU, can they set <code>'\/device:GPU:1'<\/code> through  <code>'\/device:GPU:5'<\/code> to <code>'\/device:GPU:0'<\/code>?<\/p><\/li>\n<li><p>if <strong>1<\/strong> is not possible, is there a (painless) way for me, in the custom <code>Estimator<\/code>'s <code>model_fn<\/code>, to specify how to generically distribute a graph? <\/p><\/li>\n<\/ol>\n\n<p>e.g. <\/p>\n\n<pre><code>with tf.device('available-gpu-3')\n<\/code><\/pre>\n\n<p>where <code>available-gpu-3<\/code> is the third available GPU if there are three or more GPUs, otherwise the second or first available GPU, and if no GPU it is CPU<\/p>\n\n<p>This matters because if there is a shared machine with is training two models, say one model on <code>'\/device:GPU:0'<\/code> then the other model is trained explicitly on GPUs 1 and 2... so on another 2 GPU machine, GPU 2 will not be available....<\/p>\n","answers":[{"AnswerId":"53405653","CreationDate":"2018-11-21T05:17:10.893","ParentId":null,"OwnerUserId":"7975433","Title":null,"Body":"<p>I am doing some research on this topic recently and to my knowledge, your question 1 can work only if you clear all devices when you export the model in the original tensorflow code, with flag <code>clear_devices=True<\/code>.<\/p>\n\n<p>In my own code, it looks like<\/p>\n\n<pre><code>builder = tf.saved_model.builder.SavedModelBuilder('osvos_saved')\nbuilder.add_meta_graph_and_variables(sess, ['serve'], clear_devices=True)\nbuilder.save()\n<\/code><\/pre>\n\n<p>If you only have a exported model, seems not possible. You can refer to <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/311\" rel=\"nofollow noreferrer\">this issue<\/a>. <\/p>\n\n<p>I'm currently trying to find a way to fix this, as stated in my <a href=\"https:\/\/stackoverflow.com\/questions\/53358752\/for-a-tensorflow-savedmodel-in-pbtxt-format-where-is-the-device-placement-of-an\">stackoverflow question<\/a>. Hope the workaround can help you.<\/p>\n"}]}
+{"QuestionId":53356783,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><tensorflow>","CreationDate":"2018-11-18T00:14:31.230","AcceptedAnswerId":null,"OwnerUserId":3042117.0,"Title":"How to select a batch from a preloaded dataset in Tensorflow?","Body":"<p>I have a large dataset which can fit into GPU memory, and I would like to select a random batch from it in each step of the training. The dataset consists of two arrays:<\/p>\n\n<pre><code>data1 = np.load(\"data1.npy\")\ndata2 = np.load(\"data2.npy\")\n\nt_data1 = tf.constant(data1)\nt_data2 = tf.constant(data2)\n<\/code><\/pre>\n\n<p><code>data1<\/code> and <code>data2<\/code> has a shape <code>(16000, 200)<\/code>. The batch size is 128, so I would like to select 128 elements from each array with the same indices, and provide it to the optimizer:<\/p>\n\n<pre><code>for i in range(training_steps):\n    choise = np.random.choice(data1.shape[0], batch_size)\n\n    X_batch = t_data1[choise]\n    Y_batch = t_data2[choise]\n\n    sess.run(train_step, feed_dict={X: X_batch, Y: Y_batch})\n<\/code><\/pre>\n\n<p>Unfortunately I get this error:<\/p>\n\n<pre><code>ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice' (op: 'StridedSlice') with input shapes\n<\/code><\/pre>\n\n<p>What do I wrong? How is it possible to generate batch from the data, which is already on the GPU?<\/p>\n","answers":[{"AnswerId":"53358423","CreationDate":"2018-11-18T06:22:34.453","ParentId":null,"OwnerUserId":"7389608","Title":null,"Body":"<p><code>t_data1<\/code> and <code>t_data2<\/code> are constants in tensorflow. You can do this in numpy, but tensorflow doesn't support advanced indexing, you need to use <code>tf.gather()<\/code>.<\/p>\n\n<p>Change to:<\/p>\n\n<pre><code>X_batch = tf.gather(t_data1,choise)\nY_batch = tf.gather(t_data2,choise)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53356846,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-18T00:30:15.200","AcceptedAnswerId":57444978.0,"OwnerUserId":260826.0,"Title":"Tensorflow Hparam replacement","Body":"<p>In TF 1.12 or TF 2.0 is there going to be a replacement for the following function:<\/p>\n\n<p><code>from tensorflow.contrib.training.python.training import hparam<\/code><\/p>\n\n<p>I read that <code>contrib<\/code> module will go away or merge into core.<\/p>\n","answers":[{"AnswerId":"57444978","CreationDate":"2019-08-10T19:10:08.840","ParentId":null,"OwnerUserId":"1237534","Title":null,"Body":"<p>In TF 2.0 there is a new API <code>tensorboard.plugins.hparams.api<\/code> that includes a class <code>HParam<\/code><\/p>\n\n<p>Usage of the new API is described in this guide: <a href=\"https:\/\/www.tensorflow.org\/tensorboard\/r2\/hyperparameter_tuning_with_hparams#1_experiment_setup_and_the_hparams_experiment_summary\" rel=\"nofollow noreferrer\">Hyperparameter Tuning with the HParams Dashboard<\/a><\/p>\n"}]}
+{"QuestionId":53356849,"AnswerCount":1,"Tags":"<python><machine-learning><keras><word-embedding>","CreationDate":"2018-11-18T00:31:01.017","AcceptedAnswerId":53357165.0,"OwnerUserId":866082.0,"Title":"How to train a model with only an Embedding layer in Keras and no labels","Body":"<p>I have some text without any labels. Just a bunch of text files. And I want to train an Embedding layer to map the words to embedding vectors. Most of the examples I've seen so far are like this:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Embedding, Flatten, Dense\n\nmodel = Sequential()\nmodel.add(Embedding(max_words, embedding_dim, input_length=maxlen))\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.summary()\nmodel.compile(optimizer='rmsprop',\n    loss='binary_crossentropy',\n    metrics=['acc'])\nmodel.fit(x_train, y_train,\n    epochs=10,\n    batch_size=32,\n    validation_data=(x_val, y_val))\n<\/code><\/pre>\n\n<p>They all assume that the Embedding layer is part of a bigger model which tries to predict a label. But in my case, I have no label. I'm not trying to classify anything. I just want to train the mapping from words (more precisely integers) to embedding vectors. But the <code>fit<\/code> method of the model, asks for <code>x_train<\/code> and <code>y_train<\/code> (as the example given above).<\/p>\n\n<p>How can I train a model only with an Embedding layer and no labels?<\/p>\n\n<p><strong>[UPDATE]<\/strong><\/p>\n\n<p>Based on the answer I've got from @Daniel M\u00f6ller, Embedding layer in Keras is implementing a supervised algorithm and thus cannot be trained without labels. Initially, I was thinking that it is a variation of Word2Vec and thus does not need labels to be trained. Apparently, that's not the case. Personally, I ended up using the <a href=\"https:\/\/github.com\/facebookresearch\/fastText\" rel=\"nofollow noreferrer\">FastText<\/a> which has nothing to do with Keras or Python.<\/p>\n","answers":[{"AnswerId":"53357165","CreationDate":"2018-11-18T01:37:18.050","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Does it make sense to do that without a label\/target?<\/p>\n\n<p>How will your model decide which values in the vectors are good for anything if there is no objective?<\/p>\n\n<p>All embeddings are \"trained\" for a purpose. If there is no purpose, there is no target, if there is no target, there is no training. <\/p>\n\n<p>If you really want to transform words in vectors without any purpose\/target, you've got two options:<\/p>\n\n<ul>\n<li>Make one-hot encoded vectors. You may use the Keras <code>to_categorical<\/code> function for that.        <\/li>\n<li>Use a pretrained embedding. There are some available, such as glove, embeddings from Google, etc. (All of they were trained at some point for some purpose).    <\/li>\n<\/ul>\n\n<hr>\n\n<h3>A very naive approach based on our chat, considering word distance<\/h3>\n\n<p>Warning: I don't really know anything about Word2Vec, but I'll try to show how to add the rules for your embedding using some naive kind of word distance and how to use dummy \"labels\" just to satisfy Keras' way of training.<\/p>\n\n<pre><code>from keras.layers import Input, Embedding, Subtract, Lambda\nimport keras.backend as K\nfrom keras.models import Model\n\ninput1 = Input((1,)) #word1\ninput2 = Input((1,)) #word2\n\nembeddingLayer = Embedding(...params...)\n\nword1 = embeddingLayer(input1)\nword2 = embeddingLayer(input2)\n\n#naive distance rule, subtract, expect zero difference\nword_distance = Subtract()([word1,word2])\n\n#reduce all dimensions to a single dimension\nword_distance = Lambda(lambda x: K.mean(x, axis=-1))(word_distance)\n\nmodel = Model([input1,input2], word_distance)\n<\/code><\/pre>\n\n<p>Now that our model outputs directly a word distance, our labels will be \"zero\", they're not really labels for a supervised training, but they're the expected result of the model, something necessary for Keras to work.<\/p>\n\n<p>We can have as loss function the <code>mae<\/code> (mean absolute error) or <code>mse<\/code> (mean squared error), for instance.<\/p>\n\n<pre><code>model.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>And training with word2 being the word after word1:<\/p>\n\n<pre><code>xTrain = entireText\nxTrain1 = entireText[:-1]\nxTrain2 = entireText[1:]\nyTrain = np.zeros((len(xTrain1),))\n\nmodel.fit([xTrain1,xTrain2], yTrain, .... more params.... ) \n<\/code><\/pre>\n\n<p>Although this may be completely wrong regarding what Word2Vec really does, it shows the main points that are:<\/p>\n\n<ul>\n<li>Embedding layers don't have special properties, they're just trainable lookup tables    <\/li>\n<li>Rules for creating an embedding should be defined by the model and expected outputs    <\/li>\n<li>A Keras model will need \"targets\", even if those targets are not \"labels\" but a mathematical trick for an expected result.      <\/li>\n<\/ul>\n"}]}
+{"QuestionId":53356891,"AnswerCount":0,"Tags":"<tensorflow><rnn>","CreationDate":"2018-11-18T00:39:44.180","AcceptedAnswerId":null,"OwnerUserId":10664840.0,"Title":"how does the weight update in BasicRnnCell for char level","Body":"<p>When I used the Tensorflow to build a character level RNN network, I was confused by the change of the weight in the model. I thought the weight won't update in one batch. <\/p>\n\n<p>But the kernel(which I think is Wk) was changing. And there were 6 kernels. So I am confused why it changed and why there were 6 kernels. Can I get W and U directly using Tensorflow. Here is my code. Thanks.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.InteractiveSession()\nh = [1, 0, 0, 0]\ne = [0, 1, 0, 0]\nl = [0, 0, 1, 0]\no = [0, 0, 0, 1]\n\n\nwith tf.variable_scope('two_sequances') as scope:\n    # One cell RNN input_dim (4) -&gt; output_dim (2). sequence: 5\n    hidden_size = 2\n    cell = tf.contrib.rnn.BasicRNNCell(num_units=hidden_size)\n    x_data = np.array([[h, e, l, l, o]], dtype=np.float32)\n    print(x_data.shape)\n    print(x_data)\n    outputs, _states = tf.nn.dynamic_rnn(cell, x_data, dtype=tf.float32)\n    sess.run(tf.global_variables_initializer())\n    results, state = sess.run([outputs, _states])\n\n    variable_names = [v.name for v in tf.global_variables()]\n    values = sess.run(variable_names)\n    for k, v in zip(variable_names, values):\n        print(k, v)\n<\/code><\/pre>\n\n<p>And the output of the variables are as follow.<\/p>\n\n<pre><code>two_sequances\/rnn\/basic_rnn_cell\/kernel:0 \n[[ 0.6147509   0.6268855 ]\n [ 0.34818882  0.8140872 ]\n [ 0.4074654   0.011693  ]\n [-0.5032909  -0.69920516]\n [ 0.62231725  0.18967694]\n [ 0.6888749   0.77280706]]\ntwo_sequances\/rnn\/basic_rnn_cell\/bias:0 \n[0. 0.]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53357133,"AnswerCount":0,"Tags":"<python><keras><autoencoder><loss-function>","CreationDate":"2018-11-18T01:31:21.770","AcceptedAnswerId":null,"OwnerUserId":4627855.0,"Title":"Adding custom loss function for variational autoencoder keras","Body":"<p>I am trying to add a custom loss function for variational autoencoder. Along with the reconstruction loss, KL divergence I wish to add a loss based on the difference between hamming distances of pairs of input and outputs.<\/p>\n\n<p>But the problem what I am having is that with or without this extra loss, the results are same. Could anyone point out what I should be doing to correct it? Is it something to do with dimensions or something else.<\/p>\n\n<p>Here is my code snippet:<\/p>\n\n<pre><code>def ham_loss(y_true,y_pred):\n    # calculate pairwise hamming distance matrix\n    # differences of y_pred probabilities)\n    pairwise_diff_pred = K.expand_dims(y_pred, 0) - K.expand_dims(y_pred, 1)\n    pairwise_distance_pred = K.sum(pairwise_diff_pred, axis=-1)\n\n    # calculate pairwise hamming distance matrix for inputs\n    pairwise_diff_true = K.expand_dims(y_true, 0) - K.expand_dims(y_true, 1)\n    pairwise_distance_true = K.sum(pairwise_diff_true, axis=-1)\n\n    #Difference between the distances of y_true and y_predictions\n    hamm_sum= Lambda(differences)([pairwise_distance_true, pairwise_distance_pred])\n    print(hamm_sum)\n    return K.sum(hamm_sum, axis=-1)\n\ndef vae_loss(y_true, y_pred):\n    \"\"\" Calculate loss = reconstruction loss + KL loss for each data in minibatch \"\"\"\n         # E[log P(X|z)]\n\n         recon = K.sum(K.binary_crossentropy(y_true,y_pred),axis=1)\n         # D_KL(Q(z|X) || P(z|X)); calculate in closed form as both dist. are Gaussian\n         kl = 0.5 * K.sum(K.exp(z_log_var) + K.square(z_mean) - 1. - z_log_var, axis=1)\n\n         hamming_loss = ham_loss(y_true,y_pred)\n\n         return recon + kl + hamming_loss\n<\/code><\/pre>\n\n<p>Any help much appreciated!<\/p>\n\n<p>Thanks in advance..<\/p>\n","answers":[]}
+{"QuestionId":53357272,"AnswerCount":1,"Tags":"<macos><tensorflow><bazel><tensorflow-lite>","CreationDate":"2018-11-18T02:02:41.060","AcceptedAnswerId":null,"OwnerUserId":6655715.0,"Title":"bazel build tensorflow graph transform looking for \/usr\/bin\/python3, but file does not exist (and I cannot create it)","Body":"<p>I am following <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/graph_transforms\/README.md#optimizing-for-deployment\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/graph_transforms\/README.md#optimizing-for-deployment<\/a> and running this command:<\/p>\n\n<p><code>bazel build tensorflow\/tools\/graph_transforms:transform_graph<\/code><\/p>\n\n<p>and getting this error:<\/p>\n\n<blockquote>\n  <p>ERROR: \/Users\/daniel\/software\/tensorflow\/tensorflow\/core\/BUILD:2475:1: no such package '@local_config_git\/\/': Traceback (most recent call last):\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 52\n          _fail(result.stderr)\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 14, in _fail\n          fail((\"%sGit Configuration Error:%s %...)))\n  Git Configuration Error: java.io.IOException: Cannot run program \"\/usr\/bin\/python3\" (in directory \"\/private\/var\/tmp\/_bazel_daniel\/bff3ff07c5339b6b7496cc244d119e78\/external\/local_config_git\"): error=2, No such file or directory\n   and referenced by '\/\/tensorflow\/core:version_info_gen'\n  ERROR: \/Users\/daniel\/software\/tensorflow\/tensorflow\/core\/BUILD:2475:1: no such package '@local_config_git\/\/': Traceback (most recent call last):\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 52\n          _fail(result.stderr)\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 14, in _fail\n          fail((\"%sGit Configuration Error:%s %...)))\n  Git Configuration Error: java.io.IOException: Cannot run program \"\/usr\/bin\/python3\" (in directory \"\/private\/var\/tmp\/_bazel_daniel\/bff3ff07c5339b6b7496cc244d119e78\/external\/local_config_git\"): error=2, No such file or directory\n   and referenced by '\/\/tensorflow\/core:version_info_gen'\n  ERROR: \/Users\/daniel\/software\/tensorflow\/tensorflow\/core\/BUILD:2475:1: no such package '@local_config_git\/\/': Traceback (most recent call last):\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 52\n          _fail(result.stderr)\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 14, in _fail\n          fail((\"%sGit Configuration Error:%s %...)))\n  Git Configuration Error: java.io.IOException: Cannot run program \"\/usr\/bin\/python3\" (in directory \"\/private\/var\/tmp\/_bazel_daniel\/bff3ff07c5339b6b7496cc244d119e78\/external\/local_config_git\"): error=2, No such file or directory\n   and referenced by '\/\/tensorflow\/core:version_info_gen'\n  ERROR: Analysis of target '\/\/tensorflow\/tools\/graph_transforms:summarize_graph' failed; build aborted: no such package '@local_config_git\/\/': Traceback (most recent call last):\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 52\n          _fail(result.stderr)\n      File \"\/Users\/daniel\/software\/tensorflow\/third_party\/git\/git_configure.bzl\", line 14, in _fail\n          fail((\"%sGit Configuration Error:%s %...)))\n  Git Configuration Error: java.io.IOException: Cannot run program \"\/usr\/bin\/python3\" (in directory \"\/private\/var\/tmp\/_bazel_daniel\/bff3ff07c5339b6b7496cc244d119e78\/external\/local_config_git\"): error=2, No such file or directory\n  ```<\/p>\n<\/blockquote>\n\n<p>The recurring line of interest (I think) is <code>java.io.IOException: Cannot run program \"\/usr\/bin\/python3\"<\/code>. The error message is correct:<\/p>\n\n<p><code>$ ls \/usr\/bin\/python3\nls: \/usr\/bin\/python3: No such file or directory\n<\/code><\/p>\n\n<p>But I'm not sure why it's looking for <code>\/usr\/bin\/python3<\/code> in the first place, as it's not in my PATH:<\/p>\n\n<p><code>$ type python3\npython3 is \/anaconda3\/bin\/python3\n<\/code><\/p>\n\n<p>Is bazel hardcoded to look for that path? I can't create anything in <code>\/usr\/bin<\/code>, see: <a href=\"https:\/\/stackoverflow.com\/questions\/39928053\/install-python-3-to-usr-bin-on-macos\">Install Python 3 to \/usr\/bin\/ on macOS<\/a><\/p>\n\n<p>I don't really know what to do here. Any advice welcome.<\/p>\n","answers":[{"AnswerId":"53357330","CreationDate":"2018-11-18T02:18:27.163","ParentId":null,"OwnerUserId":"6655715","Title":null,"Body":"<p>[UPDATE] -- Running <code>.\/configure<\/code> from <code>tensorflow\/<\/code> and accepting all the default options fixed my problem. <\/p>\n"}]}
+{"QuestionId":53357560,"AnswerCount":1,"Tags":"<python><tensorflow><keras><google-cloud-platform><deep-learning>","CreationDate":"2018-11-18T03:12:26.070","AcceptedAnswerId":null,"OwnerUserId":10669150.0,"Title":"gcloud ML engine - Keras not running on GPU","Body":"<p>I'm new with Google Cloud Machine Learning Engine, \nI'm trying to train a DL algorithm for image classification based on Keras in gcloud. \nIn order to configure the GPU on gcloud I have included <code>'tensorflow-gpu'<\/code> in the <code>setup.py install_requires<\/code>. \nMy <code>cloud-gpu.yaml<\/code> is the following <\/p>\n\n<pre><code>trainingInput:\n  scaleTier: BASIC_GPU\n  runtimeVersion: \"1.0\"\n<\/code><\/pre>\n\n<p>In the code I've added: <\/p>\n\n<pre><code>sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<p>at the beginning and <\/p>\n\n<pre><code>with tf.device('\/gpu:0'):\n<\/code><\/pre>\n\n<p>before any Keras code. <\/p>\n\n<p>The result is that gcloud is recognising the gpu but not using it, as you can see from <\/p>\n\n<p>Screenshot of actual cloud training:<\/p>\n\n<pre><code>INFO    2018-11-18 12:19:59 -0600   master-replica-0        Epoch 1\/20\nINFO    2018-11-18 12:20:56 -0600   master-replica-0          1\/219 [..............................] - ETA: 4:17:12 - loss: 0.8846 - acc: 0.5053 - f1_measure: 0.1043\nINFO    2018-11-18 12:21:57 -0600   master-replica-0          2\/219 [..............................] - ETA: 3:51:32 - loss: 0.8767 - acc: 0.5018 - f1_measure: 0.1013\nINFO    2018-11-18 12:22:59 -0600   master-replica-0          3\/219 [..............................] - ETA: 3:46:49 - loss: 0.8634 - acc: 0.5039 - f1_measure: 0.1010\nINFO    2018-11-18 12:23:58 -0600   master-replica-0          4\/219 [..............................] - ETA: 3:44:59 - loss: 0.8525 - acc: 0.5045 - f1_measure: 0.0991\nINFO    2018-11-18 12:24:48 -0600   master-replica-0          5\/219 [..............................] - ETA: 3:41:17 - loss: 0.8434 - acc: 0.5031 - f1_measure: 0.0992Sun Nov 18 18:24:48 2018       \nINFO    2018-11-18 12:24:48 -0600   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        | NVIDIA-SMI 396.26                 Driver Version: 396.26                    |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        |-------------------------------+----------------------+----------------------+\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        | Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        |===============================+======================+======================|\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        |   0  Tesla K80           Off  | 00000000:00:04.0 Off |                    0 |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        | N\/A   32C    P0    56W \/ 149W |  10955MiB \/ 11441MiB |      0%      Default |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        +-------------------------------+----------------------+----------------------+\nINFO    2018-11-18 12:24:48 -0600   master-replica-0                                                                                       \nINFO    2018-11-18 12:24:48 -0600   master-replica-0        +-----------------------------------------------------------------------------+\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        | Processes:                                                       GPU Memory |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        |  GPU       PID   Type   Process name                             Usage      |\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        |=============================================================================|\nINFO    2018-11-18 12:24:48 -0600   master-replica-0        +-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>Basically the GPU usage stays at 0% during the training, how can this be possible?<\/p>\n","answers":[{"AnswerId":"56402660","CreationDate":"2019-05-31T23:21:27.333","ParentId":null,"OwnerUserId":"9687748","Title":null,"Body":"<p>I recommend to use the <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/machine-types#machine_type_table\" rel=\"nofollow noreferrer\"><code>standard_gpu<\/code><\/a> which has the same n1-standard-8 with one k80 GPU in the <code>cloud-gpu.yaml<\/code>:<\/p>\n\n<pre><code>trainingInput:\n  scaleTier: CUSTOM\n  # standard_gpu provides 1 GPU. Change to complex_model_m_gpu for 4 GPUs\n  masterType: standard_gpu\n  runtimeVersion: \"1.5\"\n<\/code><\/pre>\n\n<p>this:<\/p>\n\n<pre><code>with tf.device('\/gpu:0'):\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu#supported_devices\" rel=\"nofollow noreferrer\">should be<\/a>:<\/p>\n\n<pre><code>with tf.device('\/device:GPU:0'):\n<\/code><\/pre>\n\n<p>I recommend to review this <a href=\"https:\/\/github.com\/Neeraj-Natu\/keras-cloud-test\/blob\/master\/trainer\/cnn_with_keras.py\" rel=\"nofollow noreferrer\">cnn_with_keras.py<\/a> to have a better example.<\/p>\n"}]}
+{"QuestionId":53357687,"AnswerCount":1,"Tags":"<android><tensorflow><keras><tensorflow-lite>","CreationDate":"2018-11-18T03:39:53.253","AcceptedAnswerId":null,"OwnerUserId":3138464.0,"Title":"How to test Tensorflowlite model with multiple inputs?","Body":"<p>I created a simple MLP Regression Keras model with 4 inputs and one output. I converted this model to TFlite now I'm just trying to find out how to test it on android studio. How can I input multiple 4D objects to test in Java? \nThe following gives an error when trying to run the model:<\/p>\n\n<pre><code>try{\n            tflite = new Interpreter(loadModelFile());\n        }\n        catch(Exception ex){\n            ex.printStackTrace();\n        }   \n\ndouble[][] inp= new double[1][4];\n    inp[0][1]= 0;\n    inp[0][0] = 0;\n    inp[0][2]= 0;\n    inp[0][3]=-2.01616982303105;\n\n    double[] output = new double[100];\n\n    tflite.run(inp,output);\n<\/code><\/pre>\n\n<p>EDIT: \nHere is the model I originally created:<\/p>\n\n<pre><code># create model\nmodel = Sequential()\nmodel.add(Dense(50, activation=\"tanh\", input_dim=4, \nkernel_initializer=\"random_uniform\", name=\"input_tensor\"))\nmodel.add(Dense(50, activation=\"tanh\", \nkernel_initializer=\"random_uniform\"))\nmodel.add(Dense(1, activation=\"linear\", \nkernel_initializer='random_uniform', name=\"output_tensor\"))\n<\/code><\/pre>\n","answers":[{"AnswerId":"54584247","CreationDate":"2019-02-08T00:06:20.347","ParentId":null,"OwnerUserId":"447719","Title":null,"Body":"<p>If your inputs are actually 4 separate tensors, then you should use the <code>Interpreter.runForMultipleInputsAndOutputs<\/code> API which allows multiple separate inputs. See also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/java\/src\/test\/java\/org\/tensorflow\/lite\/InterpreterTest.java#L185\" rel=\"nofollow noreferrer\">this example<\/a> from the TensorFlow Lite repository. For example:<\/p>\n\n<pre><code>double[] input0 = {...};\ndouble[] input1 = {...};\nObject[] inputs = {input0, input1};\ndouble[] output = new double[100];\nMap&lt;Integer, Object&gt; outputs = new HashMap&lt;&gt;();\noutputs.put(0, output);\ninterpreter.runForMultipleInputsOutputs(inputs, outputs);\n<\/code><\/pre>\n"}]}
+{"QuestionId":53357706,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2018-11-18T03:45:53.127","AcceptedAnswerId":53363886.0,"OwnerUserId":5204470.0,"Title":"How to update weights with TensorFlow Eager Execution?","Body":"<p>So I tried TensorFlow's eager execution and my implementation of it wasn't successful. I used <code>gradient.tape<\/code>, and while the program runs, there is no visible update in any of the weights. I've seen some sample algorithms and tutorials using <code>optimizer.apply_gradients()<\/code> in order to update all variables, but I'm assuming I'm not using it properly.<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.contrib.eager as tfe\n\n# emabling eager execution\ntf.enable_eager_execution()\n\n# establishing hyperparameters\nLEARNING_RATE = 20\nTRAINING_ITERATIONS = 3\n\n# establishing all LABLES\nLABELS = tf.constant(tf.random_normal([3, 1]))\n# print(LABELS)\n\n# stub statment for input\ninit = tf.Variable(tf.random_normal([3, 1]))\n\n# declare and intialize all weights\nweight1 = tfe.Variable(tf.random_normal([2, 3]))\nbias1 = tfe.Variable(tf.random_normal([2, 1]))\nweight2 = tfe.Variable(tf.random_normal([3, 2]))\nbias2 = tfe.Variable(tf.random_normal([3, 1]))\nweight3 = tfe.Variable(tf.random_normal([2, 3]))\nbias3 = tfe.Variable(tf.random_normal([2, 1]))\nweight4 = tfe.Variable(tf.random_normal([3, 2]))\nbias4 = tfe.Variable(tf.random_normal([3, 1]))\nweight5 = tfe.Variable(tf.random_normal([3, 3]))\nbias5 = tfe.Variable(tf.random_normal([3, 1]))\n\nVARIABLES = [weight1, bias1, weight2, bias2, weight3, bias3, weight4, bias4, weight5, bias5]\n\n\ndef thanouseEyes(input):  # nn model aka: Thanouse's Eyes\n    layerResult = tf.nn.relu(tf.matmul(weight1, input) + bias1)\n    input = layerResult\n    layerResult = tf.nn.relu(tf.matmul(weight2, input) + bias2)\n    input = layerResult\n    layerResult = tf.nn.relu(tf.matmul(weight3, input) + bias3)\n    input = layerResult\n    layerResult = tf.nn.relu(tf.matmul(weight4, input) + bias4)\n    input = layerResult\n    layerResult = tf.nn.softmax(tf.matmul(weight5, input) + bias5)\n    return layerResult\n\n\n# Begin training and update variables\noptimizer = tf.train.AdamOptimizer(LEARNING_RATE)\n\nwith tf.GradientTape(persistent=True) as tape:  # gradient calculation\n    for i in range(TRAINING_ITERATIONS):\n        COST = tf.reduce_sum(LABELS - thanouseEyes(init))\n        GRADIENTS = tape.gradient(COST, VARIABLES)\n        optimizer.apply_gradients(zip(GRADIENTS, VARIABLES))\n        print(weight1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53363886","CreationDate":"2018-11-18T17:57:28.193","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>The usage of <code>optimizer<\/code> seems fine, however the computation defined by <code>thanouseEyes()<\/code> will always return [1., 1., 1.] irrespective of the variables, thus the gradients are always 0 and thus the variables will never be updated (<code>print(thanouseEyes(init))<\/code> and <code>print(GRADIENTS)<\/code> should demonstrate that).<\/p>\n\n<p>Digging in a bit more, <code>tf.nn.softmax<\/code> is applied to <code>x = tf.matmul(weight5, input) + bias5<\/code> which has a shape of [3, 1]. So the <code>tf.nn.softmax(x)<\/code> is effectively computing <code>[softmax(x[0]), softmax(x[1]), softmax(x[2])]<\/code> as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax<\/code><\/a> applies (by default) on the last axis of the input. <code>x[0]<\/code>, <code>x[1]<\/code>, and <code>x[2]<\/code> are vectors with one element so <code>softmax(x[i])<\/code> will always be 1.0.<\/p>\n\n<p>Hope that helps.<\/p>\n\n<p>Some additional points unrelated to your question that you may be interested in:<\/p>\n\n<ul>\n<li><p>As of TensorFlow 1.11, you don't need the <code>tf.contrib.eager<\/code> module in your program. Replace all occurrences of <code>tfe<\/code> with <code>tf<\/code> (i.e., <code>tf.Variable<\/code> instead of <code>tfe.Variable<\/code>) and you'll get the same result<\/p><\/li>\n<li><p>Computation performed inside the context of a <code>GradientTape<\/code> is \"recorded\", i.e., it holds on to intermediate tensors so that gradients can be computed later on. Long story short, you'd want to move the <code>GradientTape<\/code> inside the loop body:<\/p><\/li>\n<\/ul>\n\n<p>-<\/p>\n\n<pre><code>for i in range(TRAINING_ITERATIONS):\n    with tf.GradientTape() as tape:\n        COST = tf.reduce_sum(LABELS - thanouseEyes(init))\n    GRADIENTS = tape.gradient(COST, VARIABLES)\n    optimizer.apply_gradients(zip(GRADIENTS, VARIABLES))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53357901,"AnswerCount":2,"Tags":"<r><image><keras><generator>","CreationDate":"2018-11-18T04:29:51.387","AcceptedAnswerId":53744995.0,"OwnerUserId":9399668.0,"Title":"Using a custom R generator function with fit_generator (Keras, R)","Body":"<p>I'd like to train a convolutional network to solve a multi-class, multi-label problem on image data.  Due to the nature of the data, and for reasons I'll spare you, it would be best if I could use a custom R generator function to feed to the <code>fit_generator<\/code> command, instead of its built-in <code>image_data_generator<\/code> and <code>flow_images_from_directory<\/code> commands (which I was successfully able to get working, just not for this particular problem).<\/p>\n\n<p>Here (<a href=\"https:\/\/www.rdocumentation.org\/packages\/keras\/versions\/2.2.0\/topics\/fit_generator\" rel=\"nofollow noreferrer\">https:\/\/www.rdocumentation.org\/packages\/keras\/versions\/2.2.0\/topics\/fit_generator<\/a>) it says that I can do just that, without giving any examples.  So I tried the following.  Here is an extremely stripped down example of what I'm trying to do (this code is entirely self contained):<\/p>\n\n<pre><code>library(keras)\nlibrary(reticulate)      #for py_iterator function\n\nplay.network = keras_model_sequential() %&gt;%\n  layer_dense(units = 10, activation = \"relu\", input_shape = c(10)) %&gt;%\n  layer_dense(units = 1, activation = \"relu\")\n\nplay.network %&gt;% compile(\n  optimizer = \"rmsprop\",\n  loss = \"mse\"\n)\n\nmikes.custom.generator.function = function()     #generates a 2-list of a random 1 x 10 array, and a scalar\n{\n  new.func = function()\n  {\n    arr = array(dim = c(1,10))\n    arr[,] = sample(1:10, 10, replace = TRUE)\/10\n    return(list(arr,runif(1)))\n  }\n}\n\nmikes.custom.iterator = py_iterator(mikes.custom.generator.function())          #creates a python iterator object\n\ngenerator_next(mikes.custom.iterator)                 #correctly returns a 2-member list consisting of a 1 x 10 array, and a scalar\ngenerator_next(mikes.custom.iterator)[[1]]            #a 1 x 10 array\ngenerator_next(mikes.custom.iterator)[[2]]            #a scalar\n\n#try to fit with \"fit_generator\":\n\nplay.network %&gt;% fit_generator(                       #FREEZES.\n  mikes.custom.iterator,\n  steps_per_epoch = 1,\n  epochs = 1\n)\n<\/code><\/pre>\n\n<p>The thing freezes at training time, without giving me an error message or anything.  I also tried it with a custom image data generator for my original problem, same result.<\/p>\n\n<p>Note that this network trains just fine if I just use <code>fit<\/code> and input the training data manually:<\/p>\n\n<pre><code>play.network %&gt;% fit(generator_next(mikes.custom.iterator)[[1]],generator_next(mikes.custom.iterator)[[2]], epochs = 1, batch_size = 1)      \n\n#trains just fine\n<\/code><\/pre>\n\n<p>I think I know the problem, but I don't know the solution.  If you ask it for the class of my custom iterator, it gives<\/p>\n\n<pre><code>class(mikes.custom.iterator)\n\n[1] \"python.builtin.iterator\"  \"rpytools.generator.RGenerator\"  \"python.builtin.object\" \n<\/code><\/pre>\n\n<p>whereas if I build an iterator using the builtin <code>image_data_generator<\/code> and <code>flow_images_from_directory<\/code> commands, it gives<\/p>\n\n<pre><code>train_datagen &lt;- image_data_generator(rescale = 1\/255)\n\nclass(train_datagen)\n\n[1] \"keras.preprocessing.image.ImageDataGenerator\" \"keras_preprocessing.image.ImageDataGenerator\" \"python.builtin.object\"\n\ntrain_generator &lt;- flow_images_from_directory(\n  train_dir,\n  train_datagen,\n  ....\n)\n\nclass(train_generator)\n\n[1] \"python.builtin.iterator\" \"keras_preprocessing.image.DirectoryIterator\" \"keras_preprocessing.image.Iterator\"        \"tensorflow.python.keras.utils.data_utils.Sequence\" \"python.builtin.object\"\n<\/code><\/pre>\n\n<p>So my guess is that <code>train_datagen<\/code> and\/or <code>train_generator<\/code> have attributes that <code>mikes.custom.iterator<\/code> does not, and <code>fit_generator<\/code> is trying to call upon <code>mikes.custom.iterator<\/code> using functions other than the basic <code>generator_next<\/code> (which is in theory all it should really need).  But I don't know what they may be, or how to build <code>mikes.custom.iterator<\/code> correctly, even after searching for two hours online.<\/p>\n\n<p>Help anyone? <\/p>\n","answers":[{"AnswerId":"54045991","CreationDate":"2019-01-04T21:00:29.347","ParentId":null,"OwnerUserId":"9179937","Title":null,"Body":"<p>In R, you can build an iterator using <code>&lt;&lt;-<\/code> operator. This is very helpful to build a custom generator function; and it is compatible with Keras' <code>fit_generator()<\/code> function.<\/p>\n\n<p>Some minimal example:<\/p>\n\n<pre><code># example data\ndata &lt;- data.frame(\n  x = runif(80),\n  y = runif(80),\n  z = runif(80)\n)\n\n# example generator\ndata_generator &lt;- function(data, x, y, batch_size) {\n\n  # start iterator\n  i &lt;- 1\n\n  # return an iterator function\n  function() {\n\n    # reset iterator if already seen all data\n    if ((i + batch_size - 1) &gt; nrow(data)) i &lt;&lt;- 1\n\n    # iterate current batch's rows\n    rows &lt;- c(i:min(i + batch_size - 1, nrow(data)))\n\n    # update to next iteration\n    i &lt;&lt;- i + batch_size\n\n    # create container arrays\n    x_array &lt;- array(0, dim = c(length(rows), length(x)))\n    y_array &lt;- array(0, dim = c(length(rows), length(y)))\n\n    # fill the container\n    x_array[1:length(rows), ] &lt;- data[rows, x]\n    y_array[1:length(rows), ] &lt;- data[rows, y]\n\n    # return the batch\n    list(x_array, y_array)\n\n  }\n\n}\n\n# set-up a generator\ngen &lt;- data_generator(\n  data = data.matrix(data),\n  x = 1:2, # it is flexible, you can use the column numbers,\n  y = c(\"y\", \"z\"), # or the column name\n  batch_size = 32\n)\n<\/code><\/pre>\n\n<p>From above function, you can simply check the resulting arrays by calling the generator:<\/p>\n\n<pre><code>gen()\n<\/code><\/pre>\n\n<p>Or you could also test the generator using a simple Keras model:<\/p>\n\n<pre><code># import keras\nlibrary(keras)\n\n# set up a simple keras model\nmodel &lt;- keras_model_sequential() %&gt;% \n  layer_dense(32, input_shape = c(2)) %&gt;% \n  layer_dense(2)\n\nmodel %&gt;% compile(\n  optimizer = \"rmsprop\",\n  loss = \"mse\"\n)\n\n# fit using generator\nmodel %&gt;% fit_generator(\n  generator = gen,\n  steps_per_epoch = 100, # will auto-reset after see all sample\n  epochs = 10\n)\n<\/code><\/pre>\n\n<p>I have to admit that the process is a little bit complex and requires extensive programming. You should check this featured blog post by <a href=\"https:\/\/blogs.rstudio.com\/tensorflow\/posts\/2017-12-20-time-series-forecasting-with-recurrent-neural-networks\/\" rel=\"nofollow noreferrer\">Fran\u00e7ois Chollet<\/a> himself, or <a href=\"https:\/\/kerasgenerator.netlify.com\" rel=\"nofollow noreferrer\"><code>kerasgenerator<\/code><\/a> package that I develop personally.<\/p>\n"},{"AnswerId":"53744995","CreationDate":"2018-12-12T14:16:08.863","ParentId":null,"OwnerUserId":"2466267","Title":null,"Body":"<pre><code>sampling_generator &lt;- function(X_data, Y_data, batch_size) {\n  function() {\n    rows &lt;- sample(1:nrow(X_data), batch_size, replace = TRUE)\n    list(X_data[rows,], Y_data[rows,])\n  }\n}\n\nmodel %&gt;% \n  fit_generator(sampling_generator(X_train, Y_train, batch_size = 128), \n            steps_per_epoch = nrow(X_train) \/ 128, epochs = 10)\n<\/code><\/pre>\n\n<p>I found this answer in R keras FAQs which seems to work<\/p>\n\n<p><a href=\"https:\/\/keras.rstudio.com\/articles\/faq.html#how-can-i-use-keras-with-datasets-that-dont-fit-in-memory\" rel=\"nofollow noreferrer\">https:\/\/keras.rstudio.com\/articles\/faq.html#how-can-i-use-keras-with-datasets-that-dont-fit-in-memory<\/a><\/p>\n"}]}
+{"QuestionId":53358033,"AnswerCount":1,"Tags":"<keras><deep-learning>","CreationDate":"2018-11-18T04:56:59.463","AcceptedAnswerId":null,"OwnerUserId":9160974.0,"Title":"why ResNet50 include_top argument giving error?","Body":"<p>Keras version - 2.2.4<\/p>\n\n<p>Python - 3.6.6<\/p>\n\n<p>issue:<\/p>\n\n<pre><code>from resnet50 import ResNet50\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZpH8y.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZpH8y.png\" alt=\"enter image description here\"><\/a>\n----> 6 model = ResNet50(input_tensor=image_input, \ninclude_top=True,weights='imagenet')<\/p>\n\n<p>TypeError: _obtain_input_shape() got an unexpected keyword argument 'include_top'<\/p>\n\n<p>is there a fix for this?<\/p>\n","answers":[{"AnswerId":"58097635","CreationDate":"2019-09-25T11:46:14.207","ParentId":null,"OwnerUserId":"10959693","Title":null,"Body":"<p><strong>Replace:<\/strong> <\/p>\n\n<pre><code>model = ResNet50(input_tensor=image_input, include_top=True,weights='imagenet')\n<\/code><\/pre>\n\n<p><strong>With:<\/strong><\/p>\n\n<pre><code>model = ResNet50(input_tensor=image_input,require_flatten=True,weights='imagenet')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53358250,"AnswerCount":1,"Tags":"<pytorch><mount><google-colaboratory><drive>","CreationDate":"2018-11-18T05:46:03.290","AcceptedAnswerId":53364363.0,"OwnerUserId":3131604.0,"Title":"Google-Colaboratory - How to Refresh google-drive?","Body":"<p>I am using google-colaboratory GPU to train NN models.\nMy python\/pytorch code is stored in google-drive.\nI am able to mount my drive in colaboratory  and train models.\nBut any python code changes in the \"my drive\"  is not updated to google-colaboratory even after rebooting my PC and start all again.<\/p>\n\n<p>To clear the google-colaboratory cache I tried :<\/p>\n\n<pre><code>!google-drive-ocamlfuse -cc\n<\/code><\/pre>\n\n<p>But it does not work:<\/p>\n\n<pre><code>\/bin\/bash: google-drive-ocamlfuse: command not found\n<\/code><\/pre>\n\n<p>How to clean this cache and avoid waiting hours before my code being taken into account by google-colaboratory ? Thanks in advance<\/p>\n\n<p>PS : the method I used to mount:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/drive\/')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53364363","CreationDate":"2018-11-18T18:51:19.593","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p><code>google-drive-ocamlfuse<\/code> is irrelevant to mounts using <code>google.colab.drive.mount<\/code> as described in the PS, so not surprising the -cc invocation is not helping you.\nI suspect what's happening is you have <code>.py<\/code> files stored in Google Drive, which you're <code>import<\/code>ing in your notebook, and you want to see changes to the <code>.py<\/code> files reflected in your runtime, but they're not because python's <code>import<\/code> system is idempotent (an <code>import<\/code> statement is ignored if python thinks it's already loaded a module by that name, even if the underlying file has changed). \nYou can force a reload using something like <a href=\"https:\/\/stackoverflow.com\/a\/437591\/8755609\">https:\/\/stackoverflow.com\/a\/437591\/8755609<\/a> e.g.:<\/p>\n\n<pre><code>from importlib import reload  # Py3 only; unneeded in py2.\nfoo = reload(foo)\n<\/code><\/pre>\n\n<p>(obvs replace <code>foo<\/code> with your module's name).<\/p>\n"}]}
+{"QuestionId":53358518,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><rnn><seq2seq>","CreationDate":"2018-11-18T06:42:15.077","AcceptedAnswerId":null,"OwnerUserId":10669491.0,"Title":"Decoder targets required for RNN inference","Body":"<p>I have been trying to run some experiments using the deepfix tool (<a href=\"https:\/\/bitbucket.org\/iiscseal\/deepfix\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/iiscseal\/deepfix<\/a>) which is a <code>seq2seq<\/code> model for correcting common programming errors.\nI made changes to the code so that it is compatible to <code>TF-1.12<\/code>, as the original code contains <code>tensorflow.contrib.seq2seq<\/code> functions which are not supported in version <code>TF-1.12<\/code> (only in <code>TF-1.0.x<\/code>).<\/p>\n\n<p>The main changes were in the <code>seq2seq_model<\/code> defined in <code>neural_net\/train.py<\/code>.\nBelow is the changed code. I'm new to the <code>tensorflow RNN<\/code>, and coded the decoder part using help from online codes.<\/p>\n\n<pre><code>class seq2seq_model():\n\nPAD = 0\nEOS = 1\n\ndef __init__(self, vocab_size, embedding_size, max_output_seq_len,\n             cell_type='LSTM', memory_dim=300, num_layers=4, dropout=0.2,\n             attention=True,\n             scope=None,\n             verbose=False):\n\n    assert 0 &lt;= dropout and dropout &lt;= 1, '0 &lt;= dropout &lt;= 1, you passed dropout={}'.format(\n        dropout)\n\n    tf.set_random_seed(1189)\n\n    self.attention = attention\n    self.max_output_seq_len = max_output_seq_len\n\n    self.memory_dim = memory_dim\n    self.num_layers = num_layers\n    self.dropout = dropout\n    self.scope = scope\n\n    if dropout != 0:\n        self.keep_prob = tf.placeholder(tf.float32)\n    else:\n        self.keep_prob = None\n\n    self.vocab_size = vocab_size\n    self.embedding_size = embedding_size\n\n    self.encoder_cell = _new_RNN_cell(\n        memory_dim, num_layers, cell_type, dropout, self.keep_prob)\n    self.decoder_cell = _new_RNN_cell(\n        memory_dim, num_layers, cell_type, dropout, self.keep_prob)\n\n    self._make_graph()\n\n    if self.scope is not None:\n        saver_vars = [var for var in tf.global_variables(\n        ) if var.name.startswith(self.scope)]\n    else:\n        saver_vars = tf.global_variables()\n\n    if verbose:\n        print 'root-scope:', self.scope\n        print \"\\n\\nDiscovered %d saver variables.\" % len(saver_vars)\n        for each in saver_vars:\n            print each.name\n\n    self.saver = tf.train.Saver(saver_vars, max_to_keep=5)\n\n@property\ndef decoder_hidden_units(self):\n    return self.memory_dim\n\ndef _make_graph(self):\n    self._init_placeholders()\n\n    self._init_decoder_train_connectors()\n\n    self._init_embeddings()\n\n    self._init_simple_encoder()\n\n    self._init_decoder()\n\n    self._init_optimizer()\n\ndef _init_placeholders(self):\n    \"\"\" Everything is time-major \"\"\"\n    self.encoder_inputs = tf.placeholder(\n        shape=(None, None),\n        dtype=tf.int32,\n        name='encoder_inputs',\n    )\n    self.encoder_inputs_length = tf.placeholder(\n        shape=(None,),\n        dtype=tf.int32,\n        name='encoder_inputs_length',\n    )\n\n    self.decoder_targets = tf.placeholder(\n        shape=(None, None),\n        dtype=tf.int32,\n        name='decoder_targets'\n    )\n    self.decoder_targets_length = tf.placeholder(\n        shape=(None,),\n        dtype=tf.int32,\n        name='decoder_targets_length',\n    )\n\ndef _init_decoder_train_connectors(self):\n\n    with tf.name_scope('decoderTrainFeeds'):\n        sequence_size, batch_size = tf.unstack(\n            tf.shape(self.decoder_targets), name='decoder_targets_shape')\n\n        EOS_SLICE = tf.ones([1, batch_size], dtype=tf.int32) * self.EOS\n        PAD_SLICE = tf.ones([1, batch_size], dtype=tf.int32) * self.PAD\n\n        self.decoder_train_inputs = tf.concat(\n            [EOS_SLICE, self.decoder_targets], axis=0, name=\"decoder_train_inputs\")\n        self.decoder_train_length = self.decoder_targets_length + 1\n\n        decoder_train_targets = tf.concat(\n            [self.decoder_targets, PAD_SLICE], axis=0)\n        decoder_train_targets_seq_len, _ = tf.unstack(\n            tf.shape(decoder_train_targets))\n        decoder_train_targets_eos_mask = tf.one_hot(self.decoder_train_length - 1,\n                                                    decoder_train_targets_seq_len,\n                                                    on_value=self.EOS, off_value=self.PAD,\n                                                    dtype=tf.int32)\n        decoder_train_targets_eos_mask = tf.transpose(\n            decoder_train_targets_eos_mask, [1, 0])\n\n        decoder_train_targets = tf.add(decoder_train_targets,\n                                       decoder_train_targets_eos_mask, name=\"decoder_train_targets\")\n\n        self.decoder_train_targets = decoder_train_targets\n\n        self.loss_weights = tf.ones([\n            batch_size,\n            tf.reduce_max(self.decoder_train_length)\n        ], dtype=tf.float32, name=\"loss_weights\")\n\ndef _init_embeddings(self):\n    with tf.variable_scope(\"embedding\") as scope:\n        sqrt3 = math.sqrt(3)\n        initializer = tf.random_uniform_initializer(-sqrt3, sqrt3)\n\n        self.embedding_matrix = tf.get_variable(\n            name=\"embedding_matrix\",\n            shape=[self.vocab_size, self.embedding_size],\n            initializer=initializer,\n            dtype=tf.float32)\n\n        self.encoder_inputs_embedded = tf.nn.embedding_lookup(\n            self.embedding_matrix, self.encoder_inputs,\n            name=\"encoder_inputs_embedded\")\n\n        self.decoder_train_inputs_embedded = tf.nn.embedding_lookup(\n            self.embedding_matrix, self.decoder_train_inputs,\n            name=\"decoder_train_inputs_embedded\")\n\ndef _init_simple_encoder(self):\n    with tf.variable_scope(\"Encoder\") as scope:\n        (self.encoder_outputs, self.encoder_state) = (\n            tf.nn.dynamic_rnn(cell=self.encoder_cell,\n                              inputs=self.encoder_inputs_embedded,\n                              sequence_length=self.encoder_inputs_length,\n                              time_major=True,\n                              dtype=tf.float32)\n        )\n\ndef _init_decoder(self):\n    with tf.variable_scope(\"decoder\") as scope:\n        # def output_fn(outputs):\n        #     return tf.contrib.layers.fully_connected(outputs, self.vocab_size, scope=scope,\n        #                                                 name = \"output_fn\")\n\n        sequence_size, batch_size = tf.unstack(\n            tf.shape(self.decoder_targets), name='decoder_targets_shape')\n\n        train_helper = seq2seq.TrainingHelper(\n                inputs=self.decoder_train_inputs_embedded,\n                sequence_length=self.decoder_train_length,\n                time_major=True,\n                name=\"train_helper\")\n\n\n        pred_helper = seq2seq.SampleEmbeddingHelper(\n                embedding=self.embedding_matrix,\n                start_tokens=tf.ones([batch_size], dtype=tf.int32) * self.EOS,\n                end_token=self.EOS)\n                # name=\"pred_helper\")\n\n        def _decode(helper, scope, reuse=None):\n            with tf.variable_scope(scope, reuse=reuse):\n                attention_states = tf.transpose(\n                    self.encoder_outputs, [1, 0, 2])\n\n                attention_mechanism = seq2seq.BahdanauAttention(\n                num_units=self.decoder_hidden_units, memory=attention_states,\n                name=\"attention_mechanism\")\n\n                attention_cell = seq2seq.AttentionWrapper(\n                self.decoder_cell, attention_mechanism,\n                name=\"atttention_wrapper\")\n\n                out_cell = tf.contrib.rnn.OutputProjectionWrapper(\n                    attention_cell, self.vocab_size, reuse=reuse)\n                    # name=\"output_cell\")\n\n                decoder = seq2seq.BasicDecoder(\n                    cell=out_cell, helper=helper,\n                    initial_state=out_cell.zero_state(\n                        dtype=tf.float32, batch_size=batch_size))\n                        # name=\"decoder\")\n\n                outputs = seq2seq.dynamic_decode(\n                    decoder=decoder, output_time_major=True,\n                    impute_finished=True)\n                    # name=\"outputs\")\n\n                return outputs\n\n\n\n        (self.decoder_logits_train, self.decoder_state_train, _) = _decode(train_helper, \"decoder\")\n        (self.decoder_logits_inference, self.decoder_state_inference, _) = _decode(pred_helper, \"decoder\", reuse=True)\n\n        self.decoder_logits_train = self.decoder_logits_train.rnn_output\n        self.decoder_logits_inference = self.decoder_logits_inference.rnn_output\n        # self.decoder_logits_train = output_fn(self.decoder_outputs_train)\n\n        self.decoder_prediction_train = tf.argmax(\n            self.decoder_logits_train, axis=-1, name='decoder_prediction_train')\n\n        scope.reuse_variables()\n\n        self.decoder_prediction_inference = tf.argmax(self.decoder_logits_inference, axis=-1,\n                                                      name='decoder_prediction_inference')\n\n\ndef _init_optimizer(self):\n    logits = tf.transpose(self.decoder_logits_train, [1, 0, 2])\n    targets = tf.transpose(self.decoder_train_targets, [1, 0])\n    self.loss = seq2seq.sequence_loss(logits=logits, targets=targets,\n                                      weights=self.loss_weights)\n\n    self.optimizer = tf.train.AdamOptimizer()\n    gvs = self.optimizer.compute_gradients(self.loss)\n\n    def ClipIfNotNone(grad):\n        if grad is None:\n            return grad\n        return tf.clip_by_value(grad, -1., 1)\n\n    # capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs]\n    capped_gvs = [(ClipIfNotNone(grad), var) for grad, var in gvs]\n\n    self.train_op = self.optimizer.apply_gradients(capped_gvs)\n\ndef make_feed_dict(self, x, x_len, y, y_len):\n    feed_dict = {\n        self.encoder_inputs: x,\n        self.encoder_inputs_length: x_len,\n\n        self.decoder_targets: y,\n        self.decoder_targets_length: y_len,\n    }\n\n    if self.dropout != 0:\n        feed_dict.update({self.keep_prob: 1.0 - self.dropout})\n\n    return feed_dict\n\ndef load_parameters(self, sess, filename):\n    self.saver.restore(sess, filename)\n\ndef save_parameters(self, sess, filename, global_step=None):\n    self.saver.save(sess, filename, global_step=global_step)\n\ndef train_step(self, session, x, x_len, y, y_len):\n    feed_dict = self.make_feed_dict(x, x_len, y, y_len)\n    _, loss = session.run([self.train_op, self.loss], feed_dict)\n    return loss\n\ndef validate_step(self, session, x, x_len, y, y_len):\n    feed_dict = self.make_feed_dict(x, x_len, y, y_len)\n    loss, decoder_prediction, decoder_train_targets = session.run([self.loss,\n                                                                   self.decoder_prediction_inference,\n                                                                   self.decoder_train_targets], feed_dict)\n    return loss, np.array(decoder_prediction).T, np.array(decoder_train_targets).T\n\ndef sample(self, session, X, X_len):\n    feed_dict = {self.encoder_inputs: X,\n                 self.encoder_inputs_length: X_len}\n\n    if self.dropout != 0:\n        feed_dict.update({self.keep_prob: 1.0})\n\n    decoder_prediction = session.run(\n        self.decoder_prediction_inference, feed_dict)\n    return np.array(decoder_prediction).T\n<\/code><\/pre>\n\n<p>I am having some problems with this code:<\/p>\n\n<ol>\n<li>Main problem - The <code>seq2seq.train_step()<\/code> and <code>seq2seq.validate_step()<\/code> functions are working, but when I use <code>seq2seq.sample()<\/code> for actually making inferences, I get an error that asks me to feed a value for <code>decoder_targets<\/code>. This is an unexpected behaviour as the <code>SampleEmbeddingHelper<\/code> function is used for inference which does not require decoder_targets. The error:<\/li>\n<\/ol>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): You must feed a value\n  for placeholder tensor 'ids\/decoder_targets' with dtype int32 and\n  shape [?,?]    [[node ids\/decoder_targets (defined at\n  ...\/code\/neural_net\/train.py:241)  = Placeholderdtype=DT_INT32,\n  shape=[?,?],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<ol start=\"2\">\n<li>When I try to use the <code>GreedyEmbeddingHelper<\/code> instead of <code>SampleEmbeddingHelper<\/code>, and then run <code>decoder_logits_inference<\/code> op, the machine hangs and runs out of memory after some time. Although <code>SampleEmbeddingHelper<\/code> works fine.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"53359623","CreationDate":"2018-11-18T09:52:39.303","ParentId":null,"OwnerUserId":"10356594","Title":null,"Body":"<ol>\n<li><p>Well, <code>SampleEmbeddingHelper<\/code> does need decoder targets, since it mixes part of <code>GreedyEmbeddingHelper<\/code>(infer mode) and <code>tf.contrib.seq2seq.TrainingHelper<\/code>(teacher forcing). I think you just need to use <code>GreedyEmbeddingHelper<\/code>.<\/p><\/li>\n<li><p>Since in the beginning, the parameters are totally random (if not pre-trained).\nMaybe you have seen that the results of the first few loops of seq2seq model are totally messed up.\nSo if you use <code>GreedyEmbeddingHelper<\/code>, which outputs a result based on the previous one, and of course no one teaches it \"where to stop\", so it usually goes infinitely until your memory runs out. To solve this, you need to set an upper limit for the length of sentence in <code>tf.contrib.seq2seq.dynamic_decode<\/code>.\nThe argument is <code>maximum_iterations<\/code>. as shown in\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/seq2seq\/dynamic_decode\" rel=\"nofollow noreferrer\">tf.contrib.seq2seq.dynamic_decode<\/a><\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53358752,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-18T07:26:59.330","AcceptedAnswerId":null,"OwnerUserId":7975433.0,"Title":"For a tensorflow SavedModel in pbtxt format, where is the device placement of an operation\/node defined?","Body":"<p>I have a <code>SavedModel<\/code> with <code>saved_model.pbtxt<\/code> and <code>variables\\<\/code> which was pre-trained on a single GPU from this repo: <a href=\"https:\/\/github.com\/sthalles\/deeplab_v3\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sthalles\/deeplab_v3<\/a>. I'm trying to serve this <code>SavedModel<\/code> with tensorflow-serving, and it can only utilise GPU:0 in a multi-GPU machine. I learned from <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/311\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/issues\/311<\/a> that tensorflow-serving loads the graph with tensorflow, and this model was trained on a single GPU. I tried to save the model with <code>clear_devices=True<\/code> flag but no help, still ran on GPU:0.<\/p>\n\n<p>Then I try to read the <code>GraphDef<\/code> in <code>saved_model.pbtxt<\/code>, from <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/model_files#device\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/extend\/model_files#device<\/a> I know that the device assigned to one node\/operation is defined in <code>NodeDef<\/code>.<\/p>\n\n<p>My problem is, in this <code>saved_model.pbtxt<\/code>, only CPU was assigned for some operations\/nodes in <code>NodeDef<\/code> as <code>device: \"\/device:CPU:0\"<\/code>, while no GPU was specifically assigned. All those operations executed on GPU don't have a <code>device<\/code> tag in their <code>NodeDef<\/code>.<\/p>\n\n<p>I wonder where are the device placement infomation for GPU operations was saved in <code>SavedModel<\/code> and can I change the device info in a graph? Thanks for your help.<\/p>\n\n<p>For example, in this <code>saved_model.pbtxt<\/code> a CPU op was defined as:<\/p>\n\n<pre><code>node {\n     name: \"save\/RestoreV2\/tensor_names\"\n     op: \"Const\"\n     device: \"\/device:CPU:0\"\n     ...\n     }\n<\/code><\/pre>\n\n<p>A computation op was:<\/p>\n\n<pre><code>node {\n  name: \"resnet_v2_50\/block1\/unit_1\/bottleneck_v2\/conv2\/kernel\/Regularizer\/l2_regularizer\"\n  op: \"Mul\"\n  input: \"resnet_v2_50\/block1\/unit_1\/bottleneck_v2\/conv2\/kernel\/Regularizer\/l2_regularizer\/scale\"\n  input: \"resnet_v2_50\/block1\/unit_1\/bottleneck_v2\/conv2\/kernel\/Regularizer\/l2_regularizer\/L2Loss\"\n  attr {\n    key: \"T\"\n    value {\n      type: DT_FLOAT\n    }\n  }\n  attr {\n    key: \"_class\"\n    value {\n      list {\n        s: \"loc:@resnet_v2_50\/block1\/unit_1\/bottleneck_v2\/conv2\/weights\"\n      }\n    }\n  }\n  attr {\n    key: \"_output_shapes\"\n    value {\n      list {\n        shape {\n        }\n      }\n    }\n  }\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53358838,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-18T07:42:37.687","AcceptedAnswerId":53360472.0,"OwnerUserId":10110401.0,"Title":"Predicting all zeros","Body":"<p>I've depeloped a neural network for classification and I'm getting a 0.93 of accuracy, the problem is that I'm predicting all zeros because the distribution of the data.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3bjHs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3bjHs.png\" alt=\"Data distribution\"><\/a><\/p>\n\n<p>How can I fix it? Should I change from neural network to another algorithm?<\/p>\n\n<p>Thanks in advance<\/p>\n\n<p><strong>Edit<\/strong>: i've just checked and my model is predicting the same probability for each row.<\/p>\n\n<p>The model is a NN with 5 layers, and <code>tf.nn.relu6<\/code> as activation function. The cost function is <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/p>\n\n<p>To predict the values I use:<\/p>\n\n<pre><code>predicted = tf.nn.sigmoid(Z5)\ncorrect_pred = tf.equal(tf.round(predicted), Y)\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n<\/code><\/pre>\n\n<p><strong>EDIT 2<\/strong><\/p>\n\n<p>I have 'fixed' the imbalance class problem (undersampling and upsampling 0s and 1s) but the net is still predicting the same values for each row:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CRxSc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CRxSc.png\" alt=\"Prediction\"><\/a><\/p>\n\n<p>I have tested to change activation function to tanh or sigmoid but then outputs NaN's<\/p>\n","answers":[{"AnswerId":"53360472","CreationDate":"2018-11-18T11:42:25.637","ParentId":null,"OwnerUserId":"8947739","Title":null,"Body":"<p>There are multiple solutions for unbalanced data. But first, the accuracy is not a good metric for unbalanced data, because if you only had 5 positives and 95 negatives, you accuracy will be 95% of predicting negatives. You should check <strong>sensitivity and specificity<\/strong>, or other metrics that work good with unbalanced data like the <strong>LIFT score<\/strong>.<\/p>\n\n<p>To train the model with unbalanced data, there are multiple solutions. One of them is the U<strong>p-sample Minority Class<\/strong>. <\/p>\n\n<blockquote>\n  <p>Up-sampling is the process of randomly duplicating observations from\n  the minority class in order to reinforce its signal.<\/p>\n<\/blockquote>\n\n<p>You can upsample data with a code like this:<\/p>\n\n<pre><code>from sklearn.utils import resample\n# Separate majority and minority classes\ndf_majority = df[df.balance==0]\ndf_minority = df[df.balance==1]\n\n# Upsample minority class\ndf_minority_upsampled = resample(df_minority, \n                                 replace=True,     # sample with replacement\n                                 n_samples=576,    # to match majority class\n                                 random_state=123) # reproducible results\n\n# Combine majority class with upsampled minority class\ndf_upsampled = pd.concat([df_majority, df_minority_upsampled])\n\n# Display new class counts\ndf_upsampled.balance.value_counts()\n# 1    576\n# 0    576\n# Name: balance, dtype: int64\n<\/code><\/pre>\n\n<p>You can find more information and other solutions that are well explained <a href=\"https:\/\/elitedatascience.com\/imbalanced-classes\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53359056,"AnswerCount":0,"Tags":"<python><neural-network><pytorch>","CreationDate":"2018-11-18T08:21:59.723","AcceptedAnswerId":null,"OwnerUserId":7399312.0,"Title":"Pytorch error Illegal instruction (core dumped)","Body":"<p>I have install pytorch version <code>0.4.1<\/code><\/p>\n\n<p>I installed it directly with pip without conda, I've also noted the issue is with the binary and from my research on processor incompatibility with C gcc version.\nMy version of gcc is <code>7.3.0<\/code><\/p>\n\n<p>And my processor type <code>AMD A8-7410 APU with AMD Radeon R5 Graphics<\/code><\/p>\n\n<p>gcc location.<\/p>\n\n<pre><code>(data-science) sam@sam-Lenovo-G51-35:~\/code\/data science projects\/pytorch$ which gcc\n\/usr\/bin\/gcc\n<\/code><\/pre>\n\n<p>This is how I get my error.<\/p>\n\n<p>The following code runs...<\/p>\n\n<pre><code>from torchvision import datasets, transforms\nfrom torch import nn\nimport torch\nimport torch.nn.functional as F\n\n# Define a transform to normalize the data\ntransform = transforms.Compose([transforms.ToTensor(),\n                              transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),\n                              ])\n\n# Download and load the training data\ntrainset = datasets.MNIST('MNIST_data\/', download=True, train=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)\n\ndataiter = iter(trainloader)\nimages, labels = dataiter.next()\n\nclass Network(nn.Module):\n    def __init__(self):\n        super(Network, self).__init__()\n        # Defining the layers, 128, 64, 10 units each\n        self.fc1 = nn.Linear(784, 128)\n        self.fc2 = nn.Linear(128, 64)\n        # Output layer, 10 units - one for each digit\n        self.fc3 = nn.Linear(64, 10)\n\n    def forward(self, x):\n        ''' Forward pass through the network, returns the output logits '''\n\n        x = self.fc1(x)\n        x = F.relu(x)\n        x = self.fc2(x)\n        x = F.relu(x)\n        x = self.fc3(x)\n        x = F.softmax(x, dim=1)\n\n        return x\n\n# Create the network and look at it's text representation\nmodel = Network()\n\n# print(model.fc1.weight)\n# print(model.fc1.bias)\n\n# Set biases to all zeros\nmodel.fc1.bias.data.fill_(0)\n\n# sample from random normal with standard dev = 0.01\nmodel.fc1.weight.data.normal_(std=0.01)\n\n# Grab some data \ndataiter = iter(trainloader)\nimages, labels = dataiter.next()\n\n# Resize images into a 1D vector, new shape is (batch size, color channels, image pixels) \nimages.resize_(64, 1, 784)\n# or images.resize_(images.shape[0], 1, 784) to automatically get batch size\n<\/code><\/pre>\n\n<p>all of this successfully.<\/p>\n\n<p>But when I run this other line<\/p>\n\n<pre><code># Forward pass through the network\nimg_idx = 0\nps = model.forward(images[img_idx,:])\n<\/code><\/pre>\n\n<p>I get this error<\/p>\n\n<pre><code>Illegal instruction (core dumped)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53359127,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><lstm>","CreationDate":"2018-11-18T08:34:29.557","AcceptedAnswerId":53361746.0,"OwnerUserId":3612693.0,"Title":"Keras-- low accuracy with LSTM layer but the accuracy is good without LSTM","Body":"<p>I am training a model in Keras with IMDB dataset. For this model with LSTM layer, the accuracy is about 50%:<\/p>\n\n<pre><code> model = Sequential()\n model.add(Embedding(max_features, 32))\n model.add(LSTM(32, return_sequences=True))\n model.add(LSTM(32, return_sequences=True))\n model.add(LSTM(32))\n model.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>Accuracy: <\/p>\n\n<pre><code>loss: 0.6933 - acc: 0.5007 - val_loss: 0.6932 - val_acc: 0.4947\n<\/code><\/pre>\n\n<p>I have also tried with a single LSTM layer but it also gives similar accuracy.<\/p>\n\n<p>However, if I don't use LSTM layer the accuracy reaches to around 82% <\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(layers.Dense(16, kernel_regularizer=regularizers.l1(0.001), activation='relu', input_shape=(10000,)))\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Dense(16, kernel_regularizer=regularizers.l1(0.001), activation='relu'))\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>Accuracy:<\/p>\n\n<pre><code> loss: 0.6738 - acc: 0.8214 - val_loss: 0.6250 - val_acc: 0.8320\n<\/code><\/pre>\n\n<p>This is how I compile and fit the model:<\/p>\n\n<pre><code>model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['acc'])\nmodel.fit(partial_x_train, partial_y_train, epochs=Numepochs, batch_size=Batchsize, validation_data=(x_val, y_val))\n<\/code><\/pre>\n\n<p>How can this be explained? I thought LSTM works great for sequential text data?<\/p>\n","answers":[{"AnswerId":"53361746","CreationDate":"2018-11-18T14:04:06.803","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Don't forget that LSTM is used for processing sequences such as timeseries or text data. In a sequence the order of elements is very important and if you reorder the element then the whole meaning of that sequence might completely change. <\/p>\n\n<p>Now the problem in your case is that the preprocessing step you have used is not the proper one for a LSTM model. You are encoding each sentence as a vector where each of its elements represents the presence or absence of particular word. Therefore, you are completely ignoring the order of appearance of words in a sentence, which LSTM layer is good at modeling it. There is also another issue in your LSTM model, considering the preprocessing scheme you have used, which is the fact that Embedding layer accepts word indices as input and not a vector of zero and ones (i.e. the output of the preprocessing stage).<\/p>\n\n<p>Since the IMDB data is already stored as sequences of word indices, to resolve this issue you just need to preprocess the IMDB data by only padding\/truncating the sequences with a specified length to be able to utilize batch processing. For example:<\/p>\n\n<pre><code>from keras.preprocessing.sequences import pad_sequences\n\nvocab_size = 10000 # only consider the 10000 most frequent words\n(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=vocab_size)\n\nx_train = pad_sequences(x_train, maxlen=500)  # truncate or pad sequences to make them all have a length of 500\n<\/code><\/pre>\n\n<p>Now, <code>x_train<\/code> would have a shape of <code>(25000, 500)<\/code> and it consists of 25000 sequences of length 500, encoded as integer word indices. Now you can use it for training by passing it to <code>fit<\/code> method. I guess you can reach at least 80% training accuracy with an Embedding layer and a single LSTM layer. Don't forget that to use a validation scheme to monitor overfitting (one simple option is to set <code>validation_split<\/code> argument when calling <code>fit<\/code> method). <\/p>\n"}]}
+{"QuestionId":53359785,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-18T10:18:18.897","AcceptedAnswerId":null,"OwnerUserId":10110401.0,"Title":"Neural Network predicting the same probability for each row","Body":"<p>I've depeloped a 5 layers neural network for classification and it always predict the same probability for each row which ends predicting the same class.<\/p>\n\n<p>I'm using Relu as activation fuction (if I use sigmoid or tanh it outputs NaN's) and <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/p>\n\n<p>The model will be this:<\/p>\n\n<pre><code> X, Y = create_placeholders(n_x, n_y)\n\nparameters = initialize_parameters()\n\nZ5 = forward_propagation(X, parameters)\n\ncost = cost_function(Z5, Y)\n\noptimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)\n<\/code><\/pre>\n\n<p>And the prediction:<\/p>\n\n<pre><code>predicted = tf.nn.sigmoid(Z5)\ncorrect_pred = tf.equal(tf.round(predicted), Y)\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n<\/code><\/pre>\n\n<p>The cost value drops until 0.24, in the first 100 epochs, but the values predicted don't change.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pZKnD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pZKnD.png\" alt=\"Prediction\"><\/a><\/p>\n\n<p>At the beginning I thought it was cause of an imbalance class problem but I fixed it (upsampling and wonsampling 1s and 0s) so it should not be the problem.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[]}
+{"QuestionId":53359829,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-11-18T10:22:41.467","AcceptedAnswerId":null,"OwnerUserId":10485421.0,"Title":"Some questions about Pytocrh DA-RNN","Body":"<p>I have two different kind of data,and I try to Processing the two different data separately,not all put into the model.\nI use DA-RNN and have one Encoder&amp;Decoder.I'm trying add new 'encoder_num_hidden' in Decoder to deal with the data,and merge the data from Decoders ,but I always get wrong.\nHow can I do?Or I just did it wrong?\nThis is code,it come from <a href=\"https:\/\/github.com\/Zhenye-Na\/DA-RNN\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<pre><code>class Encoder(nn.Module):\n\"\"\"encoder in DA_RNN.\"\"\"\n\ndef __init__(self, T,\n             input_size,\n             encoder_num_hidden,\n             parallel=False):\n    \"\"\"Initialize an encoder in DA_RNN.\"\"\"\n    super(Encoder, self).__init__()\n    self.encoder_num_hidden = encoder_num_hidden\n    self.input_size = input_size\n    self.parallel = parallel\n    self.T = T\n\n    # Fig 1. Temporal Attention Mechanism: Encoder is LSTM\n    self.encoder_lstm = nn.LSTM(\n        input_size=self.input_size, hidden_size=self.encoder_num_hidden)\n\n    # Construct Input Attention Mechanism via deterministic attention model\n    # Eq. 8: W_e[h_{t-1}; s_{t-1}] + U_e * x^k\n    self.encoder_attn = nn.Linear(\n        in_features=2 * self.encoder_num_hidden + self.T - 1, out_features=1, bias=True)\n\ndef forward(self, X):\n    \"\"\"forward.\n    Args:\n        X\n    \"\"\"\n    X_tilde = Variable(X.data.new(\n        X.size(0), self.T - 1, self.input_size).zero_())\n    X_encoded = Variable(X.data.new(\n        X.size(0), self.T - 1, self.encoder_num_hidden).zero_())\n\n    # Eq. 8, parameters not in nn.Linear but to be learnt\n    # v_e = torch.nn.Parameter(data=torch.empty(\n    #     self.input_size, self.T).uniform_(0, 1), requires_grad=True)\n    # U_e = torch.nn.Parameter(data=torch.empty(\n    #     self.T, self.T).uniform_(0, 1), requires_grad=True)\n\n    # hidden, cell: initial states with dimention hidden_size\n    h_n = self._init_states(X)\n    s_n = self._init_states(X)\n\n    for t in range(self.T - 1):\n        # batch_size * input_size * (2*hidden_size + T - 1)\n        x = torch.cat((h_n.repeat(self.input_size, 1, 1).permute(1, 0, 2),\n                       s_n.repeat(self.input_size, 1, 1).permute(1, 0, 2),\n                       X.permute(0, 2, 1)), dim=2)\n\n        x = self.encoder_attn(\n            x.view(-1, self.encoder_num_hidden * 2 + self.T - 1))\n\n        # get weights by softmax\n        alpha = F.softmax(x.view(-1, self.input_size))\n\n        # get new input for LSTM\n        x_tilde = torch.mul(alpha, X[:, t, :])\n\n        # encoder LSTM\n        self.encoder_lstm.flatten_parameters()\n        _, final_state = self.encoder_lstm(\n            x_tilde.unsqueeze(0), (h_n, s_n))\n        h_n = final_state[0]\n        s_n = final_state[1]\n\n        X_tilde[:, t, :] = x_tilde\n        X_encoded[:, t, :] = h_n\n\n    return X_tilde, X_encoded\n\ndef _init_states(self, X):\n    \"\"\"Initialize all 0 hidden states and cell states for encoder.\n    Args:\n        X\n    Returns:\n        initial_hidden_states\n    \"\"\"\n    # hidden state and cell state [num_layers*num_directions, batch_size, hidden_size]\n    # https:\/\/pytorch.org\/docs\/master\/nn.html?#lstm\n    initial_states = Variable(X.data.new(\n        1, X.size(0), self.encoder_num_hidden).zero_())\n    return initial_states\n<\/code><\/pre>\n\n<p>class Decoder(nn.Module):\n    \"\"\"decoder in DA_RNN.\"\"\"<\/p>\n\n<pre><code>def __init__(self, T, decoder_num_hidden, encoder_num_hidden):\n    \"\"\"Initialize a decoder in DA_RNN.\"\"\"\n    super(Decoder, self).__init__()\n    self.decoder_num_hidden = decoder_num_hidden\n    self.encoder_num_hidden = encoder_num_hidden\n    self.T = T\n\n    self.attn_layer = nn.Sequential(nn.Linear(2 * decoder_num_hidden + encoder_num_hidden, encoder_num_hidden),\n                                    nn.Tanh(),\n                                    nn.Linear(encoder_num_hidden, 1))\n    self.lstm_layer = nn.LSTM(\n        input_size=1, hidden_size=decoder_num_hidden)\n    self.fc = nn.Linear(encoder_num_hidden + 1, 1)\n    self.fc_final = nn.Linear(decoder_num_hidden + encoder_num_hidden, 1)\n\n    self.fc.weight.data.normal_()\n\ndef forward(self, X_encoed, y_prev):\n    \"\"\"forward.\"\"\"\n    d_n = self._init_states(X_encoed)\n    c_n = self._init_states(X_encoed)\n\n    for t in range(self.T - 1):\n\n        x = torch.cat((d_n.repeat(self.T - 1, 1, 1).permute(1, 0, 2),\n                       c_n.repeat(self.T - 1, 1, 1).permute(1, 0, 2),\n                       X_encoed), dim=2)\n\n        beta = F.softmax(self.attn_layer(\n            x.view(-1, 2 * self.decoder_num_hidden + self.encoder_num_hidden)).view(-1, self.T - 1))\n        # Eqn. 14: compute context vector\n        # batch_size * encoder_hidden_size\n        context = torch.bmm(beta.unsqueeze(1), X_encoed)[:, 0, :]\n        if t &lt; self.T - 1:\n            # Eqn. 15\n            # batch_size * 1\n            y_tilde = self.fc(\n                torch.cat((context, y_prev[:, t].unsqueeze(1)), dim=1))\n            # Eqn. 16: LSTM\n            self.lstm_layer.flatten_parameters()\n            _, final_states = self.lstm_layer(\n                y_tilde.unsqueeze(0), (d_n, c_n))\n            # 1 * batch_size * decoder_num_hidden\n            d_n = final_states[0]\n            # 1 * batch_size * decoder_num_hidden\n            c_n = final_states[1]\n    # Eqn. 22: final output\n    y_pred = self.fc_final(torch.cat((d_n[0], context), dim=1))\n\n    return y_pred\n\ndef _init_states(self, X):\n    \"\"\"Initialize all 0 hidden states and cell states for encoder.\n    Args:\n        X\n    Returns:\n        initial_hidden_states\n    \"\"\"\n    # hidden state and cell state [num_layers*num_directions, batch_size, hidden_size]\n    # https:\/\/pytorch.org\/docs\/master\/nn.html?#lstm\n    initial_states = Variable(X.data.new(\n        1, X.size(0), self.decoder_num_hidden).zero_())\n    return initial_states\n<\/code><\/pre>\n\n<p>class DA_rnn(nn.Module):\n    \"\"\"da_rnn.\"\"\"<\/p>\n\n<pre><code>def __init__(self, X, y, T,\n             encoder_num_hidden,\n             decoder_num_hidden,\n             batch_size,\n             learning_rate,\n             epochs,\n             parallel=False):\n    \"\"\"da_rnn initialization.\"\"\"\n    super(DA_rnn, self).__init__()\n    self.encoder_num_hidden = encoder_num_hidden\n    self.decoder_num_hidden = decoder_num_hidden\n    self.learning_rate = learning_rate\n    self.batch_size = batch_size\n    self.parallel = parallel\n    self.shuffle = False\n    self.epochs = epochs\n    self.T = T\n    self.X = X\n    self.y = y\n\n    self.Encoder = Encoder(input_size=X.shape[1],\n                           encoder_num_hidden=encoder_num_hidden,\n                           T=T)\n    self.Decoder = Decoder(encoder_num_hidden=encoder_num_hidden,\n                           decoder_num_hidden=decoder_num_hidden,\n                           T=T)\n\n    # Loss function\n    self.criterion = nn.MSELoss()\n\n    if self.parallel:\n        self.encoder = nn.DataParallel(self.encoder)\n        self.decoder = nn.DataParallel(self.decoder)\n\n    self.encoder_optimizer = optim.Adam(params=filter(lambda p: p.requires_grad,\n                                                      self.Encoder.parameters()),\n                                        lr=self.learning_rate)\n    self.decoder_optimizer = optim.Adam(params=filter(lambda p: p.requires_grad,\n                                                      self.Decoder.parameters()),\n                                        lr=self.learning_rate)\n\n    # Training set\n    self.train_timesteps = int(self.X.shape[0] * 0.7)\n    self.input_size = self.X.shape[1]\n\ndef train(self):\n    \"\"\"training process.\"\"\"\n    iter_per_epoch = int(np.ceil(self.train_timesteps * 1. \/ self.batch_size))\n    self.iter_losses = np.zeros(self.epochs * iter_per_epoch)\n    self.epoch_losses = np.zeros(self.epochs)\n\n    n_iter = 0\n\n    for epoch in range(self.epochs):\n        if self.shuffle:\n            ref_idx = np.random.permutation(self.train_timesteps - self.T)\n        else:\n            ref_idx = np.array(range(self.train_timesteps - self.T))\n\n        idx = 0\n\n        while (idx &lt; self.train_timesteps):\n            # get the indices of X_train\n            indices = ref_idx[idx:(idx + self.batch_size)]\n            # x = np.zeros((self.T - 1, len(indices), self.input_size))\n            x = np.zeros((len(indices), self.T - 1, self.input_size))\n            y_prev = np.zeros((len(indices), self.T - 1))\n            y_gt = self.y[indices + self.T]\n\n            # format x into 3D tensor\n            for bs in range(len(indices)):\n                x[bs, :, :] = self.X[indices[bs]:(indices[bs] + self.T - 1), :]\n                y_prev[bs, :] = self.y[indices[bs]:(indices[bs] + self.T - 1)]\n\n            loss = self.train_forward(x, y_prev, y_gt)\n            self.iter_losses[epoch * iter_per_epoch + idx \/ self.batch_size] = loss\n\n            idx += self.batch_size\n            n_iter += 1\n\n            if n_iter % 50000 == 0 and n_iter != 0:\n                for param_group in self.encoder_optimizer.param_groups:\n                    param_group['lr'] = param_group['lr'] * 0.9\n                for param_group in self.decoder_optimizer.param_groups:\n                    param_group['lr'] = param_group['lr'] * 0.9\n\n            self.epoch_losses[epoch] = np.mean(self.iter_losses[range(epoch * iter_per_epoch, (epoch + 1) * iter_per_epoch)])\n\n        if epoch % 10 == 0:\n            print \"Epochs: \", epoch, \" Iterations: \", n_iter, \" Loss: \", self.epoch_losses[epoch]\n\n        if epoch == self.epochs - 1:\n            y_train_pred = self.test(on_train=True)\n            y_test_pred = self.test(on_train=False)\n            y_pred = np.concatenate((y_train_pred, y_test_pred))\n            plt.ioff()\n            plt.figure()\n            plt.plot(range(1, 1 + len(self.y)),\n                     self.y, label=\"True\")\n            plt.plot(range(self.T, len(y_train_pred) + self.T),\n                     y_train_pred, label='Predicted - Train')\n            plt.plot(range(self.T + len(y_train_pred), len(self.y) + 1),\n                     y_test_pred, label='Predicted - Test')\n            plt.legend(loc='upper left')\n            plt.show()\n\n\n        # Save files in last iterations\n        # if epoch == self.epochs - 1:\n        #     np.savetxt('..\/loss.txt', np.array(self.epoch_losses), delimiter=',')\n        #     np.savetxt('..\/y_pred.txt',\n        #                np.array(self.y_pred), delimiter=',')\n        #     np.savetxt('..\/y_true.txt',\n        #                np.array(self.y_true), delimiter=',')\n\n\ndef train_forward(self, X, y_prev, y_gt):\n    # zero gradients\n    self.encoder_optimizer.zero_grad()\n    self.decoder_optimizer.zero_grad()\n\n    input_weighted, input_encoded = self.Encoder(\n        Variable(torch.from_numpy(X).type(torch.FloatTensor)))\n    y_pred = self.Decoder(input_encoded, Variable(\n        torch.from_numpy(y_prev).type(torch.FloatTensor)))\n\n    y_true = Variable(torch.from_numpy(\n        y_gt).type(torch.FloatTensor))\n\n    y_true = y_true.view(-1, 1)\n    loss = self.criterion(y_pred, y_true)\n    loss.backward()\n\n    self.encoder_optimizer.step()\n    self.decoder_optimizer.step()\n\n    return loss.item()\n\ndef val(self):\n    \"\"\"validation.\"\"\"\n    pass\n\ndef test(self, on_train=False):\n    \"\"\"test.\"\"\"\n\n    if on_train:\n        y_pred = np.zeros(self.train_timesteps - self.T + 1)\n    else:\n        y_pred = np.zeros(self.X.shape[0] - self.train_timesteps)\n\n    i = 0\n    while i &lt; len(y_pred):\n        batch_idx = np.array(range(len(y_pred)))[i : (i + self.batch_size)]\n        X = np.zeros((len(batch_idx), self.T - 1, self.X.shape[1]))\n        y_history = np.zeros((len(batch_idx), self.T - 1))\n\n        for j in range(len(batch_idx)):\n            if on_train:\n                X[j, :, :] = self.X[range(batch_idx[j], batch_idx[j] + self.T - 1), :]\n                y_history[j, :] = self.y[range(batch_idx[j],  batch_idx[j]+ self.T - 1)]\n            else:\n                X[j, :, :] = self.X[range(batch_idx[j] + self.train_timesteps - self.T, batch_idx[j] + self.train_timesteps - 1), :]\n                y_history[j, :] = self.y[range(batch_idx[j] + self.train_timesteps - self.T,  batch_idx[j]+ self.train_timesteps - 1)]\n\n        y_history = Variable(torch.from_numpy(y_history).type(torch.FloatTensor))\n        _, input_encoded = self.Encoder(Variable(torch.from_numpy(X).type(torch.FloatTensor)))\n        y_pred[i:(i + self.batch_size)] = self.Decoder(input_encoded, y_history).cpu().data.numpy()[:, 0]\n        i += self.batch_size\n\n    return y_pred\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53360012,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-18T10:47:41.183","AcceptedAnswerId":null,"OwnerUserId":26625.0,"Title":"How do I resolve an InvalidArgumentError in Classifier model?","Body":"<p>New to TensorFlow, so apologies for newbie question.<\/p>\n\n<p>Following this <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough\" rel=\"nofollow noreferrer\">tutorial<\/a> but instead of using image data I am using numerical data.<\/p>\n\n<p><strong>Load the dataset:<\/strong><\/p>\n\n<pre><code>train_dataset_url = \"xxx.csv\"\ntrain_dataset_fp = tf.keras.utils.get_file(\n  fname=os.path.basename(train_dataset_url),\n  origin=train_dataset_url)\n<\/code><\/pre>\n\n<p><strong>Make training dataset:<\/strong><\/p>\n\n<pre><code>batch_size = 32\n\ntrain_dataset = tf.contrib.data.make_csv_dataset(\n    train_dataset_fp,\n    batch_size, \n    column_names=column_names,\n    label_name=label_name,\n    num_epochs=1)\n<\/code><\/pre>\n\n<p><strong>Train classified model using:<\/strong><\/p>\n\n<p><code>model = tf.keras.Sequential([\n  tf.keras.layers.Dense(10, activation=tf.nn.relu, input_shape=(1,)),\n  tf.keras.layers.Dense(10, activation=tf.nn.relu),\n  tf.keras.layers.Dense(4)\n])<\/code><\/p>\n\n<p>But when I \"test\" the model with the same inputs:<\/p>\n\n<p><code>predictions = model(features)<\/code><\/p>\n\n<p>I receive the error:<\/p>\n\n<p><code>InvalidArgumentError: cannot compute MatMul as input #0(zero-based) was expected to be a float tensor but is a int32 tensor [Op:MatMul]<\/code><\/p>\n\n<p>It's possible I have missed something fundamental. I feel like I need to specify a type somewhere.<\/p>\n","answers":[{"AnswerId":"53360345","CreationDate":"2018-11-18T11:28:55.923","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>The data which you feed in the model is a  numpy array according to my assumption . The error states that the model requires a tensor with dtype=float32 or float64. You are providing a int32 numpy array. So, wherever you create a numpy array, just mention the dtype as float32.<\/p>\n"}]}
+{"QuestionId":53360596,"AnswerCount":1,"Tags":"<python><numpy><pytorch>","CreationDate":"2018-11-18T11:57:06.727","AcceptedAnswerId":53361357.0,"OwnerUserId":10670192.0,"Title":"Pytorch Index a tensor of [n*n*3] by two tensors of size[n,n]","Body":"<p>In numpy, I can index in the following:<\/p>\n\n<pre><code>a = np.random.randn(2,2,3)\nb = np.eye(2,2).astype(np.uint8)\nc = np.eye(2,2).astype(np.uint8)\nprint(a)\nprint(\"diff\")\nprint(a[b,c,:])\n<\/code><\/pre>\n\n<p>, in which a[b, c, :] is a tensor of 2*2. <\/p>\n\n<pre><code>[[[-1.01338087  0.70149058  0.55268617]\n  [ 2.56941124  1.12720312 -0.07219555]]\n\n [[-0.04084548  0.17018995  2.14229567]\n  [-0.68017558 -0.91788125  1.1719151 ]]]\ndiff\n[[[-0.68017558 -0.91788125  1.1719151 ]\n  [-1.01338087  0.70149058  0.55268617]]\n\n [[-1.01338087  0.70149058  0.55268617]\n  [-0.68017558 -0.91788125  1.1719151 ]]]\n<\/code><\/pre>\n\n<p>But in Pytorch, I could not do the index in the same way like <code>a[b,c,:]<\/code>. Who knows how to do this. Thanks~\n<img src=\"https:\/\/i.stack.imgur.com\/oUG4G.png\" alt=\"enter image description here\"><\/p>\n","answers":[{"AnswerId":"53361357","CreationDate":"2018-11-18T13:24:53.360","ParentId":null,"OwnerUserId":"5373738","Title":null,"Body":"<p>Indexing in PyTorch is almost similar to numpy.<\/p>\n\n<pre><code>a = torch.randn(2, 2, 3)\nb = torch.eye(2, 2, dtype=torch.long)\nc = torch.eye(2, 2, dtype=torch.long)\n\nprint(a)\nprint(a[b, c, :])\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>tensor([[[ 1.2471,  1.6571, -2.0504],\n         [-1.7502,  0.5747, -0.3451]],\n\n        [[-0.4389,  0.4482,  0.7294],\n         [-1.3051,  0.6606, -0.6960]]])\ntensor([[[-1.3051,  0.6606, -0.6960],\n         [ 1.2471,  1.6571, -2.0504]],\n\n        [[ 1.2471,  1.6571, -2.0504],\n         [-1.3051,  0.6606, -0.6960]]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53360781,"AnswerCount":1,"Tags":"<java><image><tensorflow>","CreationDate":"2018-11-18T12:19:23.590","AcceptedAnswerId":53360865.0,"OwnerUserId":10095320.0,"Title":"Tensorflow and image processing in Java","Body":"<p>Are there some examples how to load freeze graph trained in python to Java and use it for image recognition (image as an input, example of converting Image into Tensor). I am looking for a way to use existing model in Java application.<\/p>\n","answers":[{"AnswerId":"53360865","CreationDate":"2018-11-18T12:27:19.650","ParentId":null,"OwnerUserId":"1617910","Title":null,"Body":"<p>You can check <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/main\/java\/org\/tensorflow\/examples\/LabelImage.java\" rel=\"nofollow noreferrer\">LabelImage.java<\/a> example which uses Inception5h model to classify images.<\/p>\n\n<p>Inception5h is the pre-trained model which you can download from <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Note : To run this java code in your IDE, you need to download 2 libraries<\/p>\n\n<ol>\n<li>libtensorflow-1.10.1.jar<\/li>\n<li>libtensorflow_jni-1.10.0.jar<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53360866,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-18T12:27:23.433","AcceptedAnswerId":null,"OwnerUserId":1493609.0,"Title":"How to build TensorFlow C++ library on Windows XP 32-bit","Body":"<p>I'm attempting to build TensorFlow's C++ library for Windows XP. While I've been able to build and use it on Windows 10, 32-bit XP isn't working. The background: I'm working on a COM module that calls fuctions from tensorflow.dll. My build environment:<\/p>\n\n<ul>\n<li>Visual Studio 2017 15.7<\/li>\n<li>CMake 3.11.1<\/li>\n<li>TensorFlow 1.8<\/li>\n<li>Windows 10<\/li>\n<\/ul>\n\n<p>The sequence I use to build tensorflow.dll is:<\/p>\n\n<ul>\n<li>Open \"x64_x86 Cross Tools Command Prompt for VS 2017\"<\/li>\n<li>Try to force the use of functions availablbe in Win XP: <code>set CXXFLAGS=\/D_WINVER=0x0501 \/D_WIN32_WINNT=0x0501<\/code><\/li>\n<li>Add Git to path: <code>set PATH=%PATH%;C:\\Program Files (x86)\\Git\\bin<\/code><\/li>\n<li>Fix CMake file for converting *.proto files to *.pb.h files <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18931\" rel=\"nofollow noreferrer\">as described here<\/a>.<\/li>\n<li>Configure CMake: <code>cmake .. -A Win32 -T v141_xp,host=x64 -DCMAKE_SYSTEM_VERSION=7.0 -DCMAKE_BUILD_TYPE=Release -DPYTHON_EXECUTABLE=C:\\Users\\williams\\AppData\\Local\\Continuum\\Anaconda3\\envs\\tensorflow\\python.exe -Dtensorflow_BUILD_SHARED_LIB=ON -Dtensorflow_BUILD_PYTHON_BINDINGS=OFF -Dtensorflow_WIN_CPU_SIMD_OPTIONS=\"\/arch:IA32\"<\/code><\/li>\n<li>Build: <code>cmake --build . --target tensorflow --config Release -- \/fileLogger \/m:1 \/p:CL_MPCount=1<\/code><\/li>\n<\/ul>\n\n<p>The last step also involves some manual labour as the build process doesn't copy .lib files from the 3rd part dependencies to where they are needed. For whatever reason, a bunch of INSTALL projects never get run so I had to do that manually each time the build would fail while looking for a missing lib file. Once that was done, the build completed successfully.<\/p>\n\n<p>Next I copy my COM module (a DLL) and the TensorFlow DLL over to a Windows XP virtual machine for testing and try to register the COM module, but get an error <code>LoadLibrary(\"MyDLL.dll\") - The specified procedure could not be found<\/code>. I don't know what procedure it is looking for, so the best I can offer is that Dependency Walker highlights WS2_32.DLL and tells me it can't find <code>inet_ntop<\/code> and <code>inet_pton<\/code>.<\/p>\n\n<p>Any suggestions on how to build TensorFlow so that it doesn't use these two functions?<\/p>\n\n<p>P.S. suggestions of \"Stop using XP, its old and no longer supported\" don't help here. Upgrading to Windows 10 is an absolute last resort because of the disruption it would cause at the facility where this software will be tested.<\/p>\n\n<p><strong>Edit 1:<\/strong><\/p>\n\n<p>These two functions <code>inet_pton<\/code> and <code>inet_ntop<\/code> were only used in one file that forms part of Google Cloud Storage support in TensorFlow. The build process generated a tensorflow_static.lib in addition to tensorflow.dll. Linking against the static version and adding a few dependencies that aren't included in tensorflow_static.lib got rid of the code using <code>inet_*<\/code> functions.<\/p>\n\n<p>My COM module still isn't working on Windows XP though because the file <code>tensorflow\\core\\platform\\windows\\env.cc<\/code> uses functions like <code>CloseThreadpoolWork<\/code>, submit <code>SubmitThreadpoolWork<\/code>, etc. that were only introduced in Windows Vista. It looks like I'll have to replace them with something else, as I don't see an alternative implementation in TensorFlow.<\/p>\n\n<p>Additionally, I found that <code>tensorflow\\contrib\\cmake\\CMakeLists.txt<\/code> forces <code>_WIN32_WINNT=0x0A00<\/code> and that <code>CXXFLAGS<\/code> is the wrong environment variable to use. Changing it to <code>CMAKE_CXX_FLAGS<\/code> at least gets my macro definitions included, FWIW.<\/p>\n","answers":[{"AnswerId":"53871033","CreationDate":"2018-12-20T14:54:20.953","ParentId":null,"OwnerUserId":"1493609","Title":null,"Body":"<p>In the end I gave up on this as simply being impossible. Even replacing the thread pool functions with something from Boost didn't help. If someone else manages to get this working, I'll gladly accept that as the answer, but so far this looks impossible.<\/p>\n"},{"AnswerId":"55275351","CreationDate":"2019-03-21T07:10:07.773","ParentId":null,"OwnerUserId":"11235959","Title":null,"Body":"<p>It is almost impossible to port tensorflow to windows xp, because:<\/p>\n\n<ol>\n<li><p>TF's platform depedent code requires some Windows APIs later than winxp such as Thread Pool API. This would possibly bypassed by using third party thread pool libs.<\/p><\/li>\n<li><p>The nsync, protobuf and eigen, which are core parts of TF,  use C++11 thread_local, which makes them unable to run-time load as dll, see <a href=\"https:\/\/docs.microsoft.com\/en-us\/cpp\/parallel\/thread-local-storage-tls?view=vs-2017\" rel=\"nofollow noreferrer\">https:\/\/docs.microsoft.com\/en-us\/cpp\/parallel\/thread-local-storage-tls?view=vs-2017<\/a> for details. This later feature can theoretically be replaced by old windows TLS API, which requires many modifications on TF's core framework.<\/p><\/li>\n<\/ol>\n\n<p>Anyway, if you really need the xp support, good luck with that.<\/p>\n"}]}
+{"QuestionId":53361134,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-18T13:00:16.387","AcceptedAnswerId":53361559.0,"OwnerUserId":7429461.0,"Title":"generate tensor like this in tensorflow","Body":"<p>Supposed I have a vector<\/p>\n\n<pre><code>[3, 2, 4]\n<\/code><\/pre>\n\n<p>I want to generate a corresponding tensor<\/p>\n\n<pre><code>[[1,1,1,0], [1,1,0,0],[1,1,1,1]]\n<\/code><\/pre>\n\n<p>Would you please tell us how to do this? Thanks<\/p>\n","answers":[{"AnswerId":"53361559","CreationDate":"2018-11-18T13:45:18.513","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>This should work:<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\n\n# Input tensor\nindices = tf.placeholder_with_default([3, 2, 4], shape=(None,))\n\nmax_index = tf.reduce_max(indices)\n\ndef to_ones(index):\n    ones = tf.ones((index, ), dtype=indices.dtype)\n    zeros = tf.zeros((max_index - index, ), dtype=indices.dtype)\n    return tf.concat([ones, zeros], axis=0)\n\n# Output tensor\nresult = tf.map_fn(to_ones, indices)\n<\/code><\/pre>\n\n<p><code>print(result)<\/code>:<\/p>\n\n<pre><code>Tensor(\"map\/TensorArrayStack\/TensorArrayGatherV3:0\", shape=(?, ?), dtype=int32)\n<\/code><\/pre>\n\n<p><code>print(sess.run(result))<\/code>:<\/p>\n\n<pre><code>[[1 1 1 0]\n [1 1 0 0]\n [1 1 1 1]]\n<\/code><\/pre>\n\n<p><code>print(sess.run(result, feed_dict={indices: [1, 3, 3, 7]}))<\/code>:<\/p>\n\n<pre><code>[[1 0 0 0 0 0 0]\n [1 1 1 0 0 0 0]\n [1 1 1 0 0 0 0]\n [1 1 1 1 1 1 1]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53361187,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-11-18T13:06:33.193","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"How to reduce size of existing network?","Body":"<p>I have model based on MobileNet for regression task:<\/p>\n\n<pre><code>def MobileNet_v1():\n    # Keras 2.1.6\n    mobilenet = MobileNet(input_shape=(config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS),\n                          alpha=1.0,\n                          depth_multiplier=1,\n                          include_top=False,\n                          weights='imagenet'\n                          )\n\n    x = Flatten()(mobilenet.output)\n    x = Dropout(0.5)(x)\n    x = Dense(config.N_LANDMARKS * 2, activation='linear')(x)\n\n    # -------------------------------------------------------\n\n    model = Model(inputs=mobilenet.input, outputs=x)\n    optimizer = Adadelta()\n    model.compile(optimizer=optimizer, loss=mae_loss)\n\n    model.summary()\n    import sys\n    sys.exit()\n\n    return model\n<\/code><\/pre>\n\n<p>Network structure:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 128, 128, 3)       0\n_________________________________________________________________\nconv1_pad (ZeroPadding2D)    (None, 130, 130, 3)       0\n_________________________________________________________________\nconv1 (Conv2D)               (None, 64, 64, 32)        864\n_________________________________________________________________\nconv1_bn (BatchNormalization (None, 64, 64, 32)        128\n_________________________________________________________________\nconv1_relu (Activation)      (None, 64, 64, 32)        0\n_________________________________________________________________\nconv_pad_1 (ZeroPadding2D)   (None, 66, 66, 32)        0\n_________________________________________________________________\nconv_dw_1 (DepthwiseConv2D)  (None, 64, 64, 32)        288\n_________________________________________________________________\nconv_dw_1_bn (BatchNormaliza (None, 64, 64, 32)        128\n_________________________________________________________________\nconv_dw_1_relu (Activation)  (None, 64, 64, 32)        0\n_________________________________________________________________\nconv_pw_1 (Conv2D)           (None, 64, 64, 64)        2048\n_________________________________________________________________\nconv_pw_1_bn (BatchNormaliza (None, 64, 64, 64)        256\n_________________________________________________________________\nconv_pw_1_relu (Activation)  (None, 64, 64, 64)        0\n_________________________________________________________________\nconv_pad_2 (ZeroPadding2D)   (None, 66, 66, 64)        0\n_________________________________________________________________\nconv_dw_2 (DepthwiseConv2D)  (None, 32, 32, 64)        576\n_________________________________________________________________\nconv_dw_2_bn (BatchNormaliza (None, 32, 32, 64)        256\n_________________________________________________________________\nconv_dw_2_relu (Activation)  (None, 32, 32, 64)        0\n_________________________________________________________________\nconv_pw_2 (Conv2D)           (None, 32, 32, 128)       8192\n_________________________________________________________________\nconv_pw_2_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_pw_2_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pad_3 (ZeroPadding2D)   (None, 34, 34, 128)       0\n_________________________________________________________________\nconv_dw_3 (DepthwiseConv2D)  (None, 32, 32, 128)       1152\n_________________________________________________________________\nconv_dw_3_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_dw_3_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pw_3 (Conv2D)           (None, 32, 32, 128)       16384\n_________________________________________________________________\nconv_pw_3_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_pw_3_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pad_4 (ZeroPadding2D)   (None, 34, 34, 128)       0\n_________________________________________________________________\nconv_dw_4 (DepthwiseConv2D)  (None, 16, 16, 128)       1152\n_________________________________________________________________\nconv_dw_4_bn (BatchNormaliza (None, 16, 16, 128)       512\n_________________________________________________________________\nconv_dw_4_relu (Activation)  (None, 16, 16, 128)       0\n_________________________________________________________________\nconv_pw_4 (Conv2D)           (None, 16, 16, 256)       32768\n_________________________________________________________________\nconv_pw_4_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_pw_4_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pad_5 (ZeroPadding2D)   (None, 18, 18, 256)       0\n_________________________________________________________________\nconv_dw_5 (DepthwiseConv2D)  (None, 16, 16, 256)       2304\n_________________________________________________________________\nconv_dw_5_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_dw_5_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pw_5 (Conv2D)           (None, 16, 16, 256)       65536\n_________________________________________________________________\nconv_pw_5_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_pw_5_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pad_6 (ZeroPadding2D)   (None, 18, 18, 256)       0\n_________________________________________________________________\nconv_dw_6 (DepthwiseConv2D)  (None, 8, 8, 256)         2304\n_________________________________________________________________\nconv_dw_6_bn (BatchNormaliza (None, 8, 8, 256)         1024\n_________________________________________________________________\nconv_dw_6_relu (Activation)  (None, 8, 8, 256)         0\n_________________________________________________________________\nconv_pw_6 (Conv2D)           (None, 8, 8, 512)         131072\n_________________________________________________________________\nconv_pw_6_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_6_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_7 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_7 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_7_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_7_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_7 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_7_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_7_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_8 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_8 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_8_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_8_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_8 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_8_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_8_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_9 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_9 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_9_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_9_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_9 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_9_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_9_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_10 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_10 (DepthwiseConv2D) (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_10_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_10_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_10 (Conv2D)          (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_10_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_10_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_11 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_11 (DepthwiseConv2D) (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_11_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_11_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_11 (Conv2D)          (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_11_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_11_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_12 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_12 (DepthwiseConv2D) (None, 4, 4, 512)         4608\n_________________________________________________________________\nconv_dw_12_bn (BatchNormaliz (None, 4, 4, 512)         2048\n_________________________________________________________________\nconv_dw_12_relu (Activation) (None, 4, 4, 512)         0\n_________________________________________________________________\nconv_pw_12 (Conv2D)          (None, 4, 4, 1024)        524288\n_________________________________________________________________\nconv_pw_12_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_pw_12_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nconv_pad_13 (ZeroPadding2D)  (None, 6, 6, 1024)        0\n_________________________________________________________________\nconv_dw_13 (DepthwiseConv2D) (None, 4, 4, 1024)        9216\n_________________________________________________________________\nconv_dw_13_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_dw_13_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nconv_pw_13 (Conv2D)          (None, 4, 4, 1024)        1048576\n_________________________________________________________________\nconv_pw_13_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_pw_13_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nflatten_1 (Flatten)          (None, 16384)             0\n_________________________________________________________________\ndropout_1 (Dropout)          (None, 16384)             0\n_________________________________________________________________\ndense_1 (Dense)              (None, 156)               2556060\n=================================================================\nTotal params: 5,784,924\nTrainable params: 5,763,036\nNon-trainable params: 21,888\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>As we can see, about a half of network parameters is in last dense layer. So my question is if I have already trained network how to reduce model size? I have tested global average pooling instead of dense layer and for my regression task it performs poorly, so it's not an option, so I'm looking forward something like reducing dense layer size or sparsifying dense layer.<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>Example of network with global average pooling:<\/p>\n\n<pre><code>def MobileNet_v2():\n    # MobileNet with GAP layer on top\n\n    # Keras 2.1.6\n    mobilenet = MobileNet(input_shape=(config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS),\n                          alpha=1.0,\n                          depth_multiplier=1,\n                          include_top=False,\n                          weights='imagenet'\n                          )\n\n    x = Conv2D(filters=config.N_LANDMARKS * 2, kernel_size=(1,1), activation='linear')(mobilenet.output)\n    x = GlobalAveragePooling2D()(x)\n\n    # -------------------------------------------------------\n\n    model = Model(inputs=mobilenet.input, outputs=x)\n    optimizer = Adadelta()\n    model.compile(optimizer=optimizer, loss=mae_loss)\n\n    model.summary()\n    import sys\n    sys.exit()\n\n    return model\n\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 128, 128, 3)       0\n_________________________________________________________________\nconv1_pad (ZeroPadding2D)    (None, 130, 130, 3)       0\n_________________________________________________________________\nconv1 (Conv2D)               (None, 64, 64, 32)        864\n_________________________________________________________________\nconv1_bn (BatchNormalization (None, 64, 64, 32)        128\n_________________________________________________________________\nconv1_relu (Activation)      (None, 64, 64, 32)        0\n_________________________________________________________________\nconv_pad_1 (ZeroPadding2D)   (None, 66, 66, 32)        0\n_________________________________________________________________\nconv_dw_1 (DepthwiseConv2D)  (None, 64, 64, 32)        288\n_________________________________________________________________\nconv_dw_1_bn (BatchNormaliza (None, 64, 64, 32)        128\n_________________________________________________________________\nconv_dw_1_relu (Activation)  (None, 64, 64, 32)        0\n_________________________________________________________________\nconv_pw_1 (Conv2D)           (None, 64, 64, 64)        2048\n_________________________________________________________________\nconv_pw_1_bn (BatchNormaliza (None, 64, 64, 64)        256\n_________________________________________________________________\nconv_pw_1_relu (Activation)  (None, 64, 64, 64)        0\n_________________________________________________________________\nconv_pad_2 (ZeroPadding2D)   (None, 66, 66, 64)        0\n_________________________________________________________________\nconv_dw_2 (DepthwiseConv2D)  (None, 32, 32, 64)        576\n_________________________________________________________________\nconv_dw_2_bn (BatchNormaliza (None, 32, 32, 64)        256\n_________________________________________________________________\nconv_dw_2_relu (Activation)  (None, 32, 32, 64)        0\n_________________________________________________________________\nconv_pw_2 (Conv2D)           (None, 32, 32, 128)       8192\n_________________________________________________________________\nconv_pw_2_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_pw_2_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pad_3 (ZeroPadding2D)   (None, 34, 34, 128)       0\n_________________________________________________________________\nconv_dw_3 (DepthwiseConv2D)  (None, 32, 32, 128)       1152\n_________________________________________________________________\nconv_dw_3_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_dw_3_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pw_3 (Conv2D)           (None, 32, 32, 128)       16384\n_________________________________________________________________\nconv_pw_3_bn (BatchNormaliza (None, 32, 32, 128)       512\n_________________________________________________________________\nconv_pw_3_relu (Activation)  (None, 32, 32, 128)       0\n_________________________________________________________________\nconv_pad_4 (ZeroPadding2D)   (None, 34, 34, 128)       0\n_________________________________________________________________\nconv_dw_4 (DepthwiseConv2D)  (None, 16, 16, 128)       1152\n_________________________________________________________________\nconv_dw_4_bn (BatchNormaliza (None, 16, 16, 128)       512\n_________________________________________________________________\nconv_dw_4_relu (Activation)  (None, 16, 16, 128)       0\n_________________________________________________________________\nconv_pw_4 (Conv2D)           (None, 16, 16, 256)       32768\n_________________________________________________________________\nconv_pw_4_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_pw_4_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pad_5 (ZeroPadding2D)   (None, 18, 18, 256)       0\n_________________________________________________________________\nconv_dw_5 (DepthwiseConv2D)  (None, 16, 16, 256)       2304\n_________________________________________________________________\nconv_dw_5_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_dw_5_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pw_5 (Conv2D)           (None, 16, 16, 256)       65536\n_________________________________________________________________\nconv_pw_5_bn (BatchNormaliza (None, 16, 16, 256)       1024\n_________________________________________________________________\nconv_pw_5_relu (Activation)  (None, 16, 16, 256)       0\n_________________________________________________________________\nconv_pad_6 (ZeroPadding2D)   (None, 18, 18, 256)       0\n_________________________________________________________________\nconv_dw_6 (DepthwiseConv2D)  (None, 8, 8, 256)         2304\n_________________________________________________________________\nconv_dw_6_bn (BatchNormaliza (None, 8, 8, 256)         1024\n_________________________________________________________________\nconv_dw_6_relu (Activation)  (None, 8, 8, 256)         0\n_________________________________________________________________\nconv_pw_6 (Conv2D)           (None, 8, 8, 512)         131072\n_________________________________________________________________\nconv_pw_6_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_6_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_7 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_7 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_7_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_7_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_7 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_7_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_7_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_8 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_8 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_8_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_8_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_8 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_8_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_8_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_9 (ZeroPadding2D)   (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_9 (DepthwiseConv2D)  (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_9_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_9_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_9 (Conv2D)           (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_9_bn (BatchNormaliza (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_9_relu (Activation)  (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_10 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_10 (DepthwiseConv2D) (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_10_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_10_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_10 (Conv2D)          (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_10_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_10_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_11 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_11 (DepthwiseConv2D) (None, 8, 8, 512)         4608\n_________________________________________________________________\nconv_dw_11_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_dw_11_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pw_11 (Conv2D)          (None, 8, 8, 512)         262144\n_________________________________________________________________\nconv_pw_11_bn (BatchNormaliz (None, 8, 8, 512)         2048\n_________________________________________________________________\nconv_pw_11_relu (Activation) (None, 8, 8, 512)         0\n_________________________________________________________________\nconv_pad_12 (ZeroPadding2D)  (None, 10, 10, 512)       0\n_________________________________________________________________\nconv_dw_12 (DepthwiseConv2D) (None, 4, 4, 512)         4608\n_________________________________________________________________\nconv_dw_12_bn (BatchNormaliz (None, 4, 4, 512)         2048\n_________________________________________________________________\nconv_dw_12_relu (Activation) (None, 4, 4, 512)         0\n_________________________________________________________________\nconv_pw_12 (Conv2D)          (None, 4, 4, 1024)        524288\n_________________________________________________________________\nconv_pw_12_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_pw_12_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nconv_pad_13 (ZeroPadding2D)  (None, 6, 6, 1024)        0\n_________________________________________________________________\nconv_dw_13 (DepthwiseConv2D) (None, 4, 4, 1024)        9216\n_________________________________________________________________\nconv_dw_13_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_dw_13_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nconv_pw_13 (Conv2D)          (None, 4, 4, 1024)        1048576\n_________________________________________________________________\nconv_pw_13_bn (BatchNormaliz (None, 4, 4, 1024)        4096\n_________________________________________________________________\nconv_pw_13_relu (Activation) (None, 4, 4, 1024)        0\n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 4, 4, 156)         159900\n_________________________________________________________________\nglobal_average_pooling2d_1 ( (None, 156)               0\n=================================================================\nTotal params: 3,388,764\nTrainable params: 3,366,876\nNon-trainable params: 21,888\n<\/code><\/pre>\n","answers":[{"AnswerId":"53913092","CreationDate":"2018-12-24T11:54:48.530","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>Regarding reducing dense layer size:<\/p>\n\n<blockquote>\n  <p>Compress the weight matrix W of an inner product (fully connected)\n  layer using truncated SVD<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/blob\/master\/tools\/compress_net.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/blob\/master\/tools\/compress_net.py<\/a><\/p>\n"}]}
+{"QuestionId":53361294,"AnswerCount":0,"Tags":"<python><tensorflow><tensorrt>","CreationDate":"2018-11-18T13:18:07.090","AcceptedAnswerId":null,"OwnerUserId":987397.0,"Title":"Tensorflow: optimize grahp for inference. Provide shape for input placeholder","Body":"<p>This question is related to <a href=\"https:\/\/stackoverflow.com\/questions\/52319765\/swap-a-tensorflow-dataset-input-pipeline-with-a-placeholder-after-training\">Swap a TensorFlow Dataset input pipeline with a placeholder after training<\/a><\/p>\n\n<p>I have a graph that I want to optimize for inference with TensorRT. To get the graph, exluding the input pipeline part I do:<\/p>\n\n<pre><code>python optimize_for_inference.py \\\n--input=frozen_model.pb \\\n--output=inference.pb \\\n--frozen_graph=True \\\n--input_names=IteratorGetNext\n--output_names=name_of_the_output_tensor_you_want_to_use\n<\/code><\/pre>\n\n<p>\"This will replace the IteratorGetNext node with a float placeholder. You might want to choose another node, in which case just change the name. You can also change the type of the generated placeholder via the --placeholder_type_enum option. In that case, you need to provide an integer value matching the datatype you want from the DataType enum.\" (answer on this question <a href=\"https:\/\/stackoverflow.com\/questions\/52319765\/swap-a-tensorflow-dataset-input-pipeline-with-a-placeholder-after-training\">Swap a TensorFlow Dataset input pipeline with a placeholder after training<\/a>\n)<\/p>\n\n<p>However, the float placeholder does not have a shape. When doing giving the graph to TensorRT it fails because this shape is missing. Is it possible to give a shape to this float placeholder?<\/p>\n\n<p>To be more specific, the error given by TensorRT is<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.NotFoundError: No attr named\n  'shape' in NodeDef:    [[Node: input_images =\n  Placeholderdtype=DT_FLOAT]] for 'input_images' (op: 'Placeholder')\n  with input shapes: .<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53362127,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-18T14:48:28.767","AcceptedAnswerId":53362230.0,"OwnerUserId":1897935.0,"Title":"TensorFlow GPU: No Performance increase in HelloWorld code","Body":"<p><strong>Background<\/strong>:<\/p>\n\n<p>I am a Python Developer new to TensorFlow. <\/p>\n\n<p>System Spec:<\/p>\n\n<ul>\n<li>i5-7200U CPU @ 2.50GHz \u00d7 4 <\/li>\n<li>GeForce 940MX 4GB<\/li>\n<li>Ubuntu 18<\/li>\n<\/ul>\n\n<p>I am running TensorFlow on Docker (found installing cuda stuff too complicated, and long, maybe i messed up something)<\/p>\n\n<hr>\n\n<p>Basically I am running a kind of HelloWorld code on GPU and CPU and checking what kind of difference will it have and to my surprise there is hardly any!<\/p>\n\n<p><strong>docker-compose.yml<\/strong><\/p>\n\n<pre><code>version: '2.3'\n\nservices:\n  tensorflow:\n    # image: tensorflow\/tensorflow:latest-gpu-py3\n    image: tensorflow\/tensorflow:latest-py3\n    runtime: nvidia\n    volumes:\n      - .\/:\/notebooks\/TensorTest1\n    ports:\n      - 8888:8888\n<\/code><\/pre>\n\n<p>When I run with <code>image: tensorflow\/tensorflow:latest-py3<\/code> I get approx 5 seconds.<\/p>\n\n<pre><code>root@e7dc71acfa59:\/notebooks\/TensorTest1# python3 hello1.py \n2018-11-18 14:37:24.288321: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nTIME: 4.900559186935425\nresult:  [3. 3. 3. ... 3. 3. 3.]\n<\/code><\/pre>\n\n<p>when I run with <code>image: tensorflow\/tensorflow:latest-gpu-py3<\/code> I again get approx 5 seconds.<\/p>\n\n<pre><code>root@baf68fc71921:\/notebooks\/TensorTest1# python3 hello1.py \n2018-11-18 14:39:39.811575: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2018-11-18 14:39:39.877483: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-18 14:39:39.878122: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 0 with properties: \nname: GeForce 940MX major: 5 minor: 0 memoryClockRate(GHz): 1.189\npciBusID: 0000:01:00.0\ntotalMemory: 3.95GiB freeMemory: 3.56GiB\n2018-11-18 14:39:39.878148: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0\n 2018-11-18 14:44:17.101263: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-18 14:44:17.101303: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 \n2018-11-18 14:44:17.101313: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N \n2018-11-18 14:44:17.101540: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 3259 MB memory) -&gt; physical GPU (device: 0, name: GeForce 940MX, pci bus id: 0000:01:00.0, compute capability: 5.0)\nTIME: 5.82940673828125\nresult:  [3. 3. 3. ... 3. 3. 3.]\n<\/code><\/pre>\n\n<p><strong>My Code<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nimport time\n\nwith tf.Session():\n    start_time = time.time()\n\n    input1 = tf.constant([1.0, 1.0, 1.0, 1.0] * 100 * 100 * 100)\n    input2 = tf.constant([2.0, 2.0, 2.0, 2.0] * 100 * 100 * 100)\n    output = tf.add(input1, input2)\n    result = output.eval()\n\n    duration = time.time() - start_time\n    print(\"TIME:\", duration)\n\n    print(\"result: \", result)\n<\/code><\/pre>\n\n<p>Am I doing something wrong here? Based on prints it seems to be using GPU correctly<\/p>\n\n<hr>\n\n<p>Followed these steps at <a href=\"https:\/\/stackoverflow.com\/a\/37774470\/1897935\">Can I measure the execution time of individual operations with TensorFlow?<\/a> and I got this <a href=\"https:\/\/i.stack.imgur.com\/qNZSa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qNZSa.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53362230","CreationDate":"2018-11-18T14:57:39.277","ParentId":null,"OwnerUserId":"152948","Title":null,"Body":"<p>A GPU is an \"external\" processor, there's overhead involved in compiling a program for it, running it, sending it data, and retrieving the results. GPUs also have different performance tradeoffs from CPUs. While GPUs are frequently faster for large and complex number-crunching tasks, your \"hello world\" is too simple. It doesn't do very much with each data item between loading it and saving it (just pairwise addition), and it doesn't do very much at <em>all<\/em> \u2014 a million operations is nothing. That makes any setup\/teardown overhead relatively more noticeable. So while the GPU is slower for <em>this<\/em> program it's still likely to be faster for more useful programs.<\/p>\n"}]}
+{"QuestionId":53362642,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-18T15:43:34.033","AcceptedAnswerId":53362862.0,"OwnerUserId":2455494.0,"Title":"What does the first value returned from tf.metrics.accuracy represent","Body":"<p>I'd like to understand the values returned from <code>tf.metrics.accuracy<\/code><\/p>\n\n<p>Consider this minimal example:<\/p>\n\n<pre><code>predictions = tf.constant([[1,0,1], [1,0,1]], dtype=tf.int32)\nlabels = tf.constant([[0,0,1], [1,0,1]], dtype=tf.int32)\n\nacc = tf.metrics.accuracy(labels, predictions)\n\nwith tf.Session() as sess:\n  sess.run(tf.local_variables_initializer())\n  res = sess.run(acc)\nprint(res) \n<\/code><\/pre>\n\n<p>which outputs:<\/p>\n\n<pre><code>(0.0, 0.8333333)\n<\/code><\/pre>\n\n<p>The second value is what I'd expect (we have 5\/6 values where <code>predictions == labels<\/code> so 5\/6 ~= 0.83) but the first zero value is a mystery to me?<\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/metrics\/accuracy\" rel=\"nofollow noreferrer\">TensorFlow docs<\/a> have this to say:<\/p>\n\n<blockquote>\n  <p>Returns:<\/p>\n  \n  <ul>\n  <li>accuracy: A Tensor representing the accuracy, the value of\n  total divided by count. <\/li>\n  <li>update_op: An operation that increments the\n  total and count variables appropriately and whose value matches\n  accuracy.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>but this suggests that both return ops should have value of 0.83, no? (but then what would be the point of returning 2 ops?!)<\/p>\n\n<p>Feel like I am missing something fundamental!<\/p>\n","answers":[{"AnswerId":"53362862","CreationDate":"2018-11-18T16:06:05.587","ParentId":null,"OwnerUserId":"5245893","Title":null,"Body":"<p>Based on tensorflow code and the description, first value is the accuracy computed without the latest batch of data (in case of streaming data) and the second value is the final accuracy computed with all the data. This is used by tensorflow for stream processing of data.    <\/p>\n"}]}
+{"QuestionId":53363052,"AnswerCount":0,"Tags":"<import><keras>","CreationDate":"2018-11-18T16:27:11.810","AcceptedAnswerId":null,"OwnerUserId":5182570.0,"Title":"Hidden layers within sequential layer after importing model in Keras","Body":"<p>i have a \"simple\" problem after loading Keras model.<\/p>\n\n<p>During training my network has following structure<\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/clmli.png\" rel=\"nofollow noreferrer\">Output from command line after defining architecture<\/a><\/p>\n\n<p>After training + saving + loading my network appears to have following structure:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NgdDR.png\" rel=\"nofollow noreferrer\">Output from command line after loading<\/a><\/p>\n\n<p>Unfortunately I'm unable to find the way to access the sequential layer. My goal is to visualize the feature maps which are within the sequential layer.<\/p>\n\n<p>I will provide code in case it will be necessary.<\/p>\n\n<p>Thank you,\nVaclav<\/p>\n","answers":[]}
+{"QuestionId":53363060,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><prediction>","CreationDate":"2018-11-18T16:28:01.717","AcceptedAnswerId":53365738.0,"OwnerUserId":8513493.0,"Title":"keras unable to call model.predict_classes for multiple times","Body":"<pre><code>def predictOne(imgPath):\n\n    model = load_model(\"withImageMagic.h5\")\n    image = read_image(imgPath)\n    test_sample = preprocess(image)\n    predicted_class = model.predict_classes(([test_sample]))\n    return predicted_class\n<\/code><\/pre>\n\n<p>I have already trained a model. In this function, I load my model, read a new image, do some preprocessing and finally predict its label. <\/p>\n\n<p>When I run my main.py file, this function is called and everything goes smoothly. However, after a couple of seconds, this function will be called again with another image and I get this error:<\/p>\n\n<blockquote>\n<pre><code>'Cannot interpret feed_dict key as Tensor: ' + e.args[0])\n<\/code><\/pre>\n  \n  <p>TypeError: Cannot interpret feed_dict key as Tensor: Tensor Tensor(\"Placeholder:0\", shape=(5, 5, 1, 32), dtype=float32) is not an element of this graph.<\/p>\n<\/blockquote>\n\n<p>It's very strange that the function only works the first time. I tested multiple images and got the same behavior.<\/p>\n\n<p>Windows 10 - tensorflow-gpu with keras<\/p>\n","answers":[{"AnswerId":"53365738","CreationDate":"2018-11-18T21:41:29.087","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>Try loading model from file outside the function, and give the model object as argument to the function <code>def predictOne(imgPath, model)<\/code>. This will also be much faster, since the weights don't need to be loaded from disk every time a prediction is needed. <\/p>\n\n<p>If you want to keep loading model inside the function, import the backend:<\/p>\n\n<pre><code>from keras import backend as K\n<\/code><\/pre>\n\n<p>and then <\/p>\n\n<pre><code>K.clear_session() \n<\/code><\/pre>\n\n<p>before loading the model.<\/p>\n"}]}
+{"QuestionId":53363204,"AnswerCount":2,"Tags":"<python><numpy><keras><one-hot-encoding>","CreationDate":"2018-11-18T16:43:01.853","AcceptedAnswerId":53363247.0,"OwnerUserId":5231224.0,"Title":"Keras one-hot-encoder","Body":"<p>I have an array, and use the to_categorical function in keras:<\/p>\n\n<pre><code>labels = np.array([1,7,7,1,7])\nkeras.utils.to_categorical(labels)\n<\/code><\/pre>\n\n<p>I get this response:<\/p>\n\n<pre><code>array([[0., 1., 0., 0., 0., 0., 0., 0.],\n   [0., 0., 0., 0., 0., 0., 0., 1.],\n   [0., 0., 0., 0., 0., 0., 0., 1.],\n   [0., 1., 0., 0., 0., 0., 0., 0.],\n   [0., 0., 0., 0., 0., 0., 0., 1.]], dtype=float32)\n<\/code><\/pre>\n\n<p>How can I get only two columns? One for the 1 and one for the 7.<\/p>\n\n<p>This is a possible way, but not a very good one:<\/p>\n\n<pre><code>labels = np.delete(labels, np.s_[0:1], axis=1)\nnp.delete(labels, np.s_[1:6], axis=1)\n<\/code><\/pre>\n\n<p>that gives:<\/p>\n\n<pre><code>array([[1., 0.],\n       [0., 1.],\n       [0., 1.],\n       [1., 0.],\n       [0., 1.]], dtype=float32)\n<\/code><\/pre>\n\n<p>Is there a better way to achieve this? Preferably by some \"hidden\" function in Keras utils or similar? <\/p>\n","answers":[{"AnswerId":"53363219","CreationDate":"2018-11-18T16:45:21.497","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>Use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.0\/reference\/generated\/numpy.unique.html\" rel=\"nofollow noreferrer\"><code>np.unique<\/code><\/a> with <code>return_inverse<\/code> flag -<\/p>\n\n<pre><code># Get unique IDs mapped to each group of elements\nIn [73]: unql, idx = np.unique(labels, return_inverse=True)\n\n# Perform outer comparison for idx against range of unique groups\nIn [74]: (idx[:,None] == np.arange(len(unql))).astype(float)\nOut[74]: \narray([[1., 0.],\n       [0., 1.],\n       [0., 1.],\n       [1., 0.],\n       [0., 1.]])\n<\/code><\/pre>\n\n<p>Alternatively with direct usage of unique labels -<\/p>\n\n<pre><code>In [96]: (labels[:,None] == np.unique(labels)).astype(float)\nOut[96]: \narray([[1., 0.],\n       [0., 1.],\n       [0., 1.],\n       [1., 0.],\n       [0., 1.]])\n<\/code><\/pre>\n"},{"AnswerId":"53363247","CreationDate":"2018-11-18T16:48:32.330","ParentId":null,"OwnerUserId":"6671176","Title":null,"Body":"<p>IIUC, you can just index your array by <em>any<\/em> column that has a value:<\/p>\n\n<pre><code>cat = keras.utils.to_categorical(labels)\n&gt;&gt;&gt; cat\narray([[0., 1., 0., 0., 0., 0., 0., 0.],\n       [0., 0., 0., 0., 0., 0., 0., 1.],\n       [0., 0., 0., 0., 0., 0., 0., 1.],\n       [0., 1., 0., 0., 0., 0., 0., 0.],\n       [0., 0., 0., 0., 0., 0., 0., 1.]])\n\n# Select column if it has at least one value:\n&gt;&gt;&gt; cat[:,cat.any(0)]\narray([[1., 0.],\n       [0., 1.],\n       [0., 1.],\n       [1., 0.],\n       [0., 1.]])\n<\/code><\/pre>\n\n<p>You could also use <code>pandas<\/code>:<\/p>\n\n<pre><code>import pandas as pd\ncat = pd.get_dummies(labels).values\n&gt;&gt;&gt; cat\narray([[1, 0],\n       [0, 1],\n       [0, 1],\n       [1, 0],\n       [0, 1]], dtype=uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53363349,"AnswerCount":1,"Tags":"<python><tensorflow><shapes><rnn><tensor>","CreationDate":"2018-11-18T16:59:31.393","AcceptedAnswerId":53364652.0,"OwnerUserId":9918142.0,"Title":"Tensorflow in_top_kv","Body":"<p>Hi there i'm trying to build a simple rnn with 11 inputs and 2 outputs \n<code>X=tf.placeholder(tf.float32,[None,n_steps,n_inputs])\ny=tf.placeholder(tf.int32,[None,n_steps,n_outputs])<\/code><\/p>\n\n<p>I know the rnn excepts an input in the shape of [batch_size,n_steps,n_inputs] so that's why i have shaped my placeholders like this <\/p>\n\n<p>However when i run the code i get an error <\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 2 but is rank 3 for 'in_top_k\/InTopKV2' (op: 'InTopKV2') with input shapes: [1,270,2], [1,270,2], [].<\/p>\n<\/blockquote>\n\n<p>The error seems to originate here : <code>correct = tf.nn.in_top_k(logits,tf.reshape(y,[1,n_steps,n_outputs]),1)<\/code><\/p>\n\n<p>I have tried reshaping the logits, squeezing the logits, expanding the y dimensions, but nothing seems to work.<\/p>\n\n<p>One difference that i have noticed is that when i squeeze the logits with <\/p>\n\n<pre><code>tf.squeeze(logits)\n<\/code><\/pre>\n\n<p>The error now says <\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 1 but is rank 3<\/p>\n<\/blockquote>\n\n<p>That is the only 'progress' that i have been able to make, any help would be appreciated.<\/p>\n\n<p>p.s go easy on me this is my first question ever<\/p>\n","answers":[{"AnswerId":"53364652","CreationDate":"2018-11-18T19:27:51.403","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You have to reshape the inputs as 2D tensors, then you can reshape the result back to the desired shape:<\/p>\n\n<pre><code>logits_res = tf.reshape(logits, (-1, n_outputs))\ny_res = tf.reshape(y, (-1, n_outputs))\ncorrect_res = tf.nn.in_top_k(logits_res, y_res, 1)\ncorrect = tf.reshape(correct_res, (-1, n_steps))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53363360,"AnswerCount":0,"Tags":"<python><keras><word2vec><cosine-similarity><word-embedding>","CreationDate":"2018-11-18T17:00:29.533","AcceptedAnswerId":null,"OwnerUserId":4592081.0,"Title":"How is cosine similarity performed in this Keras word2vec implementation?","Body":"<p>I'm following a tutorial\n<a href=\"http:\/\/adventuresinmachinelearning.com\/word2vec-keras-tutorial\/\" rel=\"nofollow noreferrer\">http:\/\/adventuresinmachinelearning.com\/word2vec-keras-tutorial\/<\/a>\nwhich uses cosine similarity for validation model but dot product (not normalized unlike cosine similarity) for training. That's really confusing but my question is about axis of dot product, for validation model it's:<\/p>\n\n<pre><code>similarity = merge([target, context], mode='cos', dot_axes=0)\n<\/code><\/pre>\n\n<p>as a result when I log the model summary for the validation model I get:<\/p>\n\n<blockquote>\n  <p>Layer (type) Output Shape<\/p>\n  \n  <p>dot_1 (Dot) (300, 1, 1)<\/p>\n<\/blockquote>\n\n<p>but for training model<\/p>\n\n<pre><code>dot_product = merge([target, context], mode='dot', dot_axes=1)\n<\/code><\/pre>\n\n<p>not surprisingly model summary shows a different thing<\/p>\n\n<blockquote>\n  <p>Layer (type) Output Shape<\/p>\n  \n  <p>dot_2 (Dot) (None, 1, 1)<\/p>\n<\/blockquote>\n\n<p>I've read 'Ouput' in the summary (who's value is usually None) shows batch size (if it's fixed for some reason) but it still doesn't make sense to me why it's 300 for the first model because I think batch size in the example is not fixed and 300 is also my embedding dimension. Clearly something's wrong the <code>dot_axes<\/code>, can anybody confirm? Do I just fix it by putting <code>dot_axes=1<\/code> instead of zero? Why is first axis 1 and not 0? Does <code>dot_axes=0<\/code> dot over batch size or what (makes no sense to me)?<\/p>\n\n<p>code on github:\n<a href=\"https:\/\/github.com\/adventuresinML\/adventures-in-ml-code\/blob\/master\/keras_word2vec.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/adventuresinML\/adventures-in-ml-code\/blob\/master\/keras_word2vec.py<\/a><\/p>\n","answers":[]}
+{"QuestionId":53363663,"AnswerCount":1,"Tags":"<python><conv-neural-network><pytorch><google-colaboratory>","CreationDate":"2018-11-18T17:31:29.197","AcceptedAnswerId":null,"OwnerUserId":9171186.0,"Title":"Issues loading from Drive with pytorch' datasets.DatasetFolder","Body":"<p>The loading works great using jupyter and local files, but when I adapted to Colab, fetching data from a Drive folder, datasets.DatasetFolder always loads 9500 odd datapoints, never the full 10 000. Anyone had similar issues?<\/p>\n\n<pre><code>train_data = datasets.DatasetFolder('\/content\/drive\/My Drive\/4 - kaggle\/data', np.load, list(('npy'))  )\n\nprint(train_data.__len__)\n<\/code><\/pre>\n\n<p>Returns<\/p>\n\n<pre><code>&lt;bound method DatasetFolder.__len__ of Dataset DatasetFolder\n    Number of datapoints: 9554\n    Root Location: \/content\/drive\/My Drive\/4 - kaggle\/data\n    Transforms (if any): None\n    Target Transforms (if any): None&gt;\n<\/code><\/pre>\n\n<p>Where I would get the full 10 000 elements usually.<\/p>\n","answers":[{"AnswerId":"53754048","CreationDate":"2018-12-13T02:04:00.253","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p>Loading lots of files from a single folder in Drive is likely to be slow and error-prone. You'll probably end up much happier if you either stage the data on GCS or upload an archive (.zip or .tar.gz) to Drive and copy that one file to your colab VM, unarchive it there, and then run your code over the local data.<\/p>\n"}]}
+{"QuestionId":53363858,"AnswerCount":1,"Tags":"<tensorflow><anaconda><python-3.6>","CreationDate":"2018-11-18T17:54:00.000","AcceptedAnswerId":57421018.0,"OwnerUserId":9409770.0,"Title":"Getting error named Unknown Meta Architecture after a rookie mistake while training in Anaconda venv tensorflow_cpu version","Body":"<p>I am new to Tensorflow as well as Data Science. Recently, I wanted to train custom object detection.<\/p>\n\n<p>Machine and Environment Types and Other Requirements:<\/p>\n\n<ol>\n<li>Anaconda to create and operate in venv<\/li>\n<li>Tensor Flow cpu version<\/li>\n<li>ssd-mobile-net-v1-coco<\/li>\n<li>Python 3.6<\/li>\n<\/ol>\n\n<p>At first I stored all my files in <\/p>\n\n<pre><code>C:\\Users\\Documents\\TensorFlow\\models\\research\n<\/code><\/pre>\n\n<p>directory as it was giving error continuously while I try to run this from other folder.<\/p>\n\n<p>then after copying all the files and folders of my custom object's to above directory, I started the venv of tensorflow_cpu in anaconda prompt and then I wrote following command:<\/p>\n\n<pre><code>python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/pipeline.config\n<\/code><\/pre>\n\n<p>Then the training began,<\/p>\n\n<p>I have done all of this by following a blog. I wanted to keep my global loss at least 2. But suddenly without any intention to give keyboard interrupt I pressed ctrl+c and the train.py file terminated. Now when I re-run this, I get the following error:<\/p>\n\n<pre><code>(tensorflow_cpu) C:\\Users\\arfan\\Documents\\TensorFlow\\models\\research&gt;python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/pipeline.config\nWARNING:tensorflow:From C:\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\platform\\app.py:125: main (from __main__) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse object_detection\/model_main.py.\nW1118 23:02:23.971923  9792 tf_logging.py:125] From C:\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\platform\\app.py:125: main (from __main__) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse object_detection\/model_main.py.\nTraceback (most recent call last):\n  File \"train.py\", line 184, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"C:\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 306, in new_func\n    return func(*args, **kwargs)\n  File \"train.py\", line 180, in main\n    graph_hook_fn=graph_rewriter_fn)\n  File \"C:\\Users\\arfan\\Documents\\TensorFlow\\models\\research\\object_detection\\legacy\\trainer.py\", line 248, in train\n    detection_model = create_model_fn()\n  File \"C:\\Users\\arfan\\Documents\\TensorFlow\\models\\research\\object_detection\\builders\\model_builder.py\", line 122, in build\n    raise ValueError('Unknown meta architecture: {}'.format(meta_architecture))\nValueError: Unknown meta architecture: None\n<\/code><\/pre>\n\n<p>Can anyone please help? I am new to this field so please explain your answer. This is my first post here, extremely sorry if I make any unintentional mistake.<\/p>\n","answers":[{"AnswerId":"57421018","CreationDate":"2019-08-08T21:45:23.987","ParentId":null,"OwnerUserId":"10678409","Title":null,"Body":"<p>I have also faced the same error. This error occurs due to the incorrect path given into config file. if it is a Windows machine, make sure you use forward slash at the paths you write into the config file.<\/p>\n"}]}
+{"QuestionId":53363986,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network><sequence><lstm>","CreationDate":"2018-11-18T18:08:39.477","AcceptedAnswerId":null,"OwnerUserId":9303844.0,"Title":"Adding static data( not changing over time) to sequence data in LSTM","Body":"<p>I am trying to build a model like the following figure. Please see the following image:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/lGvH6.png.\" alt=\"Please see the image\"><\/p>\n\n<p>I want to pass sequence data in LSTM layer and static data (blood group, gender) in another feed forward neural network layer. Later I want to merge them. However, I am confused about the dimenstion here.<\/p>\n\n<ul>\n<li>If my understaning is right(which i depict in  the image), how the 5-dimensional sequence data can be merged with 4 dimenstional static data?<\/li>\n<li>Also, what is the difference of attention mechanism with this structure? (I found in the KERAS documentation that attention mechanism is an way to add static data with sequence data) <\/li>\n<\/ul>\n\n<p>Basically, I want  to add  the static data with sequence data. Any other suggestion is apprciated.<\/p>\n","answers":[{"AnswerId":"53373263","CreationDate":"2018-11-19T11:04:47.303","ParentId":null,"OwnerUserId":"6145828","Title":null,"Body":"<p>I am not sure if I got what you are asking, but I will try.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7IHSF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7IHSF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Example in <code>Keras<\/code>:<\/p>\n\n<pre><code>static_out = (static_input)\n\nx = LSTM(n_cell_lstm, return_sequences=True)(dynamic_input)\nx = Flatten()(x)\ndynamic_out = (x)\n\nz = concatenate([dynamic_out, static_out])\n\nz = Dense(64, activation='relu')(z)\n\nmain_output = Dense(classes, activation='softmax', name='main_output')(z)\n<\/code><\/pre>\n\n<p>Practically you are using an LSTM architecture as you would if you where using only the dynamic data, but at the end you add the info coming from the static data. Hope this helps.<\/p>\n"}]}
+{"QuestionId":53364001,"AnswerCount":1,"Tags":"<python><ios><tensorflow><tensorflow-lite><pre-trained-model>","CreationDate":"2018-11-18T18:11:02.703","AcceptedAnswerId":null,"OwnerUserId":441499.0,"Title":"Tensorflow Lite iOS Camera example does not work with retrained MobileNet model","Body":"<p>I'm trying to run <a href=\"https:\/\/www.tensorflow.org\/lite\/demo_ios\" rel=\"nofollow noreferrer\">the Tensorflow Lite Camera example<\/a> with a retrained Mobilenet model.<\/p>\n\n<p>I successfully run the iOS camera app according to <a href=\"https:\/\/www.tensorflow.org\/lite\/demo_ios\" rel=\"nofollow noreferrer\">the instructions<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/23907\" rel=\"nofollow noreferrer\">this fix<\/a>. The app works as expected with the model <code>mobilenet_v1_1.0_224.tflite<\/code>.<\/p>\n\n<p>I install Tensorflow:<\/p>\n\n<pre><code>pip3 install -U virtualenv\nvirtualenv --system-site-packages -p python3 .\/venv\nsource .\/venv\/bin\/activate\npip install --upgrade pip\npip install --upgrade tensorflow==1.12.0\npip install --upgrade tensorflow-hub==40.6.2\n<\/code><\/pre>\n\n<p>I now want to <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">retrain the model<\/a> using <a href=\"http:\/\/download.tensorflow.org\/example_images\/flower_photos.tgz\" rel=\"nofollow noreferrer\">the flowers set<\/a>. I download the flowers folder and run: <\/p>\n\n<pre><code>python retrain.py \\\n    --bottleneck_dir=bottleneck \\\n    --how_many_training_steps=400 \\\n    --model_dir=model \\\n    --output_graph=pola_retrained.pb \\\n    --output_labels=pola_retrained_labels.txt \\\n    --tfhub_module https:\/\/tfhub.dev\/google\/imagenet\/mobilenet_v1_100_224\/quantops\/feature_vector\/1 \\\n    --image_dir=flower_photos\n<\/code><\/pre>\n\n<p>Note: I can successfully test the retrained model using <code>label_image.py<\/code> script.<\/p>\n\n<p>I convert the retrained model to its <code>tflite<\/code> format:<\/p>\n\n<pre><code>toco \\\n  --graph_def_file=pola_retrained.pb \\\n  --input_format=TENSORFLOW_GRAPHDEF \\\n  --output_format=TFLITE \\\n  --output_file=mobilenet_v1_1.0_224.tflite \\\n  --inference_type=FLOAT \\\n  --input_type=FLOAT \\\n  --input_arrays=Placeholder \\\n  --output_arrays=final_result \\\n  --input_shapes=1,224,224,3\n<\/code><\/pre>\n\n<p>I copy both the new model and the labels file to the iOS app. I modify the app parameters in <code>CameraExampleViewController.mm<\/code> as follows:<\/p>\n\n<pre><code>\/\/ These dimensions need to match those the model was trained with.\nconst int wanted_input_width = 224;\nconst int wanted_input_height = 224;\nconst int wanted_input_channels = 3;\nconst float input_mean = 128.0f;\nconst float input_std = 128.0f;\nconst std::string input_layer_name = \"input\";\nconst std::string output_layer_name = \"final_result\";\n<\/code><\/pre>\n\n<p>The app crashes. The index of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/ios\/camera\/CameraExampleViewController.mm#L349:L354\" rel=\"nofollow noreferrer\">the recognized object is outside of the range of trained objects<\/a>. The confidence level is above 1.<\/p>\n","answers":[{"AnswerId":"53636407","CreationDate":"2018-12-05T16:13:30.833","ParentId":null,"OwnerUserId":"441499","Title":null,"Body":"<p>The Tensorflow Lite Camera example hardcodes the output tensor size to <code>1000<\/code>. If you test the example with a retrained model having a smaller number of outputs, the iOS app crashes. Replace the following code in <code>CameraExampleViewController.mm<\/code>:<\/p>\n\n<pre><code>const int output_size = 1000;\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>\/\/ read output size from the output sensor\nconst int output_tensor_index = interpreter-&gt;outputs()[0];\nTfLiteTensor* output_tensor = interpreter-&gt;tensor(output_tensor_index);\nTfLiteIntArray* output_dims = output_tensor-&gt;dims;\nassert(output_dims-&gt;size == 2);\nconst int output_size = output_dims-&gt;data[1]-output_dims-&gt;data[0];\n<\/code><\/pre>\n\n<p>The code above fixes the problem by reading the output size from the model output dimensions. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/24177\" rel=\"nofollow noreferrer\">The appropriate PR<\/a> has been submitted.<\/p>\n"}]}
+{"QuestionId":53364064,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><object-detection>","CreationDate":"2018-11-18T18:17:19.647","AcceptedAnswerId":null,"OwnerUserId":10379826.0,"Title":"Get Class and Probability in Tensorflow Object Detection API","Body":"<p>I have a problem with getting the class and the probability of a detected object in the Tensorflow Object Detection API. I would like to print these two values with every image.<\/p>\n\n<p>This is the code:<\/p>\n\n<pre><code>for image_path in TEST_IMAGE_PATHS:\n  image = Image.open(image_path)\n  # the array based representation of the image will be used later in order to prepare the\n  # result image with boxes and labels on it.\n  image_np = load_image_into_numpy_array(image)\n  # Expand dimensions since the model expects images to have shape: [1, None, None, 3]\n  image_np_expanded = np.expand_dims(image_np, axis=0)\n  # Actual detection.\n  output_dict = run_inference_for_single_image(image_np, detection_graph)\n  # Visualization of the results of a detection.\n  vis_util.visualize_boxes_and_labels_on_image_array(\n      image_np,\n      output_dict['detection_boxes'],\n      output_dict['detection_classes'],\n      output_dict['detection_scores'],\n      category_index,\n      instance_masks=output_dict.get('detection_masks'),\n      use_normalized_coordinates=True,\n      line_thickness=2)\n  plt.figure(figsize=IMAGE_SIZE)\n  plt.imshow(image_np)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53372976","CreationDate":"2018-11-19T10:48:22.403","ParentId":null,"OwnerUserId":"9479442","Title":null,"Body":"<p>the following code give the possibility to extract the class id and the scores for all the entities with a score higher than 50%.<\/p>\n\n<pre><code>#Create indexes list of element with a score &gt; 0.5\nindexes = [k for k,v in enumerate(output_dict['detection_scores']) if (v &gt; 0.5)]\n\n#Number of entities\nnum_entities = len(indexes)\n\n#Extract the class id\nclass_id = itemgetter(*indexes)(output_dict['detection_classes'])\nscores = itemgetter(*indexes)(output_dict['detection_scores'])\n\n#Convert the class id in their name\nclass_names = []\nif num_entities == 1:\n  class_names.append(category_index[class_id]['name'])\n  class_name = str(class_names)\nelse:\n  for i in range(0, len(indexes)):\n  class_names.append(category_index[class_id[i]]['name'])\n<\/code><\/pre>\n\n<p>The if is necessary if you detect only one element.<\/p>\n\n<p>Then you can print <code>class_names[i]<\/code> and <code>str(scores[i])<\/code><\/p>\n"}]}
+{"QuestionId":53364136,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-18T18:24:39.533","AcceptedAnswerId":null,"OwnerUserId":6341510.0,"Title":"Keras - multi-input model using flow_from_dataframe","Body":"<h2><strong>Problem<\/strong><\/h2>\n\n<p>I am trying to build a multi-input model in keras using two inputs, image and text. I am using the <code>flow_from_dataframe<\/code> method, passing it a pandas dataframe containing the image-names as well as the respective text (as a vectorized feature-represenation) for each image and the target label\/class. As such, the dataframe looks as follows:<\/p>\n\n<pre><code>ID     path                 text-features             label\n111    'cat001.jpg'       [0.0, 1.0, 0.0,...]         cat\n112    'dog001.jpg'       [1.0, 0.0, 1.0,...]         dog\n113    'bunny001.jpg'     [0.0, 1.0, 1.0,...]         bunny\n...\n<\/code><\/pre>\n\n<p>After constructing my model using the Keras functional API, I feed both inputs into the model like so:<\/p>\n\n<pre><code>model = Model(inputs=[images, text], outputs=output)\n<\/code><\/pre>\n\n<p>For the images I use an ImageDataGenerator as suggested in the docs (<a href=\"https:\/\/keras.io\/preprocessing\/image\/#flow_from_dataframe\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/#flow_from_dataframe<\/a>) :<\/p>\n\n<pre><code>datagen=ImageDataGenerator(rescale=1.\/255,validation_split=0.15)\n\ntrain_generator=datagen.flow_from_dataframe(dataframe=df, directory=data_dir, x_col=path, y_col=\"label\", has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,subset=\"training\")\n\n validation_generator=datagen.flow_from_dataframe(dataframe=df, directory=data_dir, x_col=path, y_col=\"label\", has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,subset=\"validation\")\n<\/code><\/pre>\n\n<p>So far so good, but now I am stuck on how to feed the text-features within my dataframe to the model as well during training.<\/p>\n\n<h2><strong>Question<\/strong><\/h2>\n\n<p>How can I modify the <code>flow_from_dataframe<\/code> generator in order to handle the text-feature data in the dataframe as well as the images during training? Also, since I can't find any example of this sort of modification on <code>flow_from_dataframe<\/code> I am wondering if I am approaching this problem wrong i.e. is there any better method of achieving this? <\/p>\n\n<h2>UPDATE<\/h2>\n\n<p>Meanwhile I've been trying to write my own generator following the guide I found here (<a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly<\/a>) and adjusting it to my needs. This is what I came up with:<\/p>\n\n<pre><code>from matplotlib.image import imread\n\n\nclass DataGenerator(keras.utils.Sequence):\n    def __init__(self, list_IDs, labels, batch_size=32, dim=(32,32,32), n_channels=1,\n                 n_classes=10, shuffle=True):\n        #'Initialization'\n        self.dim = dim\n        self.batch_size = batch_size\n        self.labels = labels\n        self.list_IDs = list_IDs\n        self.n_channels = n_channels\n        self.n_classes = n_classes\n        self.shuffle = shuffle\n        self.on_epoch_end() \n\n    def on_epoch_end(self):\n      #'Updates indexes after each epoch' \n        self.indexes = np.arange(len(self.list_IDs))\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n\n# method for producing batches of data.\n# takes as argument the list of IDs of the target batch\ndef __data_generation(self, list_IDs_temp):\n    #'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n    # Initialization\n    X = np.empty((self.batch_size, *self.dim, self.n_channels))\n    Xtext = np.empty((self.batch_size, 7576)) \n    y = np.empty((self.batch_size), dtype=int)\n\n    # Generate data\n    for i, ID in enumerate(list_IDs_temp):\n        # Store sample\n        X[i,] = imread('C:\/Users\/aaron\/Desktop\/training\/'+str(ID)) # &lt;--- all files are in the same DIR\n        Xtext[i,] = np.array(total_data[df.path== str(ID)][\"text-features\"].values) # &lt;--- I look-up the text-features by using the ID as a filter with the path column. This line throws the error. \n\n\n        # Store class\n        y[i] = self.labels[ID]\n\n    return X, Xtext, keras.utils.to_categorical(y, num_classes=self.n_classes)\n\ndef __len__(self):\n  #'Denotes the number of batches per epoch'\n  return int(np.floor(len(self.list_IDs) \/ self.batch_size))\n\n# Now, when the batch corresponding to a given index is called, \n# the generator executes the __getitem__ method to generate it.\ndef __getitem__(self, index):\n    #'Generate one batch of data'\n    # Generate indexes of the batch\n    indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]\n\n    # Find list of IDs\n    list_IDs_temp = [self.list_IDs[k] for k in indexes]\n\n    # Generate data\n    X,Xtext, y = self.__data_generation(list_IDs_temp)\n\n    return X,Xtext, y\n<\/code><\/pre>\n\n<p>And I initialize the generator as follows:<\/p>\n\n<pre><code>partition = {}\npartition['train'] = X_train.path.values\npartition['validation'] = X_test.path.values\n\nfrom sklearn import preprocessing\nle = preprocessing.LabelEncoder()\nencoded_labels = le.fit_transform(df.label)\n\nlabels = pd.Series(encoded_labels,index=df.path).to_dict()\n\n# Parameters\nparams = {'dim': (224,224),\n          'batch_size': 64,\n          'n_classes': 5,\n          'n_channels': 3,\n          'shuffle': True}\n\n\n# Generators\ntraining_generator = DataGenerator(partition['train'], labels, **params)\nvalidation_generator = DataGenerator(partition['validation'], labels, **params)\n<\/code><\/pre>\n\n<p>Using this generator however throws me an error:<\/p>\n\n<blockquote>\n  <p>ValueError: setting an array element with a sequence.<\/p>\n<\/blockquote>\n\n<p>caused by the line  <code>X_text[i,] = np.array(total_data[total_data.bust == str(ID)].text.values)<\/code> in my code above. Any suggestion on how to solve this?<\/p>\n","answers":[]}
+{"QuestionId":53364952,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-11-18T20:07:44.823","AcceptedAnswerId":null,"OwnerUserId":10614496.0,"Title":"Weird data type error using tensorflow and numpy","Body":"<p>I don't understand the difference betwen the two codes here below:<\/p>\n\n<p>Is there a reason why this works:<\/p>\n\n<pre><code>def evaluer_dessin ():\n    global result,img\n    listeimage=[]\n    for y in matrice:\n        for x in y:\n            listeimage.append(x)\n    img = np.array([listeimage])\n\n    afficher_resultat()  \n\n    try:\n        Cadre.delete(result)\n    except:\n        pass\n    if evaluerimg:\n        result=Cadre.create_text(Largeur\/2,Hauteur\/8,text=\"c'est un \n\"+str((prediction[0]+1)%10),font=('', '50'),fill=\"red\")\n    root.after(300,evaluer_dessin)\n\ndef afficher_resultat():\n    global prediction\n    prediction = session.run(y_pred_cls,feed_dict={ x : img})\n<\/code><\/pre>\n\n<p>But this code here doesn't :<\/p>\n\n<pre><code>def evaluer_dessin ():\n    global result\n    listeimage=[]\n    for y in matrice:\n        for x in y:\n            listeimage.append(x)\n    img = np.array([listeimage])\n\n    # this is the line that doesn't work, it says that it cant convert an\n    # int into a tensor\n    prediction = session.run(y_pred_cls,feed_dict={ x : img})\n\n    try:\n        Cadre.delete(result)\n    except:\n        pass\n    if evaluerimg:\n        result=Cadre.create_text(Largeur\/2,Hauteur\/8,text=\"c'est un \n\"+str((prediction[0]+1)%10),font=('', '50'),fill=\"red\")\n    root.after(300,evaluer_dessin)\n<\/code><\/pre>\n\n<p>It says that \"img\" is an int but it's type is clearly an array, and the weirdest thing is that when I use a seperate function to calculate the prediction, it works well!<\/p>\n\n<p>Here's the error shown for the second program:<\/p>\n\n<pre><code>&gt;Traceback (most recent call last):\n&gt;  File \"C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1092, in _run\n&gt;    subfeed, allow_tensor=True, allow_operation=False)\n&gt;  File \"C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3490, in as_graph_element\n&gt;    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)\n&gt;  File \"C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3579, in &gt;_as_graph_element_locked\n&gt;    types_str))\n&gt;TypeError: Can not convert a int into a Tensor.\n&gt;\n&gt;During handling of the above exception, another exception occurred:\n&gt;\n&gt;Traceback (most recent call last):\n&gt;  File \"C:\\Users\\nicol\\Desktop\\CNN_exe\\programmeCNNexe.py\", line 170, in &lt;module&gt;\n&gt;    evaluer_dessin()\n&gt;  File \"C:\\Users\\nicol\\Desktop\\CNN_exe\\programmeCNNexe.py\", line 123, in &gt;evaluer_dessin\n&gt;    prediction = session.run(y_pred_cls,feed_dict={ x : img})\n&gt;  File \"C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 929, in run\n&gt;    run_metadata_ptr)\n&gt;  File \"C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1095, in _run\n&gt;    'Cannot interpret feed_dict key as Tensor: ' + e.args[0])\n&gt;TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a int into &gt;a Tensor.\n&gt;[Finished in 2.1s with exit code 1]\n&gt;[shell_cmd: python -u \"C:\\Users\\nicol\\Desktop\\CNN_exe\\programmeCNNexe.py\"]\n&gt;[dir: C:\\Users\\nicol\\Desktop\\CNN_exe]\n&gt;[path: C:\\Program Files (x86)\\Intel\\iCLS Client\\;C:\\Program Files\\Intel\\iCLS &gt;Client\\;C:\\WINDOWS\\system32;C:\\WINDOWS;C:\\WINDOWS\\System32\\Wbem;C:\\WINDOWS\\Syst&gt;em32\\WindowsPowerShell\\v1.0\\;C:\\Program Files (x86)\\Intel\\Intel(R) Management &gt;Engine Components\\DAL;C:\\Program Files\\Intel\\Intel(R) Management Engine &gt;Components\\DAL;C:\\Program Files (x86)\\Intel\\Intel(R) Management Engine Components\\IPT;C:\\Program Files\\Intel\\Intel(R) Management Engine Components\\IPT;C:\\Program Files\\Intel\\WiFi\\bin\\;C:\\Program Files\\Common Files\\Intel\\WirelessCommon\\;C:\\WINDOWS\\System32\\OpenSSH\\;C:\\Users\\nicol\\AppData&gt;\\Local\\Programs\\Python\\Python35\\Scripts\\;C:\\Users\\nicol\\AppData\\Local\\Programs\\&gt;Python\\Python35\\;C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python37\\Scripts\\&gt;;C:\\Users\\nicol\\AppData\\Local\\Programs\\Python\\Python37\\;C:\\Users\\nicol\\AppData\\&gt;Local\\Microsoft\\WindowsApps;]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53417982","CreationDate":"2018-11-21T17:55:29.063","ParentId":null,"OwnerUserId":"10614496","Title":null,"Body":"<p>Thanks jdehesa for helping me solving the problem. The problem came actually from the variable x and not form img.\nAll I had to do was to give another name to the the variable in the for loop:<\/p>\n\n<pre><code>def evaluer_dessin ():\n    global listeimage,result,img\n    listeimage=[]\n    for y in matrice:\n        for z in y: # I just change the name from \"x\" to \"z\" and the problem was solved\n            listeimage.append(z)\n    img = np.array([listeimage])\n\n    prediction = session.run(y_pred_cls,feed_dict={ x : img})\n\n    try:\n        Cadre.delete(result)\n    except:\n        pass\n    if evaluerimg:\n        result=Cadre.create_text(Largeur\/2,Hauteur\/8,text=\"c'est un \"+str((prediction[0]+1)%10),font=('', '50'),fill=\"red\")\n    root.after(300,evaluer_dessin)\n<\/code><\/pre>\n\n<p>I'm new to stackoverflow and I just wanted to say how awesome this website is, you get all the answers you want by clever and nice people. <\/p>\n"}]}
+{"QuestionId":53365346,"AnswerCount":1,"Tags":"<tensorflow><image-processing><metadata><conv-neural-network>","CreationDate":"2018-11-18T20:53:51.503","AcceptedAnswerId":null,"OwnerUserId":10638765.0,"Title":"adding multiple metadata to tflearn CNN","Body":"<p>I'm using CNN for (medical) image analysis and prediction, using the typical CNN. I added one set of metadata to the CNN network like this and it seems to work:\n    network = input_data(shape=[..],..)\n    metadata_1 = input_data(shape=[..],..)<\/p>\n\n<pre><code>network = &lt;convolutions and some max pooling&gt;\nnetwork = fully_connected(network, 100,..&gt;\nnetwork = merge (network, metadata_1)\nnetwork = fully_connected ()\n...\n<\/code><\/pre>\n\n<p>Now, could i extend this to do this? Anyone has any experience? and pitfalls?<\/p>\n\n<pre><code>network = input_data(shape=[..],..)\nmetadata_1 = input_data(shape=[..],..)    \n...\nmetadata_n = input_data(shape=[..],..)\nnetwork = &lt;convolutions and some max pooling&gt;\nnetwork = fully_connected(network, 100,..&gt;\nnetwork = merge (network, metadata_1)\n...\nnetwork = merge (network, metadata_n)\nnetwork = fully_connected ()\n...\n<\/code><\/pre>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"53420185","CreationDate":"2018-11-21T20:42:53.913","ParentId":null,"OwnerUserId":"1636676","Title":null,"Body":"<p>I think you're talking about layer concatenation here. At least that's what I used in my CNNs. <\/p>\n\n<p>Now in your case you're adding metadata into consecutive layers n-times. This produces n extra layers, which can become memory intensive. What I find more intuitive is to use <em><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/concat\" rel=\"nofollow noreferrer\">concat<\/a><\/em> layer and concatenate conv and all metadata layers together. <\/p>\n\n<pre><code>network = &lt;convolutions and some max pooling&gt;\nnetwork = fully_connected(network, 100,..&gt;  \nnetwork = concat (network, metadata_1, metadata_2, ..., metadata_n)\nnetwork = fully_connected ()\n...\n<\/code><\/pre>\n\n<p>You might get different results with your approach, but I suspect there won't be much difference. If you want to know you should try both.<\/p>\n"}]}
+{"QuestionId":53365464,"AnswerCount":0,"Tags":"<python><keras><neural-network><conv-neural-network><data-science>","CreationDate":"2018-11-18T21:07:19.457","AcceptedAnswerId":null,"OwnerUserId":7933418.0,"Title":"Loading model from json object Keras","Body":"<p>My script to load Keras CNN model from saved json object is as follows:<\/p>\n\n<pre><code>import keras\nfrom keras.models import model_from_json\njson_file = open('k_model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nloaded_model = model_from_json(loaded_model_json)\n<\/code><\/pre>\n\n<p>model_from_json is throwing the following error :<\/p>\n\n<pre><code>TypeError: __init__() takes at least 4 arguments (3 given)\n<\/code><\/pre>\n\n<p>Has anyone faced similar issue? <\/p>\n\n<p>Note: In Python3 it works fine, even in Python2.7 it was working fine earlier. <\/p>\n\n<p>Complete stack trace as follows :<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"test.py\", line 63, in &lt;module&gt;\n    loaded_model = model_from_json(loaded_model_json)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/models.py\", line 197, in model_from_json\n    return layer_from_config(config, custom_objects=custom_objects)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/utils\/layer_utils.py\", line 36, in layer_from_config\n    return layer_class.from_config(config['config'])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/models.py\", line 1019, in from_config\n    layer = get_or_create_layer(first_layer)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/models.py\", line 1003, in get_or_create_layer\n    layer = layer_from_config(layer_data)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/utils\/layer_utils.py\", line 36, in layer_from_config\n    return layer_class.from_config(config['config'])\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/topology.py\", line 929, in from_config\n    return cls(**config)\nTypeError: __init__() takes at least 4 arguments (3 given)\nFailed with exit code: 1\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53365704,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-18T21:35:47.477","AcceptedAnswerId":null,"OwnerUserId":1139273.0,"Title":"Using TensorFlow Estimator in a loop","Body":"<p>I have implemented a custom estimator and now I like to use this estimator for training in a loop as follows:<\/p>\n\n<pre><code>for _ in xrange(steps):\n   tf.estimator.train(training_iteration)\n   post_process_results(...)\n<\/code><\/pre>\n\n<p>The problem is that each time tf.estimator.train is being called in the loop a new MonitoredSession is created and never being closed. As such, the allocated memory is never being released and at some point the whole process gets killed because of out of memory exception. Is there anyway to forcefully close the session being created by tf.estimator.train?<\/p>\n","answers":[]}
+{"QuestionId":53365950,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2018-11-18T22:08:48.450","AcceptedAnswerId":53366364.0,"OwnerUserId":10543278.0,"Title":"Tensorflow - Prepare a CSV file to feed a Recurrent Neural Network","Body":"<p>I have a question. At school we started a new project with Neural Networks and we had to choose what kind of AI we wanted to program. I chose for a Recurrent Neural Network which could predict if a price will be higher or lower after a few periods. I succesfully programmed that and it trained well. But now I want to try a test run, but I don't know how I can prepare a csv file to feed the RNN. This is my trainings code:<\/p>\n\n<pre><code>main_df = pd.DataFrame()\n\nratios = [\"BTC-USD\", \"LTC-USD\", \"ETH-USD\"]\nfor ratio in ratios:\n\n\n    url=\"https:\/\/www.test.nl\/get_csv_content.php?method=train&amp;ratio=\" + str(ratio)\n    dataset = requests.get(url, verify=False).content\n    df = pd.read_csv(io.StringIO(dataset.decode('utf-8')), names=[\"time\", \"low\", \"high\", \"open\", \"close\", \"volume\", \"rsi14\", \"ma5\", \"ema5\", \"ema12\", \"ema20\", \"macd\", \"signal\"])\n\n    df.rename(columns={\"close\": str(ratio)+\"_close\", \"volume\": str(ratio) + \"_volume\", \"rsi14\": str(ratio) + \"_rsi14\", \"ma5\": str(ratio) + \"_ma5\", \"ema5\": str(ratio) + \"_ema5\", \"ema12\": str(ratio) + \"_ema12\", \"ema20\": str(ratio) + \"_ema20\", \"macd\": str(ratio) + \"_macd\", \"signal\": str(ratio) + \"_signal\"}, inplace=True)\n\n    df.set_index(\"time\", inplace=True)\n    df = df[[str(ratio) + \"_close\", str(ratio) + \"_volume\", str(ratio) + \"_rsi14\", str(ratio) + \"_ma5\", str(ratio) + \"_ema5\", str(ratio) + \"_ema12\", str(ratio) + \"_ema20\", str(ratio) + \"_macd\", str(ratio) + \"_signal\"]]\n\n    if len(main_df) == 0:\n        main_df = df\n    else:\n        main_df = main_df.join(df)\n\n\nmain_df['future'] = main_df[str(RATIO_TO_PREDICT) + \"_close\"].shift(-FUTURE_PERIOD_PREDICT)\nmain_df['target'] = list(map(classify, main_df[str(RATIO_TO_PREDICT) + \"_close\"], main_df[\"future\"]))\n#print(main_df[[str(RATIO_TO_PREDICT) + \"_close\", \"future\", \"target\"]].head(10))\n\n\ntimes = sorted(main_df.index.values)\nlast_5pct = times[-int(0.05*len(times))]\n\nvalidation_main_df = main_df[(main_df.index &gt;= last_5pct)]\nmain_df = main_df[(main_df.index &lt; last_5pct)]\n\ntrain_x, train_y = preprocess_df(main_df)\nvalidation_x, validation_y = preprocess_df(validation_main_df)\n<\/code><\/pre>\n\n<p>And here are the functions:<\/p>\n\n<pre><code>#Constant Variables\nSEQ_LEN = 30\nFUTURE_PERIOD_PREDICT = 3\nRATIO_TO_PREDICT = \"LTC-USD\"\nEPOCHS = 10\nBATCH_SIZE = 64\nNAME = str(RATIO_TO_PREDICT) + \"-\" + str(SEQ_LEN) + \"-SEQ-\" + str(FUTURE_PERIOD_PREDICT) + \"-PRED-\" + str(int(time.time()))\n\ndef classify(current, future):\n    if float(future) &gt; float(current):\n        return 1\n    else:\n        return 0\n\ndef preprocess_df(df):\n    df = df.drop('future', 1)\n\n    for col in df.columns:\n        if col != \"target\":\n            df[col] = df[col].pct_change()\n            df.dropna(inplace=True)\n            df[col] = preprocessing.scale(df[col].values)\n\n    df.dropna(inplace=True)\n\n    sequential_data = []\n    prev_days = deque(maxlen=SEQ_LEN)\n\n\n\n    for i in df.values:\n        prev_days.append([n for n in i[:-1]])\n        if len(prev_days) == SEQ_LEN:\n            sequential_data.append([np.array(prev_days), i[-1]])\n\n    random.shuffle(sequential_data)\n\n    buys = []\n    sells = []\n\n    for seq, target in sequential_data:\n        if target == 0:\n            sells.append([seq, target])\n        elif target == 1:\n            buys.append([seq, target])\n\n\n    random.shuffle(buys)\n    random.shuffle(sells)\n\n    lower = min(len(buys), len(sells))\n\n\n    buys = buys[:lower]\n    sells = sells[:lower]\n\n\n    sequential_data = buys+sells\n\n    random.shuffle(sequential_data)\n\n    x = []\n    y = []\n\n    for seq, target in sequential_data:\n        x.append(seq)\n        y.append(target)\n\n    return np.array(x), y\n<\/code><\/pre>\n\n<p>Now my question is: After I trained the model, how can I prepare a new CSV file in the model?<\/p>\n","answers":[{"AnswerId":"53366364","CreationDate":"2018-11-18T23:10:53.013","ParentId":null,"OwnerUserId":"10669112","Title":null,"Body":"<p>Typically, to test your model, you would select a subset of your original dataset, and set it aside for testing purposes <em>only<\/em>. That is, you would not use that data for training at all.<\/p>\n\n<p>Now, the link that you're using in your code to fetch CSV files from a remote server does not work for me, but it does have a query param, <code>?method=train<\/code>, which you could presumably change to something like <code>?method=test<\/code> to fetch the testing dataset, and use that for your trial run. Failing that, you could just set aside 20% of your dataset for testing, and use the rest for training.<\/p>\n"}]}
+{"QuestionId":53366458,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2018-11-18T23:24:45.797","AcceptedAnswerId":null,"OwnerUserId":10672115.0,"Title":"Testing Accuracy is less than half of Trainings Accuracy, and manual testing result is way off","Body":"<p>Firstly I have to admit that I am total beginner of PyTorch and CNN image classification. <\/p>\n\n<p>I am making an app to classify cat breeds.<\/p>\n\n<p>The image sets I gathered have around 300-500 per breed, with a total of 62 breeds, plus one other set which represents non-cats which contains 600 samples. I have split the samples into training and testing respectively at 4:1 ratio.<\/p>\n\n<p>The training results are quite disappointing. The training accuracy can reach as much as 90% but the testing results are only 39%.<\/p>\n\n<p>Here are the hyperparameters:<\/p>\n\n<p>LR is 0.1, momentum is 0.1, and batch_size is 128, wideresnet is using 40 layers of widen factor of 10.<\/p>\n\n<p>Please see the source code at:<\/p>\n\n<p><a href=\"https:\/\/github.com\/silver-xu\/wideresnet-trial\" rel=\"nofollow noreferrer\">https:\/\/github.com\/silver-xu\/wideresnet-trial<\/a><\/p>\n\n<p>I have searched all over the internet and pretty much 90% of the articles are about pre-compiled datasets like cifar or MNIST. As a result lots of the codes which I found only is optimised for one type of dataset only.<\/p>\n\n<p>Thanks for all the help! Critics are also Welcome! <\/p>\n\n<p>Here is training output for Epoch 5:<\/p>\n\n<pre><code>Epoch: [5][0\/170]       Time 0.237 (0.237)      Loss 3.3054 (3.3054)    Prec@1 13.281 (13.281)\nEpoch: [5][10\/170]      Time 0.229 (0.228)      Loss 3.2665 (3.3118)    Prec@1 14.844 (13.920)\nEpoch: [5][20\/170]      Time 0.227 (0.227)      Loss 3.0962 (3.2856)    Prec@1 17.969 (14.695)\nEpoch: [5][30\/170]      Time 0.228 (0.227)      Loss 3.3670 (3.2853)    Prec@1 10.938 (14.844)\nEpoch: [5][40\/170]      Time 0.229 (0.227)      Loss 3.3259 (3.2917)    Prec@1 15.625 (15.282)\nEpoch: [5][50\/170]      Time 0.228 (0.227)      Loss 3.2016 (3.2931)    Prec@1 14.844 (14.859)\nEpoch: [5][60\/170]      Time 0.227 (0.227)      Loss 3.3739 (3.3071)    Prec@1 11.719 (14.677)\nEpoch: [5][70\/170]      Time 0.227 (0.227)      Loss 3.4417 (3.3042)    Prec@1 15.625 (14.833)\nEpoch: [5][80\/170]      Time 0.226 (0.227)      Loss 3.2507 (3.2996)    Prec@1 10.938 (14.911)\nEpoch: [5][90\/170]      Time 0.224 (0.227)      Loss 3.2627 (3.2978)    Prec@1 14.844 (15.093)\nEpoch: [5][100\/170]     Time 0.226 (0.227)      Loss 3.3668 (3.2946)    Prec@1 14.062 (15.060)\nEpoch: [5][110\/170]     Time 0.225 (0.227)      Loss 3.2839 (3.2915)    Prec@1 10.156 (14.921)\nEpoch: [5][120\/170]     Time 0.227 (0.227)      Loss 3.3308 (3.2906)    Prec@1 11.719 (14.837)\nEpoch: [5][130\/170]     Time 0.224 (0.227)      Loss 3.1656 (3.2885)    Prec@1 21.875 (14.909)\nEpoch: [5][140\/170]     Time 0.226 (0.227)      Loss 3.2521 (3.2851)    Prec@1 20.312 (14.966)\nEpoch: [5][150\/170]     Time 0.227 (0.227)      Loss 3.1261 (3.2825)    Prec@1 14.844 (14.989)\nEpoch: [5][160\/170]     Time 0.227 (0.227)      Loss 3.4400 (3.2802)    Prec@1 10.938 (15.018)\nTest: [0\/43]    Time 0.262 (0.262)      Loss 3.6978 (3.6978)    Prec@1 8.594 (8.594)\nTest: [10\/43]   Time 0.074 (0.091)      Loss 3.3584 (3.3736)    Prec@1 17.188 (13.139)\nTest: [20\/43]   Time 0.074 (0.083)      Loss 3.3834 (3.4058)    Prec@1 12.500 (12.537)\nTest: [30\/43]   Time 0.074 (0.080)      Loss 3.4457 (3.3994)    Prec@1 14.844 (12.802)\nTest: [40\/43]   Time 0.074 (0.079)      Loss 3.2851 (3.3946)    Prec@1 16.406 (13.281)\n * Prec@1 13.130\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53366491,"AnswerCount":0,"Tags":"<python><tensorflow><multiprocessing>","CreationDate":"2018-11-18T23:29:30.013","AcceptedAnswerId":null,"OwnerUserId":7353970.0,"Title":"Can the TensorFlow's map_and_batch be used with padded batches?","Body":"<p>From TensorFlow's article on <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets\" rel=\"nofollow noreferrer\">Data Input Pipeline Performance<\/a> I can see that in some cases it makes sense to use the <code>map_and_batch<\/code> function. However, I use padded shapes (<code>tf.data.padded_batch()<\/code>) and would like to know if there's still a way to use the suggested <code>map_and_batch<\/code>. The documentation doesn't explain this case.<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":53366667,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><mse>","CreationDate":"2018-11-18T23:56:04.800","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"How to write masked MSE loss in Keras?","Body":"<p>I was trying to write masked MSE loss:<\/p>\n\n<pre><code>def mae_loss_masked(mask):\n    def loss_fn(y_true, y_pred):\n        abs_vec = tf.multiply(tf.abs(y_pred-y_true), mask)\n        loss = tf.reduce_mean(abs_vec)\n        return loss\n    return loss_fn\n<\/code><\/pre>\n\n<p>My model:<\/p>\n\n<pre><code>def MobileNet_v1():\n    # MobileNet with dense layer on top\n\n    # Keras 2.1.6\n    mobilenet = MobileNet(input_shape=(config.IMAGE_H, config.IMAGE_W, config.N_CHANNELS),\n                          alpha=1.0,\n                          depth_multiplier=1,\n                          include_top=False,\n                          weights='imagenet'\n                          )\n\n    x = Flatten()(mobilenet.output)\n    x = Dropout(0.5)(x)\n    x = Dense(config.N_LANDMARKS * 2, activation='linear')(x)\n\n    # -------------------------------------------------------\n\n    model = Model(inputs=mobilenet.input, outputs=x)\n    optimizer = Adadelta()\n    model.compile(optimizer=optimizer, loss=mae_loss_masked)\n\n    model.summary()\n    import sys\n    sys.exit()\n\n    return model\n<\/code><\/pre>\n\n<p>But it give an error:\n<code>TypeError: mae_loss_masked() takes 1 positional argument but 2 were given<\/code><\/p>\n\n<p>Also a question how batch generator output should look like in this case.<\/p>\n","answers":[]}
+{"QuestionId":53366722,"AnswerCount":0,"Tags":"<r><tensorflow><keras><rstudio>","CreationDate":"2018-11-19T00:03:31.527","AcceptedAnswerId":null,"OwnerUserId":1005184.0,"Title":"Keras in R and tensorflow backend","Body":"<p>I need to limit the number of threads running my neural network using these instructions here: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4740\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4740<\/a>. <\/p>\n\n<p>However, I am using keras in R, and I am not sure how do I access the tensorflow implementation used in keras I load in R using <\/p>\n\n<p><code>library(\"keras\")<\/code><\/p>\n\n<p>I can call <code>library(tensorflow)<\/code>, however, isn't it loading a library copy unrelated to the one loaded by keras? And I cannot find any functionality in R that allows to load tensorflow backend associated with keras in Rstudio. Also I cannot find any links to anyone doing the same. <\/p>\n\n<p>Can someone suggest a way to do the operations in the link from R, given keras loaded with <code>library(\"keras\")<\/code> (in the link tensorflow backend for keras is used to set the number of threads per core). It would also be good to know how to check which version is loaded into R by keras.<\/p>\n","answers":[]}
+{"QuestionId":53366768,"AnswerCount":3,"Tags":"<python><tensorflow><keras><tensorflow-serving>","CreationDate":"2018-11-19T00:12:35.490","AcceptedAnswerId":null,"OwnerUserId":10006823.0,"Title":"Serving a Keras model with Tensorflow Serving","Body":"<p>Tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/tag\/v1.12.0\" rel=\"nofollow noreferrer\">1.12 release notes<\/a> states that: <em>\"Keras models can now be directly exported to the SavedModel format(tf.contrib.saved_model.save_keras_model()) and used with Tensorflow Serving\"<\/em>. So I gave it a shot -<\/p>\n\n<p>I have exported a simple model with this op using a single line. However, Tensorflow serving doesn't recognize the model. I guess the problem is with the docker call, and maybe with a missing 'signature_defs' in the model definition. I would be thankful for info regarding the missing steps.<\/p>\n\n<p><strong>1. Training and exporting the model to TF serving<\/strong>:<\/p>\n\n<p>Here is the code based on Jason Brownlee's <a href=\"https:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"nofollow noreferrer\">first NN<\/a> (chosen thanks to its simplicity)<\/p>\n\n<p>(the training data, as a short CSV file, is <a href=\"https:\/\/raw.githubusercontent.com\/jbrownlee\/Datasets\/master\/pima-indians-diabetes.data.csv\" rel=\"nofollow noreferrer\">here<\/a>):<\/p>\n\n<pre><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.contrib.saved_model import save_keras_model\nimport numpy\n\n# fix random seed for reproducibility\nnumpy.random.seed(7)\n\n# load pima indians dataset\ndataset = numpy.loadtxt(\"pima-indians-diabetes.csv\", delimiter=\",\")\n# split into input (X) and output (Y) variables\nX = dataset[:,0:8]\nY = dataset[:,8]\n\n# create model\nmodel = Sequential()\nmodel.add(Dense(12, input_dim=8, activation='relu'))\nmodel.add(Dense(8, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n\n# Compile model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# Fit the model\nmodel.fit(X, Y, epochs=150, batch_size=10)\n\n# evaluate the model\nscores = model.evaluate(X, Y)\nprint(\"\\n%s: %.2f%%\" % (model.metrics_names[1], scores[1]*100))\n\n# calculate predictions\npredictions = model.predict(X)\n# round predictions\nrounded = [round(x[0]) for x in predictions]\nprint(rounded)\n\n# Save the model for serving\npath = '\/TensorFlow_Models\/Keras_serving\/saved_model' # full path of where to save the model\nsave_keras_model(model, path)\n<\/code><\/pre>\n\n<p><strong>2. Setting up Tensorflow Server<\/strong>:<\/p>\n\n<p>The server can be set via docker or by its own build. TF recommends docker (<a href=\"https:\/\/www.tensorflow.org\/serving\/setup\" rel=\"nofollow noreferrer\">TF ref<\/a>). Following this, and based on <a href=\"https:\/\/medium.com\/tensorflow\/serving-ml-quickly-with-tensorflow-serving-and-docker-7df7094aa008\" rel=\"nofollow noreferrer\">TF blog<\/a> and <a href=\"https:\/\/www.tensorflow.org\/serving\/serving_basic#load_exported_model_with_standard_tensorflow_modelserver\" rel=\"nofollow noreferrer\">TF Serving Tutorial<\/a>:<\/p>\n\n<ol>\n<li>Install Docker (from <a href=\"https:\/\/docs.docker.com\/install\/\" rel=\"nofollow noreferrer\">here<\/a>)<\/li>\n<li>Get the latest TF serving version:<\/li>\n<\/ol>\n\n<blockquote>\n  <p>docker pull tensorflow\/serving<\/p>\n<\/blockquote>\n\n<ol start=\"3\">\n<li>Activate TF serving with this model (<a href=\"https:\/\/www.tensorflow.org\/serving\/docker#serving_with_docker\" rel=\"nofollow noreferrer\">TF ref<\/a>):<\/li>\n<\/ol>\n\n<blockquote>\n  <p>docker run -p 8501:8501 --name NNN --mount\n  type=bind,source=SSS,target=TTT -e MODEL_NAME=MMM -t\n  tensorflow\/serving &amp;<\/p>\n<\/blockquote>\n\n<p>I would be happy if one could confirm:<\/p>\n\n<ul>\n<li>NNN - the docker container name - which is used, for instance, to\nkill the process. It can be set arbitrarily (e.g. to: mydocker).<\/li>\n<li>MMM - the name of the model, which seem to be set arbitrarily.<\/li>\n<li>SSS - the folder where the model is located, full path.<\/li>\n<li><strong>TTT - What should be this set to ?<\/strong><\/li>\n<\/ul>\n\n<p><strong>3. the client<\/strong><\/p>\n\n<p>The server can get requests either over gRPC or RESTful API. Assuming we go with RESTful API, the model can be accessed by using curl (<a href=\"https:\/\/www.tensorflow.org\/serving\/docker#serving_example\" rel=\"nofollow noreferrer\">here is a TF example<\/a>). But how do we set the input\/output of the model? does SignatureDefs needed (<a href=\"https:\/\/www.tensorflow.org\/serving\/signature_defs\" rel=\"nofollow noreferrer\">ref<\/a>)?<\/p>\n\n<p><strong>All in all<\/strong>, while <em>\"Keras models can now be directly exported to the SavedModel format(tf.contrib.saved_model.save_keras_model()) and used with Tensorflow Serving\"<\/em>, as stated in TF1.12 release notes, there is a way to go in order to actually serve the model. I would be happy for ideas on completing this.<\/p>\n","answers":[{"AnswerId":"53380590","CreationDate":"2018-11-19T18:25:07.693","ParentId":null,"OwnerUserId":"10675469","Title":null,"Body":"<p>Thanks for your question, more or less is linked with mine \n<a href=\"https:\/\/stackoverflow.com\/questions\/53380386\/tensorflow-serving-signature-for-an-xor\">tensorflow-serving signature for an XOR<\/a><\/p>\n\n<p>I add exactly your doubts about the TTT<\/p>\n"},{"AnswerId":"53405554","CreationDate":"2018-11-21T05:04:53.750","ParentId":null,"OwnerUserId":"7975433","Title":null,"Body":"<p>You are all correct about NNN and SSS. NNN can be arbitrary, if not specified, docker will give it a random name.<\/p>\n\n<p>For MMM, better give it a meaningful name.<\/p>\n\n<p>For TTT this is general about <code>docker run<\/code> command, and you can refer <a href=\"https:\/\/docs.docker.com\/edge\/engine\/reference\/commandline\/run\/#add-bind-mounts-or-volumes-using-the---mount-flag\" rel=\"nofollow noreferrer\">docker doc<\/a>. This is where you map(bind) SSS inside the container, usually set to <code>\/models\/$MODEL_NAME<\/code>. If you get into this container and open <code>\/models\/$MODEL_NAME<\/code>, you will see the version folder(s) just as in SSS.<\/p>\n\n<p>The input of RESTful API is the same as the input to the model in TensorFlow code, in your example is <code>X = dataset[:,0:8]<\/code>.<\/p>\n\n<p>If you didn't define signature when saving the model like <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/example\/mnist_saved_model.py#L100\" rel=\"nofollow noreferrer\">the example in doc<\/a>, then it's not necessary in serving.<\/p>\n"},{"AnswerId":"58965756","CreationDate":"2019-11-21T01:08:45.710","ParentId":null,"OwnerUserId":"3286035","Title":null,"Body":"<p>If you want to create an API endpoint server with Keras out of the box.  I recommend you to check out BentoML(<a href=\"https:\/\/github.com\/bentoml\/bentoml\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bentoml\/bentoml<\/a>). <\/p>\n\n<p>It is an open source python library to make it easy to serve and deploy machine learning models in the cloud. \n<em>disclosure: I am one of the creator of BentoML<\/em>.<\/p>\n\n<p>Here is an exammple of Keras model using BentoML (<a href=\"https:\/\/github.com\/bentoml\/gallery\/blob\/master\/keras\/fashion-mnist\/keras-fashion-mnist.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bentoml\/gallery\/blob\/master\/keras\/fashion-mnist\/keras-fashion-mnist.ipynb<\/a>)<\/p>\n\n<p>With few line of codes, you could package your model, preprocessing code and dependencies into a standardized format that could distribute to different formats (Docker, CLI tool, Spark UDF, and etc) and deploy them to different cloud platforms(Sagemaker, Lambda, Kubernetes)<\/p>\n\n<p>After you finished training your model in Jupyter notebook. You can add BentoService spec to create your model in standardized bundle format<\/p>\n\n<pre><code>%%writefile my_model.py\n\nfrom bentoml import api, artifacts, env, BentoService\nfrom bentoml.artifact import KerasModelArtifact\nfrom bentoml.handlers import DataframeHandler\n\n@env(pip_dependencies=['keras', 'tensorflow==1.14.0'])\n@artifacts([KerasModelArtifact('model')])\nclass KerasModelService(BentoService):\n\n    @api(DataframHandler)\n    def predict(self, df):\n        return self.artifacts.model.predict(df)\n<\/code><\/pre>\n\n<p>In the next cell you could pack your model and save it.<\/p>\n\n<pre><code># 1) import the custom BentoService defined above\nfrom my_model import KerasModelService\n\n# 2) `pack` it with required artifacts\nkeras_svc = KerasModelService.pack(model=model)\n\n# 3) save your BentoSerivce to file archive\nsaved_path = keras_svc.save()\n<\/code><\/pre>\n\n<p>These will packaged your model, preprocessing code, dependencies and also generate Dockerfile and other for you.<\/p>\n\n<p>With this bundle, you could easily start a docker container with REST API endpoints.<\/p>\n\n<pre><code>cd saved_path &amp;&amp; docker build -t keras-model .\n<\/code><\/pre>\n\n<p>Feel free to ping me, if you have any questions.<\/p>\n"}]}
+{"QuestionId":53366921,"AnswerCount":1,"Tags":"<tensorflow><tensorflowjs-converter>","CreationDate":"2018-11-19T00:38:44.693","AcceptedAnswerId":53367421.0,"OwnerUserId":481649.0,"Title":"tensorflowjs_converter: SavedModel file does not exist at:","Body":"<p>I am following the instructions <a href=\"https:\/\/js.tensorflow.org\/tutorials\/import-saved-model.html\" rel=\"nofollow noreferrer\">here<\/a> to import my Tensorflow model generated using Python into Tensorflow.js.<\/p>\n\n<p>The model is created using a <a href=\"https:\/\/www.tensorflow.org\/guide\/estimators\" rel=\"nofollow noreferrer\"><code>tf.estimator<\/code> API<\/a>, which auto saves the model files according to <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#overview_of_saving_and_restoring_models\" rel=\"nofollow noreferrer\">this page<\/a>.<\/p>\n\n<p>After installing <code>tensorflowjs<\/code> using pip, I get the following error when running the convert script:<\/p>\n\n<pre><code>$ tensorflowjs_converter     --input_format=tf_saved_model     --output_node_names='DeepWater'     \/home\/sterengiannini\/deepwater\/model-js-import-test     \/home\/sterengiannini\/deepwater\/web-model\nUsing TensorFlow backend.\nTraceback (most recent call last):\n  File \"\/usr\/local\/bin\/tensorflowjs_converter\", line 11, in &lt;module&gt;\n    sys.exit(main())\n  File \"\/home\/sterengiannini\/.local\/lib\/python2.7\/site-packages\/tensorflowjs\/converters\/converter.py\", line 317, in main\n    strip_debug_ops=FLAGS.strip_debug_ops)\n  File \"\/home\/sterengiannini\/.local\/lib\/python2.7\/site-packages\/tensorflowjs\/converters\/tf_saved_model_conversion.py\", line 289, in convert_tf_saved_model\n    input_saved_model_dir=saved_model_dir)\n  File \"\/home\/sterengiannini\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 338, in freeze_graph\n    input_saved_model_dir, saved_model_tags).graph_def\n  File \"\/home\/sterengiannini\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/tools\/saved_model_utils.py\", line 42, in get_meta_graph_def\n    saved_model = reader.read_saved_model(saved_model_dir)\n  File \"\/home\/sterengiannini\/.local\/lib\/python2.7\/site-packages\/tensorflow\/contrib\/saved_model\/python\/saved_model\/reader.py\", line 55, in read_saved_model\n    raise IOError(\"SavedModel file does not exist at: %s\" % saved_model_dir)\nIOError: SavedModel file does not exist at: \/home\/sterengiannini\/deepwater\/model-js-import-test\n<\/code><\/pre>\n\n<p>Here are the files present in the model folder:<\/p>\n\n<pre><code>$ ls model-js-import-test\/\ngraph.pbtxt  model.ckpt-312.data-00000-of-00001  model.ckpt-312.index  model.ckpt-312.meta\n<\/code><\/pre>\n","answers":[{"AnswerId":"53367421","CreationDate":"2018-11-19T02:11:31.197","ParentId":null,"OwnerUserId":"10672427","Title":null,"Body":"<p>It looks like the model was not saved in the right format. You should expect something like the following in your model folder (<a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#structure_of_a_savedmodel_directory\" rel=\"nofollow noreferrer\">see Structure of a SavedModel directory<\/a>):<\/p>\n\n<pre><code>assets\/\nassets.extra\/\nvariables\/\n    variables.data-?????-of-?????\n    variables.index\nsaved_model.pb|saved_model.pbtxt\n<\/code><\/pre>\n\n<p>For, example, if you are using tf.estimator API, you need to export the model with the following (<a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#using_savedmodel_with_estimators\" rel=\"nofollow noreferrer\">see Using SavedModel with Estimators<\/a>): <\/p>\n\n<pre><code>estimator.export_savedmodel(export_dir_base, serving_input_receiver_fn)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53367049,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-19T00:59:39.987","AcceptedAnswerId":53368778.0,"OwnerUserId":2686980.0,"Title":"F.relu(self.fc1(x)) is causing RuntimeError problem","Body":"<p>I have implemented the following CNN for my training and validation data sets\nthat contain 90 and 20 images respectively divided into 3 classes:<\/p>\n\n<pre><code>def __init__(self):\n    super(Net, self).__init__()\n    self.conv1 = nn.Conv2d(3, 6, 5)\n    self.pool = nn.MaxPool2d(2, 2)\n    self.conv2 = nn.Conv2d(6, 16, 5)\n    self.fc1 = nn.Linear(16 * 5 * 5, 120)\n    self.fc2 = nn.Linear(120, 50)\n    self.fc3 = nn.Linear(50, len(classes))\n\n\ndef forward(self, x):\n    print(x.shape)\n    x = self.pool(F.relu(self.conv1(x)))\n    print(x.shape)\n    x = self.pool(F.relu(self.conv2(x)))\n    print(x.shape)\n    x = x.view(x.size(0),-1)\n        #x = x.view(-1,x.size(1)*x.size(2)*x.size(3))\n        #x = x.view(-1, 16 * 5 * 5)\n    x = F.relu(self.fc1(x))\n    x = F.relu(self.fc2(x))\n    x = self.fc3(x)\n    return x\n<\/code><\/pre>\n\n<p>I am getting the following error while running:<\/p>\n\n<pre><code>RuntimeError: size mismatch, m1: [1 x 214720], m2: [400 x 120] at \/opt\/conda\/conda-bld\/pytorch_1524584710464\/work\/aten\/src\/TH\/generic\/THTensorMath.c:2033 for x = F.relu(self.fc1(x))\n<\/code><\/pre>\n\n<p>Can someone please suggest what I should do to get rid of this problem? I changed the x.view(...) by following some threads. However, it did not help in this case.<\/p>\n","answers":[{"AnswerId":"53368778","CreationDate":"2018-11-19T05:27:47.700","ParentId":null,"OwnerUserId":"5373738","Title":null,"Body":"<p>The size of the <code>in_channels<\/code> to <code>self.fc1<\/code> is dependent on the input image size and not on the kernel-size.<\/p>\n\n<p>In your case, \n<code>self.fc1 = nn.Linear(16 * 5 * 5, 120)<\/code> should be <code>nn.Linear(16 * image_size * image_size)<\/code>\nwhere, <code>image_size<\/code>: is the size of the image in the last convolution layer.<\/p>\n\n<p>Sample Code:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport numpy as np\n\nclass Net(nn.Module):\n    def __init__(self, classes):\n        super().__init__()\n        self.conv1 = nn.Conv2d(3, 6, 5, padding=2)\n        self.pool = nn.MaxPool2d(2, 2)\n        self.conv2 = nn.Conv2d(6, 16, 5, padding=2)\n        self.fc1 = nn.Linear(16 * 25 * 25, 120)\n        self.fc2 = nn.Linear(120, 50)\n        self.fc3 = nn.Linear(50, classes)\n\n    def forward(self, x):\n        print('one', x.shape)\n        x = self.pool(F.relu(self.conv1(x)))\n        print('two', x.shape)\n        x = self.pool(F.relu(self.conv2(x)))\n        print('three', x.shape)\n        x = x.view(-1, np.product(x.shape[1:]))\n        x = F.relu(self.fc1(x))\n        x = F.relu(self.fc2(x))\n        x = self.fc3(x)\n        return x\n\nx = torch.rand((32, 3, 100, 100))\nnet = Net(2)\nout= net(x)\nprint('out', out.shape)\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>one torch.Size([32, 3, 100, 100])\ntwo torch.Size([32, 6, 50, 50])\nthree torch.Size([32, 16, 25, 25])\nout torch.Size([32, 2])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53367059,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-19T01:01:33.817","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"How to zero pad dense layer in Keras?","Body":"<p>I can see there is <code>ZeroPadding1D<\/code> in Keras <a href=\"https:\/\/keras.io\/layers\/convolutional\/#zeropadding1d\" rel=\"nofollow noreferrer\">[doc]<\/a>, but it require <code>3D tensor with shape (batch, axis_to_pad, features)<\/code>, but how to zero pad dense layer output on the right side with shape <code>(batch, features)<\/code>?<\/p>\n\n<p>Something like:<\/p>\n\n<pre><code>x = Dense(64, activation='linear')(x)\nx = ZeroPadding1D(padding=(0,64))(x)\n<\/code><\/pre>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>I have tried to implement custom layer:<\/p>\n\n<pre><code>class CustomZeroPadding1D(Layer):\n    def __init__(self, **kwargs):\n        super(CustomZeroPadding1D, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        super(CustomZeroPadding1D, self).build(input_shape)\n\n    def call(self, x):\n        res = concatenate([x, K.zeros_like(x)], axis=-1)\n        return res\n\n    def compute_output_shape(self, input_shape):\n        print('-'*60)\n        print('input_shape', input_shape)\n        output_shape = (input_shape[0], input_shape[1]*2)\n        print('output_shape', output_shape)\n        print('-' * 60)\n        return output_shape\n<\/code><\/pre>\n\n<p>Shapes in <code>model.summary()<\/code> looks ok, but when I run training it fails with <code>Incompatible shapes: [32,312] vs. [0,312]<\/code>, where batch_size=32 and n_features=312, so seems for some reason custom layer set batch size = 0?<\/p>\n","answers":[{"AnswerId":"53367347","CreationDate":"2018-11-19T01:57:51.070","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You can do it with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\">tensorflow.reshape()<\/a>:<\/p>\n\n<pre><code>x = Dense(64, activation='linear')(x)\nx = Reshape([-1, 1])(x)\nx = ZeroPadding1D(padding=(0,64))(x)\n<\/code><\/pre>\n\n<p><strong>A simple example:<\/strong><\/p>\n\n<pre><code>from keras.layers import Dense, ZeroPadding1D, Reshape\nimport tensorflow as tf\n\nx = tf.constant([ [1.,1.,1.,1.], [1.,1.,1.,1.], [1.,1.,1.,1.] ])\n\nx = Dense(64, activation='linear')(x)\nprint(x.shape)\nx = Reshape([-1, 1])(x)\nprint(x.shape)\nx = ZeroPadding1D(padding=(0,10))(x)\nprint(x.shape)\nx_ = x\nx = Reshape([-1])(x)\nprint(x.shape)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(x_.eval()[0])\n    print(x.eval()[0])\n<\/code><\/pre>\n\n<p><strong>Out:<\/strong><\/p>\n\n<pre><code>(3, 64)\n(3, 64, 1)\n(3, 74, 1)\n(3, 74)\n[[-0.4807737 ]\n [-0.11104995]\n [-0.13179883]\n [-0.34604812]\n [-0.51428366]\n [ 0.39941764]\n [-0.0830393 ]\n [-0.79970515]\n [-0.34765747]\n [ 0.6095309 ]\n [ 0.03473596]\n [-0.68571013]\n [-0.12576953]\n [ 0.08276424]\n [ 0.06647275]\n [ 0.1670239 ]\n [-0.26894042]\n [ 0.03662822]\n [ 0.24533364]\n [ 0.2816307 ]\n [-0.28530025]\n [ 0.33335078]\n [-0.52831376]\n [-0.5450369 ]\n [-0.30863497]\n [ 0.14870909]\n [ 0.4303183 ]\n [-0.11658342]\n [ 0.60449684]\n [-0.47217163]\n [-0.3101738 ]\n [-0.15606529]\n [ 0.25018048]\n [-0.34411854]\n [-0.6233641 ]\n [ 0.01476687]\n [-0.32950908]\n [ 0.40554196]\n [ 0.2916515 ]\n [-0.5265654 ]\n [-0.13000801]\n [ 0.45457274]\n [-0.32708472]\n [ 0.20291099]\n [ 0.15016158]\n [ 0.02729714]\n [ 0.33809263]\n [-0.67841053]\n [ 0.31094086]\n [-0.3722076 ]\n [ 0.31136334]\n [ 0.21413101]\n [-0.40144968]\n [ 0.37131637]\n [ 0.17351764]\n [ 0.1576828 ]\n [-0.299753  ]\n [ 0.32608157]\n [-0.15042162]\n [-0.2388339 ]\n [ 0.18553   ]\n [-0.4828058 ]\n [ 0.07377535]\n [-0.291501  ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]\n [ 0.        ]]\n\n[-0.4807737  -0.11104995 -0.13179883 -0.34604812 -0.51428366  0.39941764\n -0.0830393  -0.79970515 -0.34765747  0.6095309   0.03473596 -0.68571013\n -0.12576953  0.08276424  0.06647275  0.1670239  -0.26894042  0.03662822\n  0.24533364  0.2816307  -0.28530025  0.33335078 -0.52831376 -0.5450369\n -0.30863497  0.14870909  0.4303183  -0.11658342  0.60449684 -0.47217163\n -0.3101738  -0.15606529  0.25018048 -0.34411854 -0.6233641   0.01476687\n -0.32950908  0.40554196  0.2916515  -0.5265654  -0.13000801  0.45457274\n -0.32708472  0.20291099  0.15016158  0.02729714  0.33809263 -0.67841053\n  0.31094086 -0.3722076   0.31136334  0.21413101 -0.40144968  0.37131637\n  0.17351764  0.1576828  -0.299753    0.32608157 -0.15042162 -0.2388339\n  0.18553    -0.4828058   0.07377535 -0.291501    0.          0.\n  0.          0.          0.          0.          0.          0.\n  0.          0.        ]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53367063,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T01:02:04.150","AcceptedAnswerId":null,"OwnerUserId":10672298.0,"Title":"tensorflow python expected dense_input to have 2 dimensions, but got array with shape (5, 28, 5)","Body":"<p>I am a complete newbie to tensorflow, trying to learn about it and solve a problem.  I tried a lot of tutorials but they all talked about the same classify image or mnist stuff, so I followed the documentation and tried to figure something out.  <\/p>\n\n<p>The goal is to find a pattern to predict the result when the input is [[1000,10, 5, 3, 1744...etc.  There are only 5 cases when the value is 300 400, 500, 600, 700, with shape 28,5 and the result for each is 28,2 list.  The data is loaded from file and assigned to tf.tensor.  <\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><code>model = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Dense(28, activation=tf.nn.relu, input_shape=(5,)))\nmodel.add(tf.keras.layers.Dense(28, activation=tf.nn.relu, input_shape=(5,)))\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(28))\n\nmodel.compile(optimizer='adam',\n              loss='mean_squared_error',\n              metrics=['accuracy'])\n\nmodel.fit(newData, newResults, epochs=3, steps_per_epoch=5)\n<\/code><\/pre>\n\n<p>newData:<\/p>\n\n<pre><code>[[[300, 10, 5, 3, 1744], [300, 10, 5, 5, 2848], [300, 10, 5, 4, 2418], [300, 10, 5, 2, 1152], [300, 10, 5, 3, 1126], [300, 10, 5, 3, 1897], [300, 10, 5, 3, 1089], [300, 10, 5, 2, 1581], [300, 10, 5, 4, 1793], [300, 10, 5, 3, 1525], [300, 10, 5, 2, 1529], [300, 10, 5, 3, 1052], [300, 10, 5, 2, 1556], [300, 10, 5, 3, 1569], [300, 10, 5, 5, 2873], [300, 10, 5, 4, 2269], [300, 10, 5, 3, 3003], [300, 10, 5, 3, 1310], [300, 10, 5, 3, 1464], [300, 10, 5, 3, 2807], [300, 10, 5, 2, 1262], [300, 10, 5, 3, 1734], [300, 10, 5, 2, 2709], [300, 10, 5, 3, 2234], [300, 10, 5, 3, 1961], [300, 10, 5, 2, 1594], [300, 10, 5, 2, 1836], [300, 10, 5, 2, 1345]], \n[[400, 10, 5, 3, 1744], [400, 10, 5, 5, 2848], [400, 10, 5, 4, 2418], [400, 10, 5, 2, 1152], [400, 10, 5, 3, 1126], [400, 10, 5, 3, 1897], [400, 10, 5, 3, 1089], [400, 10, 5, 2, 1581], [400, 10, 5, 4, 1793], [400, 10, 5, 3, 1525], [400, 10, 5, 2, 1529], [400, 10, 5, 3, 1052], [400, 10, 5, 2, 1556], [400, 10, 5, 3, 1569], [400, 10, 5, 5, 2873], [400, 10, 5, 4, 2269], [400, 10, 5, 3, 3003], [400, 10, 5, 3, 1310], [400, 10, 5, 3, 1464], [400, 10, 5, 3, 2807], [400, 10, 5, 2, 1262], [400, 10, 5, 3, 1734], [400, 10, 5, 2, 2709], [400, 10, 5, 3, 2234], [400, 10, 5, 3, 1961], [400, 10, 5, 2, 1594], [400, 10, 5, 2, 1836], [400, 10, 5, 2, 1345]], \n[[500, 10, 5, 3, 1744], [500, 10, 5, 5, 2848], [500, 10, 5, 4, 2418], [500, 10, 5, 2, 1152], [500, 10, 5, 3, 1126], [500, 10, 5, 3, 1897], [500, 10, 5, 3, 1089], [500, 10, 5, 2, 1581], [500, 10, 5, 4, 1793], [500, 10, 5, 3, 1525], [500, 10, 5, 2, 1529], [500, 10, 5, 3, 1052], [500, 10, 5, 2, 1556], [500, 10, 5, 3, 1569], [500, 10, 5, 5, 2873], [500, 10, 5, 4, 2269], [500, 10, 5, 3, 3003], [500, 10, 5, 3, 1310], [500, 10, 5, 3, 1464], [500, 10, 5, 3, 2807], [500, 10, 5, 2, 1262], [500, 10, 5, 3, 1734], [500, 10, 5, 2, 2709], [500, 10, 5, 3, 2234], [500, 10, 5, 3, 1961], [500, 10, 5, 2, 1594], [500, 10, 5, 2, 1836], [500, 10, 5, 2, 1345]], \n[[600, 10, 5, 3, 1744], [600, 10, 5, 5, 2848], [600, 10, 5, 4, 2418], [600, 10, 5, 2, 1152], [600, 10, 5, 3, 1126], [600, 10, 5, 3, 1897], [600, 10, 5, 3, 1089], [600, 10, 5, 2, 1581], [600, 10, 5, 4, 1793], [600, 10, 5, 3, 1525], [600, 10, 5, 2, 1529], [600, 10, 5, 3, 1052], [600, 10, 5, 2, 1556], [600, 10, 5, 3, 1569], [600, 10, 5, 5, 2873], [600, 10, 5, 4, 2269], [600, 10, 5, 3, 3003], [600, 10, 5, 3, 1310], [600, 10, 5, 3, 1464], [600, 10, 5, 3, 2807], [600, 10, 5, 2, 1262], [600, 10, 5, 3, 1734], [600, 10, 5, 2, 2709], [600, 10, 5, 3, 2234], [600, 10, 5, 3, 1961], [600, 10, 5, 2, 1594], [600, 10, 5, 2, 1836], [600, 10, 5, 2, 1345]], \n[[700, 10, 5, 3, 1744], [700, 10, 5, 5, 2848], [700, 10, 5, 4, 2418], [700, 10, 5, 2, 1152], [700, 10, 5, 3, 1126], [700, 10, 5, 3, 1897], [700, 10, 5, 3, 1089], [700, 10, 5, 2, 1581], [700, 10, 5, 4, 1793], [700, 10, 5, 3, 1525], [700, 10, 5, 2, 1529], [700, 10, 5, 3, 1052], [700, 10, 5, 2, 1556], [700, 10, 5, 3, 1569], [700, 10, 5, 5, 2873], [700, 10, 5, 4, 2269], [700, 10, 5, 3, 3003], [700, 10, 5, 3, 1310], [700, 10, 5, 3, 1464], [700, 10, 5, 3, 2807], [700, 10, 5, 2, 1262], [700, 10, 5, 3, 1734], [700, 10, 5, 2, 2709], [700, 10, 5, 3, 2234], [700, 10, 5, 3, 1961], [700, 10, 5, 2, 1594], [700, 10, 5, 2, 1836], [700, 10, 5, 2, 1345]]]\n<\/code><\/pre>\n\n<p>newResult:<\/p>\n\n<pre><code>[[[29.0, 8.92], [52.0, 21.67], [41.0, 14.38], [7.0, 1.49], [26.0, 8.25], [18.0, 4.53], [24.0, 6.61], [21.0, 9.54], [17.0, 5.53], [27.0, 9.61], [11.0, 0.35], [22.0, 8.11], [7.0, 1.22], [36.0, 15.49], [57.0, 31.44], [43.0, 16.52], [34.0, 11.46], [15.0, 2.49], [20.0, 2.34], [16.0, 4.86], [10.0, 0.8], [8.0, 0.4], [1.0, 0.0], [30.0, 7.57], [24.0, 7.21], [5.0, 0.58], [14.0, 0.73], [4.0, 0.15]], \n[[45.0, 8.17], [100.0, 43.28], [54.0, 16.05], [10.0, 2.77], [37.0, 8.86], [27.0, 6.12], [33.0, 9.13], [34.0, 14.03], [20.0, 5.06], [45.0, 15.42], [21.0, 0.69], [26.0, 8.83], [11.0, 2.14], [44.0, 17.74], [73.0, 43.39], [43.0, 18.8], [46.0, 21.56], [29.0, 9.16], [21.0, 3.76], [20.0, 7.39], [16.0, 2.54], [1.0, 1.63], [1.0, 0.02], [28.0, 12.14], [30.0, 12.35], [7.0, 1.18], [19.0, 3.29], [4.0, 0.16]], \n[[59.0, 18.74], [100.0, 75.18], [69.0, 32.13], [11.0, 3.04], [49.0, 15.76], [30.0, 10.33], [45.0, 14.51], [43.0, 20.82], [37.0, 8.2], [69.0, 24.53], [1.0, 0.3], [38.0, 12.57], [1.0, 3.67], [65.0, 24.77], [91.0, 57.39], [53.0, 18.22], [47.0, 27.07], [34.0, 16.31], [25.0, 5.39], [31.0, 11.5], [23.0, 5.73], [19.0, 4.11], [2.0, 0.11], [35.0, 15.52], [41.0, 18.15], [7.0, 1.48], [25.0, 7.53], [3.0, 0.14]], \n[[80.0, 30.29], [100.0, 85.22], [94.0, 52.73], [11.0, 2.45], [72.0, 30.7], [46.0, 14.75], [70.0, 22.81], [50.0, 28.26], [40.0, 14.19], [60.0, 26.82], [14.0, 0.28], [45.0, 19.1], [16.0, 4.72], [82.0, 40.98], [100.0, 78.96], [66.0, 27.05], [67.0, 31.09], [34.0, 16.92], [23.0, 7.03], [48.0, 21.28], [27.0, 8.19], [21.0, 3.95], [2.0, 0.17], [43.0, 19.96], [55.0, 23.54], [8.0, 1.47], [28.0, 12.04], [4.0, 0.13]], \n[[95.0, 38.09], [100.0, 92.88], [99.0, 58.96], [13.0, 3.54], [96.0, 45.78], [33.0, 12.05], [87.0, 38.11], [62.0, 34.97], [48.0, 15.49], [84.0, 33.13], [10.0, 0.09], [63.0, 25.52], [16.0, 4.87], [100.0, 55.9], [100.0, 91.32], [90.0, 34.24], [96.0, 45.36], [37.0, 15.13], [27.0, 9.28], [49.0, 26.3], [30.0, 10.92], [22.0, 3.72], [3.0, 0.14], [67.0, 24.82], [73.0, 31.32], [8.0, 1.36], [31.0, 15.03], [4.0, 0.2]]]\n<\/code><\/pre>\n\n<p>Getting this error when I run it:<\/p>\n\n<pre><code>  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1536, in fit\n    validation_split=validation_split)\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 992, in _standardize_user_data\n    class_weight, batch_size)\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1117, in _standardize_weights\n    exception_prefix='input')\n  File \"C:\\Program Files\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py\", line 323, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking input: expected dense_input to have 2 dimensions, but got array with shape (5, 28, 5)\n<\/code><\/pre>\n\n<p>I know my model definitely has something wrong with it, but I can't quite figure out what.  I have trouble finding information other than the afore mentioned examples.<\/p>\n","answers":[{"AnswerId":"53367188","CreationDate":"2018-11-19T01:27:26.197","ParentId":null,"OwnerUserId":"8489404","Title":null,"Body":"<p>The code you've written starts with a Dense layer of 28 neurons expecting an array of shape <code>(*, 5)<\/code>. The error message shows that the shape it is receiving is `(5, 28, 5).<\/p>\n\n<p>Looking at your data, it is shape <code>(5, 28, 5)<\/code>. Maybe you have an extra set of brackets?<\/p>\n"}]}
+{"QuestionId":53367354,"AnswerCount":1,"Tags":"<python><keras><deep-learning><lstm><gesture-recognition>","CreationDate":"2018-11-19T01:59:00.243","AcceptedAnswerId":53367864.0,"OwnerUserId":3481696.0,"Title":"Unable to improve validation accuracy though training is 100% and how to feed images extracted from videos to deep learning model","Body":"<p>I am building a gesture recognition system using 20bn's jester dataset. Currently, I am working on only 4 classes. Dataset consists of images extracted from videos at 12 frames\/sec rate.  I built 2 models, 3D-CNN and CNN-LSTM  but getting accuracy around 25-30% only using keras and Tensorflow.<\/p>\n\n<pre><code>Dataset looks like X_train = (651, 1, 128, 128, 22)\nX_valid=(260, 1, 128, 128, 22)\nY_train=(651, 4)\nY_valid = (260, 4)\n<\/code><\/pre>\n\n<p>Image size is 128*128 with 1 channel, I am appending 22 images at a time with 651 as a total number of samples.\n3d-CNN architecture<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Convolution3D(32, (3, 3, 3), strides=(1, 1, 1), input_shape=(1, img_rows, img_cols, img_depth),\n                            activation='relu', data_format='channels_first'))\n    model.add(MaxPooling3D((3, 3, 3), data_format='channels_first'))\n    model.add(Dropout(0.5))\n    model.add(Flatten())\n    model.add(Dense(512, activation='sigmoid'))\n    model.add(Dropout(0.5))\n    model.add(Dense(nb_classes, kernel_initializer='normal'))\n\n    model.add(Activation('softmax'))\n<\/code><\/pre>\n\n<p>CNN-LSTM model architecture<\/p>\n\n<pre><code>(651, 22, 128, 128, 1)\n(260, 22, 128, 128, 1)\n(651, 4)\n(260, 4)\n    model = Sequential()\n\n    model.add(TimeDistributed(Conv2D(32, (7, 7), strides=(2, 2),\n                                     activation='relu', padding='same'), input_shape=input_shape))\n    model.add(TimeDistributed(Conv2D(32, (3, 3),\n                                     kernel_initializer=\"he_normal\", activation='relu')))\n    model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))\n\n\n    model.add(TimeDistributed(Conv2D(64, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(Conv2D(64, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))\n\n\n    model.add(TimeDistributed(Conv2D(128, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(Conv2D(128, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))\n\n\n    model.add(TimeDistributed(Conv2D(256, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(Conv2D(256, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))\n\n\n    model.add(TimeDistributed(Conv2D(512, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(Conv2D(512, (3, 3),\n                                     padding='same', activation='relu')))\n    model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))\n\n\n    model.add(TimeDistributed(Flatten()))\n\n    model.add(Dropout(0.5))\n    model.add(LSTM(256, return_sequences=False, dropout=0.5))\n    model.add(Dense(nb_classes, activation='softmax'))\n<\/code><\/pre>\n\n<p>Dataset is balanced with 100 gestures from each class and I ran the model for 100 epoch. Is the architecture correct? I also have doubts regarding feeding the images. As I am working on videos so I need an extra dimension to extract the temporal features so appended 22 images in 1 array first and kept on appending similarly, is it the right way to do? First I tried with 12 also instead of 22 but same result.<\/p>\n","answers":[{"AnswerId":"53367864","CreationDate":"2018-11-19T03:19:40.437","ParentId":null,"OwnerUserId":"868736","Title":null,"Body":"<p>Below are a few suggestions<\/p>\n\n<ul>\n<li>your training n is 400, I 'feel' like this n is too small, i recommend increasing the training dataset to maybe like 500*4. remember to shuffle the data when writing the training data and enable random shuffle per batch during training.<\/li>\n<li>is the data normalized (from 0 to 1, or zero mean unit variance)?<\/li>\n<li>are you sure each sample have 22 images in the training dataset, and you zero pad in the time channel for the samples if they do not have 22 frames? <\/li>\n<li>have you tried varying optimizers, learning rates?<\/li>\n<li>for both architecture, the bottle neck segment, I would recommend at least 2 layers of dense layers.<\/li>\n<li>for 3D-CNN, i recommend adding a few more cnn layers, like what you did with the cnn-lstm.<\/li>\n<li>for both architectures, beside the last output layer, consistently use the same activation 'relu', no need to mix and match (your CNN-LSTM have a sigmoid in the dense layer)<\/li>\n<li>for both architectures, reduce dropout probability to maybe 0.2 (if you are using keras).<\/li>\n<li>if above changes are not helping, add in batch normalization for each layer.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53367724,"AnswerCount":1,"Tags":"<python><tensorflow><pytorch><mxnet>","CreationDate":"2018-11-19T02:59:07.817","AcceptedAnswerId":null,"OwnerUserId":5157515.0,"Title":"Theoretical underpinning behind Hardmax operator","Body":"<p>In the tensor flow Github repository, in the file <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c19e29306ce1777456b2dbb3a14f511edf7883a8\/tensorflow\/contrib\/seq2seq\/python\/ops\/attention_wrapper.py#L1025\" rel=\"nofollow noreferrer\">attentionwrapper.py<\/a>, hardmax operator has been defined. On the docs, it has been mentioned <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/seq2seq\/hardmax\" rel=\"nofollow noreferrer\">tf.contrib.seq2seq.hardmax<\/a><\/p>\n\n<p>I want to know what's the theoretical underpinning behind providing this functionality for hardmax operator. Prima facie google searches for past few weeks haven't led me to concrete understanding of the concept. <\/p>\n\n<ol>\n<li><p>If softmax is differentiable (soft), why would hardmax be ever used? If it can't be used in back propagation (due to non-differentiability required in gradient calculation), where else can it be used?<\/p><\/li>\n<li><p>Reinforcement learning literature talks about Soft vs Hard attention. However I couldn't see concrete examples nor explanations of where the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/seq2seq\/hardmax\" rel=\"nofollow noreferrer\">tf.contrib.seq2seq.hardmax<\/a> can be actually used in some RL model.<\/p><\/li>\n<li><p>By the looks of it, since it is mentioned in seq2seq, it should be obviously having some application in Natural Language Processing. But exactly where? There are tonnes of NLP tasks. Couldn't find any direct task SOTA algorithm which uses hardmax.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53367768","CreationDate":"2018-11-19T03:05:38.067","ParentId":null,"OwnerUserId":"8083699","Title":null,"Body":"<p>Hardmax is used when you have no choice but to make a decision nonprobabalistically. For example, when you are using a model to generate a neural architecture as in neural module networks, you have to make a discrete choice. To make this trainable (since this would be non-differentiable as you state), you can use REINFORCE (an algorithm in RL) to train via policy gradient and estimate this loss contribution via Monte Carlo sampling. Neural module networks are an NLP construct and depend on seq2seq. I'm sure there are many examples, but this is one that immediately came to mind.<\/p>\n"}]}
+{"QuestionId":53367734,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T02:59:53.123","AcceptedAnswerId":null,"OwnerUserId":10631580.0,"Title":"How to implement average pooling in case of Conv1d in tensorflow?","Body":"<p>I want to implement the average pooling in conv1d. But <code>tf.nn.avg_pool<\/code> function can only be implemented on 4 dimensional tensor. So what should I do to overcome this problem? <\/p>\n\n<pre><code>def avg_pool(conv_out):\n    return tf.nn.avg_pool(conv_out,ksize=[1,1,2,1],strides=[1,1,2,1],padding='SAME')\n\ni = tf.constant([1, 0, 2, 3, 0, 1], dtype=tf.float32)\n\ndata   = tf.reshape(i, [1, int(i.shape[0]), 1], name='data')\n\nkernel = tf.Variable(tf.random_normal([2,1,1]))\n\nconv_out = tf.nn.conv1d(data, kernel, 2, 'VALID')\npool_out = avg_pool(conv_out)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53373859","CreationDate":"2018-11-19T11:39:47.160","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>One option is to add an additional dimension to your data and then remove it:<\/p>\n\n<pre><code>def avg_pool(conv_out):\n    conv_out_2d = conv_out[:, tf.newaxis]\n    pool_out_2d = tf.nn.avg_pool(conv_out_2d,\n                                 ksize=[1, 1, 2, 1],\n                                 strides=[1, 1, 2, 1],\n                                 padding='SAME')\n    pool_out = pool_out_2d[:, 0]\n    return pool_out\n<\/code><\/pre>\n\n<p>Another possibility is to use the generic <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/pool\" rel=\"nofollow noreferrer\"><code>tf.nn.pool<\/code><\/a>:<\/p>\n\n<pre><code>def avg_pool(conv_out):\n    return tf.nn.pool(conv_out, window_shape=[2], pooling_type='AVG', padding='SAME')\n<\/code><\/pre>\n\n<p>Note in this case I am not including the stride because the default value matches what you used for your example, but you can modify it too if you want.<\/p>\n"}]}
+{"QuestionId":53367825,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2018-11-19T03:14:50.383","AcceptedAnswerId":null,"OwnerUserId":8018636.0,"Title":"Parts of tensorflow codes cannot be directed to by Go To Declaration","Body":"<p>I am using pycharm to program a deep learning code. However, I found that some parts of Tensorflow library are not accessible by Go To Declaration in Pycharm. <\/p>\n\n<p>For example, the following one if I try to go to declaration of <code>Attentionwrapper<\/code> or <code>seq2seq<\/code><\/p>\n\n<pre><code>tf.contrib.seq2seq.AttentionWrapper\n<\/code><\/pre>\n\n<p>would have error as <code>\"cannot find declaration to go\"<\/code>. But it works if I only choose to go to declaration of <code>contrib<\/code><\/p>\n\n<p>I wonder why that happens and how I should fix this ? <\/p>\n","answers":[]}
+{"QuestionId":53368649,"AnswerCount":1,"Tags":"<tensorflow><image-processing><tensorflow.js>","CreationDate":"2018-11-19T05:10:10.627","AcceptedAnswerId":null,"OwnerUserId":2253546.0,"Title":"TensorFlow.js Resize 3D Tensor","Body":"<p>I have a 3D tensor with the the following dimensions : Width x Height x Depth. I need to resize variable sized volumes to a specific shape say 256 x 256 x 256. Unfortunately, in TensorFlow.js the set of methods they have for resizing such as <em>tf.image.resizeBilinear<\/em> &amp; <em>tf.image.resizeNearestNeighbor<\/em> only work for 2D images. Is there a workaround to get these methods to work in 3D space? <\/p>\n","answers":[{"AnswerId":"53381081","CreationDate":"2018-11-19T19:05:45.800","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>To resize a tensor, one can use <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.3\/#reshape\" rel=\"nofollow noreferrer\">tf.reshape<\/a> if the input size matches the output size<\/p>\n\n<pre class=\"lang-js prettyprint-override\"><code>const x = tf.tensor(Array.from({length :64}, (_, i) =&gt; i), [4, 4]);\nx.reshape([1, 16])\n<\/code><\/pre>\n\n<p>One application of reshape is when creating batches from an initial dataset <\/p>\n\n<p>If the input and the output size does not match, one can use <a href=\"https:\/\/js.tensorflow.org\/api\/0.13.3\/#slice\" rel=\"nofollow noreferrer\">tf.slice<\/a><\/p>\n\n<pre class=\"lang-js prettyprint-override\"><code>const x = tf.tensor(Array.from({length :64}, (_, i) =&gt; i), [4, 4, 4]);\nx.slice([1, 1, 1], [2, 2, 2]) \/\/ we are taking the 8 values at the center of the cube\n<\/code><\/pre>\n\n<p>The latter can be used to crop an image with the shape <code>[ height, width, channels]<\/code><\/p>\n\n<pre class=\"lang-js prettyprint-override\"><code>\/\/ t is a tensor\n\/\/ edge is the size of an edge of the cube\nconst cropImage = (t, edge) =&gt; {\n shape = t.shape;\n startCoord = shape.map(i =&gt; (i - edge) \/ 2)\n return t.slice(startCoord, [edge, edge, edge])\n \/\/ to keep the number of channels \n return t.slice([...startCoord.slice(0, shape.length - 1), 0], [edge, edge, channels])\n }\n<\/code><\/pre>\n"}]}
+{"QuestionId":53368717,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-11-19T05:19:40.300","AcceptedAnswerId":null,"OwnerUserId":9511772.0,"Title":"Loading Keras Model KeyError: 'weighted_metrics'","Body":"<p>I get the following error when trying to load a model through Keras:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"visualise.py\", line 82, in &lt;module&gt;\n    model = tf.keras.models.load_model(CNN_MODEL_DIR)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 251, in load_model\n    training_config['weighted_metrics'])\nKeyError: 'weighted_metrics'\n<\/code><\/pre>\n\n<p>I am using <code>model.save(\"Models\/ResNet.model\")<\/code> to save the model and <code>model = tf.keras.models.load_model(CNN_MODEL_DIR)<\/code> to load the model. <\/p>\n\n<p>The computer I am training this model on has <strong>Keras 2.1.2<\/strong> and the computer I am wanting to test it on (my home computer) has <strong>Keras 2.2.4<\/strong>. Is there a way to get around this error without retraining the model or downgrading my version of Keras?<\/p>\n","answers":[{"AnswerId":"53439843","CreationDate":"2018-11-23T02:03:05.760","ParentId":null,"OwnerUserId":"9511772","Title":null,"Body":"<p>I determined the issue:<\/p>\n\n<p>I imported the ResNet model with <code>from keras.applications.resnet50 import ResNet50<\/code> and saved it by simply using model.save(). Therefore, my model was purely Keras-related. Consequently, when I tried loading it with a Tensorflow-related import (i.e. <code>model = tf.keras.models.load_model(CNN_MODEL_DIR)<\/code>) I got the error. <\/p>\n\n<p>The fix was simply to change the import to be purely Keras.  <\/p>\n\n<p><code>model = keras.models.load_model(CNN_MODEL_DIR)<\/code><\/p>\n"}]}
+{"QuestionId":53369050,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><raspberry-pi><counter>","CreationDate":"2018-11-19T05:57:10.013","AcceptedAnswerId":null,"OwnerUserId":10646665.0,"Title":"Count people\/object using raspberry PI by OPENCV, TENSORFLOW and PYTHON","Body":"<p>My code can detect an object now, which is human and other objects. But the issue that i faced now is i want to count the moving human entering a classroom. I would like to count in and out so that i can know the number of people in a room. How can do this? something like this <a href=\"https:\/\/www.youtube.com\/watch?v=aEcBnD80nL\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=aEcBnD80nL<\/a><\/p>\n\n<p>How can i make that detected object in rectangle to count in and out when they pass through the red and blue line?<\/p>\n\n<pre><code>import os\nimport cv2\nimport numpy as np\nfrom picamera.array import PiRGBArray\nfrom picamera import PiCamera\nimport tensorflow as tf\nimport argparse\nimport sys\n\n# Set up camera constants\nIM_WIDTH = 1280\nIM_HEIGHT = 720\n#IM_WIDTH = 640    Use smaller resolution for\n#IM_HEIGHT = 480   slightly faster framerate\n\n# Select camera type (if user enters --usbcam when calling this script,\n# a USB webcam will be used)\ncamera_type = 'picamera'\nparser = argparse.ArgumentParser()\nparser.add_argument('--usbcam', help='Use a USB webcam instead of picamera',\n                action='store_true')\nargs = parser.parse_args()\nif args.usbcam:\ncamera_type = 'usb'\n\n# This is needed since the working directory is the object_detection folder.\nsys.path.append('..')\n\n# Import utilites\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util\n\n# Name of the directory containing the object detection module we're using\nMODEL_NAME = 'ssdlite_mobilenet_v2_coco_2018_05_09'\n\n# Grab path to current working directory\nCWD_PATH = os.getcwd()\n\n# Path to frozen detection graph .pb file, which contains the model that is used\n# for object detection.\nPATH_TO_CKPT = os.path.join(CWD_PATH,MODEL_NAME,'frozen_inference_graph.pb')\n\n# Path to label map file\nPATH_TO_LABELS = os.path.join(CWD_PATH,'data','mscoco_label_map.pbtxt')\n\n# Number of classes the object detector can identify\nNUM_CLASSES = 90\n\n## Load the label map.\n# Label maps map indices to category names, so that when the convolution\n# network predicts `5`, we know that this corresponds to `airplane`.\n# Here we use internal utility functions, but anything that returns a\n# dictionary mapping integers to appropriate string labels would be fine\nlabel_map = label_map_util.load_labelmap(PATH_TO_LABELS)\ncategories = label_map_util.convert_label_map_to_categories(label_map, \nmax_num_classes=NUM_CLASSES, use_display_name=True)\ncategory_index = label_map_util.create_category_index(categories)\n\n# Load the Tensorflow model into memory. \ndetection_graph = tf.Graph()\nwith detection_graph.as_default():\nod_graph_def = tf.GraphDef()\nwith tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:\n    serialized_graph = fid.read()\n    od_graph_def.ParseFromString(serialized_graph)\n    tf.import_graph_def(od_graph_def, name='')\n\nsess = tf.Session(graph=detection_graph)\n\n\n# Define input and output tensors (i.e. data) for the object detection \nclassifier\n\n# Input tensor is the image\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n\n# Output tensors are the detection boxes, scores, and classes\n# Each box represents a part of the image where a particular object was detected\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n\n# Each score represents level of confidence for each of the objects.\n# The score is shown on the result image, together with the class label.\ndetection_scores = detection_graph.get_tensor_by_name('detection_scores:0')\ndetection_classes = \ndetection_graph.get_tensor_by_name('detection_classes:0')\n\n# Number of objects detected\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\n# Initialize frame rate calculation\nframe_rate_calc = 1\nfreq = cv2.getTickFrequency()\nfont = cv2.FONT_HERSHEY_SIMPLEX\n\n# Initialize camera and perform object detection.\n# The camera has to be set up and used differently depending on if it's a\n# Picamera or USB webcam.\n\n# I know this is ugly, but I basically copy+pasted the code for the object\n# detection loop twice, and made one work for Picamera and the other work\n# for USB.\n\n### Picamera ###\nif camera_type == 'picamera':\n# Initialize Picamera and grab reference to the raw capture\ncamera = PiCamera()\ncamera.resolution = (IM_WIDTH,IM_HEIGHT)\ncamera.framerate = 10\nrawCapture = PiRGBArray(camera, size=(IM_WIDTH,IM_HEIGHT))\nrawCapture.truncate(0)\n\nfor frame1 in camera.capture_continuous(rawCapture, format=\"bgr\",use_video_port=True):\n\n    t1 = cv2.getTickCount()\n\n    # Acquire frame and expand frame dimensions to have shape: [1, None, None, 3]\n    # i.e. a single-column array, where each item in the column has the pixel RGB value\n    frame = frame1.array\n    frame.setflags(write=1)\n    frame_expanded = np.expand_dims(frame, axis=0)\n\n    # Perform the actual detection by running the model with the image as input\n    (boxes, scores, classes, num) = sess.run(\n        [detection_boxes, detection_scores, detection_classes, num_detections],\n        feed_dict={image_tensor: frame_expanded})\n\n    # Draw the results of the detection (aka 'visulaize the results')\n    vis_util.visualize_boxes_and_labels_on_image_array(\n        frame,\n        np.squeeze(boxes),\n        np.squeeze(classes).astype(np.int32),\n        np.squeeze(scores),\n        category_index,\n        use_normalized_coordinates=True,\n        line_thickness=8,\n        min_score_thresh=0.40)\n\n    # Blue line\n    cv2.line(frame, (IM_WIDTH \/\/ 2, 0), (IM_WIDTH \/\/ 2 , IM_WIDTH), (250, 0, 1), 2)\n    # Red line\n    cv2.line(frame, (IM_WIDTH \/\/ 2 - 50, 0), (IM_WIDTH \/\/ 2 - 50, IM_WIDTH), (0, 0, 255), 2)\n\n    # FPS Text\n    cv2.putText(frame,\"FPS: {0:.2f}\".format(frame_rate_calc),(30,50),font,1,(255,255,0),2,cv2.LINE_AA)\n\n    # All the results have been drawn on the frame, so it's time to display it.\n    cv2.imshow('Object detector', frame)\n\n    t2 = cv2.getTickCount()\n    time1 = (t2-t1)\/freq\n    frame_rate_calc = 1\/time1\n\n    # Press 'q' to quit\n    if cv2.waitKey(1) == ord('q'):\n        break\n\n    rawCapture.truncate(0)\n\ncamera.close()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53369111,"AnswerCount":0,"Tags":"<deep-learning><conv-neural-network><pytorch>","CreationDate":"2018-11-19T06:03:38.130","AcceptedAnswerId":null,"OwnerUserId":10673014.0,"Title":"Dice and CE loss not training network together","Body":"<p>I am training a segmentation network on the Kaggle Salt challenge. My dice and ce decrease, but then suddenly dice increases and CE jumps up a bit, this keeps happening to dice. I have been trying all day to fix this but can\u2019t get my code to run. I am running on only 10 data points to overfit my data but it just is not happening. Any help would be greatly appreciated.<\/p>\n\n<p>Plots of dice(top) and CE:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/u5mrs.png\" rel=\"nofollow noreferrer\">Loss curve<\/a><\/p>\n\n<p>Heres my dice and train:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def dice(input, target,weights=torch.tensor([1,1]).float().cuda()):\n    smooth=.001\n\n    dummy=np.zeros([batch_size,2,100,100]) # create dummy to one hot encode target for weighted dice\n    dummy[:,0,:,:][target==0]=1 # background class is 0\n    dummy[:,1,:,:][target==1]=1 # salt class is 1 \n\n\n    target=torch.tensor(dummy).float().cuda()\n\n#     print(input.size(),input[:,0,:,:].size())\n    input1=input[:,0,:,:].contiguous().view(-1) #flatten both classes seperately\n    target1=target[:,0,:,:].contiguous().view(-1)\n\n    input2=input[:,1,:,:].contiguous().view(-1)\n    target2=target[:,1,:,:].contiguous().view(-1)\n\n    score1=2*(input1*target1).sum()\/(input1.sum()+target1.sum()+smooth) #back\n    score2=2*(input2*target2).sum()\/(input2.sum()+target2.sum()+smooth) #salt\n\n\n    score=1-(weights[0]*score1+weights[1]*score2)\/2\n    if score&lt;0:\n        score=score-score\n\n    return(score)\nHeres the train:\n\n\ndef train(epoch):\n    for idx, batch_data in enumerate(dataloader) : \n        x, target=batch_data['image'].float().cuda(),batch_data['label'].float().cuda()\n\n\n        optimizer.zero_grad()\n        output = net(x)\n#         print(output.size())\n        output.squeeze_(1)\n\n#         print('out',output.size(),target.size())\n        bce_loss = criterion(output, target.long())\n        lc.append(bce_loss.item())\n\n        dice_loss = dice((output), target)\n        ld.append(dice_loss.item())\n        loss =  dice_loss + bce_loss\n        l.append(loss.item())\n\n        loss.backward()\n        optimizer.step()\n\n        print('Epoch {}, loss {}, bce {}, dice {}'.format(\n            epoch, sum(l)\/len(l), sum(lc)\/len(lc) , sum(ld)\/len(ld) ))\n<\/code><\/pre>\n\n<p>Heres the rest of the code( I downed from gaggle kernel): <a href=\"https:\/\/github.com\/bluesky314\/Salt-Segmentation\/blob\/master\/kernel-2.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bluesky314\/Salt-Segmentation\/blob\/master\/kernel-2.ipynb<\/a> 1 (the training showed here is when I ran that cell(14) a second time so the ups and downs don\u2019t appear but can be seen in the plot)<\/p>\n\n<p><code>dataset=DatasetSalt(limit_paths=10)<\/code> just limits the dataset to any number by only taking the top paths to get the images from<\/p>\n\n<p>Would really appreciate any help, have been literally struggling on this for 8+ hrs<\/p>\n","answers":[]}
+{"QuestionId":53369534,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-19T06:41:53.237","AcceptedAnswerId":null,"OwnerUserId":679688.0,"Title":"How to train and predict using keras Layers without high-level Model functions","Body":"<p>In my Tensorflow code, I want to compose my network using Keras Layers like Dense, Dropout and BatchNormalization because I don't want to write the layers myself but I don't want to rely on the sparsely-documented magic of high-level Keras Model functions like Model.fit(). I want to get the loss function, get the train op and apply the back-propagation variable update myself, mostly because I want to understand explicitly what the network is doing and be able to customize the training in ways that Keras does not provide.<\/p>\n\n<p>If I had only Layers that behave the same during both training and validation\/testing\/inference, then I could just use Approach 1 <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/models\/Model\" rel=\"nofollow noreferrer\">here<\/a> to make a Model, call it with the input tensor to get the prediction tensor and do whatever I need with it for the back-propagation and call it a day.<\/p>\n\n<p>However, with special layers like Dropout and BatchNormalization, I also need to do these:<\/p>\n\n<ul>\n<li>For Dropout, distinguish between training and non-training when getting the prediction, applying dropout only during training.<\/li>\n<li>For BatchNormalization, before each training step, update the batchnorm statistics.<\/li>\n<\/ul>\n\n<p>How can I use Model \"the hard way\" for training and prediction when Dropout and BatchNormalization are involved? I haven't been able to find any useful examples, and without a high-level explanation the only thing I can do is to dig into the Keras code and trace all manner of overloaded duck-typed functions. More precisely, I don't even need Keras Model per se. I just need a way to compose the Keras Layers for a network with which to do prediction and training.<\/p>\n","answers":[]}
+{"QuestionId":53369667,"AnswerCount":1,"Tags":"<matrix><vector><pytorch><tensor>","CreationDate":"2018-11-19T06:54:58.063","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"PyTorch element-wise product of vectors \/ matrices \/ tensors","Body":"<p>In PyTorch, how do I get the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Hadamard_product_(matrices)\" rel=\"nofollow noreferrer\">element-wise product<\/a> of two vectors \/ matrices \/ tensors?<\/p>\n\n<p>For googlers, this is product is also known as:<\/p>\n\n<ul>\n<li>Hadamard product<\/li>\n<li>Schur product<\/li>\n<li>Entrywise product<\/li>\n<\/ul>\n","answers":[{"AnswerId":"53369668","CreationDate":"2018-11-19T06:54:58.063","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>Given two tensors <code>A<\/code> and <code>B<\/code> you can use either:<\/p>\n\n<ul>\n<li><code>A * B<\/code><\/li>\n<li><a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?#torch.mul\" rel=\"nofollow noreferrer\"><code>torch.mul<\/code><\/a><code>(A, B)<\/code><\/li>\n<li><code>A.mul(B)<\/code><\/li>\n<\/ul>\n\n<p><strong>Note<\/strong>: for <em>matrix multiplication<\/em>, you want to use <code>A @ B<\/code> which is equivalent to <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=matmul#torch.matmul\" rel=\"nofollow noreferrer\"><code>torch.matmul()<\/code><\/a>.<\/p>\n"}]}
+{"QuestionId":53370000,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2018-11-19T07:22:15.947","AcceptedAnswerId":null,"OwnerUserId":2265724.0,"Title":"How to implement an ensemble of CNNs in Caffe?","Body":"<p>I would like to train an ensemble of CNNs in Caffe.<\/p>\n\n<ol>\n<li>What layers do I use to combine the softmax outputs to make decision by averaging? I am thinking perhaps adding a Concat layer followed by Pooling average layer in the TEST phase would do the trick? <\/li>\n<li>Each CNN should see different region in the image. Do I use several Crop layers to do this i.e. a Crop layer for each network ?   <\/li>\n<\/ol>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":53370003,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-11-19T07:22:43.593","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"PyTorch mapping operators to functions","Body":"<p>What are all the PyTorch operators, and what are their function equivalents?<\/p>\n\n<p>Eg, is <code>a @ b<\/code> equivalent to <code>a.mm(b)<\/code> or <code>a.matmul(b)<\/code>?<\/p>\n\n<p>I'm after a canonical listing of operator -> function mappings.<\/p>\n\n<p>I'd be happy to be given a PyTorch documentation link as an answer - my googlefu couldn't track it down.<\/p>\n","answers":[{"AnswerId":"53373333","CreationDate":"2018-11-19T11:08:46.710","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>This defines tensor operations for 0.3.1 (it does also contain the definitions of the other operators):\n<a href=\"https:\/\/pytorch.org\/docs\/0.3.1\/_modules\/torch\/tensor.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/0.3.1\/_modules\/torch\/tensor.html<\/a><\/p>\n\n<p>The code for the current stable has been rearranged (I guess they do more in C now), but since the behaviour of matrix multiplication hasn't changed,  I think it is save to  assume that this is still valid.<\/p>\n\n<p>See for the definition of <code>__matmul__<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def __matmul__(self, other):\n    if not torch.is_tensor(other):\n        return NotImplemented\n    return self.matmul(other)\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def matmul(self, other):\n    r\"\"\"Matrix product of two tensors.\n\n    See :func:`torch.matmul`.\"\"\"\n    return torch.matmul(self, other)\n<\/code><\/pre>\n\n<p>The operator <code>@<\/code> was introduced with <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0465\/\" rel=\"nofollow noreferrer\">PEP 465<\/a> and is mapped to <code>__matmul__<\/code>.<\/p>\n\n<p>See also here for this:<br>\n<a href=\"https:\/\/stackoverflow.com\/questions\/27385633\/what-is-the-symbol-for-in-python\">What is the &#39;@=&#39; symbol for in Python?<\/a><\/p>\n"},{"AnswerId":"53373484","CreationDate":"2018-11-19T11:17:28.300","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>The Python documentation table <a href=\"https:\/\/docs.python.org\/3\/library\/operator.html#mapping-operators-to-functions\" rel=\"nofollow noreferrer\">Mapping Operators to Functions<\/a> provides canonical mappings from:<\/p>\n\n<p>operator -> <code>__function__()<\/code><\/p>\n\n<p>Eg:<\/p>\n\n<blockquote>\n<pre><code>Matrix Multiplication        a @ b        matmul(a, b)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Elsewhere on the page, you will see the <code>__matmul__<\/code> name as an alternate to <code>matmul<\/code>.<\/p>\n\n<p>The definitions of the PyTorch <code>__functions__<\/code> are found either in:<\/p>\n\n<ul>\n<li><p>The <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/tensor.html\" rel=\"nofollow noreferrer\"><code>torch.Tensor<\/code> module documentation<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/e35418b3bec2f64814919f60aae9d98650771a3a\/tools\/autograd\/templates\/python_variable_methods.cpp#L625\" rel=\"nofollow noreferrer\"><code>python_variable_methods.cpp<\/code><\/a><\/p><\/li>\n<\/ul>\n\n<p>You can look up the documentation for the named functions at:<\/p>\n\n<pre><code>https:\/\/pytorch.org\/docs\/stable\/torch.html?#torch.&lt;FUNCTION-NAME&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":53370431,"AnswerCount":0,"Tags":"<python><conv-neural-network><pytorch>","CreationDate":"2018-11-19T07:55:29.983","AcceptedAnswerId":null,"OwnerUserId":8233800.0,"Title":"Pytorch - How to run inference of a model after thinning the model without the model.py file?","Body":"<p>I am having a pytorch model with the following network structure.<\/p>\n\n<pre><code>from __future__ import print_function\nimport argparse\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\n\nclass lenet_mnist(nn.Module):\n    def __init__(self):\n        super(lenet_mnist, self).__init__()\n        self.cuda()\n        self.conv1 = nn.Conv2d(1, 16, kernel_size=3,padding=1)\n        self.relu1 = nn.ReLU(inplace=True)\n        self.relu2 = nn.ReLU(inplace=True)\n        self.relu3 = nn.ReLU(inplace=True)\n        self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2)\n        self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2)\n        self.conv2 = nn.Conv2d(16, 32, kernel_size=5,padding=2)\n        self.conv2_drop = nn.Dropout2d()\n        self.fc1 = nn.Linear(1568, 256)\n        self.fc2 = nn.Linear(256, 10)\n\n    def forward(self,x):\n        #x = x.to(torch.device(\"cuda:0\"))\n        out = self.conv1(x)\n        out = self.maxpool1(out)\n        out = self.relu1(out)\n        out = self.conv2(out)\n        out = self.conv2_drop(out)\n        out = self.maxpool2(out)\n        out = self.relu2(out)\n        out = out.view(-1,1568)\n        out = self.fc1(out)\n        out = self.fc2(out)\n        return out \n<\/code><\/pre>\n\n<p>I train this model and I remove few channels in the model, for example I made the <code>self.conv2<\/code> as (<code>in=16,out=24<\/code>). I reloaded the <code>state_dict<\/code> model and modified the model for the new structure. <\/p>\n\n<p>Now , I am trying to run an inference of the model, but <code>pytorch<\/code> still picks up the old <code>model.py<\/code> file where the original structure is defined. How do I stop <code>pytorch<\/code> looking for <code>model.py<\/code>?<\/p>\n\n<p>Edit: (more information)\nFor thinning the model, I just get the state_dict of the model, edit the tensors by deleting few channels in it and then use load_state_dict on the model to load the model again. I also change the model._parameters with the updated weights and bias. I use this model to run a forward pass. While running, I get an error saying mismatch in dimensions at out.view layer. And, the backtrace shows that the pytorch is reading the model from model.py file where the above code snippet is present. <\/p>\n","answers":[]}
+{"QuestionId":53370621,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-19T08:12:54.927","AcceptedAnswerId":53370708.0,"OwnerUserId":9713902.0,"Title":"I am using keras , but it is throwing a error with the helper libraries. No module names \"lstm\"","Body":"<pre><code>from keras.layers.core import Dense,Activation,Dropout\nfrom keras.layers.recurrent import LSTM\nfrom keras.models import Sequential\nimport time,lstm\n<\/code><\/pre>\n\n<p>The error message is No module name \"lstm\" <\/p>\n","answers":[{"AnswerId":"53370708","CreationDate":"2018-11-19T08:19:20.940","ParentId":null,"OwnerUserId":"2558618","Title":null,"Body":"<p>It looks like you have already imported <code>LSTM<\/code> from <code>keras.layers<\/code>, so there's no reason to import <code>lstm<\/code> again. According to the <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/\" rel=\"nofollow noreferrer\">examples<\/a>, I don't see any instances of the (lower case) <code>lstm<\/code> import, just the <code>LSTM<\/code> from <code>keras.layers<\/code>.<\/p>\n"}]}
+{"QuestionId":53371014,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><computer-vision>","CreationDate":"2018-11-19T08:43:43.660","AcceptedAnswerId":null,"OwnerUserId":10186104.0,"Title":"How to load images for training a neural network in real time in Keras?","Body":"<p>I'm using <code>Keras<\/code> with <code>tensorflow<\/code> backend. I have a large dataset (Caltech 256) that I want to train. <\/p>\n\n<p>I've created four <code>.npy<\/code> files called <code>training<\/code>, <code>trainlbl<\/code>, <code>testing<\/code>, <code>testlbl<\/code>. <\/p>\n\n<p>Now instead of running <code>np.load()<\/code> which almost completely occupies my ram, how can I load them from disk in real time while training itself?\nFrom searching on the internet I get an idea that it can be done through Keras's data generator, but I'm not able to understand how.<\/p>\n","answers":[{"AnswerId":"53371225","CreationDate":"2018-11-19T08:59:51.977","ParentId":null,"OwnerUserId":"7301994","Title":null,"Body":"<p>Maybe it would be possible to answer your question in this post, but i think it isn't useful because you need more knowledge how to use or write deep learning networks.<\/p>\n\n<p>My suggestion would be to use existing tutorials which shows you how to create neural networks for image classification.\nA very nice example is this blog post: <a href=\"https:\/\/www.pyimagesearch.com\/2017\/12\/11\/image-classification-with-keras-and-deep-learning\/\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2017\/12\/11\/image-classification-with-keras-and-deep-learning\/<\/a><\/p>\n\n<p>This guy shows you how you use DataGenerator from Keras.\nDataGenerator can do many more things than just lazy loading images from disk.<\/p>\n"}]}
+{"QuestionId":53371591,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T09:25:03.233","AcceptedAnswerId":null,"OwnerUserId":10673786.0,"Title":"Reusing closed session in tensorflow","Body":"<p>I made my model and using <code>with<\/code> clause I used the same in a session:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n\n<p>How can I use the above session <code>sess<\/code> since it is closed outside the with clause?<\/p>\n","answers":[{"AnswerId":"53372052","CreationDate":"2018-11-19T09:52:24.510","ParentId":null,"OwnerUserId":"2482068","Title":null,"Body":"<p>You can run your model in the same <code>with<\/code> block:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    sess.run(your_model, feed_dict={...})\n<\/code><\/pre>\n\n<p>Or create session and use it without <code>with<\/code>:<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsess.run(your_model, feed_dict={...})\n<\/code><\/pre>\n"}]}
+{"QuestionId":53371650,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network><classification>","CreationDate":"2018-11-19T09:27:56.223","AcceptedAnswerId":53372737.0,"OwnerUserId":2651073.0,"Title":"MNIST Classification: mean_squared_error loss function and tanh activation function","Body":"<p>I changed the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">getting started example<\/a> of Tensorflow as following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom sklearn.metrics import roc_auc_score\nimport numpy as np\nimport commons as cm\nfrom sklearn.metrics import confusion_matrix\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport seaborn as sn\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(),\n  tf.keras.layers.Dense(512, activation=tf.nn.tanh),\n  # tf.keras.layers.Dense(512, activation=tf.nn.tanh),\n  # tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation=tf.nn.tanh)\n])\nmodel.compile(optimizer='adam',\n               loss='mean_squared_error',\n              # loss = 'sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nhistory = cm.Histories()\nh= model.fit(x_train, y_train, epochs=50, callbacks=[history])\nprint(\"history:\", history.losses)\ncm.plot_history(h)\n# cm.plot(history.losses, history.aucs)\n\n\ntest_predictions = model.predict(x_test)\n\n\n# Compute confusion matrix\npred = np.argmax(test_predictions,axis=1)\npred2 = model.predict_classes(x_test)\nconfusion = confusion_matrix(y_test, pred)\ncm.draw_confusion(confusion,range(10))\n<\/code><\/pre>\n\n<p>With its default parameters: <\/p>\n\n<ul>\n<li><code>relu<\/code> activation at hidden layers, <\/li>\n<li><code>softmax<\/code> at the output layer and <\/li>\n<li><code>sparse_categorical_crossentropy<\/code> as loss function, <\/li>\n<\/ul>\n\n<p>it works fine and the prediction for all digits are above 99%<\/p>\n\n<p>However with my parameters: <code>tanh<\/code> activation function and <code>mean_squared_error<\/code> loss function it just predict <code>0<\/code> for all test samples:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CH7HG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CH7HG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I wonder what is the problem? The accuracy rate is increasing for each epoch and it reaches 99% and loss is about 20<\/p>\n","answers":[{"AnswerId":"53372737","CreationDate":"2018-11-19T10:35:12.357","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>You need to use the proper loss function for your data. Here you have a categorical output, so you need to use <code>sparse_categorical_crossentropy<\/code>, but also set <code>from_logits<\/code> without any activation for the last layer.<\/p>\n\n<p>If you need to use <code>tanh<\/code> as your output, then you can use MSE with a one-hot encoded version of your labels + rescaling.<\/p>\n"}]}
+{"QuestionId":53372460,"AnswerCount":1,"Tags":"<tensorflow><kotlin-native>","CreationDate":"2018-11-19T10:17:37.157","AcceptedAnswerId":53500857.0,"OwnerUserId":6107009.0,"Title":"Kotlin\/Native Tensorflow","Body":"<p>I created a tensorflow library and trying get tensorflow version in my kotlin script like that <\/p>\n\n<pre><code>import tensorflow.TF_Version\n\nfun main(args: Array&lt;String&gt;) {\n    TF_Version()\n}\n<\/code><\/pre>\n\n<p>idea can link TF_Version() method to library with Ctrl+click\nand there(in library) i have method <\/p>\n\n<pre><code>fun TF_Version(): CPointer&lt;ByteVar&gt;? {\n    return interpretCPointer&lt;ByteVar&gt;(kniBridge0())\n}\n\n@SymbolName(\"tensorflow_kniBridge0\")\nprivate external fun kniBridge0(): NativePtr\n<\/code><\/pre>\n\n<p>but when i try to build and start i see this error <\/p>\n\n<blockquote>\n  <p>\/tmp\/konan_temp8584442034275821874\/combined.o:ld-temp.o:function Konan_start: error: undefined reference to 'tensorflow_kniBridge0'\n  error: \/home\/fnasibov\/.konan\/dependencies\/target-gcc-toolchain-3-linux-x86-64\/x86_64-unknown-linux-gnu\/bin\/ld.gold invocation reported errors<\/p>\n<\/blockquote>\n\n<p>Please help me fix that<\/p>\n","answers":[{"AnswerId":"53500857","CreationDate":"2018-11-27T13:30:20.360","ParentId":null,"OwnerUserId":"6107009","Title":null,"Body":"<p>Found solution:\nmy .def file was <\/p>\n\n<pre><code>headers =  \/home\/user\/.konan\/third--arty\/tensorflow\/include\/tensorflow\/c\/c_api.h\n<\/code><\/pre>\n\n<p>And after i add Opts to def file:<\/p>\n\n<pre><code>headers = \/home\/fnasibov\/.konan\/third-party\/tensorflow\/include\/tensorflow\/c\/c_api.h\n\ncompilerOpts.linux = -I\/usr\/include -I\/usr\/include\/x86_64-linux-gnu\nlinkerOpts.linux = -L\/home\/user\/.konan\/third-party\/tensorflow\/lib -ltensorflow\n<\/code><\/pre>\n\n<p>and regenerate klib - thats worked!<\/p>\n"}]}
+{"QuestionId":53374114,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-19T11:55:16.727","AcceptedAnswerId":53376239.0,"OwnerUserId":10648778.0,"Title":"Tensorflow accuracy stuck at 25% for 4 labels, text classification","Body":"<p>The accuracy starts off at around 40% and drops down during one epoch to 25%<\/p>\n\n<p>My model:<\/p>\n\n<pre><code>self._model = keras.Sequential()\nself._model.add(keras.layers.Dense(12, activation=tf.nn.sigmoid))  # hidden layer\nself._model.add(keras.layers.Dense(len(VCDNN.conventions), activation=tf.nn.softmax))  # output layer\noptimizer = tf.train.AdamOptimizer(0.01)\nself._model.compile(optimizer, loss=tf.losses.sparse_softmax_cross_entropy, metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>I have 4 labels, 60k rows of data, split evenly for each label so 15k each and 20k rows of data for evaluation<\/p>\n\n<p>my data example:<\/p>\n\n<pre><code>name label\nabcTest label1\nmete_Test label2 \nROMOBO label3\ntest label4\n<\/code><\/pre>\n\n<p>The input is turned into integers for each character and then hot encoded and output is just turned into integers [0-3]<\/p>\n\n<p>1 epoch evaluation (loss, acc):<\/p>\n\n<pre><code>[0.7436684370040894, 0.25]\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong>\nMore details about the data <\/p>\n\n<p>The strings are of up to 20 characters\nI first convert each character to int based on an alphabet dictionary (a: 1, b:2, c:3) and if a word is shorter than 20 chars i fill the rest with 0's now those values are hot encoded and reshaped so<\/p>\n\n<pre><code>assume max 5 characters\n1. [\"abc\",\"d\"]\n2. [[1,2,3,0,0],[4,0,0,0,0]]\n3. [[[0,1,0,0,0],[0,0,1,0,0],[0,0,0,1,0],[1,0,0,0,0],[1,0,0,0,0]],[[0,0,0,0,1],[1,0,0,0,0],[1,0,0,0,0],[1,0,0,0,0],[1,0,0,0,0]]]\n4. [[0,1,0,0,0,0,0,1,0,0,0,0,0,1,0,1,0,0,0,0,1,0,0,0,0],[0,0,0,0,1,1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,0,0,0,0]]\n<\/code><\/pre>\n\n<p>and labels describe the way a word is spelled basically naming convention e.g. all lowercase - unicase, testBest - camelCase, TestTest - PascalCase, test_test - snake_case<\/p>\n\n<p>With added 2 extra layers and LR reduced to 0.001\n<a href=\"https:\/\/i.stack.imgur.com\/ggXOb.png\" rel=\"nofollow noreferrer\">Pic of training<\/a><\/p>\n\n<p><strong>Update 2<\/strong><\/p>\n\n<pre><code>self._model = keras.Sequential()\nself._model.add(\n        keras.layers.Embedding(VCDNN.alphabetLen, 12, input_length=VCDNN.maxFeatureLen * VCDNN.alphabetLen))\n    self._model.add(keras.layers.LSTM(12))\n    self._model.add(keras.layers.Dense(len(VCDNN.conventions), activation=tf.nn.softmax))  # output layer\n    self._model.compile(tf.train.AdamOptimizer(self._LR), loss=\"sparse_categorical_crossentropy\",\n                        metrics=self._metrics)\n<\/code><\/pre>\n\n<p>Seems to start and immediately dies with no error (-1073740791)<\/p>\n","answers":[{"AnswerId":"53374308","CreationDate":"2018-11-19T12:06:43.943","ParentId":null,"OwnerUserId":"953674","Title":null,"Body":"<p>This means your models isn't really learning anything useful. It might be stuck in a local minima. This could be due to following reasons:<\/p>\n\n<ul>\n<li>a) you don't have enough train data to train a neural network. NNs usually require fairly large datasets to converge. Try using a RandomForest classifier at first to see what results you can get there<\/li>\n<li>b) it's possible your target data might not have anything to do with your train data and so it's impossible to train such a model that would map efficiently without overfitting<\/li>\n<li>c) your model could do with some improvements<\/li>\n<\/ul>\n\n<p>If you want to give improving your model a go I would add a few extra dense layers with a few more units. So after line 2 of your model I'd add:<\/p>\n\n<pre><code>self._model.add(keras.layers.Dense(36, activation=tf.nn.sigmoid))\nself._model.add(keras.layers.Dense(36, activation=tf.nn.sigmoid))\n<\/code><\/pre>\n\n<p>Another thing you can try is a different learning rate. I'd go with the default for AdamOptimizer which is 0.001. So just change <code>0.01<\/code> to <code>0.001<\/code> in the AdamOptimizer() call<\/p>\n\n<p>You may also want to train more than just one epoch<\/p>\n"},{"AnswerId":"53376239","CreationDate":"2018-11-19T13:58:36.053","ParentId":null,"OwnerUserId":"4554481","Title":null,"Body":"<p>The 0.25 acc means the model couldn't learn anything useful as it is same as the random guess. This means the network structure may not good for the problem.<\/p>\n\n<p>Currently, the recurring neural network, like LSTM, is more commonly used for sequence modeling. For instance:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(char_size, embedding_size))\nmodel.add(LSTM(hidden_size))\nmodel.add(Dense(len(VCDNN.conventions), activation='softmax'))\n<\/code><\/pre>\n\n<p>This will work better if the label is related to the char sequence information about the input words.<\/p>\n"}]}
+{"QuestionId":53374154,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-11-19T11:58:17.577","AcceptedAnswerId":53374243.0,"OwnerUserId":2651073.0,"Title":"How to convert a binary matrix to a bipolar one in python","Body":"<p>There is a function in <a href=\"https:\/\/keras.io\/utils\/\" rel=\"nofollow noreferrer\">Keras<\/a> to generate a binary matrix for an array of labels:<\/p>\n\n<pre><code># Consider an array of 5 labels out of a set of 3 classes {0, 1, 2}:\n&gt; labels\narray([0, 2, 1, 2, 0])\n# `to_categorical` converts this into a matrix with as many\n# columns as there are classes. The number of rows\n# stays the same.\n&gt; to_categorical(labels)\narray([[ 1.,  0.,  0.],\n       [ 0.,  0.,  1.],\n       [ 0.,  1.,  0.],\n       [ 0.,  0.,  1.],\n       [ 1.,  0.,  0.]], dtype=float32)\n<\/code><\/pre>\n\n<p>I need the above functionality, but having <code>-1<\/code> instead of zeros. I didn't find any option or other functions to do it. Is there any easy way to do that?<\/p>\n","answers":[{"AnswerId":"53374243","CreationDate":"2018-11-19T12:03:04.407","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>Just rescale your data:<\/p>\n\n<pre><code>2*to_categorical(labels)-1\n<\/code><\/pre>\n"},{"AnswerId":"53374300","CreationDate":"2018-11-19T12:06:05.367","ParentId":null,"OwnerUserId":"4001592","Title":null,"Body":"<p>You could do the following:<\/p>\n\n<pre><code>import numpy as np\n\narr = np.array([[1., 0., 0.],\n                [0., 0., 1.],\n                [0., 1., 0.],\n                [0., 0., 1.],\n                [1., 0., 0.]])\n\narr[np.isclose(arr, 0)] = -1\nprint(arr)\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>[[ 1. -1. -1.]\n [-1. -1.  1.]\n [-1.  1. -1.]\n [-1. -1.  1.]\n [ 1. -1. -1.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53374257,"AnswerCount":0,"Tags":"<java><macos><tensorflow>","CreationDate":"2018-11-19T12:03:47.897","AcceptedAnswerId":null,"OwnerUserId":10461729.0,"Title":"How to run trained Tensorflow model on webcam in Java","Body":"<p>I have trained my Tensorflow object detection model for a single class only.<\/p>\n\n<p>I have followed <a href=\"http:\/\/\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\/tree\/d1c5b59803543e48362c27c48d704d4b0d92d135\" rel=\"nofollow noreferrer\">this tutorial<\/a> to train the model.<\/p>\n\n<p>This tutorial produces a .pb file. I want to use it on a live webcam. I have trained the model on Mac, and want to use it on a webcam (live stream) in Java.\nCan anyone tell me how I can do this?<\/p>\n","answers":[]}
+{"QuestionId":53374499,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-11-19T12:19:50.347","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"Get the data type of a PyTorch tensor","Body":"<p>I understand that PyTorch tensors are homogenous, ie, each of the elements are of the same type.<\/p>\n\n<p>How do I find out the type of the elements in a PyTorch tensor?<\/p>\n","answers":[{"AnswerId":"53374500","CreationDate":"2018-11-19T12:19:50.347","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>Use the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensor_attributes.html#torch-dtype\" rel=\"nofollow noreferrer\"><code>dtype<\/code> attribute<\/a>:<\/p>\n\n<pre><code>&gt;&gt;&gt; import torch\n&gt;&gt;&gt; print(torch.Tensor(1).dtype)\ntorch.float32\n<\/code><\/pre>\n\n<p>The documentation also gives the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensor_attributes.html#torch-dtype\" rel=\"nofollow noreferrer\">canonical list of datatypes<\/a>.<\/p>\n"},{"AnswerId":"57094062","CreationDate":"2019-07-18T12:04:48.770","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>There are three kinds of things:<\/p>\n\n<pre><code>dtype                   || CPU tensor               || GPU tensor\n\ntorch.float32              torch.FloatTensor           torch.cuda.FloatTensor\n<\/code><\/pre>\n\n<p>The first one you get with <code>print(t.dtype)<\/code> if <code>t<\/code> is your tensor, else you use <code>t.type()<\/code> for the other two.<\/p>\n"}]}
+{"QuestionId":53374928,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-19T12:44:24.820","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"Check if PyTorch tensors are equal within epsilon","Body":"<p>How do I check if two PyTorch tensors are semantically equal?<\/p>\n\n<p>Given floating point errors, I want to know if the the elements differ only by a small epsilon value.<\/p>\n","answers":[{"AnswerId":"53374929","CreationDate":"2018-11-19T12:44:24.820","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>At the time of writing, this is a undocumented function in the latest stable release (0.4.1), but the documentation is in the <code>master (unstable)<\/code> branch.<\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/master\/torch.html?highlight=allclose#torch.allclose\" rel=\"nofollow noreferrer\"><code>torch.allclose()<\/code><\/a> will return a boolean indicating whether all element-wise differences are equal allowing for a margin of error.<\/p>\n\n<p>Additionally, there's the undocumented <code>isclose()<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; torch.isclose(torch.Tensor([1]), torch.Tensor([1.00000001]))\ntensor([1], dtype=torch.uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53375422,"AnswerCount":1,"Tags":"<c++><python-3.x><parallel-processing><distributed-computing><pytorch>","CreationDate":"2018-11-19T13:13:17.437","AcceptedAnswerId":null,"OwnerUserId":1612428.0,"Title":"How pytorch's parallel method and distributed method works?","Body":"<p>I'm no expert in distributed system and CUDA. But there is one really interesting feature that PyTorch support which is <code>nn.DataParallel<\/code> and <code>nn.DistributedDataParallel<\/code>. How they are actually implemented? How they separate common embeddings and synchronize data? <\/p>\n\n<p>Here is a basic example of <code>DataParallel<\/code>.<\/p>\n\n<pre><code>import torch.nn as nn\nfrom torch.autograd.variable import Variable\nimport numpy as np\n\nclass Model(nn.Module):\n    def __init__(self):\n        super().__init__(\n            embedding=nn.Embedding(1000, 10),\n            rnn=nn.Linear(10, 10),\n        )\n\n    def forward(self, x):\n        x = self.embedding(x)\n        x = self.rnn(x)\n        return x\n\nmodel = nn.DataParallel(Model())\nmodel.forward(Variable.from_numpy(np.array([1,2,3,4,5,6], dtype=np.int64)).cuda()).cpu()\n<\/code><\/pre>\n\n<p>PyTorch can split the input and send them to many GPUs and merge the results back. <\/p>\n\n<p>How does it manage embeddings and synchronization for a parallel model or a distributed model? <br>\nI wandered around PyTorch's code but it's very hard to know how the fundamentals work.<\/p>\n","answers":[{"AnswerId":"53523827","CreationDate":"2018-11-28T16:17:51.997","ParentId":null,"OwnerUserId":"2745140","Title":null,"Body":"<p>From what I can trace, the code is implemented in <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/35348dab103316a929ae55d91a9d03e049d7fc43\/torch\/nn\/parallel\/parallel_apply.py#L21\" rel=\"nofollow noreferrer\"><code>parallel_apply.py<\/code><\/a><\/p>\n\n<p>[Edit: Paste the code here of easy reference]<\/p>\n\n<pre><code>def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):\n    r\"\"\"Applies each `module` in :attr:`modules` in parallel on arguments\n    contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword)\n    on each of :attr:`devices`.\n    Args:\n        modules (Module): modules to be parallelized\n        inputs (tensor): inputs to the modules\n        devices (list of int or torch.device): CUDA devices\n    :attr:`modules`, :attr:`inputs`, :attr:`kwargs_tup` (if given), and\n    :attr:`devices` (if given) should all have same length. Moreover, each\n    element of :attr:`inputs` can either be a single object as the only argument\n    to a module, or a collection of positional arguments.\n    \"\"\"\n    assert len(modules) == len(inputs)\n    if kwargs_tup is not None:\n        assert len(modules) == len(kwargs_tup)\n    else:\n        kwargs_tup = ({},) * len(modules)\n    if devices is not None:\n        assert len(modules) == len(devices)\n    else:\n        devices = [None] * len(modules)\n    devices = list(map(lambda x: _get_device_index(x, True), devices))\n    lock = threading.Lock()\n    results = {}\n    grad_enabled = torch.is_grad_enabled()\n\n    def _worker(i, module, input, kwargs, device=None):\n        torch.set_grad_enabled(grad_enabled)\n        if device is None:\n            device = get_a_var(input).get_device()\n        try:\n            with torch.cuda.device(device):\n                # this also avoids accidental slicing of `input` if it is a Tensor\n                if not isinstance(input, (list, tuple)):\n                    input = (input,)\n                output = module(*input, **kwargs)\n            with lock:\n                results[i] = output\n        except Exception as e:\n            with lock:\n                results[i] = e\n\n    if len(modules) &gt; 1:\n        threads = [threading.Thread(target=_worker,\n                                    args=(i, module, input, kwargs, device))\n                   for i, (module, input, kwargs, device) in\n                   enumerate(zip(modules, inputs, kwargs_tup, devices))]\n\n        for thread in threads:\n            thread.start()\n        for thread in threads:\n            thread.join()\n    else:\n        _worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])\n\n    outputs = []\n    for i in range(len(inputs)):\n        output = results[i]\n        if isinstance(output, Exception):\n            raise output\n        outputs.append(output)\n    return outputs\n<\/code><\/pre>\n\n<ul>\n<li><code>modules<\/code> are modules to be paralleled.<\/li>\n<li><code>inputs<\/code> are tensors to the modules<\/li>\n<li><code>devices<\/code> are CUDA devices<\/li>\n<li><code>results<\/code> and <code>output<\/code> store the final result<\/li>\n<li><code>_worker()<\/code> is the main function to be run by thread<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53375521,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-19T13:19:40.860","AcceptedAnswerId":null,"OwnerUserId":10674863.0,"Title":"keras data generator to tensorflow input","Body":"<p>I am trying to train a deep learning model in a distributive way. I have built my model in Keras and now I need to transform my model to a tensorflow estimator in order for it to work. Since my dataset is already big enough and it will grow bigger as I collect more data, I used <code>ImageDataGenerator<\/code> followed by <code>flow_from_directory<\/code>. I specify that I want batches of size 32 with desired shape, which is (299, 299, 3). Therefore, the output are batches of (32, 299, 299, 3). Before, I just feed the objects to <code>model.fit_generator<\/code> and Keras takes care of the rest. <\/p>\n\n<p>So far, I have converted my Keras model to a tensorflow estimator, but I can't figure out if I can use the object from the <code>flow_from_directory<\/code> to train the data using tf.train. I have already tried to write my own <code>input_fn<\/code> but I got confused. Your help is highly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53375772,"AnswerCount":1,"Tags":"<tensorflow><computer-vision>","CreationDate":"2018-11-19T13:33:42.213","AcceptedAnswerId":53435173.0,"OwnerUserId":4330326.0,"Title":"Home surveillance, TensorFlow (+ OpenCV)","Body":"<p>I am considering to implement a network of security cameras in my countryside property due to the recent intrusion of some burglars. To do so, I am trying to make the most out of the equipment I already have at home, such as:<\/p>\n\n<ul>\n<li>6 IP night vision cameras <\/li>\n<li>Jetson TK1 as a processing unit,\n[<a href=\"https:\/\/www.nvidia.com\/object\/jetson-tk1-embedded-dev-kit.html][1]\" rel=\"nofollow noreferrer\">https:\/\/www.nvidia.com\/object\/jetson-tk1-embedded-dev-kit.html][1]<\/a><\/li>\n<li>Raspberry Pi 2 model B (not needed, I guess)<\/li>\n<li>PoE Switch <\/li>\n<\/ul>\n\n<p>I wanted to feed the video stream of each camera \"into\" TensorFlow+OpenCV (or other DeepLearning algorithms) to make sure I can recognise what's\/who's in my garden, garage, etc. - basically it would be:<\/p>\n\n<ul>\n<li>Human detection &amp; car detection<\/li>\n<\/ul>\n\n<p>planning to use Telegram or other APIs provided by the instant messaging platform to be notified real-time.<\/p>\n\n<p>Configuring Tensorflow on my ubuntu 16.04 LTS (workstation with GPU support) was not trivial at all nor it will be the process of recognising objects and people. <\/p>\n\n<p>Do you have any recommendation on this little project?\nWould it make sense using Docker to make the configuration smoother? <\/p>\n","answers":[{"AnswerId":"53435173","CreationDate":"2018-11-22T16:37:55.270","ParentId":null,"OwnerUserId":"6380921","Title":null,"Body":"<p>You can break it down as follow:<\/p>\n\n<p>1 - Decide between bounding box and semantic segmentation<\/p>\n\n<p>2 - Select your pre-trained model after selecting your framework (Tensorflow in your case). You need to beat around Real Time vs Accuracy\n2.1 - The best in my opinion would be MobilenetV2 based deeplab (segmentation) or NASNET (bounding box)<\/p>\n\n<p>3 - Retrain it for transfer learning if Cars &amp; Persons are not in your model, otherwise you dont need to<\/p>\n\n<p>4 - Look for model optimization if required for Jetson TK1 (compression etc)<\/p>\n\n<p>5 - Once successfully deployed and it starts looking for person and cars, develop your logic for how should it alert you. (e.g. absence of car for x consecutive frames or presence of person for x consecutive frames etc)<\/p>\n\n<p>Happy Surveillaying! <\/p>\n"}]}
+{"QuestionId":53375889,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-19T13:40:40.633","AcceptedAnswerId":53376318.0,"OwnerUserId":7050062.0,"Title":"How to use dataset.map in Tensorflow Data API correctly","Body":"<p>I'm learning how to use the Tensorflow Data API, and struggle to understand how mapping works. For the context, I want to load a dataset of images and send them to a neural network. <\/p>\n\n<p>The MWE below does that in idea (fake dataset of size 10, <code>read_image<\/code> function mapped to the dataset).<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n\ndef read_image(filename, label):\n    return np.random.rand(8, 8, 1), label # simulate data load (generate random data)\n\n# generate fake dataset of filenames (of size 10)\nfilenames = tf.constant(np.asarray([\"file\" + str(i) for i in range(10)]))\nlabels = tf.constant(np.asarray([2*i for i in range(10)]))\n\ndataset = tf.data.Dataset.from_tensor_slices((filenames, labels))\ndataset = dataset.map(read_image)\n\ndataset = dataset.repeat().batch(2)\niterator = tf.data.Iterator.from_structure(dataset.output_types, dataset.output_shapes)\n\n\nX, y = iterator.get_next()\ntrain_init_op = iterator.make_initializer(dataset)\n\nwith tf.Session() as session:\n    tf.global_variables_initializer().run()\n    session.run(train_init_op)\n    for _ in range(10):\n        print(session.run([X]))\n<\/code><\/pre>\n\n<p>When running this code (that is supposed to do nothing, just print the values generated by <code>read_image<\/code>), it ends up with always the same data: <code>read_image<\/code> is called only once. Why is that ? I used <code>dataset.map<\/code>, isn't it supposed to be called on every element of my dataset (10 here) ?<\/p>\n\n<p>Thanks in advance for any help or advice.<\/p>\n","answers":[{"AnswerId":"53376318","CreationDate":"2018-11-19T14:02:52.950","ParentId":null,"OwnerUserId":"7050062","Title":null,"Body":"<p>My understanding of it was wrong, I actually figured out how to make it work. This map function is \"integrated\" to the Tensorflow graph, hence indeed called just once. One need to use TF operations in it.<\/p>\n\n<p>If <code>read_image<\/code> is replaced by:<\/p>\n\n<pre><code>def read_image(filename, label):\n    return tf.random_normal([4]), tf.random_normal([1])\n<\/code><\/pre>\n\n<p>It works as expected (random values are generated each time <code>session.run<\/code> is called).<\/p>\n"}]}
+{"QuestionId":53376761,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2018-11-19T14:28:57.043","AcceptedAnswerId":53378238.0,"OwnerUserId":1115237.0,"Title":"LSTM - Making predictions on partial sequence","Body":"<p>This question is in continue to a <a href=\"https:\/\/stackoverflow.com\/questions\/53190253\/stateful-lstm-and-stream-predictions\">previous question<\/a> I've asked.<\/p>\n\n<p>I've trained an LSTM model to predict a binary class (1 or 0) for batches of 100 samples with 3 features each, i.e: the shape of the data is (m, 100, 3), where m is the number of batches.<\/p>\n\n<p>Data:<\/p>\n\n<pre><code>[\n    [[1,2,3],[1,2,3]... 100 sampels],\n    [[1,2,3],[1,2,3]... 100 sampels],\n    ... avaialble batches in the training data\n]\n<\/code><\/pre>\n\n<p>Target:<\/p>\n\n<pre><code>[\n   [1]\n   [0]\n   ...\n]\n<\/code><\/pre>\n\n<p>Model code:<\/p>\n\n<pre><code>def build_model(num_samples, num_features, is_training):\n    model = Sequential()\n    opt = optimizers.Adam(lr=0.0005, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)\n\n    batch_size = None if is_training else 1\n    stateful = False if is_training else True\n    first_lstm = LSTM(32, batch_input_shape=(batch_size, num_samples, num_features), return_sequences=True,\n                      activation='tanh', stateful=stateful)\n\n    model.add(first_lstm)\n    model.add(LeakyReLU())\n    model.add(Dropout(0.2))\n    model.add(LSTM(16, return_sequences=True, activation='tanh', stateful=stateful))\n    model.add(Dropout(0.2))\n    model.add(LeakyReLU())\n    model.add(LSTM(8, return_sequences=False, activation='tanh', stateful=stateful))\n    model.add(LeakyReLU())\n    model.add(Dense(1, activation='sigmoid'))\n\n    if is_training:\n        model.compile(loss='binary_crossentropy', optimizer=opt,\n                      metrics=['accuracy', keras_metrics.precision(), keras_metrics.recall(), f1])\n    return model\n<\/code><\/pre>\n\n<p>For the training stage, the model is <strong>NOT<\/strong> stateful. When predicting I'm using a <strong>stateful<\/strong> model, iterating over the data and outputting a probability for each sample:<\/p>\n\n<pre><code>for index, row in data.iterrows():\n    if index % 100 == 0:\n        predicting_model.reset_states()\n    vals = np.array([[row[['a', 'b', 'c']].values]])\n    prob = predicting_model.predict_on_batch(vals)\n<\/code><\/pre>\n\n<p>When looking at the probability at the end of a batch, it is exactly the value I get when predicting with the entire batch (not one by one). However, I've expected that the probability will always continue in the right direction when new samples arrive. What actually happens is that the probability output can spike to the wrong class on an arbitrary sample (see below).<\/p>\n\n<hr>\n\n<p>Two samples of 100 sample batches over the time of prediction (label = 1):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QDINN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QDINN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>and Label = 0:\n<a href=\"https:\/\/i.stack.imgur.com\/oIEbw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oIEbw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there a way to achieve what I want (avoid extreme spikes while predicting probability), or is that a given fact?<\/p>\n\n<p>Any explanation, advice would be appreciated.<\/p>\n\n<hr>\n\n<p><strong>Update<\/strong>\nThanks to @today advice, I've tried training the network with the hidden state output for each input time step using <code>return_sequence=True<\/code> on the last LSTM layer.<\/p>\n\n<p>So now the labels look like so (shape (100,100)):<\/p>\n\n<pre><code>[[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]\n [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\n  1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\n  1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]\n...]\n<\/code><\/pre>\n\n<p>the model summary:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nlstm_1 (LSTM)                (None, 100, 32)           4608      \n_________________________________________________________________\nleaky_re_lu_1 (LeakyReLU)    (None, 100, 32)           0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 100, 32)           0         \n_________________________________________________________________\nlstm_2 (LSTM)                (None, 100, 16)           3136      \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 100, 16)           0         \n_________________________________________________________________\nleaky_re_lu_2 (LeakyReLU)    (None, 100, 16)           0         \n_________________________________________________________________\nlstm_3 (LSTM)                (None, 100, 8)            800       \n_________________________________________________________________\nleaky_re_lu_3 (LeakyReLU)    (None, 100, 8)            0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 100, 1)            9         \n=================================================================\nTotal params: 8,553\nTrainable params: 8,553\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>However, I get an exception:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_1 to have 3 dimensions, but got array with shape (75, 100)\n<\/code><\/pre>\n\n<p>What do I need to fix?<\/p>\n","answers":[{"AnswerId":"53378238","CreationDate":"2018-11-19T15:50:06.677","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><strong>Note: This is just an idea and it might be wrong. Try it if you would like and I would appreciate any feedback.<\/strong><\/p>\n\n<hr>\n\n<blockquote>\n  <p>Is there a way to achieve what I want (avoid extreme spikes while\n  predicting probability), or is that a given fact?<\/p>\n<\/blockquote>\n\n<p>You can do this <strong>experiment<\/strong>: set the <code>return_sequences<\/code> argument of last LSTM layer to <code>True<\/code> and replicate the labels of each sample as much as the length of each sample. For example if a sample has a length of 100 and its label is 0, then create a new label for this sample which consists of 100 zeros (you can probably easily do this using numpy function like <code>np.repeat<\/code>). Then retrain your new model and test it on new samples afterwards. I am not sure of this, but I would expect more monotonically increasing\/decreasing probability graphs this time.<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong> The error you mentioned is caused by the fact that the labels should be a 3D array (look at the output shape of last layer in the model summary). Use <code>np.expand_dims<\/code> to add another axis of size one to the end. The correct way of repeating the labels would look like this, assuming <code>y_train<\/code> has a shape of <code>(num_samples,)<\/code>:<\/p>\n\n<pre><code>rep_y_train = np.repeat(y_train, num_reps).reshape(-1, num_reps, 1)\n<\/code><\/pre>\n\n<hr>\n\n<h2>The experiment on IMDB dataset:<\/h2>\n\n<p>Actually, I tried the experiment suggested above on the IMDB dataset using a simple model with one LSTM layer. One time, I used only <strong>one label per each sample<\/strong> (as in original approach of @Shlomi) and the other time I replicated the labels to have <strong>one label per each timestep of a sample<\/strong> (as I suggested above). Here is the code if you would like to try it yourself:<\/p>\n\n<pre><code>from keras.layers import *\nfrom keras.models import Sequential, Model\nfrom keras.datasets import imdb\nfrom keras.preprocessing.sequence import pad_sequences\nimport numpy as np\n\nvocab_size = 10000\nmax_len = 200\n(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=vocab_size)\nX_train = pad_sequences(x_train, maxlen=max_len)\n\ndef create_model(return_seq=False, stateful=False):\n    batch_size = 1 if stateful else None\n    model = Sequential()\n    model.add(Embedding(vocab_size, 128, batch_input_shape=(batch_size, None)))\n    model.add(CuDNNLSTM(64, return_sequences=return_seq, stateful=stateful))\n    model.add(Dense(1, activation='sigmoid'))\n\n    model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['acc'])\n    return model\n\n# train model with one label per sample\ntrain_model = create_model()\ntrain_model.fit(X_train, y_train, epochs=10, batch_size=128, validation_split=0.3)\n\n# replicate the labels\ny_train_rep = np.repeat(y_train, max_len).reshape(-1, max_len, 1)\n\n# train model with one label per timestep\nrep_train_model = create_model(True)\nrep_train_model.fit(X_train, y_train_rep, epochs=10, batch_size=128, validation_split=0.3)\n<\/code><\/pre>\n\n<p>Then we can create the stateful replicas of the training models and run them on some test data to compare their results:<\/p>\n\n<pre><code># replica of `train_model` with the same weights\ntest_model = create_model(False, True)\ntest_model.set_weights(train_model.get_weights())\ntest_model.reset_states()\n\n# replica of `rep_train_model` with the same weights\nrep_test_model = create_model(True, True)\nrep_test_model.set_weights(rep_train_model.get_weights())\nrep_test_model.reset_states()\n\ndef stateful_predict(model, samples):\n    preds = []\n    for s in samples:\n        model.reset_states()\n        ps = []\n        for ts in s:\n            p = model.predict(np.array([[ts]]))\n            ps.append(p[0,0])\n        preds.append(list(ps))\n    return preds\n\nX_test = pad_sequences(x_test, maxlen=max_len)\n<\/code><\/pre>\n\n<p>Actually, the first sample of <code>X_test<\/code> has a 0 label (i.e. belongs to negative class) and the second sample of <code>X_test<\/code> has a 1 label (i.e. belongs to positive class). So let's first see what the stateful prediction of <code>test_model<\/code> (i.e. the one that were trained using one label per sample) for these two samples would look like:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\n\npreds = stateful_predict(test_model, X_test[0:2])\n\nplt.plot(preds[0])\nplt.plot(preds[1])\nplt.legend(['Class 0', 'Class 1'])\n<\/code><\/pre>\n\n<p>The result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JEtMs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JEtMs.png\" alt=\"&lt;code&gt;test_model&lt;\/code&gt; stateful predictions\"><\/a><\/p>\n\n<p>Correct label (i.e. probability) at the end (i.e. timestep 200) but very spiky and fluctuating in between. Now let's compare it with the stateful predictions of the <code>rep_test_model<\/code> (i.e. the one that were trained using one label per each timestep):<\/p>\n\n<pre><code>preds = stateful_predict(rep_test_model, X_test[0:2])\n\nplt.plot(preds[0])\nplt.plot(preds[1])\nplt.legend(['Class 0', 'Class 1'])\n<\/code><\/pre>\n\n<p>The result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3N8mF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3N8mF.png\" alt=\"&lt;code&gt;rep_test_model&lt;\/code&gt; stateful predictions\"><\/a><\/p>\n\n<p>Again, correct label prediction at the end but this time with a much more smoother and monotonic trend, as expected. <\/p>\n\n<p>Note that this was just an example for demonstration and therefore I have used a very simple model here with just one LSTM layer and I did not attempt to tune it at all. I guess with a better tuning of the model (e.g. adjusting the number of layers, number of units in each layer, activation functions used, optimizer type and parameters, etc.), you might get far better results.<\/p>\n"}]}
+{"QuestionId":53376996,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-19T14:42:13.157","AcceptedAnswerId":53377099.0,"OwnerUserId":10663471.0,"Title":"Cannot concatenate Keras Lambda layers","Body":"<p>I need to process some layers in a different way, doing some OR operations. I've found how to do it, I create a Lambda Layer and process the data with <code>keras.backend.any<\/code>. I am also doing a split, because I need to operate 2 separates groups with my logical OR.<\/p>\n\n<pre><code>def logical_or_layer(x):\n    \"\"\"Processing an OR operation\"\"\"\n    import keras.backend\n    #normalized to 0,1 \n    aux_array = keras.backend.sign(x)\n    aux_array = keras.backend.relu(aux_array)\n    # OR operation\n    aux_array = keras.backend.any(aux_array)\n    # casting back the True\/False to 1,0\n    aux_array = keras.backend.cast(aux_array, dtype='float32')\n\n    return aux_array\n<\/code><\/pre>\n\n<p>Then I'am creating my layers like this:<\/p>\n\n<pre><code>#this is the input tensor\ninputs = Input(shape=(inputSize,))\n\n#this is the Neurule layer\nx = Dense(neurulesQt, activation='softsign')(inputs)\n#after each neurule layer, the outputs need to be put into SIGNUM (-1 or 1)\nx = Lambda(signumTransform, output_shape=lambda x:x, name='signumAfterNeurules')(x)\n\n#separating into 2 (2 possible outputs)\nlayer_split0 = Lambda( lambda x: x[:, :end_output0], output_shape=(11, ), name='layer_split0')(x)\nlayer_split1 = Lambda( lambda x: x[:, start_output1:end_output1], output_shape=(9,), name='layer_split1')(x)\n\n#this is the OR layer\ny_0 = Lambda(logical_or_layer, output_shape=(1,), name='or0')(layer_split0)\ny_1 = Lambda(logical_or_layer, output_shape=(1,), name='or1')(layer_split1)\n<\/code><\/pre>\n\n<p>Just FYI: Neurules are Neurons created based on IF-THEN rules, that's one project to work with Neurons which were trained with a TruthTable, representing Expert Knowledge.<\/p>\n\n<p>Now, when I try to put the splitted layers back like this:<\/p>\n\n<pre><code>y = concatenate([y_0,y_1])\n<\/code><\/pre>\n\n<p>This error comes:<\/p>\n\n<pre><code>ValueError: Can't concatenate scalars (use tf.stack instead) for 'concatenate_32\/concat' (op: 'ConcatV2') with input shapes: [], [], [].\n<\/code><\/pre>\n\n<p>Then ok, let's use the <code>tf.stack<\/code> as suggested:<\/p>\n\n<pre><code>y = keras.backend.stack([y_0, y_1])\n<\/code><\/pre>\n\n<p>Then it can't be used as an output in the Model anymore, when I try:<\/p>\n\n<pre><code>model = Model(inputs=inputs, outputs=y)\n<\/code><\/pre>\n\n<p>Comes the error:<\/p>\n\n<pre><code>ValueError: Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata). Found: Tensor(\"stack_14:0\", shape=(2,), dtype=float32)\n<\/code><\/pre>\n\n<p>Checking with the function <code>keras.backend.is_keras_tensor(y)<\/code> it gives me <code>False<\/code>, but with all the other layers it gives me <code>True<\/code><\/p>\n\n<p>How should I concatenate it correctly?<\/p>\n\n<p><strong>EDIT:<\/strong> Following the answer from @today I was able to create a new Lambda Layer with the <code>stack<\/code> wrapped inside of it. But the outputs are modified, it should be <code>(None,2)<\/code> and it is <code>(2,None,1)<\/code> here is the output from <code>model.summary()<\/code>:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_90 (InputLayer)           (None, 24)           0                                            \n__________________________________________________________________________________________________\ndense_90 (Dense)                (None, 20)           500         input_90[0][0]                   \n__________________________________________________________________________________________________\nsignumAfterNeurules (Lambda)    (None, 20)           0           dense_90[0][0]                   \n__________________________________________________________________________________________________\nlayer_split0 (Lambda)           (None, 11)           0           signumAfterNeurules[0][0]        \n__________________________________________________________________________________________________\nlayer_split1 (Lambda)           (None, 9)            0           signumAfterNeurules[0][0]        \n__________________________________________________________________________________________________\nor0 (Lambda)                    (None, 1)            0           layer_split0[0][0]               \n__________________________________________________________________________________________________\nor1 (Lambda)                    (None, 1)            0           layer_split1[0][0]               \n__________________________________________________________________________________________________\noutput (Lambda)                 (2, None, 1)         0           or0[0][0]                        \n                                                                 or1[0][0]                        \n==================================================================================================\nTotal params: 500\nTrainable params: 0\nNon-trainable params: 500\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>How should I define the output_shape in the layers to have the batch still there at the end?<\/p>\n\n<p><strong>EDIT2:<\/strong> Following the tips from @today I've done the following:<\/p>\n\n<pre><code>#this is the input tensor\ninputs = Input(shape=(inputSize,))\n\n#this is the Neurule layer\nx = Dense(neurulesQt, activation='softsign')(inputs)\n#after each neuron layer, the outputs need to be put into SIGNUM (-1 or 1)\nx = Lambda(signumTransform, output_shape=lambda x:x, name='signumAfterNeurules')(x)\n#separating into 2 (2 possible outputs)\nlayer_split0 = Lambda( lambda x: x[:, :end_output0], output_shape=[11], name='layer_split0')(x)    \nlayer_split1 = Lambda( lambda x: x[:, start_output1:end_output1], output_shape=[9], name='layer_split1')(x)\n#this is the OR layer\ny_0 = Lambda(logical_or_layer, output_shape=(1,), name='or0')(layer_split0)\ny_1 = Lambda(logical_or_layer, output_shape=(1,), name='or1')(layer_split1)\n\ny = Lambda(lambda x: K.stack([x[0], x[1]]),output_shape=(2,),  name=\"output\")([y_0, y_1])\n<\/code><\/pre>\n\n<p>Now it seems to work correctly, the <code>model.summary()<\/code> below:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, 24)           0                                            \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 20)           500         input_1[0][0]                    \n__________________________________________________________________________________________________\nsignumAfterNeurules (Lambda)    (None, 20)           0           dense_1[0][0]                    \n__________________________________________________________________________________________________\nlayer_split0 (Lambda)           (None, 11)           0           signumAfterNeurules[0][0]        \n__________________________________________________________________________________________________\nlayer_split1 (Lambda)           (None, 9)            0           signumAfterNeurules[0][0]        \n__________________________________________________________________________________________________\nor0 (Lambda)                    (None, 1)            0           layer_split0[0][0]               \n__________________________________________________________________________________________________\nor1 (Lambda)                    (None, 1)            0           layer_split1[0][0]               \n__________________________________________________________________________________________________\noutput (Lambda)                 (None, 2)            0           or0[0][0]                        \n                                                                 or1[0][0]                        \n==================================================================================================\nTotal params: 500\nTrainable params: 0\nNon-trainable params: 500\n__________________________________________________________________________________________________\n<\/code><\/pre>\n","answers":[{"AnswerId":"53377099","CreationDate":"2018-11-19T14:48:49.293","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Wrap the <code>K.stack<\/code> inside a <code>Lambda<\/code> layer like this:<\/p>\n\n<pre><code>from keras import backend as K\n\ny = Lambda(lambda x: K.stack([x[0], x[1]]))([y_0, y_1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53377033,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><yolo>","CreationDate":"2018-11-19T14:44:34.490","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"OSError: [WinError 8] Not enough memory resources are available to process this command in yolo","Body":"<p>I am trying to create a project on real time vehicle classification using yolo model. when executing this command<\/p>\n\n<pre><code>python flow --model cfg\/tiny-yolo-voc-1c.cfg --load bin\/tiny-yolo-voc.weights --train`\n<\/code><\/pre>\n\n<p>this error pops up:<\/p>\n\n<pre><code>mm = mmap.mmap(fid.fileno(), bytes, access=acc, offset=start)\nOSError: [WinError 8] Not enough memory resources are available to process this command\n<\/code><\/pre>\n\n<p>Source which I've used to train can be found from <a href=\"https:\/\/github.com\/markjay4k\/YOLO-series\" rel=\"nofollow noreferrer\">https:\/\/github.com\/markjay4k\/YOLO-series<\/a><\/p>\n","answers":[]}
+{"QuestionId":53377681,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-19T15:19:20.323","AcceptedAnswerId":53378186.0,"OwnerUserId":10433093.0,"Title":"Reversing an ML model","Body":"<p>I am currently teaching myself the basics of machine learning by creating a simple image classifier using Keras (with a Tensorflow backend). The model classifies a (greyscaled) image as either a cat or not a cat.<\/p>\n\n<p>My model is relatively good at this task, so I now want to see if it can <strong>generate<\/strong> images that it would classify as a cat.<\/p>\n\n<p>I have attempted to start this in a simple way, by creating a random array of the same shape as the images, with random numbers in each index:<\/p>\n\n<pre><code>    from random import randint\n\n    json_file = open('model.json', 'r')\n    loaded_model_json = json_file.read()\n    json_file.close()\n    model = model_from_json(loaded_model_json)\n    model.load_weights(\"model_weights.h5\")\n\n    confidence = 0.0\n    thresholdConfidence = 0.6\n\n    while confidence &lt; thresholdConfidence:\n        img_array = np.array([[[randint(0, 255) for z in range(1)] for y in range(64)] for x in range(64)])\n        img_array = img_array.reshape((1,) + img_array.shape)\n        confidence = model.predict(img_array)\n<\/code><\/pre>\n\n<p>This method is obviously not good at all, since it just creates random things and could potentially run eternally. Could the model somehow run in reverse by telling it that an array is 100% cat, and having it predict what the array representation of the image is?<\/p>\n\n<p>Thank you for reading.<\/p>\n\n<p>[This is my first post on StackOverflow, so please let me know if I've done something wrong!]<\/p>\n","answers":[{"AnswerId":"53378186","CreationDate":"2018-11-19T15:46:53.370","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>If you wish to generate a special type of image, you can use <strong>Generative Adversary Networks<\/strong>. This are made into two parts which need to be trained separately. The two parts are<\/p>\n\n<ol>\n<li><strong>Generator<\/strong> : Creates noise that is random images.<\/li>\n<li><strong>Discriminator<\/strong> : Gives feedback to the generator regarding the images <\/li>\n<\/ol>\n\n<p>You can refer <a href=\"https:\/\/medium.com\/@jonathan_hui\/gan-whats-generative-adversarial-networks-and-its-application-f39ed278ef09\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53377860,"AnswerCount":0,"Tags":"<python><tensorflow><hessian>","CreationDate":"2018-11-19T15:29:50.190","AcceptedAnswerId":null,"OwnerUserId":8205514.0,"Title":"Diagonal of the Hessian with Tensorflow","Body":"<p>I'm doing some machine learning and I have to deal with a custom loss function. The derivatives and the Hessian of the loss function are difficult to derive and so I've resorted to computing them automatically using Tensorflow.<\/p>\n\n<p>Here is an example.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ny_true = np.array([\n    [1, 0, 0, 0, 0],\n    [0, 1, 0, 0, 0],\n    [0, 0, 0, 1, 0],\n    [0, 0, 1, 0, 0],\n    [0, 0, 0, 0, 1],\n    [0, 0, 0, 0, 1],\n    [0, 0, 0, 0, 1]\n], dtype=float)\n\ny_pred = np.array([\n    [1, 0, 0, 0, 0],\n    [0, 1, 0, 0, 0],\n    [0, 0, 0, 1, 0],\n    [0, 0, 1, 0, 0],\n    [0, 0, 0, 0, 1],\n    [0, 0, 0, 0, 1],\n    [0, 0, 0, 0, 1]\n], dtype=float)\n\nweights = np.array([1, 1, 1, 1, 1], dtype=float)\n\nwith tf.Session():\n\n    # We first convert the numpy arrays to Tensorflow tensors\n    y_true = tf.convert_to_tensor(y_true)\n    y_pred = tf.convert_to_tensor(y_pred)\n    weights = tf.convert_to_tensor(weights)\n\n    # The following code block is a custom loss \n    ys = tf.reduce_sum(y_true, axis=0)\n    y_true = y_true \/ ys\n    ln_p = tf.nn.log_softmax(y_pred)\n    wll = tf.reduce_sum(y_true * ln_p, axis=0)\n    loss = -tf.tensordot(weights, wll, axes=1)\n\n    grad = tf.gradients(loss, y_pred)[0]\n\n    hess = tf.hessians(loss, y_pred)[0]\n    hess = tf.diag_part(hess)\n\n    print(hess.eval())\n<\/code><\/pre>\n\n<p>which prints out<\/p>\n\n<pre><code>[[0.24090069 0.12669198 0.12669198 0.12669198 0.12669198]\n [0.12669198 0.24090069 0.12669198 0.12669198 0.12669198]\n [0.12669198 0.12669198 0.12669198 0.24090069 0.12669198]\n [0.12669198 0.12669198 0.24090069 0.12669198 0.12669198]\n [0.04223066 0.04223066 0.04223066 0.04223066 0.08030023]\n [0.04223066 0.04223066 0.04223066 0.04223066 0.08030023]\n [0.04223066 0.04223066 0.04223066 0.04223066 0.08030023]]\n<\/code><\/pre>\n\n<p>I'm happy with this because it works, the problem is that it doesn't scale. For my use case I only need the diagonal of the Hessian matrix. I've managed to extract it using <code>hess = tf.diag_part(hess)<\/code> but this will still compute the full Hessian, which is unnecessary. The overhead is so bad that I can't use it for moderate sized datasets (~100k rows).<\/p>\n\n<p>My question is thus: <strong>is there any better way to extract the diagonal of the Hessian<\/strong>? I'm well aware of this <a href=\"https:\/\/stackoverflow.com\/questions\/38200982\/how-to-compute-all-second-derivatives-only-the-diagonal-of-the-hessian-matrix\">post<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/40101165\/tensorflow-compute-hessian-matrix-only-diagonal-part-with-respect-to-a-high-r\">this one<\/a> but I don't find the answers good enough.<\/p>\n","answers":[]}
+{"QuestionId":53378045,"AnswerCount":0,"Tags":"<python><unity3d><tensorflow><object-detection><tensorflowsharp>","CreationDate":"2018-11-19T15:39:14.980","AcceptedAnswerId":null,"OwnerUserId":6382445.0,"Title":"Freezing fine-tuned graph for TensorFlowSharp with TF 1.4","Body":"<p>I've fine-tuned a model (using TF 1.9) from  <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">Object Detection Zoo Model<\/a> and right now I am trying to freeze the graph for TensorFlowSharp using TF 1.9. <\/p>\n\n<pre><code>import tensorflow as tf\nimport os\nfrom tensorflow.python.tools import freeze_graph\nfrom tensorflow.core.protobuf import saver_pb2\n\n#print(\"current tensorflow version: \", tf.version)\nsess=tf.Session()\n\nmodel_path = 'latest_cp\/'\nsaver = tf.train.import_meta_graph('model.ckpt.meta')\nsaver.restore(sess,tf.train.latest_checkpoint('.')) #current dir of the checkpoint file\ntf.train.write_graph(sess.graph_def, '.', 'test.pbtxt') #output in pbtxt format\n\nfreeze_graph.freeze_graph(input_graph = 'test.pbtxt',\n    input_binary = False,\n    input_checkpoint = model_path + 'model.ckpt',\n    output_node_names =     \"num_detections,detection_boxes,detection_scores,detection_classes\",\n    output_graph = 'test.bytes' ,\n    clear_devices = True, initializer_nodes = \"\",input_saver = \"\",\n    restore_op_name = \"save\/restore_all\", filename_tensor_name = \"save\/Const:0\")\n<\/code><\/pre>\n\n<p>It worked but then after I imported it to Unity it returned the following error:<\/p>\n\n<p><code>TFException: Op type not registered 'NonMaxSuppressionV3' in binary running on AK38713. Make sure the Op and Kernel are registered in the binary running in this process.<\/code><\/p>\n\n<p>I find out that TensorFlowSharp works with TensorFlow 1.4 and when I tried to freeze graph with 1.4 it returns the same <code>NonMaxSuppressionV3<\/code> error.\nDo you know any way to solve this issue? Thank you so much for the support.<\/p>\n","answers":[]}
+{"QuestionId":53378242,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-platform><google-cloud-ml>","CreationDate":"2018-11-19T15:50:16.720","AcceptedAnswerId":53379881.0,"OwnerUserId":2383106.0,"Title":"No module named trainer, Cloud ML Engine for TensorFlow Tutorial, Running Locally","Body":"<p><a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/getting-started-training-prediction#local-train-single\" rel=\"nofollow noreferrer\">Tutorial I am following<\/a><\/p>\n\n<p>Have been attempting to follow the Google tutorial to use ML Engine for TensorFlow. Have gotten stuck where it says \"run a local training job\" with the error <\/p>\n\n<pre><code>\/usr\/bin\/python: No module named trainer\n<\/code><\/pre>\n\n<p>Full command is:<\/p>\n\n<pre><code>gcloud ml-engine local train \\\n    --module-name trainer.task \\\n    --package-path trainer\/ \\\n    --job-dir $MODEL_DIR \\\n    -- \\\n    --train-files $TRAIN_DATA \\\n    --eval-files $EVAL_DATA \\\n    --train-steps 1000 \\\n    --eval-steps 100\n<\/code><\/pre>\n\n<p>The three variables are all set up correctly to my knowledge though it doesn't even get to them right now. The tutorial doesn't specify downloading a trainer file or how it is referenced, googling for the past hour hasn't turned up any working solutions. Have found this general explanation:<\/p>\n\n<blockquote>\n  <p>--module-name specifies the name of your application's main module,\n  using your package's namespace dot notation. This is the Python file\n  that you run to start your application. For example, if your main\n  module is ...\/my_application\/trainer\/task.py (see the recommended\n  project structure), then the module name is trainer.task<\/p>\n<\/blockquote>\n\n<p>Any info would be appreciated.<\/p>\n","answers":[{"AnswerId":"53379881","CreationDate":"2018-11-19T17:33:07.290","ParentId":null,"OwnerUserId":"2383106","Title":null,"Body":"<p>Thanks to the information from Dustin in the comments have found the solution. <\/p>\n\n<p>Was simply downloading the trainer directory on the Cloud Shell and not my local environment.<\/p>\n\n<p>Now File Structure Looks Like:<\/p>\n\n<pre><code>estimator\n    |-- data\n    |   |-- adult_data.csv\n    |   |-- adult_test.csv\n    |-- output\n    |-- trainer\n    |   |-- __init__.py\n    |   |-- model.py\n    |   |-- task.ipynb\n    |   |-- task.py\n<\/code><\/pre>\n"}]}
+{"QuestionId":53378284,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><probability><reinforcement-learning>","CreationDate":"2018-11-19T15:52:42.000","AcceptedAnswerId":53378514.0,"OwnerUserId":8954691.0,"Title":"Return distribution over set of action space from Neural Network","Body":"<p>I am trying to build a neural network to output a probabilistic distribution over set of whole action space. <\/p>\n\n<p>My action space is a vector of 3 individual actions : <code>[a,b,c]<\/code> <\/p>\n\n<p><code>a<\/code> can have 3 possible actions within itself <code>a1,a2,a3<\/code> and similarly <code>b<\/code> has <code>b1,b2,b3<\/code>, <code>c<\/code> has <code>c1,c2,c3<\/code>. So In total i can have 27 different combinations of these actions <code>3^3 = 27<\/code>. Ultimately the neural network should output 27 combinations of these actions (which is a matrix of 27 x 3) : <code>[[a1,b1,c1],[a2,b2,c2],[a3,b3,c3],[a1,b1,c2],[a1,b1,c3],.....]<\/code> and so on for all 27 combinations. Just to mention the input to my network is a state which is a vector of 5 elements. <\/p>\n\n<p>I want a probability associated to each of these 27 combinations. <\/p>\n\n<p>I know I can associate probability by using  a softmax with 27 outputs but I don't understand how the network can output a matrix in this case where every row has a probability associated to it. <\/p>\n","answers":[{"AnswerId":"53378514","CreationDate":"2018-11-19T16:04:20.053","ParentId":null,"OwnerUserId":"8083699","Title":null,"Body":"<p>Is there any reason you want it to return a matrix of these actions? Why not just map each of the 27 combinations to integers 0-26? So your architecture could look like <code>[Linear(5, n), ReLU, Linear(n, .) ... Softmax(Linear(., 27))]<\/code>. Then when you need to evaluate, you can just map it back to the action sequence. This is similar to how in NLP tasks you map multidimensional word vectors to integers via <code>stoi<\/code> for training and bring them back via <code>itos<\/code>.<\/p>\n\n<p>I should point out that if your training paradigm involves some further use of these discrete choices (say you use the <code>argmax<\/code> of this upstream of another net), then the nondifferentiability of <code>argmax<\/code> means that this architechture will not learn anything. I only mention this because you use the phrase \"action space\", which is typical in DRL. If this is the case, you may want to consider algorithms like REINFORCE where action sequences can be learned discretely and used upstream via policy gradient.<\/p>\n"}]}
+{"QuestionId":53378515,"AnswerCount":1,"Tags":"<python><keras><training-data>","CreationDate":"2018-11-19T16:04:24.693","AcceptedAnswerId":53379239.0,"OwnerUserId":6421394.0,"Title":"Epochs and batches control in Keras","Body":"<p>I would like to implement an autoencoder model that acts as following:<\/p>\n\n<pre><code>for epoch in xrange(100):\n  for X_batch in batch_list:\n     model.train_on_batch(X_batch, X_batch)\n     training_error = model.evaluate(X_batch, X_batch, verbose=0)\n  average the training error by the number of the batches considered\n  save it as the epoch training error\n  call the function to get the validation error in the same fashion over  the validation data\n  compare the two errors and decide whether go on training or stopping\n<\/code><\/pre>\n\n<p>I have looked on Internet and already asked something and it has been suggested me to use fit_generator, but I have not understand how to implement it. Or should I use the method train_on_batch or the fit one with number of epochs equal to 1 to proper fit the model?<\/p>\n\n<p>Which is the best practice in this case? Do you have an example or a similar question to link me?<\/p>\n","answers":[{"AnswerId":"53379239","CreationDate":"2018-11-19T16:49:02.930","ParentId":null,"OwnerUserId":"6223135","Title":null,"Body":"<p>From what I can understand, you want to use validation error as early stopping criteria. Good news is that keras already has early stopping callback. SO all you need is to create a callback and call it during training after some epochs\/iterations.<\/p>\n\n<pre><code>keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=0, verbose=0, mode='auto', baseline=None, restore_best_weights=False)\n<\/code><\/pre>\n\n<p>Let us look at train_on_batch and fit()<\/p>\n\n<pre><code>train_on_batch(x, y, sample_weight=None, class_weight=None)\n\n\nfit(x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)\n<\/code><\/pre>\n\n<p>You can see that train_on_batch doesn't take any callback as input, so good choice is to use fit here, unless you want to implement it by yourself.<\/p>\n\n<p>Now you can call fit as following<\/p>\n\n<pre><code>callbacks = [EarlyStopping(monitor='val_loss', patience=2),\n         ModelCheckpoint(filepath='path to latest ckpt', monitor='val_loss', save_best_only=True)]\n\nhistory = model.fit(train_features,train_target, epochs=num_epochs, callbacks=callbacks, verbose=0, batch_size=your_choice, validation_data) \n<\/code><\/pre>\n"}]}
+{"QuestionId":53378990,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-11-19T16:32:30.310","AcceptedAnswerId":null,"OwnerUserId":10443847.0,"Title":"Python: problem of using categorical crosses trophy in CNN to classify multiple object","Body":"<p>I am using A.I. algorithm sample from github to study image recognition and processing. 600 jpeg images for 4 dota hero classes are imported and resize to 64*64 pixels. 400 images are training set, and 200 images are testing set for this code. The problem is the python program output give me: this dota II hero is sven and also give me the option five: this is a non dota hero,  where the image that putting in to test is the image of mirana. Here is part of my code:<\/p>\n\n<pre><code>classifier.add(Dense(units = 128, activation = 'relu'))\nclassifier.add(Dense(units = 4, activation = 'softmax'))\n\n# Compiling the CNN\n classifier.compile(optimizer = 'adam', loss =                         \n'categorical_crossentropy', metrics = ['accuracy'])\n\n# Part 2 - Fitting the CNN to the images\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\nshear_range = 0.2,\nzoom_range = 0.2,\nhorizontal_flip = True)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\ntraining_set = train_datagen.flow_from_directory('dataset\/dota_training_set',\ntarget_size = (64, 64),\nbatch_size = 32,\nclass_mode = 'categorical')\ntest_set = test_datagen.flow_from_directory('dataset\/dota_test_set',\ntarget_size = (64, 64),\nbatch_size = 32,\nclass_mode = 'categorical')\nclassifier.fit_generator(training_set,\nsteps_per_epoch = 1000,\nepochs = 3,\nvalidation_data = test_set,\nvalidation_steps = 500)\n\n# Part 3 - Making new predictions\nimport numpy as np\nfrom keras.preprocessing import image\ntest_image = image.load_img('dataset\/single_prediction\/test_2.jpg', target_size = (64, 64))\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis = 0)\nresult = classifier.predict(test_image)\ntraining_set.class_indices\nif result[0][0] == 1.0:\n    prediction = 'sven'\n    print(\"This is a Sven.\")\nif result[0][0] == 2.0:\n    prediction = 'mirana'\n    print(\"This is a mirana.\")\nif result[0][0] == 3.0:\n    prediction = 'Wraith_King'\n    print(\"This is a Wraith King.\")\nif result[0][0] == 4.0:\n    prediction = 'Phantom_Lancer'\n    print(\"This is a Phantom Lancer.\")\nelse:\n    prediction = 'non_dota hero'\n    print(\"This is a non dota hero.\")\n<\/code><\/pre>\n\n<p>I suspect is it the last part if and else problem. How I able to correct my problem, so the program can provide the correct answer. Thank You.<\/p>\n","answers":[{"AnswerId":"53379037","CreationDate":"2018-11-19T16:36:08.280","ParentId":null,"OwnerUserId":"10067173","Title":null,"Body":"<p>Your else block is only an <code>else<\/code> for the last <code>if<\/code>.<\/p>\n\n<pre><code>if result[0][0] == 4.0:\n    prediction = 'Phantom_Lancer'\n    print(\"This is a Phantom Lancer.\")\nelse:\n    prediction = 'non_dota hero'\n    print(\"This is a non dota hero.\")\n<\/code><\/pre>\n\n<p>Change the other <code>ifs<\/code> to <code>elifs<\/code>:<\/p>\n\n<pre><code>if result[0][0] == 1.0:\n    prediction = 'sven'\n    print(\"This is a Sven.\")\nelif result[0][0] == 2.0:\n    prediction = 'mirana'\n    print(\"This is a mirana.\")\nelif result[0][0] == 3.0:\n    prediction = 'Wraith_King'\n    print(\"This is a Wraith King.\")\nelif result[0][0] == 4.0:\n    prediction = 'Phantom_Lancer'\n    print(\"This is a Phantom Lancer.\")\nelse:\n    prediction = 'non_dota hero'\n    print(\"This is a non dota hero.\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":53379173,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-11-19T16:45:14.830","AcceptedAnswerId":null,"OwnerUserId":4479760.0,"Title":"Keras\/TensorFlow No registered NGraphVariable Opkernel","Body":"<p><strong>Output error<\/strong>:<\/p>\n\n<p><code>E tensorflow\/core\/framework\/op_segment.cc:54] Create kernel failed: Not found: No registered 'NGraphVariable' OpKernel for GPU devices compatible with node {{node lstm_1\/kernel}} = NGraphVariable[_class=[\"loc:@IsVariableInitialized\"], container=\"\", dtype=DT_FLOAT, just_looking=false, shape=[1,80], shared_name=\"\", _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()\n    .  Registered:  device='CPU'\n<\/code><\/p>\n\n<h3>Environment:<\/h3>\n\n<ol>\n<li>Sample <a href=\"https:\/\/machinelearningmastery.com\/develop-bidirectional-lstm-sequence-classification-python-keras\/\" rel=\"nofollow noreferrer\">BiLSTM python code<\/a><\/li>\n<li>CUDA 9.2, CuDNN 7.4, Python 3.6, TensorFlow GPU 1.12, Keras 2.2.4<\/li>\n<li>Ubuntu 18.10 with GTX1080 graphics<\/li>\n<\/ol>\n\n<h3>Question:<\/h3>\n\n<p>Is this error solved by compiling <a href=\"https:\/\/github.com\/NervanaSystems\/ngraph\" rel=\"nofollow noreferrer\">NervanaSystems\/ngraph<\/a> from source?<\/p>\n","answers":[]}
+{"QuestionId":53379681,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T17:17:42.320","AcceptedAnswerId":null,"OwnerUserId":7551149.0,"Title":"How to run two tensorflow session on the same GPU in parallel?","Body":"<p>Is it possible to run two sessions on the same GPU concurrently? I am very new to Tensorflow and don't know how to do that.<\/p>\n","answers":[]}
+{"QuestionId":53379887,"AnswerCount":2,"Tags":"<python><python-3.x><deep-learning><conv-neural-network><pytorch>","CreationDate":"2018-11-19T17:33:29.837","AcceptedAnswerId":null,"OwnerUserId":8832082.0,"Title":"load test data in pytorch","Body":"<p>All is in the title, I just want to know, how can I load my own test data (image.jpg) in pytorch in order to test my CNN.<\/p>\n","answers":[{"AnswerId":"53388516","CreationDate":"2018-11-20T07:57:00.633","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to feed images to net the same as in training: that is, you should apply exactly the same transformations to get similar results.  <\/p>\n\n<p>Assuming your net was trained using <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/imagenet\/main.py\" rel=\"nofollow noreferrer\">this code<\/a> (or similar), you can see that an input image (for validation) undergoes the following <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/imagenet\/main.py#L210-L215\" rel=\"nofollow noreferrer\">transformations<\/a>:<\/p>\n\n<blockquote>\n<pre><code>transforms.Compose([\n            transforms.Resize(256),\n            transforms.CenterCrop(224),\n            transforms.ToTensor(),\n            normalize,\n        ])),\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Following <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/transforms.html\" rel=\"nofollow noreferrer\"><code>torchvision.transforms<\/code><\/a> docs you can see that an input image goes through:<\/p>\n\n<ul>\n<li>Resizing to 256x256 pix  <\/li>\n<li>Cropping 224x224 rect from the center of the image<\/li>\n<li>The image is converted from uint8 datatype to float in range <code>[0, 1]<\/code>, and transposed to 3-by-224-by-224 array  <\/li>\n<li>The image is <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/imagenet\/main.py#L188-l189\" rel=\"nofollow noreferrer\">normalize<\/a> by subtracting mean and dividing by std.<\/li>\n<\/ul>\n\n<p>You can do all this manually to any image<\/p>\n\n<pre><code>import numpy as np\nfrom PIL import Image\n\npil_img = Image.open('image.jpg').resize((256, 256), Image.BILINEAR)  # read and resize\n# center crop\nw, h = pil_img.size\ni = int(round((h - 224) \/ 2.))\nj = int(round((w - 224) \/ 2.))\npil_img = pil_img.crop((j, i, j+224, i+224))\nnp_img = np.array(pil_img).astype(np.float32) \/ 255.\nnp_img = np.transpose(np_img, (2, 0, 1))  \n# normalize\nmean = [0.485, 0.456, 0.406]\nstd = [0.229, 0.224, 0.225]\nfor c in range(3):\n  np_img = (np_img[c, ...] - mean[c]) \/ std[c]\n<\/code><\/pre>\n\n<p>Once you have <code>np_img<\/code> ready for your model, you can run a feed forward pass:<\/p>\n\n<pre><code>pred = model(np_img[None, ...])  # note that we add a singleton leading dim for batch\n<\/code><\/pre>\n"},{"AnswerId":"53408216","CreationDate":"2018-11-21T08:47:20.513","ParentId":null,"OwnerUserId":"8832082","Title":null,"Body":"<p>thanks for your response. My problem was loading test data and <strong>I found a solution<\/strong>.<\/p>\n\n<pre><code>test_data = datasets.ImageFolder('root\/test_cnn', transform=transform)\n<\/code><\/pre>\n\n<p>For example if I have 2 directories cat &amp; dog (in the test_cnn directory) that contains images, the Object ImageFolder will assign automatically the classes cat and dog to my images.<\/p>\n\n<p>During testing, I have just to drop the classes.<\/p>\n"}]}
+{"QuestionId":53379902,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T17:34:05.583","AcceptedAnswerId":null,"OwnerUserId":4958349.0,"Title":"Fetch argument 0.0 has invalid type <class 'numpy.float32'>, must be a string or Tensor. (Can not convert a float32 into a Tensor or Operation.)","Body":"<p>i wrote a bi-lstm language model, and i used a <code>adaptive softmax<\/code> in <a href=\"https:\/\/github.com\/yangsaiyong\/tf-adaptive-softmax-lstm-lm\/blob\/master\/train_lm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yangsaiyong\/tf-adaptive-softmax-lstm-lm\/blob\/master\/train_lm.py<\/a><\/p>\n\n<pre><code>def train(train_data, vocabulary_size, args):   \n\n    with tf.Session() as sess:\n\n        model = BiRNNLanguageModel(vocabulary_size, args)\n        sess.run(tf.global_variables_initializer())\n\n        saver = tf.train.Saver()\n\n        ckpt = tf.train.get_checkpoint_state(checkpoint_dir=args.model_path)\n        if ckpt and ckpt.model_checkpoint_path:\n            saver.restore(sess, ckpt.model_checkpoint_path)\n            step = int(ckpt.model_checkpoint_path.rsplit('-', 1)[1])\n\n\n        def train_step(batch_x):\n            feed_dict = {model.x: batch_x, model.keep_prob: args.keep_prob}\n\n            step, loss, _ = sess.run([model.global_steps, model.cost, model.train_op], feed_dict=feed_dict)\n\n            if step % 100 == 0:\n                print(\"step {0}: loss = {1}\".format(step, loss))\n                saver.save(sess, args.model_path+'bilstm-model', global_step=step)\n\n        batches = batch_iter(train_data, args.batch_size, args.num_epochs)\n        for batch_x in batches:\n            train_step(batch_x)\n            step = tf.train.global_step(sess, model.global_steps)\n<\/code><\/pre>\n\n<p>But when i want to store the model, the error just happened below<\/p>\n\n<pre><code>TypeError: Fetch argument 0.0 has invalid type &lt;class 'numpy.float32'&gt;, must be a string or Tensor. (Can not convert a float32 into a Tensor or Operation.)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53380010,"AnswerCount":1,"Tags":"<python><python-3.x><python-2.7><tensorflow>","CreationDate":"2018-11-19T17:41:51.777","AcceptedAnswerId":53380055.0,"OwnerUserId":2723224.0,"Title":"Will a tensorflow model give different results if trained in a previous version","Body":"<p>I have a tensorflow model trained in tf version 1.1 and python 2.7. I'm running it in tf version 1.10 and python 3.6. After feeding the same data into the model in both versions, the outputs are slightly different. Of 512 outputs, 7 do not match. Is this a versioning issue, or something else? I appreciate your help. <\/p>\n","answers":[{"AnswerId":"53380055","CreationDate":"2018-11-19T17:44:58.123","ParentId":null,"OwnerUserId":"5202571","Title":null,"Body":"<p>Since your Tensorflow code may contain internal random number generator, you might check if you forgot to seed the random generator with the tf.random.set_random_seed(seed) function.<\/p>\n"}]}
+{"QuestionId":53380052,"AnswerCount":0,"Tags":"<python><keras><neural-network><classification><autoencoder>","CreationDate":"2018-11-19T17:44:52.383","AcceptedAnswerId":null,"OwnerUserId":2651073.0,"Title":"How to give the encoded output of an encoder to the inputs of a classifier?","Body":"<p>I am trying to use an <code>AutoEncoder<\/code> to enhance the performance of a classifier.<\/p>\n\n<p>This is the <code>AutoEncoder<\/code> I wrote using the keras documentation for MNIST data:<\/p>\n\n<pre><code>from keras.layers import Input, Dense\nfrom keras.models import Model\n\n# this is the size of our encoded representations\n# 32 floats -&gt; compression of factor 24.5, assuming the input is 784 floats\nencoding_dim = 32\n\n# this is our input placeholder\ninput_img = Input(shape=(784,))\n# \"encoded\" is the encoded representation of the input\nencoded = Dense(encoding_dim, activation='relu')(input_img)\n# \"decoded\" is the lossy reconstruction of the input\ndecoded = Dense(784, activation='sigmoid')(encoded)\n\n# this model maps an input to its reconstruction\nautoencoder = Model(input_img, decoded)\n# this model maps an input to its encoded representation\nencoder = Model(input_img, encoded)\n# create a placeholder for an encoded (32-dimensional) input\nencoded_input = Input(shape=(encoding_dim,))\n# retrieve the last layer of the autoencoder model\ndecoder_layer = autoencoder.layers[-1]\n# create the decoder model\ndecoder = Model(encoded_input, decoder_layer(encoded_input))\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\n\nautoencoder.fit(x_train, x_train,\n                epochs=50,\n                batch_size=256,\n                shuffle=True,\n                validation_data=(x_test, x_test))\n<\/code><\/pre>\n\n<p>Now I'm trying to connect it to a classifier. Using <a href=\"https:\/\/stackoverflow.com\/questions\/49565638\/how-to-initialise-weights-of-a-mlp-using-an-autoencoder-2nd-part-deep-autoenc\">this question<\/a>, I wrote:<\/p>\n\n<pre><code>x = encoder.output\n# h = Dense(3, activation='relu', name='hidden')(x)\ny = Dense(1, activation='sigmoid', name='predictions')(x)\n\nclassifier = Model(inputs=autoencoder.inputs, outputs=y)\n\n\n# Compile model\nclassifier.compile(loss='binary_crossentropy', optimizer='adam',\n                   metrics=['accuracy'])\n\n# Fit the model\nhistory = classifier.fit(x_train, y_train, \n                         epochs=10, \n                         batch_size=10,\n                         validation_split=.1)\n<\/code><\/pre>\n\n<p>First, I don't understand this code. Second, I guess the output of <code>y<\/code> must be 10 corresponding of 10 digits, but I can't set it to 10 because I get an error.<\/p>\n\n<p>Anyway, the accuracy of the above classifier is very low (10%)! Is there anything wrong in my approach?<\/p>\n","answers":[]}
+{"QuestionId":53380068,"AnswerCount":1,"Tags":"<gpu><pytorch><google-colaboratory>","CreationDate":"2018-11-19T17:46:00.857","AcceptedAnswerId":53773846.0,"OwnerUserId":3131604.0,"Title":"Pytorch on google-colaboratory GPU - Illegal memory access","Body":"<p>I am using pytorch(0.4.0) on google-colaboratory ( NVIDIA-SMI 396.44                 Driver Version: 396.44)<\/p>\n\n<p>When running my code outside any function, I am able to send pytorch tensors and model to the GPU :<\/p>\n\n<pre><code>...\nmodel.cuda()\ndata_tensor = data_tensor.cuda()\n...\n<\/code><\/pre>\n\n<p>And my CNN model is trained successfully with 98% accurancy.<\/p>\n\n<p>But when I put the same code in a function,<\/p>\n\n<pre><code>def main(...):\n    ....\n    model.cuda()\n    data_tensor= data_tensor.cuda()\n    ...\n\nif __name__ == \"__main__\":\n    main('...)\n<\/code><\/pre>\n\n<p>I have the following error:<\/p>\n\n<pre><code>cuda runtime error (77) : an illegal memory access was encountered at \/pytorch\/aten\/src\/THC\/generic\/THCTensorCopy.c:20\n<\/code><\/pre>\n\n<p>UPDATE(18\/11\/21):<\/p>\n\n<p>It turned out that being part or not of a function is irrelevant. Usually, I have first a CUDNN_STATUS_EXECUTION_FAILED error then the second time a  cuda runtime error (77) as shown below. But it sometimes works a few times before failing.<\/p>\n\n<p>CUDNN_STATUS_EXECUTION_FAILED (first try) :<\/p>\n\n<pre><code>RuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-27-53476e08e017&gt; in &lt;module&gt;()\n      1 main('mnist', 'to', 'ndd', Xd=16, epo=5, bs=100, tXn=-1, vXn=300,\n----&gt; 2      lr=0.05, suf=\"s1\", n_class=10, cuda=True)\n\n&lt;ipython-input-23-918584456207&gt; in main(ds, framework, format, Xd, epo, bs, tXn, vXn, lr, suf, n_class, cuda)\n     12     opt = torch.optim.SGD(net.parameters(), lr)\n     13 \n---&gt; 14     train(net, opt, Xd, epo, bs, cuda, tXn, tX, tT, vX, vT,lr)\n     15 \n\n&lt;ipython-input-26-6b574a9e8af6&gt; in train(model, optimizer, Xd, epo, bs, cuda, Xn, tX, tT, vX, vT, lr)\n     26             #t = t.cuda()\n     27             optimizer.zero_grad()\n---&gt; 28             z = model(x)\n     29             bat_loss = criterion(z, t)\n     30             bat_loss.backward()\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in __call__(self, *input, **kwargs)\n    489             result = self._slow_forward(*input, **kwargs)\n    490         else:\n--&gt; 491             result = self.forward(*input, **kwargs)\n    492         for hook in self._forward_hooks.values():\n    493             hook_result = hook(self, input, result)\n\n&lt;ipython-input-22-b4bc2e0b39b8&gt; in forward(self, X)\n     10         H0 = torch.zeros(self.n_H, X.size(0), self.Wh)\n     11         C0 = torch.zeros(self.n_H, X.size(0), self.Wh)\n---&gt; 12         O, (Hn, Cn), = self.lstm1(X, (H0, C0))\n     13         O = self.linear1(O[:, -1, :])\n     14         return O\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in __call__(self, *input, **kwargs)\n    489             result = self._slow_forward(*input, **kwargs)\n    490         else:\n--&gt; 491             result = self.forward(*input, **kwargs)\n    492         for hook in self._forward_hooks.values():\n    493             hook_result = hook(self, input, result)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/rnn.py in forward(self, input, hx)\n    190             flat_weight=flat_weight\n    191         )\n--&gt; 192         output, hidden = func(input, self.all_weights, hx, batch_sizes)\n    193         if is_packed:\n    194             output = PackedSequence(output, batch_sizes)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/_functions\/rnn.py in forward(input, *fargs, **fkwargs)\n    321             func = decorator(func)\n    322 \n--&gt; 323         return func(input, *fargs, **fkwargs)\n    324 \n    325     return forward\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/_functions\/rnn.py in forward(input, weight, hx, batch_sizes)\n    285             batch_first, dropout, train, bool(bidirectional),\n    286             list(batch_sizes.data) if variable_length else (),\n--&gt; 287             dropout_ts)\n    288 \n    289         if cx is not None:\n\nRuntimeError: CUDNN_STATUS_EXECUTION_FAILED\n<\/code><\/pre>\n\n<p>cuda runtime error (77)  (other tries):<\/p>\n\n<pre><code>RuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-28-53476e08e017&gt; in &lt;module&gt;()\n      1 main('mnist', 'to', 'ndd', Xd=16, epo=5, bs=100, tXn=-1, vXn=300,\n----&gt; 2      lr=0.05, suf=\"s1\", n_class=10, cuda=True)\n\n&lt;ipython-input-23-918584456207&gt; in main(ds, framework, format, Xd, epo, bs, tXn, vXn, lr, suf, n_class, cuda)\n     12     opt = torch.optim.SGD(net.parameters(), lr)\n     13 \n---&gt; 14     train(net, opt, Xd, epo, bs, cuda, tXn, tX, tT, vX, vT,lr)\n     15 \n\n&lt;ipython-input-26-6b574a9e8af6&gt; in train(model, optimizer, Xd, epo, bs, cuda, Xn, tX, tT, vX, vT, lr)\n      4     if cuda and torch.cuda.is_available():\n      5         print(\"tX type (before):\", tX.type())\n----&gt; 6         model.cuda()\n      7         tX = tX.cuda()\n      8         tT = tT.cuda()\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in cuda(self, device)\n    247             Module: self\n    248         \"\"\"\n--&gt; 249         return self._apply(lambda t: t.cuda(device))\n    250 \n    251     def cpu(self):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in _apply(self, fn)\n    174     def _apply(self, fn):\n    175         for module in self.children():\n--&gt; 176             module._apply(fn)\n    177 \n    178         for param in self._parameters.values():\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/rnn.py in _apply(self, fn)\n    109 \n    110     def _apply(self, fn):\n--&gt; 111         ret = super(RNNBase, self)._apply(fn)\n    112         self.flatten_parameters()\n    113         return ret\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in _apply(self, fn)\n    180                 # Tensors stored in modules are graph leaves, and we don't\n    181                 # want to create copy nodes, so we have to unpack the data.\n--&gt; 182                 param.data = fn(param.data)\n    183                 if param._grad is not None:\n    184                     param._grad.data = fn(param._grad.data)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py in &lt;lambda&gt;(t)\n    247             Module: self\n    248         \"\"\"\n--&gt; 249         return self._apply(lambda t: t.cuda(device))\n    250 \n    251     def cpu(self):\n\nRuntimeError: cuda runtime error (77) : an illegal memory access was encountered at \/pytorch\/aten\/src\/THC\/generic\/THCTensorCopy.c:20\n<\/code><\/pre>\n","answers":[{"AnswerId":"53773846","CreationDate":"2018-12-14T05:14:08.863","ParentId":null,"OwnerUserId":"3131604","Title":null,"Body":"<p>It now works with Pytorch 1.0 using:<\/p>\n\n<pre><code>!pip3 install https:\/\/download.pytorch.org\/whl\/cu80\/torch-1.0.0-cp36-cp36m-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
+{"QuestionId":53380386,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-19T18:09:39.617","AcceptedAnswerId":53405372.0,"OwnerUserId":10675469.0,"Title":"tensorflow-serving signature for an XOR","Body":"<p>I am trying to export my first xor NN using tensorflow serving but I am not getting any result when I call the gRPC.\nHere the code I use to predict the XOR<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\nfrom keras import backend as K\nK.set_session(sess)\nK.set_learning_phase(0)  # all new operations will be in test mode from now on\n\nfrom tensorflow.python.saved_model import builder as saved_model_builder\nfrom tensorflow.python.saved_model import tag_constants, signature_constants, signature_def_utils_impl\n\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation\nfrom keras.optimizers import SGD\nimport numpy as np\n\nmodel_version = \"2\" #Change this to export different model versions, i.e. 2, ..., 7\nepoch = 100 ## the higher this number is the more accurate the prediction will be 10000 is a good number to s\net it at just takes a while to train\n\n#Exhaustion of Different Possibilities\nX = np.array([\n    [0,0],\n    [0,1],\n    [1,0],\n    [1,1]\n])\n\n#Return values of the different inputs\nY = np.array([[0],[1],[1],[0]])\n\n#Create Model\nmodel = Sequential()\nmodel.add(Dense(8, input_dim=2))\nmodel.add(Activation('tanh'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\nsgd = SGD(lr=0.1)\n\nmodel.compile(loss='binary_crossentropy', optimizer=sgd)\nmodel.fit(X, Y, batch_size=1, nb_epoch=epoch)\n\ntest = np.array([[0.0,0.0]])\n\n#setting values for the sake of saving the model in the proper format\nx = model.input\ny = model.output\n\nprint('Results of Model', model.predict_proba(X))\n\nprediction_signature = tf.saved_model.signature_def_utils.predict_signature_def({\"inputs\": x}, {\"prediction\":\ny})\n\nvalid_prediction_signature = tf.saved_model.signature_def_utils.is_valid_signature(prediction_signature)\nif(valid_prediction_signature == False):\n    raise ValueError(\"Error: Prediction signature not valid!\")\n\nbuilder = saved_model_builder.SavedModelBuilder('.\/'+model_version)\nlegacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')\n\n# Add the meta_graph and the variables to the builder\nbuilder.add_meta_graph_and_variables(\n      sess, [tag_constants.SERVING],\n      signature_def_map={\n           signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:prediction_signature,\n      },\n      legacy_init_op=legacy_init_op)\n\n# save the graph\nbuilder.save()\n<\/code><\/pre>\n\n<p>Implement in docker<\/p>\n\n<pre><code>docker run -p 8501:8501 --mount type=bind,source=\/root\/tensorflow3\/projects\/example\/xor_keras_tensorflow_serving,target=\/models\/xor -e MODEL_NAME=xor -t tensorflow\/serving &amp;\n<\/code><\/pre>\n\n<p>then I request the prediction by the following:<\/p>\n\n<pre><code>curl -d '{\"inputs\": [1,1]}' -X POST http:\/\/localhost:8501\/v2\/models\/xor\n<\/code><\/pre>\n\n<p>the result is always this<\/p>\n\n<pre><code>&lt;HTML&gt;&lt;HEAD&gt;\n&lt;TITLE&gt;404 Not Found&lt;\/TITLE&gt;\n&lt;\/HEAD&gt;&lt;BODY&gt;\n&lt;H1&gt;Not Found&lt;\/H1&gt;\n&lt;\/BODY&gt;&lt;\/HTML&gt;\n<\/code><\/pre>\n\n<p>Can you help me to find where I wrong?\nI have tried to change \"inputs\" in the curl with \"instances\", but nothing\nThanks,\nManuel<\/p>\n","answers":[{"AnswerId":"53405372","CreationDate":"2018-11-21T04:41:11.017","ParentId":null,"OwnerUserId":"7975433","Title":null,"Body":"<p>Can you first try<\/p>\n\n<blockquote>\n  <p>curl <a href=\"http:\/\/localhost:8501\/v1\/models\/xor\" rel=\"nofollow noreferrer\">http:\/\/localhost:8501\/v1\/models\/xor<\/a><\/p>\n<\/blockquote>\n\n<p>to check if the model is running? This should return the status of your model.<\/p>\n\n<p>From <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/97687024c3b7515d2f2979c35054f44c8f84d146\/tensorflow_serving\/g3doc\/api_rest.md#model-status-api\" rel=\"nofollow noreferrer\">RESTful API doc<\/a>, the format is <code>GET http:\/\/host:port\/v1\/models\/${MODEL_NAME}[\/versions\/${MODEL_VERSION}]<\/code><\/p>\n"},{"AnswerId":"53433207","CreationDate":"2018-11-22T14:34:35.700","ParentId":null,"OwnerUserId":"10675469","Title":null,"Body":"<p>Thanks! you got the point! so I solved.<\/p>\n\n<p>Actually there were 2 errors to the curl command:<\/p>\n\n<ol>\n<li>localhost:8501\/<strong>v1<\/strong>\/models\/xor  I put v2 thinking to use the version #2, but if you put v2 it does not work. It seems v# it is not the versiono of your saved model<\/li>\n<li>Also I need to specify the predict so the exact request is: <strong><a href=\"http:\/\/localhost:8501\/v1\/models\/xor:predict\" rel=\"nofollow noreferrer\">http:\/\/localhost:8501\/v1\/models\/xor:predict<\/a><\/strong><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53380417,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2018-11-19T18:12:06.977","AcceptedAnswerId":null,"OwnerUserId":5220793.0,"Title":"In keras how to get epoch and validation loss from model checkpoint?","Body":"<p>From here (<a href=\"https:\/\/keras.io\/callbacks\/#modelcheckpoint\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#modelcheckpoint<\/a>) you can save the best model according to the validation error by setting save_best_only.<\/p>\n\n<p>I know you can save the corresponding epoch and validation error by writing it to the checkpoint's file name. However, this means a lot of models could be saved and I expect this to result in memory error on my GPU.<\/p>\n\n<p>Is there a way to get the epoch and val loss corresponding to the final best model without having to write it in the filename?   <\/p>\n","answers":[{"AnswerId":"53384385","CreationDate":"2018-11-20T00:04:50.733","ParentId":null,"OwnerUserId":"5536887","Title":null,"Body":"<p>Depends on what you want to do with the epoch and validation error and at what time in the training, but you can implement your own callback functionality quite easily. What you want to get the training metrics is the <code>logs<\/code> object, which is passed to the callback in each of the callback events (see <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L185\" rel=\"nofollow noreferrer\">here<\/a>).<\/p>\n\n<p>If for example you need to call a certain function f with the epoch and validation error at the end of every epoch, you could implement this using the <code>LambdaCallback<\/code>:<\/p>\n\n<pre><code>keras.callbacks.LambdaCallback(on_epoch_end=lambda epoch, logs: f(epoch, logs['val_loss']))\n<\/code><\/pre>\n\n<p>If instead you want to use the ModelCheckpoint callback but don't want it to write to file, you can create a custom CallBack that rewrites the ModelCheck callback and changes the save behavior (code <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L444\" rel=\"nofollow noreferrer\">here<\/a>).<\/p>\n\n<p>(Don't know if that answers your question, not entirely sure what the requirements are)<\/p>\n"}]}
+{"QuestionId":53380560,"AnswerCount":1,"Tags":"<deep-learning><keras>","CreationDate":"2018-11-18T22:56:29.727","AcceptedAnswerId":58827741.0,"OwnerUserId":7334951.0,"Title":"Combining Two CNN's","Body":"<p>I Want to Combine Two CNN Into Just One In Keras, What I Mean Is that I Want The Neural Network  To Take Two Images And Process Each One in Separate CNN, and Then Concatenate Them Together Into The Flattening Layer and Use Fully Connected Layer to Do The Last Work, Here What I Did:<\/p>\n\n<pre><code># Start With First Branch ############################################################\nbranch_one = Sequential()\n\n# Adding The Convolution\nbranch_one.add(Conv2D(32, (3,3),input_shape = (64,64,3) , activation = 'relu'))\nbranch_one.add(Conv2D(32, (3, 3), activation='relu'))\n\n# Doing The Pooling Phase\nbranch_one.add(MaxPooling2D(pool_size=(2, 2)))\nbranch_one.add(Dropout(0.25))\nbranch_one.add(Flatten())\n\n# Start With Second Branch ############################################################\n\nbranch_two = Sequential()\n\n# Adding The Convolution\nbranch_two.add(Conv2D(32, (3,3),input_shape = (64,64,3) , activation = 'relu'))\nbranch_two.add(Conv2D(32, (3, 3), activation='relu'))\n\n# Doing The Pooling Phase\nbranch_two.add(MaxPooling2D(pool_size=(2, 2)))\nbranch_two.add(Dropout(0.25))\nbranch_two.add(Flatten())\n\n# Making The Combinition ##########################################################\nfinal = Sequential()\nfinal.add(Concatenate([branch_one, branch_two]))\nfinal.add(Dense(units = 128, activation = \"relu\"))\nfinal.add(Dense(units = 1, activation = \"sigmoid\"))\n\n# Doing The Compilation\nfinal.compile(loss = 'binary_crossentropy', optimizer = 'adam', metrics = ['accuracy'])\n# Adding and Pushing The Images to CNN\n\n# use ImageDataGenerator to preprocess the data\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\n# augment the data that we have\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   horizontal_flip = True)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\n# prepare training data\nX1 = train_datagen.flow_from_directory('.\/ddsm1000_resized\/images\/train',\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'binary')\n\nX2 = train_datagen.flow_from_directory('.\/ddsm1000_resized_canny\/images\/train',\n\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'binary')\n\n# prepare test data\nY1 = test_datagen.flow_from_directory('.\/ddsm1000_resized\/images\/test',\n                                            target_size = (64, 64),\n                                            batch_size = 32,\n                                            class_mode = 'binary')\nY2 = test_datagen.flow_from_directory('.\/ddsm1000_resized_canny\/images\/test',\n                                            target_size = (64, 64),\n                                            batch_size = 32,\n                                            class_mode = 'binary')\nfinal.fit_generator([X1, X2], steps_per_epoch = (8000 \/ 32), epochs = 1, validation_data = [Y1,Y2], validation_steps = 2000)\n<\/code><\/pre>\n\n<p>Keras Telling Me <\/p>\n\n<blockquote>\n  <p>RuntimeError: You must compile your model before using it.<\/p>\n<\/blockquote>\n\n<p>I Think That is The CNN Does not the shapes of input data, so what Can I Do Here ?? Thanks<\/p>\n","answers":[{"AnswerId":"58827741","CreationDate":"2019-11-12T22:37:16.867","ParentId":null,"OwnerUserId":"5783902","Title":null,"Body":"<p>Make the change as pointed below:<\/p>\n\n<pre><code>from keras.layers import Merge\n...\n...\n\n# Making The Combinition ##########################################################\nfinal = Sequential()\nfinal.add(Merge([branch_one, branch_two], mode = 'concat'))\n\n...\n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53380911,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T18:49:24.393","AcceptedAnswerId":53381012.0,"OwnerUserId":10543278.0,"Title":"Tensorflow - How to read the prediction","Body":"<p>I have a problem.\nFor a school project, I created a Recurrent Neural Network (RNN) where I wanted to predict if the stock price will rise or fall. I also have some data from a CSV file. Training went fine, so I was ready to predict some tests. From the RNN I get some results because it has multiple predictions over a period of a week.\nHere is my code:<\/p>\n\n<pre><code>import io\nimport requests\nimport os\nimport time\nimport random\n\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport tensorflow as tf\n\nfrom sklearn.metrics import mean_absolute_error\nfrom sklearn import preprocessing\nfrom collections import deque\n\n#Constant Variables\nSEQ_LEN = 30\nFUTURE_PERIOD_PREDICT = 3\nRATIO_TO_PREDICT = \"LTC-USD\"\nBATCH_SIZE = 64\nNAME = str(RATIO_TO_PREDICT) + \"-\" + str(SEQ_LEN) + \"-SEQ-\" + str(FUTURE_PERIOD_PREDICT) + \"-PRED-\" + str(int(time.time()))\nACTIONS = [\"Sell\", \"Buy\"]\n\ndef classify(current, future):\n    if float(future) &gt; float(current):\n        return 1\n    else:\n        return 0\n\ndef preprocess_df(df):\n    df = df.drop('future', 1)\n\n    for col in df.columns:\n        if col != \"target\":\n            df[col] = df[col].pct_change()\n            df.dropna(inplace=True)\n            #df[col] = preprocessing.scale(df[col].values)\n\n    df.dropna(inplace=True)\n\n    sequential_data = []\n    prev_days = deque(maxlen=SEQ_LEN)\n\n\n\n    for i in df.values:\n        prev_days.append([n for n in i[:-1]])\n        if len(prev_days) == SEQ_LEN:\n            sequential_data.append([np.array(prev_days), i[-1]])\n\n    buys = []\n    sells = []\n\n    for seq, target in sequential_data:\n        if target == 0:\n            sells.append([seq, target])\n        elif target == 1:\n            buys.append([seq, target])\n\n\n    random.shuffle(buys)\n    random.shuffle(sells)\n\n    lower = min(len(buys), len(sells))\n\n\n    buys = buys[:lower]\n    sells = sells[:lower]\n\n\n    sequential_data = buys+sells\n\n    x = []\n    y = []\n\n    for seq, target in sequential_data:\n        x.append(seq)\n        y.append(target)\n\n    return np.array(x), y\n\n\n\n\nmain_df = pd.DataFrame()\n\nratios = [\"BTC-USD\", \"LTC-USD\", \"ETH-USD\"]\nfor ratio in ratios:\n\n\n    url=\"https:\/\/www.test.nl\/get_csv_data_onscreen.php?method=test&amp;ratio=\" + str(ratio)\n    dataset = requests.get(url, verify=False).content\n    df = pd.read_csv(io.StringIO(dataset.decode('utf-8')), names=[\"time\", \"low\", \"high\", \"open\", \"close\", \"volume\", \"rsi14\", \"ma5\", \"ema5\", \"ema12\", \"ema20\", \"macd\", \"signal\"])\n\n    df.rename(columns={\"close\": str(ratio)+\"_close\", \"volume\": str(ratio) + \"_volume\", \"rsi14\": str(ratio) + \"_rsi14\", \"ma5\": str(ratio) + \"_ma5\", \"ema5\": str(ratio) + \"_ema5\", \"ema12\": str(ratio) + \"_ema12\", \"ema20\": str(ratio) + \"_ema20\", \"macd\": str(ratio) + \"_macd\", \"signal\": str(ratio) + \"_signal\"}, inplace=True)\n\n    df.set_index(\"time\", inplace=True)\n    df = df[[str(ratio) + \"_close\", str(ratio) + \"_volume\", str(ratio) + \"_rsi14\", str(ratio) + \"_ma5\", str(ratio) + \"_ema5\", str(ratio) + \"_ema12\", str(ratio) + \"_ema20\", str(ratio) + \"_macd\", str(ratio) + \"_signal\"]]\n\n    if len(main_df) == 0:\n        main_df = df\n    else:\n        main_df = main_df.join(df)\n\n\nmain_df['future'] = main_df[str(RATIO_TO_PREDICT) + \"_close\"].shift(-FUTURE_PERIOD_PREDICT)\nmain_df['target'] = list(map(classify, main_df[str(RATIO_TO_PREDICT) + \"_close\"], main_df[\"future\"]))\n#print(main_df[[str(RATIO_TO_PREDICT) + \"_close\", \"future\", \"target\"]].head(10))\n\n\ntimes = sorted(main_df.index.values)\nlast_5pct = times[-int(0.05*len(times))]\n\nvalidation_main_df = main_df[(main_df.index &gt;= last_5pct)]\nmain_df = main_df[(main_df.index &lt; last_5pct)]\n\ntest_x, test_y = preprocess_df(main_df)\nvalidation_x, validation_y = preprocess_df(validation_main_df)\n\nmodel = tf.keras.models.load_model(\"models\\Crypto_Model_0.6337.h5\")\n\npredictions = model.predict(test_x)\nprint(predictions)\nprint(ACTIONS[int(prediction[0][0])])\n<\/code><\/pre>\n\n<p>So when I print the predictions I get a list of numbers arround 0 and 1. Here is a short version of the result:<\/p>\n\n<pre><code>[[ 0.61009574]\n [ 0.5243717 ]\n [ 0.56290686]\n [ 0.49165   ]\n [ 0.50527   ]\n [ 0.77428705]\n [ 0.62151164]\n [ 0.55098933]\n [ 0.45642132]\n [ 0.61239064]\n [ 0.69220203]\n [ 0.3707057 ]\n [ 0.5335519 ]\n [ 0.43078205]\n [ 0.57520276]\n [ 0.46626005]\n [ 0.37625414]\n [ 0.56013215]]\n<\/code><\/pre>\n\n<p>But what is the latest datapoint. For example, here is a part of the list I upload:<\/p>\n\n<pre><code>1535782500,63.41,63.63,63.47,63.52,83505,55.104896,63.574000,63.586200,63.61220000,63.454000,0.31080000,0.44500684\n1535783400,63.44,63.74,63.52,63.62,95980,56.921744,63.578000,63.597500,63.61340000,63.469800,0.28840000,0.41370000\n1535784300,63.62,63.86,63.64,63.81,71996,60.216065,63.616000,63.668300,63.64360000,63.502200,0.28270000,0.38750000\n1535785200,63.71,64.00,63.83,63.82,101652,60.387764,63.644000,63.718900,63.67070000,63.532500,0.27580000,0.36520000\n1535786100,63.64,63.87,63.82,63.84,78686,60.752590,63.722000,63.759300,63.69670000,63.561800,0.26880000,0.34590000\n1535787000,63.76,63.88,63.84,63.84,82486,60.752590,63.786000,63.786200,63.71870000,63.588300,0.26030000,0.32880000\n1535787900,63.70,63.89,63.84,63.72,71654,57.093572,63.806000,63.764100,63.71890000,63.600800,0.24110000,0.31130000\n1535788800,63.69,63.87,63.73,63.76,88931,58.001593,63.796000,63.762700,63.72520000,63.616000,0.22650000,0.29430000\n1535789700,63.71,63.86,63.79,63.82,87103,59.389894,63.796000,63.781800,63.73980000,63.635400,0.21730000,0.27890000\n1535790600,63.76,63.97,63.77,63.89,102919,61.009256,63.806000,63.817900,63.76290000,63.659600,0.21320000,0.26580000\n<\/code><\/pre>\n\n<p>I input 1 week of data of 15 minutes, So that is 672 rows. So just to be clear....<\/p>\n\n<p>Is the last value of predictions the prediction of the last row from the csv file?<\/p>\n","answers":[{"AnswerId":"53381012","CreationDate":"2018-11-19T18:59:23.010","ParentId":null,"OwnerUserId":"3761434","Title":null,"Body":"<p>Why are you shuffling the sequential time data? The date\/time index should be in each row and tell you what day it is predicting. Shuffling sequential data isn't heavily recommended for training RNNs or LSTMs. It also appears like you are trying to apply reinforcement learning which I would always recommend avoiding for training, you can get a few lucky actions and the model will just remember the data points, not generalize an algorithm.<\/p>\n"}]}
+{"QuestionId":53381593,"AnswerCount":0,"Tags":"<python><tensorflow><dll><artificial-intelligence><chatbot>","CreationDate":"2018-11-19T19:43:39.630","AcceptedAnswerId":null,"OwnerUserId":9457355.0,"Title":"ImportError: DLL load failed: The specified module could not be found. Tensorflow Chatbot","Body":"<p>I have an error saying <code>ImportError: DLL load failed: The specified module could not be found.<\/code>  I got this error when I was running main.py in this Github repository, <a href=\"https:\/\/github.com\/llSourcell\/chatbot_tutorial\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/chatbot_tutorial<\/a>.  I downloaded all the things needed.  The following picture is the error.<a href=\"https:\/\/i.stack.imgur.com\/rIk2a.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rIk2a.png\" alt=\"The error\"><\/a><\/p>\n\n<p>Can somebody please help?<\/p>\n","answers":[]}
+{"QuestionId":53382268,"AnswerCount":1,"Tags":"<python><tensorflow><png><tfrecord>","CreationDate":"2018-11-19T20:36:45.293","AcceptedAnswerId":null,"OwnerUserId":9403249.0,"Title":"PNG TO TFRecord: error : has type numpy.ndarray, but expected one of: bytes","Body":"<p>I'm doing a .png file to tfrecord. <\/p>\n\n<pre><code>def _bytes_feature(value):\n    \"\"\"Wrapper for inserting bytes features into Example proto.\"\"\"\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n<\/code><\/pre>\n\n<p>convert to example file:<\/p>\n\n<pre><code>def _convert_to_example(filename, image_buffer, label, text, height, width):\nexample = tf.train.Example(features=tf.train.Features(feature={\n    'image\/height': _int64_feature(height),\n    'image\/width': _int64_feature(width),\n    ........\n    'image\/encoded': _bytes_feature(image_buffer)})) #Error\n    return example\n<\/code><\/pre>\n\n<p>decode_png file, not sure this part<\/p>\n\n<pre><code>def decode_png(image_data):\n\n    img_str = image_data.tostring()\n    reconstructed_img_1d = np.fromstring(img_str, dtype=np.uint8)\n    reconstructed_img = reconstructed_img_1d.reshape(image_data.shape)\n\n    return reconstructed_img\n<\/code><\/pre>\n\n<p>process_image file:<\/p>\n\n<pre><code>def _process_image(filename):\n\n    # Read the image file.\n    with open(filename, 'r') as f:\n        image_data = io.imread(f)\n\n    # Decode the RGB PNG.\n    image = decode_png(image_data)\n\n    # Check that image converted to RGB\n    assert len(image.shape) == 3\n    height = image.shape[0]\n    width = image.shape[1]\n    assert image.shape[2] == 3\n\n    return image_data, height, width\n<\/code><\/pre>\n\n<p>Main, process image files batch:<\/p>\n\n<pre><code>for i in files_in_shard:\n        filename = filenames[i]\n        label = labels[i]\n        text = texts[i]\n\n        image_buffer, height, width = _process_image(filename)\n\n        example = _convert_to_example(filename, image_buffer, label,\n                                      text, height, width)\n        writer.write(example.SerializeToString())\n        shard_counter += 1\n        counter += 1\n<\/code><\/pre>\n\n<p>But, there is a error that <code>TypeError: array([[[223, 198, 219],\n        [215, 185, 209],\n        [207, 174, 201],\n        ...,\n        [230 has type numpy.ndarray, but expected one of: bytes<\/code><\/p>\n\n<p>How should I deal with this? Any help would be great.\nThank you<\/p>\n","answers":[{"AnswerId":"53382536","CreationDate":"2018-11-19T20:59:09.753","ParentId":null,"OwnerUserId":"9403249","Title":null,"Body":"<p>Ok I got it... logic pro<\/p>\n\n<pre><code>def decode_png(image_data):\n\nimg_str = image_data.tostring()\nreconstructed_img_1d = np.fromstring(img_str, dtype=np.uint8)\nreconstructed_img = reconstructed_img_1d.reshape(image_data.shape)\n\nreturn img_str\n\n\ndef _process_image(filename):\n    # Read the image file.\n    with open(filename, 'r') as f:\n        image_data = io.imread(f)\n\n    # Decode the RGB PNG.\n    img_str = image_data.tostring()\n\n    height = image_data.shape[0]\n    width = image_data.shape[1]\n\n    return img_str, height, width\n<\/code><\/pre>\n"}]}
+{"QuestionId":53382694,"AnswerCount":0,"Tags":"<python><keras><neural-network><deep-learning>","CreationDate":"2018-11-19T21:11:34.853","AcceptedAnswerId":null,"OwnerUserId":8599554.0,"Title":"How to pass one_hot encoding labels into flow_from_dataframe?","Body":"<p>I am using <code>flow_from_dataframe<\/code> to pass in my images to the generator. However I worked using cifar10 data. In the flow from dataframe method there is a parameter <code>\"y_col\"<\/code> that needs to be set which contains the column for the labels. I am having my file as the following.<\/p>\n\n<pre><code>           Name  1   2   3   4   5   6   7   8   9   10   11   12   13   14 \n00001522_000.png 0   0   0   0   0   0   0   0   0    0    0    0    0    0\n00023313_000.png 1   0   0   0   0   0   0   0   0    0    0    0    0    0\n00023313_001.png 0   0   0   0   0   0   0   0   0    0    0    0    0    0\n00023313_002.png 0   0   1   0   0   0   0   1   0    0    1    0    0    0\n<\/code><\/pre>\n\n<p>If we look at it the last image falls into multiple classes. <\/p>\n\n<p>Now the file with which I learnt the method, the file looks like the following: <\/p>\n\n<pre><code>id,label\n 1,frog\n 2,truck\n 3,truck\n 4,deer\n<\/code><\/pre>\n\n<p>Below is the code I have used <\/p>\n\n<pre><code>import pandas as pd\ndf=pd.read_csv(r\".\\train.csv\")\n\ndatagen=ImageDataGenerator(rescale=1.\/255)\ntrain_generator=datagen.flow_from_dataframe(dataframe=df,directory=\".\\train_imgs\", x_col=\"id\", y_col=\"label\", has_ext=False, class_mode=\"categorical\", target_size=(32,32), batch_size=32)\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3), padding='same',\n             input_shape=(32,32,3)))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10, activation='softmax'))\n\nmodel.compile(optimizers.rmsprop(lr=0.0001,loss=\"categorical_crossentropy\", metrics=[\"accuracy\"])\n\nSTEP_SIZE_TRAIN=train_generator.n\/\/train_generator.batch_size\nSTEP_SIZE_VALID=valid_generator.n\/\/valid_generator.batch_size\nmodel.fit_generator(generator=train_generator,\n                steps_per_epoch=STEP_SIZE_TRAIN,\n                validation_data=valid_generator,\n                validation_steps=STEP_SIZE_VALID,\n                epochs=10)\n<\/code><\/pre>\n\n<p>How can I arrange my label columns so that there can be a single column for the <code>\"y_col\"<\/code> parameter just like the second file. Or is there any other of of without changing the labels I can pass it into the <code>y_col<\/code>? <\/p>\n","answers":[]}
+{"QuestionId":53382769,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-11-19T21:17:32.843","AcceptedAnswerId":53389043.0,"OwnerUserId":1861627.0,"Title":"tf.data.Dataset with constant size batches","Body":"<p>I have a dataset with 19 elements and a batch size of 10. I set my dataset to continuously iterate over the same elements but I noticed that the last batch has only 4 elements instead of 5, and then it starts over with 5, 5, 5, 4, and so on.<\/p>\n\n<p>How is is possible to force the iterator to fill up shorter batches with elements coming from the next iteration so that all the batches have the same size?<\/p>\n\n<p>P.S. just to understand, isn't this the obvious behavior when training a model?<\/p>\n","answers":[{"AnswerId":"53389043","CreationDate":"2018-11-20T08:36:51.067","ParentId":null,"OwnerUserId":"1861627","Title":null,"Body":"<p>To have this behavior, the <code>.repeat()<\/code> method should be invoked before the <code>batch()<\/code> or <code>padded_batch()<\/code> one. So:<\/p>\n\n<pre><code>file_names = [...]\ndef my_map_func(record):\n    ....\ndataset = tf.data.TFRecordDataset(file_names)\\\n    .map(map_func=my_map_func)\\\n    .repeat()\\  # here!\n    .batch(5)\n<\/code><\/pre>\n"},{"AnswerId":"58259926","CreationDate":"2019-10-06T17:59:15.227","ParentId":null,"OwnerUserId":"4792548","Title":null,"Body":"<p>To expand on the use of <code>repeat<\/code> and <code>batch<\/code> I will say this. Actually the place where you put the <code>.repeat()<\/code> does not matter much if you want to have a fixed batch_size. If you set set <code>drop_remainder=True<\/code> in <code>.batch()<\/code> you won't get <code>None<\/code> in the first dimension, you will get a fixed batch size no matter it the position of <code>.repeat()<\/code>. Repeat and batch is very intuitively to use, for example, let's define 4 data-sets varying the position of batch and repeat:<\/p>\n\n<pre><code>import tensorflow as tf\ndataset = tf.data.Dataset.range(3)\n\ndataset1 = dataset.batch(2,drop_remainder=True)\ndataset1 = dataset1.repeat()\n\ndataset2 = dataset.repeat()\ndataset2 = dataset2.batch(2,drop_remainder=True)\n\ndataset3 = dataset.repeat()\ndataset3 = dataset3.batch(2)\n\ndataset4 = dataset.batch(2)\ndataset4 = dataset4.repeat()\n<\/code><\/pre>\n\n<p>you will get this as result:<\/p>\n\n<p><strong>Dataset1<\/strong>: Notice how 2 is never in any batch<\/p>\n\n<ul>\n<li>shape: (2,): Notice that the batch size is 2 and not None<\/li>\n<li>batch 1: [0,1] <\/li>\n<li>batch 1: [0,1]<\/li>\n<\/ul>\n\n<p><strong>Dataset2<\/strong>: This is what you want to do to achieve what you want. Notice how now 2 is in the second batch<\/p>\n\n<ul>\n<li>shape: (2,): Again, the batch size is not None and the position of <code>repeat<\/code> is different<\/li>\n<li>batch 1: [0,1]<\/li>\n<li>batch 1: [2,0]<\/li>\n<\/ul>\n\n<p><strong>Dataset3<\/strong>: <\/p>\n\n<ul>\n<li>shape: (?,): Because you are not using <code>drop_remainder=True<\/code> you get a None, nonetheless you will get batches <strong>always<\/strong> of fix size 2. This will happen no matter how many times you repeat it<\/li>\n<li>batch 1: [0,1]<\/li>\n<li>batch 2: [2,0]<\/li>\n<li>batch 3: [1,2]<\/li>\n<\/ul>\n\n<p><strong>Dataset4<\/strong>:<\/p>\n\n<ul>\n<li>shape(?,): Same a as Dataset3<\/li>\n<li>batch 1: [0,1]<\/li>\n<li>batch 2: [2] you get 'incomplete' batch sizes<\/li>\n<li>batch 3: [0,1]<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53382944,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2018-11-19T21:33:08.070","AcceptedAnswerId":53383755.0,"OwnerUserId":2651073.0,"Title":"Getting a part of a Keras Model","Body":"<p>I have a model for an <code>AutoEncoder<\/code> as follows:<\/p>\n\n<pre><code>height, width = 28, 28\n\ninput_img = Input(shape=(height * width,))\nencoding_dim = height * width\/\/256\n\nx = Dense(height * width, activation='relu')(input_img)\n\nencoded1 = Dense(height * width\/\/2, activation='relu')(x)\nencoded2 = Dense(height * width\/\/8, activation='relu')(encoded1)\n\ny = Dense(encoding_dim, activation='relu')(encoded2)\n\ndecoded2 = Dense(height * width\/\/8, activation='relu')(y)\ndecoded1 = Dense(height * width\/\/2, activation='relu')(decoded2)\n\nz = Dense(height * width, activation='sigmoid')(decoded1)\nautoencoder = Model(input_img, z)\n\n#encoder is the model of the autoencoder slice in the middle \nencoder = Model(input_img, y)\ndecoder = Model(y, z)\nprint(encoder)\nprint(decoder)\n<\/code><\/pre>\n\n<p>The encoder part is retrived using the code above, however I can't get the decoder part using the code I added above:<\/p>\n\n<p>I recieve the following error:<\/p>\n\n<pre><code>ValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_39:0\", shape=(?, 784), dtype=float32) at layer \"input_39\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>Could you please guide me how to get that part?<\/p>\n","answers":[{"AnswerId":"53383755","CreationDate":"2018-11-19T22:47:55.100","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>The <code>decoder<\/code> Model needs to have an input layer. For example, the <code>decoder_input<\/code> here:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>height, width = 28, 28\n\n# Encoder layers\ninput_img = Input(shape=(height * width,))\nencoding_dim = height * width\/\/256\n\nx = Dense(height * width, activation='relu')(input_img)\nencoded1 = Dense(height * width\/\/2, activation='relu')(x)\nencoded2 = Dense(height * width\/\/8, activation='relu')(encoded1)\ny = Dense(encoding_dim, activation='relu')(encoded2)\n\n# Decoder layers\ndecoder_input = Input(shape=(encoding_dim,))\ndecoded2 = Dense(height * width\/\/8, activation='relu')(decoder_input)\ndecoded1 = Dense(height * width\/\/2, activation='relu')(decoded2)\nz = Dense(height * width, activation='sigmoid')(decoded1)\n\n# Build the encoder and decoder models\nencoder = Model(input_img, y)\ndecoder = Model(decoder_input, z)\n\n# Finally, glue encoder and decoder together by feeding the encoder \n# output to the decoder\nautoencoder = Model(input_img, decoder(y))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53383032,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-19T21:40:53.703","AcceptedAnswerId":null,"OwnerUserId":260826.0,"Title":"Option for tf.contrib.lookup.index_table_from_tensor","Body":"<p>I heard that TF 2.0 will remove tf.contrib, what is the best alternative for <code>tf.contrib.lookup.index_table_from_tensor<\/code> function in TF 1.12?<\/p>\n","answers":[{"AnswerId":"55228236","CreationDate":"2019-03-18T18:53:30.690","ParentId":null,"OwnerUserId":"8397234","Title":null,"Body":"<p>Much of the functionality in <code>tf.contrib.lookup<\/code> has been migrated to <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/lookup\" rel=\"nofollow noreferrer\"><code>tf.lookup<\/code><\/a>, though I do not believe there is a 1:1 mapping for <code>tf.contrib.lookup.index_table_from_tensor<\/code>. To date, the available endpoints are:<\/p>\n\n<ul>\n<li><code>KeyValueTensorInitializer<\/code> - table initializers given <code>keys<\/code> and <code>values<\/code> tensors.<\/li>\n<li><code>StaticHashTable<\/code> - a generic hash table implementation.<\/li>\n<li><code>StaticVocabularyTable<\/code> - string to ID table wrapper that assigns out-of-vocabulary keys to buckets.<\/li>\n<li><code>TextFileInitializer<\/code> - table initializers from a text file.<\/li>\n<li><code>experimental.DenseHashTable<\/code> - generic mutable hash table implementation using tensors as the backing store.<\/li>\n<\/ul>\n\n<p>You can find out more about <code>tf.contrib<\/code> deprecation and functionality migration in the <a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\" rel=\"nofollow noreferrer\">Sunsetting <code>tf.contrib<\/code> RFC<\/a>. <\/p>\n"}]}
+{"QuestionId":53383151,"AnswerCount":1,"Tags":"<tensorflow><object-detection><data-augmentation>","CreationDate":"2018-11-19T21:51:24.920","AcceptedAnswerId":53392597.0,"OwnerUserId":3528819.0,"Title":"TensorFlow Object Detection API: specifying multiple data_augmentation_options","Body":"<p>I'm wondering if there's any difference between specifying the data augmentations like this:<\/p>\n\n<pre><code>data_augmentation_options {\n  random_horizontal_flip {\n  }\n}\ndata_augmentation_options {\n  ssd_random_crop {\n  }\n}\n<\/code><\/pre>\n\n<p>Or like this:<\/p>\n\n<pre><code>data_augmentation_options {\n  random_horizontal_flip {\n  }\n  ssd_random_crop {\n  }\n}\n<\/code><\/pre>\n\n<p>In the object detection pipeline file?<\/p>\n\n<p>All the samples in the models repo use the first format, but the second format is accepted as well.<\/p>\n","answers":[{"AnswerId":"53392597","CreationDate":"2018-11-20T12:02:27.533","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>The only correct format is the first one.<\/p>\n\n<p>While the second one will not break the pipeline, it will only take the first specified option.\nYou can verify this yourself by inspecting the created pipeline.config in model_dir.\nThe reason for that is that <code>data_augmentation_options<\/code> is of type <code>PreprocessingStep<\/code> which consists of a <code>oneof preprocessing_step<\/code>. Note the <code>oneof<\/code>.<\/p>\n\n<p>On the other hand, <code>data_augmentation_options<\/code> is <code>repeated<\/code>, thus you can specify <\/p>\n\n<pre><code>data_augmentation_options {\n  augmentation_option_1 {\n  }\n}\ndata_augmentation_options {\n  augmentation_option_2 {\n  }\n}\n...\n<\/code><\/pre>\n\n<p>and so on, as many as you like.<\/p>\n"}]}
+{"QuestionId":53383154,"AnswerCount":1,"Tags":"<python><numpy><matrix><torch>","CreationDate":"2018-11-19T21:51:39.287","AcceptedAnswerId":53383212.0,"OwnerUserId":9312122.0,"Title":"How many rows are equals between two matrices in numpy","Body":"<p>I'm training a neural network for a multi-label classification problem. I have two rows A and B of size BxL (B = batch size, L = number of labels) where A are the actual labels of the minibatch and B are the predictions maded by my model, somethin like that:<\/p>\n\n<pre><code>A = array([[0., 1., 0.],\n          [0., 1., 1.],\n          [0., 1., 0.],\n          [0., 0., 0.]])\n\nB = array([[1., 1., 0.],\n          [0., 1., 1.],\n          [0., 0., 0.],\n          [0., 1., 0.]])\n<\/code><\/pre>\n\n<p>And I want to count how many samples are classified correctly (that means, how many rows are equals in A and B)<\/p>\n\n<p>I would like to know if there is a way to do it with tensor\/numpy parallel functions...<\/p>\n\n<p>something like<\/p>\n\n<pre><code>sum(torch.eq(A,B, axis=0)) # that doesn't exists\n<\/code><\/pre>\n","answers":[{"AnswerId":"53383212","CreationDate":"2018-11-19T21:57:01.023","ParentId":null,"OwnerUserId":"6671176","Title":null,"Body":"<p>You can use <code>numpy<\/code> with <code>all<\/code>:<\/p>\n\n<pre><code>np.sum((A == B).all(1))\n#1\n<\/code><\/pre>\n\n<p>This works by searching whether all values in each row match. <\/p>\n\n<pre><code>&gt;&gt;&gt; A == B\narray([[False,  True,  True],\n       [ True,  True,  True],\n       [ True, False,  True],\n       [ True, False,  True]])\n<\/code><\/pre>\n\n<p>Gives you where the elements match, then <code>all(axis=1)<\/code> returns a boolean of rows where all values are <code>True<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; (A == B).all(1)\narray([False,  True, False, False])\n<\/code><\/pre>\n\n<p>Showing that the row at index <code>1<\/code> is an exact match between the two arrays. <\/p>\n\n<p>Then, summing over the boolean array gives you the count of such rows.<\/p>\n"}]}
+{"QuestionId":53383286,"AnswerCount":0,"Tags":"<python><tensorflow><precision><convergence><epsilon>","CreationDate":"2018-11-19T22:03:41.333","AcceptedAnswerId":null,"OwnerUserId":24183.0,"Title":"Switching entire network from float32 to float64 on condition","Body":"<p>Since lower precision can yield significant computational time savings, I would like to be able to switch (mid run) all variables in my partially trained network from float32 to float64 on an error condition.<\/p>\n\n<p>For example: I initialize all variables as float32, run several hundred thousand batches through the network, and observe the loss reaches a tolerance of order 1e-8. At this point, to continue converging the model, I would like to switch to double precision for all model variables.<\/p>\n\n<p>Is there a simple way to do this in python?<\/p>\n\n<p>EDIT: Also, will switching the dtype of all of the network variables (weights, biases, inputs, etc.) cause issues with the optimizer I previously used? For example if Adam is being used, and computes moment estimates in single precision, will switching to double precision cause an issue?<\/p>\n","answers":[]}
+{"QuestionId":53383559,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T22:27:38.083","AcceptedAnswerId":53383561.0,"OwnerUserId":2628369.0,"Title":"How can I compute the gradient w.r.t. a non-variable in TensorFlow's eager execution mode?","Body":"<p>I am trying to compute the gradient of my model's loss with respect to its input in order to create an adversarial example. Since the model's input is non-trainable, I need to compute the gradient with respect to a tensor, not a variable. However, I found that TensorFlow's <code>GradientTape<\/code> returns <code>None<\/code> gradients if the tensor is not a trainable variable:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ntf.enable_eager_execution()\n\na = tf.convert_to_tensor(np.array([1., 2., 3.]), dtype=tf.float32)\nb = tf.constant([1., 2., 3.])\nc = tf.Variable([1., 2., 3.], trainable=False)\nd = tf.Variable([1., 2., 3.], trainable=True)\n\nwith tf.GradientTape() as tape:\n    result = a + b + c + d\n\ngrads = tape.gradient(result, [a, b, c, d])\n<\/code><\/pre>\n\n<p><code>print(grads)<\/code> prints:<\/p>\n\n<pre><code>[None, None, None, &lt;tf.Tensor: id=26, shape=(3,), dtype=float32, numpy=array([1., 1., 1.], dtype=float32)&gt;]\n<\/code><\/pre>\n\n<p>I went through TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/eager_basics\" rel=\"nofollow noreferrer\">Eager Execution tutorial<\/a> and the <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">Eager Execution guide<\/a>, but couldn't find a solution for calculating the gradient w.r.t. a tensor.<\/p>\n","answers":[{"AnswerId":"53383561","CreationDate":"2018-11-19T22:27:38.083","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GradientTape\" rel=\"nofollow noreferrer\"><code>tf.GradientTape<\/code><\/a> documentation reveals the simple solution:<\/p>\n\n<blockquote>\n  <p>Trainable variables (created by <code>tf.Variable<\/code> or <code>tf.get_variable<\/code>, where <code>trainable=True<\/code> is default in both cases) are automatically watched. Tensors can be manually watched by invoking the <code>watch<\/code> method on this context manager.<\/p>\n<\/blockquote>\n\n<p>In this case,<\/p>\n\n<pre><code>with tf.GradientTape() as tape:\n    tape.watch(a)\n    tape.watch(b)\n    tape.watch(c)\n    result = a + b + c + d\n\ngrads = tape.gradient(result, [a, b, c, d])\n<\/code><\/pre>\n\n<p>will result in <code>print(grads)<\/code>:<\/p>\n\n<pre><code>[&lt;tf.Tensor: id=26, shape=(3,), dtype=float32, numpy=array([1., 1., 1.], dtype=float32)&gt;, \n &lt;tf.Tensor: id=26, shape=(3,), dtype=float32, numpy=array([1., 1., 1.], dtype=float32)&gt;, \n &lt;tf.Tensor: id=26, shape=(3,), dtype=float32, numpy=array([1., 1., 1.], dtype=float32)&gt;, \n &lt;tf.Tensor: id=26, shape=(3,), dtype=float32, numpy=array([1., 1., 1.], dtype=float32)&gt;]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53384110,"AnswerCount":0,"Tags":"<python><keras><image-segmentation><loss><cross-entropy>","CreationDate":"2018-11-19T23:27:01.453","AcceptedAnswerId":null,"OwnerUserId":4724057.0,"Title":"Weighted binary cross entropy dice loss for segmentation problem","Body":"<p>I am using weighted Binary cross entropy Dice loss for a segmentation problem with class imbalance (80 times more black pixels than white pixels) . <\/p>\n\n<pre><code>def weighted_bce_dice_loss(y_true, y_pred):\n    y_true = K.cast(y_true, 'float32')\n    y_pred = K.cast(y_pred, 'float32')\n    averaged_mask = K.pool2d(\n        y_true, pool_size=(50, 50), strides=(1, 1), padding='same', pool_mode='avg')\n    weight = K.ones_like(averaged_mask)\n    w0 = K.sum(weight)\n    weight = 5. * K.exp(-5. * K.abs(averaged_mask - 0.5))\n    w1 = K.sum(weight)\n    weight *= (w0 \/ w1)\n    loss = weighted_bce_loss(y_true, y_pred, weight) + dice_loss(y_true, y_pred)\n    return loss\n<\/code><\/pre>\n\n<p>Dice coeffecient increased and the loss decreased but at every epoch I am getting a black image as output (all the pixels are labelled black)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Iv2xA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Iv2xA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any ideas on why this is happening?<\/p>\n","answers":[]}
+{"QuestionId":53384168,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2018-11-19T23:33:34.693","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Keras \/ Tensorflow: Predict Using tf.data.Dataset API","Body":"<p>I'm using Keras with a Tensorflow backend for building a model for this problem: <a href=\"https:\/\/www.kaggle.com\/cfpb\/us-consumer-finance-complaints\" rel=\"noreferrer\">https:\/\/www.kaggle.com\/cfpb\/us-consumer-finance-complaints<\/a> (just practicing).<\/p>\n\n<p>I train my Keras model using the <code>tf.data.Dataset<\/code> API.  Now, I have a Pandas DataFrame, <code>df_testing<\/code>, whose columns are <code>complaint<\/code> (strings) and <code>label<\/code> (also strings).  I want to predict on these new samples.  I create a <code>tf.data.Dataset<\/code> object, perform preprocessing, make an Iterator, and call predict on my model:<\/p>\n\n<pre><code>data = df_testing[\"complaint\"].values\nlabels = df_testing[\"label\"].values\n\ndataset = tf.data.Dataset.from_tensor_slices((data))\ndataset = dataset.map(lambda x: ({'reviews': x}))\ndataset = dataset.batch(self.batch_size).repeat()\ndataset = dataset.map(lambda x: self.preprocess_text(x, self.data_table))\ndataset = dataset.map(lambda x: x['reviews'])\ndataset = dataset.make_initializable_iterator()\n<\/code><\/pre>\n\n<p>My training used a <code>tf.data.Dataset<\/code> where each element was of the form <code>({'reviews': \"movie was great\"}, \"positive\")<\/code> so I'm mimicking that here for prediction.  Also, my preprocessing just turns my string into a <code>Tensor<\/code> of integers.<\/p>\n\n<p>When I call:<\/p>\n\n<pre><code>preds = model.predict(dataset)\n<\/code><\/pre>\n\n<p>But I'm told my <code>predict<\/code> call fails:<\/p>\n\n<pre><code>ValueError: When using iterators as input to a model, you should specify the `steps` argument.\n<\/code><\/pre>\n\n<p>So I modify this call to be:<\/p>\n\n<pre><code>preds = model.predict(dataset, steps=3)\n<\/code><\/pre>\n\n<p>But now I get back:<\/p>\n\n<pre><code>ValueError: Please provide data as a list or tuple of 2 elements  - input and target pair. Received Tensor(\"IteratorGetNext_2:0\", shape=(?, 100), dtype=int32)\n<\/code><\/pre>\n\n<p>What am I doing incorrectly here?  I shouldn't have to provide a tuple of 2 elements when predicting (I shouldn't need the label).<\/p>\n\n<p>Thanks for any help you can offer!<\/p>\n","answers":[{"AnswerId":"55624442","CreationDate":"2019-04-11T03:54:52.267","ParentId":null,"OwnerUserId":"3197748","Title":null,"Body":"<p>The following code worked for me (tested on tensorflow 1.10.0):<\/p>\n\n<p><strong>[TLDR]<\/strong> Only insert empty dictionary as a dummy input and specify the number of steps:<\/p>\n\n<pre><code>model.predict(x={},steps=4)\n<\/code><\/pre>\n\n<p>Full code:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.data import Dataset\nfrom tensorflow.keras.layers import Dense, Input\nfrom tensorflow.keras.models import Model\n\n\n# dummy data:\nx = np.arange(4).reshape(-1, 1).astype('float32')\ny = np.arange(5, 9).reshape(-1, 1).astype('float32')\n\n# build the Datasets\nds_x = Dataset.from_tensor_slices(x).repeat().batch(4)\nit_x = ds_x.make_one_shot_iterator()\n\nds_y = Dataset.from_tensor_slices(y).repeat().batch(4)\nit_y = ds_y.make_one_shot_iterator()\n\n\n# build compile and train the model\ninput_vals = Input(tensor=it_x.get_next())\noutput = Dense(1, activation='relu')(input_vals)\nmodel = Model(inputs=input_vals, outputs=output)\nmodel.compile('rmsprop', 'mse', target_tensors=[it_y.get_next()])\nmodel.fit(steps_per_epoch=1, epochs=5, verbose=2)\n\n# infer using the dataset\nmodel.predict(x={},steps=4)\n\n<\/code><\/pre>\n"},{"AnswerId":"53384892","CreationDate":"2018-11-20T01:17:28.183","ParentId":null,"OwnerUserId":"5536887","Title":null,"Body":"<p>What version of Keras are you on? I cannot find that specific error message in the code base, but I think I found where it used to be.<\/p>\n\n<p>Here's the error in a version of the code that I think is close to the version you're running: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/025277a1598fa227b53ddc4e316a7a953b2006c8#diff-de9b96ac2d81503324cbbbe21732031fL931\" rel=\"nofollow noreferrer\">commit<\/a> <\/p>\n\n<p>And here's the updated version of that error: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training_eager.py#L464\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training_eager.py#L464<\/a><\/p>\n\n<p>The conditions of the input validation have changed (in the newest version your input would be accepted), but what's relevant is that the error message is much more clear:<\/p>\n\n<pre><code>raise ValueError(\n    'Please provide data as a list or tuple of 1, 2, or 3 elements '\n    ' - `(input)`, or `(input, target)`, or `(input, target,'\n    'sample_weights)`. Received %s. We do not use the `target` or'\n    '`sample_weights` value here.' % inputs.output_shapes)\n<\/code><\/pre>\n\n<p>The target value is never used in the predict function, and so can be anything. Looking at the rest of the function <code>next_element[1]<\/code> is never used.<\/p>\n\n<p>[<strong>TLDR<\/strong>] Using your current version, add a dummy target value to the data, or update your Keras.<\/p>\n"}]}
+{"QuestionId":53384179,"AnswerCount":0,"Tags":"<python><keras><thread-safety><generator>","CreationDate":"2018-11-19T23:34:57.960","AcceptedAnswerId":null,"OwnerUserId":876896.0,"Title":"threadsafe custom generator for keras `fit_generator()`","Body":"<p>I followed the example of a threadsafe generator for Keras <code>fit_generator<\/code> given here: <a href=\"https:\/\/keras.io\/utils\/#sequence\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/utils\/#sequence<\/a>\nIt looks like the index of the batch (<code>idx<\/code>) is locked to each thread. In my case, I want to lock the thread to the example index. Here is my implementation:<\/p>\n\n<pre><code>class CustomGenerator():\n\n    def __init__(self):\n        self.input_ = np.arange(0, 1000)\n        self.labels = np.arange(0, 1000) * 0.1\n        self.batch_sz = 5\n        self.example_index = 0\n\n    def __len__(self):\n        return np.ceil(len(self.input_) \/ float(self.batch_sz))\n\n    def __getitem__(self, batch_idx):\n        batch_x = np.zeros(self.batch_sz)\n        batch_y = np.zeros(self.batch_sz)\n        row = 0\n        while row &lt; self.batch_sz:\n            if self.example_index % 2 == 0:\n                batch_x[row] = self.input_[self.example_index]\n                batch_y[row] = self.labels[self.example_index]\n                row += 1\n            self.example_index += 1\n\n        return batch_x, batch_y\n\ncg = CustomGenerator()\nbatch_idx = 0\n\nwhile True:\n    print(cg.__getitem__(batch_idx))\n    batch_idx += 1\n<\/code><\/pre>\n\n<p>It prints the right output:<\/p>\n\n<pre><code>(array([0., 2., 4., 6., 8.]), array([0. , 0.2, 0.4, 0.6, 0.8]))\n(array([10., 12., 14., 16., 18.]), array([1. , 1.2, 1.4, 1.6, 1.8]))\n(array([20., 22., 24., 26., 28.]), array([2. , 2.2, 2.4, 2.6, 2.8]))\n(array([30., 32., 34., 36., 38.]), array([3. , 3.2, 3.4, 3.6, 3.8]))\n<\/code><\/pre>\n\n<p>How can I make sure that this implementation would work in a threadsafe manner, i.e the different workers are not going to use the same <code>example_index<\/code> when generating the batches.<\/p>\n","answers":[]}
+{"QuestionId":53384184,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-19T23:35:19.990","AcceptedAnswerId":53384220.0,"OwnerUserId":3875270.0,"Title":"Tensor flow throws an uninitialized error even after the variable is initialized","Body":"<p>I am trying to write a program in python using tensorflow to fit data with different patterns using a neural network with a hidden layer. I am facing an error with the code which states that the variable b2 is uninitialized. But I have initialized it already and dont understand what I am missing here.<\/p>\n\n<p>This is part of an <a href=\"http:\/\/ranger.uta.edu\/~kamangar\/CSE-5368-FA18\/Assignment05(Duedate.Nov.18,2018).html\" rel=\"nofollow noreferrer\">assignment<\/a> (the dataset can be understood here) and got stuck here while solving it.<\/p>\n\n<p>The link to the colab notebook is <a href=\"https:\/\/colab.research.google.com\/drive\/1U2CtJ3-_pJUshFbnVZJjml3ooz15aIq6\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The initialization lines look like<\/p>\n\n<pre><code>W1 = tf.Variable(np.random.uniform(low=-0.01, high=0.01, size=(hidden, 2)), name=\"W1\")\nb1 = tf.Variable(np.random.uniform(low=-0.01, high=0.01, size=(hidden, 1)), name=\"b1\")\nW2 = tf.Variable(np.random.uniform(low=-0.01, high=0.01, size=(classes, hidden)), name=\"W2\")\nb2 = tf.Variable(np.random.uniform(low=-0.01, high=0.01, size=(classes, 1)), name=\"b2\")\n<\/code><\/pre>\n\n<p>The code snippet that shows the computation graph is below.<\/p>\n\n<pre><code>operation = \"ReLU\" # \"Sigmoid\"\no = tf.add(tf.matmul(W1, p), b1)\n# ReLU or Sigmoid\nif operation == \"ReLU\":\n  z = tf.zeros((hidden, 1), dtype=tf.float64)\n  output = tf.maximum(o, z)\nelse:\n  output = tf.sigmoid(o)\nfoutput = tf.add(tf.matmul(W2, output), b2)\ncrossentropy = tf.log(tf.exp(foutput) \/ tf.reduce_sum(tf.exp(foutput), 0))\ninit = tf.initialize_all_variables()\nwith tf.Session() as sess:\n  ce = sess.run([crossentropy], feed_dict={p : inputs, t : targets, lr : 0.01})\nprint(ce)\n<\/code><\/pre>\n\n<p>Error message<\/p>\n\n<pre><code>FailedPreconditionError                   Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)\n   1333     try:\n-&gt; 1334       return fn(*args)\n   1335     except errors.OpError as e:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)\n   1318       return self._call_tf_sessionrun(\n-&gt; 1319           options, feed_dict, fetch_list, target_list, run_metadata)\n   1320 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)\n   1406         self._session, options, feed_dict, fetch_list, target_list,\n-&gt; 1407         run_metadata)\n   1408 \n\nFailedPreconditionError: Attempting to use uninitialized value b2\n     [[{{node b2\/read}} = Identity[T=DT_DOUBLE, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](b2)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53384220","CreationDate":"2018-11-19T23:40:34.580","ParentId":null,"OwnerUserId":"5495304","Title":null,"Body":"<p>You also need to run the init function in the graph before computing cross entropy:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  sess.run(init)\n  ce = sess.run([crossentropy], feed_dict={p : inputs, t : targets, lr : 0.01})\n<\/code><\/pre>\n\n<p>Also, <code>tf.initialize_all_variables()<\/code> is deprecated. Use <code>tf.global_variables_initializer()<\/code> instead.<\/p>\n"}]}
+{"QuestionId":53384185,"AnswerCount":1,"Tags":"<java><android><android-studio><tensorflow>","CreationDate":"2018-11-19T23:35:29.257","AcceptedAnswerId":null,"OwnerUserId":10677287.0,"Title":"Tensorflow not working right, not sure what is wrong","Body":"<p>I use Android Studio to code and in the tensorflowObjectDetection example, I commented out <code>@Disable<\/code> and added our Vuforia key, and when I run the program it just crashes.  I've narrowed it down to this code:<\/p>\n\n<pre><code>private void initTfod() {\n    int tfodMonitorViewId = hardwareMap.appContext.getResources().getIdentifier(\n        \"tfodMonitorViewId\", \"id\", hardwareMap.appContext.getPackageName());\n    TFObjectDetector.Parameters tfodParameters = new TFObjectDetector.Parameters(tfodMonitorViewId);\n    tfod = ClassFactory.getInstance().createTFObjectDetector(tfodParameters, vuforia);\n    tfod.loadModelFromAsset(TFOD_MODEL_ASSET, LABEL_GOLD_MINERAL, LABEL_SILVER_MINERAL);\n}\n<\/code><\/pre>\n\n<p>By commenting out the bottom two lines it doesn't crash, but it don't understand why.<\/p>\n","answers":[{"AnswerId":"53402096","CreationDate":"2018-11-20T21:49:08.097","ParentId":null,"OwnerUserId":"10677287","Title":null,"Body":"<p>Figured it out.  The vuforia needed more time to init, so we did this:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>robot.initVuforia(hardwareMap);\r\n\r\nwhile(robot.vuforiaLoaded == false &amp;&amp; !isStopRequested()){\r\n\r\n}\r\n\r\nif (ClassFactory.getInstance().canCreateTFObjectDetector()) {\r\n    robot.initTfod(hardwareMap);\r\n} else {\r\n    telemetry.addData(\"Sorry!\", \"This device is not compatible with TFOD\");\r\n}<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53384231,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning><oversampling>","CreationDate":"2018-11-19T23:41:43.957","AcceptedAnswerId":null,"OwnerUserId":1531248.0,"Title":"How to apply SMOTE technique (oversampling) before word embedding layer","Body":"<p>How to apply SMOTE algorithm before word embedding layer in LSTM.<\/p>\n\n<p>I have a problem of text binary classification (Good(9500) or Bad(500) review with total of 10000 training sample and it's unbalanced training sample), mean while i am using LSTM with pre-trained word-embeddings (100 dimension space for each word) as well, so each training input have an id's (Total of 50 ids with zero padding's as well when the text description is having lesser than 50 words and trimmed to 50 when the description is exceeded 50 characters) of word dictionary.<\/p>\n\n<p>Below is my general flow,<\/p>\n\n<ul>\n<li>Input - 1000(batch) X 50 (sequence length)<\/li>\n<li>Word Embedding - 200(Unique vocabulary word) X 100 (word representation)<\/li>\n<li>After word embedding layer (new input for LSTM) - 1000(batch) X 50(sequence) X 100 (features)<\/li>\n<li>Final State from LSTM 1000 (batch) X 100 (units)<\/li>\n<li>Apply final layer 1000(batch) X 100 X [100(units) X 2 (output class)]<\/li>\n<\/ul>\n\n<p>All i want to generate more data for Bad review with the help of SMOTE<\/p>\n","answers":[]}
+{"QuestionId":53384261,"AnswerCount":0,"Tags":"<tensorflow><loss-function>","CreationDate":"2018-11-19T23:45:47.673","AcceptedAnswerId":null,"OwnerUserId":10675506.0,"Title":"the loss of CNN in tensorflow","Body":"<p>When I train the CNN network, the loss will become very large at the beginning, and then slowly decrease. Why does it suddenly become bigger? Is it a problem with the loss function or activation function, or something else?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/VGNZf.png\" alt=\"enter image description here\"> <\/p>\n","answers":[]}
+{"QuestionId":53384454,"AnswerCount":1,"Tags":"<c++><tensorflow><machine-learning>","CreationDate":"2018-11-20T00:14:55.940","AcceptedAnswerId":53466666.0,"OwnerUserId":486350.0,"Title":"How to use Custom OP to build TensorFlow Graph in C++?","Body":"<p>From TensorFlow documentation, the following can be done to build graph using inherent OP<\/p>\n\n<pre><code>#include \"tensorflow\/cc\/client\/client_session.h\"\n#include \"tensorflow\/cc\/ops\/standard_ops.h\"\n#include \"tensorflow\/core\/framework\/tensor.h\"\n\nint main() {\n  using namespace tensorflow;\n  using namespace tensorflow::ops;\n  Scope root = Scope::NewRootScope();\n  \/\/ Matrix A = [3 2; -1 0]\n  auto A = Const(root, { {3.f, 2.f}, {-1.f, 0.f} });\n  \/\/ Vector b = [3 5]\n  auto b = Const(root, { {3.f, 5.f} });\n  \/\/ v = Ab^T\n  auto v = MatMul(root.WithOpName(\"v\"), A, b, MatMul::TransposeB(true));\n  std::vector&lt;Tensor&gt; outputs;\n  ClientSession session(root);\n  \/\/ Run and fetch v\n  TF_CHECK_OK(session.Run({v}, &amp;outputs));\n  \/\/ Expect outputs[0] == [19; -3]\n  LOG(INFO) &lt;&lt; outputs[0].matrix&lt;float&gt;();\n  return 0;\n}\n<\/code><\/pre>\n\n<p>It seems that <code>MatMul<\/code> class is auto generated as there is no <code>tensorflow\/cc\/ops\/math_ops.h<\/code> in the github source code.\nHow to do the same thing for custom op such as ZeroOut OP from <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#define_the_ops_interface\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n","answers":[{"AnswerId":"53466666","CreationDate":"2018-11-25T10:40:49.363","ParentId":null,"OwnerUserId":"486350","Title":null,"Body":"<p>Take <code>ZeroOut<\/code> from <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">here<\/a> as example, you have to do the following<\/p>\n\n<pre><code>class ZeroOut {\n public:\n  ZeroOut(const ::tensorflow::Scope&amp; scope, ::tensorflow::Input x);\n  operator ::tensorflow::Output() const { return y; }\n  operator ::tensorflow::Input() const { return y; }\n  ::tensorflow::Node* node() const { return y.node(); }\n\n  ::tensorflow::Output y;\n};\n\nZeroOut::ZeroOut(const ::tensorflow::Scope&amp; scope, ::tensorflow::Input x) {\n  if (!scope.ok()) return;\n  auto _x = ::tensorflow::ops::AsNodeOut(scope, x);\n  if (!scope.ok()) return;\n  ::tensorflow::Node* ret;\n  const auto unique_name = scope.GetUniqueNameForOp(\"ZeroOut\");\n  auto builder = ::tensorflow::NodeBuilder(unique_name, \"ZeroOut\")\n                     .Input(_x)\n  ;\n  scope.UpdateBuilder(&amp;builder);\n  scope.UpdateStatus(builder.Finalize(scope.graph(), &amp;ret));\n  if (!scope.ok()) return;\n  scope.UpdateStatus(scope.DoShapeInference(ret));\n  this-&gt;y = Output(ret, 0);\n}\n<\/code><\/pre>\n\n<p>Then you can use it to build graph<\/p>\n\n<pre><code>Scope root = Scope::NewRootScope();\n\/\/ Matrix A = [3 2; -1 0]\nauto A = Const(root, { {3, 2}, {-1, 0} });\nauto v = ZeroOut(root.WithOpName(\"v\"), A);\nstd::vector&lt;Tensor&gt; outputs;\nClientSession session(root);\n\/\/ Run and fetch v\nTF_CHECK_OK(session.Run({v}, &amp;outputs));\nLOG(INFO) &lt;&lt; outputs[0].matrix&lt;int&gt;();\n<\/code><\/pre>\n\n<p>Note: For TensorFlow inherent OP, code like <code>ZeroOut class<\/code> is autogenerated by bazel rule. We can imitate those codes(e.g. <code>tensorflow\/cc\/ops\/math_ops.h<\/code>) to hand write our own classes if we only have a few custom OPs.<\/p>\n"}]}
+{"QuestionId":53384822,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving><tensorflow-hub>","CreationDate":"2018-11-20T01:08:13.273","AcceptedAnswerId":null,"OwnerUserId":6555133.0,"Title":"Error when call prediction with base 64 input","Body":"<p>I am using Tensorflow hub's example to export a saved_model to be serve with Tensorflow serving using Docker. (<a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/hub\/blob\/master\/examples\/image_retraining\/retrain.py<\/a>)<\/p>\n\n<p>I just followed some instruction on the internet and modified the export_model like below <\/p>\n\n<pre><code>def export_model(module_spec, class_count, saved_model_dir):\n  \"\"\"Exports model for serving.\n\n  Args:\n    module_spec: The hub.ModuleSpec for the image module being used.\n    class_count: The number of classes.\n    saved_model_dir: Directory in which to save exported model and variables.\n  \"\"\"\n  # The SavedModel should hold the eval graph.\n  sess, in_image, _, _, _, _ = build_eval_session(module_spec, class_count)\n\n  # Shape of [None] means we can have a batch of images.\n  image = tf.placeholder(shape=[None], dtype=tf.string)\n\n  with sess.graph.as_default() as graph:\n    tf.saved_model.simple_save(\n        sess,\n        saved_model_dir,\n        #inputs={'image': in_image},\n        inputs = {'image_bytes': image},\n        outputs={'prediction': graph.get_tensor_by_name('final_result:0')},\n        legacy_init_op=tf.group(tf.tables_initializer(), name='legacy_init_op')\n    )\n<\/code><\/pre>\n\n<p>The problem is when i try to call the api using postman it came with this error<\/p>\n\n<pre><code>{\n    \"error\": \"Tensor Placeholder_1:0, specified in either feed_devices or fetch_devices was not found in the Graph\"\n}\n<\/code><\/pre>\n\n<p>Do I need to modify the retraining process so it can accept base64 input?<\/p>\n","answers":[]}
+{"QuestionId":53384906,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-20T01:19:52.317","AcceptedAnswerId":null,"OwnerUserId":9005671.0,"Title":"How to compute the complexity of machine learning models","Body":"<p>I am working on comparison of deep learning models with application in Vehicular network communication security. I want to know how I can compute the complexity of these models to know the performance of my proposed ones. I am making use of tensorflow <\/p>\n","answers":[{"AnswerId":"53385033","CreationDate":"2018-11-20T01:38:01.723","ParentId":null,"OwnerUserId":"1586200","Title":null,"Body":"<p>You can compare the complexity of two deep networks with respect to space and time.<\/p>\n\n<p>Regarding space complexity:<\/p>\n\n<p>Number of parameters in your model -> this is directly proportional to the amount of memory consumed by your model.<\/p>\n\n<p>Regarding time complexity:<\/p>\n\n<ol>\n<li>Amount of time it takes to train a single batch for a given batch size.<\/li>\n<li>Amount of time it takes for training to converge<\/li>\n<li>Amount of time it takes to perform inference on a single sample<\/li>\n<\/ol>\n\n<p>Some papers also discuss the architecture complexity. For example, if GoogLeNet accuracy is only marginally higher than VGG-net, some people might prefer VGG-net as it is a lot easier to implement.<\/p>\n\n<p>You can also discuss some analysis on tolerance of your network to hyperparameter tuning i.e. how your performance varies when you change the hyperparameters.<\/p>\n\n<p>If your model is in a distributed setting, there are other things to mention such as the communication interval as it is the bottleneck sometimes.<\/p>\n\n<p>In summary, you can discuss pretty much anything you feel that is implemented differently in another network and that is contributing additional complexity without much improvement in accuracy with respect to your network.<\/p>\n\n<p>I don't think you would want it but there is also an open source project called <a href=\"https:\/\/github.com\/baidu-research\/DeepBench\" rel=\"nofollow noreferrer\">deepBench<\/a> to benchmark different deep network models.<\/p>\n"}]}
+{"QuestionId":53384948,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-20T01:26:27.473","AcceptedAnswerId":53390725.0,"OwnerUserId":9491512.0,"Title":"how to calculate the euclidean distance between the vectors in one matrix\uff1f","Body":"<p>I want to calculate the Euclidean distance of the vectors in the features, which is a tf.Tensor got from the network. <\/p>\n\n<p>I tried it in the following way, but failed with error: <\/p>\n\n<pre><code>'Tensor' object is not iterable\n<\/code><\/pre>\n\n<p>So I want to calculate the distance between the rows in one matrix just through matrix\uff0cwithout iteration of every rows.<\/p>\n\n<pre><code>features, _ = mnist_net(images)\nfeature_matrix = np.zeros(shape=(FLAGS.batch_size,FLAGS.batch_size))\nfor i in range (FLAGS.batch_size):\n   for j in range (FLAGS.batch_size):\n      aa = tf.slice(features,[i,0],[1,50])\n      bb = tf.slice(features,[j,0],[1,50])\n      feature_matrix[i,j] = tf.sqrt(sum((aa-bb)**2))  \n<\/code><\/pre>\n","answers":[{"AnswerId":"53390725","CreationDate":"2018-11-20T10:14:37.630","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can achieve that simply with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/norm\" rel=\"nofollow noreferrer\"><code>tf.norm<\/code><\/a>\/<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/linalg\/norm\" rel=\"nofollow noreferrer\"><code>tf.linalg.norm<\/code><\/a>:<\/p>\n\n<pre><code>feature_matrix = tf.linalg.norm(features[:, tf.newaxis] - features, axis=-1)\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    features = tf.placeholder(tf.float32, [None, None])\n    feature_matrix = tf.linalg.norm(features[:, tf.newaxis] - features, axis=-1)\n    print(sess.run(feature_matrix, feed_dict={features: [[ 1,  2,  3],\n                                                         [ 4,  5,  6],\n                                                         [ 7,  8,  9],\n                                                         [10, 11, 12]]}))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[ 0.        5.196152 10.392304 15.588457]\n [ 5.196152  0.        5.196152 10.392304]\n [10.392304  5.196152  0.        5.196152]\n [15.588457 10.392304  5.196152  0.      ]]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT:<\/p>\n\n<p>If you cannot use <code>tf.norm<\/code>, the following is an equivalent implementation:<\/p>\n\n<pre><code>sqdiff = tf.squared_difference(features[:, tf.newaxis], features)\nfeature_matrix = tf.sqrt(tf.reduce_sum(sqdiff, axis=-1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53385937,"AnswerCount":0,"Tags":"<tensorflow><yolo>","CreationDate":"2018-11-20T03:47:58.530","AcceptedAnswerId":null,"OwnerUserId":10677945.0,"Title":"The test loss changes strangely with learning rate 0?","Body":"<p>I'm trying to study the performace of YOLOv2. I have implemented the YOLOv2 in tensorflow with the backbone of mobilenet.<\/p>\n\n<p>After setting the learning rate=0, weight decay=0, BN's is_training=False, the loss of test dataset is still changing unexpectedly. <\/p>\n\n<p>I just want to know if there is any other possible parameter may influence the test loss. <\/p>\n","answers":[]}
+{"QuestionId":53386111,"AnswerCount":0,"Tags":"<python><tensorflow><eager>","CreationDate":"2018-11-20T04:12:40.023","AcceptedAnswerId":null,"OwnerUserId":9847201.0,"Title":"How to get intermediate output for eager mode","Body":"<p>I established a transfer learning model and tried to get the intermediate output. I followed one of the suggestion <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">here<\/a> in order to get the intermediate output of a model, as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras import backend as K\n\ntf.enable_eager_execution()\n\nbase_conv = tf.keras.applications.MobileNetV2(input_shape= None, \n            alpha=1.0, depth_multiplier=1, \\\n            include_top=True, weights='imagenet', input_tensor=None, \\\n            pooling=None, classes=1000)\n\nmodel = tf.keras.models.Sequential()\n\nmodel.add(base_conv)\n\nmodel.add(tf.keras.layers.Dense(512, activation='relu'))\n\nmodel.add(tf.keras.layers.Dense(5, activation='softmax'))\n\n# with a Sequential model\nget_layer_output = K.function([model.layers[0].input],                                 \n                          [model.layers[0].get_layer('Conv_1').output])\n\n#x is the input image\nlayer_output = get_layer_output([x])\n<\/code><\/pre>\n\n<p>But then I got an error:<\/p>\n\n<pre><code>AttributeError: Tensor.name is meaningless when eager execution is enabled.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53386245,"AnswerCount":1,"Tags":"<python><scikit-learn><pytorch><skorch>","CreationDate":"2018-11-20T04:29:01.153","AcceptedAnswerId":53444695.0,"OwnerUserId":10677991.0,"Title":"Why Skorch show NAN in the every epoch?","Body":"<p>I want to create my own dataset class based on Dataset class of Skorch because I want to differentiate categorical columns and continuous columns. These categorical columns will be passed through the embedding layers in the model. The result is weird because it show NAN\nlike this:<\/p>\n\n<pre><code>  epoch    train_loss    valid_loss     dur\n-------  ------------  ------------  ------\n      1           nan           nan  0.2187\n      2           nan           nan  0.1719\n      3           nan           nan  0.1719\n      4           nan           nan  0.1562\n      5           nan           nan  0.1406\n<\/code><\/pre>\n\n<p>Can you help me fix it ? I am using data from this kaggle: \n<a href=\"https:\/\/www.kaggle.com\/c\/house-prices-advanced-regression-techniques\/data\" rel=\"nofollow noreferrer\">Here<\/a><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from skorch import NeuralNetRegressor\nfrom skorch.dataset import Dataset\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport pandas as pd\nimport numpy as np\nfrom sklearn.preprocessing import LabelEncoder\n\n\nclass TabularDataset(Dataset):\n    def __init__(self, data, cat_cols=None, output_col=None):\n        self.n = data.shape[0]\n\n        if output_col:\n            self.y = data[output_col].astype(np.float32).values.reshape(-1, 1)\n        else:\n            self.y = np.zeros((self.n, 1))\n\n        self.cat_cols = cat_cols if cat_cols else []\n        self.cont_cols = [col for col in data.columns\n                          if col not in self.cat_cols + [output_col]]\n\n        if self.cont_cols:\n            self.cont_X = data[self.cont_cols].astype(np.float32).values\n        else:\n            self.cont_X = np.zeros((self.n, 1))\n\n        if self.cat_cols:\n            self.cat_X = data[self.cat_cols].astype(np.int64).values\n        else:\n            self.cat_X = np.zeros((self.n, 1))\n\n    def __len__(self):\n        # Denotes the total number of sampoes\n        return self.n\n\n    def __getitem__(self, idx):\n        # generates one sample of data\n        return [self.cont_X[idx], self.cat_X[idx]], self.y[idx]\n\n\nclass FeedForwardNN(nn.Module):\n\n    def __init__(self, emb_dims, no_of_cont, lin_layer_sizes,\n                 output_size, emb_dropout, lin_layer_dropouts):\n\n        \"\"\"\n        Parameters\n        ----------\n        emb_dims: List of two element tuples\n          This list will contain a two element tuple for each\n          categorical feature. The first element of a tuple will\n          denote the number of unique values of the categorical\n          feature. The second element will denote the embedding\n          dimension to be used for that feature.\n        no_of_cont: Integer\n          The number of continuous features in the data.\n        lin_layer_sizes: List of integers.\n          The size of each linear layer. The length will be equal\n          to the total number\n          of linear layers in the network.\n        output_size: Integer\n          The size of the final output.\n        emb_dropout: Float\n          The dropout to be used after the embedding layers.\n        lin_layer_dropouts: List of floats\n          The dropouts to be used after each linear layer.\n        \"\"\"\n\n        super().__init__()\n\n        # Embedding layers\n        self.emb_layers = nn.ModuleList([nn.Embedding(x, y)\n                                         for x, y in emb_dims])\n\n        no_of_embs = sum([y for x, y in emb_dims])\n        self.no_of_embs = no_of_embs\n        self.no_of_cont = no_of_cont\n\n        # Linear Layers\n        first_lin_layer = nn.Linear(self.no_of_embs + self.no_of_cont,\n                                    lin_layer_sizes[0])\n\n        self.lin_layers = \\\n            nn.ModuleList([first_lin_layer] + \\\n                          [nn.Linear(lin_layer_sizes[i], lin_layer_sizes[i + 1])\n                           for i in range(len(lin_layer_sizes) - 1)])\n\n        for lin_layer in self.lin_layers:\n            nn.init.kaiming_normal_(lin_layer.weight.data)\n\n        # Output Layer\n        self.output_layer = nn.Linear(lin_layer_sizes[-1],\n                                      output_size)\n        nn.init.kaiming_normal_(self.output_layer.weight.data)\n\n        # Batch Norm Layers\n        self.first_bn_layer = nn.BatchNorm1d(self.no_of_cont)\n        self.bn_layers = nn.ModuleList([nn.BatchNorm1d(size)\n                                        for size in lin_layer_sizes])\n\n        # Dropout Layers\n        self.emb_dropout_layer = nn.Dropout(emb_dropout)\n        self.droput_layers = nn.ModuleList([nn.Dropout(size)\n                                            for size in lin_layer_dropouts])\n\n    def forward(self, X):\n        cont_data = X[0]\n        cat_data = X[1]\n        if self.no_of_embs != 0:\n            x = [emb_layer(cat_data[:, i])\n                 for i, emb_layer in enumerate(self.emb_layers)]\n            x = torch.cat(x, 1)\n            x = self.emb_dropout_layer(x)\n\n        if self.no_of_cont != 0:\n            normalized_cont_data = self.first_bn_layer(cont_data)\n\n            if self.no_of_embs != 0:\n                x = torch.cat([x, normalized_cont_data], 1)\n            else:\n                x = normalized_cont_data\n\n        for lin_layer, dropout_layer, bn_layer in \\\n                zip(self.lin_layers, self.droput_layers, self.bn_layers):\n            x = F.relu(lin_layer(x))\n            x = bn_layer(x)\n            x = dropout_layer(x)\n\n        x = self.output_layer(x)\n\n        return x\n\n\n# Read data\ndata = pd.read_csv(\"data\/train.csv\", usecols=[\"SalePrice\", \"MSSubClass\", \"MSZoning\", \"LotFrontage\", \"LotArea\",\n                                              \"Street\", \"YearBuilt\", \"LotShape\", \"1stFlrSF\", \"2ndFlrSF\"]).dropna()\n\ncategorical_features = [\"MSSubClass\", \"MSZoning\", \"Street\", \"LotShape\", \"YearBuilt\"]\noutput_feature = \"SalePrice\"\n\n# Label Encode Categorial Features\nlabel_encoders = {}\nfor cat_col in categorical_features:\n    label_encoders[cat_col] = LabelEncoder()\n    data[cat_col] = label_encoders[cat_col].fit_transform(data[cat_col])\n\n# feed Forward NN\ncat_dims = [int(data[col].nunique()) for col in categorical_features]\n\nemb_dims = [(x, min(50, (x + 1) \/\/ 2)) for x in cat_dims]\n\n\nnet = FeedForwardNN(emb_dims, no_of_cont=4, lin_layer_sizes=[50, 100],\n                    output_size=1, emb_dropout=0.04,\n                    lin_layer_dropouts=[0.001, 0.01])\n\n# Fit\nds = TabularDataset(data=data, cat_cols=categorical_features,\n                    output_col=output_feature)\nX = data.drop(['SalePrice'], axis=1)\ny = data['SalePrice'].values.reshape(-1, 1)\nnet = NeuralNetRegressor(\n    net,\n    max_epochs=5,\n    lr=0.1,\n    dataset=ds\n)\nnet.fit(X, y)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53444695","CreationDate":"2018-11-23T10:13:56.730","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>The problem is not with skorch but your data. You have to scale your inputs and, in this case, especially the targets to avoid huge losses and exploding gradients. As a start I suggest using, for example, <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html\" rel=\"nofollow noreferrer\"><code>sklearn.preprocessing.StandardScaler<\/code><\/a>:<\/p>\n\n<pre><code>from sklearn.preprocessing import StandardScaler\n\nclass TabularDataset(Dataset):\n    def __init__(self, data, cat_cols=None, output_col=None):\n        self.n = data.shape[0]\n        # [...]\n        if output_col:\n            scaler_y = StandardScaler()\n            self.y = data[output_col].astype(np.float32).values.reshape(-1, 1)\n\n            scaler_y.fit(self.y)\n            self.y = scaler_y.transform(self.y)\n        # [...]\n        if self.cont_cols:\n            scaler_X_cont = StandardScaler()\n            self.cont_X = data[self.cont_cols].astype(np.float32).values\n            scaler_X_cont.fit(self.cont_X)\n            self.cont_X = scaler_X_cont.transform(self.cont_X)\n        # [...]\n<\/code><\/pre>\n\n<p>As a side note, you don't need <code>X<\/code> and <code>y<\/code> when you have a dataset that provides the actual data, you can simply pass it to <code>net.fit<\/code> (with the exception of using a stratified CV split):<\/p>\n\n<pre><code>net = NeuralNetRegressor(\n    net,\n    max_epochs=5,\n    lr=0.00001,\n)\nnet.fit(ds, y=None)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53386385,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning><conv-neural-network>","CreationDate":"2018-11-20T04:45:41.460","AcceptedAnswerId":null,"OwnerUserId":10671973.0,"Title":"Tensorflow: Scale weights to fit new model","Body":"<p>I am trying to utilize weights from a checkpoint that have a shape of (784, 62) with a new model who's weights are (4096, 62).  I first began by trying to resize the old weights as one would with a picture, but that did not work.  Is there any way to somehow 'scale' the old weights to fit the model of the new and still maintain their accuracy?  Besides the increased shape, both models are exactly the same.<\/p>\n\n<p>I am happy to provide any code of my program upon request.<\/p>\n","answers":[]}
+{"QuestionId":53387059,"AnswerCount":1,"Tags":"<tensorflow><image-processing><object-detection><tensorflow-datasets>","CreationDate":"2018-11-20T05:57:38.813","AcceptedAnswerId":53392772.0,"OwnerUserId":3775518.0,"Title":"Faster RCNN + inception v2 input size","Body":"<p>What is the input size of faster RCNN RPN?\nI'm using an object detection API of Tensorflow which is using faster RCNN as region proposal network ( RPN ) and Inception as feature extractor ( according to the config file ). The API is using the online approach in prediction phase and detects every input image singly. however, I'm now trying to feed images to the network in the batch manner by use of Tensorflow dataset API.\nas you know for make batch out of the data, firstly we need to resize all of the images to a same size. I think the best way of resizing the images is to resize them exactly to the input size of faster RCNN to avoid duplicate resizing. Now my question is what is the input size of the faster RCNN RPN?\nthanks in advance<\/p>\n","answers":[{"AnswerId":"53392772","CreationDate":"2018-11-20T12:13:26.080","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>It depends on the input resolution which was specified in the pipeline config file, in <code>image_resizer<\/code>.\nFor example, for Faster R-CNN over InceptionV2 trained on COCO dataset, see <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/samples\/configs\/faster_rcnn_inception_v2_coco.config\" rel=\"nofollow noreferrer\">this config file<\/a>.\nThe specified resolution is 600x1024.<\/p>\n\n<p>On a side note, fully convolutional architectures (such as RFCN, SSD, YOLO) don't restrict to a single resolution, i.e. you can apply them on different input resolution without modifying the architecture.\nBut this doesn't mean that the model will be robust to it if you're training on a single resolution.<\/p>\n"}]}
+{"QuestionId":53388453,"AnswerCount":1,"Tags":"<python><tensorflow><tensorlayer>","CreationDate":"2018-11-20T07:52:37.343","AcceptedAnswerId":null,"OwnerUserId":9725577.0,"Title":"Tensorflow ValueError: No variables to optimize","Body":"<p>I am trying to use 3D unet on meical image. The interface I am using is <a href=\"https:\/\/tensorlayer.readthedocs.io\/en\/stable\/\" rel=\"nofollow noreferrer\">tensorlayer<\/a>. While I am trying to calculate the mean_squared_error of two tensors:<\/p>\n\n<pre><code>mse_loss = tl.cost.mean_squared_error(dense_unet3D.outputs, t_target_image, is_mean=True)\n<\/code><\/pre>\n\n<p>Here the tl stands for tensorlayer and I am using 3D dense_unet. The size of dense_unet3D.outputs is: (4, 128, 128, 128, 1) meaning (batch, depth, hight, width, channel), same as t_target_image. \nI know that in tensorlayer, the mean_squared_error could not calculate 5D tensor, so I added these code in the function <strong><a href=\"https:\/\/tensorlayer.readthedocs.io\/en\/stable\/modules\/cost.html\" rel=\"nofollow noreferrer\">cost<\/a><\/strong> in tensorlayer:<\/p>\n\n<pre><code>mse = tf.reduce_mean(tf.reduce_mean(tf.squared_difference(output, target), [1, 2, 3, 4]), name=name)\n<\/code><\/pre>\n\n<p>And I defined the optimizer as this:<\/p>\n\n<pre><code>g_optim_init = tf.train.RMSPropOptimizer(lr_v).minimize(mse_loss, var_list=g_vars)\n<\/code><\/pre>\n\n<p>However, when the code above is executed, the error:<\/p>\n\n<p><strong>ValueError: No variables to optimize.<\/strong><\/p>\n\n<p>came out. I don't know why. Could anyone uses tensorlayer tell me how to solve this problem? <\/p>\n","answers":[{"AnswerId":"53391184","CreationDate":"2018-11-20T10:39:52.773","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>How is your variable list <code>g_vars<\/code> looks like? That could be empty which left the optimizer with no variables to optimize. <\/p>\n"}]}
+{"QuestionId":53389331,"AnswerCount":1,"Tags":"<python><tensorflow><keras><layer><prediction>","CreationDate":"2018-11-20T08:54:10.777","AcceptedAnswerId":53390479.0,"OwnerUserId":4782060.0,"Title":"How can I get multiple outputs in an LSTM network in Python with Keras and Tensorflow?","Body":"<p>I am working by first time with LSTMs in Keras and Tensorflow in Python, and I want to create a neural network with some layers and which gives 10 output values. I generated multiple layers in a neural network, and I created an output DenseLayer of 10 elements. I have the next code:<\/p>\n\n<pre><code>from pandas import DataFrame\nfrom pandas import Series\nfrom pandas import concat\nfrom pandas import read_csv\nfrom pandas import datetime\nfrom sklearn.metrics import mean_squared_error\nfrom sklearn.preprocessing import MinMaxScaler\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom math import sqrt\nfrom matplotlib import pyplot\nimport numpy\nfrom numpy import array\nimport math\n\n\n# convert an array of values into a dataset matrix\ndef create_dataset(dataset, look_back):\n    dataX, dataY = [], []\n    for i in range(len(dataset)-look_back-1):\n        a = dataset[i:(i+look_back), 0]\n        dataX.append(a)\n        dataY.append(dataset[i + look_back, 0])\n    return numpy.array(dataX), numpy.array(dataY)\n\nlook_back = 10\nepochs = 1000\nbatch_size = 50\n\ndata = data.astype('float32')\n\nscaler = MinMaxScaler(feature_range=(0, 1))\ndataset = scaler.fit_transform(data)\n\n\n# split into train and test sets\ntrain_size = int(len(dataset) * 0.67)\ntest_size = len(dataset) - train_size\ntrain, test = dataset[0:train_size,:], dataset[train_size:len(dataset),:]\n\ntrainX, trainY = create_dataset(train, look_back)\ntestX, testY = create_dataset(test, look_back)\n\n\n# reshape input to be [samples, time steps, features]\ntrainX = numpy.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))\ntestX = numpy.reshape(testX, (testX.shape[0], 1, testX.shape[1]))\n\n\n# create and fit the LSTM network\nmodel = Sequential()\nmodel.add(LSTM(100, activation = 'tanh', inner_activation = 'hard_sigmoid', return_sequences=True))#, input_shape=(1, look_back)))\nmodel.add(LSTM(50, activation = 'tanh', inner_activation = 'hard_sigmoid', return_sequences=True))\nmodel.add(LSTM(25, activation = 'tanh', inner_activation = 'hard_sigmoid'))\n# I want 10 outputs    \nmodel.add(Dense(10))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.fit(trainX, trainY, epochs=epochs, batch_size=batch_size, verbose=2)\n<\/code><\/pre>\n\n<p>But when I execute the code I get the next error message:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_1 to have shape (10,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>What can I do to solve the problem? I want to give me predictions for the next 10 elements, that is the reason why I put a final layer of 10 elements.<\/p>\n","answers":[{"AnswerId":"53390479","CreationDate":"2018-11-20T10:01:25.823","ParentId":null,"OwnerUserId":"10238218","Title":null,"Body":"<p>From what you have said above the error <code>ValueError: Error when checking target: expected dense_1 to have shape (10,) but got array with shape (1,)<\/code> is due to a problem in your target:<\/p>\n\n<ul>\n<li>you have a list of values that are the targets.<\/li>\n<li>you try to predict ten values while having only one to compare to.<\/li>\n<\/ul>\n\n<p>you need to rework the trainY matrx to include every value you wish to predict.\nfor example if you wish to predict the 5 values in the closest futur, you'll need a target line (ie each element) of size 5 including all values.<\/p>\n\n<p>as such you'll train the network to predict the 5 futur values.\ni'll try to get you the code howerver it's just a reshaping with a roll to get futur values.<\/p>\n\n<p>to be more precise, for 1 X (one input) you'll need a <code>y=[v1,v2,v3,v4,v5]<\/code>\nso if you have <code>train = [X1,X2,..]<\/code> then <code>Y = [[v1,v2,v3,v4,v5],[v2,v3,v4,v5,v6]<\/code><\/p>\n"}]}
+{"QuestionId":53389404,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2018-11-20T08:58:24.333","AcceptedAnswerId":null,"OwnerUserId":7904697.0,"Title":"Python 3.5 - ModuleNotFoundError: No module named 'tensorflow'","Body":"<p>I have been trying to use <code>tensorflow<\/code> to build a simple neural network using PyCharm. I have managed to install the <code>tensorflow<\/code> package(see image below) but the following error appears when I try to run the application.<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"C:\/Users\/abc\/Desktop\/new-app\/neuraln.py\", line 1, in\n  \n      import tensorflow as tf ModuleNotFoundError: No module named 'tensorflow'<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vAXz3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vAXz3.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I would really appreciate some help. Thanks!<\/p>\n\n<p><em>Note - I'm not using <code>conda<\/code> env<\/em><\/p>\n","answers":[]}
+{"QuestionId":53389416,"AnswerCount":1,"Tags":"<python><python-3.x><keras><dataset>","CreationDate":"2018-11-20T08:59:11.127","AcceptedAnswerId":53425808.0,"OwnerUserId":7213391.0,"Title":"Keras ImageDataGenerator.flow_from_dataframe keeps on loading","Body":"<p>I'm working on the <a href=\"https:\/\/stanfordmlgroup.github.io\/competitions\/mura\/\" rel=\"nofollow noreferrer\">MURA dataset<\/a> by Stanford. I'm trying to load the dataset using Keras's ImageDataGenerator. The data is in the following hierarchy:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZZVs6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZZVs6.png\" alt=\"The directory hierarchy\"><\/a><\/p>\n\n<p>The <code>study1_positive<\/code> folder contains the images.<\/p>\n\n<p><code>ImageDataGenerator.flow_from_directory<\/code> cannot be used with this folder structure, therefore I tried using the <code>flow_from_dataframe<\/code> method. <\/p>\n\n<p>However, when run, the code keeps on executing and doesn't stop.<\/p>\n\n<p>Following is the format of the Pandas DataFrame that I'm passing to the <code>flow_from_directory<\/code> method:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QNeie.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QNeie.png\" alt=\"The DataFrame passed to flow_from_dataframe\"><\/a><\/p>\n\n<p>I've also tried changing the labels to 'abnormal' and 'normal' in place of 1 and 0, respectively.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I've pasted the code below:<\/p>\n\n<pre><code>train_imggen = ImageDataGenerator(rescale=1.\/255, rotation_range=30,\n                              horizontal_flip=True)\n\ntrain_loader = train_imggen.flow_from_dataframe(traindf, '.', shuffle=True,\n                                            x_col='path', y_col='label',\n                                            color_mode='grayscale',\n                                            target_size=(320,320), \n                                            class_mode='binary', \n                                            batch_size=8)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53425808","CreationDate":"2018-11-22T07:26:11.767","ParentId":null,"OwnerUserId":"7213391","Title":null,"Body":"<p>Found the solution <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/issues\/96\" rel=\"nofollow noreferrer\">here<\/a>. Apparently, the current <code>flow_from_dataframe<\/code> doesn't handle relative paths.<\/p>\n\n<blockquote>\n  <ol>\n  <li>Clone my \"fix_found_0_images\" branch.   <code>git clone -b fix_found_0_images_bug\n  https:\/\/github.com\/smurak\/keras-preprocessing.git<\/code><\/li>\n  <li>Move the \"keras_preprocessing\" subdirectory to your working directory.<\/li>\n  <li>Import it.<\/li>\n  <\/ol>\n<\/blockquote>\n\n<pre><code>import keras\nfrom keras_preprocessing import image\n...\ntrain_imggen = image.ImageDataGenerator(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53389417,"AnswerCount":1,"Tags":"<python><pytorch><torch>","CreationDate":"2018-11-20T08:59:12.043","AcceptedAnswerId":53389447.0,"OwnerUserId":497600.0,"Title":"Is it possible to split the training DataLoader (and dataset) into training and validation datasets?","Body":"<p>The torchvision package provides easy access to commonly used datasets. <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py\" rel=\"nofollow noreferrer\">You would use them like this<\/a>:<\/p>\n\n<pre><code>trainset = torchvision.datasets.CIFAR10(root='.\/data', train=True,\n                                        download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=4,\n                                          shuffle=True, num_workers=2)\n\ntestset = torchvision.datasets.CIFAR10(root='.\/data', train=False,\n                                       download=True, transform=transform)\ntestloader = torch.utils.data.DataLoader(testset, batch_size=4,\n                                         shuffle=False, num_workers=2)\n<\/code><\/pre>\n\n<p>Apparently, you can only switch between <code>train=True<\/code> and <code>train=False<\/code>. The <a href=\"https:\/\/pytorch.org\/docs\/master\/torchvision\/datasets.html#torchvision.datasets.MNIST\" rel=\"nofollow noreferrer\">docs<\/a> explain:<\/p>\n\n<blockquote>\n  <p>train (bool, optional) \u2013 If True, creates dataset from training.pt,\n  otherwise from test.pt.<\/p>\n<\/blockquote>\n\n<p>But this goes against the common practice of having a three-way split. For serious work, I need another <code>DataLoader<\/code> with a validation set. Also, it would be nice to specify the split proportions myself. They don't say what percentage of the dataset is reserved for testing, maybe I would like to change that.<\/p>\n\n<p>I assume that this is a conscious design decision. Everyone working on one of these datasets is supposed to use the same testset. That makes results comparable. But I still need to get a validation set out of the <code>trainloader<\/code>. Is it possible to split a <code>DataLoader<\/code> into two separate streams of data? <\/p>\n","answers":[{"AnswerId":"53389447","CreationDate":"2018-11-20T09:00:46.613","ParentId":null,"OwnerUserId":"497600","Title":null,"Body":"<p>Meanwhile, I stumbled upon the method <a href=\"https:\/\/pytorch.org\/docs\/master\/data.html#torch.utils.data.random_split\" rel=\"nofollow noreferrer\"><code>random_split<\/code><\/a>. So, you don't split the <code>DataLoader<\/code>, but you split the <code>Dataset<\/code>:<\/p>\n\n<pre><code>torch.utils.data.random_split(dataset, lengths)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53389871,"AnswerCount":0,"Tags":"<python><tensorflow><case>","CreationDate":"2018-11-20T09:26:49.333","AcceptedAnswerId":null,"OwnerUserId":10171000.0,"Title":"Tensorflow case function does not work when pred_fn_pairs is built from for loop","Body":"<p>I am having a weird issue here.<\/p>\n\n<p>I am simply trying to apply a mapping on tensor values according some intervals.\nFor example, let's map tensor's values that are in the interval [0, 10] to 1, in the interval [10, 20] to 2, in the interval [20, 30] to 3 and others to 0.<\/p>\n\n<p>If i build this mapping thanks to a list and iterates over it \"by hand\" it works fine.<\/p>\n\n<pre><code>def build_mult_mapping(l):\n    return lambda x: tf.case({\n                              tf.math.logical_and(tf.greater(x, tf.constant(l[0][0][0])), tf.less(x, tf.constant(l[0][0][1]))): lambda: tf.constant(l[0][1]),\n                              tf.math.logical_and(tf.greater(x, tf.constant(l[1][0][0])), tf.less(x, tf.constant(l[1][0][1]))): lambda: tf.constant(l[1][1]),\n                              tf.math.logical_and(tf.greater(x, tf.constant(l[2][0][0])), tf.less(x, tf.constant(l[2][0][1]))): lambda: tf.constant(l[2][1])\n                              },\n                              default=lambda: tf.constant(l[3]), exclusive=True)\n<\/code><\/pre>\n\n<p>It is ugly, but if i apply this <code>build_mult_mapping<\/code> version it works fine.<\/p>\n\n<pre><code>def f(x):\n    mult_mapping = build_mult_mapping([((0, 10), 1), ((10, 20), 2), ((20, 30), 3), 0])\n    return tf.map_fn(mult_mapping, x)\n\nx = tf.constant([5, 15, 25, 35])\nsess = tf.Session()\nsess.run(f(x))\n<\/code><\/pre>\n\n<p>If i run this block i get: <code>array([1, 2, 3, 0], dtype=int32)<\/code> which is the expected output.<\/p>\n\n<p>Now if i try to iterate over the <code>build_mult_mapping<\/code> argument <code>y<\/code> with a for loop, i get a weird behavior. Here is the new implementation<\/p>\n\n<pre><code>def build_mult_mapping(l):\n    return lambda x: tf.case({tf.math.logical_and(tf.greater(x, tf.constant(i[0][0])), tf.less(x, tf.constant(i[0][1]))): lambda: tf.constant(i[1])\n                              for i in l[:-1]\n                             },\n                             default=lambda: tf.constant(l[-1]), exclusive=True)\n<\/code><\/pre>\n\n<p>If i run the same block as before i get the following output <code>array([3, 3, 3, 0], dtype=int32)<\/code> whereas i am expecting to have the same as before.<\/p>\n\n<p>Any ideas on what is happening ?<\/p>\n","answers":[]}
+{"QuestionId":53389959,"AnswerCount":1,"Tags":"<tensorflow><object-detection><bounding-box><data-augmentation>","CreationDate":"2018-11-20T09:31:29.713","AcceptedAnswerId":53392673.0,"OwnerUserId":10673713.0,"Title":"Tensorflow Object Detection API Data Augmentation Bounding Boxes","Body":"<p>For Object Detection via the Tensorflow API using the model_main.py, when I use i.e. <code>random_horizontal_flip<\/code> in the <code>data_augmentation_options<\/code> in the <code>train_config<\/code> of my pipeline.config, are my bounding boxes also affected? This is very important, as otherwise these options are not applicable. <a href=\"https:\/\/stackoverflow.com\/q\/50293090\/10673713\">This<\/a> is the same question, but it was not properly answered.<\/p>\n","answers":[{"AnswerId":"53392673","CreationDate":"2018-11-20T12:06:51.847","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>Yes, the bounding boxes are affected in the same way.\nSpecifically for <code>random_horizontal_flip<\/code>, you can verify it by looking at the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/core\/preprocessor.py#L443\" rel=\"nofollow noreferrer\">function<\/a>, which also receives boxes.\nFlipping the bounding boxes is performed <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/core\/preprocessor.py#L517\" rel=\"nofollow noreferrer\">here<\/a>.\nNote not all augmentation options need bounding box altering, but those who do - alter the bounding box accordingly.<\/p>\n"}]}
+{"QuestionId":53390093,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><computer-vision><semantic-segmentation>","CreationDate":"2018-11-20T09:40:24.040","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"Metric for evaluating predicted bounding boxes from semantic segmentation on an object level outside of training","Body":"<h1>Context<\/h1>\n\n<p>For simplicity let us pretend we are performing semantic segmentation on a series of one pixel high images of width <em>w<\/em> with three channels (r, g, b) with <em>n<\/em> label classes. <\/p>\n\n<p>In other words, a single image might look like:<\/p>\n\n<pre><code>img = [\n    [r1, r2, ..., rw], # channel r\n    [g1, g2, ..., gw], # channel g\n    [b1, b2, ..., bw], # channel b\n]\n<\/code><\/pre>\n\n<p>and have dimensions <code>[3, w]<\/code>.<\/p>\n\n<p>then for a given image with <code>w=10<\/code> and <code>n=3<\/code> its labels ground truth might be:<\/p>\n\n<pre><code># ground \"truth\"\ntarget = np.array([\n  #0     1     2     3     4     5     6     7     8     9      # position\n  [0,    1,    1,    1,    0,    0,    1,    1,    1,    1],    # class 1\n  [0,    0,    0,    0,    1,    1,    1,    1,    0,    0],    # class 2\n  [1,    0,    0,    0,    0,    0,    0,    0,    0,    0],    # class 3\n])\n<\/code><\/pre>\n\n<p>and our model might predict as output:<\/p>\n\n<pre><code># prediction\noutput = np.array([\n  #0     1     2     3     4     5     6     7     8     9      # position\n  [0.11, 0.71, 0.98, 0.95, 0.20, 0.15, 0.81, 0.82, 0.95, 0.86], # class 1\n  [0.13, 0.17, 0.05, 0.42, 0.92, 0.89, 0.93, 0.93, 0.67, 0.21], # class 2\n  [0.99, 0.33, 0.20, 0.12, 0.15, 0.15, 0.20, 0.01, 0.02, 0.13], # class 3\n])\n<\/code><\/pre>\n\n<p>for further simplicity, let us transform our model's output by binarizing it with a cutoff of <code>0.9<\/code><\/p>\n\n<pre><code># binary mask with cutoff 0.9\nb_mask = np.array([\n  #0     1     2     3     4     5     6     7     8     9      # position\n  [0,    0,    1,    1,    0,    0,    0,    0,    1,    0],    # class 1\n  [0,    0,    0,    0,    1,    0,    1,    1,    0,    0],    # class 2\n  [1,    0,    0,    0,    0,    0,    0,    0,    0,    0],    # class 3\n])\n<\/code><\/pre>\n\n<p>Then if we were to look at the \"objects\" of each class the bounding boxes (or in this case just boundaries i.e. <code>[start, stop]<\/code> pixels) our predicted objects from the binary mask \"introduce\" an object:  <\/p>\n\n<pre><code># \"detected\" objects\np_obj = [\n  [[2, 3], [8, 8]],  # class 1\n  [[4, 4], [6, 7]],  # class 2\n  [[0, 0]]           # class 3\n] \n<\/code><\/pre>\n\n<p>compared to the objects of the ground truth:<\/p>\n\n<pre><code># true objects\nt_obj = [\n  [[1, 3], [6, 9]],  # class 1\n  [[4, 7]],          # class 2\n  [[0, 0]]           # class 3\n] \n<\/code><\/pre>\n\n<h1>Question<\/h1>\n\n<p>If I wanted a metric to describe the accuracy of the boundaries, on average, per object, what would be the appropriate metric?<\/p>\n\n<p>I understand IOU in the context of training a model which <em>predicts<\/em> bounding boxes, e.g. it is an object to object comparison, but what should one do when one object might be fragmented into several? <\/p>\n\n<h1>Goal<\/h1>\n\n<p>I would like a metric that, per class, gives me something like this:<\/p>\n\n<pre><code>class 1: [-1, 2]  # bounding boxes for class one, on average start one\n                  # pixel before they should and end two pixels after \n                  # they should\n\nclass 2: [ 0, 3]  # bounding boxes for class two, on average start \n                  # exactly where they should and end three pixels  \n                  # after they should\n\nclass 3: [ 3, -1] # bounding boxes for class three, on average start \n                  # three pixels after where they begin and end one \n                  # pixels too soon\n<\/code><\/pre>\n\n<p>but I am not sure how to best approach this when a single object is fragmented into several...<\/p>\n","answers":[{"AnswerId":"53420141","CreationDate":"2018-11-21T20:38:56.880","ParentId":null,"OwnerUserId":"5623899","Title":null,"Body":"<h1>Assumption<\/h1>\n\n<p>You ask specifically about the 1D case, so we will solve the 1D case here, but the method is essentially the same for 2D.<\/p>\n\n<p>Let us assume you have two ground truth bounding boxes: box 1 and box 2.<\/p>\n\n<p>Further, let us assume that our model is not so great and predicts more than 2 boxes \n(maybe it found something new, maybe it broke one box into two).<\/p>\n\n<p>For this demonstration let us consider that this is what we are working with:<\/p>\n\n<pre><code># labels\n# box 1: x----y \n# box 2: x++++y\n# 0  1  2  3  4  5  6  7  8  9  10 11 12 13 14 15 16 17 18 19 20\n#             x--------y        x+++++++++++++++++++++++++++++y     TRUTH\n#             a-----------b                                         PRED 1, BOX 1\n#                   a+++++++++++++++++b                             PRED 2, BOX 2\n#                a++++++++++++++++++++++++++++++++b                 PRED 3, BOX 2\n<\/code><\/pre>\n\n<h1>Core Problem<\/h1>\n\n<p>What you want is in effect a score on the alignment of your predictions to the targets.... but oh no! which targets \nbelong to which predictions? <\/p>\n\n<p>Pick your distance function of choice and pair each prediction with a target based on that function.\nIn this case I will use a modified intersection over union (IOU) for the 1D case. \nI chose this function as I wanted both PRED 2 and 3 from the above diagram to align to box 2.<\/p>\n\n<p>With a score for each prediction, pair it with the target that produced the best score.<\/p>\n\n<p>Now with a one-to-one prediction-target pair, calculate whatever it is that you want.<\/p>\n\n<h1>Demo with above assumption<\/h1>\n\n<p>from the above assumptions:<\/p>\n\n<pre><code>pred_boxes = [\n    [4,  8],\n    [6, 12],\n    [5, 16]\n]\n\ntrue_boxes = [\n    [4,   7],\n    [10, 20]\n]\n<\/code><\/pre>\n\n<p>a 1d version of intersection over union:<\/p>\n\n<pre><code>def iou_1d(predicted_boundary, target_boundary):\n  '''Calculates the intersection over union (IOU) based on a span.\n\n  Notes:\n    boundaries are provided in the the form of [start, stop].\n    boundaries where start = stop are accepted\n    boundaries are assumed to be only in range [0, int &lt; inf)\n\n  Args:\n    predicted_boundary (list): the [start, stop] of the predicted boundary\n    target_boundary (list): the ground truth [start, stop] for which to compare\n\n  Returns:\n    iou (float): the IOU bounded in [0, 1]\n  '''\n\n  p_lower, p_upper = predicted_boundary\n  t_lower, t_upper = target_boundary\n\n  # boundaries are in form [start, stop] and 0&lt;= start &lt;= stop\n  assert 0&lt;= p_lower &lt;= p_upper\n  assert 0&lt;= t_lower &lt;= t_upper\n\n   # no overlap, pred is too far left or pred is too far right\n  if p_upper &lt; t_lower or p_lower &gt; t_upper:\n    return 0\n\n  if predicted_boundary == target_boundary:\n    return 1\n\n  intersection_lower_bound = max(p_lower, t_lower)\n  intersection_upper_bound = min(p_upper, t_upper)\n\n\n  intersection = intersection_upper_bound - intersection_lower_bound\n  union = max(t_upper, p_upper) - min(t_lower, p_lower)  \n  union = union if union != 0 else 1  \n  return min(intersection \/ union, 1)\n<\/code><\/pre>\n\n<p>some simple helpers:<\/p>\n\n<pre><code>from math import sqrt\ndef euclidean(u, v):\n  return sqrt((u[0]-v[0])**2 + (u[1]-v[1])**2)\n\ndef mean(arr):\n  return sum(arr) \/ len(arr)\n<\/code><\/pre>\n\n<p>how we align our boundaries:<\/p>\n\n<pre><code>def align_1d(predicted_boundary, target_boundaries, alignment_scoring_fn=iou_1d, take=max):\n  '''Aligns predicted_bondary to the closest target_boundary based on the \n    alignment_scoring_fn\n\n  Args:\n    predicted_boundary (list): the predicted boundary in form of [start, stop]\n\n    target_boundaries (list): a list of all valid target boundaries each having\n      form [start, stop]\n\n    alignment_scoring_fn (function): a function taking two arguments each of \n      which is a list of two elements, the first assumed to be the predicted\n      boundary and the latter the target boundary. Should return a single number.\n\n    take (function): should either be min or max. Selects either the highest or\n      lower score according to the alignment_scoring_fn\n\n  Returns:\n    aligned_boundary (list): the aligned boundary in form [start, stop]\n  '''\n  scores = [\n      alignment_scoring_fn(predicted_boundary, target_boundary) \n      for target_boundary in target_boundaries\n  ]\n\n\n\n  # boundary did not align to any boxes, use fallback scoring mechanism to break\n  # tie\n  if not any(scores):\n    scores = [\n      1 \/ euclidean(predicted_boundary, target_boundary)\n      for target_boundary in target_boundaries\n    ]\n\n  aligned_index = scores.index(take(scores))\n  aligned = target_boundaries[aligned_index]\n  return aligned\n<\/code><\/pre>\n\n<p>how we calculate difference:<\/p>\n\n<pre><code>def diff(u, v):\n  return [u[0] - v[0], u[1] - v[1]]\n<\/code><\/pre>\n\n<p>combine it all into one:<\/p>\n\n<pre><code>def aligned_distance_1d(predicted_boundaries, target_boundaries, alignment_scoring_fn=iou_1d, take=max, distance_fn=diff, aggregate_fn=mean):\n  '''Returns the aggregated distance of predicted boundings boxes to their aligned bounding box based on alignment_scoring_fn and distance_fn\n\n  Args:\n    predicted_boundaries (list): a list of all valid target boundaries each \n      having form [start, stop]\n\n    target_boundaries (list): a list of all valid target boundaries each having\n      form [start, stop]\n\n    alignment_scoring_fn (function): a function taking two arguments each of \n      which is a list of two elements, the first assumed to be the predicted\n      boundary and the latter the target boundary. Should return a single number.\n\n    take (function): should either be min or max. Selects either the highest or\n      lower score according to the alignment_scoring_fn\n\n    distance_fn (function): a function taking two lists and should return a\n      single value.\n\n    aggregate_fn (function): a function taking a list of numbers (distances \n      calculated by distance_fn) and returns a single value (the aggregated \n      distance)\n\n  Returns:\n    aggregated_distnace (float): return the aggregated distance of the \n      aligned predicted_boundaries\n\n      aggregated_fn([distance_fn(pair) for pair in paired_boundaries(predicted_boundaries, target_boundaries)])\n  '''\n\n\n  paired = [\n      (predicted_boundary, align_1d(predicted_boundary, target_boundaries, alignment_scoring_fn))\n      for predicted_boundary in predicted_boundaries\n  ]\n  distances = [distance_fn(*pair) for pair in paired]\n  aggregated = [aggregate_fn(error) for error in zip(*distances)]\n  return aggregated\n<\/code><\/pre>\n\n<p>run:<\/p>\n\n<pre><code>aligned_distance_1d(pred_boxes, true_boxes)\n\n# [-3.0, -3.6666666666666665]\n<\/code><\/pre>\n\n<p>Note, for many predictions and many targets there are many ways to optimize the code. Here, I broke up the main functional chunks so it is clear what is going on. <\/p>\n\n<p>Now does this make sense? Well since I wanted pred 2 and 3 to align to box 2, yes, both starts are prior the truth and both end prematurely.<\/p>\n\n<h1>Solution to question asked<\/h1>\n\n<p>copy pasted your examples:<\/p>\n\n<pre><code># \"detected\" objects\np_obj = [\n  [[2, 3], [8, 8]],  # class 1\n  [[4, 4], [6, 7]],  # class 2\n  [[0, 0]]           # class 3\n] \n\n# true objects\nt_obj = [\n  [[1, 3], [6, 9]],  # class 1\n  [[4, 7]],          # class 2\n  [[0, 0]]           # class 3\n] \n<\/code><\/pre>\n\n<p>since you know the boxes per class this is easy:<\/p>\n\n<pre><code>[\n    aligned_distance_1d(p_obj[cls_no], t_obj[cls_no])\n    for cls_no in range(len(t_obj))\n]\n\n\n# [[1.5, -0.5], [1.0, -1.5], [0.0, 0.0]]\n<\/code><\/pre>\n\n<p>Does this output make sense?<\/p>\n\n<p>Starting with a sanity check, let us look at class 3. The average distance of [start, stop] are both 0. Makes sense.<\/p>\n\n<p>How about class 1? both predictions start too late (2 > 1, 8 > 6) but only one ends too soon (8 &lt; 9). So makes sense.<\/p>\n\n<p>Now let us look at class 2, which is why it seems you asked the question (more predictions than targets).<\/p>\n\n<p>If we were to draw what the score suggests it would be:<\/p>\n\n<pre><code>#  0  1  2  3  4  5  6  7  8  9\n#              ----------        # truth [4, 7]\n#                 ++             # pred  [4 + 1, 7 - 1.5]\n<\/code><\/pre>\n\n<p>It doesn't look so great, but this is just an example...<\/p>\n\n<p>Does this make sense? Yes \/ no. Yes in terms of how we calculated the metric. One stoped 3 values too soon the other started 2 too late.\nNo in the sense that neither of your predictions actually cover the value 5, and yet this metric leads you to believe that is the case...<\/p>\n\n<h1>Conclusion<\/h1>\n\n<p>Is this a faulty metric?<\/p>\n\n<p>Depends on what you are using it for \/ trying to show. \nHowever since you use a binary mask to generate you predicted boundaries, that is a non negligible root of this problem. Perhaps there is a better strategy to get boundaries from your label probabilities.<\/p>\n"}]}
+{"QuestionId":53390625,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-20T10:08:49.260","AcceptedAnswerId":53394116.0,"OwnerUserId":8723840.0,"Title":"couldn't run embedding network Keras with multiplue input","Body":"<p>I have tried to run simple keras model with one embedding layer with 9 inputs. But I always get two errors, depending the layer after embedding.\nI tried to use 2 different representations of data, but I get the same.\nNow, what I have:<\/p>\n\n<p>1.I'm using my own fit generator, which yeild data:<\/p>\n\n<pre><code>(list of shapes of input data) -\n[(25,), (25,), (25,), (25, 24), (25, 11), (25, 10), (25, 28), (25, 8), (25, 7)] \n\n\nfeatures = [['id1',1], ['id2',1],\n ['id3',1], ['id4',24],\n ['id5',11], ['id6',10], ['id7',28], ['id8',8], ['id9',7]]\n\nembeddings = []\ninputs = []\nfor idx, feature in enumerate(features):\n      meta_input = Input(shape=(feature[1],), name = feature[0] + '_input')\n      sqrt = int(np.sqrt(feature[1]))\n\n      embedding = Embedding(feature[1], 1, input_length=1,name = feature[0] + '_embed')(meta_input)\n      fl = Flatten()(embedding)\n      embeddings.append(fl)\n      inputs.append(meta_input)\n\nx = Concatenate()(embeddings)    \ndense_meta_1 =  Dense(256, activation='relu')(x) #x\ndrop_meta = Dropout(0.2)(dense_meta_1)\n\ndense_meta_2 = Dense(1)(drop_meta)\n\n\nmodel = Model(inputs, dense_meta_2)\n\nmodel.compile(optimizer='Adam', loss='mean_squared_error', metrics= \n['mae']) \nhistory = model.fit_generator(my_gen_v2(batch_size, batch_folder, steps), epochs=1, steps_per_epoch=steps,\n                         max_queue_size=1)\n<\/code><\/pre>\n\n<p>so when I use flatten layers  - I got this message (some part):<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Matrix size-incompatible: In[0]: [25,91], In[1]: [9,256]\n       [[node dense_25\/MatMul (defined at \/home\/human\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py:1076)  = MatMul[T=DT_FLOAT, _class=[\"loc:@training_7\/Adam\/gradients\/dense_25\/MatMul_grad\/MatMul\"], transpose_a=false, transpose_b=false, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](concatenate_16\/concat, dense_25\/kernel\/read)]]\n       [[{{node metrics_11\/mean_absolute_error\/Mean_1\/_1219}} = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_1116_metrics_11\/mean_absolute_error\/Mean_1\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>but when I use Reshape layers:<\/p>\n\n<pre><code>embedding = Reshape(target_shape=(1,), name = feature[0] + '_reshape')(embedding)\n<\/code><\/pre>\n\n<p>I'v got this:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Input to reshape is a tensor with 600 values, but the requested shape has 25\n       [[node race_reshape\/Reshape (defined at \/home\/human\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py:1898)  = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](race_embed_16\/GatherV2, race_reshape\/Reshape\/shape)]]\n       [[{{node metrics_12\/mean_absolute_error\/Mean_1\/_1417}} = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_1098_metrics_12\/mean_absolute_error\/Mean_1\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>There are no simillar questions on stackoverflow, only about images shapes. Please help me resolve this, coz I spend a lot of time for this(<\/p>\n","answers":[{"AnswerId":"53394116","CreationDate":"2018-11-20T13:30:47.140","ParentId":null,"OwnerUserId":"8723840","Title":null,"Body":"<p>The problem was resolved by changing input_length in Embedding layer to input shape of feature (feature[1] in my example)<\/p>\n"}]}
+{"QuestionId":53391061,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml><tensorflow-estimator>","CreationDate":"2018-11-20T10:33:26.643","AcceptedAnswerId":null,"OwnerUserId":6695801.0,"Title":"Accumulate predictions in output tensor for Google Cloud serving","Body":"<p>I'm new to TF and trying to deploy <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/recurrent_quickdraw\" rel=\"nofollow noreferrer\">Quick, Draw!<\/a> model on Google Cloud.\nIt uses custom <code>Estimator<\/code> and takes 1 <code>tfexample<\/code> and returns classification tensor of shape <code>(8)<\/code>. \nOne of the requirements for models to be deployed is that input and output tensors should support batching, e.g. their outer dimensions should be <code>None<\/code>. <\/p>\n\n<p>Doing so for input tensor is easy since we have <code>serving_input_receiver_fn<\/code> under control, as for output tensor I'm confused because whether I change output tensor I also need to refactor whole train\/evaluate operation. Are there any shortcuts to accomplish that?<\/p>\n","answers":[{"AnswerId":"56601928","CreationDate":"2019-06-14T16:16:42.637","ParentId":null,"OwnerUserId":"4231563","Title":null,"Body":"<p>I was able to confirm that indeed the output from a saved model exported from a TensorFlow estimator will carry the same batch size as the input.<\/p>\n\n<p>There are a few caveats if you try to export and deploy this particular tutorial to Google Cloud's AI Platform Prediction service though:<\/p>\n\n<ul>\n<li><p>The <code>model_fn<\/code> needs to be updated to handle the <code>PREDICT<\/code> mode properly.<\/p><\/li>\n<li><p>The Prediction service actually requires the outer dimension to be <code>None<\/code> in the saved model, this also requires some code change in the tutorial since it specifies the batch size.<\/p><\/li>\n<li><p>The tutorial code does not include the call to <code>estimator.export_saved_model<\/code>, in particular you'll need to implement a <code>serving_input_receiver_fn<\/code>.<\/p><\/li>\n<li><p>There seems to be a bug in TF that causes restoration of <code>cudnn_lstm<\/code> to fail: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23898\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23898<\/a>.  So in particular at the moment if the model is trained with <code>cudnn_lstm<\/code> it can't be deployed as a saved model.<\/p><\/li>\n<\/ul>\n\n<p>I have made these changes to the tutorial in a fork: <a href=\"https:\/\/github.com\/dizcology\/models\/tree\/quickdraw-gcp\/tutorials\/rnn\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dizcology\/models\/tree\/quickdraw-gcp\/tutorials\/rnn<\/a><\/p>\n"}]}
+{"QuestionId":53391253,"AnswerCount":1,"Tags":"<python><tensorflow><checkpoint>","CreationDate":"2018-11-20T10:43:44.483","AcceptedAnswerId":null,"OwnerUserId":4958717.0,"Title":"How to extract weights and other variable values from tensorflow checkpoint without restoring the graph?","Body":"<p>Provided a checkpoint file but no meta graph or code that produced the network, I want to extract the stored values of the variables in the checkpoint file.<\/p>\n\n<p>So without restoring the graph, how do I extract the values stored in thr checkpoint. I could potentially convert everything from the checkpoint to a dictionary of numpy arrays or something similar.<\/p>\n","answers":[{"AnswerId":"53391481","CreationDate":"2018-11-20T10:56:42.717","ParentId":null,"OwnerUserId":"4958717","Title":null,"Body":"<p>Found the solution:<\/p>\n\n<pre><code>reader = tf.train.NewCheckpointReader(\"\/path\/to\/checkpoint\")\nshapes_dict = reader.get_variable_to_shape_map()  # use it to get the variable names\nextracted_values = reader.get_tensor(shapes_dict.keys()[0])\n# array([[ 0.       , -1.8053141],\n#        [-1.5647348,  0.       ]], dtype=float32)\n<\/code><\/pre>\n\n<p>The <code>tf.train.NewCheckpointReader<\/code> is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/NewCheckpointReader\" rel=\"nofollow noreferrer\">not really documented<\/a> in current documentation of API r1.12.\nBut you can see the usage example in the source code <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a6d8ffae097d0132989ae4688d224121ec6d8f35\/tensorflow\/python\/training\/saver_test.py#L2407\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53391412,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-20T10:53:04.630","AcceptedAnswerId":null,"OwnerUserId":7275235.0,"Title":"Keras: IndexError: tuple index out of range - Conv2D","Body":"<p>I'm new in Keras and Tensorflow. I'm trying to run some specific code based on Keras and Tensorflow. In that code, I prepared dataset from tfrecordes files and reshape it with <code>tf.reshape(image, [32, 32, 3])<\/code>, in my model, I need to apply a Conv2D layer on this image output, I using Keras, in my code below:<\/p>\n\n<pre><code># Image\nx = k.layers.Input(shape=input_shape,tensor=img)\n# ReLU Conv1\nconv1 = k.layers.Conv2D(filters = 256, kernel_size = 9, strides=1,padding='valid', activation='relu', name='conv1')(x)\n<\/code><\/pre>\n\n<p>the problem is that, when I run the code, I got this error:<\/p>\n\n<blockquote>\n  <p>self.kernel_size[i],\n  IndexError: tuple index out of range<\/p>\n<\/blockquote>\n\n<p>so, after many searches, I found that this error may be fixed by change Keras ver from 2.2 to 1.2.\nI do that but actually, I must change all of my code, do anybody how can fix this error just in Keras 2.2, without using Karas 1.2? <\/p>\n","answers":[{"AnswerId":"53391655","CreationDate":"2018-11-20T11:06:46.463","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>I think you need a 9 * 9 kernel. Change <code>kernel_size=9<\/code> to <code>kernel_size=( 9 , 9 )<\/code>\nEdited code line :<\/p>\n\n<pre><code>conv1 = k.layers.Conv2D(filters = 256, kernel_size = ( 9 , 9 ), strides=1,padding='valid', activation='relu', name='conv1')(x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53391543,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-20T11:00:24.543","AcceptedAnswerId":53391863.0,"OwnerUserId":3946761.0,"Title":"Illegal instruction (core dumped) -Tensorflow GPU","Body":"<p>I have installed Tensorflow-GPU version 1.9.0 and simple tensorflow import statement gives exception \"Illegal instruction (core dumped)\". If I downgrade tensorflow version to 1.5.0, it works fine. How to fix this issue for higher version as I need to work with it? <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53391863","CreationDate":"2018-11-20T11:18:59.607","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Starting with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/v1.5.1\" rel=\"nofollow noreferrer\">v1.5.1<\/a> on Linux and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\/v1.6.0\" rel=\"nofollow noreferrer\">v1.6.0<\/a> on other platforms, the official TensorFlow distribution is compiled with <a href=\"https:\/\/en.wikipedia.org\/wiki\/Advanced_Vector_Extensions\" rel=\"nofollow noreferrer\">AVX<\/a> instructions, meaning that older CPUs will not work with it (you can look up model compatibility, but it does not have to be an ancient CPU, it happened to me on an old Core i7).<\/p>\n\n<p>If you want to use official releases, the only solution is to switch to a different hardware or to stick to the older version. There have been <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18689\" rel=\"nofollow noreferrer\">requests for support for older CPUs<\/a> (and some people have uploaded their own <a href=\"https:\/\/github.com\/maxhgerlach\/tensorflow-1.8.0-ubuntu16.04-py27-no_avx-xeon_x5650\" rel=\"nofollow noreferrer\">build for a particular configuration<\/a>, if it works for you and you trust it), but the general answer is that if you need specific support for your platform you can always build it yourself, disabling AVX optimizations (see the <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"nofollow noreferrer\">installation guide<\/a>).<\/p>\n"}]}
+{"QuestionId":53391618,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-20T11:05:00.990","AcceptedAnswerId":53393968.0,"OwnerUserId":6925077.0,"Title":"Tensor Tensor(\"predictions\/Softmax:0\", shape=(?, 1000), dtype=float32) is not an element of this graph","Body":"<p>I am trying to follow a <a href=\"https:\/\/machinelearningmastery.com\/use-pre-trained-vgg-model-classify-objects-photographs\/\" rel=\"nofollow noreferrer\">simple tutorial<\/a> on how to use a pre-trained VGG model for image classification. The code which I have: <\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nfrom keras.preprocessing.image import load_img, img_to_array\nfrom keras.applications.vgg16 import preprocess_input, decode_predictions\n\nimport numpy as np\n\nclass KerasModel(object):\n    def __init__(self):\n        self.model = VGG16()\n    def evaluate(self):\n        image = load_img('mug.jpg', target_size=(224,224))\n        image = img_to_array(image)\n        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))\n        image = preprocess_input(image)\n        yhat = self.model.predict(image)\n        label = decode_predictions(yhat)\n        label = label[0][0]\n        return ('%s (%.2f%%)' % (label[1]), label[2]*100)\n<\/code><\/pre>\n\n<p>This gives the error: Tensor Tensor(\"predictions\/Softmax:0\", shape=(?, 1000), dtype=float32) is not an element of this graph.<\/p>\n\n<p>After some searching for this error I got to this code: <\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nfrom keras.preprocessing.image import load_img, img_to_array\nfrom keras.applications.vgg16 import preprocess_input, decode_predictions\n\nimport numpy as np\n\nimport tensorflow as tf\ngraph = tf.get_default_graph()\n\n\nclass KerasModel(object):\n    def __init__(self):\n        self.model = VGG16()\n    def evaluate(self):\n        image = load_img('mug.jpg', target_size=(224,224))\n        image = img_to_array(image)\n        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))\n        image = preprocess_input(image)\n        with graph.as_default():\n            yhat = self.model.predict(image)\n        label = decode_predictions(yhat)\n        label = label[0][0]\n        return ('%s (%.2f%%)' % (label[1]), label[2]*100)\n<\/code><\/pre>\n\n<p>But this still results in the same error. Could someone please help me out? I don't understand what I am doing wrong because the tutorial seems to work for everyone. <\/p>\n\n<p>Model summary:<\/p>\n\n<pre><code> _________________________________________________________________\nxvision | Layer (type)                 Output Shape              Param #   \nxvision | =================================================================\nxvision | input_1 (InputLayer)         (None, 224, 224, 3)       0         \nxvision | _________________________________________________________________\nxvision | block1_conv1 (Conv2D)        (None, 224, 224, 64)      1792      \nxvision | _________________________________________________________________\nxvision | block1_conv2 (Conv2D)        (None, 224, 224, 64)      36928     \nxvision | _________________________________________________________________\nxvision | block1_pool (MaxPooling2D)   (None, 112, 112, 64)      0         \nxvision | _________________________________________________________________\nxvision | block2_conv1 (Conv2D)        (None, 112, 112, 128)     73856     \nxvision | _________________________________________________________________\nxvision | block2_conv2 (Conv2D)        (None, 112, 112, 128)     147584    \nxvision | _________________________________________________________________\nxvision | block2_pool (MaxPooling2D)   (None, 56, 56, 128)       0         \nxvision | _________________________________________________________________\nxvision | block3_conv1 (Conv2D)        (None, 56, 56, 256)       295168    \nxvision | _________________________________________________________________\nxvision | block3_conv2 (Conv2D)        (None, 56, 56, 256)       590080    \nxvision | _________________________________________________________________\nxvision | block3_conv3 (Conv2D)        (None, 56, 56, 256)       590080    \nxvision | _________________________________________________________________\nxvision | block3_pool (MaxPooling2D)   (None, 28, 28, 256)       0         \nxvision | _________________________________________________________________\nxvision | block4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160   \nxvision | _________________________________________________________________\nxvision | block4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808   \nxvision | _________________________________________________________________\nxvision | block4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808   \nxvision | _________________________________________________________________\nxvision | block4_pool (MaxPooling2D)   (None, 14, 14, 512)       0         \nxvision | _________________________________________________________________\nxvision | block5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808   \nxvision | _________________________________________________________________\nxvision | block5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808   \nxvision | _________________________________________________________________\nxvision | block5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808   \nxvision | _________________________________________________________________\nxvision | block5_pool (MaxPooling2D)   (None, 7, 7, 512)         0         \nxvision | _________________________________________________________________\nxvision | flatten (Flatten)            (None, 25088)             0         \nxvision | _________________________________________________________________\nxvision | fc1 (Dense)                  (None, 4096)              102764544 \nxvision | _________________________________________________________________\nxvision | fc2 (Dense)                  (None, 4096)              16781312  \nxvision | _________________________________________________________________\nxvision | predictions (Dense)          (None, 1000)              4097000   \nxvision | =================================================================\nxvision | Total params: 138,357,544\nxvision | Trainable params: 138,357,544\nxvision | Non-trainable params: 0\nxvision | _________________________________________________________________\nxvision | None\n<\/code><\/pre>\n","answers":[{"AnswerId":"53393968","CreationDate":"2018-11-20T13:23:01.453","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>As your code is fine, running with a clean environment should solve it. <br><\/p>\n\n<ul>\n<li><p>Clear keras cache at <code>~\/.keras\/<\/code><\/p><\/li>\n<li><p>Run on a new environment, with the right packages (can be done easily with anaconda)<\/p><\/li>\n<li><p>Make sure you are on a fresh session, <code>keras.backend.clear_session()<\/code> should remove all existing tf graphs.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53391669,"AnswerCount":0,"Tags":"<android><tensorflow><kotlin><tensorflow-lite>","CreationDate":"2018-11-20T11:07:37.943","AcceptedAnswerId":null,"OwnerUserId":10456906.0,"Title":"TensorFlow Lite: Batching with Mobilenet_V2","Body":"<p>I want to do batching with Mobilenet_V2_1.0_224 in TensorFlow Lite . <\/p>\n\n<p>I tried this answer <a href=\"https:\/\/stackoverflow.com\/questions\/52783747\/how-to-do-batching-with-tensorflow-lite\/53125376#53125376\">(https:\/\/stackoverflow.com\/questions\/52783747\/how-to-do-batching-with-tensorflow-lite\/53125376#53125376<\/a>), but this works only for Inception_V4. <\/p>\n\n<p>When I try to use a batch size >1 (in this case batch_size 2) for Mobilenet i get the following error:<\/p>\n\n<pre><code>  ComposedException 1 :\n    java.lang.NullPointerException: Internal error: Cannot allocate memory for the interpreter: tensorflow\/contrib\/lite\/kernels\/reshape.cc:68 num_input_elements != num_output_elements (2002 != 1001)Node 64 failed to prepare.\n\n    at org.tensorflow.lite.NativeInterpreterWrapper.allocateTensors(Native Method)\n    at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:118)\n    at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:168)\n    at org.tensorflow.lite.Interpreter.run(Interpreter.java:145)\n<\/code><\/pre>\n\n<p>Edit:\nThis is the tflite code resulting in this error:<\/p>\n\n<pre><code>    DIM_BATCH_SIZE = 4\n    var tfliteOptions: Interpreter.Options = Interpreter.Options()\n    tfliteOptions.setUseNNAPI(true)\n    var interpreter = Interpreter(loadModelFile(assetManager, \"mobilenet_v2_1.0_224.tflite\")!!, tfliteOptions)    \n    interpreter!!.resizeInput(0,intArrayOf(DIM_BATCH_SIZE,224,224,3))\n\n    interpreter!!.run(imgData, labelProb)\n<\/code><\/pre>\n\n<p>imgData is a ByteBuffer containing the preprocessed images, labelProb the FloatArray for the predictions.<\/p>\n","answers":[]}
+{"QuestionId":53392793,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><computer-vision>","CreationDate":"2018-11-20T12:14:59.030","AcceptedAnswerId":53462075.0,"OwnerUserId":10679504.0,"Title":"Tensorflow Reshape error with custom pooling\/unpooling layer","Body":"<p>I am trying to implement a smaller scale version of SegNet described in this paper (<a href=\"https:\/\/arxiv.org\/pdf\/1511.00561.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1511.00561.pdf<\/a>), but I'm trying to tailor it towards detecting edges<\/p>\n\n<p>Dataset:\nI am using the BSDS500 boundary dataset, I cropped and rotated the images so their sizes are 320x480x3 instead of 321x481x3<\/p>\n\n<p>Input shapes, 200 training images and 100 validation images:<\/p>\n\n<pre><code>x_train: (200, 320, 480, 3)\nx_val: (100, 320, 480, 3)\ny_train: (200, 153600)\ny_val: (100, 153600)\n<\/code><\/pre>\n\n<p>Framework:\nI am using keras with tensorflow backend <\/p>\n\n<p>These are the functions I am using for the custom pooling and unpooling layers:<\/p>\n\n<pre><code>def pool_argmax2D(x, pool_size=(2,2), strides=(2,2)):\n    padding = 'SAME'\n    pool_size = [1, pool_size[0], pool_size[1], 1]\n    strides = [1, strides[0], strides[1], 1]\n    ksize = [1, pool_size[0], pool_size[1], 1]\n    output, argmax = tf.nn.max_pool_with_argmax(\n        x,\n        ksize = ksize,\n        strides = strides,\n        padding = padding\n    )\n\n    return [output, argmax]\n\n\ndef unpool2D(pool, argmax, ksize=(2,2)):\n    with tf.variable_scope(\"unpool\"):\n        input_shape =  tf.shape(pool)\n        output_shape = [input_shape[0], \n                        input_shape[1] * ksize[0], \n                        input_shape[2] * ksize[1], \n                        input_shape[3]]\n\n        flat_input_size = tf.cumprod(input_shape)[-1]\n        flat_output_shape = tf.cast([output_shape[0], \n                            output_shape[1] * output_shape[2] * output_shape[3]], tf.int64)\n\n        pool_ = tf.reshape(pool, [flat_input_size])\n        batch_range = tf.reshape(tf.range(tf.cast(output_shape[0], tf.int64), dtype=tf.int64),\n                                shape=[input_shape[0], 1, 1, 1])\n\n        b = tf.ones_like(argmax) * batch_range\n        b = tf.reshape(b, [flat_input_size, 1])\n\n        ind_ = tf.reshape(argmax, [flat_input_size, 1]) % flat_output_shape[1]\n        ind_ = tf.concat([b, ind_], 1)\n        ret = tf.scatter_nd(ind_, pool_, shape=flat_output_shape)\n        ret = tf.reshape(ret, output_shape)\n        return ret\n<\/code><\/pre>\n\n<p>This is the code for the model:<\/p>\n\n<pre><code>batch_size = 4\nkernel = 3\npool_size=(2,2)\nimg_shape = (320,480,3)\n\n\ninputs = Input(shape=img_shape, name='main_input')\n\nconv_1 = Conv2D(32, (kernel, kernel), padding=\"same\")(inputs)\nconv_1 = BatchNormalization()(conv_1)\nconv_1 = Activation(\"relu\")(conv_1)\nconv_2 = Conv2D(32, (kernel, kernel), padding=\"same\")(conv_1)\nconv_2 = BatchNormalization()(conv_2)\nconv_2 = Activation(\"relu\")(conv_2)\n\npool_1, mask_1 = Lambda(pool_argmax2D, arguments={'pool_size': pool_size, 'strides': pool_size})(conv_2)\n\nconv_3 = Conv2D(64, (kernel, kernel), padding=\"same\")(pool_1)\nconv_3 = BatchNormalization()(conv_3)\nconv_3 = Activation(\"relu\")(conv_3)\nconv_4 = Conv2D(64, (kernel, kernel), padding=\"same\")(conv_3)\nconv_4 = BatchNormalization()(conv_4)\nconv_4 = Activation(\"relu\")(conv_4)\n\npool_2, mask_2 = Lambda(pool_argmax2D, arguments={'pool_size': pool_size, 'strides': pool_size})(conv_4)\n\nconv_5 = Conv2D(64, (kernel, kernel), padding=\"same\")(pool_2)\nconv_5 = BatchNormalization()(conv_5)\nconv_5 = Activation(\"relu\")(conv_5)\n\nunpool_1 = Lambda(unpool2D, output_shape = (160,240,64), arguments={'ksize':pool_size, 'argmax': mask_2})(conv_5)\n\nconv_6 = Conv2D(64, (kernel, kernel), padding=\"same\")(unpool_1)\nconv_6 = BatchNormalization()(conv_6)\nconv_6 = Activation(\"relu\")(conv_6)\nconv_7 = Conv2D(64, (kernel, kernel), padding=\"same\")(conv_6)\nconv_7 = BatchNormalization()(conv_7)\nconv_7 = Activation(\"relu\")(conv_7)\n\nunpool_2 = Lambda(unpool2D, output_shape = (320,480,64), arguments={'ksize':pool_size, 'argmax': mask_1})(conv_7)\n\nconv_8 = Conv2D(32, (kernel, kernel), padding=\"same\")(unpool_2)\nconv_8 = BatchNormalization()(conv_8)\nconv_8 = Activation(\"relu\")(conv_8)\nconv_9 = Conv2D(32, (kernel, kernel), padding=\"same\")(conv_8)\nconv_9 = BatchNormalization()(conv_9)\nconv_9 = Activation(\"relu\")(conv_9)\n\nconv_10 = Conv2D(1, (1, 1), padding=\"same\")(conv_9)\nconv_10 = BatchNormalization()(conv_10)\n\nflatten_1 = Flatten()(conv_10)\n\noutputs = Activation('softmax')(flatten_1)\n\nmodel = Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n\n<p>The model compiles properly when I run:<\/p>\n\n<pre><code>model.compile(optimizer='adam', loss='mean_absolute_error', metrics=['accuracy'])\nmodel.summary()\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nmain_input (InputLayer)      (None, 320, 480, 3)       0         \n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 320, 480, 32)      896       \n_________________________________________________________________\nbatch_normalization_1 (Batch (None, 320, 480, 32)      128       \n_________________________________________________________________\nactivation_1 (Activation)    (None, 320, 480, 32)      0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 320, 480, 32)      9248      \n_________________________________________________________________\nbatch_normalization_2 (Batch (None, 320, 480, 32)      128       \n_________________________________________________________________\nactivation_2 (Activation)    (None, 320, 480, 32)      0         \n_________________________________________________________________\nlambda_1 (Lambda)            [(None, 160, 240, 32), (N 0         \n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 160, 240, 64)      18496     \n_________________________________________________________________\nbatch_normalization_3 (Batch (None, 160, 240, 64)      256       \n_________________________________________________________________\nactivation_3 (Activation)    (None, 160, 240, 64)      0         \n_________________________________________________________________\nconv2d_4 (Conv2D)            (None, 160, 240, 64)      36928     \n_________________________________________________________________\nbatch_normalization_4 (Batch (None, 160, 240, 64)      256       \n_________________________________________________________________\nactivation_4 (Activation)    (None, 160, 240, 64)      0         \n_________________________________________________________________\nlambda_2 (Lambda)            [(None, 80, 120, 64), (No 0         \n_________________________________________________________________\nconv2d_5 (Conv2D)            (None, 80, 120, 64)       36928     \n_________________________________________________________________\nbatch_normalization_5 (Batch (None, 80, 120, 64)       256       \n_________________________________________________________________\nactivation_5 (Activation)    (None, 80, 120, 64)       0         \n_________________________________________________________________\nlambda_3 (Lambda)            (None, 160, 240, 64)      0         \n_________________________________________________________________\nconv2d_6 (Conv2D)            (None, 160, 240, 64)      36928     \n_________________________________________________________________\nbatch_normalization_6 (Batch (None, 160, 240, 64)      256       \n_________________________________________________________________\nactivation_6 (Activation)    (None, 160, 240, 64)      0         \n_________________________________________________________________\nconv2d_7 (Conv2D)            (None, 160, 240, 64)      36928     \n_________________________________________________________________\nbatch_normalization_7 (Batch (None, 160, 240, 64)      256       \n_________________________________________________________________\nactivation_7 (Activation)    (None, 160, 240, 64)      0         \n_________________________________________________________________\nlambda_4 (Lambda)            (None, 320, 480, 64)      0         \n_________________________________________________________________\nconv2d_8 (Conv2D)            (None, 320, 480, 32)      18464     \n_________________________________________________________________\nbatch_normalization_8 (Batch (None, 320, 480, 32)      128       \n_________________________________________________________________\nactivation_8 (Activation)    (None, 320, 480, 32)      0         \n_________________________________________________________________\nconv2d_9 (Conv2D)            (None, 320, 480, 32)      9248      \n_________________________________________________________________\nbatch_normalization_9 (Batch (None, 320, 480, 32)      128       \n_________________________________________________________________\nactivation_9 (Activation)    (None, 320, 480, 32)      0         \n_________________________________________________________________\nconv2d_10 (Conv2D)           (None, 320, 480, 1)       33        \n_________________________________________________________________\nbatch_normalization_10 (Batc (None, 320, 480, 1)       4         \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 153600)            0         \n_________________________________________________________________\nactivation_10 (Activation)   (None, 153600)            0         \n=================================================================\nTotal params: 205,893\nTrainable params: 204,995\nNon-trainable params: 898\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>However when trying to fit the model <\/p>\n\n<pre><code>history = model.fit(x=x_train, y=y_train, batch_size=batch_size, epochs=3, verbose=2, validation_data=(x_val,y_val))\n<\/code><\/pre>\n\n<p>I encounter this error:<\/p>\n\n<pre><code>InvalidArgumentError: Input to reshape is a tensor with 4915200 values, but the requested shape has 9830400\n     [[{{node lambda_4\/unpool\/Reshape_3}} = Reshape[T=DT_INT64, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](lambda_1\/MaxPoolWithArgmax:1, lambda_4\/unpool\/Reshape_2\/shape)]]\n     [[{{node lambda_4\/unpool\/strided_slice_6\/_515}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_1479_lambda_4\/unpool\/strided_slice_6\", tensor_type=DT_INT64, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>I have looked over all the shapes after each layers and they are what I expect. I also tested out the pooling\/unpooling functions on sample tensors and they produced expected output. What am I doing wrong here?<\/p>\n\n<p>I've been pulling my hair out trying to solve this, any help is greatly appreciated!<\/p>\n","answers":[{"AnswerId":"53462075","CreationDate":"2018-11-24T20:26:27.950","ParentId":null,"OwnerUserId":"10679504","Title":null,"Body":"<p>Found the problem, mask_1 has 32 channels while unpool_2 is trying to reshape output with 64 channels. I just re-arranged stuff so the depths lined up.<\/p>\n"}]}
+{"QuestionId":53392807,"AnswerCount":2,"Tags":"<tensorflow><object-detection><object-detection-api><satellite-image>","CreationDate":"2018-11-20T12:16:15.353","AcceptedAnswerId":null,"OwnerUserId":8596375.0,"Title":"Meaning of Tensorflow Object Detection API image_additional_channels","Body":"<p>I would like to use the Tensorflow Object Detection API for multi-channel images (e.g., 4-channels RGB + infrared). There is a <a href=\"https:\/\/github.com\/minhnhat93\/tf_object_detection_multi_channels\" rel=\"nofollow noreferrer\">tutorial<\/a> how to change the API to add additional channels. However, the tutorial was written a year ago and the API has evolved since then and it seems that the API may accept multi-channel images now. <\/p>\n\n<p>For example, in tensorflow-models\/research\/object-detection\/data_decoders\/tf_example_decoder.py in addition to fields.InputDataFields.image there is now <strong>fields.InputDataFields.image_additional_channels<\/strong>. Can it be used for any additional channels one has in an input image beyond standard 3 channels fed into fields.InputDataFields.image? I cannot figure out the purpose of this image_additional_channels and how to use it.<\/p>\n\n<p>More general, my question is how to use the Tensorflow Object Detection API for multi-channel (>3) images. Are they accepted, i.e. taken into account, by default? I can feed them to train a model, but for inference in object_detection_tutorial notebook it cannot accept more than 3 channels, which makes me wonder whether it ignores the 4th channels during training. <\/p>\n\n<p>I am using Tensorflow 1.12.0, latest commit (7a75bfc) of the Object Detection API. image_additional_channels was added in commit 9fce9c6 on 6 June 2018<\/p>\n","answers":[{"AnswerId":"53725704","CreationDate":"2018-12-11T14:00:23.347","ParentId":null,"OwnerUserId":"1351741","Title":null,"Body":"<p>I'm trying do the same thing. It seems to accept additional channels during training (you need to add them during the creation of your TfExample file(s)). You also need to set <code>num_additional_channels<\/code> in the <code>train_input_reader<\/code> portion of the pipeline config file to be the number of channels you've added.<\/p>\n\n<p>However, the script for exporting the model for inference does not seem to support exporting the model in a way that allows it to accept additional channels.<\/p>\n\n<p>As you can see here: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/exporter.py#L129\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/exporter.py#L129<\/a><\/p>\n\n<p>The input tensor is only a standard image tensor and the <code>tensor_dict[fields.InputDataFields.image_additional_channels]<\/code> is not included in the input.<\/p>\n\n<p>I'm about to fix this for my project, so I'll try to open a pull request and get them to merge it in.<\/p>\n"},{"AnswerId":"57115277","CreationDate":"2019-07-19T15:11:20.513","ParentId":null,"OwnerUserId":"11792411","Title":null,"Body":"<p>For the TFRecord creation you must edit the example here:<\/p>\n\n<pre><code>def create_tf_example(group, path):\nwith tf.gfile.GFile(os.path.join(path, '{}'.format(group.filename)), 'rb') as fid:\n    encoded_jpg = fid.read()\nencoded_jpg_io = io.BytesIO(encoded_jpg)\nimage = Image.open(encoded_jpg_io)\nwidth, height = image.size\n<\/code><\/pre>\n\n<p>This portion of code loads and opens an image file path; if you have multiple images you have to load them in different variables. For example, I use something like this:<\/p>\n\n<pre><code>with tf.gfile.GFile(dictionary[0], 'rb') as fid:\n    encoded_jpg = fid.read()\nif ndata &gt; 1:\n    with tf.gfile.GFile(dictionary[1], 'rb') as fid:\n        encoded_depth = fid.read()\n    encoded_inputs = encoded_depth\n<\/code><\/pre>\n\n<p>where <code>dictionary[0]<\/code> contains the path of the rgb image and <code>dictionary[1]<\/code> contains the path of the depth image.<\/p>\n\n<p>Then, the TFRecord must be created like this:<\/p>\n\n<pre><code>tf_example = tf.train.Example(features=tf.train.Features(feature={\n    'image\/height': dataset_util.int64_feature(height),\n    'image\/width': dataset_util.int64_feature(width),\n    'image\/filename': dataset_util.bytes_feature(filename),\n    'image\/source_id': dataset_util.bytes_feature(filename),\n    'image\/encoded': dataset_util.bytes_feature(encoded_jpg),\n    'image\/additional_channels\/encoded': dataset_util.bytes_feature(encoded_inputs),\n    'image\/format': dataset_util.bytes_feature(image_format),\n    'image\/object\/bbox\/xmin': dataset_util.float_list_feature(xmins),\n    'image\/object\/bbox\/xmax': dataset_util.float_list_feature(xmaxs),\n    'image\/object\/bbox\/ymin': dataset_util.float_list_feature(ymins),\n    'image\/object\/bbox\/ymax': dataset_util.float_list_feature(ymaxs),\n    'image\/object\/class\/text': dataset_util.bytes_list_feature(classes_text),\n    'image\/object\/class\/label': dataset_util.int64_list_feature(classes),\n}))\n<\/code><\/pre>\n\n<blockquote>\n  <p>However, the script for exporting the model for inference does not\n  seem to support exporting the model in a way that allows it to accept\n  additional channels.<\/p>\n  \n  <p>As you can see here:\n  <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/exporter.py#L129\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/exporter.py#L129<\/a><\/p>\n  \n  <p>The input tensor is only a standard image tensor and the\n  tensor_dict[fields.InputDataFields.image_additional_channels] is not\n  included in the input.<\/p>\n  \n  <p>I'm about to fix this for my project, so I'll try to open a pull\n  request and get them to merge it in.<\/p>\n<\/blockquote>\n\n<p>I would like to know that too! I managed to train without problems but I can't use the trained model since the additional channel cannot be loaded... Did you fixed that?<\/p>\n"}]}
+{"QuestionId":53393311,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2018-11-20T12:46:30.540","AcceptedAnswerId":null,"OwnerUserId":5617507.0,"Title":"Why the training of a neural network using binary cross-entropy loss function gets stuck when we use real-valued training targets?","Body":"<p>Assume that we have a binary classification problem, in which the training targets are not in {0,1} but in [0,1]. We use the following code to train a simple classifier in Keras:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(100, input_shape=(X.shape[1],), activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='rmsprop')\nmodel.fit(X,y)\n<\/code><\/pre>\n\n<p>If we pass the real training targets (in [0,1]), the training hardly proceeds, getting stuck around its initial loss value; but if we quantize the targets in {0,1} it performs better, rapidly decreases the training loss.<\/p>\n\n<p>Is this a normal phenomena? What is its reason? <\/p>\n\n<p>Edit: <a href=\"https:\/\/github.com\/h-amirkhani\/logistic-regression-temp\" rel=\"nofollow noreferrer\">Here<\/a> is the reproducible experiment.\nAnd this is the obtained plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SN28o.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SN28o.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53393871","CreationDate":"2018-11-20T13:17:25.383","ParentId":null,"OwnerUserId":"5095259","Title":null,"Body":"<p>You state that you want to solve a binary classification task, for which the target should be <strong>binary<\/strong>-valued, i.e. {0,1}.<\/p>\n\n<p>However, if your target instead is some float value in [0,1], you are actually trying to perform <strong>regression<\/strong>. <\/p>\n\n<p>This, amongst others, changes the requirements for your loss function.\nSee <a href=\"https:\/\/stats.stackexchange.com\/questions\/223256\/tensorflow-cross-entropy-for-regression\">Tensorflow Cross Entropy for Regression?<\/a>, where the usage of cross entropy loss for regression is discussed in more detail. <\/p>\n"}]}
+{"QuestionId":53393430,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-20T12:52:42.137","AcceptedAnswerId":53394212.0,"OwnerUserId":2213201.0,"Title":"Tensorflow, where (index) 'and' conditional","Body":"<p>I have the following Tensor:<\/p>\n\n<pre><code># (class, index)\nobj_class_indexes = tf.constant([(0, 0), (0, 1), (0, 2), (1, 3)])\n<\/code><\/pre>\n\n<p>And to each value I'm looking for the objects with the same class.\nFor now I'm trying the following:<\/p>\n\n<pre><code>same_classes = tf.logical_and(tf.equal(obj_classes_indexes[:, 0], obj_classes_indexes[0][0]), \\\n                                        obj_classes_indexes[:, 1]  &gt; obj_classes_indexes[0][1])\nfound_indexes = tf.where(same_classes)\n\nwith tf.Session() as sess:\n    print(sess.run(same_classes))\n    print(sess.run(indexes))\n<\/code><\/pre>\n\n<p>The expected output would be:<\/p>\n\n<pre><code>[False  True  True False]\n[1, 2]\n<\/code><\/pre>\n\n<p>But it's giving me:<\/p>\n\n<pre><code>[False  True  True False]\n[[1], [2]]\n<\/code><\/pre>\n\n<p>I don't think the <code>logical_and<\/code> output is actually the correct input to the <code>tf.where<\/code> function. Or Am I missing something?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53394212","CreationDate":"2018-11-20T13:34:47.870","ParentId":null,"OwnerUserId":"10658208","Title":null,"Body":"<p>There is nothin wrong with the output. tf.where() is expected to output a 2D tensor as quoted <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/where\" rel=\"nofollow noreferrer\">here<\/a>:\n\"The coordinates are returned in a 2-D tensor where the first dimension (rows) represents the number of true elements, and the second dimension (columns) represents the coordinates of the true elements\"<\/p>\n\n<p>If you want the output to be a 1D tensor as you have mentioned, you could just add a reshape op in your case as below:<\/p>\n\n<pre><code>found_indexes = tf.where(same_classes)\nfound_indexes = tf.reshape(found_indexes, [-1])\n<\/code><\/pre>\n\n<p>hope this helps!<\/p>\n"}]}
+{"QuestionId":53394012,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><linear-regression>","CreationDate":"2018-11-20T13:25:22.470","AcceptedAnswerId":53407968.0,"OwnerUserId":6138789.0,"Title":"Limitation of Keras\/Tensorflow for solving Linear Regression tasks","Body":"<p>I was trying to implement linear regression in Keras\/TensorFlow and was very surprised how difficult it is. The standard examples work great on random data. However, if we change the input data a little bit, all examples stop work correctly.<\/p>\n\n<p>I try to find coefficients for <code>y = 0.5 * x1 + 0.5 * x2<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>np.random.seed(1443)\nn = 100000\nx = np.zeros((n, 2))\ny = np.zeros((n, 1))\n\nx[:,0] =  sorted(preprocessing.scale( np.random.poisson(1000000, (n)) ))\nx[:,1] =  sorted(preprocessing.scale( np.random.poisson(1000000, (n)) ) )\ny = (x[:,0] + x[:,1]) \/2\n\nmodel = keras.Sequential()\nmodel.add( keras.layers.Dense(1, input_shape =(2,), dtype=\"float32\" ))\nmodel.compile(loss='mean_squared_error', optimizer='sgd')\n\nmodel.fit(x,y, epochs=1000, batch_size=64)\nprint(model.get_weights())\n<\/code><\/pre>\n\n<p>The results:<\/p>\n\n<pre><code>| epochs| batch_size |  bias      | x1         | x2\n| ------+------------+------------+------------+-----------\n| 1000  | 64         | -5.83E-05  | 0.90410435 | 0.09594361\n| 1000  | 1024       | -5.71E-06  | 0.98739249 | 0.01258729\n| 1000  | 10000      | -3.07E-07  | -0.2441376 | 1.2441349\n<\/code><\/pre>\n\n<p>My first thought was that it is a bug in Keras. So, I tried R\/Tensorflow library:<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>floatType &lt;- \"float32\"\np &lt;- 2L\nX &lt;- tf$placeholder(floatType, shape = shape(NULL, p), name = \"x-data\")\nY &lt;- tf$placeholder(floatType, name = \"y-data\")\nW &lt;- tf$Variable(tf$zeros(list(p, 1L), dtype=floatType))\nb &lt;- tf$Variable(tf$zeros(list(1L), dtype=floatType))\nY_hat &lt;- tf$add(tf$matmul(X, W), b)\ncost &lt;- tf$reduce_mean(tf$square(Y_hat - Y))\ngenerator &lt;- tf$train$GradientDescentOptimizer(learning_rate=0.01)\noptimizer &lt;- generator$minimize(cost)\n\nsession &lt;- tf$Session()\nsession$run(tf$global_variables_initializer())\n\nset.seed(1443)\nn &lt;- 10^5\nx &lt;- matrix( replicate(p, sort(scale((rpois(n, 10^6))))) , nrow = n )\ny &lt;- matrix((x[,1]+x[,2])\/2)\n\ni &lt;- 1\nbatch_size &lt;- 10000\nepoch_number  &lt;- 1000\niterationNumber &lt;- n*epoch_number \/ batch_size\n\nwhile (iterationNumber &gt; 0) {\n  feed_dict &lt;- dict(X = x[i:(i+batch_size-1),, drop = F], Y = y[i:(i+batch_size-1),, drop = F])\n  session$run(optimizer, feed_dict = feed_dict)\n\n  i &lt;- i+batch_size\n  if( i &gt; n-batch_size)\n    i &lt;- i %% batch_size \n\n  iterationNumber &lt;- iterationNumber - 1\n}\nr_model &lt;- lm(y ~ x)\ntf_coef &lt;- c(session$run(b), session$run(W))\nr_coef  &lt;- r_model$coefficients\nprint(rbind(tf_coef, r_coef))\n<\/code><\/pre>\n\n<p>The results:<\/p>\n\n<pre><code>| epochs| batch_size |  bias      | x1         | x2\n| ------+------------+------------+------------+-----------\n|2000   | 64         | -1.33E-06  | 0.500307   | 0.4996932\n|1000   | 1000       | 2.79E-08   | 0.5000809  | 0.499919\n|1000   | 10000      | -4.33E-07  | 0.5004921  | 0.499507\n|1000   | 100000     | 2.96E-18   | 0.5        | 0.5   \n<\/code><\/pre>\n\n<p>Tensorflow finds the correct result only when batch size = samples number and the optimization algorithm is SGD. If optimization algorithm was \"adam\" or \"adagrad\", errors were much larger.<\/p>\n\n<ol>\n<li>For obvious reasons, I cannot choose hyperparameter <code>batch_size = n<\/code>. Could you recommend any approaches to solve this problem with precision 1E-07 for Keras or TensorFlow?<\/li>\n<li>Why TensorFlow finds better solutions than Keras?<\/li>\n<\/ol>\n\n<p><strong><em>Comment 1.<\/em><\/strong> \nBased on post \"today\" below:\nTrain dataset shuffling will significantly improve the performance of TensorFlow version:\n<\/p>\n\n<pre><code>shuffledIndex&lt;-sample(1:(nrow(x)))\nx &lt;- x[shuffledIndex,]\ny &lt;- y[shuffledIndex,,drop=FALSE]\n<\/code><\/pre>\n\n<p>For batch size = 2000:<\/p>\n\n<pre><code>|(Intercept)     |       x1  |        x2\n|----------------+-----------+----------\n|-1.130693e-09   | 0.5000004 | 0.4999989\n<\/code><\/pre>\n","answers":[{"AnswerId":"53407968","CreationDate":"2018-11-21T08:31:35.540","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are sorting the generated random numbers for each feature value. So they end up very close to each other:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; np.mean(np.abs(x[:,0]-x[:,1]))\n0.004125721684553685\n<\/code><\/pre>\n\n<p>As a result we would have:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y = (x1 + x2) \/ 2\n ~= (x1 + x1) \/ 2\n  = x1\n  = 0.5 * x1 + 0.5 * x1\n  = 0.3 * x1 + 0.7 * x1\n  = -0.3 * x1 + 1.3 * x1\n  = 10.1 * x1 - 9.1 * x1\n  = thousands of other possible combinations\n<\/code><\/pre>\n\n<p>In this case the solution that Keras would converge to would really depend on the initial value of the weights and bias of Dense layer. With different initial values you would get different results (and possibly for some of them, it may not converge at all):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># set the initial weight of Dense layer\nmodel.layers[0].set_weights([np.array([[0], [1]]), np.array([0])])\n\n# fit the model ...\n\n# the final weights\nmodel.get_weights()\n\n[array([[0.00203656],\n        [0.9981099 ]], dtype=float32),\n array([4.5520876e-05], dtype=float32)]    # because: y = 0 * x1 + 1 * x1 = x1 ~= (x1 + x2) \/ 2\n\n# again set the weights to something different\nmodel.layers[0].set_weights([np.array([[0], [0]]), np.array([1])])\n\n# fit the model...\n\n# the final weights\nmodel.get_weights()\n\n[array([[0.49986306],\n       [0.50013727]], dtype=float32),\n array([1.4176634e-08], dtype=float32)]    # the one you were looking for!\n<\/code><\/pre>\n\n<p>However, if you don't sort the features (i.e. just remove <code>sorted<\/code>) it is very likely that the converged weights to be very close to <code>[0.5, 0.5]<\/code>.<\/p>\n"}]}
+{"QuestionId":53394015,"AnswerCount":0,"Tags":"<python><tensorflow><memory><keras>","CreationDate":"2018-11-20T13:25:25.650","AcceptedAnswerId":null,"OwnerUserId":4847250.0,"Title":"How to deal with large layers in Keras","Body":"<p>I'm trying to create a deep learning model with Keras, but when I'm trying make the model, keras uses too much memory.\nI'm trying to make a model with a large amount of inputs (24,000) and it seems too much for my computer.<\/p>\n\n<p>Here is my simple code:<\/p>\n\n<pre><code>N=20\nM=5\nn = 24310\nnbclass = 4 \n\nX_train = np.random.rand(N,n)\nX_test = np.random.rand(M,n)\nY_train = np_utils.to_categorical(np.random.randint(nbclass, size=N), nbclass)\nY_test = np_utils.to_categorical(np.random.randint(nbclass, size=M), nbclass)\n\n\nmodel.add(Dense(n, input_dim=n, activation=\"relu\", kernel_initializer=\"uniform\"))\nmodel.add(Dropout(0.25)) \nmodel.add(Dense(n\/\/2,  activation=\"relu\", kernel_initializer=\"uniform\"))\nmodel.add(Dropout(0.25))  \nmodel.add(Dense(nbclass,activation=\"softmax\", kernel_initializer=\"uniform\"))\n\n# Compiling the model \nmodel.compile(loss='binary_crossentropy',\n              optimizer='adam', metrics=['accuracy'])\n# Fitting the model \nmodel.fit(X_train, Y_train, validation_data=(X_test, Y_test),epochs=100, batch_size=5,callbacks=[PlotLosses()])\n<\/code><\/pre>\n\n<p>My issue here is, when the fit is done, it consumes all my memory (20 Go), then it starting to swap and I cannot go further.\nSomebody knows how can I modify the model to make it less memory consuming? <\/p>\n","answers":[]}
+{"QuestionId":53394499,"AnswerCount":2,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2018-11-20T13:51:24.827","AcceptedAnswerId":null,"OwnerUserId":9898479.0,"Title":"Error using Estimator export_saved_model (Not found: Key global_step not found in checkpoint)","Body":"<p>I am trying to prepare a pre-trained model for google cloud ML. I am trying to use an estimator to export the model. During the loading of the checkpoints by the estimator I get the following error:<\/p>\n\n<pre><code>2018-11-19 13:28:57.526564: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at save_restore_v2_ops.cc:184 : Not found: Key global_step not found in checkpoint\nTraceback (most recent call last):\n  File \"C:\\Users\\Administrator\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1334, in _do_call\n    return fn(*args)\n  File \"C:\\Users\\Administrator\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1319, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"C:\\Users\\Administrator\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1407, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.NotFoundError: Key global_step not found in checkpoint\n         [[{{node save_1\/RestoreV2}} = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](_arg_save_1\/Const_0_0, save_1\/RestoreV2\/tensor_names, save_1\/RestoreV2\/shape_and_slices)]]\n<\/code><\/pre>\n\n<ul>\n<li>TensorFlow version (use command below): 1.12.0<\/li>\n<li>Python version: 3.6<\/li>\n<\/ul>\n\n<p>Here is the code I am using:<\/p>\n\n<pre><code>MODEL_DIR='model\/'\ndef decode_image(image_bytes):\n    image = tf.image.decode_image(image_bytes)\n    image = tf.cast(image, dtype=tf.uint8)\n    return image\n\ndef serving_input_fn():\n    createmodel()\n    inputs = {'image_bytes': tf.placeholder(tf.string, shape=(), name=\"image_bytes\")}\n    imagebytes = tf.squeeze(inputs['image_bytes']) # make it a scalar\n    image = decode_image(imagebytes)\n    # make the outer dimension unknown (and not 1)\n    image = tf.placeholder_with_default(image, shape=[None, None, None, 3])\n\n    features = {'image_bytes' : image}\n    return tf.estimator.export.ServingInputReceiver(features, inputs)\n\ndef model_fn(features, labels, mode, params):\n    pred = tf.get_default_graph().get_tensor_by_name(\"fc1_voc12:0\")\n    return tf.estimator.EstimatorSpec(\n        mode=tf.estimator.ModeKeys.PREDICT,\n        predictions=pred,\n        export_outputs={'pred':tf.estimator.export.PredictOutput(pred)}\n        )\n\nestimator = tf.estimator.Estimator(\n    model_fn=model_fn,\n    model_dir=MODEL_DIR)\n\nestimator.export_savedmodel('deployment_gcp_1', serving_input_fn, strip_default_attrs=True)\n<\/code><\/pre>\n\n<p>I have searched this issue quite a bit. There was one bug report for an older version of tensorflow (I think 1.2.0 but I am not sure now). I am able to load and save this model using tf.saved_model.simple_save, and it works when I run predictions on it.<\/p>\n\n<p>I am not sure if this is a bug or if I am missing something really simple. I have posted the same thing on the tensorflow github repo but no responses there yet.<\/p>\n","answers":[{"AnswerId":"53486860","CreationDate":"2018-11-26T18:15:39.117","ParentId":null,"OwnerUserId":"9898479","Title":null,"Body":"<p>I have finally managed to get it to work by adding the \"global_step\" variable manually, exporting that as a new checkpoint and loading that instead.\nAfter loading the current checkpoint I ran the following code:<\/p>\n\n<pre><code>b = tf.Variable(load_step, name=\"global_step\", dtype=tf.int64)\nsess.run(b.initializer)\nsaver = tf.train.Saver()\nsaver.save(sess,'UpdatedModel\/model.ckpt', global_step=load_step)\n<\/code><\/pre>\n\n<p>Then in my previous code I used the new folder MODEL_DIR and it worked.<\/p>\n"},{"AnswerId":"53395626","CreationDate":"2018-11-20T14:51:48.377","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>In your code, where you minimize the loss using any optimizer,<\/p>\n\n<pre><code>train_op = optimizer.minimize( loss , global_step=tf.train.get_global_step())\n<\/code><\/pre>\n\n<p>You might not have mentioned the <code>global_step=<\/code> due to which you get the error.<\/p>\n"}]}
+{"QuestionId":53394824,"AnswerCount":1,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2018-11-20T14:07:14.170","AcceptedAnswerId":53395506.0,"OwnerUserId":6596737.0,"Title":"What's the difference between Android TensorFlow support and TensorFlow Lite for Android\uff1f","Body":"<p>I see a sample in Google codelabs <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-style-transfer-android\/index.html?index=..%2F..index#0\" rel=\"nofollow noreferrer\">this<\/a><\/p>\n\n<p>it requirements dependencies Android TensorFlow support<\/p>\n\n<pre><code>dependencies {\nimplementation 'org.tensorflow:tensorflow-android:1.2.0-preview'\n}\n<\/code><\/pre>\n\n<p>I know the TensorFlow Lite was help developer to use the model in mobile \ndevices<\/p>\n\n<p>What's the difference between these two<\/p>\n","answers":[{"AnswerId":"53395506","CreationDate":"2018-11-20T14:44:56.997","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>The code snippet which you provided corresponds to TensorFlow Mobile.<\/p>\n\n<ul>\n<li><p>TensorFlow Mobile is a program useful for running protocol buffers ( .pb ) files on Android , iOS and other IoT stuff. It can only be used to run inferences on a TensorFlow model which is converted to a .pb file. It can only function over specific platforms.<\/p><\/li>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/lite\" rel=\"nofollow noreferrer\">TensorFlow Lite<\/a> is a successor of TensorFlow Mobile. Lite can run inferences on models which are converted to a .tflite file. The Lite version also allows the developer to run Graphs, Sessions and Tensors over Java and Android. It also provides the Neural Networks API. It can functions over Android and iOS devices, <a href=\"https:\/\/firebase.google.com\/mlkit\" rel=\"nofollow noreferrer\">Firebase MLKit<\/a>, <a href=\"https:\/\/js.tensorflow.org\/\" rel=\"nofollow noreferrer\">TensorFlow.js<\/a> and also TensorFlow C++ APIs.<\/p><\/li>\n<\/ul>\n\n<p>Even Google recommends to use TensorFlow Lite instead of TensorFlow Mobile. <\/p>\n"}]}
+{"QuestionId":53395126,"AnswerCount":1,"Tags":"<python><keras><classification><predict>","CreationDate":"2018-11-20T14:24:42.037","AcceptedAnswerId":null,"OwnerUserId":10212759.0,"Title":"How to predict an image among 10 classes instead of two classes","Body":"<p>I am new to deep learning and I am trying to predict an image from 10 classes.<\/p>\n\n<p>I found the code below but only considers two classes. How can I adjust the code to 10 classes?<\/p>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing import image\ntest_image = image.load_img(path = 'dataset\/single_prediction\/cat_or_dog_1.jpg', \n                            target_size = (64, 64))\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis = 0)\n\nresult = classifier.predict(test_image)\nindices = training_set.class_indices\nif result[0][0] == 1:\n    prediction = 'dog'\nelse:\n    prediction = 'cat'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53395869","CreationDate":"2018-11-20T15:04:04.450","ParentId":null,"OwnerUserId":"10642035","Title":null,"Body":"<p>Omar, my suggestion is for you to study a little further the libs you're using, such as Keras (mainly) and Numpy. <\/p>\n\n<p>This kind of question is going to be downvoted from users, because goes against the best practices from Stack Overflow, since you're asking for someone to give you the full code as an answer without looking out by yourself and detailing the specific challenge you're facing with which class, and not detailing the whole aspect of your code as well (such as the classifier object, which is not detailed where it came from, and guessing is not a good method to help here).<\/p>\n\n<p>Now, jumping to your problem, I suggest:<\/p>\n\n<ol>\n<li>Understand better the model from keras and its implications, on how to load models: <a href=\"https:\/\/keras.io\/models\/about-keras-models\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/about-keras-models\/<\/a> <\/li>\n<li>Take a closer look on keras documentation and its methods\n<a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a><\/li>\n<li>Take a look on this answer: <a href=\"https:\/\/stackoverflow.com\/a\/43470074\/10642035\">https:\/\/stackoverflow.com\/a\/43470074\/10642035<\/a><\/li>\n<\/ol>\n\n<p>Take a look on the documentation from stack overflow, to ensure you won't be losing points here:<\/p>\n\n<ul>\n<li>Topics to as about: <a href=\"https:\/\/stackoverflow.com\/help\/on-topic\">https:\/\/stackoverflow.com\/help\/on-topic<\/a><\/li>\n<li>How to write a good question: <a href=\"https:\/\/stackoverflow.com\/help\/how-to-ask\">https:\/\/stackoverflow.com\/help\/how-to-ask<\/a><\/li>\n<li>What to avoid asking: <a href=\"https:\/\/stackoverflow.com\/help\/dont-ask\">https:\/\/stackoverflow.com\/help\/dont-ask<\/a><\/li>\n<li>In case you decide to delete the question: <a href=\"https:\/\/stackoverflow.com\/help\/what-to-do-instead-of-deleting-question\">https:\/\/stackoverflow.com\/help\/what-to-do-instead-of-deleting-question<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53395284,"AnswerCount":1,"Tags":"<python><keras><classification><regression><loss-function>","CreationDate":"2018-11-20T14:32:35.747","AcceptedAnswerId":null,"OwnerUserId":5617507.0,"Title":"Custom loss function in Keras for weighting missclassified samples","Body":"<p>Assume that <code>y_true<\/code> and <code>y_pred<\/code> are in [-1,1]. I want a weighted mean-square-error loss function, in which the loss for samples that are positive in the <code>y_true<\/code> and negative in <code>y_pred<\/code> or vice versa are weighted by <code>exp(alpha)<\/code>. Here is my code:<\/p>\n\n<pre><code>import keras.backend as K\nalpha = 1.0\ndef custom_loss(y_true, y_pred):\n     se = K.square(y_pred-y_true)\n     true_label = K.less_equal(y_true,0.0)\n     pred_label = K.less_equal(y_pred,0.0)\n     return K.mean(se * K.exp(alpha*K.cast(K.not_equal(true_label,pred_label), tf.float32)))\n<\/code><\/pre>\n\n<p>And here is a plot of this loss function. Different curves are for different values for <code>y_true<\/code>.\n<a href=\"https:\/\/i.stack.imgur.com\/i03cE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/i03cE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I want to know:<\/p>\n\n<ul>\n<li>Whether this is a valid loss function, since it is not differentiable in 0?<\/li>\n<li>Is my code correct?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"55727908","CreationDate":"2019-04-17T12:48:11.073","ParentId":null,"OwnerUserId":"10375049","Title":null,"Body":"<p>I suggest you this type of loss function to handle imbalance dataset<\/p>\n\n<pre><code>def focal_loss(y_true, y_pred):\n   gamma = 2.0, alpha = 0.25\n   pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))\n   pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))\n   return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1))-K.sum((1-alpha) * K.pow(pt_0, gamma) * K.log(1. - pt_0))\n<\/code><\/pre>\n\n<p>from <a href=\"https:\/\/towardsdatascience.com\/handling-imbalanced-datasets-in-deep-learning-f48407a0e758\" rel=\"nofollow noreferrer\">this<\/a> source<\/p>\n"}]}
+{"QuestionId":53395329,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><conv-neural-network>","CreationDate":"2018-11-20T14:34:56.480","AcceptedAnswerId":53396373.0,"OwnerUserId":2628369.0,"Title":"Should Dropout masks be reused during Adversarial Training?","Body":"<p>I am implementing adversarial training with the FGSM method from <a href=\"https:\/\/arxiv.org\/abs\/1412.6572\" rel=\"nofollow noreferrer\">Explaining and Harnessing Adversarial Examples<\/a> using the custom loss function:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RG7eB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RG7eB.png\" alt=\"\"><\/a><\/p>\n\n<p>Implemented in <code>tf.keras<\/code> using a custom loss function, it conceptually looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential([\n    ...\n])\n\ndef loss(labels, logits):\n    # Compute the cross-entropy on the legitimate examples\n    cross_ent = tf.losses.softmax_cross_entropy(labels, logits)\n\n    # Compute the adversarial examples\n    gradients, = tf.gradients(cross_ent, model.input)\n    inputs_adv = tf.stop_gradient(model.input + 0.3 * tf.sign(gradients))\n\n    # Compute the cross-entropy on the adversarial examples\n    logits_adv = model(inputs_adv)\n    cross_ent_adv = tf.losses.softmax_cross_entropy(labels, logits_adv)\n\n    return 0.5 * cross_ent + 0.5 * cross_ent_adv\n\nmodel.compile(optimizer='adam', loss=loss)\nmodel.fit(x_train, y_train, ...)\n<\/code><\/pre>\n\n<p>This works well for a simple convolutional neural network. <\/p>\n\n<p>During the <code>logits_adv = model(inputs_adv)<\/code> call, the model is called for the second time. This means, that it will use different dropout masks than in the original feed-forward pass with <code>model.inputs<\/code>. The <code>inputs_adv<\/code>, however, were created with <code>tf.gradients(cross_ent, model.input)<\/code>, i.e. with the dropout masks from the original feed-forward pass. This could be problematic, as allowing the model to use new dropout masks will likely dampen the effect of the adversarial batch.<\/p>\n\n<p>Since implementing the reusing of dropout masks in Keras would be cumbersome, I am interested in the actual effect of reusing the masks. Does it make a difference w.r.t. the test accuracy on both legitimate and adversarial examples?<\/p>\n","answers":[{"AnswerId":"53396373","CreationDate":"2018-11-20T15:31:39.093","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>I tried out reusing the dropout masks during the adversarial training step's feed-forward pass with a simple CNN on MNIST. I chose the same network architecture as the one used in this <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans_tutorials\/mnist_tutorial_keras_tf.py\" rel=\"nofollow noreferrer\">cleverhans tutorial<\/a> with an additional dropout layer before the softmax layer. <\/p>\n\n<p>This is the result <em>(red = reuse dropout masks, blue = naive implementation)<\/em>:\n<a href=\"https:\/\/i.stack.imgur.com\/o9tjn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/o9tjn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The solid lines represent the accuracy on legitimate test examples. The dotted lines represent the accuracy on adversarial examples generated on the test set.<\/p>\n\n<p><strong>In conclusion<\/strong>, if you only use adversarial training as a regularizer in order to improve the test accuracy itself, reusing dropout masks might not be worth the effort. For the robustness against adversarial attacks, it <strong>might<\/strong> make a  small difference. However, you would need to run further experiments on other datasets, architectures, random seeds etc. to make a more confident statement.<\/p>\n\n<p>To keep the figure above readable, I omitted the accuracy on adversarial test examples for the model trained without adversarial training. The values lay around 10%.<\/p>\n\n<p>You can find the code for this experiment in <a href=\"https:\/\/gist.github.com\/batzner\/aaed9cdb4cd1d05f85a8a88e67c57aac\" rel=\"nofollow noreferrer\">this gist<\/a>. With TensorFlow's eager mode, it was rather straightforward to implement storing and reusing the dropout masks.<\/p>\n"}]}
+{"QuestionId":53395427,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-20T14:39:39.267","AcceptedAnswerId":53402721.0,"OwnerUserId":9008988.0,"Title":"Keras: What is model.inputs in VGG16","Body":"<p>I start playing with keras and vgg16 recently, and I am using keras.applications.vgg16. <\/p>\n\n<p>But here I come with a question about what is <code>model.inputs<\/code> because I saw others using it in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/conv_filter_visualization.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/conv_filter_visualization.py<\/a> although it does not initialize it<\/p>\n\n<pre><code>    ...\n    input_img = model.input\n    ...\n    layer_output = layer_dict[layer_name].output\n    if K.image_data_format() == 'channels_first':\n        loss = K.mean(layer_output[:, filter_index, :, :])\n    else:\n        loss = K.mean(layer_output[:, :, :, filter_index])\n\n    # we compute the gradient of the input picture wrt this loss\n    grads = K.gradients(loss, input_img)[0]\n<\/code><\/pre>\n\n<p>I checked the keras site but it only said that is an input tensor with shape (1,224,224,3) But I still don't understand what is that exactly. Is that an image from ImageNet?Or a default image provided by keras for keras model?<\/p>\n\n<p>I am sorry if I don't have enough understanding of deep learning, but can someone explain it to me please. Thanks<\/p>\n","answers":[{"AnswerId":"53402721","CreationDate":"2018-11-20T22:45:59.270","ParentId":null,"OwnerUserId":"3254400","Title":null,"Body":"<p>The 4 dimensions of <code>(1,224,224,3)<\/code> are the <code>batch_size<\/code>, <code>image_width<\/code>, <code>image_height<\/code> and <code>image_channels<\/code> respectively. <code>(1,224,224,3)<\/code> means that the <code>VGG16<\/code> model accepts a batch size of <code>1<\/code> (one image at a time) of shape <code>224x224<\/code> and three channels (RGB).<\/p>\n\n<p>For more information on what a <code>batch<\/code> and therefore a <code>batch size<\/code> is, you can check <a href=\"https:\/\/stats.stackexchange.com\/questions\/153531\/what-is-batch-size-in-neural-network\">this<\/a> Cross Validated question.<\/p>\n\n<p>Returning to <code>VGG16<\/code>, the input of the architecture is <code>(1, 224, 224, 3)<\/code>. What does this mean? That in order to input a image into the network, you will need to:<\/p>\n\n<ol>\n<li>Preprocess it to reach a shape of (224, 224) and 3 channels (RGB)<\/li>\n<li>Convert this to an actual matrix of shape (224, 224, 3)<\/li>\n<li>Group together various images in a batch of the size that requires the network (in this case, the batch size is 1, but you need to add a dimension to the matrix, in order to obtain the (1, 224, 224, 3)<\/li>\n<\/ol>\n\n<p>After doing this, you can input the image to the model.<\/p>\n\n<p>Keras offers few utilitary functions to do these tasks. Below I present a modified version of the code snippet shown in <em>Extract features with VGG16<\/em> from <a href=\"https:\/\/keras.io\/applications\/#usage-examples-for-image-classification-models\" rel=\"nofollow noreferrer\">Usage examples for image classification models<\/a> in the documentation.<\/p>\n\n<p>In order to have it actually working, you need a <code>jpg<\/code> of any size named <code>elephant.jpg<\/code>. You can obtain it running this bash command:<\/p>\n\n<pre><code>wget https:\/\/upload.wikimedia.org\/wikipedia\/commons\/f\/f9\/Zoorashia_elephant.jpg -O elephant.jpg   \n<\/code><\/pre>\n\n<p>I will split the code in the image preprocesing and the model prediction for clarity:<\/p>\n\n<h2>Load the image<\/h2>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing import image\nfrom keras.applications.vgg16 import preprocess_input\n\nimg_path = 'elephant.jpg'\nimg = image.load_img(img_path, target_size=(224, 224))\nx = image.img_to_array(img)\nx = np.expand_dims(x, axis=0)\nx = preprocess_input(x)\n<\/code><\/pre>\n\n<p>You can add prints along the way to see what's going on, but here is a brief summary:<\/p>\n\n<ol>\n<li><code>image.load_img()<\/code> load a PIL image, already in RGB and already reshaping it to (224, 224)<\/li>\n<li><code>image.img_to_array()<\/code> is translating this image into a matrix of shape (224, 224, 3). If you access, x[0, 0, 0] you will get the red component of the first pixels as a number between 0 and 255<\/li>\n<li><code>np.expand_dims(x, axis=0)<\/code> is adding the first dimension. x after is has shape <code>(1, 224, 224, 3)<\/code><\/li>\n<li><code>preprocess_input<\/code> is doing an extra preprocessing required for imagenet-trained architectures. From its docstring (run <code>help(preprocess_input)<\/code>) you can see that it:\n\n<blockquote>\n  <p>will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling<\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>This seems to be the standard input for ImageNet training set.<\/p>\n\n<p>That's it for the preprocessing, now you can just input the image in the pretrained model and get a prediction<\/p>\n\n<h2>Predict<\/h2>\n\n<pre><code>y_hat = base_model.predict(x)\nprint(y_hat.shape) # res.shape (1, 1000)\n<\/code><\/pre>\n\n<p><code>y_hat<\/code> contains the probabilities for each of the 1000 imagenet classes the model assigned to this image.<\/p>\n\n<p>In order to obtain the class names and a readable output, keras provided an utility function too:<\/p>\n\n<pre><code>from keras.applications.vgg16 import decode_predictions\ndecode_predictions(y_hat)\n<\/code><\/pre>\n\n<p>Outputs, for the <code>Zoorashia_elephant.jpg<\/code> image I downloaded before:<\/p>\n\n<pre><code>[[('n02504013', 'Indian_elephant', 0.48041093),\n  ('n02504458', 'African_elephant', 0.47474155),\n  ('n01871265', 'tusker', 0.03912963),\n  ('n02437312', 'Arabian_camel', 0.0038948185),\n  ('n01704323', 'triceratops', 0.00062475674)]]\n<\/code><\/pre>\n\n<p>Which seems pretty good!<\/p>\n"}]}
+{"QuestionId":53395787,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2018-11-20T15:00:09.937","AcceptedAnswerId":null,"OwnerUserId":10528635.0,"Title":"How to Save Prediction values for whole data in keras","Body":"<p>I am using pre-trained VGG16 model to classify images located in the folder. Currently, I am able to classify only one single image. <\/p>\n\n<ol>\n<li>How can I modify the code to classify all the images in the folder<\/li>\n<li>How can I save the prediction values for each image ?<\/li>\n<\/ol>\n\n<p>Below is my code :<\/p>\n\n<pre><code>from keras.preprocessing.image import load_img\nfrom keras.preprocessing.image import img_to_array\nfrom keras.applications.imagenet_utils import decode_predictions\nimport matplotlib.pyplot as plt\n\nfilename = 'cat.jpg'\n# load an image in PIL format\noriginal = load_img(filename, target_size=(224, 224))\nprint('PIL image size',original.size)\nplt.imshow(original)\nplt.show()\n\n# convert the PIL image to a numpy array\n# IN PIL - image is in (width, height, channel)\n# In Numpy - image is in (height, width, channel)\nnumpy_image = img_to_array(original)\nplt.imshow(np.uint8(numpy_image))\nplt.show()\nprint('numpy array size',numpy_image.shape)\n\n# Convert the image \/ images into batch format\n# expand_dims will add an extra dimension to the data at a particular axis\n# We want the input matrix to the network to be of the form (batchsize, height, width, channels)\n# Thus we add the extra dimension to the axis 0.\nimage_batch = np.expand_dims(numpy_image, axis=0)\nprint('image batch size', image_batch.shape)\nplt.imshow(np.uint8(image_batch[0]))\n\n\n\n# prepare the image for the VGG model\nprocessed_image = vgg16.preprocess_input(image_batch.copy())\n\n# get the predicted probabilities for each class\npredictions = vgg_model.predict(processed_image)\nprint (predictions)\n\n# convert the probabilities to class labels\n# We will get top 5 predictions which is the default\n#label = decode_predictions(predictions)\n<\/code><\/pre>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"53398208","CreationDate":"2018-11-20T17:17:23.093","ParentId":null,"OwnerUserId":"4279661","Title":null,"Body":"<p>You just have to list filenames and loop over them. You could use i.e. os.listdir, specifying the input folder. I removed the various imshow or plot chunks of code.<\/p>\n\n<p>Code below:<\/p>\n\n<pre><code>from keras.preprocessing.image import load_img\nfrom keras.preprocessing.image import img_to_array\nfrom keras.applications.imagenet_utils import decode_predictions\nimport matplotlib.pyplot as plt\nimport os\nnfiles = os.listdir(\".\/inputfolder\") # List filenames\n\nfor filename in nfiles: # Enter the loop\n    # load an image in PIL format\n    original = load_img(filename, target_size=(224, 224))    \n    # convert the PIL image to a numpy array\n    # IN PIL - image is in (width, height, channel)\n    # In Numpy - image is in (height, width, channel)\n    numpy_image = img_to_array(original)\n    print('numpy array size',numpy_image.shape)\n\n    # Convert the image \/ images into batch format\n    # expand_dims will add an extra dimension to the data at a particular \naxis\n    # We want the input matrix to the network to be of the form (batchsize, \nheight, width, channels)\n    # Thus we add the extra dimension to the axis 0.\n    image_batch = np.expand_dims(numpy_image, axis=0)\n    print('image batch size', image_batch.shape)\n\n    # prepare the image for the VGG model\n    processed_image = vgg16.preprocess_input(image_batch.copy())\n\n    # get the predicted probabilities for each class\n    predictions = vgg_model.predict(processed_image)\n    print (predictions)\n\n    # convert the probabilities to class labels\n    # We will get top 5 predictions which is the default\n    #label = decode_predictions(predictions)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53396087,"AnswerCount":0,"Tags":"<python-3.x><docker><tensorflow><object-detection-api>","CreationDate":"2018-11-20T15:16:20.780","AcceptedAnswerId":null,"OwnerUserId":9071615.0,"Title":"Trouble creating dockerfile for Object Detection API Tensorflow","Body":"<p>As the title suggests, I'm trying to build this dockerfile. To add some context: I successfully pulled the official repository and adapted the model to my needs. Following the docs on github installing and running the model on my local machine is not a problem.<\/p>\n\n<p>Now, to put this model into production I have to create a docker for the obvious reasons. First of all, the directory looks like this:<\/p>\n\n<pre><code>TFModels\n    &gt; Dockerfile\n    &gt; pycocotools\n    &gt; research\n          &gt; object_detection\n    &gt; slim\n<\/code><\/pre>\n\n<p>Now, the dockerfile which also encompasses the instructions from github concerning protoc, looks like this:<\/p>\n\n<pre><code>FROM python:3.5\n\nRUN  apt-get update \\\n  &amp;&amp; apt-get install -y wget \\\n  unzip \\\n  protobuf-compiler \\\n  python-pil \\\n  python-lxml \\\n  python-tk\n\nRUN mkdir -p \/TFModel\/\nCOPY research \/TFModel\/research\nCOPY slim \/TFModel\/slim\n\nRUN pip install --trusted-host pypi.python.org -r \/TFModel\/research\/object_detection\/requirements.txt\n\nWORKDIR \/TFModel\/research\n\nRUN protoc object_detection\/protos\/*.proto --python_out=.\nRUN export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`\/slim\n\nRUN python object_detection\/builders\/model_builder_test.py\n\nRUN python object_detection\/objectDetectImg.py\n<\/code><\/pre>\n\n<p>Building the docker gives following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n      File \"object_detection\/objectDetectImg.py\", line 24, in &lt;module&gt;\n        from utils import label_map_util\n      File \"\/TFModel\/research\/object_detection\/utils\/label_map_util.py\", line 21, in &lt;module&gt;\n        from object_detection.protos import string_int_label_map_pb2\n    ImportError: No module named 'object_detection'\n    The command '\/bin\/sh -c python object_detection\/objectDetectImg.py' returned a non-zero code: 1\n<\/code><\/pre>\n\n<p>I think the error may be in this line:<\/p>\n\n<pre><code>RUN export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`\/slim\n<\/code><\/pre>\n\n<p>Anyone who could help?\nThanks in advance<\/p>\n","answers":[]}
+{"QuestionId":53396147,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-20T15:19:08.743","AcceptedAnswerId":null,"OwnerUserId":4803860.0,"Title":"What is the difference between Model.train_on_batch from keras and Session.run([train_optimizer]) from tensorflow?","Body":"<p>In the following Keras and Tensorflow implementations of the training of a neural network, how <code>model.train_on_batch([x], [y])<\/code> in the keras implementation is different than <code>sess.run([train_optimizer, cross_entropy, accuracy_op], feed_dict=feed_dict)<\/code> in the Tensorflow implementation? In particular: how those two lines can lead to different computation in training?:<\/p>\n\n<p><em>keras_version.py<\/em><\/p>\n\n<pre><code>input_x = Input(shape=input_shape, name=\"x\")\nc = Dense(num_classes, activation=\"softmax\")(input_x)\n\nmodel = Model([input_x], [c])\nopt = Adam(lr)\nmodel.compile(loss=['categorical_crossentropy'], optimizer=opt)\n\nnb_batchs = int(len(x_train)\/batch_size)\n\nfor epoch in range(epochs):\n    loss = 0.0\n    for batch in range(nb_batchs):\n        x = x_train[batch*batch_size:(batch+1)*batch_size]\n        y = y_train[batch*batch_size:(batch+1)*batch_size]\n\n        loss_batch, acc_batch = model.train_on_batch([x], [y])\n\n        loss += loss_batch\n    print(epoch, loss \/ nb_batchs)\n<\/code><\/pre>\n\n<p><em>tensorflow_version.py<\/em><\/p>\n\n<pre><code>input_x = Input(shape=input_shape, name=\"x\")\nc = Dense(num_classes)(input_x)\n\ninput_y = tf.placeholder(tf.float32, shape=[None, num_classes], name=\"label\")\ncross_entropy = tf.reduce_mean(\n    tf.nn.softmax_cross_entropy_with_logits_v2(labels=input_y, logits=c, name=\"xentropy\"),\n    name=\"xentropy_mean\"\n)\ntrain_optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(cross_entropy)\n\nnb_batchs = int(len(x_train)\/batch_size)\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    for epoch in range(epochs):\n        loss = 0.0\n        acc = 0.0\n\n        for batch in range(nb_batchs):\n            x = x_train[batch*batch_size:(batch+1)*batch_size]\n            y = y_train[batch*batch_size:(batch+1)*batch_size]\n\n            feed_dict = {input_x: x,\n                         input_y: y}\n            _, loss_batch = sess.run([train_optimizer, cross_entropy], feed_dict=feed_dict)\n\n            loss += loss_batch\n        print(epoch, loss \/ nb_batchs)\n<\/code><\/pre>\n\n<p><em>Note: This question follows <a href=\"https:\/\/stackoverflow.com\/questions\/53391500\/same-model-converges-in-keras-but-not-in-tensorflow\">Same (?) model converges in Keras but not in Tensorflow<\/a> , which have been considered too broad but in which I show exactly why I think those two statements are somehow different and lead to different computation.<\/em><\/p>\n","answers":[{"AnswerId":"53457574","CreationDate":"2018-11-24T11:17:11.733","ParentId":null,"OwnerUserId":"8382707","Title":null,"Body":"<p>Yes, the results can be different. The results shouldn't be surprising if you know the following things in advance:<\/p>\n\n<ol>\n<li>Implementation of <code>corss-entropy<\/code> in Tensorflow and Keras is different. Tensorflow assumes the input to <code>tf.nn.softmax_cross_entropy_with_logits_v2<\/code> as the raw unnormalized logits while <code>Keras<\/code> accepts inputs as probabilities<\/li>\n<li>Implementation of <code>optimizers<\/code> in Keras and Tensorflow are different.<\/li>\n<li>It might be the case that you are shuffling the data and the batches passed aren't in the same order. Although it doesn't matter if you run the model for long but initial few epochs can be entirely different. Make sure same batch is passed to both and then compare the results.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53396670,"AnswerCount":1,"Tags":"<python><tensorflow><random-seed>","CreationDate":"2018-11-20T15:47:55.257","AcceptedAnswerId":53400336.0,"OwnerUserId":7886651.0,"Title":"Unable to reproduce tensorflow results even after setting random seed","Body":"<p>To be able to perform hyperparameter optimization, I guess one should better set the random seed so we will better compare the models. So I set the random seed in tensorflow through: <code>tf.set_random_seed(mseed)<\/code>, and I created an initializer as: <code>kernel_initializer = tf.glorot_normal_initializer(seed=mseed)<\/code> whereI passed the later into the <code>LSTM<\/code> cell. So the model will be as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n\ndef lstm_model(lstm_model(inputs, cell_size1, kernel_initializer, m_dtype, dropout, is_training, use_peepholes, mseed, num_classes):\n\n    with tf.variable_scope('lstm_model'):\n\n        cell = tf.nn.rnn_cell.LSTMCell(cell_size1, initializer=kernel_initializer)\n        initial_state = cell.zero_state(34 if is_training else 14, dtype=tf.float32)\n\n        output, new_state = tf.nn.dynamic_rnn(cell, inputs, dtype=m_dtype, initial_state=initial_state)\n\n\n    with tf.variable_scope(\"output\"):\n        output = tf.reshape(output, shape=[-1, cell_size1])\n        output = tf.layers.dense(output, units=num_classes,\n                                 kernel_initializer=kernel_initializer)\n\n        if is_training:\n            output = tf.reshape(output, shape=[34, -1, num_classes])\n        else:\n            output = tf.reshape(output, shape=[14, -1, num_classes])\n\n        return output, new_state, initial_state, model_summary\n\n\ndef model(inputs..., mseed, kernel_initializer, reuse=False):\n    with tf.variable_scope('model', reuse=reuse):\n\n        output, new_state, initial_state, model_summary = lstm_model(inputs, num_units,\n                                                                     kernel_initializer,\n                                                                     tf.float32, dropout, not reuse, True,\n                                                                     mseed, num_classes=num_classes)\n\n    # Now I calculate the loss and used an optimizer in case of training...\n    return output, new_state, initial_state, model_summary\n\nmseed = 123\nseed(mseed)\ntf.set_random_seed(mseed)\nkernel_initializer = tf.glorot_normal_initializer(seed=mseed)\n\n# here I loaded the data as numpy arrays...\n# Here I created the placeholders...\n\n# t for train, and d for development\noutput_t, new_state_t, init_state_t = model(inputs_t..., mseed, kernel_initializer, reuse=False)\noutput_d, new_state_d, init_state_d = model(inputs_d..., mseed, kernel_initializer, reuse=True)\n\ntrain_model()...\n<\/code><\/pre>\n\n<p>So the code is summarized to include only the important part of it.<\/p>\n\n<p>Now even after setting the random seed and using a <code>mseed<\/code> when creating the kernels, yet I cannot reproduce the same results. <\/p>\n\n<p>Here is a screen shot for the accuracy and loss over time. We can see that the values are the same almost at the beginning and then they become different. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Vpz0M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Vpz0M.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vCuLR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vCuLR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I would like to know where I could be mistaken or at least what part to debug of the code. Finally, please note that through out training, I am feeding back the last hidden state into the initial state at each batch. as follows:<\/p>\n\n<pre><code>new_state_train_py = sess.run(initiale_state_train)\n    for i in range(num_iterations):\n        _, summary_t, new_state_train_py = sess.run([train_step, summary_train, new_state_train],\n                                feed_dict={train_x_ph: train_x[:, i*time_steps: (i + 1) * time_steps, :],\n                                           train_y_ph: train_y[:, i*time_steps: (i + 1) * time_steps, :],\n                                           initiale_state_train: new_state_train_py})\n\n        train_writer.add_summary(summary_t, epoch * num_iterations + i)\n<\/code><\/pre>\n\n<p>What I know so far is that the main problem should be in the model definition since this what defines any randomness in the whole model and this is what makes the prediction different between the different runs of the same model. Please correct me if I am mistaken. <\/p>\n\n<p><strong>After debugging, I noticed that when I changed the dtype for each op in the graph to tf.float64, I found that both curves totally intersect except at the end, I get another slight changes in the curves. What could be causing this behavior?<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9ZW0J.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9ZW0J.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/rBzRk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rBzRk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[{"AnswerId":"53400336","CreationDate":"2018-11-20T19:36:43.257","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>If you're running code on a GPU, it is likely due to the non-deterministic behavior of cuDNN (see <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/theano-users\/718YYXwaYEk\" rel=\"nofollow noreferrer\">this thread<\/a> for more details). The order in which some operations are executed on a GPU can be random due to performance optimizations. This means that rounding errors also occur in a different order, which leads to small differences in the result of these operations. In your case, these small differences add up over the course of training, which leads to significantly different behavior already after a few training steps. <\/p>\n\n<p>The order of magnitude of the rounding errors depends on the floating point precision used by the GPU. With <code>float64<\/code>, the rounding errors take a lot longer to add up noticeably than with <code>float32<\/code>. <\/p>\n\n<p>On a CPU, this non-deterministic behavior should not occur when python's, <code>numpy<\/code>'s and <code>tensorflow<\/code>'s random seeds are fixed (and op parallelism is deactivated, more info <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-reproducible-results-using-keras-during-development\" rel=\"nofollow noreferrer\">here<\/a>). So, if you run your code on a CPU, you should get the same results for every run (but that of course takes a lot longer).<\/p>\n"}]}
+{"QuestionId":53397589,"AnswerCount":1,"Tags":"<python><tensorflow><windows-10>","CreationDate":"2018-11-20T16:40:20.080","AcceptedAnswerId":null,"OwnerUserId":10681328.0,"Title":"No module named 'tensorflow' error showing up in Windows 10 Python 3.6","Body":"<p>I installed Tensorflow using the instructions on its website <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a>. I used the virtual environment option, but didn't use Anaconda.<br\/>\nEverything went smoothly until I tried to import tensorflow in python. I did \n    <code>import tensorflow as tf<\/code>\nin my Python file and in return I get <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"...\", line 1, in &lt;module&gt;\n    import tensorflow as tf\nModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>What's up with this? How can I fix it? I'm using Python 3.6 on a Windows 10 machine. All other answers to this I could find involved Anaconda which I didn't use. <br\/>\nThanks in advance!<\/p>\n","answers":[{"AnswerId":"53398023","CreationDate":"2018-11-20T17:06:03.073","ParentId":null,"OwnerUserId":"7221127","Title":null,"Body":"<p>I recommend you start using Anaconda. It just makes life so much easier (especially when handling different Versions of TensorFlow)<\/p>\n\n<p>Make sure that you are <em>in<\/em> the Environment you\u2018ve installed it. Note that you need to activate it first (as shown on the Installation tutorial). <\/p>\n\n<p>Another thing you can try is installing it with <code>pip3<\/code>. Although this didn\u2018t always work for me (it can lead to trouble when having python 3.7 installed)<\/p>\n\n<p>Lastly you could lookup which python is linked in the PATH and verify that the tensorflof packages are actually there.<\/p>\n"}]}
+{"QuestionId":53398291,"AnswerCount":0,"Tags":"<cmake><anaconda><caffe2>","CreationDate":"2018-11-20T17:21:38.027","AcceptedAnswerId":null,"OwnerUserId":2130515.0,"Title":"cmake undefined reference to `pthread_create' installing Caffe2","Body":"<p>I am installing Caffe2 from this <a href=\"https:\/\/jin-zhe.github.io\/guides\/installing-caffe2-with-cuda-on-anaconda\/\" rel=\"nofollow noreferrer\">Caffe2<\/a>. When I run the cmake command I get this error :<\/p>\n\n<pre><code>cmake -DCUDNN_INCLUDE_DIR=~\/cuda\/include -DCUDNN_LIBRARY=~\/cuda\/lib64\/libcudnn.so  -DCUDA_HOST_COMPILER=\/usr\/bin\/gcc-5 -DUSE_NNPACK=OFF -DCMAKE_PREFIX_PATH=~\/anaconda3\/envs\/caffe2 -DCMAKE_INSTALL_PREFIX=~\/anaconda3\/envs\/caffe2 ..\n<\/code><\/pre>\n\n<p><strong>cmake undefined reference to `pthread_create'<\/strong>\nI am using:<br>\nubuntu 18.0.4<br>\nanconda3<br>\ncmake 3.10.0  <\/p>\n\n<pre><code>ldconfig -p | grep pthread\n    libpthread.so.0 (libc6,x86-64, OS ABI: Linux 3.2.0) =&gt; \/lib\/x86_64-linux-gnu\/libpthread.so.0\n    libpthread.so.0 (libc6, OS ABI: Linux 3.2.0) =&gt; \/lib\/i386-linux-gnu\/libpthread.so.0\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53398721,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><reshape><torch>","CreationDate":"2018-11-20T17:49:37.037","AcceptedAnswerId":null,"OwnerUserId":497600.0,"Title":"tensorflow: can reshape() create a copy?","Body":"<p>In v4 of their API, torch has introduced <code>reshape()<\/code>, to align more with the style of numpy. Previously, changing the shape of a torch tensor was done with <code>view()<\/code>.<\/p>\n\n<p>I wondered whether <code>view()<\/code> was going to be deprecated now and looked at the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\">docs<\/a>. Turns out that <code>reshape()<\/code> is not just a numpy-friendly alias for <code>view()<\/code>, actually, it has a different semantic. Torch tries to give you contiguous memory where possible. If the new view dimensions violate a contiguity constraint, you have to explicitly call <code>contiguous()<\/code> before <code>view()<\/code>. <code>Reshape<\/code> will work even if this constraint is violated, but will silently make a copy of the data. \nThis is the same behaviour as in numpy, where <code>reshape<\/code> <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.1\/reference\/generated\/numpy.reshape.html#numpy-reshape\" rel=\"nofollow noreferrer\">can produce copies, too<\/a>.<\/p>\n\n<p>A question on <code>view()<\/code> vs <code>reshape()<\/code> in torch is here: <a href=\"https:\/\/stackoverflow.com\/questions\/49643225\/whats-the-difference-between-reshape-and-view-in-pytorch\">What&#39;s the difference between reshape and view in pytorch?<\/a><\/p>\n\n<blockquote>\n  <p>if you need a copy use clone() if you need the same storage use\n  view(). The semantics of reshape() are that it may or may not share\n  the storage and you don't know beforehand.<\/p>\n<\/blockquote>\n\n<p>Up until now, torch only offered <code>view()<\/code>. Maybe to intentionally force their developers to care about memory layout. Which makes me wonder how <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\"><code>reshape()<\/code><\/a> works in tensorflow.<\/p>\n\n<p>In torch, the distinction between a view and a copy could produce complicated bugs. You assume that tensors share data, but they don't.\nIn tensorflow this problem shouldn't exist. A tensorflow tensor is symbolic and doesn't hold a value. A reshape is just an Op in the tensorflow graph. While the graph is evaluated, data in placeholders and variables don't change, so it is clear what data you are working on.<\/p>\n\n<p>But I don't know if this could hurt performance. Copying a huge tensor can be very expensive. Do I have to be careful when using <code>reshape()<\/code>, not to duplicate memory ?<\/p>\n","answers":[]}
+{"QuestionId":53400472,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-20T19:46:06.647","AcceptedAnswerId":null,"OwnerUserId":7374099.0,"Title":"Keras Model Weights for some layers become all \"NaNs\"","Body":"<p>I am training a triplet model similar to -<a href=\"https:\/\/stackoverflow.com\/questions\/51292212\/keras-model-params-are-all-nans-after-reloading\">Keras model params are all &quot;NaN&quot;s after reloading<\/a>; except that this model is built over the inception_v3 model. <\/p>\n\n<p>(I am using Keras with Tensorflow backend)<\/p>\n\n<p>But after just 2 epochs the model weights turn out to be NaN. And when I try to extract learnt features by passing an Input Image the features are all 0. <\/p>\n\n<p>Model Architecture- <\/p>\n\n<p>def triplet_loss(x, ALPHA=0.2):                                         <\/p>\n\n<pre><code>anchor, positive, negative = x                                      \n\npos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), 1)\nneg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), 1)\n\nbasic_loss = tf.add(tf.subtract(pos_dist, neg_dist), ALPHA)         \nloss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)               \n\nreturn loss                                                         \n<\/code><\/pre>\n\n<p>class StyleNet():                                                       <\/p>\n\n<pre><code>def __init__(self, input_shape_x, input_shape_y, input_shape_z, n_classes, reg_lambda):\n\n    self.input_shape_x = input_shape_x                              \n    self.input_shape_y = input_shape_y                              \n    self.input_shape_z = input_shape_z                              \n    self.n_classes = n_classes                                      \n    self.reg_lambda = reg_lambda                                    \n\n\ndef create_model(self):                                             \n\n    anchor_example = Input(shape=(self.input_shape_x, self.input_shape_y, self.input_shape_z), name='input_1')\n    positive_example = Input(shape=(self.input_shape_x, self.input_shape_y, self.input_shape_z), name='input_2')\n    negative_example = Input(shape=(self.input_shape_x, self.input_shape_y, self.input_shape_z), name='input_3')\n\n    input_image = Input(shape=(self.input_shape_x, self.input_shape_y, self.input_shape_z))\n\n    base_inception = InceptionV3(input_tensor = input_image, input_shape=(self.input_shape_x, self.input_shape_y, self.input_shape_z), weights=None, include_top=False, pooling='avg')\n    base_pool5 = base_inception.output                              \n\n    ##############Adding the Bottleneck layer Here#######################################################\n    bottleneck_layer = Dense(256, kernel_regularizer=l2(self.reg_lambda), name='bottleneck_layer')(base_pool5)\n    bottleneck_norm = BatchNormalization(name='bottleneck_norm')(bottleneck_layer)\n    bottleneck_relu = Activation('relu', name='bottleneck_relu')(bottleneck_norm)\n    bottleneck_drop = Dropout(0.5)(bottleneck_relu)                 \n\n    fin = Dense(self.n_classes)(bottleneck_drop)                    \n    fin_norm = BatchNormalization(name='fin_norm')(fin)             \n    fin_softmax = Activation('softmax')(fin_norm)                   \n    ######################################################################################################\n\n    ###########Triplet Model Which learns the embedding layer relu6####################\n    self.triplet_model = Model(input_image, bottleneck_drop)        \n    positive_embedding = self.triplet_model(positive_example)       \n    negative_embedding = self.triplet_model(negative_example)       \n    anchor_embedding = self.triplet_model(anchor_example)           \n    ###########Triplet Model Which learns the embedding layer relu6####################\n\n    adam_opt = optimizers.Adam(lr=0.00001, clipnorm = 1.0, amsgrad=False)\n\n    #The Triplet Model which optimizes over the triplet loss.       \n    loss = Lambda(triplet_loss, output_shape=(1,))([anchor_embedding, positive_embedding, negative_embedding])\n    self.triplet_model_worker = Model(inputs=[anchor_example, positive_example, negative_example], outputs = loss)\n    self.triplet_model_worker.compile(loss='mean_absolute_error', optimizer=adam_opt)\n\ndef fit_model(self, pathname='.\/models\/'):                          \n    if not os.path.exists(pathname):                                \n        os.makedirs(pathname)                                       \n    if not os.path.exists(pathname+'\/weights'):                     \n        os.makedirs(pathname+'\/weights')                            \n    if not os.path.exists(pathname+'\/tb'):                          \n        os.makedirs(pathname+'\/tb')                                 \n    filepath=pathname+\"weights\/{epoch:02d}.hdf5\"                    \n    checkpoint = ModelCheckpoint(filepath, monitor='loss', verbose=1, save_best_only=False, mode='auto')\n    tensorboard = TensorBoard(log_dir=pathname+'\/tb', write_graph=True, write_images=True)\n    callbacks_list = [checkpoint, tensorboard]                      \n\n    #Parameter                                                              \n    params = {'dim': (224, 224), 'batch_size':32, 'n_classes':11, 'n_channels':3, 'shuffle':True}\n\n    #Datasets                                                               \n    partition = pickle.load(open('..\/..\/..\/data\/bam_2_partition_triplet.pkl', 'rb'))\n    labels = pickle.load(open('..\/..\/..\/data\/bam_2_labels_triplet.pkl', 'rb'))\n\n    #Generators                                                             \n    training_generator = DataGenerator(partition['train'], labels, **params)\n    self.triplet_model_worker.fit_generator(generator = training_generator,  epochs = 60, use_multiprocessing=True, workers = 10, callbacks = callbacks_list, verbose = 1)\n<\/code><\/pre>\n\n<p>What is troublesome is as answered in the link above. Even after using ***clipnorm=1.0**** the gradients explode and the weights give out \"nan\" values.<\/p>\n\n<p>Saving and loading the model, then printing the weights. The NaNs are clearly visible.\n   Loading code:<\/p>\n\n<pre><code>m = load_model('\/scratch\/models_inception_stage2\/yo\/weights\/02.hdf5', custom_objects={\"tf\":tf})\n\nfor layer in m.layers:                                                  \n    weights = layer.get_weights()                                       \n    print (weights)\n<\/code><\/pre>\n\n<p>Snippet for the printed weights<\/p>\n\n<pre><code>Here\n[array([ 3.4517611e-04,  1.3431008e-03, -1.1081886e-03,  2.6104850e-04,\n   -2.1620051e-04,  1.6816283e-03,  8.8927911e-05, -3.8964470e-04,\n    1.7968584e-03,  1.0259283e-03,  5.0400384e-04, -3.6578919e-04,\n   -1.1292399e-03,  1.1509922e-03,  3.2478449e-04, -3.6580343e-05,\n   -4.4458261e-04,  4.8210021e-04, -9.5213606e-04, -6.4406055e-04,\n    5.0959276e-04, -3.4098624e-04, -7.0486858e-05,  2.8134760e-04,\n   -8.0100907e-04,  8.2962180e-04, -6.4140803e-04,  9.4872032e-04,\n   -3.3409546e-05, -3.0277384e-04,  5.2237371e-04, -8.3427120e-04,\n   -2.5856070e-04, -1.0346439e-03,  4.3354488e-05, -8.8099617e-04,\n   -6.8233605e-04, -1.2386916e-04,  8.2019303e-04, -1.9070004e-03,\n    1.5571159e-03, -3.4599879e-04,  6.2088901e-04, -8.4720332e-06,\n    1.6024955e-04, -1.2059419e-03, -1.4946899e-04, -6.7080715e-04,\n   -2.8154058e-05,  5.1517348e-04,  5.9993083e-05,  2.8555689e-04,\n    3.9626448e-04, -5.1538437e-04,  1.9132573e-04,  1.1226863e-03,\n    1.1591403e-03, -6.3404470e-04,  2.8910063e-04, -7.9366821e-04,\n   -1.7228167e-04,  6.2899920e-04,  1.7438219e-04,  1.1385380e-04],\n  dtype=float32), array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n   nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan],\n  dtype=float32), array([0.50248814, 0.48732147, 0.64627343, 0.49432775, 0.45906776,\n   0.5168214 , 0.8561428 , 0.7308014 , 0.5067555 , 0.516555  ,\n   1.3287057 , 0.5746422 , 0.55597156, 1.0038179 , 0.9859771 ,\n   0.6110601 , 0.7357226 , 0.6123694 , 0.90676117, 0.5439505 ,\n   0.48629472, 0.5434108 , 0.4934845 , 0.5407317 , 0.6443982 ,\n   1.0403991 , 0.48624724, 0.83786434, 0.72478205, 0.7294607 ,\n   0.536994  , 0.38235992, 1.0484552 , 0.45833316, 0.48205158,\n   0.48236838, 0.71035874, 0.9472658 , 0.78085536, 1.0207686 ,\n   0.5089741 , 0.97984046, 0.86524594, 0.9828817 , 0.49027866,\n   0.7367909 , 0.57438385, 0.5011991 , 0.47189236, 0.52376693,\n   0.45648402, 0.40523565, 0.8375675 , 0.57908285, 0.6055632 ,\n   1.0325785 , 0.5377976 , 0.47033092, 0.83586556, 1.2780553 ,\n   0.503384  , 0.54509026, 0.5375585 , 0.6091993 ], dtype=float32)]\n<\/code><\/pre>\n\n<p>Would be grateful for any help.<\/p>\n","answers":[]}
+{"QuestionId":53400895,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2018-11-20T20:17:59.317","AcceptedAnswerId":53407323.0,"OwnerUserId":7043716.0,"Title":"How to add a second input argument (the first is an image) to a CNN model built with Keras?","Body":"<p>Let's say I have a list of <em>images<\/em> (converted to numpy arrays) downloaded from Instagram, along with their corresponding <em>likes<\/em> and user <em>followers<\/em>. And let's say I have a <strong>CNN model<\/strong> (using <strong>Keras<\/strong> on <strong>Tensorflow<\/strong>) which I train on these images (200x200x3 numpy arrays) and it tries to predict the number of <em>likes<\/em> an image will get.<\/p>\n\n<p>What if I want to give to this model each image's corresponding <em>followers<\/em> as a <strong>second input<\/strong>?<\/p>\n\n<p>This is my code so far:<\/p>\n\n<pre><code>IMAGESIZE = (200, 200)\n\ndef create_model():\n    # create model and add layers\n    model = Sequential()\n\n    model.add(Conv2D(10, 5, 5, activation='relu',\n                     input_shape=(IMAGESIZE[0], IMAGESIZE[1], 3)))\n\n    model.add(Conv2D(10, 5, 5, activation='relu'))\n    model.add(MaxPool2D((5, 5)))\n    model.add(Dropout(0.2))\n    model.add(Flatten())\n    model.add(Dense(50))\n    model.add(Activation('relu'))\n    model.add(Dense(1))\n\n    print(model.summary())\n\n    model.compile(loss='mse',\n                  optimizer='rmsprop', metrics=[\"accuracy\"])\n    return model\n\n# Read the likes\nlikes = getlikes(src='..\/data\/pickledump')\nlikesraw = np.array(likes)\nlikes = (likesraw - np.mean(likesraw))\/np.std(likesraw)  # normalize\n\n# Read the images and resize them\nimages = []\nfor imgfile in glob.glob('..\/data\/download\/*.jpeg'):\n    img = cv2.imread(imgfile)\n    resized = cv2.resize(img, IMAGESIZE)\n    images.append(resized)\n    break\nimages = np.array(images)\n\n# Read the followers\nfollowers= getfollowers(src='..\/data\/pickledump')\nfollowersraw= np.array(followers)\nfollowers= (followersraw- np.mean(followersraw))\/np.std(followersraw)  # normalize\n\nclassifier = KerasClassifier(build_fn=create_model, epochs=20)\nprint(\"Accuracy (Cross Validation=10): \",\n      np.mean(cross_val_score(classifier, images, likes, cv=2)))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53407323","CreationDate":"2018-11-21T07:42:13.607","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One approach is to use a two branch model, where one branch processes the image and another branch processes other non-image inputs (such as posts texts or number of followers and followings, etc.). Then you can merge the result of these two branches and possibly add a few other layers afterwards to act a the final classifier\/regressor. To build such a model in Keras you need to use the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">functional API<\/a> instead. Just for demonstration, here is an example:<\/p>\n\n<pre><code>inp_img = Input(shape=image_shape)\ninp_others = Input(shape=others_shape)\n\n# branch 1: process input image\nx = Conv2D(...)(inp_img)\nx = Conv2D(...)(x)\nx = MaxPool2D(...)(x)\nout_b1 = Flatten()(x)\n\n# branch 2: process other input\nout_b2 = Dense(...)(inp_other)\n\n\n# merge the results by concatenation\nmerged = concatenate([out_b1, out_b2])\n\n# pass merged tensor to some other layers\nx = Dense(...)(merged)\noutput = Dense(...)(x)\n\n# build the model and compile it\nmodel = Model([inp_img, inp_other], output)\nmodel.compile(...)\n\n# fit on training data\nmodel.fit([img_array, other_array], label_array, ...)\n<\/code><\/pre>\n\n<p>Note that we used <code>concatenation<\/code> layer above, but there are other <a href=\"https:\/\/keras.io\/layers\/merge\/\" rel=\"nofollow noreferrer\">merge layers<\/a> which you can use. And make sure you read the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">functional API guide<\/a>, it's a must-read guide.<\/p>\n"}]}
+{"QuestionId":53401053,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-20T20:31:26.163","AcceptedAnswerId":53401140.0,"OwnerUserId":10630432.0,"Title":"Is there .all() or .any() equivalent in python Tensorflow","Body":"<p>Trying to find a similar operation to <code>.any()<\/code>, <code>.all()<\/code> methods that will work on a tensor. Here is a scenario:<\/p>\n\n<pre><code>a = tf.Variable([True, False, True], dtype=tf.bool)\n\n# this is how I do it right now\nhas_true = a.reduce_sum(tf.cast(a, tf.float32)) \n# this is what I'm looking for\nhas_true = a.any()\n<\/code><\/pre>\n\n<p>Currently converting my boolean tensor into <code>int<\/code> using <code>reduce_sum<\/code> to see if there are any truths in it. Is there a cleaner way to perform this operation?<\/p>\n","answers":[{"AnswerId":"53401140","CreationDate":"2018-11-20T20:38:26.120","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p>There are <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/reduce_any\" rel=\"nofollow noreferrer\"><code>tf.reduce_any<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/reduce_all\" rel=\"nofollow noreferrer\"><code>tf.reduce_all<\/code><\/a> methods:<\/p>\n\n<pre><code>sess = tf.Session()\n\na = tf.Variable([True, False, True], dtype=tf.bool)\nsess.run(tf.global_variables_initializer())\n\nsess.run(tf.reduce_any(a))\n# True\nsess.run(tf.reduce_all(a))\n# False\n<\/code><\/pre>\n"}]}
+{"QuestionId":53401076,"AnswerCount":0,"Tags":"<tensorflow><global>","CreationDate":"2018-11-20T20:33:26.457","AcceptedAnswerId":null,"OwnerUserId":1161315.0,"Title":"tensorflow - map_fn - global variable access","Body":"<p>Could you tell me the right way to return from the above function with return statement :  return z[key] ? <\/p>\n\n<pre><code>from IPython import display\nimport pandas as pd\nimport tensorflow as tf\nimport numpy as np\nimport argparse\nimport sys\n\nglobal z\nz = np.arange(400, 420 + 1)\n\ndef try_mapping(key):\n    # print(key)\n    # print(key.dtype)\n    #return key*2\n    return z[key]\n\ndef inputs(data):\n    keys = tf.map_fn(try_mapping , data)\n    return keys\n\ndef main(unused_argv):\n    x = np.arange(1, 20 + 1)\n    df = pd.DataFrame(data={'Key': x})\n    keys  = inputs(df)\n    with tf.Session() as sess:\n        print(sess.run(keys))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53401117,"AnswerCount":3,"Tags":"<tensorflow><machine-learning><neural-network><tflearn>","CreationDate":"2018-11-20T20:36:13.630","AcceptedAnswerId":53785786.0,"OwnerUserId":1744914.0,"Title":"How to properly reward neural network when playing a game","Body":"<p>I'm pretty new to deep learning and neural networks and trying to implement an agent that would be able to play my simple game<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pcOuo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pcOuo.png\" alt=\"game field\"><\/a><\/p>\n\n<p>So the goal is to get the highest possible score (sum of cells visited) while reaching towards the destination (orange cell) within steps available (always gte distance from the player to the finish cell).<\/p>\n\n<p>Model for my network is really simple (I'm using <code>tflearn<\/code>)<\/p>\n\n<pre><code>network = input_data(shape=[None, 13, 1], name='input')\nnetwork = fully_connected(\n    network,\n    13**2,\n    activation='relu'\n)\nnetwork = fully_connected(network, 1, activation='linear')\nnetwork = regression(\n    network,\n    optimizer='adam',\n    learning_rate=self.lr,\n    loss='mean_square',\n    name='target',\n)\nmodel = tflearn.DNN(network, tensorboard_dir='log')\n<\/code><\/pre>\n\n<p>where <code>13<\/code> is a number of features I am able to extract from the game state. But the resulting model gives really bad behaviour when playing<\/p>\n\n<pre><code>[default] INFO:End the game with a score: 36\n[default] INFO:Path: up,up,up,up,up,up,up,up,up,up,up,up,up,up,up\n<\/code><\/pre>\n\n<p>So I want to figure out what important parts I have missed and have some open questions to clarify:<\/p>\n\n<pre><code>Training Step: 3480  | total loss: 0.11609 | time: 4.922s\n| Adam | epoch: 001 | loss: 0.11609 -- iter: 222665\/222665\n<\/code><\/pre>\n\n<ol>\n<li>What is the loss value I'm looking for? Is there a rule of thumb that will tell me a loss is good enough?<\/li>\n<li>How many epochs do I need to have? How to figure out the exact number of them?<\/li>\n<li>What if my NN architecture is completely wrong and unsuitable for this task? How to spot that?<\/li>\n<li>And finally: what is a good place to start when debugging your network and which aspects should I double-check and verify first.<\/li>\n<\/ol>\n\n<p>I understand that this is a slightly open question and it might be inappropriate to post it here, so I'll appreciate any kind of guidance or general comments.<\/p>\n","answers":[{"AnswerId":"53541633","CreationDate":"2018-11-29T14:48:28.650","ParentId":null,"OwnerUserId":"1744914","Title":null,"Body":"<p>Not an answer to the question above but a good place to start to obtain valueable information for your specific network <a href=\"https:\/\/stats.stackexchange.com\/questions\/181\/how-to-choose-the-number-of-hidden-layers-and-nodes-in-a-feedforward-neural-netw?newreg=4109ced4e5d24473aa87523a965e5dcd\">here<\/a><\/p>\n"},{"AnswerId":"53751531","CreationDate":"2018-12-12T21:20:44.143","ParentId":null,"OwnerUserId":"1744914","Title":null,"Body":"<p>Ok, so you should solve the problem with proper tools. As mentioned in comments the right way to do that is by using Reinforcement Learning. Here's the algorithm that returns optimal policy for our environment (based on <code>Q-learning<\/code>)<\/p>\n\n<pre><code>states_space_size = (game.field.leny - 2)*(game.field.lenx - 2)\nactions_space_size = len(DIRECTIONS)\nQSA = np.zeros(shape=(states_space_size, actions_space_size))\nmax_iterations = 80\ngamma = 1  # discount factor\nalpha = 0.9  # learning rate\neps = 0.99  # exploitation rate\ns = 0  # initial state\nfor i in range(max_iterations):\n    # explore the world?\n    a = choose_an_action(actions_space_size)\n    # or not?\n    if random.random() &gt; eps:\n        a = np.argmax(QSA[s])\n\n    r, s_ = perform_action(s, a, game)\n    qsa = QSA[s][a]\n    qsa_ = np.argmax(QSA[s_])\n    QSA[s][a] = qsa + alpha*(r + gamma*qsa_ - qsa)\n\n    # change state\n    s = s_\nprint(QSA)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stats.stackexchange.com\/questions\/380623\/value-iteration-does-not-converge-when-using-q-learning\">Here's<\/a> more detailed explanation with a simplified example of how to achieve this result.<\/p>\n"},{"AnswerId":"53785786","CreationDate":"2018-12-14T19:25:48.287","ParentId":null,"OwnerUserId":"9191460","Title":null,"Body":"<p>Traditionally, reinforcement learning was limited to only solving discrete state discrete action problems because continuous problems caused the problem of \"curse of dimensionality\".  For example, lets say a robot arm can move between 0 - 90 degrees.  That means you need an action for angle = 0, 0.00001, 0.00002, ..., which is infeasible for traditional tabular based RL.<\/p>\n\n<p>To solve this problem, we had to teach RL that 0.00001 and 0.00002 are more or less the same.  To achieve this, we need to use function approximation such as neural networks.  These approximations' aim is to approximate the Q-matrix in tabular RL and capture the policy (i.e., the choices of the robot).  However, even up to today, non-linear function approximation are known to be extremely difficult to train.  The first time NNs were successful in RL was by David Silver and his deterministic policy gradient (2014).  His approach was to map states to actions directly, without the Q-value.  But the loss function of the neural network would be guided by rewards.<\/p>\n\n<p>To answer the  original question of \"how to properly reward the NN\":<\/p>\n\n<ol>\n<li>Generate many trajectories of your robots movement in the state space. In your example, one trajectory is ends either when the agent reaches the goal, or if he took more than 50 steps (taking too long).  We will call each step as one episode.<\/li>\n<li>After each <strong>successful<\/strong> trajectory (reaching the goal), reward the last episode a reward of 1, and each episode before that should be discounted by your discount rate.  Example:  The 2nd last episode would get a reward of 0.95 if your discount rate is 0.95, etc.<\/li>\n<li>After adequate trajectories are collected (~50), pose this as a supervised learning problem.  Where your inputs are states, target are actions, and reward is the cross entropy reward guided loss: -log(pi)*R, where pi is the probability of taking that action in that state.  R is your current reward.<\/li>\n<li>Train using a variant of gradient descent, taking the gradient of the cross entropy reward guided loss: - (dPi\/pi * R).  Here, you can see a very negative R will have a high loss, whereas a positive R will have a low loss.<\/li>\n<li>Repeat until convergence.  This training method is called a Monte-Carlo method because we generate many trajectories and we say the average of each S, A, R, S' pair is very likely.<\/li>\n<\/ol>\n\n<p>Here is the original paper: <a href=\"http:\/\/proceedings.mlr.press\/v32\/silver14.pdf\" rel=\"nofollow noreferrer\">http:\/\/proceedings.mlr.press\/v32\/silver14.pdf<\/a><\/p>\n\n<p>The issue with monte carlo methods is their high variance, because each trajectory can be highly different from others.  So then modern RL (late 2015 - now) uses an actor-critic method where the actor is the above algorithm, but there's another critic that approximates the Q-matrix using a neural network.  This critic attempts to stabilize the actors learning by giving it information after each episode.  Thus, reducing the variance of the actor.  <\/p>\n\n<p>The 2 most popular algorithms are:  Deep deterministic policy gradient and proximal policy optimization.<\/p>\n\n<p>I would recommend you get familiar with deterministic policy gradients first before trying the other ones.<\/p>\n"}]}
+{"QuestionId":53401198,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-11-20T20:42:51.730","AcceptedAnswerId":null,"OwnerUserId":875295.0,"Title":"EvalSpec input_fn : case where the feature is an array","Body":"<p>I have some trouble understanding some details of the Estimator API and <code>tf.estimator.EvalSpec<\/code>. \nIn an <code>EvalSpec<\/code>, the user is supposed to give a <code>input_fn<\/code>. A call to input_fn is supposed to return <code>A tuple (features, labels)<\/code> .\nAs far as I understand, the features can be a dictionary keyed by \"feature name\" and whose values are a tensor of values. For instance, if I have a batch of 100 examples and a feature called \"weight\" I will create an entry in the feature dictionary that is a tensor of shape (100,1), with key weight, and with all the weights for all the examples, right?<\/p>\n\n<p>However:<\/p>\n\n<ul>\n<li>what if my initial feature is already a tensor, like \"size\" which is a an array of 3 double values? How can I input it via the input_fn ?<\/li>\n<\/ul>\n\n<p>And the question I'm mostly interested in:<\/p>\n\n<ul>\n<li>what if my initial feature is a variable-length array ? For instance my feature could be \"prices of all purchased products\", and it would be a variable-length array of doubles (this correspond to tf.io.VarLenFeature in feature specs). How can I send several examples of this via the input_fn ?<\/li>\n<\/ul>\n\n<p>Are these types of features \"compatible\" with the Estimator API ?<\/p>\n\n<p>thanks!<\/p>\n","answers":[{"AnswerId":"53414822","CreationDate":"2018-11-21T14:59:44.400","ParentId":null,"OwnerUserId":"5623899","Title":null,"Body":"<p>I am also new to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> API, but I have learned quite a lot with the S.O. community and will try to answer your question.<\/p>\n\n<p>First, I would like to point you to this <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a>. This is currently the convention I use for my <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a>s. <\/p>\n\n<p>You are correct in that the <code>input_fn<\/code> for both the <code>TRAIN<\/code> and <code>EVAL<\/code> modes are meant to be tuples in form <code>(features, labels)<\/code>. <\/p>\n\n<p>So let us tackle your first question:<\/p>\n\n<blockquote>\n  <p>what if my initial feature is already a tensor, like \"size\" which is a an array of 3 double values? How can I input it via the input_fn ?<\/p>\n<\/blockquote>\n\n<p>Well this requires me to back track a bit, to your input of:<\/p>\n\n<blockquote>\n  <p>batch of 100 examples and a feature called \"weight\" I will create an entry in the feature dictionary that is a tensor of shape (100,1),<\/p>\n<\/blockquote>\n\n<p>To make sure I understand correctly, you are saying that, what if instead of a <code>Tensor<\/code> with shape <code>[100, 1]<\/code>, you have a <code>Tensor<\/code> or <code>[100, &lt;size&gt;]<\/code>, in this case 3 doubles, so <code>[100, 3]<\/code>?<\/p>\n\n<p>Well if that is the case, that is no problem at all. In the linked <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> a single example of the input has shape <code>[20, 7]<\/code>. So a <code>Tensor<\/code> of <code>[3]<\/code> is straight forward. <\/p>\n\n<p>The short answer is that whatever you specify as the <code>features<\/code> part of the tuple is passed to <code>model_fn<\/code>. So you want to pass a <code>Tensor<\/code> of <code>[batch_size, size]<\/code> you return a tuple of <code>([batch_size, size], labels)<\/code>. However, as another user pointed out to me on S.O. I will impart you with the same advice - use dictionaries e.g.<\/p>\n\n<pre><code>my_data = # Tensor with shape [batch_size, size]\nfeatures = {'my_data': my_data}\n...\nreturn (features, labels)\n<\/code><\/pre>\n\n<p>For reference, let us example the <code>input_fn<\/code> of the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a>, where I do the same things as advised above:<\/p>\n\n<pre><code>def input_fn(filenames:list, params):\n  mode = params['mode'] if 'mode' in params else 'train'\n  batch_size = params['batch_size']\n\n\n  shuffle(filenames) # &lt;--- far more efficient than tf dataset shuffle\n  dataset = tf.data.TFRecordDataset(filenames)\n\n  # using fio's SCHEMA fill the TF Feature placeholders with values\n  dataset = dataset.map(lambda record: fio.from_record(record))\n\n  # using fio's SCHEMA restructure and unwrap (if possible) features (because tf records require wrapping everything into a list)\n  dataset = dataset.map(lambda context, features: fio.reconstitute((context, features)))\n\n  # dataset should be a tuple of (features, labels)\n  dataset = dataset.map(lambda context, features: ( \n      {\"input_tensors\": features[I_FEATURE]}, # features &lt;--- wrapping it in a dictionary\n      features[O_FEATURE]                     # labels\n    )\n  ) \n<\/code><\/pre>\n\n<p>For simplicity, I will assume you are using <code>tf.data.Dataset<\/code>. If your data is not stored as TF <code>Record<\/code>s, you will need to replace  line 1.<\/p>\n\n<pre><code>  1. dataset = tf.data.TFRecordDataset(filenames)\n  2. dataset = dataset.map(lambda record: fio.from_record(record))\n  3. dataset = dataset.map(lambda context, features: fio.reconstitute((context, features)))\n<\/code><\/pre>\n\n<p>with however you construct your dataset, be it <code>FeatureColumn<\/code>, <code>from_tensor_slices<\/code>, etc and remove lines 2 and 3 since you do not need to recover your <code>(Sequence)Example<\/code> from TF <code>Record<\/code>s.<\/p>\n\n<p>Now let us address your second question, variable length arrays.<\/p>\n\n<p>It is just the same as above! wrap it in a dictionary and return it.<\/p>\n\n<p>This is true with the notable exception of recovering your <code>SequenceExample<\/code> from TF <code>Record<\/code>s, where you will need <code>VarLenFeature<\/code><\/p>\n"}]}
+{"QuestionId":53401319,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-20T20:51:03.050","AcceptedAnswerId":53403768.0,"OwnerUserId":5256563.0,"Title":"PyTorch DataLoader and Parallelism","Body":"<p>I have created a class that extends DataSet to load images for a segmentation task, so one input and one output. Every time the method getitem is called, this class performs the necessary operations for data augmentation on both the input and the output, and it works perfectly.<\/p>\n\n<p>However, when I use this class with PyTorch DataLoader, the input transformation do not match with the output transformations. To perform the same operations, I have to get\/set the states of random operations\/classes, and my bet is that the DataLoader does the same, so there is a conflict between them.<\/p>\n\n<p>How can I fix it?<\/p>\n","answers":[{"AnswerId":"53403768","CreationDate":"2018-11-21T00:46:21.827","ParentId":null,"OwnerUserId":"5256563","Title":null,"Body":"<p>The solution is to create a local instance of all the Random classes uses, because the DataLoader does not. Doing that, all the random transformations performed are according to random values\/states that are not affected by the DataLoader. The common way to do it seems to create a class and put all the transformations inside.<\/p>\n"}]}
+{"QuestionId":53401431,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><deep-learning><pytorch>","CreationDate":"2018-11-20T20:58:56.710","AcceptedAnswerId":null,"OwnerUserId":5321842.0,"Title":"Wasserstein GAN critic training ambiguity","Body":"<p>I'm running a DCGAN-based GAN, and am experimenting with WGANs, but am a bit confused about how to train the WGAN.<\/p>\n\n<p>In the official <a href=\"https:\/\/github.com\/martinarjovsky\/WassersteinGAN\/blob\/master\/main.py\" rel=\"nofollow noreferrer\">Wasserstein GAN PyTorch implementation<\/a>, the discriminator\/critic is said to be trained <code>Diters<\/code> (usually 5) times per each generator training.<\/p>\n\n<p>Does this mean that the critic\/discriminator trains on <code>Diters<\/code> <em>batches<\/em> or the <em>whole dataset<\/em> <code>Diters<\/code> times? If I'm not mistaken, the official implementation suggests the discriminator\/critic is trained on the <em>whole dataset<\/em> <code>Diters<\/code> times, but other implementations of WGAN (in PyTorch and TensorFlow etc.) do the opposite.<\/p>\n\n<p>Which is correct? <a href=\"https:\/\/arxiv.org\/pdf\/1701.07875.pdf\" rel=\"nofollow noreferrer\">The WGAN paper<\/a> (to me, at least), indicates that it is <code>Diters<\/code> <em>batches<\/em>. Training on the whole dataset is obviously orders of magnitude slower.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"53403228","CreationDate":"2018-11-20T23:36:46.487","ParentId":null,"OwnerUserId":"10530728","Title":null,"Body":"<p>The correct is to consider an iteration as a batch. \nIn the original <a href=\"https:\/\/arxiv.org\/pdf\/1701.07875.pdf\" rel=\"nofollow noreferrer\">paper<\/a>, for each iteration of the critic\/discriminator they are sampling a batch of size <code>m<\/code> of the real data and  a batch of size <code>m<\/code> of prior samples <code>p(z)<\/code> to work it. After the critic is trained over <code>Diters<\/code> iterations, they train the generator which also starts by the sampling of a batch of prior samples of <code>p(z)<\/code>.\nTherefore, each iteration is working on a batch. <\/p>\n\n<p>In the <a href=\"https:\/\/github.com\/martinarjovsky\/WassersteinGAN\/blob\/master\/main.py\" rel=\"nofollow noreferrer\">official implementation<\/a> this is also happening. What may be confusing is that they use the variable name <code>niter<\/code> to represent the number of epochs to train the model. Although they use a different scheme to set <code>Diters<\/code> at lines <a href=\"https:\/\/github.com\/martinarjovsky\/WassersteinGAN\/blob\/f81eafd2aa41e93698f203732f8f395abc70be02\/main.py#L162\" rel=\"nofollow noreferrer\">162<\/a>-166:<\/p>\n\n<pre><code># train the discriminator Diters times\n    if gen_iterations &lt; 25 or gen_iterations % 500 == 0:\n        Diters = 100\n    else:\n        Diters = opt.Diters\n<\/code><\/pre>\n\n<p>they are, as in the paper, training the critic over <code>Diters<\/code> batches.<\/p>\n"}]}
+{"QuestionId":53401636,"AnswerCount":0,"Tags":"<python><ios><xcode><tensorflow>","CreationDate":"2018-11-20T21:14:49.820","AcceptedAnswerId":null,"OwnerUserId":9643695.0,"Title":"Tensorflow Demo iOS App with different models not working","Body":"<p><strong>System information<\/strong><\/p>\n\n<p><p>OS Platform and Distribution: macOS Mojave 10.14.2 Beta\n<p>Mobile device : iPhone 8P \/ IPhone XR\n<p>TensorFlow installed from (source or binary): SOURCE\n<p>TensorFlow version: Newest\n<p>Python version: 3.7\n<p>Installed using virtualenv? pip? conda?: Unknown\n<p>Bazel version (if compiling from source): Unknown\n<p>GCC\/Compiler version (if compiling from source): Unknown\n<p>CUDA\/cuDNN version:\n<p>GPU model and memory:\n<p><strong>Describe the problem<\/strong>\n<p>Xcode Crashed when I import my models into the TF-Lite example project. I don't know what to do or try to get the models to work. <\/p>\n\n<p><strong>Provide the exact sequence of commands \/ steps that you executed before running into the problem<\/strong><\/p>\n\n<p><p>1.Import the TFlite and txt files\n<p>2.copy the path if needle ( Xcode asks this )\n<p>3.rename model_file_name to imported tflite model\n<p>4.rename label_file_name to imported txt file\n<p><strong>Any other info \/ logs<\/strong>\n<p>Loaded model 1resolved reportertensorflow\/contrib\/lite\/kernels\/squeeze.cc:64 current >= 0 &amp;&amp; current &lt; input_num_dims &amp;&amp; input_dims->data[current] == 1 was not true.\nNode 66 failed to prepare.<\/p>\n\n<p>Failed to allocate tensors!2018-11-20 13:08:26.664470-0600 tflite_camera_example[9610:2085000] [MC] System group container for systemgroup.com.apple.configurationprofiles path is \/private\/var\/containers\/Shared\/SystemGroup\/systemgroup.com.apple.configurationprofiles\n2018-11-20 13:08:26.664721-0600 tflite_camera_example[9610:2085000] [MC] Reading from public effective user settings.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/75vFt.png\" rel=\"nofollow noreferrer\">Xcode Debugger output<\/a><\/p>\n","answers":[]}
+{"QuestionId":53401829,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-20T21:28:57.757","AcceptedAnswerId":null,"OwnerUserId":10560880.0,"Title":"Decoding Training Examples","Body":"<p>I am using a ddsm dataset to make a CNN, and the owner of the dataset gave some code to decode the training examples, but I can't figure out what first part of the code is actually doing. Still learning. \nFor reference: <a href=\"https:\/\/www.kaggle.com\/skooch\/ddsm-mammography\/home\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/skooch\/ddsm-mammography\/home<\/a><\/p>\n\n<p>The code:  features = tf.parse_single_example(\n        serialized_example,\n        features={\n            'label': tf.FixedLenFeature([], tf.int64),\n            'label_normal': tf.FixedLenFeature([], tf.int64),\n            'image': tf.FixedLenFeature([], tf.string)\n        })<\/p>\n","answers":[]}
+{"QuestionId":53402248,"AnswerCount":1,"Tags":"<javascript><tensorflow>","CreationDate":"2018-11-20T22:01:29.437","AcceptedAnswerId":53859638.0,"OwnerUserId":175242.0,"Title":"issue with multi-person tracking with PoseNet using JavaScript & TensorFlow","Body":"<p>We are using an issue with multi-person tracking with the PoseNet library for Javascript.   We are seeing the data getting associated with the wrong person sometimes. <\/p>\n\n<p>person 1<\/p>\n\n<ul>\n<li>tracking data<\/li>\n<\/ul>\n\n<p>person 2<\/p>\n\n<ul>\n<li>tracking data<\/li>\n<\/ul>\n\n<p>some time later:<\/p>\n\n<p>person 1<\/p>\n\n<ul>\n<li>tracking data of person 2<\/li>\n<\/ul>\n\n<p>person 2<\/p>\n\n<ul>\n<li>tracking data of person 1 <\/li>\n<\/ul>\n\n<p>We can add some code in Javascript to check the validity of the data.  Is this the way multi-person tracking works in PoseNet or a bug?<\/p>\n","answers":[{"AnswerId":"53859638","CreationDate":"2018-12-19T21:50:22.437","ParentId":null,"OwnerUserId":"1473227","Title":null,"Body":"<p>Posenet doesn't track people that way. It just looks for people and creates an array in somewhat random order. In other words, it starts over with every frame, with no knowledge of the previous frames. You would have to write an algorithm of your own.<\/p>\n\n<p>My suggestion would be to look at the points in the face in each frame and compare them to the points from the previous frame. A naive approach would be to simply find the face in the previous frame that is closest to the face in the current frame. If you needed to do this in a more sophisticated way, you would need do blob tracking where you are actually looking at the colors\/composition of each person and making sure they match across frames.<\/p>\n"}]}
+{"QuestionId":53402303,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><image-segmentation>","CreationDate":"2018-11-20T22:06:06.057","AcceptedAnswerId":null,"OwnerUserId":5826676.0,"Title":"How do I need to configure Keras model to predict an image?","Body":"<p>The main task is to predict a mask for the input image. So I have the following data for training:<\/p>\n\n<ul>\n<li>lot's of 768x768 original pics like this:<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Sm91L.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Sm91L.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<ul>\n<li>and output mask pics(also 768x768) like this:<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4NoWW.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4NoWW.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also I have validation original pics.<\/p>\n\n<p>I prepare some kind of neural model that should predict the output mask. I prepared <a href=\"https:\/\/gist.github.com\/keipa\/54ba9fd1c39ea6d2810dbf836c98ce25\" rel=\"nofollow noreferrer\">keras model configuaration<\/a> that should have a topology which looks like below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/931I6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/931I6.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The code I prepared for training is there. <\/p>\n\n<pre><code>import keras\nepochs=100\n\nimage_datagen = keras.preprocessing.image.ImageDataGenerator()\nmask_datagen = keras.preprocessing.image.ImageDataGenerator()\nseed = 1\nimage_generator = image_datagen.flow_from_directory(\n    'H:\/LABS\/ship_detection\/test_train\/',\n    color_mode='rgb',batch_size=32,target_size=(768,768),\n    seed=seed)\n\nmask_generator = mask_datagen.flow_from_directory(\n    'H:\/LABS\/ship_detection\/test_mask\/',\n    class_mode=\"categorical\",batch_size=32,target_size=(768,768),\n    seed=seed)\n\ntrain_generator = zip(image_generator, mask_generator)\n\nmodel.fit_generator(generator=train_generator,\n                    epochs=epochs,\n                    callbacks=callbacks,steps_per_epoch=1)\n<\/code><\/pre>\n\n<p>But when I try to fit generator for prediction I have an issue:<\/p>\n\n<pre><code>c:\\users\\harwister\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\keras\\engine\\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n    208                     batch_size = list(x.values())[0].shape[0]\n    209                 else:\n--&gt; 210                     batch_size = x.shape[0]\n    211                 batch_logs['batch'] = batch_index\n    212                 batch_logs['size'] = batch_size\n\nAttributeError: 'tuple' object has no attribute 'shape'\n<\/code><\/pre>\n\n<p>I do something wrong for sure, but I can't understand anything from these kind of errors. The simple question I can't find a response in Google is: How can I push into Keras two images (input and output images) for training and after training get an output image providing an input image?<\/p>\n","answers":[{"AnswerId":"53406984","CreationDate":"2018-11-21T07:15:48.340","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since you have separate generators for the images and the labels (i.e. masks), you need to set the <code>class_mode<\/code> argument to <code>None<\/code> to prevent the generators from producing any labels arrays:<\/p>\n\n<pre><code>image_generator = image_datagen.flow_from_directory(class_mode=None, ...)\nmask_generator = mask_datagen.flow_from_directory(class_mode=None, ...)\n<\/code><\/pre>\n\n<p>This way, <code>image_generator<\/code> would only generate the input images and the <code>mask_generator<\/code> would only generate the mask (i.e. true label) images.<\/p>\n"}]}
+{"QuestionId":53402928,"AnswerCount":1,"Tags":"<deep-learning><pytorch><fast-ai>","CreationDate":"2018-11-20T23:04:11.660","AcceptedAnswerId":53698454.0,"OwnerUserId":497600.0,"Title":"list of fastai models","Body":"<p>FastAI uses AWD-LSTM for text processing. They provide pretrained models with <a href=\"https:\/\/docs.fast.ai\/text.models.html#get_language_model\" rel=\"nofollow noreferrer\"><code>get_language_model()<\/code><\/a>. But I can't find proper documentation on what's available.<\/p>\n\n<p>Their <a href=\"https:\/\/github.com\/fastai\/fastai\/blob\/master\/examples\/text.ipynb\" rel=\"nofollow noreferrer\">github example page<\/a> is really a moving target. Model names such as <code>lstm_wt103<\/code> and <code>WT103_1<\/code> are used. In the forums I found <code>wt103RNN<\/code>.<\/p>\n\n<p>Where can I find an updated list of pretrained models and their download URLs ?<\/p>\n","answers":[{"AnswerId":"53698454","CreationDate":"2018-12-10T01:36:40.167","ParentId":null,"OwnerUserId":"1235433","Title":null,"Body":"<p>URLs is defined in <a href=\"https:\/\/github.com\/fastai\/fastai\/blob\/release-1.0.36\/fastai\/datasets.py#L7\" rel=\"nofollow noreferrer\">fastai.datasets<\/a>, there are constants for two models: WT103, WT103_1.<\/p>\n\n<p><a href=\"https:\/\/s3.amazonaws.com\/fast-ai-modelzoo\" rel=\"nofollow noreferrer\">AWS bucket<\/a> has just the two models.<\/p>\n"}]}
+{"QuestionId":53403306,"AnswerCount":2,"Tags":"<nlp><pytorch>","CreationDate":"2018-11-20T23:47:12.617","AcceptedAnswerId":53403392.0,"OwnerUserId":4542336.0,"Title":"How to batch convert sentence lengths to masks in PyTorch?","Body":"<p>For example, from<\/p>\n\n<pre><code>lens = [3, 5, 4]\n<\/code><\/pre>\n\n<p>we want to get<\/p>\n\n<pre><code>mask = [[1, 1, 1, 0, 0],\n        [1, 1, 1, 1, 1],\n        [1, 1, 1, 1, 0]]\n<\/code><\/pre>\n\n<p>Both of which are <code>torch.LongTensor<\/code>s.<\/p>\n","answers":[{"AnswerId":"53403392","CreationDate":"2018-11-20T23:58:04.453","ParentId":null,"OwnerUserId":"4542336","Title":null,"Body":"<p>One way that I found is:<\/p>\n\n<pre><code>torch.arange(max_len).expand(len(lens), max_len) &lt; lens.unsqueeze(1)\n<\/code><\/pre>\n\n<p>Please share if there are better ways!<\/p>\n"},{"AnswerId":"56573638","CreationDate":"2019-06-13T05:02:52.163","ParentId":null,"OwnerUserId":"4433431","Title":null,"Body":"<p><code>torch.arange(max_len)[None, :] &lt; lens[:, None]<\/code><\/p>\n"}]}
+{"QuestionId":53403311,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><reinforcement-learning><tensorflow-estimator>","CreationDate":"2018-11-20T23:47:30.130","AcceptedAnswerId":53432579.0,"OwnerUserId":1171204.0,"Title":"tensorflow - implementing experience replay memory with the estimator api","Body":"<p>I try to implement an <a href=\"https:\/\/arxiv.org\/pdf\/1602.01783.pdf\" rel=\"nofollow noreferrer\">experience replay memory<\/a> with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\">tf.estimator.Estimator<\/a> API. However, i am not sure what is the best way to achieve a result which works atleast for all modes (<code>TRAIN<\/code>, <code>EVALUATE<\/code>, <code>PREDICT<\/code>). I tried the following:<\/p>\n\n<ul>\n<li>Implement the memory with a <code>tf.Variable<\/code>, which causes issues with the batching and the input pipeline (i cannot input a custom experience in testing or prediction phase)<\/li>\n<\/ul>\n\n<p>and currently try to:<\/p>\n\n<ul>\n<li>Implement the memory outside the <code>tf.Graph<\/code>. Set the values after each run with a <code>tf.train.SessionRunHook<\/code>. Load the experiences with <code>tf.data.Dataset.from_generator()<\/code> during training and testing. Manage the state on your own.<\/li>\n<\/ul>\n\n<p>I am failing on several points and starting to believe that the tf.estimator.Estimator API does not provide me with the necessary interfaces to easily write this down.<\/p>\n\n<p>Some code (first approach, which fails with the batch_size, since it is fixed for the slicing of the exp, i cannot use the model for prediction or evaluation):<\/p>\n\n<pre><code> def model_fn(self, features, labels, mode, params):\n    batch_size = features[\"matrix\"].get_shape()[0].value\n\n    # get prev_exp\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        erm = tf.get_variable(\"erm\", shape=[30000, 10], initializer=tf.constant_initializer(self.erm.initial_train_erm()), trainable=False)\n        prev_exp = tf.slice(erm, [features[\"index\"][0], 0], [batch_size, 10])\n\n    # model\n    pred = model(features[\"matrix\"], prev_exp, params) \n<\/code><\/pre>\n\n<p>However: it would be better to have the erm inside the feature dict. But then i have to manage the erm outside the graph and also write back my newest experience with a SessionRunHook. Is there any better way or am i missing something?<\/p>\n","answers":[{"AnswerId":"53432579","CreationDate":"2018-11-22T13:57:28.203","ParentId":null,"OwnerUserId":"1171204","Title":null,"Body":"<p>I solved my issue by implementing the ERM outside the graph, feeding it back into the input pipeline with tf.data.Dataset.from_generator() and writing back by using SessionRunHooks. Yeah, pretty tedious but it is working.<\/p>\n"}]}
+{"QuestionId":53403445,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2018-11-21T00:03:05.613","AcceptedAnswerId":53408188.0,"OwnerUserId":8195828.0,"Title":"Cannot implement multiple stacked bidirectional RNNs","Body":"<p>I am trying to implement a Seq2Seq variant in Tensorflow, which includes two encoders and a decoder. For the encoders' first layer, I have bidirectional LSTMs. So I have implemented this method for getting bidirectional LSTMs for variable number of layers:<\/p>\n\n<pre><code>def bidirectional_lstm(batch, num_layers=2, hidden_layer=256):\n\n\nforward_lstms=[LSTMCell(num_units=hidden_layer\/2) for _ in range(num_layers)]\nbackward_lstms=[LSTMCell(num_units=hidden_layer\/2) for _ in range(num_layers)]\n\nstates_fw=[f_l.zero_state(BATCH_SIZE, tf.float64) for f_l in forward_lstms]\nstates_bw=[b_l.zero_state(BATCH_SIZE, tf.float64) for b_l in backward_lstms]\n\n\noutputs, final_state_fw, final_state_bw=tf.contrib.rnn.stack_bidirectional_dynamic_rnn(\n    forward_lstms,\n    backward_lstms,\n    batch,\n    initial_states_fw=states_fw,\n    initial_states_bw=states_bw,\n    parallel_iterations=32\n)\n\nreturn outputs\n<\/code><\/pre>\n\n<p>But when I run the lines below:<\/p>\n\n<pre><code>a=bidirectional_lstm(a_placeholder)\n\nb=bidirectional_lstm(b_placeholder, num_layers=1)\n<\/code><\/pre>\n\n<p>I get this error message:<\/p>\n\n<pre><code>ValueError\nVariable \nstack_bidirectional_rnn\/cell_0\/bidirectional_rnn\/fw\/lstm_cell\/kernel \nalready exists, disallowed. Did you mean to set reuse=True or \nreuse=tf.AUTO_REUSE in VarScope? Originally defined at:    File \n\"\/usr\/local\/lib\/python3.6\/dist- \npackages\/tensorflow\/contrib\/rnn\/python\/ops\/rnn.py\", line 233, in \nstack_bidirectional_dynamic_rnn     time_major=time_major)\n<\/code><\/pre>\n\n<p>I do not want to \"reuse\" a given stacked bidirectional LSTM. How can I run two separate encoders containing two stacked bidirectional LSTMs? <\/p>\n","answers":[{"AnswerId":"53408188","CreationDate":"2018-11-21T08:45:44.717","ParentId":null,"OwnerUserId":"8195828","Title":null,"Body":"<p>Figured it out: The two encoders need to \"run\" in two different variable scopes to avoid \"mixup\" during gradient updates<\/p>\n\n<pre><code>with tf.variable_scope(\"a\"):\n    a=bidirectional_lstm(a_placeholder)\n\n\nwith tf.variable_scope(\"b\"):\n    b=bidirectional_lstm(b_placeholder)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53403511,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T00:12:13.083","AcceptedAnswerId":53405366.0,"OwnerUserId":3259896.0,"Title":"Tensorflow Custom Estimator: How to implement a `input_fn` function that returns a list of labels, and a list of features ?","Body":"<p>I am trying to convert my Tensorflow graph to use a custom tensorflow estimator, but I am getting stuck defining the function for <code>input_fn<\/code> ; I am currently getting an error. <\/p>\n\n<p>This is the function I use to generate my input data and labels<\/p>\n\n<pre><code>data_index = 0 \nepoch_index = 0 \nrecEpoch_indexA = 0 #Used to help keep store of the total number of epoches with the models\n\ndef generate_batch(batch_size, inputCount): \n    global data_index, epoch_index\n\n    batch = np.ndarray(shape=(batch_size, inputCount), dtype=np.int32) \n    labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)\n\n    n=0\n    while n &lt; batch_size:\n      if len(    set(my_data[data_index, 1])   ) &gt;= inputCount:\n        labels[n,0] = my_data[data_index, 0]\n        batch[n] = random.sample( set(my_data[data_index, 1]),  inputCount)\n        n = n+1\n        data_index = (data_index + 1) % len(my_data) #may have to do something like len my_data[:]\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n      else:\n        data_index = (data_index + 1) % len(my_data)\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n\n    return batch, labels     \n<\/code><\/pre>\n\n<p>This is where I define my Estimator and attempt to do training<\/p>\n\n<pre><code>#Define the estimator\nword2vecEstimator = tf.estimator.Estimator(\n        model_fn=my_model,\n        params={\n            'batch_size': 1024,\n            'embedding_size': 50,\n            'num_inputs': 5,\n            'num_sampled':128\n        })\n\nbatch_size = 16\nnum_inputs = 3\n\n#Train with Estimator\nword2vecEstimator.train(\n    input_fn=generate_batch(batch_size, num_inputs),\n    steps=10)\n<\/code><\/pre>\n\n<p>This is the error message that I get<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n\/usr\/lib\/python3.6\/inspect.py in getfullargspec(func)\n   1118                                        skip_bound_arg=False,\n-&gt; 1119                                        sigcls=Signature)\n   1120     except Exception as ex:\n\n\/usr\/lib\/python3.6\/inspect.py in _signature_from_callable(obj, follow_wrapper_chains, skip_bound_arg, sigcls)\n   2185     if not callable(obj):\n-&gt; 2186         raise TypeError('{!r} is not a callable object'.format(obj))\n   2187 \n\nTypeError: (array([[1851833,  670357,  343012],\n       [ 993526,  431296,  935528],\n       [ 938067, 1155719, 2277388],\n       [ 534965, 1125669, 1665716],\n       [1412657, 2152211, 1176177],\n       [ 268114, 2097642, 2707258],\n       [1280762, 1516464,  453615],\n       [2545980, 2302607, 2421182],\n       [1706260, 2735027,  292652],\n       [1802025, 2949676,  653015],\n       [ 854228, 2626773,  225486],\n       [1747135, 1608478, 2503487],\n       [1326661,  272883, 2089444],\n       [3082922, 1359481,  621031],\n       [2636832, 1842777, 1979638],\n       [2512269, 1617986,  389356]], dtype=int32), array([[1175598],\n       [2528125],\n       [1870906],\n       [ 643521],\n       [2349752],\n       [ 754986],\n       [2277570],\n       [2121120],\n       [2384306],\n       [1881398],\n       [3046987],\n       [2505729],\n       [2908573],\n       [2438025],\n       [ 441422],\n       [2355625]], dtype=int32)) is not a callable object\n\nThe above exception was the direct cause of the following exception:\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-15-7acc939af001&gt; in &lt;module&gt;()\n      5 word2vecEstimator.train(\n      6     input_fn=generate_batch(batch_size, num_inputs),\n----&gt; 7     steps=10)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in train(self, input_fn, hooks, steps, max_steps, saving_listeners)\n    352 \n    353       saving_listeners = _check_listeners_type(saving_listeners)\n--&gt; 354       loss = self._train_model(input_fn, hooks, saving_listeners)\n    355       logging.info('Loss for final step: %s.', loss)\n    356       return self\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _train_model(self, input_fn, hooks, saving_listeners)\n   1205       return self._train_model_distributed(input_fn, hooks, saving_listeners)\n   1206     else:\n-&gt; 1207       return self._train_model_default(input_fn, hooks, saving_listeners)\n   1208 \n   1209   def _train_model_default(self, input_fn, hooks, saving_listeners):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _train_model_default(self, input_fn, hooks, saving_listeners)\n   1232       features, labels, input_hooks = (\n   1233           self._get_features_and_labels_from_input_fn(\n-&gt; 1234               input_fn, model_fn_lib.ModeKeys.TRAIN))\n   1235       worker_hooks.extend(input_hooks)\n   1236       estimator_spec = self._call_model_fn(\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _get_features_and_labels_from_input_fn(self, input_fn, mode)\n   1073     \"\"\"Extracts the `features` and labels from return values of `input_fn`.\"\"\"\n   1074     return estimator_util.parse_input_fn_result(\n-&gt; 1075         self._call_input_fn(input_fn, mode))\n   1076 \n   1077   def _extract_batch_length(self, preds_evaluated):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _call_input_fn(self, input_fn, mode)\n   1151       ValueError: if `input_fn` takes invalid arguments.\n   1152     \"\"\"\n-&gt; 1153     input_fn_args = function_utils.fn_args(input_fn)\n   1154     kwargs = {}\n   1155     if 'mode' in input_fn_args:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/function_utils.py in fn_args(fn)\n     54     if _is_callable_object(fn):\n     55       fn = fn.__call__\n---&gt; 56     args = tf_inspect.getfullargspec(fn).args\n     57     if _is_bounded_method(fn):\n     58       args.remove('self')\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/tf_inspect.py in getfullargspec(obj)\n    214   return next((d.decorator_argspec\n    215                for d in decorators\n--&gt; 216                if d.decorator_argspec is not None), _getfullargspec(target))\n    217 \n    218 \n\n\/usr\/lib\/python3.6\/inspect.py in getfullargspec(func)\n   1123         # else. So to be fully backwards compatible, we catch all\n   1124         # possible exceptions here, and reraise a TypeError.\n-&gt; 1125         raise TypeError('unsupported callable') from ex\n   1126 \n   1127     args = []\n\nTypeError: unsupported callable\n<\/code><\/pre>\n\n<p>Here is a link to the Google Colab notebook for people to run on their own. For anyone looking to execute this, this will download a data file that is ~500 mbs. <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1LjIz04xhRi5Fsw_Q3IzoG_5KkkXI3WFE\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1LjIz04xhRi5Fsw_Q3IzoG_5KkkXI3WFE<\/a><\/p>\n\n<p>And here is the full code, from the notebook. <\/p>\n\n<pre><code>import math\nimport numpy as np\nimport random\nimport zipfile\nimport shutil\nfrom collections import namedtuple\n\nimport os\nimport pprint\n\nimport tensorflow as tf\n\nimport pandas as pd\nimport pickle\nfrom numpy import genfromtxt\n\n!pip install -U -q PyDrive\n\nfrom google.colab import files\nfrom pydrive.auth import GoogleAuth\nfrom pydrive.drive import GoogleDrive\nfrom google.colab import auth\nfrom oauth2client.client import GoogleCredentials\n\nauth.authenticate_user()\ngauth = GoogleAuth()\ngauth.credentials = GoogleCredentials.get_application_default()\ndrive = GoogleDrive(gauth)\n\nvocabulary_size = 3096637 #updated 10-25-18 3096636\n\nimport gc\n\ndl_id = '19yha9Scxq4zOdfPcw5s6L2lkYQWenApC' #updated 10-22-18\n\nmyDownload = drive.CreateFile({'id': dl_id})\nmyDownload.GetContentFile('Data.npy')\nmy_data = np.load('Data.npy')\n#os.remove('Data.npy')\nnp.random.shuffle(my_data)\nprint(my_data[0:15])\n\ndata_index = 0 \nepoch_index = 0 \nrecEpoch_indexA = 0 #Used to help keep store of the total number of epoches with the models\n\ndef generate_batch(batch_size, inputCount): \n    global data_index, epoch_index\n\n    batch = np.ndarray(shape=(batch_size, inputCount), dtype=np.int32) \n    labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)\n\n    n=0\n    while n &lt; batch_size:\n      if len(    set(my_data[data_index, 1])   ) &gt;= inputCount:\n        labels[n,0] = my_data[data_index, 0]\n        batch[n] = random.sample( set(my_data[data_index, 1]),  inputCount)\n        n = n+1\n        data_index = (data_index + 1) % len(my_data) #may have to do something like len my_data[:]\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n      else:\n        data_index = (data_index + 1) % len(my_data)\n        if data_index == 0:\n          epoch_index = epoch_index + 1\n          print('Completed %d Epochs' % epoch_index)\n\n    return batch, labels     \n\ndef my_model( features, labels, mode, params):\n\n#     train_dataset = tf.placeholder(tf.int32, shape=[batch_size, num_inputs ])\n#     train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])\n\n    train_dataset = features\n    train_labels = labels\n\n    batch_sizeE=params[\"batch_size\"]\n    embedding_sizeE=params[\"embedding_size\"]\n    num_inputsE=params[\"num_inputs\"]\n    num_sampledE=params[\"num_sampled\"]\n\n    epochCount = tf.get_variable( 'epochCount', initializer= 0) #to store epoch count to total # of epochs are known\n    update_epoch = tf.assign(epochCount, epochCount + 1)\n\n    embeddings = tf.get_variable( 'embeddings', dtype=tf.float32,\n        initializer= tf.random_uniform([vocabulary_size, embedding_sizeE], -1.0, 1.0, dtype=tf.float32) )\n\n    softmax_weights = tf.get_variable( 'softmax_weights', dtype=tf.float32,\n        initializer= tf.truncated_normal([vocabulary_size, embedding_sizeE],\n                             stddev=1.0 \/ math.sqrt(embedding_sizeE), dtype=tf.float32 ) )\n\n    softmax_biases = tf.get_variable('softmax_biases', dtype=tf.float32,\n        initializer= tf.zeros([vocabulary_size], dtype=tf.float32),  trainable=False )\n\n    embed = tf.nn.embedding_lookup(embeddings, train_dataset) #train data set is\n\n    embed_reshaped = tf.reshape( embed, [batch_sizeE*num_inputs, embedding_sizeE] )\n\n    segments= np.arange(batch_size).repeat(num_inputs)\n\n    averaged_embeds = tf.segment_mean(embed_reshaped, segments, name=None)\n\n    loss = tf.reduce_mean(\n        tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=averaged_embeds, \n                                   sampled_values=tf.nn.uniform_candidate_sampler(true_classes=tf.cast(train_labels, tf.int64), num_sampled=num_sampled, num_true=1, unique=True, range_max=vocabulary_size, seed=None),\n                                   labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size)) \n\n    optimizer = tf.train.AdagradOptimizer(1.0).minimize(loss) \n\n    saver = tf.train.Saver()\n\n#Define the estimator\nword2vecEstimator = tf.estimator.Estimator(\n        model_fn=my_model,\n        params={\n            'batch_size': 1024,\n            'embedding_size': 50,\n            'num_inputs': 5,\n            'num_sampled':128\n        })\n\nbatch_size = 16\nnum_inputs = 3\n\n#Train with Estimator\nword2vecEstimator.train(\n    input_fn=generate_batch(batch_size, num_inputs),\n    steps=10)\n<\/code><\/pre>\n","answers":[{"AnswerId":"56238126","CreationDate":"2019-05-21T12:18:20.913","ParentId":null,"OwnerUserId":"9762153","Title":null,"Body":"<p>There's no way of correcting the function, cuz it can never be implemented using Tensorflow. The input_fn() function must return Tensors, not numpy arrays, cuz the input_fn() is a function constructing the graph, and it maybe is just called once when building the graph. In this context, the numpy array is just a constant value. It may seem to be weird, But it's the truth. You need to understand the mechanism of Tensorflow: the STATIC COMPUTE GRAPH!<\/p>\n"},{"AnswerId":"53405366","CreationDate":"2018-11-21T04:40:36.403","ParentId":null,"OwnerUserId":"3259896","Title":null,"Body":"<p>Answer here<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/49140164\/tensorflow-error-unsupported-callable?rq=1\">Tensorflow error : unsupported callable<\/a><\/p>\n\n<blockquote>\n  <p>train method accepts the input function, so it should be input_fn, not input_fn().<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53403722,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><keras-rl>","CreationDate":"2018-11-21T00:39:38.227","AcceptedAnswerId":null,"OwnerUserId":7411022.0,"Title":"Scale actor network output to the action space bounds in Keras Rl","Body":"<p>I am trying to implement DDPG from Keras RL and have the following actor network.<\/p>\n\n<pre><code>actor = Sequential()\nactor.add(Flatten(input_shape=(1,) + env.observation_space.shape))\nactor.add(Dense(16))\nactor.add(Activation('relu'))\nactor.add(Dense(16))\nactor.add(Activation('relu'))\nactor.add(Dense(16))\nactor.add(Activation('relu'))\nactor.add(Dense(nb_actions))\nactor.add(Activation('linear'))\n<\/code><\/pre>\n\n<p>However, I would prefer to have the output scaled to a custom gym environment action space bounds for my problem. <code>env.action_space<\/code>. <\/p>\n\n<p><a href=\"https:\/\/pemami4911.github.io\/blog\/2016\/08\/21\/ddpg-rl.html\" rel=\"nofollow noreferrer\">https:\/\/pemami4911.github.io\/blog\/2016\/08\/21\/ddpg-rl.html<\/a> shows this using the tflearn api where they use <\/p>\n\n<pre><code>def create_actor_network(self):\n        inputs = tflearn.input_data(shape=[None, self.s_dim])\n        net = tflearn.fully_connected(inputs, 400)\n        net = tflearn.layers.normalization.batch_normalization(net)\n        net = tflearn.activations.relu(net)\n        net = tflearn.fully_connected(net, 300)\n        net = tflearn.layers.normalization.batch_normalization(net)\n        net = tflearn.activations.relu(net)\n        # Final layer weights are init to Uniform[-3e-3, 3e-3]\n        w_init = tflearn.initializations.uniform(minval=-0.003, maxval=0.003)\n        out = tflearn.fully_connected(\n            net, self.a_dim, activation='tanh', weights_init=w_init)\n        # Scale output to -action_bound to action_bound\n        scaled_out = tf.multiply(out, self.action_bound)\n        return inputs, out, scaled_out\n<\/code><\/pre>\n\n<p>What is the equivalent command for scaling the output layer according to my requirements?<\/p>\n","answers":[]}
+{"QuestionId":53403812,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2018-11-21T00:52:53.553","AcceptedAnswerId":54774006.0,"OwnerUserId":5116657.0,"Title":"Using nn.ModuleList over Python list dramatically slows down training","Body":"<p>I'm training a very simple model that takes the number of hidden layers as a parameter. I originally stored these hidden layers in a vanilla python list <code>[]<\/code>, however when converting this list to a <code>nn.ModuleList<\/code>, training slows down dramatically by <strong>at least one order of magnitude<\/strong>!<\/p>\n\n<p><strong>AdderNet<\/strong><\/p>\n\n<pre><code>class AdderNet(nn.Module):\n    def __init__(self, num_hidden, hidden_width):\n        super(AdderNet, self).__init__()\n        self.relu = nn.ReLU()\n\n        self.hiddenLayers = []\n        self.inputLayer = nn.Linear(2, hidden_width)\n        self.outputLayer = nn.Linear(hidden_width, 1)\n\n        for i in range(num_hidden):\n            self.hiddenLayers.append(nn.Linear(hidden_width, hidden_width))\n\n        self.hiddenLayers = nn.ModuleList(self.hiddenLayers)  # &lt;--- causes DRAMATIC slowdown!\n\n    def forward(self, x):\n        out = self.inputLayer(x)\n        out = self.relu(out)\n\n        for layer in self.hiddenLayers:\n            out = layer(out)\n            out = self.relu(out)\n\n        return self.outputLayer(out)\n<\/code><\/pre>\n\n<p><strong>Training<\/strong><\/p>\n\n<pre><code>for epoch in range(num_epochs):\n    for i in range(0,len(data)):\n        out = model.forward(data[i].x)\n        loss = lossFunction(out, data[i].y)\n\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n<\/code><\/pre>\n","answers":[{"AnswerId":"54774006","CreationDate":"2019-02-19T19:56:24.283","ParentId":null,"OwnerUserId":"11086527","Title":null,"Body":"<p>That's because when using a normal python list, the parameters are not added to the model's parameter list, but when using a ModuleList, they are. So, in the original scenario, you were never actually training the hidden layers, which is why it was faster. (Print out model.parameters() in each case and see what happens!) <\/p>\n"}]}
+{"QuestionId":53404301,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><metrics>","CreationDate":"2018-11-21T01:58:34.847","AcceptedAnswerId":null,"OwnerUserId":9494790.0,"Title":"How to compute Spearman correlation in Tensorflow","Body":"<h1>Problem<\/h1>\n\n<p>I need to compute the Pearson and Spearman correlations, and use it as metrics in tensorflow.<\/p>\n\n<p>For Pearson, it's trivial :<\/p>\n\n<pre><code>tf.contrib.metrics.streaming_pearson_correlation(y_pred, y_true)\n<\/code><\/pre>\n\n<p>But for Spearman, I am clueless !<\/p>\n\n<h1>What I tried :<\/h1>\n\n<p>From <a href=\"https:\/\/stackoverflow.com\/a\/42730743\/9494790\">this answer<\/a> :<\/p>\n\n<pre><code>    samples = 1\n    predictions_rank = tf.nn.top_k(y_pred, k=samples, sorted=True, name='prediction_rank').indices\n    real_rank = tf.nn.top_k(y_true, k=samples, sorted=True, name='real_rank').indices\n    rank_diffs = predictions_rank - real_rank\n    rank_diffs_squared_sum = tf.reduce_sum(rank_diffs * rank_diffs)\n    six = tf.constant(6)\n    one = tf.constant(1.0)\n    numerator = tf.cast(six * rank_diffs_squared_sum, dtype=tf.float32)\n    divider = tf.cast(samples * samples * samples - samples, dtype=tf.float32)\n    spearman_batch = one - numerator \/ divider\n<\/code><\/pre>\n\n<p>But this return <code>NaN<\/code>...<\/p>\n\n<hr>\n\n<p>Following the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Spearman%27s_rank_correlation_coefficient#Definition_and_calculation\" rel=\"nofollow noreferrer\">definition of Wikipedia<\/a> :\n <a href=\"https:\/\/i.stack.imgur.com\/M8RQM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/M8RQM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I tried : <\/p>\n\n<pre><code>size = tf.size(y_pred)\nindice_of_ranks_pred = tf.nn.top_k(y_pred, k=size)[1]\nindice_of_ranks_label = tf.nn.top_k(y_true, k=size)[1]\nrank_pred = tf.nn.top_k(-indice_of_ranks_pred, k=size)[1]\nrank_label = tf.nn.top_k(-indice_of_ranks_label, k=size)[1]\nrank_pred = tf.to_float(rank_pred)\nrank_label = tf.to_float(rank_label)\nspearman = tf.contrib.metrics.streaming_pearson_correlation(rank_pred, rank_label)\n<\/code><\/pre>\n\n<p>But running this I got the following error :<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: input\n  must have at least k columns. Had 1, needed 32<\/p>\n  \n  <p>[[{{node metrics\/spearman\/TopKV2}} = TopKV2[T=DT_FLOAT, sorted=true,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](lambda_1\/add, metrics\/pearson\/pearson_r\/variance_predictions\/Size)]]<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"55651603","CreationDate":"2019-04-12T12:25:29.390","ParentId":null,"OwnerUserId":"11351522","Title":null,"Body":"<p>One thing you can do is use Tensorflow's function <code>tf.py_function<\/code> to use with the \n<code>scipy.stats.spearmanr<\/code> and define the input and output like that:<\/p>\n\n<pre><code>from scipy.stats import spearmanr\ndef get_spearman_rankcor(y_true, y_pred):\n     return ( tf.py_function(spearmanr, [tf.cast(y_pred, tf.float32), \n                       tf.cast(y_true, tf.float32)], Tout = tf.float32) )\n<\/code><\/pre>\n"}]}
+{"QuestionId":53404314,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><regression>","CreationDate":"2018-11-21T02:00:49.133","AcceptedAnswerId":null,"OwnerUserId":10520937.0,"Title":"How would I code a polynomial regression using tensorflow and python","Body":"<p>How would I do this in python? I also want to predict it using multiple variables.<\/p>\n","answers":[]}
+{"QuestionId":53404367,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><scipy><metrics>","CreationDate":"2018-11-21T02:08:15.780","AcceptedAnswerId":53412330.0,"OwnerUserId":9494790.0,"Title":"Why Pearson correlation is different between Tensorflow and Scipy","Body":"<p>I compute the Pearson correlation in 2 ways :<\/p>\n\n<hr>\n\n<p>In Tensorflow, I use the following metric :<\/p>\n\n<pre><code>tf.contrib.metrics.streaming_pearson_correlation(y_pred, y_true)\n<\/code><\/pre>\n\n<p>When I evaluate my network on test data, I got following results :<\/p>\n\n<blockquote>\n  <p>loss = 0.5289223349094391<\/p>\n  \n  <p>pearson = 0.3701728057861328<\/p>\n<\/blockquote>\n\n<p><em>(Loss is <code>mean_squared_error<\/code>)<\/em><\/p>\n\n<hr>\n\n<p>Then I predict the test data and compute the same metrics with Scipy :<\/p>\n\n<pre><code>import scipy.stats as measures\nper_coef = measures.pearsonr(y_pred, y_true)[0]\nmse_coef = np.mean(np.square(np.array(y_pred) - np.array(y_true)))\n<\/code><\/pre>\n\n<p>And I get following results :<\/p>\n\n<blockquote>\n  <p>Pearson = 0.5715300096509959 <\/p>\n  \n  <p>MSE = 0.5289223312665985<\/p>\n<\/blockquote>\n\n<hr>\n\n<p><strong>Is it a known issue ? Is it normal ?<\/strong><\/p>\n\n<h1>Minimal, complete and verifiable example<\/h1>\n\n<pre><code>import tensorflow as tf\nimport scipy.stats as measures\n\ny_pred = [2, 2, 3, 4, 5, 5, 4, 2]\ny_true = [1, 2, 3, 4, 5, 6, 7, 8]\n\n## Scipy\nval2 = measures.pearsonr(y_pred, y_true)[0]\nprint(\"Scipy's Pearson = {}\".format(val2))\n\n## Tensorflow\nlogits = tf.placeholder(tf.float32, [8])\nlabels = tf.to_float(tf.Variable(y_true))\n\nacc, acc_op = tf.contrib.metrics.streaming_pearson_correlation(logits,labels)\n\nsess = tf.Session()\nsess.run(tf.local_variables_initializer())\nsess.run(tf.global_variables_initializer())\nsess.run(acc, {logits:y_pred})\nsess.run(acc_op, {logits:y_pred})\n\nprint(\"Tensorflow's Pearson:{}\".format(sess.run(acc,{logits:y_pred})))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53412330","CreationDate":"2018-11-21T12:45:05.440","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>In the minimal verifiable example you gave, <code>y_pred<\/code> and <code>y_true<\/code> are lists of integers. In the first line of the <code>scipy.stats.measures.pearsonr<\/code> <a href=\"https:\/\/github.com\/scipy\/scipy\/blob\/v0.14.0\/scipy\/stats\/stats.py#L2392\" rel=\"nofollow noreferrer\">source<\/a>, you will see that the inputs are converted to numpy arrays with <code>x = np.asarray(x)<\/code>. We can see the resulting data types of these arrays via:<\/p>\n\n<pre><code>print(np.asarray(y_pred).dtype)  # Prints 'int64'\n<\/code><\/pre>\n\n<p>When dividing two <code>int64<\/code> numbers, SciPy uses <code>float64<\/code> precision, while TensorFlow will use <code>float32<\/code> precision in the example above. The difference can be quite large, even for a single division:<\/p>\n\n<pre><code>&gt;&gt;&gt; '%.15f' % (8.5 \/ 7)\n'1.214285714285714'\n&gt;&gt;&gt; '%.15f' % (np.array(8.5, dtype=np.float32) \/ np.array(7, dtype=np.float32))\n'1.214285731315613'\n&gt;&gt;&gt; '%.15f' % (np.array(8.5, dtype=np.float32) \/ np.array(7, dtype=np.float32) - 8.5 \/ 7)\n'0.000000017029899'\n<\/code><\/pre>\n\n<p>You can get the same results for SciPy and TensorFlow by using <code>float32<\/code> precision for <code>y_pred<\/code> and <code>y_true<\/code>:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport scipy.stats as measures\n\ny_pred = np.array([2, 2, 3, 4, 5, 5, 4, 2], dtype=np.float32)\ny_true = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.float32)\n\n## Scipy\nval2 = measures.pearsonr(y_pred, y_true)[0]\nprint(\"Scipy's Pearson: \\t\\t{}\".format(val2))\n\n## Tensorflow\nlogits = tf.placeholder(tf.float32, [8])\nlabels = tf.to_float(tf.Variable(y_true))\n\nacc, acc_op = tf.contrib.metrics.streaming_pearson_correlation(logits,labels)\n\nsess = tf.Session()\nsess.run(tf.local_variables_initializer())\nsess.run(tf.global_variables_initializer())\nsess.run(acc, {logits:y_pred})\nsess.run(acc_op, {logits:y_pred})\n\nprint(\"Tensorflow's Pearson: \\t{}\".format(sess.run(acc,{logits:y_pred})))\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Scipy's Pearson:        0.38060760498046875\nTensorflow's Pearson:   0.38060760498046875\n<\/code><\/pre>\n\n<h2>Differences between SciPy's and TensorFlow's computation<\/h2>\n\n<p>In the test scores you report, the difference is quite high. I took a look at the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/contrib\/metrics\/python\/ops\/metric_ops.py#L3300\" rel=\"nofollow noreferrer\">source<\/a> and found the following differences:<\/p>\n\n<h3>1. Update ops<\/h3>\n\n<p>The result of <code>tf.contrib.metrics.streaming_pearson_correlation<\/code> is not stateless. It returns the correlation coefficient op, together with an <code>update_op<\/code> for new incoming data. If you call the update op with different data before calling the coefficient op with the actual <code>y_pred<\/code>, it will give a completely different result:<\/p>\n\n<pre><code>sess.run(tf.global_variables_initializer())\n\nfor _ in range(20):\n    sess.run(acc_op, {logits: np.random.randn(*y_pred.shape)})\n\nprint(\"Tensorflow's Pearson: \\t{}\".format(sess.run(acc,{logits:y_pred})))\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Scipy's Pearson:        0.38060760498046875\nTensorflow's Pearson:   -0.0678008571267128\n<\/code><\/pre>\n\n<h3>2. Different formulae<\/h3>\n\n<p>SciPy:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YbWQ3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YbWQ3.png\" alt=\"\"><\/a><\/p>\n\n<p>TensorFlow:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DE29r.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DE29r.png\" alt=\"\"><\/a><\/p>\n\n<p>While mathematically the same, the computation of the correlation coefficient is different in TensorFlow. It uses the sample covariance for (x, x), (x, y) and (y, y) to compute the correlation coefficient, which can introduce different rounding errors.<\/p>\n"}]}
+{"QuestionId":53404679,"AnswerCount":3,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2018-11-21T03:00:00.457","AcceptedAnswerId":null,"OwnerUserId":7591187.0,"Title":"How do I train the Convolutional Neural Network with negative and positive elements as the input of the first layer?","Body":"<p>Just I am curious why I have to scale the testing set on the testing set, and not on the training set when I\u2019m training a model on, for example, CNN?!\nOr am I wrong? And I still have to scale it on the training set.\nAlso, can I train a dataset in the CNN that contents positive and negative elements as the first input of the network?\nAny answers with reference will be really appreciated.  <\/p>\n","answers":[{"AnswerId":"53404984","CreationDate":"2018-11-21T03:46:31.160","ParentId":null,"OwnerUserId":"1304642","Title":null,"Body":"<p>We usually have 3 types of datasets for getting a model trained,<\/p>\n\n<ol>\n<li>Training Dataset<\/li>\n<li>Validation Dataset<\/li>\n<li>Test Dataset<\/li>\n<\/ol>\n\n<p><strong>Training Dataset<\/strong><\/p>\n\n<p>This should be an evenly distributed data set which covers all varieties of data. If your train with more <strong>epochs<\/strong>, the model will get used to the training dataset and will only give proper proper prediction on the training dataset and this is called <strong>Overfitting<\/strong>. Only way to keep a check on overfitting is by having other datasets which the model has never been trained on.<\/p>\n\n<p><strong>Validation Dataset<\/strong><\/p>\n\n<p>This can be used fine tune model hyperparameters<\/p>\n\n<p><strong>Test Dataset<\/strong><\/p>\n\n<p>This is the dataset which the model has not been trained on has never been a part of deciding the hyperparameters and will give the reality of how the model is performing.<\/p>\n"},{"AnswerId":"53404823","CreationDate":"2018-11-21T03:23:36.303","ParentId":null,"OwnerUserId":"8498264","Title":null,"Body":"<p>Scaling data depends upon the requirement as well the feed\/data you got. Test data gets scaled with Test data only, because <strong>Test data don't have the Target variable<\/strong> (one less feature in Test data). If we scale our Training data with new Test data, our model will not be able to correlate with any target variable and thus fail to learn. So the key difference is the existence of Target variable.<\/p>\n"},{"AnswerId":"53412207","CreationDate":"2018-11-21T12:38:16.080","ParentId":null,"OwnerUserId":"4288795","Title":null,"Body":"<p>If scaling and normalization is used, the testing set should use the same parameters used during training. \nA good answer that links to that: <a href=\"https:\/\/datascience.stackexchange.com\/questions\/27615\/should-we-apply-normalization-to-test-data-as-well\">https:\/\/datascience.stackexchange.com\/questions\/27615\/should-we-apply-normalization-to-test-data-as-well<\/a><\/p>\n\n<p>Also, some models tend to require normalization and others do not. \nThe Neural Network architectures are normally robust and might not need normalization.<\/p>\n"}]}
+{"QuestionId":53405145,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><segmentation-fault><semantic-segmentation><deeplab>","CreationDate":"2018-11-21T04:11:12.083","AcceptedAnswerId":null,"OwnerUserId":10683482.0,"Title":"Why do all images resulting by Deeplabv 3+ become black only?","Body":"<p>I tried semanticsegmentation with deeplab v3+ but I got results all black out.<\/p>\n\n<p>I deleted the original file and put original datas in ImageSets\/,JPEGImages\/ and SegmentationClass\/ corresponding to each.<\/p>\n\n<p>I prepared SegmentationClassRaw image according to rule of PASCAL VOC 2012 color.<\/p>\n\n<p>And I edited build_voc2012_data.py and segmentation_dataset.py<\/p>\n\n<p>[build_voc2012_data.py]<\/p>\n\n<pre><code>FLAGS = tf.app.flags.FLAGS\n\ntf.app.flags.DEFINE_string('image_folder',\n                           '.\/VOCdevkit\/VOC2012\/JPEGImages',\n                           'Folder containing images.')\n\ntf.app.flags.DEFINE_string(\n    'semantic_segmentation_folder',\n    '.\/VOCdevkit\/VOC2012\/SegmentationClassRaw',\n    'Folder containing semantic segmentation annotations.')\n\ntf.app.flags.DEFINE_string(\n    'list_folder',\n    '.\/VOCdevkit\/VOC2012\/ImageSets\/Segmentation',\n    'Folder containing lists for training and validation')\n\ntf.app.flags.DEFINE_string(\n    'output_dir',\n    '.\/tfrecord',\n    'Path to save converted SSTable of TensorFlow examples.')\n\n\n_NUM_SHARDS = 4\n\n# add --&gt;&gt;\nFLAGS.image_folder = \".\/pascal_voc_seg\/VOCdevkit\/VOC2012\/JPEGImages\"\nFLAGS.semantic_segmentation_folder = \".\/pascal_voc_seg\/VOCdevkit\/VOC2012\/SegmentationClassRaw\"\nFLAGS.list_folder = \".\/pascal_voc_seg\/VOCdevkit\/VOC2012\/ImageSets\/Segmentation\"\nFLAGS.image_format = \"png\"\nFLAGS.output_dir = \".\/pascal_voc_seg\/tfrecord\"\n# add --&lt;&lt;\n<\/code><\/pre>\n\n<p>[segmentation_dataset.pu]<\/p>\n\n<pre><code># add kani 20181115 --&gt;&gt;\n_ORIGINAL_INFORMATION = DatasetDescriptor(\n    splits_to_sizes={\n        'train': 10,\n        'trainval': 2,\n        'val': 2,\n    },\n    num_classes=5,\n    ignore_label=255,\n)\n#add kani 20181115 --&lt;&lt;\n\n# mod kani 20181115 --&gt;&gt;\n# _DATASETS_INFORMATION = {\n#     'cityscapes': _CITYSCAPES_INFORMATION,\n#     'pascal_voc_seg': _PASCAL_VOC_SEG_INFORMATION,\n#     'ade20k': _ADE20K_INFORMATION,\n# }\n\n_DATASETS_INFORMATION = {\n    'cityscapes': _CITYSCAPES_INFORMATION,\n    'pascal_voc_seg': _PASCAL_VOC_SEG_INFORMATION,\n    'ade20k': _ADE20K_INFORMATION,\n    'original': _ORIGINAL_INFORMATION,\n}\n# mod kani 20181115 --&lt;&lt;\n<\/code><\/pre>\n\n<p>I run train.py and vis.py like this.<\/p>\n\n<p>[train.py command] <\/p>\n\n<pre><code>python train.py   --logtostderr   --train_split=trainval   --model_variant=xception_65   --atrous_rates=3   --atrous_rates=6   --atrous_rates=9   --output_stride=32   --decoder_output_stride=4   --train_crop_size=512   --train_crop_size=512   --train_batch_size=2   --training_number_of_steps=6000   --fine_tune_batch_norm=false   --tf_initial_checkpoint=\".\/datasets\/pascal_voc_seg\/init_models\/deeplabv3_pascal_train_aug\/model.ckpt\"  --train_logdir=\".\/datasets\/pascal_voc_seg\/exp\/train_on_trainval_set\/train\"  --dataset_dir=\".\/datasets\/pascal_voc_seg\/tfrecord\" --dataset=original\n<\/code><\/pre>\n\n<p>[vis.py command]<\/p>\n\n<pre><code>python vis.py   --logtostderr   --vis_split=\"val\"   --model_variant=\"xception_65\"   --atrous_rates=6   --atrous_rates=12   --atrous_rates=18   --output_stride=16   --decoder_output_stride=4   --vis_crop_size=513   --vis_crop_size=513   --checkpoint_dir=\".\/datasets\/pascal_voc_seg\/exp\/train_on_trainval_set\/train\"   --vis_logdir=\".\/datasets\/pascal_voc_seg\/exp\/train_on_trainval_set\/vis\"  --dataset_dir=\".\/datasets\/pascal_voc_seg\/tfrecord\"   --max_number_of_iterations=1   --dataset=original   --max_resize_value=512   --min_resize_value=128\n<\/code><\/pre>\n\n<p>Both ended without a problem but I confirmed pictures in <code>datasets\/pascal_voc_seg\/exp\/train_on_trainval_set\/vis\/raw_segmentation_results\/<\/code>, these are black out all. why?<\/p>\n\n<p>Is this because of train datas are bigger than 512x512?\n(Train datas size are so big about 15000x13500)<\/p>\n\n<p>[structure my directory]  <\/p>\n\n<pre><code>\/tmp\/models\/research\/deeplab  \n-README.md  \n-common.py  \n-datasets\/  \n--__init__.py  \n--build_data.py  \n--convert_cityscapes.sh  \n--pascal_voc_seg\/  \n---VOCdevkit\/  \n----VOC2012\/  \n-----Annotations\/  \n-----ImageSets\/  \n-----JPEGImages\/  \n-----SegmentationClass\/  \n-----SegmentationObject\/  \n---VOCtrainval_11-May-2012.tar  \n---exp\/  \n----train_on_trainval_set\/  \n-----train\/  \n------train.py  \n-----vis\/  \n------vis.py  \n---init_models\/  \n----deeplabv3_pascal_train_aug\/  \n-----frozen_inference_graph.pb  \n-----model.ckpt.data-00000-of-00001  \n-----model.ckpt.index  \n---tfrecord\/  \n----build_voc_2012.py  \n--__pycache__  \n--build_data.pyc  \n--download_and_convert_ade20k.sh  \n--remove_gt_colormap.py  \n--build_ade20k_data.py  \n--build_voc2012_data.py  \n--download_and_convert_voc2012.sh  \n--segmentation_dataset.py  \n--build_cityscapes_data.py  \n--build_voc2012_data.py.org  \n-export_model.py  \n-local_test.sh  \n-model_test.py  \n-utils\/  \n-__init__.py  \n-common_test.py  \n-deeplab_demo.ipynb  \n-g3doc\/  \n-local_test_mobilenetv2.sh  \n-train.py  \n-vis.py  \n-__pycache__  \n-core\/  \n-eval.py  \n-input_preprocess.py  \n-model.py  \n-train.py.bk  \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53405657,"AnswerCount":3,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-21T05:17:25.370","AcceptedAnswerId":53658935.0,"OwnerUserId":3259896.0,"Title":"Converting Tensorflow Graph to use Estimator, get 'TypeError: data type not understood' at loss function using `sampled_softmax_loss` or `nce_loss`","Body":"<p>I am trying to convert Tensorflow's official basic word2vec implementation to use tf.Estimator. \nThe issue is that the loss function( <code>sampled_softmax_loss<\/code> or <code>nce_loss<\/code> ) gives an error when using Tensorflow Estimators. It works perfectly fine in the original implementation. <\/p>\n\n<p>Here's is Tensorflow's official basic word2vec implementation:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py<\/a><\/p>\n\n<p>Here is the Google Colab notebook where I implemented this code, which is working. <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1nTX77dRBHmXx6PEF5pmYpkIVxj_TqT5I\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1nTX77dRBHmXx6PEF5pmYpkIVxj_TqT5I<\/a><\/p>\n\n<p>Here is the Google Colab notebook where I altered the code so that it uses Tensorflow Estimator, which is Not working.  <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1IVDqGwMx6BK5-Bgrw190jqHU6tt3ZR3e\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1IVDqGwMx6BK5-Bgrw190jqHU6tt3ZR3e<\/a><\/p>\n\n<p>For convenience, here is exact code from the Estimator version above where I define <code>model_fn<\/code><\/p>\n\n<pre><code>batch_size = 128\nembedding_size = 128  # Dimension of the embedding vector.\nskip_window = 1  # How many words to consider left and right.\nnum_skips = 2  # How many times to reuse an input to generate a label.\nnum_sampled = 64  # Number of negative examples to sample.\n\ndef my_model( features, labels, mode, params):\n\n    with tf.name_scope('inputs'):\n        train_inputs = features\n        train_labels = labels\n\n    with tf.name_scope('embeddings'):\n        embeddings = tf.Variable(\n          tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))\n        embed = tf.nn.embedding_lookup(embeddings, train_inputs)\n\n    with tf.name_scope('weights'):\n        nce_weights = tf.Variable(\n          tf.truncated_normal(\n              [vocabulary_size, embedding_size],\n              stddev=1.0 \/ math.sqrt(embedding_size)))\n    with tf.name_scope('biases'):\n        nce_biases = tf.Variable(tf.zeros([vocabulary_size]))\n\n    with tf.name_scope('loss'):\n        loss = tf.reduce_mean(\n            tf.nn.nce_loss(\n                weights=nce_weights,\n                biases=nce_biases,\n                labels=train_labels,\n                inputs=embed,\n                num_sampled=num_sampled,\n                num_classes=vocabulary_size))\n\n    tf.summary.scalar('loss', loss)\n\n    if mode == \"train\":\n        with tf.name_scope('optimizer'):\n            optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)\n\n        return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=optimizer)\n<\/code><\/pre>\n\n<p>And here is where I call the estimator and training <\/p>\n\n<pre><code>word2vecEstimator = tf.estimator.Estimator(\n        model_fn=my_model,\n        params={\n            'batch_size': 16,\n            'embedding_size': 10,\n            'num_inputs': 3,\n            'num_sampled': 128,\n            'batch_size': 16\n        })\n\nword2vecEstimator.train(\n    input_fn=generate_batch,\n    steps=10)\n<\/code><\/pre>\n\n<p>And this the error message I get when I call the Estimator training:<\/p>\n\n<pre><code>INFO:tensorflow:Calling model_fn.\n---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-22-955f44867ee5&gt; in &lt;module&gt;()\n      1 word2vecEstimator.train(\n      2     input_fn=generate_batch,\n----&gt; 3     steps=10)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in train(self, input_fn, hooks, steps, max_steps, saving_listeners)\n    352 \n    353       saving_listeners = _check_listeners_type(saving_listeners)\n--&gt; 354       loss = self._train_model(input_fn, hooks, saving_listeners)\n    355       logging.info('Loss for final step: %s.', loss)\n    356       return self\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _train_model(self, input_fn, hooks, saving_listeners)\n   1205       return self._train_model_distributed(input_fn, hooks, saving_listeners)\n   1206     else:\n-&gt; 1207       return self._train_model_default(input_fn, hooks, saving_listeners)\n   1208 \n   1209   def _train_model_default(self, input_fn, hooks, saving_listeners):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _train_model_default(self, input_fn, hooks, saving_listeners)\n   1235       worker_hooks.extend(input_hooks)\n   1236       estimator_spec = self._call_model_fn(\n-&gt; 1237           features, labels, model_fn_lib.ModeKeys.TRAIN, self.config)\n   1238       global_step_tensor = training_util.get_global_step(g)\n   1239       return self._train_with_estimator_spec(estimator_spec, worker_hooks,\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/estimator\/estimator.py in _call_model_fn(self, features, labels, mode, config)\n   1193 \n   1194     logging.info('Calling model_fn.')\n-&gt; 1195     model_fn_results = self._model_fn(features=features, **kwargs)\n   1196     logging.info('Done calling model_fn.')\n   1197 \n\n&lt;ipython-input-20-9d389437162a&gt; in my_model(features, labels, mode, params)\n     33                 inputs=embed,\n     34                 num_sampled=num_sampled,\n---&gt; 35                 num_classes=vocabulary_size))\n     36 \n     37     # Add the loss value as a scalar to summary.\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py in nce_loss(weights, biases, labels, inputs, num_sampled, num_classes, num_true, sampled_values, remove_accidental_hits, partition_strategy, name)\n   1246       remove_accidental_hits=remove_accidental_hits,\n   1247       partition_strategy=partition_strategy,\n-&gt; 1248       name=name)\n   1249   sampled_losses = sigmoid_cross_entropy_with_logits(\n   1250       labels=labels, logits=logits, name=\"sampled_losses\")\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py in _compute_sampled_logits(weights, biases, labels, inputs, num_sampled, num_classes, num_true, sampled_values, subtract_log_q, remove_accidental_hits, partition_strategy, name, seed)\n   1029   with ops.name_scope(name, \"compute_sampled_logits\",\n   1030                       weights + [biases, inputs, labels]):\n-&gt; 1031     if labels.dtype != dtypes.int64:\n   1032       labels = math_ops.cast(labels, dtypes.int64)\n   1033     labels_flat = array_ops.reshape(labels, [-1])\n\nTypeError: data type not understood\n<\/code><\/pre>\n\n<p>Edit: Upon request, here's what a typical output for input_fn looks like<\/p>\n\n<p><code>print(generate_batch(batch_size=8, num_skips=2, skip_window=1))<\/code><\/p>\n\n<pre><code>(array([3081, 3081,   12,   12,    6,    6,  195,  195], dtype=int32), array([[5234],\n       [  12],\n       [   6],\n       [3081],\n       [  12],\n       [ 195],\n       [   6],\n       [   2]], dtype=int32))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53536930","CreationDate":"2018-11-29T10:31:04.837","ParentId":null,"OwnerUserId":"10585322","Title":null,"Body":"<p>You use <code>generate_batch<\/code> like a variable here:<\/p>\n\n<pre><code>word2vecEstimator.train(\n    input_fn=generate_batch,\n    steps=10)\n<\/code><\/pre>\n\n<p>Call the function with <code>generate_batch()<\/code>.\nBut I think you must pass  some values to the function.<\/p>\n"},{"AnswerId":"53599805","CreationDate":"2018-12-03T18:36:12.480","ParentId":null,"OwnerUserId":"3259896","Title":null,"Body":"<p>It might be that tensors and ops must be in the <code>input_fn<\/code>, not in the 'model_fn'<\/p>\n\n<blockquote>\n  <p>I found this issue #4026 which solved my problem ... Maybe it is just me being stupid, but it would be great if you mention that the tensors and ops all have to be inside the input_fn somewhere in the documentation.<\/p>\n  \n  <p>You have to call read_batch_examples from somewhere inside input_fn so that the tensors it creates are in the graph that Estimator creates in fit().<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8042\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8042<\/a><\/p>\n\n<blockquote>\n  <p>Oh I feel like an idiot! I've been creating the op outside of the graph scope. It works now, can't believe I didn't think to try that. Thanks a lot! This is a non-issue and has been resolved<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4026\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4026<\/a><\/p>\n\n<p>However, there still is not enough info on what's causing the issue. This is just a lead. <\/p>\n"},{"AnswerId":"53658935","CreationDate":"2018-12-06T20:09:50.357","ParentId":null,"OwnerUserId":"3259896","Title":null,"Body":"<p>Found the answer<\/p>\n\n<blockquote>\n  <p>Error clearly says you have invalid type for labels.<\/p>\n  \n  <p>You trying to pass numpy array instead of Tensor. Sometimes Tensorflow\n  performs implicit conversion from ndarray to Tensor under the hood\n  (what's why your code works outside of Estimator), but in this case it\n  don't.<\/p>\n<\/blockquote>\n\n<p>.<\/p>\n\n<blockquote>\n  <p>No, official impl. feeds data from a placeholder. Placeholder is\n  always a Tensor, so it don't depends on implicit things.<\/p>\n  \n  <p>But if you directly call loss function with a numpy array as input\n  (Notice: call during graph construction phase, so argument content\n  gets embedded into graph), it MAY work (however, I did not check it).<\/p>\n  \n  <p>This code:<\/p>\n  \n  <p>nce_loss(labels=[1,2,3]) will be called only ONCE during graph\n  construction. Labels will be statically embedded into graph as a\n  constant and potentially can be of any Tensor-compatible type (list,\n  ndarray, etc)<\/p>\n  \n  <p>This code: ```Python def model(label_input):\n  nce_loss(labels=label_input)<\/p>\n  \n  <p>estimator(model_fun=model).train() ``` can't embed labels variable\n  statically, because it content is not defined during graph\n  construction. So if you feed anything except the Tensor, it will throw\n  an error.<\/p>\n<\/blockquote>\n\n<p>From<\/p>\n\n<p><a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/a39pef\/r_tensorflow_estimators_managing_simplicity_vs\/\" rel=\"nofollow noreferrer\">https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/a39pef\/r_tensorflow_estimators_managing_simplicity_vs\/<\/a><\/p>\n\n<p>So I used <code>labels=tf.dtypes.cast( train_labels, tf.int64)<\/code> and it worked<\/p>\n"}]}
+{"QuestionId":53405934,"AnswerCount":1,"Tags":"<python><optimization><neural-network><deep-learning><pytorch>","CreationDate":"2018-11-21T05:47:35.817","AcceptedAnswerId":53407300.0,"OwnerUserId":7433733.0,"Title":"How to print the \"actual\" learning rate in Adadelta in pytorch","Body":"<p><strong>In short<\/strong>: <\/p>\n\n<p>I can't draw lr\/epoch curve when using adadelta optimizer in pytorch because <code>optimizer.param_groups[0]['lr']<\/code> always return the same value.<\/p>\n\n<p><strong>In detail<\/strong>:<\/p>\n\n<p>Adadelta can dynamically adapts over time using only first order information and\nhas minimal computational overhead beyond vanilla stochastic gradient descent [1].<\/p>\n\n<p>In pytorch, the source code of Adadelta is here <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/adadelta.html#Adadelta\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/adadelta.html#Adadelta<\/a><\/p>\n\n<p>Since it requires no manual tuning of learning rate, in my knowledge, we don't have to set any schedular after declare the optimizer<\/p>\n\n<p><code>self.optimizer = torch.optim.Adadelta(self.model.parameters(), lr=1)<\/code><\/p>\n\n<p>The way to check learning rate is <\/p>\n\n<p><code>current_lr = self.optimizer.param_groups[0]['lr']<\/code> <\/p>\n\n<p><strong>The problem is it always return 1 (the initial lr).<\/strong> <\/p>\n\n<p>Could anyone tell me how can I get the true learning rate so that can I draw a lr\/epch curve?<\/p>\n\n<p>[1] <a href=\"https:\/\/arxiv.org\/pdf\/1212.5701.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1212.5701.pdf<\/a><\/p>\n","answers":[{"AnswerId":"53407300","CreationDate":"2018-11-21T07:40:54.807","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>Check: <code>self.optimizer.state<\/code>. This is optimized with the lr and used in optimization process. <\/p>\n\n<p>From documentation a lr is just:<\/p>\n\n<blockquote>\n  <p>lr (float, optional): coefficient that scale delta before it is\n  applied\n                  to the parameters (default: 1.0)<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/adadelta.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/adadelta.html<\/a><\/p>\n\n<p>Edited: you may find acc_delta values in self.optimizer.state values but you need to go through dictionaries contained by this dictionary:<\/p>\n\n<pre><code>dict_with_acc_delta = [self.optimizer.state[i] for i in self.optimizer.state.keys() if \"acc_delta\" in self.optimizer.state[i].keys()]\nacc_deltas = [i[\"acc_delta\"] for i in dict_with_acc_delta]\n<\/code><\/pre>\n\n<p>I have eight layers and shapes of elements in the acc_deltas list are following <\/p>\n\n<pre><code>[torch.Size([25088]),\n torch.Size([25088]),\n torch.Size([4096, 25088]),\n torch.Size([4096]),\n torch.Size([1024, 4096]),\n torch.Size([1024]),\n torch.Size([102, 1024]),\n torch.Size([102])]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53406545,"AnswerCount":1,"Tags":"<tensorflow><keras><reshape><layer><shapes>","CreationDate":"2018-11-21T06:43:19.587","AcceptedAnswerId":53406660.0,"OwnerUserId":9587055.0,"Title":"The meaning of question mark (?) in keras layers input size","Body":"<p>I'm using python2 with keras, tensorflow.<\/p>\n\n<pre><code>x = Input((32,), name=\"input1\")\n<\/code><\/pre>\n\n<p>I think x's shape is (32,) but print(x) 's result is that 'shape(?,32)'.\nWhat is means of 'shape(?,32)'?\nAnd '?' means what and 32 means what..?<\/p>\n","answers":[{"AnswerId":"53406660","CreationDate":"2018-11-21T06:51:31.357","ParentId":null,"OwnerUserId":"3222782","Title":null,"Body":"<p>When you define tour input with <code>Input((32,), name=\"input1\")<\/code> you are telling Keras that each input will be 1-dimensional with size 32. However you might send more than one input during training\/predicting. For example if you send in 10 samples, each with length 32, you will actually send in a tensor with shape <code>(10, 32)<\/code>. <\/p>\n\n<p>Since the topology of the network is not dependent on the number of samples you send in, the shape may vary and is presented as <code>(?,32)<\/code> where <code>?<\/code> is the number of samples. <\/p>\n"}]}
+{"QuestionId":53407224,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><spyder>","CreationDate":"2018-11-21T07:35:17.293","AcceptedAnswerId":null,"OwnerUserId":10684324.0,"Title":"Have anyone managed to use windows+anaconda+tensorflow combination?","Body":"<p>If anyone did it successfully, please share the procedure with simple steps.\nI am using win 8, anaconda 3.<\/p>\n\n<p>Probably one guy from Spyder team will join and write, the problem is not related to spyder, as usual, lol<\/p>\n","answers":[{"AnswerId":"53410289","CreationDate":"2018-11-21T10:40:35.070","ParentId":null,"OwnerUserId":"5810950","Title":null,"Body":"<ol>\n<li>Download Anaconda for windows from <a href=\"https:\/\/www.anaconda.com\/\" rel=\"nofollow noreferrer\">here<\/a><\/li>\n<li><p>Create the new environment called 'py35'. Open Windows command prompt and type: <\/p>\n\n<p><code>conda create -n py35 python=3.5 anaconda<\/code> <\/p>\n\n<p>This automatically installs python 3.5 in the new environment.<\/p><\/li>\n<li><p>Now, activate the environment by typing the following into the command prompt:<\/p>\n\n<p><code>activate py35<\/code><\/p><\/li>\n<li><p>The command prompt should now change to have a (py35)  at the beginning of each line. This indicates that we're working in the new py35 environment.<\/p><\/li>\n<li><p>Install the Packages:<\/p>\n\n<p><code>conda install spyder\nconda install tensorflow (cpu-version)<\/code><\/p>\n\n<p>Hopefully this should complete without errors. If you're having trouble getting tensorflow to work try:\n<code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0-cp35-cp35m-win_amd64.whl<\/code><\/p><\/li>\n<li><p>If you want to install the GPU version of tensorflow, just type:\n<code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-0.12.1-cp35-cp35m-win_amd64.whl<\/code><\/p>\n\n<p>Now, suppose you want to open spyder, open windows command prompt and type <code>activate py35<\/code>. Type <code>spyder<\/code> and press enter. <\/p>\n\n<p>Check if tensorflow is properly installed or not using <code>import tensorflow<\/code>.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53407432,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-21T07:53:02.057","AcceptedAnswerId":null,"OwnerUserId":10347970.0,"Title":"Importing TensorFlow 1.9.0 on Jupyter Notebook Causes an Error","Body":"<p>I installed TensorFlow 1.12.0 using Anaconda Navigator, and when I try to import it on Jupyter notebook I keep getting this error:<\/p>\n\n<pre><code>__init__() got an unexpected keyword argument 'serialized_options'\n<\/code><\/pre>\n\n<p>I tried searching for a similar problem here, all I found was to install TensorFlow 1.9.0 instead of any other version. I tried that and it did not work. I even created a whole new environment on Anaconda Navigator and installed nothing but TensorFlow 1.9.0 on it and it still gave me the same error. I tried uninstalling the whole thing and starting over with 1.9.0 but it is no use. I tried all the older versions and they all still gave me the same error. Does anyone have any idea what the issue might be?<\/p>\n\n<p>Here is the complete error<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-1-64156d691fe5&gt; in &lt;module&gt;\n----&gt; 1 import tensorflow as tf\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\__init__.py in &lt;module&gt;\n     22 \n     23 # pylint: disable=g-bad-import-order\n---&gt; 24 from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n     25 \n     26 try:\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\__init__.py in &lt;module&gt;\n     57 \n     58 # Protocol buffers\n---&gt; 59 from tensorflow.core.framework.graph_pb2 import *\n     60 from tensorflow.core.framework.node_def_pb2 import *\n     61 from tensorflow.core.framework.summary_pb2 import *\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py in &lt;module&gt;\n     13 \n     14 \n---&gt; 15 from tensorflow.core.framework import node_def_pb2 as tensorflow_dot_core_dot_framework_dot_node__def__pb2\n     16 from tensorflow.core.framework import function_pb2 as tensorflow_dot_core_dot_framework_dot_function__pb2\n     17 from tensorflow.core.framework import versions_pb2 as tensorflow_dot_core_dot_framework_dot_versions__pb2\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\node_def_pb2.py in &lt;module&gt;\n     13 \n     14 \n---&gt; 15 from tensorflow.core.framework import attr_value_pb2 as tensorflow_dot_core_dot_framework_dot_attr__value__pb2\n     16 \n     17 \n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\attr_value_pb2.py in &lt;module&gt;\n     13 \n     14 \n---&gt; 15 from tensorflow.core.framework import tensor_pb2 as tensorflow_dot_core_dot_framework_dot_tensor__pb2\n     16 from tensorflow.core.framework import tensor_shape_pb2 as tensorflow_dot_core_dot_framework_dot_tensor__shape__pb2\n     17 from tensorflow.core.framework import types_pb2 as tensorflow_dot_core_dot_framework_dot_types__pb2\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\tensor_pb2.py in &lt;module&gt;\n     13 \n     14 \n---&gt; 15 from tensorflow.core.framework import resource_handle_pb2 as tensorflow_dot_core_dot_framework_dot_resource__handle__pb2\n     16 from tensorflow.core.framework import tensor_shape_pb2 as tensorflow_dot_core_dot_framework_dot_tensor__shape__pb2\n     17 from tensorflow.core.framework import types_pb2 as tensorflow_dot_core_dot_framework_dot_types__pb2\n\n~\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\resource_handle_pb2.py in &lt;module&gt;\n     20   syntax='proto3',\n     21   serialized_options=_b('\\n\\030org.tensorflow.frameworkB\\016ResourceHandleP\\001Z=github.com\/tensorflow\/tensorflow\/tensorflow\/go\/core\/framework\\370\\001\\001'),\n---&gt; 22   serialized_pb=_b('\\n\/tensorflow\/core\/framework\/resource_handle.proto\\x12\\ntensorflow\\\"r\\n\\x13ResourceHandleProto\\x12\\x0e\\n\\x06\\x64\\x65vice\\x18\\x01 \\x01(\\t\\x12\\x11\\n\\tcontainer\\x18\\x02 \\x01(\\t\\x12\\x0c\\n\\x04name\\x18\\x03 \\x01(\\t\\x12\\x11\\n\\thash_code\\x18\\x04 \\x01(\\x04\\x12\\x17\\n\\x0fmaybe_type_name\\x18\\x05 \\x01(\\tBn\\n\\x18org.tensorflow.frameworkB\\x0eResourceHandleP\\x01Z=github.com\/tensorflow\/tensorflow\/tensorflow\/go\/core\/framework\\xf8\\x01\\x01\\x62\\x06proto3')\n     23 )\n     24 \n\nTypeError: __init__() got an unexpected keyword argument 'serialized_options'\n<\/code><\/pre>\n","answers":[{"AnswerId":"54860960","CreationDate":"2019-02-25T06:57:54.627","ParentId":null,"OwnerUserId":"10389355","Title":null,"Body":"<p>I had also faced this issue.\nBy updating the protobuf I got the error resolved.<\/p>\n\n<p>Please try the below code for updating and check whether the problem is resolved:<\/p>\n\n<p><strong>pip install -U protobuf<\/strong><\/p>\n"}]}
+{"QuestionId":53408453,"AnswerCount":2,"Tags":"<python><machine-learning><keras><nlp><word2vec>","CreationDate":"2018-11-21T09:01:50.670","AcceptedAnswerId":53410341.0,"OwnerUserId":8163412.0,"Title":"How to use keras embedding layer with 3D tensor input?","Body":"<p>I am facing difficulty in using Keras embedding layer with one hot encoding of my input data.<\/p>\n\n<p>Following is the toy code.<\/p>\n\n<p><strong>Import packages<\/strong><\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import Flatten\nfrom keras.layers.embeddings import Embedding\nfrom keras.optimizers import Adam\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport openpyxl\nimport pandas as pd\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.callbacks import ReduceLROnPlateau\n<\/code><\/pre>\n\n<p>The input data is text based as follows.<\/p>\n\n<p><strong>Train and Test data<\/strong><\/p>\n\n<pre><code>X_train_orignal= np.array(['OC(=O)C1=C(Cl)C=CC=C1Cl', 'OC(=O)C1=C(Cl)C=C(Cl)C=C1Cl',\n       'OC(=O)C1=CC=CC(=C1Cl)Cl', 'OC(=O)C1=CC(=CC=C1Cl)Cl',\n       'OC1=C(C=C(C=C1)[N+]([O-])=O)[N+]([O-])=O'])\n\nX_test_orignal=np.array(['OC(=O)C1=CC=C(Cl)C=C1Cl', 'CCOC(N)=O',\n       'OC1=C(Cl)C(=C(Cl)C=C1Cl)Cl'])\n\nY_train=np.array(([[2.33],\n       [2.59],\n       [2.59],\n       [2.54],\n       [4.06]]))\n\nY_test=np.array([[2.20],\n   [2.81],\n   [2.00]])\n<\/code><\/pre>\n\n<p><strong>Creating dictionaries<\/strong><\/p>\n\n<p>Now i create two dictionaries, characters to index vice. The unique character number is stored in <code>len(charset)<\/code> and maximum length of the string along with 5 additional characters is stored in <code>embed<\/code>. The start of each string will be padded with <code>!<\/code> and end will be <code>E<\/code>.<\/p>\n\n<pre><code>charset = set(\"\".join(list(X_train_orignal))+\"!E\")\nchar_to_int = dict((c,i) for i,c in enumerate(charset))\nint_to_char = dict((i,c) for i,c in enumerate(charset))\nembed = max([len(smile) for smile in X_train_orignal]) + 5\nprint (str(charset))\nprint(len(charset), embed)\n<\/code><\/pre>\n\n<p><strong>One hot encoding<\/strong> <\/p>\n\n<p>I convert all the train data into one hot encoding as follows. <\/p>\n\n<pre><code>def vectorize(smiles):\n        one_hot =  np.zeros((smiles.shape[0], embed , len(charset)),dtype=np.int8)\n        for i,smile in enumerate(smiles):\n            #encode the startchar\n            one_hot[i,0,char_to_int[\"!\"]] = 1\n            #encode the rest of the chars\n            for j,c in enumerate(smile):\n                one_hot[i,j+1,char_to_int[c]] = 1\n            #Encode endchar\n            one_hot[i,len(smile)+1:,char_to_int[\"E\"]] = 1\n\n        return one_hot[:,0:-1,:]\n\nX_train = vectorize(X_train_orignal)\nprint(X_train.shape)\nX_test = vectorize(X_test_orignal)\nprint(X_test.shape)\n<\/code><\/pre>\n\n<p>When it converts the input train data into one hot encoding, the shape of the one hot encoded data becomes <code>(5, 44, 14)<\/code> for train and <code>(3, 44, 14)<\/code> for test. For train, there are 5 example, 0-44 is the maximum length and 14 are the unique characters. The examples for which there are less number of characters, are padded with <code>E<\/code> till the maximum length.<\/p>\n\n<p><strong>Verifying the correct padding<\/strong>\nFollowing is the code to verify if we have done the padding rightly.<\/p>\n\n<pre><code>mol_str_train=[]\nmol_str_test=[]\nfor x in range(5):\n\n    mol_str_train.append(\"\".join([int_to_char[idx] for idx in np.argmax(X_train[x,:,:], axis=1)]))\n\nfor x in range(3):\n    mol_str_test.append(\"\".join([int_to_char[idx] for idx in np.argmax(X_test[x,:,:], axis=1)]))\n<\/code><\/pre>\n\n<p>and let's see, how the train set looks like.<\/p>\n\n<pre><code>mol_str_train\n\n['!OC(=O)C1=C(Cl)C=CC=C1ClEEEEEEEEEEEEEEEEEEEE',\n '!OC(=O)C1=C(Cl)C=C(Cl)C=C1ClEEEEEEEEEEEEEEEE',\n '!OC(=O)C1=CC=CC(=C1Cl)ClEEEEEEEEEEEEEEEEEEEE',\n '!OC(=O)C1=CC(=CC=C1Cl)ClEEEEEEEEEEEEEEEEEEEE',\n '!OC1=C(C=C(C=C1)[N+]([O-])=O)[N+]([O-])=OEEE']\n<\/code><\/pre>\n\n<p>Now is the time to build model.<\/p>\n\n<p><strong>Model<\/strong><\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(len(charset), 10, input_length=embed))\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='linear'))\n\ndef coeff_determination(y_true, y_pred):\n    from keras import backend as K\n    SS_res =  K.sum(K.square( y_true-y_pred ))\n    SS_tot = K.sum(K.square( y_true - K.mean(y_true) ) )\n    return ( 1 - SS_res\/(SS_tot + K.epsilon()) )\n\ndef get_lr_metric(optimizer):\n    def lr(y_true, y_pred):\n        return optimizer.lr\n    return lr\n\n\noptimizer = Adam(lr=0.00025)\nlr_metric = get_lr_metric(optimizer)\nmodel.compile(loss=\"mse\", optimizer=optimizer, metrics=[coeff_determination, lr_metric])\n\n\n\ncallbacks_list = [\n    ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=1e-15, verbose=1, mode='auto',cooldown=0),\n    ModelCheckpoint(filepath=\"weights.best.hdf5\", monitor='val_loss', save_best_only=True, verbose=1, mode='auto')]\n\n\nhistory =model.fit(x=X_train, y=Y_train,\n                              batch_size=1,\n                              epochs=10,\n                              validation_data=(X_test,Y_test),\n                              callbacks=callbacks_list)\n<\/code><\/pre>\n\n<p><strong>Error<\/strong><\/p>\n\n<pre><code>ValueError: Error when checking input: expected embedding_3_input to have 2 dimensions, but got array with shape (5, 44, 14)\n<\/code><\/pre>\n\n<p>The embedding layer expects two dimensional array. How can I deal with this issue so that it can accept the one hot vector encoded data.<\/p>\n\n<p>All the above code can be run. <\/p>\n","answers":[{"AnswerId":"53409895","CreationDate":"2018-11-21T10:21:13.093","ParentId":null,"OwnerUserId":"7168287","Title":null,"Body":"<p>The Keras embedding layer works with indices, not directly with one-hot encodings.\nSo you don't need to have (5,44,14), just (5,44) works fine.<\/p>\n\n<p>E.g. get indices with argmax:<\/p>\n\n<pre><code>X_test = np.argmax(X_test, axis=2)\nX_train = np.argmax(X_train, axis=2)\n<\/code><\/pre>\n\n<p>Although it's probably better to not one-hot encode it first =)<\/p>\n\n<p>Besides that, your 'embed' variable says size 45, while your data is size 44. <\/p>\n\n<p>If you change those, your model runs fine:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(len(charset), 10, input_length=44))\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='linear'))\n\ndef coeff_determination(y_true, y_pred):\n    from keras import backend as K\n    SS_res =  K.sum(K.square( y_true-y_pred ))\n    SS_tot = K.sum(K.square( y_true - K.mean(y_true) ) )\n    return ( 1 - SS_res\/(SS_tot + K.epsilon()) )\n\ndef get_lr_metric(optimizer):\n    def lr(y_true, y_pred):\n        return optimizer.lr\n    return lr\n\n\noptimizer = Adam(lr=0.00025)\nlr_metric = get_lr_metric(optimizer)\nmodel.compile(loss=\"mse\", optimizer=optimizer, metrics=[coeff_determination,     lr_metric])\n\n\n\ncallbacks_list = [\n    ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=1e-15,     verbose=1, mode='auto',cooldown=0),\n    ModelCheckpoint(filepath=\"weights.best.hdf5\", monitor='val_loss',         save_best_only=True, verbose=1, mode='auto')]\n\n\nhistory =model.fit(x=np.argmax(X_train, axis=2), y=Y_train,\n                              batch_size=1,\n                              epochs=10,\n                              validation_data=(np.argmax(X_test, axis=2),Y_test),\n                              callbacks=callbacks_list)    \n<\/code><\/pre>\n"},{"AnswerId":"53410341","CreationDate":"2018-11-21T10:43:26.043","ParentId":null,"OwnerUserId":"2235817","Title":null,"Body":"<p>our input shape was not defined properly in the embedding layer. The following code works for me by reducing the steps to covert your data dimensions to 2D you can directly pass the 3-D input to your embedding layer.<\/p>\n\n<pre><code>#THE MISSING STUFF\n#_________________________________________\nY_train = Y_train.reshape(5) #Dense layer contains a single unit so need to input single dimension array\nmax_len = len(charset)\nmax_features = embed-1\ninputshape = (max_features, max_len) #input shape didn't define. Embedding layer can accept 3D input by using input_shape\n#__________________________________________\n\nmodel = Sequential()\n#model.add(Embedding(len(charset), 10, input_length=14))\n\nmodel.add(Embedding(max_features, 10, input_shape=inputshape))#input_length=max_len))\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='linear'))\nprint(model.summary())\n\noptimizer = Adam(lr=0.00025)\nlr_metric = get_lr_metric(optimizer)\nmodel.compile(loss=\"mse\", optimizer=optimizer, metrics=[coeff_determination, lr_metric])\n\n\ncallbacks_list = [\n    ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=1e-15, verbose=1, mode='auto',cooldown=0),\n    ModelCheckpoint(filepath=\"weights.best.hdf5\", monitor='val_loss', save_best_only=True, verbose=1, mode='auto')]\n\nhistory =model.fit(x=X_train, y=Y_train,\n                              batch_size=10,\n                              epochs=10,\n                              validation_data=(X_test,Y_test),\n                              callbacks=callbacks_list)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53409440,"AnswerCount":1,"Tags":"<tensorflow><checkpointing>","CreationDate":"2018-11-21T09:57:17.003","AcceptedAnswerId":53414160.0,"OwnerUserId":9442808.0,"Title":"ImageProjectiveTransformV2 error in loading meta graph by import_meta_graph","Body":"<p>I am  trying to load meta graph of trained networks \"name.ckpt-1.meta\" using <code>tf.train.import_meta_graph(\".\/name.ckpt-1.meta\")<\/code>\nbut the following error appears:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/rapsodo\/.local\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 3265, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-7-634d5d15ac05&gt;\", line 1, in &lt;module&gt;\n    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=False)\n  File \"\/home\/rapsodo\/workspace_mike3352\/anaconda2\/envs\/mike_tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1960, in import_meta_graph\n    **kwargs)\n  File \"\/home\/rapsodo\/workspace_mike3352\/anaconda2\/envs\/mike_tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/meta_graph.py\", line 744, in import_scoped_meta_graph\n    producer_op_list=producer_op_list)\n  File \"\/home\/rapsodo\/workspace_mike3352\/anaconda2\/envs\/mike_tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 432, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/rapsodo\/workspace_mike3352\/anaconda2\/envs\/mike_tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/importer.py\", line 391, in import_graph_def\n    _RemoveDefaultAttrs(op_dict, producer_op_list, graph_def)\n  File \"\/home\/rapsodo\/workspace_mike3352\/anaconda2\/envs\/mike_tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/importer.py\", line 158, in _RemoveDefaultAttrs\n    op_def = op_dict[node.op]\nKeyError: 'ImageProjectiveTransformV2'\n<\/code><\/pre>\n\n<p>I did not understand the reason and did not saw same thing somewhere else, Im not sure because of the tensorflow version or something else.<\/p>\n","answers":[{"AnswerId":"53414160","CreationDate":"2018-11-21T14:22:14.323","ParentId":null,"OwnerUserId":"9442808","Title":null,"Body":"<p>I found the solution; It is because of version mismatch. Newer version of tensorflow does not match with older versions in terms of saving graph etc. \nIf we saved checkpoints with older version, we should use proper version (same version is preferred) to load meta graph or frozen graph. <\/p>\n"}]}
+{"QuestionId":53409455,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T09:58:05.460","AcceptedAnswerId":53418514.0,"OwnerUserId":5024514.0,"Title":"How do I find the output of a tensor and\/or op in tensorflow? (or \"tensorflow op.outputs only pointing to itself\")","Body":"<p>First things first - this might be a very basic and dumb question, but I've tried many things and searched all over to no avail, so here I am.\nThe problem is the following:<br>\nI have a tensor, and I'd like to find \"where it leads\" for various reasons.\nThe way to theoretically do this would be to just look at my_tensor.op.outputs, as per documentation and such, but <strong>this always seems to point back to my_tensor<\/strong> itself!<br>\nI've easily gone the other way before, meaning I can get the input tensor by using my_tensor.op.inputs, but for some reason \"outputs\" isn't doing the expected.<br>\nHere's a simple example:  <\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.placeholder(tf.uint8, name='a')\nb = tf.placeholder(tf.uint8, name='b')\nmy_sum = tf.identity(a + b, name='my_sum')\ngraph = tf.get_default_graph()\n\n# I should have 4 ops at this point, as validated by:\nprint(graph.get_operations())\n&gt;&gt; [&lt;tf.Operation 'a' type=Placeholder&gt;, &lt;tf.Operation 'b' type=Placeholder&gt;, &lt;tf.Operation 'add' type=Add&gt;, &lt;tf.Operation 'my_sum' type=Identity&gt;]\n\n# So let's try get the output of 'a':\nprint(list(a.op.outputs))\n&gt;&gt; [&lt;tf.Tensor 'a:0' shape=&lt;unknown&gt; dtype=uint8&gt;]\n<\/code><\/pre>\n\n<p>If you tried the above, you'll see you got back to 'a'...<br>\nAgain, running my_sum.op.inputs gives the 'add' op, and running even further back get's us to 'a' and 'b' as expected:  <\/p>\n\n<pre><code>input_to_my_sum = list(my_sum.op.inputs)[0]\nprint(input_to_my_sum)\n&gt;&gt; Tensor(\"add:0\", dtype=uint8)\n\nprint(list(input_to_my_sum.op.inputs))\n&gt;&gt; [&lt;tf.Tensor 'a:0' shape=&lt;unknown&gt; dtype=uint8&gt;, &lt;tf.Tensor 'b:0' shape=&lt;unknown&gt; dtype=uint8&gt;]\n<\/code><\/pre>\n\n<p>But the other way round? No such luck:<\/p>\n\n<pre><code>print(list(input_to_my_sum.op.outputs))\n&gt;&gt; [&lt;tf.Tensor 'add:0' shape=&lt;unknown&gt; dtype=uint8&gt;]\n\nprint('This is no fun at all')\n&gt;&gt; This is no fun at all\n<\/code><\/pre>\n\n<p>So <strong>what am I doing wrong?<\/strong><br>\nI've also tried using the (deprecated) op.values() with no success, and I'm confused because the documentation explicitly states that this should give me the outputs of the op (from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation<\/a>):  <\/p>\n\n<blockquote>\n  <p>outputs<br>\n  The list of Tensor objects representing the outputs of this op.<\/p>\n<\/blockquote>\n\n<p>(I checked that a.op.__class__ is the right class and that I'm reading the correct documentation).  <\/p>\n\n<p>(Just to wrap things up, the node_def of the ops also shows no signs of an output field...).<br>\nThanks in advance for any advice! <\/p>\n\n<hr>\n\n<p>Edit (due to Yuxin's answer):<br>\nJust to clarify, taking the output of the output of the etc. stays put on the same tensor. I'm trying to reach the <em>next<\/em> tensor\/op.<\/p>\n\n<p>P.S: This is my first stackoverflow question, so if I did anything \"wrong\" let me know and I'll try fix it.<\/p>\n","answers":[{"AnswerId":"53418514","CreationDate":"2018-11-21T18:34:32.543","ParentId":null,"OwnerUserId":"1601927","Title":null,"Body":"<blockquote>\n  <p>and I'm confused because the documentation explicitly states that this should give me the outputs of the op<\/p>\n<\/blockquote>\n\n<p>It does give you the output of the op:<\/p>\n\n<pre><code>print(list(input_to_my_sum.op.outputs))\n&gt;&gt; [&lt;tf.Tensor 'add:0' shape=&lt;unknown&gt; dtype=uint8&gt;]\n<\/code><\/pre>\n\n<p>So I'm not sure what you're confused about or what you're expecting.<\/p>\n\n<p>Be sure to understand the concept of \"Tensor\" and \"Operation\". In a tensorflow graph, an Operation\ntakes a list of input Tensors and outputs a list of Tensors.\nIn Python, you can <code>print(x)<\/code> to see if <code>x<\/code> is a Tensor or Operation.<\/p>\n\n<p>Operation has <code>.inputs<\/code> (which are Tensors) and <code>.outputs<\/code> (which are Tensors).\nTensor does not have inputs and outputs. Tensor has <code>.op<\/code>, which is the Operation that produces it as outputs.\nTensor also has <code>.consumers()<\/code>, which is a list of Operations which take the Tensor as inputs.<\/p>\n\n<p><code>input_to_my_sum<\/code> is a Tensor, the result of the Add.\n<code>input_to_my_sum.op<\/code> is a Operation, the Add Operation that produces the result <code>input_to_my_sum<\/code>.\nTherefore <code>input_to_my_sum.op.outputs<\/code> contains <code>input_to_my_sum<\/code>.<\/p>\n"}]}
+{"QuestionId":53409481,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><conv-neural-network><zero-padding>","CreationDate":"2018-11-21T09:59:40.140","AcceptedAnswerId":53409771.0,"OwnerUserId":10684929.0,"Title":"What is the default kernel-size, Zero-padding and stride for keras.layers.Conv2D?","Body":"<p>What are the default Kernel-Size, Zero-Padding, and Stride arguments in Conv2D (keras.layers.Conv2D)? What happens if these arguments are not specified?<\/p>\n","answers":[{"AnswerId":"53409771","CreationDate":"2018-11-21T10:14:56.770","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>You can find the documentation here: <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/convolutional\/<\/a><\/p>\n\n<p>In python you can give default values for parameters of a function, If you don't specify these parameters while calling the function, defaults are used instead. <\/p>\n\n<p>In the link above you'll find that Conv2D has the parameters:<\/p>\n\n<pre><code>filters, kernel_size, strides=(1, 1), padding='valid', data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None\n<\/code><\/pre>\n\n<p>only filters and kernel_size parameters must be given, others are optional or has default  values next to them.<\/p>\n"}]}
+{"QuestionId":53409535,"AnswerCount":1,"Tags":"<keras><deep-learning><prediction><keras-layer><transfer-learning>","CreationDate":"2018-11-21T10:02:34.020","AcceptedAnswerId":null,"OwnerUserId":10237443.0,"Title":"I want all the output of the pretrained VGG16 model as well as the new classes it is trained on","Body":"<p>I have tried transfer learning using VGG16 but only getting a result for those classes which are trained.I  want that the output consists of both the VGG16 classes+ my new trained classes. Is it possible?<\/p>\n\n<p>I have attached my whole code.<\/p>\n\n<pre><code>`enter code here`import matplotlib.pyplot as plt\nimport PIL\nimport tensorflow as tf\nimport numpy as np\nimport os\n\n\n\n\n\nimport keras \nfrom keras.models import Sequential, Model\nfrom keras.layers.core import Dense, Dropout, Flatten, Reshape, Activation\nfrom keras.layers import Embedding, Input, merge, ELU\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom keras.optimizers import SGD, Adam, RMSprop\nfrom keras.regularizers import l2\nfrom keras.utils.np_utils import to_categorical\nimport sklearn.metrics as metrics\nfrom PIL import Image, ImageDraw\nfrom keras.applications import VGG16\nfrom keras.applications.vgg16 import preprocess_input, decode_predictions\nfrom keras.preprocessing.image import ImageDataGenerator\n\n\n# # Helper Function\n\n\n\n\ndef load_images(image_paths):\n    # Load the images from disk.\n    images = [plt.imread(path) for path in image_paths]\n\n    # Convert to a numpy array and return it.\n    return np.asarray(images)\n\n\n\n\n\ndef path_join(dirname, filenames):\n    return [os.path.join(dirname, filename) for filename in filenames]\n\n\n# In[5]:\n\n\ntrain_dir = \"\/home\/priyank\/Jupyter_notebook\/plant_leaves_train_set\"\ntest_dir = \"\/home\/priyank\/Jupyter_notebook\/val_data_plant\"\n\n\n# # Pre-Trained Model: VGG16\n# Downloading the pretrained model of imagenet dataset.\n\n\n\n\nmodel = VGG16(include_top=True, weights='imagenet')\n\n\n# # Input Pipeline\n# First we need to know the shape of the tensors expected as input by the pre-trained VGG16 model. In this case it is images of shape 224 x 224 x 3.\n\n\n\ninput_shape = model.layers[0].output_shape[1:3] # the input shape of the vgg16 model\ninput_shape\n\n\n# # ImageDataGenerator \n# It will pick the image one-by-one and transform all the data each time the image is loaded in the training set.\n\n\n\n\n\ndatagen_train = ImageDataGenerator(\n      rescale=1.\/255,\n      rotation_range=180,\n      width_shift_range=0.1,\n      height_shift_range=0.1,\n      shear_range=0.1,\n      zoom_range=[0.9, 1.5],\n      horizontal_flip=True,\n      vertical_flip=True,\n      fill_mode='nearest')\n\n\n\ndatagen_test = ImageDataGenerator(rescale=1.\/255)\n\n\n# \n# The datagenerator will return the batches of the images. VGG16 model is too large so we can't create the batches too large otherwise we will run out of the RAM and GPU.\n\n\n\n\n# Taking small batch size\nbatch_size = 20\n\n\n# \n# We can save the randomly transformed images during training, so as to inspect whether they have been overly distorted, so we have to adjust the parameters for the data-generator above.\n\n\n\n\nif True:\n    save_to_dir = None\nelse:\n    save_to_dir='augmented_images\/'\n\n\n\ngenerator_train = datagen_train.flow_from_directory(directory=train_dir,\n                                                    target_size=input_shape,\n                                                    batch_size=batch_size,\n                                                    shuffle=True,\n                                                    save_to_dir=save_to_dir)\n\n\n\n\n\ngenerator_test = datagen_test.flow_from_directory(directory=test_dir,\n                                                  target_size=input_shape,\n                                                  batch_size=batch_size,\n                                                  shuffle=False)\n\n\n\n\n\nsteps_test = generator_test.n \/ batch_size\nsteps_test\n\n\n\n\n\nimage_paths_train = path_join(train_dir, generator_train.filenames)\nimage_paths_test = path_join(test_dir, generator_test.filenames)\n\n\n\n\ncls_train = generator_train.classes\ncls_test = generator_test.classes\n\n\n\n\n\nclass_names = list(generator_train.class_indices.keys())\nclass_names\n\n\n\n\n\n\n\nnum_classes = generator_train.num_classes\nnum_classes\n\n\n# The dataset we have is imbalanced so the gradients for 9.01192 will remain higher adn the gradients of 0.8080 will reamin lower so that model can learn from higher gradient more than the lower gradient.\n# \n\n\n\nfrom sklearn.utils.class_weight import compute_class_weight\nclass_weight = compute_class_weight(class_weight='balanced',\n                                    classes=np.unique(cls_train),\n                                    y=cls_train)\nclass_weight\n\n\n#  \n#  Predicting the our data image with the already trained VGG16 model. Using  a helper function which can resize the image so it can be the input to VGG16 model\n\n\n\ndef predict(image_path):\n    # Load and resize the image using PIL.\n    img = PIL.Image.open(image_path)\n    img_resized = img.resize(input_shape, PIL.Image.LANCZOS)\n\n    # Plot the image.\n    plt.imshow(img_resized)\n    plt.show()\n\n    # Convert the PIL image to a numpy-array with the proper shape.\n    img_array = np.expand_dims(np.array(img_resized), axis=0)\n\n    # Use the VGG16 model to make a prediction.\n    # This outputs an array with 1000 numbers corresponding to\n    # the classes of the ImageNet-dataset.\n    print(img_array.shape)\n    pred = model.predict(img_array)\n\n    # Decode the output of the VGG16 model.\n    print(pred)\n    print(pred.shape)\n    pred_decoded = decode_predictions(pred)[0]\n\n    # Print the predictions.\n    for code, name, score in pred_decoded:\n        print(\"{0:&gt;6.2%} : {1}\".format(score, name))\n\n\n\n\n\npredict(image_path='\/home\/priyank\/Pictures\/people.jpg')\n\n\n\n\n\npredict(image_path=image_paths_train[0])\n\n\n# The pre-trained VGG16 model was unable to classify images from the plant disease dataset. The reason is perhaps that the VGG16 model was trained on the so-called ImageNet dataset which may not have contained many images of plant diseases.\n# \n# The lower layers of a Convolutional Neural Network can recognize many different shapes or features in an image. It is the last few fully-connected layers that combine these featuers into classification of a whole image. So we can try and re-route the output of the last convolutional layer of the VGG16 model to a new fully-connected neural network that we create for doing classification \n\n\n\n\n# summary of VGG16 model.\nmodel.summary()\n\n\n# We can see that the last convolutional layer is called 'block5_pool' so we use Keras to get a reference to that layer.\n\n\n\n\ntransfer_layer = model.get_layer('block5_pool')\n\n\n# \n# \n# We refer to this layer as the Transfer Layer because its output will be re-routed to our new fully-connected neural network which will do the classification for the Knifey-Spoony dataset.\n# \n# The output of the transfer layer has the following shape:\n# \n\n\n\n\ntransfer_layer.output\n\n\n# we take the part of the VGG16 model from its input-layer to the output of the transfer-layer. We may call this the convolutional model, because it consists of all the convolutional layers from the VGG16 model.\n\n\n\n\nconv_model = Model(inputs=model.input,\n                   outputs=transfer_layer.output)\n\n\n\n\n\n# Start a new Keras Sequential model.\nnew_model = Sequential()\n\n# Add the convolutional part of the VGG16 model from above.\nnew_model.add(conv_model)\n\n# Flatten the output of the VGG16 model because it is from a\n# convolutional layer.\nnew_model.add(Flatten())\n\n# Add a dense (aka. fully-connected) layer.\n# This is for combining features that the VGG16 model has\n# recognized in the image.\nnew_model.add(Dense(1024, activation='relu'))\n\n# Add a dropout-layer which may prevent overfitting and\n# improve generalization ability to unseen data e.g. the test-set.\nnew_model.add(Dropout(0.5))\n\n# Add the final layer for the actual classification.\nnew_model.add(Dense(num_classes, activation='softmax'))\n\n\n\n\n\noptimizer = Adam(lr=1e-5)\n\n\n\n\n\nloss = 'categorical_crossentropy'\n\n\n\n\nmetrics = ['categorical_accuracy']\n\n\n# Helper-function for printing whether a layer in the VGG16 model should be trained.\n\n\n\n\ndef print_layer_trainable():\n    for layer in conv_model.layers:\n        print(\"{0}:\\t{1}\".format(layer.trainable, layer.name))\n\n\n# In[32]:\n\n\nprint_layer_trainable()\n\n\n# \n# \n# In Transfer Learning we are initially only interested in reusing the pre-trained VGG16 model as it is, so we will disable training for all its layers.\n# \n\n\n\n\nconv_model.trainable = False\n\n\n\n\nfor layer in conv_model.layers:\n    layer.trainable = False\n\n\n\n\n\nprint_layer_trainable()\n\n\n\n\n\nnew_model.compile(optimizer=optimizer, loss=loss, metrics=metrics)\n\n\n\n\nepochs = 15\nsteps_per_epoch = 100\n# Steps per epochs are multiplied with the epoch here 100*20 = 2000  means 2000 random images will be selected.\n\n\n\n\n\nhistory = new_model.fit_generator(generator=generator_train,\n                                  epochs=epochs,\n                                  steps_per_epoch=steps_per_epoch,\n                                  class_weight=class_weight,\n                                  validation_data=generator_test,\n                                  validation_steps=steps_test)\n\n\n\n\n\nnew_model.save(\"trained_new.h5\")\n\n\n\n\n\npredict(image_path = \"\/home\/priyank\/Jupyter_notebook\/pp.jpg\")\n<\/code><\/pre>\n\n<p>**IT is only predicting the 38 classes it is trained on I want, if the new image is not belongs to these 38 classes then the model should return the VGG16 class or no match found. please help ** \nThanks in advance.<\/p>\n","answers":[{"AnswerId":"53411847","CreationDate":"2018-11-21T12:16:04.673","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>use functional api instead of Sequential, \noffical guide is here:  <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/functional-api-guide\/<\/a><\/p>\n\n<p>You can find a multi input &amp; multi output model example there. What you want is very similar but uses only one input instead of multiple. <\/p>\n"}]}
+{"QuestionId":53409547,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-serving><tensorflow-estimator>","CreationDate":"2018-11-21T10:03:11.353","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow Estimator BestExporter event_file_pattern doesn't do anything","Body":"<p>I have a <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> with a very simple demo <code>Estimator<\/code> for the sole purpose of learning \/ understanding the <code>Estimator<\/code> api with the goal of making a convention for a plug-and-play model with common bells and whistles of the trade in tack (e.g. early stopping if the validation set stops improving, exporting the model, etc). <\/p>\n\n<p>In the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub#scrollTo=xGQI18i47BHQ&amp;line=2&amp;uniqifier=1\" rel=\"nofollow noreferrer\">cell<\/a> where I define the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/BestExporter\" rel=\"nofollow noreferrer\"><code>BestExporter<\/code><\/a> I have:<\/p>\n\n<pre><code>exporter = tf.estimator.BestExporter(\n    name=\"best_exporter\",\n    serving_input_receiver_fn=serving_input_receiver_fn,\n    # event_file_pattern=\"model_*\", # &lt;--- doesn't do anything?\n    exports_to_keep=5\n) # this will keep the 5 best checkpoints\n<\/code><\/pre>\n\n<p>which currently exports my best models under<\/p>\n\n<pre><code>&lt;MODEL_DIR&gt;\/export\/best_exporter\/&lt;some_number&gt;\n<\/code><\/pre>\n\n<p>where  <code>MODEL_DIR=.\/test<\/code> (under the cell <code>setup &gt; constants<\/code> and specified in the <code>RunConfig<\/code> set in <code>estimator functions &gt; run config<\/code>)<\/p>\n\n<p>If I uncomment <code>event_file_pattern=\"model_*\",<\/code> I expected to see my model be exported to <code>&lt;MODEL_DIR&gt;\/export\/best_exporter\/model_&lt;number&gt;<\/code>, but it doesn't. <\/p>\n\n<p>So how do I specify what my models should be called?<\/p>\n\n<p>For example, in the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> I track the metrics <code>mse<\/code> and <code>mae<\/code>.<\/p>\n\n<p>How would I get the 5 best <code>exports_to_keep<\/code> be saved as:<\/p>\n\n<pre><code>MODEL_DIR\/export\/best_exporter\/&lt;model_name&gt;_mae_&lt;mae-rounded-to-3-decimals&gt;\n<\/code><\/pre>\n\n<p>?<\/p>\n","answers":[]}
+{"QuestionId":53409699,"AnswerCount":0,"Tags":"<tensorflow><gpu>","CreationDate":"2018-11-21T10:11:27.653","AcceptedAnswerId":null,"OwnerUserId":3145671.0,"Title":"Why the GPU performance changed so much for different data types","Body":"<p>Use a piece of code to test the different of performance with CPU and GPU, I found the performance diff is very interesting.\n<p>\nfloat16 has the worst performance, but the improvement of GPU is most obvious.\n<a href=\"https:\/\/i.stack.imgur.com\/G6Mk1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G6Mk1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>float32 has the best performance, but the improvement of GPU is limited.\n<a href=\"https:\/\/i.stack.imgur.com\/ezd5D.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ezd5D.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>float64 is better than float16, but gpu is not as good as CPU.\n<a href=\"https:\/\/i.stack.imgur.com\/DQHzn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DQHzn.png\" alt=\"enter image description here\"><\/a>\n<p>\nCan any one here can share some insights on why?<p>\nMy hardware is Xeon E5-1650 + Quadro K620. \n<br>\nThis is my testing code.<\/p>\n\n<pre><code>from __future__ import print_function\nimport matplotlib\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport time\n\ndef get_times(maximum_time):\n\n    device_times = {\n        \"\/gpu:0\":[],\n        \"\/cpu:0\":[]\n    }\n    matrix_sizes = range(500,50000,50)\n\n    for size in matrix_sizes:\n        for device_name in device_times.keys():\n\n            print(\"####### Calculating on the \" + device_name + \" #######\")\n\n            shape = (size,size)\n            data_type = tf.float16\n            with tf.device(device_name):\n                r1 = tf.random_uniform(shape=shape, minval=0, maxval=1, dtype=data_type)\n                r2 = tf.random_uniform(shape=shape, minval=0, maxval=1, dtype=data_type)\n                dot_operation = tf.matmul(r2, r1)\n\n\n            with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as session:\n                    start_time = time.time()\n                    result = session.run(dot_operation)\n                    time_taken = time.time() - start_time\n                    print(result)\n                    device_times[device_name].append(time_taken)\n\n            print(device_times)\n\n            if time_taken &gt; maximum_time:\n                return device_times, matrix_sizes\n\n\ndevice_times, matrix_sizes = get_times(1.5)\ngpu_times = device_times[\"\/gpu:0\"]\ncpu_times = device_times[\"\/cpu:0\"]\n\nplt.plot(matrix_sizes[:len(gpu_times)], gpu_times, 'o-')\nplt.plot(matrix_sizes[:len(cpu_times)], cpu_times, 'o-')\nplt.ylabel('Time')\nplt.xlabel('Matrix size')\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53410205,"AnswerCount":1,"Tags":"<python><machine-learning><keras><generator><conv-neural-network>","CreationDate":"2018-11-21T10:36:19.850","AcceptedAnswerId":53410663.0,"OwnerUserId":9968404.0,"Title":"Image generator missing positional argument for unet keras","Body":"<p>I keep getting the following error for below code when I try to train the model: <code>TypeError: fit_generator() missing 1 required positional argument: 'generator'<\/code>. For the life of me I can not figure out what is causing this error. x_train is an rgb image of shape (400, 256, 256, 3) and for y_train i have 10 output classes making it shape (400, 256, 256, 10). What is going wrong here?<\/p>\n\n<p>If necessary the data can be downloaded with the following link:\n<a href=\"https:\/\/www49.zippyshare.com\/v\/5pR3GPv3\/file.html\" rel=\"nofollow noreferrer\">https:\/\/www49.zippyshare.com\/v\/5pR3GPv3\/file.html<\/a><\/p>\n\n<pre><code>import skimage\nfrom skimage.io import imread, imshow, imread_collection, concatenate_images\nfrom skimage.transform import resize\nfrom skimage.morphology import label\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom keras.models import Model\nfrom keras.layers import Input, merge, Convolution2D, MaxPooling2D, UpSampling2D, Reshape, core, Dropout\nfrom keras.optimizers import Adam\nfrom keras.callbacks import ModelCheckpoint, LearningRateScheduler\nfrom keras import backend as K\nfrom sklearn.metrics import jaccard_similarity_score\nfrom shapely.geometry import MultiPolygon, Polygon\nimport shapely.wkt\nimport shapely.affinity\nfrom collections import defaultdict\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.utils.np_utils import to_categorical\nfrom keras import utils as np_utils\nimport os\nfrom keras.preprocessing.image import ImageDataGenerator\ngen = ImageDataGenerator()\n#Importing image and labels\nlabels = skimage.io.imread(\"ede_subset_293_wegen.tif\")\nimages = skimage.io.imread(\"ede_subset_293_20180502_planetscope.tif\")[...,:-1]\n\n\n#scaling image\nimg_scaled = images \/ images.max()\n\n#Make non-roads 0\nlabels[labels == 15] = 0\n\n#Resizing image and mask and labels\nimg_scaled_resized = img_scaled[:6400, :6400 ]\nprint(img_scaled_resized.shape)\nlabels_resized = labels[:6400, :6400]\nprint(labels_resized.shape)\n\n#splitting images\nsplit_img = [\n    np.split(array, 25, axis=0) \n    for array in np.split(img_scaled_resized, 25, axis=1)\n]\n\nsplit_img[-1][-1].shape\n\n#splitting labels\nsplit_labels = [\n    np.split(array, 25, axis=0) \n    for array in np.split(labels_resized, 25, axis=1)\n]\n\n#Convert to np.array\nsplit_labels = np.array(split_labels)\nsplit_img = np.array(split_img)\n\ntrain_images = np.reshape(split_img, (625, 256, 256, 3))\ntrain_labels = np.reshape(split_labels, (625, 256, 256, 10))\n\ntrain_labels = np_utils.to_categorical(train_labels, 10)\n\n#Create train test and val\nx_train = train_images[:400,:,:,:]\nx_val = train_images[400:500,:,:,:]\nx_test = train_images[500:625,:,:,:]\ny_train = train_labels[:400,:,:]\ny_val = train_labels[400:500,:,:]\ny_test = train_labels[500:625,:,:]\n\n# Create image generator (credit to Ioannis Nasios)\ndata_gen_args = dict(rotation_range=5,\n                     width_shift_range=0.1,\n                     height_shift_range=0.1,\n                     validation_split=0.2)\nimage_datagen = ImageDataGenerator(**data_gen_args)\n\nseed = 1\nbatch_size = 100\n\ndef XYaugmentGenerator(X1, y, seed, batch_size):\n    genX1 = gen.flow(X1, y, batch_size=batch_size, seed=seed)\n    genX2 = gen.flow(y, X1, batch_size=batch_size, seed=seed)\n    while True:\n        X1i = genX1.next()\n        X2i = genX2.next()\n\n        yield X1i[0], X2i[0]\n\n\n# Train model\nModel.fit_generator(XYaugmentGenerator(x_train, y_train, seed, batch_size), steps_per_epoch=np.ceil(float(len(x_train)) \/ float(batch_size)),\n                validation_data = XYaugmentGenerator(x_val, y_val,seed, batch_size), \n                validation_steps = np.ceil(float(len(x_val)) \/ float(batch_size))\n, shuffle=True, epochs=20)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53410663","CreationDate":"2018-11-21T11:00:13.240","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>You have a few mistakes in your code, but considering your error:<\/p>\n\n<blockquote>\n  <p>TypeError: fit_generator() missing 1 required positional argument:\n  'generator'<\/p>\n<\/blockquote>\n\n<p>this is caused because fit_generator call XYaugmentGenerator but no augmentation generator is called inside.<\/p>\n\n<pre><code>gen.flow(...\n<\/code><\/pre>\n\n<p>won't work because gen is not declared. You should either rename image_datagen to gen as:<\/p>\n\n<pre><code>gen = ImageDataGenerator(**data_gen_args)\n<\/code><\/pre>\n\n<p>or, replace gen with image_datagen<\/p>\n\n<pre><code>genX1 = image_datagen.flow(X1, y, batch_size=batch_size, seed=seed)\ngenX2 = image_datagen.flow(y, X1, batch_size=batch_size, seed=seed)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53410413,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-21T10:47:29.050","AcceptedAnswerId":null,"OwnerUserId":7673365.0,"Title":"which file contains my deep learning saved model?","Body":"<p><br> I have trained a deep learning model using tensorflow and i saved it as an \"cnn.h5\" file using keras. Now I have 3 files that have \"cnn.h5\" in their name but all of them contain a different extension. <br>The three files are:\n<br>cnn.h5.meta\n<br>cnn.h5.index\n<br>CNN.h5.data-00000-of-00001<\/p>\n\n<p>now can anyone tell me which one of the above files is the saved model? i have to load that  model in my GUI for testing.\n<br> Thanks.<\/p>\n","answers":[{"AnswerId":"53414992","CreationDate":"2018-11-21T15:09:06.337","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>This <a href=\"https:\/\/stackabuse.com\/tensorflow-save-and-restore-models\/\" rel=\"nofollow noreferrer\">blog<\/a> post explains saving and restoring you can refer to it for details  a snippet of explanation for the type of saved files is as below.  <\/p>\n\n<blockquote>\n  <p>When saving the model, you'll notice that it takes 4 types of files to save it:<\/p>\n  \n  <blockquote>\n    <blockquote>\n      <p>\".meta\" files: containing the graph structure<\/p>\n      \n      <p>\".data\" files: containing the values of variables<\/p>\n      \n      <p>\".index\" files: identifying the checkpoint<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n"}]}
+{"QuestionId":53410419,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-21T10:48:00.430","AcceptedAnswerId":53411638.0,"OwnerUserId":1643257.0,"Title":"tf.keras.layers.Concatenate() works with a list but fails on a tuple of tensors","Body":"<p>This will work: <\/p>\n\n<pre><code>tf.keras.layers.Concatenate()([features['a'], features['b']])\n<\/code><\/pre>\n\n<p>While this:<\/p>\n\n<pre><code>tf.keras.layers.Concatenate()((features['a'], features['b']))\n<\/code><\/pre>\n\n<p>Results in:<\/p>\n\n<pre><code>TypeError: int() argument must be a string or a number, not 'TensorShapeV1'\n<\/code><\/pre>\n\n<p>Is that expected? If so - why does it matter what sequence do I pass?<\/p>\n\n<p>Thanks,\nZach<\/p>\n\n<p><strong>EDIT (adding a code example):<\/strong><\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\n\ndata = {\n    'a': [1.0, 2.0, 3.0],\n    'b': [0.1, 0.3, 0.2],\n}\n\nwith tf.Session() as sess: \n  ds = tf.data.Dataset.from_tensor_slices(data)\n  ds = ds.batch(1)\n\n  it = ds.make_one_shot_iterator()\n  features = it.get_next()\n\n  concat = tf.keras.layers.Concatenate()((features['a'], features['b']))\n\n\n  try:\n    while True:\n      print(sess.run(concat))\n  except tf.errors.OutOfRangeError:\n    pass\n\n\n---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-135-0e1a45017941&gt; in &lt;module&gt;()\n      6   features = it.get_next()\n      7 \n----&gt; 8   concat = tf.keras.layers.Concatenate()((features['a'], features['b']))\n      9 \n     10 \n\ngoogle3\/third_party\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    751           # the user has manually overwritten the build method do we need to\n    752           # build it.\n--&gt; 753           self.build(input_shapes)\n    754         # We must set self.built since user defined build functions are not\n    755         # constrained to set self.built.\n\ngoogle3\/third_party\/tensorflow\/python\/keras\/utils\/tf_utils.py in wrapper(instance, input_shape)\n    148             tuple(tensor_shape.TensorShape(x).as_list()) for x in input_shape]\n    149       else:\n--&gt; 150         input_shape = tuple(tensor_shape.TensorShape(input_shape).as_list())\n    151     output_shape = fn(instance, input_shape)\n    152     if output_shape is not None:\n\ngoogle3\/third_party\/tensorflow\/python\/framework\/tensor_shape.py in __init__(self, dims)\n    688       else:\n    689         # Got a list of dimensions\n--&gt; 690         self._dims = [as_dimension(d) for d in dims_iter]\n    691 \n    692   @property\n\ngoogle3\/third_party\/tensorflow\/python\/framework\/tensor_shape.py in as_dimension(value)\n    630     return value\n    631   else:\n--&gt; 632     return Dimension(value)\n    633 \n    634 \n\ngoogle3\/third_party\/tensorflow\/python\/framework\/tensor_shape.py in __init__(self, value)\n    183       raise TypeError(\"Cannot convert %s to Dimension\" % value)\n    184     else:\n--&gt; 185       self._value = int(value)\n    186       if (not isinstance(value, compat.bytes_or_text_types) and\n    187           self._value != value):\n\nTypeError: int() argument must be a string or a number, not 'TensorShapeV1'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53411638","CreationDate":"2018-11-21T12:03:47.430","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/merge.py#L329\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/merge.py#L329<\/a><\/p>\n\n<p>comment on the concanate class states it requires a list. \nthis class calls <code>K.backend<\/code>'s concatenate function\n <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/tensorflow_backend.py#L2041\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/tensorflow_backend.py#L2041<\/a><\/p>\n\n<p>which also states it requires a list. <\/p>\n\n<p>in tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/ops\/array_ops.py#L1034\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/ops\/array_ops.py#L1034<\/a><\/p>\n\n<p>also states it requires a list of tensors. Why? I don't know. in this function the tensors (variable called \"values\") actually gets checked if its a list or tuple. but somewhere along the way you still get an error. <\/p>\n"}]}
+{"QuestionId":53410469,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2018-11-21T10:50:01.450","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow Estimator ServingInputReceiver features vs receiver_tensors: when and why?","Body":"<p>In a <a href=\"https:\/\/stackoverflow.com\/questions\/52874647\/tensorflow-v1-10-why-is-an-input-serving-receiver-function-needed-when-checkpoi\">previous question<\/a> the purpose and structure of the <code>serving_input_receiver_fn<\/code> is explored and in the <a href=\"https:\/\/stackoverflow.com\/a\/52876786\/5623899\">answer<\/a>:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  \"\"\"For the sake of the example, let's assume your input to the network will be a 28x28 grayscale image that you'll then preprocess as needed\"\"\"\n  input_images = tf.placeholder(dtype=tf.uint8,\n                                         shape=[None, 28, 28, 1],\n                                         name='input_images')\n  # here you do all the operations you need on the images before they can be fed to the net (e.g., normalizing, reshaping, etc). Let's assume \"images\" is the resulting tensor.\n\n  features = {'input_data' : images} # this is the dict that is then passed as \"features\" parameter to your model_fn\n  receiver_tensors = {'input_data': input_images} # As far as I understand this is needed to map the input to a name you can retrieve later\n  return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n<\/code><\/pre>\n\n<p>the <a href=\"https:\/\/stackoverflow.com\/a\/52876786\/5623899\">answer<\/a>'s author states (in regards to <code>receiver_tensors<\/code>):<\/p>\n\n<blockquote>\n  <p>As far as I understand this is needed to map the input to a name you can retrieve later<\/p>\n<\/blockquote>\n\n<p>This distinction is unclear to me. In practice, (see this <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"noreferrer\">colab<\/a>), the same dictionary can be passed to both <code>features<\/code> and <code>receiver_tensors<\/code>.<\/p>\n\n<p>From the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/estimator\/export\/export.py\" rel=\"noreferrer\">source code<\/a> of <code>@estimator_export('estimator.export.ServingInputReceiver')<\/code> (or the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/ServingInputReceiver\" rel=\"noreferrer\">ServingInputReceiver docs<\/a>:<\/p>\n\n<blockquote>\n  <ul>\n  <li><strong>features<\/strong>: A <code>Tensor<\/code>, <code>SparseTensor<\/code>, or dict of string to <code>Tensor<\/code> or\n    <code>SparseTensor<\/code>, specifying the features to be passed to the model. Note:\n    if <code>features<\/code> passed is not a dict, it will be wrapped in a dict with a\n    single entry, using 'feature' as the key.  Consequently, the model must\n    accept a feature dict of the form {'feature': tensor}.  You may use\n    <code>TensorServingInputReceiver<\/code> if you want the tensor to be passed as is.<\/li>\n  <li><strong>receiver_tensors<\/strong>: A <code>Tensor<\/code>, <code>SparseTensor<\/code>, or dict of string to <code>Tensor<\/code>\n    or <code>SparseTensor<\/code>, specifying input nodes where this receiver expects to\n    be fed by default.  Typically, this is a single placeholder expecting\n    serialized <code>tf.Example<\/code> protos.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>After reading, it is clear to me what the purposes of <code>features<\/code> is. <code>features<\/code> is a dictionary of inputs that I then send through the graph. Many common models have just a single input, but you can or course have more. <\/p>\n\n<p>So then the statement regarding <code>receiver_tensors<\/code> which \"Typically, this is a single placeholder expecting serialized <code>tf.Example<\/code> protos.\", to me, suggests that <code>receiver_tensors<\/code> want a singular batched placeholder for <code>(Sequence)Example<\/code>s parsed from TF <code>Record<\/code>s. <\/p>\n\n<p>Why? If the TF <code>Record<\/code>s is fully preprocessed, then this is redundant? if it is not fully pre-processed, why would one pass it? Should the keys in the <code>features<\/code> and <code>receiver_tensors<\/code> dictionaries be the same?<\/p>\n\n<p>Can someone please provide me with a more concrete example of the difference and what goes where, as right now<\/p>\n\n<pre><code>input_tensors = tf.placeholder(tf.float32, &lt;shape&gt;, name=\"input_tensors\")\nfeatures = receiver_tensors =  {'input_tensors': input_tensors}\n<\/code><\/pre>\n\n<p>works... (even if maybe it shouldn't...)<\/p>\n","answers":[{"AnswerId":"54986315","CreationDate":"2019-03-04T15:25:21.060","ParentId":null,"OwnerUserId":"3135116","Title":null,"Body":"<p>If you do the preprocessing inside TensorServingInputReceiver than receiver_tensors and features would be different. features will be passed to the model after the preprocessing inside TensorServingInputReceiver has been made. receiver_tensors are the input for the TensorServingInputReceiver and they can be in a tf.Example format<\/p>\n"},{"AnswerId":"58257237","CreationDate":"2019-10-06T12:28:55.737","ParentId":null,"OwnerUserId":"11652623","Title":null,"Body":"<p>The job of the serving input function is to convert the <strong>received raw features<\/strong> into the <strong>processed features<\/strong> which your model function accepts.<\/p>\n\n<p><strong><em><code>receiver_tensors<\/code> : These are the input placeholders. This is opening in your graph where you will receive your raw input features.<\/em><\/strong><\/p>\n\n<p>After defining this placeholder you perform transformations on these receiver tensors to convert them into features which are model acceptable. Some of these transformations will include:<\/p>\n\n<ul>\n<li>Pre-processing received data.<\/li>\n<li>Parsing example from tfrecord. (In case you are providing tfrecord as input to serving function)<\/li>\n<\/ul>\n\n<p><strong><em><code>features<\/code> : Once you transform receive tensors features are obtained which are directly fed to your model function during prediction.<\/em><\/strong><\/p>\n\n<p>In your case pre-processing is not required for the data which you are providing to your serving input function. Hence <code>features = receiver_tensors<\/code> is working.<\/p>\n"}]}
+{"QuestionId":53410753,"AnswerCount":2,"Tags":"<python><tensorflow><pip><anaconda><fancyimpute>","CreationDate":"2018-11-21T11:05:30.223","AcceptedAnswerId":null,"OwnerUserId":10685202.0,"Title":"Anaconda installing tensorflow and fancyimpute","Body":"<p>As a premise, I would specify that I am new to Python, so please forgive eventual inaccuracies.<\/p>\n\n<p>So, I have recently installed Anaconda, and updated the Python version to 3.7.1.\nIn order to impute some missing values in my dataset using KNN, I've found a useful function in a package called fancyimpute.<\/p>\n\n<p>However, such package is not among those already available (that is, from Spyder, the IDE I'm using, I cannot simply import it), so I need to install it.<\/p>\n\n<p>Opening, as such, the Anaconda prompt and typing \"conda install fancyimpute\" doesn't work, returning the following:<\/p>\n\n<p>\"PackagesNotFoundError: The following packages are not available from current channels:<\/p>\n\n<ul>\n<li>fancyimpute<\/li>\n<\/ul>\n\n<p>Current channels:<\/p>\n\n<p>(here a list of some channels)<\/p>\n\n<p>To search for alternate channels that may provide the conda package you're\nlooking for, navigate to<\/p>\n\n<p><a href=\"https:\/\/anaconda.org\" rel=\"nofollow noreferrer\">https:\/\/anaconda.org<\/a><\/p>\n\n<p>and use the search bar at the top of the page\".<\/p>\n\n<p>Going, therefore, with \"pip install fancyimpute\" (although to me it's still not clear the difference between conda install and pip install) after a while I obtain the following:<\/p>\n\n<p>\"Collecting tensorflow (from fancyimpute)<\/p>\n\n<p>Could not find a version that satisfies the requirement tensorflow (from fancyimpute) (from versions: )\nNo matching distribution found for tensorflow (from fancyimpute)\"<\/p>\n\n<p>I have therefore now run \"conda install tensorflow\", and is already about 30 minutes that is Solving environment.<\/p>\n\n<p>What can I do? How can I obtain the desired package and avoid similar problems in future? Many thanks and hope I was clear in exposing the problem.<\/p>\n\n<p>UPDATE: <a href=\"https:\/\/anaconda.org\/search?q=fancyimpute\" rel=\"nofollow noreferrer\">https:\/\/anaconda.org\/search?q=fancyimpute<\/a> From here, it seems that fancyimpute isn't available on my platform, w-64. How can I overcome this problem?<\/p>\n","answers":[{"AnswerId":"53410983","CreationDate":"2018-11-21T11:18:35.057","ParentId":null,"OwnerUserId":"148680","Title":null,"Body":"<p>User <em>brittainhard<\/em> on <a href=\"https:\/\/anaconda.org\/brittainhard\/fancyimpute\" rel=\"nofollow noreferrer\">anaconda.org had the same idea<\/a>. To use his\/her version of the library (hosted on anaconda.org):<\/p>\n\n<pre><code>conda install -c brittainhard fancyimpute\n<\/code><\/pre>\n"},{"AnswerId":"53417778","CreationDate":"2018-11-21T17:40:34.043","ParentId":null,"OwnerUserId":"10685202","Title":null,"Body":"<p>SOLVED: Apparently, I have solved the problem. \nI have first created an ad hoc environment and installed tensorflow using conda.\nThen, I have pip installed fancyimpute: at this point, got a couple of new errors (\"Failed building wheel for fastacache\", and same for cvxpy), both solved installing Microsoft Visual C++ Build Tools. So, finally, I have been able to install also fancyimpute.\nNonetheless, at this point, I coulnd't import it (ImportError: DLL load failed: The specified module could not be found. Failed to load the native TensorFlow runtime). After uninstalling and reinstalling tensorflow using conda-forge as channel, now it works.<\/p>\n"}]}
+{"QuestionId":53410952,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><computer-vision><conv-neural-network>","CreationDate":"2018-11-21T11:16:51.383","AcceptedAnswerId":null,"OwnerUserId":4462831.0,"Title":"How to sweep a neural-network through an image with tensorflow?","Body":"<p>My question is about finding an efficient (mostly in term of parameters count) way to implement a sliding window in tensorflow (1.4) in order to apply a neural network through the image and produce a 2-d map with each pixel (or region) representing the network output for the corresponding receptive field (which in this case is the sliding window itself).<\/p>\n\n<p>In practice, I'm trying to implement either a <a href=\"https:\/\/www.ncbi.nlm.nih.gov\/pubmed\/12906178\" rel=\"nofollow noreferrer\">MTANN<\/a> or a <a href=\"https:\/\/arxiv.org\/abs\/1604.04382\" rel=\"nofollow noreferrer\">PatchGAN<\/a> using tensorflow, but I cannot understand the implementation I found.<\/p>\n\n<p>The two architectures can be briefly described as:<\/p>\n\n<ul>\n<li><p>MTANN: A linear neural network with input size of [1,N,N,1] and output size [ ] is applied to an image of size [1,M,M,1] to produce a map of size [1,G,G,1], in which every pixel of the generated map corresponds to a likelihood of the corresponding NxN patch to belong to a certain class.<\/p><\/li>\n<li><p>PatchGAN Discriminator: More general architecture, as I can understand the network that is strided through the image outputs a map itself instead of a single value, which then is combined with adjacent maps to produce the final map.<\/p><\/li>\n<\/ul>\n\n<p>While I cannot find any tensorflow implementation of MTANN, I found the PatchGAN <a href=\"https:\/\/github.com\/affinelayer\/pix2pix-tensorflow\/blob\/master\/pix2pix.py#L395\" rel=\"nofollow noreferrer\">implementation<\/a>, which is <a href=\"https:\/\/github.com\/junyanz\/pytorch-CycleGAN-and-pix2pix\/issues\/39\" rel=\"nofollow noreferrer\">considered as a convolutional network<\/a>, but I couldn't figure out how to implement this in practice.<\/p>\n\n<p>Let's say I got a pre-trained network of which I got the output tensor. I understand that convolution is the way to go, since a convolutional layer operates over a local region of the input and what is I'm trying to do can be clearly represented as a convolutional network. However, what if I already have the network that generates the sub-maps from a given window of fixed-size?<\/p>\n\n<p>E.g. I got a tensor<\/p>\n\n<pre><code>sub_map = network(input_patch)\n<\/code><\/pre>\n\n<p>which returns a [1,2,2,1] maps from a [1,8,8,3] image (corresponding to a 3-layer FCN with input size 8, filter size 3x3).\nHow can I sweep this network on [1,64,64,3] images, in order to produce a [1,64,64,1] map composed of each spatial contribution, like it happens in a convolution?<\/p>\n\n<p>I've considered these solutions:<\/p>\n\n<ul>\n<li><p>Using <code>tf.image.extract_image_patches<\/code> which explicitly extract all the image patches and channels in the depth dimension, but I think it would consume too many resources, as I'm switching to PatchGAN Discriminator from a full convolutional network due to memory constraints - also the composition of the final map is not so straight-forward. <\/p><\/li>\n<li><p>Adding a convolutional layer before the network I got, but I cannot figure out what the filter (and its size) should be in this case in order to keep the pretrained model work on 8x8 images while integrating it in a model which works on bigger images. \nFor what I can get it should be something like <code>whole_map = tf.nn.convolution(input=x64_images, filter=sub_map, ...)<\/code> but I don't think this would work as the filter is an operator which depends on the receptive field itself.<\/p><\/li>\n<\/ul>\n\n<p>The ultimate goal is to apply this small network to big images (eg. 1024x1024) in an efficient way, since my current model downscales progressively the images and wouldn't fit in memory due to the huge number of parameters. <\/p>\n\n<p>Can anyone help me to get a better understanding of what I am missing?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"53518013","CreationDate":"2018-11-28T11:06:54.083","ParentId":null,"OwnerUserId":"4462831","Title":null,"Body":"<p>I found an interesting <a href=\"https:\/\/www.coursera.org\/lecture\/convolutional-neural-networks\/convolutional-implementation-of-sliding-windows-6UnU4\" rel=\"nofollow noreferrer\">video<\/a> by Andrew Ng exactly on how to implement a sliding window using a convolutional layer. \nThe problem here was that I was thinking at the number of layers as a variable that is dependent on a fixed input\/output shape, while it should be the opposite.<\/p>\n\n<p>In principle, a saved model should only contain the learned filters for each level and as long as the filter shapes are compatible with the layers' input\/output depth. Thus, applying a different (ie. bigger) spatial resolution to the network input produces a different output shape, which can be seen as an application of the neural network to a sliding windows sweeping across the input image.<\/p>\n"}]}
+{"QuestionId":53411097,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T11:26:55.523","AcceptedAnswerId":null,"OwnerUserId":714011.0,"Title":"Extremely high CPU usage when training a model on GPU","Body":"<p>Recently I discovered something rather strange with the project that I worked on for quite a while already. The model I have is rather conventional: a convnet with a few fully connected layers. For data loading I use tf.data API, but the same thing happens with queue-based code that I had before porting to tf.data. After a few hours since the training of the model begins, the CPU usage rises to very high levels, 1500-2000% as reported by the <code>htop<\/code> util. And at the beginning of training everything is fine, the main process shows only about 200% CPU usage. Attached is the screenshot of the <code>htop<\/code> output, and another thing that's worrying is all the child processes that also have pretty high CPU load.<\/p>\n\n<p>I am using tensorflow-gpu version 1.11, running it on NVIDIA Tesla V100. I am pretty sure that the model does run on the GPU and not on the CPU: nvidia-smi shows that GPU is occupied at an about 70% rate.<\/p>\n\n<p>Obviously, I cannot ask for an exact cause of this, and it would be difficult to strip the problem down to a reproducible test case. However, may be you could point me at some debugging techniques that are applicable in such case.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rfn6N.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rfn6N.png\" alt=\"htop output\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53411566,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><keras-layer>","CreationDate":"2018-11-21T11:58:16.710","AcceptedAnswerId":53411908.0,"OwnerUserId":10254228.0,"Title":"Keras: input with size x*x generates unwanted output y*x","Body":"<p>I have the following neural network in Keras:<\/p>\n\n<pre><code>inp = layers.Input((3,))\n#Middle layers omitted\nout_prop = layers.Dense(units=3, activation='softmax')(inp)\nout_value = layers.Dense(units=1, activation = 'linear')(inp)\n<\/code><\/pre>\n\n<p>Then I prepared a pseudo-input to test my network:<\/p>\n\n<pre><code>inpu = np.array([[1,2,3],[4,5,6],[7,8,9]])\n<\/code><\/pre>\n\n<p>When I try to predict, this happens:<\/p>\n\n<pre><code>In [45]:nn.network.predict(inpu)\nOut[45]: \n[array([[0.257513  , 0.41672954, 0.32575747],\n    [0.20175152, 0.4763418 , 0.32190666],\n    [0.15986516, 0.53449154, 0.30564335]], dtype=float32),\narray([[-0.24281949],\n    [-0.10461146],\n    [ 0.11201331]], dtype=float32)]\n<\/code><\/pre>\n\n<p>So, as you can see above, I wanted two output: one should have been an array with size 3, the other should have been a normal value. Instead, I get a 3x3 matrix, and an array with 3 elements. What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"53411908","CreationDate":"2018-11-21T12:20:34.380","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You are passing three input samples to the network:<\/p>\n\n<pre><code>&gt;&gt;&gt; inpu.shape\n(3,3)  # three samples of size 3\n<\/code><\/pre>\n\n<p>And you have two output layers: one of them outputs a vector of size 3 <strong>for each sample<\/strong> and the other outputs a vector of size one (i.e. scalar), again <strong>for each sample<\/strong>. As a result the output shapes would be <code>(3, 3)<\/code> and <code>(3, 1)<\/code>.<\/p>\n\n<p><strong>Update:<\/strong> If you want your network to accept an input sample of shape <code>(3,3)<\/code> and outputs vectors of size 3 and 1, and you want to only use Dense layers in your network, then you must use a <code>Flatten<\/code> layer somewhere in the model. One possible option is to use it right after the input layer:<\/p>\n\n<pre><code>inp = layers.Input((3,3))  # don't forget to set the correct input shape\nx = Flatten()(inp)\n# pass x to other Dense layers\n<\/code><\/pre>\n\n<p>Alternatively, you could flatten your data to have a shape of <code>(num_samples, 9)<\/code> and then pass it to your network without using a <code>Flatten<\/code> layer.<\/p>\n\n<p><strong>Update 2:<\/strong> As @Mete correctly pointed out in the comments, make sure the input array have a shape of <code>(num_samples, 3, 3)<\/code> if each input sample has a shape of <code>(3,3)<\/code>.<\/p>\n"}]}
+{"QuestionId":53412391,"AnswerCount":0,"Tags":"<python><tensorflow><scope>","CreationDate":"2018-11-21T12:48:48.220","AcceptedAnswerId":null,"OwnerUserId":8263839.0,"Title":"Tensor Flow increment nested variable_scope","Body":"<p>I know I can increment a variable_scope using the 'default_name' argument:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.variable_scope(\"A\") # This is scope \"A\"\ntf.variable_scope(None, \"A\") # incremented scope \"A_1\"\n<\/code><\/pre>\n\n<p>However, this no longer works when an outer context is re-entered<\/p>\n\n<pre><code>reuse= tf.AUTO_REUSE\nwith tf.variable_scope(\"A\", reuse=reuse):\n    with tf.variable_scope(\"B\", reuse=tf.AUTO_REUSE):\n        print tf.get_variable(\"x\", (), tf.float32) # 'A\/B\/x:0'        \n    with tf.variable_scope(None, \"B\"): # Increment B, as expected\n        print tf.get_variable(\"x\", (), tf.float32) # 'A\/B_1\/x:0'\n# Re-enter A and try to increment B\nwith tf.variable_scope(\"A\", reuse=reuse):    \n    with tf.variable_scope(None, \"B\"): # Does not increment B !!!\n        print tf.get_variable(\"x\", (), tf.float32) # 'A\/B\/x:0' !!!\n<\/code><\/pre>\n\n<ul>\n<li>Is there a way to increment \"B\" after re-entering \"A\" ?<\/li>\n<li>The re-entered context shares its variable with the initial context A, but not the way it increments its inner context. I find this very confusing, and wonder about the rationale.<\/li>\n<\/ul>\n\n<p>Thank you !<\/p>\n","answers":[]}
+{"QuestionId":53413407,"AnswerCount":0,"Tags":"<python><tensorflow><lstm><tensorflow-datasets>","CreationDate":"2018-11-21T13:42:23.830","AcceptedAnswerId":null,"OwnerUserId":3767336.0,"Title":"Processing time series with the Tensorflow Dataset API for feeding an LSTM in batches","Body":"<p>I have a working solution for my problem, but I find it relatively slow, and I would be surprised if there is no better way of doing what I want. See the description of the problem below.<\/p>\n\n<p><strong>CURRENT SOLUTION:<\/strong> (comes first, as it might be simpler for experienced users of tensorflow to infer what I want to do from this code snippet)<\/p>\n\n<pre><code>def stack_batch_dim(*x):\n    return {key: tf.stack([b[key] for b in x], axis=1) for key in x[0]}\n\n\ndataset = tf.data.TFRecordDataset(filenames)\ndataset = dataset.map(map_func=parse_box2d_protobuf, num_parallel_calls=8)\ndataset = dataset.repeat()\n\n# place the \"cursors\":\ndatasets = [dataset.skip(t * inter_cursor_space) for t in range(batch_size)]\ndataset = tf.data.Dataset.zip(tuple(datasets))\n\n# consume time_depth samples at a time\ndataset = dataset.batch(time_depth)\n\n# the actual batch dimension is a tuple and not a tensor\n# need to \"stack\" the tuple into a tensor\ndataset = dataset.map(stack_batch_dim)\niterator = dataset.make_initializable_iterator()\n<\/code><\/pre>\n\n<hr>\n\n<p>So I would like train an LSTM network with a sequence of data points. The data I have is stored in tfrecords, (the recommended way of storing data with tensorflow).<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset(filenames)\ndataset = dataset.map(map_func=parse_box2d_protobuf, num_parallel_calls=8)\n\n# parse_box2d_protobuf is a function that transforms raw bytes into\n# a python structure of the form:\n# {\"positions\": tensor_of_shape_4,\n#  \"speeds\": tensor_of_shape_4,\n#  \"action\": tensor_of_shape_4,\n#  \"vision\": tensor_of_shape_h_w_3,\n#  \"index\": tensor_int64}\n\n\niterator = dataset.make_initializable_iterator()\nnext_sample = iterator.get_next()\nwith tf.Session() as sess:\n    sess.run(iterator.initializer)\n    print(sess.run(next_sample))\n    # prints a dictionary as expected.\n<\/code><\/pre>\n\n<p>Now what I want to do is to read that sequence in order to feed it in an LSTM in batches.\nAs you might expect, the \"index\" key in the dictionary corresponds to an integer that is the index of the data sample in the sequence (starting from 0):<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(iterator.initializer)\n    for i in range(10, 15):\n        print(sess.run(next_sample[\"index\"]))\n        # prints successively 0, 1, 2, 3 and 4\n<\/code><\/pre>\n\n<p>Now let's say I want a batch size of 4, and I want to unfold the LSTM net for 3 iterations. This means I need to read the sequence at 4 positions simultaneously (4 cursors), each cursor reading samples 3 by 3. Let's add a parameter that controls the space between the cursors:<\/p>\n\n<pre><code>batch_size = 4\ntime_depth = 3\ninter_cursor_space = 5\n<\/code><\/pre>\n\n<p>How can I define an iterator such that calling <code>iterator.get_next()[\"index\"]<\/code> in a session returns<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(iterator.initializer)\n\n    # first call:\n    print(sess.run(next_sample[\"index\"]))\n\n    # should print\n    # [[0, 5, 10, 15],\n    #  [1, 6, 11, 16],\n    #  [2, 7, 12, 17]]\n\n    # second call:\n    print(sess.run(next_sample[\"index\"]))\n\n    # should print\n    # [[3, 8, 13, 18],\n    #  [4, 9, 14, 19],\n    #  [5, 10, 15, 20]]\n\n    # shape = [time_depth, batch_size]\n<\/code><\/pre>\n\n<p>Such the this iterator is compatible with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/static_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.static_rnn<\/code><\/a> function.<\/p>\n\n<hr>\n","answers":[]}
+{"QuestionId":53413490,"AnswerCount":2,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2018-11-21T13:46:54.753","AcceptedAnswerId":null,"OwnerUserId":10492592.0,"Title":"IndexError: list index out of range in TensorFlow","Body":"<p>I got an error ,IndexError: list index out of range.Traceback says<\/p>\n\n<pre><code>Run id: P0W5X0\nLog directory: \/tmp\/tflearn_logs\/\nException in thread Thread-2:\nTraceback (most recent call last):\n  File \"\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/threading.py\", line 801, in __bootstrap_inner\n    self.run()\n  File \"\/usr\/local\/Cellar\/python@2\/2.7.15\/Frameworks\/Python.framework\/Versions\/2.7\/lib\/python2.7\/threading.py\", line 754, in run\n    self.__target(*self.__args, **self.__kwargs)\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tflearn\/data_flow.py\", line 201, in fill_batch_ids_queue\n    ids = self.next_batch_ids()\n  File \"\/Users\/xxx\/anaconda\/xxx\/lib\/python2.7\/site-packages\/tflearn\/data_flow.py\", line 215, in next_batch_ids\n    batch_start, batch_end = self.batches[self.batch_index]\nIndexError: list index out of range\n<\/code><\/pre>\n\n<p>I wrote codes,<\/p>\n\n<pre><code># coding: utf-8\nimport tensorflow as tf\nimport tflearn\n\nfrom tflearn.layers.core import input_data,dropout,fully_connected\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.normalization import local_response_normalization\nfrom tflearn.layers.estimator import regression\n\ntf.reset_default_graph()\nnet = input_data(shape=[None,20000, 4, 42])\nnet = conv_2d(net, 4, 16, activation='relu')\nnet = max_pool_2d(net, 1)\nnet = tflearn.activations.relu(net)\nnet = dropout(net, 0.5)\nnet = tflearn.fully_connected(net, 2, activation='softmax')\nnet = tflearn.regression(net, optimizer='adam', learning_rate=0.5, loss='categorical_crossentropy')\n\nmodel = tflearn.DNN(net)\n\nmodel.fit(np.array(trainDataSet).reshape(1,20000, 4, 42), np.array(trainLabel), n_epoch=400, batch_size=32, validation_set=0.1, show_metric=True)\n\n\npred = np.array(model.predict(np.array(testDataSet).reshape(1,20000, 4, 42)).argmax(axis=1))\n\nlabel = np.array(testLabel).argmax(axis=0)\naccuracy = np.mean(pred == label, axis=0)\n\nprint(accuracy)\n<\/code><\/pre>\n\n<p>I really cannot understand why such an error happens.I tried to rewrite  into<\/p>\n\n<pre><code>model.fit(np.array(trainDataSet).reshape(1,20000, 4, 42), np.array(trainLabel), n_epoch=400, batch_size=1, validation_set=0.1, show_metric=True) \n<\/code><\/pre>\n\n<p>because bach causes this error,but same error happens.I rewrite another number in this part but same error happens.What is wrong in my codes?How should I fix this?<\/p>\n","answers":[{"AnswerId":"55422442","CreationDate":"2019-03-29T17:13:00.187","ParentId":null,"OwnerUserId":"4023951","Title":null,"Body":"<h1>Question<\/h1>\n\n<p>How can I fix my list index out of range error?<\/p>\n\n<h1>Answer<\/h1>\n\n<p>It appears from your code that your training and test set that you are passing into the neural net only has 1 element given by reshape(1,20000, 4, 42) of shape 20000x4x42.  What I believe you meant is to have 20000 elements of 4x42.<\/p>\n\n<p>Instead of <code>reshape(1,20000, 4, 42)<\/code>, let's use <code>reshape(20000, 4, 42, 1)<\/code>.  We will also have to change <code>input_data(shape=[None, 20000, 4, 42])<\/code> to <code>input_data(shape=[None, 4, 42, 1])<\/code><\/p>\n\n<p>If you do this your code works fine.<\/p>\n\n<h1>Working Code<\/h1>\n\n<pre><code># coding: utf-8\nimport tensorflow as tf\nimport tflearn\n\nfrom tflearn.layers.core import input_data,dropout,fully_connected\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.normalization import local_response_normalization\nfrom tflearn.layers.estimator import regression\n\ntf.reset_default_graph()\nnet = input_data(shape=[None, 4, 42, 1])\nnet = conv_2d(net, 4, 16, activation='relu')\nnet = max_pool_2d(net, 1)\nnet = tflearn.activations.relu(net)\nnet = dropout(net, 0.5)\nnet = tflearn.fully_connected(net, 2, activation='softmax')\nnet = tflearn.regression(net, optimizer='adam', learning_rate=0.5, loss='categorical_crossentropy')\n\nmodel = tflearn.DNN(net)\n\nmodel.fit(np.array(trainDataSet).reshape(20000, 4, 42, 1), np.array(trainLabel), n_epoch=400, batch_size=32, validation_set=0.1, show_metric=True)\n\n\npred = np.array(model.predict(np.array(testDataSet).reshape(20000, 4, 42, 1)).argmax(axis=1))\n\nlabel = np.array(testLabel).argmax(axis=0)\naccuracy = np.mean(pred == label, axis=0)\n\nprint(accuracy)\n<\/code><\/pre>\n\n<h1>Output<\/h1>\n\n<p>To get the above code to work we have to include some training and test data. Numpy random is used like so<\/p>\n\n<pre><code>import numpy as np\n\ntrainDataSet = np.random.rand(20000, 4, 42)\ntrainLabel = ( np.random.rand(20000,2) &gt; .5 ) *1.0\n\ntestDataSet = np.random.rand(20000, 4, 42)\ntestLabel = ( np.random.rand(20000,2) &gt; .5 ) *1.0\n<\/code><\/pre>\n\n<p>Here is the output<\/p>\n\n<pre><code>Run id: JDSG88\nLog directory: \/tmp\/tflearn_logs\/\n---------------------------------\nTraining samples: 18000\nValidation samples: 2000\n--\nTraining Step: 563  | total loss: 12.13387 | time: 5.312s\n| Adam | epoch: 001 | loss: 12.13387 - acc: 0.7138 | val_loss: 11.90437 - val_acc: 0.7400 -- iter: 18000\/18000\n--\nTraining Step: 1126  | total loss: 11.58909 | time: 5.184s\n| Adam | epoch: 002 | loss: 11.58909 - acc: 0.7496 | val_loss: 11.90437 - val_acc: 0.7400 -- iter: 18000\/18000\n--\nTraining Step: 1689  | total loss: 11.93482 | time: 5.174s\n| Adam | epoch: 003 | loss: 11.93482 - acc: 0.7357 | val_loss: 11.90437 - val_acc: 0.7400 -- iter: 18000\/18000\n--\n...\n<\/code><\/pre>\n"},{"AnswerId":"53457829","CreationDate":"2018-11-24T11:50:16.850","ParentId":null,"OwnerUserId":"6184368","Title":null,"Body":"<p>I also had the same problem with you. My solution is making the number of n_epoch equal to your row's number of the dataset. For example, my array's shape is 461*5, the value of the n_epoch is 461. you can also make the value a little bit bigger or shorter than your row's number. In my code, 500 or 400 is also useful. <\/p>\n"}]}
+{"QuestionId":53413505,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2018-11-21T13:47:26.253","AcceptedAnswerId":53413667.0,"OwnerUserId":3800527.0,"Title":"Save and continue training the LSTM network","Body":"<p>I try to make a LSTM model continue were its last run left off. All compiles fine till I try to fit the network. Then it gives an error: <\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking target: expected dense_29 to have 3 dimensions, but got array with shape (672, 1)<\/p>\n<\/blockquote>\n\n<p>I checked various articles such as <a href=\"https:\/\/jovianlin.io\/saving-loading-keras-models\/\" rel=\"nofollow noreferrer\">this<\/a> and <a href=\"https:\/\/keras.io\/models\/about-keras-models\/\" rel=\"nofollow noreferrer\">this<\/a>  but I don't see what is wrong in my code.<\/p>\n\n<pre><code>from keras import Sequential\nfrom keras.preprocessing.sequence import pad_sequences\nfrom sklearn.model_selection import train_test_split\nfrom keras.models import Sequential,Model\nfrom keras.layers import LSTM, Dense, Bidirectional, Input,Dropout,BatchNormalization\nfrom keras import backend as K\nfrom keras.engine.topology import Layer\nfrom keras import initializers, regularizers, constraints\n\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.models import load_model\nimport os.path\nimport os\nfilepath=\"Train-weights.best.hdf5\"\nact = 'relu'\n\nmodel = Sequential()\nmodel.add(BatchNormalization(input_shape=(10, 128)))\nmodel.add(Bidirectional(LSTM(128, dropout=0.5, activation=act, return_sequences=True)))\nmodel.add(Dense(1,activation='sigmoid'))\n\nif (os.path.exists(filepath)):\n   print(\"extending training of previous run\")\n   model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n   with open('model_architecture.json', 'r') as f:\n      model = model_from_json(f.read())\n   model.load_weights(filepath)\nelse:\n   print(\"First run\")      \n   model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n   model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=100, batch_size=32, callbacks=callbacks_list, verbose=2)\n   model.save_weights(filepath)\n   with open('model_architecture.json', 'w') as f:\n       f.write(model.to_json())\n\n checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n callbacks_list = [checkpoint]\n\n model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=100, batch_size=32, callbacks=callbacks_list, verbose=0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53413667","CreationDate":"2018-11-21T13:56:08.670","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Try <code>model.summary()<\/code> and you would see that the output shape of last layer (i.e. the Dense layer) in the network is <code>(None, 10, 1)<\/code>. Therefore, the labels you provide to the model (i.e. <code>y_train<\/code>) must also have a shape of <code>(num_samples, 10, 1)<\/code>.<\/p>\n\n<p>If the output shape <code>(None, 10, 1)<\/code> is not what you wanted (e.g. you want <code>(None, 1)<\/code> as the output shape of your model) then you need to modify your model definition. One simple modification to achieve that is removing the <code>return_sequences=True<\/code> argument from the LSTM layer. <\/p>\n"}]}
+{"QuestionId":53413706,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-21T13:58:30.427","AcceptedAnswerId":53413842.0,"OwnerUserId":9984948.0,"Title":"Large WGAN-GP train loss","Body":"<p>This is the loss function of WGAN-GP<\/p>\n\n<pre><code>gen_sample = model.generator(input_gen)\ndisc_real = model.discriminator(real_image, reuse=False)\ndisc_fake = model.discriminator(gen_sample, reuse=True)\ndisc_concat = tf.concat([disc_real, disc_fake], axis=0)\n# Gradient penalty\nalpha = tf.random_uniform(\n    shape=[BATCH_SIZE, 1, 1, 1],\n    minval=0.,\n    maxval=1.)\ndifferences = gen_sample - real_image\ninterpolates = real_image + (alpha * differences)\ngradients = tf.gradients(model.discriminator(interpolates, reuse=True), [interpolates])[0]    # why [0]\nslopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))\ngradient_penalty = tf.reduce_mean((slopes-1.)**2)\n\nd_loss_real = tf.reduce_mean(disc_real)\nd_loss_fake = tf.reduce_mean(disc_fake)\n\ndisc_loss = -(d_loss_real - d_loss_fake) + LAMBDA * gradient_penalty\ngen_loss = - d_loss_fake\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/oFosH.jpg\" rel=\"nofollow noreferrer\">This is the training loss<\/a><\/p>\n\n<p>The generator loss is oscillating, and the value is so big. \nMy question is:\nis the generator loss normal or abnormal?<\/p>\n","answers":[{"AnswerId":"53413842","CreationDate":"2018-11-21T14:04:52.520","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>One thing to note is that your gradient penalty calculation is wrong. The following line:<\/p>\n\n<pre><code>slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))\n<\/code><\/pre>\n\n<p>should actually be:<\/p>\n\n<pre><code>slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1,2,3]))\n<\/code><\/pre>\n\n<p>You are reducing on the first axis, but the gradient is based on an image as shown by the alpha values and therefore you have to reduce on the axis <code>[1,2,3]<\/code>.<\/p>\n\n<p>Another error in your code is that the generator loss is:<\/p>\n\n<pre><code>gen_loss = d_loss_real - d_loss_fake\n<\/code><\/pre>\n\n<p>For the gradient calculation this makes no difference, due to the parameters of the generator only being contained in d_loss_fake. However, for the value of the generator loss this makes all the difference in the world and is the reason why this oszillates this much.<\/p>\n\n<p>At the end of the day you should look at your actual performance metric you care about to determine the quality of your GAN like the inception score or the  Fr\u00e9chet Inception Distance (FID), because the loss of discriminator and generator are only mildly descriptive.<\/p>\n"}]}
+{"QuestionId":53413756,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-21T14:00:41.297","AcceptedAnswerId":null,"OwnerUserId":6357916.0,"Title":"Setting proxy to let keras \/ tensorflow download required json","Body":"<p>I was trying out this <a href=\"https:\/\/blog.keras.io\/building-a-simple-keras-deep-learning-rest-api.html\" rel=\"nofollow noreferrer\">article<\/a>.<\/p>\n\n<p>It seems that while making prediction, keras fails to download <a href=\"https:\/\/s3.amazonaws.com\/deep-learning-models\/image-models\/imagenet_class_index.json\" rel=\"nofollow noreferrer\">json<\/a> because of corporate proxy at <a href=\"https:\/\/github.com\/jrosebr1\/simple-keras-rest-api\/blob\/6fb7490afca21c5923c8ce5336f3ae99167a8844\/run_keras_server.py#L62\" rel=\"nofollow noreferrer\">line<\/a>:<\/p>\n\n<pre><code>results = imagenet_utils.decode_predictions(preds)\n<\/code><\/pre>\n\n<p>The error is as follows:<\/p>\n\n<pre><code>Using TensorFlow backend.\n* Loading Keras model and Flask starting server...please wait until server has fully started\n2018-11-21 17:47:27.664222: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n * Serving Flask app \"run_keras_server\" (lazy loading)\n * Environment: production\n   WARNING: Do not use the development server in a production environment.\n   Use a production WSGI server instead.\n * Debug mode: off\n * Running on http:\/\/127.0.0.1:5000\/ (Press CTRL+C to quit)\nDownloading data from https:\/\/s3.amazonaws.com\/deep-learning-models\/image-models\/imagenet_class_index.json\n[2018-11-21 19:04:25,054] ERROR in app: Exception on \/predict [POST]\nTraceback (most recent call last):\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 1318, in do_open\n    encode_chunked=req.has_header('Transfer-encoding'))\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 1239, in request\n    self._send_request(method, url, body, headers, encode_chunked)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 1285, in _send_request\n    self.endheaders(body, encode_chunked=encode_chunked)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 1234, in endheaders\n    self._send_output(message_body, encode_chunked=encode_chunked)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 1026, in _send_output\n    self.send(msg)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 964, in send\n    self.connect()\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 1392, in connect\n    super().connect()\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\http\\client.py\", line 936, in connect\n    (self.host,self.port), self.timeout, self.source_address)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\socket.py\", line 724, in create_connection\n    raise err\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\socket.py\", line 713, in create_connection\n    sock.connect(sa)\nTimeoutError: [WinError 10060] A connection attempt failed because the connected party did not properly respond after a period of time, or established connection failed because connected host has failed to respond\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 222, in get_file\n    urlretrieve(origin, fpath, dl_progress)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 248, in urlretrieve\n    with contextlib.closing(urlopen(url, data)) as fp:\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 223, in urlopen\n    return opener.open(url, data, timeout)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 526, in open\n    response = self._open(req, data)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 544, in _open\n    '_open', req)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 504, in _call_chain\n    result = func(*args)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 1361, in https_open\n    context=self._context, check_hostname=self._check_hostname)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\urllib\\request.py\", line 1320, in do_open\n    raise URLError(err)\nurllib.error.URLError: &lt;urlopen error [WinError 10060] A connection attempt failed because the connected party did not properly respond after a period of time, or established connection failed because connected host has failed to respond&gt;\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\app.py\", line 2292, in wsgi_app\n    response = self.full_dispatch_request()\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\app.py\", line 1815, in full_dispatch_request\n    rv = self.handle_user_exception(e)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\app.py\", line 1718, in handle_user_exception\n    reraise(exc_type, exc_value, tb)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\_compat.py\", line 35, in reraise\n    raise value\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\app.py\", line 1813, in full_dispatch_request\n    rv = self.dispatch_request()\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\flask\\app.py\", line 1799, in dispatch_request\n    return self.view_functions[rule.endpoint](**req.view_args)\n  File \"D:\\Mahesh\\workspaces\\ignioworkspace6\\keras_flask\\run_keras_server.py\", line 64, in predict\n    results = imagenet_utils.decode_predictions(preds)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\keras\\applications\\__init__.py\", line 28, in wrapper\n    return base_fun(*args, **kwargs)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\keras\\applications\\imagenet_utils.py\", line 14, in decode_predictions\n    *args, **kwargs)\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\keras_applications\\imagenet_utils.py\", line 228, in decode_predictions\n    file_hash='c2c37ea517e94d9795004a39431a14cb')\n  File \"D:\\user123\\Programs\\python\\python-3.6.6-amd64\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 226, in get_file\n    raise Exception(error_msg.format(origin, e.errno, e.reason))\nException: URL fetch failure on https:\/\/s3.amazonaws.com\/deep-learning-models\/image-models\/imagenet_class_index.json: None -- [WinError 10060] A connection attempt failed because the connected party did not properly respond after a period of time, or established connection failed because connected host has failed to respond\n127.0.0.1 - - [21\/Nov\/2018 19:04:25] \"[1m[35mPOST \/predict HTTP\/1.1[0m\" 500 -\n<\/code><\/pre>\n\n<p>How can I set proxy for keras to let it download the desired file? Is there anyway I can manually download this file and make keras use it from filesystem location. Similar error occurred while downloading <code>resnet50_weights_tf_dim_ordering_tf_kernels.h5<\/code>. So, I manually downloaded it and specified it in the <code>ResNet50()<\/code>.<\/p>\n","answers":[{"AnswerId":"56756184","CreationDate":"2019-06-25T14:18:50.377","ParentId":null,"OwnerUserId":"7394973","Title":null,"Body":"<p>On Windows the Python interpreter will get a proxy from your system environmental variables:<\/p>\n\n<p>set the system variable HTTP_PROXY\nto the value <code>http:\/\/yourproxy<\/code><\/p>\n\n<p>and the system variable HTTPS_PROXY\nto the value <code>http:\/\/yourproxy<\/code><\/p>\n\n<p>Alternately, you can set the environment in a specific Python terminal session according to the answer to this post\n<a href=\"https:\/\/stackoverflow.com\/questions\/11726881\/how-to-set-an-http-proxy-in-python-2-7\">set proxy in python session<\/a><\/p>\n"}]}
+{"QuestionId":53413921,"AnswerCount":2,"Tags":"<tensorflow><dialogflow><chatbot><assistant>","CreationDate":"2018-11-21T14:09:35.040","AcceptedAnswerId":null,"OwnerUserId":10637452.0,"Title":"Can I build my custom AI assistant using dialog flow?","Body":"<p>I am so confused. I want to build a chatbot like Siri but for my own tasks. It should be able to : \n- search on the internet and get answers of questions .\n- give people specific information daily.\n- discuss with people some scientific phenomenas.<\/p>\n\n<p>I can't determine what platform should I use to build this chatbot . I thought about using Dialog flow but I can't figure out if it will give me the ability to do that or no. Also I thought about using tensor flow but I think that it will take a very long time so I was wondering if I can achieve what I want with using some thing like dialog flow and not to build it from scratch?<\/p>\n","answers":[{"AnswerId":"53415430","CreationDate":"2018-11-21T15:31:38.503","ParentId":null,"OwnerUserId":"10423558","Title":null,"Body":"<p>Actually in my opinion, DialogFlow is the best option to build an assistant, it is really easy to build a chatbot to save reminders, to check the weather or to have a simple conversation. DialogFlow has a really powerful tool called webhook that uses Cloud Functions to do the real programming, for example, call google APIs, such as Translate API, or insert data in your Cloud SQL database. <\/p>\n\n<p>Also, DialogFlow use Machine Learning algorithms to understand the customer, for example, if the client says: \"What's the weathe in Barcelona?\" It will answer the question correctly.<\/p>\n\n<p>Another great feature is that it is integrated with multiple technologies, such as, Google Assistant, Amazon Alexa, Cortana, Telegram, Line, Facebook Messenger, etc.<\/p>\n\n<p>I recommend you to follow <a href=\"https:\/\/medium.com\/google-cloud\/how-to-create-a-chatbot-using-dialogflow-enterprise-edition-and-dialogflow-api-v2-923f4a965176\" rel=\"nofollow noreferrer\">this tutorial<\/a>.<\/p>\n"},{"AnswerId":"53416440","CreationDate":"2018-11-21T16:24:58.713","ParentId":null,"OwnerUserId":"10096436","Title":null,"Body":"<p><a href=\"https:\/\/www.luis.ai\/home\" rel=\"nofollow noreferrer\">Luis from Microsoft<\/a> and <a href=\"https:\/\/dialogflow.com\" rel=\"nofollow noreferrer\">DialogFlow from Google<\/a> allow you to build models for natural language processing. These models need to be trained. So the answer is \"no\", out of the box, these tools do not \"search the internet\" to discover answers for your intents.<\/p>\n\n<p>What you need to do is figure out how to train a natural language model, and integrate search data therein. This is bleeding edge AI.  And this really is your question; \"How do I integrate search with NLP and\/or chatbot?\"<\/p>\n\n<p>Both Google and Microsoft let you hook into search. You do not need the dialog tools to do this; you can just pass the query text to do the searching (and let the engine use both ML and heuristic methods to rank results).  You mentioned <a href=\"https:\/\/www.ibm.com\/watson\" rel=\"nofollow noreferrer\">IBM Watson<\/a> and this is a tool that uses ML modelling to try and answer QnA questions.  The Google competition is DeepMind.  You can check out those yourself.<\/p>\n\n<p>But I believe curated content is often the way to go.  Tools like <a href=\"https:\/\/www.qnamaker.ai\/\" rel=\"nofollow noreferrer\">Microsoft's QnA Maker<\/a> let you build these types of applications very easily with little programming required. You can also look into the <a href=\"https:\/\/azure.microsoft.com\/en-us\/services\/search\/\" rel=\"nofollow noreferrer\">Azure or Bing<\/a> search APIs.<\/p>\n\n<p>And if you are looking to start with a bot from template, there are tons of examples on GitHub for <a href=\"https:\/\/azure.microsoft.com\/en-us\/services\/bot-service\/\" rel=\"nofollow noreferrer\">Azure Bot Service<\/a> and Actions-on-Google.  Some even integrate with search and QnA tools.  :-)<\/p>\n\n<p>(And here is the disclaimer. I work for Microsoft. My views do not represent that of my employer.)<\/p>\n"}]}
+{"QuestionId":53414168,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving><tensorflow-estimator>","CreationDate":"2018-11-21T14:22:37.367","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TensorFlow ExportOutputs, PredictOuput, and specifying signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY","Body":"<h1>Context<\/h1>\n\n<p>I have a <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> with a very simple demo <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> for the purpose of learning \/ understanding the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> API with the goal of making a convention for a plug-and-play model with useful bells and whistles of the trade in tack (e.g. early stopping if the validation set stops improving, exporting the model, etc).<\/p>\n\n<p>Each of the three <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> modes (<code>TRAIN<\/code>, <code>EVAL<\/code>, and <code>PREDICT<\/code>) return an <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EstimatorSpec\" rel=\"nofollow noreferrer\"><code>EstimatorSpec<\/code><\/a>.<\/p>\n\n<p>According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EstimatorSpec\" rel=\"nofollow noreferrer\">docs<\/a>:<\/p>\n\n<pre><code>__new__(\n    cls,\n    mode,\n    predictions=None,          # required by PREDICT\n    loss=None,                 # required by TRAIN and EVAL\n    train_op=None,             # required by TRAIN\n    eval_metric_ops=None,\n    export_outputs=None,\n    training_chief_hooks=None,\n    training_hooks=None,\n    scaffold=None,\n    evaluation_hooks=None,\n    prediction_hooks=None.     \n)\n<\/code><\/pre>\n\n<p>Of these named arguments I would like to bring attention to <code>predictions<\/code> and <code>export_outputs<\/code>, which are described in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EstimatorSpec\" rel=\"nofollow noreferrer\">docs<\/a> as:<\/p>\n\n<blockquote>\n  <ul>\n  <li><strong><code>predictions<\/code><\/strong>: Predictions Tensor or dict of Tensor.<\/li>\n  <li><strong><code>export_outputs<\/code><\/strong>: Describes the output signatures to be exported to <code>SavedModel<\/code> and used during serving. A dict <code>{name: output}<\/code> where:\n  \n  <ul>\n  <li><code>name<\/code>: An arbitrary name for this output.<\/li>\n  <li><code>output<\/code>: an <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/ExportOutput\" rel=\"nofollow noreferrer\"><code>ExportOutput<\/code><\/a> object such as <code>ClassificationOutput<\/code>, <code>RegressionOutput<\/code>, or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a>. Single-headed models only need to specify one entry in this dictionary. Multi-headed models should specify one entry for each head, one of which must be named using <code>signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY<\/code>. If no entry is provided, a default <code>PredictOutput<\/code> mapping to predictions will be created.<\/li>\n  <\/ul><\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Thus it should be clear why I bring up <code>export_outputs<\/code>; namely, as one would most likely like to use the model they trained in the future (by loading it from a <code>SavedModel<\/code>). <\/p>\n\n<p>To make this question a bit more accessible \/ add some clarity:<\/p>\n\n<ul>\n<li><p>\"single-headed\" models are the most common model one encounters where the <code>input_fn<\/code> <code>features<\/code> are transformed to a singular (batched) <code>output<\/code><\/p><\/li>\n<li><p>\"multi-headed\" models are models where there is more than one output<\/p><\/li>\n<\/ul>\n\n<p>e.g. this multi-headed model's <code>input_fn<\/code> (in accordance with the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> api) returns a tuple <code>(features, labels)<\/code> i.e. this model has two heads).<\/p>\n\n<pre><code>def input_fn():\n  features = ...\n  labels1 = ...\n  labels2 = ...\n  return features, {'head1': labels1, 'head2': labels2}\n<\/code><\/pre>\n\n<p>How one specifies the <code>signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY<\/code> is the core of this question. Namely, how does one specify it? (e.g. should it be a dict <code>{signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: head}<\/code>)<\/p>\n\n<p>Right, so in the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> you see that our model's <code>export_outputs<\/code> is actually defined in a multi-head manner (although it shouldn't be):<\/p>\n\n<p>From <code>estimator functions &gt; model_fn<\/code> of the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a>:<\/p>\n\n<pre><code>def model_fn(...):\n\n    # ...\n\n    # send the features through the graph\n    MODEL = build_fn(MODEL)\n\n    # prediction\n    MODEL['predictions'] = {'labels': MODEL['net_logits']} # &lt;--- net_logits added in the build_fn\n\n    MODEL['export_outputs'] = {\n        k: tf.estimator.export.PredictOutput(v) for k, v in MODEL['predictions'].items()\n    }\n\n    # ...\n<\/code><\/pre>\n\n<p>in this particular instance, if we expand the dictionary comprehension, we have the functional equivalent of:<\/p>\n\n<pre><code>MODEL['export_outputs'] = {\n    'labels': tf.estimator.export.PredictOutput(MODEL['net_logits'])\n}\n<\/code><\/pre>\n\n<p>It works in this instance as our dictionary has one key and hence one <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a>, where in the <a href=\"https:\/\/colab.research.google.com\/drive\/1GKAqEo7qSr6kMAgxPrOrudA5eLndQ6Ub\" rel=\"nofollow noreferrer\">colab<\/a> our <code>model_fn<\/code> has only a single head and would be more properly formatted as:<\/p>\n\n<pre><code>MODEL['export_outputs'] = {\n    'predictions': tf.estimator.export.PredictOutput(MODEL['predictions'])\n}\n<\/code><\/pre>\n\n<p>as it states in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a>:<\/p>\n\n<pre><code>__init__(outputs)\n<\/code><\/pre>\n\n<p>where <\/p>\n\n<ul>\n<li><code>outputs<\/code>: A Tensor or a dict of string to Tensor representing the predictions.<\/li>\n<\/ul>\n\n<h1>Question<\/h1>\n\n<p>Thus my question is as follows:<\/p>\n\n<ol>\n<li><p>if <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a> can be a dictionary, when \/ why would one want multiple <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a>s as their <code>export_outputs<\/code> for the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EstimatorSpec\" rel=\"nofollow noreferrer\"><code>EstimatorSpec<\/code><\/a>?<\/p><\/li>\n<li><p>If one has a multi-headed model, (say with multiple <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/export\/PredictOutput\" rel=\"nofollow noreferrer\"><code>PredictOutput<\/code><\/a>s) how does one actually specify the <code>signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY<\/code><\/p><\/li>\n<li><p>what is the point of <code>predictions<\/code> in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/EstimatorSpec\" rel=\"nofollow noreferrer\"><code>EstimatorSpec<\/code><\/a> when it is also \"required\" (for anyone who cares about using <code>SavedModel<\/code>s) in <code>export_outputs<\/code>?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"56194720","CreationDate":"2019-05-17T23:27:10.087","ParentId":null,"OwnerUserId":"1284636","Title":null,"Body":"<p>Thanks for your detailed question; you have clearly dug deep here.<\/p>\n\n<ol>\n<li><p>There are also classes for RegressionOutput and ClassificationOutput which cannot be dictionaries. The use of an export_outputs dict allows for generalizations over those use cases.<\/p><\/li>\n<li><p>The head you want to be served by default from the saved model should take the default signature key. For example:<\/p><\/li>\n<\/ol>\n\n<pre><code>export_outputs = {\n  signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:\n    PredictOutput(outputs={'some_output_1': output_1}),\n  'head-2': PredictOutput(outputs={'some_output_2': output_2}),\n  'head-3': PredictOutput(outputs={'some_output_3': output_3})\n}\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>Reason 1: Many people use the default export_outputs (which is in turn the value of predictions), or don't export to saved model. Reason 2: History. Predictions came first, and over time more and more features were added. These features required flexibility and extra info, and were therefore independently packed into the EstimatorSpec.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53414185,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory><cudnn>","CreationDate":"2018-11-21T14:23:47.953","AcceptedAnswerId":null,"OwnerUserId":8049537.0,"Title":"Google Colab Error : Failed to get convolution algorithm.This is probably because cuDNN failed to initialize","Body":"<p>UnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.<\/p>\n\n<pre><code> [[{{node conv2d_1\/convolution}} = Conv2D[T=DT_FLOAT, data_format=\"NCHW\", dilations=[1, 1, 1, 1], padding=\"VALID\", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](zero_padding2d_1\/Pad, conv2d_1\/kernel\/read)]]\n [[{{node metrics\/acc\/Mean\/_255}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_563_metrics\/acc\/Mean\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>This error I am getting on Google colab. Colab's tensorflow version is 1.12.0.<\/p>\n\n<p>I couldn't find any solution. Cause All of them were of local system.<\/p>\n","answers":[{"AnswerId":"54013463","CreationDate":"2019-01-02T21:35:22.363","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>For future passers-by, this looks like a service bug that is being addressed by the Colab folks now.<\/p>\n\n<p>Details are on <a href=\"https:\/\/github.com\/googlecolab\/colabtools\/issues\/384\" rel=\"nofollow noreferrer\">the GitHub issue<\/a>.<\/p>\n\n<p><strong>Update<\/strong>: This issue has been fixed service-side. Details: <a href=\"https:\/\/github.com\/googlecolab\/colabtools\/issues\/384#issuecomment-451001849\" rel=\"nofollow noreferrer\">https:\/\/github.com\/googlecolab\/colabtools\/issues\/384#issuecomment-451001849<\/a><\/p>\n"}]}
+{"QuestionId":53414433,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T14:37:30.497","AcceptedAnswerId":53414584.0,"OwnerUserId":1271079.0,"Title":"Tensorflow: tensor binarization","Body":"<p>I want to transform this dataset in such a way that each tensor has a given size <code>n<\/code> and that a feature at index <code>i<\/code> of this new tensor is set to 1 if and only if there is a <code>i<\/code> in the original feature (modulo n).<\/p>\n\n<p>I hope the following example will make things clearer<\/p>\n\n<p>Let's suppose I have a dataset like: <\/p>\n\n<pre><code>t = tf.constant([\n  [0, 3, 4],\n  [12, 2 ,4]])\nds = tf.data.Dataset.from_tensors(t)\n<\/code><\/pre>\n\n<p>I want to get (if <code>n<\/code> = 9)<\/p>\n\n<pre><code>t = tf.constant([\n  [1, 0, 0, 1, 1, 0, 0, 0, 0], # index set to 1 are 0, 3 and 4\n  [0, 0, 1, 1, 1, 0, 0, 0, 0]]) # index set to 1 are 2, 4, and 12%9 = 3\n<\/code><\/pre>\n\n<p>I know how to apply the modulo to a tensor, but I don't find how to do the rest of the transformation\nthanks<\/p>\n","answers":[{"AnswerId":"53414584","CreationDate":"2018-11-21T14:46:51.153","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>That is similar to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/one_hot\" rel=\"nofollow noreferrer\"><code>tf.one_hot<\/code><\/a>, only for multiple values at the same time. Here is a way to do that:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef binarization(t, n):\n    # One-hot encoding of each value\n    t_1h = tf.one_hot(t % n, n, dtype=tf.bool, on_value=True, off_value=False)\n    # Reduce across last dimension of the original tensor\n    return tf.cast(tf.reduce_any(t_1h, axis=-2), t.dtype)\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    t = tf.constant([\n        [ 0,  3,  4],\n        [12,  2,  4]\n    ])\n    t_m1h = binarization(t, 9)\n    print(sess.run(t_m1h))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[1 0 0 1 1 0 0 0 0]\n [0 0 1 1 1 0 0 0 0]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53415299,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T15:25:40.573","AcceptedAnswerId":53867259.0,"OwnerUserId":6533075.0,"Title":"Retrieving information from a Mask_RCNN Tensor","Body":"<p>I've succesfully trained a <a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\" rel=\"nofollow noreferrer\">Mask_RCNN<\/a>, and for illustration purposes, let's focus on this sample image the network generates:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FN8Ft.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FN8Ft.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It's all very good, no problem. What I'd like to achieve however is to have the following variables with their values per instance:<\/p>\n\n<pre><code>   mask:  (as an image which shows the detected object only, like a binary map)\n   box: (as a list)\n   mask_border_positions (x,y) : (as a list)\n   mask_center_position (x,y) :  (as a tuple)\n<\/code><\/pre>\n\n<p>I've also the function which visualizes the above image, from the official <a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/mrcnn\/visualize.py\" rel=\"nofollow noreferrer\">site<\/a>:<\/p>\n\n<pre><code>def display_instances(image, boxes, masks, class_ids, class_names,\n                      scores=None, title=\"\",\n                      figsize=(16, 16), ax=None,\n                      show_mask=True, show_bbox=True,\n                      colors=None, captions=None):\n    \"\"\"\n    boxes: [num_instance, (y1, x1, y2, x2, class_id)] in image coordinates.\n    masks: [height, width, num_instances]\n    class_ids: [num_instances]\n    class_names: list of class names of the dataset\n    scores: (optional) confidence scores for each box\n    title: (optional) Figure title\n    show_mask, show_bbox: To show masks and bounding boxes or not\n    figsize: (optional) the size of the image\n    colors: (optional) An array or colors to use with each object\n    captions: (optional) A list of strings to use as captions for each object\n    \"\"\"\n    # Number of instances\n    N = boxes.shape[0]\n    if not N:\n        print(\"\\n*** No instances to display *** \\n\")\n    else:\n        assert boxes.shape[0] == masks.shape[-1] == class_ids.shape[0]\n\n    # If no axis is passed, create one and automatically call show()\n    auto_show = False\n    if not ax:\n        _, ax = plt.subplots(1, figsize=figsize)\n        auto_show = True\n\n    # Generate random colors\n    colors = colors or random_colors(N)\n\n    # Show area outside image boundaries.\n    height, width = image.shape[:2]\n    ax.set_ylim(height + 10, -10)\n    ax.set_xlim(-10, width + 10)\n    ax.axis('off')\n    ax.set_title(title)\n\n    masked_image = image.astype(np.uint32).copy()\n    for i in range(N):\n        color = colors[i]\n\n        # Bounding box\n        if not np.any(boxes[i]):\n            # Skip this instance. Has no bbox. Likely lost in image cropping.\n            continue\n        y1, x1, y2, x2 = boxes[i]\n        if show_bbox:\n            p = patches.Rectangle((x1, y1), x2 - x1, y2 - y1, linewidth=2,\n                                alpha=0.7, linestyle=\"dashed\",\n                                edgecolor=color, facecolor='none')\n            ax.add_patch(p)\n\n        # Label\n        if not captions:\n            class_id = class_ids[i]\n            score = scores[i] if scores is not None else None\n            label = class_names[class_id]\n            x = random.randint(x1, (x1 + x2) \/\/ 2)\n            caption = \"{} {:.3f}\".format(label, score) if score else label\n        else:\n            caption = captions[i]\n        ax.text(x1, y1 + 8, caption,\n                color='w', size=11, backgroundcolor=\"none\")\n\n        # Mask\n        mask = masks[:, :, i]\n        if show_mask:\n            masked_image = apply_mask(masked_image, mask, color)\n\n        # Mask Polygon\n        # Pad to ensure proper polygons for masks that touch image edges.\n        padded_mask = np.zeros(\n            (mask.shape[0] + 2, mask.shape[1] + 2), dtype=np.uint8)\n        padded_mask[1:-1, 1:-1] = mask\n        contours = find_contours(padded_mask, 0.5)\n        for verts in contours:\n            # Subtract the padding and flip (y, x) to (x, y)\n            verts = np.fliplr(verts) - 1\n            p = Polygon(verts, facecolor=\"none\", edgecolor=color)\n            ax.add_patch(p)\n    ax.imshow(masked_image.astype(np.uint8))\n    if auto_show:\n        plt.show()\n<\/code><\/pre>\n\n<p>These code snippets below are then called in the main as follows:<\/p>\n\n<pre><code>file_names = glob(os.path.join(IMAGE_DIR, \"*.jpg\"))\nmasks_prediction = np.zeros((510, 510, len(file_names)))\nfor i in range(len(file_names)):\n    print(i)\n    image = skimage.io.imread(file_names[i])\n    predictions = model.detect([image],  verbose=1)\n    p = predictions[0]\n    masks = p['masks']\n    merged_mask = np.zeros((masks.shape[0], masks.shape[1]))\n    for j in range(masks.shape[2]):\n        merged_mask[masks[:,:,j]==True] = True\n        masks_prediction[:,:,i] = merged_mask\nprint(masks_prediction.shape)\n<\/code><\/pre>\n\n<p>and:<\/p>\n\n<pre><code>file_names = glob(os.path.join(IMAGE_DIR, \"*.jpg\"))\nclass_names = ['BG', 'car', 'traffic_light', 'person']\ntest_image = skimage.io.imread(file_names[random.randint(0,len(file_names)-1)])\npredictions = model.detect([test_image], verbose=1) # We are replicating the same image to fill up the batch_size\np = predictions[0]\nvisualize.display_instances(test_image, p['rois'], p['masks'], p['class_ids'], \n                            class_names, p['scores'])\n<\/code><\/pre>\n\n<p>I know it's probably a trivial question and they already exist in the code somewhere, but since I am a starter, I could not get the mask outliers or their centers. If there is a way to have these information per instance, it would be great. <\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"53867259","CreationDate":"2018-12-20T10:53:39.970","ParentId":null,"OwnerUserId":"6533075","Title":null,"Body":"<p>The following does it right:<\/p>\n\n<pre><code>masks = p['masks']\nclass_ids = p['class_ids']\nrois = p['rois']\nscores = p['scores']\nbounding_box = rois[enumerator]\n<\/code><\/pre>\n\n<p>as for the outline coordinates:<\/p>\n\n<pre><code>def getBoundaryPositions(im):\n\n    class_ids = p['class_ids']  # for usage convenience\n\n    im = im.astype(np.uint8)\n\n    # Find contours:\n\n    (im, contours, hierarchy) = cv2.findContours(im, cv2.RETR_EXTERNAL,\n            cv2.CHAIN_APPROX_NONE)\n    cnts = contours[0]\n    outline_posesXY = np.array([np.append(x[0]) for x in cnts])\n\n\n    # Calculate image moments of the detected contour\n    M = cv2.moments(contours[0])\n\n    # collect pose points (for now only position because we don't have pose) of the center\n    positionXY = []\n    positionXY.append(round(M['m10'] \/ M['m00']))\n    positionXY.append(round(M['m01'] \/ M['m00']))\n\n\n    return (im, positionXY, outline_posesXY)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53416135,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tfrecord>","CreationDate":"2018-11-21T16:09:44.180","AcceptedAnswerId":null,"OwnerUserId":8156497.0,"Title":"tfrecord type looks like txt or image","Body":"<p>I'm generating tfrecords of BDD dataset from a 20000 subset. While creating tfrecords I choose 1000 images per tfrecord, and everything is fine. <\/p>\n\n<p>However if I choose 500 images per tfrecord, some of them (very rare) have type TGA image (image\/x-tga) or MATLAB script\/function (text\/x-matlab) while normally they should have had Program (application\/octet-stream) or Binary (application\/octet-stream).<\/p>\n\n<p>Why would this happen, and does it mean the tfrecords are broken?<\/p>\n\n<pre><code>feature = self._get_tf_feature(\n                   picture_id, os.path.join(full_images_path, f),\n                   m.group(2), picture_id_annotations, new_format)\nexample = tf.train.Example(features=feature)\nwriter.write(example.SerializeToString())\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53416170,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-21T16:11:14.253","AcceptedAnswerId":53420571.0,"OwnerUserId":10160917.0,"Title":"How to print out the prediction probabilities in Tensorflow","Body":"<p>I am new to tf. I have trained an encoder - decoder using tensorflow. The program takes as input a word and prints out its phonemes.<\/p>\n\n<p>For example: Hello World  ->  ['h', 'E', 'l', '\"', '@U', ' ', 'w', '\"', '3`', 'r', '5', 'd']<\/p>\n\n<p>I would like to have access to the prediction probability of each phoneme chosen.<\/p>\n\n<p>In the prediction section, the code I am using is the following:<\/p>\n\n<pre><code>def predict(words, sess):\n\n    if len(words) &gt; hp.batch_size:\n        after = predict(words[hp.batch_size:], sess)\n        words = words[:hp.batch_size]\n\n    else:\n        after = []\n    x = np.zeros((len(words), hp.maxlen), np.int32)  # 0: &lt;PAD&gt;\n    for i, w in enumerate(words):\n        for j, g in enumerate((w + \"E\")[:hp.maxlen]):\n            x[i][j] = g2idx.get(g, 2)         \n\n\n    preds = np.zeros((len(x), hp.maxlen), np.int32)\n    for j in range(hp.maxlen):\n\n        xpreds = sess.run(graph.preds, {graph.x: x, graph.y: preds})\n        preds[:, j] = xpreds[:, j]\n<\/code><\/pre>\n\n<p>Thank you in advance!<\/p>\n\n<p>My main problem is where these probabilities are \"hidden\" and how to access them. For example, the letter \"o\" in the word \"Hello\" was mapped with the phoneme \"@U\". I would like to find out with what probability \"@U\" was chosen as the ideal phoneme.<\/p>\n","answers":[{"AnswerId":"53420571","CreationDate":"2018-11-21T21:15:36.723","ParentId":null,"OwnerUserId":"1466792","Title":null,"Body":"<p>Following the discussion, I think I can point you to where the code should be changed.\nIn train.py, line 104:<\/p>\n\n<pre><code>self.preds = tf.to_int32(tf.argmax(logits, -1))\n<\/code><\/pre>\n\n<p>They assign the preds variable to the index with highest probability.\nIn order to get the softmax predictions, you can change the code as follows:<\/p>\n\n<pre><code>self.preds = tf.nn.softmax(logits)\n<\/code><\/pre>\n\n<p>I think that should do it.<\/p>\n\n<p>How to view the probabilities:<\/p>\n\n<pre><code>preds = np.zeros((len(x), hp.maxlen), np.float32)\nfor j in range(hp.maxlen):\n    xpreds = sess.run(graph.preds, {graph.x: x, graph.y: preds})\n    # print shape of output -&gt; batch_size, max_length,number_of_output_options\n    print(xpreds.shape)\n    # print all predictions of the first output\n    print(xpreds[0, 0])\n    # print the probabilty of the network prediction\n    print(xpreds[0, 0, np.argmax(xpreds[0][0])])\n    # preds[:, j] = _preds[:, j]     Need to accumulate the results according to the correct output shape\n<\/code><\/pre>\n"}]}
+{"QuestionId":53416245,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><tensorflow.js>","CreationDate":"2018-11-21T16:14:03.780","AcceptedAnswerId":53450325.0,"OwnerUserId":10300594.0,"Title":"Best model type in TensorFlow.js for Color Prediction?","Body":"<p>I was creating a color predictor when I realized a problem arise. I got the model to work successfully, but the predictions are always within the same median range of about 2.5 to about 5.5. The model is supposed to output a 0 thru 8 corresponding to each color, and I have an even amount of data point for each color for training. Is there a better model I can use so that it will predict something to be 0 or 7? I'm assuming it won't because it thinks they are some kind of outliers.<\/p>\n\n<p>Here's my model<\/p>\n\n<pre><code>const model = tf.sequential();\n\nconst hidden = tf.layers.dense({\n  units: 3,\n  inputShape: [3] \/\/Each input has 3 values r, g, and b\n});\nconst output = tf.layers.dense({\n  units: 1 \/\/only one output (the color that corresponds to the rgb values\n    });\nmodel.add(hidden);\nmodel.add(output);\n\nmodel.compile({\n  activation: 'sigmoid',\n  loss: \"meanSquaredError\",\n  optimizer: tf.train.sgd(0.005)\n});\n<\/code><\/pre>\n\n<p>Is this a good model for my problem?<\/p>\n","answers":[{"AnswerId":"53450325","CreationDate":"2018-11-23T16:43:38.023","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The model is lacking non-linearity, for there is no activation functions.\nGiven a rgb input, the model should predict the most likely color in 8 possible values. It is a classification problem. The model as defined in the question is doing a regression, i.e it is trying to predict a numerical value given the input. <\/p>\n\n<p>For classification problem, the last layer should predict probabilities. <code>softmax<\/code> activation function is mostly used for the last layer in that case. \nThe loss function should be <code>categoricalCrossentropy<\/code> or <code>binaryCrossEntropy<\/code> (if there were only two colors to predict). <\/p>\n\n<p>Consider the following model predicting 3 classes of colors: red, green and blue<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-js lang-js prettyprint-override\"><code>const model = tf.sequential();\r\nmodel.add(tf.layers.dense({units: 10, inputShape: [3], activation: 'sigmoid' }));\r\nmodel.add(tf.layers.dense({units: 10, activation: 'sigmoid' }));\r\nmodel.add(tf.layers.dense({units: 3, activation: 'softmax' }));\r\n\r\nmodel.compile({ loss: 'categoricalCrossentropy', optimizer: 'adam' });\r\n\r\nconst xs = tf.tensor([\r\n  [255, 23, 34],\r\n  [255, 23, 43],\r\n  [12, 255, 56],\r\n  [13, 255, 56],\r\n  [12, 23, 255],\r\n  [12, 56, 255]\r\n]);\r\n\r\n\/\/ Labels\r\nconst label = ['red', 'red', 'green', 'green', 'blue', 'blue']\r\nconst setLabel = Array.from(new Set(label))\r\nconst ys = tf.oneHot(tf.tensor1d(label.map((a) =&gt; setLabel.findIndex(e =&gt; e === a)), 'int32'), 3)\r\n\r\n\/\/ Train the model using the data.\r\n  model.fit(xs, ys, {epochs: 100}).then((loss) =&gt; {\r\n  const t = model.predict(xs);\r\n  pred = t.argMax(1).dataSync(); \/\/ get the class of highest probability\r\n  labelsPred = Array.from(pred).map(e =&gt; setLabel[e])\r\n  console.log(labelsPred)\r\n}).catch((e) =&gt; {\r\n  console.log(e.message);\r\n})<\/code><\/pre>\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n  &lt;head&gt;\r\n    &lt;!-- Load TensorFlow.js --&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.13.3\/dist\/tf.min.js\"&gt; &lt;\/script&gt;\r\n  &lt;\/head&gt;\r\n\r\n  &lt;body&gt;\r\n  &lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":53416833,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><tensor>","CreationDate":"2018-11-21T16:44:52.477","AcceptedAnswerId":53420966.0,"OwnerUserId":2414957.0,"Title":"RuntimeError: Given groups=1, weight of size [64, 3, 7, 7], expected input[3, 1, 224, 224] to have 3 channels, but got 1 channels instead","Body":"<p>In the code below:<\/p>\n\n<pre><code>    model_ft.eval()\n    test_data, test_target = image_datasets['train'][idx]\n    test_data = test_data.cuda()\n    #test_target = test_target.cuda()\n    test_target = torch.tensor(test_target)\n    test_target = test_target.cuda()\n    test_data.unsqueeze_(1)\n    test_target.unsqueeze_(0)\n    print(test_data.shape)\n    output = model_ft(test_data)\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"test_loocv.py\", line 245, in &lt;module&gt;\n    output = model_ft(test_data)\n  File \"\/scratch\/sjn-p3\/anaconda\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/scratch\/sjn-p3\/anaconda\/anaconda3\/lib\/python3.6\/site-packages\/torchvision-0.2.1-py3.6.egg\/torchvision\/models\/resnet.py\", line 139, in forward\n  File \"\/scratch\/sjn-p3\/anaconda\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/scratch\/sjn-p3\/anaconda\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/conv.py\", line 301, in forward\n    self.padding, self.dilation, self.groups)\nRuntimeError: Given groups=1, weight of size [64, 3, 7, 7], expected input[3, 1, 224, 224] to have 3 channels, but got 1 channels instead\n<\/code><\/pre>\n\n<p>Also, test_data has the shape: torch.Size([3, 1, 224, 224]).<\/p>\n\n<p>How should I fix this?<\/p>\n","answers":[{"AnswerId":"53420966","CreationDate":"2018-11-21T21:52:04.800","ParentId":null,"OwnerUserId":"2414957","Title":null,"Body":"<p>Here's the fix:<\/p>\n\n<pre><code>test_data, test_target = image_datasets['train'][idx]\ntest_data = test_data.cuda()\ntest_target = torch.tensor(test_target)\ntest_target = test_target.cuda()\ntest_data.unsqueeze_(0)\ntest_target.unsqueeze_(0)\noutput = model_ft(test_data)\n<\/code><\/pre>\n\n<p>I had to change <code>test_data.unsqueeze_(1)<\/code> to <code>test_data.unsqueeze_(0)<\/code><\/p>\n"}]}
+{"QuestionId":53417537,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-21T17:23:26.090","AcceptedAnswerId":53434778.0,"OwnerUserId":2595787.0,"Title":"Keras initialize large embeddings layer with pretrained embeddings","Body":"<p>I am trying to re-train a word2vec model in Keras 2 with Tensorflow backend by using pretrained embeddings and custom corpus.<\/p>\n\n<p>This is how I initialize the embeddings layer with pretrained embeddings:<\/p>\n\n<pre><code>embedding = Embedding(vocab_size, embedding_dim,\n                      input_length=1, name='embedding',\n                      embeddings_initializer=lambda x: pretrained_embeddings)\n<\/code><\/pre>\n\n<p>where <code>pretrained_embeddings<\/code> is a big matrix of size <code>vocab_size<\/code> x <code>embedding_dim<\/code><\/p>\n\n<p>This works as long as <code>pretrained_embeddings<\/code> is not too big. <\/p>\n\n<p>In my case unfortunately this is not the case - <code>vocab_size=2270872<\/code> and <code>embedding_dim=300<\/code>.<\/p>\n\n<p>Upon initializing the Embeddings layer I get the error:<\/p>\n\n<p><code>Cannot create a tensor proto whose content is larger than 2GB.<\/code><\/p>\n\n<p>The error comes from the function <code>add_weight()<\/code> in \n <code>\/opt\/r\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py<\/code>, more specifically the following line:<\/p>\n\n<pre><code>weight = K.variable(initializer(shape),\n                    dtype=dtype,\n                    name=name,\n                    constraint=constraint)\n<\/code><\/pre>\n\n<p><code>initializer<\/code> is the lambda function from above, which returns the big matrix. <code>shape<\/code> is <code>(2270872, 300)<\/code> as already mentioned.<\/p>\n\n<p>Is it possible to solve this issue without having to go to low-level Tensorflow programming ? If I switch to Theano as a backend the code runs fine, but I'd like to use Tensorflow for its better long-term prospects.<\/p>\n\n<p>The only similar Stackoverflow question I found was <a href=\"https:\/\/stackoverflow.com\/questions\/35394103\/initializing-tensorflow-variable-with-an-array-larger-than-2gb\">this<\/a>, which proposes placeholder variables, but I am not sure how I can apply them on the level of Keras.<\/p>\n\n<p>Thanks a lot<\/p>\n\n<p><strong>Edit:<\/strong>\nI am more than willing to work around this issue on the level of the Tensorflow backend. It's just that I don't know how to combine in this case Tensorflow and Keras code in the same application. Most examples are either one or the other, not both.<\/p>\n\n<p>For example, what use are the Tensorflow placeholder variables when the initialization of the Embeddings layer in Keras will inevitably invoke the add_weight() function, which causes the issue ?<\/p>\n\n<p><strong>Solution:<\/strong><\/p>\n\n<p>As hinted by in @blue-phoenox's comment I rewrote the code like this:<\/p>\n\n<pre><code>embedding = Embedding(vocab_size, embedding_dim,\n                      input_length=1, \n                      name='embedding')\nembedding.build(input_shape=(1,)) # the input_shape here has no effect in the build function\nembedding.set_weights([pretrained_embeddings])\n<\/code><\/pre>\n\n<p>That did it. Thanks again @blue-phoenox.<\/p>\n","answers":[{"AnswerId":"53434778","CreationDate":"2018-11-22T16:11:20.733","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Instead of using the <code>embeddings_initializer<\/code> argument of the Embedding layer you can load pre-trained weights for your embedding layer using the <code>weights<\/code> argument, this way you should be able to hand over pre-trained embeddings larger than 2GB.<\/p>\n\n<p>Here is a short example:<\/p>\n\n<pre><code>from keras.layers import Embedding\n\nembedding_layer = Embedding(vocab_size,\n                            EMBEDDING_DIM,\n                            weights=[embedding_matrix],\n                            input_length=MAX_SEQUENCE_LENGTH,\n                            trainable=False)\n<\/code><\/pre>\n\n<p>Where <code>embedding_matrix<\/code> is just a regular numpy matrix containing your weights.<\/p>\n\n<p>For for examples you can also take a look here:<br>\n<a href=\"https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html<\/a><\/p>\n\n<hr>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>As <em>@PavlinMavrodiev<\/em> <em>(see end of question)<\/em> pointed out correctly the <code>weights<\/code> argument is deprecated. He instead used the <em>layer method<\/em> <a href=\"https:\/\/keras.io\/layers\/about-keras-layers\/\" rel=\"nofollow noreferrer\"><code>set_weights<\/code><\/a> to set the weights instead:<\/p>\n\n<ul>\n<li><blockquote>\n  <p><code>layer.set_weights(weights)<\/code>: sets the weights of the layer from a list\n  of Numpy arrays (with the same shapes as the output of <code>get_weights<\/code>).<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p>To get trained weights <code>get_weights<\/code> can be used:<\/p>\n\n<ul>\n<li><blockquote>\n  <p><code>layer.get_weights()<\/code>: returns the weights of the layer as a list of\n  Numpy arrays.<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p><em>Both are methods from the Keras Layer-Baseclass and can be used for all keras layers, including embeddings layer.<\/em><\/p>\n"}]}
+{"QuestionId":53418168,"AnswerCount":1,"Tags":"<python><tensorflow><pip>","CreationDate":"2018-11-21T18:09:05.643","AcceptedAnswerId":53502795.0,"OwnerUserId":8126904.0,"Title":"Upgrade protobuf leeds to warning with tensorflow","Body":"<p>Here is what I found out so far. Upgrading protobuf leads to this:<\/p>\n\n<pre><code>$ pip3 install --upgrade protobuf\nCollecting protobuf\n  Using cached https:\/\/files.pythonhosted.org\/packages\/18\/e7\/785830a65d1f1faba7dccfa8314f7afded1db8cc2860218226ba4b3f6150\/protobuf-3.6.1-cp37-cp37m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl\nRequirement already satisfied, skipping upgrade: setuptools in \/usr\/local\/lib\/python3.7\/site-packages (from protobuf) (40.5.0)\nRequirement already satisfied, skipping upgrade: six&gt;=1.9 in \/usr\/local\/lib\/python3.7\/site-packages (from protobuf) (1.11.0)\ntensorflow 0.12.0 has requirement protobuf==3.1.0, but you'll have protobuf 3.6.1 which is incompatible.\nInstalling collected packages: protobuf\n  Found existing installation: protobuf 3.1.0\n    Uninstalling protobuf-3.1.0:\n      Successfully uninstalled protobuf-3.1.0\nSuccessfully installed protobuf-3.6.1\n<\/code><\/pre>\n\n<p>In short: <code>tensorflow 0.12.0 has requirement protobuf==3.1.0<\/code>.\nWhy does the newest tensorflow version needs and old protobuf version?<\/p>\n\n<p>But this is even weirder:<\/p>\n\n<pre><code>$ pip3 install --upgrade tensorflow\n...\ntensorboard 1.12.0 has requirement protobuf&gt;=3.4.0\n<\/code><\/pre>\n\n<p>Is my conclusion correct? tensorflow and tensorboard need different versions of protobuf? how? what am I missing out?<\/p>\n\n<p>I found a similar bug but I didn't understand it completely. Something with Mac Operation System and cpython. Link to issue: <a href=\"https:\/\/github.com\/google\/or-tools\/issues\/681\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/or-tools\/issues\/681<\/a><\/p>\n\n<p>I use macOS and I've installed python with Homebrew. <\/p>\n\n<p>Will it help me to install an older python version? <\/p>\n\n<p>I am searching for an explanation and solutions. \nThank you<\/p>\n","answers":[{"AnswerId":"53502795","CreationDate":"2018-11-27T15:20:51.240","ParentId":null,"OwnerUserId":"8126904","Title":null,"Body":"<p>Turns out Tensorflow doesn\u2019t support the newest version of python. Pip cannot find the newest version of Tensorflow for python 3.7.<\/p>\n"}]}
+{"QuestionId":53419127,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2018-11-21T19:18:04.743","AcceptedAnswerId":53437377.0,"OwnerUserId":1968073.0,"Title":"how to get data from within Keras model for visualisation?","Body":"<p>I am using Tensorflow 1.12 which has Keras integrated together with Python 3.6.x<\/p>\n\n<p>I wish to use Keras for its simplicity of model building, but also would like to use data on the intermediate layer for visualization of feature maps and kernels to better understand how machine learning works(even though this is admittedly not so evident)<\/p>\n\n<p>I am using the mnist data base and a very basic Keras model to try to do what I want to do.<\/p>\n\n<p>Here is the code<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras import layers\nfrom tensorflow import keras\n\nprint(tf.VERSION)\nprint(tf.keras.__version__)\n\ntf.keras.backend.clear_session()\n\nmnist = tf.keras.datasets.mnist\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\nx_train_shaped = np.expand_dims(x_train, axis=3) \/ 255.0\nx_test_shaped = np.expand_dims(x_test, axis=3) \/ 255.0\n\ndef create_model():\n\n  model = tf.keras.models.Sequential([\n    keras.layers.Conv2D(32, kernel_size=(4, 4),strides=(1,1),activation='relu', input_shape=(28,28,1)),\n    keras.layers.Dropout(0.5),\n    keras.layers.MaxPooling2D(pool_size=(2,2), strides=(2,2)),\n    keras.layers.Conv2D(24, kernel_size=(8, 8),strides=(1,1)),\n    keras.layers.Flatten(),\n    keras.layers.Dropout(0.5),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n  ])\n\n  model.compile(optimizer=tf.keras.optimizers.Adam(), \n            loss=tf.keras.losses.sparse_categorical_crossentropy,\n            metrics=['accuracy'])\n\n  return model\n<\/code><\/pre>\n\n<p>The above sets up the dataset and the model\nNext I define my session for Tensorflow and do the training.<\/p>\n\n<p>This all works fine but now I want to get my data for the, as example, the first layer out as ideally a numpy array on which I can do the visualization.<\/p>\n\n<p>My <code>model.layers[0].output<\/code> gives me a <code>Tensor<\/code> of <code>(?,25,25,32)<\/code> as expected and now I try to do a <code>eval()<\/code> and thenafter a <code>.numpy()<\/code> method to get my result.<\/p>\n\n<p>The error message is<\/p>\n\n<pre><code>You must feed a value for placeholder tensor 'conv2d_6_input' with dtype float and shape [?,28,28,1]\n<\/code><\/pre>\n\n<p>I am looking for help on how to get my data (32 feature maps of 25x25 pixels) out as numpy array for visualization. <\/p>\n\n<pre><code>sess = tf.Session(graph=tf.get_default_graph())\ntf.keras.backend.set_session(sess)\n\nwith sess.as_default():\n   model = create_model()\n   model.summary()\n\n   model.fit(x_train_shaped[:10000], y_train[:10000], epochs=2, \n   batch_size=64, validation_split=.2,)\n\n   model.layers[0].output\n   print(model.layers[0].output.shape)\n   my_array = model.layers[0].output\n   my_array.eval()\n\ntf.keras.backend.clear_session()\nsess.close()\n<\/code><\/pre>\n","answers":[{"AnswerId":"53437377","CreationDate":"2018-11-22T19:57:28.520","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First of all, you must note that getting the output of a model or a layer only makes sense when you feed the input layers with some data. You get the model something (i.e. input data), you get something in return (i.e. output or feature map or activation map). That's why it would produce the following error:<\/p>\n\n<blockquote>\n  <p>You must <strong>feed<\/strong> a value for placeholder tensor 'conv2d_6_<strong>input<\/strong>'<\/p>\n<\/blockquote>\n\n<p>You haven't fed the baby, so it would cry :)<\/p>\n\n<blockquote>\n  <p>Now, the idea of building a new Keras model is counterproductive. When you have a large model in the first place, one would like to plug in some kind of ready-made code that can get the output of the feature maps and visualize them. So this route seems not really interesting.<\/p>\n<\/blockquote>\n\n<p>I think you are mistakenly thinking that when you construct a new model out of the layers of another model, a whole new model is cloned. That's not the case since the parameters of the layers would be shared.<\/p>\n\n<p>Concretely, what you are looking for can be achieved like this:<\/p>\n\n<pre><code>viz_conv = Model(model.input, model.layers[0].output)\nconv_active = viz_conv(my_input_data)  # my_input_data is a numpy array of shape `(num_samples,28,28,1)`\n<\/code><\/pre>\n\n<p>All the parameters of <code>viz_conv<\/code> are shared with <code>model<\/code> and they have not been copied either. Under the hood they are using the same weight Tensors.<\/p>\n\n<p>Alternatively, you could define a backend function to do this:<\/p>\n\n<pre><code>from tensorflow.keras import backend as K\n\nviz_func = K.function([model.input], [any layer(s) you would like in the model])\noutput = viz_func([my_input_data])\n<\/code><\/pre>\n\n<p>This has been covered in <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">Keras documentation<\/a> and I highly recommend to read that as well.<\/p>\n"}]}
+{"QuestionId":53419474,"AnswerCount":3,"Tags":"<neural-network><deep-learning><pytorch><dropout>","CreationDate":"2018-11-21T19:44:26.967","AcceptedAnswerId":53452827.0,"OwnerUserId":6224975.0,"Title":"Using Dropout in Pytorch: nn.Dropout vs. F.dropout","Body":"<p>By using pyTorch there is two ways to dropout\n <code>torch.nn.Dropout<\/code> and <code>torch.nn.functional.Dropout<\/code>.<\/p>\n\n<p>I struggle to see the difference between the use of them: <br><\/p>\n\n<ul>\n<li>When to use what?<br><\/li>\n<li>Does it make a difference?<br><\/li>\n<\/ul>\n\n<p>I don't see any performance difference when I switched them around.<\/p>\n","answers":[{"AnswerId":"56223207","CreationDate":"2019-05-20T14:38:49.113","ParentId":null,"OwnerUserId":"6080303","Title":null,"Body":"<p>Check <code>torch.nn.functional<\/code>'s implementation:<\/p>\n\n<pre><code> if p &lt; 0. or p &gt; 1.:\n        raise ValueError(\"dropout probability has to be between 0 and 1, \"\n                         \"but got {}\".format(p))\n    return (_VF.dropout_(input, p, training)\n            if inplace\n            else _VF.dropout(input, p, training))\n<\/code><\/pre>\n\n<p>Check: <code>torch.nn.dropout<\/code>'s implementation:<\/p>\n\n<pre><code>def forward(self, input):\n        return F.dropout(input, self.p, self.training, self.inplace)\n<\/code><\/pre>\n\n<p>So: their internal operation is the same. Interfaces are different. As for <code>_VF<\/code>, I guess that's some C\/C++ code. <\/p>\n"},{"AnswerId":"53420065","CreationDate":"2018-11-21T20:32:41.477","ParentId":null,"OwnerUserId":"7606121","Title":null,"Body":"<p>If you look at the source code of <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/dropout.html\" rel=\"nofollow noreferrer\">nn.Dropout<\/a> and <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/functional.html\" rel=\"nofollow noreferrer\">Functional.Dropout<\/a>, you can see <code>Functional<\/code> is an interface and <code>nn<\/code> module implement functions with respect to this interface.<br>\nLook at the implementations in <code>nn<\/code> class:<br><\/p>\n\n<pre><code>from .. import functional as F\nclass Dropout(_DropoutNd):\n    def forward(self, input):\n        return F.dropout(input, self.p, self.training, self.inplace)\n\nclass Dropout2d(_DropoutNd):\n    def forward(self, input):\n        return F.dropout2d(input, self.p, self.training, self.inplace)\n<\/code><\/pre>\n\n<p>And so on.<\/p>\n\n<p>Implementation of <code>Functional<\/code> class:<br><\/p>\n\n<pre><code>def dropout(input, p=0.5, training=False, inplace=False):\n    return _functions.dropout.Dropout.apply(input, p, training, inplace)\n\ndef dropout2d(input, p=0.5, training=False, inplace=False):\n    return _functions.dropout.FeatureDropout.apply(input, p, training, inplace)\n<\/code><\/pre>\n\n<p>look at the example below to understand:<br><\/p>\n\n<pre><code>class Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.conv1 = nn.Conv2d(1, 10, kernel_size=5)\n        self.conv2 = nn.Conv2d(10, 20, kernel_size=5)\n        self.conv2_drop = nn.Dropout2d()\n        self.fc1 = nn.Linear(320, 50)\n        self.fc2 = nn.Linear(50, 10)\n\n    def forward(self, x):\n        x = F.relu(F.max_pool2d(self.conv1(x), 2))\n        x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))\n        x = x.view(-1, 320)\n        x = F.relu(self.fc1(x))\n        x = F.dropout(x, training=self.training)\n        x = self.fc2(x)\n        return F.log_softmax(x)\n<\/code><\/pre>\n\n<p>There is a <code>F.dropout<\/code> in <code>forward()<\/code> function and a <code>nn.Dropout<\/code> in <code>__init__()<\/code> function. Now this is the explanation:<\/p>\n\n<p>In PyTorch you define your Models as subclasses of torch.nn.Module.<\/p>\n\n<p>In the <strong>init<\/strong> function, you are supposed to initialize the layers you want to use. Unlike keras, Pytorch goes more low level and you have to specify the sizes of your network so that everything matches.<\/p>\n\n<p>In the forward method, you specify the connections of your layers. This means that you will use the layers you already initialized, in order to re-use the same layer for each forward pass of data you make.<\/p>\n\n<p>torch.nn.Functional contains some useful functions like activation functions a convolution operations you can use. However, these are not full layers so if you want to specify a layer of any kind you should use torch.nn.Module.<\/p>\n\n<p>You would use the torch.nn.Functional conv operations to define a custom layer for example with a convolution operation, but not to define a standard convolution layer.<\/p>\n"},{"AnswerId":"53452827","CreationDate":"2018-11-23T20:53:12.180","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The technical differences have already been shown in the other answer. However the main difference is that <code>nn.Dropout<\/code> is a torch Module itself which bears some convenience:<\/p>\n\n<p>A short example for illustration of some differences: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\n\nclass Model1(nn.Module):\n    # Model 1 using functional dropout\n    def __init__(self, p=0.0):\n        super().__init__()\n        self.p = p\n\n    def forward(self, inputs):\n        return nn.functional.dropout(inputs, p=self.p, training=True)\n\nclass Model2(nn.Module):\n    # Model 2 using dropout module\n    def __init__(self, p=0.0):\n        super().__init__()\n        self.drop_layer = nn.Dropout(p=p)\n\n    def forward(self, inputs):\n        return self.drop_layer(inputs)\nmodel1 = Model1(p=0.5) # functional dropout \nmodel2 = Model2(p=0.5) # dropout module\n\n# creating inputs\ninputs = torch.rand(10)\n# forwarding inputs in train mode\nprint('Normal (train) model:')\nprint('Model 1', model1(inputs))\nprint('Model 2', model2(inputs))\nprint()\n\n# switching to eval mode\nmodel1.eval()\nmodel2.eval()\n\n# forwarding inputs in evaluation mode\nprint('Evaluation mode:')\nprint('Model 1', model1(inputs))\nprint('Model 2', model2(inputs))\n# show model summary\nprint('Print summary:')\nprint(model1)\nprint(model2)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Normal (train) model:\nModel 1 tensor([ 1.5040,  0.0000,  0.0000,  0.8563,  0.0000,  0.0000,  1.5951,\n         0.0000,  0.0000,  0.0946])\nModel 2 tensor([ 0.0000,  0.3713,  1.9303,  0.0000,  0.0000,  0.3574,  0.0000,\n         1.1273,  1.5818,  0.0946])\n\nEvaluation mode:\nModel 1 tensor([ 0.0000,  0.3713,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,\n         0.0000,  0.0000,  0.0000])\nModel 2 tensor([ 0.7520,  0.1857,  0.9651,  0.4281,  0.7883,  0.1787,  0.7975,\n         0.5636,  0.7909,  0.0473])\nPrint summary:\nModel1()\nModel2(\n  (drop_layer): Dropout(p=0.5)\n)\n<\/code><\/pre>\n\n<p><strong>So which should I use?<\/strong><\/p>\n\n<p>Both are completely equivalent in terms of applying dropout and even though the differences in usage are not that big, there are some reasons to favour the <code>nn.Dropout<\/code> over <code>nn.functional.dropout<\/code>:<\/p>\n\n<p>Dropout is designed to be only applied during training, so when doing predictions or evaluation of the model you want dropout to be turned off.<\/p>\n\n<p>The dropout module <code>nn.Dropout<\/code> conveniently handles this and shuts dropout off as soon as your model enters evaluation mode, while the functional dropout does not care about the evaluation \/ prediction mode. <\/p>\n\n<p>Even though you <em>can<\/em> set functional dropout to <code>training=False<\/code> to turn it off, it is still not such a convenient solution like with <code>nn.Dropout<\/code>.<\/p>\n\n<p>Also the drop rate is stored in the module, so you don't have to save it in an extra variable. In larger networks you might want to create different dropout layers with different drop rates - here <code>nn.Dropout<\/code> may increase readability and can bear also some convenience when using the layers multiple times.<\/p>\n\n<p>Finally, all modules which are assigned to your model are registered in your model. So you model class keeps track of them, that is why you can just turn off the dropout module by calling <code>eval()<\/code>. When using the functional dropout your model is not aware of it, thus it won't appear in any summary.<\/p>\n"}]}
+{"QuestionId":53419665,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-11-21T20:00:10.087","AcceptedAnswerId":null,"OwnerUserId":5176060.0,"Title":"random_image_scale (Data augmentation in the object detection API) does it really work in training?","Body":"<p>I am trying to find a good combination of data augmentation options for my object detection problem. After checking(<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/b2522f9bb9a559e148502a67ebd9a6592320cb0f\/research\/object_detection\/core\/preprocessor.py#L767\" rel=\"nofollow noreferrer\">object_detection\/core\/preprocessor.py<\/a>), it seems that the option random_image_scale donesn't function in the training process. The reason is that tf.image.resize_images is used to scale the image, not tf.image.resize_image_with_crop_or_pad. <\/p>\n\n<p>In object_detection\/legacy\/trainer.py, however, data augmentation is applied first to generate (<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/b2522f9bb9a559e148502a67ebd9a6592320cb0f\/research\/object_detection\/legacy\/trainer.py#L74\" rel=\"nofollow noreferrer\">input_queue<\/a>) first. And then input_queue got fed into the pre-defined mata archs for scaling again. So, if tf.image.resize_images is used, although the size of input_queue has been scaled, it will be scaled back to the wanted size when fed into the network, in this way the random_image_scale option doesn't make any sense and instead, waste computation time. <\/p>\n\n<p>Please let me know if I am wrong.<\/p>\n","answers":[]}
+{"QuestionId":53420194,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving><tensorflow-hub>","CreationDate":"2018-11-21T20:43:19.497","AcceptedAnswerId":53497240.0,"OwnerUserId":2621303.0,"Title":"Tensorflow Hub module reuse","Body":"<p>Say I want to use a specific <code>module<\/code> (text embeddings) from TF Hub to create two distinct <code>models<\/code>, that I would then like to export and serve. <\/p>\n\n<p><strong>Option 1:<\/strong>\nImport the <code>module<\/code> for each <code>model<\/code>, put each classifier on top, and export 2 <code>models<\/code>; serve each in its own docker container. These <code>models<\/code> contain both the underlying embedding module and the classifier.<\/p>\n\n<p><strong>Option 2:<\/strong>\nServe the <code>module<\/code> itself, and have its output go to 2 different served <code>models<\/code>, that themselves do not contain the embeddings. (Is this even possible?)<\/p>\n\n<p>My computer science background tells me that <strong>option 2<\/strong> is better, since we are re-using the original embeddings module for both models, also decoupling the models themselves from the embeddings module.<\/p>\n\n<p>However, from a practical standpoint, when a data scientist is coding, they are importing the <code>module<\/code> and training with the classifier on top of it, so it becomes cumbersome having to export the <code>model<\/code> itself without the underlying embeddings.<\/p>\n\n<p>Can anyone point me in the right direction? Hopefully my question makes sense, I am not a data scientist myself, I am coming more from a development background.<\/p>\n\n<p>Thanks  <\/p>\n","answers":[{"AnswerId":"53497240","CreationDate":"2018-11-27T10:08:35.663","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>Putting a classifier on top of an embedding module creates a fairly strong dependency: the classifier must be trained to the particular embedding space. Unless you make very special arrangements, just swapping in another embedding module won't work. So Option 1 is quite good: it yields two models that can be served and updated independently. They have some overlap, akin to two statically linked programs using the same library, but the source code is still modular: using Hub embedding modules through their common signature makes them interchangeable.<\/p>\n\n<p>Option 2, by comparison, gives you three moving parts with non-trivial dependencies. If your goal is simplicity, I wouldn't go there.<\/p>\n"}]}
+{"QuestionId":53420338,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><feature-extraction>","CreationDate":"2018-11-21T20:54:46.700","AcceptedAnswerId":null,"OwnerUserId":4305584.0,"Title":"Why feature vectors have a lot of zero values in Keras VGG16 model's output?","Body":"<p>I am trying to extract the features from the last layer of VGG16 model in Keras using the following code:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nfrom keras.preprocessing import image\nfrom keras.applications.vgg16 import preprocess_input\nimport numpy as np\n\nmodel = VGG16(weights='imagenet', include_top=True )\n\n\nimg_path = 'E:\\project\\KERAS DEEP\\poodle.png'\nimg = image.load_img(img_path, target_size=(224, 224))\nimg_data = image.img_to_array(img)\nimg_data = np.expand_dims(img_data, axis=0)\nimg_data = preprocess_input(img_data)\nmodel.summary()\nmodel.layers.pop();\nmodel.outputs = [model.layers[-1].output]\nmodel.layers[-1].outbound_nodes = []\nfeature = model.predict(img_data)[0]\n<\/code><\/pre>\n\n<p>feature variable supposes to be the feature vector but it has a lot of zeros, I think because of relu layer. In Matlab, for example, the extracted features vector seems to have both positive and negative values, how I can get the same with keras model?<\/p>\n\n<p>The matlab code is:<\/p>\n\n<pre><code>im=imread('poodle.png');\nim=imresize(im,[224,224]);\nnet=vgg16;\ntrainingFeatures = activations(net, im, 'fc7', ...\n'OutputAs', 'rows');\n<\/code><\/pre>\n\n<p>the two output vectors <code>feature<\/code> and <code>trainingFeatures<\/code> as following (python output to the left and Matlab's to the right<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/quz8S.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/quz8S.png\" alt=\"Samples from the Keras output\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/8EZ9X.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8EZ9X.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And here is the tested image:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qMenv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qMenv.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57814840","CreationDate":"2019-09-06T01:59:04.940","ParentId":null,"OwnerUserId":"31045","Title":null,"Body":"<p>In my own custom network I have found that zeros generally occur in the same location in the vector regardless of the input image, and that the number of these zero dimensions is related to the dropout keep probability. <\/p>\n\n<p>So I would infer dropout causes this. Lower dropout keep probability likely causes less zeros, as more non-zero features are needed to ensure important ones make it through the dropout phase.<\/p>\n\n<p>As for why dropout causes zeros, I\u2019m not sure. It would be interesting to see if the number changes during training. <\/p>\n"}]}
+{"QuestionId":53420994,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorflow-lite><quantization>","CreationDate":"2018-11-21T21:54:22.370","AcceptedAnswerId":null,"OwnerUserId":6161357.0,"Title":"Run quantized tensorflow model on FPGA \/ pure python","Body":"<p>I have a model trained in keras which is a simple model trained on MNIST dataset.<\/p>\n\n<p>What I try to do is to rewrite this model and run on FPGA device.\nIn order to do this I want to fully understand how quantized model works.<\/p>\n\n<p>First I converted this model with post training quantization to .tflite format and UINT8 precision (<a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization<\/a>).<\/p>\n\n<p>So I have quantized model and accuracy is about 90%.<\/p>\n\n<p>Now I try to get weights from quantized model and implement it in a pure python.  I use this tool for visualization and to get model weights: <a href=\"https:\/\/github.com\/lutzroeder\/netron\" rel=\"nofollow noreferrer\">https:\/\/github.com\/lutzroeder\/netron<\/a>.<\/p>\n\n<p>Although simple python code (matrix multiplication, add bias and relu) works, the one with quantized weights doesn't work. <\/p>\n\n<p>So my question is how to write  a feed forward using numpy?<\/p>\n\n<p>My model in keras looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(512, input_shape=input_shape))\nmodel.add(Activation(tf.nn.relu))\nmodel.add(Dense(100))\nmodel.add(Activation(tf.nn.relu))\nmodel.add(Dense(num_classes))\nmodel.add(Activation(tf.nn.softmax))\nmodel.compile(\n    optimizer=Adam(),\n    loss='categorical_crossentropy',\n    metrics=['accuracy'],\n)\n<\/code><\/pre>\n\n<p>I converted it with TocoConverter. And it works in tensorflow.<\/p>\n\n<p>Then I try to write feed forward in pure python:<\/p>\n\n<pre><code>for img, label in zip(x_test, y_test):\n    img = img.astype('uint8')\n    total_seen += 1\n    label = tf.keras.utils.to_categorical(label, num_classes=num_classes)\n    X = img.reshape(1, 784)\n    z1 = np.dot(X, W0.T) + b0\n    a1 = relu(z1)\n    z2 = np.dot(a1, W1.T) + b1\n    a2 = relu(z2)\n    z3 = np.dot(a2, W2.T) + b2\n    prediction = np.argmax(z3)\n    label = np.argmax(label)\n    if prediction == label:\n        num_correct += 1\n<\/code><\/pre>\n\n<p>But this model accuracy is about 10%, so something goes wrong.\nHow to correct this model?<\/p>\n\n<p>Thanks in advance.<\/p>\n\n<p>Edit:\nI've read paper about quantization in tensorflow:\n<a href=\"http:\/\/openaccess.thecvf.com\/content_cvpr_2018\/papers\/Jacob_Quantization_and_Training_CVPR_2018_paper.pdf\" rel=\"nofollow noreferrer\">http:\/\/openaccess.thecvf.com\/content_cvpr_2018\/papers\/Jacob_Quantization_and_Training_CVPR_2018_paper.pdf<\/a><\/p>\n\n<p>And I know almost everything, I know what are S and Z values for activations and kernels. But after matrix multiplication it should be multiplied by factor: M :=S1*S2\/S3.\nAnd i don't know what is S3 scale and how to get it. Because i can't see anything related in netron graph. Any suggestion?<\/p>\n","answers":[{"AnswerId":"53743231","CreationDate":"2018-12-12T12:39:12.720","ParentId":null,"OwnerUserId":"7691129","Title":null,"Body":"<p>There are two steps you'll need to do:<\/p>\n\n<ol>\n<li><p>Dequantize the input, weights and bias back into full precision (or integer equivalent)<\/p>\n\n<p>(w-w_offset)*w_scale<\/p><\/li>\n<li><p>After the Relu, quantize the activations back into integer<\/p>\n\n<p>a\/a_scale+a_offset<\/p>\n\n<p>You can probably skip step 2 that quantize-dequantize the activations with minor risk of getting different result as TFlite model. This is because Relu has no upper bound but TFlite will saturate it to a maximum value.<\/p><\/li>\n<\/ol>\n\n<p>You can check out my tutorials on TFlite in <a href=\"http:\/\/www.github.com\/soon-yau\/QNN\" rel=\"nofollow noreferrer\">my Github<\/a> where I have introduced the concept and training and is about to write out about inference.<\/p>\n"}]}
+{"QuestionId":53421050,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2018-11-21T21:59:16.810","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"How to let predict_generator make ordered prediction in deep learning?","Body":"<p>I am designing a deep learning model to classify images and I am using the following code to check the predication performance. It compares the images' labels with the predicted classes and then returns the error predication. <\/p>\n\n<p>I was using sequential model, and this code was working fine. But, now it does not work fine.. <code>predict_generator<\/code> changes the order of the images. (predict_classes and validation_generator are not in the same order)<\/p>\n\n<p>I want to make the prediction in the same order of the images in <code>validation_generator<\/code>, Do you know how can I do that??   <\/p>\n\n<pre><code>validation_generator = validation_datagen.flow_from_directory(\n            validation_dir,\n            target_size=(image_size, image_size),\n            batch_size=val_batchsize,\n            class_mode='categorical',\n            shuffle=False)\n\n    # Get the filenames from the generator\n    fnames = validation_generator.filenames\n\n    # Get the ground truth from generator\n    ground_truth = validation_generator.classes\n\n    # Get the label to class mapping from the generator\n    label2index = validation_generator.class_indices\n\n    # Getting the mapping from class index to class label\n    idx2label = dict((v,k) for k,v in label2index.items())\n\n    # Get the predictions from the model using the generator\n    predictions = model.predict_generator(validation_generator, steps=validation_generator.samples\/validation_generator.batch_size,verbose=1)\n    predicted_classes = np.argmax(predictions,axis=1)\n\n    errors = np.where(predicted_classes != ground_truth)[0]\n    print(\"No of errors = {}\/{}\".format(len(errors),validation_generator.samples))\n\n    # Show the errors\n    for i in range(len(errors)):\n        pred_class = np.argmax(predictions[errors[i]])\n        pred_label = idx2label[pred_class]\n\n        title = 'Original label:{}, Prediction :{}, confidence : {:.3f}'.format(\n            fnames[errors[i]].split('\/')[0],\n            pred_label,\n            predictions[errors[i]][pred_class])\n\n        original = load_img('{}\/{}'.format(validation_dir,fnames[errors[i]]))\n        plt.figure(figsize=[7,7])\n        plt.axis('off')\n        plt.title(title)\n        plt.imshow(original)\n        plt.show()\n<\/code><\/pre>\n\n<p>The model code <\/p>\n\n<pre><code>from keras.engine import  Model\nvgg16_model = keras.applications.vgg16.VGG16()\n\ninput_layer2 = vgg16_model.input\nvgg16_model.get_layer(index = 0).name = 'input2'\n\nlast_layer2 = vgg16_model.get_layer('block4_pool').output\n\nx = Conv2D(512, (3, 3), activation='relu', name='block5_conv1', padding='same')(last_layer2)\nx = BatchNormalization(name='bn1')(x)\nx = Conv2D(512, (3, 3), activation='relu', name='block5_conv2', padding='same')(x)\nx = BatchNormalization(name='bn2')(x)\nx = Conv2D(512, (3, 3), activation='relu', name='block5_conv3', padding='same')(x)\nx = BatchNormalization(name='bn3')(x)\nx = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name = 'block5_pool')(x)\nx = Flatten(name = 'flatten')(x)\nx = Dense(4096, activation='relu', name='fc1')(x)\nx = Dense(4096, activation='relu', name='fc2')(x)\nout2 = Dropout(0.3, name='dropout1')(x)\n\n#Classification layer\noutput_layer2 = Dense(class_no, activation='softmax', name='prediction')(out2) \nvgg16_face_model = Model(input_layer2, output_layer2)\n\nvgg16_face_model.trainable = True\nset_trainable = False\nfor layer in vgg16_face_model.layers:\n    if layer.name == 'block5_conv1':\n        set_trainable = True\n    if set_trainable:\n        layer.trainable = True\n    else:\n        layer.trainable = False\n<\/code><\/pre>\n\n<p>Please Help me.. <\/p>\n","answers":[]}
+{"QuestionId":53421074,"AnswerCount":1,"Tags":"<c++><tensorflow><deep-learning><c-api>","CreationDate":"2018-11-21T22:01:42.637","AcceptedAnswerId":53421394.0,"OwnerUserId":3858668.0,"Title":"Reshaping Tensor in C","Body":"<p>How can I reshape TF_Tensor* using Tensorflow's C_api as it's being done in C++?<\/p>\n\n<pre><code>TensorShape inputShape({1,1,80,80});\n\nTensor inputTensor;\nTensor newTensor;\n\nbool result = inputTensor-&gt;CopyFrom(newTensor, inputShape);\n<\/code><\/pre>\n\n<p>I don't see a similar method using the tensorflow's c_api.<\/p>\n","answers":[{"AnswerId":"53421394","CreationDate":"2018-11-21T22:36:30.173","ParentId":null,"OwnerUserId":"3656081","Title":null,"Body":"<p>Tensorflow C API operates with a <code>(data,dims)<\/code> model - treating data as a flat raw array supplied with the needed dimensions.<\/p>\n\n<p><strong>Step 1<\/strong>: Allocating a <code>new<\/code> Tensor<\/p>\n\n<p>Have a look at <code>TF_AllocateTensor<\/code>(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/c_api.h#L248\" rel=\"nofollow noreferrer\">ref<\/a>):<\/p>\n\n<pre><code>TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTensor(TF_DataType,\n                                                   const int64_t* dims,\n                                                   int num_dims, size_t len);\n<\/code><\/pre>\n\n<p>Here: <\/p>\n\n<ol>\n<li><code>TF_DataType<\/code>: The <code>TF<\/code> equivalent of the data type you need from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/c_api.h#L99\" rel=\"nofollow noreferrer\">here<\/a>.<\/li>\n<li><code>dims<\/code>: Array corresponding to dimensions of tensor to be allocated eg. <code>{1, 1, 80, 80}<\/code><\/li>\n<li><code>num_dims<\/code>: length of dims(<code>4<\/code> above)<\/li>\n<li><code>len<\/code>: reduce(dims, *): i.e. 1*1*80*80*sizeof(DataType) = 6400*sizeof(DataType).<\/li>\n<\/ol>\n\n<p><strong>Step 2<\/strong>: Copying data<\/p>\n\n<pre><code>\/\/ Get the tensor buffer\nauto buff = (DataType *)TF_TensorData(output_of_tf_allocate);\n\/\/ std::memcpy() ...\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/saatviks\/peloton\/blob\/master\/src\/brain\/util\/tf_session_entity\/tf_session_entity_input.cpp#L45\" rel=\"nofollow noreferrer\">Here<\/a> is some sample code from a project I did a while back on writing a very light Tensorflow C-API Wrapper.<\/p>\n\n<p>So, essentially your reshape will involve allocating your new tensor and copying the data from the original tensor into <code>buff<\/code>.<\/p>\n\n<p>The Tensorflow C API isnt meant for regular usage and thus is harder to learn + lacking documentation. I figured a lot of this out with experimentation. Any suggestions from the more experienced developers out there?<\/p>\n"}]}
+{"QuestionId":53421179,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><autoencoder><tensorflow-probability>","CreationDate":"2018-11-21T22:13:37.607","AcceptedAnswerId":53490299.0,"OwnerUserId":798689.0,"Title":"Forming conditional distributions in TensorFlow probability","Body":"<p>I am using Tensorflow Probability to build a VAE which includes image pixels as well as some other variables. The output of the VAE:<\/p>\n\n<pre><code>tfp.distributions.Independent(tfp.distributions.Bernoulli(logits), 2, name=\"decoder-dist\")\n<\/code><\/pre>\n\n<p>I am trying to understand how to form other conditional distributions based on this which I can use with the inference methods (MCMC or VI). Say the output above was P(A,B,C | Z), how would I take that distribution to form a posterior P(A|B, C, Z) that I could perform inference on? I have been trying to read through the docs but I am having some trouble grasping them.<\/p>\n","answers":[{"AnswerId":"53490299","CreationDate":"2018-11-26T22:49:40.237","ParentId":null,"OwnerUserId":"187421","Title":null,"Body":"<p>The answer to your question depends very much on the nature of the joint model within which you'd like to do the conditioning. Much has been written about the topic, and in short it's a very hard problem in general :). Without knowing a bit more about the particulars of your problem, it's near impossible to recommend a useful generic inference procedure. However, we do have a bunch of examples (scripts and jupyter\/colab notebooks) in the TFP repo here: <a href=\"https:\/\/github.com\/tensorflow\/probability\/tree\/master\/tensorflow_probability\/examples\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/tree\/master\/tensorflow_probability\/examples<\/a><\/p>\n\n<p>In particular, there's<\/p>\n\n<ul>\n<li><p>The <a href=\"https:\/\/github.com\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/examples\/jupyter_notebooks\/HLM_TFP_R_Stan.ipynb\" rel=\"nofollow noreferrer\">Hierarchical Linear Model<\/a> example, which is a sort of Rosetta stone showing how to do posterior inference using Hamiltonian Monte Carlo (an MCMC technique) in TFP, R, and Stan,<\/p><\/li>\n<li><p>The <a href=\"https:\/\/github.com\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/examples\/jupyter_notebooks\/Linear_Mixed_Effects_Model_Variational_Inference.ipynb\" rel=\"nofollow noreferrer\">Linear Mixed Effects Model<\/a> example, showing how you might use VI to solve a standard LME problem,<\/p><\/li>\n<\/ul>\n\n<p>among many others. You can click the \"Run in Google Colab\" link at the top of any of these notebooks to open and run on them on <a href=\"https:\/\/colab.research.google.com\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com<\/a>.<\/p>\n\n<p>Please feel free, also, to reach out on to us via email at tfprobability@tensorflow.org. This is a <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!forum\/tfprobability\" rel=\"nofollow noreferrer\">public Google Group<\/a> where users can engage with the team that builds TFP directly. If you provide us some more info there on what you'd like to do, we're happy to provide guidance on modeling and inference with TFP.<\/p>\n\n<p>Hope this is gives at least a start in the right direction!<\/p>\n"}]}
+{"QuestionId":53421233,"AnswerCount":1,"Tags":"<python><rest><tensorflow><keras><computer-vision>","CreationDate":"2018-11-21T22:19:37.673","AcceptedAnswerId":null,"OwnerUserId":6854425.0,"Title":"ValueError(\"Tensor %s is not an element of this graph.\" % obj)","Body":"<p>First of all, English is not my mother language so excuse me, if I don't express myself really well, please feel free to correct me.<\/p>\n\n<p>I'm making an emotion recognition system that uses rest services to send the image from the client's browser.<\/p>\n\n<p>This is the code:<\/p>\n\n<pre><code># hyper-parameters for bounding boxes shape\nframe_window = 10\nemotion_offsets = (20, 40)\n\n# loading models\nface_detection = load_detection_model(detection_model_path)\nemotion_classifier = load_model(emotion_model_path, compile=False)\nK.clear_session()\n\n# getting input model shapes for inference\nemotion_target_size = emotion_classifier.input_shape[1:3]\n\n# starting lists for calculating modes\nemotion_window = []\n<\/code><\/pre>\n\n<p>And the function:<\/p>\n\n<pre><code>def detect_emotion(self, img):\n\n    # Convert RGB to BGR\n    bgr_image = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)\n    gray_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2GRAY)\n    rgb_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2RGB)\n    faces = detect_faces(face_detection, gray_image)\n\n    for face_coordinates in faces:\n\n        x1, x2, y1, y2 = apply_offsets(face_coordinates, emotion_offsets)\n        gray_face = gray_image[y1:y2, x1:x2]\n        try:\n            gray_face = cv2.resize(gray_face, (emotion_target_size))\n        except:\n            continue\n\n        gray_face = preprocess_input(gray_face, True)\n        gray_face = np.expand_dims(gray_face, 0)\n        gray_face = np.expand_dims(gray_face, -1)\n        emotion_classifier._make_predict_function()\n\n        emotion_prediction = emotion_classifier.predict(gray_face)\n\n        emotion_probability = np.max(emotion_prediction)\n        emotion_label_arg = np.argmax(emotion_prediction)\n        emotion_text = emotion_labels[emotion_label_arg]\n        emotion_window.append(emotion_text)\n\n        if len(emotion_window) &gt; frame_window:\n            emotion_window.pop(0)\n        try:\n            emotion_mode = mode(emotion_window)\n        except:\n            continue\n\n        if emotion_text == 'angry':\n            color = emotion_probability * np.asarray((255, 0, 0))\n        elif emotion_text == 'sad':\n            color = emotion_probability * np.asarray((0, 0, 255))\n        elif emotion_text == 'happy':\n            color = emotion_probability * np.asarray((255, 255, 0))\n        elif emotion_text == 'surprise':\n            color = emotion_probability * np.asarray((0, 255, 255))\n        else:\n            color = emotion_probability * np.asarray((0, 255, 0))\n\n        color = color.astype(int)\n        color = color.tolist()\n\n        draw_bounding_box(face_coordinates, rgb_image, color)\n        draw_text(face_coordinates, rgb_image, emotion_mode,\n                  color, 0, -45, 1, 1)\n\n    img = Image.fromarray(rgb_image)\n\n    return img\n\n\nI'm facing this error when i run my code using waitress:\n\n    File \"c:\\users\\afgir\\documents\\pythonprojects\\face_reco\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3569, in _as_graph_element_locked\n    raise ValueError(\"Tensor %s is not an element of this graph.\" % obj) \n    ValueError: Tensor Tensor(\"predictions_1\/Softmax:0\", shape=(?, 7), dtype=float32) is not an element of this graph.\n<\/code><\/pre>\n\n<p>It loads the image and does all the processing well, I'm pretty sure that the error is in the <code>emotion_classifier.predict<\/code> line, just don't know how to fix it.<\/p>\n\n<p>I've tried with the two solutions in <a href=\"https:\/\/stackoverflow.com\/questions\/47115946\/tensor-is-not-an-element-of-this-graph\">this question<\/a> and none of them worked.<\/p>\n\n<p>I'm really new using <code>Tensorflow<\/code> so I'm kinda stuck with this.<\/p>\n","answers":[{"AnswerId":"53421554","CreationDate":"2018-11-21T22:52:27.967","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>I'm just trying to find out your real environment, but I guess you may use <code>Keras<\/code> amd some <strong>Keras model<\/strong> to predict emotions.  <\/p>\n\n<p>Your error message caused because of the line: <\/p>\n\n<pre><code>K.clear_session()\n<\/code><\/pre>\n\n<p>which, from the documentation: <a href=\"http:\/\/%20tf.keras.backend.clear_session\" rel=\"nofollow noreferrer\">keras.backend.clear_session()<\/a>.\nSo you clear all graph it has been created, then you try to run the classifier's <code>predict()<\/code>, which lost all its context this way.<br>\nThus just simply delete this line.<\/p>\n\n<p>This section is was about some code the Op deleted:<br>\nIn this task you don't need to use <code>tf.Graph()<\/code> at all. You just simply should call <code>emotion_classifier.predict()<\/code> as a simple python method <strong>outside<\/strong> and <strong>without<\/strong> of using any <code>tensorflow graph<\/code>:<\/p>\n\n<pre><code>def detect_emotion(self, img):\n\n    # Convert RGB to BGR\n    bgr_image = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)\n    gray_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2GRAY)\n    rgb_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2RGB)\n    faces = detect_faces(face_detection, gray_image)\n\n    for face_coordinates in faces:\n\n        x1, x2, y1, y2 = apply_offsets(face_coordinates, emotion_offsets)\n        gray_face = gray_image[y1:y2, x1:x2]\n        try:\n            gray_face = cv2.resize(gray_face, (emotion_target_size))\n        except:\n            continue\n\n        gray_face = preprocess_input(gray_face, True)\n        gray_face = np.expand_dims(gray_face, 0)\n        gray_face = np.expand_dims(gray_face, -1)\n        emotion_classifier._make_predict_function()\n\n        emotion_prediction = emotion_classifier.predict(gray_face)\n\n        emotion_probability = np.max(emotion_prediction)\n        emotion_label_arg = np.argmax(emotion_prediction)\n        emotion_text = emotion_labels[emotion_label_arg]\n        emotion_window.append(emotion_text)\n\n        if len(emotion_window) &gt; frame_window:\n            emotion_window.pop(0)\n        try:\n            emotion_mode = mode(emotion_window)\n        except:\n            continue\n\n        if emotion_text == 'angry':\n            color = emotion_probability * np.asarray((255, 0, 0))\n        elif emotion_text == 'sad':\n            color = emotion_probability * np.asarray((0, 0, 255))\n        elif emotion_text == 'happy':\n            color = emotion_probability * np.asarray((255, 255, 0))\n        elif emotion_text == 'surprise':\n            color = emotion_probability * np.asarray((0, 255, 255))\n        else:\n            color = emotion_probability * np.asarray((0, 255, 0))\n\n        color = color.astype(int)\n        color = color.tolist()\n\n        draw_bounding_box(face_coordinates, rgb_image, color)\n        draw_text(face_coordinates, rgb_image, emotion_mode,\n                  color, 0, -45, 1, 1)\n\n    img = Image.fromarray(rgb_image)\n\n    return img\n<\/code><\/pre>\n"}]}
+{"QuestionId":53421305,"AnswerCount":1,"Tags":"<python><tensorflow><keras><image-preprocessing>","CreationDate":"2018-11-21T22:26:29.340","AcceptedAnswerId":53428522.0,"OwnerUserId":6341510.0,"Title":"Custom preprocessing_function with tf.image.rgb_to_grayscale - ValueError: setting an array element with a sequence","Body":"<p>I am trying to use a custom preprocessing function to convert RGB images to grayscale during training. As such, I try to use <code>tf.image.rbg_to_grayscale<\/code> for this. My function looks as following:<\/p>\n\n<pre><code>def prep_data(x):\n    x = tf.image.rgb_to_grayscale(x)\n    return x\n\ndatagen = ImageDataGenerator(preprocessing_function=prep_data,validation_split=0.15)\n<\/code><\/pre>\n\n<p>The <code>train_generator<\/code> is defined using <code>datagen.flow_from_dataframe(...)<\/code>. Training without this custom function works just fine, however once I use it I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: setting an array element with a sequence.<\/p>\n<\/blockquote>\n\n<p>Judging from this answer <a href=\"https:\/\/stackoverflow.com\/questions\/4674473\/valueerror-setting-an-array-element-with-a-sequence\">here<\/a>, I assume I need to change my input to <code>rgb_to_grayscale<\/code>, but I don't know what's the correct way of passing <code>x<\/code> to the function. <\/p>\n\n<p>Any idea on how to solve this?<\/p>\n","answers":[{"AnswerId":"53428522","CreationDate":"2018-11-22T10:10:44.480","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Instead, you can use the <code>color_mode<\/code> argument of <code>flow_from_directory<\/code> and set it to <code>'grayscale'<\/code> to convert the images to grayscale. From <a href=\"https:\/\/keras.io\/preprocessing\/image\/#flow_from_directory\" rel=\"nofollow noreferrer\">Keras docs<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>color_mode<\/strong>: One of \"grayscale\", \"rbg\", \"rgba\". Default: \"rgb\". Whether the images will be converted to have 1, 3, or 4 channels.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53421626,"AnswerCount":6,"Tags":"<python><tensorflow><scikit-learn>","CreationDate":"2018-11-21T22:58:29.480","AcceptedAnswerId":null,"OwnerUserId":10687398.0,"Title":"ValueError: Found array with 0 sample(s) (shape=(0, 1)) while a minimum of 1 is required by MinMaxScaler","Body":"<p>This semester I started working with ML. We have only used APIs such as Microsoft's Azure and Amazon's AWS, but we have not gone in depth about how those services work. My good friend, who is a Math major senior, asked me to help him create a stock predictor with TensorFlow based on a <code>.csv<\/code> the file he provided me.<\/p>\n\n<p>There are a few problems I have. The first one is his <code>.csv<\/code> file. The file has only dates and closing values, which are not separated, therefore I had to manually separate the dates and values. I've managed to do that, and now I'm having trouble with the MinMaxScaler(). I was told I could pretty much disregard the dates and only test the closing values, normalize them, and make a prediction based off of them. \nI keep getting this error - <code>ValueError: Found array with 0 sample(s) (shape=(0, 1)) while a minimum of 1 is required by MinMaxScaler()<\/code>.<\/p>\n\n<p>I honestly have not ever used <code>SKLearning<\/code> and TensorFlow before, and it is my first time working on such a project. All the guides I see on the topic utilize pandas, but in my case, the <code>.csv<\/code> file is a mess and I don't believe I can use pandas for it.<\/p>\n\n<p>I'm following <a href=\"https:\/\/www.datacamp.com\/community\/tutorials\/lstm-python-stock-market\" rel=\"nofollow noreferrer\">this<\/a> guide: <\/p>\n\n<p>But unfortunately, due to my lack of experience, some things are not really working for me, and I would appreciate a little more clarity of how I should proceed in my case.<\/p>\n\n<p>Attached below is my (messy) code:<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nimport tensorflow as tf\nimport sklearn\nfrom sklearn.model_selection import KFold\nfrom sklearn.preprocessing import scale\nfrom sklearn.preprocessing import MinMaxScaler\nimport matplotlib\nimport matplotlib.pyplot as plt\nfrom dateutil.parser import parse\nfrom datetime import datetime, timedelta\nfrom collections import deque\n\nstock_data = []\nstock_date = []\nstock_value = []\nf = open(\"s&amp;p500closing.csv\",\"r\")\ndata = f.read()\nrows = data.split(\"\\n\")\nrows_noheader = rows[1:len(rows)]\n\n#Separating values from messy `.csv`, putting each value to it's list and also a combined list of both\nfor row in rows_noheader:\n    [date, value] = row[1:len(row)-1].split('\\t')\n    stock_date.append(date)\n    stock_value.append((value))\n    stock_data.append((date, value))\n\n#Numpy array of all closing values converted to floats and normalized against the maximum\nstock_value = np.array(stock_value, dtype=np.float32)\nnormvalue = [i\/max(stock_value) for i in stock_value]\n\n#Number of closing values and days. Since there is one closing value for each, they both match and there are 4528 of them (each)\nnclose_and_days = 0\nfor i in range(len(stock_data)):\n    nclose_and_days+=1\n\ntrain_data = stock_value[:2264]\ntest_data = stock_value[2264:]\n\nscaler = MinMaxScaler()\n\ntrain_data = train_data.reshape(-1,1)\ntest_data = test_data.reshape(-1,1)\n\n# Train the Scaler with training data and smooth data\nsmoothing_window_size = 1100\nfor di in range(0,4400,smoothing_window_size):\n    #error occurs here\n    scaler.fit(train_data[di:di+smoothing_window_size,:])\n    train_data[di:di+smoothing_window_size,:] = scaler.transform(train_data[di:di+smoothing_window_size,:])\n\n# You normalize the last bit of remaining data\nscaler.fit(train_data[di+smoothing_window_size:,:])\ntrain_data[di+smoothing_window_size:,:] = scaler.transform(train_data[di+smoothing_window_size:,:])\n\n# Reshape both train and test data\ntrain_data = train_data.reshape(-1)\n\n# Normalize test data\ntest_data = scaler.transform(test_data).reshape(-1)\n\n# Now perform exponential moving average smoothing\n# So the data will have a smoother curve than the original ragged data\nEMA = 0.0\ngamma = 0.1\nfor ti in range(1100):\n    EMA = gamma*train_data[ti] + (1-gamma)*EMA\n    train_data[ti] = EMA\n\n# Used for visualization and test purposes\nall_mid_data = np.concatenate([train_data,test_data],axis=0)\n\nwindow_size = 100\nN = train_data.size\nstd_avg_predictions = []\nstd_avg_x = []\nmse_errors = []\n\nfor pred_idx in range(window_size,N):\n    std_avg_predictions.append(np.mean(train_data[pred_idx-window_size:pred_idx]))\n    mse_errors.append((std_avg_predictions[-1]-train_data[pred_idx])**2)\n    std_avg_x.append(date)\n\nprint('MSE error for standard averaging: %.5f'%(0.5*np.mean(mse_errors)))\n<\/code><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58275108","CreationDate":"2019-10-07T18:20:11.390","ParentId":null,"OwnerUserId":"9663497","Title":null,"Body":"<p>I know that this post is old, but as I stumbled here, others will..\nAfter running in the same problem and googling quite a bit I found a post \n<a href=\"https:\/\/github.com\/llSourcell\/Make_Money_with_Tensorflow_2.0\/issues\/7\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/Make_Money_with_Tensorflow_2.0\/issues\/7<\/a><\/p>\n\n<p>so it seems that if you download a too small dataset it will throw that error.\nDownload a .csv from 1962 and it'll be big enough ;).<\/p>\n\n<p>Now,I just have to find the right parameters for my dataset..as I'm adapting this to another type o prediction..\nHope it helps <\/p>\n"},{"AnswerId":"55751656","CreationDate":"2019-04-18T18:11:57.367","ParentId":null,"OwnerUserId":"11380460","Title":null,"Body":"<p>I see that your Window is 1100 and in your for loop you go from 0 to 4400 with 1100 intervals. With that you have 0 as a remainder, which in turn leaves 0 items to normalize, so your code that has<\/p>\n\n<pre><code># You normalize the last bit of remaining data\nscaler.fit(train_data[di+smoothing_window_size:,:])\ntrain_data[di+smoothing_window_size:,:] = scaler.transform(train_data[di+smoothing_window_size:,:])\n<\/code><\/pre>\n\n<p>You do not need those lines of code, just comment those out. It should work after that<\/p>\n"},{"AnswerId":"53772497","CreationDate":"2018-12-14T01:47:19.430","ParentId":null,"OwnerUserId":"8802814","Title":null,"Body":"<p>Your issue isn't your CSV or pandas. You can actually read the CSV with pandas straight into a dataframe which is what I recommend you do. df = pd.read_csv(path)<\/p>\n\n<p>I am having the same issue with the same code.\nWhats happening is the Scaler = MinMaxScaler, then in the for di in range part and you are fitting the data in the training set then transforming it and reassigning it back to itself. <\/p>\n\n<p>The problem is, is its trying to find more data in your training set to fit to the scaler and its running out of data. Which is weird because of the way the tutorial you are following presented it. <\/p>\n"},{"AnswerId":"58446964","CreationDate":"2019-10-18T08:33:05.897","ParentId":null,"OwnerUserId":"7897976","Title":null,"Body":"<p>at the top of code i wrote %reset to deal with stackoverflow, it cleared out the memory and i got rid of this 'found array with 0 feature', <\/p>\n\n<pre><code>#fitting multiple linear regression to the training set\nfrom sklearn.linear_model import LinearRegression\nregressor =  LinearRegression()\nregressor.fit(X_train,y_train) &lt;------ earlier error occurred here\n<\/code><\/pre>\n"},{"AnswerId":"58424112","CreationDate":"2019-10-17T02:41:45.263","ParentId":null,"OwnerUserId":"10204108","Title":null,"Body":"<p>You have a column of all 0's in your data. If you try to scale it the MinMaxScaler can't assign a scale and it trips up. You need to filter out empty\/0 columns before you scale the data. Try :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    stock_value=stock_value[:,~np.all(np.isnan(d), axis=0)]\n<\/code><\/pre>\n\n<p>to filter out nan columns in your data<\/p>\n"},{"AnswerId":"58292699","CreationDate":"2019-10-08T19:14:41.927","ParentId":null,"OwnerUserId":"8503605","Title":null,"Body":"<p>The <code>train_data<\/code> variable has a length of 2264:<\/p>\n\n<pre><code>train_data = stock_value[:2264]\n<\/code><\/pre>\n\n<p>Then, when you go to fit the scaler, you go outside of <code>train_data<\/code>'s bounds on the third iteration of the for loop:<\/p>\n\n<pre><code>smoothing_window_size = 1100\nfor di in range(0, 4400, smoothing_window_size):\n<\/code><\/pre>\n\n<p>Notice the size of the data set in the tutorial. The training and testing chunks each have length 11,000, and the <code>smoothing_window_size<\/code> is 2500, so it will never go exceed <code>train_data<\/code>'s boundaries.<\/p>\n"}]}
+{"QuestionId":53421999,"AnswerCount":2,"Tags":"<python><pickle><pytorch><torch><torchtext>","CreationDate":"2018-11-21T23:44:12.203","AcceptedAnswerId":null,"OwnerUserId":497600.0,"Title":"how to save torchtext Dataset?","Body":"<p>I'm working with text and use <a href=\"https:\/\/torchtext.readthedocs.io\/en\/latest\/data.html#torchtext.data.Dataset\" rel=\"nofollow noreferrer\"><code>torchtext.data.Dataset<\/code><\/a>.\nCreating the dataset takes a considerable amount of time.\nFor just running the program this is still acceptable. But I would like to debug the torch code for the neural network. And if python is started in debug mode, the dataset creation takes roughly 20 minutes (!!). That's just to get a working environment where I can debug-step through the neural network code.<\/p>\n\n<p>I would like to save the Dataset, for example with pickle. This sample code is taken from <a href=\"https:\/\/github.com\/fastai\/fastai\/blob\/master\/courses\/dl1\/lesson4-imdb.ipynb\" rel=\"nofollow noreferrer\">here<\/a>, but I removed everything that is not necessary for this example:<\/p>\n\n<pre><code>from torchtext import data\nfrom fastai.nlp import *\n\nPATH = 'data\/aclImdb\/'\n\nTRN_PATH = 'train\/all\/'\nVAL_PATH = 'test\/all\/'\nTRN = f'{PATH}{TRN_PATH}'\nVAL = f'{PATH}{VAL_PATH}'\n\nTEXT = data.Field(lower=True, tokenize=\"spacy\")\n\nbs = 64;\nbptt = 70\n\nFILES = dict(train=TRN_PATH, validation=VAL_PATH, test=VAL_PATH)\nmd = LanguageModelData.from_text_files(PATH, TEXT, **FILES, bs=bs, bptt=bptt, min_freq=10)\n\nwith open(\"md.pkl\", \"wb\") as file:\n    pickle.dump(md, file)\n<\/code><\/pre>\n\n<p>To run the code, you need the aclImdb dataset, it can be downloaded from <a href=\"http:\/\/files.fast.ai\/data\/aclImdb.tgz\" rel=\"nofollow noreferrer\">here<\/a>. Extract it into a <code>data\/<\/code> folder next to this code snippet. The code produces an error in the last line, where pickle is used:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/lhk\/programming\/fastai_sandbox\/lesson4-imdb2.py\", line 27, in &lt;module&gt;\n    pickle.dump(md, file)\nTypeError: 'generator' object is not callable\n<\/code><\/pre>\n\n<p>The samples from fastai often use <a href=\"https:\/\/pypi.org\/project\/dill\/\" rel=\"nofollow noreferrer\">dill<\/a> instead of pickle. But that doesn't work for me either.<\/p>\n","answers":[{"AnswerId":"57296594","CreationDate":"2019-07-31T18:02:58.307","ParentId":null,"OwnerUserId":"11865445","Title":null,"Body":"<p>You can always use the pickle to dump the objects, but keep in mind one thing that dumping a list of dictionary or fields objects are not taken care of by the module, so to the best try to decompose the list first <\/p>\n\n<h3>To Store the DataSet Object to a pickle file for later easy loading<\/h3>\n\n<pre><code>def save_to_pickle(dataSetObject,PATH):\n    with open(PATH,'wb') as output:\n        for i in dataSetObject:\n            pickle.dump(vars(i), output, pickle.HIGHEST_PROTOCOL)\n<\/code><\/pre>\n\n<p>The toughest thing is yet to come, Yeah loading the pickle file.... ;)<\/p>\n\n<p>First, try to look for all field names and field attributes and then go for the kill<\/p>\n\n<h3>To load the pickle file into the DataSetObject<\/h3>\n\n<pre><code>def load_pickle(PATH, FIELDNAMES, FIELD):\n    dataList = []\n    with open(PATH, \"rb\") as input_file:\n        while True:\n            try:\n                # Taking the dictionary instance as the input Instance\n                inputInstance = pickle.load(input_file)\n                # plugging it into the list\n                dataInstance =  [inputInstance[FIELDNAMES[0]],inputInstance[FIELDNAMES[1]]]\n                # Finally creating an example objects list\n                dataList.append(Example().fromlist(dataInstance,fields=FIELD))\n            except EOFError:\n                break\n\n    # At last creating a data Set Object\n    exampleListObject = Dataset(dataList, fields=data_fields)\n    return exampleListObject \n<\/code><\/pre>\n\n<p>This hackish solution has worked in my case, hope you will find it useful in your case too.<\/p>\n\n<p>Btw any suggestion is welcome :).<\/p>\n"},{"AnswerId":"54415604","CreationDate":"2019-01-29T07:09:32.817","ParentId":null,"OwnerUserId":"5162955","Title":null,"Body":"<p>You can use dill instead of pickle. It works for me.\nYou can save a torchtext Field like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>TEXT = data.Field(sequential=True, tokenize=tokenizer, lower=True,fix_length=200,batch_first=True)\nwith open(\"model\/TEXT.Field\",\"wb\")as f:\n     dill.dump(TEXT,f)\n<\/code><\/pre>\n\n<p>And load a Field like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with open(\"model\/TEXT.Field\",\"rb\")as f:\n     TEXT=dill.load(f)\n<\/code><\/pre>\n\n<p>The offical code suppport is under development\uff0cyou can follow <a href=\"https:\/\/github.com\/pytorch\/text\/issues\/451\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/text\/issues\/451<\/a> and <a href=\"https:\/\/github.com\/pytorch\/text\/issues\/73\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/text\/issues\/73<\/a> .<\/p>\n"}]}
+{"QuestionId":53422311,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><classification>","CreationDate":"2018-11-22T00:29:49.003","AcceptedAnswerId":null,"OwnerUserId":10664643.0,"Title":"Dynamic batch size in tensorflow","Body":"<p>I have built a classifier using tesnorflow. I generate proposal regions from images and those proposals are individually classified by my classifier.<\/p>\n\n<p>My problem is that I do not have a constant batch size when evaluating my model. Because every image has a different number of proposals, the number of proposals to be evaluated for every image is not constant.<\/p>\n\n<p>Right now I have set the batch size to 1, but this is inefficient and limits the processing speed of my classifier. <\/p>\n\n<p>Below is the placeholder for the input to the model<\/p>\n\n<pre><code>self.image_op = tf.placeholder(tf.float32, shape=[batch_size, 48, 48, 3], name='input_image')\n<\/code><\/pre>\n\n<p>And this is how I feed the input to the model<\/p>\n\n<pre><code>def predict(self,image):\n    cls_prob = self.sess.run([self.cls_prob], feed_dict={self.image_op: image})\n    return cls_prob\n<\/code><\/pre>\n\n<p>Is there any way of setting the batch size to a dynamic value without having to restore the model for every image?<\/p>\n","answers":[]}
+{"QuestionId":53422359,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2018-11-22T00:35:27.917","AcceptedAnswerId":null,"OwnerUserId":10067577.0,"Title":"Installing Older Tensorflow on Windows (.whl file)","Body":"<p>I've been trying to install tensorflow 1.0.1 in a variety of ways with no luck. Currently using Python 3.7 on Windows 10 (64 bit platform). I have plenty of CPU and memory power, but essentially little gpu power (using a server). My best bet was to use install an older tf version through a .whl file (in virtualenv), but keep getting an error. I download this while in the directory of the file: <\/p>\n\n<pre><code>pip install tensorflow-1.3.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>but keep getting-<\/p>\n\n<pre><code>tensorflow-1.3.0-cp35-cp35m-win_amd64.whl is not a supported wheel on this platform.\n<\/code><\/pre>\n\n<p>I prefer to continue using tf through Windows and CPython.... unless there's no other way<\/p>\n","answers":[]}
+{"QuestionId":53422978,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-22T02:13:48.210","AcceptedAnswerId":53423661.0,"OwnerUserId":5689175.0,"Title":"Results not matching when running session multiple times","Body":"<p>When i tried to print  out1 and out2 i observed that the values coming in out2 doesnt exist in out1. But out2 is just finding the maximum from out1. Need help<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\nbox_class_probs = tf.random_normal([2, 2, 1, 2], mean=1, stddev=4, seed = 1)\nmax_ind_class=K.max(box_class_probs,axis=-1)\n\nwith tf.Session() as sess:\n    out1=sess.run(box_class_probs)\n    print(out1)\n    out2=sess.run(max_ind_class)\n    print(out2)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>[[[[-2.24527287  6.93839502]]\n\n  [[ 1.26131749 -8.77081585]]]\n\n\n [[[ 1.39699364  3.36489725]]\n\n  [[ 3.37129188 -7.49171829]]]]\n---------------------------------------------\n---------------------------------------------\n---------------------------------------------\n[[[ 1.96837616]\n  [ 3.06311464]]\n\n [[ 9.33515644]\n  [ 6.58941841]]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53423661","CreationDate":"2018-11-22T03:57:13.783","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>You need to run both of your results in one session run, because you are generating the box_class_probs randomly, and according to the random seed (default or internal), it is going to change every time you execute a session run. And also, keep in mind its always more consistent to get the current keras backend session using <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">K.get_session()<\/a> and then run your code when you are mixing keras and tensorflow.<\/p>\n\n<pre><code>sess = K.get_session()\nout1, out2 = sess.run([box_class_probs, max_ind_class])\nprint(out1)\nprint(out2)\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>[[[[-2.2452729  6.938395 ]]\n\n  [[ 1.2613175 -8.770817 ]]]\n\n\n [[[ 1.3969936  3.3648973]]\n\n  [[ 3.3712919 -7.4917183]]]]\n[[[6.938395 ]\n  [1.2613175]]\n\n [[3.3648973]\n  [3.3712919]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53423563,"AnswerCount":0,"Tags":"<python><pandas><machine-learning><scikit-learn><keras>","CreationDate":"2018-11-22T03:42:05.257","AcceptedAnswerId":null,"OwnerUserId":6818017.0,"Title":"Why is the accuracy and f1 score of this model the same no matter what?","Body":"<p>After doing some data wrangling, I have some variables that are either categorical ones labeled 0 or 1, or numerical variables that have been normalized to have values ranging from 0.0 to 1.0. Then here is where I define and run the model: <\/p>\n\n<pre><code>def baseline_model():\n    # create model\n    model = Sequential()\n    model.add(Dense(8, input_dim=6, activation='relu'))\n    model.add(Dense(2, activation='softmax'))\n    # Compile model\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n\n# Fit the model\nestimator = KerasClassifier(build_fn=baseline_model, nb_epoch=200, batch_size=10, verbose=0)\n\nkfold = KFold(n_splits=5, shuffle=True, random_state=seed)\nresults = cross_val_score(estimator, X, Y, cv=kfold, scoring=\"f1_macro\")\nprint(\"\\nF1 score: %.2f%% (%.2f%%)\" % (results.mean()*100, results.std()*100))\n<\/code><\/pre>\n\n<p>However, no matter which features I put into X, or how much of them, the F1 Score will always be 49.79%<\/p>\n\n<p>EDIT: I've also tested with different optimizers, but the result is still the same<\/p>\n","answers":[]}
+{"QuestionId":53423610,"AnswerCount":1,"Tags":"<python><windows><anaconda><theano><patch>","CreationDate":"2018-11-22T03:49:29.707","AcceptedAnswerId":55721226.0,"OwnerUserId":7241796.0,"Title":"how to update scan Cython code in Theano?","Body":"<p>when I use following ways to update ,there is some problem, system told me \"missing header for unified diff at line 1 of patch<\/p>\n\n<pre><code>cd theano\/scan_module\/\ncython scan_perform.pyx\npatch scan_perform.c numpy_api_changes.diff\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MspbY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MspbY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>by the way, the reason I want to update is that when I used theano in windows anaconda environment, system told me \"UserWarning: The file scan_perform.c is not available. This donot happen normally.\" I find the above method at:<a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/scan_module\/scan_perform_ext.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Theano\/Theano\/blob\/master\/theano\/scan_module\/scan_perform_ext.py<\/a>\nthen use following ways install patch command<\/p>\n\n<pre><code>conda install patch \n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/EilBS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/EilBS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>how to solve this problem?<\/p>\n","answers":[{"AnswerId":"55721226","CreationDate":"2019-04-17T06:29:44.220","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>I was able to disable this warning by following these steps:<\/p>\n\n<ol>\n<li>Go to the <a href=\"https:\/\/github.com\/Theano\/Theano\/releases\" rel=\"nofollow noreferrer\">releases page<\/a> of Theano, click on the commit hash corresponding to the version of Theano you have installed, and then on \"Browse Files\". In my case, since I have Theano 1.0.3, I used the hash <a href=\"https:\/\/github.com\/Theano\/Theano\/tree\/65fefc3acbdbc498e09ea6c6fa8143e2b14dd9e8\" rel=\"nofollow noreferrer\"><code>65fefc3<\/code><\/a>.<\/li>\n<li><p>Navigate to <code>theano\/scan_module\/c_code<\/code> in the github repo (corresponding to <a href=\"https:\/\/github.com\/Theano\/Theano\/tree\/65fefc3acbdbc498e09ea6c6fa8143e2b14dd9e8\/theano\/scan_module\/c_code\" rel=\"nofollow noreferrer\">this<\/a> URL in my case), and download the contents of the file <code>scan_perform.c<\/code> into a file with the same name.<\/p><\/li>\n<li><p>Go to wherever the Theano source files are located in your machine. In my case since I was using anaconda they were in:<\/p>\n\n<pre><code>\/home\/&lt;username&gt;\/miniconda3\/envs\/&lt;env_name&gt;\/lib\/python2.7\/site-packages\/theano\n<\/code><\/pre><\/li>\n<li><p>Once there, navigate to the <code>scan_module<\/code> directory, create a directory called <code>c_code<\/code>, and <code>cd<\/code> into it.<\/p><\/li>\n<li><p>Paste the <code>scan_perform.c<\/code> file you just downloaded into this directory.<\/p><\/li>\n<\/ol>\n\n<p>When executing Theano again you shouldn't see the warning anymore.<\/p>\n"}]}
+{"QuestionId":53423775,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><loss-function><yolo>","CreationDate":"2018-11-22T04:10:04.143","AcceptedAnswerId":null,"OwnerUserId":10688944.0,"Title":"YOLOv3-tensorflow not converging","Body":"<p>I am trying to implement YOLOv3 in tensorflow, I have taken help from online repositories and was successful in converting the darknet weights to tensorflow and run inference.\nNow, I am trying to train the model using the YOLO loss as implemented <a href=\"https:\/\/github.com\/maiminh1996\/YOLOv3-tensorflow\/blob\/master\/loss_function.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I am using the following code snipppet to do so:<\/p>\n\n<pre><code>    with tf.name_scope('Loss_and_Detect'):\n        yolo_loss = compute_loss(output, y_true, anchors, config.num_classes, print_loss=False)\n        tf.summary.scalar('YOLO_loss', yolo_loss)\n        variables = tf.trainable_variables()\n\n        # Variables to be optimized by train_op if the pre-trained darknet-53 is used as is\n        if config.pre_train:\n            variables = variables[312:] # Get the weights after the 52nd conv-layer (darknet-53)\n\n                  # 5e-4 as used in the paper\n        l2_loss = config.weight_decay * tf.add_n([tf.nn.l2_loss(tf.cast(v, dtype=tf.float32)) for v in variables]) \n        loss = yolo_loss + l2_loss\n        tf.summary.scalar('L2_loss', l2_loss)\n        tf.summary.scalar('Total_loss', loss)\n\n    # Define an optimizer for minimizing the computed loss\n    with tf.name_scope('Optimizer'):\n        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n        with tf.control_dependencies(update_ops):\n            optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\n\n    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n    with tf.control_dependencies(update_ops):\n        train_op = optimizer.minimize(loss=loss, global_step=global_step, var_list=variables)\n<\/code><\/pre>\n\n<p>The problem is my YOLO_loss is stuck at ~7-8 and the L2_loss keeps on increasing.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3XHyy.png\" rel=\"nofollow noreferrer\">Here is a snapshot of the tensorboard with learning rate 1e-6 with exponential decay applied to it.(decay_rate=0.8)<\/a><\/p>\n\n<p>I cannot figure out what an I missing\/doing wrong.\nAny help is appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53423858,"AnswerCount":1,"Tags":"<android><tensorflow><android-things><tensorflow-lite>","CreationDate":"2018-11-22T04:21:44.437","AcceptedAnswerId":null,"OwnerUserId":1374416.0,"Title":"TensorFlow Lite and Android Things - Locating the detected Object and store them in RectF Objects?","Body":"<p>I have an Android Tablet where I installed the TensorFlow-Lite DetectorActivity in the examples that were available. It works well on an Android Tablet. However, when I tried to deploy it on a RaspberryPi 3 Model B that ran Android Things, it didn't run. There seemed to be an issue with configuring the camera properly in terms of enabling a live camera preview and running an analysis.<\/p>\n\n<p>My original goal is to make an object detection app run on Android Things. It is also essential to draw the bounding rectangle on the detected objects.<\/p>\n\n<p>I was looking for an example of an Android App that used TensorFlow-Lite and ran on Android Things. I quickly found <a href=\"https:\/\/www.hackster.io\/cagiva851\/neuralcandy-67d2d5\" rel=\"nofollow noreferrer\">this example from hackster.io that uses Image Classification to dispense candy<\/a>. I ran it on my RaspberryPi Board, and it ran. It gives the results, the name of the object, the confidence level, as well as the ID. I was okay to build upon this sample code. Instead of a live camera feed, I could just make the app take a photo, analyze, and give the result. After which, it takes another photo and the cycle continues. <\/p>\n\n<p>However, it did not specify the location in terms of a RectF Object.<\/p>\n\n<p>What I tried to do was to adapt the <code>recognizeFunction<\/code> in the TFLite Android Example, it's in the TFLiteObjectDetectionAPIModel class. I adapted it to the <code>doIdentification<\/code> function of the Candy Dispenser Android App. My function now looks like this:<\/p>\n\n<pre><code>\/\/ outputLocations: array of shape [Batchsize, NUM_DETECTIONS,4]\n\/\/ contains the location of detected boxes\nprivate float[][][] outputLocations;\n\/\/ outputClasses: array of shape [Batchsize, NUM_DETECTIONS]\n\/\/ contains the classes of detected boxes\nprivate float[][] outputClasses;\n\/\/ outputScores: array of shape [Batchsize, NUM_DETECTIONS]\n\/\/ contains the scores of detected boxes\nprivate float[][] outputScores;\n\/\/ numDetections: array of shape [Batchsize]\n\/\/ contains the number of detected boxes\nprivate float[] numDetections;\n\nprivate static final int NUM_DETECTIONS = 10;\n\nprivate static final float IMAGE_MEAN = 128.0f;\nprivate static final float IMAGE_STD = 128.0f;\n\n\n\nprivate void doIdentification(Bitmap image) {\n    Log.e(TAG, \"doing identification!\");\n    Trace.beginSection(\"recognizeImage\");\n\n    int numBytesPerChannel;\n    if (TF_OD_API_IS_QUANTIZED) {\n        Log.e(TAG, \"model is quantized\");\n        numBytesPerChannel = 1; \/\/ Quantized\n    } else {\n        Log.e(TAG, \"model is NOT quantized\");\n        numBytesPerChannel = 4; \/\/ Floating point\n    }\n\n\n    ByteBuffer imgData = ByteBuffer.allocateDirect(1 * TF_INPUT_IMAGE_HEIGHT * TF_INPUT_IMAGE_HEIGHT\n            * 3 * numBytesPerChannel);\n\n    Trace.beginSection(\"preprocessBitmap\");\n    \/\/ Preprocess the image data from 0-255 int to normalized float based\n    \/\/ on the provided parameters.\n    int[] intValues = new int[TF_INPUT_IMAGE_HEIGHT * TF_INPUT_IMAGE_HEIGHT];\n\n    image.getPixels(intValues, 0, image.getWidth(), 0, 0, image.getWidth(), image.getHeight());\n\n    imgData.rewind();\n    for (int i = 0; i &lt; TF_INPUT_IMAGE_HEIGHT; ++i) {\n        for (int j = 0; j &lt; TF_INPUT_IMAGE_HEIGHT; ++j) {\n            int pixelValue = intValues[i * TF_INPUT_IMAGE_HEIGHT + j];\n\n            if (TF_OD_API_IS_QUANTIZED) {\n                imgData.put((byte) ((pixelValue &gt;&gt; 16) &amp; 0xFF));\n                imgData.put((byte) ((pixelValue &gt;&gt; 8) &amp; 0xFF));\n                imgData.put((byte) (pixelValue &amp; 0xFF));\n            } else {\n                imgData.putFloat((((pixelValue &gt;&gt; 16) &amp; 0xFF) - IMAGE_MEAN) \/ IMAGE_STD);\n                imgData.putFloat((((pixelValue &gt;&gt; 8) &amp; 0xFF) - IMAGE_MEAN) \/ IMAGE_STD);\n                imgData.putFloat(((pixelValue &amp; 0xFF) - IMAGE_MEAN) \/ IMAGE_STD);\n            }\n\n        }\n    }\n    Trace.endSection(); \/\/ preprocessBitmap\n\n\n    \/\/ Allocate space for the inference results\n    byte[][] confidencePerLabel = new byte[1][mLabels.size()];\n\n\n    \/\/for box detections\n    \/\/ Copy the input data into TensorFlow.\n    Trace.beginSection(\"feed\");\n    outputLocations = new float[1][NUM_DETECTIONS][4];\n    outputClasses = new float[1][NUM_DETECTIONS];\n    outputScores = new float[1][NUM_DETECTIONS];\n    numDetections = new float[1];\n\n    Object[] inputArray = {imgData};\n    Map&lt;Integer, Object&gt; outputMap = new HashMap&lt;&gt;();\n    outputMap.put(0, outputLocations);\n    outputMap.put(1, outputClasses);\n    outputMap.put(2, outputScores);\n    outputMap.put(3, numDetections);\n    Trace.endSection();\n\n\n    \/\/ Read image data into buffer formatted for the TensorFlow model\n    TensorFlowHelper.convertBitmapToByteBuffer(image, intValues, imgData);\n\n    \/\/ Run inference on the network with the image bytes in imgData as input,\n    \/\/ storing results on the confidencePerLabel array.\n\n    Trace.beginSection(\"run\");\n    mTensorFlowLite.runForMultipleInputsOutputs(inputArray, outputMap);\n    Trace.endSection();\n\n    \/\/ TODO - we try and fetch our rectF's here\n\n    final ArrayList&lt;Recognition&gt; recognitions = new ArrayList&lt;&gt;(NUM_DETECTIONS);\n    for (int i = 0; i &lt; NUM_DETECTIONS; ++i) {\n        final RectF detection =\n                new RectF(\n                        outputLocations[0][i][1] * TF_OD_API_INPUT_SIZE,\n                        outputLocations[0][i][0] * TF_OD_API_INPUT_SIZE,\n                        outputLocations[0][i][3] * TF_OD_API_INPUT_SIZE,\n                        outputLocations[0][i][2] * TF_OD_API_INPUT_SIZE);\n        \/\/ SSD Mobilenet V1 Model assumes class 0 is background class\n        \/\/ in label file and class labels start from 1 to number_of_classes+1,\n        \/\/ while outputClasses correspond to class index from 0 to number_of_classes\n        int labelOffset = 1;\n\n        Log.e(TAG, \"adding the following to our results: \");\n        Log.e(TAG, \"recognition id: \" + i);\n        Log.e(TAG, \"recognition label: \" + mLabels.get((int) outputClasses[0][i] + labelOffset));\n        Log.e(TAG, \"recognition confidence: \" + outputScores[0][i]);\n\n        recognitions.add(\n                new Recognition(\n                        \"\" + i,\n                        mLabels.get((int) outputClasses[0][i] + labelOffset),\n                        outputScores[0][i],\n                        detection));\n\n    }\n    Trace.endSection(); \/\/ \"recognizeImage\"\n\n\n        \/\/ TODO -- This is the old working code\n        \/\/ Get the results with the highest confidence and map them to their labels\n        Collection&lt;Recognition&gt; results = TensorFlowHelper.getBestResults(confidencePerLabel, mLabels);\n        Log.e(TAG, \"results count is = \" + results.size());\n\n        \/\/ Report the results with the highest confidence\n        onClassificationComplete(results);\n\n}\n<\/code><\/pre>\n\n<p>I set the Quantized constant to true, and ran the code. However, I was greeted by the following error:<\/p>\n\n<pre><code>java.lang.IllegalArgumentException: Cannot convert between a TensorFlowLite tensor with type UINT8 and a Java object of type [[[F (which is compatible with the TensorFlowLite type FLOAT32).\n<\/code><\/pre>\n\n<p>and the line responsible is:<\/p>\n\n<pre><code>mTensorFlowLite.runForMultipleInputsOutputs(inputArray, outputMap);\n<\/code><\/pre>\n\n<p>I tried changing to to just <code>mTensorFlowLite.run<\/code>, but that resulted in a different error. <\/p>\n\n<p>Has anyone implemented Object Detection (drawing a RectF on the detected objects) on Android Things? <\/p>\n","answers":[{"AnswerId":"53507663","CreationDate":"2018-11-27T20:31:21.613","ParentId":null,"OwnerUserId":"246461","Title":null,"Body":"<p>The issue here is likely you are mixing models. There is a difference between <strong>image classification<\/strong> and <strong>object detection<\/strong> models. Classification simply reports the confidence of a certain type of object within the image and detection adds on identifying the object's location. The candy dispenser sample you are starting from uses an image classification model (<code>mobilenet_quant_v1_224.tflite<\/code>) whereas the TFLite sample you mentioned runs an object detection model (<code>mobilenet_ssd.tflite<\/code>).<\/p>\n\n<p>I would recommend starting from the sample that does object detection and work through the camera issues rather than solving the problem the other way around. The candy dispenser sample (as well as the <a href=\"https:\/\/github.com\/androidthings\/sample-tensorflow-imageclassifier\" rel=\"nofollow noreferrer\">official image classifier sample<\/a>) provide a good reference for getting the camera on the RPi3 to capture an image and converting it for use with the model.<\/p>\n"}]}
+{"QuestionId":53424056,"AnswerCount":1,"Tags":"<python><pytorch><backpropagation><tensor>","CreationDate":"2018-11-22T04:48:36.917","AcceptedAnswerId":null,"OwnerUserId":2359772.0,"Title":"Apply transformation for Tensor without including it in Backward","Body":"<p>Let's say I have <code>n<\/code> layered neural network. After running <code>l<\/code> layers, I want to apply some transformation to the <code>l<\/code>^th layer output, without including that transformation in backpropagation. <\/p>\n\n<p>For e.g. : <\/p>\n\n<pre><code>output_layer_n = self.LinearLayer(output_layer_prev)\n#apply some transformation to output_layer_n, but don't want to take autograd w.r.t. this transformation, basically this transformation function doesn't have any parameter \noutput_layer_n.data = TransformationFunction(output_layer_n.data) \n<\/code><\/pre>\n\n<p>So how should I go about implementing it? What I want is not to take gradient accounted for <code>TransformationFunction()<\/code> in my code.<\/p>\n","answers":[{"AnswerId":"53427644","CreationDate":"2018-11-22T09:27:31.527","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>If you just don't want to compute gradients for your <code>TransformationFunction<\/code>, it is easiest to turn off gradient computation for all parameters involved in this computation by setting the <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#requires-grad\" rel=\"nofollow noreferrer\"><code>requires_grad<\/code><\/a> flag to <code>False<\/code>.<\/p>\n\n<ul>\n<li><strong><em><a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/autograd.html#excluding-subgraphs-from-backward\" rel=\"nofollow noreferrer\">Excluding subgraphs from backward:<\/a><\/em><\/strong>\n\n<blockquote>\n  <p>If there\u2019s a single input to an operation that requires gradient, its\n  output will also require gradient. Conversely, only if all inputs\n  don\u2019t require gradient, the output also won\u2019t require it. Backward\n  computation is never performed in the subgraphs, where all Tensors\n  didn\u2019t require gradients.<\/p>\n  \n  <p>This is especially useful when you want to freeze part of your model,\n  or you know in advance that you\u2019re not going to use gradients w.r.t.\n  some parameters. For example if you want to finetune a pretrained CNN,\n  it\u2019s enough to switch the <code>requires_grad<\/code> flags in the frozen base, and\n  no intermediate buffers will be saved, until the computation gets to\n  the last layer, where the affine transform will use weights that\n  require gradient, and the output of the network will also require\n  them.<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p>Here is a small example which would do so:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\n\n# define layers\nnormal_layer = nn.Linear(5, 5)\nTransformationFunction = nn.Linear(5, 5)\n# disable gradient computation for parameters of TransformationFunction\n# here weight and bias\nTransformationFunction.weight.requires_grad = False\nTransformationFunction.bias.requires_grad   = False\n\n# input \ninp = torch.rand(1, 5)\n\n# do computation\nout = normal_layer(inp)\nout = TransformationFunction(out)\n\n# loss\nloss = torch.sum(out)\n# backward\nloss.backward()\n\n# gradient for l1\nprint('Gradients for \"normal_layer\"', normal_layer.weight.grad, normal_layer.bias.grad)\n# gradient for l2\nprint('Gradients for \"TransformationFunction\"', TransformationFunction.weight.grad, TransformationFunction.bias.grad)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Gradients for \"normal_layer\" tensor([[0.1607, 0.0215, 0.0192, 0.2595, 0.0811],\n        [0.0788, 0.0105, 0.0094, 0.1272, 0.0398],\n        [0.1552, 0.0207, 0.0186, 0.2507, 0.0784],\n        [0.1541, 0.0206, 0.0184, 0.2489, 0.0778],\n        [0.2945, 0.0393, 0.0352, 0.4756, 0.1486]]) tensor([0.2975, 0.1458, 0.2874, 0.2853, 0.5452])\nGradients for \"TransformationFunction\" None None\n<\/code><\/pre>\n\n<p>I hope this is what you were looking for, if not please edit your question with more detail!<\/p>\n"}]}
+{"QuestionId":53424152,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-22T04:59:09.037","AcceptedAnswerId":53585477.0,"OwnerUserId":1090562.0,"Title":"How to improve the performance of this data pipeline for my tensorflow model","Body":"<p>I have a tensorflow model which I am training on <a href=\"http:\/\/colab.research.google.com\" rel=\"noreferrer\">google-colab<\/a>. The actual model is more complex, but I condensed it into a <a href=\"https:\/\/colab.research.google.com\/drive\/1H-eIdWVhN8FCeOTIirkNVwYMDIk0QNI6\" rel=\"noreferrer\">reproducible example<\/a> (removed saving\/restoring, learning rate decay, asserts, tensorboard events, gradient clipping and so on). The model works reasonably (converges to acceptable loss) and I am looking <strong>for a way to speed up the training<\/strong> (iterations per second). <\/p>\n\n<p>Currently on colab's <strong>GPU it takes 10 minutes to train for 1000 iteration<\/strong>. With my current batch size of 512 it means that the model processes <em>~850 examples per second<\/em> (I would prefer to have a batch size of 512 unless other sizes provide reasonable speedup. By itself changing batch size does not change the speed).<\/p>\n\n<hr>\n\n<p>So currently I have a data stored in tfrecord format: here is a <a href=\"https:\/\/drive.google.com\/open?id=1fUZdXo8Kxhtmwr3pKusxvng5oEOY22NK\" rel=\"noreferrer\">500Mb example file<\/a>, the total data-size is ~0.5Tb. This data passes through a reasonably heavy preprocessing step (I can't do preprocessing beforehand as it will increase the size of my tfrecords way above what I can afford). Preprocessing is done via <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\" rel=\"noreferrer\">tf.data<\/a> and the output tensors (<code>(batch_size, 8, 8, 24)<\/code> which is treated as NHWC, <code>(batch_size, 10)<\/code>) are passed into a model. The example colab does not contain a simplified model which serves just as an example.<\/p>\n\n<hr>\n\n<p>I tried a few approaches to speedup the training:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"noreferrer\">manual device placement<\/a> (data pre-processing on cpu, propagations on gpu), but all my attempts resulted in worse speed (from 10% to 50% increase).<\/li>\n<li>improve data preprocessing. I reviewed <a href=\"https:\/\/www.youtube.com\/watch?v=uIcqeP7MFH0\" rel=\"noreferrer\">tf.data video<\/a> and <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets\" rel=\"noreferrer\">data tutorials<\/a>. I tried almost every technique from that tutorial got no improvement (decrease in speed from 0% to 15%). In particular I tried:\n\n<ul>\n<li><code>dataset.prefetch(...)<\/code><\/li>\n<li>passing <code>num_parallel_calls<\/code> to map<\/li>\n<li>combining map and batch in <code>tf.contrib.data.map_and_batch<\/code><\/li>\n<li>using <code>parallel_interleave<\/code><\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>The code related to data preprocessing is here (here is a <a href=\"https:\/\/colab.research.google.com\/drive\/1H-eIdWVhN8FCeOTIirkNVwYMDIk0QNI6#scrollTo=sOLX0JqKwqZt\" rel=\"noreferrer\">full reproducible example<\/a> with <a href=\"https:\/\/drive.google.com\/open?id=1fUZdXo8Kxhtmwr3pKusxvng5oEOY22NK\" rel=\"noreferrer\">example data<\/a>):<\/p>\n\n<pre><code>_keys_to_map = {\n    'd': tf.FixedLenFeature([], tf.string),  # data\n    's': tf.FixedLenFeature([], tf.int64),   # score\n}\n\n\ndef _parser(record):][3]\n    parsed = tf.parse_single_example(record, _keys_to_map)\n    return parsed['d'], parsed['s']\n\n\ndef init_tfrecord_dataset():\n  files_train = glob.glob(DIR_TFRECORDS + '*.tfrecord')\n  random.shuffle(files_train)\n\n  with tf.name_scope('tfr_iterator'):\n    ds = tf.data.TFRecordDataset(files_train)      # define data from randomly ordered files\n    ds = ds.shuffle(buffer_size=10000)             # select elements randomly from the buffer\n    ds = ds.map(_parser)                           # map them based on tfrecord format\n    ds = ds.batch(BATCH_SIZE, drop_remainder=True) # group elements in batch (remove batch of less than BATCH_SIZE)\n    ds = ds.repeat()                               # iterate infinitely \n\n    return ds.make_initializable_iterator()        # initialize the iterator\n\n\ndef iterator_to_data(iterator):\n  \"\"\"Creates a part of the graph which reads the raw data from an iterator and transforms it to a \n  data ready to be passed to model.\n\n  Args:\n    iterator      - iterator. Created by `init_tfrecord_dataset`\n\n  Returns:\n    data_board      - (BATCH_SIZE, 8, 8, 24) you can think about as NWHC for images.\n    data_flags      - (BATCH_SIZE, 10)\n    combined_score  - (BATCH_SIZE,)\n  \"\"\"\n\n  b = tf.constant((128, 64, 32, 16, 8, 4, 2, 1), dtype=tf.uint8, name='unpacked_const')\n\n  with tf.name_scope('tfr_parse'):\n    with tf.name_scope('packed_data'):\n      next_element = iterator.get_next()\n      data_packed, score_int = next_element\n      score = tf.cast(score_int, tf.float64, name='score_float')\n\n    # https:\/\/stackoverflow.com\/q\/45454470\/1090562\n    with tf.name_scope('data_unpacked'):\n      data_unpacked = tf.reshape(tf.mod(tf.to_int32(tf.decode_raw(data_packed, tf.uint8)[:,:,None] \/\/ b), 2), [BATCH_SIZE, 1552], name='data_unpack')\n\n    with tf.name_scope('score'):\n      with tf.name_scope('is_mate'):\n        score_is_mate = tf.cast(tf.squeeze(tf.slice(data_unpacked, [0, 1546], [BATCH_SIZE, 1])), tf.float64, name='is_mate')\n      with tf.name_scope('combined'):\n        combined_score = (1 - score_is_mate) * VALUE_A * tf.tanh(score \/ VALUE_K) + score_is_mate * tf.sign(score) * (VALUE_A + (1 - VALUE_A) \/ (VALUE_B - 1) * tf.reduce_max(tf.stack([tf.zeros(BATCH_SIZE, dtype=tf.float64), VALUE_B - tf.abs(score)]), axis=0))\n\n\n    with tf.name_scope('board'):\n      with tf.name_scope('reshape_layers'):\n        data_board = tf.reshape(tf.slice(data_unpacked, [0, 0], [BATCH_SIZE, 8 * 8 * 24]), [BATCH_SIZE, 8, 8, 24], name='board_reshape')\n\n      with tf.name_scope('combine_layers'):  \n        data_board = tf.cast(tf.stack([\n          data_board[:,:,:, 0],\n          data_board[:,:,:, 4],\n          data_board[:,:,:, 8],\n          data_board[:,:,:,12],\n          data_board[:,:,:,16],\n          data_board[:,:,:,20],\n          - data_board[:,:,:, 1],\n          - data_board[:,:,:, 5],\n          - data_board[:,:,:, 9],\n          - data_board[:,:,:,13],\n          - data_board[:,:,:,17],\n          - data_board[:,:,:,21],\n          data_board[:,:,:, 2],\n          data_board[:,:,:, 6],\n          data_board[:,:,:,10],\n          data_board[:,:,:,14],\n          data_board[:,:,:,18],\n          data_board[:,:,:,22],\n          - data_board[:,:,:, 3],\n          - data_board[:,:,:, 7],\n          - data_board[:,:,:,11],\n          - data_board[:,:,:,15],\n          - data_board[:,:,:,19],\n          - data_board[:,:,:,23],\n        ], axis=3), tf.float64, name='board_compact')\n\n    with tf.name_scope('flags'):\n      data_flags = tf.cast(tf.slice(data_unpacked, [0, 1536], [BATCH_SIZE, 10]), tf.float64, name='flags')\n\n  return data_board, data_flags, combined_score\n<\/code><\/pre>\n\n<hr>\n\n<p>I am looking for practical solutions (I have tried significant amount of theoretical ideas) which can improve the the speed of training (in terms of examples\/second). I am <strong>not<\/strong> looking for a way to improve the accuracy of the model (or modify the model) as this is just a test model.<\/p>\n\n<p>I have spent significant amount of time trying to optimize this (and gave up). So I would be happy to award a bounty of 200 for a working solution with a nice explanation.<\/p>\n","answers":[{"AnswerId":"53434100","CreationDate":"2018-11-22T15:26:46.740","ParentId":null,"OwnerUserId":"10658208","Title":null,"Body":"<p>I have a couple of suggestions:<\/p>\n\n<p>1) After creating the batch, the entire batch is processed by the <code>iterator_to_data()<\/code> function. This isn't really distributing the task on multiple threads, atleast not at the api level. Instead, you could try something like this in the <code>init_tfrecord_dataset()<\/code> function:<\/p>\n\n<pre><code>ds = tf.data.TFRecordDataset(files_train)      # define data from randomly ordered files\nds = ds.shuffle(buffer_size=10000)             # select elements randomly from the buffer\nds = ds.map(_parser)  \nds = ds.map(map_func=iterator_to_data, num_parallel_calls=FLAGS.num_preprocessing_threads)\nds = ds.batch(BATCH_SIZE, drop_remainder=True) # group elements in batch (remove batch of less than BATCH_SIZE)\nds = ds.repeat()\n<\/code><\/pre>\n\n<p>you might also want to change a few lines in the iterator_to_data() fucntion as the input argument is not a iterator with the above changes.<\/p>\n\n<p>2) You might also want to get the profiling information using something like <code>tf.train.ProfilerHook<\/code>. This can tell you if the bottleneck is with the cpu or gpu. For example, if the bottleneck is with the CPU, you could see GPU ops waiting for memcpyHtoD op to complete.<\/p>\n"},{"AnswerId":"53585477","CreationDate":"2018-12-02T23:03:44.797","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/stackoverflow.com\/a\/53434100\/3574081\">suggestion from hampi<\/a> to profile your training job is a good one, and may be necessary to understand the actual bottlenecks in your pipeline. The other suggestions in the <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets\" rel=\"noreferrer\">Input Pipeline performance guide<\/a> should be useful as well.<\/p>\n\n<p>However, there is another possible \"quick fix\" that might be useful. In some cases, the amount of work in a <code>Dataset.map()<\/code> transformation can be very small, and dominated by the cost of invoking the function for each element. In those cases, we often try to <strong>vectorize<\/strong> the map function, and move it after the <code>Dataset.batch()<\/code> transformation, in order to invoke the function fewer times (1\/512 as many times, in this case), and perform larger\u2014and potentially easier-to-parallelize\u2014operations on each batch. Fortunately, your pipeline can be vectorized as follows:<\/p>\n\n<pre><code>def _batch_parser(record_batch):\n  # NOTE: Use `tf.parse_example()` to operate on batches of records.\n  parsed = tf.parse_example(record_batch, _keys_to_map)\n  return parsed['d'], parsed['s']\n\ndef init_tfrecord_dataset():\n  files_train = glob.glob(DIR_TFRECORDS + '*.tfrecord')\n  random.shuffle(files_train)\n\n  with tf.name_scope('tfr_iterator'):\n    ds = tf.data.TFRecordDataset(files_train)      # define data from randomly ordered files\n    ds = ds.shuffle(buffer_size=10000)             # select elements randomly from the buffer\n    # NOTE: Change begins here.\n    ds = ds.batch(BATCH_SIZE, drop_remainder=True) # group elements in batch (remove batch of less than BATCH_SIZE)\n    ds = ds.map(_batch_parser)                     # map batches based on tfrecord format\n    # NOTE: Change ends here.\n    ds = ds.repeat()                               # iterate infinitely \n\n    return ds.make_initializable_iterator()        # initialize the iterator\n<\/code><\/pre>\n\n<p>Currently, vectorization is a change that you have to make manually, but the <code>tf.data<\/code> team are working on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/bd962d8cdfcda01a23c7051fa05e3db86dd9c30f\/tensorflow\/core\/grappler\/optimizers\/data\/map_vectorization.cc\" rel=\"noreferrer\">an optimization pass that provides automatic vectorization<\/a>.<\/p>\n"}]}
+{"QuestionId":53424304,"AnswerCount":0,"Tags":"<python><tensorflow><resnet>","CreationDate":"2018-11-22T05:18:39.270","AcceptedAnswerId":null,"OwnerUserId":5801328.0,"Title":"Failed to make Dataset.filter() work in the model\/official\/resnet\/resnet_run_loop.py file","Body":"<p>In the official resnet model, I want to filters the dataset from test.bin by the value of 'label' when eval_only set to be True. I tried the tf.data.Dataset.filter() function to get only one class of test data but it didn't work.<\/p>\n\n<pre><code>dataset = dataset.filter(lambda inputs, label: tf.equal(label,15))\n<\/code><\/pre>\n\n<p>I put this code in the resnet_run_loop.process_record_dataset function, but it raised an error<\/p>\n\n<pre><code> raise ValueError(\"`predicate` must return a scalar boolean tensor.\")\n<\/code><\/pre>\n\n<p>I found that the shape of tensor 'label' is (?,) :'Tensor(\"arg1:0\", shape=(?,), dtype=int32, device=\/device:CPU:0)' <\/p>\n","answers":[]}
+{"QuestionId":53424747,"AnswerCount":0,"Tags":"<android><tensorflow><object-detection><object-detection-api><tensorflow-lite>","CreationDate":"2018-11-22T06:02:08.690","AcceptedAnswerId":null,"OwnerUserId":10099298.0,"Title":"Detecting Multiple Boxes in Object Detection using Tensorflow in Android App","Body":"<p>I am trying to train my custom model for detecting \"Raccoon\" in the Android App with <strong>ssd_mobilenet_v1_pets.config<\/strong>. I have used <strong>ssd_mobilenet_v2_coco_2018_03_29<\/strong> pre-trained model. After 20K training steps, I am getting successfully <code>frozen_inference_graph.pb<\/code> file. But when I used this file in my Android App result is not accurate. I am getting multiple boxes with 0.99 accuracies not only on Raccoon but also on others objects. Besides <strong>ssd_mobilenet_v2_coco_2018_03_29<\/strong>, i tried lots of pre-trained models.<\/p>\n\n<p>Please help if anyone faced this problem.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/W12nI.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/W12nI.jpg\" alt=\"I have also attached the output screen\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":53424847,"AnswerCount":2,"Tags":"<keras><ocr>","CreationDate":"2018-11-22T06:11:28.210","AcceptedAnswerId":null,"OwnerUserId":10689350.0,"Title":"Why the \"get_output_size\" is len(alphabet) + 1 not len(alphabet) in the Keras OCR example\uff1f","Body":"<p>I am just a Keras beginner and I try to implement a OCR project by Keras.So I try to learn from Keras OCR example.Here's <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/image_ocr.py\" rel=\"nofollow noreferrer\">a link<\/a>!<br>\nI do not understand why \"get_output_size\" in class TextImageGenerator is  len(alphabet) + 1 but not len(alphabet)?\nI will appreciate it if someone can tell me why ..<\/p>\n","answers":[{"AnswerId":"53427609","CreationDate":"2018-11-22T09:25:31.220","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p>It's related to the CTC layer used as cost function. Maybe reading the scientific papers will give you more perspective, but it's related to a \"extra\" class used by the model to say (\"there is no letter\"). \n<a href=\"https:\/\/papers.nips.cc\/paper\/3449-offline-handwriting-recognition-with-multidimensional-recurrent-neural-networks.pdf\" rel=\"nofollow noreferrer\">Paper by Graves explaining the algorithm behind<\/a><\/p>\n"},{"AnswerId":"53433956","CreationDate":"2018-11-22T15:18:19.657","ParentId":null,"OwnerUserId":"2679594","Title":null,"Body":"<p>There is <strong>one extra-character needed<\/strong> in neural networks trained with <strong>CTC loss<\/strong>. This extra-character essentially means \"no character seen at this position\" and is called <strong>CTC blank<\/strong>.<\/p>\n\n<p>It is used to allow different alignments of a text or to allow some white-space between characters (think of an image containing \"    hello\" or \"hello    \" with whitespace around them, for both you want to recognize \"hello\"). \nWhen recognizing the text, these blanks are removed: e.g. when using best path decoding, the best-scoring character at each position is taken, but the blanks will be removed.<\/p>\n\n<p>To get a better idea of this special CTC blank character, let's look at the illustration below. The output of the neural network contains the characters a, b and the CTC blank (denoted as \"-\"). \nLet's pick the best-scoring characters for each position t0...t4, this gives us \"aaa-b\". Best path decoding removes repeated characters, this gives us \"a-b\", and finally removes all blanks, which gives us \"ab\".\n<a href=\"https:\/\/i.stack.imgur.com\/azwao.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/azwao.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If you want some more information, you can look at my <a href=\"https:\/\/towardsdatascience.com\/3797e43a86c\" rel=\"nofollow noreferrer\">CTC article<\/a>, or <a href=\"https:\/\/distill.pub\/2017\/ctc\/\" rel=\"nofollow noreferrer\">this article<\/a>, or the <a href=\"https:\/\/www.cs.toronto.edu\/~graves\/icml_2006.pdf\" rel=\"nofollow noreferrer\">original paper<\/a>.<\/p>\n"}]}
+{"QuestionId":53426033,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><visualization><tensorboard>","CreationDate":"2018-11-22T07:41:24.637","AcceptedAnswerId":null,"OwnerUserId":4948889.0,"Title":"Projector in Tensorboard python application","Body":"<p>I have the following code and sample which is working fine and exactly I want it to:  <\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport sklearn\nimport sklearn.preprocessing\nimport datetime\nimport os\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nfrom tensorflow.contrib.tensorboard.plugins import projector\n\nvalid_set_size_percentage = 3\ntest_set_size_percentage = 3\nseq_len = 5 # choose sequence length\ndf = pd.read_csv(\"Test.csv\", encoding = 'utf-16',sep=',',index_col = 0)\ndf.head()\ndef normalize_data(df):\n    cols = list(df_stock.columns.values)\n    min_max_scaler = sklearn.preprocessing.MinMaxScaler()\n    df = pd.DataFrame(min_max_scaler.fit_transform(df.values))\n    df.columns = cols\n    return df\ndef load_data(stock, seq_len):\n    data_raw = stock.as_matrix() # convert to numpy array\n    data = []\n    print(data_raw.shape)\n    for index in range(len(data_raw) - seq_len): \n        data.append(data_raw[index: index + seq_len])\n    data = np.array(data);\n    valid_set_size = int(np.round(valid_set_size_percentage\/100*data.shape[0]));  \n    test_set_size = int(np.round(test_set_size_percentage\/100*data.shape[0]));\n    train_set_size = data.shape[0] - (valid_set_size + test_set_size);\n    x_train = data[:train_set_size,:-1,:]\n    y_train = data[:train_set_size,-1,:4]\n    x_valid = data[train_set_size:train_set_size+valid_set_size,:-1,:]\n    y_valid = data[train_set_size:train_set_size+valid_set_size,-1,:4]\n    x_test = data[train_set_size+valid_set_size:,:-1,:]\n    y_test = data[train_set_size+valid_set_size:,-1,:4]\n    return [x_train, y_train, x_valid, y_valid, x_test, y_test]\ndf_stock = df.copy()\ncols = list(df_stock.columns.values)\nprint('df_stock.columns.values = ', cols)\ndf_stock_norm = df_stock.copy()\ndf_stock_norm = normalize_data(df_stock_norm)\nx_train, y_train, x_valid, y_valid, x_test, y_test = load_data(df_stock_norm, seq_len)\nprint(y_train[:2])\nprint('x_train.shape = ',x_train.shape)\nprint('y_train.shape = ', y_train.shape)\nprint('Inputs = ',x_train.shape[2])\nprint('Outputs = ', y_train.shape[1])\nprint('x_valid.shape = ',x_valid.shape)\nprint('y_valid.shape = ', y_valid.shape)\nprint('x_test.shape = ', x_test.shape)\nprint('y_test.shape = ',y_test.shape)\nindex_in_epoch = 0;\nperm_array  = np.arange(x_train.shape[0])\nnp.random.shuffle(perm_array)\ndef get_next_batch(batch_size):\n    global index_in_epoch, x_train, perm_array  \n    if index_in_epoch &gt; x_train.shape[0]:\n        start = 0 # start next epoch\n        index_in_epoch = 0#batch_size\n    start = index_in_epoch\n    index_in_epoch += batch_size       \n    end = index_in_epoch\n    return x_train[perm_array[start:end]], y_train[perm_array[start:end]]\nn_steps = seq_len -1\nn_inputs = x_train.shape[2]\nn_neurons = 100\nn_outputs = y_train.shape[-1]\nn_layers = 2\nlearning_rate = 0.001\nbatch_size =10\nn_epochs = 100\ntrain_set_size = x_train.shape[0]\ntest_set_size = x_test.shape[0]\n\ntf.reset_default_graph()\n\nX = tf.placeholder(tf.float32, [None, n_steps, n_inputs])\ny = tf.placeholder(tf.float32, [None,n_outputs])\n\n\nlayers = [tf.contrib.rnn.LSTMCell(num_units=n_neurons, \n                                 activation=tf.nn.leaky_relu, use_peepholes = True)\n         for layer in range(n_layers)]\nmulti_layer_cell = tf.contrib.rnn.MultiRNNCell(layers)\nrnn_outputs, states = tf.nn.dynamic_rnn(multi_layer_cell, X, dtype=tf.float32)\nstacked_rnn_outputs = tf.reshape(rnn_outputs, [-1, n_neurons]) \nstacked_outputs = tf.layers.dense(stacked_rnn_outputs, n_outputs)\n\noutputs = tf.reshape(stacked_outputs, [-1, n_steps, n_outputs])\n\noutputs = outputs[:,n_steps-1,:] # keep only last output of sequence\n\nloss = tf.reduce_mean(tf.squared_difference(outputs, y)) # loss function = mean squared error \n\noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) \ntraining_op = optimizer.minimize(loss)\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess: \n    sess.run(tf.global_variables_initializer())\n    for iteration in range(int(n_epochs*train_set_size\/batch_size)):\n        x_batch, y_batch = get_next_batch(batch_size) # fetch the next training batch \n\n        sess.run(training_op, feed_dict={X: x_batch, y: y_batch}) \n\n        if iteration % int(1*train_set_size\/batch_size) == 0:\n\n            mse_train = loss.eval(feed_dict={X: x_train, y: y_train}) \n            mse_valid = loss.eval(feed_dict={X: x_valid, y: y_valid}) \n            mse_test = loss.eval(feed_dict={X: x_test, y: y_test})\n\n            print('%.2f epochs: MSE train\/valid\/test = %.3f\/%.3f\/%.3f'%(\n                iteration*batch_size\/train_set_size, mse_train, mse_valid,mse_test))\n            try:\n                save_path = saver.save(sess, \"modelfile\\\\model\"+str(iteration)+\".ckpt\")\n            except Exception as e:\n                print(e)\n                if not os.path.exists(\"modelfile\\\\\"):\n                    os.makedirs(\"modelfile\\\\\")\n\n                save_path = saver.save(sess, \"modelfile\\\\model\"+str(iteration)+\".ckpt\")\n<\/code><\/pre>\n\n<p>The following is my sample what I am trying to execute:<br>\n<a href=\"https:\/\/gist.github.com\/JafferWilson\/572fae56dcf9c5f8fa0144f64904ae0e\" rel=\"nofollow noreferrer\">Same data Please click and see<\/a>   <\/p>\n\n<p>I am willing to add the Projector of the Tensorboard in my code. But I could not understand how I can make it. I want to visualize the different inputs I am giving for my training. I am supplying the following columns and trying to predict the <code>ohlc<\/code> values.  <\/p>\n\n<pre><code>'o', 'h', 'l', 'c', 'rel1', 'rel2', 'rel3', 'rel4', 'rel5', 'rel6', 'rel7', 'rel8'  \n<\/code><\/pre>\n\n<p>I want to visualize the above columns in the projector to know how they are relating with each other to give me the output.<br>\nPlease let me know what I can do to get what I am willing to.   <\/p>\n\n<p><strong>EDITED<\/strong>:    <\/p>\n\n<p>I have tried something as follows but cannot see the projector tab:  <\/p>\n\n<pre><code>tf.reset_default_graph()\n\nX = tf.placeholder(tf.float32, [None, n_steps, n_inputs])\ny = tf.placeholder(tf.float32, [None,n_outputs])\n\nsymbols = tf.placeholder(tf.int32, [None, 1], name='stock_labels')\nembed_matrix = tf.Variable(\n                tf.random_uniform([1, n_inputs],0.0, 1.0),\n                name=\"embed_matrix\"\n            )\nstacked_symbols = tf.tile(symbols, [batch_size,n_steps], name='stacked_stock_labels')\nstacked_embeds = tf.nn.embedding_lookup(embed_matrix, stacked_symbols)\n# stacked_embeds = tf.nn.embedding_lookup(embed_matrix)\n\n# After concat, inputs.shape = (batch_size, num_steps, input_size + embed_size)\ninputs_with_embed = tf.concat([X, stacked_embeds], axis=2, name=\"inputs_with_embed\")\nembed_matrix_summ = tf.summary.histogram(\"embed_matrix\", embed_matrix)\n<\/code><\/pre>\n\n<p>And edited the following lines in the session code:  <\/p>\n\n<pre><code>merged_sum = tf.summary.merge_all()\n    global_step = 0\n        # Set up the logs folder\n    writer = tf.summary.FileWriter('logs')\n    writer.add_graph(sess.graph)\n    projector_config = projector.ProjectorConfig()\n\n            # You can add multiple embeddings. Here we add only one.\n    added_embed = projector_config.embeddings.add()\n    added_embed.tensor_name = embed_matrix.name\n    # Link this tensor to its metadata file (e.g. labels).\n    shutil.copyfile(\"logs\\\\metadata.tsv\",\n                    \"logs\\\\metadata1.tsv\")\n    added_embed.metadata_path = \"metadata.tsv\"\n\n    # The next line writes a projector_config.pbtxt in the LOG_DIR. TensorBoard will\n    # read this file during startup.\n    projector.visualize_embeddings(writer, projector_config)\n    sess.run(tf.global_variables_initializer())\n    if iteration % int(1*train_set_size\/batch_size) == 0:\n        global_step += 1\n        mse_train = loss.eval(feed_dict={X: x_train, y: y_train}) \n        mse_valid = loss.eval(feed_dict={X: x_valid, y: y_valid}) \n        mse_test = loss.eval(feed_dict={X: x_test, y: y_test})\n        _,train_merge = sess.run([outputs,merged_sum], feed_dict={X: x_train, y: y_train})\n        writer.add_summary(train_merge, global_step=global_step)\n<\/code><\/pre>\n\n<p>Here is teh <a href=\"https:\/\/gist.github.com\/JafferWilson\/67c282bfd493c8d6bfdc5eb7ca12d949\" rel=\"nofollow noreferrer\">metadata.tsv file<\/a><br>\nPlease let me know what I missed.<\/p>\n","answers":[]}
+{"QuestionId":53427362,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2018-11-22T09:12:07.447","AcceptedAnswerId":null,"OwnerUserId":5596238.0,"Title":"Is there any method to 'combine layers' in tensorflow?","Body":"<p>Let me explain my question:<\/p>\n\n<p>I assume that there are some(more than one) tensorflow layers like below: <\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.placeholder(tf.float32,[1,64,64,3])\nconv_layer_1 = tf.layers.Conv2D(16,[3,3])\nconv_layer_2 = tf.layers.Conv2D(16,[3,3])\n<\/code><\/pre>\n\n<p>Is there any method to combine these layers to one tf.Layer object?That means:<\/p>\n\n<pre><code># conv_layer is the combined result    \n# conv_layer(x) is equal to conv_layer2(conv_layer1(x))\nconv_layer = some_function([conv_layer1,conv_layer2])\n<\/code><\/pre>\n\n<p>Currently I use a ugly method:<\/p>\n\n<pre><code>class MyLayer(tf.Layer)\n    ......\n    def combine(layers_list):\n        self.layers_list += layers_list\n    def __call__(x):\n        for layer in self.layers:\n            y = layer(y)\n        return y\n<\/code><\/pre>\n","answers":[{"AnswerId":"53427500","CreationDate":"2018-11-22T09:19:37.843","ParentId":null,"OwnerUserId":"6070667","Title":null,"Body":"<p>You can use tf.layers.concatenate for this case\nFor more details visit tensorflow site <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/concatenate\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/concatenate<\/a><\/p>\n"}]}
+{"QuestionId":53427489,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-22T09:19:18.443","AcceptedAnswerId":null,"OwnerUserId":9940477.0,"Title":"ValueError: Shape must be rank 2 but is rank 3","Body":"<p>I'm trying to run my code for a RNN, but got this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 2 but is rank 3 for 'echo_state_rnn_cell\/MatMul_3' (op: 'MatMul') with input shapes: [2,1,1000], [1000,1000].<\/p>\n<\/blockquote>\n\n<p>The thing that I don't understand is why my tensor has three dimensions, it should be [1,1000]. <\/p>\n\n<p>My code is:<\/p>\n\n<pre><code>esn_init_state = np.zeros([1, res_size], dtype=\"float32\")\n<\/code><\/pre>\n\n<p><code>tf.reset_default_graph()\nstatic_graph = tf.Graph()\nwith static_graph.as_default() as g:<\/code><\/p>\n\n<pre><code>rng = np.random.RandomState(random_seed)\ncell = EchoStateRNNCell(num_inputs=input_size, num_units=res_size, decay=0.1, leaky=leak,\n                        epsilon=1e-10, alpha=0.0100, rng=rng)\n\nX = tf.placeholder(tf.float32, [input_size+res_size, trainlen-initlen])\n\nYt = data[None, initlen+1:trainlen+1]\n\ninit_state = tf.placeholder(tf.float32, [1, res_size])\ninputs = tf.placeholder(tf.float32, [trainlen,input_size])  #modifica postuma\n\nstate = init_state\n\nfor t in range(trainlen):\n    prev_state = state\n    state = cell(inputs=inputs[t:t+1,:], state=prev_state)\n    if t &gt;= initlen:\n        X[:,t-initlen] = np.vstack((inputs,state))[:,0]\n<\/code><\/pre>\n\n<p>The error should be in: <code>state = cell(inputs=inputs[t:t+1,:], state=prev_state)<\/code>.<\/p>\n\n<p>Here is the definition of cell function:<\/p>\n\n<pre><code>def call(self, inputs, state):\n    new_state = (1 - self._leaky*self.decay) * state + self.decay*self._activation(math_ops.matmul(inputs,self.W) + math_ops.matmul(self._activation(state), self.U*self.rho_one))\n    output = self._activation(new_state)\n    return output, new_state   \n<\/code><\/pre>\n\n<p>Thanks for your help!<\/p>\n","answers":[]}
+{"QuestionId":53428443,"AnswerCount":1,"Tags":"<keras><tensorflow-lite>","CreationDate":"2018-11-22T10:06:12.350","AcceptedAnswerId":null,"OwnerUserId":8231289.0,"Title":"issue with converting keras h5 model file to tflite - Type Error('keyword argument not understood: ', 'interpolation')","Body":"<p>Built my own model in keras (only modified existing VGGSegnet version) which works perfectly. Trained Model with keras in google colab\nThen downloaded ex1.model.1 to my laptop\n(Inference works great on laptop)\nConverted the model to a h5 file using:<\/p>\n\n<pre><code>from keras.models import load_model, save_model\nm = load_model('ex1.model.1')\nm.save('model.h5')\n<\/code><\/pre>\n\n<p>Because I wanted to convert the model to the tflite using the terminal command from tflites for keras models <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/cmdline_examples\" rel=\"nofollow noreferrer\">website<\/a>:\n tflite_convert --output_file=newmode.tflite --keras_model_file=model.h5<\/p>\n\n<p>gives me this error <\/p>\n\n<pre><code>    Instructions for updating:\n`normal` is a deprecated alias for `truncated_normal`\nTraceback (most recent call last):\n  File \"\/home\/otto\/miniconda3\/bin\/tflite_convert\", line 11, in &lt;module&gt;\n    sys.exit(main())\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 412, in main\n    app.run(main=run_main, argv=sys.argv[:1])\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 408, in run_main\n    _convert_model(tflite_flags)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 100, in _convert_model\n    converter = _get_toco_converter(flags)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", line 87, in _get_toco_converter\n    return converter_fn(**converter_kwargs)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", line 368, in from_keras_model_file\n    keras_model = _keras.models.load_model(model_file)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 230, in load_model\n    model = model_from_config(model_config, custom_objects=custom_objects)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 310, in model_from_config\n    return deserialize(config, custom_objects=custom_objects)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/layers\/serialization.py\", line 64, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/utils\/generic_utils.py\", line 173, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/network.py\", line 1292, in from_config\n    process_layer(layer_data)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/network.py\", line 1278, in process_layer\n    layer = deserialize_layer(layer_data, custom_objects=custom_objects)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/layers\/serialization.py\", line 64, in deserialize\n    printable_module_name='layer')\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/utils\/generic_utils.py\", line 175, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 1606, in from_config\n    return cls(**config)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/layers\/convolutional.py\", line 1896, in __init__\n    super(UpSampling2D, self).__init__(**kwargs)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/training\/checkpointable\/base.py\", line 474, in _method_wrapper\n    method(self, *args, **kwargs)\n  File \"\/home\/otto\/miniconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 138, in __init__\n    raise TypeError('Keyword argument not understood:', kwarg)\nTypeError: ('Keyword argument not understood:', 'interpolation')\n<\/code><\/pre>\n\n<p>Google Colab was using keras version 2.2.4 and tensorflow 1.12.0 (and python2)\nMy Laptop used Linux 18.10,  and same keras\/ tensorflow versions (and python 3.5)<\/p>\n\n<p>Any Ideas? Thanks for your help!<\/p>\n\n<p>edit: checked whether training and running on local machine makes any difference - but still same error\nShould I provide the train.py and model file?<\/p>\n","answers":[{"AnswerId":"53431026","CreationDate":"2018-11-22T12:27:34.603","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Using tflite_convert command tool gives a ton of errors. If you wish to convert your keras model (.h5) to TensorFlow Lite format (.tflite), then you can do it with Google Colab. Follow these steps:<\/p>\n\n<ol>\n<li>Create a Google Colab Notebook. In the left top corner, click the \"UPLOAD\" button and upload your .h5 file.<\/li>\n<li><p>Create a code cell and insert this code.<\/p>\n\n<pre><code>from tensorflow.contrib import lite\nconverter = lite.TFLiteConverter.from_keras_model_file( 'model.h5' ) # Your model's name\nmodel = converter.convert()\nfile = open( 'model.tflite' , 'wb' ) \nfile.write( model )\n<\/code><\/pre><\/li>\n<li><p>Run the cell. You will get a model.tflite file. Right click on the file and select \"DOWNLOAD\" option.<\/p><\/li>\n<\/ol>\n\n<p>You can use this <a href=\"https:\/\/colab.research.google.com\/drive\/1IUIn9ffk5ICKujqPyuGaHL2irQ9Wmtpm\" rel=\"nofollow noreferrer\">notebook<\/a> as a reference.<\/p>\n"}]}
+{"QuestionId":53429231,"AnswerCount":2,"Tags":"<python><bash><tensorflow><google-cloud-ml>","CreationDate":"2018-11-22T10:47:59.377","AcceptedAnswerId":53431418.0,"OwnerUserId":5118420.0,"Title":"ML Engine Runtime version and Python version not supported","Body":"<p>I am trying to run Google clound ml-engine <a href=\"https:\/\/medium.com\/tensorflow\/training-and-serving-a-realtime-mobile-object-detector-in-30-minutes-with-cloud-tpus-b78971cf1193\" rel=\"nofollow noreferrer\">following this tutorial<\/a>, when executing this command :<\/p>\n\n<pre><code>$ gcloud ml-engine jobs submit training `whoami`_object_detection_`date +%s` \\ \n    --job-dir=gs:\/\/${YOUR_GCS_BUCKET}\/train\\\n    --packages dist\/object_detection-0.1.tar.gz,slim\/dist\/slim-0.1.tar.gz,\/tmp\/pycocotools\/pycocotools-2.0.tar.gz\\\n    --module-name object_detection.model_tpu_main \\\n    --runtime-version 1.10\\\n    --scale-tier BASIC_TPU\\\n    --region us-central1\\\n    --\n    --model_dir=gs:\/\/${YOUR_GCS_BUCKET}\/train\n    --tpu_zone us-central1\n    --pipeline_config_path=gs:\/\/${YOUR_GCS_BUCKET}\/data\/pipeline.config \n<\/code><\/pre>\n\n<p>commands which answers me :<\/p>\n\n<pre><code>  ERROR: (gcloud.ml-engine.jobs.submit.training) INVALID_ARGUMENT: Field: runtime_version Error: The specified runtime version '1.10' with the Python version '' is not supported for TPU training.  Please specify a different runtime version. See https:\/\/cloud.google.com\/ml\/docs\/concepts\/runtime-version-list for a list of supported versions\n- '@type': type.googleapis.com\/google.rpc.BadRequest\n  fieldViolations:\n  - description: The specified runtime version '1.10' with the Python version '' is\n  not supported for TPU training.  Please specify a different runtime version.\n  See https:\/\/cloud.google.com\/ml\/docs\/concepts\/runtime-version-list for a list\n  of supported versions\nfield: runtime_version\n<\/code><\/pre>\n\n<p>Because no version of python is detected, I added a <code>--config=config.yaml<\/code> argument to the command line:<\/p>\n\n<p><code>config.yaml<\/code>:<\/p>\n\n<pre><code>trainingInput:\n  pythonVersion: \"3.5\"\n<\/code><\/pre>\n\n<p>but the problem barely changed :<\/p>\n\n<pre><code>...\n - description: The specified runtime version '1.10' with the Python version '3.5'\n...\n<\/code><\/pre>\n\n<p>the <a href=\"https:\/\/cloud.google.com\/ml\/docs\/concepts\/runtime-version-list\" rel=\"nofollow noreferrer\">runtime version list<\/a> specify that runtime-version 1.10 is compatible with python 3.5, I also tried different runtime environement \/ python version which are required to work but my command keeps failling.<\/p>\n","answers":[{"AnswerId":"53431418","CreationDate":"2018-11-22T12:49:15.357","ParentId":null,"OwnerUserId":"10423558","Title":null,"Body":"<p>@jdehesa is right, the supported version is 1.9. The version 1.10 <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/runtime-version-list\" rel=\"nofollow noreferrer\">is not supported<\/a> for <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/runtime-version-list#tpu-support\" rel=\"nofollow noreferrer\">training Cloud TPU models<\/a>. change the runtime version editing this flag:<\/p>\n\n<pre><code>--runtime-version 1.9\n<\/code><\/pre>\n"},{"AnswerId":"56357189","CreationDate":"2019-05-29T09:22:31.893","ParentId":null,"OwnerUserId":"5555563","Title":null,"Body":"<p>For information the current supported  versions are now <strong>1.11<\/strong> or <strong>1.12<\/strong> \n<a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/runtime-version-list#tpu-support\" rel=\"nofollow noreferrer\">runtime-version-list<\/a><\/p>\n"}]}
+{"QuestionId":53429896,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-22T11:24:11.563","AcceptedAnswerId":null,"OwnerUserId":10577132.0,"Title":"How do I disable TensorFlow's eager execution?","Body":"<p>I am trying to learn TensorFlow. Currently, I am working with placeholders. When I tried to create the placeholder, I got an error: <code>RuntimeError: tf.placeholder() is not compatible with eager execution<\/code>, which makes sense as placeholders are not executable immediately. <\/p>\n\n<p>So, how do I turn eager execution off?<\/p>\n\n<p>I have never turned eager execution on in the first place, so I am not sure how it happened. Is there an opposite to <code>tf.disable_eager_execution()<\/code>?<\/p>\n","answers":[{"AnswerId":"55938423","CreationDate":"2019-05-01T15:10:38.860","ParentId":null,"OwnerUserId":"11437516","Title":null,"Body":"<p>I assume the you are using TensorFlow 2.0. In TF2, eager mode is turned on by default. However, there is a <code>disable_eager_execution()<\/code> in TensorFlow 2.0.0-alpha0 but it is hidden quite deep and cannot be directly accessed from top-level module namespace (i.e tf namespace).<\/p>\n\n<p>You can call the function like so:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework.ops import disable_eager_execution\n\ndisable_eager_execution()\n\na = tf.constant(1)\nb = tf.constant(2)\nc = a + b\nprint(c)\n<\/code><\/pre>\n\n<p><code>&gt;&gt;&gt;Tensor(\"add:0\", shape=(), dtype=int32)<\/code><\/p>\n\n<pre><code>print(disable_eager_execution.__doc__) \n<\/code><\/pre>\n\n<p><code>&gt;&gt;&gt;Disables eager execution.\nThis function can only be called before any Graphs, Ops, or Tensors have been  created. It can be used at the beginning of the program for complex migration projects from TensorFlow 1.x to 2.x.<\/code><\/p>\n"},{"AnswerId":"54098172","CreationDate":"2019-01-08T19:08:55.887","ParentId":null,"OwnerUserId":"2117745","Title":null,"Body":"<p>Assume you are using Tensorflow 2.0 preview release which has eager execution enabled by default. There is a <code>disable_eager_execution()<\/code> in v1 API, which you can put in the front of your code like:<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.compat.v1.disable_eager_execution()\n<\/code><\/pre>\n\n<p>On the other hand, if you are not using <strong>2.0 preview<\/strong>, please check if you accidentally enabled eager execution somewhere.<\/p>\n"}]}
+{"QuestionId":53430205,"AnswerCount":2,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-22T11:40:15.657","AcceptedAnswerId":53430651.0,"OwnerUserId":5669372.0,"Title":"How to process a voice input in Keras that is not one of the two speaker outputs?","Body":"<p>I'm trying to create a speaker recognition with a neural network using Keras and also the Fourier transformation to process the voice samples. The voice samples are me and my friend saying 'eeeee' for 3 seconds. Now the problem is if we give the neural network an input of someone else doing that ('ee' for 3 seconds), it still gives an output that indicates that it's  100% one of us. I output data using softmax so it gives about [1, 0] for my friend and [0, 1] for me. Sometimes it's like [0.95, 0.05]. <\/p>\n\n<p>It's working well except for if we input the data from another person like I said, it still gives like [1,0] although I would expect it to give something like [0.6, 0.4] because it's another voice. Now I have also tried using 2 features of MFCC but it doesn't seem to work either. Would making a third output and train it with random samples work (I myself don't really think so because it can't train for all different inputs)? Or how can I try to face this issue otherwise? I've been struggling with this issue for quite a while now so any help would be much appreciated!<\/p>\n","answers":[{"AnswerId":"53430299","CreationDate":"2018-11-22T11:45:04.133","ParentId":null,"OwnerUserId":"4279661","Title":null,"Body":"<p>It is not a simple matter. You need a lot more training examples and to do some tests.<\/p>\n\n<p>You COULD try to train something to be you-and-your-friend vs all, but it won't be that easy and (again) you will need lots of examples.<\/p>\n\n<p>It's very broad as a question, there are a few different approaches and i'm not sure Keras and neural networks are the correct answer for this specific task.<\/p>\n"},{"AnswerId":"53430651","CreationDate":"2018-11-22T12:06:02.140","ParentId":null,"OwnerUserId":"5024514","Title":null,"Body":"<p>A neural-net is - in essence - nothing more than a fancy feature-extractor and interpolator.<br>\nThere is no reason to expect anything specific for data that it's never seen, and this doesn't have much to do with working with the DTFT, MFCC, or I-Vectors, it's a basic principle of data-driven algorithms.<br>\nJust as a methodological explanation, not to be taken literally, finding a differentiation between two speakers going \"eeeeeeeee\" can be done by finding the  mean pitch and just deciding upon that one not-so-informative feature.<br>\n<em>Then<\/em> what do you think would happen when introducing a new utterance?<br>\nOne last point - there are <em>so many<\/em> ways to solve such a degenerate case, that you will most probably get some overfit. Which can also lead to unexpected results for out-of-sample data.  <\/p>\n\n<hr>\n\n<p>Regarding the general problem, there are a few different approaches, but I'd recommend having two stages:  <\/p>\n\n<ol>\n<li>1-Class SVM or something similar, to identify that the utterance is in the class <code>you or your friend<\/code>  <\/li>\n<li>The NN you trained.  <\/li>\n<\/ol>\n\n<p>Another option would be to get enough \"general speaker\" samples and add them as a third class. This is known in some contexts as the OOS (out of sample) category for classification. That can get you some good googling material :-)<\/p>\n"}]}
+{"QuestionId":53431120,"AnswerCount":0,"Tags":"<python><tensorflow><lambda><casting><keras>","CreationDate":"2018-11-22T12:32:42.240","AcceptedAnswerId":null,"OwnerUserId":10663471.0,"Title":"How do I cast properly a boolean layer in keras?","Body":"<p>I am working with some custom layers and having problems with the shape of them, when I work with it separately it works, but when I load the model to use in another one as a layer, it doesn't work anymore. Here is my layers definition:<\/p>\n\n<pre><code>def signumTransform(x):\n    \"\"\"\n    SIGNUM function\n    if positive 1\n    if negative -1\n    \"\"\"\n    import keras.backend\n    return keras.backend.sign(x)\n\ndef logical_or_layer(x):\n    \"\"\"Processing an OR operation\"\"\"\n    import keras.backend\n    #normalized to 0,1 \n    aux_array = keras.backend.sign(x)\n    aux_array = keras.backend.relu(aux_array)\n    # OR operation\n    aux_array = keras.backend.any(aux_array)\n    # casting back the True\/False to 1,0\n    aux_array = keras.backend.cast(aux_array, dtype='float32')\n    return aux_array\n\n\n#this is the input tensor\ninputs = Input(shape=(inputSize,), name='input')\n\n#this is the Neurule layer\nx = Dense(neurulesQt, activation='softsign', name='neurules')(inputs)\n#after each neuron layer, the outputs need to be put into SIGNUM (-1 or 1)\nx = Lambda(signumTransform, output_shape=lambda x:x, name='signumAfterNeurules')(x)\n\n#separating into 2 (2 possible outputs)\nlayer_split0 = Lambda( lambda x: x[:, :11], output_shape=[11], name='layer_split0')(x)\nlayer_split1 = Lambda( lambda x: x[:, 11:20], output_shape=[9], name='layer_split1')(x)\n#this is the OR layer\ny_0 = Lambda(logical_or_layer, output_shape=[1], name='or0')(layer_split0)\ny_1 = Lambda(logical_or_layer, output_shape=[1], name='or1')(layer_split1)\n\ny = Lambda(lambda x: K.stack([x[0], x[1]]),output_shape=[2],  name=\"output\")([y_0, y_1])\n<\/code><\/pre>\n\n<p>Until the <code>layer_split<\/code> everything works properly, but in my <code>y_0<\/code> and <code>y_1<\/code> I need to do an OR operation with <code>keras.backend.any()<\/code>, as a return I receive a boolean so I cast it back with <code>keras.backend.cast()<\/code>.<\/p>\n\n<p>If I use the Model as it is here described, it works...it compiles, can be validated and so on, but if I try to save it and load it, it simply loses one dimension, the batch dimension (<code>None<\/code>). The output in the summary is shown as <code>(None, 2)<\/code>, but when used as a layer and concatenated with another one, it shows <code>(2,)<\/code> and an error is thrown:<\/p>\n\n<p><code>InvalidArgumentError: Shape must be rank 2 but is rank 1 for 'merging_layer_10\/concat' (op: 'ConcatV2') with input shapes: [?,16], [2], [].<\/code><\/p>\n\n<p>How should I properly cast it in the <code>logical_or_layer<\/code> function? Should I change the <code>output_shape<\/code> in the Lambda Layer?<\/p>\n","answers":[]}
+{"QuestionId":53432173,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-22T13:34:30.290","AcceptedAnswerId":null,"OwnerUserId":3521069.0,"Title":"NotImplementedError: Operation of type AssignVariableOp (AssignVariableOp) is not supported on the TPU for inference","Body":"<p>I'm trying to export SavedModel of a classifier of the type TPUEstimator. Since I'm trying to export the model to run predictions on a GPU\/CPU, the use_tpu parameter of TPUEstimator was set to False.<\/p>\n\n<p>When I try to save the model, the following error is thrown:<\/p>\n\n<pre><code>NotImplementedError: Operation of type AssignVariableOp\n(AssignVariableOp) is not supported on the TPU for inference. Execution \nwill fail if this op is used in the graph. Make sure your variables are \nusing variable_scope.\n<\/code><\/pre>\n\n<p>Since I plan to serve the model through a GPU\/CPU, the Op shouldn't be a problem. How can I export this estimator as SavedModel?<\/p>\n","answers":[{"AnswerId":"56283150","CreationDate":"2019-05-23T21:28:44.573","ParentId":null,"OwnerUserId":"6456481","Title":null,"Body":"<p>This might help, just before calling export_savedmodel(...) call<\/p>\n\n<p><strong>estimator._export_to_tpu = False<\/strong> <\/p>\n"},{"AnswerId":"54742449","CreationDate":"2019-02-18T07:39:55.070","ParentId":null,"OwnerUserId":"7704053","Title":null,"Body":"<p>If you actually don't need TPU inference support you can create a tf.estimator.Estimator instead of a tf.contrib.tpu.TPUEstimator one, using the same model_fn and trained model. Then, you should be able to export the model.<\/p>\n"}]}
+{"QuestionId":53432395,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-22T13:46:21.973","AcceptedAnswerId":null,"OwnerUserId":10691336.0,"Title":"Tensorflow variable keeps changing as being evaluated","Body":"<p>I do not know why the value keeps changing every time it is calculated.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>import tensorflow as tf\nsample_num = 10\nradii = tf.random_uniform([sample_num,1],0,1)\nnorm1 = tf.reduce_max(radii,axis = 0)\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\nprint(sess.run(norm1))\nprint(sess.run(norm1))\nprint(sess.run(norm1))\n<\/code><\/pre>\n\n<p>All three print statements give me different values.<\/p>\n\n<pre><code>norm1.shape\n<\/code><\/pre>\n\n<p>gives me <code>TensorShape([Dimension(1)])<\/code>, so it looks like <code>norm1<\/code> correctly caught the maximum.\nI changed the sample number to 1 to see whether it keeps changing and it did.<\/p>\n\n<p>Can anyone help me figure out what the problem is?<\/p>\n","answers":[]}
+{"QuestionId":53433926,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorboard>","CreationDate":"2018-11-22T15:16:36.920","AcceptedAnswerId":null,"OwnerUserId":592568.0,"Title":"Keras with TensorBoard error \"InvalidArgumentError: You must feed a value for placeholder tensor\"","Body":"<p>I'm trying to get TensorBoard working with Keras.\nIt looks like I'm able to run an initial model with tf (1.12.0) and keras (2.1.6-tf). I have some simple code. Listed below:<\/p>\n\n<pre><code>%matplotlib inline\nfrom io import StringIO\n\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\n\ncsv = StringIO('''a,b,c,y\n0,1,2,0\n1,2,0,1\n0,2,1,0\n3,2,1,1\n3,1,2,0''')\ndata = pd.read_csv(csv)\n\ndef tb_cb(batch_size):\n    # visualize graphs and grandient\n    tb = tf.keras.callbacks.TensorBoard(log_dir='\/tmp\/test\/',\n                                       histogram_freq=1,\n                                       batch_size=batch_size, write_graph=True,\n                                       write_grads=True)\n    return tb\n\nm = tf.keras.Sequential([\n   # going to change 1 in the line below\n    tf.keras.layers.Dense(1, activation='relu', input_shape=(3,), name='hidden1'),\n    tf.keras.layers.Dense(1, activation='linear', name='output')\n])\nm.compile(loss='mse', optimizer='adam', metrics=['mae'])\nX = data.iloc[:,:3]\ny = data.y\nhist = m.fit(X, y, epochs=10, verbose=1, callbacks=[tb_cb(10)],\n            validation_data=(X,y))\n<\/code><\/pre>\n\n<p>The first time I run this I get TensorBoard output. I have then changed the number of neurons in the hidden layer and re-run the model.<\/p>\n\n<p>I get the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-11-4e4fb6f60bf0&gt; in &lt;module&gt;\n     15 y = data.y\n     16 hist = m.fit(X, y, epochs=10, verbose=1, callbacks=[tb_cb(10)],\n---&gt; 17             validation_data=(X,y))\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, max_queue_size, workers, use_multiprocessing, **kwargs)\n   1637           initial_epoch=initial_epoch,\n   1638           steps_per_epoch=steps_per_epoch,\n-&gt; 1639           validation_steps=validation_steps)\n   1640 \n   1641   def evaluate(self,\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py in fit_loop(model, inputs, targets, sample_weights, batch_size, epochs, verbose, callbacks, val_inputs, val_targets, val_sample_weights, shuffle, initial_epoch, steps_per_epoch, validation_steps)\n    231                 sample_weights=val_sample_weights,\n    232                 batch_size=batch_size,\n--&gt; 233                 verbose=0)\n    234             if not isinstance(val_outs, list):\n    235               val_outs = [val_outs]\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py in test_loop(model, inputs, targets, sample_weights, batch_size, verbose, steps)\n    437         ins_batch[i] = ins_batch[i].toarray()\n    438 \n--&gt; 439       batch_outs = f(ins_batch)\n    440 \n    441       if isinstance(batch_outs, list):\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/backend.py in __call__(self, inputs)\n   2984 \n   2985     fetched = self._callable_fn(*array_vals,\n-&gt; 2986                                 run_metadata=self.run_metadata)\n   2987     self._call_fetch_callbacks(fetched[-len(self._fetches):])\n   2988     return fetched[:len(self.outputs)]\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1437           ret = tf_session.TF_SessionRunCallable(\n   1438               self._session._session, self._handle, args, status,\n-&gt; 1439               run_metadata_ptr)\n   1440         if run_metadata:\n   1441           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n~\/.env\/364\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)\n    526             None, None,\n    527             compat.as_text(c_api.TF_Message(self.status.status)),\n--&gt; 528             c_api.TF_GetCode(self.status.status))\n    529     # Delete the underlying status object from memory otherwise it stays alive\n    530     # as there is a reference to status from this from the traceback due to\n\nInvalidArgumentError: You must feed a value for placeholder tensor 'dense_9_target' with dtype float and shape [?,?]\n     [[{{node dense_9_target}} = Placeholder[dtype=DT_FLOAT, shape=[?,?], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"54515509","CreationDate":"2019-02-04T11:52:21.377","ParentId":null,"OwnerUserId":"7954856","Title":null,"Body":"<p>You should always call Keras.clear_session() before building your model to avoid leftover nodes from previous runs in your tf graph.<\/p>\n\n<p>so add keras.backend.clear_session() right before you create your model. <\/p>\n"}]}
+{"QuestionId":53435534,"AnswerCount":1,"Tags":"<python><machine-learning><lambda><keras><deep-learning>","CreationDate":"2018-11-22T17:02:42.387","AcceptedAnswerId":53440381.0,"OwnerUserId":9044381.0,"Title":"Keras with Lambda-Layer","Body":"<p>I'm relatively new to Keras and I am about to build a Dueling Q-Network to train a KI. I found a code snippet to build a model that surprisingly seems to work. I just have no idea why, because I'm not too familiar with lambda expressions in Keras. Can anybody explain me how exactly the creation of the lambda-layer in the following model works?\nThank you very much in advantage!<\/p>\n\n<pre><code>def build_model():\n\n    model = Sequential()\n    model.add(Dense(units=16, activation='relu', input_dim = 2))\n    model.add(Dense(units=32, activation='relu'))\n    model.add(Dense(units=9, activation='relu'))\n\n    #I definitely don't understand how the following layer works:\n    model.add(Lambda(lambda i: K.expand_dims(i[:,0],-1) + i[:,1:] - K.mean(i[:,1:], keepdims=True), output_shape=(8,)))\n\n    model.add(Dense(units=8, activation='linear'))\n    model.compile(loss='mse',\n    optimizer = RMSprop(lr=0.001) )\n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"53440381","CreationDate":"2018-11-23T03:37:18.627","ParentId":null,"OwnerUserId":"3250829","Title":null,"Body":"<p>I am not familiar with your particular area of research, but I can tell you what that layer is doing.  A <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\"><code>Lambda<\/code><\/a> layer is when you want to define a custom operation on the inputs that don't come from anything that is predefined from Keras.  Specifically, you want to apply some custom operation to a tensor coming into the layer that Keras doesn't already handle.<\/p>\n\n<p>The input into the <code>Lambda<\/code> layer is an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Anonymous_function\" rel=\"nofollow noreferrer\">anonymous function<\/a> where the input is the tensor that is coming into this layer.  Note however that you can also specify any function or operation into this layer and it doesn't have to be an anonymous function... so long as it operates on the input tensor and produces an output one. You then define the operation you want to perform with this input tensor and it creates a corresponding output tensor that would be fed to the next layer.  This behaviour is of course assuming a feedforward network which is what I see here.  You can think of an anonymous function as a one time function that would be used to perform operations, but you don't want them lingering around later as you have no need for them anymore after you specify what to do with the input tensor. <\/p>\n\n<p><code>lambda i<\/code> thus denotes that you are creating an anonymous function which the <code>Lambda<\/code> layer would be operating on the input tensor defined as <code>i<\/code>.  <code>K.expand_dims<\/code> ensures that you add in singleton dimensions for the purposes of <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.0\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">broadcasting<\/a>.  In this case, we wish to take the first column of the input tensor <code>i[:,0]<\/code> which becomes a 1D array and making sure that the input tensor is a 2D array with a single column (i.e. going from a <code>N,<\/code> array to a <code>N x 1<\/code> array).  The <code>-1<\/code> argument is the axis you want to expand on.  Setting this to <code>-1<\/code> simply expands the very last dimension which in this case the last dimension is the first (and the only) dimension.<\/p>\n\n<p>If you're not used to broadcasting, the addition operation to this expanded array is a bit hard to understand but once you get the hang of it, it's one of the most powerful mechanisms in computing.  <code>i[:,1:]<\/code> slices the input tensor so that we consider the tensor from the second column to the end.  Under the hood, adding this sliced tensor with the expanded single column <code>i[:,0]<\/code> means that this column gets <strong>replicated<\/strong> and is added to every single column in <code>i[:,1:]<\/code> individually.<\/p>\n\n<p>For example if <code>i[:,0]<\/code> was <code>[1, 2, 3, 4]<\/code> and <code>i[:,1:]<\/code> was <code>[[4, 4, 4, 4], [5, 5, 5, 5], [6, 6, 6, 6]<\/code>, performing <code>K.expand_dims(i[:,0], -1) + i[:,1:]<\/code> results in <code>[[5, 6, 7, 8], [6, 7, 8, 9], [7, 8, 9, 10]]<\/code>.<\/p>\n\n<p>The last piece of the puzzle is this: <code>K.mean(i[:,1:], keepdims=True)<\/code>.  We take <code>K.expand_dims(i[:,0], -1) + i[:,1:]<\/code> then subtract this with <code>K.mean(i[:,1:], keepdims=True)<\/code>.  <code>K.mean<\/code> in this context will find the average of <strong>all<\/strong> values in the tensor for all the rows from the second column onwards.  This is the default behaviour of the operation.  Depending on how you use <code>K.mean<\/code>, one or more of the dimensions may drop.  An additional input into <code>K.mean<\/code> is the <code>axis<\/code> which allows you to specify which dimensions you want to analyze the average of in the tensor.  For example, if you did <code>axis=0<\/code>, this finds the mean of each column individually. \n This would reduce to a 1D tensor of values.  With the <code>keepdims<\/code> keyword, if you specified <code>keepdims=True<\/code>, this would ensure that the tensor is still 2D with the number of columns being 1 (i.e. a <code>N x 1<\/code> tensor instead of a <code>N,<\/code> tensor). The default behaviour is <code>false<\/code>.<\/p>\n\n<p>Therefore, by doing the <code>K.mean<\/code> operation, we ensure that the final result is <code>1 x 1<\/code> and this value is subtracted from every value of the <code>K.expand_dims(i[:,0],-1) + i[:,1:]<\/code> result.  This is again possible due to broadcasting.<\/p>\n\n<p>Finally, we make sure that the output shape of this operation gives a 1D tensor of size 8.<\/p>\n\n<h1><code>tl;dr<\/code><\/h1>\n\n<p>This operation is a custom one where we take the first column of an input tensor, add it to all of the other columns from the second column onwards and subtract every value of this result with the average of all of the other columns from the second column onwards.  Additionally, we constrain the output size of the tensor so that it is 1D with size 8.<\/p>\n"}]}
+{"QuestionId":53435711,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2018-11-22T17:16:32.713","AcceptedAnswerId":53447864.0,"OwnerUserId":10594993.0,"Title":"Average of the two inputs in multi-input deep learning model","Body":"<p>I want to create a multi-input deep learning model. The model takes two inputs (images) from different datasets and calculates the average of them. See the code:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>input1 = keras.layers.Input(shape=(16,))\nx1 = keras.layers.Dense(8, activation='relu')(input1)\ninput2 = keras.layers.Input(shape=(32,))\nx2 = keras.layers.Dense(8, activation='relu')(input2)\na = keras.layers.average([x1, x2])\n\nout = keras.layers.Dense(4)(a)\nmodel = keras.models.Model(inputs=[input1, input2], outputs=out)\n<\/code><\/pre>\n\n<p>I tried the following code to create the generator, but I got an error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input_imgen = ImageDataGenerator( \n                                  rotation_range=10,\n                                  shear_range=0.2,\n                                  zoom_range=0.1,\n                                  width_shift_range=0.1,\n                                  height_shift_range=0.1\n                                  )\n\n\n\ntest_imgen = ImageDataGenerator()\n\n\n\ndef generate_generator_multiple(generator,dir1, dir2, batch_size, img_height,img_width):\n\n\n    genX1 = generator.flow_from_directory(dir1,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n\n    genX2 = generator.flow_from_directory(dir2,\n                                          target_size = (img_height,img_width),\n                                          class_mode = 'categorical',\n                                          batch_size = batch_size,\n                                          shuffle=False, \n                                          seed=7)\n    while True:\n        X2i = genX2.next()    \n        X1i = genX1.next()\n        yield X1i[0], X2i[0] \n\n\n\ninputgenerator=generate_generator_multiple(generator=input_imgen,\n                                           dir1=train_data1,\n                                           dir2=train_data2,\n                                           batch_size=32,\n                                           img_height=224,\n                                           img_width=224)       \n\nvalidgenerator=generate_generator_multiple(generator=test_imgen,\n                                          dir1=valid_data1,\n                                          dir2=valid_data2,\n                                          batch_size=32,\n                                          img_height=224,\n                                          img_width=224) \n\ntestgenerator=generate_generator_multiple(generator=test_imgen,\n                                          dir1=test_data1,\n                                          dir2=test_data2,\n                                          batch_size=32,\n                                          img_height=224,\n                                          img_width=224)\n\n\n    # compile the model\n    multi_model.compile(\n            loss='categorical_crossentropy',\n            optimizer=Adam(lr=0.0001),\n            metrics=['accuracy']\n        )\n\n\n# train the model and save the history\nhistory = multi_model.fit_generator(\ninputgenerator,\nsteps_per_epoch=len(train_data) \/\/ batch_size,\nepochs=10,\nverbose=1,\nvalidation_data=validgenerator,\nvalidation_steps=len(valid_data) \/\/ batch_size,\nuse_multiprocessing=True,\nshuffle=False)\n<\/code><\/pre>\n\n<p>I got this error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 2 array(s), but instead got the following list of 1 arrays: [array([[[[108.930984, 108.930984, 108.930984],\n         [113.63957 , 113.63957 , 113.63957 ],\n         [113.07516 , 113.07516 , 113.07516 ],\n         ...,\n         [ 99.46968 ,  99.46968 ,  99.46968 ...\n<\/code><\/pre>\n\n<p>How can I solve this problem and create the generator?<\/p>\n","answers":[{"AnswerId":"53447864","CreationDate":"2018-11-23T13:45:16.147","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The error is raised because your model has two inputs but in this line:<\/p>\n\n<pre><code>yield X1i[0], X2i[0]\n<\/code><\/pre>\n\n<p>The generator would return a tuple of two arrays. In <code>fit_generator<\/code> the first one would be interpreted as the model input and the second one would be interpreted as the model output. Hence you would get that error saying you have only passed one input to the model. To resolve this put the inputs in a list and also return the labels, whatever they should be:<\/p>\n\n<pre><code>yield [X1i[0], X2i[0]], the_labels_array\n<\/code><\/pre>\n"}]}
+{"QuestionId":53435846,"AnswerCount":0,"Tags":"<python><tensorflow><keras><quantization>","CreationDate":"2018-11-22T17:27:43.803","AcceptedAnswerId":null,"OwnerUserId":6946404.0,"Title":"Is it possible to add a layer in a keras model that is only applied during inference?","Body":"<p>I am trying to quantize a custom keras model myself. Next to the weights I also want to quantize activations to have all fixed point values for the computationally intensive convolutional layers. Can I add a custom layer to my model that is only used for infernce and not during training? I want to train at full precision and only quantize post-training and for activations during infernce obviously. <\/p>\n\n<p>Another alternative might be to save the weights after training to a .h5 file and build a similar model that includes the activation quantization layers. But I am guessing keras would be thrown off if the model structure is not exactly the same.<\/p>\n\n<p>I am just getting started with ML and TF. Thanks so much for any tips! <\/p>\n","answers":[]}
+{"QuestionId":53436031,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-22T17:43:59.527","AcceptedAnswerId":null,"OwnerUserId":10692366.0,"Title":"libtensorflow_framework.so: undefined symbol: cuDevicePrimaryCtxGetState","Body":"<p>I have installed tensorflow-gpu with conda successfully. When I test doing import tensorflow I have the issue mentioned above. Any idea? I have checked my drivers, the nvidia toolkit and cudnn are intalled correctly and set the values of PATH, LD_LIBRARY_PATH and CUDA_HOME respectively.\n...Fri Nov 23 12:00:18 2018<br>\n+------------------------------------------------------+<br>\n| NVIDIA-SMI 340.107    Driver Version: 340.107        |<br>\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  Quadro FX 5600      Off  | 0000:02:00.0     N\/A |                  N\/A |\n| 61%   77C    P0    N\/A \/  N\/A |      2MiB \/  1535MiB |     N\/A      Default |\n+-------------------------------+----------------------+----------------------+<\/p>\n\n<p>+-----------------------------------------------------------------------------+\n| Compute processes:                                               GPU Memory |\n|  GPU       PID  Process name                                     Usage      |\n|=============================================================================|\n|    0            Not Supported                                               |\n+-----------------------------------------------------------------------------+<\/p>\n","answers":[{"AnswerId":"53436179","CreationDate":"2018-11-22T17:58:52.070","ParentId":null,"OwnerUserId":"2266772","Title":null,"Body":"<p>You need to have the proper minimum CUDA support (seems like it's 7 <a href=\"https:\/\/askubuntu.com\/questions\/988787\/nvidia-cuda-theano-could-not-find-symbol-cudeviceprimaryctxgetstate\">https:\/\/askubuntu.com\/questions\/988787\/nvidia-cuda-theano-could-not-find-symbol-cudeviceprimaryctxgetstate<\/a>) with cuDNN 3.<\/p>\n\n<p>Upgrade your drivers if possible to get this version.<\/p>\n\n<p>Otherwise, use either <code>tensorflow-mkl<\/code> or for older CPU models <code>tensorflow-eigen<\/code>.<\/p>\n"},{"AnswerId":"54406448","CreationDate":"2019-01-28T16:35:41.637","ParentId":null,"OwnerUserId":"4137497","Title":null,"Body":"<p>I had the same problem - I had already installed the required NVIDIA drivers on my Ubuntu machine, but when I tried to import Tensorflow I got the same error. <\/p>\n\n<p>So for me the problem was that I was using the Nouveau driver instead of the NVIDIA driver.  In order to fix the issue you need to go to <code>System Settings<\/code>-><code>Software &amp; Updates<\/code>-><code>Additional Drivers<\/code> and select the option <code>Using NVIDIA binary driver ...<\/code> and then click at the <code>Apply Changes<\/code> button. Then just reboot and you're done.<\/p>\n"}]}
+{"QuestionId":53436684,"AnswerCount":0,"Tags":"<python><keras><lstm><hybrid>","CreationDate":"2018-11-22T18:48:28.433","AcceptedAnswerId":null,"OwnerUserId":10398903.0,"Title":"How to implement a hybrid model in keras python?","Body":"<p>How to implement a hybrid model in python using HMM and LSTM?, and i don't know how to start and didn't found anything helpful when searching.<\/p>\n\n<p>Can anyone help me?<\/p>\n","answers":[]}
+{"QuestionId":53436736,"AnswerCount":0,"Tags":"<python><tensorflow><image-processing><keras><vgg-net>","CreationDate":"2018-11-22T18:54:17.310","AcceptedAnswerId":null,"OwnerUserId":3337089.0,"Title":"Layer output of Keras pre-trained VGG19 model produces different outputs for same input","Body":"<p>I'm extracting intermediate layer outputs from pretrained VGG19 ConvNet for a given image. I expect that if I give the same image twice, I should get the same output. But, I'm not getting the same output. Why is this happening and how to fix this?<\/p>\n\n<p><strong>Additional Details:<\/strong>\nI'm following <a href=\"https:\/\/arxiv.org\/pdf\/1609.04802.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>. They use a VGG19 ConvNet and extract the features from some intermediate layer (VGG22 means 2nd layer before 2nd convolution) for Super-Resolved Image and Ground-Truth Image. Then they calculate the mean squared error between these 2 feature sets and use it as a loss parameter. Now, my expectation is that if I give Ground Truth Image only twice, the mean squared error should be zero. But it is not happening? I'm getting different feature values at different iteration, but with same image. Also I noticed that, when I run the program again afresh, I get the same set of values. Code below for reference:<\/p>\n\n<pre><code>import numpy\nfrom keras import backend as K\nfrom keras.applications.vgg19 import VGG19, preprocess_input\nfrom keras.preprocessing.image import img_to_array, load_img\n\n\nmodel = VGG19()\nvgg22_layer_output = K.function([model.layers[0].input], [model.layers[5].output])\n\n# image_matrix is a 224x224x3 matrix for an RGB-image.\nhr_image_obj = load_img(hr_image_path)\nhr_image_matrix = img_to_array(hr_image_obj)\ncropped_hr_image = hr_image_matrix[0:224, 0:224, :]\nexpanded_image = numpy.expand_dims(cropped_hr_image, axis=0)\npreprocessed_image = preprocess_input(expanded_image)\nfeatures1 = vgg22_layer_output ([preprocessed_image])[0]\nfeatures2 = vgg22_layer_output ([preprocessed_image])[0]\n<\/code><\/pre>\n\n<p>Here, my expectation is that features1 = features2, which isn't<\/p>\n\n<p>Results:<\/p>\n\n<p>features1:<\/p>\n\n<pre><code>array([[[[2.15184002e+01, 1.81470230e+02, 0.00000000e+00, ...,\n          0.00000000e+00, 1.98130661e+02, 0.00000000e+00],\n         [2.27056488e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          0.00000000e+00, 0.00000000e+00, 0.00000000e+00],\n         [1.54923904e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          0.00000000e+00, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [2.29082489e+02, 2.58140778e+02, 0.00000000e+00, ...,\n          3.18900665e+02, 0.00000000e+00, 0.00000000e+00],\n         [1.58660873e+02, 1.24280603e+03, 0.00000000e+00, ...,\n          2.76672821e+02, 0.00000000e+00, 0.00000000e+00],\n         [2.66982513e+02, 4.27661194e+02, 0.00000000e+00, ...,\n          4.57434418e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 1.71959274e+02, ...,\n          0.00000000e+00, 1.25863232e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          3.51934662e+02, 4.45714081e+02, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          3.73108368e+02, 7.51479004e+02, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 3.06031370e+00, 0.00000000e+00, ...,\n          3.09630096e+02, 2.15055069e+02, 1.91232590e+02],\n         [0.00000000e+00, 1.33151245e+03, 0.00000000e+00, ...,\n          2.78728699e+02, 2.91452618e+01, 4.12124878e+02],\n         [1.13750778e+02, 3.04266022e+02, 0.00000000e+00, ...,\n          4.93073273e+02, 0.00000000e+00, 1.25463562e+02]],\n        [[0.00000000e+00, 2.36886551e+02, 1.87017990e+02, ...,\n          0.00000000e+00, 5.56484497e+02, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          1.29744125e+02, 5.47009888e+02, 0.00000000e+00],\n         [2.10977726e+01, 0.00000000e+00, 5.83388855e+02, ...,\n          3.78568268e+02, 1.76858459e+03, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 2.26063950e+02, 0.00000000e+00, ...,\n          1.74201874e+02, 1.10421577e+02, 2.92625153e+02],\n         [0.00000000e+00, 1.49054639e+03, 1.73763367e+02, ...,\n          3.43214760e+01, 1.41045761e+02, 5.26752502e+02],\n         [1.79130356e+02, 4.18553101e+02, 1.12429085e+02, ...,\n          2.08473053e+02, 0.00000000e+00, 1.46159897e+02]],\n        ...,\n        [[0.00000000e+00, 0.00000000e+00, 6.14884460e+02, ...,\n          4.48683044e+02, 2.60172217e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          1.05306360e+03, 5.45696045e+02, 0.00000000e+00],\n         [5.33453941e+01, 0.00000000e+00, 6.09368164e+02, ...,\n          7.00016541e+02, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          2.49793106e+02, 0.00000000e+00, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 3.44778638e+03, ...,\n          1.97339310e+02, 0.00000000e+00, 0.00000000e+00],\n         [1.27069351e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          1.85339737e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 4.90521271e+02, ...,\n          4.68645844e+02, 3.26934399e+03, 0.00000000e+00],\n         [2.26508102e+01, 0.00000000e+00, 7.08834915e+01, ...,\n          1.11953967e+03, 1.10590857e+03, 0.00000000e+00],\n         [1.11061287e+02, 0.00000000e+00, 8.05527405e+02, ...,\n          8.03228516e+02, 2.84233459e+02, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 1.10313757e+03, ...,\n          5.78258667e+02, 1.47924316e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 1.59146082e+03, ...,\n          7.10267578e+02, 6.43671143e+02, 0.00000000e+00],\n         [3.27744568e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          4.53388458e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 1.12306348e+03, ...,\n          1.63393646e+02, 3.52517969e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 6.36935806e+01, ...,\n          4.52494598e+02, 1.94326257e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          2.83666046e+02, 4.89346985e+02, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 1.02328314e+03, ...,\n          2.65413391e+02, 2.64639990e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          4.30894745e+02, 1.33343530e+03, 0.00000000e+00],\n         [7.57115707e+01, 0.00000000e+00, 0.00000000e+00, ...,\n          2.14354630e+02, 0.00000000e+00, 0.00000000e+00]]]],\n      dtype=float32)\n<\/code><\/pre>\n\n<p>features2:<\/p>\n\n<pre><code>array([[[[2.44103737e+01, 3.35516052e+02, 0.00000000e+00, ...,\n          0.00000000e+00, 2.06830643e+02, 0.00000000e+00],\n         [4.71717712e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          2.63770996e+02, 0.00000000e+00, 0.00000000e+00],\n         [3.93549591e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          1.77212814e+02, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [5.33919487e+01, 0.00000000e+00, 0.00000000e+00, ...,\n          1.85940536e+02, 0.00000000e+00, 0.00000000e+00],\n         [0.00000000e+00, 2.96363708e+02, 0.00000000e+00, ...,\n          1.09057648e+02, 0.00000000e+00, 0.00000000e+00],\n         [2.27105503e+01, 8.29022141e+01, 0.00000000e+00, ...,\n          1.38949188e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 2.47062546e+02, ...,\n          0.00000000e+00, 1.66465466e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          6.74320862e+02, 4.15592712e+02, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          8.65957825e+02, 8.59399170e+02, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          1.66129944e+02, 0.00000000e+00, 0.00000000e+00],\n         [0.00000000e+00, 2.76259674e+02, 0.00000000e+00, ...,\n          8.00474930e+01, 0.00000000e+00, 1.08291901e+02],\n         [0.00000000e+00, 2.20606117e+01, 0.00000000e+00, ...,\n          1.28005768e+02, 0.00000000e+00, 3.49725151e+01]],\n        [[0.00000000e+00, 2.14503006e+02, 8.82690811e+01, ...,\n          0.00000000e+00, 5.60968628e+02, 0.00000000e+00],\n         [3.28399010e+01, 0.00000000e+00, 0.00000000e+00, ...,\n          3.34213745e+02, 2.90819824e+02, 0.00000000e+00],\n         [8.66472626e+01, 0.00000000e+00, 1.10250635e+03, ...,\n          6.37486572e+02, 1.67822144e+03, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          5.90463066e+01, 0.00000000e+00, 9.77452278e+00],\n         [0.00000000e+00, 3.39350586e+02, 4.62688398e+00, ...,\n          1.32679808e+00, 0.00000000e+00, 1.65987671e+02],\n         [2.47563610e+01, 7.48269196e+01, 1.33592939e+01, ...,\n          6.36582108e+01, 0.00000000e+00, 5.70933228e+01]],\n        ...,\n        [[0.00000000e+00, 0.00000000e+00, 6.27470215e+02, ...,\n          2.55267532e+02, 2.27369629e+03, 0.00000000e+00],\n         [1.52827530e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          1.20087329e+03, 0.00000000e+00, 0.00000000e+00],\n         [1.33066071e+02, 0.00000000e+00, 5.95311890e+02, ...,\n          7.66817871e+02, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          4.81101898e+02, 0.00000000e+00, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 3.99484155e+03, ...,\n          5.40802429e+02, 0.00000000e+00, 0.00000000e+00],\n         [1.93494095e+02, 0.00000000e+00, 1.16481377e+02, ...,\n          3.75594208e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 5.08203369e+02, ...,\n          3.65947357e+02, 2.66369580e+03, 0.00000000e+00],\n         [2.29821182e+02, 0.00000000e+00, 3.83578918e+02, ...,\n          1.37410413e+03, 1.28806320e+02, 0.00000000e+00],\n         [1.89210968e+02, 0.00000000e+00, 9.40994324e+02, ...,\n          8.16117615e+02, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 1.34960962e+03, ...,\n          1.03916003e+03, 6.58975891e+02, 0.00000000e+00],\n         [6.77491531e+01, 0.00000000e+00, 2.07465186e+03, ...,\n          1.13461414e+03, 0.00000000e+00, 0.00000000e+00],\n         [2.96653259e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          6.33178528e+02, 0.00000000e+00, 0.00000000e+00]],\n        [[0.00000000e+00, 0.00000000e+00, 1.20268628e+03, ...,\n          6.86023560e+01, 2.83282886e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 3.50556335e+02, ...,\n          8.04942566e+02, 7.94925537e+02, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          3.74615967e+02, 0.00000000e+00, 0.00000000e+00],\n         ...,\n         [0.00000000e+00, 0.00000000e+00, 1.17975757e+03, ...,\n          6.35223450e+02, 1.62643567e+03, 0.00000000e+00],\n         [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, ...,\n          9.20189697e+02, 4.05781097e+02, 0.00000000e+00],\n         [1.26037315e+02, 0.00000000e+00, 0.00000000e+00, ...,\n          4.17285614e+02, 0.00000000e+00, 0.00000000e+00]]]],\n      dtype=float32)\n<\/code><\/pre>\n\n<p>The image is taken from <a href=\"https:\/\/drive.google.com\/drive\/folders\/1bXbXtGfyQiMJSdDjVRKLGxEZ4BoGUltC\" rel=\"nofollow noreferrer\">here<\/a>. Path: <code>v1.3\/Code\/Ours\/Images_GroundTruth\/BSD200\/335094.png<\/code><\/p>\n\n<p><strong>Edit 1<\/strong>: Added additional code and results<\/p>\n","answers":[]}
+{"QuestionId":53436960,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><vgg-net>","CreationDate":"2018-11-22T19:16:18.920","AcceptedAnswerId":null,"OwnerUserId":6017021.0,"Title":"Display output of vgg19 layer as image","Body":"<p>I was reading this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1508.06576.pdf\" rel=\"nofollow noreferrer\">Neural Style Transfer<\/a>. In this paper author reconstructs image from output of layers of vgg19. I am using Keras. The size of output of <code>block1_conv1<\/code> layer is <code>(1, 400, 533, 64)<\/code>. Here 1 is number of images as input, 400 is number of rows, 533 number of columns and 64 number of channels. When I try to reconstruct it as an image, I get an error as size of image is 13644800 which is not a multiple of 3, so I can't display the image in three channels. How can I reconstruct this image?<\/p>\n\n<p>I want to reconstruct images from layers as shown below:\n<a href=\"https:\/\/i.stack.imgur.com\/A7aMc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/A7aMc.png\" alt=\"image reconstruction from vgg layers\"><\/a>\nBelow is the code for the same:<\/p>\n\n<pre><code>from keras.preprocessing.image import load_img, img_to_array\nfrom scipy.misc import imsave\nimport numpy as np\nfrom keras.applications import vgg19\nfrom keras import backend as K\n\nCONTENT_IMAGE_FN = store image as input here\n\ndef preprocess_image(image_path):\n    img = load_img(image_path, target_size=(img_nrows, img_ncols))\n    img = img_to_array(img)\n    img = np.expand_dims(img, axis=0)\n    img = vgg19.preprocess_input(img)\n    return img\n\nwidth, height = load_img(CONTENT_IMAGE_FN).size\nimg_nrows = 400\nimg_ncols = int(width * img_nrows \/ height)\nbase_image = K.variable(preprocess_image(CONTENT_IMAGE_FN))\n\nRESULT_DIR = \"generated\/\"\nRESULT_PREFIX = RESULT_DIR + \"gen\"\nif not os.path.exists(RESULT_DIR):\n  os.makedirs(RESULT_DIR)\nresult_prefix = RESULT_PREFIX\n\n# this will contain our generated image\nif K.image_data_format() == 'channels_first':\n    combination_image = K.placeholder((1, 3, img_nrows, img_ncols))\nelse:\n    combination_image = K.placeholder((1, img_nrows, img_ncols, 3))\n\nx = preprocess_image(CONTENT_IMAGE_FN)\n\noutputs_dict = dict([(layer.name, layer.output) for layer in model.layers])\nfeature_layers = ['block1_conv1', 'block2_conv1',\n                  'block3_conv1', 'block4_conv1',\n                  'block5_conv1']\noutputs = []\nfor layer_name in feature_layers:\n  outputs.append(outputs_dict[layer_name])\nfunctor = K.function([combination_image], outputs )   # evaluation function\n\n# Testing\ntest = x\nlayer_outs = functor([test])\nprint(layer_outs)\n\nlayer_outs[0].reshape(400, -1 , 3) \/\/getting error here\n<\/code><\/pre>\n\n<p>I am getting following error:<\/p>\n\n<pre><code>ValueError: cannot reshape array of size 13644800 into shape (400,newaxis,3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53448895","CreationDate":"2018-11-23T14:55:45.350","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p><strong>You wrote:<\/strong><\/p>\n\n<blockquote>\n  <p>\"The size of output of <code>block1_conv1<\/code> layer is <code>(1, 400, 533, 64<\/code>). Here 1\n  is number of images as input, 400 is number of rows, 533 number of\n  columns and 64 number of channels\"\n  But this is not correct. The <code>block1_conv1<\/code> output corresponds 1 channel dimension(channel first), 400 * 533 image dimension and 64 <strong>filters<\/strong>.<\/p>\n<\/blockquote>\n\n<p>The error occurs, as you try to reshape a vector of <code>VGG19<\/code> output of an image input with a 1 channel <strong>(400 * 533 * 64 = 13644800)<\/strong> to a vector which correspond to a 3 channels output. <\/p>\n\n<p>Furthermore you have to pass <strong>3 channel<\/strong> input:<\/p>\n\n<p>From the <strong>VGG19<\/strong> code:<\/p>\n\n<blockquote>\n  <p><strong>input_shape:<\/strong> optional shape tuple, only to be specified\n              if <code>include_top<\/code> is False (otherwise the input shape\n              has to be <code>(224, 224, 3)<\/code>\n              (with <code>channels_last<\/code> data format)\n              or <code>(3, 224, 224)<\/code> (with <code>channels_first<\/code> data format).\n              It should have exactly 3 inputs channels,\n              and width and height should be no smaller than 32.\n              E.g. <code>(200, 200, 3)<\/code> would be one valid value.<\/p>\n<\/blockquote>\n\n<p>Thus your input images has to be 3 channels. If you even want to feed 1 channel(grayscale) images to <code>VGG19<\/code> you should make the following, if <code>channels first<\/code>:<\/p>\n\n<pre><code>X = np.repeat(X, 3 , axis=0) \n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>X = np.repeat(X, 3 , axis=2) \n<\/code><\/pre>\n\n<p>if <code>channels last<\/code> <strong>without batch dimension<\/strong> or<\/p>\n\n<pre><code>X = np.repeat(X, 3 , axis=3) \n<\/code><\/pre>\n\n<p><strong>with batch dimension<\/strong>.<\/p>\n\n<p>If you provide more information regarding the actual dimensions of your input matrices of your images and type of it(grayscale,RGB), I can give you more help upon needing it.<\/p>\n"}]}
+{"QuestionId":53436994,"AnswerCount":2,"Tags":"<python><multithreading><tensorflow>","CreationDate":"2018-11-22T19:20:28.700","AcceptedAnswerId":null,"OwnerUserId":1082410.0,"Title":"Trouble with the predict function of my model while in a separate thread","Body":"<p>I have a saved model that I can generate a prediction when I run it on the main thread. When I try to use the same function in a thread I get the following error. <\/p>\n\n<p>Note a summary of my code will be pasted after the error. <\/p>\n\n<pre><code>Exception in thread Thread-6:\nTraceback (most recent call last):\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\threading.py\", line 916, in _bootstrap_inner\n    self.run()\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\threading.py\", line 864, in run\n    self._target(*self._args, **self._kwargs)\n  File \"D:\\PiChess\\Core\\ComModule.py\", line 49, in threader\n    b = Globals.moduleManager.GetPrediction(predictionSet)\n  File \"D:\\PiChess\\Core\\ManagementModel.py\", line 56, in GetPrediction\n    return self.Model.predict(data, steps=10)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1878, in predict\n    self, x, batch_size=batch_size, verbose=verbose, steps=steps)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 295, in predict_loop\n    batch_outs = f(ins)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 2983, in __call__\n    self._make_callable(feed_arrays, feed_symbols, symbol_vals, session)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 2928, in _make_callable\n    callable_fn = session._make_callable_from_options(callable_opts)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1471, in _make_callable_from_options\n    return BaseSession._Callable(self, callable_options)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1425, in __init__\n    session._session, options_ptr, status)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 528, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Requested tensor connection from unknown node: \"dense_input:0\".\n\nException ignored in: &lt;bound method BaseSession._Callable.__del__ of &lt;tensorflow.python.client.session.BaseSession._Callable object at 0x000001E7749C5438&gt;&gt;\nTraceback (most recent call last):\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1455, in __del__\n    self._session._session, self._handle, status)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python36_64\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 528, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: No such callable handle: 2093487919952\n<\/code><\/pre>\n\n<p>The model is loaded into memory when I start running the code. To do this I created a globals object that loads the model the following way. <\/p>\n\n<pre><code>self.ModelName = ModelName\nself.Model = load_model(\"Models\/\" + self.ModelName + '.mdl')\nself.Model._make_predict_function()\n<\/code><\/pre>\n\n<p>To use the predict function it is run in a thread <\/p>\n\n<pre><code>predict = self.Model.predict(data, steps=10)\n<\/code><\/pre>\n\n<p>I know the data supplied is correct for I get a result if I just cut out the threading aspect out of my code and make my code into a single thread. <\/p>\n\n<p>I cut out most of code since its split up into multiple files and as I tried to paste here stack overflow told me I had to much code. <\/p>\n\n<p>If needed please let me know which sections of code would be most beneficial to post in as an edit. <\/p>\n","answers":[{"AnswerId":"55556990","CreationDate":"2019-04-07T08:21:24.733","ParentId":null,"OwnerUserId":"2473022","Title":null,"Body":"<p>You may need to store both the graph (and sometimes session) that were used to load the model, and pass them to your separate thread.<\/p>\n\n<p>If you are using Keras:<\/p>\n\n<pre><code>import keras.backend as K\n\n# Load model\nself.sess = tf.Session()\nK.set_session(sess)\nself.ModelName = ModelName\nself.Model = load_model(\"Models\/\" + self.ModelName + '.mdl')\nself.Model._make_predict_function()\nself.graph = tf.get_default_graph()\n\n# Predict\nK.set_session(sess)\nwith self.sess.as_default():\n    with self.graph.as_default():\n        predict = self.Model.predict(data, steps=10)\n<\/code><\/pre>\n\n<p>If you are using TensorFlow:<\/p>\n\n<pre><code># Load model\nself.sess = tf.Session()\nself.sess.run(tf.global_variables_initializer())\nself.Model.load(self.sess, \"Models\/\" + self.ModelName + '.mdl')\nself.graph = tf.get_default_graph()\n\n# Predict\nwith self.graph.as_default():\n    predict = self.Model.predict(self.sess, ...)\n<\/code><\/pre>\n\n<p>A somewhat related answer: <a href=\"https:\/\/stackoverflow.com\/a\/52023152\">The problem you are running into is because the tensorflow model is not loaded and used in the same thread.<\/a><\/p>\n"},{"AnswerId":"57252221","CreationDate":"2019-07-29T10:47:32.770","ParentId":null,"OwnerUserId":"7658313","Title":null,"Body":"<p>not sure u guys have found the solution or not, i just encountered the similar problem, and solved it, just share here.<\/p>\n\n<p><strong>In short, the fix is very simple: after u called the <code>K.clear_session()<\/code> u must reload the model again, even in same process!<\/strong><\/p>\n\n<p>My previous problematic code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = create_input_selection_model(config.input_shape)\n\nfor i in range(batches):\n  # Do a big batch prediction (100K or more depends on ur physical memory size).\n  sub_preds = model.predict_generator()\n  # For solving the memory problem, call 'K.clear_session()' to release GPU memory.\n  K.clear_session()\n<\/code><\/pre>\n\n<p>Above code will run into the problem OP posted.<\/p>\n\n<p>Fix is quite simple: reload the model in each for loop iteration:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for i in range(batches):\n\n  model = create_input_selection_model(config.input_shape) # &lt;- reload model here.\n\n  # Do a big batch prediction (100K or more depends on ur physical memory size).\n  sub_preds = model.predict_generator()\n  # For solving the memory problem, call 'K.clear_session()' to release GPU memory.\n  K.clear_session()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53437666,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-22T20:30:43.790","AcceptedAnswerId":53497727.0,"OwnerUserId":8085890.0,"Title":"AttributeError: 'Compose' object has no attribute 'Compose' (in Pytorch 0.2.1)","Body":"<p>This is the block of code where am getting this error:<\/p>\n\n<pre><code>train_transforms = transforms.Compose([transforms.RandomRotation(30),\n                                       transforms.RandomResizedCrop(224),\n                                       transforms.RandomHorizontalFlip(),\n                                       transforms.ToTensor(),\n                                       transforms.Normalize([0.5, 0.5, 0.5], \n                                                            [0.5, 0.5, 0.5])])\n<\/code><\/pre>\n\n<p>I've tried updating my torchvision but had no luck!<\/p>\n","answers":[{"AnswerId":"53497727","CreationDate":"2018-11-27T10:34:51.550","ParentId":null,"OwnerUserId":"10123877","Title":null,"Body":"<p>The problem is that you have a variable called <code>transforms<\/code> after <code>from torchvision import transforms<\/code> which has a compose of a certain type. This override the transform you import from the <code>torchvison<\/code>. Therefore when you run the above code it calls the transforms which is a <code>variable<\/code> not the one from <code>torchvision<\/code> module. <\/p>\n\n<p>It is advisable to rename the variable or if you are using jupyter notebook run the cell where you <code>import transforms<\/code> before running the cell with the code above.<\/p>\n"}]}
+{"QuestionId":53437710,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-22T20:35:05.803","AcceptedAnswerId":53437807.0,"OwnerUserId":5036064.0,"Title":"Why is the restored TensorFlow model, made with Keras, giving me 0 predictions?","Body":"<p>I made a classifier using MNIST digit data in Keras and I'm trying save and restore the model using only Tensorflow but I'm getting zeros for all ten of my classes.<\/p>\n\n<p>Here is the classifier:<\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\nfrom keras.datasets import mnist\nfrom keras.layers.core import K\nimport tensorflow as tf\n\n(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\n\nnetwork = models.Sequential()\nnetwork.add(layers.Dense(512, activation='relu', input_shape=(28*28,)))\nnetwork.add(layers.Dense(10, activation='softmax'))\n\nnetwork.compile(optimizer='rmsprop',\n                loss='categorical_crossentropy',\n                metrics=['accuracy'])\n\ntrain_images = train_images.reshape((60000, 28*28))\ntrain_images = train_images.astype('float32')\/255\n\ntest_images = test_images.reshape((10000, 28*28))\ntest_images = test_images.astype('float32')\/255\n\nfrom keras.utils import to_categorical\n\ntrain_labels = to_categorical(train_labels)\ntest_labels = to_categorical(test_labels)\n\nnetwork.fit(train_images, train_labels, epochs=5, batch_size=128)\n\ntest_loss, test_acc = network.evaluate(test_images, test_labels)\nprint('\\n\\ntest_acc: ', test_acc)\nprint('\\n' + network.input.op.name)\nprint('\\n' + network.output.op.name)\n\nsess = K.get_session()\nsaver = tf.train.Saver()\nsaver.save(sess, '.\/digit-model')\n<\/code><\/pre>\n\n<p>The results after running the script:<\/p>\n\n<pre><code>test_acc:  0.9799\n\ndense_1_input\n\ndense_2\/Softmax\n<\/code><\/pre>\n\n<p>Here is the script where I pass my own image, resize to 28 * 28 and feed it to the saved model:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport os\nimport cv2\n\n# First, pass the path of the image\ndir_path = os.path.dirname(os.path.realpath(__file__))\nimage_path = '.\/3.png'  # sys.argv[1]\nfilename = dir_path + '\/' + image_path\nimage_size = 28\nnum_channels = 1\nimages = []\n\n# Reading the image using OpenCV\nimage = cv2.imread(filename)\n\n# Resizing the image to our desired size and pre-processing will be done exactly as done\n# during training\nimage = cv2.resize(image, (image_size, image_size), 0, 0, cv2.INTER_LINEAR)\n\nif num_channels == 1:\n    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n    image = np.reshape(image, (image.shape[0], image.shape[1], 1))\n\nimages.append(image)\nimages = np.array(images, dtype=np.uint8)\nimages = images.astype('float32')\nimages = np.multiply(images, 1.0 \/ 255.0)\n\n# The input to the network is of shape [None image_size image_size num_channels].\n# Hence we reshape\nx_batch = images.reshape(1, image_size * image_size)\n\n# Let us restore the saved model\nsess = tf.Session()\n\n# Step-1: Recreate the network graph. At this step only graph is created.\nsaver = tf.train.import_meta_graph('.\/digit-model.meta')\n\n# Step-2: Now let's load the weights saved using the restore method\nsaver.restore(sess, tf.train.latest_checkpoint('.\/'))\n\n# Accessing the default graph which we have restored\ngraph = tf.get_default_graph()\n\n# for op in graph.get_operations():\n#     print(str(op.name))\n\n# Now, let's get hold of the op that we can be processed to get the output.\n# In the original network y_pred is the tensor that is the prediction of the network\ny_pred = graph.get_tensor_by_name(\"dense_2\/Softmax:0\")\n\n# Let's feed the images to the input placeholders\nx = graph.get_tensor_by_name(\"dense_1_input:0\")\ny_true = graph.get_tensor_by_name(\"dense_2\/Softmax:0\")\ny_test_images = np.zeros((1, 10))\n\n# Creating the feed_dict that is required to be fed to calculate y_pred\nfeed_dict_testing = {x: x_batch, y_true: y_test_images}\nresult = sess.run(y_pred, feed_dict=feed_dict_testing)\n\nfor r in result:\n    for i in r:\n        print(i)\n<\/code><\/pre>\n\n<p>These are the results I get back:<\/p>\n\n<pre><code>0.0\n0.0\n0.0\n0.0\n0.0\n0.0\n0.0\n0.0\n0.0\n0.0\n<\/code><\/pre>\n\n<p>I'm guessing it has something to do with the tensors I'm using for <strong>y_true<\/strong> and <strong>y_pred<\/strong>, I don't know which other tensor to use for them.  I would appriciate any help, thanks!<\/p>\n","answers":[{"AnswerId":"53437807","CreationDate":"2018-11-22T20:46:58.390","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>The problem in your code, that you should feed only the <code>x_batch<\/code> for the graph.\nSo the solution is altering the original:<\/p>\n\n<pre><code>feed_dict_testing = {x: x_batch, y_true: y_test_images\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>  feed_dict_testing = {x: x_batch}\n<\/code><\/pre>\n\n<p>We don't need the line at all:\n    <code>y_true = graph.get_tensor_by_name(\"dense_2\/Softmax:0\")<\/code>\nActually we don't need the <code>y_true<\/code> variable, we only need <strong>true<\/strong> labels for <strong>training phase<\/strong>.<\/p>\n\n<p>You can change the rows as well:<\/p>\n\n<pre><code>images = np.array(images, dtype=np.uint8)\nimages = images.astype('float32')\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>images = np.array(images, dtype=np.float32)\n<\/code><\/pre>\n\n<p>So, you win one row and save some plus computation.<\/p>\n\n<p>As a result of our efforts we get some reasonable results for an arbitrary image. <\/p>\n\n<pre><code>INFO:tensorflow:Restoring parameters from .\/digit-model\n3.4569392e-29\n6.898592e-28\n1.0\n2.9526584e-16\n0.0\n7.1692116e-14\n2.0023208e-12\n1.9861456e-14\n2.7171926e-23\n1.3212834e-28\n<\/code><\/pre>\n"}]}
+{"QuestionId":53438062,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-22T21:15:26.770","AcceptedAnswerId":53438522.0,"OwnerUserId":10692810.0,"Title":"how can I insert a Tensor into another Tensor in pytorch","Body":"<p>I have pytorch Tensor with shape <code>(batch_size, step, vec_size)<\/code>, for example, a <code>Tensor(32, 64, 128)<\/code>, let's call it A.<\/p>\n\n<p>I have another <code>Tensor(batch_size, vec_size)<\/code>, e.g. <code>Tensor(32, 128)<\/code>, let's call it B.<\/p>\n\n<p>I want to insert B into a certain position at axis 1 of A. The insert positions are given in a <code>Tensor(batch_size)<\/code>, named P.<\/p>\n\n<p>I understand there is no Empty tensor(like an empty list) in pytorch, so, I initialize A as zeros, and add B at a certain position at axis 1 of A.<\/p>\n\n<p><code>A = Variable(torch.zeros(batch_size, step, vec_size))<\/code><\/p>\n\n<p>What I'm doing is like:<\/p>\n\n<pre><code>for i in range(batch_size):\n    pos = P[i]\n    A[i][pos] = A[i][pos] + B[i]\n<\/code><\/pre>\n\n<p>But I get an Error:<\/p>\n\n<pre><code>RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation\n<\/code><\/pre>\n\n<p>Then, I make a clone of A each inside the loop:<\/p>\n\n<pre><code>for i in range(batch_size):\n    A_clone =  A.clone()\n    pos = P[i]\n    A_clone[i][pos] = A_clone[i][pos] + B[i]\n<\/code><\/pre>\n\n<p>This is very slow for autograd, I wonder if there any better solutions? Thank you.<\/p>\n","answers":[{"AnswerId":"53438522","CreationDate":"2018-11-22T22:11:36.450","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>You can use a mask instead of cloning.<\/p>\n\n<p>See the code below<\/p>\n\n<pre><code># setup\nbatch, step, vec_size = 64, 10, 128 \nA = torch.rand((batch, step, vec_size))\nB = torch.rand((batch, vec_size))\npos = torch.randint(10, (64,)).long()\n\n# computations\n# create a mask where pos is 0 if it is to be replaced\nmask = torch.ones( (batch, step)).view(batch,step,1).float()\nmask[torch.arange(batch), pos]=0\n\n# expand B to have same dimension as A and compute the result\nresult = A*mask + B.unsqueeze(dim=1).expand([-1, step, -1])*(1-mask)\n<\/code><\/pre>\n\n<p>This way you avoid using for loops and cloning as well. <\/p>\n"}]}
+{"QuestionId":53438146,"AnswerCount":0,"Tags":"<pytorch><torchvision>","CreationDate":"2018-11-22T21:26:50.123","AcceptedAnswerId":null,"OwnerUserId":10377169.0,"Title":"How can I change the padded input size per channel in Pytorch?","Body":"<p>I am trying to set up an image classifier using Pytorch. My sample images have 4 channels and are 28x28 pixels in size. I am trying to use the built-in torchvision.models.inception_v3() as my model. Whenever I try to run my code, I get this error: <\/p>\n\n<blockquote>\n  <p>RuntimeError: Calculated padded input size per channel: (1 x 1).\n  Kernel size: (3 x 3). Kernel size can't greater than actual input size\n  at\n  \/opt\/conda\/conda-bld\/pytorch_1524584710464\/work\/aten\/src\/THNN\/generic\/SpatialConvolutionMM.c:48<\/p>\n<\/blockquote>\n\n<p>I can't find how to change the padded input size per channel or quite figure out what the error means. I figure that I must modify the padded input size per channel since I can't edit the Kernel size in the pre-made model.<\/p>\n\n<p>I have tried padding, but it didn't help.\nHere is a shortened part of my code that throws the error when I call train():<\/p>\n\n<pre><code>import torch\nimport torchvision as tv\nimport torch.optim as optim\nfrom torch import nn\nfrom torch.utils.data import DataLoader\n\nmodel = tv.models.inception_v3()\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.0001, weight_decay=0)\nlr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=4, gamma=0.9)  \n\ntrn_dataset = tv.datasets.ImageFolder(\n    \"D:\/tests\/classification_test_data\/trn\",\n    transform=tv.transforms.Compose([tv.transforms.RandomRotation((0,275)), tv.transforms.RandomHorizontalFlip(),\n                                  tv.transforms.ToTensor()]))\ntrn_dataloader = DataLoader(trn_dataset, batch_size=32, num_workers=4, shuffle=True)\n\nfor epoch in range(0, 10):\n    train(trn_dataloader, model, criterion, optimizer, lr_scheduler, 6, 32)\nprint(\"End of training\")\n\n\ndef train(train_loader, model, criterion, optimizer, scheduler, num_classes, batch_size):\n    model.train()\n    scheduler.step()\n\n    for index, data in enumerate(train_loader):\n        inputs, labels = data\n        optimizer.zero_grad()\n        outputs = model(inputs)\n        outputs_flatten = flatten_outputs(outputs, num_classes)\n        loss = criterion(outputs_flatten, labels)\n        loss.backward()\n        optimizer.step()\n\n\ndef flatten_outputs(predictions, number_of_classes):\n    logits_permuted = predictions.permute(0, 2, 3, 1)\n    logits_permuted_cont = logits_permuted.contiguous()\n    outputs_flatten = logits_permuted_cont.view(-1, number_of_classes)\n    return outputs_flatten\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53438264,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-22T21:40:30.423","AcceptedAnswerId":53438954.0,"OwnerUserId":10685804.0,"Title":"Keras concatenate with shapes [(1, 8), (None, 32)]","Body":"<p>My network consists of <strong>LSTM and Dense parts connected together<\/strong> by the other Dense part and I <strong>cannot concatenate inputs of size [(1, 8), (None, 32)]<\/strong>. <code>Reshape<\/code> and <code>Flatten<\/code> do not work.<\/p>\n\n<p>Here's the architecture:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_model_equal(dropout_rate=0.25):\n\n    curve_input_1 = Input(batch_shape=(1, None, 1), name='curve_input_1')\n    lstm_1 = LSTM(256, return_sequences=True, dropout=0.1)(curve_input_1)\n    lstm_1 = LSTM(64, dropout=0.1)(lstm_1)\n    lstm_out = Dense(8)(lstm_1)\n\n    metadata_input = Input(shape=(31,), name='metadata_input')\n\n    dense_1 = Dense(512, activation='relu')(metadata_input)\n    dense_1 = BatchNormalization()(dense_1)\n    dense_1 = Dropout(dropout_rate)(dense_1)\n\n    dense_out = Dense(32)(dense_1)\n\n    x = keras.layers.concatenate([lstm_out, dense_out], axis=1)\n\n    output_hidden = Dense(64)(x)\n    output_hidden = BatchNormalization()(output_hidden)\n    output_hidden = Dropout(dropout_rate)(output_hidden)\n\n    output = Dense(n_classes, activation='softmax', name='output')(output_hidden)\n\n    model = Model(inputs=[curve_input_1, metadata_input], outputs=output)\n    return model\n<\/code><\/pre>\n\n<p>When I train this model via <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit([x_train, x_metadata], y_train,\n    validation_data=[[x_valid, x_metadata_val], y_valid], \n    epochs=n_epoch,\n    batch_size=n_batch, shuffle=True, \n    verbose=2, callbacks=[checkPoint]\n)  \n<\/code><\/pre>\n\n<p>I get an error <\/p>\n\n<blockquote>\n  <p>ValueError: A Concatenate layer requires inputs with matching shapes except for the concat axis. <strong>Got inputs shapes: [(1, 8), (None, 32)]<\/strong><\/p>\n<\/blockquote>\n\n<p>When I add <code>Reshape<\/code> layer like <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dense_out = Dense(32)(dense_4)\ndense_out = Reshape((1, 32))(dense_out)\n\nx = keras.layers.concatenate([lstm_out, dense_out], axis=1)\n<\/code><\/pre>\n\n<p>I get <\/p>\n\n<blockquote>\n  <p>ValueError: A Concatenate layer requires inputs with matching shapes except for the concat axis. <strong>Got inputs shapes: [(1, 8), (None, 1, 32)]<\/strong><\/p>\n<\/blockquote>\n\n<p><code>Reshape<\/code> layer <code>input_shape=(32,)<\/code> or <code>input_shape=(None, 32)<\/code> parameters do not change the situation, error and shapes are the same.<\/p>\n\n<p>Adding <code>Reshape<\/code> to LSTM like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>curve_input_1 = Input(batch_shape=(1, None, 1), name='curve_input_1')\nlstm_first_1 = LSTM(256, return_sequences=True, dropout=0.1, name='lstm_first_1')(curve_input_1)\nlstm_second_1 = LSTM(64, dropout=0.1, name='lstm_second_1')(lstm_first_1)\nlstm_out = Dense(8)(lstm_second_1)\nlstm_out = Reshape((None, 8))(lstm_out)\n<\/code><\/pre>\n\n<p>Produces an error<\/p>\n\n<blockquote>\n  <p>ValueError: Tried to convert 'shape' to a tensor and failed. Error: None values not supported.<\/p>\n<\/blockquote>\n\n<p>Changing <code>concatenate<\/code> <code>axis<\/code> parameter to <code>0<\/code>, <code>1<\/code> and <code>-1<\/code> doesn't help.<\/p>\n\n<p>Changing <code>Dense<\/code> part input shape doesn't help. When I do <code>metadata_input = Input(shape=(1, 31), name='metadata_input')<\/code> instead of metadata_input = Input(shape=(31,), name='metadata_input') it produces an error with <code>[(1, 8), (None, 1, 32)]<\/code> dimensions.<\/p>\n\n<p>My guess is that I need to transform data either to <code>[(1, 8), (1, 32)]<\/code> or to <code>[(None, 8), (None, 32)]<\/code> shape, but <code>Reshape<\/code> and <code>Flatten<\/code> layers didn't help. <\/p>\n\n<p>There should be an easy way to do that that I missed.<\/p>\n","answers":[{"AnswerId":"53438954","CreationDate":"2018-11-22T23:12:40.593","ParentId":null,"OwnerUserId":"3254400","Title":null,"Body":"<p>I think the problem could be the use <code>batch_shape<\/code> for the first <code>Input<\/code> and <code>shape<\/code> for the second one. <\/p>\n\n<p>With the first input, your batch size is hardcoded as <code>1<\/code> and your input data has 2 extra dimensions, <code>None<\/code> (unspecified) and <code>1<\/code>. <\/p>\n\n<p>For the second input, since you are using <code>shape<\/code>, you are declaring that your input has batch size unespecified and data has one dimension of <code>31<\/code> values.<\/p>\n\n<p>Note that using <code>shape=(31,)<\/code> is the same as using <code>batch_shape=(None, 31)<\/code> (from  <a href=\"https:\/\/stackoverflow.com\/questions\/44792106\/when-do-you-use-input-shape-vs-batch-shape-in-keras\">here<\/a>).<\/p>\n\n<p>Aligning both is working for me, at least at model declaration time (I was unable to run the fit though, and I'm not sure if I'm missing something and this solution doesn't fit your use case.<\/p>\n\n<p>So, to summarize, you can try:<\/p>\n\n<pre><code>curve_input_1 = Input(batch_shape=(1, None, 1), name='curve_input_1')\nmetadata_input = Input(batch_shape=(1, 31), name='metadata_input')\n<\/code><\/pre>\n\n<p>Or:<\/p>\n\n<pre><code>curve_input_1 = Input(batch_shape=(1, None, 1), name='curve_input_1')\nmetadata_input = Input(batch_shape=(1, 31), name='metadata_input')\n<\/code><\/pre>\n\n<p>Which is equivalent to:<\/p>\n\n<pre><code>curve_input_1 = Input(shape=(None, 1, ), name='curve_input_1')\nmetadata_input = Input(shape=(31, ), name='metadata_input')\n<\/code><\/pre>\n\n<p>Please, let me know it this worked or lead you in a good direction!<\/p>\n"}]}
+{"QuestionId":53439070,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2018-11-22T23:31:14.827","AcceptedAnswerId":null,"OwnerUserId":2062298.0,"Title":"Understanding Keras LSTM NN input & output for binary classification","Body":"<p>I am trying to create a simple LSTM network that would - based on the last 16 time frames - provide some output. Let's say I have a dataset with 112000 rows (measurements) and 7 columns (6 features + class). What I understand is that I have to \"pack\" the dataset into X number of 16 elements long batches. With 112000 rows that would mean 112000\/16 = 7000 batches, therefore a numpy 3D array with shape (7000, 16, 7). Splitting this array for train and test data I get shapes:<\/p>\n\n<pre><code>xtrain.shape == (5000, 16, 6)\nytrain.shape == (5000, 16)\nxtest.shape == (2000, 16, 6)\nytest.shape == (2000, 16)\n<\/code><\/pre>\n\n<p>My model looks like this:<\/p>\n\n<pre><code>model.add(keras.layers.LSTM(8, input_shape=(16, 6), stateful=True, batch_size=16, name=\"input\"));\nmodel.add(keras.layers.Dense(5, activation=\"relu\", name=\"hidden1\"));\nmodel.add(keras.layers.Dense(1, activation=\"sigmoid\", name=\"output\"));\nmodel.compile(optimizer=\"rmsprop\", loss=\"binary_crossentropy\", metrics=[\"accuracy\"]);\n\nmodel.fit(xtrain, ytrain, batch_size=16, epochs=10);\n<\/code><\/pre>\n\n<p>However after trying to fit the model I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking target: expected output to have shape (1,) but got array with shape (16,)<\/p>\n<\/blockquote>\n\n<p>What I guess is wrong is that the model expects a single output per batch (so the ytrain shape should be (5000,)), instead of 16 outputs (one for every entry in a batch - (5000, 16)).<\/p>\n\n<p>If that is the case, should I, instead of packing the data like this, create a 16 elements long batch for every output? Therefore having <\/p>\n\n<pre><code>xtrain.shape == (80000, 16, 6)\nytrain.shape == (80000,)\nxtest.shape == (32000, 16, 6)\nytest.shape == (32000,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53439223","CreationDate":"2018-11-22T23:56:36.873","ParentId":null,"OwnerUserId":"3254400","Title":null,"Body":"<p>You are close with the last comments of the question. Since it's a binary classification problem, you should have <code>1<\/code> output per input, so you need to get rid of the <code>16<\/code> in you <code>y<\/code>s and replace it for a <code>1<\/code>.<\/p>\n\n<p>Besides, you need to be able to divide the train set by your batch size, so you can use <code>5008<\/code> for example.<\/p>\n\n<p>In fact:<\/p>\n\n<pre><code>ytrain.shape == (5000, 1)\n<\/code><\/pre>\n\n<p>Passes the error you mention, but raises a new one:<\/p>\n\n<blockquote>\n  <p>ValueError: In a stateful network, you should only pass inputs with a number of samples that can be divided by the batch size. Found: 5000 samples<\/p>\n<\/blockquote>\n\n<p>Which is addressed by ensuring that:<\/p>\n\n<pre><code>xtrain.shape == (5008, 16, 6)\nytrain.shape == (5008, 1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53439157,"AnswerCount":0,"Tags":"<tensorflow><ubuntu-16.04>","CreationDate":"2018-11-22T23:44:49.433","AcceptedAnswerId":null,"OwnerUserId":4330326.0,"Title":"Tensorflow libcublas.so.9.0 error - CUDA Version 10.0","Body":"<p>I cannot get around the problem regarding the configuration of tensorflow, here it is as follows: <\/p>\n\n<pre><code>(venv) home@home:~\/Downloads\/venv$ python models\/research\/object_detection\/inputs_test.py\nTraceback (most recent call last):\n  File \"\/home\/home\/Downloads\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/home\/Downloads\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/home\/Downloads\/venv\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/home\/home\/Downloads\/venv\/lib\/python3.5\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/home\/Downloads\/venv\/lib\/python3.5\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\n**ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory**\n<\/code><\/pre>\n\n<p>~\/.bashrc has been modified as follows (last lines of code):<\/p>\n\n<pre><code>export PATH=\/usr\/local\/cuda-10.0\/bin${PATH:+:${PATH}}\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64:${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}\n<\/code><\/pre>\n\n<p>and libraries tensorflow\/models\/research\/ and slim directories have been appended to PYTHONPATH: <\/p>\n\n<pre><code>export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`\/slim\n<\/code><\/pre>\n\n<p>Any idea?<\/p>\n","answers":[]}
+{"QuestionId":53439765,"AnswerCount":1,"Tags":"<python><keras><reshape><keras-layer>","CreationDate":"2018-11-23T01:47:20.333","AcceptedAnswerId":null,"OwnerUserId":389122.0,"Title":"In Keras, how to use Reshape layer with None dimension?","Body":"<p>In my model, a layer has a shape of <code>[None, None, 40, 64]<\/code>. I want to reshape this into <code>[None, None, 40*64]<\/code>. However, if I simply do the following:<\/p>\n\n<pre><code>reshaped_layer = Reshape((None, None, 40*64))(my_layer)\n<\/code><\/pre>\n\n<p>It throws an error complaining that <code>None values not supported<\/code>.<\/p>\n\n<p>(Just to be clear, this is not <code>tf.keras<\/code>, this is just Keras).<\/p>\n","answers":[{"AnswerId":"53446949","CreationDate":"2018-11-23T12:44:07.433","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First of all, the argument you pass to <code>Reshape<\/code> layer is the desired shape of one sample in the batch and not the whole batch of samples. So since each of the samples in the batch is a 3D tensor, the argument must also consider only that 3D tensor (i.e. excluding the batch axis).<\/p>\n\n<p>Second, you can use <code>-1<\/code> as the shape of only one axis. It tells to the <code>Reshape<\/code> layer to automatically infer the shape of that axis based on the shape of other axes you provide. So considering these two points, it would be:<\/p>\n\n<pre><code>reshaped_out = Reshape((-1, 40*64))(layer_out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53440128,"AnswerCount":1,"Tags":"<python><keras><generator>","CreationDate":"2018-11-23T02:54:36.070","AcceptedAnswerId":53470907.0,"OwnerUserId":5171213.0,"Title":"Keras fit_generator issue","Body":"<p>I followed <a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">this tutorial<\/a> to create a custom generator for my Keras model. Here is an MWE that shows the issues I'm facing:<\/p>\n\n<pre><code>import sys, keras\nimport numpy as np\nimport tensorflow as tf\nimport pandas as pd\nfrom keras.models import Model\nfrom keras.layers import Dense, Input\nfrom keras.optimizers import Adam\nfrom keras.losses import binary_crossentropy\n\nclass DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, list_IDs, batch_size, shuffle=False):\n        'Initialization'\n        self.batch_size = batch_size\n        self.list_IDs = list_IDs\n        self.shuffle = shuffle\n        self.on_epoch_end()\n\n    def __len__(self):\n        'Denotes the number of batches per epoch'\n        return int(np.floor(len(self.list_IDs) \/ self.batch_size))\n\n    def __getitem__(self, index):\n        'Generate one batch of data'\n        # Generate indexes of the batch\n        #print('self.batch_size: ', self.batch_size)\n        print('index: ', index)\n        sys.exit()\n\n    def on_epoch_end(self):\n        'Updates indexes after each epoch'\n        self.indexes = np.arange(len(self.list_IDs))\n        print('self.indexes: ', self.indexes)\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n\n    def __data_generation(self, list_IDs_temp):\n        'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n\n        X1 = np.empty((self.batch_size, 10), dtype=float)\n        X2 = np.empty((self.batch_size, 12),  dtype=int)\n\n        #Generate data\n        for i, ID in enumerate(list_IDs_temp):\n            print('i is: ', i, 'ID is: ', ID)\n\n            #Preprocess this sample (omitted)\n            X1[i,] = np.repeat(1, X1.shape[1])\n            X2[i,] = np.repeat(2, X2.shape[1])\n\n        Y = X1[:,:-1]\n        return X1, X2, Y\n\nif __name__=='__main__':\n    train_ids_to_use = list(np.arange(1, 321)) #1, 2, ...,320 \n    valid_ids_to_use = list(np.arange(321, 481)) #321, 322, ..., 480\n\n    params = {'batch_size': 32}\n\n    train_generator = DataGenerator(train_ids_to_use, **params)\n    valid_generator = DataGenerator(valid_ids_to_use, **params)\n\n    #Build a toy model\n    input_1 = Input(shape=(3, 10))\n    input_2 = Input(shape=(3, 12))\n    y_input = Input(shape=(3, 10))\n\n    concat_1 = keras.layers.concatenate([input_1, input_2])\n    concat_2 = keras.layers.concatenate([concat_1, y_input])\n\n    dense_1 = Dense(10, activation='relu')(concat_2)\n    output_1 = Dense(10, activation='sigmoid')(dense_1)\n\n    model = Model([input_1, input_2, y_input], output_1)\n    print(model.summary())\n\n    #Compile and fit_generator\n    model.compile(optimizer=Adam(lr=0.001), loss=binary_crossentropy)\n    model.fit_generator(generator=train_generator, validation_data = valid_generator, epochs=2, verbose=2)\n<\/code><\/pre>\n\n<p>I don't want to shuffle my input data. I thought that was getting handled, but in my code, when I print out <code>index<\/code> in <code>__get_item__<\/code>, I get random numbers. I would like consecutive numbers. Notice I'm trying to kill the process using <code>sys.exit<\/code> inside <code>__getitem__<\/code> to see what's going on.<\/p>\n\n<p>My questions:<\/p>\n\n<ol>\n<li><p>Why does <code>index<\/code> not go consecutively? How can I fix this?<\/p><\/li>\n<li><p>When I run this in the terminal using screen, why doesn't it respond to Ctrl+C?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53470907","CreationDate":"2018-11-25T19:05:56.060","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You can use <code>shuffle<\/code> argument of <code>fit_generator<\/code> method to generate batches consecutively. From <code>fit_generator()<\/code> <a href=\"https:\/\/keras.io\/models\/sequential\/#fit_generator\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>shuffle:<\/strong> Boolean. Whether to shuffle the order of the batches at the beginning of each epoch. Only used with instances of <code>Sequence<\/code> (<code>keras.utils.Sequence<\/code>). Has no effect when <code>steps_per_epoch<\/code> is not <code>None<\/code>.<\/p>\n<\/blockquote>\n\n<p>Just pass <code>shuffle=False<\/code> to <code>fit_generator<\/code>:<\/p>\n\n<pre><code>model.fit_generator(generator=train_generator, shuffle=False, ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53440338,"AnswerCount":1,"Tags":"<object><tensorflow><detection>","CreationDate":"2018-11-23T03:30:49.403","AcceptedAnswerId":null,"OwnerUserId":9501391.0,"Title":"How to detect person only instead of whole labeled objects using tensorflow object-detection API?","Body":"<p>I want to detect person only in a given picture (with person, cat, bicycle etc on it) using Tensorflow object detection API (and a pre-trained model \"ssd_mobilenet_v1_coco_2017_11_17\/frozen_inference_graph.pb\"). How should I modify the following code? Maybe I should modify this line <code>detection_graph.get_tensor_by_name('detection_classes:0')<\/code>, but I have no idea what exactly I should do. Help me my friends, please! Thank you in advance. Or some references will also be great.<\/p>\n\n<pre><code>def detect_objects(image_np, sess, detection_graph):\n# Expand dimensions since the model expects images to have shape: [1, None, None, 3]\nimage_np_expanded = np.expand_dims(image_np, axis=0)\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n\n# Each box represents a part of the image where a particular object was detected.\nboxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n\n# Each score represent how level of confidence for each of the objects.\n# Score is shown on the result image, together with the class label.\nscores = detection_graph.get_tensor_by_name('detection_scores:0')\nclasses = detection_graph.get_tensor_by_name('detection_classes:0')\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\n# Actual detection.\n(boxes, scores, classes, num_detections) = sess.run(\n    [boxes, scores, classes, num_detections],\n    feed_dict={image_tensor: image_np_expanded})\n\n# Visualization of the results of a detection.\nvis_util.visualize_boxes_and_labels_on_image_array(\n    image_np,\n    np.squeeze(boxes),\n    np.squeeze(classes).astype(np.int32),\n    np.squeeze(scores),\n    category_index,\n    use_normalized_coordinates=True,\n    line_thickness=8)\nreturn image_np\n<\/code><\/pre>\n","answers":[{"AnswerId":"53447273","CreationDate":"2018-11-23T13:05:57.003","ParentId":null,"OwnerUserId":"8188358","Title":null,"Body":"<p>If I understand it correctly, you have to know the class label for person, then you can select only  for that class in the part where you visualize the result of detection. Assuming the <code>classes<\/code> and <code>boxes<\/code>  can be sliced. <\/p>\n"}]}
+{"QuestionId":53440551,"AnswerCount":2,"Tags":"<python><tensorflow><keras><softmax>","CreationDate":"2018-11-23T04:09:09.093","AcceptedAnswerId":53447308.0,"OwnerUserId":785494.0,"Title":"Producing a softmax on two channels in Tensorflow and Keras","Body":"<p>My network's penultimate layer has shape <code>(U, C)<\/code> where <code>C<\/code> is the number of channels. I'd like to apply softmax function across each channel separately.<\/p>\n\n<p>For example, if <code>U=2<\/code> and <code>C=3<\/code>, and the layer produces <code>[ [1 2 3], [10 20 30] ]<\/code>, I'd like the output to do <code>softmax(1, 2, 3)<\/code> for channel 0 and <code>softmax(10, 20, 30)<\/code> for the channel 1.<\/p>\n\n<p>Is there a way I can do this with Keras? I'm using TensorFlow as the backend.<\/p>\n\n<h2>UPDATE<\/h2>\n\n<p>Please also explain how to ensure that the loss is the sum of both cross entropies, and how I can verify that? (That is, I don't want the optimizer to only train for loss on one of the softmax, but rather the sum of each's cross entropy loss).  The model uses Keras's built in <code>categorical_crossentropy<\/code> for loss.<\/p>\n","answers":[{"AnswerId":"53441879","CreationDate":"2018-11-23T06:50:48.410","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>use functional api for multiple outputs. <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/functional-api-guide\/<\/a><\/p>\n\n<pre><code>input = Input(...)\n...\nt = some_tensor\nt0 = t0[:,:,0]\nt1 = t0[:,:,1]\nsoft0 = Softmax(output_shape)(t0)\nsoft1 = Softmax(output_shape)(t1)\noutputs = [soft0,soft1]\nmodel = Model(inputs=input, outputs=outputs)\nmodel.compile(...)\nmodel.fit(x_train, [y_train0, ytrain1], epoch = 10, batch_size=32)\n<\/code><\/pre>\n"},{"AnswerId":"53447308","CreationDate":"2018-11-23T13:07:35.183","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Define a <code>Lambda<\/code> layer and use the <code>softmax<\/code> function from the backend with a desired axis to compute the softmax over that axis:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.layers import Lambda\n\nsoft_out = Lambda(lambda x: K.softmax(x, axis=my_desired_axis))(input_tensor)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Update:<\/strong> A numpy array with N dimension would have a shape of <code>(d1, d2, d3, ..., dn)<\/code>. Each one of them is called an axis. So the first axis (i.e. <code>axis=0<\/code>) has dimension <code>d1<\/code>, the second axis (i.e. <code>axis=1<\/code>) has dimension <code>d2<\/code> and so on. Further, the most common case of an array is a 2D array or a matrix which has a shape of <code>(m, n)<\/code>, i.e. <code>m<\/code> rows (i.e. <code>axis=0<\/code>) and <code>n<\/code> columns (i.e. <code>axis=1<\/code>). Now when we specify an axis for performing an operation, it means that the operation should be computed over that axis. Let me make this more clear by examples:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; a = np.arange(12).reshape(3,4)\n&gt;&gt;&gt; a\narray([[ 0,  1,  2,  3],\n       [ 4,  5,  6,  7],\n       [ 8,  9, 10, 11]])\n\n&gt;&gt;&gt; a.shape\n(3, 4)   # three rows and four columns\n\n&gt;&gt;&gt; np.sum(a, axis=0)  # compute the sum over the rows (i.e. for each column)\narray([12, 15, 18, 21])\n\n&gt;&gt;&gt; np.sum(a, axis=1)  # compute the sum over the columns (i.e. for each row)\narray([ 6, 22, 38])\n\n&gt;&gt;&gt; np.sum(a, axis=-1) # axis=-1 is equivalent to the last axis (i.e. columns)\narray([ 6, 22, 38])\n<\/code><\/pre>\n\n<p>Now, in your example, the same thing holds for computing softmax function. You must first determine over which axis you want to compute the softmax and then specify that using <code>axis<\/code> argument. Further, note that softmax by default is applied on the last axis (i.e. <code>axis=-1<\/code>) so if you want to compute it over the last axis you don't need the Lambda layer above. Just use the <a href=\"https:\/\/keras.io\/activations\/\" rel=\"nofollow noreferrer\"><code>Activation<\/code><\/a> layer instead:<\/p>\n\n<pre><code>from keras.layers import Activation\n\nsoft_out = Activation('softmax')(input_tensor)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Update 2:<\/strong> There is also another way of doing this using <code>Softmax<\/code> layer:<\/p>\n\n<pre><code>from keras.layers import Softmax\n\nsoft_out = Softmax(axis=desired_axis)(input_tensor)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53441343,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><regression>","CreationDate":"2018-11-23T06:01:13.060","AcceptedAnswerId":53447657.0,"OwnerUserId":10693988.0,"Title":"How to normalize Keras network output in a regression problem that demands output with unit L2 norm?","Body":"<p>My regression problem requires that the network output <code>y<\/code> has unit norm <code>||y|| = 1.<\/code>. I would like to impose that as a <code>Lambda<\/code> layer after the linear activation:<\/p>\n\n<pre><code>from keras import backend as K  \n...  \nmodel.add(Dense(numOutputs, activation='linear'))  \nmodel.add(Lambda(lambda x: K.l2_normalize(x)))  \n<\/code><\/pre>\n\n<p>The backend is TensorFlow. The code compiles but the network predicts output vectors with distinct norms (the norm is not 1 and varies).<\/p>\n\n<p>Any hints regarding what I am doing wrongly?<\/p>\n","answers":[{"AnswerId":"53447657","CreationDate":"2018-11-23T13:31:38.103","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you haven't passed the <code>axis<\/code> argument to the <code>K.l2_normalize<\/code> function. As a result it would normalize all the elements in the whole batch so that their norm would be equal to one. To resolve this, just pass <code>axis=-1<\/code> to normalize over the last axis:<\/p>\n\n<pre><code>model.add(Lambda(lambda x: K.l2_normalize(x, axis=-1)))  \n<\/code><\/pre>\n"}]}
+{"QuestionId":53441938,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2018-11-23T06:55:33.523","AcceptedAnswerId":null,"OwnerUserId":4760789.0,"Title":"Trying to install tensorflow 1.12.0 in an Anaconda 3 virtual python environment","Body":"<p>Trying to install tensorflow on a virtual environment (<code>Anaconda 3<\/code> and <code>Python 3.6<\/code>) on a Windows 10 computer.<\/p>\n\n<p>After installing Anaconda, I go to the Anaconda prompt and key in the Following:<\/p>\n\n<pre><code>conda create -n tensorflow_cpu pip python=3.6\nactivate tensorflow_cpu\npip install --ignore-installed --upgrade tensorflow\n<\/code><\/pre>\n\n<p>After that, I enter these in <code>python<\/code> shell:<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>It raises error:<\/p>\n\n<blockquote>\n  <p>ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.\n  Failed to load the native TensorFlow runtime. <\/p>\n<\/blockquote>\n\n<p>Any assistance would be appreciated <\/p>\n","answers":[]}
+{"QuestionId":53442069,"AnswerCount":1,"Tags":"<numpy><matrix><vectorization><matrix-multiplication><pytorch>","CreationDate":"2018-11-23T07:08:14.870","AcceptedAnswerId":53442187.0,"OwnerUserId":4160286.0,"Title":"How to vectorize the following python code","Body":"<p>I'm trying to obtain a matrix, where each element is calculated as follows:<\/p>\n\n<pre><code>X = torch.ones(batch_size, dim)\nX_ = torch.ones(batch_size, dim)\nY = torch.ones(batch_size, dim)\nM = torch.zeros(batch_size, batch_size)\nfor i in range(batch_size):\n    for j in range(batch_size):\n        M[i, j] = ((X[i] - X_[i] * Y[j])**2).sum()\n<\/code><\/pre>\n\n<p>It's very slow to calculate <code>M<\/code> element-wise, is there any suggestion about how to use matrix multiplication to replace the for loops?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53442187","CreationDate":"2018-11-23T07:18:02.563","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you want to <code>sum()<\/code> over dim, you can \"lift\" your 2D problem to 3D and sum there:<\/p>\n\n<pre><code>M = ((X[:, None, :] - X_[:, None, :] * Y[None, ...])**2).sum(dim=2)\n<\/code><\/pre>\n\n<p><strong>How it works<\/strong>:<\/p>\n\n<p><code>X[:, None, :]<\/code> and <code>X_[:, None, :]<\/code> are 3D of size <code>(batch_size, 1, dim)<\/code>, and <code>Y[None, ...]<\/code> is of size <code>(1, batch_size, dim)<\/code>.  <\/p>\n\n<p>When multiplying <code>X_[:, None, :] * Y[None, ...]<\/code> pytorch <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html#module-numpy.doc.broadcasting\" rel=\"nofollow noreferrer\">broadcasts<\/a> the dimensions of size 1 to the appropriate dimension to get a result of size <code>(batch_size, batch_size, dim)<\/code>.<br>\nFinally, you <code>sum()<\/code> only over the last dimension <code>(dim=2)<\/code> to get an output <code>M<\/code> of size <code>(batch_size, batch_size)<\/code>.<\/p>\n\n<p>The trick here is done by taking advantage of <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html#module-numpy.doc.broadcasting\" rel=\"nofollow noreferrer\">broadcasting<\/a>.<\/p>\n"}]}
+{"QuestionId":53442190,"AnswerCount":1,"Tags":"<python-3.x><keras><deep-learning><conv-neural-network><keras-layer>","CreationDate":"2018-11-23T07:18:27.493","AcceptedAnswerId":53442550.0,"OwnerUserId":5430317.0,"Title":"ImportError: cannot import name 'normalize_data_format'","Body":"<p>I have read an article <a href=\"https:\/\/blog.sicara.com\/keras-generative-adversarial-networks-image-deblurring-45e3ab6977b5\" rel=\"nofollow noreferrer\">Here<\/a> and its pretty nice enough to understand. Given its implementation on <a href=\"https:\/\/github.com\/RaphaelMeudec\/deblur-gan\" rel=\"nofollow noreferrer\">GitHub<\/a>. When I am trying to train at my own using given code it gives me an Import Error in <a href=\"https:\/\/github.com\/RaphaelMeudec\/deblur-gan\/blob\/master\/deblurgan\/layer_utils.py\" rel=\"nofollow noreferrer\">this file at line 117 <\/a>  like following. I am using google Colab environment. Having some search over the error i got that the following line is compatible to keras version==2.2.2. I have also installed that yet not solved with the error. Please help me to get over it. By default keras version installed in colab is 2.2.4<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-47-f8ce7e15cf87&gt; in &lt;module&gt;()\n      9 from keras.layers.merge import Add\n     10 from keras.utils import conv_utils\n---&gt; 11 from keras.utils.conv_utils import normalize_data_format\n     12 \n     13 from keras.layers.core import Dropout\n\nImportError: cannot import name 'normalize_data_format'\n\n---------------------------------------------------------------------------\n<\/code><\/pre>\n","answers":[{"AnswerId":"53442550","CreationDate":"2018-11-23T07:47:07.637","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/utils\/conv_utils.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/utils\/conv_utils.py<\/a><\/p>\n\n<p>master branch's conv_utils doesn't have normalize_data_format.\nsome of the other branches do have it such as tf-keras branch.\nIt is a trivial function here is its implementation: <\/p>\n\n<pre><code>import keras.backend as K\ndef normalize_data_format(value):\n    if value is None:\n        value = K.image_data_format()\n    data_format = value.lower()\n    if data_format not in {'channels_first', 'channels_last'}:\n        raise ValueError('The `data_format` argument must be one of '\n                         '\"channels_first\", \"channels_last\". Received: ' +\n                         str(value))\n    return data_format\n<\/code><\/pre>\n"}]}
+{"QuestionId":53442281,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-23T07:25:28.040","AcceptedAnswerId":null,"OwnerUserId":9201770.0,"Title":"Assign requires shapes of both tensors to match. lhs shape= [1,1,256,256] rhs shape= [1,1,1280,256] [[Node: save\/Assign_348","Body":"<p>I am trying to train a SSD Lite + MobileNetv2 using the model_main.py in the tensorflow\/models\/research\/objectdetection but I'm getting the following error<\/p>\n\n<pre><code>Assign requires shapes of both tensors to match. lhs shape= [1,1,256,256] rhs shape= [1,1,1280,256] [[Node: save\/Assign_348 = Assign[T=DT_FLOAT, _class=[\"loc:@FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights\"], use_locking=true, validate_shape=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights, save\/RestoreV2:348)]]\n<\/code><\/pre>\n\n<p>The full log is given here, <\/p>\n\n<pre><code>python E:\\Documents\\Projects\\tensorflow\\models\\research\\object_detection\\model_main.py --alsologtostderr --pipeline_config_path=experiments\/training_\/ssdlite_mobilenet_v2_coco.config --model_dir=experiments\/training_\/ --num_train_steps=50000 --NUM_EVAL_STEPS=2000\nWARNING:tensorflow:Forced number of epochs for all eval validations to be 1.\nW1123 09:11:18.686478 7432 tf_logging.py:125] Forced number of epochs for all eval validations to be 1.\nWARNING:tensorflow:Expected number of evaluation epochs is 1, but instead encountered eval_on_train_input_config.num_epochs = 0. Overwriting num_epochs to 1.\nW1123 09:11:18.687448 7432 tf_logging.py:125] Expected number of evaluation epochs is 1, but instead encountered eval_on_train_input_config.num_epochs = 0. Overwriting num_epochs to 1.\nWARNING:tensorflow:Estimator's model_fn (&lt;function create_model_fn..model_fn at 0x000001E641D31268&gt;) includes params argument, but params are not passed to Estimator.\nW1123 09:11:18.688472 7432 tf_logging.py:125] Estimator's model_fn (&lt;function create_model_fn..model_fn at 0x000001E641D31268&gt;) includes params argument, but params are not passed to Estimator.\n2018-11-23 09:11:23.084879: I T:\\src\\github\\tensorflow\\tensorflow\\core\\platform\\cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n2018-11-23 09:11:23.365711: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1392] Found device 0 with properties:\nname: GeForce GTX 1060 6GB major: 6 minor: 1 memoryClockRate(GHz): 1.835\npciBusID: 0000:01:00.0\ntotalMemory: 6.00GiB freeMemory: 4.97GiB\n2018-11-23 09:11:23.372771: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1471] Adding visible gpu devices: 0\n2018-11-23 09:11:24.001841: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:952] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-23 09:11:24.004967: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:958] 0\n2018-11-23 09:11:24.007058: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:971] 0: N\n2018-11-23 09:11:24.009175: I T:\\src\\github\\tensorflow\\tensorflow\\core\\common_runtime\\gpu\\gpu_device.cc:1084] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4741 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:01:00.0, compute capability: 6.1)\nTraceback (most recent call last):\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1322, in _do_call\nreturn fn(*args)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1307, in _run_fn\noptions, feed_dict, fetch_list, target_list, run_metadata)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1409, in _call_tf_sessionrun\nrun_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [1,1,256,256] rhs shape= [1,1,1280,256]\n[[Node: save\/Assign_348 = Assign[T=DT_FLOAT, _class=[\"loc:@FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights\"], use_locking=true, validate_shape=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights, save\/RestoreV2:348)]]\n[[Node: save\/RestoreV2\/_599 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_728_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\nFile \"E:\\Documents\\Projects\\tensorflow\\models\\research\\object_detection\\model_main.py\", line 109, in \ntf.app.run()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n_sys.exit(main(argv))\nFile \"E:\\Documents\\Projects\\tensorflow\\models\\research\\object_detection\\model_main.py\", line 105, in main\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 447, in train_and_evaluate\nreturn executor.run()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 531, in run\nreturn self.run_local()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 681, in run_local\neval_result, export_results = evaluator.evaluate_and_export()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 886, in evaluate_and_export\nhooks=self._eval_spec.hooks)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\estimator.py\", line 460, in evaluate\noutput_dir=self.eval_dir(name))\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\estimator.py\", line 1386, in _evaluate_run\nconfig=self._session_config)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\evaluation.py\", line 209, in _evaluate_once\nsession_creator=session_creator, hooks=hooks) as session:\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 826, in init\nstop_grace_period_secs=stop_grace_period_secs)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 549, in init\nself._sess = _RecoverableSession(self._coordinated_creator)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 1012, in init\n_WrappedSession.init(self, self._create_session())\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 1017, in _create_session\nreturn self._sess_creator.create_session()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 706, in create_session\nself.tf_sess = self._session_creator.create_session()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 477, in create_session\ninit_fn=self._scaffold.init_fn)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\session_manager.py\", line 281, in prepare_session\nconfig=config)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\session_manager.py\", line 195, in _restore_checkpoint\nsaver.restore(sess, checkpoint_filename_with_path)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1752, in restore\n{self.saver_def.filename_tensor_name: save_path})\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 900, in run\nrun_metadata_ptr)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1135, in _run\nfeed_dict_tensor, options, run_metadata)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1316, in _do_run\nrun_metadata)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1335, in _do_call\nraise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [1,1,256,256] rhs shape= [1,1,1280,256]\n[[Node: save\/Assign_348 = Assign[T=DT_FLOAT, _class=[\"loc:@FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights\"], use_locking=true, validate_shape=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights, save\/RestoreV2:348)]]\n[[Node: save\/RestoreV2\/_599 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_728_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nCaused by op 'save\/Assign_348', defined at:\nFile \"E:\\Documents\\Projects\\tensorflow\\models\\research\\object_detection\\model_main.py\", line 109, in \ntf.app.run()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n_sys.exit(main(argv))\nFile \"E:\\Documents\\Projects\\tensorflow\\models\\research\\object_detection\\model_main.py\", line 105, in main\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 447, in train_and_evaluate\nreturn executor.run()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 531, in run\nreturn self.run_local()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 681, in run_local\neval_result, export_results = evaluator.evaluate_and_export()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\training.py\", line 886, in evaluate_and_export\nhooks=self._eval_spec.hooks)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\estimator.py\", line 460, in evaluate\noutput_dir=self.eval_dir(name))\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\estimator\\estimator.py\", line 1386, in _evaluate_run\nconfig=self._session_config)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\evaluation.py\", line 209, in _evaluate_once\nsession_creator=session_creator, hooks=hooks) as session:\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 826, in init\nstop_grace_period_secs=stop_grace_period_secs)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 549, in init\nself._sess = _RecoverableSession(self._coordinated_creator)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 1012, in init\n_WrappedSession.init(self, self._create_session())\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 1017, in _create_session\nreturn self._sess_creator.create_session()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 706, in create_session\nself.tf_sess = self._session_creator.create_session()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 468, in create_session\nself._scaffold.finalize()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\monitored_session.py\", line 212, in finalize\nself._saver = training_saver._get_saver_or_default() # pylint: disable=protected-access\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 856, in _get_saver_or_default\nsaver = Saver(sharded=True, allow_empty=True)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1284, in init\nself.build()\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1296, in build\nself._build(self._filename, build_save=True, build_restore=True)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1333, in _build\nbuild_save=build_save, build_restore=build_restore)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 775, in _build_internal\nrestore_sequentially, reshape)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 453, in _AddShardedRestoreOps\nname=\"restore_shard\"))\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 422, in _AddRestoreOps\nassign_ops.append(saveable.restore(saveable_tensors, shapes))\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 113, in restore\nself.op.get_shape().is_fully_defined())\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\ops\\state_ops.py\", line 219, in assign\nvalidate_shape=validate_shape)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\ops\\gen_state_ops.py\", line 63, in assign\nuse_locking=use_locking, name=name)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\nop_def=op_def)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3414, in create_op\nop_def=op_def)\nFile \"D:\\ProgramData\\Anaconda3\\envs\\tfod\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1740, in init\nself._traceback = self._graph._extract_stack() # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [1,1,256,256] rhs shape= [1,1,1280,256]\n[[Node: save\/Assign_348 = Assign[T=DT_FLOAT, _class=[\"loc:@FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights\"], use_locking=true, validate_shape=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](FeatureExtractor\/MobilenetV2\/layer_19_1_Conv2d_2_1x1_256\/weights, save\/RestoreV2:348)]]\n[[Node: save\/RestoreV2\/_599 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_728_save\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n<\/code><\/pre>\n\n<p>My config can be found <a href=\"https:\/\/gist.github.com\/zubairahmed-ai\/489ff298a221ab5e44b2855996cacb5e\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Also note that I originally <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5806\" rel=\"nofollow noreferrer\">posted it on Tensorflow\/models\/issues<\/a> but the team suggested that I post here on SO<\/p>\n","answers":[{"AnswerId":"53457472","CreationDate":"2018-11-24T11:04:21.073","ParentId":null,"OwnerUserId":"9201770","Title":null,"Body":"<p>The error disappeared after I upgraded to TF 1.12 and cuDNN 7.3 for my CUDA v9<\/p>\n"},{"AnswerId":"55920562","CreationDate":"2019-04-30T12:03:06.137","ParentId":null,"OwnerUserId":"11432394","Title":null,"Body":"<p>Your error is because of the dimensions in your image, you should make them equal as the ones you used for training<\/p>\n"}]}
+{"QuestionId":53442343,"AnswerCount":3,"Tags":"<python><tensorflow><tegra><tensorrt><nvidia-jetson>","CreationDate":"2018-11-23T07:30:16.653","AcceptedAnswerId":53461245.0,"OwnerUserId":2249026.0,"Title":"How to do inference using TesnorFlow-GPU models on Tegra X2?","Body":"<p>I am new to Jetson tegra x2 board. <\/p>\n\n<p>I have a plan to run my tensorflow-gpu models on TX2 board and see how they perform there. These models are trained and tested on GTX GPU machine. <\/p>\n\n<p>On tx2 board, Jetpack full does not have tensorflow in it. So tensorflow needs to be built\/installed which I have seen several tutorials on and tried. My python files train.py and test.py expect tensorflow-gpu. <\/p>\n\n<p>Now I suspect, if tensorflow-gpu buiding on tx2 board is the right way to go? <\/p>\n\n<p>Oh, there is Nvidia TensorRT on TX2, that will do part of the job, but how? and is that right? <\/p>\n\n<p>Will tensorflow and tensorRT work together to replace tensorflow-gpu? but how? then what modifications will i have to make in my train and test python files?<\/p>\n\n<p>Do I really need to build tensorflow for tx2 at all? I only need inference I don't want to do training there.<\/p>\n\n<p>I have studied different blogs and tried a several options but now things are bit messed up. <\/p>\n\n<p>My simple question is: <\/p>\n\n<p><strong>What are steps to get inference done on Jetson TX2 board by using TensorFlow-GPU deep learning models trained on GTX machine?<\/strong><\/p>\n","answers":[{"AnswerId":"53882578","CreationDate":"2018-12-21T09:53:24.197","ParentId":null,"OwnerUserId":"4034566","Title":null,"Body":"<p>You can find the tensorflow-gpu wheel files of TX2 for both python 2.7 and python 3.5 in this link of Nvidia's Developer Forum. <\/p>\n\n<p><a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/1031300\/jetson-tx2\/tensorflow-1-8-wheel-with-jetpack-3-2-\/\" rel=\"nofollow noreferrer\">https:\/\/devtalk.nvidia.com\/default\/topic\/1031300\/jetson-tx2\/tensorflow-1-8-wheel-with-jetpack-3-2-\/<\/a><\/p>\n"},{"AnswerId":"53558557","CreationDate":"2018-11-30T13:31:31.827","ParentId":null,"OwnerUserId":"7556865","Title":null,"Body":"<p>I built tensorflow on JetsonTX2 following this guide. It provides instructions and wheels for both Python 2 and Python3. <\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/jetsonhacks\/installTensorFlowTX2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jetsonhacks\/installTensorFlowTX2<\/a><\/li>\n<\/ul>\n\n<p>If you are new to Jetson TX2, also take a look at this \"<strong><em>Guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson<\/em><\/strong>\". (*This does not require tensorflow installation since Jetpack already builds TensorRT)<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/dusty-nv\/jetson-inference#building-from-source-on-jetson\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dusty-nv\/jetson-inference#building-from-source-on-jetson<\/a><\/li>\n<\/ul>\n\n<p>If you have tensorflow trained graphs that you want to run inference on Jetson then you need to first install tensorflow. Afterwards, it is recommended (not compulsory for inference) that you optimize your trained models with tensorRT.Check out these repos for object detection\/classification examples that uses TensorRT optimization.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/NVIDIA-AI-IOT\/tf_trt_models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA-AI-IOT\/tf_trt_models<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/NVIDIA-AI-IOT\/tf_to_trt_image_classification\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA-AI-IOT\/tf_to_trt_image_classification<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"53461245","CreationDate":"2018-11-24T18:36:44.957","ParentId":null,"OwnerUserId":"2624719","Title":null,"Body":"<p>The easiest way is to install the NVIDIA provided wheel: <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/dgx\/install-tf-jetsontx2\/index.html\" rel=\"nofollow noreferrer\">https:\/\/docs.nvidia.com\/deeplearning\/dgx\/install-tf-jetsontx2\/index.html<\/a><\/p>\n\n<p>All the dependencies are already installed by JetPack.<\/p>\n\n<p>After you install Tensorflow using the wheel, you can use it however you use Tensorflow on other platforms. For running inference, you can download a Tensorflow model into TX2 memory, and run your Tensorflow inference scripts on them.<\/p>\n\n<p>You can also optimize your Tensorflow models by passing them through TF-TRT: <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/dgx\/integrate-tf-trt\/index.html\" rel=\"nofollow noreferrer\">https:\/\/docs.nvidia.com\/deeplearning\/dgx\/integrate-tf-trt\/index.html<\/a>\nThere is just one API call that does the optimization: create_inference_graph(...)\nThis will optimize the Tensorflow graph (by mostly fusing nodes), and also let you build the models for lower precision to get better speedup.<\/p>\n"}]}
+{"QuestionId":53442510,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-23T07:43:21.507","AcceptedAnswerId":null,"OwnerUserId":7955314.0,"Title":"How to handle Multi Label DataSet from Directory for image captioning in PyTorch","Body":"<p>I need  a  help in PyTorch,\nRegarding Dataloader, and dataset\nCan someone aid\/guide me <\/p>\n\n<p>Here is my query :\nI am trying for Image Captioning using <a href=\"https:\/\/github.com\/yunjey\/pytorch-tutorial\/tree\/master\/tutorials\/03-advanced\/image_captioning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yunjey\/pytorch-tutorial\/tree\/master\/tutorials\/03-advanced\/image_captioning<\/a>.<\/p>\n\n<p>Here they have used Standard COCO Dataset. <\/p>\n\n<p>I have dataset as images\/ and captions\/ directory .<\/p>\n\n<p>Example <\/p>\n\n<p><strong>Directory Structure:<\/strong> <\/p>\n\n<pre><code>images\/T001.jpg \nimages\/T002.jpg \n...\n...\ncaptions\/T001.txt\ncaptions\/T002.txt\n....\n....\n<\/code><\/pre>\n\n<p>The above is the relation. Caption file has 'n' number of captions in each separate line. <\/p>\n\n<p>I am able to create a custom Dataset class, in that the complete caption file content is being returned. But I want only one line alone gas to be returned. <\/p>\n\n<p>Any guidance\/suggestion on how to achieving this.       <\/p>\n\n<p>++++++++++++++++++++++++++++++++++++++++++++++++\nHere is the class that i have designed: <\/p>\n\n<pre><code>from __future__ import print_function\nimport torch\nfrom torchvision import datasets, models, transforms\nfrom torchvision import transforms\nfrom torch.autograd import Variable\nfrom torch.nn.utils.rnn import pack_padded_sequence\nimport torch.optim as optim\nimport torch.nn as nn\n#from torch import np\nimport numpy as np\nimport utils_c\nfrom data_loader_c import get_cust_data_loader \nfrom models import CNN, RNN\nfrom vocab_custom import Vocabulary, load_vocab\nimport os\n\nclass ImageCaptionDataSet(data.Dataset):\n    def __init__(self, path, json, vocab=None, transform=None):\n        self.vocab = vocab \n        self.transform = transform\n        self.img_dir_path = path  \n        self.cap_dir_path = json \n        self.all_imgs_path = glob.glob(os.path.join(self.img_dir_path,'*.jpg'))\n        self.all_caps_path = glob.glob(os.path.join(self.cap_dir_path,'*.txt'))\n        pass\n\n    def __getitem__(self,index):\n        vocab = self.vocab\n\n        img_path = self.all_imgs_path[index]\n        img_base_name = os.path.basename(img_path) \n        cap_base_name = img_base_name.replace(\".jpg\",\".txt\")\n        cap_path  = os.path.join(self.cap_dir_path,cap_base_name)\n\n        caption_all_for_a_image = open(cap_path).read().split(\"\\n\")\n\n        image = Image.open(img_path)\n        image = image.convert('RGB')\n\n        if self.transform != None:\n            # apply image preprocessing\n            image = self.transform(image)\n\n        #captions_combined = []\n        #max_len = 0  \n        #for caption in caption_all_for_a_image:\n        #    caption_str = str(caption).lower()\n        #    tokens = nltk.tokenize.word_tokenize(caption_str)\n        #    m = len(tokens) + 2 \n        #    if m&gt;max_len:\n        #        max_len = m \n        #    caption = torch.Tensor([vocab(vocab.start_token())] +\n        #                           [vocab(token) for token in tokens] +\n        #                           [vocab(vocab.end_token())])\n        #    captions_combined.append(caption) \n        #    #yield image, caption\n        #return image,torch.Tensor(captions_combined)\n\n        caption_str = str(caption_all_for_a_image).lower()\n        tokens = nltk.tokenize.word_tokenize(caption_str)\n        caption = torch.Tensor([vocab(vocab.start_token())] +\n                                   [vocab(token) for token in tokens] +\n                                   [vocab(vocab.end_token())])\n\n        return image,caption\n\n    def __len__(self):\n        return len(self.all_imgs_path)\n<\/code><\/pre>\n\n<p>+++++++++++++++++++++++++++++++++<\/p>\n","answers":[{"AnswerId":"53444263","CreationDate":"2018-11-23T09:51:29.463","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>First, using <code>str()<\/code> to convert the <em>list<\/em> of captions into a single string (<code>caption_str = str(caption_all_for_a_image)<\/code>) is a <strong>bad<\/strong> idea: <\/p>\n\n<pre><code>cap = ['a sentence', 'bla bla bla']\nstr(cap)\n<\/code><\/pre>\n\n<p>Returns this <strong>sting<\/strong>:<\/p>\n\n<blockquote>\n<pre><code>\"['a sentence', 'bla bla bla']\"\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Note that <code>['<\/code>, and <code>', '<\/code> are part of the resulting string!<\/p>\n\n<p>You can pick one of the captions at random:<\/p>\n\n<pre><code>import random\n...\ncap_idx = random.randi(0, len(caption_all_for_a_image)-1)  # pick one at random\ncaption_str = caption_all_for_a_image[cap_idx].lower()  # actual selection\n<\/code><\/pre>\n"}]}
+{"QuestionId":53442801,"AnswerCount":1,"Tags":"<python><image><opencv><tensorflow><image-processing>","CreationDate":"2018-11-23T08:09:49.547","AcceptedAnswerId":53447477.0,"OwnerUserId":6533075.0,"Title":"Averaging pixels in image ROI","Body":"<p>I have the following image and I would like to apply averaging on each masked region (<strong>not the bounding box<\/strong>) in the image. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TgXHp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TgXHp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As you can see, right now, the regions have changing values inside them on the heat map. Some pixels are yellowish, some are purplish. I want this to be not the case inside the masks. <\/p>\n\n<p>So what I need to do is (I guess):<\/p>\n\n<ul>\n<li>to find which coordinates correspond to the masks<\/li>\n<li>average the pixels within these coordinates<\/li>\n<\/ul>\n\n<p>Here is how the masks are found:<\/p>\n\n<pre><code>file_names = glob(os.path.join(IMAGE_DIR, \"*.jpg\"))\nmasks_prediction = np.zeros((521, 768, len(file_names)))\nfor i in range(len(file_names)):\n    print(i)\n    image = skimage.io.imread(file_names[i])\n    predictions = model.detect([image],  verbose=1)\n    p = predictions[0]\n    masks = p['masks']\n    merged_mask = np.zeros((masks.shape[0], masks.shape[1]))\n    for j in range(masks.shape[2]):\n        merged_mask[masks[:,:,j]==True] = True\n        masks_prediction[:,:,i] = merged_mask\n<\/code><\/pre>\n\n<p>Here is the function which applies the masks:<\/p>\n\n<pre><code>def apply_mask(image, mask, color, alpha=0.5):\n    \"\"\"Apply the given mask to the image.\n    \"\"\"\n    for c in range(3):\n        image[:, :, c] = np.where(mask == 1,\n                                  image[:, :, c] *\n                                  (1 - alpha) + alpha * color[c] * 255,\n                                  image[:, :, c])\n    return image\n<\/code><\/pre>\n\n<p>and in the main file, here is how it is used:<\/p>\n\n<pre><code>mask = masks[:, :, i]\nif show_mask:\n     masked_image = apply_mask(masked_image, mask, color)\n<\/code><\/pre>\n\n<p>So I need to make a modification in somewhere here but I don't know where exactly.<\/p>\n","answers":[{"AnswerId":"53447477","CreationDate":"2018-11-23T13:19:10.753","ParentId":null,"OwnerUserId":"8930347","Title":null,"Body":"<p>I think the mask you are looking for is provided by the code :<\/p>\n\n<pre><code>mask = masks[:, :, i]\n<\/code><\/pre>\n\n<p>where <em>i<\/em> refers to the number of mask you have.<\/p>\n\n<p>You can obtain the average values of the mask region from the original image using the openCV function <code>mean<\/code><\/p>\n\n<p>Here is how your code should look like:<\/p>\n\n<pre><code>mask = masks[:, :, i]\navg_masked_value = cv2.mean(original_image,mask)\n<\/code><\/pre>\n\n<p>where <em>original_image<\/em> is the original image you loaded and avg_masked_value will contain 3x1 array of averaged value.<\/p>\n"}]}
+{"QuestionId":53442887,"AnswerCount":2,"Tags":"<python><tensorflow><keras><reshape>","CreationDate":"2018-11-23T08:17:05.770","AcceptedAnswerId":null,"OwnerUserId":10663471.0,"Title":"How do I add the None Dimension back to a Tensor?","Body":"<p>I've made some transformation in a Lambda Layer and now I have shape <code>(1,)<\/code>, how do I get back to <code>(None, 1)<\/code>?<\/p>\n\n<p>Here are my operations<\/p>\n\n<pre><code>def function_lambda(x):\n    import keras.backend\n\n    aux_array = keras.backend.sign(x)\n    #the shape is (?, 11) here OK\n\n    aux_array = keras.backend.relu(aux_array)\n    #the shape is (?, 11) here still OK\n\n    aux_array = keras.backend.any(aux_array)\n    #the shape is () here not OK\n\n    aux_array = keras.backend.reshape(aux_array, [-1])\n    #now the shape is (1,) almost OK\n\n    return aux_array\n<\/code><\/pre>\n\n<p>How do I reshape and put back the <code>None<\/code> Batch Dimension?<\/p>\n","answers":[{"AnswerId":"53446457","CreationDate":"2018-11-23T12:08:10.107","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>If you have a fully-defined shape (like <code>(1,)<\/code>) then you don't need <code>None<\/code> dimensions, because you know exactly how many elements your tensor has. But you can reshape to have two dimensions replacing you reshaping with:<\/p>\n\n<pre><code>aux_array = keras.backend.reshape(aux_array, [-1, 1])\n<\/code><\/pre>\n\n<p>This will give you an array with shape <code>(1, 1)<\/code>, which is compatible with the shape <code>(None, 1)<\/code>, so it should be okay. Note that <code>-1<\/code> in <code>reshape<\/code> means \"however many elements are necessary to accommodate the tensor in this shape\", but in this case you know you only have one element, so it would be the same to use:<\/p>\n\n<pre><code>aux_array = keras.backend.reshape(aux_array, [1, 1])\n<\/code><\/pre>\n\n<p>Because, again, the shape is fully defined and you know what the size of each dimension should be exactly. However, using <code>-1<\/code> is convenient because it works whether you know the shape fully or not, and Keras\/TensorFlow will work out what the size should be (the dimension will have the necessary size if it can be computed or <code>None<\/code> if part of the shape is unknown).<\/p>\n"},{"AnswerId":"55437510","CreationDate":"2019-03-31T03:01:46.493","ParentId":null,"OwnerUserId":"11036442","Title":null,"Body":"<p>The tf.reshape has a limitation that it cannot work with \"None\" dimension. So far as I know, the only place where we can define a \"None\" dimension is in tf.placeholder. The following should work:<\/p>\n\n<pre><code>def function_lambda(x):\n    import keras.backend\n\n    aux_array = keras.backend.sign(x)\n    #the shape is (?, 11) here OK\n\n    aux_array = keras.backend.relu(aux_array)\n    #the shape is (?, 11) here still OK\n\n    aux_array = keras.backend.any(aux_array)\n    #the shape is () here not OK\n\n    aux_array = keras.backend.placeholder_with_default(aux_array, [None,1])\n    #now the shape is (?,1) OK\n\n    return aux_array\n<\/code><\/pre>\n\n<p>Additional note: creating a new placeholder for reintroducing the \"None\" dimension is useful when trying to use Google Machine Learning Engine. MLE requires that the first dimension (batch) always remain \"None\" or unknown.<\/p>\n"}]}
+{"QuestionId":53443415,"AnswerCount":1,"Tags":"<python><bash><tensorflow><protocol-buffers><tensorflow-serving>","CreationDate":"2018-11-23T08:58:22.220","AcceptedAnswerId":53447716.0,"OwnerUserId":10324651.0,"Title":"Running tensorflow_model_server from bash is throwing 'command not found error'","Body":"<p>I am running tensorflow server from bash:\nCommand used:<\/p>\n\n<pre><code>tensorflow_model_server --rest_api_port=8501 --model_name=saved_model --model_base_path=$(&lt;path to .pb file in my system&gt;\/saved_model.pb)\n<\/code><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<pre><code>&lt;path to .pb file&gt;\/saved_model.pb:\n line 1: $'\\b\\001\\022\\344\\307\\376\\016': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 2: $'\\203\\020\\022\\345\\017': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 4: $'\\003Add\\022\\006': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 6: $'\\001x\\001T\\022\\006\\n\\001y\\001T\\032\\006': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 7: $'\\001z\\001T\\E': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 8: $'\\001T\\022\\004type:\\020': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 9: $'\\0162\\f\\016\\023\\001\\002\\004\\006\\005\\003': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 10: 8: command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 11: $'\\004Cast\\022': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 14: $'\\001x\\004SrcT\\032\\t\\n\\001y\\004DstT\\f\\n\\004SrcT\\022\\004type\\f': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 15: $'\\004DstT\\022\\004type': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 16: 8: command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 17: $'\\005Const\\032\\017': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 18: $'\\006output\\005dtype\\017': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 24: $'\\005value\\022\\006tensor\\r\\n\\005dtype\\022\\004type\\n\\354\\001\\n\\006Conv2D\\022\\n\\n\\005input\\001T\\022\\v': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 28: $'\\006filter\\001T\\032\\v\\n\\006output\\001T\\023\\n\\001T\\022\\004type:\\b\\n\\0062\\004\\023\\016\\001\\002\\024': command not found &lt;path to .pb file&gt;\/saved_model.pb:\n line 29: syntax error near unexpected token (' &lt;path to .pb file&gt;\/saved_model.pb:\n line 29: strides\ufffd  list(int)\"\ufffd'\n<\/code><\/pre>\n\n<p>I am not sure how to fix the issue.\nI am using anaconda in my system.\nAny lead would be really helpful.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"53447716","CreationDate":"2018-11-23T13:35:14.017","ParentId":null,"OwnerUserId":"7975433","Title":null,"Body":"<p><code>model_base_path<\/code> should be a directory, the parent folder of the version folder of <code>saved_model.pb<\/code> and <code>variables\\<\/code>, not a file. For example, for the following directory, <code>model_base_path<\/code> should be <code>path_to_versions\\versions<\/code>.<\/p>\n\n<pre><code>`-- versions\n    `-- 1\n        |-- saved_model.pbtxt\n        `-- variables\n            |-- variables.data-00000-of-00001\n            `-- variables.index\n<\/code><\/pre>\n"}]}
+{"QuestionId":53443475,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><nlp>","CreationDate":"2018-11-23T09:02:07.893","AcceptedAnswerId":53446723.0,"OwnerUserId":4566277.0,"Title":"Keras ValueError: Error when checking target: expected dense_15 to have 3 dimensions, but got array with shape (301390, 8)","Body":"<p>I have 8 classes that I want to predict from input text. Here is my code for preprocessing the data: <\/p>\n\n<pre><code>num_max = 1000\ntok = Tokenizer(num_words=num_max)\ntok.fit_on_texts(x_train)\nmat_texts = tok.texts_to_matrix(x_train,mode='count')\nnum_max = 1000\ntok = Tokenizer(num_words=num_max)\ntok.fit_on_texts(x_train)\nmax_len = 100\ncnn_texts_seq = tok.texts_to_sequences(x_train)\nprint(cnn_texts_seq[0])\n\n[12, 4, 303]\n\n# padding the sequences\ncnn_texts_mat = sequence.pad_sequences(cnn_texts_seq,maxlen=max_len)\nprint(cnn_texts_mat[0])\nprint(cnn_texts_mat.shape)\n\n[  0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0\n   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0\n   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0\n   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0\n   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0\n   0   0   0   0   0   0   0  12   4 303]\n\n(301390, 100)\n<\/code><\/pre>\n\n<p>Below is the structure of my model which contains an embedding layer:<\/p>\n\n<pre><code>max_features = 20000\nmax_features = cnn_texts_mat.shape[1]\nmaxlen = 100\nembedding_size = 128\n\n\nmodel = Sequential()\nmodel.add(Embedding(max_features, embedding_size, input_length=maxlen))\nmodel.add(Dropout(0.2))\n\nmodel.add(Dense(5000, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(600, activation='relu'))\nmodel.add(Dropout(0.1))\nmodel.add(Dense(units=y_train.shape[1], activation='softmax'))\n\nsgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss='binary_crossentropy',\n              optimizer=sgd)\n<\/code><\/pre>\n\n<p>Below is the model summary:<\/p>\n\n<pre><code>model.summary()\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nembedding_5 (Embedding)      (None, 100, 128)          12800     \n_________________________________________________________________\ndropout_13 (Dropout)         (None, 100, 128)          0         \n_________________________________________________________________\ndense_13 (Dense)             (None, 100, 5000)         645000    \n_________________________________________________________________\ndropout_14 (Dropout)         (None, 100, 5000)         0         \n_________________________________________________________________\ndense_14 (Dense)             (None, 100, 600)          3000600   \n_________________________________________________________________\ndropout_15 (Dropout)         (None, 100, 600)          0         \n_________________________________________________________________\ndense_15 (Dense)             (None, 100, 8)            4808      \n=================================================================\nTotal params: 3,663,208\nTrainable params: 3,663,208\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>After this, I am getting below error when I try to run the model:<\/p>\n\n<pre><code>model.fit(x=cnn_texts_mat, y=y_train, epochs=2, batch_size=100)\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-41-4b9da9914e7e&gt; in &lt;module&gt;\n----&gt; 1 model.fit(x=cnn_texts_mat, y=y_train, epochs=2, batch_size=100)\n\n~\/.local\/lib\/python3.5\/site-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n    950             sample_weight=sample_weight,\n    951             class_weight=class_weight,\n--&gt; 952             batch_size=batch_size)\n    953         # Prepare validation data.\n    954         do_validation = False\n\n~\/.local\/lib\/python3.5\/site-packages\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    787                 feed_output_shapes,\n    788                 check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 789                 exception_prefix='target')\n    790 \n    791             # Generate sample-wise weight values given the `sample_weight` and\n\n~\/.local\/lib\/python3.5\/site-packages\/keras\/engine\/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    126                         ': expected ' + names[i] + ' to have ' +\n    127                         str(len(shape)) + ' dimensions, but got array '\n--&gt; 128                         'with shape ' + str(data_shape))\n    129                 if not check_batch_axis:\n    130                     data_shape = data_shape[1:]\n\nValueError: Error when checking target: expected dense_15 to have 3 dimensions, but got array with shape (301390, 8)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53446723","CreationDate":"2018-11-23T12:26:55.790","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Look at the output shape of the last layer in model summary: it is <code>(None, 100, 8)<\/code>. This is not what you are looking for. Your labels for each sample have a shape of <code>(8,)<\/code> and not <code>(100,8)<\/code>. Why this happened? That's because <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">the Dense layer is applied on the last axis of its input<\/a> and therefore since the output of Embedding layer is 3D, the output of all the following Dense layers would also have 3 dimension.<\/p>\n\n<p>How to resolve this? One approach is to use <code>Flatten<\/code> layer somewhere in your model (possibly right after the embedding layer). This way you would have a 2D ouput of shape <code>(None, 8)<\/code> which is what you want and is consistent with the shape of labels.<\/p>\n\n<p>However, note that you may end up with a very big model (i.e. too many parameters) which would be highly prone to overfitting. Either reduce the number of units in the Dense layers or alternatively use <code>Conv1D<\/code> and <code>MaxPooling1D<\/code> layers or even RNN layers to process the embeddings and reduce the dimensionality of resulting tensors (which is also probable that using them would also increase the accuracy of the model as well).<\/p>\n"}]}
+{"QuestionId":53443553,"AnswerCount":1,"Tags":"<c++><tensorflow><modelica><openmodelica>","CreationDate":"2018-11-23T09:07:35.730","AcceptedAnswerId":53444704.0,"OwnerUserId":10622607.0,"Title":"Check failed: GetOpGradFactory()->insert({op, func}).second Duplicated gradient for MapAccumulate","Body":"<p><em>Note:\nI also posted this question here <a href=\"https:\/\/openmodelica.org\/forum\/default-topic\/2570-linux-mint,-python-3-6-omc-server-is-down-please-start-it\" rel=\"nofollow noreferrer\">https:\/\/openmodelica.org\/forum\/default-topic\/2570-linux-mint,-python-3-6-omc-server-is-down-please-start-it<\/a> and I hope to reach a larger group with tensorflow experience as well.<\/em><\/p>\n\n<hr>\n\n<p>I want to use the <strong>Tensorflow C++ API<\/strong> in a <strong>Modelica Model<\/strong>.<\/p>\n\n<p>For this, I created a class in C++ that encapsulates the Tensorflow details.\nMoreover, the same file contains 3 functions (to call the constructor, destructor and a function that uses an existing instance of my class. These functions are declared \"extern\" in the respective header file (only if read by a C++ compiler, I am using __cplusplus to differentiate).<\/p>\n\n<p>I then compiled said file (the class and the 3 functions) into a .so using g++ (I did make use of -fPIC and -shared). In \/usr\/lib, I created a symlink to said shared object library.<\/p>\n\n<p>(For completeness, I should mention that I had trouble with the \"LibraryDirectory\" and \"IncludeDirectory\" annotations. This is why I created symlinks to the contents of all the directories I would pass to g++\/gcc using -L and -I (for my test program) right into Model\/Resources\/Library and Model\/Resources\/Include directories. As there were no more compilation errors, I thought that this was okay.)<\/p>\n\n<p>Now the strange thing is:<\/p>\n\n<p>I can use the 3 functions in a test program compiled with gcc. I linked it against all Tensorflow libraries (and libm, libstdc++ and my .so from above) and it just works as expected.\nHowever, when I try to use these 3 functions in a) a Modelica Class (constructor and destructor), and b) a function using an \"external\" function call, it will compile fine but a runtime error occurs:<\/p>\n\n<hr>\n\n<p>stdout | OMEditInfo | <p>\/tmp\/OpenModelica_fred\/OMEdit\/TestCpp -port=44382 -logFormat=xmltcp -override=startTime=0,stopTime=1,stepSize=0.002,tolerance=1e-6,solver=dassl,outputFormat=mat,variableFilter=.* -r=TestCpp_res.mat -w -lv=LOG_STATS<\/p>\nstdout | error | <p>2018-11-22 10:47:00.153977: F tensorflow\/core\/framework\/function.cc:1440] Check failed: GetOpGradFactory()-&gt;insert({op, func}).second Duplicated gradient for MapAccumulate<br><\/p>\n\n<p>stdout | error | <p>Process crashed<\/p>\nstdout | error | <p>Process crashed<br>\nSimulation process failed. Exited with code 6.<\/p><\/p>\n\n<hr>\n\n<p>How is this possible? I am not able to do anything in my test program with the 3 functions that leads to the same error. What could the OMC-generated executable be doing with them that I am not in my test program? <\/p>\n","answers":[{"AnswerId":"53444704","CreationDate":"2018-11-23T10:14:23.857","ParentId":null,"OwnerUserId":"10622607","Title":null,"Body":"<p>I found a solution pointing the symlink \"clang\" in \/usr\/bin to \"gcc\".<\/p>\n\n<p>For instructions, check <a href=\"https:\/\/openmodelica.org\/forum\/default-topic\/2570-linux-mint,-python-3-6-omc-server-is-down-please-start-it\" rel=\"nofollow noreferrer\">https:\/\/openmodelica.org\/forum\/default-topic\/2570-linux-mint,-python-3-6-omc-server-is-down-please-start-it<\/a> .<\/p>\n"}]}
+{"QuestionId":53443596,"AnswerCount":1,"Tags":"<caffe><multilabel-classification><multiclass-classification>","CreationDate":"2018-11-23T09:10:24.457","AcceptedAnswerId":53482586.0,"OwnerUserId":4749728.0,"Title":"Caffe: Accuracy stuck on 0","Body":"<p>I'm trying to train AlexNet with multi class lmdb data.\nThe input layers are:<\/p>\n\n<pre><code>layer { \n name: \"data\" \n type: \"Data\"\n top:  \"data\" \n data_param { \n   source: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/TrainImage\" \n   backend: LMDB \n   batch_size: 64 \n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/TrainImage\/TrainImage.binaryproto\"\n    mirror: true \n  } \n  include: { phase: TRAIN } \n} \nlayer {\n  name: \"labels\"\n  type: \"Data\"\n  top: \"labels\"\n  include {\n    phase: TRAIN\n  }\n  data_param {\n    source: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/TrainLabel\" \n    batch_size: 64\n    backend: LMDB\n  }\n}\nlayer { \n name: \"data\" \n type: \"Data\"\n top:  \"data\" \n data_param { \n   source: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/ValImage\"\n   backend: LMDB \n   batch_size: 64\n } \n transform_param { \n    crop_size: 227 \n    mean_file: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/ValImage\/ValImage.binaryproto\"\n    mirror: false \n  } \n  include: { phase: TEST } \n} \nlayer {\n  name: \"labels\"\n  type: \"Data\"\n  top: \"labels\"\n  include {\n    phase: TEST\n  }\n  data_param {\n    source: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/ValLabel\" \n    batch_size: 64\n    backend: LMDB\n  }\n}\nlayer {\n  name: \"slice\"\n  type: \"Slice\"\n  bottom: \"labels\"\n  top: \"class\"\n  top: \"subclass\"\n  top: \"color\"\n  slice_param {\n    axis: 1\n    slice_point: 1\n    slice_point: 2\n  }\n}\n<\/code><\/pre>\n\n<p>And the outputs layers are:<\/p>\n\n<pre><code>layer {\n  name: \"fc8_1_class\"\n  type: \"InnerProduct\"\n  bottom: \"fc7\"\n  top: \"fc8_1_class\"\n  param {\n    lr_mult: 5\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 10\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 2\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayer {\n  name: \"fc8_2_subclass\"\n  type: \"InnerProduct\"\n  bottom: \"fc7\"\n  top: \"fc8_2_subclass\"\n  param {\n    lr_mult: 5\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 10\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 15\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayer {\n  name: \"fc8_3_color\"\n  type: \"InnerProduct\"\n  bottom: \"fc7\"\n  top: \"fc8_3_color\"\n  param {\n    lr_mult: 5\n    decay_mult: 1\n  }\n  param {\n    lr_mult: 10\n    decay_mult: 0\n  }\n  inner_product_param {\n    num_output: 8\n    weight_filler {\n      type: \"gaussian\"\n      std: 0.01\n    }\n    bias_filler {\n      type: \"constant\"\n      value: 0\n    }\n  }\n}\nlayer {\n  name: \"accuracy_class\"\n  type: \"Accuracy\"\n  bottom: \"fc8_1_class\"\n  bottom: \"class\"\n  top: \"accuracy_class\"\n  include: { phase: TEST }\n}\nlayer {\n  name: \"loss_class\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"fc8_1_class\"\n  bottom: \"class\"\n  top: \"loss_class\"\n  loss_weight: 0.5\n}\nlayer {\n  name: \"accuracy_subclass\"\n  type: \"Accuracy\"\n  bottom: \"fc8_2_subclass\"\n  bottom: \"subclass\"\n  top: \"accuracy_subclass\"\n  include: { phase: TEST }\n}\nlayer {\n  name: \"loss_subclass\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"fc8_2_subclass\"\n  bottom: \"subclass\"\n  top: \"loss_subclass\"\n  loss_weight: 0.5\n}\nlayer {\n  name: \"accuracy_color\"\n  type: \"Accuracy\"\n  bottom: \"fc8_3_color\"\n  bottom: \"color\"\n  top: \"accuracy_color\"\n  include: { phase: TEST }\n}\nlayer {\n  name: \"loss_color\"\n  type: \"SoftmaxWithLoss\"\n  bottom: \"fc8_3_color\"\n  bottom: \"color\"\n  top: \"loss_color\"\n  loss_weight: 0.5\n}\n<\/code><\/pre>\n\n<p>\"Class\" is a dominant class with 2 labels which gives good accuracy, on the other hand \"SubClass\" has 15 labels and gives poor results, \"Color\" keeps giving me accuracy=0 no matter what I tryed.\nThis is part of train output:<\/p>\n\n<pre><code>I1123 10:47:52.258751  3392 solver.cpp:273] Learning Rate Policy: step\nI1123 10:47:52.422647  3392 solver.cpp:218] Iteration 0 (7.96684e-08 iter\/s, 0.163777s\/100 iters), loss = 3.50787\nI1123 10:47:52.422716  3392 solver.cpp:237]     Train net output #0: loss_class = 1.83086 (* 0.5 = 0.915429 loss)\nI1123 10:47:52.422742  3392 solver.cpp:237]     Train net output #1: loss_color = 2.39675 (* 0.5 = 1.19837 loss)\nI1123 10:47:52.422768  3392 solver.cpp:237]     Train net output #2: loss_subclass = 2.78814 (* 0.5 = 1.39407 loss)\nI1123 10:47:52.422803  3392 sgd_solver.cpp:105] Iteration 0, lr = 0.0001\nI1123 10:48:01.282282  3392 solver.cpp:330] Iteration 100, Testing net (#0)\nI1123 10:48:01.554111  3392 blocking_queue.cpp:49] Waiting for data\nI1123 10:48:03.717224  3406 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:03.884392  3405 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:04.193727  3392 solver.cpp:397]     Test net output #0: accuracy_class = 0.958594\nI1123 10:48:04.193784  3392 solver.cpp:397]     Test net output #1: accuracy_color = 0\nI1123 10:48:04.193802  3392 solver.cpp:397]     Test net output #2: accuracy_subclass = 0.495313\nI1123 10:48:04.193841  3392 solver.cpp:397]     Test net output #3: loss_class = 0.143002 (* 0.5 = 0.0715012 loss)\nI1123 10:48:04.193871  3392 solver.cpp:397]     Test net output #4: loss_color = 1.53319 (* 0.5 = 0.766597 loss)\nI1123 10:48:04.193894  3392 solver.cpp:397]     Test net output #5: loss_subclass = 1.40099 (* 0.5 = 0.700494 loss)\nI1123 10:48:04.290772  3392 solver.cpp:218] Iteration 100 (8.42592 iter\/s, 11.8681s\/100 iters), loss = 1.50249\nI1123 10:48:04.302697  3392 solver.cpp:237]     Train net output #0: loss_class = 0.145322 (* 0.5 = 0.0726612 loss)\nI1123 10:48:04.302740  3392 solver.cpp:237]     Train net output #1: loss_color = 1.44309 (* 0.5 = 0.721547 loss)\nI1123 10:48:04.302767  3392 solver.cpp:237]     Train net output #2: loss_subclass = 1.41657 (* 0.5 = 0.708286 loss)\nI1123 10:48:04.302788  3392 sgd_solver.cpp:105] Iteration 100, lr = 0.0001\nI1123 10:48:08.030251  3404 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:08.077759  3403 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:13.195049  3392 solver.cpp:330] Iteration 200, Testing net (#0)\nI1123 10:48:15.287199  3406 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:15.466104  3405 data_layer.cpp:73] Restarting data prefetching from start.\nI1123 10:48:16.056556  3392 solver.cpp:397]     Test net output #0: accuracy_class = 0.958594\nI1123 10:48:16.056615  3392 solver.cpp:397]     Test net output #1: accuracy_color = 0\nI1123 10:48:16.056633  3392 solver.cpp:397]     Test net output #2: accuracy_subclass = 0.520703\nI1123 10:48:16.056658  3392 solver.cpp:397]     Test net output #3: loss_class = 0.11685 (* 0.5 = 0.0584251 loss)\nI1123 10:48:16.056679  3392 solver.cpp:397]     Test net output #4: loss_color = 1.53322 (* 0.5 = 0.766609 loss)\nI1123 10:48:16.056707  3392 solver.cpp:397]     Test net output #5: loss_subclass = 1.42879 (* 0.5 = 0.714395 loss)\nI1123 10:48:16.151207  3392 solver.cpp:218] Iteration 200 (8.43993 iter\/s, 11.8484s\/100 iters), loss = 1.74936\nI1123 10:48:16.163187  3392 solver.cpp:237]     Train net output #0: loss_class = 0.253774 (* 0.5 = 0.126887 loss)\nI1123 10:48:16.163224  3392 solver.cpp:237]     Train net output #1: loss_color = 1.69651 (* 0.5 = 0.848255 loss)\nI1123 10:48:16.163246  3392 solver.cpp:237]     Train net output #2: loss_subclass = 1.54843 (* 0.5 = 0.774216 loss)\nI1123 10:48:16.163265  3392 sgd_solver.cpp:105] Iteration 200, lr = 0.0001\n<\/code><\/pre>\n\n<p>The data contains around 9000 images for train and another 2000 for validation, labels are in this form (starts from 0):<\/p>\n\n<pre><code>\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/15036.png 0 4 18 \n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/31658.png 1 7 17\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/26971.png 0 4 17\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/19301.png 0 4 20\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/35906.png 1 7 17\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/24690.png 1 11 23\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/27442.png 0 9 24\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/26689.png 1 11 23\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/38403.png 0 4 19\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/37141.png 0 4 19\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/16961.png 0 4 21\n\/media\/gal\/USB\/MAFAT\/crops\/train\/img\/23952.png 1 16 24\n<\/code><\/pre>\n\n<p>This is the solver file:<\/p>\n\n<pre><code>net: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/AlexNetMulti.prototxt\"  \ntest_initialization: false\ntest_iter: 40\ntest_interval: 100\nbase_lr: 0.0001\nmomentum: 0.9\ngamma: 0.1  \nlr_policy: \"step\"    \nstepsize: 1600   # drop the learning rate every .... iterations\nmax_iter: 4800\ndisplay: 100\nweight_decay: 0.0005\nsnapshot: 300000\nsnapshot_prefix: \"\/home\/gal\/caffe-1.0\/models\/mixDCNN\/hdf5\/snapshot\/GoogleNet_class\"\nsolver_mode: GPU\n<\/code><\/pre>\n\n<p>Why am I keep getting accuracy 0 on \"color\" class?<\/p>\n","answers":[{"AnswerId":"53482586","CreationDate":"2018-11-26T13:52:09.600","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You set <code>num_output: 8<\/code> for your <code>\"fc8_3_color\"<\/code> layer, that is, you are trying to predict \"color\" with 8 discrete classes 0..7. However, looking at your data the labels for color ranges from 17 to 24 -- this cannot work. the labels MUST be from 0..7 if you are using <code>\"SoftmaxWithLoss\"<\/code> and <code>\"Accuracy\"<\/code> layers.<\/p>\n"}]}
+{"QuestionId":53443996,"AnswerCount":1,"Tags":"<python><machine-learning><keras><classification><lstm>","CreationDate":"2018-11-23T09:35:37.727","AcceptedAnswerId":null,"OwnerUserId":5572627.0,"Title":"'NoneType' object has no attribute '_inbound_nodes' ? Keras seq2seq classification","Body":"<p>I have a problem with creating a Keras model. I found a simple encoder decoder and try to fix it like below:<\/p>\n\n<pre><code>   # some encoder code ... .... above is not shown here, where it is too obvious\n    encoder_outputs, state_h, state_c = encoder(encoder_inputs)\n\n    encoder_states = [state_h, state_c]\n\n    decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)\n    decoder_outputs, _, _ = decoder_lstm(encoder_outputs[-1:], initial_state=encoder_states)\n    decoder_dense = Dense(num_decoder_tokens, activation='softmax')\n    decoder_outputs = decoder_dense(decoder_outputs)\n\n    model = Model(inputs=[encoder_inputs], outputs=decoder_outputs)\n<\/code><\/pre>\n\n<p>which will use the last encoder output as decoder input and only goes for a single output. <\/p>\n\n<p>I wonder why it creates such a problem at:<\/p>\n\n<pre><code>model = Model(inputs=[encoder_inputs], outputs=decoder_outputs)\n<\/code><\/pre>\n\n<p>and creating message:<\/p>\n\n<blockquote>\n  <p>'NoneType' object has no attribute '_inbound_nodes'<\/p>\n<\/blockquote>\n\n<p>How do I solve it? I tried to find similar questions' answers but I didn't get a good one to fix mine.<\/p>\n","answers":[{"AnswerId":"53446480","CreationDate":"2018-11-23T12:09:38.900","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>First of all, <code>encoder_outputs[-1:]<\/code> would give you the last batch, not the last output of each batch which is <code>encoder_outputs[:,-1:]<\/code>.<\/p>\n\n<p>Second, since you need to pass Keras Tensors to layers in Keras, you need to use a <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\"><code>Lambda<\/code><\/a> layer to do the slicing:<\/p>\n\n<pre><code>last_input = Lambda(lambda x: x[:,-1:])(encoder_inputs)\ndecoder_outputs, _, _ = decoder_lstm(last_input,\n                                     initial_state=encoder_states)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53444241,"AnswerCount":1,"Tags":"<python><image-processing><machine-learning><keras><conv-neural-network>","CreationDate":"2018-11-23T09:50:34.703","AcceptedAnswerId":53446019.0,"OwnerUserId":10573543.0,"Title":"How to convert a image dimensions which are directed through model.flow_from_directory?","Body":"<p>I am trying to build an image classifier using <code>Keras 2.2.0<\/code> and <code>tensorflow 1.9.0<\/code><\/p>\n\n<p>I am getting an error of this sort:<\/p>\n\n<pre><code>str(data_shape))\nValueError: Error when checking input: expected conv2d_1_input to have shape (1, 224, 224) but got array with shape (224, 224, 3)\n<\/code><\/pre>\n\n<p>Here is the code:<\/p>\n\n<pre><code>train_datagen=ImageDataGenerator(rotation_range=40,\n            width_shift_range=0.2,\n            height_shift_range=0.2,\n            rescale=1.\/255,\n            shear_range=0.2,\n            zoom_range=0.2,\n            horizontal_flip=True,\n            fill_mode='nearest')\n\nvalidation_datagen=ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory('\/media\/centura\/DANISH\/mobile backup\/moles\/train\/',class_mode='binary',target_size=(224, 224),batch_size=batch_size)\n\nvalidation_generator = validation_datagen.flow_from_directory('\/media\/centura\/DANISH\/mobile backup\/moles\/test\/',class_mode='binary',target_size=(224, 224),batch_size=batch_size)\n\n#Data Dimensions\nimg_rows,img_cols=224,224\n\ninput_shape1=(1,img_rows,img_cols)\n\n#initialising the model\n\nmodel=Sequential()\n\n#layer 1\nmodel.add(Conv2D(filters=32, kernel_size=(3,3), strides=(1, 1), padding='same',input_shape=input_shape1,data_format=\"channels_last\"))\nmodel.add(BatchNormalization())\nmodel.add(Activation('relu'))\n#model.add(AveragePooling2D(pool_size=(2,2)))\nmodel.add(Dropout(0.25))\n\n\n#fully connected first layer\n\nmodel.add(Flatten())\nmodel.add(Dense(500,use_bias=False))\nmodel.add(BatchNormalization())\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.25))      \n\n\n#Fully connected final layer\nmodel.add(Dense(1)) \nmodel.add(Activation('sigmoid')) \n\ntensorboard=TensorBoard(log_dir='logs\/{}'.format(name))\n\n\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])\n#model.summary()\nmodel.fit_generator(train_generator,epochs=50,validation_data=validation_generator,callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>I believe the error is coming from the <code>train_generator<\/code>. I searched in stack overflow for similar problems. I found certain solutions but it was not working for me. How can I change the dimensions of the image if it is being called from the <code>.flow_from_directory<\/code>? <\/p>\n","answers":[{"AnswerId":"53446019","CreationDate":"2018-11-23T11:40:02.163","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Let's break down the error step-by-step to find out what it is telling us:<\/p>\n\n<blockquote>\n  <p>Error when checking input:<\/p>\n<\/blockquote>\n\n<p>So it is related to the input data and input layer of the model.<\/p>\n\n<blockquote>\n  <p>expected conv2d_1_input to have shape (1, 224, 224)<\/p>\n<\/blockquote>\n\n<p>If we look at the code for the first convolution layer we see that:<\/p>\n\n<pre><code>Conv2D(..., input_shape=input_shape1, ...)\n<\/code><\/pre>\n\n<p>And the value of <code>input_shape1<\/code> as you have defined it is <code>(1,img_rows,img_cols)<\/code> which is <code>(1, 224, 224)<\/code>. But:<\/p>\n\n<blockquote>\n  <p>but got array with shape (224, 224, 3)<\/p>\n<\/blockquote>\n\n<p>Which means the images generated by the <code>train_generator<\/code> have a shape of <code>(224, 224, 3)<\/code> (which is correct and expected).<\/p>\n\n<p>As a result, we see these two shapes, the shape of generated images and the given shape to <code>input_shape<\/code> argument, must be the same. Therefore, you need to modify the value of <code>input_shape1<\/code> as follows:<\/p>\n\n<pre><code>input_shape1=(img_rows, img_cols, 3)\n<\/code><\/pre>\n\n<p>which is exactly what a convolution layer expects as its input shape (i.e. <code>(image_height, image_width, image_channels)<\/code>).<\/p>\n"}]}
+{"QuestionId":53444265,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><mnist><activation-function>","CreationDate":"2018-11-23T09:51:44.567","AcceptedAnswerId":null,"OwnerUserId":8821999.0,"Title":"using tanh as activation function in MNIST dataset in tensorflow","Body":"<p>I am working on simple MLP neural network for MNIST dataset using tensorflow as my homework. in the question we should implement a multilayer perceptron with tanh as activation function. I should use the data label with [-1,+1].For example for number 3 we have:<\/p>\n\n<pre><code>[-1,-1,-1,+1,-1,-1,-1,-1,-1,-1]\n<\/code><\/pre>\n\n<p>I know that for sigmoid function we can use on_hot such as:<\/p>\n\n<pre><code>mnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n<\/code><\/pre>\n\n<p>in order to putting data in [0,1] like the following for number 3:<\/p>\n\n<pre><code>[0,0,0,1,0,0,0,0,0,0]\n<\/code><\/pre>\n\n<p>how can I encode label between [-1 ,+1].\nthanks in advance for every help<\/p>\n","answers":[{"AnswerId":"53447034","CreationDate":"2018-11-23T12:48:48.607","ParentId":null,"OwnerUserId":"10193890","Title":null,"Body":"<p>Unnecessary downvotes to the question. BTW.. if I understood it correctly, here's the answer.<\/p>\n\n<p>What I understood is, instead of using sigmoid, you have to use tanh and so you want the output data in format of +1s and -1s instead of 0s and 1s.<\/p>\n\n<p>Note that one hot encoding is specifically designed for getting outputs of 1s and 0s. That's why it is called one hot encoding - it outputs 1 for right answer and 0 for others.<\/p>\n\n<p>Now, there is no built-in function to get the output you want. But I prefer a short and simple way by writing my own code. Don't get afraid - that's only 1 line of code.<\/p>\n\n<pre><code>import numpy as np\na = np.array([1, 1, 1, 0, 1, 0])\na[a==0]=-1\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>array([1, 1, 1, -1, 1, -1])\n<\/code><\/pre>\n\n<p>You can use the same.. Take the one hot encoding labels as output using your code and then use this one line of code to get what you want.<\/p>\n\n<pre><code>a[a==0]=-1\n<\/code><\/pre>\n\n<p>Thank you..<\/p>\n"}]}
+{"QuestionId":53444377,"AnswerCount":1,"Tags":"<python><tensorflow><nlp>","CreationDate":"2018-11-23T09:57:48.843","AcceptedAnswerId":53447514.0,"OwnerUserId":10694929.0,"Title":"What is differece between token-level and segment-level in NLP task?","Body":"<p>Actually, I am not so understand about token..\nWhen I read googleresearch\/bert model, I see these words.<\/p>\n\n<pre><code># In the demo, we are doing a simple classification task on the entire   \n# segment.  \n#   \n# If you want to use the token-level output, use model.get_sequence_output()   # instead.\n<\/code><\/pre>\n\n<p>Can anyone make an example about token-level and segment-level classification?<\/p>\n","answers":[{"AnswerId":"53447514","CreationDate":"2018-11-23T13:22:08.307","ParentId":null,"OwnerUserId":"9651974","Title":null,"Body":"<p>Segment-level classification means that each segment will have one label, for example, a classifier which categorises a movie review as good or bad. There is only one output label for the entire input sequence.<\/p>\n\n<p>Token-level classification means that each token will be given a label, for example a part-of-speech tagger will classify each word as one particular part of speech. Each token (element in the sequence) will have a corresponding label in the output.<\/p>\n\n<p>If you're not sure what a token is, you can start by thinking of it as each word in a sentence, but to be more correct, look at <a href=\"https:\/\/nlp.stanford.edu\/IR-book\/html\/htmledition\/tokenization-1.html\" rel=\"nofollow noreferrer\">https:\/\/nlp.stanford.edu\/IR-book\/html\/htmledition\/tokenization-1.html<\/a>. Depending on how you tokenise and preprocess your text, tokens can be words, punctuation symbols, special markers, subword-level symbols, etc.<\/p>\n"}]}
+{"QuestionId":53445345,"AnswerCount":1,"Tags":"<keras><pytorch>","CreationDate":"2018-11-23T10:55:33.473","AcceptedAnswerId":53446975.0,"OwnerUserId":996366.0,"Title":"What is the PyTorch alternative for Keras input_shape, output_shape, get_weights, get_config and summary","Body":"<p>In Keras, after creating a model, we can see its input and output shapes using <code>model.input_shape<\/code>, <code>model.output_shape<\/code>. For weights and config  we can use <code>model.get_weights()<\/code> and <code>model.get_config()<\/code>, respectively.<\/p>\n\n<p>What are the similar alternatives for PyTorch? Also is there any other functions we need to know for inspecting a PyTorch model?<\/p>\n\n<p>To get the summary in PyTorch, we print model <code>print(model)<\/code> but this gives lesser information than <code>model.summary()<\/code>. Is there a better summary for PyTorch? <\/p>\n","answers":[{"AnswerId":"53446975","CreationDate":"2018-11-23T12:45:15.563","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>There is no \"model.summary()\" method in pytorch. You need to use built_in methods and fields of the model. <\/p>\n\n<p>For example, I have customized inception_v3 model. To get the information I need to use other many different fields. For instance: <\/p>\n\n<p>IN:<\/p>\n\n<pre><code>print(model) # print network architecture\n<\/code><\/pre>\n\n<p>OUT    <\/p>\n\n<pre><code>Inception3(\n  (Conv2d_1a_3x3): BasicConv2d(\n    (conv): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), bias=False)\n    (bn): BatchNorm2d(32, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)\n  )\n  (Conv2d_2a_3x3): BasicConv2d(\n    (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), bias=False)\n    (bn): BatchNorm2d(32, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)\n  )\n  (Conv2d_2b_3x3): BasicConv2d(\n    (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n    (bn): BatchNorm2d(64, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)\n  )\n  (Conv2d_3b_1x1): BasicConv2d(\n    (conv): Conv2d(64, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)\n    (bn): BatchNorm2d(80, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)\n  )\n  (Conv2d_4a_3x3): BasicConv2d(\n    (conv): Conv2d(80, 192, kernel_size=(3, 3), stride=(1, 1), bias=False)\n   ...\n<\/code><\/pre>\n\n<p>IN:<\/p>\n\n<pre><code>for i in model.state_dict().keys():\n    print(i) \n#print keys of dict with values of learned weights, bias, parameters\n<\/code><\/pre>\n\n<p>OUT:  <\/p>\n\n<pre><code>    Conv2d_1a_3x3.conv.weight\n    Conv2d_1a_3x3.bn.weight\n    Conv2d_1a_3x3.bn.bias\n    Conv2d_1a_3x3.bn.running_mean\n    Conv2d_1a_3x3.bn.running_var\n    Conv2d_1a_3x3.bn.num_batches_tracked\n    Conv2d_2a_3x3.conv.weight\n    Conv2d_2a_3x3.bn.weight\n    Conv2d_2a_3x3.bn.bias\n    Conv2d_2a_3x3.bn.running_mean \n    ...\n<\/code><\/pre>\n\n<p>So if I would like to get weights for the layer CNN at Conv2d_1a_3x3, I look for the key \"Conv2d_1a_3x3.conv.weight\":<\/p>\n\n<pre><code>print(\"model.save_dict()[\"Conv2d_1a_3x3.conv.weight\"])\n<\/code><\/pre>\n\n<p>OUT:    <\/p>\n\n<pre><code>tensor([[[[-0.2103, -0.3441, -0.0344],\n          [-0.1420, -0.2520, -0.0280],\n          [ 0.0736,  0.0183,  0.0381]],\n\n         [[ 0.1417,  0.1593,  0.0506],\n          [ 0.0828,  0.0854,  0.0186],\n          [ 0.0283,  0.0144,  0.0508]],\n...\n<\/code><\/pre>\n\n<p>If you want to see used hyperparameters from optimizer:<\/p>\n\n<pre><code>optimizer.param_groups\n<\/code><\/pre>\n\n<p>OUT:<\/p>\n\n<pre><code>[{'dampening': 0,\n  'lr': 0.01,\n  'momentum': 0.01,\n  'nesterov': False,\n  'params': [Parameter containing:\n   tensor([[[[-0.2103, -0.3441, -0.0344],\n             [-0.1420, -0.2520, -0.0280],\n             [ 0.0736,  0.0183,  0.0381]],\n          ...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53446532,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-11-23T12:13:42.207","AcceptedAnswerId":53448575.0,"OwnerUserId":6583936.0,"Title":"Update values in tensor along deeper dimensions","Body":"<p>I have tensor A of shape MxNxC where M stands for number of examples, N is number of features and C is 3 euler rotation angles. Also i have a tensor B of similar shape, but instead of angles there are coordinates.\nWhat is needed is to convert both of these tensors to one containing affine transformation matrices so its shape would be like MxNx4x4. I dont know how to iterate over these tensors together, I've looked for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\">tf.map_fn<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scan\" rel=\"nofollow noreferrer\">tf.scan<\/a> but they iterate only with the first dimension. What im looking for is some method to apply function like the one below to all of the elements along the last axes.<\/p>\n\n<pre><code>def f(angles, vector): #dimensions 3 or 3x1\n    ...\n    return matrix # dimension 4x4\n<\/code><\/pre>\n\n<p>Any help would be useful, thanks!<\/p>\n","answers":[{"AnswerId":"53448575","CreationDate":"2018-11-23T14:33:10.233","ParentId":null,"OwnerUserId":"10658208","Title":null,"Body":"<p>you can try something like this:<\/p>\n\n<pre><code>A_flattened = tf.reshape(A, [-1, 3])# flatten it out\nB_flattened = tf.reshape(B, [-1, 3])\nAB_flattened = tf.map_fn(convert_to_mat, (A_flattened, B_flattened))# convert_to_mat should return a 4x4 matrix\nAB = tf.reshape(AB_flattened, [M, N, 4, 4])\n<\/code><\/pre>\n\n<p>this should do the trick!<\/p>\n"}]}
+{"QuestionId":53446875,"AnswerCount":1,"Tags":"<python><opencv><keras><deep-learning><deeplab>","CreationDate":"2018-11-23T12:38:29.703","AcceptedAnswerId":null,"OwnerUserId":10269571.0,"Title":"DeepLab v3 - how to process images with any size ratio. (h * w)","Body":"<p>Deeplab v3 predicts only with default ratio 512 * 512. how to process images with different size ratio.<\/p>\n\n<p><strong>please look over the error:-<\/strong><\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 35, in \n      predicted = model_dlv3.predict(img[np.newaxis, ...])<\/p>\n  \n  <p>File\n  \"\/home\/developer\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\",\n  line 1149, in predict\n      x, _, _ = self._standardize_user_data(x)<\/p>\n  \n  <p>File\n  \"\/home\/developer\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\",\n  line 751, in _standardize_user_data\n      exception_prefix='input')<\/p>\n  \n  <p>File\n  \"\/home\/developer\/anaconda2\/envs\/py36\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\",\n  line 138, in standardize_input_data\n      str(data_shape))<\/p>\n  \n  <p><strong>ValueError: Error when checking input: expected input_3 to have shape (512, 512, 3) but got array with shape (2448, 3264, 3)<\/strong><\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53447362","CreationDate":"2018-11-23T13:10:48.343","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>You can resize the images to 512*512 or turn images into a square using padding and then resize them into 512*512. Latter option is better since it won't change the shape of the objects inside the image. You can use different padding modes and see what gives the best result.<\/p>\n\n<p>You can use opencv or scikit-image for resizing and padding<\/p>\n\n<p>opencv documentation for padding: <a href=\"https:\/\/docs.opencv.org\/2.4\/modules\/imgproc\/doc\/filtering.html?highlight=copymakeborder#copymakeborder\" rel=\"nofollow noreferrer\">https:\/\/docs.opencv.org\/2.4\/modules\/imgproc\/doc\/filtering.html?highlight=copymakeborder#copymakeborder<\/a><\/p>\n\n<p>scikit-image documentation for padding : <a href=\"http:\/\/scikit-image.org\/docs\/dev\/api\/skimage.util.html#skimage.util.pad\" rel=\"nofollow noreferrer\">http:\/\/scikit-image.org\/docs\/dev\/api\/skimage.util.html#skimage.util.pad<\/a><\/p>\n"}]}
+{"QuestionId":53446905,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><loss>","CreationDate":"2018-11-23T12:40:34.727","AcceptedAnswerId":53447141.0,"OwnerUserId":9468014.0,"Title":"KERAS \"sparse_categorical_crossentropy\" question","Body":"<p>As an input a have a float 1.0 or 0.0. When I try to predict with my model and the <code>sparse_categorical_crossentropy<\/code> loss I get something like: \n<code>[[0.4846592 0.5153408]]<\/code>.<\/p>\n\n<p>How do I know what category it predicts?<\/p>\n","answers":[{"AnswerId":"53447141","CreationDate":"2018-11-23T12:55:52.480","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>These numbers you see are the probability of each class for the given input sample. For example, <code>[[0.4846592 0.5153408]]<\/code> means that the given sample belongs to class 0 with probability of around 0.48 and it belongs to class 1 with probability of around 0.51. So you want to take the class with the highest probability and therefore you can use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.15.1\/reference\/generated\/numpy.argmax.html\" rel=\"noreferrer\"><code>np.argmax<\/code><\/a> to find which index (i.e. 0 or 1) is the maximum one:<\/p>\n\n<pre><code>import numpy as np\n\npred_class = np.argmax(probs, axis=-1) \n<\/code><\/pre>\n\n<p>Further, this has nothing to do with the loss function of the model. These probabilities are given by the last layer in your model which is very likely that it uses <code>softmax<\/code> as the activation function to normalize the output as a probability distribution.<\/p>\n"}]}
+{"QuestionId":53447341,"AnswerCount":0,"Tags":"<validation><keras><prediction>","CreationDate":"2018-11-23T13:09:38.220","AcceptedAnswerId":null,"OwnerUserId":258483.0,"Title":"Is it possible to obtain prediction at the end of epoch in Keras?","Body":"<p>At the end of epoch, Keras prints validation loss and probably other metrics, like validation accuracy.<\/p>\n\n<p>Doesn't this mean it was to compute predicions against validation set? Can I obtain results of this predicion?<\/p>\n\n<p>At the end of epoch, I see field <code>validation_data<\/code> which probably contains concatenation of my <code>x<\/code> and <code>y<\/code> and also additional column of ones. What are they?<\/p>\n","answers":[]}
+{"QuestionId":53447345,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2018-11-23T13:09:50.687","AcceptedAnswerId":53447634.0,"OwnerUserId":5353461.0,"Title":"PyTorch set_grad_enabled(False) vs with no_grad():","Body":"<p>Assuming autograd is on (as it is by default), is there any difference (besides  indent) between doing:<\/p>\n\n<pre><code>with torch.no_grad():\n    &lt;code&gt;\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>torch.set_grad_enabled(False)\n&lt;code&gt;\ntorch.set_grad_enabled(True)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53447634","CreationDate":"2018-11-23T13:30:28.847","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Actually no, there no difference in the way used in the question. When you take a look at the source code of <code>no_grad<\/code>. You see that it is actually using <code>torch.set_grad_enabled<\/code> to archive this behaviour:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class no_grad(object):\n    r\"\"\"Context-manager that disabled gradient calculation.\n\n    Disabling gradient calculation is useful for inference, when you are sure\n    that you will not call :meth:`Tensor.backward()`. It will reduce memory\n    consumption for computations that would otherwise have `requires_grad=True`.\n    In this mode, the result of every computation will have\n    `requires_grad=False`, even when the inputs have `requires_grad=True`.\n\n    Also functions as a decorator.\n\n\n    Example::\n\n        &gt;&gt;&gt; x = torch.tensor([1], requires_grad=True)\n        &gt;&gt;&gt; with torch.no_grad():\n        ...   y = x * 2\n        &gt;&gt;&gt; y.requires_grad\n        False\n        &gt;&gt;&gt; @torch.no_grad()\n        ... def doubler(x):\n        ...     return x * 2\n        &gt;&gt;&gt; z = doubler(x)\n        &gt;&gt;&gt; z.requires_grad\n        False\n    \"\"\"\n\n    def __init__(self):\n        self.prev = torch.is_grad_enabled()\n\n    def __enter__(self):\n        torch._C.set_grad_enabled(False)\n\n    def __exit__(self, *args):\n        torch.set_grad_enabled(self.prev)\n        return False\n\n    def __call__(self, func):\n        @functools.wraps(func)\n        def decorate_no_grad(*args, **kwargs):\n            with self:\n                return func(*args, **kwargs)\n        return decorate_no_grad\n<\/code><\/pre>\n\n<p>However there is an additional functionality of <code>torch.set_grad_enabled<\/code>  over <code>torch.no_grad<\/code> when used in a <code>with<\/code>-statement which lets you control to switch on or off gradient computation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    &gt;&gt;&gt; x = torch.tensor([1], requires_grad=True)\n    &gt;&gt;&gt; is_train = False\n    &gt;&gt;&gt; with torch.set_grad_enabled(is_train):\n    ...   y = x * 2\n    &gt;&gt;&gt; y.requires_grad\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/autograd\/grad_mode.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/autograd\/grad_mode.html<\/a><\/p>\n\n<hr>\n\n<p>Edit:<\/p>\n\n<p>@TomHale Regarding your comment. I just made a short test with PyTorch 1.0 and it turned out that the gradient will be active:<\/p>\n\n<pre><code>import torch\nw = torch.rand(5, requires_grad=True)\nprint('Grad Before:', w.grad)\ntorch.set_grad_enabled(False)\nwith torch.enable_grad():\n    scalar = w.sum()\n    scalar.backward()\n    # Gradient tracking will be enabled here.\ntorch.set_grad_enabled(True)\n\nprint('Grad After:', w.grad)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Grad Before: None\nGrad After: tensor([1., 1., 1., 1., 1.])\n<\/code><\/pre>\n\n<p>So gradients will be computed in this setting.<\/p>\n\n<p>The other setting you posted in your answer also yields to the same result:<\/p>\n\n<pre><code>import torch\nw = torch.rand(5, requires_grad=True)\nprint('Grad Before:', w.grad)\nwith torch.no_grad():\n    with torch.enable_grad():\n        # Gradient tracking IS enabled here.\n        scalar = w.sum()\n        scalar.backward()\n\nprint('Grad After:', w.grad)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Grad Before: None\nGrad After: tensor([1., 1., 1., 1., 1.])\n<\/code><\/pre>\n"},{"AnswerId":"55629560","CreationDate":"2019-04-11T09:49:32.720","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>The <a href=\"https:\/\/pytorch.org\/docs\/stable\/autograd.html#torch.autograd.enable_grad\" rel=\"nofollow noreferrer\"><code>torch.autograd.enable_grad<\/code> documentation<\/a> says:<\/p>\n\n<blockquote>\n  <p>Enables gradient calculation inside a <code>no_grad<\/code> context. This has no effect outside of <code>no_grad<\/code>.<\/p>\n<\/blockquote>\n\n<p>Given this wording, the following is expected:<\/p>\n\n<pre><code>torch.set_grad_enabled(False)\nwith torch.enable_grad:\n    # Gradient tracking will NOT be enabled here.\ntorch.set_grad_enabled(True)\n<\/code><\/pre>\n\n<p>vs:<\/p>\n\n<pre><code>with torch.no_grad():\n    with torch.enable_grad:\n        # Gradient tracking IS enabled here.\n<\/code><\/pre>\n\n<p>But as <a href=\"https:\/\/stackoverflow.com\/a\/53447634\/5353461\">blue-phoenix shows<\/a>, this is <strong>not<\/strong> the case.<\/p>\n\n<p>I raised an issue <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/19189\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53447406,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-23T13:13:38.210","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"Save and restore autograd enabled state","Body":"<p>Is there a way of saving and restoring the current state of autograd?<\/p>\n\n<p>eg:<\/p>\n\n<pre><code>saved_state = torch.autograd.get_grad_enabled()  # How do I do this?\ntorch.autograd.set_grad_enabled(temporary_state)\n# do things for great justice\ntorch.autograd.set_grad_enabled(saved_state)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53447899","CreationDate":"2018-11-23T13:47:43.967","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>The <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/autograd\/grad_mode.html\" rel=\"nofollow noreferrer\">Source code for <code>torch.autograd.grad_mode<\/code><\/a> hints at a solution:<\/p>\n\n<pre><code>torch.is_grad_enabled()\n<\/code><\/pre>\n\n<p>It seems that <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/4474#issuecomment-441245282\" rel=\"nofollow noreferrer\">this function is currently undocumented<\/a>.<\/p>\n"}]}
+{"QuestionId":53447411,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><classification>","CreationDate":"2018-11-23T13:14:08.403","AcceptedAnswerId":null,"OwnerUserId":4107990.0,"Title":"Tensorflow Network Test Accuracy Fails to Improve on CIFAR 10","Body":"<p>I'm new to Tensorflow. Right now, I'm trying to create a simple 4 layer fully connected neural network to classify the CIFAR-10 dataset. However, on my testset, the neural network accuracy on the test set is completely static, and is stuck at 11%. <\/p>\n\n<p>I know that a fully connected neural network is probably not ideal fo this task, but it is weird that the network is not improving\/changing at all. So I was wondering if anyone knows the solution to my problem. I have copied my code below, any help is appreciated! Thank you very much.<\/p>\n\n<pre><code>import tensorflow as tf \nimport numpy as np\n\n(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n\ny_train_one_hot = np.zeros((y_train.shape[0], 10))\nfor i in range(y_train.shape[0]):\n    y_train_one_hot[i][y_train[i]] = 1\n\ny_test_one_hot = np.zeros((y_test.shape[0], 10))\nfor i in range(y_test.shape[0]):\n    y_test_one_hot[i][y_test[i]] = 1\n\n\nx = tf.placeholder(dtype=tf.float32, shape=(None, 32, 32, 3), name='X')\ny = tf.placeholder(dtype=tf.float32, shape=(None, 10), name='Y')\nkeep_prob = tf.placeholder(tf.float32)\n\nx_flatten = tf.reshape(x, [-1, 32*32*3])\nnn = tf.layers.dense(x_flatten, 1028, activation=tf.nn.relu)\nnn = tf.nn.dropout(nn, keep_prob)\nnn = tf.layers.dense(nn, 1028, activation=tf.nn.relu)\nnn = tf.nn.dropout(nn, keep_prob)\nnn = tf.layers.dense(nn, 512, activation=tf.nn.relu)\nnn = tf.nn.dropout(nn, keep_prob)\nprediction = tf.layers.dense(nn, 10, activation=tf.nn.relu)\n\ncross_entropy = tf.losses.softmax_cross_entropy(onehot_labels=y, logits=prediction)\nloss = tf.reduce_mean(cross_entropy)\ntrain_step = tf.train.AdamOptimizer().minimize(loss)\n\ncorrect_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n\n    for iteration in range(1000):\n        sess.run(train_step, feed_dict={x: x_train[:1000], y: y_train_one_hot[:1000], keep_prob: 0.5})\n        if iteration % 10 == 0:\n            acc = sess.run(accuracy, feed_dict={x: x_test[:100], y: y_test_one_hot[:100], keep_prob: 1.0})\n            print(acc)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53448008,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><lstm>","CreationDate":"2018-11-23T13:54:28.683","AcceptedAnswerId":53449017.0,"OwnerUserId":5419282.0,"Title":"LSTM multiple features prediction : shape to give to data (Tensorflow)","Body":"<p>After long research, I was unable to find an answer for this question.<\/p>\n\n<p>My data is a series of items, each having a number of features. The data is in this form:\n<a href=\"https:\/\/i.stack.imgur.com\/fwEWt.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fwEWt.jpg\" alt=\"Data schema\"><\/a><\/p>\n\n<p>I want to guess the features of the n+1 element.\nShould the network have a particular number of nodes? What should be the shape of the data fed into the LSTM Tensorflow network and its output?<\/p>\n","answers":[{"AnswerId":"53449017","CreationDate":"2018-11-23T15:06:14.543","ParentId":null,"OwnerUserId":"8394915","Title":null,"Body":"<p>As for the input and output sizes, you have defined them yourself in the picture: you want to predict a vector of length <code>num_features<\/code> (4 in your picture) from an input matrix of shape <code>(N, num_features)<\/code> where <code>N<\/code> is the number of samples to use for one prediction.<\/p>\n\n<p>In tensorflow, the placeholders usually have the shape <code>(batch_size, number_of_steps, number_of_features)<\/code>. For you, they could look like this<\/p>\n\n<pre><code>input_holder = tf.placeholder(shape=(batch_size, None, num_features), dtype=..., name='inputs')\ntarget_holder = tf.placeholder(shape=(batch_size, num_features), dtype=..., name='targets')\n<\/code><\/pre>\n\n<p>in case you don't want to predefine <code>N<\/code> - then you need to have a fixed batch size. Otherwise you could also have<\/p>\n\n<pre><code>input_holder = tf.placeholder(shape=(None, N, num_features), dtype=..., name='inputs')\ntarget_holder = tf.placeholder(shape=(None, num_features), dtype=..., name='targets')\n<\/code><\/pre>\n\n<p>in which <code>N<\/code> would be fixed and your batch size variable. (You cannot have both variable.)<\/p>\n\n<hr>\n\n<p>As for your network size, you need to experiment yourself: it is impossible to make an a priori judgement on the require number of nodes, layers, etc.<\/p>\n"}]}
+{"QuestionId":53448059,"AnswerCount":1,"Tags":"<python><tensorflow><types><keras>","CreationDate":"2018-11-23T13:58:29.797","AcceptedAnswerId":53448332.0,"OwnerUserId":7290815.0,"Title":"Python SET to DICT results in unhashable type: 'set'","Body":"<p>since one week I am building up knowledge in machine learning, using python + tensorflow. My main target is to get the tutorials presented by keras up and running, so my own introduction to python is getting these load-functions to work properly, bc I am stucked behind a firewall and pip ain't downloading. \nNow I've arrived at a problem which makes me loose my mind. \nCurrently I am working at a basic regression, just like <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_text_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_text_classification<\/a> - convert integer back to words<\/p>\n\n<p>Every other tutorial just worked fine after a little modification, but this one is driving me crazy.<\/p>\n\n<pre><code>def get_word_index(path='training\\imdb_word_index.json'):\nwith open('training\\imdb_word_index.json') as f: \n    json1_str = f.read()\n    json1_data = json.loads(json1_str)\n    return dict(json1_data)\n<\/code><\/pre>\n\n<p>this piece of code opens the dataset and returns it as a dictionary. \nthe content looks like this:<\/p>\n\n<pre><code>\"the\":1402\n<\/code><\/pre>\n\n<p>that's it. \nI am trying to reverse it, bc the built-in function doesn't seem to work like intended. \nNow the best practice I found was\n<a href=\"https:\/\/stackoverflow.com\/questions\/13464152\/typeerror-unhashable-type-list-when-using-built-in-set-function\">TypeError: unhashable type: &#39;list&#39; when using built-in set function<\/a>\nwhich didn't help much. \nMy code is<\/p>\n\n<pre><code>reverse_word_ind = {}\nreverse_word_ind = dict.fromkeys(word_index.items(), 0)\n<\/code><\/pre>\n\n<p>which raises the error:<\/p>\n\n<blockquote>\n  <p>TypeError: unhashable type: 'set'<\/p>\n<\/blockquote>\n\n<p>After multiple experiments, turning it manually over and typecasting it to 'dict' my patience failed.\nPlease help...<\/p>\n","answers":[{"AnswerId":"53448332","CreationDate":"2018-11-23T14:17:06.060","ParentId":null,"OwnerUserId":"425458","Title":null,"Body":"<p>You can reverse a dictionary with a <a href=\"https:\/\/docs.python.org\/3\/tutorial\/datastructures.html#dictionaries\" rel=\"nofollow noreferrer\">dict comprehension<\/a>:<\/p>\n\n<pre><code>reverse_word_ind = {v:k for k,v in word_index.items()}\n<\/code><\/pre>\n\n<p>Not every dictionary can be reversed. This is because while there are no restrictions on the types of objects that can be dictionary values, there are <a href=\"https:\/\/wiki.python.org\/moin\/DictionaryKeys\" rel=\"nofollow noreferrer\">restrictions as to what types can be keys<\/a>.<\/p>\n\n<p>The error that you're getting implies that at least one of the values in your dictionary is a set, contrary to your expectations. You can debug what's going on and figure out the bad value by doing the dictionary reversal within a <code>try...except<\/code> clause:<\/p>\n\n<pre><code>reverse_word_ind = {}\nfor k,v in word_index.items():\n    try:\n        reverse_word_ind[v] = k\n    except TypeError as e:\n        print('could not reverse pair (%s, %s) of types (%s, %s)' % (k, v, type(k), type(v)))\n        raise e\n<\/code><\/pre>\n\n<p>Running the above code should result in the same <code>TypeError<\/code> as you were getting before, but now it'll first print out a helpful error message pointing out the bad value in your dictionary.<\/p>\n"}]}
+{"QuestionId":53449907,"AnswerCount":1,"Tags":"<python><tensorflow><visual-studio-code><pip>","CreationDate":"2018-11-23T16:11:49.140","AcceptedAnswerId":null,"OwnerUserId":10696374.0,"Title":"I can't seem to import tensorflow to python VSCODE","Body":"<p>Traceback (<em>most recent call last<\/em>):<\/p>\n\n<pre><code>File \"d:\\kuliah\\aa\\aaa\\aaaa\\bpnn.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n\nFile \"C:\\Users\\xxx\\AppData\\Roaming\\Python\\Python36\\sitepackages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt; from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n\nFile \"C:\\Users\\xxx\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\python\\__init__.py\", line 59, in &lt;module&gt; from tensorflow.core.framework.graph_pb2 import *\n\nFile \"C:\\Users\\xxx\\AppData\\Roaming\\Python\\Python36\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 6, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n\nFile \"C:\\Program Files\\Python36\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in &lt;module&gt;\n    from google.protobuf.pyext import _message\n\nImportError: DLL load failed: The specified procedure could not be found.\n<\/code><\/pre>\n\n<p>This happened when I'm importing my tensorflow on a .py code, what am i missing?<\/p>\n","answers":[{"AnswerId":"57623083","CreationDate":"2019-08-23T09:08:30.527","ParentId":null,"OwnerUserId":"11834494","Title":null,"Body":"<p>This problem occurs beacuse <code>VScode<\/code> and <code>Coderunner<\/code> extension can not activate the <code>conda<\/code> environment you are using. To fix this. <\/p>\n\n<ol>\n<li>Start <code>anaconda command prompt<\/code> then type in <code>activate &lt;your environment name&gt;<\/code><\/li>\n<li>Type in <code>code<\/code> to start <code>VSCode<\/code>. Your import should be working now.<\/li>\n<\/ol>\n\n<p>Thanks!<\/p>\n"}]}
+{"QuestionId":53449927,"AnswerCount":0,"Tags":"<pytorch><torchvision>","CreationDate":"2018-11-23T16:13:23.673","AcceptedAnswerId":null,"OwnerUserId":1051505.0,"Title":"RuntimeError: DataLoader worker is killed by signal: Illegal instruction.","Body":"<p>I am using the following code. I run this on a GPU machine with below config:<\/p>\n\n<pre><code>nvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2017 NVIDIA Corporation\nBuilt on Fri_Sep__1_21:08:03_CDT_2017\nCuda compilation tools, release 9.0, V9.0.176\n<\/code><\/pre>\n\n<p>I am using pytorch version 1.0.0.dev20181123. <\/p>\n\n<p>I face this error: RuntimeError: DataLoader worker (pid 23646) is killed by signal: Illegal instruction.<\/p>\n\n<p>Thinking that its a shared memory problem, I tried the solution here at: <a href=\"https:\/\/www.lucidarme.me\/increase-shared-memory-limit\/\" rel=\"nofollow noreferrer\">https:\/\/www.lucidarme.me\/increase-shared-memory-limit\/<\/a><\/p>\n\n<p>It didn't help. Any pointers?<\/p>\n\n<pre><code>import argparse\nimport json\nimport os\nimport torch\n\n#=====START: ADDED FOR DISTRIBUTED======\nfrom distributed import init_distributed, apply_gradient_allreduce, reduce_tensor\nfrom torch.utils.data.distributed import DistributedSampler\n#=====END:   ADDED FOR DISTRIBUTED======\n\nfrom torch.utils.data import DataLoader\nfrom glow import WaveGlow, WaveGlowLoss\nfrom mel2samp import Mel2Samp\n\ndef load_checkpoint(checkpoint_path, model, optimizer):\n    assert os.path.isfile(checkpoint_path)\n    checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')\n    iteration = checkpoint_dict['iteration']\n    optimizer.load_state_dict(checkpoint_dict['optimizer'])\n    model_for_loading = checkpoint_dict['model']\n    model.load_state_dict(model_for_loading.state_dict())\n    print(\"Loaded checkpoint '{}' (iteration {})\" .format(\n          checkpoint_path, iteration))\n    return model, optimizer, iteration\n\ndef save_checkpoint(model, optimizer, learning_rate, iteration, filepath):\n    print(\"Saving model and optimizer state at iteration {} to {}\".format(\n          iteration, filepath))\n    model_for_saving = WaveGlow(**waveglow_config).cuda()\n    model_for_saving.load_state_dict(model.state_dict())\n    torch.save({'model': model_for_saving,\n                'iteration': iteration,\n                'optimizer': optimizer.state_dict(),\n                'learning_rate': learning_rate}, filepath)\n\ndef train(num_gpus, rank, group_name, output_directory, epochs, learning_rate,\n          sigma, iters_per_checkpoint, batch_size, seed, checkpoint_path):\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    #=====START: ADDED FOR DISTRIBUTED======\n    if num_gpus &gt; 1:\n        init_distributed(rank, num_gpus, group_name, **dist_config)\n    #=====END:   ADDED FOR DISTRIBUTED======\n\n    criterion = WaveGlowLoss(sigma)\n    model = WaveGlow(**waveglow_config).cuda()\n\n    #=====START: ADDED FOR DISTRIBUTED======\n    if num_gpus &gt; 1:\n        model = apply_gradient_allreduce(model)\n    #=====END:   ADDED FOR DISTRIBUTED======\n\n    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)\n\n    # Load checkpoint if one exists\n    iteration = 0\n    if checkpoint_path != \"\":\n        model, optimizer, iteration = load_checkpoint(checkpoint_path, model,\n                                                      optimizer)\n        iteration += 1  # next iteration is iteration + 1\n\n    trainset = Mel2Samp(**data_config)\n    # =====START: ADDED FOR DISTRIBUTED======\n    train_sampler = DistributedSampler(trainset) if num_gpus &gt; 1 else None\n    # =====END:   ADDED FOR DISTRIBUTED======\n    train_loader = DataLoader(trainset, num_workers=1, shuffle=False,\n                              sampler=train_sampler,\n                              batch_size=batch_size,\n                              pin_memory=False,\n                              drop_last=True)\n\n    # Get shared output_directory ready\n    if rank == 0:\n        if not os.path.isdir(output_directory):\n            os.makedirs(output_directory)\n            os.chmod(output_directory, 0o775)\n        print(\"output directory\", output_directory)\n\n    model.train()\n    epoch_offset = max(0, int(iteration \/ len(train_loader)))\n    # ================ MAIN TRAINNIG LOOP! ===================\n    for epoch in range(epoch_offset, epochs):\n        print(\"Epoch: {}\".format(epoch))\n        for i, batch in enumerate(train_loader):\n            model.zero_grad()\n\n            mel, audio = batch\n            mel = torch.autograd.Variable(mel.cuda())\n            audio = torch.autograd.Variable(audio.cuda())\n            outputs = model((mel, audio))\n\n            loss = criterion(outputs)\n            if num_gpus &gt; 1:\n                reduced_loss = reduce_tensor(loss.data, num_gpus).item()\n            else:\n                reduced_loss = loss.item()\n            loss.backward()\n            optimizer.step()\n\n            print(\"{}:\\t{:.9f}\".format(iteration, reduced_loss))\n\n            if (iteration % iters_per_checkpoint == 0):\n                if rank == 0:\n                    checkpoint_path = \"{}\/waveglow_{}\".format(\n                        output_directory, iteration)\n                    save_checkpoint(model, optimizer, learning_rate, iteration,\n                                    checkpoint_path)\n\n            iteration += 1\n\nif __name__ == \"__main__\":\n    parser = argparse.ArgumentParser()\n    parser.add_argument('-c', '--config', type=str,\n                        help='JSON file for configuration')\n    parser.add_argument('-r', '--rank', type=int, default=0,\n                        help='rank of process for distributed')\n    parser.add_argument('-g', '--group_name', type=str, default='',\n                        help='name of group for distributed')\n    args = parser.parse_args()\n\n    # Parse configs.  Globals nicer in this case\n    with open(args.config) as f:\n        data = f.read()\n    config = json.loads(data)\n    train_config = config[\"train_config\"]\n    global data_config\n    data_config = config[\"data_config\"]\n    global dist_config\n    dist_config = config[\"dist_config\"]\n    global waveglow_config\n    waveglow_config = config[\"waveglow_config\"]\n\n    num_gpus = torch.cuda.device_count()\n    if num_gpus &gt; 1:\n        if args.group_name == '':\n            print(\"WARNING: Multiple GPUs detected but no distributed group set\")\n            print(\"Only running 1 GPU.  Use distributed.py for multiple GPUs\")\n            num_gpus = 1\n\n    if num_gpus == 1 and args.rank != 0:\n        raise Exception(\"Doing single GPU training on rank &gt; 0\")\n\n    torch.backends.cudnn.enabled = True\n    torch.backends.cudnn.benchmark = False\n    train(num_gpus, args.rank, args.group_name, **train_config)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53450158,"AnswerCount":0,"Tags":"<tensorflow><object-recognition>","CreationDate":"2018-11-23T16:29:45.647","AcceptedAnswerId":null,"OwnerUserId":1543076.0,"Title":"Tensorflow Lite - detect face parts","Body":"<p>I'm looking for a way to detect parts of a face (like eyes, nose, lips) with Tensorflow Lite. So far, I haven't seen much info. Is there a way to actually retrieve coordinates that describe this kind of data?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":53450218,"AnswerCount":1,"Tags":"<python><tensorflow><sparse-matrix>","CreationDate":"2018-11-23T16:34:35.503","AcceptedAnswerId":53450816.0,"OwnerUserId":1271079.0,"Title":"Tensorflow dense tensor to sparse binarized hash trick tensor","Body":"<p>I want to transform this dataset in such a way that each tensor has a given size <code>n<\/code> and that a feature at index <code>i<\/code> of this new tensor is set to 1 if and only if there is a <code>i<\/code> in the original feature (modulo n).<\/p>\n\n<p>I hope the following example will make things clearer<\/p>\n\n<p>Let's suppose I have a dataset like: <\/p>\n\n<pre><code>t = tf.constant([\n  [0, 3, 4],\n  [12, 2 ,4]])\nds = tf.data.Dataset.from_tensors(t)\n<\/code><\/pre>\n\n<p>I want to get the sparse equivalent of (if <code>n<\/code> = 9)<\/p>\n\n<pre><code>t = tf.constant([\n  [1, 0, 0, 1, 1, 0, 0, 0, 0], # index set to 1 are 0, 3 and 4\n  [0, 0, 1, 1, 1, 0, 0, 0, 0]]) # index set to 1 are 2, 4, and 12%9 = 3\n<\/code><\/pre>\n\n<p>I already know how to obtain a not sparse representation (<a href=\"https:\/\/stackoverflow.com\/questions\/53414433\/tensorflow-tensor-binarization\">Tensorflow: tensor binarization<\/a>) and as I will end up with n > 1 million, I do not want to pass by the dense tensor to get the sparse one<\/p>\n\n<p>thanks<\/p>\n","answers":[{"AnswerId":"53450816","CreationDate":"2018-11-23T17:25:41.420","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a possible implementation for that:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef binarization_sparse(t, n):\n    # Input size\n    t_shape = tf.shape(t)\n    t_rows = t_shape[0]\n    t_cols = t_shape[1]\n    # Make sparse row indices for each value\n    row_idx = tf.tile(tf.range(t_rows)[: ,tf.newaxis], [1, t_cols])\n    # Sparse column indices\n    col_idx = t % n\n    # \"Flat\" indices - needed to discard repetitions\n    total_idx = row_idx * n + col_idx\n    # Remove repeated elements\n    out_idx, _ = tf.unique(tf.reshape(total_idx, [-1]))\n    # Back to row and column indices\n    sparse_idx = tf.stack([out_idx \/\/ n, out_idx % n], axis=-1)\n    # Sparse values\n    sparse_values = tf.ones([tf.shape(sparse_idx)[0]], dtype=t.dtype)\n    # Make sparse tensor\n    out = tf.sparse.SparseTensor(tf.cast(sparse_idx, tf.int64),\n                                 sparse_values,\n                                 [t_rows, n])\n    # Reorder indices\n    out = tf.sparse.reorder(out)\n    return out\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    t = tf.constant([\n        [ 0,  3,  4],\n        [12,  2,  4]\n    ])\n    # Sparse result\n    t_m1h_sp = binarization_sparse(t, 9)\n    # Convert to dense to check output\n    t_m1h = tf.sparse.to_dense(t_m1h_sp)\n    print(sess.run(t_m1h))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[1 0 0 1 1 0 0 0 0]\n [0 0 1 1 1 0 0 0 0]]\n<\/code><\/pre>\n\n<p>I added the logic to remove repeated elements because in principle it could happen, but if you have a guarantee that there are no repetitions (including modulo), you may skip that step. Also, I reorder the sparse tensor at the end. That is not strictly necessary here, but (I think) sparse operations sometimes expect the indices to be ordered (and <code>sparse_idx<\/code> may not be ordered).<\/p>\n\n<p>Also, this solution is specific to 2D inputs. For 1D inputs would be simpler, and it can be written for higher-dimensional inputs as well if necessary. I think a completely general solution is possible but it would be more complicated (specially if you want to consider tensors with unknown number of dimensions).<\/p>\n"}]}
+{"QuestionId":53450466,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><amazon-s3>","CreationDate":"2018-11-23T16:54:46.200","AcceptedAnswerId":null,"OwnerUserId":10696490.0,"Title":"How TensorFlow read file from s3 bytestream","Body":"<p>I have done a deep learning model in TensorFlow for image recognition, and this one works reading an image file from local directory with <code>tf.read_file()<\/code> method, but I need now that the file be read by TensorFlow since a variable that is a Byte-Streaming that extract the image file since an S3 Bucket of Amazon without storage the streaming in local directory  <\/p>\n","answers":[{"AnswerId":"58649310","CreationDate":"2019-10-31T18:25:23.757","ParentId":null,"OwnerUserId":"1082267","Title":null,"Body":"<p>You should be able to pass in the fully formed s3 path to tf.read_file(), like:\n<code>s3:\/\/bucket-name\/path\/to\/file.jpeg<\/code> where <code>bucket-name<\/code> is the name of your s3 bucket, and <code>path\/to\/file.jpeg<\/code> is where it's stored in your bucket. It seems possible you might be running into some access permissions issue, depending on if your bucket is private. You can follow <a href=\"https:\/\/github.com\/tensorflow\/examples\/blob\/master\/community\/en\/docs\/deploy\/s3.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/examples\/blob\/master\/community\/en\/docs\/deploy\/s3.md<\/a> to set up your credentials<\/p>\n\n<p>Is there an error you ran into when doing this?<\/p>\n"}]}
+{"QuestionId":53450495,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><tensorflow-datasets>","CreationDate":"2018-11-23T16:57:27.147","AcceptedAnswerId":53452692.0,"OwnerUserId":1757224.0,"Title":"tf.data API cannot print all the batches","Body":"<p>I am self-teaching myself about <code>tf.data<\/code> API. I am using <code>MNIST<\/code> dataset for binary classification. The training x and y data is zipped together in the full train_dataset. Chained along together with this zip method is first the <code>batch()<\/code> dataset method. the data is batched with a batch size of 30. Since my training set size is 11623, with batch size 128, I will have 91 batches. The size of the last batch will be 103 which is fine since this is LSTM. Additionally, I am using drop-out. When I compute batch accuracy, I am turning off the drop-out.<\/p>\n\n<p>The full code is given below:<\/p>\n\n<pre><code>#Ignore the warnings\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nimport pandas as pd\nimport tensorflow as tf\nimport numpy as np\n\nimport matplotlib.pyplot as plt\nplt.rcParams['figure.figsize'] = (8,7)\n\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\")\n\nXtrain = mnist.train.images[mnist.train.labels &lt; 2]\nytrain = mnist.train.labels[mnist.train.labels &lt; 2]\n\nprint(Xtrain.shape)\nprint(ytrain.shape)\n\n#Data parameters\nnum_inputs = 28\nnum_classes = 2\nnum_steps=28\n\n# create the training dataset\nXtrain = tf.data.Dataset.from_tensor_slices(Xtrain).map(lambda x: tf.reshape(x,(num_steps, num_inputs)))\n# apply a one-hot transformation to each label for use in the neural network\nytrain = tf.data.Dataset.from_tensor_slices(ytrain).map(lambda z: tf.one_hot(z, num_classes))\n# zip the x and y training data together and batch and Prefetch data for faster consumption\ntrain_dataset = tf.data.Dataset.zip((Xtrain, ytrain)).batch(128).prefetch(128)\n\niterator = tf.data.Iterator.from_structure(train_dataset.output_types,train_dataset.output_shapes)\nX, y = iterator.get_next()\n\ntraining_init_op = iterator.make_initializer(train_dataset)\n\n\n#### model is here ####\n\n#Network parameters\nnum_epochs = 2\nbatch_size = 128\noutput_keep_var = 0.5\n\nwith tf.Session() as sess:\n    init.run()\n\n    print(\"Initialized\")\n    # Training cycle\n    for epoch in range(0, num_epochs):\n        num_batch = 0\n        print (\"Epoch: \", epoch)\n        avg_cost = 0.\n        avg_accuracy =0\n        total_batch = int(11623 \/ batch_size + 1)\n        sess.run(training_init_op)\n       while True:\n            try:\n                _, miniBatchCost = sess.run([trainer, loss], feed_dict={output_keep_prob: output_keep_var})\n                miniBatchAccuracy = sess.run(accuracy, feed_dict={output_keep_prob: 1.0})\n               print('Batch %d: loss = %.2f, acc = %.2f' % (num_batch, miniBatchCost, miniBatchAccuracy * 100))\n                num_batch +=1\n            except tf.errors.OutOfRangeError:\n                break\n<\/code><\/pre>\n\n<p>When I run this code, it seems it is working and printing:<\/p>\n\n<pre><code>Batch 0: loss = 0.67276, acc = 0.94531\nBatch 1: loss = 0.65672, acc = 0.92969\nBatch 2: loss = 0.65927, acc = 0.89062\nBatch 3: loss = 0.63996, acc = 0.99219\nBatch 4: loss = 0.63693, acc = 0.99219\nBatch 5: loss = 0.62714, acc = 0.9765\n......\n......\nBatch 39: loss = 0.16812, acc = 0.98438\nBatch 40: loss = 0.10677, acc = 0.96875\nBatch 41: loss = 0.11704, acc = 0.99219\nBatch 42: loss = 0.10592, acc = 0.98438\nBatch 43: loss = 0.09682, acc = 0.97656\nBatch 44: loss = 0.16449, acc = 1.00000\n<\/code><\/pre>\n\n<p>However, as one can see easily, there is something wrong. Only 45 batches are printed not 91 and I do not know why this is happening. I tried so many things and I think I am missing something out. <\/p>\n\n<p>I can use <code>repeat()<\/code> function but I do not want that because I have redundant observations for last batches and I want LSTM to handle it. <\/p>\n","answers":[{"AnswerId":"53452692","CreationDate":"2018-11-23T20:36:45.643","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>This is an annoying pitfall when defining a model based directly on the <code>get_next()<\/code> output of a <code>tf.data<\/code> iterator. In your loop, you have two <code>sess.run<\/code> calls, <em>both<\/em> of which will advance the iterator by one step. This means each loop iteration actually consumes two batches (and also your loss and accuracy calculations are computed on different batches).<\/p>\n\n<p>Not entirely sure if there is a \"canonical\" way of fixing this, but you could<\/p>\n\n<ul>\n<li>compute the accuracy in the same <code>run<\/code> call as the cost\/training step. This would mean that the accuracy calculation is also affected by the dropout mask, but since it's an approximate value based on only one batch, that shouldn't be a huge issue.<\/li>\n<li>define your model based on a placeholder instead, and in each loop iteration <code>run<\/code> the <code>get_next<\/code> op itself, then feed the resulting numpy arrays (i.e. the batch) into the loss\/accuracy computations.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53450677,"AnswerCount":1,"Tags":"<tensorflow><activation-function>","CreationDate":"2018-11-23T17:12:50.170","AcceptedAnswerId":null,"OwnerUserId":10675506.0,"Title":"How to use maxout in Tensorflow?","Body":"<p>guys! I have a question to ask.If I want to use maxout as activation function , how should I write the codes in Tensorflow? An input parameter is required in the slim.maxout() function, so it cannot be used for \nslim.arg_scope([slim.conv], activation_fn=slim.maxout)?What should I do?<\/p>\n","answers":[{"AnswerId":"54394237","CreationDate":"2019-01-28T00:37:59.317","ParentId":null,"OwnerUserId":"9912116","Title":null,"Body":"<p>You may have to define <code>maxout<\/code> in a separate function. For example:<\/p>\n\n<pre><code>def maxout(inputs,num_inputs):\n  return slim.maxout(inputs,num_inputs)\n\nslim.arg_scope([slim.conv],activation_fn=maxout)\n<\/code><\/pre>\n\n<p><em>(I may have the arguments defined incorrectly in the <code>maxout<\/code> function.)<\/em><\/p>\n\n<p>In any case, I recommend that you switch to <code>tf.layers<\/code> (tf core API) because <code>tf.slim<\/code> seems to be in the works of being phased out.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16182#issuecomment-372397483\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16182#issuecomment-372397483<\/a><\/p>\n"}]}
+{"QuestionId":53450891,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-23T17:32:33.097","AcceptedAnswerId":null,"OwnerUserId":9337813.0,"Title":"Use all GPU devices available for tf.estimator.Estimator().predict()","Body":"<p>I am running a model with tensorflow 1.12, I am using the tf.estimator.Estimator API along with MirroredStrategy as train strategy.<\/p>\n\n<p>The train and evaluate methods from the estimator work just fine, using all available devices to produce the results. However, when I try to use predict to get the logits produced by my model, I only see activity in the first available GPU device. I am using predict to process my validation set, and gather the resulting logits, and using just one device to do that is not efficient in my case.<\/p>\n\n<p>So my question: is there any way to specify a distribution strategy to the predict method, so I can make use of all the available GPUs to obtain my result?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[]}
+{"QuestionId":53451052,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning><keras>","CreationDate":"2018-11-23T17:47:18.953","AcceptedAnswerId":null,"OwnerUserId":9566690.0,"Title":"Predictive model giving only one class on holdout set keras","Body":"<p>Currently, using a CNN to predict a model for some black and white images.  The model is working well for what I want it to do. This is the model:<\/p>\n\n<pre><code>lr = 0.2\nAdam(lr=lr, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.001)\nclassifier = Sequential()\n\n# Step 1 - Convolution\nclassifier.add(Conv2D(32, (3, 3), input_shape = (64, 64, 3), activation = 'relu'))\ndropout = 0.3\n# Step 2 - Pooling\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\nclassifier.add(Dropout(rate=dropout))\n# Adding a second convolutional layer\nclassifier.add(Conv2D(32, (3, 3), activation = 'sigmoid'))\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\nclassifier.add(Dropout(rate=dropout))\nclassifier.add(Conv2D(32, (3, 3), activation = 'sigmoid'))\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n# Step 3 - Flattening\nclassifier.add(Flatten())\n\n# Step 4 - Full connection\nclassifier.add(Dense(units = 128, activation = 'relu'))\nclassifier.add(Dense(units = 1, activation = 'sigmoid'))\n\n# Compiling the CNN\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\nepoch_number = 40\nepochs = 25\nsteps_validation = 100\nhistory = classifier.fit_generator(training_set,\n                         steps_per_epoch = epoch_number,\n                         epochs = epochs,\n                         validation_data = test_set,\nvalidation_steps = steps_validation)  \nclassifier.save('epoch40epochs25softmax.h5')\n<\/code><\/pre>\n\n<p>Now, after I train this model I want to use it in the 'real world' so I have a holdout set that I use it on. The problem is it only predicts class 1 when I do a <code>classifier.predict_classes(test_image)<\/code> I figured I was doing something wrong so I tried doing a <code>classifier.predict_classes(test_image)<\/code> for all the elements in the testing set since I knew I was getting 75% accuracy on that set and still upon doing a:<\/p>\n\n<pre><code>for test_image in test_set_dir:\n    classifier.predict_classes(test_image)\n<\/code><\/pre>\n\n<p>I am still only getting numbers very close to <code>[[1]]<\/code>. My question is am I using keras correctly, or if not, what is going on?  I should get some <code>[[0]]<\/code> classifications. <\/p>\n\n<p>I am absolutely sure that the model is not predicting just class 1 since I have a 50% - 50% split on the two classes and I am getting 75% accuracy. <\/p>\n\n<p>Edit:\nI am normalizing the images thusly<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale = 1.\/255,\n                                   shear_range = 0,  \n                                   zoom_range = 0,\n                                   horizontal_flip = True)\n\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ntraining_set = train_datagen.flow_from_directory('datasetI\/training_set',\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'binary')\n\ntest_set = test_datagen.flow_from_directory('datasetI\/test_set',\n                                            target_size = (64, 64),\n                                            batch_size = 32,\n                                            class_mode = 'binary')\n<\/code><\/pre>\n\n<p>If I want to get the class prediction on the test_set what should I do to get it?<\/p>\n","answers":[]}
+{"QuestionId":53451188,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-23T18:01:26.483","AcceptedAnswerId":null,"OwnerUserId":10696125.0,"Title":"Tensorflow - 'str' object has no attribute 'op' - Adam Optimizer - var_list","Body":"<p>I am trying to use Adam with var_list, but it is not working.<\/p>\n\n<pre><code> def get_name_trainable_variables(self):\n    var_names = []\n    for var in tf.trainable_variables():\n\n        var_names.append(var.name)\n\n    return var_names\n<\/code><\/pre>\n\n<pre class=\"lang-none prettyprint-override\"><code>\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/optimizer.py in _get_processor(v)\n    200     # True if and only if `v` was initialized eagerly.\n    201     return _DenseResourceVariableProcessor(v)\n--&gt; 202   if v.op.type == \"VarHandleOp\":\n    203     return _DenseResourceVariableProcessor(v)\n    204   if isinstance(v, variables.Variable):\n\nAttributeError: 'str' object has no attribute 'op'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53451728,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><computer-vision>","CreationDate":"2018-11-23T18:54:56.760","AcceptedAnswerId":null,"OwnerUserId":868736.0,"Title":"n-dimensional sliding window operation in python using gpu accelerated libraries, preferably Tensorflow?","Body":"<p>I am looking for a gpu-accelerated n-dimensional sliding window operation implementation in Python using Tensorflow. You can post your implementation in Torch, Caffe or Theano, but I'll choose the Tensorflow implementation as the accepted answer.  Please post working code snippet that performs a 2d median filter operation (hopefully, with no code change or minimal code change, can be applied to n-dimensional images)<\/p>\n\n<p>With my limited knowledge on Tensorflow, I believe the 2 potential modules to start with are <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/data\/sliding_window_batch\" rel=\"nofollow noreferrer\"><code>sliding_window_batch<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/extract_image_patches\" rel=\"nofollow noreferrer\"><code>extract_image_patches<\/code><\/a> and then with some <code>map<\/code>,<code>apply<\/code>,<code>reshape<\/code> magic?<\/p>\n\n<p>My failed attempt is posted below, for entertainment.  Please note I have posted a similar <a href=\"https:\/\/stackoverflow.com\/questions\/35691947\/3d-sliding-window-operation-in-theano\">question<\/a> 2 years ago, asking for a Theano implementation, nowadays, most people are using tf\/keras or torch.<\/p>\n\n<pre><code>import time\nimport matplotlib.pyplot as plt\n%matplotlib inline\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.contrib.data.python.ops import sliding\nfrom skimage import img_as_float, data\nfrom scipy.signal import medfilt\n\nimgs = img_as_float(data.camera())\n\n### SCIPY median ###\nstime = time.time()\nscipysmoothed = medfilt(imgs,(9,9))\netime = time.time()\nprint('scipy smoothed: {:1.4f} seconds'.format(etime-stime))\n\n### Failed attempt of TF median ###\nmethod = 'Tensorflow'\nstime = time.time()\n\nwindow_func = lambda x: tf.contrib.distributions.percentile(x, 50.0)\n\n# create TensorFlow Dataset object\ndata = tf.data.Dataset.from_tensor_slices(imgs)\n\n# sliding window - only 1d is allowed?\nwindow = 3\nstride = 1\ndata = data.apply(sliding.sliding_window_batch(window, stride)).map(lambda x: window_func(x))\n\n# create TensorFlow Iterator object\niterator =  tf.data.Iterator.from_structure(data.output_types)\nnext_element = iterator.get_next()\n\n# create initialization ops \ninit_op = iterator.make_initializer(data)\nc=0\nsmoothed = np.zeros(imgs.shape)\nwith tf.Session() as sess:\n    # initialize the iterator on the data\n    sess.run(init_op)\n    while True:\n        try:\n            elem = sess.run(next_element)\n            smoothed[c,:]=elem\n            # obviously WRONG.\n            c+=1\n        except tf.errors.OutOfRangeError:\n            #print(\"End of dataset.\")\n            break\n#print(c)\netime = time.time()\nprint('tf smoothed: {:1.4f} seconds'.format(etime-stime))\n\n\nplt.figure(figsize=(20,20))\nplt.subplot(131)\nplt.imshow(imgs,cmap='gray',interpolation='none')\nplt.title('original')\nplt.subplot(132)\nplt.imshow(smoothed,cmap='gray',interpolation='none')\nplt.title('actual smoothed\\nwith {}'.format(method))\nplt.subplot(133)\nplt.imshow(scipysmoothed,cmap='gray',interpolation='none')\n_=plt.title('expected smoothed')\n<\/code><\/pre>\n\n<p>.<\/p>\n\n<pre><code>scipy smoothed: 1.1899 seconds\ntf smoothed: 0.7485 seconds\n<\/code><\/pre>\n","answers":[{"AnswerId":"53452423","CreationDate":"2018-11-23T20:07:45.107","ParentId":null,"OwnerUserId":"868736","Title":null,"Body":"<p>Proposal 1: My attempt is the below, since it just uses <code>tf.image.extract_image_patches<\/code> and <code>tf.extract_volume_patches<\/code>, the implementation supports only 2d and 3d images.<\/p>\n\n<p>Proposal 2: one could just format the data as a preprocessing step (via <code>tf.data.Dataset.map<\/code>), however this also takes alot of time, I am not sure why yet ( example <a href=\"https:\/\/gist.github.com\/pangyuteng\/ca5cb07fe383ebe59b521c832f2e2918\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/pangyuteng\/ca5cb07fe383ebe59b521c832f2e2918<\/a> ).<\/p>\n\n<p>Proposal 3: use convolutional blocks to parallelize processing, see \"Hypercolumns for Object Segmentation and Fine-grained Localization\" <a href=\"https:\/\/arxiv.org\/abs\/1411.5752\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1411.5752<\/a> .<\/p>\n\n<p>--<\/p>\n\n<p>Proposal 1 code:<\/p>\n\n<pre><code>import time\nimport matplotlib.pyplot as plt\n%matplotlib inline\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.contrib.data.python.ops import sliding\nfrom skimage import img_as_float, data\nfrom scipy.signal import medfilt\n\ndtype = 2\nif dtype==2:\n    imgs = img_as_float(data.camera())\nelif dtype==3:\n    imgs = np.random.rand(28,28,28)\n\nimgs = img_as_float(data.camera())\n\n### SCIPY median ###\nstime = time.time()\nscipysmoothed = medfilt(imgs,(9,9))\netime = time.time()\nprint('scipy smoothed: {:1.4f} seconds'.format(etime-stime))\n\n### TF median ###\nmethod = 'Tensorflow'\nimgs = np.expand_dims(imgs,axis=-1)\nimgs = np.expand_dims(imgs,axis=0)\nprint('imgs.shape:{}'.format(imgs.shape))\nimgs = tf.cast(imgs,tf.float32)\n\nstime = time.time()\n\nif len(imgs.shape) == 4:\n    kernel=(1,9,9,1)\n    stride=(1,1,1,1)\n    rates=(1,1,1,1)\n    padding='SAME'\n    patches=tf.image.extract_image_patches(\n        imgs,kernel,stride,rates,padding,\n    )    \n    _,x,y,n = patches.shape\n    _,sx,sy,_ = kernel\n    window_func = lambda x: tf.contrib.distributions.percentile(x, 50.0)\n    patches = tf.reshape(patches,[x*y,sx,sy])\n    smoothed = tf.map_fn(lambda x: window_func(patches[x,:,:]), tf.range(x*y), dtype=tf.float32)\n    smoothed = tf.reshape(smoothed,[x,y])\n\nelif len(imgs.shape) == 5:\n\n    kernel=(1,12,12,12,1)\n    stride=(1,1,1,1,1)    \n    padding='SAME'\n    patches=tf.extract_volume_patches(\n        imgs,kernel,stride,padding,\n    )\n    _,x,y,z,n = patches.shape\n    _,sx,sy,sz,_ = kernel\n    window_func = lambda x: tf.contrib.distributions.percentile(x, 50.0)\n    patches = tf.reshape(patches,[x*y*z,sx,sy,sz])\n    smoothed = tf.map_fn(lambda x: window_func(patches[x,:,:]), tf.range(x*y*z), dtype=tf.float32)\n    smoothed = tf.reshape(smoothed,[x,y,z])\n\nelse:\n    raise NotImplemented()\n\nwith tf.Session() as sess:\n    output = sess.run(smoothed)\n\netime = time.time()\nprint('tf smoothed: {:1.4f} seconds'.format(etime-stime))\n\nprint(output.shape)\n\nplt.figure(figsize=(20,20))\nplt.subplot(131)\nimgs = img_as_float(data.camera())\nplt.imshow(imgs.squeeze(),cmap='gray',interpolation='none')\nplt.title('original')\nplt.subplot(132)\nplt.imshow(output.squeeze(),cmap='gray',interpolation='none')\nplt.title('actual smoothed\\nwith {}'.format(method))\nplt.subplot(133)\nplt.imshow(scipysmoothed,cmap='gray',interpolation='none')\n_=plt.title('expected smoothed')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53451927,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2018-11-23T19:14:54.703","AcceptedAnswerId":null,"OwnerUserId":1171204.0,"Title":"tensorflow - tf.data.Dataset randomly skip samples before batching to get different batches","Body":"<p>My model consumes chronologically ordered sequences within each input batch. Therefore, i am creating batches before shuffling my input data. This brings the issue that batches always include the same data samples across the whole dataset (starting with the same indices - shifted by <code>batch_size<\/code>), i solved this issue by caching the initial dataset and sampling from skipped datasets, however this eats up memory pretty fast (though my dataset has only 150MB):<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(data)\ndataset = dataset.window(size=window_size, shift=window_shift, stride=window_stride, drop_remainder=True).flat_map(lambda x: x.batch(window_size))\ndataset = dataset.map(process_fn, num_parallel_calls=8)\ndataset = dataset.cache()\ndatasets = []\nfor i in range(0, batch_size):\n    d = dataset.skip(i)\n    d = d.batch(batch_size, drop_remainder=True)\n    datasets.append(d)\ndataset = tf.data.experimental.sample_from_datasets(datasets)\ndataset = dataset.shuffle(buffer_size=30000, reshuffle_each_iteration=False)\ndataset = dataset.repeat()\n<\/code><\/pre>\n\n<p>Is there another way to achieve this behaviour? I want to cover all possible indices for the start of the first sequence inside a batch.<\/p>\n","answers":[{"AnswerId":"55509708","CreationDate":"2019-04-04T07:03:03.573","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>You are eating up memory because you are shuffling entire batches -- also skipping may not be very efficient. Since your data seems to be entire in memory, you could possibly sample your data directly in python without too much concern about performance:<\/p>\n\n<pre><code>def make_batch(start_idx):\n  batch = np.empty((batch_size, window_size), dtype=data.dtype)\n  for batch_idx, data_idx in enumerate(\n      range(start_idx, start_idx + window_shift * batch_size, window_shift)):\n    batch[batch_idx] = data[data_idx:data_idx + window_size * window_stride:window_stride]\n  return batch\n\ndataset = (tf.data.Dataset\n  .range(len(data) - window_stride * (window_size - 1) - window_shift * (batch_size- 1))\n  .shuffle(buffer_size=30000, reshuffle_each_iteration=False)\n  .map(lambda x: tf.py_func(make_batch, [x], tf.float32)) # assuming your data is float32\n  .repeat()\n  .prefetch(1)) # you might want to consider prefetching for performance\n<\/code><\/pre>\n\n<p>The shuffling now occurs on indices, not on entire batches, so it has a much lower memory footprint.<\/p>\n"}]}
+{"QuestionId":53452049,"AnswerCount":1,"Tags":"<random><pytorch><random-seed>","CreationDate":"2018-11-23T19:27:55.970","AcceptedAnswerId":53453809.0,"OwnerUserId":8898218.0,"Title":"understanding seed of a ByteTensor in PyTorch","Body":"<p>I understand that a seed is a number used to initialize pseudo-random number generator. in pytorch, <code>torch.get_rng_state<\/code> documentation states as follows \"Returns the random number generator state as a torch.ByteTensor.\". and when i print it i get a 1-d tensor of size 5048 whose values are as shown below<\/p>\n\n<blockquote>\n  <p>tensor([ 80,  78, 248,  ...,   0,   0,   0], dtype=torch.uint8)<\/p>\n<\/blockquote>\n\n<p>why does a seed have 5048 values and how is this different from usual seed which we can get using  <code>torch.initial_seed<\/code><\/p>\n","answers":[{"AnswerId":"53453809","CreationDate":"2018-11-23T23:21:11.280","ParentId":null,"OwnerUserId":"2166798","Title":null,"Body":"<p>It sounds like you're thinking of the seed and the state as equivalent.  For older pseudo-random number generators (PRNGs) that was true, but with more modern PRNGs tend to work as described <a href=\"https:\/\/stackoverflow.com\/a\/48505649\/2166798\">here<\/a>. (The answer in the link was written with respect to Mersenne Twister, but the concepts apply equally to other generators.)<\/p>\n\n<p>Why is it a good idea to not have a 32- or 64-bit state space and report the state as the generator's output?  Because if you do that, as soon as you see any value repeat the entire sequence will repeat.  PRNGs were designed to be \"full cycle,\" i.e., to iterate through the maximum number of values possible before repeating.  <a href=\"https:\/\/www.sciencedirect.com\/science\/article\/pii\/0167637786900921\" rel=\"nofollow noreferrer\">This paper<\/a> showed that the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Birthday_problem\" rel=\"nofollow noreferrer\">birthday problem<\/a> could quickly (O(sqrt(cycle-length)) identify such PRNGs as non-random.  This meant, for instance, that with 32-bit integers you shouldn't use more than ~50000 values before a statistician could call you out with a better than 99% level of confidence.  The solution, used by many modern PRNGs, is to have a larger state space and collapse it down to output a 32- or 64-bit result.  Since multiple states can produce the same output, duplicates will occur in the output stream without the entire stream being replicated.  It looks like that's what PyTorch is doing.<\/p>\n\n<p>Given the larger state space, why allow seeding with a single integer?  Convenience.  For instance, Mersenne Twister has a 19,937 bit state space, but most people don't want to enter that much info to kick-start it.  You can if you want to, but most people use the front-end which populates the full state space from a single integer input.<\/p>\n"}]}
+{"QuestionId":53452540,"AnswerCount":1,"Tags":"<lambda><keras>","CreationDate":"2018-11-23T20:20:55.323","AcceptedAnswerId":53452695.0,"OwnerUserId":1716733.0,"Title":"load_model with Lamda layer created in a loop in Keras","Body":"<p>I have a functional model that creates several gates in a loop in keras with <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>predictions = []\nfor ii, kk in enumerate(label_cols):\n    slicer = Lambda(lambda x: x[:,:,:,ii:ii+1],\n                    output_shape=gates_shape[:-2]+(1,),\n                    name='slice_'+kk)\n    gate_  = slicer(gates)\n    ...\n    prediction = Dense(n_classes[kk], activation=final_activation, name=kk)(x)\n    predictions.append(prediction)\n<\/code><\/pre>\n\n<p>When I'm trying to load the model, it complains:<\/p>\n\n<pre><code>\/anaconda3\/lib\/python3.6\/site-packages\/keras\/layers\/core.py in call(self, inputs, mask)\n    691         if has_arg(self.function, 'mask'):\n    692             arguments['mask'] = mask\n--&gt; 693         return self.function(inputs, **arguments)\n    694 \n    695     def compute_mask(self, inputs, mask=None):\n\n\/anaconda3\/lib\/python3.6\/site-packages\/keras\/layers\/core.py in &lt;lambda&gt;(x)\n    334 print('gates_shape', gates_shape)\n    335 for ii, kk in enumerate(prms.label_cols):\n--&gt; 336     slicer = Lambda(lambda x: x[:,:,:,ii:ii+1], \n    337                     output_shape=gates_shape[:-2]+(1,),\n    338                     name='slice_'+kk)\n\nNameError: name 'ii' is not defined\n<\/code><\/pre>\n\n<p>Everything worked before I added the <code>Lambda<\/code>.\nWhat is the right way to do it?<\/p>\n","answers":[{"AnswerId":"53452695","CreationDate":"2018-11-23T20:37:19.617","ParentId":null,"OwnerUserId":"5536887","Title":null,"Body":"<p>The <code>ii<\/code> variable is not in scope of the lambda, so you have to pass it in through <code>arguments<\/code>.<\/p>\n\n<p>Try:<\/p>\n\n<pre><code>x = Lambda(lambda x,ii: x[:,:,:,ii:ii+1], arguments={'ii':ii})  \n<\/code><\/pre>\n"}]}
+{"QuestionId":53453663,"AnswerCount":1,"Tags":"<python><image-processing><deep-learning><bigdata><pytorch>","CreationDate":"2018-11-23T22:52:45.360","AcceptedAnswerId":53460973.0,"OwnerUserId":7606121.0,"Title":"Load huge image dataset and some data augmentation using Pytorch","Body":"<p>I have a question. I have about 2 million images (place365-standard dataset) and I want to do some data augmentation like transforming, cropping etc. Also, I have to make my own target image (y) based on some color model algorithms (CMYK) for example.<\/p>\n\n<p>So Actually, my preprocessing step includes augmentation and making terget image (y). And then I should feed these images to a deep network. When should I do this based on <code>Dataset<\/code> class? Should I do my processing step in <code>__getitem__()<\/code>? If yes, would it be parallel and fast? <\/p>\n\n<p>Here is my template of <code>Dataset(data.Dataset)<\/code> class:<\/p>\n\n<pre><code>import torch\nfrom torch.utils import data\n\nclass Dataset(data.Dataset):\n    \"\"\"\n    Return Dataset class representing our data set\n    \"\"\"\n    def __int__(self, list_IDs, labels):\n        \"\"\"\n        Initialize data set as a list of IDs corresponding to each item of data set and labels of each data\n\n        Args:\n            list_IDs: a list of IDs for each data point in data set\n            labels: label of an item in data set with respect to the ID\n        \"\"\"\n\n        self.labels = labels\n        self.list_IDs = list_IDs\n\n    def __len__(self):\n        \"\"\"\n        Return the length of data set using list of IDs\n\n        :return: number of samples in data set\n        \"\"\"\n        return len(self.list_IDs)\n\n    def __getitem__(self, item):\n        \"\"\"\n        Generate one item of data set. Here we apply our preprocessing things like halftone styles and subtractive color process using CMYK color model etc. (See the paper for operations)\n\n        :param item: index of item in IDs list\n\n        :return: a sample of data\n        \"\"\"\n        ID = self.list_IDs[item]\n\n        # Code to load data\n        X = None #\n\n        # code to apply your custom function to make y image (time consuming task - some algorithms)\n        y = None #\n\n        return X, y\n<\/code><\/pre>\n\n<p>Thanks for any advice <\/p>\n\n<p>Best regards<\/p>\n","answers":[{"AnswerId":"53460973","CreationDate":"2018-11-24T18:03:14.477","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you look at e.g., <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/datasets.html#imagefolder\" rel=\"nofollow noreferrer\"><code>torchvision.dataset.ImageFolder<\/code><\/a> you'll see that it works quite similar to your design: the class has <code>transform<\/code> member that lists all sorts of augmentations (resizing, cropping, flipping etc.) and these are carried out on the images in the <code>__getitem__<\/code> method.<br>\nRegarding parallelism, the <code>Dataset<\/code> itself is not parallel, but the <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>DataLoader<\/code><\/a> can be (see <code>num_workers<\/code> argument), so if you use your dataset inside a parallel dataloader you get the parallelism for free, Cool!<\/p>\n"}]}
+{"QuestionId":53453675,"AnswerCount":0,"Tags":"<keras><pyspark>","CreationDate":"2018-11-23T22:54:54.370","AcceptedAnswerId":null,"OwnerUserId":7295231.0,"Title":"PicklingError in when using KerasClassifier for Cross-Validation in Spark","Body":"<p>Though I am able to run the following code perfectly on my local:<\/p>\n\n<pre><code>`from keras.wrappers.scikit_learn import KerasClassifier\n from sklearn.model_selection import cross_val_score\n from keras.models import Sequential\n from keras.layers import Dense\n def build_classifier():\n     classifier = Sequential()\n     classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu', input_dim = 11))\n     classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu'))\n     classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n     classifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n     return classifier\nclassifier = KerasClassifier(build_fn = build_classifier, batch_size = 10, epochs = 100)\naccuracies = cross_val_score(estimator = classifier, X = X_train, y = y_train, cv = 10, n_jobs = -1)`\n<\/code><\/pre>\n\n<p>I am getting <code>PicklingError<\/code> : <code>Can't pickle &lt;function build_classifier at 0x7f98f3cbabf8&gt;: attribute lookup build_classifier on __main__ failed<\/code><\/p>\n\n<p>Based on my understanding of the <code>PicklingError<\/code> when trying to parallelize in <code>Python<\/code>, it looks like there is some issue with the position where I have defined the function, but I am not able to make it run. How can I troubleshoot this?<\/p>\n","answers":[]}
+{"QuestionId":53454320,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><spyder>","CreationDate":"2018-11-24T01:02:33.370","AcceptedAnswerId":null,"OwnerUserId":6538281.0,"Title":"Spyder IDE won't open","Body":"<p>so I had Python 3.7.0 and my Spyder IDE was working fine.\nHowever, because there was no version of tensorflow for that version of python I decided to downgrade my python to version 3.5.0.\nNow realising that was a trade-off, I'm assuming there's a correlation there because my  Spyder IDE no longer launches, neither does my Anaconda Navigator.\nI have already done a Spyder --reset command by the way, no luck.\nHow can I fix this issue while maintaining to use tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":53454589,"AnswerCount":0,"Tags":"<python><deep-learning><time-series><pytorch><rnn>","CreationDate":"2018-11-24T02:04:16.853","AcceptedAnswerId":null,"OwnerUserId":5182620.0,"Title":"Pytorch RNN always gives the same output for multivariate time series","Body":"<p>I have a time series data looking something like this:\n<a href=\"https:\/\/i.stack.imgur.com\/4vaBL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4vaBL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am trying to model this with a sequence to sequence RNN in pytorch. It trains well and I can see the loss going down. But on testing it gives the same out put irrespective of the input.<\/p>\n\n<p>My Model:<\/p>\n\n<pre><code>class RNNModel(nn.Module):\n\ndef __init__(self, predictor_size, hidden_size, num_layers, dropout = 0.3, output_size=83):\n    super(RNNModel, self).__init__()\n    self.drop = nn.Dropout(dropout)\n    self.rnn = nn.GRU(predictor_size, hidden_size, num_layers=num_layers, dropout = dropout)\n    self.decoder = nn.Linear(hidden_size, output_size)\n    self.init_weights()\n    self.hidden_size = hidden_size\n    self.num_layers = num_layers\n\ndef init_weights(self):\n    initrange = 0.1\n    self.decoder.bias.data.fill_(0)\n    self.decoder.weight.data.uniform_(-initrange, initrange)\n\ndef forward(self, input, hidden):\n    output, hidden = self.rnn(input, hidden)\n    output = self.drop(output)\n    decoded = self.decoder(output.view(output.size(0) * output.size(1), output.size(2)))\n    return decoded.view(output.size(0), output.size(1), decoded.size(1)), hidden\n\ndef init_hidden(self, batch_size):\n    weight = next(self.parameters()).data\n    return Variable(weight.new(self.num_layers, batch_size, self.hidden_size).zero_())\n<\/code><\/pre>\n\n<p>Train Method:<\/p>\n\n<pre><code>def train(data_source, lr):\n    # turn on training mode that enables dropout\n\n    model.train()\n    total_loss = 0\n    hidden = model.init_hidden(bs_train)\n    optimizer = optim.Adam(model.parameters(), lr = lr)\n\n    for batch, i in enumerate(range(0, data_source.size(0) - 1, bptt_size)):\n\n        data, targets = get_batch(data_source, i)\n\n        # Starting each batch, we detach the hidden state from how it was previously produced\n        # so that model doesen't ry to backprop to all the way start of the dataset\n        # unrolling of the graph will go from the last iteration to the first iteration\n        hidden = Variable(hidden.data)\n        if cuda.is_available():\n            hidden = hidden.cuda()\n        optimizer.zero_grad()\n\n        output, hidden = model(data, hidden)\n        loss = criterion(output, targets)\n        loss.backward()\n\n        # clip_grad_norm to prevent gradient explosion\n        torch.nn.utils.clip_grad_norm(model.parameters(), clip)\n\n        optimizer.step()\n        total_loss += len(data) * loss.data\n        # return accumulated loss for all the iterations\n        return total_loss[0] \/ len(data_source)\n<\/code><\/pre>\n\n<p>Evaluation Method:<\/p>\n\n<pre><code>def evaluate(data_source):\n    # turn on evaluation to disable dropout\n    model.eval()\n    model.train(False)\n    total_loss = 0\n    hidden = model.init_hidden(bs_valid)\n\n    for i in range(0, data_source.size(0) - 1, bptt_size):\n        data, targets = get_batch(data_source, i, evaluation = True)\n\n        if cuda.is_available():\n            hidden = hidden.cuda()\n\n        output, hidden = model(data, hidden)\n        total_loss += len(data) * criterion(output, targets).data\n        hidden = Variable(hidden.data)\n\n    return total_loss[0]\/len(data_source)\n<\/code><\/pre>\n\n<p>Training Loop:<\/p>\n\n<pre><code>best_val_loss = None\nbest_epoch = 0\ndef run(epochs, lr):\n    val_losses = []\n    num_epochs = []\n    global best_val_loss\n    global best_epoch\n    for epoch in range(0, epochs):\n        train_loss = train(train_set, lr)\n        val_loss = evaluate(test_set)\n        num_epochs.append(epoch)\n        val_losses.append(val_loss)\n        print(\"Train Loss: \", train_loss, \" Validation Loss: \", val_loss)\n\n        if not best_val_loss or val_loss &lt; best_val_loss:\n            best_val_loss = val_loss\n            torch.save(model.state_dict(), \".\/4.model.pth\")\n            best_epoch = epoch\n    return num_epochs, val_losses\n<\/code><\/pre>\n\n<p>Loss with epochs:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/M0nwj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/M0nwj.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Getting the output:<\/p>\n\n<pre><code>model = RNNModel(predictor_size, hidden_size, num_layers, dropout_pct, output_size)\nmodel.load_state_dict(torch.load(\".\/4.model.pth\"))\n\nif cuda.is_available():\n    model.cuda()\n\nmodel.eval()\nmodel.train(False)\nhidden = model.init_hidden(1)\ninp = torch.Tensor(var[105])\ninput = Variable(inp.contiguous().view(1,1,predictor_size), volatile=True)\nif cuda.is_available():\n    input.data = input.data.cuda()\noutput, hidden = model(input, hidden)\nop = output.squeeze().data.cpu()\nprint(op)\n<\/code><\/pre>\n\n<p>Here I always get the same output irrespective of datapoint I give as input. Can somebody please tell me what I am doing wrong.<\/p>\n","answers":[]}
+{"QuestionId":53454667,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2018-11-24T02:23:42.070","AcceptedAnswerId":null,"OwnerUserId":6939251.0,"Title":"Why do my Tensorflow neural networks act so slow after I start training it?","Body":"<p>My <strong>Tensorflow neural network<\/strong> gets progressively slower as it trains. Every time I train the network, it acts slower and slower. <\/p>\n\n<p><strong>After the 30th or so iteration of training, it becomes unbearably slow<\/strong> and it's virtually unusable. By the 60th or so iteration, the program stop responding.<\/p>\n\n<p>I didn't think that this neural network is that complex. It's a simple three layer network put together with Tensorflow.<\/p>\n\n<p>Do you guys have any idea how to go about this problem?<\/p>\n\n<pre><code>import tensorflow as tf\n\nhidden_1_layer = {'weights': tf.Variable(tf.random_normal([37500, 500])),\n                  'biases': tf.Variable(tf.random_normal([500]))}\nhidden_2_layer = {'weights': tf.Variable(tf.random_normal([500, 250])),\n                  'biases': tf.Variable(tf.random_normal([250]))}\nhidden_3_layer = {'weights': tf.Variable(tf.random_normal([250, 125])),\n                  'biases': tf.Variable(tf.random_normal([125]))}\noutput_layer = {'weights': tf.Variable(tf.random_normal([125, 1])),\n                'biases': tf.Variable(tf.random_normal([1]))}\n\nclass ImageNN():\n    def train(self, array, target):\n        x = tf.placeholder('float', name='x')\n        l1 = tf.add(tf.matmul(x, hidden_1_layer['weights']), hidden_1_layer['biases'])\n        l1 = tf.nn.relu(l1)\n        l2 = tf.add(tf.matmul(l1, hidden_2_layer['weights']), hidden_2_layer['biases'])\n        l2 = tf.nn.relu(l2)\n        l3 = tf.add(tf.matmul(l2, hidden_3_layer['weights']), hidden_3_layer['biases'])\n        l3 = tf.nn.relu(l3)\n        output = tf.add(tf.matmul(l3, output_layer['weights']), output_layer['biases'])\n        output = tf.nn.sigmoid(output)\n        cost = tf.square(output-target)\n        optimizer = tf.train.AdamOptimizer().minimize(cost)\n        array = array.reshape(1, 37500)\n        with tf.Session() as sess:\n            sess.run(tf.global_variables_initializer())\n            sess.run(optimizer, feed_dict={x: array})\n            sess.close()\n        del x, l1, l2, output, cost, optimizer\n\n    #Do computations with our artificial nueral network\n    def predict(self, data):          #Input data is of size (37500,)\n        x = tf.placeholder('float', name='x')    #get data into the right rank (dimensions), this is just a placeholder, it has no values\n        l1 = tf.add(tf.matmul(x, hidden_1_layer['weights']), hidden_1_layer['biases'])\n        l1 = tf.nn.relu(l1)\n        l2 = tf.add(tf.matmul(l1, hidden_2_layer['weights']), hidden_2_layer['biases'])\n        l2 = tf.nn.relu(l2)\n        l3 = tf.add(tf.matmul(l2, hidden_3_layer['weights']), hidden_3_layer['biases'])\n        l3 = tf.nn.relu(l3)\n        output = tf.add(tf.matmul(l3, output_layer['weights']), output_layer['biases'])\n        output = tf.nn.sigmoid(output)\n        data = data.reshape(1, 37500)\n        with tf.Session() as sess:\n            sess.run(tf.global_variables_initializer())\n            theOutput = sess.run(output, feed_dict={x: data})\n            sess.close()\n        del x, l1, l2, output, data\n        return theOutput\n<\/code><\/pre>\n","answers":[{"AnswerId":"53454720","CreationDate":"2018-11-24T02:36:47.553","ParentId":null,"OwnerUserId":"6679549","Title":null,"Body":"<p>It sounds like a memory issue. You're not deleting l3 or array in your train method, or l3 in your predict method. I don't think this is the cause of the issue as this should get thrown away by the Python interpreter anyway I believe.<\/p>\n\n<p>How are you calling this class? It might be that you're holding onto the output in memory and it's getting very large. <\/p>\n"}]}
+{"QuestionId":53454781,"AnswerCount":0,"Tags":"<python><tensorflow><gpu>","CreationDate":"2018-11-24T02:54:40.400","AcceptedAnswerId":null,"OwnerUserId":10655065.0,"Title":"How to confirm that my GPU is running at full capacity? The GPU load is so low","Body":"<p>My environment: windows10, tensorflow-gpu, single Nvidia RTX2070 (8GB memory), using tflearn to run AlexNet with the OxFlowers17 data set, batch_size=128.\nI got this information:<a href=\"https:\/\/i.stack.imgur.com\/SGuQ7.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>This means the GPU is running, right?\nHowever, when I check the GPU usage with GPU-Z, the GPU load is low (The code is doing training, epoch information is being printed in the output window). Does this mean that my GPU is not running at full capacity?<a href=\"https:\/\/i.stack.imgur.com\/heldO.gif\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>After I closed PyCharm and with no big program running on my computer, the GPU usage is this:<a href=\"https:\/\/i.stack.imgur.com\/ULyFf.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>Then I have three questions:<\/p>\n\n<ol>\n<li><p>The GPU load is low. Does this mean the GPU is not running at full capacity?<\/p><\/li>\n<li><p>I forgot to set the gpu_memory_fraction. If I want to allocate a suitable gpu memory, how much should I set? (e.g. AlexNet with OxFlowers17 data set, the size of the data set is about 60MB.) <\/p><\/li>\n<li><p>Why was the System Memory Used also so high? It has already used all the free gpu memory.<\/p><\/li>\n<\/ol>\n\n<p>PS: This is the CPU (i5-8500) information when training:<a href=\"https:\/\/i.stack.imgur.com\/7JlkT.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n","answers":[]}
+{"QuestionId":53455780,"AnswerCount":1,"Tags":"<lstm><pytorch><one-hot-encoding><cross-entropy>","CreationDate":"2018-11-24T06:34:23.643","AcceptedAnswerId":53458159.0,"OwnerUserId":6755360.0,"Title":"Pytorch LSTM: Target Dimension in Calculating Cross Entropy Loss","Body":"<p>I've been trying to get an LSTM (LSTM followed by a linear layer in a custom model), working in Pytorch, but was getting the following error when calculating the loss:<\/p>\n\n<p><code>Assertion cur_target &gt;= 0 &amp;&amp; cur_target &lt; n_classes' failed.<\/code><\/p>\n\n<p>I defined the loss function with:<\/p>\n\n<pre><code>criterion = nn.CrossEntropyLoss()\n<\/code><\/pre>\n\n<p>and then called with <\/p>\n\n<pre><code>loss += criterion(output, target)\n<\/code><\/pre>\n\n<p>I was giving the target with dimensions [sequence_length, number_of_classes], and output has dimensions [sequence_length, 1, number_of_classes].<\/p>\n\n<p>The examples I was following seemed to be doing the same thing, but it was different on the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#crossentropyloss\" rel=\"nofollow noreferrer\">Pytorch docs on cross entropy loss.<\/a><\/p>\n\n<p>The docs say the target should be of dimension (N), where each value is  0 \u2264 targets[i] \u2264 C\u22121 and C is the number of classes. I changed the target to be in that form, but now I'm getting an error saying (The sequence length is 75, and there are 55 classes):<\/p>\n\n<pre><code>Expected target size (75, 55), got torch.Size([75])\n<\/code><\/pre>\n\n<p>I've tried looking at solutions for both errors, but still can't get this working properly. I'm confused as to the proper dimensions of target, as well as the actual meaning behind the first error (different searches gave very different meanings for the error, none of the fixes worked).<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53458159","CreationDate":"2018-11-24T12:30:05.363","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>You can use <code>squeeze()<\/code> on your <code>output<\/code> tensor, this returns a tensor with all the dimensions of size 1 removed.<\/p>\n\n<p>This short code uses the shapes you mentioned in your question:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sequence_length   = 75\nnumber_of_classes = 55\n# creates random tensor of your output shape\noutput = torch.rand(sequence_length, 1, number_of_classes)\n# creates tensor with random targets\ntarget = torch.randint(55, (75,)).long()\n\n# define loss function and calculate loss\ncriterion = nn.CrossEntropyLoss()\nloss = criterion(output, target)\nprint(loss)\n<\/code><\/pre>\n\n<p>Results in the error you described:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ValueError: Expected target size (75, 55), got torch.Size([75])\n<\/code><\/pre>\n\n<p>So using <code>squeeze()<\/code> on your <code>output<\/code> tensor solves your problem by getting it to correct shape.<\/p>\n\n<p><strong>Example with corrected shape:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sequence_length   = 75\nnumber_of_classes = 55\n# creates random tensor of your output shape\noutput = torch.rand(sequence_length, 1, number_of_classes)\n# creates tensor with random targets\ntarget = torch.randint(55, (75,)).long()\n\n# define loss function and calculate loss\ncriterion = nn.CrossEntropyLoss()\n\n# apply squeeze() on output tensor to change shape form [75, 1, 55] to [75, 55]\nloss = criterion(output.squeeze(), target)\nprint(loss)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor(4.0442)\n<\/code><\/pre>\n\n<p>Using <code>squeeze()<\/code> changes your tensor shape from <code>[75, 1, 55]<\/code> to <code>[75, 55]<\/code> so it that output and target shape matches! <\/p>\n\n<p>You can also use other methods to reshape your tensor, it is just important that you have the shape of <code>[sequence_length, number_of_classes]<\/code> instead of <code>[sequence_length, 1, number_of_classes]<\/code>.<\/p>\n\n<p>Your targets should be a <code>LongTensor<\/code> resp. a tensor of type <code>torch.long<\/code> containing the classes. Shape here is <code>[sequence_length]<\/code>.<\/p>\n\n<p><strong>Edit:<br><\/strong>\n<em>Shapes from above example when passing to cross-entropy function:<\/em><\/p>\n\n<p>Outputs: <code>torch.Size([75, 55])<\/code><br>\nTargets: <code>torch.Size([75])<\/code><\/p>\n\n<hr>\n\n<p>Here is a more general example what outputs and targets should look like for CE. In this case we assume we have 5 different target classes, there are three examples for sequences of length 1, 2 and 3:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># init CE Loss function\ncriterion = nn.CrossEntropyLoss()\n\n# sequence of length 1\noutput = torch.rand(1, 5)\n# in this case the 1th class is our target, index of 1th class is 0\ntarget = torch.LongTensor([0])\nloss = criterion(output, target)\nprint('Sequence of length 1:')\nprint('Output:', output, 'shape:', output.shape)\nprint('Target:', target, 'shape:', target.shape)\nprint('Loss:', loss)\n\n# sequence of length 2\noutput = torch.rand(2, 5)\n# targets are here 1th class for the first element and 2th class for the second element\ntarget = torch.LongTensor([0, 1])\nloss = criterion(output, target)\nprint('\\nSequence of length 2:')\nprint('Output:', output, 'shape:', output.shape)\nprint('Target:', target, 'shape:', target.shape)\nprint('Loss:', loss)\n\n# sequence of length 3\noutput = torch.rand(3, 5)\n# targets here 1th class, 2th class and 2th class again for the last element of the sequence\ntarget = torch.LongTensor([0, 1, 1])\nloss = criterion(output, target)\nprint('\\nSequence of length 3:')\nprint('Output:', output, 'shape:', output.shape)\nprint('Target:', target, 'shape:', target.shape)\nprint('Loss:', loss)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Sequence of length 1:\nOutput: tensor([[ 0.1956,  0.0395,  0.6564,  0.4000,  0.2875]]) shape: torch.Size([1, 5])\nTarget: tensor([ 0]) shape: torch.Size([1])\nLoss: tensor(1.7516)\n\nSequence of length 2:\nOutput: tensor([[ 0.9905,  0.2267,  0.7583,  0.4865,  0.3220],\n        [ 0.8073,  0.1803,  0.5290,  0.3179,  0.2746]]) shape: torch.Size([2, 5])\nTarget: tensor([ 0,  1]) shape: torch.Size([2])\nLoss: tensor(1.5469)\n\nSequence of length 3:\nOutput: tensor([[ 0.8497,  0.2728,  0.3329,  0.2278,  0.1459],\n        [ 0.4899,  0.2487,  0.4730,  0.9970,  0.1350],\n        [ 0.0869,  0.9306,  0.1526,  0.2206,  0.6328]]) shape: torch.Size([3, 5])\nTarget: tensor([ 0,  1,  1]) shape: torch.Size([3])\nLoss: tensor(1.3918)\n<\/code><\/pre>\n\n<p>I hope this helps!<\/p>\n"}]}
+{"QuestionId":53456238,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-24T07:48:37.203","AcceptedAnswerId":53456428.0,"OwnerUserId":10698302.0,"Title":"How to derive weights for bucketized_column in tf.estimator.LinearRegressor in tensorflow?","Body":"<p>I am studying Google Crash ML cause.\nI have trouble in chapter \u201cFeature Cross\u201d.<\/p>\n\n<p><a href=\"https:\/\/developers.google.com\/machine-learning\/crash-course\/feature-crosses\/programming-exercise\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/machine-learning\/crash-course\/feature-crosses\/programming-exercise<\/a><\/p>\n\n<p>I tried to get the weight of cross feature from <code>linear_regressor<\/code>. <\/p>\n\n<pre><code># here I change _ to linear_model \nlinear_model = train_model(\n               learning_rate=1.0,\n               steps=500,\n               batch_size=100,\n               feature_columns=construct_feature_columns(),\n               training_examples=training_examples,\n               training_targets=training_targets,\n               validation_examples=validation_examples,\n               validation_targets=validation_targets)\n\nWeight_bucketized_longitude= linear_model.get_variable_value('linear\/linear_model\/bucketized_longitude\/weights')   \nprint(Weight_bucketized_longitude)\n<\/code><\/pre>\n\n<p>However, I got error message as below:<\/p>\n\n<p>Error Message:<\/p>\n\n<blockquote>\n  <p>NotFoundError: Key linear\/linear_model\/bucketized_longitude\/weights\n  not found in checkpoint<\/p>\n<\/blockquote>\n\n<p>It looks like the path is wrong.\nThe path works for <code>numeric_column<\/code>, but it doesn\u2019t for <code>bucketized_column<\/code>. <\/p>\n\n<p>Could you help to indicate the correct path?\nThanks.<\/p>\n\n<h3>#<\/h3>\n\n<p>I tried Geeocode's method.\nHowever, I still got error message.<\/p>\n\n<pre><code>Weight_bucketized_longitude= linear_model.get_variable_value('linear\/linear_model\/bucketized_longitude\/weights')   \n<\/code><\/pre>\n\n<blockquote>\n  <p>AttributeErrorTraceback (most recent call last)\n   in ()\n  ----> 1 Weight_bucketized_longitude= >linear_model.get_variable_value([\"linear\", \"linear_model\", >\"bucketized_longitude\", \"weights\"])<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist->packages\/tensorflow\/python\/estimator\/estimator.pyc in >get_variable_value(self, name)\n     252     _check_checkpoint_available(self.model_dir)\n     253     with context.graph_mode():\n  --> 254       return training.load_variable(self.model_dir, name)\n     255 \n     256   def get_variable_names(self):<\/p>\n  \n  <p>\/usr\/local\/lib\/python2.7\/dist->packages\/tensorflow\/python\/training\/checkpoint_utils.pyc in >load_variable(ckpt_dir_or_file, name)\n      77   \"\"\"\n      78   # TODO(b\/29227106): Fix this in the right place and remove >this.\n  ---> 79   if name.endswith(\":0\"):\n      80     name = name[:-2]\n      81   reader = load_checkpoint(ckpt_dir_or_file)<\/p>\n  \n  <p>AttributeError: 'list' object has no attribute 'endswith'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53456428","CreationDate":"2018-11-24T08:20:39.820","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>The problem is that <code>linear_model.get_variable_value()<\/code> have to pass a list of string with variables' name. From the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/LinearRegressor\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>get_variable_value<\/strong> <\/p>\n  \n  <p><code>get_variable_value(name)<\/code> <\/p>\n  \n  <p>Returns value of the variable given by name.<\/p>\n  \n  <p>Args: name: <strong>string or a list of string, name of the tensor<\/strong>. Returns:\n  Numpy array - value of the tensor.<\/p>\n  \n  <p>Raises: ValueError: If the Estimator has not produced a checkpoint\n  yet.<\/p>\n<\/blockquote>\n\n<p>Thus your code should changes as follow:<\/p>\n\n<pre><code>Weight_bucketized_longitude= linear_model.get_variable_value([\"linear\", \"linear_model\", \"bucketized_longitude\", \"weights\"])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53456413,"AnswerCount":1,"Tags":"<keras><deep-learning><tensorflow-datasets>","CreationDate":"2018-11-24T08:17:23.723","AcceptedAnswerId":null,"OwnerUserId":9395856.0,"Title":"deep learning data preparation","Body":"<p>I have a text dataset, that contains 6 classes. for each sample, I have the percent value and sum of the 6 percent values is 100% (features are related to each other). For example :<\/p>\n\n<pre><code>{A:16, B:35, C:7, D:0, E:3, F:40}\n<\/code><\/pre>\n\n<p>how can I feed a deep learning algorithm with this dataset?\nI actually want the prediction to be exactly in the shape of training data.<\/p>\n","answers":[{"AnswerId":"53456558","CreationDate":"2018-11-24T08:41:15.847","ParentId":null,"OwnerUserId":"1462770","Title":null,"Body":"<p>Here is what you can do:<\/p>\n\n<ol>\n<li>First of all, <strong>normalize<\/strong> all of your labels and scale them between 0-1.<\/li>\n<li>Use a <code>softmax<\/code> layer for prediction.<\/li>\n<\/ol>\n\n<p>Here is some code in <code>Keras<\/code> for intuition:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(100, input_dim = x.shape[1], activation='relu'))\nmodel.add(Dense(y.shape[1], activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam')\n<\/code><\/pre>\n"}]}
+{"QuestionId":53456668,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-24T08:59:34.807","AcceptedAnswerId":null,"OwnerUserId":8764887.0,"Title":"Better way to shuffle patches for image dataset- tf.data input pipeline","Body":"<p>I have a image dataset of 4644 color images which i reshape to patches of size 50 x 50 and pass to my deep neural network.<\/p>\n\n<p>The total no of patches which gets generated are 369765. I am using tf.data input pipeline for patches_generation. <\/p>\n\n<p>My question is how to efficiently shuffle the patches before passing to network. <\/p>\n\n<p>Is buffer size = 10000 in shuffle operation sufficient enough before passing to network or is there any other efficient way to shuffle among 369765 patches?<\/p>\n\n<p>Steps that i followed:\n1. Created a single tf-record which stores all 4644 images.\n2. Use tf.data pipeline to decode each image and create patches from it.\n3. Shuffle every 10000 patches and pass to network.<\/p>\n\n<p>This is the code that i am using:\n I am using buffer_size=10000, parallel_calls=4<\/p>\n\n<pre><code>dataset = (tf.data.TFRecordDataset( tfrecords_filename_image )\n           .repeat( no_epochs )\n           .map( read_and_decode, num_parallel_calls=num_parallel_calls )  \n\n           .map( get_patches_fn, num_parallel_calls=num_parallel_calls )  \n\n            .apply( tf.data.experimental.unbatch())  # unbatch the patches we just produced\n\n           .shuffle( buffer_size=buffer_size, seed=random_number_1 )  \n           .batch( batch_size )  \n           .prefetch( 1 )  \n           )\nget_patches_function definition:\n\nget_patches_fn = lambda image: get_patches( image, patch_size=patch_size )\n\ndef get_patches( image, patch_size=16 ):\n    # Function to compute patches for given image\n    # Input- image - Image which has to be converted to patches\n    #        patch_size- size of each patch\n    # Output-patches of image(4d Tensor)\n    # with tf.device('\/cpu:0'):\n    pad = [ [ 0, 0 ], [ 0, 0 ] ]\n    patches_image = tf.space_to_batch_nd( [ image ], [ patch_size, patch_size ], pad )\n    patches_image = tf.split( patches_image, patch_size * patch_size, 0 )\n    patches_image = tf.stack( patches_image, 3 )\n    patches_image = tf.reshape( patches_image, [ -1, patch_size, patch_size, 3 ] )\n   )\n    return patches_image\n\nread and decode function definition:\n\ndef read_and_decode( tf_record_file ):\n    # Function to read the tensorflow record and return image suitable for patching\n    #  Input: tf_record_file - tf record file in which image can be extracted\n    #  Output: Image\n\n    features = {\n        'height': tf.FixedLenFeature( [ ], tf.int64 ),\n        'width': tf.FixedLenFeature( [ ], tf.int64 ),\n        'image_raw': tf.FixedLenFeature( [ ], tf.string )\n    }\n    parsed = tf.parse_single_example( tf_record_file, features )\n    image = tf.decode_raw( parsed[ 'image_raw' ], tf.uint8 )\n    height = tf.cast( parsed[ 'height' ], tf.int32 )\n    width = tf.cast( parsed[ 'width' ], tf.int32 )\n    image_shape = tf.stack( [ height, width, -1 ] )\n    image = tf.reshape( image, image_shape )\n    image = image[ :, :, :3 ]\n    image = tf.cast( image, tf.float32 )\n\n    return image\n<\/code><\/pre>\n\n<p>Please also suggest if it's better to create separate tf-records for each images rather than a single tf-record for all images .<\/p>\n\n<p>Thanks in Advance.<\/p>\n","answers":[{"AnswerId":"53600513","CreationDate":"2018-12-03T19:23:59.770","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>A single tf-record file for all images is probably good enough given the number of images you have. If you have multiple disks you can try to split the file into one file per disk for higher throughput, but I don't think this should substantially slow a pipeline with the size of yours.<\/p>\n\n<p>Re the shuffle buffer size, that's an empirical question. A shuffle buffer as big as the dataset will give you true IID sampling; a smaller shuffle buffer will approximate it. Usually more randomness is better, but up to a point, so I recommend trying out a few different buffer sizes (assuming you can't have a buffer which fits the entire dataset) and see what works for you.<\/p>\n"}]}
+{"QuestionId":53456689,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network><rnn>","CreationDate":"2018-11-24T09:03:12.917","AcceptedAnswerId":null,"OwnerUserId":5539313.0,"Title":"InvalidArgumentError while training long short term memory network with Python\/Keras (Audio classification)","Body":"<p>I am new to Keras and I am trying to build a recurrent neural network to classify Audio files.<\/p>\n\n<p>During the training, I am receiving an <code>InvalidArgumentError: indices[28,0] = -711 is not in [0, 20000)<\/code>.<\/p>\n\n<p>I have found various topics talking about this error but, to be honest, I did not understand what I have to change in the parameters I am passing to the network in order to help it managing the negative values I have in the training array.<\/p>\n\n<p>Below the code:   <\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Embedding, LSTM, Bidirectional\n\nmax_features = 20000\nmaxlen = 40\nbatch_size = 32\n\nmodel = Sequential()\nmodel.add(Embedding(max_features, 128, input_length=maxlen))\nmodel.add(Bidirectional(LSTM(64)))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\n\n# try using different optimizers and different optimizer configs\nmodel.compile('adam', 'binary_crossentropy', metrics=['accuracy'])\n\nprint('Train...')\nmodel.fit(X_train, y_train,\n          batch_size=batch_size,\n          epochs=4,\n          validation_data=[X_test, y_test])\n<\/code><\/pre>\n\n<p>Error below:<\/p>\n\n<pre><code>Train...\nTrain on 964 samples, validate on 476 samples\nEpoch 1\/4\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-22-3452d23cb8b5&gt; in &lt;module&gt;()\n     12           batch_size=batch_size,\n     13           epochs=4,\n---&gt; 14           validation_data=[X_test, y_test])\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1037                                         initial_epoch=initial_epoch,\n   1038                                         steps_per_epoch=steps_per_epoch,\n-&gt; 1039                                         validation_steps=validation_steps)\n   1040 \n   1041     def evaluate(self, x=None, y=None,\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training_arrays.py in fit_loop(model, f, ins, out_labels, batch_size, epochs, verbose, callbacks, val_f, val_ins, shuffle, callback_metrics, initial_epoch, steps_per_epoch, validation_steps)\n    197                     ins_batch[i] = ins_batch[i].toarray()\n    198 \n--&gt; 199                 outs = f(ins_batch)\n    200                 outs = to_list(outs)\n    201                 for l, o in zip(out_labels, outs):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/backend\/tensorflow_backend.py in __call__(self, inputs)\n   2713                 return self._legacy_call(inputs)\n   2714 \n-&gt; 2715             return self._call(inputs)\n   2716         else:\n   2717             if py_any(is_tensor(x) for x in inputs):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/backend\/tensorflow_backend.py in _call(self, inputs)\n   2673             fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)\n   2674         else:\n-&gt; 2675             fetched = self._callable_fn(*array_vals)\n   2676         return fetched[:len(self.outputs)]\n   2677 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1437           ret = tf_session.TF_SessionRunCallable(\n   1438               self._session._session, self._handle, args, status,\n-&gt; 1439               run_metadata_ptr)\n   1440         if run_metadata:\n   1441           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)\n    526             None, None,\n    527             compat.as_text(c_api.TF_Message(self.status.status)),\n--&gt; 528             c_api.TF_GetCode(self.status.status))\n    529     # Delete the underlying status object from memory otherwise it stays alive\n    530     # as there is a reference to status from this from the traceback due to\n\nInvalidArgumentError: indices[28,0] = -711 is not in [0, 20000)\n     [[{{node embedding_3\/embedding_lookup}} = GatherV2[Taxis=DT_INT32, Tindices=DT_INT32, Tparams=DT_FLOAT, _class=[\"loc:@training_1\/Adam\/Assign_2\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](embedding_3\/embeddings\/read, embedding_3\/Cast, training_1\/Adam\/gradients\/embedding_3\/embedding_lookup_grad\/concat\/axis)]]\n<\/code><\/pre>\n\n<p><strong>EDIT1<\/strong>: X_train is the float32 array below<\/p>\n\n<pre><code>array([[-5.79938449e+02,  6.63875936e+01, -6.75054944e+00, ...,\n        -2.89458464e+00, -2.30009868e+00, -2.34216322e+00],\n       [-3.38924973e+02,  1.60668197e+01, -5.39871140e+01, ...,\n         1.27180395e+00,  4.28090614e+00,  2.01538667e+00],\n       [-5.53199739e+02,  3.45314936e+01, -1.68711443e+01, ...,\n        -9.47345310e-02, -1.04780706e-02,  1.69060756e-01],\n       ...,\n       [-5.91902354e+02,  6.14329122e+01,  1.43761675e+00, ...,\n        -4.38644438e+00, -3.67977820e+00, -1.89899207e+00],\n       [-7.04889969e+02,  6.24931510e+01,  1.90338300e+01, ...,\n        -1.47540089e+00, -1.75498741e+00, -4.55713837e-01],\n       [-8.24296641e+02,  7.43124586e+01,  1.43319513e+01, ...,\n        -7.60749297e-01, -1.05324700e+00, -8.54044186e-01]])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53456742,"AnswerCount":0,"Tags":"<leveldb><caffe2>","CreationDate":"2018-11-24T09:12:04.510","AcceptedAnswerId":null,"OwnerUserId":10697137.0,"Title":"How can I get the size of a leveldb dataset using Python?","Body":"<p>is there a function or an efficient way to get the size of a leveldb dataset using python? I am working with big datasets which contain more than 400,000 elements.<\/p>\n\n<p>I am using Caffe2 for deep learning.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53456874,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-24T09:33:03.930","AcceptedAnswerId":54192123.0,"OwnerUserId":6453802.0,"Title":"ImportError: cannot import name 'convert_kernel'","Body":"<p>When i try to use tensorflow to train model, i get this error message. <\/p>\n\n<p>File \"\/Users\/ABC\/anaconda3\/lib\/python3.6\/site-packages\/keras\/utils\/layer_utils.py\", line 7, in \n    from .conv_utils import convert_kernel<\/p>\n\n<p>ImportError: cannot import name 'convert_kernel'<\/p>\n\n<p>i have already install Keras<\/p>\n","answers":[{"AnswerId":"54192123","CreationDate":"2019-01-15T02:54:36.143","ParentId":null,"OwnerUserId":"10914880","Title":null,"Body":"<p>I got the same issue. The filename of my python code was \"tensorflow.py\". After I changed the name to \"test.py\". The issue was resolved.<\/p>\n\n<p>I guess there is already a \"tensorflow.py\" in the tensorflow package. If anyone uses the same name, it may lead to the conflict.<\/p>\n\n<p>If your python code is also called \"tensorflow.py\", you can try to use other names and see if it helps.<\/p>\n"}]}
+{"QuestionId":53458051,"AnswerCount":0,"Tags":"<machine-learning><keras><neural-network><deep-learning>","CreationDate":"2018-11-24T12:15:17.350","AcceptedAnswerId":null,"OwnerUserId":10594993.0,"Title":"How to prevent Keras predict_generator from shuffling data?","Body":"<p>I created a deep learning model, and I want to check the performance of the model by using <code>predict_generator<\/code>. I am using the following code which compares the images' labels with the predicted classes and then returns the prediction error. <\/p>\n\n<pre><code>validation_generator = validation_datagen.flow_from_directory(\n            validation_dir,\n            target_size=(image_size, image_size),\n            batch_size=val_batchsize,\n            class_mode='categorical',\n            shuffle=False)\n\n    # Get the filenames from the generator\n    fnames = validation_generator.filenames\n\n    # Get the ground truth from generator\n    ground_truth = validation_generator.classes\n\n    # Get the label to class mapping from the generator\n    label2index = validation_generator.class_indices\n\n    # Getting the mapping from class index to class label\n    idx2label = dict((v,k) for k,v in label2index.items())\n\n    # Get the predictions from the model using the generator\n    predictions = model.predict_generator(validation_generator, steps=validation_generator.samples\/validation_generator.batch_size,verbose=1)\n    predicted_classes = np.argmax(predictions,axis=1)\n\n    errors = np.where(predicted_classes != ground_truth)[0]\n    print(\"No of errors = {}\/{}\".format(len(errors),validation_generator.samples))\n\n    # Show the errors\n    for i in range(len(errors)):\n        pred_class = np.argmax(predictions[errors[i]])\n        pred_label = idx2label[pred_class]\n\n        title = 'Original label:{}, Prediction :{}, confidence : {:.3f}'.format(\n            fnames[errors[i]].split('\/')[0],\n            pred_label,\n            predictions[errors[i]][pred_class])\n\n        original = load_img('{}\/{}'.format(validation_dir,fnames[errors[i]]))\n        plt.figure(figsize=[7,7])\n        plt.axis('off')\n        plt.title(title)\n        plt.imshow(original)\n        plt.show()\n<\/code><\/pre>\n\n<p><code>validation_generator.classes<\/code> is arranged but <code>predicted_classes<\/code> is not arranged. <\/p>\n\n<p>I take the code from here <a href=\"https:\/\/www.learnopencv.com\/keras-tutorial-fine-tuning-using-pre-trained-models\/\" rel=\"nofollow noreferrer\">https:\/\/www.learnopencv.com\/keras-tutorial-fine-tuning-using-pre-trained-models\/<\/a><\/p>\n\n<p>How can I prevent <code>predict_generator<\/code> from shuffling data?<\/p>\n","answers":[]}
+{"QuestionId":53458205,"AnswerCount":0,"Tags":"<r><keras><infopath2010>","CreationDate":"2018-11-24T12:35:32.770","AcceptedAnswerId":null,"OwnerUserId":9721085.0,"Title":"install_keras() Error: Keras installation failed (no conda binary found)","Body":"<p>I am facing error in this command though I have python3 x in my system and have anaconda as well. I tried all the option given in \n<a href=\"https:\/\/stackoverflow.com\/questions\/52372425\/error-unable-to-find-conda-binary-is-anaconda-installed-reticulate-rstudio\/53434444#53434444\">Error: Unable to find conda binary. Is Anaconda installed? reticulate Rstudio<\/a>\n<a href=\"https:\/\/stackoverflow.com\/posts\/53434444\/revisions\">https:\/\/stackoverflow.com\/posts\/53434444\/revisions<\/a>, but nothing working for me. \nIn fact, I am getting error for H2o as well.  <\/p>\n\n<p>Warning in install.packages :<\/p>\n\n<blockquote>\n  <p><strong>unable to move temporary installation<\/strong> \u2018C:\\Users\\kumaashi\\Documents\\R\\win-library\\3.5\\file124f052421e1\\h2o\u2019 to \u2018C:\\Users\\kumaashi\\Documents\\R\\win-library\\3.5\\h2o\u2019<\/p>\n<\/blockquote>\n\n<p>I got this error earlier also but used to get fixed with <code>trace(utils:::unpackPkgZip, edit=TRUE)<\/code> by changing <code>sys.time<\/code> - 2.5 but <strong>since i updated R<\/strong> that option is not coming at all<\/p>\n","answers":[]}
+{"QuestionId":53458266,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><image-processing><keras><conv-neural-network>","CreationDate":"2018-11-24T12:43:33.937","AcceptedAnswerId":53459334.0,"OwnerUserId":10573543.0,"Title":"Type error for merge() function in Keras for 2d convolutional layer","Body":"<p>I am trying to recreate the Inception model version 4. But i want to train it on my image data set standard shape <code>(224,224,3)<\/code> ,so i am not taking in any pretrained weights.\nBut I am getting an error like this.<\/p>\n\n<pre><code>x = merge([x1, x2], mode='concat', concat_axis=channel_axis)\nTypeError: 'module' object is not callable\n<\/code><\/pre>\n\n<p>Here is the code:<\/p>\n\n<pre><code>def inception_stem(input):\n    if K.image_dim_ordering() == \"th\":\n        channel_axis = 1\n    else:\n        channel_axis = -1\n\n    # Input Shape is 299 x 299 x 3 (th) or 3 x 299 x 299 (th)\n    x = conv_block(input, 32, 3, 3, subsample=(2, 2), border_mode='valid')\n    x = conv_block(x, 32, 3, 3, border_mode='valid')\n    x = conv_block(x, 64, 3, 3)\n\n    x1 = MaxPooling2D((3, 3), strides=(2, 2), border_mode='valid')(x)\n    x2 = conv_block(x, 96, 3, 3, subsample=(2, 2), border_mode='valid')\n    x = tf.concat([x1,x2],axis=channel_axis)\n    #x = merge([x1, x2], mode='concat', concat_axis=channel_axis) #here is the error occuring try find out the reason behind it\n\n    x1 = conv_block(x, 64, 1, 1)\n    x1 = conv_block(x1, 96, 3, 3, border_mode='valid')\n\n    x2 = conv_block(x, 64, 1, 1)\n    x2 = conv_block(x2, 64, 1, 7)\n    x2 = conv_block(x2, 64, 7, 1)\n    x2 = conv_block(x2, 96, 3, 3, border_mode='valid')\n\n    x = merge([x1, x2], mode='concat', concat_axis=channel_axis)\n\n    x1 = conv_block(x, 192, 3, 3, subsample=(2, 2), border_mode='valid')\n    x2 = MaxPooling2D((3, 3), strides=(2, 2), border_mode='valid')(x)\n\n    x = merge([x1, x2], mode='concat', concat_axis=channel_axis)\n    return x\n<\/code><\/pre>\n\n<p>I am using <code>python 3.6<\/code>,<code>keras 2.2.2<\/code> , <code>tensorflow-gpu 1.9.0<\/code>.<\/p>\n\n<p><em>I followed the GitHub for the issue, but the answers were not clear and exact.<\/em>\nCan anyone find the solution.<\/p>\n","answers":[{"AnswerId":"53459334","CreationDate":"2018-11-24T14:48:56.803","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>Use the concatenate layer, that should help you<\/p>\n\n<pre><code>from tensorflow.python.keras.layers import concatenate\nx = concatenate([x1, x2], axis=channel_axis)\nreturn x\n<\/code><\/pre>\n"}]}
+{"QuestionId":53458379,"AnswerCount":1,"Tags":"<keras><neural-network><deep-learning><conv-neural-network><python-chess>","CreationDate":"2018-11-24T12:57:08.613","AcceptedAnswerId":null,"OwnerUserId":10500621.0,"Title":"Problem with dimensions of input array in Keras (TF) with CNN to play chess","Body":"<p>I want to feed a numpy array into a CNN, that contains 2 chess positions, one before a move and the second after a certain move. I want to train the CNN to estimate the evaluation of that move by a conventional chess program. These evaluations are int values.<\/p>\n\n<p>The shapes of the <code>x<\/code> and <code>y<\/code> are: <code>x: (2000000, 8, 8, 2) , y: (2000000,)<\/code><\/p>\n\n<p>The model code:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>#define model\nmodel = Sequential()\n#model.add(Dense(1024, activation='relu', input_dim=864))\n\nmodel.add(Conv2D(128, kernel_size=(3, 3), strides=(1, 1), activation='relu', input_shape=(8,8,2)))\nmodel.add(Conv2D(128, kernel_size=(3, 3), strides=(1, 1), activation='relu'))\n\nmodel.add(Dense(128, activation='relu', init='uniform'))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(1))\n\nmodel.compile(loss='mean_squared_error', optimizer='adam',metrics=['mae'])\nprint(model.summary())\n<\/code><\/pre>\n\n<p>The training is done with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>history = model.fit(x, y, validation_split=0.1, epochs=5, batch_size=20000, verbose=2)\n<\/code><\/pre>\n\n<p>It gives me the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-7-619de3f1be1b&gt; in &lt;module&gt;()\n    171         for i in range(5):\n    172             print(\"Fitting begins\", x.shape, y.shape)\n--&gt; 173             history = model.fit(x, y, validation_split=0.1, epochs=5, batch_size=20000, verbose=2)\n    174             #score = model.evaluate(x, y, verbose=2)\n    175             #print(score)\n\n\/usr\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n    950             sample_weight=sample_weight,\n    951             class_weight=class_weight,\n--&gt; 952             batch_size=batch_size)\n    953         # Prepare validation data.\n    954         do_validation = False\n\n\/usr\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    787                 feed_output_shapes,\n    788                 check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 789                 exception_prefix='target')\n    790 \n    791             # Generate sample-wise weight values given the `sample_weight` and\n\n\/usr\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    126                         ': expected ' + names[i] + ' to have ' +\n    127                         str(len(shape)) + ' dimensions, but got array '\n--&gt; 128                         'with shape ' + str(data_shape))\n    129                 if not check_batch_axis:\n    130                     data_shape = data_shape[1:]\n\nValueError: Error when checking target: expected dense_10 to have 4 dimensions, but got array with shape (2000000, 1)\n<\/code><\/pre>\n\n<p>What do I do wrong? How can I fix this?<\/p>\n\n<hr>\n\n<p>OK, I realized that the problem is related to the output shape of the last layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_13 (Conv2D)           (None, 6, 6, 128)         2432      \n_________________________________________________________________\nconv2d_14 (Conv2D)           (None, 4, 4, 128)         147584    \n_________________________________________________________________\ndense_14 (Dense)             (None, 4, 4, 128)         16512     \n_________________________________________________________________\nbatch_normalization_7 (Batch (None, 4, 4, 128)         512       \n_________________________________________________________________\ndense_15 (Dense)             (None, 4, 4, 1)           129       \n=================================================================\n<\/code><\/pre>\n\n<p>But why is it <code>(None, 4, 4, 1)<\/code>? Shouldn't it be <code>(None, 1)<\/code>? It's a single neuron with 1 value!<\/p>\n","answers":[{"AnswerId":"53518810","CreationDate":"2018-11-28T11:49:50.247","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<blockquote>\n  <p>But why is it (None, 4, 4, 1)? Shouldn't it be (None, 1) ???<\/p>\n<\/blockquote>\n\n<p>No, it shouldn't. Because the <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">Dense layer is applied on the last axis of its input<\/a>, and therefore since in this case it is applied on the output of <code>Conv2D<\/code> layer, which is is a 4D tensor, the output of the <code>Dense<\/code> layer would also be 4D tensor. To resolve this, you can first flatten the output of <code>Conv2D<\/code> layer using a <code>Flatten<\/code> layer and then use the <code>Dense<\/code> layer, like this:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add(Conv2D(128, kernel_size=(3, 3), strides=(1, 1), activation='relu'))\nmodel.add(Flatten())  # flatten the output of `Conv2D` to a 2D tensor\nmodel.add(Dense(128, activation='relu', init='uniform'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53458522,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-24T13:15:07.217","AcceptedAnswerId":null,"OwnerUserId":2677667.0,"Title":"How to save parameters on local server when using remote grpc session on Tensorflow?","Body":"<p>I first start a grpc server in server A. <\/p>\n\n<pre><code>server = tf.train.Server.create_local_server()\nserver.join()\n<\/code><\/pre>\n\n<p>Then I execute the training process on server B:<\/p>\n\n<pre><code>sess = tf.Session(\"grpc:\/\/172.31.222.83:34217\")\nsess.run(init)\n\nfor i in range(1000):\n    _, l = sess.run([train_op, loss], feed)\n\nsaver.save(sess, '.\/ckpts\/model')\n<\/code><\/pre>\n\n<p>When the training process is finished, I find the checkpoints have saved on server A. But I want server A just used as <strong>computational node<\/strong>. That is to say, I want the parameters are all saved on server B, server A is used only to compute. How can I achieve this?<\/p>\n","answers":[{"AnswerId":"53458618","CreationDate":"2018-11-24T13:27:24.677","ParentId":null,"OwnerUserId":"2882597","Title":null,"Body":"<p>Here is one possibility.<\/p>\n\n<pre><code>service TrainerService {\n    rpc Train(TrainRequest) returns (TrainResponse);\n}\n\nfunc (s *trainerServer) Train(..., req *pb.TrainRequest) (resp *pb.TrainResponse) {\n    return &amp;pb.TrainResponse{tensorTrainer.Train(req.data)}\n}\n<\/code><\/pre>\n\n<p>Here is another<\/p>\n\n<pre><code>service TrainerService {\n    rpc Train(TrainRequest) returns (TrainResponse);\n    rpc Results(ResultsRequest) returns (ResultsResponse);\n}\n\nfunc (s *trainerServer) Train(..., req *pb.TrainRequest) (resp *pb.TrainResponse) {\n    session, err := NewTrainingSession(req.data)\n    if err != nil { panic() }\n    go session.Train()\n    return &amp;pb.TrainResponse{session.id}\n}\n\nfunc (s *trainerServer) Results(..., req *pb.ResultsRequest) (resp *pb.ResultsResponse) {\n    results, err := GetResults(req.id)\n    if err != nil { panic() }\n    return &amp;pb.ResultsResponse{results}\n}\n<\/code><\/pre>\n\n<p>The client can call <code>Train<\/code> and poll <code>Results<\/code> until success. Perhaps <code>TrainResponse<\/code> returns an estimate.<\/p>\n"}]}
+{"QuestionId":53458833,"AnswerCount":1,"Tags":"<python><tensorflow><google-compute-engine><google-cloud-tpu>","CreationDate":"2018-11-24T13:52:17.213","AcceptedAnswerId":53840048.0,"OwnerUserId":2368505.0,"Title":"Memory reduction Tensorflow TPU v2\/v3 bfloat16","Body":"<p>My model is too big to get a batch >64 with the normal v2 TPU devices. On the <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting#memory-usage\" rel=\"nofollow noreferrer\">troubleshooting<\/a> site it is mentioned that upcoming tensorflow versions will have bfloat16 support. Are the newly supported tf versions 1.9-1.12 capable to use bfloat16 now and if yes, is there a limited set of optimizers I can use? I did not find any further documentation on this but saw the usage of bfloat16 in the tensor2tensor model, so I guess there must be a way.<\/p>\n\n<p>Furthermore I read that <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/deciding-tpu-version\" rel=\"nofollow noreferrer\">TPU v3 supports bigger models as well<\/a> but that the model would need minimal changes, but I don't find any documentation what needs to be changed.<\/p>\n\n<p>I'm already using <a href=\"https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/utils\/optimize.py\" rel=\"nofollow noreferrer\">Adafactor<\/a> and tried to reduce my layers, if you have any further reduction tips, that would be great too. I'm using picture matrices and word vectors (float32 as of now) as input. <\/p>\n","answers":[{"AnswerId":"53840048","CreationDate":"2018-12-18T19:47:54.060","ParentId":null,"OwnerUserId":"10355421","Title":null,"Body":"<p>You can use <code>bfloat16<\/code> with TPUs. There are two main things to do:<\/p>\n\n<ol>\n<li>Cast the input to bfloat16 in your input pipeline<\/li>\n<li>Surround your network within a bfloat16 scope and cast the outputs as F32 for further calculations. <\/li>\n<\/ol>\n\n<p>Here is a code snippet that illustrates the necessary changes: <\/p>\n\n<pre><code>def input_fn():\n\n  def dataset_parser(self, value):\n    \"\"\"Parse an ImageNet record from a serialized string Tensor.\"\"\"\n    image = self.image_preprocessing_fn(\n        image_bytes=image_bytes,\n        is_training=self.is_training,\n    )\n\n    if self.use_bfloat16:\n      image = tf.cast(image, tf.bfloat16)\n\n    return image, label\n\n\ndef resnet_model_fn(features, labels, mode, params):\n  \"\"\"The model_fn for ResNet to be used with TPUEstimator.\"\"\"\n\n  # This nested function allows us to avoid duplicating the logic which\n  # builds the network, for different values of --precision.\n  def build_network():\n    network = resnet_model.resnet_v1(\n        resnet_depth=FLAGS.resnet_depth,\n        num_classes=LABEL_CLASSES,\n        data_format=FLAGS.data_format)\n    return network(\n        inputs=features, is_training=(mode == tf.estimator.ModeKeys.TRAIN))\n\n  if FLAGS.precision == 'bfloat16':\n    with bfloat16.bfloat16_scope():\n      logits = build_network()\n    logits = tf.cast(logits, tf.float32)\n  elif FLAGS.precision == 'float32':\n    logits = build_network()\n<\/code><\/pre>\n\n<p>You can also see the second condition illustrated in <a href=\"https:\/\/github.com\/tensorflow\/tpu\/blob\/0e3cfbdfbcf321681c1ac1c387baf7a1a41d8d21\/models\/official\/resnet\/resnet_main.py#L315-L320\" rel=\"nofollow noreferrer\">this TPU model<\/a>.<\/p>\n"}]}
+{"QuestionId":53459677,"AnswerCount":1,"Tags":"<python><tensorflow><padding>","CreationDate":"2018-11-24T15:31:59.380","AcceptedAnswerId":null,"OwnerUserId":8734624.0,"Title":"Tensorflow tf.image.resize_image_with_crop_or_pad For 3D Images","Body":"<p>I want to train a convolutional neural network on 3 dimensional nifti images i.e. they have a width, height, and depth. A sample shape is (166, 256, 256). However, I have read that when implementing fully connected layers, you need to have all the images of the same size. The first dimension is one of 160,166,170, the second and third dimensions are one of 240,256,192. I want to pad all images to (170, 256, 256) so I do not lose any information, but the <\/p>\n\n<pre><code>tf.image.resize_image_with_crop_or_pad\n<\/code><\/pre>\n\n<p>function only seems to have arguments for two functions. How can I go about padding these images? <\/p>\n","answers":[{"AnswerId":"53459885","CreationDate":"2018-11-24T15:57:28.630","ParentId":null,"OwnerUserId":"868736","Title":null,"Body":"<p>i think your best bet is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad\" rel=\"nofollow noreferrer\"><code>tf.pad<\/code><\/a>, below code was not tested.<\/p>\n\n<pre><code>target_z = 170\n# x is shape of (166, 256, 256)\nzp = 170-x.get_shape().as_list()[0]\n# what if zp is negative ?\npaddings = tf.constant([[0, zp], [0, 0], [0,0]])\ntf.pad(x, paddings, \"CONSTANT\") \n<\/code><\/pre>\n"}]}
+{"QuestionId":53459764,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-24T15:43:30.923","AcceptedAnswerId":null,"OwnerUserId":10160917.0,"Title":"Problem predicting probabilities using tf.nn.softmax instead of tf.argmax","Body":"<p>I have searched for this problem and couldn't find any answer that helps me.<\/p>\n\n<p>I have trained a model and I use tf.argmax(logits, -1) in order to predict the right class (among 98 classes). This works fine and I succeed to predict the result.<\/p>\n\n<p>Then, I try to see the probabilities that led to this prediction, replacing the tf.argmax command with tf.nn.softmax. When I do this, it predicts an other class that has probability of 0,9 and the class that should predict (which is that when using argmax) has probability of 0.3. Which is wrong, because the result predicted with argmax is correct. Why when i use softmax the probabilities choose an other (not correct) class? Any ideas on what could be the issue there? Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":53459804,"AnswerCount":2,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-24T15:47:46.723","AcceptedAnswerId":null,"OwnerUserId":4324293.0,"Title":"Tensorflow dense_image_warp Failed to convert object of type <type 'list'> to Tensor","Body":"<p>Using up-to-date Keras and Tensorflow on Ubuntu 16.04 and 14.04. For the following code:<\/p>\n\n<pre><code>img2D = Input(shape=(100, 100, 3))\nrefPosX = Input(shape=(100, 100, 1))\nrefPosY = Input(shape=(100, 100, 1))\n# refImg = Input(shape=(100, 100, 3))\n\n# Passing values to depth net\ndepth_map = depth_net(dFeatures)\ncurX = tf.multiply(depth_map, refPosX)\n# curY = K.dot(depth_map, refPosY)\ncurY = tf.multiply(depth_map, refPosY)\n# dMove = concatenate([curX, curY])\ndMove = tf.concat([curX, curY], axis=3)\n\nwarped = tfc.image.dense_image_warp(img2D, dMove)\n<\/code><\/pre>\n\n<p>I'm receiving the error output:<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\nFile \"\/home\/carson\/ws\/dla\/test_network_5\/depth_and_color_nets.py\", line 89, in &lt;module&gt;\nFile \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/image\/python\/ops\/dense_image_warp.py\", line 195, in dense_image_warp\n    [batch_size, height * width, 2])\n File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 6482, in reshape\n\"Reshape\", tensor=tensor, shape=shape, name=name)\nFile \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 513, in _apply_op_helper\nraise err\nTypeError: Failed to convert object of type &lt;type 'list'&gt; to Tensor. Contents: [None, 10000, 2]. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>I'm a bit confused as I don't know where I'm going wrong here. Why would the elements not be of a supported type? At what point am I passing a list object to the function?<\/p>\n","answers":[{"AnswerId":"53479260","CreationDate":"2018-11-26T10:35:39.930","ParentId":null,"OwnerUserId":"4324293","Title":null,"Body":"<p>For those interested, you must define input as <\/p>\n\n<pre><code>Input(batch_shape=(batch_size, h, w, c))\n<\/code><\/pre>\n\n<p>The function requires batch_size. which you can generally get away without when using Keras.<\/p>\n"},{"AnswerId":"55302615","CreationDate":"2019-03-22T15:09:26.633","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>The error comes from the shape of the inputs. The function <code>tfc.image.dense_image_warp<\/code> requires inputs of shapes <code>[batch, height, width, channels]<\/code> and <code>[batch, height, width, 2]<\/code> respectively with a specified batch size. (cf <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/dense_image_warp\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/dense_image_warp<\/a>)<\/p>\n\n<p>Specifically, <code>None<\/code> values for batch size are not supported and lead to the error you have encountered.<\/p>\n"}]}
+{"QuestionId":53460026,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-estimator><finetunning>","CreationDate":"2018-11-24T16:14:49.563","AcceptedAnswerId":null,"OwnerUserId":779837.0,"Title":"How to persist non-trainable variables in tf.Estimator checkpoint?","Body":"<p>I'm trying to include a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/dense\" rel=\"nofollow noreferrer\">Dense<\/a> layer that is not trainable and initialized as an identity matrix, as part of my tensorflow Estimator.  The intuition is for this Dense layer to pass through its inputs during standard training and a fine tuning step afterward.  The catch is that I don't want these weights updated at all during the initial round, only during fine tuning.<\/p>\n\n<p>I can do several things to make these weights non-trainable, including using the trainable argument in the Dense constructor or by filtering out anything with dense in its name before passing to MomentumOptimizer.compute_gradients().<\/p>\n\n<p>But in either case (make dense non-trainable or just don't pass it to optimizer), tf will throw an error saying that it cannot find a key related to the dense layer.<\/p>\n\n<p>I understand that since on the first run, where dense is non-trainable, that it won't be persisted in the checkpoint file.  Likewise, if it's filtered out from being passed to compute_gradients, then the same issue occurs.<\/p>\n\n<p>Is there any method to just persist untrained variables, even with just their initialized values, across runs?<\/p>\n\n<blockquote>\n  <p>NotFoundError (see above for traceback): Key dense\/kernel\/Momentum not\n  found in checkpoint<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53502040","CreationDate":"2018-11-27T14:35:54.107","ParentId":null,"OwnerUserId":"779837","Title":null,"Body":"<p>I will answer my own question here because it wasn't immediately obvious to me, as tf documentation doesn't seem to make it clear.  If you want to introduce a new trainable variable, then it will need to fundamentally be a different model in later ones.  Thus in order to handle fine-tuning of existing weights, those existing weights in the new model must be resolved from the warm start settings.<\/p>\n\n<p>So train a model, conditionally not including the fine-tune layer when your Estimator's model function runs.  Train the existing model and then create another model that is separate.  Technically this just means you need to use a fresh model directory, but warm start settings should point to the model you trained beforehand.<\/p>\n\n<p>On fine-tuning runs, your model function should conditionally include the fine-tuning layers, but it should restore the weights from the previous run setting the warm start settings to look at previous model directory.<\/p>\n"}]}
+{"QuestionId":53460392,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><lstm>","CreationDate":"2018-11-24T16:55:12.130","AcceptedAnswerId":53469583.0,"OwnerUserId":6798599.0,"Title":"Keras functional api gives error \"expected ndim=3, found ndim=4\"","Body":"<p>I am trying to build a multi-input multi-output model using <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#multi-input-and-multi-output-models\" rel=\"nofollow noreferrer\">keras functional api<\/a> and I am following their code\nbut I got that error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_54: expected\n  ndim=3, found ndim=4<\/p>\n<\/blockquote>\n\n<p>I got that error during creating the lstm_out layer, here is the code:<\/p>\n\n<pre><code>def build_model(self):\n    main_input = Input(shape=(self.seq_len, 1), name='main_input')\n    #seq_len = 50, vocab_len = 1000\n    x = Embedding(output_dim=512, input_dim=self.vocab_len()+1, input_length=self.seq_len)(main_input)\n\n    # A LSTM will transform the vector sequence into a single vector,\n    # containing information about the entire sequence\n    lstm_out = LSTM(50)(x)\n    self.auxiliary_output = Dense(1, activation='sigmoid', name='aux_output')(lstm_out)\n\n    auxiliary_input = Input(shape=(self.seq_len,1), name='aux_input')\n    x = concatenate([lstm_out, auxiliary_input])\n\n    # We stack a deep densely-connected network on top\n    x = Dense(64, activation='relu')(x)\n    x = Dense(64, activation='relu')(x)\n    x = Dense(64, activation='relu')(x)\n\n    # And finally we add the main logistic regression layer\n    main_output = Dense(1, activation='sigmoid', name='main_output')(x)\n\n    self.model = Model(inputs=[main_input, auxiliary_input], outputs=[main_output, auxiliary_output])\n\n    print(self.model.summary())\n    self.model.compile(optimizer='rmsprop', loss='binary_crossentropy',\n              loss_weights=[1., 0.2])\n<\/code><\/pre>\n\n<p>I thought the problem is with the input_dim in the Embedding layer but I read in the <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">keras Embedding documentation<\/a> that (input_dim should equal size of vocabulary + 1).<\/p>\n\n<p>I do not know exactly why I got this, and what is the exact error in my input_dim, and how to solve it ?<\/p>\n","answers":[{"AnswerId":"53469583","CreationDate":"2018-11-25T16:34:13.567","ParentId":null,"OwnerUserId":"6560413","Title":null,"Body":"<p>Input shape for Embedding should be a 2D tensor with shape: (batch_size, sequence_length). In your code snippet main_input is provided which is a 3D tensor. To fix it change the following lines:<\/p>\n\n<pre><code>main_input = Input(shape=(self.seq_len, 1), name='main_input')\n&lt;...&gt;\nauxiliary_input = Input(shape=(self.seq_len,1), name='aux_input')\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>main_input = Input(shape=(self.seq_len, ), name='main_input')\n&lt;...&gt;\nauxiliary_input = Input(shape=(self.seq_len, ), name='aux_input')\n<\/code><\/pre>\n\n<p>It should solve the issue with different dimensions<\/p>\n"}]}
+{"QuestionId":53460728,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><linear-regression>","CreationDate":"2018-11-24T17:33:51.107","AcceptedAnswerId":53461334.0,"OwnerUserId":10579805.0,"Title":"How to pass input to placeholders?","Body":"<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport pylab as pl\nimport numpy as np\nimport tensorflow as tf\nimport matplotlib.patches as mpatches\nimport matplotlib.pyplot as plt\nplt.rcParams['figure.figsize'] = (20, 6)\n\ndf1 = pd.read_csv(\"TrainData.csv\")\ndf2 = pd.read_csv(\"TestData.csv\")\n\n\ntrain_data_X = np.asanyarray(df1['ENGINE SIZE'])\ntrain_data_Y = np.asanyarray(df1['CO2 EMISSIONS'])\ntest_data_X = np.asanyarray(df2['ENGINE SIZE'])\ntest_data_Y = np.asanyarray(df2['CO2 EMISSIONS'])\n\nW = tf.Variable(20.0, name= 'Weight')\nb = tf.Variable(30.0, name= 'Bias')\nX = tf.placeholder(tf.float32, name= 'Input')\nY = tf.placeholder(tf.float32, name= 'Output')\n\nY = W*X + b\n\nloss = tf.reduce_mean(tf.square(Y - train_data_Y))\noptimizer = tf.train.GradientDescentOptimizer(0.05)\ntrain = optimizer.minimize(loss)\n\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n\nloss_values = []\ntrain_data = []\nfor step in range(100):\n    _, loss_val, a_val, b_val = sess.run([train, loss, W, b], feed_dict={X:train_data_X, Y:train_data_Y})\n    loss_values.append(loss_val)\n    if step % 5 == 0:\n      print(step, loss_val, a_val, b_val)\n      train_data.append([a_val, b_val])\n\nplt.plot(loss_values, 'ro')\nplt.show()\n<\/code><\/pre>\n\n<p>I am trying to make a linear regression model to detect CO2 emission by giving size of engine as input. I am using the above code in tensorflow. \n1) When I use this code Weight and Bias remains unchanged. What is the problem in code?\n2) Also if I want engine size and milage both as inputs. what code changes should be made<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"53461334","CreationDate":"2018-11-24T18:48:00.097","ParentId":null,"OwnerUserId":"7850371","Title":null,"Body":"<p>There were few mistakes in the code which are mentioned below :<\/p>\n\n<ul>\n<li>You were using placeholder <code>Y = W*X + b<\/code>, which in later section of the code was used to feed data (<code>feed_dict={X:train_data_X, Y:train_data_Y}<\/code>). You should have used another variable for prediction (not the placeholder which you were using to feed data) and then you should have been able to calculate loss function. However, required changes have been made. Check <code>prediction= W*X + b<\/code> in the below code<\/li>\n<li>You were passing complete data in feed_dict at once (<code>feed_dict={X:train_data_X, Y:train_data_Y}<\/code>). However, you need to pass single data value at a time (<code>feed_dict={X:x, Y:y}<\/code>)<\/li>\n<\/ul>\n\n<p>Below code with the needful correction should be working fine. <\/p>\n\n<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport pylab as pl\nimport numpy as np\nimport tensorflow as tf\nimport matplotlib.patches as mpatches\nimport matplotlib.pyplot as plt\nplt.rcParams['figure.figsize'] = (20, 6)\n\ndf1 = pd.read_csv(\"TrainData.csv\")\ndf2 = pd.read_csv(\"TestData.csv\")\n\n\ntrain_data_X = np.asanyarray(df1['ENGINE SIZE'])\ntrain_data_Y = np.asanyarray(df1['CO2 EMISSIONS'])\ntest_data_X = np.asanyarray(df2['ENGINE SIZE'])\ntest_data_Y = np.asanyarray(df2['CO2 EMISSIONS'])\n\nW = tf.Variable(20.0, name= 'Weight')\nb = tf.Variable(30.0, name= 'Bias')\nX = tf.placeholder(tf.float32, name= 'Input')\nY = tf.placeholder(tf.float32, name= 'Output')\nprediction= W*X + b\n\n\nloss = tf.reduce_mean(tf.square(prediction - Y))\noptimizer = tf.train.GradientDescentOptimizer(0.05)\ntrain = optimizer.minimize(loss)\nloss_values = []\ntrain_data = []\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    for step in range(100):\n        for (x,y) in zip(train_data_X,train_data_Y):\n            _, loss_val, a_val, b_val = sess.run([train, loss, W, b], feed_dict={X:x, Y:y})\n            loss_values.append(loss_val)\n        if step % 5 == 0:\n            print(step, loss_val, a_val, b_val)\n            train_data.append([a_val, b_val])\n\nplt.plot(loss_values, 'ro')\nplt.show()\n<\/code><\/pre>\n\n<blockquote>\n  <p><strong>Note<\/strong> : Because of incorrect choice of loss function, your loss keeps on increasing with every step.<\/p>\n<\/blockquote>\n\n<p>I have mentioned a loss function below, which might work for your data. I am not sure how your data looks like, but you can give this a try if you want and let me know if this worked.<\/p>\n\n<pre><code>n_samples = train_data_X.shape[0]\nloss = tf.reduce_sum(tf.pow(prediction - Y, 2)) \/ (2 * n_samples)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Response to your second query.<\/p>\n<\/blockquote>\n\n<p>Assuming that your data has column name as <strong>MILEAGE<\/strong>, You can perform the below changes in <code>train_data_X<\/code> and <code>test_data_X<\/code>. Rest of the code will remain the same as above.<\/p>\n\n<pre><code>train_data_X = np.asanyarray(df1[['ENGINE SIZE','MILEAGE']])\ntrain_data_Y = np.asanyarray(df1['CO2 EMISSIONS'])\ntest_data_X = np.asanyarray(df2[['ENGINE SIZE','MILEAGE']])\ntest_data_Y = np.asanyarray(df2['CO2 EMISSIONS'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53461304,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-24T18:44:25.040","AcceptedAnswerId":53461463.0,"OwnerUserId":2651073.0,"Title":"How to use mean_squared_error loss in tensorflow session","Body":"<p>I am new to tensorflow<\/p>\n\n<p>In a part of a code for a tensorflow session, there is :<\/p>\n\n<pre><code> loss = tf.nn.softmax_cross_entropy_with_logits_v2(\n                    logits=net, labels=self.out_placeholder, name='cross_entropy')\n                self.loss = tf.reduce_mean(loss, name='mean_squared_error')\n<\/code><\/pre>\n\n<p>I want to use <code>mean_squared_error<\/code> loss function for this purpose. I found this loss function in tensorflow website:<\/p>\n\n<pre><code>tf.losses.mean_squared_error(\nlabels,\npredictions,\nweights=1.0,\nscope=None,\nloss_collection=tf.GraphKeys.LOSSES,\nreduction=Reduction.SUM_BY_NONZERO_WEIGHTS\n)\n<\/code><\/pre>\n\n<p>I need this loss function for a regression problem.<\/p>\n\n<p>I tried:<\/p>\n\n<pre><code>loss = tf.losses.mean_squared_error(predictions=net, labels=self.out_placeholder)\nself.loss = tf.reduce_mean(loss, name='mean_squared_error')\n<\/code><\/pre>\n\n<p>Where <code>net = tf.matmul(input_tensor, weights) + biases<\/code><\/p>\n\n<p>However, I'm not sure if it's the correct way.<\/p>\n","answers":[{"AnswerId":"53461463","CreationDate":"2018-11-24T19:03:42.730","ParentId":null,"OwnerUserId":"9280994","Title":null,"Body":"<p>First of all keep in mind that cross-entropy is mainly used for classification, while MSE is used for regression.<\/p>\n\n<p>In your case cross entropy measures the difference between two distributions (the real occurences, called labels - and your predictions)<\/p>\n\n<p>So while the first loss functions works on the result of the softmax layer (which can be seen as a probability distribution), the second one works directly on the floating point output of your network (which is no probability distribution) - therefore they cannot be simply exchanged.<\/p>\n"}]}
+{"QuestionId":53461311,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-24T18:45:45.277","AcceptedAnswerId":null,"OwnerUserId":8585655.0,"Title":"Cannot import Tensorflow v1.12","Body":"<p>I've installed Tensorflow 1.12 on Python 3.6.7 using:<\/p>\n\n<pre><code>conda install tensorflow-gpu\n<\/code><\/pre>\n\n<p>Running Windows Server 2016 with an CPU that supports AVX and AVX2. I\nget the following when importing tensorflow:<\/p>\n\n<pre><code> import tensorflow as tf\n<\/code><\/pre>\n\n<blockquote>\n  <p>ImportError                               Traceback (most recent call\n  last)\n  ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\n  in \n       57 \n  ---> 58   from tensorflow.python.pywrap_tensorflow_internal import *\n       59   from tensorflow.python.pywrap_tensorflow_internal import <strong>version<\/strong><\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\n  in \n       27             return _mod\n  ---> 28     _pywrap_tensorflow_internal = swig_import_helper()\n       29     del swig_import_helper<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\n  in swig_import_helper()\n       23             try:\n  ---> 24                 _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname,\n  description)\n       25             finally:<\/p>\n  \n  <p>~\\Anaconda3\\lib\\imp.py in load_module(name, file, filename, details)\n      242         else:\n  --> 243             return load_dynamic(name, filename, file)\n      244     elif type_ == PKG_DIRECTORY:<\/p>\n  \n  <p>~\\Anaconda3\\lib\\imp.py in load_dynamic(name, path, file)\n      342             name=name, loader=loader, origin=path)\n  --> 343         return _load(spec)\n      344 <\/p>\n  \n  <p>ImportError: DLL load failed: The specified module could not be found.<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>ImportError                               Traceback (most recent call\n  last)  in \n  ----> 1 import tensorflow as tf<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow__init__.py in \n       22 \n       23 # pylint: disable=g-bad-import-order\n  ---> 24 from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n       25 \n       26 try:<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py in\n  \n       47 import numpy as np\n       48 \n  ---> 49 from tensorflow.python import pywrap_tensorflow\n       50 \n       51 from tensorflow.python.tools import component_api_helper<\/p>\n  \n  <p>~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\n  in \n       72 for some common reasons and solutions.  Include the entire stack trace\n       73 above this error message when asking for help.\"\"\" % traceback.format_exc()\n  ---> 74   raise ImportError(msg)\n       75 \n       76 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long<\/p>\n  \n  <p>ImportError: Traceback (most recent call last):   File\n  \"C:\\Users\\tkassis\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"C:\\Users\\tkassis\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"C:\\Users\\tkassis\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File \"C:\\Users\\tkassis\\Anaconda3\\lib\\imp.py\",\n  line 243, in load_module\n      return load_dynamic(name, filename, file)   File \"C:\\Users\\tkassis\\Anaconda3\\lib\\imp.py\", line 343, in load_dynamic\n      return _load(spec) ImportError: DLL load failed: The specified module could not be found.<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":53461561,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><lstm>","CreationDate":"2018-11-24T19:17:01.543","AcceptedAnswerId":53466460.0,"OwnerUserId":6798599.0,"Title":"trying to concatenate two layers in keras with the same shape giving error in shapes matching","Body":"<p>I am trying to build a multi-input multi-output model using <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#multi-input-and-multi-output-models\" rel=\"nofollow noreferrer\">keras functional api<\/a> and I am following their code but I got that error:<\/p>\n\n<blockquote>\n  <p>ValueError: A <code>Concatenate<\/code> layer requires inputs with matching\n  shapes except for the concat axis. Got inputs shapes: [(None, 50),\n  (None, 50, 1)]<\/p>\n<\/blockquote>\n\n<p>I have skipped the Embedding layer, here is the code:<\/p>\n\n<pre><code>def build_model(self):\n    main_input = Input(shape=(self.seq_len, 1), name='main_input')\n    print(main_input.shape)\n    # seq_len = 50\n    # A LSTM will transform the vector sequence into a single vector,\n    # containing information about the entire sequence\n    lstm_out = LSTM(self.seq_len,input_shape=(self.seq_len,1) )(main_input)\n    self.auxiliary_output = Dense(1, activation='sigmoid', name='aux_output')(lstm_out)\n\n    auxiliary_input = Input(shape=(self.seq_len,1), name='aux_input')\n    print(auxiliary_input.shape)\n    x = concatenate([lstm_out, auxiliary_input])\n\n    # We stack a deep densely-connected network on top\n    x = Dense(64, activation='relu')(x)\n    x = Dense(64, activation='relu')(x)\n    x = Dense(64, activation='relu')(x)\n\n    # And finally we add the main logistic regression layer\n    main_output = Dense(1, activation='sigmoid', name='main_output')(x)\n    self.model = Model(inputs=[main_input, auxiliary_input], outputs=[main_output, auxiliary_output])\n    print(self.model.summary())\n    self.model.compile(optimizer='rmsprop', loss='binary_crossentropy',\n              loss_weights=[1., 0.2])\n<\/code><\/pre>\n\n<p>I got that error in the concatenation step, although printing the shape of both layers are (?,50,1).\nI do not know exactly why I got this, and what is the exact error in the input_shape of the first layer and why it does not give me the same shape as it should be using <code>print(main_input.shape)<\/code>, and how to solve it ?<\/p>\n\n<blockquote>\n  <p>UPDATE:<\/p>\n<\/blockquote>\n\n<p>I found a solution for the error by changing the shape of the second input layer<\/p>\n\n<pre><code>auxiliary_input = Input(shape=(self.seq_len,), name='aux_input')\n<\/code><\/pre>\n\n<p>so now they can concatenate smoothly, but still not clear to me why ?<\/p>\n","answers":[{"AnswerId":"53466460","CreationDate":"2018-11-25T10:07:23.657","ParentId":null,"OwnerUserId":"10701497","Title":null,"Body":"<p>For the second input, you specified before the bug that,<\/p>\n\n<pre><code>input_shape = (50,1)# seq_length=50\n<\/code><\/pre>\n\n<p>This means final shape is:<\/p>\n\n<pre><code>(None,50,1)\n<\/code><\/pre>\n\n<p>Now, when the first input passes through <code>LSTM<\/code> , since you didn't specify <code>return_sequences=True<\/code> it will return a tensor of shape <code>(batch_size, units)<\/code> viz. <code>(None, 50)<\/code> which you are concatenating with the above mentioned <code>(None, 50, 1)<\/code> <\/p>\n\n<p>Your error went away because you changed the input shape for the second input as <code>(50,)<\/code> so the final shape becomes <code>(None,50)<\/code> which matches with output of <code>LSTM<\/code> and hence it concatenated smoothly<\/p>\n"}]}
+{"QuestionId":53461869,"AnswerCount":2,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2018-11-24T19:57:18.447","AcceptedAnswerId":55028339.0,"OwnerUserId":5620642.0,"Title":"Pytorch: \"Model Weights not Changing\"","Body":"<p>Can someone help me understand why the weights are not updating?<\/p>\n\n<pre><code>    unet = Unet()\n    optimizer = torch.optim.Adam(unet.parameters(), lr=0.001)\n    loss_fn = torch.nn.MSELoss()\n    input =  Variable(torch.randn(32, 1, 64, 64, 64 ), requires_grad=True)\n    target = Variable(torch.randn(32, 1, 64, 64, 64), requires_grad=False)\n\n    optimizer.zero_grad()\n    y_pred = unet(input)\n    y = target[: , : , 20:44, 20:44, 20:44]\n\n    loss = loss_fn(y_pred, y)\n    print(unet.conv1.weight.data[0][0]) # weights of the first layer in the unet\n    loss.backward()\n    optimizer.step()\n    print(unet.conv1.weight.data[0][0]) # weights havent changed\n<\/code><\/pre>\n\n<p>The model is defined like:<\/p>\n\n<pre><code>class Unet(nn.Module):\n\ndef __init__(self):\n  super(Unet, self).__init__()\n\n  # Down hill1\n  self.conv1 = nn.Conv3d(1, 2, kernel_size=3,  stride=1)\n  self.conv2 = nn.Conv3d(2, 2, kernel_size=3,  stride=1)\n\n  # Down hill2\n  self.conv3 = nn.Conv3d(2, 4, kernel_size=3,  stride=1)\n  self.conv4 = nn.Conv3d(4, 4, kernel_size=3,  stride=1)\n\n  #bottom\n  self.convbottom1 = nn.Conv3d(4, 8, kernel_size=3,  stride=1)\n  self.convbottom2 = nn.Conv3d(8, 8, kernel_size=3,  stride=1)\n\n  #up hill1\n  self.upConv0 = nn.Conv3d(8, 4, kernel_size=3,  stride=1)\n  self.upConv1 = nn.Conv3d(4, 4, kernel_size=3,  stride=1)\n  self.upConv2 = nn.Conv3d(4, 2, kernel_size=3,  stride=1)\n\n  #up hill2\n  self.upConv3 = nn.Conv3d(2, 2, kernel_size=3, stride=1)\n  self.upConv4 = nn.Conv3d(2, 1, kernel_size=1, stride=1)\n\n  self.mp = nn.MaxPool3d(kernel_size=3, stride=2, padding=1)\n  # some more irrelevant properties...\n<\/code><\/pre>\n\n<p>The forward function looks like:<\/p>\n\n<pre><code>def forward(self, input):\n    # Use U-net Theory to Update the filters.\n    # Example Approach...\n    input = F.relu(self.conv1(input))\n    input = F.relu(self.conv2(input))\n\n    input = self.mp(input)\n\n    input = F.relu(self.conv3(input))\n    input = F.relu(self.conv4(input))\n\n    input = self.mp(input)\n\n    input = F.relu(self.convbottom1(input))\n    input = F.relu(self.convbottom2(input))\n\n    input = F.interpolate(input, scale_factor=2, mode='trilinear')\n\n    input = F.relu(self.upConv0(input))\n    input = F.relu(self.upConv1(input))\n\n    input = F.interpolate(input, scale_factor=2, mode='trilinear')\n\n\n    input = F.relu(self.upConv2(input))\n    input = F.relu(self.upConv3(input))\n\n    input = F.relu(self.upConv4(input))\n\n    return input\n<\/code><\/pre>\n\n<p>I have followed the approach of any example and documentation i could find and it is beyound me why that doesn't work?  <\/p>\n\n<p>I can figure out as much that <code>y_pred.grad<\/code> after the backward call is none which it shouldn't be. If we have no gradient then ofcourse the optimizer can't change the weights in any direction but why is there no gradient?<\/p>\n","answers":[{"AnswerId":"55028339","CreationDate":"2019-03-06T16:54:33.240","ParentId":null,"OwnerUserId":"5620642","Title":null,"Body":"<p>I identified this problem to be of \"The Dying ReLu Problem\" Due to the data being Hounsfield units and Pytorch uniform distribution of initial weights meant that many neurons would start out in ReLu's zero region leaving them paralyzed and dependable on other neurons to produce a gradient that could pull them out of the zero region. This unlikely to happen as training progresses all neurons gets pushed into ReLu's zero region.<\/p>\n\n<p>There are several solutions to this problem. You can use Leaky_relu or other activation functions that do not have a zero region. <\/p>\n\n<p>You can also normalize the input data using Batch Normalization and initialize the weights to only be of the positive kind. <\/p>\n\n<p>Solution number two is probably the most optimal solution since both will solve the problem but leaky_relu will prolong training whereas Batch normalization will do the opposite and increase accuracy. On the other hand, Leaky_relu is an easy fix whereas the other solution requires a little extra work. <\/p>\n\n<p>For Hounsfield data, one could also add a constant of 1000 to the input eliminating negative units from the data. This still requires a different weight initialization than Pytorch's standard initialization. <\/p>\n"},{"AnswerId":"53462514","CreationDate":"2018-11-24T21:29:18.217","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>I do not think that weights should be printed with command you use. Try <code>print(unet.conv1.state_dict()[\"weight\"])<\/code> instead of <code>print(unet.conv1.weight.data[0][0])<\/code>. <\/p>\n"}]}
+{"QuestionId":53461972,"AnswerCount":1,"Tags":"<tensorflow><gcc><archlinux>","CreationDate":"2018-11-24T20:14:27.590","AcceptedAnswerId":null,"OwnerUserId":10700187.0,"Title":"install gcc6 in archlinux","Body":"<p>The gcc version in my machine is 8.3. However, I need to use gcc6. I git cloned the source from AUR, and I run:<\/p>\n\n<pre><code>makepkg -si\n<\/code><\/pre>\n\n<p>It fails. It jumped into a look, checking everything again and again. The screen flashed so fast and I couldn't have a screenshot.\nI tried to use yay to install gcc6 from AUR, it is the same problem.<\/p>\n","answers":[{"AnswerId":"53462671","CreationDate":"2018-11-24T21:50:40.250","ParentId":null,"OwnerUserId":"10700187","Title":null,"Body":"<p>problem resolved.\nI just download the package from the archive in archlinux AUR. and pacman.\nI am not sure why I cannot use yay or makepkg -si to install it.\nTotally no idea about it.<\/p>\n"}]}
+{"QuestionId":53462087,"AnswerCount":1,"Tags":"<keras><neural-network><artificial-intelligence><autoencoder><anomaly-detection>","CreationDate":"2018-11-24T20:28:09.363","AcceptedAnswerId":null,"OwnerUserId":10436627.0,"Title":"Use internal representation of autoencoder for anomaly detection","Body":"<p>I've trained an autoencoder to recognize 'positive' time series (the network is a simple fully connected network, no recurrent layers). The problem is that from what my advisor says, I should try to detect anomalies using some statistics on the latent space (like difference between histograms of latent space between good and outlier data), but when I predict time series with outliers I get the same internal representation as with the good data. I believe this is due to the fact that my network can only reproduce the normal data. \nDo you have any hints?\nThanks <\/p>\n","answers":[{"AnswerId":"53508825","CreationDate":"2018-11-27T21:57:07.373","ParentId":null,"OwnerUserId":"296549","Title":null,"Body":"<p>I guess you are trying to detect anomaly by using the reconstruction error of your network, i.e. train on some normal timeseries, then detect on a timeseries which includes outliers. If I am further guessing right, your advisor is suggesting that the internal representation will tell you more about the nature of the anomaly, i.e. which features of the input data is most abnormal.<\/p>\n\n<p>First thing to note, is that the intermediate features your network is building with only legitimate data will not change in detection mode, thus the internal representation (the weights of each of your intermediate layer's neurons) will not change when processing a new datapoint in this mode.<\/p>\n\n<p>You will only be able to reason about the root cause of the outlier (which dimension contributes the most to the reconstruction error) if you have a good idea of which original features your intermediate features represent. This can be quite hard if you have one fully connected auto-encoder and several hidden layers, where the contribution of each feature to each neuron is increasingly interleaved with the other features. A trick is to build one auto-encoder per set of features, and use them as an ensemble for anomaly prediction. That way, each auto-encoder in the ensemble is known to be responsible for a set of features, and it makes it easier to known how each set of features contribute to the anomaly. See an example <a href=\"https:\/\/arxiv.org\/abs\/1802.09089\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53462143,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-24T20:34:38.867","AcceptedAnswerId":null,"OwnerUserId":7590932.0,"Title":"Pip3 won't update Tensorflow","Body":"<p>I have been trying to upgrade tensorflow on my Apple OS to the newest version. I used the following command<\/p>\n\n<pre><code>pip3.6 install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>and received the following message<\/p>\n\n<pre><code>Requirement already up-to-date: tensorflow in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (0.12.0)\nRequirement already satisfied, skipping upgrade: protobuf==3.1.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (from tensorflow) (3.1.0)\nRequirement already satisfied, skipping upgrade: numpy&gt;=1.11.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (from tensorflow) (1.15.4)\nRequirement already satisfied, skipping upgrade: wheel&gt;=0.26 in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (from tensorflow) (0.32.3)\nRequirement already satisfied, skipping upgrade: six&gt;=1.10.0 in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (from tensorflow) (1.11.0)\nRequirement already satisfied, skipping upgrade: setuptools in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages (from protobuf==3.1.0-&gt;tensorflow) (40.6.2)\n<\/code><\/pre>\n\n<p>I believe that 0.12.0 is not the latest version of Tensorflow? Is there a way to get pip3.6 to update install tensorflow? I have already tried uninstalling and then reinstalling - which is an issue in itself. When I installed and used the following command<\/p>\n\n<p><code>pip3.6 install --user --upgrade tensorflow<\/code><\/p>\n\n<p>I received the following error <\/p>\n\n<p><code>Could not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow<\/code><\/p>\n\n<p>I have also tried (from: <a href=\"https:\/\/stackoverflow.com\/questions\/38896424\/tensorflow-not-found-using-pip\">TensorFlow not found using pip<\/a>)<\/p>\n\n<p><code>python3.6 -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.0-py3-none-any.whl<\/code><\/p>\n\n<p>to install, which works but this installs 0.12.0. I switched the 0.12.0 with 1.10.0 in the URL and while the installation worked, I got this error when importing tensorflow <\/p>\n\n<pre><code> File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: dlopen(\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so, 6): Symbol not found: _SecKeyCopyExternalRepresentation\n  Referenced from: \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so (which was built for Mac OS X 10.12)\n  Expected in: \/System\/Library\/Frameworks\/Security.framework\/Versions\/A\/Security\n in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: dlopen(\/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so, 6): Symbol not found: _SecKeyCopyExternalRepresentation\n  Referenced from: \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so (which was built for Mac OS X 10.12)\n  Expected in: \/System\/Library\/Frameworks\/Security.framework\/Versions\/A\/Security\n in \/Library\/Frameworks\/Python.framework\/Versions\/3.6\/lib\/python3.6\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help\n<\/code><\/pre>\n\n<p>Is there any solution to this problem?<\/p>\n\n<p>Thanks<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>After seeing (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10107\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10107<\/a>), it seems that TF >= 1.10.0 doesn't work for this issue. Trying 1.9.0 worked:<\/p>\n\n<pre><code>python3.6 -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-1.9.0-py3-none-any.whl\n<\/code><\/pre>\n","answers":[{"AnswerId":"56604699","CreationDate":"2019-06-14T20:22:09.787","ParentId":null,"OwnerUserId":"7555390","Title":null,"Body":"<p>I just tried in my Mac and there is no problem at all for pip to find the latest tensorflow<\/p>\n\n<p>Here is my pip version:\npip3.7 --19.0.3<\/p>\n\n<pre><code>$ pip install --user --upgrade tensorflow\nCollecting tensorflow\n  Downloading https:\/\/files.pythonhosted.org\/packages\/d5\/1c\/3ac472009a5c54ae7ec5a3294520ca36d1908cd1e5cf3e3fd923f9b7b31f\/tensorflow-1.13.1-cp37-cp37m-macosx_10_11_x86_64.whl (73.6MB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 73.6MB 62.2MB\/s\nCollecting tensorboard&lt;1.14.0,&gt;=1.13.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/0f\/39\/bdd75b08a6fba41f098b6cb091b9e8c7a80e1b4d679a581a0ccd17b10373\/tensorboard-1.13.1-py3-none-any.whl (3.2MB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 3.2MB 38.7MB\/s\nRequirement already satisfied, skipping upgrade: termcolor&gt;=1.1.0 in \/usr\/local\/lib\/python3.7\/site-packages (from tensorflow) (1.1.0)\nCollecting protobuf&gt;=3.6.1 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/fd\/ed\/e53989e7b08274334ddb41dac51294d270b3db7b3d1e354f5b548714c7c6\/protobuf-3.8.0-cp37-cp37m-macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (1.4MB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 1.4MB 48.1MB\/s\nCollecting astor&gt;=0.6.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/d1\/4f\/950dfae467b384fc96bc6469de25d832534f6b4441033c39f914efd13418\/astor-0.8.0-py2.py3-none-any.whl\nRequirement already satisfied, skipping upgrade: six&gt;=1.10.0 in \/usr\/local\/lib\/python3.7\/site-packages (from tensorflow) (1.12.0)\nCollecting absl-py&gt;=0.1.6 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/da\/3f\/9b0355080b81b15ba6a9ffcf1f5ea39e307a2778b2f2dc8694724e8abd5b\/absl-py-0.7.1.tar.gz (99kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 102kB 21.0MB\/s\nCollecting tensorflow-estimator&lt;1.14.0rc0,&gt;=1.13.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/bb\/48\/13f49fc3fa0fdf916aa1419013bb8f2ad09674c275b4046d5ee669a46873\/tensorflow_estimator-1.13.0-py2.py3-none-any.whl (367kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 368kB 34.3MB\/s\nRequirement already satisfied, skipping upgrade: numpy&gt;=1.13.3 in \/usr\/local\/lib\/python3.7\/site-packages (from tensorflow) (1.16.1)\nCollecting keras-preprocessing&gt;=1.0.5 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/28\/6a\/8c1f62c37212d9fc441a7e26736df51ce6f0e38455816445471f10da4f0a\/Keras_Preprocessing-1.1.0-py2.py3-none-any.whl (41kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 51kB 13.4MB\/s\nCollecting grpcio&gt;=1.8.6 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/5e\/71\/92d142dcc3337dbafb6f1d936b19b94a1f65efc9bc3ddfb2315710068025\/grpcio-1.21.1-cp37-cp37m-macosx_10_9_x86_64.whl (2.0MB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 2.0MB 24.7MB\/s\nRequirement already satisfied, skipping upgrade: wheel&gt;=0.26 in .\/Python\/lib\/python3.7\/site-packages (from tensorflow) (0.33.1)\nCollecting keras-applications&gt;=1.0.6 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/71\/e3\/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f\/Keras_Applications-1.0.8-py3-none-any.whl (50kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 51kB 20.2MB\/s\nCollecting gast&gt;=0.2.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/4e\/35\/11749bf99b2d4e3cceb4d55ca22590b0d7c2c62b9de38ac4a4a7f4687421\/gast-0.2.2.tar.gz\nCollecting werkzeug&gt;=0.11.15 (from tensorboard&lt;1.14.0,&gt;=1.13.0-&gt;tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/9f\/57\/92a497e38161ce40606c27a86759c6b92dd34fcdb33f64171ec559257c02\/Werkzeug-0.15.4-py2.py3-none-any.whl (327kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 327kB 42.3MB\/s\nCollecting markdown&gt;=2.6.8 (from tensorboard&lt;1.14.0,&gt;=1.13.0-&gt;tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/c0\/4e\/fd492e91abdc2d2fcb70ef453064d980688762079397f779758e055f6575\/Markdown-3.1.1-py2.py3-none-any.whl (87kB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 92kB 22.8MB\/s\nRequirement already satisfied, skipping upgrade: setuptools in .\/Python\/lib\/python3.7\/site-packages (from protobuf&gt;=3.6.1-&gt;tensorflow) (41.0.1)\nCollecting mock&gt;=2.0.0 (from tensorflow-estimator&lt;1.14.0rc0,&gt;=1.13.0-&gt;tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/05\/d2\/f94e68be6b17f46d2c353564da56e6fb89ef09faeeff3313a046cb810ca9\/mock-3.0.5-py2.py3-none-any.whl\nCollecting h5py (from keras-applications&gt;=1.0.6-&gt;tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/80\/c5\/eec74d7324628f1b640c6e706981c4ed51afcaa1656ece26cb08d862598e\/h5py-2.9.0-cp37-cp37m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (6.2MB)\n     |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 6.2MB 43.1MB\/s\nBuilding wheels for collected packages: absl-py, gast\n  Building wheel for absl-py (setup.py) ... done\n  Stored in directory: \/Users\/xinyao\/Library\/Caches\/pip\/wheels\/ee\/98\/38\/46cbcc5a93cfea5492d19c38562691ddb23b940176c14f7b48\n  Building wheel for gast (setup.py) ... done\n  Stored in directory: \/Users\/xinyao\/Library\/Caches\/pip\/wheels\/5c\/2e\/7e\/a1d4d4fcebe6c381f378ce7743a3ced3699feb89bcfbdadadd\nSuccessfully built absl-py gast\nInstalling collected packages: grpcio, protobuf, werkzeug, absl-py, markdown, tensorboard, astor, mock, tensorflow-estimator, keras-preprocessing, h5py, keras-applications, gast, tensorflow\n  WARNING: The script markdown_py is installed in '\/Users\/xinyao\/.local\/bin' which is not on PATH.\n  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\n  WARNING: The script tensorboard is installed in '\/Users\/xinyao\/.local\/bin' which is not on PATH.\n  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\n  WARNING: The scripts freeze_graph, saved_model_cli, tensorboard, tf_upgrade_v2, tflite_convert, toco and toco_from_protos are installed in '\/Users\/xinyao\/.local\/bin' which is not on PATH.\n  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\nSuccessfully installed absl-py-0.7.1 astor-0.8.0 gast-0.2.2 grpcio-1.21.1 h5py-2.9.0 keras-applications-1.0.8 keras-preprocessing-1.1.0 markdown-3.1.1 mock-3.0.5 protobuf-3.8.0 tensorboard-1.13.1 tensorflow-1.13.1 tensorflow-estimator-1.13.0 werkzeug-0.15.4\n<\/code><\/pre>\n\n<p>Maybe you should try to upgrade your python3.6 to python3.7?<\/p>\n"}]}
+{"QuestionId":53462493,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2018-11-24T21:25:33.543","AcceptedAnswerId":53462819.0,"OwnerUserId":8171916.0,"Title":"Shape of pytorch model.parameter is inconsistent with how it's defined in the model","Body":"<p>I'm attempting to extract the weights and biases from a simple network built in PyTorch. My entire network is composed of nn.Linear layers. When I create a layer by calling <code>nn.Linear(in_dim, out_dim)<\/code>, I expect the parameters that I get from calling <code>model.parameters()<\/code> for that model to be of shape <code>(in_dim, out_dim)<\/code> for the weight and <code>(out_dim)<\/code> for the bias. However, the weights that come out of <code>model.parameters()<\/code> are instead of shape <code>(out_dim, in_dim)<\/code>.<\/p>\n\n<p>The intention of my code is to be able to use matrix multiplication to perform a forward pass using only numpy, not any PyTorch. Because of the shape inconsistency, matrix multiplications throw an error. How can I fix this?<\/p>\n\n<p>Here is my exact code:<\/p>\n\n<pre><code>class RNN(nn.Module):\n\n    def __init__(self, dim_input, dim_recurrent, dim_output):\n\n        super(RNN, self).__init__()\n\n        self.dim_input = dim_input\n        self.dim_recurrent = dim_recurrent\n        self.dim_output = dim_output\n\n        self.dense1 = nn.Linear(self.dim_input, self.dim_recurrent)\n        self.dense2 = nn.Linear(self.dim_recurrent, self.dim_recurrent, bias = False)\n        self.dense3 = nn.Linear(self.dim_input, self.dim_recurrent)\n        self.dense4 = nn.Linear(self.dim_recurrent, self.dim_recurrent, bias = False)\n        self.dense5 = nn.Linear(self.dim_recurrent, self.dim_output)\n\n#There is a defined forward pass\n\nmodel = RNN(12, 100, 6)\n\nfor i in model.parameters():\n    print(i.shape())\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>torch.Size([100, 12])\ntorch.Size([100])\ntorch.Size([100, 100])\ntorch.Size([100, 12])\ntorch.Size([100])\ntorch.Size([100, 100])\ntorch.Size([6, 100])\ntorch.Size([6])\n<\/code><\/pre>\n\n<p>The output should, if I'm correct, be:<\/p>\n\n<pre><code>torch.Size([12, 100])\ntorch.Size([100])\ntorch.Size([100, 100])\ntorch.Size([12, 100])\ntorch.Size([100])\ntorch.Size([100, 100])\ntorch.Size([100, 6])\ntorch.Size([6])\n<\/code><\/pre>\n\n<p>What is my issue?<\/p>\n","answers":[{"AnswerId":"53462819","CreationDate":"2018-11-24T22:16:05.467","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>What you see there is not the (out_dim, in_dim), it is just the shape of the weight matrix. When you call <code>print(model)<\/code> you can see that input and output features are correct:<\/p>\n\n<pre><code>RNN(\n  (dense1): Linear(in_features=12, out_features=100, bias=True)\n  (dense2): Linear(in_features=100, out_features=100, bias=False)\n  (dense3): Linear(in_features=12, out_features=100, bias=True)\n  (dense4): Linear(in_features=100, out_features=100, bias=False)\n  (dense5): Linear(in_features=100, out_features=6, bias=True)\n)\n<\/code><\/pre>\n\n<p>You can check the source code to see that the <strong>weights<\/strong> are actually <strong>transposed<\/strong> before calling <code>matmul<\/code>.<\/p>\n\n<p><br><\/p>\n\n<p><em><code>nn.Linear<\/code> is define here:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html#Linear\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html#Linear<\/a><\/em><\/p>\n\n<p>You can check the <code>forward<\/code>, it looks like this:<\/p>\n\n<pre><code>def forward(self, input):\n    return F.linear(input, self.weight, self.bias)\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p><em><code>F.linear<\/code> is define here:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/functional.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/functional.html<\/a><\/em><\/p>\n\n<p>The respective line for multiplying the weights is:<\/p>\n\n<pre><code>output = input.matmul(weight.t())\n<\/code><\/pre>\n\n<p>As mentioned above you can see that the <strong>weights<\/strong> are <strong>transposed<\/strong> before applying <code>matmul<\/code> and therefore the shape of the weights is different than you expected.\n<br><\/p>\n\n<p>So if you want to do the matrix multiplication manually, you do:<\/p>\n\n<pre><code># dummy input of length 5\ninput = torch.rand(5, 12)\n# apply layer dense1 (without bias, for bias just add + model.dense1.bias)\noutput_first_layer = input.matmul(model.dense1.weight.t())\nprint(output_first_layer.shape)\n<\/code><\/pre>\n\n<p>Just as you would expect from your <code>dense1<\/code> it returns:<\/p>\n\n<pre><code>torch.Size([5, 100])\n<\/code><\/pre>\n\n<p>I hope this explains your observations with the shape :)<\/p>\n"}]}
+{"QuestionId":53462965,"AnswerCount":0,"Tags":"<python><tensorflow><data-visualization><tensorboard><google-colaboratory>","CreationDate":"2018-11-24T22:38:43.433","AcceptedAnswerId":null,"OwnerUserId":7818052.0,"Title":"How to get Histogram from tensorboard in Google Colab","Body":"<p>I'm using the package called 'tensorboardcolab'. But by now I only can get graph from tensorboard based on the following code:<\/p>\n\n<pre><code>!pip install -U tensorboardcolab\nfrom tensorboardcolab import *\n\ntbc=TensorBoardColab()\n\ntf.reset_default_graph()\n\nwith tf.get_default_graph().as_default():\n    x = tf.placeholder(tf.float32, [], name=\"x\")\n    y = tf.placeholder(tf.float32, [], name=\"y\")\n    k = tf.get_variable(\"k\", shape=[])\n    b = tf.get_variable(\"b\", shape=[])\n    loss = tf.square(k * x + b - y, name=\"square_error\")\n\noptimizer = tf.train.GradientDescentOptimizer(0.5)\ntrain_op = optimizer.minimize(loss)\n\nwith tf.Session() as sess:\n# Initialize `k` and `b`\nsess.run(tf.global_variables_initializer())\nfor i in range(100):\n    # y = 3x-1 + noise\n    data_x = np.random.random()\n    data_y = 3.0 * data_x - 1.0 + 0.0001*np.random.randn()\n\n    k_val, b_val, loss_val, _ = sess.run([k, b, loss, train_op],\n                                         feed_dict={x: data_x, y: data_y})\n    if i % 20 == 0:\n        print(\"k, b, loss: \", k_val, b_val, loss_val)\n\nsess.close()\n\ntrain_writer = tbc.get_writer();\ntrain_writer.add_graph(sess.graph)\n\ntrain_writer.flush();\ntbc.close()\n<\/code><\/pre>\n\n<hr>\n\n<p>Can I add some code to visualize the histogram?? I didn't have any idea to find to fill in the parameters like 'write_grads = True' etc.<\/p>\n","answers":[]}
+{"QuestionId":53463156,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2018-11-24T23:08:37.507","AcceptedAnswerId":53465190.0,"OwnerUserId":470184.0,"Title":"Sparse tensors to decrease training time","Body":"<p>I've learning about PyTorch sparse tensors : <a href=\"https:\/\/pytorch.org\/docs\/stable\/sparse.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/sparse.html<\/a><\/p>\n\n<p>From the docs (<a href=\"https:\/\/pytorch.org\/docs\/stable\/sparse.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/sparse.html<\/a>) : \"Torch supports sparse tensors in COO(rdinate) format, which can efficiently store and process tensors for which the majority of elements are zeros.\"<\/p>\n\n<p>Is one of the intended use of sparse tensors instead of regular PyTorch tensors to decrease training time ?<\/p>\n","answers":[{"AnswerId":"53465190","CreationDate":"2018-11-25T06:27:29.923","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>Yes but indirectly. <\/p>\n\n<p>sparse tensors can reduce the complexity of computations and hence training\/inference time. Complexity of matrix multiplication depends on number of elements in matrix whereas complexity of sparse matrix multiplication would depend on the number of non-zero elements which are less (due to sparsity)<\/p>\n"}]}
+{"QuestionId":53463361,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-24T23:38:10.307","AcceptedAnswerId":53463971.0,"OwnerUserId":9912116.0,"Title":"Copying Tensorflow - __enter__ and __exit__ errors using context manager","Body":"<p>I tried to use with numpy.load('file.npy') as bar but it gives me an error of attributeError: _enter__ and at a different time attributeError: _exit__. I was just trying to copy what the example shown on Tensorflow's page did. I don't understand why their code would work but mine -which should be the same- wouldn't.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YWeXj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YWeXj.png\" alt=\"my attempt\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/v93eT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/v93eT.png\" alt=\"Tensorflow pic\"><\/a><\/p>\n","answers":[{"AnswerId":"53463971","CreationDate":"2018-11-25T01:41:31.227","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<pre><code>In [29]: f = open('con.npy')\nIn [30]: f.__enter__\nOut[30]: &lt;function TextIOWrapper.__enter__&gt;\nIn [31]: f.__exit__\nOut[31]: &lt;function TextIOWrapper.__exit__&gt;\nIn [32]: f.close()\nIn [33]: arr = np.load('con.npy')\n<\/code><\/pre>\n\n<p><code>arr<\/code> is a <code>ndarray<\/code> and does not have <code>__enter__<\/code> method.<\/p>\n\n<p>A <code>npz<\/code> file is a zip archive, and does have a <code>__enter__<\/code>.  It should useable in a <code>with<\/code>:<\/p>\n\n<pre><code>In [34]: ls *.npz\nMcoo.npz  Mcsr.npz  test.npz\nIn [35]: d = np.load('test.npz')\nIn [36]: d.__enter__\nOut[36]: &lt;bound method NpzFile.__enter__ of &lt;numpy.lib.npyio.NpzFile object at 0x7fc5e03f7cf8&gt;&gt;\n<\/code><\/pre>\n\n<p>SO using the <code>npz<\/code> in a with:<\/p>\n\n<pre><code>In [39]: with np.load('test.npz') as d:\n    ...:     print(list(d.keys()))\n    ...:     a = d['a']\n    ...:     \n['a', 'b']\nIn [40]: a\nOut[40]: array(1)\n<\/code><\/pre>\n\n<p>Usually we don't bother using the <code>with<\/code> for a <code>npz<\/code>, because there's no cost (that I know of) of hanging on to the dictionary like object.  <\/p>\n\n<hr>\n\n<p>In the Tensorflow example, <code>features = data['features']<\/code> is consistent with <code>data<\/code> being an opened <code>NPZ<\/code> archive.  So despite the <code>npy<\/code>, this <code>load<\/code> is an archive.  The distinction is based on data bits in the file itself (a <code>_ZIP_PREFIX<\/code>), not the filename.<\/p>\n\n<p>From the <code>load<\/code> documentation:<\/p>\n\n<pre><code>- If the file is a ``.npz`` file, the returned value supports the\n  context manager protocol in a similar fashion to the open function::\n\n    with load('foo.npz') as data:\n        a = data['a']\n<\/code><\/pre>\n"}]}
+{"QuestionId":53463532,"AnswerCount":2,"Tags":"<neural-network><gradient><pytorch><backpropagation>","CreationDate":"2018-11-25T00:08:00.200","AcceptedAnswerId":53464798.0,"OwnerUserId":1434041.0,"Title":"Auto updating custom layer parameters while backpropagating in pytorch","Body":"<p>I have a pytorch custom layer defined as:<\/p>\n\n<pre><code>class MyCustomLayer(nn.Module):\n  def __init__(self):\n    super(MyCustomLayer, self).__init__()\n\n    self.my_parameter = torch.rand(1, requires_grad = True)\n\n    # the following allows the previously defined parameter to be recognized as a network parameter when instantiating the model\n    self.my_registered_parameter = nn.ParameterList([nn.Parameter(self.my_parameter)])\n\n  def forward(self, x):\n    return x*self.my_parameter\n<\/code><\/pre>\n\n<p>I then define my network where the custom layer is used:<\/p>\n\n<pre><code>class MyNet(nn.Module):\n  def __init__(self):\n    super(MyNet, self).__init__()\n    self.layer1 = MyCustomLayer()\n\n  def forward(self, x):\n    x = self.layer1(x)\n    return x\n<\/code><\/pre>\n\n<p>Now Let's instantiate MyNet and observe the issue:<\/p>\n\n<pre><code># instantiate MyNet and run it over one input value\nmodel = MyNet()\nx = torch.tensor(torch.rand(1))\noutput = model(x)\ncriterion = nn.MSELoss()\nloss = criterion(1, output)\nloss.backward()\n<\/code><\/pre>\n\n<p>Iterating through model parameters shows <code>None<\/code> for custom layer parameter:<\/p>\n\n<pre><code>for p in model.parameters():\n    print (p.grad)\n\nNone\n<\/code><\/pre>\n\n<p>while directly accessing that parameter shows the correct <code>grad<\/code> value:<\/p>\n\n<pre><code>print(model.layer1.my_parameter.grad)\n\ntensor([-1.4370])\n<\/code><\/pre>\n\n<p>This, in turn, prevents the optim step from updating the inner parameters automatically and leaves me with the hassle of having to update those manually. Anyone knows how I can address this issue?<\/p>\n","answers":[{"AnswerId":"53464194","CreationDate":"2018-11-25T02:37:26.833","ParentId":null,"OwnerUserId":"1434041","Title":null,"Body":"<p>Alright! \nI had to switch the parameter variable calls within the custom layer to the <code>nn.ParameterList<\/code> object (i.e. <code>return x*self.my_registered_parameter[0]<\/code> instead of x*self.my_parameter ). In this example that meant changing the custom layer's parameter call in forward method to:<\/p>\n\n<pre><code>  def forward(self, x):\n    return x*self.my_registered_parameter[0]\n<\/code><\/pre>\n\n<p>This is where it would've been nice to have <em>pass by reference<\/em>!<\/p>\n\n<p>Now optim updates all the parameters as expected!<\/p>\n"},{"AnswerId":"53464798","CreationDate":"2018-11-25T05:03:40.347","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>What you did i.e. <code>return x*self.my_registered_parameter[0]<\/code> worked because you use the registered param for calculating the gradient. <\/p>\n\n<p>When you call <code>nn.Parameter<\/code> it returns a new object and hence <code>self.my_parameter<\/code> that you use for the operation and the one registered are not same.<\/p>\n\n<p>You can fix this by declaring the <code>my_parameter<\/code> as <code>nn.Parameter<\/code><\/p>\n\n<pre><code>self.my_parameter = nn.Parameter(torch.rand(1, requires_grad = True))\nself.my_registered_parameter= nn.ParameterList([self.some_parameter])\n<\/code><\/pre>\n\n<p>or you don't need to create <code>my_registered_parameter<\/code> variable at all. When you declare <code>self.my_parameter<\/code> as <code>nn.Parameter<\/code> it gets registered as a parameter. <\/p>\n"}]}
+{"QuestionId":53463782,"AnswerCount":1,"Tags":"<python><pandas><numpy><tensorflow><keras>","CreationDate":"2018-11-25T00:58:32.253","AcceptedAnswerId":53463806.0,"OwnerUserId":3079439.0,"Title":"Preparing y_target variable for Neural Network training","Body":"<p>I have the array for <code>y_target<\/code> variable, which have the shape <code>(5000,)<\/code> and I would like to use it for Keras NN models for time series classification.<\/p>\n\n<p>Using numpy <code>newaxis<\/code> I have converted the dimension from 1D to 2D by doing the following: <code>y_target = y_target[..., newaxis]<\/code><\/p>\n\n<p>My target variable has 2 classes: <code>[0,1]<\/code> so the array from the last converting (shape <code>(5000,1)<\/code>). And looks like this:<\/p>\n\n<pre><code>[[0],\n [0],\n [1],\n [0],\n [1]]\n<\/code><\/pre>\n\n<p>I would like to transform this array using the following logics: <code>(5000,2)<\/code> which will have binary structure like this:<\/p>\n\n<pre><code> [[1, 0],\n  [1, 0],\n  [0, 1],\n  [1, 0],\n  [0, 1]]\n<\/code><\/pre>\n\n<p>Is there a straight forward approach to do that in pandas?<\/p>\n","answers":[{"AnswerId":"53463806","CreationDate":"2018-11-25T01:03:28.120","ParentId":null,"OwnerUserId":"3079439","Title":null,"Body":"<p>I've got the point how to do that :)<\/p>\n\n<p>One possible way is to use <code>Keras<\/code> library and <code>to_categorical()<\/code> function<\/p>\n\n<pre><code>from keras.utils import to_categorical\ny_target = to_categorical(y_target)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53464139,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><neural-network><deep-learning>","CreationDate":"2018-11-25T02:22:34.897","AcceptedAnswerId":53466818.0,"OwnerUserId":10588061.0,"Title":"I'm pca with Deep neural network using TenserFlow with MNISt database, getting errors with shape of the data","Body":"<p>I'm trying to train the mnist database with the neural network after applying PCA. and I keep getting errors because of the data shape after applying the PCA. I'm not sure how to fit everything together. and how to go through the whole database, not just a small patch.<\/p>\n\n<p>here is my code:<\/p>\n\n<pre><code>    &lt;pre&gt; &lt;code&gt;\n\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport random\nfrom sklearn.preprocessing import StandardScaler\nfrom tensorflow.examples.tutorials.mnist import input_data\nfrom sklearn.decomposition import PCA\n\ndatadir='\/data' \ndata= input_data.read_data_sets(datadir, one_hot=True)\ntrain_x = data.train.images[:55000]\ntrain_y= data.train.labels[:55000]\ntest_x = data.test.images[:10000]\ntest_y = data.test.labels[:10000]\nprint(\"original shape:   \", data.train.images.shape)\n\npercent=600\npca=PCA(percent)\ntrain_x=pca.fit_transform(train_x)\ntest_x=pca.fit_transform(test_x)\nprint(\"transformed shape:\", data.train.images.shape)\ntrain_x=pca.inverse_transform(train_x)\ntest_x=pca.inverse_transform(test_x)\nc=pca.n_components_\n\nplt.figure(figsize=(8,4));\nplt.subplot(1, 2, 1);\nimage=np.reshape(data.train.images[3],[28,28])\nplt.imshow(image, cmap='Greys_r')\nplt.title(\"Original Data\")\n\nplt.subplot(1, 2, 2);\nimage1=train_x[3].reshape(28,28)\nimage.shape\nplt.imshow(image1, cmap='Greys_r')\nplt.title(\"Original Data after 0.8 PCA\")\n\nplt.figure(figsize=(10,8))\nplt.plot(range(c), np.cumsum(pca.explained_variance_ratio_))\nplt.grid()\nplt.title(\"Cumulative Explained Variance\")\nplt.xlabel('number of components')\nplt.ylabel('cumulative explained variance');\n\n\nnum_iters=10\nhidden_1=1024\nhidden_2=1024\ninput_l=percent\nout_l=10\n'''input layer'''\nx=tf.placeholder(tf.float32, [None, 28,28,1])\nx=tf.reshape(x,[-1, input_l])\n\nw1=tf.Variable(tf.random_normal([input_l,hidden_1])) \nw2=tf.Variable(tf.random_normal([hidden_1,hidden_2]))\nw3=tf.Variable(tf.random_normal([hidden_2,out_l]))\n\nb1=tf.Variable(tf.random_normal([hidden_1]))\nb2=tf.Variable(tf.random_normal([hidden_2]))\nb3=tf.Variable(tf.random_normal([out_l]))\n\nLayer1=tf.nn.relu_layer(x,w1,b1)\nLayer2=tf.nn.relu_layer(Layer1,w2,b2)\ny_pred=tf.matmul(Layer2,w3)+b3\ny_true=tf.placeholder(tf.float32,[None,out_l])\n\n\nloss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, \nlabels=y_true))\noptimizer= tf.train.AdamOptimizer(0.006).minimize(loss)\ncorrect_pred=tf.equal(tf.argmax(y_pred,1), tf.argmax(y_true,1))\naccuracy= tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\nstore_training=[]\nstore_step=[]\nm = 10000\n\ninit=tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    sess.run(init)\n\n    for epoch in range(num_iters):\n        indices = random.sample(range(0, m), 100)\n        batch_xs = train_x[indices]\n        batch_ys = train_y[indices]\n        sess.run(optimizer, feed_dict={x:batch_xs, y_true:batch_ys})\n        training=sess.run(accuracy, feed_dict={x:test_x, y_true:test_y})\n        store_training.append(training)  \n    testing=sess.run(accuracy, feed_dict={x:test_x, y_true:test_y})\n\nprint('Accuracy :{:.4}%'.format(testing*100))\nz_reg=len(store_training)\nx_reg=np.arange(0,z_reg,1)\ny_reg=store_training\nplt.figure(1)\nplt.plot(x_reg, y_reg,label='Regular Accuracy')\n<\/code><\/pre>\n\n<p>that is the error I got :\n    <pre> <code>\n    \"Traceback (most recent call last):<\/p>\n\n File \"&lt;ipython-input-2-ff57ada92ef5&gt;\", line 135, in &lt;module&gt;\nsess.run(optimizer, feed_dict={x:batch_xs, y_true:batch_ys})\n\n File \"C:\\anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", \n line 929, in run\n  run_metadata_ptr)\n\n File \"C:\\anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", \n line 1128, in _run\n str(subfeed_t.get_shape())))\n\n ValueError: Cannot feed value of shape (100, 784) for Tensor 'Reshape:0', \n which has shape '(?, 600)'\"\n<\/code><\/pre>\n","answers":[{"AnswerId":"53466818","CreationDate":"2018-11-25T11:03:48.623","ParentId":null,"OwnerUserId":"6560413","Title":null,"Body":"<p>First of all, I\u00b4d recommend to fit PCA only for train set since you may get different PCA components for train and test. So the easiest fix is to change the following piece of code:<\/p>\n\n<pre><code>percent=600\npca=PCA(percent)\ntrain_x=pca.fit_transform(train_x)\ntest_x=pca.fit_transform(test_x)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>percent=.80\npca=PCA(percent)\npca.fit(train_x)\ntrain_x=pca.transform(train_x)\ntest_x=pca.transform(test_x)\n<\/code><\/pre>\n\n<p>Secondly, you use <code>percent=600<\/code> while doing PCA and then applying PCA inverse transform which means that you return to space with the original number of features. In order to start learning with the reduced number of PCA components you may also try to change this piece of code:<\/p>\n\n<pre><code>train_x=pca.inverse_transform(train_x)\ntest_x=pca.inverse_transform(test_x)\nc=pca.n_components_\n&lt;plotting code&gt;    \ninput_l=percent\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>c=pca.n_components_\n#plotting commented out   \ninput_l=c\n<\/code><\/pre>\n\n<p>It should give you the correct tensor dimensions for subsequent optimization procedure.<\/p>\n"},{"AnswerId":"53466574","CreationDate":"2018-11-25T10:25:01.217","ParentId":null,"OwnerUserId":"10701497","Title":null,"Body":"<p>The error says that your are reshaping your tensor <code>x<\/code> which has the shape <code>(None, 28 , 28, 1)<\/code> to shape <code>(None, percent)<\/code>\nand you gave the value of <code>percent<\/code> as 600, and later you are giving the value of x as <code>(100, 28*28*1)<\/code> and feeding it to <code>x<\/code> which has the shape of <code>(None, 600)<\/code> and that doesn't match.<\/p>\n"}]}
+{"QuestionId":53464247,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2018-11-25T02:51:17.507","AcceptedAnswerId":53464320.0,"OwnerUserId":5819072.0,"Title":"How does `images, labels = dataiter.next() ` work in PyTorch Tutorial?","Body":"<p>Newer Python &amp; PyTorch practitioner here. <\/p>\n\n<p>From the tutorial <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html\" rel=\"nofollow noreferrer\">cifar10_tutorial<\/a>, how is <code>images, labels<\/code> assigned?<\/p>\n\n<pre><code>trainset = torchvision.datasets.CIFAR10(root='.\/data', train=True,\n                                        download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=4,\n                                          shuffle=True, num_workers=2)\n\ndef imshow(img):\n    img = img \/ 2 + 0.5     # unnormalize\n    npimg = img.numpy()\n    plt.imshow(np.transpose(npimg, (1, 2, 0)))\n\n\n# get some random training images\ndataiter = iter(trainloader)\n\nimages, labels = dataiter.next()\n<\/code><\/pre>\n\n<p>How does the last line know how to automatically assign <code>images, label<\/code> in <code>images, labels = dataiter.next()<\/code>?<\/p>\n\n<p>I checked the <code>DataLoader<\/code> class and the <code>DataLoaderIter<\/code> class, but think I need a bit more knowledge of iters in general.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53464320","CreationDate":"2018-11-25T03:11:12.790","ParentId":null,"OwnerUserId":"7343721","Title":null,"Body":"<p>I think it is crucial to understand the difference between an iterable and an iterator. An iterable is an object that you can iterate over. An Iterator is an object which is used to iterate over an iterable object using the __next__ method, which returns the next item of the object.<\/p>\n\n<p>A simple example is the following. Consider an iterable and use the next method to call the next item in the list. This will print the next item until the end of the list is reached. If the end is reached it will raise a StopIteration error.<\/p>\n\n<pre><code>test = (1,2,3)\ntester = iter(test)\n\nwhile True:\n    nextItem = next(tester)\n    print(nextItem)\n<\/code><\/pre>\n\n<p>The class you refer to above probably has an implementation similar to this, however it returns a tuple containing the image and the label.<\/p>\n"}]}
+{"QuestionId":53464665,"AnswerCount":0,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2018-11-25T04:32:37.977","AcceptedAnswerId":null,"OwnerUserId":10693988.0,"Title":"Keras \/ tensorflow custom layer gradient calculation that requires loop and if","Body":"<p>My question is about custom gradient calculation. I have a custom layer and could not figure out how to properly program its gradient. This gradient corresponds to a transfer function that is similar to a hard-limiter (or sigmoid with very steep transition): the output is <code>y=x<\/code> if the absolute value <code>|x| &lt; th<\/code> is less than a threshold <code>th<\/code> or <code>sign(x)<\/code> otherwise. So, the gradient <code>g<\/code> is <code>g=0<\/code> if <code>|x| &gt;= th<\/code> or <code>g=1<\/code> otherwise. After searching Web I tried the function _MySignGrad described below. The code runs but I face two problems: 1) I tried to \"register\" the grad calculation but for some reason it is not called and 2) whenever it gets called I am sure I will get runtime error given the code is potentially wrong. Any hints?<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom keras.layers import Input, Dense, Flatten\nfrom keras.models import Model\nimport keras.backend as K\nfrom keras.engine.topology import Layer\nfrom tensorflow.python.framework import ops\n\n# training parameters\nepochs = 10\nbatch_size = 3\ndim_x = 2\ndim_y = 4\nN = 100 #half training examples\n\n#define some training data and labels\ntrain_class_0 = np.random.normal(size=(N,dim_x,dim_y)) + 3\ntrain_class_1 = np.random.normal(size=(N,dim_x,dim_y)) - 3\ntrain_data = np.concatenate((train_class_0,train_class_1),axis=0)\noutput_labels = np.concatenate((0*np.ones(N,),1*np.ones(N,)))\n\n# Define custom function which takes also a grad op as argument:\ndef ak_py_func(func, inp, Tout, stateful=True, name=None, grad=None):\n    # Need to generate a unique name to avoid duplicates:\n    rnd_name = 'AkPyFuncGrad' + str(np.random.randint(0, 1E+7))\n    tf.RegisterGradient(rnd_name)(grad)\n    g = tf.get_default_graph()\n    #tf.py_func allows to use a python code as an op\n    with g.gradient_override_map({\"PyFunc\": rnd_name, \"PyFuncStateless\": rnd_name}):\n        return tf.py_func(func, inp, Tout, stateful=stateful, name=name)\n\n# My custom gradient calculation that is wrong\ndef _MySignGrad(op, grad):\n    threshold = 0.01\n    newgrad = grad #ok, I guess I cannot do this\n    for i in range(batch_size):\n        mask = K.ones((op.inputs[i].shape)) #initialize with 1's\n        mask[K.abs(op.inputs[i]) &gt; threshold] = 0 #mas is zero when input is above threshold\n        newgrad[i] = grad[i] * mask #zero selected gradient entries\n    return newgrad\n\n# function that binds forward and backward pass of tensorflow ops\ndef myfunc(x, name=None):\n    with ops.name_scope(name, \"Myfunc\", [x]) as name:\n        output_x = ak_py_func(np.sign, #try also square and exp\n                            [x],\n                            [tf.float32],\n                            name=name,\n                            grad=_MySignGrad)  # &lt;-- here's the call to the gradient\n        return output_x[0]\n\n#Custom layer\nclass LimiterLayer(Layer):\n    def __init__(self, **kwargs):\n        super(LimiterLayer, self).__init__(**kwargs)\n    def call(self, x):\n        return myfunc(x)\n    def get_output_shape_for(self, input_shape):\n        return tuple(input_shape)\n\n# Define all layers\nl1_input = Input(shape=(dim_x,dim_y),name='l1_input')\nl2_flatten = Flatten(name='l2_flatten')(l1_input)\nl3_limiter = LimiterLayer(name='l3_limiter')(l2_flatten)\nl4_dense = Dense(10, activation='tanh', name='l4_dense')(l3_limiter)\nl5_output = Dense(1, activation='sigmoid', name='l5_output') (l4_dense)\n# put it all together, creating model from layers\nmodel = Model(l1_input, l5_output)\nprint(model.summary())\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(train_data, output_labels, epochs=epochs, batch_size=batch_size, validation_split=0.3)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53464692,"AnswerCount":0,"Tags":"<python><mpi><pytorch><scatter>","CreationDate":"2018-11-25T04:39:11.507","AcceptedAnswerId":null,"OwnerUserId":10700971.0,"Title":"PyTorch TypeError: scatter() missing 1 required positional argument: 'scatter_list'","Body":"<p>all. I have a problem with starting a distributed program while using \u201cmpi\u201d as the backend. The program is as follows:<\/p>\n\n<pre><code>def run(rank, size, hostname):\n    print(\"I am {} of {} in {}\".format(rank, size, hostname))\n    tensor = torch.zeros(1)\n    group=dist.new_group([0,1,2])\n    if rank == 0:\n        scatter_list=[torch.zeros(1) for _ in range(3)]\n        dist.scatter(tensor= tensor, src=0, scatter_list=scatter_list, group=group)\n        print(\"Master has completed Scatter\")\n    else:\n        tensor += 1\n        dist.scatter(tensor= tensor, src=0, group=group)\n        print(\"worker has completed scatter\")\n    print('Rank', rank, 'has data', tensor[0])\n\ndef init_process(rank, size, hostname, fn, backend='tcp'):\n    dist.init_process_group(backend, rank=rank, world_size=size)\n    fn(rank, size,hostname)\n\nif __name__ == \"__main__\":\n    world_size = int(os.environ['OMPI_COMM_WORLD_SIZE'])\n    world_rank = int(os.environ['OMPI_COMM_WORLD_RANK'])\n    hostname = socket.gethostname()\n    p = Process(target = init_process, \n                     args=(world_rank, world_size, hostname, run, 'mpi'))\n    p.start()\n    p.join()\n<\/code><\/pre>\n\n<p>However, when the program starts, it always throws an error like:<\/p>\n\n<pre><code>  File \"mpi_test.py\", line 17, in run\ndist.scatter(tensor= tensor, src=0, group=group)\n<\/code><\/pre>\n\n<p><code>TypeError: scatter() missing 1 required positional argument: 'scatter_list'<\/code><\/p>\n\n<p>The error is emitted by ranks 1 and 2 which don\u2019t need argument: \u2018scatter_list\u2019.\nI tried many ways, but failed. Does anybody know why?\nThanks for your reading.<\/p>\n","answers":[]}
+{"QuestionId":53465006,"AnswerCount":0,"Tags":"<numpy><keras>","CreationDate":"2018-11-25T05:50:22.607","AcceptedAnswerId":null,"OwnerUserId":1390640.0,"Title":"How to use Keras' evaluate_generator() when \"Axis must be specified...\"","Body":"<p>Why is Keras telling me weights and batch sizes are different? How can I fix this? (adding <code>next(...)<\/code> does not help here).\nThanks in advance; there's something I'm just not getting here.<\/p>\n\n<p>Error is on evaluate_generator(): <code>TypeError: Axis must be specified when shapes of a and weights differ.<\/code><\/p>\n\n<pre><code>from sklearn.utils import shuffle as identical_shuffle\n\nSAMPLES_PER_BATCH=1\nBATCHES_PER_EPOCH=1\nBATCHES_PER_VALIDATION=1\nEPOCHS_PER_SIMULATION=2\n\ndef generate_training(batch_size=64):\n    validation_length = (int(len(data.train)*0.25) \/\/ batch_size) * batch_size\n    while True:\n        for i in range(validation_length,len(data.train),batch_size):\n            x,y,n = identical_shuffle(data.train[i:i+batch_size],data.target[i:i+batch_size],data.context[i:i+batch_size])\n            yield {'input':x,'target':n}, y\n\ndef generate_validation(batch_size=64):\n    validation_length = (int(len(data.train)*0.25) \/\/ batch_size) * batch_size\n    while True:\n        for i in range(0,validation_length,batch_size):\n            x,y,n = identical_shuffle(data.train[i:i+batch_size],data.target[i:i+batch_size],data.context[i:i+batch_size])\n            yield {'input':x,'target':n}, y\n\nfor epoch in range(EPOCHS_PER_SIMULATION):\n  for batch in range(BATCHES_PER_EPOCH):\n    result_training = model.train_on_batch( *next(generate_training(batch_size=SAMPLES_PER_BATCH)) )\n    # &lt;redacted operations on result_training&gt;\n  result_validation = model.evaluate_generator( generate_validation(batch_size=SAMPLES_PER_BATCH), steps=BATCHES_PER_VALIDATION )\n<\/code><\/pre>\n\n<p>For comparison, the below runs okay with the same generator.<\/p>\n\n<pre><code>model.fit_generator(generator=generate_training(batch_size=SAMPLES_PER_BATCH),\n                    validation_data=next(generate_validation(batch_size=SAMPLES_PER_BATCH)),\n                    validation_steps=1,\n                    steps_per_epoch=1,\n                    epochs=1)\n<\/code><\/pre>\n\n<p>Full Traceback<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-37-962a23d537c3&gt; in &lt;module&gt;()\n     79     print(result_training)\n     80   model.reset_states()\n---&gt; 81   result_validation = model.evaluate_generator( generate_validation(batch_size=SAMPLES_PER_BATCH), steps=BATCHES_PER_VALIDATION )\n     82   #result_validation = model.test_on_batch( *next(generate_validation(batch_size=SAMPLES_PER_BATCH)) )\n     83   print(\"VALIDATION OF EPOCH \"+str(epoch))\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/legacy\/interfaces.py in wrapper(*args, **kwargs)\n     89                 warnings.warn('Update your `' + object_name + '` call to the ' +\n     90                               'Keras 2 API: ' + signature, stacklevel=2)\n---&gt; 91             return func(*args, **kwargs)\n     92         wrapper._original_function = func\n     93         return wrapper\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training.py in evaluate_generator(self, generator, steps, max_queue_size, workers, use_multiprocessing, verbose)\n   1470             workers=workers,\n   1471             use_multiprocessing=use_multiprocessing,\n-&gt; 1472             verbose=verbose)\n   1473 \n   1474     @interfaces.legacy_generator_methods_support\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training_generator.py in evaluate_generator(model, generator, steps, max_queue_size, workers, use_multiprocessing, verbose)\n    375         if i not in stateful_metric_indices:\n    376             averages.append(np.average([out[i] for out in outs_per_batch],\n--&gt; 377                                        weights=batch_sizes))\n    378         else:\n    379             averages.append(np.float64(outs_per_batch[-1][i]))\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/lib\/function_base.py in average(a, axis, weights, returned)\n   1140             if axis is None:\n   1141                 raise TypeError(\n-&gt; 1142                     \"Axis must be specified when shapes of a and weights \"\n   1143                     \"differ.\")\n   1144             if wgt.ndim != 1:\n\nTypeError: Axis must be specified when shapes of a and weights differ.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53465066,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network>","CreationDate":"2018-11-25T06:03:01.317","AcceptedAnswerId":53465497.0,"OwnerUserId":10227815.0,"Title":"Keras CNN Error: expected Sequence to have 3 dimensions, but got array with shape (500, 400)","Body":"<p>I'm getting this error. I'm quite new to ML.<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected Sequence to have 3 dimensions, but got array with shape (500, 400)<\/p>\n<\/blockquote>\n\n<p>These are the below codes that I'm using.<\/p>\n\n<pre><code>print(X1_Train.shape)\nprint(X2_Train.shape)\nprint(y_train.shape)\n\n====================================\nOutput (here I've 500 rows in each):\n(500, 400)\n(500, 1500)\n(500,)\n\n400 =&gt; timesteps (below)\n1500 =&gt; n (below)\n====================================\n\n\ntimesteps = 50 * 8\nn = 50 * 30\n\ndef createClassifier():\n    sequence = Input(shape=(timesteps, 1), name='Sequence')\n    features = Input(shape=(n,), name='Features')\n\n    conv = Sequential()\n    conv.add(Conv1D(10, 5, activation='relu', input_shape=(timesteps, 1)))\n    conv.add(Conv1D(10, 5, activation='relu'))\n    conv.add(MaxPool1D(2))\n    conv.add(Dropout(0.5))\n\n    conv.add(Conv1D(5, 6, activation='relu'))\n    conv.add(Conv1D(5, 6, activation='relu'))\n    conv.add(MaxPool1D(2))\n    conv.add(Dropout(0.5))\n    conv.add(Flatten())\n    part1 = conv(sequence)\n\n    merged = concatenate([part1, features])\n\n    final = Dense(512, activation='relu')(merged)\n    final = Dropout(0.5)(final)\n    final = Dense(num_class, activation='softmax')(final)\n\n    model = Model(inputs=[sequence, features], outputs=[final])\n    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n\nmodel = createClassifier()\n# print(model.summary())\nhistory = model.fit([X1_Train, X2_Train], y_train, epochs =5)\n<\/code><\/pre>\n\n<p>Any insight please ?\nThanks in advance.<\/p>\n","answers":[{"AnswerId":"53465497","CreationDate":"2018-11-25T07:27:46.777","ParentId":null,"OwnerUserId":"4466962","Title":null,"Body":"<p>Two things -<\/p>\n\n<p>Conv1D layer expects input to be in the shape <code>(batch_size, x, filters)<\/code>, in your case <code>(500,400,1)<\/code>. <br>\nYou need to reshape your input layer, add another axis, of size 1. (this does not change anything in your data). <br><\/p>\n\n<p>You are trying to use multiple inputs, Sequential API is not the best choice for that. I would recommend using the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Functional API<\/a><\/p>\n\n<p><strong>Edit:<\/strong>\nRegarding your comment, not sure what you did wrong, but this is a working version of your code (with fake data), with a reshape:<\/p>\n\n<pre><code>import keras\n\nimport numpy as np\n\n\n\nX1_Train = np.ones((500,400))\nX2_Train = np.ones((500,1500))\ny_train = np.ones((500))\nprint(X1_Train.shape)\nprint(X2_Train.shape)\nprint(y_train.shape)\n\nnum_class = 1\n\n\ntimesteps = 50 * 8\nn = 50 * 30\n\ndef createClassifier():\n    sequence = keras.layers.Input(shape=(timesteps, 1), name='Sequence')\n    features = keras.layers.Input(shape=(n,), name='Features')\n\n    conv = keras.Sequential()\n    conv.add(keras.layers.Conv1D(10, 5, activation='relu', input_shape=(timesteps, 1)))\n    conv.add(keras.layers.Conv1D(10, 5, activation='relu'))\n    conv.add(keras.layers.MaxPool1D(2))\n    conv.add(keras.layers.Dropout(0.5))\n\n    conv.add(keras.layers.Conv1D(5, 6, activation='relu'))\n    conv.add(keras.layers.Conv1D(5, 6, activation='relu'))\n    conv.add(keras.layers.MaxPool1D(2))\n    conv.add(keras.layers.Dropout(0.5))\n    conv.add(keras.layers.Flatten())\n    part1 = conv(sequence)\n\n    merged = keras.layers.concatenate([part1, features])\n\n    final = keras.layers.Dense(512, activation='relu')(merged)\n    final = keras.layers.Dropout(0.5)(final)\n    final = keras.layers.Dense(num_class, activation='softmax')(final)\n\n    model = keras.Model(inputs=[sequence, features], outputs=[final])\n    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n\nmodel = createClassifier()\n# print(model.summary())\nX1_Train = X1_Train.reshape((500,400,1))\nhistory = model.fit([X1_Train, X2_Train], y_train, epochs =5)\n<\/code><\/pre>\n\n<p>With output:<\/p>\n\n<pre><code>Using TensorFlow backend.\n(500, 400)\n(500, 1500)\n(500,)\nEpoch 1\/5\n500\/500 [==============================] - 1s 3ms\/step - loss: 1.1921e-07 - acc: 1.0000\nEpoch 2\/5\n500\/500 [==============================] - 0s 160us\/step - loss: 1.1921e-07 - acc: 1.0000\nEpoch 3\/5\n500\/500 [==============================] - 0s 166us\/step - loss: 1.1921e-07 - acc: 1.0000\nEpoch 4\/5\n500\/500 [==============================] - 0s 154us\/step - loss: 1.1921e-07 - acc: 1.0000\nEpoch 5\/5\n500\/500 [==============================] - 0s 157us\/step - loss: 1.1921e-07 - acc: 1.0000\n<\/code><\/pre>\n"}]}
+{"QuestionId":53465376,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-25T07:02:11.467","AcceptedAnswerId":null,"OwnerUserId":9511772.0,"Title":"Session graph empty - multiple graphs","Body":"<p>I am trying to use a Policy Network and CNN within the same Python file. I am only training the PN and using the CNN as a feed-forward feature extractor. I am getting an issue where the Train-PC isn't \"building\" the imported Keras graph into the session, and thus when I try to run a feed-forward function on the CNN (<code>feature_vector = self.extract_feature_vector([X_img_norm])[0]<\/code>), I get the error posted below. <\/p>\n\n<p>The only difference between the files on the Home-PC I am developing on and the Train-PC is the way I import the model which I believe may be causing this (possibly it's not recognised). I posted previously about this error <a href=\"https:\/\/stackoverflow.com\/questions\/53368717\/loading-keras-model-keyerror-weighted-metrics\">here<\/a>. <\/p>\n\n<p>Since the Train-PC is not my own, I cannot change the version of the packages. So I am looking for a solution which does not involve downgrading versions, as the issue with importing wasn't solved through that also.  <\/p>\n\n<p>Home-PC (Mac):<\/p>\n\n<pre><code>tensorflow==1.11.0\nkeras 2.2.4\n<\/code><\/pre>\n\n<p>Train-PC (Ubuntu):<\/p>\n\n<pre><code>tensorflow-gpu==1.10.1\nkeras 2.1.2\n<\/code><\/pre>\n\n<p>Code of CNN:<\/p>\n\n<pre><code>class ResNetCNN:\n\n    def __init__(self):\n        self.graph = tf.Graph()\n        self.sess = tf.Session(graph=self.graph)\n\n        with self.sess.graph.as_default():\n            # Load pre-trained model from file\n            self.CNN_ResNet_Pascal = models.load_model(CNN_MODEL_DIR)\n            # self.CNN_ResNet_Pascal = tf.keras.models.load_model(CNN_MODEL_DIR)\n\n            self.extract_feature_vector = K.function([self.CNN_ResNet_Pascal.layers[0].input],\n                                                     [self.CNN_ResNet_Pascal.layers[-4].output])\n\n    def feed_forward(self, img):\n        \"\"\"Extract the feature vector of the passed image using the pre-trained CNN.\"\"\"\n\n        # Resize observable region into input volume\n        X_img = np.array(img, dtype=np.float32).reshape(-1, IMG_SIZE, IMG_SIZE, 3)\n\n        # Normalise feature data\n        X_img_norm = X_img \/ 255.0\n\n        # Extract feature vector from the pre-trained CNN (1, 4096)\n        feature_vector = self.extract_feature_vector([X_img_norm])[0]\n\n        # Reshape tensor to (4096, )\n        feature_vector = np.array(feature_vector).reshape(4096, )\n\n        # Confirm that the correct output layer was used\n        assert feature_vector.shape == (4096, ), \"Incorrect CNN output layer: shape = {}\".format(feature_vector.shape)\n\n        return feature_vector\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"keras_pn.py\", line 856, in &lt;module&gt;\n    s, a, r, d_r, n = rollout(epsilon, RENDER, PolicyNetwork, ResNetCNN)\n  File \"keras_pn.py\", line 682, in rollout\n    feature_vector = ResNetCNN.feed_forward(observable_region)\n  File \"keras_pn.py\", line 87, in feed_forward\n    feature_vector = self.extract_feature_vector([X_img_norm])[0]\n  File \"\/home\/name\/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2357, in __call__\n    **self.session_kwargs)\n  File \"\/home\/name\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 877, in run\n    run_metadata_ptr)\n  File \"\/home\/name\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1025, in _run\n    raise RuntimeError('The Session graph is empty.  Add operations to the '\nRuntimeError: The Session graph is empty.  Add operations to the graph before calling run().\n<\/code><\/pre>\n\n<p>Edit:<\/p>\n\n<pre><code>def __init__(self):\n    self.graph = tf.Graph()\n\n    with self.graph.as_default():\n        # Load pre-trained model from file\n        self.CNN_ResNet_Pascal = models.load_model(CNN_MODEL_DIR)\n        # self.CNN_ResNet_Pascal = tf.keras.models.load_model(CNN_MODEL_DIR)\n\n        self.extract_feature_vector = K.function([self.CNN_ResNet_Pascal.layers[0].input],\n                                                 [self.CNN_ResNet_Pascal.layers[-4].output])\n\n    self.sess = tf.Session(graph=self.graph)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53465463,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2018-11-25T07:20:08.173","AcceptedAnswerId":null,"OwnerUserId":8310534.0,"Title":"pytorch CUDA out of memory when training on big image","Body":"<p>When I train a model with <code>512*512<\/code> images with <code>batch_size=4<\/code>, it works,\nbut when I train it with <code>1024*1024<\/code> images with <code>batch_size=1<\/code>, the <code>cuda<\/code> out of memory error happens.<\/p>\n\n<p>I am confused, the total number of pixels are equal. Why one <code>1024*1024<\/code> image uses more memory than four <code>512*512<\/code> images?<\/p>\n","answers":[]}
+{"QuestionId":53465475,"AnswerCount":1,"Tags":"<python><machine-learning><keras><word2vec><word-embedding>","CreationDate":"2018-11-25T07:22:01.257","AcceptedAnswerId":null,"OwnerUserId":1355120.0,"Title":"Confusion about input shape for Keras Embedding layer","Body":"<p>I'm trying to use the Keras embedding layer to create my own CBoW implementation to see how it works.<\/p>\n\n<p>I've generated outputs represented by a vector of the context word I'm searching for with size equal to my vocab. I've also generated inputs so that each context word has X many nearby words represented by their one-hot encoded vectors.<\/p>\n\n<p>So for example if my sentence is:<\/p>\n\n<blockquote>\n  <p>\"I ran over the fence to find my dog\"<\/p>\n<\/blockquote>\n\n<p>using window size 2, I could generate the following input\/output:<\/p>\n\n<p>[[over, the, to, find], fence] where 'fence' is my context word, 'over', 'the', 'to', 'find' are my nearby words with window 2 (2 in front, 2 in back).<\/p>\n\n<p>Using sample vocab size of 500 and 100 training samples, after one-hot encoding my input and output, it would have the following dimensions:<\/p>\n\n<pre><code>y.shape -&gt; (100,500)\nX.shape -&gt; (100,4,500)\n<\/code><\/pre>\n\n<p>That is, I have 100 outputs each represented by a 500-sized vector. I have 100 inputs each represented by a series of 4 500-sized vectors.<\/p>\n\n<p>I have a simple model defined as:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(input_dim=vocabulary_size, output_dim=embedding_size, input_length=2*window_size))\n#take average of context words at hidden layer\nmodel.add(Lambda(lambda x: K.mean(x, axis = 1), output_shape=(embedding_size,)))\nmodel.add(Dense(vocabulary_size, activation='softmax'))\nmodel.compile(loss = 'categorical_crossentropy', optimizer = 'adam')\n<\/code><\/pre>\n\n<p>However, when I try to fit my model, I get a dimensional exception:<\/p>\n\n<pre><code>model.fit(X, y, batch_size=10, epochs=2, verbose=1)\nValueError: Error when checking input: expected embedding_6_input to have 2 dimensions, but got array with shape (100, 4, 500)\n<\/code><\/pre>\n\n<p>Now, I can only assume I'm using the embedding layer wrongly. I've read both this <a href=\"https:\/\/stats.stackexchange.com\/questions\/270546\/how-does-keras-embedding-layer-work\">CrossValidated Question<\/a> and the <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">Keras documentation<\/a>.<\/p>\n\n<p>I'm still not sure exactly how the inputs of this embedding layer works. I'm fairly certain my <code>input_dim<\/code> and <code>output_dim<\/code> are correct, which leaves <code>input_length<\/code>. According to the CrossValidated, my <code>input_length<\/code> is the length of my sequence. According to Keras, my input should be of dimension <code>(batch_size, input_length)<\/code>.<\/p>\n\n<p>If my inputs are 4 words each represented by a word vector of size <code>vocab_size<\/code>, how do I input this to the model? <\/p>\n","answers":[{"AnswerId":"53519199","CreationDate":"2018-11-28T12:11:59.887","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you are thinking about the embedding layer in a wrong way. An Embedding layer is just a trainable look-up table: you give it <strong>an integer<\/strong>, which is the index of the word in the vocabulary, and it returns the word-vector (i.e. word embedding) of the given index. Therefore, its input must be the indices of the words in a sentence.<\/p>\n\n<p>As an example, if the indices of the words \"over\", \"the\", \"to\" and \"find\" are 43, 6, 9 and 33 respectively, then the input of the Embedding layer would be an array of those indices, i.e. <code>[43, 6, 9, 33]<\/code>. Therefore, the training data must have a shape of <code>(num_samples, num_words_in_a_sentence)<\/code>. In your case, it would be <code>(100, 4)<\/code>. In other words, you don't need to one-hot encode the words for the input data. You can also use word indices as the labels as well if you use <code>sparse_categorical_crossentropy<\/code> as the loss function instead.<\/p>\n"}]}
+{"QuestionId":53465477,"AnswerCount":2,"Tags":"<tensorflow><keras><pycharm><cudnn>","CreationDate":"2018-11-25T07:22:03.253","AcceptedAnswerId":53838473.0,"OwnerUserId":7231796.0,"Title":"Keras in PyCharm not using GPU","Body":"<p>These threads did not solve my problem:\n<a href=\"https:\/\/stackoverflow.com\/questions\/47639587\/keras-does-not-use-gpu-on-pycharm-having-python-3-5-and-tensorflow-1-4\">Keras does not use GPU on Pycharm having python 3.5 and Tensorflow 1.4<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/46080634\/keras-with-tensorflow-backend-not-using-gpu\">Keras with TensorFlow backend not using GPU<\/a><\/p>\n\n<p>I have installed Tensorflow and Tensorflow-gpu (v.1.12.0) on my PC which is running Windows 10 and has GTX 750 Ti graphics card, So it does support CUDA. I have also installed CUDA Toolkit v10 and cuDNN libraries and when I run <code>nvcc -V<\/code> on command prompt I get:<\/p>\n\n<blockquote>\n  <p>nvcc: NVIDIA (R) Cuda compiler...<\/p>\n<\/blockquote>\n\n<p>I am using PyCharm and I don't have any problem running Keras on CPU. But it doesn't use my GPU.<\/p>\n\n<p>When I type<\/p>\n\n<pre><code>from keras import backend as K\nK.tensorflow_backend._get_available_gpus()\n<\/code><\/pre>\n\n<p>it says<\/p>\n\n<blockquote>\n  <p>2018-11-25 10:47:57.448275: I tensorflow\/core\/platform\/cpu_feature_gaurd.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2<\/p>\n  \n  <p>[ ]<\/p>\n<\/blockquote>\n\n<p><strong>What I have tried:<\/strong><\/p>\n\n<p>1) I tried uninstalling Tensorflow and Tensorflow-gpu and reinstalling Tensorflow-gpu like the above thread says. Didn't work and my code didn't run on CPU anymore and gave an error regarding Tensorflow. Once I reinstalled Tensorflow, It was OK again.<\/p>\n\n<p>2) I tried using a library named Theano, which is mentioned on Keras official documentation <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-run-keras-on-gpu\" rel=\"noreferrer\">here<\/a>. When I add the following lines<\/p>\n\n<pre><code>import theano\ntheano.config.device='gpu'\ntheano.config.floatX='float32'\n<\/code><\/pre>\n\n<p>it says<\/p>\n\n<blockquote>\n  <p>WARNING (theano.configdefaults): g++ not available, if using conda: 'conda install m2w64-toolchain'\n  C:\\Users\\HOME-PC\\PyCharmProjects\\env\\lib\\site-packages\\theano\\configdefaults.py.=:560: UserWarning: DeprecationWarning: there is no c++ compiler.This is deprecated and with theano 0.11 a c++ compiler will be mandatory. warning.warn(\"DeprecationWarning: there is no c++ compiler.\"<\/p>\n<\/blockquote>\n\n<p>3) I tried adding these lines and also nothing happened.<\/p>\n\n<pre><code>import os\nos.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\" \n<\/code><\/pre>\n\n<p>What am I missing here? How should I introduce CUDA to PyCharm?<\/p>\n","answers":[{"AnswerId":"53838473","CreationDate":"2018-12-18T17:46:10.640","ParentId":null,"OwnerUserId":"7231796","Title":null,"Body":"<p>The problem was with CUDA's version. I had installed CUDA v10.0 but Tensorflow seems to work only with v9.0. Installed it and it works like a charm.<\/p>\n"},{"AnswerId":"53572780","CreationDate":"2018-12-01T16:22:31.727","ParentId":null,"OwnerUserId":"6911920","Title":null,"Body":"<p>It might not be the case but, installing and importing the same librariy can sometimes be confusing.\nMore clearly, my guess is your pycharm environment is different than the default python environment, youre installing packages in the default environment and importing from the pycharm's environment. \nTo ensure you have a package installed in pycharm's environment, <\/p>\n\n<p>you can try the following, \nfrom pycharm's python console<\/p>\n\n<p>!python -m pip  install --upgrade tensorflow, keras <\/p>\n\n<p>this shall properly install the packages, and you can be sure of that installation  isn't at fault<\/p>\n"}]}
+{"QuestionId":53465608,"AnswerCount":1,"Tags":"<matrix><neural-network><deep-learning><matrix-multiplication><pytorch>","CreationDate":"2018-11-25T07:48:06.550","AcceptedAnswerId":53466176.0,"OwnerUserId":7483494.0,"Title":"PyTorch - shape of nn.Linear weights","Body":"<p>Yesterday I came across <a href=\"https:\/\/stackoverflow.com\/questions\/53462493\/shape-of-pytorch-model-parameter-is-inconsistent-with-how-its-defined-in-the-mo\/\">this question<\/a> and for the first time noticed that the weights of the linear layer <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html#Linear\" rel=\"nofollow noreferrer\"><code>nn.Linear<\/code><\/a> need to be transposed before applying <code>matmul<\/code>.<\/p>\n\n<p><em><a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/functional.html\" rel=\"nofollow noreferrer\">Code<\/a> for applying the weights:<\/em><\/p>\n\n<pre><code>output = input.matmul(weight.t())\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p>What is the reason for this? <\/p>\n\n<p>Why are the weights not in the transposed shape just from the beginning, so they don't need to be transposed every time before applying the layer?<\/p>\n","answers":[{"AnswerId":"53466176","CreationDate":"2018-11-25T09:26:00.880","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>I found an answer here:\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/2159\" rel=\"nofollow noreferrer\">Efficient forward pass in nn.Linear #2159<\/a> <\/p>\n\n<p>It seems like there is no real reasoning behind this. However the transpose operation doesn't seem to be slowing down the computation.<\/p>\n\n<p>According to the issue mentioned above, during the <strong>forward<\/strong> pass the transpose operation <strong>is (almost) free<\/strong> in terms of computation. While during the backward pass <strong>leaving out the transpose operation<\/strong> would actually make computation <strong>less efficient<\/strong> with the current implementation.<\/p>\n\n<p>The last post in that issue sums it up quite nicely:<\/p>\n\n<blockquote>\n  <p>It's historical weight layout, changing it is backward-incompatible.\n  Unless there is some BIG benefit in terms of speed or convenience, we\n  wont break userland.<\/p>\n<\/blockquote>\n\n<p><em><a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/2159#issuecomment-390068272\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/2159#issuecomment-390068272<\/a><\/em><\/p>\n"}]}
+{"QuestionId":53466396,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-25T09:57:35.540","AcceptedAnswerId":null,"OwnerUserId":9137210.0,"Title":"tensorflow dataset api input for training tuple object has no ndims attribute","Body":"<p>so I'm trying to train a GAN to color images using a the new TensorFlow data set API \nand I cant get it to work <\/p>\n\n<p>I'm trying to use the simple one shot iterator for my data set and I think it might be causing the problem but I can't figure out why <\/p>\n\n<p>so what I'm asking is <\/p>\n\n<p>can someone tell me whats wrong with the code <\/p>\n\n<p><strong>code:<\/strong><\/p>\n\n<p>creating the data set <\/p>\n\n<pre><code>def get_next():\n\n    #where gray_ls is just a list of image paths \n    gray_ds   = tf.data.Dataset.from_tensor_slices(gray_ls).shuffle(50).map(in_parser).batch(30).repeat()\n    print(f\"output types = {gray_ds.output_types}\")   # --&gt; output types = &lt;dtype: 'float32'&gt;\n    print(f\"output shapes = {gray_ds.output_shapes}\") # --&gt; output shapes = (?, ?, ?, ?)\n\n    gray_iter  = gray_ds.make_one_shot_iterator()\n    next_gray  = gray_iter.get_next()\n\n    # next_color is the same as next gray just different images\n    return next_color, next_gray\n\n# mapping function \ndef in_parser(img_path):\n\n    img_file = tf.read_file(img_path)\n    img = tf.image.decode_image(img_file,channels=3)\n    img = tf.image.random_flip_left_right(img)\n    img = tf.image.random_brightness(img, max_delta = 0.1)\n    img = tf.image.random_contrast(img, lower = 0.9, upper = 1.1)\n    img = tf.cast(img, tf.float32)\n    img = img\/255.0\n    print(img)\n    return img\n#some global vars \nstddev  = 0.02\ndecay   = 0.9\nepsilon = 1e-4\nk_size  = [5,5]\nstrides = [2,2]\ndef gen(input, is_train):\n\n#chanel number\nc1 , c2 ,c3 ,c4 = 64, 128, 256, 512\n\nwith tf.variable_scope(\"gen\",reuse=tf.AUTO_REUSE):\n\n    #this is where it crashes\n    conv1 = tf.layers.conv2d(input,c1,k_size,strides,'SAME',\n                             kernel_initializer=tf.truncated_normal_initializer(stddev=stddev),\n                             name='conv1')\n\n    bn1 = tf.contrib.layers.batch_norm(conv1,is_training=is_train, updates_collections=None,\n                                       decay=decay,epsilon=epsilon,scope='bn1')\n    ac1 = lrelu(bn1,'ac1')\n#there is more code after this\n<\/code><\/pre>\n\n<p>trying to run it:<\/p>\n\n<pre><code>next_color, next_gray = get_next()\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsess.run(tf.local_variables_initializer())\nfoo = sess.run(next_gray)\n\nprint(f\"foo ndims : {foo.ndim}\") # --&gt; foo ndims : 4\ngen_image = gen(foo, True)\n# some more code after this \n<\/code><\/pre>\n\n<p><strong>now this rasises an error:<\/strong><\/p>\n\n<pre><code> AttributeError: 'tuple' object has no attribute 'ndims'\n     ---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-1-701a9276e633&gt; in &lt;module&gt;()\n     94 \n     95 \n---&gt; 96 train()\n\n&lt;ipython-input-1-701a9276e633&gt; in train()\n     41 #     print(foo.shape)\n     42     print(\"==========================+==============\")\n---&gt; 43     gen_image = gen(foo, True)\n     44 #     gen_image = gen(next_gray, True)\n     45     print(\"==========================+==============\")\n\n~\\Desktop\\code\\python\\image_processing\\Untitled Folder\\Untitled Folder\\testing1_2\\my_gen.py in gen(input, is_train)\n     30         conv1 = tf.layers.conv2d(input,c1,k_size,strides,'SAME',\n     31                                  kernel_initializer=tf.truncated_normal_initializer(stddev=stddev),\n---&gt; 32                                  name='conv1')\n     33 \n     34         bn1 = tf.contrib.layers.batch_norm(conv1,is_training=is_train, updates_collections=None,\n\n~\\Anaconda2\\envs\\image_rec\\lib\\site-packages\\tensorflow\\python\\layers\\convolutional.py in conv2d(inputs, filters, kernel_size, strides, padding, data_format, dilation_rate, activation, use_bias, kernel_initializer, bias_initializer, kernel_regularizer, bias_regularizer, activity_regularizer, kernel_constraint, bias_constraint, trainable, name, reuse)\n    423       _reuse=reuse,\n    424       _scope=name)\n--&gt; 425   return layer.apply(inputs)\n    426 \n    427 \n\n~\\Anaconda2\\envs\\image_rec\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py in apply(self, inputs, *args, **kwargs)\n    803       Output tensor(s).\n    804     \"\"\"\n--&gt; 805     return self.__call__(inputs, *args, **kwargs)\n    806 \n    807   def _set_learning_phase_metadata(self, inputs, outputs):\n\n~\\Anaconda2\\envs\\image_rec\\lib\\site-packages\\tensorflow\\python\\layers\\base.py in __call__(self, inputs, *args, **kwargs)\n    360 \n    361       # Actually call layer\n--&gt; 362       outputs = super(Layer, self).__call__(inputs, *args, **kwargs)\n    363 \n    364     if not context.executing_eagerly():\n\n~\\Anaconda2\\envs\\image_rec\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py in __call__(self, inputs, *args, **kwargs)\n    718 \n    719         # Check input assumptions set before layer building, e.g. input rank.\n--&gt; 720         self._assert_input_compatibility(inputs)\n    721         if input_list and self._dtype is None:\n    722           try:\n\n~\\Anaconda2\\envs\\image_rec\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py in _assert_input_compatibility(self, inputs)\n   1408           spec.min_ndim is not None or\n   1409           spec.max_ndim is not None):\n-&gt; 1410         if x.shape.ndims is None:\n   1411           raise ValueError('Input ' + str(input_index) + ' of layer ' +\n   1412                            self.name + ' is incompatible with the layer: '\n\nAttributeError: 'tuple' object has no attribute 'ndims'\n<\/code><\/pre>\n\n<p>thanks in advance <\/p>\n","answers":[{"AnswerId":"53522842","CreationDate":"2018-11-28T15:25:48.097","ParentId":null,"OwnerUserId":"9137210","Title":null,"Body":"<p>so apparently casting the out put to a tf.float32 solves the problem  <\/p>\n\n<pre><code>next_color, next_gray = get_next()\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsess.run(tf.local_variables_initializer())  \nfoo = sess.run(next_gray)\ngray_batch = tf.cast(foo, dtype = tf.float32) \n\ngen_image = gen(gray_batch, True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53466500,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network><tflearn>","CreationDate":"2018-11-25T10:14:35.507","AcceptedAnswerId":null,"OwnerUserId":8023082.0,"Title":"Can't save and load a trained CNN model for binary image classification","Body":"<p>I have built a Binary Image classifier using Convolutional Neural Networks using TensorFlow.It is running fine, however, each time it takes too long to train from scratch. So, I want to save the trained model and load it next time. I can't seem to understand how to implement <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">these<\/a> guides in my program as shown in the TensorFlow documentation. \nHere's the full code: <\/p>\n\n<pre><code># Python program to create\n# Image Classifier using CNN\n\n# Importing the required libraries\nimport cv2\nimport os\nimport numpy as np\nfrom random import shuffle\nfrom tqdm import tqdm\nfrom keras.models import Sequential\n'''Setting up the env'''\n\nTRAIN_DIR = 'D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\train\\\\'\nTEST_DIR = 'D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\test0\\\\'\ncheck_point = 'D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\chkpt\\\\'\nIMG_SIZE = 80\nLR = 1e-4\n\n'''Setting up the model which will help with tensorflow models'''\nMODEL_NAME = 'NormalVsAbnormalXRays-{}-{}.model'.format(LR, '6conv-basic')\n\n'''Labelling the dataset'''\n\n\ndef label_img(img):\n    word_label = img.split('.')[-3]\n    # DIY One hot encoder\n    if word_label == 'Nor':\n        return [1, 0]\n    elif word_label == 'Pne':\n        return [0, 1]\n    else :\n        return[0, 0]\n\n'''Creating the training data'''\n\n\ndef create_train_data():\n    # Creating an empty list where we should the store the training data\n    # after a little preprocessing of the data\n    training_data = []\n\n    # tqdm is only used for interactive loading\n    # loading the training data\n    for img in tqdm(os.listdir(TRAIN_DIR)):\n        # labeling the images\n        label = label_img(img)\n\n        path = os.path.join(TRAIN_DIR, img)\n\n        # loading the image from the path and then converting them into\n        # greyscale for easier covnet prob\n        img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n\n        # resizing the image for processing them in the covnet\n        img = cv2.resize(img, (IMG_SIZE, IMG_SIZE))\n\n        # final step-forming the training data list with numpy array of the images\n        training_data.append([np.array(img), np.array(label)])\n\n        # shuffling of the training data to preserve the random state of our data\n    shuffle(training_data)\n\n    # saving our trained data for further uses if required\n    np.save('train_data.npy', training_data)\n    return training_data\n\n\n'''Processing the given test data'''\n\n\n# Almost same as processing the traning data but\n# we dont have to label it.\ndef process_test_data():\n    testing_data = []\n    for img in tqdm(os.listdir(TEST_DIR)):\n        path = os.path.join(TEST_DIR, img)\n        img_num = img.split('.')[0]\n        img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)\n        img = cv2.resize(img, (IMG_SIZE, IMG_SIZE))\n        testing_data.append([np.array(img), img_num])\n\n    shuffle(testing_data)\n    np.save('test_data.npy', testing_data)\n    return testing_data\n\n\n'''Running the training and the testing in the dataset for our model'''\n#train_data = create_train_data()\n#test_data = process_test_data()\n\ntrain_data = np.load('train_data.npy')\ntest_data = np.load('test_data.npy')\n'''Creating the neural network using tensorflow'''\n# Importing the required libraries\nimport tflearn\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\n\nimport tensorflow as tf\nmodel = Sequential()\ntf.reset_default_graph()\n\n\nsaver = tf.train.import_meta_graph('D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\check_point-78.meta')\nconvnet = input_data(shape=[None,IMG_SIZE, IMG_SIZE, 1], name='input')\n\nconvnet = conv_2d(convnet, 32, 4, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\n\nconvnet = conv_2d(convnet, 64, 4, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\n\nconvnet = conv_2d(convnet, 128, 4, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = conv_2d(convnet, 64, 4, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = conv_2d(convnet, 32, 4, activation='relu')\nconvnet = max_pool_2d(convnet, 2)\n\nconvnet = fully_connected(convnet, 1024, activation='relu')\nconvnet = dropout(convnet, 0.3)\n\nconvnet = fully_connected(convnet, 2, activation='softmax')\nconvnet = regression(convnet, optimizer='adam', learning_rate=LR,\n                     loss='categorical_crossentropy', name='targets')\n\nmodel = tflearn.DNN(convnet, tensorboard_dir='log', checkpoint_path='check_point',best_checkpoint_path= 'check_point',max_checkpoints= 5)\n\n# Splitting the testing data and training data\ntrain = train_data\ntest = train_data\n\n'''Setting up the features and lables'''\n# X-Features &amp; Y-Labels\n\nX = np.array([i[0] for i in train]).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\nY = [i[1] for i in train]\ntest_x = np.array([i[0] for i in test]).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\ntest_y = [i[1] for i in test]\n\n'''Fitting the data into our model'''\n# epoch = 40 taken\nmodel.fit({'input': X}, {'targets': Y}, n_epoch=1,\n          validation_set=0.05,\n          snapshot_step=500, show_metric=True, run_id=MODEL_NAME)\nmodel.save(MODEL_NAME)\n\n'''Testing the data'''\nimport matplotlib.pyplot as plt\n\n# if you need to create the data:\n# test_data = process_test_data()\n# if you already have some saved:\ntest_data = np.load('test_data.npy')\n\nfig = plt.figure(figsize=(80,80))\n\nfor num, data in enumerate(test_data[:1]):\n\n\n    img_num = data[1]\n    img_data = data[0]\n\n    y = fig.add_subplot(1, 1, num + 1)\n    orig = img_data\n    data = img_data.reshape(IMG_SIZE, IMG_SIZE, 1)\n\n    # model_out = model.predict([data])[0]\n    model_out = model.predict([data])[0]\n\n    if np.argmax(model_out) == 1:\n        str_label = 'Abnormal'\n    else:\n        str_label = 'Normal'\n\n    y.imshow(orig, cmap='gray')\n    plt.title(str_label,fontsize=20)\n    y.axes.get_xaxis().set_visible(False)\n    y.axes.get_yaxis().set_visible(False)\n plt.show()\n<\/code><\/pre>\n\n<p>I have tried to use saver = <code>tf.train.import_meta_graph('D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\check_point-78.meta')<\/code> to import the graph but I get this error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/Project\/Final_Project\/chest_xray\/Final_CNN.py\", line 104, in &lt;module&gt;\n    saver = tf.train.import_meta_graph('D:\\\\Project\\\\Final_Project\\\\chest_xray\\\\check_point-78.meta')\n  File \"C:\\Users\\waqar\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1674, in import_meta_graph\n    meta_graph_or_file, clear_devices, import_scope, **kwargs)[0]\n  File \"C:\\Users\\waqar\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1696, in _import_meta_graph_with_return_elements\n    **kwargs))\n  File \"C:\\Users\\waqar\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\meta_graph.py\", line 852, in import_scoped_meta_graph_with_return_elements\n    ops.prepend_name_scope(value, scope_to_prepend_to_names))\n  File \"C:\\Users\\waqar\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3490, in as_graph_element\n    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)\n  File \"C:\\Users\\waqar\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3550, in _as_graph_element_locked\n    \"graph.\" % repr(name))\nKeyError: \"The name 'Adam' refers to an Operation not in the graph.\"\n<\/code><\/pre>\n\n<p>Process finished with exit code 1<\/p>\n","answers":[]}
+{"QuestionId":53467011,"AnswerCount":1,"Tags":"<python><numpy><pytorch>","CreationDate":"2018-11-25T11:33:35.067","AcceptedAnswerId":53467161.0,"OwnerUserId":6299165.0,"Title":"PyTorch - multiplying tensor with scalar results in zero vector","Body":"<p>I have no idea why the result is all 0 with tensor. Anything wrong here?<\/p>\n\n<pre><code>&gt;&gt;&gt; import torch\n&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; import math\n\n&gt;&gt;&gt; torch.__version__\n'0.4.1'\n&gt;&gt;&gt; np.__version__\n'1.15.4'\n\n&gt;&gt;&gt; torch.arange(0, 10, 2) *-(math.log(10000.0) \/ 10)\ntensor([0, 0, 0, 0, 0])\n&gt;&gt;&gt; np.arange(0, 10, 2) *-(math.log(10000.0) \/ 10)\narray([-0.        , -1.84206807, -3.68413615, -5.52620422, -7.3682723 ])\n\n&gt;&gt;&gt; torch.arange(0, 10, 2)\ntensor([0, 2, 4, 6, 8])\n&gt;&gt;&gt; np.arange(0, 10, 2)\narray([0, 2, 4, 6, 8])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53467161","CreationDate":"2018-11-25T11:54:42.630","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>As written in the comment when using <strong>0.4.0<\/strong> get the same results as with numpy:<\/p>\n\n<pre><code>tensor([-0.0000, -1.8421, -3.6841, -5.5262, -7.3683])\n<\/code><\/pre>\n\n<p>However with <code>0.4.1<\/code> I'm getting a zero vector too. <\/p>\n\n<p>The reason for this is that <code>torch.arange(0, 10, 2)<\/code> returns a tensor of type <code>float<\/code> for <strong>0.4.0<\/strong> while it returns a tensor of type <code>long<\/code> for <strong>0.4.1<\/strong>. <\/p>\n\n<p>So casting your tensor to <code>float<\/code> should work for you:<\/p>\n\n<pre><code>torch.arange(0, 10, 2).float() *-(math.log(10000.0) \/ 10)\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p>Multiplying <code>long<\/code> and <code>float<\/code> works by heavy rounding, as the result is still a tensor of type <code>long<\/code>. So when converting a <code>FloatTensor<\/code> to a <code>LongTensor<\/code> values between -1 and 1 will be rounded to 0.<\/p>\n\n<p>Since <code>-(math.log(10000.0) \/ 10)<\/code> results in <code>-0.9210340371976183<\/code> your result is <code>0<\/code>. So effectively <code>-0.9210340371976183<\/code> is converted to type <code>long<\/code> before multiplying. But when converting it will be round down to <code>0<\/code>, see this example:<\/p>\n\n<pre><code>t = torch.tensor((-(math.log(10000.0) \/ 10)))\nprint('FloatTensor:', t)\nprint('Converted to Long:', t.long())\n<\/code><\/pre>\n\n<p>Outout:<\/p>\n\n<pre><code>FloatTensor: tensor(-0.9210)\nConverted to Long: tensor(0)\n<\/code><\/pre>\n\n<p>Thus:<\/p>\n\n<pre><code>torch.arange(0, 10, 2).float() *-(math.log(10000.0) \/ 10)\n<\/code><\/pre>\n\n<p>becomes:<\/p>\n\n<pre><code>torch.arange(0, 10, 2).float() * 0\n<\/code><\/pre>\n\n<p>Therefore you get a tensor of zeros as result.<\/p>\n\n<p><br><\/p>\n\n<hr>\n\n<p><em>Some more examples:<\/em><\/p>\n\n<p>If you multiply it with a value between 1 and 2, lets say 1.7, it will always been rounded down to 1:<\/p>\n\n<pre><code>t = torch.tensor(range(5), dtype=torch.long)\nprint(t)\nprint(t * 1.7)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor([ 0,  1,  2,  3,  4])\ntensor([ 0,  1,  2,  3,  4])\n<\/code><\/pre>\n\n<p>And similarly when multiplying with <code>2.7<\/code> results in an effective multiplication of <code>2<\/code>:<\/p>\n\n<pre><code>t = torch.tensor(range(5), dtype=torch.long)\nprint(t)\nprint(t * 2.7)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor([ 0,  1,  2,  3,  4])\ntensor([ 0,  2,  4,  6,  8])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53467109,"AnswerCount":2,"Tags":"<python><tensorflow><keras><sentiment-analysis>","CreationDate":"2018-11-25T11:47:16.510","AcceptedAnswerId":null,"OwnerUserId":9417884.0,"Title":"Significantly lower accuracy while using class_weight for imbalanced dataset in keras","Body":"<p>I have rather unintuitive problem. I am doing the sentiment analysis on Amazon Book reviews and the data set is heavily imbalanced. Positive reviews are almost 10 times the negative reviews, accuracy for both training and testing are around 90% (with imbalanced dataset). However, when I try to balance the dataset with the help of <code>class_weight = {0:10 , 1:1}<\/code> both training and testing accuracy drops to around 65%. Again if I do <code>class_weight = {0:1 , 1:10}<\/code> accuracy booms again, so apparently I am setting the <code>class_weight<\/code> wrong but as I understood because the number of positive reviews(1) are 10 times the number of negative reviews(0), shouldn't the <code>class_weight<\/code> be set as {0:10 , 1:1} ?<\/p>\n\n<p>This is how I classify training and testing data:<\/p>\n\n<pre><code>x_train, x_test, y_train, y_test = train_test_split(sequences,labels, test_size = 0.33, random_state = 42)\n<\/code><\/pre>\n\n<p>This is my model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(max_words, embedding_dim))\nmodel.add(Dropout(0.5))\nmodel.add(LSTM(32, dropout=0.5, recurrent_dropout=0.5))\nmodel.add(Dense(2, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy',metrics=['acc'])\nmodel.fit(x_train, y_train, epochs=10, batch_size=320, class_weight = {0:1 , 1:10})\n<\/code><\/pre>\n","answers":[{"AnswerId":"53467469","CreationDate":"2018-11-25T12:32:17.943","ParentId":null,"OwnerUserId":"6347629","Title":null,"Body":"<p>Setting higher class weight for a class with lower frequency in a data set is the right approach. More than Accuracy you can look at other more <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/model_evaluation.html#common-cases-predefined-values\" rel=\"nofollow noreferrer\">useful metrics<\/a> like Precision, Recall, F1 score, auc_roc score (Concordance), Confusion matrix to actually understand what the model is learning. <\/p>\n"},{"AnswerId":"53468502","CreationDate":"2018-11-25T14:30:17.350","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>Of course if you do not balance the loss you'll get better accuracy than if you balance it. Actually this is the reason for balancing. Otherwise the model that predict only positive class for all reviews will give you 90% accuracy. But this model is useless. Accuracy is generally bad metric for such strongly unbalanced datasets. Use F1 instead and you'll see that unbalanced model gives much worse F1. <\/p>\n"}]}
+{"QuestionId":53467215,"AnswerCount":2,"Tags":"<python><numpy><pytorch>","CreationDate":"2018-11-25T12:01:05.587","AcceptedAnswerId":null,"OwnerUserId":7424807.0,"Title":"Convert Pytorch Tensor to Numpy Array using Cuda","Body":"<p>I would like to convert a Pytorch tensor to numpy array using cuda:<\/p>\n\n<p>this is the code line while not using cuda:<\/p>\n\n<blockquote>\n  <p>A = self.tensor.weight.data.numpy()<\/p>\n<\/blockquote>\n\n<p>How can I do the same operation using cuda?\nAccording to this: <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-transform-variable-into-numpy\/104\/3\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-to-transform-variable-into-numpy\/104\/3<\/a>\nit seems:<\/p>\n\n<blockquote>\n  <p>A = self.tensor.weight.data.cpu().numpy()<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53469846","CreationDate":"2018-11-25T17:08:28.030","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>If the tensor is on <code>gpu<\/code> or <code>cuda<\/code> as you say. <\/p>\n\n<p>You can use <code>self.tensor.weight.data.cpu().numpy()<\/code> It will copy the tensor to <code>cpu<\/code> and convert it to numpy array.<\/p>\n\n<p>If the tensor is on <code>cpu<\/code> already you can do <code>self.tensor.weight.data.numpy()<\/code> as you correctly figured out, but you can also do <code>self.tensor.weight.data.cpu().numpy()<\/code> in this case since tensor is already on <code>cpu<\/code>, <code>.cpu()<\/code> operation will have no effect. and this could be used as a device-agnostic way to convert the tensor to numpy array.<\/p>\n"},{"AnswerId":"54585632","CreationDate":"2019-02-08T03:27:48.753","ParentId":null,"OwnerUserId":"9273406","Title":null,"Body":"<p>I believe you also have to use <strong>.detach()<\/strong>. I had to convert my Tensor to a numpy array on Colab which uses CUDA and GPU. I did it like the following:<\/p>\n\n<pre><code>embedding = learn.model.u_weight\n\nembedding_list = list(range(0, 64382))\n\ninput = torch.cuda.LongTensor(embedding_list)\ntensor_array = embedding(input)\n# the output of the line bwlow is a numpy array\ntensor_array.cpu().detach().numpy() \n<\/code><\/pre>\n"}]}
+{"QuestionId":53467460,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-25T12:31:16.843","AcceptedAnswerId":null,"OwnerUserId":8005524.0,"Title":"one of the variables needed for gradient computation has been modified by an inplace operation","Body":"<p>Here is my LossFunction, when I use this function, it will make this error.<br\/>\nAnd I have tested that using the <code>nn.L1Loss()<\/code> instead of my LossFunction, and the network is ok.<br\/>\nwhat should I do? Thanks for your help!<\/p>\n\n<pre><code>class LossV1(nn.Module):\n    def __init__(self,weight=1,pos_weight=1,scale_factor=2.5):\n        super(LossV1,self).__init__()\n        self.weight = weight\n        self.pos_weight = pos_weight\n        self.scale_factor = scale_factor\n\n    def forward(self,pred,truth):\n        objmask = torch.tensor(truth[:,:6,:,:],dtype=torch.float32,requires_grad=False)\n        #\u6ca1\u6709\u7269\u4f53\u7684Boxes\uff0c\u7f6e\u4fe1\u5ea6\u635f\u5931*0.4\n        objmask[objmask&lt;0.65] = 0.4\n        #\u8f85\u52a9Boxes\uff0c\u7cfb\u65700.8\n        objmask[(objmask&gt;0.649)*(objmask&lt;0.949)] = 0.8\n        objLoss = torch.sum(objmask*self.myBCEWithLogitsLoss(pred[:,:6,:,:],truth[:,:6,:,:]))\n        #\u6ca1\u6709\u7269\u4f53\u7684Boxes\uff0c\u53ea\u8ba1\u7b97\u7f6e\u4fe1\u5ea6\u635f\u5931\n        objmask[objmask&lt;0.41] = 0\n        personLoss = torch.sum(objmask*self.myBCEWithLogitsLoss(pred[:,6:12,:,:],truth[:,6:12,:,:]))\n        carLoss = torch.sum(objmask*self.myBCEWithLogitsLoss(pred[:,12:18,:,:],truth[:,12:18,:,:]))\n        wLoss = torch.sum(objmask*self.myL2Loss(pred[:,18:24,:,:],truth[:,18:24,:,:]))\n        hLoss = torch.sum(objmask*self.myL2Loss(pred[:,24:,:,:],truth[:,24:,:,:]))\n        return objLoss+personLoss+carLoss+wLoss+hLoss\n\n    def myBCEWithLogitsLoss(self,x,y):\n        #pos_weight&gt;1\u589e\u52a0\u53ec\u56de,pos_weight&lt;1\u63d0\u9ad8\u7cbe\u5ea6\n        return -self.weight*(self.pos_weight*y*torch.log(torch.sigmoid(x))+(1-y)*torch.log(1-torch.sigmoid(x)))\n\n    def myL2Loss(self,x,y):\n        return torch.pow(self.scale_factor*torch.sigmoid(x\/self.scale_factor) - y,2)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53475512","CreationDate":"2018-11-26T06:03:47.100","ParentId":null,"OwnerUserId":"8005524","Title":null,"Body":"<p>I just remove the <code>objmask<\/code>,and compute it in my generation function,then pass it on to LossFunction with truth label,and the network works.<br\/>\nI can't understand I have already done that making the <code>requires_grad=False<\/code>,why pytroch still compute the gradient.<\/p>\n"}]}
+{"QuestionId":53467764,"AnswerCount":1,"Tags":"<tensorflow><keras><categorical-data><embedding>","CreationDate":"2018-11-25T13:08:08.580","AcceptedAnswerId":53468127.0,"OwnerUserId":5927701.0,"Title":"Extract the output of the embedding layer","Body":"<p>I am trying to build a regression model, for which I have a nominal variable with very high cardinality. I am trying to get the categorical embedding of the column.<\/p>\n\n<p>Input:<\/p>\n\n<pre><code>df[\"nominal_column\"]\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>the embeddings of the column.\n<\/code><\/pre>\n\n<p>I want to use the op of the embedding column alone since I would require that as a input to my traditional regression model. Is there a way to extract that output alone. <\/p>\n\n<p>P.S I am not asking for code, any suggestion on the approach would be great.<\/p>\n","answers":[{"AnswerId":"53468127","CreationDate":"2018-11-25T13:48:26.390","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>If the embedding is part of the model and you train it, then you can use functional API of keras to get output of any intermediate operation in your graph:<\/p>\n\n<pre><code>x=Input((number_of_categories,))\ny=Embedding(parameters_of_your_embeddings)(x)\noutput=Rest_of_your_model()(y)\nmodel=Model(inputs=[x],outputs=[output,y])\n<\/code><\/pre>\n\n<p>if you do it before you train the model, you'll have to define custom loss function, that deals only with part of the output. The other way is to train the model with just one output, then create identical model with two outputs and set the weights of the second model from the trained one.<\/p>\n\n<p>If you want to get the embedding matrix from your model, you can just use method get_weights of the embedding layer which returns the weights in numpy array.<\/p>\n"}]}
+{"QuestionId":53467817,"AnswerCount":1,"Tags":"<apache-spark><hadoop><tensorflow><hdfs><distribution>","CreationDate":"2018-11-25T13:13:09.933","AcceptedAnswerId":null,"OwnerUserId":8110337.0,"Title":"Save the prediction output of Tensorflow model into hdfs file","Body":"<p>I have a trained a tf model and I want to apply it to big dataset in hdfs which is about billion of samples. The main point is I need to write the prediction of tf model into hdfs file. However I can't find the relative API in tensorflow about how to save data in hdfs file, only find the api about reading hdfs file <br\/>\nUntil now the way I did it is to save the trained tf model into pb file in local and then load the pb file using Java api in spark or Mapreduce code. The problem of both spark or mapreduce is the running speed is very slow and failed with exceeds memory error. \nHere is my demo:<\/p>\n\n<pre><code>public class TF_model implements Serializable{\n\npublic Session session;\n\npublic TF_model(String model_path){\n    try{\n        Graph graph = new Graph();\n        InputStream stream = this.getClass().getClassLoader().getResourceAsStream(model_path);\n        byte[] graphBytes = IOUtils.toByteArray(stream);\n        graph.importGraphDef(graphBytes);\n        this.session = new Session(graph);\n    }\n    catch (Exception e){\n        System.out.println(\"failed to load tensorflow model\");\n    }\n}\n\/\/ this is the function to predict a sample in hdfs\npublic int[][] predict(int[] token_id_array){\n    Tensor z = session.runner()\n            .feed(\"words_ids_placeholder\", Tensor.create(new int[][]{token_id_array}))\n            .fetch(\"softmax_prediction\").run().get(0);\n    double[][][] softmax_prediction = new double[1][token_id_array.length][2];\n    z.copyTo(softmax_prediction);\n    return softmax_prediction[0];\n}}\n<\/code><\/pre>\n\n<p>below is my spark code:<\/p>\n\n<pre><code>val rdd = spark.sparkContext.textFile(file_path)\nval predct_result= rdd.mapPartitions(pa=&gt;{\n    val tf_model = new TF_model(\"model.pb\")\n    pa.map(line=&gt;{\n      val transformed = transform(line) \/\/ omitted the transform code\n      val rs = tf_model .predict(transformed)\n      rs\n    })\n  })\n<\/code><\/pre>\n\n<p>I also tried tensorflow deployed in hadoop, but can't find a way to write big dataset into HDFS.<\/p>\n","answers":[{"AnswerId":"53808574","CreationDate":"2018-12-17T03:20:52.607","ParentId":null,"OwnerUserId":"10799377","Title":null,"Body":"<p>You may read model file from hdfs one time, then use sc.broadcast your bytes array of your graph to partitions. Finally, start load graph and predict. Just to avoid read file multiple time from hdfs.<\/p>\n"}]}
+{"QuestionId":53468124,"AnswerCount":1,"Tags":"<docker><tensorflow><deployment><cloud><devops>","CreationDate":"2018-11-25T13:48:11.483","AcceptedAnswerId":null,"OwnerUserId":6267719.0,"Title":"Contingent Tensorflow Model Deployment","Body":"<p>I am looking to deploy a Tensorflow model so that it can be accessed for inference using a RESTful API.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/serving\/\" rel=\"nofollow noreferrer\">Tensorflow Serving<\/a> looks like the ideal tools for this job.  I got some MWEs running using Tensorflow Serving now.  I also have my model working.<\/p>\n\n<p>Users will access the model through a Flask website, and upload data via that.  The Flask website would send the data to the model container, and return the result.<\/p>\n\n<p>However, for my use case, the model will be accessed at most a couple of times a month, and latency is not a massive concern.  Having the model running on a cloud server 24\/7 would represent a large waste of resources.  <\/p>\n\n<p>Are there any tools, workflows, or guides, that would allow the model to be deployed then shutdown by users via the web interface?  They would be willing to wait the time for an instance to be spun up.<\/p>\n\n<p>I'm cloud service agnostic, and not even married to the Tensorflow Serving system.  <a href=\"https:\/\/stackoverflow.com\/questions\/52621627\/tensorflow-model-deployment-in-gcp-without-tensorflow-serving?rq=1\">This question<\/a> raised similar concerns, but was not answered.<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"55639180","CreationDate":"2019-04-11T18:39:04.680","ParentId":null,"OwnerUserId":"3286035","Title":null,"Body":"<p>You could make the prediction on the flask app instead of sending data to the tensorflow serving server for the prediction.<\/p>\n\n<p>This is a basic example:<\/p>\n\n<pre><code>@app.route('\/predict\/', methods=['GET','POST'])\ndef predict():\n    # get data from drawing canvas and save as image\n    data = parseData(request.get_data())\n\n    graph = tf.get_default_graph()\n    with graph.as_default():\n        out = model.predict(x)\n        response = np.array_str(np.argmax(out, axis=1))\n        return response \n<\/code><\/pre>\n\n<p>We also build an open source toolkit for packaging models for packaging and deploying models, <a href=\"http:\/\/www.github.com\/bentoml\/bentoml\" rel=\"nofollow noreferrer\">www.github.com\/bentoml\/bentoml<\/a>. With our toolkit, you can export your model to a file archive and import it as a python module in your flask app. This way, you will share the resource with your flask app when you make predictions.<\/p>\n"}]}
+{"QuestionId":53469433,"AnswerCount":2,"Tags":"<keras><tensorflow-serving><resnet>","CreationDate":"2018-11-25T16:18:48.970","AcceptedAnswerId":53724447.0,"OwnerUserId":727219.0,"Title":"Tensorflow Serving on pretrained Keras ResNet50 model returning always same predictions","Body":"<p>I'm using the following code to export a pre-trained ResNet50 keras' model to tensorflow, for tensorflow-serving:<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\nfrom keras import backend as K\nK.set_session(sess)\nK.set_learning_phase(0)\n\n# Modelo resnet con pesos entrenados en imagenet\nfrom keras.applications.resnet50 import ResNet50\nmodel = ResNet50(weights='imagenet')\n\n# exportar en tensorflow\nimport os\nversion_number = max([ int(x) for x in os.listdir('.\/resnet-classifier') ]) + 1\nexport_path = '.\/resnet-classifier\/{}'.format(version_number)\nwith tf.keras.backend.get_session() as sess:\n    init_op = tf.global_variables_initializer()\n    sess.run(init_op)\n    tf.saved_model.simple_save(sess, export_path,\n            inputs=dict(input_image=model.input),\n            outputs={t.name:t for t in model.outputs}\n    )\n<\/code><\/pre>\n\n<p>I tried some variations of the above, all of them with the same results (same prediction while served by tensorflow serving).<\/p>\n\n<p>Then i run tensorflow-serving like:<\/p>\n\n<pre><code>docker run -p 8501:8501 \\\n  -v .\/resnet-classifier:\/models\/resnet-classifier \\\n  -e MODEL_NAME=resnet-classifier -e MODEL_BASE_PATH=\/models \\\n  -t tensorflow\/serving\n<\/code><\/pre>\n\n<p>Finally, i'm using the following function to make predictions against tensorflow serving:<\/p>\n\n<pre><code>def imagepath_to_tfserving_payload(img_path):\n    import numpy as np\n    from keras.preprocessing import image\n    from keras.applications.resnet50 import preprocess_input\n    img = image.img_to_array(image.load_img(img_path, target_size=(224, 224)))\n    X = np.expand_dims(img, axis=0).astype('float32')\n    X = preprocess_input(X)\n    payload = dict(instances=X.tolist())\n    payload = json.dumps(payload)\n    return payload\n\ndef tfserving_predict(image_payload, url=None):\n    import requests\n    if url is None:\n        url = 'http:\/\/localhost:8501\/v1\/models\/resnet-classifier:predict'\n    r = requests.post(url, data=image_payload)\n    pred_json = json.loads(r.content.decode('utf-8'))\n    from keras.applications.resnet50 import decode_predictions\n    predictions = decode_predictions(np.asarray(pred_json['predictions']), top=3)[0]\n    return predictions\n<\/code><\/pre>\n\n<p>Then i use both functions above from an ipython shell to select random imagenes from imagenet's val set, which i've stored locally. The problem is that tensorflow serving is always returning the same prediction for all images i send.<\/p>\n\n<p>Each time i export the model with the first script above, i'm getting slightly different classes, with '1' confidence for the first class and '0' for others, for example:<\/p>\n\n<pre><code># Serialization 1, in .\/resnet-classifier\/1 always returning:\n[\n  [\n    \"n07745940\",\n    \"strawberry\",\n    1.0\n  ],\n  [\n    \"n02104029\",\n    \"kuvasz\",\n    1.4013e-36\n  ],\n  [\n    \"n15075141\",\n    \"toilet_tissue\",\n    0.0\n  ]\n]\n\n# Serialization 2, in .\/resnet-classifier\/2 always returning:\n[\n  [\n    \"n01530575\",\n    \"brambling\",\n    1.0\n  ],\n  [\n    \"n15075141\",\n    \"toilet_tissue\",\n    0.0\n  ],\n  [\n    \"n02319095\",\n    \"sea_urchin\",\n    0.0\n  ]\n]\n<\/code><\/pre>\n\n<p>This might be related with <a href=\"https:\/\/stackoverflow.com\/questions\/47912161\/tensorflow-serving-model-return-always-the-same-prediction\">Tensorflow : serving model return always the same prediction<\/a>, but i don't know how the answers there (no accepted one) may help.<\/p>\n\n<p>Anybody knows what's wrong above, and how to fix it?<\/p>\n","answers":[{"AnswerId":"53470701","CreationDate":"2018-11-25T18:41:54.947","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>I had sometimes this kind of problems when forgot to normalize image. I think resnet accepts images in the format of float numbers between 0. and 1. (or maybe -1. to 1.). I don't know what preprocess_input function does but you can check if it returns array in the expected format.<\/p>\n"},{"AnswerId":"53724447","CreationDate":"2018-12-11T12:40:50.723","ParentId":null,"OwnerUserId":"727219","Title":null,"Body":"<p>I've found that calling sess.run(tf.global_variables_initializer()) overrides pretrained weights, clue found at <a href=\"http:\/\/zachmoshe.com\/2017\/11\/11\/use-keras-models-with-tf.html\" rel=\"nofollow noreferrer\">http:\/\/zachmoshe.com\/2017\/11\/11\/use-keras-models-with-tf.html<\/a>.<\/p>\n\n<p>The solution for me was really simple, just change the first block of code in the original question by the following, which calls tf.global_variables_initializer() <em>before<\/em> model instantiation \/ weight load:<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nfrom keras import backend as K\nK.set_session(sess)\nK.set_learning_phase(0)\n\n# Modelo resnet con pesos entrenados en imagenet\nfrom keras.applications.resnet50 import ResNet50\nmodel = ResNet50(weights='imagenet')\n\n# exportar en tensorflow\nimport os\nversions = [ int(x) for x in os.listdir('.\/resnet-classifier') ]\nversion_number = max(versions) + 1 if versions else 1\nexport_path = '.\/resnet-classifier\/{}'.format(version_number)\n\ntf.saved_model.simple_save(sess, export_path,\n        inputs=dict(input_image=model.input),\n        outputs={t.name:t for t in model.outputs}\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53469998,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2018-11-25T17:25:47.040","AcceptedAnswerId":53472614.0,"OwnerUserId":4835204.0,"Title":"TensorFlow C++ what does \"NOLINT(build\/namespaces)\" mean?","Body":"<p>Odd question here, at various places in the C++ code portions of TensorFlow on Github there is the following comment:<\/p>\n\n<pre><code>\/\/ NOLINT(build\/namespaces)\n<\/code><\/pre>\n\n<p>For example, the C++ label image example has this comment:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc#L120\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc#L120<\/a><\/p>\n\n<pre><code>Status ReadTensorFromImageFile(const string&amp; file_name, const int input_height,\n                               const int input_width, const float input_mean,\n                               const float input_std,\n                               std::vector&lt;Tensor&gt;* out_tensors) {\n  auto root = tensorflow::Scope::NewRootScope();\n  using namespace ::tensorflow::ops;  \/\/ NOLINT(build\/namespaces)\n\n  string input_name = \"file_reader\";\n  string output_name = \"normalized\";\n  .\n  .\n  .\n<\/code><\/pre>\n\n<p>Can anybody explain what <code>\/\/ NOLINT(build\/namespaces)<\/code> means ?? I would try to research it more myself but I'm not even sure where to begin.<\/p>\n","answers":[{"AnswerId":"53472614","CreationDate":"2018-11-25T22:23:43.677","ParentId":null,"OwnerUserId":"206302","Title":null,"Body":"<p>Linters are tools to check the code for style and correctness in a static way. One of those tools used by TensorFlow developers and other is <a href=\"http:\/\/clang.llvm.org\/extra\/clang-tidy\/\" rel=\"nofollow noreferrer\">clang-tidy<\/a>.<\/p>\n\n<p>While those tools are very useful to spot mistakes, sometimes the general checks don't apply (for instance in generated code, highly optimized code or when purposely using outdated APIs for backwards compatibility) in such cases an annotation like <code>\/\/NOLINT<\/code> can be used to tell the linter not to report errors,to keep false-positives low and the results  meaningful.<\/p>\n\n<p>Each of these places should state a reason, so developers know what is being ignored and why, an should regularly be reviewed, so changes due to updates are detected.<\/p>\n\n<p>See also the section on suppression: <a href=\"http:\/\/clang.llvm.org\/extra\/clang-tidy\/#suppressing-undesired-diagnostics\" rel=\"nofollow noreferrer\">http:\/\/clang.llvm.org\/extra\/clang-tidy\/#suppressing-undesired-diagnostics<\/a><\/p>\n"}]}
+{"QuestionId":53470029,"AnswerCount":1,"Tags":"<python><keras><keras-2>","CreationDate":"2018-11-25T17:28:33.363","AcceptedAnswerId":null,"OwnerUserId":3837778.0,"Title":"@property: AttributeError: can't set attribute","Body":"<p>I am trying this <a href=\"https:\/\/github.com\/bstriner\/keras-adversarial\/blob\/master\/examples\/example_gan_cifar10.py\" rel=\"nofollow noreferrer\">example<\/a> and got the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-1-9f46197aa363&gt; in &lt;module&gt;()\n    206 \n    207 if __name__ == \"__main__\":\n--&gt; 208     main()\n\n&lt;ipython-input-1-9f46197aa363&gt; in main()\n    202                 opt_d=Adam(1e-3, decay=1e-5),\n    203                 nb_epoch=100, generator=generator, discriminator=discriminator,\n--&gt; 204                 latent_dim=latent_dim)\n    205 \n    206 \n\n&lt;ipython-input-1-9f46197aa363&gt; in example_gan(adversarial_optimizer, path, opt_g, opt_d, nb_epoch, generator, discriminator, latent_dim, targets, loss)\n    155     model = AdversarialModel(base_model=gan,\n    156                              player_params=[generator.trainable_weights, discriminator.trainable_weights],\n--&gt; 157                              player_names=[\"generator\", \"discriminator\"])\n    158     model.adversarial_compile(adversarial_optimizer=adversarial_optimizer,\n    159                               player_optimizers=[opt_g, opt_d],\n\n\/Users\/cpei\/ebay\/dev\/lib\/python2.7\/site-packages\/keras_adversarial-0.0.3-py2.7.egg\/keras_adversarial\/adversarial_model.pyc in __init__(self, player_params, base_model, player_models, player_names)\n     45             raise ValueError(\"Specify base_model or player_models, not both\")\n     46         if base_model is not None:\n---&gt; 47             self.layers = []\n     48             for i in range(self.player_count):\n     49                 # duplicate base model\n\n\/Users\/cpei\/ebay\/dev\/lib\/python2.7\/site-packages\/keras\/engine\/network.pyc in __setattr__(self, name, value)\n    314                 if value not in self._layers:\n    315                     self._layers.append(value)\n--&gt; 316         super(Network, self).__setattr__(name, value)\n    317 \n    318     @property\n\nAttributeError: can't set attribute\n<\/code><\/pre>\n\n<p>Any comments? Thanks<\/p>\n","answers":[{"AnswerId":"55185132","CreationDate":"2019-03-15T14:44:22.170","ParentId":null,"OwnerUserId":"11209233","Title":null,"Body":"<p>Check your <code>Keras<\/code> version. I experienced the same issue with <code>Keras 2.2.4<\/code> After I installed <code>Keras==2.1.1<\/code> it started working well.<\/p>\n"}]}
+{"QuestionId":53470309,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-25T18:00:08.363","AcceptedAnswerId":null,"OwnerUserId":10702857.0,"Title":"Using tf.batch_scatter_add in Keras","Body":"<p>I'm looking for a way to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_add\" rel=\"nofollow noreferrer\">tf.scatter_add<\/a> with Keras batches. \nShape of outputs: <code>(?, 1000)<\/code> and shapes of indices and updates are <code>(?, 100)<\/code> each. <\/p>\n\n<p>Try1: Using Keras tensors <\/p>\n\n<pre><code>vals = tf.scatter_add(outputs, indices, updates)\n<\/code><\/pre>\n\n<p>This throws an error: <\/p>\n\n<blockquote>\n  <p>'Tensor' object has no attribute '_lazy_read'<\/p>\n<\/blockquote>\n\n<p>Try2: Tried using k.variable which should be updatable <\/p>\n\n<pre><code>vals = K.variable(outputs)\nvals = tf.scatter_add(vals, inputs[1], inputs[2]) \n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: initial_value must have a shape specified: <br>\n      Tensor(\"scatter_add_43\/zeros_like:0\", shape=(?, 1000), dtype=float32))<\/p>\n<\/blockquote>\n\n<p>Any clues? <code>Scatter_add<\/code> and <code>batch_scatter_add<\/code> result in the same errors. Will I need to write a custom layer for this? Seems like even that will run into one of the above errors. <\/p>\n","answers":[{"AnswerId":"55066266","CreationDate":"2019-03-08T15:25:34.310","ParentId":null,"OwnerUserId":"2936263","Title":null,"Body":"<p>You don't need a Layer for this. You can use <code>tf.tensor_scatter_add<\/code> to update <strong>Tensors<\/strong> directly or <code>scatter_nd_add<\/code> (method from <code>tf.Variable<\/code>) to update <strong>Variables<\/strong>. <\/p>\n\n<p>References:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tensor_scatter_add\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tensor_scatter_add<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable#scatter_nd_add\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable#scatter_nd_add<\/a><\/p>\n\n<p>Unfortunately, Tensorflow API does not have a <code>tf.batched_scatter_nd_add<\/code> yet. So we have to try mimic this behaviour. Both methods work with individual or slices. In order to work with batches, we need to transform the indices to the expected multi-dimensional expected.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\nbatch_size = 2\nvocab_size = 6\nseq_len = 3\n\npredictions = tf.Variable(tf.zeros([batch_size, vocab_size], dtype=tf.float32))\n\nindices = tf.constant([[0, 2, 4], [1, 3, 5]], dtype=tf.int32)\nupdates = tf.ones([batch_size, seq_len], dtype=tf.float32)\n\nbatched_indices = [[i, j.numpy()] for i, indexes in enumerate(indices) for j in indexes]\nbatched_updates = tf.reshape(updates, [-1])  # flatten\n\npredictions.scatter_nd_add(indices=batched_indices, updates=batched_updates)\n\nprint(predictions)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>&lt;tf.Variable 'Variable:0' shape=(2, 6) dtype=float32, numpy=\narray([[1., 0., 1., 0., 1., 0.],\n       [0., 1., 0., 1., 0., 1.]], dtype=float32)&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":53470688,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><image-recognition>","CreationDate":"2018-11-25T18:40:32.640","AcceptedAnswerId":53472446.0,"OwnerUserId":5357221.0,"Title":"Give an example visual recognition task where a fully connected network would be more suitable than a convolution neural networks","Body":"<p>I know CNN has a lot of good features like weight sharing, save memory and feature extracting. However, this question makes me very confused. Is there any possible situation that fully connected network better than CNN? Why?<\/p>\n\n<p>Thanks a lot guys! <\/p>\n","answers":[{"AnswerId":"53472446","CreationDate":"2018-11-25T22:01:30.723","ParentId":null,"OwnerUserId":"1866656","Title":null,"Body":"<blockquote>\n  <p>Is there any possible situation that fully connected network better than CNN? <\/p>\n<\/blockquote>\n\n<p>Well, I think we should first define what we mean by \"better\". Accuracy and precision are not the only things to consider: computational time, degrees of freedom and difficulty of the optimization should also be taken into account.<\/p>\n\n<p>First, consider an input of size <code>h*w*c<\/code>. Feeding this input to a convolutional layer with <code>F<\/code> featuremaps and kernel size <code>s<\/code> will result in at about <code>F*s*s*c<\/code> learnable parameters (assuming there are no constraints on the ranks of the convolutions, otherwise we even have less parameters.). Feeding the same input into a fully connected layer with the same number of featuremaps will result in <code>F*d_1*d_2*w*h*c<\/code>, (where <code>d_1,d_2<\/code> are the dimensions of each featuremap) which is clearly in the order of billions of learnable parameters given any input image with decent resolution.<\/p>\n\n<p>While it can be tempting to think that we can get away with shallower networks (we already have lots of parameters, right?), fully connected layers are just linear layers after all, so we still need to insert many non-linearities in order for the network to gain reasonable representational power. So, this will mean that you will still need a deep network, however with so many parameters that it would be untractable. In addition, a larger network will have more degrees of freedom, and will therefore model much more than what we want: it will model noise unless we feed it some data or constrain it.<\/p>\n\n<p>So yes, there might be a fully connected network that in theory could give us better performance, but we don't know how to train it yet. Finally, and this is purely based on intuition and therefore might be wrong, but it seems unlikely to me that such a fully connected network would converge to a dense solution. Since many convolutional networks achieve very high levels of accuracy (99% and up) on many tasks, I think that the optimal solution the fully connected network would converge to would be close to the convolutional network. So, we don't really need to train the fully connected one, but just a subset of its architecture.<\/p>\n"}]}
+{"QuestionId":53471105,"AnswerCount":1,"Tags":"<python><pandas><opencv><tensorflow><computer-vision>","CreationDate":"2018-11-25T19:31:52.267","AcceptedAnswerId":null,"OwnerUserId":7638835.0,"Title":"Creating a Tensorflow Model Based on Hough Lines","Body":"<p>So I have a project of creating an autonomous rc car. In order to do virtual training, I have set up a simulator to collect picture data, along with a steering angle. I then use an OpenCV program, using Hough line Detection, to detect the 3 lines on the road. I want to create a model that when images are inputted the output will be a steering angle. Using TensorFlow, I trained a CNN with these images and the result was extremely poor. <\/p>\n\n<p>What kind of model should I use ? <\/p>\n\n<p>How should I approach training such images?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/A0Vu2.jpg\" alt=\"Original Simulator Image\"> and\n<img src=\"https:\/\/i.stack.imgur.com\/02mmF.jpg\" alt=\"Hough Lines\"><\/p>\n\n<p>I can also train the data using the actual data for each line which looks like this for one line detected:<\/p>\n\n<blockquote>\n  <p>[[[0.0, 63.0, 54.0, 31.0, -0.5925925925925926]]]<\/p>\n<\/blockquote>\n\n<p>And this for 2 lines detected:<\/p>\n\n<blockquote>\n  <p>[[[0.0, 61.0, 34.0, 32.0, -0.8529411764705882], [41.0, 42.0, 43.0, 77.0, 17.5]]]<\/p>\n<\/blockquote>\n\n<p>The only problem is I don't know how to deal with inputs of variable size. Help?<\/p>\n","answers":[{"AnswerId":"56718195","CreationDate":"2019-06-22T18:45:55.757","ParentId":null,"OwnerUserId":"9331201","Title":null,"Body":"<p>now is quite late, but I wish this answer could still be useful.<\/p>\n\n<p>The CNN is a NN with bidimensional input, so is feed with constant length matrices (images for example); however most of their learning capabilities comes from the fact that CNN's are able of learning from correlated data, so it will work better with the non processed image (in fact the convolutional layer perform an adaptive feature extraction).<\/p>\n\n<p>I think that if you need to use CNN you should focus on labelling your training images in sets of turning right, left and not turning at all for classification of the turning side and letting the steering control to some fuzzy logic based controller for example.<\/p>\n\n<p>If you wish to do the control with a NN, I'll use a recurrent one for the decision as you'll need to compare the change of the picture's curve through time<\/p>\n"}]}
+{"QuestionId":53471237,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-25T19:47:12.227","AcceptedAnswerId":null,"OwnerUserId":4697430.0,"Title":"Installing packages with older version of Python","Body":"<p>I have Python 3.7, which currently doesn't support Tensorflow.<\/p>\n\n<p>So I installed Python 3.6.7, went to the command line and typed in <code>py -m pip install tensorflow<\/code>, and still got the same error:<\/p>\n\n<p><code>Could not find a version that satisfies the requirement tensorflow (from versions: ). No matching distribution found for tensorflow<\/code><\/p>\n\n<p>What can I do ensure that I'm trying to install Tensorflow using Python 3.6.7?<\/p>\n","answers":[{"AnswerId":"53471278","CreationDate":"2018-11-25T19:51:40.083","ParentId":null,"OwnerUserId":"4697430","Title":null,"Body":"<p>Browsing the documentation: <a href=\"https:\/\/docs.python.org\/3\/installing\/index.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/installing\/index.html<\/a><\/p>\n\n<p>The solution is to invoke version 3.6 by specifying it as follows:<\/p>\n\n<p><code>py -3.6 -m pip install tensorflow<\/code><\/p>\n"}]}
+{"QuestionId":53471393,"AnswerCount":1,"Tags":"<anaconda><pytorch>","CreationDate":"2018-11-25T20:03:18.150","AcceptedAnswerId":null,"OwnerUserId":10427536.0,"Title":"Anoconda does not allow Torchvision to be installed","Body":"<p>I tried the following commands \nconda create -n torch_env -c pytorch pytorch torchvision\nconda install -c soumith\/label\/pytorch torchvision\nconda install -c soumith torchvision<\/p>\n\n<p>Provided by Anaconda but none of them works. Please help me I am stack !<\/p>\n\n<p>Error message:\nSolving environment: failed<\/p>\n\n<p>PackagesNotFoundError: The following packages are not available from current channels:<\/p>\n\n<ul>\n<li>torchvision<\/li>\n<\/ul>\n\n<p>Current channels:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/conda.anaconda.org\/pytorch\/win-64\" rel=\"nofollow noreferrer\">https:\/\/conda.anaconda.org\/pytorch\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/conda.anaconda.org\/pytorch\/noarch\" rel=\"nofollow noreferrer\">https:\/\/conda.anaconda.org\/pytorch\/noarch<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/main\/win-64\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/main\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/main\/noarch\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/main\/noarch<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/free\/win-64\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/free\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/free\/noarch\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/free\/noarch<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/r\/win-64\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/r\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/r\/noarch\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/r\/noarch<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/pro\/win-64\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/pro\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/pro\/noarch\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/pro\/noarch<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/msys2\/win-64\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/msys2\/win-64<\/a><\/li>\n<li><a href=\"https:\/\/repo.anaconda.com\/pkgs\/msys2\/noarch\" rel=\"nofollow noreferrer\">https:\/\/repo.anaconda.com\/pkgs\/msys2\/noarch<\/a><\/li>\n<\/ul>\n\n<p>To search for alternate channels that may provide the conda package you're\nlooking for, navigate to<\/p>\n\n<pre><code>https:\/\/anaconda.org\n<\/code><\/pre>\n\n<p>and use the search bar at the top of the page.<\/p>\n","answers":[{"AnswerId":"53471441","CreationDate":"2018-11-25T20:09:29.493","ParentId":null,"OwnerUserId":"10427536","Title":null,"Body":"<p>I just found the answer to my question. <\/p>\n\n<p>torchvision is not available in Anaconda Cloud. Instead, I used this command and it worked fine:<\/p>\n\n<pre><code>pip install torchvision\n<\/code><\/pre>\n"}]}
+{"QuestionId":53471716,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2018-11-25T20:37:20.860","AcceptedAnswerId":53534017.0,"OwnerUserId":5812238.0,"Title":"Index pytorch 4d tensor by values in 2d tensor","Body":"<p>I have two pytorch tensors:<\/p>\n\n<ul>\n<li><code>X<\/code> with shape <code>(A, B, C, D)<\/code><\/li>\n<li><code>I<\/code> with shape <code>(A, B)<\/code><\/li>\n<\/ul>\n\n<p>Values in <code>I<\/code> are integers in range <code>[0, C)<\/code>.<\/p>\n\n<hr>\n\n<p>What is the most efficient way to get tensor <code>Y<\/code> with shape <code>(A, B, D)<\/code>, such that:<\/p>\n\n<pre><code>Y[i][j][k] = X[i][j][ I[i][j] ][k]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53534017","CreationDate":"2018-11-29T07:40:03.103","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You probably want to use <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.gather\" rel=\"nofollow noreferrer\"><code>torch.gather<\/code><\/a> for the indexing and <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.expand\" rel=\"nofollow noreferrer\"><code>expand<\/code><\/a> to adjust <code>I<\/code> to the required size:<\/p>\n\n<pre><code>eI = I[..., None, None].expand(-1, -1, 1, X.size(3))  # make eI the same for the last dimension\nY = torch.gather(X, dim=2, index=eI).squeeze()\n<\/code><\/pre>\n\n<p>testing the code:<\/p>\n\n<pre><code>A = 3 \nB = 4 \nC = 5 \nD = 7\n\nX = torch.rand(A, B, C, D)\nI = torch.randint(0, C, (A, B), dtype=torch.long)\n\neI = I[..., None, None].expand(-1, -1, 1, X.size(3))\nY = torch.gather(X, dim=2, index=eI).squeeze()\n\n# manually gather\nrefY = torch.empty(A, B, D)\nfor i in range(A):\n    for j in range(B):\n        refY[i, j, :] = X[i, j, I[i,j], :]\n\n(refY == Y).all()\n# Out[]: tensor(1, dtype=torch.uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53471761,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><conv-neural-network>","CreationDate":"2018-11-25T20:43:44.290","AcceptedAnswerId":53474922.0,"OwnerUserId":7857121.0,"Title":"Training a model with tensorflow,but loss just won't go down?","Body":"<p>I am training a model that classifies 62 alphanumeric characters.But The loss value first drastically drops during first few batches and then hit the ground and never go down.I don't know where it is wrong or how to debug the model.<\/p>\n\n<p>Here is a snapshot of training log:  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wMGUM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wMGUM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here is examples of traning data:  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zimwS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zimwS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I use 4 conv layers followed by 1 fc layer,adam optimizer to minimize log loss.I double checked that the image label is right.So i don't know where else it is wrong.<\/p>\n\n<p>Here is the code:  <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport os\nfrom PIL import Image\nimport shutil\nimport time\n\ninput = temp= tf.placeholder(dtype='float32', shape=(None,32,32,1), name='input')#(None,62)\nlabel = tf.placeholder(dtype='float32',shape=(None,62))#(None,62)\n\ntemp = tf.layers.conv2d(inputs=temp,filters=32,kernel_size=(3,3),padding=\"SAME\",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,32,32,32)\n#temp = tf.layers.dropout(inputs=temp,rate=0.5)\ntemp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,16,16,32)\n\ntemp = tf.layers.conv2d(inputs=temp,filters=64,kernel_size=(3,3),padding=\"SAME\",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,16,16,64)\n#temp = tf.layers.dropout(inputs=temp,rate=0.2)\ntemp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,8,8,64)\n\ntemp = tf.layers.conv2d(inputs=temp,filters=128,kernel_size=(3,3),padding=\"SAME\",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,8,8,128)\ntemp = tf.layers.dropout(inputs=temp,rate=0.2)\ntemp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,4,4,128)\n\ntemp = tf.layers.conv2d(inputs=temp,filters=256,kernel_size=(3,3),padding=\"SAME\",activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.he_normal())#(None,4,4,256)\ntemp = tf.layers.dropout(inputs=temp,rate=0.2)\ntemp = tf.layers.max_pooling2d(temp,pool_size=[2, 2], strides=2)#(None,2,2,256)\n\ntemp = tf.layers.conv2d(inputs=temp,filters=62,kernel_size=(2,2),padding=\"VALID\",kernel_initializer=tf.keras.initializers.he_normal())#fc (None,1,1,62)\noutput = temp = tf.layers.flatten(temp)#(None,62)\noutput = tf.nn.softmax(output)\n\n#loss\noutput_clip = tf.clip_by_value(output,1e-7,1-1e-7)\nloss = tf.reduce_mean(tf.reduce_sum(-label*tf.log(output_clip)-(1-label)*tf.log(1-output_clip),axis=-1))#scaler\noptimizer = tf.train.AdamOptimizer().minimize(loss)\n\n#accuracy\nindexoutput = tf.argmax(output,axis=-1)#(None,)\nlabelindex = tf.argmax(label,axis=-1)#(None,)\nequals = tf.equal(indexoutput,labelindex)#(None,)\nequals = tf.reduce_sum(tf.cast(equals,dtype='int8'),axis=-1)#scaler\nacc = tf.cast(equals,dtype='float32')\/tf.cast(tf.shape(output)[0],dtype='float32')#scaler\n\ndef train(epochs):\n    saver = tf.train.Saver()\n    lossrec=[]\n    accrec = []\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        #saver.restore(sess, os.path.join(os.getcwd(),'model_logloss','captchabreak.ckpt'))\n        valimg,vallabel = next(validategenerator(os.path.join(os.getcwd(),'tests')))\n        for i in range(epochs):\n            for j, (trainimg, trainlabel) in enumerate(traingenerator(os.path.join(os.getcwd(),'pics'),32)):\n                _, trainacc, trainloss = sess.run([optimizer, acc, loss],feed_dict={input: trainimg, label: trainlabel});\n                valacc, valloss = sess.run([acc, loss], feed_dict={input: np.array(valimg), label: np.array(vallabel)})\n                print(\"epoch:{} batch:{} trainloss:{:.4f} validateloss:{:.4f} trainacc:{:.2f} validateacc:{:.2f}\"\n                      .format(i, j, trainloss, valloss, trainacc, valacc))\n            #keep some logs\n            lossrec.append(valloss)\n            accrec.append(valacc)\n            print(lossrec);\n            print(accrec)\n            if len(lossrec) &gt;= 3 and valloss &gt;= lossrec[-2] and valloss &gt;= lossrec[-3]: break;\n            shutil.rmtree(\"model_logloss\")\n            saver.save(sess, \"model_logloss\/captchabreak.ckpt\")\n\ndef traingenerator(path,batch_size):\n    fs = os.listdir(path);\n    fs=np.random.permutation(fs)\n    if batch_size == 0:\n        batch_size = len(fs)\n    offset=0\n    while offset&lt;len(fs):\n        yield fetch(fs[offset:offset+batch_size],path)\n        offset+=batch_size\n\ndef validategenerator(path):\n    fs = os.listdir(path);\n    yield fetch(fs,path)\n\ndef fetch(fs,path):\n    imgs = []\n    labels = []\n    for i, fname in enumerate(fs):\n        fp = os.path.join(path, fname)\n        imp = Image.open(fp).resize((32, 32));\n        imp = imp.convert('L')\n        imp = imp.point(lambda p: p &gt; 210 and 255)\n        im = np.array(imp)\n        im = np.expand_dims(im, axis=-1)\n        imp.close()\n\n        c = fname[0]\n        lb = np.zeros((62))\n        if ord(c) &gt;= 48 and ord(c) &lt;= 57:\n            lb[ord(c) - 48] = 1\n        if ord(c) &gt;= 65 and ord(c) &lt;= 90:\n            lb[ord(c) - 65 + 10] = 1\n        if ord(c) &gt;= 97 and ord(c) &lt;= 122:\n            lb[ord(c) - 97 + 36] = 1\n        imgs.append(im)\n        labels.append(lb)\n    return np.array(imgs), np.array(labels)\n\nif __name__ == \"__main__\":\n    train(30)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53474922","CreationDate":"2018-11-26T04:48:31.677","ParentId":null,"OwnerUserId":"7857121","Title":null,"Body":"<p>It's solved . I forgot to  divide the pixel value by 255.<\/p>\n"}]}
+{"QuestionId":53471855,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-25T20:52:53.717","AcceptedAnswerId":53472040.0,"OwnerUserId":5071577.0,"Title":"Why does Keras halt at the first epoch when I attempt to train it using fit_generator?","Body":"<p>I'm using Keras to fine tune an existing VGG16 model and am using a fit_generator to train the last 4 layers. Here's the relevant code that I'm working with:<\/p>\n\n<pre><code># Create the model\nmodel = models.Sequential()\n\n# Add the vgg convolutional base model\nmodel.add(vgg_conv)\n\n# Add new layers\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(1024, activation='relu'))\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Dense(5, activation='softmax'))\n\n# Show a summary of the model. Check the number of trainable params\nmodel.summary()\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(\n    rescale=1.\/255,\n    rotation_range=20,\n    width_shift_range=0.2,\n    height_shift_range=0.2,\n    horizontal_flip=True,\n    fill_mode='nearest')\n\nvalidation_datagen = ImageDataGenerator(rescale=1.\/255)\n\n#Change the batchsize according to the system RAM\ntrain_batchsize = 100\nval_batchsize = 10\n\ntrain_dir='training_data\/train'\nvalidation_dir='training_data\/validation'\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_dir,\n    target_size=(image_size1, image_size2),\n    batch_size=train_batchsize,\n    class_mode='categorical')\n\nvalidation_generator = validation_datagen.flow_from_directory(\n    validation_dir,\n    target_size=(image_size1, image_size2),\n    batch_size=val_batchsize,\n    class_mode='categorical',\n    shuffle=False)\n\n# Compile the model\nmodel.compile(loss='categorical_crossentropy',\n              optimizer=optimizers.RMSprop(lr=1e-4),\n              metrics=['acc'])\n\n# Train the model\nhistory = model.fit_generator(\n    train_generator,\n    steps_per_epoch=train_generator.samples\/train_generator.batch_size,\n    epochs=30,\n    validation_data=validation_generator,\n    validation_steps=validation_generator.samples\/validation_generator.batch_size,\n    verbose=1)\n<\/code><\/pre>\n\n<p>The issue is that when I run my script to train the model, it works fine until the actual training begins. Here, it gets stuck at epoch 1\/30.<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #\n=================================================================\nvgg16 (Model)                (None, 15, 20, 512)       14714688\n_________________________________________________________________\nflatten_1 (Flatten)          (None, 153600)            0\n_________________________________________________________________\ndense_1 (Dense)              (None, 1024)              157287424\n_________________________________________________________________\ndropout_1 (Dropout)          (None, 1024)              0\n_________________________________________________________________\ndense_2 (Dense)              (None, 5)                 5125\n=================================================================\nTotal params: 172,007,237\nTrainable params: 164,371,973\nNon-trainable params: 7,635,264\n_________________________________________________________________\nFound 1989 images belonging to 5 classes.\nFound 819 images belonging to 5 classes.\nEpoch 1\/30\n<\/code><\/pre>\n\n<p>This is no good unfortunately. I looked around online and I believe that the problem is in using fit_generator. There's something about the code for fit_generator in Keras being buggy. However, most of the other people experiencing issues with the epochs end up getting stuck on later epochs (ex. somebody wants to run it for 20 epochs and it halts on epoch 19\/20). <\/p>\n\n<p>How would I go about fixing this issue? This is my first time doing deep learning so I'm incredibly confused and would appreciate any help. Do I just need to move to using model.fit()?<\/p>\n","answers":[{"AnswerId":"53472040","CreationDate":"2018-11-25T21:13:03.523","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You have to pass a valid <strong>integer<\/strong> number to <code>fit_generator()<\/code> as <code>steps_per_epoch<\/code> and <code>validation_steps<\/code> parameters.\nSo you can use as follows:<\/p>\n\n<pre><code>history = model.fit_generator(\n    train_generator,\n    steps_per_epoch=train_generator.samples\/\/train_generator.batch_size,\n    epochs=30,\n    validation_data=validation_generator, validation_steps=validation_generator.samples\/\/validation_generator.batch_size,\n    verbose=1)\n<\/code><\/pre>\n\n<p>The second factor I can see that your model has <code>165M<\/code> trainable <strong>parameter<\/strong> which has huge memory consumption particularly coupled with a high <strong>batchsize<\/strong>. \nYou should use images with lower resolution, note that in many case we can get better results with them.<\/p>\n"}]}
+{"QuestionId":53472227,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><google-cloud-platform><apache-beam><tensorflow-transform>","CreationDate":"2018-11-25T21:33:27.243","AcceptedAnswerId":null,"OwnerUserId":10703486.0,"Title":"Problem with Tensorflow Transform(TFX) compute_and_apply_vocabulary\/sparse_tensor_to_dense_with_shape","Body":"<p>I have some problems running a Apache beam job on Dataflow. The code runs fine on a small dataset but when runing a bigger batch job on Dataflow i get the following message: <\/p>\n\n<p>RuntimeError: InvalidArgumentError: indices[10925] = [889,43] is out of bounds: need 0 &lt;= index &lt; [1238,43]\n     [[{{node transform\/transform\/SparseToDense}} = SparseToDense[T=DT_INT64, Tindices=DT_INT64, validate_indices=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](transform\/transform\/StringSplit, transform\/transform\/SparseToDense\/output_shape, transform\/transform\/compute_and_apply_vocabulary\/apply_vocab\/hash_table_Lookup, transform\/transform\/compute_and_apply_vocabulary\/apply_vocab\/string_to_index\/hash_table\/Const)]]<\/p>\n\n<p>Taken from the log in Dataflow. <\/p>\n\n<p>I have figured out that it is connected to my AnalyzeAndTransformDataset function:<\/p>\n\n<pre><code>def preprocessing_fn(inputs):\n      words = tf.string_split(inputs['tweet'],DELIMITERS)\n      int_representation =tft.compute_and_apply_vocabulary(words,default_value=0,top_k=10000)\n      # The shape out here is the problem, i think\n\n      int_representation = \n      tft.sparse_tensor_to_dense_with_shape(int_representation, \n        [None,43])\n      outputs = inputs\n      outputs[\"int_representation\"] = int_representation \n      return outputs\n<\/code><\/pre>\n\n<p>The goal is to get a dense output vector of length 43 out for each example. This will be used for Sentimental Analysis on Twitter data as a home project :) <\/p>\n\n<p>Thanks for any help!<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>I didn't manage to solve to the problem but instead converted the vector to a dense vector inside my Tensorflow estimator model. Not the solution I had hoped for but it works! Have also reported it to Tensorflow transforms on GitHub. <\/p>\n","answers":[]}
+{"QuestionId":53473110,"AnswerCount":1,"Tags":"<python><tensorflow><keras><generator>","CreationDate":"2018-11-25T23:39:59.533","AcceptedAnswerId":53499391.0,"OwnerUserId":6341510.0,"Title":"Same ImageDataGenerator but different class_indices - How to remap the classes within a generator?","Body":"<h2>Background<\/h2>\n\n<p>I am training a model that takes as inputs two images. As the data is too large to fit into my machines RAM, I am using <code>flow_from_dataframe<\/code> to create the generators for training and validation - two training generators which each provide one of the respective images (front view &amp; back view as indicated by the <code>x_col<\/code> parameter) and two generators for validation respectively.<\/p>\n\n<p>Like this:<\/p>\n\n<pre><code>datagen=ImageDataGenerator(rescale=1. \/ 255)\n\nX1_train_generator =datagen.flow_from_dataframe(dataframe=train, directory=data_dir, x_col=\"front\", y_col=target, has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,seed = 1)\nX2_train_generator=datagen.flow_from_dataframe(dataframe=train, directory=data_dir, x_col=\"back\", y_col=target, has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,seed = 1)\n\nX1_validation_generator =datagen.flow_from_dataframe(dataframe=test, directory=data_dir, x_col=\"front\", y_col=target, has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,seed = 1)\nX2_validation_generator=datagen.flow_from_dataframe(dataframe=test, directory=data_dir, x_col=\"back\", y_col=target, has_ext=True, class_mode=\"categorical\", target_size=(224,224), batch_size=batch_size,seed = 1)\n<\/code><\/pre>\n\n<p>In order to combine the train and validation generator for training I am using: <\/p>\n\n<pre><code>def format_gen_outputs(gen1,gen2):\n    x1 = gen1[0]\n    x2 = gen2[0]\n    y1 = gen1[1]\n    return [x1, x2], y1\n\ntrain_combo_gen= map(format_gen_outputs, X1_train_generator , X2_train_generator )\nvalidation_combo_gen= map(format_gen_outputs, X1_validation_generator , X2_validation_generator )\n<\/code><\/pre>\n\n<p>Now I use <code>fit_generator<\/code> to train my model passing <code>train_combo_gen<\/code> for training purpose and <code>validation_combo_gen<\/code> to the <code>validation_data<\/code> parameter for validation purpose<\/p>\n\n<h2>Problem<\/h2>\n\n<p>However, I realize that both my <code>X1_train_generator<\/code> and <code>X2_train_generator<\/code> show a different <code>.class_indices<\/code> mapping than my other two validation generators <code>X1_validation_generator<\/code> and <code>X2_validation_generator<\/code>.<\/p>\n\n<p>Like this (notice how cat &amp; dog are assigned to different classes):<\/p>\n\n<pre><code>X1_train_generator.class_indices\n&gt;&gt; {'cat': 0, 'dog': 1, 'car': 2, 'bike': 3}\n\nX1_validation_generato.class_indices\n&gt;&gt; {'dog': 0, 'cat': 1, 'car': 2, 'bike': 3}\n<\/code><\/pre>\n\n<h2>Question<\/h2>\n\n<p>Hence, I don't trust my <code>val_loss<\/code> and <code>val_acc<\/code> during training. Is there any way to fix this i.e. remap the classes within the generators?<\/p>\n","answers":[{"AnswerId":"53499391","CreationDate":"2018-11-27T12:09:55.227","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>When you don't explicitly set the classes using <code>classes<\/code> argument, the <code>flow_from_dataframe<\/code> internally uses pandas Series <a href=\"https:\/\/pandas.pydata.org\/pandas-docs\/stable\/generated\/pandas.Series.unique.html\" rel=\"nofollow noreferrer\"><code>unique<\/code><\/a> method on the <code>y_col<\/code> column <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/7b7af2690026b51b5a7a50b4b359ff12960ac8b7\/keras_preprocessing\/image.py#L2107\" rel=\"nofollow noreferrer\">to find the classes<\/a>:<\/p>\n\n<pre><code>if not classes:\n    classes = []\n    if class_mode not in [\"other\", \"input\", None]:\n        classes = list(self.df[y_col].unique())\n<\/code><\/pre>\n\n<p>The <code>unique<\/code> method would return the unique values in order of appearance in the column. Since the order of appearance of labels in your train and test dataframe are different from each other, you would get different indices for classes.<\/p>\n\n<p>One workaround is to explicitly set the <code>classes<\/code> argument for all the <code>flow_from_dataframe<\/code> calls to guarantee the same class indices mapping in train and validation generators:<\/p>\n\n<pre><code>X1_train_generator =datagen.flow_from_dataframe(dataframe=train, \n                                                classes=['cat', 'dog', 'car', 'bike'], ...)\n\n# do the same for `X2_train_generator`, `X1_validation_generator` and `X2_validation_generator`\n<\/code><\/pre>\n"}]}
+{"QuestionId":53473184,"AnswerCount":1,"Tags":"<python><tensorflow><random-forest><cleverhans>","CreationDate":"2018-11-25T23:51:53.117","AcceptedAnswerId":53475584.0,"OwnerUserId":8568673.0,"Title":"Random Forest Black Box with CleverHans","Body":"<p>I am new to this stuff and trying to attack Random Forest with Black Box FGSM (from clever hans)<\/p>\n\n<p>But I'm not sure how to implement it. They've a <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans_tutorials\/mnist_blackbox.py\" rel=\"nofollow noreferrer\">blackbox example for Mnist data<\/a> but I dont understand where should I put my random forest and where should I attack. Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"53475584","CreationDate":"2018-11-26T06:10:47.127","ParentId":null,"OwnerUserId":"10560104","Title":null,"Body":"<p>In the current tutorial, the black-box model is a neural network implemented with TensorFlow and its predictions (the labels) are used to train a substitute model (a copy of the black-box model). The substitute model is then used to craft adversarial examples that transfer to the black-box model.  <\/p>\n\n<p>In your case, you would have to replace bbox_val in <\/p>\n\n<pre><code>bbox_val = batch_eval(sess, [x], [bbox_preds], [x_sub_prev],\n                        args=eval_params)[0]\n<\/code><\/pre>\n\n<p>by the predictions of your random forest on the numpy array of substitute training data <code>x_sub_prev<\/code>. <\/p>\n\n<p>You can find more information about the attack implemented in this tutorial in the following paper: <a href=\"https:\/\/arxiv.org\/abs\/1602.02697\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1602.02697<\/a><\/p>\n"}]}
+{"QuestionId":53473264,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-26T00:06:14.867","AcceptedAnswerId":null,"OwnerUserId":5190484.0,"Title":"Tensorflow gradient returning null","Body":"<p>I am trying to implement the tutorial given on <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training\" rel=\"nofollow noreferrer\">Tensorflow tutorial on custom training<\/a>. Due to some reason dW and DB is None. I am not getting why t.gradient() is returning None.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\nclass Model(object):\n    def __init__(self):\n        self.W = tf.Variable(5.0)\n        self.b = tf.Variable(0.0)\n    def __call__(self,x):\n        return self.W*x+self.b\n    def loss_function(self, y_true, y_predicted):\n        return tf.reduce_mean(tf.square(y_predicted-y_true))\n    def train(self, inputs, outputs, learning_rate):\n        with tf.GradientTape() as t:\n            current_loss = self.loss_function(inputs,outputs)\n        dW,db = t.gradient(current_loss,[self.W, self.b])\n        ## dW and db returns None\n        self.W.assign_sub(learning_rate*dW)\n        self.b.assign_sub(learning_rate*db)\n<\/code><\/pre>\n\n<p>But the following code works fine when train is not a method of model. Any reason why?<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\nclass Model(object):\n    def __init__(self):\n        self.W = tf.Variable(5.0)\n        self.b = tf.Variable(0.0)\n    def __call__(self,x):\n        return self.W*x+self.b\n    def loss_function(self, y_true, y_predicted):\n        return tf.reduce_mean(tf.square(y_predicted-y_true))\n\ndef train(model, inputs, outputs, learning_rate):\n    with tf.GradientTape() as t:\n        current_loss = model.loss_function(model(inputs),outputs)\n    dW,db = t.gradient(current_loss,[model.W, model.b])\n    ## dW and db returns None\n    model.W.assign_sub(learning_rate*dW)\n    model.b.assign_sub(learning_rate*db)\n<\/code><\/pre>\n","answers":[{"AnswerId":"55330488","CreationDate":"2019-03-25T02:17:10.763","ParentId":null,"OwnerUserId":"10903402","Title":null,"Body":"<p>For <code>gradient<\/code> to work, the whole graph needs to be crafted within the scope of <code>GradientTape<\/code>.<\/p>\n\n<p>For example, in the code provided in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training\" rel=\"nofollow noreferrer\">Tensorflow tutorial on custom training<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.GradientTape() as t:\n    current_loss = model.loss_function(model(inputs),outputs)\n<\/code><\/pre>\n\n<p>The graph where <code>current_loss<\/code> is connected to the model variables (<code>model.W<\/code> and <code>model.B<\/code>) are constructed within the scope of <code>t<\/code>.<\/p>\n\n<p>If you change the codes provided by the tutorial as follows:<\/p>\n\n<pre><code>logits = model(inputs)\nwith tf.GradientTape() as t:\n    current_loss = model.loss_function(logits, outputs)\n<\/code><\/pre>\n\n<p>You will get <code>None<\/code> for <code>dW<\/code> and <code>db<\/code>.<\/p>\n"}]}
+{"QuestionId":53473336,"AnswerCount":3,"Tags":"<python><tensorflow><keras><anaconda><virtualenv>","CreationDate":"2018-11-26T00:18:46.827","AcceptedAnswerId":null,"OwnerUserId":5302331.0,"Title":"Installing Keras broke Conda virtual environment (Windows 10)","Body":"<p>I was trying to install Keras on my python 3.7 environment then I noticed that neither tensorflow nor keras supports python 3.7, they support 3.6 <br>\nThen I created a new virtual environment with python 3.6 using conda\/anaconda.<br>\nI was able to activate this virtual environment by entering\n<code>activate deeplearning<\/code> in the cmd and then type <code>python<\/code> to enter python shell.<br>I installed Keras and its dependencies using <code>conda install -c conda-forge keras<\/code> , it seemed okay I entered the python shell imported keras and other libraries tinkered around with the libraries, they were working as intended. <br>\nI closed down the window to try again if i can open up the virtual environment again manually myself.<br> Typed <code>activate deeplearning<\/code>, and then I ran into this problem(\n<a href=\"https:\/\/i.stack.imgur.com\/dc79q.jpg\" rel=\"nofollow noreferrer\">Screenshot<\/a>\n)<br>I only typed the <code>activate deeplearning<\/code> part on the cmd screenshot.\n<br>\nThe problem here is that I can not enter the virtual environment, I get this commands? on the cmd and then it shuts itself down. I can not enter the virtual environment therefore preventing me from installing new libraries. But I can enter the python shell of this environment from where it is installed.<br> However there is not a .bat file in the folder so I can not activate the virtual environment from there as well.\n<br>\nI do not know what to do, I would really appreciate it if someone could help, thanks in advance.<\/p>\n","answers":[{"AnswerId":"54820325","CreationDate":"2019-02-22T04:58:53.963","ParentId":null,"OwnerUserId":"414125","Title":null,"Body":"<p>On my system, the error was caused by another activation script, probably installed at the same time. To confirm that this is the case for you, look into the batch file generated dynamically at startup. To find out its name, change <code>_conda_activate.bat<\/code> as so:<\/p>\n\n<pre><code>...\n\n:FIXUP43\n@FOR \/F \"delims=\" %%i IN ('@CALL \"%CONDA_EXE%\" shell.cmd.exe %*') DO @SET \"_TEMP_SCRIPT_PATH=%%i\"\n@IF \"%_TEMP_SCRIPT_PATH%\"==\"\" @EXIT \/B 1\n@IF NOT \"%CONDA_PROMPT_MODIFIER%\" == \"\" @CALL SET \"PROMPT=%%PROMPT:%CONDA_PROMPT_MODIFIER%=%_empty_not_set_%%%\"\n\n:: Add these lines:\nset _TEMP_SCRIPT_PATH\npause\n\n@CALL \"%_TEMP_SCRIPT_PATH%\"\n...\n\n<\/code><\/pre>\n\n<p>In the temp script comment out the invocation of other scripts\nuntil you find the culprit. For me it was something to do with vs2015.<\/p>\n\n<pre><code>@SET \"PYTHONIOENCODING=1252\"\n@CALL \"C:\\Users\\x030551\\AppData\\Local\\Continuum\\miniconda3\\etc\\conda\\activate.d\\keras_activate.bat\"\n:: @CALL \"C:\\Users\\x030551\\AppData\\Local\\Continuum\\miniconda3\\etc\\conda\\activate.d\\vs2015_compiler_vars.bat\"\n<\/code><\/pre>\n\n<p>Once I had this information, Google pointed me to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/23310\" rel=\"nofollow noreferrer\">this issue<\/a>. The solution to that depends too much on your environment to post here.<\/p>\n"},{"AnswerId":"54085047","CreationDate":"2019-01-08T04:00:59.340","ParentId":null,"OwnerUserId":"3701521","Title":null,"Body":"<p>I fixed this on Python 2.7 by removing via Anaconda Navigator the item that goes by the name vs2008_win-32, that fixed my environment.<\/p>\n"},{"AnswerId":"53474925","CreationDate":"2018-11-26T04:48:44.630","ParentId":null,"OwnerUserId":"10355733","Title":null,"Body":"<p>It seems that the scripts for activating keras are exiting prematurely. One of the reasons can be that the environment is not set in python 3.6.<\/p>\n\n<p>Try creating a virtual environment with command:<\/p>\n\n<blockquote>\n  <p>conda create --name deeplearning python=3.6<\/p>\n<\/blockquote>\n\n<p>And then you can follow instructions from:\n<a href=\"http:\/\/inmachineswetrust.com\/posts\/deep-learning-setup\/\" rel=\"nofollow noreferrer\">http:\/\/inmachineswetrust.com\/posts\/deep-learning-setup\/<\/a><\/p>\n\n<p>Note: While installing keras use command:<\/p>\n\n<blockquote>\n  <p>conda install keras<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53473565,"AnswerCount":0,"Tags":"<python><keras><neural-network><classification><multilabel-classification>","CreationDate":"2018-11-26T01:01:30.277","AcceptedAnswerId":null,"OwnerUserId":2434875.0,"Title":"Training multi-label classifier with unbalanced samples in Keras","Body":"<p>I'm trying to train a keras model that takes in samples, let's say <code>x_i<\/code> for sample <code>i<\/code>, and predicts multiple independent labels, <code>{y_hat}_ij<\/code>, such that <code>{y_hat}_ij = 1<\/code> if the model predicts sample <code>x_i<\/code> to have label <code>j<\/code>, and <code>0<\/code> otherwise.<\/p>\n\n<p>The problem I'm running into repeatedly is that the number of samples per label is not balanced across labels. Particularly, here's a count of how many samples each label has (there's double counting, of course, since we're dealing with multi-label classification). My model has 23 labels.<\/p>\n\n<pre><code>total: 6790\nCounter({22: 4702, 0: 1749, 12: 130, 8: 43, 16: 39,\n         15: 30, 17: 24, 20: 17, 4: 13, 5: 13, 19: 9,\n         6: 7, 7: 6, 2: 4, 10: 4})\n<\/code><\/pre>\n\n<p>I'm not sure how to weight them by size, however. I saw <a href=\"https:\/\/datascience.stackexchange.com\/questions\/13490\/how-to-set-class-weights-for-imbalanced-classes-in-keras\">this<\/a> question, but it uses one-hot encodings of labels, i.e., assumes there's only one true label per sample. It is essentially a sparse categorical classification problem. I also ran into <a href=\"https:\/\/stackoverflow.com\/questions\/48485870\/multi-label-classification-with-class-weights-in-keras\">this<\/a> question, but I believe that question is concerned with per-sample weighting of classes, somehow. The accepted answer does not state how to compute weights across training samples.<\/p>\n\n<p>What I tried so far:\nI computed arbitrary weights of each label relative to its occurrences vs the total samples.\n<code>class_weight = dict((c,round((1\/v)*total,1)) for c,v in class_occ.items())<\/code>, where <code>class_occ<\/code> is the counter shown above. Here's what I get: <\/p>\n\n<pre><code>class_weights: {0: 3.9, 2: 1697.5, 4: 522.3, 5: 522.3, 6: 970.0, 7: 1131.7, 8: 157.9, 10: 1697.5, 12: 52.2, 15: 226.3, 16: 174.1, 17: 282.9, 19: 754.4, 20: 399.4, 22: 1.4}\n<\/code><\/pre>\n\n<p>Then, while training the model, I supply it as a kwarg:<\/p>\n\n<p><code>model.fit(x=...., class_weight=class_weight)<\/code><\/p>\n\n<p>However, the model doesn't really learn thereafter, and accuracy remains at a low of 0.1 or so. To give some context, here's how I compiled the model:<\/p>\n\n<pre><code>out = Dense(numclasses+1, activation='sigmoid', name='out')(dense_top)\n\nmodel = Model(inputs=[input], outputs=[out])\nmodel.compile(optimizer=\"adam\", loss=\"binary_crossentropy\",\n              metrics=[\"categorical_accuracy\"])\n<\/code><\/pre>\n\n<p>And here's how the training goes (blue is the loss; all layers trainable): \n<a href=\"https:\/\/i.stack.imgur.com\/lZpM7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lZpM7.png\" alt=\"model-fit-history\"><\/a><\/p>\n\n<p>And from another training instance (<em>yellow<\/em> is the loss; initial layers are frozen):\n<a href=\"https:\/\/i.stack.imgur.com\/klraX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/klraX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I'm not sure what's going wrong- am I even using the class weights correctly? Without the class weights, the model would just predict the dominant class(es) all the time and the accuracy would a baseline of the proportion of samples that had the dominant class(es). But looks like it does poorer than random now.<\/p>\n\n<p>Just in case it helps, here's my <code>model.summary()<\/code> output:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, None, 128)         0         \n_________________________________________________________________\nlstm_1 (LSTM)                (None, 192)               246528    \n_________________________________________________________________\ndense_1 (Dense)              (None, 160)               30880     \n_________________________________________________________________\ndense_2 (Dense)              (None, 128)               20608     \n_________________________________________________________________\ndense_3 (Dense)              (None, 128)               16512     \n_________________________________________________________________\ndense_4 (Dense)              (None, 64)                8256      \n_________________________________________________________________\ndense_top_1 (Dense)          (None, 64)                4160      \n_________________________________________________________________\ndropout_42 (Dropout)         (None, 64)                0         \n_________________________________________________________________\ndense_top_2 (Dense)          (None, 64)                4160      \n_________________________________________________________________\ndropout_43 (Dropout)         (None, 64)                0         \n_________________________________________________________________\ndense_top_3 (Dense)          (None, 64)                4160      \n_________________________________________________________________\ndropout_44 (Dropout)         (None, 64)                0         \n_________________________________________________________________\ndense_top_4 (Dense)          (None, 64)                4160      \n_________________________________________________________________\ndropout_45 (Dropout)         (None, 64)                0         \n_________________________________________________________________\ndense_top_5 (Dense)          (None, 32)                2080      \n_________________________________________________________________\nout (Dense)                  (None, 23)                759       \n=================================================================\nTotal params: 342,263\nTrainable params: 19,479\nNon-trainable params: 322,784\n<\/code><\/pre>\n\n<p>The reason for the non-trainable params is that I'm using transfer learning: the first bunch of layers are from another model that did binary classification on the same data, for sets of related labels ('does a sample <code>x<\/code> have any label from label-set A vs. label-set B?').\nJust in case <em>that's<\/em> the problem, I also set trainable to <code>True<\/code> in a separate instance, but it did not help with training and getting the model to work.<\/p>\n","answers":[]}
+{"QuestionId":53473606,"AnswerCount":1,"Tags":"<callback><keras><google-colaboratory>","CreationDate":"2018-11-26T01:10:19.030","AcceptedAnswerId":null,"OwnerUserId":1276374.0,"Title":"Keras history order problem when saving to GDrive","Body":"<p>My aim is to save Keras logs (accuracies, losses, etc) to Google Drive after every epoch<\/p>\n\n<p>I am using the following code:<\/p>\n\n<pre><code>from google.colab import drive\ndrive.mount('\/content\/drive')\n\nclass HistoryCallback(callbacks.Callback):\n    def on_epoch_end(self, epoch, logs={}):\n        with open(\"drive\/My Drive\/\"+csv_path, \"a\") as myfile:\n            myfile.write(str(epoch)+\",\"+str(logs)+\"\\n\")\n<\/code><\/pre>\n\n<p>classifier.fit_generator(..., callbacks=[HistoryCallback()])<\/p>\n\n<p>On my drive, I <strong><em>sometimes<\/em><\/strong> get strange output, like this:<\/p>\n\n<pre><code>25  {'acc': 0.963835932997043   loss': 0.10425430848152908  val_acc': 0.7071953016230713    val_loss': 1.1386645126622854   lr': 2.3961632e-06}\n\n2366427 loss': 1.1117452404459112   val_acc': 0.5577092514076597    val_loss': 1.0548135743792362   lr': 4.980681e-06}  \n2   {'acc': 0.6330444829612712  loss': 0.9205646275682026   val_acc': 0.5994126287500939    val_loss': 0.9518886575614829   lr': 4.956604e-06}\n3   {'acc': 0.6983824379057777  loss': 0.7753418573921365   val_acc': 0.6314243757777277    val_loss': 0.8930798317542336   lr': 4.923029e-06}\n4   {'acc': 0.7437319468601393  loss': 0.6659318362681732   val_acc': 0.6464023491359492    val_loss': 0.866631023106596    lr': 4.8800885e-06}\n5   {'acc': 0.7798526863016054  loss': 0.5813610065455109   val_acc': 0.6637298091742786    val_loss': 0.8554121221564764   lr': 4.8279526e-06}\n6   {'acc': 0.8090410167584868  loss': 0.5098161270401851   val_acc': 0.6657856092579039    val_loss': 0.8496283279291509   lr': 4.7668264e-06}\n7   {'acc': 0.8317157712132858  loss': 0.45465362796302755  val_acc': 0.6734214392352559    val_loss': 0.8745797048056179   lr': 4.6969512e-06}\n8   {'acc': 0.8491478913819287  loss': 0.4042509938624124   val_acc': 0.688986784105959 val_loss': 0.8465897937878288   lr': 4.6186033e-06}\n....\n<\/code><\/pre>\n\n<p>As you can see, the ordering is messed up, epoch 0 is missing, and 2366427 is a portion of the accuracy with missing epoch number and a few preceding numbers  <\/p>\n\n<p>Has anyone encountered this before and knows how to deal with it?<\/p>\n\n<p><strong>EDIT:<\/strong>\nI have noticed that after some epochs, the file on the GDrive is also 0 bytes and then on the next epoch it is filled back in<\/p>\n","answers":[{"AnswerId":"53534960","CreationDate":"2018-11-29T08:46:31.387","ParentId":null,"OwnerUserId":"8755609","Title":null,"Body":"<p>A guess: your training is proceeding in multiple threads\/subprocesses, so your on_epoch_end is fighting itself - concurrently-executing copies of it are all opening the same file in 'a'ppend mode, and (partially) overwriting each other's outputs.\nDoes the problem go away if you write to a per-epoch file, or serialize output another way?<\/p>\n"}]}
+{"QuestionId":53474056,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-26T02:32:05.243","AcceptedAnswerId":53475089.0,"OwnerUserId":9633432.0,"Title":"How to convert 3D tensor to 2D tensor in pytorch?","Body":"<p>I am new to pytorch. I have 3D tensor (32,10,64) and I want a 2D tensor (32, 64).\nI tried <code>view()<\/code> and used after passing to linear layer <code>squeeze()<\/code> which converted it to (32,10). <\/p>\n","answers":[{"AnswerId":"53475089","CreationDate":"2018-11-26T05:14:25.490","ParentId":null,"OwnerUserId":"9049053","Title":null,"Body":"<p>Try this<\/p>\n\n<pre><code>t = torch.rand(32, 10, 64).permute(0, 2, 1)[:, :, -1]\n<\/code><\/pre>\n\n<p>or, as pointed out by <a href=\"https:\/\/stackoverflow.com\/questions\/53474056\/how-to-convert-3d-tensor-to-2d-tensor-in-pytorch\/53475089?noredirect=1#comment93868946_53475089\">Shai<\/a>, you could also<\/p>\n\n<pre><code>t = torch.rand(32, 10, 64)[:, -1, :]\n\n\nprint(t.size()) # torch.Size([32, 64])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53474236,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><conv-neural-network><max-pooling>","CreationDate":"2018-11-26T02:58:15.447","AcceptedAnswerId":null,"OwnerUserId":8172168.0,"Title":"how max pool layer decrease\/increase the numbers of feature channels?","Body":"<p>When reading some deep learning papers, which sometimes mentioned that max-pooling layer for downsampling can also be used for decreasing\/increasing the number of feature channels(maps). This confused me a lot. It looks to me the max-pooling layer can down sample the size, but should keep the number of original feature maps.\nI search the usage of max pool in tensorflow, but there is on arguments for change the feature channels. So, could anybody tell me how to decrease the feature channels by max pool layer?\nfor example, max pool changes the number of feature channels in the following picture. \n<a href=\"https:\/\/i.stack.imgur.com\/DVmkr.jpg\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>Hiroshi Fukui, Robust Pedestrian Attribute Recognition for an Unbalanced Dataset Using Mini-Batch Training with Rarity Rate<\/p>\n","answers":[]}
+{"QuestionId":53474497,"AnswerCount":1,"Tags":"<python><machine-learning><keras><generator><rnn>","CreationDate":"2018-11-26T03:39:11.210","AcceptedAnswerId":53478122.0,"OwnerUserId":866082.0,"Title":"What should the generator return if it is used in a multi-model functional API?","Body":"<p>Following <a href=\"https:\/\/blog.keras.io\/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html\" rel=\"nofollow noreferrer\">this article<\/a>, I'm trying to implement a generative RNN. In the mentioned article, the training and validation data are passed as fully loaded <code>np.array<\/code>s. But I'm trying to use the <code>model.fit_generator<\/code> method and provide a generator instead.<\/p>\n\n<p>I know that if it was a straightforward model, the generator should return:<\/p>\n\n<pre><code>def generator():\n    ...\n    yield (samples, targets)\n<\/code><\/pre>\n\n<p>But this is a generative model which means there are two models involved:<\/p>\n\n<pre><code>encoder_inputs = Input(shape=(None,))\nx = Embedding(num_encoder_tokens, embedding_dim)(encoder_inputs)\nx.set_weights([embedding_matrix])\nx.trainable = False\nx, state_h, state_c = LSTM(embedding_dim, return_state=True)(x)\nencoder_states = [state_h, state_c]\n\ndecoder_inputs = Input(shape=(None,))\nx = Embedding(num_decoder_tokens, embedding_dim)(decoder_inputs)\nx.set_weights([embedding_matrix])\nx.trainable = False\nx = LSTM(embedding_dim, return_sequences=True)(x, initial_state=encoder_states)\ndecoder_outputs = Dense(num_decoder_tokens, activation='softmax')(x)\n\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)\n\nmodel.fit([encoder_input_data, decoder_input_data], decoder_target_data,\n          batch_size=batch_size,\n          epochs=epochs,\n          validation_split=0.2)\n<\/code><\/pre>\n\n<p>As mentioned before, I'm trying to use a generator:<\/p>\n\n<pre><code>model.fit_generator(generator(),\n                   steps_per_epoch=500,\n                   epochs=20,\n                   validation_data=generator(),\n                   validation_steps=val_steps)\n<\/code><\/pre>\n\n<p>But what should the <code>generator()<\/code> return? I'm a little confused since there are two input collections and one target.<\/p>\n","answers":[{"AnswerId":"53478122","CreationDate":"2018-11-26T09:30:26.073","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since your model has two inputs and one output, the generator should return a tuple with two elements where the first element is a <strong>list<\/strong> containing two arrays, which corresponds to two input layers, and the second element is an array corresponding to output layer:<\/p>\n\n<pre><code>def generator():\n    ...\n    yield [input_samples1, input_samples2], targets\n<\/code><\/pre>\n\n<p>Generally, in a model with <code>M<\/code> inputs and <code>N<\/code> outputs, the generator should return a tuple of two lists where the first one has <code>M<\/code> arrays and the second one has <code>N<\/code> arrays:<\/p>\n\n<pre><code>def generator():\n        ...\n        yield [in1, in2, ..., inM], [out1, out2, ..., outN]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53474533,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><hdf5><h5py><python-wheel>","CreationDate":"2018-11-26T03:44:51.160","AcceptedAnswerId":null,"OwnerUserId":8782672.0,"Title":"Getting error installing custom build tensorflow 1.12 wheel with python 3.6","Body":"<p>I am trying to install a custom tensorflow 1.12 wheel on ppc64le\nIBM Power9 system with NVIDIA Pascal V100 but getting the following error:<\/p>\n\n<pre><code>gcc -pthread -B \/ccs\/home\/shubhankar\/miniconda3\/compiler_compat -Wl,--sysroot=\/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -DH5_USE_16_API -I.\/h5py -I\/tmp\/pip-install-d__8pdpg\/h5py\/lzf -I\/opt\/local\/include -I\/usr\/local\/include -I\/ccs\/home\/shubhankar\/miniconda3\/lib\/python3.6\/site-packages\/numpy\/core\/include -I\/ccs\/home\/shubhankar\/miniconda3\/include\/python3.6m -c \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/defs.c -o build\/temp.linux-ppc64le-3.6\/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/defs.o\n    In file included from \/ccs\/home\/shubhankar\/miniconda3\/lib\/python3.6\/site-packages\/numpy\/core\/include\/numpy\/ndarraytypes.h:1821:0,\n                     from \/ccs\/home\/shubhankar\/miniconda3\/lib\/python3.6\/site-packages\/numpy\/core\/include\/numpy\/ndarrayobject.h:18,\n                     from \/ccs\/home\/shubhankar\/miniconda3\/lib\/python3.6\/site-packages\/numpy\/core\/include\/numpy\/arrayobject.h:4,\n                     from \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/api_compat.h:26,\n                     from \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/defs.c:657:\n    \/ccs\/home\/shubhankar\/miniconda3\/lib\/python3.6\/site-packages\/numpy\/core\/include\/numpy\/npy_1_7_deprecated_api.h:15:2: warning: #warning \"Using deprecated NumPy API, disable it by \" \"#defining NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION\" [-Wcpp]\n     #warning \"Using deprecated NumPy API, disable it by \" \\\n      ^\n    In file included from \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/defs.c:657:0:\n    \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/api_compat.h:27:18: fatal error: hdf5.h: No such file or directory\n     #include \"hdf5.h\"\n                      ^\n    compilation terminated.\n    error: command 'gcc' failed with exit status 1\n\n    ----------------------------------------\n  Rolling back uninstall of h5py\nCommand \"\/ccs\/home\/shubhankar\/miniconda3\/bin\/python -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-install-d__8pdpg\/h5py\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-record-5_3ofwib\/install-record.txt --single-version-externally-managed --compile\" failed with error code 1 in \/tmp\/pip-install-d__8pdpg\/h5py\/\n<\/code><\/pre>\n\n<p>It breaks at the installation of h5py for some reason. It all starts with a warning message. I am not sure if it is related to it.<\/p>\n\n<pre><code>Cythonizing \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/h5z.pyx\n    \/tmp\/pip-install-d__8pdpg\/h5py\/.eggs\/Cython-0.29.1-py3.6-linux-ppc64le.egg\/Cython\/Compiler\/Main.py:367: FutureWarning: Cython directive 'language_level' not set, using 2 for now (Py2). This will change in a later release! File: \/tmp\/pip-install-d__8pdpg\/h5py\/h5py\/h5z.pxd\n<\/code><\/pre>\n\n<p>And finally<\/p>\n\n<pre><code>Rolling back uninstall of h5py\nCommand \"\/ccs\/home\/shubhankar\/miniconda3\/bin\/python -u -c \"import setuptools, tokenize;__file__='\/tmp\/pip-install-d__8pdpg\/h5py\/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\\r\\n', '\\n');f.close();exec(compile(code, __file__, 'exec'))\" install --record \/tmp\/pip-record-5_3ofwib\/install-record.txt --single-version-externally-managed --compile\" failed with error code 1 in \/tmp\/pip-install-d__8pdpg\/h5py\/\n<\/code><\/pre>\n\n<p>I am using miniconda3 for package management. The system installed gcc version is 4.8.5<\/p>\n","answers":[{"AnswerId":"53484323","CreationDate":"2018-11-26T15:26:41.153","ParentId":null,"OwnerUserId":"7976758","Title":null,"Body":"<p>You're installing <code>tensorflow<\/code> on a non-standard architecture (there are no precompiled wheels <a href=\"https:\/\/pypi.org\/project\/tensorflow\/1.12.0\/#files\" rel=\"nofollow noreferrer\">at PyPI<\/a>) so <code>pip<\/code> has to compile everything (including dependencies, including <code>h5py<\/code>) from sources.<\/p>\n\n<p>To compile <code>h5py<\/code> from sources you need to install HDF5 1.8.4 or newer <strong>with development headers<\/strong>. See <a href=\"http:\/\/docs.h5py.org\/en\/stable\/build.html#source-installation\" rel=\"nofollow noreferrer\">http:\/\/docs.h5py.org\/en\/stable\/build.html#source-installation<\/a> for details.<\/p>\n\n<p>Install <code>libhdf5-dev<\/code> using your OS' package manager if the package is available. Or <a href=\"http:\/\/micro.stanford.edu\/wiki\/Install_HDF5#Build_and_Installation_from_Sources\" rel=\"nofollow noreferrer\">install HDF5 from sources<\/a>.<\/p>\n"}]}
+{"QuestionId":53474967,"AnswerCount":0,"Tags":"<c++11><pytorch><torch><fbconnect><autoreconf>","CreationDate":"2018-11-26T04:55:00.507","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"fblualib installation in Ubunt16.04","Body":"<p>I am trying to install <code>fblualib<\/code> from this <a href=\"https:\/\/github.com\/facebookarchive\/fblualib\" rel=\"nofollow noreferrer\">link<\/a>.<\/p>\n\n<p>When I run <code>.\/install_all.sh<\/code> inside <code>fblualib<\/code>, I have errors as<\/p>\n\n<pre><code>+ echo Building folly\nBuilding folly\n+ echo\n\n+ cd \/tmp\/fblualib-build.zofU4g\/folly\/folly\n+ autoreconf -ivf\nautoreconf: 'configure.ac' or 'configure.in' is required\n<\/code><\/pre>\n\n<p>According to last discussion <a href=\"https:\/\/github.com\/facebook\/folly\/issues\/613\" rel=\"nofollow noreferrer\">here<\/a>, there is a discussion as \"autotools based build system has been removed from Folly in favor of CMake, so you'll need to build Folly with CMake rather than autotools.\" so I go to folly folder <\/p>\n\n<pre><code>\/tmp\/fblualib-build.zofU4g\/folly\/build and try to build using `cmake`.\n<\/code><\/pre>\n\n<p>Then I have errors as<\/p>\n\n<pre><code>[ 36%] Building CXX object CMakeFiles\/folly_base.dir\/folly\/detail\/ThreadLocalDetail.cpp.o\nIn file included from \/tmp\/fblualib-build.zofU4g\/folly\/folly\/detail\/ThreadLocalDetail.h:33:0,\n                 from \/tmp\/fblualib-build.zofU4g\/folly\/folly\/detail\/ThreadLocalDetail.cpp:16:\n\/tmp\/fblualib-build.zofU4g\/folly\/folly\/SharedMutex.h: In static member function \u2018static void folly::threadlocal_detail::StaticMetaBase::onThreadExit(void*)\u2019:\n\/tmp\/fblualib-build.zofU4g\/folly\/folly\/SharedMutex.h:457:33: error: \u2018*((void*)&amp; rlock +10)\u2019 may be used uninitialized in this function [-Werror=maybe-uninitialized]\n         !tryUnlockSharedDeferred(token.slot_)) {\n                                 ^\n\/tmp\/fblualib-build.zofU4g\/folly\/folly\/detail\/ThreadLocalDetail.cpp:169:31: note: \u2018*((void*)&amp; rlock +10)\u2019 was declared here\n       SharedMutex::ReadHolder rlock(nullptr);\n                               ^\ncc1plus: all warnings being treated as errors\nCMakeFiles\/folly_base.dir\/build.make:1886: recipe for target 'CMakeFiles\/folly_base.dir\/folly\/detail\/ThreadLocalDetail.cpp.o' failed\nmake[2]: *** [CMakeFiles\/folly_base.dir\/folly\/detail\/ThreadLocalDetail.cpp.o] Error 1\nCMakeFiles\/Makefile2:67: recipe for target 'CMakeFiles\/folly_base.dir\/all' failed\nmake[1]: *** [CMakeFiles\/folly_base.dir\/all] Error 2\nMakefile:127: recipe for target 'all' failed\nmake: *** [all] Error 2\n<\/code><\/pre>\n\n<p>The error is happening in <code>SharedMutex.h<\/code>, what could be worng><\/p>\n","answers":[]}
+{"QuestionId":53475580,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2018-11-26T06:10:28.387","AcceptedAnswerId":null,"OwnerUserId":5353461.0,"Title":"PyTorch: get input layer size","Body":"<p>I want to programatically find the size of my input layer.<\/p>\n\n<p>If my first layer is called <code>fc1<\/code>, how do I find out its input?<\/p>\n","answers":[{"AnswerId":"53475678","CreationDate":"2018-11-26T06:20:47.203","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>Assuming your model is called <code>model<\/code>, this will give the number of input features of the <code>fc1<\/code> layer:<\/p>\n\n<pre><code>model.fc1.in_features\n<\/code><\/pre>\n\n<p>This is useful inside the <code>.forward()<\/code> method:<\/p>\n\n<pre><code>def forward(self, x):\n    x = x.view(-1, self.fc1.in_features)  # resize the input to match the input layer\n    ...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53475803,"AnswerCount":1,"Tags":"<python><lstm><pytorch><rnn><softmax>","CreationDate":"2018-11-26T06:32:43.673","AcceptedAnswerId":53477502.0,"OwnerUserId":10678961.0,"Title":"Understanding the softmax output layer of RNN","Body":"<p>Here's a simple LSTM model in Keras:<\/p>\n\n<pre><code>input = Input(shape=(max_len,))\nmodel = Embedding(input_dim=input_dim, output_dim=embed_dim, input_length=max_len)(input)\nmodel = Dropout(0.1)(model)\nmodel = Bidirectional(LSTM(units=blstm_dim, return_sequences=True, recurrent_dropout=0.1))(model)\nout =Dense(label_dim, activation=\"softmax\")(model)\n<\/code><\/pre>\n\n<p>Here's my attempt to translate it to a Pytorch model:<\/p>\n\n<pre><code>class RNN(nn.Module):\n    def __init__(self, input_dim, embed_dim, blstm_dim, label_dim):\n        super(RNN, self).__init__()\n        self.embed = nn.Embedding(input_dim, embed_dim)\n        self.blstm = nn.LSTM(embed_dim, blstm_dim,  bidirectional=True, batch_first=True)\n        self.fc = nn.Linear(2*blstm_dim, label_dim)\n\n\n    def forward(self, x):\n        h0 = torch.zeros(2, x.size(0), blstm_dim).to(device) \n        c0 = torch.zeros(2, x.size(0), blstm_dim).to(device)\n        x = self.embed(x)\n        x = F.dropout(x, p=0.1, training=self.training)\n        x,_ = self.blstm(x, (h0, c0))\n        x = self.fc(x)\n        return F.softmax(x, dim=1)\n#        return x\n<\/code><\/pre>\n\n<p>Now running the Keras model gives the following:<\/p>\n\n<pre><code>Epoch 5\/5\n38846\/38846 [==============================] - 87s 2ms\/step - loss: 0.0374 - acc: 0.9889 - val_loss: 0.0473 - val_acc: 0.9859\n<\/code><\/pre>\n\n<p>But running the PyTorch model gives this:<\/p>\n\n<pre><code>Train Epoch: 10\/10 [6400\/34532 (19%)]   Loss: 2.788933\nTrain Epoch: 10\/10 [12800\/34532 (37%)]  Loss: 2.788880\nTrain Epoch: 10\/10 [19200\/34532 (56%)]  Loss: 2.785547\nTrain Epoch: 10\/10 [25600\/34532 (74%)]  Loss: 2.796180\nTrain Epoch: 10\/10 [32000\/34532 (93%)]  Loss: 2.790446\nValidation: Average loss: 0.0437, Accuracy: 308281\/431600 (71%)\n<\/code><\/pre>\n\n<p>I've made sure the loss &amp; optimiser are the same (cross entropy &amp; RMSprop). Now interestingly if I remove the softmax from the PyTorch model (i.e. use the hashed output in the code, I get what seems to be right:<\/p>\n\n<pre><code>Train Epoch: 10\/10 [32000\/34532 (93%)]  Loss: 0.022118\nValidation: Average loss: 0.0009, Accuracy: 424974\/431600 (98%)\n<\/code><\/pre>\n\n<p>So here are my questions:<\/p>\n\n<p>1) Are the two models I've printed above equivalent (let's ignore the recurrent_dropout since I haven't figure out how to do that in PyTorch)?<\/p>\n\n<p>2) What am I doing wrong with the softmax output layer in PyTorch?<\/p>\n\n<p>Thanks a lot!<\/p>\n","answers":[{"AnswerId":"53477502","CreationDate":"2018-11-26T08:52:01.700","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<ol>\n<li><strong><em>Are the two models I've printed above equivalent (let's ignore the recurrent_dropout since I haven't figure out how to do that in PyTorch)?<\/em><\/strong><\/li>\n<\/ol>\n\n<p>Besides the dropout I can see no difference. So they should be completely equivalent in terms of structure. <\/p>\n\n<p><em>One note: You don't have to initialize the states if you use it this way (if you're not reusing the states). You can just forward the LSTM with <code>x,_ = self.blstm(x)<\/code> - it will automatically initialize states with zeros.<\/em> <\/p>\n\n<ol start=\"2\">\n<li><strong><em>What am I doing wrong with the softmax output layer in PyTorch?<\/em><\/strong><\/li>\n<\/ol>\n\n<p>PyTorch <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\"><code>torch.nn.CrossEntropyLoss<\/code><\/a> already includes softmax: <\/p>\n\n<blockquote>\n  <p>This criterion combines <code>nn.LogSoftmax()<\/code> and <code>nn.NLLLoss()<\/code> in one\n  single class.<\/p>\n<\/blockquote>\n\n<p>So it is actually a CE with logits. I guess this makes it more efficient. So you can just leave out the softmax activation at the end.<\/p>\n"}]}
+{"QuestionId":53476098,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-26T07:00:34.220","AcceptedAnswerId":53531590.0,"OwnerUserId":10004761.0,"Title":"Training VGG16 with CIFAR-10 dataset in TensorFlow, can't get good accuracy","Body":"<p>I am a beginner to study ML. I am studying the VGG16 model in TensorFlow. My study is <a href=\"https:\/\/github.com\/NienfengYao\/TensorFlow_Study\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Now I meet a question and still can't find the reason. I studied the two case.<\/p>\n\n<ol>\n<li>Study pre-trained VGG16-A model to do predict. (<a href=\"https:\/\/www.cs.toronto.edu\/~frossard\/post\/vgg16\/\" rel=\"nofollow noreferrer\">VGG in TensorFlow<\/a>)<\/li>\n<li>Study VGG16-B model with CIFAR-10 dataset to do training. (<a href=\"https:\/\/github.com\/Ashing00\/Tensorflow_VGG\" rel=\"nofollow noreferrer\">Ashing00\/Tensorflow_VGG<\/a>)<\/li>\n<\/ol>\n\n<p>They all work fine. But now I want to simulate the Case 2.'s procedure, try to use VGG16-A model to do training with CIFAR-10 dataset. But I am failed, I can't get good accuracy.<\/p>\n\n<p>The testing is in <a href=\"https:\/\/github.com\/NienfengYao\/TensorFlow_Study\/blob\/master\/vgg16\/vgg16_final\/vgg16f_training_cifar10.ipynb\" rel=\"nofollow noreferrer\">vgg16f_training_cifar10.ipynb<\/a>. Can someone give me some hint?\nThanks a lot.<\/p>\n","answers":[{"AnswerId":"53531590","CreationDate":"2018-11-29T03:53:03.463","ParentId":null,"OwnerUserId":"10004761","Title":null,"Body":"<p>I add batch normalization in each hidden layer. Now we can get good accuracy. <\/p>\n"}]}
+{"QuestionId":53476305,"AnswerCount":2,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2018-11-26T07:18:08.080","AcceptedAnswerId":53476678.0,"OwnerUserId":4409660.0,"Title":"AttributeError: 'tuple' object has no attribute 'log_softmax'","Body":"<p>while trying to finetune inception_V3 for my own dataset by changing the last fc layer like <\/p>\n\n<pre><code>last_layer =nn.Linear(n_inputs, len(classes))\ninception_v3.fc = last_layer\n<\/code><\/pre>\n\n<p>after that when I train it got this error on this position<\/p>\n\n<pre><code>    # on training loop\n    output = inception_v3(data)\n    # calculate the batch loss\n    loss = criterion(output, target)\n<\/code><\/pre>\n\n<p>Error is <\/p>\n\n<pre><code> AttributeError: 'tuple' object has no attribute 'log_softmax'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53476678","CreationDate":"2018-11-26T07:48:37.287","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>This is a well known problem.  <\/p>\n\n<p>Try one the following solutions:  <\/p>\n\n<ol>\n<li><p>disable aux_logits when the model is created here by also passing <code>aux_logits=False<\/code> to the inception_v3 function.<\/p><\/li>\n<li><p>edit your train function to accept and unpack the returned tuple to be something like: <code>output, aux = model(input_var)<\/code><\/p><\/li>\n<\/ol>\n\n<p>Check the following <a href=\"https:\/\/github.com\/pytorch\/vision\/issues\/302#issuecomment-341163548\" rel=\"nofollow noreferrer\">link<\/a> for more info.<\/p>\n"},{"AnswerId":"56943518","CreationDate":"2019-07-08T23:09:02.193","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>This problem seams to me like instead you define <code>F<\/code>:<\/p>\n\n<pre><code>import torch.nn.functional as F\n<\/code><\/pre>\n\n<p>You in accident have set <code>F<\/code> to some tuple<\/p>\n\n<pre><code>F=(1,2)\n<\/code><\/pre>\n\n<p>And then when you call <code>F.log_softmax<\/code> you get exactly this error.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fGlhw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fGlhw.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53476633,"AnswerCount":1,"Tags":"<keras><lstm><text-classification>","CreationDate":"2018-11-26T07:45:11.873","AcceptedAnswerId":null,"OwnerUserId":6272481.0,"Title":"LSTM model weights to train data for text classification","Body":"<p>I built a LSTM model for text classification using Keras. Now I have new data to be trained. instead of appending to the original data and retrain the model, I thought of training the data using the model weights. i.e. making the weights to get trained with the new data. \nHowever, irrespective of the volume i train, the model is not predicting the correct classification (even if i give the same sentence for prediction). What could be the reason?\nKindly help me.<\/p>\n","answers":[{"AnswerId":"53490074","CreationDate":"2018-11-26T22:28:11.617","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>Are you using the following to save the trained model?<\/p>\n\n<pre><code>model.save('model.h5')\nmodel.save_weights('model_weights.h5')\n<\/code><\/pre>\n\n<p>And the following to load it?<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('model.h5') # Load the architecture\nmodel = model.load_weights('model_weights.h5') # Set the weights\n\n# train on new data\nmodel.compile...\nmodel.fit...\n<\/code><\/pre>\n\n<p>The model loaded is the exact same as the model being saved here. If you are doing this, then there must be something different in the data (in comparison with what it is trained on).<\/p>\n"}]}
+{"QuestionId":53476720,"AnswerCount":0,"Tags":"<python><pandas><opencv><tensorflow><keras>","CreationDate":"2018-11-26T07:51:31.817","AcceptedAnswerId":null,"OwnerUserId":7810784.0,"Title":"Reading images to create dataset for image classification","Body":"<p>I want to train a classifier based in keras and tensorflow on my images data.<\/p>\n\n<p>The images are located as sub folders inside folders:<\/p>\n\n<pre><code>data\n    train\n        terror\n        no_terror\n    test\n        terror\n        no_terror\n<\/code><\/pre>\n\n<p>The problem here is that, I have all the images which are of different sizes.<\/p>\n\n<p>I am using the code below.<\/p>\n\n<pre><code>TERROR_TRAIN_DIR = 'C:\/\/Users\/shiva\/OneDrive\/Desktop\/Dhivy\/data\/train\/terror\/'\nTERROR_TEST_DIR = 'C:\/\/Users\/shiva\/OneDrive\/Desktop\/Dhivy\/data\/test\/terror\/'\n\nNON_TERROR_TRAIN_DIR = 'C:\/\/Users\/shiva\/OneDrive\/Desktop\/Dhivy\/data\/train\/non-terror\/'\nNON_TERROR_TEST_DIR = 'C:\/\/Users\/shiva\/OneDrive\/Desktop\/Dhivy\/data\/test\/non-terror\/'\n\nROWS = 64\nCOLS = 64\nCHANNELS = 3\n\nterror_train_images = [TERROR_TRAIN_DIR+i for i in os.listdir(TERROR_TRAIN_DIR)] \nterror_test_images = [TERROR_TEST_DIR+i for i in os.listdir(TERROR_TEST_DIR)] \n\nnon_terror_train_images = [NON_TERROR_TRAIN_DIR+i for i in os.listdir(NON_TERROR_TRAIN_DIR)] \nnon_terror_test_images = [NON_TERROR_TEST_DIR+i for i in os.listdir(NON_TERROR_TEST_DIR)] \n\n\n\ndef read_image(file_path):\n    img = cv2.imread(file_path, cv2.IMREAD_COLOR) #cv2.IMREAD_GRAYSCALE\n    return cv2.resize(img, (ROWS, COLS), interpolation=cv2.INTER_CUBIC)\n\n\ndef prep_data(images):\n    count = len(images)\n    data = np.ndarray((count, CHANNELS, ROWS, COLS), dtype=np.uint8)\n\n    for i, image_file in enumerate(images):\n        image = read_image(image_file)\n        data[i] = image.T\n        if i%250 == 0: print('Processed {} of {}'.format(i, count))\n\n    return data\n\nterror_train_images = prep_data(terror_train_images)\nterror_test_images = prep_data(terror_test_images)\n\nnon_terror_train_images = prep_data(non_terror_train_images)\nnon_terror_test_images = prep_data(non_terror_test_images)\n<\/code><\/pre>\n\n<p>I am getting the error shown below.<\/p>\n\n<pre><code>&lt;ipython-input-27-4eadf011e89d&gt; in read_image(file_path)\n      1 def read_image(file_path):\n      2     img = cv2.imread(file_path, cv2.IMREAD_COLOR) #cv2.IMREAD_GRAYSCALE\n----&gt; 3     return cv2.resize(img, (ROWS, COLS), interpolation=cv2.INTER_CUBIC)\n      4 \n      5 \n\nerror: OpenCV(3.4.3) C:\\projects\\opencv-python\\opencv\\modules\\imgproc\\src\\resize.cpp:4044: error: (-215:Assertion failed) !ssize.empty() in function 'cv::resize'\n<\/code><\/pre>\n\n<p>Is there any way, I can create train and test dataset with labels as well, from the folder structure itself and create X_train, y_train, X_test and y_test, which I can feed directly into VGG and Resnet, my main gaol is to test how these images perform on these pre-trained models.<\/p>\n\n<p>Thanks in Advance<\/p>\n","answers":[]}
+{"QuestionId":53476735,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><conv-neural-network><tensorflow-estimator>","CreationDate":"2018-11-26T07:52:13.203","AcceptedAnswerId":null,"OwnerUserId":9044016.0,"Title":"InvalidArgumentError : Shape Mismatch during Computing Ops","Body":"<p>I am creating Custome Estimator API for a Time Series data. There occurs no error during TensorFlow graph Construction, but I get the mismatch shape error during the graph Computation. Before posting the code let me explain the little bit about the dataset I been using. The data I am using to create Custom Estimator is <strong>Time Series data<\/strong> which have two years history. <\/p>\n\n<p>Dataset Shape is <strong>(13, 731)<\/strong><\/p>\n\n<p>My Goal is to predict the next 30 days store traffic.<\/p>\n\n<p>I decided to use  <strong>Convolutional Neural Network for High dimensional Time-series Forecasting.<\/strong> I took 701 data points for training and next 30 its corresponding label data points. I wrote following <strong>csv_input_fn<\/strong> to pass input data into Estimator.<\/p>\n\n<pre><code>def csv_input_fn(csv_path, batch_size):\n    #input function\n    def _input_fn():\n        input_filename= tf.train.match_filenames_once(csv_path)\n\n        filename_queue= tf.train.string_input_producer(\n            input_filename, num_epochs=None, shuffle=True)\n\n        reader = tf.TextLineReader()\n        _, value =  reader.read_up_to(filename_queue, num_records= batch_size)\n        value_column = tf.expand_dims(value, -1)\n\n        all_data= tf.decode_csv(value_column, record_defaults=CSV_TYPES)\n        inputs= all_data[:len(all_data) - pred_steps]\n        labels= all_data[len(all_data) - pred_steps: ]\n\n        inputs= tf.concat(inputs, axis=1)\n        labels= tf.concat(labels, axis=1)\n\n        return{'raw_data': inputs}, labels\n\n\n    return _input_fn\n<\/code><\/pre>\n\n<p>where <code>pred_steps=30<\/code> as I have to predict the last 30 days traffic. I have created my custom NN model which have <strong>8 conv1d layers with dilation rates<\/strong> followed by <strong>dense<\/strong>, followed by a <strong>dropout<\/strong>, followed by another <strong>dense<\/strong> layer, and at last final layer. This the <strong>cnn_model_fn<\/strong> i created<\/p>\n\n<pre><code>def cnn_model_fn(features, labels, mode):\n\n    #setup  the mode\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        tf.logging.info(\"My Model Function: PREDICT, {}\".format(mode))\n    elif mode == tf.estimator.ModeKeys.EVAL:\n        tf.logging.info(\"My Model Function: EVAL, {}\".format(mode))\n    elif mode ==  tf.estimator.ModeKeys.TRAIN:\n        tf.logging.info(\"My Model Function: TRAIN, {}\".format(mode))\n\n\n    #set up the initializer\n    #Input layer\n\n    input_layer = tf.reshape(features['raw_data'], [-1, total_len_inp_features,1])\n\n    #conv1 layer\n    conv1 = tf.layers.conv1d(input_layer, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[0])\n\n    #conv2 layer\n    conv2 = tf.layers.conv1d(conv1, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[1])\n\n    #conv3 layer\n    conv3 = tf.layers.conv1d(conv2, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[2])\n\n    #conv4 layer\n    conv4 = tf.layers.conv1d(conv3, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[3])\n\n    #conv5 layer\n    conv5 = tf.layers.conv1d(conv4, filters=n_filters,\n        kernel_size= filter_width,\n        padding= 'same',\n        dilation_rate= dilation_rates[4])\n\n    #conv6 layer\n    conv6 = tf.layers.conv1d(conv5, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[5])\n\n    #conv7 layer\n    conv7 = tf.layers.conv1d(conv6, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[6])\n\n    #conv8 layer\n    conv8 = tf.layers.conv1d(conv7, filters=n_filters,\n        kernel_size= filter_width,\n        padding='same',\n        dilation_rate= dilation_rates[7])\n\n    #add dense layer\n    dense_1 = tf.layers.dense(conv8, units=128,\n        activation= tf.nn.relu)\n\n    #add dropout\n    dropout= tf.layers.dropout(dense_1, rate=0.4)\n\n    #add dense layer\n    dense_2 = tf.layers.dense(dropout, units=1)\n\n    #output layer\n    outlen= tf.reshape(dense_2, [-1, len(OUTPUT_COLUMN_NAMES)])\n\n    #predictions\n    predictions= tf.layers.dense(outlen, 30, activation= None)\n<\/code><\/pre>\n\n<p>while running the model, I face no error during construction of the graph, but facing the following error during the computation of the graph.<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Input to reshape is a\n  tensor with 9113 values, but the requested shape requires a multiple\n  of 30\n           [[Node: Reshape_1 = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](dense_1\/BiasAdd,\n  Reshape_1\/shape)]]\n           [[Node: mean_squared_error\/value\/_227 = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\",\n  send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\",\n  send_device_incarnation=1,\n  tensor_name=\"edge_1825_mean_squared_error\/value\",\n  tensor_type=DT_FLOAT,\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>while Troubleshooting this problem I referred the link <a href=\"https:\/\/stackoverflow.com\/questions\/49335990\/no-shape-error-in-tensorflow-graph-construction-but-getting-shape-mismatch-error\">Invalid Argument Error. Shape Mismatch while computing gradients<\/a>. I am sure, that my program is throwing error only because of <strong>Shape Mismatch.<\/strong> Since I am new to build Custom Estimators I have problems in figuring out the exact solutions.<\/p>\n\n<p>For your reference, I am attaching the link <a href=\"https:\/\/github.com\/Madhivarman\/Custom-Estimator-API\/blob\/master\/Source\/Conv1D%2Bdilation.py\" rel=\"nofollow noreferrer\">Full code<\/a><\/p>\n","answers":[]}
+{"QuestionId":53476874,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2018-11-26T08:05:31.877","AcceptedAnswerId":null,"OwnerUserId":1300084.0,"Title":"Using GPU error when use TensorFlow to train image","Body":"<p>When I am runing a tensorflow image train job in the container tensorflow\/tensorflow:latest-gpu, it doesn't work.<\/p>\n\n<p>Error message:<\/p>\n\n<pre><code>Cannot assign a device for operation InceptionV3\/InceptionV3\/Conv2d_1a_3x3\/Conv2D: Operation was explicitly assigned to \/device:GPU:0 but available devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0, \/job:localhost\/replica:0\/task:0\/device:XLA_CPU:0, \/job:localhost\/replica:0\/task:0\/device:XLA_GPU:0 ]. Make sure the device specification refers to a valid device.\n[[node InceptionV3\/InceptionV3\/Conv2d_1a_3x3\/Conv2D (defined at \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py:1057)  = Conv2D[T=DT_FLOAT, data_format=\"NHWC\", dilations=[1, 1, 1, 1], padding=\"VALID\", strides=[1, 2, 2, 1], use_cudnn_on_gpu=true, _device=\"\/device:GPU:0\"](fifo_queue_Dequeue, InceptionV3\/Conv2d_1a_3x3\/weights\/read)]]\n<\/code><\/pre>\n\n<p>GPU info:\nnvidia-smi\nMon Nov 26 07:48:59 2018<br>\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 410.72       Driver Version: 410.72       CUDA Version: 10.0     |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  GeForce GT 630      Off  | 00000000:01:00.0 N\/A |                  N\/A |\n| 25%   47C    P0    N\/A \/  N\/A |      0MiB \/  1998MiB |     N\/A      Default |\n+-------------------------------+----------------------+----------------------+<\/p>\n\n<p>+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0                    Not Supported                                       |\n+-----------------------------------------------------------------------------+<\/p>\n","answers":[{"AnswerId":"56442520","CreationDate":"2019-06-04T10:59:13.293","ParentId":null,"OwnerUserId":"8472265","Title":null,"Body":"<p>It seems that you Tensorflow is not detecting any gpu as available but maps the operations to GPU:0. First try this: <\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n<\/code><\/pre>\n\n<p>And you'll get the available devices. Is there <code>\/device:GPU:0<\/code> ?<\/p>\n"}]}
+{"QuestionId":53477005,"AnswerCount":3,"Tags":"<python-3.x><tensorflow><spyder>","CreationDate":"2018-11-26T08:15:26.160","AcceptedAnswerId":54864247.0,"OwnerUserId":9805724.0,"Title":"Tensorflow r1.12: TypeError: Type already registered for SparseTensorValue when running a 2nd script","Body":"<p>I have just built Tensorflow r1.12 from source in Ubuntu 16.04. The installation is successful.<\/p>\n\n<p>When I run a certain script in Spyder at the 1st time, everything flows smoothly.\nHowever, when I continue to run another script, following errors occur (which didn't happen previously):<\/p>\n\n<blockquote>\n  <p>File \"\/home\/haohua\/tf_env\/lib\/python3.6\/site-packages\/tensorflow\/<strong>init<\/strong>.py\", line 24, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import<\/p>\n  \n  <p>File \"\/home\/haohua\/tf_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/<strong>init<\/strong>.py\", line 70, in \n      from tensorflow.python.framework.framework_lib import *  # pylint: disable=redefined-builtin<\/p>\n  \n  <p>File \"\/home\/haohua\/tf_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/framework_lib.py\", line 30, in \n      from tensorflow.python.framework.sparse_tensor import SparseTensor<\/p>\n  \n  <p>File \"\/home\/haohua\/tf_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/sparse_tensor.py\", line 248, in \n      pywrap_tensorflow.RegisterType(\"SparseTensorValue\", SparseTensorValue)<\/p>\n  \n  <p>TypeError: Type already registered for SparseTensorValue <\/p>\n<\/blockquote>\n\n<p>The temporary solution to avoid such TypeError is to restart the kernel.<\/p>\n\n<p>But I don't want to restart kernel at every single step of running a script.\nThus, I would like to ask for a critical solution for such kind of issue. Thank you in advance. <\/p>\n","answers":[{"AnswerId":"54121680","CreationDate":"2019-01-10T03:44:43.933","ParentId":null,"OwnerUserId":"10893132","Title":null,"Body":"<p>Get rid of the import statement:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>import tensorflow as tf<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Seems to work after that... it's kinda janky<\/p>\n"},{"AnswerId":"54860749","CreationDate":"2019-02-25T06:40:30.540","ParentId":null,"OwnerUserId":"8288145","Title":null,"Body":"<pre><code>import tensorflow as tf\n\nnode1 = tf.constant(3.0,tf.float32)\nnode2 = tf.constant(4.0)\nsess = tf.Session()\n\nprint(sess.run([node1,node2]))\n\nsess.close()\n<\/code><\/pre>\n\n<p>When you run this code the first time it will show the output, but when you run it for the second time it will show the error. For that, you have to select the whole program except <code>import tensorflow as tf<\/code> and run in <\/p>\n\n<blockquote>\n  <p>run in current cell mode<\/p>\n<\/blockquote>\n\n<p>it will work and it will show the output; <\/p>\n\n<p>Otherwise, restart your kernel it will work.<\/p>\n"},{"AnswerId":"54864247","CreationDate":"2019-02-25T10:35:32.080","ParentId":null,"OwnerUserId":"438386","Title":null,"Body":"<p>(<em>Spyder maintainer here<\/em>) This error was fixed in Spyder <strong>3.3.3<\/strong>, released on February\/2019.<\/p>\n"}]}
+{"QuestionId":53477565,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2018-11-26T08:56:32.150","AcceptedAnswerId":53491762.0,"OwnerUserId":6693969.0,"Title":"Input Shape for 1D CNN (Keras)","Body":"<p>I'm building a CNN using Keras, with the following Conv1D as my first layer:<\/p>\n\n<pre><code>cnn.add(Conv1D(\n    filters=512,\n    kernel_size=3,\n    strides=2,\n    activation=hyperparameters[\"activation_fn\"],\n    kernel_regularizer=getattr(regularizers, hyperparameters[\"regularization\"])(hyperparameters[\"regularization_rate\"]),\n    input_shape=(1000, 1),\n))\n<\/code><\/pre>\n\n<p>I'm training with the function:<\/p>\n\n<pre><code>cnn.fit(\n    x=train_df[\"payload\"].tolist(),\n    y=train_df[\"label\"].tolist(),\n    batch_size=hyperparameters[\"batch_size\"],\n    epochs=hyperparameters[\"epochs\"],\n)\n<\/code><\/pre>\n\n<p>In which train_df is a pandas dataframe of two columns where, for each row, label is an int (0 or 1) and payload is a ndarray of floats padded with zeros\/truncated to a length of 1000.\nThe total # of training examples within train_df is 15641.<\/p>\n\n<p>The model compiles, but during training, I get this error:<\/p>\n\n<pre><code>ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 15641 arrays: [array([[0.09019608],\n   [0.01176471],\n   [0.01176471],\n   [0.        ],\n   [0.30196078],\n   [0.        ],\n   [0.        ],\n   [0.        ],\n   [0.        ],\n   [0....\n<\/code><\/pre>\n\n<p>I looked at <a href=\"https:\/\/stackoverflow.com\/questions\/42596057\/keras-error-expected-to-see-1-array\">this post<\/a> and tried changing my input to a ndarray of 1000-float-long lists, but ended up with another error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected conv1d_1_input to have 3 dimensions, but got array with shape (15641, 1000)\n<\/code><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"53491762","CreationDate":"2018-11-27T02:09:35.527","ParentId":null,"OwnerUserId":"6693969","Title":null,"Body":"<p>So I set the input_shape to (1000, 1)<\/p>\n\n<p>I also converted the input that's fed to fit() into a single ndarray of n ndarrays (each ndarray is a vector of 1000 floats, n is the total count of samples\/vectors) and reshaped each of those ndarrays to (1, 1000, 1) during preprocessing after reading <a href=\"https:\/\/stackoverflow.com\/a\/51854675\/6693969\">this<\/a> explanation on inputs &amp; input shape<\/p>\n\n<p>The final shape of my input data was (15641, 1000, 1)<\/p>\n\n<p>** All of this should apply to validation data too (if specified) **<\/p>\n\n<p>EDIT: Sry I didn't state it, but this fixed my issue<\/p>\n"}]}
+{"QuestionId":53477828,"AnswerCount":0,"Tags":"<python><tensorflow><deeplab>","CreationDate":"2018-11-26T09:12:37.563","AcceptedAnswerId":null,"OwnerUserId":7102778.0,"Title":"deeplabv3+ mean subtraction implementated?","Body":"<p>I'm currently using deeplabv3+ on my own dataset, therefore I took a closer look at code of some utility functions for preprocessing. It seems, that the mean subtraction is defined in feature_extractor.py but not used in input_preprocess.py which is called by input_generator.py. The train.py does not do the mean subtraction either, it just calls input_generator where no mean subtraction function is called.\nI'm wondering did I miss something? Is it implemented in somewhere else or? Thanks for the reply! <\/p>\n","answers":[]}
+{"QuestionId":53477861,"AnswerCount":2,"Tags":"<python-3.x><image-processing><machine-learning><deep-learning><pytorch>","CreationDate":"2018-11-26T09:14:38.563","AcceptedAnswerId":53481980.0,"OwnerUserId":8279690.0,"Title":"Pytorch DataLoader multiple data source","Body":"<p>I am trying to use Pytorch dataloader to define my own dataset, but I am not sure how to load multiple data source:<\/p>\n\n<p>My current code:<\/p>\n\n<pre><code>class MultipleSourceDataSet(Dataset):\n    def __init__ (self, json_file, root_dir, transform = None):\n        with open(root_dir + 'block0.json') as f:\n            self.result = torch.Tensor(json.load(f))\n\n    self.root_dir = root_dir\n    self.transform = transform\n\n    def __len__(self):\n        return len(self.result[0])\n\n    def __getitem__ (self):\n        None\n<\/code><\/pre>\n\n<p>The data source is 50 blocks under <code>root_dir = ~\/Documents\/blocks\/<\/code><\/p>\n\n<p>I split them and avoid to combine them directly before since this is a very big dataset.<\/p>\n\n<p>How can I load them into a single dataloader? <\/p>\n","answers":[{"AnswerId":"53481980","CreationDate":"2018-11-26T13:17:27.633","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>For <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>DataLoader<\/code><\/a> you need to have a single <code>Dataset<\/code>, your problem is that you have multiple <code>'json'<\/code> files and you only know how to create a <code>Dataset<\/code> from each <code>'json'<\/code> separately.<br>\nWhat you can do in this case is to use <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.ConcatDataset\" rel=\"nofollow noreferrer\"><code>ConcatDataset<\/code><\/a> that contains all the single-<code>'json'<\/code> datasets you create:<\/p>\n\n<pre><code>import os\nimport torch.utils.data as data\n\nclass SingeJsonDataset(data.Dataset):\n    # implement a single json dataset here...\n\nlist_of_datasets = []\nfor j in os.path.listdir(root_dir):\n    if not j.endswith('.json'):\n        continue  # skip non-json files\n    list_of_datasets.append(SingeJsonDataset(json_file=j, root_dir=root_dir, transform=None))\n# once all single json datasets are created you can concat them into a single one:\nmultiple_json_dataset = data.ConcatDataset(list_of_datasets)\n<\/code><\/pre>\n\n<p>Now you can feed the concatenated dataset into <code>data.DataLoader<\/code>.<\/p>\n"},{"AnswerId":"53513689","CreationDate":"2018-11-28T06:50:05.293","ParentId":null,"OwnerUserId":"8279690","Title":null,"Body":"<p>I should revise my question as 2 different sub-questions:<\/p>\n\n<ol>\n<li>How to deal with large datasets in PyTorch to avoid memory error<\/li>\n<li><p>If I am separating large a dataset into small chunks, how can I load multiple mini-datasets<\/p>\n\n<p>For <strong>question 1<\/strong>: <\/p>\n\n<p>PyTorch DataLoader can prevent this issue by creating mini-batches. <a href=\"https:\/\/discuss.pytorch.org\/t\/dealing-with-large-dataset-without-out-of-memory-error\/19057\" rel=\"nofollow noreferrer\">Here<\/a> you can find further explanations.<\/p>\n\n<p>For <strong>question 2<\/strong>:<\/p>\n\n<p>Please refer to Shai's answer above.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":53478622,"AnswerCount":2,"Tags":"<python><machine-learning><keras><neural-network><training-data>","CreationDate":"2018-11-26T09:58:49.160","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"Stop Keras Training when the network has fully converge","Body":"<p>How will I configure Keras to stop training until convergence or when the loss is 0? I intendedly want to overfit it. I don't want to set number of epochs. I just wanted it to stop when it converges.<\/p>\n","answers":[{"AnswerId":"53479368","CreationDate":"2018-11-26T10:42:41.827","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Use an <a href=\"https:\/\/keras.io\/callbacks\/#earlystopping\" rel=\"nofollow noreferrer\">EarlyStopping<\/a> callback. You may freely choose which loss\/metric to observe and when to stop.<\/p>\n\n<p>Usually, you would look at the \"validation loss\" (<code>val_loss<\/code>), as this is the most important var that tells that your model is still learning to generalize.   <\/p>\n\n<p>But since you said you want to overfit, then you may look at the \"training loss\" (<code>loss<\/code>). <\/p>\n\n<p>The callback works with \"deltas\", not with absolute values, which is good, because the loss doesn't necessarily has \"zero\" as its goal. But you can use the <code>baseline<\/code> argument for setting absolute values.<\/p>\n\n<p>So, usually, a callback that looks at the validation loss:<\/p>\n\n<pre><code>from keras.callbacks import EarlyStopping\nusualCallback = EarlyStopping()\n<\/code><\/pre>\n\n<p>This is the same as <code>EarlyStopping(monitor='val_loss', min_delta=0, patience=0)<\/code><\/p>\n\n<p>One that will overfit: <\/p>\n\n<pre><code>overfitCallback = EarlyStopping(monitor='loss', min_delta=0, patience = 20)\n<\/code><\/pre>\n\n<p>Watch out for the <code>paticence<\/code> argument, it's important as the loss value doesn't always decrease at every epoch. Let the model keep trying for a few more epochs before ending.<\/p>\n\n<p>Finally, just pass the callback to <code>fit<\/code> along with a huge number of epochs:<\/p>\n\n<pre><code>model.fit(X, Y, epochs=100000000, callbacks=[overfitCallback])\n<\/code><\/pre>\n"},{"AnswerId":"53483346","CreationDate":"2018-11-26T14:33:09.347","ParentId":null,"OwnerUserId":"2942154","Title":null,"Body":"<p>An EarlyStopping will do exactly what you want: it helps you to stop the training when the monitored quantity (loss) has stopped improving. This is done using the <strong>patience<\/strong> parameter giving the number of epochs after which, if no improvement is noticed (~possible convergence), the training should stop.\nEarlyStopping usage information can be also found at the <a href=\"https:\/\/stackoverflow.com\/questions\/43906048\/keras-early-stopping\">possible duplicate<\/a>.<\/p>\n\n<p>What is helpful is also <a href=\"https:\/\/machinelearningmastery.com\/display-deep-learning-model-training-history-in-keras\/\" rel=\"nofollow noreferrer\">to vizualize the training process<\/a>. <\/p>\n"}]}
+{"QuestionId":53479007,"AnswerCount":2,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-26T10:20:53.223","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"How to Setup Adaptive Learning Rate in Keras","Body":"<p>Below is my code: <\/p>\n\n<pre><code>model = Sequential([\n    Dense(32, input_shape=(32,), activation = 'relu'),\n    Dense(100, activation='relu'),\n    Dense(65, input_shape=(65,), activation='softmax')\n])\n\n\n\n\nmodel.summary()\nmodel.compile(SGD(lr=.1), loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(train_samples, train_labels, batch_size=1000, epochs=1000,shuffle = True, verbose=2)\n<\/code><\/pre>\n\n<p>How will I set an adaptive learning rate of my model?<\/p>\n","answers":[{"AnswerId":"53479052","CreationDate":"2018-11-26T10:23:17.850","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>You need to replace <code>SGD<\/code> here <\/p>\n\n<pre><code>model.compile(SGD(lr=.1), loss='binary_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>with one of the provided optimizers, e.g. Adam:<\/p>\n\n<pre><code>model.compile(Adam(lr=.1), loss='binary_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Read this <a href=\"https:\/\/keras.io\/optimizers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/optimizers\/<\/a><\/p>\n"},{"AnswerId":"56625691","CreationDate":"2019-06-17T06:24:09.090","ParentId":null,"OwnerUserId":"11108743","Title":null,"Body":"<p>You may use a workaround.<\/p>\n\n<p>For each_iteration in range(0, MaxEpoch):<\/p>\n\n<ol>\n<li><p>Specify your own learning rate function that outputs a learning rate lr with respect to per epoch. The lr is then passed to your_optimiser<\/p><\/li>\n<li><p>run model.compile(...optimizer=your_optimiser...)<\/p><\/li>\n<li><p>run model.fit(...epochs = 1...)<\/p><\/li>\n<li><p>After the ONE epoch, use model.save_weights(...)<\/p><\/li>\n<li><p>Load weights by model.load_weights(...) for next iteration. See here for details <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model<\/a><\/p><\/li>\n<\/ol>\n\n<p>In fact, #4 and #5 enables you to do a transfer learning<\/p>\n"}]}
+{"QuestionId":53479315,"AnswerCount":2,"Tags":"<python><numpy><reshape><pytorch><tensor>","CreationDate":"2018-11-26T10:39:17.473","AcceptedAnswerId":53940813.0,"OwnerUserId":610569.0,"Title":"Munging PyTorch's tensor shape from (C, B, H) to (B, C*H)","Body":"<p>Given an input tensor of shape (C, B, H) <code>torch.Size([2, 5, 32])<\/code> of some neural net layers, where<\/p>\n\n<ul>\n<li><code>channels<\/code> = 2<\/li>\n<li><code>batch_size<\/code> = 5<\/li>\n<li><code>hidden_size<\/code> = 32<\/li>\n<\/ul>\n\n<p>The goal is to flatten the channels and manipulate the input tensor to the shape (B, C*H) <code>torch.Size([5, 2 * 32])<\/code>, where:<\/p>\n\n<ul>\n<li><code>batch_size<\/code> = 5<\/li>\n<li><code>hidden_size<\/code> = 32 * 2<\/li>\n<\/ul>\n\n<p>I've tried to do the following: <\/p>\n\n<pre><code>import torch\n\nt = torch.rand([2, 5, 32])\n\n# Changed from (channels, batch_size, hidden_size) \n# -&gt; (batch_size, channels, hidden_size)\nt = t.permute(1, 0, 2)\n\n# Reshape using view(), where batch_size is t.size(0) \n# and -1 is to flatten the left over values to the other dimension.\nz = t.contiguous().view(t.size(0), -1)\n\nprint(z.shape)\nprint(z)\n<\/code><\/pre>\n\n<p>[out]:<\/p>\n\n<pre><code>torch.Size([5, 64])\ntensor([[0.3911, 0.9586, 0.2104, 0.3937, 0.9976, 0.3378, 0.0630, 0.6676, 0.0806,\n         0.9311, 0.5219, 0.1697, 0.7442, 0.5162, 0.2555, 0.0826, 0.5502, 0.9700,\n         0.3375, 0.5012, 0.9025, 0.8176, 0.1465, 0.1848, 0.3460, 0.9999, 0.7892,\n         0.7577, 0.6615, 0.2620, 0.6868, 0.2003, 0.4840, 0.8354, 0.9253, 0.3172,\n         0.9516, 0.8962, 0.1272, 0.2268, 0.6510, 0.5166, 0.6772, 0.9616, 0.9826,\n         0.5254, 0.9191, 0.4378, 0.7048, 0.8808, 0.0299, 0.1102, 0.9710, 0.8714,\n         0.7256, 0.9684, 0.6117, 0.1957, 0.8663, 0.4742, 0.2843, 0.6548, 0.9592,\n         0.1559],\n        [0.2333, 0.0858, 0.5284, 0.2965, 0.3863, 0.3370, 0.6940, 0.3387, 0.3513,\n         0.1022, 0.3731, 0.3575, 0.7095, 0.0053, 0.7024, 0.4091, 0.3289, 0.5808,\n         0.5640, 0.8847, 0.7584, 0.8878, 0.9873, 0.0525, 0.7731, 0.2501, 0.9926,\n         0.5226, 0.0925, 0.0300, 0.4176, 0.0456, 0.4643, 0.4497, 0.5920, 0.9519,\n         0.6647, 0.2379, 0.4927, 0.9666, 0.1675, 0.9887, 0.7741, 0.5668, 0.7376,\n         0.4452, 0.7449, 0.1298, 0.9065, 0.3561, 0.5813, 0.1439, 0.2115, 0.5874,\n         0.2038, 0.1066, 0.3843, 0.6179, 0.8321, 0.9428, 0.1067, 0.5045, 0.9324,\n         0.3326],\n        [0.6556, 0.1479, 0.9288, 0.9238, 0.1324, 0.0718, 0.6620, 0.2659, 0.7162,\n         0.7559, 0.7564, 0.2120, 0.3943, 0.9497, 0.7520, 0.8455, 0.4444, 0.4708,\n         0.8371, 0.6365, 0.3616, 0.0326, 0.1581, 0.4973, 0.6701, 0.9245, 0.8274,\n         0.3464, 0.7044, 0.5376, 0.0441, 0.5210, 0.8603, 0.7396, 0.2544, 0.3514,\n         0.5686, 0.3283, 0.7248, 0.4303, 0.9531, 0.5587, 0.8703, 0.1585, 0.9161,\n         0.9043, 0.9778, 0.4489, 0.9463, 0.8655, 0.5576, 0.1135, 0.1268, 0.3424,\n         0.1504, 0.2265, 0.1734, 0.1872, 0.3995, 0.1191, 0.0532, 0.6109, 0.1662,\n         0.6937],\n        [0.6342, 0.1922, 0.1758, 0.4625, 0.7654, 0.6509, 0.2908, 0.1546, 0.4768,\n         0.3779, 0.2490, 0.0086, 0.6170, 0.5425, 0.6953, 0.4730, 0.5834, 0.8326,\n         0.0165, 0.8236, 0.0023, 0.7479, 0.5621, 0.9894, 0.5957, 0.0857, 0.6087,\n         0.5667, 0.5478, 0.8197, 0.9228, 0.7329, 0.4434, 0.5894, 0.9860, 0.6133,\n         0.2395, 0.4718, 0.8830, 0.6361, 0.6104, 0.6630, 0.5084, 0.7604, 0.7591,\n         0.3601, 0.6888, 0.6767, 0.9178, 0.5291, 0.0591, 0.4320, 0.7875, 0.5038,\n         0.4419, 0.0319, 0.3719, 0.5843, 0.0334, 0.3525, 0.0023, 0.1205, 0.4040,\n         0.7908],\n        [0.0989, 0.8436, 0.0425, 0.6247, 0.6091, 0.4778, 0.2692, 0.4785, 0.9217,\n         0.9604, 0.6355, 0.4686, 0.9414, 0.7722, 0.8013, 0.1660, 0.6578, 0.6414,\n         0.6814, 0.6212, 0.4124, 0.7102, 0.7416, 0.7404, 0.9842, 0.6542, 0.0106,\n         0.3826, 0.5529, 0.8079, 0.9855, 0.3012, 0.2341, 0.9353, 0.6597, 0.7177,\n         0.8214, 0.1438, 0.4729, 0.6747, 0.9310, 0.4167, 0.3689, 0.8464, 0.9395,\n         0.9407, 0.8419, 0.5486, 0.1786, 0.1423, 0.9900, 0.9365, 0.3996, 0.1862,\n         0.6232, 0.7547, 0.7779, 0.4767, 0.6218, 0.9079, 0.6153, 0.1488, 0.5960,\n         0.4015]])\n<\/code><\/pre>\n\n<p>Although the <code>permute()<\/code> + <code>view()<\/code> achieve the desired output, are there other ways to perform the same operation? Is there a better way that can directly rehape without first permutating the order of the shape?<\/p>\n","answers":[{"AnswerId":"53940813","CreationDate":"2018-12-27T06:40:38.017","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Let's look \"behind the curtain\" and see why one <em>must<\/em> have both <a href=\"https:\/\/pytorch.org\/docs\/master\/tensors.html#torch.Tensor.permute\" rel=\"nofollow noreferrer\"><code>permute<\/code><\/a>\/<a href=\"https:\/\/pytorch.org\/docs\/master\/tensors.html#torch.Tensor.transpose\" rel=\"nofollow noreferrer\"><code>transpose<\/code><\/a> and <a href=\"https:\/\/pytorch.org\/docs\/master\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><code>view<\/code><\/a> in order to go from a <code>C<\/code>-<code>B<\/code>-<code>H<\/code> to <code>B<\/code>-<code>C*H<\/code>:<\/p>\n\n<p>Elements of tensors are stored as a long <em>contiguous<\/em> vector in memory. For instance, if you look at a 2-3-4 tensor it has 24 elements stored at 24 consecutive places in memory. This tensor also has a \"header\" that tells pytorch to treat these 24 values as a 2-by-3-by-4 tensor. This is done by storing not only the <code>size<\/code> of the tensor, but also \"strides\": what is the \"stride\" one need to jump in order to get to the next element along each dimension. In our example, <code>size=(2,3,4)<\/code> and <code>strides=(12, 4, 1)<\/code> (you can check this out yourself, and you can see more about it <a href=\"http:\/\/www.jessicayung.com\/numpy-arrays-memory-and-strides\/\" rel=\"nofollow noreferrer\">here<\/a>).  <\/p>\n\n<p>Now, if you only want to change the <code>size<\/code> to <code>2-(3*4)<\/code> you do not need to move any item of the tensor in memory, only to update the \"header\" of the tensor. By setting <code>size=(2, 12)<\/code> and <code>strides=(12, 1)<\/code> you are done!  <\/p>\n\n<p>Alternatively, if you want to \"transpose\" the tensor to <code>3-2-4<\/code> that's a bit more tricky, but you can still do that by manipulating the strides. Setting <code>size=(3, 2, 4)<\/code> and <code>strides=(4, 12, 1)<\/code> gives you exactly what you want <em>without<\/em> moving any of the real tensor elements in memory.<\/p>\n\n<p>However, once you manipulated the strides, you cannot trivially change the size of the tensor - because now you will need to have two different \"stride\" values for one (or more) dimensions. This is why you must call <a href=\"https:\/\/pytorch.org\/docs\/master\/tensors.html#torch.Tensor.contiguous\" rel=\"nofollow noreferrer\"><code>contiguous()<\/code><\/a> at this point.<\/p>\n\n<p><strong>Summary<\/strong><br>\nIf you want to move from shape <code>(C, B, H)<\/code> to <code>(B, C*H)<\/code> you must have <code>permute<\/code>, <code>contiguous<\/code> and <code>view<\/code> operations, otherwise you just scramble the entries of your tensor.<\/p>\n\n<p>A small example with <code>2-3-4<\/code> tensor:  <\/p>\n\n<pre><code>a = \narray([[[ 0,  1,  2,  3],\n        [ 4,  5,  6,  7],\n        [ 8,  9, 10, 11]],\n\n       [[12, 13, 14, 15],\n        [16, 17, 18, 19],\n        [20, 21, 22, 23]]])\n<\/code><\/pre>\n\n<p>If you just change the <code>view<\/code> of the tensor you get<\/p>\n\n<pre><code>a.view(3,8) \narray([[ 0,  1,  2,  3,  4,  5,  6,  7],\n       [ 8,  9, 10, 11, 12, 13, 14, 15],\n       [16, 17, 18, 19, 20, 21, 22, 23]])\n<\/code><\/pre>\n\n<p>Which is <strong>not<\/strong> what you want!<br>\nYou need to have <\/p>\n\n<pre><code>a.permute(1,0,2).contiguous().view(3, 8)\narray([[ 0,  1,  2,  3, 12, 13, 14, 15],\n       [ 4,  5,  6,  7, 16, 17, 18, 19],\n       [ 8,  9, 10, 11, 20, 21, 22, 23]])\n<\/code><\/pre>\n"},{"AnswerId":"56827457","CreationDate":"2019-06-30T19:23:26.480","ParentId":null,"OwnerUserId":"498892","Title":null,"Body":"<p>Einops allows doing such element rearrangements in one (readable) line<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from einops import rearrange\nimport torch\n\nt = torch.rand([2, 5, 32])\ny = rearrange(t, 'c b h -&gt; b (c h)')\ny.shape  # prints torch.Size([5, 64])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53479523,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><computer-vision><pytorch>","CreationDate":"2018-11-26T10:50:45.213","AcceptedAnswerId":null,"OwnerUserId":5321842.0,"Title":"Increasingly large, positive WGAN-GP loss","Body":"<p>I'm investigating the use of a Wasserstein GAN with gradient penalty in PyTorch, but consistently get large, positive generator losses that increase over epochs.\nI'm heavily borrowing from <a href=\"https:\/\/github.com\/caogang\/wgan-gp\/blob\/master\/gan_cifar10.py\" rel=\"nofollow noreferrer\">Caogang's implementation<\/a>, but am using the discriminator and generator losses used in <a href=\"https:\/\/github.com\/znxlwm\/pytorch-generative-model-collections\/blob\/master\/WGAN_GP.py\" rel=\"nofollow noreferrer\">this implementation<\/a> because I get <code>Invalid gradient at index 0 - expected shape[] but got [1]<\/code> if I try to call <code>.backward()<\/code> with the <code>one<\/code> and <code>mone<\/code> args used in the Caogang implementation.<\/p>\n\n<p>I'm training on a augmented WikiArt dataset (>400k 64x64 images) and CIFAR-10, and have gotten a normal WGAN (with weight clipping to work) [i.e. it produces passable images after 25 epochs], despite the fact that the D and G losses both hover around 3 [I calculate them using <code>torch.mean(D_real)<\/code> etc.] for all epochs. However, in the WGAN-GP version, the generator loss increases dramatically on both the WikiArt and CIFAR-10 datasets, and completely fails to generate anything other than noise on WikiArt.<\/p>\n\n<p>Here's an example of the loss after 25 epochs on CIFAR-10:\n<a href=\"https:\/\/i.stack.imgur.com\/AS05S.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AS05S.png\" alt=\"WGAN-GP loss\"><\/a><\/p>\n\n<p>I don't use any tricks like one-sided label smoothing, and I train with the default learning rate of 0.001, the Adam optimizer and I train the discriminator 5 times for every generator update. Why does this crazy loss behaviour happen, and why does the normal weight-clipping WGAN still 'work' on WikiArt but WGANGP completely fail?<\/p>\n\n<p>This happens irrespective of the structure, whether both G and D are DCGANs or when using <a href=\"https:\/\/arxiv.org\/pdf\/1706.07068.pdf\" rel=\"nofollow noreferrer\">this modified DCGAN, the Creative Adversarial Network<\/a>, which requires that D be able to classify images and G generate ambiguous images.<\/p>\n\n<p>Below is the relevant part of my current <code>train<\/code>method:<\/p>\n\n<pre><code>self.generator = Can64Generator(self.z_noise, self.channels, self.num_gen_filters).to(self.device)\nself.discriminator =WCan64Discriminator(self.channels,self.y_dim, self.num_disc_filters).to(self.device)\nstyle_criterion = nn.CrossEntropyLoss()\n\nself.disc_optimizer = optim.Adam(self.discriminator.parameters(), lr=self.lr, betas=(self.beta1, 0.9))\nself.gen_optimizer = optim.Adam(self.generator.parameters(), lr=self.lr, betas=(self.beta1, 0.9))\n\n\nwhile i &lt; len(dataloader):\n            j = 0\n            disc_loss_epoch = []\n            gen_loss_epoch = []\n            if self.type == \"can\":\n                disc_class_loss_epoch = []\n                gen_class_loss_epoch = []\n\n            if self.gradient_penalty == False:\n                # critic training methodology in official WGAN implementation\n                if gen_iterations &lt; 25 or (gen_iterations % 500 == 0):\n                    disc_iters = 100\n            else:\n                disc_iters = self.disc_iterations\n\n            while j &lt; disc_iters and (i &lt; len(dataloader)):\n                # if using wgan with weight clipping\n                if self.gradient_penalty == False:\n                    # Train Discriminator\n                    for param in self.discriminator.parameters():\n                        param.data.clamp_(self.lower_clamp,self.upper_clamp)\n\n\n                for param in self.discriminator.parameters():\n                    param.requires_grad_(True)\n\n                j+=1\n                i+=1\n                data = data_iterator.next()\n                self.discriminator.zero_grad()\n                real_images, image_labels = data\n                # image labels are the the image's classes (e.g. Impressionism)\n                real_images = real_images.to(self.device) \n                batch_size = real_images.size(0)\n                real_image_labels = torch.LongTensor(batch_size).to(self.device)\n                real_image_labels.copy_(image_labels)\n\n                labels = torch.full((batch_size,),real_label,device=self.device)\n\n                if self.type == 'can':\n                    predicted_output_real, predicted_styles_real = self.discriminator(real_images.detach())\n                    predicted_styles_real = predicted_styles_real.to(self.device)\n                    disc_class_loss = style_criterion(predicted_styles_real,real_image_labels)\n                    disc_class_loss.backward(retain_graph=True)\n\n                else:\n                    predicted_output_real = self.discriminator(real_images.detach())\n\n                disc_loss_real = -torch.mean(predicted_output_real)\n\n\n                # fake\n\n                noise = torch.randn(batch_size,self.z_noise,1,1,device=self.device)\n                with torch.no_grad():\n                    noise_g = noise.detach()\n                fake_images = self.generator(noise_g)\n                labels.fill_(fake_label)\n\n                if self.type == 'can':\n                    predicted_output_fake, predicted_styles_fake = self.discriminator(fake_images)\n\n                else:\n                    predicted_output_fake = self.discriminator(fake_images)\n\n\n\n                disc_gen_z_1 = predicted_output_fake.mean().item()\n\n                disc_loss_fake = torch.mean(predicted_output_fake)\n\n\n                #via https:\/\/github.com\/znxlwm\/pytorch-generative-model-collections\/blob\/master\/WGAN_GP.py\n                if self.gradient_penalty:\n                    # gradient penalty\n                    alpha = torch.rand((real_images.size()[0], 1, 1, 1)).to(self.device) \n                    x_hat = alpha * real_images.data + (1 - alpha) * fake_images.data\n                    x_hat.requires_grad_(True)\n                    if self.type == 'can':\n                        pred_hat, _ = self.discriminator(x_hat)\n                    else:\n                        pred_hat = self.discriminator(x_hat)\n                    gradients = grad(outputs=pred_hat, inputs=x_hat, grad_outputs=torch.ones(pred_hat.size()).to(self.device),\n                                    create_graph=True, retain_graph=True, only_inputs=True)[0]\n\n                    gradient_penalty = lambda_ * ((gradients.view(gradients.size()[0], -1).norm(2, 1) - 1) ** 2).mean()\n                    disc_loss = disc_loss_fake + disc_loss_real + gradient_penalty\n                else:\n                    disc_loss  =  disc_loss_fake  + disc_loss_real\n\n\n                if self.type == 'can':\n                    disc_loss += disc_class_loss.mean()\n\n                disc_x = disc_loss.mean().item()\n                disc_loss.backward(retain_graph=True)\n                self.disc_optimizer.step()\n\n\n\n            # train generator\n            for param in self.discriminator.parameters():\n                param.requires_grad_(False)\n\n            self.generator.zero_grad()\n            labels.fill_(real_label)\n\n            if self.type == 'can':\n                predicted_output_fake, predicted_styles_fake = self.discriminator(fake_images)\n                predicted_styles_fake = predicted_styles_fake.to(self.device)\n\n            else:\n                predicted_output_fake = self.discriminator(fake_images)\n\n            gen_loss = -torch.mean(predicted_output_fake)\n            disc_gen_z_2 = gen_loss.mean().item()\n\n            if self.type == 'can':\n                fake_batch_labels = 1.0\/self.y_dim * torch.ones_like(predicted_styles_fake)\n                fake_batch_labels = torch.mean(fake_batch_labels,1).long().to(self.device)\n                gen_class_loss = style_criterion(predicted_styles_fake,fake_batch_labels)\n                gen_class_loss.backward(retain_graph=True)\n                gen_loss += gen_class_loss.mean()\n\n            gen_loss.backward()\n            gen_iterations += 1\n<\/code><\/pre>\n\n<p>This is the code for the (DCGAN) generator:<\/p>\n\n<pre><code>class Can64Generator(nn.Module):\ndef __init__(self, z_noise, channels, num_gen_filters):\n    super(Can64Generator,self).__init__()\n    self.ngpu = 1\n    self.main = nn.Sequential(\n    nn.ConvTranspose2d(z_noise, num_gen_filters * 16, 4, 1, 0, bias=False),\n    nn.BatchNorm2d(num_gen_filters * 16),\n    nn.ReLU(True),\n    nn.ConvTranspose2d(num_gen_filters * 16, num_gen_filters * 4, 4, 2, 1, bias=False),\n    nn.BatchNorm2d(num_gen_filters * 4),\n    nn.ReLU(True),\n    nn.ConvTranspose2d(num_gen_filters * 4, num_gen_filters * 2, 4, 2, 1, bias=False),\n    nn.BatchNorm2d(num_gen_filters * 2),\n    nn.ReLU(True),\n    nn.ConvTranspose2d(num_gen_filters * 2, num_gen_filters, 4, 2, 1, bias=False),\n    nn.BatchNorm2d(num_gen_filters),\n    nn.ReLU(True),\n    nn.ConvTranspose2d(num_gen_filters, 3, 4, 2, 1, bias=False),\n    nn.Tanh()\n    )\ndef forward(self, inp):\n    output = self.main(inp)\n    return output\n<\/code><\/pre>\n\n<p>And this is the (current) CAN discriminator, which has extra layers for \nstyle (image class) classification):<\/p>\n\n<pre><code>class Can64Discriminator(nn.Module):\n\ndef __init__(self, channels,y_dim, num_disc_filters):\n        super(Can64Discriminator, self).__init__()\n        self.ngpu = 1\n        self.conv = nn.Sequential(\n                nn.Conv2d(channels, num_disc_filters \/\/ 2, 4, 2, 1, bias=False),\n                nn.LeakyReLU(0.2, inplace=True),\n\n                nn.Conv2d(num_disc_filters \/\/ 2, num_disc_filters, 4, 2, 1, bias=False),\n                nn.BatchNorm2d(num_disc_filters),\n                nn.LeakyReLU(0.2, inplace=True),\n\n                nn.Conv2d(num_disc_filters, num_disc_filters * 2, 4, 2, 1, bias=False),\n                nn.BatchNorm2d(num_disc_filters * 2),\n                nn.LeakyReLU(0.2, inplace=True),\n\n                nn.Conv2d(num_disc_filters * 2, num_disc_filters * 4, 4, 2, 1, bias=False),\n                nn.BatchNorm2d(num_disc_filters * 4),\n                nn.LeakyReLU(0.2, inplace=True),\n\n                nn.Conv2d(num_disc_filters * 4, num_disc_filters * 8, 4, 1, 0, bias=False),\n                nn.BatchNorm2d(num_disc_filters * 8),\n                nn.LeakyReLU(0.2, inplace=True),\n\n            )\n        # was this\n        #self.final_conv = nn.Conv2d(num_disc_filters * 8, num_disc_filters * 8, 4, 2, 1, bias=False)\n\n        self.real_fake_head = nn.Linear(num_disc_filters * 8, 1)\n\n        # no bn and lrelu needed\n        self.sig = nn.Sigmoid()\n        self.fc = nn.Sequential() \n        self.fc.add_module(\"linear_layer{0}\".format(num_disc_filters*16),nn.Linear(num_disc_filters*8,num_disc_filters*16))\n        self.fc.add_module(\"linear_layer{0}\".format(num_disc_filters*8),nn.Linear(num_disc_filters*16,num_disc_filters*8))\n        self.fc.add_module(\"linear_layer{0}\".format(num_disc_filters),nn.Linear(num_disc_filters*8,y_dim))\n        self.fc.add_module('softmax',nn.Softmax(dim=1))\n\ndef forward(self, inp):\n    x = self.conv(inp)\n    x = x.view(x.size(0),-1) \n    real_out = self.sig(self.real_fake_head(x))\n    real_out = real_out.view(-1,1).squeeze(1)\n    style = self.fc(x) \n    #style = torch.mean(style,1) # CrossEntropyLoss requires input be (N,C)\n    return real_out,style\n<\/code><\/pre>\n\n<p>The only differences between the WGANGP version and the WGAN version of my GAN is the WGAN version uses <code>RMSprop<\/code> with <code>lr=0.00005<\/code> and clips the weights of the discriminator, as per the WGAN paper.<\/p>\n\n<p>What could be causing this? I'd like to make as minimal change as possible, as I want to compare loss functions alone. The same problem is encountered even when using an unmodified DCGAN discriminator on CIFAR-10. Am I encountering this perhaps because I am only training currently for 25 epochs, or is there another reason? Interestingly, my GAN also completely fails to generate anything other than noise when using LSGAN (<code>nn.MSELoss()<\/code>).<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"53485246","CreationDate":"2018-11-26T16:24:36.027","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Batch Normalization in the discriminator breaks Wasserstein GANs with gradient penalty. The authors themselves advocate the usage of layer normalization instead, but this is clearly written in bold in their paper (<a href=\"https:\/\/papers.nips.cc\/paper\/7159-improved-training-of-wasserstein-gans.pdf\" rel=\"noreferrer\">https:\/\/papers.nips.cc\/paper\/7159-improved-training-of-wasserstein-gans.pdf<\/a>). It is hard to say if there are other bugs in your code, but I urge you to thoroughly read the DCGAN and the Wasserstein GAN paper and really take notes on the hyperparameters. Getting them wrong really destroys the performance of the GAN and doing a hyperparameter search gets expensive quite quickly.<\/p>\n\n<p>By the way transposed convolutions produce stairway artifacts in your output images. Use image resizing instead. For an indepth explanation of that phenomenon I can recommend the following resource (<a href=\"https:\/\/distill.pub\/2016\/deconv-checkerboard\/\" rel=\"noreferrer\">https:\/\/distill.pub\/2016\/deconv-checkerboard\/<\/a>).<\/p>\n"}]}
+{"QuestionId":53479600,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2018-11-26T10:55:45.743","AcceptedAnswerId":53485304.0,"OwnerUserId":6933148.0,"Title":"How to make the enabling of a given layer, in a keras model, trainable?","Body":"<p>I have the following siamese model:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DpaDd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DpaDd.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I would like to make the enabling\/disabling of layers <code>a-L1<\/code> and <code>b-L1<\/code> trainable. ie: <code>a-L1<\/code> and\/or <code>b-L1<\/code> should be transparent (not used or disabled) for the current input if necessary. So, the model after training, will learn when it should enable\/disable one or both of the layers <code>a-L1<\/code> and <code>b-L1<\/code>.<br>\nI managed to train this model with 4 cases, so I got 4 different models accordingly:<\/p>\n\n<ul>\n<li>model-1: without a-L1 and b-L1<\/li>\n<li>model-2: without a-L1<\/li>\n<li>model-3: without b-L1<\/li>\n<li>model-4: with both a-L1 and b-L1<\/li>\n<\/ul>\n\n<p>the performances of these models complement each other and I would like to combine them. Do you have some suggestions, please? <\/p>\n","answers":[{"AnswerId":"53485304","CreationDate":"2018-11-26T16:28:29.093","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>Let's consider you have trained four models and them let's call them m1, m2, m3 and m4<\/p>\n\n<p>first define the input layer which is common for all of them.<\/p>\n\n<pre><code>inputs = Input(shape=your_inputs_shape)\n\nmodel_1_output = m1(inputs)\nmodel_2_output = m2(inputs)\nmodel_3_output = m3(inputs)\nmodel_4_output = m4(inputs)\nmerged_layer = Concatenate(axis=your_concatanation_axis)([model_1_output, model_2_output, model_3_output,model_4_output)\nnew_model = Model(inputs=inputs, outputs=merged_layer)\n<\/code><\/pre>\n\n<p>I hope this will solve your problem.<\/p>\n\n<h2>EDIT:<\/h2>\n\n<p>To answer your question on comment, It is possible to combine only the layers before L2. But you have to decide which model's layers starting from L2, you are going to use(Since you are not combining layers starting from L2). Let's assume you want to use m1 model's layers after L2. In addition I want to add the weighting mechanism I've specified above in comments of the answer. <\/p>\n\n<h3>First let's define new models with common new inputs<\/h3>\n\n<pre><code>new_inputs = Input(shape=(inputs_shape))\nnew_m1 = keras.models.Model(inputs = new_inputs, outputs = m1(new_inputs))\nnew_m2 = keras.models.Model(inputs = new_inputs, outputs = m2(new_inputs))\nnew_m3 = keras.models.Model(inputs = new_inputs, outputs = m3(new_inputs))\nnew_m4 = keras.models.Model(inputs = new_inputs, outputs = m4(new_inputs))\n<\/code><\/pre>\n\n<h3>Now get the L2 layer for all models<\/h3>\n\n<pre><code>model1_l2 = new_m1.layers[1].get_layer(\"L2\").output\nmodel2_l2 = new_m2.layers[1].get_layer(\"L2\").output\nmodel3_l2 = new_m3.layers[1].get_layer(\"L2\").output\nmodel4_l2 = new_m4.layers[1].get_layer(\"L2\").output\n<\/code><\/pre>\n\n<h3>weighted merge<\/h3>\n\n<pre><code>merged = Concatenate(axis=your_concatanation_axis)([model1_l2, model2_l2, model3_l2,model4_l2])\nmerged_layer_shape = merged.get_shape().as_list()\n\n# specify number of channels you want the output to have after merging\n\ndesired_output_channels = 32\n\nnew_trainable_weights =  keras.backend.random_normal_variable(shape=(merged_layer_shape[-1], desired_output_channels),mean=0,scale=1)\nweighted_output = keras.backend.dot(merged,new_trainable_weights)\n<\/code><\/pre>\n\n<h3>now connect the layer of model1(m1) next to L2 with this new weighted_output<\/h3>\n\n<pre><code># I'm using some protected properties of layer. But it is not recommended way to do it.\n# get the index of l2 layer in new_m1\nfor i in range(len(new_m1.layers[1].layers)):\n    if new_m1.layers[1].layers[i].name==\"L2\":\n        index = i\nx = weighted_output\nfor i in range(index+1, len(new_m1.layers[1].layers)):\n    x = new_m1.layers[1].layers[i](x)\n\nnew_model = keras.models.Model(inputs=new_inputs, outputs=x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53479623,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow><keras>","CreationDate":"2018-11-26T10:56:44.380","AcceptedAnswerId":53482542.0,"OwnerUserId":8042888.0,"Title":"Keras model.predict, InvalidArgumentError, ConcatOp : Dimensions of inputs should match","Body":"<p>I am using the unit standard network and I can easily train it with keras. But when I want to use model.predict I get the following error massage, InvalidArgument. I have no idea why this is happening.\nI have my input shape when training network as [H,W,3],and the same shape when doing the model.predict<\/p>\n\n<p>Can anyone help?<\/p>\n\n<p>I also put my unit network and the way I train.<\/p>\n\n<pre><code>def UNet(n_input_channels, n_output_channels):\nfrom keras.layers import Input, Dropout, UpSampling2D, MaxPooling2D, BatchNormalization, Conv2D, Concatenate\nfrom keras.models import Model\n\ninputs = Input((None, None, n_input_channels))\nconv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)\nconv1 = BatchNormalization()(conv1)\nconv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv1)\nconv1 = BatchNormalization()(conv1)\npool1 = MaxPooling2D(pool_size=(2, 2))(conv1)\nconv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)\nconv2 = BatchNormalization()(conv2)\nconv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv2)\nconv2 = BatchNormalization()(conv2)\npool2 = MaxPooling2D(pool_size=(2, 2))(conv2)\nconv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)\nconv3 = BatchNormalization()(conv3)\nconv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv3)\nconv3 = BatchNormalization()(conv3)\npool3 = MaxPooling2D(pool_size=(2, 2))(conv3)\nconv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)\nconv4 = BatchNormalization()(conv4)\nconv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv4)\nconv4 = BatchNormalization()(conv4)\ndrop4 = Dropout(0.5)(conv4)\npool4 = MaxPooling2D(pool_size=(2, 2))(drop4)\n\nconv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)\nconv5 = BatchNormalization()(conv5)\nconv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv5)\nconv5 = BatchNormalization()(conv5)\ndrop5 = Dropout(0.5)(conv5)\n\nup6 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(drop5))\nmerge6 = Concatenate(axis=-1)([conv4,up6])\nconv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge6)\nconv6 = BatchNormalization()(conv6)\nconv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv6)\nconv6 = BatchNormalization()(conv6)\n\nup7 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv6))\nmerge7 = Concatenate(axis=-1)([conv3,up7])\nconv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge7)\nconv7 = BatchNormalization()(conv7)\nconv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv7)\nconv7 = BatchNormalization()(conv7)\n\nup8 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv7))\nmerge8 = Concatenate(axis=-1)([conv2,up8])\nconv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge8)\nconv8 = BatchNormalization()(conv8)\nconv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv8)\nconv8 = BatchNormalization()(conv8)\n\nup9 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv8))\nmerge9 = Concatenate(axis=-1)([conv1,up9])\nconv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge9)\nconv9 = BatchNormalization()(conv9)\nconv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)\nconv9 = BatchNormalization()(conv9)\nconv9 = Conv2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)\nconv9 = BatchNormalization()(conv9)\nconv10 = Conv2D(n_output_channels, 1, activation = 'softmax')(conv9)\n\nreturn Model(inputs = inputs, outputs = conv10)\n<\/code><\/pre>\n\n<p>and do the training as:<\/p>\n\n<pre><code>model_unet = UNet(n_input_channels=3, n_output_channels=3)\nmodel_unet.compile(optimizer='adam', loss=dice_coef_loss, metrics=[dice_coef])\nhistory_unet = model_unet.fit(x_train, y_train, batch_size=batchsize, epochs=epochs, \n                    verbose=1, shuffle=True, \n                    validation_data=(x_validation, y_validation))\n<\/code><\/pre>\n\n<p>and the model.predict<\/p>\n\n<pre><code>predictions = model.predict(sample_images, batch_size=4)\n\nInvalidArgumentError\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/IPython\/core\/interactiveshell.pyc in run_cell_magic(self, magic_name, line, cell)\n   2115             magic_arg_s = self.var_expand(line, stack_depth)\n   2116             with self.builtin_trap:\n-&gt; 2117                 result = fn(magic_arg_s, cell)\n   2118             return result\n   2119 \n\n&lt;decorator-gen-60&gt; in time(self, line, cell, local_ns)\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/IPython\/core\/magic.pyc in &lt;lambda&gt;(f, *a, **k)\n    186     # but it's overkill for just that one bit of state.\n    187     def magic_deco(arg):\n--&gt; 188         call = lambda f, *a, **k: f(*a, **k)\n    189 \n    190         if callable(arg):\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/IPython\/core\/magics\/execution.pyc in time(self, line, cell, local_ns)\n   1191         else:\n   1192             st = clock2()\n-&gt; 1193             exec(code, glob, local_ns)\n   1194             end = clock2()\n   1195             out = None\n\n&lt;timed exec&gt; in &lt;module&gt;()\n\n&lt;ipython-input-39-6b97b7df800f&gt; in predict(model, num_samples, indices)\n     10     sample_masks = masks[samples_ind, :142, :92]\n     11     print samplep_images.shape\n---&gt; 12     predictions = model.predict(sample_images, batch_size=4)\n     13     toc = time.time()\n     14     print('Time per image = {:.4f} sec'.format((toc-tic) \/ num_samples))\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/keras\/engine\/training.pyc in predict(self, x, batch_size, verbose, steps)\n   1167                                             batch_size=batch_size,\n   1168                                             verbose=verbose,\n-&gt; 1169                                             steps=steps)\n   1170 \n   1171     def train_on_batch(self, x, y,\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/keras\/engine\/training_arrays.pyc in predict_loop(model, f, ins, batch_size, verbose, steps)\n    292                 ins_batch[i] = ins_batch[i].toarray()\n    293 \n--&gt; 294             batch_outs = f(ins_batch)\n    295             batch_outs = to_list(batch_outs)\n    296             if batch_index == 0:\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/keras\/backend\/tensorflow_backend.pyc in __call__(self, inputs)\n   2713                 return self._legacy_call(inputs)\n   2714 \n-&gt; 2715             return self._call(inputs)\n   2716         else:\n   2717             if py_any(is_tensor(x) for x in inputs):\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/keras\/backend\/tensorflow_backend.pyc in _call(self, inputs)\n   2673             fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)\n   2674         else:\n-&gt; 2675             fetched = self._callable_fn(*array_vals)\n   2676         return fetched[:len(self.outputs)]\n   2677 \n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.pyc in __call__(self, *args, **kwargs)\n   1380           ret = tf_session.TF_SessionRunCallable(\n   1381               self._session._session, self._handle, args, status,\n-&gt; 1382               run_metadata_ptr)\n   1383         if run_metadata:\n   1384           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n\/work\/banafsheh.beheshtipour\/virenv\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/errors_impl.pyc in __exit__(self, type_arg, value_arg, traceback_arg)\n    517             None, None,\n    518             compat.as_text(c_api.TF_Message(self.status.status)),\n--&gt; 519             c_api.TF_GetCode(self.status.status))\n    520     # Delete the underlying status object from memory otherwise it stays alive\n    521     # as there is a reference to status from this from the traceback due to\n\nInvalidArgumentError: ConcatOp : Dimensions of inputs should match: shape[0] = [1,512,17,11] vs. shape[1] = [1,512,16,10]\n     [[Node: concatenate_1\/concat = ConcatV2[N=2, T=DT_FLOAT, Tidx=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](batch_normalization_8\/cond\/Merge, conv2d_11\/Relu, concatenate_1\/concat-2-LayoutOptimizer)]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53479831","CreationDate":"2018-11-26T11:08:31.777","ParentId":null,"OwnerUserId":"4670317","Title":null,"Body":"<p>Double check your input shape you want to give to your model. <\/p>\n\n<blockquote>\n  <p>inputs = Input((None, None, n_input_channels))<\/p>\n<\/blockquote>\n\n<p>Input shape while training model and prediction should be same.(H, W , channel)<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: ConcatOp : Dimensions of inputs should match:\n  shape[0] = [1,512,17,11] vs. shape[1] = [1,512,16,10]<\/p>\n<\/blockquote>\n\n<p><strong>Updated:<\/strong>\nResize Input size to (32,32,3)<\/p>\n\n<pre><code>img = cv2.resize(img, (32, 32)) \n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> Change variable according to your need.<\/p>\n"},{"AnswerId":"53482542","CreationDate":"2018-11-26T13:49:40.223","ParentId":null,"OwnerUserId":"8042888","Title":null,"Body":"<p>Thanks to @Saugat Bhattarai to help me about finding the answer. I found that to do the prediction, your image size needs to be any power of 2. So the H and W can be 8,32,64,128,.... There is no need to have the exact same image size for the training and prediction as long as the size is a power of two. So I tried my case with training data images with size (32,32,3) and prediction with (128,128,3) and it perfectly worked. \nHope this helps other people.<\/p>\n"}]}
+{"QuestionId":53479930,"AnswerCount":1,"Tags":"<python-3.x><windows><tensorflow>","CreationDate":"2018-11-26T11:14:57.643","AcceptedAnswerId":null,"OwnerUserId":10705841.0,"Title":"tensorflow 1.8 with python 3.6 in windows64","Body":"<p>I am trying to install tensorflow 1.8 cpu version with python 3.6 on my windows 64 bit os. Anyone konw that  whether the tensorflow 1.8 support python 3.6 in windows 64 bit system\uff1f The table here <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">TF install guide<\/a> didn't mention it. I had met the same problems bofore and it seems that the tensorflow version is too new for python 3.x to support on windows.<\/p>\n\n<p>here are the issues:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 14, in swig_import_helper\n    return importlib.import_module(mname)\n  File \"D:\\python36\\lib\\importlib\\__init__.py\", line 126, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 994, in _gcd_import\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 971, in _find_and_load\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 955, in _find_and_load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 658, in _load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 571, in module_from_spec\n  File \"&lt;frozen importlib._bootstrap_external&gt;\", line 922, in create_module\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 219, in _call_with_frames_removed\nImportError: DLL load failed: \u52a8\u6001\u94fe\u63a5\u5e93(DLL)\u521d\u59cb\u5316\u4f8b\u7a0b\u5931\u8d25\u3002\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 17, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 16, in swig_import_helper\n    return importlib.import_module('_pywrap_tensorflow_internal')\n  File \"D:\\python36\\lib\\importlib\\__init__.py\", line 126, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\nModuleNotFoundError: No module named '_pywrap_tensorflow_internal'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 74, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 14, in swig_import_helper\n    return importlib.import_module(mname)\n  File \"D:\\python36\\lib\\importlib\\__init__.py\", line 126, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 994, in _gcd_import\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 971, in _find_and_load\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 955, in _find_and_load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 658, in _load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 571, in module_from_spec\n  File \"&lt;frozen importlib._bootstrap_external&gt;\", line 922, in create_module\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 219, in _call_with_frames_removed\nImportError: DLL load failed: \u52a8\u6001\u94fe\u63a5\u5e93(DLL)\u521d\u59cb\u5316\u4f8b\u7a0b\u5931\u8d25\u3002\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 17, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"D:\\python36\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 16, in swig_import_helper\n    return importlib.import_module('_pywrap_tensorflow_internal')\n  File \"D:\\python36\\lib\\importlib\\__init__.py\", line 126, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\nModuleNotFoundError: No module named '_pywrap_tensorflow_internal'\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\n\nfor some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help. \n<\/code><\/pre>\n","answers":[{"AnswerId":"53480414","CreationDate":"2018-11-26T11:45:44.487","ParentId":null,"OwnerUserId":"10705841","Title":null,"Body":"<p>Tragedy. It turns out that my computer with a <i>Intel(R) Celeron(R) cpu N2940<\/i> and it doesn't support AVX instructions. TF version >= 1.6 will need AVX. That's it.<\/p>\n"}]}
+{"QuestionId":53480034,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-hub>","CreationDate":"2018-11-26T11:21:33.020","AcceptedAnswerId":null,"OwnerUserId":10705955.0,"Title":"what's the definition of tensorflow hub imagenet classification module output classes","Body":"<p>such as <a href=\"https:\/\/tfhub.dev\/google\/imagenet\/inception_resnet_v2\/classification\/1\" rel=\"nofollow noreferrer\">inception_resnet_v2<\/a> module,in this page,it says there are 1001 classes see above ,but I can't find this definition of class indices\n<a href=\"https:\/\/i.stack.imgur.com\/nbeGz.png\" rel=\"nofollow noreferrer\">1001 classes<\/a><\/p>\n","answers":[{"AnswerId":"53496917","CreationDate":"2018-11-27T09:52:18.837","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>Please see <a href=\"https:\/\/github.com\/tensorflow\/hub\/issues\/133\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/hub\/issues\/133<\/a> for a quick answer (and the request to document that more properly)<\/p>\n"}]}
+{"QuestionId":53480116,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2018-11-26T11:26:45.870","AcceptedAnswerId":null,"OwnerUserId":6733064.0,"Title":"`get_variable()` doesn't recognize existing variables for tf.estimator","Body":"<p>This question has been asked <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1325\" rel=\"nofollow noreferrer\">here<\/a>, difference is my problem is focused on <code>Estimator<\/code>.<\/p>\n\n<p>Some context: We have trained a model using estimator and get some variable defined within Estimator <code>input_fn<\/code>, this function preprocesses data to batches. Now, we are moving to prediction. During the prediction, we use the same <code>input_fn<\/code> to read in and process the data. <strong>But got error saying variable (word_embeddings) does not exist<\/strong> (variables exist in the chkp graph), here's the relevant bit of code in <code>input_fn<\/code>:<\/p>\n\n<pre><code>with tf.variable_scope('vocabulary', reuse=tf.AUTO_REUSE):\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        word_to_index, word_to_vec = load_embedding(graph_params[\"word_to_vec\"])\n        word_embeddings = tf.get_variable(initializer=tf.constant(word_to_vec, dtype=tf.float32),\n                                          trainable=False,\n                                          name=\"word_to_vec\",\n                                          dtype=tf.float32)\n    else:\n        word_embeddings = tf.get_variable(\"word_to_vec\", dtype=tf.float32)\n<\/code><\/pre>\n\n<p>basically, when it's in prediction mode, <code>else<\/code> is invoked to load up variables in checkpoint. Failure of recognizing this variable indicates a) inappropriate usage of scope; b) graph is not restored. I don't think scope matters that much here as long as <code>reuse<\/code> is set properly.<\/p>\n\n<p>I suspect that is because the graph is not yet restored at <code>input_fn<\/code> phase. Usually, the graph is restored by calling <code>saver.restore(sess, \"\/tmp\/model.ckpt\")<\/code> <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">reference<\/a>. Investigation of estimator <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/estimator\/estimator.py#L508\" rel=\"nofollow noreferrer\">source code<\/a> doesn't get me anything relating to restore, the best shot is MonitoredSession, a wrapper of training. It's already been stretch so much from the original problem, not confident if I'm on the right path, I'm looking for help here if anyone has any insights.<\/p>\n\n<p>One line summary of my question: How does graph get restored within <code>tf.estimator<\/code>, via <code>input_fn<\/code> or <code>model_fn<\/code>?<\/p>\n","answers":[{"AnswerId":"53733868","CreationDate":"2018-12-11T23:23:39.153","ParentId":null,"OwnerUserId":"8593338","Title":null,"Body":"<p>Hi I think that you error comes simply because you didn't specify the shape in the tf.get_variable (at predict) , it seems that you need to specify the shape even if the variable is going to be restored.<\/p>\n\n<p>I've made the following test with a simple linear regressor estimator that simply needs to predict x + 5 <\/p>\n\n<pre><code>def input_fn(mode):\n    def _input_fn():\n        with tf.variable_scope('all_input_fn', reuse=tf.AUTO_REUSE):\n            if mode == tf.estimator.ModeKeys.TRAIN:\n                var_to_follow = tf.get_variable('var_to_follow', initializer=tf.constant(20))\n                x_data = np.random.randn(1000)\n                labels = x_data + 5\n                return {'x':x_data}, labels\n            elif mode == tf.estimator.ModeKeys.PREDICT:\n                var_to_follow = tf.get_variable(\"var_to_follow\", dtype=tf.int32, shape=[])\n                return {'x':[0,10,100,var_to_follow]}\n    return _input_fn\n\nfeatcols = [tf.feature_column.numeric_column('x')]\nmodel = tf.estimator.LinearRegressor(featcols, '.\/outdir')\n<\/code><\/pre>\n\n<p>This code works perfectly fine, the value of the const is 20 and also for fun use it in my test set to confirm :p<\/p>\n\n<p>However if you remove the shape=[] , it breaks, you can also give another initializer such as tf.constant(500) and everything will work and 20 will be used.<\/p>\n\n<p>By running<\/p>\n\n<pre><code>model.train(input_fn(tf.estimator.ModeKeys.TRAIN), max_steps=10000)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>preds = model.predict(input_fn(tf.estimator.ModeKeys.PREDICT))\nprint(next(preds))\n<\/code><\/pre>\n\n<p>You can visualize the graph and you'll see that  a) the scoping is normal and b) the graph is restored.<\/p>\n\n<p>Hope this will help you.<\/p>\n"}]}
+{"QuestionId":53480391,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2018-11-26T11:44:14.700","AcceptedAnswerId":null,"OwnerUserId":7985466.0,"Title":"Tensorflow: How can I design custom cost function?","Body":"<p>I am a student studying Object Detection.\nI implemented a simple object finding model with FCN. The accuracy is 95% but the recall is 0. In short, the model has determined that there are no objects for all areas. I used softmax_cross_entropy as the error function. However, if the model judges that there is no object for the part where the object exists, I want to make a large error. I think I have to design the loss function myself, what should I do?<\/p>\n\n<pre><code>def model(X, P):\n    x = tf.cast(X, tf.float32)\n    x = x \/ 255.0\n\n    net = slim.conv2d(x, 32, kernel_size =(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 128 160\n    net = slim.conv2d(net, 64, kernel_size =(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 64 80\n    net = slim.conv2d(net, 128, kernel_size =(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 32 40\n    net = slim.conv2d(net, 256, kernel_size =(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 16 20\n    net = slim.conv2d(net, 512, kernel_size=(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 8 10\n    net = tf.nn.dropout(net, keep_prob=P)\n    net = slim.conv2d(net, 2, kernel_size=(3, 3))\n    net = slim.max_pool2d(net, (2, 2)) # 4 5\n\nX = tf.placeholder(tf.uint8 ,[None, 320, 256, 3])\nY = tf.placeholder(tf.int64, [None, 20])\nP = tf.placeholder(tf.float32)\ny = tf.reshape(Y, (-1, 5, 4))\nlogits = model(X,P)\narg = tf.argmax(logits, -1)\nwith tf.name_scope('Optimizer'):\n    cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y))\n    optimizer = tf.train.AdadeltaOptimizer(0.001).minimize(cost)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TMOW2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TMOW2.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53480713","CreationDate":"2018-11-26T12:03:17.717","ParentId":null,"OwnerUserId":"497600","Title":null,"Body":"<p>You are using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/losses\/sparse_softmax_cross_entropy\" rel=\"nofollow noreferrer\"><code>sparse_softmax_cross_entropy<\/code><\/a>, this already allows you to add weights to your loss:<\/p>\n\n<pre><code>tf.losses.sparse_softmax_cross_entropy(\n    labels,\n    logits,\n    weights=1.0,\n    ...\n<\/code><\/pre>\n\n<blockquote>\n  <p>weights acts as a coefficient for the loss. If a scalar is provided,\n  then the loss is simply scaled by the given value. If weights is a\n  tensor of shape [batch_size], then the loss weights apply to each\n  corresponding sample.<\/p>\n<\/blockquote>\n\n<p>If your y is binary (1 for object, 0 for empty), you could simply multiply this with a desired factor and pass it in as weights:<\/p>\n\n<pre><code>penalty_empty = 0.5\npenalty_object = 1\n\nweights = penalty_empty + y * (penalty_object - penalty_empty)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53480889,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2018-11-26T12:12:32.520","AcceptedAnswerId":null,"OwnerUserId":10706109.0,"Title":"Using preprocessing function of ImageDataGenerator to convert color space","Body":"<p>Ive encountered a problem that using the preprocessing function of ImageDataGenerator to convert the color space of my image. I tried to convert RGB to CMYK and feed it into the CNN. But the preprocessing function can only return the same shape as the input. For example, the image in RGB mode has a shape(224,224,3), after the convertion, the shape should be converted into (224,224,4). And this seems to be a problem to using the preprocessing function.\nCan anyone help?<\/p>\n\n<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport matplotlib.image as img\nfrom keras.models import Sequential\nfrom keras.layers import Flatten, Dense, Input, Lambda\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D, ZeroPadding2D\nfrom keras.utils import to_categorical, multi_gpu_model\nfrom keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img\nfrom keras import optimizers\nfrom keras.applications.vgg16 import VGG16, preprocess_input\nfrom keras.models import Model\nimport os\nimport keras.backend.tensorflow_backend as KTF\nimport numpy as np\nimport cv2\nfrom PIL import Image\nimport pickle\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\n\n\n#cmyk convetion\ndef rgb_to_cmyk(img_arr):\n    aaa = []\n\n    img_arr_min = (255 - img_arr)\n    K = np.min(img_arr_min, axis=2)\n\n    C_init = (img_arr_min[:, :, 0])\n    M_init = (img_arr_min[:, :, 1])\n    Y_init = (img_arr_min[:, :, 2])\n\n    C = C_init - K\n    M = M_init - K\n    Y = Y_init - K\n\n    aaa.append(C)\n    aaa.append(M)\n    aaa.append(Y)\n    aaa.append(K)\n    new_arr = np.array(aaa).transpose(1,2,0)\n    return new_arr\n\n\n#customize function used for color convetion\ndef myFunc(image):\n    image = np.array(image)\n    converted_img = rgb_to_cmyk(image)\n    converted_img \/= 255\n    return converted_img\n\n\n\n\n#parameters\nbatch_size = 32\nn_epoch = 150\nlr_rate = 0.001\n\n\n\n#data augementation\ntrain_datagen = ImageDataGenerator(rescale = None,\n                                   rotation_range = 30,\n                                   width_shift_range = 0.2,\n                                   height_shift_range = 0.2,\n                                   shear_range = 0.2,\n                                   zoom_range = 0.2,\n                                   horizontal_flip = True,\n                                   preprocessing_function=myFunc)\n\ntest_datagen = ImageDataGenerator(rescale = None,\n                                  preprocessing_function=myFunc)\n\n\n\nwith tf.device('\/cpu:0'):\n    # load data\n    train_generator = train_datagen.flow_from_directory('train_dir',\n                                                        target_size=(224,224),\n                                                        color_mode='rgb',\n                                                        batch_size=batch_size,\n                                                        class_mode='categorical')\n\n    test_generator = test_datagen.flow_from_directory('test_dir',\n                                                      target_size=(224,224),\n                                                      color_mode='rgb',\n                                                      batch_size=batch_size,\n                                                      class_mode='categorical')\n\n\n\n\n\n\n\n\n#initialize the model\nmodel = Sequential()\n\n# Conv Block 1\nmodel.add(Conv2D(64, (3, 3), input_shape=(224,224,1), activation='relu', padding='same'))\nmodel.add(Conv2D(64, (3, 3), activation='relu', padding='same'))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))\n\n# Conv Block 2\nmodel.add(Conv2D(128, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(128, (3, 3), activation='relu', padding='same'))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))\n\n# Conv Block 3\nmodel.add(Conv2D(256, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(256, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(256, (3, 3), activation='relu', padding='same'))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))\n\n# Conv Block 4\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))\n\n# Conv Block 5\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(Conv2D(512, (3, 3), activation='relu', padding='same'))\nmodel.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))\n\n# FC layers\nmodel.add(Flatten())\nmodel.add(Dense(4096, activation='relu'))\nmodel.add(Dense(4096, activation='relu'))\nmodel.add(Dense(101, activation='softmax'))\n\n\n\n\n\n\nnetwork = model\n\nsgd = optimizers.SGD(lr=lr_rate, decay=1e-6, momentum=0.9, nesterov=True)\nnetwork.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\nnetwork.summary()\n\n\n\n# train\nhistory_callback = network.fit_generator(train_generator,\n                      steps_per_epoch=2000,\n                      epochs=n_epoch,\n                      validation_data=test_generator,\n                      validation_steps=800, verbose=2)\n                      #use_multiprocessing=False)\n\n\n\n\nall_history = history_callback.history\npickle.dump(all_history, open('save_dir\/results', 'wb'), 0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53481189","CreationDate":"2018-11-26T12:31:25.067","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>Keras only works with rgb or grayscale images.\nFor example a snippet from the <code>array_to_img()<\/code> method:<\/p>\n\n<pre><code>if x.shape[2] == 4:\n    # RGBA\n    return pil_image.fromarray(x.astype('uint8'), 'RGBA')\nelif x.shape[2] == 3:\n    # RGB\n    return pil_image.fromarray(x.astype('uint8'), 'RGB')\nelif x.shape[2] == 1:\n    # grayscale\n<\/code><\/pre>\n\n<p>Thus you have to convert CMYK images to RGB first.<\/p>\n"}]}
+{"QuestionId":53481153,"AnswerCount":1,"Tags":"<tensorflow><rsync><tensorboard>","CreationDate":"2018-11-26T12:29:25.347","AcceptedAnswerId":53481465.0,"OwnerUserId":10196913.0,"Title":"Why does Tensorboard not refresh when used with rsync?","Body":"<p>I am running a tensorflow experiment on a remote machine continuously writing to the same <code>events.out.tfevents.xxx<\/code> file. I would expect tensorboard to refresh automatically every minute or so displaying the new logs. This does work when using sshfs to mount the remote machine on my laptop and using the mounted directory to run tensorboard on.<\/p>\n\n<p>However, when using rsync to copy the files over and run tensorboard on the local files, the tensorboard never refreshes, I have to restart it in order to get the updates.<\/p>\n\n<p>This is my rsync command:<\/p>\n\n<pre><code>rsync -aP --del -e ssh server_name:folder_on_server local_folder --exclude='*checkpoints*' --exclude='*.json' --exclude='*.DS_Store'\n<\/code><\/pre>\n\n<p>Any help would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"53481465","CreationDate":"2018-11-26T12:48:31.657","ParentId":null,"OwnerUserId":"7370153","Title":null,"Body":"<p>It's a known issue with the Tensorboard, see this <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/349\" rel=\"nofollow noreferrer\">issue<\/a> on github. <\/p>\n\n<p>Here's an quote from the issue (emphasis is mine) : <\/p>\n\n<blockquote>\n  <p>It looks like when the tensorboard reads an event file from local directory - it <strong>will not notice that the event file was deleted and recreated<\/strong> (which is quite valid case when you are using [...] rsync to sync the data)<\/p>\n<\/blockquote>\n\n<p>One workaround is to use <code>--inplace<\/code> as an option in your rsync command. <\/p>\n"}]}
+{"QuestionId":53481163,"AnswerCount":1,"Tags":"<tensorflow><profiling>","CreationDate":"2018-11-26T12:30:01.847","AcceptedAnswerId":null,"OwnerUserId":2594778.0,"Title":"TensorFlow: how to start the Profiler","Body":"<p>I want to profile my model. This is a tutorial on how to do it: <a href=\"https:\/\/towardsdatascience.com\/howto-profile-tensorflow-1a49fb18073d\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/howto-profile-tensorflow-1a49fb18073d<\/a>. But I would like to use the TensorFlow profiler, as shown in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/profiler\/README.md#quick-start\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/profiler\/README.md#quick-start<\/a>. According to this post, the following code should start the profiler:<\/p>\n\n<pre><code># When using high-level API, session is usually hidden.\n#\n# Under the default ProfileContext, run a few hundred steps.\n# The ProfileContext will sample some steps and dump the profiles\n# to files. Users can then use command line tool or Web UI for\n# interactive profiling.\nwith tf.contrib.tfprof.ProfileContext('\/tmp\/train_dir') as pctx:\n  # High level API, such as slim, Estimator, etc.\n  train_loop()\n\nbazel-bin\/tensorflow\/core\/profiler\/profiler \\\n--profile_path=\/tmp\/train_dir\/profile_xx\ntfprof&gt; op -select micros,bytes,occurrence -order_by micros\n\n# To be open sourced...\nbazel-bin\/tensorflow\/python\/profiler\/profiler_ui \\\n--profile_path=\/tmp\/profiles\/profile_1\n<\/code><\/pre>\n\n<p>I generated the file profile_100 and located the directory profiler. So this is what I typed in my terminal:<\/p>\n\n<pre><code>bazel-\/Users\/mencia\/anaconda\/envs\/tensorflow_py36\/lib\/python3.6\/site-packages\/tensorflow\/profiler \\ \n--profile_path=~\/tmp\/train_dir\/profile_100\n<\/code><\/pre>\n\n<p>This raised the following error:<\/p>\n\n<pre><code>-bash: \nbazel-\/Users\/mencia\/anaconda\/envs\/tensorflow_py36\/lib\/python3.6\/site- packages\/tensorflow\/profiler: No such file or directory\n<\/code><\/pre>\n\n<p>My directory profiler contains:<\/p>\n\n<pre><code>__init__.py\n__pycache__\n<\/code><\/pre>\n\n<p>But according to the above code, there should be <\/p>\n\n<pre><code>profiler\/profiler\n<\/code><\/pre>\n\n<p>Which I don't have.<\/p>\n\n<p>What do I do to start the Profiler?<\/p>\n","answers":[{"AnswerId":"54109067","CreationDate":"2019-01-09T11:26:10.650","ParentId":null,"OwnerUserId":"10190297","Title":null,"Body":"<p>You have to build the profiler first. Clone the TensorFlow repo (<code>git clone https:\/\/github.com\/tensorflow\/tensorflow.git<\/code>) and run <code>bazel build --config opt tensorflow\/core\/profiler:profiler<\/code> in the root directory.<\/p>\n"}]}
+{"QuestionId":53481231,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2018-11-26T12:34:25.130","AcceptedAnswerId":null,"OwnerUserId":3145671.0,"Title":"Tensorflow reads multi csv needs too much time","Body":"<p>I plan to read the feature data from multi csv files.<br>\nEach csv file has 150 columns, and the batch size is 256. <\/p>\n\n<p>The time cost for just read 1000 iteration needs roughly 12s.<br>\nI feel the time cost for that should not be that much, any one here can have some suggestion?<\/p>\n\n<pre><code>def _parse_csv_row(*vals):\n    features = tf.convert_to_tensor(vals[0:f_size * 5])\n    class_label = tf.cast(vals[f_size * 5] + tf.convert_to_tensor(1.0, tf.float64), tf.int64)\n    return features, class_label\n\n\ndef get_batch_data(name):\n    root_path=\"g:\\\\market\\\\2018-11-12\\\\feature_{}\\\\\".format(name)\n    file_queue = list(map(lambda x: \"{}{}\".format(root_path, x), fnmatch.filter(os.listdir(\"g:\\\\market\\\\2018-11-12\\\\feature_{}\\\\\".format(name)), \"*.sz_result.csv\")))\n    record_defaults = [tf.float32] * f_size * 5 + [tf.float64]\n    selected_cols = reduce(lambda x, y: x + y, [list(range(1 + x * 29, 1 + x * 29 + 9)) for x in range(0, 5)]) + [146]\n    dataset = tf.contrib.data.CsvDataset(\n        file_queue,\n        record_defaults,\n        buffer_size=1024 * 1024 * 10,\n        header=True,\n        na_value='0.0',\n        select_cols=selected_cols)\n    dataset = dataset.apply(tf.contrib.data.map_and_batch(\n        map_func=_parse_csv_row, batch_size=train_config.BATCH_SIZE))\n    dataset = dataset.prefetch(256 * 1024)\n    dataset = dataset.repeat()\n\n    dataset = dataset.shuffle(buffer_size=32)\n    iterator = dataset.make_one_shot_iterator()\n    feature_batch, label_batch = iterator.get_next()\n    return feature_batch, label_batch\n\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nwith tf.Session(config=config) as sess:\n    sess.run(tf.global_variables_initializer())\n    a, b = get_batch_data(\"train\")\n    start_time = time.time()\n    for x in range(1000):\n        v = sess.run([a,b])\n    print(time.time() - start_time)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53482953","CreationDate":"2018-11-26T14:12:14.247","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>The line <code>dataset = dataset.prefetch(256 * 1024)<\/code>. It's written after <code>map_and_batch<\/code> operation. That means that your are prefetching 256 * 1024 batches.  So, when your program tries to load 1st record, it actually first loads 256 * 1024 * 256 records. Probably your intention was to prefetch only 1024 batches. In real life it's enough to prefetch only one record.<\/p>\n\n<p>I would put <code>ds.prefetch(1)<\/code> line as the last operation on the dataset. See <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets#map_and_interleave_prefetch_shuffle\" rel=\"nofollow noreferrer\">Summary of Best Practices<\/a>.<\/p>\n\n<p>Also you allocate pretty big buffer for reading CSV file (<code>buffer_size<\/code> parameter). If your intention is to cache the whole csv file, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/data\/CsvDataset#cache\" rel=\"nofollow noreferrer\">ds.cache()<\/a> operation. Without parameters it caches contents in the memory. Put it before the <code>ds.repeat()<\/code> operation.<\/p>\n"}]}
+{"QuestionId":53481675,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2018-11-26T12:59:44.747","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"Keras Always Output a Constant Value","Body":"<p>I've been training a simple Artificial Neural Network using Keras<\/p>\n\n<pre><code>model = Sequential([\n    Dense(32, input_shape=(32,), activation = 'relu'),\n    Dense(20, activation='relu'),\n    Dense(65, input_shape=(65,), activation='softmax')\n])\n\nmodel.summary()\nmodel.compile(Adam(lr=.001), loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(train_samples, train_labels, batch_size=1000, epochs=10000,shuffle = True, verbose=2)\n<\/code><\/pre>\n\n<p>After training, I test the model, and it always outputs a constant value. What shall I do?<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/1727\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/1727<\/a><\/p>\n\n<p>The link above tells that I have to center my data to have a zero mean. I don't have any idea to do that.<\/p>\n","answers":[{"AnswerId":"53482045","CreationDate":"2018-11-26T13:21:06.237","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>If you want to center your data to have zero mean, just subtract sample mean from each samples.<\/p>\n\n<p>e.g if you have <code>features<\/code> data and want to center it to have zero mean, you can do that using <a href=\"http:\/\/lagrange.univ-lyon1.fr\/docs\/numpy\/1.11.0\/reference\/generated\/numpy.mean.html\" rel=\"nofollow noreferrer\">numpy's mean function<\/a> as follows.<\/p>\n\n<pre><code>features = features - np.mean(features)\n<\/code><\/pre>\n\n<p>You may also need to normalize using standard deviation. This can be achieved by using numpy as follows.  <\/p>\n\n<pre><code>normalized_features = (features - np.mean(features))\/ np.std(features)\n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"},{"AnswerId":"53482783","CreationDate":"2018-11-26T14:02:18.233","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You're using <code>'relu'<\/code> activations without caring for it going to all zeros. <\/p>\n\n<p>Relu has a \"zero region\" when its input is negative, and this region, naturally has no gradient, so there is no possibility of it changing over training. <\/p>\n\n<p>If all neurons in a layer go to the zero region, your model is frozen forever. <\/p>\n\n<p>One thing you can do is to replace <code>'relu'<\/code> by <code>'sigmoid'<\/code> or <code>'tanh'<\/code>.<br>\nAnother thing is to use a <code>BatchNormalization<\/code> layer before it.    <\/p>\n\n<p>The <code>BatchNormalization<\/code> does the centering for you, besides adding some speed to training and a little regularization. <\/p>\n\n\n\n<pre><code>model = Sequential([\n\n    #optional BatchNormalization(input_shape(32,))\n\n    Dense(32, input_shape=(32,)),\n    BatchNormalization(),\n    Activation('relu')\n    Dense(20),\n    BatchNormalization(),\n    Activation('relu'),\n    Dense(65, input_shape=(65,), activation='softmax')\n])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53481899,"AnswerCount":1,"Tags":"<machine-learning><keras><lstm><rnn><word-embedding>","CreationDate":"2018-11-26T13:13:06.597","AcceptedAnswerId":53482469.0,"OwnerUserId":8122842.0,"Title":"keras Bidirectional layer using 4 dimension data","Body":"<p>I'am designing keras model for classification based on article data.<\/p>\n\n<p>I have data with 4 dimension as follows <\/p>\n\n<pre><code>[batch, article_num, word_num, word embedding size]\n<\/code><\/pre>\n\n<p>and i want to feed each (word_num, word embedding) data into keras Bidirectional layer<\/p>\n\n<p>in order to get result with 3 dimension as follows.<\/p>\n\n<pre><code>[batch, article_num, bidirectional layer output size]\n<\/code><\/pre>\n\n<p>when i tried to feed 4 dimension data for testing like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inp = Input(shape=(article_num, word_num, ))\n# dims = [batch, article_num, word_num]\n\nx = Reshape((article_num * word_num, ), input_shape = (article_num, word_num))(inp)\n# dims = [batch, article_num * word_num]\n\nx = Embedding(word_num, word_embedding_size, input_length = article_num * word_num)(x)\n# dims = [batch, article_num * word_num, word_embedding_size]\n\nx = Reshape((article_num , word_num, word_embedding_size), \n             input_shape = (article_num * word_num, word_embedding_size))(x)\n# dims = [batch, article_num, word_num, word_embedding_size]\n\nx = Bidirectional(CuDNNLSTM(50, return_sequences = True), \n                  input_shape=(article_num , word_num, word_embedding_size))(x)\n<\/code><\/pre>\n\n<p>and i got the error<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer bidirectional_12: expected ndim=3, found ndim=4\n<\/code><\/pre>\n\n<p>how can i achieve this?<\/p>\n","answers":[{"AnswerId":"53482469","CreationDate":"2018-11-26T13:45:34.270","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>If you don't want it to touch the <code>article_num<\/code> dimension, you can try using a <code>TimeDistributed<\/code> wrapper. But I'm not certain that it will be compatible with bidirectional and other stuff.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inp = Input(shape=(article_num, word_num))    \n\nx = TimeDistributed(Embedding(word_num, word_embedding_size)(x))\n\n#option 1\n#x1 shape : (batch, article_num, word_num, 50)\nx1 = TimeDistributed(Bidirectional(CuDNNLSTM(50, return_sequences = True)))(x)\n\n#option 2\n#x2 shape : (batch, article_num, 50)\nx2 = TimeDistributed(Bidirectional(CuDNNLSTM(50)))(x)\n<\/code><\/pre>\n\n<p>Hints: <\/p>\n\n<ul>\n<li>Don't use <code>input_shape<\/code> everywhere, you only need it at the <code>Input<\/code> tensor.    <\/li>\n<li>You probably don't need any of the reshapes if you also use a <code>TimeDistributed<\/code> in the embedding.    <\/li>\n<li>If you don't want <code>word_num<\/code> in the final dimension, use <code>return_sequences=False<\/code>. <\/li>\n<\/ul>\n"}]}
+{"QuestionId":53482251,"AnswerCount":3,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-26T13:33:01.007","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"How to Convert Keras Prediction Output to desired Binary Value","Body":"<p>I have a network with 32 input nodes, 20 hidden nodes and 65 output nodes.\nMy network input actually is a hash code of length 32 and the output is the word.<\/p>\n\n<p>The input is the ascii value of each character in the Hash divided by 256.\nThe output of the network is a binary representation I have made. Say for example a is equal to 00000 and b is equal to 00001 and so on and so forth. It only includes the alphabet and the space that why it's only 5 bits per character. I have a maximum limit of only 13 characters in my training input . SO my output nodes is 13 * 5 = 65. And Im expecting a binary output like <code>10101010101010101010101010101010101010101010101010101010101001011<\/code> . The bit sequence can predict at most 16 characters word given a hash code of 32 length as an input.<\/p>\n\n<pre><code>model = Sequential([\n    Dense(32, input_shape=(32,), activation = 'relu'),\n    Dense(20, activation='relu'),\n    Dense(65, input_shape=(65,), activation='softmax')\n])\n\nmodel.summary()\nmodel.compile(Adam(lr=.001), loss='binary_crossentropy', metrics=  ['accuracy'])\nmodel.fit(train_samples, train_labels, batch_size=1000, epochs=10000,shuffle = True, verbose=2)\n<\/code><\/pre>\n\n<p>When I tried predicting using the code below:<\/p>\n\n<pre><code>clf.predict(X)\n<\/code><\/pre>\n\n<p>It always outputs me small decimal values less than 0.5.<\/p>\n\n<pre><code>[[8.95109400e-03 1.11340620e-02 1.27389077e-02 1.90807953e-02\n 1.56925414e-02 7.47500360e-03 1.30378362e-02 1.67052317e-02\n 1.07944654e-02 9.68935993e-03 9.82633699e-03 1.29385451e-02\n 1.56633276e-02 1.38113154e-02 1.50949452e-02 8.81231762e-03\n 1.26177669e-02 1.46279763e-02 1.42763760e-02 1.31389238e-02\n 8.32264405e-03 1.52036361e-02 1.52883027e-02 1.47563582e-02\n 1.19247697e-02 1.16073946e-02 1.72672570e-02 1.35995271e-02\n 1.77132934e-02 1.33292647e-02 1.41840307e-02 1.78522542e-02\n 9.77656059e-03 1.82192177e-02 9.86329466e-03 1.62205566e-02\n 1.95278302e-02 9.18696448e-03 2.06225738e-02 1.01496875e-02\n 2.08229423e-02 2.36334335e-02 6.02523983e-03 2.36746706e-02\n 6.56269025e-03 2.44314633e-02 2.70614270e-02 4.14136378e-03\n 2.72923186e-02 3.86772421e-03 2.90246904e-02 2.92722285e-02\n 3.06371972e-03 2.97660977e-02 1.89558265e-03 3.17853205e-02\n 3.13901827e-02 1.13886443e-03 3.24600078e-02 1.15508994e-03\n 3.36604454e-02 3.36041413e-02 4.59054590e-08 3.35478485e-02\n 4.63940282e-08]]\n<\/code><\/pre>\n\n<p>I'm expecting a binary output. How will I get my desired binary value? I have tried approximating it to 0 when its near to 0 and approximate it to 1 , when its near 1. Is it right? If so, then my output is always 0, because all are close to 0. Which I think is not right. Please help.<\/p>\n","answers":[{"AnswerId":"53482379","CreationDate":"2018-11-26T13:40:18.860","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>There is probably a misconception in your activation function. <\/p>\n\n<p>The <code>softmax<\/code> is designed for \"one correct class\", not for 65 possibly correct classes.<br>\nThe sum of the softmax results will always be 1, thus you will probably not have (m)any thing(s) above .5 indeed.<\/p>\n\n<p>Use a <code>sigmoid<\/code> activation. <\/p>\n"},{"AnswerId":"53482639","CreationDate":"2018-11-26T13:54:06.483","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>Your activation of the last layer is causing the problem. When softmax activation is used the model outputs in a way that the model's output sums to one. This is not the behavior you want. You have two options for binary activation. The first choice is sigmoid activation(It outputs values between 0 and 1). Second options is tanh function(It outputs values between -1 and 1). To convert to binary values, for sigmoid function  use <code>greather than or equals to 0.5<\/code> predicate and for tanh <code>greather than or equals to 0<\/code> predicate.<\/p>\n\n<p>The way you encode the characters is not efficient way for neural networks. Use embedding vector or one hot encoding for your inputs, and also consider using one-hot encoding for your output nodes.<\/p>\n\n<p>My suggestion:  <\/p>\n\n<h2>With Embedding<\/h2>\n\n<pre><code> model = Sequential([\n        Embedding(input_dim, output_dim, input_length=input_length),\n        Dense(32, activation = 'relu'),\n        Dense(20, activation='relu'),\n        Dense(num_of_classes, activation='softmax')\n    ])\n<\/code><\/pre>\n\n<h2>With one hot encoding<\/h2>\n\n<pre><code>   model = Sequential([\n\n        Dense(32,input_shape=(32, number_of_classes), activation = 'relu'),\n        Dense(20, activation='relu'),\n        Dense(num_of_classes, activation='softmax')\n    ])\n<\/code><\/pre>\n"},{"AnswerId":"58454296","CreationDate":"2019-10-18T15:50:37.063","ParentId":null,"OwnerUserId":"6551580","Title":null,"Body":"<pre><code>def classify_local(sentence):\n    ERROR_THRESHOLD = 0.15\n    input_data = pd.DataFrame([bow(sentence, words)], dtype=float, index=['input'])\n    results = model.predict([input_data])[0]\n    results = [[i,r] for i,r in enumerate(results) if r &gt; ERROR_THRESHOLD]\n    results.sort(key=lambda x: x[1], reverse=True)\n    return_list = []\n    for r in results:\n        return_list.append((classes[r[0]], str(r[1])))\n    return return_list\n<\/code><\/pre>\n"}]}
+{"QuestionId":53482975,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensor>","CreationDate":"2018-11-26T14:13:06.137","AcceptedAnswerId":53483606.0,"OwnerUserId":7503252.0,"Title":"Changing the Values of a Certain Column of a Tensor","Body":"<p>I want to overwrite the values of a certain column of a 3-dimensional tensor(batch size, sequence length, number of classes) with a 2-dimensional tensor (batch size, sequence length). I have tried the following value assignment on numpy while debugging and worked perfectly fine but not sure how to do the same on tensor. <\/p>\n\n<pre><code>Numpy Solution:\n\n    Tensor A shape [50,4,4]\n    Tensor B shape [50,4]\n\n  * A[:,:,0]=b[:,:] \n    Tensor A shape is [50,4,4]\n\nExample: \n    A[1]: \n        [[0.2,0.6,0.1,0.02],\n        [0.3,0.4,0.5,0.12],\n        [0.2,0.46,0.31,0.02],\n        [0.2,0.1,0.2,0.03]]\n    B[1]:\n        [0,1,1,0]\n    A*[1]:\n        [[0,0.6,0.1,0.02],\n        [1,0.4,0.5,0.12],\n        [1,0.46,0.31,0.02],\n        [0,0.1,0.2,0.03]]\n<\/code><\/pre>\n\n<p>I know that item assignment is not supported on tensors but was wondering if there is a way without losing the data of ref tensor.<\/p>\n","answers":[{"AnswerId":"53483606","CreationDate":"2018-11-26T14:47:37.920","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think the easiest thing in this case would be:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.placeholder(tf.float32, [None, None, None])\nb = tf.placeholder(tf.float32, [None, None])\na_star = tf.concat([b[:, :, tf.newaxis], a[:, :, 1:]], axis=-1)\n# Test\nwith tf.Session() as sess:\n    print(sess.run(a_star, feed_dict={\n        a: [[[0.2 , 0.6 , 0.1 , 0.02],\n             [0.3 , 0.4 , 0.5 , 0.12],\n             [0.2 , 0.46, 0.31, 0.02],\n             [0.2 , 0.1 , 0.2 , 0.03]]],\n        b: [[0, 1, 1, 0]]\n    }))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[[0.   0.6  0.1  0.02]\n  [1.   0.4  0.5  0.12]\n  [1.   0.46 0.31 0.02]\n  [0.   0.1  0.2  0.03]]]\n<\/code><\/pre>\n\n<p>I proposed a more flexible replacement operation for contiguous slices in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18383#issuecomment-439459312\" rel=\"nofollow noreferrer\">TensorFlow issue #18383<\/a>, but this is probably simpler and faster in this case.<\/p>\n"}]}
+{"QuestionId":53483153,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2018-11-26T14:22:25.050","AcceptedAnswerId":53483547.0,"OwnerUserId":1212320.0,"Title":"Tensorflow: while loop on tensor","Body":"<p>I'm trying apply a while loop over a tensor's value. For example for variable \"a\" I am trying to increase the tensor's value incrementally till a certain condition is met. However, I keep getting this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 0 but is rank 3 for 'while_12\/LoopCond'\n  (op: 'LoopCond') with input shapes: [3,1,1].<\/p>\n<\/blockquote>\n\n<pre><code>a = array([[[0.76393723]],\n       [[0.93270312]],\n       [[0.08361106]]])\n\na = np.random.random((3,1,1))\na1 = tf.constant(np.float64(a))\ni = tf.constant(np.float64(6.14))\n\nc = lambda i: tf.less(i, a1)\nb = lambda x: tf.add(x, 0.1)\nr = tf.while_loop(c, b, [a1])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53483547","CreationDate":"2018-11-26T14:44:21.453","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>The first argument of the  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">tf.while_loop()<\/a> should return scalar (the tensor of rank 0 is, actually, a scalar - that's what the error message is about). In your example you probably want to make the condition return <code>true<\/code> in case if all the numbers in the <code>a1<\/code> tensor are less than <code>6.14<\/code>. This can be achieved by <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/reduce_all\" rel=\"nofollow noreferrer\">tf.reduce_all()<\/a> (logical AND) and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/reduce_any\" rel=\"nofollow noreferrer\">tf.reduce_any()<\/a> (logical OR).<\/p>\n\n<p>That snippet has worked for me:<\/p>\n\n<pre><code>tf.reset_default_graph()\n\na = np.random.random_integers(3, size=(3,2))\nprint(a)\n# [[1 1]\n#  [2 3]\n#  [1 1]]\n\na1 = tf.constant(a)\ni = 6\n\n# condition returns True till any number in `x` is less than 6\ncondition = lambda x : tf.reduce_any(tf.less(x, i))\nbody      = lambda x : tf.add(x, 1)\nloop = tf.while_loop(\n    condition,\n    body,\n    [a1],\n)\n\nwith tf.Session() as sess:\n    result = sess.run(loop)\n    print(result)\n    # [[6 6]\n    #  [7 8]\n    #  [6 6]]\n    # All numbers now are greater than 6\n<\/code><\/pre>\n"}]}
+{"QuestionId":53483631,"AnswerCount":0,"Tags":"<opencv><tensorflow><dnn-module>","CreationDate":"2018-11-26T14:48:57.370","AcceptedAnswerId":null,"OwnerUserId":8560600.0,"Title":"CNN model does not load inside Visual Studio 2017 using dnn.readNetFromTensorflow","Body":"<p>I have created a simple CNN model which I froze into <code>tf_model.pb<\/code> file and I have tried loading it using open cv <code>dnn.readNetFromTesnorflow(model_path)<\/code> method inside Jupyter Notebook and it works perfectly. The problem is when I try to load it inside Visual Studio 2017, for some reason it does not load correctly.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>#include &lt;iostream&gt;\n#include &lt;opencv2\/opencv.hpp&gt;\n#include &lt;opencv2\/dnn.hpp&gt;\n#include &lt;vector&gt;\n\n#include &lt;stdio.h&gt;\n#include &lt;stdlib.h&gt;\n#include &lt;string.h&gt;\n\nusing namespace std;\nusing namespace cv;\nusing namespace cv::dnn;\n\nint main(int argc, char** argv) {\n\n\n    class CharNet {\n\n        private:\n            Net network;\n            string output_layer;\n\n        public:\n\n            CharNet(string path) {\n                try {\n                    cout &lt;&lt; \"Path: \"&lt;&lt; path &lt;&lt; endl;\n\n                    network = readNetFromTensorflow(path);\n\n                    vector&lt;String&gt; network_layers = network.getLayerNames();\n                    cout &lt;&lt; \"There are \"&lt;&lt;network_layers.size()&lt;&lt; \" layers\" &lt;&lt; endl;\n                    for (int i = 0; i &lt; network_layers.size(); i++) {\n                        cout &lt;&lt; network_layers[i] &lt;&lt; endl;\n                        if ( network_layers[i].find(\"Softmax\") != std::string::npos ) {\n                            cout &lt;&lt; \"Output Layer: \"&lt;&lt; network_layers[i] &lt;&lt; endl;\n                            output_layer = network_layers[i];\n                            break;\n                        }\n                    }\n\n                }\n                catch (cv::Exception&amp; e) {\n                    cerr &lt;&lt; \"Exception: \" &lt;&lt; e.what() &lt;&lt; endl;\n                    if (network.empty()) {\n                        cerr &lt;&lt; \"Can't load the model\" &lt;&lt; endl;\n                    }\n                }\n            }\n\n    };\n\n    string model = \"C:\/Users\/stefan_cepa995\/source\/repos\/OpenCV_Test\/OpenCV_Test\/tf_model.pb\";\n\n    CharNet* obj = new CharNet(model);\n\n    return 0;\n\n}\n<\/code><\/pre>\n\n<p>Since the model is not loaded correctly I never enter the for loop. Does anyone have any idea whats going on?<\/p>\n","answers":[]}
+{"QuestionId":53483685,"AnswerCount":6,"Tags":"<python><keras><anaconda>","CreationDate":"2018-11-26T14:51:26.623","AcceptedAnswerId":null,"OwnerUserId":6173146.0,"Title":"Keras breaks Anaconda Prompt","Body":"<p>I am switching from tensorflow to keras on my Anaconda distribution and am having some problems with the latter. I install it through Anaconda prompt with the command<\/p>\n\n<pre><code>conda install keras\n<\/code><\/pre>\n\n<p>and I do not think the installation is properly finished since it runs the command<\/p>\n\n<pre><code>python -c \"import keras\"  1&gt;nul 2&gt;&amp;1\n<\/code><\/pre>\n\n<p>and closes the prompt. Afterwards, if I am to open the command line it automatically runs the command above and closes it, so I am unable to use the prompt. This has happened for both Anaconda 5.3.1 (Python 3.7) and Anaconda 5.2.0 (Python 3.6).<\/p>\n\n<p>Thank you very much in advance. Any help will be much appreciated.<\/p>\n","answers":[{"AnswerId":"55992534","CreationDate":"2019-05-05T13:39:19.227","ParentId":null,"OwnerUserId":"10010695","Title":null,"Body":"<p>I tried almost every solution to this problem (erasing \"nul\" from activate.d seemed to work at first, but then conda commands related to packages still crashed the prompt). So this is what I did. The problems seems to originate in the way that conda installs keras.<\/p>\n\n<p>1) Uninstall keras <a href=\"https:\/\/datatofish.com\/pip-uninstall-package\/\" rel=\"nofollow noreferrer\">using pip<\/a>. Use the Scripts folder in the Anaconda installation folder.<\/p>\n\n<p>2) Manually delete every remaining folder from Keras. Most of them are located in the Anaconda installation folder. Don't forget to delete the <code>keras_activate.bat<\/code> and <code>keras_deactive.bat<\/code> files in the <code>activate.d<\/code> and <code>deactivate.d<\/code> folders.<\/p>\n\n<p>3) Install keras using pip.<\/p>\n\n<p>4) That solves the <code>activate.d<\/code> problem. However, Anaconda Prompt still crashes because of the other file in the <code>%UserProfile%Anaconda3\/etc\/conda<\/code> folder, which is called <code>vs2015_compiler_vars.bat<\/code>. Delete that file and everything will work just fine (weird error messages that appeared while using Keras will also go away).<\/p>\n\n<p>P.S. I went through one extra step to make Anaconda Prompt work perfectly, but I don't know if it is related to installing Keras (that's the reason why I am not including it in the answer). As conda commands were getting stuck in \"Solving environment\", I enabled strich channel priority with <code>conda config --set channel_priority strict<\/code>. Now Anaconda is totally functional!<\/p>\n"},{"AnswerId":"54193036","CreationDate":"2019-01-15T05:16:43.900","ParentId":null,"OwnerUserId":"10915210","Title":null,"Body":"<p>I had the same problem, took me 2 days to finally get things working and many re-installs. This message comes from the following file. <code>%UserProfile%Anaconda3\/etc\/conda\/activate.d\/keras_activate.bat.<\/code>  There is some issue in the code written in this file which closes the Anaconda prompt every time. I dont know how to solve it, perhaps someone else can suggest something, but one way to still do some essential tasks on your prompt window is -> RIGHT CLICK ON THE ABOVE MENTIONED FILE AND SELECT EDIT -> WRITE YOUR ANACONDA PROMPT COMMAND AT THE TOP OF THE FILE. (It will execute and then close.)<\/p>\n"},{"AnswerId":"54279712","CreationDate":"2019-01-20T18:39:50.770","ParentId":null,"OwnerUserId":"10941601","Title":null,"Body":"<p>IN below line in .bat file nul is creating the problem and it is closing the terminal please remove nul and this will be solved. So change this<\/p>\n\n<pre><code>python -c \"import keras\" 1&gt;nul 2&gt;&amp;1\n<\/code><\/pre>\n\n<p>to this<\/p>\n\n<pre><code>python -c \"import keras\" 1&gt; 2&gt;&amp;1\n<\/code><\/pre>\n"},{"AnswerId":"54304639","CreationDate":"2019-01-22T09:03:53.687","ParentId":null,"OwnerUserId":"10949095","Title":null,"Body":"<p>I have the same problem. I didn't find a permanent fix, but before the script finishes running, you can escape it using <code>ctrl+d<\/code>, and you should be able to do whatever you want after that.<\/p>\n"},{"AnswerId":"54327432","CreationDate":"2019-01-23T12:35:30.510","ParentId":null,"OwnerUserId":"5929443","Title":null,"Body":"<p>I figured out the answer after combining answers from GAURAV and GYAN ARORA. The solution is this:<\/p>\n\n<p>1) Go to <code>%UserProfile%Anaconda3\/etc\/conda\/activate.d<\/code>and right click on <code>keras_activate.bat<\/code>\n2) Click on edit. This is what the .bat file looks like:<\/p>\n\n<pre><code>:: Figure out the default Keras backend by reading the config file.\npython %CONDA_PREFIX%\\etc\\keras\\load_config.py &gt; temp.txt\nset \/p KERAS_BACKEND=&lt;temp.txt\ndel temp.txt\n\n:: Try to use the default Keras backend.\n:: Fallback to Theano if it fails (Theano always works).\npython -c \"import keras\" 1&gt;nul 2&gt;&amp;1\nif errorlevel 1 (\n    ver &gt; nul\n    set \"KERAS_BACKEND=theano\"\n    python -c \"import keras\" 1&gt;nul 2&gt;&amp;1\n)\n<\/code><\/pre>\n\n<p>Change both <code>1&gt;nul<\/code> to <code>1&gt;<\/code>. The final file should look like this:<\/p>\n\n<pre><code>:: Figure out the default Keras backend by reading the config file.\npython %CONDA_PREFIX%\\etc\\keras\\load_config.py &gt; temp.txt\nset \/p KERAS_BACKEND=&lt;temp.txt\ndel temp.txt\n\n:: Try to use the default Keras backend.\n:: Fallback to Theano if it fails (Theano always works).\npython -c \"import keras\" 1&gt; 2&gt;&amp;1\nif errorlevel 1 (\n    ver &gt; nul\n    set \"KERAS_BACKEND=theano\"\n    python -c \"import keras\" 1&gt; 2&gt;&amp;1\n)\n<\/code><\/pre>\n\n<p>3) Save and close<\/p>\n"},{"AnswerId":"54672741","CreationDate":"2019-02-13T14:37:16.650","ParentId":null,"OwnerUserId":"7900034","Title":null,"Body":"<p>The best solution I found was ,<\/p>\n\n<p>1.Uninstall Keras first<\/p>\n\n<p>2.goto the location \n<code>C:\\Users\\username\\AppData\\Local\\Continuum\\anaconda3\\etc\\conda\\activate.d<\/code><\/p>\n\n<p>You can see the keras batch files inside in both <code>activate.d<\/code> and <code>deactivate.d<\/code>, which runs every time the anaconda prompt is <code>opened.DELETE<\/code> them.<\/p>\n\n<p>3.Reinstall Keras.<\/p>\n\n<p>This worked well for me.<\/p>\n"}]}
+{"QuestionId":53484544,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><tensorboard>","CreationDate":"2018-11-26T15:40:37.833","AcceptedAnswerId":53499607.0,"OwnerUserId":5577386.0,"Title":"Display Keras graph in Tensorboard without using the callback in the fit method","Body":"<p>Is it possible to display a Keras graph in Tensorboard without using the tensorboard callback in the fit method?<\/p>\n\n<p>Is it possible to extract the graph from Keras and display the graph using the tensorflow FileWriter?\n<code>tf.summary.FileWriter(logdir='logdir', graph=graph)<\/code><\/p>\n\n<p>I want to do this to check all the connections of this part of the graph are as expected (this model is part of a larger model that is far from finished).<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"53499607","CreationDate":"2018-11-27T12:22:20.570","ParentId":null,"OwnerUserId":"5577386","Title":null,"Body":"<p>It turned out to be very simple by extracting the Tensorflow graph from the backend and using the file writer.<\/p>\n\n<pre><code>import tensorflow as tf\n# Used to get the graph\nfrom tensorflow.python.keras import backend as K\n\n\ntb_path = \"logs\/\"\n\n# Simple model to test the tensorboard plotting\nmodel = SimpleModel(50, 20, 10).build_model()\n\n# Get the sessions graph\ngraph = K.get_session().graph\n\n# Display with the tensorflow file writer\nwriter = tf.summary.FileWriter(logdir=tb_path, graph=graph)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53484975,"AnswerCount":0,"Tags":"<keras><loss-function><crf>","CreationDate":"2018-11-26T16:07:02.903","AcceptedAnswerId":null,"OwnerUserId":5107841.0,"Title":"loss function returns Nan in join mode of CRF keras-contrib","Body":"<p>I use a BiLSTM-CRF architecture to assign some labels to a sequence of the sentences in a paper. We have 150 papers each of which contains 380 sentences and each sentence is represented by a double array with size 11 in range (0,1) and the number of class labels is 11.<\/p>\n\n<pre><code>input = Input(shape=(None,11))\nmask = Masking (mask_value=0)(input)\nlstm = Bidirectional(LSTM(50, return_sequences=True))(mask)\nlstm = Dropout(0.3)(lstm)\nlstm= TimeDistributed(Dense(50, activation=\"relu\"))(lstm)\ncrf = CRF(11 , sparse_target=False , learn_mode='join')  # CRF layer\nout = crf(lstm)\nmodel = Model(input, out)\nmodel.summary()\nmodel.compile('adam', loss=crf.loss_function, metrics=[crf.accuracy])\n<\/code><\/pre>\n\n<p>I use keras-contrib package to implement CRF layer. CRF layer has two learning modes: join mode and marginal mode. I know that  join mode is a real CRF that uses viterbi algorithm to predict the best path. While, marginal mode is not a real CRF that uses categorical-crossentropy for computing loss function.\nWhen I use marginal mode, the output is as follows:<\/p>\n\n<pre><code>Epoch 4\/250:  - 6s - loss: 1.2289 - acc: 0.5657 - val_loss: 1.3459 - val_acc: 0.5262\n<\/code><\/pre>\n\n<p>But, in the join mode, the value of loss function is nan:<\/p>\n\n<pre><code>Epoch 2\/250 :  - 5s - loss: nan - acc: 0.1880 - val_loss: nan - val_acc: 0.2120\n<\/code><\/pre>\n\n<p>I do not understand why this happens and would be grateful to anybody who could give a hint.<\/p>\n","answers":[]}
+{"QuestionId":53485724,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-11-26T16:54:33.100","AcceptedAnswerId":53485987.0,"OwnerUserId":1179925.0,"Title":"Keras : is it possible to give lamda layer a name?","Body":"<p>I have following code inside my model definition<\/p>\n\n<pre><code># (None, 6) -&gt; (None, 8)\ndef affine_to_projective(x):\n    y = tf.pad(x,\n               paddings=tf.constant([[0, 0], [0, 2]]),\n               mode='CONSTANT',\n               constant_values=0\n               )\n    return y\n\nprojective_transform = Lambda(affine_to_projective)(affine_transform)\n<\/code><\/pre>\n\n<p>In model summary it's seen as <\/p>\n\n<pre><code>lambda_1 (Lambda)               (None, 8)            0           affine_transform[0][0]\n<\/code><\/pre>\n\n<p>Is it possible to give this layer a name instead of <code>lambda_1<\/code>?<\/p>\n","answers":[{"AnswerId":"53485987","CreationDate":"2018-11-26T17:12:27.920","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<pre><code>projective_transform = Lambda(affine_to_projective, name='lambda_name')(affine_transform)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53486016,"AnswerCount":1,"Tags":"<keras><conv-neural-network><max-pooling>","CreationDate":"2018-11-26T17:14:20.613","AcceptedAnswerId":null,"OwnerUserId":9421400.0,"Title":"Does a maxpooling layer reduce the number of parameters in a network?","Body":"<p>I have a simple network defined: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv1D(5, 3, activation='relu', input_shape=(10, 1),name=\"conv1\",padding=\"same\"))\nmodel.add(MaxPooling1D())\nmodel.add(Conv1D(5, 3, activation='relu',name=\"conv2\",padding=\"same\"))\nmodel.add(MaxPooling1D())\nmodel.add(Dense(1, activation='relu',name=\"dense1\"))\nmodel.compile(loss='mse', optimizer='rmsprop')\n<\/code><\/pre>\n\n<p>The shape of the layers is as follows:<\/p>\n\n<pre><code>conv1-(None, 10, 5)\n\nmax1-(None, 5, 5)\n\nconv2-(None,5,5)\n\nmax2-(None,2,5)\n\ndense1-(None,2,1)\n<\/code><\/pre>\n\n<p>The model has a total of 106 parameters, however if I remove max pooling layer then the model summary looks as follows:<\/p>\n\n<pre><code>conv1-(None, 10, 5) \n\nconv2-(None,10,5)\n\ndense1-(None,10,1)\n<\/code><\/pre>\n\n<p>In both the cases total parameters remain 106, but why is it commonly written that the max-pooling layer reduces the number of parameters?  <\/p>\n","answers":[{"AnswerId":"53486105","CreationDate":"2018-11-26T17:20:52.753","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Which kind of network? It's all up to you. <\/p>\n\n<ul>\n<li>Conv layers: no    <\/li>\n<li>Dense layers:   \n\n<ul>\n<li>Directly after Conv or Pooling:\n\n<ul>\n<li>With \"channels_last\": no   <\/li>\n<li>With \"channels_first\": yes    <\/li>\n<\/ul><\/li>\n<li>After Flatten layers: yes    <\/li>\n<li>After GlobalPooling layers: no    <\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Your network: no. <\/p>\n\n<h2>Explanations<\/h2>\n\n<ul>\n<li>Poolings and GlobalPoolings change the image sizes, but don't change the number of channels    <\/li>\n<li>Conv layers are fixed size filters that stride along the images.The filter size is independent of the image size, thus there is no change. Filters depend on kernel size and channels    <\/li>\n<li>Dense layers work on the last dimension only.    \n\n<ul>\n<li>If the last dimension is channels, the pooling layers don't affect it    <\/li>\n<li>If the last dimension is an image side, it's affected    <\/li>\n<\/ul><\/li>\n<li>Flatten layers transform the image sizes and channels into a single dimension.     <\/li>\n<\/ul>\n"}]}
+{"QuestionId":53486118,"AnswerCount":1,"Tags":"<keras><amazon-sagemaker>","CreationDate":"2018-11-26T17:21:42.240","AcceptedAnswerId":null,"OwnerUserId":7496670.0,"Title":"Early Stopping and Callbacks with Keras when using SageMaker","Body":"<p>I am using sagemaker to train a keras model. I need to implement early stoping approach when training the model. <\/p>\n\n<p>Is there a way to pass callbacks such as EarlyStopping, Histories..etc. <\/p>\n\n<p>In traditional way, we used to pass this as a parameter to keras's fit function:<\/p>\n\n<pre><code>results = model.fit(train_x_trim, train_y_trim, \n                    validation_data=(test_x, test_y), \n                    epochs=FLAGS.epoch,  \n                    verbose=0, \n                    callbacks=[tboard, checkpointer, early_stopping, history])\n<\/code><\/pre>\n\n<p>However, if using SageMaker, we need to call SageMaker's fit function instead which doesn't support callbacks. <\/p>\n\n<pre><code>from sagemaker.tensorflow import TensorFlow \niris_estimator = TensorFlow(entry_point='training_code.py', \n                            role=role, output_path=model_location, \n                            code_location=custom_code_upload_location, \n                            train_instance_count=1, \n                            train_instance_type='ml.c4.xlarge', \n                            training_steps=1000, \n                            evaluation_steps=100)\n<\/code><\/pre>\n\n<p>Any idea how to implement callbacks in SageMaker ? <\/p>\n","answers":[{"AnswerId":"53824799","CreationDate":"2018-12-18T00:27:02.383","ParentId":null,"OwnerUserId":"9090582","Title":null,"Body":"<p>I apologize for the late response.<\/p>\n\n<p>It looks like the Keras code you specified above is essentially your algorithm code. This would be defined in your user script, which would be \"training_code.py\" in the SageMaker Python SDK example you provided.<\/p>\n\n<p>Starting with TensorFlow 1.11, the SageMaker predefined TensorFlow containers have support for \"script mode\". You should be able to specify your Keras callbacks within your user script.<\/p>\n\n<p>For more information: <a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/README.rst#tensorflow-sagemaker-estimators-and-models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/README.rst#tensorflow-sagemaker-estimators-and-models<\/a><\/p>\n"}]}
+{"QuestionId":53486421,"AnswerCount":1,"Tags":"<installation><protocol-buffers><caffe>","CreationDate":"2018-11-26T17:43:10.980","AcceptedAnswerId":null,"OwnerUserId":2651062.0,"Title":"Which protobuf version to use with caffe?","Body":"<p>I am trying to install Caffe on Ubuntu 16.04. As I want to use it with OpenPose, I don't want to use Anaconda for the installation of Caffe. After having installed a lot of Caffe dependencies (following multiple tutorials online), I discovered that protobuf was installed using python:<\/p>\n\n<blockquote>\n  <p><strong>$pip show protobuf<\/strong><br>\n  Name: protobuf <br\/>\n  Version: 3.6.1 <br\/>\n  Summary: Protocol<br\/>\n  Buffers Home-page: <a href=\"https:\/\/developers.google.com\/protocol-buffers\/\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/protocol-buffers\/<\/a><br\/>\n  Author: None Author-email: None License: 3-Clause BSD License<br\/>\n  Location: \/home\/userq\/.local\/lib\/python2.7\/site-packages <br\/>\n  Requires: setuptools, six<br\/> \n  Required-by:<\/p>\n<\/blockquote>\n\n<p>I know from past experiences that using <strong>apt-get install libprotobuf-dev protobuf-compiler<\/strong> installs a protobuf version that is older than 3.6.1, so I downloaded protobuf 3.6.1 binaries and put them in \/usr\/bin and \/usr\/include. Now I have:<\/p>\n\n<blockquote>\n  <p><strong>$ protoc --version<br\/><\/strong>\n  libprotoc 3.6.1<\/p>\n<\/blockquote>\n\n<p>But when I try to build Caffe, It fails:<\/p>\n\n<blockquote>\n  <p>CXX .build_release\/src\/caffe\/proto\/caffe.pb.cc\n  In file included from .build_release\/src\/caffe\/proto\/caffe.pb.cc:4:0:\n  .build_release\/src\/caffe\/proto\/caffe.pb.h:12:2: error: #error This <strong>file was generated by a newer version of protoc which is<\/strong>\n   <strong>#error This file was generated by a newer version of protoc which is<\/strong>\n    ^\n  .build_release\/src\/caffe\/proto\/caffe.pb.h:13:2: error: #error incompatible with your Protocol Buffer headers. Please update\n   #error incompatible with your Protocol Buffer headers.  Please update\n    ^\n  .build_release\/src\/caffe\/proto\/caffe.pb.h:14:2: error: #error your headers.\n   #error your headers.\n    ^\n  In file included from .build_release\/src\/caffe\/proto\/caffe.pb.cc:4:0:\n  .build_release\/src\/caffe\/proto\/caffe.pb.h:23:35: fatal error: google\/protobuf\/arena.h: No such file or directory\n  compilation terminated.\n  Makefile:598: recipe for target '.build_release\/src\/caffe\/proto\/caffe.pb.o' failed\n  make: *** [.build_release\/src\/caffe\/proto\/caffe.pb.o] Error 1<\/p>\n<\/blockquote>\n\n<p>When I open the file <strong>.build_release\/src\/caffe\/proto\/caffe.pb.h<\/strong>, I find that such an error message is generated when the protobuf version is older than 3.6.1. Here's an excerpt from this file:<\/p>\n\n<pre><code>#if GOOGLE_PROTOBUF_VERSION &lt; 3006001\n#error This file was generated by a newer version of protoc which is\n#error incompatible with your Protocol Buffer headers.  Please update\n#error your headers.\n#endif\n#if 3006001 &lt; GOOGLE_PROTOBUF_MIN_PROTOC_VERSION\n#error This file was generated by an older version of protoc which is\n#error incompatible with your Protocol Buffer headers.  Please\n#error regenerate this file with a newer version of protoc.\n#endif\n<\/code><\/pre>\n\n<p>I don't think there is a version conflict in this case. And although I have the version 3.6.1, it looks like the system doesn't detect it. What could be the problem?<\/p>\n","answers":[{"AnswerId":"54167323","CreationDate":"2019-01-13T08:49:44.390","ParentId":null,"OwnerUserId":"10400041","Title":null,"Body":"<p>I also have the same issue but I think it is because of protoc --version 3.6 try to install protoc 3.0 for me it solved.<\/p>\n"}]}
+{"QuestionId":53486527,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-26T17:51:16.717","AcceptedAnswerId":53486604.0,"OwnerUserId":7457450.0,"Title":"What does tf.reduce_sum do with axis = -1?","Body":"<p>I don't understand why the output of the following code is <code>[7 56]<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.constant([[1, 2, 4], [8, 16, 32]])\na = tf.reduce_sum(x, -1)  # [ 9 18 36]\n\nwith tf.Session() as sess:\n  output_a = sess.run(a)\n  print(output_a)\n<\/code><\/pre>\n\n<p>I get that row-wise addition has been done. But can someone shed some light on why <code>-1<\/code> in the <code>reduce_sum<\/code> function has been treated to sum all values in a row?<\/p>\n","answers":[{"AnswerId":"53486604","CreationDate":"2018-11-26T17:57:17.730","ParentId":null,"OwnerUserId":"4983450","Title":null,"Body":"<p><code>-1<\/code> means the last axis; Since you have a rank 2 tensor, the last axis is the second axis, that is, along the rows; <code>tf.reduce_sum<\/code> with <code>axis=-1<\/code> will thus reduce (sum) the second dimension.<\/p>\n"},{"AnswerId":"58121193","CreationDate":"2019-09-26T16:33:13.390","ParentId":null,"OwnerUserId":"11126311","Title":null,"Body":"<p>I run your code and actually gave me a different answer:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.constant([[1, 2, 4], [8, 16, 32]])\na = tf.reduce_sum(x, -1)  \ntf.print(a)\n<\/code><\/pre>\n\n<p>The answer is [7,56], which is added 1+2+4 =7, and 8+16+32=56. <\/p>\n\n<p>axis: The dimensions to reduce. \nMy understanding:<\/p>\n\n<ol>\n<li>'-1' means the last axis, or dimension. <\/li>\n<\/ol>\n\n<p>tf.reduce_sum(x, -1) is equal to tf.reduce_sum(x, 1) here since only 2 dimensions. <\/p>\n\n<ol start=\"2\">\n<li>The dimension is 2*3  (2 rows*3 cols). Goal is to remove the last dimension '3' and the result should be 2 rows:<\/li>\n<\/ol>\n\n<p><code>[[7]\n[56]]<\/code><\/p>\n\n<p>Since no 'keepdims=True' here, [] will be removed and we get result [7,56]<\/p>\n\n<ol start=\"3\">\n<li>You can test this example. Note: '-1' and '1' are different here. <\/li>\n<\/ol>\n\n<p><code>y = tf.constant([[[1, 2, 4], [1, 0, 3]],[[1,2,3],[2,2,1]]])\nc = tf.reduce_sum(y, 1)  # if (y,-1) will be [[7,4],[6,5]]\ntf.print(c)   #[[2 2 7], [3 4 4]]\n<\/code><\/p>\n"}]}
+{"QuestionId":53486698,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><tensorflow-estimator>","CreationDate":"2018-11-26T18:04:23.663","AcceptedAnswerId":null,"OwnerUserId":10707745.0,"Title":"tensorflow LinearClassifier predict function","Body":"<p>I've been playing around with tensorflow and because of lack of documentation I'm having a lot of troubles to understand some of the function outputs. <\/p>\n\n<p>In this case, for the module <strong>tf.estimator.LinearClassifier<\/strong> I don't understand what the parameters in the dictionary returned by the function <strong>predict()<\/strong> mean. Below there's an example:<\/p>\n\n<blockquote>\n  <p>{'probabilities': array([0.92966837, 0.07033166], dtype=float32),\n  'logits': array([-2.581606], dtype=float32), 'classes': array(['0'],\n  dtype=object), 'class_ids': array([0]), 'logistic':\n  array([0.07033166], dtype=float32)}<\/p>\n<\/blockquote>\n\n<p>I assume that the array \"probabilities\" contains = [1-p, p], where \"p\" is the probability and I understand the \"classes\" and \"class_ids\" meaning. <\/p>\n\n<p>However I get lost with the other dict keys. Can someone point me to some good documentation where the outputs of functions are explained?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":53486701,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><mnist>","CreationDate":"2018-11-26T18:04:44.087","AcceptedAnswerId":null,"OwnerUserId":10707719.0,"Title":"mnist train and test accuracy only 0.49","Body":"<p>I am trying to implement the basic ANN on MNIST dataset, and I use the same code in tensorflow: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/<\/a>, but I get very low accuracy on training and testing dataset. I can't see what cause the problem. below is the output:<\/p>\n\n<pre><code>    2018-11-26 11:55:02.526281: I \ntensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports \ninstructions that this TensorFlow binary was not compiled to use: AVX2 \nFMA\nEpoch 1\/10\n   32\/7000 [..............................] - ETA: 1:12 - loss: 1.4159 \n- acc: 0.4062\n 960\/7000 [===&gt;..........................] - ETA: 2s - loss: 1.1616 - \nacc: 0.4052  \n1920\/7000 [=======&gt;......................] - ETA: 1s - loss: 1.0438 - \nacc: 0.4297\n2816\/7000 [===========&gt;..................] - ETA: 0s - loss: 1.0004 - \nacc: 0.4592\n3712\/7000 [==============&gt;...............] - ETA: 0s - loss: 0.9788 - \nacc: 0.4749\n4448\/7000 [==================&gt;...........] - ETA: 0s - loss: 0.9691 - \nacc: 0.4795\n5408\/7000 [======================&gt;.......] - ETA: 0s - loss: 0.9602 - \nacc: 0.4741\n6368\/7000 [==========================&gt;...] - ETA: 0s - loss: 0.9567 - \nacc: 0.4763\n7000\/7000 [==============================] - 1s 105us\/step - loss: \n0.9543\nacc: 0.4773\n.................(I omitted some of the outputs, because they are almost the same.)\n\nEpoch 10\/10\n  32\/7000 [..............................] - ETA: 0s - loss: 0.9383 - \nacc: 0.5000\n 896\/7000 [==&gt;...........................] - ETA: 0s - loss: 0.9201 - \nacc: 0.5022\n1696\/7000 [======&gt;.......................] - ETA: 0s - loss: 0.9120 - \nacc: 0.5083\n1984\/7000 [=======&gt;......................] - ETA: 0s - loss: 0.9153 - \nacc: 0.5055\n2560\/7000 [=========&gt;....................] - ETA: 0s - loss: 0.9210 - \nacc: 0.5043\n3328\/7000 [=============&gt;................] - ETA: 0s - loss: 0.9208 - \nacc: 0.5072\n4096\/7000 [================&gt;.............] - ETA: 0s - loss: 0.9199 - \nacc: 0.5054\n4672\/7000 [===================&gt;..........] - ETA: 0s - loss: 0.9235 - \nacc: 0.4991\n5344\/7000 [=====================&gt;........] - ETA: 0s - loss: 0.9219 - \nacc: 0.5011\n6144\/7000 [=========================&gt;....] - ETA: 0s - loss: 0.9240 - \nacc: 0.4977\n7000\/7000 [==============================] - 1s 75us\/step - loss: \n0.9235 \nacc: 0.4946\n32\/3000 [..............................] - ETA: 8s\n1920\/3000 [==================&gt;...........] - ETA: 0s\n3000\/3000 [==============================] - 0s 54us\/step\nTest accuracy: 0.49966666666666665\n<\/code><\/pre>\n\n<hr>\n\n<p>I just use the code below, that is from that website:<a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/<\/a>. The dataset is MNIST.<\/p>\n\n<pre><code>import tensorflow as tf\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\nmodel = tf.keras.models.Sequential([\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(512, activation=tf.nn.relu),\n    tf.keras.layers.Dropout(0.2),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=5)\n\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n\n<p>Actually, I just tried the code on google colab, and the result is normal:<\/p>\n\n<pre><code>Epoch 1\/1\n60000\/60000 [==============================] - 16s 270us\/step - loss: 0.2015 - acc: 0.9406\n10000\/10000 [==============================] - 1s 61us\/step\n[0.11236203697770834, 0.9653]\n<\/code><\/pre>\n\n<p>Now I assume that this problem is caused by my computer, which is Macbook early 2015, and the OS is Mac OS X 10.14.1.<\/p>\n","answers":[]}
+{"QuestionId":53487595,"AnswerCount":1,"Tags":"<tensorflow><johnsnowlabs-spark-nlp>","CreationDate":"2018-11-26T19:13:05.877","AcceptedAnswerId":null,"OwnerUserId":4992374.0,"Title":"Where can I find a list of class labels for pretrained SparkNLP NerDLModel?","Body":"<p>I have been searching for a while but no luck finding out what NER labels are included in the pretrained NerDL(tensorflow) model. I would think the training data can provide such information, but I do not see it mentioned in any documentation.<\/p>\n\n<p>downloadable model:\n<a href=\"https:\/\/s3.amazonaws.com\/auxdata.johnsnowlabs.com\/public\/models\/ner_precise_en_1.7.0_2_1539623388047.zip\" rel=\"nofollow noreferrer\">https:\/\/s3.amazonaws.com\/auxdata.johnsnowlabs.com\/public\/models\/ner_precise_en_1.7.0_2_1539623388047.zip<\/a><\/p>\n\n<p>Any direction would be appreciated!<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>I indeed filed an issue in SparkNLP github following the advice here :)\nI just heard back from them. Here is the answer:<\/p>\n\n<blockquote>\n  <p>For practical purposes, the pretrained NER model has<\/p>\n  \n  <p>B-ORG<\/p>\n  \n  <p>I-ORG<\/p>\n  \n  <p>B-PER<\/p>\n  \n  <p>I-PER<\/p>\n  \n  <p>B-LOC<\/p>\n  \n  <p>I-LOC<\/p>\n  \n  <p>and it has been trained from:\n  <a href=\"https:\/\/raw.githubusercontent.com\/patverga\/torch-ner-nlp-from-scratch\/master\/data\/conll2003\/eng.train\" rel=\"nofollow noreferrer\">https:\/\/raw.githubusercontent.com\/patverga\/torch-ner-nlp-from-scratch\/master\/data\/conll2003\/eng.train<\/a><\/p>\n<\/blockquote>\n\n<p>See original issue <a href=\"https:\/\/github.com\/JohnSnowLabs\/spark-nlp\/issues\/312\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"53614105","CreationDate":"2018-12-04T13:31:16.750","ParentId":null,"OwnerUserId":"537068","Title":null,"Body":"<p>that model is trained on the CONLL2003 dataset for NER,<\/p>\n\n<p><a href=\"http:\/\/aclweb.org\/anthology\/W03-0419\" rel=\"nofollow noreferrer\">http:\/\/aclweb.org\/anthology\/W03-0419<\/a><\/p>\n\n<p>That dataset basically has PERSON, ORGANIZATION, and LOCATION.<\/p>\n\n<p>hope this helps,\nAlberto.<\/p>\n"}]}
+{"QuestionId":53487718,"AnswerCount":0,"Tags":"<tensorflow><regression><loss><acc>","CreationDate":"2018-11-26T19:22:43.120","AcceptedAnswerId":null,"OwnerUserId":10281305.0,"Title":"How solve non Loss problem in tensorflow in regression model","Body":"<p>I have a data base with this shape: (1400000, 44)<\/p>\n\n<p>which the 44th column is output.<\/p>\n\n<p>all numbers are float and between 0 and 1. I used a Tensorflow like below but the loss function is non and the acc is zero.<\/p>\n\n<pre><code># Create network with Keras\nimport tensorflow as tf\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nimport numpy as np\n\ndataset=np.loadtxt(\"Dataset5.txt\")\n\n\ns=dataset.size\ntr_size=int( 0.7*s) \n\nX = dataset[0:tr_size,0:43]\n\nY = dataset[0:tr_size,43]\n\n# create model\nmodel = Sequential()\nmodel.add(Dense(64, input_dim=43, init='uniform', activation='relu'))\nmodel.add(Dense(16, init='uniform', activation='relu'))\nmodel.add(Dense(4, init='uniform', activation='sigmoid'))\nmodel.add(Dense(1, init='uniform', activation='relu'))\n# Compile model\n\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\n# Fit the model\nmodel.fit(X, Y, epochs=150, batch_size=1000,  verbose=2)\n# calculate predictions\npredictions = model.predict(X)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53488063,"AnswerCount":3,"Tags":"<python><windows><cmd><keras><anaconda>","CreationDate":"2018-11-26T19:52:33.153","AcceptedAnswerId":null,"OwnerUserId":10496423.0,"Title":"Anaconda Prompt Stuck\/Closing after Keras installation","Body":"<p>So I recently downloaded keras through the anaconda prompt (I've downloaded other packages successfully). It actually worked, as I was able to import keras in Jupyter Notebook. However, now I get the following issue when I open the Anaconda Prompt:<\/p>\n\n<pre><code>C:\\Users\\[User Name]&gt;python C:\\Users\\[User Name]\\Anaconda3\\etc\\keras\\load_config.py  1&gt;temp.txt\npython: can't open file 'C:\\Users\\[User]': [Errno 2] No such file or directory\n\nC:\\Users\\[User Name]&gt;set \/p KERAS_BACKEND= 0&lt;temp.txt\n\nC:\\Users\\[User Name]&gt;del temp.txt\n\nC:\\Users\\[User Name]&gt;python -c \"import keras\"  1&gt;nul 2&gt;&amp;1\n<\/code><\/pre>\n\n<p>Things to Note:<\/p>\n\n<ul>\n<li>At first, I simply cannot type or enter anything into the window.<\/li>\n<li>After closing and reopening it, the same message opens up, but is followed by a bunch of text I can't read because the window immediately closes afterward.<\/li>\n<li>In the error message provided, the third line has [User] instead of user. That is because my users folder name on my computer is FirstName LastName. There is a space which I'm wondering might be the issue. <\/li>\n<li>I don't want to uninstall anaconda, as I have a project due in a few days and don't want any installation\/loss of data issues.<\/li>\n<\/ul>\n\n<p>So, is my folder name having a space the root of the issue? If it is, how do I change it without messing up my computer? If it isn't, do you guys have any ideas of what is the solution to this? Thank you so much, I have two projects due in just a few days :(<\/p>\n","answers":[{"AnswerId":"53488104","CreationDate":"2018-11-26T19:55:17.527","ParentId":null,"OwnerUserId":"8188846","Title":null,"Body":"<p>Use quotes.  This is a <code>cmd.exe<\/code> parsing question.<\/p>\n\n<pre><code>python.exe \"%USERPROFILE%\\Anaconda3\\etc\\keras\\load_config.py\" &gt; temp.txt\n<\/code><\/pre>\n"},{"AnswerId":"56471598","CreationDate":"2019-06-06T05:55:04.553","ParentId":null,"OwnerUserId":"1787555","Title":null,"Body":"<p>delete files in \nC:\\Users\\your_name\\AppData\\Local\\Continuum\\anaconda3\\etc\\conda\\activate.d<\/p>\n\n<p>or you can reinstall\/remove keras from anaconda navigator:\n- right click env name\n- start terminal\n- ctrl-c to stop any batch job.<\/p>\n"},{"AnswerId":"54237245","CreationDate":"2019-01-17T13:42:54.750","ParentId":null,"OwnerUserId":"8902456","Title":null,"Body":"<p>After a few days of searching, I found an alternative solution. <\/p>\n\n<ol>\n<li><p>Open Anaconda Navigator<\/p><\/li>\n<li><p>Click on <code>Environments<\/code> tab ( as shown in the picture)<a href=\"https:\/\/i.stack.imgur.com\/0b0dg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0b0dg.png\" alt=\"enter image description here\"><\/a><\/p><\/li>\n<li><p>Click <code>base(root)<\/code> (triangle icon) and then click on <code>Open Terminal<\/code><a href=\"https:\/\/i.stack.imgur.com\/aNvDv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aNvDv.png\" alt=\"enter image description here\"><\/a><\/p><\/li>\n<\/ol>\n\n<p>By using this solution, you'll be able to open <code>Anaconda Prompt<\/code> window and do whatever you want with <code>Anaconda Prompt<\/code>.\n<a href=\"https:\/\/i.stack.imgur.com\/3sZko.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3sZko.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53488359,"AnswerCount":1,"Tags":"<keras><lstm><rnn>","CreationDate":"2018-11-26T20:14:16.303","AcceptedAnswerId":53490437.0,"OwnerUserId":7297511.0,"Title":"CNN-LSTM Image Classification","Body":"<p>Is it possible to reshape 512x512 rgb image to (timestep, dim)? Otherwards, I am trying to convert this reshape layer: <code>Reshape((23, 3887))<\/code> to 512 vice 299. Also, is there any documentation explaining how to determine input_dim and timestep for Keras? <\/p>\n","answers":[{"AnswerId":"53490437","CreationDate":"2018-11-26T23:04:07.063","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>It seems like your problem is similar to one that i had earlier today. Look at it here: <a href=\"https:\/\/stackoverflow.com\/questions\/53488768\/keras-functional-api-combine-cnn-model-with-a-rnn-to-to-look-at-sequences-of-im\">Keras functional API: Combine CNN model with a RNN to to look at sequences of images<\/a><\/p>\n\n<p>Now to add to the answer from the question i linked too. Let <code>number_of_images<\/code> be <code>n<\/code>. In your case the original data format would be <code>(n, 512, 512, 3)<\/code>. All you then need to do decide how many images you want per sequence. Say you want a sequence of 5 images and have gotten 5000 images in total. Then reshaping to <code>(1000, 5, 512, 512, 3)<\/code> should do. This way the model sees 1000 sequences of 5 images. <\/p>\n"}]}
+{"QuestionId":53488493,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2018-11-26T20:24:10.073","AcceptedAnswerId":null,"OwnerUserId":9655387.0,"Title":"How to pass sequence of image through Conv2D in Keras?","Body":"<p>I have a sequence of 5 images that I want to pass through a CNN sequentially. A single input will have size: <code>(5, width, height, channels)<\/code> and I want to pass each image in the sequence in order to a 2D CNN, concatenate all 5 outputs at some layer and then feed to an LSTM. My model looks something like this:<\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Dense, Input, LSTM, Flatten, Conv2D, MaxPooling2D\n\n# Feed images in sequential order here\ninputs = Input(shape=(128, 128, 3))\nx = Conv2D(16, 3, activation='relu')(inputs)\nx = MaxPooling2D((2, 2))(x)\n...\n# Concatenate sequence outputs here\nx = LSTM(8)(x)\nx = Flatten()(x)\noutputs = Dense(5, activation='sigmoid')\n\nmodel = Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n\n<p>Eventually I want to concatenate all 5 outputs together at some point in the network and feed them to an LSTM but I am having trouble figuring out how to feed sequence of images in order to a 2D convolutional layer. I have looked into 3D convolutional layers and the <code>ConvLSTM2D<\/code> layer but I want to figure out how I can do it this way instead.<\/p>\n","answers":[]}
+{"QuestionId":53488524,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><keras>","CreationDate":"2018-11-26T20:26:23.290","AcceptedAnswerId":null,"OwnerUserId":5208232.0,"Title":"Keras model has very good loss after 1 epoch but doesn't really get better with more epochs","Body":"<p>Hello I just wanted to ask this theoretical question.\nWhat could be the causes of a model that has already a very good loss (0.004 on normalized data) after one single epoch but this loss doesn't really decrease over time (after 10 epochs it's still 0.0032).<\/p>\n\n<p>Shouldn't it normally decrease way more over time?<\/p>\n\n<p>The dataset is pretty big with bit more than a million datapoints and I didn't expect this very good loss just after 1 epoch.<\/p>\n\n<p>So what could I change about this model or what am I doing wrong? (it's a densely connected NN predicting regression with adam and mse)<\/p>\n","answers":[{"AnswerId":"53488732","CreationDate":"2018-11-26T20:42:12.710","ParentId":null,"OwnerUserId":"1410665","Title":null,"Body":"<p>Try changing (increase or decrease) the learning rate. Thus you can ensure the noticeable changes to weights. You can also check if there is a missing of activation (e.g., relu) after convolution or not. You can also try some other network like Alexnet or VGG16. I recommend you to take some online courses about deep learning, it would be helpful. I have summarised the gist <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2711\" rel=\"nofollow noreferrer\">from github<\/a>.<\/p>\n"},{"AnswerId":"53490006","CreationDate":"2018-11-26T22:21:18.577","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>There are multiple possibilities, but the problem needs some clarification. <\/p>\n\n<p>Could you specify the range of your target?<br>\n0.004 might sound low as a loss, but it's not if your target ranges from 0 to 0.0001 for example. <\/p>\n\n<p>What are the metrics of your validation &amp; test data set? Loss on itself does not say much without knowing the validation loss.   <\/p>\n\n<p>Guessing that the 0.004 is too good to be true, your model might be over fitting. \nTry implementing dropout to avoid over fitting. <\/p>\n\n<p>In case your model is not over fitting, it might be the case that Adam is overshooting a (local) minima. Try lowering its learning rate, or try sgd with custom hyper-parameters. This does take a lot of tuning.<\/p>\n\n<p>There is a free course on Coursera called Machine Learning by Stanford. This covers theory on these concepts (and more) in a good way.<\/p>\n"}]}
+{"QuestionId":53488768,"AnswerCount":1,"Tags":"<keras><time-series><conv-neural-network><recurrent-neural-network>","CreationDate":"2018-11-26T20:44:54.210","AcceptedAnswerId":null,"OwnerUserId":3482325.0,"Title":"Keras functional API: Combine CNN model with a RNN to to look at sequences of images","Body":"<p>So i was stuck with a question on how to combine a CNN with a RNN in Keras. While posting the question someone pointed me out that this is the correct way to approach the problem. Apparently i just overlooked something in the original code, which made me answer my own question. <\/p>\n\n<p>The original problem is as follows:<\/p>\n\n<p>How do you create a model in Keras that has sequences of images as the input, with a CNN 'looking' at each individual image and the sequence of the CNN output being fed into a RNN?<\/p>\n\n<p>To make it more clear:  <\/p>\n\n<p>Model one: a CNN that looks at single images.<br>\nModel two: a RNN that at the sequences of the output of the CNN from model one.<\/p>\n\n<p>So for example the CNN should see 5 images and this sequence of 5 outputs from the CNN should be passed on to the RNN. <\/p>\n\n<p>The input data is in the following format:<br>\n(number_of_images, width, height, channels) = (4000, 120, 60, 1)<\/p>\n","answers":[{"AnswerId":"53489433","CreationDate":"2018-11-26T21:36:27.310","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>The answer to this question is as follows. <\/p>\n\n<p>Take this oversimplified CNN model: <\/p>\n\n<pre><code>cnn = Sequential()\ncnn.add(Conv2D(16, (50, 50), input_shape=(120, 60, 1)))\n\ncnn.add(Conv2D(16, (40, 40)))\n\ncnn.add(Flatten()) # Not sure if this if the proper way to do this.\n<\/code><\/pre>\n\n<p>Then there is this simple RNN model: <\/p>\n\n<pre><code>rnn = Sequential()\n\nrnn = GRU(64, return_sequences=False, input_shape=(120, 60))\n<\/code><\/pre>\n\n<p>Which should be connected to a dense network:<\/p>\n\n<pre><code>dense = Sequential()\ndense.add(Dense(128))\ndense.add(Dense(64))\n\ndense.add(Dense(1)) # Model output\n<\/code><\/pre>\n\n<p>Notice that activation functions and such have been left out for readability. <\/p>\n\n<p>Now all that is left is combining these 3 main models. <\/p>\n\n<pre><code>main_input = Input(shape=(5, 120, 60, 1)) # Data has been reshaped to (800, 5, 120, 60, 1)\n\nmodel = TimeDistributed(cnn)(main_input) # this should make the cnn 'run' 5 times?\nmodel = rnn(model) # combine timedistributed cnn with rnn\nmodel = dense(model) # add dense\n<\/code><\/pre>\n\n<p>Then finally<\/p>\n\n<pre><code>final_model = Model(inputs=main_input, outputs=model)\n\nfinal_model.compile...\nfinal_model.fit...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53488870,"AnswerCount":2,"Tags":"<tensorflow><keras><amazon-sagemaker>","CreationDate":"2018-11-26T20:51:46.823","AcceptedAnswerId":53489851.0,"OwnerUserId":758836.0,"Title":"SageMaker fails when using Multi-GPU with keras.utils.multi_gpu_model","Body":"<p>Running AWS SageMaker with a custom model, the TrainingJob fails with an <strong>Algorithm Error<\/strong> when using Keras plus a Tensorflow backend in multi-gpu configuration:<\/p>\n\n<pre><code>from keras.utils import multi_gpu_model\n\nparallel_model = multi_gpu_model(model, gpus=K)\nparallel_model.compile(loss='categorical_crossentropy',\noptimizer='rmsprop')\nparallel_model.fit(x, y, epochs=20, batch_size=256)\n<\/code><\/pre>\n\n<p>This simple parallel model loading will fail. There is no further error or exception from CloudWatch logging. This configuration works properly on local machine with 2x NVIDIA GTX 1080, same Keras Tensorflow backend.<\/p>\n\n<p>According to SageMaker documentation and <a href=\"https:\/\/github.com\/awslabs\/keras-apache-mxnet\/wiki\/Multi-GPU-Model-Training-with-Keras-MXNet\" rel=\"nofollow noreferrer\">tutorials<\/a> the <code>multi_gpu_model<\/code> utility will work ok when Keras backend is MXNet, but I did not find any mention when the backend is Tensorflow with the same multi gpu configuration.<\/p>\n\n<p><strong>[UPDATE]<\/strong><\/p>\n\n<p>I have updated the code with the suggested answer below, and I'm adding some logging before the TrainingJob hangs<\/p>\n\n<p>This logging repeats twice<\/p>\n\n<pre><code>2018-11-27 10:02:49.878414: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0, 1, 2, 3\n2018-11-27 10:02:49.878462: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-27 10:02:49.878471: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988] 0 1 2 3\n2018-11-27 10:02:49.878477: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0: N Y Y Y\n2018-11-27 10:02:49.878481: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 1: Y N Y Y\n2018-11-27 10:02:49.878486: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 2: Y Y N Y\n2018-11-27 10:02:49.878492: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 3: Y Y Y N\n2018-11-27 10:02:49.879340: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/device:GPU:0 with 14874 MB memory) -&gt; physical GPU (device: 0, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:1b.0, compute capability: 7.0)\n2018-11-27 10:02:49.879486: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/device:GPU:1 with 14874 MB memory) -&gt; physical GPU (device: 1, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:1c.0, compute capability: 7.0)\n2018-11-27 10:02:49.879694: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/device:GPU:2 with 14874 MB memory) -&gt; physical GPU (device: 2, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:1d.0, compute capability: 7.0)\n2018-11-27 10:02:49.879872: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/device:GPU:3 with 14874 MB memory) -&gt; physical GPU (device: 3, name: Tesla V100-SXM2-16GB, pci bus id: 0000:00:1e.0, compute capability: 7.0)\n<\/code><\/pre>\n\n<p>Before there is some logging info about each GPU, that repeats 4 times<\/p>\n\n<pre><code>2018-11-27 10:02:46.447639: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 3 with properties:\nname: Tesla V100-SXM2-16GB major: 7 minor: 0 memoryClockRate(GHz): 1.53\npciBusID: 0000:00:1e.0\ntotalMemory: 15.78GiB freeMemory: 15.37GiB\n<\/code><\/pre>\n\n<p>According to the logging all the 4 GPUs are visible and loaded in the Tensorflow Keras backend. After that no application logging follows, the TrainingJob status is <strong>inProgress<\/strong> for a while, after that it becomes <strong>Failed<\/strong> with the same <strong>Algorithm Error<\/strong>.<\/p>\n\n<p>Looking at CloudWatch logging I can see some metrics at work. Specifically <code>GPU Memory Utilization<\/code>, <code>CPU Utilization<\/code> are ok, while <code>GPU utilization<\/code> is 0%.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/260hL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/260hL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>[UPDATE]<\/strong><\/p>\n\n<p>Due to a <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8123#issuecomment-354857044\" rel=\"nofollow noreferrer\">known<\/a> bug on Keras that is about saving a multi gpu model, I'm using this override of the <strong>multi_gpu_model<\/strong> utility in <strong>keras.utils<\/strong><\/p>\n\n<pre><code>from keras.layers import Lambda, concatenate\nfrom keras import Model\nimport tensorflow as tf\n\ndef multi_gpu_model(model, gpus):\n    #source: https:\/\/github.com\/keras-team\/keras\/issues\/8123#issuecomment-354857044\n  if isinstance(gpus, (list, tuple)):\n    num_gpus = len(gpus)\n    target_gpu_ids = gpus\n  else:\n    num_gpus = gpus\n    target_gpu_ids = range(num_gpus)\n\n  def get_slice(data, i, parts):\n    shape = tf.shape(data)\n    batch_size = shape[:1]\n    input_shape = shape[1:]\n    step = batch_size \/\/ parts\n    if i == num_gpus - 1:\n      size = batch_size - step * i\n    else:\n      size = step\n    size = tf.concat([size, input_shape], axis=0)\n    stride = tf.concat([step, input_shape * 0], axis=0)\n    start = stride * i\n    return tf.slice(data, start, size)\n\n  all_outputs = []\n  for i in range(len(model.outputs)):\n    all_outputs.append([])\n\n  # Place a copy of the model on each GPU,\n  # each getting a slice of the inputs.\n  for i, gpu_id in enumerate(target_gpu_ids):\n    with tf.device('\/gpu:%d' % gpu_id):\n      with tf.name_scope('replica_%d' % gpu_id):\n        inputs = []\n        # Retrieve a slice of the input.\n        for x in model.inputs:\n          input_shape = tuple(x.get_shape().as_list())[1:]\n          slice_i = Lambda(get_slice,\n                           output_shape=input_shape,\n                           arguments={'i': i,\n                                      'parts': num_gpus})(x)\n          inputs.append(slice_i)\n\n        # Apply model on slice\n        # (creating a model replica on the target device).\n        outputs = model(inputs)\n        if not isinstance(outputs, list):\n          outputs = [outputs]\n\n        # Save the outputs for merging back together later.\n        for o in range(len(outputs)):\n          all_outputs[o].append(outputs[o])\n\n  # Merge outputs on CPU.\n  with tf.device('\/cpu:0'):\n    merged = []\n    for name, outputs in zip(model.output_names, all_outputs):\n      merged.append(concatenate(outputs,\n                                axis=0, name=name))\n    return Model(model.inputs, merged)\n<\/code><\/pre>\n\n<p>This works ok on local <code>2x NVIDIA GTX 1080 \/ Intel Xeon \/ Ubuntu 16.04<\/code>. It will fails on SageMaker Training Job.<\/p>\n\n<p>I have posted this issue on AWS Sagemaker forum in <\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/forums.aws.amazon.com\/thread.jspa?messageID=879769\" rel=\"nofollow noreferrer\">TrainingJob custom algorithm with Keras backend and multi GPU<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/forums.aws.amazon.com\/thread.jspa?threadID=294095&amp;tstart=0\" rel=\"nofollow noreferrer\">SageMaker Fails when using Multi-GPU with\nkeras.utils.multi_gpu_model<\/a><\/p><\/li>\n<\/ul>\n\n<p><strong>[UPDATE]<\/strong><\/p>\n\n<p>I have slightly modified the <code>tf.session<\/code> code adding some initializers<\/p>\n\n<pre><code>with tf.Session() as session:\n    K.set_session(session)\n    session.run(tf.global_variables_initializer())\n    session.run(tf.tables_initializer())\n<\/code><\/pre>\n\n<p>and now at least I can see that one GPU (I assume device <code>gpu:0<\/code>) is used from the instance metrics. The multi-gpu does not work anyways.<\/p>\n","answers":[{"AnswerId":"53489851","CreationDate":"2018-11-26T22:07:03.360","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>This might not be the best answer for your problem, but this is what I am using for a multi-gpu model with Tensorflow backend. First i initialize using: <\/p>\n\n<pre><code>def setup_multi_gpus():\n    \"\"\"\n    Setup multi GPU usage\n\n    Example usage:\n    model = Sequential()\n    ...\n    multi_model = multi_gpu_model(model, gpus=num_gpu)\n    multi_model.fit()\n\n    About memory usage:\n    https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory\n    \"\"\"\n    import tensorflow as tf\n    from keras.utils.training_utils import multi_gpu_model\n    from tensorflow.python.client import device_lib\n\n    # IMPORTANT: Tells tf to not occupy a specific amount of memory\n    from keras.backend.tensorflow_backend import set_session  \n    config = tf.ConfigProto()  \n    config.gpu_options.allow_growth = True  # dynamically grow the memory used on the GPU  \n    sess = tf.Session(config=config)  \n    set_session(sess)  # set this TensorFlow session as the default session for Keras.\n\n\n    # getting the number of GPUs \n    def get_available_gpus():\n       local_device_protos = device_lib.list_local_devices()\n       return [x.name for x in local_device_protos if x.device_type    == 'GPU']\n\n    num_gpu = len(get_available_gpus())\n    print('Amount of GPUs available: %s' % num_gpu)\n\n    return num_gpu\n<\/code><\/pre>\n\n<p>Then i call<\/p>\n\n<pre><code># Setup multi GPU usage\nnum_gpu = setup_multi_gpus()\n<\/code><\/pre>\n\n<p>and create a model.<\/p>\n\n<pre><code>...\n<\/code><\/pre>\n\n<p>After which you're able to make it a multi GPU model.<\/p>\n\n<pre><code>multi_model = multi_gpu_model(model, gpus=num_gpu)\nmulti_model.compile...\nmulti_model.fit...\n<\/code><\/pre>\n\n<p>The only thing here that is different from what you are doing is the way Tensorflow is initializing the GPU's. I can't imagine it being the problem, but it might be worth trying out. <\/p>\n\n<p>Good luck! <\/p>\n\n<p>Edit: I noticed sequence to sequence not being able to work with multi GPU. Is that the type of model you are trying to train?<\/p>\n"},{"AnswerId":"53754450","CreationDate":"2018-12-13T03:01:20.703","ParentId":null,"OwnerUserId":"9090582","Title":null,"Body":"<p>I apologize for the slow response.<\/p>\n\n<p>It seems there are a lot of threads that are running in parallel, and I want to link them together, so that other individuals who have the same issue can see the progress and discussion going on.<\/p>\n\n<p><a href=\"https:\/\/forums.aws.amazon.com\/thread.jspa?messageID=881541\" rel=\"nofollow noreferrer\">https:\/\/forums.aws.amazon.com\/thread.jspa?messageID=881541<\/a>\n<a href=\"https:\/\/forums.aws.amazon.com\/thread.jspa?messageID=881540\" rel=\"nofollow noreferrer\">https:\/\/forums.aws.amazon.com\/thread.jspa?messageID=881540<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/issues\/512\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/issues\/512<\/a><\/p>\n\n<p>There a few questions in regards to this.<\/p>\n\n<p>What version of TensorFlow and Keras?<\/p>\n\n<p>I am not too sure what is causing this problem. Does your container have all of the needed dependencies such as CUDA and etc? <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a><\/p>\n\n<p>Were you able to train using single GPU with Keras?<\/p>\n"}]}
+{"QuestionId":53489253,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision>","CreationDate":"2018-11-26T21:21:10.757","AcceptedAnswerId":null,"OwnerUserId":1136018.0,"Title":"TensorFlow: How to load images asynchronously for streaming video predictions in Python","Body":"<p>I am training a model that takes a video as input and makes per-frame predictions (object tracking, to be specific).<\/p>\n\n<p>TensorFlow provides the <code>data.Dataset.prefetch()<\/code> operation to load images asynchronously during training. This works well because my training sequences are only a few frames long. However, I would also like to load images asynchronously during testing, where the sequences can be thousands of frames long.<\/p>\n\n<p>Is there any convenient way to achieve this in TensorFlow? I suppose I could define each video as a <code>Dataset<\/code> object? This seems kind of ugly because I need to create one operation in the graph for each video in my dataset?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":53489352,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2018-11-26T21:29:32.170","AcceptedAnswerId":null,"OwnerUserId":3002473.0,"Title":"Keras - get the exact training examples used in a batch in a Callback","Body":"<p>I'm having a problem training a neural network in Keras. Every epoch, the loss will decrease steadily, reaching about 1e-9, then somewhere in the middle of the epoch (could be anywhere), the loss jumps up to 5e-5 and eventually just stabilizes at the same final loss each epoch. I believe this is due to some dirty data in my dataset causing the model to fail to train beyond a certain point, though I'm really not sure.<\/p>\n\n<p>To test my hypothesis, I wanted to create a custom Keras callback object, that would determine if there's a big enough jump in the loss after a batch, and indicate which batch caused the jump. The problem is that the <code>batch<\/code> argument supplied to <code>keras.callbacks.Callback.on_batch_end<\/code> is just the batch <em>number<\/em>, and not actually the training examples <em>used<\/em> in that batch. Furthermore, the <code>logs<\/code> dict passed in as well only contains <code>loss<\/code> and <code>acc<\/code>.<\/p>\n\n<p>This means I can't actually determine which pieces of data are causing the jump in the loss. Is there a way I can determine the exact training example that is causing the jump in each epoch? Is there some way I could access it in a callback?<\/p>\n","answers":[]}
+{"QuestionId":53489606,"AnswerCount":0,"Tags":"<pandas><keras><time-series><lstm>","CreationDate":"2018-11-26T21:48:39.290","AcceptedAnswerId":null,"OwnerUserId":2791816.0,"Title":"Keras - Variable batch size for stateful LSTM","Body":"<p>I have a list of 986 arrays generated using the below line to select the Weeks only<\/p>\n\n<pre><code>weeks = [g for n, g in m15.groupby(pd.Grouper(key='datebuff',freq='W'))]\n<\/code><\/pre>\n\n<p>I need two achieve two things. First, prepare the X_train and y_train, the X_train is gonna be each array excluding window size from the end. and y_train is the shift of the window size till the end of each array. <em>still researching<\/em><\/p>\n\n<p>Second, is there a way and the most important, is there a way to train LSTM, on each array separately, like feed it ? is it going to be the stateful loop ?<\/p>\n\n<p><br><\/p>\n\n<pre><code>[               datebuff    High     Low    Open   Close\n 0   2000-01-03 00:00:00  1.0080  1.0073  1.0073  1.0077\n 1   2000-01-03 00:15:00  1.0087  1.0076  1.0078  1.0086\n 2   2000-01-03 00:30:00  1.0089  1.0079  1.0087  1.0079\n 3   2000-01-03 00:45:00  1.0132  1.0078  1.0078  1.0128\n 4   2000-01-03 01:00:00  1.0133  1.0120  1.0129  1.0122\n 5   2000-01-03 01:15:00  1.0125  1.0120  1.0123  1.0124\n 6   2000-01-03 01:30:00  1.0137  1.0129  1.0132  1.0133\n 7   2000-01-03 01:45:00  1.0141  1.0133  1.0135  1.0137\n 8   2000-01-03 02:00:00  1.0145  1.0134  1.0140  1.0138\n 9   2000-01-03 02:15:00  1.0142  1.0135  1.0135  1.0141\n 10  2000-01-03 02:30:00  1.0147  1.0137  1.0142  1.0145\n 11  2000-01-03 02:45:00  1.0173  1.0142  1.0142  1.0171\n 12  2000-01-03 03:00:00  1.0190  1.0170  1.0170  1.0185\n 13  2000-01-03 03:15:00  1.0189  1.0176  1.0188  1.0180\n 14  2000-01-03 03:30:00  1.0183  1.0173  1.0180  1.0180\n 15  2000-01-03 03:45:00  1.0180  1.0173  1.0179  1.0175\n 16  2000-01-03 04:00:00  1.0177  1.0164  1.0173  1.0171\n 17  2000-01-03 04:15:00  1.0171  1.0165  1.0167  1.0170\n 18  2000-01-03 04:30:00  1.0171  1.0166  1.0166  1.0168\n 19  2000-01-03 04:45:00  1.0171  1.0166  1.0167  1.0167\n 20  2000-01-03 05:00:00  1.0171  1.0165  1.0171  1.0165\n 21  2000-01-03 05:15:00  1.0171  1.0164  1.0171  1.0167\n 22  2000-01-03 05:30:00  1.0168  1.0165  1.0168  1.0167\n 23  2000-01-03 05:45:00  1.0172  1.0166  1.0166  1.0169\n 24  2000-01-03 06:00:00  1.0180  1.0167  1.0171  1.0178\n 25  2000-01-03 06:15:00  1.0181  1.0175  1.0181  1.0175\n 26  2000-01-03 06:30:00  1.0179  1.0172  1.0177  1.0172\n 27  2000-01-03 06:45:00  1.0177  1.0170  1.0173  1.0170\n 28  2000-01-03 07:00:00  1.0174  1.0158  1.0174  1.0163\n 29  2000-01-03 07:15:00  1.0167  1.0160  1.0163  1.0165\n ..                  ...     ...     ...     ...     ...\n 448 2000-01-07 16:30:00  1.0313  1.0299  1.0306  1.0299\n 449 2000-01-07 16:45:00  1.0307  1.0288  1.0298  1.0292\n 450 2000-01-07 17:00:00  1.0305  1.0284  1.0293  1.0304\n 451 2000-01-07 17:15:00  1.0306  1.0288  1.0304  1.0297\n 452 2000-01-07 17:30:00  1.0306  1.0294  1.0295  1.0303\n 453 2000-01-07 17:45:00  1.0306  1.0293  1.0303  1.0293\n 454 2000-01-07 18:00:00  1.0299  1.0289  1.0299  1.0293\n 455 2000-01-07 18:15:00  1.0299  1.0292  1.0295  1.0297\n 456 2000-01-07 18:30:00  1.0300  1.0280  1.0297  1.0280\n 457 2000-01-07 18:45:00  1.0302  1.0281  1.0283  1.0299\n 458 2000-01-07 19:00:00  1.0306  1.0293  1.0297  1.0293\n 459 2000-01-07 19:15:00  1.0299  1.0290  1.0293  1.0292\n 460 2000-01-07 19:30:00  1.0294  1.0284  1.0292  1.0287\n 461 2000-01-07 19:45:00  1.0293  1.0285  1.0286  1.0289\n 462 2000-01-07 20:00:00  1.0289  1.0280  1.0286  1.0283\n 463 2000-01-07 20:15:00  1.0284  1.0273  1.0279  1.0275\n 464 2000-01-07 20:30:00  1.0284  1.0275  1.0276  1.0283\n 465 2000-01-07 20:45:00  1.0289  1.0272  1.0284  1.0281\n 466 2000-01-07 21:00:00  1.0289  1.0279  1.0280  1.0281\n 467 2000-01-07 21:15:00  1.0288  1.0276  1.0277  1.0277\n 468 2000-01-07 21:30:00  1.0290  1.0276  1.0276  1.0287\n 469 2000-01-07 21:45:00  1.0298  1.0289  1.0292  1.0294\n 470 2000-01-07 22:00:00  1.0294  1.0287  1.0292  1.0291\n 471 2000-01-07 22:15:00  1.0292  1.0284  1.0290  1.0287\n 472 2000-01-07 22:30:00  1.0298  1.0289  1.0289  1.0296\n 473 2000-01-07 22:45:00  1.0299  1.0295  1.0295  1.0295\n 474 2000-01-07 23:00:00  1.0297  1.0292  1.0292  1.0294\n 475 2000-01-07 23:15:00  1.0298  1.0294  1.0298  1.0297\n 476 2000-01-07 23:30:00  1.0299  1.0299  1.0299  1.0299\n 477 2000-01-07 23:45:00  1.0300  1.0298  1.0300  1.0298\n\n [478 rows x 5 columns],\n                datebuff    High     Low    Open   Close\n 478 2000-01-10 00:00:00  1.0298  1.0286  1.0286  1.0298\n 479 2000-01-10 00:15:00  1.0300  1.0292  1.0297  1.0300\n 480 2000-01-10 00:30:00  1.0301  1.0300  1.0301  1.0300\n 481 2000-01-10 00:45:00  1.0305  1.0297  1.0305  1.0297\n 482 2000-01-10 01:00:00  1.0298  1.0289  1.0295  1.0291\n 483 2000-01-10 01:15:00  1.0302  1.0289  1.0294  1.0301\n 484 2000-01-10 01:30:00  1.0302  1.0287  1.0300  1.0291\n 485 2000-01-10 01:45:00  1.0293  1.0282  1.0290  1.0283\n 486 2000-01-10 02:00:00  1.0287  1.0274  1.0282  1.0284\n 487 2000-01-10 02:15:00  1.0292  1.0280  1.0283  1.0284\n 488 2000-01-10 02:30:00  1.0290  1.0284  1.0287  1.0286\n 489 2000-01-10 02:45:00  1.0292  1.0285  1.0287  1.0289\n 490 2000-01-10 03:00:00  1.0291  1.0283  1.0291  1.0289\n 491 2000-01-10 03:15:00  1.0292  1.0285  1.0291  1.0286\n 492 2000-01-10 03:30:00  1.0292  1.0283  1.0289  1.0289\n 493 2000-01-10 03:45:00  1.0301  1.0288  1.0290  1.0289\n 494 2000-01-10 04:00:00  1.0295  1.0289  1.0290  1.0290\n 495 2000-01-10 04:15:00  1.0298  1.0289  1.0291  1.0292\n 496 2000-01-10 04:30:00  1.0295  1.0292  1.0293  1.0293\n 497 2000-01-10 04:45:00  1.0295  1.0292  1.0294  1.0293\n 498 2000-01-10 05:00:00  1.0295  1.0292  1.0292  1.0292\n 499 2000-01-10 05:15:00  1.0294  1.0290  1.0291  1.0292\n 500 2000-01-10 05:30:00  1.0294  1.0290  1.0294  1.0292\n 501 2000-01-10 05:45:00  1.0294  1.0287  1.0294  1.0290\n 502 2000-01-10 06:00:00  1.0292  1.0287  1.0287  1.0289\n 503 2000-01-10 06:15:00  1.0293  1.0287  1.0293  1.0292\n 504 2000-01-10 06:30:00  1.0298  1.0290  1.0291  1.0292\n 505 2000-01-10 06:45:00  1.0295  1.0291  1.0293  1.0295\n 506 2000-01-10 07:00:00  1.0297  1.0289  1.0294  1.0291\n 507 2000-01-10 07:15:00  1.0298  1.0285  1.0292  1.0289\n ..                  ...     ...     ...     ...     ...\n 927 2000-01-14 16:30:00  1.0142  1.0121  1.0130  1.0136\n 928 2000-01-14 16:45:00  1.0140  1.0115  1.0135  1.0117\n 929 2000-01-14 17:00:00  1.0134  1.0114  1.0117  1.0130\n 930 2000-01-14 17:15:00  1.0138  1.0125  1.0129  1.0137\n 931 2000-01-14 17:30:00  1.0144  1.0132  1.0135  1.0144\n 932 2000-01-14 17:45:00  1.0146  1.0125  1.0142  1.0128\n 933 2000-01-14 18:00:00  1.0152  1.0126  1.0128  1.0152\n 934 2000-01-14 18:15:00  1.0156  1.0141  1.0152  1.0145\n 935 2000-01-14 18:30:00  1.0147  1.0137  1.0144  1.0139\n 936 2000-01-14 18:45:00  1.0150  1.0139  1.0139  1.0143\n 937 2000-01-14 19:00:00  1.0148  1.0133  1.0143  1.0146\n 938 2000-01-14 19:15:00  1.0151  1.0145  1.0147  1.0149\n 939 2000-01-14 19:30:00  1.0152  1.0146  1.0149  1.0150\n 940 2000-01-14 19:45:00  1.0156  1.0145  1.0149  1.0154\n 941 2000-01-14 20:00:00  1.0157  1.0140  1.0154  1.0141\n 942 2000-01-14 20:15:00  1.0147  1.0138  1.0141  1.0142\n 943 2000-01-14 20:30:00  1.0147  1.0141  1.0142  1.0145\n 944 2000-01-14 20:45:00  1.0157  1.0143  1.0146  1.0157\n 945 2000-01-14 21:00:00  1.0155  1.0144  1.0154  1.0146\n 946 2000-01-14 21:15:00  1.0147  1.0144  1.0145  1.0144\n 947 2000-01-14 21:30:00  1.0145  1.0141  1.0145  1.0142\n 948 2000-01-14 21:45:00  1.0145  1.0140  1.0141  1.0145\n 949 2000-01-14 22:00:00  1.0145  1.0135  1.0144  1.0138\n 950 2000-01-14 22:15:00  1.0140  1.0127  1.0140  1.0134\n 951 2000-01-14 22:30:00  1.0132  1.0122  1.0131  1.0124\n 952 2000-01-14 22:45:00  1.0124  1.0121  1.0122  1.0121\n 953 2000-01-14 23:00:00  1.0133  1.0122  1.0122  1.0125\n 954 2000-01-14 23:15:00  1.0128  1.0116  1.0128  1.0118\n 955 2000-01-14 23:30:00  1.0122  1.0118  1.0119  1.0119\n 956 2000-01-14 23:45:00  1.0123  1.0120  1.0121  1.0122\n\n [479 rows x 5 columns],\n                 datebuff    High     Low    Open   Close\n 957  2000-01-17 00:00:00  1.0131  1.0127  1.0129  1.0127\n 958  2000-01-17 00:15:00  1.0127  1.0124  1.0126  1.0124\n 959  2000-01-17 00:30:00  1.0124  1.0121  1.0123  1.0122\n 960  2000-01-17 00:45:00  1.0122  1.0120  1.0121  1.0122\n 961  2000-01-17 01:00:00  1.0131  1.0123  1.0123  1.0129\n 962  2000-01-17 01:15:00  1.0138  1.0130  1.0133  1.0134\n 963  2000-01-17 01:30:00  1.0137  1.0131  1.0134  1.0136\n 964  2000-01-17 01:45:00  1.0140  1.0129  1.0137  1.0129\n 965  2000-01-17 02:00:00  1.0135  1.0127  1.0128  1.0132\n 966  2000-01-17 02:15:00  1.0134  1.0131  1.0133  1.0131\n 967  2000-01-17 02:30:00  1.0135  1.0128  1.0130  1.0134\n 968  2000-01-17 02:45:00  1.0137  1.0133  1.0134  1.0135\n 969  2000-01-17 03:00:00  1.0136  1.0131  1.0133  1.0131\n 970  2000-01-17 03:15:00  1.0137  1.0130  1.0133  1.0137\n 971  2000-01-17 03:30:00  1.0140  1.0132  1.0140  1.0132\n 972  2000-01-17 03:45:00  1.0135  1.0132  1.0135  1.0134\n 973  2000-01-17 04:00:00  1.0136  1.0130  1.0136  1.0133\n 974  2000-01-17 04:15:00  1.0134  1.0132  1.0134  1.0133\n 975  2000-01-17 04:30:00  1.0135  1.0131  1.0132  1.0133\n 976  2000-01-17 04:45:00  1.0134  1.0132  1.0132  1.0132\n 977  2000-01-17 05:00:00  1.0136  1.0132  1.0134  1.0133\n 978  2000-01-17 05:15:00  1.0136  1.0132  1.0132  1.0133\n 979  2000-01-17 05:30:00  1.0136  1.0132  1.0136  1.0134\n 980  2000-01-17 05:45:00  1.0137  1.0134  1.0136  1.0134\n 981  2000-01-17 06:00:00  1.0136  1.0132  1.0134  1.0133\n 982  2000-01-17 06:15:00  1.0138  1.0130  1.0134  1.0137\n 983  2000-01-17 06:30:00  1.0142  1.0137  1.0138  1.0140\n 984  2000-01-17 06:45:00  1.0145  1.0138  1.0140  1.0141\n 985  2000-01-17 07:00:00  1.0148  1.0139  1.0142  1.0140\n 986  2000-01-17 07:15:00  1.0145  1.0134  1.0139  1.0138\n ...                  ...     ...     ...     ...     ...\n 1407 2000-01-21 16:30:00  1.0080  1.0061  1.0075  1.0074\n 1408 2000-01-21 16:45:00  1.0083  1.0066  1.0076  1.0070\n 1409 2000-01-21 17:00:00  1.0080  1.0065  1.0071  1.0068\n 1410 2000-01-21 17:15:00  1.0079  1.0065  1.0068  1.0069\n 1411 2000-01-21 17:30:00  1.0080  1.0067  1.0069  1.0071\n 1412 2000-01-21 17:45:00  1.0095  1.0070  1.0070  1.0094\n 1413 2000-01-21 18:00:00  1.0102  1.0086  1.0095  1.0089\n 1414 2000-01-21 18:15:00  1.0091  1.0079  1.0089  1.0086\n 1415 2000-01-21 18:30:00  1.0094  1.0083  1.0085  1.0085\n 1416 2000-01-21 18:45:00  1.0096  1.0085  1.0085  1.0092\n 1417 2000-01-21 19:00:00  1.0095  1.0083  1.0091  1.0087\n 1418 2000-01-21 19:15:00  1.0100  1.0085  1.0085  1.0095\n 1419 2000-01-21 19:30:00  1.0102  1.0093  1.0096  1.0101\n 1420 2000-01-21 19:45:00  1.0103  1.0096  1.0101  1.0096\n 1421 2000-01-21 20:00:00  1.0099  1.0090  1.0095  1.0097\n 1422 2000-01-21 20:15:00  1.0101  1.0093  1.0098  1.0095\n 1423 2000-01-21 20:30:00  1.0099  1.0092  1.0094  1.0095\n 1424 2000-01-21 20:45:00  1.0103  1.0093  1.0094  1.0103\n 1425 2000-01-21 21:00:00  1.0100  1.0092  1.0098  1.0096\n 1426 2000-01-21 21:15:00  1.0100  1.0092  1.0096  1.0098\n 1427 2000-01-21 21:30:00  1.0099  1.0093  1.0094  1.0093\n 1428 2000-01-21 21:45:00  1.0097  1.0088  1.0092  1.0088\n 1429 2000-01-21 22:00:00  1.0091  1.0084  1.0087  1.0088\n 1430 2000-01-21 22:15:00  1.0094  1.0086  1.0091  1.0088\n 1431 2000-01-21 22:30:00  1.0092  1.0084  1.0085  1.0085\n 1432 2000-01-21 22:45:00  1.0094  1.0086  1.0088  1.0094\n 1433 2000-01-21 23:00:00  1.0097  1.0089  1.0097  1.0091\n 1434 2000-01-21 23:15:00  1.0101  1.0090  1.0090  1.0101\n 1435 2000-01-21 23:30:00  1.0089  1.0089  1.0089  1.0089\n 1436 2000-01-21 23:45:00  1.0087  1.0087  1.0087  1.0087\n\n [480 rows x 5 columns],\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p><strong>Update 1<\/strong><br>\nfor the first point, i have tried the below code, but i keep getting 'ValueError: all the input arrays must have same number of dimensions'\n weeks is array of 986 arrays, each is around 478-480, the window size is 10<\/p>\n\n<pre><code>X = np.array([])\ny = np.array([])\nfor i in range(0, len(weeks)):\n    z = weeks[i]\n    z = np.asarray(z)\n    z = z[:,[1,2,3,4,5]]\n    X = np.append(X,[np.atleast_3d(np.array([z[index:index + window_size] for index in range(0, z.shape[0] - window_size)]))],axis=0) # The inputs to a predictor\n    y = np.append(y,[z[window_size:,[4]]],axis=0)\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<p><strong>Update 2<\/strong><br>\nFinally solved the first issue by using python's lists instead of numpy arrays.\n<br><\/p>\n\n<pre><code>window_size = 10\nX = []\ny = []\n\nfor i in range(0, len(weeks)):\n    z = weeks[i]\n    z = np.asarray(z)\n    z = z[:,[1,2,3,4,5]]\n    X.append([np.atleast_3d(np.array([z[index:index + window_size] for index in range(0, z.shape[0] - window_size)]))]) # The inputs to a predictor\n    y.append([z[window_size:,[4]]])\n<\/code><\/pre>\n\n<p><br>\nAnd as per the keras documentation, it seem like the stateful is what i am looking for:<br>\nstateful: Boolean (default False). If True, the last state for each sample at index i in a batch will be used as initial state for the sample of index i in the following batch.<\/p>\n\n<p><strong>Update 3<\/strong><br>\nNow i am trying to work with the below model - using stateful - the issue is in setting batch_input_shape, as it requires the batch size, which i need to change in every stateful iteration or it will get the below error:<\/p>\n\n<pre><code>InvalidArgumentError: Incompatible shapes: [276,1] vs. [277,1]\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\n\nneurons = 50\nbatch_size = len(X)\nwindow_size = X[0][0].shape[1]\nnb_features = X[0][0].shape[2]\n\nmodel = Sequential()\nmodel.add(LSTM(neurons, batch_input_shape=(len(X[0][0]), window_size, nb_features), stateful=True))\nmodel.add(Dense(1))\nmodel.compile(loss='mae', optimizer='adam', metrics=['mae'])\n\n\nfor i in range(batch_size):\n    model.fit(X[i], y[i], epochs=1, batch_size=len(X[i][0]), verbose=0, shuffle=False)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53490450,"AnswerCount":0,"Tags":"<python><keras><tensor><tensorflow-datasets>","CreationDate":"2018-11-26T23:05:45.747","AcceptedAnswerId":null,"OwnerUserId":10708757.0,"Title":"What is causing keras value error using tensorflow dataset?","Body":"<p>I have been trying to run a neural net model using Keras on some .tfrecord files I have already generated. To do this I am passing them in as command line arguments and storing in a tensorflow dataset which I am then using to fit the model. However when I run the code I get the following error: <code>ValueError: Please provide either inputs and targets or inputs, targets, and sample_weights<\/code>. It seems like Keras is angry I am not passing separate input and label tensors but I have been led to believe you can use the dataset as a single argument instead? The code is shown below:<\/p>\n\n<pre><code>import tensorflow as tf\nimport sys\nimport tensorflow.data\nfrom tensorflow import keras\nfrom tensorflow.keras import layers\n\ntf.enable_eager_execution()\n\ninputList = []\nfor file in sys.argv[0:]:\n    inputList.append(file)    \nfilenames = tf.Variable(inputList, tf.string)\ndataset = tf.data.TFRecordDataset(filenames)\ndataset.shuffle(1600000)\nmodel = tf.keras.Sequential()\nmodel.add(layers.Dense(13, input_shape=(13,), activation='relu'))\nmodel.add(layers.Dense(20, activation='relu'))\nmodel.add(layers.Dense(20, activation='relu'))\nmodel.add(layers.Dense(10, activation='relu'))\nmodel.add(layers.Dense(2, activation='relu'))\nmodel.compile(optimizer=tf.train.AdamOptimizer(0.001), loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(dataset, epochs=10, steps_per_epoch=30)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53490609,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-26T23:24:12.593","AcceptedAnswerId":null,"OwnerUserId":10458815.0,"Title":"Propagate value from nested class to base class","Body":"<p>I am working on a TensorFlow based project where the goal is to be very abstract in implementation.<\/p>\n\n<p>I have reached a problem however with adding GoogleNet to my project.\nAs opposed to most other networks I have implemented thus far (alexnet, vggnet, etc), GoogleNet is different in that it uses an \"Auxiliary\" classifier at two points in the network. This means that rather than being a straight sequential pass through the network, I need to diverge and remember data at these \"Auxiliary\" points for use later.<\/p>\n\n<p>The problem arises because I use the class <code>tf.keras.Sequential<\/code> specifically because it makes the project much easier to read and debug. I would really prefer to keep using this class for those reasons, however its behaviour basically prevents the \"Auxiliary\" classifier from doing what it needs to do. This is because the output of a given layer wrapped in <code>tf.keras.Sequential<\/code> is passed as the input to the following wrapped layer. <\/p>\n\n<p>Since \"Auxiliary\" is a class in my implementation, this means I can only return one item from its call function at a time, as to not break <code>Sequential<\/code>. And since it is a diverging classifier, I currently return its unaltered input, as to not mess with the following layers. Thus it currently performs the necessary calculations, but I have no where to store the output.<\/p>\n\n<p>I have thought about this problem for a while, and came up with several solutions that involve restructuring my code, such as putting everything in one file, removing <code>tf.keras.Sequential<\/code>, or ignoring the \"Auxiliary\" altogether since it works without it, but I have not been able to figure out a way to keep my structure and perform this variable update.<\/p>\n\n<p>The current structure of the project is as follows:<\/p>\n\n<pre><code>cnn.py\n______________________________\nfrom models.py import googlenet\nclass CNN() -&gt; calls function googlenet() inside __init__()\n\n    models.py\n    _______________________________\n    from layers.py import Auxiliary\n    def googlenet() -&gt; invokes class Auxiliary()\n\n        layers.py\n        ________________\n        class Auxiliary() -&gt; performs calculations\n                          -&gt; needs to propagate value to CNN() class\n                          -&gt; has to return unaltered input\n<\/code><\/pre>\n\n<p>I have tried every way I could think of to import \"CNN\" into \"Auxiliary\", each causing an import error. I have tried to pass the \"CNN\" class instance itself to the \"Auxiliary\" class to no avail, since I am passing the class inside the constructor of said class. And I have tried to make the \"Auxiliary\" class inherit from the \"CNN\" class, even though that doesn't make sense from a design perspective.<\/p>\n\n<p>So how do I propagate a result from one class to another, which are both in separate files, without using a return statement or creating a circular dependency?<\/p>\n","answers":[]}
+{"QuestionId":53491293,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><linear-regression>","CreationDate":"2018-11-27T00:57:03.070","AcceptedAnswerId":null,"OwnerUserId":7906378.0,"Title":"Linear regression and Estimator returning wrong loses. Tensorflow","Body":"<p>I implemented this model using Keras, and the result was as expected. Now im trying with Tensorflow and I just can't get it right.\nAs you can see at bellow my loss is just not right.<\/p>\n\n<p>what am I doing wrong here?<\/p>\n\n<p>ps: I prefer to use estimators instead of multiply tensors and etc.<\/p>\n\n<pre><code>X = numpy.array([ 1.1, 1.3, 1.5, 2.0, 2.2, 2.9, 3.0, 3.2, 3.2, 3.7, 3.9, 4.0, 4.0, 4.1, 4.5, 4.9, 5.1, 5.3, 5.9, 6.0, 6.8, 7.1, 7.9, 8.2, 8.7, 9.0, 9.5, 9.6, 10.3, 10.5])\n\ny = numpy.array([ 39.343, 46.205, 37.731, 43.525, 39.891, 56.642, 60.15, 54.445, 64.445, 57.189, 63.218, 55.794, 56.957, 57.081, 61.111, 67.938, 66.029, 83.088, 81.363, 93.94, 91.738, 98.273, 101.302, 113.812, 109.431, 105.582, 116.969, 112.635, 122.391, 121.872])\n\n\n#reduce salaries to unit of thousands\n#Split 70% training, 30% test\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=1)\n\n#Create estimator\nfeat_cols = [ tf.feature_column.numeric_column('X', shape=[1]) ]\nestimator = tf.estimator.LinearRegressor(feature_columns=feat_cols)\n\n#input functions\ntrain_input_func = tf.estimator.inputs.numpy_input_fn({'X': X_train}, y_train, shuffle=False)\ntest_input_func = tf.estimator.inputs.numpy_input_fn({'X': X_test}, y_test, shuffle=False)\n\n#Train and test\nestimator.train(input_fn=train_input_func)\ntrain_metrics = estimator.evaluate(input_fn=train_input_func)\ntest_metrics = estimator.evaluate(input_fn=test_input_func)\n\n#Predict salary for arbitrary years of experience\nX_single_data = np.array([4.6])\npred_input_func = tf.estimator.inputs.numpy_input_fn({'X': X_single_data}, shuffle=False)\nsingle_pred = estimator.predict(pred_input_func)\n\nprint('--Train metrics--')\nprint(train_metrics)\nprint(' ')\nprint('--Test metrics--')\nprint(test_metrics)\n<\/code><\/pre>\n\n<blockquote>\n  <p>--Train metrics--<\/p>\n  \n  <p>{'average_loss': 5795.477, 'label\/mean': 72.32367, 'loss': 121705.016, 'prediction\/mean': 1.2057142, 'global_step': 1}<\/p>\n  \n  <p>--Test metrics--<\/p>\n<\/blockquote>\n\n<p>{'average_loss': 7422.221, 'label\/mean': 84.588104, 'loss': 66799.99, 'prediction\/mean': 1.3955557, 'global_step': 1}<\/p>\n\n<p>FYI:\nThis is what I got with keras:\n<a href=\"https:\/\/www.kaggleusercontent.com\/kf\/7701684\/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..GXmz9TfZ6ggQ0VAkb_PQhQ.qiRSLhMrkEV3vbSOm7TNallAudgtF2YILb8Ue8CrMfUFqzRgWrkyt78R21aqvP34qz6hHamz3P3uTUxeq7gs4CeEswfgpERrZMvFiX1-usm8kvkmA1XacQ2QrtHtjjwUvRKyo1WyD-2_oD2l2Kq_ZMhoxiG_BuDn4PVKnDW9fbnKeTC6bwCIVq0ySeUiJlaG.sOzBaCcqc73AYv9Z1WW_xg\/__results___files\/__results___10_1.png\" rel=\"nofollow noreferrer\">Link to image<\/a><\/p>\n","answers":[{"AnswerId":"53529933","CreationDate":"2018-11-28T23:59:56.807","ParentId":null,"OwnerUserId":"6560413","Title":null,"Body":"<p>You are missing the number of epochs to train together with a batch size. Add these parameters to your definition of the input functions, e.g., in case of train input for linear regression it may look like this<\/p>\n\n<pre><code>train_input_func = \\\n    tf.estimator.inputs.numpy_input_fn({'X': X_train},\\\n                                       y_train,\\\n                                       num_epochs=500,\\\n                                       batch_size=1,\\\n                                       shuffle=False)\n<\/code><\/pre>\n\n<p>After that it should start training and loss function will go down.<\/p>\n"}]}
+{"QuestionId":53491527,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T01:36:08.820","AcceptedAnswerId":53493097.0,"OwnerUserId":5869383.0,"Title":"How to save model in .pb format and then load it for inference in Tensorflow?","Body":"<p>I am newbie of Tensorflow and trying to run one tutorial code located in <a href=\"https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/02_Convolutional_Neural_Network.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/02_Convolutional_Neural_Network.ipynb<\/a><\/p>\n\n<p>Based on this code, I would like to try to save the model in .pb format using <code>simple_save<\/code> and restore it for testing but I have no idea how to modify this piece of code. I have browsed some web pages but still didn't get the idea. Can anyone help me change this piece of code so that I can save the trained model and then load it for inference? Thank you!<\/p>\n","answers":[{"AnswerId":"53493097","CreationDate":"2018-11-27T05:08:38.917","ParentId":null,"OwnerUserId":"3786711","Title":null,"Body":"<p>For saving model, you need two things - input and output tensor names. In your case, input tensor is called <code>x<\/code> and output tensor is <code>y_pred<\/code> and <code>y_pred_cls<\/code> (mentioned in <code>In [29]<\/code> in <a href=\"https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/02_Convolutional_Neural_Network.ipynb\" rel=\"nofollow noreferrer\">notebook<\/a>). Here's simple example to save your model:<\/p>\n\n<pre><code>simple_save(session,\n            export_dir,\n            inputs={\"x\": x,},\n            outputs={\"y_pred\": y_pred,\n                     \"y_pred_cls\": y_pred_class})\n<\/code><\/pre>\n\n<p>EDIT:\nRestoring-<\/p>\n\n<pre><code>restoring_graph = tf.Graph()\nwith restoring_graph.as_default():\n    with tf.Session(graph=restoring_graph) as sess:\n       # Restore saved values\n       tf.saved_model.loader.load(\n          sess,\n          [tag_constants.TRAINING],\n          export_dir  # Path to SavedModel\n       )\n      # Pass inputs to model and do predictions below\n<\/code><\/pre>\n"}]}
+{"QuestionId":53491595,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-27T01:45:21.983","AcceptedAnswerId":null,"OwnerUserId":7667138.0,"Title":"Prebuilt binaries of TensorFlow for processors without AVX instruction support","Body":"<p>i have just spent a day just building tensorflow from source code, but have not have any luck. Is there any prebuilt binaries for old processors available. I am runnning Windows 10(64bit) , Cuda 10 and Cudnn7.<\/p>\n\n<p>The processor i am using is an Intel W3680 and GPU is Nvidia Gtx 1070.\nI have a university project, for that i need tensorflow. Pls guide towards something...\nOh and btw i am using Anaconda.\nThanks.<\/p>\n","answers":[]}
+{"QuestionId":53492000,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-27T02:43:29.100","AcceptedAnswerId":null,"OwnerUserId":10360845.0,"Title":"How to create distributed cluster environment for Distributed Tensorflow?","Body":"<p>I have went through documentation of Tensorflow Distributed Environment, But I cant find anywhere how to setup a cluster suitable to Tensorflow.<\/p>\n\n<p>Currently I have 10 number of servers where I want to start training a model, But I would like to know the dependencies that I need to install on all the cluster servers except tensorflow library. <\/p>\n\n<p>OR <\/p>\n\n<p>Is it enough if I specify the cluster specifications in the python code that I write for distributed training with the IP address of workers and parameter servers Or is there anything that I am missing?<\/p>\n\n<pre><code>tf.train.ClusterSpec({\n\"worker\": [\n    \"172.09.09.09:1290\",\n    \"172.10.10.10:2040\",\n    \"172.11.11.11:3060\"\n],\n\"ps\": [\n    \"172.12.12.12:8080\",\n    \"172.13.13.13:8080\"\n]})\n<\/code><\/pre>\n\n<p>I can find everything else in the tensorflow documentation. Should I build a cluster similar to apache spark?<\/p>\n\n<p>Please help for the right documented flow of setup for tensorflow cluster or if I  am missing something help me with the simple answer you can.<\/p>\n","answers":[{"AnswerId":"53492082","CreationDate":"2018-11-27T02:57:13.380","ParentId":null,"OwnerUserId":"7343117","Title":null,"Body":"<p>Perhaps you can create a cluster and install Cloudera Manager on it, then one can try to install tensorflow library on it and use Spark to perform operations on your data:<\/p>\n\n<p><a href=\"https:\/\/www.cloudera.com\/documentation\/enterprise\/5-9-x\/topics\/install_cluster_template.html\" rel=\"nofollow noreferrer\">https:\/\/www.cloudera.com\/documentation\/enterprise\/5-9-x\/topics\/install_cluster_template.html<\/a><\/p>\n\n<p>Once you have your cluster you can do as you were initially doing and run tensorflow on Hadoop:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/deploy\/hadoop\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/deploy\/hadoop<\/a> <\/p>\n"},{"AnswerId":"55240078","CreationDate":"2019-03-19T11:34:44.263","ParentId":null,"OwnerUserId":"11225970","Title":null,"Body":"<p>You just need to run one script on each server. With same code. Each script gets some input as arguments to setup itself. You can have a look here: <a href=\"https:\/\/github.com\/tensorflow\/examples\/blob\/master\/community\/en\/docs\/deploy\/distributed.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/examples\/blob\/master\/community\/en\/docs\/deploy\/distributed.md<\/a><\/p>\n"}]}
+{"QuestionId":53492018,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-27T02:47:06.423","AcceptedAnswerId":null,"OwnerUserId":5171213.0,"Title":"Keras fit_generator: input ValueError","Body":"<p>I have already tried changing the list to a tuple, from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2568\" rel=\"nofollow noreferrer\">here<\/a>, and I have read <a href=\"https:\/\/stackoverflow.com\/questions\/50240768\/input-shape-error-when-using-keras-functional-api-with-fit-generator\">this<\/a>, but it wasn't helpful. <\/p>\n\n<p>The code below is based on <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8078\" rel=\"nofollow noreferrer\">this<\/a>. <\/p>\n\n<pre><code>from __future__ import division\nfrom __future__ import print_function\n\nfrom pandas import concat\nfrom pandas import DataFrame\n\nimport keras\nimport numpy as np\nimport tensorflow as tf\nimport pandas as pd\n\nfrom keras.models import Model\nfrom keras.layers import Dense, Input, LSTM, Dropout\nfrom keras.optimizers import Adam\nfrom keras.losses import binary_crossentropy\n\n#https:\/\/machinelearningmastery.com\/multivariate-time-series-forecasting-lstms-keras\/\ndef series_to_supervised(data, n_lags=1, n_out=1, dropnan=True):\n    n_vars = 1 if type(data) is list else data.shape[1]\n    df = DataFrame(data)\n    cols, names = list(), list()\n    #input sequence t-n, ..., t-1\n    for i in range(n_lags, 0, -1): #for i in 3 to 0 not including 0\n        cols.append(df.shift(i))\n        names += [('var%d(t-%d)' % (j+1, i)) for j in range (n_vars)]\n    #forecast sequence t, t+1, ..., t+n\n    for i in range(0, n_out):\n        cols.append(df.shift(-i))\n        if i==0:\n            names += [('var%d(t)' % (j+1)) for j in range(n_vars)]\n        else:\n            names += [('var%d(t+%d)' % (j+1, i)) for j in range(n_vars)]\n    #put it all together\n    agg = concat(cols, axis=1)\n    agg.columns = names\n    #drop rows with NaN values\n    if dropnan:\n        agg.dropna(inplace=True)\n    #print('agg: ', agg)\n    return agg\n\ndef data_gen(sample_IDs, batch_size, nn, num_features, n_lags, n_vars):\n\n    while True:\n\n        in_1 = np.empty((batch_size, n_vars), dtype=float)\n        in_2 = np.empty((batch_size, int(n_vars*num_features)),  dtype=int)\n\n        for count in range(batch_size):\n            in_1[count,] = np.arange(n_vars)\n            in_2[count,] = np.arange(int(n_vars*num_features))\n            count = count + 1\n\n        in_1_s2s = np.asarray(series_to_supervised(in_1, n_lags))\n\n        in_1_data = in_1_s2s[:, :int(n_vars*n_lags)] #expect shape (29, 20*3)\n        target = in_1_s2s[:, int(n_vars*n_lags):] #expect shape (29, 20)\n\n        in_1_data = in_1_data.reshape((in_1_data.shape[0], n_lags, n_vars))\n\n        #Replicate in_2 over time.\n        in_2 = in_2[n_lags:, :] #to match in_1_data shape; expect shape (29, 100)\n        in_2 = np.repeat(in_2, n_lags, axis=0)\n        in_2 = in_2.reshape(((batch_size-n_lags), n_lags, n_vars*num_features))\n\n        yield (np.asarray(in_1), np.asarray(in_2), np.asarray(target))\n\nif __name__=='__main__':\n    nn = 10\n    n_vars = 20\n    num_features = 5\n    n_lags = 3\n    batch_size = 32\n\n    sample_IDs = np.arange(1, 34) #1-32 inclusive and 2-33 inclusive; 2 batches?\n\n    steps_val = int(np.ceil(sample_IDs.shape[0]\/batch_size))\n\n    my_gen = data_gen(sample_IDs, batch_size, nn, num_features, n_lags, n_vars)\n\n    #Build a toy model\n    adj_input = Input(shape=(n_lags, n_vars))\n    feat_input = Input(shape=(n_lags, num_features*n_vars))\n    y_input = Input(shape=(n_lags, n_vars))\n\n    concat_1 = keras.layers.concatenate([adj_input, feat_input])\n    dense_1 = Dense(4, activation='relu')(concat_1)\n    enc_1 = LSTM(4, activation='relu')(dense_1)\n    drop_1 = Dropout(0.2)(enc_1)\n\n    y_dense = Dense(4, name='y_dense')(y_input)\n\n    model = Model([adj_input, feat_input, y_input], [drop_1, y_dense])\n\n    print(model.summary())\n\n    #Compile and fit_generator\n    model.compile(optimizer=Adam(lr=0.001), loss=binary_crossentropy)\n    model.fit_generator(my_gen, steps_per_epoch=steps_val, shuffle=False, epochs=2, verbose=2)\n<\/code><\/pre>\n\n<p>After printing out the model summary, it gives the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"Nov26SO.py\", line 113, in &lt;module&gt;\n    model.fit_generator(my_gen, steps_per_epoch=steps_val, shuffle=False, epochs=2, verbose=2)\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1418, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training_generator.py\", line 217, in fit_generator\n    class_weight=class_weight)\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1211, in train_on_batch\n    class_weight=class_weight)\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 751, in _standardize_user_data\n    exception_prefix='input')\n  File \"\/user\/pkgs\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training_utils.py\", line 102, in standardize_input_data\n    str(len(data)) + ' arrays: ' + str(data)[:200] + '...')\nValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 3 array(s), but instead got the following list of 1 arrays: [array([[ 0.,  1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9., 10., 11., 12.,\n        13., 14., 15., 16., 17., 18., 19.],\n       [ 0.,  1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9., 10., 11., 12.,\n        1...\n<\/code><\/pre>\n\n<p>Question: why am I not passing three numpy arrays into the model? I thought the yield statement was yielding three numpy arrays.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>This is not the same as the question <a href=\"https:\/\/stackoverflow.com\/questions\/53474497\/what-should-the-generator-return-if-it-is-used-in-a-multi-model-functional-api\">here<\/a>, although I appreciate the pointer. The problem I'm having is that I have three inputs: in_1, in_2, and targets. When I change my <code>yield<\/code> line to be <code>yield [np.asarray(in_1), np.asarray(in_2), np.asarray(target)]<\/code>, I get exactly the same problem.<\/p>\n","answers":[]}
+{"QuestionId":53492060,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-27T02:54:12.663","AcceptedAnswerId":null,"OwnerUserId":9511772.0,"Title":"TensorBoard Callbacks in Keras Backend Function","Body":"<p>In order to get training and graph information of a CNN in Keras, I simply define <code>tensorboard = TensorBoard(log_dir=\"CNN\/Logs\/{}\".format(MODEL_ID))<\/code> and then add <code>callbacks=[tensorboard]<\/code> to the <code>fit()<\/code> method of my model. <\/p>\n\n<p>I am implementing a Policy Network following <a href=\"https:\/\/gist.github.com\/kkweon\/c8d1caabaf7b43317bc8825c226045d2\" rel=\"nofollow noreferrer\">this<\/a> example, and am wondering now how I can get training and graph information of the PN in Tensorboard. <\/p>\n\n<p>Possibly there is a callbacks equivalent in Keras Backend's <code>function()<\/code>, I have not had any hope with the documentation. <\/p>\n\n<p>Specifically, my equivalent of <code>fit()<\/code> looks as follows:<\/p>\n\n<pre><code>def build_train_fn(self):\n    \"\"\" Train function for the Policy Network.\n\n    This function replaces model.fit(X, y).\n    \"\"\"\n\n    with self.graph.as_default():\n\n        K.set_session(self.session)\n\n        action_prob_placeholder = self.model.output\n        action_onehot_placeholder = K.placeholder(shape=(None, ACTIONS), name=\"action_onehot\")\n        discount_reward_placeholder = K.placeholder(shape=(None,), name=\"discount_reward\")\n\n        action_prob = K.sum(action_prob_placeholder * action_onehot_placeholder, axis=1)\n        log_action_prob = K.log(action_prob)\n\n        loss = - log_action_prob * discount_reward_placeholder\n        loss = K.mean(loss)\n\n        adam = optimizers.Adam()\n\n        updates = adam.get_updates(params=self.model.trainable_weights,\n                                   loss=loss)\n\n        self.train_fn = K.function(inputs=[self.model.input,\n                                           action_onehot_placeholder,\n                                           discount_reward_placeholder],\n                                   outputs=[],\n                                   updates=updates)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53492199,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-11-27T03:14:46.443","AcceptedAnswerId":53493615.0,"OwnerUserId":3101671.0,"Title":"How to concat a scalar to a 1D vector in Keras?","Body":"<p>I have a Keras layer of Shape (None, 8) and I would like to append a single scalar (value = 1) to the end of the Tensor. However I haven't been successful.<\/p>\n\n<p>Here is my code (simplified):<\/p>\n\n<pre><code>print(layers)\n# Tensor(\"feature_layer_2_89\/Relu:0\", shape=(?, 8), dtype=float32)\npad_tensor = tf.constant([1.0])\nconcat = concatenate([layers, pad_tensor])\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 8), (1,)]\n<\/code><\/pre>\n\n<p>EDIT:\nBasically I want to append a scalar (with the value of 1) to a vector. For example given a vector (1D Tensor) <code>[1 3 3 0 2]<\/code>, how to produce <code>[1 3 3 0 2 1]<\/code>, and I don't want to mess with the batch size which is <code>None<\/code> here.<\/p>\n","answers":[{"AnswerId":"53493615","CreationDate":"2018-11-27T06:01:26.663","ParentId":null,"OwnerUserId":"3101671","Title":null,"Body":"<p>You could create the <code>pad_tensor<\/code> such that it has the rank 2 as @Psidom suggested. So first we need to get the <code>batch_size<\/code> as follows:<\/p>\n\n<pre><code>batch_size = tf.shape(layers)[0]\npadding_tensor = tf.ones([batch_size, 1])\n<\/code><\/pre>\n\n<p>Now we can use <code>concatenate<\/code> function to concat  the two Tensors:<\/p>\n\n<pre><code>concat = concatenate([layers, padding_tensor])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53492970,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2018-11-27T04:53:53.867","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"Python - Keras Model doesnt converge","Body":"<p>I have a network with <code>32<\/code> input nodes, <code>20<\/code> hidden nodes and <code>65<\/code> output nodes. My network input actually is a hash code of length <code>32<\/code> and the output is the word.\nThe input is the ascii value of each character of the Hash code. The output of the network is a binary representation I have made. Say for example <code>a<\/code> is equal to <code>00000<\/code> and <code>b<\/code> is equal to <code>00001<\/code> and so on and so forth. It only includes the alphabet and the space that why it's only <code>5<\/code> bits per character. I have a maximum limit of only <code>13<\/code> characters in my training input, so my output nodes is <code>13 * 5 = 65<\/code>. And Im expecting a binary output like <code>10101010101010101010101010101010101010101010101010101010101001011<\/code> . The bit sequence can predict at most 16 characters word given a hash code of <code>32<\/code> length as an input. Below is my current code:<\/p>\n\n<pre><code>scaler = MinMaxScaler(feature_range=(0,1))\nscaled_train_samples = scaler.fit_transform((train_samples).reshape(-1, 32))\ntrain_labels = train_labels.reshape(-1, 65)\n\n\nmodel = Sequential([\n    Dense(32, input_shape=(32,), activation = 'sigmoid'),\n    BatchNormalization(),\n    Dense(25, activation='tanh'),\n    BatchNormalization(),\n    Dense(65, input_shape=(65,), activation='sigmoid')\n])\n\noverfitCallback = EarlyStopping(monitor='loss', min_delta=0, patience = 1000)\n\nmodel.summary()\nmodel.compile(SGD(lr=.01, decay=1e-6,  momentum=0.9),     loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(train_samples, train_labels, batch_size=1000, epochs=1000000, callbacks=[overfitCallback], shuffle = True, verbose=2)\n<\/code><\/pre>\n\n<p>I plan to overfit the model, so that it can memorize all the hash codes of the words in the dictionary. As an initial, my training samples is only <code>5,000<\/code> something. I just wanted to see if it will learn from a small dataset. How will I make network converge faster? I think its running more than one hour, and its loss function is still .5004 something and the accuracy is .7301. It gets up and down but when I check every 10 minutes or so, I can see only alittle improvement. How will I fine tune it?<\/p>\n\n<p>UPDATE :<\/p>\n\n<p>The training had already stopped but it didn't converge. It's loss is <code>.4614<\/code> and accuracy is <code>.7422<\/code><\/p>\n","answers":[{"AnswerId":"53506006","CreationDate":"2018-11-27T18:30:02.313","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>There are some hyper parameters that i would suggest to change first.  <\/p>\n\n<p>Try <code>'relu'<\/code> or <code>LeakyReLU()<\/code> as the activation function for the non-output layers. Basically <code>relu<\/code> is the standard activation function for baseline models.  <\/p>\n\n<p>The standard optimizer (for most cases) currently is Adam, try using this. Tweak its learning rate when needed. You could get better results with sgd, but it often takes a lot of epochs and a lot of hyper parameter tuning. Adam is basically the quickest (in general) optimizer to reach a 'low' loss. <\/p>\n\n<p>To prevent overfitting you might also want to implement <code>Dropout(0.5)<\/code>, where the 0.5 is as an example. <\/p>\n\n<p>Once you have reached the lowest loss, you might start changing these hyper parameters even more, to try and egt a lower loss. <\/p>\n\n<p>Apart from this, the first thing i actually suggest is trying and add multiple hidden layers with different sizes. This might have a way larger impact then trying to optimize all the hyper parameters. <\/p>\n\n<p>Edit: Maybe you could post a screenshot of your training loss vs epochs for the train &amp; val data? This might make things more clear for others.<\/p>\n"}]}
+{"QuestionId":53493296,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><keras><cudnn>","CreationDate":"2018-11-27T05:30:51.963","AcceptedAnswerId":53499317.0,"OwnerUserId":2290436.0,"Title":"OpenCV DNN fails with keras DenseNet121","Body":"<p>following <a href=\"http:\/\/answers.opencv.org\/question\/183507\/opencv-dnn-import-error-for-keras-pretrained-vgg16-model\/\" rel=\"nofollow noreferrer\">http:\/\/answers.opencv.org\/question\/183507\/opencv-dnn-import-error-for-keras-pretrained-vgg16-model\/<\/a> I'm trying to get densenet to work on openCV DNN but getting:<\/p>\n\n<blockquote>\n  <p>\"error: OpenCV(3.4.2)\n  \/io\/opencv\/modules\/dnn\/src\/tensorflow\/tf_graph_simplifier.cpp:712:\n  error: (-2:Unspecified error) Tensor's\n  data type is not supported in function\n  'getTensorContent'\"<\/p>\n<\/blockquote>\n\n<pre><code>import numpy as np\nfrom keras import applications\nfrom keras import backend as K\nimport cv2 as cv\nimport tensorflow as tf\n\nmodel = applications.densenet.DenseNet121(input_shape=(224, 224, 3), weights='imagenet', include_top=True)\nsess = K.get_session()\n\nprint(model.input, model.outputs)\n## Tensor(\"input_1:0\", shape=(?, 224, 224, 3), dtype=float32) [&lt;tf.Tensor 'fc1000\/Softmax:0' shape=(?, 1000) dtype=float32&gt;]\n\nfrom tensorflow.python.tools import freeze_graph\nfrom tensorflow.python.tools import optimize_for_inference_lib\n\nMODEL_PATH = 'out'\nMODEL_NAME = 'test'\ninput_node_name = 'input_1'\noutput_node_name = 'fc1000\/Softmax'\n!rm -rf {MODEL_PATH}\/\n\ntf.train.write_graph(sess.graph_def, MODEL_PATH, f'{MODEL_NAME}_graph.pb', as_text=False)\ntf.train.write_graph(sess.graph_def, MODEL_PATH, f'{MODEL_NAME}_graph.pbtxt')\ntf.train.Saver().save(sess, f'{MODEL_PATH}\/{MODEL_NAME}.chkp')\n\nfreeze_graph.freeze_graph(f'{MODEL_PATH}\/{MODEL_NAME}_graph.pbtxt',\n                          None, False,\n                          f'{MODEL_PATH}\/{MODEL_NAME}.chkp',\n                          output_node_name,\n                          \"save\/restore_all\",\n                          \"save\/Const:0\",\n                          f'{MODEL_PATH}\/frozen_{MODEL_NAME}.pb',\n                          True, \"\")\n\ngraph_def = tf.GraphDef()\nwith tf.gfile.Open(f'{MODEL_PATH}\/frozen_{MODEL_NAME}.pb', \"rb\") as f:\n    graph_def.ParseFromString(f.read())\n\noutput_graph_def = optimize_for_inference_lib.optimize_for_inference(\n    graph_def, [input_node_name], [output_node_name], tf.float32.as_datatype_enum)\n\nwith tf.gfile.GFile(f'{MODEL_PATH}\/opt_{MODEL_NAME}.pb', \"wb\") as f:\n    f.write(output_graph_def.SerializeToString())\n\n# Strip Const nodes.\nfor i in reversed(range(len(graph_def.node))):\n    if graph_def.node[i].op == 'Const':\n        del graph_def.node[i]\n#     for attr in ['T', 'data_format', 'Tshape', 'N', 'Tidx', 'Tdim',\n#                  'use_cudnn_on_gpu', 'Index', 'Tperm', 'is_training',\n#                  'Tpaddings']:\n#         if attr in graph_def.node[i].attr:\n#             del graph_def.node[i].attr[attr]\n\n# Save stripped model.\ntf.train.write_graph(graph_def, \"\", f'{MODEL_PATH}\/stripped_{MODEL_NAME}.pbtxt', as_text=True)\n\nnet = cv.dnn.readNetFromTensorflow(f'{MODEL_PATH}\/opt_{MODEL_NAME}.pb', f'{MODEL_PATH}\/stripped_{MODEL_NAME}.pbtxt')\n## error: OpenCV(3.4.2) \/io\/opencv\/modules\/dnn\/src\/tensorflow\/tf_graph_simplifier.cpp:712: error: (-2:Unspecified error) Tensor's data type is not supported in function 'getTensorContent'\n<\/code><\/pre>\n\n<p>It's hard to tell which node is causing that from the error. \nAny idea please? \nCheers<\/p>\n\n<pre>\ntensorflow 1.12.0\nopencv 3.4.3\n<\/pre>\n","answers":[{"AnswerId":"53499317","CreationDate":"2018-11-27T12:06:03.597","ParentId":null,"OwnerUserId":"2290436","Title":null,"Body":"<p>Although, I still wasn't able to get optimize_for_inference to work due to the FusedBatchNorm, thanks for the feedback from @dkurt and <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6775\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/6775<\/a>  which explains the keras learning_phase. <strong>You have to set the learning phase before loading the model!<\/strong><\/p>\n\n<p>load the model with set_learning_phase(0):<\/p>\n\n<pre><code>import numpy as np\nfrom keras import applications\nfrom keras import backend as K\nimport tensorflow as tf\n\nK.set_learning_phase(0)  ##\nmodel = applications.densenet.DenseNet121(input_shape=(224, 224, 3), weights='imagenet', include_top=True)\nsess = K.get_session()\n\nprint(model.input, model.outputs)\n## Tensor(\"input_1:0\", shape=(?, 224, 224, 3), dtype=float32) [&lt;tf.Tensor 'fc1000\/Softmax:0' shape=(?, 1000) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>freeze it:<\/p>\n\n<pre><code>from tensorflow.python.tools import freeze_graph\nfrom tensorflow.python.tools import optimize_for_inference_lib\n\nMODEL_PATH = 'out'\nMODEL_NAME = 'test'\ninput_node_name = 'input_1'\noutput_node_name = 'fc1000\/Softmax'\n!rm -rf {MODEL_PATH}\/\n\ntf.train.write_graph(sess.graph_def, MODEL_PATH, f'{MODEL_NAME}_graph.pb', as_text=False)\ntf.train.write_graph(sess.graph_def, MODEL_PATH, f'{MODEL_NAME}_graph.pbtxt')\ntf.train.Saver().save(sess, f'{MODEL_PATH}\/{MODEL_NAME}.chkp')\n\nfreeze_graph.freeze_graph(f'{MODEL_PATH}\/{MODEL_NAME}_graph.pbtxt',\n                          None, False,\n                          f'{MODEL_PATH}\/{MODEL_NAME}.chkp',\n                          output_node_name,\n                          \"save\/restore_all\",\n                          \"save\/Const:0\",\n                          f'{MODEL_PATH}\/frozen_{MODEL_NAME}.pb',\n                          True, \"\")\n<\/code><\/pre>\n\n<p>then load it with dnn:<\/p>\n\n<pre><code>import cv2 as cv\nnet = cv.dnn.readNetFromTensorflow(f'{MODEL_PATH}\/frozen_{MODEL_NAME}.pb')\n\n# Smoke test\ninp = np.ones([1, 3, 224, 224]).astype(np.float32)\nnet.setInput(inp)\ndnn_out = net.forward()\nprint(dnn_out.shape, dnn_out[0,:5])\n## (1, 1000) [2.0760612e-04 2.6876197e-04 5.9680151e-05 5.5908626e-05 1.4762023e-04]\n<\/code><\/pre>\n\n<p>As said, I wasn't able to get optimize_for_inference to work due to the FusedBatchNorm: \n<code>WARNING:tensorflow:Didn't find expected Conv2D input to 'conv2_block1_0_bn\/FusedBatchNorm_1'\nopencv-4.0.0\/modules\/dnn\/src\/tensorflow\/tf_importer.cpp:497: error: (-2:Unspecified error) Input layer not found: conv2_block1_1_bn\/FusedBatchNorm_1 in function 'connect'<\/code>\nSo please let me know if you know a solution for that. Thanks<\/p>\n"}]}
+{"QuestionId":53493320,"AnswerCount":0,"Tags":"<python><tensorflow><artificial-intelligence><chatbot>","CreationDate":"2018-11-27T05:33:11.890","AcceptedAnswerId":null,"OwnerUserId":10579153.0,"Title":"Creating training data by training more than one files","Body":"<p>I am trying to create training data for my intelligent chatbot. Each time a user logs into the webpage he\/she could enter data to the file. These file is used to create training data. I don't have an idea that if there are more than one file how will I take data from all files and create a training data.<\/p>\n\n<p>Below is the I used for training:<\/p>\n\n<pre><code># things we need for NLP\nimport nltk\nfrom nltk.stem.lancaster import LancasterStemmer\nstemmer = LancasterStemmer()\n\n# things we need for Tensorflow\nimport numpy as np\nimport tflearn\nimport tensorflow as tf\nimport random\nimport scipy\n\n# import our chat-bot intents file\nimport json\nwith open('sales.json') as json_data:\n    intents = json.load(json_data)\nwords = []\nclasses = []\ndocuments = []\nignore_words = ['?']\n# loop through each sentence in our intents patterns\nfor intent in intents['intents']:\n    for pattern in intent['patterns']:\n        # tokenize each word in the sentence\n        w = nltk.word_tokenize(pattern)\n        # add to our words list\n        words.extend(w)\n        # add to documents in our corpus\n        documents.append((w, intent['tag']))\n        # add to our classes list\n        if intent['tag'] not in classes:\n            classes.append(intent['tag'])\n# stem and lower each word and remove duplicates\nwords = [stemmer.stem(w.lower()) for w in words if w not in ignore_words]\nwords = sorted(list(set(words)))\n\n# remove duplicates\nclasses = sorted(list(set(classes)))\nprint (len(documents), \"documents\")\nprint (len(classes), \"classes\", classes)\nprint (len(words), \"unique stemmed words\", words)\n\n# create our training data\ntraining = []\noutput = []\n# create an empty array for our output\noutput_empty = [0] * len(classes)\n\n# training set, bag of words for each sentence\nfor doc in documents:\n    # initialize our bag of words\n    bag = []\n    # list of tokenized words for the pattern\n    pattern_words = doc[0]\n    # stem each word\n    pattern_words = [stemmer.stem(word.lower()) for word in pattern_words]\n    # create our bag of words array\n    for w in words:\n        bag.append(1) if w in pattern_words else bag.append(0)\n\n\n    # output is a '0' for each tag and '1' for current tag\n    output_row = list(output_empty)\n    output_row[classes.index(doc[1])] = 1\n\n    training.append([bag, output_row])\n\n# shuffle our features and turn into np.array\nrandom.shuffle(training)\ntraining = np.array(training)\n\n# create train and test lists\ntrain_x = list(training[:,0])\ntrain_y = list(training[:,1])\n\n# reset underlying graph data\ntf.reset_default_graph()\n# Build neural network\nnet = tflearn.input_data(shape=[None, len(train_x[0])])\nnet = tflearn.fully_connected(net, 8)\nnet = tflearn.fully_connected(net, 8)\nnet = tflearn.fully_connected(net, len(train_y[0]), activation='softmax')\nnet = tflearn.regression(net)\n\n# Define model and setup tensorboard\nmodel = tflearn.DNN(net, tensorboard_dir='tflearn_logs')\n# Start training (apply gradient descent algorithm)\nmodel.fit(train_x, train_y, n_epoch=1000, batch_size=8, show_metric=True)\nmodel.save('model.tflearn')\n\nimport pickle\npickle.dump( {'words':words, 'classes':classes, 'train_x':train_x, 'train_y':train_y}, open( \"training_data\", \"wb\" ) )\n<\/code><\/pre>\n\n<p>Here I have created training data by using only one json file, however, if there are more than one file, how can I use all files to create training data.<\/p>\n","answers":[]}
+{"QuestionId":53493874,"AnswerCount":1,"Tags":"<tensorflow><tensor>","CreationDate":"2018-11-27T06:22:32.597","AcceptedAnswerId":null,"OwnerUserId":10709658.0,"Title":"ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32","Body":"<p><strong>ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor(\"embedding_1\/random_uniform:0\", shape=(4518, 1), dtype=float32)'<\/strong><\/p>\n\n<pre><code>File \"C:\\PythonProjects\\combined\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1050, in convert_to_tensor as_ref=False):\n\n@tf_export(\"convert_to_tensor\")\ndef convert_to_tensor(value, dtype=None, name=None, preferred_dtype=None):\n\nFile \"C:\\PythonProjects\\combined\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1146, in internal_convert_to_tensor\nret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n\ninternal_convert_to_tensor(value=value,dtype=dtype,name=name,preferred_dtype=preferred_dtype,as_ref=False)\n<\/code><\/pre>\n\n<p>My question is how to change dtype into int32?<\/p>\n","answers":[{"AnswerId":"53494087","CreationDate":"2018-11-27T06:40:59.460","ParentId":null,"OwnerUserId":"5810950","Title":null,"Body":"<p>you can cast a tensor from float32 to int32 either using <code>tf.cast(given_tensor, tf.int32)<\/code> or <code>tf.to_int32(given_tensor)<\/code>.<\/p>\n"}]}
+{"QuestionId":53493942,"AnswerCount":1,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2018-11-27T06:28:18.820","AcceptedAnswerId":null,"OwnerUserId":4622487.0,"Title":"How can I give the video input to PoseNet using TensorFlow.js without a HTML page?","Body":"<p>I am working on <a href=\"https:\/\/github.com\/tensorflow\/tfjs-models\/tree\/master\/posenet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tfjs-models\/tree\/master\/posenet<\/a>  which works on real time .My aim is to save the extracted keypoints from a prerecorded video to a .json file. So is there a way i can give the video input directly to the camera.js without HTML  so that i can run camera.js directly from cmd like   node camera.js.Can anyone tell me how this can be done.? Currently i am able to extract the keypoints for the poses from a prerecorded video and  display using console.<\/p>\n","answers":[{"AnswerId":"55443276","CreationDate":"2019-03-31T17:01:34.057","ParentId":null,"OwnerUserId":"5978234","Title":null,"Body":"<p>Are you still working on this Swati?  I think this could all be done easier in the browser. Posenet gets it's information from a canvas and it should be fairly easy to load a pre-recorded video onto a canvas.<\/p>\n"}]}
+{"QuestionId":53493970,"AnswerCount":1,"Tags":"<tensorflow><keras><amazon-sagemaker>","CreationDate":"2018-11-27T06:30:19.930","AcceptedAnswerId":null,"OwnerUserId":1007922.0,"Title":"How to ensure libraries installed using sagemaker?","Body":"<p>I have a jupyter notebook with a standard template code like so<\/p>\n\n<p>from sagemaker.tensorflow import TensorFlow<\/p>\n\n<pre><code>import sagemaker\nfrom sagemaker import get_execution_role\nsagemaker_session = sagemaker.Session()\nrole = get_execution_role()\n\ntf_estimator = TensorFlow(entry_point='sagemaker_predict_2.py', role=role,\n                          training_steps=10000, evaluation_steps=100,\n                          train_instance_count=1, train_instance_type='ml.p2.xlarge',\n                          framework_version='1.10.0')\ntf_estimator.fit('s3:\/\/XXX-sagemaker\/XXX')\n<\/code><\/pre>\n\n<p>This kicks off fine but eventually throws an error<\/p>\n\n<pre><code>2018-11-27 06:21:12 Starting - Starting the training job...\n2018-11-27 06:21:15 Starting - Launching requested ML instances.........\n2018-11-27 06:22:44 Starting - Preparing the instances for training...\n2018-11-27 06:23:35 Downloading - Downloading input data...\n2018-11-27 06:24:03 Training - Downloading the training image......\n2018-11-27 06:25:12 Training - Training image download completed. Training in progress..\n2018-11-27 06:25:11,813 INFO - root - running container entrypoint\n2018-11-27 06:25:11,813 INFO - root - starting train task\n2018-11-27 06:25:11,833 INFO - container_support.training - Training starting\n2018-11-27 06:25:15,306 ERROR - container_support.training - uncaught exception during training: No module named keras\nTraceback (most recent call last):\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/container_support\/training.py\", line 36, in start\n    fw.train()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tf_container\/train_entry_point.py\", line 143, in train\n    customer_script = env.import_user_module()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/container_support\/environment.py\", line 101, in import_user_module\n    user_module = importlib.import_module(script)\n  File \"\/usr\/lib\/python2.7\/importlib\/__init__.py\", line 37, in import_module\n    __import__(name)\n  File \"\/opt\/ml\/code\/sagemaker_predict_2.py\", line 7, in &lt;module&gt;\n    import keras\nImportError: No module named keras  \n<\/code><\/pre>\n\n<p>My <code>sagemaker_predict_2.py<\/code> needs some of these libraries:<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nimport sys\nimport keras\nfrom keras.models import Model, Input\nfrom keras.layers import LSTM, Embedding, Dense, TimeDistributed, Dropout, Bidirectional\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom keras_contrib.layers import CRF\n<\/code><\/pre>\n\n<p>I suppose it has no problem importing <code>pandas<\/code> and <code>numpy<\/code>, but dies when importing <code>keras<\/code>. I thought <code>keras<\/code> was standard in the notebook. When I kick this script off, does it have some other uninitialized environment?<\/p>\n\n<p>Also, I believe <code>keras_contrib<\/code> is not standard, so I will need a way to install that. How do I do that?<\/p>\n\n<p>I tried <code>!pip install keras<\/code> in the cell above but it reported that <code>Requirement already satisfied<\/code>, so it seems my jupyter environment has the library. But kicking off the <code>sagemaker_predict_2.py<\/code> must be in a different environment?<\/p>\n","answers":[{"AnswerId":"53734740","CreationDate":"2018-12-12T01:18:38.740","ParentId":null,"OwnerUserId":"9090582","Title":null,"Body":"<p>You are correct. The sagemaker_predict_2.py runs in a different environment from your notebook instance. That particular code runs on SageMaker executed inside of our predefined TensorFlow Docker container.<\/p>\n\n<p>Installing your dependencies in the notebook instance will only allow access to the installed libraries within the notebook kernel.<\/p>\n\n<p>As for installing your dependencies inside of the running Docker container, that can be achieved by <a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/estimator.py#L176\" rel=\"nofollow noreferrer\">specifying your dependencies<\/a> in a <a href=\"https:\/\/pip.pypa.io\/en\/stable\/user_guide\/#requirements-files\" rel=\"nofollow noreferrer\">requirements.txt<\/a>.<\/p>\n\n<p>Since iterations can take from 8-10 minutes, it is recommended to use local mode to make sure that your training job can run locally before sending the training job to SageMaker. This can be done by specifying the training_instance_type as 'local' or please reference this notebook: <a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/tensorflow_distributed_mnist\/tensorflow_local_mode_mnist.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/tensorflow_distributed_mnist\/tensorflow_local_mode_mnist.ipynb<\/a><\/p>\n\n<p>What local mode does is essentially run your docker container on the localhost that is executing the Python code. This can be on our SageMaker notebook instance or your own local machine.<\/p>\n"}]}
+{"QuestionId":53494039,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-27T06:36:23.943","AcceptedAnswerId":null,"OwnerUserId":6159371.0,"Title":"How can tensorflow_model_server bind its HTTP\/REST on 0.0.0.0 (default localhost)?","Body":"<p>I found that docker image TensorFlow\/serving actually use the executable <code>tensorflow_model_server<\/code> to start a server. So I attach into a docker container to \nstart the server manually.<\/p>\n\n<p>My command:<\/p>\n\n<pre><code>tensorflow_model_server --model_base_path=\/root\/serving\/my_image_classifier\/ --rest_api_port=8501 --model_name=ImageClassifier\n<\/code><\/pre>\n\n<p>This command start the server, with logs:<\/p>\n\n<pre><code>2018-11-27 03:56:43.302391: I tensorflow_serving\/core\/loader_harness.cc:86] Successfully loaded servable version {name: ImageClassifier version: 2}\n\n2018-11-27 03:56:43.304439: I tensorflow_serving\/model_servers\/server.cc:286] Running gRPC ModelServer at 0.0.0.0:8500 ...\n\n[warn] getaddrinfo: address family for nodename not supported\n\n2018-11-27 03:56:43.306009: I tensorflow_serving\/model_servers\/server.cc:302] Exporting HTTP\/REST API at:localhost:8501 \n<\/code><\/pre>\n\n<p>As it says\uff0cits HTTP\/REST API exported as localhost. That's not right, cause if a sever bind on localhost in docker, it cannot be attach from outside.<\/p>\n\n<p>I read --help arefully, but didn't find anything usefully.<\/p>\n\n<p><strong>So, how can I let it bind on 0.0.0.0?<\/strong><\/p>\n\n<p>TensorFlow ModelServer version:<\/p>\n\n<pre><code>TensorFlow ModelServer: 1.12.0-rc0+dev.sha.87470f0\nTensorFlow Library: 1.12.0\n<\/code><\/pre>\n","answers":[{"AnswerId":"56395176","CreationDate":"2019-05-31T12:30:03.343","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>In the below command, since you are mentioning <code>rest_api_port=8501<\/code>, it is binding it with <code>localhost<\/code> (default value). <\/p>\n\n<pre><code>tensorflow_model_server --model_base_path=\/root\/serving\/my_image_classifier\/ --rest_api_port=8501 --model_name=ImageClassifier\n<\/code><\/pre>\n\n<p>If you want to bind it to some IP address, say <code>0.0.0.0<\/code>, you can try passing its value as <\/p>\n\n<p><code>rest_api_port=0.0.0.0:8501<\/code><\/p>\n"},{"AnswerId":"53530835","CreationDate":"2018-11-29T02:07:24.523","ParentId":null,"OwnerUserId":"6159371","Title":null,"Body":"<p>In fact, even the logs show that the HTTP\/REST API bind on \"localhost\", it actucally <strong>can be attached<\/strong> outside of the docker container.<\/p>\n\n<p>Why I could not attach that is another strange problem (whitch automatic disapeared after the machine rebooted).<\/p>\n\n<p>However, I still think the tensorflow_model_server uses \"localhost\" in log is not right, especially when it tells RPC bind on \"0.0.0.0\".<\/p>\n\n<p>I should open an issue.<\/p>\n"}]}
+{"QuestionId":53494498,"AnswerCount":2,"Tags":"<out-of-memory><pytorch>","CreationDate":"2018-11-27T07:14:41.390","AcceptedAnswerId":null,"OwnerUserId":7242982.0,"Title":"Pytorch RuntimeError: CUDA error: out of memory at loss.backward() , No error when using CPU","Body":"<p>I'm training a Fully convolutional network (FCN32) for semantic segmentation on Tesla K80 with more than 11G memory.<\/p>\n\n<p>The input image is pretty large: 352x1216. Network structure is shown below. I used <em>batch_size=1<\/em>, but still encounter the \nout_of_memory error. <\/p>\n\n<p>Criterion is nn.BCEWithLogitsLoss()<\/p>\n\n<p>The network works fine when I run on CPU.<\/p>\n\n<hr>\n\n<pre><code>    Layer (type)               Output Shape       #  Param \n        Conv2d-1        [-1, 64, 352, 1216]           1,792\n        Conv2d-2        [-1, 64, 352, 1216]          36,928\n     MaxPool2d-3         [-1, 64, 176, 608]               0\n        Conv2d-4        [-1, 128, 176, 608]          73,856\n        Conv2d-5        [-1, 128, 176, 608]         147,584\n     MaxPool2d-6         [-1, 128, 88, 304]               0\n        Conv2d-7         [-1, 256, 88, 304]         295,168\n        Conv2d-8         [-1, 256, 88, 304]         590,080\n        Conv2d-9         [-1, 256, 88, 304]         590,080\n    MaxPool2d-10         [-1, 256, 44, 152]               0\n       Conv2d-11         [-1, 512, 44, 152]       1,180,160\n       Conv2d-12         [-1, 512, 44, 152]       2,359,808\n       Conv2d-13         [-1, 512, 44, 152]       2,359,808\n    MaxPool2d-14          [-1, 512, 22, 76]               0\n       Conv2d-15          [-1, 512, 22, 76]       2,359,808\n       Conv2d-16          [-1, 512, 22, 76]       2,359,808\n       Conv2d-17          [-1, 512, 22, 76]       2,359,808\n    MaxPool2d-18          [-1, 512, 11, 38]               0\n       Conv2d-19         [-1, 4096, 11, 38]     102,764,544\n       Conv2d-20         [-1, 4096, 11, 38]      16,781,312\n       Conv2d-21          [-1, 1, 11, 38]           4,097  ConvTranspose2d-22         [-1, 1, 352, 1216]             4,096\n<\/code><\/pre>\n\n<hr>\n\n<p>Error message:<\/p>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- RuntimeError                              Traceback (most recent call\n  last)  in ()\n       36         print (loss)\n       37         #torch.cuda.empty_cache()\n  ---> 38         loss.backward()\n       39         optimizer.step()\n       40 <\/p>\n  \n  <p>\/anaconda\/envs\/py35\/lib\/python3.5\/site-packages\/torch\/tensor.py in\n  backward(self, gradient, retain_graph, create_graph)\n       91                 products. Defaults to <code>False<\/code>.\n       92         \"\"\"\n  ---> 93         torch.autograd.backward(self, gradient, retain_graph, create_graph)\n       94 \n       95     def register_hook(self, hook):<\/p>\n  \n  <p>\/anaconda\/envs\/py35\/lib\/python3.5\/site-packages\/torch\/autograd\/<strong>init<\/strong>.py\n  in backward(tensors, grad_tensors, retain_graph, create_graph,\n  grad_variables)\n       88     Variable._execution_engine.run_backward(\n       89         tensors, grad_tensors, retain_graph, create_graph,\n  ---> 90         allow_unreachable=True)  # allow_unreachable flag\n       91 \n       92 <\/p>\n  \n  <p>RuntimeError: CUDA error: out of memory<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53566538","CreationDate":"2018-12-01T00:03:32.013","ParentId":null,"OwnerUserId":"7242982","Title":null,"Body":"<p>I found out the reason... It's hardware related. I changed to another machine and the error disappeared.<\/p>\n"},{"AnswerId":"53568701","CreationDate":"2018-12-01T07:25:25.003","ParentId":null,"OwnerUserId":"7606121","Title":null,"Body":"<p>Usually this happens because of <strong>memory on your GPU<\/strong>. If you have more powerful GPUs, your problem could be solved (as you mentioned in your answer).<br>\nBut if you do not have, you can <strong>scale down your images<\/strong> into about <code>256*x<\/code> sizes.\nIt is also good practice for <strong>performance<\/strong>'s sake.<\/p>\n"}]}
+{"QuestionId":53494627,"AnswerCount":1,"Tags":"<machine-learning><keras><loss-function>","CreationDate":"2018-11-27T07:24:13.877","AcceptedAnswerId":53497601.0,"OwnerUserId":1511274.0,"Title":"Loss in Keras Model evaluation","Body":"<p>I am doing binary classification with Keras\n<code>loss='binary_crossentropy'<\/code>, <code>optimizer=tf.keras.optimizers.Adam<\/code> and final layer is <code>keras.layers.Dense(1, activation=tf.nn.sigmoid)<\/code>.<\/p>\n\n<p>As I know, <code>loss<\/code> value is used to evaluate the model during training phase. However, when I use <code>Keras<\/code> model evaluation for my <strong>testing dataset<\/strong> (e.g. <code>m_recall.evaluate(testData,testLabel)<\/code>, there are also <code>loss<\/code> values, accompanied by <code>accuracy<\/code> values like the output below<\/p>\n\n<pre><code>test size:  (1889, 18525)\n1889\/1889 [==============================] - 1s 345us\/step\nm_acc:  [0.5690245978371045, 0.9523557437797776]\n1889\/1889 [==============================] - 1s 352us\/step\nm_recall:  [0.24519687695911097, 0.9359449444150344]\n1889\/1889 [==============================] - 1s 350us\/step\nm_f1:  [0.502442331737344, 0.9216516675489677]\n1889\/1889 [==============================] - 1s 360us\/step\nmetric name:  ['loss', 'acc']\n<\/code><\/pre>\n\n<p>What is the meaning\/usage of <code>loss<\/code> during testing? Why it is so high (e.g. <code>0.5690<\/code> in <code>m_acc<\/code>)? The accuracy evaluation seems fine to me (e.g. <code>0.9523<\/code> in <code>m_acc<\/code>) but I am concerned about the <code>loss<\/code> too, does it make my model perform badly?<\/p>\n\n<p><strong>P.S.<\/strong> \n<code>m_acc<\/code>, <code>m_recall<\/code>, etc. are just the way I name my models (they were trained by on different metrics in <code>GridSearchCV<\/code>)<\/p>\n\n<p><strong>Update:<\/strong>\nI just realized that <code>loss<\/code> values are not in percentage, so how are they calculated? And with current values, are they good enough or do I need to optimize them more?<\/p>\n\n<p>Suggestions for further reading are appreciated too!<\/p>\n","answers":[{"AnswerId":"53497601","CreationDate":"2018-11-27T10:28:13.997","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>When defining a machine learning model, we want a way to measure the performance of our model so that we could compare it with other models to choose the best one and also make sure that it is good enough. Therefore, we define some <strong>metrics<\/strong> like accuracy (in the context of classification), which is the proportion of correctly classified samples by the model, to measure how our model performs and whether it is good enough for our task or not.<\/p>\n\n<p>Although these metrics are truly comprehensible by us, however the problem is that they cannot be directly used by the learning process of our models to tune the parameters of the model. Instead, we define other measures, which are usually called <strong>loss functions<\/strong> or <strong>objective functions<\/strong>, which can be directly used by the training process (i.e. optimization). These functions are usually defined such that we expect that when their values are low we would have a high accuracy. That's why you would commonly see that the machine learning algorithms are trying to minimize a loss function with the expectation that the accuracy increases. In other words, the models are <strong>indirectly learning<\/strong> by optimizing the loss functions. The loss values are important during training of the model, e.g. if they are not decreasing or fluctuating then this means there is a problem somewhere that needs to be fixed.<\/p>\n\n<p>As a result, what we are ultimately (i.e. when testing a model) concerned about is the value of metrics (like accuracy) we have initially defined and we don't care about the final value of loss functions. That's why you don't hear things like \"the <strong>loss value<\/strong> of a [specific model] on the ImageNet dataset is 8.732\"! That does not tell you anything whether the model is great, good, bad or terrible. Rather, you would hear that \"this model performs with 87% <strong>accuracy<\/strong> on the ImageNet dataset\".<\/p>\n"}]}
+{"QuestionId":53494667,"AnswerCount":1,"Tags":"<tensorflow><cpu>","CreationDate":"2018-11-27T07:27:25.090","AcceptedAnswerId":null,"OwnerUserId":9873377.0,"Title":"Can traing deep learning model using Intel Xeon CPU in tensorflow , solving short of gpu memory","Body":"<p>NVidia GPU has 16GB memory at most, which limits large model training. Model parallism may needs modification of deep learning framework. Is it feasible to train tensorflow models using Intel multi-core CPUs? Could you give some advice about the hardware configuration and the performance?<\/p>\n","answers":[{"AnswerId":"53671008","CreationDate":"2018-12-07T13:56:38.813","ParentId":null,"OwnerUserId":"9465768","Title":null,"Body":"<p>You can try using Intel AI Devcloud which is cloud hosted hardware and software platform available to developers, researchers and startups to learn and get started on their Artificial Intelligence projects. It has Intel\u00ae Xeon\u00ae Scalable Processors and each processor has 24 cores with 2-way hyper-threading. Each processor has access to 96 GB of on-platform RAM. <\/p>\n\n<p>Refer the below link for more details.<\/p>\n\n<p><a href=\"https:\/\/ai.intel.com\/devcloud\/\" rel=\"nofollow noreferrer\">https:\/\/ai.intel.com\/devcloud\/<\/a><\/p>\n\n<p>You can access this platform for 30 days by registering in the following link.<\/p>\n\n<p><a href=\"https:\/\/software.intel.com\/en-us\/ai-academy\/devcloud\" rel=\"nofollow noreferrer\">https:\/\/software.intel.com\/en-us\/ai-academy\/devcloud<\/a><\/p>\n\n<p>You will get a welcome mail which gives the user name and password. Open the hyperlink in the welcome mail to get more details on how to connect and use the Devcloud.\nTo get the best performance on Devcloud, change the parallelism threads and OpenMP settings (either inside the code or in the terminal) as below: <\/p>\n\n<p>In the terminal:<\/p>\n\n<p>export OMP_NUM_THREADS=\"NUM_PARALLEL_EXEC_UNITS\"<\/p>\n\n<p>export KMP_BLOCKTIME=\"0\" <\/p>\n\n<p>export KMP_SETTINGS=\"1\" <\/p>\n\n<p>export KMP_AFFINITY=\"granularity=fine,verbose,compact,1,0\"<\/p>\n\n<p>Inside code:<\/p>\n\n<p>import os<\/p>\n\n<p>os.environ[\"OMP_NUM_THREADS\"] = \"NUM_PARALLEL_EXEC_UNITS\"<\/p>\n\n<p>os.environ[\"KMP_BLOCKTIME\"] = \"0\" <\/p>\n\n<p>os.environ[\"KMP_SETTINGS\"] = \"1\" <\/p>\n\n<p>os.environ[\"KMP_AFFINITY\"]= \"granularity=fine,verbose,compact,1,0\"<\/p>\n\n<p>For more details regarding optimization, please refer:<\/p>\n\n<p><a href=\"https:\/\/communities.intel.com\/docs\/DOC-112392\" rel=\"nofollow noreferrer\">https:\/\/communities.intel.com\/docs\/DOC-112392<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":53495424,"AnswerCount":0,"Tags":"<keras><neural-network>","CreationDate":"2018-11-27T08:23:45.507","AcceptedAnswerId":null,"OwnerUserId":10710212.0,"Title":"approximate simple function given parameters with neural network in Keras","Body":"<p>I need to create a neural network that approximates a function given its parameters. I give four parameters to my neural network (A, x0, phi, omega) and I want to obtain, as output, <\/p>\n\n<blockquote>\n  <p>A sin(omega x + phi) + x0<\/p>\n<\/blockquote>\n\n<p>(I need this net as a part of another network)\nHowever, I am not able to train the network as I obtain a very poor convergence. Why is that?<\/p>\n\n<p>I use a fully connected network with three layers. This is the code<\/p>\n\n<pre><code>    def get_batches(N_batches):\n        A = tan( random.uniform(low=0.,high=2*pi,size=[N_batches,1]))\n        x0 = random.randn(N_batches,1)*10\n        omega = random.uniform(low=0.,high=10*pi, size=[N_batches,1])\n        phi = random.uniform(low=0.,high=2*pi, size=[N_batches,1])  \n        x = linspace(0,t_max, n_max)\n        x = tile(x,N_batches).reshape(N_batches,n_max)\n        return (A*sin(omega*x+phi) + x0, hstack([A,x0,phi,omega]) )\n\n    N_batches = 80\n    N_epochs = 50\n    t_max = 5.0\n    n_max = 100\n    n_par = 4\n\n    net_layers = []\n\n    net_inp = Input(shape=(n_par,))\n    net_layers.append(Dense(25, input_shape=(n_par,),  activation=\"relu\"))\n    net_layers.append(Dense(25,  activation=\"relu\"))\n    net_layers.append(Dense(25,  activation=\"relu\"))\n    net_layers.append(Dense(n_max, activation=\"linear\"))\n\n    net_l = net_inp\n    for i in range(len(net_layers)):\n        net_l = net_layers[i](net_l)\n    net = Model(net_inp, net_l)\n    net.compile(loss=\"mean_squared_error\", optimizer=\"adam\") \n\n    costs = zeros(N_epochs)\n    for i in range(N_epochs):\n        y_true, y_in = get_batches(N_batches)\n        costs[i]=net.train_on_batch(y_in,y_true)\n<\/code><\/pre>\n\n<p>Even if I train more, I don't get better results than this\npicture (approximated function and real function plot for a test sample):<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/nJFJP.png\" alt=\"plot1\"> <\/p>\n\n<p>The plot of the cost function is quite strange:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Zo6to.png\" alt=\"plot2\"><\/p>\n\n<p>What mistakes did I do? Thank you!<\/p>\n","answers":[]}
+{"QuestionId":53495522,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2018-11-27T08:31:01.580","AcceptedAnswerId":53495633.0,"OwnerUserId":5247304.0,"Title":"Get a prediction from Tensor Flow Model","Body":"<p>I want to get predictions from my trained tensor flow model. The following is the code I have for training my model.<\/p>\n\n<pre><code>def train_model(self, train, test, learning_rate=0.0001, num_epochs=16, minibatch_size=32, print_cost=True, graph_filename='costs'):\n\n        # Ensure that model can be rerun without overwriting tf variables\n        ops.reset_default_graph()\n        # For reproducibility\n        tf.set_random_seed(42)\n        seed = 42\n        # Get input and output shapes\n        (n_x, m) = train.images.T.shape\n        n_y = train.labels.T.shape[0]\n\n        costs = []\n\n        # Create placeholders of shape (n_x, n_y)\n        X, Y = self.create_placeholders(n_x, n_y)\n        # Initialize parameters\n        parameters = self.initialize_parameters()\n\n        # Forward propagation\n        Z3 = self.forward_propagation(X, parameters)\n        # Cost function\n        cost = self.compute_cost(Z3, Y)\n        # Backpropagation (using Adam optimizer)\n        optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)\n\n        # Initialize variables\n        init = tf.global_variables_initializer()\n\n        # Start session to compute Tensorflow graph\n        with tf.Session() as sess:\n\n            # Run initialization\n            sess.run(init)\n\n            # Training loop\n            for epoch in range(num_epochs):\n\n                epoch_cost = 0.\n                num_minibatches = int(m \/ minibatch_size)\n                seed = seed + 1\n\n                for i in range(num_minibatches):\n\n                    # Get next batch of training data and labels\n                    minibatch_X, minibatch_Y = train.next_batch(minibatch_size)\n\n                    # Execute optimizer and cost function\n                    _, minibatch_cost = sess.run([optimizer, cost], feed_dict={X: minibatch_X.T, Y: minibatch_Y.T})\n\n                    # Update epoch cost\n                    epoch_cost += minibatch_cost \/ num_minibatches\n\n                # Print the cost every epoch\n                if print_cost == True:\n                    print(\"Cost after epoch {epoch_num}: {cost}\".format(epoch_num=epoch, cost=epoch_cost))\n                    costs.append(epoch_cost)\n\n            # Plot costs\n            plt.figure(figsize=(16,5))\n            plt.plot(np.squeeze(costs), color='#2A688B')\n            plt.xlim(0, num_epochs-1)\n            plt.ylabel(\"cost\")\n            plt.xlabel(\"iterations\")\n            plt.title(\"learning rate = {rate}\".format(rate=learning_rate))\n            plt.savefig(graph_filename, dpi=300)\n            plt.show()\n\n            # Save parameters\n            parameters = sess.run(parameters)\n            print(\"Parameters have been trained!\")\n\n            # Calculate correct predictions\n            correct_prediction = tf.equal(tf.argmax(Z3), tf.argmax(Y))\n\n            # Calculate accuracy on test set\n            accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n\n            print (\"Train Accuracy:\", accuracy.eval({X: train.images.T, Y: train.labels.T}))\n            print (\"Test Accuracy:\", accuracy.eval({X: test.images.T, Y: test.labels.T}))\n\n        return parameters\n<\/code><\/pre>\n\n<p>After training the model, I want to extract the prediction from the model.\nSo I add <\/p>\n\n<p><code>print(sess.run(accuracy, feed_dict={X: test.images.T}))<\/code><\/p>\n\n<p>But I am seeing the below error after running the above code:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: You must feed a value for placeholder tensor 'Y'\n  with dtype float and shape [10,?]\n           [[{{node Y}} = Placeholderdtype=DT_FLOAT, shape=[10,?], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>Any help is welcome..<\/p>\n","answers":[{"AnswerId":"53495633","CreationDate":"2018-11-27T08:39:25.617","ParentId":null,"OwnerUserId":"5024514","Title":null,"Body":"<p>The tensor <code>accuracy<\/code> is a function of the tensor <code>correct_prediction<\/code>, which in turn is a function of (among the rest) <code>Y<\/code>.<br>\nSo you're correctly being told that you should feed values for that placeholder too.<br>\nI'm assuming <code>Y<\/code> hold your labels, so it should also make intuitive sense that your feed_dict would also contain the correct <code>Y<\/code> values.<br>\nHope that helps.<br>\nGood luck!<\/p>\n"}]}
+{"QuestionId":53496095,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><rnn>","CreationDate":"2018-11-27T09:08:02.803","AcceptedAnswerId":53496805.0,"OwnerUserId":6187009.0,"Title":"Keras reports TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'","Body":"<p>I'm a beginner in Keras and just write a toy example. It reports a <code>TypeError<\/code>. The code and error are as follows:<\/p>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre><code>inputs = keras.Input(shape=(3, ))\n\ncell = keras.layers.SimpleRNNCell(units=5, activation='softmax')\nlabel = keras.layers.RNN(cell)(inputs)\n\nmodel = keras.models.Model(inputs=inputs, outputs=label)\nmodel.compile(optimizer='rmsprop',\n              loss='mae',\n              metrics=['acc'])\n\ndata = np.array([[1, 2, 3], [3, 4, 5]])\nlabels = np.array([1, 2])\nmodel.fit(x=data, y=labels)\n<\/code><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><code>Traceback (most recent call last):\n    File \"\/Users\/david\/Documents\/code\/python\/Tensorflow\/test.py\", line 27, in &lt;module&gt;\n        run()\n    File \"\/Users\/david\/Documents\/code\/python\/Tensorflow\/test.py\", line 21, in run\n        label = keras.layers.RNN(cell)(inputs)\n    File \"\/Users\/david\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/recurrent.py\", line 619, in __call__\n...\n    File \"\/Users\/david\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/init_ops.py\", line 473, in __call__\n        scale \/= max(1., (fan_in + fan_out) \/ 2.)\nTypeError: unsupported operand type(s) for +: 'NoneType' and 'int'\n<\/code><\/pre>\n\n<p>So how can I deal with it? <\/p>\n","answers":[{"AnswerId":"53496805","CreationDate":"2018-11-27T09:46:58.080","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The input to a RNN layer would have a shape of <code>(num_timesteps, num_features)<\/code>, i.e. each sample consists of <code>num_timesteps<\/code> timesteps where each timestep is a vector of length <code>num_features<\/code>. Further, the number of timesteps (i.e. <code>num_timesteps<\/code>) could be variable or unknown (i.e. <code>None<\/code>) but the number of features (i.e. <code>num_features<\/code>) should be fixed and specified from the beginning. Therefore, you need to change the shape of Input layer to be consistent with the RNN layer. For example:<\/p>\n\n<pre><code>inputs = keras.Input(shape=(None, 3))  # variable number of timesteps each with length 3\ninputs = keras.Input(shape=(4, 3))     # 4 timesteps each with length 3\ninputs = keras.Input(shape=(4, None))  # this is WRONG! you can't do this. Number of features must be fixed\n<\/code><\/pre>\n\n<p>Then, you also need to change the shape of input data (i.e. <code>data<\/code>) as well to be consistent with the input shape you have specified (i.e. it must have a shape of <code>(num_samples, num_timesteps, num_features)<\/code>).<\/p>\n\n<p>As a side note, you could define the RNN layer more simply by using the <code>SimpleRNN<\/code> layer directly:<\/p>\n\n<pre><code>label = keras.layers.SimpleRNN(units=5, activation='softmax')(inputs)\n<\/code><\/pre>\n"},{"AnswerId":"55660438","CreationDate":"2019-04-12T23:02:28.717","ParentId":null,"OwnerUserId":"4468490","Title":null,"Body":"<p>I think @today's answer is very clear. However, not complete. The key thing here is that, if your input doesn't contain <code>num_features<\/code>, you have to make a <code>Embedding<\/code> layer next to the input.<\/p>\n\n<p>So if you use:<\/p>\n\n<pre><code>inputs = keras.Input(shape=(3,))\nembedding = Embedding(voc_size, embed_dim, ..)\nX = embedding(inputs)\n<\/code><\/pre>\n\n<p>it also works.<\/p>\n"}]}
+{"QuestionId":53496373,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-27T09:24:06.167","AcceptedAnswerId":null,"OwnerUserId":5677268.0,"Title":"Train Keras Model on large number of .mat files","Body":"<p>I want to train a model using neural network on Python using Keras. The data I have are a bunch of .mat files that contain data associated with ECG signals. <\/p>\n\n<p>All .mat files have same structure but different values. However there is some differences in some array sizes in the files. For example, one file contain an array called \"P_Signal\" having a size of 9000, another file have it with a size of 18000.<\/p>\n\n<p>I want to train a neural network model using Keras. As far as I know I should prepare a CSV file containing all the necessary attributes + label on each column so Keras module can understand it.<\/p>\n\n<p>Unfortunately, this is impossible due to large arrays and fields inside the .mat files (CSV file can not load more than 16,384 columns and I have some arrays of 18000 each). So I should find another way to load those data in Keras.<\/p>\n\n<p>Appreciate your help a lot.<\/p>\n","answers":[]}
+{"QuestionId":53496570,"AnswerCount":2,"Tags":"<python><neural-network><pytorch><activation-function>","CreationDate":"2018-11-27T09:34:18.350","AcceptedAnswerId":53496823.0,"OwnerUserId":8323793.0,"Title":"Matrix multiplication in pyTorch","Body":"<p>I'm writing a simple neural network in pyTorch, where features and weights both are (1, 5) tensors. What are the differences between the two methods that I mention below?<\/p>\n\n<pre><code>y = activation(torch.sum(features*weights) + bias)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>yy = activation(torch.mm(features, weights.view(5,1)) + bias)\n<\/code><\/pre>\n","answers":[{"AnswerId":"56665305","CreationDate":"2019-06-19T10:15:26.100","ParentId":null,"OwnerUserId":"9995930","Title":null,"Body":"<pre><code>features = torch.rand(1, 5) \nweights = torch.Tensor([1, 2, 3, 4, 5])\nprint(features)\nprint(weights)\n\n# Element-wise multiplication of shape (1 x 5)\n# out = [f1*w1, f2*w2, f3*w3, f4*w4, f5*w5]\nprint(features*weights)\n\n# weights has been reshaped to (5, 1)\n# Element-wise multiplication of shape (5 x 5)\n# out =   [f1*w1, f2*w1, f3*w1, f4*w1, f5*w1]\n#         [f1*w2, f2*w2, f3*w2, f4*w2, f5*w2]\n#         [f1*w3, f2*w3, f3*w3, f4*w3, f5*w3]\n#         [f1*w4, f2*w4, f3*w4, f4*w4, f5*w4]\n#         [f1*w5, f2*w5, f3*w5, f4*w5, f5*w5]\nprint(features*weights.view(5, 1))\n\n# Matrix-multiplication\n# (1, 5) * (5, 1) -&gt; (1, 1)\n# out = [f1*w1 + f2*w2 + f3*w3 + f4*w4 + f5*w5]\nprint(torch.mm(features, weights.view(5, 1)))\n<\/code><\/pre>\n\n<p>output<\/p>\n\n<pre><code>tensor([[0.1467, 0.6925, 0.0987, 0.5244, 0.6491]])  # features\ntensor([1., 2., 3., 4., 5.])                        # weights\n\ntensor([[0.1467, 1.3851, 0.2961, 2.0976, 3.2455]])  # features*weights\ntensor([[0.1467, 0.6925, 0.0987, 0.5244, 0.6491],\n        [0.2934, 1.3851, 0.1974, 1.0488, 1.2982],\n        [0.4400, 2.0776, 0.2961, 1.5732, 1.9473],\n        [0.5867, 2.7701, 0.3947, 2.0976, 2.5964],\n        [0.7334, 3.4627, 0.4934, 2.6220, 3.2455]])  # features*weights.view(5,1)\ntensor([[7.1709]])                                  # torch.mm(features, weights.view(5, 1))\n<\/code><\/pre>\n"},{"AnswerId":"53496823","CreationDate":"2018-11-27T09:47:58.577","ParentId":null,"OwnerUserId":"6799297","Title":null,"Body":"<p>Consider it step by step:<\/p>\n\n<pre><code>x = torch.tensor([[10, 2], [3,5]])\ny = torch.tensor([[1,3], [5,6]])\n\nx * y\n# tensor([[10,  6],\n#         [15, 30]])\n\ntorch.sum(x*y)\n\n#tensor(61)\n\nx = torch.tensor([[10, 2], [3,5]])\ny = torch.tensor([[1,3], [5,6]])\n\nnp.matmul(x, y)\n# array([[20, 42],\n#       [28, 39]])\n<\/code><\/pre>\n\n<p>So there is a difference betweeen <code>matmul<\/code> and <code>* operator<\/code>. Furthermore, torch.sum makes an entire sum from the tensor, not row or columnwisely. <\/p>\n"}]}
+{"QuestionId":53496913,"AnswerCount":1,"Tags":"<python><tensorflow><nvidia><ubuntu-18.04>","CreationDate":"2018-11-27T09:52:06.037","AcceptedAnswerId":null,"OwnerUserId":10710778.0,"Title":"Import error tensorflow-gpu in ubuntu18.04","Body":"<p>I'm getting some error while importing tensorflow.<\/p>\n\n<p>My computer's Specifications:<\/p>\n\n<p>OS:ubuntu 18.04<\/p>\n\n<p>Nvidia RTX 2080 Ti*2 <\/p>\n\n<p>Nvidia driver-415<\/p>\n\n<p>CUDA\uff1a10.0<\/p>\n\n<p>cuDNN\uff1a7.3.0\ntensorflow\uff1a1.11.0<\/p>\n\n<pre><code>import tensorflow\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File \"\/usr\/lib\/python3.6\/imp.py\", line 243,\n  in load_module\n      return load_dynamic(name, filename, file)   File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n      return _load(spec) ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n  \n  <p>Traceback (most recent call last):   File \"\", line 1, in\n     File\n  \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/<strong>init<\/strong>.py\", line\n  22, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import   File\n  \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/<strong>init<\/strong>.py\",\n  line 49, in \n      from tensorflow.python import pywrap_tensorflow   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\",\n  line 74, in \n      raise ImportError(msg) ImportError: Traceback (most recent call last):   File\n  \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\",\n  line 24, in swig_import_helper\n      _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)   File \"\/usr\/lib\/python3.6\/imp.py\", line 243,\n  in load_module\n      return load_dynamic(name, filename, file)   File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n      return _load(spec) ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory<\/p>\n  \n  <p>Failed to load the native TensorFlow runtime.<\/p>\n  \n  <p>See\n  <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources#common_installation_problems<\/a><\/p>\n  \n  <p>for some common reasons and solutions.  Include the entire stack trace\n  above this error message when asking for help.<\/p>\n<\/blockquote>\n\n<p>I already tried to use ubuntu 16.04 ,but GPU didn't support.\nInstallation of CUDA9.0 &amp; CUDA9.2 was not supported too.<\/p>\n\n<p>how can I use tensorflow-gpu?<\/p>\n\n<p>I already app path in ~\/.bashrc<\/p>\n\n<pre><code>export PATH=\/usr\/local\/cuda-10.0\/bin${PATH:+:${PATH}}\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}\nexport CUDA_HOME=\/usr\/local\/cuda-10.0\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/lib64\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/lib64$LD_LIBRARY_PATH\n<\/code><\/pre>\n","answers":[{"AnswerId":"53500218","CreationDate":"2018-11-27T12:54:33.700","ParentId":null,"OwnerUserId":"7370153","Title":null,"Body":"<p>The <code>tensorflow-gpu<\/code> package is built against Cuda 9.0, but you have Cuda 10.0 installed. <\/p>\n\n<p>You need either to downgrade your version of Cuda to 9.0 (but if I recall, that's not possible with a 2080Ti), or build tensorflow from the sources. There is extensive documentation on how to do so on the <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">tensorflow webiste<\/a><\/p>\n\n<p>You can also try to install the package <code>tf-nightly-gpu<\/code>. You should note that this version is more experimental though, as it has not been tested as extensively. <\/p>\n"}]}
+{"QuestionId":53496948,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-27T09:53:23.610","AcceptedAnswerId":53541695.0,"OwnerUserId":6925077.0,"Title":"Tensorflow to Keras: import graph def error on Keras model","Body":"<p>I have a Tensorflow code for classifying images which I want to convert to Keras code. But I'm having trouble with the higher level API not having all codes which I desire. The problem which I have been stuck at is: <\/p>\n\n<pre><code>#net = get_vgg_model() &lt;- got tf.VGG16 model\nnet = tf.keras.applications.VGG16()\n\n\ng1 = tf.Graph()\nwith tf.Session(graph=g1, config=config) as sess, g1.device('\/cpu:0'):\n    tf.import_graph_def(net['graph_def'], name='vgg')\n<\/code><\/pre>\n\n<p>this code gives the error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"app.py\", line 16, in &lt;module&gt;\n    from modules.xvision import Xvision\n    File \"\/app\/modules\/xvision.py\", line 84, in &lt;module&gt;\n       tf.import_graph_def(net['graph_def'], name='vgg')\n   TypeError: 'Model' object has no attribute '__getitem__'\n<\/code><\/pre>\n\n<p>Could someone help me with this graph?<\/p>\n","answers":[{"AnswerId":"53541695","CreationDate":"2018-11-29T14:51:32.237","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h2>Getting the graph<\/h2>\n\n<p>You can get the graph from Keras with: <\/p>\n\n<pre><code>import keras.backend as K\nK.get_session().graph\n<\/code><\/pre>\n\n<p>You can probably pass it to <code>import_graph_def<\/code>, but I suspect it's already Tensorflow's default graph, since in the link below, the creator of Keras says there is only one graph. <\/p>\n\n<p>More in: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3223\" rel=\"noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/3223<\/a><\/p>\n\n<h2>Working suggestion<\/h2>\n\n<p>I don't know what you're trying to achieve, but if the idea is using Keras regularly, you'd probably never need to grab the graph.<\/p>\n\n<p>In Keras, once you created your model with <code>net = tf.keras.applications.VGG16()<\/code>, you'd start using Keras methods from this model, such as:<\/p>\n\n<pre><code>#compile for training\nnet.compile(optimizer=someKerasOptimizer, loss=someKerasLoss, metrics=[m1,m2])\n\n#training\nnet.fit(trainingInputs, trainingTargets, epochs=..., batch_size=..., ...)    \nnet.fit_generator(someGeneratorThatLoadsBatches, steps_per_epoch=...., ....)\n\n#predicting\nnet.predict(inputs)\nnet.predict_generator(someGeneratorThatLoadsInputImages, steps=howManyBatches)    \n<\/code><\/pre>\n\n<p>Accessing weights and layers would be done by:<\/p>\n\n<pre><code>layer = net.layers[index]\nlayer = net.get_layer('layer_name')\n\nweights = layer.get_weights()\nlayer.set_weights(someWeightsList)\n\nallWeights = net.get_weights()\nnet.set_weights(listWithAllWeights)\n<\/code><\/pre>\n"},{"AnswerId":"53581718","CreationDate":"2018-12-02T15:29:12.423","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>If what you are trying to do is import trained tensorflow to keras, First you have to consider that the naming of each keras model's variable matches with tensorflows model's variables. To show this I've created simple model with single input and single output without hidden layer. To deal with naming issue I've created the model using keras's layers and used keras to train the model.<\/p>\n\n<pre><code>inputs = tf.keras.layers.Input(shape=(1,), name=\"inputs\")\noutputs = tf.keras.layers.Dense(1, activation=\"linear\", name=\"outputs\")(inputs)\nmodel = tf.keras.models.Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n\n<p>Now fit the network with dummy dataset, then save the model using tensorflow<\/p>\n\n<pre><code>model.compile(loss=\"mse\", optimizer=tf.keras.optimizers.Adam(1e-1))\nx = np.random.randn(1000) * 1000\ny = x * .5 + 3\nmodel.fit(x, y, epochs=20, batch_size=32)\n\nwith tf.keras.backend.get_session() as sess:\n    saver = tf.train.Saver()\n    meta_graph_def = tf.train.export_meta_graph(filename='.\/model.meta')\n    save_path = saver.save(sess, \".\/model.ckpt\")\n<\/code><\/pre>\n\n<p>Now you can create the same model using keras and load weights using tensorflow as follows. <\/p>\n\n<pre><code>inputs = tf.keras.layers.Input(shape=(1,), name=\"inputs\")\noutputs = tf.keras.layers.Dense(1, activation=\"linear\", name=\"outputs\")(inputs)\nmodel = tf.keras.models.Model(inputs=inputs, outputs=outputs)\n\nsess = tf.keras.backend.get_session()\nsaver = tf.train.Saver()\nsaver.restore(sess, \".\/model.ckpt\")\n<\/code><\/pre>\n\n<p>Now you can use your model for predicting or what ever you want.<\/p>\n\n<pre><code>print(model.predict([10, 2,4,5,6]))\n# [[8.000007 ]\n# [4.0000067]\n# [5.0000067]\n# [5.5000067]\n# [6.0000067]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53497357,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T10:14:55.967","AcceptedAnswerId":53498263.0,"OwnerUserId":3971621.0,"Title":"Tensorflow: Compute inverse counts of labels","Body":"<p>I have a tensor <code>labels<\/code> of shape (b,n) of integers with values in [0,1,2,3,4,5].\nI want to create a tensor <code>weights<\/code> of shape (b,n) that carries at position (i,j) the inverse of the number of times that the integer <code>labels[i,j]<\/code> occurs in labels. <\/p>\n\n<p>Working example code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\ntf.InteractiveSession()\nlabels=tf.convert_to_tensor(np.array([[1,0,0,1,2,4],[2,2,2,4,2,1]]), dtype=tf.int32)\nweights=tf.ones_like(labels, dtype=tf.float32)\nbc=tf.bincount(labels, minlength=6, maxlength=6)\nfor i in range(6):\n    cur_count = 1.0\/(1e-10+tf.cast(bc[i], tf.float32))\n    count_tensor = tf.ones_like(labels, dtype=tf.float32)*cur_count\n    weights = tf.where(tf.equal(labels,i), count_tensor, weights)\nweights.eval()\n# array([[0.3333, 0.5, 0.5, 0.3333, 0.2, 0.5],\n#        [0.2, 0.2, 0.2, 0.5, 0.2, 0.3333]], dtype=float32)\n<\/code><\/pre>\n\n<p>So for example, the label <code>1<\/code> appears three times in the <code>labels<\/code> tensor so that in <code>weights<\/code> the value 1\/3 appears at each position, where a <code>1<\/code> is in <code>labels<\/code>.<\/p>\n\n<p>Now what I don't like about this code is that tf.bincount does not work on the GPU in my tensorflow version (1.4.0) and I cannot update.\nAlso, I am not sure how the for loop is handled by tensorflow and how much overhead is created by this. <\/p>\n\n<p>I suppose that there is a more elegant solution to my problem. Any ideas?<\/p>\n","answers":[{"AnswerId":"53498263","CreationDate":"2018-11-27T11:04:25.193","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>About the loop, you can replace it with a call to <code>tf.gather<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntf.InteractiveSession()\nlabels = tf.convert_to_tensor(\n    np.array([[1, 0, 0, 1, 2, 4], [2, 2, 2, 4, 2, 1]]), dtype=tf.int32)\nbc = tf.bincount(labels, minlength=6, maxlength=6)\nweights = tf.gather(1.0 \/ (1e-10 + tf.cast(bc, tf.float32)), labels)\nprint(weights.eval())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>[[0.33333334 0.5        0.5        0.33333334 0.2        0.5       ]\n [0.2        0.2        0.2        0.5        0.2        0.33333334]]\n<\/code><\/pre>\n\n<p>About <code>tf.bincount<\/code> being CPU-only, that does not seem to be the case currently. In fact, a GPU implementation seems to have been available <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.5.0\/tensorflow\/core\/kernels\/bincount_op_gpu.cu.cc\" rel=\"nofollow noreferrer\">since v1.5.0<\/a>.<\/p>\n\n<p>If you wanted an alternative implementation, you could do something like this:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ntf.InteractiveSession()\nlabels = tf.convert_to_tensor(\n    np.array([[1, 0, 0, 1, 2, 4], [2, 2, 2, 4, 2, 1]]), dtype=tf.int32)\neq = tf.equal(labels[:, :, tf.newaxis], tf.range(6, dtype=labels.dtype))\nbc = tf.reduce_sum(tf.cast(eq, tf.float32), axis=[0, 1])\nweights = tf.gather(1.0 \/ (1e-10 + tf.cast(bc, tf.float32)), labels)\nprint(weights.eval())\n# Same output\n<\/code><\/pre>\n\n<p>However <code>tf.bincount<\/code> is likely to be more efficient than this.<\/p>\n"}]}
+{"QuestionId":53497902,"AnswerCount":1,"Tags":"<keras><embedding><attention-model>","CreationDate":"2018-11-27T10:45:12.153","AcceptedAnswerId":null,"OwnerUserId":863713.0,"Title":"Keras how to add an attention layer for a weighted sum","Body":"<p>I have the following a network architecture (only the relevant part of the network is shown below) <\/p>\n\n<pre><code>vocab_dimension = 1500\nembed_dimension = 10\n\n\nx = [Input(shape=(None, ), name='input', dtype='int32'),\n     Input(shape=(None, ), name='weights'),\n     Input(shape=(None, ), name='neg_examples', dtype='int32')]\n\n\nembedding_layer = Embedding(input_dim=vocab_dimension, output_dim=embed_dimension)\n\n\ndef _weighted_sum(x):\n    return K.sum(x[0] * K.expand_dims(x[1], -1), axis=1, keepdims=True)\n\n\nweighted_sum = Lambda(_weighted_sum, name='weighted_sum')\n\nitem_vecs = embedding_layer(x[2])\nuser_vecs = weighted_sum([embedding_layer(x[0]), x[1]])\n<\/code><\/pre>\n\n<p>The problem here is that I would like to not pass the weights as input, but I would like to 'learn' them, like in an attention layer.<\/p>\n\n<p>I know that attention layers could be created this way<\/p>\n\n<pre><code>attention_probs = Dense(h, activation='softmax', name='attention_probs')(x[0])\nweighted_sum = Lambda(_weighted_sum)([x[0], attention_probs])\n<\/code><\/pre>\n\n<p><code>h<\/code> is equal to the dimension of the length of the input, which I set to 5. However, if I do the above I get the following error\n<code>TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'<\/code><\/p>\n\n<p>I think this has to do with the dimensions of the input parameters, but I'm not sure how to fix this.<\/p>\n","answers":[{"AnswerId":"53500164","CreationDate":"2018-11-27T12:51:45.303","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>If you fix <code>h<\/code> you'll have to fix the length of the input. You put it to be <code>None<\/code> but your dimensions are going to match only in case of <code>h==input_size<\/code>. It strange that you even can pass undefined size to Dense layer usually you'll get error like <code>ValueError: The last dimension of the inputs to 'Dense' should be defined. Found 'None'.<\/code> <\/p>\n\n<p>Note that in the Input layer you do not have batch_size. So if your backend tensor should be of size (batch_size,dim1) you create input like <code>Input((dim1,))<\/code><\/p>\n"}]}
+{"QuestionId":53498952,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><mpi>","CreationDate":"2018-11-27T11:45:54.993","AcceptedAnswerId":null,"OwnerUserId":8255399.0,"Title":"TensorFlow Horovod: NCCL and MPI","Body":"<p><a href=\"https:\/\/github.com\/uber\/horovod\" rel=\"nofollow noreferrer\">Horovod<\/a> is combining <a href=\"https:\/\/developer.nvidia.com\/nccl\" rel=\"nofollow noreferrer\">NCCL<\/a> and MPI into an wrapper for Distributed Deep Learning in for example TensorFlow.\nI haven't head of NCCL previously and was looking into the functionality and from the Nvidia website they state the following about NCCL:<\/p>\n\n<blockquote>\n  <p>The NVIDIA Collective Communications Library (NCCL) implements multi-GPU and multi-node collective communication primitives that are performance optimized for NVIDIA GPUs. <\/p>\n<\/blockquote>\n\n<p>From the <a href=\"http:\/\/on-demand.gputechconf.com\/gtc\/2018\/video\/S8462\/\" rel=\"nofollow noreferrer\">introduction video<\/a> about NCCL I understood that NCCL works via PCIe, NVLink, Native Infiniband, Ethernet and it can even detect if GPU Direct via RDMA makes sense in the current hardware topology and uses it transparently.<\/p>\n\n<p>So I am questioning why MPI is needed in Horovod? As far as I understand MPI is also used for efficiently exchanging the gradients among distributed nodes via an allreduce paradigm. But as I understand NCCL already supports those functionalities. <\/p>\n\n<p>So is MPI only used for easily scheduling the jobs on a cluster? For Distributed Deep Learning on CPU, since we cannot use NCCL there? <\/p>\n\n<p>I would highly appreciate if someone could explain in which scenarios MPI and\/or NCCL is used for Distributed Deep Learning and what are their responsibilities during the training job.<\/p>\n","answers":[]}
+{"QuestionId":53499321,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T12:06:13.430","AcceptedAnswerId":null,"OwnerUserId":10579805.0,"Title":"Could not find a version that satisfies the requirement tensorflow for python 3.6.7","Body":"<pre><code>C:\\WINDOWS\\system32&gt;python --version\nPython 3.6.7\n\nC:\\WINDOWS\\system32&gt;pip3 list\nPackage    Version\n---------- -------\npip        10.0.1\nsetuptools 39.0.1\n\nC:\\WINDOWS\\system32&gt;PATH\nPATH=C:\\Program Files (x86)\\Common Files\\Oracle\\Java\\javapath;C:\\HighTec\\toolchains\\tricore\\v4.9.1.0-infineon-1.1\\bin;C:\\HighTec\\licensemanager;C:\\WINDOWS\\system32;C:\\WINDOWS;C:\\WINDOWS\\System32\\Wbem;C:\\WINDOWS\\System32\\WindowsPowerShell\\v1.0\\;C:\\Program Files (x86)\\WebEx\\Productivity Tools;C:\\Program Files\\Micro Focus\\StarTeam SDK 16.0\\lib;C:\\Program Files\\Micro Focus\\StarTeam SDK 16.0\\bin;C:\\Program Files\\Git\\cmd;c:\\users\\snaglapu\\appdata\\local\\programs\\python\\python36\\;C:\\Users\\snaglapu\\AppData\\Local\\Programs\\Python\\Python36\\Scripts\\;C:\\Users\\snaglapu\\AppData\\Local\\Programs\\Python\\Python36\\;C:\\Users\\snaglapu\\AppData\\Local\\Microsoft\\WindowsApps;C:\\Users\\snaglapu\\AppData\\Local\\Android\\Sdk\\platform-tools\n\nC:\\WINDOWS\\system32&gt;py -m pip install --upgrade pip\nRetrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001A3A6CDFC50&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/pip\/\nRetrying (Retry(total=3, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001A3A6CDFDD8&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/pip\/\nRetrying (Retry(total=2, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001A3A6CDFD68&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/pip\/\nRetrying (Retry(total=1, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001A3A6CDF160&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/pip\/\nRetrying (Retry(total=0, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001A3A6CDFEB8&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/pip\/\nRequirement already up-to-date: pip in c:\\users\\snaglapu\\appdata\\local\\programs\\python\\python36\\lib\\site-packages (10.0.1)\n\nC:\\WINDOWS\\system32&gt;pip3 install --upgrade tensorflow\nCollecting tensorflow\nRetrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001F1F7B99470&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/tensorflow\/\nRetrying (Retry(total=3, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001F1F7BCD4E0&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/tensorflow\/\nRetrying (Retry(total=2, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001F1F7BCD588&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/tensorflow\/\nRetrying (Retry(total=1, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001F1F7BCD7B8&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/tensorflow\/\nRetrying (Retry(total=0, connect=None, read=None, redirect=None, status=None)) after connection broken by 'NewConnectionError('&lt;pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0x000001F1F7BCD2B0&gt;: Failed to establish a new connection: [Errno 11001] getaddrinfo failed',)': \/simple\/tensorflow\/\n\nCould not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow\n\nC:\\WINDOWS\\system32&gt;\n<\/code><\/pre>\n\n<p>I am trying to install tensorflow on top of python 3.6.7. I am getting the above mentioned error message. What could be the problem? I uninstalled python 3.6.5 and reinstalled 3.6.7.Also, tried upgrading pip.Both of them didn't work for me.I tried every possible solutions mentioned in the following link. <a href=\"https:\/\/stackoverflow.com\/questions\/48720833\/could-not-find-a-version-that-satisfies-the-requirement-tensorflow\">Could not find a version that satisfies the requirement tensorflow<\/a><\/p>\n\n<pre><code> C:\\WINDOWS\\system32&gt;python -m pip install --upgrade pip --proxy http:\/\/snaglapu:******** @&lt;Proxy_name&gt;:&lt;port&gt;--user\nCollecting pip\n Downloading https:\/\/files.pythonhosted.org\/packages\/c2\/d7\/90f34cb0d83a6c5631cf71dfe64cc1054598c843a92b400e55675cc2ac37\/pip-18.1-py2.py3-none-any.whl (1.3MB)\n100% |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 1.3MB 819kB\/s\nInstalling collected packages: pip\nSuccessfully installed pip-18.1\n\nC:\\WINDOWS\\system32&gt;pip list\nTraceback (most recent call last):\nFile \"c:\\users\\snaglapu\\appdata\\local\\programs\\python\\python36\\lib\\runpy.py\", line 193, in _run_module_as_main\n\"__main__\", mod_spec)\nFile \"c:\\users\\snaglapu\\appdata\\local\\programs\\python\\python36\\lib\\runpy.py\", line 85, in _run_code\nexec(code, run_globals)\nFile \"C:\\Users\\snaglapu\\AppData\\Local\\Programs\\Python\\Python36\\Scripts\\pip.exe\\__main__.py\", line 5, in &lt;module&gt;\n<\/code><\/pre>\n\n<p>ImportError: cannot import name 'main'<\/p>\n\n<p>after running the code behind the proxy error is popping when trying to read pip3 version or read the list of pip3<\/p>\n","answers":[]}
+{"QuestionId":53499409,"AnswerCount":2,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2018-11-27T12:11:02.487","AcceptedAnswerId":53504855.0,"OwnerUserId":2455494.0,"Title":"Tensorflow TFRecord: Can't parse serialized example","Body":"<p>I am trying to follow <a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/tf-records\" rel=\"noreferrer\">this guide<\/a> in order to serialize my input data into the TFRecord format but I keep hitting this error when trying to read it:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Key: my_key.  Can't parse serialized Example.<\/p>\n<\/blockquote>\n\n<p>I am not sure where I'm going wrong. Here is a minimal reproduction of the issue I cannot get past.<\/p>\n\n<p><strong>Serialise some sample data:<\/strong><\/p>\n\n<pre><code>with tf.python_io.TFRecordWriter('train.tfrecords') as writer:\n  for idx in range(10):\n        example = tf.train.Example(\n            features=tf.train.Features(\n                feature={\n                    'label': tf.train.Feature(int64_list=tf.train.Int64List(value=[1,2,3])),\n                    'test': tf.train.Feature(float_list=tf.train.FloatList(value=[0.1,0.2,0.3])) \n                }\n            )\n        )\n\n        writer.write(example.SerializeToString())\n  writer.close()\n<\/code><\/pre>\n\n<p><strong>Parsing function &amp; deserialise:<\/strong><\/p>\n\n<pre><code>def parse(tfrecord):\n  features = {\n      'label': tf.FixedLenFeature([], tf.int64, default_value=0),\n      'test': tf.FixedLenFeature([], tf.float32, default_value=0.0),\n  }\n  return tf.parse_single_example(tfrecord, features)\n\ndataset = tf.data.TFRecordDataset('train.tfrecords').map(parse)\ngetnext = dataset.make_one_shot_iterator().get_next()\n<\/code><\/pre>\n\n<p>When trying to run this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  v = sess.run(getnext)\n  print (v)\n<\/code><\/pre>\n\n<p>I trigger the above error message. <\/p>\n\n<p>Is it possible to get past this error and deserialize my data?<\/p>\n","answers":[{"AnswerId":"53504855","CreationDate":"2018-11-27T17:15:06.390","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/io\/FixedLenFeature\" rel=\"noreferrer\">tf.FixedLenFeature()<\/a> is used for reading the fixed size arrays of data. And the shape of the data should be defined beforehand. Updating the parse function to<\/p>\n\n<pre><code>def parse(tfrecord):\n   return tf.parse_single_example(tfrecord, features={\n       'label': tf.FixedLenFeature([3], tf.int64, default_value=[0,0,0]),\n       'test': tf.FixedLenFeature([3], tf.float32, default_value=[0.0, 0.0, 0.0]),\n   })\n<\/code><\/pre>\n\n<p>Should do the job.<\/p>\n"},{"AnswerId":"57990442","CreationDate":"2019-09-18T10:18:09.950","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>As an alternative, if your input features lengths are not fixed and are of arbitrary sizes then you can also use <code>tf.io.FixedLenSequenceFeature()<\/code> with arguments <code>allow_missing = True<\/code> and <code>default_value=0<\/code> (in case of type int and 0.0 for float) which does not require the input feature to be of fixed size unlike <code>tf.io.FixedLenFeature()<\/code>. You can find more information <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/io\/FixedLenSequenceFeature\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53500047,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-27T12:45:38.390","AcceptedAnswerId":null,"OwnerUserId":4780863.0,"Title":"Stop Training in Keras when Accuracy is already 1.0","Body":"<p>How will I stop Keras Training when the accuracy already reached 1.0? I tried monitoring loss value, but I haven't tried stopping the training when the accuracy is already 1.<\/p>\n\n<p>I tried the code below with no luck:<\/p>\n\n<pre><code>stopping_criterions =[\n    EarlyStopping(monitor='loss', min_delta=0, patience = 1000),\n    EarlyStopping(monitor='acc', base_line=1.0, patience =0)\n\n]\n\nmodel.summary()\nmodel.compile(Adam(), loss='binary_crossentropy', metrics=['accuracy']) \nmodel.fit(scaled_train_samples, train_labels, batch_size=1000, epochs=1000000, callbacks=[stopping_criterions], shuffle = True, verbose=2)\n<\/code><\/pre>\n\n<p>UPDATE: <\/p>\n\n<p>The training immediately stops st first epoch and even if the accuracy is still not 1.0.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/29Z3Z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/29Z3Z.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Please help.<\/p>\n","answers":[{"AnswerId":"53501868","CreationDate":"2018-11-27T14:26:09.407","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><strong>Update:<\/strong> I don't know why <code>EarlyStopping<\/code> does not work in this case. Instead, I defined a custom callback that stops training when <code>acc<\/code> (or <code>val_acc<\/code>) reaches a specified baseline:<\/p>\n\n<pre><code>from keras.callbacks import Callback\n\nclass TerminateOnBaseline(Callback):\n    \"\"\"Callback that terminates training when either acc or val_acc reaches a specified baseline\n    \"\"\"\n    def __init__(self, monitor='acc', baseline=0.9):\n        super(TerminateOnBaseline, self).__init__()\n        self.monitor = monitor\n        self.baseline = baseline\n\n    def on_epoch_end(self, epoch, logs=None):\n        logs = logs or {}\n        acc = logs.get(self.monitor)\n        if acc is not None:\n            if acc &gt;= self.baseline:\n                print('Epoch %d: Reached baseline, terminating training' % (epoch))\n                self.model.stop_training = True\n<\/code><\/pre>\n\n<p>You can use it like this:<\/p>\n\n<pre><code>callbacks = [TerminateOnBaseline(monitor='acc', baseline=0.8)]\ncallbacks = [TerminateOnBaseline(monitor='val_acc', baseline=0.95)]\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Note: This solution does not work.<\/strong><\/p>\n\n<p>If you want to stop training when the training (or validation) accuracy exactly reaches 100%, then use <code>EarlyStopping<\/code> callback and set the <code>baseline<\/code> argument to 1.0 and <code>patience<\/code> to zero: <\/p>\n\n<pre><code>EarlyStopping(monitor='acc', baseline=1.0, patience=0)  # use 'val_acc' instead to monitor validation accuarcy\n<\/code><\/pre>\n"}]}
+{"QuestionId":53500185,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-lite>","CreationDate":"2018-11-27T12:53:02.227","AcceptedAnswerId":53542804.0,"OwnerUserId":2174040.0,"Title":"Convert Keras MobileNet model to TFLite with 8-bit quantization","Body":"<p>I have used Keras to finetune MobileNet v1. Now I have <code>model.h5<\/code> and I need to convert it to TensorFlow Lite to use it in Android app. <\/p>\n\n<p>I use TFLite conversion <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/cmdline_examples\" rel=\"nofollow noreferrer\">script<\/a> <code>tflite_convert<\/code>. I can convert it without quantization but I need more performance so I need to make quantization.<\/p>\n\n<p>If I run this script:<\/p>\n\n<pre><code>tflite_convert --output_file=model_quant.tflite \\\n --keras_model_file=model.h5 \\\n --inference_type=QUANTIZED_UINT8 \\\n --input_arrays=input_1 \\\n --output_arrays=predictions\/Softmax \\\n --mean_values=128 \\\n --std_dev_values=127 \\\n --input_shape=\"1,224,224,3\" \n<\/code><\/pre>\n\n<p>It fails:<\/p>\n\n<blockquote>\n  <p>F tensorflow\/contrib\/lite\/toco\/tooling_util.cc:1634] Array\n  conv1_relu\/Relu6, which is an input to the DepthwiseConv operator\n  producing the output array conv_dw_1_relu\/Relu6, is lacking min\/max\n  data, which is necessary for quantization. If accuracy matters, either\n  target a non-quantized output format, or run quantized training with\n  your model from a floating point checkpoint to change the input graph\n  to contain min\/max information. If you don't care about accuracy, you\n  can pass --default_ranges_min= and --default_ranges_max= for easy\n  experimentation.\\nAborted (core dumped)\\n\"<\/p>\n<\/blockquote>\n\n<p>If I use <code>default_ranges_min<\/code> and <code>default_ranges_max<\/code> (called as \"dummy-quantization\"), it works but it is only for debugging performance without accuracy as it is described in error log. <\/p>\n\n<p>So what I need to do to make Keras model correctly quantizable? Do I need to find best <code>default_ranges_min<\/code> and <code>default_ranges_max<\/code>? How? Or it is about changes in Keras training phase?<\/p>\n\n<p>Library versions:<\/p>\n\n<pre><code>Python 3.6.4\nTensorFlow 1.12.0\nKeras 2.2.4\n<\/code><\/pre>\n","answers":[{"AnswerId":"53542804","CreationDate":"2018-11-29T15:50:30.200","ParentId":null,"OwnerUserId":"2221231","Title":null,"Body":"<p>Unfortunately, Tensorflow does not provide the tooling for post-training per layer quantization in flatbuffer (tflite) yet, but only in protobuf. The only available way now is to introduce <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/quantization\/fake_quant_with_min_max_vars\" rel=\"nofollow noreferrer\">fakeQuantization<\/a> layers in your graph and re-train \/ fine-tune your model on the train or a calibration set. This is called \"<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/quantize\" rel=\"nofollow noreferrer\">Quantization-aware training<\/a>\".<\/p>\n\n<p>Once the fakeQuant layers are introduced, then you can feed the training set and TF is going to use them on Feed-Forward as simulated quantisation layers (fp-32 datatypes that represent 8-bit values) and back-propagate using full precision values. This way, you can get back the accuracy loss that caused by quantization.<\/p>\n\n<p>In addition, the fakeQuant layers are going to capture the ranges per layer or per channel through moving average and store them in min \/ max variables. <\/p>\n\n<p>Later, you can extract the graph definition and get rid of the fakeQuant nodes through <code>freeze_graph<\/code> tool.<\/p>\n\n<p>Finally, the model can be fed into tf_lite_converter (cross-fingers it won't brake) and extract the u8_tflite with captured ranges.<\/p>\n\n<p>A very good white-paper, explaining all these is provided by Google here : <a href=\"https:\/\/arxiv.org\/pdf\/1806.08342.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1806.08342.pdf<\/a><\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
+{"QuestionId":53500264,"AnswerCount":1,"Tags":"<keras><neural-network><deep-learning>","CreationDate":"2018-11-27T12:57:22.287","AcceptedAnswerId":null,"OwnerUserId":8924389.0,"Title":"my simple regression model(by keras) doesn't work","Body":"<p>i'm beginner studying neural network.\ni'm doing kaggle project - bike sharing demand.\ni want to use simple neural network by keras, but loss is not decreasing.\nwhat should i do?<\/p>\n\n<p>-----------code--------------<\/p>\n\n<pre><code># dataset from pandas\n\nfeature_names = [\"season\", \"holiday\", \"workingday\", \"weather\",\n             \"temp\", \"atemp\", \"humidity\", \"windspeed\",\n             \"datetime_year\", \"datetime_hour\", \"datetime_dayofweek\"]\nlabel_name = [\"count\"]\nX_train = train[feature_names]   #shape (10886, 11)\nY_train = train[label_name]      #shape (10886, 1)\nX_test = test[feature_names]\n\n# layers\n\nmodel = Sequential()\nmodel.add(Dense(units = 50, kernel_initializer = 'uniform', activation = 'relu', input_dim=11))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units = 50, kernel_initializer = 'uniform', activation = 'relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units = 5, kernel_initializer = 'uniform', activation = 'relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n\n\nmodel.compile(optimizer = 'adam', loss = 'mean_squared_error', metrics = ['accuracy'])\n\n# Train\nmodel.fit(X_train, Y_train, batch_size = 100, epochs = 200)\n<\/code><\/pre>\n\n<p>---------result------------<\/p>\n\n<pre><code>Epoch 1\/200\n10886\/10886 [==============================] - 4s 325us\/step - loss: 69206.2478 - acc: 0.0094\nEpoch 2\/200\n10886\/10886 [==============================] - 1s 93us\/step - loss: 69184.5435 - acc: 0.0096\nEpoch 3\/200\n10886\/10886 [==============================] - 1s 89us\/step - loss: 69181.6330 - acc: 0.0096\nEpoch 4\/200\n10886\/10886 [==============================] - 1s 93us\/step - loss: 69179.0222 - acc: 0.0096\nEpoch 5\/200\n10886\/10886 [==============================] - 1s 91us\/step - loss: 69175.7442 - acc: 0.0096\nEpoch 6\/200\n10886\/10886 [==============================] - 1s 109us\/step - loss: 69171.9052 - acc: 0.0096\nEpoch 7\/200\n10886\/10886 [==============================] - 1s 122us\/step - loss: 69171.6164 - acc: 0.0096\nEpoch 8\/200\n10886\/10886 [==============================] - 1s 92us\/step - loss: 69167.6923 - acc: 0.0096\nEpoch 9\/200\n10886\/10886 [==============================] - 1s 91us\/step - loss: 69166.2911 - acc: 0.0096\nEpoch 10\/200\n10886\/10886 [==============================] - 1s 94us\/step - loss: 69164.1145 - acc: 0.0096\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"53500503","CreationDate":"2018-11-27T13:10:32.553","ParentId":null,"OwnerUserId":"3114765","Title":null,"Body":"<p>Try setting the learning rate in the <code>model.compile<\/code> section. Start with <code>0.0001<\/code>, then <code>0.001<\/code> and <code>0.01<\/code> and see what happens.<\/p>\n"}]}
+{"QuestionId":53500375,"AnswerCount":1,"Tags":"<tensorflow><object-detection>","CreationDate":"2018-11-27T13:03:41.980","AcceptedAnswerId":null,"OwnerUserId":10698687.0,"Title":"Why its result is very poor using ssd_mobilenet_v1_pnp with my own dataset?","Body":"<p>Tensorflow 1.12.0<\/p>\n\n<p>I am currently trying to train SSD_Mobilenet_V1_pnp model (pre-trained with COCO) with my dataset. My data set has 490 images for training, and 210 for evaluating, 23 classes<\/p>\n\n<p>label_map.pbtxt:<\/p>\n\n<p>item {\n  id: 1\n  name: 'a'\n}\nitem {\n  id: 2\n  name: 'b'\n}<\/p>\n\n<p>...<\/p>\n\n<p>pipeline.config:<\/p>\n\n<pre><code>model {\n  ssd {\n    num_classes: 24\n    image_resizer {\n      fixed_shape_resizer {\n        height: 300\n        width: 300\n      }\n    }\n    feature_extractor {\n      type: \"ssd_mobilenet_v1_ppn\"\n      conv_hyperparams {\n        regularizer {\n          l2_regularizer {\n            weight: 3.99999989895e-05\n          }\n        }\n        initializer {\n          random_normal_initializer {\n            mean: 0.0\n            stddev: 0.00999999977648\n          }\n        }\n        activation: RELU_6\n        batch_norm {\n          decay: 0.97000002861\n          center: true\n          scale: true\n          epsilon: 0.0010000000475\n        }\n      }\n      override_base_feature_extractor_hyperparams: true\n    }\n    box_coder {\n      faster_rcnn_box_coder {\n        y_scale: 10.0\n        x_scale: 10.0\n        height_scale: 5.0\n        width_scale: 5.0\n      }\n    }\n    matcher {\n      argmax_matcher {\n        matched_threshold: 0.5\n        unmatched_threshold: 0.5\n        ignore_thresholds: false\n        negatives_lower_than_unmatched: true\n        force_match_for_each_row: true\n        use_matmul_gather: true\n      }\n    }\n    similarity_calculator {\n      iou_similarity {\n      }\n    }\n    box_predictor {\n      weight_shared_convolutional_box_predictor {\n        conv_hyperparams {\n          regularizer {\n            l2_regularizer {\n              weight: 3.99999989895e-05\n            }\n          }\n          initializer {\n            random_normal_initializer {\n              mean: 0.0\n              stddev: 0.00999999977648\n            }\n          }\n          activation: RELU_6\n          batch_norm {\n            decay: 0.97000002861\n            center: true\n            scale: true\n            epsilon: 0.0010000000475\n            train: true\n          }\n        }\n        depth: 512\n        num_layers_before_predictor: 1\n        kernel_size: 1\n        class_prediction_bias_init: -4.59999990463\n        share_prediction_tower: true\n      }\n    }\n    anchor_generator {\n      ssd_anchor_generator {\n        num_layers: 6\n        min_scale: 0.15000000596\n        max_scale: 0.949999988079\n        aspect_ratios: 1.0\n        aspect_ratios: 2.0\n        aspect_ratios: 0.5\n        aspect_ratios: 3.0\n        aspect_ratios: 0.333299994469\n        reduce_boxes_in_lowest_layer: false\n      }\n    }\n    post_processing {\n      batch_non_max_suppression {\n        score_threshold: 0.300000011921\n        iou_threshold: 0.600000023842\n        max_detections_per_class: 100\n        max_total_detections: 100\n      }\n      score_converter: SIGMOID\n    }\n    normalize_loss_by_num_matches: true\n    loss {\n      localization_loss {\n        weighted_smooth_l1 {\n        }\n      }\n      classification_loss {\n        weighted_sigmoid_focal {\n          gamma: 2.0\n          alpha: 0.75\n        }\n      }\n      classification_weight: 1.0\n      localization_weight: 1.5\n    }\n    encode_background_as_zeros: true\n    normalize_loc_loss_by_codesize: true\n    inplace_batchnorm_update: true\n    freeze_batchnorm: false\n  }\n}\ntrain_config {\n  batch_size: 512\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n  data_augmentation_options {\n    ssd_random_crop {\n    }\n  }\n  sync_replicas: true\n  optimizer {\n    momentum_optimizer {\n      learning_rate {\n        cosine_decay_learning_rate {\n          learning_rate_base: 0.699999988079\n          total_steps: 50000\n          warmup_learning_rate: 0.13330000639\n          warmup_steps: 2000\n        }\n      }\n      momentum_optimizer_value: 0.899999976158\n    }\n    use_moving_average: false\n  }\n  fine_tune_checkpoint: \"model.ckpt\"\n  num_steps: 50000\n  startup_delay_steps: 0.0\n  replicas_to_aggregate: 8\n  max_number_of_boxes: 100\n  unpad_groundtruth_tensors: false\n  from_detection_checkpoint: true\n}\ntrain_input_reader {\n  label_map_path: \"annotations\\label_map.pbtxt\"\n  tf_record_input_reader {\n    input_path: \"train.record\"\n  }\n}\neval_config {\n  num_examples: 210\n  max_evals: 10\n  metrics_set: \"coco_detection_metrics\"\n  use_moving_averages: false\n}\neval_input_reader {\n  label_map_path: \"annotations\\label_map.pbtxt\"\n  shuffle: false\n  num_epochs: 1\n  num_readers: 1\n  tf_record_input_reader {\n    input_path: \"val.record\"\n  }\n}\n<\/code><\/pre>\n\n<p>train:<\/p>\n\n<pre><code>python object_detection\/model_main.py --logtostderr --pipeline_config_path=pipeline.config --model_dir=train\n<\/code><\/pre>\n\n<p>log:<\/p>\n\n<pre><code>Accumulating evaluation results...\nDONE (t=0.05s).\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000\nAverage Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000\nAverage Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000\nAverage Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.000\n<\/code><\/pre>\n\n<p>Is this a normal?\nHow to solve?<\/p>\n","answers":[{"AnswerId":"53524197","CreationDate":"2018-11-28T16:39:27.120","ParentId":null,"OwnerUserId":"2383106","Title":null,"Body":"<p>Few things I notice that may help you out:<\/p>\n\n<p>num_classes in your config file is at 24 but you are training 23 classes.<\/p>\n\n<p>Also see you are using a fixed image resizer, depending on the dimensions of your photos this may be an issue since you are not maintaining aspect ratio.<\/p>\n\n<p>Might see slight improvement by training for fewer steps (20k) since your data set is fairly small.<\/p>\n\n<p>If none of this helps out would look into adding a hard data miner argument to config file to introduce a minimum number of negative examples.<\/p>\n"}]}
+{"QuestionId":53500511,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2018-11-27T13:11:11.687","AcceptedAnswerId":53501869.0,"OwnerUserId":8733369.0,"Title":"ValueError: expected 2D or 3D input (got 1D input) PyTorch","Body":"<pre><code>class VAE(torch.nn.Module):\n\n def __init__(self, input_size, hidden_sizes, batch_size):\n\n    super(VAE, self).__init__()\n\n    self.input_size = input_size\n    self.hidden_sizes = hidden_sizes\n    self.batch_size = batch_size\n    self.fc = torch.nn.Linear(input_size, hidden_sizes[0])\n    self.BN = torch.nn.BatchNorm1d(hidden_sizes[0])\n    self.fc1 = torch.nn.Linear(hidden_sizes[0], hidden_sizes[1])\n    self.BN1 = torch.nn.BatchNorm1d(hidden_sizes[1])\n    self.fc2 = torch.nn.Linear(hidden_sizes[1], hidden_sizes[2])\n    self.BN2 = torch.nn.BatchNorm1d(hidden_sizes[2])\n    self.fc3_mu = torch.nn.Linear(hidden_sizes[2], hidden_sizes[3])\n    self.fc3_sig = torch.nn.Linear(hidden_sizes[2], hidden_sizes[3])\n\n    self.fc4 = torch.nn.Linear(hidden_sizes[3], hidden_sizes[2])\n    self.BN4 = torch.nn.BatchNorm1d(hidden_sizes[2])\n    self.fc5 = torch.nn.Linear(hidden_sizes[2], hidden_sizes[1])\n    self.BN5 = torch.nn.BatchNorm1d(hidden_sizes[1])\n    self.fc6 = torch.nn.Linear(hidden_sizes[1], hidden_sizes[0])\n    self.BN6 = torch.nn.BatchNorm1d(hidden_sizes[0])\n    self.fc7 = torch.nn.Linear(hidden_sizes[0], input_size)\n\ndef sample_z(self, x_size, mu, log_var):\n\n     eps = torch.randn(x_size, self.hidden_sizes[-1])\n     return(mu + torch.exp(log_var\/2) * eps)\n\n def forward(self, x):\n\n    ###########\n    # Encoder #\n    ###########\n\n    out1 = self.fc(x)\n    out1 = nn.relu(self.BN(out1))\n    out2 = self.fc1(out1)\n    out2 = nn.relu(self.BN1(out2))\n    out3 = self.fc2(out2)\n    out3 = nn.relu(self.BN2(out3))\n    mu = self.fc3_mu(out3)\n    sig = nn.softplus(self.fc3_sig(out3))\n\n    ###########\n    # Decoder  #\n    ###########\n\n    # sample from the distro\n    sample = self.sample_z(x.size(0), mu, sig)\n    out4 = self.fc4(sample)\n    out4 = nn.relu(self.BN4(out4))\n    out5 = self.fc5(out4)\n    out5 = nn.relu(self.BN5(out5))\n    out6 = self.fc6(out5)\n    out6 = nn.relu(self.BN6(out6))\n    out7 = nn.sigmoid(self.fc7(out6))\n\n    return(out7, mu, sig)\n\nvae = VAE(input_size, hidden_sizes, batch_size)\n\nvae.eval()\n\nx_sample, z_mu, z_var = vae(X)\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<blockquote>\n<pre><code>File \"VAE_LongTensor.py\", line 200, in &lt;module&gt;    \nx_sample, z_mu, z_var = vae(X)      \nValueError: expected 2D or 3D input (got 1D input)\n<\/code><\/pre>\n<\/blockquote>\n","answers":[{"AnswerId":"53501869","CreationDate":"2018-11-27T14:26:10.340","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When you build a <code>nn.Module<\/code> in pytorch for processing 1D signals, pytorch actually expects the input to be 2D: first dimension is the \"mini batch\" dimension.<br>\nThus you need to add a singleton dimesion to your <code>X<\/code>:<\/p>\n\n<pre><code>x_sample, z_mu, z_var = vae(X[None, ...])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53500684,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-27T13:20:59.623","AcceptedAnswerId":null,"OwnerUserId":1946465.0,"Title":"Dataset from generator that yields multiple elements at a time","Body":"<p>I'm testing the waters if it's time to migrate to the Dataset API in TensorFlow from the deprecated queue-based API.<\/p>\n\n<p>One use case I can't seem to find the equivalent for is the <code>enqueue_many<\/code> parameter of <code>tf.train.batch<\/code>.<\/p>\n\n<p>In particular I would like to create a Python generator that can yield \"batched\" arrays, where the \"batch size\" is not necessarily the same as the one used for SGD training updates and then apply batching on that stream of data (i.e. as it is with enqueue_many in tf.train.batch).<\/p>\n\n<p>Is there any workaround to achieve this in the new Dataset API?<\/p>\n","answers":[{"AnswerId":"54019020","CreationDate":"2019-01-03T08:55:36.467","ParentId":null,"OwnerUserId":"3534616","Title":null,"Body":"<p>Try using flatmap<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nn_reads=10\nread_batch_size=20\ntraining_batch_size = 2\n\ndef mnist_gen():\n    mnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n    for i in range(n_reads):\n        batch_x, batch_y = mnist.train.next_batch(read_batch_size)\n        # Yielding a batch instead of single record\n        yield batch_x,batch_y\ndata = tf.data.Dataset.from_generator(mnist_gen,output_types=(tf.float32,tf.float32))\ndata = data.flat_map(lambda *x: tf.data.Dataset.zip(tuple(map(tf.data.Dataset.from_tensor_slices,x)))).batch(training_batch_size)\n# if u yield only batch_x change lambda function to data.flat_map(lambda x: tf.data.Dataset.from_tensor_slices(x)))\niter = data.make_one_shot_iterator()\nnext_item = iter.get_next()\n\nX= next_item[0]\nY = next_item[1]\n\nwith tf.Session() as sess:\n    for i in range(n_reads*read_batch_size \/\/ training_batch_size):\n        print(i, sess.run(X))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53500701,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-27T13:21:58.713","AcceptedAnswerId":null,"OwnerUserId":7735772.0,"Title":"Face recognition with keras","Body":"<p>Below is my face recognition model. I get several issues while training my training data on it. My dataset contains images of me. When I train it, validation accuracy is 100%. And also its prediction is bad. What can I do to solve this problem?<\/p>\n\n<pre><code>from keras import layers\nfrom keras import models \n\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32,(3,3),activation='relu',\n    input_shape = (150,150,3)))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Conv2D(64,(3,3),activation='relu'))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Conv2D(128,(3,3),activation='relu'))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Dropout(0.5))\nmodel.add(layers.Conv2D(128,(3,3),activation='relu'))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(512,activation='relu'))\nmodel.add(layers.Dense(1,activation='sigmoid'))\n\nprint(model.summary())\n\nfrom keras import optimizers\n\nmodel.compile(loss='binary_crossentropy',\n    optimizer=optimizers.RMSprop(lr=1e-4),\n    metrics=['acc'])\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255)\nval_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_dir,\n    target_size = (150,150),\n    batch_size=20)\n\nvalidation_generator = val_datagen.flow_from_directory(\n    validation_dir,\n    target_size = (150,150),\n    batch_size=20)\n\nhistory = model.fit_generator(\ntrain_generator,\nsteps_per_epoch = 100,\nepochs = 3,\nvalidation_data = validation_generator,\nvalidation_steps = 50)\n\nmodel.save('\/home\/monojit\/Desktop\/me3.h5')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53505798","CreationDate":"2018-11-27T18:16:55.953","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>How large is the dataset that is being used? A small dataset might be the problem, or your model architecture if the model is not generalizing well. Also, you might look at image augmentation using the <code>ImageDataGenerator<\/code>, see the blog post that i'll link to in a bit.<\/p>\n\n<p>If the purpose of this project is to get as high accuracy as possible, without explicitly learning how CNN and their different layers work, then i would suggest the following. As you are working with images, you might not want to re-invent the wheel. Go on and take a pre-trained convolutional neural network, then train that on your imaged. This will yield you way higher accuracy with less epochs, than an untrained network. A great blog post can be found here <a href=\"https:\/\/jkjung-avt.github.io\/keras-tutorial\/\" rel=\"nofollow noreferrer\">Keras Cats vs dogs<\/a>.\nNow this tutorial is cats vs dogs. But you can use (almost, depending on your input images) the exact same code for your problem.<\/p>\n"}]}
+{"QuestionId":53500838,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><pytorch>","CreationDate":"2018-11-27T13:29:22.300","AcceptedAnswerId":53501763.0,"OwnerUserId":8778755.0,"Title":"Pytorch RuntimeError: size mismatch, m1: [1 x 7744], m2: [400 x 120]","Body":"<p>In a simple CNN that classifies 5 objects, I get a size mis-match error:<\/p>\n\n<pre><code>\"RuntimeError: size mismatch, m1: [1 x 7744], m2: [400 x 120]\" in the convolutional layer . \n<\/code><\/pre>\n\n<p>my model.py file:<\/p>\n\n<pre><code>import torch.nn as nn\nimport torch.nn.functional as F\n\nclass FNet(nn.Module):\n\n\n    def __init__(self,device):\n        # make your convolutional neural network here\n        # use regularization\n        # batch normalization\n        super(FNet, self).__init__()\n        num_classes = 5\n        self.conv1 = nn.Conv2d(3, 6, 5)\n        self.conv2 = nn.Conv2d(6, 16, 5)\n        # an affine operation: y = Wx + b\n        self.fc1 = nn.Linear(16 * 5 * 5, 120)\n        self.fc2 = nn.Linear(120, 84)\n        self.fc3 = nn.Linear(84, 5)\n\n    def forward(self, x):\n\n        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))\n\n        x = F.max_pool2d(F.relu(self.conv2(x)), 2)\n        x = x.view(-1, self.num_flat_features(x))\n        x = F.relu(self.fc1(x))\n        x = F.relu(self.fc2(x))\n        x = self.fc3(x)\n        return x\n\n\n    def num_flat_features(self, x):\n        size = x.size()[1:]  # all dimensions except the batch dimension\n        num_features = 1\n        for s in size:\n            num_features *= s\n        return num_features\n\nif __name__ == \"__main__\":\n    net = FNet()\n<\/code><\/pre>\n\n<p>Complete Error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"main.py\", line 98, in &lt;module&gt;\n    train_model('..\/Data\/fruits\/', save=True, destination_path='\/home\/mitesh\/E yantra\/task1#hc\/Task 1\/Task 1B\/Data\/fruits')\n  File \"main.py\", line 66, in train_model\n    outputs = model(images)\n  File \"\/home\/mitesh\/anaconda3\/envs\/HC#850_stage1\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/mitesh\/E yantra\/task1#hc\/Task 1\/Task 1B\/Code\/model.py\", line 28, in forward\n    x = F.relu(self.fc1(x))\n  File \"\/home\/mitesh\/anaconda3\/envs\/HC#850_stage1\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/mitesh\/anaconda3\/envs\/HC#850_stage1\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/linear.py\", line 55, in forward\n    return F.linear(input, self.weight, self.bias)\n  File \"\/home\/mitesh\/anaconda3\/envs\/HC#850_stage1\/lib\/python3.6\/site-packages\/torch\/nn\/functional.py\", line 1024, in linear\n    return torch.addmm(bias, input, weight.t())\nRuntimeError: size mismatch, m1: [1 x 7744], m2: [400 x 120] at \/opt\/conda\/conda-bld\/pytorch-cpu_1532576596369\/work\/aten\/src\/TH\/generic\/THTensorMath.cpp:2070\n<\/code><\/pre>\n","answers":[{"AnswerId":"53501763","CreationDate":"2018-11-27T14:20:20.177","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you have a <code>nn.Linear<\/code> layer in your net, you cannot decide \"on-the-fly\" what the input size for this layer would be.<br>\nIn your net you compute <code>num_flat_features<\/code> for every <code>x<\/code> and expect your <code>self.fc1<\/code> to handle whatever size of <code>x<\/code> you feed the net. However, <code>self.fc1<\/code> has a <strong>fixed size<\/strong> weight matrix of size 400x120 (that is expecting input of dimension 16*5*5=400 and outputs 120 dim feature). In your case the size of <code>x<\/code> translated to 7744 dim feature vector that <code>self.fc1<\/code> simply cannot handle. <\/p>\n\n<p>If you do want your network to be able to handle any size <code>x<\/code>, you can have a <em>parameter-free<\/em> interpolation layer resizing all <code>x<\/code> to the right size before <code>self.fc1<\/code>:<\/p>\n\n<pre><code>x = F.max_pool2d(F.relu(self.conv2(x)), 2)  # output of conv layers\nx = F.interpolate(x, size=(5, 5), mode='bilinear')  # resize to the size expected by the linear unit\nx = x.view(x.size(0), 5 * 5 * 16)\nx = F.relu(self.fc1(x))  # you can go on from here...\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.functional.interpolate\" rel=\"nofollow noreferrer\"><code>torch.nn.functional.interpolate<\/code><\/a> for more information.<\/p>\n"}]}
+{"QuestionId":53500853,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T13:30:08.760","AcceptedAnswerId":null,"OwnerUserId":10711793.0,"Title":"Does it make sense to multithread two different neural networks on one GPU?","Body":"<p>Let's say I have two neural networks represented as Python classes A and B. Methods <code>A.run()<\/code> and <code>B.run()<\/code> represent a feedforward inference for one image.<\/p>\n\n<p>As an example, <code>A.run()<\/code> takes 100 ms, and <code>B.run()<\/code> takes 50 ms.<\/p>\n\n<p>When ran one after another, i.e.<\/p>\n\n<pre><code>img = cap.read()[1]  # e.g. cv2.VideoCapture instance\nstart_time = time.time()\nA.run(img)  # 100 ms\nB.run(img)  # 50 ms\ntime_diff = time.time() - start_time  # 100 + 50 = 150 ms\n<\/code><\/pre>\n\n<p>the inference times just add up to 150 ms.<\/p>\n\n<p>To make this faster, we can try parallelizing so that they start at the same time. An implementation that uses Python's threading is outlined below:<\/p>\n\n<pre><code>class A:\n    # This method is spawned using Python's threading library\n    def run_queue(self, input_queue, output_queue):\n        while True:\n            img = input_queue.get()\n            start_time = time.time()\n            output = self.run(img)  \n            time_diff = time.time() - start_time # Supposedly 100 ms for class A, and 50 ms for class B\n\n\n# in main program flow:            \n# Assume that a_input_queue and a_output_queue are tied to an instance of class A\n# And similar for class B\n\nimg = cap.read()[1]\na_input_queue.put(img)\nb_input_queue.put(img)\n\nstart_time = time.time()\na_output = a_output_queue.get()  # Should take 100 ms\nb_output = b_output_queue.get()  # B.run() should take 50 ms, but since it started at the same time as A.run(), this get() should effectively return immediately\ntime_diff = time.time() - start_time  # Should theoretically be 100 ms\n<\/code><\/pre>\n\n<p>So theoretically, we should only be bottlenecked by A, and end up with 100 ms for the whole system.<\/p>\n\n<p>However, it seems that B.run() takes around 100 ms as well when measured in <code>B.run_queue()<\/code>. Since they started at around the same time, the whole system also takes around a 100 ms.<\/p>\n\n<p>Does this make sense? Is trying to thread the two neural networks sensible, if the resulting total inference time is about the same (or possibly incrementally faster at least)?<\/p>\n\n<p>My guess is that the GPU is maxed at 100% for one neural network, so when trying to inference two networks at the same time, it just rearranges the instructions but can only do the same number of computations anyway:<\/p>\n\n<pre><code>Illustration:\nA.run() executes 8 blocks of instructions:\n| X | X | X | X | X | X | X | X |\nB.run() executes only 4 blocks of instructions:\n| Y | Y | Y | Y |\n\nNow, say that the GPU can process 2 blocks of instructions per second.\n\nSo, in the case that A.run() and B.run() are ran one after the other (non-threaded):\n| X | X | X | X | X | X | X | X | Y | Y | Y | Y | -&gt; A.run() takes 4 s, B.run() takes 2 s, everything takes 6 s\n\nIn the threaded case, the instructions are rearranged so both start at the same time, but get stretched out:\n| X | X | Y | X | X | Y | X | X | Y | X | X | Y  -&gt; A.run() roughly takes 6 s, B.run() roughly takes 6 s, everything seems to take 6 s\n<\/code><\/pre>\n\n<p>Is the above illustration the case?<\/p>\n\n<p>Finally, let's consider a class C similar to B (e.g. inference time=50 ms), except that it uses the CPU. Thus, it shouldn't compete with A in GPU usage, but from experiments, it just behaved like B; its inference time seemed to be stretched to match A's.<\/p>\n\n<p>Thoughts? Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":53501084,"AnswerCount":1,"Tags":"<python><tensorflow><dataset>","CreationDate":"2018-11-27T13:42:18.017","AcceptedAnswerId":null,"OwnerUserId":4642578.0,"Title":"Tensorflow estimator issue with datasets","Body":"<p>I'm having a strange issue with TF estimators and trying to use tf.Dataset in my input function.<\/p>\n\n<p>First off, my model looks like this:<\/p>\n\n<pre><code>    model = tf.estimator.DNNClassifier(\n        feature_columns=my_feature_column,\n        hidden_units=[hidden_layers, hidden_layers],\n        n_classes=n_classes)\n<\/code><\/pre>\n\n<p>and my feature colum is like this<\/p>\n\n<pre><code>    my_feature_column = [tf.feature_column.numeric_column(key='image', shape=[32, 32, 3])]\n<\/code><\/pre>\n\n<p>Now, if I train like this, it all works fine and the training runs in a couple of seconds:<\/p>\n\n<pre><code>    model.train(\n        input_fn=tf.estimator.inputs.numpy_input_fn(\n            dict({'image':X_train}),\n            y_train,\n            shuffle=True),\n        steps=nb_epoch)\n<\/code><\/pre>\n\n<p>But when I try to add in tf.Datasets in the input function, it takes forever to run:<\/p>\n\n<pre><code>def input_fn(features, labels, batch_size):\n    dataset = tf.data.Dataset.from_tensor_slices(({'image':features}, labels))\n    return dataset.shuffle(1000).batch(batch_size).repeat()\n\nmodel.train(\n    input_fn=lambda:input_fn(X_train, y_train, batch_size),\n    steps=nb_epoch)\n<\/code><\/pre>\n\n<p>Can anyone see what I'm doing wrong please? It should be identical, right?<\/p>\n\n<p>Thanks,\nPaul<\/p>\n","answers":[{"AnswerId":"53600468","CreationDate":"2018-12-03T19:20:32.747","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Your dataset is repeating infinitely and there's no default maximum number of iterations, so tensorflow does not know when to stop.<\/p>\n\n<p>Replace the line which says <code>return dataset.shuffle(1000).batch(batch_size).repeat()<\/code> with something like  <code>return dataset.shuffle(1000).batch(batch_size).repeat(10)<\/code>, which will train for 10 epochs, and you'll be fine.<\/p>\n"}]}
+{"QuestionId":53501315,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T13:55:33.363","AcceptedAnswerId":null,"OwnerUserId":5634636.0,"Title":"How to understand tensorflow error message?","Body":"<p>I found that the error message from TensorFlow, especially at <em>run time<\/em> (i.e. in <code>sess.run()<\/code>). There'is few document explaining how to understand the error message.<\/p>\n\n<p>For example, there is a error message:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1322, in _do_call\n    return fn(*args)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1307, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1409, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 10669 values, but the requested shape has 11172\n     [[Node: optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape\/tensor, optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Shape)]]\n     [[Node: cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\/_1897 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_4151_cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\", tensor_type=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/main.py\", line 155, in &lt;module&gt;\n    res = runManager.modRun(i)\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/Utils\/RunManager.py\", line 97, in modRun\n    return self.runAndMerge(feed_dict, options=options if options is not None else self.options, run_metadata=run_metadata if run_metadata is not None else self.run_metadata)\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/Utils\/RunManager.py\", line 71, in runAndMerge\n    res = self.sess.run(self.inputTensors, feed_dict=feed_dict, options=options, run_metadata=run_metadata)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 900, in run\n    run_metadata_ptr)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1135, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1316, in _do_run\n    run_metadata)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1335, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 10669 values, but the requested shape has 11172\n     [[Node: optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape\/tensor, optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Shape)]]\n     [[Node: cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\/_1897 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_4151_cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\", tensor_type=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\nCaused by op 'optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape', defined at:\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/main.py\", line 118, in &lt;module&gt;\n    trainOp = createUpdateOp()\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/main.py\", line 104, in createUpdateOp\n    grads = optimizer.compute_gradients(totalLoss, var_list=net.getVariables())\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 526, in compute_gradients\n    colocate_gradients_with_ops=colocate_gradients_with_ops)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 494, in gradients\n    gate_gradients, aggregation_method, stop_gradients)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 636, in _GradientsHelper\n    lambda: grad_fn(op, *out_grads))\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 385, in _MaybeCompile\n    return grad_fn()  # Exit early\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 636, in &lt;lambda&gt;\n    lambda: grad_fn(op, *out_grads))\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_grad.py\", line 521, in _ReshapeGrad\n    return [array_ops.reshape(grad, array_ops.shape(op.inputs[0])), None]\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 6113, in reshape\n    \"Reshape\", tensor=tensor, shape=shape, name=name)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3392, in create_op\n    op_def=op_def)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1718, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\n...which was originally created as op 'RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2', defined at:\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/main.py\", line 96, in &lt;module&gt;\n    tf.losses.add_loss(net.getLoss(boxes, classes))\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/BoxEngine\/BoxNetwork.py\", line 50, in getLoss\n    return self.rpn.loss(refBoxes) + self.boxRefiner.loss(self.proposals, refBoxes, refClasses)\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/BoxEngine\/RPN.py\", line 186, in loss\n    return tf.cond(tf.shape(refBoxes)[0] &gt; 0, lambda: calcLoss(), lambda: tf.constant(0.0))\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 432, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2063, in cond\n    orig_res_t, res_t = context_t.BuildCondBranch(true_fn)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 1913, in BuildCondBranch\n    original_result = fn()\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/BoxEngine\/RPN.py\", line 186, in &lt;lambda&gt;\n    return tf.cond(tf.shape(refBoxes)[0] &gt; 0, lambda: calcLoss(), lambda: tf.constant(0.0))\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/BoxEngine\/RPN.py\", line 173, in calcLoss\n    positiveLosses, negativeLosses = calcAllLosses(inAnchros, inBoxes, inRawSizes, inScores, inBoxSizes)\n  File \"\/home\/hyh\/projects\/RFCN-tensorflow\/BoxEngine\/RPN.py\", line 145, in calcAllLosses\n    classificationLoss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=scores, labels=refScores, name=\"classification_loss\")\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 1878, in softmax_cross_entropy_with_logits_v2\n    cost = array_ops.reshape(cost, output_shape)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 6113, in reshape\n    \"Reshape\", tensor=tensor, shape=shape, name=name)\n  File \"\/home\/hyh\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n\nInvalidArgumentError (see above for traceback): Input to reshape is a tensor with 10669 values, but the requested shape has 11172\n     [[Node: optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Reshape\/tensor, optimizer\/gradients\/RPNloss\/cond\/calcRPNLoss\/calcAllRPNLosses\/classification_loss\/Reshape_2_grad\/Shape)]]\n     [[Node: cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\/_1897 = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\", send_device_incarnation=1, tensor_name=\"edge_4151_cond\/getRefinementLoss\/posLoss\/getPosLoss\/Reshape\", tensor_type=DT_INT32, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>I have two questions:<\/p>\n\n<ol>\n<li><p>Where there is so many calling stack? First is <code>Trackback<\/code> and then <code>During handling of the above exception, another exception occurred:<\/code>, and <code>Caused by...<\/code>, finally <code>...which was originally created as op<\/code>. What do they mean respectively?<\/p><\/li>\n<li><p>Why there is so many error node? In the message above, it seems that there are two nodes that have gone wrong. What does it mean? Which node caused this error?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"53502703","CreationDate":"2018-11-27T15:16:00.107","ParentId":null,"OwnerUserId":"9710391","Title":null,"Body":"<p>Tensorflow error messages are always quite verbose and this is mainly due to how TF works (because of the Computation Graph it builds).\nIn your case, it seems that you are reshaping a tensor with the wrong shape:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 10669 values, but the requested shape has 11172\n<\/code><\/pre>\n\n<p>To see if that is the case try printing the shape of the tensor given to reshape op, i.e.:<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32, [None, 28, 28, 1])\nx = tf.layers.dense(input, units=64, activation=tf.nn.relu)\nx = tf.Print(x, [x])\nx_rs = tf.reshape(x, [-1, 28*28])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53501575,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-27T14:09:44.203","AcceptedAnswerId":53929725.0,"OwnerUserId":5030834.0,"Title":"Tensorflow\/Keras multi-label classifier","Body":"<p>I have just started developing some simple classifier in Tenosrflow and I've started using this example on Tensorflow site: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a><\/p>\n\n<p>Now I want my model to get images like this as features:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/E2pyt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/E2pyt.png\" alt=\"Corresponding label: [1, 0]\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OfhTV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OfhTV.png\" alt=\"Corresponding label: [3, 0]\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vczRe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vczRe.png\" alt=\"Corresponding label: [1, 3]\"><\/a><\/p>\n\n<p>These images should have, as corresponding labels, three arrays: [1,0], [3,0] and [1,3].\nMy problem is: how can I load into the model these kind of labels (i.e. labels that are arrays and not a single scalar)?\nWhen I try as in the example down here, the only thing I got is an error message that I won't report here because they are generated from my lack of knowledge on the thing that I'm trying to do.<\/p>\n\n<p>Additional question: how should the last neural layer be? How many neurons should it have?<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport skimage\nfrom skimage.color import rgb2gray\nimport csv\nimport numpy as np\n\nnames = ['Cerchio', 'Quadrato', 'Stella']\n\nimages = []\nlabels = [[]]\n\ntest_images = []\ntest_labels = [[]]\nfinal_images = []\n\nfor i in range(1, 501):\n    images.append(skimage.data.imread(\"{0}.bmp\".format(i)))\nfor i in range(501, 601):\n    test_images.append(skimage.data.imread(\"{0}.bmp\".format(i)))\n\nfor i in range(601, 701):\n    final_images.append(skimage.data.imread(\"{0}.bmp\".format(i)))\n\nfile = open(\"labels.csv\", \"rU\")\n\nreader = csv.reader(file, delimiter=\",\")\n\n\nfor row in reader:\n    for i in range(0, 499):\n        if int(row[i]) &lt; 10:\n            labels.append([int(int(row[i])\/10), 0])\n        else:\n            labels.append([int(int(row[i])\/10), int(row[i])%10])\n    for i in range(500, 600):\n        if int(row[i]) &lt; 10:\n            test_labels.append([int(int(row[i])\/10), 0])\n        else:\n            test_labels.append([int(int(row[i])\/10), int(row[i])%10])\n\n\nfile.close()\n\nimages28 = np.array(images)\nimages28 = rgb2gray(images28)\ntest_images28 = np.array(test_images)\ntest_images28 = rgb2gray(test_images28)\nfinal_images28 = np.array(final_images)\nfinal_images28 = rgb2gray(final_images28)\n\nlabels = np.array(labels)\ntest_labels = np.array(test_labels)\nprint(labels)\n\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(28, 56)),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(4, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(), \n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(images28, labels, epochs=5)\n\ntest_loss, test_acc = model.evaluate(test_images28, test_labels)\n\nprint('Test accuracy:', test_acc)\na = input()\nimg = final_images28[int(a)]\n\nprint(img.shape)\n\nimg = (np.expand_dims(img, 0))\nprint(img.shape)\n\npredictions_single = model.predict(img)\nprint(predictions_single)\nprint(names[np.argmax(predictions_single)])\n<\/code><\/pre>\n","answers":[{"AnswerId":"53929725","CreationDate":"2018-12-26T08:59:34.360","ParentId":null,"OwnerUserId":"5030834","Title":null,"Body":"<p>After Googling around and toying with my program, I found the solution: a multi-hot encoded array.\nIn this array, if I have a position for a circle, a square, a star and the blank space (hence a 4 position array), I can feed to my model labels that have a '1' in each corresponding space.\nE.g. (referring to the example above):<\/p>\n\n<ul>\n<li>[1, 0, 1, 0]<\/li>\n<li>[1, 0, 0, 1]<\/li>\n<li>[0, 0, 1, 1]<\/li>\n<\/ul>\n\n<p>This did work perfectly.<\/p>\n"},{"AnswerId":"53502270","CreationDate":"2018-11-27T14:50:16.090","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>One way is just map the array labels into an index, like [[0,0],[0,0],[0,0]]->0, [[1,0],[0,0],[0,0]]->1,... etc. You'll have 3^6=729 possible labels. If these forms on the images are standard you probably can use just simplest classificator with no hidden layers so it's gonna be dim1xdim2x729 trainable weights. If they are not standard you will be better off using convolutional layers. <\/p>\n\n<p>You can probably also use fully convolutional model for this problem that is returning 3 dimensional tensor as an output. In this case you can use multidimensional labels. But then you'll have to write custom loss function for it.<\/p>\n"}]}
+{"QuestionId":53501946,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2018-11-27T14:30:02.063","AcceptedAnswerId":null,"OwnerUserId":5946607.0,"Title":"Custom loss function which depends on another neural network in keras","Body":"<p>I have a \"How can I do that\" question with keras :<\/p>\n\n<p>Assuming that I have a first neural network, say NNa which has 4 inputs (x,y,z,t) which <strong>is already trained<\/strong>.\nIf I have a second neural network, say NNb, and that its loss function depends on the first neural network.<\/p>\n\n<p>The custom loss function of NNb <code>customLossNNb<\/code> calls the prediction of NNa with a fixed grid (x,y,z) and just modify the last variable t.<\/p>\n\n<p>Here in pseudo-python-code what I would like to do to traine the second NN : NNb:<\/p>\n\n<pre><code>grid=np.mgrid[0:10:1,0:10:1,0:10:1].reshape(3,-1).T\n\nY[:,0]=time\nY[:,1]=something\n\ndef customLossNNb(NNa,grid):\n     def diff(y_true,y_pred): \n         for ii in range(y_true.shape[0]):\n               currentInput=concatenation of grid and y_true[ii,0]\n               toto[ii,:]=NNa.predict(currentInput)\n               #some stuff with toto\n         return #...\n     return diff\n<\/code><\/pre>\n\n<p>Then <\/p>\n\n<pre><code>NNb.compile(loss=customLossNNb(NNa,K.variable(grid)),optimizer='Adam')\nNNb.fit(input,Y)\n<\/code><\/pre>\n\n<p>In fact the line that cause me troubles is <code>currentInput=concatenation of grid and y_true[ii,0]<\/code><\/p>\n\n<p>I tried to send to customLossNNb the grid as a tensor with <code>K.variable(grid)<\/code>. But I can't defined a new tensor inside the loss function, something like <code>CurrentY<\/code> which has a shape <code>(grid.shape[0],1)<\/code> fill with <code>y[ii,0]<\/code>(<em>i.e.<\/em> the current t) and then concatenate <code>grid<\/code> and <code>currentY<\/code> to build <code>currentInput<\/code><\/p>\n\n<p>Any ideas?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"53503230","CreationDate":"2018-11-27T15:44:35.623","ParentId":null,"OwnerUserId":"7907205","Title":null,"Body":"<p>You can include your custom loss function into the graph using functional API of keras. The model in this case can be used as a function, something like this:<\/p>\n\n<pre><code>for l in NNa.layers: \n    l.trainable=False\nx=Input(size)\ny=NNb(x)\nz=NNa(y)\n<\/code><\/pre>\n\n<p>Predict method will not work, since loss function should be part of the graph, and predict method returns np.array<\/p>\n"},{"AnswerId":"53503653","CreationDate":"2018-11-27T16:05:52.940","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>First, make <code>NNa<\/code> untrainable. Notice that you should do this recursively if your model has inner models.<\/p>\n\n<pre><code>def makeUntrainable(layer):\n    layer.trainable = False\n\n    if hasattr(layer, 'layers'):\n        for l in layer.layers:\n            makeUntrainable(l)\n\nmakeUntrainable(NNa)\n<\/code><\/pre>\n\n<p>Then you have two options:<\/p>\n\n<ul>\n<li>Attach NNa to the end of your model (notice that both <code>y_true<\/code> and <code>y_pred<\/code> will be changed)   \n\n<ul>\n<li>Then change your targets (predict with NNa) for correct results since your model is now expecting the output of NNa, not NNb.  <\/li>\n<\/ul><\/li>\n<li>Create a custom loss function that uses <code>NNa<\/code> inside it, without changing your targets<\/li>\n<\/ul>\n\n<h2>Option 1 - Attaching models<\/h2>\n\n<pre><code>inputs = NNb.inputs   \noutputs = NNa(NNb.outputs) #make sure NNb is outputing 4 tensors to match NNa inputs   \nfullModel = Model(inputs,outputs)\n\n#changing the targets:\nnewY_train = NNa.predict(oldY_train)    \n<\/code><\/pre>\n\n<h2>Option 2 - Creating a custom loss<\/h2>\n\n<blockquote>\n  <p>Warning: please test whether NNa's weights are really frozen while training this configuration<\/p>\n<\/blockquote>\n\n<pre><code>from keras.losses import binary_crossentropy\n\ndef customLoss(true,pred):\n    true = NNa(true)\n    pred = NNa(pred)\n\n    #use some of the usual losses or create your own\n    binary_crossentropy(true,pred)\n\nNNb.compile(optimizer=anything, loss = customLoss)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53502785,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><tensorflow-lite>","CreationDate":"2018-11-27T15:20:29.097","AcceptedAnswerId":null,"OwnerUserId":2248896.0,"Title":"Converted my own Keras CNN model to TensorFlow Lite. Detection results in TensorFlow Lite Demo App wrong","Body":"<p>I am working with a Keras CNN for image classification I built myself. Detection results on my PC machine are fine for now (about 290 out of 300 test images detected correctly). Now I need to get my model running on an Android phone. I installed the <a href=\"https:\/\/www.tensorflow.org\/lite\/demo_android\" rel=\"nofollow noreferrer\">TFLite Demo App<\/a> on my phone. So far it works well with the preinstalled model as well as with some random pretrained models I found online. Sadly, if I convert the model do TFLite it will show wrong detection results.<\/p>\n\n<p>Now to convert my Keras (.h5-file) to TFLite (.tflite-file) I followed the description on the official <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/python_api\" rel=\"nofollow noreferrer\">Tensorflow<\/a> website, using this code:<\/p>\n\n<pre><code># Convert to TensorFlow Lite model.\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(keras_file)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>As I said above, the just created keras-model (<code>keras_file<\/code>) I used to convert to TFLite works perfectly fine on testing, but the TFLite-model does not. To be exact. For testing purposes there are currently 4 possible categories to detect. In 3 out of 4 cases the app will show me a wrong detection result, saying it is 100% certain. I thought there could be a problem with the label file I created. But there is no scheme to find (like maybe the order of the strings in the file are wrong or something). So I am almost certain that the label-file is OK.<\/p>\n\n<p>Now what I did on the Android app:<\/p>\n\n<ul>\n<li>(As mentioned in the <a href=\"https:\/\/www.tensorflow.org\/lite\/demo_android\" rel=\"nofollow noreferrer\">description<\/a>) change the classifier to <code>ImageClassifierFloatInception<\/code>.<\/li>\n<li>Create the label-file containing 4 strings for my 4 categories<\/li>\n<li>Change the image size<\/li>\n<li>Change the return value of <code>getNumBytesPerChannel()<\/code> to 4, to match my category-count<\/li>\n<\/ul>\n\n<p>Specs:<\/p>\n\n<ul>\n<li>Windows 10<\/li>\n<li>Tf-nightly-gpu 1.13.0a20181125 (TensorFlow-gpu 1.12.0 won't support the convert method)<\/li>\n<li>Galaxy S6 on Android 7.0<\/li>\n<\/ul>\n\n<p>Am I missing something? If you need more information I am happy to provide.<\/p>\n\n<p>Stephan<\/p>\n","answers":[{"AnswerId":"57189065","CreationDate":"2019-07-24T18:21:12.833","ParentId":null,"OwnerUserId":"7337609","Title":null,"Body":"<p>There may be a problem with your <code>convertBitmaptoByteBuffer()<\/code> function, or maybe with your <code>.tflite<\/code> model.<\/p>\n"}]}
+{"QuestionId":53502836,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><sparse-matrix><categorical-data><caffe2>","CreationDate":"2018-11-27T15:23:14.293","AcceptedAnswerId":null,"OwnerUserId":10712346.0,"Title":"How to use sparselengthssum in caffe2 for categorical data ?","Body":"<p>I am trying to convert categorical data to numerical data. caffe2 has few new operators for the same. I am interested in using sparselenthssum. But the caffe2 docs does not have an example use case. How to use sparelengthssum for categorical data?<\/p>\n\n<p><a href=\"https:\/\/caffe2.ai\/docs\/operators-catalogue.html#sparselengthssum\" rel=\"nofollow noreferrer\">https:\/\/caffe2.ai\/docs\/operators-catalogue.html#sparselengthssum<\/a>. <\/p>\n","answers":[]}
+{"QuestionId":53503094,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2018-11-27T15:37:46.783","AcceptedAnswerId":53504122.0,"OwnerUserId":8926217.0,"Title":"Increasing Accuracy of Image Classification Model","Body":"<p>I'm trying to use do image classification on two different classes using the pre-trained Inception V3 model. I have a data set of around 1400 images which are roughly balanced. When I run my program I get results that are off at the first couple epochs. Is this normal when training the model?<\/p>\n\n<pre><code>epochs =  175\n\nbatch_size = 64\n\n#include_top = false to accomodate new classes \nbase_model = keras.applications.InceptionV3(\n        weights ='imagenet',\n        include_top=False, \n        input_shape = (img_width,img_height,3))\n\n#Classifier Model ontop of Convolutional Model\nmodel_top = keras.models.Sequential()\nmodel_top.add(keras.layers.GlobalAveragePooling2D(input_shape=base_model.output_shape[1:], data_format=None)),\nmodel_top.add(keras.layers.Dense(350,activation='relu'))\nmodel_top.add(keras.layers.Dropout(0.4))\nmodel_top.add(keras.layers.Dense(1,activation = 'sigmoid'))\nmodel = keras.models.Model(inputs = base_model.input, outputs = model_top(base_model.output))\n\n#freeze the convolutional layers of InceptionV3\nfor layer in model.layers[:30]:\nlayer.trainable = False\n\n#Compiling model using Adam Optimizer \nmodel.compile(optimizer = keras.optimizers.Adam(\n                    lr=0.000001,\n                    beta_1=0.9,\n                    beta_2=0.999,\n                    epsilon=1e-08),\n                    loss='binary_crossentropy',\n                    metrics=['accuracy'])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/d23Fn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d23Fn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>With my current parameters I only get an accuracy of 89% with a test loss of 0.3 when testing on a separated set of images. Do I need to add more layers to my model to increase this accuracy?<\/p>\n","answers":[{"AnswerId":"53504122","CreationDate":"2018-11-27T16:31:39.550","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>There are several issues with your code... <\/p>\n\n<p>To start with, your way to build <code>model_top<\/code> is quite unconventional (and IMHO quite messy as well); in such cases, the <a href=\"https:\/\/keras.io\/applications\/#usage-examples-for-image-classification-models\" rel=\"nofollow noreferrer\">documentation examples<\/a> are your best friend. So, start with replacing your <code>model_top<\/code> part with:<\/p>\n\n<pre><code>x = base_model.output\nx = GlobalAveragePooling2D()(x)\nx = Dense(350, activation='relu')(x)\nx = Dropout(0.4)(x)\npredictions = Dense(1, activation='sigmoid')(x)\n\n# this is the model we will train\nmodel = Model(inputs=base_model.input, outputs=predictions)\n<\/code><\/pre>\n\n<p>Notice that I have not changed your parameters of choice - you could certainly experiment with more units in the dense layer (the example in the docs uses 1024)...<\/p>\n\n<p>Second, it is not clear why you choose to freeze only 30 layers of the InceptionV3, which has no less than 311 layers:<\/p>\n\n<pre><code>len(base_model.layers)\n# 311\n<\/code><\/pre>\n\n<p>So, replace also this part with<\/p>\n\n<pre><code>for layer in base_model.layers:\n    layer.trainable = False\n<\/code><\/pre>\n\n<p>Third, your learning rate seems way too small; the Adam optimizer is supposed to work well enough out of the box with its default parameters, so I also suggest to compile your model simply as <\/p>\n\n<pre><code>model.compile(optimizer = keras.optimizers.Adam(),\n                    loss='binary_crossentropy',\n                    metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53503234,"AnswerCount":1,"Tags":"<python><machine-learning><statistics><pytorch>","CreationDate":"2018-11-27T15:44:44.057","AcceptedAnswerId":53519781.0,"OwnerUserId":4615566.0,"Title":"PyTorch - parameters not changing","Body":"<p>In an effort to learn how pytorch works, I am trying to do maximum likelihood estimation of some of the parameters in a multivariate normal distribution. However it does not seem to work for any of the covariance related parameters.<\/p>\n\n<p>So my question is: why does this code not work?<\/p>\n\n<pre><code>import torch\n\n\ndef make_covariance_matrix(sigma, rho):\n    return torch.tensor([[sigma[0]**2, rho * torch.prod(sigma)],\n                         [rho * torch.prod(sigma), sigma[1]**2]])\n\n\nmu_true = torch.randn(2)\nrho_true = torch.rand(1)\nsigma_true = torch.exp(torch.rand(2))\n\ncov_true = make_covariance_matrix(sigma_true, rho_true)\ndist_true = torch.distributions.MultivariateNormal(mu_true, cov_true)\n\nsamples = dist_true.sample((1_000,))\n\nmu = torch.zeros(2, requires_grad=True)\nlog_sigma = torch.zeros(2, requires_grad=True)\natanh_rho = torch.zeros(1, requires_grad=True)\n\nlbfgs = torch.optim.LBFGS([mu, log_sigma, atanh_rho])\n\n\ndef closure():\n    lbfgs.zero_grad()\n    sigma = torch.exp(log_sigma)\n    rho = torch.tanh(atanh_rho)\n    cov = make_covariance_matrix(sigma, rho)\n    dist = torch.distributions.MultivariateNormal(mu, cov)\n    loss = -torch.mean(dist.log_prob(samples))\n    loss.backward()\n    return loss\n\n\nlbfgs.step(closure)\n\nprint(\"mu: {}, mu_hat: {}\".format(mu_true, mu))\nprint(\"sigma: {}, sigma_hat: {}\".format(sigma_true, torch.exp(log_sigma)))\nprint(\"rho: {}, rho_hat: {}\".format(rho_true, torch.tanh(atanh_rho)))\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>mu: tensor([0.4168, 0.1580]), mu_hat: tensor([0.4127, 0.1454], requires_grad=True)\nsigma: tensor([1.1917, 1.7290]), sigma_hat: tensor([1., 1.], grad_fn=&lt;ExpBackward&gt;)\nrho: tensor([0.3589]), rho_hat: tensor([0.], grad_fn=&lt;TanhBackward&gt;)\n\n&gt;&gt;&gt; torch.__version__\n'1.0.0.dev20181127'\n<\/code><\/pre>\n\n<p>In other words, why have the estimates of <code>log_sigma<\/code> and <code>atanh_rho<\/code> not moved from their initial value?<\/p>\n","answers":[{"AnswerId":"53519781","CreationDate":"2018-11-28T12:43:56.903","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The way you create your covariance matrix is not <em>backprob-able<\/em>:<\/p>\n\n<pre><code>def make_covariance_matrix(sigma, rho):\n    return torch.tensor([[sigma[0]**2, rho * torch.prod(sigma)],\n                         [rho * torch.prod(sigma), sigma[1]**2]])\n<\/code><\/pre>\n\n<p>When creating a new tensor from (multiple) tensors, only the values of your input tensors will be kept. All additional information from the input tensors is stripped away, thus all <em>graph-connection<\/em> to your parameters is cut from this point, therefore backpropagation cannot get through.<\/p>\n\n<p><strong>Here is a short example to illustrate this:<\/strong><\/p>\n\n<pre><code>import torch\n\nparam1 = torch.rand(1, requires_grad=True)\nparam2 = torch.rand(1, requires_grad=True)\ntensor_from_params = torch.tensor([param1, param2])\n\nprint('Original parameter 1:')\nprint(param1, param1.requires_grad)\nprint('Original parameter 2:')\nprint(param2, param2.requires_grad)\nprint('New tensor form params:')\nprint(tensor_from_params, tensor_from_params.requires_grad)\n<\/code><\/pre>\n\n<p><em>Output:<\/em><\/p>\n\n<pre><code>Original parameter 1:\ntensor([ 0.8913]) True\nOriginal parameter 2:\ntensor([ 0.4785]) True\nNew tensor form params:\ntensor([ 0.8913,  0.4785]) False\n<\/code><\/pre>\n\n<p>As you can see the tensor, created from the parameters <code>param1<\/code> and <code>param2<\/code>, does not keep track of the gradients of <code>param1<\/code> and <code>param2<\/code>.<\/p>\n\n<p><br><\/p>\n\n<p><strong>So instead you can use this code that keeps the <em>graph connection<\/em> and is <em>backprob-able<\/em>:<\/strong><\/p>\n\n<pre><code>def make_covariance_matrix(sigma, rho):\n    conv = torch.cat([(sigma[0]**2).view(-1), rho * torch.prod(sigma), rho * torch.prod(sigma), (sigma[1]**2).view(-1)])\n    return conv.view(2, 2)\n<\/code><\/pre>\n\n<p>The values are concatenated to a flat tensor using <code>torch.cat<\/code>. Then they are brought into right shape using <code>view()<\/code>. <br>This results in the same matrix output as in your function, but it keeps the connection to your parameters <code>log_sigma<\/code> and <code>atanh_rho<\/code>.<\/p>\n\n<p>Here is an output before and after the step with the changed <code>make_covariance_matrix<\/code>. As you can see, now you can optimize your parameters and the values do change:<\/p>\n\n<pre><code>Before:\nmu: tensor([ 0.1191,  0.7215]), mu_hat: tensor([ 0.,  0.])\nsigma: tensor([ 1.4222,  1.0949]), sigma_hat: tensor([ 1.,  1.])\nrho: tensor([ 0.2558]), rho_hat: tensor([ 0.])\n\nAfter:\nmu: tensor([ 0.1191,  0.7215]), mu_hat: tensor([ 0.0712,  0.7781])\nsigma: tensor([ 1.4222,  1.0949]), sigma_hat: tensor([ 1.4410,  1.0807])\nrho: tensor([ 0.2558]), rho_hat: tensor([ 0.2235])\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":53503389,"AnswerCount":4,"Tags":"<python><keras><deep-learning>","CreationDate":"2018-11-27T15:52:34.177","AcceptedAnswerId":53512269.0,"OwnerUserId":10712335.0,"Title":"How to set parameters in keras to be non-trainable?","Body":"<p>I am new to Keras and I am building a model. I want to freeze the weights of the last few layers of the model while training the previous layers. I tried to set the trainable property of the lateral model to be False, but it dosen't seem to work. Here is the code and the model summary:<\/p>\n\n<pre><code>opt = optimizers.Adam(1e-3)\ndomain_layers = self._build_domain_regressor()\ndomain_layers.trainble = False\nfeature_extrator = self._build_common()\nimg_inputs = Input(shape=(160, 160, 3))\nconv_out = feature_extrator(img_inputs)\ndomain_label = domain_layers(conv_out)\nself.domain_regressor = Model(img_inputs, domain_label)\nself.domain_regressor.compile(optimizer = opt, loss='binary_crossentropy', metrics=['accuracy'])\nself.domain_regressor.summary()\n<\/code><\/pre>\n\n<p>The model summary: <a href=\"https:\/\/i.stack.imgur.com\/nWR8q.png\" rel=\"nofollow noreferrer\">model summary<\/a>.<\/p>\n\n<p>As you can see, <code>model_1<\/code> is trainable. But according to the code, it is set to be non-trainable.<\/p>\n","answers":[{"AnswerId":"53512269","CreationDate":"2018-11-28T04:42:21.067","ParentId":null,"OwnerUserId":"10712335","Title":null,"Body":"<p>There is a typo in the Word \"trainble\"(missing an \"a\"). Saddly keras doesn't warn me that the model doesn't have the property \"trainble\". The question could be closed.<\/p>\n"},{"AnswerId":"53508529","CreationDate":"2018-11-27T21:35:06.740","ParentId":null,"OwnerUserId":"4836149","Title":null,"Body":"<p>Change the last 3 lines in your code:<\/p>\n\n<pre><code>last_few_layers = 20 #number of the last few layers to freeze\nself.domain_regressor = Model(img_inputs, domain_label)\nfor layer in model.layers[:-last_few_layers]:\n    layer.trainable = False\nself.domain_regressor.compile(optimizer = opt, loss='binary_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n"},{"AnswerId":"53503615","CreationDate":"2018-11-27T16:03:48.927","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>You can simple assign a boolean value to the layer property <code>trainable<\/code>.<\/p>\n\n<pre><code>model.layers[n].trainable = False\n<\/code><\/pre>\n\n<p>You can visualize which layer is trainable:<\/p>\n\n<pre><code>for l in model.layers:\n    print(l.name, l.trainable)\n<\/code><\/pre>\n\n<p>You can pass it by the model definition too:<\/p>\n\n<pre><code>frozen_layer = Dense(32, trainable=False)\n<\/code><\/pre>\n\n<p>From Keras <a href=\"https:\/\/keras.io\/getting-started\/faq\/\" rel=\"noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>To \"freeze\" a layer means to exclude it from training, i.e. its\n  weights will never be updated. This is useful in the context of\n  fine-tuning a model, or using fixed embeddings for a text input.<br>\n  You can pass a trainable argument (boolean) to a layer constructor to\n  set a layer to be non-trainable.\n  Additionally, you can set the trainable property of a layer to True or\n  False after instantiation. For this to take effect, you will need to\n  call compile() on your model after modifying the trainable property.<\/p>\n<\/blockquote>\n"},{"AnswerId":"58771327","CreationDate":"2019-11-08T17:33:28.307","ParentId":null,"OwnerUserId":"1133157","Title":null,"Body":"<p>Despite the fact that the original question's solution is a typo fix, let me add some information on keras trainables.<\/p>\n\n<p>Modern Keras contains the following facilities to view and manipulate trainable state:<\/p>\n\n<ul>\n<li><code>tf.keras.Layer._get_trainable_state()<\/code> function - prints the dictinary where keys are  model components and values are booleans. Note that <code>tf.keras.Model<\/code> is also a <code>tf.Keras.Layer<\/code>.<\/li>\n<li><code>tf.keras.Layer.trainable<\/code> property - to manipulate trainable state of individual layers.<\/li>\n<\/ul>\n\n<p>So the typical actions look like following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Print current trainable map:\nprint(model._get_trainable_state())\n\n# Set every layer to be non-trainable:\nfor k,v in model._get_trainable_state().items():\n    k.trainable = False\n\n# Don't forget to re-compile the model\nmodel.compile(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53503856,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><gated-recurrent-unit>","CreationDate":"2018-11-27T16:16:37.197","AcceptedAnswerId":null,"OwnerUserId":9552077.0,"Title":"How to implement 1-sigmoid in Keras?","Body":"<p>As I want to implement a structure which is similar to the update gate of GRU:<\/p>\n\n<p>h<sub>t<\/sub> = (1-z<sub>t<\/sub>)h<sub>t-1<\/sub> + z<sub>t<\/sub>h<sub>t<\/sub><\/p>\n\n<p>And I am trying to implement it with these code but it doesn't work. I am sure the problem are in the following code:<\/p>\n\n<pre><code>one = K.ones(shape=(1, len, 128))\nzt=Subtract([one,zt])\ntemp_conv2=multiply([reset_conv,zt])\noutput=Add([temp_conv1,temp_conv2])\n<\/code><\/pre>\n\n<p>I have the following error:<\/p>\n\n<blockquote>\n  <p>AttributeError:'Tensor' object has no attribute '_keras_history'<\/p>\n<\/blockquote>\n\n<p>I have already tried some other method such as using <code>Lambda<\/code> layer but it doesn't work.<\/p>\n","answers":[{"AnswerId":"53504024","CreationDate":"2018-11-27T16:26:07.433","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>one<\/code> is not a Keras Tensor therefore you would get that error. You can wrap this in a <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>zt = Lambda(lambda x: Subtract([K.ones(shape=(1, len, 128)), x]))(zt)\n<\/code><\/pre>\n\n<p>Even you don't need to construct that Tensor of ones. Simply use <code>1-x<\/code>:<\/p>\n\n<pre><code>zt = Lambda(lambda x: 1-x)(zt)\n<\/code><\/pre>\n\n<p>It will be automatically broadcasted and the subtraction would be element-wise.<\/p>\n"}]}
+{"QuestionId":53503889,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T16:18:22.927","AcceptedAnswerId":54452226.0,"OwnerUserId":6583936.0,"Title":"Getting error with tensorflow tf.control_dependencies and tf.layers.batch_normalization","Body":"<p>I'm trying to build a neural net with two loss functions, which are combined like weighed sum. First one simply computes <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/losses\/mean_squared_error\" rel=\"nofollow noreferrer\">mean square error<\/a> of linear output of dense layer and given labels, but another makes heavy use of nested <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\">tf.map_fn<\/a>. There are batch norm layers used with<code>tf.layers.batch_normalization()<\/code> so i had to add these lines to optimization objective: <\/p>\n\n<pre><code>    with tf.name_scope(\"Optimizer\"):\n        with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):\n            adam = tf.train.AdamOptimizer()\n            self.train_op = adam.minimize(self.total_loss)\n<\/code><\/pre>\n\n<p>However i get error:<\/p>\n\n<pre><code>   AttributeError: 'NoneType' object has no attribute 'op'\n<\/code><\/pre>\n\n<p>It comes from <code>minimize()<\/code> method. If i remove control dependencies, there is no error. Also if i remove second optimization objective which relies on loops, there is no error. I've tested second loss function in forward propagation and it worked just fine.  <\/p>\n\n<p>Any ideas how to track the problem? Full error log: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 3267, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-2-6d5efdb6d091&gt;\", line 1, in &lt;module&gt;\n    runfile('\/home\/mtarasov\/PycharmProjects\/ML\/src\/utils\/model.py', wdir='\/home\/mtarasov\/PycharmProjects\/ML\/src\/utils')\n  File \"\/home\/mtarasov\/Installations\/pycharm-2018.2.4\/helpers\/pydev\/_pydev_bundle\/pydev_umd.py\", line 198, in runfile\n    pydev_imports.execfile(filename, global_vars, local_vars)  # execute the script\n  File \"\/home\/mtarasov\/Installations\/pycharm-2018.2.4\/helpers\/pydev\/_pydev_imps\/_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"\/home\/mtarasov\/PycharmProjects\/ML\/src\/utils\/model.py\", line 168, in &lt;module&gt;\n    model = Model().build()\n  File \"\/home\/mtarasov\/PycharmProjects\/ML\/src\/utils\/model.py\", line 60, in build\n    self.train_op = adam.minimize(self.total_loss)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 400, in minimize\n    grad_loss=grad_loss)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 514, in compute_gradients\n    colocate_gradients_with_ops=colocate_gradients_with_ops)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 596, in gradients\n    gate_gradients, aggregation_method, stop_gradients)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 663, in _GradientsHelper\n    to_ops, from_ops, colocate_gradients_with_ops, func_graphs, xs)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gradients_impl.py\", line 190, in _PendingCount\n    between_op_list, between_ops, colocate_gradients_with_ops)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 1432, in MaybeCreateControlFlowState\n    loop_state.AddWhileContext(op, between_op_list, between_ops)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 1244, in AddWhileContext\n    grad_state = GradLoopState(forward_ctxt, outer_grad_state)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 846, in __init__\n    real_cnt, outer_grad_state)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2585, in AddBackpropLoopCounter\n    name=\"b_count\")\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 248, in _Enter\n    data, frame_name, is_constant, parallel_iterations, name=name)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_control_flow_ops.py\", line 178, in enter\n    parallel_iterations=parallel_iterations, name=name)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 454, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3155, in create_op\n    op_def=op_def)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1746, in __init__\n    self._control_flow_post_processing()\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1757, in _control_flow_post_processing\n    self._control_flow_context.AddOp(self)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2438, in AddOp\n    self._AddOpInternal(op)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2475, in _AddOpInternal\n    for x in external_inputs if x.outputs]\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2475, in &lt;listcomp&gt;\n    for x in external_inputs if x.outputs]\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 80, in identity\n    return gen_array_ops.identity(input, name=name)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 3264, in identity\n    \"Identity\", input=input, name=name)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 454, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3155, in create_op\n    op_def=op_def)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1746, in __init__\n    self._control_flow_post_processing()\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1757, in _control_flow_post_processing\n    self._control_flow_context.AddOp(self)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2438, in AddOp\n    self._AddOpInternal(op)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2466, in _AddOpInternal\n    self._MaybeAddControlDependency(op)\n  File \"\/home\/mtarasov\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2504, in _MaybeAddControlDependency\n    op._add_control_input(self.GetControlPivot().op)\nAttributeError: 'NoneType' object has no attribute 'op\n<\/code><\/pre>\n","answers":[{"AnswerId":"53515513","CreationDate":"2018-11-28T08:54:20.383","ParentId":null,"OwnerUserId":"6583936","Title":null,"Body":"<p>I dont know what exactly happened, but i have replaced all layers including batch norm from <code>tf.layers<\/code> to <code>tf.keras.layers<\/code> and the error has gone!<\/p>\n"},{"AnswerId":"54452226","CreationDate":"2019-01-31T01:53:04.537","ParentId":null,"OwnerUserId":"6217859","Title":null,"Body":"<p>Adding on to mcstarioni's answer. As pointed out, replacing the batch norm layers with <code>tf.keras.layers.BatchNormalization<\/code> appears to remove the error. However, this is because  BatchNormalization in keras does not add the batch norm parameters in <code>UPDATE_OPS<\/code> as is noted <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19643\" rel=\"nofollow noreferrer\">here<\/a>, as it uses a different way of training. If you check the moving mean and variance, you'll notice they are not updated during training just by running <code>train_op<\/code>. <strong>It is important to run <code>layer.update<\/code> in addition to <code>train_op<\/code><\/strong>, which should solve the problem.<\/p>\n\n<p>Alternatively, if possible, try to remove the nested <code>map_fn<\/code>.<\/p>\n"}]}
+{"QuestionId":53503926,"AnswerCount":0,"Tags":"<tensorflow><lstm><information-extraction><seq2seq>","CreationDate":"2018-11-27T16:20:28.113","AcceptedAnswerId":null,"OwnerUserId":9437254.0,"Title":"Training with tensorflow seq2seq model","Body":"<p>I am currently working with lstm. I have a dataset of a number of sentences about transactional info and I want to extract information, suppose amount, date and transactionWith. I already tried with basic lstm where my system tried to predict each word of a given sequence as amount, date, transactionWith or irrelevant. <\/p>\n\n<p>I have made my training data like this:<\/p>\n\n<p>Input:<\/p>\n\n<blockquote>\n  <p>You gave 100.00 to John on 13-08-2018<\/p>\n<\/blockquote>\n\n<p>Target:(labelled every word)<\/p>\n\n<blockquote>\n  <p>ir ir amount ir transactionWith ir date<\/p>\n<\/blockquote>\n\n<p>You can see that the entire dataset has a lot of \"ir\" or irrelevant tag and I think that will make my system biased to predict \"ir\" for test data.<\/p>\n\n<p>Now I want to try using seq2seq model of tensorflow where the input is a transactional sentence and the target is a seq of the extracted information. An example would be like this - <\/p>\n\n<p>Input: <\/p>\n\n<blockquote>\n  <p>You gave 100.00 to John on 13-08-2018.<\/p>\n<\/blockquote>\n\n<p>Target:<\/p>\n\n<blockquote>\n  <p>100.00 13-08-2018 John <\/p>\n<\/blockquote>\n\n<p>Here all my target seq will maintain a fixed format like the first one is the amount, the second one is date the third one is the transactionWith etc.<\/p>\n\n<p>Can I do this like a language translation model with encoder for input seq and decoder for target seq and how can I make sure that my predicted seq for test data is from within the vocabulary of the given single input sentence and not from the entire target vocabulary?<\/p>\n\n<p>Thank you all the awesome people in advance. :)<\/p>\n","answers":[]}
+{"QuestionId":53503985,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning><softmax><activation-function>","CreationDate":"2018-11-27T16:23:34.873","AcceptedAnswerId":53504317.0,"OwnerUserId":309798.0,"Title":"What is the difference between keras.activations.softmax and keras.layers.Softmax?","Body":"<p>What is the difference between <code>keras.activations.softmax<\/code> and <code>keras.layers.Softmax<\/code>? Why are there two definitions of the same activation function?<\/p>\n\n<p><code>keras.activations.softmax<\/code>: <a href=\"https:\/\/keras.io\/activations\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/activations\/<\/a><\/p>\n\n<p><code>keras.layers.Softmax<\/code>: <a href=\"https:\/\/keras.io\/layers\/advanced-activations\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/advanced-activations\/<\/a><\/p>\n","answers":[{"AnswerId":"53504317","CreationDate":"2018-11-27T16:43:18.950","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>They are equivalent to each other in terms of what they do. Actually, the <code>Softmax<\/code> layer would call the <code>activations.softmax<\/code> <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/920e8af34a43ad2cd11190a21200a2acbfd83e11\/keras\/layers\/advanced_activations.py#L253\" rel=\"nofollow noreferrer\">under the hood<\/a>:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>def call(self, inputs):\n    return activations.softmax(inputs, axis=self.axis)\n<\/code><\/pre>\n\n<p>However, their difference is that the <code>Softmax<\/code> layer could be directly used as a layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Softmax\n\nsoft_out = Softmax()(input_tensor)\n<\/code><\/pre>\n\n<p>But, <code>activations.softmax<\/code> could not be used directly as a layer. Rather, you can pass it as the activation function of other layers through <code>activation<\/code> argument:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import activations\n\ndense_out = Dense(n_units, activation=activations.softmax)\n<\/code><\/pre>\n\n<p>Further, note that the good thing about using <code>Softmax<\/code> layer is that it takes an <code>axis<\/code> argument and you can compute the softmax over another axis of the input instead of its last axis (which is the default):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>soft_out = Softmax(axis=desired_axis)(input_tensor)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53504149,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-27T16:33:17.570","AcceptedAnswerId":null,"OwnerUserId":6109283.0,"Title":"Train large number of independent NNs with tensorflow","Body":"<p>I have a setting where I need to train a large number of independent NNs. They all share the same architecture, but train and test data is different for each one. The problem I'm finding is that RAM use increases and performance degrades very quickly. Question: how to do this in a efficient and scalable way?<\/p>\n\n<p>I tried a simple method:<\/p>\n\n<pre><code>for i in range(1000):\n    # Create NN\n    model = keras.Sequential(...)\n    optimizer = keras.optimizers.Adam()\n    model.compile(loss=loss, optimizer=optimizer, metrics=metrics)\n\n    # Train and test\n    Xtrain, ytrain, Xtest, ytest = loadData(i)\n    history = model.fit(XTrain, yTrain, epochs=epochs, validation_data=(Xtest, ytest))\n\n    # Save data\n    allHistory.append(history)\n<\/code><\/pre>\n\n<p>and then tried to reuse a single NN, reset weights and train:<\/p>\n\n<pre><code># Create NN\nmodel = keras.Sequential(...)\noptimizer = keras.optimizers.Adam()\nmodel.compile(loss=loss, optimizer=optimizer, metrics=metrics)\n# Save fresh model\nmodel.save(savePath)\n\nfor i in range(1000):\n    # Load fresh model\n    model = keras.models.load_model(savePath)\n\n    # Train and test\n    Xtrain, ytrain, Xtest, ytest = loadData(i)\n    history = model.fit(XTrain, yTrain, epochs=epochs, validation_data=(Xtest, ytest))\n\n    # Save data\n    allHistory.append(history)\n<\/code><\/pre>\n\n<p>performance and RAM use seem similar in both cases, so I guess TensorFlow keeps growing its graph in every iteration.<\/p>\n","answers":[{"AnswerId":"53504368","CreationDate":"2018-11-27T16:45:42.367","ParentId":null,"OwnerUserId":"10585322","Title":null,"Body":"<p>Maybe you can use <code>import thread<\/code> and train every NN in an own Thread. This should work since multithreading makes operations sequential and helps to improve its scalability.<\/p>\n\n<p>I think you'll get good results with:<\/p>\n\n<pre><code>for i in range(0,1000,5):\n    t1 = Thread(target=NN, args=arguments(i+1))\n    ...\n    t5 = Thread(target=NN, args=arguments(i+5))\n    t1.start()\n    t1.join()\n    ...\n    t5.start()\n    t5.join()\n<\/code><\/pre>\n"}]}
+{"QuestionId":53504813,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2018-11-27T17:12:27.480","AcceptedAnswerId":53513741.0,"OwnerUserId":8767101.0,"Title":"Custom Keras metrics and losses' names","Body":"<p>I have a very big model with many losses and metrics.\nWhen I do <code>print(np.array(self.model.metrics_names))<\/code>\nI get this:<\/p>\n\n<pre><code>['loss' 'autoencoder_loss' 'autoencoder_loss' 'autoencoder_loss'\n 'autoencoder_loss' 's_regularisation_phase_loss'\n 'gen_regularisation_phase_loss' 's_regularisation_phase_loss'\n 'z_regularisation_phase_loss' 'gen_regularisation_phase_loss'\n 'z_regularisation_phase_loss' 'gen_regularisation_phase_loss'\n 'gen_regularisation_phase_loss' 'autoencoder_categorical_accuracy'\n 'autoencoder_output' 'autoencoder_categorical_accuracy_1'\n 'autoencoder_output_1' 'autoencoder_categorical_accuracy_2'\n 'autoencoder_output_2' 'autoencoder_categorical_accuracy_3'\n 'autoencoder_output_3' 's_regularisation_phase_categorical_accuracy'\n 's_regularisation_phase_output'\n 'gen_regularisation_phase_categorical_accuracy'\n 'gen_regularisation_phase_output'\n 's_regularisation_phase_categorical_accuracy_1'\n 's_regularisation_phase_output_1'\n 'z_regularisation_phase_categorical_accuracy'\n 'z_regularisation_phase_output'\n 'gen_regularisation_phase_categorical_accuracy_1'\n 'gen_regularisation_phase_output_1'\n 'z_regularisation_phase_categorical_accuracy_1'\n 'z_regularisation_phase_output_1'\n 'gen_regularisation_phase_categorical_accuracy_2'\n 'gen_regularisation_phase_output_2'\n 'gen_regularisation_phase_categorical_accuracy_3'\n 'gen_regularisation_phase_output_3']\n<\/code><\/pre>\n\n<p>Is there a way to give them more meaningful names?<\/p>\n","answers":[{"AnswerId":"53513741","CreationDate":"2018-11-28T06:53:38.583","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>The names before each <code>_loss<\/code> and <code>_accuracy<\/code> came from the names of output layers.  If you want to modify this names you should rename the output layers. <\/p>\n\n<p>Consider the following model.<\/p>\n\n<pre><code>input_ =  keras.layers.Input(shape=(8,))\nx =  keras.layers.Dense(16)(input_)\noutput1 = keras.layers.Dense(32, name=\"output1\")(x)\noutput2 = keras.layers.Dense(32, name=\"output2\")(x)\nmodel = keras.models.Model(inputs=input_, outputs=[output1, output2])\nmodel.compile(loss=[\"mse\", \"mae\"], optimizer=\"adam\", metrics={\"output1\":\"accuracy\",\"output2\":\"accuracy\"})\n<\/code><\/pre>\n\n<p>Now <code>model.metrics_names<\/code> will give you the following list<\/p>\n\n<pre><code>['loss', 'output1_loss', 'output2_loss', 'output1_acc', 'output2_acc']\n<\/code><\/pre>\n"}]}
+{"QuestionId":53504937,"AnswerCount":1,"Tags":"<python><tensorflow><pip><ubuntu-14.04>","CreationDate":"2018-11-27T17:19:58.743","AcceptedAnswerId":53541704.0,"OwnerUserId":4068678.0,"Title":"pip canot find tensorflow when in virtualenv with specified python","Body":"<p>I am experiencing a weird issue, on ubuntu 14.04 with python2.7.9 installed in a custom folder when trying to install tensorflow from inside a virtualenv that is using the custom python build pip doesn't seem to find tensorflow.<\/p>\n\n<pre><code> virtualenv venv --python=\/opt\/python279\/bin\/python2.7\n cd venv\n source bin\/activate\n pip install tensorflow\n<\/code><\/pre>\n\n<blockquote>\n  <p>Collecting tensorflow\n    Could not find a version that satisfies the requirement tensorflow (from versions: )\n  No matching distribution found for tensorflow<\/p>\n<\/blockquote>\n\n<p>other packages install normally :\/ I am utterly confused<\/p>\n","answers":[{"AnswerId":"53541704","CreationDate":"2018-11-29T14:52:12.607","ParentId":null,"OwnerUserId":"2650249","Title":null,"Body":"<p>Some time ago, I wrote <a href=\"https:\/\/stackoverflow.com\/a\/49342341\/2650249\">a more or less comprehensive checklist<\/a> for possible mismatches leading to the <code>Could not find a version that satisfies requirement<\/code> error. Although the question is specifically about MacOS, the answer is applicable to Linux also. In your case, you have an ABI mismatch: as found out in the comments,<\/p>\n\n<pre><code>\/opt\/python279\/bin\/python2.7 -c \"import sysconfig;\\\n    abiflags=('m' if sysconfig.get_config_var('WITH_PYMALLOC') else '')+\\\n    ('u' if sysconfig.get_config_var('WITH_WIDE_UNICODE') else '')+\\\n    ('d' if sysconfig.get_config_var('WITH_PYDEBUG') else ''); print(abiflags)\" \n<\/code><\/pre>\n\n<p>returned <code>m<\/code>, indicating that this Python distribution was compiled without the wide unicode support (missing the <code>u<\/code> flag). This means that <code>pip<\/code> will install only <code>cp27-cp27m-manylinux1_x86_64<\/code> wheels for this Python distribution. However, the only wheel <a href=\"https:\/\/pypi.org\/project\/tensorflow\/#files\" rel=\"nofollow noreferrer\"><code>tensorflow<\/code> offers<\/a> for Python 2.7 on Linux is the <code>cp27-cp27mu-manylinux1_x86_64<\/code> one. To be able to install <code>tensorflow<\/code>, you'll need to build a Python 2.7 distribution with wide unicode support:<\/p>\n\n<pre><code>$ cd Python2.7-src\n$ configure --enable-unicode=ucs4\n$ make &amp;&amp; make install\n<\/code><\/pre>\n"}]}
+{"QuestionId":53505149,"AnswerCount":0,"Tags":"<python><vector><pytorch><xor><scalar>","CreationDate":"2018-11-27T17:34:18.403","AcceptedAnswerId":null,"OwnerUserId":6408523.0,"Title":"Getting error: only one element tensors can be converted to Python scalars","Body":"<p>Could you please help how to solve this problem. Basically, I'm trying to get into Pytorch tensor function data that is vector not scalar. X1 and X2 are basically columns in CSV file that contains many strings. \nHow to kind of iterate through every single data from x1 and x2 and not just trying to parse the whole vector. I'm a newbie at Python and Pytorch as well.<\/p>\n\n<pre><code>import torch\nimport random\nimport pandas\n\ndata = pandas.read_csv('train\/train.tsv', sep='\\t')\n\n\nlearningrate = torch.tensor(0.01)\nW = torch.rand([2, 2], dtype=torch.float, requires_grad=True)\nb = torch.rand(2, dtype=torch.float, requires_grad=True)\nU = torch.rand(2, dtype=torch.float, requires_grad=True)\nc = torch.rand(1, dtype=torch.float, requires_grad=True)\n\n\ndef get_item():\n x1 = torch.tensor(data['Powierzchnia w m2'],\n                  dtype=torch.float, requires_grad=True)\n x2 = torch.tensor(data['Liczba pokoi'],\n                  dtype=torch.float, requires_grad=True)\n x = torch.tensor([x1, x2], dtype=torch.float)\n yexpected = torch.tensor(data['cena'].values, dtype=torch.float)\n return x, yexpected\n\n\nfor _ in range(100000):\n\n x, yexpected = get_item()\n h = torch.sigmoid(W @ x+b)\n\nprint(x)\nprint(yexpected)\nprint(h)\ny = torch.sigmoid(U@h+c)\nloss = (y-yexpected)**2\nprint(loss)\nloss.backward()\nwith torch.no_grad():\n    W -= learningrate * W.grad\n    b -= learningrate * b.grad\n    c -= learningrate * c.grad\n    U -= learningrate * U.grad\n    b.grad.zero_()\n    W.grad.zero_()\n    c.grad.zero_()\n    U.grad.zero_()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53505150,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2018-11-27T17:34:22.707","AcceptedAnswerId":53661358.0,"OwnerUserId":3486060.0,"Title":"Using tensorflow dataset with stratified sampling","Body":"<p>Given a tensorflow dataset<\/p>\n\n<pre><code>Train_dataset = tf.data.Dataset.from_tensor_slices((Train_Image_Filenames,Train_Image_Labels))\nTrain_dataset = Train_dataset.map(Parse_JPEG_Augmented)\n...\n<\/code><\/pre>\n\n<p>I would like to stratify my batches to deal with class imbalance. I found tf.contrib.training.stratified_sample and thought I could use it in the following way:<\/p>\n\n<pre><code>Train_dataset_iter = Train_dataset.make_one_shot_iterator()\nTrain_dataset_Image_Batch,Train_dataset_Label_Batch = Train_dataset_iter.get_next()\nTrain_Stratified_Images,Train_Stratified_Labels = tf.contrib.training.stratified_sample(Train_dataset_Image_Batch,Train_dataset_Label_Batch,[1\/Classes]*Classes,Batch_Size)\n<\/code><\/pre>\n\n<p>But it gives the following error and I'm not sure that this would allow me to keep the performance benefits of tensorflow dataset as I may have then have to pass <code>Train_Stratified_Images<\/code> and <code>Train_Stratified_Labels<\/code> via feed_dict ?<\/p>\n\n<pre><code>File \"\/xxx\/xxx\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/training\/python\/training\/sampling_ops.py\", line 192, in stratified_sample\nwith ops.name_scope(name, 'stratified_sample', list(tensors) + [labels]):\nFile \"\/xxx\/xxx\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 459, in __iter__\n\"Tensor objects are only iterable when eager execution is \"\nTypeError: Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.\n<\/code><\/pre>\n\n<p>What would be the \"best practice\" way of using dataset with stratified batches?<\/p>\n","answers":[{"AnswerId":"53661358","CreationDate":"2018-12-06T23:47:24.800","ParentId":null,"OwnerUserId":"4838303","Title":null,"Body":"<p>I am looking into a similar issue and I found <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/experimental\/rejection_resample\" rel=\"nofollow noreferrer\">this<\/a>. I haven't tried it in my pipeline though. Maybe it will work for your purpose?<\/p>\n"},{"AnswerId":"58053992","CreationDate":"2019-09-22T23:01:30.800","ParentId":null,"OwnerUserId":"3965888","Title":null,"Body":"<p>Here is below a simple example to demonstrate the usage of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/experimental\/sample_from_datasets\" rel=\"nofollow noreferrer\">sample_from_datasets<\/a> (thanks @Agade for the idea).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import math\nimport tensorflow as tf\nimport numpy as np\n\n\ndef print_dataset(name, dataset):\n    elems = np.array([v.numpy() for v in dataset])\n    print(\"Dataset {} contains {} elements :\".format(name, len(elems)))\n    print(elems)\n\n\ndef combine_datasets_balanced(dataset_smaller, size_smaller, dataset_bigger, size_bigger, batch_size):\n    ds_smaller_repeated = dataset_smaller.repeat(count=int(math.ceil(size_bigger \/ size_smaller)))\n    # we repeat the smaller dataset so that the 2 datasets are about the same size\n    balanced_dataset = tf.data.experimental.sample_from_datasets([ds_smaller_repeated, dataset_bigger], weights=[0.5, 0.5])\n    # each element in the resulting dataset is randomly drawn (without replacement) from dataset even with proba 0.5 or from odd with proba 0.5\n    balanced_dataset = balanced_dataset.take(2 * size_bigger).batch(batch_size)\n    return balanced_dataset\n\n\nN, M = 3, 10\neven = tf.data.Dataset.range(0, 2 * N, 2).repeat(count=int(math.ceil(M \/ N)))\nodd = tf.data.Dataset.range(1, 2 * M, 2)\neven_odd = combine_datasets_balanced(even, N, odd, M, 2)\n\nprint_dataset(\"even\", even)\nprint_dataset(\"odd\", odd)\nprint_dataset(\"even_odd_all\", even_odd)\n<\/code><\/pre>\n\n<pre><code>Output :\n\nDataset even contains 12 elements :  # 12 = 4 x N  (because of .repeat)\n[0 2 4 0 2 4 0 2 4 0 2 4]\nDataset odd contains 10 elements :\n[ 1  3  5  7  9 11 13 15 17 19]\nDataset even_odd contains 10 elements :  # 10 = 2 x M \/ 2  (2xM because of .take(2 * M) and \/2 because of .batch(2))\n[[ 0  2]\n [ 1  4]\n [ 0  2]\n [ 3  4]\n [ 0  2]\n [ 4  0]\n [ 5  2]\n [ 7  4]\n [ 0  9]\n [ 2 11]] \n<\/code><\/pre>\n"}]}
+{"QuestionId":53505176,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network><lstm><rnn>","CreationDate":"2018-11-27T17:35:53.593","AcceptedAnswerId":53639353.0,"OwnerUserId":9145046.0,"Title":"How to use Conv2D and LSTM layers with Keras?","Body":"<p>I have a model which works with <strong>Conv2D<\/strong> using Keras but I would like to <strong>add a LSTM<\/strong> layer. This is the data I am using: <\/p>\n\n<ul>\n<li>x_train with shape (13984, <strong>334, 35, 1<\/strong>)<\/li>\n<li>y_train with shape (13984, <strong>5<\/strong>)<\/li>\n<\/ul>\n\n<p>My model without LSTM is: <\/p>\n\n<pre><code>inputs = Input(name='input',shape=(334,35,1))\nlayer = Conv2D(64, kernel_size=3,activation='relu',data_format='channels_last')(inputs)\nlayer = Flatten()(layer)\npredictions = Dense(5, activation='softmax')(layer)\nnetwork = Model(inputs=inputs, outputs=predictions)\nnetwork.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>What is <strong>the correct way of adding a LSTM<\/strong> layer just before the Dense layer? <\/p>\n\n<p>I tried to use TimeDistributed or Reshape\/Permute but I always get errors. <\/p>\n","answers":[{"AnswerId":"53505638","CreationDate":"2018-11-27T18:06:35.253","ParentId":null,"OwnerUserId":"3482325","Title":null,"Body":"<p>It seems like your question is similar to one that i had yesterday. The answer can be found here: <a href=\"https:\/\/stackoverflow.com\/questions\/53488768\/keras-functional-api-combine-cnn-model-with-a-rnn-to-to-look-at-sequences-of-im\">Keras functional API: Combine CNN model with a RNN to to look at sequences of images<\/a><\/p>\n"},{"AnswerId":"53639353","CreationDate":"2018-12-05T19:21:58.150","ParentId":null,"OwnerUserId":"9145046","Title":null,"Body":"<p>The method explained by the user deKeijzer works. I found another way to solve the problem. It is to use a Reshape layer (reshaping by (334,35)) just after the last Conv2D layer and then add LSTM layers. <\/p>\n"}]}
+{"QuestionId":53505244,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T17:40:13.087","AcceptedAnswerId":null,"OwnerUserId":10712924.0,"Title":"No module named 'tensorflow.python._impl","Body":"<p>I want to run code for try neural network with python but \n<strong>No module named 'tensorflow.python._impl<\/strong><\/p>\n\n<p>part of my code:<\/p>\n\n<p>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.python._impl import activations<\/p>\n\n<p>I did all solutions that some sites said such as\n - Uninstall and install package like tensorflow,..\nalso:\n - Upgrade tensorflow\n - Upgrade keras <\/p>\n\n<p>but it doesn't solve yet. I have this error on co lab too.\nplease help me if you have any solution<\/p>\n\n<p>thanks<\/p>\n","answers":[]}
+{"QuestionId":53505690,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2018-11-27T18:09:58.777","AcceptedAnswerId":null,"OwnerUserId":1316501.0,"Title":"Tensorflow: Dataset Map With num_parallel_calls Offers No Speedup","Body":"<p>I'm using TensorFlow and the <code>tf.data.Dataset<\/code> API to perform some text preprocessing.  Without using <code>num_parallel_calls<\/code> in my <code>dataset.map<\/code> call, it takes 0.03s to preprocess 10K records.<\/p>\n\n<p>When I use <code>num_parallel_trials=8<\/code> (the number of cores on my machine), it also takes 0.03s to preprocess 10K records.<\/p>\n\n<p>I googled around and came across this: <a href=\"https:\/\/stackoverflow.com\/questions\/48484689\/parallelism-isnt-reducing-the-time-in-dataset-map\">Parallelism isn&#39;t reducing the time in dataset map<\/a><\/p>\n\n<p>where they indicate that you need to use TensorFlow operations to see a speedup.  Here's the thing: I <em>am<\/em> using only TensorFlow operations.  Specifically, I'm mapping this function:<\/p>\n\n<pre><code>def preprocess(self, x, data_table):\n    x['reviews'] = tf.string_split(x['reviews'], delimiter=' ')\n    x['reviews'] = tf.sparse_tensor_to_dense(x['reviews'], default_value=' ')\n    x['reviews'] = tf.cast(data_table.lookup(x['reviews']), tf.int32)\n    nbatch = tf.cast(tf.shape(x['reviews'])[0], tf.int32)\n    nseq = tf.cast(tf.shape(x['reviews'])[1], tf.int32)\n    padding = tf.cond(tf.less(nseq, 100),\n                      lambda: 0 * tf.ones([nbatch, 100 - nseq], tf.int32),\n                      lambda: 0 * tf.ones([nbatch, 0], tf.int32))\n    x['reviews'] = tf.concat((x['reviews'], padding), axis=1)[:, :100]\n    x['reviews'].set_shape([None, 100])\n    return x\n<\/code><\/pre>\n\n<p>Any idea why I don't see any speedup?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53549042","CreationDate":"2018-11-29T23:21:10.953","ParentId":null,"OwnerUserId":"8418486","Title":null,"Body":"<p>My first assumption would be the calls to lambda are crushing your speed as lazy initialization for each iteration and core.   According to this url he has a similar problem to speed and core use . <a href=\"https:\/\/stackoverflow.com\/questions\/47045026\/is-there-a-way-to-use-tensorflow-map-fn-on-gpu\">Is there a way to use tensorflow map_fn on GPU?<\/a> \nI am pretty much a beginner with tensor and piping but I\u2019ll look into later when I have access to computer,  I would like to know what executables are being run where though.<\/p>\n"}]}
+{"QuestionId":53505863,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-27T18:20:45.920","AcceptedAnswerId":null,"OwnerUserId":9854132.0,"Title":"Explanation of parallel arguments of tf.while_loop in TensorFlow","Body":"<p>I want to implement an algorithm which allows a parallel implementation in TensorFlow. My question is what the arguments parallel_iterations, swap_memory and maximum_iterations actually do and which are their appropriate values according the situation. Specifically, in the documentation on TensorFlow's site <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop<\/a> says that parallel_iterations are the number of iterations allowed to run in parallel. Is this number the number of threads? When someone should allow CPU-GPU swap memory and for what reason? What are the advantages and disadvantages from this choice? What is the purpose of maximum_iterations? Can it be combined with parallel_iterations?<\/p>\n","answers":[{"AnswerId":"53508389","CreationDate":"2018-11-27T21:26:10.147","ParentId":null,"OwnerUserId":"3670165","Title":null,"Body":"<p><strong>swap_memory<\/strong> is used when you want to have extra memory on the GPU device. Usually when you are training a model some activations are saved in the GPU mem. for later use. With swap_memory, you can store those activations in the CPU memory and use the GPU mem. to fit e.g. larger batch sizes. And this is an advantage. You would choose this if you need big batch_sizes or have long sequences and want to avoid OOM exceptions. Disadvantage is computation time since you need to transfer the data from CPU mem. to GPU mem.<\/p>\n\n<p>The <strong>maximum iterations<\/strong> is smth. like this:<\/p>\n\n<pre><code>while num_iter &lt; 100 and &lt;some condition&gt;:\n    do something\n    num_iter += 1 \n<\/code><\/pre>\n\n<p>So it is useful when you check a condition, but also want to have an upper bound (one example is to check if your model converges. If it doesn't you still want to end after k iterations.)<\/p>\n\n<p>As for <strong>parallel_iterations<\/strong> I am not sure, but it sounds like multiple threads, yes. You can try and see the effect in a sample script.<\/p>\n"}]}
+{"QuestionId":53505865,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network>","CreationDate":"2018-11-27T18:20:52.913","AcceptedAnswerId":null,"OwnerUserId":6266336.0,"Title":"Keras. How to change model for it not to use Embedding layer","Body":"<p>There is a BLSTM RNN that is currectly trained with an embedding layer. I want to change it to train purely on the input sequences. I am somehow not completely sure where to start making changes. <\/p>\n\n<p>Does anyone have sugestions? <\/p>\n\n<pre><code>embedding_layer = Embedding(len(word2int) + 1,\n                            EMBEDDING_DIM,\n                            weights=[embedding_matrix],\n                            input_length=MAX_SEQUENCE_LENGTH,\n                            trainable=True)\nsequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')\nembedded_sequences = embedding_layer(sequence_input)\n\nl_lstm = Bidirectional(LSTM(64, return_sequences=True))(embedded_sequences)\npreds = TimeDistributed(Dense(n_tags + 1, activation='softmax'))(l_lstm)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53508532","CreationDate":"2018-11-27T21:35:26.440","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>I'm interpreting <em>change it to train purely on the input sequences<\/em> two ways:<\/p>\n\n<ul>\n<li><p>Either you want your LSTM to actually run on integers <code>[34, 21, 8, ...]<\/code> directly in which case you need to say that every timestep there is only 1 integer, i.e. convert <code>(batch_size, max_seq_len)<\/code> to <code>(batch_size, max_seq_len, 1)<\/code> which can be achieved using <code>embedded_seqs = Lambda(lambda x: K.expand_dims(x, -1))(sequence_input)<\/code>. This will force the model to see actual integer values but I wouldn't recommend it.<\/p><\/li>\n<li><p>Another option would be one hot encode it so there is no trainable embedding but a large vector for each word with a 1 at the corresponding index. You can do this by removing <code>weights=[embedding_matrix]<\/code> and adding <code>embeddings_initializer='eye'<\/code> to your <code>Embedding<\/code> layer setting <code>trainable=False<\/code>.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53506781,"AnswerCount":1,"Tags":"<tensorflow><keras><google-colaboratory><google-cloud-tpu><mlp>","CreationDate":"2018-11-27T19:27:13.197","AcceptedAnswerId":null,"OwnerUserId":1919641.0,"Title":"Google Colab TPU takes more time than GPU","Body":"<p>Below is the code I am using. I commented out the line to convert my model to the TPU model. With GPU for the same amount of data it's taking 7 seconds for an epoch while using TPU it takes 90 secs. <\/p>\n\n<pre><code>    Inp = tf.keras.Input(name='input', shape=(input_dim,), dtype=tf.float32)\n    x = tf.keras.layers.Dense(900, kernel_initializer='uniform',  activation='relu', input_dim=input_dim, name = 'Dense_01')(Inp)\n    x = tf.keras.layers.Dropout(0.3, name = 'Dropout_02')(x)\n    output = tf.keras.layers.Dense(stop_criteria, activation='softmax',name = 'Dense_02')(x)\n\n    model = tf.keras.Model(inputs=[Inp], outputs=[output])\n    opt = tf.train.AdamOptimizer(.001)\n    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['acc'])\n\n    '''tpu_model = tf.contrib.tpu.keras_to_tpu_model(model,\n                                                  strategy=tf.contrib.tpu.TPUDistributionStrategy(\n                                                      tf.contrib.cluster_resolver.TPUClusterResolver(TPU_ADDRESS)))'''\n    model.fit(X_tra, y_tra, epochs=5, batch_size=batch_size, shuffle=False,\n              validation_split=0.1, verbose=2)\n<\/code><\/pre>\n\n<p>Here is the link to the <a href=\"https:\/\/colab.research.google.com\/drive\/15hEdj-YIGFLoYWlpkCwMtavbKxgdLqUe\" rel=\"nofollow noreferrer\">notebook<\/a><\/p>\n","answers":[{"AnswerId":"53508092","CreationDate":"2018-11-27T21:02:24.110","ParentId":null,"OwnerUserId":"3670165","Title":null,"Body":"<p>Have you tried the <code>tpu_model.fit_generator<\/code> method like in the example below?\nThe other part looks fine. \nAlso, one problem could be the use of Adam Optimizer. There was smth. about it, but I forgot where the link is. Try another optimizer and the code below and if a different optimizer worked, you know it must be smth. with the Adam Optimizer.<\/p>\n\n<pre><code>tf.keras.backend.clear_session()\n\ntraining_model = lstm_model(seq_len=100, batch_size=128, stateful=False)\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    training_model,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))\n\ntpu_model.fit_generator(\n    training_generator(seq_len=100, batch_size=1024),\n    steps_per_epoch=100,\n    epochs=10,\n)\ntpu_model.save_weights('\/tmp\/bard.h5', overwrite=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53507039,"AnswerCount":1,"Tags":"<c++><pytorch>","CreationDate":"2018-11-27T19:45:36.527","AcceptedAnswerId":53532295.0,"OwnerUserId":300908.0,"Title":"Row-wise Element Indexing in PyTorch for C++","Body":"<p>I am using the <a href=\"https:\/\/pytorch.org\/cppdocs\/frontend.html\" rel=\"nofollow noreferrer\">C++ frontend for PyTorch<\/a> and am struggling with a relatively basic indexing problem.<\/p>\n\n<p>I have an <code>8<\/code> by <code>6<\/code> Tensor such as the one below:<\/p>\n\n<pre><code>[ Variable[CUDAFloatType]{8,6} ] \n                 0           1           2           3           4           5\n0       1.7107e-14  4.0448e-17  4.9708e-06  1.1664e-08  9.9999e-01  2.1857e-20\n1       1.8288e-14  5.9356e-17  5.3042e-06  1.2369e-08  9.9999e-01  2.4799e-20\n2       2.6828e-04  9.0390e-18  1.7517e-02  1.0529e-03  9.8116e-01  6.7854e-26\n3       5.7521e-10  3.1037e-11  1.5021e-03  1.2304e-06  9.9850e-01  1.4888e-17\n4       1.7811e-13  1.8383e-15  1.6733e-05  3.8466e-08  9.9998e-01  5.2815e-20\n5       9.6191e-06  2.6217e-23  3.1345e-02  2.3024e-04  9.6842e-01  2.9435e-34\n6       2.2653e-04  8.4642e-18  1.6085e-02  9.7405e-04  9.8271e-01  6.3059e-26\n7       3.8951e-14  2.9903e-16  8.3518e-06  1.7974e-08  9.9999e-01  3.6993e-20\n<\/code><\/pre>\n\n<p>I have another Tensor with just <code>8<\/code> elements in it such as:<\/p>\n\n<pre><code>[ Variable[CUDALongType]{8} ] \n 0\n 3\n 4\n 4\n 4\n 4\n 4\n 4\n<\/code><\/pre>\n\n<p>I would like to index the rows of my first tensor using the second to produce:<\/p>\n\n<pre><code>        0           \n0       1.7107e-14  \n1       1.2369e-08\n2       9.8116e-01  \n3       9.9850e-01  \n4       9.9998e-01\n5       9.6842e-01  \n6       9.8271e-01  \n7       9.9999e-01\n<\/code><\/pre>\n\n<p>I have tried a few different approaches including <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.index_select\" rel=\"nofollow noreferrer\"><code>index_select<\/code><\/a> but it seems to produce an output that has the same dimensions as the input (<code>8x6<\/code>).<\/p>\n\n<p>In Python I think I could index with Python's built-in indexing as discussed here: <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/1080\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/1080<\/a><\/p>\n\n<p>Unfortunately, in C++ I can only index a Tensor with a scalar (zero-dimensional Tensor) so I don't think that approach works for me here.<\/p>\n\n<p>How can I achieve my desired result without resorting to loops?<\/p>\n","answers":[{"AnswerId":"53532295","CreationDate":"2018-11-29T05:16:16.240","ParentId":null,"OwnerUserId":"300908","Title":null,"Body":"<p>It turns out you can do this in a couple different ways. One with <code>gather<\/code> and one with <code>index<\/code>. From the <a href=\"https:\/\/discuss.pytorch.org\/t\/row-wise-element-indexing-in-pytorch-for-c\/30705\" rel=\"nofollow noreferrer\">PyTorch discussions<\/a> where I asked the same question:<\/p>\n\n<p>Using <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.gather\" rel=\"nofollow noreferrer\"><code>torch::gather<\/code><\/a><\/p>\n\n<pre><code>auto x = torch::randn({8, 6});\nint64_t idx_data[8] = { 0, 3, 4, 4, 4, 4, 4, 4 };\nauto idx = x.type().toScalarType(torch::kLong).tensorFromBlob(idx_data, 8);\nauto result = x.gather(1, idx.unsqueeze(1));\n<\/code><\/pre>\n\n<p>Using the C++ specific <a href=\"https:\/\/pytorch.org\/cppdocs\/api\/function_namespaceat_1aa35c2bf04b75e2a7e61fcbe034acc5d4.html?highlight=index\" rel=\"nofollow noreferrer\"><code>torch::index<\/code><\/a><\/p>\n\n<pre><code>auto x = torch::randn({8, 6});\nint64_t idx_data[8] = { 0, 3, 4, 4, 4, 4, 4, 4 };\nauto idx = x.type().toScalarType(torch::kLong).tensorFromBlob(idx_data, 8);\nauto rows = torch::arange(0, x.size(0), torch::kLong);\nauto result = x.index({rows, idx});\n<\/code><\/pre>\n"}]}
+{"QuestionId":53507255,"AnswerCount":1,"Tags":"<deep-learning><pytorch>","CreationDate":"2018-11-27T20:01:26.070","AcceptedAnswerId":null,"OwnerUserId":5223033.0,"Title":"How to reshape a pytorch matrix without mixing elements of items in a batch","Body":"<p>In my Neural network model, I represent  an 8 word-sentence with a <code>8x256<\/code> dimensional embedding matrix. I want to give it to a LSTM as a input where LSTM takes a single word embedding at a time as input and process it. According to pytorch <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.LSTM\" rel=\"nofollow noreferrer\">documentation<\/a>, the input should be in the shape of  <code>(seq_len, batch, input_size)<\/code>. What is the correct way to convert my input to desired shape ? I don't want to mixup the numbers by mistake. I am quite new in PyTorch and row-major calculations, therefore I wanted to ask it here. I do it as follows, is it correct ? <\/p>\n\n<pre><code>x = torch.rand(8,256)\nlstm_input = torch.reshape(x,(8,1,256))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53508817","CreationDate":"2018-11-27T21:56:32.383","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your solution is correct: you added a Singleton dimension for the \"batch\" dimension, leaving <code>x<\/code> to be with temporal dimension 8 and input dimension 256.<\/p>\n\n<p>Since you are new to pytorch, here are a few equivalent ways of doing the same thing:<\/p>\n\n<pre><code>x = x[:, None, :]\n<\/code><\/pre>\n\n<p>Putting <code>None<\/code> in the dim=1 indicates to pytorch to add a singelton dimension.<br>\nAnother way is to use <code>view<\/code>:<\/p>\n\n<pre><code>x = x.view(8, 1, 256)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53507266,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-27T20:02:04.077","AcceptedAnswerId":53734856.0,"OwnerUserId":8100895.0,"Title":"Failing to fetch Tensorflow's convolution algorithm","Body":"<p>I am printing the following the error message:<\/p>\n\n<pre><code>UnknownError                              Traceback (most recent call last)\n&lt;ipython-input-11-e73400b11710&gt; in &lt;module&gt;()\n      1 earlystopper = EarlyStopping(patience=6, verbose=1)\n----&gt; 2 history = parallel_model.fit(X_train, Y_train, validation_split=0.25, batch_size = 16, verbose=1, epochs=30, callbacks=[earlystopper])\n      3 model_out = parallel_model.layers[-2]\n      4 model_out.save_weights(filepath=\".\/multi_class.hdf5\")\n\n~\/anaconda\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)\n   1037                                         initial_epoch=initial_epoch,\n   1038                                         steps_per_epoch=steps_per_epoch,\n-&gt; 1039                                         validation_steps=validation_steps)\n   1040 \n   1041     def evaluate(self, x=None, y=None,\n\n~\/anaconda\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py in fit_loop(model, f, ins, out_labels, batch_size, epochs, verbose, callbacks, val_f, val_ins, shuffle, callback_metrics, initial_epoch, steps_per_epoch, validation_steps)\n    197                     ins_batch[i] = ins_batch[i].toarray()\n    198 \n--&gt; 199                 outs = f(ins_batch)\n    200                 outs = to_list(outs)\n    201                 for l, o in zip(out_labels, outs):\n\n~\/anaconda\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in __call__(self, inputs)\n   2713                 return self._legacy_call(inputs)\n   2714 \n-&gt; 2715             return self._call(inputs)\n   2716         else:\n   2717             if py_any(is_tensor(x) for x in inputs):\n\n~\/anaconda\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in _call(self, inputs)\n   2673             fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)\n   2674         else:\n-&gt; 2675             fetched = self._callable_fn(*array_vals)\n   2676         return fetched[:len(self.outputs)]\n   2677 \n\n~\/anaconda\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1437           ret = tf_session.TF_SessionRunCallable(\n   1438               self._session._session, self._handle, args, status,\n-&gt; 1439               run_metadata_ptr)\n   1440         if run_metadata:\n   1441           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n~\/anaconda\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)\n    526             None, None,\n    527             compat.as_text(c_api.TF_Message(self.status.status)),\n--&gt; 528             c_api.TF_GetCode(self.status.status))\n    529     # Delete the underlying status object from memory otherwise it stays alive\n    530     # as there is a reference to status from this from the traceback due to\n\nUnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.\n     [[{{node replica_0\/model_1\/conv2d_1\/convolution}} = Conv2D[T=DT_FLOAT, _class=[\"loc:@train...propFilter\"], data_format=\"NCHW\", dilations=[1, 1, 1, 1], padding=\"SAME\", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](training\/Adamax\/gradients\/replica_0\/model_1\/conv2d_1\/convolution_grad\/Conv2DBackpropFilter-0-TransposeNHWCToNCHW-LayoutOptimizer, conv2d_1\/kernel\/read)]]\n     [[{{node training\/Adamax\/gradients\/conv2d_transpose_5_1\/concat_grad\/Slice_2\/_1191}} = _Recv[client_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:2\", send_device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\", send_device_incarnation=1, tensor_name=\"edge_3830_training\/Adamax\/gradients\/conv2d_transpose_5_1\/concat_grad\/Slice_2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:2\"]()]]\n<\/code><\/pre>\n\n<p>However, I searched up the error message but couldn't find much information. Is this because cudnn wasn't downloaded properly or is this a different issue?<\/p>\n\n<p>I'm using the following code to download and setup cuda on my cluster. This was working perfectly fine until a week ago. <code>curl -O http:\/\/developer.download.nvidia.com\/compute\/redist\/cudnn\/v7.0.5\/cudnn-9.0-linux-x64-v7.tgz &amp;&amp; tar -xzvf cudnn-9.0-linux-x64-v7.tgz &amp;&amp; mkdir \/usr\/local\/cuda\/include &amp;&amp; cp cuda\/include\/cudnn.h \/usr\/local\/cuda\/include &amp;&amp; cp cuda\/lib64\/libcudnn* \/usr\/local\/cuda\/lib64 &amp;&amp; chmod a+r \/usr\/local\/cuda\/include\/cudnn.h &amp;&amp; chmod a+r \/usr\/local\/cuda\/lib64\/libcudnn* &amp;&amp; cp -P cuda\/include\/cudnn.h \/usr\/include &amp;&amp; cp -P cuda\/lib64\/libcudnn* \/usr\/lib\/x86_64-linux-gnu\/ &amp;&amp; chmod a+r \/usr\/lib\/x86_64-linux-gnu\/libcudnn* &amp;&amp; rm -r cuda<\/code><\/p>\n","answers":[{"AnswerId":"53734856","CreationDate":"2018-12-12T01:34:22.810","ParentId":null,"OwnerUserId":"8600470","Title":null,"Body":"<p>Requirement of the version of cudnn changed:\n<a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a><\/p>\n\n<p>cudnn version >= 7.2<\/p>\n\n<p>I update cudnn and it works well.<\/p>\n"}]}
+{"QuestionId":53507346,"AnswerCount":2,"Tags":"<python><python-3.x><pytorch><gradient>","CreationDate":"2018-11-27T20:07:46.880","AcceptedAnswerId":null,"OwnerUserId":3008221.0,"Title":"Gradient is none in pytorch when it shouldn't","Body":"<p>I am trying to get\/trace the gradient of a variable using pytorch, where I have that variable, pass it to a first function that looks for some minimum value of some other variable, then the output of the first function is inputted to a second function, and the whole thing repeats multiple times. <\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>import torch\n\ndef myFirstFunction(parameter_current_here):\n    optimalValue = 100000000000000\n    Optimal = 100000000000000\n    for j in range(2, 10):\n        i = torch.ones(1, requires_grad=True)*j\n        with torch.enable_grad():\n            optimalValueNow = i*parameter_current_here.sum()\n        if (optimalValueNow &lt; optimalValue):\n            optimalValue = optimalValueNow\n            Optimal = i\n    return optimalValue, Optimal\n\ndef mySecondFunction(Current):\n    with torch.enable_grad():\n        y = (20*Current)\/2 + (Current**2)\/10\n    return y\n\ncounter = 0\nwhile counter &lt; 5:\n    parameter_current = torch.randn(2, 2, requires_grad=True)\n\n    outputMyFirstFunction = myFirstFunction(parameter_current)\n    outputmySecondFunction = mySecondFunction(outputMyFirstFunction[1])\n    outputmySecondFunction.backward()\n\n    print(\"outputMyFirstFunction after backward:\",\n               outputMyFirstFunction)\n    print(\"outputmySecondFunction after backward:\",\n               outputmySecondFunction)\n    print(\"parameter_current Gradient after backward:\",\n               parameter_current.grad)\n\n    counter = counter + 1\n<\/code><\/pre>\n\n<p>The parameter_current.grad is none for all iterations when it obviously shouldn't be none. What am I doing wrong? And how can I fix it?<\/p>\n\n<p>Your help on this would be highly appreciated. Thanks a lot!<\/p>\n\n<p>Aly<\/p>\n","answers":[{"AnswerId":"53546401","CreationDate":"2018-11-29T19:41:13.797","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>Since it is not really clear to me what you actually want to archive, besides computing gradients for <code>parameter_current<\/code>,\nI just focus on describing why it doesn't work and what you can do to acutally compute gradients.<\/p>\n\n<p>I've added some comments in the code to make it more clear what the problem is. <\/p>\n\n<p>But in short the problem is that your <code>parameter_current<\/code> is not part of the computation of your <em>loss<\/em> resp. the tensor you call <code>backward()<\/code> on which is <code>outputmySecondFunction<\/code>.<\/p>\n\n<p>So currently you are only computing gradients for <code>i<\/code> as you have set <code>requires_grad=True<\/code> for it.<\/p>\n\n<p>Please check the comments, for detailes:<\/p>\n\n<pre><code>import torch\n\ndef myFirstFunction(parameter_current_here):\n    # I removed some stuff to reduce it to the core features\n    # removed torch.enable_grad(), since it is enabled by default\n    # removed Optimal=100000000000000 and Optimal=i, they are not used\n    optimalValue=100000000000000\n    for j in range(2,10):\n        # Are you sure you want to compute gradients this tensor i? \n        # Because this is actually what requires_grad=True does.\n        # Just as a side note, this isn't your problem, but affects performance of the model.\n        i= torch.ones(1,requires_grad=True)*j\n        optimalValueNow=i*parameter_current_here.sum()\n        if (optimalValueNow&lt;optimalValue):\n            optimalValue=optimalValueNow\n\n    # Part Problem 1:\n    # optimalValueNow is multiplied with your parameter_current\n    # i is just your parameter i, nothing else\n    # lets jump now the output below in the loop: outputMyFirstFunction\n    return optimalValueNow,i\n\ndef mySecondFunction(Current):\n    y=(20*Current)\/2 + (Current**2)\/10\n    return y\n\ncounter=0\nwhile counter&lt;5:\n    parameter_current = torch.randn(2, 2,requires_grad=True)\n\n    # Part Problem 2:\n    # this is a tuple (optimalValueNow,i) like described above\n    outputMyFirstFunction=myFirstFunction(parameter_current)\n    # now you are taking i as an input\n    # and i is just torch.ones(1,requires_grad=True)*j\n    # it as no connection to parameter_current\n    # thus nothing is optimized\n    outputmySecondFunction=mySecondFunction(outputMyFirstFunction[1])\n\n    # calculating gradients, since parameter_current is not part of the computation \n    # no gradients will be computed, you only get gradients for i\n    # Btw. if you would not have set requires_grad=True for i, you actually would get an error message\n    # for calling backward on this\n    outputmySecondFunction.backward()\n\n    print(\"outputMyFirstFunction after backward:\",outputMyFirstFunction)\n    print(\"outputmySecondFunction after backward:\",outputmySecondFunction)\n    print(\"parameter_current Gradient after backward:\",parameter_current.grad)\n\n    counter=counter+1\n<\/code><\/pre>\n\n<p>So if you want to compute gradients for <code>parameter_current<\/code> you simply have to make sure it is part of the computation \nof the tensor you call <code>backward()<\/code> on, you can do so for example by changing:<\/p>\n\n<pre><code>outputmySecondFunction=mySecondFunction(outputMyFirstFunction[1])\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>outputmySecondFunction=mySecondFunction(outputMyFirstFunction[0])\n<\/code><\/pre>\n\n<p>Will have this effect, as soon as you change it you will get gradients for <code>parameter_current<\/code>!<\/p>\n\n<p>I hope it helps!<\/p>\n\n<hr>\n\n<p><br><\/p>\n\n<p><strong>Full working code:<\/strong><\/p>\n\n<pre><code>import torch\n\ndef myFirstFunction(parameter_current_here):\n    optimalValue=100000000000000\n    for j in range(2,10):\n        i= torch.ones(1,requires_grad=True)*j\n        optimalValueNow=i*parameter_current_here.sum()\n        if (optimalValueNow&lt;optimalValue):\n            optimalValue=optimalValueNow\n\n    return optimalValueNow,i\n\ndef mySecondFunction(Current):\n    y=(20*Current)\/2 + (Current**2)\/10\n    return y\n\ncounter=0\nwhile counter&lt;5:\n    parameter_current = torch.randn(2, 2,requires_grad=True)\n    outputMyFirstFunction=myFirstFunction(parameter_current)\n    outputmySecondFunction=mySecondFunction(outputMyFirstFunction[0]) # changed line\n    outputmySecondFunction.backward()\n\n    print(\"outputMyFirstFunction after backward:\",outputMyFirstFunction)\n    print(\"outputmySecondFunction after backward:\",outputmySecondFunction)\n    print(\"parameter_current Gradient after backward:\",parameter_current.grad)\n\n    counter=counter+1\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>outputMyFirstFunction after backward: (tensor([ 1.0394]), tensor([ 9.]))\noutputmySecondFunction after backward: tensor([ 10.5021])\nparameter_current Gradient after backward: tensor([[ 91.8709,  91.8709],\n        [ 91.8709,  91.8709]])\noutputMyFirstFunction after backward: (tensor([ 13.1481]), tensor([ 9.]))\noutputmySecondFunction after backward: tensor([ 148.7688])\nparameter_current Gradient after backward: tensor([[ 113.6667,  113.6667],\n        [ 113.6667,  113.6667]])\noutputMyFirstFunction after backward: (tensor([ 5.7205]), tensor([ 9.]))\noutputmySecondFunction after backward: tensor([ 60.4772])\nparameter_current Gradient after backward: tensor([[ 100.2969,  100.2969],\n        [ 100.2969,  100.2969]])\noutputMyFirstFunction after backward: (tensor([-13.9846]), tensor([ 9.]))\noutputmySecondFunction after backward: tensor([-120.2888])\nparameter_current Gradient after backward: tensor([[ 64.8278,  64.8278],\n        [ 64.8278,  64.8278]])\noutputMyFirstFunction after backward: (tensor([-10.5533]), tensor([ 9.]))\noutputmySecondFunction after backward: tensor([-94.3959])\nparameter_current Gradient after backward: tensor([[ 71.0040,  71.0040],\n        [ 71.0040,  71.0040]])\n<\/code><\/pre>\n"},{"AnswerId":"53546294","CreationDate":"2018-11-29T19:34:05.530","ParentId":null,"OwnerUserId":"7078425","Title":null,"Body":"<p>I'm guessing the problem is the <code>with torch.enable_grad():<\/code> statements. After to exited the <code>with<\/code> statement, the <code>torch.enable_grad()<\/code> no longer applies and torch will clear the grads after the functions are run.<\/p>\n"}]}
+{"QuestionId":53508247,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2018-11-27T21:15:37.343","AcceptedAnswerId":53508444.0,"OwnerUserId":2478819.0,"Title":"Error when checking target: expected activation_2 to have shape (512,) but got array with shape (1,)","Body":"<p>Having trouble with the error in the title. I do actually know why I'm receiving this error but after researching I struggle to understand the concept as I'm new to ML. The error is relating to not having X_TRAIN and Y_TRAIN but I'm unable to understand what they consist of. All images have been resized to the same dimensions. I'm using the CUB 2011 dataset with a split of 90:10 training\/validation, here is my code below that relates to the error. I'm using Keras.<\/p>\n\n<pre><code>classifier.add(Conv2D(256, (3, 3), input_shape = (467,386,3)))\nclassifier.add(Activation('relu'))\n\n#Smaller for precise pooling\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n\nclassifier.add(Flatten())\n\n#Add 512 unit layer\nclassifier.add(Dense(units = 512, activation = 'relu'))\n<\/code><\/pre>\n\n \n","answers":[{"AnswerId":"53508444","CreationDate":"2018-11-27T21:29:34.423","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>The problem is in the discrepancy between your final <code>Dense(512, ...)<\/code> layer and the fact that your data targets are just 1 dimension. You can either fix by converting the model to output the expected dimensions, or adjust the data to the expected model outputs. That depends on what you are trying to achieve. To fix the model you can add another <code>Dense<\/code> layer:<\/p>\n\n<pre><code>classifier.add(Dense(units = 512, activation = 'relu'))\nclassifier.add(Dropout(0.2)) #prevent overfitting\nclassifier.add(Dense(units=1, activation='sigmoid')) # HERE ---\n<\/code><\/pre>\n\n<p>Or have a look at the what generator is outputting by <code>print(next(train_generator))<\/code> to see why the target array is 1 dimensional.<\/p>\n"}]}
+{"QuestionId":53508686,"AnswerCount":1,"Tags":"<tensorflow><rnn><tensorflow-datasets>","CreationDate":"2018-11-27T21:47:23.050","AcceptedAnswerId":53510578.0,"OwnerUserId":1079311.0,"Title":"Tensorflow 1.12 apply dataset.filter to dataset.window","Body":"<p>I have a sequential dataset from which I create windows in order to train an RNN. There are cases, where I want to throw away certain windows. However, when I use <code>dataset.window<\/code> followed by <code>dataset.filter<\/code> something breaks inside the pipeline. Here is a toy example, could someone please advise me on how to do this properly. This toy example demonstrates my issues. The below code creates windows of size 4 and then creates batches of 4 windows per batch. I want to throw away a window if the last element in the window is odd, so my batches will always be of size 4 but the windows inside the batch always have to end with an event element.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\nds = tf.data.Dataset.range(100)\nds = ds.window(size=4, shift=1,\n               stride=1,\n               drop_remainder=True).flat_map(lambda x: x.batch(4))\n\n#*I want to keep the window if the last element in the window is even*\nds = ds.filter(lambda x: x[3] % 2 == 0) \n\nds = ds.repeat()\nds = ds.batch(4, drop_remainder=True)\nit = ds.make_one_shot_iterator()\ndata = it.get_next()\nfor i in range(100):\n    print(sess.run([data]))\nsess.close()\n<\/code><\/pre>\n\n<p>And this throws the following error:<\/p>\n\n<pre><code>OutOfRangeError: End of sequence\n     [[{{node IteratorGetNext_6}} = IteratorGetNext[output_shapes=[[4,?]], output_types=[DT_INT64], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](OneShotIterator_6)]]\n\nDuring handling of the above exception, another exception occurred:\n\nOutOfRangeError                           Traceback (most recent call last)\n&lt;ipython-input-54-d6d959b5be78&gt; in &lt;module&gt;\n      1 for i in range(100):\n----&gt; 2     print(sess.run([data]))\n<\/code><\/pre>\n","answers":[{"AnswerId":"53510578","CreationDate":"2018-11-28T01:05:13.550","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>If you look at the return type of predicate in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#filter\" rel=\"nofollow noreferrer\">filter<\/a> method, you need to return a scalar tf bool tensor, which I suspect is not happening with your pythonic predicate. Modifying the code to return such a tensor, gives me results. <\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\nds = tf.data.Dataset.range(100)\nds = ds.window(size=4, shift=1,\n               stride=1,\n               drop_remainder=True).flat_map(lambda x: x.batch(4))\n\n#*I want to keep the window if the last element in the window is even*\nds = ds.filter(lambda x: tf.equal(x[3] % 2, 0)) \n\nds = ds.repeat()\nds = ds.batch(4, drop_remainder=True)\nit = ds.make_one_shot_iterator()\ndata = it.get_next()\nfor i in range(100):\n    print(sess.run([data]))\nsess.close()\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>[array([[ 1,  2,  3,  4],\n       [ 3,  4,  5,  6],\n       [ 5,  6,  7,  8],\n       [ 7,  8,  9, 10]])]\n[array([[ 9, 10, 11, 12],\n       [11, 12, 13, 14],\n       [13, 14, 15, 16],\n       [15, 16, 17, 18]])]\n[array([[17, 18, 19, 20],\n       [19, 20, 21, 22],\n       [21, 22, 23, 24],\n       [23, 24, 25, 26]])]\n[array([[25, 26, 27, 28],\n       [27, 28, 29, 30],\n       [29, 30, 31, 32],\n       [31, 32, 33, 34]])]\n[array([[33, 34, 35, 36],\n       [35, 36, 37, 38],\n       [37, 38, 39, 40],\n       [39, 40, 41, 42]])]\n<\/code><\/pre>\n\n<p>and so on<\/p>\n"}]}
+{"QuestionId":53509213,"AnswerCount":2,"Tags":"<python><tensorflow><import><install>","CreationDate":"2018-11-27T22:28:38.527","AcceptedAnswerId":53509445.0,"OwnerUserId":3677168.0,"Title":"Error importing Tensorflow when import tensorflow as tf","Body":"<p>I've recently gone through the installation of tensorflow (and struggled a little) and when I believed I had got it, I now get these import errors when running a file that ONLY contains <code>import tensorflow as tf<\/code>, I ran the cmd: <code>python3 tftest.py<\/code> and get these import errors:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow = swig_import_helper()\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow', fp, pathname, description)\n  File \"\/usr\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so: invalid ELF header\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"stockprice.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/__init__.py\", line 60, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow = swig_import_helper()\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow', fp, pathname, description)\n  File \"\/usr\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so: invalid ELF header\n\n\nError importing tensorflow.  Unless you are using bazel,\nyou should not try to import tensorflow from its source directory;\nplease exit the tensorflow source tree, and relaunch your python interpreter\nfrom there.\n<\/code><\/pre>\n\n<p>I do have tensorflow installed<\/p>\n\n<pre><code>Name: tensorflow\nVersion: 0.12.0\nLocation: \/usr\/local\/lib\/python3.4\/dist-packages\nRequires: numpy, six, protobuf, wheel\n<\/code><\/pre>\n\n<p>I've been trying to figure this out for hours, has anyone encountered this or a similar error? I've looked at installation manuals, and followed every single step I could. I've tried reintalling protobuf<\/p>\n\n<p>I am not running my test file on the tensorflow directory.<\/p>\n\n<p>I would really REALLY appreciate any help since this error is draining my brain out.<\/p>\n","answers":[{"AnswerId":"53509445","CreationDate":"2018-11-27T22:49:23.180","ParentId":null,"OwnerUserId":"9873654","Title":null,"Body":"<p>Try to delete the library file manually and reinstall it using pip<\/p>\n"},{"AnswerId":"55245982","CreationDate":"2019-03-19T16:40:14.793","ParentId":null,"OwnerUserId":"6157415","Title":null,"Body":"<p>check through <code>pip list<\/code> not to have installed the <code>tensorflow-gpu<\/code> library because some GPUs are not supported.<\/p>\n\n<p>If this is the case, <strong><em>uninstall tensor flow-gpu<\/em><\/strong> and <strong><em>tensorflow-estimator<\/em><\/strong> and <strong><em>re-install tensorflow<\/em><\/strong> :<br><br>\n<code>pip uninstall tensorflow-gpu<\/code><br> <br>\n<code>pip uninstall tensorflow-estimator<\/code><br> <br>\n<code>pip install tensorflow<\/code><br> <br>\nmake sure you use python 3.6 with the <code>pip -V<\/code> command<\/p>\n"}]}
+{"QuestionId":53509654,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-27T23:12:11.057","AcceptedAnswerId":null,"OwnerUserId":8234286.0,"Title":"Will TensorFlow 2.0 support common Machine Learning libraries stated below?","Body":"<p>With the exciting new features of TensorFlow 2.0, may I know if TensorFlow 2.0 will still support the common machine learning techniques as stated below? <\/p>\n\n<ul>\n<li><strong>Linear Regression<\/strong><\/li>\n<li><strong>Logistic Regression<\/strong><\/li>\n<li><strong>K-Means Clustering<\/strong><\/li>\n<li><strong>K-Nearest Neighbors<\/strong> <\/li>\n<li><strong>Random Forest<\/strong> <\/li>\n<li><strong>Naive Bayes<\/strong> <\/li>\n<li><strong>Support Vector Machine<\/strong><\/li>\n<\/ul>\n\n<p>Thank you very much for your time and clarification in advance. <\/p>\n","answers":[{"AnswerId":"54637990","CreationDate":"2019-02-11T19:47:47.197","ParentId":null,"OwnerUserId":"8397234","Title":null,"Body":"<p>The models that you mention above will still definitely be supported in TensorFlow 2.0 as part of <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/estimator\" rel=\"nofollow noreferrer\"><code>tf.estimator<\/code><\/a>. If you are using a symbol in TensorFlow 1.x that is currently housed in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\" rel=\"nofollow noreferrer\"><code>tf.contrib<\/code><\/a>, it will either be deprecated, transitioned to the <a href=\"https:\/\/github.com\/tensorflow\/addons\" rel=\"nofollow noreferrer\"><code>tensorflow\/addons<\/code><\/a> repo, or moved into the core TF 2.0 API.<\/p>\n\n<p>You can check out the RFCs below for more detailed changes:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180827-api-names.md\" rel=\"nofollow noreferrer\">TensorFlow Namespaces<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\" rel=\"nofollow noreferrer\">Sunsetting <code>tf.contrib<\/code><\/a><\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20181214-move-to-addons.md\" rel=\"nofollow noreferrer\">Move from <code>tf.contrib<\/code> to <code>tensorflow\/addons<\/code><\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53509791,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><object-detection><google-cloud-ml>","CreationDate":"2018-11-27T23:26:22.713","AcceptedAnswerId":53545526.0,"OwnerUserId":2383106.0,"Title":"Faster RCNN Model training stops running on GCP, runs locally without issue","Body":"<p>Attempting to run program based on the Tensorflow Object Detection API. Faster RCNN Model stops training on GCP but runs locally without issue. Any feedback would be appreciated. Have tried Logs Writer role permission for Service Agent as suggested in different posts. Have not been able to find any more feedback.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KArBk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KArBk.png\" alt=\"GCP Log\"><\/a><\/p>\n\n<p>Full Error Message:<\/p>\n\n<blockquote>\n  <p>The replica master 0 exited with a non-zero status of 1. Termination\n  reason: Error. Traceback (most recent call last): File\n  \"\/usr\/lib\/python2.7\/runpy.py\", line 174, in _run_module_as_main\n  \"<strong>main<\/strong>\", fname, loader, pkg_name) File\n  \"\/usr\/lib\/python2.7\/runpy.py\", line 72, in _run_code exec code in\n  run_globals File\n  \"\/root\/.local\/lib\/python2.7\/site-packages\/object_detection\/train.py\",\n  line 198, in  tf.app.run() File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/platform\/app.py\",\n  line 48, in run _sys.exit(main(_sys.argv[:1] + flags_passthrough))\n  File\n  \"\/root\/.local\/lib\/python2.7\/site-packages\/object_detection\/train.py\",\n  line 194, in main worker_job_name, is_chief, FLAGS.train_dir) File\n  \"\/root\/.local\/lib\/python2.7\/site-packages\/object_detection\/trainer.py\",\n  line 296, in train saver=saver) File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/slim\/python\/slim\/learning.py\",\n  line 763, in train sess, train_op, global_step, train_step_kwargs)\n  File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/slim\/python\/slim\/learning.py\",\n  line 487, in train_step run_metadata=run_metadata) File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 889, in run run_metadata_ptr) File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 1120, in _run feed_dict_tensor, options, run_metadata) File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 1317, in _do_run options, run_metadata) File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 1336, in _do_call raise type(e)(node_def, op, message)\n  UnavailableError: Endpoint read failed To find out more about why your\n  job exited please check the logs:\n  <a href=\"https:\/\/console.cloud.google.com\/logs\/viewer?project=1086278442266&amp;resource=ml_job%2Fjob_id%2Fuav_object_detection_1543356760&amp;advancedFilter=resource.type%3D%22ml_job%22%0Aresource.labels.job_id%3D%22uav_object_detection_1543356760%22\" rel=\"nofollow noreferrer\">https:\/\/console.cloud.google.com\/logs\/viewer?project=1086278442266&amp;resource=ml_job%2Fjob_id%2Fuav_object_detection_1543356760&amp;advancedFilter=resource.type%3D%22ml_job%22%0Aresource.labels.job_id%3D%22uav_object_detection_1543356760%22<\/a><\/p>\n<\/blockquote>\n\n<p>This is what I am running in Terminal to start training:<\/p>\n\n<pre><code>gcloud ml-engine jobs submit training `whoami`_object_detection_`date +%s` \\\n   --job-dir=gs:\/\/my_gcs_bucket\/train \\\n   --packages dist\/object_detection-0.1.tar.gz,slim\/dist\/slim-0.1.tar.gz \\\n   --module-name object_detection.train \\\n   --region us-central1 \\\n   --config object_detection\/samples\/cloud\/cloud.yml \\\n   --runtime-version=1.4 \\\n   -- \\\n   --train_dir=gs:\/\/my_gcs_bucket\/train \\\n   --pipeline_config_path=gs:\/\/my_gcs_bucket\/data\/faster_rcnn_resnet101.config\n<\/code><\/pre>\n\n<p>This is my file structure in the GCP Bucket<\/p>\n\n<pre><code>+ data\/\n  - faster_rcnn_resnet101.config\n  - model.ckpt.index\n  - model.ckpt.meta\n  - model.ckpt.data-00000-of-00001\n  - pet_label_map.pbtxt\n  - train.record\n  - val.record\n+ train\/\n<\/code><\/pre>\n\n<p>This is my file structure in the folder I am running from<\/p>\n\n<pre><code>+dist\/\n  -object_detection-0.1.tar.gz\n+object_detection\/\n+object_detection.egg-info\/\n+slim\/\nsetup.py\n<\/code><\/pre>\n\n<p>Config File:<\/p>\n\n<pre><code># Faster R-CNN with Resnet-101 (v1) configured for the Oxford-IIIT Pet Dataset.\n# Users should configure the fine_tune_checkpoint field in the train config as\n# well as the label_map_path and input_path fields in the train_input_reader and\n# eval_input_reader. Search for \"PATH_TO_BE_CONFIGURED\" to find the fields that\n# should be configured.\n\nmodel {\n  faster_rcnn {\n    num_classes: 1\n    image_resizer {\n      keep_aspect_ratio_resizer {\n        min_dimension: 600\n        max_dimension: 1024\n      }\n    }\n    feature_extractor {\n      type: 'faster_rcnn_resnet101'\n      first_stage_features_stride: 16\n    }\n    first_stage_anchor_generator {\n      grid_anchor_generator {\n        scales: [0.25, 0.5, 1.0, 2.0]\n        aspect_ratios: [0.5, 1.0, 2.0]\n        height_stride: 16\n        width_stride: 16\n      }\n    }\n    first_stage_box_predictor_conv_hyperparams {\n      op: CONV\n      regularizer {\n        l2_regularizer {\n          weight: 0.0\n        }\n      }\n      initializer {\n        truncated_normal_initializer {\n          stddev: 0.01\n        }\n      }\n    }\n    first_stage_nms_score_threshold: 0.0\n    first_stage_nms_iou_threshold: 0.7\n    first_stage_max_proposals: 300\n    first_stage_localization_loss_weight: 2.0\n    first_stage_objectness_loss_weight: 1.0\n    initial_crop_size: 14\n    maxpool_kernel_size: 2\n    maxpool_stride: 2\n    second_stage_box_predictor {\n      mask_rcnn_box_predictor {\n        use_dropout: false\n        dropout_keep_probability: 1.0\n        fc_hyperparams {\n          op: FC\n          regularizer {\n            l2_regularizer {\n              weight: 0.0\n            }\n          }\n          initializer {\n            variance_scaling_initializer {\n              factor: 1.0\n              uniform: true\n              mode: FAN_AVG\n            }\n          }\n        }\n      }\n    }\n    second_stage_post_processing {\n      batch_non_max_suppression {\n        score_threshold: 0.0\n        iou_threshold: 0.6\n        max_detections_per_class: 100\n        max_total_detections: 300\n      }\n      score_converter: SOFTMAX\n    }\n    second_stage_localization_loss_weight: 2.0\n    second_stage_classification_loss_weight: 1.0\n  }\n}\n\ntrain_config: {\n  batch_size: 1\n  batch_queue_capacity: 1\n  num_batch_queue_threads: 1\n  prefetch_queue_capacity: 1\n  optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0003\n          schedule {\n            step: 0\n            learning_rate: .0003\n          }\n          schedule {\n            step: 900000\n            learning_rate: .00003\n          }\n          schedule {\n            step: 1200000\n            learning_rate: .000003\n          }\n        }\n      }\n      momentum_optimizer_value: 0.9\n    }\n    use_moving_average: false\n  }\n  gradient_clipping_by_norm: 10.0\n  fine_tune_checkpoint: \"gs:\/\/my_gcs_bucket\/data\/model.ckpt\"\n  from_detection_checkpoint: true\n  # Note: The below line limits the training process to 200K steps, which we\n  # empirically found to be sufficient enough to train the pets dataset. This\n  # effectively bypasses the learning rate schedule (the learning rate will\n  # never decay). Remove the below line to train indefinitely.\n  num_steps:2000\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n}\n\ntrain_input_reader: {\n  tf_record_input_reader {\n    input_path: \"gs:\/\/my_gcs_bucket\/data\/data\/train.record\"\n  }\n  label_map_path: \"gs:\/\/my_gcs_bucket\/data\/data\/label_map.pbtxt\"\n  queue_capacity: 10\n  min_after_dequeue: 5\n}\n\neval_config: {\n  num_examples: 4\n  # Note: The below line limits the evaluation process to 10 evaluations.\n  # Remove the below line to evaluate indefinitely.\n  max_evals: 10\n}\n\neval_input_reader: {\n  tf_record_input_reader {\n    input_path: \"gs:\/\/my_gcs_bucket\/data\/data\/val.record\"\n  }\n  label_map_path: \"gs:\/\/my_gcs_bucket\/data\/data\/label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"53545526","CreationDate":"2018-11-29T18:37:01.170","ParentId":null,"OwnerUserId":"2383106","Title":null,"Body":"<p>Changed runtime version to 1.2 in cloud.yml and initial request.<\/p>\n"}]}
+{"QuestionId":53509823,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-probability>","CreationDate":"2018-11-27T23:28:58.860","AcceptedAnswerId":53675410.0,"OwnerUserId":4258928.0,"Title":"Tensorflow Implement Multivariate Student T diagonal distribution","Body":"<p>I'm implementing a diagonal multivariate student t distribution (so logP(x1,x2,x3,..xD) = logP(x1) + logP(x2)+ ....+ logP(xD) ) such that it can be used as base distribution for bijectors in TensorFlow<\/p>\n\n<pre><code>import tensorflow_probability as tfp\ntfd = tfp.distributions\n\nD = 2 # number of dimension\ndf = 5. # degree of freedom\n\n# construct D univariate student t distribution\n\nbase_dist = tfd.StudentT(loc=tf.constant([0.] * D,dtype=DTYPE),\n                         scale = tf.constant([1.] * D,dtype=DTYPE),\n                         df = tf.constant([df],dtype=DTYPE))\n\nQ = tfd.TransformedDistribution(distribution=base_dist,bijector=Chain)\n# where Chain is a tfb.Chain() object that a sequence of bisector numbers\n<\/code><\/pre>\n\n<p>I change <code>tfd.StudentT.log_prob()<\/code> such that it sum over last axes. It takes shape <code>[batch_size,dim]<\/code> as input and return pdf with shape<code>[batch_size,]<\/code><\/p>\n\n<p>However, when I call <code>Q.log_prob(x)<\/code>; I got error <code>ValueError: event_ndims (0) must be larger than min_event_ndims (1)<\/code><\/p>\n\n<p>I'm not sure how to fix this error; can some one helps me ?<\/p>\n","answers":[{"AnswerId":"53675410","CreationDate":"2018-12-07T19:00:48.650","ParentId":null,"OwnerUserId":"6716334","Title":null,"Body":"<p>TensorFlow Probability provides a way to create vector-valued distributions from scalar distributions via the <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/distributions\/Independent\" rel=\"nofollow noreferrer\">tfd.Independent<\/a> meta-distribution. That would automatically perform the summation in <code>log_prob<\/code> that you want.<\/p>\n\n<p>If you really wanted to implement things yourself, the issue you're having sounds like you didn't override the <code>event_shape<\/code> and <code>event_shape_tensor<\/code> methods (along with the <code>batch_shape<\/code> and <code>batch_shape_tensor<\/code>).<\/p>\n\n<p>Lastly, typically when people talk about the multivariate student t distribution, they mean the elliptical distribution described <a href=\"https:\/\/en.wikipedia.org\/wiki\/Multivariate_t-distribution\" rel=\"nofollow noreferrer\">here<\/a>, which is <em>not<\/em> the same thing as taking a product of 1D Student-t's distributions and then linearly transforming them. Recently, TFP added an implementation of the elliptical variant of this distribution <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/distributions\/MultivariateStudentTLinearOperator\" rel=\"nofollow noreferrer\">here<\/a>. It takes an affine transformation as input that you can use to setup the location\/correlation structure of the distribution.<\/p>\n"}]}
+{"QuestionId":53509911,"AnswerCount":2,"Tags":"<python><keras><time-series><lstm><autoencoder>","CreationDate":"2018-11-27T23:38:34.863","AcceptedAnswerId":null,"OwnerUserId":8932902.0,"Title":"Keras LSTM Autoencoder time-series reconstruction","Body":"<p>I am trying to reconstruct time series data with LSTM Autoencoder (Keras).\nNow I want train autoencoder on small amount of samples (5 samples, every sample is 500 time-steps long and have 1 dimension). I want to make sure that model can reconstruct that 5 samples and after that I will use all data (6000 samples).<\/p>\n\n<pre><code>window_size = 500\nfeatures = 1\ndata = data.reshape(5, window_size, features)\n\nmodel = Sequential()\n\nmodel.add(LSTM(256, input_shape=(window_size, features), \nreturn_sequences=True))\nmodel.add(LSTM(128, input_shape=(window_size, features), \nreturn_sequences=False))\nmodel.add(RepeatVector(window_size))\n\nmodel.add(LSTM(128, input_shape=(window_size, features), \nreturn_sequences=True))\nmodel.add(LSTM(256, input_shape=(window_size, features), \nreturn_sequences=True))\nmodel.add(TimeDistributed(Dense(1)))\n\nmodel.compile(optimizer='adam', loss='mse')\nmodel.fit(data, data, epochs=100, verbose=1)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/myTSa.png\" rel=\"noreferrer\">Model<\/a><\/p>\n\n<p>Training:<\/p>\n\n<pre><code>Epoch 1\/100\n5\/5 [==============================] - 2s 384ms\/step - loss: 0.1603\n...\nEpoch 100\/100\n5\/5 [==============================] - 2s 388ms\/step - loss: 0.0018\n<\/code><\/pre>\n\n<p>After training, I tried reconstruct one of 5 samples:<\/p>\n\n<pre><code>yhat = model.predict(np.expand_dims(data[1,:,:], axis=0), verbose=0)\n<\/code><\/pre>\n\n<p>Reconstitution: Blue<br>\nInput: Orange   <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Tzsdg.png\" alt=\"Reconstion (blue) vs Input (orange)\"><\/p>\n\n<p>Why is reconstruction so bad when loss is small? How can I make model better? Thanks.<\/p>\n","answers":[{"AnswerId":"57179382","CreationDate":"2019-07-24T09:17:31.697","ParentId":null,"OwnerUserId":"6325830","Title":null,"Body":"<p>I tried running your code on the following data<\/p>\n\n<pre><code>data = np.random.uniform(-0.1, 0.1, size=(5, 500))\ndata = data.cumsum(axis=1)\n<\/code><\/pre>\n\n<p>so the data is just the cumalative sum of some random uniform noise. I ran for 1000 epochs, and my results are not as bad as yours, the LSTM seems to make some effort to follow the line, though it seems to just be hovering around the running mean (as one might expect).<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/kRQivaF.png\" alt=\"blah\"><\/p>\n\n<p>Note that this is running the model on the TRAINING data (which you seem to imply you were doing in your question) - if we try to look at performance on data that the model was not trained on, we can get bad results.<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/Lrmz0Nk.png\" alt=\"blah\"><\/p>\n\n<p>This is not surprising in the least, with such a small training set, we should fully expect the model to overfit, and not generalise to new data.<\/p>\n"},{"AnswerId":"57200195","CreationDate":"2019-07-25T10:47:31.540","ParentId":null,"OwnerUserId":"5210098","Title":null,"Body":"<p>It seems to me, a time series should be given to the LSTMs in this format:<\/p>\n\n<pre><code> (samples, features , window_size)\n<\/code><\/pre>\n\n<p>So, if you change the format, for example I exchanged the variables, and look at the results: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5JUDN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5JUDN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Code for reproducing the result(I didn't change the name of the variables, so please don't be confused :)):<\/p>\n\n<pre><code>import numpy as np\nimport keras\nfrom keras import Sequential\nfrom keras.layers import Dense, RepeatVector,        TimeDistributed\nfrom keras.layers import LSTM\n\nN = 10000\ndata = np.random.uniform(-0.1, 0.1, size=(N, 500))\ndata = data.cumsum(axis=1)\nprint(data.shape)\nwindow_size = 1\nfeatures = 500\ndata = data.reshape(N, window_size, features)\n\nmodel = Sequential()\n\nmodel.add(LSTM(32, input_shape=\n(window_size,features), \nreturn_sequences=True))\nmodel.add(LSTM(16, input_shape=(window_size,   \nfeatures), \nreturn_sequences=False))\nmodel.add(RepeatVector(window_size))\n\nmodel.add(LSTM(16, input_shape=(window_size, \nfeatures), \nreturn_sequences=True))\nmodel.add(LSTM(32, input_shape=(window_size,   \nfeatures), \nreturn_sequences=True))\nmodel.add(TimeDistributed(Dense(500)))\n\nmodel.compile(optimizer='adam', loss='mse')\nmodel.fit(data, data, epochs=100, verbose=1)\n\n\nyhat = model.predict(np.expand_dims(data[1,:,:],   axis=0), verbose=0)\nplot(np.arange(500), yhat[0,0,:])\nplot(np.arange(500), data[1,0,:])\n<\/code><\/pre>\n\n<p>Credit to sobe86: I used the proposed data by him\/her.<\/p>\n"}]}
+{"QuestionId":53510458,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T00:47:31.353","AcceptedAnswerId":null,"OwnerUserId":10253771.0,"Title":"tensorflow - same program on different computers allocate different GPU memory","Body":"<p>ubuntu 16.04, python 2.7.12, tensorflow 1.10.1 (gpu version), cuda 9.0, cudnn 7.2<\/p>\n\n<p>I have built and trained a CNN model, and now I am using a while loop to repeatedly let my model make predictions.<\/p>\n\n<p>In order to limit the memory usage, I am using the following code to create my classifier:<\/p>\n\n<pre><code>import tensorflow as tf\n\nsession_config = tf.ConfigProto(log_device_placement=False)\nsession_config.gpu_options.allow_growth = True \nrun_config = tf.estimator.RunConfig().replace(session_config=session_config)\n\n\nclassifier = tf.estimator.Estimator(\n    model_fn = my_model_fn, \n    model_dir = my_trained_model_dir,\n    config = run_config,\n    params={}\n<\/code><\/pre>\n\n<p>)<\/p>\n\n<p>And I call <code>classifier.predict(my_input_fn)<\/code> in a <code>while<\/code> loop to repeatedly make predictions.<\/p>\n\n<p><strong>Issue:<\/strong><\/p>\n\n<p>I am running my codes on two computers, both with the same software environment as I listed above.<\/p>\n\n<p>However, the two computers have different GPUs:<\/p>\n\n<p>Computer A: 1050 2G<\/p>\n\n<p>Computer B: 1070 8G<\/p>\n\n<p>My code works well on both computer.<\/p>\n\n<p><strong>However, when I use <code>nvidia-smi<\/code> to check the GPU memory allocation, I found that my code will allocate 1.4G GPU memory on Computer A, while it becomes 3.6G on Computer B.<\/strong><\/p>\n\n<p><strong>So, Why would this happen?<\/strong><\/p>\n\n<p><strong>I think <code>session_config.gpu_options.allow_growth = True<\/code> tells the program to allocate as much as it needs. Computer A has proved that 1.4G is enough, then why would the same code allocate 3.6G on Computer B?<\/strong> <\/p>\n","answers":[{"AnswerId":"53510814","CreationDate":"2018-11-28T01:41:15.047","ParentId":null,"OwnerUserId":"686184","Title":null,"Body":"<p>It may be that 1.4gb is actually not enough, and some of the required memory is swapped into main memory. Video cards drivers do that.<\/p>\n"}]}
+{"QuestionId":53510628,"AnswerCount":1,"Tags":"<python><tensorflow><keras><generative-adversarial-network>","CreationDate":"2018-11-28T01:12:42.517","AcceptedAnswerId":null,"OwnerUserId":4360167.0,"Title":"Delete some neurons at output layer (Keras)","Body":"<p>I'm using Keras to make a GAN for synthetic data generating. So, i want use a pre-treined classifier model to be the Discriminator (Because I would not have to train the discriminator, only the generator). <\/p>\n\n<p>The pre-treined model was made in keras and has originaly 4 outputs at last layer <code>Dense(4, activation='softmax')<\/code> (A, B, C, D). I'want generate synthetic data from class C. The Discriminator model must have only 1 output (fake or true),  so I need change the last layer from pre-treinade model to be only the C output. So, how i can do it in keras or maybe using Tensorflow at backend?<\/p>\n\n<p>I'm looking something like this...<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/U3o83.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/U3o83.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>model = load_model('pre_treined_model.h5')\ndiscriminator = model\ndiscriminator.layers[-1] = \"only C weight\/output\"\n<\/code><\/pre>\n","answers":[{"AnswerId":"53521511","CreationDate":"2018-11-28T14:18:06.510","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>You're going to have to do something like this:<\/p>\n\n<pre><code>model = load_model('pre_trained_model.h5')\nmodel.layers.pop()\n\nx = Dense(1, activation='sigmoid')(model.layers[-1].output)\nmodel = Model(inputs=model.input, outputs=[x])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53510644,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><cudnn>","CreationDate":"2018-11-28T01:14:52.533","AcceptedAnswerId":null,"OwnerUserId":6890926.0,"Title":"CuDNN version does not match","Body":"<p>I am very close to configuring a gpu enabled environment using the keras\/tensorflow python library.  When I try to train my model I get a long error message:<\/p>\n\n<pre><code>2018-11-27 18:34:47.776387: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0\n2018-11-27 18:34:48.769258: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-27 18:34:48.769471: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 \n2018-11-27 18:34:48.769595: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N \n2018-11-27 18:34:48.769825: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 3024 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1050, pci bus id: 0000:01:00.0, compute capability: 6.1)\n2018-11-27 18:34:50.405201: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:363] Loaded runtime CuDNN library: 7.1.4 but source was compiled with: 7.2.1.  CuDNN library major and minor version needs to match or have higher minor version in case of CuDNN 7.0 or later version. If using a binary install, upgrade your CuDNN library.  If building from sources, make sure the library loaded at runtime is compatible with the version specified during compile configuration.\n<\/code><\/pre>\n\n<p>I've looked at a couple similar stack overflow posts and it appears that I need to either adjust the CuDNN version or the tensorflow-gpu version.  I downloaded the correct version of CuDNN from Nvidia's website but it did not appear to do anything.  I have also found several posts about changing my tensorflow-gpu version, but WHICH version should I download and HOW.  I am using WIndows 10.<\/p>\n","answers":[{"AnswerId":"54810982","CreationDate":"2019-02-21T15:45:37.010","ParentId":null,"OwnerUserId":"2005576","Title":null,"Body":"<p>Hi this powershell Script should work to update your driver.<\/p>\n\n<pre><code>$ur='https:\/\/dsvmteststore.blob.core.windows.net\/patches\/cuda\/cudnnpatch.zip?st=2019-02-20T04%3A10%3A00Z&amp;se=2019-03-01T04%3A10%3A00Z&amp;sp=r&amp;sv=2017-07-29&amp;sr=c&amp;sig=w1VqK70ZcWWbbRW2K4Y8q5298dNxBqsoP71%2F4nF6uYM%3D'\nInvoke-WebRequest -Uri  $ur  -OutFile '.\\cudnnpatch.zip' -UseBasicParsing\n\n$from='.\\cudnnpatch.zip'\n$to='.\\'\ncmd \/c \"c:\\7-Zip\\7z.exe x $from -o$to -y\"\n\n\n$root='C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v9.0'\n\n\n$dll='\\bin\\cudnn64_7.dll' \n$header='\\include\\cudnn.h' \n$lib='\\lib\\x64\\cudnn.lib' \n\n$from='.\\cuda'\nCopy-Item \"$from$dll\" \"$root$dll\"  -Force\nCopy-Item \"$from$header\" \"$root$header\"  -Force\nCopy-Item \"$from$lib\" \"$root$lib\"  -Force\n<\/code><\/pre>\n"}]}
+{"QuestionId":53511949,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-28T04:04:10.453","AcceptedAnswerId":null,"OwnerUserId":3994228.0,"Title":"Tensorflow Android - Ops can be multi-threaded but with multi-threaded application?","Body":"<p>As per the <a href=\"https:\/\/www.tensorflow.org\/lite\/tfmobile\/optimizing#threading\" rel=\"nofollow noreferrer\">documentation<\/a>, ops can be multi-threaded on mobile (default <code>inter_op_parallelism_threads<\/code> is set to <code>1<\/code>) based on the number of cores on that particular device. I am building an Android application that has three concurrent (threads) running three different models, would it be possible to still enable <code>inter_op_parallelism_threads<\/code> and set to <code>4<\/code> for a quad-core device? <\/p>\n","answers":[{"AnswerId":"53512308","CreationDate":"2018-11-28T04:45:24.837","ParentId":null,"OwnerUserId":"6866309","Title":null,"Body":"<p>When it comes to mobile,  <code>inter_op_parallelism_threads<\/code> is not big of use, but based on the documentation, there is a significant benefit from number of threads with in the operation (<strong>\"intra\"<\/strong> - not inter). And the variable to be used for it is <code>intra_op_parallelism_threads<\/code>.<\/p>\n\n<blockquote>\n  <p>On mobile, how many threads an op will use is set to the number of\n  cores by default, or 2 when the number of cores can't be determined.\n  You can override the default number of threads that ops are using by\n  setting intra_op_parallelism_threads in the session options<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":53512062,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2018-11-28T04:18:04.873","AcceptedAnswerId":null,"OwnerUserId":10714928.0,"Title":"why are my tensorflow events files empty?","Body":"<p>I am running the tensorflow object detection API and using the SSD_mobilenet model.I have the model.cpkt as well as the graph.pbtxt in my training dir. But in my training dir I found that my events files are empty. It seems that no data was written to my events. Could anyone help me,please!!!<\/p>\n","answers":[{"AnswerId":"53520791","CreationDate":"2018-11-28T13:40:22.333","ParentId":null,"OwnerUserId":"10330116","Title":null,"Body":"<p>Tensorflow event files will be generated based on the summaries what we have added in code.<\/p>\n\n<p>For example, suppose you are training a convolutional neural network for recognizing MNIST digits. You'd like to record how the learning rate varies over time, and how the objective function is changing. Collect these by attaching tf.summary.scalar ops to the nodes that output the learning rate and loss respectively. Then, give each scalar_summary a meaningful tag, like 'learning rate' or 'loss function'.<\/p>\n\n<p>For example:<\/p>\n\n<h1>Add a scalar summary for the snapshot loss.<\/h1>\n\n<p>tf.summary.scalar('loss', loss)<\/p>\n\n<p>Please refer the below link:\n<a href=\"https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard<\/a><\/p>\n"}]}
+{"QuestionId":53512281,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><pytorch><tensor>","CreationDate":"2018-11-28T04:43:35.487","AcceptedAnswerId":null,"OwnerUserId":10555987.0,"Title":"Concatenate Two Tensors in Pytorch","Body":"<pre><code>RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 2. Got 32 and 71 in dimension 0 at \/pytorch\/aten\/src\/THC\/generic\/THCTensorMath.cu:87\n<\/code><\/pre>\n\n<p>I have a tensor of shape <code>[71 32 1]<\/code>.<\/p>\n\n<p>I want to make it of shape <code>[100 32 1]<\/code> by padding zero vectors.<\/p>\n\n<p>I tried by concatenating a padding vector of zeros of shape [29 32 1]. I get the error above.<\/p>\n\n<p>I try with a padding vector of zeros of shape <code>[29 32 1]<\/code>, I still get an error.<\/p>\n\n<p>How could I create the required tensor?<\/p>\n","answers":[{"AnswerId":"53513331","CreationDate":"2018-11-28T06:24:09.097","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In order to help you better, you need to post the <a href=\"https:\/\/stackoverflow.com\/help\/mcve\">code<\/a> that caused the error, without it we are just guessing here...<\/p>\n\n<p>Guessing from the error message you got:<\/p>\n\n<h3>1.<\/h3>\n\n<blockquote>\n<pre><code>Sizes of tensors must match except in dimension 2\n<\/code><\/pre>\n<\/blockquote>\n\n<p>pytorch tries to concat along the 2nd dimension, whereas you try to concat along the first.<\/p>\n\n<h3>2.<\/h3>\n\n<blockquote>\n<pre><code>Got 32 and 71 in dimension 0\n<\/code><\/pre>\n<\/blockquote>\n\n<p>It seems like the dimensions of the tensor you want to concat are not as you expect, you have one with size <code>(72, ...)<\/code> while the other is <code>(32, ...)<\/code>.<br>\nYou need to check this as well.<\/p>\n\n<h3>Working code<\/h3>\n\n<p>Here's an example of concat<\/p>\n\n<pre><code>import torch\n\nx = torch.rand((71, 32, 1))\n# x.shape = torch.Size([71, 32, 1])\npx = torch.cat((torch.zeros(29, 32, 1, dtype=x.dtype, device=x.device), x), dim=0)\n# px.shape = torch.Size([100, 32, 1])\n<\/code><\/pre>\n\n<p>Alternatively, you can use <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.functional.pad\" rel=\"noreferrer\"><code>functional.pad<\/code><\/a>:<\/p>\n\n<pre><code>from torch.nn import functional as F\n\npx = F.pad(x, (0, 0, 0, 0, 29, 0))\n<\/code><\/pre>\n"}]}
+{"QuestionId":53512395,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T04:56:16.663","AcceptedAnswerId":null,"OwnerUserId":10714993.0,"Title":"tensorflow - relating the accuracy back to the graph","Body":"<p>This is from a tutorial, but there's a large piece of it that I can't see the logic in - and I'd rather not just accept it.<\/p>\n\n<p>It's a tensorflow example to train and test against mnist images of digits.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot = True)\n\nn_nodes_hl1 = 500\nn_nodes_hl2 = 500\nn_nodes_hl3 = 500\n\nn_classes = 10\nbatch_size = 100\n\nx = tf.placeholder('float', [None, 784])\ny = tf.placeholder('float')\n\ndef neural_network_model(data):\n    hidden_1_layer = {'weights':tf.Variable(tf.random_normal([784, n_nodes_hl1])),\n                      'biases':tf.Variable(tf.random_normal([n_nodes_hl1]))}\n\n    hidden_2_layer = {'weights':tf.Variable(tf.random_normal([n_nodes_hl1, n_nodes_hl2])),\n                      'biases':tf.Variable(tf.random_normal([n_nodes_hl2]))}\n\n    hidden_3_layer = {'weights':tf.Variable(tf.random_normal([n_nodes_hl2, n_nodes_hl3])),\n                      'biases':tf.Variable(tf.random_normal([n_nodes_hl3]))}\n\n    output_layer = {'weights':tf.Variable(tf.random_normal([n_nodes_hl3, n_classes])),\n                    'biases':tf.Variable(tf.random_normal([n_classes])),}\n\n\n    l1 = tf.add(tf.matmul(data,hidden_1_layer['weights']), hidden_1_layer['biases'])\n    l1 = tf.nn.relu(l1)\n\n    l2 = tf.add(tf.matmul(l1,hidden_2_layer['weights']), hidden_2_layer['biases'])\n    l2 = tf.nn.relu(l2)\n\n    l3 = tf.add(tf.matmul(l2,hidden_3_layer['weights']), hidden_3_layer['biases'])\n    l3 = tf.nn.relu(l3)\n\n    output = tf.matmul(l3,output_layer['weights']) + output_layer['biases']\n\n    return output\n\ndef train_neural_network(x):\n    prediction = neural_network_model(x)\n    cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y) )\n    optimizer = tf.train.AdamOptimizer().minimize(cost)\n\n    hm_epochs = 10\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        for epoch in range(hm_epochs):\n            epoch_loss = 0\n            for _ in range(int(mnist.train.num_examples\/batch_size)):\n                epoch_x, epoch_y = mnist.train.next_batch(batch_size)\n                _, c = sess.run([optimizer, cost], feed_dict={x: epoch_x, y: epoch_y})\n                epoch_loss += c\n\n            print('Epoch', epoch, 'completed out of',hm_epochs,'loss:',epoch_loss)\n\n        correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\n        accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n        print('Accuracy:',accuracy.eval({x:mnist.test.images, y:mnist.test.labels}))\n\ntrain_neural_network(x)\n<\/code><\/pre>\n\n<p>My question is specifically about these three lines:<\/p>\n\n<pre><code>        correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\n        accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n        print('Accuracy:',accuracy.eval({x:mnist.test.images, y:mnist.test.labels}))\n<\/code><\/pre>\n\n<p>How does 'y' get its data - unless it merely holds the values for the last batch of the last epoch?<\/p>\n\n<p>How do these 3 lines tie together and connect back to the graph created? <\/p>\n\n<p>If the variable 'correct' is already an assessment of the output prediction against 'y', why is the apparent result of that comparison used to test against test data?<\/p>\n","answers":[]}
+{"QuestionId":53512769,"AnswerCount":1,"Tags":"<python><keras><lstm><sequential>","CreationDate":"2018-11-28T05:35:43.347","AcceptedAnswerId":53523708.0,"OwnerUserId":4542815.0,"Title":"Trouble understanding LSTM output.","Body":"<p>I have two data sets that contains sequence of numbers. One of them is my X and the other one is Y. <\/p>\n\n<p>For example,<\/p>\n\n<pre><code>X:\n1 0 7 2\n4 8 2 0\n5 9 2 1\n.\n.\n.\n\nShape of X is: (10000, 4)\n\nY:\n10 24 5 15\n7  6  10 4\n13 22 6  2\n.\n.\n.\nShape of Y is: (10000, 4)\n<\/code><\/pre>\n\n<p>The values in X lies in the range of 0-10 while the values in Y lies in range 0-24. <\/p>\n\n<p>I'm using LSTM implementation in Keras to train X and Y. Since LSTM model requires the input to be in 3 dimensional, I preprocessed the data and changed X to (10000, 4, 10) and Y to (10000, 4, 24).<\/p>\n\n<p>After preprocessing into one-hot encoding (these are the actual data and doesn't represent the one used for example):<\/p>\n\n<pre><code>X:\n[[[1 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 1 0 ... 0 0 0]\n  [1 0 0 ... 0 0 0]]\n\n [[0 0 1 ... 0 0 0]\n  [0 0 0 ... 0 1 0]\n  [1 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]\n  [0 0 0 ... 0 0 0]]\n\n ...\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]]\n\n [[0 0 0 ... 1 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [1 0 0 ... 0 0 0]\n  [1 0 0 ... 0 0 0]\n  [0 1 0 ... 0 0 0]]]\n\n\n\nY:\n[[[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]\n  [0 0 0 ... 0 0 0]]\n\n ...\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 1 0 0]]\n\n [[0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]\n  [0 0 0 ... 0 0 0]]]\n<\/code><\/pre>\n\n<p>This is the code for my LSTM model:<\/p>\n\n<pre><code>    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size = 0.2, random_state = 1)\n    model = Sequential()\n    model.add(LSTM(output_dim = 24, input_shape = X_train.shape[1:], return_sequences = True, init = 'glorot_normal', inner_init = 'glorot_normal', activation = 'sigmoid'))\n    model.add(LSTM(output_dim = 24, input_shape = X_train.shape[1:], return_sequences = True, init = 'glorot_normal', inner_init = 'glorot_normal', activation = 'sigmoid'))\n    model.add(LSTM(output_dim = 24, input_shape = X_train.shape[1:], return_sequences = True, init = 'glorot_normal', inner_init = 'glorot_normal', activation = 'sigmoid'))\n    model.add(LSTM(output_dim = 24, input_shape = X_train.shape[1:], return_sequences = True, init = 'glorot_normal', inner_init = 'glorot_normal', activation = 'sigmoid'))\n    model.compile(loss = 'categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])\n    model.fit(X_train, Y_train, nb_epoch = 500, validation_data = (X_test, Y_test))\n    model.save('LSTM500.h5')\n\n    predictions = model.predict(X_test)\n    predictions = np.array(predictions, dtype = np.float64)\n    predictions = predictions.reshape(2000, 4, 24)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[[[0.1552688  0.15805855 0.2013046  ... 0.16005482 0.19403476\n    0.        ]\n   [0.0458279  0.09995601 0.06456595 ... 0.09573169 0.07952237\n    0.        ]\n   [0.19871283 0.19968285 0.06270849 ... 0.14653654 0.18313469\n    0.        ]\n   [0.08407309 0.091876   0.09707277 ... 0.12661831 0.12858406\n    0.        ]]\n\n  [[0.15482235 0.14433247 0.18260191 ... 0.15641384 0.20746264\n    0.        ]\n   [0.03096719 0.05375536 0.05373315 ... 0.05018555 0.07592873\n    0.        ]\n   [0.20420487 0.17884348 0.13145864 ... 0.17901334 0.19768076\n    0.        ]\n   [0.03465272 0.06732351 0.02182322 ... 0.06144218 0.07827628\n    0.        ]]\n\n  [[0.15116604 0.15068266 0.18474537 ... 0.17088319 0.15841168\n    0.        ]\n   [0.09633015 0.11277901 0.10069521 ... 0.09309217 0.11326427\n    0.        ]\n   [0.17512578 0.13187788 0.10418645 ... 0.10735759 0.10635827\n    0.        ]\n   [0.13673681 0.12714103 0.06212005 ... 0.03213149 0.14153068\n    0.        ]]\n\n  ...\n<\/code><\/pre>\n\n<p>Shape of the prediction: (1, 2000, 4, 24) <\/p>\n\n<p>I reshaped the prediction array to (2000, 4, 24). <\/p>\n\n<pre><code>[[[0.1552688  0.15805855 0.2013046  ... 0.16005482 0.19403476 0.        ]\n  [0.0458279  0.09995601 0.06456595 ... 0.09573169 0.07952237 0.        ]\n  [0.19871283 0.19968285 0.06270849 ... 0.14653654 0.18313469 0.        ]\n  [0.08407309 0.091876   0.09707277 ... 0.12661831 0.12858406 0.        ]]\n\n [[0.15482235 0.14433247 0.18260191 ... 0.15641384 0.20746264 0.        ]\n  [0.03096719 0.05375536 0.05373315 ... 0.05018555 0.07592873 0.        ]\n  [0.20420487 0.17884348 0.13145864 ... 0.17901334 0.19768076 0.        ]\n  [0.03465272 0.06732351 0.02182322 ... 0.06144218 0.07827628 0.        ]]\n\n [[0.15116604 0.15068266 0.18474537 ... 0.17088319 0.15841168 0.        ]\n  [0.09633015 0.11277901 0.10069521 ... 0.09309217 0.11326427 0.        ]\n  [0.17512578 0.13187788 0.10418645 ... 0.10735759 0.10635827 0.        ]\n  [0.13673681 0.12714103 0.06212005 ... 0.03213149 0.14153068 0.        ]]\n\n ...\n<\/code><\/pre>\n\n<p>I do not seem to understand the output that I got. What are these numbers? Shouldn't the prediction array contains values only in 0 and 1 (this way, I can retrieve the actual value) just as Y_test? Thanks.<\/p>\n","answers":[{"AnswerId":"53523708","CreationDate":"2018-11-28T16:11:38.583","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>Those numbers are the numbers you got from your last layer, which is a sigmoid layer, sigmoid will return values between 0 and 1 which is what we see in your output. <\/p>\n\n<p>How to interpret these values? <\/p>\n\n<p>Since you are feeding and looking for one-hot outputs, you could select the maximum number in last axis and get its index value on that axis with <code>np.argmax(prediction, axis = -1)<\/code> this should give you a numpy array with shape (2000,4) , each element is a number between [0,24) which is the same format as your original data. These values are what your what your LSTM model predicted as the most likely outcome. <\/p>\n\n<p>Second biggest number would be the second most likely outcome.<\/p>\n"}]}
+{"QuestionId":53513439,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2018-11-28T06:31:19.540","AcceptedAnswerId":53513605.0,"OwnerUserId":10715353.0,"Title":"Error while input an index (Python 3.6) only integers, slices (`:`), ellipsis (`...`)","Body":"<p>I was trying to run <a href=\"https:\/\/github.com\/AppliedDataSciencePartners\/DeepReinforcementLearning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/AppliedDataSciencePartners\/DeepReinforcementLearning<\/a> on jupyter notebook and everything was fine until I execute this block.<\/p>\n\n<pre><code>from game import Game\nfrom funcs import playMatchesBetweenVersions\n\nimport loggers as lg\n\nenv = Game()\n\nplayMatchesBetweenVersions(env, 2, -1, 26, 1, lg.logger_tourney,0, 1)\n<\/code><\/pre>\n\n<p>Where the important part is that -1 is for a human player if I run this, this happens. <a href=\"https:\/\/i.stack.imgur.com\/9JcbB.png\" rel=\"nofollow noreferrer\">Ask for imput<\/a><\/p>\n\n<p>when entering a valid input this is what happens. <a href=\"https:\/\/i.stack.imgur.com\/HLVaY.png\" rel=\"nofollow noreferrer\">IndexError: only integers, slices (<code>:<\/code>), ellipsis (<code>...<\/code>), numpy.newaxis (<code>None<\/code>) and integer or boolean arrays are valid indices<\/a><\/p>\n","answers":[{"AnswerId":"53513605","CreationDate":"2018-11-28T06:43:55.947","ParentId":null,"OwnerUserId":"10702969","Title":null,"Body":"<p>In file<\/p>\n\n<blockquote>\n  <p>D:\\Codigo fuente\\agent.py<\/p>\n<\/blockquote>\n\n<p>on line 26<\/p>\n\n<pre><code>action = input('Enter your chosen action: ')\n<\/code><\/pre>\n\n<p><code>action<\/code> variable is of type string. You need to convert <code>action<\/code> variable to int for indexing.<\/p>\n\n<p>It should be<\/p>\n\n<pre><code>action = int(input('Enter your chosen action: '))\n<\/code><\/pre>\n"},{"AnswerId":"53513676","CreationDate":"2018-11-28T06:49:11.000","ParentId":null,"OwnerUserId":"6403406","Title":null,"Body":"<p>Change line 26 to:<\/p>\n\n<p><code>action=int(input(\"Enter your chosen answer: \"))<\/code><\/p>\n\n<p>as I mentioned in the comments.<\/p>\n"}]}
+{"QuestionId":53513515,"AnswerCount":1,"Tags":"<classification><conv-neural-network><caffe>","CreationDate":"2018-11-28T06:37:39.867","AcceptedAnswerId":53520065.0,"OwnerUserId":8433989.0,"Title":"Caffe-terminate called after throwing an instance of 'std::bad_alloc' while building a Cat\/Dog Classifier using a Convolutional Neural Network","Body":"<p>I am following <a href=\"http:\/\/adilmoujahid.com\/posts\/2016\/06\/introduction-deep-learning-python-caffe\/\" rel=\"nofollow noreferrer\">this tutorial<\/a> to build cat-dog classifier using caffe. But while training I got this error;<a href=\"https:\/\/i.stack.imgur.com\/iTomT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iTomT.png\" alt=\"enter image description here\"><\/a> why?<\/p>\n","answers":[{"AnswerId":"53520065","CreationDate":"2018-11-28T13:00:03.717","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The error you got <code>\"std::bad_alloc\"<\/code> usually refers to CPU memory allocation: you are out of memory.<br>\nJudging by the stack trace, you got the error when trying to read an element from the LMDB file. You need to check your LMDB: is it possible you store HUGE images there? is it possible the LMDB is corrupt? can you read it with different software?<\/p>\n"}]}
+{"QuestionId":53513527,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T06:38:29.710","AcceptedAnswerId":null,"OwnerUserId":10715396.0,"Title":"ValueError: Shape must be rank 3 but is rank 4 for 'adjust_hue\/Slice' (op: 'Slice') with input shapes: [384,12,12,3], [3], [3]","Body":"<p>While working on a TensorFlow project, I've encountered a problem and I don't know how to solve it. I have <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZZ8DI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZZ8DI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 3 but is rank 4 for 'adjust_hue\/Slice' (op: 'Slice') with input shapes: [384,12,12,3], [3], [3].<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"53662769","CreationDate":"2018-12-07T03:24:07.713","ParentId":null,"OwnerUserId":"3012517","Title":null,"Body":"<p>according to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8926\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8926<\/a><\/p>\n\n<p>some <code>tf.image<\/code> functions only accept a single image as an input.<\/p>\n\n<p>but some image ops (e.g. <code>tf.image.random_brightness()<\/code>, <code>tf.image.random_contrast()<\/code>) allow batches.<\/p>\n\n<p>the solution is using lambda to work on batch:<\/p>\n\n<pre><code>    hue = lambda x: tf.image.random_hue(x, 0.5)\n    inputs = tf.map_fn(hue, inputs)\n<\/code><\/pre>\n"},{"AnswerId":"53913932","CreationDate":"2018-12-24T13:11:02.853","ParentId":null,"OwnerUserId":"7283818","Title":null,"Body":"<p>Just delete\/comment  tf.image.random_hue and image.random_saturation in function image_color_distort at train.py like the code bellow:<\/p>\n\n<pre><code>def image_color_distort(inputs):\n    inputs = tf.image.random_contrast(inputs, lower=0.5, upper=1.5)\n    inputs = tf.image.random_brightness(inputs, max_delta=0.2)\n    # inputs = tf.image.random_hue(inputs,max_delta= 0.2)\n    # inputs = tf.image.random_saturation(inputs,lower = 0.5, upper= 1.5)\n\n    return inputs\n<\/code><\/pre>\n"}]}
+{"QuestionId":53513707,"AnswerCount":1,"Tags":"<pytorch><amazon-ami>","CreationDate":"2018-11-28T06:51:04.443","AcceptedAnswerId":null,"OwnerUserId":5012631.0,"Title":"PyTorch not detecting CUDA in Amazon Deep learning AMI","Body":"<p>I am using Amazon Deep Learning AMI (Ubuntu) Version 18.0 (with Ubuntu 16.04 LTS). After activating pytorch environment, PyTorch is not detecting CUDA. Also, nvidia-smi shows \"NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.\"\nHow to solve this problem?<\/p>\n","answers":[{"AnswerId":"53794321","CreationDate":"2018-12-15T14:50:39.660","ParentId":null,"OwnerUserId":"7128719","Title":null,"Body":"<p>What EC2 instance type are you using? <\/p>\n\n<p>Are you sure that the EC2 instance type has GPUs?<\/p>\n"}]}
+{"QuestionId":53513892,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-28T07:04:31.570","AcceptedAnswerId":53534547.0,"OwnerUserId":260826.0,"Title":"Replace \"from tensorflow.contrib import layers\"","Body":"<p>How to replace <code>from tensorflow.contrib import layers<\/code> with new core functionality. I want to move my TF 1.4 code to 1.12 in preparation for TF 2.0.<\/p>\n","answers":[{"AnswerId":"53534547","CreationDate":"2018-11-29T08:20:36.943","ParentId":null,"OwnerUserId":"2260153","Title":null,"Body":"<p>The core functionality corresponding to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\" rel=\"nofollow noreferrer\"><code>tf.contrib.layers<\/code><\/a> is in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\" rel=\"nofollow noreferrer\"><code>tf.layers<\/code><\/a>. Some of the differences are discussed in <a href=\"https:\/\/stackoverflow.com\/questions\/44805675\/tensorflow-tf-layers-vs-tf-contrib-layers\">this question<\/a>. However this will not prepare you for TF 2.0.<\/p>\n\n<p>If your goal is to prepare your code for TF 2.0, consider that <code>tf.contrib<\/code> will be removed entirely (either split from TF or integrated into it) and that <code>tf.layers<\/code> too will be removed and the high-level API will reside under <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\" rel=\"nofollow noreferrer\"><code>tf.keras<\/code><\/a>. So to best prepare for TF 2.0 you should start using <code>tf.keras.layers<\/code> instead.<\/p>\n\n<p>Here is a <a href=\"https:\/\/pgaleone.eu\/tensorflow\/gan\/2018\/11\/04\/tensorflow-2-models-migration-and-new-design\/\" rel=\"nofollow noreferrer\">blog post<\/a> about some of the practical differences to expect with TF 2.0.<\/p>\n"}]}
+{"QuestionId":53514495,"AnswerCount":2,"Tags":"<tensorflow><dataset>","CreationDate":"2018-11-28T07:47:03.473","AcceptedAnswerId":53517848.0,"OwnerUserId":10492790.0,"Title":"What does batch, repeat, and shuffle do with TensorFlow Dataset?","Body":"<p>I'm currently learning TensorFlow but i come across a confusion within this code:<\/p>\n\n<pre><code>dataset = dataset.shuffle(buffer_size = 10 * batch_size) \ndataset = dataset.repeat(num_epochs).batch(batch_size)\nreturn dataset.make_one_shot_iterator().get_next()\n<\/code><\/pre>\n\n<p>i know first the dataset will hold all the data but what shuffle(),repeat(), and batch() do to the dataset? please give me an explanation with an example<\/p>\n","answers":[{"AnswerId":"53517848","CreationDate":"2018-11-28T10:57:36.350","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>Imagine, you have a dataset: <code>[1, 2, 3, 4, 5, 6]<\/code>, then:<\/p>\n\n<p><strong>How ds.shuffle() works<\/strong><\/p>\n\n<p><code>dataset.shuffle(buffer_size=3)<\/code> will allocate a buffer of size 3 for picking random entries. This buffer will be connected to the source dataset.\nWe could image it like this:<\/p>\n\n<pre><code>Random buffer\n   |\n   |   Source dataset where all other elements live\n   |         |\n   \u2193         \u2193\n[1,2,3] &lt;= [4,5,6]\n<\/code><\/pre>\n\n<p>Let's assume that the entry <code>2<\/code> was taken from the random buffer. Free space is filled by the next element from the source buffer, that is <code>4<\/code>:<\/p>\n\n<pre><code>2 &lt;= [1,3,4] &lt;= [5,6]\n<\/code><\/pre>\n\n<p>We continue reading till nothing is left:<\/p>\n\n<pre><code>1 &lt;= [3,4,5] &lt;= [6]\n5 &lt;= [3,4,6] &lt;= []\n3 &lt;= [4,6]   &lt;= []\n6 &lt;= [4]      &lt;= []\n4 &lt;= []      &lt;= []\n<\/code><\/pre>\n\n<p><strong>How ds.repeat() works<\/strong><\/p>\n\n<p>As soon as all the entries are read from the dataset and you try to read the next element, the dataset will throw an error.\nThat's where <code>ds.repeat()<\/code> comes into play. It will re-initialize the dataset, making it again like this:<\/p>\n\n<pre><code>[1,2,3] &lt;= [4,5,6]\n<\/code><\/pre>\n\n<p><strong>What will ds.batch() produce<\/strong><\/p>\n\n<p>The <code>ds.batch()<\/code> will take first <code>batch_size<\/code> entries and make a batch out of them. So, batch size of 3 for our example dataset will produce two batch records:<\/p>\n\n<pre><code>[2,1,5]\n[3,6,4]\n<\/code><\/pre>\n\n<p>As we have a <code>ds.repeat()<\/code> before the batch, the generation of the data will continue. But the order of the elements will be different, due to the <code>ds.random()<\/code>. What should be taken into account is that <code>6<\/code> will never be present in the first batch, due to the size of the random buffer.<\/p>\n"},{"AnswerId":"53517772","CreationDate":"2018-11-28T10:53:19.677","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>The following methods in tf.Dataset :<\/p>\n\n<ol>\n<li><code>repeat( count=0 )<\/code> The method repeats the dataset <code>count<\/code> number of times.<\/li>\n<li><code>shuffle( buffer_size, seed=None, reshuffle_each_iteration=None)<\/code> The method shuffles the samples in the dataset. The <code>buffer_size<\/code> is the number of samples which are randomized and returned as <code>tf.Dataset<\/code>.<\/li>\n<li><code>batch(batch_size,drop_remainder=False)<\/code> Creates batches of the dataset with batch size given as <code>batch_size<\/code> which is also the length of the batches.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":53514751,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><classification>","CreationDate":"2018-11-28T08:04:42.543","AcceptedAnswerId":53514927.0,"OwnerUserId":7999915.0,"Title":"ValueError: Error when checking input: expected flatten_1_input to have shape (4, 4, 512) but got array with shape (128, 128, 3)","Body":"<p>I have the following code for an image classification problem. And I keep running into this error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected flatten_1_input to have shape (4, 4, 512) but got array with shape (128, 128, 3)\n<\/code><\/pre>\n\n<p>I've seen a similar problem with someone who butchered their model loading process, but I'm not doing that. Here's my code:<\/p>\n\n<pre><code>def save_bottlebeck_features():`\n    datagen = ImageDataGenerator(rescale=1. \/ 255)\n    \u200b\n    # build the VGG16 network\n    model = applications.VGG16(include_top=False, weights='imagenet')\n    \u200b\n    generator = datagen.flow_from_directory(\n            train_data_dir,\n            target_size=(img_width, img_height),\n            batch_size=batch_size,\n            class_mode=None,\n            shuffle=False)\n\n   bottleneck_features_train = model.predict_generator(\n            generator, nb_train_samples \/\/ batch_size)\n   np.save('bottleneck_features_train.npy', bottleneck_features_train)\n    \u200b\n   generator = datagen.flow_from_directory(\n            validation_data_dir,\n            target_size=(img_width, img_height),\n            batch_size=batch_size,\n            class_mode=None,\n            shuffle=False)\n\n   bottleneck_features_validation = model.predict_generator(\n            generator, nb_validation_samples \/\/ batch_size)\n   np.save('bottleneck_features_validation.npy',bottleneck_features_validation)\n\ndef train_top_model():\n\n  train_data = np.load('bottleneck_features_train.npy',\"r+\")\n  train_labels = np.array([0] * (nb_train_samples \/\/ 2) + [1] * (nb_train_samples \/\/ 2))\n  validation_data = np.load('bottleneck_features_validation.npy',\"r+\")\n  validation_labels = np.array([0] * (nb_validation_samples \/\/ 2) + [1] * (nb_validation_samples \/\/ 2))\n\n  model = Sequential()\n  model.add(Flatten(input_shape=train_data.shape[1:]))\n  model.add(Dense(256, activation='relu'))\n  model.add(Dropout(0.5))\n  model.add(Dense(1, activation='sigmoid'))\n\n  model.compile(optimizer='rmsprop',\n                  loss='binary_crossentropy', metrics=['accuracy'])\n\n  model.fit(train_data, train_labels,\n              epochs=epochs,\n              batch_size=batch_size,\n              validation_data=(validation_data, validation_labels))\n  model.save_weights(top_model_weights_path)\n\n  model.save('my_model.model')\n\nsave_bottlebeck_features()\ntrain_top_model()\n\nfrom keras.preprocessing import image\nimport numpy as np\nfrom keras.models import load_model\nimport os\n\nresnet_50 = load_model(\"my_model.model\")\n\nTEST_DIR = 'test\/'\n\nwith open('better_score.csv','w') as f:\n    f.write('Id,Expected\\n')\n    for x in range(1,7091): \n        mystr = \"test_\" + str(x) +\".jpg\"\n        path = os.path.join(TEST_DIR, mystr)\n\n        if (os.path.exists(path)):\n            img = image.load_img(path, target_size=(128, 128))\n            img = image.img_to_array(img)\n            img = np.expand_dims(img, axis=0)\n            model_out  = resnet_50.predict(img\/255)\n            f.write('{},{}\\n'.format(mystr, model_out[0][0]))\n<\/code><\/pre>\n\n<p>I have printed the <code>train_data<\/code> shape and that is <code>(2000, 4, 4, 512)<\/code>\nand <code>validation_data<\/code> shape is <code>(800, 4, 4, 512)<\/code>. I am able to train my model, and save it. The problem occurs when I get to the line before last, when I'm trying to output my results to a csv file.<\/p>\n","answers":[{"AnswerId":"53514927","CreationDate":"2018-11-28T08:17:30.087","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You first need to extract features of your input image using VGG model and then pass the features to <code>resnet_50<\/code> model. So it would look like this:<\/p>\n\n<pre><code>model = applications.VGG16(include_top=False, weights='imagenet')\nfeatures = model.predict(img\/255.0)\nmodel_out = resnet_50.predict(features)\n<\/code><\/pre>\n\n<p>Don't forget that the input of your <code>resent_50<\/code> model is the features of the image obtained by VGG model.<\/p>\n"}]}
+{"QuestionId":53515096,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-28T08:27:53.670","AcceptedAnswerId":null,"OwnerUserId":6431509.0,"Title":"How to normalization params when use tensorflow to inference","Body":"<p>I'm study tensorflow,for example,I have a array contains 1~10000 linear number to be a train data,before I train data I normalization this array like below<\/p>\n\n<pre><code>from sklearn import preprocessing\n\nmin_max_scaler = preprocessing.MinMaxScaler()\ndata_array = min_max_scaler.fit_transform(data_array)\n<\/code><\/pre>\n\n<p>now,I get a model,but how to use this model to inference? my app user input a param maybe 20000,how should I normalize 20000\uff1f <\/p>\n","answers":[{"AnswerId":"53517567","CreationDate":"2018-11-28T10:43:26.480","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>You can use the simple normalization formula : <\/p>\n\n<ol>\n<li>Calculate the min and max from your data. Min=1 and max=20000<\/li>\n<li>For every element in the array, which is suppose 'x' , subtract the min and divide the answer by max -min<\/li>\n<li>The formula : x( new ) = x( old ) - min \/ max - min<\/li>\n<\/ol>\n\n<p>Refer <a href=\"https:\/\/medium.com\/@equipintelligence\/normalization-of-crazy-data-python-4fa6611e7b46\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":53515395,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow><tensorflow-serving>","CreationDate":"2018-11-28T08:47:27.270","AcceptedAnswerId":53516328.0,"OwnerUserId":1111479.0,"Title":"Failed on freeze graph for resnet_v1_50 using TF-slim","Body":"<p>I would like to freeze the model and deploy them. I firstly followed the tutorial shipped with slim from <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim#exporting-the-inference-graph\" rel=\"nofollow noreferrer\">1<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim#freezing-the-exported-graph\" rel=\"nofollow noreferrer\">2<\/a>, using InceptionV3, and it was good. Then I am trying to do the same for resnet_v1_50, downloaded from <a href=\"http:\/\/download.tensorflow.org\/models\/resnet_v1_50_2016_08_28.tar.gz\" rel=\"nofollow noreferrer\">the provided url<\/a>. <\/p>\n\n<p>Any suggestions would be appreciated! <\/p>\n\n<p>Thanks!<\/p>\n\n<p>Following is the steps:<\/p>\n\n<pre><code>cd \/path\/to\/tensorflow-models\/research\/slim\n# export inference graph, success\npython export_inference_graph.py --alsologtostderr --model_name=resnet_v1_50 --output_file=resnet_v1_50_inf_graph.pb\n# freeze the inference graph with the trained weights, failed\npython ~\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py \\\n    --input_graph=resnet_v1_50_inf_graph.pb \\\n    --input_checkpoint=\/path\/to\/models\/resnet_v1_50.ckpt \\\n    --input_binary=true --output_graph=frozen_resnet_v1_50.pb \\\n    --output_node_names=resnet_v1_50\/predictions\/Reshape\n<\/code><\/pre>\n\n<p>The error message is too long and I redirect one to refer it <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5824#issuecomment-442367921\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n","answers":[{"AnswerId":"53516328","CreationDate":"2018-11-28T09:37:37.480","ParentId":null,"OwnerUserId":"1111479","Title":null,"Body":"<p>As the following line suggests, the <code>labels_offset<\/code> should be set to <code>1<\/code> when exporting inference graph.<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Assign requires shapes of both \ntensors to match. lhs shape= [1001] rhs shape= [1000]\n     [[Node: save\/Assign_265 = Assign[T=DT_FLOAT, _class= . \n[\"loc:@resnet_v1_50\/logits\/biases\"], use_locking=true, validate_shape=true, \n_device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](resnet_v1_50\/logits\/biases, \nsave\/RestoreV2_265\/_1)]]\n<\/code><\/pre>\n\n<p>The command will be like:<\/p>\n\n<pre><code>python export_inference_graph.py --model_name=resnet_v1_50 --labels_offset=1 --output_file=resnet_v1_50_inf_graph.pb\n<\/code><\/pre>\n"}]}
+{"QuestionId":53515547,"AnswerCount":1,"Tags":"<keras><nlp><lstm><language-model><perplexity>","CreationDate":"2018-11-28T08:56:22.023","AcceptedAnswerId":null,"OwnerUserId":6057371.0,"Title":"Check perplexity of a Language Model","Body":"<p>I created a language model with Keras LSTM and now I want to assess wether it's good so I want to calculate perplexity.<\/p>\n\n<p>What is the best way to calc perplexity of a model in Python?<\/p>\n","answers":[{"AnswerId":"53564032","CreationDate":"2018-11-30T19:52:05.617","ParentId":null,"OwnerUserId":"2931946","Title":null,"Body":"<p>I've come up with two versions and attached their corresponding source, please feel free to check the links out.<\/p>\n\n<pre><code>def perplexity_raw(y_true, y_pred):\n    \"\"\"\n    The perplexity metric. Why isn't this part of Keras yet?!\n    https:\/\/stackoverflow.com\/questions\/41881308\/how-to-calculate-perplexity-of-rnn-in-tensorflow\n    https:\/\/github.com\/keras-team\/keras\/issues\/8267\n    \"\"\"\n#     cross_entropy = K.sparse_categorical_crossentropy(y_true, y_pred)\n    cross_entropy = K.cast(K.equal(K.max(y_true, axis=-1),\n                          K.cast(K.argmax(y_pred, axis=-1), K.floatx())),\n                  K.floatx())\n    perplexity = K.exp(cross_entropy)\n    return perplexity\n\ndef perplexity(y_true, y_pred):\n    \"\"\"\n    The perplexity metric. Why isn't this part of Keras yet?!\n    https:\/\/stackoverflow.com\/questions\/41881308\/how-to-calculate-perplexity-of-rnn-in-tensorflow\n    https:\/\/github.com\/keras-team\/keras\/issues\/8267\n    \"\"\"\n    cross_entropy = K.sparse_categorical_crossentropy(y_true, y_pred)\n    perplexity = K.exp(cross_entropy)\n    return perplexity\n<\/code><\/pre>\n"}]}
+{"QuestionId":53515719,"AnswerCount":0,"Tags":"<python><lstm><pytorch><rnn>","CreationDate":"2018-11-28T09:06:46.403","AcceptedAnswerId":null,"OwnerUserId":3726958.0,"Title":"DataLoader issue : RNN module weights are not part of single contiguous chunk of memory","Body":"<p>I am trying to build an LSTM model for time series data. The details of dataset:<br>\nInput data is time series with 800 subjects, each having a 2D array data of 60 rows and 200 columns. I loaded the entire data as a Tensor of shape <code>[800,60,200]<\/code> and labels for the classification problem is of shape <code>[800,1]<\/code>.\nI made a dictionary of the data using the following code :<\/p>\n\n<pre><code>class DataCurate(Dataset):\n    def __init__(self, l1,l2, transform=None):\n        self.l1=l1\n        self.l2=l2\n        self.transform=transform\n    def __len__(self):\n        return len(self.l1)\n    def __getitem__(self, index):\n        array=self.l1[index]\n        label=self.l2[index]\n        sample = {'time_data': array, 'labels': label}\n        return sample\n<\/code><\/pre>\n\n<p>the data and labels are in variables x &amp; y. I call data=Datacurate(x,y)<\/p>\n\n<p>Later on I build and LSTM model for classification problem using the code:<\/p>\n\n<pre><code>class RNNModel(nn.Module):\n    def __init__(self, input_size, hidden_size, num_layers, num_classes):\n    super(RNNModel, self).__init__()\n    self.hidden_size = hidden_size\n    self.num_layers = num_layers\n    self.lstm = nn.LSTM(input_size, hidden_size, num_layers,batch_first=True) \n    self.linear = nn.Linear(hidden_size, num_layers, bias=True) \n    self.softmax = nn.LogSoftmax()\n\n    def forward(self, x)\n        self.lstm.flatten_parameters()\n        out_packed, state = self.lstm(x)  # RNN\n        print(\"lstm output size: {out.size()}\"+str(out_packed.size()))\n        out = self.linear(out_packed[-1])  # linear transform\n    print(\"linear output size {out.size()} \"+str(out.size()))\n    log_probs = F.log_softmax(out,dim=1)\n    print(\"softmax output size {log_probs.size()}\"+str(log_probs.size()))\n        return log_probs\n<\/code><\/pre>\n\n<p>This gives me an error when I run the training script: <\/p>\n\n<pre><code>UserWarning: RNN module weights are not part of single contiguous chunk of memory. This means they need to be compacted at every call, possibly greatly increasing memory usage. To compact weights again call flatten_parameters().\n  out_packed, state = self.lstm(x)  # RNN\nTraceback (most recent call last):\n  File \"main_2.py\", line 100, in &lt;module&gt;\n    output = model(train_inputs.transpose(0,1))\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/media\/iab\/disk_a\/meghal\/test\/quickdraw_tutorial_dataset_v1\/pytorch_RNN_examples\/model.py\", line 26, in forward\n    out_packed, state = self.lstm(x)  # RNN\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/module.py\", line 477, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/modules\/rnn.py\", line 192, in forward\n    output, hidden = func(input, self.all_weights, hx, batch_sizes)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/_functions\/rnn.py\", line 324, in forward\n    return func(input, *fargs, **fkwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/torch\/nn\/_functions\/rnn.py\", line 288, in forward\n    dropout_ts)\nRuntimeError: param_from.type() == param_to.type() ASSERT FAILED at \/pytorch\/aten\/src\/ATen\/native\/cudnn\/RNN.cpp:491, please report a bug to PyTorch. parameter types mismatch\n<\/code><\/pre>\n\n<p>I don't know what this means and how do I resolve it. I am completely new to LSTM.<\/p>\n","answers":[]}
+{"QuestionId":53516046,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><conv-neural-network>","CreationDate":"2018-11-28T09:24:17.383","AcceptedAnswerId":null,"OwnerUserId":10610292.0,"Title":"Keras yields 97 percent on test data for binary classification, but in the end all 1s are predicts","Body":"<p>With a simple binary classification, the model constantly gives answer 1 (even if we input data from different types of folders on which we were trained to it), while the accuracy of both test and validation samples is high - about 97\/92 percent.\nThere are about 500 images for each class, but I also use generators.<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale=1. \/ 255,\n                                  rotation_range=40,\n                                  width_shift_range=0.2,\n                                  height_shift_range=0.2,\n                                  zoom_range=0.2,\n                                  horizontal_flip=True,\n                                  fill_mode='nearest')`\n\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=input_shape))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(32, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(64, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\nmodel.compile(loss='binary_crossentropy',\n               optimizer='adam',\n               metrics=['accuracy'])\n\nmodel.fit_generator(\ntrain_generator,\nsteps_per_epoch=100,\nepochs=1,\nvalidation_data=val_generator,\nvalidation_steps=20)\n<\/code><\/pre>\n\n<p>image type 0<\/p>\n\n<p>real: 0.0\npred: 1.<\/p>\n\n<p>image type 1<\/p>\n\n<p>real: 1.0\npred: 1.<\/p>\n","answers":[{"AnswerId":"53521211","CreationDate":"2018-11-28T14:02:05.903","ParentId":null,"OwnerUserId":"9919799","Title":null,"Body":"<p>You are rescaling your test images in your generator <code>ImageDataGenerator(rescale=1.\/255, ...)<\/code> and probably forgetting to do this to your test images. Try <code>model.predict(test_images\/255)<\/code><\/p>\n"}]}
+{"QuestionId":53516662,"AnswerCount":1,"Tags":"<python><image><vector><keras>","CreationDate":"2018-11-28T09:56:20.753","AcceptedAnswerId":53517528.0,"OwnerUserId":10127088.0,"Title":"Two inputs to one model in Keras","Body":"<p>Is it possible in Keras to feed both an image and a vector of values as inputs to one model? If yes, how? <\/p>\n\n<p>What I want is to create a CNN with an image and a vector of 6 values on the input.<\/p>\n\n<p>The output is the vector of 3 values. <\/p>\n","answers":[{"AnswerId":"53517528","CreationDate":"2018-11-28T10:41:41.807","ParentId":null,"OwnerUserId":"5495381","Title":null,"Body":"<p>Yes, please have a look at Keras' <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"noreferrer\">Functional API<\/a> for many examples on how to build models with multiple inputs. <\/p>\n\n<p>Your code will look something like this, where you will probably want to pass the image through a convolutional layer, flatten the output and concatenate it with your vector input:<\/p>\n\n<pre><code>from keras.layers import Input, Concatenate, Conv2D, Flatten, Dense\nfrom keras.models import Model\n\n# Define two input layers\nimage_input = Input((32, 32, 3))\nvector_input = Input((6,))\n\n# Convolution + Flatten for the image\nconv_layer = Conv2D(32, (3,3))(image_input)\nflat_layer = Flatten()(conv_layer)\n\n# Concatenate the convolutional features and the vector input\nconcat_layer= Concatenate()([vector_input, flat_layer])\noutput = Dense(3)(concat_layer)\n\n# define a model with a list of two inputs\nmodel = Model(inputs=[image_input, vector_input], outputs=output)\n<\/code><\/pre>\n\n<p>This will give you a model with the following specs:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_8 (InputLayer)            (None, 32, 32, 3)    0                                            \n__________________________________________________________________________________________________\nconv2d_4 (Conv2D)               (None, 30, 30, 32)   896         input_8[0][0]                    \n__________________________________________________________________________________________________\ninput_9 (InputLayer)            (None, 6)            0                                            \n__________________________________________________________________________________________________\nflatten_3 (Flatten)             (None, 28800)        0           conv2d_4[0][0]                   \n__________________________________________________________________________________________________\nconcatenate_3 (Concatenate)     (None, 28806)        0           input_9[0][0]                    \n                                                                 flatten_3[0][0]                  \n__________________________________________________________________________________________________\ndense_3 (Dense)                 (None, 3)            86421       concatenate_3[0][0]              \n==================================================================================================\nTotal params: 87,317\nTrainable params: 87,317\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Another way to visualize it is through <a href=\"https:\/\/keras.io\/visualization\/\" rel=\"noreferrer\">Keras' visualization utilities<\/a>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AhbED.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AhbED.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53516966,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-28T10:12:44.627","AcceptedAnswerId":53569444.0,"OwnerUserId":7275235.0,"Title":"Keras - 'Node' object has no attribute 'output_masks'","Body":"<p>I'm new in Tensorflow and Keras. I trying to run code in Keras. I have labels that convert to one hot encoding whit 2 depth and after that convert it to Keras layers format. \nnow when I rum my code I got this error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n  File \"\/Users\/vahid\/Documents\/Thesis\/Code\/new_code\/CapsNet_model.py\", line       224, in \n      CapsNet((None,28,28,1),2,3,trainImg,trainLbl)\n    File \"\/Users\/vahid\/Documents\/Thesis\/Code\/new_code\/CapsNet_model.py\", line 164, in CapsNet\n      x_recon = k.layers.Dense(512, activation='relu')(masked)\n    File \"\/Users\/vahid\/TensorProject\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py\", line 443, in <strong>call<\/strong>\n      previous_mask = _collect_previous_mask(inputs)\n    File \"\/Users\/vahid\/TensorProject\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py\", line 1311, in _collect_previous_mask\n  mask = node.output_masks[tensor_index]\n      AttributeError: 'Node' object has no attribute 'output_masks'<\/p>\n<\/blockquote>\n\n<p>and the part of the code that I run and got error:<\/p>\n\n<pre><code>def CapsNet(input_shape, n_class, num_routing,img,lbl):\n\"\"\"\n:param input_shape: (None, width, height, channels)\n:param n_class: number of classes\n:param num_routing: number of routing iterations\n:return: A Keras Model with 2 inputs (image, label) and\n         2 outputs (capsule output and reconstruct image)\n\"\"\"\n# Image\n#x = k.layers.Input(shape=input_shape,tensor=img)\n# ReLU Conv1\n\nimages, labels = tf.train.shuffle_batch([img, lbl], batch_size=50, capacity=30, num_threads=1, min_after_dequeue=10)\n\ninp = k.layers.Lambda(lambda x: x)(images)\n#inpy = k.layers.Lambda(lambda x: x,output_shape=(n_class,2))(labels)\n\n#conv1 = k.layers.Conv2D(nb_filter= 256,nb_row= 9 , nb_col= 9, subsample =(1,1),activation='relu', name='conv1')(x)\nconv1 = k.layers.Conv2D(filters = 256, kernel_size =9 , strides=(1,1),padding='valid', activation='relu', name='conv1')(inp)\n\n# PrimaryCapsules: Conv2D layer with `squash` activation,\n# reshape to [None, num_capsule, dim_vector]\nprimarycaps = PrimaryCap(conv1, dim_vector=8, n_channels=32,kernel_size=9, strides=2, padding='valid')\n#primarycaps = PrimaryCap(conv1, dim_vector=8, n_channels=32 , subsample =(2,2) )\n\n# DigiCap: Capsule layer. Routing algorithm works here.\n\ndigitcaps = DigiCap(num_capsule = n_class, dim_vector = 16, num_routing = num_routing, name='digitcaps')(primarycaps)\n\n# The length of the capsule's output vector\nout_caps = Length(name='out_caps')(digitcaps)\n\n# Decoder network.\ny = k.layers.Input(shape=(n_class,),tensor=labels)\n\n# The true label is used to extract the corresponding vj\nmasked = Mask()([digitcaps, y])\nx_recon = k.layers.Dense(512, activation='relu')(masked)\nx_recon = k.layers.Dense(1024, activation='relu')(x_recon)\nx_recon = k.layers.Dense(784, activation='sigmoid')(x_recon)\nx_recon = k.layers.Reshape(target_shape=[28, 28, 1], name='out_recon')(x_recon)\n\n# two-input-two-output keras Model\nreturn k.models.Model([inp, y], [out_caps, x_recon])\n<\/code><\/pre>\n\n<p>code of Mask:<\/p>\n\n<pre><code>class Mask(tf.layers.Layer):\n\"\"\"\nMask a Tensor with shape=[None, d1, d2] by the max value in axis=1.\nOutput shape: [None, d2]\n\"\"\"\ndef call(self, inputs, **kwargs):\n    # use true label to select target capsule, shape=[batch_size, num_capsule]\n    if type(inputs) is list:  # true label is provided with shape = [batch_size, n_classes], i.e. one-hot code.\n        assert len(inputs) == 2\n        inputs, mask = inputs\n\n    else:  # if no true label, mask by the max length of vectors of capsules\n        x = inputs\n        x = tf.cast(x,tf.float32)\n        # Enlarge the range of values in x to make max(new_x)=1 and others &lt; 0\n        x = (x - K.max(x, 1, True)) \/ K.epsilon() + 1\n        mask = K.clip(x, 0, 1)  # the max value in x clipped to 1 and other to 0\n\n    # masked inputs, shape = [batch_size, dim_vector]\n    inputs_masked = K.batch_dot(inputs, mask, [1, 1])\n    return inputs_masked\n<\/code><\/pre>\n","answers":[{"AnswerId":"53569444","CreationDate":"2018-12-01T09:31:21.170","ParentId":null,"OwnerUserId":"7275235","Title":null,"Body":"<p>well, I find out how to fix this problem. actually, we can use two kinds of Keras in our code. Keras package or just use tf.keras. In this code I using Keras package when using \"Dense\". for example in   <code>x_recon = k.layers.Dense(512, activation='relu')(masked)<\/code> so it seems that tf.keras and Keras have different source and when I've changed <code>k.layers<\/code>to <code>tf.keras<\/code> problem was fixed.<\/p>\n"}]}
+{"QuestionId":53517114,"AnswerCount":0,"Tags":"<python><tensorflow><segmentation-fault><tensorflow-estimator>","CreationDate":"2018-11-28T10:20:01.237","AcceptedAnswerId":null,"OwnerUserId":6859185.0,"Title":"How to track down the cause a SIGSEGV error in an ML Engine training job?","Body":"<p>I am training a custom tensorflow estimator on ML engine and running into this error:<\/p>\n\n<pre><code>The replica master 0 exited with a non-zero status of 11(SIGSEGV)\n<\/code><\/pre>\n\n<p>with the only other error log being:<\/p>\n\n<pre><code>Command '['python3', '-m', 'train_model.train', ... ']' returned non-zero exit status -11\n<\/code><\/pre>\n\n<p>There is no longer traceback so this 'invalid memory reference or segmentation fault <a href=\"https:\/\/www.codechef.com\/wiki\/status-codes\" rel=\"nofollow noreferrer\">code<\/a>' is all I have to go on. <\/p>\n\n<p>This SIGSEGV error does not always occur. Some training jobs run without problems, others throw this error after 4 hours, and others after 15 minutes.<\/p>\n\n<p>I've varied parts of the estimator training code to try to trial and error my way to finding the cause but have had no success.<\/p>\n\n<p>I thought the <code>11<\/code> code may correspond to <a href=\"https:\/\/github.com\/googleapis\/googleapis\/blob\/master\/google\/rpc\/code.proto#L157\" rel=\"nofollow noreferrer\">this error code<\/a> in the google apis and found a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19541\" rel=\"nofollow noreferrer\">number of people have experienced<\/a> <code>OutOfSequence<\/code> and <code>OutOfRange<\/code> errors when using custom metrics in the estimator <code>EvalSpec<\/code> but I do not think that is what is causing the error here as I use a <code>tf.metric<\/code>.<\/p>\n\n<p>I am using the <code>BASIC<\/code> scale tier and looking at the CPU utilisation it never goes above 80% and the memory utilisation graph shows around 25%. <\/p>\n\n<p>I am caching the tensorflow dataset, but I also receive this error when not caching the dataset. The error occurs when running the <code>train_and_evaluate<\/code> method as well as the <code>train<\/code> method. <\/p>\n\n<p>Is there any advice as to how I can track down the route cause of this crash in the training jobs? Or what some of the common causes of this crash are? Is the solution just to use a larger memory machine?<\/p>\n","answers":[]}
+{"QuestionId":53517618,"AnswerCount":0,"Tags":"<tensorflow><installation>","CreationDate":"2018-11-28T10:45:58.973","AcceptedAnswerId":null,"OwnerUserId":2194381.0,"Title":"Install tensorflow in editable mode","Body":"<p>I would like to contribute to the project tensorflow. To do so, I would like to install tensorflow in editable mode. Does someone know how to do it ?<\/p>\n\n<p>Many thanks !<\/p>\n","answers":[]}
+{"QuestionId":53517661,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2018-11-28T10:47:56.373","AcceptedAnswerId":null,"OwnerUserId":10092730.0,"Title":"Python Keras: Overfitting a Keras MLP although loss curve doesn't look overfitted?","Body":"<p>I'm using Keras and Python to train a MLP Sequential model for classification of two classes. My Training Data has 247 features and I've got 17 Samples of class 1, 922 Samples of class 2. I use Smote Borderline Oversampling Algorithm to balance the dataset. I use Cross Validation with k=4 to validate the performance of precision and recall. For Training on each fold I plot the loss curve over training and validation, to estimate if the model is under- or overfitted.<\/p>\n\n<p>I trained a model with 3 hidden layers, and reached 95% of precision and 71% for recall. The loss function plot for each fold doesn't seem to be overfitted. (I'm not allowed by stackoverflow to post the image). <strong>But the evaluation of this model is worse, than with a model whose training precision and recall is worse.<\/strong><\/p>\n\n<p><strong>Is this overfitting? And how can I detect it before evaluation?<\/strong><\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"53518201","CreationDate":"2018-11-28T11:16:32.520","ParentId":null,"OwnerUserId":"10586790","Title":null,"Body":"<p>You can't detect overfitting simply from the learning curve. The definition of overfitting is when your model does extremely well on the training set, and poorly on your evaluation set, which is exactly what you're reporting. \nIn this case, I suspect the main problem is the unbalanced dataset. You can verify the spread of both classes in each of your sets (training, validation folds and test set) and see how your model performs on the minority class.<\/p>\n"}]}
+{"QuestionId":53517730,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-28T10:51:19.710","AcceptedAnswerId":null,"OwnerUserId":10716620.0,"Title":"How to store images for Tensorflow and Keras?","Body":"<p>I have about 100,000 images of 10 classes. Reading these image everytime is a time consuming.I want to store these image as CIFAR10 or any other dataset so I can read them quickly and fit them into models easily.<\/p>\n","answers":[{"AnswerId":"53517846","CreationDate":"2018-11-28T10:57:35.250","ParentId":null,"OwnerUserId":"10330116","Title":null,"Body":"<p>You can convert images into binary format which is understandable by tensorflow called \"tfrecord\" format. <\/p>\n\n<p>Please refer to the below links<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/datasets<\/a>\n<a href=\"http:\/\/www.machinelearninguru.com\/deep_learning\/tensorflow\/basics\/tfrecord\/tfrecord.html\" rel=\"nofollow noreferrer\">http:\/\/www.machinelearninguru.com\/deep_learning\/tensorflow\/basics\/tfrecord\/tfrecord.html<\/a><\/p>\n"},{"AnswerId":"53519929","CreationDate":"2018-11-28T12:52:04.323","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>There are multiple ways of doing it.<\/p>\n\n<p>You can store images in ten folders with class names as their name, You can store ten numpy arrays (<code>numpy.save()<\/code> and <code>numpy.load()<\/code>) containing images for each class. You can store one numpy array containing all of your images. You can use tfrecords which is mentioned by @Rajeswari <\/p>\n\n<p>fastest option is tfrecords tfrecord file, it can also be saved and read as multiple shards which helps if you don't have enough memory to hold all of the data in memory at once, worst option is keeping them as image files. Saving a  numpy array is easier than creating tfrecords while still being fast.<\/p>\n"}]}
+{"QuestionId":53517857,"AnswerCount":2,"Tags":"<ubuntu><tensorflow><tensorboard>","CreationDate":"2018-11-28T10:57:56.167","AcceptedAnswerId":56537286.0,"OwnerUserId":2466407.0,"Title":"Can I delete events.out.tfevents.XXXXXXXXXX.computer_name files from training folder","Body":"<p>I am training faster_rcnn_inception module for object detection on custom dataset.\nIn training directory,we found folder called eval_0 and tensorflow generated events.out.tfevents.xxxxxx files.<\/p>\n\n<p><strong>Training Directory structure as follows<\/strong><\/p>\n\n<pre><code>+training_dir\n    +eval_0\n     -events.out.tfevents.1542309785.instance-1  1.2GB\n     -events.out.tfevents.1542367255.instance-1  5.3GB\n     -events.out.tfevents.1542369886.instance-1  3.6GB\n     -events.out.tfevents.1542624154.instance-1  31MB\n     -events.out.tfevents.1543060258.instance-1  19MB\n     -events.out.tfevents.1543066775.instance-2  1.6GB\n -events.out.tfevents.1542308099.instance-1  17MB\n -events.out.tfevents.1542308928.instance-1  17MB\n -events.out.tfevents.1542366369.instance-1  17MB\n -events.out.tfevents.1542369000.instance-1  17MB\n -events.out.tfevents.1542623262.instance-1  17MB\n -events.out.tfevents.1543064936.instance-2  17MB\n -events.out.tfevents.1543065796.instance-2  17MB\n -events.out.tfevents.1543065880.instance-2  17MB\n -model.ckpt-96004.data-00000-of-00001\n -model.ckpt-96004.data-00000-of-00001\n -model.ckpt-96004.index\n -model.ckpt-96004.meta\n -model.ckpt-96108.data-00000-of-00001\n -model.ckpt-96108.index\n -model.ckpt-96108.meta\n<\/code><\/pre>\n\n<p>As per my understanding, <strong>tfevents<\/strong> files in eval_0 folder are summery files of <strong><em>evaluation<\/em><\/strong> and tfevents files in <strong>training_dir<\/strong> are summery files of <strong><em>training<\/em><\/strong>.<\/p>\n\n<p>I have interrupted training process several times and continued from recent checkpoint.\nI also understand restarting training process generates new tfevents files.<\/p>\n\n<p><strong>My Questions as follows:<\/strong><\/p>\n\n<ul>\n<li><p>Why training <strong>tfevents_files<\/strong> have same size, but in case if\n<strong>eval_0\/tfevents_files size varies<\/strong> ?<\/p><\/li>\n<li><p>Why interrupting training generates new <strong>tfevents_file<\/strong> in training\nfolder, but same not observed in case of <strong>eval_0<\/strong>?<\/p><\/li>\n<li><p>Can I delete all <strong>tfevents<\/strong> files in <strong>eval_0<\/strong> except latest one? Does it\naffect on training or evolution history?<\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"56537286","CreationDate":"2019-06-11T05:52:15.450","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p><strong>tfevents<\/strong> files are not essential for training and can be safely removed.<\/p>\n\n<p>In Tensorflow <strong>tfevents<\/strong> are created by FileWriters and are generally used to store summary output. Here are some common examples of how tf.summaries are used:<\/p>\n\n<ul>\n<li>storing a description of the tensorflow graph before training starts<\/li>\n<li>writing a value of the loss function for every training step<\/li>\n<li>storing a histogram of activations or weights for a layer once per epoch<\/li>\n<li>storing an example of output image of the network once on every validation<\/li>\n<li>storing average precision (or any other metric) for the whole validation set<\/li>\n<\/ul>\n\n<p>This information is not essential for training and can therefore be deleted. Yet, it might come in handy for debugging or studying behavior of the model. TensorBoard is the most common tool to read and visualize data stored in tfevent files. Anyone can read and interpret TFRecord files manually using protobuf protocol and it's implementation for Python, C++ and other.<\/p>\n\n<p>tfevents are written in TFRecord format. <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/tf_records\" rel=\"nofollow noreferrer\">TFRecord<\/a> is a simple format for storing a sequence of binary records. Tensorflow always appends new events\/summaries to the end of the file if file already exists. This explains file grows.<\/p>\n\n<p>Due to details of implementation of optimization routine provided with <em>tensorflow\/models\/reserach\/object_detection<\/em> training and evaluation event files have different behaviour. Namely, evaluation event file is created using a FileWriter directly, which will reuse latest existing event file in the log_dir whenever one exists. Implementation also has large number of summaries that are collected regularly, which increases event file during training. <\/p>\n\n<p>For the training routine, on the other hand, developers explicitly specify an empty list of summaries when training is done on <a href=\"https:\/\/en.wikipedia.org\/wiki\/Tensor_processing_unit\" rel=\"nofollow noreferrer\">TPU<\/a>. Which means that event file is created once and is never used afterwards. This behaviour can be different when training is performed on non-TPU hardware or <em>summarize_gradients<\/em> option is enabled for training.<\/p>\n"},{"AnswerId":"56536535","CreationDate":"2019-06-11T04:33:06.607","ParentId":null,"OwnerUserId":"2699907","Title":null,"Body":"<p>TFEvent files are mainly used by TensorBoard. If you open a terminal and start it (ie: <code>tensorboard --logdir .<\/code>) what you see is found in these event files.<\/p>\n\n<p>Of course, you can have multiple \"summary writers\". In your case, events logged during training are logged in the root \"training_dir\" while the ones from the eval phase are put under \"eval_0\". You would want to do this because tensorboard plots each folder as a separate group in charts.<\/p>\n\n<p>Your training data is different from your eval data, so the event files will of course be different as well.<\/p>\n\n<p>As for checkpoints, all you need are the <code>model.ckpt*<\/code> files to restore the weights. The event files are not used at all, so you can safely delete them. Actually, you probably want to start with a clean log folder whenever you start a new training process if you are actually planning to use tensorboard.<\/p>\n\n<p>The event files are not actually part of the checkpoints, they are log files. As such, whenever a logging method is called, new entries will be added. You don't see new entries in the eval_0 folder probably because you stopped the process during the training phase, before the evaluation phase was reached.<\/p>\n"}]}
+{"QuestionId":53518414,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2018-11-28T11:28:01.597","AcceptedAnswerId":null,"OwnerUserId":10716809.0,"Title":"Install tensorflow on pycharm","Body":"<p>I'm using Pycharm 2.5, but I can't install tensorflow on Pycharm, this is the error.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Y6BiB.png\" \/><\/p>\n\n<p>What should I do now ?<\/p>\n","answers":[]}
+{"QuestionId":53518437,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><lstm><rnn>","CreationDate":"2018-11-28T11:29:17.867","AcceptedAnswerId":null,"OwnerUserId":10716516.0,"Title":"How do I initialize LSTM weights from Numpy arrays correctly in Tensorflow?","Body":"<p>I have the same question for <code>tf.contrib.rnn.LSTMBlockCell<\/code> and <code>tf.contrib.cudnn_rnn.CudnnCompatibleLSTMCell<\/code>:<\/p>\n\n<p>How do I initialize the LSTM weights from numpy arrays correctly? The following code-snipped executes, but does not seem\nto do what I am looking for:<\/p>\n\n<pre><code>train_data = np.load('mnist_train_data.npy').reshape(-1,28,28)\ntrain_label = np.load('mnist_train_label.npy')\n\nparams = [np.random.randn(28+128, 4*128), np.zeros(4*128)]\n\nX = tf.placeholder(tf.float32, shape=[54999, 28, 28])\ny = tf.placeholder(tf.int64, None)\nstate = LSTMStateTuple(*(tf.zeros((54999, 128), dtype=tf.float32) for _ in range(2)))\n\ncell = tf.contrib.rnn.LSTMBlockCell(128)\ncell.build(tf.TensorShape((None, 28)))\ncell.set_weights(params)\n\ninitial_weights = cell.get_weights()\n\nprint(np.array_equal(params[0], initial_weights[0]))\n\nw1 = tf.Variable(np.random.randn(128, 10), dtype=tf.float32)\nb1 = tf.Variable(np.zeros(10), dtype=tf.float32)\n\nfull_seq, current_state = tf.nn.dynamic_rnn(cell, X, initial_state=state, dtype=tf.float32)\noutput = tf.matmul(current_state[1], w1)\noutput += b1\nloss = tf.losses.softmax_cross_entropy(y, output)\n\ntrain_step = tf.train.AdamOptimizer(0.01).minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(1):\n        feed_dict = {X: train_data, y: train_label}\n        sess.run(train_step, feed_dict=feed_dict)\n\nfinal_weights = cell.get_weights()\n\nprint(np.array_equal(initial_weights[0], final_weights[0]))\n<\/code><\/pre>\n\n<p>This prints out  <code>False<\/code> in the first print-statement, so the numpy arrays do not actually seem to be used as weights. <\/p>\n\n<p>Moreover, after the training session, this prints out <code>True<\/code> thus implying, that these weights are not actually updated during training.<\/p>\n\n<p>Thanks in advance for any help on the subject.<\/p>\n","answers":[]}
+{"QuestionId":53518977,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-28T11:59:12.267","AcceptedAnswerId":null,"OwnerUserId":9209379.0,"Title":"How to freeze TensorFlow ckpt with queuerunner?","Body":"<p>I used <code>tf.train.shuffle_batch<\/code> while training, and do not use any placeholder, when I freezed the ckpt to pb file, I did not get any input tensor, though I can feed thing to <code>shuffle_batch<\/code> tensor, but it need the feed has the same size with the <code>shuffle_batch<\/code> data. How to fix it? I know I can rewrite a net and restore params, then freeze, but it is not wise?<\/p>\n\n<p>The train <\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorlayer as tl\nfrom tensorlayer.layers import *\nimport os\ntrainsize=35680\ntestsize=889\n\nbatch_size=32\ninputW=224\ninputH=480\nTRAIN_TFRECORD='.\/train.tfrecords'\nTEST_TFRECORD='.\/test.tfrecords'\nBATCH_CAPACITY=512\nMIN_AFTER_DEQU=256\nMAX_Cycle=100000\nTRAIN_CYCLE=int(trainsize\/batch_size)\nTEST_CYCLE=int(testsize\/batch_size)\nlearning_rt = 0.001\nsavepath='.\/ckpt\/'\nlogpath='.\/logs\/'\n\n\ndef network(inputs,is_train,reuse):\n    BITW =8\n    BITA=8\n    Decay=0.99\n    Epsi=1e-5\n    with tf.variable_scope('Model',reuse=reuse):\n        net=InputLayer(inputs,name='input') #224*480\n\n        net=QuanConv2dWithBN(net,32,(3,3),(1,1),'SAME',tf.nn.relu, decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv1_1')\n\n        net=QuanConv2dWithBN(net,64,(3,3),(2,2),'SAME',tf.nn.relu, decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv1_2') #112*240\n\n        net=QuanConv2dWithBN(net,64,(3,3),(1,1),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv2_1')\n        net=QuanConv2dWithBN(net,128,(3,3),(2,2),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv2_2') #56*120\n\n        net=QuanConv2dWithBN(net,128,(3,3),(1,1),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv3_1')\n\n        net=QuanConv2dWithBN(net,64,(1,1),(1,1),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv3_2')\n\n        net=QuanConv2dWithBN(net,128,(3,3),(2,2),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv3_3') #28*60\n\n        net=QuanConv2dWithBN(net,64,(3,3),(1,1),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv4_1')\n\n        net=QuanConv2dWithBN(net,96,(3,3),(2,2),'VALID',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv4_2') #14*30\n\n        print(net.outputs)\n        net=QuanConv2dWithBN(net,128,(3,3),(2,2),'SAME',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv5_1') #7*30\n\n        net=QuanConv2dWithBN(net,128,(3,3),(1,2),'VALID',tf.nn.relu,decay=Decay, epsilon=Epsi,  is_train=is_train,  bitW=BITW,bitA=BITA,name='Conv5_2') #3*30\n\n        net=QuanConv2d(net,128,(3,3),(1,2),'VALID',tf.nn.leaky_relu,bitW=BITW,bitA=BITA,name='Conv5_3') #1*30\n        print(net.outputs)\n        net=FlattenLayer(net,name='flat1')\n        net=QuanDenseLayer(net,128,act=tf.nn.leaky_relu,bitW=BITW,bitA=BITA,name='dense1')\n\n        net=DropoutLayer(net,0.5,is_fix=True,is_train=is_train,name='drop1')\n        net=DenseLayer(net,1,name='dense2')\n        outnet=net\n        volcume=net.outputs\n        print(volcume)\n        return outnet,net.outputs,volcume\n\n\ndef inference(inputs,is_train,reuse):\n    return network(inputs,is_train,reuse)\n\n\n\ndef read_and_decode(filename):\n    filename_queue = tf.train.string_input_producer([filename])\n    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(serialized_example,\n                                       features={\n                                           'img': tf.FixedLenFeature([], tf.string),\n                                           'num' : tf.FixedLenFeature([], tf.float32),\n                                       })\n\n    img = tf.decode_raw(features['img'], tf.uint8)\n    img = tf.reshape(img, [480, 240, 3])\n    img=tf.random_crop(img,[480,224,3])\n    img = tf.image.random_brightness(img, max_delta=0.3)\n    img = tf.image.random_contrast(img, lower=0.1, upper=0.5)\n    # img = tf.image.random_hue(img, max_delta=0.1)\n    # img = tf.image.random_saturation(img, lower=0, upper=2.5)\n    img = tf.image.per_image_standardization(img)\n    label = tf.reshape( tf.cast(features['num'], tf.float32)*(1.\/230.)-0.5,[1])\n    return img, label\n\ndef read_and_decode_test(filename):\n    filename_queue = tf.train.string_input_producer([filename])\n    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(serialized_example,\n                                       features={\n                                           'img': tf.FixedLenFeature([], tf.string),\n                                           'num' : tf.FixedLenFeature([], tf.float32),\n                                       })\n\n    img = tf.decode_raw(features['img'], tf.uint8)\n    img = tf.reshape(img, [480, 240, 3])\n    img=img[:,8:232,:]\n    img = tf.image.per_image_standardization(img)\n    label = tf.reshape( tf.cast(features['num'], tf.float32)*(1.\/230.)-0.5,[1])\n    return img, label\n\n\ndef smooth_L1(x):\n    return tf.where(tf.less_equal(tf.abs(x), 1.0), tf.multiply(0.5, tf.pow(x, 2.0)), tf.subtract(tf.abs(x), 0.5))\n\ndef cal_loss(logits,labels):\n    # return tf.clip_by_value(tf.reduce_mean(tf.losses.mean_squared_error(labels,logits)) ,0.000001,10000000.)\n    return tf.reduce_mean(tf.where(tf.less_equal(tf.abs(logits-labels), 0.02),0.00001*tf.ones_like(logits-labels),  tf.multiply(1., tf.pow(logits-labels, 2.0))))\n    # return tf.clip_by_value(tf.reduce_sum(smooth_L1(labels-logits)),0.0000001,100.)\n\ndef cal_acc(logits,labels):\n    return tf.reduce_mean( tf.cast( tf.less_equal(tf.abs(labels-logits),tf.ones_like(labels)*.1),tf.float32))\n\nif __name__ == '__main__':\n\n    img_train,num_train  = read_and_decode(TRAIN_TFRECORD)\n    img_test,num_test  = read_and_decode(TEST_TFRECORD)\n    img_train_batch,  num_train_batch = tf.train.shuffle_batch(\n        [img_train, num_train], batch_size=batch_size, capacity=BATCH_CAPACITY,\n        min_after_dequeue=MIN_AFTER_DEQU)\n    img_test_batch,  num_test_batch = tf.train.batch(\n        [img_test,num_test], batch_size=batch_size)\n    net,_,logits_train=inference(img_train_batch,True,None)\n    _,_,logits_test=inference(img_test_batch,False,True)\n    loss_train=cal_loss(logits_train,num_train_batch)\n    loss_test=cal_loss(logits_test,num_test_batch)\n    acc_test=cal_acc(logits_test,num_test_batch)\n    acc_train=cal_acc(logits_train,num_train_batch)\n    global_step=tf.train.create_global_step()\n    #tf.train.get_global_step()\n    learning_rate=tf.train.exponential_decay(learning_rt, global_step,\n                                           5000, 0.9, staircase=True)\n    train = tf.train.MomentumOptimizer(learning_rate,momentum=0.9).minimize(loss_train,global_step=global_step)\n    # train = tf.train.AdamOptimizer(learning_rt).minimize(loss_train)\n\n\n    tf.summary.scalar('loss_train', loss_train)\n    tf.summary.scalar('acc_train', acc_train)\n    merged = tf.summary.merge_all()\n    with tf.Session(config=tf.ConfigProto()) as sess:\n        trainwrite = tf.summary.FileWriter(logpath, sess.graph)\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n        sess.run(tf.global_variables_initializer())\n        saver = tf.train.Saver()\n        run_cycle=0\n        if os.path.exists(savepath+'313.ckpt.index') :\n            print('\\nStart Restore')\n            saver.restore(sess, savepath+'33.ckpt')\n            print('\\nEnd Restore')\n        print('\\nStart Training')\n        try:\n\n            while not coord.should_stop():\n\n                while run_cycle &lt; MAX_Cycle:\n\n                    run_cycle+=1\n                    # if run_cycle%10==0:\n                    #     learning_rt*=0.6\n                    # if run_cycle%200==0:\n                    #     learning_rt*=2.\n                    l_tall=0\n                    a_tall=0\n\n                    l_teall=0\n                    a_teall=0\n\n                    for train_c in range(TRAIN_CYCLE):\n\n\n                        _,l_train,a_train=sess.run([train,loss_train,acc_train])\n                        l_tall+=l_train\n                        a_tall+=a_train\n\n                        if (train_c+1)%100==0:\n                            print('train_loss:%f'%(l_tall\/100.))\n                            print('train_acc:%f'%(a_tall\/100.))\n                            l_tall = 0\n                            a_tall = 0\n                        if (train_c+1)%500==0:\n                            print('Global Step:',sess.run(global_step))\n                            result_merged=sess.run(merged)\n                            trainwrite.add_summary(result_merged, run_cycle*TRAIN_CYCLE+train_c)\n                    for test_c in range(TEST_CYCLE):\n                        l_test,a_test=sess.run([loss_test,acc_test])\n                        l_teall+=l_test\n                        a_teall+=a_test\n                        if (test_c+1)%TEST_CYCLE==0:\n                            print('------------------')\n                            print('test_loss:%f'%(l_teall\/TEST_CYCLE))\n                            print('test_acc:%f'%(a_teall\/TEST_CYCLE))\n                            print('------------------')\n                            l_teall = 0\n                            l_teall = 0\n                    saver.save(sess, savepath+ str(run_cycle) + '.ckpt')\n\n\n\n\n        except tf.errors.OutOfRangeError:\n            print('Done training!!!')\n        finally:\n            # When done, ask the threads to stop.\n            coord.request_stop()\n\n        coord.join(threads)\n        sess.close()\n<\/code><\/pre>\n\n<p>Freeze code<\/p>\n\n<p>import os, argparse<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework import graph_util\n\ndir = os.path.dirname(os.path.realpath(__file__))\n\n\ndef freeze_graph(model_folder, output_nodes='y_hat',\n                 output_filename='frozen-graph.pb',\n                 rename_outputs=None):\n    # Load checkpoint\n    checkpoint = tf.train.get_checkpoint_state(model_folder)\n    input_checkpoint = checkpoint.model_checkpoint_path\n\n    output_graph = output_filename\n\n    # Devices should be cleared to allow Tensorflow to control placement of\n    # graph when loading on different machines\n    saver = tf.train.import_meta_graph(input_checkpoint + '.meta',\n                                       clear_devices=True)\n\n    graph = tf.get_default_graph()\n    onames = output_nodes.split(',')\n\n    # https:\/\/stackoverflow.com\/a\/34399966\/4190475\n    if rename_outputs is not None:\n        nnames = rename_outputs.split(',')\n        with graph.as_default():\n            for o, n in zip(onames, nnames):\n                _out = tf.identity(graph.get_tensor_by_name(o + ':0'), name=n)\n            onames = nnames\n\n    input_graph_def = graph.as_graph_def()\n\n    # fix batch norm nodes\n    for node in input_graph_def.node:\n\n        if node.op == 'RefSwitch':\n            node.op = 'Switch'\n            for index in range(len(node.input)):\n                if 'moving_' in node.input[index]:\n                    node.input[index] = node.input[index] + '\/read'\n        elif node.op == 'AssignSub':\n            node.op = 'Sub'\n            if 'use_locking' in node.attr: del node.attr['use_locking']\n\n    with tf.Session(graph=graph) as sess:\n        saver.restore(sess, input_checkpoint)\n\n        # In production, graph weights no longer need to be updated\n        # graph_util provides utility to change all variables to constants\n        output_graph_def = graph_util.convert_variables_to_constants(\n            sess, input_graph_def,\n            onames  # unrelated nodes will be discarded\n        )\n\n        # Serialize and write to file\n        with tf.gfile.GFile(output_graph, \"wb\") as f:\n            f.write(output_graph_def.SerializeToString())\n        print(\"%d ops in the final graph.\" % len(output_graph_def.node))\n\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser(\n        description='Prune and freeze weights from checkpoints into production models')\n    parser.add_argument(\"--checkpoint_path\",\n                        default='.\/regressionDir\/',\n                        type=str, help=\"Path to checkpoint files\")\n    parser.add_argument(\"--output_nodes\",\n                        default='Model\/dense2\/bias_add',\n                        type=str, help=\"Names of output node, comma seperated\")\n    parser.add_argument(\"--output_graph\",\n                        default='reg.pb',\n                        type=str, help=\"Output graph filename\")\n    parser.add_argument(\"--rename_outputs\",\n                        default='out_vol',\n                        type=str, help=\"Rename output nodes for better \\\n                            readability in production graph, to be specified in \\\n                            the same order as output_nodes\")\n    args = parser.parse_args()\n\n    freeze_graph(args.checkpoint_path, args.output_nodes, args.output_graph, args.rename_outputs)\n<\/code><\/pre>\n\n<p>Test inference code<\/p>\n\n<p>import tensorflow as tf\n    import numpy as np\n    from PIL import Image\n    import time<\/p>\n\n<pre><code>gf = tf.GraphDef()\ngf.ParseFromString(open('reg.pb', 'rb').read())\nprint([n.name + '=&gt;' + n.op for n in gf.node])\n\noutput_graph_path = '.\/reg.pb'\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    output_graph_def = tf.GraphDef()\n    with open(output_graph_path, \"rb\") as f:\n        output_graph_def.ParseFromString(f.read())\n        _ = tf.import_graph_def(output_graph_def, name=\"\")\n\n    input_img = sess.graph.get_tensor_by_name(\"shuffle_batch:0\")\n    print(input_img)\n    out_vol = sess.graph.get_tensor_by_name(\"out_vol:0\")\n    out_voln = sess.graph.get_tensor_by_name(\"shuffle_batch:0\")\n    a = np.random.random([48, 480, 224, 3])\n    for x in range(100):\n        ntime1 = time.time()\n        vol = sess.run(out_vol, {input_img: a})\n        ntime2 = time.time()\n        print(ntime2 - ntime1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53520576,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2018-11-28T13:28:47.820","AcceptedAnswerId":null,"OwnerUserId":5378310.0,"Title":"Is it possible to feed dynamic shape matrix in Tensorflow?","Body":"<p>I have data set with unstatic shape like <code>(batch_size, None, None, None, 92)<\/code><\/p>\n\n<pre><code>for x in x_data:\n    print(x.shape)\n(4, 4, 8, 92)\n(3, 3, 7, 92)\n(4, 4, 8, 92)\n(3, 3, 7, 92)\n(4, 4, 8, 92)\n(4, 4, 7, 92)\n(3, 3, 7, 92)\n(4, 4, 8, 92)\n(4, 4, 8, 92)\n(3, 3, 8, 92)\n<\/code><\/pre>\n\n<p>But when I try to feed this <strong>x_data<\/strong> to my <strong>X place holder<\/strong> I faced error<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32, [None, None, None, None, 92])\nwith tf.Session() as sess:\n    c, _ = sess.run([cost, optimizer], feed_dict={X: x_data, Y: y_data})\n<\/code><\/pre>\n\n<p>This error may caused from unstatic shape of input data.\n<br>Error message<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/bsjun\/Documents\/GitHub\/CCpyNN\/CCpyNN\/Inception_v.2.py\", line 274, in &lt;module&gt;\n    c, hy, _ = sess.run([cost, logit_layer, optimizer], feed_dict={X: batch_x[i], Y: batch_y[i], keep_prob: 0.8})\n  File \"C:\\Users\\bsjun\\AppData\\Local\\conda\\conda\\envs\\tf_normal\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"C:\\Users\\bsjun\\AppData\\Local\\conda\\conda\\envs\\tf_normal\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1121, in _run\n    np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)\n  File \"C:\\Users\\bsjun\\AppData\\Local\\conda\\conda\\envs\\tf_normal\\lib\\site-packages\\numpy\\core\\numeric.py\", line 501, in asarray\n    return array(a, dtype, copy=False, order=order)\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>Is it impossible to feed dynamic shape matrix ?<\/p>\n\n<p>Here is a summary of my code.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nshapes = [(4, 4, 8, 92), (3, 3, 7, 92), (4, 4, 8, 92), (3, 3, 7, 92)]\nx_data = []\nfor s in shapes:\n    x = np.zeros(shape=s)\n    print(x.shape)\n    x_data.append(x)\n\nX = tf.placeholder(tf.float32, [None, None, None, None, 92])\nwith tf.Session() as sess:\n    sess.run(X, feed_dict={X: x_data})\n<\/code><\/pre>\n","answers":[{"AnswerId":"53520933","CreationDate":"2018-11-28T13:48:18.603","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>It won't work for the data which you provide, but there are some options how to deal with that.<\/p>\n\n<p><strong>Why it doesn't work<\/strong><\/p>\n\n<p>The line<\/p>\n\n<pre><code> x = tf.placeholder(tf.float32, [None, None, None, 92])\n<\/code><\/pre>\n\n<p>means that the shape of the input array is not known beforehand. But it must be an object which could be converted numpy array. As your input data is a sequence of numpy arrays of different shapes, it won't be converted.<\/p>\n\n<p><strong>How to deal with that<\/strong><\/p>\n\n<p><strong>1. Provide separate input for your model.<\/strong> You could, probably, modify the code of your model, providing it two inputs:<\/p>\n\n<ul>\n<li>Input data with the \"maximum\" shape. That means, that any array from your sequence would fit into this shape;<\/li>\n<li>Actual shapes of the input arrays.<\/li>\n<\/ul>\n\n<p>Such approach is used, for example, in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"nofollow noreferrer\">tf.nn.dynamic_rnn()<\/a>, where one of the parameters is the actual data, and another is <code>sequence_length<\/code> - the length of each sequence.<\/p>\n\n<p><strong>2. Vary shape of data from batch to batch.<\/strong>  Another option would be to feed arrays of different shapes on each batch. So, for example, you group <code>batch_size<\/code> arrays of shape (4, 4, 8, 92) into one batch and feed it into the model. Then you take <code>batch_size<\/code> arrays of shape (3, 3, 8, 92) and do one more pass and so on. So, the shape can vary in the dataset, but it should be constant in the single batch.<\/p>\n"}]}
+{"QuestionId":53521149,"AnswerCount":0,"Tags":"<python><c++><tensorflow><keras>","CreationDate":"2018-11-28T13:58:41.257","AcceptedAnswerId":null,"OwnerUserId":10632819.0,"Title":"Invalid argument: specified in either feed_devices or fetch_devices was not found in the Graph","Body":"<p>I am trying to run a model exported from Keras in Tensorflow using the C++ API. The model is exported and imported correctly, however when I try to run the session in C++, the names of the variables I pass in the <code>Run<\/code> member of <code>tensorflow::Session<\/code> get modified, causing the error:<\/p>\n\n<pre><code>Invalid argument: Tensor Output0:0, specified in either feed_devices or fetch_devices was not found in the Graph\n<\/code><\/pre>\n\n<p>Here is the model I have built in Keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>S = Input(shape=[state_size], name='Input')\nh0 = Dense(self.HIDDEN1_UNITS, name='Hidden1', kernel_initializer=keras.initializers.RandomUniform())(S)\nh0_a = Activation('relu')(h0)\nh1 = Dense(self.HIDDEN2_UNITS, name='Hidden2', kernel_initializer=keras.initializers.RandomUniform())(h0_a)\nh1_a = Activation('relu')(h1)\noutput = Dense(1)(h1_a)\noutput_a = Activation('tanh', name='Output0')(output)\nmodel = Model(S,output_a)\n<\/code><\/pre>\n\n<p>The model is exported using <code>tensorflow.python.framework.graph_util<\/code>\nand <code>tensorflow.python.framework.graph_io<\/code>, following the approach implemented in <a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"noreferrer\">https:\/\/github.com\/amir-abdi\/keras_to_tensorflow<\/a><\/p>\n\n<p>The model is then imported in Tensorflow using the C++ API:<\/p>\n\n<pre class=\"lang-cc prettyprint-override\"><code>ParseProtoUnlimited(&amp;graph_def, &lt;graph_protoprotobuffer&gt;, &lt;graph_size&gt; );\n<\/code><\/pre>\n\n<p>The graph seems to be imported successfully, in fact when I print the nodes name contained in the graph in Tensorflow I get the following:<\/p>\n\n<pre><code>Names : Input_5\nNames : Hidden1_5\/kernel\nNames : Hidden1_5\/kernel\/read\nNames : Hidden1_5\/bias\nNames : Hidden1_5\/bias\/read\nNames : Hidden1_5\/MatMul\nNames : Hidden1_5\/BiasAdd\nNames : activation_5_1\/Relu\nNames : Hidden2_5\/kernel\nNames : Hidden2_5\/kernel\/read\nNames : Hidden2_5\/bias\nNames : Hidden2_5\/bias\/read\nNames : Hidden2_5\/MatMul\nNames : Hidden2_5\/BiasAdd\nNames : activation_6_1\/Relu\nNames : dense_3_1\/kernel\nNames : dense_3_1\/kernel\/read\nNames : dense_3_1\/bias\nNames : dense_3_1\/bias\/read\nNames : dense_3_1\/MatMul\nNames : dense_3_1\/BiasAdd\nNames : Output_5\/Tanh\nNames : output0\n<\/code><\/pre>\n\n<p>Now, if I want to try and run the session, I do the following:<\/p>\n\n<pre class=\"lang-cc prettyprint-override\"><code>Tensor input(DT_FLOAT, TensorShape( { 4 }));\n...\n[init tensor]\n...\nsession-&gt;Run({{\"Input_5\", input}}, {\"Output0\"}, {},\n                &amp;outputs);\n<\/code><\/pre>\n\n<p>But the application throws the following error:<\/p>\n\n<pre><code>Invalid argument: Tensor Output0:0, specified in either feed_devices or fetch_devices was not found in the Graph\n<\/code><\/pre>\n\n<p>Why the function Run session appends the :0 to the output tensor I want to retrieve? <\/p>\n\n<p>Also, I have tried beforehand to do the above with a simple graph built with Tensorflow, and the session on C++ run with no problem. Here is the graph that worked:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:\n    a = tf.Variable(5.0, name='a')\n    b = tf.Variable(6.0, name='b')\n    c = tf.multiply(a, b, name=\"c\")\n    sess.run(tf.global_variables_initializer())\n    exported = sess.graph_def.SerializeToString()\n    size = sess.graph_def.ByteSize()\n<\/code><\/pre>\n\n<pre class=\"lang-cc prettyprint-override\"><code>Tensor a(DT_FLOAT, TensorShape());\na.scalar&lt;float&gt;()() = 5.0;\n\nTensor b(DT_FLOAT, TensorShape());\nb.scalar&lt;float&gt;()() = 6.0;\n\nstd::vector&lt;std::pair&lt;string, tensorflow::Tensor&gt;&gt; inputs = {\n                { \"a\", a }, { \"b\", b }, };\n\nstd::vector&lt;tensorflow::Tensor&gt; outputs;\nRun the session, evaluating our \"c\" operation from the graph\nstatus = session-&gt;Run(inputs, { \"c\" }, { }, &amp;outputs);\n<\/code><\/pre>\n\n<p>Why in the second case I manage to run the session whereas in the first case I get that error?<\/p>\n\n<p>Keras version is <code>2.2.2<\/code>\nTensorflow version is <code>1.11.0<\/code><\/p>\n\n<p>Any help would be much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":53521339,"AnswerCount":1,"Tags":"<tensorflow><bazel><avx2><avx512><intel-tensorflow>","CreationDate":"2018-11-28T14:08:04.907","AcceptedAnswerId":53625888.0,"OwnerUserId":10658634.0,"Title":"build tensorflow for intel xeon gold 6148","Body":"<p>I have a server with two Intel xeon gold 6148 and tensorflow running on it.\nWhen I install tf with pip I get a message that AVX2 and AVX512 is not used with my installation.\nSo, to get the best performance I tried to build tf from source using docker.\nI did so following <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/source<\/a> but for the bazel build command I used: <br\/><\/p>\n\n<pre><code>bazel build --config=mkl -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mavx512f --copt=-mavx512pf --copt=-mavx512cd --copt=-mavx512er \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>following <a href=\"https:\/\/software.intel.com\/en-us\/articles\/intel-optimization-for-tensorflow-installation-guide\" rel=\"nofollow noreferrer\">https:\/\/software.intel.com\/en-us\/articles\/intel-optimization-for-tensorflow-installation-guide<\/a>. <br\/>\nBut this installation performs much worse than the standard pip installation.<\/p>\n\n<p>So, to sum this up: what is the best way to install tensorflow an xeon gold architecture?<\/p>\n","answers":[{"AnswerId":"53625888","CreationDate":"2018-12-05T05:45:47.457","ParentId":null,"OwnerUserId":"9465768","Title":null,"Body":"<p>Try out using tensorflow wheel from the Intel website.<\/p>\n\n<p>Please follow the below steps to create your own conda environment and to install the wheel.<\/p>\n\n<p>1) conda create -n ENV_NAME -c intel python=3.6 pip numpy<\/p>\n\n<p>2) source activate ENV_NAME <\/p>\n\n<p>3) pip install <a href=\"https:\/\/storage.googleapis.com\/intel-optimized-tensorflow\/tensorflow-1.11.0-cp36-cp36m-linux_x86_64.whl\" rel=\"noreferrer\">https:\/\/storage.googleapis.com\/intel-optimized-tensorflow\/tensorflow-1.11.0-cp36-cp36m-linux_x86_64.whl<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":53523669,"AnswerCount":0,"Tags":"<python><dataframe><machine-learning><pytorch><rnn>","CreationDate":"2018-11-28T16:09:41.463","AcceptedAnswerId":null,"OwnerUserId":10718077.0,"Title":"Pytorch RNN data from three different csv. Python 3","Body":"<p>i need to train a RNN on pythorch on features vectors taken from 3 different frames of the same object. The features dimensions are alway the same. <\/p>\n\n<p>I have a csv for first frames, csv for seconds frames, csv for third one.<\/p>\n\n<p>Need to take the same raw from different csvs in order to get the features about the specific frame. \n(raw 1 in csv1 is frame 1 of object x, raw 1 in csv2 is frame 2 of object x, raw 1 in csv3 is frame 3 of object x)<\/p>\n\n<p>I need to create a RNN that at state t0, t1, t2 take frame 1 features, frame 2 features and frame 3. <\/p>\n\n<p>I saw character examples and time prediction ones where for each state are fed a single character of a world or a single time value of the dataset.<\/p>\n\n<p>How can I organize the training in order to take a whole features vector to train that state, then the second feature and so on? <\/p>\n\n<p>thanks in advance<\/p>\n","answers":[]}
+{"QuestionId":53523729,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><lstm><rnn>","CreationDate":"2018-11-28T16:12:44.697","AcceptedAnswerId":53523873.0,"OwnerUserId":3484477.0,"Title":"How to select a proper API\/implementation of an LSTM cell in TensorFlow?","Body":"<p>In Tensorflow there are many different implementations of the same task under different APIs. For LSTM cell one can find many implementations, e.g.,<\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/rnn_cell\/LSTMCell#class_lstmcell\" rel=\"nofollow noreferrer\">tf.nn.rnn_cell.LSTMCell<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/LSTMBlockCell\" rel=\"nofollow noreferrer\">tf.contrib.rnn.LSTMBlockCell<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/LSTMCell\" rel=\"nofollow noreferrer\">tf.keras.layers.LSTMCell<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/LSTM\" rel=\"nofollow noreferrer\">tf.keras.layers.LSTM<\/a><\/li>\n<\/ol>\n\n<p>Which of the above implementation shall I select? Is there a general guideline? For example always use tf.nn > tf.keras > tf.layers > tf. contrib.<\/p>\n\n<p>In a slightly related question <a href=\"https:\/\/stackoverflow.com\/questions\/48001759\/what-is-right-batch-normalization-function-in-tensorflow\">here<\/a> (for batch normalization), the approved answer says tf.contrib is not a good choice as it is for early implementation. Therein, the KERAS API also uses tf.nn. So it looks like tf.nn > tf.keras > tf.contrib.<\/p>\n","answers":[{"AnswerId":"53523873","CreationDate":"2018-11-28T16:20:17.853","ParentId":null,"OwnerUserId":"3009130","Title":null,"Body":"<p>Here <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/overview#rnn_performance\" rel=\"nofollow noreferrer\">RNN Performance<\/a> are some recommendations for TensorFlow. A couple of important points from the doc:<\/p>\n\n<ul>\n<li>use <code>tf.contrib.cudnn_rnn()<\/code> on NVIDIA GPUs;<\/li>\n<li>use <code>tf.nn.dynamic_rnn()<\/code> instead of <code>tf.nn.static_rnn()<\/code>. Probably they mean, that we need to add <code>sequence_length<\/code> to avoid extra computations.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53523913,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T16:22:57.753","AcceptedAnswerId":null,"OwnerUserId":8106257.0,"Title":"classifying integer data by tensorflow","Body":"<p>I want to classify \nif input data is under 200 than output is (0, 1)\nand if input data is over 200 than output is (1, 0)<\/p>\n\n<p>input value is sequential integer value and layer is 5.<\/p>\n\n<p>hidden layer use sigmoid and last hidden layer use softmax function<\/p>\n\n<p>loss function is reduce_mean and training with gradient descendent<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ndef set_x_data():\n    x_data = np.array([[50]\n                     , [60]\n                     , [70]\n                     , [80]\n                     , [90]\n                     , [110]\n                     , [120]\n                     , [130]\n                     , [140]\n                     , [150]\n                     , [160]\n                     , [170]\n                     , [180]\n                     , [190]\n                     , [200]\n                     , [210]\n                     , [220]\n                     , [230]\n                     , [240]\n                     , [250]\n                     , [260]\n                     , [270]\n                     , [280]\n                     , [290]\n                     , [300]\n                     , [310]\n                     , [320]\n                     , [330]\n                     , [340]\n                     , [350]\n                     , [360]\n                     , [370]\n                     , [380]\n                     , [390]])\n\n    return x_data\n\ndef set_y_data(x):\n    y_data = np.array([[0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [0, 1]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]\n                     , [1, 0]])\n    return y_data\n\ndef set_bias(efficiency):\n    arr = np.array([efficiency])\n\n    return arr\n\nW1 = tf.Variable(tf.random_normal([1, 5]), name='weight1')\nW2 = tf.Variable(tf.random_normal([5, 5]), name='weight2')\nW3 = tf.Variable(tf.random_normal([5, 5]), name='weight3')\nW4 = tf.Variable(tf.random_normal([5, 5]), name='weight4')\nW5 = tf.Variable(tf.random_normal([5, 2]), name='weight5')\n\ndef inference(input, b):\n    hidden_layer1 = tf.sigmoid(tf.matmul(input, W1) + b)\n    hidden_layer2 = tf.sigmoid(tf.matmul(hidden_layer1, W2) + b)\n    hidden_layer3 = tf.sigmoid(tf.matmul(hidden_layer2, W3) + b)\n    hidden_layer4 = tf.sigmoid(tf.matmul(hidden_layer3, W4) + b)\n    out_layer = tf.nn.softmax(tf.matmul(hidden_layer4, W5) + b)\n\n    return out_layer\n\ndef loss(hypothesis, y):\n    cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(hypothesis), reduction_indices=[1]))\n\n    return cross_entropy\n\ndef train(loss):\n    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)\n    train = optimizer.minimize(loss)\n\n    return train\n\nx_data = set_x_data(1)\ny_data = set_y_data(0)\nb_data = set_bias(0.8)\n\nx= tf.placeholder(tf.float32, shape=[None, 1])\ny= tf.placeholder(tf.float32, shape=[None, 2])\nb = tf.placeholder(tf.float32, shape=[None])\n\nhypothesis = inference(x, b)\nloss = loss(hypothesis, y)\ntrain = train(loss)\n\nsess = tf.Session()\ninit = tf.global_variables_initializer()\nsess.run(init)\n\nprint(sess.run(W1))\n\nfor step in range(2000):\n    sess.run(train, feed_dict={x:x_data, y:y_data, b:b_data})\n\nprint(sess.run(W1))\nprint(sess.run(hypothesis, feed_dict={x:np.array([[1000]]), b:b_data}))\n<\/code><\/pre>\n\n<p>when I print W1 before training and after training, value doesn't change specially and testing when input = 1000, that value doesn't currect what I expect. I think value nearly close to (1, 0), but result is almost (0.5, 0.5)<\/p>\n\n<p>I guess that mistakes come from loss function because it was copied from here and there, but I can't be sure about it<\/p>\n\n<hr>\n\n<p>upper code is just simplified of my code but I think I have to show my real code<\/p>\n\n<p>the code is too long so I create new post<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/53790116\/classifying-data-by-tensorflow-but-accuracy-value-didnt-change\">classifying data by tensorflow but accuracy value didn&#39;t change<\/a><\/p>\n","answers":[{"AnswerId":"53527868","CreationDate":"2018-11-28T20:48:34.937","ParentId":null,"OwnerUserId":"5541673","Title":null,"Body":"<p>There are a few issues in the training of the above network, but with a few changes you can achieve a network that gets <a href=\"https:\/\/i.stack.imgur.com\/caLCi.png\" rel=\"nofollow noreferrer\">this decision function<\/a><\/p>\n\n<p>(<a href=\"https:\/\/i.stack.imgur.com\/caLCi.png\" rel=\"nofollow noreferrer\">The plot in the link<\/a> shows the score of class 2, i.e. if x > 200)<\/p>\n\n<p>The list of issues subject to improvement in this network: <\/p>\n\n<ul>\n<li><p><strong>The training data is very scarce<\/strong> (only 34 points!) This is typically too small, especially for a 5-layer network as in your case. You typically want many more input samples than parameters in the network. Try adding more input values and reducing the number of layers (as in the code below - I've used floats instead of integers to get more points, but I think it is still compatible). <\/p><\/li>\n<li><p><strong>The input ranges typically require scaling<\/strong> (below I've tried a super-simple scaling by dividing by a constant). This is because you typically want to avoid high ranges of variables (especially of you pass many layers with a soft-max non-linearity, this would destroy the information contained in the very high or very low values). In more advanced cases you might want to do Min-Max Scaling or z-scores. <\/p><\/li>\n<li><p><strong>Try more epochs<\/strong> (and try plotting the evolution of the loss function value). With the given number of epochs, the optimization of the loss function had not converged. Below I do 10x more epochs. See how the code below now almost converges in <a href=\"https:\/\/i.stack.imgur.com\/5OhZg.png\" rel=\"nofollow noreferrer\">this plot<\/a> (and see how 2000 epochs were not enough):<\/p><\/li>\n<li><p>Something that helped was <strong>shuffling the (x,y) data<\/strong>. Though this is not crucial in this case, it converges faster (see the paper \"<a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/lecun-98b.pdf\" rel=\"nofollow noreferrer\">Efficient Backprop<\/a>\" by Le Cun). And in more serious examples it is typically needed. <\/p><\/li>\n<li><p>Importantly, <strong>I think you want <code>b<\/code> to be a parameter<\/strong>, not a constant, don't you? The bias of a network is typically also optimized together with the multiplicative weights. (Also, it is not common to use a single, shared bias for all the hidden layers. )<\/p><\/li>\n<\/ul>\n\n<p>Below is the code. Note there might be further improvements but these few tricks end up with the desired decision function. <\/p>\n\n<p>I've added some inline comments to indicate changes with respect to the original. I hope you find these pieces of advice insightful! <\/p>\n\n<p>The code: <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\n# I've modified the functions set_x_data and set_y_data\n# so as to generate a larger set of numbers. \n\n# Generate a range of numbers from 50 to 390\ndef set_x_data():\n    x_data = np.arange(50, 390, 0.1)\n    return x_data[:,None]\n\n# Assign labels depending on x_data\ndef set_y_data(x_data):\n    ydata1 = x_data &gt;= 200\n    ydata2 = x_data &lt; 200\n    return np.hstack((ydata1, ydata2))\n\ndef set_bias(efficiency):\n    arr = np.array([efficiency])\n\n    return arr\n\n# Let's keep W1 and W5 (one hidden layer only)\n# BTW, in this problem you could do with 0 hidden layers. But keeping\n# 1 to show it works\nW1 = tf.Variable(tf.random_normal([1, 5]), name='weight1')\nW5 = tf.Variable(tf.random_normal([5, 2]), name='weight5')\n\n# BTW, b should be a parameter, too. \nb = tf.Variable(tf.constant(0.0))\n\n# Just keeping 1 hidden layer\ndef inference(input):\n    hidden_layer1 = tf.sigmoid(tf.matmul(input, W1) + b)\n    out_layer = tf.nn.softmax(tf.matmul(hidden_layer1, W5) + b)\n\n    return out_layer\n\n# This is unchanged\ndef loss(hypothesis, y):\n    cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(hypothesis), reduction_indices=[1]))\n\n    return cross_entropy\n\n# This is unchanged\ndef train(loss):\n    optimizer = \ntf.train.GradientDescentOptimizer(learning_rate=0.1)\n    train = optimizer.minimize(loss)\n\n    return train\n\n# Using SCALE to normalize the input variables (range of inputs too big)\n# This is a simple normalization in this case. Other examples are \n# Min-Max normalization or z-scores. \n\nSCALE = 1000\nx_data = set_x_data()\ny_data = set_y_data(x_data)\nx_data \/= SCALE\n\n# Now only placeholders are x and y (b is a parameter)\nx= tf.placeholder(tf.float32, shape=[None, 1])\ny= tf.placeholder(tf.float32, shape=[None, 2])\n\nhypothesis = inference(x)\nloss = loss(hypothesis, y)\ntrain = train(loss)\n\nsess = tf.Session()\ninit = tf.global_variables_initializer()\nsess.run(init)\n\nprint(sess.run(W1))\n\n# Epochs x 10, it did not converge with fewer epochs\nepochs = 20000\nlosses = np.zeros(epochs)\nfor step in range(epochs):\n# Shuffle data\n    r = np.random.permutation(x_data.shape[0])\n    x_data = x_data[r]\n    y_data = y_data[r,:]\n    # Small modification here to capture the loss. \n    _, l = sess.run([train, loss], feed_dict={x:x_data, y:y_data})\n    losses[step] = l\n\nprint(sess.run(W1))\nprint(sess.run(b))\n<\/code><\/pre>\n\n<p>The code to display the decision function above: <\/p>\n\n<pre><code>%matplotlib inline\nimport matplotlib.pyplot as plt\nystar = np.arange(50, 400, 10)[:,None]\nplt.plot(ystar, sess.run(hypothesis, feed_dict={x:ystar\/SCALE})[:,0])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53524312,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T16:46:47.057","AcceptedAnswerId":53524683.0,"OwnerUserId":6246880.0,"Title":"TensorFlow: offset when rotating image and points by a random angle","Body":"<p>I have an image and 3 points. I want to rotate the image and the points together. To this end, I rotate the image by some angle <em>a<\/em> and the points by the same angle.\nWhen <em>a<\/em> is fixed to a python scalar (say pi\/3), the rotation works fine (cf. image below, the blue dots are on the dark squares).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TsLAj.jpg?s=256\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TsLAj.jpg?s=256\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When the angle is randomly chosen with <code>angle = tf.random_uniform([])<\/code>, there is an offset between the rotated image and the rotated points.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eJnGR.jpg?s=256\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eJnGR.jpg?s=256\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Below is a the full code reproducing this behaviour.<\/p>\n\n<p>My question is: how to explain this behaviour and correct it?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# create toy image\nsquare = np.zeros((1, 800, 800, 3))\nsquare[:, 100:400, 100:400] = 1\nsquare[:, 140:180, 140:180] = 0\nsquare[:, 240:280, 240:280] = 0\nsquare[:, 280:320, 280:320] = 0\nkp = np.array([[160, 160], [260, 260], [300, 300]])\nkp = np.expand_dims(kp, axis=0)\n\ndef _rotate(image, keypoints, angle, keypoints_num):\n    image = tf.contrib.image.rotate(image, angle)\n    cos, sin = tf.cos(angle), tf.sin(angle)\n    x0, y0  = .5, .5\n    rot_mat = tf.Variable([[cos, -sin], [sin, cos]], trainable=False)\n    keypoints -= (x0, y0)\n    keypoints = tf.reshape(keypoints, shape=[-1, 2])\n    keypoints = tf.matmul(keypoints, rot_mat)\n    keypoints = tf.reshape(keypoints, shape=[-1, keypoints_num, 2])\n    keypoints += (x0, y0)\n    return image, keypoints\n\n\nimage = tf.placeholder(tf.float32, [None, 800, 800, 3])\nkeypoints = tf.placeholder(tf.float32, [None, 3, 2])\n\nangle = np.pi \/ 3 # fix angle, works fine\n#angle = tf.random_uniform([]) # random angle, does not work\nimage_r, keypoints_r = _rotate(image, keypoints \/ 800, angle, 3)\nkeypoints_r *= 800\n\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\n\nimr, kr = sess.run([image_r, keypoints_r], feed_dict={image: square, keypoints:kp})\n\n# displaying output\nplt.imshow(imr[0])\nplt.scatter(*zip(*kr[0]))\nplt.savefig('rotation.jpg')\n<\/code><\/pre>\n","answers":[{"AnswerId":"53524683","CreationDate":"2018-11-28T17:07:26.350","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The problem is here:<\/p>\n\n<pre><code>rot_mat = tf.Variable([[cos, -sin], [sin, cos]], trainable=False)\n<\/code><\/pre>\n\n<p>Since <code>rot_mat<\/code> is a variable, its value is being set only when variables are initialized, here:<\/p>\n\n<pre><code>sess.run(tf.initialize_all_variables())\n<\/code><\/pre>\n\n<p>So at that point <code>rot_mat<\/code> gets some value (using <code>cos<\/code> and <code>sin<\/code>, which in turn depend on <code>angle<\/code>, which is random) and it does not change anymore. Then when you do:<\/p>\n\n<pre><code>imr, kr = sess.run([image_r, keypoints_r], feed_dict={image: squares, keypoints:kps})\n<\/code><\/pre>\n\n<p>It is a different call to <code>run<\/code>, so <code>tf.random_uniform<\/code> produces a new value, but <code>rot_mat<\/code> still keeps the same value from when it was initialized. Since the image is rotated with:<\/p>\n\n<pre><code>image = tf.contrib.image.rotate(image, angle)\n<\/code><\/pre>\n\n<p>And the key points are rotated with:<\/p>\n\n<pre><code>keypoints = tf.matmul(keypoints, rot_mat)\n<\/code><\/pre>\n\n<p>The rotations do not match. The easiest fix is not to use a variable for <code>rot_mat<\/code>:<\/p>\n\n<pre><code>rot_mat = [[cos, -sin], [sin, cos]]\n<\/code><\/pre>\n\n<p>With this, the code works fine. If you really need <code>rot_mat<\/code> to be a variable, it is possible, but it is a bit more of work and it does not seem to be needed here. If you do not like <code>rot_mat<\/code> being a list and want to have a proper tensor instead, you can use <code>tf.convert_to_tensor<\/code>:<\/p>\n\n<pre><code>rot_mat = tf.convert_to_tensor([[cos, -sin], [sin, cos]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53524558,"AnswerCount":0,"Tags":"<python><tensorflow><keras><multi-gpu>","CreationDate":"2018-11-28T17:00:59.123","AcceptedAnswerId":null,"OwnerUserId":3426270.0,"Title":"Keras anomaly in its training time","Body":"<p>I am using Keras in multi-gpu, with Tensorflow backend on 2 gpus. I am using a generator (keras.utils.Sequence) to load my data in batch mode (BS = 64). Therefore I am using the <code>fit_generator<\/code> class, providing it with my train and validation data and steps.\nI noticed a strange behaviour starting from the 2nd epoch on. Basically, the first 3 steps of each epoch are completed in just 8\/9 seconds each, then the network starts taking longer and longer (as it should do). Logs are the following:<\/p>\n\n<pre><code>Epoch 00001: val_acc improved from -inf to 0.46875, saving model to data\/subs_best_model.h5\nEpoch 2\/32\n 1\/29 [&gt;.............................] - ETA: 8s - loss: 1.0664 - acc: 0.5000\n 2\/29 [=&gt;............................] - ETA: 8s - loss: 1.1384 - acc: 0.4531\n 3\/29 [==&gt;...........................] - ETA: 9s - loss: 1.0915 - acc: 0.5052\n 4\/29 [===&gt;..........................] - ETA: 42:03 - loss: 1.1064 - acc: 0.5117\n 5\/29 [====&gt;.........................] - ETA: 56:02 - loss: 1.1173 - acc: 0.4969\n 6\/29 [=====&gt;........................] - ETA: 1:03:13 - loss: 1.0964 - acc: 0.4974\n 7\/29 [======&gt;.......................] - ETA: 1:06:45 - loss: 1.0740 - acc: 0.5067\n 8\/29 [=======&gt;......................] - ETA: 1:08:35 - loss: 1.0592 - acc: 0.5195\n 9\/29 [========&gt;.....................] - ETA: 1:08:53 - loss: 1.0580 - acc: 0.5191\n<\/code><\/pre>\n\n<p>Do you know what could cause this anomaly\/strange behaviour? <\/p>\n\n<p>EDIT:<\/p>\n\n<p>My <code>DataGenerator<\/code> is inspired by <a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">this implementation<\/a><\/p>\n\n<p>The code I use for the fit_generator is as follows:<\/p>\n\n<pre><code>params = {'batch_size':  TrainConfig.BATCH_SIZE,\n              'dim' : ( TrainConfig.BATCH_SIZE, 1, TrainConfig.SAMPLES),\n              'labels_dim': ( TrainConfig.BATCH_SIZE,),\n              'n_classes' : TrainConfig.OUTPUT_DIM}\n\ntraining_generator = DataGenerator(train_set, **params)\nvalidation_generator = DataGenerator(val_set, **params)\n\ntraining_steps_per_epoch = int(1.*len(train_set) \/ batch_size)\nvalidation_steps_per_epoch = int(1.*len(val_set) \/ batch_size)\n\nhistory = model.fit_generator(generator=training_generator,\n                              verbose=1,\n                              use_multiprocessing=False,\n                              workers=1,\n                              steps_per_epoch=training_steps_per_epoch,\n                              epochs=epochs,\n                              validation_data=validation_generator,\n                              validation_steps =validation_steps_per_epoch,\n                              callbacks=callbacks)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53525440,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-28T17:57:50.033","AcceptedAnswerId":null,"OwnerUserId":10718070.0,"Title":"Change output shape of two merged layers in Keras","Body":"<p>I am trying to create a model for argument mining, which takes two sentences as input and return a single output, whether the two are neural, disagreeing or agreeing. The model is build with two bidirectional LSTM layers which get merged into a single layer, <a href=\"https:\/\/i.stack.imgur.com\/a7dbQ.png\" rel=\"nofollow noreferrer\">Model can be seen here<\/a>. The problem is that the output layers shape is (timestamps, label) where I would like it to be a single output of shape (1, label).<\/p>\n\n<pre><code>    LstmLayer = LSTM(32, activation='tanh', return_sequences=False)\n    left_input = Input(shape=(1000,), dtype='float32', name='left_input')\n    right_input = Input(shape=(1000,), dtype='float32', name='right_input')\n\n    left = Embedding(len(self.word_index) + 1,\n                     100,\n                     weights=[self.embeddingMatrix],\n                     input_length=1000,\n                     trainable=False)(left_input)\n    left = Bidirectional(LstmLayer, merge_mode='concat')(left)\n\n    right = Embedding(len(self.word_index) + 1,\n                     100,\n                     weights=[self.embeddingMatrix],\n                     input_length=1000,\n                     trainable=False)(right_input)\n    right = Bidirectional(LstmLayer, merge_mode='concat')(right)\n\n    merged = kl.concatenate([left, right], axis=1)\n    merged = Dense(32, activation='tanh')(merged)\n    main_output = Dense(self.OUTPUT_SIZE, activation='softmax', name='main_output')(merged)\n    self.model = Model(inputs=[left_input, right_input], outputs=[main_output])\n<\/code><\/pre>\n\n<p>The summary of the model can be seen below:<\/p>\n\n<pre><code>    Layer (type)                     Output Shape      Param #         Connected to                     \n    =====================================================================\n    left_input (InputLayer)         (None, 1000)         0            \n    right_input (InputLayer)        (None, 1000)         0                    \n    embedding_1 (Embedding)         (None, 1000, 100)    629600      left_input[0][0]\n    embedding_2 (Embedding)         (None, 1000, 100)    629600      right_input[0][0]\n    bidirectional_1 (Bidirectional) (None, 64)           34048       embedding_1[0][0]\n    bidirectional_2 (Bidirectional) (None, 64)           34048       embedding_2[0][0]\n    concatenate_1 (Concatenate)     (None, 128)          0           bidirectional_1[0][0]\n                                                                     bidirectional_2[0][0]\n    dense_1 (Dense)                 (None, 32)           4128        concatenate_1[0][0]\n    main_output (Dense)             (None, 3)            99          dense_1[0][0]  \n    ======================================================================\n    Total params: 1,331,523\n    Trainable params: 72,323\n    Non-trainable params: 1,259,200\n<\/code><\/pre>\n\n<p>The prediction function encodes the two sentences to the same length, and tokenize them with the pre-trained tokenizer which the model used.<\/p>\n\n<pre><code> pred = self.model.predict([self._encode_sentences(sentence1), self._encode_sentences(sentence2)])\n    return pred\n<\/code><\/pre>\n\n<p>If i predict on (\"I am looking at a dog\", \"I am looking at a cat\"), the output shape is (21,3)<\/p>\n","answers":[{"AnswerId":"53529046","CreationDate":"2018-11-28T22:28:21.233","ParentId":null,"OwnerUserId":"4592059","Title":null,"Body":"<p>What you probably want:<\/p>\n\n<blockquote>\n  <p>I would like it to be a single output of shape <code>(1, label)<\/code>.<\/p>\n<\/blockquote>\n\n<p>You can obtain it getting the last output in the output sequence i.e <code>result_matrix[-1]<\/code>.<\/p>\n\n<pre><code>pred = self.model.predict([self._encode_sentences(sentence1), self._encode_sentences(sentence2)])\nreturn pred[-1]\n<\/code><\/pre>\n\n<p>as this is the <strong>last state<\/strong> of model prediction with any of the given two input sentences. <\/p>\n"}]}
+{"QuestionId":53525994,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning><nlp><tokenize>","CreationDate":"2018-11-28T18:37:00.260","AcceptedAnswerId":53526098.0,"OwnerUserId":498604.0,"Title":"How to find \"num_words\" or vocabulary size of Keras tokenizer when one is not assigned?","Body":"<p>So if I were to not pass <code>num_words<\/code> argument when initializing <code>Tokenizer()<\/code>, how do I find the vocabulary size after it is used to tokenize the training dataset?<\/p>\n\n<p>Why this way, I don't want to limit the tokenizer vocabulary size to know how well my Keras model perform without it. But then I need to pass on this vocabulary size as the argument in the model's first layer definition.<\/p>\n","answers":[{"AnswerId":"53526098","CreationDate":"2018-11-28T18:44:19.730","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>All the words and their indices will be stored in a dictionary which you can access it using <code>tokenizer.word_index<\/code>. Therefore, you can find the number of the unique words based on the number of elements in this dictionary:<\/p>\n\n<pre><code>num_words = len(tokenizer.word_index) + 1\n<\/code><\/pre>\n\n<p>That <code>+ 1<\/code> is because of reserving padding (i.e. index zero).<\/p>\n\n<p><strong>Note<\/strong>: This solution (obviously) is applicable when you have not set <code>num_words<\/code> argument (i.e. you don't know or want to limit the number of words), since <code>word_index<\/code> contains <strong>all the words<\/strong> (and not only the most frequent words) no matter you set <code>num_words<\/code> or not.<\/p>\n"}]}
+{"QuestionId":53526233,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2018-11-28T18:53:51.790","AcceptedAnswerId":null,"OwnerUserId":10592842.0,"Title":"TensorFlow: What is the easiest way to incorporate predictions from one model in the training of a new model?","Body":"<p>What is the simplest way to use <code>tf.estimator<\/code> trained model A during the training of another model B?<\/p>\n\n<p>The weights in model A are fixed. In model B, I would like to take some inputs, compute, feed these results into model A, then do some more computations on the output.<\/p>\n\n<p>A simple example:<\/p>\n\n<p>ModelA returns <code>tf.matmul(input,weights)<\/code><\/p>\n\n<p>In ModelB, I would like to do the following:<\/p>\n\n<pre><code>  x1 = tf.matmul(new_inputs,new_weights1)\n  x2 = modelA(x1) # with fixed weights\n  return tf.matmul(x2,new_weights2)\n<\/code><\/pre>\n\n<p>But with more complicated models A and B, each of which is trained as a <code>tf.estimator<\/code> (though I'm happy to not use estimators if there's another easy solution -- I'm using them because I would like to use ML Engine).<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50084579\/tensorflow-use-a-tf-estimator-trained-model-within-another-tf-estimator-model-f\">This<\/a> question is related, but the proposed solution does not work for training model B, because the gradients of tf.py_func are [None]. I have tried registering a gradient for tf.py_func, but this fails with <\/p>\n\n<blockquote>\n  <p>Unsupported object type Tensor<\/p>\n<\/blockquote>\n\n<p>I have also tried <code>tf.import_graph_def<\/code> for model A, but this seems to load the pretrained graph, but not the actual weights.<\/p>\n","answers":[{"AnswerId":"53561485","CreationDate":"2018-11-30T16:34:27.087","ParentId":null,"OwnerUserId":"3645038","Title":null,"Body":"<p>For model composability, Keras works a whole lot better. You can convert a Keras model to estimator:<\/p>\n\n<p><a href=\"https:\/\/cloud.google.com\/blog\/products\/gcp\/new-in-tensorflow-14-converting-a-keras-model-to-a-tensorflow-estimator\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/blog\/products\/gcp\/new-in-tensorflow-14-converting-a-keras-model-to-a-tensorflow-estimator<\/a><\/p>\n\n<p>So you can still train on ML Engine.<\/p>\n\n<p>With Keras, it is then just a matter of loading the intermediate layers' weights and biases from a checkpoint and make that layer non-trainable. See:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/44644800\/is-it-possible-to-save-a-trained-layer-to-use-layer-on-keras\">Is it possible to save a trained layer to use layer on Keras?<\/a><\/p>\n"}]}
+{"QuestionId":53526259,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T18:55:20.683","AcceptedAnswerId":53526811.0,"OwnerUserId":5154274.0,"Title":"Dynamically determine a tensor with which the variable has been initialized in TensorFlow","Body":"<p>I want to dynamically determine a tensor with which the variable has been initialized (NOT an initialized value, but a tensor-initializer!). For example:<\/p>\n\n<pre><code>var1 = tf.Variable(tf.random_normal(shape=(2, 2)))\nvar2 = tf.Variable(var1.&lt;some method that returns tf.random_normal(shape=(2, 2)&gt;)\n<\/code><\/pre>\n\n<p><code>Variable.initialized_value()<\/code> and <code>Variable.initial_value<\/code> both return the value itself. <\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"53526811","CreationDate":"2018-11-28T19:32:39.080","ParentId":null,"OwnerUserId":"5154274","Title":null,"Body":"<p>I found a bit an ugly way to do that and I'm posting in case someone else will be interested:<\/p>\n\n<pre><code>import tensorflow as tf\nimport re\n\nvar1 = tf.Variable(tf.random_normal(shape=(2, 2)))\n\nshape = var1.initializer.outputs[0].shape\ninput_ = var1.initializer.node_def.input[1]\ntensor_name = (input_ if re.search(r'\\d+$', input_) is None\n               else '_'.join(input_.split('_')[:-1]))\n\nvar2 = tf.Variable(tf.__dict__[tensor_name](shape=shape))\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(var1.eval())\n    print(var2.eval())\n# [[ 0.23845878 -1.1440094 ]\n#  [ 0.593299   -1.1108586 ]]\n# [[ 1.1235769   1.1481414 ]\n#  [ 1.8934027  -0.33171055]]\n<\/code><\/pre>\n\n<p>In any case, I'll be glad if someone could provide a cleaner way to do this.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>This works only if the initializing tensors have default names.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>This is a much better way:<\/p>\n\n<pre><code>import tensorflow as tf\nvar1 = tf.Variable(tf.random_normal(shape=(2, 2), mean=1, stddev=3))\nvar2 = tf.contrib.copy_graph.copy_variable_to_graph(var1, var1.graph)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(var1.eval())\n    print(var2.eval())\n# [[1.7009485 0.9412894]\n#  [1.0769905 1.3085879]]\n# [[ 5.8595214  8.652523 ]\n#  [ 1.86671   -3.170361 ]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53527260,"AnswerCount":0,"Tags":"<tensorflow><cudnn><nvlink>","CreationDate":"2018-11-28T20:02:44.390","AcceptedAnswerId":null,"OwnerUserId":1363774.0,"Title":"Don't see any transfers on NVLINK with NCCL all_sum test","Body":"<p>With the following code (uses <code>tensorflow.contrib.nccl.all_sum<\/code>), I expected to see bytes being transferred over NVLINK. \nIn reality, I don't. <\/p>\n\n<pre><code>from tensorflow.contrib.nccl import all_sum \n\nwith tf.device('\/gpu:0'):\n        a = tf.get_variable(\n            \"a\", initializer=tf.constant(1.0, shape=(args.dim, args.dim)))\n\nwith tf.device('\/gpu:1'):\n        b = tf.get_variable(\n            \"b\", initializer=tf.constant(2.0, shape=(args.dim, args.dim)))\n\nwith tf.device('\/gpu:0'):\n        summed_node = all_sum([a, b])\n         sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,\n                                            log_device_placement=True))\n\ninit = tf.global_variables_initializer() \nsess.run(init)\n\nwith tf.device('\/gpu:0'):\n        summed = sess.run(summed_node)\n<\/code><\/pre>\n\n<p>My machine is an AWS instance of <code>p3.8xlarge<\/code>. My understanding is, this configuration supports NVLINK.<\/p>\n\n<p>The execution is fine but when I use <code>nvidia-smi nvlink -g 0 -i 0<\/code> the link Tx\/Rx counts are zero. <\/p>\n","answers":[]}
+{"QuestionId":53527368,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><scipy>","CreationDate":"2018-11-28T20:11:34.563","AcceptedAnswerId":null,"OwnerUserId":4505889.0,"Title":"Translating tensorflows conv2d to numpy\/scipy operations?","Body":"<p>This is the documentation for tf.nn.conv2d: Given an input tensor of shape [batch, in_height, in_width, in_channels] and a filter \/ kernel tensor of shape [filter_height, filter_width, in_channels, out_channels], this op performs the following<\/p>\n\n<ol>\n<li>Flattens the filter to a 2-D matrix with shape [filter_height *\nfilter_width * in_channels, <\/li>\n<li>Extracts image patches from the input tensor to form a virtual tensor of shape [batch, out_height, out_width, filter_height *\nfilter_width * in_channels].<\/li>\n<li>For each patch, right-multiplies the filter matrix and the image patch vector.<\/li>\n<\/ol>\n\n<p>In other words, it takes in a tensor of n images and does convolution with out_channel filters. <\/p>\n\n<p>I am trying to translate to code that uses only numpy operations and the code is the following:<\/p>\n\n<pre><code>def my_conv2d(x, kernel):\n   nf = kernel.shape[-1]  # number of filters\n   rf = kernel.shape[0]  # filter size\n   w = kernel\n   s = 1 # stride\n\n   h_range = int((x.shape[2] - rf) \/ s) + 1  # (W - F + 2P) \/ S\n   w_range = int((x.shape[1] - rf) \/ s) + 1  # (W - F + 2P) \/ S\n   np_o = np.zeros((1, h_range, w_range, nf))\n   for i in range(x.shape[0]):\n     for z in range(nf):\n       for _h in range(h_range):\n         for _w in range(w_range):\n           np_o[0, _h, _w, z] = np.sum(x[i, _h * s:_h * s + rf, _w * s:_w * s \n                                + rf, * w[:, :, :, z])                     \n    return np_o\n<\/code><\/pre>\n\n<p>The problem is that code is extremely slow. Are there any numpy or scipy functions that can replicate what tensorflows' conv2d is doing that is of similar efficiency? I have looked at <a href=\"https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.signal.convolve2d.html\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.signal.convolve2d.html<\/a> and it does convolution ONCE, meaning I have to pass a 2d tensor alongside a 2d kernel (it does not do multiple filters).\nNone of the previous stackoverflow questions helped much with this.<\/p>\n\n<p>Thanks<\/p>\n\n<p>Edit: did some testing and my code is about 44000% slower than doing tf.nn.conv2d!<\/p>\n","answers":[{"AnswerId":"53539143","CreationDate":"2018-11-29T12:31:17.223","ParentId":null,"OwnerUserId":"10658208","Title":null,"Body":"<p>Things are slow for you because you are using loops. Implementing with vector operations will be much faster but not as efficient as the high-level APIs such as tf.nn.conv2d or tf.nn.convolution. This post should be able to help you with the vectorized implementation of the same in numpy : <a href=\"https:\/\/wiseodd.github.io\/techblog\/2016\/07\/16\/convnet-conv-layer\/\" rel=\"nofollow noreferrer\">https:\/\/wiseodd.github.io\/techblog\/2016\/07\/16\/convnet-conv-layer\/<\/a><\/p>\n"}]}
+{"QuestionId":53527902,"AnswerCount":1,"Tags":"<python><keras><neural-network><linear-regression><least-squares>","CreationDate":"2018-11-28T20:50:34.317","AcceptedAnswerId":null,"OwnerUserId":5085211.0,"Title":"Linear singly-celled two-layer ANN with produces constant predictions","Body":"<p>Say that we want to fit a straight line in the plane through the origin and the point (1, 2). We can view this as linear regression with a sample of size 1 and no intercept. This, on the other hand, can be represented as a rather trivial neural network with no hidden layers and no activation. In Keras, this would look as follows:<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Dense(1, input_shape=(1,), use_bias=False,\n                       kernel_initializer=keras.initializers.Zeros()),\n])\n\nmodel.compile(loss='mse', optimizer=keras.optimizers.Adam(lr=.1))\n<\/code><\/pre>\n\n<p>Running<\/p>\n\n<pre><code>model.fit([[1]], [[2]], epochs=50)\n<\/code><\/pre>\n\n<p>we get our desired result. Namely, <code>model.predict([[1]])<\/code> produces a number close to 2.<\/p>\n\n<p>Now, include a second layer in the model:<\/p>\n\n<pre><code>model2 = keras.Sequential([\n    keras.layers.Dense(1, input_shape=(1,), use_bias=False,\n                       kernel_initializer=keras.initializers.Zeros()),\n    keras.layers.Dense(1, use_bias=False,\n                       kernel_initializer=keras.initializers.Zeros())\n])\n\nmodel2.compile(loss='mse',\n               optimizer=keras.optimizers.Adam(lr=.1))\n<\/code><\/pre>\n\n<p>This time around, no matter what we throw after <code>model2.fit<\/code>, we'll find that <code>model2.predict<\/code> is constantly equal to zero. Moreover, this happens regardless of which one of Keras' gradient descent algorithms is used.<\/p>\n\n<p>This is surprising since the two models are equivalent in the sense that the parameter of the first model can be realized as the product of the two parameters in the second one. That is, we are fitting a straight line <em>y<\/em> = <em>abx<\/em> by trying to find values of <em>a<\/em> and <em>b<\/em>. Taking, for instance, <em>a<\/em> = 2 and <em>b<\/em> = 1 would produce a perfect fit.<\/p>\n\n<p>So, what's the difference?<\/p>\n","answers":[{"AnswerId":"53527903","CreationDate":"2018-11-28T20:50:34.317","ParentId":null,"OwnerUserId":"5085211","Title":null,"Body":"<p>This turns out to be a rather unfortunate side effect of the chosen initial values. It is easy to see that for the given sample, the gradient of the mean squared error optimization objective is<\/p>\n\n<p>(<em>a<\/em>, <em>b<\/em>) \u21a6 2(<em>b<\/em>(<em>ab<\/em> \u2212 2), <em>a<\/em>(<em>ab<\/em> \u2212 2))<\/p>\n\n<p>which happens to equal zero when (<em>a<\/em>, <em>b<\/em>) = (0, 0). Choosing any other initial values for the parameters, and everything goes through as expected. In particular, using<\/p>\n\n<pre><code>kernel_initializer=keras.initializers.Constant(.1)\n<\/code><\/pre>\n\n<p>in both layers would produce the desired result.<\/p>\n\n<p>The same issue does not occur in the one-layer case since here the gradient is<\/p>\n\n<p><em>a<\/em> \u21a6 2(<em>a<\/em> \u2212 2)<\/p>\n\n<p>which is zero only when <em>a<\/em> = 2.<\/p>\n"}]}
+{"QuestionId":53528950,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-28T22:18:00.950","AcceptedAnswerId":null,"OwnerUserId":798689.0,"Title":"Map function to partial tensor in TensorFlow","Body":"<p>In the TensorFlow architecture, how do you apply a function to only some elements in a tensor? For example, on the final output of a layer, some variables represent pixel densities which I would like to run through a sigmoid, and a few variables at the beginning of the tensor represent unbounded continuous values, which I would not like to apply an activation function at all.<\/p>\n\n<p>Here is what I am trying:<\/p>\n\n<pre><code>    output_activation=tf.nn.sigmoid\n    x = tf.layers.dense(z, hidden_size, activation=hidden_activation)\n    x = tf.layers.dense(x, hidden_size, activation=hidden_activation)\n    _logits = tf.layers.dense(x, output_size, activation=None)\n    # Apply sigmoid only to image variables\n    logits = tf.concat(_logits[:functional_inputs], tf.map_fn(output_activation, _logits[functional_inputs:]), 0)\n<\/code><\/pre>\n\n<p>Which is giving the following value error:<\/p>\n\n<pre><code>ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor(\"build_decoder\/map\/TensorArrayStack\/TensorArrayGatherV3:0\", shape=(?, 288755), dtype=float32)'\n<\/code><\/pre>\n\n<p>Is this the best way to do this?<\/p>\n","answers":[]}
+{"QuestionId":53529112,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2018-11-28T22:34:03.750","AcceptedAnswerId":53532648.0,"OwnerUserId":10719567.0,"Title":"Editing layers in trained network keras","Body":"<p>I have a trained model for image demosaicing, and I want to make it smaller, by removing filters in over spec'd layers.<\/p>\n\n<p>For example, I want to take the following model (exract):<\/p>\n\n<pre><code>    conv1 = Conv2D(32, self.kernel_size, activation='relu', padding='same')(chnl4_input)\n    conv2 = Conv2D(32, self.kernel_size, strides=(2, 2), activation='relu', padding='same')(conv1)\n\n    conv5 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv2)\n    conv6 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(conv5)\n\n    up1 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv1], axis=-1)\n    conv7 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(up1)\n<\/code><\/pre>\n\n<p>and I want to change the conv5 and conv6 layers to this:<\/p>\n\n<pre><code>    conv1 = Conv2D(32, self.kernel_size, activation='relu', padding='same')(chnl4_input)\n    conv2 = Conv2D(32, self.kernel_size, strides=(2, 2), activation='relu', padding='same')(conv1)\n\n    conv5 = Conv2D(32, self.kernel_size, activation='relu', padding='same')(conv2)\n    conv6 = Conv2D(32, self.kernel_size, activation='relu', padding='same')(conv5)\n\n    up1 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv1], axis=-1)\n    conv7 = Conv2D(64, self.kernel_size, activation='relu', padding='same')(up1)\n<\/code><\/pre>\n\n<p>I've looked around, but haven't seen any glaringly obvious ways to do this.\n<a href=\"https:\/\/stackoverflow.com\/questions\/45306433\/removing-then-inserting-a-new-middle-layer-in-a-keras-model\">I found this<\/a> example of a similar problem, but the solution specifically mentions that the new layers must have the same amount of filters as the old layers, which is no good for me.<\/p>\n\n<p>If anyone has any idea how I could do this, I'd really appreciate it.<\/p>\n\n<p>[EDIT]: To clarify, I have an existing model, say 'model A'.\nI want to create a new model, 'model B'. These two models will be the same, with the exception of the layers I mentioned above.\nI am looking for a way to initialise the new model with the old models weights for all layers except the ones that have been changed. The new model would then be trained to convergence as per usual.<\/p>\n","answers":[{"AnswerId":"53532648","CreationDate":"2018-11-29T05:50:00.437","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Build a new model (with exactly the same structure, only changing the number of filters) and transfer weights properly:<\/p>\n\n<pre><code>transferLayers = [0,1,2,3,4,8,9] #indices must be chosen by you properly\n\nfor layer in transferLayers:\n    newModel.layers[layer].set_weights(oldModel.layers[layer].get_weights())\n<\/code><\/pre>\n\n<p>There will be a problem with <code>conv7<\/code> that will receive a different input ant thus also have a different size for its weight matrix.<\/p>\n\n<h3>What if you change the model structure<\/h3>\n\n<p>Then you should probably create two lists of indices, one for the old model, one for the new.   <\/p>\n\n<p>Or you could recreate the old model adding names to its layers: <\/p>\n\n<ul>\n<li>Recreate the training model exactly the same, but add names to every layer    <\/li>\n<li>Transfer weights: <code>namedTrainingModel.set_weights(unnamedTraininModel.get_weights())<\/code>   <\/li>\n<li>Then create the changed model adding the same names to layers that aren't changed and new names to changed layers    <\/li>\n<\/ul>\n\n<p>Transfer weights by name:   <\/p>\n\n<pre><code>namedTrainingModel.save_weights('filename')    \nchangedModel.load_weights('filename', by_name=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53529614,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2018-11-28T23:24:03.157","AcceptedAnswerId":53532509.0,"OwnerUserId":10300594.0,"Title":"Keras Cant Seem To Configure inputShape","Body":"<p>I'm trying to do the \"Hello World\" of Keras with Iris-Flowers and I cannot configure the input_shape. This is my error message:<\/p>\n\n<pre><code>\"ValueError: Error when checking input: expected dense_1_input to have 3 dimensions, but got array with shape (120, 4)\"\n<\/code><\/pre>\n\n<p>When I change my input_shape to the received input shape, it just changes what it says the received input is. As you know there are 4 inputs and one output with this data set. Does anyone know how I would configure my input_shape?<\/p>\n\n<p>Here's my code<\/p>\n\n<pre><code>import tensorflow as tf\n\ntext_file = open(\"iris.data.txt\")\nrawData = text_file.read().split('\\n')\ntext_file.close()\n\nfor x in range(0,150):\n    rawData[x] = rawData[x].split(',')\n\nxs_train = []\nys_train = []\nfor i in range (0,40):\n    ys_train.append(rawData[i][4])\n    xs_train.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\nfor i in range (50,90):\n    ys_train.append(rawData[i][4])\n    xs_train.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\nfor i in range (100,140):\n    ys_train.append(rawData[i][4])\n    xs_train.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\n\nxs_test = []\nys_test = []\nfor i in range (40,50):\n    ys_test.append(rawData[i][4])\n    xs_test.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\nfor i in range (90,100):\n    ys_test.append(rawData[i][4])\n    xs_test.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\nfor i in range (140,150):\n    ys_test.append(rawData[i][4])\n    xs_test.append([rawData[i][0], rawData[i][1], rawData[i][2], rawData[i][3]])\n\n# print(xs_train)   \n\nfor i in range(0, len(ys_train)):\n    if ys_train[i] == \"Iris-setosa\":\n        ys_train[i] = [1,0,0]\n    if ys_train[i] == \"Iris-versicolor\":\n        ys_train[i] = [0,1,0]\n    if ys_train[i] == \"Iris-virginica\":\n        ys_train[i] = [0,0,1]\n\n# print(ys_train)\n\n\nmodel = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Dense(4, input_shape=(1,4), activation= 'relu'))\nmodel.add(tf.keras.layers.Dense(3, activation=tf.nn.softmax))\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(xs_train, ys_train, epochs=3)\n<\/code><\/pre>\n\n<p>My inputs are formatted so that each array is a dataset, each set including 4 \ndata points, it goes as so:<\/p>\n\n<pre><code>[['5.1', '3.5', '1.4', '0.2'],\n ['4.9', '3.0', '1.4', '0.2'],\n ['4.7', '3.2', '1.3', '0.2']...]\n<\/code><\/pre>\n","answers":[{"AnswerId":"53532509","CreationDate":"2018-11-29T05:37:48.130","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<pre><code>input_shape=(4,)\n<\/code><\/pre>\n\n<p>Input shape doesn't care for the batch size, only for \"each\" sample size.<\/p>\n"}]}
+{"QuestionId":53529632,"AnswerCount":1,"Tags":"<python><tensorflow><graph><detection><tensorboard>","CreationDate":"2018-11-28T23:25:38.230","AcceptedAnswerId":53536629.0,"OwnerUserId":10719729.0,"Title":"How to display the accuracy graph on TensorBoard?","Body":"<p>I'm new to TensorFlow. I am currently working on a project that uses the TF Object detection API. I'm training a model with two classes on my custom images. So far I have successfully run <code>train.py<\/code> and <code>eval.py<\/code> and executed TensorBoard at the same time to see how the training processes is progressing. <\/p>\n\n<p>Here is the image of my work:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/QR9Ve.png\" alt=\"enter image description here\"><\/p>\n\n<p>How do I display a graph in which I can see the accuracy of the model being developed ?<\/p>\n\n<p>Any help is appreciated!<\/p>\n","answers":[{"AnswerId":"53536629","CreationDate":"2018-11-29T10:16:39.930","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>You can do this by switching from the deprecated train.py and eval.py to the recent model_main.py. It interleaves evaluation schedule into the training session, so that you can evaluate the training progress of the model without manually doing so.\nThe flags of model_main.py are very similar to those of train.py, and you can see an example in here:\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/running_locally.md#running-the-training-job\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/running_locally.md#running-the-training-job<\/a><\/p>\n"}]}
+{"QuestionId":53530036,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform>","CreationDate":"2018-11-29T00:13:49.880","AcceptedAnswerId":null,"OwnerUserId":10496625.0,"Title":"Tensorflow on GoogleML with TFrecords: OutOfRangeError (see above for traceback): End of sequence","Body":"<p>Our Tensorflow trains successfully (but slowly) when we run it locally (on a PC without a GPU). But it fails to train when we send it to GoogleML (single GPU). No change in the code or the data fed in. <\/p>\n\n<p>The error occurs while reading data from the TFrecords. Some of the relevant code:<\/p>\n\n<pre><code>files = tf.data.Dataset.list_files(XXXX)\ndataset = files.apply(tf.contrib.data.parallel_interleave(tf.data.TFRecordDataset, cycle_length=XXXX, sloppy=True, buffer_output_elements=XXXX))\n\ndataset = dataset.apply(tf.contrib.data.map_and_batch(map_func=_parse_transform_function, batch_size=config.batch_size, num_parallel_calls=config.num_parallel_calls))\ndataset = dataset.repeat()\ndataset = dataset.prefetch(config.prefetch)\n\nfiles = tf.data.Dataset.list_files(XXXX)\ndataset = files.apply(tf.contrib.data.parallel_interleave(tf.data.TFRecordDataset, cycle_length=XXXX, sloppy=True, buffer_output_elements=XXXX))\n\ndataset = dataset.apply(tf.contrib.data.map_and_batch(map_func=_parse_transform_function, batch_size=config.batch_size, num_parallel_calls=config.num_parallel_calls))\ndataset = dataset.repeat()\ndataset = dataset.prefetch(config.prefetch)\n\ntraining_iterator = dataset.make_one_shot_iterator()\nhandle = tf.placeholder(tf.string, shape=[])\niterator = tf.data.Iterator.from_string_handle(handle, dataset.output_types, dataset.output_shapes)\nnext_element = iterator.get_next()\n\n...\n\nwith tf.Session() as sess:\n    training_handle = sess.run(training_iterator.string_handle())\n    for i in range(config.epochs):\n        _image, _landmarks, _height, _width, _size, _kp_true, _short_offsets_true, _long_offsets_true = sess.run(next_element, feed_dict={handle: training_handle})\n<\/code><\/pre>\n\n<p>Here is the error we get:<\/p>\n\n<blockquote>\n  <p>ERROR 2018-11-28 15:32:55 -0800   service     OutOfRangeError (see above\n  for traceback): End of sequence ERROR 2018-11-28 15:32:55\n  -0800 service          [[Node: IteratorGetNext = IteratorGetNextoutput_shapes=[[?,416,416,3], [?,34,5], [?], [?], [?],\n  [?,52,52,11], [?,52,52,22], [?,52,52,40]], output_types=[DT_FLOAT,\n  DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_FLOAT, DT_FLOAT, DT_FLOAT],\n  _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]]<\/p>\n<\/blockquote>\n\n<p>Again we did not change the data or the code, so we believe we have no bugs in the Tensorflow code.  Why would the same implementation fail to train in GoogleML?<\/p>\n","answers":[{"AnswerId":"53562452","CreationDate":"2018-11-30T17:47:51.257","ParentId":null,"OwnerUserId":"10496625","Title":null,"Body":"<p>\"OutOfRangeError (see above for traceback): End of sequence\" is misleading. It seems to indicate we reached the end of the tfrecords, but that was not the case. <\/p>\n\n<p>The issue was that an operation in the \"_parse_transform_function\" (not shown) was using an operation not supported in a GPU.<\/p>\n\n<p>Solution: Use \"with tf.device('\/cpu:0'):\" around the dataset portion of the code to use the CPU for preprocessing.  The GPU still in use for training.<\/p>\n"}]}
+{"QuestionId":53530354,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2018-11-29T01:00:49.873","AcceptedAnswerId":55061462.0,"OwnerUserId":1371314.0,"Title":"Tensorflow Serving predictions mapped to labels","Body":"<p>I am serving up the inception model using TensorFlow serving. I am doing this on Azure Kubernetes so not via the more standard and well documented google cloud.<\/p>\n\n<p>In any event, this is all working however the bit i am confused about is the predictions come back as an array of floats. These values map to the original labels passed in during training but without the original labels file there is no way to reverse engineer what each probability relates to.<\/p>\n\n<p>Before I moved to serving i was simply using an inference script that then cross references against the labels file which i stored along with the frozen model at time of training. But with serving this does not work.<\/p>\n\n<p>So my question is how can i get the labels associated with the model and ideally get the prediction to return the labels and probabilities?<\/p>\n","answers":[{"AnswerId":"55061462","CreationDate":"2019-03-08T10:41:31.003","ParentId":null,"OwnerUserId":"3816569","Title":null,"Body":"<p>i tried the approach suggested by @user1371314 but i couldn't get it to work. An other solution that worked is creating a tensor (instead of a constant) and map it with only the first element of the output layer when saving the model. When you put it together it looks like this :<\/p>\n\n<pre><code># get labels names and create a tensor from it\n.... \nlabel_names_tensor = tf.convert_to_tensor(label_names)\n\n# save the model and map the labels to the output layer\ntf.saved_model.simple_save(\n        sess,\n        \".\/saved_models\",\n        inputs={'image': model.input},\n        outputs={'label' : label_names_tensor,'prediction': model.output[0]})\n<\/code><\/pre>\n\n<p>When you make a prediction after serving your model you will get the following result:<\/p>\n\n<pre><code>{\n    \"predictions\": [\n        {\n            \"label\": \"label-name\",\n            \"prediction\": 0.114107\n        },\n        {\n            \"label\": \"label-name\",\n            \"prediction\": 0.288598\n        },\n        {\n            \"label\": \"label-name\",\n            \"prediction\": 0.17436\n        },\n        {\n            \"label\": \"label-name\",\n            \"prediction\": 0.186366\n        },\n        {\n            \"label\": \"label-name\",\n            \"prediction\": 0.236568\n        }\n    ]\n<\/code><\/pre>\n"},{"AnswerId":"53585750","CreationDate":"2018-12-02T23:47:19.687","ParentId":null,"OwnerUserId":"1371314","Title":null,"Body":"<p>I am sure there is a way to return a mapping directly for this using the various TF ops however I have managed to at least package the labels into the model and return them in the prediction along with the probabilities.<\/p>\n\n<p>What i did was create a tf.constant from the labels array and then added that tensor to the array of output tensors in tf.saved_model.signature_def_utils.build_signature_def<\/p>\n\n<p>Now when i get a prediction i get the float array and also an array of labels and i can match them up on the client side.<\/p>\n"}]}
+{"QuestionId":53530554,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><artificial-intelligence>","CreationDate":"2018-11-29T01:27:37.977","AcceptedAnswerId":null,"OwnerUserId":3492226.0,"Title":"Applying Normalization to Inputs in Tensorflow","Body":"<p>I have created a custom class to be an ML model, and it is working fine, but I would like to normalize the inputs as they have a wide range of values (e.g. 0, 20000, 500, 10, 8). Currently, as a way of normalizing the inputs, I'm applying <code>lambda x: np.log(x + 1)<\/code> to each input (the +1 is so it doesn't error out when 0 is passed in). Would a normalization layer be better than my current approach? If so, how would I go about implementing it? My code for the model is below:<\/p>\n\n<pre><code>class FollowModel:\n    def __init__(self, input_shape, output_shape, hidden_layers, input_labels, learning_rate=0.001):\n        tf.reset_default_graph()\n        assert len(input_labels) == input_shape[1], 'Incorrect number of input labels!'\n\n        # Placeholders for input and output data\n        self.input_labels = input_labels\n        self.input_shape = input_shape\n        self.output_shape = output_shape\n        self.X = tf.placeholder(shape=input_shape, dtype=tf.float64, name='X')\n        self.y = tf.placeholder(shape=output_shape, dtype=tf.float64, name='y')\n        self.hidden_layers = hidden_layers\n        self.learning_rate = learning_rate\n\n        # Variables for two group of weights between the three layers of the network\n        self.W1 = tf.Variable(np.random.rand(input_shape[1], hidden_layers), dtype=tf.float64)\n        self.W2 = tf.Variable(np.random.rand(hidden_layers, output_shape[1]), dtype=tf.float64)\n\n        # Create the neural net graph\n        self.A1 = tf.sigmoid(tf.matmul(self.X, self.W1))\n        self.y_est = tf.sigmoid(tf.matmul(self.A1, self.W2))\n\n        # Define a loss function\n        self.deltas = tf.square(self.y_est - self.y)  # want this to be 0\n        self.loss = tf.reduce_sum(self.deltas)\n\n        # Define a train operation to minimize the loss\n        self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.loss)\n\n        #initialize\n        self.model_init = tf.global_variables_initializer()\n        self.trained = False\n\n\n    def train(self, Xtrain, ytrain, Xtest, ytest, training_steps, batch_size, print_progress=True):\n        #intiialize session\n        self.trained = True\n        self.training_steps = training_steps\n        self.batch_size = batch_size\n        self.sess = tf.Session()\n        self.sess.run(self.model_init)\n        self.losses = []\n        self.accs = []\n        self.testing_accuracies = []\n\n        for i in range(training_steps*batch_size):\n            self.sess.run(self.optimizer, feed_dict={self.X: Xtrain, self.y: ytrain})\n            local_loss = self.sess.run(self.loss, feed_dict={self.X: Xtrain.values, self.y: ytrain.values})\n            self.losses.append(local_loss)\n            self.weights1 = self.sess.run(self.W1)\n            self.weights2 = self.sess.run(self.W2)\n\n            y_est_np = self.sess.run(self.y_est, feed_dict={self.X: Xtrain.values, self.y: ytrain.values})\n            correct = [estimate.argmax(axis=0) == target.argmax(axis=0)\n                       for estimate, target in zip(y_est_np, ytrain.values)]\n            acc = 100 * sum(correct) \/ len(correct)\n            self.accs.append(acc)\n\n            if i % batch_size == 0:\n                batch_num = i \/ batch_size\n                if batch_num % 5 == 0:\n                    self.testing_accuracies.append(self.test_accuracy(Xtest, ytest, False, True))\n                temp_table = pd.concat([Xtrain, ytrain], axis=1).sample(frac=1)\n                column_names = list(temp_table.columns.values)\n                X_columns, y_columns = column_names[0:len(column_names) - 2], column_names[len(column_names) - 2:]\n                Xtrain = temp_table[X_columns]\n                ytrain = temp_table[y_columns]\n                if print_progress: print('Step: %d, Accuracy: %.2f, Loss: %.2f' % (int(i\/batch_size), acc, local_loss))\n\n        if print_progress: print(\"Training complete!\\nloss: {}, hidden nodes: {}, steps: {}, epoch size: {}, total steps: {}\".format(int(self.losses[-1]*100)\/100, self.hidden_layers, training_steps, batch_size, training_steps*batch_size))\n        self.follow_accuracy = acc\n        return acc\n\n\n    def test_accuracy(self, Xtest, ytest, print_progress=True, return_accuracy=False):\n        if self.trained:\n            X = tf.placeholder(shape=Xtest.shape, dtype=tf.float64, name='X')\n            y = tf.placeholder(shape=ytest.shape, dtype=tf.float64, name='y')\n            W1 = tf.Variable(self.weights1)\n            W2 = tf.Variable(self.weights2)\n            A1 = tf.sigmoid(tf.matmul(X, W1))\n            y_est = tf.sigmoid(tf.matmul(A1, W2))\n\n            # Calculate the predicted outputs\n            init = tf.global_variables_initializer()\n            with tf.Session() as sess:\n                sess.run(init)\n                y_est_np = sess.run(y_est, feed_dict={X: Xtest, y: ytest})\n\n            correctly_followed = 0\n            incorrectly_followed = 0\n            missed_follows = 0\n            correctly_skipped = 0\n\n            for estimate, actual in zip(y_est_np, ytest.values):\n                est = estimate.argmax(axis=0)\n                # print(estimate)\n                actual = actual.argmax(axis=0)\n                if est == 1 and actual == 0: incorrectly_followed += 1\n                elif est == 1 and actual == 1: correctly_followed += 1\n                elif est == 0 and actual == 1: missed_follows += 1\n                else: correctly_skipped += 1\n\n            # correct = [estimate.argmax(axis=0) == target.argmax(axis=0) for estimate, target in zip(y_est_np, ytest.values)]\n            total_followed = incorrectly_followed + correctly_followed\n\n            total_correct = correctly_followed + correctly_skipped\n            total_incorrect = incorrectly_followed + missed_follows\n\n            try: total_accuracy = int(total_correct * 10000 \/ (total_correct + total_incorrect)) \/ 100\n            except: total_accuracy = 0\n\n\n            total_skipped = correctly_skipped + missed_follows\n            try: follow_accuracy = int(correctly_followed * 10000 \/ total_followed) \/ 100\n            except: follow_accuracy = 0\n            try: skip_accuracy = int(correctly_skipped * 10000 \/ total_skipped) \/ 100\n            except: skip_accuracy = 0\n\n            if print_progress: print('Correctly followed {} \/ {} ({}%), correctly skipped {} \/ {} ({}%)'.format(\n                correctly_followed, total_followed, follow_accuracy, correctly_skipped, total_skipped, skip_accuracy))\n\n            self.follow_accuracy = follow_accuracy\n\n            if return_accuracy:\n                return total_accuracy\n\n        else:\n            print('The model is not trained!')\n\n\n    def make_prediction_on_normal_data(self, input_list):\n        assert len(input_list) == len(self.input_labels), 'Incorrect number of inputs (had {} should have {})'.format(len(input_list), len(self.input_labels))\n        # from ProcessData import normalize_list\n        # normalize_list(input_list)\n        input_array = np.array([input_list])\n\n        X = tf.placeholder(shape=(1, len(input_list)), dtype=tf.float64, name='X')\n        y = tf.placeholder(shape=(1, 2), dtype=tf.float64, name='y')\n        W1 = tf.Variable(self.weights1)\n        W2 = tf.Variable(self.weights2)\n        A1 = tf.sigmoid(tf.matmul(X, W1))\n        y_est = tf.sigmoid(tf.matmul(A1, W2))\n\n        with tf.Session() as sess:\n            sess.run(tf.global_variables_initializer())\n            y_est_np = sess.run(y_est, feed_dict={X: input_array, y: self.create_blank_outputs()})\n            predicted_value = y_est_np[0].argmax(axis=0)\n            return predicted_value\n\n\n    def make_prediction_on_abnormal_data(self, input_list):\n        from ProcessData import normalize_list\n        normalize_list(input_list)\n        return self.make_prediction_on_normal_data(input_list)\n\n\n    def create_blank_outputs(self):\n        blank_outputs = np.zeros(shape=(1,2), dtype=np.int)\n        for i in range(len(blank_outputs[0])):\n            blank_outputs[0][i] = float(blank_outputs[0][i])\n        return blank_outputs\n<\/code><\/pre>\n","answers":[{"AnswerId":"53561146","CreationDate":"2018-11-30T16:13:40.767","ParentId":null,"OwnerUserId":"6431410","Title":null,"Body":"<p>I don't see see why you want to create a layer that does that. The common practice of preprocessing your inputs is as you are currently doing. \nUsing the log operator is quite common for skewed data, but there are other preprocessing solutions such as sklearn's MinMaxScaler and StandardScaler<\/p>\n\n<p><a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html<\/a>\n<a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html<\/a><\/p>\n\n<p>Those are just examples of two other ways to scale your data.\nThere is such a thing called BatchNorm but it is not recommended as the first layer of the network as distribution of the data is fixed and doesn\u2019t vary during training.<\/p>\n"}]}
+{"QuestionId":53530561,"AnswerCount":0,"Tags":"<tensorflow><tensorboard><object-detection-api>","CreationDate":"2018-11-29T01:28:33.450","AcceptedAnswerId":null,"OwnerUserId":3625770.0,"Title":"Why LOSS is constantly increasing in tensorflow Object Detection API?","Body":"<p>I am using <code>faster_rcnn_resnet50_coco<\/code> model from the model zoo of tensorflow. I trained the model on my own data (solar images) and as I am observing the training (in tensorboard) the total loss is constantly increasing.\n<a href=\"https:\/\/i.stack.imgur.com\/6haN9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6haN9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is surprising to me because the model seems to be pretty successful in detecting the objects.<\/p>\n","answers":[]}
+{"QuestionId":53530628,"AnswerCount":1,"Tags":"<python><tensorflow><feature-extraction>","CreationDate":"2018-11-29T01:37:33.153","AcceptedAnswerId":null,"OwnerUserId":10404053.0,"Title":"How can I get a TF feature column from a dataframe column?","Body":"<p>I have a dataframe:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eHkgG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eHkgG.png\" alt=\"enter image description here\"><\/a>\nI want to create a feature column out of the integer \"length\" column of this dataframe. How can I achieve that?<\/p>\n","answers":[{"AnswerId":"55087916","CreationDate":"2019-03-10T12:59:21.213","ParentId":null,"OwnerUserId":"1234909","Title":null,"Body":"<pre><code>featcols = [\n    tf.feature_column.numeric_column(\"length\")\n]\n<\/code><\/pre>\n\n<p>This is a <a href=\"https:\/\/colab.research.google.com\/notebooks\/mlcc\/first_steps_with_tensor_flow.ipynb?hl=en#scrollTo=RKZ9zNcHJtwc\" rel=\"nofollow noreferrer\">reference.<\/a><\/p>\n"}]}
+{"QuestionId":53530751,"AnswerCount":1,"Tags":"<python><dataset><pytorch>","CreationDate":"2018-11-29T01:55:47.470","AcceptedAnswerId":54096225.0,"OwnerUserId":7027783.0,"Title":"How make customised dataset in Pytorch for images and their masks?","Body":"<p>I have two dataset folder of tif images, one is a folder called BMMCdata, and the other one is the mask of BMMCdata images called BMMCmasks(the name of images are corresponds). I am trying to make a customised dataset and also split the data randomly to train and test. at the moment I am getting an error <\/p>\n\n<pre><code>self.filenames.append(fn)\nAttributeError: 'CustomDataset' object has no attribute 'filenames'\n<\/code><\/pre>\n\n<p>Any comment will be appreciated a lot. <\/p>\n\n<pre><code>import torch\nfrom torch.utils.data.dataset import Dataset  # For custom data-sets\nfrom torchvision import transforms\nfrom PIL import Image\nimport os.path as osp\nimport glob\n\nfolder_data = \"\/Users\/parto\/PycharmProjects\/U-net\/BMMCdata\/data\"\n\nclass CustomDataset(Dataset):\ndef __init__(self, root):\n\n    self.filename = folder_data\n    self.root = root\n    self.to_tensor = transforms.ToTensor()\n    filenames = glob.glob(osp.join(folder_data, '*.tif'))\n    for fn in filenames:\n        self.filenames.append(fn)\n    self.len = len(self.filenames)\n    print(fn)\n\ndef __getitem__(self, index):\n    image = Image.open(self.filenames[index])\n    return self.transform(image)\n\ndef __len__(self):\n\n    return self.len\ncustom_img = CustomDataset(folder_data)\n# total images in set\nprint(custom_img.len)\n\ntrain_len = int(0.6*custom_img.len)\ntest_len = custom_img.len - train_len\ntrain_set, test_set = CustomDataset.random_split(custom_img, lengths=[train_len, test_len])\n# check lens of subset\nlen(train_set), len(test_set)\n\ntrain_set = CustomDataset(folder_data)\ntrain_set = torch.utils.data.TensorDataset(train_set, train=True, batch_size=4)\ntrain_loader = torch.utils.data.DataLoader(train_set, batch_size=4, shuffle=True, num_workers=1)\nprint(train_set)\nprint(train_loader)\n\ntest_set = torch.utils.data.DataLoader(Dataset, batch_size=4, sampler= train_sampler)\ntest_loader = torch.utils.data.DataLoader(Dataset, batch_size=4)\n<\/code><\/pre>\n","answers":[{"AnswerId":"54096225","CreationDate":"2019-01-08T16:48:05.993","ParentId":null,"OwnerUserId":"7027783","Title":null,"Body":"<p>answer given by @ptrblck in pytorch community. thank you<\/p>\n\n<pre><code> # get all the image and mask path and number of images\n folder_data = glob.glob(\"D:\\\\Neda\\\\Pytorch\\\\U-net\\\\BMMCdata\\\\data\\\\*.tif\")\n folder_mask = glob.glob(\"D:\\\\Neda\\\\Pytorch\\\\U-net\\\\BMMCmasks\\\\masks\\\\*.tif\")\n\n # split these path using a certain percentage\n len_data = len(folder_data)\n print(len_data)\n train_size = 0.6\n\n train_image_paths = folder_data[:int(len_data*train_size)]\n test_image_paths = folder_data[int(len_data*train_size):]\n\n train_mask_paths = folder_mask[:int(len_data*train_size)]\n test_mask_paths = folder_mask[int(len_data*train_size):]\n\n\n class CustomDataset(Dataset):\n    def __init__(self, image_paths, target_paths, train=True):   # initial logic \n      happens like transform\n\n         self.image_paths = image_paths\n         self.target_paths = target_paths\n         self.transforms = transforms.ToTensor()\n\n    def __getitem__(self, index):\n\n        image = Image.open(self.image_paths[index])\n        mask = Image.open(self.target_paths[index])\n        t_image = self.transforms(image)\n     return t_image, mask\n\ndef __len__(self):  # return count of sample we have\n\n    return len(self.image_paths)\n\ntrain_dataset = CustomDataset(train_image_paths, train_mask_paths, train=True)\ntrain_loader = torch.utils.data.DataLoader(train_dataset, batch_size=4, shuffle=True, num_workers=1)\n\ntest_dataset = CustomDataset(test_image_paths, test_mask_paths, train=False)\ntest_loader = torch.utils.data.DataLoader(test_dataset, batch_size=4, shuffle=False, num_workers=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53530810,"AnswerCount":1,"Tags":"<python-3.x><pytorch>","CreationDate":"2018-11-29T02:04:17.983","AcceptedAnswerId":null,"OwnerUserId":6328927.0,"Title":"RuntimeError: reduce failed to synchronize: device-side assert triggered","Body":"<pre><code>File \"\/home\/username\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/loss.py\", line 433, in forward     reduce=self.reduce)   \nFile \"\/home\/username\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/functional.py\", line 1483, in binary_cross_entropy     \nreturn torch._C._nn.binary_cross_entropy(input, target, weight, size_average, reduce)\n\nRuntimeError: reduce failed to synchronize: device-side assert triggered \n<\/code><\/pre>\n","answers":[{"AnswerId":"55720103","CreationDate":"2019-04-17T04:49:11.103","ParentId":null,"OwnerUserId":"1434041","Title":null,"Body":"<p>When using <code>cuda<\/code> you may get this generic error which is not very helpful. Try switching to <code>cpu<\/code> device instead, <code>device = torch.device(\"cpu\")<\/code>, to see the actual error stack trace.<\/p>\n\n<p>In my case, the issue was caused because binary cross entropy expected the input values to be between 0~1, but I was sending in values between -1 and 1. Applying <code>sigmoid<\/code> to the output resolved that issue.<\/p>\n"}]}
+{"QuestionId":53530867,"AnswerCount":3,"Tags":"<tensorflow><amazon-sagemaker>","CreationDate":"2018-11-29T02:12:12.190","AcceptedAnswerId":null,"OwnerUserId":1007922.0,"Title":"How do you install modules within sagemaker training jobs?","Body":"<p>I don't think I'm asking this question right but I have jupyter notebook that launches a Tensorflow training job with a python training script I wrote.<\/p>\n\n<p>That training script requires certain modules. Seems my sagemaker training job is failing because some of the modules don't exist.<\/p>\n\n<p>How can I ensure that my training job script has all the modules it needs?<\/p>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>An example of one of these modules is <code>keras<\/code>.<\/p>\n\n<p>The odd thing is, I can <code>import keras<\/code> in the jupyter notebook, but when that import statement is in my training script then I get the <code>No module named keras<\/code> error<\/p>\n","answers":[{"AnswerId":"53532238","CreationDate":"2018-11-29T05:09:52.747","ParentId":null,"OwnerUserId":"6530653","Title":null,"Body":"<p>The module script runs within a docker container which obviously does not have the dependency installed. Jupyter notebook on the other hand has keras pre-installed. \nEasy way to do this is to have a requirements.txt file with all the requirements and then pass that on when creating your model. <\/p>\n\n<pre><code>env = {\n    'SAGEMAKER_REQUIREMENTS': 'requirements.txt', # path relative to `source_dir` below.\n}\nsagemaker_model = TensorFlowModel(model_data = 's3:\/\/mybucket\/modelTarFile,\n                                  role = role,\n                                  entry_point = 'entry.py',\n                                  code_location = 's3:\/\/mybucket\/runtime-code\/',\n                                  source_dir = 'src',\n                                  env = env,\n                                  name = 'model_name',\n                                  sagemaker_session = sagemaker_session,\n                                 )\n<\/code><\/pre>\n"},{"AnswerId":"53824894","CreationDate":"2018-12-18T00:43:38.433","ParentId":null,"OwnerUserId":"9090582","Title":null,"Body":"<p>The environment on your notebook instance is exclusive from the environment of your training job on SageMaker, unless it is local mode.<\/p>\n\n<p>If you're using a custom docker image, then most likely your docker image doesn't have Keras installed.<\/p>\n\n<p>If you are using the SageMaker predefined TensorFlow container, which is most likely invoked through the following code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/estimator.py#L170\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/blob\/master\/src\/sagemaker\/tensorflow\/estimator.py#L170<\/a><\/p>\n\n<pre><code>TensorFlow(entry_point='training_code.py',\n           blah,\n           blah\n          )\n<\/code><\/pre>\n\n<p>Then you will need to install your dependencies within that container. There are currently two modes for training for TensorFlow on SageMaker, \"framework\" and \"script\" mode.<\/p>\n\n<hr>\n\n<p>If training through \"framework\" mode, which is only available with 1.12 and below, then you will be limited to using a keras_model_fn defined here:\n<a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/tree\/v1.12.0\/src\/sagemaker\/tensorflow#preparing-the-tensorflow-training-script\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/tree\/v1.12.0\/src\/sagemaker\/tensorflow#preparing-the-tensorflow-training-script<\/a><\/p>\n\n<p>Installing your dependencies would be done by passing in a requirements.txt.<\/p>\n\n<hr>\n\n<p>On \"script mode\", which is introduced with TensorFlow 1.11 and above:\n<a href=\"https:\/\/github.com\/aws\/sagemaker-python-sdk\/tree\/master\/src\/sagemaker\/tensorflow#training-with-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aws\/sagemaker-python-sdk\/tree\/master\/src\/sagemaker\/tensorflow#training-with-tensorflow<\/a><\/p>\n\n<p>Requirements.txt is not supported for \"script\" mode and instead it is recommended to install your dependencies within your user script, which would be your Python file that contains all of your Keras code.<\/p>\n\n<p>Please let me know if there is anything I can clarify.<\/p>\n\n<p>For examples:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/tree\/master\/sagemaker-python-sdk\/tensorflow_script_mode_quickstart\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/tree\/master\/sagemaker-python-sdk\/tensorflow_script_mode_quickstart<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/tree\/master\/sagemaker-python-sdk\/tensorflow_iris_dnn_classifier_using_estimators\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/tree\/master\/sagemaker-python-sdk\/tensorflow_iris_dnn_classifier_using_estimators<\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"57693268","CreationDate":"2019-08-28T13:11:28.627","ParentId":null,"OwnerUserId":"3415910","Title":null,"Body":"<ol>\n<li><p>You can upload your requirements.txt file to s3 bucket which can be\naccessible by sagemaker and download the file to your working\ndirectory of the container using boto3. Install the libraries from\nrequirements.txt the <strong>entry file<\/strong>.<\/p>\n\n<pre><code>    import os\n    import boto3\n\n    s3 = boto3.client('s3')\n    s3.download_file('BUCKET_NAME', 'OBJECT_NAME', '\/opt\/ml\/code\/requirements.txt')\n    os.command('pip install -r \/opt\/ml\/code\/requirements.txt')\n<\/code><\/pre><\/li>\n<li><p>The other way you can do it is by building your own container using\n<strong>bring your own algorithm<\/strong> option provided by aws.<\/p><\/li>\n<\/ol>\n\n<p>Ref-links:<br>\n<a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/advanced_functionality\/scikit_bring_your_own\/scikit_bring_your_own.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/advanced_functionality\/scikit_bring_your_own\/scikit_bring_your_own.ipynb<\/a><\/p>\n"}]}
+{"QuestionId":53531820,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-29T04:25:21.627","AcceptedAnswerId":53532448.0,"OwnerUserId":10720392.0,"Title":"Keras: notImplementedError\/RuntimeError when using fit_generator","Body":"<p>I am having troubles with keras and tensorflow, using the following code:<\/p>\n\n<pre><code>from tensorflow.keras.layers import Activation, Conv2D\nfrom tensorflow.keras import Model\n\nfrom data import DataGenerator\nfrom config import train_datapath, test_datapath\n\n\ntraining_generator = DataGenerator(train_datapath)\nvalidation_generator = DataGenerator(test_datapath)\n\nclass model(Model):\n    def __init__(self):\n        super(model, self).__init__()\n        self.conv1 = Conv2D(filters=2, kernel_size=1, strides=1, padding='same', input_shape=(256, 256, 1))\n        self.act1 = Activation('relu')\n\n    def call(self, input):\n        \"\"\"Run the model.\"\"\"\n        return self.act1(self.conv1(input))\n\nmodel = model()\nmodel.compile(optimizer='adam', loss='mean_squared_error')\nhistory = model.fit_generator(training_generator, epochs=5000, verbose=1, validation_data=(validation_generator),\n                              use_multiprocessing=False)\n<\/code><\/pre>\n\n<p>Running that code gives the following error:<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\/Users\/...py\", line 23, in &lt;module&gt;\n    use_multiprocessing=False)\n  File \"C:\\Users\\...\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 2161, in fit_generator\n    '`fit_generator` is not yet enabled for unbuilt Model subclasses')\nNotImplementedError: `fit_generator` is not yet enabled for unbuilt Model subclasses\n<\/code><\/pre>\n\n<p>I've tried to find some hints, if I really try to do something, that isn't intended, but it's seems unlogical to me, since this way is very comfortable for defining more complex nets and since I've allready worked with pytorch, I'm pretty sure, this should work with tf and keras also.<\/p>\n\n<p>There is a <a href=\"https:\/\/stackoverflow.com\/questions\/50530460\/notimplementederror-fit-generator-is-not-yet-enabled-for-unbuilt-model-subcla\">thread<\/a> with the same error, but for a sequential net implementation, which ist obviously not, what I'm aiming for.<\/p>\n\n<p>If I'm using a direct import for keras, I'm getting a different error:<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\/Users\/...py\", line 23, in &lt;module&gt;\n    use_multiprocessing=False)\n  File \"C:\\Users\\...\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"C:\\Users\\...\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\engine\\training.py\", line 1418, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"C:\\Users\\...\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\engine\\training_generator.py\", line 40, in fit_generator\n    model._make_train_function()\n  File \"C:\\Users\\...\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\keras\\engine\\training.py\", line 496, in _make_train_function\n    raise RuntimeError('You must compile your model before using it.')\nRuntimeError: You must compile your model before using it.\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>But I'm compiling the model one row above calling the model.fit_generator() function ...\nAnd when I'm not completely wrong, I've also given the correct input-shape, so that shouldn't be the problem either ...<\/p>\n\n<p>Working with windows, python 3.6, tensorflow-gpu 1.12.<\/p>\n","answers":[{"AnswerId":"53532448","CreationDate":"2018-11-29T05:31:32.697","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You're not creating a model anywhere. <\/p>\n\n<p>You need:   <\/p>\n\n<pre><code>model = Model(inputTensors, outputTensors)\n<\/code><\/pre>\n\n<p>Or at least at some point in your class <code>super(model,self).__init__(inputTensors,outputTensors)<\/code>.<\/p>\n\n<p>Ideally:<\/p>\n\n<pre><code>def createModel():\n    inputs = Input((256,256,1))\n    outputs = Conv2D(filters=2, kernel_size=1, strides=1, \n                     padding='same', activation='relu')(inputs)\n    return Model(inputs,outputs)\n<\/code><\/pre>\n\n<p>If you do want to have a subclass for some reason:<\/p>\n\n<pre><code>class MyModel(Model):\n    def init(self):\n        inputs = Input((256,256,1))\n        self.conv1 = Conv2D(filters=2, kernel_size=1, strides=1, padding='same')\n        self.act1 = Activation('relu')\n        outputs = self.act1(self.conv1(inputs))\n\n        super(MyModel,self).__init__(inputs,outputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":53531824,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><text-classification>","CreationDate":"2018-11-29T04:26:08.707","AcceptedAnswerId":53537125.0,"OwnerUserId":7942047.0,"Title":"Making a prediction after training and fitting a RNN Sequential model","Body":"<p>I am trying to get predictions from my sentiment analysis models that classify 500 worded News articles. The models validation loss and training loss is in are about the same and their scores are relatively high. However when I try to make predictions with them I get the same classification result in all of them regardless of the text input. \nI believe that the problem might be on the way I am trying to make a prediction (I pad my string with spaced characters). I was hoping that someone here could shed some light on this issue (my code below). Thank you for your help<\/p>\n\n<pre><code>comment = 'SAMPLE TEXT STRING'\nfor i in range(300-len(comment.split(' '))):\n    apad += ' A'\ncomment = comment + apad\ntok.fit_on_texts([comment])\nX = tokenizer.texts_to_sequences([comment])\nX = preprocessing.sequence.pad_sequences(X)\nyhat = b.predict_classes(X)\nprint(yhat)\nprediction = b.predict(X, batch_size=None, verbose=0, steps=None)\nprint(prediction)\n<\/code><\/pre>\n\n<p>The output of this script is below. Both prediction and predicted classes, are regardless of the text input always 0 for some reason:<\/p>\n\n<p>[[0]] [[0.00645966]]<\/p>\n","answers":[{"AnswerId":"53537125","CreationDate":"2018-11-29T10:41:35.950","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The problem seems to be with the tokenizer. <\/p>\n\n<p>You can't <code>fit<\/code> the tokenizer again, because you will have different tokens for each word. You should fit the tokenizer only once before training and then save the tokens to be used with all new text. <\/p>\n"}]}
+{"QuestionId":53532352,"AnswerCount":2,"Tags":"<machine-learning><dataset><conv-neural-network><pytorch>","CreationDate":"2018-11-29T05:22:05.893","AcceptedAnswerId":53533140.0,"OwnerUserId":10720610.0,"Title":"How do I split the training dataset into training, validation and test datasets?","Body":"<p>I have a custom data set of images and its target. I have created a training data set in PyTorch. I want to split it into 3 parts: training, validation and test. How do I do it?<\/p>\n","answers":[{"AnswerId":"53533140","CreationDate":"2018-11-29T06:35:18.447","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Once you have the \"master\" dataset you can use <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.Subset\" rel=\"nofollow noreferrer\"><code>data.Subset<\/code><\/a> to split it.<br>\nHere's an example for random split<\/p>\n\n<pre><code>import torch\nfrom torch.utils import data\nimport random\n\nmaster = data.Dataset( ... )  # your \"master\" dataset\nn = len(master)  # how many total elements you have\nn_test = int( n * .05 )  # number of test\/val elements\nn_train = n - 2 * n_test\nidx = list(range(n))  # indices to all elements\nrandom.shuffle(idx)  # in-place shuffle the indices to facilitate random splitting\ntrain_idx = idx[:n_train]\nval_idx = idx[n_train:(n_train + n_test)]\ntest_idx = idx[(n_train + n_test):]\n\ntrain_set = data.Subset(master, train_idx)\nval_set = data.Subset(master, val_idx)\ntest_set = data.Subset(master, test_idx)\n<\/code><\/pre>\n\n<p>This can also be achieved using <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.random_split\" rel=\"nofollow noreferrer\"><code>data.random_split<\/code><\/a>:<\/p>\n\n<pre><code>train_set, val_set, test_set = data.random_split(master, (n_train, n_val, n_test))\n<\/code><\/pre>\n"},{"AnswerId":"55878954","CreationDate":"2019-04-27T09:09:45.840","ParentId":null,"OwnerUserId":"5353461","Title":null,"Body":"<p>Given parameter <code>train_frac=0.8<\/code>, this function will split the <code>dataset<\/code> into 80%, 10%, 10%:<\/p>\n\n<pre><code>import torch, itertools\nfrom torch.utils.data import TensorDataset\n\ndef dataset_split(dataset, train_frac):\n    '''\n    param dataset:    Dataset object to be split\n    param train_frac: Ratio of train set to whole dataset\n\n    Randomly split dataset into a dictionary with keys, based on these ratios:\n        'train': train_frac\n        'valid': (1-split_frac) \/ 2\n        'test': (1-split_frac) \/ 2\n    '''\n    assert split_frac &gt;= 0 and split_frac &lt;= 1, \"Invalid training set fraction\"\n\n    length = len(dataset)\n\n    # Use int to get the floor to favour allocation to the smaller valid and test sets    \n    train_length = int(length * train_frac)\n    valid_length = int((length - train_length) \/ 2)\n    test_length  = length - train_length - valid_length\n\n    dataset = random_split(dataset, (train_length, valid_length, test_length))\n    dataset = {name: set for name, set in zip(('train', 'valid', 'test'), sets)}\n    return dataset\n<\/code><\/pre>\n"}]}
+{"QuestionId":53532381,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook><object-detection>","CreationDate":"2018-11-29T05:25:35.710","AcceptedAnswerId":null,"OwnerUserId":10486968.0,"Title":"No module named object_detection on jupyter notebook","Body":"<p>I am running Ubuntu on a Virtual Machine. I installed object_detection through the terminal in the directory ~\/models\/research\/<\/p>\n\n<pre><code>protoc object_detection\/protos\/*.proto --python_out=.\n<\/code><\/pre>\n\n<p>Then adding libraries to PYTHONPATH (still in the same directory ~\/models\/research\/)<\/p>\n\n<pre><code>export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`\/slim\n<\/code><\/pre>\n\n<p>And (still in the same directory ~\/models\/research\/) successfully tested that it is installed using:<\/p>\n\n<pre><code>python object_detection\/builders\/model_builder_test.py\n<\/code><\/pre>\n\n<p>But when I try to import it in jupyter notebook, it returns the ModuleNotFoundError: No module named 'object_detection'.<\/p>\n\n<p>Note: I run my jupyter notebook through the terminal in the ~ directory. And python version is 3.6<\/p>\n\n<p>Also, don't know if this is related, but whenever I restart my virtual machine and run the code above to test installation of object_detection, it fails. And I have to re-install it using the above steps.<\/p>\n","answers":[{"AnswerId":"53534078","CreationDate":"2018-11-29T07:43:52.360","ParentId":null,"OwnerUserId":"10056036","Title":null,"Body":"<p>I faced similar issue. Jupyter PYTHONPATH appears to be different. I add the following to the start of the script.<\/p>\n\n<pre><code>import sys\nsys.path.append(\".\")\n<\/code><\/pre>\n\n<p>This should work.<\/p>\n"}]}
+{"QuestionId":53532666,"AnswerCount":1,"Tags":"<deep-learning><caffe>","CreationDate":"2018-11-29T05:51:48.517","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"How netwrok layers are defined as output in Caffe's net.cpp?","Body":"<p>I roughly know that network's outputs are defined in <code>net.cpp Init<\/code>.<\/p>\n\n<p><code>void Net&lt;Dtype&gt;::Init(const NetParameter&amp; in_param)<\/code><\/p>\n\n<p>Exactly is here at line 240.<\/p>\n\n<pre><code>\/\/ In the end, all remaining blobs are considered output blobs.\n  for (set&lt;string&gt;::iterator it = available_blobs.begin();\n      it != available_blobs.end(); ++it) {\n    LOG_IF(INFO, Caffe::root_solver())\n        &lt;&lt; \"This network produces output \" &lt;&lt; *it;\n    net_output_blobs_.push_back(blobs_[blob_name_to_idx[*it]].get());\n    net_output_blob_indices_.push_back(blob_name_to_idx[*it]);\n  }\n<\/code><\/pre>\n\n<p>I have a customized network using Inception Net for faster-rcnn approach.<\/p>\n\n<p>My network has five convolution layers for RPN.\nThen RCNN has two inception layers and one FC layer.<\/p>\n\n<p>Now the problem is second inception layer is considered in outputs.\nLog file shows as<\/p>\n\n<pre><code>I1129 13:21:48.189395 24653 net.cpp:243] This network produces output loss_bbox\nI1129 13:21:48.189399 24653 net.cpp:243] This network produces output loss_cls\nI1129 13:21:48.189404 24653 net.cpp:243] This network produces output rpn_cls_loss\nI1129 13:21:48.189407 24653 net.cpp:243] This network produces output rpn_loss_bbox\nI1129 13:21:48.189411 24653 net.cpp:243] This network produces output tower_2\/conv\nI1129 13:21:48.189415 24653 net.cpp:243] This network produces output tower_2\/tower1\/conv1_1\nI1129 13:21:48.189419 24653 net.cpp:243] This network produces output tower_2\/tower2\/conv2_2\n<\/code><\/pre>\n\n<p>I don't understand why second inception layer is considered as output.<\/p>\n\n<p>What is wrong with my network architecture?<\/p>\n\n<p>My whole network can be seen <a href=\"https:\/\/www.dropbox.com\/s\/2hkzem76klos7ly\/train.txt?dl=0\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"54554396","CreationDate":"2019-02-06T13:07:57.527","ParentId":null,"OwnerUserId":"213615","Title":null,"Body":"<p>Your second inception layer's final output is the ReLU activation:<\/p>\n\n<pre><code>layer {\n  name: \"tower_2\/tower2\/conv2_2_relu\"\n  type: \"ReLU\"\n  bottom: \"tower_2\/tower2\/conv2_2\"\n  top: \"tower_2\/tower2\/conv2_2\"\n}\n<\/code><\/pre>\n\n<p>Which produces a <code>top<\/code> of <code>tower_2\/tower2\/conv2_2<\/code>. This top isn't connected to any other layer as a bottom and hence automatically considered as an output.<\/p>\n"}]}
+{"QuestionId":53533091,"AnswerCount":1,"Tags":"<r><keras>","CreationDate":"2018-11-29T06:31:56.820","AcceptedAnswerId":null,"OwnerUserId":8039357.0,"Title":"How to use specific GPU device for learning in keras R?","Body":"<p>I`m Da-Bin.<\/p>\n\n<p>I want to use specific GPU device for learning in keras R not python.\nBecause when I tried learning two program simultaneously, one program is good working but another program is not working.\nI think that it seems another program is waiting until end of learning.\nSo, I have two GPU 1080 ti, I want use specific device for each program.\nBut, multi_gpu_model function use When use the more than two device, right?<\/p>\n\n<p>If i can use the multi_gpu_model function for learning using one device.\nHow can I know device name for parameter \"gpu=\" name ?<\/p>\n\n<p>And, how can I use the specific device for each program?<\/p>\n","answers":[{"AnswerId":"53534238","CreationDate":"2018-11-29T07:56:05.000","ParentId":null,"OwnerUserId":"10454790","Title":null,"Body":"<p>By default gpu names are \"\/gpu:0\", \"\/gpu:1\" ... \"\/gpu:(n-1)\" if you have n gpus. <\/p>\n\n<p>You can pass a list of gpu names instead of an integer for gpus parameter of multi_gpu_model function. <\/p>\n"}]}
+{"QuestionId":53533974,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2018-11-29T07:37:20.803","AcceptedAnswerId":53535142.0,"OwnerUserId":4986593.0,"Title":"How do I get Keras to train a model on a specific GPU?","Body":"<p>There is a shared server with 2 GPUs in my institution. Suppose there are two team members each wants to train a model at the same time, then how do they get Keras to train their model on a specific GPU so as to avoid resource conflict?<\/p>\n\n<p>Ideally, Keras should figure out which GPU is currently busy training a model and then use the other GPU to train the other model. However, this doesn't seem to be the case. It seems that by default Keras only uses the first GPU (since the <code>Volatile GPU-Util<\/code> of the second GPU is always 0%).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Jg1P0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Jg1P0.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53535142","CreationDate":"2018-11-29T08:58:56.480","ParentId":null,"OwnerUserId":"6280540","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/q\/52444944\/6280540\">Possibly duplicate with my previous question<\/a><\/p>\n\n<p>It's a bit more complicated. Keras will the memory in both GPUs althugh it will only use one GPU by default. Check <code>keras.utils.multi_gpu_model<\/code> for using several GPUs.<\/p>\n\n<p>I found the solution by choosing the GPU using the environment variable CUDA_VISIBLE_DEVICES.<\/p>\n\n<p>You can add this manually before importing keras or tensorflow to choose your gpu<\/p>\n\n<pre><code>os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\" # first gpu\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\" # second gpu\n<\/code><\/pre>\n\n<p>To make it automatically, I made a function that parses <code>nvidia-smi<\/code> and detects automatically which GPU is being already used and sets the appropriate value to the variable.<\/p>\n"}]}
+{"QuestionId":53534631,"AnswerCount":1,"Tags":"<keras><installation><deep-learning><anaconda><data-science>","CreationDate":"2018-11-29T08:25:54.493","AcceptedAnswerId":null,"OwnerUserId":10690142.0,"Title":"Problem while installing keras through anaconda prompt?","Body":"<p>I installed keras by using this action on anaconda prompt \"conda install -c conda-forge keras\". However, when installing Keras the anaconda prompt closed. Now, whenever I try to start the anaconda prompt it closes after 1 second and displays a short message for a short time<br><br><a href=\"https:\/\/i.stack.imgur.com\/axOsu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/axOsu.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53542040","CreationDate":"2018-11-29T15:10:09.183","ParentId":null,"OwnerUserId":"10721407","Title":null,"Body":"<p>pip install keras is always safest.<\/p>\n\n<p>To solve your problem, as the error says, download the visual studio community from <a href=\"https:\/\/visualstudio.microsoft.com\/downloads\/\" rel=\"nofollow noreferrer\">https:\/\/visualstudio.microsoft.com\/downloads\/<\/a><\/p>\n\n<p>When done, open visual studio installer, under visual studio community you should have a 'modify' button, install all C++ workloads and individual components depending on your machine.<\/p>\n\n<p>You should then be able to access anaconda application by typing for example 'jupyter notebook' in the prompt. normally.<\/p>\n"}]}
+{"QuestionId":53535198,"AnswerCount":1,"Tags":"<python><tensorflow><keras><ocr><rnn>","CreationDate":"2018-11-29T09:02:23.717","AcceptedAnswerId":null,"OwnerUserId":10689350.0,"Title":"How can I use Keras OCR example to inference a new image?","Body":"<p>I am trying to implement a OCR project by Keras.So I try to learn from <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/image_ocr.py\" rel=\"nofollow noreferrer\">Keras OCR example<\/a>.I have use my own train data to train a new model and get the .H5 modelfile.\nNow I want to test a new image to see my model performance\uff0cso I code a\n test.py like this:  <\/p>\n\n<pre><code>from keras.models import Model\nimport cv2\nfrom keras.preprocessing.image import img_to_array\nimport numpy as np\nfrom keras.models import load_model\nfrom keras import backend as K\nfrom allNumList import alphabet\n\ndef labels_to_text(labels):\n    ret = []\n    for c in labels:\n        if c == len(alphabet):  # CTC Blank\n            ret.append(\"\")\n        else:\n            ret.append(alphabet[c])\n    return \"\".join(ret)\n\ndef decode_predict_ctc(out, top_paths = 1):\n    results = []\n    beam_width = 5\n    if beam_width &lt; top_paths:\n      beam_width = top_paths\n    for i in range(top_paths):\n      lables = K.get_value(K.ctc_decode(out, input_length=np.ones(out.shape[0])*out.shape[1],\n                           greedy=False, beam_width=beam_width, top_paths=top_paths)[0][i])[0]\n      text = labels_to_text(lables)\n      results.append(text)\n    return results\n\ndef test(modelPath,testPicTest):\n    img=cv2.imread(testPicTest)\n    img=cv2.resize(img,(128,64))\n    img=img_to_array(img)\n    img=np.array(img,dtype='float')\/255.0\n    img=np.expand_dims(img, axis=0)\n    img=img.swapaxes(1,2)   \n\n    model=load_model(modelPath,custom_objects = {'&lt;lambda&gt;': lambda y_true, y_pred: y_pred})\n    net_out_value = model.predict(img)\n    top_pred_texts = decode_predict_ctc(net_out_value)\n    return top_pred_texts\n\nresult=test(r'D:\\code\\testAndExperiment\\py\\KerasOcr\\weights.h5',r'D:\\code\\testAndExperiment\\py\\KerasOcr\\test\\avo.jpg') \nprint(result)  \n<\/code><\/pre>\n\n<p>but I get a error like this:<\/p>\n\n<pre><code>Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 4 array(s), but instead got the following list of 1 arrays: [array([[[[1., 1., 1.],          [1., 1., 1.],          [1., 1., 1.],          ...,          [1., 1., 1.],          [1., 1., 1.],          [1., 1., 1.]],          [[1., 1., 1.],          [1., 1., 1.],...\n<\/code><\/pre>\n\n<p>I have references some material\uff1a<br>\n<a href=\"https:\/\/stackoverflow.com\/a\/49537697\/10689350\">https:\/\/stackoverflow.com\/a\/49537697\/10689350<\/a><br>\n<a href=\"https:\/\/www.dlology.com\/blog\/how-to-train-a-keras-model-to-recognize-variable-length-text\/\" rel=\"nofollow noreferrer\">https:\/\/www.dlology.com\/blog\/how-to-train-a-keras-model-to-recognize-variable-length-text\/<\/a><br>\n<a href=\"https:\/\/stackoverflow.com\/questions\/46461701\/how-to-predict-the-results-for-ocr-using-keras-image-ocr-example\">How to predict the results for OCR using keras image_ocr example?<\/a>   <\/p>\n\n<p>some answer show that we should use 4 inputs <code>[input_data, labels, input_length, label_length]<\/code> in training but besides <code>input_data<\/code>, everything else is information used only for calculating the loss,so in testing maybe use the input_data is enough.So I just use a picture without <code>labels, input_length, label_length<\/code>.But I get the error above.<\/p>\n\n<p>I am confused about if the model needs 4 inputs or 1 in testing?<br>\nIt doesn't seem reasonable to require 4 inputs during the testing process.and now I have model.h5,what should I do next?<br>\nThanks in advance.<\/p>\n\n<p>My code is Here:<a href=\"https:\/\/github.com\/hqabcxyxz\/KerasOCR\/tree\/master\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hqabcxyxz\/KerasOCR\/tree\/master<\/a><\/p>\n","answers":[{"AnswerId":"53551206","CreationDate":"2018-11-30T04:22:58.750","ParentId":null,"OwnerUserId":"10689350","Title":null,"Body":"<p>maybe I know why.Because in the OCR example,we make a lambda layer to count CTC loss.This Layer need 4 inputs!\nThe right way to do test is we make a  model without this lambda layer during inference.Then load the model weight by name to do inference.After we get inference result,just use CTC decode it!\nI will update my code in github later.....<\/p>\n"}]}
+{"QuestionId":53535805,"AnswerCount":1,"Tags":"<python><tensorflow><keras><nlp><sentiment-analysis>","CreationDate":"2018-11-29T09:34:12.657","AcceptedAnswerId":53537068.0,"OwnerUserId":7810784.0,"Title":"Making Predictions on single review from input text using saved CNN model","Body":"<p>I am making a classifier based on a CNN model in Keras.<\/p>\n\n<p>I will use it in an application, where the user can load the application and enter input text and the model will be loaded from the weights and make predictions.<\/p>\n\n<p>The thing is I am using GloVe embeddings as well and the CNN model uses padded text sequences as well.<\/p>\n\n<p>I used Keras tokenizer as following:<\/p>\n\n<pre><code>tokenizer = text.Tokenizer(num_words=max_features, lower=True, char_level=False)\ntokenizer.fit_on_texts(list(train_x))\n\ntrain_x = tokenizer.texts_to_sequences(train_x)\ntest_x = tokenizer.texts_to_sequences(test_x)\n\ntrain_x = sequence.pad_sequences(train_x, maxlen=maxlen)\ntest_x = sequence.pad_sequences(test_x, maxlen=maxlen)\n<\/code><\/pre>\n\n<p>I trained the model and predicted on test data, but now I want to test the same with loaded model which I loaded and working.<\/p>\n\n<p>But my problem here is If I provide a single review, it has to be passed through the <code>tokeniser.text_to_sequences()<\/code> which is returning 2D array, with a shape of <code>(num_chars, maxlength)<\/code> and hence followed by a <code>num_chars<\/code> predictions, but I need it in <code>(1, max_length)<\/code> shape.<\/p>\n\n<p>I am using the following code for prediction:<\/p>\n\n<pre><code>review = 'well free phone cingular broke stuck not abl offer kind deal number year contract up realli want razr so went look cheapest one could find so went came euro charger small adpat made fit american outlet, gillett fusion power replac cartridg number count packagemay not greatest valu out have agillett fusion power razor'\nxtest = tokenizer.texts_to_sequences(review)\nxtest = sequence.pad_sequences(xtest, maxlen=maxlen)\n\nmodel.predict(xtest)\n<\/code><\/pre>\n\n<p>Output is:<\/p>\n\n<pre><code>array([[0.29289   , 0.36136267, 0.6205081 ],\n       [0.362869  , 0.31441122, 0.539749  ],\n       [0.32059124, 0.3231736 , 0.5552745 ],\n       ...,\n       [0.34428033, 0.3363668 , 0.57663095],\n       [0.43134686, 0.33979046, 0.48991954],\n       [0.22115968, 0.27314988, 0.6188136 ]], dtype=float32)\n<\/code><\/pre>\n\n<p>I need a single prediction here <code>array([0.29289   , 0.36136267, 0.6205081 ])<\/code> as I have a single review.<\/p>\n","answers":[{"AnswerId":"53537068","CreationDate":"2018-11-29T10:38:36.490","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The problem is that you need to pass a list of strings to <code>texts_to_sequences<\/code> method. So you need to put the single review in a list like this:<\/p>\n\n<pre><code>xtest = tokenizer.texts_to_sequences([review])\n<\/code><\/pre>\n\n<p>If you don't do that (i.e. pass a string, not a list of string(s)), considering the strings in Python are iterable, <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/7b7af2690026b51b5a7a50b4b359ff12960ac8b7\/keras_preprocessing\/text.py#L297\" rel=\"nofollow noreferrer\">it would iterate<\/a> over the characters of the given string and consider the characters, not words, as the tokens:<\/p>\n\n<pre><code>oov_token_index = self.word_index.get(self.oov_token)\nfor text in texts:  # &lt;-- it would iterate over the string instead\n    if self.char_level or isinstance(text, list):\n<\/code><\/pre>\n\n<p>That's why you would get an array of shape <code>(num_chars, maxlength)<\/code> as the return value of <code>texts_to_sequences<\/code> method.<\/p>\n"}]}
+{"QuestionId":53536468,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><classification><conv-neural-network>","CreationDate":"2018-11-29T10:08:37.590","AcceptedAnswerId":53536568.0,"OwnerUserId":7066799.0,"Title":"CNN image classification: accuracy values shakes greatly","Body":"<p>I try 2 class (dog\/cat) classification with cnn.\nBut I found its graph of training is strange.\nWhy accuracy values shakes greatly? And is it correct training?<\/p>\n\n<p>optimizer: adam\nlearning rate: 1e-4<\/p>\n\n<p>network: <a href=\"https:\/\/gist.github.com\/elect000\/130acbdb0a3779910082593db4296254\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/elect000\/130acbdb0a3779910082593db4296254<\/a><a href=\"https:\/\/i.stack.imgur.com\/rNvaq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rNvaq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>optimizer: adam\nlearning rate: 1e-6\n<a href=\"https:\/\/i.stack.imgur.com\/0vYVn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0vYVn.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53536568","CreationDate":"2018-11-29T10:13:20.903","ParentId":null,"OwnerUserId":"10719567","Title":null,"Body":"<p>Likely your learning rate is too high. \nWhen the learning rate is too high, the network takes large leaps when changing the weights, and this can cause it to overshoot the local minimum it's approaching.<\/p>\n\n<p>Have a read of this article for a better description, and a nice diagram:\n<a href=\"https:\/\/www.quora.com\/In-an-artificial-neural-network-algorithm-what-happens-if-my-learning-rate-is-wrong-too-high-or-too-low\" rel=\"nofollow noreferrer\">https:\/\/www.quora.com\/In-an-artificial-neural-network-algorithm-what-happens-if-my-learning-rate-is-wrong-too-high-or-too-low<\/a><\/p>\n"}]}
+{"QuestionId":53536672,"AnswerCount":0,"Tags":"<c++><ios><objective-c><tensorflow>","CreationDate":"2018-11-29T10:18:22.947","AcceptedAnswerId":null,"OwnerUserId":4511899.0,"Title":"No OPKernel was registered to support Op 'Equal' with these attrs","Body":"<p>Im getting the following error when trying to load my TensorFlow model, im using the following tools:<\/p>\n\n<p>Xcode (Latest)<\/p>\n\n<p>TensorFlow Experimental V1.1.1 (Installed using pods)<\/p>\n\n<p>C++<\/p>\n\n<p>Im able to create the TensorFlow session with no issues, and I'm able to read the TensorFlow model into a GraphDef, but when I try to create the model using the active TensorFlow session I get the following issue.<\/p>\n\n<blockquote>\n  <p>Invalid argument: No OpKernel was registered to support Op 'Equal' with these attrs.  <\/p>\n  \n  <p>Registered devices: <\/p>\n  \n  <p>[CPU]<\/p>\n  \n  <p>Registered kernels: <\/p>\n  \n  <p>device='CPU'; T in [DT_FLOAT] [[Node: Equal = Equal[T=DT_INT32, _output_shapes=[[]]](Rank, Equal)]]<\/p>\n<\/blockquote>\n\n<p>This is how my code currently looks<\/p>\n\n<pre><code>tensorflow::Session* sessionPointer = nullptr;\nconst tensorflow::SessionOptions options;\n\ntensorflow::Status newSessionStatus = tensorflow::NewSession(options, &amp;sessionPointer);\n\n[MachineLearning checkTensorflowStatus:newSessionStatus];\n\nstd::unique_ptr&lt;tensorflow::Session&gt; session(sessionPointer);\n\n\/\/ Create a point to the tensorflow graph\ntensorflow::GraphDef tensorflowGraph;\n\n\/\/ Get the model file path\nNSString* filePath = FilePathForResourceName(@\"frozen_inference_graph\", @\"pb\");\n\n\/\/ Read the data into the TensorflowGraph\nauto fileConverteredToProto = PortableReadFileToProto([filePath UTF8String], &amp;tensorflowGraph);\n\n\/\/ Create the tensorflow graph\ntensorflow::Status createGraphStatus = session-&gt;Create(tensorflowGraph);\n\n\/\/ Check the status of the tensorflow graph\n[MachineLearning checkTensorflowStatus:createGraphStatus];\n<\/code><\/pre>\n\n<p>I have checked that the file path it is getting is correct and it seems fine with loading the model data into the GraphDef. This is how I do it<\/p>\n\n<pre><code>\/\/ Finds the file path for the given file name\nNSString* FilePathForResourceName(NSString* name, NSString* extension)\n{\n    NSString* filePath = [[NSBundle mainBundle] pathForResource:name ofType:extension];\n    if (filePath == NULL)\n    {\n        LOG(ERROR) &lt;&lt; \"Could not find the TensorFlow model\";\n    }\n    return filePath;\n}\n\n\/\/ Reads the file into the protobuf. This was taken from the TensorFlow example code\nbool PortableReadFileToProto(const std::string&amp; fileName, ::google::protobuf::MessageLite* proto)\n{\n    \/\/ Create the input stream with the given file name\n    ::google::protobuf::io::CopyingInputStreamAdaptor stream(new IfstreamInputStream(fileName));\n\n    \/\/ Set that this object owns the copying stream\n    stream.SetOwnsCopyingStream(true);\n\n    \/\/ Create a coded input stream\n    ::google::protobuf::io::CodedInputStream codedStream(&amp;stream);\n\n    \/\/ Set the byte limit, aparently having protos too big can causes issues, so we limit it.\n    codedStream.SetTotalBytesLimit(1024LL &lt;&lt; 20, 512LL &lt;&lt; 20);\n\n    \/\/ Parse the coded stream\n    return proto-&gt;ParseFromCodedStream(&amp;codedStream);\n}\n<\/code><\/pre>\n\n<p>I did some reading about OP's not been supported in other posts, but Equal is already supported but it doesn't work in the way that I need it too for my TensorFlow model.<\/p>\n\n<p>I was also reading this guide on how to add a new OP.\n<a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">Tensorflow how to add a new OP<\/a><\/p>\n\n<p>But I don't really understand how it all works, so im struggling to get my model running.<\/p>\n\n<p>I would have preferred to convert my TensorFlow model to CoreML and used the Swift CoreML tools, but there is an issue with the conversion, where the Slice OP isn't fully supported and it can't convert my model.<\/p>\n","answers":[]}
+{"QuestionId":53537064,"AnswerCount":1,"Tags":"<python><keras><generator>","CreationDate":"2018-11-29T10:38:23.937","AcceptedAnswerId":53537771.0,"OwnerUserId":6215584.0,"Title":"Keras: Correct use of fit_generator, predict_generator, and evaluate_generator","Body":"<p>I'm encountering weird behaviour when using <code>fit_generator<\/code>, <code>predict_generator<\/code>, and <code>evaluate_generator<\/code>, and I would like to ask the following questions, for which I could find no answer in the documentation:<\/p>\n\n<ol>\n<li>Is it ok to have batches of different sizes when using <code>fit_generator<\/code>?<\/li>\n<\/ol>\n\n<p>My batches are defined time-wise: they group the events that took place in the same hour. Each batch can therefore group a different number of events. For clarity, this is how my generators look like (following the logic in <a href=\"https:\/\/stackoverflow.com\/questions\/46647008\/keras-aggregated-objective-function\">this thread<\/a>):<\/p>\n\n<pre><code>def grouper(g,x,y):\n     while True:\n        for gr in g.unique():\n            # this assigns indices to the entire set of values in g,\n            # the subsects to all the rows in which g == gr\n            indices = g == gr\n            yield (x[indices],y[indices])\n\nall_data_generator = grouper(df['batch_id'], X, Y)\ntrain_generator = grouper(df.loc[df['set'] == 'train', 'batch_id'], X_train, Y_train)\nvalidation_generator = grouper(df.loc[df['set'] == 'val', 'batch_id'], X_val, Y_val)\ntest_generator = grouper(df.loc[df['set'] == 'test', 'batch_id'], X_test, Y_test)\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Is it ok to have different number of batches in <code>train_generator<\/code> and <code>validation_generator<\/code>? <\/li>\n<\/ol>\n\n<p>For clarity, I pass those two (different) numbers explicitly to <code>fit_generator<\/code> in the call:<\/p>\n\n<pre><code>train_batches = df.loc[df['set'] == 'train', 'batch_id'].nunique()\nval_batches = df.loc[df['set'] == 'val', 'batch_id'].nunique()\n\nhistory = fmodel.fit_generator(train_generator, \n                             steps_per_epoch=train_batches, \n                             validation_data=validation_generator,\n                             validation_steps=val_batches,\n                             epochs=20, verbose = 0)\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>Predictions are wildly different depending on whether I use <code>predict_classes<\/code> or <code>predict_generator<\/code>, which baffles me.<\/li>\n<\/ol>\n\n<p>Here's the code:<\/p>\n\n<pre><code>df['pred'] = fmodel.predict_classes(X)\n\n# returns different results from\ntotal_batches = df['batch_id'].nunique()\ndf['pred_gen'] = fmodel.predict_generator(all_data_generator, steps = total_batches)\n<\/code><\/pre>\n\n<ol start=\"4\">\n<li>Similarly, <code>evaluate<\/code> and <code>evaluate_generator<\/code> return different results.<\/li>\n<\/ol>\n\n<p>The code:<\/p>\n\n<pre><code>scores = model.evaluate(X_test, Y_test, verbose = 0)\n\n# returns different results from\nscores_generator = fmodel.evaluate_generator(test_generator, steps=test_batches)\n<\/code><\/pre>\n\n<p>I know there are already many issues referring to my points 3. and 4. (e.g., <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3477\" rel=\"nofollow noreferrer\">3477<\/a>, <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6499\" rel=\"nofollow noreferrer\">6499<\/a>) but the main takeaways there seem to refer to <\/p>\n\n<ul>\n<li>using ImageDataGenerator with\/without rescaling;<\/li>\n<li>shuffling data, which according to <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">fit_generator documentation<\/a> \"Has no effect when steps_per_epoch is not None.\"<\/li>\n<li>using <code>workers &gt; 1<\/code>, which by default is not the case.<\/li>\n<\/ul>\n\n<p>So I'm wondering whether points 1. and 2. might be the culprits here.<\/p>\n","answers":[{"AnswerId":"53537771","CreationDate":"2018-11-29T11:14:49.730","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h1>1 and 2<\/h1>\n\n<p>Yes, totally ok.<br>\nIt's even expected that 2 be true. <\/p>\n\n<h1>3<\/h1>\n\n<p>Predict classes is not documented. What does it do exactly? I think it predicts indices while all other prediction methods predict the actual model's outputs, right?<\/p>\n\n<h1>4<\/h1>\n\n<p>This is sensible... <\/p>\n\n<p>Are you pretty sure your generator is outputting exactly what you want?   <\/p>\n\n<p>You may try to see a few batches to compare them with <code>x<\/code> and <code>y<\/code>:<\/p>\n\n<pre><code>for i in range(aFewBatches):\n    print(next(train_generator))\n    #or create some comparisons\n<\/code><\/pre>\n\n<p>Even if the generators are correct, you are definitely reshuffling (actually sorting) your data, when you select your batches for the generator. <\/p>\n\n<p>While <code>evaluate<\/code> will take the entire <code>x<\/code>, <code>y<\/code> data as it is, usually in batches of 32, <code>evaluate_generator<\/code> will take your selected batches. So, the metrics per batch will certainly vary, and the final result which is a mean of the batch metrics will also be different. So, unless the difference too big, it's ok.<\/p>\n\n<p>PS: I'm not entirely sure whether <code>evaluate<\/code> will give you mean batch metrics or entire data metrics, but <code>evaluate_generator<\/code> will bring mean batch metrics, which is enough for a difference.   <\/p>\n"}]}
+{"QuestionId":53538138,"AnswerCount":1,"Tags":"<python-3.x><pytorch><rnn><checkpointing>","CreationDate":"2018-11-29T11:35:43.017","AcceptedAnswerId":53539183.0,"OwnerUserId":10722221.0,"Title":"Save and load checkpoint pytorch","Body":"<p>i make a model and save the configuration as:<\/p>\n\n<pre><code>def checkpoint(state, ep, filename='.\/Risultati\/checkpoint.pth'):  \n    if ep == (n_epoch-1):\n        print('Saving state...')\n        torch.save(state,filename)\ncheckpoint({'state_dict':rnn.state_dict()},ep) \n<\/code><\/pre>\n\n<p>and then i want load this configuration :<\/p>\n\n<pre><code>state_dict= torch.load('.\/Risultati\/checkpoint.pth')\n    rnn.state_dict(state_dict)\n<\/code><\/pre>\n\n<p>when i try, this is the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"train.py\", line 288, in &lt;module&gt;\nrnn.state_dict(state_dict)\nFile \"\/home\/marco\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 593, in state_dict\ndestination._metadata[prefix[:-1]] = dict(version=self._version)\nAttributeError: 'dict' object has no attribute '_metadata'\n<\/code><\/pre>\n\n<p>where i do wrong?<\/p>\n\n<p>thx in advance<\/p>\n","answers":[{"AnswerId":"53539183","CreationDate":"2018-11-29T12:33:20.997","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You need to load <code>rnn.state_dict()<\/code> stored in the <em>dictionary<\/em> you loaded:<\/p>\n\n<pre><code>rnn.load_state_dict(state_dict['state_dict'])\n<\/code><\/pre>\n\n<p>Look at <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.load_state_dict\" rel=\"nofollow noreferrer\"><code>load_state_dict<\/code><\/a> method for more information.<\/p>\n"}]}
+{"QuestionId":53539040,"AnswerCount":0,"Tags":"<tensorflow><rnn>","CreationDate":"2018-11-29T12:25:46.943","AcceptedAnswerId":null,"OwnerUserId":8940996.0,"Title":"tf.nn.bidirectional_dynamic_rnn returns single value in Tensorflow","Body":"<p>tf.nn.bidirectional_dynamic_rnn is usually used as follows.<\/p>\n\n<pre><code> encoder_outputs, encoder_state = tf.nn.bidirectional_dynamic_rnn(cell_fw=fw_cell, cell_bw=bw_cell, \n                                                                  inputs=inputs, sequence_length=sequence_size,\n                                                                  dtype=tf.float64)\n<\/code><\/pre>\n\n<p>However, when I see the following code, I'm not sure what value is returned in _.<\/p>\n\n<pre><code>output = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, dtype=tf.float32,\n                                          inputs = X, sequence_length = seq_len)\n_, ((_, output_fw), (_, output_bw)) = output\n<\/code><\/pre>\n\n<p>Can you tell me what is returned in _?<\/p>\n","answers":[]}
+{"QuestionId":53539348,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><caffe><pytorch>","CreationDate":"2018-11-29T12:42:39.177","AcceptedAnswerId":53539777.0,"OwnerUserId":7677894.0,"Title":"What is the backward process of max operation in deep learning?","Body":"<p>I know that the backward process of deep learning follows the gradient descent algorithm. However, there is never a gradient concept for <code>max<\/code> operation.<\/p>\n\n<p>How does deep learning frameworks like tensorflow, pytorch deal with the backward of 'max' operation like <code>maxpooling<\/code>?<\/p>\n","answers":[{"AnswerId":"53539777","CreationDate":"2018-11-29T13:08:32.077","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>You have to think of  what the <code>max<\/code> operator actually does? That is:<\/p>\n\n<ul>\n<li><strong>It returns or lets better say it propagates the maximum.<\/strong> <\/li>\n<\/ul>\n\n<p>And that's exactly what it does here - it takes two or more tensors and propagates forward <em>(only)<\/em> the maximum.<\/p>\n\n<p>It is often helpful to take a look at a short example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>t1 = torch.rand(10, requires_grad=True)\nt2 = torch.rand(10, requires_grad=True)\n\n\ns1 = torch.sum(t1)\ns2 = torch.sum(t2)\nprint('sum t1:', s1, 'sum t2:', s2)\nm = torch.max(s1, s2)\nprint('max:', m, 'requires_grad:', m.requires_grad)\nm.backward()\nprint('t1 gradients:', t1.grad)\nprint('t2 gradients:', t2.grad)\n<\/code><\/pre>\n\n<p>This code creates two random tensors sums them up and puts them through a max function. Then <code>backward()<\/code> is called upon the result.<\/p>\n\n<p>Lets take a look at the two possible outcomes:<\/p>\n\n<ul>\n<li><p>Outcome 1 - sum of <code>t1<\/code> is larger:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sum t1: tensor(5.6345) sum t2: tensor(4.3965)\nmax: tensor(5.6345) requires_grad: True\nt1 gradients: tensor([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.])\nt2 gradients: tensor([ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.])\n<\/code><\/pre><\/li>\n<li><p>Outcome 2 - sum of <code>t2<\/code> is larger:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sum t1: tensor(3.3263) sum t2: tensor(4.0517)\nmax: tensor(4.0517) requires_grad: True\nt1 gradients: tensor([ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.])\nt2 gradients: tensor([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.])\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>As you would expect in the case <code>s1<\/code> represents the maximum gradients will be calculated for <code>t1<\/code>. Similarly when <code>s2<\/code> is the maximum gradients will be calculated for <code>t2<\/code>. <\/p>\n\n<ul>\n<li><strong>Similarly like in the forward step, back propagation is propagating backwards through the maximum.<\/strong><\/li>\n<\/ul>\n\n<hr>\n\n<p>One thing worth mentioning is that the other <strong>tensors<\/strong> which <strong>do not represent<\/strong> the <strong>maximum are still part of the graph<\/strong>. Only the gradients are set to zero then. If they wouldn't be part of the graph you would get <code>None<\/code> as gradient, instead of a zero vector.<\/p>\n\n<p>You can check what happens if you use python-<code>max<\/code> instead of <code>torch.max<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>t1 = torch.rand(10, requires_grad=True)\nt2 = torch.rand(10, requires_grad=True)\n\n\ns1 = torch.sum(t1)\ns2 = torch.sum(t2)\nprint('sum t1:', s1, 'sum t2:', s2)\nm = max(s1, s2)\nprint('max:', m, 'requires_grad:', m.requires_grad)\nm.backward()\nprint('t1 gradients:', t1.grad)\nprint('t2 gradients:', t2.grad)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sum t1: tensor(4.7661) sum t2: tensor(4.4166)\nmax: tensor(4.7661) requires_grad: True\nt1 gradients: tensor([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.])\nt2 gradients: None \n<\/code><\/pre>\n"}]}
+{"QuestionId":53539529,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><object-detection-api>","CreationDate":"2018-11-29T12:54:33.247","AcceptedAnswerId":null,"OwnerUserId":10512735.0,"Title":"Output predicted bounding boxes trained on TensorFlow Object Detection API","Body":"<p>i used the Google TensorFlow Object detection API [<a href=\"https:\/\/github.com\/tensorflow\/models][1]\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models][1]<\/a> to train on my own dataset using Faster RCNN inception v2 model and bty writing some of my own scripts in python 3. It works fairly well on my videos and now I want to output the predicted bounding boxes to calculate mAP. IS there any way to do this? <\/p>\n\n<p>I have three files generated from training:<\/p>\n\n<ol>\n<li>model.ckpt-6839.data-00000-of-00001<\/li>\n<li>model.ckpt-6839.index<\/li>\n<li>model.ckpt-6839.meta<\/li>\n<\/ol>\n\n<p>Is the predicted boxes contained in one of these files? Or are they stored somewhere else? Or do they need to be coded separately for the coordinates to be extracted?<\/p>\n","answers":[{"AnswerId":"53542589","CreationDate":"2018-11-29T15:39:00.137","ParentId":null,"OwnerUserId":"1621414","Title":null,"Body":"<p>The files you listed are checkpoint files, you can use then to export a frozen graph and then do prediction of input images.<\/p>\n\n<p>Once you obtained the frozen graph, you can then use this file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\" rel=\"nofollow noreferrer\"><code>object_detection_tutorial.ipynb<\/code><\/a> to do prediction of input images.\nIn this file, the function <strong>run_inference_for_single_image<\/strong> will return a output dict for each image and it contains detection boxes in it.<\/p>\n"}]}
+{"QuestionId":53539641,"AnswerCount":2,"Tags":"<tensorflow><base64><serving>","CreationDate":"2018-11-29T13:00:39.580","AcceptedAnswerId":null,"OwnerUserId":4699315.0,"Title":"Tensorflow serving: Unable to base64 decode","Body":"<p>I use the slim package <code>resnet_v2_152<\/code> to train a classification model.\nThen it is exported to .pb file to provide a service.\nBecause the input is image, so it would be encoded with web-safe base64 encoding. It looks like:<\/p>\n\n<p><code>serialized_tf_example = tf.placeholder(dtype=tf.string, name='tf_example')<\/code>\n<code>decoded = tf.decode_base64(serialized_tf_example)<\/code><\/p>\n\n<p>I then encode an image with base64 such that:<\/p>\n\n<p><code>img_path = '\/Users\/wuyanxue\/Desktop\/not_emoji1.jpeg'<\/code><\/p>\n\n<p><code>img_b64 = base64.b64encode(open(img_path, 'rb').read())<\/code><\/p>\n\n<p><code>s = str(img_b64, encoding='utf-8')<\/code><\/p>\n\n<p><code>s = s.replace('+', '-').replace(r'\/', '_')<\/code><\/p>\n\n<p>My post data is as structured as follow:\n<code>post_data = {\n'signature_name': 'predict',\n  'instances':[ { \n      'inputs': \n          { 'b64': s }\n      }]\n}<\/code>\nFinally, I post a HTTP request to this server:<\/p>\n\n<p><code>res = requests.post('server_address', json=post_data)<\/code><\/p>\n\n<p>It gives me:<\/p>\n\n<p><code>'{ \"error\": \"Failed to process element: 0 key: inputs of \\\\\\'instances\\\\\\' list. Error: Invalid argument: Unable to base64 decode\" }'<\/code><\/p>\n\n<p>I want to know how it could be encountered? And are there some solutions for that?<\/p>\n","answers":[{"AnswerId":"53968169","CreationDate":"2018-12-29T09:01:26.033","ParentId":null,"OwnerUserId":"4699315","Title":null,"Body":"<p>This question is already solved.<\/p>\n\n<pre><code>post_data = {\n'signature_name': 'predict',\n'instances':[ { \n  'inputs': \n      { 'b64': s }\n  }]\n}\n<\/code><\/pre>\n\n<p>We see that inputs is with 'b64' flag, which illustrates that tensorflow serving will <strong>decode<\/strong> s with base64 code. \nIt belongs to the tensorflow serving internal method.\nSo, the placeholder:<\/p>\n\n<pre><code>serialized_tf_example = tf.placeholder(dtype=tf.string, name='tf_example')\n<\/code><\/pre>\n\n<p>will directly receive the binary format of the input data BUT NOT base64 format.<\/p>\n\n<p>So, finally, <\/p>\n\n<pre><code>decoded = tf.decode_base64(serialized_tf_example)\n<\/code><\/pre>\n\n<p>is NOT necessary.<\/p>\n"},{"AnswerId":"53942742","CreationDate":"2018-12-27T09:32:23.387","ParentId":null,"OwnerUserId":"1283063","Title":null,"Body":"<p>I had the same issue when using python3. I solved it by adding a 'b' - a byte-like object instead of the default str to the encode function:\n<code>b'{\"instances\" : [{\"b64\": \"%s\"}]}' % base64.b64encode(\n      dl_request.content)<\/code><\/p>\n\n<p>Hope that helps, please see <a href=\"https:\/\/stackoverflow.com\/questions\/8908287\/why-do-i-need-b-to-encode-a-python-string-with-base64\">this answer<\/a> for extra info.<\/p>\n"}]}
+{"QuestionId":53539786,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2018-11-29T13:08:57.390","AcceptedAnswerId":53539911.0,"OwnerUserId":10722313.0,"Title":"Does tensorflow need GPU","Body":"<p>I am trying to install TensorFlow on Linux, Ubuntu 18.10 by following these instructions: <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/<\/a><br>\nI am using python 2.7<\/p>\n\n<p>I first tried to install it using the pip package and after: <\/p>\n\n\n\n<pre class=\"lang-bash prettyprint-override\"><code>pip install --user --upgrade tensorflow  # install in $HOME\n<\/code><\/pre>\n\n<p>I got a message to say that tensorflow 1.12.0 was successfully installed, however when I did the next line: <\/p>\n\n<pre class=\"lang-bash prettyprint-override\"><code>python -c \"import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))\"\n<\/code><\/pre>\n\n<p>I got the error<\/p>\n\n<blockquote>\n  <p>The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine<\/p>\n<\/blockquote>\n\n<p>I then tried to build from source and it failed when I tried to checkout the current branch<\/p>\n\n<pre class=\"lang-bash prettyprint-override\"><code>bazel test -c opt -- \/\/tensorflow\/... -\/\/tensorflow\/compiler\/... -\/\/tensorflow\/lite\/...\n<\/code><\/pre>\n\n<p>I skipped the GPU steps as this said optional but am wondering if it is needed?<\/p>\n","answers":[{"AnswerId":"53539971","CreationDate":"2018-11-29T13:18:50.473","ParentId":null,"OwnerUserId":"934239","Title":null,"Body":"<blockquote>\n  <p>The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine<\/p>\n<\/blockquote>\n\n<p>Of note that this is a <em>warning only<\/em> and would not stop TF from running proprely.<\/p>\n\n<p>TF version you can install from PIP is compiled to be usable on a broad range of systems, but TF can be compiled with system-specific optimizations from source. This warning is intended to remind you of it.<\/p>\n\n<p>See also <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778<\/a><\/p>\n"},{"AnswerId":"53539911","CreationDate":"2018-11-29T13:15:55.577","ParentId":null,"OwnerUserId":"6112394","Title":null,"Body":"<p>Not 100% certain what you have going on but in short no Tensorflow does not require a GPU and you shouldn't have to build it from source unless you just feel like it. <\/p>\n\n<p>Might I suggest you try uninstalling whatever version of Tenforflow you might have, and then reinstall it. <\/p>\n\n<pre><code># Try both of these just to be safe\npip uninstall tensorflow\npip uninstall tensorflow-gpu \n\n# I recommend using conda when possible \nconda install tensorflow\n\n# but if you don't have conda then pip will work just fine\npip install tensorflow\n<\/code><\/pre>\n"}]}
+{"QuestionId":53540305,"AnswerCount":1,"Tags":"<python><tensorflow><reshape><tfrecord>","CreationDate":"2018-11-29T13:36:20.863","AcceptedAnswerId":null,"OwnerUserId":10655065.0,"Title":"Shape error when reading tfrecords with tf.data.TFRecordDataset?","Body":"<p>I've created a tfrecords file with my own images, and when I try to read it using tf.data.TFRecordDataset, a shape error comes:\n<a href=\"https:\/\/i.stack.imgur.com\/ZkPpN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZkPpN.png\" alt=\"enter image description here\"><\/a>\nI created the tfrecords with codes:<\/p>\n\n<pre><code>def _int64_feature(value):\nreturn tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\ndef _bytes_feature(value):\nreturn tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef _float_feature(value):\nreturn tf.train.Feature(float_list=tf.train.FloatList(value=[value]))\n\ndef img_to_tfrecord(data_path):\nwriter = tf.python_io.TFRecordWriter('test_imgs\/test.tfrecords')\n\nfile = open('test_imgs\/test.txt')\nfor line in file.readlines():\n    img_name = line.split(' ')[0]\n    label = int(line.split(' ')[1])\n    img_path = data_path + '\/test_imgs\/' + img_name\n\n    img = Image.open(img_path)\n    img = img.resize((224, 224))\n    img_bytes = img.tobytes()\n\n    feature={'train_img': _bytes_feature(img_bytes),\n             'train_label': _int64_feature(label)}\n    example = tf.train.Example(features=tf.train.Features(feature=feature))\n    writer.write(example.SerializeToString())\n\nwriter.close()\n<\/code><\/pre>\n\n<p>And read it with codes:<\/p>\n\n<pre><code>def parser(record):\nparsed = tf.parse_single_example(record, {'train_img': tf.FixedLenFeature((), tf.string),\n                                          'train_label': tf.FixedLenFeature((), tf.int64)})\n\nimage = tf.decode_raw(parsed['train_img'], tf.uint8)\nimage = tf.reshape(image, [224, 224, 3])\nlabel = tf.cast(parsed['train_label'], tf.int32)\n\nreturn image, label\n\nif __name__ == '__main__':\n    os.environ['CUDA_VISIBLE_DEVICES'] = '-1'\n    tf.reset_default_graph()\n\n    dataset = tf.data.TFRecordDataset('test_imgs\/test.tfrecords')\n    dataset = dataset.map(parser)\n    dataset = dataset.shuffle(buffer_size=6).batch(4).repeat(2)\n    iterator = dataset.make_one_shot_iterator()\n    img, label = iterator.get_next()\n\n    with tf.Session() as sess:\n       a,b=sess.run([img, label])\n       print(a.shape)\n<\/code><\/pre>\n\n<p>Since 150528=224*224*3, how does 200704 comes from\uff1f I've read lots of tutorials but I still can't solve this problem. I've already paid attention to the parsing type of the code: image = tf.decode_raw(parsed['train_img'], tf.uint8). Anyone could help me, I'm about to break down.\nThe images using to create the tfrecords are like this:\n<a href=\"https:\/\/i.stack.imgur.com\/3v661.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3v661.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"53540939","CreationDate":"2018-11-29T14:12:40.150","ParentId":null,"OwnerUserId":"10655065","Title":null,"Body":"<p>OK, I solved it. My god. It's about the pictures themselves. The bit depth of one of the pictures is 32. Then its shape would be (224,224,4).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MtkH2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MtkH2.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53540424,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><reinforcement-learning>","CreationDate":"2018-11-29T13:43:18.747","AcceptedAnswerId":53540809.0,"OwnerUserId":10722768.0,"Title":"How to use Tensorflow tf.nn.Conv2d simultaneously for training and prediction?","Body":"<p>I am currently diving deeper into tensorflow and I am a bit puzzled with the proper use of <code>tf.nn.Conv2d(input, filter, strides, padding)<\/code>. Although it looks simple at first glance I cannot get my hear around the following issue:<\/p>\n\n<p>The use of <code>filter, strides, padding<\/code> is clear to me. However what is not clear is the correct application of the <code>input<\/code>. <\/p>\n\n<p>I am coming from a reinforcement learning Atari (Pong) problem in which I want to use the network for batch training AND (with a certain probability) also for predictions in each step. That means, for training I am feeding the network a full batch of let's say 100 , each single unit consisting of 3 frames with size 160, 128. Using the NHWC format of tensorflow my input to <code>input<\/code> would be a <code>tf.placeholder<\/code> of shape <code>(100,160,128,3)<\/code>. So for training I am feeding 100 160x128x3 packages.<\/p>\n\n<p>However, when predicting outputs from my network (going up or down with the pong paddle) in a certain situation I am only feeding the network <strong>one package<\/strong> of 160x128x3 (i.e. one package of three frames). Now this is where tensorflow crashes. It expects <code>(100,160,128,3)<\/code> but receives <code>(1,160,128,3)<\/code>. <\/p>\n\n<p>Now I am puzzled. I obviously do not want to set the batch size to 1 and always feed only one package for training. But how can I proceed here? How shall this be implemented with <code>tf.nn.conv2d<\/code>? <\/p>\n\n<p>Very much appreciated if someone cann steer me into the right direction here<\/p>\n\n<p>Thank you in advance for your help!\nKevin<\/p>\n","answers":[{"AnswerId":"53540809","CreationDate":"2018-11-29T14:05:24.187","ParentId":null,"OwnerUserId":"10718989","Title":null,"Body":"<p>You need to set your Placeholder like follows\n tf.placeholder(shape=(None,160,128,3) ....) , with having None in the first dimension , your placeholder will be flexible to any value you feed either 1 or 100.<\/p>\n"}]}
+{"QuestionId":53540615,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2018-11-29T13:53:31.237","AcceptedAnswerId":null,"OwnerUserId":4642578.0,"Title":"tensorflow estimator training only runs half of the steps","Body":"<p>I'm having a problem with TF Estimators generated from keras. I've added in a SessionRunHook to display which step is being run. Both evaluate() and predict() both run the correct number of steps, but train() only runs half the number of steps that I tell it to, e.g. I if specify enough steps for 2 epochs, it runs 1. If I specify enough steps for 4 epochs, it runs 2.<\/p>\n\n<p>Here's the relevant code snippets. First my input function. I use the same input function for train, evaluate and predict:<\/p>\n\n<pre><code>def input_function(features, labels=None, shuffle=False, batch_size=batch_size, num_epochs=1):\ninput_fn = tf.estimator.inputs.numpy_input_fn(\n    x={model.input_names[0] : features},\n    y=labels,\n    batch_size=batch_size,\n    num_epochs=num_epochs,\n    shuffle=shuffle)\nreturn input_fn\n<\/code><\/pre>\n\n<p>Here's how I call training:<\/p>\n\n<pre><code>steps = np.ceil((nb_epoch * np.ceil(X_train.shape[0] \/ batch_size).astype(np.int32)) \/ num_gpus).astype(np.int32)\nprint(\"  Steps : \", steps, \"\\n\")\ntime_hist = TimeHistory(num_steps = steps, stage_name='Training')\nest_model.train(input_fn=input_function(X_train, y_train, shuffle=True, \n     num_epochs=nb_epoch), steps=steps, hooks=[time_hist])\n<\/code><\/pre>\n\n<p>time_hist is my SessionRunHook, and basically counts the steps and print progress. So when I run training on my dataset I get:<\/p>\n\n<pre><code>TRAINING MODEL\n\n  Training dataset size :  39209\n  Batch size :  128\n  Epochs :  2\n  Num GPUs :  1\n\n  Steps :  614\n\nBegin  Training ...\nStep :  10 \/ 614\nStep :  20 \/ 614\n...\nStep :  290 \/ 614\nStep :  300 \/ 614\nTraining  finished.\n  total time :  25.3624458313\n  avg batch time :  0.08234560334837282\n  num steps :  308\n<\/code><\/pre>\n\n<p>So it calculated (correctly) that it wold take 614 steps to run 2 epochs of 128 batch_size on the dataset of 39209 elements. I pass steps=614 into train() but it only runs for 308 steps!<\/p>\n\n<p>When I run evaluate, it works fine:<\/p>\n\n<pre><code>eval_steps = np.ceil((np.ceil(X_test.shape[0] \/ batch_size).astype(np.int32)) \/ num_gpus).astype(np.int32)\ntime_hist = TimeHistory(num_steps = eval_steps, stage_name='Testing')\nscore = est_model.evaluate(input_fn=input_function(X_test, y_test, shuffle=False), steps=eval_steps, hooks=[time_hist])\n<\/code><\/pre>\n\n<p>and I get:<\/p>\n\n<pre><code>TESTING MODEL\n   Test dataset size :  12630 \nBegin  Testing ... \nStep :  10 \/ 99 \n... \nStep :  80 \/ 99 \nStep :  90 \/ 99 \nTesting  finished.   total time :  3.7828950882   \navg batch time :  0.03821106149692728   \nnum steps :  99 \nTest score :  {'loss': 0.63948023, 'global_step': 616, 'accuracy': 0.8405659}\n<\/code><\/pre>\n\n<p>Can anyone please see what I'm doing wrong? Driving me nuts! :)<\/p>\n\n<p>Update: after some further debug (see notes below), it seems that Tensorboard tells me it has run 616 steps as expected, but the SessionRunHook has only counted half that. Is it possible that SessionRunHook only runs every second step? Some real confusion here between the input_fn num_epochs, the train steps, and tensorboard global_step. Grrrrrr....<\/p>\n\n<p>Update2 : I added some code to print the global_step_value in my hook, and now it looks like:<\/p>\n\n<pre><code>Begin  Training ...\nStep :  10 \/ 1228\n  global_step :  20\nStep :  20 \/ 1228\n  global_step :  40\nStep :  30 \/ 1228\n  global_step :  60\n...\n<\/code><\/pre>\n\n<p>So, given that global_step is the number of batches seen by the graph, it seems there is some disjoint between it the steps I'm counting in my hook. Looks like the hook is only called every second global_step. Hrmmm<\/p>\n\n<p>Thanks,\nPaul<\/p>\n","answers":[]}
+{"QuestionId":53540822,"AnswerCount":0,"Tags":"<python><tensorflow><coreml><coremltools>","CreationDate":"2018-11-29T14:06:00.727","AcceptedAnswerId":null,"OwnerUserId":4511899.0,"Title":"Converting Tensorflow model to CoreML model. Shape Translator missing for OP Slice","Body":"<p>Im using TFCoreml in python to convert my Tensorflow model into CoreML for development on an iOS device using the CoreML Libs.<\/p>\n\n<p>Im using the following python code to try and convert the model to CoreML.<\/p>\n\n<p>import tfcoreml as tf_converter<\/p>\n\n<pre><code>tf_converter.convert(tf_model_path = 'frozen_inference_graph.pb', \n                     mlmodel_path = 'ml_model.mlmodel', \n                     output_feature_names = ['SemanticPredictions:0'],\n                      input_name_shape_dict =  {'ImageTensor:0' : [1, 512, 512, 3]})\n<\/code><\/pre>\n\n<p>This gives the following error:<\/p>\n\n<blockquote>\n  <p>Shape Translator missing for OP of type Slice.<\/p>\n<\/blockquote>\n\n<p>I read the docs a bit further for TFCoreml and it states that Slice isn't fully supported, and some custom conversion code is required for this to work. In the TFCoreml documentation it suggests breaking the frozen graph into sub graphs and converting them individually then merging them back together post conversion.<\/p>\n\n<p>I updated my code to use custom layers, but I don't really understand how the custom conversion functions work. <\/p>\n\n<p>Just need a few pointers on where to look to begin understanding how to write these custom conversion methods so I can solve my problem with converting the Tensorflow model to CoreML.<\/p>\n\n<p><strong>[EDIT]<\/strong><\/p>\n\n<p>I did some more reading into the TFCoreml examples and documentation and have adapted my solution to this.<\/p>\n\n<pre><code>import tfcoreml as tf_converter\n\ndef _convert_slice(**kwargs):\n    tf_op = kwargs[\"op\"]\n    coreml_nn_builder = kwargs[\"nn_builder\"]\n    constant_inputs = kwargs[\"constant_inputs\"]\n\n    params = NeuralNetwork_pb2.CustomLayerParams()\n    params.className = 'Slice'\n    params.description = \"Custom layer that corresponds to the slice TF op\"\n    # get the value of begin\n    begin = constant_inputs.get(tf_op.inputs[1].name, [0,0,0,0])\n    size = constant_inputs.get(tf_op.inputs[2].name, [0,0,0,0])\n    # add begin and size as two repeated weight fields\n    begin_as_weights = params.weights.add()\n    begin_as_weights.floatValue.extend(map(float, begin))\n    size_as_weights = params.weights.add()\n    size_as_weights.floatValue.extend(map(float, size))\n    coreml_nn_builder.add_custom(name=tf_op.name,\n                                input_names=[tf_op.inputs[0].name],\n                                output_names=[tf_op.outputs[0].name],\n                                custom_proto_spec=params)\n\ncoreml_model = tfcoreml.convert(\n        tf_model_path='frozen_inference_graph.pb',\n        mlmodel_path='my_model.mlmodel',\n        input_name_shape_dict={'ImageTensor:0':[1, 512, 512, 3]},\n        output_feature_names=['SemanticPredictions:0'],\n        add_custom_layers=True,\n        custom_conversion_functions={'Slice': _convert_slice}) # dictionary has op name as the key\n\nprint(\"\\n \\n ML Model layers info: \\n\")\n# inspect the CoreML model: this should be same as the one we got above\nspec = coreml_model.get_spec()\n_print_coreml_nn_layer_info(spec)\n<\/code><\/pre>\n\n<p>I'm still getting the same error as before<\/p>\n\n<blockquote>\n  <p>Shape Translator missing for OP of type Slice.<\/p>\n<\/blockquote>\n\n<p>But i did notice I'm also getting this error\/warning<\/p>\n\n<blockquote>\n  <p>custom_conversion_functions={'Slice': _convert_slice}) # dictionary has op name as the key<\/p>\n<\/blockquote>\n\n<p>Any assistance would be appreciated <\/p>\n","answers":[]}
+{"QuestionId":53541361,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><gpu-programming>","CreationDate":"2018-11-29T14:34:17.070","AcceptedAnswerId":null,"OwnerUserId":8389470.0,"Title":"Tensorflow unique on batches","Body":"<p>I would like to find the size of a set of integers (no repetitions) for many sets. Currently\nI use<\/p>\n\n<pre><code>...\nun_3 = tf.unique(dataset_iter.get_next())\nwith tf.Session() as sess:\n    while True:\n       result = sess.run(un_3, feed_dict={Y: some_other_data})\n       ...save result..\n<\/code><\/pre>\n\n<p>to achieve that. But this works only on 1D tensors. I have two questions:<\/p>\n\n<p>1) Will moving to batches (matrices) of say 100 vectors run the code on parallel on the GPU thus save performance?<\/p>\n\n<p>2) If so then how to use a batch-supported version of tf.unique()?<\/p>\n\n<ul>\n<li>Can I use <em>map_fn<\/em> somehow to run <em>unique()<\/em> on parallel?<\/li>\n<\/ul>\n\n<p>3) If not then how to make sure the above code runs efficiently on a GPU?<\/p>\n\n<p>--------- Edit -----<\/p>\n\n<p>I was able to run <em>tf.unique()<\/em> using<\/p>\n\n<pre><code>@tf.contrib.eager.defun\ndef func(tensors):\n    y, idx = tf.unique(tensors) #, dtype=tf.int32)\n    return tf.size(y)\n\nun_3 = tf.map_fn(func, un_2, infer_shape=False)\n<\/code><\/pre>\n\n<p>But the question remains - will it run parallel on the GPU?<\/p>\n","answers":[{"AnswerId":"53595773","CreationDate":"2018-12-03T14:25:53.447","ParentId":null,"OwnerUserId":"8389470","Title":null,"Body":"<p>The map_fn version is not parallel. Hence I used the following (not so clean) code to handle batches:<\/p>\n\n<pre><code>un_3_00, _ = tf.unique(un_2[0,:])\nun_3_01, _ = tf.unique(un_2[1,:])\nun_3_02, _ = tf.unique(un_2[2,:])\n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":53541797,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><theano><pytorch><recurrent-neural-network>","CreationDate":"2018-11-29T14:57:32.563","AcceptedAnswerId":null,"OwnerUserId":10723131.0,"Title":"Can Theano \/ Pytorch \/ Tensorflow compute the following gradient automatically?","Body":"<p>I am trying to run a recurrent neural network where the state update function for each neuron is the following<\/p>\n\n<pre><code>z = g*y\n<\/code><\/pre>\n\n<p>given that <\/p>\n\n<pre><code>g = (x&lt;x_max &amp; x&gt;x_max-e) | (x&gt;-x_max &amp; x&lt;-x_max+e)\n<\/code><\/pre>\n\n<p>Note that all the variables here are just scalars.<\/p>\n\n<p>The variable x is defined in a way that it will always  update continually so that g will always be a pulse as shown in the <a href=\"https:\/\/i.stack.imgur.com\/8BDYc.png\" rel=\"nofollow noreferrer\">this picture<\/a>. That is, g won't be 1 for a single update but it will be 1 for several consecutive updates.<\/p>\n\n<p>Can any of these packages implement an automatic gradient computation given this transfer function?<\/p>\n","answers":[{"AnswerId":"53544160","CreationDate":"2018-11-29T17:05:41.950","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>The gradient can't be computed. <\/p>\n\n<p><code>g<\/code> as you have shown is a binary variable. So it's gradient can't be computed. Even the wave-form you have plotted has gradient 0 everywhere except at two points (where its infinite, function is discontinuous)<\/p>\n"}]}
+{"QuestionId":53541803,"AnswerCount":1,"Tags":"<python><tensorflow><sparse-matrix>","CreationDate":"2018-11-29T14:57:47.327","AcceptedAnswerId":53609892.0,"OwnerUserId":10596177.0,"Title":"Rows or elements selection on sparse tensor","Body":"<p>In tensorflow, how can we do tf.gather or tf.gather_nd in sparse tensor? How can we extract select specific rows or specific elements from sparse tensor without converting it into dense tensor?<\/p>\n","answers":[{"AnswerId":"53609892","CreationDate":"2018-12-04T09:43:54.573","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>This is one possible solution, although it is still expensive in time and memory so it is probably not feasible for a big use case:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef sparse_select_indices(sp_input, indices, axis=0):\n    # Only necessary if indices may have non-unique elements\n    indices, _ = tf.unique(indices)\n    n_indices = tf.size(indices)\n    # Only necessary if indices may not be sorted\n    indices, _ = tf.math.top_k(indices, n_indices)\n    indices = tf.reverse(indices, [0])\n    # Get indices for the axis\n    idx = sp_input.indices[:, axis]\n    # Find where indices match the selection\n    eq = tf.equal(tf.expand_dims(idx, 1), tf.cast(indices, tf.int64))\n    # Mask for selected values\n    sel = tf.reduce_any(eq, axis=1)\n    # Selected values\n    values_new = tf.boolean_mask(sp_input.values, sel, axis=0)\n    # New index value for selected elements\n    n_indices = tf.cast(n_indices, tf.int64)\n    idx_new = tf.reduce_sum(tf.cast(eq, tf.int64) * tf.range(n_indices), axis=1)\n    idx_new = tf.boolean_mask(idx_new, sel, axis=0)\n    # New full indices tensor\n    indices_new = tf.boolean_mask(sp_input.indices, sel, axis=0)\n    indices_new = tf.concat([indices_new[:, :axis],\n                             tf.expand_dims(idx_new, 1),\n                             indices_new[:, axis + 1:]], axis=1)\n    # New shape\n    shape_new = tf.concat([sp_input.dense_shape[:axis],\n                           [n_indices],\n                           sp_input.dense_shape[axis + 1:]], axis=0)\n    return tf.SparseTensor(indices_new, values_new, shape_new)\n<\/code><\/pre>\n\n<p>Here is an example of use:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    # Input\n    sp1 = tf.SparseTensor([[0, 1], [2, 3], [4, 5]], [10, 20, 30], [6, 7])\n    print(sess.run(tf.sparse.to_dense(sp1)))\n    # [[ 0 10  0  0  0  0  0]\n    #  [ 0  0  0  0  0  0  0]\n    #  [ 0  0  0 20  0  0  0]\n    #  [ 0  0  0  0  0  0  0]\n    #  [ 0  0  0  0  0 30  0]\n    #  [ 0  0  0  0  0  0  0]]\n\n    # Select rows 0, 1, 2\n    sp2 = sparse_select_indices(sp1, [0, 1, 2])\n    print(sess.run(tf.sparse.to_dense(sp2)))\n    # [[ 0 10  0  0  0  0  0]\n    #  [ 0  0  0  0  0  0  0]\n    #  [ 0  0  0 20  0  0  0]]\n\n    # Select columns 4, 5\n    sp3 = sparse_select_indices(sp1, [4, 5], axis=1)\n    print(sess.run(tf.sparse.to_dense(sp3)))\n    # [[ 0  0]\n    #  [ 0  0]\n    #  [ 0  0]\n    #  [ 0  0]\n    #  [ 0 30]\n    #  [ 0  0]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":53541973,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2018-11-29T15:06:46.527","AcceptedAnswerId":53542274.0,"OwnerUserId":3320135.0,"Title":"TensorFlow while loop where the prediction is the next iteration's input","Body":"<p>I am trying to get a while loop working in TensorFlow, but am struggling to understand the <code>body<\/code> parameter of <code>tf.while_loop<\/code>.<\/p>\n\n<p>In my code, I have a neural network defined as follows:<\/p>\n\n<pre><code>input_placeholder = tf.placeholder(dtype=np.float32, shape=[None, 2])\n\nhidden_layer = tf.layers.dense(inputs=input_placeholder, units=10, activation=tf.nn.relu)\nprediction_op = tf.layers.dense(inputs=hidden_layer, units=2)\n<\/code><\/pre>\n\n<p>I then want to run this prediction operation 10 times, feeding the prediction from one iteration as the input into the next iteration. My code for this is as follows:<\/p>\n\n<pre><code>num_steps = 10\ncondition = lambda step_num, input_placeholder: tf.less(step_num, num_steps)\n_, final_prediction = tf.while_loop(condition, body, [step_num, latest_prediction])\n<\/code><\/pre>\n\n<p>What I am confused about is what to use for the <code>body<\/code> parameter. I need the body to take the prediction of the current iteration, and set this as the input for the next iteration.<\/p>\n\n<p>Can I just use <code>prediction_op<\/code> for <code>body<\/code>? If so, how should I tell the while loop to update the value of <code>input_placeholder<\/code> after each call of <code>prediction_op<\/code>? Or should I do this a different way?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"53542274","CreationDate":"2018-11-29T15:23:18.537","ParentId":null,"OwnerUserId":"10658208","Title":null,"Body":"<p>You could have the body definition like this:<\/p>\n\n<pre><code>def body(step_num, latest_prediction):\n    hidden_layer = tf.layers.dense(inputs=latest_prediction, units=10, activation=tf.nn.relu)\n    prediction_op = tf.layers.dense(inputs=hidden_layer, units=2)\n    return step_num+1, prediction_op\n<\/code><\/pre>\n\n<p>and you can modify the <code>while_loop<\/code> as:<\/p>\n\n<pre><code>step_num = 0\ninput_placeholder = tf.placeholder(dtype=np.float32, shape=[None, 2])\ncondition = lambda step_num, input_placeholder: tf.less(step_num, num_steps)\n_, final_prediction = tf.while_loop(condition, body, [step_num, input_placeholder])\n<\/code><\/pre>\n"}]}
+{"QuestionId":53542061,"AnswerCount":1,"Tags":"<optimization><keras>","CreationDate":"2018-11-29T15:11:40.953","AcceptedAnswerId":null,"OwnerUserId":3592007.0,"Title":"Keras report highly inconsistent loss between 2 optimization runs","Body":"<p>I am trying to classify some images if they contain object : Yes or No.<\/p>\n\n<p>I firstly run 50 epochs with rmsprop optimizer and continue with second run of SGD optimizer with more 50 epochs.<\/p>\n\n<p>My first run ends with loss ~ 0.4 and model is saved . Second run starts with the saved model.<\/p>\n\n<p>The problem that at start of second run Keras shows that loss is ~ 0.8 at 1 epoch.<\/p>\n\n<p>Why could it happen for same loss function ?<\/p>\n","answers":[{"AnswerId":"53543343","CreationDate":"2018-11-29T16:19:50.723","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Your new SGD optimizer is not optimized for that model. The moments and the adaptive learning rate were forgotten. <\/p>\n\n<p>Thus, there is indeed a high chance that this new compilation start badly. <\/p>\n\n<p>You can try to restart with lower learning rates, and also try to add moment to the SGD. <\/p>\n"}]}
+{"QuestionId":53542522,"AnswerCount":3,"Tags":"<tensorflow><tensorflow.js>","CreationDate":"2018-11-29T15:36:02.183","AcceptedAnswerId":53545030.0,"OwnerUserId":6256711.0,"Title":"Access to single values of tensor2d in TensorFlow.js","Body":"<p>I'm trying to access the elements of the second row of the 2d matrix using TensorFlow.js package. Here below I access to the first and second element of the first row of the matrix by using <code>[0]<\/code> and <code>[1]<\/code>, but honestly I would expect to access those element like <code>[0][0]<\/code> for get <strong>1<\/strong> and <code>[1][1]<\/code> for get <strong>4<\/strong>. <\/p>\n\n<pre><code>const t1 = tf.tensor2d([\n                         [1,2],\n                         [3,4]\n                         ]);\n\n    const tensorData = t1.dataSync();\n    console.log(\"data[0] is \" + tensorData[0]); \/\/ I get 1\n    console.log(\"data[1] is \" + tensorData[1]); \/\/ I get 2\n    console.log(\"data[0][0] is \" + tensorData[0][0]); \/\/ I get undefined, but I would expect to get 1\n<\/code><\/pre>\n\n<p>Can someone explain me why? And how could I can access to the single elements of the second row of the matrix?<\/p>\n","answers":[{"AnswerId":"53542925","CreationDate":"2018-11-29T15:56:55.673","ParentId":null,"OwnerUserId":"1650811","Title":null,"Body":"<p>Sintax tensorData[0][0] is correct, but looks like your tensorData var is equal to  [1, 2, 3, 4] without second row at all.<\/p>\n\n<p>Just check what tensorData exactly contains before accessing.  <\/p>\n"},{"AnswerId":"53545030","CreationDate":"2018-11-29T18:01:31.423","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>If you want to get the tensor value by coordinate, there is the <code>get<\/code> method. Here is an example<\/p>\n\n<pre><code>const a=tf.tensor([1, 2, 3, 4], [2, 2])\nb=a.get(0,0)\nconsole.log(b)\n<\/code><\/pre>\n\n<p>You can also have a look at the following <a href=\"https:\/\/stackoverflow.com\/questions\/51848367\/how-to-get-the-value-of-the-element-i-j-of-a-tensor\/51848542#51848542\">answer<\/a><\/p>\n"},{"AnswerId":"53709966","CreationDate":"2018-12-10T16:40:34.407","ParentId":null,"OwnerUserId":"551465","Title":null,"Body":"<p>Please do not use <code>Tensor.get<\/code>, this will actually be deprecated when we do a major version update.<\/p>\n\n<p>You should use:<\/p>\n\n<pre><code>const buffer = tensor.buffer();\nconst value = buffer.get(i, j, k);\n<\/code><\/pre>\n\n<p>If you want to index into multiple values, please store the <code>TensorBuffer<\/code> as an object. This will keep all values on the CPU and allow you to index into them quickly.<\/p>\n"}]}
+{"QuestionId":53542974,"AnswerCount":2,"Tags":"<python><deep-learning><computer-vision><conv-neural-network><pytorch>","CreationDate":"2018-11-29T15:59:13.680","AcceptedAnswerId":53549050.0,"OwnerUserId":10720610.0,"Title":"Pytorch: Normalize Image data set","Body":"<p>I want to normalize custom dataset of images. For that i need to compute mean and standard deviation by iterating over the dataset. How can I normalize my entire dataset before creating the data set?<\/p>\n","answers":[{"AnswerId":"53549050","CreationDate":"2018-11-29T23:22:25.113","ParentId":null,"OwnerUserId":"9873654","Title":null,"Body":"<p>Well, let's take this image as an example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LENWe.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LENWe.jpg\" alt=\"Test image: Lena\"><\/a><\/p>\n\n<p>The first thing you need to do is decide which library you want to use: Pillow or OpenCV. In this example I'll use Pillow:<\/p>\n\n<pre><code>from PIL import Image\nimport numpy as np\n\nimg = Image.open(\"test.jpg\")\npix = np.asarray(img.convert(\"RGB\")) # Open the image as RGB\n\nRchan = pix[:,:,0]  # Red color channel\nGchan = pix[:,:,1]  # Green color channel\nBchan = pix[:,:,2]  # Blue color channel\n\nRchan_mean = Rchan.mean()\nGchan_mean = Gchan.mean()\nBchan_mean = Bchan.mean()\n\nRchan_var = Rchan.var()\nGchan_var = Gchan.var()\nBchan_var = Bchan.var()\n<\/code><\/pre>\n\n<p>And the results are:<\/p>\n\n<ul>\n<li>Red Channel Mean: 134.80585625 <\/li>\n<li>Red Channel Variance: 3211.35843945<\/li>\n<li>Green Channel Mean: 81.0884125<\/li>\n<li>Green Channel Variance: 1672.63200823<\/li>\n<li>Blue Channel Mean: 68.1831375<\/li>\n<li>Blue Channel Variance: 1166.20433566<\/li>\n<\/ul>\n\n<p>Hope it helps for your needs.<\/p>\n"},{"AnswerId":"53543608","CreationDate":"2018-11-29T16:34:18.687","ParentId":null,"OwnerUserId":"9873654","Title":null,"Body":"<p>What normalization tries to do is mantain the overall information on your dataset, even when there exists differences in the values, in the case of images it tries to set apart some issues like brightness and contrast that in certain case does not contribute to the general information that the image has. There are several ways to do this, each one with pros and cons, depending on the image set you have and the processing effort you want to do on them, just to name a few:<\/p>\n\n<ul>\n<li>Linear Histogram stetching: where you do a linear map on the current\nrange of values in your image and stetch it to match the 0 and 255\nvalues in RGB <\/li>\n<li>Nonlinear Histogram stetching: Where you use a\nnonlinear function to map the input pixels to a new image. Commonly\nused functions are logarithms and exponentials. My favorite function\nis the cumulative probability function of the original histogram, it\nworks pretty well.<\/li>\n<li>Adaptive Histogram equalization: Where you do a linear\nhistogram stretching in certain places of your image to avoid doing\nan identity mapping where you have the max range of values in your original\nimage.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":53543754,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2018-11-29T16:42:12.813","AcceptedAnswerId":null,"OwnerUserId":1410665.0,"Title":"Expected binary or unicode string, got {'labels': <tf.Tensor 'tower_0\/Reshape_1:0' shape=(5,) dtype=float16>}","Body":"<p>I'm working on the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py#L274\" rel=\"nofollow noreferrer\">CIFAR10<\/a> and using different dataset. I am getting below error due to loss function and label data. I am not getting the idea to transform my labels into 1D or binary so that the issue can be resolved. Any type of help will be highly appreciated. I already checked existing posts but didn't help me. <\/p>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><code>Expected binary or unicode string, got {'labels': &lt;tf.Tensor 'tower_0\/Reshape_1:0' shape=(5,) dtype=float16&gt;}\n<\/code><\/pre>\n\n<p><strong>loss function:<\/strong><\/p>\n\n<pre><code>def loss(logits, labels):\n  \"\"\"Add L2Loss to all the trainable variables.\n  Add summary for \"Loss\" and \"Loss\/avg\".\n  Args:\n    logits: Logits from inference().\n    labels: Labels from distorted_inputs or inputs(). 1-D tensor\n            of shape [batch_size]\n  Returns:\n    Loss tensor of type float.\n  \"\"\"\n  # Calculate the average cross entropy loss across the batch.\n  labels = tf.cast(labels, tf.int64)\n  cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(\n      labels=labels, logits=logits, name='cross_entropy_per_example')\n  cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')\n  tf.add_to_collection('losses', cross_entropy_mean)\n\n  # The total loss is defined as the cross entropy loss plus all of the weight\n  # decay terms (L2 loss).\n  return tf.add_n(tf.get_collection('losses'), name='total_loss')\n<\/code><\/pre>\n\n<p>My labels are like this <a href=\"https:\/\/i.stack.imgur.com\/gX8zt.png\" rel=\"nofollow noreferrer\">Labels<\/a><\/p>\n\n<p><strong>Training code <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py#L82\" rel=\"nofollow noreferrer\">link<\/a><\/strong><\/p>\n\n<pre><code># Build the portion of the Graph calculating the losses. Note that we will\n  # assemble the total_loss using a custom function below.\n  _ = cifar10.loss(logits, labels)\n\n  # Assemble all of the losses for the current tower only.\n  losses = tf.get_collection('losses', scope)\n\n  # Calculate the total loss for the current tower.\n  total_loss = tf.add_n(losses, name='total_loss')\n<\/code><\/pre>\n\n<p>N.B: This may be a very stupid post for you but Its a very serious issue for me. Please avoid Negative voting rather encourage people who are asking such type of silly questions. <\/p>\n","answers":[]}
+{"QuestionId":53543872,"AnswerCount":3,"Tags":"<python><ubuntu><tensorflow><tensorflow-lite>","CreationDate":"2018-11-29T16:48:39.953","AcceptedAnswerId":53750563.0,"OwnerUserId":10266109.0,"Title":"How do i get the input_shape of a frozen Tensorflow-Model for TOCO tf_convert","Body":"<p>I'm trying to convert a frozen model i got from <a href=\"https:\/\/github.com\/davidsandberg\/facenet\" rel=\"nofollow noreferrer\">davidsandberg\/facenet<\/a> to .tflite on Ubuntu 18.04.1 LTS (VirtualBox) using the <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/\" rel=\"nofollow noreferrer\">TF Lite Converter<\/a> <a href=\"https:\/\/drive.google.com\/file\/d\/1EXPBSXwTaqrSC0OhUdXNmKSh9qJUQ55-\/view\" rel=\"nofollow noreferrer\">(this is the specific model i am using)<\/a>.\nWhen i try to run the command:<\/p>\n\n<pre><code>\/home\/nils\/.local\/bin\/tflite_convert \n--output_file=\/home\/nils\/Documents\/frozen.tflite \n--graph_def_file=\/home\/nils\/Documents\/20180402-114759\/20180402-114759.pb \n--input_arrays=input --output_array=embeddings\n<\/code><\/pre>\n\n<p>i get the following error:<\/p>\n\n<pre><code>2018-11-29 16:36:21.774098: I \ntensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports \ninstructions that this TensorFlow binary was not compiled to use: AVX2\nTraceback (most recent call last):\nFile \"\/home\/nils\/.local\/bin\/tflite_convert\", line 11, in &lt;module&gt;\n sys.exit(main())\nFile \n\"\/home\/nils\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib   \/lite\/python\/tflite_convert.py\", \nline 412, in main\n app.run(main=run_main, argv=sys.argv[:1])\n   File \n\"\/home\/nils\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/platform\/app.py\", \nline 125, in run\n _sys.exit(main(argv))\nFile \n\"\/home\/nils\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", \nline 408, in run_main\n _convert_model(tflite_flags)\nFile \n\"\/home\/nils\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/tflite_convert.py\", \nline 162, in _convert_model\n output_data = converter.convert()\nFile \n\"\/home\/nils\/.local\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/lite\/python\/lite.py\", \nline 404, in convert\n \"'{0}'.\".format(_tensor_name(tensor)))\nValueError: Provide an input shape for input array 'input'.\n<\/code><\/pre>\n\n<p>Since i havn't trained the model myself, i don't know what exact shape the input had. Probably one can extract it from the classifier.py and the facenet.py found in David Sandberg's GitHubRep., at facenet\/src but i don't understand the code enough to do this myself.\nI even tried to analyse the graph via tensorboard. I couldn't figure it out anyway, but maybe u can: <a href=\"https:\/\/i.stack.imgur.com\/tG94K.png\" rel=\"nofollow noreferrer\">Tensorboard-Screenshot<\/a>\nAs you might have already noticed, I'm quite new to Ubuntu, Tensorflow and everything related, so I am happy to take any kind of advice on this issue.\nThank you in advance!<\/p>\n\n<p>This is the relevant part of classifier.py, where the model gets loaded and set up:<\/p>\n\n<pre><code> # Load the model\n        print('Loading feature extraction model')\n        facenet.load_model(args.model)\n\n        # Get input and output tensors\n        images_placeholder = tf.get_default_graph().get_tensor_by_name(\"input:0\")\n        embeddings = tf.get_default_graph().get_tensor_by_name(\"embeddings:0\")\n        phase_train_placeholder = tf.get_default_graph().get_tensor_by_name(\"phase_train:0\")\n        embedding_size = embeddings.get_shape()[1]\n\n        # Run forward pass to calculate embeddings\n        print('Calculating features for images')\n        nrof_images = len(paths)\n        nrof_batches_per_epoch = int(math.ceil(1.0*nrof_images \/ args.batch_size))\n        emb_array = np.zeros((nrof_images, embedding_size))\n        for i in range(nrof_batches_per_epoch):\n            start_index = i*args.batch_size\n            end_index = min((i+1)*args.batch_size, nrof_images)\n            paths_batch = paths[start_index:end_index]\n            images = facenet.load_data(paths_batch, False, False, args.image_size)\n            feed_dict = { images_placeholder:images, phase_train_placeholder:False }\n            emb_array[start_index:end_index,:] = sess.run(embeddings, feed_dict=feed_dict)\n\n        classifier_filename_exp = os.path.expanduser(args.classifier_filename)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53750563","CreationDate":"2018-12-12T20:04:06.080","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>If you start Tensorboard again, go back to the graph that you've seen, there should be a search icon (I think at the upperleft corner) where you can type in \"input\" and find the input tensor. It'll give you the desired shape. I'm guessing it'll be something in the form of '[1,image_size, image_size, 3]'.<\/p>\n\n<p>Alternatively, you can examine the code<\/p>\n\n<pre><code>feed_dict = { images_placeholder:images, phase_train_placeholder:False }\n<\/code><\/pre>\n\n<p>Note that we're feeding in the 'images' object into the images_placeholder, which maps to the \"input:0\" tensor. You essentially then want the shape of the images object. <\/p>\n\n<p>images comes from the call to facenet.load_data(). If you go into facenet.py and examine the load_data function, you can observe that the shape is something like what I suggested above. If you print the image_size value, it should match what you see in Tensorboard.<\/p>\n"},{"AnswerId":"53832055","CreationDate":"2018-12-18T11:29:28.887","ParentId":null,"OwnerUserId":"10266109","Title":null,"Body":"<p>Thank you for your help, i did like Alan Chiao said and followed load_data() to the facenet.py, where i eventually found the shape [1,160, 160, 3]. Also, <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/cmdline_reference\" rel=\"nofollow noreferrer\">Tensorflow's command line reference for the tf lite converter<\/a> showed me what i had to look out for: <\/p>\n\n<blockquote>\n  <p>--input_shapes. Type: colon-separated list of comma-separated lists\n  of integers. Each comma-separated list of integers gives the shape of\n  one of the input arrays specified in TensorFlow convention.<\/p>\n  \n  <p>Example: --input_shapes=1,60,80,3 for a typical vision model means a batch size of 1, an input image height of 60, an input image widthof 80, and an input image depth of 3 (representing RGB channels).<\/p>\n<\/blockquote>\n"},{"AnswerId":"54123992","CreationDate":"2019-01-10T07:45:41.270","ParentId":null,"OwnerUserId":"10893881","Title":null,"Body":"<p>I went through the code for tflite converter. I found out that you need to give the input shape as a dictionary in the format <code>{\"input_tensor_name\": [input shape]}<\/code>.<\/p>\n\n<p>Here's an example that resolves the issue:<\/p>\n\n<pre><code>`graph_def_file = \"20180402-114759\/20180402-114759.pb\"\ninput_arrays = [\"input\"]\noutput_arrays = [\"embeddings\"]\n\nconverter = tf.lite.TFLiteConverter.from_frozen_graph(\n  graph_def_file, input_arrays, output_arrays,input_shapes={\"input\":[1,160,160,3]})\n\ntflite_model = converter.convert()\nopen(\"model.tflite\", \"wb\").write(tflite_model)\n`\n<\/code><\/pre>\n"}]}
+{"QuestionId":53544809,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><dataset><data-handling>","CreationDate":"2018-11-29T17:45:49.780","AcceptedAnswerId":53545970.0,"OwnerUserId":4487238.0,"Title":"Tensorflow Dataset using many compressed numpy files","Body":"<p>I have a large dataset that I would like to use for training in Tensorflow. <\/p>\n\n<p>The data is stored in compressed numpy format (using <code>numpy.savez_compressed<\/code>). There are variable numbers of images per file due to the way they are produced.<\/p>\n\n<p>Currently I use a Keras Sequence based generator object to train, but I'd like to move entirely to Tensorflow without Keras. <\/p>\n\n<p>I'm looking at the Dataset API on the TF website, but it is not obvious how I might use this to read numpy data. <\/p>\n\n<p>My first idea was this <\/p>\n\n<pre><code>import glob\nimport tensorflow as tf\nimport numpy as np\n\ndef get_data_from_filename(filename):\n   npdata = np.load(open(filename))\n   return npdata['features'],npdata['labels']\n\n# get files\nfilelist = glob.glob('*.npz')\n\n# create dataset of filenames\nds = tf.data.Dataset.from_tensor_slices(filelist)\nds.flat_map(get_data_from_filename)\n<\/code><\/pre>\n\n<p>However, this passes a TF Tensor placeholder to a real numpy function and numpy is expecting a standard string. This results in the error:<\/p>\n\n<pre><code>File \"test.py\", line 6, in get_data_from_filename\n   npdata = np.load(open(filename))\nTypeError: coercing to Unicode: need string or buffer, Tensor found\n<\/code><\/pre>\n\n<p>The other option I'm considering (but seems messy) is to create a Dataset object built on TF placeholders which I then fill during my epoch-batch loop from my numpy files.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"53545970","CreationDate":"2018-11-29T19:09:42.333","ParentId":null,"OwnerUserId":"10111931","Title":null,"Body":"<p>You can define a wrapper and use pyfunc like this:<\/p>\n\n<pre><code>def get_data_from_filename(filename):\n   npdata = np.load(filename)\n   return npdata['features'], npdata['labels']\n\ndef get_data_wrapper(filename):\n   # Assuming here that both your data and label is float type.\n   features, labels = tf.py_func(\n       get_data_from_filename, [filename], (tf.float32, tf.float32)) \n   return tf.data.Dataset.from_tensor_slices((features, labels))\n\n# Create dataset of filenames.\nds = tf.data.Dataset.from_tensor_slices(filelist)\nds.flat_map(get_data_wrapper)\n<\/code><\/pre>\n\n<p>If your dataset is very large and you have memory issues, you can consider using a combination of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#interleave\" rel=\"nofollow noreferrer\">interleave<\/a> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/data\/parallel_interleave\" rel=\"nofollow noreferrer\">parallel_interleave<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#from_generator\" rel=\"nofollow noreferrer\">from_generator<\/a> methods instead. The from_generator method uses py_func internally so you can directly read your np file and then define your generator in python.<\/p>\n"}]}
+{"QuestionId":53544901,"AnswerCount":1,"Tags":"<python><pytorch><lstm>","CreationDate":"2018-11-29T17:51:50.793","AcceptedAnswerId":53545699.0,"OwnerUserId":7537870.0,"Title":"How to mask weights in PyTorch weight parameters?","Body":"<p>I am attempting to mask (force to zero) specific weight values in PyTorch. The weights I am trying to mask are defined as so in the <code>def __init__<\/code><\/p>\n\n<pre><code>class LSTM_MASK(nn.Module):\n        def __init__(self, options, inp_dim):\n            super(LSTM_MASK, self).__init__()\n            ....\n            self.wfx = nn.Linear(input_dim, curernt_output, bias=add_bias)\n<\/code><\/pre>\n\n<p>The mask is also defined in <code>def __init__<\/code> as<\/p>\n\n<pre><code>self.mask_use = torch.Tensor(curernt_output, input_dim)\n<\/code><\/pre>\n\n<p>The mask is a constant and the <code>.requires_grad_()<\/code> is <code>False<\/code> for the mask parameter. Now in the <code>def forward<\/code> part of the class I attempt to do an element-wise multiplication of the weight parameter and the mask before the linear operation is completed<\/p>\n\n<pre><code>def forward(self, x):\n    ....\n    self.wfx.weight = self.wfx.weight * self.mask_use\n    wfx_out = self.wfx(x)\n<\/code><\/pre>\n\n<p>I get ann error message:<\/p>\n\n<pre><code>self.wfx.weight = self.wfx.weight * self.mask_use\n  File \"\/home\/xyz\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/modules\/module.py\", line 537, in __setattr__\n    .format(torch.typename(value), name))\nTypeError: cannot assign 'torch.cuda.FloatTensor' as parameter 'weight' (torch.nn.Parameter or None expected)\n<\/code><\/pre>\n\n<p>But when I check on the both parameters with <code>.type()<\/code> both of them come up as <code>torch.cuda.FloatTensor<\/code>. I am not sure why there is an error here.<\/p>\n","answers":[{"AnswerId":"53545699","CreationDate":"2018-11-29T18:49:03.173","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>The element-wise operation always returns a <code>FloatTensor<\/code>. It is not possible to assign normal tensors as <code>weight<\/code> of layers. <\/p>\n\n<p>There are two possible options to deal with it. You can assign it to the <code>data<\/code> attribute of your weight, there it is possible assign normal tensors. <\/p>\n\n<p>Or alternatively you convert your result to an <code>nn.Parameter<\/code> itself, then you can assign it to <code>wfx.weight<\/code>.<\/p>\n\n<p>Here is an example which shows both ways:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\n\nwfx = nn.Linear(10, 10)\nmask_use = torch.rand(10, 10)\n#wfx.weight = wfx.weight * mask_use #your example - this raises an error\n\n# Option 1: write directly to data\nwfx.weight.data = wfx.weight * mask_use\n\n# Option 2: convert result to nn.Parameter and write to weight\nwfx.weight = nn.Parameter(wfx.weight * mask_use)\n<\/code><\/pre>\n\n<p><em>Disclaimer: When using an <code>=<\/code> (assignment) on the weights you are replacing the weights tensor of your parameter. This may have unwanted effects on the graph resp. the optimization step.<\/em><\/p>\n"}]}
+{"QuestionId":53545316,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network><deep-learning>","CreationDate":"2018-11-29T18:21:47.653","AcceptedAnswerId":53545578.0,"OwnerUserId":10594993.0,"Title":"Average layer in multi input deep learning","Body":"<p>I am working to create a multi-input Convolutional Neural Network (CNN) model in Keras for Images Classification that takes two images and gives one output which is the class of the two images.<\/p>\n\n<p>I have two datasets: type1 and type2, and each dataset contains the same classes. The model should take one image from Type1 dataset and one image from Type2 dataset and then classify these images to one class (ClassA or ClassB or------).<\/p>\n\n<p><strong>I want to create a model that predict the two images and then compute the average of the prediction similar to the following image:<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7n9en.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7n9en.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>How can I create this model?\nHow can I create the generator in fit_generator?<\/strong><\/p>\n","answers":[{"AnswerId":"53545578","CreationDate":"2018-11-29T18:40:02.590","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h2>Option 1 - Both sides are the same model, just using different inputs<\/h2>\n\n<p>Suppose you have a model that goes up to \"predication\", called <code>predModel<\/code>.<br>\nCreate two input tensors:<\/p>\n\n<pre><code>input1 = Input(shape)   \ninput2 = Input(shape)\n<\/code><\/pre>\n\n<p>Get the outputs for each input:<\/p>\n\n<pre><code>pred1 = predModel(input1)\npred2 = predModel(input2)   \n<\/code><\/pre>\n\n<p>Average the outputs:<\/p>\n\n<pre><code>output = Average()([pred1,pred2])\n<\/code><\/pre>\n\n<p>Create the final model:<\/p>\n\n<pre><code>model = Model([input1,input2], output)\n<\/code><\/pre>\n\n<h2>Option2 - Both sides are similar models, but use different weights<\/h2>\n\n<p>Basically the same as above, but create the layers individually for each side.<\/p>\n\n<pre><code>def createCommonPart(inputTensor):\n    out = ZeroPadding2D(...)(inputTensor)\n    out = Conv2D(...)(out)\n\n    ...\n    out = Flatten()(out)\n    return Dense(...)(out)\n<\/code><\/pre>\n\n<p>Make the two inputs:<\/p>\n\n<pre><code>input1 = Input(shape)   \ninput2 = Input(shape)\n<\/code><\/pre>\n\n<p>Get the two outputs:<\/p>\n\n<pre><code>pred1 = createCommonPart(input1)\npred2 = createCommonPart(input2)\n<\/code><\/pre>\n\n<p>Average the outputs:<\/p>\n\n<pre><code>output = Average()([pred1,pred2])\n<\/code><\/pre>\n\n<p>Create the final model:<\/p>\n\n<pre><code>model = Model([input1,input2], output)\n<\/code><\/pre>\n\n<h2>The generator<\/h2>\n\n<p>Anything that yields <code>[xTrain1,xTrain2], y<\/code>.   <\/p>\n\n<p>You can create one like this:<\/p>\n\n<pre><code>def generator(files1,files2, batch_size):\n\n    while True: #must be infinite\n\n        for i in range(len(files1)\/\/batch_size)):\n            bStart = i*batch_size\n            bEnd = bStart+batch_size\n\n            x1 = loadImagesSomehow(files1[bStart:bEnd])\n            x2 = loadImagesSomehow(files2[bStart:bEnd])\n            y = loadPredictionsSomeHow(forSamples[bStart:bEnd])\n\n            yield [x1,x2], y\n<\/code><\/pre>\n\n<p>You can also implement a <code>keras.utils.Sequence<\/code> in a similar way.<\/p>\n\n<pre><code>class gen(Sequence):\n    def __init__(self, files1, files2, batchSize):\n        self.files1 = files1\n        self.files2 = files2\n        self.batchSize = batchSize\n\n    def __len__(self):\n        return self.len(files1) \/\/ self.batchSize\n\n    def __getitem__(self,i):\n\n        bStart = i*self.batchSize\n        bEnd = bStart+self.batchSize \n\n        x1 = loadImagesSomehow(files1[bStart:bEnd])\n        x2 = loadImagesSomehow(files2[bStart:bEnd])\n        y = loadPredictionsSomeHow(forSamples[bStart:bEnd])\n\n        return [x1,x2], y\n<\/code><\/pre>\n"}]}
+{"QuestionId":53545350,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><neo4j>","CreationDate":"2018-11-29T18:23:53.063","AcceptedAnswerId":53545557.0,"OwnerUserId":5723098.0,"Title":"Do I need a neural network or graph database like neo4j for suggestion engine?","Body":"<p>I am building a simple recommendation\/suggestion engine for a demo application which maintains a list of people. For each person, it keeps track of their food habits with the following preferences:<\/p>\n\n<ol>\n<li>Diet type: Vegetarian\/non-vegetarian\/vegan<\/li>\n<li>Cuisine likings: Indian, Mexican, Italian, etc. (a person can like more than one)<\/li>\n<li>Type of meals with exact time: Breakfast, lunch, supper, dinner<\/li>\n<li>Specialized diets: Keto, Blood-group type, Atkins, etc.<\/li>\n<li>Favorite vegetables: Spinach, Broccoli, etc.<\/li>\n<li>Food allergy requirements<\/li>\n<li>Location - City, area, street, etc.<\/li>\n<\/ol>\n\n<p>Once this data is available to the system, I need to build a simple suggestion engine - <\/p>\n\n<ol>\n<li>For any person chosen, suggest 10 other people who are most compatible in terms of food habits.<\/li>\n<li>For a group of people chosen (say x, max 5), suggest x + 10 (here 15) people from the system in such a way that each person in the chosen group has compatible habit with at least one other person. Order of people in a group doesn't matter.<\/li>\n<\/ol>\n\n<p>My understanding here is that I don't need future predictions on unknown data and thus really no need for machine learning. All I need is suggestions based on statistical compatibility of existing data set. The rules are mostly based on relationships that people have with their food habits.<\/p>\n\n<p>Is my understanding correct? Is this problem entirely solvable with graph database like Neo4j? Or do I really to build a Neural model for this using Tensorflow?<\/p>\n","answers":[{"AnswerId":"53545557","CreationDate":"2018-11-29T18:38:58.100","ParentId":null,"OwnerUserId":"1243762","Title":null,"Body":"<p>The easiest way to know which one to choose is to answer this more specific question. Is the data open world or <a href=\"https:\/\/en.wikipedia.org\/wiki\/Closed-world_assumption\" rel=\"nofollow noreferrer\">closed world<\/a>. <\/p>\n\n<p>If you are an original <a href=\"https:\/\/en.wikipedia.org\/wiki\/The_Corbomite_Maneuver#Plot\" rel=\"nofollow noreferrer\">Star Trek<\/a> fan then think of it like how Spock would inference and needs all of the data before being able to answer (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Closed-world_assumption\" rel=\"nofollow noreferrer\">Closed world<\/a>) or Captain Kirk who doesn't need all of the information before answering (Open World).<\/p>\n\n<p>Seeing that you have all of the data needed before giving a solution I would say that you have a closed world problem and this should use Neo4j. <\/p>\n\n<p>However being a fan of Logic programming, my personal preference would be to use either Prolog, specifically <a href=\"http:\/\/www.swi-prolog.org\/\" rel=\"nofollow noreferrer\">SWI-Prolog<\/a> or <a href=\"https:\/\/www.mercurylang.org\/\" rel=\"nofollow noreferrer\">Mercury<\/a>.<\/p>\n"}]}
+{"QuestionId":53545439,"AnswerCount":1,"Tags":"<linux><database><deep-learning><caffe><lmdb>","CreationDate":"2018-11-29T18:30:25.563","AcceptedAnswerId":null,"OwnerUserId":2379009.0,"Title":"Caffe loading LMDB batches very slowly","Body":"<p>I generated an LMDB database using the SSD-Caffe fork <a href=\"https:\/\/github.com\/chuanqi305\/ssd\" rel=\"nofollow noreferrer\">here<\/a>. I have successfully generated the VOC LMDB trainval\/test LMDB directories and am able to train the model. <\/p>\n\n<p>However, during training, it takes inordinatly long to load data from the LMDB database. For example when profiling using Caffe's <code>time<\/code> function using this command:<\/p>\n\n<p><code>ssdcaffe time --model \"jobs\/VGGNet\/VOC0712\/SSD_300x300\/train.prototxt\" --gpu 0 --iterations 20<\/code><\/p>\n\n<p>I get that the forward pass takes on average 8.9s, and the backward pass takes on average 0.5s. On a layer-by-layer inspection, the data injestion layer takes the bulk of that time at 8.7s. See below:<\/p>\n\n<pre><code>I1129 10:14:11.094445  8011 caffe.cpp:404]       data   forward: 8660.38 ms.\n...\nI1129 10:14:11.095383  8011 caffe.cpp:412] Average Forward pass: 8933.31 ms.\nI1129 10:14:11.095389  8011 caffe.cpp:414] Average Backward pass: 519.549 ms.\n<\/code><\/pre>\n\n<p>If I half the batchsize from 32 to 16, then the data injestion layer time decreases roughly in half:<\/p>\n\n<pre><code>I1129 10:20:07.975527  8093 caffe.cpp:404]       data   forward: 3906.53 ms.\n<\/code><\/pre>\n\n<p>This is clearly not the intended speed, and something is wrong. Any help would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"53547073","CreationDate":"2018-11-29T20:29:10.217","ParentId":null,"OwnerUserId":"2379009","Title":null,"Body":"<p>Found my issue: <\/p>\n\n<p>My images were too big. The standard VOC images which the repo used were ~350x500 pixels, whereas my images were 1080x1920. When I downsized my images by 3x (eg 9x less pixels), my data ingestion layer took only 181ms (a 48x speedup over previous time of 8.6s)<\/p>\n"}]}
+{"QuestionId":53545481,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2018-11-29T18:33:50.177","AcceptedAnswerId":53546612.0,"OwnerUserId":10712670.0,"Title":"\"No Values to save\" when initializing tf.train.Saver","Body":"<p>I have a class DDPG that is a tensorflow neural network. When I add <code>self.saver = tf.train.Saver()<\/code> it reports an error <\/p>\n\n<blockquote>\n  <p>\"  File \"C:\\Users\\Harry\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1131, in _build\n      raise ValueError(\"No variables to save\")\"<\/p>\n<\/blockquote>\n\n<p>Here is the part of my code.<\/p>\n\n<pre><code> def __init__(self, action_dimension, state_dimension):\n    self.memory = np.zeros((MEMORY_CAPACITY, state_dimension*2+action_dimension+1), dtype = np.float32)\n    self.memory_pointer = 0\n    self.sess = tf.Session()\n    self.action_dimension = action_dimension\n    self.state_dimension = state_dimension\n    # define state space as x * state dimension matrix\n    self.current_state = tf.placeholder(tf.float32, [None, state_dimension], 'current_state')\n    self.next_state = tf.placeholder(tf.float32, [None, state_dimension], 'next_state')  # same as above\n    self.reward = tf.placeholder(tf.float32, [None, 1], 'reward')\n    self.saver = tf.train.Saver()\n<\/code><\/pre>\n\n<h1>Here is two functions I declared in this class.<\/h1>\n\n<pre><code>    def save_model(self):\n    path = self.saver.save(self.sess, \"\/saved_model\/model.ckpt\")\n    print(\"Model saved in path: %s\" % path)\n\ndef load_model(self):\n    self.saver.restore(self.sess, \"\/saved_model\/model.ckpt\")\n    print(\"Model restored.\")\n<\/code><\/pre>\n\n<p>I am wondering what I am doing is incorrect? Thanks.<\/p>\n","answers":[{"AnswerId":"53546612","CreationDate":"2018-11-29T19:55:50.167","ParentId":null,"OwnerUserId":"3532564","Title":null,"Body":"<p>Probably because you didn't initialize your variables yet. if you run <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/initializers\/global_variables\" rel=\"nofollow noreferrer\">tf global variables initializer<\/a> after declaring all your variables and before your \n<code>self.saver<\/code> you should be fine<\/p>\n"}]}
+{"QuestionId":53546141,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-11-29T19:21:38.647","AcceptedAnswerId":53554091.0,"OwnerUserId":6636181.0,"Title":"Plot the derivative of a function with PyTorch?","Body":"<p>I have this code:<\/p>\n\n<pre><code>import torch\nimport matplotlib.pyplot as plt  \nx=torch.linspace(-10, 10, 10, requires_grad=True)\ny = torch.sum(x**2)\ny.backward()\nplt.plot(x.detach().numpy(), y.detach().numpy(), label='function')\nplt.legend()\n<\/code><\/pre>\n\n<p>But, I got this error:<\/p>\n\n<pre><code>ValueError: x and y must have same first dimension, but have shapes (10,) and (1,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"53551537","CreationDate":"2018-11-30T05:03:28.550","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>I'm assuming you want to plot the graph of the derivative of <code>x**2<\/code>.<\/p>\n\n<p>Then, you need to plot the graph between <code>x<\/code> and <code>x.grad<\/code> <strong>NOT<\/strong> <code>x<\/code> and <code>y<\/code> i.e.<\/p>\n\n<p><code>plt.plot(x.detach().numpy(), x.grad.detach().numpy(), label='function')<\/code>.<\/p>\n"},{"AnswerId":"53554091","CreationDate":"2018-11-30T08:46:12.083","ParentId":null,"OwnerUserId":"7483494","Title":null,"Body":"<p>I think the main problem is that your dimensions do not match. Why do you wan't to use <code>torch.sum<\/code>?<\/p>\n\n<p>This should work for you:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># %matplotlib inline added this line only for jupiter notebook\nimport torch\nimport matplotlib.pyplot as plt  \nx = torch.linspace(-10, 10, 10, requires_grad=True)\n\ny = x**2      # removed the sum to stay with the same dimensions\ny.backward(x) # handing over the parameter x, as y isn't a scalar anymore\n# your function\nplt.plot(x.detach().numpy(), x.detach().numpy(), label='x**2')\n# gradients\nplt.plot(x.detach().numpy(), x.grad.detach().numpy(), label='grad')\nplt.legend()\n<\/code><\/pre>\n\n<p>Output Graph:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CIF0E.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CIF0E.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You get a nicer picture though with more steps, I also changed the interval a bit to <code>torch.linspace(-2.5, 2.5, 50, requires_grad=True)<\/code>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cuWFf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cuWFf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Edit regarding comment:<\/strong><\/p>\n\n<p>This version plots you the gradients with <code>torch.sum<\/code> included:<\/p>\n\n<pre><code># %matplotlib inline added this line only for jupiter notebook\nimport torch\nimport matplotlib.pyplot as plt  \nx = torch.linspace(-10, 10, 10, requires_grad=True)\n\ny = torch.sum(x**2) \ny.backward() \nprint(x.grad)\nplt.plot(x.detach().numpy(), x.grad.detach().numpy(), label='grad')\nplt.legend()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tensor([-20.0000, -15.5556, -11.1111,  -6.6667,  -2.2222,   2.2222,\n      6.6667,  11.1111,  15.5556,  20.0000])\n<\/code><\/pre>\n\n<p>Plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AfLUK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AfLUK.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":53546276,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2018-11-29T19:32:52.577","AcceptedAnswerId":53566729.0,"OwnerUserId":896678.0,"Title":"CloudML: Check failed: size >= 0","Body":"<p>I came across a strange error happening when I try to train my model using Google CloudML. It doesn't happen when I train the model locally.<\/p>\n\n<pre><code>Check failed: size &gt;= 0 (-1 vs. 0)\n<\/code><\/pre>\n\n<p>In <a href=\"https:\/\/stackoverflow.com\/questions\/45423134\/distributed-tensorflow-check-failed-size-0\">this post<\/a> the problem was solved by changing batch size, which I tried and it didn't help. Any ideas? I am using TF 1.9.0  <\/p>\n","answers":[{"AnswerId":"53566729","CreationDate":"2018-12-01T00:39:10.893","ParentId":null,"OwnerUserId":"896678","Title":null,"Body":"<p>I figured out the problem. One of my training parameters was set to -1 when training was done in cloud. I had to look through all of them because there was no stack trace.<\/p>\n"}]}
+{"QuestionId":53546636,"AnswerCount":2,"Tags":"<python><keras><neural-network><mnist>","CreationDate":"2018-11-29T19:57:22.570","AcceptedAnswerId":null,"OwnerUserId":10072120.0,"Title":"My keras neural network does not predict handwritten numbers","Body":"<p>I trained a feedforward neural network in Keras module, but there are some problems with it. The problem is the incorrect prediction of images with self-written digits. <\/p>\n\n<pre><code>(X_train, y_train), (X_test, y_test) = mnist.load_data()\nRESHAPED = 784\nX_train = X_train.reshape(60000, RESHAPED)\nX_test = X_test.reshape(10000, RESHAPED)\nX_train = X_train.astype('float32')\nX_test = X_test.astype('float32')\nX_train \/= 255\nX_test \/= 255\nY_train = np_utils.to_categorical(y_train, NB_CLASSES)\nY_test = np_utils.to_categorical(y_test, NB_CLASSES)\nmodel = Sequential()\nmodel.add(Dense(N_HIDDEN, input_shape=(RESHAPED,)))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(DROPOUT))\nmodel.add(Dense(N_HIDDEN))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(DROPOUT))\nmodel.add(Dense(NB_CLASSES))\nmodel.add(Activation('softmax'))\nmodel.summary()\nmodel.compile(loss='categorical_crossentropy',\n          optimizer=OPTIMIZER,\n          metrics=['accuracy'])\n\nhistory = model.fit(X_train, Y_train,\n                batch_size=BATCH_SIZE, epochs=NB_EPOCH,\n                verbose=VERBOSE, validation_split=VALIDATION_SPLIT)\nscore = model.evaluate(X_test, Y_test, verbose=VERBOSE) \n\/\/output\n....\nEpoch 20\/20\n48000\/48000 [==============================] - 2s 49us\/step - loss: 0.0713 - \nacc: 0.9795 - val_loss: 0.1004 - val_acc: 0.9782\n10000\/10000 [==============================] - 1s 52us\/step\nTest score: 0.09572517980986121 \nTest accuracy: 0.9781\n<\/code><\/pre>\n\n<p>Next, I upload my image, of dimension 28x28, which was self-written, for example - 2, and executed this script:<\/p>\n\n<pre><code>img_array = imageio.imread('2.png', as_gray=True)\npredictions=model.predict(img_array.reshape(1,784))\nprint (np.argmax(predictions))\n\/\/output for example 3, but I expect - 2\n<\/code><\/pre>\n\n<p>I tried other pictures with different numbers, which also gave wrong predictions. What is wrong? Model shows-Test accuracy: 0.9781. Help me please))) <\/p>\n","answers":[{"AnswerId":"53547216","CreationDate":"2018-11-29T20:39:51.913","ParentId":null,"OwnerUserId":"10718989","Title":null,"Body":"<p>This is a common problem of data mismatch , in order to have a prefect accuracy on your own images , you need to make some preprocessing before , if you noticed one thing in the mnist dataset all images are white in  black background , this is one of the modification you should make to your own image , for more information check this awesome medium article :<\/p>\n\n<p><a href=\"https:\/\/medium.com\/@o.kroeger\/tensorflow-mnist-and-your-own-handwritten-digits-4d1cd32bbab4\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@o.kroeger\/tensorflow-mnist-and-your-own-handwritten-digits-4d1cd32bbab4<\/a><\/p>\n"},{"AnswerId":"53546808","CreationDate":"2018-11-29T20:10:26.713","ParentId":null,"OwnerUserId":"10585322","Title":null,"Body":"<p>You are using the MNIST Dataset. This set was originally created for the american post office. You'll have to deal with <a href=\"https:\/\/en.wikipedia.org\/wiki\/Regional_handwriting_variation\" rel=\"nofollow noreferrer\">regional handwriting variation<\/a> so make sure you have written american numbers.<\/p>\n\n<p>On the other hand errors are normal. Make a bigger set of handwritten digits and make some tests.<\/p>\n"}]}
+{"QuestionId":53547031,"AnswerCount":3,"Tags":"<tensorflow><gcloud><google-cloud-ml><tensorflow-estimator><hyperparameters>","CreationDate":"2018-11-29T20:26:04.497","AcceptedAnswerId":null,"OwnerUserId":2218905.0,"Title":"Hyperparameter tuning locally -- Tensorflow Google Cloud ML Engine","Body":"<p>Is it possible to tune hyperparameters using ML Engine to train the model locally? The documentation only mentions training with hyperparameter tuning in the cloud (submitting a job), and has no mention to doing so locally. <\/p>\n\n<p>Otherwise, is there another commonly used hyperparameter tuning that passes in command arguments to task.py as in the census estimator tutorial?<\/p>\n\n<p><a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census\" rel=\"nofollow noreferrer\">https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census<\/a><\/p>\n","answers":[{"AnswerId":"53761866","CreationDate":"2018-12-13T12:28:39.210","ParentId":null,"OwnerUserId":"7335458","Title":null,"Body":"<p>Check <a href=\"https:\/\/github.com\/LarsHH\/sherpa\" rel=\"nofollow noreferrer\">Sherpa<\/a>, excellent Hyperparameter optimization library.<\/p>\n\n<p>It says:<\/p>\n\n<blockquote>\n  <p>Hyperparameter optimization that enables researchers to experiment, visualize, and scale quickly<\/p>\n<\/blockquote>\n\n<p>There are many HyperParameter optimization libraries out there, but using Sherpa one can visualize the results.<\/p>\n"},{"AnswerId":"53582472","CreationDate":"2018-12-02T16:51:27.097","ParentId":null,"OwnerUserId":"2218905","Title":null,"Body":"<p>As Puneith said, hyperparamater tuning cannot run locally in ML-Engine.<\/p>\n\n<p>SciKit Optimize provides an easy to use wrapper that works with any model including estimators. Just put the code that runs training for N epochs into its own function, which returns the evaluation 1-accuracy, 1-auroc or loss metric for minimizing. <\/p>\n\n<pre><code>import numpy as np\nfrom skopt import gp_minimize\n\ndef train(hyperparam_config):\n    # set from passed in hyperparameters\n    learning_rate = hyperparam_config[0]\n    num_layers = hyperparam_config[2]\n    # run training\n    res = estimator.train_and_evaluate()...\n    return res['loss']  # return metric to minimize\n\nhyperparam_config = [Real(0.0001, 0.01, name=\"learning_rate\"),\n                      Integer(3, 10, name=\"num_layers\")]\nres = gp_minimize(train, hyperparam_config)\nwith open('results.txt', 'w') as wf:\n    wf.write(str(res))\nprint(res)\n<\/code><\/pre>\n\n<p>Source:\n<a href=\"https:\/\/github.com\/scikit-optimize\/scikit-optimize\/blob\/master\/examples\/hyperparameter-optimization.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/scikit-optimize\/scikit-optimize\/blob\/master\/examples\/hyperparameter-optimization.ipynb<\/a><\/p>\n"},{"AnswerId":"53549518","CreationDate":"2018-11-30T00:20:57.277","ParentId":null,"OwnerUserId":"6031363","Title":null,"Body":"<p>You cannot perform HPTuning (<a href=\"https:\/\/cloud.google.com\/blog\/products\/gcp\/hyperparameter-tuning-cloud-machine-learning-engine-using-bayesian-optimization\" rel=\"nofollow noreferrer\">Bayesian Optimization<\/a> based HPTuning which Cloud ML Engine supports) locally, since it's a managed service which Cloud ML Engine offers. There are other ways to perform Hyperparameter tuning e.g., Scikit-learn GridSearch but they are far less effective in this task.<\/p>\n"}]}
+{"QuestionId":53547066,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2018-11-29T20:28:33.290","AcceptedAnswerId":53600419.0,"OwnerUserId":7886651.0,"Title":"What happens when setting reuse=True in tensorflow","Body":"<p>I would like to know what happens in the background when we set the reuse to True in tensorflow. <\/p>\n\n<p>Basically, when building models in tensorflow for training and testing, I have to create the model first within a method and then call that within a variable scope as follows:<\/p>\n\n<pre><code>def model(inputs, return_top=True):\n    #.... Here I have several conv layers\n\n    if return_top:\n        output = tf.layers.dense(output, units=8, name='outputs')\n\n    return output\n\nwith tf.variable_scope('model'):\n    output_train = model(inputs_train)\n    mse_train = cal_loss(output_train, labels_train) # This is a function that calculates the loss\n    train_step = optimize(mse_train)    # This is a function that implements the optimizer\n\nwith tf.variable_scope('model', reuse=True):\n    output_validation = model(inputs_validation)\n    mse_validation = cal_loss(output_validation, labels_validation)\n<\/code><\/pre>\n\n<p>When creating models in tensorflow for training and testing, we usually create one model for training; and let's assume that we give it a name \"model\"; i.e. we created the whole model under a <code>tf.variable_scope(\"model\", reuse=False)<\/code>; and then we reused the model for testing where we set reuse to True. Thus, we use <code>with tf.variable_scope(\"model\", reuse=True)<\/code>. Now if I look into tensorboard, I find two copies for the whole model, one under name \"model\", and the other under \"model_1\". Also, I found that \"model_1\" references \"model\"; i.e., the weights of \"model_1\" are taken from \"model\" (That's my assumption; I would like to know if this is true). Also, I found that the \"model\" outputs goes into the optimizer, which is not the case with \"model_1\". I wonder why. In other words, if \"model_1\" references \"model\"; and the optimizer modifies the weights of \"model\"; should it modify the weights of \"model_1\"?<\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[{"AnswerId":"53552551","CreationDate":"2018-11-30T06:43:10.397","ParentId":null,"OwnerUserId":"10056036","Title":null,"Body":"<p>First, you have namespace conflict for variable_scope. Since variable_scope 'model' is already present, the second variable_scope creation needs to be unique. Tensorflow automatically uniquify it as 'model_1'. You try repeating the definition again, if will create 'model_2' variable_scope.<\/p>\n\n<p>Second, the reuse=True is not for variable_scope names. It is for the tf.Variable inside the tensorflow variable_scope. <\/p>\n\n<p>Suppose you want to use a tf.Variable across 2 layers. In this case, you will use 2 python variables pointing to same TF variable. <\/p>\n\n<p>Without reuse=True, it will throw an error saying something like Variable already exists.\nWith reuse=True, it gives a pass.<\/p>\n"},{"AnswerId":"53600419","CreationDate":"2018-12-03T19:17:12.233","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>First, reuse and variable scopes in general are deprecated and will be removed in tf2. They can be very confusing, as you see here. We instead recommend you use the tf.keras layers to build your model, which you can reuse by just reusing the objects.<\/p>\n\n<p>tf.get_variable and tf.variable_scope together can be used to create and reuse the variables in your model. Inside a variable_scope, once you've called get_variable with a variable name, calling it again with the same variable name is problematic, in that TF cannot tell whether you mean to create a new variable or to reuse the existing one. If you pass reuse=False, the default option, we raise an error. If you pass reuse=True, we give you the same old variable back. However, if you call get_variable with a new variable name and pass reuse=True, we also raise an error since there's no variable to reuse. We also have reuse=tf.AUTO_REUSE which never raises an error (returns a variable if it exists and creates if not).<\/p>\n\n<p>You can also pass reuse as a parameter to variable scopes, which means you'll pass it implicitly to all get_variable calls in that scope.<\/p>\n"}]}
+{"QuestionId":53547304,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2018-11-29T20:45:34.833","AcceptedAnswerId":null,"OwnerUserId":5154274.0,"Title":"Gradients are zero in manual weights update in PyTorch","Body":"<p>I'm trying to implement a simple neural net with manual update of weights for MNIST with AUTOGRAD, similar to AUTOGRAD example given <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/pytorch_with_examples.html\" rel=\"nofollow noreferrer\">here<\/a>. This is my code:<\/p>\n\n<pre><code>import os\nimport sys\n\nimport torch\nimport torchvision\nclass Datasets:\n    \"\"\"Helper for extracting datasets.\"\"\"\n\n    def __init__(self, root='data\/', batch_size=25):\n        if not os.path.exists(root):\n            os.mkdir(root)\n        self.root = root\n        self.batch_size = batch_size\n\n    def get_mnist_loaders(self):\n        train_data = torchvision.datasets.MNIST(\n                root=self.root, train=True, download=True)\n        test_data = torchvision.datasets.MNIST(\n                root=self.root, train=False, download=True)\n\n\n        train_loader = torch.utils.data.DataLoader(\n                dataset=train_data, batch_size=self.batch_size, shuffle=True)\n        test_loader = torch.utils.data.DataLoader(\n                dataset=test_data, batch_size=self.batch_size, shuffle=False)\n\n        return train_loader, test_loader\n\n    def create_batches(self, data, labels, batch_size):\n        return [(data[i:i+batch_size], labels[i:i+batch_size])\n            for i in range(0, len(data), max(1, batch_size))]\n\ndef train1():\n    dtype = torch.float\n    n_inputs = 28*28\n    n_hidden1 = 300\n    n_hidden2 = 100\n    n_outputs = 10\n    batch_size = 200\n    n_epochs = 25\n    learning_rate = 0.01\n    test_step = 100 \n    device = torch.device(\"cpu\")\n\n    datasets = Datasets(batch_size=batch_size)\n    train_loader, test_loader = datasets.get_mnist_loaders()\n\n    def feed_forward(X):\n        x_shape = list(X.size())\n        X = X.view(x_shape[0], x_shape[1]*x_shape[2])\n        hidden1 = torch.mm(X, w1)\n        hidden1 += b1\n        hidden1 = hidden1.clamp(min=0)\n        hidden2 = torch.mm(hidden1, w2) + b2\n        hidden2 = hidden2.clamp(min=0)\n        logits = torch.mm(hidden2, w3) + b3\n        softmax = pytorch_softmax(logits)\n        return softmax\n\n    def accuracy(y_pred, y):\n\n        if list(y_pred.size()) != list(y.size()):\n            raise ValueError('Inputs have different shapes.')\n\n        total_correct = 0\n        total = 0\n        for i, (y1, y2) in enumerate(zip(y_pred, y)):\n            if y1 == y2:\n                total_correct += 1\n            total += 1\n\n        return total_correct \/ total\n\n    w1 = torch.randn(n_inputs, n_hidden1, device=device, dtype=dtype, requires_grad=True)\n    b1 = torch.nn.Parameter(torch.zeros(n_hidden1), requires_grad=True)\n\n    w2 = torch.randn(n_hidden1, n_hidden2, requires_grad=True)\n    b2 = torch.nn.Parameter(torch.zeros(n_hidden2), requires_grad=True)\n\n    w3 = torch.randn(n_hidden2, n_outputs, dtype=dtype, requires_grad=True)\n    b3 = torch.nn.Parameter(torch.zeros(n_outputs), requires_grad=True)\n\n    pytorch_softmax = torch.nn.Softmax(0)\n    pytorch_cross_entropy = torch.nn.CrossEntropyLoss(reduction='elementwise_mean')\n\n    step = 0\n    for epoch in range(n_epochs):\n        batches = datasets.create_batches(train_loader.dataset.train_data,\n                                          train_loader.dataset.train_labels,\n                                          batch_size)\n        for x, y in batches:\n            step += 1\n\n            softmax = feed_forward(x.float())\n            vals, y_pred = torch.max(softmax, 1)\n            accuracy_ = accuracy(y_pred, y)\n            cross_entropy = pytorch_cross_entropy(softmax, y)\n\n            print(epoch, step, cross_entropy.item(), accuracy_)\n\n            cross_entropy.backward()\n\n            with torch.no_grad():\n                w1 -= learning_rate * w1.grad\n                w2 -= learning_rate * w2.grad\n                w3 -= learning_rate * w3.grad\n\n                b1 -= learning_rate * b1.grad\n                b2 -= learning_rate * b2.grad\n                b3 -= learning_rate * b3.grad\n\n                w1.grad.zero_()\n                w2.grad.zero_()\n                w3.grad.zero_()\n\n                b1.grad.zero_()\n                b2.grad.zero_()\n                b3.grad.zero_()\n\nif __name__ == '__main__':\n    train1()\n<\/code><\/pre>\n\n<p>However the network doesn't seem to train. When I print parts of gradients (e.g. <code>w1.grad.data[:10, :10]<\/code>) they consist of zeros. I've tried to use <code>weight.data<\/code> and <code>weight.grad.data<\/code> to update the weights and tried to remove the <code>w.grad.zero_()<\/code> part (even though it is in the example) but it doesn't help. What is the problem here?<\/p>\n","answers":[{"AnswerId":"53570190","CreationDate":"2018-12-01T11:10:29.800","ParentId":null,"OwnerUserId":"4330591","Title":null,"Body":"<p>When you feed the network with Tensor, gradients are not calculating by default. To make it works you may wrap your FloatTensor into <code>torch.autograd.Variable<\/code> or set property <code>requires_grad<\/code> of tensor. <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/examples_autograd\/two_layer_net_autograd.html\" rel=\"nofollow noreferrer\">Here is an example.<\/a><\/p>\n"},{"AnswerId":"58762624","CreationDate":"2019-11-08T07:59:59.753","ParentId":null,"OwnerUserId":"4570466","Title":null,"Body":"<p>There are 3 issues here.<\/p>\n\n<p>Firstly, the axis in which you're taking the softmax is wrong. It should be taken on the last axis.<\/p>\n\n<pre><code>pytorch_softmax = torch.nn.Softmax(-1)\n<\/code><\/pre>\n\n<p>Secondly, your <code>logits<\/code> consists of very large numbers. The resulting derivative of this is a very small number, hence the zeros that you are seeing.<\/p>\n\n<pre><code>tensor([[ -95782.0859,  -30961.9023,   -3614.0188,  ..., -328240.6250,\n          -40818.2227, -160598.5469],\n        [-182128.5938,  -76499.2969,  143654.6250,  ..., -300924.1250,\n          -74291.3125, -109025.0391],\n        [-163018.4062,  -71817.1172, -134466.0156,  ...,  -49884.1211,\n          -19183.3691,  116674.1406],\n         ...,\n        [ 225013.4219,  -37008.6484,  244807.2188,  ..., -466822.8750,\n           63626.5625, -147146.0781],\n        [ 122045.7031,  -90937.7344,   77259.1641,  ..., -397063.9375,\n         -188736.9688,  -78475.5000],\n        [  23139.7578,  -14914.8359, -205065.0625,  ...,  -65808.6562,\n           31458.8906,  -11362.2344]], grad_fn=&lt;AddBackward0&gt;)\n<\/code><\/pre>\n\n<p>Several things you can do to includes normalising your data, adding BatchNorm, clamping etc. I can see that your data <code>X<\/code> is a tensor with values ranging from 0 to 255.<\/p>\n\n<p>Thirdly, you shouldn't need to wrap your tensors with <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/parameter.html\" rel=\"nofollow noreferrer\"><code>nn.Parameter<\/code><\/a> as they are only used in conjunction with the <code>nn.Module<\/code> class.<\/p>\n"}]}
+{"QuestionId":53547711,"AnswerCount":1,"Tags":"<python-3.x><keras><text-classification><keras-layer>","CreationDate":"2018-11-29T21:15:47.697","AcceptedAnswerId":null,"OwnerUserId":8679410.0,"Title":"keras parametes for multilabel text classification","Body":"<p>I am using keras in my multiclass text classifcation, the dataset contains 25000 arabic tweets with 10 class labels\nI use this code :<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(512, input_shape=(10902,)))#10902\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(512))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(10))\nmodel.add(Activation('softmax'))\nmodel.summary()\n #categorical_crossentropy\nmodel.compile(loss='sparse_categorical_crossentropy',        optimizer='rmsprop',\n              metrics=['accuracy'])\n..\nhistory = model.fit(X_train, y_train,\n                    batch_size=100,\n                    epochs=30,\n                    verbose=1,\n                    validation_split=0.5)\n<\/code><\/pre>\n\n<p>Summary:<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ndense_23 (Dense)             (None, 512)               5582336   \n_________________________________________________________________\nactivation_22 (Activation)   (None, 512)               0         \n_________________________________________________________________\ndropout_15 (Dropout)         (None, 512)               0         \n_________________________________________________________________\ndense_24 (Dense)             (None, 512)               262656    \n_________________________________________________________________\nactivation_23 (Activation)   (None, 512)               0         \n_________________________________________________________________\ndropout_16 (Dropout)         (None, 512)               0         \n_________________________________________________________________\ndense_25 (Dense)             (None, 10)                5130      \n_________________________________________________________________\nactivation_24 (Activation)   (None, 10)                0         \n=================================================================\nTotal params: 5,850,122\nTrainable params: 5,850,122\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>but i get error:\n <em>could not convert string to float: 'food'<\/em>\nwhere food is a class name<\/p>\n\n<p>when i change loss to categorical_crossentropy  i get the error\n <em>Error when checking target: expected activation_24 to have shape (10,) but got array with shape (1,)<\/em><\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<pre><code>'\nnd=data.replace(['ads', 'Politic', 'eco', 'food', 'health', 'porno', 'religion', 'sports', 'tech','tv'], \n                     [1, 2, 3, 4, 5,6,7,8,9,10]) \nmodel = Sequential()\n    model.add(Dense(512, input_shape=(10902,10)))#no. of words\n    model.add(Activation('relu'))\n    model.add(Dropout(0.3))\n    model.add(Dense(512))\n    model.add(Activation('relu'))\n    model.add(Dropout(0.3))\n    model.add(Dense(10))\n    model.add(Activation('softmax'))\n    model.summary()\n     #categorical_crossentropy\n    model.compile(loss='categorical_crossentropy',        optimizer='rmsprop',\n                  metrics=['accuracy'])\n    y_train=keras.utils.to_categorical(y_train) \n    history = model.fit(X_train, y_train,\n                        batch_size=100,\n                        epochs=30,\n                        verbose=1,\n                        validation_split=0.5)'\n<\/code><\/pre>\n","answers":[{"AnswerId":"53548887","CreationDate":"2018-11-29T23:04:17.650","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You correctly used <code>Dense(10)<\/code> at the end, in order to produce ten results, one for each class.<\/p>\n\n<p>But you should have your output <code>y_train<\/code> shaped also with 10 classes.    <\/p>\n\n<p>It should have shape <code>(numberOfTweets, 10)<\/code>. <\/p>\n\n<p>For this you should:<\/p>\n\n<ul>\n<li>If you have an array with indices, transform them using the keras function <code>y_train=to_categorical(y_train)<\/code>.    <\/li>\n<li>If you have them as strings, you must transform them in indices, and then use <code>to_categorical<\/code><\/li>\n<\/ul>\n"}]}
+{"QuestionId":53547872,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2018-11-29T21:27:01.407","AcceptedAnswerId":null,"OwnerUserId":10724777.0,"Title":"Creating of custom activation function in keras","Body":"<p>I am trying to create my own custom activation function in keras, which would return 0 if x &lt; 0 and 1 if x >= 0<\/p>\n\n<p><i><\/p>\n\n<pre><code> from keras.layers import Dense\n from keras.models import Sequential\n from keras.layers import Activation\n import tensorflow as tf\n\n\ndef hard_lim(x):\n\n     zero = tf.convert_to_tensor(0., x.dtype.base_dtype)\n\n     one = tf.convert_to_tensor(1., x.dtype.base_dtype)\n\n     sess = tf.Session()\n\n     if sess.run(tf.greater_equal(x, zero)):\n         return one\n     else:\n         return zero\n\n     model = Sequential()\n\n     model.add(Dense(4, input_dim=2, activation=Activation(hard_lim))\n     model.add(Dense(2, activation=Activation(hard_lim))\n     model.add(Dense(1, activation=Activation(hard_lim))\n<\/code><\/pre>\n\n<p><\/i><\/p>\n\n<p>It's giving me this error <\/p>\n\n<p><i><\/p>\n\n<pre><code> InvalidArgumentError (see above for traceback): You must feed a value           \n for placeholder tensor '1_input' with dtype float and shape [?,2]\n<\/code><\/pre>\n\n<p><\/i><\/p>\n\n<p>How can I fix it?<\/p>\n","answers":[{"AnswerId":"53548827","CreationDate":"2018-11-29T22:57:29.240","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<blockquote>\n  <p><strong>Warning<\/strong>: this operation you want has no gradients and will not be trainable. You will see error messages like \"an operation has None for gradient\" or something like \"None type not supported\".<\/p>\n  \n  <p>As a workaround for your activation, I believe the 'relu' activation would be the closest and best option, with the advantage of being very popular and used in most models.   <\/p>\n<\/blockquote>\n\n<p>In Keras, you don't usually run sessions. For custom operations, you create a function using <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">backend<\/a> functions.  <\/p>\n\n<p>So, you'd use a <code>Lambda<\/code> layer:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef hardlim(x):\n   return K.cast(K.greater_equal(x,0), K.floatx())\n<\/code><\/pre>\n\n<p>You can then use <code>activation=hardlim<\/code> in layers.    <\/p>\n"}]}
+{"QuestionId":53547893,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2018-11-29T21:28:55.833","AcceptedAnswerId":null,"OwnerUserId":10461429.0,"Title":"AttributeError: 'module' object has no attribute 'get_compile_flags'","Body":"<p>I try to start the following tensorflow model word2vec example from the official github repository:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/embedding\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/embedding<\/a><\/p>\n\n<p>The data preparation works:<\/p>\n\n<pre><code>curl http:\/\/mattmahoney.net\/dc\/text8.zip &gt; text8.zip\nunzip text8.zip\ncurl https:\/\/storage.googleapis.com\/google-code-archive-source\/v2\/code.google.com\/word2vec\/source-archive.zip &gt; source-archive.zip\nunzip -p source-archive.zip  word2vec\/trunk\/questions-words.txt &gt; questions-words.txt\nrm text8.zip source-archive.zip\n<\/code><\/pre>\n\n<p>But the attempt to add the custom tensorflow operation fails like described in the following:<\/p>\n\n<pre><code>TF_CFLAGS=( $(python -c 'import tensorflow as tf; print(\" \".join(tf.sysconfig.get_compile_flags()))') )\n<\/code><\/pre>\n\n<p>Results in:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\nAttributeError: 'module' object has no attribute 'get_compile_flags'\n<\/code><\/pre>\n\n<p>The newest tensorflow version (v1.12.0) is installed via pip:<\/p>\n\n<pre><code>pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>The python version is v2.7.15.<\/p>\n","answers":[]}
+{"QuestionId":53548368,"AnswerCount":0,"Tags":"<tensorflow><segmentation-fault><nvidia-docker>","CreationDate":"2018-11-29T22:10:47.477","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"How to debug tensorflow in nvidia-docker segfaulting?","Body":"<p>Am on ubuntu 18.04 running in interactive env like this:<\/p>\n\n<pre><code>docker run --runtime=nvidia -it --rm -v $PWD:\/root\/stuff -w \/root tensorflow\/tensorflow:latest-gpu-py3 bash\n<\/code><\/pre>\n\n<p>Curiously, I don't get segfaults when I run non-interactively i.e. docker run ... python stuff\/mnist.py<\/p>\n\n<p>nvidia details:<\/p>\n\n<pre><code>$ nvidia-smi\nThu Nov 29 22:09:25 2018       \n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 415.18       Driver Version: 415.18       CUDA Version: 10.0     |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |\n| Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |\n|===============================+======================+======================|\n|   0  GeForce RTX 2070    Off  | 00000000:01:00.0  On |                  N\/A |\n| 30%   32C    P8    11W \/ 175W |    358MiB \/  7949MiB |      3%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0      1471      G   \/usr\/lib\/xorg\/Xorg                            18MiB |\n|    0      1523      G   \/usr\/bin\/gnome-shell                          50MiB |\n|    0      1919      G   \/usr\/lib\/xorg\/Xorg                           129MiB |\n|    0      2063      G   \/usr\/bin\/gnome-shell                         114MiB |\n|    0      3762      G   ...quest-channel-token=2440404091774701506    43MiB |\n+-----------------------------------------------------------------------------+\n\n\n\nroot@4a46cc9acb73:~# python -X faulthandler -vv stuff\/mnist.py\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/15\n2018-11-29 22:06:26.371579: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2018-11-29 22:06:26.500120: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:964] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2018-11-29 22:06:26.500670: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1432] Found device 0 with properties: \nname: GeForce RTX 2070 major: 7 minor: 5 memoryClockRate(GHz): 1.62\npciBusID: 0000:01:00.0\ntotalMemory: 7.76GiB freeMemory: 7.29GiB\n2018-11-29 22:06:26.500686: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1511] Adding visible gpu devices: 0\n2018-11-29 22:06:26.723360: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:\n2018-11-29 22:06:26.723400: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:988]      0 \n2018-11-29 22:06:26.723407: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1001] 0:   N \n2018-11-29 22:06:26.723859: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 7015 MB memory) -&gt; physical GPU (device: 0, name: GeForce RTX 2070, pci bus id: 0000:01:00.0, compute capability: 7.5)\nFatal Python error: Segmentation fault\n\nThread 0x00007f82a1277700 (most recent call first):\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1439 in __call__\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/backend.py\", line 2986 in __call__\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py\", line 215 in fit_loop\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1639 in fit\n  File \"stuff\/mnist.py\", line 36 in &lt;module&gt;\nSegmentation fault (core dumped)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":53548405,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><batch-normalization>","CreationDate":"2018-11-29T22:14:06.820","AcceptedAnswerId":null,"OwnerUserId":9633432.0,"Title":"Where to apply Batch Normalisation","Body":"<p>I have input data and I am passing it through the relu-hidden layer. I wanted to ask where should I apply Batch Normalisation? Following is the model architecture:<\/p>\n\n<pre><code>def function(self, data):\n    hidden_layer = F.relu((self.layer1(data)))\n    pred = self.pred(hidden_layer)\n    pred = self.softmax(pred)\n    return pred\n<\/code><\/pre>\n\n<p>My best guess is that it should go on the hidden layer like <code>self.batch_norm(hidden_layer)<\/code>. The input data is already scaled using <code>X = StandardScaler().fit_transform(X)<\/code><\/p>\n","answers":[{"AnswerId":"53563381","CreationDate":"2018-11-30T19:02:38.993","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>It's a simple menmonic ... once your data pass through <em>any<\/em> layer, they're no longer in their original form -- and, most important, they're no longer in a <em>batch<\/em>.<\/p>\n\n<p>The only place you have available for an effective batch-norm is just before you enter the first hidden layer.<\/p>\n\n<p>In many (most?) implementations, the observations of a batch are run through the layers in series, one at a time; results are accumulated at the last layer.  \"Batch\" in this case refers most to the back-prop staging: we defer back-prop until we've run all the observations in the batch, and then update per the batch averages.<\/p>\n\n<p>For any layer interface, it's possible to counteract vanishing or exploding weights with normalization -- this may be what's confusing you.  However, it's usually best to (1) not get your fingers in there until you know you need to; (2) do it as part of back-propagation -- again bringing in the \"batch\" synchronization.<\/p>\n"}]}
+{"QuestionId":53548672,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2018-11-29T22:41:05.970","AcceptedAnswerId":53548775.0,"OwnerUserId":9944708.0,"Title":"Understanding Keras LSTM input shape","Body":"<p>I am learning to use Keras LSTM model.  I have looked at <a href=\"https:\/\/machinelearningmastery.com\/multivariate-time-series-forecasting-lstms-keras\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial, <a href=\"https:\/\/machinelearningmastery.com\/use-timesteps-lstm-networks-time-series-forecasting\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial and <a href=\"https:\/\/towardsdatascience.com\/using-lstms-to-forecast-time-series-4ab688386b1f\" rel=\"nofollow noreferrer\">this<\/a> tutorial and am feeling unsure about my understanding of LSTM model's input shape.  My question is if one is to shape one's data like the first tutorial (8760, 1, 8) and the data is inputted to the network 1 timestep at a time i.e. the input_shape=(1, 8) does the network learn the temporal dependencies between samples?  <\/p>\n","answers":[{"AnswerId":"53548775","CreationDate":"2018-11-29T22:51:40.090","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It only makes sense to have batches of 1 timestep when you're using <code>stateful=True<\/code>. Otherwise there is no temporal dependency, as you presumed.<\/p>\n\n<p>The difference is:<\/p>\n\n<ul>\n<li><code>stateful=False<\/code>, <code>input_shape=(1,any)<\/code>:   \n\n<ul>\n<li>first batch of shape (N, 1, any): contains N different sequences of length 1    <\/li>\n<li>second batch: contains another N different sequences of length 1   <\/li>\n<li>total of the two batches: 2N sequences of length 1   <\/li>\n<li>more batches: more independent sequences   <\/li>\n<li>yes, there is no point in using steps=1 when <code>stateful=False<\/code>   <\/li>\n<\/ul><\/li>\n<li><code>stateful=True<\/code>, <code>input_shape=(1,any)<\/code>:    \n\n<ul>\n<li>first batch of shape (N, 1, any): contains the first step of N different sequences   <\/li>\n<li>second batch: contains the second step of the same N sequences   <\/li>\n<li>total of the two batches: N sequences of length 2    <\/li>\n<li>more batches = more steps of the same sequences, until you call <code>model.reset_states()<\/code>   <\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Usually, it's more complicated to handle <code>stateful=True<\/code> layers, and if you can put entire sequences in a batch, like <code>input_shape=(allSteps, any)<\/code>, there is no reason to turn stateful on.    <\/p>\n\n<p>If you want a detailed explanation of RNNs on Keras, <a href=\"https:\/\/stackoverflow.com\/questions\/38714959\/understanding-keras-lstms\/50235563#50235563\">see this answer<\/a><\/p>\n"}]}
+{"QuestionId":53548976,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><deep-learning><tensorflow.js>","CreationDate":"2018-11-29T23:14:11.513","AcceptedAnswerId":54340534.0,"OwnerUserId":6288172.0,"Title":"is there a way to download the frozen graph ( .pb file ) of PoseNet?","Body":"<p>I intend to use posenet in python and not in browser, for that I need the model as a frozen graph to do inference on. Is there a way to do that?<\/p>\n","answers":[{"AnswerId":"53709777","CreationDate":"2018-12-10T16:30:13.430","ParentId":null,"OwnerUserId":"551465","Title":null,"Body":"<p>We currently do not have the frozen graph for inference publicly, however you could download the assets and run them in a Node.js environment.<\/p>\n"},{"AnswerId":"54340534","CreationDate":"2019-01-24T06:26:54.980","ParentId":null,"OwnerUserId":"1519264","Title":null,"Body":"<p>I ported Google's tfjs PoseNet to Python over the holidays. The demo apps in the repository automatically download the weights, freeze a graph, and save to a model file. You can grab this model and use in any TF variant.<\/p>\n\n<p>I wrote a Python version of the multi-person post processing code that uses  vectorized scipy\/numpy ops to speed a few parts. I have not done exhaustive testing of this part, but with a number of spot checks on various test images against the reference, and using it for some other sources, it seems to be reasonably close to the original, and faster :)<\/p>\n\n<p>Python + TF at <a href=\"https:\/\/github.com\/rwightman\/posenet-python\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rwightman\/posenet-python<\/a><\/p>\n\n<p>I also did a PyTorch conversion at <a href=\"https:\/\/github.com\/rwightman\/posenet-pytorch\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rwightman\/posenet-pytorch<\/a><\/p>\n\n<p>And if you happen to be looking for a CoreML port at some point, I started off with the weight conversion code from this project <a href=\"https:\/\/github.com\/infocom-tpo\/PoseNet-CoreML\" rel=\"nofollow noreferrer\">https:\/\/github.com\/infocom-tpo\/PoseNet-CoreML<\/a><\/p>\n"}]}